ft-ops.cc 224 KB
Newer Older
1 2
/* -*- mode: C++; c-basic-offset: 4; indent-tabs-mode: nil -*- */
// vim: ft=cpp:expandtab:ts=8:sw=4:softtabstop=4:
Bradley C. Kuszmaul's avatar
Bradley C. Kuszmaul committed
3
#ident "$Id$"
4
#ident "Copyright (c) 2007-2012 Tokutek Inc.  All rights reserved."
5
#ident "The technology is licensed by the Massachusetts Institute of Technology, Rutgers State University of New Jersey, and the Research Foundation of State University of New York at Stony Brook under United States of America Serial No. 11/760379 and to the patents and/or patent applications resulting from it."
6

7 8
/*

9
Managing the tree shape:  How insertion, deletion, and querying work
10

11
When we insert a message into the FT_HANDLE, here's what happens.
12

13
to insert a message at the root
14

15 16 17 18
    - find the root node
    - capture the next msn of the root node and assign it to the message
    - split the root if it needs to be split
    - insert the message into the root buffer
19
    - if the root is too full, then flush_some_child() of the root on a flusher thread
20

21
flusher functions use an advice struct with provides some functions to
22
call that tell it what to do based on the context of the flush. see ft-flusher.h
23 24 25 26 27

to flush some child, given a parent and some advice
    - pick the child using advice->pick_child()
    - remove that childs buffer from the parent
    - flush the buffer to the child
28
    - if the child has stable reactivity and
29 30 31 32 33
      advice->should_recursively_flush() is true, then
      flush_some_child() of the child
    - otherwise split the child if it needs to be split
    - otherwise maybe merge the child if it needs to be merged

34
flusher threads:
35

36 37 38 39 40 41 42 43 44 45 46 47 48 49
    flusher threads are created on demand as the result of internal nodes
    becoming gorged by insertions. this allows flushing to be done somewhere
    other than the client thread. these work items are enqueued onto
    the cachetable kibbutz and are done in a first in first out order.

cleaner threads:

    the cleaner thread wakes up every so often (say, 1 second) and chooses
    a small number (say, 5) of nodes as candidates for a flush. the one
    with the largest cache pressure is chosen to be flushed. cache pressure
    is a function of the size of the node in the cachetable plus the work done.
    the cleaner thread need not actually do a flush when awoken, so only
    nodes that have sufficient cache pressure are flushed.

50
checkpointing:
51 52 53 54 55 56 57 58 59

    the checkpoint thread wakes up every minute to checkpoint dirty nodes
    to disk. at the time of this writing, nodes during checkpoint are
    locked and cannot be queried or flushed to. a design in which nodes
    are copied before checkpoint is being considered as a way to reduce
    the performance variability caused by a checkpoint locking too
    many nodes and preventing other threads from traversing down the tree,
    for a query or otherwise.

60
To shrink a file: Let X be the size of the reachable data.
61 62 63
    We define an acceptable bloat constant of C.  For example we set C=2 if we are willing to allow the file to be as much as 2X in size.
    The goal is to find the smallest amount of stuff we can move to get the file down to size CX.
    That seems like a difficult problem, so we use the following heuristics:
64 65 66 67 68 69 70 71
       If we can relocate the last block to an lower location, then do so immediately.        (The file gets smaller right away, so even though the new location
         may even not be in the first CX bytes, we are making the file smaller.)
       Otherwise all of the earlier blocks are smaller than the last block (of size L).         So find the smallest region that has L free bytes in it.
         (This can be computed in one pass)
         Move the first allocated block in that region to some location not in the interior of the region.
               (Outside of the region is OK, and reallocating the block at the edge of the region is OK).
            This has the effect of creating a smaller region with at least L free bytes in it.
         Go back to the top (because by now some other block may have been allocated or freed).
72 73 74 75 76 77 78 79 80
    Claim: if there are no other allocations going on concurrently, then this algorithm will shrink the file reasonably efficiently.  By this I mean that
       each block of shrinkage does the smallest amount of work possible.  That doesn't mean that the work overall is minimized.
    Note: If there are other allocations and deallocations going on concurrently, we might never get enough space to move the last block.  But it takes a lot
      of allocations and deallocations to make that happen, and it's probably reasonable for the file not to shrink in this case.

To split or merge a child of a node:
Split_or_merge (node, childnum) {
  If the child needs to be split (it's a leaf with too much stuff or a nonleaf with too much fanout)
    fetch the node and the child into main memory.
81
    split the child, producing two nodes A and B, and also a pivot.   Don't worry if the resulting child is still too big or too small.         Fix it on the next pass.
82
    fixup node to point at the two new children.  Don't worry about the node getting too much fanout.
83
    return;
84 85 86 87
  If the child needs to be merged (it's a leaf with too little stuff (less than 1/4 full) or a nonleaf with too little fanout (less than 1/4)
    fetch node, the child  and a sibling of the child into main memory.
    move all messages from the node to the two children (so that the FIFOs are empty)
    If the two siblings together fit into one node then
88
      merge the two siblings.
89 90 91
      fixup the node to point at one child
    Otherwise
      load balance the content of the two nodes
92
    Don't worry about the resulting children having too many messages or otherwise being too big or too small.        Fix it on the next pass.
93 94
  }
}
95

96 97
Here's how querying works:

98 99
lookups:
    - As of Dr. No, we don't do any tree shaping on lookup.
100
    - We don't promote eagerly or use aggressive promotion or passive-aggressive
101
    promotion.        We just push messages down according to the traditional FT_HANDLE
102 103
    algorithm on insertions.
    - when a node is brought into memory, we apply ancestor messages above it.
104 105 106 107

basement nodes, bulk fetch,  and partial fetch:
    - leaf nodes are comprised of N basement nodes, each of nominal size. when
    a query hits a leaf node. it may require one or more basement nodes to be in memory.
108
    - for point queries, we do not read the entire node into memory. instead,
109 110
      we only read in the required basement node
    - for range queries, cursors may return cursor continue in their callback
111 112 113 114
      to take a the shortcut path until the end of the basement node.
    - for range queries, cursors may prelock a range of keys (with or without a txn).
      the fractal tree will prefetch nodes aggressively until the end of the range.
    - without a prelocked range, range queries behave like successive point queries.
115 116

*/
117

118
#include "checkpoint.h"
119 120 121 122 123 124 125 126 127
#include "ft.h"
#include "ft-cachetable-wrappers.h"
#include "ft-flusher.h"
#include "ft-internal.h"
#include "ft_layout_version.h"
#include "key.h"
#include "log-internal.h"
#include "sub_block.h"
#include "txn_manager.h"
128 129
#include "ule.h"
#include "xids.h"
130

131
#include <toku_race_tools.h>
132
#include <portability/toku_atomic.h>
133 134 135 136 137 138
#include <util/mempool.h>
#include <util/partitioned_counter.h>
#include <util/rwlock.h>
#include <util/sort.h>

#include <stdint.h>
139

140 141 142 143 144 145 146 147 148
#if defined(HAVE_CILK)
#include <cilk/cilk.h>
#define cilk_worker_count (__cilkrts_get_nworkers())
#else
#define cilk_spawn
#define cilk_sync
#define cilk_for for
#define cilk_worker_count 1
#endif
149

150
static const uint32_t this_version = FT_LAYOUT_VERSION;
151

152 153 154
/* Status is intended for display to humans to help understand system behavior.
 * It does not need to be perfectly thread-safe.
 */
155
static FT_STATUS_S ft_status;
156

157 158 159 160 161 162 163 164
#define STATUS_INIT(k,t,l) do {                                         \
        ft_status.status[k].keyname = #k;                               \
        ft_status.status[k].type    = t;                                \
        ft_status.status[k].legend  = "brt: " l;                        \
        if (t == PARCOUNT) {                                            \
            ft_status.status[k].value.parcount = create_partitioned_counter(); \
        }                                                               \
    } while (0)
165

166 167
static toku_mutex_t ft_open_close_lock;

168 169 170 171 172
static void
status_init(void)
{
    // Note, this function initializes the keyname, type, and legend fields.
    // Value fields are initialized to zero by compiler.
173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216
    STATUS_INIT(FT_UPDATES,                                PARCOUNT, "dictionary updates");
    STATUS_INIT(FT_UPDATES_BROADCAST,                      PARCOUNT, "dictionary broadcast updates");
    STATUS_INIT(FT_DESCRIPTOR_SET,                         PARCOUNT, "descriptor set");
    STATUS_INIT(FT_PARTIAL_EVICTIONS_NONLEAF,              PARCOUNT, "nonleaf node partial evictions");
    STATUS_INIT(FT_PARTIAL_EVICTIONS_LEAF,                 PARCOUNT, "leaf node partial evictions");
    STATUS_INIT(FT_MSN_DISCARDS,                           PARCOUNT, "messages ignored by leaf due to msn");
    //STATUS_INIT(FT_MAX_WORKDONE,                           UINT64, "max workdone over all buffers");
    STATUS_INIT(FT_TOTAL_RETRIES,                          PARCOUNT, "total search retries due to TRY_AGAIN");
    //STATUS_INIT(FT_MAX_SEARCH_EXCESS_RETRIES,              UINT64, "max excess search retries (retries - tree height) due to TRY_AGAIN");
    STATUS_INIT(FT_SEARCH_TRIES_GT_HEIGHT,                 PARCOUNT, "searches requiring more tries than the height of the tree");
    STATUS_INIT(FT_SEARCH_TRIES_GT_HEIGHTPLUS3,            PARCOUNT, "searches requiring more tries than the height of the tree plus three");
    STATUS_INIT(FT_DISK_FLUSH_LEAF,                        PARCOUNT, "leaf nodes flushed to disk (not for checkpoint)");
    STATUS_INIT(FT_DISK_FLUSH_NONLEAF,                     PARCOUNT, "nonleaf nodes flushed to disk (not for checkpoint)");
    STATUS_INIT(FT_DISK_FLUSH_LEAF_FOR_CHECKPOINT,         PARCOUNT, "leaf nodes flushed to disk (for checkpoint)");
    STATUS_INIT(FT_DISK_FLUSH_NONLEAF_FOR_CHECKPOINT,      PARCOUNT, "nonleaf nodes flushed to disk (for checkpoint)");
    STATUS_INIT(FT_CREATE_LEAF,                            PARCOUNT, "leaf nodes created");
    STATUS_INIT(FT_CREATE_NONLEAF,                         PARCOUNT, "nonleaf nodes created");
    STATUS_INIT(FT_DESTROY_LEAF,                           PARCOUNT, "leaf nodes destroyed");
    STATUS_INIT(FT_DESTROY_NONLEAF,                        PARCOUNT, "nonleaf nodes destroyed");
    STATUS_INIT(FT_MSG_BYTES_IN,                           PARCOUNT, "bytes of messages injected at root (all trees)");
    STATUS_INIT(FT_MSG_BYTES_OUT,                          PARCOUNT, "bytes of messages flushed from h1 nodes to leaves");
    STATUS_INIT(FT_MSG_BYTES_CURR,                         PARCOUNT, "bytes of messages currently in trees (estimate)");
    //STATUS_INIT(FT_MSG_BYTES_MAX,                          UINT64, "max bytes of messages ever in trees (estimate)");
    STATUS_INIT(FT_MSG_NUM,                                PARCOUNT, "messages injected at root");
    STATUS_INIT(FT_MSG_NUM_BROADCAST,                      PARCOUNT, "broadcast messages injected at root");
    STATUS_INIT(FT_NUM_BASEMENTS_DECOMPRESSED_NORMAL,      PARCOUNT, "basements decompressed as a target of a query");
    STATUS_INIT(FT_NUM_BASEMENTS_DECOMPRESSED_AGGRESSIVE,  PARCOUNT, "basements decompressed for prelocked range");
    STATUS_INIT(FT_NUM_BASEMENTS_DECOMPRESSED_PREFETCH,    PARCOUNT, "basements decompressed for prefetch");
    STATUS_INIT(FT_NUM_BASEMENTS_DECOMPRESSED_WRITE,       PARCOUNT, "basements decompressed for write");
    STATUS_INIT(FT_NUM_MSG_BUFFER_DECOMPRESSED_NORMAL,     PARCOUNT, "buffers decompressed as a target of a query");
    STATUS_INIT(FT_NUM_MSG_BUFFER_DECOMPRESSED_AGGRESSIVE, PARCOUNT, "buffers decompressed for prelocked range");
    STATUS_INIT(FT_NUM_MSG_BUFFER_DECOMPRESSED_PREFETCH,   PARCOUNT, "buffers decompressed for prefetch");
    STATUS_INIT(FT_NUM_MSG_BUFFER_DECOMPRESSED_WRITE,      PARCOUNT, "buffers decompressed for write");
    STATUS_INIT(FT_NUM_PIVOTS_FETCHED_QUERY,               PARCOUNT, "pivots fetched for query");
    STATUS_INIT(FT_NUM_PIVOTS_FETCHED_PREFETCH,            PARCOUNT, "pivots fetched for prefetch");
    STATUS_INIT(FT_NUM_PIVOTS_FETCHED_WRITE,               PARCOUNT, "pivots fetched for write");
    STATUS_INIT(FT_NUM_BASEMENTS_FETCHED_NORMAL,           PARCOUNT, "basements fetched as a target of a query");
    STATUS_INIT(FT_NUM_BASEMENTS_FETCHED_AGGRESSIVE,       PARCOUNT, "basements fetched for prelocked range");
    STATUS_INIT(FT_NUM_BASEMENTS_FETCHED_PREFETCH,         PARCOUNT, "basements fetched for prefetch");
    STATUS_INIT(FT_NUM_BASEMENTS_FETCHED_WRITE,            PARCOUNT, "basements fetched for write");
    STATUS_INIT(FT_NUM_MSG_BUFFER_FETCHED_NORMAL,          PARCOUNT, "buffers fetched as a target of a query");
    STATUS_INIT(FT_NUM_MSG_BUFFER_FETCHED_AGGRESSIVE,      PARCOUNT, "buffers fetched for prelocked range");
    STATUS_INIT(FT_NUM_MSG_BUFFER_FETCHED_PREFETCH,        PARCOUNT, "buffers fetched for prefetch");
    STATUS_INIT(FT_NUM_MSG_BUFFER_FETCHED_WRITE,           PARCOUNT, "buffers fetched for write");
217

218 219 220 221 222 223 224 225 226 227 228 229 230 231
    STATUS_INIT(FT_PRO_NUM_ROOT_SPLIT,                     PARCOUNT, "promotion: roots split");
    STATUS_INIT(FT_PRO_NUM_ROOT_H0_INJECT,                 PARCOUNT, "promotion: leaf roots injected into");
    STATUS_INIT(FT_PRO_NUM_ROOT_H1_INJECT,                 PARCOUNT, "promotion: h1 roots injected into");
    STATUS_INIT(FT_PRO_NUM_INJECT_DEPTH_0,                 PARCOUNT, "promotion: injections at depth 0");
    STATUS_INIT(FT_PRO_NUM_INJECT_DEPTH_1,                 PARCOUNT, "promotion: injections at depth 1");
    STATUS_INIT(FT_PRO_NUM_INJECT_DEPTH_2,                 PARCOUNT, "promotion: injections at depth 2");
    STATUS_INIT(FT_PRO_NUM_INJECT_DEPTH_3,                 PARCOUNT, "promotion: injections at depth 3");
    STATUS_INIT(FT_PRO_NUM_INJECT_DEPTH_GT3,               PARCOUNT, "promotion: injections lower than depth 3");
    STATUS_INIT(FT_PRO_NUM_STOP_NONEMPTY_BUF,              PARCOUNT, "promotion: stopped because of a nonempty buffer");
    STATUS_INIT(FT_PRO_NUM_STOP_H1,                        PARCOUNT, "promotion: stopped at height 1");
    STATUS_INIT(FT_PRO_NUM_STOP_LOCK_CHILD,                PARCOUNT, "promotion: stopped because the child was locked or not at all in memory");
    STATUS_INIT(FT_PRO_NUM_STOP_CHILD_INMEM,               PARCOUNT, "promotion: stopped because the child was not fully in memory");
    STATUS_INIT(FT_PRO_NUM_DIDNT_WANT_PROMOTE,             PARCOUNT, "promotion: stopped anyway, after locking the child");

232
    ft_status.initialized = true;
233
}
234 235 236 237 238 239 240
static void status_destroy(void) {
    for (int i = 0; i < FT_STATUS_NUM_ROWS; ++i) {
        if (ft_status.status[i].type == PARCOUNT) {
            destroy_partitioned_counter(ft_status.status[i].value.parcount);
        }
    }
}
241 242 243
#undef STATUS_INIT

void
244 245
toku_ft_get_status(FT_STATUS s) {
    *s = ft_status;
246 247
}

248
#define STATUS_INC(x, d) increment_partitioned_counter(ft_status.status[x].value.parcount, d)
249

250
bool is_entire_node_in_memory(FTNODE node) {
251 252 253 254 255 256 257
    for (int i = 0; i < node->n_children; i++) {
        if(BP_STATE(node,i) != PT_AVAIL) {
            return false;
        }
    }
    return true;
}
258

259
void
260 261
toku_assert_entire_node_in_memory(FTNODE UU() node) {
    paranoid_invariant(is_entire_node_in_memory(node));
262 263
}

Yoni Fogel's avatar
Yoni Fogel committed
264
static uint32_t
265
get_leaf_num_entries(FTNODE node) {
Yoni Fogel's avatar
Yoni Fogel committed
266
    uint32_t result = 0;
267
    int i;
268
    toku_assert_entire_node_in_memory(node);
269
    for ( i = 0; i < node->n_children; i++) {
270
        result += toku_omt_size(BLB_BUFFER(node, i));
271 272 273 274
    }
    return result;
}

275
static enum reactivity
Zardosht Kasheff's avatar
Zardosht Kasheff committed
276
get_leaf_reactivity (FTNODE node, uint32_t nodesize) {
277
    enum reactivity re = RE_STABLE;
Leif Walsh's avatar
Leif Walsh committed
278
    toku_assert_entire_node_in_memory(node);
279
    paranoid_invariant(node->height==0);
280
    unsigned int size = toku_serialize_ftnode_size(node);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
281
    if (size > nodesize && get_leaf_num_entries(node) > 1) {
Leif Walsh's avatar
Leif Walsh committed
282 283
        re = RE_FISSIBLE;
    }
Zardosht Kasheff's avatar
Zardosht Kasheff committed
284
    else if ((size*4) < nodesize && !BLB_SEQINSERT(node, node->n_children-1)) {
Leif Walsh's avatar
Leif Walsh committed
285
        re = RE_FUSIBLE;
286 287
    }
    return re;
288 289
}

290
enum reactivity
291
get_nonleaf_reactivity (FTNODE node) {
292
    paranoid_invariant(node->height>0);
293
    int n_children = node->n_children;
294 295 296 297 298
    if (n_children > TREE_FANOUT) return RE_FISSIBLE;
    if (n_children*4 < TREE_FANOUT) return RE_FUSIBLE;
    return RE_STABLE;
}

299
enum reactivity
Zardosht Kasheff's avatar
Zardosht Kasheff committed
300
get_node_reactivity (FTNODE node, uint32_t nodesize) {
301
    toku_assert_entire_node_in_memory(node);
302
    if (node->height==0)
Zardosht Kasheff's avatar
Zardosht Kasheff committed
303
        return get_leaf_reactivity(node, nodesize);
304
    else
305
        return get_nonleaf_reactivity(node);
306 307
}

308 309 310
unsigned int
toku_bnc_nbytesinbuf(NONLEAF_CHILDINFO bnc)
{
311
    return toku_fifo_buffer_size_in_use(bnc->buffer);
312 313
}

Yoni Fogel's avatar
Yoni Fogel committed
314
// return true if the size of the buffers plus the amount of work done is large enough.   (But return false if there is nothing to be flushed (the buffers empty)).
315
bool
Zardosht Kasheff's avatar
Zardosht Kasheff committed
316
toku_ft_nonleaf_is_gorged (FTNODE node, uint32_t nodesize) {
Yoni Fogel's avatar
Yoni Fogel committed
317
    uint64_t size = toku_serialize_ftnode_size(node);
318

Yoni Fogel's avatar
Yoni Fogel committed
319
    bool buffers_are_empty = true;
320
    toku_assert_entire_node_in_memory(node);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
321 322 323 324
    //
    // the nonleaf node is gorged if the following holds true:
    //  - the buffers are non-empty
    //  - the total workdone by the buffers PLUS the size of the buffers
Zardosht Kasheff's avatar
Zardosht Kasheff committed
325
    //     is greater than nodesize (which as of Maxwell should be
Zardosht Kasheff's avatar
Zardosht Kasheff committed
326 327
    //     4MB)
    //
328
    paranoid_invariant(node->height > 0);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
329 330 331
    for (int child = 0; child < node->n_children; ++child) {
        size += BP_WORKDONE(node, child);
    }
332
    for (int child = 0; child < node->n_children; ++child) {
333
        if (toku_bnc_nbytesinbuf(BNC(node, child)) > 0) {
Yoni Fogel's avatar
Yoni Fogel committed
334
            buffers_are_empty = false;
335 336 337
            break;
        }
    }
Zardosht Kasheff's avatar
Zardosht Kasheff committed
338
    return ((size > nodesize)
339 340
            &&
            (!buffers_are_empty));
341 342
}

343 344
static void ft_verify_flags(FT UU(ft), FTNODE UU(node)) {
    paranoid_invariant(ft->h->flags == node->flags);
345
}
346

347
int toku_ft_debug_mode = 0;
348

Yoni Fogel's avatar
Yoni Fogel committed
349
uint32_t compute_child_fullhash (CACHEFILE cf, FTNODE node, int childnum) {
350 351
    paranoid_invariant(node->height>0);
    paranoid_invariant(childnum<node->n_children);
352
    return toku_cachetable_hash(cf, BP_BLOCKNUM(node, childnum));
353 354
}

355 356 357 358 359 360
int
toku_bnc_n_entries(NONLEAF_CHILDINFO bnc)
{
    return toku_fifo_n_entries(bnc->buffer);
}

361
static const DBT *prepivotkey (FTNODE node, int childnum, const DBT * const lower_bound_exclusive) {
362
    if (childnum==0)
363
        return lower_bound_exclusive;
364
    else {
365
        return &node->childkeys[childnum-1];
366 367 368
    }
}

369
static const DBT *postpivotkey (FTNODE node, int childnum, const DBT * const upper_bound_inclusive) {
370
    if (childnum+1 == node->n_children)
371
        return upper_bound_inclusive;
372
    else {
373
        return &node->childkeys[childnum];
374 375
    }
}
376
static struct pivot_bounds next_pivot_keys (FTNODE node, int childnum, struct pivot_bounds const * const old_pb) {
377
    struct pivot_bounds pb = {.lower_bound_exclusive = prepivotkey(node, childnum, old_pb->lower_bound_exclusive),
378
                              .upper_bound_inclusive = postpivotkey(node, childnum, old_pb->upper_bound_inclusive)};
379 380
    return pb;
}
381

382
// how much memory does this child buffer consume?
383
long
384
toku_bnc_memory_size(NONLEAF_CHILDINFO bnc)
385 386 387
{
    return (sizeof(*bnc) +
            toku_fifo_memory_footprint(bnc->buffer) +
388 389 390
            bnc->fresh_message_tree.memory_size() +
            bnc->stale_message_tree.memory_size() +
            bnc->broadcast_list.memory_size());
391 392 393 394 395 396
}

// how much memory in this child buffer holds useful data?
// originally created solely for use by test program(s).
long
toku_bnc_memory_used(NONLEAF_CHILDINFO bnc)
397 398
{
    return (sizeof(*bnc) +
399
            toku_fifo_memory_size_in_use(bnc->buffer) +
400 401 402
            bnc->fresh_message_tree.memory_size() +
            bnc->stale_message_tree.memory_size() +
            bnc->broadcast_list.memory_size());
403 404
}

Zardosht Kasheff's avatar
Zardosht Kasheff committed
405
static long
406
get_avail_internal_node_partition_size(FTNODE node, int i)
Zardosht Kasheff's avatar
Zardosht Kasheff committed
407
{
408
    paranoid_invariant(node->height > 0);
409
    return toku_bnc_memory_size(BNC(node, i));
Zardosht Kasheff's avatar
Zardosht Kasheff committed
410 411
}

Zardosht Kasheff's avatar
Zardosht Kasheff committed
412

413
static long
414
ftnode_cachepressure_size(FTNODE node)
Zardosht Kasheff's avatar
Zardosht Kasheff committed
415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433
{
    long retval = 0;
    bool totally_empty = true;
    if (node->height == 0) {
        goto exit;
    }
    else {
        for (int i = 0; i < node->n_children; i++) {
            if (BP_STATE(node,i) == PT_INVALID || BP_STATE(node,i) == PT_ON_DISK) {
                continue;
            }
            else if (BP_STATE(node,i) == PT_COMPRESSED) {
                SUB_BLOCK sb = BSB(node, i);
                totally_empty = false;
                retval += sb->compressed_size;
            }
            else if (BP_STATE(node,i) == PT_AVAIL) {
                totally_empty = totally_empty && (toku_bnc_n_entries(BNC(node, i)) == 0);
                retval += get_avail_internal_node_partition_size(node, i);
434
                retval += BP_WORKDONE(node, i);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
435 436
            }
            else {
437
                abort();
Zardosht Kasheff's avatar
Zardosht Kasheff committed
438 439 440 441 442 443 444 445 446 447
            }
        }
    }
exit:
    if (totally_empty) {
        return 0;
    }
    return retval;
}

Zardosht Kasheff's avatar
Zardosht Kasheff committed
448
long
449
ftnode_memory_size (FTNODE node)
450
// Effect: Estimate how much main memory a node requires.
451
{
452 453 454
    long retval = 0;
    int n_children = node->n_children;
    retval += sizeof(*node);
455
    retval += (n_children)*(sizeof(node->bp[0]));
456
    retval += node->totalchildkeylens;
457 458 459 460 461 462 463

    // now calculate the sizes of the partitions
    for (int i = 0; i < n_children; i++) {
        if (BP_STATE(node,i) == PT_INVALID || BP_STATE(node,i) == PT_ON_DISK) {
            continue;
        }
        else if (BP_STATE(node,i) == PT_COMPRESSED) {
464
            SUB_BLOCK sb = BSB(node, i);
465 466 467 468 469
            retval += sizeof(*sb);
            retval += sb->compressed_size;
        }
        else if (BP_STATE(node,i) == PT_AVAIL) {
            if (node->height > 0) {
Zardosht Kasheff's avatar
Zardosht Kasheff committed
470
                retval += get_avail_internal_node_partition_size(node, i);
471 472
            }
            else {
473
                BASEMENTNODE bn = BLB(node, i);
474
                retval += sizeof(*bn);
475 476 477 478 479 480 481
                {
                    // include fragmentation overhead but do not include space in the
                    // mempool that has not yet been allocated for leaf entries
                    size_t poolsize = toku_mempool_footprint(&bn->buffer_mempool);
                    invariant (poolsize >= BLB_NBYTESINBUF(node,i));
                    retval += poolsize;
                }
482 483 484 485 486
                OMT curr_omt = BLB_BUFFER(node, i);
                retval += (toku_omt_memory_size(curr_omt));
            }
        }
        else {
487
            abort();
488
        }
489
    }
490
    return retval;
491
}
Bradley C. Kuszmaul's avatar
Rename  
Bradley C. Kuszmaul committed
492

493 494 495
PAIR_ATTR make_ftnode_pair_attr(FTNODE node) {
    long size = ftnode_memory_size(node);
    long cachepressure_size = ftnode_cachepressure_size(node);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
496
    PAIR_ATTR result={
497 498 499 500 501
        .size = size,
        .nonleaf_size = (node->height > 0) ? size : 0,
        .leaf_size = (node->height > 0) ? 0 : size,
        .rollback_size = 0,
        .cache_pressure_size = cachepressure_size,
Yoni Fogel's avatar
Yoni Fogel committed
502
        .is_valid = true
503 504
    };
    return result;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
505 506
}

507 508
PAIR_ATTR make_invalid_pair_attr(void) {
    PAIR_ATTR result={
509 510 511 512 513
        .size = 0,
        .nonleaf_size = 0,
        .leaf_size = 0,
        .rollback_size = 0,
        .cache_pressure_size = 0,
Yoni Fogel's avatar
Yoni Fogel committed
514
        .is_valid = false
515 516
    };
    return result;
517
}
Zardosht Kasheff's avatar
Zardosht Kasheff committed
518 519


520
// assign unique dictionary id
521
static uint64_t dict_id_serial = 1;
522 523
static DICTIONARY_ID
next_dict_id(void) {
524
    uint64_t i = toku_sync_fetch_and_add(&dict_id_serial, 1);
525
    assert(i);        // guarantee unique dictionary id by asserting 64-bit counter never wraps
526 527 528 529
    DICTIONARY_ID d = {.dictid = i};
    return d;
}

530
//
531
// Given a bfe and a childnum, returns whether the query that constructed the bfe
532 533 534
// wants the child available.
// Requires: bfe->child_to_read to have been set
//
535
bool
536
toku_bfe_wants_child_available (struct ftnode_fetch_extra* bfe, int childnum)
537
{
538 539
    if (bfe->type == ftnode_fetch_all ||
        (bfe->type == ftnode_fetch_subset && bfe->child_to_read == childnum))
540
    {
541
        return true;
542 543
    }
    else {
544
        return false;
545 546 547
    }
}

548
int
549
toku_bfe_leftmost_child_wanted(struct ftnode_fetch_extra *bfe, FTNODE node)
550
{
551
    paranoid_invariant(bfe->type == ftnode_fetch_subset || bfe->type == ftnode_fetch_prefetch);
552 553 554 555 556
    if (bfe->left_is_neg_infty) {
        return 0;
    } else if (bfe->range_lock_left_key == NULL) {
        return -1;
    } else {
557
        return toku_ftnode_which_child(node, bfe->range_lock_left_key, &bfe->h->cmp_descriptor, bfe->h->compare_fun);
558 559 560 561
    }
}

int
562
toku_bfe_rightmost_child_wanted(struct ftnode_fetch_extra *bfe, FTNODE node)
563
{
564
    paranoid_invariant(bfe->type == ftnode_fetch_subset || bfe->type == ftnode_fetch_prefetch);
565 566 567 568 569
    if (bfe->right_is_pos_infty) {
        return node->n_children - 1;
    } else if (bfe->range_lock_right_key == NULL) {
        return -1;
    } else {
570
        return toku_ftnode_which_child(node, bfe->range_lock_right_key, &bfe->h->cmp_descriptor, bfe->h->compare_fun);
571 572
    }
}
573

574
static int
575
ft_cursor_rightmost_child_wanted(FT_CURSOR cursor, FT_HANDLE brt, FTNODE node)
576 577 578 579 580 581
{
    if (cursor->right_is_pos_infty) {
        return node->n_children - 1;
    } else if (cursor->range_lock_right_key.data == NULL) {
        return -1;
    } else {
582
        return toku_ftnode_which_child(node, &cursor->range_lock_right_key, &brt->ft->cmp_descriptor, brt->ft->compare_fun);
583 584
    }
}
585

586
STAT64INFO_S
587
toku_get_and_clear_basement_stats(FTNODE leafnode) {
588 589 590
    invariant(leafnode->height == 0);
    STAT64INFO_S deltas = ZEROSTATS;
    for (int i = 0; i < leafnode->n_children; i++) {
Zardosht Kasheff's avatar
Zardosht Kasheff committed
591 592 593 594 595
        BASEMENTNODE bn = BLB(leafnode, i);
        invariant(BP_STATE(leafnode,i) == PT_AVAIL);
        deltas.numrows  += bn->stat64_delta.numrows;
        deltas.numbytes += bn->stat64_delta.numbytes;
        bn->stat64_delta = ZEROSTATS;
596 597 598 599
    }
    return deltas;
}

Yoni Fogel's avatar
Yoni Fogel committed
600
static void ft_status_update_flush_reason(FTNODE node, bool for_checkpoint) {
Zardosht Kasheff's avatar
Zardosht Kasheff committed
601 602
    if (node->height == 0) {
        if (for_checkpoint) {
603
            STATUS_INC(FT_DISK_FLUSH_LEAF_FOR_CHECKPOINT, 1);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
604 605
        }
        else {
606
            STATUS_INC(FT_DISK_FLUSH_LEAF, 1);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
607 608 609 610
        }
    }
    else {
        if (for_checkpoint) {
611
            STATUS_INC(FT_DISK_FLUSH_NONLEAF_FOR_CHECKPOINT, 1);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
612 613
        }
        else {
614
            STATUS_INC(FT_DISK_FLUSH_NONLEAF, 1);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
615 616 617 618
        }
    }
}

619 620
static void ftnode_update_disk_stats(
    FTNODE ftnode,
621
    FT ft,
Yoni Fogel's avatar
Yoni Fogel committed
622
    bool for_checkpoint
623
    )
Zardosht Kasheff's avatar
Zardosht Kasheff committed
624 625 626
{
    STAT64INFO_S deltas = ZEROSTATS;
    // capture deltas before rebalancing basements for serialization
627
    deltas = toku_get_and_clear_basement_stats(ftnode);
628 629 630 631 632 633 634 635 636 637
    // locking not necessary here with respect to checkpointing
    // in Clayface (because of the pending lock and cachetable lock
    // in toku_cachetable_begin_checkpoint)
    // essentially, if we are dealing with a for_checkpoint 
    // parameter in a function that is called by the flush_callback,
    // then the cachetable needs to ensure that this is called in a safe
    // manner that does not interfere with the beginning
    // of a checkpoint, which it does with the cachetable lock
    // and pending lock
    toku_ft_update_stats(&ft->h->on_disk_stats, deltas);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
638
    if (for_checkpoint) {
639
        toku_ft_update_stats(&ft->checkpoint_header->on_disk_stats, deltas);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
640 641 642
    }
}

643
static void ftnode_clone_partitions(FTNODE node, FTNODE cloned_node) {
Zardosht Kasheff's avatar
Zardosht Kasheff committed
644 645
    for (int i = 0; i < node->n_children; i++) {
        BP_BLOCKNUM(cloned_node,i) = BP_BLOCKNUM(node,i);
646
        paranoid_invariant(BP_STATE(node,i) == PT_AVAIL);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
647 648 649
        BP_STATE(cloned_node,i) = PT_AVAIL;
        BP_WORKDONE(cloned_node, i) = BP_WORKDONE(node, i);
        if (node->height == 0) {
650
            set_BLB(cloned_node, i, toku_clone_bn(BLB(node,i)));
Zardosht Kasheff's avatar
Zardosht Kasheff committed
651 652 653 654 655 656 657
        }
        else {
            set_BNC(cloned_node, i, toku_clone_nl(BNC(node,i)));
        }
    }
}

658 659 660 661 662 663 664 665 666 667 668 669 670
void toku_ftnode_checkpoint_complete_callback(void *value_data) {
    FTNODE node = static_cast<FTNODE>(value_data);
    if (node->height > 0) {
        for (int i = 0; i < node->n_children; ++i) {
            if (BP_STATE(node, i) == PT_AVAIL) {
                NONLEAF_CHILDINFO bnc = BNC(node, i);
                bnc->flow[1] = bnc->flow[0];
                bnc->flow[0] = 0;
            }
        }
    }
}

671
void toku_ftnode_clone_callback(
672 673 674
    void* value_data,
    void** cloned_value_data,
    PAIR_ATTR* new_attr,
Yoni Fogel's avatar
Yoni Fogel committed
675
    bool for_checkpoint,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
676 677 678
    void* write_extraargs
    )
{
679
    FTNODE node = static_cast<FTNODE>(value_data);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
680
    toku_assert_entire_node_in_memory(node);
681 682
    FT ft = static_cast<FT>(write_extraargs);
    FTNODE XCALLOC(cloned_node);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
683
    if (node->height == 0) {
684
        // set header stats, must be done before rebalancing
685
        ftnode_update_disk_stats(node, ft, for_checkpoint);
686
        // rebalance the leaf node
687
        rebalance_ftnode_leaf(node, ft->h->basementnodesize);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706
    }

    cloned_node->max_msn_applied_to_node_on_disk = node->max_msn_applied_to_node_on_disk;
    cloned_node->flags = node->flags;
    cloned_node->thisnodename = node->thisnodename;
    cloned_node->layout_version = node->layout_version;
    cloned_node->layout_version_original = node->layout_version_original;
    cloned_node->layout_version_read_from_disk = node->layout_version_read_from_disk;
    cloned_node->build_id = node->build_id;
    cloned_node->height = node->height;
    cloned_node->dirty = node->dirty;
    cloned_node->fullhash = node->fullhash;
    cloned_node->n_children = node->n_children;
    cloned_node->totalchildkeylens = node->totalchildkeylens;

    XMALLOC_N(node->n_children-1, cloned_node->childkeys);
    XMALLOC_N(node->n_children, cloned_node->bp);
    // clone pivots
    for (int i = 0; i < node->n_children-1; i++) {
707
        toku_clone_dbt(&cloned_node->childkeys[i], node->childkeys[i]);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
708 709
    }
    // clone partition
710
    ftnode_clone_partitions(node, cloned_node);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
711 712 713 714

    // clear dirty bit
    node->dirty = 0;
    cloned_node->dirty = 0;
715
    node->layout_version_read_from_disk = FT_LAYOUT_VERSION;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
716 717
    // set new pair attr if necessary
    if (node->height == 0) {
718
        *new_attr = make_ftnode_pair_attr(node);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
719 720
    }
    else {
Yoni Fogel's avatar
Yoni Fogel committed
721
        new_attr->is_valid = false;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
722 723 724 725 726
    }
    *cloned_value_data = cloned_node;
}


727
void toku_ftnode_flush_callback (
728
    CACHEFILE UU(cachefile),
729 730
    int fd,
    BLOCKNUM nodename,
731
    void *ftnode_v,
732 733 734 735
    void** disk_data,
    void *extraargs,
    PAIR_ATTR size __attribute__((unused)),
    PAIR_ATTR* new_size,
Yoni Fogel's avatar
Yoni Fogel committed
736 737 738
    bool write_me,
    bool keep_me,
    bool for_checkpoint,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
739
    bool is_clone
740
    )
Zardosht Kasheff's avatar
Zardosht Kasheff committed
741
{
742 743
    FT h = (FT) extraargs;
    FTNODE ftnode = (FTNODE) ftnode_v;
744 745 746
    FTNODE_DISK_DATA* ndd = (FTNODE_DISK_DATA*)disk_data;
    assert(ftnode->thisnodename.b==nodename.b);
    int height = ftnode->height;
747
    if (write_me) {
Zardosht Kasheff's avatar
Zardosht Kasheff committed
748
        if (height == 0 && !is_clone) {
749
            ftnode_update_disk_stats(ftnode, h, for_checkpoint);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
750
        }
751 752 753 754
        toku_assert_entire_node_in_memory(ftnode);
        int r = toku_serialize_ftnode_to(fd, ftnode->thisnodename, ftnode, ndd, !is_clone, h, for_checkpoint);
        assert_zero(r);
        ftnode->layout_version_read_from_disk = FT_LAYOUT_VERSION;
755
        ft_status_update_flush_reason(ftnode, for_checkpoint);
756 757
    }
    if (!keep_me) {
758
        if (!is_clone) {
759 760 761
            toku_free(*disk_data);
        }
        else {
762 763 764 765 766
            if (ftnode->height == 0) {
                for (int i = 0; i < ftnode->n_children; i++) {
                    if (BP_STATE(ftnode,i) == PT_AVAIL) {
                        BASEMENTNODE bn = BLB(ftnode, i);
                        toku_ft_decrease_stats(&h->in_memory_stats, bn->stat64_delta);
767 768 769 770
                    }
                }
            }
        }
771
        toku_ftnode_free(&ftnode);
772
    }
Zardosht Kasheff's avatar
Zardosht Kasheff committed
773
    else {
774
        *new_size = make_ftnode_pair_attr(ftnode);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
775
    }
776 777
}

778
void
779
toku_ft_status_update_pivot_fetch_reason(struct ftnode_fetch_extra *bfe)
780
{
781
    if (bfe->type == ftnode_fetch_prefetch) {
782
        STATUS_INC(FT_NUM_PIVOTS_FETCHED_PREFETCH, 1);
783
    } else if (bfe->type == ftnode_fetch_all) {
784
        STATUS_INC(FT_NUM_PIVOTS_FETCHED_WRITE, 1);
785
    } else if (bfe->type == ftnode_fetch_subset) {
786
        STATUS_INC(FT_NUM_PIVOTS_FETCHED_QUERY, 1);
787 788
    }
}
789

Zardosht Kasheff's avatar
Zardosht Kasheff committed
790
int toku_ftnode_fetch_callback (CACHEFILE UU(cachefile), PAIR p, int fd, BLOCKNUM nodename, uint32_t fullhash,
791
                                 void **ftnode_pv,  void** disk_data, PAIR_ATTR *sizep, int *dirtyp, void *extraargs) {
792
    assert(extraargs);
793 794 795 796
    assert(*ftnode_pv == NULL);
    FTNODE_DISK_DATA* ndd = (FTNODE_DISK_DATA*)disk_data;
    struct ftnode_fetch_extra *bfe = (struct ftnode_fetch_extra *)extraargs;
    FTNODE *node=(FTNODE*)ftnode_pv;
797 798 799
    // deserialize the node, must pass the bfe in because we cannot
    // evaluate what piece of the the node is necessary until we get it at
    // least partially into memory
800 801 802
    int r = toku_deserialize_ftnode_from(fd, nodename, fullhash, node, ndd, bfe);
    if (r != 0) {
        if (r == TOKUDB_BAD_CHECKSUM) {
803
            fprintf(stderr,
804 805
                    "Checksum failure while reading node in file %s.\n",
                    toku_cachefile_fname_in_env(cachefile));
806
        } else {
807
            fprintf(stderr, "Error deserializing node, errno = %d", r);
808 809
        }
        // make absolutely sure we crash before doing anything else.
810
        abort();
811
    }
812

813
    if (r == 0) {
814
        *sizep = make_ftnode_pair_attr(*node);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
815
        (*node)->ct_pair = p;
816
        *dirtyp = (*node)->dirty;  // deserialize could mark the node as dirty (presumably for upgrade)
817
    }
818 819 820
    return r;
}

821 822
void toku_ftnode_pe_est_callback(
    void* ftnode_pv,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
823
    void* disk_data,
824 825
    long* bytes_freed_estimate,
    enum partial_eviction_cost *cost,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
826 827 828
    void* UU(write_extraargs)
    )
{
829
    paranoid_invariant(ftnode_pv != NULL);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
830
    long bytes_to_free = 0;
831
    FTNODE node = static_cast<FTNODE>(ftnode_pv);
832
    if (node->dirty || node->height == 0 ||
833
        node->layout_version_read_from_disk < FT_FIRST_LAYOUT_VERSION_WITH_BASEMENT_NODES) {
Zardosht Kasheff's avatar
Zardosht Kasheff committed
834 835 836 837 838 839 840 841 842 843 844 845 846
        *bytes_freed_estimate = 0;
        *cost = PE_CHEAP;
        goto exit;
    }

    //
    // we are dealing with a clean internal node
    //
    *cost = PE_EXPENSIVE;
    // now lets get an estimate for how much data we can free up
    // we estimate the compressed size of data to be how large
    // the compressed data is on disk
    for (int i = 0; i < node->n_children; i++) {
John Esmet's avatar
John Esmet committed
847
        if (BP_STATE(node,i) == PT_AVAIL && BP_SHOULD_EVICT(node,i)) {
Zardosht Kasheff's avatar
Zardosht Kasheff committed
848 849 850 851
            // calculate how much data would be freed if
            // we compress this node and add it to
            // bytes_to_free

852
            // first get an estimate for how much space will be taken
Zardosht Kasheff's avatar
Zardosht Kasheff committed
853 854
            // after compression, it is simply the size of compressed
            // data on disk plus the size of the struct that holds it
855
            FTNODE_DISK_DATA ndd = (FTNODE_DISK_DATA) disk_data;
Yoni Fogel's avatar
Yoni Fogel committed
856
            uint32_t compressed_data_size = BP_SIZE(ndd, i);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
857 858 859
            compressed_data_size += sizeof(struct sub_block);

            // now get the space taken now
Yoni Fogel's avatar
Yoni Fogel committed
860
            uint32_t decompressed_data_size = get_avail_internal_node_partition_size(node,i);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
861 862 863 864 865 866 867 868 869
            bytes_to_free += (decompressed_data_size - compressed_data_size);
        }
    }

    *bytes_freed_estimate = bytes_to_free;
exit:
    return;
}

870
static void
871
compress_internal_node_partition(FTNODE node, int i, enum toku_compression_method compression_method)
872 873 874 875 876
{
    // if we should evict, compress the
    // message buffer into a sub_block
    assert(BP_STATE(node, i) == PT_AVAIL);
    assert(node->height > 0);
877
    SUB_BLOCK XMALLOC(sb);
878
    sub_block_init(sb);
879
    toku_create_compressed_partition_from_available(node, i, compression_method, sb);
880

881 882 883 884 885 886
    // now free the old partition and replace it with this
    destroy_nonleaf_childinfo(BNC(node,i));
    set_BSB(node, i, sb);
    BP_STATE(node,i) = PT_COMPRESSED;
}

887
void toku_evict_bn_from_memory(FTNODE node, int childnum, FT h) {
888 889 890
    // free the basement node
    assert(!node->dirty);
    BASEMENTNODE bn = BLB(node, childnum);
891
    toku_ft_decrease_stats(&h->in_memory_stats, bn->stat64_delta);
892 893 894 895 896 897 898
    struct mempool * mp = &bn->buffer_mempool;
    toku_mempool_destroy(mp);
    destroy_basement_node(bn);
    set_BNULL(node, childnum);
    BP_STATE(node, childnum) = PT_ON_DISK;
}

899
// callback for partially evicting a node
900 901
int toku_ftnode_pe_callback (void *ftnode_pv, PAIR_ATTR UU(old_attr), PAIR_ATTR* new_attr, void* extraargs) {
    FTNODE node = (FTNODE)ftnode_pv;
902
    FT ft = (FT) extraargs;
903
    // Don't partially evict dirty nodes
Zardosht Kasheff's avatar
Zardosht Kasheff committed
904 905 906
    if (node->dirty) {
        goto exit;
    }
907 908
    // Don't partially evict nodes whose partitions can't be read back
    // from disk individually
909
    if (node->layout_version_read_from_disk < FT_FIRST_LAYOUT_VERSION_WITH_BASEMENT_NODES) {
910 911
        goto exit;
    }
Zardosht Kasheff's avatar
Zardosht Kasheff committed
912
    //
913
    // partial eviction for nonleaf nodes
Zardosht Kasheff's avatar
Zardosht Kasheff committed
914 915 916 917 918
    //
    if (node->height > 0) {
        for (int i = 0; i < node->n_children; i++) {
            if (BP_STATE(node,i) == PT_AVAIL) {
                if (BP_SHOULD_EVICT(node,i)) {
919
                    STATUS_INC(FT_PARTIAL_EVICTIONS_NONLEAF, 1);
920
                    cilk_spawn compress_internal_node_partition(node, i, ft->h->compression_method);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
921 922 923 924 925 926 927 928 929
                }
                else {
                    BP_SWEEP_CLOCK(node,i);
                }
            }
            else {
                continue;
            }
        }
930
        cilk_sync;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
931 932 933 934 935 936 937 938 939 940
    }
    //
    // partial eviction strategy for basement nodes:
    //  if the bn is compressed, evict it
    //  else: check if it requires eviction, if it does, evict it, if not, sweep the clock count
    //
    else {
        for (int i = 0; i < node->n_children; i++) {
            // Get rid of compressed stuff no matter what.
            if (BP_STATE(node,i) == PT_COMPRESSED) {
941
                STATUS_INC(FT_PARTIAL_EVICTIONS_LEAF, 1);
942
                SUB_BLOCK sb = BSB(node, i);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
943
                toku_free(sb->compressed_ptr);
944
                toku_free(sb);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
945
                set_BNULL(node, i);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
946 947 948 949
                BP_STATE(node,i) = PT_ON_DISK;
            }
            else if (BP_STATE(node,i) == PT_AVAIL) {
                if (BP_SHOULD_EVICT(node,i)) {
950
                    STATUS_INC(FT_PARTIAL_EVICTIONS_LEAF, 1);
951
                    toku_evict_bn_from_memory(node, i, ft);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
952 953 954 955 956 957 958 959 960
                }
                else {
                    BP_SWEEP_CLOCK(node,i);
                }
            }
            else if (BP_STATE(node,i) == PT_ON_DISK) {
                continue;
            }
            else {
961
                abort();
Zardosht Kasheff's avatar
Zardosht Kasheff committed
962 963 964
            }
        }
    }
Zardosht Kasheff's avatar
Zardosht Kasheff committed
965 966

exit:
967
    *new_attr = make_ftnode_pair_attr(node);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
968 969 970
    return 0;
}

971

Zardosht Kasheff's avatar
Zardosht Kasheff committed
972
// Callback that states if a partial fetch of the node is necessary
973 974 975 976
// Currently, this function is responsible for the following things:
//  - reporting to the cachetable whether a partial fetch is required (as required by the contract of the callback)
//  - A couple of things that are NOT required by the callback, but we do for efficiency and simplicity reasons:
//   - for queries, set the value of bfe->child_to_read so that the query that called this can proceed with the query
977
//      as opposed to having to evaluate toku_ft_search_which_child again. This is done to make the in-memory query faster
978 979 980
//   - touch the necessary partition's clock. The reason we do it here is so that there is one central place it is done, and not done
//      by all the various callers
//
Yoni Fogel's avatar
Yoni Fogel committed
981
bool toku_ftnode_pf_req_callback(void* ftnode_pv, void* read_extraargs) {
982
    // placeholder for now
Yoni Fogel's avatar
Yoni Fogel committed
983
    bool retval = false;
984 985
    FTNODE node = (FTNODE) ftnode_pv;
    struct ftnode_fetch_extra *bfe = (struct ftnode_fetch_extra *) read_extraargs;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
986 987
    //
    // The three types of fetches that the brt layer may request are:
988 989 990
    //  - ftnode_fetch_none: no partitions are necessary (example use: stat64)
    //  - ftnode_fetch_subset: some subset is necessary (example use: toku_ft_search)
    //  - ftnode_fetch_all: entire node is necessary (example use: flush, split, merge)
991
    // The code below checks if the necessary partitions are already in memory,
Yoni Fogel's avatar
Yoni Fogel committed
992
    // and if they are, return false, and if not, return true
Zardosht Kasheff's avatar
Zardosht Kasheff committed
993
    //
994
    if (bfe->type == ftnode_fetch_none) {
Yoni Fogel's avatar
Yoni Fogel committed
995
        retval = false;
996
    }
997
    else if (bfe->type == ftnode_fetch_all) {
Yoni Fogel's avatar
Yoni Fogel committed
998
        retval = false;
999
        for (int i = 0; i < node->n_children; i++) {
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1000 1001 1002 1003
            BP_TOUCH_CLOCK(node,i);
            // if we find a partition that is not available,
            // then a partial fetch is required because
            // the entire node must be made available
1004
            if (BP_STATE(node,i) != PT_AVAIL) {
Yoni Fogel's avatar
Yoni Fogel committed
1005
                retval = true;
1006 1007 1008
            }
        }
    }
1009
    else if (bfe->type == ftnode_fetch_subset) {
1010 1011 1012 1013 1014
        // we do not take into account prefetching yet
        // as of now, if we need a subset, the only thing
        // we can possibly require is a single basement node
        // we find out what basement node the query cares about
        // and check if it is available
1015 1016
        paranoid_invariant(bfe->h->compare_fun);
        paranoid_invariant(bfe->search);
1017
        bfe->child_to_read = toku_ft_search_which_child(
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1018
            &bfe->h->cmp_descriptor,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1019
            bfe->h->compare_fun,
1020 1021 1022
            node,
            bfe->search
            );
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1023
        BP_TOUCH_CLOCK(node,bfe->child_to_read);
Yoni Fogel's avatar
Yoni Fogel committed
1024
        // child we want to read is not available, must set retval to true
1025 1026
        retval = (BP_STATE(node, bfe->child_to_read) != PT_AVAIL);
    }
1027
    else if (bfe->type == ftnode_fetch_prefetch) {
1028 1029
        // makes no sense to have prefetching disabled
        // and still call this function
1030
        paranoid_invariant(!bfe->disable_prefetching);
1031 1032 1033 1034
        int lc = toku_bfe_leftmost_child_wanted(bfe, node);
        int rc = toku_bfe_rightmost_child_wanted(bfe, node);
        for (int i = lc; i <= rc; ++i) {
            if (BP_STATE(node, i) != PT_AVAIL) {
Yoni Fogel's avatar
Yoni Fogel committed
1035
                retval = true;
1036 1037
            }
        }
1038 1039 1040
    }
    else {
        // we have a bug. The type should be known
1041
        abort();
1042 1043
    }
    return retval;
1044 1045
}

1046
static void
1047 1048
ft_status_update_partial_fetch_reason(
    struct ftnode_fetch_extra* UU(bfe),
1049 1050
    int UU(i),
    int UU(state),
Yoni Fogel's avatar
Yoni Fogel committed
1051
    bool UU(is_leaf)
1052 1053 1054 1055
    )
{
    invariant(state == PT_COMPRESSED || state == PT_ON_DISK);
    if (is_leaf) {
1056
        if (bfe->type == ftnode_fetch_prefetch) {
1057
            if (state == PT_COMPRESSED) {
1058
                STATUS_INC(FT_NUM_BASEMENTS_DECOMPRESSED_PREFETCH, 1);
1059
            } else {
1060
                STATUS_INC(FT_NUM_BASEMENTS_FETCHED_PREFETCH, 1);
1061
            }
1062
        } else if (bfe->type == ftnode_fetch_all) {
1063
            if (state == PT_COMPRESSED) {
1064
                STATUS_INC(FT_NUM_BASEMENTS_DECOMPRESSED_WRITE, 1);
1065
            } else {
1066
                STATUS_INC(FT_NUM_BASEMENTS_FETCHED_WRITE, 1);
1067 1068 1069
            }
        } else if (i == bfe->child_to_read) {
            if (state == PT_COMPRESSED) {
1070
                STATUS_INC(FT_NUM_BASEMENTS_DECOMPRESSED_NORMAL, 1);
1071
            } else {
1072
                STATUS_INC(FT_NUM_BASEMENTS_FETCHED_NORMAL, 1);
1073 1074 1075
            }
        } else {
            if (state == PT_COMPRESSED) {
1076
                STATUS_INC(FT_NUM_BASEMENTS_DECOMPRESSED_AGGRESSIVE, 1);
1077
            } else {
1078
                STATUS_INC(FT_NUM_BASEMENTS_FETCHED_AGGRESSIVE, 1);
1079 1080 1081 1082
            }
        }
    }
    else {
1083
        if (bfe->type == ftnode_fetch_prefetch) {
1084
            if (state == PT_COMPRESSED) {
1085
                STATUS_INC(FT_NUM_MSG_BUFFER_DECOMPRESSED_PREFETCH, 1);
1086
            } else {
1087
                STATUS_INC(FT_NUM_MSG_BUFFER_FETCHED_PREFETCH, 1);
1088
            }
1089
        } else if (bfe->type == ftnode_fetch_all) {
1090
            if (state == PT_COMPRESSED) {
1091
                STATUS_INC(FT_NUM_MSG_BUFFER_DECOMPRESSED_WRITE, 1);
1092
            } else {
1093
                STATUS_INC(FT_NUM_MSG_BUFFER_FETCHED_WRITE, 1);
1094 1095 1096
            }
        } else if (i == bfe->child_to_read) {
            if (state == PT_COMPRESSED) {
1097
                STATUS_INC(FT_NUM_MSG_BUFFER_DECOMPRESSED_NORMAL, 1);
1098
            } else {
1099
                STATUS_INC(FT_NUM_MSG_BUFFER_FETCHED_NORMAL, 1);
1100 1101 1102
            }
        } else {
            if (state == PT_COMPRESSED) {
1103
                STATUS_INC(FT_NUM_MSG_BUFFER_DECOMPRESSED_AGGRESSIVE, 1);
1104
            } else {
1105
                STATUS_INC(FT_NUM_MSG_BUFFER_FETCHED_AGGRESSIVE, 1);
1106 1107 1108 1109 1110
            }
        }
    }
}

1111
// callback for partially reading a node
1112 1113
// could have just used toku_ftnode_fetch_callback, but wanted to separate the two cases to separate functions
int toku_ftnode_pf_callback(void* ftnode_pv, void* disk_data, void* read_extraargs, int fd, PAIR_ATTR* sizep) {
1114
    int r = 0;
1115 1116 1117
    FTNODE node = (FTNODE) ftnode_pv;
    FTNODE_DISK_DATA ndd = (FTNODE_DISK_DATA) disk_data;
    struct ftnode_fetch_extra *bfe = (struct ftnode_fetch_extra *) read_extraargs;
1118
    // there must be a reason this is being called. If we get a garbage type or the type is ftnode_fetch_none,
1119
    // then something went wrong
1120
    assert((bfe->type == ftnode_fetch_subset) || (bfe->type == ftnode_fetch_all) || (bfe->type == ftnode_fetch_prefetch));
1121 1122
    // determine the range to prefetch
    int lc, rc;
1123
    if (!bfe->disable_prefetching &&
1124
        (bfe->type == ftnode_fetch_subset || bfe->type == ftnode_fetch_prefetch)
1125
        )
1126
    {
1127 1128 1129 1130 1131 1132
        lc = toku_bfe_leftmost_child_wanted(bfe, node);
        rc = toku_bfe_rightmost_child_wanted(bfe, node);
    } else {
        lc = -1;
        rc = -1;
    }
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1133 1134
    // TODO: possibly cilkify expensive operations in this loop
    // TODO: review this with others to see if it can be made faster
1135 1136 1137 1138
    for (int i = 0; i < node->n_children; i++) {
        if (BP_STATE(node,i) == PT_AVAIL) {
            continue;
        }
1139
        if ((lc <= i && i <= rc) || toku_bfe_wants_child_available(bfe, i)) {
1140
            ft_status_update_partial_fetch_reason(bfe, i, BP_STATE(node, i), (node->height == 0));
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1141
            if (BP_STATE(node,i) == PT_COMPRESSED) {
1142
                r = toku_deserialize_bp_from_compressed(node, i, &bfe->h->cmp_descriptor, bfe->h->compare_fun);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1143 1144
            }
            else if (BP_STATE(node,i) == PT_ON_DISK) {
1145
                r = toku_deserialize_bp_from_disk(node, ndd, i, fd, bfe);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1146 1147
            }
            else {
1148
                abort();
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1149
            }
1150
        }
1151

1152 1153
        if (r != 0) {
            if (r == TOKUDB_BAD_CHECKSUM) {
1154 1155 1156
                fprintf(stderr,
                        "Checksum failure while reading node partition in file %s.\n",
                        toku_cachefile_fname_in_env(bfe->h->cf));
1157
            } else {
1158 1159
                fprintf(stderr,
                        "Error while reading node partition %d\n",
1160
                        get_maybe_error_errno());
1161
            }
1162
            abort();
1163
        }
1164
    }
1165

1166
    *sizep = make_ftnode_pair_attr(node);
1167

1168
    return 0;
1169 1170
}

1171
static int
Yoni Fogel's avatar
Yoni Fogel committed
1172
leafval_heaviside_le (uint32_t klen, void *kval,
1173
                      struct cmd_leafval_heaviside_extra *be) {
1174
    DBT dbt;
1175
    DBT const * const key = be->key;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1176
    FAKE_DB(db, be->desc);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1177
    return be->compare_fun(&db,
1178 1179
                           toku_fill_dbt(&dbt, kval, klen),
                           key);
1180 1181
}

1182 1183 1184
//TODO: #1125 optimize
int
toku_cmd_leafval_heaviside (OMTVALUE lev, void *extra) {
1185 1186
    LEAFENTRY CAST_FROM_VOIDP(le, lev);
    struct cmd_leafval_heaviside_extra *CAST_FROM_VOIDP(be, extra);
Yoni Fogel's avatar
Yoni Fogel committed
1187
    uint32_t keylen;
1188 1189
    void*     key = le_key_and_len(le, &keylen);
    return leafval_heaviside_le(keylen, key,
1190
                                be);
1191 1192
}

1193
static int
1194
ft_compare_pivot(DESCRIPTOR desc, ft_compare_func cmp, const DBT *key, const DBT *pivot)
1195
{
1196
    int r;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1197
    FAKE_DB(db, desc);
1198
    r = cmp(&db, key, pivot);
1199
    return r;
1200 1201
}

1202

1203
// destroys the internals of the ftnode, but it does not free the values
1204
// that are stored
1205
// this is common functionality for toku_ftnode_free and rebalance_ftnode_leaf
1206
// MUST NOT do anything besides free the structures that have been allocated
1207
void toku_destroy_ftnode_internals(FTNODE node)
1208
{
1209
    for (int i=0; i<node->n_children-1; i++) {
1210
        toku_free(node->childkeys[i].data);
1211 1212 1213
    }
    toku_free(node->childkeys);
    node->childkeys = NULL;
1214 1215 1216 1217

    for (int i=0; i < node->n_children; i++) {
        if (BP_STATE(node,i) == PT_AVAIL) {
            if (node->height > 0) {
1218
                destroy_nonleaf_childinfo(BNC(node,i));
1219
            } else {
1220
                destroy_basement_node(BLB(node, i));
1221
            }
1222
        } else if (BP_STATE(node,i) == PT_COMPRESSED) {
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1223 1224
            SUB_BLOCK sb = BSB(node,i);
            toku_free(sb->compressed_ptr);
1225
            toku_free(sb);
1226
        } else {
1227
            paranoid_invariant(is_BNULL(node, i));
1228
        }
1229
        set_BNULL(node, i);
1230 1231 1232 1233
    }
    toku_free(node->bp);
    node->bp = NULL;

1234 1235
}

1236

1237
/* Frees a node, including all the stuff in the hash table. */
1238
void toku_ftnode_free (FTNODE *nodep) {
1239

1240
    //TODO: #1378 Take omt lock (via ftnode) around call to toku_omt_destroy().
1241

1242
    FTNODE node=*nodep;
1243
    if (node->height == 0) {
1244
        for (int i = 0; i < node->n_children; i++) {
1245
            if (BP_STATE(node,i) == PT_AVAIL) {
1246 1247
                struct mempool * mp = &(BLB_BUFFER_MEMPOOL(node, i));
                toku_mempool_destroy(mp);
1248
            }
1249
        }
1250
        STATUS_INC(FT_DESTROY_LEAF, 1);
1251
    } else {
1252
        STATUS_INC(FT_DESTROY_NONLEAF, 1);
1253
    }
1254
    toku_destroy_ftnode_internals(node);
1255 1256 1257 1258
    toku_free(node);
    *nodep=0;
}

1259
void
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1260
toku_initialize_empty_ftnode (FTNODE n, BLOCKNUM nodename, int height, int num_children, int layout_version, unsigned int flags)
1261
// Effect: Fill in N as an empty ftnode.
1262
{
1263 1264
    paranoid_invariant(layout_version != 0);
    paranoid_invariant(height >= 0);
1265

1266
    if (height == 0)
1267
        STATUS_INC(FT_CREATE_LEAF, 1);
1268
    else
1269
        STATUS_INC(FT_CREATE_NONLEAF, 1);
1270

1271
    n->max_msn_applied_to_node_on_disk = ZERO_MSN;    // correct value for root node, harmless for others
1272
    n->flags = flags;
Bradley C. Kuszmaul's avatar
Rename  
Bradley C. Kuszmaul committed
1273
    n->thisnodename = nodename;
1274
    n->layout_version               = layout_version;
1275 1276 1277
    n->layout_version_original = layout_version;
    n->layout_version_read_from_disk = layout_version;
    n->height = height;
1278
    n->totalchildkeylens = 0;
1279
    n->childkeys = 0;
1280
    n->bp = 0;
1281
    n->n_children = num_children;
1282

1283
    if (num_children > 0) {
1284 1285
        XMALLOC_N(num_children-1, n->childkeys);
        XMALLOC_N(num_children, n->bp);
1286
        for (int i = 0; i < num_children; i++) {
1287 1288
            BP_BLOCKNUM(n,i).b=0;
            BP_STATE(n,i) = PT_INVALID;
1289
            BP_WORKDONE(n,i) = 0;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1290
            BP_INIT_TOUCHED_CLOCK(n, i);
1291
            set_BNULL(n,i);
1292
            if (height > 0) {
1293 1294 1295
                set_BNC(n, i, toku_create_empty_nl());
            } else {
                set_BLB(n, i, toku_create_empty_bn());
1296
            }
1297
        }
Bradley C. Kuszmaul's avatar
Rename  
Bradley C. Kuszmaul committed
1298
    }
1299
    n->dirty = 1;  // special case exception, it's okay to mark as dirty because the basements are empty
Bradley C. Kuszmaul's avatar
Rename  
Bradley C. Kuszmaul committed
1300 1301
}

1302
static void
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1303 1304 1305
ft_init_new_root(FT ft, FTNODE oldroot, FTNODE *newrootp)
// Effect:  Create a new root node whose two children are the split of oldroot.
//  oldroot is unpinned in the process.
1306 1307
//  Leave the new root pinned.
{
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328
    FTNODE newroot;

    BLOCKNUM old_blocknum = oldroot->thisnodename;
    uint32_t old_fullhash = oldroot->fullhash;
    PAIR old_pair = oldroot->ct_pair;
    
    int new_height = oldroot->height+1;
    uint32_t new_fullhash;
    BLOCKNUM new_blocknum;
    PAIR new_pair = NULL;

    cachetable_put_empty_node_with_dep_nodes(
        ft,
        1,
        &oldroot,
        &new_blocknum,
        &new_fullhash,
        &newroot
        );
    new_pair = newroot->ct_pair;
    
1329 1330
    assert(newroot);
    assert(new_height > 0);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1331 1332 1333 1334 1335 1336 1337 1338 1339 1340
    toku_initialize_empty_ftnode (
        newroot, 
        new_blocknum, 
        new_height, 
        1, 
        ft->h->layout_version, 
        ft->h->flags
        );
    MSN msna = oldroot->max_msn_applied_to_node_on_disk;
    newroot->max_msn_applied_to_node_on_disk = msna;
1341
    BP_STATE(newroot,0) = PT_AVAIL;
1342
    newroot->dirty = 1;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359

    // now do the "switcheroo"
    BP_BLOCKNUM(newroot,0) = new_blocknum;
    newroot->thisnodename = old_blocknum;
    newroot->fullhash = old_fullhash;
    newroot->ct_pair = old_pair;

    oldroot->thisnodename = new_blocknum;
    oldroot->fullhash = new_fullhash;
    oldroot->ct_pair = new_pair;

    toku_cachetable_swap_pair_values(old_pair, new_pair);

    toku_ft_split_child(
        ft,
        newroot,
        0, // childnum to split
1360 1361
        oldroot,
        SPLIT_EVENLY
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378
        );

    // ft_split_child released locks on newroot
    // and oldroot, so now we repin and
    // return to caller
    struct ftnode_fetch_extra bfe;
    fill_bfe_for_full_read(&bfe, ft);
    toku_pin_ftnode_off_client_thread(
        ft,
        old_blocknum,
        old_fullhash,
        &bfe,
        PL_WRITE_EXPENSIVE, // may_modify_node
        0,
        NULL,
        newrootp
        );
1379
}
Bradley C. Kuszmaul's avatar
Rename  
Bradley C. Kuszmaul committed
1380

1381
static void
1382
init_childinfo(FTNODE node, int childnum, FTNODE child) {
1383 1384
    BP_BLOCKNUM(node,childnum) = child->thisnodename;
    BP_STATE(node,childnum) = PT_AVAIL;
1385 1386
    BP_WORKDONE(node, childnum)   = 0;
    set_BNC(node, childnum, toku_create_empty_nl());
1387 1388 1389
}

static void
1390
init_childkey(FTNODE node, int childnum, const DBT *pivotkey) {
1391
    toku_copy_dbt(&node->childkeys[childnum], *pivotkey);
1392
    node->totalchildkeylens += pivotkey->size;
1393 1394
}

1395
// Used only by test programs: append a child node to a parent node
1396
void
1397
toku_ft_nonleaf_append_child(FTNODE node, FTNODE child, const DBT *pivotkey) {
1398 1399
    int childnum = node->n_children;
    node->n_children++;
1400
    XREALLOC_N(node->n_children, node->bp);
1401
    init_childinfo(node, childnum, child);
1402
    XREALLOC_N(node->n_children-1, node->childkeys);
1403
    if (pivotkey) {
1404 1405
        invariant(childnum > 0);
        init_childkey(node, childnum-1, pivotkey);
1406
    }
1407
    node->dirty = 1;
1408 1409
}

1410
static void
1411
ft_leaf_delete_leafentry (
1412
    BASEMENTNODE bn,
Yoni Fogel's avatar
Yoni Fogel committed
1413
    uint32_t idx,
1414 1415
    LEAFENTRY le
    )
1416 1417 1418 1419 1420 1421 1422
// Effect: Delete leafentry
//   idx is the location where it is
//   le is the leafentry to be deleted
{
    // Figure out if one of the other keys is the same key

    {
1423
        int r = toku_omt_delete_at(bn->buffer, idx);
1424
        assert_zero(r);
1425 1426
    }

1427
    bn->n_bytes_in_buffer -= leafentry_disksize(le);
1428

1429
    toku_mempool_mfree(&bn->buffer_mempool, 0, leafentry_memsize(le)); // Must pass 0, since le is no good any more.
1430 1431
}

1432
void
1433
toku_ft_bn_apply_cmd_once (
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1434
    BASEMENTNODE bn,
1435
    const FT_MSG cmd,
Yoni Fogel's avatar
Yoni Fogel committed
1436
    uint32_t idx,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1437
    LEAFENTRY le,
1438 1439
    uint64_t *workdone,
    STAT64INFO stats_to_update
1440 1441
    )
// Effect: Apply cmd to leafentry (msn is ignored)
1442
//         Calculate work done by message on leafentry and add it to caller's workdone counter.
1443 1444 1445
//   idx is the location where it goes
//   le is old leafentry
{
1446
    size_t newsize=0, oldsize=0, workdone_this_le=0;
1447
    LEAFENTRY new_le=0;
1448
    void *maybe_free = 0;
1449 1450
    int64_t numbytes_delta = 0;  // how many bytes of user data (not including overhead) were added or deleted from this row
    int64_t numrows_delta = 0;   // will be +1 or -1 or 0 (if row was added or deleted or not)
1451 1452

    if (le)
1453
        oldsize = leafentry_memsize(le);
1454

1455
    // apply_msg_to_leafentry() may call mempool_malloc_from_omt() to allocate more space.
1456
    // That means le is guaranteed to not cause a sigsegv but it may point to a mempool that is
1457
    // no longer in use.  We'll have to release the old mempool later.
1458
    {
1459
        int r = apply_msg_to_leafentry(cmd, le, &newsize, &new_le, &bn->buffer, &bn->buffer_mempool, &maybe_free, &numbytes_delta);
1460
        invariant(r==0);
1461
    }
1462

1463 1464 1465
    if (new_le) {
        paranoid_invariant(newsize == leafentry_disksize(new_le));
    }
1466
    if (le && new_le) {
1467 1468
        bn->n_bytes_in_buffer -= oldsize;
        bn->n_bytes_in_buffer += newsize;
1469

1470 1471 1472 1473
        // This mfree must occur after the mempool_malloc so that when
        // the mempool is compressed everything is accounted for.  But
        // we must compute the size before doing the mempool mfree
        // because otherwise the le pointer is no good.
1474
        toku_mempool_mfree(&bn->buffer_mempool, 0, oldsize); // Must pass 0, since le may be no good any more.
1475

1476 1477 1478 1479 1480 1481
        {
            int r = toku_omt_set_at(bn->buffer, new_le, idx);
            invariant(r==0);
        }

        workdone_this_le = (oldsize > newsize ? oldsize : newsize);  // work done is max of le size before and after message application
1482

1483
    } else {           // we did not just replace a row, so ...
1484 1485
        if (le) {
            //            ... we just deleted a row ...
1486
            // It was there, note that it's gone and remove it from the mempool
1487
            ft_leaf_delete_leafentry (bn, idx, le);
1488 1489 1490 1491 1492 1493 1494
            workdone_this_le = oldsize;
            numrows_delta = -1;
        }
        if (new_le) {
            //            ... or we just added a row
            int r = toku_omt_insert_at(bn->buffer, new_le, idx);
            invariant(r==0);
1495
            bn->n_bytes_in_buffer += newsize;
1496 1497
            workdone_this_le = newsize;
            numrows_delta = 1;
1498
        }
1499
    }
1500
    if (workdone) {  // test programs may call with NULL
1501
        *workdone += workdone_this_le;
1502
    }
1503

1504 1505 1506
    // if we created a new mempool buffer, free the old one
    if (maybe_free) toku_free(maybe_free);

1507 1508 1509
    // now update stat64 statistics
    bn->stat64_delta.numrows  += numrows_delta;
    bn->stat64_delta.numbytes += numbytes_delta;
1510 1511 1512 1513 1514
    // the only reason stats_to_update may be null is for tests
    if (stats_to_update) {
        stats_to_update->numrows += numrows_delta;
        stats_to_update->numbytes += numbytes_delta;
    }
1515

1516 1517
}

1518 1519
static const uint32_t setval_tag = 0xee0ccb99; // this was gotten by doing "cat /dev/random|head -c4|od -x" to get a random number.  We want to make sure that the user actually passes us the setval_extra_s that we passed in.
struct setval_extra_s {
Yoni Fogel's avatar
Yoni Fogel committed
1520 1521
    uint32_t  tag;
    bool did_set_val;
1522
    int         setval_r;    // any error code that setval_fun wants to return goes here.
1523
    // need arguments for toku_ft_bn_apply_cmd_once
1524
    BASEMENTNODE bn;
1525
    MSN msn;              // captured from original message, not currently used
1526 1527
    XIDS xids;
    const DBT *key;
Yoni Fogel's avatar
Yoni Fogel committed
1528
    uint32_t idx;
1529
    LEAFENTRY le;
1530
    uint64_t * workdone;  // set by toku_ft_bn_apply_cmd_once()
1531
    STAT64INFO stats_to_update;
1532 1533 1534 1535 1536
};

/*
 * If new_val == NULL, we send a delete message instead of an insert.
 * This happens here instead of in do_delete() for consistency.
1537 1538
 * setval_fun() is called from handlerton, passing in svextra_v
 * from setval_extra_s input arg to brt->update_fun().
1539 1540
 */
static void setval_fun (const DBT *new_val, void *svextra_v) {
1541
    struct setval_extra_s *CAST_FROM_VOIDP(svextra, svextra_v);
1542 1543
    paranoid_invariant(svextra->tag==setval_tag);
    paranoid_invariant(!svextra->did_set_val);
Yoni Fogel's avatar
Yoni Fogel committed
1544
    svextra->did_set_val = true;
1545 1546

    {
1547
        // can't leave scope until toku_ft_bn_apply_cmd_once if
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1548 1549
        // this is a delete
        DBT val;
1550
        FT_MSG_S msg = { FT_NONE, svextra->msn, svextra->xids,
1551
                         .u = { .id = {svextra->key, NULL} } };
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1552
        if (new_val) {
1553
            msg.type = FT_INSERT;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1554 1555
            msg.u.id.val = new_val;
        } else {
1556
            msg.type = FT_DELETE_ANY;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1557 1558 1559
            toku_init_dbt(&val);
            msg.u.id.val = &val;
        }
1560
        toku_ft_bn_apply_cmd_once(svextra->bn, &msg,
1561 1562
                                  svextra->idx, svextra->le,
                                  svextra->workdone, svextra->stats_to_update);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1563
        svextra->setval_r = 0;
1564
    }
1565 1566
}

1567
// We are already past the msn filter (in toku_ft_bn_apply_cmd(), which calls do_update()),
1568
// so capturing the msn in the setval_extra_s is not strictly required.         The alternative
1569 1570
// would be to put a dummy msn in the messages created by setval_fun(), but preserving
// the original msn seems cleaner and it preserves accountability at a lower layer.
1571
static int do_update(ft_update_func update_fun, DESCRIPTOR desc, BASEMENTNODE bn, FT_MSG cmd, uint32_t idx,
1572
                     LEAFENTRY le,
1573 1574
                     uint64_t * workdone,
                     STAT64INFO stats_to_update) {
1575 1576 1577 1578 1579 1580 1581 1582 1583
    LEAFENTRY le_for_update;
    DBT key;
    const DBT *keyp;
    const DBT *update_function_extra;
    DBT vdbt;
    const DBT *vdbtp;

    // the location of data depends whether this is a regular or
    // broadcast update
1584
    if (cmd->type == FT_UPDATE) {
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1585 1586
        // key is passed in with command (should be same as from le)
        // update function extra is passed in with command
1587
        STATUS_INC(FT_UPDATES, 1);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1588 1589
        keyp = cmd->u.id.key;
        update_function_extra = cmd->u.id.val;
1590
    } else if (cmd->type == FT_UPDATE_BROADCAST_ALL) {
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1591 1592
        // key is not passed in with broadcast, it comes from le
        // update function extra is passed in with command
1593
        paranoid_invariant(le);  // for broadcast updates, we just hit all leafentries
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1594
                     // so this cannot be null
1595
        paranoid_invariant(cmd->u.id.key->size == 0);
1596
        STATUS_INC(FT_UPDATES_BROADCAST, 1);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1597 1598
        keyp = toku_fill_dbt(&key, le_key(le), le_keylen(le));
        update_function_extra = cmd->u.id.val;
1599
    } else {
1600
        abort();
1601 1602 1603
    }

    if (le && !le_latest_is_del(le)) {
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1604
        // if the latest val exists, use it, and we'll use the leafentry later
Yoni Fogel's avatar
Yoni Fogel committed
1605
        uint32_t vallen;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1606 1607 1608
        void *valp = le_latest_val_and_len(le, &vallen);
        vdbtp = toku_fill_dbt(&vdbt, valp, vallen);
        le_for_update = le;
1609
    } else {
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1610 1611 1612
        // otherwise, the val and leafentry are both going to be null
        vdbtp = NULL;
        le_for_update = NULL;
1613 1614
    }

Yoni Fogel's avatar
Yoni Fogel committed
1615
    struct setval_extra_s setval_extra = {setval_tag, false, 0, bn, cmd->msn, cmd->xids,
1616
                                          keyp, idx, le_for_update, workdone, stats_to_update};
1617
    // call handlerton's brt->update_fun(), which passes setval_extra to setval_fun()
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1618
    FAKE_DB(db, desc);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1619 1620 1621 1622 1623 1624 1625
    int r = update_fun(
        &db,
        keyp,
        vdbtp,
        update_function_extra,
        setval_fun, &setval_extra
        );
1626 1627 1628 1629 1630

    if (r == 0) { r = setval_extra.setval_r; }
    return r;
}

1631
// Should be renamed as something like "apply_cmd_to_basement()."
1632
void
1633 1634 1635
toku_ft_bn_apply_cmd (
    ft_compare_func compare_fun,
    ft_update_func update_fun,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1636
    DESCRIPTOR desc,
1637
    BASEMENTNODE bn,
1638
    FT_MSG cmd,
1639 1640
    uint64_t *workdone,
    STAT64INFO stats_to_update
1641
    )
1642
// Effect:
1643 1644
//   Put a cmd into a leaf.
//   Calculate work done by message on leafnode and add it to caller's workdone counter.
1645
// The leaf could end up "too big" or "too small".  The caller must fix that up.
1646 1647 1648 1649
{
    LEAFENTRY storeddata;
    OMTVALUE storeddatav=NULL;

Yoni Fogel's avatar
Yoni Fogel committed
1650
    uint32_t omt_size;
1651
    int r;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1652
    struct cmd_leafval_heaviside_extra be = {compare_fun, desc, cmd->u.id.key};
1653

1654 1655
    unsigned int doing_seqinsert = bn->seqinsert;
    bn->seqinsert = 0;
1656

1657
    switch (cmd->type) {
1658 1659
    case FT_INSERT_NO_OVERWRITE:
    case FT_INSERT: {
Yoni Fogel's avatar
Yoni Fogel committed
1660
        uint32_t idx;
1661 1662 1663 1664
        if (doing_seqinsert) {
            idx = toku_omt_size(bn->buffer);
            r = toku_omt_fetch(bn->buffer, idx-1, &storeddatav);
            if (r != 0) goto fz;
1665
            CAST_FROM_VOIDP(storeddata, storeddatav);
1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676
            int cmp = toku_cmd_leafval_heaviside(storeddata, &be);
            if (cmp >= 0) goto fz;
            r = DB_NOTFOUND;
        } else {
        fz:
            r = toku_omt_find_zero(bn->buffer, toku_cmd_leafval_heaviside, &be,
                                   &storeddatav, &idx);
        }
        if (r==DB_NOTFOUND) {
            storeddata = 0;
        } else {
1677
            assert_zero(r);
1678
            CAST_FROM_VOIDP(storeddata, storeddatav);
1679
        }
1680
        toku_ft_bn_apply_cmd_once(bn, cmd, idx, storeddata, workdone, stats_to_update);
1681 1682 1683 1684 1685

        // if the insertion point is within a window of the right edge of
        // the leaf then it is sequential
        // window = min(32, number of leaf entries/16)
        {
Yoni Fogel's avatar
Yoni Fogel committed
1686 1687
            uint32_t s = toku_omt_size(bn->buffer);
            uint32_t w = s / 16;
1688 1689 1690 1691 1692 1693 1694 1695
            if (w == 0) w = 1;
            if (w > 32) w = 32;

            // within the window?
            if (s - idx <= w)
                bn->seqinsert = doing_seqinsert + 1;
        }
        break;
1696
    }
1697 1698 1699
    case FT_DELETE_ANY:
    case FT_ABORT_ANY:
    case FT_COMMIT_ANY: {
Yoni Fogel's avatar
Yoni Fogel committed
1700
        uint32_t idx;
1701 1702 1703 1704 1705
        // Apply to all the matches

        r = toku_omt_find_zero(bn->buffer, toku_cmd_leafval_heaviside, &be,
                               &storeddatav, &idx);
        if (r == DB_NOTFOUND) break;
1706
        assert_zero(r);
1707
        CAST_FROM_VOIDP(storeddata, storeddatav);
1708 1709

        while (1) {
Yoni Fogel's avatar
Yoni Fogel committed
1710
            uint32_t num_leafentries_before = toku_omt_size(bn->buffer);
1711

1712
            toku_ft_bn_apply_cmd_once(bn, cmd, idx, storeddata, workdone, stats_to_update);
1713 1714 1715

            {
                // Now we must find the next leafentry.
Yoni Fogel's avatar
Yoni Fogel committed
1716
                uint32_t num_leafentries_after = toku_omt_size(bn->buffer);
1717 1718 1719 1720 1721
                //idx is the index of the leafentry we just modified.
                //If the leafentry was deleted, we will have one less leafentry in
                //the omt than we started with and the next leafentry will be at the
                //same index as the deleted one. Otherwise, the next leafentry will
                //be at the next index (+1).
1722 1723
                paranoid_invariant(num_leafentries_before   == num_leafentries_after ||
                                   num_leafentries_before-1 == num_leafentries_after);
1724 1725
                if (num_leafentries_after==num_leafentries_before) idx++; //Not deleted, advance index.

1726
                paranoid_invariant(idx <= num_leafentries_after);
1727 1728 1729 1730
                if (idx == num_leafentries_after) break; //Reached the end of the leaf
                r = toku_omt_fetch(bn->buffer, idx, &storeddatav);
                assert_zero(r);
            }
1731
            CAST_FROM_VOIDP(storeddata, storeddatav);
1732 1733
            {        // Continue only if the next record that we found has the same key.
                DBT adbt;
Yoni Fogel's avatar
Yoni Fogel committed
1734
                uint32_t keylen;
1735 1736 1737 1738 1739 1740 1741 1742 1743 1744
                void *keyp = le_key_and_len(storeddata, &keylen);
                FAKE_DB(db, desc);
                if (compare_fun(&db,
                                toku_fill_dbt(&adbt, keyp, keylen),
                                cmd->u.id.key) != 0)
                    break;
            }
        }

        break;
1745
    }
1746
    case FT_OPTIMIZE_FOR_UPGRADE:
1747
        // fall through so that optimize_for_upgrade performs rest of the optimize logic
1748 1749
    case FT_COMMIT_BROADCAST_ALL:
    case FT_OPTIMIZE:
1750 1751
        // Apply to all leafentries
        omt_size = toku_omt_size(bn->buffer);
Yoni Fogel's avatar
Yoni Fogel committed
1752
        for (uint32_t idx = 0; idx < omt_size; ) {
1753 1754
            r = toku_omt_fetch(bn->buffer, idx, &storeddatav);
            assert_zero(r);
1755
            CAST_FROM_VOIDP(storeddata, storeddatav);
1756 1757
            int deleted = 0;
            if (!le_is_clean(storeddata)) { //If already clean, nothing to do.
1758
                toku_ft_bn_apply_cmd_once(bn, cmd, idx, storeddata, workdone, stats_to_update);
Yoni Fogel's avatar
Yoni Fogel committed
1759
                uint32_t new_omt_size = toku_omt_size(bn->buffer);
1760
                if (new_omt_size != omt_size) {
1761
                    paranoid_invariant(new_omt_size+1 == omt_size);
1762 1763 1764 1765 1766 1767 1768 1769 1770
                    //Item was deleted.
                    deleted = 1;
                }
            }
            if (deleted)
                omt_size--;
            else
                idx++;
        }
1771
        paranoid_invariant(toku_omt_size(bn->buffer) == omt_size);
1772 1773

        break;
1774 1775
    case FT_COMMIT_BROADCAST_TXN:
    case FT_ABORT_BROADCAST_TXN:
1776 1777
        // Apply to all leafentries if txn is represented
        omt_size = toku_omt_size(bn->buffer);
Yoni Fogel's avatar
Yoni Fogel committed
1778
        for (uint32_t idx = 0; idx < omt_size; ) {
1779 1780
            r = toku_omt_fetch(bn->buffer, idx, &storeddatav);
            assert_zero(r);
1781
            CAST_FROM_VOIDP(storeddata, storeddatav);
1782 1783
            int deleted = 0;
            if (le_has_xids(storeddata, cmd->xids)) {
1784
                toku_ft_bn_apply_cmd_once(bn, cmd, idx, storeddata, workdone, stats_to_update);
Yoni Fogel's avatar
Yoni Fogel committed
1785
                uint32_t new_omt_size = toku_omt_size(bn->buffer);
1786
                if (new_omt_size != omt_size) {
1787
                    paranoid_invariant(new_omt_size+1 == omt_size);
1788 1789 1790 1791 1792 1793 1794 1795 1796
                    //Item was deleted.
                    deleted = 1;
                }
            }
            if (deleted)
                omt_size--;
            else
                idx++;
        }
1797
        paranoid_invariant(toku_omt_size(bn->buffer) == omt_size);
1798 1799

        break;
1800
    case FT_UPDATE: {
Yoni Fogel's avatar
Yoni Fogel committed
1801
        uint32_t idx;
1802 1803 1804 1805 1806
        r = toku_omt_find_zero(bn->buffer, toku_cmd_leafval_heaviside, &be,
                               &storeddatav, &idx);
        if (r==DB_NOTFOUND) {
            r = do_update(update_fun, desc, bn, cmd, idx, NULL, workdone, stats_to_update);
        } else if (r==0) {
1807
            CAST_FROM_VOIDP(storeddata, storeddatav);
1808 1809 1810
            r = do_update(update_fun, desc, bn, cmd, idx, storeddata, workdone, stats_to_update);
        } // otherwise, a worse error, just return it
        break;
1811
    }
1812
    case FT_UPDATE_BROADCAST_ALL: {
1813
        // apply to all leafentries.
Yoni Fogel's avatar
Yoni Fogel committed
1814 1815
        uint32_t idx = 0;
        uint32_t num_leafentries_before;
1816 1817
        while (idx < (num_leafentries_before = toku_omt_size(bn->buffer))) {
            r = toku_omt_fetch(bn->buffer, idx, &storeddatav);
1818
            assert_zero(r);
1819
            CAST_FROM_VOIDP(storeddata, storeddatav);
1820
            r = do_update(update_fun, desc, bn, cmd, idx, storeddata, workdone, stats_to_update);
1821
            assert_zero(r);
1822 1823 1824 1825 1826 1827 1828

            if (num_leafentries_before == toku_omt_size(bn->buffer)) {
                // we didn't delete something, so increment the index.
                idx++;
            }
        }
        break;
1829
    }
1830
    case FT_NONE: break; // don't do anything
1831 1832
    }

1833
    return;
1834 1835
}

1836 1837
static inline int
key_msn_cmp(const DBT *a, const DBT *b, const MSN amsn, const MSN bmsn,
1838
            DESCRIPTOR descriptor, ft_compare_func key_cmp)
1839
{
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1840
    FAKE_DB(db, descriptor);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
1841
    int r = key_cmp(&db, a, b);
1842
    if (r == 0) {
1843 1844 1845 1846 1847 1848 1849
        if (amsn.msn > bmsn.msn) {
            r = +1;
        } else if (amsn.msn < bmsn.msn) {
            r = -1;
        } else {
            r = 0;
        }
1850 1851 1852 1853 1854
    }
    return r;
}

int
1855
toku_fifo_entry_key_msn_heaviside(const int32_t &offset, const struct toku_fifo_entry_key_msn_heaviside_extra &extra)
1856
{
1857
    const struct fifo_entry *query = toku_fifo_get_entry(extra.fifo, offset);
1858
    DBT qdbt;
1859
    const DBT *query_key = fill_dbt_for_fifo_entry(&qdbt, query);
1860 1861 1862
    const DBT *target_key = extra.key;
    return key_msn_cmp(query_key, target_key, query->msn, extra.msn,
                       extra.desc, extra.cmp);
1863 1864 1865
}

int
1866
toku_fifo_entry_key_msn_cmp(const struct toku_fifo_entry_key_msn_cmp_extra &extra, const int32_t &ao, const int32_t &bo)
1867
{
1868 1869
    const struct fifo_entry *a = toku_fifo_get_entry(extra.fifo, ao);
    const struct fifo_entry *b = toku_fifo_get_entry(extra.fifo, bo);
1870 1871 1872 1873
    DBT adbt, bdbt;
    const DBT *akey = fill_dbt_for_fifo_entry(&adbt, a);
    const DBT *bkey = fill_dbt_for_fifo_entry(&bdbt, b);
    return key_msn_cmp(akey, bkey, a->msn, b->msn,
1874
                       extra.desc, extra.cmp);
1875 1876
}

1877
void toku_bnc_insert_msg(NONLEAF_CHILDINFO bnc, const void *key, ITEMLEN keylen, const void *data, ITEMLEN datalen, enum ft_msg_type type, MSN msn, XIDS xids, bool is_fresh, DESCRIPTOR desc, ft_compare_func cmp)
1878 1879 1880
// Effect: Enqueue the message represented by the parameters into the
//   bnc's buffer, and put it in either the fresh or stale message tree,
//   or the broadcast list.
1881 1882
//
// This is only exported for tests.
1883
{
1884
    int32_t offset;
1885 1886
    int r = toku_fifo_enq(bnc->buffer, key, keylen, data, datalen, type, msn, xids, is_fresh, &offset);
    assert_zero(r);
1887
    if (ft_msg_type_applies_once(type)) {
1888 1889
        DBT keydbt;
        struct toku_fifo_entry_key_msn_heaviside_extra extra = { .desc = desc, .cmp = cmp, .fifo = bnc->buffer, .key = toku_fill_dbt(&keydbt, key, keylen), .msn = msn };
1890
        if (is_fresh) {
1891
            r = bnc->fresh_message_tree.insert<struct toku_fifo_entry_key_msn_heaviside_extra, toku_fifo_entry_key_msn_heaviside>(offset, extra, nullptr);
1892
            assert_zero(r);
1893
        } else {
1894
            r = bnc->stale_message_tree.insert<struct toku_fifo_entry_key_msn_heaviside_extra, toku_fifo_entry_key_msn_heaviside>(offset, extra, nullptr);
1895
            assert_zero(r);
1896
        }
1897 1898
    } else {
        invariant(ft_msg_type_applies_all(type) || ft_msg_type_does_nothing(type));
1899 1900
        const uint32_t idx = bnc->broadcast_list.size();
        r = bnc->broadcast_list.insert_at(offset, idx);
1901
        assert_zero(r);
1902
    }
1903 1904 1905 1906
}

// append a cmd to a nonleaf node's child buffer
// should be static, but used by test programs
1907
void toku_ft_append_to_child_buffer(ft_compare_func compare_fun, DESCRIPTOR desc, FTNODE node, int childnum, enum ft_msg_type type, MSN msn, XIDS xids, bool is_fresh, const DBT *key, const DBT *val) {
1908
    paranoid_invariant(BP_STATE(node,childnum) == PT_AVAIL);
1909
    toku_bnc_insert_msg(BNC(node, childnum), key->data, key->size, val->data, val->size, type, msn, xids, is_fresh, desc, compare_fun);
1910
    node->dirty = 1;
1911 1912
}

1913
static void ft_nonleaf_cmd_once_to_child(ft_compare_func compare_fun, DESCRIPTOR desc, FTNODE node, int target_childnum, FT_MSG cmd, bool is_fresh, size_t flow_deltas[])
1914 1915 1916
// Previously we had passive aggressive promotion, but that causes a lot of I/O a the checkpoint.  So now we are just putting it in the buffer here.
// Also we don't worry about the node getting overfull here.  It's the caller's problem.
{
1917 1918 1919
    unsigned int childnum = (target_childnum >= 0
                             ? target_childnum
                             : toku_ftnode_which_child(node, cmd->u.id.key, desc, compare_fun));
1920
    toku_ft_append_to_child_buffer(compare_fun, desc, node, childnum, cmd->type, cmd->msn, cmd->xids, is_fresh, cmd->u.id.key, cmd->u.id.val);
1921 1922 1923
    NONLEAF_CHILDINFO bnc = BNC(node, childnum);
    bnc->flow[0] += flow_deltas[0];
    bnc->flow[1] += flow_deltas[1];
1924
}
1925

1926 1927 1928 1929
/* Find the leftmost child that may contain the key.
 * If the key exists it will be in the child whose number
 * is the return value of this function.
 */
1930 1931
int toku_ftnode_which_child(FTNODE node, const DBT *k,
                            DESCRIPTOR desc, ft_compare_func cmp) {
1932 1933
#define DO_PIVOT_SEARCH_LR 0
#if DO_PIVOT_SEARCH_LR
1934
    int i;
1935
    for (i=0; i<node->n_children-1; i++) {
1936
        int c = ft_compare_pivot(desc, cmp, k, d, &node->childkeys[i]);
1937 1938 1939
        if (c > 0) continue;
        if (c < 0) return i;
        return i;
1940
    }
1941
    return node->n_children-1;
1942
#else
1943 1944 1945
#endif
#define DO_PIVOT_SEARCH_RL 0
#if DO_PIVOT_SEARCH_RL
1946
    // give preference for appending to the dictionary.         no change for
1947
    // random keys
1948
    int i;
1949
    for (i = node->n_children-2; i >= 0; i--) {
1950
        int c = ft_compare_pivot(desc, cmp, k, d, &node->childkeys[i]);
1951
        if (c > 0) return i+1;
1952 1953 1954
    }
    return 0;
#endif
1955 1956 1957
#define DO_PIVOT_BIN_SEARCH 1
#if DO_PIVOT_BIN_SEARCH
    // a funny case of no pivots
1958
    if (node->n_children <= 1) return 0;
1959 1960

    // check the last key to optimize seq insertions
1961
    int n = node->n_children-1;
1962
    int c = ft_compare_pivot(desc, cmp, k, &node->childkeys[n-1]);
1963
    if (c > 0) return n;
1964 1965 1966 1967 1968 1969

    // binary search the pivots
    int lo = 0;
    int hi = n-1; // skip the last one, we checked it above
    int mi;
    while (lo < hi) {
1970
        mi = (lo + hi) / 2;
1971
        c = ft_compare_pivot(desc, cmp, k, &node->childkeys[mi]);
1972 1973 1974 1975 1976 1977 1978 1979 1980
        if (c > 0) {
            lo = mi+1;
            continue;
        }
        if (c < 0) {
            hi = mi;
            continue;
        }
        return mi;
1981 1982 1983
    }
    return lo;
#endif
1984 1985
}

Leif Walsh's avatar
Leif Walsh committed
1986
// Used for HOT.
1987
int
1988
toku_ftnode_hot_next_child(FTNODE node,
1989 1990 1991
                           const DBT *k,
                           DESCRIPTOR desc,
                           ft_compare_func cmp) {
Leif Walsh's avatar
Leif Walsh committed
1992 1993 1994 1995 1996
    int low = 0;
    int hi = node->n_children - 1;
    int mi;
    while (low < hi) {
        mi = (low + hi) / 2;
1997
        int r = ft_compare_pivot(desc, cmp, k, &node->childkeys[mi]);
Leif Walsh's avatar
Leif Walsh committed
1998 1999 2000 2001 2002 2003 2004 2005 2006 2007 2008 2009 2010 2011
        if (r > 0) {
            low = mi + 1;
        } else if (r < 0) {
            hi = mi;
        } else {
            // if they were exactly equal, then we want the sub-tree under
            // the next pivot.
            return mi + 1;
        }
    }
    invariant(low == hi);
    return low;
}

2012 2013 2014
// TODO Use this function to clean up other places where bits of messages are passed around
//      such as toku_bnc_insert_msg() and the call stack above it.
static size_t
2015
ft_msg_size(FT_MSG msg) {
2016 2017 2018
    size_t keylen = msg->u.id.key->size;
    size_t vallen = msg->u.id.val->size;
    size_t xids_size = xids_get_serialize_size(msg->xids);
2019
    size_t rval = keylen + vallen + KEY_VALUE_OVERHEAD + FT_CMD_OVERHEAD + xids_size;
2020 2021 2022 2023
    return rval;
}


2024
static void
2025
ft_nonleaf_cmd_all (ft_compare_func compare_fun, DESCRIPTOR desc, FTNODE node, FT_MSG cmd, bool is_fresh, size_t flow_deltas[])
2026 2027
// Effect: Put the cmd into a nonleaf node.  We put it into all children, possibly causing the children to become reactive.
//  We don't do the splitting and merging.  That's up to the caller after doing all the puts it wants to do.
2028
//  The re_array[i] gets set to the reactivity of any modified child i.         (And there may be several such children.)
2029
{
2030 2031
    for (int i = 0; i < node->n_children; i++) {
        ft_nonleaf_cmd_once_to_child(compare_fun, desc, node, i, cmd, is_fresh, flow_deltas);
2032 2033 2034
    }
}

Yoni Fogel's avatar
Yoni Fogel committed
2035
static bool
2036
ft_msg_applies_once(FT_MSG cmd)
2037
{
2038
    return ft_msg_type_applies_once(cmd->type);
2039 2040
}

Yoni Fogel's avatar
Yoni Fogel committed
2041
static bool
2042
ft_msg_applies_all(FT_MSG cmd)
2043
{
2044
    return ft_msg_type_applies_all(cmd->type);
2045 2046
}

Yoni Fogel's avatar
Yoni Fogel committed
2047
static bool
2048
ft_msg_does_nothing(FT_MSG cmd)
2049
{
2050
    return ft_msg_type_does_nothing(cmd->type);
2051 2052 2053
}

static void
2054
ft_nonleaf_put_cmd (ft_compare_func compare_fun, DESCRIPTOR desc, FTNODE node, int target_childnum, FT_MSG cmd, bool is_fresh, size_t flow_deltas[])
2055 2056
// Effect: Put the cmd into a nonleaf node.  We may put it into a child, possibly causing the child to become reactive.
//  We don't do the splitting and merging.  That's up to the caller after doing all the puts it wants to do.
2057
//  The re_array[i] gets set to the reactivity of any modified child i.         (And there may be several such children.)
2058 2059
//
{
2060 2061

    //
2062
    // see comments in toku_ft_leaf_apply_cmd
2063 2064
    // to understand why we handle setting
    // node->max_msn_applied_to_node_on_disk here,
2065
    // and don't do it in toku_ft_node_put_cmd
2066
    //
2067
    MSN cmd_msn = cmd->msn;
2068 2069
    invariant(cmd_msn.msn > node->max_msn_applied_to_node_on_disk.msn);
    node->max_msn_applied_to_node_on_disk = cmd_msn;
2070

2071 2072 2073 2074 2075 2076
    if (ft_msg_applies_once(cmd)) {
        ft_nonleaf_cmd_once_to_child(compare_fun, desc, node, target_childnum, cmd, is_fresh, flow_deltas);
    } else if (ft_msg_applies_all(cmd)) {
        ft_nonleaf_cmd_all(compare_fun, desc, node, cmd, is_fresh, flow_deltas);
    } else {
        paranoid_invariant(ft_msg_applies_none(cmd));
2077 2078 2079
    }
}

2080 2081
// Garbage collect one leaf entry.
static void
2082
ft_basement_node_gc_once(BASEMENTNODE bn,
Yoni Fogel's avatar
Yoni Fogel committed
2083
                          uint32_t index,
2084
                          LEAFENTRY leaf_entry,
2085 2086 2087
                          const xid_omt_t &snapshot_xids,
                          const rx_omt_t &referenced_xids,
                          const xid_omt_t &live_root_txns,
2088 2089
                          STAT64INFO_S * delta)
{
2090
    paranoid_invariant(leaf_entry);
2091 2092 2093 2094 2095 2096 2097

    // There is no point in running GC if there is only one committed
    // leaf entry.
    if (leaf_entry->type != LE_MVCC || leaf_entry->u.mvcc.num_cxrs <= 1) { // MAKE ACCESSOR
        goto exit;
    }

2098 2099 2100 2101 2102 2103
    size_t oldsize;
    oldsize = 0;
    size_t newsize;
    newsize = 0;
    LEAFENTRY new_leaf_entry;
    new_leaf_entry = NULL;
2104 2105 2106 2107

    // The mempool doesn't free itself.  When it allocates new memory,
    // this pointer will be set to the older memory that must now be
    // freed.
2108 2109
    void * maybe_free;
    maybe_free = NULL;
2110 2111 2112 2113 2114 2115

    // Cache the size of the leaf entry.
    oldsize = leafentry_memsize(leaf_entry);
    garbage_collect_leafentry(leaf_entry,
                              &new_leaf_entry,
                              &newsize,
2116
                              &bn->buffer,
2117 2118 2119
                              &bn->buffer_mempool,
                              &maybe_free,
                              snapshot_xids,
2120
                              referenced_xids,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
2121
                              live_root_txns);
2122 2123 2124

    // These will represent the number of bytes and rows changed as
    // part of the garbage collection.
2125 2126 2127 2128
    int64_t numbytes_delta;
    numbytes_delta = newsize - oldsize;
    int64_t numrows_delta;
    numrows_delta = 0;
2129 2130 2131 2132 2133 2134 2135 2136 2137 2138 2139
    if (new_leaf_entry) {
        // If we have a new leaf entry, we must update the size of the
        // memory object.
        bn->n_bytes_in_buffer -= oldsize;
        bn->n_bytes_in_buffer += newsize;
        toku_mempool_mfree(&bn->buffer_mempool, 0, oldsize);
        toku_omt_set_at(bn->buffer, new_leaf_entry, index);
        numrows_delta = 0;
    } else {
        // Our garbage collection removed the leaf entry so we must
        // remove it from the mempool.
2140
        ft_leaf_delete_leafentry (bn, index, leaf_entry);
2141 2142 2143 2144 2145 2146 2147 2148 2149 2150 2151 2152 2153 2154 2155
        numrows_delta = -1;
    }

    // If we created a new mempool buffer we must free the
    // old/original buffer.
    if (maybe_free) {
        toku_free(maybe_free);
    }

    // Update stats.
    bn->stat64_delta.numrows += numrows_delta;
    bn->stat64_delta.numbytes += numbytes_delta;
    delta->numrows += numrows_delta;
    delta->numbytes += numbytes_delta;

2156
exit:
2157 2158 2159 2160 2161 2162
    return;
}

// Garbage collect all leaf entries for a given basement node.
static void
basement_node_gc_all_les(BASEMENTNODE bn,
2163 2164 2165
                         const xid_omt_t &snapshot_xids,
                         const rx_omt_t &referenced_xids,
                         const xid_omt_t &live_root_txns,
2166 2167 2168
                         STAT64INFO_S * delta)
{
    int r = 0;
Yoni Fogel's avatar
Yoni Fogel committed
2169 2170
    uint32_t index = 0;
    uint32_t num_leafentries_before;
2171 2172 2173 2174
    while (index < (num_leafentries_before = toku_omt_size(bn->buffer))) {
        OMTVALUE storedatav = NULL;
        LEAFENTRY leaf_entry;
        r = toku_omt_fetch(bn->buffer, index, &storedatav);
2175
        assert_zero(r);
2176
        CAST_FROM_VOIDP(leaf_entry, storedatav);
2177
        ft_basement_node_gc_once(bn, index, leaf_entry, snapshot_xids, referenced_xids, live_root_txns, delta);
2178 2179 2180 2181 2182 2183 2184 2185 2186
        // Check if the leaf entry was deleted or not.
        if (num_leafentries_before == toku_omt_size(bn->buffer)) {
            ++index;
        }
    }
}

// Garbage collect all leaf entires.
static void
2187 2188
ft_leaf_gc_all_les(FTNODE node,
                    FT h,
2189 2190 2191
                    const xid_omt_t &snapshot_xids,
                    const rx_omt_t &referenced_xids,
                    const xid_omt_t &live_root_txns)
2192 2193
{
    toku_assert_entire_node_in_memory(node);
2194
    paranoid_invariant_zero(node->height);
2195 2196 2197 2198 2199 2200 2201
    // Loop through each leaf entry, garbage collecting as we go.
    for (int i = 0; i < node->n_children; ++i) {
        // Perform the garbage collection.
        BASEMENTNODE bn = BLB(node, i);
        STAT64INFO_S delta;
        delta.numrows = 0;
        delta.numbytes = 0;
2202
        basement_node_gc_all_les(bn, snapshot_xids, referenced_xids, live_root_txns, &delta);
2203
        toku_ft_update_stats(&h->in_memory_stats, delta);
2204 2205 2206
    }
}

2207
void toku_bnc_flush_to_child(
2208
    FT h,
2209
    NONLEAF_CHILDINFO bnc,
2210
    FTNODE child
Zardosht Kasheff's avatar
Zardosht Kasheff committed
2211
    )
2212
{
2213 2214
    paranoid_invariant(bnc);
    paranoid_invariant(toku_fifo_n_entries(bnc->buffer)>0);
2215
    STAT64INFO_S stats_delta = {0,0};
2216
    size_t remaining_memsize = toku_fifo_buffer_size_in_use(bnc->buffer);
2217 2218 2219 2220
    FIFO_ITERATE(
        bnc->buffer, key, keylen, val, vallen, type, msn, xids, is_fresh,
        ({
            DBT hk,hv;
2221 2222
            FT_MSG_S ftcmd = { type, msn, xids, .u = { .id = { toku_fill_dbt(&hk, key, keylen),
                                                               toku_fill_dbt(&hv, val, vallen) } } };
2223 2224 2225 2226 2227 2228 2229 2230 2231 2232
            size_t flow_deltas[] = { 0, 0 };
            if (remaining_memsize <= bnc->flow[0]) {
                // this message is in the current checkpoint's worth of
                // the end of the fifo
                flow_deltas[0] = FIFO_CURRENT_ENTRY_MEMSIZE;
            } else if (remaining_memsize <= bnc->flow[0] + bnc->flow[1]) {
                // this message is in the last checkpoint's worth of the
                // end of the fifo
                flow_deltas[1] = FIFO_CURRENT_ENTRY_MEMSIZE;
            }
2233
            toku_ft_node_put_cmd(
2234 2235 2236
                h->compare_fun,
                h->update_fun,
                &h->cmp_descriptor,
2237
                child,
2238
                -1,
2239
                &ftcmd,
2240
                is_fresh,
2241
                flow_deltas,
2242
                &stats_delta
Zardosht Kasheff's avatar
Zardosht Kasheff committed
2243
                );
2244
            remaining_memsize -= FIFO_CURRENT_ENTRY_MEMSIZE;
2245
        }));
2246
    invariant(remaining_memsize == 0);
2247
    if (stats_delta.numbytes || stats_delta.numrows) {
2248
        toku_ft_update_stats(&h->in_memory_stats, stats_delta);
2249
    }
2250
    // Run garbage collection, if we are a leaf entry.
2251
    TOKULOGGER logger = toku_cachefile_logger(h->cf);
2252
    if (child->height == 0 && logger) {
2253 2254 2255
        xid_omt_t snapshot_txnids;
        rx_omt_t referenced_xids;
        xid_omt_t live_root_txns;
2256 2257
        toku_txn_manager_clone_state_for_gc(
            logger->txn_manager,
2258 2259 2260
            snapshot_txnids,
            referenced_xids,
            live_root_txns
2261 2262
            );
        size_t buffsize = toku_fifo_buffer_size_in_use(bnc->buffer);
2263
        STATUS_INC(FT_MSG_BYTES_OUT, buffsize);
2264
        // may be misleading if there's a broadcast message in there
2265
        STATUS_INC(FT_MSG_BYTES_CURR, -buffsize);
2266
        // Perform the garbage collection.
2267
        ft_leaf_gc_all_les(child, h, snapshot_txnids, referenced_xids, live_root_txns);
2268 2269

        // Free the OMT's we used for garbage collecting.
2270 2271 2272
        snapshot_txnids.destroy();
        live_root_txns.destroy();
        referenced_xids.destroy();
Zardosht Kasheff's avatar
Zardosht Kasheff committed
2273
    }
2274 2275
}

2276 2277 2278 2279 2280 2281
bool toku_bnc_should_promote(FT ft, NONLEAF_CHILDINFO bnc) {
    static const double factor = 0.125;
    const uint64_t flow_threshold = ft->h->nodesize * factor;
    return bnc->flow[0] >= flow_threshold || bnc->flow[1] >= flow_threshold;
}

2282
void bring_node_fully_into_memory(FTNODE node, FT h)
Zardosht Kasheff's avatar
Zardosht Kasheff committed
2283
{
2284
    if (!is_entire_node_in_memory(node)) {
2285
        struct ftnode_fetch_extra bfe;
2286
        fill_bfe_for_full_read(&bfe, h);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
2287 2288
        toku_cachetable_pf_pinned_pair(
            node,
2289
            toku_ftnode_pf_callback,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
2290 2291 2292 2293 2294
            &bfe,
            h->cf,
            node->thisnodename,
            toku_cachetable_hash(h->cf, node->thisnodename)
            );
2295
    }
Zardosht Kasheff's avatar
Zardosht Kasheff committed
2296 2297
}

Leif Walsh's avatar
Leif Walsh committed
2298
void
2299 2300 2301
toku_ft_node_put_cmd (
    ft_compare_func compare_fun,
    ft_update_func update_fun,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
2302
    DESCRIPTOR desc,
2303
    FTNODE node,
2304
    int target_childnum,
2305
    FT_MSG cmd,
2306
    bool is_fresh,
2307
    size_t flow_deltas[],
2308
    STAT64INFO stats_to_update
Zardosht Kasheff's avatar
Zardosht Kasheff committed
2309
    )
2310 2311
// Effect: Push CMD into the subtree rooted at NODE.
//   If NODE is a leaf, then
2312
//   put CMD into leaf, applying it to the leafentries
2313 2314
//   If NODE is a nonleaf, then push the cmd into the FIFO(s) of the relevent child(ren).
//   The node may become overfull.  That's not our problem.
2315
{
2316
    toku_assert_entire_node_in_memory(node);
2317
    //
2318
    // see comments in toku_ft_leaf_apply_cmd
2319 2320 2321 2322
    // to understand why we don't handle setting
    // node->max_msn_applied_to_node_on_disk here,
    // and instead defer to these functions
    //
2323
    if (node->height==0) {
2324
        toku_ft_leaf_apply_cmd(compare_fun, update_fun, desc, node, target_childnum, cmd, nullptr, stats_to_update);
2325
    } else {
2326
        ft_nonleaf_put_cmd(compare_fun, desc, node, target_childnum, cmd, is_fresh, flow_deltas);
2327
    }
2328 2329
}

2330
static const struct pivot_bounds infinite_bounds = {.lower_bound_exclusive=NULL,
2331
                                                    .upper_bound_inclusive=NULL};
2332

Zardosht Kasheff's avatar
Zardosht Kasheff committed
2333

2334
// Effect: applies the cmd to the leaf if the appropriate basement node is in memory.
2335 2336
//           This function is called during message injection and/or flushing, so the entire
//           node MUST be in memory.
2337 2338 2339
void toku_ft_leaf_apply_cmd(
    ft_compare_func compare_fun,
    ft_update_func update_fun,
2340
    DESCRIPTOR desc,
2341
    FTNODE node,
2342
    int target_childnum,  // which child to inject to, or -1 if unknown
2343
    FT_MSG cmd,
2344 2345
    uint64_t *workdone,
    STAT64INFO stats_to_update
2346
    )
Zardosht Kasheff's avatar
Zardosht Kasheff committed
2347
{
2348
    VERIFY_NODE(t, node);
2349
    toku_assert_entire_node_in_memory(node);
2350

2351
    //
2352
    // Because toku_ft_leaf_apply_cmd is called with the intent of permanently
2353 2354
    // applying a message to a leaf node (meaning the message is permanently applied
    // and will be purged from the system after this call, as opposed to
2355
    // toku_apply_ancestors_messages_to_node, which applies a message
2356 2357
    // for a query, but the message may still reside in the system and
    // be reapplied later), we mark the node as dirty and
2358 2359
    // take the opportunity to update node->max_msn_applied_to_node_on_disk.
    //
2360
    node->dirty = 1;
2361

2362 2363 2364
    //
    // we cannot blindly update node->max_msn_applied_to_node_on_disk,
    // we must check to see if the msn is greater that the one already stored,
2365
    // because the cmd may have already been applied earlier (via
2366
    // toku_apply_ancestors_messages_to_node) to answer a query
2367
    //
2368
    // This is why we handle node->max_msn_applied_to_node_on_disk both here
2369
    // and in ft_nonleaf_put_cmd, as opposed to in one location, toku_ft_node_put_cmd.
2370
    //
2371
    MSN cmd_msn = cmd->msn;
2372 2373 2374
    if (cmd_msn.msn > node->max_msn_applied_to_node_on_disk.msn) {
        node->max_msn_applied_to_node_on_disk = cmd_msn;
    }
2375

2376
    if (ft_msg_applies_once(cmd)) {
2377 2378 2379 2380 2381 2382
        unsigned int childnum = (target_childnum >= 0
                                 ? target_childnum
                                 : toku_ftnode_which_child(node, cmd->u.id.key, desc, compare_fun));
        BASEMENTNODE bn = BLB(node, childnum);
        if (cmd->msn.msn > bn->max_msn_applied.msn) {
            bn->max_msn_applied = cmd->msn;
2383
            toku_ft_bn_apply_cmd(compare_fun,
2384 2385 2386 2387 2388 2389
                                 update_fun,
                                 desc,
                                 bn,
                                 cmd,
                                 workdone,
                                 stats_to_update);
2390
        } else {
2391
            STATUS_INC(FT_MSN_DISCARDS, 1);
2392 2393
        }
    }
2394
    else if (ft_msg_applies_all(cmd)) {
2395
        for (int childnum=0; childnum<node->n_children; childnum++) {
2396 2397
            if (cmd->msn.msn > BLB(node, childnum)->max_msn_applied.msn) {
                BLB(node, childnum)->max_msn_applied = cmd->msn;
2398
                toku_ft_bn_apply_cmd(compare_fun,
2399 2400 2401 2402 2403 2404
                                      update_fun,
                                      desc,
                                      BLB(node, childnum),
                                      cmd,
                                      workdone,
                                      stats_to_update);
2405
            } else {
2406
                STATUS_INC(FT_MSN_DISCARDS, 1);
2407
            }
2408
        }
2409
    }
2410
    else if (!ft_msg_does_nothing(cmd)) {
2411
        abort();
2412 2413 2414 2415
    }
    VERIFY_NODE(t, node);
}

2416 2417 2418 2419 2420 2421
static void inject_message_in_locked_node(FT ft, FTNODE node, int childnum, FT_MSG_S *cmd, size_t flow_deltas[]) {
    // No guarantee that we're the writer, but oh well.
    // TODO(leif): Implement "do I have the lock or is it someone else?"
    // check in frwlock.  Should be possible with TOKU_PTHREAD_DEBUG, nop
    // otherwise.
    invariant(toku_ctpair_is_write_locked(node->ct_pair));
2422
    toku_assert_entire_node_in_memory(node);
2423 2424 2425 2426 2427
    // Get the MSN from the header.  Now that we have a write lock on the
    // node we're injecting into, we know no other thread will get an MSN
    // after us and get that message into our subtree before us.
    cmd->msn.msn = toku_sync_add_and_fetch(&ft->h->max_msn_in_ft.msn, 1);
    paranoid_invariant(cmd->msn.msn > node->max_msn_applied_to_node_on_disk.msn);
2428
    STAT64INFO_S stats_delta = {0,0};
2429
    toku_ft_node_put_cmd(
2430 2431 2432
        ft->compare_fun,
        ft->update_fun,
        &ft->cmp_descriptor,
2433
        node,
2434
        childnum,
2435
        cmd,
2436
        true,
2437
        flow_deltas,
2438
        &stats_delta
2439
        );
2440
    if (stats_delta.numbytes || stats_delta.numrows) {
2441
        toku_ft_update_stats(&ft->in_memory_stats, stats_delta);
2442
    }
2443
    //
2444
    // assumption is that toku_ft_node_put_cmd will
2445 2446 2447
    // mark the node as dirty.
    // enforcing invariant here.
    //
2448
    paranoid_invariant(node->dirty != 0);
2449 2450 2451

    // update some status variables
    if (node->height != 0) {
2452
        uint64_t msgsize = ft_msg_size(cmd);
2453 2454 2455 2456 2457 2458
        STATUS_INC(FT_MSG_BYTES_IN, msgsize);
        STATUS_INC(FT_MSG_BYTES_CURR, msgsize);
        //if (STATUS_VALUE(FT_MSG_BYTES_CURR) > STATUS_VALUE(FT_MSG_BYTES_MAX)) {
        //    STATUS_VALUE(FT_MSG_BYTES_MAX) = STATUS_VALUE(FT_MSG_BYTES_CURR);
        //}
        STATUS_INC(FT_MSG_NUM, 1);
2459
        if (ft_msg_applies_all(cmd)) {
2460
            STATUS_INC(FT_MSG_NUM_BROADCAST, 1);
2461
        }
2462
    }
2463 2464 2465 2466 2467 2468 2469 2470 2471 2472 2473 2474

    // verify that msn of latest message was captured in root node
    paranoid_invariant(cmd->msn.msn == node->max_msn_applied_to_node_on_disk.msn);

    // if we call flush_some_child, then that function unpins the root
    // otherwise, we unpin ourselves
    if (node->height > 0 && toku_ft_nonleaf_is_gorged(node, ft->h->nodesize)) {
        flush_node_on_background_thread(ft, node);
    }
    else {
        toku_unpin_ftnode(ft, node);
    }
Bradley C. Kuszmaul's avatar
Rename  
Bradley C. Kuszmaul committed
2475 2476
}

2477 2478 2479 2480 2481 2482 2483 2484 2485
// seqinsert_loc is a bitmask.
// The root counts as being both on the "left extreme" and on the "right extreme".
// Therefore, at the root, you're at LEFT_EXTREME | RIGHT_EXTREME.
typedef char seqinsert_loc;
static const seqinsert_loc NEITHER_EXTREME = 0;
static const seqinsert_loc LEFT_EXTREME = 1;
static const seqinsert_loc RIGHT_EXTREME = 2;

static bool process_maybe_reactive_child(FT ft, FTNODE parent, FTNODE child, int childnum, seqinsert_loc loc)
2486
// Effect:
2487 2488 2489 2490 2491 2492 2493
//  If child needs to be split or merged, do that.
//  parent and child will be unlocked if this happens
//  also, the batched pin will have ended if this happens
// Requires: parent and child are read locked
// Returns:
//  true if relocking is needed
//  false otherwise
2494
{
2495 2496 2497 2498 2499 2500 2501 2502 2503 2504 2505 2506 2507 2508 2509 2510 2511 2512 2513 2514 2515 2516 2517 2518 2519 2520 2521 2522 2523 2524 2525 2526 2527 2528 2529 2530 2531 2532 2533 2534 2535 2536 2537 2538 2539 2540 2541 2542 2543 2544 2545 2546 2547 2548 2549 2550 2551 2552 2553 2554 2555 2556 2557 2558 2559 2560 2561 2562 2563 2564 2565 2566 2567 2568 2569 2570 2571 2572 2573 2574 2575 2576 2577 2578 2579 2580 2581 2582 2583 2584 2585 2586 2587 2588 2589 2590 2591 2592 2593
    enum reactivity re = get_node_reactivity(child, ft->h->nodesize);
    enum reactivity newre;
    BLOCKNUM child_blocknum;
    uint32_t child_fullhash;
    switch (re) {
    case RE_STABLE:
        return false;
    case RE_FISSIBLE:
        {
            // We only have a read lock on the parent.  We need to drop both locks, and get write locks.
            BLOCKNUM parent_blocknum = parent->thisnodename;
            uint32_t parent_fullhash = toku_cachetable_hash(ft->cf, parent_blocknum);
            int parent_height = parent->height;
            int parent_n_children = parent->n_children;
            toku_unpin_ftnode_read_only(ft, child);
            toku_unpin_ftnode_read_only(ft, parent);
            struct ftnode_fetch_extra bfe;
            fill_bfe_for_full_read(&bfe, ft);
            FTNODE newparent, newchild;
            toku_pin_ftnode_off_client_thread_batched(ft, parent_blocknum, parent_fullhash, &bfe, PL_WRITE_CHEAP, 0, nullptr, &newparent);
            if (newparent->height != parent_height || newparent->n_children != parent_n_children ||
                childnum >= newparent->n_children || toku_bnc_n_entries(BNC(newparent, childnum))) {
                // If the height changed or childnum is now off the end, something clearly got split or merged out from under us.
                // If something got injected in this node, then it got split or merged and we shouldn't be splitting it.
                // But we already unpinned the child so we need to have the caller re-try the pins.
                toku_unpin_ftnode_read_only(ft, newparent);
                return true;
            }
            // It's ok to reuse the same childnum because if we get something
            // else we need to split, well, that's crazy, but let's go ahead
            // and split it.
            child_blocknum = BP_BLOCKNUM(newparent, childnum);
            child_fullhash = compute_child_fullhash(ft->cf, newparent, childnum);
            toku_pin_ftnode_off_client_thread_batched(ft, child_blocknum, child_fullhash, &bfe, PL_WRITE_CHEAP, 1, &newparent, &newchild);
            newre = get_node_reactivity(newchild, ft->h->nodesize);
            if (newre == RE_FISSIBLE) {
                enum split_mode split_mode;
                if (newparent->height == 1 && (loc & LEFT_EXTREME) && childnum == 0) {
                    split_mode = SPLIT_RIGHT_HEAVY;
                } else if (newparent->height == 1 && (loc & RIGHT_EXTREME) && childnum == newparent->n_children - 1) {
                    split_mode = SPLIT_LEFT_HEAVY;
                } else {
                    split_mode = SPLIT_EVENLY;
                }
                toku_ft_split_child(ft, newparent, childnum, newchild, split_mode);
            } else {
                // some other thread already got it, just unpin and tell the
                // caller to retry
                toku_unpin_ftnode_read_only(ft, newchild);
                toku_unpin_ftnode_read_only(ft, newparent);
            }
            return true;
        }
    case RE_FUSIBLE:
        {
            if (parent->height == 1) {
                // prevent re-merging of recently unevenly-split nodes
                if (((loc & LEFT_EXTREME) && childnum <= 1) ||
                    ((loc & RIGHT_EXTREME) && childnum >= parent->n_children - 2)) {
                    return false;
                }
            }

            int parent_height = parent->height;
            BLOCKNUM parent_blocknum = parent->thisnodename;
            uint32_t parent_fullhash = toku_cachetable_hash(ft->cf, parent_blocknum);
            toku_unpin_ftnode_read_only(ft, child);
            toku_unpin_ftnode_read_only(ft, parent);
            struct ftnode_fetch_extra bfe;
            fill_bfe_for_full_read(&bfe, ft);
            FTNODE newparent, newchild;
            toku_pin_ftnode_off_client_thread_batched(ft, parent_blocknum, parent_fullhash, &bfe, PL_WRITE_CHEAP, 0, nullptr, &newparent);
            if (newparent->height != parent_height || childnum >= newparent->n_children) {
                // looks like this is the root and it got merged, let's just start over (like in the split case above)
                toku_unpin_ftnode_read_only(ft, newparent);
                return true;
            }
            child_blocknum = BP_BLOCKNUM(newparent, childnum);
            child_fullhash = compute_child_fullhash(ft->cf, newparent, childnum);
            toku_pin_ftnode_off_client_thread_batched(ft, child_blocknum, child_fullhash, &bfe, PL_READ, 1, &newparent, &newchild);
            newre = get_node_reactivity(newchild, ft->h->nodesize);
            if (newre == RE_FUSIBLE) {
                if (newparent->n_children < 2) {
                    // weird case where newparent is the root node and
                    // can't merge, so it might only have one child.  In
                    // this case, just return false so we can insert
                    // anyway.
                    return false;
                }
                toku_unpin_ftnode_read_only(ft, newchild);
                toku_ft_merge_child(ft, newparent, childnum);
            } else {
                // some other thread already got it, just unpin and tell the
                // caller to retry
                toku_unpin_ftnode_read_only(ft, newchild);
                toku_unpin_ftnode_read_only(ft, newparent);
            }
            return true;
        }
2594
    }
2595 2596
    abort();
}
2597

2598 2599 2600 2601 2602
static void inject_message_at_this_blocknum(FT ft, CACHEKEY cachekey, uint32_t fullhash, FT_MSG_S *cmd, size_t flow_deltas[])
// Effect:
//  Inject cmd into the node at this blocknum (cachekey).
//  Gets a write lock on the node for you.
{
2603
    FTNODE node;
2604 2605 2606 2607 2608 2609 2610 2611
    struct ftnode_fetch_extra bfe;
    fill_bfe_for_full_read(&bfe, ft);
    toku_pin_ftnode_off_client_thread_batched(ft, cachekey, fullhash, &bfe, PL_WRITE_CHEAP, 0, NULL, &node);
    toku_assert_entire_node_in_memory(node);
    paranoid_invariant(node->fullhash==fullhash);
    ft_verify_flags(ft, node);
    inject_message_in_locked_node(ft, node, -1, cmd, flow_deltas);
}
Zardosht Kasheff's avatar
Zardosht Kasheff committed
2612

2613 2614 2615 2616 2617 2618 2619 2620 2621 2622 2623 2624 2625 2626 2627 2628 2629 2630 2631 2632 2633 2634 2635 2636 2637 2638 2639 2640 2641 2642 2643 2644 2645 2646 2647 2648 2649 2650 2651 2652 2653 2654 2655 2656 2657 2658 2659 2660 2661 2662 2663 2664 2665 2666 2667 2668 2669 2670 2671 2672 2673 2674 2675 2676 2677 2678 2679 2680 2681 2682 2683 2684 2685 2686 2687 2688 2689 2690 2691 2692 2693 2694 2695 2696 2697 2698 2699 2700 2701 2702 2703 2704 2705 2706 2707 2708 2709 2710 2711 2712 2713 2714 2715 2716 2717 2718 2719 2720 2721 2722 2723 2724 2725 2726 2727 2728 2729 2730 2731 2732 2733 2734 2735 2736 2737 2738 2739 2740 2741 2742 2743 2744 2745 2746 2747 2748 2749 2750 2751 2752
__attribute__((const))
static inline bool should_inject_in_node(seqinsert_loc loc, int height, int depth)
// We should inject directly in a node if:
//  - it's a leaf, or
//  - it's a height 1 node not at either extreme, or
//  - it's a depth 2 node not at either extreme
{
    return (height == 0 || (loc == NEITHER_EXTREME && (height <= 1 || depth >= 2)));
}

static void push_something_in_subtree(FT ft, FTNODE subtree_root, int target_childnum, FT_MSG_S *cmd, size_t flow_deltas[], int depth, seqinsert_loc loc, bool just_did_split_or_merge)
// Effects:
//  Assign cmd an MSN from ft->h.
//  Put cmd in the subtree rooted at node.  Due to promotion the message may not be injected directly in this node.
//  Unlock node or schedule it to be unlocked (after a background flush).
//   Either way, the caller is not responsible for unlocking node.
// Requires:
//  subtree_root is read locked and fully in memory.
// Notes:
//  In Ming, the basic rules of promotion are as follows:
//   Don't promote broadcast messages.
//   Don't promote past non-empty buffers.
//   Otherwise, promote at most to height 1 or depth 2 (whichever is highest), as far as the birdie asks you to promote.
//    We don't promote to leaves because injecting into leaves is expensive, mostly because of #5605 and some of #5552.
//    We don't promote past depth 2 because we found that gives us enough parallelism without costing us too much pinning work.
//
//    This is true with the following caveats:
//     We always promote all the way to the leaves on the rightmost and leftmost edges of the tree, for sequential insertions.
//      (That means we can promote past depth 2 near the edges of the tree.)
//
//   When the birdie is still saying we should promote, we use get_and_pin so that we wait to get the node.
//   If the birdie doesn't say to promote, we try maybe_get_and_pin.  If we get the node cheaply, and it's dirty, we promote anyway.
{
    toku_assert_entire_node_in_memory(subtree_root);
    if (should_inject_in_node(loc, subtree_root->height, depth)) {
        switch (depth) {
        case 0:
            STATUS_INC(FT_PRO_NUM_INJECT_DEPTH_0, 1); break;
        case 1:
            STATUS_INC(FT_PRO_NUM_INJECT_DEPTH_1, 1); break;
        case 2:
            STATUS_INC(FT_PRO_NUM_INJECT_DEPTH_2, 1); break;
        case 3:
            STATUS_INC(FT_PRO_NUM_INJECT_DEPTH_3, 1); break;
        default:
            STATUS_INC(FT_PRO_NUM_INJECT_DEPTH_GT3, 1); break;
        }
        inject_message_in_locked_node(ft, subtree_root, target_childnum, cmd, flow_deltas);
    } else {
        int r;
        int childnum;
        NONLEAF_CHILDINFO bnc;

        // toku_ft_root_put_cmd should not have called us otherwise.
        paranoid_invariant(ft_msg_applies_once(cmd));

        childnum = (target_childnum >= 0 ? target_childnum
                    : toku_ftnode_which_child(subtree_root, cmd->u.id.key, &ft->cmp_descriptor, ft->compare_fun));
        bnc = BNC(subtree_root, childnum);

        if (toku_bnc_n_entries(bnc) > 0) {
            // The buffer is non-empty, give up on promoting.
            STATUS_INC(FT_PRO_NUM_STOP_NONEMPTY_BUF, 1);
            goto relock_and_push_here;
        }

        seqinsert_loc next_loc;
        if ((loc & LEFT_EXTREME) && childnum == 0) {
            next_loc = LEFT_EXTREME;
        } else if ((loc & RIGHT_EXTREME) && childnum == subtree_root->n_children - 1) {
            next_loc = RIGHT_EXTREME;
        } else {
            next_loc = NEITHER_EXTREME;
        }

        if (next_loc == NEITHER_EXTREME && subtree_root->height <= 1) {
            // Never promote to leaf nodes except on the edges
            STATUS_INC(FT_PRO_NUM_STOP_H1, 1);
            goto relock_and_push_here;
        }

        {
            const BLOCKNUM child_blocknum = BP_BLOCKNUM(subtree_root, childnum);
            toku_verify_blocknum_allocated(ft->blocktable, child_blocknum);
            const uint32_t child_fullhash = toku_cachetable_hash(ft->cf, child_blocknum);

            FTNODE child;
            {
                const int child_height = subtree_root->height - 1;
                const int child_depth = depth + 1;
                // If we're locking a leaf, or a height 1 node or depth 2
                // node in the middle, we know we won't promote further
                // than that, so just get a write lock now.
                const pair_lock_type lock_type = (should_inject_in_node(next_loc, child_height, child_depth)
                                                  ? PL_WRITE_CHEAP
                                                  : PL_READ);
                if (next_loc != NEITHER_EXTREME || (toku_bnc_should_promote(ft, bnc) && depth <= 1)) {
                    // If we're on either extreme, or the birdie wants to
                    // promote and we're in the top two levels of the
                    // tree, don't stop just because someone else has the
                    // node locked.
                    struct ftnode_fetch_extra bfe;
                    fill_bfe_for_full_read(&bfe, ft);
                    toku_pin_ftnode_off_client_thread_batched(ft, child_blocknum, child_fullhash, &bfe, lock_type, 0, nullptr, &child);
                } else {
                    r = toku_maybe_pin_ftnode_clean(ft, child_blocknum, child_fullhash, lock_type, &child);
                    if (r != 0) {
                        // We couldn't get the child cheaply, so give up on promoting.
                        STATUS_INC(FT_PRO_NUM_STOP_LOCK_CHILD, 1);
                        goto relock_and_push_here;
                    }
                    if (is_entire_node_in_memory(child)) {
                        // toku_pin_ftnode... touches the clock but toku_maybe_pin_ftnode... doesn't.
                        // This prevents partial eviction.
                        for (int i = 0; i < child->n_children; ++i) {
                            BP_TOUCH_CLOCK(child, i);
                        }
                    } else {
                        // We got the child, but it's not fully in memory.  Give up on promoting.
                        STATUS_INC(FT_PRO_NUM_STOP_CHILD_INMEM, 1);
                        goto unlock_child_and_push_here;
                    }
                }
            }
            paranoid_invariant_notnull(child);

            if (!just_did_split_or_merge) {
                BLOCKNUM subtree_root_blocknum = subtree_root->thisnodename;
                uint32_t subtree_root_fullhash = toku_cachetable_hash(ft->cf, subtree_root_blocknum);
                const bool did_split_or_merge = process_maybe_reactive_child(ft, subtree_root, child, childnum, loc);
                if (did_split_or_merge) {
                    // Need to re-pin this node and try at this level again.
                    FTNODE newparent;
                    struct ftnode_fetch_extra bfe;
                    fill_bfe_for_full_read(&bfe, ft); // should be fully in memory, we just split it
                    toku_pin_ftnode_off_client_thread_batched(ft, subtree_root_blocknum, subtree_root_fullhash, &bfe, PL_READ, 0, nullptr, &newparent);
                    push_something_in_subtree(ft, newparent, -1, cmd, flow_deltas, depth, loc, true);
                    return;
                }
            }
2753

2754 2755 2756 2757 2758 2759 2760 2761
            if (next_loc != NEITHER_EXTREME || child->dirty || toku_bnc_should_promote(ft, bnc)) {
                push_something_in_subtree(ft, child, -1, cmd, flow_deltas, depth + 1, next_loc, false);
                toku_sync_fetch_and_add(&bnc->flow[0], flow_deltas[0]);
                // The recursive call unpinned the child, but
                // we're responsible for unpinning subtree_root.
                toku_unpin_ftnode_read_only(ft, subtree_root);
                return;
            }
2762

2763 2764 2765 2766 2767 2768 2769 2770 2771 2772 2773 2774 2775 2776 2777 2778 2779 2780 2781 2782 2783 2784 2785 2786 2787 2788 2789 2790 2791 2792 2793 2794 2795 2796 2797 2798 2799
            STATUS_INC(FT_PRO_NUM_DIDNT_WANT_PROMOTE, 1);
        unlock_child_and_push_here:
            // We locked the child, but we decided not to promote.
            // Unlock the child, and fall through to the next case.
            toku_unpin_ftnode_read_only(ft, child);
        }
    relock_and_push_here:
        // Give up on promoting.
        // We have subtree_root read-locked and we don't have a child locked.
        // Drop the read lock, grab a write lock, and inject here.
        {
            // Right now we have a read lock on subtree_root, but we want
            // to inject into it so we get a write lock instead.
            BLOCKNUM subtree_root_blocknum = subtree_root->thisnodename;
            uint32_t subtree_root_fullhash = toku_cachetable_hash(ft->cf, subtree_root_blocknum);
            toku_unpin_ftnode_read_only(ft, subtree_root);
            switch (depth) {
            case 0:
                STATUS_INC(FT_PRO_NUM_INJECT_DEPTH_0, 1); break;
            case 1:
                STATUS_INC(FT_PRO_NUM_INJECT_DEPTH_1, 1); break;
            case 2:
                STATUS_INC(FT_PRO_NUM_INJECT_DEPTH_2, 1); break;
            case 3:
                STATUS_INC(FT_PRO_NUM_INJECT_DEPTH_3, 1); break;
            default:
                STATUS_INC(FT_PRO_NUM_INJECT_DEPTH_GT3, 1); break;
            }
            inject_message_at_this_blocknum(ft, subtree_root_blocknum, subtree_root_fullhash, cmd, flow_deltas);
        }
    }
}

void toku_ft_root_put_cmd(FT ft, FT_MSG_S *cmd)
// Effect:
//  - assign msn to cmd and update msn in the header
//  - push the cmd into the ft
2800

2801 2802 2803 2804 2805 2806 2807 2808 2809 2810 2811 2812 2813 2814 2815 2816 2817 2818 2819
// As of Clayface, the root blocknum is a constant, so preventing a
// race between message injection and the split of a root is the job
// of the cachetable's locking rules.
//
// We also hold the MO lock for a number of reasons, but an important
// one is to make sure that a begin_checkpoint may not start while
// this code is executing. A begin_checkpoint does (at least) two things
// that can interfere with the operations here:
//  - Copies the header to a checkpoint header. Because we may change
//    the max_msn_in_ft below, we don't want the header to be copied in
//    the middle of these operations.
//  - Takes note of the log's LSN. Because this put operation has
//    already been logged, this message injection must be included
//    in any checkpoint that contains this put's logentry.
//    Holding the mo lock throughout this function ensures that fact.
{
    // blackhole fractal trees drop all messages, so do nothing.
    if (ft->blackhole) {
        return;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
2820
    }
Zardosht Kasheff's avatar
Zardosht Kasheff committed
2821

2822 2823 2824 2825 2826 2827 2828 2829 2830 2831 2832 2833 2834 2835 2836 2837 2838 2839 2840 2841 2842 2843 2844 2845 2846 2847 2848 2849 2850 2851 2852 2853 2854 2855 2856 2857 2858 2859 2860 2861 2862 2863 2864 2865 2866 2867 2868 2869 2870 2871 2872 2873 2874 2875 2876 2877 2878 2879 2880 2881
    FTNODE node;

    uint32_t fullhash;
    CACHEKEY root_key;
    toku_calculate_root_offset_pointer(ft, &root_key, &fullhash);
    struct ftnode_fetch_extra bfe;
    fill_bfe_for_full_read(&bfe, ft);

    size_t flow_deltas[] = { toku_ft_msg_memsize_in_fifo(cmd), 0 };

    pair_lock_type lock_type;
    lock_type = PL_READ; // try first for a read lock
    // If we need to split the root, we'll have to change from a read lock
    // to a write lock and check again.  We change the variable lock_type
    // and jump back to here.
 change_lock_type:
    // get the root node
    toku_pin_ftnode_off_client_thread_batched(ft, root_key, fullhash, &bfe, lock_type, 0, NULL, &node);
    toku_assert_entire_node_in_memory(node);
    paranoid_invariant(node->fullhash==fullhash);
    ft_verify_flags(ft, node);

    // First handle a reactive root.
    // This relocking for split algorithm will cause every message
    // injection thread to change lock type back and forth, when only one
    // of them needs to in order to handle the split.  That's not great,
    // but root splits are incredibly rare.
    enum reactivity re = get_node_reactivity(node, ft->h->nodesize);
    switch (re) {
    case RE_STABLE:
    case RE_FUSIBLE: // cannot merge anything at the root
        if (lock_type != PL_READ) {
            // We thought we needed to split, but someone else got to
            // it before us.  Downgrade to a read lock.
            toku_unpin_ftnode_read_only(ft, node);
            lock_type = PL_READ;
            goto change_lock_type;
        }
        break;
    case RE_FISSIBLE:
        if (lock_type == PL_READ) {
            // Here, we only have a read lock on the root.  In order
            // to split it, we need a write lock, but in the course of
            // gaining the write lock, someone else may have gotten in
            // before us and split it.  So we upgrade to a write lock
            // and check again.
            toku_unpin_ftnode_read_only(ft, node);
            lock_type = PL_WRITE_CHEAP;
            goto change_lock_type;
        } else {
            // We have a write lock, now we can split.
            ft_init_new_root(ft, node, &node);
            // Then downgrade back to a read lock, and we can finally
            // do the injection.
            toku_unpin_ftnode_off_client_thread(ft, node);
            lock_type = PL_READ;
            STATUS_INC(FT_PRO_NUM_ROOT_SPLIT, 1);
            goto change_lock_type;
        }
        break;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
2882
    }
2883 2884 2885 2886 2887 2888 2889 2890 2891 2892 2893 2894 2895 2896 2897 2898 2899 2900 2901 2902 2903 2904 2905 2906 2907 2908 2909 2910 2911 2912
    // If we get to here, we have a read lock and the root doesn't
    // need to be split.  It's safe to inject the message.
    paranoid_invariant(lock_type == PL_READ);
    // We cannot assert that we have the read lock because frwlock asserts
    // that its mutex is locked when we check if there are any readers.
    // That wouldn't give us a strong guarantee that we have the read lock
    // anyway.

    // Now, either inject here or promote.  We decide based on a heuristic:
    if (node->height == 0 || !ft_msg_applies_once(cmd)) {
        // If the root's a leaf or we're injecting a broadcast, drop the read lock and inject here.
        toku_unpin_ftnode_read_only(ft, node);
        STATUS_INC(FT_PRO_NUM_ROOT_H0_INJECT, 1);
        inject_message_at_this_blocknum(ft, root_key, fullhash, cmd, flow_deltas);
    } else if (node->height > 1) {
        // If the root's above height 1, we are definitely eligible for promotion.
        push_something_in_subtree(ft, node, -1, cmd, flow_deltas, 0, LEFT_EXTREME | RIGHT_EXTREME, false);
    } else {
        // The root's height 1.  We may be eligible for promotion here.
        // On the extremes, we want to promote, in the middle, we don't.
        int childnum = toku_ftnode_which_child(node, cmd->u.id.key, &ft->cmp_descriptor, ft->compare_fun);
        if (childnum == 0 || childnum == node->n_children - 1) {
            // On the extremes, promote.  We know which childnum we're going to, so pass that down too.
            push_something_in_subtree(ft, node, childnum, cmd, flow_deltas, 0, LEFT_EXTREME | RIGHT_EXTREME, false);
        } else {
            // At height 1 in the middle, don't promote, drop the read lock and inject here.
            toku_unpin_ftnode_read_only(ft, node);
            STATUS_INC(FT_PRO_NUM_ROOT_H1_INJECT, 1);
            inject_message_at_this_blocknum(ft, root_key, fullhash, cmd, flow_deltas);
        }
2913
    }
2914 2915
}

2916
// Effect: Insert the key-val pair into brt.
2917 2918
void toku_ft_insert (FT_HANDLE brt, DBT *key, DBT *val, TOKUTXN txn) {
    toku_ft_maybe_insert(brt, key, val, txn, false, ZERO_LSN, true, FT_INSERT);
2919 2920
}

2921
void toku_ft_load_recovery(TOKUTXN txn, FILENUM old_filenum, char const * new_iname, int do_fsync, int do_log, LSN *load_lsn) {
2922
    paranoid_invariant(txn);
Yoni Fogel's avatar
Yoni Fogel committed
2923
    toku_txn_force_fsync_on_commit(txn);  //If the txn commits, the commit MUST be in the log
2924
                                          //before the (old) file is actually unlinked
Yoni Fogel's avatar
Yoni Fogel committed
2925 2926
    TOKULOGGER logger = toku_txn_logger(txn);

2927
    BYTESTRING new_iname_bs = {.len=(uint32_t) strlen(new_iname), .data=(char*)new_iname};
2928 2929
    toku_logger_save_rollback_load(txn, old_filenum, &new_iname_bs);
    if (do_log && logger) {
2930
        TXNID xid = toku_txn_get_txnid(txn);
2931
        toku_log_load(logger, load_lsn, do_fsync, txn, xid, old_filenum, new_iname_bs);
Yoni Fogel's avatar
Yoni Fogel committed
2932 2933 2934
    }
}

2935 2936 2937 2938
// 2954
// this function handles the tasks needed to be recoverable
//  - write to rollback log
//  - write to recovery log
2939
void toku_ft_hot_index_recovery(TOKUTXN txn, FILENUMS filenums, int do_fsync, int do_log, LSN *hot_index_lsn)
2940
{
2941
    paranoid_invariant(txn);
2942 2943 2944
    TOKULOGGER logger = toku_txn_logger(txn);

    // write to the rollback log
2945 2946
    toku_logger_save_rollback_hot_index(txn, &filenums);
    if (do_log && logger) {
2947 2948
        TXNID xid = toku_txn_get_txnid(txn);
        // write to the recovery log
2949
        toku_log_hot_index(logger, hot_index_lsn, do_fsync, txn, xid, filenums);
2950 2951
    }
}
2952

2953 2954
// Effect: Optimize the ft.
void toku_ft_optimize (FT_HANDLE brt) {
2955
    TOKULOGGER logger = toku_cachefile_logger(brt->ft->cf);
2956
    if (logger) {
2957
        TXNID oldest = toku_txn_manager_get_oldest_living_xid(logger->txn_manager);
2958

2959 2960 2961 2962 2963 2964
        XIDS root_xids = xids_get_root_xids();
        XIDS message_xids;
        if (oldest == TXNID_NONE_LIVING) {
            message_xids = root_xids;
        }
        else {
2965 2966
            int r = xids_create_child(root_xids, &message_xids, oldest);
            invariant(r == 0);
2967
        }
2968

2969 2970 2971 2972
        DBT key;
        DBT val;
        toku_init_dbt(&key);
        toku_init_dbt(&val);
2973
        FT_MSG_S ftcmd = { FT_OPTIMIZE, ZERO_MSN, message_xids, .u = { .id = {&key,&val} } };
2974
        toku_ft_root_put_cmd(brt->ft, &ftcmd);
2975 2976
        xids_destroy(&message_xids);
    }
2977 2978
}

2979
void toku_ft_load(FT_HANDLE brt, TOKUTXN txn, char const * new_iname, int do_fsync, LSN *load_lsn) {
2980
    FILENUM old_filenum = toku_cachefile_filenum(brt->ft->cf);
Yoni Fogel's avatar
Yoni Fogel committed
2981
    int do_log = 1;
2982
    toku_ft_load_recovery(txn, old_filenum, new_iname, do_fsync, do_log, load_lsn);
Yoni Fogel's avatar
Yoni Fogel committed
2983 2984
}

2985 2986
// ft actions for logging hot index filenums
void toku_ft_hot_index(FT_HANDLE brt __attribute__ ((unused)), TOKUTXN txn, FILENUMS filenums, int do_fsync, LSN *lsn) {
2987
    int do_log = 1;
2988
    toku_ft_hot_index_recovery(txn, filenums, do_fsync, do_log, lsn);
2989 2990
}

2991
void
2992
toku_ft_log_put (TOKUTXN txn, FT_HANDLE brt, const DBT *key, const DBT *val) {
2993
    TOKULOGGER logger = toku_txn_logger(txn);
2994
    if (logger) {
2995 2996
        BYTESTRING keybs = {.len=key->size, .data=(char *) key->data};
        BYTESTRING valbs = {.len=val->size, .data=(char *) val->data};
2997
        TXNID xid = toku_txn_get_txnid(txn);
2998
        toku_log_enq_insert(logger, (LSN*)0, 0, txn, toku_cachefile_filenum(brt->ft->cf), xid, keybs, valbs);
2999 3000 3001
    }
}

3002
void
3003
toku_ft_log_put_multiple (TOKUTXN txn, FT_HANDLE src_ft, FT_HANDLE *brts, uint32_t num_fts, const DBT *key, const DBT *val) {
3004
    assert(txn);
3005
    assert(num_fts > 0);
3006 3007
    TOKULOGGER logger = toku_txn_logger(txn);
    if (logger) {
3008
        FILENUM         fnums[num_fts];
3009
        uint32_t i;
3010
        for (i = 0; i < num_fts; i++) {
3011
            fnums[i] = toku_cachefile_filenum(brts[i]->ft->cf);
3012
        }
3013 3014 3015 3016 3017 3018
        FILENUMS filenums = {.num = num_fts, .filenums = fnums};
        BYTESTRING keybs = {.len=key->size, .data=(char *) key->data};
        BYTESTRING valbs = {.len=val->size, .data=(char *) val->data};
        TXNID xid = toku_txn_get_txnid(txn);
        FILENUM src_filenum = src_ft ? toku_cachefile_filenum(src_ft->ft->cf) : FILENUM_NONE;
        toku_log_enq_insert_multiple(logger, (LSN*)0, 0, txn, src_filenum, filenums, xid, keybs, valbs);
3019 3020 3021
    }
}

3022
void toku_ft_maybe_insert (FT_HANDLE ft_h, DBT *key, DBT *val, TOKUTXN txn, bool oplsn_valid, LSN oplsn, bool do_logging, enum ft_msg_type type) {
3023
    paranoid_invariant(type==FT_INSERT || type==FT_INSERT_NO_OVERWRITE);
3024
    XIDS message_xids = xids_get_root_xids(); //By default use committed messages
3025
    TXNID xid = toku_txn_get_txnid(txn);
3026
    if (txn) {
3027 3028 3029 3030
        BYTESTRING keybs = {key->size, (char *) key->data};
        toku_logger_save_rollback_cmdinsert(txn, toku_cachefile_filenum(ft_h->ft->cf), &keybs);
        toku_txn_maybe_note_ft(txn, ft_h->ft);
        message_xids = toku_txn_get_xids(txn);
Bradley C. Kuszmaul's avatar
Rename  
Bradley C. Kuszmaul committed
3031
    }
3032
    TOKULOGGER logger = toku_txn_logger(txn);
3033
    if (do_logging && logger) {
3034 3035
        BYTESTRING keybs = {.len=key->size, .data=(char *) key->data};
        BYTESTRING valbs = {.len=val->size, .data=(char *) val->data};
3036
        if (type == FT_INSERT) {
3037
            toku_log_enq_insert(logger, (LSN*)0, 0, txn, toku_cachefile_filenum(ft_h->ft->cf), xid, keybs, valbs);
3038 3039
        }
        else {
3040
            toku_log_enq_insert_no_overwrite(logger, (LSN*)0, 0, txn, toku_cachefile_filenum(ft_h->ft->cf), xid, keybs, valbs);
3041
        }
3042 3043
    }

3044
    LSN treelsn;
3045
    if (oplsn_valid && oplsn.lsn <= (treelsn = toku_ft_checkpoint_lsn(ft_h->ft)).lsn) {
3046
        // do nothing
3047
    } else {
3048
        toku_ft_send_insert(ft_h, key, val, message_xids, type);
3049
    }
3050
}
3051

3052
static void
3053
ft_send_update_msg(FT_HANDLE brt, FT_MSG_S *msg, TOKUTXN txn) {
3054
    msg->xids = (txn
3055 3056
                 ? toku_txn_get_xids(txn)
                 : xids_get_root_xids());
3057
    toku_ft_root_put_cmd(brt->ft, msg);
3058 3059
}

3060
void toku_ft_maybe_update(FT_HANDLE ft_h, const DBT *key, const DBT *update_function_extra,
Yoni Fogel's avatar
Yoni Fogel committed
3061 3062
                      TOKUTXN txn, bool oplsn_valid, LSN oplsn,
                      bool do_logging) {
3063 3064
    TXNID xid = toku_txn_get_txnid(txn);
    if (txn) {
3065
        BYTESTRING keybs = { key->size, (char *) key->data };
3066
        toku_logger_save_rollback_cmdupdate(
3067 3068
            txn, toku_cachefile_filenum(ft_h->ft->cf), &keybs);
        toku_txn_maybe_note_ft(txn, ft_h->ft);
3069 3070
    }

3071 3072
    TOKULOGGER logger;
    logger = toku_txn_logger(txn);
3073
    if (do_logging && logger) {
3074
        BYTESTRING keybs = {.len=key->size, .data=(char *) key->data};
3075
        BYTESTRING extrabs = {.len=update_function_extra->size,
3076
                              .data = (char *) update_function_extra->data};
3077
        toku_log_enq_update(logger, NULL, 0, txn,
3078
                                toku_cachefile_filenum(ft_h->ft->cf),
3079
                                xid, keybs, extrabs);
3080 3081 3082
    }

    LSN treelsn;
3083 3084
    if (oplsn_valid && oplsn.lsn <= (treelsn = toku_ft_checkpoint_lsn(ft_h->ft)).lsn) {
        // do nothing
3085
    } else {
3086
        FT_MSG_S msg = { FT_UPDATE, ZERO_MSN, NULL,
3087
                         .u = { .id = { key, update_function_extra } } };
3088
        ft_send_update_msg(ft_h, &msg, txn);
3089 3090 3091
    }
}

3092
void toku_ft_maybe_update_broadcast(FT_HANDLE ft_h, const DBT *update_function_extra,
Yoni Fogel's avatar
Yoni Fogel committed
3093 3094
                                TOKUTXN txn, bool oplsn_valid, LSN oplsn,
                                bool do_logging, bool is_resetting_op) {
3095
    TXNID xid = toku_txn_get_txnid(txn);
Yoni Fogel's avatar
Yoni Fogel committed
3096
    uint8_t  resetting = is_resetting_op ? 1 : 0;
3097
    if (txn) {
3098
        toku_logger_save_rollback_cmdupdatebroadcast(txn, toku_cachefile_filenum(ft_h->ft->cf), resetting);
3099
        toku_txn_maybe_note_ft(txn, ft_h->ft);
3100 3101
    }

3102 3103
    TOKULOGGER logger;
    logger = toku_txn_logger(txn);
3104
    if (do_logging && logger) {
3105
        BYTESTRING extrabs = {.len=update_function_extra->size,
3106
                              .data = (char *) update_function_extra->data};
3107
        toku_log_enq_updatebroadcast(logger, NULL, 0, txn,
3108
                                         toku_cachefile_filenum(ft_h->ft->cf),
3109
                                         xid, extrabs, resetting);
3110 3111
    }

3112
    //TODO(yoni): remove treelsn here and similar calls (no longer being used)
3113 3114
    LSN treelsn;
    if (oplsn_valid &&
3115
        oplsn.lsn <= (treelsn = toku_ft_checkpoint_lsn(ft_h->ft)).lsn) {
3116

3117
    } else {
3118 3119
        DBT nullkey;
        const DBT *nullkeyp = toku_init_dbt(&nullkey);
3120
        FT_MSG_S msg = { FT_UPDATE_BROADCAST_ALL, ZERO_MSN, NULL,
3121
                         .u = { .id = { nullkeyp, update_function_extra } } };
3122
        ft_send_update_msg(ft_h, &msg, txn);
3123 3124 3125
    }
}

3126
void toku_ft_send_insert(FT_HANDLE brt, DBT *key, DBT *val, XIDS xids, enum ft_msg_type type) {
3127
    FT_MSG_S ftcmd = { type, ZERO_MSN, xids, .u = { .id = { key, val } } };
3128
    toku_ft_root_put_cmd(brt->ft, &ftcmd);
3129 3130
}

3131
void toku_ft_send_commit_any(FT_HANDLE brt, DBT *key, XIDS xids) {
3132
    DBT val;
3133
    FT_MSG_S ftcmd = { FT_COMMIT_ANY, ZERO_MSN, xids, .u = { .id = { key, toku_init_dbt(&val) } } };
3134
    toku_ft_root_put_cmd(brt->ft, &ftcmd);
3135 3136
}

3137 3138
void toku_ft_delete(FT_HANDLE brt, DBT *key, TOKUTXN txn) {
    toku_ft_maybe_delete(brt, key, txn, false, ZERO_LSN, true);
3139 3140
}

3141
void
3142
toku_ft_log_del(TOKUTXN txn, FT_HANDLE brt, const DBT *key) {
3143
    TOKULOGGER logger = toku_txn_logger(txn);
3144
    if (logger) {
3145
        BYTESTRING keybs = {.len=key->size, .data=(char *) key->data};
3146
        TXNID xid = toku_txn_get_txnid(txn);
3147
        toku_log_enq_delete_any(logger, (LSN*)0, 0, txn, toku_cachefile_filenum(brt->ft->cf), xid, keybs);
3148 3149 3150
    }
}

3151
void
3152
toku_ft_log_del_multiple (TOKUTXN txn, FT_HANDLE src_ft, FT_HANDLE *brts, uint32_t num_fts, const DBT *key, const DBT *val) {
3153
    assert(txn);
3154
    assert(num_fts > 0);
3155 3156
    TOKULOGGER logger = toku_txn_logger(txn);
    if (logger) {
3157
        FILENUM         fnums[num_fts];
3158
        uint32_t i;
3159
        for (i = 0; i < num_fts; i++) {
3160
            fnums[i] = toku_cachefile_filenum(brts[i]->ft->cf);
3161
        }
3162 3163 3164 3165 3166 3167
        FILENUMS filenums = {.num = num_fts, .filenums = fnums};
        BYTESTRING keybs = {.len=key->size, .data=(char *) key->data};
        BYTESTRING valbs = {.len=val->size, .data=(char *) val->data};
        TXNID xid = toku_txn_get_txnid(txn);
        FILENUM src_filenum = src_ft ? toku_cachefile_filenum(src_ft->ft->cf) : FILENUM_NONE;
        toku_log_enq_delete_multiple(logger, (LSN*)0, 0, txn, src_filenum, filenums, xid, keybs, valbs);
3168 3169 3170
    }
}

3171
void toku_ft_maybe_delete(FT_HANDLE ft_h, DBT *key, TOKUTXN txn, bool oplsn_valid, LSN oplsn, bool do_logging) {
3172
    XIDS message_xids = xids_get_root_xids(); //By default use committed messages
3173
    TXNID xid = toku_txn_get_txnid(txn);
3174
    if (txn) {
3175 3176 3177 3178
        BYTESTRING keybs = {key->size, (char *) key->data};
        toku_logger_save_rollback_cmddelete(txn, toku_cachefile_filenum(ft_h->ft->cf), &keybs);
        toku_txn_maybe_note_ft(txn, ft_h->ft);
        message_xids = toku_txn_get_xids(txn);
3179
    }
3180
    TOKULOGGER logger = toku_txn_logger(txn);
3181
    if (do_logging && logger) {
3182
        BYTESTRING keybs = {.len=key->size, .data=(char *) key->data};
3183
        toku_log_enq_delete_any(logger, (LSN*)0, 0, txn, toku_cachefile_filenum(ft_h->ft->cf), xid, keybs);
3184
    }
3185

3186
    LSN treelsn;
3187
    if (oplsn_valid && oplsn.lsn <= (treelsn = toku_ft_checkpoint_lsn(ft_h->ft)).lsn) {
3188
        // do nothing
3189
    } else {
3190
        toku_ft_send_delete(ft_h, key, message_xids);
3191
    }
3192 3193
}

3194
void toku_ft_send_delete(FT_HANDLE brt, DBT *key, XIDS xids) {
3195
    DBT val; toku_init_dbt(&val);
3196
    FT_MSG_S ftcmd = { FT_DELETE_ANY, ZERO_MSN, xids, .u = { .id = { key, &val } } };
3197
    toku_ft_root_put_cmd(brt->ft, &ftcmd);
3198 3199
}

3200 3201 3202 3203
/* mempool support */

struct omt_compressor_state {
    struct mempool *new_kvspace;
3204
    OMTVALUE *newvals;
3205 3206
};

Yoni Fogel's avatar
Yoni Fogel committed
3207
static int move_it (OMTVALUE lev, uint32_t idx, void *v) {
3208 3209
    LEAFENTRY CAST_FROM_VOIDP(le, lev);
    struct omt_compressor_state *CAST_FROM_VOIDP(oc, v);
Yoni Fogel's avatar
Yoni Fogel committed
3210
    uint32_t size = leafentry_memsize(le);
3211
    LEAFENTRY CAST_FROM_VOIDP(newdata, toku_mempool_malloc(oc->new_kvspace, size, 1));
3212
    paranoid_invariant_notnull(newdata); // we do this on a fresh mempool, so nothing bad should happen
3213
    memcpy(newdata, le, size);
3214
    oc->newvals[idx] = newdata;
3215 3216 3217 3218
    return 0;
}

// Compress things, and grow the mempool if needed.
3219
static void omt_compress_kvspace (OMT *omtp, struct mempool *memp, size_t added_size, void **maybe_free) {
Yoni Fogel's avatar
Yoni Fogel committed
3220
    uint32_t total_size_needed = memp->free_offset-memp->frag_size + added_size;
3221 3222
    if (total_size_needed+total_size_needed >= memp->size) {
        memp->size = total_size_needed+total_size_needed;
3223 3224 3225 3226
    }
    void *newmem = toku_xmalloc(memp->size);
    struct mempool new_kvspace;
    toku_mempool_init(&new_kvspace, newmem, memp->size);
3227 3228 3229 3230 3231 3232
    uint32_t numvals = toku_omt_size(*omtp);
    OMTVALUE *XMALLOC_N(numvals, newvals);
    struct omt_compressor_state oc = { &new_kvspace, newvals };
    toku_omt_iterate(*omtp, move_it, &oc);
    toku_omt_destroy(omtp);
    toku_omt_create_steal_sorted_array(omtp, &newvals, numvals, numvals);
3233 3234 3235 3236 3237 3238 3239 3240 3241 3242

    if (maybe_free) {
        *maybe_free = memp->base;
    } else {
        toku_free(memp->base);
    }
    *memp = new_kvspace;
}

void *
3243
mempool_malloc_from_omt(OMT *omtp, struct mempool *mp, size_t size, void **maybe_free) {
3244
    void *v = toku_mempool_malloc(mp, size, 1);
3245
    if (v == NULL) {
3246
        omt_compress_kvspace(omtp, mp, size, maybe_free);
3247
        v = toku_mempool_malloc(mp, size, 1);
3248
        paranoid_invariant_notnull(v);
3249 3250 3251 3252
    }
    return v;
}

3253 3254
/* ******************** open,close and create  ********************** */

3255
// Test only function (not used in running system). This one has no env
3256
int toku_open_ft_handle (const char *fname, int is_create, FT_HANDLE *ft_handle_p, int nodesize,
3257 3258 3259 3260
                   int basementnodesize,
                   enum toku_compression_method compression_method,
                   CACHETABLE cachetable, TOKUTXN txn,
                   int (*compare_fun)(DB *, const DBT*,const DBT*)) {
3261
    FT_HANDLE brt;
3262 3263
    const int only_create = 0;

3264
    toku_ft_handle_create(&brt);
3265 3266 3267
    toku_ft_handle_set_nodesize(brt, nodesize);
    toku_ft_handle_set_basementnodesize(brt, basementnodesize);
    toku_ft_handle_set_compression_method(brt, compression_method);
3268
    toku_ft_set_bt_compare(brt, compare_fun);
3269

3270
    int r = toku_ft_handle_open(brt, fname, is_create, only_create, cachetable, txn);
3271
    if (r != 0) {
3272
        return r;
3273 3274
    }

3275
    *ft_handle_p = brt;
3276 3277 3278
    return r;
}

3279 3280
// open a file for use by the brt
// Requires:  File does not exist.
3281
static int ft_create_file(FT_HANDLE UU(brt), const char *fname, int *fdp) {
3282
    mode_t mode = S_IRWXU|S_IRWXG|S_IRWXO;
3283
    int r;
3284
    int fd;
3285
    int er;
3286
    fd = open(fname, O_RDWR | O_BINARY, mode);
3287
    assert(fd==-1);
3288 3289
    if ((er = get_maybe_error_errno()) != ENOENT) {
        return er;
3290 3291
    }
    fd = open(fname, O_RDWR | O_CREAT | O_BINARY, mode);
3292
    if (fd==-1) {
3293
        r = get_error_errno();
3294
        return r;
3295
    }
3296 3297

    r = toku_fsync_directory(fname);
3298
    resource_assert_zero(r);
3299

3300 3301 3302 3303 3304
    *fdp = fd;
    return 0;
}

// open a file for use by the brt.  if the file does not exist, error
3305
static int ft_open_file(const char *fname, int *fdp) {
3306 3307 3308 3309
    mode_t mode = S_IRWXU|S_IRWXG|S_IRWXO;
    int fd;
    fd = open(fname, O_RDWR | O_BINARY, mode);
    if (fd==-1) {
3310
        return get_error_errno();
3311 3312 3313 3314 3315
    }
    *fdp = fd;
    return 0;
}

3316 3317
void
toku_ft_handle_set_compression_method(FT_HANDLE t, enum toku_compression_method method)
3318
{
3319 3320 3321 3322 3323 3324
    if (t->ft) {
        toku_ft_set_compression_method(t->ft, method);
    }
    else {
        t->options.compression_method = method;
    }
3325 3326
}

3327 3328
void
toku_ft_handle_get_compression_method(FT_HANDLE t, enum toku_compression_method *methodp)
3329
{
3330 3331 3332 3333 3334 3335
    if (t->ft) {
        toku_ft_get_compression_method(t->ft, methodp);
    }
    else {
        *methodp = t->options.compression_method;
    }
3336 3337
}

3338
static int
Yoni Fogel's avatar
Yoni Fogel committed
3339
verify_builtin_comparisons_consistent(FT_HANDLE t, uint32_t flags) {
3340
    if ((flags & TOKU_DB_KEYCMP_BUILTIN) && (t->options.compare_fun != toku_builtin_compare_fun))
3341
        return EINVAL;
3342 3343 3344
    return 0;
}

3345 3346 3347 3348
//
// See comments in toku_db_change_descriptor to understand invariants 
// in the system when this function is called
//
3349
void toku_ft_change_descriptor(
3350
    FT_HANDLE ft_h,
3351 3352
    const DBT* old_descriptor,
    const DBT* new_descriptor,
Yoni Fogel's avatar
Yoni Fogel committed
3353
    bool do_log,
3354
    TOKUTXN txn,
Yoni Fogel's avatar
Yoni Fogel committed
3355
    bool update_cmp_descriptor
3356
    )
3357 3358 3359
{
    DESCRIPTOR_S new_d;

3360 3361 3362 3363 3364
    // if running with txns, save to rollback + write to recovery log
    if (txn) {
        // put information into rollback file
        BYTESTRING old_desc_bs = { old_descriptor->size, (char *) old_descriptor->data };
        BYTESTRING new_desc_bs = { new_descriptor->size, (char *) new_descriptor->data };
3365
        toku_logger_save_rollback_change_fdescriptor(
3366
            txn,
3367
            toku_cachefile_filenum(ft_h->ft->cf),
3368
            &old_desc_bs
3369
            );
3370 3371 3372 3373 3374
        toku_txn_maybe_note_ft(txn, ft_h->ft);

        if (do_log) {
            TOKULOGGER logger = toku_txn_logger(txn);
            TXNID xid = toku_txn_get_txnid(txn);
3375
            toku_log_change_fdescriptor(
3376 3377 3378 3379 3380 3381 3382 3383 3384
                logger, NULL, 0,
                txn,
                toku_cachefile_filenum(ft_h->ft->cf),
                xid,
                old_desc_bs,
                new_desc_bs,
                update_cmp_descriptor
                );
        }
3385
    }
3386 3387 3388

    // write new_descriptor to header
    new_d.dbt = *new_descriptor;
3389
    toku_ft_update_descriptor(ft_h->ft, &new_d);
3390
    // very infrequent operation, worth precise threadsafe count
3391
    STATUS_INC(FT_DESCRIPTOR_SET, 1);
3392

3393
    if (update_cmp_descriptor) {
3394
        toku_ft_update_cmp_descriptor(ft_h->ft);
3395
    }
3396 3397
}

3398 3399 3400
static void
toku_ft_handle_inherit_options(FT_HANDLE t, FT ft) {
    struct ft_options options = {
3401 3402 3403 3404
        .nodesize = ft->h->nodesize,
        .basementnodesize = ft->h->basementnodesize,
        .compression_method = ft->h->compression_method,
        .flags = ft->h->flags,
3405 3406 3407 3408
        .compare_fun = ft->compare_fun,
        .update_fun = ft->update_fun
    };
    t->options = options;
Yoni Fogel's avatar
Yoni Fogel committed
3409
    t->did_set_flags = true;
3410 3411
}

3412
// This is the actual open, used for various purposes, such as normal use, recovery, and redirect.
3413 3414
// fname_in_env is the iname, relative to the env_dir  (data_dir is already in iname as prefix).
// The checkpointed version (checkpoint_lsn) of the dictionary must be no later than max_acceptable_lsn .
3415
// Requires: The multi-operation client lock must be held to prevent a checkpoint from occuring.
3416
static int
3417
ft_handle_open(FT_HANDLE ft_h, const char *fname_in_env, int is_create, int only_create, CACHETABLE cachetable, TOKUTXN txn, FILENUM use_filenum, DICTIONARY_ID use_dictionary_id, LSN max_acceptable_lsn) {
3418
    int r;
Yoni Fogel's avatar
Yoni Fogel committed
3419
    bool txn_created = false;
3420 3421
    char *fname_in_cwd = NULL;
    CACHEFILE cf = NULL;
3422
    FT ft = NULL;
Yoni Fogel's avatar
Yoni Fogel committed
3423
    bool did_create = false;
3424
    toku_ft_open_close_lock();
3425

3426 3427
    if (ft_h->did_set_flags) {
        r = verify_builtin_comparisons_consistent(ft_h, ft_h->options.flags);
3428 3429
        if (r!=0) { goto exit; }
    }
3430

3431
    assert(is_create || !only_create);
3432 3433
    FILENUM reserved_filenum;
    reserved_filenum = use_filenum;
3434
    fname_in_cwd = toku_cachetable_get_fname_in_cwd(cachetable, fname_in_env);
3435
    {
3436
        int fd = -1;
3437
        r = ft_open_file(fname_in_cwd, &fd);
John Esmet's avatar
John Esmet committed
3438 3439 3440
        if (reserved_filenum.fileid == FILENUM_NONE.fileid) {
            reserved_filenum = toku_cachetable_reserve_filenum(cachetable);
        }
3441
        if (r==ENOENT && is_create) {
Yoni Fogel's avatar
Yoni Fogel committed
3442
            did_create = true;
3443 3444
            mode_t mode = S_IRWXU|S_IRWXG|S_IRWXO;
            if (txn) {
3445
                BYTESTRING bs = { .len=(uint32_t) strlen(fname_in_env), .data = (char*)fname_in_env };
3446
                toku_logger_save_rollback_fcreate(txn, reserved_filenum, &bs); // bs is a copy of the fname relative to the environment
3447
            }
Yoni Fogel's avatar
Yoni Fogel committed
3448
            txn_created = (bool)(txn!=NULL);
3449
            toku_logger_log_fcreate(txn, fname_in_env, reserved_filenum, mode, ft_h->options.flags, ft_h->options.nodesize, ft_h->options.basementnodesize, ft_h->options.compression_method);
3450
            r = ft_create_file(ft_h, fname_in_cwd, &fd);
3451 3452 3453
            assert_zero(r);
        }
        if (r) { goto exit; }
John Esmet's avatar
John Esmet committed
3454
        r=toku_cachetable_openfd_with_filenum(&cf, cachetable, fd, fname_in_env, reserved_filenum);
3455
        if (r) { goto exit; }
3456
    }
3457
    assert(ft_h->options.nodesize>0);
Yoni Fogel's avatar
Yoni Fogel committed
3458
    bool was_already_open;
3459
    if (is_create) {
3460
        r = toku_read_ft_and_store_in_cachefile(ft_h, cf, max_acceptable_lsn, &ft, &was_already_open);
3461
        if (r==TOKUDB_DICTIONARY_NO_HEADER) {
3462
            toku_ft_create(&ft, &ft_h->options, cf, txn);
3463 3464 3465 3466 3467 3468 3469 3470 3471 3472
        }
        else if (r!=0) {
            goto exit;
        }
        else if (only_create) {
            assert_zero(r);
            r = EEXIST;
            goto exit;
        }
        // if we get here, then is_create was true but only_create was false,
3473 3474
        // so it is ok for toku_read_ft_and_store_in_cachefile to have read
        // the header via toku_read_ft_and_store_in_cachefile
3475
    } else {
3476
        r = toku_read_ft_and_store_in_cachefile(ft_h, cf, max_acceptable_lsn, &ft, &was_already_open);
3477 3478
        if (r) { goto exit; }
    }
3479 3480
    if (!ft_h->did_set_flags) {
        r = verify_builtin_comparisons_consistent(ft_h, ft_h->options.flags);
3481
        if (r) { goto exit; }
3482
    } else if (ft_h->options.flags != ft->h->flags) {                  /* if flags have been set then flags must match */
3483 3484
        r = EINVAL;
        goto exit;
3485
    }
3486
    toku_ft_handle_inherit_options(ft_h, ft);
3487 3488

    if (!was_already_open) {
3489
        if (!did_create) { //Only log the fopen that OPENs the file.  If it was already open, don't log.
3490
            toku_logger_log_fopen(txn, fname_in_env, toku_cachefile_filenum(cf), ft_h->options.flags);
3491
        }
3492
    }
3493 3494
    int use_reserved_dict_id;
    use_reserved_dict_id = use_dictionary_id.dictid != DICTIONARY_ID_NONE.dictid;
3495
    if (!was_already_open) {
3496 3497 3498 3499 3500 3501 3502
        DICTIONARY_ID dict_id;
        if (use_reserved_dict_id) {
            dict_id = use_dictionary_id;
        }
        else {
            dict_id = next_dict_id();
        }
3503
        ft->dict_id = dict_id;
3504 3505
    }
    else {
3506 3507
        // dict_id is already in header
        if (use_reserved_dict_id) {
3508
            assert(ft->dict_id.dictid == use_dictionary_id.dictid);
3509
        }
3510
    }
3511 3512 3513
    assert(ft);
    assert(ft->dict_id.dictid != DICTIONARY_ID_NONE.dictid);
    assert(ft->dict_id.dictid < dict_id_serial);
3514

3515 3516 3517
    // important note here,
    // after this point, where we associate the header
    // with the brt, the function is not allowed to fail
3518 3519
    // Code that handles failure (located below "exit"),
    // depends on this
3520
    toku_ft_note_ft_handle_open(ft, ft_h);
3521
    if (txn_created) {
3522
        assert(txn);
3523
        toku_txn_maybe_note_ft(txn, ft);
3524
    }
3525

3526
    //Opening a brt may restore to previous checkpoint.         Truncate if necessary.
3527
    {
3528
        int fd = toku_cachefile_get_fd (ft->cf);
3529
        toku_maybe_truncate_file_on_open(ft->blocktable, fd);
3530
    }
3531 3532 3533 3534 3535 3536

    r = 0;
exit:
    if (fname_in_cwd) {
        toku_free(fname_in_cwd);
    }
John Esmet's avatar
John Esmet committed
3537
    if (r != 0 && cf) {
3538
        if (ft) {
John Esmet's avatar
John Esmet committed
3539 3540 3541 3542 3543 3544
            // we only call toku_ft_note_ft_handle_open
            // when the function succeeds, so if we are here,
            // then that means we have a reference to the header
            // but we have not linked it to this brt. So,
            // we can simply try to remove the header.
            // We don't need to unlink this brt from the header
3545
            toku_ft_grab_reflock(ft);
Yoni Fogel's avatar
Yoni Fogel committed
3546
            bool needed = toku_ft_needed_unlocked(ft);
3547 3548
            toku_ft_release_reflock(ft);
            if (!needed) {
3549 3550
                // close immediately.
                toku_ft_evict_from_memory(ft, false, ZERO_LSN);
3551 3552
            }
        }
John Esmet's avatar
John Esmet committed
3553
        else {
3554
            toku_cachefile_close(&cf, false, ZERO_LSN);
John Esmet's avatar
John Esmet committed
3555
        }
3556
    }
3557
    toku_ft_open_close_unlock();
3558
    return r;
3559 3560
}

3561
// Open a brt for the purpose of recovery, which requires that the brt be open to a pre-determined FILENUM
3562
// and may require a specific checkpointed version of the file.
3563
// (dict_id is assigned by the ft_handle_open() function.)
3564
int
3565
toku_ft_handle_open_recovery(FT_HANDLE t, const char *fname_in_env, int is_create, int only_create, CACHETABLE cachetable, TOKUTXN txn, FILENUM use_filenum, LSN max_acceptable_lsn) {
3566
    int r;
3567
    assert(use_filenum.fileid != FILENUM_NONE.fileid);
3568
    r = ft_handle_open(t, fname_in_env, is_create, only_create, cachetable,
3569
                 txn, use_filenum, DICTIONARY_ID_NONE, max_acceptable_lsn);
3570 3571 3572
    return r;
}

3573
// Open a brt in normal use.  The FILENUM and dict_id are assigned by the ft_handle_open() function.
3574
// Requires: The multi-operation client lock must be held to prevent a checkpoint from occuring.
3575
int
3576
toku_ft_handle_open(FT_HANDLE t, const char *fname_in_env, int is_create, int only_create, CACHETABLE cachetable, TOKUTXN txn) {
3577
    int r;
3578
    r = ft_handle_open(t, fname_in_env, is_create, only_create, cachetable, txn, FILENUM_NONE, DICTIONARY_ID_NONE, MAX_LSN);
3579 3580 3581
    return r;
}

3582 3583 3584 3585
// clone an ft handle. the cloned handle has a new dict_id but refers to the same fractal tree
int 
toku_ft_handle_clone(FT_HANDLE *cloned_ft_handle, FT_HANDLE ft_handle, TOKUTXN txn) {
    FT_HANDLE result_ft_handle; 
3586
    toku_ft_handle_create(&result_ft_handle);
3587 3588 3589 3590 3591 3592 3593 3594 3595 3596 3597 3598

    // we're cloning, so the handle better have an open ft and open cf
    invariant(ft_handle->ft);
    invariant(ft_handle->ft->cf);

    // inherit the options of the ft whose handle is being cloned.
    toku_ft_handle_inherit_options(result_ft_handle, ft_handle->ft);

    // we can clone the handle by creating a new handle with the same fname
    CACHEFILE cf = ft_handle->ft->cf;
    CACHETABLE ct = toku_cachefile_get_cachetable(cf);
    const char *fname_in_env = toku_cachefile_fname_in_env(cf);
3599
    int r = toku_ft_handle_open(result_ft_handle, fname_in_env, false, false, ct, txn); 
3600 3601 3602 3603 3604 3605 3606 3607
    if (r != 0) {
        toku_ft_handle_close(result_ft_handle);
        result_ft_handle = NULL;
    }
    *cloned_ft_handle = result_ft_handle;
    return r;
}

3608
// Open a brt in normal use.  The FILENUM and dict_id are assigned by the ft_handle_open() function.
3609
int
3610 3611
toku_ft_handle_open_with_dict_id(
    FT_HANDLE t,
3612 3613 3614 3615 3616
    const char *fname_in_env,
    int is_create,
    int only_create,
    CACHETABLE cachetable,
    TOKUTXN txn,
3617
    DICTIONARY_ID use_dictionary_id
3618
    )
3619
{
3620
    int r;
3621
    r = ft_handle_open(
3622 3623 3624 3625 3626 3627 3628 3629
        t,
        fname_in_env,
        is_create,
        only_create,
        cachetable,
        txn,
        FILENUM_NONE,
        use_dictionary_id,
3630 3631
        MAX_LSN
        );
3632 3633 3634 3635
    return r;
}

DICTIONARY_ID
3636
toku_ft_get_dictionary_id(FT_HANDLE brt) {
3637
    FT h = brt->ft;
3638 3639 3640 3641
    DICTIONARY_ID dict_id = h->dict_id;
    return dict_id;
}

3642 3643 3644
void toku_ft_set_flags(FT_HANDLE ft_handle, unsigned int flags) {
    ft_handle->did_set_flags = true;
    ft_handle->options.flags = flags;
3645 3646
}

3647 3648
void toku_ft_get_flags(FT_HANDLE ft_handle, unsigned int *flags) {
    *flags = ft_handle->options.flags;
3649 3650
}

3651
void toku_ft_get_maximum_advised_key_value_lengths (unsigned int *max_key_len, unsigned int *max_val_len)
3652 3653 3654 3655 3656 3657
// return the maximum advisable key value lengths.  The brt doesn't enforce these.
{
    *max_key_len = 32*1024;
    *max_val_len = 32*1024*1024;
}

3658 3659 3660 3661 3662 3663 3664 3665

void toku_ft_handle_set_nodesize(FT_HANDLE ft_handle, unsigned int nodesize) {
    if (ft_handle->ft) {
        toku_ft_set_nodesize(ft_handle->ft, nodesize);
    }
    else {
        ft_handle->options.nodesize = nodesize;
    }
3666 3667
}

3668 3669 3670 3671 3672 3673 3674 3675 3676 3677 3678 3679 3680 3681 3682 3683 3684 3685 3686 3687 3688 3689 3690 3691 3692
void toku_ft_handle_get_nodesize(FT_HANDLE ft_handle, unsigned int *nodesize) {
    if (ft_handle->ft) {
        toku_ft_get_nodesize(ft_handle->ft, nodesize);
    }
    else {
        *nodesize = ft_handle->options.nodesize;
    }
}

void toku_ft_handle_set_basementnodesize(FT_HANDLE ft_handle, unsigned int basementnodesize) {
    if (ft_handle->ft) {
        toku_ft_set_basementnodesize(ft_handle->ft, basementnodesize);
    }
    else {
        ft_handle->options.basementnodesize = basementnodesize;
    }
}

void toku_ft_handle_get_basementnodesize(FT_HANDLE ft_handle, unsigned int *basementnodesize) {
    if (ft_handle->ft) {
        toku_ft_get_basementnodesize(ft_handle->ft, basementnodesize);
    }
    else {
        *basementnodesize = ft_handle->options.basementnodesize;
    }
3693 3694
}

3695
void toku_ft_set_bt_compare(FT_HANDLE brt, int (*bt_compare)(DB*, const DBT*, const DBT*)) {
3696
    brt->options.compare_fun = bt_compare;
3697 3698
}

3699
void toku_ft_set_redirect_callback(FT_HANDLE brt, on_redirect_callback redir_cb, void* extra) {
3700 3701 3702 3703
    brt->redirect_callback = redir_cb;
    brt->redirect_callback_extra = extra;
}

3704
void toku_ft_set_update(FT_HANDLE brt, ft_update_func update_fun) {
3705
    brt->options.update_fun = update_fun;
3706 3707
}

3708
ft_compare_func toku_ft_get_bt_compare (FT_HANDLE brt) {
3709
    return brt->options.compare_fun;
3710 3711
}

3712 3713
static void
ft_remove_handle_ref_callback(FT UU(ft), void *extra) {
3714
    FT_HANDLE CAST_FROM_VOIDP(handle, extra);
3715 3716 3717
    toku_list_remove(&handle->live_ft_handle_link);
}

3718 3719 3720 3721 3722 3723
// close an ft handle during normal operation. the underlying ft may or may not close,
// depending if there are still references. an lsn for this close will come from the logger.
void
toku_ft_handle_close(FT_HANDLE ft_handle) {
    // There are error paths in the ft_handle_open that end with ft_handle->ft==NULL.
    FT ft = ft_handle->ft;
3724
    if (ft) {
3725 3726
        const bool oplsn_valid = false;
        toku_ft_remove_reference(ft, oplsn_valid, ZERO_LSN, ft_remove_handle_ref_callback, ft_handle);
3727
    }
3728
    toku_free(ft_handle);
3729 3730
}

3731 3732 3733 3734 3735 3736 3737 3738 3739 3740 3741 3742 3743 3744 3745 3746 3747
// close an ft handle during recovery. the underlying ft must close, and will use the given lsn.
void 
toku_ft_handle_close_recovery(FT_HANDLE ft_handle, LSN oplsn) {
    FT ft = ft_handle->ft;
    // the ft must exist if closing during recovery. error paths during 
    // open for recovery should close handles using toku_ft_handle_close()
    assert(ft);
    const bool oplsn_valid = true;
    toku_ft_remove_reference(ft, oplsn_valid, oplsn, ft_remove_handle_ref_callback, ft_handle);
    toku_free(ft_handle);
}

// TODO: remove this, callers should instead just use toku_ft_handle_close()
int 
toku_close_ft_handle_nolsn (FT_HANDLE ft_handle, char** UU(error_string)) {
    toku_ft_handle_close(ft_handle);
    return 0;
3748 3749
}

3750 3751
void toku_ft_handle_create(FT_HANDLE *ft_handle_ptr) {
    FT_HANDLE XMALLOC(brt);
3752
    memset(brt, 0, sizeof *brt);
3753
    toku_list_init(&brt->live_ft_handle_link);
3754
    brt->options.flags = 0;
Yoni Fogel's avatar
Yoni Fogel committed
3755
    brt->did_set_flags = false;
3756 3757 3758 3759 3760
    brt->options.nodesize = FT_DEFAULT_NODE_SIZE;
    brt->options.basementnodesize = FT_DEFAULT_BASEMENT_NODE_SIZE;
    brt->options.compression_method = TOKU_DEFAULT_COMPRESSION_METHOD;
    brt->options.compare_fun = toku_builtin_compare_fun;
    brt->options.update_fun = NULL;
3761
    *ft_handle_ptr = brt;
3762
}
3763

3764 3765
/* ************* CURSORS ********************* */

3766
static inline void
3767
ft_cursor_cleanup_dbts(FT_CURSOR c) {
3768 3769 3770 3771
    if (c->key.data) toku_free(c->key.data);
    if (c->val.data) toku_free(c->val.data);
    memset(&c->key, 0, sizeof(c->key));
    memset(&c->val, 0, sizeof(c->val));
3772 3773
}

3774 3775 3776 3777 3778 3779 3780 3781 3782 3783 3784
//
// This function is used by the leafentry iterators.
// returns TOKUDB_ACCEPT if live transaction context is allowed to read a value
// that is written by transaction with LSN of id
// live transaction context may read value if either id is the root ancestor of context, or if
// id was committed before context's snapshot was taken.
// For id to be committed before context's snapshot was taken, the following must be true:
//  - id < context->snapshot_txnid64 AND id is not in context's live root transaction list
// For the above to NOT be true:
//  - id > context->snapshot_txnid64 OR id is in context's live root transaction list
//
3785
static int
3786
does_txn_read_entry(TXNID id, TOKUTXN context) {
3787 3788 3789
    int rval;
    TXNID oldest_live_in_snapshot = toku_get_oldest_in_live_root_txn_list(context);
    if (id < oldest_live_in_snapshot || id == context->ancestor_txnid64) {
3790
        rval = TOKUDB_ACCEPT;
3791
    }
3792
    else if (id > context->snapshot_txnid64 || toku_is_txn_in_live_root_txn_list(*context->live_root_txn_list, id)) {
3793
        rval = 0;
3794 3795
    }
    else {
3796
        rval = TOKUDB_ACCEPT;
3797 3798 3799 3800
    }
    return rval;
}

3801
static inline void
3802 3803
ft_cursor_extract_key_and_val(LEAFENTRY le,
                               FT_CURSOR cursor,
Yoni Fogel's avatar
Yoni Fogel committed
3804
                               uint32_t *keylen,
3805
                               void            **key,
Yoni Fogel's avatar
Yoni Fogel committed
3806
                               uint32_t *vallen,
3807
                               void            **val) {
3808
    if (toku_ft_cursor_is_leaf_mode(cursor)) {
3809 3810 3811
        *key = le_key_and_len(le, keylen);
        *val = le;
        *vallen = leafentry_memsize(le);
3812
    } else if (cursor->is_snapshot_read) {
3813 3814 3815 3816 3817 3818 3819 3820
        le_iterate_val(
            le,
            does_txn_read_entry,
            val,
            vallen,
            cursor->ttxn
            );
        *key = le_key_and_len(le, keylen);
3821
    } else {
3822 3823
        *key = le_key_and_len(le, keylen);
        *val = le_latest_val_and_len(le, vallen);
3824
    }
3825 3826
}

3827 3828 3829
int toku_ft_cursor (
    FT_HANDLE brt,
    FT_CURSOR *cursorptr,
3830
    TOKUTXN ttxn,
Yoni Fogel's avatar
Yoni Fogel committed
3831 3832
    bool is_snapshot_read,
    bool disable_prefetching
3833
    )
3834 3835
{
    if (is_snapshot_read) {
3836
        invariant(ttxn != NULL);
3837
        int accepted = does_txn_read_entry(brt->ft->h->root_xid_that_created, ttxn);
3838 3839 3840 3841
        if (accepted!=TOKUDB_ACCEPT) {
            invariant(accepted==0);
            return TOKUDB_MVCC_DICTIONARY_TOO_NEW;
        }
3842
    }
3843 3844
    FT_CURSOR XCALLOC(cursor);
#if 0
3845
    // if this cursor is to do read_committed fetches, then the txn objects must be valid.
3846
    if (cursor == 0)
3847
        return ENOMEM;
3848
    memset(cursor, 0, sizeof(*cursor));
3849
#endif
3850
    cursor->ft_handle = brt;
Yoni Fogel's avatar
Yoni Fogel committed
3851
    cursor->prefetching = false;
3852 3853
    toku_init_dbt(&cursor->range_lock_left_key);
    toku_init_dbt(&cursor->range_lock_right_key);
Yoni Fogel's avatar
Yoni Fogel committed
3854 3855
    cursor->left_is_neg_infty = false;
    cursor->right_is_pos_infty = false;
3856
    cursor->is_snapshot_read = is_snapshot_read;
Yoni Fogel's avatar
Yoni Fogel committed
3857
    cursor->is_leaf_mode = false;
3858
    cursor->ttxn = ttxn;
3859
    cursor->disable_prefetching = disable_prefetching;
Yoni Fogel's avatar
Yoni Fogel committed
3860
    cursor->is_temporary = false;
3861 3862 3863 3864
    *cursorptr = cursor;
    return 0;
}

3865
void
3866
toku_ft_cursor_set_temporary(FT_CURSOR ftcursor) {
Yoni Fogel's avatar
Yoni Fogel committed
3867
    ftcursor->is_temporary = true;
3868 3869
}

3870
void
3871
toku_ft_cursor_set_leaf_mode(FT_CURSOR ftcursor) {
Yoni Fogel's avatar
Yoni Fogel committed
3872
    ftcursor->is_leaf_mode = true;
3873 3874 3875
}

int
3876 3877
toku_ft_cursor_is_leaf_mode(FT_CURSOR ftcursor) {
    return ftcursor->is_leaf_mode;
3878 3879
}

3880
void
3881
toku_ft_cursor_set_range_lock(FT_CURSOR cursor, const DBT *left, const DBT *right,
Yoni Fogel's avatar
Yoni Fogel committed
3882
                               bool left_is_neg_infty, bool right_is_pos_infty)
3883
{
3884 3885
    // Destroy any existing keys and then clone the given left, right keys
    toku_destroy_dbt(&cursor->range_lock_left_key);
3886
    if (left_is_neg_infty) {
Yoni Fogel's avatar
Yoni Fogel committed
3887
        cursor->left_is_neg_infty = true;
3888
    } else {
3889
        toku_clone_dbt(&cursor->range_lock_left_key, *left);
3890
    }
3891 3892

    toku_destroy_dbt(&cursor->range_lock_right_key);
3893
    if (right_is_pos_infty) {
Yoni Fogel's avatar
Yoni Fogel committed
3894
        cursor->right_is_pos_infty = true;
3895
    } else {
3896
        toku_clone_dbt(&cursor->range_lock_right_key, *right);
3897 3898 3899
    }
}

3900
void toku_ft_cursor_close(FT_CURSOR cursor) {
3901
    ft_cursor_cleanup_dbts(cursor);
3902 3903
    toku_destroy_dbt(&cursor->range_lock_left_key);
    toku_destroy_dbt(&cursor->range_lock_right_key);
3904
    toku_free(cursor);
3905
}
3906

3907
static inline void ft_cursor_set_prefetching(FT_CURSOR cursor) {
Yoni Fogel's avatar
Yoni Fogel committed
3908
    cursor->prefetching = true;
3909 3910
}

Yoni Fogel's avatar
Yoni Fogel committed
3911
static inline bool ft_cursor_prefetching(FT_CURSOR cursor) {
3912 3913 3914
    return cursor->prefetching;
}

Yoni Fogel's avatar
Yoni Fogel committed
3915 3916
//Return true if cursor is uninitialized.  false otherwise.
static bool
3917
ft_cursor_not_set(FT_CURSOR cursor) {
3918
    assert((cursor->key.data==NULL) == (cursor->val.data==NULL));
Yoni Fogel's avatar
Yoni Fogel committed
3919
    return (bool)(cursor->key.data == NULL);
3920 3921 3922
}

static int
Yoni Fogel's avatar
Yoni Fogel committed
3923
pair_leafval_heaviside_le (uint32_t klen, void *kval,
3924
                           ft_search_t *search) {
3925
    DBT x;
3926
    int cmp = search->compare(search,
3927
                              search->k ? toku_fill_dbt(&x, kval, klen) : 0);
3928 3929
    // The search->compare function returns only 0 or 1
    switch (search->direction) {
3930 3931
    case FT_SEARCH_LEFT:   return cmp==0 ? -1 : +1;
    case FT_SEARCH_RIGHT:  return cmp==0 ? +1 : -1; // Because the comparison runs backwards for right searches.
3932
    }
3933
    abort(); return 0;
3934 3935 3936 3937
}


static int
3938
heaviside_from_search_t (OMTVALUE lev, void *extra) {
3939 3940
    LEAFENTRY CAST_FROM_VOIDP(le, lev);
    ft_search_t *CAST_FROM_VOIDP(search, extra);
Yoni Fogel's avatar
Yoni Fogel committed
3941
    uint32_t keylen;
3942
    void* key = le_key_and_len(le, &keylen);
3943

3944
    return pair_leafval_heaviside_le (keylen, key,
3945
                                      search);
3946 3947 3948
}


3949 3950 3951
//
// Returns true if the value that is to be read is empty.
//
3952
static inline int
3953
is_le_val_del(LEAFENTRY le, FT_CURSOR ftcursor) {
3954
    int rval;
3955
    if (ftcursor->is_snapshot_read) {
Yoni Fogel's avatar
Yoni Fogel committed
3956
        bool is_del;
3957 3958 3959 3960
        le_iterate_is_del(
            le,
            does_txn_read_entry,
            &is_del,
3961
            ftcursor->ttxn
3962 3963
            );
        rval = is_del;
3964 3965
    }
    else {
3966
        rval = le_latest_is_del(le);
3967
    }
3968
    return rval;
3969 3970
}

3971 3972
static const DBT zero_dbt = {0,0,0,0};

3973
static void search_save_bound (ft_search_t *search, DBT *pivot) {
3974
    if (search->have_pivot_bound) {
3975
        toku_free(search->pivot_bound.data);
3976 3977 3978 3979 3980
    }
    search->pivot_bound = zero_dbt;
    search->pivot_bound.data = toku_malloc(pivot->size);
    search->pivot_bound.size = pivot->size;
    memcpy(search->pivot_bound.data, pivot->data, pivot->size);
Yoni Fogel's avatar
Yoni Fogel committed
3981
    search->have_pivot_bound = true;
3982 3983
}

Yoni Fogel's avatar
Yoni Fogel committed
3984 3985 3986
static bool search_pivot_is_bounded (ft_search_t *search, DESCRIPTOR desc, ft_compare_func cmp, DBT *pivot) __attribute__((unused));
static bool search_pivot_is_bounded (ft_search_t *search, DESCRIPTOR desc, ft_compare_func cmp, DBT *pivot)
// Effect:  Return true iff the pivot has already been searched (for fixing #3522.)
3987 3988 3989
//  If searching from left to right, if we have already searched all the values less than pivot, we don't want to search again.
//  If searching from right to left, if we have already searched all the vlaues greater than pivot, we don't want to search again.
{
Yoni Fogel's avatar
Yoni Fogel committed
3990
    if (!search->have_pivot_bound) return true; // isn't bounded.
Zardosht Kasheff's avatar
Zardosht Kasheff committed
3991
    FAKE_DB(db, desc);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
3992
    int comp = cmp(&db, pivot, &search->pivot_bound);
3993
    if (search->direction == FT_SEARCH_LEFT) {
3994 3995
        // searching from left to right.  If the comparison function says the pivot is <= something we already compared, don't do it again.
        return comp>0;
3996
    } else {
3997
        return comp<0;
3998 3999 4000
    }
}

4001
struct store_fifo_offset_extra {
4002
    int32_t *offsets;
4003 4004 4005
    int i;
};

4006 4007
int store_fifo_offset(const int32_t &offset, const uint32_t UU(idx), struct store_fifo_offset_extra *const extra) __attribute__((nonnull(3)));
int store_fifo_offset(const int32_t &offset, const uint32_t UU(idx), struct store_fifo_offset_extra *const extra)
4008
{
4009 4010
    extra->offsets[extra->i] = offset;
    extra->i++;
4011 4012 4013
    return 0;
}

4014 4015 4016 4017 4018
/**
 * Given pointers to offsets within a FIFO where we can find messages,
 * figure out the MSN of each message, and compare those MSNs.  Returns 1,
 * 0, or -1 if a is larger than, equal to, or smaller than b.
 */
4019 4020
int fifo_offset_msn_cmp(FIFO &fifo, const int32_t &ao, const int32_t &bo);
int fifo_offset_msn_cmp(FIFO &fifo, const int32_t &ao, const int32_t &bo)
4021
{
4022 4023
    const struct fifo_entry *a = toku_fifo_get_entry(fifo, ao);
    const struct fifo_entry *b = toku_fifo_get_entry(fifo, bo);
4024 4025 4026 4027 4028 4029 4030
    if (a->msn.msn > b->msn.msn) {
        return +1;
    }
    if (a->msn.msn < b->msn.msn) {
        return -1;
    }
    return 0;
4031 4032
}

4033 4034
/**
 * Given a fifo_entry, either decompose it into its parameters and call
4035
 * toku_ft_bn_apply_cmd, or discard it, based on its MSN and the MSN of the
4036 4037
 * basement node.
 */
4038
static void
4039
do_bn_apply_cmd(FT_HANDLE t, BASEMENTNODE bn, struct fifo_entry *entry, uint64_t *workdone, STAT64INFO stats_to_update)
4040
{
4041 4042 4043 4044
    // The messages are being iterated over in (key,msn) order or just in
    // msn order, so all the messages for one key, from one buffer, are in
    // ascending msn order.  So it's ok that we don't update the basement
    // node's msn until the end.
4045 4046 4047
    if (entry->msn.msn > bn->max_msn_applied.msn) {
        ITEMLEN keylen = entry->keylen;
        ITEMLEN vallen = entry->vallen;
4048
        enum ft_msg_type type = fifo_entry_get_msg_type(entry);
4049 4050 4051
        MSN msn = entry->msn;
        const XIDS xids = (XIDS) &entry->xids_s;
        bytevec key = xids_get_end_of_array(xids);
Yoni Fogel's avatar
Yoni Fogel committed
4052
        bytevec val = (uint8_t*)key + entry->keylen;
4053 4054 4055 4056

        DBT hk;
        toku_fill_dbt(&hk, key, keylen);
        DBT hv;
4057
        FT_MSG_S ftcmd = { type, msn, xids, .u = { .id = { &hk, toku_fill_dbt(&hv, val, vallen) } } };
4058
        toku_ft_bn_apply_cmd(
4059 4060 4061
            t->ft->compare_fun,
            t->ft->update_fun,
            &t->ft->cmp_descriptor,
4062
            bn,
4063
            &ftcmd,
4064
            workdone,
4065 4066
            stats_to_update
            );
4067
    } else {
4068
        STATUS_INC(FT_MSN_DISCARDS, 1);
4069
    }
4070 4071 4072 4073 4074
    // We must always mark entry as stale since it has been marked
    // (using omt::iterate_and_mark_range)
    // It is possible to call do_bn_apply_cmd even when it won't apply the message because
    // the node containing it could have been evicted and brought back in.
    entry->is_fresh = false;
4075 4076
}

4077
struct iterate_do_bn_apply_cmd_extra {
4078
    FT_HANDLE t;
4079
    BASEMENTNODE bn;
4080 4081
    NONLEAF_CHILDINFO bnc;
    uint64_t *workdone;
4082
    STAT64INFO stats_to_update;
4083 4084
};

4085 4086
int iterate_do_bn_apply_cmd(const int32_t &offset, const uint32_t UU(idx), struct iterate_do_bn_apply_cmd_extra *const e) __attribute__((nonnull(3)));
int iterate_do_bn_apply_cmd(const int32_t &offset, const uint32_t UU(idx), struct iterate_do_bn_apply_cmd_extra *const e)
4087
{
4088 4089
    struct fifo_entry *entry = toku_fifo_get_entry(e->bnc->buffer, offset);
    do_bn_apply_cmd(e->t, e->bn, entry, e->workdone, e->stats_to_update);
4090 4091 4092
    return 0;
}

4093 4094 4095 4096 4097 4098 4099 4100 4101 4102 4103 4104 4105 4106 4107
/**
 * Given the bounds of the basement node to which we will apply messages,
 * find the indexes within message_tree which contain the range of
 * relevant messages.
 *
 * The message tree contains offsets into the buffer, where messages are
 * found.  The pivot_bounds are the lower bound exclusive and upper bound
 * inclusive, because they come from pivot keys in the tree.  We want OMT
 * indices, which must have the lower bound be inclusive and the upper
 * bound exclusive.  We will get these by telling toku_omt_find to look
 * for something strictly bigger than each of our pivot bounds.
 *
 * Outputs the OMT indices in lbi (lower bound inclusive) and ube (upper
 * bound exclusive).
 */
4108
template<typename find_bounds_omt_t>
4109
static void
4110 4111
find_bounds_within_message_tree(
    DESCRIPTOR desc,       /// used for cmp
4112
    ft_compare_func cmp,  /// used to compare keys
4113
    const find_bounds_omt_t &message_tree,      /// tree holding FIFO offsets, in which we want to look for indices
4114 4115
    FIFO buffer,           /// buffer in which messages are found
    struct pivot_bounds const * const bounds,  /// key bounds within the basement node we're applying messages to
4116 4117
    uint32_t *lbi,        /// (output) "lower bound inclusive" (index into message_tree)
    uint32_t *ube         /// (output) "upper bound exclusive" (index into message_tree)
4118 4119 4120
    )
{
    int r = 0;
4121 4122

    if (bounds->lower_bound_exclusive) {
4123 4124 4125 4126 4127
        // By setting msn to MAX_MSN and by using direction of +1, we will
        // get the first message greater than (in (key, msn) order) any
        // message (with any msn) with the key lower_bound_exclusive.
        // This will be a message we want to try applying, so it is the
        // "lower bound inclusive" within the message_tree.
4128 4129 4130 4131 4132 4133 4134 4135 4136
        struct toku_fifo_entry_key_msn_heaviside_extra lbi_extra;
        ZERO_STRUCT(lbi_extra);
        lbi_extra.desc = desc;
        lbi_extra.cmp = cmp;
        lbi_extra.fifo = buffer;
        lbi_extra.key = bounds->lower_bound_exclusive;
        lbi_extra.msn = MAX_MSN;
        int32_t found_lb;
        r = message_tree.template find<struct toku_fifo_entry_key_msn_heaviside_extra, toku_fifo_entry_key_msn_heaviside>(lbi_extra, +1, &found_lb, lbi);
4137
        if (r == DB_NOTFOUND) {
4138 4139 4140
            // There is no relevant data (the lower bound is bigger than
            // any message in this tree), so we have no range and we're
            // done.
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4141 4142
            *lbi = 0;
            *ube = 0;
4143
            return;
4144 4145
        }
        if (bounds->upper_bound_inclusive) {
4146
            // Check if what we found for lbi is greater than the upper
4147 4148
            // bound inclusive that we have.  If so, there are no relevant
            // messages between these bounds.
4149
            const DBT *ubi = bounds->upper_bound_inclusive;
4150
            const int32_t offset = found_lb;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4151 4152
            DBT found_lbidbt;
            fill_dbt_for_fifo_entry(&found_lbidbt, toku_fifo_get_entry(buffer, offset));
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4153
            FAKE_DB(db, desc);
4154
            int c = cmp(&db, &found_lbidbt, ubi);
4155 4156 4157 4158 4159
            // These DBTs really are both inclusive bounds, so we need
            // strict inequality in order to determine that there's
            // nothing between them.  If they're equal, then we actually
            // need to apply the message pointed to by lbi, and also
            // anything with the same key but a bigger msn.
4160
            if (c > 0) {
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4161 4162
                *lbi = 0;
                *ube = 0;
4163
                return;
4164 4165
            }
        }
4166
    } else {
4167 4168
        // No lower bound given, it's negative infinity, so we start at
        // the first message in the OMT.
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4169
        *lbi = 0;
4170 4171
    }
    if (bounds->upper_bound_inclusive) {
4172 4173 4174 4175
        // Again, we use an msn of MAX_MSN and a direction of +1 to get
        // the first thing bigger than the upper_bound_inclusive key.
        // This is therefore the smallest thing we don't want to apply,
        // and toku_omt_iterate_on_range will not examine it.
4176 4177 4178 4179 4180 4181 4182 4183
        struct toku_fifo_entry_key_msn_heaviside_extra ube_extra;
        ZERO_STRUCT(ube_extra);
        ube_extra.desc = desc;
        ube_extra.cmp = cmp;
        ube_extra.fifo = buffer;
        ube_extra.key = bounds->upper_bound_inclusive;
        ube_extra.msn = MAX_MSN;
        r = message_tree.template find<struct toku_fifo_entry_key_msn_heaviside_extra, toku_fifo_entry_key_msn_heaviside>(ube_extra, +1, nullptr, ube);
4184
        if (r == DB_NOTFOUND) {
4185 4186 4187
            // Couldn't find anything in the buffer bigger than our key,
            // so we need to look at everything up to the end of
            // message_tree.
4188
            *ube = message_tree.size();
4189
        }
4190
    } else {
4191 4192
        // No upper bound given, it's positive infinity, so we need to go
        // through the end of the OMT.
4193
        *ube = message_tree.size();
4194 4195 4196
    }
}

4197 4198 4199 4200 4201 4202 4203
/**
 * For each message in the ancestor's buffer (determined by childnum) that
 * is key-wise between lower_bound_exclusive and upper_bound_inclusive,
 * apply the message to the basement node.  We treat the bounds as minus
 * or plus infinity respectively if they are NULL.  Do not mark the node
 * as dirty (preserve previous state of 'dirty' bit).
 */
4204
static void
4205
bnc_apply_messages_to_basement_node(
4206
    FT_HANDLE t,             // used for comparison function
4207
    BASEMENTNODE bn,   // where to apply messages
4208
    FTNODE ancestor,  // the ancestor node where we can find messages to apply
4209
    int childnum,      // which child buffer of ancestor contains messages we want
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4210
    struct pivot_bounds const * const bounds,  // contains pivot key bounds of this basement node
Yoni Fogel's avatar
Yoni Fogel committed
4211
    bool* msgs_applied
4212 4213 4214 4215
    )
{
    int r;
    NONLEAF_CHILDINFO bnc = BNC(ancestor, childnum);
4216

4217 4218
    // Determine the offsets in the message trees between which we need to
    // apply messages from this buffer
4219
    STAT64INFO_S stats_delta = {0,0};
4220 4221
    uint64_t workdone_this_ancestor = 0;

4222
    uint32_t stale_lbi, stale_ube;
4223
    if (!bn->stale_ancestor_messages_applied) {
4224
        find_bounds_within_message_tree(&t->ft->cmp_descriptor, t->ft->compare_fun, bnc->stale_message_tree, bnc->buffer, bounds, &stale_lbi, &stale_ube);
4225
    } else {
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4226 4227
        stale_lbi = 0;
        stale_ube = 0;
4228
    }
Yoni Fogel's avatar
Yoni Fogel committed
4229
    uint32_t fresh_lbi, fresh_ube;
4230
    find_bounds_within_message_tree(&t->ft->cmp_descriptor, t->ft->compare_fun, bnc->fresh_message_tree, bnc->buffer, bounds, &fresh_lbi, &fresh_ube);
4231

4232 4233
    // We now know where all the messages we must apply are, so one of the
    // following 4 cases will do the application, depending on which of
4234 4235
    // the lists contains relevant messages:
    //
4236
    // 1. broadcast messages and anything else, or a mix of fresh and stale
4237 4238
    // 2. only fresh messages
    // 3. only stale messages
4239 4240 4241
    if (bnc->broadcast_list.size() > 0 ||
        (stale_lbi != stale_ube && fresh_lbi != fresh_ube)) {
        // We have messages in multiple trees, so we grab all
4242 4243
        // the relevant messages' offsets and sort them by MSN, then apply
        // them in MSN order.
4244
        const int buffer_size = ((stale_ube - stale_lbi) + (fresh_ube - fresh_lbi) + bnc->broadcast_list.size());
4245
        int32_t *XMALLOC_N(buffer_size, offsets);
4246
        struct store_fifo_offset_extra sfo_extra = { .offsets = offsets, .i = 0 };
4247

4248
        // Populate offsets array with offsets to stale messages
4249
        r = bnc->stale_message_tree.iterate_on_range<struct store_fifo_offset_extra, store_fifo_offset>(stale_lbi, stale_ube, &sfo_extra);
4250
        assert_zero(r);
4251

4252 4253
        // Then store fresh offsets, and mark them to be moved to stale later.
        r = bnc->fresh_message_tree.iterate_and_mark_range<struct store_fifo_offset_extra, store_fifo_offset>(fresh_lbi, fresh_ube, &sfo_extra);
4254 4255
        assert_zero(r);

4256
        // Store offsets of all broadcast messages.
4257
        r = bnc->broadcast_list.iterate<struct store_fifo_offset_extra, store_fifo_offset>(&sfo_extra);
4258
        assert_zero(r);
4259
        invariant(sfo_extra.i == buffer_size);
4260 4261

        // Sort by MSN.
4262
        r = toku::sort<int32_t, FIFO, fifo_offset_msn_cmp>::mergesort_r(offsets, buffer_size, bnc->buffer);
4263 4264
        assert_zero(r);

4265
        // Apply the messages in MSN order.
4266
        for (int i = 0; i < buffer_size; ++i) {
Yoni Fogel's avatar
Yoni Fogel committed
4267
            *msgs_applied = true;
4268
            struct fifo_entry *entry = toku_fifo_get_entry(bnc->buffer, offsets[i]);
4269
            do_bn_apply_cmd(t, bn, entry, &workdone_this_ancestor, &stats_delta);
4270
        }
4271

4272
        toku_free(offsets);
4273
    } else if (stale_lbi == stale_ube) {
4274
        // No stale messages to apply, we just apply fresh messages, and mark them to be moved to stale later.
4275
        struct iterate_do_bn_apply_cmd_extra iter_extra = { .t = t, .bn = bn, .bnc = bnc, .workdone = &workdone_this_ancestor, .stats_to_update = &stats_delta };
Yoni Fogel's avatar
Yoni Fogel committed
4276
        if (fresh_ube - fresh_lbi > 0) *msgs_applied = true;
4277
        r = bnc->fresh_message_tree.iterate_and_mark_range<struct iterate_do_bn_apply_cmd_extra, iterate_do_bn_apply_cmd>(fresh_lbi, fresh_ube, &iter_extra);
4278
        assert_zero(r);
4279 4280
    } else {
        invariant(fresh_lbi == fresh_ube);
4281 4282
        // No fresh messages to apply, we just apply stale messages.

Yoni Fogel's avatar
Yoni Fogel committed
4283
        if (stale_ube - stale_lbi > 0) *msgs_applied = true;
4284
        struct iterate_do_bn_apply_cmd_extra iter_extra = { .t = t, .bn = bn, .bnc = bnc, .workdone = &workdone_this_ancestor, .stats_to_update = &stats_delta };
4285

4286
        r = bnc->stale_message_tree.iterate_on_range<struct iterate_do_bn_apply_cmd_extra, iterate_do_bn_apply_cmd>(stale_lbi, stale_ube, &iter_extra);
4287
        assert_zero(r);
4288
    }
4289 4290 4291
    //
    // update stats
    //
4292 4293 4294
    if (workdone_this_ancestor > 0) {
        (void) toku_sync_fetch_and_add(&BP_WORKDONE(ancestor, childnum), workdone_this_ancestor);
    }
4295
    if (stats_delta.numbytes || stats_delta.numrows) {
4296
        toku_ft_update_stats(&t->ft->in_memory_stats, stats_delta);
4297
    }
4298 4299
}

Zardosht Kasheff's avatar
Zardosht Kasheff committed
4300
void
4301
toku_apply_ancestors_messages_to_node (FT_HANDLE t, FTNODE node, ANCESTORS ancestors, struct pivot_bounds const * const bounds, bool* msgs_applied)
4302
// Effect:
4303
//   Bring a leaf node up-to-date according to all the messages in the ancestors.
4304
//   If the leaf node is already up-to-date then do nothing.
4305 4306 4307 4308 4309
//   If the leaf node is not already up-to-date, then record the work done
//   for that leaf in each ancestor.
// Requires:
//   This is being called when pinning a leaf node for the query path.
//   The entire root-to-leaf path is pinned and appears in the ancestors list.
4310 4311
{
    VERIFY_NODE(t, node);
4312
    invariant(node->height == 0);
4313
    // know we are a leaf node
4314 4315 4316 4317 4318
    // An important invariant:
    // We MUST bring every available basement node up to date.
    // flushing on the cleaner thread depends on this. This invariant
    // allows the cleaner thread to just pick an internal node and flush it
    // as opposed to being forced to start from the root.
4319
    for (int i = 0; i < node->n_children; i++) {
4320
        if (BP_STATE(node, i) != PT_AVAIL) { continue; }
4321 4322 4323 4324
        BASEMENTNODE curr_bn = BLB(node, i);
        struct pivot_bounds curr_bounds = next_pivot_keys(node, i, bounds);
        for (ANCESTORS curr_ancestors = ancestors; curr_ancestors; curr_ancestors = curr_ancestors->next) {
            if (curr_ancestors->node->max_msn_applied_to_node_on_disk.msn > curr_bn->max_msn_applied.msn) {
4325
                paranoid_invariant(BP_STATE(curr_ancestors->node, curr_ancestors->childnum) == PT_AVAIL);
4326
                bnc_apply_messages_to_basement_node(
4327 4328 4329 4330
                    t,
                    curr_bn,
                    curr_ancestors->node,
                    curr_ancestors->childnum,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4331 4332
                    &curr_bounds,
                    msgs_applied
4333
                    );
4334 4335
                // We don't want to check this ancestor node again if the
                // next time we query it, the msn hasn't changed.
4336 4337 4338
                curr_bn->max_msn_applied = curr_ancestors->node->max_msn_applied_to_node_on_disk;
            }
        }
4339 4340 4341 4342 4343
        // At this point, we know all the stale messages above this
        // basement node have been applied, and any new messages will be
        // fresh, so we don't need to look at stale messages for this
        // basement node, unless it gets evicted (and this field becomes
        // false when it's read in again).
4344
        curr_bn->stale_ancestor_messages_applied = true;
4345 4346
    }
    VERIFY_NODE(t, node);
4347 4348
}

Leif Walsh's avatar
Leif Walsh committed
4349 4350 4351 4352 4353 4354 4355 4356 4357 4358 4359 4360 4361 4362 4363 4364 4365 4366 4367 4368 4369 4370 4371 4372 4373 4374 4375 4376
bool toku_ft_leaf_needs_ancestors_messages(FT ft, FTNODE node, ANCESTORS ancestors, struct pivot_bounds const * const bounds, MSN *const max_msn_in_path)
// Effect: Determine whether there are messages in a node's ancestors
//  which must be applied to it.  These messages are in the correct
//  keyrange for any available basement nodes, and are in nodes with the
//  correct max_msn_applied_to_node_on_disk.
// Notes:
//  This is an approximate query.
// Output:
//  max_msn_in_path: max of "max_msn_applied_to_node_on_disk" over
//    ancestors.  This is used later to update basement nodes'
//    max_msn_applied values in case we don't do the full algorithm.
// Returns:
//  true if there may be some such messages
//  false only if there are definitely no such messages
// Rationale:
//  When we pin a node with a read lock, we want to quickly determine if
//  we should exchange it for a write lock in preparation for applying
//  messages.  If there are no messages, we don't need the write lock.
{
    invariant(node->height == 0);
    MSN max_msn_applied = ZERO_MSN;
    bool needs_ancestors_messages = false;
    for (int i = 0; i < node->n_children; ++i) {
        if (BP_STATE(node, i) != PT_AVAIL) { continue; }
        BASEMENTNODE bn = BLB(node, i);
        struct pivot_bounds curr_bounds = next_pivot_keys(node, i, bounds);
        for (ANCESTORS curr_ancestors = ancestors; curr_ancestors; curr_ancestors = curr_ancestors->next) {
            if (curr_ancestors->node->max_msn_applied_to_node_on_disk.msn > bn->max_msn_applied.msn) {
4377
                paranoid_invariant(BP_STATE(curr_ancestors->node, curr_ancestors->childnum) == PT_AVAIL);
Leif Walsh's avatar
Leif Walsh committed
4378 4379 4380 4381 4382 4383 4384 4385 4386 4387 4388 4389 4390 4391 4392 4393 4394 4395 4396 4397 4398 4399 4400 4401 4402 4403 4404 4405 4406 4407 4408 4409 4410 4411 4412 4413 4414 4415 4416 4417 4418 4419 4420 4421 4422 4423 4424 4425
                NONLEAF_CHILDINFO bnc = BNC(curr_ancestors->node, curr_ancestors->childnum);
                if (bnc->broadcast_list.size() > 0) {
                    needs_ancestors_messages = true;
                    goto cleanup;
                }
                if (!bn->stale_ancestor_messages_applied) {
                    uint32_t stale_lbi, stale_ube;
                    find_bounds_within_message_tree(&ft->cmp_descriptor,
                                                    ft->compare_fun,
                                                    bnc->stale_message_tree,
                                                    bnc->buffer,
                                                    &curr_bounds,
                                                    &stale_lbi,
                                                    &stale_ube);
                    if (stale_lbi < stale_ube) {
                        needs_ancestors_messages = true;
                        goto cleanup;
                    }
                }
                uint32_t fresh_lbi, fresh_ube;
                find_bounds_within_message_tree(&ft->cmp_descriptor,
                                                ft->compare_fun,
                                                bnc->fresh_message_tree,
                                                bnc->buffer,
                                                &curr_bounds,
                                                &fresh_lbi,
                                                &fresh_ube);
                if (fresh_lbi < fresh_ube) {
                    needs_ancestors_messages = true;
                    goto cleanup;
                }
                if (curr_ancestors->node->max_msn_applied_to_node_on_disk.msn > max_msn_applied.msn) {
                    max_msn_applied = curr_ancestors->node->max_msn_applied_to_node_on_disk;
                }
            }
        }
    }
    *max_msn_in_path = max_msn_applied;
cleanup:
    return needs_ancestors_messages;
}

void toku_ft_bn_update_max_msn(FTNODE node, MSN max_msn_applied) {
    invariant(node->height == 0);
    for (int i = 0; i < node->n_children; ++i) {
        if (BP_STATE(node, i) != PT_AVAIL) { continue; }
        BASEMENTNODE bn = BLB(node, i);
        if (max_msn_applied.msn > bn->max_msn_applied.msn) {
4426 4427 4428 4429 4430
            // This function runs in a shared access context, so to silence tools
            // like DRD, we use a CAS and ignore the result.
            // Any threads trying to update these basement nodes should be
            // updating them to the same thing (since they all have a read lock on
            // the same root-to-leaf path) so this is safe.
4431
            (void) toku_sync_val_compare_and_swap(&bn->max_msn_applied.msn, bn->max_msn_applied.msn, max_msn_applied.msn);
Leif Walsh's avatar
Leif Walsh committed
4432 4433 4434 4435
        }
    }
}

4436 4437 4438 4439 4440
struct copy_to_stale_extra {
    FT ft;
    NONLEAF_CHILDINFO bnc;
};

4441 4442
int copy_to_stale(const int32_t &offset, const uint32_t UU(idx), struct copy_to_stale_extra *const extra) __attribute__((nonnull(3)));
int copy_to_stale(const int32_t &offset, const uint32_t UU(idx), struct copy_to_stale_extra *const extra)
4443 4444 4445 4446 4447 4448 4449 4450 4451 4452 4453 4454 4455 4456 4457 4458 4459 4460 4461 4462 4463 4464 4465 4466 4467 4468 4469 4470 4471 4472
{
    struct fifo_entry *entry = toku_fifo_get_entry(extra->bnc->buffer, offset);
    DBT keydbt;
    DBT *key = fill_dbt_for_fifo_entry(&keydbt, entry);
    struct toku_fifo_entry_key_msn_heaviside_extra heaviside_extra = { .desc = &extra->ft->cmp_descriptor, .cmp = extra->ft->compare_fun, .fifo = extra->bnc->buffer, .key = key, .msn = entry->msn };
    int r = extra->bnc->stale_message_tree.insert<struct toku_fifo_entry_key_msn_heaviside_extra, toku_fifo_entry_key_msn_heaviside>(offset, heaviside_extra, nullptr);
    invariant_zero(r);
    return 0;
}

__attribute__((nonnull))
void
toku_move_ftnode_messages_to_stale(FT ft, FTNODE node) {
    invariant(node->height > 0);
    // TODO: could be cilkified
    for (int i = 0; i < node->n_children; ++i) {
        if (BP_STATE(node, i) != PT_AVAIL) {
            continue;
        }
        NONLEAF_CHILDINFO bnc = BNC(node, i);
        // We can't delete things out of the fresh tree inside the above
        // procedures because we're still looking at the fresh tree.  Instead
        // we have to move messages after we're done looking at it.
        struct copy_to_stale_extra cts_extra = { .ft = ft, .bnc = bnc };
        int r = bnc->fresh_message_tree.iterate_over_marked<struct copy_to_stale_extra, copy_to_stale>(&cts_extra);
        invariant_zero(r);
        bnc->fresh_message_tree.delete_all_marked();
    }
}

4473
static int
4474 4475
ft_cursor_shortcut (
    FT_CURSOR cursor,
4476
    int direction,
4477
    FT_GET_CALLBACK_FUNCTION getf,
4478
    void *getf_v,
Yoni Fogel's avatar
Yoni Fogel committed
4479
    uint32_t *keylen,
4480
    void **key,
Yoni Fogel's avatar
Yoni Fogel committed
4481
    uint32_t *vallen,
4482 4483 4484
    void **val
    );

4485 4486
// This is a bottom layer of the search functions.
static int
4487
ft_search_basement_node(
4488
    BASEMENTNODE bn,
4489 4490
    ft_search_t *search,
    FT_GET_CALLBACK_FUNCTION getf,
4491
    void *getf_v,
Yoni Fogel's avatar
Yoni Fogel committed
4492
    bool *doprefetch,
4493
    FT_CURSOR ftcursor,
Yoni Fogel's avatar
Yoni Fogel committed
4494
    bool can_bulk_fetch
4495
    )
4496
{
4497
    // Now we have to convert from ft_search_t to the heaviside function with a direction.  What a pain...
4498

4499 4500
    int direction;
    switch (search->direction) {
4501 4502
    case FT_SEARCH_LEFT:   direction = +1; goto ok;
    case FT_SEARCH_RIGHT:  direction = -1; goto ok;
4503 4504
    }
    return EINVAL;  // This return and the goto are a hack to get both compile-time and run-time checking on enum
4505
ok: ;
4506
    OMTVALUE datav;
Yoni Fogel's avatar
Yoni Fogel committed
4507
    uint32_t idx = 0;
4508
    int r = toku_omt_find(bn->buffer,
4509 4510 4511
                          heaviside_from_search_t,
                          search,
                          direction,
4512
                          &datav, &idx);
4513 4514
    if (r!=0) return r;

4515
    LEAFENTRY CAST_FROM_VOIDP(le, datav);
4516
    if (toku_ft_cursor_is_leaf_mode(ftcursor))
4517
        goto got_a_good_value;        // leaf mode cursors see all leaf entries
4518
    if (is_le_val_del(le,ftcursor)) {
4519 4520 4521 4522
        // Provisionally deleted stuff is gone.
        // So we need to scan in the direction to see if we can find something
        while (1) {
            switch (search->direction) {
4523
            case FT_SEARCH_LEFT:
4524
                idx++;
4525 4526
                if (idx >= toku_omt_size(bn->buffer))
                    return DB_NOTFOUND;
4527
                break;
4528
            case FT_SEARCH_RIGHT:
4529
                if (idx == 0)
4530
                    return DB_NOTFOUND;
4531 4532 4533
                idx--;
                break;
            default:
4534
                abort();
4535
            }
4536
            r = toku_omt_fetch(bn->buffer, idx, &datav);
4537
            assert_zero(r); // we just validated the index
4538
            CAST_FROM_VOIDP(le, datav);
4539
            if (!is_le_val_del(le,ftcursor)) goto got_a_good_value;
4540
        }
4541 4542
    }
got_a_good_value:
4543
    {
Yoni Fogel's avatar
Yoni Fogel committed
4544
        uint32_t keylen;
4545
        void *key;
Yoni Fogel's avatar
Yoni Fogel committed
4546
        uint32_t vallen;
4547 4548
        void *val;

4549 4550
        ft_cursor_extract_key_and_val(le,
                                       ftcursor,
4551 4552 4553 4554
                                       &keylen,
                                       &key,
                                       &vallen,
                                       &val
4555
            );
4556

4557
        r = getf(keylen, key, vallen, val, getf_v, false);
4558
        if (r==0 || r == TOKUDB_CURSOR_CONTINUE) {
4559 4560
            ftcursor->leaf_info.to_be.omt   = bn->buffer;
            ftcursor->leaf_info.to_be.index = idx;
4561

4562
            if (r == TOKUDB_CURSOR_CONTINUE && can_bulk_fetch) {
4563 4564
                r = ft_cursor_shortcut(
                    ftcursor,
4565 4566 4567 4568 4569 4570 4571 4572 4573 4574
                    direction,
                    getf,
                    getf_v,
                    &keylen,
                    &key,
                    &vallen,
                    &val
                    );
            }

4575 4576 4577 4578 4579 4580
            ft_cursor_cleanup_dbts(ftcursor);
            if (!ftcursor->is_temporary) {
                ftcursor->key.data = toku_memdup(key, keylen);
                ftcursor->val.data = toku_memdup(val, vallen);
                ftcursor->key.size = keylen;
                ftcursor->val.size = vallen;
4581
            }
4582
            //The search was successful.  Prefetching can continue.
Yoni Fogel's avatar
Yoni Fogel committed
4583
            *doprefetch = true;
4584
        }
4585
    }
4586
    if (r == TOKUDB_CURSOR_CONTINUE) r = 0;
4587
    return r;
4588 4589 4590
}

static int
4591 4592 4593 4594
ft_search_node (
    FT_HANDLE brt,
    FTNODE node,
    ft_search_t *search,
4595
    int child_to_search,
4596
    FT_GET_CALLBACK_FUNCTION getf,
4597
    void *getf_v,
Yoni Fogel's avatar
Yoni Fogel committed
4598
    bool *doprefetch,
4599
    FT_CURSOR ftcursor,
4600 4601
    UNLOCKERS unlockers,
    ANCESTORS,
4602
    struct pivot_bounds const * const bounds,
Yoni Fogel's avatar
Yoni Fogel committed
4603
    bool can_bulk_fetch
4604
    );
4605 4606

// the number of nodes to prefetch
4607
#define TOKU_DO_PREFETCH 1
4608 4609
#if TOKU_DO_PREFETCH

4610
static int
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4611
ftnode_fetch_callback_and_free_bfe(CACHEFILE cf, PAIR p, int fd, BLOCKNUM nodename, uint32_t fullhash, void **ftnode_pv, void** UU(disk_data), PAIR_ATTR *sizep, int *dirtyp, void *extraargs)
4612
{
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4613
    int r = toku_ftnode_fetch_callback(cf, p, fd, nodename, fullhash, ftnode_pv, disk_data, sizep, dirtyp, extraargs);
4614
    struct ftnode_fetch_extra *CAST_FROM_VOIDP(ffe, extraargs);
4615 4616
    destroy_bfe_for_prefetch(ffe);
    toku_free(ffe);
4617 4618 4619 4620
    return r;
}

static int
4621
ftnode_pf_callback_and_free_bfe(void *ftnode_pv, void* disk_data, void *read_extraargs, int fd, PAIR_ATTR *sizep)
4622
{
4623
    int r = toku_ftnode_pf_callback(ftnode_pv, disk_data, read_extraargs, fd, sizep);
4624
    struct ftnode_fetch_extra *CAST_FROM_VOIDP(ffe, read_extraargs);
4625 4626
    destroy_bfe_for_prefetch(ffe);
    toku_free(ffe);
4627 4628 4629
    return r;
}

4630
static void
Yoni Fogel's avatar
Yoni Fogel committed
4631
ft_node_maybe_prefetch(FT_HANDLE brt, FTNODE node, int childnum, FT_CURSOR ftcursor, bool *doprefetch) {
Rich Prohaska's avatar
Rich Prohaska committed
4632

4633
    // if we want to prefetch in the tree
4634
    // then prefetch the next children if there are any
4635 4636
    if (*doprefetch && ft_cursor_prefetching(ftcursor) && !ftcursor->disable_prefetching) {
        int rc = ft_cursor_rightmost_child_wanted(ftcursor, brt, node);
4637 4638
        for (int i = childnum + 1; (i <= childnum + TOKU_DO_PREFETCH) && (i <= rc); i++) {
            BLOCKNUM nextchildblocknum = BP_BLOCKNUM(node, i);
Yoni Fogel's avatar
Yoni Fogel committed
4639
            uint32_t nextfullhash = compute_child_fullhash(brt->ft->cf, node, i);
4640
            struct ftnode_fetch_extra *MALLOC(bfe);
4641
            fill_bfe_for_prefetch(bfe, brt->ft, ftcursor);
Yoni Fogel's avatar
Yoni Fogel committed
4642
            bool doing_prefetch = false;
4643
            toku_cachefile_prefetch(
4644
                brt->ft->cf,
4645 4646
                nextchildblocknum,
                nextfullhash,
4647
                get_write_callbacks_for_node(brt->ft),
4648 4649 4650
                ftnode_fetch_callback_and_free_bfe,
                toku_ftnode_pf_req_callback,
                ftnode_pf_callback_and_free_bfe,
4651 4652 4653 4654 4655 4656 4657
                bfe,
                &doing_prefetch
                );
            if (!doing_prefetch) {
                destroy_bfe_for_prefetch(bfe);
                toku_free(bfe);
            }
Yoni Fogel's avatar
Yoni Fogel committed
4658
            *doprefetch = false;
4659
        }
4660 4661 4662 4663
    }
}

#endif
4664

4665 4666 4667
struct unlock_ftnode_extra {
    FT_HANDLE ft_handle;
    FTNODE node;
Yoni Fogel's avatar
Yoni Fogel committed
4668
    bool msgs_applied;
4669 4670 4671
};
// When this is called, the cachetable lock is held
static void
4672
unlock_ftnode_fun (PAIR p, void *v) {
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4673 4674
    struct unlock_ftnode_extra *x = NULL;
    CAST_FROM_VOIDP(x, v);
4675 4676
    FT_HANDLE brt = x->ft_handle;
    FTNODE node = x->node;
4677
    // CT lock is held
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4678
    int r = toku_cachetable_unpin_ct_prelocked_no_flush(
4679
        p,
4680
        brt->ft->cf,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4681
        node->ct_pair,
4682
        (enum cachetable_dirty) node->dirty,
4683
        x->msgs_applied ? make_ftnode_pair_attr(node) : make_invalid_pair_attr()
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4684
        );
4685
    assert_zero(r);
4686 4687
}

4688
/* search in a node's child */
4689
static int
Yoni Fogel's avatar
Yoni Fogel committed
4690 4691
ft_search_child(FT_HANDLE brt, FTNODE node, int childnum, ft_search_t *search, FT_GET_CALLBACK_FUNCTION getf, void *getf_v, bool *doprefetch, FT_CURSOR ftcursor, UNLOCKERS unlockers,
                 ANCESTORS ancestors, struct pivot_bounds const * const bounds, bool can_bulk_fetch)
4692
// Effect: Search in a node's child.  Searches are read-only now (at least as far as the hardcopy is concerned).
4693
{
4694
    struct ancestors next_ancestors = {node, childnum, ancestors};
4695

4696
    BLOCKNUM childblocknum = BP_BLOCKNUM(node,childnum);
Yoni Fogel's avatar
Yoni Fogel committed
4697
    uint32_t fullhash = compute_child_fullhash(brt->ft->cf, node, childnum);
4698
    FTNODE childnode;
4699

4700
    struct ftnode_fetch_extra bfe;
4701
    fill_bfe_for_subset_read(
4702
        &bfe,
4703
        brt->ft,
4704
        search,
4705 4706 4707 4708 4709
        &ftcursor->range_lock_left_key,
        &ftcursor->range_lock_right_key,
        ftcursor->left_is_neg_infty,
        ftcursor->right_is_pos_infty,
        ftcursor->disable_prefetching
4710
        );
Yoni Fogel's avatar
Yoni Fogel committed
4711
    bool msgs_applied = false;
4712
    {
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4713 4714 4715 4716
        int rr = toku_pin_ftnode_batched(brt, childblocknum, fullhash,
                                         unlockers,
                                         &next_ancestors, bounds,
                                         &bfe,
Leif Walsh's avatar
Leif Walsh committed
4717
                                         PL_READ, // we try to get a read lock, but we may upgrade to a write lock on a leaf for message application.
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4718 4719 4720 4721 4722 4723 4724 4725 4726
                                         true,
                                         &childnode,
                                         &msgs_applied);
        if (rr==TOKUDB_TRY_AGAIN) {
            return rr;
        }
        // We end the batch before applying ancestor messages if we get
        // all the way to a leaf.
        invariant_zero(rr);
4727 4728
    }

4729
    struct unlock_ftnode_extra unlock_extra   = {brt,childnode,msgs_applied};
Yoni Fogel's avatar
Yoni Fogel committed
4730
    struct unlockers next_unlockers = {true, unlock_ftnode_fun, (void*)&unlock_extra, unlockers};
4731

4732
    int r = ft_search_node(brt, childnode, search, bfe.child_to_read, getf, getf_v, doprefetch, ftcursor, &next_unlockers, &next_ancestors, bounds, can_bulk_fetch);
4733
    if (r!=TOKUDB_TRY_AGAIN) {
4734
#if TOKU_DO_PREFETCH
4735 4736
        // maybe prefetch the next child
        if (r == 0 && node->height == 1) {
4737
            ft_node_maybe_prefetch(brt, node, childnum, ftcursor, doprefetch);
4738
        }
4739 4740
#endif

4741
        assert(next_unlockers.locked);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4742
        if (msgs_applied) {
4743
            toku_unpin_ftnode(brt->ft, childnode);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4744 4745
        }
        else {
4746
            toku_unpin_ftnode_read_only(brt->ft, childnode);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4747
        }
4748
    } else {
4749
        // try again.
4750 4751

        // there are two cases where we get TOKUDB_TRY_AGAIN
4752
        //  case 1 is when some later call to toku_pin_ftnode returned
4753
        //  that value and unpinned all the nodes anyway. case 2
4754
        //  is when ft_search_node had to stop its search because
4755 4756
        //  some piece of a node that it needed was not in memory. In this case,
        //  the node was not unpinned, so we unpin it here
4757
        if (next_unlockers.locked) {
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4758
            if (msgs_applied) {
4759
                toku_unpin_ftnode(brt->ft, childnode);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4760 4761
            }
            else {
4762
                toku_unpin_ftnode_read_only(brt->ft, childnode);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4763
            }
4764
        }
4765
    }
4766

4767 4768 4769
    return r;
}

4770
static inline int
4771
search_which_child_cmp_with_bound(DB *db, ft_compare_func cmp, FTNODE node, int childnum, ft_search_t *search, DBT *dbt)
4772
{
4773
    return cmp(db, toku_copy_dbt(dbt, node->childkeys[childnum]), &search->pivot_bound);
4774 4775
}

4776
int
4777
toku_ft_search_which_child(
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4778
    DESCRIPTOR desc,
4779 4780 4781
    ft_compare_func cmp,
    FTNODE node,
    ft_search_t *search
4782
    )
4783
{
4784 4785 4786 4787 4788 4789 4790 4791 4792 4793 4794
#define DO_SEARCH_WHICH_CHILD_BINARY 1
#if DO_SEARCH_WHICH_CHILD_BINARY
    if (node->n_children <= 1) return 0;

    DBT pivotkey;
    toku_init_dbt(&pivotkey);
    int lo = 0;
    int hi = node->n_children - 1;
    int mi;
    while (lo < hi) {
        mi = (lo + hi) / 2;
4795
        toku_copy_dbt(&pivotkey, node->childkeys[mi]);
4796 4797 4798 4799 4800 4801 4802 4803 4804 4805 4806 4807
        // search->compare is really strange, and only works well with a
        // linear search, it makes binary search a pita.
        //
        // if you are searching left to right, it returns
        //   "0" for pivots that are < the target, and
        //   "1" for pivots that are >= the target
        // if you are searching right to left, it's the opposite.
        //
        // so if we're searching from the left and search->compare says
        // "1", we want to go left from here, if it says "0" we want to go
        // right.  searching from the right does the opposite.
        bool c = search->compare(search, &pivotkey);
4808 4809
        if (((search->direction == FT_SEARCH_LEFT) && c) ||
            ((search->direction == FT_SEARCH_RIGHT) && !c)) {
4810 4811
            hi = mi;
        } else {
4812 4813
            assert(((search->direction == FT_SEARCH_LEFT) && !c) ||
                   ((search->direction == FT_SEARCH_RIGHT) && c));
4814 4815 4816 4817 4818 4819
            lo = mi + 1;
        }
    }
    // ready to return something, if the pivot is bounded, we have to move
    // over a bit to get away from what we've already searched
    if (search->have_pivot_bound) {
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4820
        FAKE_DB(db, desc);
4821
        if (search->direction == FT_SEARCH_LEFT) {
4822 4823 4824 4825 4826 4827 4828 4829 4830 4831 4832 4833 4834 4835 4836 4837 4838 4839 4840 4841 4842 4843
            while (lo < node->n_children - 1 &&
                   search_which_child_cmp_with_bound(&db, cmp, node, lo, search, &pivotkey) <= 0) {
                // searching left to right, if the comparison says the
                // current pivot (lo) is left of or equal to our bound,
                // don't search that child again
                lo++;
            }
        } else {
            while (lo > 0 &&
                   search_which_child_cmp_with_bound(&db, cmp, node, lo - 1, search, &pivotkey) >= 0) {
                // searching right to left, same argument as just above
                // (but we had to pass lo - 1 because the pivot between lo
                // and the thing just less than it is at that position in
                // the childkeys array)
                lo--;
            }
        }
    }
    return lo;
#endif
#define DO_SEARCH_WHICH_CHILD_LINEAR 0
#if DO_SEARCH_WHICH_CHILD_LINEAR
4844
    int c;
4845 4846
    DBT pivotkey;
    toku_init_dbt(&pivotkey);
4847

4848 4849
    /* binary search is overkill for a small array */
    int child[node->n_children];
4850

4851 4852
    /* scan left to right or right to left depending on the search direction */
    for (c = 0; c < node->n_children; c++) {
4853
        child[c] = (search->direction == FT_SEARCH_LEFT) ? c : node->n_children - 1 - c;
4854 4855
    }
    for (c = 0; c < node->n_children-1; c++) {
4856
        int p = (search->direction == FT_SEARCH_LEFT) ? child[c] : child[c] - 1;
4857
        toku_copy_dbt(&pivotkey, node->childkeys[p]);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4858
        if (search_pivot_is_bounded(search, desc, cmp, &pivotkey) && search->compare(search, &pivotkey)) {
4859 4860 4861 4862 4863
            return child[c];
        }
    }
    /* check the first (left) or last (right) node if nothing has been found */
    return child[c];
4864
#endif
4865
}
4866

4867 4868
static void
maybe_search_save_bound(
4869
    FTNODE node,
4870
    int child_searched,
4871
    ft_search_t *search)
4872
{
4873
    int p = (search->direction == FT_SEARCH_LEFT) ? child_searched : child_searched - 1;
4874
    if (p >= 0 && p < node->n_children-1) {
4875
        search_save_bound(search, &node->childkeys[p]);
4876 4877 4878 4879
    }
}

static int
4880 4881 4882 4883
ft_search_node(
    FT_HANDLE brt,
    FTNODE node,
    ft_search_t *search,
4884
    int child_to_search,
4885
    FT_GET_CALLBACK_FUNCTION getf,
4886
    void *getf_v,
Yoni Fogel's avatar
Yoni Fogel committed
4887
    bool *doprefetch,
4888
    FT_CURSOR ftcursor,
4889
    UNLOCKERS unlockers,
4890
    ANCESTORS ancestors,
4891
    struct pivot_bounds const * const bounds,
Yoni Fogel's avatar
Yoni Fogel committed
4892
    bool can_bulk_fetch
4893
    )
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4894 4895
{
    int r = 0;
4896
    // assert that we got a valid child_to_search
4897 4898
    invariant(child_to_search >= 0);
    invariant(child_to_search < node->n_children);
4899 4900 4901 4902 4903
    //
    // At this point, we must have the necessary partition available to continue the search
    //
    assert(BP_STATE(node,child_to_search) == PT_AVAIL);
    while (child_to_search >= 0 && child_to_search < node->n_children) {
4904 4905
        //
        // Normally, the child we want to use is available, as we checked
4906
        // before entering this while loop. However, if we pass through
4907
        // the loop once, getting DB_NOTFOUND for this first value
4908
        // of child_to_search, we enter the while loop again with a
4909 4910
        // child_to_search that may not be in memory. If it is not,
        // we need to return TOKUDB_TRY_AGAIN so the query can
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4911
        // read the appropriate partition into memory
4912 4913 4914 4915
        //
        if (BP_STATE(node,child_to_search) != PT_AVAIL) {
            return TOKUDB_TRY_AGAIN;
        }
4916 4917
        const struct pivot_bounds next_bounds = next_pivot_keys(node, child_to_search, bounds);
        if (node->height > 0) {
4918
            r = ft_search_child(
4919 4920 4921 4922 4923 4924 4925
                brt,
                node,
                child_to_search,
                search,
                getf,
                getf_v,
                doprefetch,
4926
                ftcursor,
4927 4928
                unlockers,
                ancestors,
4929 4930
                &next_bounds,
                can_bulk_fetch
4931 4932 4933
                );
        }
        else {
4934
            r = ft_search_basement_node(
4935
                BLB(node, child_to_search),
4936 4937 4938 4939
                search,
                getf,
                getf_v,
                doprefetch,
4940
                ftcursor,
4941
                can_bulk_fetch
4942 4943 4944
                );
        }
        if (r == 0) return r; //Success
4945

4946 4947
        if (r != DB_NOTFOUND) {
            return r; //Error (or message to quit early, such as TOKUDB_FOUND_BUT_REJECTED or TOKUDB_TRY_AGAIN)
4948
        }
4949 4950 4951 4952
        // not really necessary, just put this here so that reading the
        // code becomes simpler. The point is at this point in the code,
        // we know that we got DB_NOTFOUND and we have to continue
        assert(r == DB_NOTFOUND);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4953 4954 4955 4956 4957 4958 4959 4960 4961 4962 4963 4964 4965 4966 4967 4968 4969 4970 4971 4972 4973 4974 4975 4976
        // we have a new pivotkey
        if (node->height == 0) {
            // when we run off the end of a basement, try to lock the range up to the pivot. solves #3529
            const DBT *pivot = NULL;
            if (search->direction == FT_SEARCH_LEFT)
                pivot = next_bounds.upper_bound_inclusive; // left -> right
            else
                pivot = next_bounds.lower_bound_exclusive; // right -> left
            if (pivot) {
                int rr = getf(pivot->size, pivot->data, 0, NULL, getf_v, true);
                if (rr != 0)
                    return rr; // lock was not granted
            }
        }

        // If we got a DB_NOTFOUND then we have to search the next record.        Possibly everything present is not visible.
        // This way of doing DB_NOTFOUND is a kludge, and ought to be simplified.  Something like this is needed for DB_NEXT, but
        //        for point queries, it's overkill.  If we got a DB_NOTFOUND on a point query then we should just stop looking.
        // When releasing locks on I/O we must not search the same subtree again, or we won't be guaranteed to make forward progress.
        // If we got a DB_NOTFOUND, then the pivot is too small if searching from left to right (too large if searching from right to left).
        // So save the pivot key in the search object.
        maybe_search_save_bound(node, child_to_search, search);

        // We're about to pin some more nodes, but we thought we were done before.
4977
        if (search->direction == FT_SEARCH_LEFT) {
4978 4979 4980 4981 4982
            child_to_search++;
        }
        else {
            child_to_search--;
        }
4983
    }
4984
    return r;
4985 4986 4987
}

static int
Yoni Fogel's avatar
Yoni Fogel committed
4988
toku_ft_search (FT_HANDLE brt, ft_search_t *search, FT_GET_CALLBACK_FUNCTION getf, void *getf_v, FT_CURSOR ftcursor, bool can_bulk_fetch)
4989 4990 4991
// Effect: Perform a search.  Associate cursor with a leaf if possible.
// All searches are performed through this function.
{
4992
    int r;
4993
    uint trycount = 0;     // How many tries did it take to get the result?
Zardosht Kasheff's avatar
Zardosht Kasheff committed
4994
    FT ft = brt->ft;
4995

Zardosht Kasheff's avatar
Zardosht Kasheff committed
4996
try_again:
4997

4998
    trycount++;
4999

Zardosht Kasheff's avatar
Zardosht Kasheff committed
5000 5001 5002 5003 5004 5005 5006 5007 5008 5009
    //
    // Here is how searches work
    // At a high level, we descend down the tree, using the search parameter
    // to guide us towards where to look. But the search parameter is not
    // used here to determine which child of a node to read (regardless
    // of whether that child is another node or a basement node)
    // The search parameter is used while we are pinning the node into
    // memory, because that is when the system needs to ensure that
    // the appropriate partition of the child we are using is in memory.
    // So, here are the steps for a search (and this applies to this function
5010 5011 5012
    // as well as ft_search_child:
    //  - Take the search parameter, and create a ftnode_fetch_extra, that will be used by toku_pin_ftnode(_holding_lock)
    //  - Call toku_pin_ftnode(_holding_lock) with the bfe as the extra for the fetch callback (in case the node is not at all in memory)
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5013
    //       and the partial fetch callback (in case the node is perhaps partially in memory) to the fetch the node
5014
    //  - This eventually calls either toku_ftnode_fetch_callback or  toku_ftnode_pf_req_callback depending on whether the node is in
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5015
    //     memory at all or not.
5016 5017 5018
    //  - Within these functions, the "ft_search_t search" parameter is used to evaluate which child the search is interested in.
    //     If the node is not in memory at all, toku_ftnode_fetch_callback will read the node and decompress only the partition for the
    //     relevant child, be it a message buffer or basement node. If the node is in memory, then toku_ftnode_pf_req_callback
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5019
    //     will tell the cachetable that a partial fetch is required if and only if the relevant child is not in memory. If the relevant child
5020
    //     is not in memory, then toku_ftnode_pf_callback is called to fetch the partition.
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5021
    //  - These functions set bfe->child_to_read so that the search code does not need to reevaluate it.
5022 5023 5024
    //  - Just to reiterate, all of the last item happens within toku_ftnode_pin(_holding_lock)
    //  - At this point, toku_ftnode_pin_holding_lock has returned, with bfe.child_to_read set,
    //  - ft_search_node is called, assuming that the node and its relevant partition are in memory.
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5025
    //
5026
    struct ftnode_fetch_extra bfe;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5027 5028
    fill_bfe_for_subset_read(
        &bfe,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5029
        ft,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5030
        search,
5031 5032 5033 5034 5035
        &ftcursor->range_lock_left_key,
        &ftcursor->range_lock_right_key,
        ftcursor->left_is_neg_infty,
        ftcursor->right_is_pos_infty,
        ftcursor->disable_prefetching
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5036
        );
5037
    FTNODE node = NULL;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5038
    {
Yoni Fogel's avatar
Yoni Fogel committed
5039
        uint32_t fullhash;
5040
        CACHEKEY root_key;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5041 5042 5043
        toku_calculate_root_offset_pointer(ft, &root_key, &fullhash);
        toku_pin_ftnode_off_client_thread_batched(
            ft,
5044
            root_key,
5045
            fullhash,
5046
            &bfe,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5047
            PL_READ, // may_modify_node set to false, because root cannot change during search
5048 5049 5050 5051
            0,
            NULL,
            &node
            );
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5052
    }
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5053

5054
    uint tree_height = node->height + 1;  // How high is the tree?  This is the height of the root node plus one (leaf is at height 0).
5055

Zardosht Kasheff's avatar
Zardosht Kasheff committed
5056

Yoni Fogel's avatar
Yoni Fogel committed
5057 5058
    struct unlock_ftnode_extra unlock_extra   = {brt,node,false};
    struct unlockers                unlockers      = {true, unlock_ftnode_fun, (void*)&unlock_extra, (UNLOCKERS)NULL};
5059

5060
    {
Yoni Fogel's avatar
Yoni Fogel committed
5061
        bool doprefetch = false;
5062
        //static int counter = 0;         counter++;
5063
        r = ft_search_node(brt, node, search, bfe.child_to_read, getf, getf_v, &doprefetch, ftcursor, &unlockers, (ANCESTORS)NULL, &infinite_bounds, can_bulk_fetch);
5064
        if (r==TOKUDB_TRY_AGAIN) {
5065
            // there are two cases where we get TOKUDB_TRY_AGAIN
5066
            //  case 1 is when some later call to toku_pin_ftnode returned
5067
            //  that value and unpinned all the nodes anyway. case 2
5068
            //  is when ft_search_node had to stop its search because
5069
            //  some piece of a node that it needed was not in memory.
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5070
            //  In this case, the node was not unpinned, so we unpin it here
5071
            if (unlockers.locked) {
5072
                toku_unpin_ftnode_read_only(brt->ft, node);
5073
            }
5074 5075 5076 5077
            goto try_again;
        } else {
            assert(unlockers.locked);
        }
5078 5079
    }

5080
    assert(unlockers.locked);
5081
    toku_unpin_ftnode_read_only(brt->ft, node);
5082

5083

5084 5085
    //Heaviside function (+direction) queries define only a lower or upper
    //bound.  Some queries require both an upper and lower bound.
5086
    //They do this by wrapping the FT_GET_CALLBACK_FUNCTION with another
5087 5088 5089 5090 5091 5092
    //test that checks for the other bound.  If the other bound fails,
    //it returns TOKUDB_FOUND_BUT_REJECTED which means not found, but
    //stop searching immediately, as opposed to DB_NOTFOUND
    //which can mean not found, but keep looking in another leaf.
    if (r==TOKUDB_FOUND_BUT_REJECTED) r = DB_NOTFOUND;
    else if (r==DB_NOTFOUND) {
5093
        //We truly did not find an answer to the query.
5094
        //Therefore, the FT_GET_CALLBACK_FUNCTION has NOT been called.
5095 5096 5097 5098 5099 5100 5101
        //The contract specifies that the callback function must be called
        //for 'r= (0|DB_NOTFOUND|TOKUDB_FOUND_BUT_REJECTED)'
        //TODO: #1378 This is not the ultimate location of this call to the
        //callback.  It is surely wrong for node-level locking, and probably
        //wrong for the STRADDLE callback for heaviside function(two sets of key/vals)
        int r2 = getf(0,NULL, 0,NULL, getf_v, false);
        if (r2!=0) r = r2;
5102
    }
5103
    {   // accounting (to detect and measure thrashing)
5104
        uint retrycount = trycount - 1;         // how many retries were needed?
5105 5106 5107
        if (retrycount) {
            STATUS_INC(FT_TOTAL_RETRIES, retrycount);
        }
5108
        if (retrycount > tree_height) {         // if at least one node was read from disk more than once
5109 5110 5111 5112
            STATUS_INC(FT_SEARCH_TRIES_GT_HEIGHT, 1);
            //uint excess_tries = retrycount - tree_height;
            //if (excess_tries > STATUS_VALUE(FT_MAX_SEARCH_EXCESS_RETRIES))
            //    STATUS_VALUE(FT_MAX_SEARCH_EXCESS_RETRIES) = excess_tries;
5113
            if (retrycount > (tree_height+3))
5114
                STATUS_INC(FT_SEARCH_TRIES_GT_HEIGHTPLUS3, 1);
5115
        }
5116
    }
5117 5118 5119
    return r;
}

5120 5121
struct ft_cursor_search_struct {
    FT_GET_CALLBACK_FUNCTION getf;
5122
    void *getf_v;
5123 5124
    FT_CURSOR cursor;
    ft_search_t *search;
5125
};
5126

5127 5128
/* search for the first kv pair that matches the search object */
static int
Yoni Fogel's avatar
Yoni Fogel committed
5129
ft_cursor_search(FT_CURSOR cursor, ft_search_t *search, FT_GET_CALLBACK_FUNCTION getf, void *getf_v, bool can_bulk_fetch)
5130
{
5131
    int r = toku_ft_search(cursor->ft_handle, search, getf, getf_v, cursor, can_bulk_fetch);
5132 5133 5134
    return r;
}

5135
static inline int compare_k_x(FT_HANDLE brt, const DBT *k, const DBT *x) {
5136 5137
    FAKE_DB(db, &brt->ft->cmp_descriptor);
    return brt->ft->compare_fun(&db, k, x);
5138 5139
}

5140
static int
5141
ft_cursor_compare_one(ft_search_t *search __attribute__((__unused__)), DBT *x __attribute__((__unused__)))
5142
{
5143 5144 5145
    return 1;
}

5146
static int ft_cursor_compare_set(ft_search_t *search, DBT *x) {
5147
    FT_HANDLE CAST_FROM_VOIDP(brt, search->context);
5148
    return compare_k_x(brt, search->k, x) <= 0; /* return min xy: kv <= xy */
5149 5150
}

5151
static int
5152
ft_cursor_current_getf(ITEMLEN keylen,                 bytevec key,
5153 5154
                        ITEMLEN vallen,                 bytevec val,
                        void *v, bool lock_only) {
5155
    struct ft_cursor_search_struct *CAST_FROM_VOIDP(bcss, v);
5156 5157
    int r;
    if (key==NULL) {
5158
        r = bcss->getf(0, NULL, 0, NULL, bcss->getf_v, lock_only);
5159
    } else {
5160
        FT_CURSOR cursor = bcss->cursor;
5161 5162
        DBT newkey;
        toku_fill_dbt(&newkey, key, keylen);
5163
        if (compare_k_x(cursor->ft_handle, &cursor->key, &newkey) != 0) {
5164 5165 5166 5167 5168
            r = bcss->getf(0, NULL, 0, NULL, bcss->getf_v, lock_only); // This was once DB_KEYEMPTY
            if (r==0) r = TOKUDB_FOUND_BUT_REJECTED;
        }
        else
            r = bcss->getf(keylen, key, vallen, val, bcss->getf_v, lock_only);
5169 5170 5171 5172 5173
    }
    return r;
}

int
5174
toku_ft_cursor_current(FT_CURSOR cursor, int op, FT_GET_CALLBACK_FUNCTION getf, void *getf_v)
5175
{
5176
    if (ft_cursor_not_set(cursor))
5177
        return EINVAL;
5178
    if (op == DB_CURRENT) {
5179 5180
        struct ft_cursor_search_struct bcss = {getf, getf_v, cursor, 0};
        ft_search_t search; ft_search_init(&search, ft_cursor_compare_set, FT_SEARCH_LEFT, &cursor->key, cursor->ft_handle);
Yoni Fogel's avatar
Yoni Fogel committed
5181
        int r = toku_ft_search(cursor->ft_handle, &search, ft_cursor_current_getf, &bcss, cursor, false);
5182
        ft_search_finish(&search);
5183
        return r;
5184
    }
5185
    return getf(cursor->key.size, cursor->key.data, cursor->val.size, cursor->val.data, getf_v, false); // ft_cursor_copyout(cursor, outkey, outval);
5186 5187
}

5188
int
5189
toku_ft_cursor_first(FT_CURSOR cursor, FT_GET_CALLBACK_FUNCTION getf, void *getf_v)
5190
{
5191
    ft_search_t search; ft_search_init(&search, ft_cursor_compare_one, FT_SEARCH_LEFT, 0, cursor->ft_handle);
Yoni Fogel's avatar
Yoni Fogel committed
5192
    int r = ft_cursor_search(cursor, &search, getf, getf_v, false);
5193
    ft_search_finish(&search);
5194
    return r;
5195 5196
}

5197
int
5198
toku_ft_cursor_last(FT_CURSOR cursor, FT_GET_CALLBACK_FUNCTION getf, void *getf_v)
5199
{
5200
    ft_search_t search; ft_search_init(&search, ft_cursor_compare_one, FT_SEARCH_RIGHT, 0, cursor->ft_handle);
Yoni Fogel's avatar
Yoni Fogel committed
5201
    int r = ft_cursor_search(cursor, &search, getf, getf_v, false);
5202
    ft_search_finish(&search);
5203
    return r;
5204 5205
}

5206
static int ft_cursor_compare_next(ft_search_t *search, DBT *x) {
5207
    FT_HANDLE CAST_FROM_VOIDP(brt, search->context);
5208
    return compare_k_x(brt, search->k, x) < 0; /* return min xy: kv < xy */
5209 5210
}

5211

5212
static int
5213 5214
ft_cursor_shortcut (
    FT_CURSOR cursor,
5215
    int direction,
5216
    FT_GET_CALLBACK_FUNCTION getf,
5217
    void *getf_v,
Yoni Fogel's avatar
Yoni Fogel committed
5218
    uint32_t *keylen,
5219
    void **key,
Yoni Fogel's avatar
Yoni Fogel committed
5220
    uint32_t *vallen,
5221
    void **val
5222
    )
5223
{
5224
    int r = 0;
Yoni Fogel's avatar
Yoni Fogel committed
5225
    uint32_t index = cursor->leaf_info.to_be.index;
5226 5227 5228
    OMT omt = cursor->leaf_info.to_be.omt;
    // if we are searching towards the end, limit is last element
    // if we are searching towards the beginning, limit is the first element
Yoni Fogel's avatar
Yoni Fogel committed
5229
    uint32_t limit = (direction > 0) ? (toku_omt_size(omt) - 1) : 0;
5230 5231

    //Starting with the prev, find the first real (non-provdel) leafentry.
5232
    OMTVALUE lev = NULL;
5233 5234
    while (index != limit) {
        index += direction;
5235
        r = toku_omt_fetch(omt, index, &lev);
5236
        assert_zero(r);
5237
        LEAFENTRY CAST_FROM_VOIDP(le, lev);
5238

5239
        if (toku_ft_cursor_is_leaf_mode(cursor) || !is_le_val_del(le, cursor)) {
5240

5241
            ft_cursor_extract_key_and_val(
5242 5243 5244 5245 5246 5247 5248
                le,
                cursor,
                keylen,
                key,
                vallen,
                val
                );
5249

5250 5251
            r = getf(*keylen, *key, *vallen, *val, getf_v, false);
            if (r == 0 || r == TOKUDB_CURSOR_CONTINUE) {
5252 5253 5254
                //Update cursor.
                cursor->leaf_info.to_be.index = index;
            }
5255
            if (r == TOKUDB_CURSOR_CONTINUE) {
5256 5257 5258 5259 5260
                continue;
            }
            else {
                break;
            }
5261
        }
5262
    }
5263

5264
    return r;
5265 5266 5267
}

int
5268
toku_ft_cursor_next(FT_CURSOR cursor, FT_GET_CALLBACK_FUNCTION getf, void *getf_v)
5269
{
5270
    ft_search_t search; ft_search_init(&search, ft_cursor_compare_next, FT_SEARCH_LEFT, &cursor->key, cursor->ft_handle);
Yoni Fogel's avatar
Yoni Fogel committed
5271
    int r = ft_cursor_search(cursor, &search, getf, getf_v, true);
5272 5273
    ft_search_finish(&search);
    if (r == 0) ft_cursor_set_prefetching(cursor);
5274
    return r;
5275 5276
}

5277
static int
5278
ft_cursor_search_eq_k_x_getf(ITEMLEN keylen,               bytevec key,
5279 5280
                              ITEMLEN vallen,               bytevec val,
                              void *v, bool lock_only) {
5281
    struct ft_cursor_search_struct *CAST_FROM_VOIDP(bcss, v);
5282 5283
    int r;
    if (key==NULL) {
5284
        r = bcss->getf(0, NULL, 0, NULL, bcss->getf_v, false);
5285
    } else {
5286
        FT_CURSOR cursor = bcss->cursor;
5287 5288
        DBT newkey;
        toku_fill_dbt(&newkey, key, keylen);
5289
        if (compare_k_x(cursor->ft_handle, bcss->search->k, &newkey) == 0) {
5290 5291 5292 5293 5294
            r = bcss->getf(keylen, key, vallen, val, bcss->getf_v, lock_only);
        } else {
            r = bcss->getf(0, NULL, 0, NULL, bcss->getf_v, lock_only);
            if (r==0) r = TOKUDB_FOUND_BUT_REJECTED;
        }
5295 5296 5297
    }
    return r;
}
5298 5299 5300

/* search for the kv pair that matches the search object and is equal to k */
static int
5301
ft_cursor_search_eq_k_x(FT_CURSOR cursor, ft_search_t *search, FT_GET_CALLBACK_FUNCTION getf, void *getf_v)
5302
{
5303
    struct ft_cursor_search_struct bcss = {getf, getf_v, cursor, search};
Yoni Fogel's avatar
Yoni Fogel committed
5304
    int r = toku_ft_search(cursor->ft_handle, search, ft_cursor_search_eq_k_x_getf, &bcss, cursor, false);
5305
    return r;
5306 5307
}

5308
static int ft_cursor_compare_prev(ft_search_t *search, DBT *x) {
5309
    FT_HANDLE CAST_FROM_VOIDP(brt, search->context);
5310
    return compare_k_x(brt, search->k, x) > 0; /* return max xy: kv > xy */
5311 5312
}

5313
int
5314
toku_ft_cursor_prev(FT_CURSOR cursor, FT_GET_CALLBACK_FUNCTION getf, void *getf_v)
5315
{
5316
    ft_search_t search; ft_search_init(&search, ft_cursor_compare_prev, FT_SEARCH_RIGHT, &cursor->key, cursor->ft_handle);
Yoni Fogel's avatar
Yoni Fogel committed
5317
    int r = ft_cursor_search(cursor, &search, getf, getf_v, true);
5318
    ft_search_finish(&search);
5319
    return r;
5320 5321
}

5322
static int ft_cursor_compare_set_range(ft_search_t *search, DBT *x) {
5323
    FT_HANDLE CAST_FROM_VOIDP(brt, search->context);
5324
    return compare_k_x(brt, search->k,        x) <= 0; /* return kv <= xy */
5325 5326 5327
}

int
5328
toku_ft_cursor_set(FT_CURSOR cursor, DBT *key, FT_GET_CALLBACK_FUNCTION getf, void *getf_v)
5329
{
5330 5331 5332
    ft_search_t search; ft_search_init(&search, ft_cursor_compare_set_range, FT_SEARCH_LEFT, key, cursor->ft_handle);
    int r = ft_cursor_search_eq_k_x(cursor, &search, getf, getf_v);
    ft_search_finish(&search);
5333
    return r;
5334 5335 5336
}

int
5337
toku_ft_cursor_set_range(FT_CURSOR cursor, DBT *key, FT_GET_CALLBACK_FUNCTION getf, void *getf_v)
5338
{
5339
    ft_search_t search; ft_search_init(&search, ft_cursor_compare_set_range, FT_SEARCH_LEFT, key, cursor->ft_handle);
Yoni Fogel's avatar
Yoni Fogel committed
5340
    int r = ft_cursor_search(cursor, &search, getf, getf_v, false);
5341
    ft_search_finish(&search);
5342
    return r;
5343 5344
}

5345
static int ft_cursor_compare_set_range_reverse(ft_search_t *search, DBT *x) {
5346
    FT_HANDLE CAST_FROM_VOIDP(brt, search->context);
5347
    return compare_k_x(brt, search->k, x) >= 0; /* return kv >= xy */
5348 5349 5350
}

int
5351
toku_ft_cursor_set_range_reverse(FT_CURSOR cursor, DBT *key, FT_GET_CALLBACK_FUNCTION getf, void *getf_v)
5352
{
5353
    ft_search_t search; ft_search_init(&search, ft_cursor_compare_set_range_reverse, FT_SEARCH_RIGHT, key, cursor->ft_handle);
Yoni Fogel's avatar
Yoni Fogel committed
5354
    int r = ft_cursor_search(cursor, &search, getf, getf_v, false);
5355
    ft_search_finish(&search);
5356
    return r;
5357 5358 5359
}


5360 5361 5362
//TODO: When tests have been rewritten, get rid of this function.
//Only used by tests.
int
5363
toku_ft_cursor_get (FT_CURSOR cursor, DBT *key, FT_GET_CALLBACK_FUNCTION getf, void *getf_v, int get_flags)
5364
{
5365
    int op = get_flags & DB_OPFLAGS_MASK;
5366
    if (get_flags & ~DB_OPFLAGS_MASK)
5367
        return EINVAL;
5368 5369 5370 5371

    switch (op) {
    case DB_CURRENT:
    case DB_CURRENT_BINDING:
5372
        return toku_ft_cursor_current(cursor, op, getf, getf_v);
5373
    case DB_FIRST:
5374
        return toku_ft_cursor_first(cursor, getf, getf_v);
5375
    case DB_LAST:
5376
        return toku_ft_cursor_last(cursor, getf, getf_v);
5377
    case DB_NEXT:
John Esmet's avatar
John Esmet committed
5378
        if (ft_cursor_not_set(cursor)) {
5379
            return toku_ft_cursor_first(cursor, getf, getf_v);
John Esmet's avatar
John Esmet committed
5380
        } else {
5381
            return toku_ft_cursor_next(cursor, getf, getf_v);
John Esmet's avatar
John Esmet committed
5382
        }
5383
    case DB_PREV:
John Esmet's avatar
John Esmet committed
5384
        if (ft_cursor_not_set(cursor)) {
5385
            return toku_ft_cursor_last(cursor, getf, getf_v);
John Esmet's avatar
John Esmet committed
5386
        } else {
5387
            return toku_ft_cursor_prev(cursor, getf, getf_v);
John Esmet's avatar
John Esmet committed
5388
        }
5389
    case DB_SET:
5390
        return toku_ft_cursor_set(cursor, key, getf, getf_v);
5391
    case DB_SET_RANGE:
5392
        return toku_ft_cursor_set_range(cursor, key, getf, getf_v);
5393
    default: ;// Fall through
5394
    }
5395
    return EINVAL;
5396 5397
}

5398
void
5399
toku_ft_cursor_peek(FT_CURSOR cursor, const DBT **pkey, const DBT **pval)
5400 5401
// Effect: Retrieves a pointer to the DBTs for the current key and value.
// Requires:  The caller may not modify the DBTs or the memory at which they points.
5402
// Requires:  The caller must be in the context of a
5403
// FT_GET_(STRADDLE_)CALLBACK_FUNCTION
5404 5405 5406 5407 5408 5409
{
    *pkey = &cursor->key;
    *pval = &cursor->val;
}

//We pass in toku_dbt_fake to the search functions, since it will not pass the
5410
//key(or val) to the heaviside function if key(or val) is NULL.
5411 5412 5413
//It is not used for anything else,
//the actual 'extra' information for the heaviside function is inside the
//wrapper.
5414
static const DBT __toku_dbt_fake = {};
5415 5416
static const DBT* const toku_dbt_fake = &__toku_dbt_fake;

Yoni Fogel's avatar
Yoni Fogel committed
5417
bool toku_ft_cursor_uninitialized(FT_CURSOR c) {
5418
    return ft_cursor_not_set(c);
5419 5420 5421
}


5422 5423
/* ********************************* lookup **************************************/

5424
int
5425
toku_ft_lookup (FT_HANDLE brt, DBT *k, FT_GET_CALLBACK_FUNCTION getf, void *getf_v)
5426
{
5427
    int r, rr;
5428
    FT_CURSOR cursor;
5429

Yoni Fogel's avatar
Yoni Fogel committed
5430
    rr = toku_ft_cursor(brt, &cursor, NULL, false, false);
5431 5432
    if (rr != 0) return rr;

5433
    int op = DB_SET;
5434
    r = toku_ft_cursor_get(cursor, k, getf, getf_v, op);
5435

5436
    toku_ft_cursor_close(cursor);
5437 5438 5439 5440

    return r;
}

5441
/* ********************************* delete **************************************/
5442
static int
5443
getf_nothing (ITEMLEN UU(keylen), bytevec UU(key), ITEMLEN UU(vallen), bytevec UU(val), void *UU(pair_v), bool UU(lock_only)) {
5444 5445 5446 5447
    return 0;
}

int
5448
toku_ft_cursor_delete(FT_CURSOR cursor, int flags, TOKUTXN txn) {
5449 5450 5451
    int r;

    int unchecked_flags = flags;
Yoni Fogel's avatar
Yoni Fogel committed
5452
    bool error_if_missing = (bool) !(flags&DB_DELETE_ANY);
5453 5454
    unchecked_flags &= ~DB_DELETE_ANY;
    if (unchecked_flags!=0) r = EINVAL;
5455
    else if (ft_cursor_not_set(cursor)) r = EINVAL;
5456
    else {
5457 5458
        r = 0;
        if (error_if_missing) {
5459
            r = toku_ft_cursor_current(cursor, DB_CURRENT, getf_nothing, NULL);
5460 5461
        }
        if (r == 0) {
5462
            toku_ft_delete(cursor->ft_handle, &cursor->key, txn);
5463
        }
5464 5465 5466 5467 5468 5469 5470
    }
    return r;
}

/* ********************* keyrange ************************ */


5471
struct keyrange_compare_s {
5472
    FT_HANDLE ft_handle;
5473 5474 5475
    DBT *key;
};

5476
static int
5477
keyrange_compare (OMTVALUE lev, void *extra) {
5478
    LEAFENTRY CAST_FROM_VOIDP(le, lev);
Yoni Fogel's avatar
Yoni Fogel committed
5479
    uint32_t keylen;
5480 5481 5482
    void* key = le_key_and_len(le, &keylen);
    DBT   omt_dbt;
    toku_fill_dbt(&omt_dbt, key, keylen);
5483
    struct keyrange_compare_s *CAST_FROM_VOIDP(s, extra);
5484
    // TODO: maybe put a const fake_db in the header
5485 5486
    FAKE_DB(db, &s->ft_handle->ft->cmp_descriptor);
    return s->ft_handle->ft->compare_fun(&db, &omt_dbt, s->key);
5487 5488
}

5489
static void
Yoni Fogel's avatar
Yoni Fogel committed
5490 5491
keyrange_in_leaf_partition (FT_HANDLE brt, FTNODE node, DBT *key, int child_number, uint64_t estimated_num_rows,
                            uint64_t *less, uint64_t *equal, uint64_t *greater)
5492
// If the partition is in main memory then estimate the number
5493
// If KEY==NULL then use an arbitrary key (leftmost or zero)
5494
{
5495
    paranoid_invariant(node->height == 0); // we are in a leaf
5496
    if (BP_STATE(node, child_number) == PT_AVAIL) {
5497 5498 5499 5500
        // If the partition is in main memory then get an exact count.
        struct keyrange_compare_s s = {brt,key};
        BASEMENTNODE bn = BLB(node, child_number);
        OMTVALUE datav;
Yoni Fogel's avatar
Yoni Fogel committed
5501
        uint32_t idx = 0;
5502 5503 5504 5505 5506 5507 5508 5509 5510 5511 5512 5513
        // if key is NULL then set r==-1 and idx==0.
        int r = key ? toku_omt_find_zero(bn->buffer, keyrange_compare, &s, &datav, &idx) : -1;
        if (r==0) {
            *less    = idx;
            *equal   = 1;
            *greater = toku_omt_size(bn->buffer)-idx-1;
        } else {
            // If not found, then the idx says where it's between.
            *less    = idx;
            *equal   = 0;
            *greater = toku_omt_size(bn->buffer)-idx;
        }
5514
    } else {
5515 5516 5517
        *less    = estimated_num_rows / 2;
        *equal   = 0;
        *greater = *less;
5518 5519 5520
    }
}

5521
static int
5522
toku_ft_keyrange_internal (FT_HANDLE brt, FTNODE node,
Yoni Fogel's avatar
Yoni Fogel committed
5523 5524
                            DBT *key, uint64_t *less, uint64_t *equal, uint64_t *greater,
                            uint64_t estimated_num_rows,
5525
                            struct ftnode_fetch_extra *bfe, // set up to read a minimal read.
5526
                            struct unlockers *unlockers, ANCESTORS ancestors, struct pivot_bounds const * const bounds)
5527 5528 5529
// Implementation note: Assign values to less, equal, and greater, and then on the way out (returning up the stack) we add more values in.
{
    int r = 0;
5530
    // if KEY is NULL then use the leftmost key.
5531
    int child_number = key ? toku_ftnode_which_child (node, key, &brt->ft->cmp_descriptor, brt->ft->compare_fun) : 0;
5532 5533 5534
    uint64_t rows_per_child = estimated_num_rows / node->n_children;
    if (node->height == 0) {

5535
        keyrange_in_leaf_partition(brt, node, key, child_number, rows_per_child, less, equal, greater);
5536 5537 5538

        *less    += rows_per_child * child_number;
        *greater += rows_per_child * (node->n_children - child_number - 1);
5539

5540
    } else {
5541 5542 5543
        // do the child.
        struct ancestors next_ancestors = {node, child_number, ancestors};
        BLOCKNUM childblocknum = BP_BLOCKNUM(node, child_number);
Yoni Fogel's avatar
Yoni Fogel committed
5544
        uint32_t fullhash = compute_child_fullhash(brt->ft->cf, node, child_number);
5545
        FTNODE childnode;
Yoni Fogel's avatar
Yoni Fogel committed
5546
        bool msgs_applied = false;
5547
        r = toku_pin_ftnode_batched(
5548 5549 5550 5551 5552 5553
            brt,
            childblocknum,
            fullhash,
            unlockers,
            &next_ancestors,
            bounds,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5554
            bfe,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5555
            PL_READ, // may_modify_node is false, because node guaranteed to not change
Yoni Fogel's avatar
Yoni Fogel committed
5556
            false,
5557
            &childnode,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5558 5559
            &msgs_applied
            );
5560
        paranoid_invariant(!msgs_applied);
5561
        if (r != TOKUDB_TRY_AGAIN) {
5562
            assert_zero(r);
5563

Yoni Fogel's avatar
Yoni Fogel committed
5564 5565
            struct unlock_ftnode_extra unlock_extra   = {brt,childnode,false};
            struct unlockers next_unlockers = {true, unlock_ftnode_fun, (void*)&unlock_extra, unlockers};
5566
            const struct pivot_bounds next_bounds = next_pivot_keys(node, child_number, bounds);
5567

5568
            r = toku_ft_keyrange_internal(brt, childnode, key, less, equal, greater, rows_per_child,
5569
                                           bfe, &next_unlockers, &next_ancestors, &next_bounds);
5570
            if (r != TOKUDB_TRY_AGAIN) {
5571
                assert_zero(r);
5572

5573 5574
                *less    += rows_per_child * child_number;
                *greater += rows_per_child * (node->n_children - child_number - 1);
5575

5576
                assert(unlockers->locked);
5577
                toku_unpin_ftnode_read_only(brt->ft, childnode);
5578 5579
            }
        }
5580
    }
5581
    return r;
5582 5583
}

5584
void toku_ft_keyrange(FT_HANDLE brt, DBT *key, uint64_t *less_p, uint64_t *equal_p, uint64_t *greater_p)
5585 5586
// Effect: Return an estimate  of the number of keys to the left, the number equal, and the number to the right of the key.
//   The values are an estimate.
5587
//   If you perform a keyrange on two keys that are in the same in-memory and uncompressed basement,
5588 5589 5590
//   you can use the keys_right numbers (or the keys_left) numbers to get an exact number keys in the range,
//   if the basement does not change between the keyrange queries.
//   TODO 4184: What to do with a NULL key?
5591
//   If KEY is NULL then the system picks an arbitrary key and returns it.
5592
{
5593
    struct ftnode_fetch_extra bfe;
5594
    fill_bfe_for_min_read(&bfe, brt->ft);  // read pivot keys but not message buffers
5595
try_again:
5596
    {
Yoni Fogel's avatar
Yoni Fogel committed
5597
        uint64_t less = 0, equal = 0, greater = 0;
5598
        FTNODE node = NULL;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5599
        {
Yoni Fogel's avatar
Yoni Fogel committed
5600
            uint32_t fullhash;
5601
            CACHEKEY root_key;
5602
            toku_calculate_root_offset_pointer(brt->ft, &root_key, &fullhash);
5603
            toku_pin_ftnode_off_client_thread_batched(
5604
                brt->ft,
5605
                root_key,
5606
                fullhash,
5607
                &bfe,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5608
                PL_READ, // may_modify_node, cannot change root during keyrange
5609 5610 5611 5612
                0,
                NULL,
                &node
                );
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5613
        }
5614

Yoni Fogel's avatar
Yoni Fogel committed
5615 5616
        struct unlock_ftnode_extra unlock_extra = {brt,node,false};
        struct unlockers unlockers = {true, unlock_ftnode_fun, (void*)&unlock_extra, (UNLOCKERS)NULL};
5617 5618

        {
5619
            int64_t numrows = brt->ft->in_memory_stats.numrows;
5620 5621
            if (numrows < 0)
                numrows = 0;  // prevent appearance of a negative number
5622
            int r = toku_ft_keyrange_internal (brt, node, key,
5623 5624 5625 5626 5627 5628 5629 5630 5631 5632
                                                &less, &equal, &greater,
                                                numrows,
                                                &bfe, &unlockers, (ANCESTORS)NULL, &infinite_bounds);
            assert(r == 0 || r == TOKUDB_TRY_AGAIN);
            if (r == TOKUDB_TRY_AGAIN) {
                assert(!unlockers.locked);
                goto try_again;
            }
        }
        assert(unlockers.locked);
5633
        toku_unpin_ftnode_read_only(brt->ft, node);
5634 5635 5636
        *less_p    = less;
        *equal_p   = equal;
        *greater_p = greater;
5637
    }
5638 5639
}

5640
void toku_ft_handle_stat64 (FT_HANDLE brt, TOKUTXN UU(txn), struct ftstat64_s *s) {
5641
    toku_ft_stat64(brt->ft, s);
5642 5643
}

5644 5645
/* ********************* debugging dump ************************ */
static int
5646
toku_dump_ftnode (FILE *file, FT_HANDLE brt, BLOCKNUM blocknum, int depth, const DBT *lorange, const DBT *hirange) {
5647
    int result=0;
5648
    FTNODE node;
5649
    toku_get_node_for_verify(blocknum, brt, &node);
5650
    result=toku_verify_ftnode(brt, brt->ft->h->max_msn_in_ft, brt->ft->h->max_msn_in_ft, false, node, -1, lorange, hirange, NULL, NULL, 0, 1, 0);
Yoni Fogel's avatar
Yoni Fogel committed
5651
    uint32_t fullhash = toku_cachetable_hash(brt->ft->cf, blocknum);
5652
    struct ftnode_fetch_extra bfe;
5653
    fill_bfe_for_full_read(&bfe, brt->ft);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5654 5655
    toku_pin_ftnode_off_client_thread(
        brt->ft,
5656
        blocknum,
5657
        fullhash,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5658 5659 5660
        &bfe,
        PL_WRITE_EXPENSIVE,
        0,
5661
        NULL,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5662
        &node
5663
        );
5664
    assert(node->fullhash==fullhash);
5665
    fprintf(file, "%*sNode=%p\n", depth, "", node);
5666

Zardosht Kasheff's avatar
Zardosht Kasheff committed
5667 5668
    fprintf(file, "%*sNode %" PRId64 " height=%d n_children=%d  keyrange=%s %s\n",
            depth, "", blocknum.b, node->height, node->n_children, (char*)(lorange ? lorange->data : 0), (char*)(hirange ? hirange->data : 0));
5669
    {
5670 5671 5672
        int i;
        for (i=0; i+1< node->n_children; i++) {
            fprintf(file, "%*spivotkey %d =", depth+1, "", i);
5673
            toku_print_BYTESTRING(file, node->childkeys[i].size, (char *) node->childkeys[i].data);
5674 5675 5676 5677
            fprintf(file, "\n");
        }
        for (i=0; i< node->n_children; i++) {
            if (node->height > 0) {
5678
                NONLEAF_CHILDINFO bnc = BNC(node, i);
5679 5680 5681 5682
                fprintf(file, "%*schild %d buffered (%d entries):", depth+1, "", i, toku_bnc_n_entries(bnc));
                FIFO_ITERATE(bnc->buffer, key, keylen, data, datalen, type, msn, xids, UU(is_fresh),
                             {
                                 data=data; datalen=datalen; keylen=keylen;
5683
                                 fprintf(file, "%*s xid=%" PRIu64 " %u (type=%d) msn=0x%" PRIu64 "\n", depth+2, "", xids_get_innermost_xid(xids), (unsigned)toku_dtoh32(*(int*)key), type, msn.msn);
5684 5685 5686 5687 5688 5689 5690 5691 5692
                                 //assert(strlen((char*)key)+1==keylen);
                                 //assert(strlen((char*)data)+1==datalen);
                             });
            }
            else {
                int size = toku_omt_size(BLB_BUFFER(node, i));
                if (0)
                    for (int j=0; j<size; j++) {
                        OMTVALUE v = 0;
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5693
                        int r = toku_omt_fetch(BLB_BUFFER(node, i), j, &v);
5694
                        assert_zero(r);
5695
                        LEAFENTRY CAST_FROM_VOIDP(le, v);
5696
                        fprintf(file, " [%d]=", j);
5697
                        print_leafentry(file, le);
5698 5699 5700 5701 5702 5703 5704 5705 5706 5707
                        fprintf(file, "\n");
                    }
                //               printf(" (%d)%u ", len, *(int*)le_key(data)));
                fprintf(file, "\n");
            }
        }
        if (node->height > 0) {
            for (i=0; i<node->n_children; i++) {
                fprintf(file, "%*schild %d\n", depth, "", i);
                if (i>0) {
5708
                    char *CAST_FROM_VOIDP(key, node->childkeys[i-1].data);
5709 5710
                    fprintf(file, "%*spivot %d len=%u %u\n", depth+1, "", i-1, node->childkeys[i-1].size, (unsigned)toku_dtoh32(*(int*)key));
                }
5711
                toku_dump_ftnode(file, brt, BP_BLOCKNUM(node, i), depth+4,
5712 5713 5714 5715
                                  (i==0) ? lorange : &node->childkeys[i-1],
                                  (i==node->n_children-1) ? hirange : &node->childkeys[i]);
            }
        }
5716
    }
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5717
    toku_unpin_ftnode_off_client_thread(brt->ft, node);
5718 5719 5720
    return result;
}

5721
int toku_dump_ft (FILE *f, FT_HANDLE brt) {
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5722
    int r;
5723 5724
    assert(brt->ft);
    toku_dump_translation_table(f, brt->ft->blocktable);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5725
    {
Yoni Fogel's avatar
Yoni Fogel committed
5726
        uint32_t fullhash = 0;
5727
        CACHEKEY root_key;
5728
        toku_calculate_root_offset_pointer(brt->ft, &root_key, &fullhash);
5729
        r = toku_dump_ftnode(f, brt, root_key, 0, 0, 0);
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5730 5731
    }
    return r;
5732
}
5733

5734
int toku_ft_layer_init(void) {
Yoni Fogel's avatar
Yoni Fogel committed
5735 5736
    int r = 0;
    //Portability must be initialized first
5737 5738 5739
    r = toku_portability_init();
    if (r) { goto exit; }

5740
    partitioned_counters_init();
5741
    status_init();
5742 5743
    toku_checkpoint_init();
    toku_ft_serialize_layer_init();
5744 5745
    toku_mutex_init(&ft_open_close_lock, NULL);
exit:
Yoni Fogel's avatar
Yoni Fogel committed
5746
    return r;
Yoni Fogel's avatar
Yoni Fogel committed
5747 5748
}

5749
void toku_ft_layer_destroy(void) {
5750
    toku_mutex_destroy(&ft_open_close_lock);
5751 5752
    toku_ft_serialize_layer_destroy();
    toku_checkpoint_destroy();
5753
    status_destroy();
5754
    partitioned_counters_destroy();
Yoni Fogel's avatar
Yoni Fogel committed
5755
    //Portability must be cleaned up last
5756
    toku_portability_destroy();
Yoni Fogel's avatar
Yoni Fogel committed
5757 5758
}

5759 5760
// This lock serializes all opens and closes because the cachetable requires that clients do not try to open or close a cachefile in parallel.  We made
// it coarser by not allowing any cachefiles to be open or closed in parallel.
5761 5762 5763 5764 5765 5766 5767 5768
void toku_ft_open_close_lock(void) {
    toku_mutex_lock(&ft_open_close_lock);
}

void toku_ft_open_close_unlock(void) {
    toku_mutex_unlock(&ft_open_close_lock);
}

5769 5770
// Prepare to remove a dictionary from the database when this transaction is committed:
//  - mark transaction as NEED fsync on commit
5771
//  - make entry in rollback log
5772
//  - make fdelete entry in recovery log
5773 5774 5775 5776 5777 5778 5779 5780 5781 5782 5783 5784
//
// Effect: when the txn commits, the ft's cachefile will be marked as unlink
//         on close. see toku_commit_fdelete and how unlink on close works
//         in toku_cachefile_close();
// Requires: serialized with begin checkpoint
//           this does not need to take the open close lock because
//           1.) the ft/cf cannot go away because we have a live handle.
//           2.) we're not setting the unlink on close bit _here_. that
//           happens on txn commit (as the name suggests).
//           3.) we're already holding the multi operation lock to 
//           synchronize with begin checkpoint.
// Contract: the iname of the ft should never be reused.
5785
void toku_ft_unlink_on_commit(FT_HANDLE handle, TOKUTXN txn) {
5786
    assert(txn);
5787 5788

    CACHEFILE cf = handle->ft->cf;
5789
    FT CAST_FROM_VOIDP(ft, toku_cachefile_get_userdata(cf));
5790

5791 5792
    toku_txn_maybe_note_ft(txn, ft);

5793 5794 5795 5796
    // If the txn commits, the commit MUST be in the log before the file is actually unlinked
    toku_txn_force_fsync_on_commit(txn); 
    // make entry in rollback log
    FILENUM filenum = toku_cachefile_filenum(cf);
5797
    toku_logger_save_rollback_fdelete(txn, filenum);
5798
    // make entry in recovery log
5799
    toku_logger_log_fdelete(txn, filenum);
5800 5801
}

5802 5803 5804 5805 5806 5807
// Non-transactional version of fdelete
//
// Effect: The ft file is unlinked when the handle closes and it's ft is not
//         pinned by checkpoint. see toku_remove_ft_ref() and how unlink on
//         close works in toku_cachefile_close();
// Requires: serialized with begin checkpoint
5808
void toku_ft_unlink(FT_HANDLE handle) {
5809
    CACHEFILE cf;
5810 5811
    cf = handle->ft->cf;
    toku_cachefile_unlink_on_close(cf);
5812 5813
}

5814
int
5815
toku_ft_get_fragmentation(FT_HANDLE brt, TOKU_DB_FRAGMENTATION report) {
5816 5817
    int r;

5818
    int fd = toku_cachefile_get_fd(brt->ft->cf);
5819
    toku_ft_lock(brt->ft);
5820 5821

    int64_t file_size;
5822
    r = toku_os_get_file_size(fd, &file_size);
5823
    if (r==0) {
5824
        report->file_size_bytes = file_size;
5825
        toku_block_table_get_fragmentation_unlocked(brt->ft->blocktable, report);
5826
    }
5827
    toku_ft_unlock(brt->ft);
5828 5829
    return r;
}
5830

Yoni Fogel's avatar
Yoni Fogel committed
5831
static bool is_empty_fast_iter (FT_HANDLE brt, FTNODE node) {
5832
    if (node->height > 0) {
5833
        for (int childnum=0; childnum<node->n_children; childnum++) {
5834
            if (toku_bnc_nbytesinbuf(BNC(node, childnum)) != 0) {
5835 5836
                return 0; // it's not empty if there are bytes in buffers
            }
5837
            FTNODE childnode;
5838 5839
            {
                BLOCKNUM childblocknum = BP_BLOCKNUM(node,childnum);
Yoni Fogel's avatar
Yoni Fogel committed
5840
                uint32_t fullhash =  compute_child_fullhash(brt->ft->cf, node, childnum);
5841
                struct ftnode_fetch_extra bfe;
5842
                fill_bfe_for_full_read(&bfe, brt->ft);
5843
                // don't need to pass in dependent nodes as we are not
5844
                // modifying nodes we are pinning
5845
                toku_pin_ftnode_off_client_thread(
5846
                    brt->ft,
5847 5848 5849
                    childblocknum,
                    fullhash,
                    &bfe,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5850
                    PL_READ, // may_modify_node set to false, as nodes not modified
5851 5852 5853
                    0,
                    NULL,
                    &childnode
5854
                    );
5855 5856
            }
            int child_is_empty = is_empty_fast_iter(brt, childnode);
5857
            toku_unpin_ftnode(brt->ft, childnode);
5858 5859 5860
            if (!child_is_empty) return 0;
        }
        return 1;
5861
    } else {
5862 5863 5864
        // leaf:  If the omt is empty, we are happy.
        for (int i = 0; i < node->n_children; i++) {
            if (toku_omt_size(BLB_BUFFER(node, i))) {
Yoni Fogel's avatar
Yoni Fogel committed
5865
                return false;
5866 5867
            }
        }
Yoni Fogel's avatar
Yoni Fogel committed
5868
        return true;
5869 5870 5871
    }
}

Yoni Fogel's avatar
Yoni Fogel committed
5872
bool toku_ft_is_empty_fast (FT_HANDLE brt)
5873 5874 5875
// A fast check to see if the tree is empty.  If there are any messages or leafentries, we consider the tree to be nonempty.  It's possible that those
// messages and leafentries would all optimize away and that the tree is empty, but we'll say it is nonempty.
{
Yoni Fogel's avatar
Yoni Fogel committed
5876
    uint32_t fullhash;
5877
    FTNODE node;
5878
    {
5879
        CACHEKEY root_key;
5880
        toku_calculate_root_offset_pointer(brt->ft, &root_key, &fullhash);
5881
        struct ftnode_fetch_extra bfe;
5882
        fill_bfe_for_full_read(&bfe, brt->ft);
5883
        toku_pin_ftnode_off_client_thread(
5884
            brt->ft,
5885
            root_key,
5886
            fullhash,
5887
            &bfe,
Zardosht Kasheff's avatar
Zardosht Kasheff committed
5888
            PL_READ, // may_modify_node set to false, node does not change
5889 5890 5891
            0,
            NULL,
            &node
5892
            );
5893
    }
Yoni Fogel's avatar
Yoni Fogel committed
5894
    bool r = is_empty_fast_iter(brt, node);
5895
    toku_unpin_ftnode(brt->ft, node);
5896 5897
    return r;
}
5898

5899
// test-only
5900
int toku_ft_strerror_r(int error, char *buf, size_t buflen)
5901 5902
{
    if (error>=0) {
5903
        return (long) strerror_r(error, buf, buflen);
5904
    } else {
5905 5906 5907 5908 5909 5910 5911 5912 5913
        switch (error) {
        case DB_KEYEXIST:
            snprintf(buf, buflen, "Key exists");
            return 0;
        case TOKUDB_CANCELED:
            snprintf(buf, buflen, "User canceled operation");
            return 0;
        default:
            snprintf(buf, buflen, "Unknown error %d", error);
5914
            return EINVAL;
5915
        }
5916 5917
    }
}
5918

5919
#include <toku_race_tools.h>
5920
void __attribute__((__constructor__)) toku_ft_helgrind_ignore(void);
5921
void
5922
toku_ft_helgrind_ignore(void) {
5923
    TOKU_VALGRIND_HG_DISABLE_CHECKING(&ft_status, sizeof ft_status);
5924
}
5925

5926
#undef STATUS_INC