virtmem.c 29 KB
Newer Older
1
/* Wendelin.bigfile | Virtual memory
2
 * Copyright (C) 2014-2020  Nexedi SA and Contributors.
3 4 5 6 7 8 9
 *                          Kirill Smelkov <kirr@nexedi.com>
 *
 * This program is free software: you can Use, Study, Modify and Redistribute
 * it under the terms of the GNU General Public License version 3, or (at your
 * option) any later version, as published by the Free Software Foundation.
 *
 * You can also Link and Combine this program with other software covered by
10 11 12 13
 * the terms of any of the Free Software licenses or any of the Open Source
 * Initiative approved licenses and Convey the resulting work. Corresponding
 * source of such a combination shall include the source code for all other
 * software used.
14 15 16 17 18
 *
 * This program is distributed WITHOUT ANY WARRANTY; without even the implied
 * warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.
 *
 * See COPYING file for full licensing terms.
19
 * See https://www.nexedi.com/licensing for rationale and options.
20 21 22
 *
 *
 * TODO description
23 24 25
 */

#include <wendelin/bigfile/virtmem.h>
26 27
#include <wendelin/bigfile/file.h>
#include <wendelin/bigfile/pagemap.h>
28
#include <wendelin/bigfile/ram.h>
29
#include <wendelin/utils.h>
30 31 32
#include <wendelin/bug.h>

#include <ccan/minmax/minmax.h>
33
#include <ccan/bitmap/bitmap.h>
34 35 36

#include <sys/mman.h>
#include <errno.h>
37 38
#include <signal.h>
#include <stdio.h>
39
#include <unistd.h>
40 41 42

static size_t   page_size(const Page *page);
static void     page_drop_memory(Page *page);
43
static void     page_del(Page *page);
44 45
static void    *vma_page_addr(VMA *vma, Page *page);
static pgoff_t  vma_addr_fpgoffset(VMA *vma, uintptr_t addr);
46
static void     vma_mmap_page(VMA *vma, Page *page);
47
static int      vma_page_infilerange(VMA *vma, Page *page);
48 49 50
static int      vma_page_ismapped(VMA *vma, Page *page);
static void     vma_page_ensure_unmapped(VMA *vma, Page *page);
static void     vma_page_ensure_notmappedrw(VMA *vma, Page *page);
51
static int      __ram_reclaim(RAM *ram);
52 53 54 55 56 57 58 59

#define VIRT_DEBUG   0
#if VIRT_DEBUG
# define TRACE(msg, ...) do { fprintf(stderr, msg, ## __VA_ARGS__); } while (0)
#else
# define TRACE(msg, ...) do {} while(0)
#endif

60 61 62 63

/* global lock which protects manipulating virtmem data structures
 *
 * NOTE not scalable, but this is temporary solution - as we are going to move
Kirill Smelkov's avatar
Kirill Smelkov committed
64
 * memory management back into the kernel, where it is done properly. */
65
static pthread_mutex_t virtmem_lock = PTHREAD_ERRORCHECK_MUTEX_INITIALIZER_NP;
66 67
static const VirtGilHooks *virtmem_gilhooks;

68
void *virt_gil_ensure_unlocked(void)
69 70 71 72 73 74
{
    void *gilstate = NULL;

    if (virtmem_gilhooks)
        gilstate = virtmem_gilhooks->gil_ensure_unlocked();

75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91
    return gilstate;
}


void virt_gil_retake_if_waslocked(void *gilstate)
{
    if (gilstate)
        virtmem_gilhooks->gil_retake_if_waslocked(gilstate);
}

void virt_lock()
{
    void *gilstate = NULL;

    /* make sure we don't hold e.g. python GIL (not to deadlock, as GIL oscillates) */
    gilstate = virt_gil_ensure_unlocked();

92 93 94 95
    /* acquire virtmem lock */
    xpthread_mutex_lock(&virtmem_lock);

    /* retake GIL if we were holding it originally */
96
    virt_gil_retake_if_waslocked(gilstate);
97 98 99 100 101 102 103 104 105 106 107 108 109 110 111
}

void virt_unlock()
{
    xpthread_mutex_unlock(&virtmem_lock);
}


void virt_lock_hookgil(const VirtGilHooks *gilhooks)
{
    BUG_ON(virtmem_gilhooks);       /* prevent registering multiple times */
    virtmem_gilhooks = gilhooks;
}


112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127
/* block/restore SIGSEGV for current thread - non on-pagefault code should not
 * access any not-mmapped memory -> so on any pagefault we should just die with
 * coredump, not try to incorrectly handle the pagefault.
 *
 * NOTE sigmask is per-thread. When blocking there is no race wrt other threads
 * correctly accessing data via pagefaulting.   */
static void sigsegv_block(sigset_t *save_sigset)
{
    sigset_t mask_segv;
    xsigemptyset(&mask_segv);
    xsigaddset(&mask_segv, SIGSEGV);
    xpthread_sigmask(SIG_BLOCK, &mask_segv, save_sigset);
}

static void sigsegv_restore(const sigset_t *save_sigset)
{
128 129 130 131 132 133
    int how = xsigismember(save_sigset, SIGSEGV) ? SIG_BLOCK : SIG_UNBLOCK;
    sigset_t mask_segv;

    xsigemptyset(&mask_segv);
    xsigaddset(&mask_segv, SIGSEGV);
    xpthread_sigmask(how, &mask_segv, NULL);
134 135
}

136 137 138 139 140 141 142

/****************
 * OPEN / CLOSE *
 ****************/

int fileh_open(BigFileH *fileh, BigFile *file, RAM *ram)
{
143 144 145 146
    int err = 0;
    sigset_t save_sigset;

    sigsegv_block(&save_sigset);
147
    virt_lock();
148

149 150
    bzero(fileh, sizeof(*fileh));
    fileh->ramh = ramh_open(ram);
151 152
    if (!fileh->ramh) {
        err = -1;
153
        goto out;
154
    }
155 156 157

    fileh->file = file;
    INIT_LIST_HEAD(&fileh->mmaps);
158
    INIT_LIST_HEAD(&fileh->dirty_pages);
159
    fileh->writeout_inprogress = 0;
160 161 162
    pagemap_init(&fileh->pagemap, ilog2_exact(ram->pagesize));

out:
163
    virt_unlock();
164 165
    sigsegv_restore(&save_sigset);
    return err;
166 167 168 169 170 171
}


void fileh_close(BigFileH *fileh)
{
    Page *page;
172 173 174
    sigset_t save_sigset;

    sigsegv_block(&save_sigset);
175
    virt_lock();
176 177 178 179 180 181

    /* it's an error to close fileh with existing mappings */
    // XXX implement the same semantics usual files have wrt mmaps - if we release
    // fileh, but mapping exists - real fileh release is delayed to last unmap ?
    BUG_ON(!list_empty(&fileh->mmaps));

182 183 184
    /* it's an error to close fileh while writeout is in progress */
    BUG_ON(fileh->writeout_inprogress);

185 186
    /* drop all pages (dirty or not) associated with this fileh */
    pagemap_for_each(page, &fileh->pagemap) {
187 188 189
        /* it's an error to close fileh to mapping of which an access is
         * currently being done in another thread */
        BUG_ON(page->state == PAGE_LOADING);
190
        page_drop_memory(page);
191
        page_del(page);
192 193
    }

194 195
    BUG_ON(!list_empty(&fileh->dirty_pages));

196 197 198 199 200 201 202
    /* and clear pagemap */
    pagemap_clear(&fileh->pagemap);

    if (fileh->ramh)
        ramh_close(fileh->ramh);

    bzero(fileh, sizeof(*fileh));
203
    virt_unlock();
204
    sigsegv_restore(&save_sigset);
205 206 207 208 209 210 211 212 213 214 215 216
}



/****************
 * MMAP / UNMAP *
 ****************/

int fileh_mmap(VMA *vma, BigFileH *fileh, pgoff_t pgoffset, pgoff_t pglen)
{
    void *addr;
    size_t len = pglen * fileh->ramh->ram->pagesize;
217 218 219 220
    int err = 0;
    sigset_t save_sigset;

    sigsegv_block(&save_sigset);
221
    virt_lock();
222

223
    /* start preparing vma */
224
    bzero(vma, sizeof(*vma));
225 226
    vma->fileh       = fileh;
    vma->f_pgoffset  = pgoffset;
227

228
    /* alloc vma->page_ismappedv[] */
229 230
    vma->page_ismappedv = bitmap_alloc0(pglen);
    if (!vma->page_ismappedv)
231
        goto fail;
232 233 234 235

    /* allocate address space somewhere */
    addr = mem_valloc(NULL, len);
    if (!addr)
236
        goto fail;
237 238 239 240 241 242 243 244 245 246 247 248 249

    /* everything allocated - link it up */
    vma->addr_start  = (uintptr_t)addr;
    vma->addr_stop   = vma->addr_start + len;


    // XXX need to init vma->virt_list first?
    /* hook vma to fileh->mmaps */
    list_add_tail(&vma->same_fileh, &fileh->mmaps);

    /* register vma for pagefault handling */
    virt_register_vma(vma);

250
out:
251
    virt_unlock();
252 253
    sigsegv_restore(&save_sigset);
    return err;
254

255
fail:
256
    free(vma->page_ismappedv);
257
    bzero(vma, sizeof(*vma));
258 259
    err = -1;
    goto out;
260 261 262 263 264 265 266 267 268 269 270
}


void vma_unmap(VMA *vma)
{
    BigFileH *fileh = vma->fileh;
    size_t   len    = vma->addr_stop - vma->addr_start;
    size_t   pglen  = len / fileh->ramh->ram->pagesize;
    int i;
    pgoff_t  pgoffset;
    Page *page;
271
    sigset_t save_sigset;
272

273
    sigsegv_block(&save_sigset);
274
    virt_lock();
275

276 277 278 279 280 281 282 283 284 285
    /* unregister from vmamap - so that pagefault handler does not recognize
     * this area as valid */
    virt_unregister_vma(vma);

    /* unlink from fileh.mmaps   XXX place ok ? */
    list_del_init(&vma->same_fileh);

    /* unmap whole vma at once - the kernel unmaps each mapping in turn.
     * NOTE error here would mean something is broken */
    xmunmap((void *)vma->addr_start, len);
286

287 288 289 290 291 292 293 294 295 296 297 298 299 300 301
    /* scan through mapped-to-this-vma pages and release them */
    for (i=0; i < pglen; ++i) {
        if (!bitmap_test_bit(vma->page_ismappedv, i))
            continue;

        pgoffset = vma->f_pgoffset + i;
        page = pagemap_get(&fileh->pagemap, pgoffset);
        BUG_ON(!page);
        page_decref(page);
    }

    /* free memory and be done */
    free(vma->page_ismappedv);

    bzero(vma, sizeof(*vma));
302
    virt_unlock();
303
    sigsegv_restore(&save_sigset);
304 305 306 307 308 309 310
}


/**********************
 * WRITEOUT / DISCARD *
 **********************/

311 312 313 314 315 316 317 318 319 320 321 322 323
/* helper for sorting dirty pages by ->f_pgoffset */
static int hpage_indirty_cmp_bypgoffset(struct list_head *hpage1, struct list_head *hpage2, void *_)
{
    Page *page1 = list_entry(hpage1, typeof(*page1), in_dirty);
    Page *page2 = list_entry(hpage2, typeof(*page2), in_dirty);

    if (page1->f_pgoffset < page2->f_pgoffset)
        return -1;
    if (page1->f_pgoffset > page2->f_pgoffset)
        return +1;
    return 0;
}

324 325 326 327
int fileh_dirty_writeout(BigFileH *fileh, enum WriteoutFlags flags)
{
    Page *page;
    BigFile *file = fileh->file;
328
    struct list_head *hpage, *hpage_next, *hmmap;
329
    sigset_t save_sigset;
330 331 332 333 334 335 336
    int err = 0;

    /* check flags */
    if (!(flags &  (WRITEOUT_STORE | WRITEOUT_MARKSTORED))   ||
          flags & ~(WRITEOUT_STORE | WRITEOUT_MARKSTORED))
        return -EINVAL;

337
    sigsegv_block(&save_sigset);
338
    virt_lock();
339

340 341 342 343
    /* concurrent writeouts are not allowed */
    BUG_ON(fileh->writeout_inprogress);
    fileh->writeout_inprogress = 1;

344 345 346 347
    /* pages are stored (if stored) in sorted order */
    if (flags & WRITEOUT_STORE)
        list_sort(&fileh->dirty_pages, hpage_indirty_cmp_bypgoffset, NULL);

348
    /* write out dirty pages */
349 350 351
    list_for_each_safe(hpage, hpage_next, &fileh->dirty_pages) {
        page = list_entry(hpage, typeof(*page), in_dirty);
        BUG_ON(page->state != PAGE_DIRTY);
352 353 354 355 356 357 358 359

        /* ->storeblk() */
        if (flags & WRITEOUT_STORE) {
            TODO (file->blksize != page_size(page));
            blk_t blk = page->f_pgoffset;   // NOTE assumes blksize = pagesize

            void *pagebuf;

360 361 362 363 364 365 366 367 368 369 370 371 372 373 374
            /* mmap page temporarily somewhere
             *
             * ( we cannot use present page mapping in some vma directly,
             *   because while storeblk is called with virtmem lock released that
             *   mapping can go away ) */
            pagebuf = page_mmap(page, NULL, PROT_READ);
            TODO(!pagebuf); // XXX err

            /* unlock virtmem before calling storeblk()
             *
             * that call is potentially slow and external code can take other
             * locks. If that "other locks" are also taken before external code
             * calls e.g. fileh_invalidate_page() in different codepath a deadlock
             * can happen. (similar to loadblk case) */
            virt_unlock();
375 376 377

            err = file->file_ops->storeblk(file, blk, pagebuf);

378 379 380 381
            /* relock virtmem */
            virt_lock();

            xmunmap(pagebuf, page_size(page));
382 383 384 385 386 387 388 389

            if (err)
                goto out;
        }

        /* page.state -> PAGE_LOADED and correct mappings RW -> R */
        if (flags & WRITEOUT_MARKSTORED) {
            page->state = PAGE_LOADED;
390
            list_del_init(&page->in_dirty);
391 392 393 394 395 396 397 398 399

            list_for_each(hmmap, &fileh->mmaps) {
                VMA *vma = list_entry(hmmap, typeof(*vma), same_fileh);
                vma_page_ensure_notmappedrw(vma, page);
            }
        }
    }


400 401
    /* if we successfully finished with markstored flag set - all dirty pages
     * should become non-dirty */
402
    if (flags & WRITEOUT_MARKSTORED)
403
        BUG_ON(!list_empty(&fileh->dirty_pages));
404

405
out:
406 407
    fileh->writeout_inprogress = 0;

408
    virt_unlock();
409
    sigsegv_restore(&save_sigset);
410 411 412 413 414 415 416
    return err;
}


void fileh_dirty_discard(BigFileH *fileh)
{
    Page *page;
417
    struct list_head *hpage, *hpage_next;
418 419 420
    sigset_t save_sigset;

    sigsegv_block(&save_sigset);
421
    virt_lock();
422

423 424
    /* discard is not allowed to run in parallel to writeout */
    BUG_ON(fileh->writeout_inprogress);
425

426 427 428 429 430 431 432 433
    list_for_each_safe(hpage, hpage_next, &fileh->dirty_pages) {
        page = list_entry(hpage, typeof(*page), in_dirty);
        BUG_ON(page->state != PAGE_DIRTY);

        page_drop_memory(page);
    }

    BUG_ON(!list_empty(&fileh->dirty_pages));
434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451

    virt_unlock();
    sigsegv_restore(&save_sigset);
}


/****************
 * INVALIDATION *
 ****************/

void fileh_invalidate_page(BigFileH *fileh, pgoff_t pgoffset)
{
    Page *page;
    sigset_t save_sigset;

    sigsegv_block(&save_sigset);
    virt_lock();

452 453 454
    /* it's an error to invalidate fileh while writeout is in progress */
    BUG_ON(fileh->writeout_inprogress);

455
    page = pagemap_get(&fileh->pagemap, pgoffset);
456 457 458 459 460 461 462 463 464 465 466 467 468 469
    if (page) {
        /* for pages where loading is in progress, we just remove the page from
         * pagemap and mark it to be dropped by their loaders after it is done.
         * In the mean time, as pagemap entry is now empty, on next access to
         * the memory the page will be created/loaded anew */
        if (page->state == PAGE_LOADING) {
            pagemap_del(&fileh->pagemap, pgoffset);
            page->state = PAGE_LOADING_INVALIDATED;
        }
        /* else we just make sure to drop page memory */
        else {
            page_drop_memory(page);
        }
    }
470

471
    virt_unlock();
472
    sigsegv_restore(&save_sigset);
473 474 475 476 477 478 479 480 481 482 483 484 485 486 487
}


/************************
 *  Lookup VMA by addr  *
 ************************/

/* list of all registered VMA(s) */
static LIST_HEAD(vma_list);

/* protects ^^^  XXX */
//spinlock_t vma_list_lock;



488 489
/* lookup VMA covering `addr`. NULL if not found
 * (should be called with virtmem lock held) */
490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510
VMA *virt_lookup_vma(void *addr)
{
    uintptr_t uaddr = (uintptr_t)addr;
    struct list_head *h;
    VMA *vma;

    list_for_each(h, &vma_list) {
        // XXX -> list_for_each_entry
        vma = list_entry(h, typeof(*vma), virt_list);
        if (uaddr < vma->addr_stop)
            /*
             * here:  vma->addr_start ? uaddr < vma->addr_stop
             * vma->addr_stop is first such addr_stop
             */
            return (vma->addr_start <= uaddr) ? vma : NULL;
    }

    return NULL;    /* not found at all or no overlap */
}


511 512
/* register VMA `vma` as covering some file view
 * (should be called with virtmem lock held) */
513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529
void virt_register_vma(VMA *vma)
{
    uintptr_t uaddr = vma->addr_start;
    struct list_head *h;
    struct VMA *a;

    list_for_each(h, &vma_list) {
        a = list_entry(h, typeof(*a), virt_list);
        if (uaddr < a->addr_stop)
            break;
    }

    /* either before found vma or, if not found, at the end of the list */
    list_add_tail(&vma->virt_list, h);
}


530 531
/* remove `area` from VMA registry. `area` must be registered before
 * (should be called with virtmem lock held) */
532 533 534 535 536 537 538 539
void virt_unregister_vma(VMA *vma)
{
    /* _init - to clear links, just in case */
    list_del_init(&vma->virt_list);
}


/*****************************************/
540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578

/*
 * allocate virtual memory address space
 * the pages are initially protected to prevent any access
 *
 * @addr    NULL - at anywhere,     !NULL - exactly there
 * @return  !NULL - mapped there    NULL - error
 */
void *mem_valloc(void *addr, size_t len)
{
    void *a;
    a = mmap(addr, len, PROT_NONE,
            MAP_PRIVATE | MAP_ANONYMOUS
            /* don't try to (pre-)allocate memory - just virtual address space */
            | MAP_NORESERVE
            | (addr ? MAP_FIXED : 0),
            -1, 0);

    if (a == MAP_FAILED)
        a = NULL;

    if (a && addr)
        /* verify OS respected our MAP_FIXED request */
        BUG_ON(a != addr);

    return a;
}


/* like mem_valloc() but allocation must not fail */
void *mem_xvalloc(void *addr, size_t len)
{
    void *a;
    a = mem_valloc(addr, len);
    BUG_ON(!a);
    return a;
}


579 580 581 582
/*********************
 * PAGEFAULT HANDLER *
 *********************/

583 584 585
/* pagefault entry when we know request came to our memory area
 *
 * (virtmem_lock already taken by caller)   */
586
VMFaultResult vma_on_pagefault(VMA *vma, uintptr_t addr, int write)
587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607
{
    pgoff_t pagen;
    Page *page;
    BigFileH *fileh;

    /* continuing on_pagefault() - see (1) there ... */

    /* (2) vma, addr -> fileh, pagen    ;idx of fileh page covering addr */
    fileh = vma->fileh;
    pagen = vma_addr_fpgoffset(vma, addr);

    /* (3) fileh, pagen -> page  (via pagemap) */
    page = pagemap_get(&fileh->pagemap, pagen);

    /* (4) no page found - allocate new from ram */
    while (!page) {
        page = ramh_alloc_page(fileh->ramh, pagen);
        if (!page) {
            /* try to release some memory back to OS */
            // XXX do we need and how to distinguish "no ram page" vs "no memory for `struct page`"?
            //     -> no we don't -- better allocate memory for struct pages for whole RAM at ram setup
608
            if (!__ram_reclaim(fileh->ramh->ram))
609 610 611 612 613 614 615 616 617 618 619 620 621
                OOM();
            continue;
        }

        /* ramh set up .ramh, .ramh_pgoffset, .state?
         * now setup rest (link to fileh)  */
        page->fileh      = fileh;
        page->f_pgoffset = pagen;

        /* remember page in fileh->pagemap[pagen] */
        pagemap_set(&fileh->pagemap, pagen, page);
    }

622 623
    /* (5a) if page was not yet loaded - start loading it */
    if (page->state == PAGE_EMPTY) {
624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639
        /* NOTE if we load data in-place, there would be a race with concurrent
         * access to the page here - after first enabling memory-access to
         * the page, other threads could end up reading corrupt data, while
         * loading had not finished.
         *
         * so to avoid it we first load data to separate memory address, then
         * mmap-duplicate that page into here, but it is more work compared to
         * what kernel internally does.
         *
         * TODO try to use remap_anon_pages() when it is ready
         *      (but unfortunately it is only for anonymous memory)
         * NOTE remap_file_pages() is going away...
         */
        blk_t blk;
        void *pageram;
        int err;
640
        BigFile *file;
641 642 643 644 645 646

        /*
         * if pagesize < blksize - need to prepare several adjacent pages for blk;
         * if pagesize > blksize - will need to either 1) rescan which blk got
         *    dirty, or 2) store not-even-touched blocks adjacent to modified one.
         */
647 648
        file = fileh->file;
        TODO (file->blksize != page_size(page));
649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670

        // FIXME doing this mmap-to-temp/unmap is somewhat costly. Better
        // constantly have whole RAM mapping somewhere R/W and load there.
        // (XXX but then we'll either have
        //    - VMA fragmented  (if we manage whole RAM as 1 file of physram size),
        //    - or need to waste a lot of address space (size of each ramh can be very large)
        //
        //    generally this way it also has major problems)
        //
        // Also this way, we btw don't need to require python code to drop all
        // references to loading buf.

        /* mmap page memory temporarily somewhere
         * XXX better pre-map all ram pages r/w in another area to not need to mmap/unmap it here
         *     -> will run slightly faster (but major slowdown is in clear_page in kernel)
         */
        // TODO MAP_UNINITIALIZED somehow? (we'll overwrite that memory)
        pageram = page_mmap(page, NULL, PROT_READ | PROT_WRITE);
        TODO(!pageram); // XXX err

        /* loadblk() -> pageram memory */
        blk = page->f_pgoffset;     // NOTE because blksize = pagesize
671 672 673 674 675 676

        /* mark page as loading and unlock virtmem before calling loadblk()
         *
         * that call is potentially slow and external code can take other
         * locks. If that "other locks" are also taken before external code
         * calls e.g. fileh_invalidate_page() in different codepath a deadlock
677
         * can happen. (similar to storeblk case) */
678 679 680 681 682
        page->state = PAGE_LOADING;
        virt_unlock();

        err = file->file_ops->loadblk(file, blk, pageram);

683 684 685 686 687 688 689 690
        /* TODO on error -> try to throw exception somehow to the caller, so
         *      that it can abort current transaction, but not die.
         *
         * NOTE for analogue situation when read for mmaped file fails, the
         *      kernel sends SIGBUS
         */
        TODO (err);

691 692 693
        /* relock virtmem */
        virt_lock();

694 695
        xmunmap(pageram, page_size(page));

696 697 698 699 700
        /* if the page was invalidated while we were loading it, we have to drop
         * it's memory and the Page structure completely - invalidater already
         * removed it from pagemap */
        if (page->state == PAGE_LOADING_INVALIDATED) {
            page_drop_memory(page);
701
            page_del(page);
702 703 704 705 706 707 708 709 710 711 712 713 714
        }

        /* else just mark the page as loaded ok */
        else
            page->state = PAGE_LOADED;

        /* we have to retry the whole fault, because the vma could have been
         * changed while we were loading page with virtmem lock released */
        return VM_RETRY;
    }

    /* (5b) page is currently being loaded by another thread - wait for load to complete
     *
Kirill Smelkov's avatar
Kirill Smelkov committed
715
     * NOTE a page is protected from being concurrently loaded by two threads at
716 717 718 719 720 721 722 723 724 725 726 727 728 729
     * the same time via:
     *
     *   - virtmem lock - we get/put pages from fileh->pagemap only under it
     *   - page->state is set PAGE_LOADING for loading in progress pages
     *   - such page is inserted in fileh->pagepam
     *
     * so if second thread faults at the same memory page, and the page is
     * still loading, it will find the page in PAGE_LOADING state and will just
     * wait for it to complete. */
    if (page->state == PAGE_LOADING) {
        /* XXX polling instead of proper completion */
        void *gilstate;
        virt_unlock();
        gilstate = virt_gil_ensure_unlocked();
Kirill Smelkov's avatar
Kirill Smelkov committed
730
        usleep(10000);  // XXX with 1000 usleep still busywaits
731 732 733
        virt_gil_retake_if_waslocked(gilstate);
        virt_lock();
        return VM_RETRY;
734 735 736 737 738 739 740 741 742 743
    }

    /* (6) page data ready. Mmap it atomically into vma address space, or mprotect
     * appropriately if it was already mmaped. */
    PageState newstate = PAGE_LOADED;
    if (write || page->state == PAGE_DIRTY) {
        newstate = PAGE_DIRTY;
    }

    // XXX also call page->markdirty() ?
744 745 746 747
    if (newstate == PAGE_DIRTY  &&  newstate != page->state) {
        /* it is not allowed to modify pages while writeout is in progress */
        BUG_ON(fileh->writeout_inprogress);

748
        list_add_tail(&page->in_dirty, &fileh->dirty_pages);
749
    }
750 751
    page->state = max(page->state, newstate);

752 753
    vma_mmap_page(vma, page);

754 755 756 757 758 759 760 761 762
    /* mark page as used recently */
    // XXX = list_move_tail()
    list_del(&page->lru);
    list_add_tail(&page->lru, &page->ramh->ram->lru_list);

    /*
     * (7) access to page prepared - now it is ok to return from signal handler
     *     - the caller will re-try executing faulting instruction.
     */
763
    return VM_HANDLED;
764 765 766 767 768 769 770 771
}


/***********
 * RECLAIM *
 ***********/

#define RECLAIM_BATCH   64      /* how many pages to reclaim at once */
772
static int __ram_reclaim(RAM *ram)
773 774 775 776 777 778 779 780 781 782 783 784 785 786
{
    struct list_head *lru_list = &ram->lru_list;
    struct list_head *hlru;
    Page *page;
    int batch = RECLAIM_BATCH, scanned = 0;

    TRACE("RAM_RECLAIM\n");
    hlru = lru_list->next;

    while (batch && hlru != lru_list) {
        page = list_entry(hlru, typeof(*page), lru);
        hlru = hlru->next;
        scanned++;

787 788
        /* can release ram only from loaded non-dirty pages
         * NOTE PAGE_LOADING pages are not dropped - they just continue to load */
789 790 791 792 793 794 795 796 797 798 799
        if (page->state == PAGE_LOADED) {
            page_drop_memory(page);
            batch--;
        }

        /* PAGE_EMPTY pages without mappers go away */
        if (page->state == PAGE_EMPTY) {
            BUG_ON(page->refcnt != 0);  // XXX what for then we have refcnt? -> vs discard

            /* delete page & its entry in fileh->pagemap */
            pagemap_del(&page->fileh->pagemap, page->f_pgoffset);
800
            page_del(page);
801 802 803 804 805 806 807 808 809
        }

    }

    TRACE("\t-> reclaimed %i  scanned %i\n", RECLAIM_BATCH - batch, scanned);
    return RECLAIM_BATCH - batch;
}


810 811 812 813 814 815 816 817 818 819 820 821 822 823 824
int ram_reclaim(RAM *ram)
{
    int ret;
    sigset_t save_sigset;

    sigsegv_block(&save_sigset);
    virt_lock();

    ret = __ram_reclaim(ram);

    virt_unlock();
    sigsegv_restore(&save_sigset);
    return ret;
}

825

826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851
/********************
 * Internal helpers *
 ********************/

static size_t page_size(const Page *page)
{
    return page->ramh->ram->pagesize;
}

void page_incref(Page *page)
{
    page->refcnt++;     // XXX atomically ?
}

void page_decref(Page *page)
{
    page->refcnt--;     // XXX atomically ?
    BUG_ON(page->refcnt < 0);

    // TODO if unused delete self && clear pagemap ?
    // XXX  if dirty -> delete = not ok
}

void *page_mmap(Page *page, void *addr, int prot)
{
    RAMH *ramh = page->ramh;
Kirill Smelkov's avatar
Kirill Smelkov committed
852
    // XXX better call ramh_mmap_page() without tinkering with ramh_ops?
853 854
    return ramh->ramh_ops->mmap_page(ramh, page->ramh_pgoffset, addr, prot);
}
855 856


857 858 859
/* page_drop_memory frees RAM associated with page and transitions page state into PAGE_EMPTY.
 *
 * The Page struct itself is not released - see page_del for that. */
860 861 862 863 864
static void page_drop_memory(Page *page)
{
    /* Memory for this page goes out. 1) unmap it from all mmaps */
    struct list_head *hmmap;

865 866 867
    /* NOTE we try not to drop memory for loading-in-progress pages.
     *      so if this is called for such a page - it is a bug. */
    BUG_ON(page->state == PAGE_LOADING);
868 869
    /* same for storing-in-progress */
    BUG_ON(page->fileh->writeout_inprogress && page->state == PAGE_DIRTY);
870

871 872 873 874 875 876 877 878 879 880
    if (page->state == PAGE_EMPTY)
        return;

    list_for_each(hmmap, &page->fileh->mmaps) {
        VMA *vma = list_entry(hmmap, typeof(*vma), same_fileh);
        vma_page_ensure_unmapped(vma, page);
    }

    /* 2) release memory to ram */
    ramh_drop_memory(page->ramh, page->ramh_pgoffset);
881
    if (page->state == PAGE_DIRTY)
882
        list_del_init(&page->in_dirty);
883 884 885 886 887
    page->state = PAGE_EMPTY;

    // XXX touch lru?
}

888 889 890 891 892 893 894 895
/* page_del deletes Page struct (but not page memory - see page_drop_memory).
 *
 * The page must be in PAGE_EMPTY state.
 * The page is removed from ram->lru.
 */
static void page_del(Page *page) {
    BUG_ON(page->refcnt != 0);
    BUG_ON(page->state != PAGE_EMPTY);
896
    BUG_ON(!list_empty(&page->in_dirty));
897 898 899 900 901 902

    list_del(&page->lru);
    bzero(page, sizeof(*page)); /* just in case */
    free(page);
}

903 904 905 906 907 908


/* vma: page -> addr  where it should-be mmaped in vma */
static void *vma_page_addr(VMA *vma, Page *page)
{
    uintptr_t addr;
909
    ASSERT(vma->fileh == page->fileh);
910 911 912 913 914 915 916 917 918 919 920 921 922 923 924

    addr = vma->addr_start + (page->f_pgoffset - vma->f_pgoffset) * page_size(page);
    ASSERT(vma->addr_start <= addr  &&
                              addr < vma->addr_stop);
    return (void *)addr;
}


/* vma: addr -> fileh pgoffset  with page containing addr */
static pgoff_t vma_addr_fpgoffset(VMA *vma, uintptr_t addr)
{
    return vma->f_pgoffset + (addr - vma->addr_start) / vma->fileh->ramh->ram->pagesize;
}


925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942
/* vma_mmap_page mmaps page into vma.
 *
 * the page must belong to covered file.
 * mmap protection is PROT_READ if page is PAGE_LOADED or PROT_READ|PROT_WRITE
 * if page is PAGE_DIRTY.
 *
 * must be called under virtmem lock.
 */
static void vma_mmap_page(VMA *vma, Page *page) {
    pgoff_t pgoff_invma;
    int prot = (page->state == PAGE_DIRTY ? PROT_READ|PROT_WRITE : PROT_READ);

    ASSERT(page->state == PAGE_LOADED || page->state == PAGE_DIRTY);
    ASSERT(vma->f_pgoffset <= page->f_pgoffset &&
                              page->f_pgoffset < vma_addr_fpgoffset(vma, vma->addr_stop));

    pgoff_invma = page->f_pgoffset - vma->f_pgoffset;
    if (!bitmap_test_bit(vma->page_ismappedv, pgoff_invma)) {
943 944
        void *mmapped = page_mmap(page, vma_page_addr(vma, page), prot);
        BUG_ON(!mmapped);
945 946 947 948 949 950 951 952
        bitmap_set_bit(vma->page_ismappedv, pgoff_invma);
        page_incref(page);
    }
    else {
        /* just changing protection bits should not fail, if parameters ok */
        xmprotect(vma_page_addr(vma, page), page_size(page), prot);
    }
}
953

954 955
/* is `page->fpgoffset` in file-range covered by `vma` */
static int vma_page_infilerange(VMA *vma, Page *page)
956 957 958 959 960
{
    pgoff_t vma_fpgstop;
    ASSERT(vma->fileh == page->fileh);

    vma_fpgstop = vma_addr_fpgoffset(vma, vma->addr_stop);
961 962 963 964 965 966 967 968 969 970
    return (vma->f_pgoffset <= page->f_pgoffset  &&
                               page->f_pgoffset < vma_fpgstop);
}

/* is `page` mapped to `vma` */
static int vma_page_ismapped(VMA *vma, Page *page)
{
    ASSERT(vma->fileh == page->fileh);

    if (!vma_page_infilerange(vma, page))
971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012
        return 0;

    return bitmap_test_bit(vma->page_ismappedv, page->f_pgoffset - vma->f_pgoffset);
}


/* ensure `page` is not mapped to `vma` */
static void vma_page_ensure_unmapped(VMA *vma, Page *page)
{
    if (!vma_page_ismapped(vma, page))
        return;

    /* mmap empty PROT_NONE address space instead of page memory */
    mem_xvalloc(vma_page_addr(vma, page), page_size(page));

    bitmap_clear_bit(vma->page_ismappedv, page->f_pgoffset - vma->f_pgoffset);
    page_decref(page);
}


/* ensure `page` is not mapped RW to `vma`
 *
 * if mapped -> should be mapped as R
 * if not mapped - leave as is
 */
static void vma_page_ensure_notmappedrw(VMA *vma, Page *page)
{
    if (!vma_page_ismapped(vma, page))
        return;

    /* just changing protection - should not fail */
    // XXX PROT_READ always? (it could be mmaped with PROT_NONE before without
    // first access) - then it should not be mapped in page_ismappedv -> ok.
    xmprotect(vma_page_addr(vma, page), page_size(page), PROT_READ);
}


// XXX stub
void OOM(void)
{
    BUG();
}