[t:3113] Merge from tokudb.3113, select latest header based on...

[t:3113] Merge from tokudb.3113, select latest header based on checkpoint_count, not checkpoint_lsn (file can be closed without intervening checkpoint).  Merge command in toku/tokudb was svn merge -r27478:HEAD ../tokudb.3113 .

git-svn-id: file:///svn/toku/tokudb@27500 c7de825b-a66e-492c-adef-691d508d4ae1

newbrt/brt-serialize.c

@@ -1749,78 +1749,66 @@ deserialize_brtheader_from_fd_into_rbuf(int fd, toku_off_t offset_of_header, str
 
 
 // Read brtheader from file into struct.  Read both headers and use one.
-// If a required_lsn is specified, then use the header with that lsn,
-// or if a header with that lsn is not available, then use the latest
-// header with an earlier lsn than required_lsn.  (There may not be a header
-// with required_lsn if the file did not change since the last checkpoint.)
-// If a required_lsn is not specified, then use the latest header.
+// We want the latest acceptable header whose checkpoint_lsn is no later
+// than max_acceptable_lsn.
 int 
-toku_deserialize_brtheader_from (int fd, LSN required_lsn, struct brt_header **brth) {
+toku_deserialize_brtheader_from (int fd, LSN max_acceptable_lsn, struct brt_header **brth) {
     struct rbuf rb_0;
     struct rbuf rb_1;
     u_int64_t checkpoint_count_0;
     u_int64_t checkpoint_count_1;
     LSN checkpoint_lsn_0;
     LSN checkpoint_lsn_1;
-    int r = 0;
-    int r0;
-    int r1;
     u_int32_t version_0, version_1, version = 0;
+    BOOL h0_acceptable = FALSE;
+    BOOL h1_acceptable = FALSE;
+    struct rbuf *rb = NULL;
+    int r0, r1, r;
 
     {
         toku_off_t header_0_off = 0;
         r0 = deserialize_brtheader_from_fd_into_rbuf(fd, header_0_off, &rb_0, &checkpoint_count_0, &checkpoint_lsn_0, &version_0);
+	if ( (r0==0) && (checkpoint_lsn_0.lsn <= max_acceptable_lsn.lsn) )
+	    h0_acceptable = TRUE;
     }
     {
         toku_off_t header_1_off = BLOCK_ALLOCATOR_HEADER_RESERVE;
         r1 = deserialize_brtheader_from_fd_into_rbuf(fd, header_1_off, &rb_1, &checkpoint_count_1, &checkpoint_lsn_1, &version_1);
+	if ( (r1==0) && (checkpoint_lsn_1.lsn <= max_acceptable_lsn.lsn) )
+	    h1_acceptable = TRUE;
     }
-    struct rbuf *rb = NULL;
 
+    // if either header is too new, the dictionary is unreadable
     if (r0!=TOKUDB_DICTIONARY_TOO_NEW && r1!=TOKUDB_DICTIONARY_TOO_NEW) {
-        if (r0==0) {
+	if (h0_acceptable && h1_acceptable) {
+	    if (checkpoint_count_0 > checkpoint_count_1) {
+		invariant(checkpoint_count_0 == checkpoint_count_1 + 1);
+		invariant(version_0 >= version_1);
 		rb = &rb_0;
 		version = version_0;
+		r = 0;
 	    }
-        if (r1==0 && (r0!=0 || checkpoint_count_1 > checkpoint_count_0)) {
+	    else {
+		invariant(checkpoint_count_1 == checkpoint_count_0 + 1);
+		invariant(version_1 >= version_0);
 		rb = &rb_1;
 		version = version_1;
+		r = 0;
 	    }
-        if (r0==0 && r1==0) {
-	    lazy_assert(checkpoint_count_1 != checkpoint_count_0);
-	    if (rb == &rb_0) lazy_assert(version_0 >= version_1);
-	    else lazy_assert(version_0 <= version_1);
-	}
-    }
-  
-    if (required_lsn.lsn != ZERO_LSN.lsn) {  // an upper bound lsn was requested
-	if ((r0==0 && checkpoint_lsn_0.lsn <= required_lsn.lsn) &&
-	    (r1 || checkpoint_lsn_1.lsn > required_lsn.lsn)) {
-	    rb = &rb_0;
-	    version = version_0;
-	}
-	else if ((r1==0 && checkpoint_lsn_1.lsn <= required_lsn.lsn) &&
-		 (r0 || checkpoint_lsn_0.lsn > required_lsn.lsn)) {
-	    rb = &rb_1;
-	    version = version_1;
 	}
-	else if (r0==0 && r1==0) {  // read two good headers and both have qualified lsn
-	    if (checkpoint_lsn_0.lsn > checkpoint_lsn_1.lsn) {
+	else if (h0_acceptable) {
 	    rb = &rb_0;
 	    version = version_0;
+	    r = 0;
 	}
-	    else {
+	else if (h1_acceptable) {
 	    rb = &rb_1;
 	    version = version_1;
-	    }
-	}
-	else {
-	    rb = NULL; 
-	    r = -1;  // may need new error code for unable to get required version of file
+	    r = 0;
 	}
     }
 
-    if (rb==NULL && r==0) {
+    if (rb==NULL) {
         // We were unable to read either header or at least one is too new.
         // Certain errors are higher priority than others. Order of these if/else if is important.
         if (r0==TOKUDB_DICTIONARY_TOO_NEW || r1==TOKUDB_DICTIONARY_TOO_NEW)
@@ -1831,8 +1819,12 @@ toku_deserialize_brtheader_from (int fd, LSN required_lsn, struct brt_header **b
         else if (r0==TOKUDB_DICTIONARY_NO_HEADER || r1==TOKUDB_DICTIONARY_NO_HEADER) {
             r = TOKUDB_DICTIONARY_NO_HEADER;
         }
-        else r = r0; //Arbitrarily report the error from the first header.
-        lazy_assert(r!=0);
+        else r = r0 ? r0 : r1; //Arbitrarily report the error from the first header, unless it's readable
+
+	// it should not be possible for both headers to be later than the max_acceptable_lsn
+	invariant(!( (r0==0 && checkpoint_lsn_0.lsn > max_acceptable_lsn.lsn) &&
+		     (r1==0 && checkpoint_lsn_1.lsn > max_acceptable_lsn.lsn) ));
+        invariant(r!=0);
     }
 
     if (r==0) r = deserialize_brtheader_versioned(fd, rb, brth, version);

newbrt/brt-test-helpers.c

@@ -8,7 +8,7 @@
 BOOL ignore_if_was_already_open;
 int toku_testsetup_leaf(BRT brt, BLOCKNUM *blocknum) {
     BRTNODE node;
-    int r = toku_read_brt_header_and_store_in_cachefile(brt->cf, ZERO_LSN, &brt->h, &ignore_if_was_already_open);
+    int r = toku_read_brt_header_and_store_in_cachefile(brt->cf, MAX_LSN, &brt->h, &ignore_if_was_already_open);
     if (r!=0) return r;
     toku_create_new_brtnode(brt, &node, 0, 0);
 
@@ -22,7 +22,7 @@ int toku_testsetup_leaf(BRT brt, BLOCKNUM *blocknum) {
 int toku_testsetup_nonleaf (BRT brt, int height, BLOCKNUM *blocknum, int n_children, BLOCKNUM *children, u_int32_t *subtree_fingerprints, char **keys, int *keylens) {
     BRTNODE node;
     assert(n_children<=BRT_FANOUT);
-    int r = toku_read_brt_header_and_store_in_cachefile(brt->cf, ZERO_LSN, &brt->h, &ignore_if_was_already_open);
+    int r = toku_read_brt_header_and_store_in_cachefile(brt->cf, MAX_LSN, &brt->h, &ignore_if_was_already_open);
     if (r!=0) return r;
     toku_create_new_brtnode(brt, &node, height, 0);
     node->u.n.n_children=n_children;
@@ -47,7 +47,7 @@ int toku_testsetup_nonleaf (BRT brt, int height, BLOCKNUM *blocknum, int n_child
 }
 
 int toku_testsetup_root(BRT brt, BLOCKNUM blocknum) {
-    int r = toku_read_brt_header_and_store_in_cachefile(brt->cf, ZERO_LSN, &brt->h, &ignore_if_was_already_open);
+    int r = toku_read_brt_header_and_store_in_cachefile(brt->cf, MAX_LSN, &brt->h, &ignore_if_was_already_open);
     if (r!=0) return r;
     brt->h->root = blocknum;
     brt->h->root_hash.valid = FALSE;

newbrt/brt.c

@@ -3156,10 +3156,10 @@ int toku_brt_alloc_init_header(BRT t, TOKUTXN txn) {
     return r;
 }
 
-int toku_read_brt_header_and_store_in_cachefile (CACHEFILE cf, LSN required_lsn, struct brt_header **header, BOOL* was_open)
+int toku_read_brt_header_and_store_in_cachefile (CACHEFILE cf, LSN max_acceptable_lsn, struct brt_header **header, BOOL* was_open)
 // If the cachefile already has the header, then just get it.
 // If the cachefile has not been initialized, then don't modify anything.
-// required_lsn is used if a particular version of the file is required based on checkpoint_lsn
+// max_acceptable_lsn is the latest acceptable checkpointed version of the file.
 {
     {
         struct brt_header *h;
@@ -3174,7 +3174,7 @@ int toku_read_brt_header_and_store_in_cachefile (CACHEFILE cf, LSN required_lsn,
     int r;
     {
         int fd = toku_cachefile_get_and_pin_fd (cf);
-        r = toku_deserialize_brtheader_from(fd, required_lsn, &h);
+        r = toku_deserialize_brtheader_from(fd, max_acceptable_lsn, &h);
         toku_cachefile_unpin_fd(cf);
     }
     if (r!=0) return r;
@@ -3273,10 +3273,10 @@ cleanup:
 }
 
 // This is the actual open, used for various purposes, such as normal use, recovery, and redirect.  
-// fname_in_env is the iname, relative to the env_dir  (data_dir is already in iname as prefix)
-// If required_lsn is specified (not ZERO_LSN), then that specific checkpointed version is required.
+// 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 .
 static int
-brt_open(BRT t, const char *fname_in_env, int is_create, int only_create, CACHETABLE cachetable, TOKUTXN txn, DB *db, FILENUM use_filenum, DICTIONARY_ID use_dictionary_id, LSN required_lsn) {
+brt_open(BRT t, const char *fname_in_env, int is_create, int only_create, CACHETABLE cachetable, TOKUTXN txn, DB *db, FILENUM use_filenum, DICTIONARY_ID use_dictionary_id, LSN max_acceptable_lsn) {
     int r;
     BOOL txn_created = FALSE;
 
@@ -3342,7 +3342,7 @@ brt_open(BRT t, const char *fname_in_env, int is_create, int only_create, CACHET
     }
     BOOL was_already_open;
     if (is_create) {
-        r = toku_read_brt_header_and_store_in_cachefile(t->cf, required_lsn, &t->h, &was_already_open);
+        r = toku_read_brt_header_and_store_in_cachefile(t->cf, max_acceptable_lsn, &t->h, &was_already_open);
         if (r==TOKUDB_DICTIONARY_NO_HEADER) {
             r = toku_brt_alloc_init_header(t, txn);
             if (r != 0) goto died_after_read_and_pin;
@@ -3357,7 +3357,7 @@ brt_open(BRT t, const char *fname_in_env, int is_create, int only_create, CACHET
         }
         else goto found_it;
     } else {
-        if ((r = toku_read_brt_header_and_store_in_cachefile(t->cf, required_lsn, &t->h, &was_already_open))!=0) goto died_after_open;
+        if ((r = toku_read_brt_header_and_store_in_cachefile(t->cf, max_acceptable_lsn, &t->h, &was_already_open))!=0) goto died_after_open;
         found_it:
         t->nodesize = t->h->nodesize;                 /* inherit the pagesize from the file */
         if (!t->did_set_flags) {
@@ -3433,11 +3433,11 @@ brt_open(BRT t, const char *fname_in_env, int is_create, int only_create, CACHET
 // (dict_id is assigned by the brt_open() function.)
 int
 toku_brt_open_recovery(BRT t, const char *fname_in_env, int is_create, int only_create, CACHETABLE cachetable, TOKUTXN txn, 
-		       DB *db, FILENUM use_filenum, LSN required_lsn) {
+		       DB *db, FILENUM use_filenum, LSN max_acceptable_lsn) {
     int r;
     lazy_assert(use_filenum.fileid != FILENUM_NONE.fileid);
     r = brt_open(t, fname_in_env, is_create, only_create, cachetable,
-                 txn, db, use_filenum, DICTIONARY_ID_NONE, required_lsn);
+                 txn, db, use_filenum, DICTIONARY_ID_NONE, max_acceptable_lsn);
     return r;
 }
 
@@ -3445,7 +3445,7 @@ toku_brt_open_recovery(BRT t, const char *fname_in_env, int is_create, int only_
 int
 toku_brt_open(BRT t, const char *fname_in_env, int is_create, int only_create, CACHETABLE cachetable, TOKUTXN txn, DB *db) {
     int r;
-    r = brt_open(t, fname_in_env, is_create, only_create, cachetable, txn, db, FILENUM_NONE, DICTIONARY_ID_NONE, ZERO_LSN);
+    r = brt_open(t, fname_in_env, is_create, only_create, cachetable, txn, db, FILENUM_NONE, DICTIONARY_ID_NONE, MAX_LSN);
     return r;
 }
 
@@ -3467,7 +3467,7 @@ brt_open_for_redirect(BRT *new_brtp, const char *fname_in_env, TOKUTXN txn, BRT
         lazy_assert_zero(r);
     }
     CACHETABLE ct = toku_cachefile_get_cachetable(old_brt->cf);
-    r = brt_open(t, fname_in_env, 0, 0, ct, txn, old_brt->db, FILENUM_NONE, old_h->dict_id, ZERO_LSN);
+    r = brt_open(t, fname_in_env, 0, 0, ct, txn, old_brt->db, FILENUM_NONE, old_h->dict_id, MAX_LSN);
     lazy_assert_zero(r);
     if (old_h->descriptor.version==0) {
         lazy_assert(t->h->descriptor.version == 0);

newbrt/brtdump.c

@@ -76,7 +76,7 @@ static void
 dump_header (int f, struct brt_header **header, CACHEFILE cf) {
     struct brt_header *h;
     int r;
-    r = toku_deserialize_brtheader_from (f, ZERO_LSN, &h);
+    r = toku_deserialize_brtheader_from (f, MAX_LSN, &h);
     assert(r==0);
     h->cf = cf;
     printf("brtheader:\n");

newbrt/recover.c

@@ -238,7 +238,7 @@ static void recover_yield(voidfp f, void *fpthunk, void *UU(yieldthunk)) {
 
 // Open the file if it is not already open.  If it is already open, then do nothing.
 static int internal_recover_fopen_or_fcreate (RECOVER_ENV renv, BOOL must_create, int mode, BYTESTRING *bs_iname, FILENUM filenum, u_int32_t treeflags, 
-                                              u_int32_t descriptor_version, BYTESTRING* descriptor, TOKUTXN txn, uint32_t nodesize, LSN required_lsn) {
+                                              u_int32_t descriptor_version, BYTESTRING* descriptor, TOKUTXN txn, uint32_t nodesize, LSN max_acceptable_lsn) {
     int r;
     char *iname = fixup_fname(bs_iname);
 
@@ -271,7 +271,7 @@ static int internal_recover_fopen_or_fcreate (RECOVER_ENV renv, BOOL must_create
         r = toku_brt_set_descriptor(brt, descriptor_version, &descriptor_dbt);
         if (r != 0) goto close_brt;
     }
-    r = toku_brt_open_recovery(brt, iname, must_create, must_create, renv->ct, txn, fake_db, filenum, required_lsn);
+    r = toku_brt_open_recovery(brt, iname, must_create, must_create, renv->ct, txn, fake_db, filenum, max_acceptable_lsn);
     if (r != 0) {
     close_brt:
         ;
@@ -409,14 +409,15 @@ static int toku_recover_fassociate (struct logtype_fassociate *l, RECOVER_ENV re
     case FORWARD_BETWEEN_CHECKPOINT_BEGIN_END:
 	renv->ss.checkpoint_num_fassociate++;
         assert(r==DB_NOTFOUND); //Not open
-        // open it if it exists
-	// if rollback file, specify which checkpointed version of file we need (not just the latest)
+        // Open it if it exists.
+	// If rollback file, specify which checkpointed version of file we need (not just the latest)
+	// because we cannot use a rollback log that is later than the last complete checkpoint.  See #3113.
 	{
 	    BOOL rollback_file = !strcmp(fname, ROLLBACK_CACHEFILE_NAME);
-	    LSN required_lsn = ZERO_LSN;
+	    LSN max_acceptable_lsn = MAX_LSN;
 	    if (rollback_file)
-		required_lsn = renv->ss.checkpoint_begin_lsn;
-	    r = internal_recover_fopen_or_fcreate(renv, FALSE, 0, &l->iname, l->filenum, l->treeflags, 0, NULL, NULL, 0, required_lsn);
+		max_acceptable_lsn = renv->ss.checkpoint_begin_lsn;
+	    r = internal_recover_fopen_or_fcreate(renv, FALSE, 0, &l->iname, l->filenum, l->treeflags, 0, NULL, NULL, 0, max_acceptable_lsn);
 	    if (r==0 && rollback_file) {
 		//Load rollback cachefile
 		r = file_map_find(&renv->fmap, l->filenum, &tuple);
@@ -645,7 +646,7 @@ static int toku_recover_fcreate (struct logtype_fcreate *l, RECOVER_ENV renv) {
     toku_free(iname);
 
     BOOL must_create = TRUE;
-    r = internal_recover_fopen_or_fcreate(renv, must_create, l->mode, &l->iname, l->filenum, l->treeflags, l->descriptor_version, &l->descriptor, txn, l->nodesize, ZERO_LSN);
+    r = internal_recover_fopen_or_fcreate(renv, must_create, l->mode, &l->iname, l->filenum, l->treeflags, l->descriptor_version, &l->descriptor, txn, l->nodesize, MAX_LSN);
     return r;
 }
 
@@ -672,7 +673,7 @@ static int toku_recover_fopen (struct logtype_fopen *l, RECOVER_ENV renv) {
 
     if (strcmp(fname, ROLLBACK_CACHEFILE_NAME)) {
         //Rollback cachefile can only be opened via fassociate.
-        r = internal_recover_fopen_or_fcreate(renv, must_create, 0, &l->iname, l->filenum, l->treeflags, descriptor_version, descriptor, txn, 0, ZERO_LSN);
+        r = internal_recover_fopen_or_fcreate(renv, must_create, 0, &l->iname, l->filenum, l->treeflags, descriptor_version, descriptor, txn, 0, MAX_LSN);
     }
     toku_free(fname);
     return r;