]> git.ozlabs.org Git - ccan/blobdiff - ccan/tdb2/transaction.c
tdb2: overallocate the recovery area.
[ccan] / ccan / tdb2 / transaction.c
index 73ab1e39ff4ad0c4f260be0bf1dc6d9eca95bcd2..a27c027365eec3fc7be001d9693bde42ec6f0b45 100644 (file)
@@ -110,7 +110,7 @@ struct tdb_transaction {
        /* when inside a transaction we need to keep track of any
           nested tdb_transaction_start() calls, as these are allowed,
           but don't create a new transaction */
-       int nesting;
+       unsigned int nesting;
 
        /* set when a prepare has already occurred */
        bool prepared;
@@ -120,6 +120,8 @@ struct tdb_transaction {
        tdb_len_t old_map_size;
 };
 
+/* This doesn't really need to be pagesize, but we use it for similar reasons. */
+#define PAGESIZE 4096
 
 /*
   read while in a transaction. We need to check first if the data is in our list
@@ -132,8 +134,8 @@ static enum TDB_ERROR transaction_read(struct tdb_context *tdb, tdb_off_t off,
        enum TDB_ERROR ecode;
 
        /* break it down into block sized ops */
-       while (len + (off % getpagesize()) > getpagesize()) {
-               tdb_len_t len2 = getpagesize() - (off % getpagesize());
+       while (len + (off % PAGESIZE) > PAGESIZE) {
+               tdb_len_t len2 = PAGESIZE - (off % PAGESIZE);
                ecode = transaction_read(tdb, off, buf, len2);
                if (ecode != TDB_SUCCESS) {
                        return ecode;
@@ -147,7 +149,7 @@ static enum TDB_ERROR transaction_read(struct tdb_context *tdb, tdb_off_t off,
                return TDB_SUCCESS;
        }
 
-       blk = off / getpagesize();
+       blk = off / PAGESIZE;
 
        /* see if we have it in the block list */
        if (tdb->transaction->num_blocks <= blk ||
@@ -169,7 +171,7 @@ static enum TDB_ERROR transaction_read(struct tdb_context *tdb, tdb_off_t off,
        }
 
        /* now copy it out of this block */
-       memcpy(buf, tdb->transaction->blocks[blk] + (off % getpagesize()), len);
+       memcpy(buf, tdb->transaction->blocks[blk] + (off % PAGESIZE), len);
        return TDB_SUCCESS;
 
 fail:
@@ -198,8 +200,8 @@ static enum TDB_ERROR transaction_write(struct tdb_context *tdb, tdb_off_t off,
        }
 
        /* break it up into block sized chunks */
-       while (len + (off % getpagesize()) > getpagesize()) {
-               tdb_len_t len2 = getpagesize() - (off % getpagesize());
+       while (len + (off % PAGESIZE) > PAGESIZE) {
+               tdb_len_t len2 = PAGESIZE - (off % PAGESIZE);
                ecode = transaction_write(tdb, off, buf, len2);
                if (ecode != TDB_SUCCESS) {
                        return -1;
@@ -215,8 +217,8 @@ static enum TDB_ERROR transaction_write(struct tdb_context *tdb, tdb_off_t off,
                return TDB_SUCCESS;
        }
 
-       blk = off / getpagesize();
-       off = off % getpagesize();
+       blk = off / PAGESIZE;
+       off = off % PAGESIZE;
 
        if (tdb->transaction->num_blocks <= blk) {
                uint8_t **new_blocks;
@@ -244,20 +246,20 @@ static enum TDB_ERROR transaction_write(struct tdb_context *tdb, tdb_off_t off,
 
        /* allocate and fill a block? */
        if (tdb->transaction->blocks[blk] == NULL) {
-               tdb->transaction->blocks[blk] = (uint8_t *)calloc(getpagesize(), 1);
+               tdb->transaction->blocks[blk] = (uint8_t *)calloc(PAGESIZE, 1);
                if (tdb->transaction->blocks[blk] == NULL) {
                        ecode = tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
                                           "transaction_write:"
                                           " failed to allocate");
                        goto fail;
                }
-               if (tdb->transaction->old_map_size > blk * getpagesize()) {
-                       tdb_len_t len2 = getpagesize();
-                       if (len2 + (blk * getpagesize()) > tdb->transaction->old_map_size) {
-                               len2 = tdb->transaction->old_map_size - (blk * getpagesize());
+               if (tdb->transaction->old_map_size > blk * PAGESIZE) {
+                       tdb_len_t len2 = PAGESIZE;
+                       if (len2 + (blk * PAGESIZE) > tdb->transaction->old_map_size) {
+                               len2 = tdb->transaction->old_map_size - (blk * PAGESIZE);
                        }
                        ecode = tdb->transaction->io_methods->tread(tdb,
-                                       blk * getpagesize(),
+                                       blk * PAGESIZE,
                                        tdb->transaction->blocks[blk],
                                        len2);
                        if (ecode != TDB_SUCCESS) {
@@ -306,8 +308,8 @@ static void transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
        size_t blk;
 
        /* break it up into block sized chunks */
-       while (len + (off % getpagesize()) > getpagesize()) {
-               tdb_len_t len2 = getpagesize() - (off % getpagesize());
+       while (len + (off % PAGESIZE) > PAGESIZE) {
+               tdb_len_t len2 = PAGESIZE - (off % PAGESIZE);
                transaction_write_existing(tdb, off, buf, len2);
                len -= len2;
                off += len2;
@@ -320,8 +322,8 @@ static void transaction_write_existing(struct tdb_context *tdb, tdb_off_t off,
                return;
        }
 
-       blk = off / getpagesize();
-       off = off % getpagesize();
+       blk = off / PAGESIZE;
+       off = off % PAGESIZE;
 
        if (tdb->transaction->num_blocks <= blk ||
            tdb->transaction->blocks[blk] == NULL) {
@@ -379,10 +381,10 @@ static enum TDB_ERROR transaction_expand_file(struct tdb_context *tdb,
 static void *transaction_direct(struct tdb_context *tdb, tdb_off_t off,
                                size_t len, bool write_mode)
 {
-       size_t blk = off / getpagesize(), end_blk;
+       size_t blk = off / PAGESIZE, end_blk;
 
        /* This is wrong for zero-length blocks, but will fail gracefully */
-       end_blk = (off + len - 1) / getpagesize();
+       end_blk = (off + len - 1) / PAGESIZE;
 
        /* Can only do direct if in single block and we've already copied. */
        if (write_mode) {
@@ -392,17 +394,17 @@ static void *transaction_direct(struct tdb_context *tdb, tdb_off_t off,
                        return NULL;
                if (tdb->transaction->blocks[blk] == NULL)
                        return NULL;
-               return tdb->transaction->blocks[blk] + off % getpagesize();
+               return tdb->transaction->blocks[blk] + off % PAGESIZE;
        }
 
        /* Single which we have copied? */
        if (blk == end_blk
            && blk < tdb->transaction->num_blocks
            && tdb->transaction->blocks[blk])
-               return tdb->transaction->blocks[blk] + off % getpagesize();
+               return tdb->transaction->blocks[blk] + off % PAGESIZE;
 
        /* Otherwise must be all not copied. */
-       while (blk < end_blk) {
+       while (blk <= end_blk) {
                if (blk >= tdb->transaction->num_blocks)
                        break;
                if (tdb->transaction->blocks[blk])
@@ -437,7 +439,7 @@ static enum TDB_ERROR transaction_sync(struct tdb_context *tdb,
        }
 #ifdef MS_SYNC
        if (tdb->file->map_ptr) {
-               tdb_off_t moffset = offset & ~(getpagesize()-1);
+               tdb_off_t moffset = offset & ~(PAGESIZE-1);
                if (msync(moffset + (char *)tdb->file->map_ptr,
                          length + (offset - moffset), MS_SYNC) != 0) {
                        return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_ERROR,
@@ -519,32 +521,46 @@ enum TDB_ERROR tdb_transaction_start(struct tdb_context *tdb)
 
        /* some sanity checks */
        if (tdb->read_only || (tdb->flags & TDB_INTERNAL)) {
-               return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
-                                 "tdb_transaction_start: cannot start a"
-                                 " transaction on a read-only or internal db");
+               return tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL,
+                                                   TDB_LOG_USE_ERROR,
+                                                   "tdb_transaction_start:"
+                                                   " cannot start a"
+                                                   " transaction on a "
+                                                   "read-only or internal db");
        }
 
        /* cope with nested tdb_transaction_start() calls */
        if (tdb->transaction != NULL) {
-               return tdb_logerr(tdb, TDB_ERR_IO, TDB_LOG_USE_ERROR,
-                                 "tdb_transaction_start:"
-                                 " already inside transaction");
+               if (!(tdb->flags & TDB_ALLOW_NESTING)) {
+                       return tdb->last_error
+                               = tdb_logerr(tdb, TDB_ERR_IO,
+                                            TDB_LOG_USE_ERROR,
+                                            "tdb_transaction_start:"
+                                            " already inside transaction");
+               }
+               tdb->transaction->nesting++;
+               return 0;
        }
 
        if (tdb_has_hash_locks(tdb)) {
                /* the caller must not have any locks when starting a
                   transaction as otherwise we'll be screwed by lack
                   of nested locks in POSIX */
-               return tdb_logerr(tdb, TDB_ERR_LOCK, TDB_LOG_USE_ERROR,
-                                 "tdb_transaction_start: cannot start a"
-                                 " transaction with locks held");
+               return tdb->last_error = tdb_logerr(tdb, TDB_ERR_LOCK,
+                                                   TDB_LOG_USE_ERROR,
+                                                   "tdb_transaction_start:"
+                                                   " cannot start a"
+                                                   " transaction with locks"
+                                                   " held");
        }
 
        tdb->transaction = (struct tdb_transaction *)
                calloc(sizeof(struct tdb_transaction), 1);
        if (tdb->transaction == NULL) {
-               return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
-                                 "tdb_transaction_start: cannot allocate");
+               return tdb->last_error = tdb_logerr(tdb, TDB_ERR_OOM,
+                                                   TDB_LOG_ERROR,
+                                                   "tdb_transaction_start:"
+                                                   " cannot allocate");
        }
 
        /* get the transaction write lock. This is a blocking lock. As
@@ -554,7 +570,7 @@ enum TDB_ERROR tdb_transaction_start(struct tdb_context *tdb)
        if (ecode != TDB_SUCCESS) {
                SAFE_FREE(tdb->transaction->blocks);
                SAFE_FREE(tdb->transaction);
-               return ecode;
+               return tdb->last_error = ecode;
        }
 
        /* get a read lock over entire file. This is upgraded to a write
@@ -573,13 +589,13 @@ enum TDB_ERROR tdb_transaction_start(struct tdb_context *tdb)
           transaction specific methods */
        tdb->transaction->io_methods = tdb->methods;
        tdb->methods = &transaction_methods;
-       return TDB_SUCCESS;
+       return tdb->last_error = TDB_SUCCESS;
 
 fail_allrecord_lock:
        tdb_transaction_unlock(tdb, F_WRLCK);
        SAFE_FREE(tdb->transaction->blocks);
        SAFE_FREE(tdb->transaction);
-       return ecode;
+       return tdb->last_error = ecode;
 }
 
 
@@ -601,7 +617,7 @@ static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
 
        recovery_size = sizeof(tdb_len_t);
        for (i=0;i<tdb->transaction->num_blocks;i++) {
-               if (i * getpagesize() >= tdb->transaction->old_map_size) {
+               if (i * PAGESIZE >= tdb->transaction->old_map_size) {
                        break;
                }
                if (tdb->transaction->blocks[i] == NULL) {
@@ -611,7 +627,7 @@ static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
                if (i == tdb->transaction->num_blocks-1) {
                        recovery_size += tdb->transaction->last_block_size;
                } else {
-                       recovery_size += getpagesize();
+                       recovery_size += PAGESIZE;
                }
        }
 
@@ -670,9 +686,10 @@ static enum TDB_ERROR tdb_recovery_allocate(struct tdb_context *tdb,
           us an area that is being currently used (as of the start of
           the transaction) */
        if (recovery_head != 0) {
-               add_stat(tdb, frees, 1);
+               tdb->stats.frees++;
                ecode = add_free_record(tdb, recovery_head,
-                                       sizeof(rec) + rec.max_len);
+                                       sizeof(rec) + rec.max_len,
+                                       TDB_LOCK_WAIT);
                if (ecode != TDB_SUCCESS) {
                        return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
                                          "tdb_recovery_allocate:"
@@ -684,10 +701,11 @@ static enum TDB_ERROR tdb_recovery_allocate(struct tdb_context *tdb,
        /* the tdb_free() call might have increased the recovery size */
        *recovery_size = tdb_recovery_size(tdb);
 
-       /* round up to a multiple of page size */
+       /* round up to a multiple of page size. Overallocate, since each
+        * such allocation forces us to expand the file. */
        *recovery_max_size
-               = (((sizeof(rec) + *recovery_size) + getpagesize()-1)
-                  & ~(getpagesize()-1))
+               = (((sizeof(rec) + *recovery_size + *recovery_size / 2)
+                   + PAGESIZE-1) & ~(PAGESIZE-1))
                - sizeof(rec);
        *recovery_offset = tdb->file->map_size;
        recovery_head = *recovery_offset;
@@ -788,8 +806,8 @@ static enum TDB_ERROR transaction_setup_recovery(struct tdb_context *tdb,
                        continue;
                }
 
-               offset = i * getpagesize();
-               length = getpagesize();
+               offset = i * PAGESIZE;
+               length = PAGESIZE;
                if (i == tdb->transaction->num_blocks-1) {
                        length = tdb->transaction->last_block_size;
                }
@@ -895,7 +913,6 @@ static enum TDB_ERROR _tdb_transaction_prepare_commit(struct tdb_context *tdb)
 
 
        if (tdb->transaction->nesting != 0) {
-               tdb->transaction->nesting--;
                return TDB_SUCCESS;
        }
 
@@ -909,10 +926,6 @@ static enum TDB_ERROR _tdb_transaction_prepare_commit(struct tdb_context *tdb)
        /* upgrade the main transaction lock region to a write lock */
        ecode = tdb_allrecord_upgrade(tdb);
        if (ecode != TDB_SUCCESS) {
-               tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
-                        "tdb_transaction_prepare_commit:"
-                        " failed to upgrade hash locks");
-               _tdb_transaction_cancel(tdb);
                return ecode;
        }
 
@@ -920,10 +933,6 @@ static enum TDB_ERROR _tdb_transaction_prepare_commit(struct tdb_context *tdb)
           during the commit */
        ecode = tdb_lock_open(tdb, TDB_LOCK_WAIT|TDB_LOCK_NOCHECK);
        if (ecode != TDB_SUCCESS) {
-               tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
-                          "tdb_transaction_prepare_commit:"
-                          " failed to get open lock");
-               _tdb_transaction_cancel(tdb);
                return ecode;
        }
 
@@ -934,10 +943,6 @@ static enum TDB_ERROR _tdb_transaction_prepare_commit(struct tdb_context *tdb)
                                                   &tdb->transaction
                                                   ->magic_offset);
                if (ecode != TDB_SUCCESS) {
-                       tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
-                                "tdb_transaction_prepare_commit:"
-                                " failed to setup recovery data");
-                       _tdb_transaction_cancel(tdb);
                        return ecode;
                }
        }
@@ -953,10 +958,6 @@ static enum TDB_ERROR _tdb_transaction_prepare_commit(struct tdb_context *tdb)
                tdb->file->map_size = tdb->transaction->old_map_size;
                ecode = methods->expand_file(tdb, add);
                if (ecode != TDB_SUCCESS) {
-                       tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
-                                "tdb_transaction_prepare_commit:"
-                                " expansion failed");
-                       _tdb_transaction_cancel(tdb);
                        return ecode;
                }
        }
@@ -983,27 +984,31 @@ enum TDB_ERROR tdb_transaction_commit(struct tdb_context *tdb)
        enum TDB_ERROR ecode;
 
        if (tdb->transaction == NULL) {
-               return tdb_logerr(tdb, TDB_ERR_EINVAL, TDB_LOG_USE_ERROR,
-                                 "tdb_transaction_commit: no transaction");
+               return tdb->last_error = tdb_logerr(tdb, TDB_ERR_EINVAL,
+                                                   TDB_LOG_USE_ERROR,
+                                                   "tdb_transaction_commit:"
+                                                   " no transaction");
        }
 
        tdb_trace(tdb, "tdb_transaction_commit");
 
        if (tdb->transaction->nesting != 0) {
                tdb->transaction->nesting--;
-               return TDB_SUCCESS;
+               return tdb->last_error = TDB_SUCCESS;
        }
 
        /* check for a null transaction */
        if (tdb->transaction->blocks == NULL) {
                _tdb_transaction_cancel(tdb);
-               return TDB_SUCCESS;
+               return tdb->last_error = TDB_SUCCESS;
        }
 
        if (!tdb->transaction->prepared) {
                ecode = _tdb_transaction_prepare_commit(tdb);
-               if (ecode != TDB_SUCCESS)
-                       return ecode;
+               if (ecode != TDB_SUCCESS) {
+                       _tdb_transaction_cancel(tdb);
+                       return tdb->last_error = ecode;
+               }
        }
 
        methods = tdb->transaction->io_methods;
@@ -1017,8 +1022,8 @@ enum TDB_ERROR tdb_transaction_commit(struct tdb_context *tdb)
                        continue;
                }
 
-               offset = i * getpagesize();
-               length = getpagesize();
+               offset = i * PAGESIZE;
+               length = PAGESIZE;
                if (i == tdb->transaction->num_blocks-1) {
                        length = tdb->transaction->last_block_size;
                }
@@ -1026,10 +1031,6 @@ enum TDB_ERROR tdb_transaction_commit(struct tdb_context *tdb)
                ecode = methods->twrite(tdb, offset,
                                        tdb->transaction->blocks[i], length);
                if (ecode != TDB_SUCCESS) {
-                       tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
-                                  "tdb_transaction_commit:"
-                                  " write failed during commit");
-
                        /* we've overwritten part of the data and
                           possibly expanded the file, so we need to
                           run the crash recovery code */
@@ -1038,7 +1039,7 @@ enum TDB_ERROR tdb_transaction_commit(struct tdb_context *tdb)
 
                        _tdb_transaction_cancel(tdb);
 
-                       return ecode;
+                       return tdb->last_error = ecode;
                }
                SAFE_FREE(tdb->transaction->blocks[i]);
        }
@@ -1049,7 +1050,7 @@ enum TDB_ERROR tdb_transaction_commit(struct tdb_context *tdb)
        /* ensure the new data is on disk */
        ecode = transaction_sync(tdb, 0, tdb->file->map_size);
        if (ecode != TDB_SUCCESS) {
-               return ecode;
+               return tdb->last_error = ecode;
        }
 
        /*
@@ -1068,10 +1069,11 @@ enum TDB_ERROR tdb_transaction_commit(struct tdb_context *tdb)
 #endif
 
        /* use a transaction cancel to free memory and remove the
-          transaction locks */
+          transaction locks: it "restores" map_size, too. */
+       tdb->transaction->old_map_size = tdb->file->map_size;
        _tdb_transaction_cancel(tdb);
 
-       return TDB_SUCCESS;
+       return tdb->last_error = TDB_SUCCESS;
 }