fsync/msync calls are made.
*/
-
/*
hold the context of any current transaction
*/
};
/* This doesn't really need to be pagesize, but we use it for similar reasons. */
-#define PAGESIZE 4096
+#define PAGESIZE 65536
/*
read while in a transaction. We need to check first if the data is in our list
}
/*
- work out how much space the linearised recovery data will consume
+ work out how much space the linearised recovery data will consume (worst case)
*/
static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
{
return recovery_size;
}
-/*
- allocate the recovery area, or use an existing recovery area if it is
- large enough
-*/
-static enum TDB_ERROR tdb_recovery_allocate(struct tdb_context *tdb,
- tdb_len_t *recovery_size,
- tdb_off_t *recovery_offset,
- tdb_len_t *recovery_max_size)
+static enum TDB_ERROR tdb_recovery_area(struct tdb_context *tdb,
+ const struct tdb_methods *methods,
+ tdb_off_t *recovery_offset,
+ struct tdb_recovery_record *rec)
{
- struct tdb_recovery_record rec;
- const struct tdb_methods *methods = tdb->transaction->io_methods;
- tdb_off_t recovery_head;
- size_t addition;
enum TDB_ERROR ecode;
- recovery_head = tdb_read_off(tdb, offsetof(struct tdb_header,recovery));
- if (TDB_OFF_IS_ERR(recovery_head)) {
- return tdb_logerr(tdb, recovery_head, TDB_LOG_ERROR,
- "tdb_recovery_allocate:"
- " failed to read recovery head");
+ *recovery_offset = tdb_read_off(tdb,
+ offsetof(struct tdb_header, recovery));
+ if (TDB_OFF_IS_ERR(*recovery_offset)) {
+ return *recovery_offset;
}
- if (recovery_head != 0) {
- ecode = methods->tread(tdb, recovery_head, &rec, sizeof(rec));
- if (ecode != TDB_SUCCESS) {
- return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
- "tdb_recovery_allocate:"
- " failed to read recovery record");
- }
- tdb_convert(tdb, &rec, sizeof(rec));
- /* ignore invalid recovery regions: can happen in crash */
- if (rec.magic != TDB_RECOVERY_MAGIC &&
- rec.magic != TDB_RECOVERY_INVALID_MAGIC) {
- recovery_head = 0;
+ if (*recovery_offset == 0) {
+ rec->max_len = 0;
+ return TDB_SUCCESS;
+ }
+
+ ecode = methods->tread(tdb, *recovery_offset, rec, sizeof(*rec));
+ if (ecode != TDB_SUCCESS)
+ return ecode;
+
+ tdb_convert(tdb, rec, sizeof(*rec));
+ /* ignore invalid recovery regions: can happen in crash */
+ if (rec->magic != TDB_RECOVERY_MAGIC &&
+ rec->magic != TDB_RECOVERY_INVALID_MAGIC) {
+ *recovery_offset = 0;
+ rec->max_len = 0;
+ }
+ return TDB_SUCCESS;
+}
+
+static unsigned int same(const unsigned char *new,
+ const unsigned char *old,
+ unsigned int length)
+{
+ unsigned int i;
+
+ for (i = 0; i < length; i++) {
+ if (new[i] != old[i])
+ break;
+ }
+ return i;
+}
+
+static unsigned int different(const unsigned char *new,
+ const unsigned char *old,
+ unsigned int length,
+ unsigned int min_same,
+ unsigned int *samelen)
+{
+ unsigned int i;
+
+ *samelen = 0;
+ for (i = 0; i < length; i++) {
+ if (new[i] == old[i]) {
+ (*samelen)++;
+ } else {
+ if (*samelen >= min_same) {
+ return i - *samelen;
+ }
+ *samelen = 0;
}
}
- *recovery_size = tdb_recovery_size(tdb);
+ if (*samelen < min_same)
+ *samelen = 0;
+ return length - *samelen;
+}
- if (recovery_head != 0 && *recovery_size <= rec.max_len) {
- /* it fits in the existing area */
- *recovery_max_size = rec.max_len;
- *recovery_offset = recovery_head;
- return TDB_SUCCESS;
+/* Allocates recovery blob, without tdb_recovery_record at head set up. */
+static struct tdb_recovery_record *alloc_recovery(struct tdb_context *tdb,
+ tdb_len_t *len)
+{
+ struct tdb_recovery_record *rec;
+ size_t i;
+ enum TDB_ERROR ecode;
+ unsigned char *p;
+ const struct tdb_methods *old_methods = tdb->methods;
+
+ rec = malloc(sizeof(*rec) + tdb_recovery_size(tdb));
+ if (!rec) {
+ tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
+ "transaction_setup_recovery:"
+ " cannot allocate");
+ return TDB_ERR_PTR(TDB_ERR_OOM);
}
- /* we need to free up the old recovery area, then allocate a
- new one at the end of the file. Note that we cannot use
- normal allocation to allocate the new one as that might return
- us an area that is being currently used (as of the start of
- the transaction) */
- if (recovery_head != 0) {
- tdb->stats.frees++;
- ecode = add_free_record(tdb, recovery_head,
- sizeof(rec) + rec.max_len,
- TDB_LOCK_WAIT, true);
- if (ecode != TDB_SUCCESS) {
- return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
- "tdb_recovery_allocate:"
- " failed to free previous"
- " recovery area");
+ /* We temporarily revert to the old I/O methods, so we can use
+ * tdb_access_read */
+ tdb->methods = tdb->transaction->io_methods;
+
+ /* build the recovery data into a single blob to allow us to do a single
+ large write, which should be more efficient */
+ p = (unsigned char *)(rec + 1);
+ for (i=0;i<tdb->transaction->num_blocks;i++) {
+ tdb_off_t offset;
+ tdb_len_t length;
+ unsigned int off;
+ const unsigned char *buffer;
+
+ if (tdb->transaction->blocks[i] == NULL) {
+ continue;
+ }
+
+ offset = i * PAGESIZE;
+ length = PAGESIZE;
+ if (i == tdb->transaction->num_blocks-1) {
+ length = tdb->transaction->last_block_size;
+ }
+
+ if (offset >= tdb->transaction->old_map_size) {
+ continue;
+ }
+
+ if (offset + length > tdb->file->map_size) {
+ ecode = tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
+ "tdb_transaction_setup_recovery:"
+ " transaction data over new region"
+ " boundary");
+ goto fail;
+ }
+ if (offset + length > tdb->transaction->old_map_size) {
+ /* Short read at EOF. */
+ length = tdb->transaction->old_map_size - offset;
+ }
+ buffer = tdb_access_read(tdb, offset, length, false);
+ if (TDB_PTR_IS_ERR(buffer)) {
+ ecode = TDB_PTR_ERR(buffer);
+ goto fail;
}
+
+ /* Skip over anything the same at the start. */
+ off = same(tdb->transaction->blocks[i], buffer, length);
+ offset += off;
+
+ while (off < length) {
+ tdb_len_t len;
+ unsigned int samelen;
+
+ len = different(tdb->transaction->blocks[i] + off,
+ buffer + off, length - off,
+ sizeof(offset) + sizeof(len) + 1,
+ &samelen);
+
+ memcpy(p, &offset, sizeof(offset));
+ memcpy(p + sizeof(offset), &len, sizeof(len));
+ tdb_convert(tdb, p, sizeof(offset) + sizeof(len));
+ p += sizeof(offset) + sizeof(len);
+ memcpy(p, buffer + off, len);
+ p += len;
+ off += len + samelen;
+ offset += len + samelen;
+ }
+ tdb_access_release(tdb, buffer);
}
- /* the tdb_free() call might have increased the recovery size */
- *recovery_size = tdb_recovery_size(tdb);
+ *len = p - (unsigned char *)(rec + 1);
+ tdb->methods = old_methods;
+ return rec;
+
+fail:
+ free(rec);
+ tdb->methods = old_methods;
+ return TDB_ERR_PTR(ecode);
+}
+
+static tdb_off_t create_recovery_area(struct tdb_context *tdb,
+ tdb_len_t rec_length,
+ struct tdb_recovery_record *rec)
+{
+ tdb_off_t off, recovery_off;
+ tdb_len_t addition;
+ enum TDB_ERROR ecode;
+ const struct tdb_methods *methods = tdb->transaction->io_methods;
/* 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 + *recovery_size / 2)
+ rec->max_len
+ = (((sizeof(*rec) + rec_length + rec_length / 2)
+ PAGESIZE-1) & ~(PAGESIZE-1))
- - sizeof(rec);
- *recovery_offset = tdb->file->map_size;
- recovery_head = *recovery_offset;
+ - sizeof(*rec);
+ off = tdb->file->map_size;
/* Restore ->map_size before calling underlying expand_file.
Also so that we don't try to expand the file again in the
transaction commit, which would destroy the recovery
area */
addition = (tdb->file->map_size - tdb->transaction->old_map_size) +
- sizeof(rec) + *recovery_max_size;
+ sizeof(*rec) + rec->max_len;
tdb->file->map_size = tdb->transaction->old_map_size;
ecode = methods->expand_file(tdb, addition);
if (ecode != TDB_SUCCESS) {
/* write the recovery header offset and sync - we can sync without a race here
as the magic ptr in the recovery record has not been set */
- tdb_convert(tdb, &recovery_head, sizeof(recovery_head));
+ recovery_off = off;
+ tdb_convert(tdb, &recovery_off, sizeof(recovery_off));
ecode = methods->twrite(tdb, offsetof(struct tdb_header, recovery),
- &recovery_head, sizeof(tdb_off_t));
+ &recovery_off, sizeof(tdb_off_t));
if (ecode != TDB_SUCCESS) {
return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
"tdb_recovery_allocate:"
" failed to write recovery head");
}
transaction_write_existing(tdb, offsetof(struct tdb_header, recovery),
- &recovery_head,
+ &recovery_off,
sizeof(tdb_off_t));
- return TDB_SUCCESS;
-}
-
-/* Set up header for the recovery record. */
-static void set_recovery_header(struct tdb_recovery_record *rec,
- uint64_t magic,
- uint64_t datalen, uint64_t actuallen,
- uint64_t oldsize)
-{
- rec->magic = magic;
- rec->max_len = actuallen;
- rec->len = datalen;
- rec->eof = oldsize;
+ return off;
}
/*
setup the recovery data that will be used on a crash during commit
*/
-static enum TDB_ERROR transaction_setup_recovery(struct tdb_context *tdb,
- tdb_off_t *magic_offset)
+static enum TDB_ERROR transaction_setup_recovery(struct tdb_context *tdb)
{
- /* Initialized for GCC's 4.4.5 overzealous uninitialized warnings. */
tdb_len_t recovery_size = 0;
- tdb_off_t recovery_offset = 0, recovery_max_size = 0;
- unsigned char *data, *p;
- const struct tdb_methods *methods = tdb->transaction->io_methods;
- struct tdb_recovery_record *rec;
+ tdb_off_t recovery_off = 0;
tdb_off_t old_map_size = tdb->transaction->old_map_size;
+ struct tdb_recovery_record *recovery;
+ const struct tdb_methods *methods = tdb->transaction->io_methods;
uint64_t magic;
- int i;
enum TDB_ERROR ecode;
- /*
- check that the recovery area has enough space
- */
- ecode = tdb_recovery_allocate(tdb, &recovery_size,
- &recovery_offset, &recovery_max_size);
- if (ecode != TDB_SUCCESS) {
- return ecode;
- }
+ recovery = alloc_recovery(tdb, &recovery_size);
+ if (TDB_PTR_IS_ERR(recovery))
+ return TDB_PTR_ERR(recovery);
- data = (unsigned char *)malloc(recovery_size + sizeof(*rec));
- if (data == NULL) {
- return tdb_logerr(tdb, TDB_ERR_OOM, TDB_LOG_ERROR,
- "transaction_setup_recovery:"
- " cannot allocate");
+ ecode = tdb_recovery_area(tdb, methods, &recovery_off, recovery);
+ if (ecode) {
+ free(recovery);
+ return ecode;
}
- rec = (struct tdb_recovery_record *)data;
- set_recovery_header(rec, TDB_RECOVERY_INVALID_MAGIC,
- recovery_size, recovery_max_size, old_map_size);
- tdb_convert(tdb, rec, sizeof(*rec));
-
- /* build the recovery data into a single blob to allow us to do a single
- large write, which should be more efficient */
- p = data + sizeof(*rec);
- for (i=0;i<tdb->transaction->num_blocks;i++) {
- tdb_off_t offset;
- tdb_len_t length;
-
- if (tdb->transaction->blocks[i] == NULL) {
- continue;
- }
-
- offset = i * PAGESIZE;
- length = PAGESIZE;
- if (i == tdb->transaction->num_blocks-1) {
- length = tdb->transaction->last_block_size;
- }
+ if (recovery->max_len < recovery_size) {
+ /* Not large enough. Free up old recovery area. */
+ if (recovery_off) {
+ tdb->stats.frees++;
+ ecode = add_free_record(tdb, recovery_off,
+ sizeof(*recovery)
+ + recovery->max_len,
+ TDB_LOCK_WAIT, true);
+ free(recovery);
+ if (ecode != TDB_SUCCESS) {
+ return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
+ "tdb_recovery_allocate:"
+ " failed to free previous"
+ " recovery area");
+ }
- if (offset >= old_map_size) {
- continue;
+ /* Refresh recovery after add_free_record above. */
+ recovery = alloc_recovery(tdb, &recovery_size);
+ if (TDB_PTR_IS_ERR(recovery))
+ return TDB_PTR_ERR(recovery);
}
- if (offset + length > tdb->file->map_size) {
- free(data);
- return tdb_logerr(tdb, TDB_ERR_CORRUPT, TDB_LOG_ERROR,
- "tdb_transaction_setup_recovery:"
- " transaction data over new region"
- " boundary");
- }
- memcpy(p, &offset, sizeof(offset));
- memcpy(p + sizeof(offset), &length, sizeof(length));
- tdb_convert(tdb, p, sizeof(offset) + sizeof(length));
-
- /* the recovery area contains the old data, not the
- new data, so we have to call the original tdb_read
- method to get it */
- if (offset + length > old_map_size) {
- /* Short read at EOF, and zero fill. */
- unsigned int len = old_map_size - offset;
- ecode = methods->tread(tdb, offset,
- p + sizeof(offset) + sizeof(length),
- len);
- memset(p + sizeof(offset) + sizeof(length) + len, 0,
- length - len);
- } else {
- ecode = methods->tread(tdb, offset,
- p + sizeof(offset) + sizeof(length),
- length);
- }
- if (ecode != TDB_SUCCESS) {
- free(data);
- return ecode;
+ recovery_off = create_recovery_area(tdb, recovery_size,
+ recovery);
+ if (TDB_OFF_IS_ERR(recovery_off)) {
+ free(recovery);
+ return recovery_off;
}
- p += sizeof(offset) + sizeof(length) + length;
}
+ /* Now we know size, convert rec header. */
+ recovery->magic = TDB_RECOVERY_INVALID_MAGIC;
+ recovery->len = recovery_size;
+ recovery->eof = old_map_size;
+ tdb_convert(tdb, recovery, sizeof(*recovery));
+
/* write the recovery data to the recovery area */
- ecode = methods->twrite(tdb, recovery_offset, data,
- sizeof(*rec) + recovery_size);
+ ecode = methods->twrite(tdb, recovery_off, recovery, recovery_size);
if (ecode != TDB_SUCCESS) {
- free(data);
+ free(recovery);
return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
"tdb_transaction_setup_recovery:"
" failed to write recovery data");
}
- transaction_write_existing(tdb, recovery_offset, data,
- sizeof(*rec) + recovery_size);
+ transaction_write_existing(tdb, recovery_off, recovery, recovery_size);
+
+ free(recovery);
/* as we don't have ordered writes, we have to sync the recovery
data before we update the magic to indicate that the recovery
data is present */
- ecode = transaction_sync(tdb, recovery_offset,
- sizeof(*rec) + recovery_size);
- if (ecode != TDB_SUCCESS) {
- free(data);
+ ecode = transaction_sync(tdb, recovery_off, recovery_size);
+ if (ecode != TDB_SUCCESS)
return ecode;
- }
-
- free(data);
magic = TDB_RECOVERY_MAGIC;
tdb_convert(tdb, &magic, sizeof(magic));
- *magic_offset = recovery_offset + offsetof(struct tdb_recovery_record,
- magic);
+ tdb->transaction->magic_offset
+ = recovery_off + offsetof(struct tdb_recovery_record, magic);
- ecode = methods->twrite(tdb, *magic_offset, &magic, sizeof(magic));
+ ecode = methods->twrite(tdb, tdb->transaction->magic_offset,
+ &magic, sizeof(magic));
if (ecode != TDB_SUCCESS) {
return tdb_logerr(tdb, ecode, TDB_LOG_ERROR,
"tdb_transaction_setup_recovery:"
" failed to write recovery magic");
}
- transaction_write_existing(tdb, *magic_offset, &magic, sizeof(magic));
+ transaction_write_existing(tdb, tdb->transaction->magic_offset,
+ &magic, sizeof(magic));
/* ensure the recovery magic marker is on disk */
- return transaction_sync(tdb, *magic_offset, sizeof(magic));
+ return transaction_sync(tdb, tdb->transaction->magic_offset,
+ sizeof(magic));
}
static enum TDB_ERROR _tdb_transaction_prepare_commit(struct tdb_context *tdb)
/* Since we have whole db locked, we don't need the expansion lock. */
if (!(tdb->flags & TDB_NOSYNC)) {
- /* write the recovery data to the end of the file */
- ecode = transaction_setup_recovery(tdb,
- &tdb->transaction
- ->magic_offset);
+ /* Sets up tdb->transaction->recovery and
+ * tdb->transaction->magic_offset. */
+ ecode = transaction_setup_recovery(tdb);
if (ecode != TDB_SUCCESS) {
return ecode;
}
projects / ccan / blobdiff