#include "tdb_private.h"
+/* 'right' merges can involve O(n^2) cost when combined with a
+ traverse, so they are disabled until we find a way to do them in
+ O(1) time
+*/
+#define USE_RIGHT_MERGES 0
+
/* read a freelist record and check for simple errors */
int tdb_rec_free_read(struct tdb_context *tdb, tdb_off_t off, struct list_struct *rec)
{
}
-
+#if USE_RIGHT_MERGES
/* Remove an element from the freelist. Must have alloc lock. */
static int remove_from_freelist(struct tdb_context *tdb, tdb_off_t off, tdb_off_t next)
{
TDB_LOG((tdb, TDB_DEBUG_FATAL,"remove_from_freelist: not on list at off=%d\n", off));
return TDB_ERRCODE(TDB_ERR_CORRUPT, -1);
}
+#endif
/* update a record tailer (must hold allocation lock) */
neccessary. */
int tdb_free(struct tdb_context *tdb, tdb_off_t offset, struct list_struct *rec)
{
- tdb_off_t right, left;
-
/* Allocation and tailer lock */
if (tdb_lock(tdb, -1, F_WRLCK) != 0)
return -1;
goto fail;
}
+#if USE_RIGHT_MERGES
/* Look right first (I'm an Australian, dammit) */
- right = offset + sizeof(*rec) + rec->rec_len;
- if (right + sizeof(*rec) <= tdb->map_size) {
+ if (offset + sizeof(*rec) + rec->rec_len + sizeof(*rec) <= tdb->map_size) {
+ tdb_off_t right = offset + sizeof(*rec) + rec->rec_len;
struct list_struct r;
if (tdb->methods->tdb_read(tdb, right, &r, sizeof(r), DOCONV()) == -1) {
goto left;
}
rec->rec_len += sizeof(r) + r.rec_len;
+ if (update_tailer(tdb, offset, rec) == -1) {
+ TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: update_tailer failed at %u\n", offset));
+ goto fail;
+ }
}
}
-
left:
+#endif
+
/* Look left */
- left = offset - sizeof(tdb_off_t);
- if (left > TDB_DATA_START(tdb->header.hash_size)) {
+ if (offset - sizeof(tdb_off_t) > TDB_DATA_START(tdb->header.hash_size)) {
+ tdb_off_t left = offset - sizeof(tdb_off_t);
struct list_struct l;
tdb_off_t leftsize;
left = offset - leftsize;
- /* Now read in record */
+ if (leftsize > offset ||
+ left < TDB_DATA_START(tdb->header.hash_size)) {
+ goto update;
+ }
+
+ /* Now read in the left record */
if (tdb->methods->tdb_read(tdb, left, &l, sizeof(l), DOCONV()) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: left read failed at %u (%u)\n", left, leftsize));
goto update;
/* If it's free, expand to include it. */
if (l.magic == TDB_FREE_MAGIC) {
- if (remove_from_freelist(tdb, left, l.next) == -1) {
- TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: left free failed at %u\n", left));
- goto update;
- } else {
- offset = left;
- rec->rec_len += leftsize;
+ /* we now merge the new record into the left record, rather than the other
+ way around. This makes the operation O(1) instead of O(n). This change
+ prevents traverse from being O(n^2) after a lot of deletes */
+ l.rec_len += sizeof(*rec) + rec->rec_len;
+ if (tdb_rec_write(tdb, left, &l) == -1) {
+ TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: update_left failed at %u\n", left));
+ goto fail;
}
+ if (update_tailer(tdb, left, &l) == -1) {
+ TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: update_tailer failed at %u\n", offset));
+ goto fail;
+ }
+ tdb_unlock(tdb, -1, F_WRLCK);
+ return 0;
}
}
update:
- if (update_tailer(tdb, offset, rec) == -1) {
- TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_free: update_tailer failed at %u\n", offset));
- goto fail;
- }
/* Now, prepend to free list */
rec->magic = TDB_FREE_MAGIC;
tdb_off_t rec_ptr, last_ptr;
tdb_len_t rec_len;
} bestfit;
+ float multiplier = 1.0;
if (tdb_lock(tdb, -1, F_WRLCK) == -1)
return 0;
bestfit.rec_len = rec->rec_len;
bestfit.rec_ptr = rec_ptr;
bestfit.last_ptr = last_ptr;
- /* consider a fit to be good enough if
- we aren't wasting more than half
- the space */
- if (bestfit.rec_len < 2*length) {
- break;
- }
}
}
/* move to the next record */
last_ptr = rec_ptr;
rec_ptr = rec->next;
+
+ /* if we've found a record that is big enough, then
+ stop searching if its also not too big. The
+ definition of 'too big' changes as we scan
+ through */
+ if (bestfit.rec_len > 0 &&
+ bestfit.rec_len < length * multiplier) {
+ break;
+ }
+
+ /* this multiplier means we only extremely rarely
+ search more than 50 or so records. At 50 records we
+ accept records up to 11 times larger than what we
+ want */
+ multiplier *= 1.05;
}
if (bestfit.rec_ptr != 0) {
return 0;
}
+
+
+/*
+ return the size of the freelist - used to decide if we should repack
+*/
+int tdb_freelist_size(struct tdb_context *tdb)
+{
+ tdb_off_t ptr;
+ int count=0;
+
+ if (tdb_lock(tdb, -1, F_RDLCK) == -1) {
+ return -1;
+ }
+
+ ptr = FREELIST_TOP;
+ while (tdb_ofs_read(tdb, ptr, &ptr) == 0 && ptr != 0) {
+ count++;
+ }
+
+ tdb_unlock(tdb, -1, F_RDLCK);
+ return count;
+}
*/
-struct tdb_transaction_el {
- struct tdb_transaction_el *next, *prev;
- tdb_off_t offset;
- tdb_len_t length;
- unsigned char *data;
-};
-
/*
hold the context of any current transaction
*/
/* the original io methods - used to do IOs to the real db */
const struct tdb_methods *io_methods;
- /* the list of transaction elements. We use a doubly linked
- list with a last pointer to allow us to keep the list
- ordered, with first element at the front of the list. It
- needs to be doubly linked as the read/write traversals need
- to be backwards, while the commit needs to be forwards */
- struct tdb_transaction_el *elements, *elements_last;
+ /* the list of transaction blocks. When a block is first
+ written to, it gets created in this list */
+ uint8_t **blocks;
+ uint32_t num_blocks;
+ uint32_t block_size; /* bytes in each block */
+ uint32_t last_block_size; /* number of valid bytes in the last block */
/* non-zero when an internal transaction error has
occurred. All write operations will then fail until the
static int transaction_read(struct tdb_context *tdb, tdb_off_t off, void *buf,
tdb_len_t len, int cv)
{
- struct tdb_transaction_el *el;
-
- /* we need to walk the list backwards to get the most recent data */
- for (el=tdb->transaction->elements_last;el;el=el->prev) {
- tdb_len_t partial;
+ uint32_t blk;
- if (off+len <= el->offset) {
- continue;
- }
- if (off >= el->offset + el->length) {
- continue;
+ /* break it down into block sized ops */
+ while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
+ tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
+ if (transaction_read(tdb, off, buf, len2, cv) != 0) {
+ return -1;
}
+ len -= len2;
+ off += len2;
+ buf = (void *)(len2 + (char *)buf);
+ }
- /* an overlapping read - needs to be split into up to
- 2 reads and a memcpy */
- if (off < el->offset) {
- partial = el->offset - off;
- if (transaction_read(tdb, off, buf, partial, cv) != 0) {
- goto fail;
- }
- len -= partial;
- off += partial;
- buf = (void *)(partial + (char *)buf);
- }
- if (off + len <= el->offset + el->length) {
- partial = len;
- } else {
- partial = el->offset + el->length - off;
- }
- memcpy(buf, el->data + (off - el->offset), partial);
- if (cv) {
- tdb_convert(buf, len);
- }
- len -= partial;
- off += partial;
- buf = (void *)(partial + (char *)buf);
-
- if (len != 0 && transaction_read(tdb, off, buf, len, cv) != 0) {
+ if (len == 0) {
+ return 0;
+ }
+
+ blk = off / tdb->transaction->block_size;
+
+ /* see if we have it in the block list */
+ if (tdb->transaction->num_blocks <= blk ||
+ tdb->transaction->blocks[blk] == NULL) {
+ /* nope, do a real read */
+ if (tdb->transaction->io_methods->tdb_read(tdb, off, buf, len, cv) != 0) {
goto fail;
}
-
return 0;
}
- /* its not in the transaction elements - do a real read */
- return tdb->transaction->io_methods->tdb_read(tdb, off, buf, len, cv);
+ /* it is in the block list. Now check for the last block */
+ if (blk == tdb->transaction->num_blocks-1) {
+ if (len > tdb->transaction->last_block_size) {
+ goto fail;
+ }
+ }
+
+ /* now copy it out of this block */
+ memcpy(buf, tdb->transaction->blocks[blk] + (off % tdb->transaction->block_size), len);
+ if (cv) {
+ tdb_convert(buf, len);
+ }
+ return 0;
fail:
TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_read: failed at off=%d len=%d\n", off, len));
static int transaction_write(struct tdb_context *tdb, tdb_off_t off,
const void *buf, tdb_len_t len)
{
- struct tdb_transaction_el *el, *best_el=NULL;
+ uint32_t blk;
- if (len == 0) {
- return 0;
- }
-
/* if the write is to a hash head, then update the transaction
hash heads */
if (len == sizeof(tdb_off_t) && off >= FREELIST_TOP &&
memcpy(&tdb->transaction->hash_heads[chain], buf, len);
}
- /* first see if we can replace an existing entry */
- for (el=tdb->transaction->elements_last;el;el=el->prev) {
- tdb_len_t partial;
-
- if (best_el == NULL && off == el->offset+el->length) {
- best_el = el;
+ /* break it up into block sized chunks */
+ while (len + (off % tdb->transaction->block_size) > tdb->transaction->block_size) {
+ tdb_len_t len2 = tdb->transaction->block_size - (off % tdb->transaction->block_size);
+ if (transaction_write(tdb, off, buf, len2) != 0) {
+ return -1;
}
-
- if (off+len <= el->offset) {
- continue;
- }
- if (off >= el->offset + el->length) {
- continue;
+ len -= len2;
+ off += len2;
+ if (buf != NULL) {
+ buf = (const void *)(len2 + (const char *)buf);
}
+ }
- /* an overlapping write - needs to be split into up to
- 2 writes and a memcpy */
- if (off < el->offset) {
- partial = el->offset - off;
- if (transaction_write(tdb, off, buf, partial) != 0) {
- goto fail;
- }
- len -= partial;
- off += partial;
- buf = (const void *)(partial + (const char *)buf);
- }
- if (off + len <= el->offset + el->length) {
- partial = len;
+ if (len == 0) {
+ return 0;
+ }
+
+ blk = off / tdb->transaction->block_size;
+ off = off % tdb->transaction->block_size;
+
+ if (tdb->transaction->num_blocks <= blk) {
+ uint8_t **new_blocks;
+ /* expand the blocks array */
+ if (tdb->transaction->blocks == NULL) {
+ new_blocks = malloc((blk+1)*sizeof(uint8_t *));
} else {
- partial = el->offset + el->length - off;
+ new_blocks = realloc(tdb->transaction->blocks, (blk+1)*sizeof(uint8_t *));
}
- memcpy(el->data + (off - el->offset), buf, partial);
- len -= partial;
- off += partial;
- buf = (const void *)(partial + (const char *)buf);
-
- if (len != 0 && transaction_write(tdb, off, buf, len) != 0) {
+ if (new_blocks == NULL) {
+ tdb->ecode = TDB_ERR_OOM;
goto fail;
}
-
- return 0;
+ memset(&new_blocks[tdb->transaction->num_blocks], 0,
+ (1+(blk - tdb->transaction->num_blocks))*sizeof(uint8_t *));
+ tdb->transaction->blocks = new_blocks;
+ tdb->transaction->num_blocks = blk+1;
+ tdb->transaction->last_block_size = 0;
}
- /* see if we can append the new entry to an existing entry */
- if (best_el && best_el->offset + best_el->length == off &&
- (off+len < tdb->transaction->old_map_size ||
- off > tdb->transaction->old_map_size)) {
- unsigned char *data = best_el->data;
- el = best_el;
- el->data = (unsigned char *)realloc(el->data,
- el->length + len);
- if (el->data == NULL) {
+ /* allocate and fill a block? */
+ if (tdb->transaction->blocks[blk] == NULL) {
+ tdb->transaction->blocks[blk] = (uint8_t *)calloc(tdb->transaction->block_size, 1);
+ if (tdb->transaction->blocks[blk] == NULL) {
tdb->ecode = TDB_ERR_OOM;
tdb->transaction->transaction_error = 1;
- el->data = data;
- return -1;
+ return -1;
}
- if (buf) {
- memcpy(el->data + el->length, buf, len);
- } else {
- memset(el->data + el->length, TDB_PAD_BYTE, len);
+ if (tdb->transaction->old_map_size > blk * tdb->transaction->block_size) {
+ tdb_len_t len2 = tdb->transaction->block_size;
+ if (len2 + (blk * tdb->transaction->block_size) > tdb->transaction->old_map_size) {
+ len2 = tdb->transaction->old_map_size - (blk * tdb->transaction->block_size);
+ }
+ if (tdb->transaction->io_methods->tdb_read(tdb, blk * tdb->transaction->block_size,
+ tdb->transaction->blocks[blk],
+ len2, 0) != 0) {
+ SAFE_FREE(tdb->transaction->blocks[blk]);
+ tdb->ecode = TDB_ERR_IO;
+ goto fail;
+ }
+ if (blk == tdb->transaction->num_blocks-1) {
+ tdb->transaction->last_block_size = len2;
+ }
}
- el->length += len;
- return 0;
}
-
- /* add a new entry at the end of the list */
- el = (struct tdb_transaction_el *)malloc(sizeof(*el));
- if (el == NULL) {
- tdb->ecode = TDB_ERR_OOM;
- tdb->transaction->transaction_error = 1;
- return -1;
- }
- el->next = NULL;
- el->prev = tdb->transaction->elements_last;
- el->offset = off;
- el->length = len;
- el->data = (unsigned char *)malloc(len);
- if (el->data == NULL) {
- free(el);
- tdb->ecode = TDB_ERR_OOM;
- tdb->transaction->transaction_error = 1;
- return -1;
- }
- if (buf) {
- memcpy(el->data, buf, len);
+
+ /* overwrite part of an existing block */
+ if (buf == NULL) {
+ memset(tdb->transaction->blocks[blk] + off, 0, len);
} else {
- memset(el->data, TDB_PAD_BYTE, len);
+ memcpy(tdb->transaction->blocks[blk] + off, buf, len);
}
- if (el->prev) {
- el->prev->next = el;
- } else {
- tdb->transaction->elements = el;
+ if (blk == tdb->transaction->num_blocks-1) {
+ if (len + off > tdb->transaction->last_block_size) {
+ tdb->transaction->last_block_size = len + off;
+ }
}
- tdb->transaction->elements_last = el;
+
return 0;
fail:
- TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: failed at off=%d len=%d\n", off, len));
- tdb->ecode = TDB_ERR_IO;
+ TDB_LOG((tdb, TDB_DEBUG_FATAL, "transaction_write: failed at off=%d len=%d\n",
+ (blk*tdb->transaction->block_size) + off, len));
tdb->transaction->transaction_error = 1;
return -1;
}
return -1;
}
+ /* a page at a time seems like a reasonable compromise between compactness and efficiency */
+ tdb->transaction->block_size = tdb->page_size;
+
/* get the transaction write lock. This is a blocking lock. As
discussed with Volker, there are a number of ways we could
make this async, which we will probably do in the future */
if (tdb_transaction_lock(tdb, F_WRLCK) == -1) {
+ SAFE_FREE(tdb->transaction->blocks);
SAFE_FREE(tdb->transaction);
return -1;
}
tdb->transaction->io_methods = tdb->methods;
tdb->methods = &transaction_methods;
- /* by calling this transaction write here, we ensure that we don't grow the
- transaction linked list due to hash table updates */
- if (transaction_write(tdb, FREELIST_TOP, tdb->transaction->hash_heads,
- TDB_HASHTABLE_SIZE(tdb)) != 0) {
- TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_start: failed to prime hash table\n"));
- tdb->ecode = TDB_ERR_IO;
- tdb->methods = tdb->transaction->io_methods;
- goto fail;
- }
-
return 0;
fail:
tdb_brlock(tdb, FREELIST_TOP, F_UNLCK, F_SETLKW, 0, 0);
tdb_transaction_unlock(tdb);
+ SAFE_FREE(tdb->transaction->blocks);
SAFE_FREE(tdb->transaction->hash_heads);
SAFE_FREE(tdb->transaction);
return -1;
*/
int tdb_transaction_cancel(struct tdb_context *tdb)
{
+ int i;
+
if (tdb->transaction == NULL) {
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_cancel: no transaction\n"));
return -1;
tdb->map_size = tdb->transaction->old_map_size;
- /* free all the transaction elements */
- while (tdb->transaction->elements) {
- struct tdb_transaction_el *el = tdb->transaction->elements;
- tdb->transaction->elements = el->next;
- free(el->data);
- free(el);
+ /* free all the transaction blocks */
+ for (i=0;i<tdb->transaction->num_blocks;i++) {
+ if (tdb->transaction->blocks[i] != NULL) {
+ free(tdb->transaction->blocks[i]);
+ }
}
+ SAFE_FREE(tdb->transaction->blocks);
/* remove any global lock created during the transaction */
if (tdb->global_lock.count != 0) {
/* remove any locks created during the transaction */
if (tdb->num_locks != 0) {
- int i;
for (i=0;i<tdb->num_lockrecs;i++) {
tdb_brlock(tdb,FREELIST_TOP+4*tdb->lockrecs[i].list,
F_UNLCK,F_SETLKW, 0, 1);
*/
static tdb_len_t tdb_recovery_size(struct tdb_context *tdb)
{
- struct tdb_transaction_el *el;
tdb_len_t recovery_size = 0;
+ int i;
recovery_size = sizeof(uint32_t);
- for (el=tdb->transaction->elements;el;el=el->next) {
- if (el->offset >= tdb->transaction->old_map_size) {
+ for (i=0;i<tdb->transaction->num_blocks;i++) {
+ if (i * tdb->transaction->block_size >= tdb->transaction->old_map_size) {
+ break;
+ }
+ if (tdb->transaction->blocks[i] == NULL) {
continue;
}
- recovery_size += 2*sizeof(tdb_off_t) + el->length;
- }
+ recovery_size += 2*sizeof(tdb_off_t);
+ if (i == tdb->transaction->num_blocks-1) {
+ recovery_size += tdb->transaction->last_block_size;
+ } else {
+ recovery_size += tdb->transaction->block_size;
+ }
+ }
return recovery_size;
}
static int transaction_setup_recovery(struct tdb_context *tdb,
tdb_off_t *magic_offset)
{
- struct tdb_transaction_el *el;
tdb_len_t recovery_size;
unsigned char *data, *p;
const struct tdb_methods *methods = tdb->transaction->io_methods;
tdb_off_t recovery_offset, recovery_max_size;
tdb_off_t old_map_size = tdb->transaction->old_map_size;
uint32_t magic, tailer;
+ int i;
/*
check that the recovery area has enough space
/* 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 (el=tdb->transaction->elements;el;el=el->next) {
- if (el->offset >= old_map_size) {
+ for (i=0;i<tdb->transaction->num_blocks;i++) {
+ tdb_off_t offset;
+ tdb_len_t length;
+
+ if (tdb->transaction->blocks[i] == NULL) {
continue;
}
- if (el->offset + el->length > tdb->transaction->old_map_size) {
+
+ offset = i * tdb->transaction->block_size;
+ length = tdb->transaction->block_size;
+ if (i == tdb->transaction->num_blocks-1) {
+ length = tdb->transaction->last_block_size;
+ }
+
+ if (offset >= old_map_size) {
+ continue;
+ }
+ if (offset + length > tdb->transaction->old_map_size) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_setup_recovery: transaction data over new region boundary\n"));
free(data);
tdb->ecode = TDB_ERR_CORRUPT;
return -1;
}
- memcpy(p, &el->offset, 4);
- memcpy(p+4, &el->length, 4);
+ memcpy(p, &offset, 4);
+ memcpy(p+4, &length, 4);
if (DOCONV()) {
tdb_convert(p, 8);
}
/* 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 (methods->tdb_read(tdb, el->offset, p + 8, el->length, 0) != 0) {
+ if (methods->tdb_read(tdb, offset, p + 8, length, 0) != 0) {
free(data);
tdb->ecode = TDB_ERR_IO;
return -1;
}
- p += 8 + el->length;
+ p += 8 + length;
}
/* and the tailer */
const struct tdb_methods *methods;
tdb_off_t magic_offset = 0;
uint32_t zero = 0;
+ int i;
if (tdb->transaction == NULL) {
TDB_LOG((tdb, TDB_DEBUG_ERROR, "tdb_transaction_commit: no transaction\n"));
}
/* check for a null transaction */
- if (tdb->transaction->elements == NULL) {
+ if (tdb->transaction->blocks == NULL) {
tdb_transaction_cancel(tdb);
return 0;
}
}
/* perform all the writes */
- while (tdb->transaction->elements) {
- struct tdb_transaction_el *el = tdb->transaction->elements;
+ for (i=0;i<tdb->transaction->num_blocks;i++) {
+ tdb_off_t offset;
+ tdb_len_t length;
- if (methods->tdb_write(tdb, el->offset, el->data, el->length) == -1) {
+ if (tdb->transaction->blocks[i] == NULL) {
+ continue;
+ }
+
+ offset = i * tdb->transaction->block_size;
+ length = tdb->transaction->block_size;
+ if (i == tdb->transaction->num_blocks-1) {
+ length = tdb->transaction->last_block_size;
+ }
+
+ if (methods->tdb_write(tdb, offset, tdb->transaction->blocks[i], length) == -1) {
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed during commit\n"));
/* we've overwritten part of the data and
TDB_LOG((tdb, TDB_DEBUG_FATAL, "tdb_transaction_commit: write failed\n"));
return -1;
}
- tdb->transaction->elements = el->next;
- free(el->data);
- free(el);
+ SAFE_FREE(tdb->transaction->blocks[i]);
}
+ SAFE_FREE(tdb->transaction->blocks);
+ tdb->transaction->num_blocks = 0;
+
if (!(tdb->flags & TDB_NOSYNC)) {
/* ensure the new data is on disk */
if (transaction_sync(tdb, 0, tdb->map_size) == -1) {
git.samba.org / bbaumbach / samba-autobuild / .git / commitdiff