/*
* __wt_page_in_func --
* Acquire a hazard pointer to a page; if the page is not in-memory,
* read it from the disk and build an in-memory version.
*/
int
__wt_page_in_func(WT_SESSION_IMPL *session, WT_REF *ref, uint32_t flags
#ifdef HAVE_DIAGNOSTIC
, const char *file, int line
#endif
)
{
WT_BTREE *btree;
WT_DECL_RET;
WT_PAGE *page;
u_int sleep_cnt, wait_cnt;
bool busy, cache_work, oldgen, stalled;
int force_attempts;
btree = S2BT(session);
for (oldgen = stalled = false,
force_attempts = 0, sleep_cnt = wait_cnt = 0;;) {
switch (ref->state) {
case WT_REF_DELETED:
if (LF_ISSET(WT_READ_NO_EMPTY) &&
__wt_delete_page_skip(session, ref, false))
return (WT_NOTFOUND);
/* FALLTHROUGH */
case WT_REF_DISK:
if (LF_ISSET(WT_READ_CACHE))
return (WT_NOTFOUND);
/*
* The page isn't in memory, read it. If this thread is
* allowed to do eviction work, check for space in the
* cache.
*/
if (!LF_ISSET(WT_READ_NO_EVICT))
WT_RET(__wt_cache_eviction_check(
session, 1, NULL));
WT_RET(__page_read(session, ref));
oldgen = LF_ISSET(WT_READ_WONT_NEED) ||
F_ISSET(session, WT_SESSION_NO_CACHE);
continue;
case WT_REF_READING:
if (LF_ISSET(WT_READ_CACHE))
return (WT_NOTFOUND);
if (LF_ISSET(WT_READ_NO_WAIT))
return (WT_NOTFOUND);
/* Waiting on another thread's read, stall. */
WT_STAT_FAST_CONN_INCR(session, page_read_blocked);
stalled = true;
break;
case WT_REF_LOCKED:
if (LF_ISSET(WT_READ_NO_WAIT))
return (WT_NOTFOUND);
/* Waiting on eviction, stall. */
WT_STAT_FAST_CONN_INCR(session, page_locked_blocked);
stalled = true;
break;
case WT_REF_SPLIT:
return (WT_RESTART);
case WT_REF_MEM:
/*
* The page is in memory.
*
* Get a hazard pointer if one is required. We cannot
* be evicting if no hazard pointer is required, we're
* done.
*/
if (F_ISSET(btree, WT_BTREE_IN_MEMORY))
goto skip_evict;
/*
* The expected reason we can't get a hazard pointer is
* because the page is being evicted, yield, try again.
*/
#ifdef HAVE_DIAGNOSTIC
WT_RET(
__wt_hazard_set(session, ref, &busy, file, line));
#else
WT_RET(__wt_hazard_set(session, ref, &busy));
#endif
if (busy) {
WT_STAT_FAST_CONN_INCR(
session, page_busy_blocked);
break;
}
/*
* If eviction is configured for this file, check to see
* if the page qualifies for forced eviction and update
* the page's generation number. If eviction isn't being
* done on this file, we're done.
*/
if (LF_ISSET(WT_READ_NO_EVICT) ||
F_ISSET(session, WT_SESSION_NO_EVICTION) ||
F_ISSET(btree, WT_BTREE_NO_EVICTION))
goto skip_evict;
/*
* Forcibly evict pages that are too big.
*/
if (force_attempts < 10 &&
__evict_force_check(session, ref)) {
++force_attempts;
ret = __wt_page_release_evict(session, ref);
/* If forced eviction fails, stall. */
if (ret == EBUSY) {
ret = 0;
WT_STAT_FAST_CONN_INCR(session,
page_forcible_evict_blocked);
stalled = true;
break;
}
WT_RET(ret);
/*
* The result of a successful forced eviction
* is a page-state transition (potentially to
* an in-memory page we can use, or a restart
* return for our caller), continue the outer
* page-acquisition loop.
*/
continue;
}
/*
* If we read the page and we are configured to not
* trash the cache, set the oldest read generation so
* the page is forcibly evicted as soon as possible.
*
* Otherwise, update the page's read generation.
*/
page = ref->page;
if (oldgen && page->read_gen == WT_READGEN_NOTSET)
__wt_page_evict_soon(page);
else if (!LF_ISSET(WT_READ_NO_GEN) &&
page->read_gen != WT_READGEN_OLDEST &&
page->read_gen < __wt_cache_read_gen(session))
page->read_gen =
__wt_cache_read_gen_bump(session);
skip_evict:
/*
* Check if we need an autocommit transaction.
* Starting a transaction can trigger eviction, so skip
* it if eviction isn't permitted.
*/
return (LF_ISSET(WT_READ_NO_EVICT) ? 0 :
__wt_txn_autocommit_check(session));
WT_ILLEGAL_VALUE(session);
}
/*
* We failed to get the page -- yield before retrying, and if
* we've yielded enough times, start sleeping so we don't burn
* CPU to no purpose.
*/
if (stalled)
wait_cnt += WT_THOUSAND;
else if (++wait_cnt < WT_THOUSAND) {
__wt_yield();
continue;
}
/*
* If stalling and this thread is allowed to do eviction work,
* check if the cache needs help. If we do work for the cache,
* substitute that for a sleep.
*/
if (!LF_ISSET(WT_READ_NO_EVICT)) {
WT_RET(
__wt_cache_eviction_check(session, 1, &cache_work));
if (cache_work)
continue;
}
sleep_cnt = WT_MIN(sleep_cnt + WT_THOUSAND, 10000);
WT_STAT_FAST_CONN_INCRV(session, page_sleep, sleep_cnt);
__wt_sleep(0, sleep_cnt);
}
}
/*
* __wt_bt_write --
* Write a buffer into a block, returning the block's addr/size and
* checksum.
*/
int
__wt_bt_write(WT_SESSION_IMPL *session, WT_ITEM *buf,
uint8_t *addr, size_t *addr_sizep, int checkpoint, int compressed)
{
WT_BM *bm;
WT_BTREE *btree;
WT_ITEM *ip;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_PAGE_HEADER *dsk;
size_t len, src_len, dst_len, result_len, size;
int data_cksum, compression_failed;
uint8_t *src, *dst;
btree = S2BT(session);
bm = btree->bm;
/* Checkpoint calls are different than standard calls. */
WT_ASSERT(session,
(checkpoint == 0 && addr != NULL && addr_sizep != NULL) ||
(checkpoint == 1 && addr == NULL && addr_sizep == NULL));
#ifdef HAVE_DIAGNOSTIC
/*
* We're passed a table's disk image. Decompress if necessary and
* verify the image. Always check the in-memory length for accuracy.
*/
dsk = buf->mem;
if (compressed) {
WT_ERR(__wt_scr_alloc(session, dsk->mem_size, &tmp));
memcpy(tmp->mem, buf->data, WT_BLOCK_COMPRESS_SKIP);
WT_ERR(btree->compressor->decompress(
btree->compressor, &session->iface,
(uint8_t *)buf->data + WT_BLOCK_COMPRESS_SKIP,
buf->size - WT_BLOCK_COMPRESS_SKIP,
(uint8_t *)tmp->data + WT_BLOCK_COMPRESS_SKIP,
tmp->memsize - WT_BLOCK_COMPRESS_SKIP,
&result_len));
WT_ASSERT(session,
dsk->mem_size == result_len + WT_BLOCK_COMPRESS_SKIP);
tmp->size = (uint32_t)result_len + WT_BLOCK_COMPRESS_SKIP;
ip = tmp;
} else {
WT_ASSERT(session, dsk->mem_size == buf->size);
ip = buf;
}
WT_ERR(__wt_verify_dsk(session, "[write-check]", ip));
__wt_scr_free(&tmp);
#endif
/*
* Optionally stream-compress the data, but don't compress blocks that
* are already as small as they're going to get.
*/
if (btree->compressor == NULL ||
btree->compressor->compress == NULL || compressed)
ip = buf;
else if (buf->size <= btree->allocsize) {
ip = buf;
WT_STAT_FAST_DATA_INCR(session, compress_write_too_small);
} else {
/* Skip the header bytes of the source data. */
src = (uint8_t *)buf->mem + WT_BLOCK_COMPRESS_SKIP;
src_len = buf->size - WT_BLOCK_COMPRESS_SKIP;
/*
* Compute the size needed for the destination buffer. We only
* allocate enough memory for a copy of the original by default,
* if any compressed version is bigger than the original, we
* won't use it. However, some compression engines (snappy is
* one example), may need more memory because they don't stop
* just because there's no more memory into which to compress.
*/
if (btree->compressor->pre_size == NULL)
len = src_len;
else
WT_ERR(btree->compressor->pre_size(btree->compressor,
&session->iface, src, src_len, &len));
size = len + WT_BLOCK_COMPRESS_SKIP;
WT_ERR(bm->write_size(bm, session, &size));
WT_ERR(__wt_scr_alloc(session, size, &tmp));
/* Skip the header bytes of the destination data. */
dst = (uint8_t *)tmp->mem + WT_BLOCK_COMPRESS_SKIP;
dst_len = len;
compression_failed = 0;
WT_ERR(btree->compressor->compress(btree->compressor,
&session->iface,
src, src_len,
dst, dst_len,
&result_len, &compression_failed));
result_len += WT_BLOCK_COMPRESS_SKIP;
/*
* If compression fails, or doesn't gain us at least one unit of
* allocation, fallback to the original version. This isn't
* unexpected: if compression doesn't work for some chunk of
* data for some reason (noting likely additional format/header
* information which compressed output requires), it just means
* the uncompressed version is as good as it gets, and that's
* what we use.
*/
if (compression_failed ||
buf->size / btree->allocsize ==
result_len / btree->allocsize) {
ip = buf;
WT_STAT_FAST_DATA_INCR(session, compress_write_fail);
} else {
compressed = 1;
WT_STAT_FAST_DATA_INCR(session, compress_write);
/*
* Copy in the skipped header bytes, set the final data
* size.
*/
memcpy(tmp->mem, buf->mem, WT_BLOCK_COMPRESS_SKIP);
tmp->size = result_len;
ip = tmp;
}
}
dsk = ip->mem;
/* If the buffer is compressed, set the flag. */
if (compressed)
F_SET(dsk, WT_PAGE_COMPRESSED);
/*
* We increment the block's write generation so it's easy to identify
* newer versions of blocks during salvage. (It's common in WiredTiger,
* at least for the default block manager, for multiple blocks to be
* internally consistent with identical first and last keys, so we need
* a way to know the most recent state of the block. We could check
* which leaf is referenced by a valid internal page, but that implies
* salvaging internal pages, which I don't want to do, and it's not
* as good anyway, because the internal page may not have been written
* after the leaf page was updated. So, write generations it is.
*
* Nothing is locked at this point but two versions of a page with the
* same generation is pretty unlikely, and if we did, they're going to
* be roughly identical for the purposes of salvage, anyway.
*/
dsk->write_gen = ++btree->write_gen;
/*
* Checksum the data if the buffer isn't compressed or checksums are
* configured.
*/
switch (btree->checksum) {
case CKSUM_ON:
data_cksum = 1;
break;
case CKSUM_OFF:
data_cksum = 0;
break;
case CKSUM_UNCOMPRESSED:
default:
data_cksum = !compressed;
break;
}
/* Call the block manager to write the block. */
WT_ERR(checkpoint ?
bm->checkpoint(bm, session, ip, btree->ckpt, data_cksum) :
bm->write(bm, session, ip, addr, addr_sizep, data_cksum));
WT_STAT_FAST_CONN_INCR(session, cache_write);
WT_STAT_FAST_DATA_INCR(session, cache_write);
WT_STAT_FAST_CONN_INCRV(session, cache_bytes_write, ip->size);
WT_STAT_FAST_DATA_INCRV(session, cache_bytes_write, ip->size);
err: __wt_scr_free(&tmp);
return (ret);
}
/*
* __wt_log_wrlsn --
* Process written log slots and attempt to coalesce them if the LSNs
* are contiguous. The purpose of this function is to advance the
* write_lsn in LSN order after the buffer is written to the log file.
*/
void
__wt_log_wrlsn(WT_SESSION_IMPL *session, int *yield)
{
WT_CONNECTION_IMPL *conn;
WT_LOG *log;
WT_LOG_WRLSN_ENTRY written[WT_SLOT_POOL];
WT_LOGSLOT *coalescing, *slot;
WT_LSN save_lsn;
size_t written_i;
uint32_t i, save_i;
conn = S2C(session);
log = conn->log;
__wt_spin_lock(session, &log->log_writelsn_lock);
restart:
coalescing = NULL;
WT_INIT_LSN(&save_lsn);
written_i = 0;
i = 0;
/*
* Walk the array once saving any slots that are in the
* WT_LOG_SLOT_WRITTEN state.
*/
while (i < WT_SLOT_POOL) {
save_i = i;
slot = &log->slot_pool[i++];
if (slot->slot_state != WT_LOG_SLOT_WRITTEN)
continue;
written[written_i].slot_index = save_i;
written[written_i++].lsn = slot->slot_release_lsn;
}
/*
* If we found any written slots process them. We sort them
* based on the release LSN, and then look for them in order.
*/
if (written_i > 0) {
if (yield != NULL)
*yield = 0;
WT_INSERTION_SORT(written, written_i,
WT_LOG_WRLSN_ENTRY, WT_WRLSN_ENTRY_CMP_LT);
/*
* We know the written array is sorted by LSN. Go
* through them either advancing write_lsn or coalesce
* contiguous ranges of written slots.
*/
for (i = 0; i < written_i; i++) {
slot = &log->slot_pool[written[i].slot_index];
/*
* The log server thread pushes out slots periodically.
* Sometimes they are empty slots. If we find an
* empty slot, where empty means the start and end LSN
* are the same, free it and continue.
*/
if (__wt_log_cmp(&slot->slot_start_lsn,
&slot->slot_release_lsn) == 0 &&
__wt_log_cmp(&slot->slot_start_lsn,
&slot->slot_end_lsn) == 0) {
__wt_log_slot_free(session, slot);
continue;
}
if (coalescing != NULL) {
/*
* If the write_lsn changed, we may be able to
* process slots. Try again.
*/
if (__wt_log_cmp(
&log->write_lsn, &save_lsn) != 0)
goto restart;
if (__wt_log_cmp(&coalescing->slot_end_lsn,
&written[i].lsn) != 0) {
coalescing = slot;
continue;
}
/*
* If we get here we have a slot to coalesce
* and free.
*/
coalescing->slot_last_offset =
slot->slot_last_offset;
coalescing->slot_end_lsn = slot->slot_end_lsn;
WT_STAT_FAST_CONN_INCR(
session, log_slot_coalesced);
/*
* Copy the flag for later closing.
*/
if (F_ISSET(slot, WT_SLOT_CLOSEFH))
F_SET(coalescing, WT_SLOT_CLOSEFH);
} else {
/*
* If this written slot is not the next LSN,
* try to start coalescing with later slots.
* A synchronous write may update write_lsn
* so save the last one we saw to check when
* coalescing slots.
*/
save_lsn = log->write_lsn;
if (__wt_log_cmp(
&log->write_lsn, &written[i].lsn) != 0) {
coalescing = slot;
continue;
}
/*
* If we get here we have a slot to process.
* Advance the LSN and process the slot.
*/
WT_ASSERT(session, __wt_log_cmp(&written[i].lsn,
&slot->slot_release_lsn) == 0);
/*
* We need to maintain the starting offset of
* a log record so that the checkpoint LSN
* refers to the beginning of a real record.
* The last offset in a slot is kept so that
* the checkpoint LSN is close to the end of
* the record.
*/
if (slot->slot_start_lsn.l.offset !=
slot->slot_last_offset)
slot->slot_start_lsn.l.offset =
(uint32_t)slot->slot_last_offset;
log->write_start_lsn = slot->slot_start_lsn;
log->write_lsn = slot->slot_end_lsn;
__wt_cond_signal(session, log->log_write_cond);
WT_STAT_FAST_CONN_INCR(session, log_write_lsn);
/*
* Signal the close thread if needed.
*/
if (F_ISSET(slot, WT_SLOT_CLOSEFH))
__wt_cond_signal(
session, conn->log_file_cond);
}
__wt_log_slot_free(session, slot);
}
}
__wt_spin_unlock(session, &log->log_writelsn_lock);
}
/*
* __wt_realloc_aligned --
* ANSI realloc function that aligns to buffer boundaries, configured with
* the "buffer_alignment" key to wiredtiger_open.
*/
int
__wt_realloc_aligned(WT_SESSION_IMPL *session,
size_t *bytes_allocated_ret, size_t bytes_to_allocate, void *retp)
{
#if defined(HAVE_POSIX_MEMALIGN)
WT_DECL_RET;
/*
* !!!
* This function MUST handle a NULL WT_SESSION_IMPL handle.
*/
if (session != NULL && S2C(session)->buffer_alignment > 0) {
void *p, *newp;
size_t bytes_allocated;
/*
* Sometimes we're allocating memory and we don't care about the
* final length -- bytes_allocated_ret may be NULL.
*/
p = *(void **)retp;
bytes_allocated =
(bytes_allocated_ret == NULL) ? 0 : *bytes_allocated_ret;
WT_ASSERT(session,
(p == NULL && bytes_allocated == 0) ||
(p != NULL &&
(bytes_allocated_ret == NULL || bytes_allocated != 0)));
WT_ASSERT(session, bytes_to_allocate != 0);
WT_ASSERT(session, bytes_allocated < bytes_to_allocate);
if (session != NULL)
WT_STAT_FAST_CONN_INCR(session, memory_allocation);
if ((ret = posix_memalign(&newp,
S2C(session)->buffer_alignment,
bytes_to_allocate)) != 0)
WT_RET_MSG(session, ret, "memory allocation");
if (p != NULL)
memcpy(newp, p, bytes_allocated);
__wt_free(session, p);
p = newp;
/* Clear the allocated memory (see above). */
memset((uint8_t *)p + bytes_allocated, 0,
bytes_to_allocate - bytes_allocated);
/* Update caller's bytes allocated value. */
if (bytes_allocated_ret != NULL)
*bytes_allocated_ret = bytes_to_allocate;
*(void **)retp = p;
return (0);
}
#endif
/*
* If there is no posix_memalign function, or no alignment configured,
* fall back to realloc.
*/
return (__wt_realloc(
session, bytes_allocated_ret, bytes_to_allocate, retp));
}
/*
* __wt_btcur_insert --
* Insert a record into the tree.
*/
int
__wt_btcur_insert(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_SESSION_IMPL *session;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
WT_STAT_FAST_CONN_INCR(session, cursor_insert);
WT_STAT_FAST_DATA_INCR(session, cursor_insert);
WT_STAT_FAST_DATA_INCRV(session,
cursor_insert_bytes, cursor->key.size + cursor->value.size);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
WT_RET(__cursor_size_chk(session, &cursor->value));
/*
* The tree is no longer empty: eviction should pay attention to it,
* and it's no longer possible to bulk-load into it.
*/
if (btree->bulk_load_ok) {
btree->bulk_load_ok = 0;
__wt_btree_evictable(session, 1);
}
retry: WT_RET(__cursor_func_init(cbt, 1));
switch (btree->type) {
case BTREE_COL_FIX:
case BTREE_COL_VAR:
/*
* If WT_CURSTD_APPEND is set, insert a new record (ignoring
* the application's record number). First we search for the
* maximum possible record number so the search ends on the
* last page. The real record number is assigned by the
* serialized append operation.
*/
if (F_ISSET(cursor, WT_CURSTD_APPEND))
cbt->iface.recno = UINT64_MAX;
WT_ERR(__cursor_col_search(session, cbt, NULL));
if (F_ISSET(cursor, WT_CURSTD_APPEND))
cbt->iface.recno = 0;
/*
* If not overwriting, fail if the key exists. Creating a
* record past the end of the tree in a fixed-length
* column-store implicitly fills the gap with empty records.
* Fail in that case, the record exists.
*/
if (!F_ISSET(cursor, WT_CURSTD_OVERWRITE) &&
((cbt->compare == 0 && __cursor_valid(cbt, NULL)) ||
(cbt->compare != 0 && __cursor_fix_implicit(btree, cbt))))
WT_ERR(WT_DUPLICATE_KEY);
WT_ERR(__cursor_col_modify(session, cbt, 0));
if (F_ISSET(cursor, WT_CURSTD_APPEND))
cbt->iface.recno = cbt->recno;
break;
case BTREE_ROW:
WT_ERR(__cursor_row_search(session, cbt, NULL, 1));
/*
* If not overwriting, fail if the key exists, else insert the
* key/value pair.
*/
if (!F_ISSET(cursor, WT_CURSTD_OVERWRITE) &&
cbt->compare == 0 && __cursor_valid(cbt, NULL))
WT_ERR(WT_DUPLICATE_KEY);
ret = __cursor_row_modify(session, cbt, 0);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
err: if (ret == WT_RESTART) {
WT_STAT_FAST_CONN_INCR(session, cursor_restart);
WT_STAT_FAST_DATA_INCR(session, cursor_restart);
goto retry;
}
/* Insert doesn't maintain a position across calls, clear resources. */
if (ret == 0)
WT_TRET(__curfile_leave(cbt));
if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __wt_btcur_update --
* Update a record in the tree.
*/
int
__wt_btcur_update(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_SESSION_IMPL *session;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
WT_STAT_FAST_CONN_INCR(session, cursor_update);
WT_STAT_FAST_DATA_INCR(session, cursor_update);
WT_STAT_FAST_DATA_INCRV(
session, cursor_update_bytes, cursor->value.size);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
WT_RET(__cursor_size_chk(session, &cursor->value));
/*
* The tree is no longer empty: eviction should pay attention to it,
* and it's no longer possible to bulk-load into it.
*/
if (btree->bulk_load_ok) {
btree->bulk_load_ok = 0;
__wt_btree_evictable(session, 1);
}
retry: WT_RET(__cursor_func_init(cbt, 1));
switch (btree->type) {
case BTREE_COL_FIX:
case BTREE_COL_VAR:
WT_ERR(__cursor_col_search(session, cbt, NULL));
/*
* If not overwriting, fail if the key doesn't exist. If we
* find an update for the key, check for conflicts. Update the
* record if it exists. Creating a record past the end of the
* tree in a fixed-length column-store implicitly fills the gap
* with empty records. Update the record in that case, the
* record exists.
*/
if (!F_ISSET(cursor, WT_CURSTD_OVERWRITE)) {
WT_ERR(__curfile_update_check(cbt));
if ((cbt->compare != 0 || !__cursor_valid(cbt, NULL)) &&
!__cursor_fix_implicit(btree, cbt))
WT_ERR(WT_NOTFOUND);
}
ret = __cursor_col_modify(session, cbt, 0);
break;
case BTREE_ROW:
WT_ERR(__cursor_row_search(session, cbt, NULL, 1));
/*
* If not overwriting, check for conflicts and fail if the key
* does not exist.
*/
if (!F_ISSET(cursor, WT_CURSTD_OVERWRITE)) {
WT_ERR(__curfile_update_check(cbt));
if (cbt->compare != 0 || !__cursor_valid(cbt, NULL))
WT_ERR(WT_NOTFOUND);
}
ret = __cursor_row_modify(session, cbt, 0);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
err: if (ret == WT_RESTART) {
WT_STAT_FAST_CONN_INCR(session, cursor_restart);
WT_STAT_FAST_DATA_INCR(session, cursor_restart);
goto retry;
}
/*
* If successful, point the cursor at internal copies of the data. We
* could shuffle memory in the cursor so the key/value pair are in local
* buffer memory, but that's a data copy. We don't want to do another
* search (and we might get a different update structure if we race).
* To make this work, we add a field to the btree cursor to pass back a
* pointer to the modify function's allocated update structure.
*/
if (ret == 0)
WT_TRET(__wt_kv_return(session, cbt, cbt->modify_update));
if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __wt_btcur_search_near --
* Search for a record in the tree.
*/
int
__wt_btcur_search_near(WT_CURSOR_BTREE *cbt, int *exactp)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
int exact;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
exact = 0;
WT_STAT_FAST_CONN_INCR(session, cursor_search_near);
WT_STAT_FAST_DATA_INCR(session, cursor_search_near);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
WT_RET(__cursor_func_init(cbt, 1));
/*
* Set the "insert" flag for the btree row-store search; we may intend
* to position our cursor at the end of the tree, rather than match an
* existing record.
*/
WT_ERR(btree->type == BTREE_ROW ?
__cursor_row_search(session, cbt, 1) :
__cursor_col_search(session, cbt));
/*
* If we find an valid key, return it.
*
* Else, creating a record past the end of the tree in a fixed-length
* column-store implicitly fills the gap with empty records. In this
* case, we instantiate the empty record, it's an exact match.
*
* Else, move to the next key in the tree (bias for prefix searches).
* Cursor next skips invalid rows, so we don't have to test for them
* again.
*
* Else, redo the search and move to the previous key in the tree.
* Cursor previous skips invalid rows, so we don't have to test for
* them again.
*
* If that fails, quit, there's no record to return.
*/
if (__cursor_valid(cbt, &upd)) {
exact = cbt->compare;
ret = __wt_kv_return(session, cbt, upd);
} else if (__cursor_fix_implicit(btree, cbt)) {
cbt->recno = cursor->recno;
cbt->v = 0;
cursor->value.data = &cbt->v;
cursor->value.size = 1;
exact = 0;
} else if ((ret = __wt_btcur_next(cbt, 0)) != WT_NOTFOUND)
exact = 1;
else {
WT_ERR(btree->type == BTREE_ROW ?
__cursor_row_search(session, cbt, 1) :
__cursor_col_search(session, cbt));
if (__cursor_valid(cbt, &upd)) {
exact = cbt->compare;
ret = __wt_kv_return(session, cbt, upd);
} else if ((ret = __wt_btcur_prev(cbt, 0)) != WT_NOTFOUND)
exact = -1;
}
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
if (exactp != NULL && (ret == 0 || ret == WT_NOTFOUND))
*exactp = exact;
return (ret);
}
/*
* __wt_btcur_search --
* Search for a matching record in the tree.
*/
int
__wt_btcur_search(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
int valid;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
upd = NULL; /* -Wuninitialized */
WT_STAT_FAST_CONN_INCR(session, cursor_search);
WT_STAT_FAST_DATA_INCR(session, cursor_search);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
/*
* If we have a page pinned, search it; if we don't have a page pinned,
* or the search of the pinned page doesn't find an exact match, search
* from the root.
*/
valid = 0;
if (F_ISSET(cbt, WT_CBT_ACTIVE) &&
cbt->ref->page->read_gen != WT_READGEN_OLDEST) {
__wt_txn_cursor_op(session);
WT_ERR(btree->type == BTREE_ROW ?
__cursor_row_search(session, cbt, cbt->ref, 0) :
__cursor_col_search(session, cbt, cbt->ref));
valid = cbt->compare == 0 && __cursor_valid(cbt, &upd);
}
if (!valid) {
WT_ERR(__cursor_func_init(cbt, 1));
WT_ERR(btree->type == BTREE_ROW ?
__cursor_row_search(session, cbt, NULL, 0) :
__cursor_col_search(session, cbt, NULL));
valid = cbt->compare == 0 && __cursor_valid(cbt, &upd);
}
if (valid)
ret = __wt_kv_return(session, cbt, upd);
else if (__cursor_fix_implicit(btree, cbt)) {
/*
* Creating a record past the end of the tree in a fixed-length
* column-store implicitly fills the gap with empty records.
*/
cbt->recno = cursor->recno;
cbt->v = 0;
cursor->value.data = &cbt->v;
cursor->value.size = 1;
} else
ret = WT_NOTFOUND;
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __wt_delete_page --
* If deleting a range, try to delete the page without instantiating it.
*/
int
__wt_delete_page(WT_SESSION_IMPL *session, WT_REF *ref, bool *skipp)
{
WT_DECL_RET;
WT_PAGE *parent;
*skipp = false;
/* If we have a clean page in memory, attempt to evict it. */
if (ref->state == WT_REF_MEM &&
__wt_atomic_casv32(&ref->state, WT_REF_MEM, WT_REF_LOCKED)) {
if (__wt_page_is_modified(ref->page)) {
WT_PUBLISH(ref->state, WT_REF_MEM);
return (0);
}
(void)__wt_atomic_addv32(&S2BT(session)->evict_busy, 1);
ret = __wt_evict(session, ref, false);
(void)__wt_atomic_subv32(&S2BT(session)->evict_busy, 1);
WT_RET_BUSY_OK(ret);
}
/*
* Atomically switch the page's state to lock it. If the page is not
* on-disk, other threads may be using it, no fast delete.
*
* Possible optimization: if the page is already deleted and the delete
* is visible to us (the delete has been committed), we could skip the
* page instead of instantiating it and figuring out there are no rows
* in the page. While that's a huge amount of work to no purpose, it's
* unclear optimizing for overlapping range deletes is worth the effort.
*/
if (ref->state != WT_REF_DISK ||
!__wt_atomic_casv32(&ref->state, WT_REF_DISK, WT_REF_LOCKED))
return (0);
/*
* We cannot fast-delete pages that have overflow key/value items as
* the overflow blocks have to be discarded. The way we figure that
* out is to check the page's cell type, cells for leaf pages without
* overflow items are special.
*
* To look at an on-page cell, we need to look at the parent page, and
* that's dangerous, our parent page could change without warning if
* the parent page were to split, deepening the tree. It's safe: the
* page's reference will always point to some valid page, and if we find
* any problems we simply fail the fast-delete optimization.
*/
parent = ref->home;
if (__wt_off_page(parent, ref->addr) ?
((WT_ADDR *)ref->addr)->type != WT_ADDR_LEAF_NO :
__wt_cell_type_raw(ref->addr) != WT_CELL_ADDR_LEAF_NO)
goto err;
/*
* This action dirties the parent page: mark it dirty now, there's no
* future reconciliation of the child leaf page that will dirty it as
* we write the tree.
*/
WT_ERR(__wt_page_parent_modify_set(session, ref, false));
/*
* Record the change in the transaction structure and set the change's
* transaction ID.
*/
WT_ERR(__wt_calloc_one(session, &ref->page_del));
ref->page_del->txnid = session->txn.id;
WT_ERR(__wt_txn_modify_ref(session, ref));
*skipp = true;
WT_STAT_FAST_CONN_INCR(session, rec_page_delete_fast);
WT_STAT_FAST_DATA_INCR(session, rec_page_delete_fast);
WT_PUBLISH(ref->state, WT_REF_DELETED);
return (0);
err: __wt_free(session, ref->page_del);
/*
* Restore the page to on-disk status, we'll have to instantiate it.
*/
WT_PUBLISH(ref->state, WT_REF_DISK);
return (ret);
}
/*将buffer的数据写入到block对应的文件中，并计算checksum和size*/
int __wt_block_write_off(WT_SESSION_IMPL *session, WT_BLOCK *block, WT_ITEM *buf, wt_off_t *offsetp,
uint32_t *sizep, uint32_t *cksump, int data_cksum, int caller_locked)
{
WT_BLOCK_HEADER *blk;
WT_DECL_RET;
WT_FH *fh;
size_t align_size;
wt_off_t offset;
int local_locked;
blk = WT_BLOCK_HEADER_REF(buf->mem);
fh = block->fh;
local_locked = 0;
/*buf不是对齐模式，不能进行写,因为这个是和磁盘相关的写入，必须是对齐的*/
if(!F_ISSET(buf, WT_ITEM_ALIGNED)){
WT_ASSERT(session, F_ISSET(buf, WT_ITEM_ALIGNED));
WT_RET_MSG(session, EINVAL, "direct I/O check: write buffer incorrectly allocated");
}
/*计算buf->size按block对齐,对齐后有可能会比现有的buf->memsize大，如果大的话，不能进行写，有可能会缓冲区溢出*/
align_size = WT_ALIGN(buf->size, block->allocsize);
if (align_size > buf->memsize) {
WT_ASSERT(session, align_size <= buf->memsize);
WT_RET_MSG(session, EINVAL, "buffer size check: write buffer incorrectly allocated");
}
/*超过4G*/
if (align_size > UINT32_MAX) {
WT_ASSERT(session, align_size <= UINT32_MAX);
WT_RET_MSG(session, EINVAL, "buffer size check: write buffer too large to write");
}
/*将对其后pading的buffer位置进行清0*/
memset((uint8_t*)buf->mem + buf->size, 0, align_size - buf->size);
/*设置block header,计算存储的数据长度*/
blk->disk_size = WT_STORE_SIZE(align_size);
blk->flags = 0;
if(data_cksum)
F_SET(blk, WT_BLOCK_DATA_CKSUM);
/*计算buf的cksum*/
blk->cksum = __wt_cksum(buf->mem, data_cksum ? align_size : WT_BLOCK_COMPRESS_SKIP);
if (!caller_locked) {
WT_RET(__wt_block_ext_prealloc(session, 5));
__wt_spin_lock(session, &block->live_lock);
local_locked = 1;
}
ret = __wt_block_alloc(session, block, &offset, (wt_off_t)align_size);
/*判断文件是否需要进行扩大,如果不扩大就有可能存不下写入的block数据*/
if(ret == 0 && fh->extend_len != 0 && (fh->extend_size <= fh->size ||
(offset + fh->extend_len <= fh->extend_size && offset + fh->extend_len + (wt_off_t)align_size >= fh->extend_size))){
/*调整extend_size为原来的offset + extend_len的两倍*/
fh->extend_size = offset + fh->extend_len * 2;
if (fh->fallocate_available != WT_FALLOCATE_NOT_AVAILABLE) {
/*释放block->live_lock的自旋锁，因为重设文件大小会时间比较长，需要先释放自旋锁，防止CPU空转*/
if (!fh->fallocate_requires_locking && local_locked) {
__wt_spin_unlock(session, &block->live_lock);
local_locked = 0;
}
/*扩大文件的占用空间*/
if ((ret = __wt_fallocate(session,fh, offset, fh->extend_len * 2)) == ENOTSUP) {
ret = 0;
goto extend_truncate;
}
}
else{
extend_truncate:
if (!caller_locked && local_locked == 0) {
__wt_spin_lock(session, &block->live_lock);
local_locked = 1;
}
/*直接调整文件大小,这个比__wt_fallocate更慢*/
if ((ret = __wt_ftruncate(session, fh, offset + fh->extend_len * 2)) == EBUSY)
ret = 0;
}
}
if(local_locked){
__wt_spin_unlock(session, &block->live_lock);
local_locked = 0;
}
WT_RET(ret);
/*进行block的数据写入*/
ret =__wt_write(session, fh, offset, align_size, buf->mem);
if (ret != 0) {
if (!caller_locked)
__wt_spin_lock(session, &block->live_lock);
/*没写成功，将ext对应的数据返回给avail list*/
WT_TRET(__wt_block_off_free(session, block, offset, (wt_off_t)align_size));
if (!caller_locked)
__wt_spin_unlock(session, &block->live_lock);
WT_RET(ret);
}
#ifdef HAVE_SYNC_FILE_RANGE
/*需要进行fsync操作,脏页太多,进行一次异步刷盘*/
if (block->os_cache_dirty_max != 0 && (block->os_cache_dirty += align_size) > block->os_cache_dirty_max && __wt_session_can_wait(session)) {
block->os_cache_dirty = 0;
WT_RET(__wt_fsync_async(session, fh));
}
#endif
#ifdef HAVE_POSIX_FADVISE
/*清理fh->fd文件对应的system page cache中的数据,这个过程可能会有IO操作,相当于同步的sync调用*/
if (block->os_cache_max != 0 && (block->os_cache += align_size) > block->os_cache_max) {
block->os_cache = 0;
if ((ret = posix_fadvise(fh->fd, (wt_off_t)0, (wt_off_t)0, POSIX_FADV_DONTNEED)) != 0)
WT_RET_MSG( session, ret, "%s: posix_fadvise", block->name);
}
#endif
WT_STAT_FAST_CONN_INCR(session, block_write);
WT_STAT_FAST_CONN_INCRV(session, block_byte_write, align_size);
WT_RET(__wt_verbose(session, WT_VERB_WRITE, "off %" PRIuMAX ", size %" PRIuMAX ", cksum %" PRIu32,
(uintmax_t)offset, (uintmax_t)align_size, blk->cksum));
*offsetp = offset;
*sizep = WT_STORE_SIZE(align_size);
*cksump = blk->cksum;
return ret;
}
/*
* __wt_bt_read --
* Read a cookie referenced block into a buffer.
*/
int
__wt_bt_read(WT_SESSION_IMPL *session,
WT_ITEM *buf, const uint8_t *addr, size_t addr_size)
{
WT_BM *bm;
WT_BTREE *btree;
WT_DECL_ITEM(etmp);
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_ENCRYPTOR *encryptor;
WT_ITEM *ip;
const WT_PAGE_HEADER *dsk;
const char *fail_msg;
size_t result_len;
btree = S2BT(session);
bm = btree->bm;
fail_msg = NULL; /* -Wuninitialized */
/*
* If anticipating a compressed or encrypted block, read into a scratch
* buffer and decompress into the caller's buffer. Else, read directly
* into the caller's buffer.
*/
if (btree->compressor == NULL && btree->kencryptor == NULL) {
WT_RET(bm->read(bm, session, buf, addr, addr_size));
dsk = buf->data;
ip = NULL;
} else {
WT_RET(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(bm->read(bm, session, tmp, addr, addr_size));
dsk = tmp->data;
ip = tmp;
}
/*
* If the block is encrypted, copy the skipped bytes of the original
* image into place, then decrypt.
*/
if (F_ISSET(dsk, WT_PAGE_ENCRYPTED)) {
if (btree->kencryptor == NULL ||
(encryptor = btree->kencryptor->encryptor) == NULL ||
encryptor->decrypt == NULL) {
fail_msg =
"encrypted block in file for which no encryption "
"configured";
goto corrupt;
}
WT_ERR(__wt_scr_alloc(session, 0, &etmp));
if ((ret = __wt_decrypt(session,
encryptor, WT_BLOCK_ENCRYPT_SKIP, ip, etmp)) != 0) {
fail_msg = "block decryption failed";
goto corrupt;
}
ip = etmp;
dsk = ip->data;
} else if (btree->kencryptor != NULL) {
fail_msg =
"unencrypted block in file for which encryption configured";
goto corrupt;
}
if (F_ISSET(dsk, WT_PAGE_COMPRESSED)) {
if (btree->compressor == NULL ||
btree->compressor->decompress == NULL) {
fail_msg =
"compressed block in file for which no compression "
"configured";
goto corrupt;
}
/*
* Size the buffer based on the in-memory bytes we're expecting
* from decompression.
*/
WT_ERR(__wt_buf_initsize(session, buf, dsk->mem_size));
/*
* Note the source length is NOT the number of compressed bytes,
* it's the length of the block we just read (minus the skipped
* bytes). We don't store the number of compressed bytes: some
* compression engines need that length stored externally, they
* don't have markers in the stream to signal the end of the
* compressed bytes. Those engines must store the compressed
* byte length somehow, see the snappy compression extension for
* an example.
*/
memcpy(buf->mem, ip->data, WT_BLOCK_COMPRESS_SKIP);
ret = btree->compressor->decompress(
btree->compressor, &session->iface,
(uint8_t *)ip->data + WT_BLOCK_COMPRESS_SKIP,
tmp->size - WT_BLOCK_COMPRESS_SKIP,
(uint8_t *)buf->mem + WT_BLOCK_COMPRESS_SKIP,
dsk->mem_size - WT_BLOCK_COMPRESS_SKIP, &result_len);
/*
* If checksums were turned off because we're depending on the
* decompression to fail on any corrupted data, we'll end up
* here after corruption happens. If we're salvaging the file,
* it's OK, otherwise it's really, really bad.
*/
if (ret != 0 ||
result_len != dsk->mem_size - WT_BLOCK_COMPRESS_SKIP) {
fail_msg = "block decompression failed";
goto corrupt;
}
} else
/*
* If we uncompressed above, the page is in the correct buffer.
* If we get here the data may be in the wrong buffer and the
* buffer may be the wrong size. If needed, get the page
* into the destination buffer.
*/
if (ip != NULL)
WT_ERR(__wt_buf_set(
session, buf, ip->data, dsk->mem_size));
/* If the handle is a verify handle, verify the physical page. */
if (F_ISSET(btree, WT_BTREE_VERIFY)) {
if (tmp == NULL)
WT_ERR(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(bm->addr_string(bm, session, tmp, addr, addr_size));
WT_ERR(__wt_verify_dsk(session, tmp->data, buf));
}
WT_STAT_FAST_CONN_INCR(session, cache_read);
WT_STAT_FAST_DATA_INCR(session, cache_read);
if (F_ISSET(dsk, WT_PAGE_COMPRESSED))
WT_STAT_FAST_DATA_INCR(session, compress_read);
WT_STAT_FAST_CONN_INCRV(session, cache_bytes_read, dsk->mem_size);
WT_STAT_FAST_DATA_INCRV(session, cache_bytes_read, dsk->mem_size);
if (0) {
corrupt: if (ret == 0)
ret = WT_ERROR;
if (!F_ISSET(btree, WT_BTREE_VERIFY) &&
!F_ISSET(session, WT_SESSION_QUIET_CORRUPT_FILE)) {
__wt_err(session, ret, "%s", fail_msg);
ret = __wt_illegal_value(session, btree->dhandle->name);
}
}
err: __wt_scr_free(session, &tmp);
__wt_scr_free(session, &etmp);
return (ret);
}
/*
* __wt_cond_wait_signal --
* Wait on a mutex, optionally timing out. If we get it
* before the time out period expires, let the caller know.
*/
int
__wt_cond_wait_signal(
WT_SESSION_IMPL *session, WT_CONDVAR *cond, uint64_t usecs, bool *signalled)
{
DWORD err, milliseconds;
WT_DECL_RET;
uint64_t milliseconds64;
bool locked;
locked = false;
/* Fast path if already signalled. */
*signalled = true;
if (__wt_atomic_addi32(&cond->waiters, 1) == 0)
return (0);
/*
* !!!
* This function MUST handle a NULL session handle.
*/
if (session != NULL) {
WT_RET(__wt_verbose(session, WT_VERB_MUTEX,
"wait %s cond (%p)", cond->name, cond));
WT_STAT_FAST_CONN_INCR(session, cond_wait);
}
EnterCriticalSection(&cond->mtx);
locked = true;
if (usecs > 0) {
milliseconds64 = usecs / 1000;
/*
* Check for 32-bit unsigned integer overflow
* INFINITE is max unsigned int on Windows
*/
if (milliseconds64 >= INFINITE)
milliseconds64 = INFINITE - 1;
milliseconds = (DWORD)milliseconds64;
/*
* 0 would mean the CV sleep becomes a TryCV which we do not
* want
*/
if (milliseconds == 0)
milliseconds = 1;
ret = SleepConditionVariableCS(
&cond->cond, &cond->mtx, milliseconds);
} else
ret = SleepConditionVariableCS(
&cond->cond, &cond->mtx, INFINITE);
/*
* SleepConditionVariableCS returns non-zero on success, 0 on timeout
* or failure. Check for timeout, else convert to a WiredTiger error
* value and fail.
*/
if (ret == 0) {
if ((err = GetLastError()) == ERROR_TIMEOUT)
*signalled = false;
else
ret = __wt_errno();
} else
ret = 0;
(void)__wt_atomic_subi32(&cond->waiters, 1);
if (locked)
LeaveCriticalSection(&cond->mtx);
if (ret == 0)
return (0);
WT_RET_MSG(session, ret, "SleepConditionVariableCS");
}
/*
* __wt_open --
* Open a file handle.
*/
int
__wt_open(WT_SESSION_IMPL *session,
const char *name, int ok_create, int exclusive, int dio_type, WT_FH **fhp)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_FH *fh, *tfh;
mode_t mode;
int direct_io, f, fd, matched;
const char *path;
conn = S2C(session);
fh = NULL;
fd = -1;
path = NULL;
WT_VERBOSE_RET(session, fileops, "%s: open", name);
/* Increment the reference count if we already have the file open. */
matched = 0;
__wt_spin_lock(session, &conn->fh_lock);
TAILQ_FOREACH(tfh, &conn->fhqh, q)
if (strcmp(name, tfh->name) == 0) {
++tfh->refcnt;
*fhp = tfh;
matched = 1;
break;
}
__wt_spin_unlock(session, &conn->fh_lock);
if (matched)
return (0);
WT_RET(__wt_filename(session, name, &path));
f = O_RDWR;
#ifdef O_BINARY
/* Windows clones: we always want to treat the file as a binary. */
f |= O_BINARY;
#endif
#ifdef O_CLOEXEC
/*
* Security:
* The application may spawn a new process, and we don't want another
* process to have access to our file handles.
*/
f |= O_CLOEXEC;
#endif
#ifdef O_NOATIME
/* Avoid updating metadata for read-only workloads. */
if (dio_type == WT_FILE_TYPE_DATA)
f |= O_NOATIME;
#endif
if (ok_create) {
f |= O_CREAT;
if (exclusive)
f |= O_EXCL;
mode = 0666;
} else
mode = 0;
direct_io = 0;
#ifdef O_DIRECT
if (dio_type && FLD_ISSET(conn->direct_io, dio_type)) {
f |= O_DIRECT;
direct_io = 1;
}
#endif
if (dio_type == WT_FILE_TYPE_LOG &&
FLD_ISSET(conn->txn_logsync, WT_LOG_DSYNC))
#ifdef O_DSYNC
f |= O_DSYNC;
#elif defined(O_SYNC)
f |= O_SYNC;
#else
WT_ERR_MSG(session, ENOTSUP,
"Unsupported log sync mode requested");
#endif
WT_SYSCALL_RETRY(((fd = open(path, f, mode)) == -1 ? 1 : 0), ret);
if (ret != 0)
WT_ERR_MSG(session, ret,
direct_io ?
"%s: open failed with direct I/O configured, some "
"filesystem types do not support direct I/O" : "%s", path);
#if defined(HAVE_FCNTL) && defined(FD_CLOEXEC) && !defined(O_CLOEXEC)
/*
* Security:
* The application may spawn a new process, and we don't want another
* process to have access to our file handles. There's an obvious
* race here, so we prefer the flag to open if available.
*/
if ((f = fcntl(fd, F_GETFD)) == -1 ||
fcntl(fd, F_SETFD, f | FD_CLOEXEC) == -1)
WT_ERR_MSG(session, __wt_errno(), "%s: fcntl", name);
#endif
#if defined(HAVE_POSIX_FADVISE)
/* Disable read-ahead on trees: it slows down random read workloads. */
if (dio_type == WT_FILE_TYPE_DATA)
WT_ERR(posix_fadvise(fd, 0, 0, POSIX_FADV_RANDOM));
#endif
if (F_ISSET(conn, WT_CONN_CKPT_SYNC))
WT_ERR(__open_directory_sync(session));
WT_ERR(__wt_calloc(session, 1, sizeof(WT_FH), &fh));
WT_ERR(__wt_strdup(session, name, &fh->name));
fh->fd = fd;
fh->refcnt = 1;
fh->direct_io = direct_io;
/* Set the file's size. */
WT_ERR(__wt_filesize(session, fh, &fh->size));
/* Configure file extension. */
if (dio_type == WT_FILE_TYPE_DATA)
fh->extend_len = conn->data_extend_len;
/*
* Repeat the check for a match, but then link onto the database's list
* of files.
*/
matched = 0;
__wt_spin_lock(session, &conn->fh_lock);
TAILQ_FOREACH(tfh, &conn->fhqh, q)
if (strcmp(name, tfh->name) == 0) {
++tfh->refcnt;
*fhp = tfh;
matched = 1;
break;
}
if (!matched) {
TAILQ_INSERT_TAIL(&conn->fhqh, fh, q);
WT_STAT_FAST_CONN_INCR(session, file_open);
*fhp = fh;
}
__wt_spin_unlock(session, &conn->fh_lock);
if (matched) {
err: if (fh != NULL) {
__wt_free(session, fh->name);
__wt_free(session, fh);
}
if (fd != -1)
(void)close(fd);
}
__wt_free(session, path);
return (ret);
}
/*
* __wt_bt_read --
* Read a cookie referenced block into a buffer.
*/
int
__wt_bt_read(WT_SESSION_IMPL *session,
WT_ITEM *buf, const uint8_t *addr, size_t addr_size)
{
WT_BM *bm;
WT_BTREE *btree;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
const WT_PAGE_HEADER *dsk;
size_t result_len;
btree = S2BT(session);
bm = btree->bm;
/*
* If anticipating a compressed block, read into a scratch buffer and
* decompress into the caller's buffer. Else, read directly into the
* caller's buffer.
*/
if (btree->compressor == NULL) {
WT_RET(bm->read(bm, session, buf, addr, addr_size));
dsk = buf->data;
} else {
WT_RET(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(bm->read(bm, session, tmp, addr, addr_size));
dsk = tmp->data;
}
/*
* If the block is compressed, copy the skipped bytes of the original
* image into place, then decompress.
*/
if (F_ISSET(dsk, WT_PAGE_COMPRESSED)) {
if (btree->compressor == NULL ||
btree->compressor->decompress == NULL)
WT_ERR_MSG(session, WT_ERROR,
"read compressed block where no compression engine "
"configured");
/*
* We're allocating the exact number of bytes we're expecting
* from decompression.
*/
WT_ERR(__wt_buf_initsize(session, buf, dsk->mem_size));
/*
* Note the source length is NOT the number of compressed bytes,
* it's the length of the block we just read (minus the skipped
* bytes). We don't store the number of compressed bytes: some
* compression engines need that length stored externally, they
* don't have markers in the stream to signal the end of the
* compressed bytes. Those engines must store the compressed
* byte length somehow, see the snappy compression extension for
* an example.
*/
memcpy(buf->mem, tmp->data, WT_BLOCK_COMPRESS_SKIP);
WT_ERR(btree->compressor->decompress(
btree->compressor, &session->iface,
(uint8_t *)tmp->data + WT_BLOCK_COMPRESS_SKIP,
tmp->size - WT_BLOCK_COMPRESS_SKIP,
(uint8_t *)buf->mem + WT_BLOCK_COMPRESS_SKIP,
dsk->mem_size - WT_BLOCK_COMPRESS_SKIP,
&result_len));
/*
* If checksums were turned off because we're depending on the
* decompression to fail on any corrupted data, we'll end up
* here after corruption happens. If we're salvaging the file,
* it's OK, otherwise it's really, really bad.
*/
if (result_len != dsk->mem_size - WT_BLOCK_COMPRESS_SKIP)
WT_ERR(
F_ISSET(btree, WT_BTREE_VERIFY) ||
F_ISSET(session, WT_SESSION_SALVAGE_CORRUPT_OK) ?
WT_ERROR :
__wt_illegal_value(session, btree->dhandle->name));
} else
if (btree->compressor == NULL)
buf->size = dsk->mem_size;
else
/*
* We guessed wrong: there was a compressor, but this
* block was not compressed, and now the page is in the
* wrong buffer and the buffer may be of the wrong size.
* This should be rare, but happens with small blocks
* that aren't worth compressing.
*/
WT_ERR(__wt_buf_set(
session, buf, tmp->data, dsk->mem_size));
/* If the handle is a verify handle, verify the physical page. */
if (F_ISSET(btree, WT_BTREE_VERIFY)) {
if (tmp == NULL)
WT_ERR(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(bm->addr_string(bm, session, tmp, addr, addr_size));
WT_ERR(__wt_verify_dsk(session, (const char *)tmp->data, buf));
}
WT_STAT_FAST_CONN_INCR(session, cache_read);
WT_STAT_FAST_DATA_INCR(session, cache_read);
if (F_ISSET(dsk, WT_PAGE_COMPRESSED))
WT_STAT_FAST_DATA_INCR(session, compress_read);
WT_STAT_FAST_CONN_INCRV(session, cache_bytes_read, dsk->mem_size);
WT_STAT_FAST_DATA_INCRV(session, cache_bytes_read, dsk->mem_size);
err: __wt_scr_free(&tmp);
return (ret);
}
/*
* __wt_log_wrlsn --
* Process written log slots and attempt to coalesce them if the LSNs
* are contiguous. Returns 1 if slots were freed, 0 if no slots were
* freed in the progress arg. Must be called with the log slot lock held.
*/
int
__wt_log_wrlsn(WT_SESSION_IMPL *session, uint32_t *free_i, int *yield)
{
WT_CONNECTION_IMPL *conn;
WT_LOG *log;
WT_LOG_WRLSN_ENTRY written[WT_SLOT_POOL];
WT_LOGSLOT *coalescing, *slot;
size_t written_i;
uint32_t i, save_i;
conn = S2C(session);
log = conn->log;
coalescing = NULL;
written_i = 0;
i = 0;
if (free_i != NULL)
*free_i = WT_SLOT_POOL;
/*
* Walk the array once saving any slots that are in the
* WT_LOG_SLOT_WRITTEN state.
*/
while (i < WT_SLOT_POOL) {
save_i = i;
slot = &log->slot_pool[i++];
if (free_i != NULL && *free_i == WT_SLOT_POOL &&
slot->slot_state == WT_LOG_SLOT_FREE)
*free_i = save_i;
if (slot->slot_state != WT_LOG_SLOT_WRITTEN)
continue;
written[written_i].slot_index = save_i;
written[written_i++].lsn = slot->slot_release_lsn;
}
/*
* If we found any written slots process them. We sort them
* based on the release LSN, and then look for them in order.
*/
if (written_i > 0) {
/*
* If wanted, reset the yield variable to indicate that we
* have found written slots.
*/
if (yield != NULL)
*yield = 0;
WT_INSERTION_SORT(written, written_i,
WT_LOG_WRLSN_ENTRY, WT_WRLSN_ENTRY_CMP_LT);
/*
* We know the written array is sorted by LSN. Go
* through them either advancing write_lsn or coalesce
* contiguous ranges of written slots.
*/
for (i = 0; i < written_i; i++) {
slot = &log->slot_pool[written[i].slot_index];
if (coalescing != NULL) {
if (WT_LOG_CMP(&coalescing->slot_end_lsn,
&written[i].lsn) != 0) {
coalescing = slot;
continue;
}
/*
* If we get here we have a slot to coalesce
* and free.
*/
coalescing->slot_end_lsn = slot->slot_end_lsn;
WT_STAT_FAST_CONN_INCR(
session, log_slot_coalesced);
/*
* Copy the flag for later closing.
*/
if (F_ISSET(slot, WT_SLOT_CLOSEFH))
F_SET(coalescing, WT_SLOT_CLOSEFH);
} else {
/*
* If this written slot is not the next LSN,
* try to start coalescing with later slots.
*/
if (WT_LOG_CMP(
&log->write_lsn, &written[i].lsn) != 0) {
coalescing = slot;
continue;
}
/*
* If we get here we have a slot to process.
* Advance the LSN and process the slot.
*/
WT_ASSERT(session, WT_LOG_CMP(&written[i].lsn,
&slot->slot_release_lsn) == 0);
log->write_start_lsn = slot->slot_start_lsn;
log->write_lsn = slot->slot_end_lsn;
WT_RET(__wt_cond_signal(
session, log->log_write_cond));
WT_STAT_FAST_CONN_INCR(session, log_write_lsn);
/*
* Signal the close thread if needed.
*/
if (F_ISSET(slot, WT_SLOT_CLOSEFH))
WT_RET(__wt_cond_signal(
session, conn->log_file_cond));
}
WT_RET(__wt_log_slot_free(session, slot));
if (free_i != NULL && *free_i == WT_SLOT_POOL &&
slot->slot_state == WT_LOG_SLOT_FREE)
*free_i = save_i;
}
}
return (0);
}
/*
* __wt_page_in_func --
* Acquire a hazard pointer to a page; if the page is not in-memory,
* read it from the disk and build an in-memory version.
*/
int
__wt_page_in_func(WT_SESSION_IMPL *session, WT_REF *ref, uint32_t flags
#ifdef HAVE_DIAGNOSTIC
, const char *file, int line
#endif
)
{
WT_DECL_RET;
WT_PAGE *page;
u_int sleep_cnt, wait_cnt;
int busy, force_attempts, oldgen;
for (force_attempts = oldgen = 0, wait_cnt = 0;;) {
switch (ref->state) {
case WT_REF_DISK:
case WT_REF_DELETED:
if (LF_ISSET(WT_READ_CACHE))
return (WT_NOTFOUND);
/*
* The page isn't in memory, attempt to read it.
* Make sure there is space in the cache.
*/
WT_RET(__wt_cache_full_check(session));
WT_RET(__wt_cache_read(session, ref));
oldgen = LF_ISSET(WT_READ_WONT_NEED) ||
F_ISSET(session, WT_SESSION_NO_CACHE);
continue;
case WT_REF_READING:
if (LF_ISSET(WT_READ_CACHE))
return (WT_NOTFOUND);
if (LF_ISSET(WT_READ_NO_WAIT))
return (WT_NOTFOUND);
WT_STAT_FAST_CONN_INCR(session, page_read_blocked);
break;
case WT_REF_LOCKED:
if (LF_ISSET(WT_READ_NO_WAIT))
return (WT_NOTFOUND);
WT_STAT_FAST_CONN_INCR(session, page_locked_blocked);
break;
case WT_REF_SPLIT:
return (WT_RESTART);
case WT_REF_MEM:
/*
* The page is in memory: get a hazard pointer, update
* the page's LRU and return. The expected reason we
* can't get a hazard pointer is because the page is
* being evicted; yield and try again.
*/
#ifdef HAVE_DIAGNOSTIC
WT_RET(
__wt_hazard_set(session, ref, &busy, file, line));
#else
WT_RET(__wt_hazard_set(session, ref, &busy));
#endif
if (busy) {
WT_STAT_FAST_CONN_INCR(
session, page_busy_blocked);
break;
}
page = ref->page;
WT_ASSERT(session, page != NULL);
/* Forcibly evict pages that are too big. */
if (!LF_ISSET(WT_READ_NO_EVICT) &&
force_attempts < 10 &&
__evict_force_check(session, page)) {
++force_attempts;
if ((ret = __wt_page_release_busy(
session, ref, flags)) == EBUSY) {
/* If forced eviction fails, stall. */
ret = 0;
wait_cnt += 1000;
} else
WT_RET(ret);
WT_STAT_FAST_CONN_INCR(
session, page_forcible_evict_blocked);
break;
}
/* Check if we need an autocommit transaction. */
if ((ret = __wt_txn_autocommit_check(session)) != 0) {
WT_TRET(__wt_hazard_clear(session, page));
return (ret);
}
/*
* If we read the page and we are configured to not
* trash the cache, set the oldest read generation so
* the page is forcibly evicted as soon as possible.
*
* Otherwise, update the page's read generation.
*/
if (oldgen && page->read_gen == WT_READGEN_NOTSET)
__wt_page_evict_soon(page);
else if (!LF_ISSET(WT_READ_NO_GEN) &&
page->read_gen < __wt_cache_read_gen(session))
page->read_gen =
__wt_cache_read_gen_set(session);
return (0);
WT_ILLEGAL_VALUE(session);
}
/*
* We failed to get the page -- yield before retrying, and if
* we've yielded enough times, start sleeping so we don't burn
* CPU to no purpose.
*/
if (++wait_cnt < 1000)
__wt_yield();
else {
sleep_cnt = WT_MIN(wait_cnt, 10000);
wait_cnt *= 2;
WT_STAT_FAST_CONN_INCRV(session, page_sleep, sleep_cnt);
__wt_sleep(0, sleep_cnt);
}
}
}
/*
* __wt_log_slot_join --
* Join a consolidated logging slot. Must be called with
* the read lock held.
*/
void
__wt_log_slot_join(WT_SESSION_IMPL *session, uint64_t mysize,
uint32_t flags, WT_MYSLOT *myslot)
{
WT_CONNECTION_IMPL *conn;
WT_LOG *log;
WT_LOGSLOT *slot;
int64_t flag_state, new_state, old_state, released;
int32_t join_offset, new_join;
#ifdef HAVE_DIAGNOSTIC
bool unbuf_force;
#endif
conn = S2C(session);
log = conn->log;
/*
* Make sure the length cannot overflow. The caller should not
* even call this function if it doesn't fit but use direct
* writes.
*/
WT_ASSERT(session, !F_ISSET(session, WT_SESSION_LOCKED_SLOT));
/*
* There should almost always be a slot open.
*/
#ifdef HAVE_DIAGNOSTIC
unbuf_force = (++log->write_calls % 1000) == 0;
#endif
for (;;) {
WT_BARRIER();
slot = log->active_slot;
old_state = slot->slot_state;
/*
* Try to join our size into the existing size and
* atomically write it back into the state.
*/
flag_state = WT_LOG_SLOT_FLAGS(old_state);
released = WT_LOG_SLOT_RELEASED(old_state);
join_offset = WT_LOG_SLOT_JOINED(old_state);
#ifdef HAVE_DIAGNOSTIC
if (unbuf_force || mysize > WT_LOG_SLOT_BUF_MAX) {
#else
if (mysize > WT_LOG_SLOT_BUF_MAX) {
#endif
new_join = join_offset + WT_LOG_SLOT_UNBUFFERED;
F_SET(myslot, WT_MYSLOT_UNBUFFERED);
myslot->slot = slot;
} else
new_join = join_offset + (int32_t)mysize;
new_state = (int64_t)WT_LOG_SLOT_JOIN_REL(
(int64_t)new_join, (int64_t)released, (int64_t)flag_state);
/*
* Check if the slot is open for joining and we are able to
* swap in our size into the state.
*/
if (WT_LOG_SLOT_OPEN(old_state) &&
__wt_atomic_casiv64(
&slot->slot_state, old_state, new_state))
break;
/*
* The slot is no longer open or we lost the race to
* update it. Yield and try again.
*/
WT_STAT_FAST_CONN_INCR(session, log_slot_races);
__wt_yield();
}
/*
* We joined this slot. Fill in our information to return to
* the caller.
*/
if (mysize != 0)
WT_STAT_FAST_CONN_INCR(session, log_slot_joins);
if (LF_ISSET(WT_LOG_DSYNC | WT_LOG_FSYNC))
F_SET(slot, WT_SLOT_SYNC_DIR);
if (LF_ISSET(WT_LOG_FLUSH))
F_SET(slot, WT_SLOT_FLUSH);
if (LF_ISSET(WT_LOG_FSYNC))
F_SET(slot, WT_SLOT_SYNC);
if (F_ISSET(myslot, WT_MYSLOT_UNBUFFERED)) {
WT_ASSERT(session, slot->slot_unbuffered == 0);
WT_STAT_FAST_CONN_INCR(session, log_slot_unbuffered);
slot->slot_unbuffered = (int64_t)mysize;
}
myslot->slot = slot;
myslot->offset = join_offset;
myslot->end_offset = (wt_off_t)((uint64_t)join_offset + mysize);
}
/*
* __wt_log_slot_release --
* Each thread in a consolidated group releases its portion to
* signal it has completed copying its piece of the log into
* the memory buffer.
*/
int64_t
__wt_log_slot_release(WT_SESSION_IMPL *session, WT_MYSLOT *myslot, int64_t size)
{
WT_LOGSLOT *slot;
wt_off_t cur_offset, my_start;
int64_t my_size, rel_size;
WT_UNUSED(session);
slot = myslot->slot;
my_start = slot->slot_start_offset + myslot->offset;
while ((cur_offset = slot->slot_last_offset) < my_start) {
/*
* Set our offset if we are larger.
*/
if (__wt_atomic_casiv64(
&slot->slot_last_offset, cur_offset, my_start))
break;
/*
* If we raced another thread updating this, try again.
*/
WT_BARRIER();
}
/*
* Add my size into the state and return the new size.
*/
rel_size = size;
if (F_ISSET(myslot, WT_MYSLOT_UNBUFFERED))
rel_size = WT_LOG_SLOT_UNBUFFERED;
my_size = (int64_t)WT_LOG_SLOT_JOIN_REL((int64_t)0, rel_size, 0);
return (__wt_atomic_addiv64(&slot->slot_state, my_size));
}
/*
* __wt_btcur_search_near --
* Search for a record in the tree.
*/
int
__wt_btcur_search_near(WT_CURSOR_BTREE *cbt, int *exactp)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_SESSION_IMPL *session;
WT_UPDATE *upd;
int exact, valid;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
upd = NULL; /* -Wuninitialized */
exact = 0;
WT_STAT_FAST_CONN_INCR(session, cursor_search_near);
WT_STAT_FAST_DATA_INCR(session, cursor_search_near);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
/*
* If we have a row-store page pinned, search it; if we don't have a
* page pinned, or the search of the pinned page doesn't find an exact
* match, search from the root. Unlike WT_CURSOR.search, ignore pinned
* pages in the case of column-store, search-near isn't an interesting
* enough case for column-store to add the complexity needed to avoid
* the tree search.
*
* Set the "insert" flag for the btree row-store search; we may intend
* to position the cursor at the end of the tree, rather than match an
* existing record.
*/
valid = 0;
if (btree->type == BTREE_ROW &&
F_ISSET(cbt, WT_CBT_ACTIVE) &&
cbt->ref->page->read_gen != WT_READGEN_OLDEST) {
__wt_txn_cursor_op(session);
WT_ERR(__cursor_row_search(session, cbt, cbt->ref, 1));
/*
* Search-near is trickier than search when searching an already
* pinned page. If search returns the first or last page slots,
* discard the results and search the full tree as the neighbor
* pages might offer better matches. This test is simplistic as
* we're ignoring append lists (there may be no page slots or we
* might be legitimately positioned after the last page slot).
* Ignore those cases, it makes things too complicated.
*/
if (cbt->slot != 0 &&
cbt->slot != cbt->ref->page->pg_row_entries - 1)
valid = __cursor_valid(cbt, &upd);
}
if (!valid) {
WT_ERR(__cursor_func_init(cbt, 1));
WT_ERR(btree->type == BTREE_ROW ?
__cursor_row_search(session, cbt, NULL, 1) :
__cursor_col_search(session, cbt, NULL));
valid = __cursor_valid(cbt, &upd);
}
/*
* If we find a valid key, return it.
*
* Else, creating a record past the end of the tree in a fixed-length
* column-store implicitly fills the gap with empty records. In this
* case, we instantiate the empty record, it's an exact match.
*
* Else, move to the next key in the tree (bias for prefix searches).
* Cursor next skips invalid rows, so we don't have to test for them
* again.
*
* Else, redo the search and move to the previous key in the tree.
* Cursor previous skips invalid rows, so we don't have to test for
* them again.
*
* If that fails, quit, there's no record to return.
*/
if (valid) {
exact = cbt->compare;
ret = __wt_kv_return(session, cbt, upd);
} else if (__cursor_fix_implicit(btree, cbt)) {
cbt->recno = cursor->recno;
cbt->v = 0;
cursor->value.data = &cbt->v;
cursor->value.size = 1;
exact = 0;
} else if ((ret = __wt_btcur_next(cbt, 0)) != WT_NOTFOUND)
exact = 1;
else {
WT_ERR(btree->type == BTREE_ROW ?
__cursor_row_search(session, cbt, NULL, 1) :
__cursor_col_search(session, cbt, NULL));
if (__cursor_valid(cbt, &upd)) {
exact = cbt->compare;
ret = __wt_kv_return(session, cbt, upd);
} else if ((ret = __wt_btcur_prev(cbt, 0)) != WT_NOTFOUND)
exact = -1;
}
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
if (exactp != NULL && (ret == 0 || ret == WT_NOTFOUND))
*exactp = exact;
return (ret);
}
/*
* __log_slot_close --
* Close out the slot the caller is using. The slot may already be
* closed or freed by another thread.
*/
static int
__log_slot_close(
WT_SESSION_IMPL *session, WT_LOGSLOT *slot, bool *releasep, bool forced)
{
WT_CONNECTION_IMPL *conn;
WT_LOG *log;
int64_t end_offset, new_state, old_state;
WT_ASSERT(session, F_ISSET(session, WT_SESSION_LOCKED_SLOT));
WT_ASSERT(session, releasep != NULL);
conn = S2C(session);
log = conn->log;
*releasep = 0;
if (slot == NULL)
return (WT_NOTFOUND);
retry:
old_state = slot->slot_state;
/*
* If this close is coming from a forced close and a thread is in
* the middle of using the slot, return EBUSY. The caller can
* decide if retrying is necessary or not.
*/
if (forced && WT_LOG_SLOT_INPROGRESS(old_state))
return (EBUSY);
/*
* If someone else is switching out this slot we lost. Nothing to
* do but return. Return WT_NOTFOUND anytime the given slot was
* processed by another closing thread. Only return 0 when we
* actually closed the slot.
*/
if (WT_LOG_SLOT_CLOSED(old_state))
return (WT_NOTFOUND);
/*
* If someone completely processed this slot, we're done.
*/
if (FLD64_ISSET((uint64_t)slot->slot_state, WT_LOG_SLOT_RESERVED))
return (WT_NOTFOUND);
new_state = (old_state | WT_LOG_SLOT_CLOSE);
/*
* Close this slot. If we lose the race retry.
*/
if (!__wt_atomic_casiv64(&slot->slot_state, old_state, new_state))
goto retry;
/*
* We own the slot now. No one else can join.
* Set the end LSN.
*/
WT_STAT_FAST_CONN_INCR(session, log_slot_closes);
if (WT_LOG_SLOT_DONE(new_state))
*releasep = 1;
slot->slot_end_lsn = slot->slot_start_lsn;
end_offset =
WT_LOG_SLOT_JOINED_BUFFERED(old_state) + slot->slot_unbuffered;
slot->slot_end_lsn.offset += (wt_off_t)end_offset;
WT_STAT_FAST_CONN_INCRV(session,
log_slot_consolidated, end_offset);
/*
* XXX Would like to change so one piece of code advances the LSN.
*/
log->alloc_lsn = slot->slot_end_lsn;
WT_ASSERT(session, log->alloc_lsn.file >= log->write_lsn.file);
return (0);
}
/*
* __wt_btcur_remove --
* Remove a record from the tree.
*/
int
__wt_btcur_remove(WT_CURSOR_BTREE *cbt)
{
WT_BTREE *btree;
WT_CURSOR *cursor;
WT_DECL_RET;
WT_SESSION_IMPL *session;
btree = cbt->btree;
cursor = &cbt->iface;
session = (WT_SESSION_IMPL *)cursor->session;
WT_STAT_FAST_CONN_INCR(session, cursor_remove);
WT_STAT_FAST_DATA_INCR(session, cursor_remove);
WT_STAT_FAST_DATA_INCRV(session, cursor_remove_bytes, cursor->key.size);
if (btree->type == BTREE_ROW)
WT_RET(__cursor_size_chk(session, &cursor->key));
retry: WT_RET(__cursor_func_init(cbt, 1));
switch (btree->type) {
case BTREE_COL_FIX:
case BTREE_COL_VAR:
WT_ERR(__cursor_col_search(session, cbt, NULL));
/*
* If we find a matching record, check whether an update would
* conflict. Do this before checking if the update is visible
* in __cursor_valid, or we can miss conflict.
*/
WT_ERR(__curfile_update_check(cbt));
/* Remove the record if it exists. */
if (cbt->compare != 0 || !__cursor_valid(cbt, NULL)) {
if (!__cursor_fix_implicit(btree, cbt))
WT_ERR(WT_NOTFOUND);
/*
* Creating a record past the end of the tree in a
* fixed-length column-store implicitly fills the
* gap with empty records. Return success in that
* case, the record was deleted successfully.
*
* Correct the btree cursor's location: the search
* will have pointed us at the previous/next item,
* and that's not correct.
*/
cbt->recno = cursor->recno;
} else
ret = __cursor_col_modify(session, cbt, 1);
break;
case BTREE_ROW:
/* Remove the record if it exists. */
WT_ERR(__cursor_row_search(session, cbt, NULL, 0));
/* Check whether an update would conflict. */
WT_ERR(__curfile_update_check(cbt));
if (cbt->compare != 0 || !__cursor_valid(cbt, NULL))
WT_ERR(WT_NOTFOUND);
ret = __cursor_row_modify(session, cbt, 1);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
err: if (ret == WT_RESTART) {
WT_STAT_FAST_CONN_INCR(session, cursor_restart);
WT_STAT_FAST_DATA_INCR(session, cursor_restart);
goto retry;
}
/*
* If the cursor is configured to overwrite and the record is not
* found, that is exactly what we want.
*/
if (F_ISSET(cursor, WT_CURSTD_OVERWRITE) && ret == WT_NOTFOUND)
ret = 0;
if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __wt_cond_wait_signal --
* Wait on a mutex, optionally timing out. If we get it
* before the time out period expires, let the caller know.
*/
int
__wt_cond_wait_signal(
WT_SESSION_IMPL *session, WT_CONDVAR *cond, uint64_t usecs, bool *signalled)
{
BOOL sleepret;
DWORD milliseconds, windows_error;
bool locked;
uint64_t milliseconds64;
locked = false;
/* Fast path if already signalled. */
*signalled = true;
if (__wt_atomic_addi32(&cond->waiters, 1) == 0)
return (0);
/*
* !!!
* This function MUST handle a NULL session handle.
*/
if (session != NULL) {
WT_RET(__wt_verbose(session, WT_VERB_MUTEX,
"wait %s cond (%p)", cond->name, cond));
WT_STAT_FAST_CONN_INCR(session, cond_wait);
}
EnterCriticalSection(&cond->mtx);
locked = true;
if (usecs > 0) {
milliseconds64 = usecs / 1000;
/*
* Check for 32-bit unsigned integer overflow
* INFINITE is max unsigned int on Windows
*/
if (milliseconds64 >= INFINITE)
milliseconds64 = INFINITE - 1;
milliseconds = (DWORD)milliseconds64;
/*
* 0 would mean the CV sleep becomes a TryCV which we do not
* want
*/
if (milliseconds == 0)
milliseconds = 1;
sleepret = SleepConditionVariableCS(
&cond->cond, &cond->mtx, milliseconds);
} else
sleepret = SleepConditionVariableCS(
&cond->cond, &cond->mtx, INFINITE);
/*
* SleepConditionVariableCS returns non-zero on success, 0 on timeout
* or failure.
*/
if (sleepret == 0) {
windows_error = __wt_getlasterror();
if (windows_error == ERROR_TIMEOUT) {
*signalled = false;
sleepret = 1;
}
}
(void)__wt_atomic_subi32(&cond->waiters, 1);
if (locked)
LeaveCriticalSection(&cond->mtx);
if (sleepret != 0)
return (0);
__wt_errx(session, "SleepConditionVariableCS: %s",
__wt_formatmessage(session, windows_error));
return (__wt_map_windows_error(windows_error));
}
/*
* __wt_btcur_next --
* Move to the next record in the tree.
*/
int
__wt_btcur_next(WT_CURSOR_BTREE *cbt, int truncating)
{
WT_DECL_RET;
WT_PAGE *page;
WT_SESSION_IMPL *session;
uint32_t flags;
int newpage;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_STAT_FAST_CONN_INCR(session, cursor_next);
WT_STAT_FAST_DATA_INCR(session, cursor_next);
flags = WT_READ_SKIP_INTL; /* Tree walk flags. */
if (truncating)
LF_SET(WT_READ_TRUNCATE);
WT_RET(__cursor_func_init(cbt, 0));
/*
* If we aren't already iterating in the right direction, there's
* some setup to do.
*/
if (!F_ISSET(cbt, WT_CBT_ITERATE_NEXT))
__wt_btcur_iterate_setup(cbt, 1);
/*
* Walk any page we're holding until the underlying call returns not-
* found. Then, move to the next page, until we reach the end of the
* file.
*/
page = cbt->ref == NULL ? NULL : cbt->ref->page;
for (newpage = 0;; newpage = 1) {
if (F_ISSET(cbt, WT_CBT_ITERATE_APPEND)) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_append_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_append_next(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret == 0)
break;
F_CLR(cbt, WT_CBT_ITERATE_APPEND);
if (ret != WT_NOTFOUND)
break;
} else if (page != NULL) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_next(cbt, newpage);
break;
case WT_PAGE_ROW_LEAF:
ret = __cursor_row_next(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret != WT_NOTFOUND)
break;
/*
* The last page in a column-store has appended entries.
* We handle it separately from the usual cursor code:
* it's only that one page and it's in a simple format.
*/
if (page->type != WT_PAGE_ROW_LEAF &&
(cbt->ins_head = WT_COL_APPEND(page)) != NULL) {
F_SET(cbt, WT_CBT_ITERATE_APPEND);
continue;
}
}
WT_ERR(__wt_tree_walk(session, &cbt->ref, flags));
WT_ERR_TEST(cbt->ref == NULL, WT_NOTFOUND);
page = cbt->ref->page;
WT_ASSERT(session,
page->type != WT_PAGE_COL_INT &&
page->type != WT_PAGE_ROW_INT);
}
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __page_read --
* Read a page from the file.
*/
static int
__page_read(WT_SESSION_IMPL *session, WT_REF *ref)
{
const WT_PAGE_HEADER *dsk;
WT_BTREE *btree;
WT_DECL_RET;
WT_ITEM tmp;
WT_PAGE *page;
size_t addr_size;
uint32_t previous_state;
const uint8_t *addr;
btree = S2BT(session);
page = NULL;
/*
* Don't pass an allocated buffer to the underlying block read function,
* force allocation of new memory of the appropriate size.
*/
WT_CLEAR(tmp);
/*
* Attempt to set the state to WT_REF_READING for normal reads, or
* WT_REF_LOCKED, for deleted pages. If successful, we've won the
* race, read the page.
*/
if (__wt_atomic_casv32(&ref->state, WT_REF_DISK, WT_REF_READING))
previous_state = WT_REF_DISK;
else if (__wt_atomic_casv32(&ref->state, WT_REF_DELETED, WT_REF_LOCKED))
previous_state = WT_REF_DELETED;
else
return (0);
/*
* Get the address: if there is no address, the page was deleted, but a
* subsequent search or insert is forcing re-creation of the name space.
*/
WT_ERR(__wt_ref_info(session, ref, &addr, &addr_size, NULL));
if (addr == NULL) {
WT_ASSERT(session, previous_state == WT_REF_DELETED);
WT_ERR(__wt_btree_new_leaf_page(session, &page));
ref->page = page;
goto done;
}
/*
* There's an address, read or map the backing disk page and build an
* in-memory version of the page.
*/
WT_ERR(__wt_bt_read(session, &tmp, addr, addr_size));
WT_ERR(__wt_page_inmem(session, ref, tmp.data, tmp.memsize,
WT_DATA_IN_ITEM(&tmp) ?
WT_PAGE_DISK_ALLOC : WT_PAGE_DISK_MAPPED, &page));
/*
* Clear the local reference to an allocated copy of the disk image on
* return; the page steals it, errors in this code should not free it.
*/
tmp.mem = NULL;
/*
* If reading for a checkpoint, there's no additional work to do, the
* page on disk is correct as written.
*/
if (session->dhandle->checkpoint != NULL)
goto done;
/* If the page was deleted, instantiate that information. */
if (previous_state == WT_REF_DELETED)
WT_ERR(__wt_delete_page_instantiate(session, ref));
/*
* Instantiate updates from the database's lookaside table. The page
* flag was set when the page was written, potentially a long time ago.
* We only care if the lookaside table is currently active, check that
* before doing any work.
*/
dsk = tmp.data;
if (F_ISSET(dsk, WT_PAGE_LAS_UPDATE) && __wt_las_is_written(session)) {
WT_STAT_FAST_CONN_INCR(session, cache_read_lookaside);
WT_STAT_FAST_DATA_INCR(session, cache_read_lookaside);
WT_ERR(__las_page_instantiate(
session, ref, btree->id, addr, addr_size));
}
done:
WT_PUBLISH(ref->state, WT_REF_MEM);
return (0);
err: /*
* If the function building an in-memory version of the page failed,
* it discarded the page, but not the disk image. Discard the page
* and separately discard the disk image in all cases.
*/
if (ref->page != NULL)
__wt_ref_out(session, ref);
WT_PUBLISH(ref->state, previous_state);
__wt_buf_free(session, &tmp);
return (ret);
}
/*
* __wt_btcur_next --
* Move to the next record in the tree.
*/
int
__wt_btcur_next(WT_CURSOR_BTREE *cbt, int truncating)
{
WT_DECL_RET;
WT_PAGE *page;
WT_SESSION_IMPL *session;
uint32_t flags;
int newpage;
session = (WT_SESSION_IMPL *)cbt->iface.session;
WT_STAT_FAST_CONN_INCR(session, cursor_next);
WT_STAT_FAST_DATA_INCR(session, cursor_next);
flags = WT_READ_SKIP_INTL; /* Tree walk flags. */
if (truncating)
LF_SET(WT_READ_TRUNCATE);
WT_RET(__cursor_func_init(cbt, 0));
/*
* If we aren't already iterating in the right direction, there's
* some setup to do.
*/
if (!F_ISSET(cbt, WT_CBT_ITERATE_NEXT))
__wt_btcur_iterate_setup(cbt, 1);
/*
* Walk any page we're holding until the underlying call returns not-
* found. Then, move to the next page, until we reach the end of the
* file.
*/
for (newpage = 0;; newpage = 1) {
page = cbt->ref == NULL ? NULL : cbt->ref->page;
WT_ASSERT(session, page == NULL || !WT_PAGE_IS_INTERNAL(page));
if (F_ISSET(cbt, WT_CBT_ITERATE_APPEND)) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_append_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_append_next(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret == 0)
break;
F_CLR(cbt, WT_CBT_ITERATE_APPEND);
if (ret != WT_NOTFOUND)
break;
} else if (page != NULL) {
switch (page->type) {
case WT_PAGE_COL_FIX:
ret = __cursor_fix_next(cbt, newpage);
break;
case WT_PAGE_COL_VAR:
ret = __cursor_var_next(cbt, newpage);
break;
case WT_PAGE_ROW_LEAF:
ret = __cursor_row_next(cbt, newpage);
break;
WT_ILLEGAL_VALUE_ERR(session);
}
if (ret != WT_NOTFOUND)
break;
/*
* The last page in a column-store has appended entries.
* We handle it separately from the usual cursor code:
* it's only that one page and it's in a simple format.
*/
if (page->type != WT_PAGE_ROW_LEAF &&
(cbt->ins_head = WT_COL_APPEND(page)) != NULL) {
F_SET(cbt, WT_CBT_ITERATE_APPEND);
continue;
}
}
/*
* If we saw a lot of deleted records on this page, or we went
* all the way through a page and only saw deleted records, try
* to evict the page when we release it. Otherwise repeatedly
* deleting from the beginning of a tree can have quadratic
* performance. Take care not to force eviction of pages that
* are genuinely empty, in new trees.
*/
if (page != NULL &&
(cbt->page_deleted_count > WT_BTREE_DELETE_THRESHOLD ||
(newpage && cbt->page_deleted_count > 0)))
__wt_page_evict_soon(page);
cbt->page_deleted_count = 0;
WT_ERR(__wt_tree_walk(session, &cbt->ref, NULL, flags));
WT_ERR_TEST(cbt->ref == NULL, WT_NOTFOUND);
}
err: if (ret != 0)
WT_TRET(__cursor_reset(cbt));
return (ret);
}
/*
* __wt_curfile_create --
* Open a cursor for a given btree handle.
*/
int
__wt_curfile_create(WT_SESSION_IMPL *session,
WT_CURSOR *owner, const char *cfg[], bool bulk, bool bitmap,
WT_CURSOR **cursorp)
{
WT_CURSOR_STATIC_INIT(iface,
__wt_cursor_get_key, /* get-key */
__wt_cursor_get_value, /* get-value */
__wt_cursor_set_key, /* set-key */
__wt_cursor_set_value, /* set-value */
__curfile_compare, /* compare */
__curfile_equals, /* equals */
__curfile_next, /* next */
__curfile_prev, /* prev */
__curfile_reset, /* reset */
__curfile_search, /* search */
__curfile_search_near, /* search-near */
__curfile_insert, /* insert */
__curfile_update, /* update */
__curfile_remove, /* remove */
__wt_cursor_reconfigure, /* reconfigure */
__curfile_close); /* close */
WT_BTREE *btree;
WT_CONFIG_ITEM cval;
WT_CURSOR *cursor;
WT_CURSOR_BTREE *cbt;
WT_CURSOR_BULK *cbulk;
WT_DECL_RET;
size_t csize;
WT_STATIC_ASSERT(offsetof(WT_CURSOR_BTREE, iface) == 0);
cbt = NULL;
btree = S2BT(session);
WT_ASSERT(session, btree != NULL);
csize = bulk ? sizeof(WT_CURSOR_BULK) : sizeof(WT_CURSOR_BTREE);
WT_RET(__wt_calloc(session, 1, csize, &cbt));
cursor = &cbt->iface;
*cursor = iface;
cursor->session = &session->iface;
cursor->internal_uri = btree->dhandle->name;
cursor->key_format = btree->key_format;
cursor->value_format = btree->value_format;
cbt->btree = btree;
if (bulk) {
F_SET(cursor, WT_CURSTD_BULK);
cbulk = (WT_CURSOR_BULK *)cbt;
/* Optionally skip the validation of each bulk-loaded key. */
WT_ERR(__wt_config_gets_def(
session, cfg, "skip_sort_check", 0, &cval));
WT_ERR(__wt_curbulk_init(
session, cbulk, bitmap, cval.val == 0 ? 0 : 1));
}
/*
* random_retrieval
* Random retrieval cursors only support next, reset and close.
*/
WT_ERR(__wt_config_gets_def(session, cfg, "next_random", 0, &cval));
if (cval.val != 0) {
__wt_cursor_set_notsup(cursor);
cursor->next = __curfile_next_random;
cursor->reset = __curfile_reset;
}
/* __wt_cursor_init is last so we don't have to clean up on error. */
WT_ERR(__wt_cursor_init(
cursor, cursor->internal_uri, owner, cfg, cursorp));
WT_STAT_FAST_CONN_INCR(session, cursor_create);
WT_STAT_FAST_DATA_INCR(session, cursor_create);
if (0) {
err:
__wt_free(session, cbt);
}
return (ret);
}
