/*
* __wt_cond_signal --
* Signal a waiting thread.
*/
int
__wt_cond_signal(WT_SESSION_IMPL *session, WT_CONDVAR *cond)
{
WT_DECL_RET;
int locked;
locked = 0;
/*
* !!!
* This function MUST handle a NULL session handle.
*/
if (session != NULL && WT_VERBOSE_ISSET(session, mutex))
WT_RET(__wt_verbose(
session, "signal %s cond (%p)", cond->name, cond));
/* Fast path if already signalled. */
if (cond->waiters == -1)
return (0);
if (cond->waiters > 0 || !WT_ATOMIC_CAS(cond->waiters, 0, -1)) {
WT_ERR(pthread_mutex_lock(&cond->mtx));
locked = 1;
WT_ERR(pthread_cond_broadcast(&cond->cond));
}
err: if (locked)
WT_TRET(pthread_mutex_unlock(&cond->mtx));
if (ret == 0)
return (0);
WT_RET_MSG(session, ret, "pthread_cond_broadcast");
}
/*
* __wt_rec_track_init --
* Initialize the page's list of tracked objects when reconciliation
* starts.
*/
int
__wt_rec_track_init(WT_SESSION_IMPL *session, WT_PAGE *page)
{
if (WT_VERBOSE_ISSET(session, reconcile))
WT_RET(__track_dump(session, page, "reconcile init"));
return (0);
}
/*
* __wt_ovfl_reuse_search --
* Search the page's list of overflow records for a match.
*/
int
__wt_ovfl_reuse_search(WT_SESSION_IMPL *session, WT_PAGE *page,
uint8_t **addrp, size_t *addr_sizep,
const void *value, size_t value_size)
{
WT_OVFL_REUSE **head, *reuse;
*addrp = NULL;
*addr_sizep = 0;
if (page->modify->ovfl_track == NULL)
return (0);
head = page->modify->ovfl_track->ovfl_reuse;
/*
* The search function returns the first matching record in the list
* which does not have the in-use flag set, or NULL.
*/
if ((reuse = __ovfl_reuse_skip_search(head, value, value_size)) == NULL)
return (0);
*addrp = WT_OVFL_REUSE_ADDR(reuse);
*addr_sizep = reuse->addr_size;
F_SET(reuse, WT_OVFL_REUSE_INUSE);
if (WT_VERBOSE_ISSET(session, WT_VERB_OVERFLOW))
WT_RET(__ovfl_reuse_verbose(session, page, reuse, "reclaim"));
return (1);
}
/*
* __wt_cond_signal --
* Signal a waiting thread.
*/
int
__wt_cond_signal(WT_SESSION_IMPL *session, WT_CONDVAR *cond)
{
WT_DECL_RET;
int locked;
locked = 0;
/*
* !!!
* This function MUST handle a NULL session handle.
*/
if (session != NULL && WT_VERBOSE_ISSET(session, mutex))
WT_RET(__wt_verbose(
session, "signal %s cond (%p)", cond->name, cond));
WT_ERR(pthread_mutex_lock(&cond->mtx));
locked = 1;
if (!cond->signalled) {
cond->signalled = 1;
WT_ERR(pthread_cond_signal(&cond->cond));
}
err: if (locked)
WT_TRET(pthread_mutex_unlock(&cond->mtx));
if (ret == 0)
return (0);
WT_RET_MSG(session, ret, "pthread_cond_signal");
}
/*
* __wt_ovfl_txnc_add --
* Add a new entry to the page's list of transaction-cached overflow
* records.
*/
int
__wt_ovfl_txnc_add(WT_SESSION_IMPL *session, WT_PAGE *page,
const uint8_t *addr, size_t addr_size,
const void *value, size_t value_size)
{
WT_OVFL_TXNC **head, **stack[WT_SKIP_MAXDEPTH], *txnc;
size_t size;
u_int i, skipdepth;
uint8_t *p;
if (page->modify->ovfl_track == NULL)
WT_RET(__ovfl_track_init(session, page));
head = page->modify->ovfl_track->ovfl_txnc;
/* Choose a skiplist depth for this insert. */
skipdepth = __wt_skip_choose_depth(session);
/*
* Allocate the WT_OVFL_TXNC structure, next pointers for the skip
* list, room for the address and value, then copy everything into
* place.
*
* To minimize the WT_OVFL_TXNC structure size, the address offset
* and size are single bytes: that's safe because the address follows
* the structure (which can't be more than about 100B), and address
* cookies are limited to 255B.
*/
size = sizeof(WT_OVFL_TXNC) +
skipdepth * sizeof(WT_OVFL_TXNC *) + addr_size + value_size;
WT_RET(__wt_calloc(session, 1, size, &txnc));
p = (uint8_t *)txnc +
sizeof(WT_OVFL_TXNC) + skipdepth * sizeof(WT_OVFL_TXNC *);
txnc->addr_offset = (uint8_t)WT_PTRDIFF(p, txnc);
txnc->addr_size = (uint8_t)addr_size;
memcpy(p, addr, addr_size);
p += addr_size;
txnc->value_offset = WT_PTRDIFF32(p, txnc);
txnc->value_size = WT_STORE_SIZE(value_size);
memcpy(p, value, value_size);
txnc->current = __wt_txn_new_id(session);
__wt_cache_page_inmem_incr(
session, page, WT_OVFL_SIZE(txnc, WT_OVFL_TXNC));
/* Insert the new entry into the skiplist. */
__ovfl_txnc_skip_search_stack(head, stack, addr, addr_size);
for (i = 0; i < skipdepth; ++i) {
txnc->next[i] = *stack[i];
*stack[i] = txnc;
}
if (WT_VERBOSE_ISSET(session, WT_VERB_OVERFLOW))
WT_RET(__ovfl_txnc_verbose(session, page, txnc, "add"));
return (0);
}
/*
* __recovery_set_checkpoint_timestamp --
* Set the checkpoint timestamp as retrieved from the metadata file.
*/
static int
__recovery_set_checkpoint_timestamp(WT_RECOVERY *r)
{
WT_CONFIG_ITEM cval;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_SESSION_IMPL *session;
wt_timestamp_t ckpt_timestamp;
char ts_string[WT_TS_INT_STRING_SIZE], *sys_config;
sys_config = NULL;
session = r->session;
conn = S2C(session);
/*
* Read the system checkpoint information from the metadata file and
* save the stable timestamp of the last checkpoint for later query.
* This gets saved in the connection.
*/
ckpt_timestamp = 0;
/* Search in the metadata for the system information. */
WT_ERR_NOTFOUND_OK(
__wt_metadata_search(session, WT_SYSTEM_CKPT_URI, &sys_config));
if (sys_config != NULL) {
WT_CLEAR(cval);
WT_ERR_NOTFOUND_OK(__wt_config_getones(
session, sys_config, "checkpoint_timestamp", &cval));
if (cval.len != 0) {
__wt_verbose(session, WT_VERB_RECOVERY,
"Recovery timestamp %.*s",
(int)cval.len, cval.str);
WT_ERR(__wt_txn_parse_timestamp_raw(session,
"recovery", &ckpt_timestamp, &cval));
}
}
/*
* Set the recovery checkpoint timestamp and the metadata checkpoint
* timestamp so that the checkpoint after recovery writes the correct
* value into the metadata.
*/
conn->txn_global.meta_ckpt_timestamp =
conn->txn_global.recovery_timestamp = ckpt_timestamp;
if (WT_VERBOSE_ISSET(session,
WT_VERB_RECOVERY | WT_VERB_RECOVERY_PROGRESS)) {
__wt_timestamp_to_string(
conn->txn_global.recovery_timestamp, ts_string);
__wt_verbose(session,
WT_VERB_RECOVERY | WT_VERB_RECOVERY_PROGRESS,
"Set global recovery timestamp: %s", ts_string);
}
err: __wt_free(session, sys_config);
return (ret);
}
/*
* __wt_block_compact_skip --
* Return if compaction will shrink the file.
*/
int
__wt_block_compact_skip(WT_SESSION_IMPL *session, WT_BLOCK *block, int *skipp)
{
WT_DECL_RET;
WT_EXT *ext;
WT_EXTLIST *el;
WT_FH *fh;
off_t avail, ninety;
*skipp = 1; /* Return a default skip. */
fh = block->fh;
/*
* We do compaction by copying blocks from the end of the file to the
* beginning of the file, and we need some metrics to decide if it's
* worth doing. Ignore small files, and files where we are unlikely
* to recover 10% of the file.
*/
if (fh->size <= 10 * 1024)
return (0);
__wt_spin_lock(session, &block->live_lock);
if (WT_VERBOSE_ISSET(session, compact))
WT_ERR(__block_dump_avail(session, block));
/* Sum the number of available bytes in the first 90% of the file. */
avail = 0;
ninety = fh->size - fh->size / 10;
el = &block->live.avail;
WT_EXT_FOREACH(ext, el->off)
if (ext->off < ninety)
avail += ext->size;
/*
* If at least 10% of the total file is available and in the first 90%
* of the file, we'll try compaction.
*/
if (avail >= fh->size / 10)
*skipp = 0;
WT_VERBOSE_ERR(session, compact,
"%s: %" PRIuMAX "MB (%" PRIuMAX ") available space in the first "
"90%% of the file, require 10%% or %" PRIuMAX "MB (%" PRIuMAX
") to perform compaction, compaction %s",
block->name,
(uintmax_t)avail / WT_MEGABYTE, (uintmax_t)avail,
(uintmax_t)(fh->size / 10) / WT_MEGABYTE, (uintmax_t)fh->size / 10,
*skipp ? "skipped" : "proceeding");
err: __wt_spin_unlock(session, &block->live_lock);
return (ret);
}
/*
* __wt_rec_track_ovfl_reuse --
* Search for a matching overflow record and reactivate it.
*/
int
__wt_rec_track_ovfl_reuse(
WT_SESSION_IMPL *session, WT_PAGE *page,
const void *data, uint32_t data_size,
uint8_t **addrp, uint32_t *addr_sizep, int *foundp)
{
WT_PAGE_MODIFY *mod;
WT_PAGE_TRACK *track;
uint32_t i;
*foundp = 0;
mod = page->modify;
for (track = mod->track, i = 0; i < mod->track_entries; ++track, ++i) {
/* Ignore empty slots */
if (!F_ISSET(track, WT_TRK_OBJECT))
continue;
/*
* Ignore discarded objects, objects already in-use, or cached
* overflow values. We don't care about whether or not the
* object came from a page, we can re-use objects from the page
* or objects created in a previous reconciliation.
*/
if (F_ISSET(track,
WT_TRK_DISCARD | WT_TRK_INUSE | WT_TRK_OVFL_VALUE))
continue;
/*
* Ignore objects without data (must be block objects). This is
* not really necessary (presumably, our caller is matching on a
* non-zero-length data item), but paranoia is healthy.
*/
if (track->data == NULL)
continue;
/* Check to see if the data matches. */
if (track->size != data_size ||
memcmp(data, track->data, data_size) != 0)
continue;
/*
* Reactivate the record.
* Return the block addr/size pair to our caller.
*/
F_SET(track, WT_TRK_INUSE);
*addrp = track->addr.addr;
*addr_sizep = track->addr.size;
*foundp = 1;
if (WT_VERBOSE_ISSET(session, reconcile))
WT_RET(__track_msg(
session, page, "reactivate overflow", track));
return (0);
}
return (0);
}
/*
* __ovfl_txnc_wrapup --
* Resolve the page's transaction-cache list.
*/
static int
__ovfl_txnc_wrapup(WT_SESSION_IMPL *session, WT_PAGE *page)
{
WT_OVFL_TXNC **e, **head, *txnc;
uint64_t oldest_txn;
size_t decr;
int i;
head = page->modify->ovfl_track->ovfl_txnc;
/*
* Take a snapshot of the oldest transaction ID we need to keep alive.
* Since we do two passes through entries in the structure, the normal
* visibility check could give different results as the global ID moves
* forward.
*/
oldest_txn = __wt_txn_oldest_id(session);
/*
* Discard any transaction-cache records with transaction IDs earlier
* than any in the system.
*
* First, walk the overflow transaction-cache skip lists (except for
* the lowest level), fixing up links.
*/
for (i = WT_SKIP_MAXDEPTH - 1; i > 0; --i)
for (e = &head[i]; (txnc = *e) != NULL;) {
if (WT_TXNID_LE(oldest_txn, txnc->current)) {
e = &txnc->next[i];
continue;
}
*e = txnc->next[i];
}
/* Second, discard any no longer needed transaction-cache records. */
decr = 0;
for (e = &head[0]; (txnc = *e) != NULL;) {
if (WT_TXNID_LE(oldest_txn, txnc->current)) {
e = &txnc->next[0];
continue;
}
*e = txnc->next[0];
if (WT_VERBOSE_ISSET(session, WT_VERB_OVERFLOW))
WT_RET(
__ovfl_txnc_verbose(session, page, txnc, "free"));
decr += WT_OVFL_SIZE(txnc, WT_OVFL_TXNC);
__wt_free(session, txnc);
}
if (decr != 0)
__wt_cache_page_inmem_decr(session, page, decr);
return (0);
}
/*
* __ovfl_reuse_wrapup_err --
* Resolve the page's overflow reuse list after an error occurs.
*/
static int
__ovfl_reuse_wrapup_err(WT_SESSION_IMPL *session, WT_PAGE *page)
{
WT_BM *bm;
WT_DECL_RET;
WT_OVFL_REUSE **e, **head, *reuse;
size_t decr;
int i;
bm = S2BT(session)->bm;
head = page->modify->ovfl_track->ovfl_reuse;
/*
* Discard any overflow records that were just added, freeing underlying
* blocks.
*
* First, walk the overflow reuse lists (except for the lowest one),
* fixing up skiplist links.
*/
for (i = WT_SKIP_MAXDEPTH - 1; i > 0; --i)
for (e = &head[i]; (reuse = *e) != NULL;) {
if (!F_ISSET(reuse, WT_OVFL_REUSE_JUST_ADDED)) {
e = &reuse->next[i];
continue;
}
*e = reuse->next[i];
}
/*
* Second, discard any overflow record with a just-added flag, clear the
* flags for the next run.
*/
decr = 0;
for (e = &head[0]; (reuse = *e) != NULL;) {
if (!F_ISSET(reuse, WT_OVFL_REUSE_JUST_ADDED)) {
F_CLR(reuse, WT_OVFL_REUSE_INUSE);
e = &reuse->next[0];
continue;
}
*e = reuse->next[0];
if (WT_VERBOSE_ISSET(session, WT_VERB_OVERFLOW))
WT_RET(
__ovfl_reuse_verbose(session, page, reuse, "free"));
WT_TRET(bm->free(
bm, session, WT_OVFL_REUSE_ADDR(reuse), reuse->addr_size));
decr += WT_OVFL_SIZE(reuse, WT_OVFL_REUSE);
__wt_free(session, reuse);
}
if (decr != 0)
__wt_cache_page_inmem_decr(session, page, decr);
return (0);
}
/*
* __ovfl_txnc_wrapup --
* Resolve the page's transaction-cache list.
*/
static int
__ovfl_txnc_wrapup(WT_SESSION_IMPL *session, WT_PAGE *page)
{
WT_OVFL_TXNC **e, **head, *txnc;
size_t decr;
int i;
head = page->modify->ovfl_track->ovfl_txnc;
/*
* Discard any transaction-cache records with transaction IDs earlier
* than any in the system.
*
* First, walk the overflow transaction-cache skip lists (except for
* the lowest level), fixing up links.
*/
for (i = WT_SKIP_MAXDEPTH - 1; i > 0; --i)
for (e = &head[i]; *e != NULL;) {
if (!__wt_txn_visible_all(session, (*e)->current)) {
e = &(*e)->next[i];
continue;
}
*e = (*e)->next[i];
}
/* Second, discard any no longer needed transaction-cache records. */
decr = 0;
for (e = &head[0]; (txnc = *e) != NULL;) {
if (!__wt_txn_visible_all(session, txnc->current)) {
e = &(*e)->next[0];
continue;
}
*e = (*e)->next[0];
decr += WT_OVFL_SIZE(WT_OVFL_TXNC) +
txnc->addr_size + txnc->value_size;
if (WT_VERBOSE_ISSET(session, WT_VERB_OVERFLOW))
WT_RET(
__ovfl_txnc_verbose(session, page, txnc, "free"));
__wt_free(session, txnc);
}
if (decr != 0)
__wt_cache_page_inmem_decr(session, page, decr);
return (0);
}
/*
* __wt_ovfl_discard_add --
* Add a new entry to the page's list of overflow records that have been
* discarded.
*/
int
__wt_ovfl_discard_add(WT_SESSION_IMPL *session, WT_PAGE *page, WT_CELL *cell)
{
WT_OVFL_TRACK *track;
if (page->modify->ovfl_track == NULL)
WT_RET(__ovfl_track_init(session, page));
track = page->modify->ovfl_track;
WT_RET(__wt_realloc_def(session, &track->discard_allocated,
track->discard_entries + 1, &track->discard));
track->discard[track->discard_entries++] = cell;
if (WT_VERBOSE_ISSET(session, WT_VERB_OVERFLOW))
WT_RET(__ovfl_discard_verbose(session, page, cell, "add"));
return (0);
}
/*
* __ovfl_discard_wrapup --
* Resolve the page's overflow discard list after a page is written.
*/
static int
__ovfl_discard_wrapup(WT_SESSION_IMPL *session, WT_PAGE *page)
{
WT_CELL **cellp;
WT_DECL_RET;
WT_OVFL_TRACK *track;
uint32_t i;
track = page->modify->ovfl_track;
for (i = 0, cellp = track->discard;
i < track->discard_entries; ++i, ++cellp) {
if (WT_VERBOSE_ISSET(session, WT_VERB_OVERFLOW))
WT_RET(__ovfl_discard_verbose(
session, page, *cellp, "free"));
/* Discard each cell's overflow item. */
WT_RET(__wt_ovfl_discard(session, *cellp));
}
__wt_free(session, track->discard);
track->discard_entries = track->discard_allocated = 0;
return (ret);
}
/*
* __sync_file --
* Flush pages for a specific file.
*/
static int
__sync_file(WT_SESSION_IMPL *session, WT_CACHE_OP syncop)
{
struct timespec end, start;
WT_BTREE *btree;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_PAGE *page;
WT_PAGE_MODIFY *mod;
WT_REF *walk;
WT_TXN *txn;
uint64_t internal_bytes, internal_pages, leaf_bytes, leaf_pages;
uint64_t oldest_id, saved_snap_min;
uint32_t flags;
conn = S2C(session);
btree = S2BT(session);
walk = NULL;
txn = &session->txn;
saved_snap_min = WT_SESSION_TXN_STATE(session)->snap_min;
flags = WT_READ_CACHE | WT_READ_NO_GEN;
internal_bytes = leaf_bytes = 0;
internal_pages = leaf_pages = 0;
if (WT_VERBOSE_ISSET(session, WT_VERB_CHECKPOINT))
WT_RET(__wt_epoch(session, &start));
switch (syncop) {
case WT_SYNC_WRITE_LEAVES:
/*
* Write all immediately available, dirty in-cache leaf pages.
*
* Writing the leaf pages is done without acquiring a high-level
* lock, serialize so multiple threads don't walk the tree at
* the same time.
*/
if (!btree->modified)
return (0);
__wt_spin_lock(session, &btree->flush_lock);
if (!btree->modified) {
__wt_spin_unlock(session, &btree->flush_lock);
return (0);
}
/*
* Save the oldest transaction ID we need to keep around.
* Otherwise, in a busy system, we could be updating pages so
* fast that write leaves never catches up. We deliberately
* have no transaction running at this point that would keep
* the oldest ID from moving forwards as we walk the tree.
*/
oldest_id = __wt_txn_oldest_id(session);
flags |= WT_READ_NO_WAIT | WT_READ_SKIP_INTL;
for (walk = NULL;;) {
WT_ERR(__wt_tree_walk(session, &walk, flags));
if (walk == NULL)
break;
/*
* Write dirty pages if nobody beat us to it. Don't
* try to write hot pages (defined as pages that have
* been updated since the write phase leaves started):
* checkpoint will have to visit them anyway.
*/
page = walk->page;
if (__wt_page_is_modified(page) &&
WT_TXNID_LT(page->modify->update_txn, oldest_id)) {
if (txn->isolation == WT_ISO_READ_COMMITTED)
__wt_txn_get_snapshot(session);
leaf_bytes += page->memory_footprint;
++leaf_pages;
WT_ERR(__wt_reconcile(session, walk, NULL, 0));
}
}
break;
case WT_SYNC_CHECKPOINT:
/*
* If we are flushing a file at read-committed isolation, which
* is of particular interest for flushing the metadata to make
* schema-changing operation durable, get a transactional
* snapshot now.
*
* All changes committed up to this point should be included.
* We don't update the snapshot in between pages because (a)
* the metadata shouldn't be that big, and (b) if we do ever
*/
if (txn->isolation == WT_ISO_READ_COMMITTED)
__wt_txn_get_snapshot(session);
/*
* We cannot check the tree modified flag in the case of a
* checkpoint, the checkpoint code has already cleared it.
*
* Writing the leaf pages is done without acquiring a high-level
* lock, serialize so multiple threads don't walk the tree at
* the same time. We're holding the schema lock, but need the
* lower-level lock as well.
*/
__wt_spin_lock(session, &btree->flush_lock);
/*
* In the final checkpoint pass, child pages cannot be evicted
* from underneath internal pages nor can underlying blocks be
* freed until the checkpoint's block lists are stable. Also,
* we cannot split child pages into parents unless we know the
* final pass will write a consistent view of that namespace.
* Set the checkpointing flag to block such actions and wait for
* any problematic eviction or page splits to complete.
*/
WT_PUBLISH(btree->checkpointing, WT_CKPT_PREPARE);
WT_ERR(__wt_evict_file_exclusive_on(session));
__wt_evict_file_exclusive_off(session);
WT_PUBLISH(btree->checkpointing, WT_CKPT_RUNNING);
/* Write all dirty in-cache pages. */
flags |= WT_READ_NO_EVICT;
for (walk = NULL;;) {
WT_ERR(__wt_tree_walk(session, &walk, flags));
if (walk == NULL)
break;
/* Skip clean pages. */
if (!__wt_page_is_modified(walk->page))
continue;
/*
* Take a local reference to the page modify structure
* now that we know the page is dirty. It needs to be
* done in this order otherwise the page modify
* structure could have been created between taking the
* reference and checking modified.
*/
page = walk->page;
mod = page->modify;
/*
* Write dirty pages, unless we can be sure they only
* became dirty after the checkpoint started.
*
* We can skip dirty pages if:
* (1) they are leaf pages;
* (2) there is a snapshot transaction active (which
* is the case in ordinary application checkpoints
* but not all internal cases); and
* (3) the first dirty update on the page is
* sufficiently recent that the checkpoint
* transaction would skip them.
*
* Mark the tree dirty: the checkpoint marked it clean
* and we can't skip future checkpoints until this page
* is written.
*/
if (!WT_PAGE_IS_INTERNAL(page) &&
F_ISSET(txn, WT_TXN_HAS_SNAPSHOT) &&
WT_TXNID_LT(txn->snap_max, mod->first_dirty_txn)) {
__wt_page_modify_set(session, page);
continue;
}
if (WT_PAGE_IS_INTERNAL(page)) {
internal_bytes += page->memory_footprint;
++internal_pages;
} else {
leaf_bytes += page->memory_footprint;
++leaf_pages;
}
WT_ERR(__wt_reconcile(session, walk, NULL, 0));
}
break;
case WT_SYNC_CLOSE:
case WT_SYNC_DISCARD:
WT_ILLEGAL_VALUE_ERR(session);
}
if (WT_VERBOSE_ISSET(session, WT_VERB_CHECKPOINT)) {
WT_ERR(__wt_epoch(session, &end));
WT_ERR(__wt_verbose(session, WT_VERB_CHECKPOINT,
"__sync_file WT_SYNC_%s wrote:\n\t %" PRIu64
" bytes, %" PRIu64 " pages of leaves\n\t %" PRIu64
" bytes, %" PRIu64 " pages of internal\n\t"
"Took: %" PRIu64 "ms",
syncop == WT_SYNC_WRITE_LEAVES ?
"WRITE_LEAVES" : "CHECKPOINT",
leaf_bytes, leaf_pages, internal_bytes, internal_pages,
WT_TIMEDIFF_MS(end, start)));
}
err: /* On error, clear any left-over tree walk. */
if (walk != NULL)
WT_TRET(__wt_page_release(session, walk, flags));
/*
* If we got a snapshot in order to write pages, and there was no
* snapshot active when we started, release it.
*/
if (txn->isolation == WT_ISO_READ_COMMITTED &&
saved_snap_min == WT_TXN_NONE)
__wt_txn_release_snapshot(session);
if (btree->checkpointing != WT_CKPT_OFF) {
/*
* Update the checkpoint generation for this handle so visible
* updates newer than the checkpoint can be evicted.
*
* This has to be published before eviction is enabled again,
* so that eviction knows that the checkpoint has completed.
*/
WT_PUBLISH(btree->checkpoint_gen,
conn->txn_global.checkpoint_gen);
WT_STAT_FAST_DATA_SET(session,
btree_checkpoint_generation, btree->checkpoint_gen);
/*
* Clear the checkpoint flag and push the change; not required,
* but publishing the change means stalled eviction gets moving
* as soon as possible.
*/
btree->checkpointing = WT_CKPT_OFF;
WT_FULL_BARRIER();
/*
* If this tree was being skipped by the eviction server during
* the checkpoint, clear the wait.
*/
btree->evict_walk_period = 0;
/*
* Wake the eviction server, in case application threads have
* stalled while the eviction server decided it couldn't make
* progress. Without this, application threads will be stalled
* until the eviction server next wakes.
*/
WT_TRET(__wt_evict_server_wake(session));
}
__wt_spin_unlock(session, &btree->flush_lock);
/*
* Leaves are written before a checkpoint (or as part of a file close,
* before checkpointing the file). Start a flush to stable storage,
* but don't wait for it.
*/
if (ret == 0 &&
syncop == WT_SYNC_WRITE_LEAVES && F_ISSET(conn, WT_CONN_CKPT_SYNC))
WT_RET(btree->bm->sync(btree->bm, session, true));
return (ret);
}
/*
* __sync_file --
* Flush pages for a specific file.
*/
static int
__sync_file(WT_SESSION_IMPL *session, int syncop)
{
struct timespec end, start;
WT_BTREE *btree;
WT_DECL_RET;
WT_PAGE *page;
WT_PAGE_MODIFY *mod;
WT_REF *walk;
WT_TXN *txn;
uint64_t internal_bytes, leaf_bytes;
uint64_t internal_pages, leaf_pages;
uint32_t flags;
bool evict_reset;
btree = S2BT(session);
flags = WT_READ_CACHE | WT_READ_NO_GEN;
walk = NULL;
txn = &session->txn;
internal_bytes = leaf_bytes = 0;
internal_pages = leaf_pages = 0;
if (WT_VERBOSE_ISSET(session, WT_VERB_CHECKPOINT))
WT_RET(__wt_epoch(session, &start));
switch (syncop) {
case WT_SYNC_WRITE_LEAVES:
/*
* Write all immediately available, dirty in-cache leaf pages.
*
* Writing the leaf pages is done without acquiring a high-level
* lock, serialize so multiple threads don't walk the tree at
* the same time.
*/
if (!btree->modified)
return (0);
__wt_spin_lock(session, &btree->flush_lock);
if (!btree->modified) {
__wt_spin_unlock(session, &btree->flush_lock);
return (0);
}
flags |= WT_READ_NO_WAIT | WT_READ_SKIP_INTL;
for (walk = NULL;;) {
WT_ERR(__wt_tree_walk(session, &walk, NULL, flags));
if (walk == NULL)
break;
/*
* Write dirty pages if nobody beat us to it. Don't
* try to write the hottest pages: checkpoint will have
* to visit them anyway.
*/
page = walk->page;
if (__wt_page_is_modified(page) &&
__wt_txn_visible_all(
session, page->modify->update_txn)) {
if (txn->isolation == WT_ISO_READ_COMMITTED)
__wt_txn_get_snapshot(session);
leaf_bytes += page->memory_footprint;
++leaf_pages;
WT_ERR(__wt_reconcile(session, walk, NULL, 0));
}
}
break;
case WT_SYNC_CHECKPOINT:
/*
* We cannot check the tree modified flag in the case of a
* checkpoint, the checkpoint code has already cleared it.
*
* Writing the leaf pages is done without acquiring a high-level
* lock, serialize so multiple threads don't walk the tree at
* the same time. We're holding the schema lock, but need the
* lower-level lock as well.
*/
__wt_spin_lock(session, &btree->flush_lock);
/*
* When internal pages are being reconciled by checkpoint their
* child pages cannot disappear from underneath them or be split
* into them, nor can underlying blocks be freed until the block
* lists for the checkpoint are stable. Set the checkpointing
* flag to block eviction of dirty pages until the checkpoint's
* internal page pass is complete, then wait for any existing
* eviction to complete.
*/
btree->checkpointing = 1;
WT_FULL_BARRIER();
WT_ERR(__wt_evict_file_exclusive_on(session, &evict_reset));
if (evict_reset)
__wt_evict_file_exclusive_off(session);
/* Write all dirty in-cache pages. */
flags |= WT_READ_NO_EVICT;
for (walk = NULL;;) {
/*
* If we have a page, and it was ever modified, track
* the highest transaction ID in the tree. We do this
* here because we want the value after reconciling
* dirty pages.
*/
if (walk != NULL && walk->page != NULL &&
(mod = walk->page->modify) != NULL &&
WT_TXNID_LT(btree->rec_max_txn, mod->rec_max_txn))
btree->rec_max_txn = mod->rec_max_txn;
WT_ERR(__wt_tree_walk(session, &walk, NULL, flags));
if (walk == NULL)
break;
page = walk->page;
mod = page->modify;
/* Skip clean pages. */
if (!__wt_page_is_modified(page))
continue;
/*
* Write dirty pages, unless we can be sure they only
* became dirty after the checkpoint started.
*
* We can skip dirty pages if:
* (1) they are leaf pages;
* (2) there is a snapshot transaction active (which
* is the case in ordinary application checkpoints
* but not all internal cases); and
* (3) the first dirty update on the page is
* sufficiently recent that the checkpoint
* transaction would skip them.
*
* Mark the tree dirty: the checkpoint marked it clean
* and we can't skip future checkpoints until this page
* is written.
*/
if (!WT_PAGE_IS_INTERNAL(page) &&
F_ISSET(txn, WT_TXN_HAS_SNAPSHOT) &&
WT_TXNID_LT(txn->snap_max, mod->first_dirty_txn) &&
mod->rec_result != WT_PM_REC_REWRITE) {
__wt_page_modify_set(session, page);
continue;
}
if (WT_PAGE_IS_INTERNAL(page)) {
internal_bytes += page->memory_footprint;
++internal_pages;
} else {
leaf_bytes += page->memory_footprint;
++leaf_pages;
}
WT_ERR(__wt_reconcile(session, walk, NULL, 0));
}
break;
}
if (WT_VERBOSE_ISSET(session, WT_VERB_CHECKPOINT)) {
WT_ERR(__wt_epoch(session, &end));
WT_ERR(__wt_verbose(session, WT_VERB_CHECKPOINT,
"__sync_file WT_SYNC_%s wrote:\n\t %" PRIu64
" bytes, %" PRIu64 " pages of leaves\n\t %" PRIu64
" bytes, %" PRIu64 " pages of internal\n\t"
"Took: %" PRIu64 "ms",
syncop == WT_SYNC_WRITE_LEAVES ?
"WRITE_LEAVES" : "CHECKPOINT",
leaf_bytes, leaf_pages, internal_bytes, internal_pages,
WT_TIMEDIFF(end, start) / WT_MILLION));
}
err: /* On error, clear any left-over tree walk. */
if (walk != NULL)
WT_TRET(__wt_page_release(session, walk, flags));
if (txn->isolation == WT_ISO_READ_COMMITTED && session->ncursors == 0)
__wt_txn_release_snapshot(session);
if (btree->checkpointing) {
/*
* Update the checkpoint generation for this handle so visible
* updates newer than the checkpoint can be evicted.
*
* This has to be published before eviction is enabled again,
* so that eviction knows that the checkpoint has completed.
*/
WT_PUBLISH(btree->checkpoint_gen,
S2C(session)->txn_global.checkpoint_gen);
WT_STAT_FAST_DATA_SET(session,
btree_checkpoint_generation, btree->checkpoint_gen);
/*
* Clear the checkpoint flag and push the change; not required,
* but publishing the change means stalled eviction gets moving
* as soon as possible.
*/
btree->checkpointing = 0;
WT_FULL_BARRIER();
/*
* If this tree was being skipped by the eviction server during
* the checkpoint, clear the wait.
*/
btree->evict_walk_period = 0;
/*
* Wake the eviction server, in case application threads have
* stalled while the eviction server decided it couldn't make
* progress. Without this, application threads will be stalled
* until the eviction server next wakes.
*/
WT_TRET(__wt_evict_server_wake(session));
}
__wt_spin_unlock(session, &btree->flush_lock);
/*
* Leaves are written before a checkpoint (or as part of a file close,
* before checkpointing the file). Start a flush to stable storage,
* but don't wait for it.
*/
if (ret == 0 && syncop == WT_SYNC_WRITE_LEAVES)
WT_RET(btree->bm->sync(btree->bm, session, true));
return (ret);
}
/*
* __open_verbose --
* Optionally output a verbose message on handle open.
*/
static inline int
__open_verbose(
WT_SESSION_IMPL *session, const char *name, int file_type, u_int flags)
{
#ifdef HAVE_VERBOSE
WT_DECL_RET;
WT_DECL_ITEM(tmp);
const char *file_type_tag, *sep;
if (!WT_VERBOSE_ISSET(session, WT_VERB_FILEOPS))
return (0);
/*
* It's useful to track file opens when debugging platforms, take some
* effort to output good tracking information.
*/
switch (file_type) {
case WT_FS_OPEN_FILE_TYPE_CHECKPOINT:
file_type_tag = "checkpoint";
break;
case WT_FS_OPEN_FILE_TYPE_DATA:
file_type_tag = "data";
break;
case WT_FS_OPEN_FILE_TYPE_DIRECTORY:
file_type_tag = "directory";
break;
case WT_FS_OPEN_FILE_TYPE_LOG:
file_type_tag = "log";
break;
case WT_FS_OPEN_FILE_TYPE_REGULAR:
file_type_tag = "regular";
break;
default:
file_type_tag = "unknown open type";
break;
}
WT_RET(__wt_scr_alloc(session, 0, &tmp));
sep = " (";
#define WT_FS_OPEN_VERBOSE_FLAG(f, name) \
if (LF_ISSET(f)) { \
WT_ERR(__wt_buf_catfmt( \
session, tmp, "%s%s", sep, name)); \
sep = ", "; \
}
WT_FS_OPEN_VERBOSE_FLAG(WT_FS_OPEN_CREATE, "create");
WT_FS_OPEN_VERBOSE_FLAG(WT_FS_OPEN_DIRECTIO, "direct-IO");
WT_FS_OPEN_VERBOSE_FLAG(WT_FS_OPEN_EXCLUSIVE, "exclusive");
WT_FS_OPEN_VERBOSE_FLAG(WT_FS_OPEN_FIXED, "fixed");
WT_FS_OPEN_VERBOSE_FLAG(WT_FS_OPEN_READONLY, "readonly");
if (tmp->size != 0)
WT_ERR(__wt_buf_catfmt(session, tmp, ")"));
__wt_verbose(session, WT_VERB_FILEOPS,
"%s: file-open: type %s%s",
name, file_type_tag, tmp->size == 0 ? "" : (char *)tmp->data);
err: __wt_scr_free(session, &tmp);
return (ret);
#else
WT_UNUSED(session);
WT_UNUSED(name);
WT_UNUSED(file_type);
WT_UNUSED(flags);
return (0);
#endif
}
/*
* __wt_txn_update_oldest --
* Sweep the running transactions to update the oldest ID required.
* !!!
* If a data-source is calling the WT_EXTENSION_API.transaction_oldest
* method (for the oldest transaction ID not yet visible to a running
* transaction), and then comparing that oldest ID against committed
* transactions to see if updates for a committed transaction are still
* visible to running transactions, the oldest transaction ID may be
* the same as the last committed transaction ID, if the transaction
* state wasn't refreshed after the last transaction committed. Push
* past the last committed transaction.
*/
void
__wt_txn_update_oldest(WT_SESSION_IMPL *session, int force)
{
WT_CONNECTION_IMPL *conn;
WT_SESSION_IMPL *oldest_session;
WT_TXN_GLOBAL *txn_global;
WT_TXN_STATE *s;
uint64_t current_id, id, last_running, oldest_id, prev_oldest_id;
uint32_t i, session_cnt;
int32_t count;
int last_running_moved;
conn = S2C(session);
txn_global = &conn->txn_global;
current_id = last_running = txn_global->current;
oldest_session = NULL;
prev_oldest_id = txn_global->oldest_id;
/*
* For pure read-only workloads, or if the update isn't forced and the
* oldest ID isn't too far behind, avoid scanning.
*/
if (prev_oldest_id == current_id ||
(!force && WT_TXNID_LT(current_id, prev_oldest_id + 100)))
return;
/*
* We're going to scan. Increment the count of scanners to prevent the
* oldest ID from moving forwards. Spin if the count is negative,
* which indicates that some thread is moving the oldest ID forwards.
*/
do {
if ((count = txn_global->scan_count) < 0)
WT_PAUSE();
} while (count < 0 ||
!WT_ATOMIC_CAS4(txn_global->scan_count, count, count + 1));
/* The oldest ID cannot change until the scan count goes to zero. */
prev_oldest_id = txn_global->oldest_id;
current_id = oldest_id = last_running = txn_global->current;
/* Walk the array of concurrent transactions. */
WT_ORDERED_READ(session_cnt, conn->session_cnt);
for (i = 0, s = txn_global->states; i < session_cnt; i++, s++) {
/*
* Update the oldest ID.
*
* Ignore: IDs older than the oldest ID we saw. This can happen
* if we race with a thread that is allocating an ID -- the ID
* will not be used because the thread will keep spinning until
* it gets a valid one.
*/
if ((id = s->id) != WT_TXN_NONE &&
WT_TXNID_LE(prev_oldest_id, id) &&
WT_TXNID_LT(id, last_running))
last_running = id;
/*
* !!!
* Note: Don't ignore snap_min values older than the previous
* oldest ID. Read-uncommitted operations publish snap_min
* values without incrementing scan_count to protect the global
* table. See the comment in __wt_txn_cursor_op for
* more details.
*/
if ((id = s->snap_min) != WT_TXN_NONE &&
WT_TXNID_LT(id, oldest_id)) {
oldest_id = id;
oldest_session = &conn->sessions[i];
}
}
if (WT_TXNID_LT(last_running, oldest_id))
oldest_id = last_running;
/* The oldest ID can't move past any named snapshots. */
if ((id = txn_global->nsnap_oldest_id) != WT_TXN_NONE &&
WT_TXNID_LT(id, oldest_id))
oldest_id = id;
/* Update the last running ID. */
last_running_moved =
WT_TXNID_LT(txn_global->last_running, last_running);
/* Update the oldest ID. */
if ((WT_TXNID_LT(prev_oldest_id, oldest_id) || last_running_moved) &&
WT_ATOMIC_CAS4(txn_global->scan_count, 1, -1)) {
WT_ORDERED_READ(session_cnt, conn->session_cnt);
for (i = 0, s = txn_global->states; i < session_cnt; i++, s++) {
if ((id = s->id) != WT_TXN_NONE &&
WT_TXNID_LT(id, last_running))
last_running = id;
if ((id = s->snap_min) != WT_TXN_NONE &&
WT_TXNID_LT(id, oldest_id))
oldest_id = id;
}
if (WT_TXNID_LT(last_running, oldest_id))
oldest_id = last_running;
#ifdef HAVE_DIAGNOSTIC
/*
* Make sure the ID doesn't move past any named snapshots.
*
* Don't include the read/assignment in the assert statement.
* Coverity complains if there are assignments only done in
* diagnostic builds, and when the read is from a volatile.
*/
id = txn_global->nsnap_oldest_id;
WT_ASSERT(session,
id == WT_TXN_NONE || !WT_TXNID_LT(id, oldest_id));
#endif
if (WT_TXNID_LT(txn_global->last_running, last_running))
txn_global->last_running = last_running;
if (WT_TXNID_LT(txn_global->oldest_id, oldest_id))
txn_global->oldest_id = oldest_id;
WT_ASSERT(session, txn_global->scan_count == -1);
txn_global->scan_count = 0;
} else {
if (WT_VERBOSE_ISSET(session, WT_VERB_TRANSACTION) &&
current_id - oldest_id > 10000 && last_running_moved &&
oldest_session != NULL) {
(void)__wt_verbose(session, WT_VERB_TRANSACTION,
"old snapshot %" PRIu64
" pinned in session %d [%s]"
" with snap_min %" PRIu64 "\n",
oldest_id, oldest_session->id,
oldest_session->lastop,
oldest_session->txn.snap_min);
}
WT_ASSERT(session, txn_global->scan_count > 0);
(void)WT_ATOMIC_SUB4(txn_global->scan_count, 1);
}
}
/*
* __wt_rec_track --
* Add an object to the page's list of tracked objects.
*/
int
__wt_rec_track(WT_SESSION_IMPL *session, WT_PAGE *page,
const uint8_t *addr, uint32_t addr_size,
const void *data, uint32_t data_size, uint32_t flags)
{
WT_PAGE_MODIFY *mod;
WT_PAGE_TRACK *empty, *track;
uint8_t *p;
uint32_t i;
mod = page->modify;
/* Find an empty slot. */
empty = NULL;
for (track = mod->track, i = 0; i < mod->track_entries; ++track, ++i)
if (!F_ISSET(track, WT_TRK_OBJECT)) {
empty = track;
break;
}
/* Reallocate space as necessary. */
if (empty == NULL) {
WT_RET(__rec_track_extend(session, page));
empty = &mod->track[mod->track_entries - 1];
}
track = empty;
/*
* Minor optimization: allocate a single chunk of space instead of two
* separate ones: be careful when it's freed.
*/
WT_RET(__wt_calloc_def(session, addr_size + data_size, &p));
/*
* Set the just-added flag so we clean up should reconciliation fail,
* except for cached overflow values, which don't get discarded, even
* if reconciliation fails.
*/
track->flags = (uint8_t)flags | WT_TRK_OBJECT;
if (!F_ISSET(track, WT_TRK_OVFL_VALUE))
F_SET(track, WT_TRK_JUST_ADDED);
track->addr.addr = p;
track->addr.size = addr_size;
memcpy(track->addr.addr, addr, addr_size);
if (data_size) {
p += addr_size;
track->data = p;
track->size = data_size;
memcpy(track->data, data, data_size);
}
/*
* Overflow items are potentially large and on-page items remain in the
* tracking list until the page is evicted. If we're tracking a lot of
* them, their memory might matter: increment the page and cache memory
* totals. This is unlikely to matter, but it's inexpensive (unless
* there are lots of them, in with case I guess the memory matters).
*
* If this reconciliation were to fail, we would reasonably perform the
* inverse operation in __wt_rec_track_wrapup_err. I'm not bothering
* with that because we'd have to crack the structure itself to figure
* out how much to decrement and I don't think it's worth the effort.
* The potential problem is repeatedly failing reconciliation of a page
* with a large number of overflow items, which causes the page's memory
* memory footprint to become incorrectly high, causing us to push the
* page out of cache unnecessarily. Like I said, not worth the effort.
*/
if (LF_ISSET(WT_TRK_ONPAGE))
__wt_cache_page_inmem_incr(
session, page, addr_size + data_size);
if (WT_VERBOSE_ISSET(session, reconcile))
WT_RET(__track_msg(session, page, "add", track));
return (0);
}
/*
* __wt_lsm_merge --
* Merge a set of chunks of an LSM tree.
*/
int
__wt_lsm_merge(WT_SESSION_IMPL *session, WT_LSM_TREE *lsm_tree, u_int id)
{
WT_BLOOM *bloom;
WT_CURSOR *dest, *src;
WT_DECL_RET;
WT_ITEM key, value;
WT_LSM_CHUNK *chunk;
uint32_t generation;
uint64_t insert_count, record_count;
u_int dest_id, end_chunk, i, nchunks, start_chunk, start_id, verb;
int tret;
bool created_chunk, create_bloom, locked, in_sync;
const char *cfg[3];
const char *drop_cfg[] =
{ WT_CONFIG_BASE(session, WT_SESSION_drop), "force", NULL };
bloom = NULL;
chunk = NULL;
dest = src = NULL;
start_id = 0;
created_chunk = create_bloom = locked = in_sync = false;
/* Fast path if it's obvious no merges could be done. */
if (lsm_tree->nchunks < lsm_tree->merge_min &&
lsm_tree->merge_aggressiveness < WT_LSM_AGGRESSIVE_THRESHOLD)
return (WT_NOTFOUND);
/*
* Use the lsm_tree lock to read the chunks (so no switches occur), but
* avoid holding it while the merge is in progress: that may take a
* long time.
*/
WT_RET(__wt_lsm_tree_writelock(session, lsm_tree));
locked = true;
WT_ERR(__lsm_merge_span(session,
lsm_tree, id, &start_chunk, &end_chunk, &record_count));
nchunks = (end_chunk + 1) - start_chunk;
WT_ASSERT(session, nchunks > 0);
start_id = lsm_tree->chunk[start_chunk]->id;
/* Find the merge generation. */
for (generation = 0, i = 0; i < nchunks; i++)
generation = WT_MAX(generation,
lsm_tree->chunk[start_chunk + i]->generation + 1);
WT_ERR(__wt_lsm_tree_writeunlock(session, lsm_tree));
locked = false;
/* Allocate an ID for the merge. */
dest_id = __wt_atomic_add32(&lsm_tree->last, 1);
/*
* We only want to do the chunk loop if we're running with verbose,
* so we wrap these statements in the conditional. Avoid the loop
* in the normal path.
*/
if (WT_VERBOSE_ISSET(session, WT_VERB_LSM)) {
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"Merging %s chunks %u-%u into %u (%" PRIu64 " records)"
", generation %" PRIu32,
lsm_tree->name,
start_chunk, end_chunk, dest_id, record_count, generation));
for (verb = start_chunk; verb <= end_chunk; verb++)
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"Merging %s: Chunk[%u] id %u, gen: %" PRIu32
", size: %" PRIu64 ", records: %" PRIu64,
lsm_tree->name, verb, lsm_tree->chunk[verb]->id,
lsm_tree->chunk[verb]->generation,
lsm_tree->chunk[verb]->size,
lsm_tree->chunk[verb]->count));
}
WT_ERR(__wt_calloc_one(session, &chunk));
created_chunk = true;
chunk->id = dest_id;
if (FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_MERGED) &&
(FLD_ISSET(lsm_tree->bloom, WT_LSM_BLOOM_OLDEST) ||
start_chunk > 0) && record_count > 0)
create_bloom = true;
/*
* Special setup for the merge cursor:
* first, reset to open the dependent cursors;
* then restrict the cursor to a specific number of chunks;
* then set MERGE so the cursor doesn't track updates to the tree.
*/
WT_ERR(__wt_open_cursor(session, lsm_tree->name, NULL, NULL, &src));
F_SET(src, WT_CURSTD_RAW);
WT_ERR(__wt_clsm_init_merge(src, start_chunk, start_id, nchunks));
WT_WITH_SCHEMA_LOCK(session,
ret = __wt_lsm_tree_setup_chunk(session, lsm_tree, chunk));
WT_ERR(ret);
if (create_bloom) {
WT_ERR(__wt_lsm_tree_setup_bloom(session, lsm_tree, chunk));
WT_ERR(__wt_bloom_create(session, chunk->bloom_uri,
lsm_tree->bloom_config,
record_count, lsm_tree->bloom_bit_count,
lsm_tree->bloom_hash_count, &bloom));
}
/* Discard pages we read as soon as we're done with them. */
F_SET(session, WT_SESSION_NO_CACHE);
cfg[0] = WT_CONFIG_BASE(session, WT_SESSION_open_cursor);
cfg[1] = "bulk,raw,skip_sort_check";
cfg[2] = NULL;
WT_ERR(__wt_open_cursor(session, chunk->uri, NULL, cfg, &dest));
#define LSM_MERGE_CHECK_INTERVAL WT_THOUSAND
for (insert_count = 0; (ret = src->next(src)) == 0; insert_count++) {
if (insert_count % LSM_MERGE_CHECK_INTERVAL == 0) {
if (!F_ISSET(lsm_tree, WT_LSM_TREE_ACTIVE))
WT_ERR(EINTR);
WT_STAT_FAST_CONN_INCRV(session,
lsm_rows_merged, LSM_MERGE_CHECK_INTERVAL);
++lsm_tree->merge_progressing;
}
WT_ERR(src->get_key(src, &key));
dest->set_key(dest, &key);
WT_ERR(src->get_value(src, &value));
dest->set_value(dest, &value);
WT_ERR(dest->insert(dest));
if (create_bloom)
WT_ERR(__wt_bloom_insert(bloom, &key));
}
WT_ERR_NOTFOUND_OK(ret);
WT_STAT_FAST_CONN_INCRV(session,
lsm_rows_merged, insert_count % LSM_MERGE_CHECK_INTERVAL);
++lsm_tree->merge_progressing;
WT_ERR(__wt_verbose(session, WT_VERB_LSM,
"Bloom size for %" PRIu64 " has %" PRIu64 " items inserted.",
record_count, insert_count));
/*
* Closing and syncing the files can take a while. Set the
* merge_syncing field so that compact knows it is still in
* progress.
*/
(void)__wt_atomic_add32(&lsm_tree->merge_syncing, 1);
in_sync = true;
/*
* We've successfully created the new chunk. Now install it. We need
* to ensure that the NO_CACHE flag is cleared and the bloom filter
* is closed (even if a step fails), so track errors but don't return
* until we've cleaned up.
*/
WT_TRET(src->close(src));
WT_TRET(dest->close(dest));
src = dest = NULL;
F_CLR(session, WT_SESSION_NO_CACHE);
/*
* We're doing advisory reads to fault the new trees into cache.
* Don't block if the cache is full: our next unit of work may be to
* discard some trees to free space.
*/
F_SET(session, WT_SESSION_NO_EVICTION);
if (create_bloom) {
if (ret == 0)
WT_TRET(__wt_bloom_finalize(bloom));
/*
* Read in a key to make sure the Bloom filters btree handle is
* open before it becomes visible to application threads.
* Otherwise application threads will stall while it is opened
* and internal pages are read into cache.
*/
if (ret == 0) {
WT_CLEAR(key);
WT_TRET_NOTFOUND_OK(__wt_bloom_get(bloom, &key));
}
WT_TRET(__wt_bloom_close(bloom));
bloom = NULL;
}
WT_ERR(ret);
/*
* Open a handle on the new chunk before application threads attempt
* to access it, opening it pre-loads internal pages into the file
* system cache.
*/
cfg[1] = "checkpoint=" WT_CHECKPOINT;
WT_ERR(__wt_open_cursor(session, chunk->uri, NULL, cfg, &dest));
WT_TRET(dest->close(dest));
dest = NULL;
++lsm_tree->merge_progressing;
(void)__wt_atomic_sub32(&lsm_tree->merge_syncing, 1);
in_sync = false;
WT_ERR_NOTFOUND_OK(ret);
WT_ERR(__wt_lsm_tree_set_chunk_size(session, chunk));
WT_ERR(__wt_lsm_tree_writelock(session, lsm_tree));
locked = true;
/*
* Check whether we raced with another merge, and adjust the chunk
* array offset as necessary.
*/
if (start_chunk >= lsm_tree->nchunks ||
lsm_tree->chunk[start_chunk]->id != start_id)
for (start_chunk = 0;
start_chunk < lsm_tree->nchunks;
start_chunk++)
if (lsm_tree->chunk[start_chunk]->id == start_id)
break;
/*
* It is safe to error out here - since the update can only fail
* prior to making updates to the tree.
*/
WT_ERR(__wt_lsm_merge_update_tree(
session, lsm_tree, start_chunk, nchunks, chunk));
if (create_bloom)
F_SET(chunk, WT_LSM_CHUNK_BLOOM);
chunk->count = insert_count;
chunk->generation = generation;
F_SET(chunk, WT_LSM_CHUNK_ONDISK);
/*
* We have no current way of continuing if the metadata update fails,
* so we will panic in that case. Put some effort into cleaning up
* after ourselves here - so things have a chance of shutting down.
*
* Any errors that happened after the tree was locked are
* fatal - we can't guarantee the state of the tree.
*/
if ((ret = __wt_lsm_meta_write(session, lsm_tree)) != 0)
WT_PANIC_ERR(session, ret, "Failed finalizing LSM merge");
lsm_tree->dsk_gen++;
/* Update the throttling while holding the tree lock. */
__wt_lsm_tree_throttle(session, lsm_tree, true);
/* Schedule a pass to discard old chunks */
WT_ERR(__wt_lsm_manager_push_entry(
session, WT_LSM_WORK_DROP, 0, lsm_tree));
err: if (locked)
WT_TRET(__wt_lsm_tree_writeunlock(session, lsm_tree));
if (in_sync)
(void)__wt_atomic_sub32(&lsm_tree->merge_syncing, 1);
if (src != NULL)
WT_TRET(src->close(src));
if (dest != NULL)
WT_TRET(dest->close(dest));
if (bloom != NULL)
WT_TRET(__wt_bloom_close(bloom));
if (ret != 0 && created_chunk) {
/* Drop the newly-created files on error. */
if (chunk->uri != NULL) {
WT_WITH_SCHEMA_LOCK(session, tret =
__wt_schema_drop(session, chunk->uri, drop_cfg));
WT_TRET(tret);
}
if (create_bloom && chunk->bloom_uri != NULL) {
WT_WITH_SCHEMA_LOCK(session,
tret = __wt_schema_drop(
session, chunk->bloom_uri, drop_cfg));
WT_TRET(tret);
}
__wt_free(session, chunk->bloom_uri);
__wt_free(session, chunk->uri);
__wt_free(session, chunk);
if (ret == EINTR)
WT_TRET(__wt_verbose(session, WT_VERB_LSM,
"Merge aborted due to close"));
else
WT_TRET(__wt_verbose(session, WT_VERB_LSM,
"Merge failed with %s",
__wt_strerror(session, ret, NULL, 0)));
}
F_CLR(session, WT_SESSION_NO_CACHE | WT_SESSION_NO_EVICTION);
return (ret);
}
/*
* __wt_block_checkpoint_load --
* Load a checkpoint.
*/
int
__wt_block_checkpoint_load(WT_SESSION_IMPL *session, WT_BLOCK *block,
const uint8_t *addr, size_t addr_size,
uint8_t *root_addr, size_t *root_addr_sizep, bool checkpoint)
{
WT_BLOCK_CKPT *ci, _ci;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
uint8_t *endp;
ci = NULL;
/*
* Sometimes we don't find a root page (we weren't given a checkpoint,
* or the checkpoint was empty). In that case we return an empty root
* address, set that up now.
*/
*root_addr_sizep = 0;
#ifdef HAVE_VERBOSE
if (WT_VERBOSE_ISSET(session, WT_VERB_CHECKPOINT)) {
if (addr != NULL) {
WT_ERR(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(__ckpt_string(session, block, addr, tmp));
}
__wt_verbose(session, WT_VERB_CHECKPOINT,
"%s: load-checkpoint: %s", block->name,
addr == NULL ? "[Empty]" : (const char *)tmp->data);
}
#endif
/*
* There's a single checkpoint in the file that can be written, all of
* the others are read-only. We use the same initialization calls for
* readonly checkpoints, but the information doesn't persist.
*/
if (checkpoint) {
ci = &_ci;
WT_ERR(__wt_block_ckpt_init(session, ci, "checkpoint"));
} else {
/*
* We depend on the btree level for locking: things will go bad
* fast if we open the live system in two handles, or salvage,
* truncate or verify the live/running file.
*/
#ifdef HAVE_DIAGNOSTIC
__wt_spin_lock(session, &block->live_lock);
WT_ASSERT(session, block->live_open == false);
block->live_open = true;
__wt_spin_unlock(session, &block->live_lock);
#endif
ci = &block->live;
WT_ERR(__wt_block_ckpt_init(session, ci, "live"));
}
/*
* If the checkpoint has an on-disk root page, load it. Otherwise, size
* the file past the description information.
*/
if (addr == NULL || addr_size == 0)
ci->file_size = block->allocsize;
else {
/* Crack the checkpoint cookie. */
WT_ERR(__wt_block_buffer_to_ckpt(session, block, addr, ci));
/* Verify sets up next. */
if (block->verify)
WT_ERR(__wt_verify_ckpt_load(session, block, ci));
/* Read any root page. */
if (ci->root_offset != WT_BLOCK_INVALID_OFFSET) {
endp = root_addr;
WT_ERR(__wt_block_addr_to_buffer(block, &endp,
ci->root_offset, ci->root_size, ci->root_checksum));
*root_addr_sizep = WT_PTRDIFF(endp, root_addr);
}
/*
* Rolling a checkpoint forward requires the avail list, the
* blocks from which we can allocate.
*/
if (!checkpoint)
WT_ERR(__wt_block_extlist_read_avail(
session, block, &ci->avail, ci->file_size));
}
/*
* If the checkpoint can be written, that means anything written after
* the checkpoint is no longer interesting, truncate the file. Don't
* bother checking the avail list for a block at the end of the file,
* that was done when the checkpoint was first written (re-writing the
* checkpoint might possibly make it relevant here, but it's unlikely
* enough I don't bother).
*/
if (!checkpoint)
WT_ERR(__wt_block_truncate(session, block, ci->file_size));
if (0) {
err: /*
* Don't call checkpoint-unload: unload does real work including
* file truncation. If we fail early enough that the checkpoint
* information isn't correct, bad things would happen. The only
* allocated memory was in the service of verify, clean that up.
*/
if (block->verify)
WT_TRET(__wt_verify_ckpt_unload(session, block));
}
/* Checkpoints don't need the original information, discard it. */
if (checkpoint && ci != NULL)
__wt_block_ckpt_destroy(session, ci);
__wt_scr_free(session, &tmp);
return (ret);
}
/*
* __ckpt_process --
* Process the list of checkpoints.
*/
static int
__ckpt_process(
WT_SESSION_IMPL *session, WT_BLOCK *block, WT_CKPT *ckptbase)
{
WT_BLOCK_CKPT *a, *b, *ci;
WT_CKPT *ckpt, *next_ckpt;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
uint64_t ckpt_size;
int deleting, locked;
ci = &block->live;
locked = 0;
/*
* We've allocated our last page, update the checkpoint size. We need
* to calculate the live system's checkpoint size before reading and
* merging checkpoint allocation and discard information from the
* checkpoints we're deleting, those operations change the underlying
* byte counts.
*/
ckpt_size = ci->ckpt_size;
ckpt_size += ci->alloc.bytes;
ckpt_size -= ci->discard.bytes;
/*
* Extents newly available as a result of deleting previous checkpoints
* are added to a list of extents. The list should be empty, but there
* is no explicit "free the checkpoint information" call into the block
* manager; if there was an error in an upper level resulting in some
* previous checkpoint never being resolved, the list may not be empty.
*
* XXX
* This isn't sufficient, actually: we're going to leak all the blocks
* written as part of the last checkpoint because it was never resolved.
*/
__wt_block_extlist_free(session, &ci->ckpt_avail);
WT_RET(__wt_block_extlist_init(
session, &ci->ckpt_avail, "live", "ckpt_avail"));
/*
* To delete a checkpoint, we'll need checkpoint information for it and
* the subsequent checkpoint into which it gets rolled; read them from
* disk before we lock things down.
*/
deleting = 0;
WT_CKPT_FOREACH(ckptbase, ckpt) {
if (F_ISSET(ckpt, WT_CKPT_FAKE) ||
!F_ISSET(ckpt, WT_CKPT_DELETE))
continue;
deleting = 1;
/*
* Read the checkpoint and next checkpoint extent lists if we
* haven't already read them (we may have already read these
* extent blocks if there is more than one deleted checkpoint).
*/
if (ckpt->bpriv == NULL)
WT_ERR(__ckpt_extlist_read(session, block, ckpt));
for (next_ckpt = ckpt + 1;; ++next_ckpt)
if (!F_ISSET(next_ckpt, WT_CKPT_FAKE))
break;
/*
* The "next" checkpoint may be the live tree which has no
* extent blocks to read.
*/
if (next_ckpt->bpriv == NULL &&
!F_ISSET(next_ckpt, WT_CKPT_ADD))
WT_ERR(__ckpt_extlist_read(session, block, next_ckpt));
}
/*
* Hold a lock so the live extent lists and the file size can't change
* underneath us. I suspect we'll tighten this if checkpoints take too
* much time away from real work: we read the historic checkpoint
* information without a lock, but we could also merge and re-write the
* delete checkpoint information without a lock, except for ranges
* merged into the live tree.
*/
__wt_spin_lock(session, &block->live_lock);
locked = 1;
/* Skip the additional processing if we aren't deleting checkpoints. */
if (!deleting)
goto live_update;
/*
* Delete any no-longer-needed checkpoints: we do this first as it frees
* blocks to the live lists, and the freed blocks will then be included
* when writing the live extent lists.
*/
WT_CKPT_FOREACH(ckptbase, ckpt) {
if (F_ISSET(ckpt, WT_CKPT_FAKE) ||
!F_ISSET(ckpt, WT_CKPT_DELETE))
continue;
if (WT_VERBOSE_ISSET(session, ckpt)) {
if (tmp == NULL)
WT_ERR(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(__ckpt_string(
session, block, ckpt->raw.data, tmp));
WT_VERBOSE_ERR(session, ckpt,
"%s: delete-checkpoint: %s: %s",
block->name, ckpt->name, (char *)tmp->data);
}
/*
* Find the checkpoint into which we'll roll this checkpoint's
* blocks: it's the next real checkpoint in the list, and it
* better have been read in (if it's not the add slot).
*/
for (next_ckpt = ckpt + 1;; ++next_ckpt)
if (!F_ISSET(next_ckpt, WT_CKPT_FAKE))
break;
/*
* Set the from/to checkpoint structures, where the "to" value
* may be the live tree.
*/
a = ckpt->bpriv;
if (F_ISSET(next_ckpt, WT_CKPT_ADD))
b = &block->live;
else
b = next_ckpt->bpriv;
/*
* Free the root page: there's nothing special about this free,
* the root page is allocated using normal rules, that is, it
* may have been taken from the avail list, and was entered on
* the live system's alloc list at that time. We free it into
* the checkpoint's discard list, however, not the live system's
* list because it appears on the checkpoint's alloc list and so
* must be paired in the checkpoint.
*/
if (a->root_offset != WT_BLOCK_INVALID_OFFSET)
WT_ERR(__wt_block_insert_ext(session,
&a->discard, a->root_offset, a->root_size));
/*
* Free the blocks used to hold the "from" checkpoint's extent
* lists, including the avail list.
*/
WT_ERR(__ckpt_extlist_fblocks(session, block, &a->alloc));
WT_ERR(__ckpt_extlist_fblocks(session, block, &a->avail));
WT_ERR(__ckpt_extlist_fblocks(session, block, &a->discard));
/*
* Roll the "from" alloc and discard extent lists into the "to"
* checkpoint's lists.
*/
if (a->alloc.entries != 0)
WT_ERR(__wt_block_extlist_merge(
session, &a->alloc, &b->alloc));
if (a->discard.entries != 0)
WT_ERR(__wt_block_extlist_merge(
session, &a->discard, &b->discard));
/*
* If the "to" checkpoint is also being deleted, we're done with
* it, it's merged into some other checkpoint in the next loop.
* This means the extent lists may aggregate over a number of
* checkpoints, but that's OK, they're disjoint sets of ranges.
*/
if (F_ISSET(next_ckpt, WT_CKPT_DELETE))
continue;
/*
* Find blocks for re-use: wherever the "to" checkpoint's
* allocate and discard lists overlap, move the range to
* the live system's checkpoint available list.
*/
WT_ERR(__wt_block_extlist_overlap(session, block, b));
/*
* If we're updating the live system's information, we're done.
*/
if (F_ISSET(next_ckpt, WT_CKPT_ADD))
continue;
/*
* We have to write the "to" checkpoint's extent lists out in
* new blocks, and update its cookie.
*
* Free the blocks used to hold the "to" checkpoint's extent
* lists; don't include the avail list, it's not changing.
*/
WT_ERR(__ckpt_extlist_fblocks(session, block, &b->alloc));
WT_ERR(__ckpt_extlist_fblocks(session, block, &b->discard));
F_SET(next_ckpt, WT_CKPT_UPDATE);
}
/* Update checkpoints marked for update. */
WT_CKPT_FOREACH(ckptbase, ckpt)
if (F_ISSET(ckpt, WT_CKPT_UPDATE)) {
WT_ASSERT(session, !F_ISSET(ckpt, WT_CKPT_ADD));
WT_ERR(__ckpt_update(
session, block, ckpt, ckpt->bpriv, 0, 0));
}
live_update:
ci = &block->live;
/* Truncate the file if that's possible. */
WT_ERR(__wt_block_extlist_truncate(session, block, &ci->avail));
/* Update the final, added checkpoint based on the live system. */
WT_CKPT_FOREACH(ckptbase, ckpt)
if (F_ISSET(ckpt, WT_CKPT_ADD)) {
WT_ERR(__ckpt_update(
session, block, ckpt, ci, ckpt_size, 1));
/*
* XXX
* Our caller wants the final checkpoint size. Setting
* the size here violates layering, but the alternative
* is a call for the btree layer to crack the checkpoint
* cookie into its components, and that's a fair amount
* of work.
*/
ckpt->ckpt_size = ci->ckpt_size;
}
/*
* Reset the live system's alloc and discard extent lists, leave the
* avail list alone.
*/
__wt_block_extlist_free(session, &ci->alloc);
WT_ERR(__wt_block_extlist_init(session, &ci->alloc, "live", "alloc"));
__wt_block_extlist_free(session, &ci->discard);
WT_ERR(
__wt_block_extlist_init(session, &ci->discard, "live", "discard"));
#ifdef HAVE_DIAGNOSTIC
/*
* The first checkpoint in the system should always have an empty
* discard list. If we've read that checkpoint and/or created it,
* check.
*/
WT_CKPT_FOREACH(ckptbase, ckpt)
if (!F_ISSET(ckpt, WT_CKPT_DELETE))
break;
if ((a = ckpt->bpriv) == NULL)
a = &block->live;
if (a->discard.entries != 0) {
__wt_errx(session,
"first checkpoint incorrectly has blocks on the discard "
"list");
WT_ERR(WT_ERROR);
}
#endif
err: if (locked)
__wt_spin_unlock(session, &block->live_lock);
/* Discard any checkpoint information we loaded. */
WT_CKPT_FOREACH(ckptbase, ckpt)
if ((ci = ckpt->bpriv) != NULL)
__wt_block_ckpt_destroy(session, ci);
__wt_scr_free(&tmp);
return (ret);
}
/*
* __sync_file --
* Flush pages for a specific file.
*/
static int
__sync_file(WT_SESSION_IMPL *session, WT_CACHE_OP syncop)
{
WT_BTREE *btree;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_PAGE *page;
WT_PAGE_MODIFY *mod;
WT_REF *prev, *walk;
WT_TXN *txn;
uint64_t internal_bytes, internal_pages, leaf_bytes, leaf_pages;
uint64_t oldest_id, saved_pinned_id, time_start, time_stop;
uint32_t flags;
bool timer, tried_eviction;
conn = S2C(session);
btree = S2BT(session);
prev = walk = NULL;
txn = &session->txn;
tried_eviction = false;
time_start = time_stop = 0;
/* Only visit pages in cache and don't bump page read generations. */
flags = WT_READ_CACHE | WT_READ_NO_GEN;
/*
* Skip all deleted pages. For a page to be marked deleted, it must
* have been evicted from cache and marked clean. Checkpoint should
* never instantiate deleted pages: if a truncate is not visible to the
* checkpoint, the on-disk version is correct. If the truncate is
* visible, we skip over the child page when writing its parent. We
* check whether a truncate is visible in the checkpoint as part of
* reconciling internal pages (specifically in __rec_child_modify).
*/
LF_SET(WT_READ_DELETED_SKIP);
internal_bytes = leaf_bytes = 0;
internal_pages = leaf_pages = 0;
saved_pinned_id = WT_SESSION_TXN_STATE(session)->pinned_id;
timer = WT_VERBOSE_ISSET(session, WT_VERB_CHECKPOINT);
if (timer)
time_start = __wt_clock(session);
switch (syncop) {
case WT_SYNC_WRITE_LEAVES:
/*
* Write all immediately available, dirty in-cache leaf pages.
*
* Writing the leaf pages is done without acquiring a high-level
* lock, serialize so multiple threads don't walk the tree at
* the same time.
*/
if (!btree->modified)
return (0);
__wt_spin_lock(session, &btree->flush_lock);
if (!btree->modified) {
__wt_spin_unlock(session, &btree->flush_lock);
return (0);
}
/*
* Save the oldest transaction ID we need to keep around.
* Otherwise, in a busy system, we could be updating pages so
* fast that write leaves never catches up. We deliberately
* have no transaction running at this point that would keep
* the oldest ID from moving forwards as we walk the tree.
*/
oldest_id = __wt_txn_oldest_id(session);
LF_SET(WT_READ_NO_WAIT | WT_READ_SKIP_INTL);
for (;;) {
WT_ERR(__wt_tree_walk(session, &walk, flags));
if (walk == NULL)
break;
/*
* Write dirty pages if nobody beat us to it. Don't
* try to write hot pages (defined as pages that have
* been updated since the write phase leaves started):
* checkpoint will have to visit them anyway.
*/
page = walk->page;
if (__wt_page_is_modified(page) &&
WT_TXNID_LT(page->modify->update_txn, oldest_id)) {
if (txn->isolation == WT_ISO_READ_COMMITTED)
__wt_txn_get_snapshot(session);
leaf_bytes += page->memory_footprint;
++leaf_pages;
WT_ERR(__wt_reconcile(session,
walk, NULL, WT_REC_CHECKPOINT, NULL));
}
}
break;
case WT_SYNC_CHECKPOINT:
/*
* If we are flushing a file at read-committed isolation, which
* is of particular interest for flushing the metadata to make
* a schema-changing operation durable, get a transactional
* snapshot now.
*
* All changes committed up to this point should be included.
* We don't update the snapshot in between pages because the
* metadata shouldn't have many pages. Instead, read-committed
* isolation ensures that all metadata updates completed before
* the checkpoint are included.
*/
if (txn->isolation == WT_ISO_READ_COMMITTED)
__wt_txn_get_snapshot(session);
/*
* We cannot check the tree modified flag in the case of a
* checkpoint, the checkpoint code has already cleared it.
*
* Writing the leaf pages is done without acquiring a high-level
* lock, serialize so multiple threads don't walk the tree at
* the same time. We're holding the schema lock, but need the
* lower-level lock as well.
*/
__wt_spin_lock(session, &btree->flush_lock);
/*
* In the final checkpoint pass, child pages cannot be evicted
* from underneath internal pages nor can underlying blocks be
* freed until the checkpoint's block lists are stable. Also,
* we cannot split child pages into parents unless we know the
* final pass will write a consistent view of that namespace.
* Set the checkpointing flag to block such actions and wait for
* any problematic eviction or page splits to complete.
*/
WT_ASSERT(session, btree->syncing == WT_BTREE_SYNC_OFF &&
btree->sync_session == NULL);
btree->sync_session = session;
btree->syncing = WT_BTREE_SYNC_WAIT;
(void)__wt_gen_next_drain(session, WT_GEN_EVICT);
btree->syncing = WT_BTREE_SYNC_RUNNING;
/* Write all dirty in-cache pages. */
LF_SET(WT_READ_NO_EVICT);
/* Read pages with lookaside entries and evict them asap. */
LF_SET(WT_READ_LOOKASIDE | WT_READ_WONT_NEED);
for (;;) {
WT_ERR(__sync_dup_walk(session, walk, flags, &prev));
WT_ERR(__wt_tree_walk(session, &walk, flags));
if (walk == NULL)
break;
/*
* Skip clean pages, but need to make sure maximum
* transaction ID is always updated.
*/
if (!__wt_page_is_modified(walk->page)) {
if (((mod = walk->page->modify) != NULL) &&
mod->rec_max_txn > btree->rec_max_txn)
btree->rec_max_txn = mod->rec_max_txn;
if (mod != NULL &&
btree->rec_max_timestamp <
mod->rec_max_timestamp)
btree->rec_max_timestamp =
mod->rec_max_timestamp;
continue;
}
/*
* Take a local reference to the page modify structure
* now that we know the page is dirty. It needs to be
* done in this order otherwise the page modify
* structure could have been created between taking the
* reference and checking modified.
*/
page = walk->page;
/*
* Write dirty pages, if we can't skip them. If we skip
* a page, mark the tree dirty. The checkpoint marked it
* clean and we can't skip future checkpoints until this
* page is written.
*/
if (__sync_checkpoint_can_skip(session, page)) {
__wt_tree_modify_set(session);
continue;
}
if (WT_PAGE_IS_INTERNAL(page)) {
internal_bytes += page->memory_footprint;
++internal_pages;
} else {
leaf_bytes += page->memory_footprint;
++leaf_pages;
}
/*
* If the page was pulled into cache by our read, try
* to evict it now.
*
* For eviction to have a chance, we first need to move
* the walk point to the next page checkpoint will
* visit. We want to avoid this code being too special
* purpose, so try to reuse the ordinary eviction path.
*
* Regardless of whether eviction succeeds or fails,
* the walk continues from the previous location. We
* remember whether we tried eviction, and don't try
* again. Even if eviction fails (the page may stay in
* cache clean but with history that cannot be
* discarded), that is not wasted effort because
* checkpoint doesn't need to write the page again.
*/
if (!WT_PAGE_IS_INTERNAL(page) &&
page->read_gen == WT_READGEN_WONT_NEED &&
!tried_eviction) {
WT_ERR_BUSY_OK(
__wt_page_release_evict(session, walk));
walk = prev;
prev = NULL;
tried_eviction = true;
continue;
}
tried_eviction = false;
WT_ERR(__wt_reconcile(
session, walk, NULL, WT_REC_CHECKPOINT, NULL));
/*
* Update checkpoint IO tracking data if configured
* to log verbose progress messages.
*/
if (conn->ckpt_timer_start.tv_sec > 0) {
conn->ckpt_write_bytes +=
page->memory_footprint;
++conn->ckpt_write_pages;
/* Periodically log checkpoint progress. */
if (conn->ckpt_write_pages % 5000 == 0)
__wt_checkpoint_progress(
session, false);
}
}
break;
case WT_SYNC_CLOSE:
case WT_SYNC_DISCARD:
WT_ERR(__wt_illegal_value(session, syncop));
break;
}
if (timer) {
time_stop = __wt_clock(session);
__wt_verbose(session, WT_VERB_CHECKPOINT,
"__sync_file WT_SYNC_%s wrote: %" PRIu64
" leaf pages (%" PRIu64 "B), %" PRIu64
" internal pages (%" PRIu64 "B), and took %" PRIu64 "ms",
syncop == WT_SYNC_WRITE_LEAVES ?
"WRITE_LEAVES" : "CHECKPOINT",
leaf_pages, leaf_bytes, internal_pages, internal_bytes,
WT_CLOCKDIFF_MS(time_stop, time_start));
}
err: /* On error, clear any left-over tree walk. */
WT_TRET(__wt_page_release(session, walk, flags));
WT_TRET(__wt_page_release(session, prev, flags));
/*
* If we got a snapshot in order to write pages, and there was no
* snapshot active when we started, release it.
*/
if (txn->isolation == WT_ISO_READ_COMMITTED &&
saved_pinned_id == WT_TXN_NONE)
__wt_txn_release_snapshot(session);
/* Clear the checkpoint flag. */
btree->syncing = WT_BTREE_SYNC_OFF;
btree->sync_session = NULL;
__wt_spin_unlock(session, &btree->flush_lock);
/*
* Leaves are written before a checkpoint (or as part of a file close,
* before checkpointing the file). Start a flush to stable storage,
* but don't wait for it.
*/
if (ret == 0 &&
syncop == WT_SYNC_WRITE_LEAVES && F_ISSET(conn, WT_CONN_CKPT_SYNC))
WT_RET(btree->bm->sync(btree->bm, session, false));
return (ret);
}
/*
* __snapshot_process --
* Process the list of snapshots.
*/
static int
__snapshot_process(
WT_SESSION_IMPL *session, WT_BLOCK *block, WT_SNAPSHOT *snapbase)
{
WT_BLOCK_SNAPSHOT *a, *b, *si;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_SNAPSHOT *snap;
uint64_t snapshot_size;
int deleting, locked;
si = &block->live;
locked = 0;
/*
* We've allocated our last page, update the snapshot size. We need to
* calculate the live system's snapshot size before reading and merging
* snapshot allocation and discard information from the snapshots we're
* deleting, those operations will change the underlying byte counts.
*/
snapshot_size = si->snapshot_size;
snapshot_size += si->alloc.bytes;
snapshot_size -= si->discard.bytes;
/*
* Extents that become newly available as a result of deleting previous
* snapshots are added to a list of extents. The list should be empty,
* but there's no explicit "free the snapshot information" call into the
* block manager; if there was an error in an upper level resulting in
* the snapshot never being "resolved", the list might not be empty.
*
* XXX
* This isn't sufficient, actually: we're going to leak all the blocks
* that were written as part of the last snapshot because it was never
* resolved.
*/
__wt_block_extlist_free(session, &si->snapshot_avail);
WT_RET(__wt_block_extlist_init(
session, &si->snapshot_avail, "live", "snapshot_avail"));
/*
* To delete a snapshot, we'll need snapshot information for it, and we
* have to read that from the disk.
*/
deleting = 0;
WT_SNAPSHOT_FOREACH(snapbase, snap) {
/*
* To delete a snapshot, we'll need snapshot information for it
* and the subsequent snapshot. The test is tricky, we have to
* load the current snapshot's information if it's marked for
* deletion, or if it follows a snapshot marked for deletion,
* where the boundary cases are the first snapshot in the list
* and the last snapshot in the list: if we're deleting the last
* snapshot in the list, there's no next snapshot, the snapshot
* will be merged into the live tree.
*/
if (!F_ISSET(snap, WT_SNAP_DELETE) &&
(snap == snapbase ||
F_ISSET(snap, WT_SNAP_ADD) ||
!F_ISSET(snap - 1, WT_SNAP_DELETE)))
continue;
deleting = 1;
/*
* Allocate a snapshot structure, crack the cookie and read the
* snapshot's extent lists.
*
* Ignore the avail list: snapshot avail lists are only useful
* if we are rolling forward from the particular snapshot and
* they represent our best understanding of what blocks can be
* allocated. If we are not operating on the live snapshot,
* subsequent snapshots might have allocated those blocks, and
* the avail list is useless. We don't discard it, because it
* is useful as part of verification, but we don't re-write it
* either.
*/
WT_ERR(__wt_calloc(
session, 1, sizeof(WT_BLOCK_SNAPSHOT), &snap->bpriv));
si = snap->bpriv;
WT_ERR(__wt_block_snap_init(session, block, si, snap->name, 0));
WT_ERR(__wt_block_buffer_to_snapshot(
session, block, snap->raw.data, si));
WT_ERR(__wt_block_extlist_read(session, block, &si->alloc));
WT_ERR(__wt_block_extlist_read(session, block, &si->discard));
}
/*
* Hold a lock so the live extent lists and the file size can't change
* underneath us. I suspect we'll tighten this if snapshots take too
* much time away from real work: we read historic snapshot information
* without a lock, but we could also merge and re-write the delete
* snapshot information without a lock, except for ranges merged into
* the live tree.
*/
__wt_spin_lock(session, &block->live_lock);
locked = 1;
/* Skip the additional processing if we aren't deleting snapshots. */
if (!deleting)
goto live_update;
/*
* Delete any no-longer-needed snapshots: we do this first as it frees
* blocks to the live lists, and the freed blocks will then be included
* when writing the live extent lists.
*/
WT_SNAPSHOT_FOREACH(snapbase, snap) {
if (!F_ISSET(snap, WT_SNAP_DELETE))
continue;
if (WT_VERBOSE_ISSET(session, snapshot)) {
if (tmp == NULL)
WT_ERR(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(__snapshot_string(
session, block, snap->raw.data, tmp));
WT_VERBOSE_ERR(session, snapshot,
"%s: delete-snapshot: %s: %s",
block->name, snap->name, (char *)tmp->data);
}
/*
* Set the from/to snapshot structures, where the "to" value
* may be the live tree.
*/
a = snap->bpriv;
if (F_ISSET(snap + 1, WT_SNAP_ADD))
b = &block->live;
else
b = (snap + 1)->bpriv;
/*
* Free the root page: there's nothing special about this free,
* the root page is allocated using normal rules, that is, it
* may have been taken from the avail list, and was entered on
* the live system's alloc list at that time. We free it into
* the snapshot's discard list, however, not the live system's
* list because it appears on the snapshot's alloc list and so
* must be paired in the snapshot.
*/
if (a->root_offset != WT_BLOCK_INVALID_OFFSET)
WT_ERR(__wt_block_insert_ext(session,
&a->discard, a->root_offset, a->root_size));
/*
* Free the blocks used to hold the "from" snapshot's extent
* lists directly to the live system's avail list, they were
* never on any alloc list. Include the "from" snapshot's
* avail list, it's going away.
*/
WT_ERR(__snapshot_extlist_fblocks(session, block, &a->alloc));
WT_ERR(__snapshot_extlist_fblocks(session, block, &a->avail));
WT_ERR(__snapshot_extlist_fblocks(session, block, &a->discard));
/*
* Roll the "from" alloc and discard extent lists into the "to"
* snapshot's lists.
*/
if (a->alloc.entries != 0)
WT_ERR(__wt_block_extlist_merge(
session, &a->alloc, &b->alloc));
if (a->discard.entries != 0)
WT_ERR(__wt_block_extlist_merge(
session, &a->discard, &b->discard));
/*
* If the "to" snapshot is also being deleted, we're done with
* it, it's merged into some other snapshot in the next loop.
* This means the extent lists may aggregate over a number of
* snapshots, but that's OK, they're disjoint sets of ranges.
*/
if (F_ISSET(snap + 1, WT_SNAP_DELETE))
continue;
/*
* Find blocks for re-use: wherever the "to" snapshot's allocate
* and discard lists overlap is fair game, move ranges appearing
* on both lists to the live snapshot's newly available list.
*/
WT_ERR(__wt_block_extlist_overlap(session, block, b));
/*
* If we're updating the live system's information, we're done.
*/
if (F_ISSET(snap + 1, WT_SNAP_ADD))
continue;
/*
* We have to write the "to" snapshot's extent lists out in new
* blocks, and update its cookie.
*
* Free the blocks used to hold the "to" snapshot's extent lists
* directly to the live system's avail list, they were never on
* any alloc list. Do not include the "to" snapshot's avail
* list, it's not changing.
*/
WT_ERR(__snapshot_extlist_fblocks(session, block, &b->alloc));
WT_ERR(__snapshot_extlist_fblocks(session, block, &b->discard));
F_SET(snap + 1, WT_SNAP_UPDATE);
}
/* Update snapshots marked for update. */
WT_SNAPSHOT_FOREACH(snapbase, snap)
if (F_ISSET(snap, WT_SNAP_UPDATE)) {
WT_ASSERT(session, !F_ISSET(snap, WT_SNAP_ADD));
WT_ERR(__snapshot_update(
session, block, snap, snap->bpriv, 0, 0));
}
live_update:
si = &block->live;
/* Truncate the file if that's possible. */
WT_ERR(__wt_block_extlist_truncate(session, block, &si->avail));
/* Update the final, added snapshot based on the live system. */
WT_SNAPSHOT_FOREACH(snapbase, snap)
if (F_ISSET(snap, WT_SNAP_ADD)) {
WT_ERR(__snapshot_update(
session, block, snap, si, snapshot_size, 1));
/*
* XXX
* Our caller wants two pieces of information: the time
* the snapshot was taken and the final snapshot size.
* This violates layering but the alternative is a call
* for the btree layer to crack the snapshot cookie into
* its components, and that's a fair amount of work.
* (We could just read the system time in the session
* layer when updating the metadata file, but that won't
* work for the snapshot size, and so we do both here.)
*/
snap->snapshot_size = si->snapshot_size;
WT_ERR(__wt_epoch(session, &snap->sec, NULL));
}
/*
* Reset the live system's alloc and discard extent lists, leave the
* avail list alone.
*/
__wt_block_extlist_free(session, &si->alloc);
WT_ERR(__wt_block_extlist_init(session, &si->alloc, "live", "alloc"));
__wt_block_extlist_free(session, &si->discard);
WT_ERR(
__wt_block_extlist_init(session, &si->discard, "live", "discard"));
#ifdef HAVE_DIAGNOSTIC
/*
* The first snapshot in the system should always have an empty discard
* list. If we've read that snapshot and/or created it, check.
*/
WT_SNAPSHOT_FOREACH(snapbase, snap)
if (!F_ISSET(snap, WT_SNAP_DELETE))
break;
if ((a = snap->bpriv) == NULL)
a = &block->live;
if (a->discard.entries != 0) {
__wt_errx(session,
"snapshot incorrectly has blocks on the discard list");
WT_ERR(WT_ERROR);
}
#endif
err: if (locked)
__wt_spin_unlock(session, &block->live_lock);
/* Discard any snapshot information we loaded, we no longer need it. */
WT_SNAPSHOT_FOREACH(snapbase, snap)
if ((si = snap->bpriv) != NULL) {
__wt_block_extlist_free(session, &si->alloc);
__wt_block_extlist_free(session, &si->avail);
__wt_block_extlist_free(session, &si->discard);
}
__wt_scr_free(&tmp);
return (ret);
}
/*
* __snapshot_update --
* Update a snapshot.
*/
static int
__snapshot_update(
WT_SESSION_IMPL *session, WT_BLOCK *block, WT_SNAPSHOT *snap,
WT_BLOCK_SNAPSHOT *si, uint64_t snapshot_size, int is_live)
{
WT_DECL_ITEM(tmp);
WT_DECL_RET;
uint8_t *endp;
#ifdef HAVE_DIAGNOSTIC
/* Check the extent list combinations for overlaps. */
WT_RET(__wt_block_extlist_check(session, &si->alloc, &si->avail));
WT_RET(__wt_block_extlist_check(session, &si->discard, &si->avail));
WT_RET(__wt_block_extlist_check(session, &si->alloc, &si->discard));
#endif
/*
* Write the snapshot's extent lists; we only write an avail list for
* the live system, other snapshot's avail lists are static and never
* change. When we do write the avail list for the live system it's
* two lists: the current avail list plus the list of blocks that are
* being made available as of the new snapshot. We can't merge that
* second list into the real list yet, it's not truly available until
* the new snapshot location has been saved to the metadata.
*/
WT_RET(__wt_block_extlist_write(session, block, &si->alloc, NULL));
if (is_live)
WT_RET(__wt_block_extlist_write(
session, block, &si->avail, &si->snapshot_avail));
WT_RET(__wt_block_extlist_write(session, block, &si->discard, NULL));
/*
* Set the file size for the live system.
*
* XXX
* We do NOT set the file size when re-writing snapshots because we want
* to test the snapshot's blocks against a reasonable maximum file size
* during verification. This is not good: imagine a snapshot appearing
* early in the file, re-written, and then the snapshot requires blocks
* at the end of the file, blocks after the listed file size. If the
* application opens that snapshot for writing (discarding subsequent
* snapshots), we would truncate the file to the early chunk, discarding
* the re-written snapshot information. The alternative, updating the
* file size has its own problems, in that case we'd work correctly, but
* we'd lose all of the blocks between the original snapshot and the
* re-written snapshot. Currently, there's no API to roll-forward
* intermediate snapshots, if there ever is, this will need to be fixed.
*/
if (is_live)
WT_RET(__wt_filesize(session, block->fh, &si->file_size));
/* Set the snapshot size for the live system. */
if (is_live)
si->snapshot_size = snapshot_size;
/*
* Copy the snapshot information into the snapshot array's address
* cookie.
*/
WT_RET(__wt_buf_init(session, &snap->raw, WT_BTREE_MAX_ADDR_COOKIE));
endp = snap->raw.mem;
WT_RET(__wt_block_snapshot_to_buffer(session, block, &endp, si));
snap->raw.size = WT_PTRDIFF32(endp, snap->raw.mem);
if (WT_VERBOSE_ISSET(session, snapshot)) {
WT_RET(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(__snapshot_string(session, block, snap->raw.data, tmp));
WT_VERBOSE_ERR(session, snapshot,
"%s: create-snapshot: %s: %s",
block->name, snap->name, (char *)tmp->data);
}
err: __wt_scr_free(&tmp);
return (ret);
}
/*
* __wt_block_snapshot_load --
* Load a snapshot.
*/
int
__wt_block_snapshot_load(WT_SESSION_IMPL *session,
WT_BLOCK *block, WT_ITEM *dsk, const uint8_t *addr, uint32_t addr_size,
int readonly)
{
WT_BLOCK_SNAPSHOT *si;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_UNUSED(addr_size);
/*
* Sometimes we don't find a root page (we weren't given a snapshot,
* or the referenced snapshot was empty). In that case we return a
* root page size of 0. Set that up now.
*/
dsk->size = 0;
si = &block->live;
WT_RET(__wt_block_snap_init(session, block, si, "live", 1));
if (WT_VERBOSE_ISSET(session, snapshot)) {
if (addr != NULL) {
WT_ERR(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(__snapshot_string(session, block, addr, tmp));
}
WT_VERBOSE_ERR(session, snapshot,
"%s: load-snapshot: %s", block->name,
addr == NULL ? "[Empty]" : (char *)tmp->data);
}
/* If not loading a snapshot from disk, we're done. */
if (addr == NULL || addr_size == 0)
return (0);
/* Crack the snapshot cookie. */
if (addr != NULL)
WT_ERR(__wt_block_buffer_to_snapshot(session, block, addr, si));
/* Verify sets up next. */
if (block->verify)
WT_ERR(__wt_verify_snap_load(session, block, si));
/* Read, and optionally verify, any root page. */
if (si->root_offset != WT_BLOCK_INVALID_OFFSET) {
WT_ERR(__wt_block_read_off(session, block,
dsk, si->root_offset, si->root_size, si->root_cksum));
if (block->verify) {
if (tmp == NULL) {
WT_ERR(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(__snapshot_string(
session, block, addr, tmp));
}
WT_ERR(
__wt_verify_dsk(session, (char *)tmp->data, dsk));
}
}
/*
* Rolling a snapshot forward requires the avail list, the blocks from
* which we can allocate.
*/
if (!readonly)
WT_ERR(__wt_block_extlist_read(session, block, &si->avail));
/*
* If the snapshot can be written, that means anything written after
* the snapshot is no longer interesting. Truncate the file.
*/
if (!readonly) {
WT_VERBOSE_ERR(session, snapshot,
"truncate file to %" PRIuMAX, (uintmax_t)si->file_size);
WT_ERR(__wt_ftruncate(session, block->fh, si->file_size));
}
if (0) {
err: (void)__wt_block_snapshot_unload(session, block);
}
__wt_scr_free(&tmp);
return (ret);
}
/*
* __ovfl_reuse_wrapup --
* Resolve the page's overflow reuse list after a page is written.
*/
static int
__ovfl_reuse_wrapup(WT_SESSION_IMPL *session, WT_PAGE *page)
{
WT_BM *bm;
WT_OVFL_REUSE **e, **head, *reuse;
size_t decr;
int i;
bm = S2BT(session)->bm;
head = page->modify->ovfl_track->ovfl_reuse;
/*
* Discard any overflow records that aren't in-use, freeing underlying
* blocks.
*
* First, walk the overflow reuse lists (except for the lowest one),
* fixing up skiplist links.
*/
for (i = WT_SKIP_MAXDEPTH - 1; i > 0; --i)
for (e = &head[i]; (reuse = *e) != NULL;) {
if (F_ISSET(reuse, WT_OVFL_REUSE_INUSE)) {
e = &reuse->next[i];
continue;
}
*e = reuse->next[i];
}
/*
* Second, discard any overflow record without an in-use flag, clear
* the flags for the next run.
*
* As part of the pass through the lowest level, figure out how much
* space we added/subtracted from the page, and update its footprint.
* We don't get it exactly correct because we don't know the depth of
* the skiplist here, but it's close enough, and figuring out the
* memory footprint change in the reconciliation wrapup code means
* fewer atomic updates and less code overall.
*/
decr = 0;
for (e = &head[0]; (reuse = *e) != NULL;) {
if (F_ISSET(reuse, WT_OVFL_REUSE_INUSE)) {
F_CLR(reuse,
WT_OVFL_REUSE_INUSE | WT_OVFL_REUSE_JUST_ADDED);
e = &reuse->next[0];
continue;
}
*e = reuse->next[0];
WT_ASSERT(session, !F_ISSET(reuse, WT_OVFL_REUSE_JUST_ADDED));
if (WT_VERBOSE_ISSET(session, WT_VERB_OVERFLOW))
WT_RET(
__ovfl_reuse_verbose(session, page, reuse, "free"));
WT_RET(bm->free(
bm, session, WT_OVFL_REUSE_ADDR(reuse), reuse->addr_size));
decr += WT_OVFL_SIZE(reuse, WT_OVFL_REUSE);
__wt_free(session, reuse);
}
if (decr != 0)
__wt_cache_page_inmem_decr(session, page, decr);
return (0);
}
/*
* __ckpt_process --
* Process the list of checkpoints.
*/
static int
__ckpt_process(WT_SESSION_IMPL *session, WT_BLOCK *block, WT_CKPT *ckptbase)
{
WT_BLOCK_CKPT *a, *b, *ci;
WT_CKPT *ckpt, *next_ckpt;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
uint64_t ckpt_size;
bool deleting, fatal, locked;
ci = &block->live;
fatal = locked = false;
#ifdef HAVE_DIAGNOSTIC
WT_RET(__ckpt_verify(session, ckptbase));
#endif
/*
* Checkpoints are a two-step process: first, write a new checkpoint to
* disk (including all the new extent lists for modified checkpoints
* and the live system). As part of this, create a list of file blocks
* newly available for reallocation, based on checkpoints being deleted.
* We then return the locations of the new checkpoint information to our
* caller. Our caller has to write that information into some kind of
* stable storage, and once that's done, we can actually allocate from
* that list of newly available file blocks. (We can't allocate from
* that list immediately because the allocation might happen before our
* caller saves the new checkpoint information, and if we crashed before
* the new checkpoint location was saved, we'd have overwritten blocks
* still referenced by checkpoints in the system.) In summary, there is
* a second step: after our caller saves the checkpoint information, we
* are called to add the newly available blocks into the live system's
* available list.
*
* This function is the first step, the second step is in the resolve
* function.
*
* If we're called to checkpoint the same file twice (without the second
* resolution step), or re-entered for any reason, it's an error in our
* caller, and our choices are all bad: leak blocks or potentially crash
* with our caller not yet having saved previous checkpoint information
* to stable storage.
*/
__wt_spin_lock(session, &block->live_lock);
if (block->ckpt_inprogress)
ret = __wt_block_panic(session, EINVAL,
"%s: unexpected checkpoint ordering", block->name);
else
block->ckpt_inprogress = true;
__wt_spin_unlock(session, &block->live_lock);
WT_RET(ret);
/*
* Extents newly available as a result of deleting previous checkpoints
* are added to a list of extents. The list should be empty, but as
* described above, there is no "free the checkpoint information" call
* into the block manager; if there was an error in an upper level that
* resulted in some previous checkpoint never being resolved, the list
* may not be empty. We should have caught that with the "checkpoint
* in progress" test, but it doesn't cost us anything to be cautious.
*
* We free the checkpoint's allocation and discard extent lists as part
* of the resolution step, not because they're needed at that time, but
* because it's potentially a lot of work, and waiting allows the btree
* layer to continue eviction sooner. As for the checkpoint-available
* list, make sure they get cleaned out.
*/
__wt_block_extlist_free(session, &ci->ckpt_avail);
WT_RET(__wt_block_extlist_init(
session, &ci->ckpt_avail, "live", "ckpt_avail", true));
__wt_block_extlist_free(session, &ci->ckpt_alloc);
__wt_block_extlist_free(session, &ci->ckpt_discard);
/*
* To delete a checkpoint, we'll need checkpoint information for it and
* the subsequent checkpoint into which it gets rolled; read them from
* disk before we lock things down.
*/
deleting = false;
WT_CKPT_FOREACH(ckptbase, ckpt) {
if (F_ISSET(ckpt, WT_CKPT_FAKE) ||
!F_ISSET(ckpt, WT_CKPT_DELETE))
continue;
deleting = true;
/*
* Read the checkpoint and next checkpoint extent lists if we
* haven't already read them (we may have already read these
* extent blocks if there is more than one deleted checkpoint).
*/
if (ckpt->bpriv == NULL)
WT_ERR(__ckpt_extlist_read(session, block, ckpt));
for (next_ckpt = ckpt + 1;; ++next_ckpt)
if (!F_ISSET(next_ckpt, WT_CKPT_FAKE))
break;
/*
* The "next" checkpoint may be the live tree which has no
* extent blocks to read.
*/
if (next_ckpt->bpriv == NULL &&
!F_ISSET(next_ckpt, WT_CKPT_ADD))
WT_ERR(__ckpt_extlist_read(session, block, next_ckpt));
}
/*
* Failures are now fatal: we can't currently back out the merge of any
* deleted checkpoint extent lists into the live system's extent lists,
* so continuing after error would leave the live system's extent lists
* corrupted for any subsequent checkpoint (and potentially, should a
* subsequent checkpoint succeed, for recovery).
*/
fatal = true;
/*
* Hold a lock so the live extent lists and the file size can't change
* underneath us. I suspect we'll tighten this if checkpoints take too
* much time away from real work: we read the historic checkpoint
* information without a lock, but we could also merge and re-write the
* deleted and merged checkpoint information without a lock, except for
* the final merge of ranges into the live tree.
*/
__wt_spin_lock(session, &block->live_lock);
locked = true;
/*
* We've allocated our last page, update the checkpoint size. We need
* to calculate the live system's checkpoint size before merging
* checkpoint allocation and discard information from the checkpoints
* we're deleting, those operations change the underlying byte counts.
*/
ckpt_size = ci->ckpt_size;
ckpt_size += ci->alloc.bytes;
ckpt_size -= ci->discard.bytes;
/* Skip the additional processing if we aren't deleting checkpoints. */
if (!deleting)
goto live_update;
/*
* Delete any no-longer-needed checkpoints: we do this first as it frees
* blocks to the live lists, and the freed blocks will then be included
* when writing the live extent lists.
*/
WT_CKPT_FOREACH(ckptbase, ckpt) {
if (F_ISSET(ckpt, WT_CKPT_FAKE) ||
!F_ISSET(ckpt, WT_CKPT_DELETE))
continue;
#ifdef HAVE_VERBOSE
if (WT_VERBOSE_ISSET(session, WT_VERB_CHECKPOINT)) {
if (tmp == NULL)
WT_ERR(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(__ckpt_string(
session, block, ckpt->raw.data, tmp));
__wt_verbose(session, WT_VERB_CHECKPOINT,
"%s: delete-checkpoint: %s: %s",
block->name, ckpt->name, (const char *)tmp->data);
}
#endif
/*
* Find the checkpoint into which we'll roll this checkpoint's
* blocks: it's the next real checkpoint in the list, and it
* better have been read in (if it's not the add slot).
*/
for (next_ckpt = ckpt + 1;; ++next_ckpt)
if (!F_ISSET(next_ckpt, WT_CKPT_FAKE))
break;
/*
* Set the from/to checkpoint structures, where the "to" value
* may be the live tree.
*/
a = ckpt->bpriv;
if (F_ISSET(next_ckpt, WT_CKPT_ADD))
b = &block->live;
else
b = next_ckpt->bpriv;
/*
* Free the root page: there's nothing special about this free,
* the root page is allocated using normal rules, that is, it
* may have been taken from the avail list, and was entered on
* the live system's alloc list at that time. We free it into
* the checkpoint's discard list, however, not the live system's
* list because it appears on the checkpoint's alloc list and so
* must be paired in the checkpoint.
*/
if (a->root_offset != WT_BLOCK_INVALID_OFFSET)
WT_ERR(__wt_block_insert_ext(session, block,
&a->discard, a->root_offset, a->root_size));
/*
* Free the blocks used to hold the "from" checkpoint's extent
* lists, including the avail list.
*/
WT_ERR(__ckpt_extlist_fblocks(session, block, &a->alloc));
WT_ERR(__ckpt_extlist_fblocks(session, block, &a->avail));
WT_ERR(__ckpt_extlist_fblocks(session, block, &a->discard));
/*
* Roll the "from" alloc and discard extent lists into the "to"
* checkpoint's lists.
*/
if (a->alloc.entries != 0)
WT_ERR(__wt_block_extlist_merge(
session, block, &a->alloc, &b->alloc));
if (a->discard.entries != 0)
WT_ERR(__wt_block_extlist_merge(
session, block, &a->discard, &b->discard));
/*
* If the "to" checkpoint is also being deleted, we're done with
* it, it's merged into some other checkpoint in the next loop.
* This means the extent lists may aggregate over a number of
* checkpoints, but that's OK, they're disjoint sets of ranges.
*/
if (F_ISSET(next_ckpt, WT_CKPT_DELETE))
continue;
/*
* Find blocks for re-use: wherever the "to" checkpoint's
* allocate and discard lists overlap, move the range to
* the live system's checkpoint available list.
*/
WT_ERR(__wt_block_extlist_overlap(session, block, b));
/*
* If we're updating the live system's information, we're done.
*/
if (F_ISSET(next_ckpt, WT_CKPT_ADD))
continue;
/*
* We have to write the "to" checkpoint's extent lists out in
* new blocks, and update its cookie.
*
* Free the blocks used to hold the "to" checkpoint's extent
* lists; don't include the avail list, it's not changing.
*/
WT_ERR(__ckpt_extlist_fblocks(session, block, &b->alloc));
WT_ERR(__ckpt_extlist_fblocks(session, block, &b->discard));
F_SET(next_ckpt, WT_CKPT_UPDATE);
}
/* Update checkpoints marked for update. */
WT_CKPT_FOREACH(ckptbase, ckpt)
if (F_ISSET(ckpt, WT_CKPT_UPDATE))
WT_ERR(__ckpt_update(
session, block, ckpt, ckpt->bpriv, false));
live_update:
/* Truncate the file if that's possible. */
WT_ERR(__wt_block_extlist_truncate(session, block, &ci->avail));
/* Update the final, added checkpoint based on the live system. */
WT_CKPT_FOREACH(ckptbase, ckpt)
if (F_ISSET(ckpt, WT_CKPT_ADD)) {
/*
* !!!
* Our caller wants the final checkpoint size. Setting
* the size here violates layering, but the alternative
* is a call for the btree layer to crack the checkpoint
* cookie into its components, and that's a fair amount
* of work.
*/
ckpt->ckpt_size = ckpt_size;
/*
* Set the rolling checkpoint size for the live system.
* The current size includes the current checkpoint's
* root page size (root pages are on the checkpoint's
* block allocation list as root pages are allocated
* with the usual block allocation functions). That's
* correct, but we don't want to include it in the size
* for the next checkpoint.
*/
ckpt_size -= ci->root_size;
/*
* Additionally, we had a bug for awhile where the live
* checkpoint size grew without bound. We can't sanity
* check the value, that would require walking the tree
* as part of the checkpoint. Bound any bug at the size
* of the file.
* It isn't practical to assert that the value is within
* bounds since databases created with older versions
* of WiredTiger (2.8.0) would likely see an error.
*/
ci->ckpt_size =
WT_MIN(ckpt_size, (uint64_t)block->size);
WT_ERR(__ckpt_update(session, block, ckpt, ci, true));
}
/*
* Reset the live system's alloc and discard extent lists, leave the
* avail list alone. This includes freeing a lot of extents, so do it
* outside of the system's lock by copying and resetting the original,
* then doing the work later.
*/
ci->ckpt_alloc = ci->alloc;
WT_ERR(__wt_block_extlist_init(
session, &ci->alloc, "live", "alloc", false));
ci->ckpt_discard = ci->discard;
WT_ERR(__wt_block_extlist_init(
session, &ci->discard, "live", "discard", false));
#ifdef HAVE_DIAGNOSTIC
/*
* The first checkpoint in the system should always have an empty
* discard list. If we've read that checkpoint and/or created it,
* check.
*/
WT_CKPT_FOREACH(ckptbase, ckpt)
if (!F_ISSET(ckpt, WT_CKPT_DELETE))
break;
if ((a = ckpt->bpriv) == NULL)
a = &block->live;
if (a->discard.entries != 0)
WT_ERR_MSG(session, WT_ERROR,
"first checkpoint incorrectly has blocks on the discard "
"list");
#endif
err: if (ret != 0 && fatal)
ret = __wt_block_panic(session, ret,
"%s: fatal checkpoint failure", block->name);
if (locked)
__wt_spin_unlock(session, &block->live_lock);
/* Discard any checkpoint information we loaded. */
WT_CKPT_FOREACH(ckptbase, ckpt)
if ((ci = ckpt->bpriv) != NULL)
__wt_block_ckpt_destroy(session, ci);
__wt_scr_free(session, &tmp);
return (ret);
}
/*
* __wt_txn_update_oldest --
* Sweep the running transactions to update the oldest ID required.
*/
int
__wt_txn_update_oldest(WT_SESSION_IMPL *session, uint32_t flags)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_SESSION_IMPL *oldest_session;
WT_TXN_GLOBAL *txn_global;
uint64_t current_id, last_running, oldest_id;
uint64_t prev_last_running, prev_oldest_id;
bool strict, wait;
conn = S2C(session);
txn_global = &conn->txn_global;
strict = LF_ISSET(WT_TXN_OLDEST_STRICT);
wait = LF_ISSET(WT_TXN_OLDEST_WAIT);
current_id = last_running = txn_global->current;
prev_last_running = txn_global->last_running;
prev_oldest_id = txn_global->oldest_id;
/*
* For pure read-only workloads, or if the update isn't forced and the
* oldest ID isn't too far behind, avoid scanning.
*/
if (prev_oldest_id == current_id ||
(!strict && WT_TXNID_LT(current_id, prev_oldest_id + 100)))
return (0);
/* First do a read-only scan. */
if (wait)
__wt_readlock(session, txn_global->scan_rwlock);
else if ((ret =
__wt_try_readlock(session, txn_global->scan_rwlock)) != 0)
return (ret == EBUSY ? 0 : ret);
__txn_oldest_scan(session, &oldest_id, &last_running, &oldest_session);
__wt_readunlock(session, txn_global->scan_rwlock);
/*
* If the state hasn't changed (or hasn't moved far enough for
* non-forced updates), give up.
*/
if ((oldest_id == prev_oldest_id ||
(!strict && WT_TXNID_LT(oldest_id, prev_oldest_id + 100))) &&
((last_running == prev_last_running) ||
(!strict && WT_TXNID_LT(last_running, prev_last_running + 100))))
return (0);
/* It looks like an update is necessary, wait for exclusive access. */
if (wait)
__wt_writelock(session, txn_global->scan_rwlock);
else if ((ret =
__wt_try_writelock(session, txn_global->scan_rwlock)) != 0)
return (ret == EBUSY ? 0 : ret);
/*
* If the oldest ID has been updated while we waited, don't bother
* scanning.
*/
if (WT_TXNID_LE(oldest_id, txn_global->oldest_id) &&
WT_TXNID_LE(last_running, txn_global->last_running))
goto done;
/*
* Re-scan now that we have exclusive access. This is necessary because
* threads get transaction snapshots with read locks, and we have to be
* sure that there isn't a thread that has got a snapshot locally but
* not yet published its snap_min.
*/
__txn_oldest_scan(session, &oldest_id, &last_running, &oldest_session);
#ifdef HAVE_DIAGNOSTIC
{
/*
* Make sure the ID doesn't move past any named snapshots.
*
* Don't include the read/assignment in the assert statement. Coverity
* complains if there are assignments only done in diagnostic builds,
* and when the read is from a volatile.
*/
uint64_t id = txn_global->nsnap_oldest_id;
WT_ASSERT(session,
id == WT_TXN_NONE || !WT_TXNID_LT(id, oldest_id));
}
#endif
/* Update the oldest ID. */
if (WT_TXNID_LT(txn_global->oldest_id, oldest_id))
txn_global->oldest_id = oldest_id;
if (WT_TXNID_LT(txn_global->last_running, last_running)) {
txn_global->last_running = last_running;
#ifdef HAVE_VERBOSE
/* Output a verbose message about long-running transactions,
* but only when some progress is being made. */
if (WT_VERBOSE_ISSET(session, WT_VERB_TRANSACTION) &&
current_id - oldest_id > 10000 && oldest_session != NULL) {
__wt_verbose(session, WT_VERB_TRANSACTION,
"old snapshot %" PRIu64
" pinned in session %" PRIu32 " [%s]"
" with snap_min %" PRIu64 "\n",
oldest_id, oldest_session->id,
oldest_session->lastop,
oldest_session->txn.snap_min);
}
#endif
}
done: __wt_writeunlock(session, txn_global->scan_rwlock);
return (ret);
}
/*
* __ckpt_update --
* Update a checkpoint.
*/
static int
__ckpt_update(WT_SESSION_IMPL *session,
WT_BLOCK *block, WT_CKPT *ckpt, WT_BLOCK_CKPT *ci, bool is_live)
{
WT_DECL_ITEM(tmp);
WT_DECL_RET;
uint8_t *endp;
#ifdef HAVE_DIAGNOSTIC
/* Check the extent list combinations for overlaps. */
WT_RET(__wt_block_extlist_check(session, &ci->alloc, &ci->avail));
WT_RET(__wt_block_extlist_check(session, &ci->discard, &ci->avail));
WT_RET(__wt_block_extlist_check(session, &ci->alloc, &ci->discard));
#endif
/*
* Write the checkpoint's alloc and discard extent lists. After each
* write, remove any allocated blocks from the system's allocation
* list, checkpoint extent blocks don't appear on any extent lists.
*/
WT_RET(__wt_block_extlist_write(session, block, &ci->alloc, NULL));
WT_RET(__wt_block_extlist_write(session, block, &ci->discard, NULL));
/*
* We only write an avail list for the live system, other checkpoint's
* avail lists are static and never change.
*
* Write the avail list last so it reflects changes due to allocating
* blocks for the alloc and discard lists. Second, when we write the
* live system's avail list, it's two lists: the current avail list
* plus the list of blocks to be made available when the new checkpoint
* completes. We can't merge that second list into the real list yet,
* it's not truly available until the new checkpoint locations have been
* saved to the metadata.
*/
if (is_live)
WT_RET(__wt_block_extlist_write(
session, block, &ci->avail, &ci->ckpt_avail));
/*
* Set the file size for the live system.
*
* !!!
* We do NOT set the file size when re-writing checkpoints because we
* want to test the checkpoint's blocks against a reasonable maximum
* file size during verification. This is bad: imagine a checkpoint
* appearing early in the file, re-written, and then the checkpoint
* requires blocks at the end of the file, blocks after the listed file
* size. If the application opens that checkpoint for writing
* (discarding subsequent checkpoints), we would truncate the file to
* the early chunk, discarding the re-written checkpoint information.
* The alternative, updating the file size has its own problems, in
* that case we'd work correctly, but we'd lose all of the blocks
* between the original checkpoint and the re-written checkpoint.
* Currently, there's no API to roll-forward intermediate checkpoints,
* if there ever is, this will need to be fixed.
*/
if (is_live)
ci->file_size = block->size;
/*
* Copy the checkpoint information into the checkpoint array's address
* cookie.
*/
WT_RET(__wt_buf_init(session, &ckpt->raw, WT_BTREE_MAX_ADDR_COOKIE));
endp = ckpt->raw.mem;
WT_RET(__wt_block_ckpt_to_buffer(session, block, &endp, ci));
ckpt->raw.size = WT_PTRDIFF(endp, ckpt->raw.mem);
if (WT_VERBOSE_ISSET(session, WT_VERB_CHECKPOINT)) {
WT_RET(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(__ckpt_string(session, block, ckpt->raw.data, tmp));
__wt_verbose(session, WT_VERB_CHECKPOINT,
"%s: create-checkpoint: %s: %s",
block->name, ckpt->name, (const char *)tmp->data);
}
err: __wt_scr_free(session, &tmp);
return (ret);
}
/*
* __wt_block_checkpoint_load --
* Load a checkpoint.
*/
int
__wt_block_checkpoint_load(WT_SESSION_IMPL *session,
WT_BLOCK *block, WT_ITEM *dsk, const uint8_t *addr, uint32_t addr_size,
int readonly)
{
WT_BLOCK_CKPT *ci;
WT_DECL_ITEM(tmp);
WT_DECL_RET;
WT_UNUSED(addr_size);
/*
* Sometimes we don't find a root page (we weren't given a checkpoint,
* or the referenced checkpoint was empty). In that case we return a
* root page size of 0. Set that up now.
*/
dsk->size = 0;
ci = &block->live;
WT_RET(__wt_block_ckpt_init(session, block, ci, "live", 1));
if (WT_VERBOSE_ISSET(session, ckpt)) {
if (addr != NULL) {
WT_ERR(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(__ckpt_string(session, block, addr, tmp));
}
WT_VERBOSE_ERR(session, ckpt,
"%s: load-checkpoint: %s", block->name,
addr == NULL ? "[Empty]" : (char *)tmp->data);
}
/* If not loading a checkpoint from disk, we're done. */
if (addr == NULL || addr_size == 0)
return (0);
/* Crack the checkpoint cookie. */
if (addr != NULL)
WT_ERR(__wt_block_buffer_to_ckpt(session, block, addr, ci));
/* Verify sets up next. */
if (block->verify)
WT_ERR(__wt_verify_ckpt_load(session, block, ci));
/* Read, and optionally verify, any root page. */
if (ci->root_offset != WT_BLOCK_INVALID_OFFSET) {
WT_ERR(__wt_block_read_off(session, block,
dsk, ci->root_offset, ci->root_size, ci->root_cksum));
if (block->verify) {
if (tmp == NULL) {
WT_ERR(__wt_scr_alloc(session, 0, &tmp));
WT_ERR(__ckpt_string(
session, block, addr, tmp));
}
WT_ERR(
__wt_verify_dsk(session, (char *)tmp->data, dsk));
}
}
/*
* Rolling a checkpoint forward requires the avail list, the blocks from
* which we can allocate.
*/
if (!readonly)
WT_ERR(
__wt_block_extlist_read_avail(session, block, &ci->avail));
/*
* If the checkpoint can be written, that means anything written after
* the checkpoint is no longer interesting, truncate the file. Don't
* bother checking the avail list for a block at the end of the file,
* that was done when the checkpoint was first written (re-writing the
* checkpoint might possibly make it relevant here, but it's unlikely
* enough that I'm not bothering).
*/
if (!readonly) {
WT_VERBOSE_ERR(session, ckpt,
"truncate file to %" PRIuMAX, (uintmax_t)ci->file_size);
WT_ERR(__wt_ftruncate(session, block->fh, ci->file_size));
}
if (0) {
err: (void)__wt_block_checkpoint_unload(session, block);
}
__wt_scr_free(&tmp);
return (ret);
}
