/*
* __log_server --
* The log server thread.
*/
static WT_THREAD_RET
__log_server(void *arg)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_LOG *log;
WT_SESSION_IMPL *session;
int freq_per_sec, signalled;
session = arg;
conn = S2C(session);
log = conn->log;
signalled = 0;
/*
* Set this to the number of times per second we want to force out the
* log slot buffer.
*/
#define WT_FORCE_PER_SECOND 20
freq_per_sec = WT_FORCE_PER_SECOND;
/*
* The log server thread does a variety of work. It forces out any
* buffered log writes. It pre-allocates log files and it performs
* log archiving. The reason the wrlsn thread does not force out
* the buffered writes is because we want to process and move the
* write_lsn forward as quickly as possible. The same reason applies
* to why the log file server thread does not force out the writes.
* That thread does fsync calls which can take a long time and we
* don't want log records sitting in the buffer over the time it
* takes to sync out an earlier file.
*/
while (F_ISSET(conn, WT_CONN_LOG_SERVER_RUN)) {
/*
* Slots depend on future activity. Force out buffered
* writes in case we are idle. This cannot be part of the
* wrlsn thread because of interaction advancing the write_lsn
* and a buffer may need to wait for the write_lsn to advance
* in the case of a synchronous buffer. We end up with a hang.
*/
WT_ERR_BUSY_OK(__wt_log_force_write(session, 0));
/*
* We don't want to archive or pre-allocate files as often as
* we want to force out log buffers. Only do it once per second
* or if the condition was signalled.
*/
if (--freq_per_sec <= 0 || signalled != 0) {
freq_per_sec = WT_FORCE_PER_SECOND;
/*
* Perform log pre-allocation.
*/
if (conn->log_prealloc > 0)
WT_ERR(__log_prealloc_once(session));
/*
* Perform the archive.
*/
if (FLD_ISSET(conn->log_flags, WT_CONN_LOG_ARCHIVE)) {
if (__wt_try_writelock(
session, log->log_archive_lock) == 0) {
ret = __log_archive_once(session, 0);
WT_TRET(__wt_writeunlock(
session, log->log_archive_lock));
WT_ERR(ret);
} else
WT_ERR(
__wt_verbose(session, WT_VERB_LOG,
"log_archive: Blocked due to open "
"log cursor holding archive lock"));
}
}
/* Wait until the next event. */
WT_ERR(__wt_cond_wait_signal(session, conn->log_cond,
WT_MILLION / WT_FORCE_PER_SECOND, &signalled));
}
if (0) {
err: __wt_err(session, ret, "log server error");
}
return (WT_THREAD_RET_VALUE);
}
/*
* __log_server --
* The log server thread.
*/
static WT_THREAD_RET
__log_server(void *arg)
{
struct timespec start, now;
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_LOG *log;
WT_SESSION_IMPL *session;
uint64_t timediff;
bool did_work, locked, signalled;
session = arg;
conn = S2C(session);
log = conn->log;
locked = signalled = false;
/*
* Set this to the number of milliseconds we want to run archive and
* pre-allocation. Start it so that we run on the first time through.
*/
timediff = WT_THOUSAND;
/*
* The log server thread does a variety of work. It forces out any
* buffered log writes. It pre-allocates log files and it performs
* log archiving. The reason the wrlsn thread does not force out
* the buffered writes is because we want to process and move the
* write_lsn forward as quickly as possible. The same reason applies
* to why the log file server thread does not force out the writes.
* That thread does fsync calls which can take a long time and we
* don't want log records sitting in the buffer over the time it
* takes to sync out an earlier file.
*/
did_work = true;
while (F_ISSET(conn, WT_CONN_LOG_SERVER_RUN)) {
/*
* Slots depend on future activity. Force out buffered
* writes in case we are idle. This cannot be part of the
* wrlsn thread because of interaction advancing the write_lsn
* and a buffer may need to wait for the write_lsn to advance
* in the case of a synchronous buffer. We end up with a hang.
*/
WT_ERR_BUSY_OK(__wt_log_force_write(session, 0, &did_work));
/*
* We don't want to archive or pre-allocate files as often as
* we want to force out log buffers. Only do it once per second
* or if the condition was signalled.
*/
if (timediff >= WT_THOUSAND || signalled) {
/*
* Perform log pre-allocation.
*/
if (conn->log_prealloc > 0) {
/*
* Log file pre-allocation is disabled when a
* hot backup cursor is open because we have
* agreed not to rename or remove any files in
* the database directory.
*/
WT_ERR(__wt_readlock(
session, conn->hot_backup_lock));
locked = true;
if (!conn->hot_backup)
WT_ERR(__log_prealloc_once(session));
WT_ERR(__wt_readunlock(
session, conn->hot_backup_lock));
locked = false;
}
/*
* Perform the archive.
*/
if (FLD_ISSET(conn->log_flags, WT_CONN_LOG_ARCHIVE)) {
if (__wt_try_writelock(
session, log->log_archive_lock) == 0) {
ret = __log_archive_once(session, 0);
WT_TRET(__wt_writeunlock(
session, log->log_archive_lock));
WT_ERR(ret);
} else
WT_ERR(
__wt_verbose(session, WT_VERB_LOG,
"log_archive: Blocked due to open "
"log cursor holding archive lock"));
}
}
/* Wait until the next event. */
WT_ERR(__wt_epoch(session, &start));
WT_ERR(__wt_cond_auto_wait_signal(session, conn->log_cond,
did_work, &signalled));
WT_ERR(__wt_epoch(session, &now));
timediff = WT_TIMEDIFF_MS(now, start);
}
if (0) {
err: __wt_err(session, ret, "log server error");
if (locked)
WT_TRET(__wt_readunlock(
session, conn->hot_backup_lock));
}
return (WT_THREAD_RET_VALUE);
}
/*
* __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);
}
/*
* __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;
WT_UNUSED(addr_size);
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;
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_ERR(__wt_verbose(session, WT_VERB_CHECKPOINT,
"%s: load-checkpoint: %s", block->name,
addr == NULL ? "[Empty]" : (const char *)tmp->data));
}
/*
* 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 should we open the live system in two handles, or
* if we create, salvage, truncate or verify the live/running
* file, for that matter.
*/
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_cksum));
*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) {
/*
* The truncate might fail if there's a file mapping (if there's
* an open checkpoint on the file), that's OK.
*/
WT_ERR(__wt_verbose(session, WT_VERB_CHECKPOINT,
"truncate file to %" PRIuMAX, (uintmax_t)ci->file_size));
WT_ERR_BUSY_OK(
__wt_block_truncate(session, block->fh, 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);
}
