/*
* __log_wrlsn_server --
* The log wrlsn server thread.
*/
static WT_THREAD_RET
__log_wrlsn_server(void *arg)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_LOG *log;
WT_LSN prev;
WT_SESSION_IMPL *session;
int yield;
bool did_work;
session = arg;
conn = S2C(session);
log = conn->log;
yield = 0;
WT_INIT_LSN(&prev);
did_work = false;
while (F_ISSET(conn, WT_CONN_LOG_SERVER_RUN)) {
/*
* Write out any log record buffers if anything was done
* since last time. Only call the function to walk the
* slots if the system is not idle. On an idle system
* the alloc_lsn will not advance and the written lsn will
* match the alloc_lsn.
*/
if (__wt_log_cmp(&prev, &log->alloc_lsn) != 0 ||
__wt_log_cmp(&log->write_lsn, &log->alloc_lsn) != 0)
WT_ERR(__wt_log_wrlsn(session, &yield));
else
WT_STAT_FAST_CONN_INCR(session, log_write_lsn_skip);
prev = log->alloc_lsn;
if (yield == 0)
did_work = true;
else
did_work = false;
/*
* If __wt_log_wrlsn did work we want to yield instead of sleep.
*/
if (yield++ < WT_THOUSAND)
__wt_yield();
else
/*
* Send in false because if we did any work we would
* not be on this path.
*/
WT_ERR(__wt_cond_auto_wait(
session, conn->log_wrlsn_cond, did_work));
}
/*
* On close we need to do this one more time because there could
* be straggling log writes that need to be written.
*/
WT_ERR(__wt_log_force_write(session, 1, NULL));
WT_ERR(__wt_log_wrlsn(session, NULL));
if (0) {
err: __wt_err(session, ret, "log wrlsn server error");
}
return (WT_THREAD_RET_VALUE);
}
/*
* __recovery_cursor --
* Get a cursor for a recovery operation.
*/
static int
__recovery_cursor(WT_SESSION_IMPL *session, WT_RECOVERY *r,
WT_LSN *lsnp, u_int id, bool duplicate, WT_CURSOR **cp)
{
WT_CURSOR *c;
bool metadata_op;
const char *cfg[] = { WT_CONFIG_BASE(
session, WT_SESSION_open_cursor), "overwrite", NULL };
c = NULL;
/*
* Metadata operations have an id of 0. Match operations based
* on the id and the current pass of recovery for metadata.
*
* Only apply operations in the correct metadata phase, and if the LSN
* is more recent than the last checkpoint. If there is no entry for a
* file, assume it was dropped or missing after a hot backup.
*/
metadata_op = id == WT_METAFILE_ID;
if (r->metadata_only != metadata_op)
;
else if (id >= r->nfiles || r->files[id].uri == NULL) {
/* If a file is missing, output a verbose message once. */
if (!r->missing)
WT_RET(__wt_verbose(session, WT_VERB_RECOVERY,
"No file found with ID %u (max %u)",
id, r->nfiles));
r->missing = true;
} else if (__wt_log_cmp(lsnp, &r->files[id].ckpt_lsn) >= 0) {
/*
* We're going to apply the operation. Get the cursor, opening
* one if none is cached.
*/
if ((c = r->files[id].c) == NULL) {
WT_RET(__wt_open_cursor(
session, r->files[id].uri, NULL, cfg, &c));
r->files[id].c = c;
}
}
if (duplicate && c != NULL)
WT_RET(__wt_open_cursor(
session, r->files[id].uri, NULL, cfg, &c));
*cp = c;
return (0);
}
/*
* __recovery_setup_file --
* Set up the recovery slot for a file.
*/
static int
__recovery_setup_file(WT_RECOVERY *r, const char *uri, const char *config)
{
WT_CONFIG_ITEM cval;
WT_LSN lsn;
uint32_t fileid, lsnfile, lsnoffset;
WT_RET(__wt_config_getones(r->session, config, "id", &cval));
fileid = (uint32_t)cval.val;
/* Track the largest file ID we have seen. */
if (fileid > r->max_fileid)
r->max_fileid = fileid;
if (r->nfiles <= fileid) {
WT_RET(__wt_realloc_def(
r->session, &r->file_alloc, fileid + 1, &r->files));
r->nfiles = fileid + 1;
}
WT_RET(__wt_strdup(r->session, uri, &r->files[fileid].uri));
WT_RET(
__wt_config_getones(r->session, config, "checkpoint_lsn", &cval));
/* If there is checkpoint logged for the file, apply everything. */
if (cval.type != WT_CONFIG_ITEM_STRUCT)
WT_INIT_LSN(&lsn);
/* NOLINTNEXTLINE(cert-err34-c) */
else if (sscanf(cval.str,
"(%" SCNu32 ",%" SCNu32 ")", &lsnfile, &lsnoffset) == 2)
WT_SET_LSN(&lsn, lsnfile, lsnoffset);
else
WT_RET_MSG(r->session, EINVAL,
"Failed to parse checkpoint LSN '%.*s'",
(int)cval.len, cval.str);
r->files[fileid].ckpt_lsn = lsn;
__wt_verbose(r->session, WT_VERB_RECOVERY,
"Recovering %s with id %" PRIu32 " @ (%" PRIu32 ", %" PRIu32 ")",
uri, fileid, lsn.l.file, lsn.l.offset);
if ((!WT_IS_MAX_LSN(&lsn) && !WT_IS_INIT_LSN(&lsn)) &&
(WT_IS_MAX_LSN(&r->max_ckpt_lsn) ||
__wt_log_cmp(&lsn, &r->max_ckpt_lsn) > 0))
r->max_ckpt_lsn = lsn;
return (0);
}
/*
* __txn_log_recover --
* Roll the log forward to recover committed changes.
*/
static int
__txn_log_recover(WT_SESSION_IMPL *session,
WT_ITEM *logrec, WT_LSN *lsnp, WT_LSN *next_lsnp,
void *cookie, int firstrecord)
{
WT_DECL_RET;
WT_RECOVERY *r;
uint64_t txnid_unused;
uint32_t rectype;
const uint8_t *end, *p;
r = cookie;
p = WT_LOG_SKIP_HEADER(logrec->data);
end = (const uint8_t *)logrec->data + logrec->size;
WT_UNUSED(firstrecord);
/* First, peek at the log record type. */
WT_RET(__wt_logrec_read(session, &p, end, &rectype));
/*
* Record the highest LSN we process during the metadata phase.
* If not the metadata phase, then stop at that LSN.
*/
if (r->metadata_only)
r->max_rec_lsn = *next_lsnp;
else if (__wt_log_cmp(lsnp, &r->max_rec_lsn) >= 0)
return (0);
switch (rectype) {
case WT_LOGREC_CHECKPOINT:
if (r->metadata_only)
WT_RET(__wt_txn_checkpoint_logread(
session, &p, end, &r->ckpt_lsn));
break;
case WT_LOGREC_COMMIT:
if ((ret = __wt_vunpack_uint(
&p, WT_PTRDIFF(end, p), &txnid_unused)) != 0)
WT_RET_MSG(
session, ret, "txn_log_recover: unpack failure");
WT_RET(__txn_commit_apply(r, lsnp, &p, end));
break;
}
return (0);
}
/*
* __curlog_compare --
* WT_CURSOR.compare method for the log cursor type.
*/
static int
__curlog_compare(WT_CURSOR *a, WT_CURSOR *b, int *cmpp)
{
WT_CURSOR_LOG *acl, *bcl;
WT_DECL_RET;
WT_SESSION_IMPL *session;
CURSOR_API_CALL(a, session, compare, NULL);
acl = (WT_CURSOR_LOG *)a;
bcl = (WT_CURSOR_LOG *)b;
WT_ASSERT(session, cmpp != NULL);
*cmpp = __wt_log_cmp(acl->cur_lsn, bcl->cur_lsn);
/*
* If both are on the same LSN, compare step counter.
*/
if (*cmpp == 0)
*cmpp = (acl->step_count != bcl->step_count ?
(acl->step_count < bcl->step_count ? -1 : 1) : 0);
err: API_END_RET(session, ret);
}
/*
* __log_slot_dump --
* Dump the entire slot state.
*/
static void
__log_slot_dump(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
WT_LOG *log;
WT_LOGSLOT *slot;
int earliest, i;
conn = S2C(session);
log = conn->log;
earliest = 0;
for (i = 0; i < WT_SLOT_POOL; i++) {
slot = &log->slot_pool[i];
if (__wt_log_cmp(&slot->slot_release_lsn,
&log->slot_pool[earliest].slot_release_lsn) < 0)
earliest = i;
__wt_errx(session, "Slot %d:", i);
__wt_errx(session, " State: %" PRIx64 " Flags: %" PRIx32,
(uint64_t)slot->slot_state, slot->flags);
__wt_errx(session, " Start LSN: %" PRIu32 "/%" PRIu32,
slot->slot_start_lsn.l.file, slot->slot_start_lsn.l.offset);
__wt_errx(session, " End LSN: %" PRIu32 "/%" PRIu32,
slot->slot_end_lsn.l.file, slot->slot_end_lsn.l.offset);
__wt_errx(session, " Release LSN: %" PRIu32 "/%" PRIu32,
slot->slot_release_lsn.l.file,
slot->slot_release_lsn.l.offset);
__wt_errx(session, " Offset: start: %" PRIuMAX
" last:%" PRIuMAX, (uintmax_t)slot->slot_start_offset,
(uintmax_t)slot->slot_last_offset);
__wt_errx(session, " Unbuffered: %" PRId64
" error: %" PRId32, slot->slot_unbuffered,
slot->slot_error);
}
__wt_errx(session, "Earliest slot: %d", earliest);
}
/*
* __wt_log_wrlsn --
* Process written log slots and attempt to coalesce them if the LSNs
* are contiguous. The purpose of this function is to advance the
* write_lsn in LSN order after the buffer is written to the log file.
*/
int
__wt_log_wrlsn(WT_SESSION_IMPL *session)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_LOG *log;
WT_LOG_WRLSN_ENTRY written[WT_SLOT_POOL];
WT_LOGSLOT *coalescing, *slot;
WT_LSN save_lsn;
size_t written_i;
uint32_t i, save_i;
conn = S2C(session);
log = conn->log;
__wt_spin_lock(session, &log->log_writelsn_lock);
restart:
coalescing = NULL;
WT_INIT_LSN(&save_lsn);
written_i = 0;
i = 0;
/*
* Walk the array once saving any slots that are in the
* WT_LOG_SLOT_WRITTEN state.
*/
while (i < WT_SLOT_POOL) {
save_i = i;
slot = &log->slot_pool[i++];
/*
* XXX - During debugging I saw slot 0 become orphaned.
* I believe it is fixed, but check for now.
* This assertion should catch that.
*/
if (slot->slot_state == 0)
WT_ASSERT(session,
slot->slot_release_lsn.file >= log->write_lsn.file);
if (slot->slot_state != WT_LOG_SLOT_WRITTEN)
continue;
written[written_i].slot_index = save_i;
written[written_i++].lsn = slot->slot_release_lsn;
}
/*
* If we found any written slots process them. We sort them
* based on the release LSN, and then look for them in order.
*/
if (written_i > 0) {
WT_INSERTION_SORT(written, written_i,
WT_LOG_WRLSN_ENTRY, WT_WRLSN_ENTRY_CMP_LT);
/*
* We know the written array is sorted by LSN. Go
* through them either advancing write_lsn or coalesce
* contiguous ranges of written slots.
*/
for (i = 0; i < written_i; i++) {
slot = &log->slot_pool[written[i].slot_index];
/*
* The log server thread pushes out slots periodically.
* Sometimes they are empty slots. If we find an
* empty slot, where empty means the start and end LSN
* are the same, free it and continue.
*/
if (__wt_log_cmp(&slot->slot_start_lsn,
&slot->slot_release_lsn) == 0 &&
__wt_log_cmp(&slot->slot_start_lsn,
&slot->slot_end_lsn) == 0) {
__wt_log_slot_free(session, slot);
continue;
}
if (coalescing != NULL) {
/*
* If the write_lsn changed, we may be able to
* process slots. Try again.
*/
if (__wt_log_cmp(
&log->write_lsn, &save_lsn) != 0)
goto restart;
if (__wt_log_cmp(&coalescing->slot_end_lsn,
&written[i].lsn) != 0) {
coalescing = slot;
continue;
}
/*
* If we get here we have a slot to coalesce
* and free.
*/
coalescing->slot_last_offset =
slot->slot_last_offset;
coalescing->slot_end_lsn = slot->slot_end_lsn;
WT_STAT_FAST_CONN_INCR(
session, log_slot_coalesced);
/*
* Copy the flag for later closing.
*/
if (F_ISSET(slot, WT_SLOT_CLOSEFH))
F_SET(coalescing, WT_SLOT_CLOSEFH);
} else {
/*
* If this written slot is not the next LSN,
* try to start coalescing with later slots.
* A synchronous write may update write_lsn
* so save the last one we saw to check when
* coalescing slots.
*/
save_lsn = log->write_lsn;
if (__wt_log_cmp(
&log->write_lsn, &written[i].lsn) != 0) {
coalescing = slot;
continue;
}
/*
* If we get here we have a slot to process.
* Advance the LSN and process the slot.
*/
WT_ASSERT(session, __wt_log_cmp(&written[i].lsn,
&slot->slot_release_lsn) == 0);
if (slot->slot_start_lsn.offset !=
slot->slot_last_offset)
slot->slot_start_lsn.offset =
slot->slot_last_offset;
log->write_start_lsn = slot->slot_start_lsn;
log->write_lsn = slot->slot_end_lsn;
WT_ERR(__wt_cond_signal(
session, log->log_write_cond));
WT_STAT_FAST_CONN_INCR(session, log_write_lsn);
/*
* Signal the close thread if needed.
*/
if (F_ISSET(slot, WT_SLOT_CLOSEFH))
WT_ERR(__wt_cond_signal(
session, conn->log_file_cond));
}
__wt_log_slot_free(session, slot);
}
}
err: __wt_spin_unlock(session, &log->log_writelsn_lock);
return (ret);
}
/*
* __log_file_server --
* The log file server thread. This worker thread manages
* log file operations such as closing and syncing.
*/
static WT_THREAD_RET
__log_file_server(void *arg)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_FH *close_fh;
WT_LOG *log;
WT_LSN close_end_lsn, min_lsn;
WT_SESSION_IMPL *session;
uint32_t filenum;
int locked;
session = arg;
conn = S2C(session);
log = conn->log;
locked = 0;
while (F_ISSET(conn, WT_CONN_LOG_SERVER_RUN)) {
/*
* If there is a log file to close, make sure any outstanding
* write operations have completed, then fsync and close it.
*/
if ((close_fh = log->log_close_fh) != NULL) {
WT_ERR(__wt_log_extract_lognum(session, close_fh->name,
&filenum));
/*
* We update the close file handle before updating the
* close LSN when changing files. It is possible we
* could see mismatched settings. If we do, yield
* until it is set. This should rarely happen.
*/
while (log->log_close_lsn.file < filenum)
__wt_yield();
if (__wt_log_cmp(
&log->write_lsn, &log->log_close_lsn) >= 0) {
/*
* We've copied the file handle, clear out the
* one in the log structure to allow it to be
* set again. Copy the LSN before clearing
* the file handle.
* Use a barrier to make sure the compiler does
* not reorder the following two statements.
*/
close_end_lsn = log->log_close_lsn;
WT_FULL_BARRIER();
log->log_close_fh = NULL;
/*
* Set the close_end_lsn to the LSN immediately
* after ours. That is, the beginning of the
* next log file. We need to know the LSN
* file number of our own close in case earlier
* calls are still in progress and the next one
* to move the sync_lsn into the next file for
* later syncs.
*/
close_end_lsn.file++;
close_end_lsn.offset = 0;
WT_ERR(__wt_fsync(session, close_fh));
__wt_spin_lock(session, &log->log_sync_lock);
locked = 1;
WT_ERR(__wt_close(session, &close_fh));
WT_ASSERT(session, __wt_log_cmp(
&close_end_lsn, &log->sync_lsn) >= 0);
log->sync_lsn = close_end_lsn;
WT_ERR(__wt_cond_signal(
session, log->log_sync_cond));
locked = 0;
__wt_spin_unlock(session, &log->log_sync_lock);
}
}
/*
* If a later thread asked for a background sync, do it now.
*/
if (__wt_log_cmp(&log->bg_sync_lsn, &log->sync_lsn) > 0) {
/*
* Save the latest write LSN which is the minimum
* we will have written to disk.
*/
min_lsn = log->write_lsn;
/*
* We have to wait until the LSN we asked for is
* written. If it isn't signal the wrlsn thread
* to get it written.
*/
if (__wt_log_cmp(&log->bg_sync_lsn, &min_lsn) <= 0) {
WT_ERR(__wt_fsync(session, log->log_fh));
__wt_spin_lock(session, &log->log_sync_lock);
locked = 1;
/*
* The sync LSN could have advanced while we
* were writing to disk.
*/
if (__wt_log_cmp(
&log->sync_lsn, &min_lsn) <= 0) {
log->sync_lsn = min_lsn;
WT_ERR(__wt_cond_signal(
session, log->log_sync_cond));
}
locked = 0;
__wt_spin_unlock(session, &log->log_sync_lock);
} else {
WT_ERR(__wt_cond_signal(
session, conn->log_wrlsn_cond));
/*
* We do not want to wait potentially a second
* to process this. Yield to give the wrlsn
* thread a chance to run and try again in
* this case.
*/
__wt_yield();
continue;
}
}
/* Wait until the next event. */
WT_ERR(__wt_cond_wait(
session, conn->log_file_cond, WT_MILLION));
}
if (0) {
err: __wt_err(session, ret, "log close server error");
}
if (locked)
__wt_spin_unlock(session, &log->log_sync_lock);
return (WT_THREAD_RET_VALUE);
}
/*
* __log_file_server --
* The log file server thread. This worker thread manages
* log file operations such as closing and syncing.
*/
static WT_THREAD_RET
__log_file_server(void *arg)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_FH *close_fh;
WT_LOG *log;
WT_LSN close_end_lsn, min_lsn;
WT_SESSION_IMPL *session;
uint32_t filenum;
bool locked;
session = arg;
conn = S2C(session);
log = conn->log;
locked = false;
while (F_ISSET(conn, WT_CONN_LOG_SERVER_RUN)) {
/*
* If there is a log file to close, make sure any outstanding
* write operations have completed, then fsync and close it.
*/
if ((close_fh = log->log_close_fh) != NULL) {
WT_ERR(__wt_log_extract_lognum(session, close_fh->name,
&filenum));
/*
* We update the close file handle before updating the
* close LSN when changing files. It is possible we
* could see mismatched settings. If we do, yield
* until it is set. This should rarely happen.
*/
while (log->log_close_lsn.l.file < filenum)
__wt_yield();
if (__wt_log_cmp(
&log->write_lsn, &log->log_close_lsn) >= 0) {
/*
* We've copied the file handle, clear out the
* one in the log structure to allow it to be
* set again. Copy the LSN before clearing
* the file handle.
* Use a barrier to make sure the compiler does
* not reorder the following two statements.
*/
close_end_lsn = log->log_close_lsn;
WT_FULL_BARRIER();
log->log_close_fh = NULL;
/*
* Set the close_end_lsn to the LSN immediately
* after ours. That is, the beginning of the
* next log file. We need to know the LSN
* file number of our own close in case earlier
* calls are still in progress and the next one
* to move the sync_lsn into the next file for
* later syncs.
*/
WT_ERR(__wt_fsync(session, close_fh, true));
/*
* We want to have the file size reflect actual
* data with minimal pre-allocated zeroed space.
* We can't truncate the file during hot backup,
* or the underlying file system may not support
* truncate: both are OK, it's just more work
* during cursor traversal.
*/
if (!conn->hot_backup) {
__wt_readlock(
session, conn->hot_backup_lock);
if (!conn->hot_backup)
WT_ERR_ERROR_OK(
__wt_ftruncate(session,
close_fh,
close_end_lsn.l.offset),
ENOTSUP);
__wt_readunlock(
session, conn->hot_backup_lock);
}
WT_SET_LSN(&close_end_lsn,
close_end_lsn.l.file + 1, 0);
__wt_spin_lock(session, &log->log_sync_lock);
locked = true;
WT_ERR(__wt_close(session, &close_fh));
WT_ASSERT(session, __wt_log_cmp(
&close_end_lsn, &log->sync_lsn) >= 0);
log->sync_lsn = close_end_lsn;
__wt_cond_signal(session, log->log_sync_cond);
locked = false;
__wt_spin_unlock(session, &log->log_sync_lock);
}
}
/*
* If a later thread asked for a background sync, do it now.
*/
if (__wt_log_cmp(&log->bg_sync_lsn, &log->sync_lsn) > 0) {
/*
* Save the latest write LSN which is the minimum
* we will have written to disk.
*/
min_lsn = log->write_lsn;
/*
* We have to wait until the LSN we asked for is
* written. If it isn't signal the wrlsn thread
* to get it written.
*
* We also have to wait for the written LSN and the
* sync LSN to be in the same file so that we know we
* have synchronized all earlier log files.
*/
if (__wt_log_cmp(&log->bg_sync_lsn, &min_lsn) <= 0) {
/*
* If the sync file is behind either the one
* wanted for a background sync or the write LSN
* has moved to another file continue to let
* this worker thread process that older file
* immediately.
*/
if ((log->sync_lsn.l.file <
log->bg_sync_lsn.l.file) ||
(log->sync_lsn.l.file < min_lsn.l.file))
continue;
WT_ERR(__wt_fsync(session, log->log_fh, true));
__wt_spin_lock(session, &log->log_sync_lock);
locked = true;
/*
* The sync LSN could have advanced while we
* were writing to disk.
*/
if (__wt_log_cmp(
&log->sync_lsn, &min_lsn) <= 0) {
WT_ASSERT(session,
min_lsn.l.file ==
log->sync_lsn.l.file);
log->sync_lsn = min_lsn;
__wt_cond_signal(
session, log->log_sync_cond);
}
locked = false;
__wt_spin_unlock(session, &log->log_sync_lock);
} else {
__wt_cond_auto_signal(
session, conn->log_wrlsn_cond);
/*
* We do not want to wait potentially a second
* to process this. Yield to give the wrlsn
* thread a chance to run and try again in
* this case.
*/
__wt_yield();
continue;
}
}
/* Wait until the next event. */
__wt_cond_wait(session, conn->log_file_cond, WT_MILLION / 10);
}
if (0) {
err: __wt_err(session, ret, "log close server error");
}
if (locked)
__wt_spin_unlock(session, &log->log_sync_lock);
return (WT_THREAD_RET_VALUE);
}
