/*
* __curlog_reset --
* WT_CURSOR.reset method for the log cursor type.
*/
static int
__curlog_reset(WT_CURSOR *cursor)
{
WT_CURSOR_LOG *cl;
cl = (WT_CURSOR_LOG *)cursor;
cl->stepp = cl->stepp_end = NULL;
cl->step_count = 0;
INIT_LSN(cl->cur_lsn);
INIT_LSN(cl->next_lsn);
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;
WT_RET(__wt_config_getones(r->session, config, "id", &cval));
fileid = (uint32_t)cval.val;
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)
INIT_LSN(&lsn);
else if (sscanf(cval.str, "(%" PRIu32 ",%" PRIdMAX ")",
&lsn.file, (intmax_t*)&lsn.offset) != 2)
WT_RET_MSG(r->session, EINVAL,
"Failed to parse checkpoint LSN '%.*s'",
(int)cval.len, cval.str);
r->files[fileid].ckpt_lsn = lsn;
WT_RET(__wt_verbose(r->session, WT_VERB_RECOVERY,
"Recovering %s with id %u @ (%" PRIu32 ", %" PRIu64 ")",
uri, fileid, lsn.file, lsn.offset));
return (0);
}
/*
* __wt_curlog_open --
* Initialize a log cursor.
*/
int
__wt_curlog_open(WT_SESSION_IMPL *session,
const char *uri, const char *cfg[], WT_CURSOR **cursorp)
{
WT_CONNECTION_IMPL *conn;
WT_CURSOR_STATIC_INIT(iface,
NULL, /* get-key */
NULL, /* get-value */
NULL, /* set-key */
NULL, /* set-value */
__curlog_compare, /* compare */
__curlog_next, /* next */
__wt_cursor_notsup, /* prev */
__curlog_reset, /* reset */
__curlog_search, /* search */
__wt_cursor_notsup, /* search-near */
__wt_cursor_notsup, /* insert */
__wt_cursor_notsup, /* update */
__wt_cursor_notsup, /* remove */
__curlog_close); /* close */
WT_CURSOR *cursor;
WT_CURSOR_LOG *cl;
WT_DECL_RET;
STATIC_ASSERT(offsetof(WT_CURSOR_LOG, iface) == 0);
conn = S2C(session);
if (!conn->logging)
WT_RET_MSG(session, EINVAL,
"Cannot open a log cursor without logging enabled");
cl = NULL;
WT_RET(__wt_calloc_def(session, 1, &cl));
cursor = &cl->iface;
*cursor = iface;
cursor->session = &session->iface;
WT_ERR(__wt_calloc_def(session, 1, &cl->cur_lsn));
WT_ERR(__wt_calloc_def(session, 1, &cl->next_lsn));
WT_ERR(__wt_scr_alloc(session, 0, &cl->logrec));
WT_ERR(__wt_scr_alloc(session, 0, &cl->opkey));
WT_ERR(__wt_scr_alloc(session, 0, &cl->opvalue));
cursor->key_format = LOGC_KEY_FORMAT;
cursor->value_format = LOGC_VALUE_FORMAT;
INIT_LSN(cl->cur_lsn);
INIT_LSN(cl->next_lsn);
WT_ERR(__wt_cursor_init(cursor, uri, NULL, cfg, cursorp));
/* Log cursors default to read only. */
WT_ERR(__wt_cursor_config_readonly(cursor, cfg, 1));
if (0) {
err: if (F_ISSET(cursor, WT_CURSTD_OPEN))
WT_TRET(cursor->close(cursor));
else {
__wt_free(session, cl->cur_lsn);
__wt_free(session, cl->next_lsn);
__wt_scr_free(&cl->logrec);
__wt_scr_free(&cl->opkey);
__wt_scr_free(&cl->opvalue);
__wt_free(session, cl);
}
*cursorp = NULL;
}
return (ret);
}
/*
* __wt_txn_checkpoint_log --
* Write a log record for a checkpoint operation.
*/
int
__wt_txn_checkpoint_log(
WT_SESSION_IMPL *session, int full, uint32_t flags, WT_LSN *lsnp)
{
WT_DECL_RET;
WT_DECL_ITEM(logrec);
WT_LSN *ckpt_lsn;
WT_TXN *txn;
const char *fmt = WT_UNCHECKED_STRING(IIQIU);
uint8_t *end, *p;
size_t recsize;
uint32_t i, rectype = WT_LOGREC_CHECKPOINT;
txn = &session->txn;
ckpt_lsn = &txn->ckpt_lsn;
/*
* If this is a file sync, log it unless there is a full checkpoint in
* progress.
*/
if (!full) {
if (txn->full_ckpt) {
if (lsnp != NULL)
*lsnp = *ckpt_lsn;
return (0);
} else
return (__txn_log_file_sync(session, flags, lsnp));
}
switch (flags) {
case WT_TXN_LOG_CKPT_PREPARE:
txn->full_ckpt = 1;
*ckpt_lsn = S2C(session)->log->alloc_lsn;
break;
case WT_TXN_LOG_CKPT_START:
/* Take a copy of the transaction snapshot. */
txn->ckpt_nsnapshot = txn->snapshot_count;
recsize = txn->ckpt_nsnapshot * WT_INTPACK64_MAXSIZE;
WT_ERR(__wt_scr_alloc(session, recsize, &txn->ckpt_snapshot));
p = txn->ckpt_snapshot->mem;
end = p + recsize;
for (i = 0; i < txn->snapshot_count; i++)
WT_ERR(__wt_vpack_uint(
&p, WT_PTRDIFF(end, p), txn->snapshot[i]));
break;
case WT_TXN_LOG_CKPT_STOP:
/*
* During a clean connection close, we get here without the
* prepare or start steps. In that case, log the current LSN
* as the checkpoint LSN.
*/
if (!txn->full_ckpt) {
txn->ckpt_nsnapshot = 0;
*ckpt_lsn = S2C(session)->log->alloc_lsn;
}
/* Write the checkpoint log record. */
WT_ERR(__wt_struct_size(session, &recsize, fmt,
rectype, ckpt_lsn->file, ckpt_lsn->offset,
txn->ckpt_nsnapshot, &txn->ckpt_snapshot));
WT_ERR(__wt_logrec_alloc(session, recsize, &logrec));
WT_ERR(__wt_struct_pack(session,
(uint8_t *)logrec->data + logrec->size, recsize, fmt,
rectype, ckpt_lsn->file, ckpt_lsn->offset,
txn->ckpt_nsnapshot, &txn->ckpt_snapshot));
logrec->size += (uint32_t)recsize;
WT_ERR(__wt_log_write(session, logrec, lsnp, 0));
/*
* If this full checkpoint completed successfully and there is
* no hot backup in progress, tell the logging subsystem the
* checkpoint LSN so that it can archive.
*/
if (!S2C(session)->hot_backup)
WT_ERR(__wt_log_ckpt(session, ckpt_lsn));
/* FALLTHROUGH */
case WT_TXN_LOG_CKPT_FAIL:
/* Cleanup any allocated resources */
INIT_LSN(ckpt_lsn);
txn->ckpt_nsnapshot = 0;
__wt_scr_free(&txn->ckpt_snapshot);
txn->full_ckpt = 0;
break;
}
err: __wt_logrec_free(session, &logrec);
return (ret);
}
/*
* __wt_curlog_open --
* Initialize a log cursor.
*/
int
__wt_curlog_open(WT_SESSION_IMPL *session,
const char *uri, const char *cfg[], WT_CURSOR **cursorp)
{
WT_CONNECTION_IMPL *conn;
WT_CURSOR_STATIC_INIT(iface,
__wt_cursor_get_key, /* get-key */
__wt_cursor_get_value, /* get-value */
__wt_cursor_set_key, /* set-key */
__wt_cursor_set_value, /* set-value */
__curlog_compare, /* compare */
__curlog_next, /* next */
__wt_cursor_notsup, /* prev */
__curlog_reset, /* reset */
__curlog_search, /* search */
__wt_cursor_notsup, /* search-near */
__wt_cursor_notsup, /* insert */
__wt_cursor_notsup, /* update */
__wt_cursor_notsup, /* remove */
__curlog_close); /* close */
WT_CURSOR *cursor;
WT_CURSOR_LOG *cl;
WT_DECL_RET;
WT_LOG *log;
WT_STATIC_ASSERT(offsetof(WT_CURSOR_LOG, iface) == 0);
conn = S2C(session);
if (!FLD_ISSET(conn->log_flags, WT_CONN_LOG_ENABLED))
WT_RET_MSG(session, EINVAL,
"Cannot open a log cursor without logging enabled");
log = conn->log;
cl = NULL;
WT_RET(__wt_calloc_def(session, 1, &cl));
cursor = &cl->iface;
*cursor = iface;
cursor->session = &session->iface;
WT_ERR(__wt_calloc_def(session, 1, &cl->cur_lsn));
WT_ERR(__wt_calloc_def(session, 1, &cl->next_lsn));
WT_ERR(__wt_scr_alloc(session, 0, &cl->logrec));
WT_ERR(__wt_scr_alloc(session, 0, &cl->opkey));
WT_ERR(__wt_scr_alloc(session, 0, &cl->opvalue));
cursor->key_format = LOGC_KEY_FORMAT;
cursor->value_format = LOGC_VALUE_FORMAT;
INIT_LSN(cl->cur_lsn);
INIT_LSN(cl->next_lsn);
WT_ERR(__wt_cursor_init(cursor, uri, NULL, cfg, cursorp));
/* Log cursors block archiving. */
WT_ERR(__wt_readlock(session, log->log_archive_lock));
if (0) {
err: if (F_ISSET(cursor, WT_CURSTD_OPEN))
WT_TRET(cursor->close(cursor));
else {
__wt_free(session, cl->cur_lsn);
__wt_free(session, cl->next_lsn);
__wt_scr_free(&cl->logrec);
__wt_scr_free(&cl->opkey);
__wt_scr_free(&cl->opvalue);
/*
* NOTE: We cannot get on the error path with the
* readlock held. No need to unlock it unless that
* changes above.
*/
__wt_free(session, cl);
}
*cursorp = NULL;
}
return (ret);
}
/*
* __wt_log_write --
* Write a record into the log.
*/
int
__wt_log_write(WT_SESSION_IMPL *session, WT_ITEM *record, WT_LSN *lsnp,
uint32_t flags)
{
WT_CONNECTION_IMPL *conn;
WT_DECL_RET;
WT_LOG *log;
WT_LOG_RECORD *logrec;
WT_LSN lsn;
WT_MYSLOT myslot;
uint32_t rdup_len;
int locked;
conn = S2C(session);
log = conn->log;
locked = 0;
INIT_LSN(&lsn);
myslot.slot = NULL;
/*
* Assume the WT_ITEM the user passed is a WT_LOG_RECORD, which has
* a header at the beginning for us to fill in.
*
* If using direct_io, the caller should pass us an aligned record.
* But we need to make sure it is big enough and zero-filled so
* that we can write the full amount. Do this whether or not
* direct_io is in use because it makes the reading code cleaner.
*/
WT_STAT_FAST_CONN_INCRV(session, log_bytes_user, record->size);
rdup_len = __wt_rduppo2((uint32_t)record->size, log->allocsize);
WT_ERR(__wt_buf_grow(session, record, rdup_len));
WT_ASSERT(session, record->data == record->mem);
/*
* If the caller's record only partially fills the necessary
* space, we need to zero-fill the remainder.
*/
if (record->size != rdup_len) {
memset((uint8_t *)record->mem + record->size, 0,
rdup_len - record->size);
record->size = rdup_len;
}
logrec = (WT_LOG_RECORD *)record->mem;
logrec->len = (uint32_t)record->size;
logrec->checksum = 0;
logrec->checksum = __wt_cksum(logrec, record->size);
WT_STAT_FAST_CONN_INCR(session, log_writes);
if (!F_ISSET(log, WT_LOG_FORCE_CONSOLIDATE)) {
ret = __log_direct_write(session, record, lsnp, flags);
if (ret == 0)
return (0);
if (ret != EAGAIN)
WT_ERR(ret);
/*
* An EAGAIN return means we failed to get the try lock -
* fall through to the consolidation code in that case.
*/
}
/*
* As soon as we see contention for the log slot, disable direct
* log writes. We get better performance by forcing writes through
* the consolidation code. This is because individual writes flood
* the I/O system faster than they contend on the log slot lock.
*/
F_SET(log, WT_LOG_FORCE_CONSOLIDATE);
if ((ret = __wt_log_slot_join(
session, rdup_len, flags, &myslot)) == ENOMEM) {
/*
* If we couldn't find a consolidated slot for this record
* write the record directly.
*/
while ((ret = __log_direct_write(
session, record, lsnp, flags)) == EAGAIN)
;
WT_ERR(ret);
/*
* Increase the buffer size of any slots we can get access
* to, so future consolidations are likely to succeed.
*/
WT_ERR(__wt_log_slot_grow_buffers(session, 4 * rdup_len));
return (0);
}
WT_ERR(ret);
if (myslot.offset == 0) {
__wt_spin_lock(session, &log->log_slot_lock);
locked = 1;
WT_ERR(__wt_log_slot_close(session, myslot.slot));
WT_ERR(__log_acquire(
session, myslot.slot->slot_group_size, myslot.slot));
__wt_spin_unlock(session, &log->log_slot_lock);
locked = 0;
WT_ERR(__wt_log_slot_notify(session, myslot.slot));
} else
WT_ERR(__wt_log_slot_wait(session, myslot.slot));
WT_ERR(__log_fill(session, &myslot, 0, record, &lsn));
if (__wt_log_slot_release(myslot.slot, rdup_len) == WT_LOG_SLOT_DONE) {
WT_ERR(__log_release(session, myslot.slot));
WT_ERR(__wt_log_slot_free(myslot.slot));
} else if (LF_ISSET(WT_LOG_FSYNC)) {
/* Wait for our writes to reach disk */
while (LOG_CMP(&log->sync_lsn, &lsn) <= 0 &&
myslot.slot->slot_error == 0)
(void)__wt_cond_wait(
session, log->log_sync_cond, 10000);
}
err:
if (locked)
__wt_spin_unlock(session, &log->log_slot_lock);
if (ret == 0 && lsnp != NULL)
*lsnp = lsn;
/*
* If we're synchronous and some thread had an error, we don't know
* if our write made it out to the file or not. The error could be
* before or after us. So, if anyone got an error, we report it.
* If we're not synchronous, only report if our own operation got
* an error.
*/
if (LF_ISSET(WT_LOG_DSYNC | WT_LOG_FSYNC) && ret == 0 &&
myslot.slot != NULL)
ret = myslot.slot->slot_error;
return (ret);
}
/*
* __wt_txn_recover --
* Run recovery.
*/
int
__wt_txn_recover(WT_CONNECTION_IMPL *conn)
{
WT_CURSOR *metac;
WT_DECL_RET;
WT_RECOVERY r;
WT_SESSION_IMPL *session;
struct WT_RECOVERY_FILE *metafile;
char *config;
int was_backup;
WT_CLEAR(r);
INIT_LSN(&r.ckpt_lsn);
was_backup = F_ISSET(conn, WT_CONN_WAS_BACKUP) ? 1 : 0;
/* We need a real session for recovery. */
WT_RET(__wt_open_session(conn, NULL, NULL, &session));
F_SET(session, WT_SESSION_NO_LOGGING);
r.session = session;
WT_ERR(__wt_metadata_search(session, WT_METAFILE_URI, &config));
WT_ERR(__recovery_setup_file(&r, WT_METAFILE_URI, config));
WT_ERR(__wt_metadata_cursor(session, NULL, &metac));
metafile = &r.files[WT_METAFILE_ID];
metafile->c = metac;
/*
* First, do a pass through the log to recover the metadata, and
* establish the last checkpoint LSN. Skip this when opening a hot
* backup: we already have the correct metadata in that case.
*/
if (!was_backup) {
r.metadata_only = 1;
if (IS_INIT_LSN(&metafile->ckpt_lsn))
WT_ERR(__wt_log_scan(session,
NULL, WT_LOGSCAN_FIRST, __txn_log_recover, &r));
else
WT_ERR(__wt_log_scan(session,
&metafile->ckpt_lsn, 0, __txn_log_recover, &r));
WT_ASSERT(session,
LOG_CMP(&r.ckpt_lsn, &conn->log->first_lsn) >= 0);
}
/* Scan the metadata to find the live files and their IDs. */
WT_ERR(__recovery_file_scan(&r));
/*
* We no longer need the metadata cursor: close it to avoid pinning any
* resources that could block eviction during recovery.
*/
r.files[0].c = NULL;
WT_ERR(metac->close(metac));
/*
* Now, recover all the files apart from the metadata.
* Pass WT_LOGSCAN_RECOVER so that old logs get truncated.
*/
r.metadata_only = 0;
WT_ERR(__wt_verbose(session, WT_VERB_RECOVERY,
"Main recovery loop: starting at %u/%" PRIuMAX,
r.ckpt_lsn.file, (uintmax_t)r.ckpt_lsn.offset));
if (IS_INIT_LSN(&r.ckpt_lsn))
WT_ERR(__wt_log_scan(session, NULL,
WT_LOGSCAN_FIRST | WT_LOGSCAN_RECOVER,
__txn_log_recover, &r));
else
WT_ERR(__wt_log_scan(session, &r.ckpt_lsn,
WT_LOGSCAN_RECOVER,
__txn_log_recover, &r));
conn->next_file_id = r.max_fileid;
/*
* If recovery ran successfully forcibly log a checkpoint so the next
* open is fast and keep the metadata up to date with the checkpoint
* LSN and archiving.
*/
WT_ERR(session->iface.checkpoint(&session->iface, "force=1"));
err: WT_TRET(__recovery_free(&r));
__wt_free(session, config);
WT_TRET(session->iface.close(&session->iface, NULL));
return (ret);
}
