--- a/usr/src/uts/common/fs/zfs/zil.c Tue Nov 25 09:39:25 2008 -0800
+++ b/usr/src/uts/common/fs/zfs/zil.c Tue Nov 25 13:18:25 2008 -0700
@@ -1220,7 +1220,7 @@
ASSERT(zilog->zl_stop_sync == 0);
- zh->zh_replay_seq = zilog->zl_replay_seq[txg & TXG_MASK];
+ zh->zh_replay_seq = zilog->zl_replayed_seq[txg & TXG_MASK];
if (zilog->zl_destroy_txg == txg) {
blkptr_t blk = zh->zh_log;
@@ -1229,7 +1229,7 @@
ASSERT(spa_sync_pass(spa) == 1);
bzero(zh, sizeof (zil_header_t));
- bzero(zilog->zl_replay_seq, sizeof (zilog->zl_replay_seq));
+ bzero(zilog->zl_replayed_seq, sizeof (zilog->zl_replayed_seq));
if (zilog->zl_keep_first) {
/*
@@ -1466,9 +1466,7 @@
typedef struct zil_replay_arg {
objset_t *zr_os;
zil_replay_func_t **zr_replay;
- zil_replay_cleaner_t *zr_replay_cleaner;
void *zr_arg;
- uint64_t *zr_txgp;
boolean_t zr_byteswap;
char *zr_lrbuf;
} zil_replay_arg_t;
@@ -1481,9 +1479,9 @@
uint64_t reclen = lr->lrc_reclen;
uint64_t txtype = lr->lrc_txtype;
char *name;
- int pass, error, sunk;
+ int pass, error;
- if (zilog->zl_stop_replay)
+ if (!zilog->zl_replay) /* giving up */
return;
if (lr->lrc_txg < claim_txg) /* already committed */
@@ -1495,6 +1493,11 @@
/* Strip case-insensitive bit, still present in log record */
txtype &= ~TX_CI;
+ if (txtype == 0 || txtype >= TX_MAX_TYPE) {
+ error = EINVAL;
+ goto bad;
+ }
+
/*
* Make a copy of the data so we can revise and extend it.
*/
@@ -1545,69 +1548,16 @@
}
/*
- * Replay of large truncates can end up needing additional txs
- * and a different txg. If they are nested within the replay tx
- * as below then a hang is possible. So we do the truncate here
- * and redo the truncate later (a no-op) and update the sequence
- * number whilst in the replay tx. Fortunately, it's safe to repeat
- * a truncate if we crash and the truncate commits. A create over
- * an existing file will also come in as a TX_TRUNCATE record.
- *
- * Note, remove of large files and renames over large files is
- * handled by putting the deleted object on a stable list
- * and if necessary force deleting the object outside of the replay
- * transaction using the zr_replay_cleaner.
- */
- if (txtype == TX_TRUNCATE) {
- *zr->zr_txgp = TXG_NOWAIT;
- error = zr->zr_replay[TX_TRUNCATE](zr->zr_arg, zr->zr_lrbuf,
- zr->zr_byteswap);
- if (error)
- goto bad;
- zr->zr_byteswap = 0; /* only byteswap once */
- }
-
- /*
* We must now do two things atomically: replay this log record,
- * and update the log header to reflect the fact that we did so.
- * We use the DMU's ability to assign into a specific txg to do this.
+ * and update the log header sequence number to reflect the fact that
+ * we did so. At the end of each replay function the sequence number
+ * is updated if we are in replay mode.
*/
- for (pass = 1, sunk = B_FALSE; /* CONSTANTCONDITION */; pass++) {
- uint64_t replay_txg;
- dmu_tx_t *replay_tx;
-
- replay_tx = dmu_tx_create(zr->zr_os);
- error = dmu_tx_assign(replay_tx, TXG_WAIT);
- if (error) {
- dmu_tx_abort(replay_tx);
- break;
- }
-
- replay_txg = dmu_tx_get_txg(replay_tx);
-
- if (txtype == 0 || txtype >= TX_MAX_TYPE) {
- error = EINVAL;
- } else {
- /*
- * On the first pass, arrange for the replay vector
- * to fail its dmu_tx_assign(). That's the only way
- * to ensure that those code paths remain well tested.
- *
- * Only byteswap (if needed) on the 1st pass.
- */
- *zr->zr_txgp = replay_txg - (pass == 1);
- error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lrbuf,
- zr->zr_byteswap && pass == 1);
- *zr->zr_txgp = TXG_NOWAIT;
- }
-
- if (error == 0) {
- dsl_dataset_dirty(dmu_objset_ds(zr->zr_os), replay_tx);
- zilog->zl_replay_seq[replay_txg & TXG_MASK] =
- lr->lrc_seq;
- }
-
- dmu_tx_commit(replay_tx);
+ for (pass = 1; pass <= 2; pass++) {
+ zilog->zl_replaying_seq = lr->lrc_seq;
+ /* Only byteswap (if needed) on the 1st pass. */
+ error = zr->zr_replay[txtype](zr->zr_arg, zr->zr_lrbuf,
+ zr->zr_byteswap && pass == 1);
if (!error)
return;
@@ -1615,37 +1565,22 @@
/*
* The DMU's dnode layer doesn't see removes until the txg
* commits, so a subsequent claim can spuriously fail with
- * EEXIST. So if we receive any error other than ERESTART
- * we try syncing out any removes then retrying the
- * transaction.
+ * EEXIST. So if we receive any error we try syncing out
+ * any removes then retry the transaction.
*/
- if (error != ERESTART && !sunk) {
- if (zr->zr_replay_cleaner)
- zr->zr_replay_cleaner(zr->zr_arg);
+ if (pass == 1)
txg_wait_synced(spa_get_dsl(zilog->zl_spa), 0);
- sunk = B_TRUE;
- continue; /* retry */
- }
-
- if (error != ERESTART)
- break;
-
- if (pass != 1)
- txg_wait_open(spa_get_dsl(zilog->zl_spa),
- replay_txg + 1);
-
- dprintf("pass %d, retrying\n", pass);
}
bad:
- ASSERT(error && error != ERESTART);
+ ASSERT(error);
name = kmem_alloc(MAXNAMELEN, KM_SLEEP);
dmu_objset_name(zr->zr_os, name);
cmn_err(CE_WARN, "ZFS replay transaction error %d, "
"dataset %s, seq 0x%llx, txtype %llu %s\n",
error, name, (u_longlong_t)lr->lrc_seq, (u_longlong_t)txtype,
(lr->lrc_txtype & TX_CI) ? "CI" : "");
- zilog->zl_stop_replay = 1;
+ zilog->zl_replay = B_FALSE;
kmem_free(name, MAXNAMELEN);
}
@@ -1660,9 +1595,7 @@
* If this dataset has a non-empty intent log, replay it and destroy it.
*/
void
-zil_replay(objset_t *os, void *arg, uint64_t *txgp,
- zil_replay_func_t *replay_func[TX_MAX_TYPE],
- zil_replay_cleaner_t *replay_cleaner)
+zil_replay(objset_t *os, void *arg, zil_replay_func_t *replay_func[TX_MAX_TYPE])
{
zilog_t *zilog = dmu_objset_zil(os);
const zil_header_t *zh = zilog->zl_header;
@@ -1675,9 +1608,7 @@
zr.zr_os = os;
zr.zr_replay = replay_func;
- zr.zr_replay_cleaner = replay_cleaner;
zr.zr_arg = arg;
- zr.zr_txgp = txgp;
zr.zr_byteswap = BP_SHOULD_BYTESWAP(&zh->zh_log);
zr.zr_lrbuf = kmem_alloc(2 * SPA_MAXBLOCKSIZE, KM_SLEEP);
@@ -1686,7 +1617,7 @@
*/
txg_wait_synced(zilog->zl_dmu_pool, 0);
- zilog->zl_stop_replay = 0;
+ zilog->zl_replay = B_TRUE;
zilog->zl_replay_time = lbolt;
ASSERT(zilog->zl_replay_blks == 0);
(void) zil_parse(zilog, zil_incr_blks, zil_replay_log_record, &zr,
@@ -1695,6 +1626,7 @@
zil_destroy(zilog, B_FALSE);
txg_wait_synced(zilog->zl_dmu_pool, zilog->zl_destroy_txg);
+ zilog->zl_replay = B_FALSE;
}
/*