diff options
| -rw-r--r-- | man/man4/zfs.4 | 7 | ||||
| -rw-r--r-- | module/zfs/zil.c | 36 |
2 files changed, 41 insertions, 2 deletions
diff --git a/man/man4/zfs.4 b/man/man4/zfs.4 index a3664367a..e20d60134 100644 --- a/man/man4/zfs.4 +++ b/man/man4/zfs.4 @@ -2209,6 +2209,13 @@ On very fragmented pools, lowering this .Pq typically to Sy 36 KiB can improve performance. . +.It Sy zil_min_commit_timeout Ns = Ns Sy 5000 Pq u64 +This sets the minimum delay in nanoseconds ZIL care to delay block commit, +waiting for more records. +If ZIL writes are too fast, kernel may not be able sleep for so short interval, +increasing log latency above allowed by +.Sy zfs_commit_timeout_pct . +. .It Sy zil_nocacheflush Ns = Ns Sy 0 Ns | Ns 1 Pq int Disable the cache flush commands that are normally sent to disk by the ZIL after an LWB write has completed. diff --git a/module/zfs/zil.c b/module/zfs/zil.c index 02e6f4b83..2e017992f 100644 --- a/module/zfs/zil.c +++ b/module/zfs/zil.c @@ -93,6 +93,14 @@ static uint_t zfs_commit_timeout_pct = 5; /* + * Minimal time we care to delay commit waiting for more ZIL records. + * At least FreeBSD kernel can't sleep for less than 2us at its best. + * So requests to sleep for less then 5us is a waste of CPU time with + * a risk of significant log latency increase due to oversleep. + */ +static uint64_t zil_min_commit_timeout = 5000; + +/* * See zil.h for more information about these fields. */ static zil_kstat_values_t zil_stats = { @@ -1295,7 +1303,8 @@ zil_lwb_flush_vdevs_done(zio_t *zio) lwb->lwb_buf = NULL; ASSERT3U(lwb->lwb_issued_timestamp, >, 0); - zilog->zl_last_lwb_latency = gethrtime() - lwb->lwb_issued_timestamp; + zilog->zl_last_lwb_latency = (zilog->zl_last_lwb_latency * 3 + + gethrtime() - lwb->lwb_issued_timestamp) / 4; lwb->lwb_root_zio = NULL; @@ -2463,8 +2472,9 @@ zil_process_commit_list(zilog_t *zilog) spa_t *spa = zilog->zl_spa; list_t nolwb_itxs; list_t nolwb_waiters; - lwb_t *lwb; + lwb_t *lwb, *plwb; itx_t *itx; + boolean_t first = B_TRUE; ASSERT(MUTEX_HELD(&zilog->zl_issuer_lock)); @@ -2491,6 +2501,9 @@ zil_process_commit_list(zilog_t *zilog) ASSERT3S(lwb->lwb_state, !=, LWB_STATE_ISSUED); ASSERT3S(lwb->lwb_state, !=, LWB_STATE_WRITE_DONE); ASSERT3S(lwb->lwb_state, !=, LWB_STATE_FLUSH_DONE); + first = (lwb->lwb_state != LWB_STATE_OPENED) && + ((plwb = list_prev(&zilog->zl_lwb_list, lwb)) == NULL || + plwb->lwb_state == LWB_STATE_FLUSH_DONE); } while ((itx = list_head(&zilog->zl_itx_commit_list)) != NULL) { @@ -2661,7 +2674,23 @@ zil_process_commit_list(zilog_t *zilog) * try and pack as many itxs into as few lwbs as * possible, without significantly impacting the latency * of each individual itx. + * + * If we had no already running or open LWBs, it can be + * the workload is single-threaded. And if the ZIL write + * latency is very small or if the LWB is almost full, it + * may be cheaper to bypass the delay. */ + if (lwb->lwb_state == LWB_STATE_OPENED && first) { + hrtime_t sleep = zilog->zl_last_lwb_latency * + zfs_commit_timeout_pct / 100; + if (sleep < zil_min_commit_timeout || + lwb->lwb_sz - lwb->lwb_nused < lwb->lwb_sz / 8) { + lwb = zil_lwb_write_issue(zilog, lwb); + zilog->zl_cur_used = 0; + if (lwb == NULL) + zil_commit_writer_stall(zilog); + } + } } } @@ -3949,6 +3978,9 @@ EXPORT_SYMBOL(zil_kstat_values_update); ZFS_MODULE_PARAM(zfs, zfs_, commit_timeout_pct, UINT, ZMOD_RW, "ZIL block open timeout percentage"); +ZFS_MODULE_PARAM(zfs_zil, zil_, min_commit_timeout, U64, ZMOD_RW, + "Minimum delay we care for ZIL block commit"); + ZFS_MODULE_PARAM(zfs_zil, zil_, replay_disable, INT, ZMOD_RW, "Disable intent logging replay"); |