svn commit: r324205 - in stable/11/sys/cddl/contrib/opensolaris/uts/common/fs/zfs: . sys
Andriy Gapon
avg at FreeBSD.org
Mon Oct 2 12:54:03 UTC 2017
Author: avg
Date: Mon Oct 2 12:54:01 2017
New Revision: 324205
URL: https://svnweb.freebsd.org/changeset/base/324205
Log:
MFC r323433,r323793,r323915: MFV r323110: 8558 lwp_create() returns EAGAIN
on system with more than 80K ZFS filesystems, and followups
r323433: MFV r323110: 8558 lwp_create() returns EAGAIN on system with more than 80K ZFS filesystems
r323793: MFV r323792: 8602 remove unused "dp_early_sync_tasks" field from "dsl_pool" structure
r323915: MFV r323914: 8661 remove "zil-cw2" dtrace probe
Modified:
stable/11/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dsl_pool.c
stable/11/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/sys/dsl_pool.h
stable/11/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/sys/zil_impl.h
stable/11/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zil.c
Directory Properties:
stable/11/ (props changed)
Modified: stable/11/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dsl_pool.c
==============================================================================
--- stable/11/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dsl_pool.c Mon Oct 2 12:47:35 2017 (r324204)
+++ stable/11/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/dsl_pool.c Mon Oct 2 12:54:01 2017 (r324205)
@@ -137,6 +137,36 @@ uint64_t zfs_delay_scale = 1000 * 1000 * 1000 / 2000;
*/
int zfs_sync_taskq_batch_pct = 75;
+/*
+ * These tunables determine the behavior of how zil_itxg_clean() is
+ * called via zil_clean() in the context of spa_sync(). When an itxg
+ * list needs to be cleaned, TQ_NOSLEEP will be used when dispatching.
+ * If the dispatch fails, the call to zil_itxg_clean() will occur
+ * synchronously in the context of spa_sync(), which can negatively
+ * impact the performance of spa_sync() (e.g. in the case of the itxg
+ * list having a large number of itxs that needs to be cleaned).
+ *
+ * Thus, these tunables can be used to manipulate the behavior of the
+ * taskq used by zil_clean(); they determine the number of taskq entries
+ * that are pre-populated when the taskq is first created (via the
+ * "zfs_zil_clean_taskq_minalloc" tunable) and the maximum number of
+ * taskq entries that are cached after an on-demand allocation (via the
+ * "zfs_zil_clean_taskq_maxalloc").
+ *
+ * The idea being, we want to try reasonably hard to ensure there will
+ * already be a taskq entry pre-allocated by the time that it is needed
+ * by zil_clean(). This way, we can avoid the possibility of an
+ * on-demand allocation of a new taskq entry from failing, which would
+ * result in zil_itxg_clean() being called synchronously from zil_clean()
+ * (which can adversely affect performance of spa_sync()).
+ *
+ * Additionally, the number of threads used by the taskq can be
+ * configured via the "zfs_zil_clean_taskq_nthr_pct" tunable.
+ */
+int zfs_zil_clean_taskq_nthr_pct = 100;
+int zfs_zil_clean_taskq_minalloc = 1024;
+int zfs_zil_clean_taskq_maxalloc = 1024 * 1024;
+
#if defined(__FreeBSD__) && defined(_KERNEL)
extern int zfs_vdev_async_write_active_max_dirty_percent;
@@ -272,6 +302,12 @@ dsl_pool_open_impl(spa_t *spa, uint64_t txg)
zfs_sync_taskq_batch_pct, minclsyspri, 1, INT_MAX,
TASKQ_THREADS_CPU_PCT);
+ dp->dp_zil_clean_taskq = taskq_create("dp_zil_clean_taskq",
+ zfs_zil_clean_taskq_nthr_pct, minclsyspri,
+ zfs_zil_clean_taskq_minalloc,
+ zfs_zil_clean_taskq_maxalloc,
+ TASKQ_PREPOPULATE | TASKQ_THREADS_CPU_PCT);
+
mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
@@ -422,6 +458,7 @@ dsl_pool_close(dsl_pool_t *dp)
txg_list_destroy(&dp->dp_sync_tasks);
txg_list_destroy(&dp->dp_dirty_dirs);
+ taskq_destroy(dp->dp_zil_clean_taskq);
taskq_destroy(dp->dp_sync_taskq);
/*
Modified: stable/11/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/sys/dsl_pool.h
==============================================================================
--- stable/11/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/sys/dsl_pool.h Mon Oct 2 12:47:35 2017 (r324204)
+++ stable/11/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/sys/dsl_pool.h Mon Oct 2 12:54:01 2017 (r324205)
@@ -122,6 +122,7 @@ typedef struct dsl_pool {
txg_list_t dp_dirty_dirs;
txg_list_t dp_sync_tasks;
taskq_t *dp_sync_taskq;
+ taskq_t *dp_zil_clean_taskq;
/*
* Protects administrative changes (properties, namespace)
Modified: stable/11/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/sys/zil_impl.h
==============================================================================
--- stable/11/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/sys/zil_impl.h Mon Oct 2 12:47:35 2017 (r324204)
+++ stable/11/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/sys/zil_impl.h Mon Oct 2 12:54:01 2017 (r324205)
@@ -124,7 +124,6 @@ struct zilog {
list_t zl_lwb_list; /* in-flight log write list */
kmutex_t zl_vdev_lock; /* protects zl_vdev_tree */
avl_tree_t zl_vdev_tree; /* vdevs to flush in zil_commit() */
- taskq_t *zl_clean_taskq; /* runs lwb and itx clean tasks */
avl_tree_t zl_bp_tree; /* track bps during log parse */
clock_t zl_replay_time; /* lbolt of when replay started */
uint64_t zl_replay_blks; /* number of log blocks replayed */
Modified: stable/11/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zil.c
==============================================================================
--- stable/11/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zil.c Mon Oct 2 12:47:35 2017 (r324204)
+++ stable/11/sys/cddl/contrib/opensolaris/uts/common/fs/zfs/zil.c Mon Oct 2 12:54:01 2017 (r324205)
@@ -1390,8 +1390,7 @@ zil_clean(zilog_t *zilog, uint64_t synced_txg)
return;
}
ASSERT3U(itxg->itxg_txg, <=, synced_txg);
- ASSERT(itxg->itxg_txg != 0);
- ASSERT(zilog->zl_clean_taskq != NULL);
+ ASSERT3U(itxg->itxg_txg, !=, 0);
clean_me = itxg->itxg_itxs;
itxg->itxg_itxs = NULL;
itxg->itxg_txg = 0;
@@ -1402,7 +1401,9 @@ zil_clean(zilog_t *zilog, uint64_t synced_txg)
* free it in-line. This should be rare. Note, using TQ_SLEEP
* created a bad performance problem.
*/
- if (taskq_dispatch(zilog->zl_clean_taskq,
+ ASSERT3P(zilog->zl_dmu_pool, !=, NULL);
+ ASSERT3P(zilog->zl_dmu_pool->dp_zil_clean_taskq, !=, NULL);
+ if (taskq_dispatch(zilog->zl_dmu_pool->dp_zil_clean_taskq,
(void (*)(void *))zil_itxg_clean, clean_me, TQ_NOSLEEP) == 0)
zil_itxg_clean(clean_me);
}
@@ -1556,7 +1557,6 @@ zil_commit_writer(zilog_t *zilog)
kmem_free(itx, offsetof(itx_t, itx_lr)
+ itx->itx_lr.lrc_reclen);
}
- DTRACE_PROBE1(zil__cw2, zilog_t *, zilog);
/* write the last block out */
if (lwb != NULL && lwb->lwb_zio != NULL)
@@ -1831,13 +1831,10 @@ zil_open(objset_t *os, zil_get_data_t *get_data)
{
zilog_t *zilog = dmu_objset_zil(os);
- ASSERT(zilog->zl_clean_taskq == NULL);
ASSERT(zilog->zl_get_data == NULL);
ASSERT(list_is_empty(&zilog->zl_lwb_list));
zilog->zl_get_data = get_data;
- zilog->zl_clean_taskq = taskq_create("zil_clean", 1, minclsyspri,
- 2, 2, TASKQ_PREPOPULATE);
return (zilog);
}
@@ -1871,8 +1868,6 @@ zil_close(zilog_t *zilog)
zfs_dbgmsg("zil (%p) is dirty, txg %llu", zilog, txg);
VERIFY(!zilog_is_dirty(zilog));
- taskq_destroy(zilog->zl_clean_taskq);
- zilog->zl_clean_taskq = NULL;
zilog->zl_get_data = NULL;
/*
More information about the svn-src-stable
mailing list