Add Module Parameter Regarding Log Size Limit

zfs_wrlog_data_max
The upper limit of TX_WRITE log data. Once it is reached,
write operation is blocked, until log data is cleared out
after txg sync. It only counts TX_WRITE log with WR_COPIED
or WR_NEED_COPY.

Reviewed-by: Prakash Surya <prakash.surya@delphix.com>
Reviewed-by: Brian Behlendorf <behlendorf1@llnl.gov>
Signed-off-by: jxdking <lostking2008@hotmail.com>
Closes #12284

include/sys/dmu_tx.h

@@ -124,6 +124,7 @@ typedef struct dmu_tx_stats {
 	kstat_named_t dmu_tx_dirty_throttle;
 	kstat_named_t dmu_tx_dirty_delay;
 	kstat_named_t dmu_tx_dirty_over_max;
+	kstat_named_t dmu_tx_wrlog_over_max;
 	kstat_named_t dmu_tx_dirty_frees_delay;
 	kstat_named_t dmu_tx_quota;
 } dmu_tx_stats_t;

include/sys/dsl_pool.h

@@ -40,6 +40,7 @@
 #include <sys/rrwlock.h>
 #include <sys/dsl_synctask.h>
 #include <sys/mmp.h>
+#include <sys/aggsum.h>
 
 #ifdef	__cplusplus
 extern "C" {
@@ -58,6 +59,7 @@ struct dsl_deadlist;
 
 extern unsigned long zfs_dirty_data_max;
 extern unsigned long zfs_dirty_data_max_max;
+extern unsigned long zfs_wrlog_data_max;
 extern int zfs_dirty_data_sync_percent;
 extern int zfs_dirty_data_max_percent;
 extern int zfs_dirty_data_max_max_percent;
@@ -118,6 +120,9 @@ typedef struct dsl_pool {
 	uint64_t dp_mos_compressed_delta;
 	uint64_t dp_mos_uncompressed_delta;
 
+	aggsum_t dp_wrlog_pertxg[TXG_SIZE];
+	aggsum_t dp_wrlog_total;
+
 	/*
 	 * Time of most recently scheduled (furthest in the future)
 	 * wakeup for delayed transactions.
@@ -158,6 +163,8 @@ uint64_t dsl_pool_adjustedsize(dsl_pool_t *dp, zfs_space_check_t slop_policy);
 uint64_t dsl_pool_unreserved_space(dsl_pool_t *dp,
     zfs_space_check_t slop_policy);
 uint64_t dsl_pool_deferred_space(dsl_pool_t *dp);
+void dsl_pool_wrlog_count(dsl_pool_t *dp, int64_t size, uint64_t txg);
+boolean_t dsl_pool_wrlog_over_max(dsl_pool_t *dp);
 void dsl_pool_dirty_space(dsl_pool_t *dp, int64_t space, dmu_tx_t *tx);
 void dsl_pool_undirty_space(dsl_pool_t *dp, int64_t space, uint64_t txg);
 void dsl_free(dsl_pool_t *dp, uint64_t txg, const blkptr_t *bpp);

man/man4/zfs.4

@@ -1096,6 +1096,18 @@ Start syncing out a transaction group if there's at least this much dirty data
 This should be less than
 .Sy zfs_vdev_async_write_active_min_dirty_percent .
 .
+.It Sy zfs_wrlog_data_max Ns = Pq int
+The upper limit of write-transaction zil log data size in bytes.
+Once it is reached, write operation is blocked, until log data is cleared out
+after transaction group sync. Because of some overhead, it should be set
+at least 2 times the size of
+.Sy zfs_dirty_data_max
+.No to prevent harming normal write throughput.
+It also should be smaller than the size of the slog device if slog is present.
+.Pp
+Defaults to
+.Sy zfs_dirty_data_max*2
+.
 .It Sy zfs_fallocate_reserve_percent Ns = Ns Sy 110 Ns % Pq uint
 Since ZFS is a copy-on-write filesystem with snapshots, blocks cannot be
 preallocated for a file in order to guarantee that later writes will not

module/zfs/arc.c

@@ -8062,6 +8062,18 @@ arc_init(void)
 		zfs_dirty_data_max = MIN(zfs_dirty_data_max,
 		    zfs_dirty_data_max_max);
 	}
+
+	if (zfs_wrlog_data_max == 0) {
+
+		/*
+		 * dp_wrlog_total is reduced for each txg at the end of
+		 * spa_sync(). However, dp_dirty_total is reduced every time
+		 * a block is written out. Thus under normal operation,
+		 * dp_wrlog_total could grow 2 times as big as
+		 * zfs_dirty_data_max.
+		 */
+		zfs_wrlog_data_max = zfs_dirty_data_max * 2;
+	}
 }
 
 void

module/zfs/dmu_tx.c

@@ -53,6 +53,7 @@ dmu_tx_stats_t dmu_tx_stats = {
 	{ "dmu_tx_dirty_throttle",	KSTAT_DATA_UINT64 },
 	{ "dmu_tx_dirty_delay",		KSTAT_DATA_UINT64 },
 	{ "dmu_tx_dirty_over_max",	KSTAT_DATA_UINT64 },
+	{ "dmu_tx_wrlog_over_max",	KSTAT_DATA_UINT64 },
 	{ "dmu_tx_dirty_frees_delay",	KSTAT_DATA_UINT64 },
 	{ "dmu_tx_quota",		KSTAT_DATA_UINT64 },
 };
@@ -884,6 +885,12 @@ dmu_tx_try_assign(dmu_tx_t *tx, uint64_t txg_how)
 		return (SET_ERROR(ERESTART));
 	}
 
+	if (!tx->tx_dirty_delayed &&
+	    dsl_pool_wrlog_over_max(tx->tx_pool)) {
+		DMU_TX_STAT_BUMP(dmu_tx_wrlog_over_max);
+		return (SET_ERROR(ERESTART));
+	}
+
 	if (!tx->tx_dirty_delayed &&
 	    dsl_pool_need_dirty_delay(tx->tx_pool)) {
 		tx->tx_wait_dirty = B_TRUE;

module/zfs/dsl_pool.c

@@ -104,6 +104,14 @@ unsigned long zfs_dirty_data_max_max = 0;
 int zfs_dirty_data_max_percent = 10;
 int zfs_dirty_data_max_max_percent = 25;
 
+/*
+ * zfs_wrlog_data_max, the upper limit of TX_WRITE log data.
+ * Once it is reached, write operation is blocked,
+ * until log data is cleared out after txg sync.
+ * It only counts TX_WRITE log with WR_COPIED or WR_NEED_COPY.
+ */
+unsigned long zfs_wrlog_data_max = 0;
+
 /*
  * If there's at least this much dirty data (as a percentage of
  * zfs_dirty_data_max), push out a txg.  This should be less than
@@ -220,6 +228,11 @@ dsl_pool_open_impl(spa_t *spa, uint64_t txg)
 	mutex_init(&dp->dp_lock, NULL, MUTEX_DEFAULT, NULL);
 	cv_init(&dp->dp_spaceavail_cv, NULL, CV_DEFAULT, NULL);
 
+	aggsum_init(&dp->dp_wrlog_total, 0);
+	for (int i = 0; i < TXG_SIZE; i++) {
+		aggsum_init(&dp->dp_wrlog_pertxg[i], 0);
+	}
+
 	dp->dp_zrele_taskq = taskq_create("z_zrele", 100, defclsyspri,
 	    boot_ncpus * 8, INT_MAX, TASKQ_PREPOPULATE | TASKQ_DYNAMIC |
 	    TASKQ_THREADS_CPU_PCT);
@@ -416,6 +429,14 @@ dsl_pool_close(dsl_pool_t *dp)
 	rrw_destroy(&dp->dp_config_rwlock);
 	mutex_destroy(&dp->dp_lock);
 	cv_destroy(&dp->dp_spaceavail_cv);
+
+	ASSERT0(aggsum_value(&dp->dp_wrlog_total));
+	aggsum_fini(&dp->dp_wrlog_total);
+	for (int i = 0; i < TXG_SIZE; i++) {
+		ASSERT0(aggsum_value(&dp->dp_wrlog_pertxg[i]));
+		aggsum_fini(&dp->dp_wrlog_pertxg[i]);
+	}
+
 	taskq_destroy(dp->dp_unlinked_drain_taskq);
 	taskq_destroy(dp->dp_zrele_taskq);
 	if (dp->dp_blkstats != NULL)
@@ -590,6 +611,36 @@ dsl_pool_dirty_delta(dsl_pool_t *dp, int64_t delta)
 		cv_signal(&dp->dp_spaceavail_cv);
 }
 
+void
+dsl_pool_wrlog_count(dsl_pool_t *dp, int64_t size, uint64_t txg)
+{
+	ASSERT3S(size, >=, 0);
+
+	aggsum_add(&dp->dp_wrlog_pertxg[txg & TXG_MASK], size);
+	aggsum_add(&dp->dp_wrlog_total, size);
+
+	/* Choose a value slightly bigger than min dirty sync bytes */
+	uint64_t sync_min =
+	    zfs_dirty_data_max * (zfs_dirty_data_sync_percent + 10) / 100;
+	if (aggsum_compare(&dp->dp_wrlog_pertxg[txg & TXG_MASK], sync_min) > 0)
+		txg_kick(dp, txg);
+}
+
+boolean_t
+dsl_pool_wrlog_over_max(dsl_pool_t *dp)
+{
+	return (aggsum_compare(&dp->dp_wrlog_total, zfs_wrlog_data_max) > 0);
+}
+
+static void
+dsl_pool_wrlog_clear(dsl_pool_t *dp, uint64_t txg)
+{
+	int64_t delta;
+	delta = -(int64_t)aggsum_value(&dp->dp_wrlog_pertxg[txg & TXG_MASK]);
+	aggsum_add(&dp->dp_wrlog_pertxg[txg & TXG_MASK], delta);
+	aggsum_add(&dp->dp_wrlog_total, delta);
+}
+
 #ifdef ZFS_DEBUG
 static boolean_t
 dsl_early_sync_task_verify(dsl_pool_t *dp, uint64_t txg)
@@ -814,6 +865,9 @@ dsl_pool_sync_done(dsl_pool_t *dp, uint64_t txg)
 		ASSERT(!dmu_objset_is_dirty(zilog->zl_os, txg));
 		dmu_buf_rele(ds->ds_dbuf, zilog);
 	}
+
+	dsl_pool_wrlog_clear(dp, txg);
+
 	ASSERT(!dmu_objset_is_dirty(dp->dp_meta_objset, txg));
 }
 
@@ -1409,6 +1463,9 @@ ZFS_MODULE_PARAM(zfs, zfs_, delay_min_dirty_percent, INT, ZMOD_RW,
 ZFS_MODULE_PARAM(zfs, zfs_, dirty_data_max, ULONG, ZMOD_RW,
 	"Determines the dirty space limit");
 
+ZFS_MODULE_PARAM(zfs, zfs_, wrlog_data_max, ULONG, ZMOD_RW,
+	"The size limit of write-transaction zil log data");
+
 /* zfs_dirty_data_max_max only applied at module load in arc_init(). */
 ZFS_MODULE_PARAM(zfs, zfs_, dirty_data_max_max, ULONG, ZMOD_RD,
 	"zfs_dirty_data_max upper bound in bytes");

module/zfs/zfs_log.c

@@ -538,6 +538,7 @@ zfs_log_write(zilog_t *zilog, dmu_tx_t *tx, int txtype,
 	itx_wr_state_t write_state;
 	uintptr_t fsync_cnt;
 	uint64_t gen = 0;
+	ssize_t size = resid;
 
 	if (zil_replaying(zilog, tx) || zp->z_unlinked ||
 	    zfs_xattr_owner_unlinked(zp)) {
@@ -623,6 +624,10 @@ zfs_log_write(zilog_t *zilog, dmu_tx_t *tx, int txtype,
 		off += len;
 		resid -= len;
 	}
+
+	if (write_state == WR_COPIED || write_state == WR_NEED_COPY) {
+		dsl_pool_wrlog_count(zilog->zl_dmu_pool, size, tx->tx_txg);
+	}
 }
 
 /*

module/zfs/zvol.c

@@ -84,10 +84,8 @@
 #include <sys/zfs_rlock.h>
 #include <sys/spa_impl.h>
 #include <sys/zvol.h>
-
 #include <sys/zvol_impl.h>
 
-
 unsigned int zvol_inhibit_dev = 0;
 unsigned int zvol_volmode = ZFS_VOLMODE_GEOM;
 
@@ -577,6 +575,7 @@ zvol_log_write(zvol_state_t *zv, dmu_tx_t *tx, uint64_t offset,
 	uint32_t blocksize = zv->zv_volblocksize;
 	zilog_t *zilog = zv->zv_zilog;
 	itx_wr_state_t write_state;
+	uint64_t sz = size;
 
 	if (zil_replaying(zilog, tx))
 		return;
@@ -628,6 +627,10 @@ zvol_log_write(zvol_state_t *zv, dmu_tx_t *tx, uint64_t offset,
 		offset += len;
 		size -= len;
 	}
+
+	if (write_state == WR_COPIED || write_state == WR_NEED_COPY) {
+		dsl_pool_wrlog_count(zilog->zl_dmu_pool, sz, tx->tx_txg);
+	}
 }
 
 /*