Detect a slow raidz child during reads

If it's an persistent outlier then sit out reads for a period.

Signed-off-by: Don Brady <[email protected]>

cmd/ztest.c

@@ -7374,6 +7374,13 @@ ztest_run_zdb(uint64_t guid)
 	char *set_gvars_args_joined = join_strings(set_gvars_args, " ");
 	free(set_gvars_args);
 
+	printf("%s -bcc%s%s -G -d -Y -e -y %s -p %s %"PRIu64,
+	    bin,
+	    ztest_opts.zo_verbose >= 3 ? "s" : "",
+	    ztest_opts.zo_verbose >= 4 ? "v" : "",
+	    set_gvars_args_joined,
+	    ztest_opts.zo_dir,
+	    guid);
 	size_t would = snprintf(zdb, len,
 	    "%s -bcc%s%s -G -d -Y -e -y %s -p %s %"PRIu64,
 	    bin,
@@ -7386,7 +7393,7 @@ ztest_run_zdb(uint64_t guid)
 
 	umem_free(set_gvars_args_joined, strlen(set_gvars_args_joined) + 1);
 
-	if (ztest_opts.zo_verbose >= 5)
+//	if (ztest_opts.zo_verbose >= 5)
 		(void) printf("Executing %s\n", zdb);
 
 	fp = popen(zdb, "r");

include/sys/vdev_impl.h

@@ -432,6 +432,8 @@ struct vdev {
 	hrtime_t	vdev_mmp_pending; /* 0 if write finished	*/
 	uint64_t	vdev_mmp_kstat_id;	/* to find kstat entry */
 	uint64_t	vdev_expansion_time;	/* vdev's last expansion time */
+	uint64_t	vdev_outlier_count;	/* peer outlier events */
+	hrtime_t	vdev_read_sit_out_expire;
 	list_node_t	vdev_leaf_node;		/* leaf vdev list */
 
 	/*

include/sys/vdev_raidz.h

@@ -60,6 +60,7 @@ void vdev_raidz_checksum_error(zio_t *, struct raidz_col *, abd_t *);
 struct raidz_row *vdev_raidz_row_alloc(int, zio_t *);
 void vdev_raidz_reflow_copy_scratch(spa_t *);
 void raidz_dtl_reassessed(vdev_t *);
+boolean_t vdev_skip_latency_outlier(vdev_t *, zio_flag_t);
 
 extern const zio_vsd_ops_t vdev_raidz_vsd_ops;
 

include/sys/vdev_raidz_impl.h

@@ -104,6 +104,8 @@ typedef struct raidz_impl_ops {
 	char name[RAIDZ_IMPL_NAME_MAX];	/* Name of the implementation */
 } raidz_impl_ops_t;
 
+typedef struct raidz_row raidz_row_t;
+typedef struct raidz_col raidz_col_t;
 
 typedef struct raidz_col {
 	int rc_devidx;			/* child device index for I/O */
@@ -121,6 +123,9 @@ typedef struct raidz_col {
 	int rc_shadow_devidx;		/* for double write during expansion */
 	int rc_shadow_error;		/* for double write during expansion */
 	uint64_t rc_shadow_offset;	/* for double write during expansion */
+	uint64_t rc_latency_val;	/* leaf device scaled latency value */
+	raidz_col_t *rc_latency_next;	/* sorted next column in list */
+	raidz_row_t *rc_raidz_row;
 } raidz_col_t;
 
 typedef struct raidz_row {
@@ -132,6 +137,7 @@ typedef struct raidz_row {
 	int rr_firstdatacol;		/* First data column/parity count */
 	abd_t *rr_abd_empty;		/* dRAID empty sector buffer */
 	int rr_nempty;			/* empty sectors included in parity */
+	raidz_col_t *rr_latency_list_start; /* head of child latency columns */
 #ifdef ZFS_DEBUG
 	uint64_t rr_offset;		/* Logical offset for *_io_verify() */
 	uint64_t rr_size;		/* Physical size for *_io_verify() */

module/zfs/vdev_draid.c

@@ -1048,6 +1048,7 @@ vdev_draid_map_alloc_row(zio_t *zio, raidz_row_t **rrp, uint64_t io_offset,
 
 		rc->rc_devidx = vdev_draid_permute_id(vdc, base, iter, c);
 		rc->rc_offset = physical_offset;
+		rc->rc_raidz_row = rr;
 
 		if (q == 0 && i >= bc)
 			rc->rc_size = 0;
@@ -1889,6 +1890,17 @@ vdev_draid_io_start_read(zio_t *zio, raidz_row_t *rr)
 	/* Sequential rebuild must do IO at redundancy group boundary. */
 	IMPLY(zio->io_priority == ZIO_PRIORITY_REBUILD, rr->rr_nempty == 0);
 
+	/*
+	 * Calculate how much parity is available for sitting out reads
+	 */
+	int parity_avail = rr->rr_firstdatacol;
+	for (int p = 0; p < rr->rr_firstdatacol; p++) {
+		raidz_col_t *rc = &rr->rr_col[p];
+		if (!vdev_draid_readable(vd->vdev_child[rc->rc_devidx],
+		    rc->rc_offset)) {
+			parity_avail--;
+		}
+	}
 	/*
 	 * Iterate over the columns in reverse order so that we hit the parity
 	 * last.  Any errors along the way will force us to read the parity.
@@ -1993,6 +2005,13 @@ vdev_draid_io_start_read(zio_t *zio, raidz_row_t *rr)
 				rc->rc_force_repair = 1;
 				rc->rc_allow_repair = 1;
 			}
+		} else if (parity_avail > 0 && c >= rr->rr_firstdatacol &&
+		    rr->rr_missingdata == 0 &&
+		    vdev_skip_latency_outlier(cvd, zio->io_flags)) {
+			rr->rr_missingdata++;
+			rc->rc_error = SET_ERROR(EAGAIN);
+			rc->rc_skipped = 1;
+			continue;
 		}
 	}