OpenZFS 9486 - reduce memory used by device removal on fragmented pools

Device removal allocates a new location for each allocated segment on
the disk that's being removed.  Each allocation results in one entry in
the mapping table, which maps from old location + length to new
location.  When a fragmented disk is removed, this can result in a large
number of mapping entries, and thus a large amount of memory consumed by
the mapping table.  In the worst real-world cases, we've seen around 1GB
of RAM per 1TB of storage removed.

We can improve on this situation by allocating larger segments, which
span across both allocated and free regions of the device being removed.
By including free regions in the allocation (and thus mapping), we
reduce the number of mapping entries.  For example, if we have a 4K
allocation followed by 1K free and then 4K allocated, we would allocate
4+1+4 = 9KB, and then move the entire region (including allocated and
free parts).  In this case we used one mapping where previously we would
have used two, but often the ratio is much higher (up to 20:1 in
real-world use).  We then need to mark the regions that were free on the
removing device as free in the new locations, and also obsolete in the
mapping entry.

This method preserves the fragmentation of the removing device, rather
than consolidating its allocated space into a small number of chunks
where possible.  But it results in drastic reduction of memory used by
the mapping table - around 20x in the most-fragmented cases.

In the most fragmented real-world cases, this reduces memory used by the
mapping from ~1GB to ~50MB of RAM per 1TB of storage removed.  Less
fragmented cases will typically also see around 50-100MB of RAM per 1TB
of storage.

Porting notes:

* Add the following as module parameters:
    * zfs_condense_indirect_vdevs_enable
    * zfs_condense_max_obsolete_bytes

* Document the following module parameters:
   * zfs_condense_indirect_vdevs_enable
   * zfs_condense_max_obsolete_bytes
   * zfs_condense_min_mapping_bytes

Authored by: Matthew Ahrens <[email protected]>
Reviewed by: Brian Behlendorf <[email protected]>
Ported-by: Tim Chase <[email protected]>
Signed-off-by: Tim Chase <[email protected]>

OpenZFS-issue: https://illumos.org/issues/9486
OpenZFS-commit: https://github.com/ahrens/illumos/commit/07152e142e44c
External-issue: DLPX-57962
Closes #7536

1 files changed, 190 insertions, 23 deletions

diff --git a/module/zfs/vdev_removal.c b/module/zfs/vdev_removal.c
index 826e5c421..07c556ed0 100644
--- a/module/zfs/vdev_removal.c
+++ b/module/zfs/vdev_removal.c
@@ -99,6 +99,24 @@ int zfs_remove_max_copy_bytes = 64 * 1024 * 1024;
  */
 int zfs_remove_max_segment = SPA_MAXBLOCKSIZE;
 
+/*
+ * Allow a remap segment to span free chunks of at most this size. The main
+ * impact of a larger span is that we will read and write larger, more
+ * contiguous chunks, with more "unnecessary" data -- trading off bandwidth
+ * for iops.  The value here was chosen to align with
+ * zfs_vdev_read_gap_limit, which is a similar concept when doing regular
+ * reads (but there's no reason it has to be the same).
+ *
+ * Additionally, a higher span will have the following relatively minor
+ * effects:
+ *  - the mapping will be smaller, since one entry can cover more allocated
+ *    segments
+ *  - more of the fragmentation in the removing device will be preserved
+ *  - we'll do larger allocations, which may fail and fall back on smaller
+ *    allocations
+ */
+int vdev_removal_max_span = 32 * 1024;
+
 #define	VDEV_REMOVAL_ZAP_OBJS	"lzap"
 
 static void spa_vdev_remove_thread(void *arg);
@@ -710,13 +728,52 @@ vdev_mapping_sync(void *arg, dmu_tx_t *tx)
 	spa_sync_removing_state(spa, tx);
 }
 
+typedef struct vdev_copy_segment_arg {
+	spa_t *vcsa_spa;
+	dva_t *vcsa_dest_dva;
+	uint64_t vcsa_txg;
+	range_tree_t *vcsa_obsolete_segs;
+} vdev_copy_segment_arg_t;
+
+static void
+unalloc_seg(void *arg, uint64_t start, uint64_t size)
+{
+	vdev_copy_segment_arg_t *vcsa = arg;
+	spa_t *spa = vcsa->vcsa_spa;
+	blkptr_t bp = { { { {0} } } };
+
+	BP_SET_BIRTH(&bp, TXG_INITIAL, TXG_INITIAL);
+	BP_SET_LSIZE(&bp, size);
+	BP_SET_PSIZE(&bp, size);
+	BP_SET_COMPRESS(&bp, ZIO_COMPRESS_OFF);
+	BP_SET_CHECKSUM(&bp, ZIO_CHECKSUM_OFF);
+	BP_SET_TYPE(&bp, DMU_OT_NONE);
+	BP_SET_LEVEL(&bp, 0);
+	BP_SET_DEDUP(&bp, 0);
+	BP_SET_BYTEORDER(&bp, ZFS_HOST_BYTEORDER);
+
+	DVA_SET_VDEV(&bp.blk_dva[0], DVA_GET_VDEV(vcsa->vcsa_dest_dva));
+	DVA_SET_OFFSET(&bp.blk_dva[0],
+	    DVA_GET_OFFSET(vcsa->vcsa_dest_dva) + start);
+	DVA_SET_ASIZE(&bp.blk_dva[0], size);
+
+	zio_free(spa, vcsa->vcsa_txg, &bp);
+}
+
 /*
  * All reads and writes associated with a call to spa_vdev_copy_segment()
  * are done.
  */
 static void
-spa_vdev_copy_nullzio_done(zio_t *zio)
+spa_vdev_copy_segment_done(zio_t *zio)
 {
+	vdev_copy_segment_arg_t *vcsa = zio->io_private;
+
+	range_tree_vacate(vcsa->vcsa_obsolete_segs,
+	    unalloc_seg, vcsa);
+	range_tree_destroy(vcsa->vcsa_obsolete_segs);
+	kmem_free(vcsa, sizeof (*vcsa));
+
 	spa_config_exit(zio->io_spa, SCL_STATE, zio->io_spa);
 }
 
@@ -833,7 +890,8 @@ spa_vdev_copy_one_child(vdev_copy_arg_t *vca, zio_t *nzio,
  * read from the old location and write to the new location.
  */
 static int
-spa_vdev_copy_segment(vdev_t *vd, uint64_t start, uint64_t size, uint64_t txg,
+spa_vdev_copy_segment(vdev_t *vd, range_tree_t *segs,
+    uint64_t maxalloc, uint64_t txg,
     vdev_copy_arg_t *vca, zio_alloc_list_t *zal)
 {
 	metaslab_group_t *mg = vd->vdev_mg;
@@ -841,8 +899,39 @@ spa_vdev_copy_segment(vdev_t *vd, uint64_t start, uint64_t size, uint64_t txg,
 	spa_vdev_removal_t *svr = spa->spa_vdev_removal;
 	vdev_indirect_mapping_entry_t *entry;
 	dva_t dst = {{ 0 }};
+	uint64_t start = range_tree_min(segs);
+
+	ASSERT3U(maxalloc, <=, SPA_MAXBLOCKSIZE);
 
-	ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
+	uint64_t size = range_tree_span(segs);
+	if (range_tree_span(segs) > maxalloc) {
+		/*
+		 * We can't allocate all the segments.  Prefer to end
+		 * the allocation at the end of a segment, thus avoiding
+		 * additional split blocks.
+		 */
+		range_seg_t search;
+		avl_index_t where;
+		search.rs_start = start + maxalloc;
+		search.rs_end = search.rs_start;
+		range_seg_t *rs = avl_find(&segs->rt_root, &search, &where);
+		if (rs == NULL) {
+			rs = avl_nearest(&segs->rt_root, where, AVL_BEFORE);
+		} else {
+			rs = AVL_PREV(&segs->rt_root, rs);
+		}
+		if (rs != NULL) {
+			size = rs->rs_end - start;
+		} else {
+			/*
+			 * There are no segments that end before maxalloc.
+			 * I.e. the first segment is larger than maxalloc,
+			 * so we must split it.
+			 */
+			size = maxalloc;
+		}
+	}
+	ASSERT3U(size, <=, maxalloc);
 
 	int error = metaslab_alloc_dva(spa, mg->mg_class, size,
 	    &dst, 0, NULL, txg, 0, zal);
@@ -850,6 +939,31 @@ spa_vdev_copy_segment(vdev_t *vd, uint64_t start, uint64_t size, uint64_t txg,
 		return (error);
 
 	/*
+	 * Determine the ranges that are not actually needed.  Offsets are
+	 * relative to the start of the range to be copied (i.e. relative to the
+	 * local variable "start").
+	 */
+	range_tree_t *obsolete_segs = range_tree_create(NULL, NULL);
+
+	range_seg_t *rs = avl_first(&segs->rt_root);
+	ASSERT3U(rs->rs_start, ==, start);
+	uint64_t prev_seg_end = rs->rs_end;
+	while ((rs = AVL_NEXT(&segs->rt_root, rs)) != NULL) {
+		if (rs->rs_start >= start + size) {
+			break;
+		} else {
+			range_tree_add(obsolete_segs,
+			    prev_seg_end - start,
+			    rs->rs_start - prev_seg_end);
+		}
+		prev_seg_end = rs->rs_end;
+	}
+	/* We don't end in the middle of an obsolete range */
+	ASSERT3U(start + size, <=, prev_seg_end);
+
+	range_tree_clear(segs, start, size);
+
+	/*
 	 * We can't have any padding of the allocated size, otherwise we will
 	 * misunderstand what's allocated, and the size of the mapping.
 	 * The caller ensures this will be true by passing in a size that is
@@ -860,13 +974,22 @@ spa_vdev_copy_segment(vdev_t *vd, uint64_t start, uint64_t size, uint64_t txg,
 	entry = kmem_zalloc(sizeof (vdev_indirect_mapping_entry_t), KM_SLEEP);
 	DVA_MAPPING_SET_SRC_OFFSET(&entry->vime_mapping, start);
 	entry->vime_mapping.vimep_dst = dst;
+	if (spa_feature_is_enabled(spa, SPA_FEATURE_OBSOLETE_COUNTS)) {
+		entry->vime_obsolete_count = range_tree_space(obsolete_segs);
+	}
+
+	vdev_copy_segment_arg_t *vcsa = kmem_zalloc(sizeof (*vcsa), KM_SLEEP);
+	vcsa->vcsa_dest_dva = &entry->vime_mapping.vimep_dst;
+	vcsa->vcsa_obsolete_segs = obsolete_segs;
+	vcsa->vcsa_spa = spa;
+	vcsa->vcsa_txg = txg;
 
 	/*
 	 * See comment before spa_vdev_copy_one_child().
 	 */
 	spa_config_enter(spa, SCL_STATE, spa, RW_READER);
 	zio_t *nzio = zio_null(spa->spa_txg_zio[txg & TXG_MASK], spa, NULL,
-	    spa_vdev_copy_nullzio_done, NULL, 0);
+	    spa_vdev_copy_segment_done, vcsa, 0);
 	vdev_t *dest_vd = vdev_lookup_top(spa, DVA_GET_VDEV(&dst));
 	if (dest_vd->vdev_ops == &vdev_mirror_ops) {
 		for (int i = 0; i < dest_vd->vdev_children; i++) {
@@ -1069,39 +1192,79 @@ spa_vdev_copy_impl(vdev_t *vd, spa_vdev_removal_t *svr, vdev_copy_arg_t *vca,
 
 	mutex_enter(&svr->svr_lock);
 
-	range_seg_t *rs = avl_first(&svr->svr_allocd_segs->rt_root);
-	if (rs == NULL) {
+	/*
+	 * Determine how big of a chunk to copy.  We can allocate up
+	 * to max_alloc bytes, and we can span up to vdev_removal_max_span
+	 * bytes of unallocated space at a time.  "segs" will track the
+	 * allocated segments that we are copying.  We may also be copying
+	 * free segments (of up to vdev_removal_max_span bytes).
+	 */
+	range_tree_t *segs = range_tree_create(NULL, NULL);
+	for (;;) {
+		range_seg_t *rs = range_tree_first(svr->svr_allocd_segs);
+
+		if (rs == NULL)
+			break;
+
+		uint64_t seg_length;
+
+		if (range_tree_is_empty(segs)) {
+			/* need to truncate the first seg based on max_alloc */
+			seg_length =
+			    MIN(rs->rs_end - rs->rs_start, *max_alloc);
+		} else {
+			if (rs->rs_start - range_tree_max(segs) >
+			    vdev_removal_max_span) {
+				/*
+				 * Including this segment would cause us to
+				 * copy a larger unneeded chunk than is allowed.
+				 */
+				break;
+			} else if (rs->rs_end - range_tree_min(segs) >
+			    *max_alloc) {
+				/*
+				 * This additional segment would extend past
+				 * max_alloc. Rather than splitting this
+				 * segment, leave it for the next mapping.
+				 */
+				break;
+			} else {
+				seg_length = rs->rs_end - rs->rs_start;
+			}
+		}
+
+		range_tree_add(segs, rs->rs_start, seg_length);
+		range_tree_remove(svr->svr_allocd_segs,
+		    rs->rs_start, seg_length);
+	}
+
+	if (range_tree_is_empty(segs)) {
 		mutex_exit(&svr->svr_lock);
+		range_tree_destroy(segs);
 		return;
 	}
-	uint64_t offset = rs->rs_start;
-	uint64_t length = MIN(rs->rs_end - rs->rs_start, *max_alloc);
-
-	range_tree_remove(svr->svr_allocd_segs, offset, length);
 
 	if (svr->svr_max_offset_to_sync[txg & TXG_MASK] == 0) {
 		dsl_sync_task_nowait(dmu_tx_pool(tx), vdev_mapping_sync,
 		    svr, 0, ZFS_SPACE_CHECK_NONE, tx);
 	}
 
-	svr->svr_max_offset_to_sync[txg & TXG_MASK] = offset + length;
+	svr->svr_max_offset_to_sync[txg & TXG_MASK] = range_tree_max(segs);
 
 	/*
 	 * Note: this is the amount of *allocated* space
 	 * that we are taking care of each txg.
 	 */
-	svr->svr_bytes_done[txg & TXG_MASK] += length;
+	svr->svr_bytes_done[txg & TXG_MASK] += range_tree_space(segs);
 
 	mutex_exit(&svr->svr_lock);
 
 	zio_alloc_list_t zal;
 	metaslab_trace_init(&zal);
-	uint64_t thismax = *max_alloc;
-	while (length > 0) {
-		uint64_t mylen = MIN(length, thismax);
-
+	uint64_t thismax = SPA_MAXBLOCKSIZE;
+	while (!range_tree_is_empty(segs)) {
 		int error = spa_vdev_copy_segment(vd,
-		    offset, mylen, txg, vca, &zal);
+		    segs, thismax, txg, vca, &zal);
 
 		if (error == ENOSPC) {
 			/*
@@ -1115,18 +1278,17 @@ spa_vdev_copy_impl(vdev_t *vd, spa_vdev_removal_t *svr, vdev_copy_arg_t *vca,
 			 */
 			ASSERT3U(spa->spa_max_ashift, >=, SPA_MINBLOCKSHIFT);
 			ASSERT3U(spa->spa_max_ashift, ==, spa->spa_min_ashift);
-			thismax = P2ROUNDUP(mylen / 2,
+			uint64_t attempted =
+			    MIN(range_tree_span(segs), thismax);
+			thismax = P2ROUNDUP(attempted / 2,
 			    1 << spa->spa_max_ashift);
-			ASSERT3U(thismax, <, mylen);
 			/*
 			 * The minimum-size allocation can not fail.
 			 */
-			ASSERT3U(mylen, >, 1 << spa->spa_max_ashift);
-			*max_alloc = mylen - (1 << spa->spa_max_ashift);
+			ASSERT3U(attempted, >, 1 << spa->spa_max_ashift);
+			*max_alloc = attempted - (1 << spa->spa_max_ashift);
 		} else {
 			ASSERT0(error);
-			length -= mylen;
-			offset += mylen;
 
 			/*
 			 * We've performed an allocation, so reset the
@@ -1137,6 +1299,7 @@ spa_vdev_copy_impl(vdev_t *vd, spa_vdev_removal_t *svr, vdev_copy_arg_t *vca,
 		}
 	}
 	metaslab_trace_fini(&zal);
+	range_tree_destroy(segs);
 }
 
 /*
@@ -1944,6 +2107,10 @@ module_param(zfs_remove_max_segment, int, 0644);
 MODULE_PARM_DESC(zfs_remove_max_segment,
 	"Largest contiguous segment to allocate when removing device");
 
+module_param(vdev_removal_max_span, int, 0644);
+MODULE_PARM_DESC(vdev_removal_max_span,
+	"Largest span of free chunks a remap segment can span");
+
 EXPORT_SYMBOL(free_from_removing_vdev);
 EXPORT_SYMBOL(spa_removal_get_stats);
 EXPORT_SYMBOL(spa_remove_init);