zfs: merge openzfs/zfs@37af89948

Notable upstream pull request merges:
 #18509 f16b3744d zstream: refactor common functions
 #18573 -multiple Persist z_seq across znode eviction
 s18611 eb0c674c2 zfs_ioctl: fix EBUSY race between quota queries and mount
 #18637 77e64d86e Fix self-deadlock when setting the "allocating"/"path"
                  vdev property
 #18645 e3082b923 freebsd: set mnt_time on the rootfs at mountroot time
 #18652 50d012b2a zbookmark_compare: handle "marker" bookmarks with negative
                  levels
 #18664 520eeeaa6 Improve performance of "zpool offline" for log devices
 #18668 6b8f79877 Avoid more abd_t allocations in RAIDZ/dRAID
 #18669 99ab859c3 Optimize metaslab_set_selected_txg()
 #18673 97b9ba7a9 delegate: add 'send:encrypted' permission
 #18687 2ea519c2a Avoid lookup overhead for nonexistent xattr directories
 #18688 87593ea2b Fix handling of _PC_HAS_HIDDENSYSTEM for FreeBSD
 #18693 0483a8e0c Clean up embedded slog metaslab across txgs
 #18695 41311c665 RAIDZ: Optimize single data column writes
 #18706 37af89948 ddt_log: Fix refcount tagging for begin/commit

Obtained from:	OpenZFS
OpenZFS commit: 37af899488652c55c456d3f160fb8b295db2ec70

1 files changed, 45 insertions, 10 deletions

diff --git a/sys/contrib/openzfs/module/zfs/vdev_raidz.c b/sys/contrib/openzfs/module/zfs/vdev_raidz.c
index 2db7422e772e..ab9f62c3c2d0 100644
--- a/sys/contrib/openzfs/module/zfs/vdev_raidz.c
+++ b/sys/contrib/openzfs/module/zfs/vdev_raidz.c
@@ -419,7 +419,16 @@ static int zfs_scrub_partial_writes = 1;
 static void
 vdev_raidz_row_free(raidz_row_t *rr)
 {
-	for (int c = 0; c < rr->rr_cols; c++) {
+	abd_t *dabd = rr->rr_col[rr->rr_firstdatacol].rc_abd;
+	for (int c = 0; c < rr->rr_firstdatacol; c++) {
+		raidz_col_t *rc = &rr->rr_col[c];
+
+		if (rc->rc_size != 0 && rc->rc_abd != dabd)
+			abd_free(rc->rc_abd);
+		if (rc->rc_orig_data != NULL)
+			abd_free(rc->rc_orig_data);
+	}
+	for (int c = rr->rr_firstdatacol; c < rr->rr_cols; c++) {
 		raidz_col_t *rc = &rr->rr_col[c];
 
 		if (rc->rc_size != 0)
@@ -532,6 +541,22 @@ vdev_raidz_map_alloc_write(zio_t *zio, raidz_map_t *rm, uint64_t ashift)
 	 */
 	int skipped = rr->rr_scols - rr->rr_cols;
 
+	/*
+	 * When there is only a single data column the parity is a copy of
+	 * it, so point all parity columns at the data ABD directly to avoid
+	 * allocating buffers and computing parity.
+	 */
+	if (rr->rr_cols == rr->rr_firstdatacol + 1) {
+		ASSERT0(nwrapped);
+		ASSERT0(rm->rm_nskip);
+		raidz_col_t *dc = &rr->rr_col[rr->rr_firstdatacol];
+		dc->rc_abd = abd_get_offset_struct(&dc->rc_abdstruct,
+		    zio->io_abd, 0, dc->rc_size);
+		for (c = 0; c < rr->rr_firstdatacol; c++)
+			rr->rr_col[c].rc_abd = dc->rc_abd;
+		return;
+	}
+
 	/* Allocate buffers for the parity columns */
 	for (c = 0; c < rr->rr_firstdatacol; c++) {
 		raidz_col_t *rc = &rr->rr_col[c];
@@ -546,12 +571,13 @@ vdev_raidz_map_alloc_write(zio_t *zio, raidz_map_t *rm, uint64_t ashift)
 		 * VDEV queue locks (vq_lock).
 		 */
 		if (c < nwrapped) {
-			rc->rc_abd = abd_alloc_linear(
+			rc->rc_abd = abd_alloc_linear_struct(&rc->rc_abdstruct,
 			    rc->rc_size + (1ULL << ashift), B_FALSE);
 			abd_zero_off(rc->rc_abd, rc->rc_size, 1ULL << ashift);
 			skipped++;
 		} else {
-			rc->rc_abd = abd_alloc_linear(rc->rc_size, B_FALSE);
+			rc->rc_abd = abd_alloc_linear_struct(&rc->rc_abdstruct,
+			    rc->rc_size, B_FALSE);
 		}
 	}
 
@@ -599,9 +625,11 @@ vdev_raidz_map_alloc_read(zio_t *zio, raidz_map_t *rm)
 	ASSERT3U(rm->rm_nrows, ==, 1);
 
 	/* Allocate buffers for the parity columns */
-	for (c = 0; c < rr->rr_firstdatacol; c++)
-		rr->rr_col[c].rc_abd =
-		    abd_alloc_linear(rr->rr_col[c].rc_size, B_FALSE);
+	for (c = 0; c < rr->rr_firstdatacol; c++) {
+		raidz_col_t *rc = &rr->rr_col[c];
+		rc->rc_abd = abd_alloc_linear_struct(&rc->rc_abdstruct,
+		    rc->rc_size, B_FALSE);
+	}
 
 	for (uint64_t off = 0; c < rr->rr_cols; c++) {
 		raidz_col_t *rc = &rr->rr_col[c];
@@ -1046,8 +1074,8 @@ vdev_raidz_map_alloc_expanded(zio_t *zio,
 				continue;
 
 			prc->rc_abd =
-			    abd_alloc_linear(rm->rm_phys_col[i].rc_size,
-			    B_FALSE);
+			    abd_alloc_linear_struct(&prc->rc_abdstruct,
+			    prc->rc_size, B_FALSE);
 		}
 
 		/*
@@ -1075,8 +1103,8 @@ vdev_raidz_map_alloc_expanded(zio_t *zio,
 			for (int c = 0; c < rr->rr_firstdatacol; c++) {
 				raidz_col_t *rc = &rr->rr_col[c];
 				rc->rc_abd =
-				    abd_alloc_linear(rc->rc_size,
-				    B_TRUE);
+				    abd_alloc_linear_struct(&rc->rc_abdstruct,
+				    rc->rc_size, B_TRUE);
 			}
 		}
 	}
@@ -1272,6 +1300,13 @@ vdev_raidz_generate_parity_row(raidz_map_t *rm, raidz_row_t *rr)
 		return;
 	}
 
+	/*
+	 * Single data column: parity is the data itself.
+	 */
+	if (rr->rr_col[VDEV_RAIDZ_P].rc_abd ==
+	    rr->rr_col[rr->rr_firstdatacol].rc_abd)
+		return;
+
 	/* Generate using the new math implementation */
 	if (vdev_raidz_math_generate(rm, rr) != RAIDZ_ORIGINAL_IMPL)
 		return;