Skip to contentSkip to conversation

libfrog: hoist bitmap out of scrub

PATCH In review Darrick J. Wong 2 participants 5 messages Jan 30, 2019 – Feb 1, 2019

lore
Share & tools
lore
#0 lore raw

PATCH 1/2 libfrog: hoist bitmap out of scrub

From: Darrick J. Wong <[email protected]>

Move the bitmap code to libfrog so that we can use bitmaps in xfs_repair.

---
 include/bitmap.h |   24 +++
 libfrog/Makefile |    1 
 libfrog/bitmap.c |  393 ++++++++++++++++++++++++++++++++++++++++++++++++++++++
 scrub/Makefile   |    2 
 scrub/bitmap.c   |  393 ------------------------------------------------------
 scrub/bitmap.h   |   24 ---
 6 files changed, 418 insertions(+), 419 deletions(-)
 create mode 100644 include/bitmap.h
 create mode 100644 libfrog/bitmap.c
 delete mode 100644 scrub/bitmap.c
 delete mode 100644 scrub/bitmap.h
include/bitmap.h new file +24 -0
new file mode 100644
index 00000000..e29a4335
@@ -0,0 +1,24 @@
+// SPDX-License-Identifier: GPL-2.0+
+/*
+ * Copyright (C) 2018 Oracle.  All Rights Reserved.
+ * Author: Darrick J. Wong <[email protected]>
+ */
+#ifndef LIBFROG_BITMAP_H_
+#define LIBFROG_BITMAP_H_
+
+struct bitmap {
+	pthread_mutex_t		bt_lock;
+	struct avl64tree_desc	*bt_tree;
+};
+
+bool bitmap_init(struct bitmap **bmap);
+void bitmap_free(struct bitmap **bmap);
+bool bitmap_set(struct bitmap *bmap, uint64_t start, uint64_t length);
+bool bitmap_iterate(struct bitmap *bmap,
+		bool (*fn)(uint64_t, uint64_t, void *), void *arg);
+bool bitmap_test(struct bitmap *bmap, uint64_t start,
+		uint64_t len);
+bool bitmap_empty(struct bitmap *bmap);
+void bitmap_dump(struct bitmap *bmap);
+
+#endif /* LIBFROG_BITMAP_H_ */
libfrog/Makefile +1 -0
index dbff9596..f5a0539b 100644
@@ -12,6 +12,7 @@ LT_AGE = 0
 
 CFILES = \
 avl64.c \
+bitmap.c \
 convert.c \
 crc32.c \
 fsgeom.c \
libfrog/bitmap.c new file +393 -0

393 lines not on this page · raw message

scrub/Makefile +0 -2
index c6473c12..9c56d550 100644
@@ -31,7 +31,6 @@ endif
 endif	# scrub_prereqs
 
 HFILES = \
-bitmap.h \
 common.h \
 counter.h \
 disk.h \
@@ -48,7 +47,6 @@ vfs.h \
 xfs_scrub.h
 
 CFILES = \
-bitmap.c \
 common.c \
 counter.c \
 disk.c \
scrub/bitmap.c deleted +0 -393

393 lines not on this page · raw message

scrub/bitmap.h deleted +0 -24
deleted file mode 100644
index e9746a12..00000000
@@ -1,24 +0,0 @@
-// SPDX-License-Identifier: GPL-2.0+
-/*
- * Copyright (C) 2018 Oracle.  All Rights Reserved.
- * Author: Darrick J. Wong <[email protected]>
- */
-#ifndef XFS_SCRUB_BITMAP_H_
-#define XFS_SCRUB_BITMAP_H_
-
-struct bitmap {
-	pthread_mutex_t		bt_lock;
-	struct avl64tree_desc	*bt_tree;
-};
-
-bool bitmap_init(struct bitmap **bmap);
-void bitmap_free(struct bitmap **bmap);
-bool bitmap_set(struct bitmap *bmap, uint64_t start, uint64_t length);
-bool bitmap_iterate(struct bitmap *bmap,
-		bool (*fn)(uint64_t, uint64_t, void *), void *arg);
-bool bitmap_test(struct bitmap *bmap, uint64_t start,
-		uint64_t len);
-bool bitmap_empty(struct bitmap *bmap);
-void bitmap_dump(struct bitmap *bmap);
-
-#endif /* XFS_SCRUB_BITMAP_H_ */
#1 lore raw

PATCH 2/2 xfs_repair: correctly account for free space btree shrinks when fixing freelist

From: Darrick J. Wong <[email protected]>

When we fix the freelist at the end of build_agf_agfl in phase 5 of repair, we need to create incore rmap records for the blocks that get added to the AGFL. We can't let the regular freelist fixing code use the regular on-disk rmapbt update code because the rmapbt isn't fully set up yet.

Unfortunately, the original code fails to account for the fact that the free space btrees can shrink when we allocate blocks to fix the freelist; those blocks are also put on the freelist, but there are already incore rmaps for all the free space btree blocks. We must not create (redundant) incore rmaps for those blocks. If we do, repair fails with a complaint that rebuilding the rmapbt failed during phase 5. xfs/137 on a 1k block size occasionally triggers this bug.

To fix the problem, construct a bitmap of all OWN_AG blocks that we know about before traversing the AGFL, and only create new incore rmaps for those AGFL blocks that are not already tracked in the bitmap.

---
 repair/rmap.c |   54 ++++++++++++++++++++++++++++++++++++++++++++----------
 1 file changed, 44 insertions(+), 10 deletions(-)
repair/rmap.c +44 -10
index c5a86646..3bb05065 100644
@@ -12,6 +12,7 @@
 #include "dinode.h"
 #include "slab.h"
 #include "rmap.h"
+#include "bitmap.h"
 
 #undef RMAP_DEBUG
 
@@ -450,6 +451,8 @@ rmap_store_ag_btree_rec(
 	struct xfs_buf		*agflbp = NULL;
 	struct xfs_trans	*tp;
 	__be32			*agfl_bno, *b;
+	struct xfs_ag_rmap	*ag_rmap = &ag_rmaps[agno];
+	struct bitmap		*own_ag_bitmap = NULL;
 	int			error = 0;
 	struct xfs_owner_info	oinfo;
 
@@ -457,9 +460,8 @@ rmap_store_ag_btree_rec(
 		return 0;
 
 	/* Release the ar_rmaps; they were put into the rmapbt during p5. */
-	free_slab(&ag_rmaps[agno].ar_rmaps);
-	error = init_slab(&ag_rmaps[agno].ar_rmaps,
-				  sizeof(struct xfs_rmap_irec));
+	free_slab(&ag_rmap->ar_rmaps);
+	error = init_slab(&ag_rmap->ar_rmaps, sizeof(struct xfs_rmap_irec));
 	if (error)
 		goto err;
 
@@ -479,19 +481,50 @@ rmap_store_ag_btree_rec(
 	 * rmap, we only need to add rmap records for AGFL blocks past
 	 * that point in the AGFL because those blocks are a result of a
 	 * no-rmap no-shrink freelist fixup that we did earlier.
+	 *
+	 * However, some blocks end up on the AGFL because the free space
+	 * btrees shed blocks as a result of allocating space to fix the
+	 * freelist.  We already created in-core rmap records for the free
+	 * space btree blocks, so we must be careful not to create those
+	 * records again.  Create a bitmap of already-recorded OWN_AG rmaps.
 	 */
+	error = init_slab_cursor(ag_rmap->ar_raw_rmaps, rmap_compare, &rm_cur);
+	if (error)
+		goto err;
+	if (!bitmap_init(&own_ag_bitmap)) {
+		error = -ENOMEM;
+		goto err;
+	}
+	while ((rm_rec = pop_slab_cursor(rm_cur)) != NULL) {
+		if (rm_rec->rm_owner != XFS_RMAP_OWN_AG)
+			continue;
+		if (!bitmap_set(own_ag_bitmap, rm_rec->rm_startblock,
+					rm_rec->rm_blockcount)) {
+			error = EFSCORRUPTED;
+			goto err;
+		}
+	}
+	free_slab_cursor(&rm_cur);
+
+	/* Create rmaps for any AGFL blocks that aren't already rmapped. */
 	agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp);
-	b = agfl_bno + ag_rmaps[agno].ar_flcount;
+	b = agfl_bno + ag_rmap->ar_flcount;
 	while (*b != cpu_to_be32(NULLAGBLOCK) &&
 	       b - agfl_bno < libxfs_agfl_size(mp)) {
-		error = rmap_add_ag_rec(mp, agno, be32_to_cpu(*b), 1,
-				XFS_RMAP_OWN_AG);
-		if (error)
-			goto err;
+		xfs_agblock_t	agbno;
+
+		agbno = be32_to_cpu(*b);
+		if (!bitmap_test(own_ag_bitmap, agbno, 1)) {
+			error = rmap_add_ag_rec(mp, agno, agbno, 1,
+					XFS_RMAP_OWN_AG);
+			if (error)
+				goto err;
+		}
 		b++;
 	}
 	libxfs_putbuf(agflbp);
 	agflbp = NULL;
+	bitmap_free(&own_ag_bitmap);
 
 	/* Merge all the raw rmaps into the main list */
 	error = rmap_fold_raw_recs(mp, agno);
@@ -499,8 +532,7 @@ rmap_store_ag_btree_rec(
 		goto err;
 
 	/* Create cursors to refcount structures */
-	error = init_slab_cursor(ag_rmaps[agno].ar_rmaps, rmap_compare,
-			&rm_cur);
+	error = init_slab_cursor(ag_rmap->ar_rmaps, rmap_compare, &rm_cur);
 	if (error)
 		goto err;
 
@@ -541,6 +573,8 @@ rmap_store_ag_btree_rec(
 err:
 	if (agflbp)
 		libxfs_putbuf(agflbp);
+	if (own_ag_bitmap)
+		bitmap_free(&own_ag_bitmap);
 	return error;
 }
 
#2 lore raw

Re: [PATCH 2/2] xfs_repair: correctly account for free space btree shrinks when fixing freelist

Quoted 66 lines · Darrick J. Wong

From: Darrick J. Wong <[email protected]>

When we fix the freelist at the end of build_agf_agfl in phase 5 of repair, we need to create incore rmap records for the blocks that get added to the AGFL. We can't let the regular freelist fixing code use the regular on-disk rmapbt update code because the rmapbt isn't fully set up yet.

Unfortunately, the original code fails to account for the fact that the free space btrees can shrink when we allocate blocks to fix the freelist; those blocks are also put on the freelist, but there are already incore rmaps for all the free space btree blocks. We must not create (redundant) incore rmaps for those blocks. If we do, repair fails with a complaint that rebuilding the rmapbt failed during phase 5. xfs/137 on a 1k block size occasionally triggers this bug.

To fix the problem, construct a bitmap of all OWN_AG blocks that we know about before traversing the AGFL, and only create new incore rmaps for those AGFL blocks that are not already tracked in the bitmap.

---
 repair/rmap.c |   54 ++++++++++++++++++++++++++++++++++++++++++++----------
 1 file changed, 44 insertions(+), 10 deletions(-)
repair/rmap.c +8 -3
index c5a86646..3bb05065 100644
@@ -12,6 +12,7 @@
 #include "dinode.h"
 #include "slab.h"
 #include "rmap.h"
+#include "bitmap.h"
 
 #undef RMAP_DEBUG
 
@@ -450,6 +451,8 @@ rmap_store_ag_btree_rec(
 	struct xfs_buf		*agflbp = NULL;
 	struct xfs_trans	*tp;
 	__be32			*agfl_bno, *b;
+	struct xfs_ag_rmap	*ag_rmap = &ag_rmaps[agno];
+	struct bitmap		*own_ag_bitmap = NULL;
 	int			error = 0;
 	struct xfs_owner_info	oinfo;
 
@@ -457,9 +460,8 @@ rmap_store_ag_btree_rec(
 		return 0;
 
 	/* Release the ar_rmaps; they were put into the rmapbt during p5. */
-	free_slab(&ag_rmaps[agno].ar_rmaps);
-	error = init_slab(&ag_rmaps[agno].ar_rmaps,
-				  sizeof(struct xfs_rmap_irec));
+	free_slab(&ag_rmap->ar_rmaps);
+	error = init_slab(&ag_rmap->ar_rmaps, sizeof(struct xfs_rmap_irec));
 	if (error)
 		goto err;
 
@@ -479,19 +481,50 @@ rmap_store_ag_btree_rec(
 	 * rmap, we only need to add rmap records for AGFL blocks past
 	 * that point in the AGFL because those blocks are a result of a
 	 * no-rmap no-shrink freelist fixup that we did earlier.
+	 *
+	 * However, some blocks end up on the AGFL because the free space
+	 * btrees shed blocks as a result of allocating space to fix the

repair/rmap.c rmap_store_ag_btree_rec( · line 496

Quoted 10 lines
repair/rmap.c +9 -0
+	 * freelist.  We already created in-core rmap records for the free
+	 * space btree blocks, so we must be careful not to create those
+	 * records again.  Create a bitmap of already-recorded OWN_AG rmaps.
 	 */
+	error = init_slab_cursor(ag_rmap->ar_raw_rmaps, rmap_compare, &rm_cur);
+	if (error)
+		goto err;
+	if (!bitmap_init(&own_ag_bitmap)) {
+		error = -ENOMEM;
+		goto err;

I think that this ^^

repair/rmap.c rmap_store_ag_btree_rec( · line 504

Quoted 8 lines
repair/rmap.c +8 -0
+	}
+	while ((rm_rec = pop_slab_cursor(rm_cur)) != NULL) {
+		if (rm_rec->rm_owner != XFS_RMAP_OWN_AG)
+			continue;
+		if (!bitmap_set(own_ag_bitmap, rm_rec->rm_startblock,
+					rm_rec->rm_blockcount)) {
+			error = EFSCORRUPTED;
+			goto err;

and this need to free up rm_cur to be tidy, right?

-Eric

#3 lore raw

Re: [PATCH 2/2] xfs_repair: correctly account for free space btree shrinks when fixing freelist

Quoted 80 lines · Eric Sandeen
Quoted 66 lines · Darrick J. Wong

From: Darrick J. Wong <[email protected]>

When we fix the freelist at the end of build_agf_agfl in phase 5 of repair, we need to create incore rmap records for the blocks that get added to the AGFL. We can't let the regular freelist fixing code use the regular on-disk rmapbt update code because the rmapbt isn't fully set up yet.

Unfortunately, the original code fails to account for the fact that the free space btrees can shrink when we allocate blocks to fix the freelist; those blocks are also put on the freelist, but there are already incore rmaps for all the free space btree blocks. We must not create (redundant) incore rmaps for those blocks. If we do, repair fails with a complaint that rebuilding the rmapbt failed during phase 5. xfs/137 on a 1k block size occasionally triggers this bug.

To fix the problem, construct a bitmap of all OWN_AG blocks that we know about before traversing the AGFL, and only create new incore rmaps for those AGFL blocks that are not already tracked in the bitmap.

---
 repair/rmap.c |   54 ++++++++++++++++++++++++++++++++++++++++++++----------
 1 file changed, 44 insertions(+), 10 deletions(-)
repair/rmap.c +8 -3
index c5a86646..3bb05065 100644
@@ -12,6 +12,7 @@
 #include "dinode.h"
 #include "slab.h"
 #include "rmap.h"
+#include "bitmap.h"
 
 #undef RMAP_DEBUG
 
@@ -450,6 +451,8 @@ rmap_store_ag_btree_rec(
 	struct xfs_buf		*agflbp = NULL;
 	struct xfs_trans	*tp;
 	__be32			*agfl_bno, *b;
+	struct xfs_ag_rmap	*ag_rmap = &ag_rmaps[agno];
+	struct bitmap		*own_ag_bitmap = NULL;
 	int			error = 0;
 	struct xfs_owner_info	oinfo;
 
@@ -457,9 +460,8 @@ rmap_store_ag_btree_rec(
 		return 0;
 
 	/* Release the ar_rmaps; they were put into the rmapbt during p5. */
-	free_slab(&ag_rmaps[agno].ar_rmaps);
-	error = init_slab(&ag_rmaps[agno].ar_rmaps,
-				  sizeof(struct xfs_rmap_irec));
+	free_slab(&ag_rmap->ar_rmaps);
+	error = init_slab(&ag_rmap->ar_rmaps, sizeof(struct xfs_rmap_irec));
 	if (error)
 		goto err;
 
@@ -479,19 +481,50 @@ rmap_store_ag_btree_rec(
 	 * rmap, we only need to add rmap records for AGFL blocks past
 	 * that point in the AGFL because those blocks are a result of a
 	 * no-rmap no-shrink freelist fixup that we did earlier.
+	 *
+	 * However, some blocks end up on the AGFL because the free space
+	 * btrees shed blocks as a result of allocating space to fix the

repair/rmap.c rmap_store_ag_btree_rec( · line 496

Quoted 10 lines
repair/rmap.c +9 -0
+	 * freelist.  We already created in-core rmap records for the free
+	 * space btree blocks, so we must be careful not to create those
+	 * records again.  Create a bitmap of already-recorded OWN_AG rmaps.
 	 */
+	error = init_slab_cursor(ag_rmap->ar_raw_rmaps, rmap_compare, &rm_cur);
+	if (error)
+		goto err;
+	if (!bitmap_init(&own_ag_bitmap)) {
+		error = -ENOMEM;
+		goto err;

I think that this ^^

Quoted 10 lines

repair/rmap.c rmap_store_ag_btree_rec( · line 504

Quoted 8 lines
repair/rmap.c +8 -0
+	}
+	while ((rm_rec = pop_slab_cursor(rm_cur)) != NULL) {
+		if (rm_rec->rm_owner != XFS_RMAP_OWN_AG)
+			continue;
+		if (!bitmap_set(own_ag_bitmap, rm_rec->rm_startblock,
+					rm_rec->rm_blockcount)) {
+			error = EFSCORRUPTED;
+			goto err;

and this need to free up rm_cur to be tidy, right?

Yep, they do need to be 'goto err_slab', thanks for catching that.

--D

Quoted 1 line

-Eric

#4 lore raw

PATCH v2 2/2 xfs_repair: correctly account for free space btree shrinks when fixing freelist

In reply to Darrick J. Wong

From: Darrick J. Wong <[email protected]>

When we fix the freelist at the end of build_agf_agfl in phase 5 of repair, we need to create incore rmap records for the blocks that get added to the AGFL. We can't let the regular freelist fixing code use the regular on-disk rmapbt update code because the rmapbt isn't fully set up yet.

Unfortunately, the original code fails to account for the fact that the free space btrees can shrink when we allocate blocks to fix the freelist; those blocks are also put on the freelist, but there are already incore rmaps for all the free space btree blocks. We must not create (redundant) incore rmaps for those blocks. If we do, repair fails with a complaint that rebuilding the rmapbt failed during phase 5. xfs/137 on a 1k block size occasionally triggers this bug.

To fix the problem, construct a bitmap of all OWN_AG blocks that we know about before traversing the AGFL, and only create new incore rmaps for those AGFL blocks that are not already tracked in the bitmap.

---
v2: fix error jumps

---
 repair/rmap.c |   54 ++++++++++++++++++++++++++++++++++++++++++++----------
 1 file changed, 44 insertions(+), 10 deletions(-)
repair/rmap.c +44 -10
index c5a86646..b48701df 100644
@@ -12,6 +12,7 @@
 #include "dinode.h"
 #include "slab.h"
 #include "rmap.h"
+#include "bitmap.h"
 
 #undef RMAP_DEBUG
 
@@ -450,6 +451,8 @@ rmap_store_ag_btree_rec(
 	struct xfs_buf		*agflbp = NULL;
 	struct xfs_trans	*tp;
 	__be32			*agfl_bno, *b;
+	struct xfs_ag_rmap	*ag_rmap = &ag_rmaps[agno];
+	struct bitmap		*own_ag_bitmap = NULL;
 	int			error = 0;
 	struct xfs_owner_info	oinfo;
 
@@ -457,9 +460,8 @@ rmap_store_ag_btree_rec(
 		return 0;
 
 	/* Release the ar_rmaps; they were put into the rmapbt during p5. */
-	free_slab(&ag_rmaps[agno].ar_rmaps);
-	error = init_slab(&ag_rmaps[agno].ar_rmaps,
-				  sizeof(struct xfs_rmap_irec));
+	free_slab(&ag_rmap->ar_rmaps);
+	error = init_slab(&ag_rmap->ar_rmaps, sizeof(struct xfs_rmap_irec));
 	if (error)
 		goto err;
 
@@ -479,19 +481,50 @@ rmap_store_ag_btree_rec(
 	 * rmap, we only need to add rmap records for AGFL blocks past
 	 * that point in the AGFL because those blocks are a result of a
 	 * no-rmap no-shrink freelist fixup that we did earlier.
+	 *
+	 * However, some blocks end up on the AGFL because the free space
+	 * btrees shed blocks as a result of allocating space to fix the
+	 * freelist.  We already created in-core rmap records for the free
+	 * space btree blocks, so we must be careful not to create those
+	 * records again.  Create a bitmap of already-recorded OWN_AG rmaps.
 	 */
+	error = init_slab_cursor(ag_rmap->ar_raw_rmaps, rmap_compare, &rm_cur);
+	if (error)
+		goto err;
+	if (!bitmap_init(&own_ag_bitmap)) {
+		error = -ENOMEM;
+		goto err_slab;
+	}
+	while ((rm_rec = pop_slab_cursor(rm_cur)) != NULL) {
+		if (rm_rec->rm_owner != XFS_RMAP_OWN_AG)
+			continue;
+		if (!bitmap_set(own_ag_bitmap, rm_rec->rm_startblock,
+					rm_rec->rm_blockcount)) {
+			error = EFSCORRUPTED;
+			goto err_slab;
+		}
+	}
+	free_slab_cursor(&rm_cur);
+
+	/* Create rmaps for any AGFL blocks that aren't already rmapped. */
 	agfl_bno = XFS_BUF_TO_AGFL_BNO(mp, agflbp);
-	b = agfl_bno + ag_rmaps[agno].ar_flcount;
+	b = agfl_bno + ag_rmap->ar_flcount;
 	while (*b != cpu_to_be32(NULLAGBLOCK) &&
 	       b - agfl_bno < libxfs_agfl_size(mp)) {
-		error = rmap_add_ag_rec(mp, agno, be32_to_cpu(*b), 1,
-				XFS_RMAP_OWN_AG);
-		if (error)
-			goto err;
+		xfs_agblock_t	agbno;
+
+		agbno = be32_to_cpu(*b);
+		if (!bitmap_test(own_ag_bitmap, agbno, 1)) {
+			error = rmap_add_ag_rec(mp, agno, agbno, 1,
+					XFS_RMAP_OWN_AG);
+			if (error)
+				goto err;
+		}
 		b++;
 	}
 	libxfs_putbuf(agflbp);
 	agflbp = NULL;
+	bitmap_free(&own_ag_bitmap);
 
 	/* Merge all the raw rmaps into the main list */
 	error = rmap_fold_raw_recs(mp, agno);
@@ -499,8 +532,7 @@ rmap_store_ag_btree_rec(
 		goto err;
 
 	/* Create cursors to refcount structures */
-	error = init_slab_cursor(ag_rmaps[agno].ar_rmaps, rmap_compare,
-			&rm_cur);
+	error = init_slab_cursor(ag_rmap->ar_rmaps, rmap_compare, &rm_cur);
 	if (error)
 		goto err;
 
@@ -541,6 +573,8 @@ rmap_store_ag_btree_rec(
 err:
 	if (agflbp)
 		libxfs_putbuf(agflbp);
+	if (own_ag_bitmap)
+		bitmap_free(&own_ag_bitmap);
 	return error;
 }