Skip to contentSkip to conversation

xfs: skip background cowblock trims on inodes open for write

PATCH Applied Brian Foster 3 participants 7 messages Sep 3, 2024 – Oct 11, 2024 XFS filesystem

lore
Share & tools
lore
#0 lore raw

PATCH v2 1/1 xfs: skip background cowblock trims on inodes open for write

The background blockgc scanner runs on a 5m interval by default and trims preallocation (post-eof and cow fork) from inodes that are otherwise idle. Idle effectively means that iolock can be acquired without blocking and that the inode has no dirty pagecache or I/O in flight.

This simple mechanism and heuristic has worked fairly well for post-eof speculative preallocations. Support for reflink and COW fork preallocations came sometime later and plugged into the same mechanism, with similar heuristics. Some recent testing has shown that COW fork preallocation may be notably more sensitive to blockgc processing than post-eof preallocation, however.

For example, consider an 8GB reflinked file with a COW extent size hint of 1MB. A worst case fully randomized overwrite of this file results in ~8k extents of an average size of ~1MB. If the same workload is interrupted a couple times for blockgc processing (assuming the file goes idle), the resulting extent count explodes to over 100k extents with an average size <100kB. This is significantly worse than ideal and essentially defeats the COW extent size hint mechanism.

While this particular test is instrumented, it reflects a fairly reasonable pattern in practice where random I/Os might spread out over a large period of time with varying periods of (in)activity. For example, consider a cloned disk image file for a VM or container with long uptime and variable and bursty usage. A background blockgc scan that races and processes the image file when it happens to be clean and idle can have a significant effect on the future fragmentation level of the file, even when still in use.

To help combat this, update the heuristic to skip cowblocks inodes that are currently opened for write access during non-sync blockgc scans. This allows COW fork preallocations to persist for as long as possible unless otherwise needed for functional purposes (i.e. a sync scan), the file is idle and closed, or the inode is being evicted from cache. While here, update the comments to help distinguish performance oriented heuristics from the logic that exists to maintain functional correctness.

---

v2:
- Reorder logic and update comments in xfs_prep_free_cowblocks().
v1: https://lore.kernel.org/linux-xfs/[email protected]/

 fs/xfs/xfs_icache.c | 31 +++++++++++++++++++++++--------
 1 file changed, 23 insertions(+), 8 deletions(-)
fs/xfs/xfs_icache.c +23 -8
index cf629302d48e..900a6277d931 100644
@@ -1241,14 +1241,17 @@ xfs_inode_clear_eofblocks_tag(
 }
 
 /*
- * Set ourselves up to free CoW blocks from this file.  If it's already clean
- * then we can bail out quickly, but otherwise we must back off if the file
- * is undergoing some kind of write.
+ * Prepare to free COW fork blocks from an inode.
  */
 static bool
 xfs_prep_free_cowblocks(
-	struct xfs_inode	*ip)
+	struct xfs_inode	*ip,
+	struct xfs_icwalk	*icw)
 {
+	bool			sync;
+
+	sync = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
+
 	/*
 	 * Just clear the tag if we have an empty cow fork or none at all. It's
 	 * possible the inode was fully unshared since it was originally tagged.
@@ -1260,9 +1263,21 @@ xfs_prep_free_cowblocks(
 	}
 
 	/*
-	 * If the mapping is dirty or under writeback we cannot touch the
-	 * CoW fork.  Leave it alone if we're in the midst of a directio.
+	 * A cowblocks trim of an inode can have a significant effect on
+	 * fragmentation even when a reasonable COW extent size hint is set.
+	 * Therefore, we prefer to not process cowblocks unless they are clean
+	 * and idle. We can never process a cowblocks inode that is dirty or has
+	 * in-flight I/O under any circumstances, because outstanding writeback
+	 * or dio expects targeted COW fork blocks exist through write
+	 * completion where they can be remapped into the data fork.
+	 *
+	 * Therefore, the heuristic used here is to never process inodes
+	 * currently opened for write from background (i.e. non-sync) scans. For
+	 * sync scans, use the pagecache/dio state of the inode to ensure we
+	 * never free COW fork blocks out from under pending I/O.
 	 */
+	if (!sync && inode_is_open_for_write(VFS_I(ip)))
+		return false;
 	if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
 	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
 	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
@@ -1298,7 +1313,7 @@ xfs_inode_free_cowblocks(
 	if (!xfs_iflags_test(ip, XFS_ICOWBLOCKS))
 		return 0;
 
-	if (!xfs_prep_free_cowblocks(ip))
+	if (!xfs_prep_free_cowblocks(ip, icw))
 		return 0;
 
 	if (!xfs_icwalk_match(ip, icw))
@@ -1327,7 +1342,7 @@ xfs_inode_free_cowblocks(
 	 * Check again, nobody else should be able to dirty blocks or change
 	 * the reflink iflag now that we have the first two locks held.
 	 */
-	if (xfs_prep_free_cowblocks(ip))
+	if (xfs_prep_free_cowblocks(ip, icw))
 		ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);
 	return ret;
 }
Signature · 1 line
2.45.0
#1 lore raw

[PATCH 2/1] xfs: don't free cowblocks from under dirty pagecache on unshare

fallocate unshare mode explicitly breaks extent sharing. When a command completes, it checks the data fork for any remaining shared extents to determine whether the reflink inode flag and COW fork preallocation can be removed. This logic doesn't consider in-core pagecache and I/O state, however, which means we can unsafely remove COW fork blocks that are still needed under certain conditions.

For example, consider the following command sequence:

xfs_io -fc "pwrite 0 1k" -c "reflink <file> 0 256k 1k" \
	-c "pwrite 0 32k" -c "funshare 0 1k" <file>

This allocates a data block at offset 0, shares it, and then overwrites it with a larger buffered write. The overwrite triggers COW fork preallocation, 32 blocks by default, which maps the entire 32k write to delalloc in the COW fork. All but the shared block at offset 0 remains hole mapped in the data fork. The unshare command redirties and flushes the folio at offset 0, removing the only shared extent from the inode. Since the inode no longer maps shared extents, unshare purges the COW fork before the remaining 28k may have written back.

This leaves dirty pagecache backed by holes, which writeback quietly skips, thus leaving clean, non-zeroed pagecache over holes in the file. To verify, fiemap shows holes in the first 32k of the file and reads return different data across a remount:

$ xfs_io -c "fiemap -v" <file>
<file>:
 EXT: FILE-OFFSET      BLOCK-RANGE      TOTAL FLAGS
   ...
   1: [8..511]:        hole               504
   ...
$ xfs_io -c "pread -v 4k 8" <file>
00001000:  cd cd cd cd cd cd cd cd  ........
$ umount <mnt>; mount <dev> <mnt>
$ xfs_io -c "pread -v 4k 8" <file>
00001000:  00 00 00 00 00 00 00 00  ........

To avoid this problem, make unshare follow the same rules used for background cowblock scanning and never purge the COW fork for inodes with dirty pagecache or in-flight I/O.

Fixes 46afb0628b ("xfs: only flush the unshared range in xfs_reflink_unshare")

---

Here's another COW issue I came across via some unshare testing. A quick hack to enable unshare in fsx uncovered it. I'll follow up with a proper patch for that.

I'm sending this as a 2/1 here just to reflect patch order in my local tree. Also note that I haven't explicitly tested the fixes commit, but a quick test to switch back to the old full flush behavior on latest master also makes the problem go away, so I suspect that's where the regression was introduced.

Brian

 fs/xfs/xfs_icache.c  |  8 +-------
 fs/xfs/xfs_reflink.c |  3 +++
 fs/xfs/xfs_reflink.h | 19 +++++++++++++++++++
 3 files changed, 23 insertions(+), 7 deletions(-)
fs/xfs/xfs_icache.c +1 -7
index 900a6277d931..a1b34e6ccfe2 100644
@@ -1278,13 +1278,7 @@ xfs_prep_free_cowblocks(
 	 */
 	if (!sync && inode_is_open_for_write(VFS_I(ip)))
 		return false;
-	if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
-	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
-	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
-	    atomic_read(&VFS_I(ip)->i_dio_count))
-		return false;
-
-	return true;
+	return xfs_can_free_cowblocks(ip);
 }
 
 /*
fs/xfs/xfs_reflink.c +3 -0
index 6fde6ec8092f..5bf6682e701b 100644
@@ -1595,6 +1595,9 @@ xfs_reflink_clear_inode_flag(
 
 	ASSERT(xfs_is_reflink_inode(ip));
 
+	if (!xfs_can_free_cowblocks(ip))
+		return 0;
+
 	error = xfs_reflink_inode_has_shared_extents(*tpp, ip, &needs_flag);
 	if (error || needs_flag)
 		return error;
fs/xfs/xfs_reflink.h +19 -0
index fb55e4ce49fa..4a58e4533671 100644
@@ -6,6 +6,25 @@
 #ifndef __XFS_REFLINK_H
 #define __XFS_REFLINK_H 1
 
+/*
+ * Check whether it is safe to free COW fork blocks from an inode. It is unsafe
+ * to do so when an inode has dirty cache or I/O in-flight, even if no shared
+ * extents exist in the data fork, because outstanding I/O may target blocks
+ * that were speculatively allocated to the COW fork.
+ */
+static inline bool
+xfs_can_free_cowblocks(struct xfs_inode *ip)
+{
+	struct inode *inode = VFS_I(ip);
+
+	if ((inode->i_state & I_DIRTY_PAGES) ||
+	    mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY) ||
+	    mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
+	    atomic_read(&inode->i_dio_count))
+		return false;
+	return true;
+}
+
 extern int xfs_reflink_trim_around_shared(struct xfs_inode *ip,
 		struct xfs_bmbt_irec *irec, bool *shared);
 int xfs_bmap_trim_cow(struct xfs_inode *ip, struct xfs_bmbt_irec *imap,
Signature · 1 line
2.45.0
Reviewed-by
#2 lore raw

Re: [PATCH 2/1] xfs: don't free cowblocks from under dirty pagecache on unshare

Quoted 39 lines · Brian Foster

fallocate unshare mode explicitly breaks extent sharing. When a command completes, it checks the data fork for any remaining shared extents to determine whether the reflink inode flag and COW fork preallocation can be removed. This logic doesn't consider in-core pagecache and I/O state, however, which means we can unsafely remove COW fork blocks that are still needed under certain conditions.

For example, consider the following command sequence:

xfs_io -fc "pwrite 0 1k" -c "reflink <file> 0 256k 1k" \
	-c "pwrite 0 32k" -c "funshare 0 1k" <file>

This allocates a data block at offset 0, shares it, and then overwrites it with a larger buffered write. The overwrite triggers COW fork preallocation, 32 blocks by default, which maps the entire 32k write to delalloc in the COW fork. All but the shared block at offset 0 remains hole mapped in the data fork. The unshare command redirties and flushes the folio at offset 0, removing the only shared extent from the inode. Since the inode no longer maps shared extents, unshare purges the COW fork before the remaining 28k may have written back.

This leaves dirty pagecache backed by holes, which writeback quietly skips, thus leaving clean, non-zeroed pagecache over holes in the file. To verify, fiemap shows holes in the first 32k of the file and reads return different data across a remount:

$ xfs_io -c "fiemap -v" <file>
<file>:
 EXT: FILE-OFFSET      BLOCK-RANGE      TOTAL FLAGS
   ...
   1: [8..511]:        hole               504
   ...
$ xfs_io -c "pread -v 4k 8" <file>
00001000:  cd cd cd cd cd cd cd cd  ........
$ umount <mnt>; mount <dev> <mnt>
$ xfs_io -c "pread -v 4k 8" <file>
00001000:  00 00 00 00 00 00 00 00  ........
Quoted 6 lines

To avoid this problem, make unshare follow the same rules used for background cowblock scanning and never purge the COW fork for inodes with dirty pagecache or in-flight I/O.

Fixes 46afb0628b ("xfs: only flush the unshared range in xfs_reflink_unshare")

Question: Does xfs_repair report orphaned cow staging blocks after this? There's a longstanding bug that I've seen in the long soak xfs/286 VM where we slowly leak cow fork blocks (~80 per ~1 billion fsxops over 7 days).

Anyhow this looks correct on its own so

--D

Quoted 86 lines

---

Here's another COW issue I came across via some unshare testing. A quick hack to enable unshare in fsx uncovered it. I'll follow up with a proper patch for that.

I'm sending this as a 2/1 here just to reflect patch order in my local tree. Also note that I haven't explicitly tested the fixes commit, but a quick test to switch back to the old full flush behavior on latest master also makes the problem go away, so I suspect that's where the regression was introduced.

Brian

 fs/xfs/xfs_icache.c  |  8 +-------
 fs/xfs/xfs_reflink.c |  3 +++
 fs/xfs/xfs_reflink.h | 19 +++++++++++++++++++
 3 files changed, 23 insertions(+), 7 deletions(-)
fs/xfs/xfs_icache.c +1 -7
index 900a6277d931..a1b34e6ccfe2 100644
@@ -1278,13 +1278,7 @@ xfs_prep_free_cowblocks(
 	 */
 	if (!sync && inode_is_open_for_write(VFS_I(ip)))
 		return false;
-	if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
-	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
-	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
-	    atomic_read(&VFS_I(ip)->i_dio_count))
-		return false;
-
-	return true;
+	return xfs_can_free_cowblocks(ip);
 }
 
 /*
fs/xfs/xfs_reflink.c +3 -0
index 6fde6ec8092f..5bf6682e701b 100644
@@ -1595,6 +1595,9 @@ xfs_reflink_clear_inode_flag(
 
 	ASSERT(xfs_is_reflink_inode(ip));
 
+	if (!xfs_can_free_cowblocks(ip))
+		return 0;
+
 	error = xfs_reflink_inode_has_shared_extents(*tpp, ip, &needs_flag);
 	if (error || needs_flag)
 		return error;
fs/xfs/xfs_reflink.h +19 -0
index fb55e4ce49fa..4a58e4533671 100644
@@ -6,6 +6,25 @@
 #ifndef __XFS_REFLINK_H
 #define __XFS_REFLINK_H 1
 
+/*
+ * Check whether it is safe to free COW fork blocks from an inode. It is unsafe
+ * to do so when an inode has dirty cache or I/O in-flight, even if no shared
+ * extents exist in the data fork, because outstanding I/O may target blocks
+ * that were speculatively allocated to the COW fork.
+ */
+static inline bool
+xfs_can_free_cowblocks(struct xfs_inode *ip)
+{
+	struct inode *inode = VFS_I(ip);
+
+	if ((inode->i_state & I_DIRTY_PAGES) ||
+	    mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY) ||
+	    mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
+	    atomic_read(&inode->i_dio_count))
+		return false;
+	return true;
+}
+
 extern int xfs_reflink_trim_around_shared(struct xfs_inode *ip,
 		struct xfs_bmbt_irec *irec, bool *shared);
 int xfs_bmap_trim_cow(struct xfs_inode *ip, struct xfs_bmbt_irec *imap,
Signature · 1 line
2.45.0
#3 lore raw

Re: [PATCH 2/1] xfs: don't free cowblocks from under dirty pagecache on unshare

Quoted 44 lines · Darrick J. Wong
Quoted 43 lines · Brian Foster

fallocate unshare mode explicitly breaks extent sharing. When a command completes, it checks the data fork for any remaining shared extents to determine whether the reflink inode flag and COW fork preallocation can be removed. This logic doesn't consider in-core pagecache and I/O state, however, which means we can unsafely remove COW fork blocks that are still needed under certain conditions.

For example, consider the following command sequence:

xfs_io -fc "pwrite 0 1k" -c "reflink <file> 0 256k 1k" \
	-c "pwrite 0 32k" -c "funshare 0 1k" <file>

This allocates a data block at offset 0, shares it, and then overwrites it with a larger buffered write. The overwrite triggers COW fork preallocation, 32 blocks by default, which maps the entire 32k write to delalloc in the COW fork. All but the shared block at offset 0 remains hole mapped in the data fork. The unshare command redirties and flushes the folio at offset 0, removing the only shared extent from the inode. Since the inode no longer maps shared extents, unshare purges the COW fork before the remaining 28k may have written back.

This leaves dirty pagecache backed by holes, which writeback quietly skips, thus leaving clean, non-zeroed pagecache over holes in the file. To verify, fiemap shows holes in the first 32k of the file and reads return different data across a remount:

$ xfs_io -c "fiemap -v" <file>
<file>:
 EXT: FILE-OFFSET      BLOCK-RANGE      TOTAL FLAGS
   ...
   1: [8..511]:        hole               504
   ...
$ xfs_io -c "pread -v 4k 8" <file>
00001000:  cd cd cd cd cd cd cd cd  ........
$ umount <mnt>; mount <dev> <mnt>
$ xfs_io -c "pread -v 4k 8" <file>
00001000:  00 00 00 00 00 00 00 00  ........

To avoid this problem, make unshare follow the same rules used for background cowblock scanning and never purge the COW fork for inodes with dirty pagecache or in-flight I/O.

Quoted 8 lines
Quoted 2 lines
Fixes 46afb0628b ("xfs: only flush the unshared range in xfs_reflink_unshare")

Question: Does xfs_repair report orphaned cow staging blocks after this? There's a longstanding bug that I've seen in the long soak xfs/286 VM where we slowly leak cow fork blocks (~80 per ~1 billion fsxops over 7 days).

I've not seen that, at least in the test case I have. I think what's happening here is more that we clean up the COW fork correctly from an accounting standpoint, but we do so prematurely because the pagecache is dirty in ranges that are still only backed by COW fork blocks.

Quoted 3 lines

Anyhow this looks correct on its own so

Thanks!

Brian

Quoted 89 lines

--D

Quoted 86 lines

---

Here's another COW issue I came across via some unshare testing. A quick hack to enable unshare in fsx uncovered it. I'll follow up with a proper patch for that.

I'm sending this as a 2/1 here just to reflect patch order in my local tree. Also note that I haven't explicitly tested the fixes commit, but a quick test to switch back to the old full flush behavior on latest master also makes the problem go away, so I suspect that's where the regression was introduced.

Brian

 fs/xfs/xfs_icache.c  |  8 +-------
 fs/xfs/xfs_reflink.c |  3 +++
 fs/xfs/xfs_reflink.h | 19 +++++++++++++++++++
 3 files changed, 23 insertions(+), 7 deletions(-)
fs/xfs/xfs_icache.c +1 -7
index 900a6277d931..a1b34e6ccfe2 100644
@@ -1278,13 +1278,7 @@ xfs_prep_free_cowblocks(
 	 */
 	if (!sync && inode_is_open_for_write(VFS_I(ip)))
 		return false;
-	if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
-	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
-	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
-	    atomic_read(&VFS_I(ip)->i_dio_count))
-		return false;
-
-	return true;
+	return xfs_can_free_cowblocks(ip);
 }
 
 /*
fs/xfs/xfs_reflink.c +3 -0
index 6fde6ec8092f..5bf6682e701b 100644
@@ -1595,6 +1595,9 @@ xfs_reflink_clear_inode_flag(
 
 	ASSERT(xfs_is_reflink_inode(ip));
 
+	if (!xfs_can_free_cowblocks(ip))
+		return 0;
+
 	error = xfs_reflink_inode_has_shared_extents(*tpp, ip, &needs_flag);
 	if (error || needs_flag)
 		return error;
fs/xfs/xfs_reflink.h +19 -0
index fb55e4ce49fa..4a58e4533671 100644
@@ -6,6 +6,25 @@
 #ifndef __XFS_REFLINK_H
 #define __XFS_REFLINK_H 1
 
+/*
+ * Check whether it is safe to free COW fork blocks from an inode. It is unsafe
+ * to do so when an inode has dirty cache or I/O in-flight, even if no shared
+ * extents exist in the data fork, because outstanding I/O may target blocks
+ * that were speculatively allocated to the COW fork.
+ */
+static inline bool
+xfs_can_free_cowblocks(struct xfs_inode *ip)
+{
+	struct inode *inode = VFS_I(ip);
+
+	if ((inode->i_state & I_DIRTY_PAGES) ||
+	    mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY) ||
+	    mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
+	    atomic_read(&inode->i_dio_count))
+		return false;
+	return true;
+}
+
 extern int xfs_reflink_trim_around_shared(struct xfs_inode *ip,
 		struct xfs_bmbt_irec *irec, bool *shared);
 int xfs_bmap_trim_cow(struct xfs_inode *ip, struct xfs_bmbt_irec *imap,
Signature · 1 line
2.45.0
Reviewed-by on v2 1/1
#4 lore raw
In reply to Brian Foster
Quoted 85 lines · Brian Foster

The background blockgc scanner runs on a 5m interval by default and trims preallocation (post-eof and cow fork) from inodes that are otherwise idle. Idle effectively means that iolock can be acquired without blocking and that the inode has no dirty pagecache or I/O in flight.

This simple mechanism and heuristic has worked fairly well for post-eof speculative preallocations. Support for reflink and COW fork preallocations came sometime later and plugged into the same mechanism, with similar heuristics. Some recent testing has shown that COW fork preallocation may be notably more sensitive to blockgc processing than post-eof preallocation, however.

For example, consider an 8GB reflinked file with a COW extent size hint of 1MB. A worst case fully randomized overwrite of this file results in ~8k extents of an average size of ~1MB. If the same workload is interrupted a couple times for blockgc processing (assuming the file goes idle), the resulting extent count explodes to over 100k extents with an average size <100kB. This is significantly worse than ideal and essentially defeats the COW extent size hint mechanism.

While this particular test is instrumented, it reflects a fairly reasonable pattern in practice where random I/Os might spread out over a large period of time with varying periods of (in)activity. For example, consider a cloned disk image file for a VM or container with long uptime and variable and bursty usage. A background blockgc scan that races and processes the image file when it happens to be clean and idle can have a significant effect on the future fragmentation level of the file, even when still in use.

To help combat this, update the heuristic to skip cowblocks inodes that are currently opened for write access during non-sync blockgc scans. This allows COW fork preallocations to persist for as long as possible unless otherwise needed for functional purposes (i.e. a sync scan), the file is idle and closed, or the inode is being evicted from cache. While here, update the comments to help distinguish performance oriented heuristics from the logic that exists to maintain functional correctness.

---

v2:
- Reorder logic and update comments in xfs_prep_free_cowblocks().
v1: https://lore.kernel.org/linux-xfs/[email protected]/

 fs/xfs/xfs_icache.c | 31 +++++++++++++++++++++++--------
 1 file changed, 23 insertions(+), 8 deletions(-)
fs/xfs/xfs_icache.c +9 -6
index cf629302d48e..900a6277d931 100644
@@ -1241,14 +1241,17 @@ xfs_inode_clear_eofblocks_tag(
 }
 
 /*
- * Set ourselves up to free CoW blocks from this file.  If it's already clean
- * then we can bail out quickly, but otherwise we must back off if the file
- * is undergoing some kind of write.
+ * Prepare to free COW fork blocks from an inode.
  */
 static bool
 xfs_prep_free_cowblocks(
-	struct xfs_inode	*ip)
+	struct xfs_inode	*ip,
+	struct xfs_icwalk	*icw)
 {
+	bool			sync;
+
+	sync = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
+
 	/*
 	 * Just clear the tag if we have an empty cow fork or none at all. It's
 	 * possible the inode was fully unshared since it was originally tagged.
@@ -1260,9 +1263,21 @@ xfs_prep_free_cowblocks(
 	}
 
 	/*
-	 * If the mapping is dirty or under writeback we cannot touch the
-	 * CoW fork.  Leave it alone if we're in the midst of a directio.
+	 * A cowblocks trim of an inode can have a significant effect on
+	 * fragmentation even when a reasonable COW extent size hint is set.

fs/xfs/xfs_icache.c xfs_prep_free_cowblocks( · line 1277

Quoted 10 lines
fs/xfs/xfs_icache.c +10 -0
+	 * Therefore, we prefer to not process cowblocks unless they are clean
+	 * and idle. We can never process a cowblocks inode that is dirty or has
+	 * in-flight I/O under any circumstances, because outstanding writeback
+	 * or dio expects targeted COW fork blocks exist through write
+	 * completion where they can be remapped into the data fork.
+	 *
+	 * Therefore, the heuristic used here is to never process inodes
+	 * currently opened for write from background (i.e. non-sync) scans. For
+	 * sync scans, use the pagecache/dio state of the inode to ensure we
+	 * never free COW fork blocks out from under pending I/O.

Sounds good to me!

--D

Quoted 28 lines
fs/xfs/xfs_icache.c +4 -2
 	 */
+	if (!sync && inode_is_open_for_write(VFS_I(ip)))
+		return false;
 	if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
 	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
 	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
@@ -1298,7 +1313,7 @@ xfs_inode_free_cowblocks(
 	if (!xfs_iflags_test(ip, XFS_ICOWBLOCKS))
 		return 0;
 
-	if (!xfs_prep_free_cowblocks(ip))
+	if (!xfs_prep_free_cowblocks(ip, icw))
 		return 0;
 
 	if (!xfs_icwalk_match(ip, icw))
@@ -1327,7 +1342,7 @@ xfs_inode_free_cowblocks(
 	 * Check again, nobody else should be able to dirty blocks or change
 	 * the reflink iflag now that we have the first two locks held.
 	 */
-	if (xfs_prep_free_cowblocks(ip))
+	if (xfs_prep_free_cowblocks(ip, icw))
 		ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);
 	return ret;
 }
Signature · 1 line
2.45.0
#5 lore raw
In reply to Brian Foster
Quoted 5 lines · Brian Foster

The background blockgc scanner runs on a 5m interval by default and trims preallocation (post-eof and cow fork) from inodes that are otherwise idle. Idle effectively means that iolock can be acquired without blocking and that the inode has no dirty pagecache or I/O in flight.

Hey, can we get this data corruption fix staged for 6.12-rc2, please? Yesterday's fstests push contained the exerciser for this bug.

--D

Quoted 117 lines

This simple mechanism and heuristic has worked fairly well for post-eof speculative preallocations. Support for reflink and COW fork preallocations came sometime later and plugged into the same mechanism, with similar heuristics. Some recent testing has shown that COW fork preallocation may be notably more sensitive to blockgc processing than post-eof preallocation, however.

For example, consider an 8GB reflinked file with a COW extent size hint of 1MB. A worst case fully randomized overwrite of this file results in ~8k extents of an average size of ~1MB. If the same workload is interrupted a couple times for blockgc processing (assuming the file goes idle), the resulting extent count explodes to over 100k extents with an average size <100kB. This is significantly worse than ideal and essentially defeats the COW extent size hint mechanism.

While this particular test is instrumented, it reflects a fairly reasonable pattern in practice where random I/Os might spread out over a large period of time with varying periods of (in)activity. For example, consider a cloned disk image file for a VM or container with long uptime and variable and bursty usage. A background blockgc scan that races and processes the image file when it happens to be clean and idle can have a significant effect on the future fragmentation level of the file, even when still in use.

To help combat this, update the heuristic to skip cowblocks inodes that are currently opened for write access during non-sync blockgc scans. This allows COW fork preallocations to persist for as long as possible unless otherwise needed for functional purposes (i.e. a sync scan), the file is idle and closed, or the inode is being evicted from cache. While here, update the comments to help distinguish performance oriented heuristics from the logic that exists to maintain functional correctness.

---

v2:
- Reorder logic and update comments in xfs_prep_free_cowblocks().
v1: https://lore.kernel.org/linux-xfs/[email protected]/

 fs/xfs/xfs_icache.c | 31 +++++++++++++++++++++++--------
 1 file changed, 23 insertions(+), 8 deletions(-)
fs/xfs/xfs_icache.c +23 -8
index cf629302d48e..900a6277d931 100644
@@ -1241,14 +1241,17 @@ xfs_inode_clear_eofblocks_tag(
 }
 
 /*
- * Set ourselves up to free CoW blocks from this file.  If it's already clean
- * then we can bail out quickly, but otherwise we must back off if the file
- * is undergoing some kind of write.
+ * Prepare to free COW fork blocks from an inode.
  */
 static bool
 xfs_prep_free_cowblocks(
-	struct xfs_inode	*ip)
+	struct xfs_inode	*ip,
+	struct xfs_icwalk	*icw)
 {
+	bool			sync;
+
+	sync = icw && (icw->icw_flags & XFS_ICWALK_FLAG_SYNC);
+
 	/*
 	 * Just clear the tag if we have an empty cow fork or none at all. It's
 	 * possible the inode was fully unshared since it was originally tagged.
@@ -1260,9 +1263,21 @@ xfs_prep_free_cowblocks(
 	}
 
 	/*
-	 * If the mapping is dirty or under writeback we cannot touch the
-	 * CoW fork.  Leave it alone if we're in the midst of a directio.
+	 * A cowblocks trim of an inode can have a significant effect on
+	 * fragmentation even when a reasonable COW extent size hint is set.
+	 * Therefore, we prefer to not process cowblocks unless they are clean
+	 * and idle. We can never process a cowblocks inode that is dirty or has
+	 * in-flight I/O under any circumstances, because outstanding writeback
+	 * or dio expects targeted COW fork blocks exist through write
+	 * completion where they can be remapped into the data fork.
+	 *
+	 * Therefore, the heuristic used here is to never process inodes
+	 * currently opened for write from background (i.e. non-sync) scans. For
+	 * sync scans, use the pagecache/dio state of the inode to ensure we
+	 * never free COW fork blocks out from under pending I/O.
 	 */
+	if (!sync && inode_is_open_for_write(VFS_I(ip)))
+		return false;
 	if ((VFS_I(ip)->i_state & I_DIRTY_PAGES) ||
 	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_DIRTY) ||
 	    mapping_tagged(VFS_I(ip)->i_mapping, PAGECACHE_TAG_WRITEBACK) ||
@@ -1298,7 +1313,7 @@ xfs_inode_free_cowblocks(
 	if (!xfs_iflags_test(ip, XFS_ICOWBLOCKS))
 		return 0;
 
-	if (!xfs_prep_free_cowblocks(ip))
+	if (!xfs_prep_free_cowblocks(ip, icw))
 		return 0;
 
 	if (!xfs_icwalk_match(ip, icw))
@@ -1327,7 +1342,7 @@ xfs_inode_free_cowblocks(
 	 * Check again, nobody else should be able to dirty blocks or change
 	 * the reflink iflag now that we have the first two locks held.
 	 */
-	if (xfs_prep_free_cowblocks(ip))
+	if (xfs_prep_free_cowblocks(ip, icw))
 		ret = xfs_reflink_cancel_cow_range(ip, 0, NULLFILEOFF, false);
 	return ret;
 }
Signature · 1 line
2.45.0
maintainer Applied
#6 lore raw
In reply to Brian Foster
Quoted 6 lines · Brian Foster

The background blockgc scanner runs on a 5m interval by default and trims preallocation (post-eof and cow fork) from inodes that are otherwise idle. Idle effectively means that iolock can be acquired without blocking and that the inode has no dirty pagecache or I/O in flight.

Quoted 8 lines

This simple mechanism and heuristic has worked fairly well for post-eof speculative preallocations. Support for reflink and COW fork preallocations came sometime later and plugged into the same mechanism, with similar heuristics. Some recent testing has shown that COW fork preallocation may be notably more sensitive to blockgc processing than post-eof preallocation, however.

[...]

Applied to for-next, thanks!

[1/1] xfs: skip background cowblock trims on inodes open for write

      commit: 90a71daaf73f5d39bb0cbb3c7ab6af942fe6233e
[2/2] xfs: don't free cowblocks from under dirty pagecache on unshare
      commit: 4390f019ad7866c3791c3d768d2ff185d89e8ebe

Best regards,

Signature · 1 line
Carlos Maiolino <[email protected]>