Skip to contentSkip to conversation

xfs: fix an agbno overflow in __xfs_getfsmap_datadev

PATCH In review Darrick J. Wong 3 participants 10 messages Aug 23, 2023 – Aug 28, 2023 XFS filesystem

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 xfs: fix an agbno overflow in __xfs_getfsmap_datadev

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

Dave Chinner reported that xfs/273 fails if the AG size happens to be an exact power of two. I traced this to an agbno integer overflow when the current GETFSMAP call is a continuation of a previous GETFSMAP call, and the last record returned was non-shareable space at the end of an AG.

__xfs_getfsmap_datadev sets up a data device query by converting the incoming fmr_physical into an xfs_fsblock_t and cracking it into an agno and agbno pair. In the (failing) case of where fmr_blockcount of the low key is nonzero and the record was for a non-shareable extent, it will add fmr_blockcount to start_fsb and info->low.rm_startblock.

If the low key was actually the last record for that AG, then this addition causes info->low.rm_startblock to point beyond EOAG. When the rmapbt range query starts, it'll return an empty set, and fsmap moves on to the next AG.

Or so I thought. Remember how we added to start_fsb?

If agsize < 1<<agblklog, start_fsb points to the same AG as the original fmr_physical from the low key. We run the rmapbt query, which returns nothing, so getfsmap zeroes info->low and moves on to the next AG.

If agsize == 1<<agblklog, start_fsb now points to the next AG. We run the rmapbt query on the next AG with the excessively large rm_startblock. If this next AG is actually the last AG, we'll set info->high to EOFS (which is now has a lower rm_startblock than info->low), and the ranged btree query code will return -EINVAL. If it's not the last AG, we ignore all records for the intermediate AGs.

Oops.

Fix this by decoding start_fsb into agno and agbno only after making adjustments to start_fsb. This means that info->low.rm_startblock will always be set to a valid agbno, and we always start the rmapbt iteration in the correct AG.

While we're at it, fix the predicate for determining if an fsmap record represents non-shareable space to include file data on pre-reflink filesystems.

Fixes 63ef7a35912dd ("xfs: fix interval filtering in multi-step fsmap queries")
---
 fs/xfs/xfs_fsmap.c |   25 ++++++++++++++++++-------
 1 file changed, 18 insertions(+), 7 deletions(-)
fs/xfs/xfs_fsmap.c +18 -7
index 10403ba9b58f..736e5545f584 100644
@@ -565,6 +565,19 @@ xfs_getfsmap_rtdev_rtbitmap(
 }
 #endif /* CONFIG_XFS_RT */
 
+static inline bool
+rmap_not_shareable(struct xfs_mount *mp, const struct xfs_rmap_irec *r)
+{
+	if (!xfs_has_reflink(mp))
+		return true;
+	if (XFS_RMAP_NON_INODE_OWNER(r->rm_owner))
+		return true;
+	if (r->rm_flags & (XFS_RMAP_ATTR_FORK | XFS_RMAP_BMBT_BLOCK |
+			   XFS_RMAP_UNWRITTEN))
+		return true;
+	return false;
+}
+
 /* Execute a getfsmap query against the regular data device. */
 STATIC int
 __xfs_getfsmap_datadev(
@@ -598,7 +611,6 @@ __xfs_getfsmap_datadev(
 	 * low to the fsmap low key and max out the high key to the end
 	 * of the AG.
 	 */
-	info->low.rm_startblock = XFS_FSB_TO_AGBNO(mp, start_fsb);
 	info->low.rm_offset = XFS_BB_TO_FSBT(mp, keys[0].fmr_offset);
 	error = xfs_fsmap_owner_to_rmap(&info->low, &keys[0]);
 	if (error)
@@ -608,12 +620,9 @@ __xfs_getfsmap_datadev(
 
 	/* Adjust the low key if we are continuing from where we left off. */
 	if (info->low.rm_blockcount == 0) {
-		/* empty */
-	} else if (XFS_RMAP_NON_INODE_OWNER(info->low.rm_owner) ||
-		   (info->low.rm_flags & (XFS_RMAP_ATTR_FORK |
-					  XFS_RMAP_BMBT_BLOCK |
-					  XFS_RMAP_UNWRITTEN))) {
-		info->low.rm_startblock += info->low.rm_blockcount;
+		/* No previous record from which to continue */
+	} else if (rmap_not_shareable(mp, &info->low)) {
+		/* Last record seen was an unshareable extent */
 		info->low.rm_owner = 0;
 		info->low.rm_offset = 0;
 
@@ -621,8 +630,10 @@ __xfs_getfsmap_datadev(
 		if (XFS_FSB_TO_DADDR(mp, start_fsb) >= eofs)
 			return 0;
 	} else {
+		/* Last record seen was a shareable file data extent */
 		info->low.rm_offset += info->low.rm_blockcount;
 	}
+	info->low.rm_startblock = XFS_FSB_TO_AGBNO(mp, start_fsb);
 
 	info->high.rm_startblock = -1U;
 	info->high.rm_owner = ULLONG_MAX;
#1 lore raw

RFC PATCH 1/1 fstests: test fix for an agbno overflow in __xfs_getfsmap_datadev

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

Dave Chinner reported that xfs/273 fails if the AG size happens to be an exact power of two. I traced this to an agbno integer overflow when the current GETFSMAP call is a continuation of a previous GETFSMAP call, and the last record returned was non-shareable space at the end of an AG.

This is the regression test for that bug.

---
 tests/xfs/935     |   55 +++++++++++++++++++++++++++++++++++++++++++++++++++++
 tests/xfs/935.out |    2 ++
 2 files changed, 57 insertions(+)
 create mode 100755 tests/xfs/935
 create mode 100644 tests/xfs/935.out
tests/xfs/935 new file +55 -0
new file mode 100755
index 0000000000..a06f2fc8dc
@@ -0,0 +1,55 @@
+#! /bin/bash
+# SPDX-License-Identifier: GPL-2.0
+# Copyright (c) 2023 Oracle.  All Rights Reserved.
+#
+# FS QA Test 935
+#
+# Regression test for an agbno overflow bug in XFS GETFSMAP involving an
+# fsmap_advance call.  Userspace can indicate that a GETFSMAP call is actually
+# a continuation of a previous call by setting the "low" key to the last record
+# returned by the previous call.
+#
+# If the last record returned by GETFSMAP is a non-shareable extent at the end
+# of an AG and the AG size is exactly a power of two, the startblock in the low
+# key of the rmapbt query can be set to a value larger than EOAG.  When this
+# happens, GETFSMAP will return EINVAL instead of returning records for the
+# next AG.
+#
+. ./common/preamble
+_begin_fstest auto quick fsmap
+
+. ./common/filter
+
+_fixed_by_git_commit kernel XXXXXXXXXXXXX \
+	"xfs: fix an agbno overflow in __xfs_getfsmap_datadev"
+
+# Modify as appropriate.
+_supported_fs generic
+_require_xfs_io_command fsmap
+_require_xfs_scratch_rmapbt
+
+_scratch_mkfs | _filter_mkfs 2> $tmp.mkfs >> $seqres.full
+source $tmp.mkfs
+
+# Find the next power of two agsize smaller than whatever the default is.
+for ((p = 31; p > 0; p--)); do
+	desired_agsize=$((2 ** p))
+	test "$desired_agsize" -lt "$agsize" && break
+done
+
+echo "desired asize=$desired_agsize" >> $seqres.full
+_scratch_mkfs -d "agsize=${desired_agsize}b" | _filter_mkfs 2> $tmp.mkfs >> $seqres.full
+source $tmp.mkfs
+
+test "$desired_agsize" -eq "$agsize" || _notrun "wanted agsize=$desired_agsize, got $agsize"
+
+_scratch_mount
+$XFS_IO_PROG -c 'fsmap -n 1024 -v' $SCRATCH_MNT >> $tmp.big
+$XFS_IO_PROG -c 'fsmap -n 1 -v' $SCRATCH_MNT >> $tmp.small
+
+diff -Naurpw $tmp.big $tmp.small
+
+# success, all done
+echo Silence is golden
+status=0
+exit
tests/xfs/935.out new file +2 -0
new file mode 100644
index 0000000000..1b5422d1e3
@@ -0,0 +1,2 @@
+QA output created by 935
+Silence is golden
Reviewed-by on v1 1/1
#2 lore raw

Re: [RFC PATCH] fstests: test fix for an agbno overflow in __xfs_getfsmap_datadev

Quoted 53 lines · Darrick J. Wong

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

Dave Chinner reported that xfs/273 fails if the AG size happens to be an exact power of two. I traced this to an agbno integer overflow when the current GETFSMAP call is a continuation of a previous GETFSMAP call, and the last record returned was non-shareable space at the end of an AG.

This is the regression test for that bug.

---
 tests/xfs/935     |   55 +++++++++++++++++++++++++++++++++++++++++++++++++++++
 tests/xfs/935.out |    2 ++
 2 files changed, 57 insertions(+)
 create mode 100755 tests/xfs/935
 create mode 100644 tests/xfs/935.out
tests/xfs/935 new file +30 -0
new file mode 100755
index 0000000000..a06f2fc8dc
@@ -0,0 +1,55 @@
+#! /bin/bash
+# SPDX-License-Identifier: GPL-2.0
+# Copyright (c) 2023 Oracle.  All Rights Reserved.
+#
+# FS QA Test 935
+#
+# Regression test for an agbno overflow bug in XFS GETFSMAP involving an
+# fsmap_advance call.  Userspace can indicate that a GETFSMAP call is actually
+# a continuation of a previous call by setting the "low" key to the last record
+# returned by the previous call.
+#
+# If the last record returned by GETFSMAP is a non-shareable extent at the end
+# of an AG and the AG size is exactly a power of two, the startblock in the low
+# key of the rmapbt query can be set to a value larger than EOAG.  When this
+# happens, GETFSMAP will return EINVAL instead of returning records for the
+# next AG.
+#
+. ./common/preamble
+_begin_fstest auto quick fsmap
+
+. ./common/filter
+
+_fixed_by_git_commit kernel XXXXXXXXXXXXX \
+	"xfs: fix an agbno overflow in __xfs_getfsmap_datadev"
+
+# Modify as appropriate.
+_supported_fs generic
+_require_xfs_io_command fsmap
+_require_xfs_scratch_rmapbt
+

tests/xfs/935 line 40

Quoted 10 lines
tests/xfs/935 new file +10 -0
+_scratch_mkfs | _filter_mkfs 2> $tmp.mkfs >> $seqres.full
+source $tmp.mkfs
+
+# Find the next power of two agsize smaller than whatever the default is.
+for ((p = 31; p > 0; p--)); do
+	desired_agsize=$((2 ** p))
+	test "$desired_agsize" -lt "$agsize" && break
+done
+
+echo "desired asize=$desired_agsize" >> $seqres.full

agsize

Otherwise looks fine.

Signature · 2 lines
Dave Chinner
[email protected]
#3 lore raw

Re: [RFC PATCH] fstests: test fix for an agbno overflow in __xfs_getfsmap_datadev

Quoted 57 lines · Dave Chinner
Quoted 56 lines · Darrick J. Wong

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

Dave Chinner reported that xfs/273 fails if the AG size happens to be an exact power of two. I traced this to an agbno integer overflow when the current GETFSMAP call is a continuation of a previous GETFSMAP call, and the last record returned was non-shareable space at the end of an AG.

This is the regression test for that bug.

---
 tests/xfs/935     |   55 +++++++++++++++++++++++++++++++++++++++++++++++++++++
 tests/xfs/935.out |    2 ++
 2 files changed, 57 insertions(+)
 create mode 100755 tests/xfs/935
 create mode 100644 tests/xfs/935.out
tests/xfs/935 new file +33 -0
new file mode 100755
index 0000000000..a06f2fc8dc
@@ -0,0 +1,55 @@
+#! /bin/bash
+# SPDX-License-Identifier: GPL-2.0
+# Copyright (c) 2023 Oracle.  All Rights Reserved.
+#
+# FS QA Test 935
+#
+# Regression test for an agbno overflow bug in XFS GETFSMAP involving an
+# fsmap_advance call.  Userspace can indicate that a GETFSMAP call is actually
+# a continuation of a previous call by setting the "low" key to the last record
+# returned by the previous call.
+#
+# If the last record returned by GETFSMAP is a non-shareable extent at the end
+# of an AG and the AG size is exactly a power of two, the startblock in the low
+# key of the rmapbt query can be set to a value larger than EOAG.  When this
+# happens, GETFSMAP will return EINVAL instead of returning records for the
+# next AG.
+#
+. ./common/preamble
+_begin_fstest auto quick fsmap
+
+. ./common/filter
+
+_fixed_by_git_commit kernel XXXXXXXXXXXXX \
+	"xfs: fix an agbno overflow in __xfs_getfsmap_datadev"
+
+# Modify as appropriate.
+_supported_fs generic
+_require_xfs_io_command fsmap
+_require_xfs_scratch_rmapbt
+
+_scratch_mkfs | _filter_mkfs 2> $tmp.mkfs >> $seqres.full
+source $tmp.mkfs
+
Quoted 8 lines

tests/xfs/935 line 40

Quoted 7 lines
tests/xfs/935 new file +7 -0
+# Find the next power of two agsize smaller than whatever the default is.
+for ((p = 31; p > 0; p--)); do
+	desired_agsize=$((2 ** p))
+	test "$desired_agsize" -lt "$agsize" && break
+done
+
+echo "desired asize=$desired_agsize" >> $seqres.full

agsize

Fixed.

Quoted 3 lines

Otherwise looks fine.

Does the kernel patch fix the bug on your end too?

--D

Quoted 3 lines
Signature · 2 lines
Dave Chinner
[email protected]
#4 lore raw

Re: [RFC PATCH] fstests: test fix for an agbno overflow in __xfs_getfsmap_datadev

Quoted 67 lines · Darrick J. Wong
Quoted 57 lines · Dave Chinner
Quoted 56 lines · Darrick J. Wong

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

Dave Chinner reported that xfs/273 fails if the AG size happens to be an exact power of two. I traced this to an agbno integer overflow when the current GETFSMAP call is a continuation of a previous GETFSMAP call, and the last record returned was non-shareable space at the end of an AG.

This is the regression test for that bug.

---
 tests/xfs/935     |   55 +++++++++++++++++++++++++++++++++++++++++++++++++++++
 tests/xfs/935.out |    2 ++
 2 files changed, 57 insertions(+)
 create mode 100755 tests/xfs/935
 create mode 100644 tests/xfs/935.out
tests/xfs/935 new file +33 -0
new file mode 100755
index 0000000000..a06f2fc8dc
@@ -0,0 +1,55 @@
+#! /bin/bash
+# SPDX-License-Identifier: GPL-2.0
+# Copyright (c) 2023 Oracle.  All Rights Reserved.
+#
+# FS QA Test 935
+#
+# Regression test for an agbno overflow bug in XFS GETFSMAP involving an
+# fsmap_advance call.  Userspace can indicate that a GETFSMAP call is actually
+# a continuation of a previous call by setting the "low" key to the last record
+# returned by the previous call.
+#
+# If the last record returned by GETFSMAP is a non-shareable extent at the end
+# of an AG and the AG size is exactly a power of two, the startblock in the low
+# key of the rmapbt query can be set to a value larger than EOAG.  When this
+# happens, GETFSMAP will return EINVAL instead of returning records for the
+# next AG.
+#
+. ./common/preamble
+_begin_fstest auto quick fsmap
+
+. ./common/filter
+
+_fixed_by_git_commit kernel XXXXXXXXXXXXX \
+	"xfs: fix an agbno overflow in __xfs_getfsmap_datadev"
+
+# Modify as appropriate.
+_supported_fs generic
+_require_xfs_io_command fsmap
+_require_xfs_scratch_rmapbt
+
+_scratch_mkfs | _filter_mkfs 2> $tmp.mkfs >> $seqres.full
+source $tmp.mkfs
+
Quoted 8 lines

tests/xfs/935 line 40

Quoted 7 lines
tests/xfs/935 new file +7 -0
+# Find the next power of two agsize smaller than whatever the default is.
+for ((p = 31; p > 0; p--)); do
+	desired_agsize=$((2 ** p))
+	test "$desired_agsize" -lt "$agsize" && break
+done
+
+echo "desired asize=$desired_agsize" >> $seqres.full

agsize

Quoted 7 lines

Fixed.

Quoted 3 lines

Otherwise looks fine.

Does the kernel patch fix the bug on your end too?

Haven't had a chance to test it yet. I'll let you know when I do.

-Dave.

Signature · 2 lines
Dave Chinner
[email protected]
#5 lore raw

Re: [RFC PATCH] fstests: test fix for an agbno overflow in __xfs_getfsmap_datadev

In reply to Darrick J. Wong
Quoted 61 lines · Darrick J. Wong

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

Dave Chinner reported that xfs/273 fails if the AG size happens to be an exact power of two. I traced this to an agbno integer overflow when the current GETFSMAP call is a continuation of a previous GETFSMAP call, and the last record returned was non-shareable space at the end of an AG.

This is the regression test for that bug.

---
 tests/xfs/935     |   55 +++++++++++++++++++++++++++++++++++++++++++++++++++++
 tests/xfs/935.out |    2 ++
 2 files changed, 57 insertions(+)
 create mode 100755 tests/xfs/935
 create mode 100644 tests/xfs/935.out
tests/xfs/935 new file +38 -0
new file mode 100755
index 0000000000..a06f2fc8dc
@@ -0,0 +1,55 @@
+#! /bin/bash
+# SPDX-License-Identifier: GPL-2.0
+# Copyright (c) 2023 Oracle.  All Rights Reserved.
+#
+# FS QA Test 935
+#
+# Regression test for an agbno overflow bug in XFS GETFSMAP involving an
+# fsmap_advance call.  Userspace can indicate that a GETFSMAP call is actually
+# a continuation of a previous call by setting the "low" key to the last record
+# returned by the previous call.
+#
+# If the last record returned by GETFSMAP is a non-shareable extent at the end
+# of an AG and the AG size is exactly a power of two, the startblock in the low
+# key of the rmapbt query can be set to a value larger than EOAG.  When this
+# happens, GETFSMAP will return EINVAL instead of returning records for the
+# next AG.
+#
+. ./common/preamble
+_begin_fstest auto quick fsmap
+
+. ./common/filter
+
+_fixed_by_git_commit kernel XXXXXXXXXXXXX \
+	"xfs: fix an agbno overflow in __xfs_getfsmap_datadev"
+
+# Modify as appropriate.
+_supported_fs generic
+_require_xfs_io_command fsmap
+_require_xfs_scratch_rmapbt
+
+_scratch_mkfs | _filter_mkfs 2> $tmp.mkfs >> $seqres.full
+source $tmp.mkfs
+
+# Find the next power of two agsize smaller than whatever the default is.
+for ((p = 31; p > 0; p--)); do
+	desired_agsize=$((2 ** p))
+	test "$desired_agsize" -lt "$agsize" && break
+done

tests/xfs/935 line 48

Quoted 10 lines
tests/xfs/935 new file +10 -0
+
+echo "desired asize=$desired_agsize" >> $seqres.full
+_scratch_mkfs -d "agsize=${desired_agsize}b" | _filter_mkfs 2> $tmp.mkfs >> $seqres.full
+source $tmp.mkfs
+
+test "$desired_agsize" -eq "$agsize" || _notrun "wanted agsize=$desired_agsize, got $agsize"
+
+_scratch_mount
+$XFS_IO_PROG -c 'fsmap -n 1024 -v' $SCRATCH_MNT >> $tmp.big
+$XFS_IO_PROG -c 'fsmap -n 1 -v' $SCRATCH_MNT >> $tmp.small

This line reports:

xfs_io: xfsctl(XFS_IOC_GETFSMAP) iflags=0x0 ["/mnt/xfstests/scratch"]: Invalid argument

when the test case fails. Is that normal?

Quoted 16 lines
tests/xfs/935 new file +7 -0
+
+diff -Naurpw $tmp.big $tmp.small
+
+# success, all done
+echo Silence is golden
+status=0
+exit
tests/xfs/935.out new file +2 -0
new file mode 100644
index 0000000000..1b5422d1e3
@@ -0,0 +1,2 @@
+QA output created by 935
+Silence is golden
#6 lore raw

Re: [RFC PATCH] fstests: test fix for an agbno overflow in __xfs_getfsmap_datadev

Quoted 75 lines · Zorro Lang
Quoted 61 lines · Darrick J. Wong

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

Dave Chinner reported that xfs/273 fails if the AG size happens to be an exact power of two. I traced this to an agbno integer overflow when the current GETFSMAP call is a continuation of a previous GETFSMAP call, and the last record returned was non-shareable space at the end of an AG.

This is the regression test for that bug.

---
 tests/xfs/935     |   55 +++++++++++++++++++++++++++++++++++++++++++++++++++++
 tests/xfs/935.out |    2 ++
 2 files changed, 57 insertions(+)
 create mode 100755 tests/xfs/935
 create mode 100644 tests/xfs/935.out
tests/xfs/935 new file +38 -0
new file mode 100755
index 0000000000..a06f2fc8dc
@@ -0,0 +1,55 @@
+#! /bin/bash
+# SPDX-License-Identifier: GPL-2.0
+# Copyright (c) 2023 Oracle.  All Rights Reserved.
+#
+# FS QA Test 935
+#
+# Regression test for an agbno overflow bug in XFS GETFSMAP involving an
+# fsmap_advance call.  Userspace can indicate that a GETFSMAP call is actually
+# a continuation of a previous call by setting the "low" key to the last record
+# returned by the previous call.
+#
+# If the last record returned by GETFSMAP is a non-shareable extent at the end
+# of an AG and the AG size is exactly a power of two, the startblock in the low
+# key of the rmapbt query can be set to a value larger than EOAG.  When this
+# happens, GETFSMAP will return EINVAL instead of returning records for the
+# next AG.
+#
+. ./common/preamble
+_begin_fstest auto quick fsmap
+
+. ./common/filter
+
+_fixed_by_git_commit kernel XXXXXXXXXXXXX \
+	"xfs: fix an agbno overflow in __xfs_getfsmap_datadev"
+
+# Modify as appropriate.
+_supported_fs generic
+_require_xfs_io_command fsmap
+_require_xfs_scratch_rmapbt
+
+_scratch_mkfs | _filter_mkfs 2> $tmp.mkfs >> $seqres.full
+source $tmp.mkfs
+
+# Find the next power of two agsize smaller than whatever the default is.
+for ((p = 31; p > 0; p--)); do
+	desired_agsize=$((2 ** p))
+	test "$desired_agsize" -lt "$agsize" && break
+done

tests/xfs/935 line 48

Quoted 10 lines
tests/xfs/935 new file +10 -0
+
+echo "desired asize=$desired_agsize" >> $seqres.full
+_scratch_mkfs -d "agsize=${desired_agsize}b" | _filter_mkfs 2> $tmp.mkfs >> $seqres.full
+source $tmp.mkfs
+
+test "$desired_agsize" -eq "$agsize" || _notrun "wanted agsize=$desired_agsize, got $agsize"
+
+_scratch_mount
+$XFS_IO_PROG -c 'fsmap -n 1024 -v' $SCRATCH_MNT >> $tmp.big
+$XFS_IO_PROG -c 'fsmap -n 1 -v' $SCRATCH_MNT >> $tmp.small

This line reports:

Quoted 3 lines

xfs_io: xfsctl(XFS_IOC_GETFSMAP) iflags=0x0 ["/mnt/xfstests/scratch"]: Invalid argument

when the test case fails. Is that normal?

Yes. The attached bugfix should make that go away.

--D

Quoted 17 lines
Quoted 16 lines
tests/xfs/935 new file +7 -0
+
+diff -Naurpw $tmp.big $tmp.small
+
+# success, all done
+echo Silence is golden
+status=0
+exit
tests/xfs/935.out new file +2 -0
new file mode 100644
index 0000000000..1b5422d1e3
@@ -0,0 +1,2 @@
+QA output created by 935
+Silence is golden
#7 lore raw

Re: [RFC PATCH] fstests: test fix for an agbno overflow in __xfs_getfsmap_datadev

Quoted 12 lines · Darrick J. Wong
Quoted 11 lines · Zorro Lang
Quoted 10 lines · Darrick J. Wong

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

Dave Chinner reported that xfs/273 fails if the AG size happens to be an exact power of two. I traced this to an agbno integer overflow when the current GETFSMAP call is a continuation of a previous GETFSMAP call, and the last record returned was non-shareable space at the end of an AG.

This is the regression test for that bug.

.....

Quoted 12 lines
Quoted 12 lines

tests/xfs/935 line 48

Quoted 9 lines
tests/xfs/935 new file +9 -0
+echo "desired asize=$desired_agsize" >> $seqres.full
+_scratch_mkfs -d "agsize=${desired_agsize}b" | _filter_mkfs 2> $tmp.mkfs >> $seqres.full
+source $tmp.mkfs
+
+test "$desired_agsize" -eq "$agsize" || _notrun "wanted agsize=$desired_agsize, got $agsize"
+
+_scratch_mount
+$XFS_IO_PROG -c 'fsmap -n 1024 -v' $SCRATCH_MNT >> $tmp.big
+$XFS_IO_PROG -c 'fsmap -n 1 -v' $SCRATCH_MNT >> $tmp.small

This line reports:

Quoted 5 lines
Quoted 3 lines

xfs_io: xfsctl(XFS_IOC_GETFSMAP) iflags=0x0 ["/mnt/xfstests/scratch"]: Invalid argument

when the test case fails. Is that normal?

Yes. The attached bugfix should make that go away.

The kernel bug fix fixes the same problem with xfs/273; I haven't tested this specific new regression test.

-Dave.

Signature · 2 lines
Dave Chinner
[email protected]
#8 lore raw

Re: [RFC PATCH] fstests: test fix for an agbno overflow in __xfs_getfsmap_datadev

Quoted 26 lines · Dave Chinner
Quoted 12 lines · Darrick J. Wong
Quoted 11 lines · Zorro Lang
Quoted 10 lines · Darrick J. Wong

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

Dave Chinner reported that xfs/273 fails if the AG size happens to be an exact power of two. I traced this to an agbno integer overflow when the current GETFSMAP call is a continuation of a previous GETFSMAP call, and the last record returned was non-shareable space at the end of an AG.

This is the regression test for that bug.

.....

Quoted 12 lines
Quoted 12 lines

tests/xfs/935 line 48

Quoted 9 lines
tests/xfs/935 new file +9 -0
+echo "desired asize=$desired_agsize" >> $seqres.full
+_scratch_mkfs -d "agsize=${desired_agsize}b" | _filter_mkfs 2> $tmp.mkfs >> $seqres.full
+source $tmp.mkfs
+
+test "$desired_agsize" -eq "$agsize" || _notrun "wanted agsize=$desired_agsize, got $agsize"
+
+_scratch_mount
+$XFS_IO_PROG -c 'fsmap -n 1024 -v' $SCRATCH_MNT >> $tmp.big
+$XFS_IO_PROG -c 'fsmap -n 1 -v' $SCRATCH_MNT >> $tmp.small

This line reports:

Quoted 8 lines
Quoted 5 lines
Quoted 3 lines

xfs_io: xfsctl(XFS_IOC_GETFSMAP) iflags=0x0 ["/mnt/xfstests/scratch"]: Invalid argument

when the test case fails. Is that normal?

Yes. The attached bugfix should make that go away.

The kernel bug fix fixes the same problem with xfs/273; I haven't tested this specific new regression test.

Thanks for the details from both of you, I'll merge this patch in next fstests release if there's not more updates.

Thanks,
Zorro

Quoted 6 lines

-Dave.

Signature · 2 lines
Dave Chinner
[email protected]
Reviewed-by on v1 1/1
#9 lore raw
In reply to Darrick J. Wong
Quoted 38 lines · Darrick J. Wong

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

Dave Chinner reported that xfs/273 fails if the AG size happens to be an exact power of two. I traced this to an agbno integer overflow when the current GETFSMAP call is a continuation of a previous GETFSMAP call, and the last record returned was non-shareable space at the end of an AG.

__xfs_getfsmap_datadev sets up a data device query by converting the incoming fmr_physical into an xfs_fsblock_t and cracking it into an agno and agbno pair. In the (failing) case of where fmr_blockcount of the low key is nonzero and the record was for a non-shareable extent, it will add fmr_blockcount to start_fsb and info->low.rm_startblock.

If the low key was actually the last record for that AG, then this addition causes info->low.rm_startblock to point beyond EOAG. When the rmapbt range query starts, it'll return an empty set, and fsmap moves on to the next AG.

Or so I thought. Remember how we added to start_fsb?

If agsize < 1<<agblklog, start_fsb points to the same AG as the original fmr_physical from the low key. We run the rmapbt query, which returns nothing, so getfsmap zeroes info->low and moves on to the next AG.

If agsize == 1<<agblklog, start_fsb now points to the next AG. We run the rmapbt query on the next AG with the excessively large rm_startblock. If this next AG is actually the last AG, we'll set info->high to EOFS (which is now has a lower rm_startblock than info->low), and the ranged btree query code will return -EINVAL. If it's not the last AG, we ignore all records for the intermediate AGs.

Oops.

Fix this by decoding start_fsb into agno and agbno only after making adjustments to start_fsb. This means that info->low.rm_startblock will always be set to a valid agbno, and we always start the rmapbt iteration in the correct AG.

Quoted 7 lines

While we're at it, fix the predicate for determining if an fsmap record represents non-shareable space to include file data on pre-reflink filesystems.

Fixes 63ef7a35912dd ("xfs: fix interval filtering in multi-step fsmap queries")

Fixes the regression, code looks fine.

Signature · 2 lines
Dave Chinner
[email protected]