Skip to contentSkip to conversation

exfat: fix iomap mapping vs truncate race

PATCH Applied Youling Tang 3 participants 5 messages Nov 16 – Nov 17

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

v2 cover exfat: fix iomap mapping vs truncate race

From: Youling Tang <[email protected]>

This series fixes two related problems in the exfat iomap path.

Patch 1 fixes a race between iomap mapping and truncate: i_size was read twice, and a shrink in between could make the mapping length negative, which underflowed the cluster count.

Patch 2 fixes the handling of the EXFAT_EOF_CLUSTER sentinel. The iomap code used to convert it into a physical address; now exfat_map_cluster() maps only the valid prefix when a range straddles the end of the chain, and iomap reports the remaining range as a hole, matching the old buffer_head code.

Patch 2 relies on the non-negative length guaranteed by patch 1 for the hole mapping.

Changes since v1:
- Move the valid-prefix clamp into exfat_map_cluster().

v1: https://lore.kernel.org/exfat/[email protected]/T/#t

Youling Tang (2):
exfat: fix iomap mapping when writeback races with truncate
exfat: don't map EXFAT_EOF_CLUSTER as a physical cluster

 fs/exfat/inode.c | 15 ++++++++++++---
 fs/exfat/iomap.c | 21 +++++++++++++++++----
 2 files changed, 29 insertions(+), 7 deletions(-)
Signature · 1 line
2.43.0
#1 lore raw

PATCH v2 1/2 exfat: fix iomap mapping when writeback races with truncate

From: Youling Tang <[email protected]>

__exfat_iomap_begin() reads i_size twice: once to detect a range that starts beyond EOF, and once to clamp the mapping to EOF. A concurrent truncate can shrink i_size between the two reads, so the clamp can produce a negative length. The unsigned cluster index subtraction then underflows, making exfat_map_cluster() believe it has to allocate a huge number of clusters and report a hole.

Fix this by reading i_size once for both the EOF check and the block-aligned clamp.

Fixes 82a81a7352bc ("exfat: add iomap buffered I/O support")
---
 fs/exfat/iomap.c | 9 +++++----
 1 file changed, 5 insertions(+), 4 deletions(-)
fs/exfat/iomap.c +5 -4
index 4e6f3ed9dcc8..11823591b3a8 100644
@@ -51,8 +51,10 @@ static int __exfat_iomap_begin(struct inode *inode, loff_t offset, loff_t length
 	bool balloc = false;
 
 	if (!may_alloc) {
+		loff_t isize = i_size_read(inode);
+
 		/* Completely beyond EOF. Treat as hole */
-		if (i_size_read(inode) <= offset) {
+		if (offset >= isize) {
 			iomap->type = IOMAP_HOLE;
 			iomap->addr = IOMAP_NULL_ADDR;
 			iomap->offset = offset;
@@ -61,9 +63,8 @@ static int __exfat_iomap_begin(struct inode *inode, loff_t offset, loff_t length
 		}
 
 		/* Clamp length if the requested range goes beyond i_size */
-		if (offset + length > i_size_read(inode))
-			length = round_up(i_size_read(inode),
-					  i_blocksize(inode)) - offset;
+		if (offset + length > isize)
+			length = round_up(isize, i_blocksize(inode)) - offset;
 	}
 
 	num_clusters = exfat_bytes_to_cluster_round_up(sbi,
Signature · 1 line
2.43.0
#2 lore raw

PATCH v2 2/2 exfat: don't map EXFAT_EOF_CLUSTER as a physical cluster

In reply to Youling Tang

From: Youling Tang <[email protected]>

exfat_map_cluster() reports a hole by returning success with *clu == EXFAT_EOF_CLUSTER for !create. __exfat_iomap_begin() does not check that sentinel and converts it into a physical address far beyond the device. A writeback that races with the i_size/i_blocks window in exfat_cont_expand() can hit this: the writeback submits a bio past the end of the device, and fsync() or msync() returns -EIO. If the resulting iomap length is zero, iomap_add_to_ioend() can loop on a zero-length bio and hang a concurrent truncate.

When a range starts before the end of the cluster chain but extends past it, exfat_map_cluster() returns the sentinel for the whole range. Make it map the valid prefix in that case, and return the sentinel only if the whole range is beyond the chain:

    offset         chain end       offset+len
      |--- prefix ----|--- tail ---|
      |== allocated ==|
            MAPPED         HOLE

Reporting the whole straddling range as a hole instead would skip valid dirty data.

Report the remaining range as a hole instead of turning the sentinel into an address, matching the old exfat_get_block() behavior.

Fixes 82a81a7352bc ("exfat: add iomap buffered I/O support")

---
v2: move the clamp into exfat_map_cluster().

 fs/exfat/inode.c | 15 ++++++++++++---
 fs/exfat/iomap.c | 12 ++++++++++++
 2 files changed, 24 insertions(+), 3 deletions(-)
fs/exfat/inode.c +12 -3
index c1d15d4fc5e7..b175d2f3e459 100644
@@ -134,6 +134,9 @@ void exfat_sync_inode(struct inode *inode)
  * Input: inode, (logical) clu_offset, target allocation area
  * Output: errcode, cluster number
  * *clu = (~0), if it's unable to allocate a new cluster
+ *
+ * For !create, *count is limited to the part of the range backed by the
+ * cluster chain; the caller reports the remainder as a hole.
  */
 int exfat_map_cluster(struct inode *inode, unsigned int clu_offset,
 		unsigned int *clu, unsigned int *count, int create,
@@ -153,9 +156,15 @@ int exfat_map_cluster(struct inode *inode, unsigned int clu_offset,
 	    *count > num_clusters - clu_offset)
 		num_to_be_allocated = clu_offset + *count - num_clusters;
 
-	if (!create && (num_to_be_allocated > 0)) {
-		*clu = EXFAT_EOF_CLUSTER;
-		return 0;
+	if (!create && num_to_be_allocated > 0) {
+		if (clu_offset >= num_clusters) {
+			*clu = EXFAT_EOF_CLUSTER;
+			return 0;
+		}
+
+		/* Range straddles the chain end: map the valid prefix only. */
+		*count = num_clusters - clu_offset;
+		num_to_be_allocated = 0;
 	}
 
 	*clu = last_clu = ei->start_clu;
fs/exfat/iomap.c +12 -0
index 11823591b3a8..ed17b367f820 100644
@@ -90,6 +90,18 @@ static int __exfat_iomap_begin(struct inode *inode, loff_t offset, loff_t length
 	if (err)
 		goto out;
 
+	/*
+	 * exfat_map_cluster() can return EXFAT_EOF_CLUSTER when racing with
+	 * exfat_cont_expand().  Report it as a hole.
+	 */
+	if (!may_alloc && cluster == EXFAT_EOF_CLUSTER) {
+		iomap->type = IOMAP_HOLE;
+		iomap->addr = IOMAP_NULL_ADDR;
+		iomap->offset = offset;
+		iomap->length = length;
+		goto out;
+	}
+
 	cluster_offset = exfat_cluster_offset(sbi, offset);
 	cluster_length = exfat_cluster_to_bytes(sbi, num_clusters);
 
Signature · 1 line
2.43.0
Reviewed-by on v2 2/2
#3 lore raw

Re: [PATCH v2 2/2] exfat: don't map EXFAT_EOF_CLUSTER as a physical cluster

Quoted 24 lines · Youling Tang

From: Youling Tang <[email protected]>

exfat_map_cluster() reports a hole by returning success with *clu == EXFAT_EOF_CLUSTER for !create. __exfat_iomap_begin() does not check that sentinel and converts it into a physical address far beyond the device. A writeback that races with the i_size/i_blocks window in exfat_cont_expand() can hit this: the writeback submits a bio past the end of the device, and fsync() or msync() returns -EIO. If the resulting iomap length is zero, iomap_add_to_ioend() can loop on a zero-length bio and hang a concurrent truncate.

When a range starts before the end of the cluster chain but extends past it, exfat_map_cluster() returns the sentinel for the whole range. Make it map the valid prefix in that case, and return the sentinel only if the whole range is beyond the chain:

     offset         chain end       offset+len
       |--- prefix ----|--- tail ---|
       |== allocated ==|
             MAPPED         HOLE

Reporting the whole straddling range as a hole instead would skip valid dirty data.

Quoted 6 lines

Report the remaining range as a hole instead of turning the sentinel into an address, matching the old exfat_get_block() behavior.

Fixes 82a81a7352bc ("exfat: add iomap buffered I/O support")

Looks good to me:

Quoted 63 lines

---
v2: move the clamp into exfat_map_cluster().

  fs/exfat/inode.c | 15 ++++++++++++---
  fs/exfat/iomap.c | 12 ++++++++++++
  2 files changed, 24 insertions(+), 3 deletions(-)
fs/exfat/inode.c +12 -3
index c1d15d4fc5e7..b175d2f3e459 100644
@@ -134,6 +134,9 @@ void exfat_sync_inode(struct inode *inode)
   * Input: inode, (logical) clu_offset, target allocation area
   * Output: errcode, cluster number
   * *clu = (~0), if it's unable to allocate a new cluster
+ *
+ * For !create, *count is limited to the part of the range backed by the
+ * cluster chain; the caller reports the remainder as a hole.
   */
  int exfat_map_cluster(struct inode *inode, unsigned int clu_offset,
  		unsigned int *clu, unsigned int *count, int create,
@@ -153,9 +156,15 @@ int exfat_map_cluster(struct inode *inode, unsigned int clu_offset,
  	    *count > num_clusters - clu_offset)
  		num_to_be_allocated = clu_offset + *count - num_clusters;
 
-	if (!create && (num_to_be_allocated > 0)) {
-		*clu = EXFAT_EOF_CLUSTER;
-		return 0;
+	if (!create && num_to_be_allocated > 0) {
+		if (clu_offset >= num_clusters) {
+			*clu = EXFAT_EOF_CLUSTER;
+			return 0;
+		}
+
+		/* Range straddles the chain end: map the valid prefix only. */
+		*count = num_clusters - clu_offset;
+		num_to_be_allocated = 0;
  	}
 
  	*clu = last_clu = ei->start_clu;
fs/exfat/iomap.c +12 -0
index 11823591b3a8..ed17b367f820 100644
@@ -90,6 +90,18 @@ static int __exfat_iomap_begin(struct inode *inode, loff_t offset, loff_t length
  	if (err)
  		goto out;
 
+	/*
+	 * exfat_map_cluster() can return EXFAT_EOF_CLUSTER when racing with
+	 * exfat_cont_expand().  Report it as a hole.
+	 */
+	if (!may_alloc && cluster == EXFAT_EOF_CLUSTER) {
+		iomap->type = IOMAP_HOLE;
+		iomap->addr = IOMAP_NULL_ADDR;
+		iomap->offset = offset;
+		iomap->length = length;
+		goto out;
+	}
+
  	cluster_offset = exfat_cluster_offset(sbi, offset);
  	cluster_length = exfat_cluster_to_bytes(sbi, num_clusters);
 
maintainer Applied
#4 lore raw
In reply to Youling Tang
Quoted 18 lines · Youling Tang

From: Youling Tang <[email protected]>

This series fixes two related problems in the exfat iomap path.

Patch 1 fixes a race between iomap mapping and truncate: i_size was read twice, and a shrink in between could make the mapping length negative, which underflowed the cluster count.

Patch 2 fixes the handling of the EXFAT_EOF_CLUSTER sentinel. The iomap code used to convert it into a physical address; now exfat_map_cluster() maps only the valid prefix when a range straddles the end of the chain, and iomap reports the remaining range as a hole, matching the old buffer_head code.

Patch 2 relies on the non-negative length guaranteed by patch 1 for the hole mapping.

Quoted 8 lines

Changes since v1:
- Move the valid-prefix clamp into exfat_map_cluster().

v1: https://lore.kernel.org/exfat/[email protected]/T/#t

Youling Tang (2):
exfat: fix iomap mapping when writeback races with truncate
exfat: don't map EXFAT_EOF_CLUSTER as a physical cluster

Applied them to #dev.
Thanks!