Skip to contentSkip to conversation

Fix NULL pointer dereference in read_cache_folio

PATCH Superseded Gianfranco Trad 3 participants 6 messages Sep 29, 2024 – Oct 15, 2024

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 Fix NULL pointer dereference in read_cache_folio

Add check on filler to prevent NULL pointer dereference condition in read_cache_folio[1].

[1] https://syzkaller.appspot.com/bug?extid=4089e577072948ac5531

---
 mm/filemap.c | 5 ++++-
 1 file changed, 4 insertions(+), 1 deletion(-)
mm/filemap.c +4 -1
index 4f3753f0a158..960f389e2d3b 100644
@@ -2360,7 +2360,10 @@ static int filemap_read_folio(struct file *file, filler_t filler,
 	/* Start the actual read. The read will unlock the page. */
 	if (unlikely(workingset))
 		psi_memstall_enter(&pflags);
-	error = filler(file, folio);
+	if (filler)
+		error = filler(file, folio);
+	else
+		return -EIO;
 	if (unlikely(workingset))
 		psi_memstall_leave(&pflags);
 	if (error)
Signature · 1 line
2.43.0
#1 lore raw

PATCH v2 1/1 Fix NULL pointer dereference in read_cache_folio

Add check on filler to prevent NULL pointer dereference condition in read_cache_folio[1].

[1] https://syzkaller.appspot.com/bug?extid=4089e577072948ac5531

---

Notes:
changes in v2:
- refactored check on filler.

 mm/filemap.c | 2 ++
 1 file changed, 2 insertions(+)
mm/filemap.c +2 -0
index 4f3753f0a158..88de8029133c 100644
@@ -2360,6 +2360,8 @@ static int filemap_read_folio(struct file *file, filler_t filler,
 	/* Start the actual read. The read will unlock the page. */
 	if (unlikely(workingset))
 		psi_memstall_enter(&pflags);
+	if (!filler)
+		return -EIO;
 	error = filler(file, folio);
 	if (unlikely(workingset))
 		psi_memstall_leave(&pflags);
Signature · 1 line
2.43.0
maintainer
#2 lore raw
Quoted 4 lines · On Mon, 30 Sep 2024 11:02:26 +0200 Gianfranco Trad <[email protected]> wrote:

Add check on filler to prevent NULL pointer dereference condition in read_cache_folio[1].

[1] https://syzkaller.appspot.com/bug?extid=4089e577072948ac5531

Test case https://syzkaller.appspot.com/x/repro.c?x=10a0d880580000

mm/filemap.c static int filemap_read_folio(struct file *file, filler_t filler, · line 2367

Quoted 13 lines
mm/filemap.c +2 -0
index 4f3753f0a158..88de8029133c 100644
@@ -2360,6 +2360,8 @@ static int filemap_read_folio(struct file *file, filler_t filler,
 	/* Start the actual read. The read will unlock the page. */
 	if (unlikely(workingset))
 		psi_memstall_enter(&pflags);
+	if (!filler)
+		return -EIO;
 	error = filler(file, folio);
 	if (unlikely(workingset))
 		psi_memstall_leave(&pflags);

do_read_cache_folio() already handles a NULL filler (from freader_get_folio()'s read_cache_folio() call).

	if (!filler)
		filler = mapping->a_ops->read_folio;

so I'm suspecting that an appropriate fix is to teach the underlying address_space_operations (appears to be from /proc/pid/maps) to implement ->read_folio().

maintainer
#3 lore raw
In reply to Gianfranco Trad

mm/filemap.c static int filemap_read_folio(struct file *file, filler_t filler, · line 2364

Quoted 6 lines · Gianfranco Trad
mm/filemap.c +2 -0
@@ -2360,6 +2360,8 @@ static int filemap_read_folio(struct file *file, filler_t filler,
 	/* Start the actual read. The read will unlock the page. */
 	if (unlikely(workingset))
 		psi_memstall_enter(&pflags);
+	if (!filler)
+		return -EIO;

This is definitely wrong because you enter memstall, but do not exit it.

As Andrew says, the underlying problem is that the filesystem does not implement ->read_folio. Which filesystem is this?

Quoted 6 lines
mm/filemap.c +0 -0
 	error = filler(file, folio);
 	if (unlikely(workingset))
 		psi_memstall_leave(&pflags);
Signature · 1 line
2.43.0
#4 lore raw
Quoted 9 lines · Matthew Wilcox

mm/filemap.c static int filemap_read_folio(struct file *file, filler_t filler, · line 2364

Quoted 6 lines · Gianfranco Trad
mm/filemap.c +2 -0
@@ -2360,6 +2360,8 @@ static int filemap_read_folio(struct file *file, filler_t filler,
  	/* Start the actual read. The read will unlock the page. */
  	if (unlikely(workingset))
  		psi_memstall_enter(&pflags);
+	if (!filler)
+		return -EIO;

This is definitely wrong because you enter memstall, but do not exit it.

Got it, thanks.

Quoted 3 lines

As Andrew says, the underlying problem is that the filesystem does not implement ->read_folio. Which filesystem is this?

Reproducer via procfs accesses a bpf map backed by an anonymous inode (anon_inode_fs_type), with mapping->a_ops pointing to anon_aops, hence, read_folio() undefined.

Quoted 7 lines
Quoted 7 lines
mm/filemap.c +0 -0
  	error = filler(file, folio);
  	if (unlikely(workingset))
  		psi_memstall_leave(&pflags);
Signature · 1 line
2.43.0

I suppose the next step would be to contact the proper maintainers(?) If you have any additional suggestions, I'd be more than glad to listen.

Thanks to both of you for your time,

--Gian

#5 lore raw

Re: [PATCH v2] mm: fix null pointer dereference in read_cache_folio

Quoted 30 lines · Gianfranco Trad
Quoted 10 lines · Matthew Wilcox

Review static int filemap_read_folio(struct file · line 2365

Quoted 7 lines · Gianfranco Trad
(quoted diff) +2 -0
@@ -2360,6 +2360,8 @@ static int filemap_read_folio(struct file 
 *file, filler_t filler,
       /* Start the actual read. The read will unlock the page. */
       if (unlikely(workingset))
           psi_memstall_enter(&pflags);
+    if (!filler)
+        return -EIO;

This is definitely wrong because you enter memstall, but do not exit it.

Got it, thanks.

Quoted 3 lines

As Andrew says, the underlying problem is that the filesystem does not implement ->read_folio.  Which filesystem is this?

Reproducer via procfs accesses a bpf map backed by an anonymous inode (anon_inode_fs_type), with mapping->a_ops pointing to anon_aops, hence, read_folio() undefined.

Quoted 7 lines
Quoted 7 lines
(quoted diff) +0 -0
       error = filler(file, folio);
       if (unlikely(workingset))

          psi_memstall_leave(&pflags);

Signature · 1 line
2.43.0
Quoted 7 lines

I suppose the next step would be to contact the proper maintainers(?) If you have any additional suggestions, I'd be more than glad to listen.

Thanks to both of you for your time,

--Gian

Hello,

While studying how to implement read_folio in anon_aops for this specific case (bpf map backed by anon_inode_fs_type) I've come up with an intermediate solution that mitigates the null pointer dereference and avoids the memstall issue (compared to my previous patch) immediately, for all filesystems that do not implement read_folio in their address_space_operations.

The patch [1] looks like this:

mm/filemap.c +2 -0
index 4f3753f0a158..680d98086c00 100644
@@ -3775,6 +3775,8 @@ static struct folio *do_read_cache_folio(struct address_space *mapping,
 	struct folio *folio;
 	int err;
 
+	if (!filler && (!mapping->a_ops || !mapping->a_ops->read_folio))
+		return ERR_PTR(-ENOSYS);
 	if (!filler)
 		filler = mapping->a_ops->read_folio;
 repeat:

Patch was already tested with syzbot on the same reproducer case. Reproducer did not trigger any issue [2].

Let me know if for now this patch looks good enough, therefore I'll send it to you, or if I should work on it more.

Thanks for your time,

[1] https://syzkaller.appspot.com/text?tag=Patch&x=142e045f980000
[2] https://syzkaller.appspot.com/x/log.txt?x=1551045f980000

-- Gian