Skip to contentSkip to conversation

kernfs: fix false-positive WARN(nr_mmapped) in kernfs_drain_open_files

PATCH In review Neel Natu 2 participants 3 messages Jan 27, 2024 – Jan 29, 2024 kernfs

lore
Share & tools
lore

Neel Natu

author
#0 lore raw

PATCH 1/1 kernfs: fix false-positive WARN(nr_mmapped) in kernfs_drain_open_files

Prior to this change 'on->nr_mmapped' tracked the total number of mmaps across all of its associated open files via kernfs_fop_mmap(). Thus if the file descriptor associated with a kernfs_open_file was mmapped 10 times then we would have: 'of->mmapped = true' and 'of_on(of)->nr_mmapped = 10'.

The problem is that closing or draining a 'of->mmapped' file would only decrement one from the 'of_on(of)->nr_mmapped' counter.

For e.g. we have this from kernfs_unlink_open_file():

        if (of->mmapped)
                on->nr_mmapped--;

The WARN_ON_ONCE(on->nr_mmapped) in kernfs_drain_open_files() is easy to reproduce by:
1. opening a (mmap-able) kernfs file.
2. mmap-ing that file more than once (mapping just once masks the issue).
3. trigger a drain of that kernfs file.

Modulo out-of-tree patches I was able to trigger this reliably by identifying pci device nodes in sysfs that have resource regions that are mmap-able and that don't have any driver attached to them (steps 1 and 2). For step 3 we can "echo 1 > remove" to trigger a kernfs_drain.

---
 fs/kernfs/file.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
fs/kernfs/file.c +5 -3
index ffa4565c275a..e9df2f87072c 100644
@@ -483,9 +483,11 @@ static int kernfs_fop_mmap(struct file *file, struct vm_area_struct *vma)
 		goto out_put;
 
 	rc = 0;
-	of->mmapped = true;
-	of_on(of)->nr_mmapped++;
-	of->vm_ops = vma->vm_ops;
+	if (!of->mmapped) {
+		of->mmapped = true;
+		of_on(of)->nr_mmapped++;
+		of->vm_ops = vma->vm_ops;
+	}
 	vma->vm_ops = &kernfs_vm_ops;
 out_put:
 	kernfs_put_active(of->kn);
Signature · 1 line
2.43.0.429.g432eaa2c6b-goog
#1 lore raw
Quoted 39 lines · Neel Natu

Prior to this change 'on->nr_mmapped' tracked the total number of mmaps across all of its associated open files via kernfs_fop_mmap(). Thus if the file descriptor associated with a kernfs_open_file was mmapped 10 times then we would have: 'of->mmapped = true' and 'of_on(of)->nr_mmapped = 10'.

The problem is that closing or draining a 'of->mmapped' file would only decrement one from the 'of_on(of)->nr_mmapped' counter.

For e.g. we have this from kernfs_unlink_open_file():

        if (of->mmapped)
                on->nr_mmapped--;

The WARN_ON_ONCE(on->nr_mmapped) in kernfs_drain_open_files() is easy to reproduce by:
1. opening a (mmap-able) kernfs file.
2. mmap-ing that file more than once (mapping just once masks the issue).
3. trigger a drain of that kernfs file.

Modulo out-of-tree patches I was able to trigger this reliably by identifying pci device nodes in sysfs that have resource regions that are mmap-able and that don't have any driver attached to them (steps 1 and 2). For step 3 we can "echo 1 > remove" to trigger a kernfs_drain.

---
 fs/kernfs/file.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
fs/kernfs/file.c +0 -1
index ffa4565c275a..e9df2f87072c 100644
@@ -483,9 +483,11 @@ static int kernfs_fop_mmap(struct file *file, struct vm_area_struct *vma)
 		goto out_put;
 
 	rc = 0;
-	of->mmapped = true;

fs/kernfs/file.c static int kernfs_fop_mmap(struct file *file, struct vm_area_struct *vma) · line 493

Quoted 10 lines
fs/kernfs/file.c +5 -2
-	of_on(of)->nr_mmapped++;
-	of->vm_ops = vma->vm_ops;
+	if (!of->mmapped) {
+		of->mmapped = true;
+		of_on(of)->nr_mmapped++;
+		of->vm_ops = vma->vm_ops;
+	}
 	vma->vm_ops = &kernfs_vm_ops;
 out_put:
 	kernfs_put_active(of->kn);

What commit id does this fix?

thanks,

greg k-h

author
#2 lore raw

Hi,

Quoted 39 lines · Greg Kroah-Hartman
Quoted 37 lines · Neel Natu

Prior to this change 'on->nr_mmapped' tracked the total number of mmaps across all of its associated open files via kernfs_fop_mmap(). Thus if the file descriptor associated with a kernfs_open_file was mmapped 10 times then we would have: 'of->mmapped = true' and 'of_on(of)->nr_mmapped = 10'.

The problem is that closing or draining a 'of->mmapped' file would only decrement one from the 'of_on(of)->nr_mmapped' counter.

For e.g. we have this from kernfs_unlink_open_file():

        if (of->mmapped)
                on->nr_mmapped--;

The WARN_ON_ONCE(on->nr_mmapped) in kernfs_drain_open_files() is easy to reproduce by:
1. opening a (mmap-able) kernfs file.
2. mmap-ing that file more than once (mapping just once masks the issue).
3. trigger a drain of that kernfs file.

Modulo out-of-tree patches I was able to trigger this reliably by identifying pci device nodes in sysfs that have resource regions that are mmap-able and that don't have any driver attached to them (steps 1 and 2). For step 3 we can "echo 1 > remove" to trigger a kernfs_drain.

---
 fs/kernfs/file.c | 8 +++++---
 1 file changed, 5 insertions(+), 3 deletions(-)
fs/kernfs/file.c +0 -0
index ffa4565c275a..e9df2f87072c 100644
@@ -483,9 +483,11 @@ static int kernfs_fop_mmap(struct file *file, struct vm_area_struct *vma)
              goto out_put;
 
Quoted 14 lines

fs/kernfs/file.c static int kernfs_fop_mmap(struct file *file, struct vm_area_struct *vma) · line 493

Quoted 12 lines
fs/kernfs/file.c +5 -3
      rc = 0;
-     of->mmapped = true;
-     of_on(of)->nr_mmapped++;
-     of->vm_ops = vma->vm_ops;
+     if (!of->mmapped) {
+             of->mmapped = true;
+             of_on(of)->nr_mmapped++;
+             of->vm_ops = vma->vm_ops;
+     }
      vma->vm_ops = &kernfs_vm_ops;
 out_put:
      kernfs_put_active(of->kn);

What commit id does this fix?

bdb2fd7fc56e197a63c0b0e7e07d25d5e20e7c72
kernfs: Skip kernfs_drain_open_files() more aggressively

best
Neel

Quoted 4 lines

thanks,

greg k-h