Skip to contentSkip to conversation

ovl: restore vma->vm_file to old file

PATCH In review Chengguang Xu 3 participants 6 messages Apr 20, 2021 – Apr 21, 2021 Overlay filesystem

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 ovl: restore vma->vm_file to old file

In the error case of ->mmap() we should also restore vma->vm_file to old file in order to keep correct file reference in error path.

---
 fs/overlayfs/file.c | 1 +
 1 file changed, 1 insertion(+)
fs/overlayfs/file.c +1 -0
index 6e454a294046..046a7adb02c5 100644
@@ -439,6 +439,7 @@ static int ovl_mmap(struct file *file, struct vm_area_struct *vma)
 	if (ret) {
 		/* Drop reference count from new vm_file value */
 		fput(realfile);
+		vma->vm_file = file;
 	} else {
 		/* Drop reference count from previous vm_file value */
 		fput(file);
Signature · 1 line
2.27.0
maintainer
#1 lore raw
Quoted 14 lines · Chengguang Xu

In the error case of ->mmap() we should also restore vma->vm_file to old file in order to keep correct file reference in error path.

---
 fs/overlayfs/file.c | 1 +
 1 file changed, 1 insertion(+)
fs/overlayfs/file.c +0 -0
index 6e454a294046..046a7adb02c5 100644

fs/overlayfs/file.c static int ovl_mmap(struct file *file, struct vm_area_struct *vma) · line 442

Quoted 4 lines
fs/overlayfs/file.c +1 -0
@@ -439,6 +439,7 @@ static int ovl_mmap(struct file *file, struct vm_area_struct *vma)
        if (ret) {
                /* Drop reference count from new vm_file value */
                fput(realfile);
+               vma->vm_file = file;

That's interesting: commit 1527f926fd04 ("mm: mmap: fix fput in error path v2") which went into 5.11-rc1 seems to have broke the refcounting
in overlayfs in the name of cleaning up a workaround. Wondering if
there's any other damage done by this "fix"?

Changing refcounting rules in core kernel is no easy matter, a full audit of ->mmap instances (>200) should have been done beforehand.

I suggest reverting this commit as a first step.

Thanks,
Miklos

#2 lore raw

Am 21.04.21 um 11:47 schrieb Miklos Szeredi:

Quoted 9 lines · Chengguang Xu
Quoted 8 lines

In the error case of ->mmap() we should also restore vma->vm_file to old file in order to keep correct file reference in error path.

---
  fs/overlayfs/file.c | 1 +
  1 file changed, 1 insertion(+)
Quoted 13 lines

fs/overlayfs/file.c static int ovl_mmap(struct file *file, struct vm_area_struct *vma) · line 442

Quoted 9 lines
fs/overlayfs/file.c +1 -0
index 6e454a294046..046a7adb02c5 100644
@@ -439,6 +439,7 @@ static int ovl_mmap(struct file *file, struct vm_area_struct *vma)
         if (ret) {
                 /* Drop reference count from new vm_file value */
                 fput(realfile);
+               vma->vm_file = file;

That's interesting: commit 1527f926fd04 ("mm: mmap: fix fput in error path v2") which went into 5.11-rc1 seems to have broke the refcounting
in overlayfs in the name of cleaning up a workaround. Wondering if
there's any other damage done by this "fix"?

Can you give wider context? In other words why did the patch broke the reference counting in overlayfs?

Quoted 2 lines

Changing refcounting rules in core kernel is no easy matter, a full audit of ->mmap instances (>200) should have been done beforehand.

Which is pretty much what was done, see the follow up commit:

commit 295992fb815e791d14b18ef7cdbbaf1a76211a31 (able/vma_file)
Author: Christian König <[email protected]> Date:   Mon Sep 14 15:09:33 2020 +0200

    mm: introduce vma_set_file function v5

    Add the new vma_set_file() function to allow changing vma->vm_file with the necessary refcount dance.

It just looks like I missed the case in overlayfs while doing this.

Sorry for the noise.

Regards,
Christian.

Quoted 5 lines

I suggest reverting this commit as a first step.

Thanks,
Miklos

maintainer
#3 lore raw
Quoted 25 lines · Christian König

Am 21.04.21 um 11:47 schrieb Miklos Szeredi:

Quoted 9 lines · Chengguang Xu
Quoted 8 lines

In the error case of ->mmap() we should also restore vma->vm_file to old file in order to keep correct file reference in error path.

---
  fs/overlayfs/file.c | 1 +
  1 file changed, 1 insertion(+)
Quoted 13 lines

fs/overlayfs/file.c static int ovl_mmap(struct file *file, struct vm_area_struct *vma) · line 442

Quoted 9 lines
fs/overlayfs/file.c +1 -0
index 6e454a294046..046a7adb02c5 100644
@@ -439,6 +439,7 @@ static int ovl_mmap(struct file *file, struct vm_area_struct *vma)
         if (ret) {
                 /* Drop reference count from new vm_file value */
                 fput(realfile);
+               vma->vm_file = file;

That's interesting: commit 1527f926fd04 ("mm: mmap: fix fput in error path v2") which went into 5.11-rc1 seems to have broke the refcounting
in overlayfs in the name of cleaning up a workaround. Wondering if
there's any other damage done by this "fix"?

Quoted 2 lines

Can you give wider context? In other words why did the patch broke the reference counting in overlayfs?

In the error case overlayfs would put the reference on realfile (which is vma->vm_file at that point) and mmap_region() would put the reference to the original file (which was vma->vm_file before being overridden).

After your commit mmap_region() puts the ref on the override vm_file, but not on the original file.

Quoted 10 lines
Quoted 3 lines

Changing refcounting rules in core kernel is no easy matter, a full audit of ->mmap instances (>200) should have been done beforehand.

Which is pretty much what was done, see the follow up commit:

commit 295992fb815e791d14b18ef7cdbbaf1a76211a31 (able/vma_file)

Author: Christian König <[email protected]>
Date:   Mon Sep 14 15:09:33 2020 +0200
Quoted 6 lines

mm: introduce vma_set_file function v5

Add the new vma_set_file() function to allow changing vma->vm_file with the necessary refcount dance.

It just looks like I missed the case in overlayfs while doing this.

Yes. And apparently a number of other cases where vm_file is assigned...

Thanks,
Miklos

#4 lore raw

Am 21.04.21 um 13:14 schrieb Miklos Szeredi:

Quoted 6 lines
Quoted 4 lines · Christian König

Am 21.04.21 um 11:47 schrieb Miklos Szeredi:
[SNIP]
Can you give wider context? In other words why did the patch broke the reference counting in overlayfs?

Quoted 7 lines

In the error case overlayfs would put the reference on realfile (which is vma->vm_file at that point) and mmap_region() would put the reference to the original file (which was vma->vm_file before being overridden).

After your commit mmap_region() puts the ref on the override vm_file, but not on the original file.

Ah, of course. Double checking the mmap callback implementation of overlayfs that is rather obvious.

Quoted 8 lines
Quoted 8 lines
Quoted 2 lines

Changing refcounting rules in core kernel is no easy matter, a full audit of ->mmap instances (>200) should have been done beforehand.

Which is pretty much what was done, see the follow up commit:

commit 295992fb815e791d14b18ef7cdbbaf1a76211a31 (able/vma_file)

Author: Christian König <[email protected]>
Date:   Mon Sep 14 15:09:33 2020 +0200
Quoted 7 lines
Quoted 6 lines

mm: introduce vma_set_file function v5

Add the new vma_set_file() function to allow changing vma->vm_file with the necessary refcount dance.

It just looks like I missed the case in overlayfs while doing this.

Yes. And apparently a number of other cases where vm_file is assigned...

Yeah, I wasn't aware that filesystems do that as well and only concentrated on the drivers.

Just did a "grep -R 'vm_file[[:space:]]*=' on the full kernel source and it only showed one more case in fs/coda/file.c.

Do you see any other occurrences I potentially missed?

Otherwise I would say I provide patches for those two cases in a minute.

Thanks,
Christian.

Quoted 3 lines

Thanks,
Miklos

maintainer
#5 lore raw
Quoted 35 lines · Christian König

Am 21.04.21 um 13:14 schrieb Miklos Szeredi:

Quoted 6 lines
Quoted 4 lines · Christian König

Am 21.04.21 um 11:47 schrieb Miklos Szeredi:
[SNIP]
Can you give wider context? In other words why did the patch broke the reference counting in overlayfs?

Quoted 7 lines

In the error case overlayfs would put the reference on realfile (which is vma->vm_file at that point) and mmap_region() would put the reference to the original file (which was vma->vm_file before being overridden).

After your commit mmap_region() puts the ref on the override vm_file, but not on the original file.

Ah, of course. Double checking the mmap callback implementation of overlayfs that is rather obvious.

Quoted 8 lines
Quoted 8 lines
Quoted 2 lines

Changing refcounting rules in core kernel is no easy matter, a full audit of ->mmap instances (>200) should have been done beforehand.

Which is pretty much what was done, see the follow up commit:

commit 295992fb815e791d14b18ef7cdbbaf1a76211a31 (able/vma_file)

Author: Christian König <[email protected]>
Date:   Mon Sep 14 15:09:33 2020 +0200
Quoted 7 lines
Quoted 6 lines

mm: introduce vma_set_file function v5

Add the new vma_set_file() function to allow changing vma->vm_file with the necessary refcount dance.

It just looks like I missed the case in overlayfs while doing this.

Yes. And apparently a number of other cases where vm_file is assigned...

Quoted 7 lines

Yeah, I wasn't aware that filesystems do that as well and only concentrated on the drivers.

Just did a "grep -R 'vm_file[[:space:]]*=' on the full kernel source and it only showed one more case in fs/coda/file.c.

Do you see any other occurrences I potentially missed?

No, the others seem to be okay.

Thanks,
Miklos