Skip to contentSkip to conversation

Btrfs: fix regression in btrfs_page_mkwrite() from vm_fault_t conversion

PATCH Superseded Chris Mason 2 participants 3 messages Jun 25, 2018 Btrfs file system

v1 of 2 Superseded Latest is v2 →

lore
Share & tools
lore

Chris Mason

author
#0 lore raw

PATCH 1/1 Btrfs: fix regression in btrfs_page_mkwrite() from vm_fault_t conversion

The vm_fault_t conversion commit introduced a ret2 variable for tracking the integer return values from internal btrfs functions. It was sometimes returning VM_FAULT_LOCKED for pages that were actually invalid and had been removed from the radix. Something like this:

	ret2 = btrfs_delalloc_reserve_space() // returns zero on success

	lock_page(page)
	if (page->mapping != inode->i_mapping)
	    goto out_unlock;

...

out_unlock:

	if (!ret2) {
		...
		return VM_FAULT_LOCKED;
	}

This ends up triggering this WARNING in btrfs_destroy_inode()

	WARN_ON(BTRFS_I(inode)->block_rsv.size);

The fix used here is to use ret instead of ret2 in our check for success.

Fixes a528a2415087 (btrfs: change return type of btrfs_page_mkwrite to vm_fault_t)
Signed-off-by Chris Mason <[email protected]>
---
 fs/btrfs/inode.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
fs/btrfs/inode.c +2 -1
index 193f933..38403aa 100644
@@ -9013,6 +9013,7 @@ vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
 		goto out_unlock;
 	}
 	ret2 = 0;
+	ret = 0;
 
 	/* page is wholly or partially inside EOF */
 	if (page_start + PAGE_SIZE > size)

@@ -9037,7 +9038,7 @@ vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
 	unlock_extent_cached(io_tree, page_start, page_end, &cached_state);
 
 out_unlock:
-	if (!ret2) {
+	if (!ret) {
 		btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, true);
 		sb_end_pagefault(inode->i_sb);
 		extent_changeset_free(data_reserved);
Signature · 1 line
2.9.5
#1 lore raw
Quoted 35 lines · Chris Mason

The vm_fault_t conversion commit introduced a ret2 variable for tracking the integer return values from internal btrfs functions. It was sometimes returning VM_FAULT_LOCKED for pages that were actually invalid and had been removed from the radix. Something like this:

	ret2 = btrfs_delalloc_reserve_space() // returns zero on success

	lock_page(page)
	if (page->mapping != inode->i_mapping)
	    goto out_unlock;

...

out_unlock:

	if (!ret2) {
		...
		return VM_FAULT_LOCKED;
	}

This ends up triggering this WARNING in btrfs_destroy_inode()

	WARN_ON(BTRFS_I(inode)->block_rsv.size);

The fix used here is to use ret instead of ret2 in our check for success.

Fixes a528a2415087 (btrfs: change return type of btrfs_page_mkwrite to vm_fault_t)
Signed-off-by Chris Mason <[email protected]>
---
 fs/btrfs/inode.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
fs/btrfs/inode.c +0 -0
index 193f933..38403aa 100644

fs/btrfs/inode.c vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf) · line 9016

Quoted 4 lines
fs/btrfs/inode.c +1 -0
@@ -9013,6 +9013,7 @@ vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
 		goto out_unlock;
 	}
 	ret2 = 0;
+	ret = 0;

That's something I'd rather avoid, now there are two error values that need to be tracked. Shouldn't the 'ret2 = 0' be removed completely?

fs/btrfs/inode.c vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf) · line 9041

Quoted 9 lines
fs/btrfs/inode.c +1 -1
 	/* page is wholly or partially inside EOF */
 	if (page_start + PAGE_SIZE > size)

@@ -9037,7 +9038,7 @@ vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
 	unlock_extent_cached(io_tree, page_start, page_end, &cached_state);
 
 out_unlock:
-	if (!ret2) {
+	if (!ret) {

I'm not sure is safe, the new ret is one of the VM_FAULT_ values and it's not 0 by default and in fact I don't see anything that would set it to 0 directly or indirectly. Which means that the code in the out: label also needs to be revisited:

9016 out:
9017 btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, (ret != 0));
9018 btrfs_delalloc_release_space(inode, data_reserved, page_start,
9019 reserved_space, (ret != 0));

Quoted 3 lines
fs/btrfs/inode.c +0 -0
 		btrfs_delalloc_release_extents(BTRFS_I(inode), PAGE_SIZE, true);
 		sb_end_pagefault(inode->i_sb);
 		extent_changeset_free(data_reserved);
author
#2 lore raw
Quoted 14 lines · David Sterba

fs/btrfs/inode.c vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf) · line 9016

Quoted 9 lines · Chris Mason
fs/btrfs/inode.c +1 -0
index 193f933..38403aa 100644
@@ -9013,6 +9013,7 @@ vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
 		goto out_unlock;
 	}
 	ret2 = 0;
+	ret = 0;

That's something I'd rather avoid, now there are two error values that need to be tracked. Shouldn't the 'ret2 = 0' be removed completely?

Yeah, the whole thing hurts my eyes. Fixing.

Quoted 7 lines
Quoted 7 lines
fs/btrfs/inode.c +0 -0
 	/* page is wholly or partially inside EOF */
 	if (page_start + PAGE_SIZE > size)

@@ -9037,7 +9038,7 @@ vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf)
 	unlock_extent_cached(io_tree, page_start, page_end, &cached_state);
Quoted 8 lines

fs/btrfs/inode.c vm_fault_t btrfs_page_mkwrite(struct vm_fault *vmf) · line 9041

Quoted 3 lines
fs/btrfs/inode.c +1 -1
 out_unlock:
-	if (!ret2) {
+	if (!ret) {

I'm not sure is safe, the new ret is one of the VM_FAULT_ values and it's not 0 by default and in fact I don't see anything that would set it to 0 directly or indirectly. Which means that the code in the out: label also needs to be revisited:

Good point, ok.

-chris