Skip to contentSkip to conversation

btrfs-progs: fix cross stripe boundary check

PATCH In review David Sterba 4 participants 5 messages Sep 10, 2015 – Sep 11, 2015 1 message not in the archive Btrfs file system

lore
Share & tools
lore

David Sterba

author maintainer
#0 lore raw

PATCH 1/1 btrfs-progs: fix cross stripe boundary check

Commit 854437ca3c228d8ab3eb24d2efc1c21b5d56a635 ("btrfs-progs:
extent-tree: avoid allocating tree block that crosses stripe boundary") does not work for 64k nodesize. Due to an off-by-one error, all queries to check_crossing_stripes will return that all extents cross a stripe and this will lead to a false ENOSPC. This crashes later

$ ./mkfs.btrfs -n 64k image

./mkfs.btrfs(btrfs_reserve_extent+0xb77)[0x417f38] ./mkfs.btrfs(btrfs_alloc_free_block+0x57)[0x417fe0] ./mkfs.btrfs(__btrfs_cow_block+0x163)[0x408eb7] ./mkfs.btrfs(btrfs_cow_block+0xd0)[0x4097c4] ./mkfs.btrfs(btrfs_search_slot+0x16f)[0x40be4d] ./mkfs.btrfs(btrfs_insert_empty_items+0xc0)[0x40d5f9] ./mkfs.btrfs(btrfs_insert_item+0x99)[0x40da5f] ./mkfs.btrfs(btrfs_make_block_group+0x4d)[0x41705c] ./mkfs.btrfs(main+0xeef)[0x434b56]

---
 volumes.h | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
volumes.h +1 -1
index f7761311e8ce..4ecb99314a0c 100644
@@ -156,7 +156,7 @@ struct map_lookup {
 static inline int check_crossing_stripes(u64 start, u64 len)
 {
 	return (start / BTRFS_STRIPE_LEN) !=
-	       ((start + len) / BTRFS_STRIPE_LEN);
+	       ((start + len - 1) / BTRFS_STRIPE_LEN);
 }
 
 int __btrfs_map_block(struct btrfs_mapping_tree *map_tree, int rw,
Signature · 1 line
2.1.3
Reviewed-by on v1 1/1
#1 lore raw

David Sterba wrote on 2015/09/10 17:02 +0200:

Quoted 18 lines

Commit 854437ca3c228d8ab3eb24d2efc1c21b5d56a635 ("btrfs-progs:
extent-tree: avoid allocating tree block that crosses stripe boundary") does not work for 64k nodesize. Due to an off-by-one error, all queries to check_crossing_stripes will return that all extents cross a stripe and this will lead to a false ENOSPC. This crashes later

$ ./mkfs.btrfs -n 64k image

./mkfs.btrfs(btrfs_reserve_extent+0xb77)[0x417f38] ./mkfs.btrfs(btrfs_alloc_free_block+0x57)[0x417fe0] ./mkfs.btrfs(__btrfs_cow_block+0x163)[0x408eb7] ./mkfs.btrfs(btrfs_cow_block+0xd0)[0x4097c4] ./mkfs.btrfs(btrfs_search_slot+0x16f)[0x40be4d] ./mkfs.btrfs(btrfs_insert_empty_items+0xc0)[0x40d5f9] ./mkfs.btrfs(btrfs_insert_item+0x99)[0x40da5f] ./mkfs.btrfs(btrfs_make_block_group+0x4d)[0x41705c] ./mkfs.btrfs(main+0xeef)[0x434b56]

Quoted 2 lines

Nice one, yep, my fault.

BTW, any idea to add mkfs test?

Thanks,
Qu

Quoted 18 lines
---
  volumes.h | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
volumes.h +1 -1
index f7761311e8ce..4ecb99314a0c 100644
@@ -156,7 +156,7 @@ struct map_lookup {
  static inline int check_crossing_stripes(u64 start, u64 len)
  {
  	return (start / BTRFS_STRIPE_LEN) !=
-	       ((start + len) / BTRFS_STRIPE_LEN);
+	       ((start + len - 1) / BTRFS_STRIPE_LEN);
  }
 
  int __btrfs_map_block(struct btrfs_mapping_tree *map_tree, int rw,
#2 lore raw
Quoted 1 line · Qu Wenruo

BTW, any idea to add mkfs test?

Yes, I'll add a test that will cycle through combinations of various options (nodesize, raid profiles).

#3 lore raw
In reply to David Sterba
Quoted 8 lines · David Sterba

Commit 854437ca3c228d8ab3eb24d2efc1c21b5d56a635 ("btrfs-progs:
extent-tree: avoid allocating tree block that crosses stripe boundary") does not work for 64k nodesize. Due to an off-by-one error, all queries to check_crossing_stripes will return that all extents cross a stripe and this will lead to a false ENOSPC. This crashes later

$ ./mkfs.btrfs -n 64k image
Quoted 9 lines

./mkfs.btrfs(btrfs_reserve_extent+0xb77)[0x417f38] ./mkfs.btrfs(btrfs_alloc_free_block+0x57)[0x417fe0] ./mkfs.btrfs(__btrfs_cow_block+0x163)[0x408eb7] ./mkfs.btrfs(btrfs_cow_block+0xd0)[0x4097c4] ./mkfs.btrfs(btrfs_search_slot+0x16f)[0x40be4d] ./mkfs.btrfs(btrfs_insert_empty_items+0xc0)[0x40d5f9] ./mkfs.btrfs(btrfs_insert_item+0x99)[0x40da5f] ./mkfs.btrfs(btrfs_make_block_group+0x4d)[0x41705c] ./mkfs.btrfs(main+0xeef)[0x434b56]

Am I correct that this also causes false positives with btrfs check? I just ran a sanity check on an fs that had no problems whatsoever and was definitely not converted (so 16k nodesize) and got thousands of cross-stripe complaints; repair didn't help. Applying the patch seems to have fixed those; it completes without problems now.

Holger

A message that is not in the archive (sent off-list or not yet received): CAHji1518hEv2_H08ucBYNmvGA2qA=fLs4J+FAFSXEHWv5G8KUA@mail.gmail.com

#5 lore raw
Quoted 5 lines · Holger Hoffstätte

Am I correct that this also causes false positives with btrfs check? I just ran a sanity check on an fs that had no problems whatsoever and was definitely not converted (so 16k nodesize) and got thousands of cross-stripe complaints; repair didn't help. Applying the patch seems to have fixed those; it completes without problems now.

Yes you are. If the node blocks end at the stripe boundary, they're incorrectly marked as stripe crossing. The 64k nodesize case was quick to detect that because this holds for all nodes. Thanks for your report, this means the bug is not that rare. I'll do a release within a few days including this patch.