btrfs: small fixes related to qgroup deletion
cover btrfs: small fixes related to qgroup deletion
There is a recent bug report that when deleting a qgroup that is not fully exclusively owned, it can make all future qgroup operations to fail.
It turns out to be a problem handling "ret > 0" cases when deleting a qgroup.
The first patch is the root fix.
The seconds patch is to improve the error reporting of quick_update_accounting().
Finally there is a pre-existing problem exposed by that "ret > 0" handling, that if we use btrfs_remove_qgroup() on a qgroup that only exists in the rb-tree, it will return -ENOENT meanwhile still deletes the in-memory qgroup.
Fix that -ENOENT leak in the last patch.
Qu Wenruo (3):
btrfs: qgroup: do not treat "ret > 0" as error when deleting a qgroup
btrfs: qgroup: use qgroup_mark_inconsistent() in
quick_update_accounting()
btrfs: qgroup: do not leak -ENOENT from btrfs_remove_qgroup()
fs/btrfs/qgroup.c | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-)
Signature · 1 line
2.55.0
PATCH 1/3 btrfs: qgroup: do not treat "ret > 0" as error when deleting a qgroup
[BUG]
There is a bug report that when removing a child qgroup, if removing the qgroup requires a rescan, it will delete the on-disk relation and qgroup items but not the in-memory qgroup rb entry.
This makes all later qgroup rescan to fail.
[CAUSE]
The call chain looks like the following:
btrfs_remove_qgroup()
|- __del_qgroup_relation() | |- quick_update_accounting() | Which returns 1, showing this operation will require | a qgroup rescan |- if (ret) goto out;
Which skips the del_qgroup_rb() call, leaving the in-memory qgroup rb entry still intact.
Then later qgroup rescan/update will always fail and mark qgroup inconsistent:
btrfs_run_qgroups()
|- update_qgroup_info_item() | |- btrfs_search_slot() | Which returns -ENOENT, as the qgroup info item is already | deleted. |- qgroup_mark_inconsistent() |- return -ENOENT;
[FIX]
Do not treat "ret > 0" as an error.
--- fs/btrfs/qgroup.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-)
fs/btrfs/qgroup.c +1 -1
@@ -1853,7 +1853,7 @@ int btrfs_remove_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid) struct btrfs_qgroup_list, next_group); ret =__del_qgroup_relation (trans, qgroupid, list->group->qgroupid); -if (ret) +if (ret <0 ) goto out; }
Signature · 1 line
2.55.0
PATCH 2/3 btrfs: qgroup: use qgroup_mark_inconsistent() in quick_update_accounting()
In reply to Qu WenruoWhen quick_update_accounting() fails to handle a qgroup which is not fully exclusive (its exclusive number doesn't match the reference one), we have to mark the qgroup inconsistent.
Unlike other locations we didn't use qgroup_mark_inconsistent(), thus when such problem happens no extra error message is shown.
Use the helper in quick_update_accounting(), which does not only show the reason of marking qgroup inconsistent, but also cancels any running rescan and skips dirty accounting, avoiding wasting CPU time on already inconsistent qgroup numbers.
--- fs/btrfs/qgroup.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-)
fs/btrfs/qgroup.c +4 -1
@@ -1562,7 +1562,10 @@ static int quick_update_accounting(struct btrfs_fs_info *fs_info, } out:if (ret) -set_bit (BTRFS_QGROUP_STATUS_BIT_INCONSISTENT , &fs_info->qgroup_flags); +qgroup_mark_inconsistent (fs_info, +"unable to do quick excl updating for qgroup %hu/%llu, ret=%d" , +btrfs_qgroup_level (src), +btrfs_qgroup_subvolid (src), ret);return ret; }
Signature · 1 line
2.55.0
PATCH 3/3 btrfs: qgroup: do not leak -ENOENT from btrfs_remove_qgroup()
In reply to Qu WenruoIf btrfs_remove_qgroup() is called on a qgroup that only exists in the rb-tree, then del_qgroup_item() will return -ENOENT, and we tolerate that -ENOENT.
However if the to-be-deleted qgroup has no parent qgroups, the @ret will not be updated, and will leak to user space.
It will make the ioctl() caller incorrectly assume the operation failed.
Fix it by resetting @ret to 0 after a successful or tolerated del_qgroup_item() call, so the -ENOENT from del_qgroup_item() won't be leaked.
--- fs/btrfs/qgroup.c | 1 + 1 file changed, 1 insertion(+)
fs/btrfs/qgroup.c +1 -0
@@ -1850,6 +1850,7 @@ int btrfs_remove_qgroup(struct btrfs_trans_handle *trans, u64 qgroupid) ret =del_qgroup_item (trans, qgroupid);if (ret && ret != -ENOENT ) goto out; + ret =0 ;while (!list_empty (&qgroup->groups)) { list = list_first_entry(&qgroup->groups,
Signature · 1 line
2.55.0
Quoted 17 lines · Qu Wenruo
There is a recent bug report that when deleting a qgroup that is not fully exclusively owned, it can make all future qgroup operations to fail.
It turns out to be a problem handling "ret > 0" cases when deleting a qgroup.
The first patch is the root fix.The seconds patch is to improve the error reporting of
quick_update_accounting().Finally there is a pre-existing problem exposed by that "ret > 0" handling, that if we use
btrfs_remove_qgroup()on a qgroup that only exists in the rb-tree, it will return-ENOENTmeanwhile still deletes the in-memory qgroup.
Quoted 7 lines
Fix that
-ENOENTleak in the last patch.Qu Wenruo (3):
btrfs: qgroup: do not treat "ret > 0" as error when deleting a qgroup
btrfs: qgroup: useqgroup_mark_inconsistent()in
quick_update_accounting()
btrfs: qgroup: do not leak-ENOENTfrombtrfs_remove_qgroup()
Looks good, thanks.
Quoted 8 lines
fs/btrfs/qgroup.c | 8 ++++++-- 1 file changed, 6 insertions(+), 2 deletions(-)--
2.55.0