On 2017年10月24日 20:01, Nikolay Borisov wrote:
>
>
> On 24.10.2017 11:39, Qu Wenruo wrote:
>> When modifying qgroup relationship, for qgroup which only owns exclusive
>> extents, we will go through quick update path.
>>
>> In quick update path, we will just adding/removing exclusive and reference
>> number.
>>
>> However we did the opposite for qgroup reservation from the very
>> beginning.
>
> I'm afraid this sentence doesn't give much information about what's
> really going on.
I'll try to reorganize it to give a better explanation on this.
>
>>
>> In fact, we should also inherit the qgroup reservation space, just like
>> exclusive and reference numbers.
>>
>> Fix by using the newly introduced
>> qgroup_rsv_increase/decrease_by_qgroup() function call.
>>
>> Signed-off-by: Qu Wenruo <wqu@xxxxxxxx>
>> ---
>> fs/btrfs/qgroup.c | 39 ++++++++++++++++++---------------------
>> 1 file changed, 18 insertions(+), 21 deletions(-)
>>
>> diff --git a/fs/btrfs/qgroup.c b/fs/btrfs/qgroup.c
>> index 7b89da9589c1..ba6f60fd0e96 100644
>> --- a/fs/btrfs/qgroup.c
>> +++ b/fs/btrfs/qgroup.c
>> @@ -1069,21 +1069,24 @@ static void report_reserved_underflow(struct btrfs_fs_info *fs_info,
>> #endif
>> qgroup->reserved = 0;
>> }
>> +
>> /*
>> - * The easy accounting, if we are adding/removing the only ref for an extent
>> - * then this qgroup and all of the parent qgroups get their reference and
>> - * exclusive counts adjusted.
>> + * The easy accounting, we're updating qgroup relationship whose child qgroup
>> + * only have exclusive extents.
>> + * In this case, we only need to update the rfer/excl, and inherit rsv from
>> + * child qgroup (@src)
>> *
>> * Caller should hold fs_info->qgroup_lock.
>> */
>> static int __qgroup_excl_accounting(struct btrfs_fs_info *fs_info,
>> struct ulist *tmp, u64 ref_root,
>> - u64 num_bytes, int sign)
>> + struct btrfs_qgroup *src, int sign)
>> {
>> struct btrfs_qgroup *qgroup;
>> struct btrfs_qgroup_list *glist;
>> struct ulist_node *unode;
>> struct ulist_iterator uiter;
>> + u64 num_bytes = src->excl;
>> int ret = 0;
>>
>> qgroup = find_qgroup_rb(fs_info, ref_root);
>> @@ -1096,13 +1099,12 @@ static int __qgroup_excl_accounting(struct btrfs_fs_info *fs_info,
>> WARN_ON(sign < 0 && qgroup->excl < num_bytes);
>> qgroup->excl += sign * num_bytes;
>> qgroup->excl_cmpr += sign * num_bytes;
>> - if (sign > 0) {
>> - trace_qgroup_update_reserve(fs_info, qgroup, -(s64)num_bytes);
>> - if (qgroup->reserved < num_bytes)
>> - report_reserved_underflow(fs_info, qgroup, num_bytes);
>> - else
>> - qgroup->reserved -= num_bytes;
>> - }
>> +
>> + /* *Inherit* qgroup rsv info from @src */
>> + if (sign > 0)
>> + qgroup_rsv_increase_by_qgroup(qgroup, src);
>> + else
>> + qgroup_rsv_decrease_by_qgroup(qgroup, src);
>
>
> I'm a bit confused by the semantics of the 'sign' variable. So what you
> are doing is that if sign is > 0 then you are "adding a relationship"
> i.e. adding 'src reservation to 'qgroup', presumably because the src is
> a child of qgroup? So you are handling both adding and deletion in the
> if statement?
Yes, the original design of @sign is to allow single function to handle
both relationship adding and deleting.
just like the rest code, which uses @sign to handle both adding and
deleting without using if.
>
> However, before that apparently only deleting a relation ship was
> handled by that same if (And I believe that was wrong since if sign > 0
> then we should be adding bytes but here we are subtracting). SO the bug
> being fixed by this commit are actually 2 bugs:
>
> 1. Completely missing the "adding a relation ship case"
> 2. Incorrect hanlding of sign < 0, since this was handled by the sign >
> 0 case?
Yes, in fact 2 bugs.
Although the original code is acting like it's allocating space inside
the new parent, so it reduces parent's reserved, and adding new excl/refer.
However it's not the case, it should do inheriting, not allocating from
parent.
For sign > 0, (adding relationship) parent should inherit all excl/rfer
and reserved space.
For sign < 0, (deleting relationshio) parent should have all its
excl/rfer along with reserved space removed.
^^^ This should be the correct behavior.
The original code is just a copy of older code, as you can see in commit
9c8b35b1ba21 ("btrfs: quota: Automatically update related qgroups or
mark INCONSISTENT flags when assigning/deleting a qgroup relations.").
So it's a bug dating back to ancient days and it's my fault I didn't
expose it in the very beginning.
Thanks,
Qu
>
>
>
>>
>> qgroup_dirty(fs_info, qgroup);
>>
>> @@ -1122,15 +1124,10 @@ static int __qgroup_excl_accounting(struct btrfs_fs_info *fs_info,
>> qgroup->rfer_cmpr += sign * num_bytes;
>> WARN_ON(sign < 0 && qgroup->excl < num_bytes);
>> qgroup->excl += sign * num_bytes;
>> - if (sign > 0) {
>> - trace_qgroup_update_reserve(fs_info, qgroup,
>> - -(s64)num_bytes);
>> - if (qgroup->reserved < num_bytes)
>> - report_reserved_underflow(fs_info, qgroup,
>> - num_bytes);
>> - else
>> - qgroup->reserved -= num_bytes;
>> - }
>> + if (sign > 0)
>> + qgroup_rsv_increase_by_qgroup(qgroup, src);
>> + else
>> + qgroup_rsv_decrease_by_qgroup(qgroup, src);
>> qgroup->excl_cmpr += sign * num_bytes;
>> qgroup_dirty(fs_info, qgroup);
>>
>> @@ -1173,7 +1170,7 @@ static int quick_update_accounting(struct btrfs_fs_info *fs_info,
>> if (qgroup->excl == qgroup->rfer) {
>> ret = 0;
>> err = __qgroup_excl_accounting(fs_info, tmp, dst,
>> - qgroup->excl, sign);
>> + qgroup, sign);
>> if (err < 0) {
>> ret = err;
>> goto out;
>>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-btrfs" in
> the body of a message to majordomo@xxxxxxxxxxxxxxx
> More majordomo info at http://vger.kernel.org/majordomo-info.html
>
Attachment:
signature.asc
Description: OpenPGP digital signature
