[0/5] btrfs: qgroup: reduce GFP_ATOMIC usage for ulist

Message

Qu Wenruo Aug. 30, 2023, 10:37 a.m. UTC

[REPO]
https://github.com/adam900710/linux/tree/qgroup_mutex

Yep, the branch name shows my previous failed attempt to solve the
problem.

[PROBLEM]
There are quite some GFP_ATOMIC usage for btrfs qgroups, most of them
are for ulists.

Those ulists are just used as a temporary memory to trace the involved
qgroups.

[ENHANCEMENT]
This patchset would address the problem by adding a new list_head called
iterator for btrfs_qgroup.

And all call sites but one of ulist allocation can be migrated to use
the new qgroup_iterator facility to iterate qgroups without any memory
allocation.

The only remaining ulist call site is @qgroups ulist utilized inside
btrfs_qgroup_account_extent(), which is utilized to get all involved
qgroups for both old and new roots.

I tried to extract the qgroups collection code into a dedicate loop
out of qgroup_update_refcnt(), but it would lead to test case failure of
btrfs/028 (accounts underflow).

Thus for now only the safe part is sent to the list.

And BTW since we can skip quite some memory allocation failure handling
(since there is no memory allocation), we also save some lines of code.

Qu Wenruo (5):
  btrfs: qgroup: iterate qgroups without memory allocation for
    qgroup_reserve()
  btrfs: qgroup: use qgroup_iterator facility for
    btrfs_qgroup_free_refroot()
  btrfs: qgroup: use qgroup_iterator facility for qgroup_convert_meta()
  btrfs: qgroup: use qgroup_iterator facility for
    __qgroup_excl_accounting()
  btrfs: qgroup: use qgroup_iterator facility to replace @tmp ulist of
    qgroup_update_refcnt()

 fs/btrfs/qgroup.c | 252 ++++++++++++++++------------------------------
 fs/btrfs/qgroup.h |   9 ++
 2 files changed, 94 insertions(+), 167 deletions(-)

Comments

Boris Burkov Aug. 30, 2023, 10:06 p.m. UTC | #1

On Wed, Aug 30, 2023 at 06:37:22PM +0800, Qu Wenruo wrote:
> [REPO]
> https://github.com/adam900710/linux/tree/qgroup_mutex
> 
> Yep, the branch name shows my previous failed attempt to solve the
> problem.
> 
> [PROBLEM]
> There are quite some GFP_ATOMIC usage for btrfs qgroups, most of them
> are for ulists.
> 
> Those ulists are just used as a temporary memory to trace the involved
> qgroups.
> 
> [ENHANCEMENT]
> This patchset would address the problem by adding a new list_head called
> iterator for btrfs_qgroup.
> 
> And all call sites but one of ulist allocation can be migrated to use
> the new qgroup_iterator facility to iterate qgroups without any memory
> allocation.
> 
> The only remaining ulist call site is @qgroups ulist utilized inside
> btrfs_qgroup_account_extent(), which is utilized to get all involved
> qgroups for both old and new roots.
> 
> I tried to extract the qgroups collection code into a dedicate loop
> out of qgroup_update_refcnt(), but it would lead to test case failure of
> btrfs/028 (accounts underflow).
> 
> Thus for now only the safe part is sent to the list.
> 
> And BTW since we can skip quite some memory allocation failure handling
> (since there is no memory allocation), we also save some lines of code.

These all LGTM, thanks!

Reviewed-by: Boris Burkov <[e-mail]>

> 
> Qu Wenruo (5):
>   btrfs: qgroup: iterate qgroups without memory allocation for
>     qgroup_reserve()
>   btrfs: qgroup: use qgroup_iterator facility for
>     btrfs_qgroup_free_refroot()
>   btrfs: qgroup: use qgroup_iterator facility for qgroup_convert_meta()
>   btrfs: qgroup: use qgroup_iterator facility for
>     __qgroup_excl_accounting()
>   btrfs: qgroup: use qgroup_iterator facility to replace @tmp ulist of
>     qgroup_update_refcnt()
> 
>  fs/btrfs/qgroup.c | 252 ++++++++++++++++------------------------------
>  fs/btrfs/qgroup.h |   9 ++
>  2 files changed, 94 insertions(+), 167 deletions(-)
> 
> -- 
> 2.41.0
>