[Intel-gfx] [PATCH 5/9] lib/ref_tracker: improve allocation flags

[Andrzej Hajda]

On 17.02.2022 16:13, Eric Dumazet wrote:
> On Thu, Feb 17, 2022 at 6:05 AM Andrzej Hajda <andrzej.hajda at intel.com> wrote:
>> Library can be called in non-sleeping context, so it should not use
>> __GFP_NOFAIL. Instead it should calmly handle allocation fails, for
>> this __GFP_NOWARN has been added as well.
> Your commit changelog is misleading .
>
> The GFP_NOFAIL issue has been fixed already in
> commit c12837d1bb31032bead9060dec99ef310d5b9fb7 ("ref_tracker: use
> __GFP_NOFAIL more carefully")

I based the patchset on drm-tip, which do not have this commit, I will 
take a look how to keep drm-tip base (to allow intel CI tests) and take 
patch above into account - maybe simple cherry-picking?

>
>
>> Signed-off-by: Andrzej Hajda <andrzej.hajda at intel.com>
>> ---
>>   lib/ref_tracker.c | 11 ++++++-----
>>   1 file changed, 6 insertions(+), 5 deletions(-)
>>
>> diff --git a/lib/ref_tracker.c b/lib/ref_tracker.c
>> index 7b00bca300043..c8441ffbb058a 100644
>> --- a/lib/ref_tracker.c
>> +++ b/lib/ref_tracker.c
>> @@ -59,7 +59,7 @@ __ref_tracker_dir_pr_ostream(struct ref_tracker_dir *dir,
>>          if (list_empty(&dir->list))
>>                  return;
>>
>> -       sbuf = kmalloc(STACK_BUF_SIZE, GFP_NOWAIT);
>> +       sbuf = kmalloc(STACK_BUF_SIZE, GFP_NOWAIT | __GFP_NOWARN);
> This belongs to patch 3 in your series.

OK, again historical reason.

>
>>          list_for_each_entry(tracker, &dir->list, head)
>>                  ++total;
>> @@ -154,11 +154,11 @@ int ref_tracker_alloc(struct ref_tracker_dir *dir,
>>          unsigned long entries[REF_TRACKER_STACK_ENTRIES];
>>          struct ref_tracker *tracker;
>>          unsigned int nr_entries;
>> -       gfp_t gfp_mask = gfp;
> Simply change this line to : gfp_t gfp_mask = gfp | __GFP_NOFAIL;

and "| __GFP_NOWARN".

>
>> +       gfp_t gfp_mask;
>>          unsigned long flags;
>>
> Then leave all this code as is ? I find current code more readable.

Yep you are right.

>
>> -       if (gfp & __GFP_DIRECT_RECLAIM)
>> -               gfp_mask |= __GFP_NOFAIL;
>> +       gfp |= __GFP_NOWARN;
>> +       gfp_mask = (gfp & __GFP_DIRECT_RECLAIM) ? (gfp | __GFP_NOFAIL) : gfp;
>>          *trackerp = tracker = kzalloc(sizeof(*tracker), gfp_mask);
>>          if (unlikely(!tracker)) {
>>                  pr_err_once("memory allocation failure, unreliable refcount tracker.\n");
>> @@ -191,7 +191,8 @@ int ref_tracker_free(struct ref_tracker_dir *dir,
>>          }
>>          nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 1);
>>          nr_entries = filter_irq_stacks(entries, nr_entries);
> lib/ref_tracker.c got patches in net-next, your patch series is going
> to add conflicts.
>
> git log --oneline 5740d0689096..4d449bdc5b26 --no-merges -- lib/ref_tracker.c
> c2d1e3df4af59261777b39c2e47476acd4d1cbeb ref_tracker: remove
> filter_irq_stacks() call
> 8fd5522f44dcd7f05454ddc4f16d0f821b676cd9 ref_tracker: add a count of
> untracked references
> e3ececfe668facd87d920b608349a32607060e66 ref_tracker: implement
> use-after-free detection


So I will cherry-pick these patches into next version of patchset, with 
"NO MERGE" annotation (to allow testing), and if my ref_track patches 
will be accepted then they can go via net-dev tree and intel patches 
will wait till update of drm-tip.
Is this scenario OK?


Regards
Andrzej

>
>
>> -       stack_handle = stack_depot_save(entries, nr_entries, GFP_ATOMIC);
>> +       stack_handle = stack_depot_save(entries, nr_entries,
>> +                                       GFP_NOWAIT | __GFP_NOWARN);
> This is fine.
>
>>          spin_lock_irqsave(&dir->lock, flags);
>>          if (tracker->dead) {
>> --
>> 2.25.1
>>