[PATCH v5 3/5] drm/amdkfd: set activated flag true when event age unmatchs

Yat Sin, David David.YatSin at amd.com
Mon Jun 12 16:19:03 UTC 2023


[AMD Official Use Only - General]

The current ROCr patches already address my previous feedback. I am ok with the current ROCr patches.

Currently, there is no ROCrtst that would stress this multiple-waiters issue. I was thinking something like the KFDTest, but with by calling the waiters from different threads. @Zhu, James Would you have time to look into this?

~David

> -----Original Message-----
> From: Kuehling, Felix <Felix.Kuehling at amd.com>
> Sent: Friday, June 9, 2023 6:44 PM
> To: Zhu, James <James.Zhu at amd.com>; amd-gfx at lists.freedesktop.org
> Cc: Yat Sin, David <David.YatSin at amd.com>; Zhu, James
> <James.Zhu at amd.com>
> Subject: Re: [PATCH v5 3/5] drm/amdkfd: set activated flag true when event
> age unmatchs
>
>  From the KFD perspective, the series is
>
> Reviewed-by: Felix Kuehling <Felix.Kuehling at amd.com>
>
> David, I looked at the ROCr and Thunk changes as well, and they look
> reasonable to me. Do you have any feedback on these patches from a ROCr
> point of view? Is there a reasonable stress test that could be used check that
> this handles the race conditions as expected and allows all waiters to sleep?
>
> Regards,
>    Felix
>
>
> On 2023-06-09 16:43, James Zhu wrote:
> > Set waiter's activated flag true when event age unmatchs with
> last_event_age.
> >
> > -v4: add event type check
> > -v5: improve on event age enable and activated flags
> >
> > Signed-off-by: James Zhu <James.Zhu at amd.com>
> > ---
> >   drivers/gpu/drm/amd/amdkfd/kfd_events.c | 17 +++++++++++++----
> >   1 file changed, 13 insertions(+), 4 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_events.c
> > b/drivers/gpu/drm/amd/amdkfd/kfd_events.c
> > index c7689181cc22..b2586a1dd35d 100644
> > --- a/drivers/gpu/drm/amd/amdkfd/kfd_events.c
> > +++ b/drivers/gpu/drm/amd/amdkfd/kfd_events.c
> > @@ -41,6 +41,7 @@ struct kfd_event_waiter {
> >     wait_queue_entry_t wait;
> >     struct kfd_event *event; /* Event to wait for */
> >     bool activated;          /* Becomes true when event is signaled */
> > +   bool event_age_enabled;  /* set to true when last_event_age is
> > +non-zero */
> >   };
> >
> >   /*
> > @@ -797,9 +798,9 @@ static struct kfd_event_waiter
> > *alloc_event_waiters(uint32_t num_events)
> >
> >   static int init_event_waiter(struct kfd_process *p,
> >             struct kfd_event_waiter *waiter,
> > -           uint32_t event_id)
> > +           struct kfd_event_data *event_data)
> >   {
> > -   struct kfd_event *ev = lookup_event_by_id(p, event_id);
> > +   struct kfd_event *ev = lookup_event_by_id(p, event_data->event_id);
> >
> >     if (!ev)
> >             return -EINVAL;
> > @@ -808,6 +809,15 @@ static int init_event_waiter(struct kfd_process *p,
> >     waiter->event = ev;
> >     waiter->activated = ev->signaled;
> >     ev->signaled = ev->signaled && !ev->auto_reset;
> > +
> > +   /* last_event_age = 0 reserved for backward compatible */
> > +   if (waiter->event->type == KFD_EVENT_TYPE_SIGNAL &&
> > +           event_data->signal_event_data.last_event_age) {
> > +           waiter->event_age_enabled = true;
> > +           if (ev->event_age != event_data-
> >signal_event_data.last_event_age)
> > +                   waiter->activated = true;
> > +   }
> > +
> >     if (!waiter->activated)
> >             add_wait_queue(&ev->wq, &waiter->wait);
> >     spin_unlock(&ev->lock);
> > @@ -948,8 +958,7 @@ int kfd_wait_on_events(struct kfd_process *p,
> >                     goto out_unlock;
> >             }
> >
> > -           ret = init_event_waiter(p, &event_waiters[i],
> > -                                   event_data.event_id);
> > +           ret = init_event_waiter(p, &event_waiters[i], &event_data);
> >             if (ret)
> >                     goto out_unlock;
> >     }


More information about the amd-gfx mailing list