[PATCH] drm/nouveau: wait for the exclusive fence after the shared ones v2

[PATCH] drm/nouveau: wait for the exclusive fence after the shared ones v2

[Christian König]

Always waiting for the exclusive fence resulted on some performance
regressions. So try to wait for the shared fences first, then the
exclusive fence should always be signaled already.

v2: fix incorrectly placed "(", add some comment why we do this.

Signed-off-by: Christian König <christian.koenig@amd.com>
---
 drivers/gpu/drm/nouveau/nouveau_fence.c | 28 +++++++++++++------------
 1 file changed, 15 insertions(+), 13 deletions(-)

diff --git a/drivers/gpu/drm/nouveau/nouveau_fence.c b/drivers/gpu/drm/nouveau/nouveau_fence.c
index 05d0b3eb3690..0ae416aa76dc 100644
--- a/drivers/gpu/drm/nouveau/nouveau_fence.c
+++ b/drivers/gpu/drm/nouveau/nouveau_fence.c
@@ -353,15 +353,22 @@ nouveau_fence_sync(struct nouveau_bo *nvbo, struct nouveau_channel *chan, bool e
 
 		if (ret)
 			return ret;
-	}
 
-	fobj = dma_resv_shared_list(resv);
-	fence = dma_resv_excl_fence(resv);
+		fobj = NULL;
+	} else {
+		fobj = dma_resv_shared_list(resv);
+	}
 
-	if (fence) {
+	/* Waiting for the exclusive fence first causes performance regressions
+	 * under some circumstances. So manually wait for the shared ones first.
+	 */
+	for (i = 0; i < (fobj ? fobj->shared_count : 0) && !ret; ++i) {
 		struct nouveau_channel *prev = NULL;
 		bool must_wait = true;
 
+		fence = rcu_dereference_protected(fobj->shared[i],
+						dma_resv_held(resv));
+
 		f = nouveau_local_fence(fence, chan->drm);
 		if (f) {
 			rcu_read_lock();
@@ -373,20 +380,13 @@ nouveau_fence_sync(struct nouveau_bo *nvbo, struct nouveau_channel *chan, bool e
 
 		if (must_wait)
 			ret = dma_fence_wait(fence, intr);
-
-		return ret;
 	}
 
-	if (!exclusive || !fobj)
-		return ret;
-
-	for (i = 0; i < fobj->shared_count && !ret; ++i) {
+	fence = dma_resv_excl_fence(resv);
+	if (fence) {
 		struct nouveau_channel *prev = NULL;
 		bool must_wait = true;
 
-		fence = rcu_dereference_protected(fobj->shared[i],
-						dma_resv_held(resv));
-
 		f = nouveau_local_fence(fence, chan->drm);
 		if (f) {
 			rcu_read_lock();
@@ -398,6 +398,8 @@ nouveau_fence_sync(struct nouveau_bo *nvbo, struct nouveau_channel *chan, bool e
 
 		if (must_wait)
 			ret = dma_fence_wait(fence, intr);
+
+		return ret;
 	}
 
 	return ret;
-- 
2.25.1

Re: [Nouveau] [PATCH] drm/nouveau: wait for the exclusive fence after the shared ones v2

[Thorsten Leemhuis]

Hi, this is your Linux kernel regression tracker speaking.

On 09.12.21 11:23, Christian König wrote:
> Always waiting for the exclusive fence resulted on some performance
> regressions. So try to wait for the shared fences first, then the
> exclusive fence should always be signaled already.
> 
> v2: fix incorrectly placed "(", add some comment why we do this.
> 
> Signed-off-by: Christian König <christian.koenig@amd.com>

FWIW: In case you need to send an improved patch, could you please add
this (see (¹) below for the reasoning):

Link:
https://lore.kernel.org/dri-devel/da142fb9-07d7-24fe-4533-0247b8d16cdd@sfritsch.de/

And if the patch is already good to go: could the subsystem maintainer
please add it when applying? See (¹) for the reasoning.

BTW, these two lines afaics are missing as well:

Fixes: 3e1ad79bf661 ("drm/nouveau: always wait for the exclusive fence")
Reported-by: Stefan Fritsch <sf@sfritsch.de>

Ciao, Thorsten

(¹) Long story: The commit message would benefit from a link to the
regression report, for reasons explained in
Documentation/process/submitting-patches.rst. To quote:

```
If related discussions or any other background information behind the
change can be found on the web, add 'Link:' tags pointing to it. In case
your patch fixes a bug, for example, add a tag with a URL referencing
the report in the mailing list archives or a bug tracker;
```

This concept is old, but the text was reworked recently to make this use
case for the Link: tag clearer. For details see:
https://git.kernel.org/linus/1f57bd42b77c

Yes, that "Link:" is not really crucial; but it's good to have if
someone needs to look into the backstory of this change sometime in the
future. But I care for a different reason. I'm tracking this regression
(and others) with regzbot, my Linux kernel regression tracking bot. This
bot will notice if a patch with a Link: tag to a tracked regression gets
posted and record that, which allowed anyone looking into the regression
to quickly gasp the current status from regzbot's webui
(https://linux-regtracking.leemhuis.info/regzbot ) or its reports. The
bot will also notice if a commit with a Link: tag to a regression report
is applied by Linus and then automatically mark the regression as
resolved then.

IOW: this tag makes my life a regression tracker a lot easier, as I
otherwise have to tell regzbot manually when the fix lands. :-/

#regzbot ^backmonitor:
https://lore.kernel.org/dri-devel/da142fb9-07d7-24fe-4533-0247b8d16cdd@sfritsch.de/

Re: [PATCH] drm/nouveau: wait for the exclusive fence after the shared ones v2

[Stefan Fritsch]

On 09.12.21 11:23, Christian König wrote:
> Always waiting for the exclusive fence resulted on some performance
> regressions. So try to wait for the shared fences first, then the
> exclusive fence should always be signaled already.
> 
> v2: fix incorrectly placed "(", add some comment why we do this.
> 
> Signed-off-by: Christian König <christian.koenig@amd.com>

Tested-by: Stefan Fritsch <sf@sfritsch.de>

Please also add a cc for linux-stable, so that this is fixed in 5.15.x

Cheers,
Stefan

> ---
>   drivers/gpu/drm/nouveau/nouveau_fence.c | 28 +++++++++++++------------
>   1 file changed, 15 insertions(+), 13 deletions(-)
> 
> diff --git a/drivers/gpu/drm/nouveau/nouveau_fence.c b/drivers/gpu/drm/nouveau/nouveau_fence.c
> index 05d0b3eb3690..0ae416aa76dc 100644
> --- a/drivers/gpu/drm/nouveau/nouveau_fence.c
> +++ b/drivers/gpu/drm/nouveau/nouveau_fence.c
> @@ -353,15 +353,22 @@ nouveau_fence_sync(struct nouveau_bo *nvbo, struct nouveau_channel *chan, bool e
>   
>   		if (ret)
>   			return ret;
> -	}
>   
> -	fobj = dma_resv_shared_list(resv);
> -	fence = dma_resv_excl_fence(resv);
> +		fobj = NULL;
> +	} else {
> +		fobj = dma_resv_shared_list(resv);
> +	}
>   
> -	if (fence) {
> +	/* Waiting for the exclusive fence first causes performance regressions
> +	 * under some circumstances. So manually wait for the shared ones first.
> +	 */
> +	for (i = 0; i < (fobj ? fobj->shared_count : 0) && !ret; ++i) {
>   		struct nouveau_channel *prev = NULL;
>   		bool must_wait = true;
>   
> +		fence = rcu_dereference_protected(fobj->shared[i],
> +						dma_resv_held(resv));
> +
>   		f = nouveau_local_fence(fence, chan->drm);
>   		if (f) {
>   			rcu_read_lock();
> @@ -373,20 +380,13 @@ nouveau_fence_sync(struct nouveau_bo *nvbo, struct nouveau_channel *chan, bool e
>   
>   		if (must_wait)
>   			ret = dma_fence_wait(fence, intr);
> -
> -		return ret;
>   	}
>   
> -	if (!exclusive || !fobj)
> -		return ret;
> -
> -	for (i = 0; i < fobj->shared_count && !ret; ++i) {
> +	fence = dma_resv_excl_fence(resv);
> +	if (fence) {
>   		struct nouveau_channel *prev = NULL;
>   		bool must_wait = true;
>   
> -		fence = rcu_dereference_protected(fobj->shared[i],
> -						dma_resv_held(resv));
> -
>   		f = nouveau_local_fence(fence, chan->drm);
>   		if (f) {
>   			rcu_read_lock();
> @@ -398,6 +398,8 @@ nouveau_fence_sync(struct nouveau_bo *nvbo, struct nouveau_channel *chan, bool e
>   
>   		if (must_wait)
>   			ret = dma_fence_wait(fence, intr);
> +
> +		return ret;
>   	}
>   
>   	return ret;

Re: [PATCH] drm/nouveau: wait for the exclusive fence after the shared ones v2

[Christian König]

Am 11.12.21 um 10:59 schrieb Stefan Fritsch:
> On 09.12.21 11:23, Christian König wrote:
>> Always waiting for the exclusive fence resulted on some performance
>> regressions. So try to wait for the shared fences first, then the
>> exclusive fence should always be signaled already.
>>
>> v2: fix incorrectly placed "(", add some comment why we do this.
>>
>> Signed-off-by: Christian König <christian.koenig@amd.com>
>
> Tested-by: Stefan Fritsch <sf@sfritsch.de>

Thanks.

>
> Please also add a cc for linux-stable, so that this is fixed in 5.15.x

Sure, but I still need some acked-by or rb from one of the Nouveau guys. 
So gentle ping on that.

Regards,
Christian.

>
> Cheers,
> Stefan
>
>> ---
>>   drivers/gpu/drm/nouveau/nouveau_fence.c | 28 +++++++++++++------------
>>   1 file changed, 15 insertions(+), 13 deletions(-)
>>
>> diff --git a/drivers/gpu/drm/nouveau/nouveau_fence.c 
>> b/drivers/gpu/drm/nouveau/nouveau_fence.c
>> index 05d0b3eb3690..0ae416aa76dc 100644
>> --- a/drivers/gpu/drm/nouveau/nouveau_fence.c
>> +++ b/drivers/gpu/drm/nouveau/nouveau_fence.c
>> @@ -353,15 +353,22 @@ nouveau_fence_sync(struct nouveau_bo *nvbo, 
>> struct nouveau_channel *chan, bool e
>>             if (ret)
>>               return ret;
>> -    }
>>   -    fobj = dma_resv_shared_list(resv);
>> -    fence = dma_resv_excl_fence(resv);
>> +        fobj = NULL;
>> +    } else {
>> +        fobj = dma_resv_shared_list(resv);
>> +    }
>>   -    if (fence) {
>> +    /* Waiting for the exclusive fence first causes performance 
>> regressions
>> +     * under some circumstances. So manually wait for the shared 
>> ones first.
>> +     */
>> +    for (i = 0; i < (fobj ? fobj->shared_count : 0) && !ret; ++i) {
>>           struct nouveau_channel *prev = NULL;
>>           bool must_wait = true;
>>   +        fence = rcu_dereference_protected(fobj->shared[i],
>> +                        dma_resv_held(resv));
>> +
>>           f = nouveau_local_fence(fence, chan->drm);
>>           if (f) {
>>               rcu_read_lock();
>> @@ -373,20 +380,13 @@ nouveau_fence_sync(struct nouveau_bo *nvbo, 
>> struct nouveau_channel *chan, bool e
>>             if (must_wait)
>>               ret = dma_fence_wait(fence, intr);
>> -
>> -        return ret;
>>       }
>>   -    if (!exclusive || !fobj)
>> -        return ret;
>> -
>> -    for (i = 0; i < fobj->shared_count && !ret; ++i) {
>> +    fence = dma_resv_excl_fence(resv);
>> +    if (fence) {
>>           struct nouveau_channel *prev = NULL;
>>           bool must_wait = true;
>>   -        fence = rcu_dereference_protected(fobj->shared[i],
>> -                        dma_resv_held(resv));
>> -
>>           f = nouveau_local_fence(fence, chan->drm);
>>           if (f) {
>>               rcu_read_lock();
>> @@ -398,6 +398,8 @@ nouveau_fence_sync(struct nouveau_bo *nvbo, 
>> struct nouveau_channel *chan, bool e
>>             if (must_wait)
>>               ret = dma_fence_wait(fence, intr);
>> +
>> +        return ret;
>>       }
>>         return ret;

Re: [Nouveau] [PATCH] drm/nouveau: wait for the exclusive fence after the shared ones v2

[Ben Skeggs]

On Tue, 14 Dec 2021 at 19:19, Christian König
<ckoenig.leichtzumerken@gmail.com> wrote:
>
> Am 11.12.21 um 10:59 schrieb Stefan Fritsch:
> > On 09.12.21 11:23, Christian König wrote:
> >> Always waiting for the exclusive fence resulted on some performance
> >> regressions. So try to wait for the shared fences first, then the
> >> exclusive fence should always be signaled already.
> >>
> >> v2: fix incorrectly placed "(", add some comment why we do this.
> >>
> >> Signed-off-by: Christian König <christian.koenig@amd.com>
> >
> > Tested-by: Stefan Fritsch <sf@sfritsch.de>
>
> Thanks.
>
> >
> > Please also add a cc for linux-stable, so that this is fixed in 5.15.x
>
> Sure, but I still need some acked-by or rb from one of the Nouveau guys.
> So gentle ping on that.
Acked-by: Ben Skeggs <bskeggs@redhat.com>

>
> Regards,
> Christian.
>
> >
> > Cheers,
> > Stefan
> >
> >> ---
> >>   drivers/gpu/drm/nouveau/nouveau_fence.c | 28 +++++++++++++------------
> >>   1 file changed, 15 insertions(+), 13 deletions(-)
> >>
> >> diff --git a/drivers/gpu/drm/nouveau/nouveau_fence.c
> >> b/drivers/gpu/drm/nouveau/nouveau_fence.c
> >> index 05d0b3eb3690..0ae416aa76dc 100644
> >> --- a/drivers/gpu/drm/nouveau/nouveau_fence.c
> >> +++ b/drivers/gpu/drm/nouveau/nouveau_fence.c
> >> @@ -353,15 +353,22 @@ nouveau_fence_sync(struct nouveau_bo *nvbo,
> >> struct nouveau_channel *chan, bool e
> >>             if (ret)
> >>               return ret;
> >> -    }
> >>   -    fobj = dma_resv_shared_list(resv);
> >> -    fence = dma_resv_excl_fence(resv);
> >> +        fobj = NULL;
> >> +    } else {
> >> +        fobj = dma_resv_shared_list(resv);
> >> +    }
> >>   -    if (fence) {
> >> +    /* Waiting for the exclusive fence first causes performance
> >> regressions
> >> +     * under some circumstances. So manually wait for the shared
> >> ones first.
> >> +     */
> >> +    for (i = 0; i < (fobj ? fobj->shared_count : 0) && !ret; ++i) {
> >>           struct nouveau_channel *prev = NULL;
> >>           bool must_wait = true;
> >>   +        fence = rcu_dereference_protected(fobj->shared[i],
> >> +                        dma_resv_held(resv));
> >> +
> >>           f = nouveau_local_fence(fence, chan->drm);
> >>           if (f) {
> >>               rcu_read_lock();
> >> @@ -373,20 +380,13 @@ nouveau_fence_sync(struct nouveau_bo *nvbo,
> >> struct nouveau_channel *chan, bool e
> >>             if (must_wait)
> >>               ret = dma_fence_wait(fence, intr);
> >> -
> >> -        return ret;
> >>       }
> >>   -    if (!exclusive || !fobj)
> >> -        return ret;
> >> -
> >> -    for (i = 0; i < fobj->shared_count && !ret; ++i) {
> >> +    fence = dma_resv_excl_fence(resv);
> >> +    if (fence) {
> >>           struct nouveau_channel *prev = NULL;
> >>           bool must_wait = true;
> >>   -        fence = rcu_dereference_protected(fobj->shared[i],
> >> -                        dma_resv_held(resv));
> >> -
> >>           f = nouveau_local_fence(fence, chan->drm);
> >>           if (f) {
> >>               rcu_read_lock();
> >> @@ -398,6 +398,8 @@ nouveau_fence_sync(struct nouveau_bo *nvbo,
> >> struct nouveau_channel *chan, bool e
> >>             if (must_wait)
> >>               ret = dma_fence_wait(fence, intr);
> >> +
> >> +        return ret;
> >>       }
> >>         return ret;
>

Re: [Nouveau] [PATCH] drm/nouveau: wait for the exclusive fence after the shared ones v2

[Thorsten Leemhuis]

Hi, this is your Linux kernel regression tracker speaking.

CCing Dave and Daniel.

On 15.12.21 23:32, Ben Skeggs wrote:
> On Tue, 14 Dec 2021 at 19:19, Christian König
> <ckoenig.leichtzumerken@gmail.com> wrote:
>>
>> Am 11.12.21 um 10:59 schrieb Stefan Fritsch:
>>> On 09.12.21 11:23, Christian König wrote:
>>>> Always waiting for the exclusive fence resulted on some performance
>>>> regressions. So try to wait for the shared fences first, then the
>>>> exclusive fence should always be signaled already.
>>>>
>>>> v2: fix incorrectly placed "(", add some comment why we do this.
>>>>
>>>> Signed-off-by: Christian König <christian.koenig@amd.com>
>>>
>>> Tested-by: Stefan Fritsch <sf@sfritsch.de>
>>
>> Thanks.
>>
>>>
>>> Please also add a cc for linux-stable, so that this is fixed in 5.15.x
>>
>> Sure, but I still need some acked-by or rb from one of the Nouveau guys.
>> So gentle ping on that.
> Acked-by: Ben Skeggs <bskeggs@redhat.com>

What's the status of this patch? I checked a few git trees, but either
it's not there or it missed it.

Reminder, it's a regression already introduced in v5.15, hence all users
of the current stable kernel are affected by it, so it would be nice to
get the fix on its way now that Ben acked it and Dan tested it.

Ciao, Thorsten

P.S.: As a Linux kernel regression tracker I'm getting a lot of reports
on my table. I can only look briefly into most of them. Unfortunately
therefore I sometimes will get things wrong or miss something important.
I hope that's not the case here; if you think it is, don't hesitate to
tell me about it in a public reply. That's in everyone's interest, as
what I wrote above might be misleading to everyone reading this; any
suggestion I gave thus might sent someone reading this down the wrong
rabbit hole, which none of us wants.

BTW, I have no personal interest in this issue, which is tracked using
regzbot, my Linux kernel regression tracking bot
(https://linux-regtracking.leemhuis.info/regzbot/). I'm only posting
this mail to get things rolling again and hence don't need to be CC on
all further activities wrt to this regression.

#regzbot poke

>>>> ---
>>>>   drivers/gpu/drm/nouveau/nouveau_fence.c | 28 +++++++++++++------------
>>>>   1 file changed, 15 insertions(+), 13 deletions(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/nouveau/nouveau_fence.c
>>>> b/drivers/gpu/drm/nouveau/nouveau_fence.c
>>>> index 05d0b3eb3690..0ae416aa76dc 100644
>>>> --- a/drivers/gpu/drm/nouveau/nouveau_fence.c
>>>> +++ b/drivers/gpu/drm/nouveau/nouveau_fence.c
>>>> @@ -353,15 +353,22 @@ nouveau_fence_sync(struct nouveau_bo *nvbo,
>>>> struct nouveau_channel *chan, bool e
>>>>             if (ret)
>>>>               return ret;
>>>> -    }
>>>>   -    fobj = dma_resv_shared_list(resv);
>>>> -    fence = dma_resv_excl_fence(resv);
>>>> +        fobj = NULL;
>>>> +    } else {
>>>> +        fobj = dma_resv_shared_list(resv);
>>>> +    }
>>>>   -    if (fence) {
>>>> +    /* Waiting for the exclusive fence first causes performance
>>>> regressions
>>>> +     * under some circumstances. So manually wait for the shared
>>>> ones first.
>>>> +     */
>>>> +    for (i = 0; i < (fobj ? fobj->shared_count : 0) && !ret; ++i) {
>>>>           struct nouveau_channel *prev = NULL;
>>>>           bool must_wait = true;
>>>>   +        fence = rcu_dereference_protected(fobj->shared[i],
>>>> +                        dma_resv_held(resv));
>>>> +
>>>>           f = nouveau_local_fence(fence, chan->drm);
>>>>           if (f) {
>>>>               rcu_read_lock();
>>>> @@ -373,20 +380,13 @@ nouveau_fence_sync(struct nouveau_bo *nvbo,
>>>> struct nouveau_channel *chan, bool e
>>>>             if (must_wait)
>>>>               ret = dma_fence_wait(fence, intr);
>>>> -
>>>> -        return ret;
>>>>       }
>>>>   -    if (!exclusive || !fobj)
>>>> -        return ret;
>>>> -
>>>> -    for (i = 0; i < fobj->shared_count && !ret; ++i) {
>>>> +    fence = dma_resv_excl_fence(resv);
>>>> +    if (fence) {
>>>>           struct nouveau_channel *prev = NULL;
>>>>           bool must_wait = true;
>>>>   -        fence = rcu_dereference_protected(fobj->shared[i],
>>>> -                        dma_resv_held(resv));
>>>> -
>>>>           f = nouveau_local_fence(fence, chan->drm);
>>>>           if (f) {
>>>>               rcu_read_lock();
>>>> @@ -398,6 +398,8 @@ nouveau_fence_sync(struct nouveau_bo *nvbo,
>>>> struct nouveau_channel *chan, bool e
>>>>             if (must_wait)
>>>>               ret = dma_fence_wait(fence, intr);
>>>> +
>>>> +        return ret;
>>>>       }
>>>>         return ret;
>>

Re: [Nouveau] [PATCH] drm/nouveau: wait for the exclusive fence after the shared ones v2

[Christian König]

Am 21.12.21 um 11:11 schrieb Thorsten Leemhuis:
> Hi, this is your Linux kernel regression tracker speaking.
>
> CCing Dave and Daniel.
>
> On 15.12.21 23:32, Ben Skeggs wrote:
>> On Tue, 14 Dec 2021 at 19:19, Christian König
>> <ckoenig.leichtzumerken@gmail.com> wrote:
>>> Am 11.12.21 um 10:59 schrieb Stefan Fritsch:
>>>> On 09.12.21 11:23, Christian König wrote:
>>>>> Always waiting for the exclusive fence resulted on some performance
>>>>> regressions. So try to wait for the shared fences first, then the
>>>>> exclusive fence should always be signaled already.
>>>>>
>>>>> v2: fix incorrectly placed "(", add some comment why we do this.
>>>>>
>>>>> Signed-off-by: Christian König <christian.koenig@amd.com>
>>>> Tested-by: Stefan Fritsch <sf@sfritsch.de>
>>> Thanks.
>>>
>>>> Please also add a cc for linux-stable, so that this is fixed in 5.15.x
>>> Sure, but I still need some acked-by or rb from one of the Nouveau guys.
>>> So gentle ping on that.
>> Acked-by: Ben Skeggs <bskeggs@redhat.com>
> What's the status of this patch? I checked a few git trees, but either
> it's not there or it missed it.

You missed it. I've pushed it to drm-misc-fixes about 2 hours ago: 
https://cgit.freedesktop.org/drm/drm-misc/log/?h=drm-misc-fixes

Regards,
Christian.

>
> Reminder, it's a regression already introduced in v5.15, hence all users
> of the current stable kernel are affected by it, so it would be nice to
> get the fix on its way now that Ben acked it and Dan tested it.
>
> Ciao, Thorsten
>
> P.S.: As a Linux kernel regression tracker I'm getting a lot of reports
> on my table. I can only look briefly into most of them. Unfortunately
> therefore I sometimes will get things wrong or miss something important.
> I hope that's not the case here; if you think it is, don't hesitate to
> tell me about it in a public reply. That's in everyone's interest, as
> what I wrote above might be misleading to everyone reading this; any
> suggestion I gave thus might sent someone reading this down the wrong
> rabbit hole, which none of us wants.
>
> BTW, I have no personal interest in this issue, which is tracked using
> regzbot, my Linux kernel regression tracking bot
> (https://linux-regtracking.leemhuis.info/regzbot/). I'm only posting
> this mail to get things rolling again and hence don't need to be CC on
> all further activities wrt to this regression.
>
> #regzbot poke
>
>>>>> ---
>>>>>    drivers/gpu/drm/nouveau/nouveau_fence.c | 28 +++++++++++++------------
>>>>>    1 file changed, 15 insertions(+), 13 deletions(-)
>>>>>
>>>>> diff --git a/drivers/gpu/drm/nouveau/nouveau_fence.c
>>>>> b/drivers/gpu/drm/nouveau/nouveau_fence.c
>>>>> index 05d0b3eb3690..0ae416aa76dc 100644
>>>>> --- a/drivers/gpu/drm/nouveau/nouveau_fence.c
>>>>> +++ b/drivers/gpu/drm/nouveau/nouveau_fence.c
>>>>> @@ -353,15 +353,22 @@ nouveau_fence_sync(struct nouveau_bo *nvbo,
>>>>> struct nouveau_channel *chan, bool e
>>>>>              if (ret)
>>>>>                return ret;
>>>>> -    }
>>>>>    -    fobj = dma_resv_shared_list(resv);
>>>>> -    fence = dma_resv_excl_fence(resv);
>>>>> +        fobj = NULL;
>>>>> +    } else {
>>>>> +        fobj = dma_resv_shared_list(resv);
>>>>> +    }
>>>>>    -    if (fence) {
>>>>> +    /* Waiting for the exclusive fence first causes performance
>>>>> regressions
>>>>> +     * under some circumstances. So manually wait for the shared
>>>>> ones first.
>>>>> +     */
>>>>> +    for (i = 0; i < (fobj ? fobj->shared_count : 0) && !ret; ++i) {
>>>>>            struct nouveau_channel *prev = NULL;
>>>>>            bool must_wait = true;
>>>>>    +        fence = rcu_dereference_protected(fobj->shared[i],
>>>>> +                        dma_resv_held(resv));
>>>>> +
>>>>>            f = nouveau_local_fence(fence, chan->drm);
>>>>>            if (f) {
>>>>>                rcu_read_lock();
>>>>> @@ -373,20 +380,13 @@ nouveau_fence_sync(struct nouveau_bo *nvbo,
>>>>> struct nouveau_channel *chan, bool e
>>>>>              if (must_wait)
>>>>>                ret = dma_fence_wait(fence, intr);
>>>>> -
>>>>> -        return ret;
>>>>>        }
>>>>>    -    if (!exclusive || !fobj)
>>>>> -        return ret;
>>>>> -
>>>>> -    for (i = 0; i < fobj->shared_count && !ret; ++i) {
>>>>> +    fence = dma_resv_excl_fence(resv);
>>>>> +    if (fence) {
>>>>>            struct nouveau_channel *prev = NULL;
>>>>>            bool must_wait = true;
>>>>>    -        fence = rcu_dereference_protected(fobj->shared[i],
>>>>> -                        dma_resv_held(resv));
>>>>> -
>>>>>            f = nouveau_local_fence(fence, chan->drm);
>>>>>            if (f) {
>>>>>                rcu_read_lock();
>>>>> @@ -398,6 +398,8 @@ nouveau_fence_sync(struct nouveau_bo *nvbo,
>>>>> struct nouveau_channel *chan, bool e
>>>>>              if (must_wait)
>>>>>                ret = dma_fence_wait(fence, intr);
>>>>> +
>>>>> +        return ret;
>>>>>        }
>>>>>          return ret;

Re: [Nouveau] [PATCH] drm/nouveau: wait for the exclusive fence after the shared ones v2

[Dan Moulding]

Tested-by: Dan Moulding <dmoulding@me.com>

Thanks!

-- Dan