Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 2620:137:e000::1:20 as permitted sender) client-ip=2620:137:e000::1:20;
Message-ID: <74add824-76b6-96eb-b4b2-67d908474c8b@linux.intel.com>
Date:   Tue, 4 Apr 2023 17:22:39 +0100
MIME-Version: 1.0
User-Agent: Mozilla/5.0 (X11; Linux x86_64; rv:102.0) Gecko/20100101
 Thunderbird/102.8.0
Subject: Re: [PATCH v2] drm/i915/gt: Hold a wakeref for the active VM
Content-Language: en-US
To:     Andi Shyti <andi.shyti@linux.intel.com>
Cc:     Andrzej Hajda <andrzej.hajda@intel.com>,
        Jani Nikula <jani.nikula@linux.intel.com>,
        Joonas Lahtinen <joonas.lahtinen@linux.intel.com>,
        Rodrigo Vivi <rodrigo.vivi@intel.com>,
        David Airlie <airlied@gmail.com>,
        Daniel Vetter <daniel@ffwll.ch>,
        intel-gfx@lists.freedesktop.org, dri-devel@lists.freedesktop.org,
        linux-kernel@vger.kernel.org,
        Chris Wilson <chris.p.wilson@linux.intel.com>,
        Chris Wilson <chris@chris-wilson.co.uk>
References: <20230330-hold_wakeref_for_active_vm-v2-1-724d201499c2@intel.com>
 <ZCxEoelnOSpRiy6A@ashyti-mobl2.lan>
 <f193eb7f-19f9-b9a8-7edf-438dafb74da4@linux.intel.com>
 <ZCxJitP+XtC2/run@ashyti-mobl2.lan>
From:   Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Organization: Intel Corporation UK Plc
In-Reply-To: <ZCxJitP+XtC2/run@ashyti-mobl2.lan>
Content-Type: text/plain; charset=UTF-8; format=flowed
Content-Transfer-Encoding: 7bit
Precedence: bulk


On 04/04/2023 17:00, Andi Shyti wrote:
> Hi Tvrtko,
> 
>>>> diff --git a/drivers/gpu/drm/i915/gt/intel_context.h b/drivers/gpu/drm/i915/gt/intel_context.h
>>>> index 0a8d553da3f439..48f888c3da083b 100644
>>>> --- a/drivers/gpu/drm/i915/gt/intel_context.h
>>>> +++ b/drivers/gpu/drm/i915/gt/intel_context.h
>>>> @@ -14,6 +14,7 @@
>>>>    #include "i915_drv.h"
>>>>    #include "intel_context_types.h"
>>>>    #include "intel_engine_types.h"
>>>> +#include "intel_gt_pm.h"
>>>>    #include "intel_ring_types.h"
>>>>    #include "intel_timeline_types.h"
>>>>    #include "i915_trace.h"
>>>> @@ -207,8 +208,11 @@ void intel_context_exit_engine(struct intel_context *ce);
>>>>    static inline void intel_context_enter(struct intel_context *ce)
>>>>    {
>>>>    	lockdep_assert_held(&ce->timeline->mutex);
>>>> -	if (!ce->active_count++)
>>>> -		ce->ops->enter(ce);
>>>> +	if (ce->active_count++)
>>>> +		return;
>>>> +
>>>> +	ce->ops->enter(ce);
>>>> +	intel_gt_pm_get(ce->vm->gt);
>>>>    }
>>>>    static inline void intel_context_mark_active(struct intel_context *ce)
>>>> @@ -222,8 +226,11 @@ static inline void intel_context_exit(struct intel_context *ce)
>>>>    {
>>>>    	lockdep_assert_held(&ce->timeline->mutex);
>>>>    	GEM_BUG_ON(!ce->active_count);
>>>> -	if (!--ce->active_count)
>>>> -		ce->ops->exit(ce);
>>>> +	if (--ce->active_count)
>>>> +		return;
>>>> +
>>>> +	intel_gt_pm_put_async(ce->vm->gt);
>>>> +	ce->ops->exit(ce);
>>>
>>> shouldn't these two be swapped?
> 
> maybe I wasn't clear here... shouldn't it be

I missed this one.

> 	ce->ops->exit(ce);
> 	intel_gt_pm_put_async(ce->vm->gt);
> 
> Don't we need to hold the pm until exiting?

I think it doesn't matter. The problematic edge case this is fixing is 
when ce->engine->gt is different from ce->vm->gt but at this point if it 
is safe to release one it must be safe to release the other too.

Regards,

Tvrtko


> 
>>>>    }
>>>>    static inline struct intel_context *intel_context_get(struct intel_context *ce)
>>>> diff --git a/drivers/gpu/drm/i915/gt/intel_engine_pm.c b/drivers/gpu/drm/i915/gt/intel_engine_pm.c
>>>> index e971b153fda976..ee531a5c142c77 100644
>>>> --- a/drivers/gpu/drm/i915/gt/intel_engine_pm.c
>>>> +++ b/drivers/gpu/drm/i915/gt/intel_engine_pm.c
>>>> @@ -114,6 +114,15 @@ __queue_and_release_pm(struct i915_request *rq,
>>>>    	ENGINE_TRACE(engine, "parking\n");
>>>> +	/*
>>>> +	 * Open coded one half of intel_context_enter, which we have to omit
>>>> +	 * here (see the large comment below) and because the other part must
>>>> +	 * not be called due constructing directly with __i915_request_create
>>>> +	 * which increments active count via intel_context_mark_active.
>>>> +	 */
>>>> +	GEM_BUG_ON(rq->context->active_count != 1);
>>>> +	__intel_gt_pm_get(engine->gt);
>>>
>>> where is it's brother "put"?
>>
>> It's in request retire via intel_context_exit. Ie. request construction is
>> special here, while retirement is standard.
> 
> Thank you!
> Andi