[Intel-gfx] [PATCH v3 3/3] drm/doc/rfc: VM_BIND uapi definition

Tvrtko Ursulin tvrtko.ursulin at linux.intel.com
Mon Jun 27 08:30:00 UTC 2022


On 24/06/2022 21:23, Zeng, Oak wrote:
> Let's compare "tlb invalidate at vm unbind" vs "tlb invalidate at backing storage":
> 
> Correctness:
> consider this sequence of:
> 1. unbind va1 from pa1,
> 2. then bind va1 to pa2. //user space has the freedom to do this as it manages virtual address space
> 3. Submit shader code using va1,
> 4. Then retire pa1.
> 
> If you don't perform tlb invalidate at step #1, in step #3, shader will use stale entries in tlb and pa1 will be used for the shader. User want to use pa2. So I don't think invalidate tlb at step #4 make correctness.

Define step 3. Is it a new execbuf? If so then there will be a TLB flush 
there. Unless the plan is to stop doing that with eb3 but I haven't 
picked up on that anywhere so far.

> Performance:
> It is straight forward to invalidate tlb at step 1. If platform support range based tlb invalidation, we can perform range based invalidation easily at step1.

If the platform supports range base yes. If it doesn't _and_ the flush 
at unbind is not needed for 99% of use cases then it is simply a waste.

> If you do it at step 4, you either need to perform a whole gt tlb invalidation (worse performance), or you need to record all the VAs that this pa has been bound to and invalidate all the VA ranges - ugly program.

Someone can setup some benchmarking? :)

Regards,

Tvrtko

> 
> 
> Thanks,
> Oak
> 
>> -----Original Message-----
>> From: Tvrtko Ursulin <tvrtko.ursulin at linux.intel.com>
>> Sent: June 24, 2022 4:32 AM
>> To: Zeng, Oak <oak.zeng at intel.com>; Landwerlin, Lionel G
>> <lionel.g.landwerlin at intel.com>; Vishwanathapura, Niranjana
>> <niranjana.vishwanathapura at intel.com>
>> Cc: Zanoni, Paulo R <paulo.r.zanoni at intel.com>; intel-
>> gfx at lists.freedesktop.org; dri-devel at lists.freedesktop.org; Hellstrom,
>> Thomas <thomas.hellstrom at intel.com>; Wilson, Chris P
>> <chris.p.wilson at intel.com>; Vetter, Daniel <daniel.vetter at intel.com>;
>> christian.koenig at amd.com; Auld, Matthew <matthew.auld at intel.com>
>> Subject: Re: [Intel-gfx] [PATCH v3 3/3] drm/doc/rfc: VM_BIND uapi definition
>>
>>
>> On 23/06/2022 22:05, Zeng, Oak wrote:
>>>> -----Original Message-----
>>>> From: Intel-gfx <intel-gfx-bounces at lists.freedesktop.org> On Behalf
>>>> Of Tvrtko Ursulin
>>>> Sent: June 23, 2022 7:06 AM
>>>> To: Landwerlin, Lionel G <lionel.g.landwerlin at intel.com>;
>>>> Vishwanathapura, Niranjana <niranjana.vishwanathapura at intel.com>
>>>> Cc: Zanoni, Paulo R <paulo.r.zanoni at intel.com>;
>>>> intel-gfx at lists.freedesktop.org; dri-devel at lists.freedesktop.org;
>>>> Hellstrom, Thomas <thomas.hellstrom at intel.com>; Wilson, Chris P
>>>> <chris.p.wilson at intel.com>; Vetter, Daniel <daniel.vetter at intel.com>;
>>>> christian.koenig at amd.com; Auld, Matthew <matthew.auld at intel.com>
>>>> Subject: Re: [Intel-gfx] [PATCH v3 3/3] drm/doc/rfc: VM_BIND uapi
>>>> definition
>>>>
>>>>
>>>> On 23/06/2022 09:57, Lionel Landwerlin wrote:
>>>>> On 23/06/2022 11:27, Tvrtko Ursulin wrote:
>>>>>>>
>>>>>>> After a vm_unbind, UMD can re-bind to same VA range against an
>>>>>>> active VM.
>>>>>>> Though I am not sue with Mesa usecase if that new mapping is
>>>>>>> required for running GPU job or it will be for the next
>>>>>>> submission. But ensuring the tlb flush upon unbind, KMD can ensure
>>>>>>> correctness.
>>>>>>
>>>>>> Isn't that their problem? If they re-bind for submitting _new_ work
>>>>>> then they get the flush as part of batch buffer pre-amble.
>>>>>
>>>>> In the non sparse case, if a VA range is unbound, it is invalid to
>>>>> use that range for anything until it has been rebound by something else.
>>>>>
>>>>> We'll take the fence provided by vm_bind and put it as a wait fence
>>>>> on the next execbuffer.
>>>>>
>>>>> It might be safer in case of memory over fetching?
>>>>>
>>>>>
>>>>> TLB flush will have to happen at some point right?
>>>>>
>>>>> What's the alternative to do it in unbind?
>>>>
>>>> Currently TLB flush happens from the ring before every BB_START and
>>>> also when i915 returns the backing store pages to the system.
>>>
>>>
>>> Can you explain more why tlb flush when i915 retire the backing storage? I
>> never figured that out when I looked at the codes. As I understand it, tlb
>> caches the gpu page tables which map a va to a pa. So it is straight forward to
>> me that we perform a tlb flush when we change the page table (either at vm
>> bind time or unbind time. Better at unbind time for performance reason).
>>
>> I don't know what performs better - someone can measure the two
>> approaches? Certainly on platforms where we only have global TLB flushing
>> the cost is quite high so my thinking was to allow i915 to control when it will
>> be done and not guarantee it in the uapi if it isn't needed for security reasons.
>>
>>> But it is rather tricky to me to flush tlb when we retire a backing storage. I
>> don't see how backing storage can be connected to page table. Let's say user
>> unbind va1 from pa1, then bind va1 to pa2. Then retire pa1. Submit shader
>> code using va1. If we don't tlb flush after unbind va1, the new shader code
>> which is supposed to use pa2 will still use pa1 due to the stale entries in tlb,
>> right? The point is, tlb cached is tagged with virtual address, not physical
>> address. so after we unbind va1 from pa1, regardless we retire pa1 or not,
>> va1 can be bound to another pa2.
>>
>> When you say "retire pa1" I will assume you meant release backing storage
>> for pa1. At this point i915 currently does do the TLB flush and that ensures no
>> PTE can point to pa1.
>>
>> This approach deals with security of the system as a whole. Client may still
>> cause rendering corruption or a GPU hang for itself but that should be
>> completely isolated. (This is the part where you say "regardless if we retire
>> pa1 or not" I think.)
>>
>> But I think those are advanced use cases where userspace wants to
>> manipulate PTEs while something is running on the GPU in parallel. AFAIK
>> limited to compute "infinite batch" so my thinking is to avoid adding a
>> performance penalty to the common case. Especially on platforms which only
>> have global flush.
>>
>> But.. to circle back on the measuring angle. Until someone invests time and
>> effort to benchmark the two approaches (flush on unbind vs flush on backing
>> store release) we don't really know. All I know is the perf hit with the current
>> solution was significant, AFAIR up to teen digits on some games. And
>> considering the flushes were driven only by the shrinker activity, my thinking
>> was they would be less frequent than the unbinds, therefore have the
>> potential for a smaller perf hit.
>>
>> Regards,
>>
>> Tvrtko


More information about the dri-devel mailing list