[PATCH] drm/amdgpu: add ih call to process until checkpoint
Felix Kuehling
felix.kuehling at amd.com
Thu Feb 25 18:33:36 UTC 2021
Am 2021-02-25 um 11:48 a.m. schrieb Christian König:
>
>
> Am 25.02.21 um 16:35 schrieb Felix Kuehling:
>> Am 2021-02-25 um 8:53 a.m. schrieb Christian König:
>>>
>>> Am 25.02.21 um 04:15 schrieb Felix Kuehling:
>>>> On 2021-02-24 10:54 a.m., Kim, Jonathan wrote:
>>>>> [AMD Official Use Only - Internal Distribution Only]
>>>>>
>>>>>> -----Original Message-----
>>>>>> From: Koenig, Christian <Christian.Koenig at amd.com>
>>>>>> Sent: Wednesday, February 24, 2021 4:17 AM
>>>>>> To: Kim, Jonathan <Jonathan.Kim at amd.com>; amd-
>>>>>> gfx at lists.freedesktop.org
>>>>>> Cc: Yang, Philip <Philip.Yang at amd.com>; Kuehling, Felix
>>>>>> <Felix.Kuehling at amd.com>
>>>>>> Subject: Re: [PATCH] drm/amdgpu: add ih call to process until
>>>>>> checkpoint
>>>>>>
>>>>>> Am 23.02.21 um 22:10 schrieb Jonathan Kim:
>>>>>>> Add IH function to allow caller to process ring entries until the
>>>>>>> checkpoint write pointer.
>>>>>> This needs a better description of what this will be used for.
>>>>> Felix or Philip could elaborate better for HMM needs.
>>>>> Debugging tools requires this but it's in experimental mode at the
>>>>> moment so probably not the best place to describe here.
>>>> On the HMM side we're planning to use this to drain pending page
>>>> fault interrupts before we unmap memory. That should address phantom
>>>> VM faults after memory is unmapped.
>>> Thought so. I suggest to use a wait_event() here which on the waiter
>>> side checks ih->lock and add a wake_up_all() at the end of
>>> amdgpu_ih_process.
>> Right. I thought about that and it should be easy to add. The reason to
>> suggest busy waiting first is, that interrupt processing is supposed to
>> be fast. The IRQ handler itself doesn't sleep. So I'd expect the wait
>> time to be short enough that sleeping and scheduling is not worth it.
>
> Well the page fault IRQs are processed in a work item, so we busy wait
> for another thread here and not interrupt context.
Good point.
>
> This in turn can lead to starvation of the work handler and so a life
> lock as well.
>
>>
>>
>>> I won't touch rptr or wptr at all for this.
>> Not sure what's your idea here, using ih->lock. Is it to completely
>> drain all IRQs until the IH ring is completely empty?
>
> Correct.
>
>> That can
>> live-lock, if the GPU produces IRQs faster than the kernel can process
>> them. Therefore I was looking at rptr and wptr to drain only IRQs that
>> were already in the queue when the drain call was made. That also
>> ensures that the wait time is bounded and should be short (unless the
>> ring size is huge).
>
> Correct as well, but the problem here is that Jonathan's
> implementation is not even remotely correct.
>
> See when you look at the rptr and wptr you can't be sure that they
> haven't wrapped around between two looks.
>
> What you could do is look at both the rptr as well as the original
> wptr, but that is tricky.
The code in Jon's patch was suggested by me. I check for wrapping by
comparing rptr with the rptr from the previous loop iteration. Comparing
rptr with wptr you can never be sure whether rptr has already passed
wptr, or whether rptr has to wrap first.
I could see a compromise where we sleep and wake up the waiting threads when
1. the IH ring is empty
2. the IH rptr wraps
That should be good enough to ensure a quick response in the common case
(no interrupt storm), and a reasonable response in the interrupt storm case.
But then it's still not clear what's the correct condition for checking
that the interrupts the caller cares about have been drained. Looking at
just rptr and wptr is always ambiguous. Maybe we could use timestamps of
the last processed interrupt? Those 48-bit time stamps wrap much less
frequently. The idea is this:
* At the start get the timestamp of the last written IH ring entry
(checkpoint)
* Wait until the last_processed IH timestamp passes the checkpoint:
sign_extend(last_processed - checkpoint) >= 0
Regards,
Felix
>
> Regards,
> Christian.
>
>>
>> Regards,
>> Felix
>>
>>
>>> Regards,
>>> Christian.
>>>
>>>> Regards,
>>>> Felix
>>>>
>>>>
>>>>>>> Suggested-by: Felix Kuehling <felix.kuehling at amd.com>
>>>>>>> Signed-off-by: Jonathan Kim <jonathan.kim at amd.com>
>>>>>>> ---
>>>>>>> drivers/gpu/drm/amd/amdgpu/amdgpu_ih.c | 46
>>>>>> +++++++++++++++++++++++++-
>>>>>>> drivers/gpu/drm/amd/amdgpu/amdgpu_ih.h | 2 ++
>>>>>>> 2 files changed, 47 insertions(+), 1 deletion(-)
>>>>>>>
>>>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ih.c
>>>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ih.c
>>>>>>> index dc852af4f3b7..cae50af9559d 100644
>>>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ih.c
>>>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ih.c
>>>>>>> @@ -22,7 +22,7 @@
>>>>>>> */
>>>>>>>
>>>>>>> #include <linux/dma-mapping.h>
>>>>>>> -
>>>>>>> +#include <linux/processor.h>
>>>>>>> #include "amdgpu.h"
>>>>>>> #include "amdgpu_ih.h"
>>>>>>>
>>>>>>> @@ -160,6 +160,50 @@ void amdgpu_ih_ring_write(struct
>>>>>> amdgpu_ih_ring *ih, const uint32_t *iv,
>>>>>>> }
>>>>>>> }
>>>>>>>
>>>>>>> +/**
>>>>>>> + * amdgpu_ih_wait_on_checkpoint_process - wait to process IVs
>>>>>>> up to
>>>>>>> +checkpoint
>>>>>>> + *
>>>>>>> + * @adev: amdgpu_device pointer
>>>>>>> + * @ih: ih ring to process
>>>>>>> + *
>>>>>>> + * Used to ensure ring has processed IVs up to the checkpoint
>>>>>>> write
>>>>>> pointer.
>>>>>>> + */
>>>>>>> +int amdgpu_ih_wait_on_checkpoint_process(struct amdgpu_device
>>>>>> *adev,
>>>>>>> +struct amdgpu_ih_ring *ih)
>>>>>>> +{
>>>>>>> +u32 prev_rptr, cur_rptr, checkpoint_wptr;
>>>>>>> +
>>>>>>> +if (!ih->enabled || adev->shutdown)
>>>>>>> +return -ENODEV;
>>>>>>> +
>>>>>>> +cur_rptr = READ_ONCE(ih->rptr);
>>>>>>> +/* Order read of current rptr with checktpoint wptr. */
>>>>>>> +mb();
>>>>>>> +checkpoint_wptr = amdgpu_ih_get_wptr(adev, ih);
>>>>>>> +
>>>>>>> +/* allow rptr to wrap around */
>>>>>>> +if (cur_rptr > checkpoint_wptr) {
>>>>>>> +spin_begin();
>>>>>>> +do {
>>>>>>> +spin_cpu_relax();
>>>>>>> +prev_rptr = cur_rptr;
>>>>>>> +cur_rptr = READ_ONCE(ih->rptr);
>>>>>>> +} while (cur_rptr >= prev_rptr);
>>>>>>> +spin_end();
>>>>>> That's a certain NAK since it busy waits for IH processing. We need
>>>>>> some
>>>>>> event to trigger here.
>>>>> The function is meant to be just a waiter up to the checkpoint.
>>>>> There's a need to guarantee that "stale" interrupts have been
>>>>> processed on check before doing other stuff after call.
>>>>> The description could be improved to clarify that.
>>>>>
>>>>> Would busy waiting only on a locked ring help? I assume an unlocked
>>>>> ring means nothing to process so no need to wait and we can exit
>>>>> early. Or is it better to just to process the entries up to the
>>>>> checkpoint (maybe adjust amdgpu_ih_process for this need like adding
>>>>> a bool arg to skip restart or something)?
>>>>>
>>>>> Thanks,
>>>>>
>>>>> Jon
>>>>>
>>>>>>> +}
>>>>>>> +
>>>>>>> +/* wait for rptr to catch up to or pass checkpoint. */
>>>>>>> +spin_begin();
>>>>>>> +do {
>>>>>>> +spin_cpu_relax();
>>>>>>> +prev_rptr = cur_rptr;
>>>>>>> +cur_rptr = READ_ONCE(ih->rptr);
>>>>>>> +} while (cur_rptr >= prev_rptr && cur_rptr < checkpoint_wptr);
>>>>>> Same of course here.
>>>>>>
>>>>>> Christian.
>>>>>>
>>>>>>> +spin_end();
>>>>>>> +
>>>>>>> +return 0;
>>>>>>> +}
>>>>>>> +
>>>>>>> /**
>>>>>>> * amdgpu_ih_process - interrupt handler
>>>>>>> *
>>>>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ih.h
>>>>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ih.h
>>>>>>> index 6ed4a85fc7c3..6817f0a812d2 100644
>>>>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ih.h
>>>>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ih.h
>>>>>>> @@ -87,6 +87,8 @@ int amdgpu_ih_ring_init(struct amdgpu_device
>>>>>>> *adev,
>>>>>> struct amdgpu_ih_ring *ih,
>>>>>>> void amdgpu_ih_ring_fini(struct amdgpu_device *adev, struct
>>>>>> amdgpu_ih_ring *ih);
>>>>>>> void amdgpu_ih_ring_write(struct amdgpu_ih_ring *ih, const
>>>>>>> uint32_t *iv,
>>>>>>> unsigned int num_dw);
>>>>>>> +int amdgpu_ih_wait_on_checkpoint_process(struct amdgpu_device
>>>>>> *adev,
>>>>>>> +struct amdgpu_ih_ring *ih);
>>>>>>> int amdgpu_ih_process(struct amdgpu_device *adev, struct
>>>>>> amdgpu_ih_ring *ih);
>>>>>>> void amdgpu_ih_decode_iv_helper(struct amdgpu_device *adev,
>>>>>>> struct amdgpu_ih_ring *ih,
>>>> _______________________________________________
>>>> amd-gfx mailing list
>>>> amd-gfx at lists.freedesktop.org
>>>> https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&data=04%7C01%7Cfelix.kuehling%40amd.com%7C84d85e54bdcb4593e07f08d8d994be77%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637498580167313193%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C1000&sdata=RvRHB9l4O%2BpbpogUFKUmnMGkqKnecwQCYRHrkxICDqU%3D&reserved=0
>>>>
>>>>
>
More information about the amd-gfx
mailing list