[PATCH] drm/amdgpu: fix a kcq hang issue for SRIOV
Alex Deucher
alexdeucher at gmail.com
Tue Mar 27 15:52:15 UTC 2018
On Tue, Mar 27, 2018 at 11:43 AM, Christian König
<ckoenig.leichtzumerken at gmail.com> wrote:
> Am 27.03.2018 um 17:37 schrieb Alex Deucher:
>>
>> On Tue, Mar 27, 2018 at 11:15 AM, Alex Deucher <alexdeucher at gmail.com>
>> wrote:
>>>
>>> On Tue, Mar 27, 2018 at 1:58 AM, Emily Deng <Emily.Deng at amd.com> wrote:
>>>>
>>>> issue:
>>>> the vmflush in KCQ could be preempted (not like GFX ring
>>>> which doesn't allow preemption in ring buffer) and this lead
>>>> to vm flush fail when there is a world switch during
>>>> the vm flush procedure (between write invalidate request
>>>> and query invalidate ack)
>>>>
>>>> fix:
>>>> separate vm flush for gfx and compute ring, and use
>>>> the new format command in compute's vm flush which
>>>> use only one package so no preemption could allowed
>>>>
>>>> Signed-off-by: Monk Liu <Monk.Liu at amd.com>
>>>> Signed-off-by: Emily Deng <Emily.Deng at amd.com>
>>>> ---
>>>> drivers/gpu/drm/amd/amdgpu/amdgpu.h | 1 +
>>>> drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 2 ++
>>>> drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c | 10 +++++++++-
>>>> drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c | 18 +++++++++++++-----
>>>> 4 files changed, 25 insertions(+), 6 deletions(-)
>>>>
>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>>>> index a7e2229..986659f 100644
>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu.h
>>>> @@ -1790,6 +1790,7 @@ amdgpu_get_sdma_instance(struct amdgpu_ring *ring)
>>>> #define amdgpu_ring_emit_rreg(r, d) (r)->funcs->emit_rreg((r), (d))
>>>> #define amdgpu_ring_emit_wreg(r, d, v) (r)->funcs->emit_wreg((r), (d),
>>>> (v))
>>>> #define amdgpu_ring_emit_reg_wait(r, d, v, m)
>>>> (r)->funcs->emit_reg_wait((r), (d), (v), (m))
>>>> +#define amdgpu_ring_emit_reg_wait1(r, d0, d1, v, m)
>>>> (r)->funcs->emit_reg_wait1((r), (d0), (d1), (v), (m))
>>>> #define amdgpu_ring_emit_tmz(r, b) (r)->funcs->emit_tmz((r), (b))
>>>> #define amdgpu_ring_pad_ib(r, ib) ((r)->funcs->pad_ib((r), (ib)))
>>>> #define amdgpu_ring_init_cond_exec(r) (r)->funcs->init_cond_exec((r))
>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
>>>> b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
>>>> index 1d0d250..d85df5d 100644
>>>> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
>>>> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
>>>> @@ -152,6 +152,8 @@ struct amdgpu_ring_funcs {
>>>> void (*emit_wreg)(struct amdgpu_ring *ring, uint32_t reg,
>>>> uint32_t val);
>>>> void (*emit_reg_wait)(struct amdgpu_ring *ring, uint32_t reg,
>>>> uint32_t val, uint32_t mask);
>>>> + void (*emit_reg_wait1)(struct amdgpu_ring *ring, uint32_t reg0,
>>>> + uint32_t reg1, uint32_t val, uint32_t
>>>> mask);
>>>
>>> This is pretty ugly. We don't need a new callback for this
>>> workaround. It will just confuse things for other IPs. Just call
>>> gfx_v9_0_wait_reg_mem() directly if you need it in
>>> gmc_v9_0_emit_flush_gpu_tlb().
>>
>> Nevermind, I misread the patch and thought the change was in the gfx9
>> module. In this case, I'd rather split this into three patches.
>>
>> 1. add the new ring callback with a better name. E.g.,
>> emit_reg_wait_oneshot()
>
>
> Maybe "emit_reg_write_wait()", but I still think that is an absolutely worse
> idea.
Well apparently the issue is that world switches can't happen between
the req and the ack, so this would consolidate the operation to one
packet which should avoid that issue.
>
>> 2. add the new callback implementation to gfx9 and gfx8 (I think gfx8
>> will need this as well since we support sr-iov there too)
>
>
> gfx8 doesn't have the hardware bug which seems to make this necessary, not
> does it have the same VMHUB design as gfx9.
Oh, right, in this case it's the req/ack engines which were new for
soc15. We may want the same fix for sdma4 though.
>
>> 3. in the gmc8 and 9 code tlb_flush callbacks, check if the ring
>> supports the new callback and if so, use that, otherwise, fall back to
>> the existing code.
>
>
> Actually I would rather say we provide a stub in amdgpu_ring.c which just
> uses the separate write and wait callbacks.
yeah, that may be cleaner.
Alex
>
> Christian.
>
>
>>
>>
>>> Alex
>>>
>>>> void (*emit_tmz)(struct amdgpu_ring *ring, bool start);
>>>> /* priority functions */
>>>> void (*set_priority) (struct amdgpu_ring *ring,
>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c
>>>> b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c
>>>> index 1ae3de1..509c9d2 100644
>>>> --- a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c
>>>> +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c
>>>> @@ -4078,6 +4078,13 @@ static void gfx_v9_0_ring_emit_reg_wait(struct
>>>> amdgpu_ring *ring, uint32_t reg,
>>>> gfx_v9_0_wait_reg_mem(ring, 0, 0, 0, reg, 0, val, mask, 0x20);
>>>> }
>>>>
>>>> +static void gfx_v9_0_ring_emit_reg_wait_compute(struct amdgpu_ring
>>>> *ring,
>>>> + uint32_t reg0, uint32_t reg1,
>>>> + uint32_t val, uint32_t mask)
>>>> +{
>>>> + gfx_v9_0_wait_reg_mem(ring, 0, 0, 1, reg0, reg1, val, mask,
>>>> 0x20);
>>>> +}
>>>> +
>>>> static void gfx_v9_0_set_gfx_eop_interrupt_state(struct amdgpu_device
>>>> *adev,
>>>> enum
>>>> amdgpu_interrupt_state state)
>>>> {
>>>> @@ -4415,7 +4422,7 @@ static const struct amdgpu_ring_funcs
>>>> gfx_v9_0_ring_funcs_compute = {
>>>> 7 + /* gfx_v9_0_ring_emit_hdp_flush */
>>>> 5 + /* hdp invalidate */
>>>> 7 + /* gfx_v9_0_ring_emit_pipeline_sync */
>>>> - SOC15_FLUSH_GPU_TLB_NUM_WREG * 5 +
>>>> + (SOC15_FLUSH_GPU_TLB_NUM_WREG - 1) * 5 +
>>>> SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 7 +
>>>> 2 + /* gfx_v9_0_ring_emit_vm_flush */
>>>> 8 + 8 + 8, /* gfx_v9_0_ring_emit_fence x3 for user
>>>> fence, vm fence */
>>>> @@ -4433,6 +4440,7 @@ static const struct amdgpu_ring_funcs
>>>> gfx_v9_0_ring_funcs_compute = {
>>>> .set_priority = gfx_v9_0_ring_set_priority_compute,
>>>> .emit_wreg = gfx_v9_0_ring_emit_wreg,
>>>> .emit_reg_wait = gfx_v9_0_ring_emit_reg_wait,
>>>> + .emit_reg_wait1 = gfx_v9_0_ring_emit_reg_wait_compute,
>>>> };
>>>>
>>>> static const struct amdgpu_ring_funcs gfx_v9_0_ring_funcs_kiq = {
>>>> diff --git a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
>>>> b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
>>>> index e687363..968447d 100644
>>>> --- a/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
>>>> +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v9_0.c
>>>> @@ -385,11 +385,19 @@ static uint64_t gmc_v9_0_emit_flush_gpu_tlb(struct
>>>> amdgpu_ring *ring,
>>>> amdgpu_ring_emit_wreg(ring, hub->ctx0_ptb_addr_hi32 + (2 *
>>>> vmid),
>>>> upper_32_bits(pd_addr));
>>>>
>>>> - amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_req + eng, req);
>>>> -
>>>> - /* wait for the invalidate to complete */
>>>> - amdgpu_ring_emit_reg_wait(ring, hub->vm_inv_eng0_ack + eng,
>>>> - 1 << vmid, 1 << vmid);
>>>> + /* The world switch cannot be allowed to occur while
>>>> + some invalidation controller code is waiting for an ack.
>>>> + To workaround the hardware restriction, replace the original
>>>> + two command with one command for compute ring */
>>>> + if (ring->funcs->type == AMDGPU_RING_TYPE_COMPUTE &&
>>>> amdgpu_sriov_vf(adev)) {
>>
>> This will break gfx 8 with sr-iov.
>>
>> Alex
>>
>>>> + amdgpu_ring_emit_reg_wait1(ring, hub->vm_inv_eng0_req +
>>>> eng,
>>>> + hub->vm_inv_eng0_ack + eng, req, 1 <<
>>>> vmid);
>>>> + } else {
>>>> + amdgpu_ring_emit_wreg(ring, hub->vm_inv_eng0_req + eng,
>>>> req);
>>>> + /* wait for the invalidate to complete */
>>>> + amdgpu_ring_emit_reg_wait(ring, hub->vm_inv_eng0_ack +
>>>> eng,
>>>> + 1 << vmid, 1 << vmid);
>>>> + }
>>>>
>>>> return pd_addr;
>>>> }
>>>> --
>>>> 2.7.4
>>>>
>>>> _______________________________________________
>>>> amd-gfx mailing list
>>>> amd-gfx at lists.freedesktop.org
>>>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
>>
>> _______________________________________________
>> amd-gfx mailing list
>> amd-gfx at lists.freedesktop.org
>> https://lists.freedesktop.org/mailman/listinfo/amd-gfx
>
>
More information about the amd-gfx
mailing list