[PATCH 3/7] drm/amdgpu: implement tlbs invalidate on gfx9 gfx10

[Felix Kuehling]

On 2020-01-11 1:39 p.m., Alex Sierra wrote:
> tlbs invalidate pointer function added to kiq_pm4_funcs struct.
> This way, tlb flush can be done through kiq member.
> TLBs invalidatation implemented for gfx9 and gfx10.
>
> Change-Id: I1b77b364f3ae0038ff3e70e869be5f2ef6e6d293
> Signed-off-by: Alex Sierra <alex.sierra at amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h |  4 +++
>   drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c  | 33 +++++++++++++++++++++++++
>   drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c   | 32 ++++++++++++++++++++++++
>   3 files changed, 69 insertions(+)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h
> index 8e88e0411662..2927837bd401 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gfx.h
> @@ -76,11 +76,15 @@ struct kiq_pm4_funcs {
>   					struct amdgpu_ring *ring,
>   					u64 addr,
>   					u64 seq);
> +	int (*kiq_invalidate_tlbs)(struct amdgpu_ring *kiq_ring,
> +				uint16_t pasid, uint32_t flush_type,
> +				bool all_hub);
>   	/* Packet sizes */
>   	int set_resources_size;
>   	int map_queues_size;
>   	int unmap_queues_size;
>   	int query_status_size;
> +	int invalidate_tlbs_size;
>   };
>   
>   struct amdgpu_kiq {
> diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c
> index 379e46c1b7f6..2e82213f57eb 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v10_0.c
> @@ -40,6 +40,7 @@
>   #include "ivsrcid/gfx/irqsrcs_gfx_10_1.h"
>   
>   #include "soc15.h"
> +#include "soc15d.h"
>   #include "soc15_common.h"
>   #include "clearstate_gfx10.h"
>   #include "v10_structs.h"
> @@ -346,15 +347,47 @@ static void gfx10_kiq_query_status(struct amdgpu_ring *kiq_ring,
>   	amdgpu_ring_write(kiq_ring, upper_32_bits(seq));
>   }
>   
> +static int gfx10_kiq_invalidate_tlbs(struct amdgpu_ring *kiq_ring,
> +				uint16_t pasid, uint32_t flush_type,
> +				bool all_hub)
> +{
> +	signed long r;
> +	uint32_t seq;
> +	struct amdgpu_device *adev = kiq_ring->adev;
> +	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
> +
> +	spin_lock(&adev->gfx.kiq.ring_lock);
> +	amdgpu_ring_alloc(kiq_ring, kiq->pmf->invalidate_tlbs_size);
> +	amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_INVALIDATE_TLBS, 0));
> +	amdgpu_ring_write(kiq_ring,
> +			PACKET3_INVALIDATE_TLBS_DST_SEL(1) |
> +			PACKET3_INVALIDATE_TLBS_ALL_HUB(all_hub) |
> +			PACKET3_INVALIDATE_TLBS_PASID(pasid) |
> +			PACKET3_INVALIDATE_TLBS_FLUSH_TYPE(flush_type));
> +	amdgpu_fence_emit_polling(kiq_ring, &seq);
> +	amdgpu_ring_commit(kiq_ring);
> +	spin_unlock(&adev->gfx.kiq.ring_lock);

The other KIQ functions don't include the emit_polling, commit and 
locking. I think the way the KIQ-funcs interface is meant to be used, 
all that should be outside the IP-version-specific functions. For 
consistency all you should do here is the amdgpu_ring_write calls with 
IP-version-specific packets.

Regards,
   Felix


> +
> +	r = amdgpu_fence_wait_polling(kiq_ring, seq, adev->usec_timeout);
> +	if (r < 1) {
> +		DRM_ERROR("wait for kiq fence error: %ld.\n", r);
> +		return -ETIME;
> +	}
> +
> +	return 0;
> +}
> +
>   static const struct kiq_pm4_funcs gfx_v10_0_kiq_pm4_funcs = {
>   	.kiq_set_resources = gfx10_kiq_set_resources,
>   	.kiq_map_queues = gfx10_kiq_map_queues,
>   	.kiq_unmap_queues = gfx10_kiq_unmap_queues,
>   	.kiq_query_status = gfx10_kiq_query_status,
> +	.kiq_invalidate_tlbs = gfx10_kiq_invalidate_tlbs,
>   	.set_resources_size = 8,
>   	.map_queues_size = 7,
>   	.unmap_queues_size = 6,
>   	.query_status_size = 7,
> +	.invalidate_tlbs_size = 12,
>   };
>   
>   static void gfx_v10_0_set_kiq_pm4_funcs(struct amdgpu_device *adev)
> diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c
> index ad0179ea2cc5..5be6fab55b73 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v9_0.c
> @@ -837,15 +837,47 @@ static void gfx_v9_0_kiq_query_status(struct amdgpu_ring *kiq_ring,
>   	amdgpu_ring_write(kiq_ring, upper_32_bits(seq));
>   }
>   
> +static int gfx_v9_0_kiq_invalidate_tlbs(struct amdgpu_ring *kiq_ring,
> +				uint16_t pasid, uint32_t flush_type,
> +				bool all_hub)
> +{
> +	signed long r;
> +	uint32_t seq;
> +	struct amdgpu_device *adev = kiq_ring->adev;
> +	struct amdgpu_kiq *kiq = &adev->gfx.kiq;
> +
> +	spin_lock(&adev->gfx.kiq.ring_lock);
> +	amdgpu_ring_alloc(kiq_ring, kiq->pmf->invalidate_tlbs_size);
> +	amdgpu_ring_write(kiq_ring, PACKET3(PACKET3_INVALIDATE_TLBS, 0));
> +	amdgpu_ring_write(kiq_ring,
> +			PACKET3_INVALIDATE_TLBS_DST_SEL(1) |
> +			PACKET3_INVALIDATE_TLBS_ALL_HUB(all_hub) |
> +			PACKET3_INVALIDATE_TLBS_PASID(pasid) |
> +			PACKET3_INVALIDATE_TLBS_FLUSH_TYPE(flush_type));
> +	amdgpu_fence_emit_polling(kiq_ring, &seq);
> +	amdgpu_ring_commit(kiq_ring);
> +	spin_unlock(&adev->gfx.kiq.ring_lock);
> +
> +	r = amdgpu_fence_wait_polling(kiq_ring, seq, adev->usec_timeout);
> +	if (r < 1) {
> +		DRM_ERROR("wait for kiq fence error: %ld.\n", r);
> +		return -ETIME;
> +	}
> +
> +	return 0;
> +}
> +
>   static const struct kiq_pm4_funcs gfx_v9_0_kiq_pm4_funcs = {
>   	.kiq_set_resources = gfx_v9_0_kiq_set_resources,
>   	.kiq_map_queues = gfx_v9_0_kiq_map_queues,
>   	.kiq_unmap_queues = gfx_v9_0_kiq_unmap_queues,
>   	.kiq_query_status = gfx_v9_0_kiq_query_status,
> +	.kiq_invalidate_tlbs = gfx_v9_0_kiq_invalidate_tlbs,
>   	.set_resources_size = 8,
>   	.map_queues_size = 7,
>   	.unmap_queues_size = 6,
>   	.query_status_size = 7,
> +	.invalidate_tlbs_size = 12,
>   };
>   
>   static void gfx_v9_0_set_kiq_pm4_funcs(struct amdgpu_device *adev)