[PATCH] drm/amdgpu: Add sdma single packet invalidation

[Christian König]

Am 22.04.21 um 12:49 schrieb Victor Zhao:
> Add sdma single packet invalidation

Well NAK on many aspects of this.

>
> Signed-off-by: Victor Zhao <Victor.Zhao at amd.com>
> ---
>   drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c  |  8 ++++----
>   drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h |  4 ++++
>   drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c   | 14 +++++++++-----
>   drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c   | 12 ++++++++++++
>   4 files changed, 29 insertions(+), 9 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c
> index c39ed9eb0987..b2cf0e00ce38 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_gmc.c
> @@ -435,13 +435,13 @@ void amdgpu_gmc_ras_fini(struct amdgpu_device *adev)
>   
>   	/*
>   	 * The latest engine allocation on gfx9/10 is:
> -	 * Engine 2, 3: firmware
> -	 * Engine 0, 1, 4~16: amdgpu ring,
> +	 * Engine 1, 2, 3, 4: firmware
> +	 * Engine 0, 5~16: amdgpu ring,
>   	 *                    subject to change when ring number changes
>   	 * Engine 17: Gart flushes
>   	 */
> -#define GFXHUB_FREE_VM_INV_ENGS_BITMAP		0x1FFF3
> -#define MMHUB_FREE_VM_INV_ENGS_BITMAP		0x1FFF3
> +#define GFXHUB_FREE_VM_INV_ENGS_BITMAP		0x1FFE1
> +#define MMHUB_FREE_VM_INV_ENGS_BITMAP		0x1FFE1
>   
>   int amdgpu_gmc_allocate_vm_inv_eng(struct amdgpu_device *adev)
>   {
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
> index ca1622835296..fb3de16321ff 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
> @@ -192,6 +192,9 @@ struct amdgpu_ring_funcs {
>   	void (*emit_reg_write_reg_wait)(struct amdgpu_ring *ring,
>   					uint32_t reg0, uint32_t reg1,
>   					uint32_t ref, uint32_t mask);
> +	void (*vm_invalidate)(struct amdgpu_ring *ring,
> +				uint32_t reg0, uint32_t reg1,
> +				uint32_t req, uint32_t mask, uint32_t hub);

What is that good for? You need a much wider explanation if you want to 
change the design here.

Please rather implement the emit_reg_write_reg_wait() callback properly.

>   	void (*emit_frame_cntl)(struct amdgpu_ring *ring, bool start,
>   				bool secure);
>   	/* Try to soft recover the ring to make the fence signal */
> @@ -270,6 +273,7 @@ struct amdgpu_ring {
>   #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_write_reg_wait(r, d0, d1, v, m) (r)->funcs->emit_reg_write_reg_wait((r), (d0), (d1), (v), (m))
> +#define amdgpu_ring_vm_invalidate(r, d0, d1, v, m, h) ((r)->funcs->vm_invalidate((r), (d0), (d1), (v), (m), (h)))
>   #define amdgpu_ring_emit_frame_cntl(r, b, s) (r)->funcs->emit_frame_cntl((r), (b), (s))
>   #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/gmc_v10_0.c b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
> index 498b28a35f5b..0fdf60a7c53d 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gmc_v10_0.c
> @@ -481,11 +481,15 @@ static uint64_t gmc_v10_0_emit_flush_gpu_tlb(struct amdgpu_ring *ring,
>   			      (hub->ctx_addr_distance * vmid),
>   			      upper_32_bits(pd_addr));
>   
> -	amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req +
> -					    hub->eng_distance * eng,
> -					    hub->vm_inv_eng0_ack +
> -					    hub->eng_distance * eng,
> -					    req, 1 << vmid);
> +	if ((!strcmp("sdma0", ring->name)) || (!strcmp("sdma1", ring->name)))

Seriously?

> +		amdgpu_ring_vm_invalidate(ring, 0xffffffff, 0x1f,
> +					  req, 1 << vmid, 1 << ring->funcs->vmhub);
> +	else
> +		amdgpu_ring_emit_reg_write_reg_wait(ring, hub->vm_inv_eng0_req +
> +						    hub->eng_distance * eng,
> +						    hub->vm_inv_eng0_ack +
> +						    hub->eng_distance * eng,
> +						    req, 1 << vmid);
>   
>   	/* TODO: It needs to continue working on debugging with semaphore for GFXHUB as well. */
>   	if (use_semaphore)
> diff --git a/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c b/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
> index 920fc6d4a127..fd7df9194f05 100644
> --- a/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/sdma_v5_0.c
> @@ -1219,6 +1219,17 @@ static void sdma_v5_0_ring_emit_reg_write_reg_wait(struct amdgpu_ring *ring,
>   	amdgpu_ring_emit_reg_wait(ring, reg1, mask, mask);
>   }
>   
> +static void sdma_v5_0_ring_vm_invalidate(struct amdgpu_ring *ring,
> +		uint32_t reg0, uint32_t reg1,
> +		uint32_t req, uint32_t mask, uint32_t hub)
> +{
> +	amdgpu_ring_write(ring, SDMA_PKT_HEADER_OP(SDMA_OP_POLL_REGMEM) |
> +			SDMA_PKT_POLL_REGMEM_HEADER_SUB_OP(4));
> +	amdgpu_ring_write(ring, req);
> +	amdgpu_ring_write(ring, reg0);
> +	amdgpu_ring_write(ring, ((reg1 & 0x1f) << 16) | (mask & 0xffff) | (hub << 21)); /* reference */

That's not sufficient. That you use the single packet doesn't protects 
you from concurrent access of the invalidation control registers.

Christian.

> +}
> +
>   static int sdma_v5_0_early_init(void *handle)
>   {
>   	struct amdgpu_device *adev = (struct amdgpu_device *)handle;
> @@ -1654,6 +1665,7 @@ static const struct amdgpu_ring_funcs sdma_v5_0_ring_funcs = {
>   	.emit_wreg = sdma_v5_0_ring_emit_wreg,
>   	.emit_reg_wait = sdma_v5_0_ring_emit_reg_wait,
>   	.emit_reg_write_reg_wait = sdma_v5_0_ring_emit_reg_write_reg_wait,
> +	.vm_invalidate = sdma_v5_0_ring_vm_invalidate,
>   	.init_cond_exec = sdma_v5_0_ring_init_cond_exec,
>   	.patch_cond_exec = sdma_v5_0_ring_patch_cond_exec,
>   	.preempt_ib = sdma_v5_0_ring_preempt_ib,