[PATCH] drm/amdgpu:guarantee 128dws after vm flush (v5)
Christian König
deathsimple at vodafone.de
Mon Jan 23 13:15:28 UTC 2017
Am 22.01.2017 um 04:55 schrieb Monk Liu:
> previously we always insert 128nops behind vm_flush, which
> may lead to DAMframe size above 256 dw and automatially aligned
> to 512 dw.
>
> now we calculate how many DWs already inserted after vm_flush
> and make up for the reset to pad up to 128dws before emit_ib.
> that way we only take 256 dw per submit.
>
> v2:
> drop insert_nops in vm_flush
> re-calculate the estimate frame size for gfx8 ring
> v3:
> calculate the gap between vm_flush and IB in cntx_cntl
> on an member field of ring structure
> v4:
> set last_vm_flush_pos even for case of no vm flush.
> v5:
> move pipeline sync out of VM flush and always insert it.
Didn't we've already found that always doing the pipeline sync also has
some negative impact on performance as well?
I wonder if there isn't any better way of handling this. How about using
a separate frame for the VM flush as well?
If you don't have time to take a deeper look and this really fixes a bug
for you, feel free to add my Acked-by and commit it.
We can sort out the performance problems later on.
Regards,
Christian.
>
> Change-Id: I670adf8c5e7bdfe5f4fd191c24f62a5d7a170d3a
> Signed-off-by: Monk Liu <Monk.Liu at amd.com>
> ---
> drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c | 8 +++++++-
> drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h | 1 +
> drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 31 -------------------------------
> drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c | 19 +++++++++++++++----
> 4 files changed, 23 insertions(+), 36 deletions(-)
>
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
> index ff39858..a1cd429 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ib.c
> @@ -154,6 +154,7 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
>
> alloc_size = ring->funcs->emit_frame_size + num_ibs *
> ring->funcs->emit_ib_size;
> + need_ctx_switch = ring->current_ctx != fence_ctx;
>
> r = amdgpu_ring_alloc(ring, alloc_size);
> if (r) {
> @@ -164,7 +165,12 @@ int amdgpu_ib_schedule(struct amdgpu_ring *ring, unsigned num_ibs,
> if (ring->funcs->init_cond_exec)
> patch_offset = amdgpu_ring_init_cond_exec(ring);
>
> - need_ctx_switch = ring->current_ctx != fence_ctx;
> + /* pipeline sync is to prevent following DE change SH register
> + * before previous submit totally done (EOP signaled)
> + */
> + if (ring->funcs->emit_pipeline_sync)
> + amdgpu_ring_emit_pipeline_sync(ring);
> +
> if (vm) {
> r = amdgpu_vm_flush(ring, job);
> if (r) {
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
> index e0f8061..4c546fe 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h
> @@ -173,6 +173,7 @@ struct amdgpu_ring {
> #if defined(CONFIG_DEBUG_FS)
> struct dentry *ent;
> #endif
> + u32 last_vm_flush_pos;
> };
>
> int amdgpu_ring_alloc(struct amdgpu_ring *ring, unsigned ndw);
> diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> index d05546e..4c1102f 100644
> --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
> @@ -343,32 +343,6 @@ int amdgpu_vm_grab_id(struct amdgpu_vm *vm, struct amdgpu_ring *ring,
> return r;
> }
>
> -static bool amdgpu_vm_ring_has_compute_vm_bug(struct amdgpu_ring *ring)
> -{
> - struct amdgpu_device *adev = ring->adev;
> - const struct amdgpu_ip_block *ip_block;
> -
> - if (ring->funcs->type != AMDGPU_RING_TYPE_COMPUTE)
> - /* only compute rings */
> - return false;
> -
> - ip_block = amdgpu_get_ip_block(adev, AMD_IP_BLOCK_TYPE_GFX);
> - if (!ip_block)
> - return false;
> -
> - if (ip_block->version->major <= 7) {
> - /* gfx7 has no workaround */
> - return true;
> - } else if (ip_block->version->major == 8) {
> - if (adev->gfx.mec_fw_version >= 673)
> - /* gfx8 is fixed in MEC firmware 673 */
> - return false;
> - else
> - return true;
> - }
> - return false;
> -}
> -
> /**
> * amdgpu_vm_flush - hardware flush the vm
> *
> @@ -391,11 +365,6 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job)
> id->oa_size != job->oa_size);
> int r;
>
> - if (ring->funcs->emit_pipeline_sync && (
> - job->vm_needs_flush || gds_switch_needed ||
> - amdgpu_vm_ring_has_compute_vm_bug(ring)))
> - amdgpu_ring_emit_pipeline_sync(ring);
> -
> if (ring->funcs->emit_vm_flush && (job->vm_needs_flush ||
> amdgpu_vm_is_gpu_reset(adev, id))) {
> struct fence *fence;
> diff --git a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c
> index d5719eb2..8421da2 100644
> --- a/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c
> +++ b/drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c
> @@ -6578,7 +6578,6 @@ static void gfx_v8_0_ring_emit_fence_gfx(struct amdgpu_ring *ring, u64 addr,
> DATA_SEL(write64bit ? 2 : 1) | INT_SEL(int_sel ? 2 : 0));
> amdgpu_ring_write(ring, lower_32_bits(seq));
> amdgpu_ring_write(ring, upper_32_bits(seq));
> -
> }
>
> static void gfx_v8_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
> @@ -6595,6 +6594,8 @@ static void gfx_v8_0_ring_emit_pipeline_sync(struct amdgpu_ring *ring)
> amdgpu_ring_write(ring, upper_32_bits(addr) & 0xffffffff);
> amdgpu_ring_write(ring, seq);
> amdgpu_ring_write(ring, 0xffffffff);
> + if (usepfp)
> + ring->last_vm_flush_pos = ring->wptr;
> amdgpu_ring_write(ring, 4); /* poll interval */
> }
>
> @@ -6641,9 +6642,9 @@ static void gfx_v8_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
> if (usepfp) {
> /* sync PFP to ME, otherwise we might get invalid PFP reads */
> amdgpu_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
> + /* mark down last pos of vm_flush for padding NOPs before CE ib */
> + ring->last_vm_flush_pos = ring->wptr;
> amdgpu_ring_write(ring, 0x0);
> - /* GFX8 emits 128 dw nop to prevent CE access VM before vm_flush finish */
> - amdgpu_ring_insert_nop(ring, 128);
> }
> }
>
> @@ -6749,6 +6750,16 @@ static void gfx_v8_ring_emit_cntxcntl(struct amdgpu_ring *ring, uint32_t flags)
> if (amdgpu_sriov_vf(ring->adev))
> gfx_v8_0_ring_emit_de_meta_init(ring,
> (flags & AMDGPU_VM_DOMAIN) ? AMDGPU_CSA_VADDR : ring->adev->virt.csa_vmid0_addr);
> +
> + /* TODO: can we use one simple fomular to cover below logic? */
> + if (ring->wptr > ring->last_vm_flush_pos) {
> + if (ring->wptr - ring->last_vm_flush_pos < 128)
> + amdgpu_ring_insert_nop(ring, 128 - (ring->wptr - ring->last_vm_flush_pos));
> + } else {
> + if (ring->ptr_mask - ring->last_vm_flush_pos + ring->wptr < 128)
> + amdgpu_ring_insert_nop(ring,
> + 128 - (ring->ptr_mask - ring->last_vm_flush_pos + ring->wptr));
> + }
> }
>
> static void gfx_v8_0_ring_emit_rreg(struct amdgpu_ring *ring, uint32_t reg)
> @@ -7035,7 +7046,7 @@ static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_gfx = {
> 5 + /* gfx_v8_0_ring_emit_hdp_invalidate */
> 6 + 6 + 6 +/* gfx_v8_0_ring_emit_fence_gfx x3 for user fence, vm fence */
> 7 + /* gfx_v8_0_ring_emit_pipeline_sync */
> - 128 + 19 + /* gfx_v8_0_ring_emit_vm_flush */
> + 128 - 63 + 19 + /* gfx_v8_0_ring_emit_vm_flush, and padding NOPs prior to IB */
> 2 + /* gfx_v8_ring_emit_sb */
> 3 + 4 + 29, /* gfx_v8_ring_emit_cntxcntl including vgt flush/meta-data */
> .emit_ib_size = 4, /* gfx_v8_0_ring_emit_ib_gfx */
More information about the amd-gfx
mailing list