Skip to contentSkip to conversation

drm/amdgpu:guarantee 128dws between vm flush and IB(v3)

PATCH In review Monk Liu 3 participants 10 messages Jan 18, 2017 – Jan 22, 2017 7 messages not in the archive

Subsystems (2)
lore
Share & tools
lore

Monk Liu

author
#0 lore raw

PATCH 1/1 drm/amdgpu:guarantee 128dws between vm flush and IB(v3)

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.

Change-Id I0a1845de07c0b86ac1b77ac1a007e893144144fd
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h |  1 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c   |  7 +++++++
 drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c    | 16 ++++++++++++----
 3 files changed, 20 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h +1 -0
index c813cbe..332969f 100644
@@ -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);
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +7 -0
index d05546e..53fc5e0 100644
@@ -396,6 +396,13 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job)
 	    amdgpu_vm_ring_has_compute_vm_bug(ring)))
 		amdgpu_ring_emit_pipeline_sync(ring);
 
+	/* when no vm-flush this frame, we still need to mark down
+	 * the position of the tail of hdp_flush, because we still
+	 * need to make sure there are 128 DWs between last SWITCH_BUFFER and
+	 * the emit_ib this frame. otherwise there is still VM fault occured on
+	 * constant engine.
+	 */
+	ring->last_vm_flush_pos = ring->wptr;
 	if (ring->funcs->emit_vm_flush && (job->vm_needs_flush ||
 	    amdgpu_vm_is_gpu_reset(adev, id))) {
 		struct fence *fence;
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +12 -4
index 5f37313..dde4177 100644
@@ -6572,7 +6572,6 @@ 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)
@@ -6636,9 +6635,8 @@ static void gfx_v8_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
 		/* sync PFP to ME, otherwise we might get invalid PFP reads */
 		amdgpu_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
 		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);
 	}
+	ring->last_vm_flush_pos = ring->wptr;
 }
 
 static u32 gfx_v8_0_ring_get_wptr_compute(struct amdgpu_ring *ring)
@@ -6743,6 +6741,15 @@ 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);
+
+	/* We need to pad some NOPs before emit_ib to prevent CE run ahead of
+	 * vm_flush, which may trigger VM fault. */
+	if (ring->wptr > ring->last_vm_flush_pos) /* no wptr wrapping to RB head */
+		amdgpu_ring_insert_nop(ring, 128 - (ring->wptr - ring->last_vm_flush_pos));
+	else
+		if (ring->ptr_mask + 1 - ring->last_vm_flush_pos + ring->wptr < 128)
+			amdgpu_ring_insert_nop(ring,
+				128 - (ring->ptr_mask + 1 - ring->last_vm_flush_pos + ring->wptr));
 }
 
 static void gfx_v8_0_ring_emit_rreg(struct amdgpu_ring *ring, uint32_t reg)
@@ -7018,7 +7025,8 @@ static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_gfx = {
 		7 + /* gfx_v8_0_ring_emit_pipeline_sync */
 		128 + 19 + /* gfx_v8_0_ring_emit_vm_flush */
 		2 + /* gfx_v8_ring_emit_sb */
-		3 + 4 + 29, /* gfx_v8_ring_emit_cntxcntl including vgt flush/meta-data */
+		3 + 4 + 29 - /* gfx_v8_ring_emit_cntxcntl including vgt flush/meta-data */
+		20 - 7 - 6 - 3 - 4 - 29, /* no need to count gds/hdp_flush/vm_flush fence/cntx_cntl/vgt_flush/meta-data anymore */
 	.emit_ib_size =	4, /* gfx_v8_0_ring_emit_ib_gfx */
 	.emit_ib = gfx_v8_0_ring_emit_ib_gfx,
 	.emit_fence = gfx_v8_0_ring_emit_fence_gfx,
Signature · 6 lines
2.7.4

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 1484739739-30080-1-git-send-email-Monk.Liu-5C7GfCeVMHo@public.gmane.org

#2 lore raw

Am 18.01.2017 um 12:42 schrieb Monk Liu:

Quoted 79 lines

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.

Change-Id I0a1845de07c0b86ac1b77ac1a007e893144144fd
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h |  1 +
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c   |  7 +++++++
  drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c    | 16 ++++++++++++----
  3 files changed, 20 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h +1 -0
index c813cbe..332969f 100644
@@ -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);
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +7 -0
index d05546e..53fc5e0 100644
@@ -396,6 +396,13 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job)
  	    amdgpu_vm_ring_has_compute_vm_bug(ring)))
  		amdgpu_ring_emit_pipeline_sync(ring);
 
+	/* when no vm-flush this frame, we still need to mark down
+	 * the position of the tail of hdp_flush, because we still
+	 * need to make sure there are 128 DWs between last SWITCH_BUFFER and
+	 * the emit_ib this frame. otherwise there is still VM fault occured on
+	 * constant engine.
+	 */
+	ring->last_vm_flush_pos = ring->wptr;
  	if (ring->funcs->emit_vm_flush && (job->vm_needs_flush ||
  	    amdgpu_vm_is_gpu_reset(adev, id))) {
  		struct fence *fence;
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +1 -3
index 5f37313..dde4177 100644
@@ -6572,7 +6572,6 @@ 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)
@@ -6636,9 +6635,8 @@ static void gfx_v8_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
  		/* sync PFP to ME, otherwise we might get invalid PFP reads */
  		amdgpu_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
  		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);
  	}
+	ring->last_vm_flush_pos = ring->wptr;
  }
 
  static u32 gfx_v8_0_ring_get_wptr_compute(struct amdgpu_ring *ring)

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c static void gfx_v8_ring_emit_cntxcntl(struct amdgpu_ring *ring, uint32_t flags) · line 6748

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +5 -0
@@ -6743,6 +6741,15 @@ 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);
+
+	/* We need to pad some NOPs before emit_ib to prevent CE run ahead of
+	 * vm_flush, which may trigger VM fault. */
+	if (ring->wptr > ring->last_vm_flush_pos) /* no wptr wrapping to RB head */
+		amdgpu_ring_insert_nop(ring, 128 - (ring->wptr - ring->last_vm_flush_pos));

This can easily result in a negative number, couldn't it?

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c static void gfx_v8_ring_emit_cntxcntl(struct amdgpu_ring *ring, uint32_t flags) · line 6752

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +4 -0
+	else
+		if (ring->ptr_mask + 1 - ring->last_vm_flush_pos + ring->wptr < 128)
+			amdgpu_ring_insert_nop(ring,
+				128 - (ring->ptr_mask + 1 - ring->last_vm_flush_pos + ring->wptr));

I think it would be cleaner if you calculate the number of NOPs needed first for both cases and then check if the number isn't negative for both cases.

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_gfx = { · line 7028

Quoted 9 lines
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +1 -1
  }
 
  static void gfx_v8_0_ring_emit_rreg(struct amdgpu_ring *ring, uint32_t reg)
@@ -7018,7 +7025,8 @@ static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_gfx = {
  		7 + /* gfx_v8_0_ring_emit_pipeline_sync */
  		128 + 19 + /* gfx_v8_0_ring_emit_vm_flush */
  		2 + /* gfx_v8_ring_emit_sb */
-		3 + 4 + 29, /* gfx_v8_ring_emit_cntxcntl including vgt flush/meta-data */
+		3 + 4 + 29 - /* gfx_v8_ring_emit_cntxcntl including vgt flush/meta-data */

Please put the - on the next line. That is easy to miss and I asked myself for a moment why you want to add 20 here.

Apart from that the patch looks good to me,
Christian.

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_gfx = { · line 7032

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +1 -0
+		20 - 7 - 6 - 3 - 4 - 29, /* no need to count gds/hdp_flush/vm_flush fence/cntx_cntl/vgt_flush/meta-data anymore */
  	.emit_ib_size =	4, /* gfx_v8_0_ring_emit_ib_gfx */
  	.emit_ib = gfx_v8_0_ring_emit_ib_gfx,
  	.emit_fence = gfx_v8_0_ring_emit_fence_gfx,

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 7ff8505a-a6c2-8258-5cfa-45a5344ba2df-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#4 lore raw
Quoted 27 lines · Christian König

Am 18.01.2017 um 12:42 schrieb Monk Liu:

Quoted 11 lines
(quoted diff) +2 -0
@@ -6743,6 +6741,15 @@ 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);
+
+       /* We need to pad some NOPs before emit_ib to prevent CE run ahead

of

(quoted diff) +2 -0
+        * vm_flush, which may trigger VM fault. */
+       if (ring->wptr > ring->last_vm_flush_pos) /* no wptr wrapping to
Quoted 3 lines

RB head */
+ amdgpu_ring_insert_nop(ring, 128 - (ring->wptr -
ring->last_vm_flush_pos));

This can easily result in a negative number, couldn't it?

Quoted 6 lines
(quoted diff) +2 -0
+       else
+               if (ring->ptr_mask + 1 - ring->last_vm_flush_pos +

ring->wptr < 128)

(quoted diff) +2 -0
+                       amdgpu_ring_insert_nop(ring,
+                               128 - (ring->ptr_mask + 1 -

ring->last_vm_flush_pos + ring->wptr));

Quoted 2 lines

I think it would be cleaner if you calculate the number of NOPs needed first for both cases and then check if the number isn't negative for both cases.

What about this:
128 - ((ring->wptr - ring->last_vm_flush_pos) & 127)

Gražvydas
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): CANOLnOOwu2Fd0Svx5tcp7kqb2b28Ar30gTd-FeY4xQwm32J5YQ-JsoAwUIsXosN+BqQ9rBEUg@public.gmane.org

#6 lore raw

Am 19.01.2017 um 14:51 schrieb Grazvydas Ignotas:

Quoted 27 lines
Quoted 25 lines · Christian König

Am 18.01.2017 um 12:42 schrieb Monk Liu:

Quoted 11 lines
(quoted diff) +2 -0
@@ -6743,6 +6741,15 @@ 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);
+
+       /* We need to pad some NOPs before emit_ib to prevent CE run ahead

of

(quoted diff) +2 -0
+        * vm_flush, which may trigger VM fault. */
+       if (ring->wptr > ring->last_vm_flush_pos) /* no wptr wrapping to
Quoted 3 lines

RB head */
+ amdgpu_ring_insert_nop(ring, 128 - (ring->wptr -
ring->last_vm_flush_pos));

This can easily result in a negative number, couldn't it?

Quoted 6 lines
(quoted diff) +2 -0
+       else
+               if (ring->ptr_mask + 1 - ring->last_vm_flush_pos +

ring->wptr < 128)

(quoted diff) +2 -0
+                       amdgpu_ring_insert_nop(ring,
+                               128 - (ring->ptr_mask + 1 -

ring->last_vm_flush_pos + ring->wptr));

Quoted 4 lines
Quoted 2 lines

I think it would be cleaner if you calculate the number of NOPs needed first for both cases and then check if the number isn't negative for both cases.

What about this:
128 - ((ring->wptr - ring->last_vm_flush_pos) & 127)

That won't handle the case for negative nop count correctly either.

See when we already emitted more than 128 dw we don't want to add some more.

Christian.

Quoted 2 lines

Gražvydas

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): fd28a4f9-80ce-49d4-d6e7-9765988e0181-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

#8 lore raw
Quoted 39 lines · Christian König

Am 19.01.2017 um 14:51 schrieb Grazvydas Ignotas:

Quoted 38 lines
Quoted 34 lines · Christian König

Am 18.01.2017 um 12:42 schrieb Monk Liu:

Quoted 13 lines
(quoted diff) +2 -0
@@ -6743,6 +6741,15 @@ 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);
+
+       /* We need to pad some NOPs before emit_ib to prevent CE run

ahead
of

(quoted diff) +2 -0
+        * vm_flush, which may trigger VM fault. */
+       if (ring->wptr > ring->last_vm_flush_pos) /* no wptr wrapping to
Quoted 3 lines

RB head */
+ amdgpu_ring_insert_nop(ring, 128 - (ring->wptr -
ring->last_vm_flush_pos));

This can easily result in a negative number, couldn't it?

Quoted 6 lines
(quoted diff) +2 -0
+       else
+               if (ring->ptr_mask + 1 - ring->last_vm_flush_pos +

ring->wptr < 128)

(quoted diff) +2 -0
+                       amdgpu_ring_insert_nop(ring,
+                               128 - (ring->ptr_mask + 1 -

ring->last_vm_flush_pos + ring->wptr));

I think it would be cleaner if you calculate the number of NOPs needed first
for both cases and then check if the number isn't negative for both cases.

Quoted 7 lines
Quoted 2 lines

What about this:
128 - ((ring->wptr - ring->last_vm_flush_pos) & 127)

That won't handle the case for negative nop count correctly either.

See when we already emitted more than 128 dw we don't want to add some more.

Let me try again then:

count_added = (ring->wptr - ring->last_vm_flush_pos) & ring->ptr_mask; if (count_added < 128)

    amdgpu_ring_insert_nop(ring, 128 - count_added);

Gražvydas
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

author
#9 lore raw
In reply to a message not in the archive

Christian

I just found one issue in your previous patch ( the one you remove inserting 128nop before vm_fush and meanwhile set align_mask of GFX ring to 0xff)

Because I checked amdgpu_ring_commit() again, this function will not guarantee current submit size aligned with 256 dw, instead it only guarantee the ring->wptr will be 256 dw aligned after amdgpu_ring_commit()

So that means your approach can not make sure there will be 128 NOP before vm_flush at all !

e.g. this frame start at 200 position of RB, and it submitted 50 dw so it ends at 249 position, you call amdgpu_ring_commit() now and it only padding 6 more NOPs to ring buffer, so wptr finnaly end at 255, that clearly not what we want originally ~ we want to insert 128 nop after SWITCH_BUFFER (or say, before vm_flush), but you only add 6 NOPs ....

your approach only works with wptr == 0 for the first UMD submit

BR Monk

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Christian K?nig
Sent: Thursday, January 19, 2017 5:11 PM
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH] drm/amdgpu:guarantee 128dws between vm flush and IB(v3)

Am 18.01.2017 um 12:42 schrieb Monk Liu:

Quoted 81 lines

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.

Change-Id I0a1845de07c0b86ac1b77ac1a007e893144144fd
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h |  1 +
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c   |  7 +++++++
  drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c    | 16 ++++++++++++----
  3 files changed, 20 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h +1 -1
index c813cbe..332969f 100644
@@ -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..53fc5e0 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +7 -0
@@ -396,6 +396,13 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job)
  	    amdgpu_vm_ring_has_compute_vm_bug(ring)))
  		amdgpu_ring_emit_pipeline_sync(ring);
 
+	/* when no vm-flush this frame, we still need to mark down
+	 * the position of the tail of hdp_flush, because we still
+	 * need to make sure there are 128 DWs between last SWITCH_BUFFER and
+	 * the emit_ib this frame. otherwise there is still VM fault occured on
+	 * constant engine.
+	 */
+	ring->last_vm_flush_pos = ring->wptr;
  	if (ring->funcs->emit_vm_flush && (job->vm_needs_flush ||
  	    amdgpu_vm_is_gpu_reset(adev, id))) {
  		struct fence *fence;
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +0 -1
index 5f37313..dde4177 100644
@@ -6572,7 +6572,6 @@ 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) @@ -6636,9 +6635,8 @@ static void gfx_v8_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
/* sync PFP to ME, otherwise we might get invalid PFP reads */ amdgpu_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0)); amdgpu_ring_write(ring, 0x0);

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +1 -2
-		/* GFX8 emits 128 dw nop to prevent CE access VM before vm_flush finish */
-		amdgpu_ring_insert_nop(ring, 128);
  	}
+	ring->last_vm_flush_pos = ring->wptr;
  }
 
  static u32 gfx_v8_0_ring_get_wptr_compute(struct amdgpu_ring *ring) 

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c static void gfx_v8_ring_emit_cntxcntl(struct amdgpu_ring *ring, uint32_t flags) · line 6750

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +6 -0
@@ -6743,6 +6741,15 @@ 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);
+
+	/* We need to pad some NOPs before emit_ib to prevent CE run ahead of
+	 * vm_flush, which may trigger VM fault. */
+	if (ring->wptr > ring->last_vm_flush_pos) /* no wptr wrapping to RB head */
+		amdgpu_ring_insert_nop(ring, 128 - (ring->wptr - 
+ring->last_vm_flush_pos));

This can easily result in a negative number, couldn't it?

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c static void gfx_v8_ring_emit_cntxcntl(struct amdgpu_ring *ring, uint32_t flags)

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +5 -0
+	else
+		if (ring->ptr_mask + 1 - ring->last_vm_flush_pos + ring->wptr < 128)
+			amdgpu_ring_insert_nop(ring,
+				128 - (ring->ptr_mask + 1 - ring->last_vm_flush_pos + 
+ring->wptr));

I think it would be cleaner if you calculate the number of NOPs needed first for both cases and then check if the number isn't negative for both cases.

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c

Quoted 10 lines
  }

static void gfx_v8_0_ring_emit_rreg(struct amdgpu_ring *ring,
uint32_t reg) @@ -7018,7 +7025,8 @@ static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_gfx = {
7 + /* gfx_v8_0_ring_emit_pipeline_sync */
128 + 19 + /* gfx_v8_0_ring_emit_vm_flush */
2 + /* gfx_v8_ring_emit_sb */

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +2 -1
-		3 + 4 + 29, /* gfx_v8_ring_emit_cntxcntl including vgt flush/meta-data */
+		3 + 4 + 29 - /* gfx_v8_ring_emit_cntxcntl including vgt 
+flush/meta-data */

Please put the - on the next line. That is easy to miss and I asked myself for a moment why you want to add 20 here.

Apart from that the patch looks good to me, Christian.

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_gfx = { · line 7032

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +2 -0
+		20 - 7 - 6 - 3 - 4 - 29, /* no need to count gds/hdp_flush/vm_flush 
+fence/cntx_cntl/vgt_flush/meta-data anymore */
  	.emit_ib_size =	4, /* gfx_v8_0_ring_emit_ib_gfx */
  	.emit_ib = gfx_v8_0_ring_emit_ib_gfx,
  	.emit_fence = gfx_v8_0_ring_emit_fence_gfx,

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

author
#10 lore raw
In reply to Christian König

And with more think on it, even the very first UMD cmd start with wptr at 256/512 position ( because ib test ends up at 255 aligned position), your approach still cannot satisfy 128 DW between SWTICH_BUFFER and Preamble IB (if no vm_flush) or 128 Nops before VM_flush (if needed) for the next frame.

e.g. we insert 180 dw this frame, start from0, end up at 179, and ring_commit() padding NOPs and make wptr end up at 255, so only 75 NOPs is between SWITCH_BUFFER and next frame (assume we don't have vm-flush next frame)

BR Monk

-----Original Message-----
From: Liu, Monk
Sent: Thursday, January 19, 2017 10:39 PM
To: 'Christian König' <[email protected]>; [email protected]
Subject: RE: [PATCH] drm/amdgpu:guarantee 128dws between vm flush and IB(v3)

Christian

I just found one issue in your previous patch ( the one you remove inserting 128nop before vm_fush and meanwhile set align_mask of GFX ring to 0xff)

Because I checked amdgpu_ring_commit() again, this function will not guarantee current submit size aligned with 256 dw, instead it only guarantee the ring->wptr will be 256 dw aligned after amdgpu_ring_commit()

So that means your approach can not make sure there will be 128 NOP before vm_flush at all !

e.g. this frame start at 200 position of RB, and it submitted 50 dw so it ends at 249 position, you call amdgpu_ring_commit() now and it only padding 6 more NOPs to ring buffer, so wptr finnaly end at 255, that clearly not what we want originally ~ we want to insert 128 nop after SWITCH_BUFFER (or say, before vm_flush), but you only add 6 NOPs ....

your approach only works with wptr == 0 for the first UMD submit

BR Monk

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Christian K?nig
Sent: Thursday, January 19, 2017 5:11 PM
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH] drm/amdgpu:guarantee 128dws between vm flush and IB(v3)

Am 18.01.2017 um 12:42 schrieb Monk Liu:

Quoted 81 lines

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.

Change-Id I0a1845de07c0b86ac1b77ac1a007e893144144fd
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h |  1 +
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c   |  7 +++++++
  drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c    | 16 ++++++++++++----
  3 files changed, 20 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h +1 -1
index c813cbe..332969f 100644
@@ -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..53fc5e0 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +7 -0
@@ -396,6 +396,13 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job)
  	    amdgpu_vm_ring_has_compute_vm_bug(ring)))
  		amdgpu_ring_emit_pipeline_sync(ring);
 
+	/* when no vm-flush this frame, we still need to mark down
+	 * the position of the tail of hdp_flush, because we still
+	 * need to make sure there are 128 DWs between last SWITCH_BUFFER and
+	 * the emit_ib this frame. otherwise there is still VM fault occured on
+	 * constant engine.
+	 */
+	ring->last_vm_flush_pos = ring->wptr;
  	if (ring->funcs->emit_vm_flush && (job->vm_needs_flush ||
  	    amdgpu_vm_is_gpu_reset(adev, id))) {
  		struct fence *fence;
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +0 -1
index 5f37313..dde4177 100644
@@ -6572,7 +6572,6 @@ 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) @@ -6636,9 +6635,8 @@ static void gfx_v8_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
/* sync PFP to ME, otherwise we might get invalid PFP reads */ amdgpu_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0)); amdgpu_ring_write(ring, 0x0);

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +1 -2
-		/* GFX8 emits 128 dw nop to prevent CE access VM before vm_flush finish */
-		amdgpu_ring_insert_nop(ring, 128);
  	}
+	ring->last_vm_flush_pos = ring->wptr;
  }
 
  static u32 gfx_v8_0_ring_get_wptr_compute(struct amdgpu_ring *ring) 

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c static void gfx_v8_ring_emit_cntxcntl(struct amdgpu_ring *ring, uint32_t flags) · line 6750

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +6 -0
@@ -6743,6 +6741,15 @@ 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);
+
+	/* We need to pad some NOPs before emit_ib to prevent CE run ahead of
+	 * vm_flush, which may trigger VM fault. */
+	if (ring->wptr > ring->last_vm_flush_pos) /* no wptr wrapping to RB head */
+		amdgpu_ring_insert_nop(ring, 128 - (ring->wptr -
+ring->last_vm_flush_pos));

This can easily result in a negative number, couldn't it?

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c static void gfx_v8_ring_emit_cntxcntl(struct amdgpu_ring *ring, uint32_t flags)

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +5 -0
+	else
+		if (ring->ptr_mask + 1 - ring->last_vm_flush_pos + ring->wptr < 128)
+			amdgpu_ring_insert_nop(ring,
+				128 - (ring->ptr_mask + 1 - ring->last_vm_flush_pos + 
+ring->wptr));

I think it would be cleaner if you calculate the number of NOPs needed first for both cases and then check if the number isn't negative for both cases.

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c

Quoted 10 lines
  }

static void gfx_v8_0_ring_emit_rreg(struct amdgpu_ring *ring,
uint32_t reg) @@ -7018,7 +7025,8 @@ static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_gfx = {
7 + /* gfx_v8_0_ring_emit_pipeline_sync */
128 + 19 + /* gfx_v8_0_ring_emit_vm_flush */
2 + /* gfx_v8_ring_emit_sb */

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +2 -1
-		3 + 4 + 29, /* gfx_v8_ring_emit_cntxcntl including vgt flush/meta-data */
+		3 + 4 + 29 - /* gfx_v8_ring_emit_cntxcntl including vgt 
+flush/meta-data */

Please put the - on the next line. That is easy to miss and I asked myself for a moment why you want to add 20 here.

Apart from that the patch looks good to me, Christian.

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_gfx = { · line 7032

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +2 -0
+		20 - 7 - 6 - 3 - 4 - 29, /* no need to count gds/hdp_flush/vm_flush 
+fence/cntx_cntl/vgt_flush/meta-data anymore */
  	.emit_ib_size =	4, /* gfx_v8_0_ring_emit_ib_gfx */
  	.emit_ib = gfx_v8_0_ring_emit_ib_gfx,
  	.emit_fence = gfx_v8_0_ring_emit_fence_gfx,

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): BY2PR1201MB1110457767084D1B300C67D8847E0-O28G1zQ8oGliQkyLPkmea2rFom/aUZj6nBOFsp37pqbUKgpGm//BTAC/[email protected]

#12 lore raw

Yeah, thought about that possibility as well when we change the padding to 256.

This happens every time our command submission uses more than 128 dw, in this case you obviously have less than 128 padding.

But why doesn't this happen on Windows then?

Regards,
Christian.

Am 19.01.2017 um 15:49 schrieb Liu, Monk:

Quoted 147 lines

And with more think on it, even the very first UMD cmd start with wptr at 256/512 position ( because ib test ends up at 255 aligned position), your approach still cannot satisfy 128 DW between SWTICH_BUFFER and Preamble IB (if no vm_flush) or 128 Nops before VM_flush (if needed) for the next frame.

e.g. we insert 180 dw this frame, start from0, end up at 179, and ring_commit() padding NOPs and make wptr end up at 255, so only 75 NOPs is between SWITCH_BUFFER and next frame (assume we don't have vm-flush next frame)

BR Monk

-----Original Message-----
From: Liu, Monk
Sent: Thursday, January 19, 2017 10:39 PM
To: 'Christian König' <[email protected]>; [email protected]
Subject: RE: [PATCH] drm/amdgpu:guarantee 128dws between vm flush and IB(v3)

Christian

I just found one issue in your previous patch ( the one you remove inserting 128nop before vm_fush and meanwhile set align_mask of GFX ring to 0xff)

Because I checked amdgpu_ring_commit() again, this function will not guarantee current submit size aligned with 256 dw, instead it only guarantee the ring->wptr will be 256 dw aligned after amdgpu_ring_commit()

So that means your approach can not make sure there will be 128 NOP before vm_flush at all !

e.g. this frame start at 200 position of RB, and it submitted 50 dw so it ends at 249 position, you call amdgpu_ring_commit() now and it only padding 6 more NOPs to ring buffer, so wptr finnaly end at 255, that clearly not what we want originally ~ we want to insert 128 nop after SWITCH_BUFFER (or say, before vm_flush), but you only add 6 NOPs ....

your approach only works with wptr == 0 for the first UMD submit

BR Monk

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Christian K?nig
Sent: Thursday, January 19, 2017 5:11 PM
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH] drm/amdgpu:guarantee 128dws between vm flush and IB(v3)

Am 18.01.2017 um 12:42 schrieb Monk Liu:

Quoted 81 lines

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.

Change-Id I0a1845de07c0b86ac1b77ac1a007e893144144fd
---
   drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h |  1 +
   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c   |  7 +++++++
   drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c    | 16 ++++++++++++----
   3 files changed, 20 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h +1 -1
index c813cbe..332969f 100644
@@ -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..53fc5e0 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +7 -0
@@ -396,6 +396,13 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job)
   	    amdgpu_vm_ring_has_compute_vm_bug(ring)))
   		amdgpu_ring_emit_pipeline_sync(ring);
 
+	/* when no vm-flush this frame, we still need to mark down
+	 * the position of the tail of hdp_flush, because we still
+	 * need to make sure there are 128 DWs between last SWITCH_BUFFER and
+	 * the emit_ib this frame. otherwise there is still VM fault occured on
+	 * constant engine.
+	 */
+	ring->last_vm_flush_pos = ring->wptr;
   	if (ring->funcs->emit_vm_flush && (job->vm_needs_flush ||
   	    amdgpu_vm_is_gpu_reset(adev, id))) {
   		struct fence *fence;
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +0 -1
index 5f37313..dde4177 100644
@@ -6572,7 +6572,6 @@ 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) @@ -6636,9 +6635,8 @@ static void gfx_v8_0_ring_emit_vm_flush(struct amdgpu_ring *ring,
/* sync PFP to ME, otherwise we might get invalid PFP reads */

   		amdgpu_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
   		amdgpu_ring_write(ring, 0x0);
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +1 -2
-		/* GFX8 emits 128 dw nop to prevent CE access VM before vm_flush finish */
-		amdgpu_ring_insert_nop(ring, 128);
   	}
+	ring->last_vm_flush_pos = ring->wptr;
   }
 
   static u32 gfx_v8_0_ring_get_wptr_compute(struct amdgpu_ring *ring)

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c static void gfx_v8_ring_emit_cntxcntl(struct amdgpu_ring *ring, uint32_t flags) · line 6750

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +6 -0
@@ -6743,6 +6741,15 @@ 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);
+
+	/* We need to pad some NOPs before emit_ib to prevent CE run ahead of
+	 * vm_flush, which may trigger VM fault. */
+	if (ring->wptr > ring->last_vm_flush_pos) /* no wptr wrapping to RB head */
+		amdgpu_ring_insert_nop(ring, 128 - (ring->wptr -
+ring->last_vm_flush_pos));

This can easily result in a negative number, couldn't it?

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c static void gfx_v8_ring_emit_cntxcntl(struct amdgpu_ring *ring, uint32_t flags)

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +5 -0
+	else
+		if (ring->ptr_mask + 1 - ring->last_vm_flush_pos + ring->wptr < 128)
+			amdgpu_ring_insert_nop(ring,
+				128 - (ring->ptr_mask + 1 - ring->last_vm_flush_pos +
+ring->wptr));

I think it would be cleaner if you calculate the number of NOPs needed first for both cases and then check if the number isn't negative for both cases.

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c

Quoted 10 lines
   }

static void gfx_v8_0_ring_emit_rreg(struct amdgpu_ring *ring,

uint32_t reg) @@ -7018,7 +7025,8 @@ static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_gfx = {
   		7 + /* gfx_v8_0_ring_emit_pipeline_sync */
   		128 + 19 + /* gfx_v8_0_ring_emit_vm_flush */
   		2 + /* gfx_v8_ring_emit_sb */
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +2 -1
-		3 + 4 + 29, /* gfx_v8_ring_emit_cntxcntl including vgt flush/meta-data */
+		3 + 4 + 29 - /* gfx_v8_ring_emit_cntxcntl including vgt
+flush/meta-data */

Please put the - on the next line. That is easy to miss and I asked myself for a moment why you want to add 20 here.

Apart from that the patch looks good to me, Christian.

Quoted 10 lines

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_gfx = { · line 7032

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +2 -0
+		20 - 7 - 6 - 3 - 4 - 29, /* no need to count gds/hdp_flush/vm_flush
+fence/cntx_cntl/vgt_flush/meta-data anymore */
   	.emit_ib_size =	4, /* gfx_v8_0_ring_emit_ib_gfx */
   	.emit_ib = gfx_v8_0_ring_emit_ib_gfx,
   	.emit_fence = gfx_v8_0_ring_emit_fence_gfx,

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): 72ce673f-a232-290d-62c0-082e6ede3bbd-ANTagKRnAhcb1SvskN2V4Q@public.gmane.org

author
#14 lore raw

1, windows separate vm flush and gfx submit, so if with vm flush windows actually submit 256 dw totally (128 for vm flush and 128 for gfx submit)

2, for gfx submit frame: windows always emit pipeline sync(name is "w0ait_3d_idle" for windows)

3, for vm flush frame: windows always emit pipeline sync(name is drain_pfp_to_EOP for windows) prior to pd update and following vm invalidate.

Quoted 1 line

But why doesn't this happen on Windows then?

[ML] good question, I guess the answer is windows always insert pipeline sync.

for the case without vm flush: a pipeline sync in gfx submit can prevent ME ruin SH/3d register before previous pipeline finished (DE ib will programing lots of 3d/sh register; and CE ib will also change many resource descriptor if there is no context switch this frame)

for the case with a vm flush: a pipeline sync before vm invalidate can also prevent ME invalidate VM before previous pipeline finished.

I think a vm invalidate will trigger VM fault if 3d pipeline is working on the invalidated VMID *or* invalidated PD address.

so I think we should move pipeline sync out of VM flush next step, and always put it prior to IB as well as vm invalidate

BR Monk

________________________________
发件人: Christian König <[email protected]>
发送时间: 2017年1月20日 16:48:17
收件人: Liu, Monk; [email protected]
主题: Re: [PATCH] drm/amdgpu:guarantee 128dws between vm flush and IB(v3)

Yeah, thought about that possibility as well when we change the padding to 256.

This happens every time our command submission uses more than 128 dw, in this case you obviously have less than 128 padding.

But why doesn't this happen on Windows then?

Regards,
Christian.

Am 19.01.2017 um 15:49 schrieb Liu, Monk:

Quoted 153 lines

And with more think on it, even the very first UMD cmd start with wptr at 256/512 position ( because ib test ends up at 255 aligned position), your approach still cannot satisfy 128 DW between SWTICH_BUFFER and Preamble IB (if no vm_flush) or 128 Nops before VM_flush (if needed) for the next frame.

e.g. we insert 180 dw this frame, start from0, end up at 179, and ring_commit() padding NOPs and make wptr end up at 255, so only 75 NOPs is between SWITCH_BUFFER and next frame (assume we don't have vm-flush next frame)

BR Monk

-----Original Message-----
From: Liu, Monk
Sent: Thursday, January 19, 2017 10:39 PM
To: 'Christian König' <[email protected]>; [email protected]
Subject: RE: [PATCH] drm/amdgpu:guarantee 128dws between vm flush and IB(v3)

Christian

I just found one issue in your previous patch ( the one you remove inserting 128nop before vm_fush and meanwhile set align_mask of GFX ring to 0xff)

Because I checked amdgpu_ring_commit() again, this function will not guarantee current submit size aligned with 256 dw, instead it only guarantee the ring->wptr will be 256 dw aligned after amdgpu_ring_commit()

So that means your approach can not make sure there will be 128 NOP before vm_flush at all !

e.g. this frame start at 200 position of RB, and it submitted 50 dw so it ends at 249 position, you call amdgpu_ring_commit() now and it only padding 6 more NOPs to ring buffer, so wptr finnaly end at 255, that clearly not what we want originally ~ we want to insert 128 nop after SWITCH_BUFFER (or say, before vm_flush), but you only add 6 NOPs ....

your approach only works with wptr == 0 for the first UMD submit

BR Monk

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Christian K?nig
Sent: Thursday, January 19, 2017 5:11 PM
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH] drm/amdgpu:guarantee 128dws between vm flush and IB(v3)

Am 18.01.2017 um 12:42 schrieb Monk Liu:

Quoted 81 lines

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.

Change-Id I0a1845de07c0b86ac1b77ac1a007e893144144fd
---
   drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h |  1 +
   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c   |  7 +++++++
   drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c    | 16 ++++++++++++----
   3 files changed, 20 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h +1 -1
index c813cbe..332969f 100644
@@ -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..53fc5e0 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +7 -0
@@ -396,6 +396,13 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job)
          amdgpu_vm_ring_has_compute_vm_bug(ring)))
              amdgpu_ring_emit_pipeline_sync(ring);
 
+    /* when no vm-flush this frame, we still need to mark down
+     * the position of the tail of hdp_flush, because we still
+     * need to make sure there are 128 DWs between last SWITCH_BUFFER and
+     * the emit_ib this frame. otherwise there is still VM fault occured on
+     * constant engine.
+     */
+    ring->last_vm_flush_pos = ring->wptr;
      if (ring->funcs->emit_vm_flush && (job->vm_needs_flush ||
          amdgpu_vm_is_gpu_reset(adev, id))) {
              struct fence *fence;
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +0 -1
index 5f37313..dde4177 100644
@@ -6572,7 +6572,6 @@ 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) @@ -6636,9 +6635,8 @@ static void gfx_v8_0_ring_emit_vm_flush(struct amdgpu_ring *ring,

              /* sync PFP to ME, otherwise we might get invalid PFP reads */
              amdgpu_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
              amdgpu_ring_write(ring, 0x0);
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +1 -2
-            /* GFX8 emits 128 dw nop to prevent CE access VM before vm_flush finish */
-            amdgpu_ring_insert_nop(ring, 128);
      }
+    ring->last_vm_flush_pos = ring->wptr;
   }
 
   static u32 gfx_v8_0_ring_get_wptr_compute(struct amdgpu_ring *ring)

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c static void gfx_v8_ring_emit_cntxcntl(struct amdgpu_ring *ring, uint32_t flags) · line 6750

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +6 -0
@@ -6743,6 +6741,15 @@ 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);
+
+    /* We need to pad some NOPs before emit_ib to prevent CE run ahead of
+     * vm_flush, which may trigger VM fault. */
+    if (ring->wptr > ring->last_vm_flush_pos) /* no wptr wrapping to RB head */
+            amdgpu_ring_insert_nop(ring, 128 - (ring->wptr -
+ring->last_vm_flush_pos));

This can easily result in a negative number, couldn't it?

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c static void gfx_v8_ring_emit_cntxcntl(struct amdgpu_ring *ring, uint32_t flags) · line 6755

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +5 -0
+    else
+            if (ring->ptr_mask + 1 - ring->last_vm_flush_pos + ring->wptr < 128)
+                    amdgpu_ring_insert_nop(ring,
+                            128 - (ring->ptr_mask + 1 - ring->last_vm_flush_pos +
+ring->wptr));

I think it would be cleaner if you calculate the number of NOPs needed first for both cases and then check if the number isn't negative for both cases.

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c

Quoted 10 lines
   }

static void gfx_v8_0_ring_emit_rreg(struct amdgpu_ring *ring,

uint32_t reg) @@ -7018,7 +7025,8 @@ static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_gfx = {
              7 + /* gfx_v8_0_ring_emit_pipeline_sync */
              128 + 19 + /* gfx_v8_0_ring_emit_vm_flush */
              2 + /* gfx_v8_ring_emit_sb */
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +2 -1
-            3 + 4 + 29, /* gfx_v8_ring_emit_cntxcntl including vgt flush/meta-data */
+            3 + 4 + 29 - /* gfx_v8_ring_emit_cntxcntl including vgt
+flush/meta-data */

Please put the - on the next line. That is easy to miss and I asked myself for a moment why you want to add 20 here.

Apart from that the patch looks good to me, Christian.

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +2 -0
+            20 - 7 - 6 - 3 - 4 - 29, /* no need to count gds/hdp_flush/vm_flush
+fence/cntx_cntl/vgt_flush/meta-data anymore */
      .emit_ib_size = 4, /* gfx_v8_0_ring_emit_ib_gfx */
      .emit_ib = gfx_v8_0_ring_emit_ib_gfx,
      .emit_fence = gfx_v8_0_ring_emit_fence_gfx,
Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

A message that is not in the archive (sent off-list or not yet received): BY2PR1201MB11108334E6D9C3AAE1DF30D384730-O28G1zQ8oGliQkyLPkmea2rFom/aUZj6nBOFsp37pqbUKgpGm//BTAC/[email protected]

author
#16 lore raw

fix typo:

[ML] good question, I guess the answer is windows always insert pipeline sync.

for the case without vm flush: a pipeline sync in gfx submit can prevent ME ruin SH/3d register before previous pipeline finished (DE ib will programing lots of 3d/sh register and CE ib will also change many resource descriptor if context switch)

for the case with a vm flush: a pipeline sync before vm invalidate can also prevent ME invalidate VM before previous pipeline finished. (windows use drain_pfp_EOP)

________________________________
发件人: amd-gfx <[email protected]> 代表 Liu, Monk <[email protected]> 发送时间: 2017年1月22日 10:47:03
收件人: Christian König; [email protected]
主题: 答复: [PATCH] drm/amdgpu:guarantee 128dws between vm flush and IB(v3)

1, windows separate vm flush and gfx submit, so if with vm flush windows actually submit 256 dw totally (128 for vm flush and 128 for gfx submit)

2, for gfx submit frame: windows always emit pipeline sync(name is "w0ait_3d_idle" for windows)

3, for vm flush frame: windows always emit pipeline sync(name is drain_pfp_to_EOP for windows) prior to pd update and following vm invalidate.

Quoted 1 line

But why doesn't this happen on Windows then?

[ML] good question, I guess the answer is windows always insert pipeline sync.

for the case without vm flush: a pipeline sync in gfx submit can prevent ME ruin SH/3d register before previous pipeline finished (DE ib will programing lots of 3d/sh register; and CE ib will also change many resource descriptor if there is no context switch this frame)

for the case with a vm flush: a pipeline sync before vm invalidate can also prevent ME invalidate VM before previous pipeline finished.

I think a vm invalidate will trigger VM fault if 3d pipeline is working on the invalidated VMID *or* invalidated PD address.

so I think we should move pipeline sync out of VM flush next step, and always put it prior to IB as well as vm invalidate

BR Monk

________________________________
发件人: Christian König <[email protected]>
发送时间: 2017年1月20日 16:48:17
收件人: Liu, Monk; [email protected]
主题: Re: [PATCH] drm/amdgpu:guarantee 128dws between vm flush and IB(v3)

Yeah, thought about that possibility as well when we change the padding to 256.

This happens every time our command submission uses more than 128 dw, in this case you obviously have less than 128 padding.

But why doesn't this happen on Windows then?

Regards,
Christian.

Am 19.01.2017 um 15:49 schrieb Liu, Monk:

Quoted 153 lines

And with more think on it, even the very first UMD cmd start with wptr at 256/512 position ( because ib test ends up at 255 aligned position), your approach still cannot satisfy 128 DW between SWTICH_BUFFER and Preamble IB (if no vm_flush) or 128 Nops before VM_flush (if needed) for the next frame.

e.g. we insert 180 dw this frame, start from0, end up at 179, and ring_commit() padding NOPs and make wptr end up at 255, so only 75 NOPs is between SWITCH_BUFFER and next frame (assume we don't have vm-flush next frame)

BR Monk

-----Original Message-----
From: Liu, Monk
Sent: Thursday, January 19, 2017 10:39 PM
To: 'Christian König' <[email protected]>; [email protected]
Subject: RE: [PATCH] drm/amdgpu:guarantee 128dws between vm flush and IB(v3)

Christian

I just found one issue in your previous patch ( the one you remove inserting 128nop before vm_fush and meanwhile set align_mask of GFX ring to 0xff)

Because I checked amdgpu_ring_commit() again, this function will not guarantee current submit size aligned with 256 dw, instead it only guarantee the ring->wptr will be 256 dw aligned after amdgpu_ring_commit()

So that means your approach can not make sure there will be 128 NOP before vm_flush at all !

e.g. this frame start at 200 position of RB, and it submitted 50 dw so it ends at 249 position, you call amdgpu_ring_commit() now and it only padding 6 more NOPs to ring buffer, so wptr finnaly end at 255, that clearly not what we want originally ~ we want to insert 128 nop after SWITCH_BUFFER (or say, before vm_flush), but you only add 6 NOPs ....

your approach only works with wptr == 0 for the first UMD submit

BR Monk

-----Original Message-----
From: amd-gfx [mailto:[email protected]] On Behalf Of Christian K?nig
Sent: Thursday, January 19, 2017 5:11 PM
To: Liu, Monk <[email protected]>; [email protected]
Subject: Re: [PATCH] drm/amdgpu:guarantee 128dws between vm flush and IB(v3)

Am 18.01.2017 um 12:42 schrieb Monk Liu:

Quoted 81 lines

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.

Change-Id I0a1845de07c0b86ac1b77ac1a007e893144144fd
---
   drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h |  1 +
   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c   |  7 +++++++
   drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c    | 16 ++++++++++++----
   3 files changed, 20 insertions(+), 4 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_ring.h +1 -1
index c813cbe..332969f 100644
@@ -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..53fc5e0 100644

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +7 -0
@@ -396,6 +396,13 @@ int amdgpu_vm_flush(struct amdgpu_ring *ring, struct amdgpu_job *job)
          amdgpu_vm_ring_has_compute_vm_bug(ring)))
              amdgpu_ring_emit_pipeline_sync(ring);
 
+    /* when no vm-flush this frame, we still need to mark down
+     * the position of the tail of hdp_flush, because we still
+     * need to make sure there are 128 DWs between last SWITCH_BUFFER and
+     * the emit_ib this frame. otherwise there is still VM fault occured on
+     * constant engine.
+     */
+    ring->last_vm_flush_pos = ring->wptr;
      if (ring->funcs->emit_vm_flush && (job->vm_needs_flush ||
          amdgpu_vm_is_gpu_reset(adev, id))) {
              struct fence *fence;
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +0 -1
index 5f37313..dde4177 100644
@@ -6572,7 +6572,6 @@ 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) @@ -6636,9 +6635,8 @@ static void gfx_v8_0_ring_emit_vm_flush(struct amdgpu_ring *ring,

              /* sync PFP to ME, otherwise we might get invalid PFP reads */
              amdgpu_ring_write(ring, PACKET3(PACKET3_PFP_SYNC_ME, 0));
              amdgpu_ring_write(ring, 0x0);
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +1 -2
-            /* GFX8 emits 128 dw nop to prevent CE access VM before vm_flush finish */
-            amdgpu_ring_insert_nop(ring, 128);
      }
+    ring->last_vm_flush_pos = ring->wptr;
   }
 
   static u32 gfx_v8_0_ring_get_wptr_compute(struct amdgpu_ring *ring)

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c static void gfx_v8_ring_emit_cntxcntl(struct amdgpu_ring *ring, uint32_t flags) · line 6750

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +6 -0
@@ -6743,6 +6741,15 @@ 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);
+
+    /* We need to pad some NOPs before emit_ib to prevent CE run ahead of
+     * vm_flush, which may trigger VM fault. */
+    if (ring->wptr > ring->last_vm_flush_pos) /* no wptr wrapping to RB head */
+            amdgpu_ring_insert_nop(ring, 128 - (ring->wptr -
+ring->last_vm_flush_pos));

This can easily result in a negative number, couldn't it?

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c static void gfx_v8_ring_emit_cntxcntl(struct amdgpu_ring *ring, uint32_t flags) · line 6755

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +5 -0
+    else
+            if (ring->ptr_mask + 1 - ring->last_vm_flush_pos + ring->wptr < 128)
+                    amdgpu_ring_insert_nop(ring,
+                            128 - (ring->ptr_mask + 1 - ring->last_vm_flush_pos +
+ring->wptr));

I think it would be cleaner if you calculate the number of NOPs needed first for both cases and then check if the number isn't negative for both cases.

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c

Quoted 10 lines
   }

static void gfx_v8_0_ring_emit_rreg(struct amdgpu_ring *ring,

uint32_t reg) @@ -7018,7 +7025,8 @@ static const struct amdgpu_ring_funcs gfx_v8_0_ring_funcs_gfx = {
              7 + /* gfx_v8_0_ring_emit_pipeline_sync */
              128 + 19 + /* gfx_v8_0_ring_emit_vm_flush */
              2 + /* gfx_v8_ring_emit_sb */
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +2 -1
-            3 + 4 + 29, /* gfx_v8_ring_emit_cntxcntl including vgt flush/meta-data */
+            3 + 4 + 29 - /* gfx_v8_ring_emit_cntxcntl including vgt
+flush/meta-data */

Please put the - on the next line. That is easy to miss and I asked myself for a moment why you want to add 20 here.

Apart from that the patch looks good to me, Christian.

drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c

Quoted 5 lines
drivers/gpu/drm/amd/amdgpu/gfx_v8_0.c +2 -0
+            20 - 7 - 6 - 3 - 4 - 29, /* no need to count gds/hdp_flush/vm_flush
+fence/cntx_cntl/vgt_flush/meta-data anymore */
      .emit_ib_size = 4, /* gfx_v8_0_ring_emit_ib_gfx */
      .emit_ib = gfx_v8_0_ring_emit_ib_gfx,
      .emit_fence = gfx_v8_0_ring_emit_fence_gfx,
Quoted 4 lines

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx