Skip to contentSkip to conversation

drm/amdgpu: Flush TLB after mapping for VG20+XGMI

PATCH In review Philip Yang 7 participants 9 messages Apr 1, 2022 – Apr 4, 2022

Subsystems (2)
lore
Share & tools
lore
#0 lore raw

PATCH 1/1 drm/amdgpu: Flush TLB after mapping for VG20+XGMI

For VG20 + XGMI bridge, all mappings PTEs cache in TC, this may have stall invalid PTEs in TC because one cache line has 8 pages. Need always flush_tlb after updating mapping.

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 6 ++++++
 1 file changed, 6 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +6 -0
index f0aec04111a3..687c9a140645 100644
@@ -837,6 +837,12 @@ int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
 		goto error_unlock;
 	}
 
+	/* Vega20+XGMI where PTEs get inadvertently cached in L2 texture cache,
+	 * heavy-weight flush TLB unconditionally.
+	 */
+	flush_tlb |= (adev->gmc.xgmi.num_physical_nodes &&
+		      adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 0));
+
 	memset(&params, 0, sizeof(params));
 	params.adev = adev;
 	params.vm = vm;
Signature · 1 line
2.35.1
#1 lore raw

Am 01.04.22 um 21:57 schrieb Philip Yang:

Quoted 5 lines

For VG20 + XGMI bridge, all mappings PTEs cache in TC, this may have stall invalid PTEs in TC because one cache line has 8 pages. Need always flush_tlb after updating mapping.

You could drop the extra (), but either way Reviewed-by: Christian König <[email protected]>.

Quoted 21 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 6 ++++++
  1 file changed, 6 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +6 -0
index f0aec04111a3..687c9a140645 100644
@@ -837,6 +837,12 @@ int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
  		goto error_unlock;
  	}
 
+	/* Vega20+XGMI where PTEs get inadvertently cached in L2 texture cache,
+	 * heavy-weight flush TLB unconditionally.
+	 */
+	flush_tlb |= (adev->gmc.xgmi.num_physical_nodes &&
+		      adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 0));
+
  	memset(&params, 0, sizeof(params));
  	params.adev = adev;
  	params.vm = vm;
#2 lore raw
In reply to Philip Yang

Dear Philip,

Thank you for your patch.

Am 01.04.22 um 21:57 schrieb Philip Yang:

Quoted 2 lines

For VG20 + XGMI bridge, all mappings PTEs cache in TC, this may have stall invalid PTEs in TC because one cache line has 8 pages. Need always

Can you please rephrase. “may have stall …” is really hard to understand.

Quoted 1 line

flush_tlb after updating mapping.

Maybe:

So, always flush_tlb after updating the mapping.

Quoted 12 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 6 ++++++
  1 file changed, 6 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index f0aec04111a3..687c9a140645 100644
@@ -837,6 +837,12 @@ int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
  		goto error_unlock;
  	}

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm, · line 848

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +6 -0
 
+	/* Vega20+XGMI where PTEs get inadvertently cached in L2 texture cache,
+	 * heavy-weight flush TLB unconditionally.
+	 */
+	flush_tlb |= (adev->gmc.xgmi.num_physical_nodes &&
+		      adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 0));
+
  	memset(&params, 0, sizeof(params));
  	params.adev = adev;
  	params.vm = vm;

Did you do any performance measurement, if that flushing affects anything?

Kind regards,

Paul

#3 lore raw

On 2022-04-04 12:19, Paul Menzel wrote:
Dear Philip,

Thank you for your patch.

Am 01.04.22 um 21:57 schrieb Philip Yang:
For VG20 + XGMI bridge, all mappings PTEs cache in TC, this may have stall invalid PTEs in TC because one cache line has 8 pages. Need always

Can you please rephrase. “may have stall …” is really hard to understand. The patch already pushed and merged.

flush_tlb after updating mapping.

Maybe:

So, always flush_tlb after updating the mapping.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 6 ++++++

  1 file changed, 6 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index f0aec04111a3..687c9a140645 100644
@@ -837,6 +837,12 @@ int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,

          goto error_unlock;
      }
  +    /* Vega20+XGMI where PTEs get inadvertently cached in L2 texture cache, +     * heavy-weight flush TLB unconditionally.
+     */
+    flush_tlb |= (adev->gmc.xgmi.num_physical_nodes &&
+              adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 0)); +
      memset(&params, 0, sizeof(params));
      params.adev = adev;
      params.vm = vm;

Did you do any performance measurement, if that flushing affects anything?
There was another patch to optimize TLB flush to improve map to GPU performance, for this config, always flush TLB after updating mapping is the original performance before the optimization.
Regards,
Philip

Kind regards,

Paul

#5 lore raw
In reply to Paul Menzel

Am 2022-04-04 um 12:19 schrieb Paul Menzel:

Quoted 33 lines

Dear Philip,

Thank you for your patch.

Am 01.04.22 um 21:57 schrieb Philip Yang:

Quoted 2 lines

For VG20 + XGMI bridge, all mappings PTEs cache in TC, this may have stall invalid PTEs in TC because one cache line has 8 pages. Need always

Can you please rephrase. “may have stall …” is really hard to understand.

Quoted 1 line

flush_tlb after updating mapping.

Maybe:

So, always flush_tlb after updating the mapping.

Quoted 16 lines

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 6 ++++++

  1 file changed, 6 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +2 -0
index f0aec04111a3..687c9a140645 100644
@@ -837,6 +837,12 @@ int amdgpu_vm_update_range(struct amdgpu_device 
 *adev, struct amdgpu_vm *vm,
           goto error_unlock;
       }
   +    /* Vega20+XGMI where PTEs get inadvertently cached in L2 texture cache,
+     * heavy-weight flush TLB unconditionally.
+     */
Quoted 8 lines
Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +3 -0
+    flush_tlb |= (adev->gmc.xgmi.num_physical_nodes &&
+              adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 0));
+
       memset(&params, 0, sizeof(params));
       params.adev = adev;

      params.vm = vm;

Did you do any performance measurement, if that flushing affects anything?

There probably is a performance impact. However, this fix is needed for correctness. And it is very specific to this GPU configuration (at least two Vega20 GPUs with an XGMI bridge between them). This fixes a regression that was introduced by recent changes to the TLB flushing code. The workaround was originally introduced maybe 4 years ago.

Regards,
  Felix

Quoted 5 lines

Kind regards,

Paul

#4 lore raw
In reply to philip yang

Dear Philip,

Thank you for your response.

Am 04.04.22 um 18:52 schrieb philip yang:

Quoted 8 lines · Paul Menzel
Quoted 6 lines

Am 01.04.22 um 21:57 schrieb Philip Yang:

Quoted 2 lines

For VG20 + XGMI bridge, all mappings PTEs cache in TC, this may have stall invalid PTEs in TC because one cache line has 8 pages. Need always

Can you please rephrase. “may have stall …” is really hard to understand.

The patch already pushed and merged.

Sorry, but that really sucks, that commits with such a hard to understand commit message are merged.

Quoted 31 lines
Quoted 22 lines
Quoted 1 line

flush_tlb after updating mapping.

Maybe:

So, always flush_tlb after updating the mapping.

Quoted 16 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 6 ++++++
  1 file changed, 6 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +2 -0
index f0aec04111a3..687c9a140645 100644
@@ -837,6 +837,12 @@ int amdgpu_vm_update_range(struct amdgpu_device *adev, 
 struct amdgpu_vm *vm,
          goto error_unlock;
      }
  +    /* Vega20+XGMI where PTEs get inadvertently cached in L2 texture cache,
+     * heavy-weight flush TLB unconditionally.
+     */
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_update_range(struct amdgpu_device *adev,

Quoted 6 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +3 -0
+    flush_tlb |= (adev->gmc.xgmi.num_physical_nodes &&
+              adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 0));
+
      memset(&params, 0, sizeof(params));
      params.adev = adev;
      params.vm = vm;

Did you do any performance measurement, if that flushing affects anything?

Quoted 3 lines

There was another patch to optimize TLB flush to improve map to GPU performance, for this config, always flush TLB after updating mapping is the original performance before the optimization.

So, why didn’t you reference this commit in the commit message, and also did not give that information in the commit message?

Kind regards,

Paul

#6 lore raw
In reply to Philip Yang
Quoted 16 lines · Philip Yang

For VG20 + XGMI bridge, all mappings PTEs cache in TC, this may have stall invalid PTEs in TC because one cache line has 8 pages. Need always flush_tlb after updating mapping.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 6 ++++++
  1 file changed, 6 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index f0aec04111a3..687c9a140645 100644
@@ -837,6 +837,12 @@ int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
  		goto error_unlock;
  	}

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm, · line 848

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +6 -0
 
+	/* Vega20+XGMI where PTEs get inadvertently cached in L2 texture cache,
+	 * heavy-weight flush TLB unconditionally.
+	 */
+	flush_tlb |= (adev->gmc.xgmi.num_physical_nodes &&
+		      adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 0));
+
  	memset(&params, 0, sizeof(params));
  	params.adev = adev;
  	params.vm = vm;

On top of what commit does this work?

It does not apply at top of v5.18-rc1.

It does apply, but fails to compile, on top of "drm/amdkfd: Create file descriptor after client is added to smi_clients list" that is commit:
* cbe879c87245ce6272afe6456dbc8ce2c8f38d64 in amd-staging-drm-next
* e45422695c196dbc665a95526c85ff4b8752aff2 in drm-next
fetched from https://gitlab.freedesktop.org/agd5f/linux.git

The compile error is due to flush_tlb being undeclared.

Best Regards,
Tomasz Mon

#7 lore raw
Quoted 31 lines · Tomasz Moń
Quoted 16 lines · Philip Yang

For VG20 + XGMI bridge, all mappings PTEs cache in TC, this may have stall invalid PTEs in TC because one cache line has 8 pages. Need always flush_tlb after updating mapping.

---
  drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 6 ++++++
  1 file changed, 6 insertions(+)
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +0 -0
index f0aec04111a3..687c9a140645 100644
@@ -837,6 +837,12 @@ int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,
              goto error_unlock;
      }

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm, · line 848

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +6 -0
 
+     /* Vega20+XGMI where PTEs get inadvertently cached in L2 texture cache,
+      * heavy-weight flush TLB unconditionally.
+      */
+     flush_tlb |= (adev->gmc.xgmi.num_physical_nodes &&
+                   adev->ip_versions[GC_HWIP][0] == IP_VERSION(9, 4, 0));
+
      memset(&params, 0, sizeof(params));
      params.adev = adev;
      params.vm = vm;

On top of what commit does this work?

Quoted 1 line

It does not apply at top of v5.18-rc1.

This is stuff for drm-next
(https://gitlab.freedesktop.org/agd5f/linux/-/commits/drm-next).
E.g., 5.19.

Alex

Quoted 11 lines

It does apply, but fails to compile, on top of "drm/amdkfd: Create file descriptor after client is added to smi_clients list" that is commit:
* cbe879c87245ce6272afe6456dbc8ce2c8f38d64 in amd-staging-drm-next
* e45422695c196dbc665a95526c85ff4b8752aff2 in drm-next
fetched from https://gitlab.freedesktop.org/agd5f/linux.git

The compile error is due to flush_tlb being undeclared.

Best Regards,
Tomasz Mon

#8 lore raw
Quoted 8 lines · Alex Deucher
Quoted 3 lines · Tomasz Moń

On top of what commit does this work?

It does not apply at top of v5.18-rc1.

This is stuff for drm-next
(https://gitlab.freedesktop.org/agd5f/linux/-/commits/drm-next).
E.g., 5.19.

This does not compile on current drm-next, just like I noted in the next paragraph of my email.

Quoted 10 lines
Quoted 10 lines

It does apply, but fails to compile, on top of "drm/amdkfd: Create file
descriptor after client is added to smi_clients list" that is commit:
   * cbe879c87245ce6272afe6456dbc8ce2c8f38d64 in amd-staging-drm- next
   * e45422695c196dbc665a95526c85ff4b8752aff2 in drm-next fetched from https://gitlab.freedesktop.org/agd5f/linux.git

The compile error is due to flush_tlb being undeclared.

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c: In function ‘amdgpu_vm_bo_update_mapping’:
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c:845:9: error: ‘flush_tlb’ undeclared (first use in this function); did you mean ‘kfd_flush_tlb’?

  845 |         flush_tlb |= (adev->gmc.xgmi.num_physical_nodes &&
      |         ^~~~~~~~~
      |         kfd_flush_tlb

drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c:845:9: note: each undeclared identifier is reported only once for each function it appears in

Best Regards,
Tomasz Mon