Skip to contentSkip to conversation

Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

PATCH In review Kent Russell 6 participants 14 messages Apr 13, 2020 – Apr 17, 2020

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

PATCH 1/1 Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected!
amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues
Failed to suspend process 0x8000
Finished evicting pasid 0x8000
Started restoring pasid 0x8000
Finished restoring pasid 0x8000
[drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu features!
amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

---
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
 1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -26
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos,
 	uint32_t hi = ~0;
 	uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-	last = min(pos + size, adev->gmc.visible_vram_size);
-	if (last > pos) {
-		void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-		size_t count = last - pos;
-
-		if (write) {
-			memcpy_toio(addr, buf, count);
-			mb();
-			amdgpu_asic_flush_hdp(adev, NULL);
-		} else {
-			amdgpu_asic_invalidate_hdp(adev, NULL);
-			mb();
-			memcpy_fromio(buf, addr, count);
-		}
-
-		if (count == size)
-			return;
-
-		pos += count;
-		buf += count / 4;
-		size -= count;
-	}
-#endif
-
 	spin_lock_irqsave(&adev->mmio_idx_lock, flags);
 	for (last = pos + size; pos < last; pos += 4) {
 		uint32_t tmp = pos >> 31;
Signature · 6 lines
2.17.1

_______________________________________________
amd-gfx mailing list
[email protected]
https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Reviewed-by on v1 1/1
#1 lore raw

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 18 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected!
amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues
Failed to suspend process 0x8000
Finished evicting pasid 0x8000
Started restoring pasid 0x8000
Finished restoring pasid 0x8000
[drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT

Quoted 8 lines

amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu features!
amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines
Quoted 31 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -19
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos,
  	uint32_t hi = ~0;
  	uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-	last = min(pos + size, adev->gmc.visible_vram_size);
-	if (last > pos) {
-		void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-		size_t count = last - pos;
-
-		if (write) {
-			memcpy_toio(addr, buf, count);
-			mb();
-			amdgpu_asic_flush_hdp(adev, NULL);
-		} else {
-			amdgpu_asic_invalidate_hdp(adev, NULL);
-			mb();
-			memcpy_fromio(buf, addr, count);
-		}
-
-		if (count == size)
-			return;

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, · line 259

Quoted 10 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -7
-
-		pos += count;
-		buf += count / 4;
-		size -= count;
-	}
-#endif
-
  	spin_lock_irqsave(&adev->mmio_idx_lock, flags);
  	for (last = pos + size; pos < last; pos += 4) {
  		uint32_t tmp = pos >> 31;

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

#2 lore raw

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]>; [email protected]
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
  	uint32_t hi = ~0;
  	uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-	last = min(pos + size, adev->gmc.visible_vram_size);
-	if (last > pos) {
-		void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-		size_t count = last - pos;
-
-		if (write) {
-			memcpy_toio(addr, buf, count);
-			mb();
-			amdgpu_asic_flush_hdp(adev, NULL);
-		} else {
-			amdgpu_asic_invalidate_hdp(adev, NULL);
-			mb();
-			memcpy_fromio(buf, addr, count);
-		}
-
-		if (count == size)
-			return;
-
-		pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-		buf += count / 4;
-		size -= count;
-	}
-#endif
-
  	spin_lock_irqsave(&adev->mmio_idx_lock, flags);
  	for (last = pos + size; pos < last; pos += 4) {
  		uint32_t tmp = pos >> 31;

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

#3 lore raw

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]> on behalf of Russell, Kent <[email protected]>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]>; [email protected] <[email protected]>
Cc: Kuehling, Felix <[email protected]>; Kim, Jonathan <[email protected]>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]>; [email protected]
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

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

#4 lore raw

That's exactly my concern as well.

This looks a bit like the test creates erroneous data somehow, but there doesn't seems to be a RAS check in the MM data path.

And now that we use the BAR path it goes up in flames.

I just don't see how we can create erroneous data in a test case?

Christian.

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]> on behalf of Russell, Kent <[email protected]>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]>; [email protected] <[email protected]>
Cc: Kuehling, Felix <[email protected]>; Kim, Jonathan <[email protected]>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]>; [email protected]
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]> on behalf of Russell, Kent <[email protected]>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]>; [email protected] <[email protected]>
Cc: Kuehling, Felix <[email protected]>; Kim, Jonathan <[email protected]>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]>; [email protected]
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]> on behalf of Russell, Kent <[email protected]>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]>; [email protected] <[email protected]>
Cc: Kuehling, Felix <[email protected]>; Kim, Jonathan <[email protected]>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]>; [email protected]
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]> on behalf of Russell, Kent <[email protected]>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]>; [email protected] <[email protected]>
Cc: Kuehling, Felix <[email protected]>; Kim, Jonathan <[email protected]>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]>; [email protected]
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]> on behalf of Russell, Kent <[email protected]>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]>; [email protected] <[email protected]>
Cc: Kuehling, Felix <[email protected]>; Kim, Jonathan <[email protected]>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]>; [email protected]
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list
[email protected]
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

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

#5 lore raw

[AMD Official Use Only - Internal Distribution Only]

I think it's premature to push this revert.

With more testing, I'm getting failures from different tests or sometimes none at all on my machine.

Kent, let's continue the discussion on the original thread.

Thanks,

Jon

From: Koenig, Christian <[email protected]>
Sent: Tuesday, April 14, 2020 10:47 AM
To: Deucher, Alexander <[email protected]>
Cc: Russell, Kent <[email protected]>; [email protected]; Kuehling, Felix <[email protected]>; Kim, Jonathan <[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

That's exactly my concern as well.

This looks a bit like the test creates erroneous data somehow, but there doesn't seems to be a RAS check in the MM data path.

And now that we use the BAR path it goes up in flames.

I just don't see how we can create erroneous data in a test case?

Christian.

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

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

#6 lore raw

I wouldn't call it premature. Revert is a usual practice when there is a serious regression that isn't fully understood or root-caused. As far as I can tell, the problem has been reproduced on multiple systems, different GPUs, and clearly regressed to Christian's commit. I think that justifies reverting it for now.

I agree with Christian that a general HDP memory access problem causing RAS errors would potentially cause problems in other tests as well. For example common operations like GART table updates, and GPUVM page table updates and PCIe peer2peer accesses in ROCm applications use HDP. But we're not seeing obvious problems from those. So we need to understand what's special about this test. I asked questions to that effect on our other email thread.

Regards,
  Felix

Am 2020-04-14 um 10:51 a.m. schrieb Kim, Jonathan:

Quoted 798 lines

[AMD Official Use Only - Internal Distribution Only]

I think it’s premature to push this revert.

With more testing, I’m getting failures from different tests or sometimes none at all on my machine.

Kent, let’s continue the discussion on the original thread.

Thanks,

Jon

*From:* Koenig, Christian <[email protected]> *Sent:* Tuesday, April 14, 2020 10:47 AM
*To:* Deucher, Alexander <[email protected]>
*Cc:* Russell, Kent <[email protected]>; [email protected]; Kuehling, Felix <[email protected]>; Kim, Jonathan <[email protected]> *Subject:* Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

That's exactly my concern as well.

This looks a bit like the test creates erroneous data somehow, but there doesn't seems to be a RAS check in the MM data path.

And now that we use the BAR path it goes up in flames.

I just don't see how we can create erroneous data in a test case?

Christian.

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected] <mailto:[email protected]>>:

    [AMD Public Use]

     

    If this causes an issue, any access to vram via the BAR could
    cause an issue.

     

    Alex

    ------------------------------------------------------------------------

    *From:*amd-gfx <[email protected]
    <mailto:[email protected]>> on behalf of
    Russell, Kent <[email protected] <mailto:[email protected]>>
    *Sent:* Tuesday, April 14, 2020 10:19 AM
    *To:* Koenig, Christian <[email protected]
    <mailto:[email protected]>>; [email protected]
    <mailto:[email protected]>
    <[email protected] <mailto:[email protected]>>
    *Cc:* Kuehling, Felix <[email protected]
    <mailto:[email protected]>>; Kim, Jonathan
    <[email protected] <mailto:[email protected]>>
    *Subject:* RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible
    in amdgpu_device_vram_access v2"

     

    [AMD Official Use Only - Internal Distribution Only]

    On VG20 or MI100, as soon as we run the subtest, we get the dmesg
    output below, and then the kernel ends up hanging. I don't know
    enough about the test itself to know why this is occurring, but
    Jon Kim and Felix were discussing it on a separate thread when the
    issue was first reported, so they can hopefully provide some
    additional information.

     Kent

    > -----Original Message-----
    > From: Christian König <[email protected]
    <mailto:[email protected]>>
    > Sent: Tuesday, April 14, 2020 9:52 AM
    > To: Russell, Kent <[email protected]
    <mailto:[email protected]>>; [email protected]
    <mailto:[email protected]>
    > Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
    > amdgpu_device_vram_access v2"
    >
    > Am 13.04.20 um 20:20 schrieb Kent Russell:
    > > This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e.
    > > The original patch causes a RAS event and subsequent kernel
    hard-hang
    > > when running the KFDMemoryTest.PtraceAccessInvisibleVram on
    VG20 and
    > > Arcturus
    > >
    > > dmesg output at hang time:
    > > [drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected!
    > > amdgpu 0000:67:00.0: GPU reset begin!
    > > Evicting PASID 0x8000 queues
    > > Started evicting pasid 0x8000
    > > qcm fence wait loop timeout expired
    > > The cp might be in an unrecoverable state due to an unsuccessful
    > > queues preemption Failed to evict process queues Failed to suspend
    > > process 0x8000 Finished evicting pasid 0x8000 Started
    restoring pasid
    > > 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state
    may lost
    > > due to RAS ERREVENT_ATHUB_INTERRUPT
    > > amdgpu: [powerplay] Failed to send message 0x26, response 0x0
    > > amdgpu: [powerplay] Failed to set soft min gfxclk !
    > > amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
    > > amdgpu: [powerplay] Failed to send message 0x7, response 0x0
    > > amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable
    all smu
    > features!
    > > amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all
    smu features!
    > > amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM!
    > > [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend
    of IP
    > > block <powerplay> failed -5
    >
    > Do you have more information on what's going wrong here since
    this is a really
    > important patch for KFD debugging.
    >
    > >
    > > Signed-off-by: Kent Russell <[email protected]
    <mailto:[email protected]>>
    >
    > Reviewed-by: Christian König <[email protected]
    <mailto:[email protected]>>
    >
    > > ---
    > >   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26
    ----------------------
    > >   1 file changed, 26 deletions(-)
    > >
    > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
    > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
    > > index cf5d6e585634..a3f997f84020 100644
    > > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
    > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
    > > @@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct
    > amdgpu_device *adev, loff_t pos,
    > >      uint32_t hi = ~0;
    > >      uint64_t last;
    > >
    > > -
    > > -#ifdef CONFIG_64BIT
    > > -   last = min(pos + size, adev->gmc.visible_vram_size);
    > > -   if (last > pos) {
    > > -           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
    > > -           size_t count = last - pos;
    > > -
    > > -           if (write) {
    > > -                   memcpy_toio(addr, buf, count);
    > > -                   mb();
    > > -                   amdgpu_asic_flush_hdp(adev, NULL);
    > > -           } else {
    > > -                   amdgpu_asic_invalidate_hdp(adev, NULL);
    > > -                   mb();
    > > -                   memcpy_fromio(buf, addr, count);
    > > -           }
    > > -
    > > -           if (count == size)
    > > -                   return;
    > > -
    > > -           pos += count;
    > > -           buf += count / 4;
    > > -           size -= count;
    > > -   }
    > > -#endif
    > > -
    > >      spin_lock_irqsave(&adev->mmio_idx_lock, flags);
    > >      for (last = pos + size; pos < last; pos += 4) {
    > >              uint32_t tmp = pos >> 31;
    _______________________________________________
    amd-gfx mailing list
    [email protected] <mailto:[email protected]>
    https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected] <mailto:[email protected]>>:

    [AMD Public Use]

     

    If this causes an issue, any access to vram via the BAR could
    cause an issue.

     

    Alex

    ------------------------------------------------------------------------

    *From:*amd-gfx <[email protected]
    <mailto:[email protected]>> on behalf of
    Russell, Kent <[email protected] <mailto:[email protected]>>
    *Sent:* Tuesday, April 14, 2020 10:19 AM
    *To:* Koenig, Christian <[email protected]
    <mailto:[email protected]>>; [email protected]
    <mailto:[email protected]>
    <[email protected] <mailto:[email protected]>>
    *Cc:* Kuehling, Felix <[email protected]
    <mailto:[email protected]>>; Kim, Jonathan
    <[email protected] <mailto:[email protected]>>
    *Subject:* RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible
    in amdgpu_device_vram_access v2"

     

    [AMD Official Use Only - Internal Distribution Only]

    On VG20 or MI100, as soon as we run the subtest, we get the dmesg
    output below, and then the kernel ends up hanging. I don't know
    enough about the test itself to know why this is occurring, but
    Jon Kim and Felix were discussing it on a separate thread when the
    issue was first reported, so they can hopefully provide some
    additional information.

     Kent

    > -----Original Message-----
    > From: Christian König <[email protected]
    <mailto:[email protected]>>
    > Sent: Tuesday, April 14, 2020 9:52 AM
    > To: Russell, Kent <[email protected]
    <mailto:[email protected]>>; [email protected]
    <mailto:[email protected]>
    > Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
    > amdgpu_device_vram_access v2"
    >
    > Am 13.04.20 um 20:20 schrieb Kent Russell:
    > > This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e.
    > > The original patch causes a RAS event and subsequent kernel
    hard-hang
    > > when running the KFDMemoryTest.PtraceAccessInvisibleVram on
    VG20 and
    > > Arcturus
    > >
    > > dmesg output at hang time:
    > > [drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected!
    > > amdgpu 0000:67:00.0: GPU reset begin!
    > > Evicting PASID 0x8000 queues
    > > Started evicting pasid 0x8000
    > > qcm fence wait loop timeout expired
    > > The cp might be in an unrecoverable state due to an unsuccessful
    > > queues preemption Failed to evict process queues Failed to suspend
    > > process 0x8000 Finished evicting pasid 0x8000 Started
    restoring pasid
    > > 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state
    may lost
    > > due to RAS ERREVENT_ATHUB_INTERRUPT
    > > amdgpu: [powerplay] Failed to send message 0x26, response 0x0
    > > amdgpu: [powerplay] Failed to set soft min gfxclk !
    > > amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
    > > amdgpu: [powerplay] Failed to send message 0x7, response 0x0
    > > amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable
    all smu
    > features!
    > > amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all
    smu features!
    > > amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM!
    > > [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend
    of IP
    > > block <powerplay> failed -5
    >
    > Do you have more information on what's going wrong here since
    this is a really
    > important patch for KFD debugging.
    >
    > >
    > > Signed-off-by: Kent Russell <[email protected]
    <mailto:[email protected]>>
    >
    > Reviewed-by: Christian König <[email protected]
    <mailto:[email protected]>>
    >
    > > ---
    > >   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26
    ----------------------
    > >   1 file changed, 26 deletions(-)
    > >
    > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
    > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
    > > index cf5d6e585634..a3f997f84020 100644
    > > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
    > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
    > > @@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct
    > amdgpu_device *adev, loff_t pos,
    > >      uint32_t hi = ~0;
    > >      uint64_t last;
    > >
    > > -
    > > -#ifdef CONFIG_64BIT
    > > -   last = min(pos + size, adev->gmc.visible_vram_size);
    > > -   if (last > pos) {
    > > -           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
    > > -           size_t count = last - pos;
    > > -
    > > -           if (write) {
    > > -                   memcpy_toio(addr, buf, count);
    > > -                   mb();
    > > -                   amdgpu_asic_flush_hdp(adev, NULL);
    > > -           } else {
    > > -                   amdgpu_asic_invalidate_hdp(adev, NULL);
    > > -                   mb();
    > > -                   memcpy_fromio(buf, addr, count);
    > > -           }
    > > -
    > > -           if (count == size)
    > > -                   return;
    > > -
    > > -           pos += count;
    > > -           buf += count / 4;
    > > -           size -= count;
    > > -   }
    > > -#endif
    > > -
    > >      spin_lock_irqsave(&adev->mmio_idx_lock, flags);
    > >      for (last = pos + size; pos < last; pos += 4) {
    > >              uint32_t tmp = pos >> 31;
    _______________________________________________
    amd-gfx mailing list
    [email protected] <mailto:[email protected]>
    https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected] <mailto:[email protected]>>:

    [AMD Public Use]

     

    If this causes an issue, any access to vram via the BAR could
    cause an issue.

     

    Alex

    ------------------------------------------------------------------------

    *From:*amd-gfx <[email protected]
    <mailto:[email protected]>> on behalf of
    Russell, Kent <[email protected] <mailto:[email protected]>>
    *Sent:* Tuesday, April 14, 2020 10:19 AM
    *To:* Koenig, Christian <[email protected]
    <mailto:[email protected]>>; [email protected]
    <mailto:[email protected]>
    <[email protected] <mailto:[email protected]>>
    *Cc:* Kuehling, Felix <[email protected]
    <mailto:[email protected]>>; Kim, Jonathan
    <[email protected] <mailto:[email protected]>>
    *Subject:* RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible
    in amdgpu_device_vram_access v2"

     

    [AMD Official Use Only - Internal Distribution Only]

    On VG20 or MI100, as soon as we run the subtest, we get the dmesg
    output below, and then the kernel ends up hanging. I don't know
    enough about the test itself to know why this is occurring, but
    Jon Kim and Felix were discussing it on a separate thread when the
    issue was first reported, so they can hopefully provide some
    additional information.

     Kent

    > -----Original Message-----
    > From: Christian König <[email protected]
    <mailto:[email protected]>>
    > Sent: Tuesday, April 14, 2020 9:52 AM
    > To: Russell, Kent <[email protected]
    <mailto:[email protected]>>; [email protected]
    <mailto:[email protected]>
    > Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
    > amdgpu_device_vram_access v2"
    >
    > Am 13.04.20 um 20:20 schrieb Kent Russell:
    > > This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e.
    > > The original patch causes a RAS event and subsequent kernel
    hard-hang
    > > when running the KFDMemoryTest.PtraceAccessInvisibleVram on
    VG20 and
    > > Arcturus
    > >
    > > dmesg output at hang time:
    > > [drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected!
    > > amdgpu 0000:67:00.0: GPU reset begin!
    > > Evicting PASID 0x8000 queues
    > > Started evicting pasid 0x8000
    > > qcm fence wait loop timeout expired
    > > The cp might be in an unrecoverable state due to an unsuccessful
    > > queues preemption Failed to evict process queues Failed to suspend
    > > process 0x8000 Finished evicting pasid 0x8000 Started
    restoring pasid
    > > 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state
    may lost
    > > due to RAS ERREVENT_ATHUB_INTERRUPT
    > > amdgpu: [powerplay] Failed to send message 0x26, response 0x0
    > > amdgpu: [powerplay] Failed to set soft min gfxclk !
    > > amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
    > > amdgpu: [powerplay] Failed to send message 0x7, response 0x0
    > > amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable
    all smu
    > features!
    > > amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all
    smu features!
    > > amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM!
    > > [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend
    of IP
    > > block <powerplay> failed -5
    >
    > Do you have more information on what's going wrong here since
    this is a really
    > important patch for KFD debugging.
    >
    > >
    > > Signed-off-by: Kent Russell <[email protected]
    <mailto:[email protected]>>
    >
    > Reviewed-by: Christian König <[email protected]
    <mailto:[email protected]>>
    >
    > > ---
    > >   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26
    ----------------------
    > >   1 file changed, 26 deletions(-)
    > >
    > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
    > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
    > > index cf5d6e585634..a3f997f84020 100644
    > > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
    > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
    > > @@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct
    > amdgpu_device *adev, loff_t pos,
    > >      uint32_t hi = ~0;
    > >      uint64_t last;
    > >
    > > -
    > > -#ifdef CONFIG_64BIT
    > > -   last = min(pos + size, adev->gmc.visible_vram_size);
    > > -   if (last > pos) {
    > > -           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
    > > -           size_t count = last - pos;
    > > -
    > > -           if (write) {
    > > -                   memcpy_toio(addr, buf, count);
    > > -                   mb();
    > > -                   amdgpu_asic_flush_hdp(adev, NULL);
    > > -           } else {
    > > -                   amdgpu_asic_invalidate_hdp(adev, NULL);
    > > -                   mb();
    > > -                   memcpy_fromio(buf, addr, count);
    > > -           }
    > > -
    > > -           if (count == size)
    > > -                   return;
    > > -
    > > -           pos += count;
    > > -           buf += count / 4;
    > > -           size -= count;
    > > -   }
    > > -#endif
    > > -
    > >      spin_lock_irqsave(&adev->mmio_idx_lock, flags);
    > >      for (last = pos + size; pos < last; pos += 4) {
    > >              uint32_t tmp = pos >> 31;
    _______________________________________________
    amd-gfx mailing list
    [email protected] <mailto:[email protected]>
    https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected] <mailto:[email protected]>>:

    [AMD Public Use]

     

    If this causes an issue, any access to vram via the BAR could
    cause an issue.

     

    Alex

    ------------------------------------------------------------------------

    *From:*amd-gfx <[email protected]
    <mailto:[email protected]>> on behalf of
    Russell, Kent <[email protected] <mailto:[email protected]>>
    *Sent:* Tuesday, April 14, 2020 10:19 AM
    *To:* Koenig, Christian <[email protected]
    <mailto:[email protected]>>; [email protected]
    <mailto:[email protected]>
    <[email protected] <mailto:[email protected]>>
    *Cc:* Kuehling, Felix <[email protected]
    <mailto:[email protected]>>; Kim, Jonathan
    <[email protected] <mailto:[email protected]>>
    *Subject:* RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible
    in amdgpu_device_vram_access v2"

     

    [AMD Official Use Only - Internal Distribution Only]

    On VG20 or MI100, as soon as we run the subtest, we get the dmesg
    output below, and then the kernel ends up hanging. I don't know
    enough about the test itself to know why this is occurring, but
    Jon Kim and Felix were discussing it on a separate thread when the
    issue was first reported, so they can hopefully provide some
    additional information.

     Kent

    > -----Original Message-----
    > From: Christian König <[email protected]
    <mailto:[email protected]>>
    > Sent: Tuesday, April 14, 2020 9:52 AM
    > To: Russell, Kent <[email protected]
    <mailto:[email protected]>>; [email protected]
    <mailto:[email protected]>
    > Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
    > amdgpu_device_vram_access v2"
    >
    > Am 13.04.20 um 20:20 schrieb Kent Russell:
    > > This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e.
    > > The original patch causes a RAS event and subsequent kernel
    hard-hang
    > > when running the KFDMemoryTest.PtraceAccessInvisibleVram on
    VG20 and
    > > Arcturus
    > >
    > > dmesg output at hang time:
    > > [drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected!
    > > amdgpu 0000:67:00.0: GPU reset begin!
    > > Evicting PASID 0x8000 queues
    > > Started evicting pasid 0x8000
    > > qcm fence wait loop timeout expired
    > > The cp might be in an unrecoverable state due to an unsuccessful
    > > queues preemption Failed to evict process queues Failed to suspend
    > > process 0x8000 Finished evicting pasid 0x8000 Started
    restoring pasid
    > > 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state
    may lost
    > > due to RAS ERREVENT_ATHUB_INTERRUPT
    > > amdgpu: [powerplay] Failed to send message 0x26, response 0x0
    > > amdgpu: [powerplay] Failed to set soft min gfxclk !
    > > amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
    > > amdgpu: [powerplay] Failed to send message 0x7, response 0x0
    > > amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable
    all smu
    > features!
    > > amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all
    smu features!
    > > amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM!
    > > [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend
    of IP
    > > block <powerplay> failed -5
    >
    > Do you have more information on what's going wrong here since
    this is a really
    > important patch for KFD debugging.
    >
    > >
    > > Signed-off-by: Kent Russell <[email protected]
    <mailto:[email protected]>>
    >
    > Reviewed-by: Christian König <[email protected]
    <mailto:[email protected]>>
    >
    > > ---
    > >   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26
    ----------------------
    > >   1 file changed, 26 deletions(-)
    > >
    > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
    > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
    > > index cf5d6e585634..a3f997f84020 100644
    > > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
    > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
    > > @@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct
    > amdgpu_device *adev, loff_t pos,
    > >      uint32_t hi = ~0;
    > >      uint64_t last;
    > >
    > > -
    > > -#ifdef CONFIG_64BIT
    > > -   last = min(pos + size, adev->gmc.visible_vram_size);
    > > -   if (last > pos) {
    > > -           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
    > > -           size_t count = last - pos;
    > > -
    > > -           if (write) {
    > > -                   memcpy_toio(addr, buf, count);
    > > -                   mb();
    > > -                   amdgpu_asic_flush_hdp(adev, NULL);
    > > -           } else {
    > > -                   amdgpu_asic_invalidate_hdp(adev, NULL);
    > > -                   mb();
    > > -                   memcpy_fromio(buf, addr, count);
    > > -           }
    > > -
    > > -           if (count == size)
    > > -                   return;
    > > -
    > > -           pos += count;
    > > -           buf += count / 4;
    > > -           size -= count;
    > > -   }
    > > -#endif
    > > -
    > >      spin_lock_irqsave(&adev->mmio_idx_lock, flags);
    > >      for (last = pos + size; pos < last; pos += 4) {
    > >              uint32_t tmp = pos >> 31;
    _______________________________________________
    amd-gfx mailing list
    [email protected] <mailto:[email protected]>
    https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected] <mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

------------------------------------------------------------------------

*From:*amd-gfx <[email protected] <mailto:[email protected]>> on behalf of Russell, Kent <[email protected] <mailto:[email protected]>> *Sent:* Tuesday, April 14, 2020 10:19 AM
*To:* Koenig, Christian <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]> <[email protected] <mailto:[email protected]>>
*Cc:* Kuehling, Felix <[email protected] <mailto:[email protected]>>; Kim, Jonathan <[email protected] <mailto:[email protected]>>
*Subject:* RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

 Kent

Quoted 2 lines

-----Original Message-----
From: Christian König <[email protected]

<mailto:[email protected]>>

Quoted 2 lines

Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]

<mailto:[email protected]>>; [email protected] <mailto:[email protected]>

Quoted 25 lines

Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

Quoted 2 lines

features!

Quoted 1 line

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu

features!

Quoted 5 lines
Quoted 3 lines

amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this

is a really

Quoted 4 lines

important patch for KFD debugging.

Quoted 2 lines

<mailto:[email protected]>>

Quoted 2 lines

<mailto:[email protected]>>

Quoted 3 lines
Quoted 3 lines

---
   drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26

----------------------

Quoted 37 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 8 lines
   1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 28 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -25
       uint32_t hi = ~0;
       uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
-           buf += count / 4;
-           size -= count;
-   }
-#endif
Quoted 9 lines
Quoted 4 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct amdgpu_device *adev, loff_t pos, · line 259

Quoted 4 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -1
-
       spin_lock_irqsave(&adev->mmio_idx_lock, flags);
       for (last = pos + size; pos < last; pos += 4) {
               uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list
[email protected] <mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

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

#7 lore raw

[AMD Official Use Only - Internal Distribution Only]

If we're passing the test on the revert, then the only thing that's different is we're not invalidating HDP and doing a copy to host anymore in amdgpu_device_vram_access since the function is still called in ttm access_memory with BAR.

Also cwsr tests fail on Vega20 with or without the revert with the same RAS error.

Thanks,

Jon

From: Kuehling, Felix <[email protected]>
Sent: Tuesday, April 14, 2020 2:32 PM
To: Kim, Jonathan <[email protected]>; Koenig, Christian <[email protected]>; Deucher, Alexander <[email protected]>
Cc: Russell, Kent <[email protected]>; [email protected]
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

I wouldn't call it premature. Revert is a usual practice when there is a serious regression that isn't fully understood or root-caused. As far as I can tell, the problem has been reproduced on multiple systems, different GPUs, and clearly regressed to Christian's commit. I think that justifies reverting it for now.

I agree with Christian that a general HDP memory access problem causing RAS errors would potentially cause problems in other tests as well. For example common operations like GART table updates, and GPUVM page table updates and PCIe peer2peer accesses in ROCm applications use HDP. But we're not seeing obvious problems from those. So we need to understand what's special about this test. I asked questions to that effect on our other email thread.

Regards,
Felix
Am 2020-04-14 um 10:51 a.m. schrieb Kim, Jonathan:

[AMD Official Use Only - Internal Distribution Only]

I think it's premature to push this revert.

With more testing, I'm getting failures from different tests or sometimes none at all on my machine.

Kent, let's continue the discussion on the original thread.

Thanks,

Jon

From: Koenig, Christian <[email protected]><mailto:[email protected]>
Sent: Tuesday, April 14, 2020 10:47 AM
To: Deucher, Alexander <[email protected]><mailto:[email protected]>
Cc: Russell, Kent <[email protected]><mailto:[email protected]>; [email protected]<mailto:[email protected]>; Kuehling, Felix <[email protected]><mailto:[email protected]>; Kim, Jonathan <[email protected]><mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

That's exactly my concern as well.

This looks a bit like the test creates erroneous data somehow, but there doesn't seems to be a RAS check in the MM data path.

And now that we use the BAR path it goes up in flames.

I just don't see how we can create erroneous data in a test case?

Christian.

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

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

maintainer
#8 lore raw

Hi Jon,

Quoted 1 line

Also cwsr tests fail on Vega20 with or without the revert with the same RAS error.

That sounds like the system/setup has a more general problem.

Could it be that we are seeing RAS errors because there really is some hardware failure, but with the MM path we don't trigger a RAS interrupt?

Thanks,
Christian.

Am 14.04.20 um 22:30 schrieb Kim, Jonathan:

Quoted 792 lines

[AMD Official Use Only - Internal Distribution Only]

If we’re passing the test on the revert, then the only thing that’s different is we’re not invalidating HDP and doing a copy to host anymore in amdgpu_device_vram_access since the function is still called in ttm access_memory with BAR.

Also cwsr tests fail on Vega20 with or without the revert with the same RAS error.

Thanks,

Jon

*From:* Kuehling, Felix <[email protected]>
*Sent:* Tuesday, April 14, 2020 2:32 PM
*To:* Kim, Jonathan <[email protected]>; Koenig, Christian <[email protected]>; Deucher, Alexander <[email protected]> *Cc:* Russell, Kent <[email protected]>; [email protected]
*Subject:* Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

I wouldn't call it premature. Revert is a usual practice when there is a serious regression that isn't fully understood or root-caused. As far as I can tell, the problem has been reproduced on multiple systems, different GPUs, and clearly regressed to Christian's commit. I think that justifies reverting it for now.

I agree with Christian that a general HDP memory access problem causing RAS errors would potentially cause problems in other tests as well. For example common operations like GART table updates, and GPUVM page table updates and PCIe peer2peer accesses in ROCm applications use HDP. But we're not seeing obvious problems from those. So we need to understand what's special about this test. I asked questions to that effect on our other email thread.

Regards,
  Felix

Am 2020-04-14 um 10:51 a.m. schrieb Kim, Jonathan:

    [AMD Official Use Only - Internal Distribution Only]

    I think it’s premature to push this revert.

    With more testing, I’m getting failures from different tests or
    sometimes none at all on my machine.

    Kent, let’s continue the discussion on the original thread.

    Thanks,

    Jon

    *From:* Koenig, Christian <[email protected]>
    <mailto:[email protected]>
    *Sent:* Tuesday, April 14, 2020 10:47 AM
    *To:* Deucher, Alexander <[email protected]>
    <mailto:[email protected]>
    *Cc:* Russell, Kent <[email protected]>
    <mailto:[email protected]>; [email protected]
    <mailto:[email protected]>; Kuehling, Felix
    <[email protected]> <mailto:[email protected]>; Kim,
    Jonathan <[email protected]> <mailto:[email protected]>
    *Subject:* Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible
    in amdgpu_device_vram_access v2"

    That's exactly my concern as well.

    This looks a bit like the test creates erroneous data somehow, but
    there doesn't seems to be a RAS check in the MM data path.

    And now that we use the BAR path it goes up in flames.

    I just don't see how we can create erroneous data in a test case?

    Christian.

    Am 14.04.2020 16:35 schrieb "Deucher, Alexander"
    <[email protected] <mailto:[email protected]>>:

        [AMD Public Use]

        If this causes an issue, any access to vram via the BAR could
        cause an issue.

        Alex

        ------------------------------------------------------------------------

        *From:*amd-gfx <[email protected]
        <mailto:[email protected]>> on behalf of
        Russell, Kent <[email protected] <mailto:[email protected]>>
        *Sent:* Tuesday, April 14, 2020 10:19 AM
        *To:* Koenig, Christian <[email protected]
        <mailto:[email protected]>>;
        [email protected]
        <mailto:[email protected]>
        <[email protected]
        <mailto:[email protected]>>
        *Cc:* Kuehling, Felix <[email protected]
        <mailto:[email protected]>>; Kim, Jonathan
        <[email protected] <mailto:[email protected]>>
        *Subject:* RE: [PATCH] Revert "drm/amdgpu: use the BAR if
        possible in amdgpu_device_vram_access v2"

        [AMD Official Use Only - Internal Distribution Only]

        On VG20 or MI100, as soon as we run the subtest, we get the
        dmesg output below, and then the kernel ends up hanging. I
        don't know enough about the test itself to know why this is
        occurring, but Jon Kim and Felix were discussing it on a
        separate thread when the issue was first reported, so they can
        hopefully provide some additional information.

         Kent

        > -----Original Message-----
        > From: Christian König <[email protected]
        <mailto:[email protected]>>
        > Sent: Tuesday, April 14, 2020 9:52 AM
        > To: Russell, Kent <[email protected]
        <mailto:[email protected]>>; [email protected]
        <mailto:[email protected]>
        > Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if
        possible in
        > amdgpu_device_vram_access v2"
        >
        > Am 13.04.20 um 20:20 schrieb Kent Russell:
        > > This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e.
        > > The original patch causes a RAS event and subsequent
        kernel hard-hang
        > > when running the KFDMemoryTest.PtraceAccessInvisibleVram
        on VG20 and
        > > Arcturus
        > >
        > > dmesg output at hang time:
        > > [drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected!
        > > amdgpu 0000:67:00.0: GPU reset begin!
        > > Evicting PASID 0x8000 queues
        > > Started evicting pasid 0x8000
        > > qcm fence wait loop timeout expired
        > > The cp might be in an unrecoverable state due to an
        unsuccessful
        > > queues preemption Failed to evict process queues Failed to
        suspend
        > > process 0x8000 Finished evicting pasid 0x8000 Started
        restoring pasid
        > > 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU
        state may lost
        > > due to RAS ERREVENT_ATHUB_INTERRUPT
        > > amdgpu: [powerplay] Failed to send message 0x26, response 0x0
        > > amdgpu: [powerplay] Failed to set soft min gfxclk !
        > > amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
        > > amdgpu: [powerplay] Failed to send message 0x7, response 0x0
        > > amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to
        disable all smu
        > features!
        > > amdgpu: [powerplay] [DisableDpmTasks] Failed to disable
        all smu features!
        > > amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM!
        > > [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR*
        suspend of IP
        > > block <powerplay> failed -5
        >
        > Do you have more information on what's going wrong here
        since this is a really
        > important patch for KFD debugging.
        >
        > >
        > > Signed-off-by: Kent Russell <[email protected]
        <mailto:[email protected]>>
        >
        > Reviewed-by: Christian König <[email protected]
        <mailto:[email protected]>>
        >
        > > ---
        > > drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26
        ----------------------
        > >   1 file changed, 26 deletions(-)
        > >
        > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
        > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
        > > index cf5d6e585634..a3f997f84020 100644
        > > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
        > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
        > > @@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct
        > amdgpu_device *adev, loff_t pos,
        > >      uint32_t hi = ~0;
        > >      uint64_t last;
        > >
        > > -
        > > -#ifdef CONFIG_64BIT
        > > -   last = min(pos + size, adev->gmc.visible_vram_size);
        > > -   if (last > pos) {
        > > -           void __iomem *addr =
        adev->mman.aper_base_kaddr + pos;
        > > -           size_t count = last - pos;
        > > -
        > > -           if (write) {
        > > - memcpy_toio(addr, buf, count);
        > > -                   mb();
        > > - amdgpu_asic_flush_hdp(adev, NULL);
        > > -           } else {
        > > - amdgpu_asic_invalidate_hdp(adev, NULL);
        > > -                   mb();
        > > - memcpy_fromio(buf, addr, count);
        > > -           }
        > > -
        > > -           if (count == size)
        > > -                   return;
        > > -
        > > -           pos += count;
        > > -           buf += count / 4;
        > > -           size -= count;
        > > -   }
        > > -#endif
        > > -
        > > spin_lock_irqsave(&adev->mmio_idx_lock, flags);
        > >      for (last = pos + size; pos < last; pos += 4) {
        > >              uint32_t tmp = pos >> 31;
        _______________________________________________
        amd-gfx mailing list
        [email protected]
        <mailto:[email protected]>
        https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

    Am 14.04.2020 16:35 schrieb "Deucher, Alexander"
    <[email protected] <mailto:[email protected]>>:

        [AMD Public Use]

        If this causes an issue, any access to vram via the BAR could
        cause an issue.

        Alex

        ------------------------------------------------------------------------

        *From:*amd-gfx <[email protected]
        <mailto:[email protected]>> on behalf of
        Russell, Kent <[email protected] <mailto:[email protected]>>
        *Sent:* Tuesday, April 14, 2020 10:19 AM
        *To:* Koenig, Christian <[email protected]
        <mailto:[email protected]>>;
        [email protected]
        <mailto:[email protected]>
        <[email protected]
        <mailto:[email protected]>>
        *Cc:* Kuehling, Felix <[email protected]
        <mailto:[email protected]>>; Kim, Jonathan
        <[email protected] <mailto:[email protected]>>
        *Subject:* RE: [PATCH] Revert "drm/amdgpu: use the BAR if
        possible in amdgpu_device_vram_access v2"

        [AMD Official Use Only - Internal Distribution Only]

        On VG20 or MI100, as soon as we run the subtest, we get the
        dmesg output below, and then the kernel ends up hanging. I
        don't know enough about the test itself to know why this is
        occurring, but Jon Kim and Felix were discussing it on a
        separate thread when the issue was first reported, so they can
        hopefully provide some additional information.

         Kent

        > -----Original Message-----
        > From: Christian König <[email protected]
        <mailto:[email protected]>>
        > Sent: Tuesday, April 14, 2020 9:52 AM
        > To: Russell, Kent <[email protected]
        <mailto:[email protected]>>; [email protected]
        <mailto:[email protected]>
        > Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if
        possible in
        > amdgpu_device_vram_access v2"
        >
        > Am 13.04.20 um 20:20 schrieb Kent Russell:
        > > This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e.
        > > The original patch causes a RAS event and subsequent
        kernel hard-hang
        > > when running the KFDMemoryTest.PtraceAccessInvisibleVram
        on VG20 and
        > > Arcturus
        > >
        > > dmesg output at hang time:
        > > [drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected!
        > > amdgpu 0000:67:00.0: GPU reset begin!
        > > Evicting PASID 0x8000 queues
        > > Started evicting pasid 0x8000
        > > qcm fence wait loop timeout expired
        > > The cp might be in an unrecoverable state due to an
        unsuccessful
        > > queues preemption Failed to evict process queues Failed to
        suspend
        > > process 0x8000 Finished evicting pasid 0x8000 Started
        restoring pasid
        > > 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU
        state may lost
        > > due to RAS ERREVENT_ATHUB_INTERRUPT
        > > amdgpu: [powerplay] Failed to send message 0x26, response 0x0
        > > amdgpu: [powerplay] Failed to set soft min gfxclk !
        > > amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
        > > amdgpu: [powerplay] Failed to send message 0x7, response 0x0
        > > amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to
        disable all smu
        > features!
        > > amdgpu: [powerplay] [DisableDpmTasks] Failed to disable
        all smu features!
        > > amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM!
        > > [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR*
        suspend of IP
        > > block <powerplay> failed -5
        >
        > Do you have more information on what's going wrong here
        since this is a really
        > important patch for KFD debugging.
        >
        > >
        > > Signed-off-by: Kent Russell <[email protected]
        <mailto:[email protected]>>
        >
        > Reviewed-by: Christian König <[email protected]
        <mailto:[email protected]>>
        >
        > > ---
        > > drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26
        ----------------------
        > >   1 file changed, 26 deletions(-)
        > >
        > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
        > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
        > > index cf5d6e585634..a3f997f84020 100644
        > > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
        > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
        > > @@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct
        > amdgpu_device *adev, loff_t pos,
        > >      uint32_t hi = ~0;
        > >      uint64_t last;
        > >
        > > -
        > > -#ifdef CONFIG_64BIT
        > > -   last = min(pos + size, adev->gmc.visible_vram_size);
        > > -   if (last > pos) {
        > > -           void __iomem *addr =
        adev->mman.aper_base_kaddr + pos;
        > > -           size_t count = last - pos;
        > > -
        > > -           if (write) {
        > > - memcpy_toio(addr, buf, count);
        > > -                   mb();
        > > - amdgpu_asic_flush_hdp(adev, NULL);
        > > -           } else {
        > > - amdgpu_asic_invalidate_hdp(adev, NULL);
        > > -                   mb();
        > > - memcpy_fromio(buf, addr, count);
        > > -           }
        > > -
        > > -           if (count == size)
        > > -                   return;
        > > -
        > > -           pos += count;
        > > -           buf += count / 4;
        > > -           size -= count;
        > > -   }
        > > -#endif
        > > -
        > > spin_lock_irqsave(&adev->mmio_idx_lock, flags);
        > >      for (last = pos + size; pos < last; pos += 4) {
        > >              uint32_t tmp = pos >> 31;
        _______________________________________________
        amd-gfx mailing list
        [email protected]
        <mailto:[email protected]>
        https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

    Am 14.04.2020 16:35 schrieb "Deucher, Alexander"
    <[email protected] <mailto:[email protected]>>:

        [AMD Public Use]

        If this causes an issue, any access to vram via the BAR could
        cause an issue.

        Alex

        ------------------------------------------------------------------------

        *From:*amd-gfx <[email protected]
        <mailto:[email protected]>> on behalf of
        Russell, Kent <[email protected] <mailto:[email protected]>>
        *Sent:* Tuesday, April 14, 2020 10:19 AM
        *To:* Koenig, Christian <[email protected]
        <mailto:[email protected]>>;
        [email protected]
        <mailto:[email protected]>
        <[email protected]
        <mailto:[email protected]>>
        *Cc:* Kuehling, Felix <[email protected]
        <mailto:[email protected]>>; Kim, Jonathan
        <[email protected] <mailto:[email protected]>>
        *Subject:* RE: [PATCH] Revert "drm/amdgpu: use the BAR if
        possible in amdgpu_device_vram_access v2"

        [AMD Official Use Only - Internal Distribution Only]

        On VG20 or MI100, as soon as we run the subtest, we get the
        dmesg output below, and then the kernel ends up hanging. I
        don't know enough about the test itself to know why this is
        occurring, but Jon Kim and Felix were discussing it on a
        separate thread when the issue was first reported, so they can
        hopefully provide some additional information.

         Kent

        > -----Original Message-----
        > From: Christian König <[email protected]
        <mailto:[email protected]>>
        > Sent: Tuesday, April 14, 2020 9:52 AM
        > To: Russell, Kent <[email protected]
        <mailto:[email protected]>>; [email protected]
        <mailto:[email protected]>
        > Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if
        possible in
        > amdgpu_device_vram_access v2"
        >
        > Am 13.04.20 um 20:20 schrieb Kent Russell:
        > > This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e.
        > > The original patch causes a RAS event and subsequent
        kernel hard-hang
        > > when running the KFDMemoryTest.PtraceAccessInvisibleVram
        on VG20 and
        > > Arcturus
        > >
        > > dmesg output at hang time:
        > > [drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected!
        > > amdgpu 0000:67:00.0: GPU reset begin!
        > > Evicting PASID 0x8000 queues
        > > Started evicting pasid 0x8000
        > > qcm fence wait loop timeout expired
        > > The cp might be in an unrecoverable state due to an
        unsuccessful
        > > queues preemption Failed to evict process queues Failed to
        suspend
        > > process 0x8000 Finished evicting pasid 0x8000 Started
        restoring pasid
        > > 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU
        state may lost
        > > due to RAS ERREVENT_ATHUB_INTERRUPT
        > > amdgpu: [powerplay] Failed to send message 0x26, response 0x0
        > > amdgpu: [powerplay] Failed to set soft min gfxclk !
        > > amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
        > > amdgpu: [powerplay] Failed to send message 0x7, response 0x0
        > > amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to
        disable all smu
        > features!
        > > amdgpu: [powerplay] [DisableDpmTasks] Failed to disable
        all smu features!
        > > amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM!
        > > [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR*
        suspend of IP
        > > block <powerplay> failed -5
        >
        > Do you have more information on what's going wrong here
        since this is a really
        > important patch for KFD debugging.
        >
        > >
        > > Signed-off-by: Kent Russell <[email protected]
        <mailto:[email protected]>>
        >
        > Reviewed-by: Christian König <[email protected]
        <mailto:[email protected]>>
        >
        > > ---
        > > drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26
        ----------------------
        > >   1 file changed, 26 deletions(-)
        > >
        > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
        > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
        > > index cf5d6e585634..a3f997f84020 100644
        > > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
        > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
        > > @@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct
        > amdgpu_device *adev, loff_t pos,
        > >      uint32_t hi = ~0;
        > >      uint64_t last;
        > >
        > > -
        > > -#ifdef CONFIG_64BIT
        > > -   last = min(pos + size, adev->gmc.visible_vram_size);
        > > -   if (last > pos) {
        > > -           void __iomem *addr =
        adev->mman.aper_base_kaddr + pos;
        > > -           size_t count = last - pos;
        > > -
        > > -           if (write) {
        > > - memcpy_toio(addr, buf, count);
        > > -                   mb();
        > > - amdgpu_asic_flush_hdp(adev, NULL);
        > > -           } else {
        > > - amdgpu_asic_invalidate_hdp(adev, NULL);
        > > -                   mb();
        > > - memcpy_fromio(buf, addr, count);
        > > -           }
        > > -
        > > -           if (count == size)
        > > -                   return;
        > > -
        > > -           pos += count;
        > > -           buf += count / 4;
        > > -           size -= count;
        > > -   }
        > > -#endif
        > > -
        > > spin_lock_irqsave(&adev->mmio_idx_lock, flags);
        > >      for (last = pos + size; pos < last; pos += 4) {
        > >              uint32_t tmp = pos >> 31;
        _______________________________________________
        amd-gfx mailing list
        [email protected]
        <mailto:[email protected]>
        https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

    Am 14.04.2020 16:35 schrieb "Deucher, Alexander"
    <[email protected] <mailto:[email protected]>>:

        [AMD Public Use]

        If this causes an issue, any access to vram via the BAR could
        cause an issue.

        Alex

        ------------------------------------------------------------------------

        *From:*amd-gfx <[email protected]
        <mailto:[email protected]>> on behalf of
        Russell, Kent <[email protected] <mailto:[email protected]>>
        *Sent:* Tuesday, April 14, 2020 10:19 AM
        *To:* Koenig, Christian <[email protected]
        <mailto:[email protected]>>;
        [email protected]
        <mailto:[email protected]>
        <[email protected]
        <mailto:[email protected]>>
        *Cc:* Kuehling, Felix <[email protected]
        <mailto:[email protected]>>; Kim, Jonathan
        <[email protected] <mailto:[email protected]>>
        *Subject:* RE: [PATCH] Revert "drm/amdgpu: use the BAR if
        possible in amdgpu_device_vram_access v2"

        [AMD Official Use Only - Internal Distribution Only]

        On VG20 or MI100, as soon as we run the subtest, we get the
        dmesg output below, and then the kernel ends up hanging. I
        don't know enough about the test itself to know why this is
        occurring, but Jon Kim and Felix were discussing it on a
        separate thread when the issue was first reported, so they can
        hopefully provide some additional information.

         Kent

        > -----Original Message-----
        > From: Christian König <[email protected]
        <mailto:[email protected]>>
        > Sent: Tuesday, April 14, 2020 9:52 AM
        > To: Russell, Kent <[email protected]
        <mailto:[email protected]>>; [email protected]
        <mailto:[email protected]>
        > Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if
        possible in
        > amdgpu_device_vram_access v2"
        >
        > Am 13.04.20 um 20:20 schrieb Kent Russell:
        > > This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e.
        > > The original patch causes a RAS event and subsequent
        kernel hard-hang
        > > when running the KFDMemoryTest.PtraceAccessInvisibleVram
        on VG20 and
        > > Arcturus
        > >
        > > dmesg output at hang time:
        > > [drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected!
        > > amdgpu 0000:67:00.0: GPU reset begin!
        > > Evicting PASID 0x8000 queues
        > > Started evicting pasid 0x8000
        > > qcm fence wait loop timeout expired
        > > The cp might be in an unrecoverable state due to an
        unsuccessful
        > > queues preemption Failed to evict process queues Failed to
        suspend
        > > process 0x8000 Finished evicting pasid 0x8000 Started
        restoring pasid
        > > 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU
        state may lost
        > > due to RAS ERREVENT_ATHUB_INTERRUPT
        > > amdgpu: [powerplay] Failed to send message 0x26, response 0x0
        > > amdgpu: [powerplay] Failed to set soft min gfxclk !
        > > amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
        > > amdgpu: [powerplay] Failed to send message 0x7, response 0x0
        > > amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to
        disable all smu
        > features!
        > > amdgpu: [powerplay] [DisableDpmTasks] Failed to disable
        all smu features!
        > > amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM!
        > > [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR*
        suspend of IP
        > > block <powerplay> failed -5
        >
        > Do you have more information on what's going wrong here
        since this is a really
        > important patch for KFD debugging.
        >
        > >
        > > Signed-off-by: Kent Russell <[email protected]
        <mailto:[email protected]>>
        >
        > Reviewed-by: Christian König <[email protected]
        <mailto:[email protected]>>
        >
        > > ---
        > > drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26
        ----------------------
        > >   1 file changed, 26 deletions(-)
        > >
        > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
        > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
        > > index cf5d6e585634..a3f997f84020 100644
        > > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
        > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
        > > @@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct
        > amdgpu_device *adev, loff_t pos,
        > >      uint32_t hi = ~0;
        > >      uint64_t last;
        > >
        > > -
        > > -#ifdef CONFIG_64BIT
        > > -   last = min(pos + size, adev->gmc.visible_vram_size);
        > > -   if (last > pos) {
        > > -           void __iomem *addr =
        adev->mman.aper_base_kaddr + pos;
        > > -           size_t count = last - pos;
        > > -
        > > -           if (write) {
        > > - memcpy_toio(addr, buf, count);
        > > -                   mb();
        > > - amdgpu_asic_flush_hdp(adev, NULL);
        > > -           } else {
        > > - amdgpu_asic_invalidate_hdp(adev, NULL);
        > > -                   mb();
        > > - memcpy_fromio(buf, addr, count);
        > > -           }
        > > -
        > > -           if (count == size)
        > > -                   return;
        > > -
        > > -           pos += count;
        > > -           buf += count / 4;
        > > -           size -= count;
        > > -   }
        > > -#endif
        > > -
        > > spin_lock_irqsave(&adev->mmio_idx_lock, flags);
        > >      for (last = pos + size; pos < last; pos += 4) {
        > >              uint32_t tmp = pos >> 31;
        _______________________________________________
        amd-gfx mailing list
        [email protected]
        <mailto:[email protected]>
        https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

    Am 14.04.2020 16:35 schrieb "Deucher, Alexander"
    <[email protected] <mailto:[email protected]>>:

    [AMD Public Use]

    If this causes an issue, any access to vram via the BAR could
    cause an issue.

    Alex

    ------------------------------------------------------------------------

    *From:*amd-gfx <[email protected]
    <mailto:[email protected]>> on behalf of
    Russell, Kent <[email protected] <mailto:[email protected]>>
    *Sent:* Tuesday, April 14, 2020 10:19 AM
    *To:* Koenig, Christian <[email protected]
    <mailto:[email protected]>>; [email protected]
    <mailto:[email protected]>
    <[email protected] <mailto:[email protected]>>
    *Cc:* Kuehling, Felix <[email protected]
    <mailto:[email protected]>>; Kim, Jonathan
    <[email protected] <mailto:[email protected]>>
    *Subject:* RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible
    in amdgpu_device_vram_access v2"

    [AMD Official Use Only - Internal Distribution Only]

    On VG20 or MI100, as soon as we run the subtest, we get the dmesg
    output below, and then the kernel ends up hanging. I don't know
    enough about the test itself to know why this is occurring, but
    Jon Kim and Felix were discussing it on a separate thread when the
    issue was first reported, so they can hopefully provide some
    additional information.

     Kent

    > -----Original Message-----
    > From: Christian König <[email protected]
    <mailto:[email protected]>>
    > Sent: Tuesday, April 14, 2020 9:52 AM
    > To: Russell, Kent <[email protected]
    <mailto:[email protected]>>; [email protected]
    <mailto:[email protected]>
    > Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
    > amdgpu_device_vram_access v2"
    >
    > Am 13.04.20 um 20:20 schrieb Kent Russell:
    > > This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e.
    > > The original patch causes a RAS event and subsequent kernel
    hard-hang
    > > when running the KFDMemoryTest.PtraceAccessInvisibleVram on
    VG20 and
    > > Arcturus
    > >
    > > dmesg output at hang time:
    > > [drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected!
    > > amdgpu 0000:67:00.0: GPU reset begin!
    > > Evicting PASID 0x8000 queues
    > > Started evicting pasid 0x8000
    > > qcm fence wait loop timeout expired
    > > The cp might be in an unrecoverable state due to an unsuccessful
    > > queues preemption Failed to evict process queues Failed to suspend
    > > process 0x8000 Finished evicting pasid 0x8000 Started
    restoring pasid
    > > 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state
    may lost
    > > due to RAS ERREVENT_ATHUB_INTERRUPT
    > > amdgpu: [powerplay] Failed to send message 0x26, response 0x0
    > > amdgpu: [powerplay] Failed to set soft min gfxclk !
    > > amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
    > > amdgpu: [powerplay] Failed to send message 0x7, response 0x0
    > > amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable
    all smu
    > features!
    > > amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all
    smu features!
    > > amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM!
    > > [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend
    of IP
    > > block <powerplay> failed -5
    >
    > Do you have more information on what's going wrong here since
    this is a really
    > important patch for KFD debugging.
    >
    > >
    > > Signed-off-by: Kent Russell <[email protected]
    <mailto:[email protected]>>
    >
    > Reviewed-by: Christian König <[email protected]
    <mailto:[email protected]>>
    >
    > > ---
    > > drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26
    ----------------------
    > >   1 file changed, 26 deletions(-)
    > >
    > > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
    > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
    > > index cf5d6e585634..a3f997f84020 100644
    > > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
    > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
    > > @@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct
    > amdgpu_device *adev, loff_t pos,
    > >      uint32_t hi = ~0;
    > >      uint64_t last;
    > >
    > > -
    > > -#ifdef CONFIG_64BIT
    > > -   last = min(pos + size, adev->gmc.visible_vram_size);
    > > -   if (last > pos) {
    > > -           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
    > > -           size_t count = last - pos;
    > > -
    > > -           if (write) {
    > > -                   memcpy_toio(addr, buf, count);
    > > -                   mb();
    > > - amdgpu_asic_flush_hdp(adev, NULL);
    > > -           } else {
    > > - amdgpu_asic_invalidate_hdp(adev, NULL);
    > > -                   mb();
    > > -                   memcpy_fromio(buf, addr, count);
    > > -           }
    > > -
    > > -           if (count == size)
    > > -                   return;
    > > -
    > > -           pos += count;
    > > -           buf += count / 4;
    > > -           size -= count;
    > > -   }
    > > -#endif
Quoted 9 lines
    > > -
    > > spin_lock_irqsave(&adev->mmio_idx_lock, flags);
    > >      for (last = pos + size; pos < last; pos += 4) {
    > >              uint32_t tmp = pos >> 31;
    _______________________________________________
    amd-gfx mailing list
    [email protected] <mailto:[email protected]>
    https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

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

#9 lore raw

[AMD Public Use]

Hi Christian,

That could potentially be it. With additional testing, 2 of 3 Vega20 machines never hit error over BAR access with the PTRACE test. 3 of 3 machines (from the same pool) always hit error with CWSR.
To elaborate on the PTRACE test, we PEEK 2 DWORDs inside thunk allocated mapped memory and 2 DWORDS outside that boundary (it's only about 4MB to the boundary). Then we POKE to swap the DWORD positions across the boundary. The RAS event on the single failing machine happens on the out of boundary PEEK.

Felix mentioned we don't hit errors over general HDP access but that may not true. An Arcturus failure sys logs posted (which wasn't tested by me) shows someone launched rocm bandwidth test, hit a VM fault and a RAS event ensued during evictions (I can point the internal ticket or log snippet offline if interested). Whether the RAS event is BAR access triggered or the result of HW instability is beyond me since I don't have access to the machine.

Thanks,

Jon

From: Koenig, Christian <[email protected]>
Sent: Wednesday, April 15, 2020 4:11 AM
To: Kim, Jonathan <[email protected]>; Kuehling, Felix <[email protected]>; Deucher, Alexander <[email protected]>
Cc: Russell, Kent <[email protected]>; [email protected]
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

Hi Jon,

Also cwsr tests fail on Vega20 with or without the revert with the same RAS error.

That sounds like the system/setup has a more general problem.

Could it be that we are seeing RAS errors because there really is some hardware failure, but with the MM path we don't trigger a RAS interrupt?

Thanks,
Christian.

Am 14.04.20 um 22:30 schrieb Kim, Jonathan:

[AMD Official Use Only - Internal Distribution Only]

If we're passing the test on the revert, then the only thing that's different is we're not invalidating HDP and doing a copy to host anymore in amdgpu_device_vram_access since the function is still called in ttm access_memory with BAR.

Also cwsr tests fail on Vega20 with or without the revert with the same RAS error.

Thanks,

Jon

From: Kuehling, Felix <[email protected]><mailto:[email protected]>
Sent: Tuesday, April 14, 2020 2:32 PM
To: Kim, Jonathan <[email protected]><mailto:[email protected]>; Koenig, Christian <[email protected]><mailto:[email protected]>; Deucher, Alexander <[email protected]><mailto:[email protected]>
Cc: Russell, Kent <[email protected]><mailto:[email protected]>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

I wouldn't call it premature. Revert is a usual practice when there is a serious regression that isn't fully understood or root-caused. As far as I can tell, the problem has been reproduced on multiple systems, different GPUs, and clearly regressed to Christian's commit. I think that justifies reverting it for now.

I agree with Christian that a general HDP memory access problem causing RAS errors would potentially cause problems in other tests as well. For example common operations like GART table updates, and GPUVM page table updates and PCIe peer2peer accesses in ROCm applications use HDP. But we're not seeing obvious problems from those. So we need to understand what's special about this test. I asked questions to that effect on our other email thread.

Regards,
Felix
Am 2020-04-14 um 10:51 a.m. schrieb Kim, Jonathan:

[AMD Official Use Only - Internal Distribution Only]

I think it's premature to push this revert.

With more testing, I'm getting failures from different tests or sometimes none at all on my machine.

Kent, let's continue the discussion on the original thread.

Thanks,

Jon

From: Koenig, Christian <[email protected]><mailto:[email protected]>
Sent: Tuesday, April 14, 2020 10:47 AM
To: Deucher, Alexander <[email protected]><mailto:[email protected]>
Cc: Russell, Kent <[email protected]><mailto:[email protected]>; [email protected]<mailto:[email protected]>; Kuehling, Felix <[email protected]><mailto:[email protected]>; Kim, Jonathan <[email protected]><mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

That's exactly my concern as well.

This looks a bit like the test creates erroneous data somehow, but there doesn't seems to be a RAS check in the MM data path.

And now that we use the BAR path it goes up in flames.

I just don't see how we can create erroneous data in a test case?

Christian.

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex
________________________________
From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

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

maintainer
#10 lore raw
Quoted 2 lines

To elaborate on the PTRACE test, we PEEK 2 DWORDs inside thunk allocated mapped memory and 2 DWORDS outside that boundary (it’s only about 4MB to the boundary).  Then we POKE to swap the DWORD positions across the boundary.  The RAS event on the single failing machine happens on the out of boundary PEEK.

Well when you access outside of an allocated buffer I would expect that we never get as far as even touching the hardware because the kernel should block the access with an -EPERM or -EFAULT. So sounds like I'm not understanding something correctly here.

Apart from that I completely agree that we need to sort out any other RAS event first to make sure that the system is simply not failing randomly.

Regards,
Christian.

Am 15.04.20 um 11:49 schrieb Kim, Jonathan:

Quoted 880 lines

[AMD Public Use]

Hi Christian,

That could potentially be it.  With additional testing, 2 of 3 Vega20 machines never hit error over BAR access with the PTRACE test.  3 of 3 machines (from the same pool) always hit error with CWSR.

To elaborate on the PTRACE test, we PEEK 2 DWORDs inside thunk allocated mapped memory and 2 DWORDS outside that boundary (it’s only about 4MB to the boundary).  Then we POKE to swap the DWORD positions across the boundary.  The RAS event on the single failing machine happens on the out of boundary PEEK.

Felix mentioned we don’t hit errors over general HDP access but that may not true.  An Arcturus failure sys logs posted (which wasn’t tested by me) shows someone launched rocm bandwidth test, hit a VM fault and a RAS event ensued during evictions (I can point the internal ticket or log snippet offline if interested).  Whether the RAS event is BAR access triggered or the result of HW instability is beyond me since I don’t have access to the machine.

Thanks,

Jon

*From:*Koenig, Christian <[email protected]>
*Sent:* Wednesday, April 15, 2020 4:11 AM
*To:* Kim, Jonathan <[email protected]>; Kuehling, Felix <[email protected]>; Deucher, Alexander <[email protected]> *Cc:* Russell, Kent <[email protected]>; [email protected]
*Subject:* Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

Hi Jon,

Also cwsr tests fail on Vega20 with or without the revert with the same RAS error.

That sounds like the system/setup has a more general problem.

Could it be that we are seeing RAS errors because there really is some hardware failure, but with the MM path we don't trigger a RAS interrupt?

Thanks,
Christian.

Am 14.04.20 um 22:30 schrieb Kim, Jonathan:

[AMD Official Use Only - Internal Distribution Only]

If we’re passing the test on the revert, then the only thing that’s different is we’re not invalidating HDP and doing a copy to host anymore in amdgpu_device_vram_access since the function is still called in ttm access_memory with BAR.

Also cwsr tests fail on Vega20 with or without the revert with the same RAS error.

Thanks,

Jon

*From:* Kuehling, Felix <[email protected]> <mailto:[email protected]>
*Sent:* Tuesday, April 14, 2020 2:32 PM
*To:* Kim, Jonathan <[email protected]> <mailto:[email protected]>; Koenig, Christian <[email protected]> <mailto:[email protected]>; Deucher, Alexander <[email protected]> <mailto:[email protected]>
*Cc:* Russell, Kent <[email protected]> <mailto:[email protected]>; [email protected] <mailto:[email protected]>
*Subject:* Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

I wouldn't call it premature. Revert is a usual practice when there is a serious regression that isn't fully understood or root-caused. As far as I can tell, the problem has been reproduced on multiple systems, different GPUs, and clearly regressed to Christian's commit. I think that justifies reverting it for now.

I agree with Christian that a general HDP memory access problem causing RAS errors would potentially cause problems in other tests as well. For example common operations like GART table updates, and GPUVM page table updates and PCIe peer2peer accesses in ROCm applications use HDP. But we're not seeing obvious problems from those. So we need to understand what's special about this test. I asked questions to that effect on our other email thread.

Regards,
  Felix

Am 2020-04-14 um 10:51 a.m. schrieb Kim, Jonathan:

[AMD Official Use Only - Internal Distribution Only]

I think it’s premature to push this revert.

With more testing, I’m getting failures from different tests or sometimes none at all on my machine.

Kent, let’s continue the discussion on the original thread.

Thanks,

Jon

*From:* Koenig, Christian <[email protected]> <mailto:[email protected]>
*Sent:* Tuesday, April 14, 2020 10:47 AM
*To:* Deucher, Alexander <[email protected]> <mailto:[email protected]>
*Cc:* Russell, Kent <[email protected]> <mailto:[email protected]>; [email protected] <mailto:[email protected]>; Kuehling, Felix <[email protected]> <mailto:[email protected]>; Kim, Jonathan <[email protected]> <mailto:[email protected]> *Subject:* Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

That's exactly my concern as well.

This looks a bit like the test creates erroneous data somehow, but there doesn't seems to be a RAS check in the MM data path.

And now that we use the BAR path it goes up in flames.

I just don't see how we can create erroneous data in a test case?

Christian.

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected] <mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

------------------------------------------------------------------------

*From:* amd-gfx <[email protected] <mailto:[email protected]>> on behalf of Russell, Kent <[email protected] <mailto:[email protected]>>
*Sent:* Tuesday, April 14, 2020 10:19 AM
*To:* Koenig, Christian <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]> <[email protected] <mailto:[email protected]>>
*Cc:* Kuehling, Felix <[email protected] <mailto:[email protected]>>; Kim, Jonathan <[email protected] <mailto:[email protected]>> *Subject:* RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

 Kent

> -----Original Message-----
> From: Christian König <[email protected] <mailto:[email protected]>>
> Sent: Tuesday, April 14, 2020 9:52 AM
> To: Russell, Kent <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]>
> Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
> amdgpu_device_vram_access v2"
>
> Am 13.04.20 um 20:20 schrieb Kent Russell:
> > This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e.
> > The original patch causes a RAS event and subsequent kernel hard-hang
> > when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and > > Arcturus
> >
> > dmesg output at hang time:
> > [drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! > > amdgpu 0000:67:00.0: GPU reset begin!
> > Evicting PASID 0x8000 queues
> > Started evicting pasid 0x8000
> > qcm fence wait loop timeout expired
> > The cp might be in an unrecoverable state due to an unsuccessful
> > queues preemption Failed to evict process queues Failed to suspend
> > process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid
> > 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost
> > due to RAS ERREVENT_ATHUB_INTERRUPT
> > amdgpu: [powerplay] Failed to send message 0x26, response 0x0
> > amdgpu: [powerplay] Failed to set soft min gfxclk ! > > amdgpu: [powerplay] Failed to upload DPM Bootup Levels! > > amdgpu: [powerplay] Failed to send message 0x7, response 0x0
> > amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu
> features!
> > amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
> > amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! > > [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP
> > block <powerplay> failed -5
>
> Do you have more information on what's going wrong here since this is a really
> important patch for KFD debugging.
>
> >
> > Signed-off-by: Kent Russell <[email protected] <mailto:[email protected]>>
>
> Reviewed-by: Christian König <[email protected] <mailto:[email protected]>>
>
> > ---
> > drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26
----------------------
> >   1 file changed, 26 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> > index cf5d6e585634..a3f997f84020 100644
> > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > @@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct > amdgpu_device *adev, loff_t pos,
> >      uint32_t hi = ~0;
> >      uint64_t last;
> >
> > -
> > -#ifdef CONFIG_64BIT
> > -   last = min(pos + size, adev->gmc.visible_vram_size); > > -   if (last > pos) {
> > -           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
> > -           size_t count = last - pos;
> > -
> > -           if (write) {
> > - memcpy_toio(addr, buf, count);
> > -                   mb();
> > - amdgpu_asic_flush_hdp(adev, NULL);
> > -           } else {
> > - amdgpu_asic_invalidate_hdp(adev, NULL);
> > -                   mb();
> > - memcpy_fromio(buf, addr, count);
> > -           }
> > -
> > -           if (count == size)
> > - return;
> > -
> > -           pos += count;
> > -           buf += count / 4;
> > -           size -= count;
> > -   }
> > -#endif
> > -
> > spin_lock_irqsave(&adev->mmio_idx_lock, flags); > >      for (last = pos + size; pos < last; pos += 4) { > >              uint32_t tmp = pos >> 31;
_______________________________________________
amd-gfx mailing list
[email protected] <mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected] <mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

------------------------------------------------------------------------

*From:* amd-gfx <[email protected] <mailto:[email protected]>> on behalf of Russell, Kent <[email protected] <mailto:[email protected]>>
*Sent:* Tuesday, April 14, 2020 10:19 AM
*To:* Koenig, Christian <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]> <[email protected] <mailto:[email protected]>>
*Cc:* Kuehling, Felix <[email protected] <mailto:[email protected]>>; Kim, Jonathan <[email protected] <mailto:[email protected]>> *Subject:* RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

 Kent

> -----Original Message-----
> From: Christian König <[email protected] <mailto:[email protected]>>
> Sent: Tuesday, April 14, 2020 9:52 AM
> To: Russell, Kent <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]>
> Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
> amdgpu_device_vram_access v2"
>
> Am 13.04.20 um 20:20 schrieb Kent Russell:
> > This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e.
> > The original patch causes a RAS event and subsequent kernel hard-hang
> > when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and > > Arcturus
> >
> > dmesg output at hang time:
> > [drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! > > amdgpu 0000:67:00.0: GPU reset begin!
> > Evicting PASID 0x8000 queues
> > Started evicting pasid 0x8000
> > qcm fence wait loop timeout expired
> > The cp might be in an unrecoverable state due to an unsuccessful
> > queues preemption Failed to evict process queues Failed to suspend
> > process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid
> > 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost
> > due to RAS ERREVENT_ATHUB_INTERRUPT
> > amdgpu: [powerplay] Failed to send message 0x26, response 0x0
> > amdgpu: [powerplay] Failed to set soft min gfxclk ! > > amdgpu: [powerplay] Failed to upload DPM Bootup Levels! > > amdgpu: [powerplay] Failed to send message 0x7, response 0x0
> > amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu
> features!
> > amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
> > amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! > > [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP
> > block <powerplay> failed -5
>
> Do you have more information on what's going wrong here since this is a really
> important patch for KFD debugging.
>
> >
> > Signed-off-by: Kent Russell <[email protected] <mailto:[email protected]>>
>
> Reviewed-by: Christian König <[email protected] <mailto:[email protected]>>
>
> > ---
> > drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26
----------------------
> >   1 file changed, 26 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> > index cf5d6e585634..a3f997f84020 100644
> > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > @@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct > amdgpu_device *adev, loff_t pos,
> >      uint32_t hi = ~0;
> >      uint64_t last;
> >
> > -
> > -#ifdef CONFIG_64BIT
> > -   last = min(pos + size, adev->gmc.visible_vram_size); > > -   if (last > pos) {
> > -           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
> > -           size_t count = last - pos;
> > -
> > -           if (write) {
> > - memcpy_toio(addr, buf, count);
> > -                   mb();
> > - amdgpu_asic_flush_hdp(adev, NULL);
> > -           } else {
> > - amdgpu_asic_invalidate_hdp(adev, NULL);
> > -                   mb();
> > - memcpy_fromio(buf, addr, count);
> > -           }
> > -
> > -           if (count == size)
> > -                   return;
> > -
> > -           pos += count;
> > -           buf += count / 4;
> > -           size -= count;
> > -   }
> > -#endif
> > -
> > spin_lock_irqsave(&adev->mmio_idx_lock, flags); > >      for (last = pos + size; pos < last; pos += 4) { > >              uint32_t tmp = pos >> 31;
_______________________________________________
amd-gfx mailing list
[email protected] <mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected] <mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

------------------------------------------------------------------------

*From:* amd-gfx <[email protected] <mailto:[email protected]>> on behalf of Russell, Kent <[email protected] <mailto:[email protected]>>
*Sent:* Tuesday, April 14, 2020 10:19 AM
*To:* Koenig, Christian <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]> <[email protected] <mailto:[email protected]>>
*Cc:* Kuehling, Felix <[email protected] <mailto:[email protected]>>; Kim, Jonathan <[email protected] <mailto:[email protected]>> *Subject:* RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

 Kent

> -----Original Message-----
> From: Christian König <[email protected] <mailto:[email protected]>>
> Sent: Tuesday, April 14, 2020 9:52 AM
> To: Russell, Kent <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]>
> Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
> amdgpu_device_vram_access v2"
>
> Am 13.04.20 um 20:20 schrieb Kent Russell:
> > This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e.
> > The original patch causes a RAS event and subsequent kernel hard-hang
> > when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and > > Arcturus
> >
> > dmesg output at hang time:
> > [drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! > > amdgpu 0000:67:00.0: GPU reset begin!
> > Evicting PASID 0x8000 queues
> > Started evicting pasid 0x8000
> > qcm fence wait loop timeout expired
> > The cp might be in an unrecoverable state due to an unsuccessful
> > queues preemption Failed to evict process queues Failed to suspend
> > process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid
> > 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost
> > due to RAS ERREVENT_ATHUB_INTERRUPT
> > amdgpu: [powerplay] Failed to send message 0x26, response 0x0
> > amdgpu: [powerplay] Failed to set soft min gfxclk ! > > amdgpu: [powerplay] Failed to upload DPM Bootup Levels! > > amdgpu: [powerplay] Failed to send message 0x7, response 0x0
> > amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu
> features!
> > amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
> > amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! > > [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP
> > block <powerplay> failed -5
>
> Do you have more information on what's going wrong here since this is a really
> important patch for KFD debugging.
>
> >
> > Signed-off-by: Kent Russell <[email protected] <mailto:[email protected]>>
>
> Reviewed-by: Christian König <[email protected] <mailto:[email protected]>>
>
> > ---
> > drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26
----------------------
> >   1 file changed, 26 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> > index cf5d6e585634..a3f997f84020 100644
> > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > @@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct > amdgpu_device *adev, loff_t pos,
> >      uint32_t hi = ~0;
> >      uint64_t last;
> >
> > -
> > -#ifdef CONFIG_64BIT
> > -   last = min(pos + size, adev->gmc.visible_vram_size); > > -   if (last > pos) {
> > -           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
> > -           size_t count = last - pos;
> > -
> > -           if (write) {
> > - memcpy_toio(addr, buf, count);
> > -                   mb();
> > - amdgpu_asic_flush_hdp(adev, NULL);
> > -           } else {
> > - amdgpu_asic_invalidate_hdp(adev, NULL);
> > -                   mb();
> > - memcpy_fromio(buf, addr, count);
> > -           }
> > -
> > -           if (count == size)
> > -                   return;
> > -
> > -           pos += count;
> > -           buf += count / 4;
> > -           size -= count;
> > -   }
> > -#endif
> > -
> > spin_lock_irqsave(&adev->mmio_idx_lock, flags); > >      for (last = pos + size; pos < last; pos += 4) { > >              uint32_t tmp = pos >> 31;
_______________________________________________
amd-gfx mailing list
[email protected] <mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected] <mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

------------------------------------------------------------------------

*From:* amd-gfx <[email protected] <mailto:[email protected]>> on behalf of Russell, Kent <[email protected] <mailto:[email protected]>>
*Sent:* Tuesday, April 14, 2020 10:19 AM
*To:* Koenig, Christian <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]> <[email protected] <mailto:[email protected]>>
*Cc:* Kuehling, Felix <[email protected] <mailto:[email protected]>>; Kim, Jonathan <[email protected] <mailto:[email protected]>> *Subject:* RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

 Kent

> -----Original Message-----
> From: Christian König <[email protected] <mailto:[email protected]>>
> Sent: Tuesday, April 14, 2020 9:52 AM
> To: Russell, Kent <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]>
> Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
> amdgpu_device_vram_access v2"
>
> Am 13.04.20 um 20:20 schrieb Kent Russell:
> > This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e.
> > The original patch causes a RAS event and subsequent kernel hard-hang
> > when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and > > Arcturus
> >
> > dmesg output at hang time:
> > [drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! > > amdgpu 0000:67:00.0: GPU reset begin!
> > Evicting PASID 0x8000 queues
> > Started evicting pasid 0x8000
> > qcm fence wait loop timeout expired
> > The cp might be in an unrecoverable state due to an unsuccessful
> > queues preemption Failed to evict process queues Failed to suspend
> > process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid
> > 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost
> > due to RAS ERREVENT_ATHUB_INTERRUPT
> > amdgpu: [powerplay] Failed to send message 0x26, response 0x0
> > amdgpu: [powerplay] Failed to set soft min gfxclk ! > > amdgpu: [powerplay] Failed to upload DPM Bootup Levels! > > amdgpu: [powerplay] Failed to send message 0x7, response 0x0
> > amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu
> features!
> > amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
> > amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! > > [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP
> > block <powerplay> failed -5
>
> Do you have more information on what's going wrong here since this is a really
> important patch for KFD debugging.
>
> >
> > Signed-off-by: Kent Russell <[email protected] <mailto:[email protected]>>
>
> Reviewed-by: Christian König <[email protected] <mailto:[email protected]>>
>
> > ---
> > drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26
----------------------
> >   1 file changed, 26 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> > index cf5d6e585634..a3f997f84020 100644
> > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > @@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct > amdgpu_device *adev, loff_t pos,
> >      uint32_t hi = ~0;
> >      uint64_t last;
> >
> > -
> > -#ifdef CONFIG_64BIT
> > -   last = min(pos + size, adev->gmc.visible_vram_size); > > -   if (last > pos) {
> > -           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
> > -           size_t count = last - pos;
> > -
> > -           if (write) {
> > - memcpy_toio(addr, buf, count);
> > -                   mb();
> > - amdgpu_asic_flush_hdp(adev, NULL);
> > -           } else {
> > - amdgpu_asic_invalidate_hdp(adev, NULL);
> > -                   mb();
> > - memcpy_fromio(buf, addr, count);
> > -           }
> > -
> > -           if (count == size)
> > -                   return;
> > -
> > -           pos += count;
> > -           buf += count / 4;
> > -           size -= count;
> > -   }
> > -#endif
> > -
> > spin_lock_irqsave(&adev->mmio_idx_lock, flags); > >      for (last = pos + size; pos < last; pos += 4) { > >              uint32_t tmp = pos >> 31;
_______________________________________________
amd-gfx mailing list
[email protected] <mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected] <mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

------------------------------------------------------------------------

*From:* amd-gfx <[email protected] <mailto:[email protected]>> on behalf of Russell, Kent <[email protected] <mailto:[email protected]>> *Sent:* Tuesday, April 14, 2020 10:19 AM
*To:* Koenig, Christian <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]> <[email protected] <mailto:[email protected]>>
*Cc:* Kuehling, Felix <[email protected] <mailto:[email protected]>>; Kim, Jonathan <[email protected] <mailto:[email protected]>> *Subject:* RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

 Kent

> -----Original Message-----
> From: Christian König <[email protected] <mailto:[email protected]>>
> Sent: Tuesday, April 14, 2020 9:52 AM
> To: Russell, Kent <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]>
> Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
> amdgpu_device_vram_access v2"
>
> Am 13.04.20 um 20:20 schrieb Kent Russell:
> > This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. > > The original patch causes a RAS event and subsequent kernel hard-hang
> > when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and
> > Arcturus
> >
> > dmesg output at hang time:
> > [drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! > > amdgpu 0000:67:00.0: GPU reset begin!
> > Evicting PASID 0x8000 queues
> > Started evicting pasid 0x8000
> > qcm fence wait loop timeout expired
> > The cp might be in an unrecoverable state due to an unsuccessful
> > queues preemption Failed to evict process queues Failed to suspend
> > process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid
> > 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost
> > due to RAS ERREVENT_ATHUB_INTERRUPT
> > amdgpu: [powerplay] Failed to send message 0x26, response 0x0 > > amdgpu: [powerplay] Failed to set soft min gfxclk ! > > amdgpu: [powerplay] Failed to upload DPM Bootup Levels! > > amdgpu: [powerplay] Failed to send message 0x7, response 0x0 > > amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu
> features!
> > amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
> > amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! > > [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP
> > block <powerplay> failed -5
>
> Do you have more information on what's going wrong here since this is a really
> important patch for KFD debugging.
>
> >
> > Signed-off-by: Kent Russell <[email protected] <mailto:[email protected]>>
>
> Reviewed-by: Christian König <[email protected] <mailto:[email protected]>>
>
> > ---
> > drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26
----------------------
> >   1 file changed, 26 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> > index cf5d6e585634..a3f997f84020 100644
> > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > @@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct > amdgpu_device *adev, loff_t pos,
> >      uint32_t hi = ~0;
> >      uint64_t last;
> >
> > -
> > -#ifdef CONFIG_64BIT
> > -   last = min(pos + size, adev->gmc.visible_vram_size); > > -   if (last > pos) {
> > -           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
> > -           size_t count = last - pos;
> > -
> > -           if (write) {
> > -                   memcpy_toio(addr, buf, count); > > -                   mb();
> > - amdgpu_asic_flush_hdp(adev, NULL);
> > -           } else {
> > - amdgpu_asic_invalidate_hdp(adev, NULL);
> > -                   mb();
> > -                   memcpy_fromio(buf, addr, count); > > -           }
> > -
> > -           if (count == size)
> > -                   return;
> > -
> > -           pos += count;
> > -           buf += count / 4;
> > -           size -= count;
> > -   }
> > -#endif
> > -

Quoted 9 lines

> > spin_lock_irqsave(&adev->mmio_idx_lock, flags); > >      for (last = pos + size; pos < last; pos += 4) { > >              uint32_t tmp = pos >> 31;
_______________________________________________
amd-gfx mailing list [email protected] <mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

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

#11 lore raw

[AMD Official Use Only - Internal Distribution Only]

The test does not access outside of the allocated memory. But it deliberately crosses a boundary where memory can be allocated non-contiguously. This is meant to catch problems where the access function doesn't handle non-contiguous VRAM allocations correctly. However, the way that VRAM allocation has been optimized, I expect that most allocations are contiguous nowadays. However, the more interesting aspect of the test is, that it performs misaligned memory accesses. The MMIO method of accessing VRAM explicitly handles misaligned accesses and breaks them down into dword aligned accesses with proper masking and shifting.

Could the unaligned nature of the memory access have something to do with hitting RAS errors? That's something unique to this test that we wouldn't see on a normal page table update or memory eviction.

Regards,
Felix

________________________________
From: Koenig, Christian <[email protected]>
Sent: Wednesday, April 15, 2020 6:58 AM
To: Kim, Jonathan <[email protected]>; Kuehling, Felix <[email protected]>; Deucher, Alexander <[email protected]>
Cc: Russell, Kent <[email protected]>; [email protected] <[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

To elaborate on the PTRACE test, we PEEK 2 DWORDs inside thunk allocated mapped memory and 2 DWORDS outside that boundary (it’s only about 4MB to the boundary). Then we POKE to swap the DWORD positions across the boundary. The RAS event on the single failing machine happens on the out of boundary PEEK.

Well when you access outside of an allocated buffer I would expect that we never get as far as even touching the hardware because the kernel should block the access with an -EPERM or -EFAULT. So sounds like I'm not understanding something correctly here.

Apart from that I completely agree that we need to sort out any other RAS event first to make sure that the system is simply not failing randomly.

Regards,
Christian.

Am 15.04.20 um 11:49 schrieb Kim, Jonathan:

[AMD Public Use]

Hi Christian,

That could potentially be it. With additional testing, 2 of 3 Vega20 machines never hit error over BAR access with the PTRACE test. 3 of 3 machines (from the same pool) always hit error with CWSR.

To elaborate on the PTRACE test, we PEEK 2 DWORDs inside thunk allocated mapped memory and 2 DWORDS outside that boundary (it’s only about 4MB to the boundary). Then we POKE to swap the DWORD positions across the boundary. The RAS event on the single failing machine happens on the out of boundary PEEK.

Felix mentioned we don’t hit errors over general HDP access but that may not true. An Arcturus failure sys logs posted (which wasn’t tested by me) shows someone launched rocm bandwidth test, hit a VM fault and a RAS event ensued during evictions (I can point the internal ticket or log snippet offline if interested). Whether the RAS event is BAR access triggered or the result of HW instability is beyond me since I don’t have access to the machine.

Thanks,

Jon

From: Koenig, Christian <[email protected]><mailto:[email protected]>
Sent: Wednesday, April 15, 2020 4:11 AM
To: Kim, Jonathan <[email protected]><mailto:[email protected]>; Kuehling, Felix <[email protected]><mailto:[email protected]>; Deucher, Alexander <[email protected]><mailto:[email protected]>
Cc: Russell, Kent <[email protected]><mailto:[email protected]>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

Hi Jon,

Also cwsr tests fail on Vega20 with or without the revert with the same RAS error.

That sounds like the system/setup has a more general problem.

Could it be that we are seeing RAS errors because there really is some hardware failure, but with the MM path we don't trigger a RAS interrupt?

Thanks,
Christian.

Am 14.04.20 um 22:30 schrieb Kim, Jonathan:

[AMD Official Use Only - Internal Distribution Only]

If we’re passing the test on the revert, then the only thing that’s different is we’re not invalidating HDP and doing a copy to host anymore in amdgpu_device_vram_access since the function is still called in ttm access_memory with BAR.

Also cwsr tests fail on Vega20 with or without the revert with the same RAS error.

Thanks,

Jon

From: Kuehling, Felix <[email protected]><mailto:[email protected]>
Sent: Tuesday, April 14, 2020 2:32 PM
To: Kim, Jonathan <[email protected]><mailto:[email protected]>; Koenig, Christian <[email protected]><mailto:[email protected]>; Deucher, Alexander <[email protected]><mailto:[email protected]>
Cc: Russell, Kent <[email protected]><mailto:[email protected]>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

I wouldn't call it premature. Revert is a usual practice when there is a serious regression that isn't fully understood or root-caused. As far as I can tell, the problem has been reproduced on multiple systems, different GPUs, and clearly regressed to Christian's commit. I think that justifies reverting it for now.

I agree with Christian that a general HDP memory access problem causing RAS errors would potentially cause problems in other tests as well. For example common operations like GART table updates, and GPUVM page table updates and PCIe peer2peer accesses in ROCm applications use HDP. But we're not seeing obvious problems from those. So we need to understand what's special about this test. I asked questions to that effect on our other email thread.

Regards,
Felix

Am 2020-04-14 um 10:51 a.m. schrieb Kim, Jonathan:

[AMD Official Use Only - Internal Distribution Only]

I think it’s premature to push this revert.

With more testing, I’m getting failures from different tests or sometimes none at all on my machine.

Kent, let’s continue the discussion on the original thread.

Thanks,

Jon

From: Koenig, Christian <[email protected]><mailto:[email protected]>
Sent: Tuesday, April 14, 2020 10:47 AM
To: Deucher, Alexander <[email protected]><mailto:[email protected]>
Cc: Russell, Kent <[email protected]><mailto:[email protected]>; [email protected]<mailto:[email protected]>; Kuehling, Felix <[email protected]><mailto:[email protected]>; Kim, Jonathan <[email protected]><mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

That's exactly my concern as well.

This looks a bit like the test creates erroneous data somehow, but there doesn't seems to be a RAS check in the MM data path.

And now that we use the BAR path it goes up in flames.

I just don't see how we can create erroneous data in a test case?

Christian.

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

________________________________

From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

________________________________

From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

________________________________

From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

________________________________

From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

________________________________

From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

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

#12 lore raw

[AMD Official Use Only - Internal Distribution Only]

Hi Felix,

You're probably right.

Passing Vega20 system:

[   56.683273] amdgpu: [vram dbg] addr         3e7ffff8, val         deadbeef
[   56.683349] amdgpu: [vram dbg] addr         3efed000, val         cafebabe <- potential misalign access

Failing Vega20 system:

[Apr16 12:00] amdgpu: [vram dbg] addr         be7ffff8, val         deadbeef
[  +0.000082] amdgpu: [vram dbg] addr         befed000, val         ffffffff <- potential misalign access

Thanks,

Jon

From: Kuehling, Felix <[email protected]>
Sent: Wednesday, April 15, 2020 11:02 AM
To: Koenig, Christian <[email protected]>; Kim, Jonathan <[email protected]>; Deucher, Alexander <[email protected]>
Cc: Russell, Kent <[email protected]>; [email protected]
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

The test does not access outside of the allocated memory. But it deliberately crosses a boundary where memory can be allocated non-contiguously. This is meant to catch problems where the access function doesn't handle non-contiguous VRAM allocations correctly. However, the way that VRAM allocation has been optimized, I expect that most allocations are contiguous nowadays. However, the more interesting aspect of the test is, that it performs misaligned memory accesses. The MMIO method of accessing VRAM explicitly handles misaligned accesses and breaks them down into dword aligned accesses with proper masking and shifting.

Could the unaligned nature of the memory access have something to do with hitting RAS errors? That's something unique to this test that we wouldn't see on a normal page table update or memory eviction.

Regards,
Felix

________________________________
From: Koenig, Christian <[email protected]<mailto:[email protected]>>
Sent: Wednesday, April 15, 2020 6:58 AM
To: Kim, Jonathan <[email protected]<mailto:[email protected]>>; Kuehling, Felix <[email protected]<mailto:[email protected]>>; Deucher, Alexander <[email protected]<mailto:[email protected]>>
Cc: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

To elaborate on the PTRACE test, we PEEK 2 DWORDs inside thunk allocated mapped memory and 2 DWORDS outside that boundary (it's only about 4MB to the boundary). Then we POKE to swap the DWORD positions across the boundary. The RAS event on the single failing machine happens on the out of boundary PEEK.

Well when you access outside of an allocated buffer I would expect that we never get as far as even touching the hardware because the kernel should block the access with an -EPERM or -EFAULT. So sounds like I'm not understanding something correctly here.

Apart from that I completely agree that we need to sort out any other RAS event first to make sure that the system is simply not failing randomly.

Regards,
Christian.

Am 15.04.20 um 11:49 schrieb Kim, Jonathan:

[AMD Public Use]

Hi Christian,

That could potentially be it. With additional testing, 2 of 3 Vega20 machines never hit error over BAR access with the PTRACE test. 3 of 3 machines (from the same pool) always hit error with CWSR.

To elaborate on the PTRACE test, we PEEK 2 DWORDs inside thunk allocated mapped memory and 2 DWORDS outside that boundary (it's only about 4MB to the boundary). Then we POKE to swap the DWORD positions across the boundary. The RAS event on the single failing machine happens on the out of boundary PEEK.

Felix mentioned we don't hit errors over general HDP access but that may not true. An Arcturus failure sys logs posted (which wasn't tested by me) shows someone launched rocm bandwidth test, hit a VM fault and a RAS event ensued during evictions (I can point the internal ticket or log snippet offline if interested). Whether the RAS event is BAR access triggered or the result of HW instability is beyond me since I don't have access to the machine.

Thanks,

Jon

From: Koenig, Christian <[email protected]><mailto:[email protected]>
Sent: Wednesday, April 15, 2020 4:11 AM
To: Kim, Jonathan <[email protected]><mailto:[email protected]>; Kuehling, Felix <[email protected]><mailto:[email protected]>; Deucher, Alexander <[email protected]><mailto:[email protected]>
Cc: Russell, Kent <[email protected]><mailto:[email protected]>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

Hi Jon,

Also cwsr tests fail on Vega20 with or without the revert with the same RAS error.

That sounds like the system/setup has a more general problem.

Could it be that we are seeing RAS errors because there really is some hardware failure, but with the MM path we don't trigger a RAS interrupt?

Thanks,
Christian.

Am 14.04.20 um 22:30 schrieb Kim, Jonathan:

[AMD Official Use Only - Internal Distribution Only]

If we're passing the test on the revert, then the only thing that's different is we're not invalidating HDP and doing a copy to host anymore in amdgpu_device_vram_access since the function is still called in ttm access_memory with BAR.

Also cwsr tests fail on Vega20 with or without the revert with the same RAS error.

Thanks,

Jon

From: Kuehling, Felix <[email protected]><mailto:[email protected]>
Sent: Tuesday, April 14, 2020 2:32 PM
To: Kim, Jonathan <[email protected]><mailto:[email protected]>; Koenig, Christian <[email protected]><mailto:[email protected]>; Deucher, Alexander <[email protected]><mailto:[email protected]>
Cc: Russell, Kent <[email protected]><mailto:[email protected]>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

I wouldn't call it premature. Revert is a usual practice when there is a serious regression that isn't fully understood or root-caused. As far as I can tell, the problem has been reproduced on multiple systems, different GPUs, and clearly regressed to Christian's commit. I think that justifies reverting it for now.

I agree with Christian that a general HDP memory access problem causing RAS errors would potentially cause problems in other tests as well. For example common operations like GART table updates, and GPUVM page table updates and PCIe peer2peer accesses in ROCm applications use HDP. But we're not seeing obvious problems from those. So we need to understand what's special about this test. I asked questions to that effect on our other email thread.

Regards,
Felix

Am 2020-04-14 um 10:51 a.m. schrieb Kim, Jonathan:

[AMD Official Use Only - Internal Distribution Only]

I think it's premature to push this revert.

With more testing, I'm getting failures from different tests or sometimes none at all on my machine.

Kent, let's continue the discussion on the original thread.

Thanks,

Jon

From: Koenig, Christian <[email protected]><mailto:[email protected]>
Sent: Tuesday, April 14, 2020 10:47 AM
To: Deucher, Alexander <[email protected]><mailto:[email protected]>
Cc: Russell, Kent <[email protected]><mailto:[email protected]>; [email protected]<mailto:[email protected]>; Kuehling, Felix <[email protected]><mailto:[email protected]>; Kim, Jonathan <[email protected]><mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

That's exactly my concern as well.

This looks a bit like the test creates erroneous data somehow, but there doesn't seems to be a RAS check in the MM data path.

And now that we use the BAR path it goes up in flames.

I just don't see how we can create erroneous data in a test case?

Christian.

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

________________________________

From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

________________________________

From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

________________________________

From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

________________________________

From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected]<mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

________________________________

From: amd-gfx <[email protected]<mailto:[email protected]>> on behalf of Russell, Kent <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 10:19 AM
To: Koenig, Christian <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]> <[email protected]<mailto:[email protected]>>
Cc: Kuehling, Felix <[email protected]<mailto:[email protected]>>; Kim, Jonathan <[email protected]<mailto:[email protected]>>
Subject: RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

Kent

Quoted 78 lines

-----Original Message-----
From: Christian König <[email protected]<mailto:[email protected]>>
Sent: Tuesday, April 14, 2020 9:52 AM
To: Russell, Kent <[email protected]<mailto:[email protected]>>; [email protected]<mailto:[email protected]>
Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
amdgpu_device_vram_access v2"

Am 13.04.20 um 20:20 schrieb Kent Russell:

Quoted 5 lines

This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e. The original patch causes a RAS event and subsequent kernel hard-hang when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and Arcturus

Quoted 16 lines

dmesg output at hang time:
[drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! amdgpu 0000:67:00.0: GPU reset begin!
Evicting PASID 0x8000 queues
Started evicting pasid 0x8000
qcm fence wait loop timeout expired
The cp might be in an unrecoverable state due to an unsuccessful queues preemption Failed to evict process queues Failed to suspend process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost due to RAS ERREVENT_ATHUB_INTERRUPT
amdgpu: [powerplay] Failed to send message 0x26, response 0x0
amdgpu: [powerplay] Failed to set soft min gfxclk !
amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
amdgpu: [powerplay] Failed to send message 0x7, response 0x0
amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu

features!

Quoted 4 lines

amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP block <powerplay> failed -5

Do you have more information on what's going wrong here since this is a really important patch for KFD debugging.

Quoted 2 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct

Quoted 10 lines
---
  drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26 ----------------------
  1 file changed, 26 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -0
index cf5d6e585634..a3f997f84020 100644
@@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct

amdgpu_device *adev, loff_t pos,

Quoted 24 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -21
     uint32_t hi = ~0;
     uint64_t last;
 
-
-#ifdef CONFIG_64BIT
-   last = min(pos + size, adev->gmc.visible_vram_size);
-   if (last > pos) {
-           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
-           size_t count = last - pos;
-
-           if (write) {
-                   memcpy_toio(addr, buf, count);
-                   mb();
-                   amdgpu_asic_flush_hdp(adev, NULL);
-           } else {
-                   amdgpu_asic_invalidate_hdp(adev, NULL);
-                   mb();
-                   memcpy_fromio(buf, addr, count);
-           }
-
-           if (count == size)
-                   return;
-
-           pos += count;
Quoted 8 lines

drivers/gpu/drm/amd/amdgpu/amdgpu_device.c void amdgpu_device_vram_access(struct · line 259

Quoted 8 lines
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +0 -5
-           buf += count / 4;
-           size -= count;
-   }
-#endif
-
     spin_lock_irqsave(&adev->mmio_idx_lock, flags);
     for (last = pos + size; pos < last; pos += 4) {
             uint32_t tmp = pos >> 31;

_______________________________________________
amd-gfx mailing list [email protected]<mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

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

maintainer
#13 lore raw

Why do you think that 3efed000 and befed000 are misaligned addresses?

And see amdgpu_ttm_access_memory(), misaligned accesses are always routed to the MM path.

Regards,
Christian.

Am 16.04.20 um 18:08 schrieb Kim, Jonathan:

Quoted 1,013 lines

[AMD Official Use Only - Internal Distribution Only]

Hi Felix,

You’re probably right.

Passing Vega20 system:

[   56.683273] amdgpu: [vram dbg] addr         3e7ffff8, val         deadbeef

[   56.683349] amdgpu: [vram dbg] addr         3efed000, val         cafebabe <- potential misalign access

Failing Vega20 system:

[Apr16 12:00] amdgpu: [vram dbg] addr         be7ffff8, val         deadbeef

[  +0.000082] amdgpu: [vram dbg] addr         befed000, val         ffffffff <- potential misalign access

Thanks,

Jon

*From:* Kuehling, Felix <[email protected]>
*Sent:* Wednesday, April 15, 2020 11:02 AM
*To:* Koenig, Christian <[email protected]>; Kim, Jonathan <[email protected]>; Deucher, Alexander <[email protected]> *Cc:* Russell, Kent <[email protected]>; [email protected]
*Subject:* Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

The test does not access outside of the allocated memory. But it deliberately crosses a boundary where memory can be allocated non-contiguously. This is meant to catch problems where the access function doesn't handle non-contiguous VRAM allocations correctly. However, the way that VRAM allocation has been optimized, I expect that most allocations are contiguous nowadays. However, the more interesting aspect of the test is, that it performs misaligned memory accesses. The MMIO method of accessing VRAM explicitly handles misaligned accesses and breaks them down into dword aligned accesses with proper masking and shifting.

Could the unaligned nature of the memory access have something to do with hitting RAS errors? That's something unique to this test that we wouldn't see on a normal page table update or memory eviction.

Regards,
  Felix

------------------------------------------------------------------------

*From:*Koenig, Christian <[email protected] <mailto:[email protected]>>
*Sent:* Wednesday, April 15, 2020 6:58 AM
*To:* Kim, Jonathan <[email protected] <mailto:[email protected]>>; Kuehling, Felix <[email protected] <mailto:[email protected]>>; Deucher, Alexander <[email protected] <mailto:[email protected]>> *Cc:* Russell, Kent <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]> <[email protected] <mailto:[email protected]>> *Subject:* Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

To elaborate on the PTRACE test, we PEEK 2 DWORDs inside thunk allocated mapped memory and 2 DWORDS outside that boundary (it’s only about 4MB to the boundary). Then we POKE to swap the DWORD positions across the boundary.  The RAS event on the single failing machine happens on the out of boundary PEEK.

Well when you access outside of an allocated buffer I would expect that we never get as far as even touching the hardware because the kernel should block the access with an -EPERM or -EFAULT. So sounds like I'm not understanding something correctly here.

Apart from that I completely agree that we need to sort out any other RAS event first to make sure that the system is simply not failing randomly.

Regards,
Christian.

Am 15.04.20 um 11:49 schrieb Kim, Jonathan:

[AMD Public Use]

Hi Christian,

That could potentially be it.  With additional testing, 2 of 3 Vega20 machines never hit error over BAR access with the PTRACE test.  3 of 3 machines (from the same pool) always hit error with CWSR.

To elaborate on the PTRACE test, we PEEK 2 DWORDs inside thunk allocated mapped memory and 2 DWORDS outside that boundary (it’s only about 4MB to the boundary). Then we POKE to swap the DWORD positions across the boundary.  The RAS event on the single failing machine happens on the out of boundary PEEK.

Felix mentioned we don’t hit errors over general HDP access but that may not true.  An Arcturus failure sys logs posted (which wasn’t tested by me) shows someone launched rocm bandwidth test, hit a VM fault and a RAS event ensued during evictions (I can point the internal ticket or log snippet offline if interested). Whether the RAS event is BAR access triggered or the result of HW instability is beyond me since I don’t have access to the machine.

Thanks,

Jon

*From:* Koenig, Christian <[email protected]> <mailto:[email protected]>
*Sent:* Wednesday, April 15, 2020 4:11 AM
*To:* Kim, Jonathan <[email protected]> <mailto:[email protected]>; Kuehling, Felix <[email protected]> <mailto:[email protected]>; Deucher, Alexander <[email protected]> <mailto:[email protected]>
*Cc:* Russell, Kent <[email protected]> <mailto:[email protected]>; [email protected] <mailto:[email protected]>
*Subject:* Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

Hi Jon,

Also cwsr tests fail on Vega20 with or without the revert with the same RAS error.

That sounds like the system/setup has a more general problem.

Could it be that we are seeing RAS errors because there really is some hardware failure, but with the MM path we don't trigger a RAS interrupt?

Thanks,
Christian.

Am 14.04.20 um 22:30 schrieb Kim, Jonathan:

[AMD Official Use Only - Internal Distribution Only]

If we’re passing the test on the revert, then the only thing that’s different is we’re not invalidating HDP and doing a copy to host anymore in amdgpu_device_vram_access since the function is still called in ttm access_memory with BAR.

Also cwsr tests fail on Vega20 with or without the revert with the same RAS error.

Thanks,

Jon

*From:* Kuehling, Felix <[email protected]> <mailto:[email protected]>
*Sent:* Tuesday, April 14, 2020 2:32 PM
*To:* Kim, Jonathan <[email protected]> <mailto:[email protected]>; Koenig, Christian <[email protected]> <mailto:[email protected]>; Deucher, Alexander <[email protected]> <mailto:[email protected]>
*Cc:* Russell, Kent <[email protected]> <mailto:[email protected]>; [email protected] <mailto:[email protected]>
*Subject:* Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

I wouldn't call it premature. Revert is a usual practice when there is a serious regression that isn't fully understood or root-caused. As far as I can tell, the problem has been reproduced on multiple systems, different GPUs, and clearly regressed to Christian's commit. I think that justifies reverting it for now.

I agree with Christian that a general HDP memory access problem causing RAS errors would potentially cause problems in other tests as well. For example common operations like GART table updates, and GPUVM page table updates and PCIe peer2peer accesses in ROCm applications use HDP. But we're not seeing obvious problems from those. So we need to understand what's special about this test. I asked questions to that effect on our other email thread.

Regards,
  Felix

Am 2020-04-14 um 10:51 a.m. schrieb Kim, Jonathan:

[AMD Official Use Only - Internal Distribution Only]

I think it’s premature to push this revert.

With more testing, I’m getting failures from different tests or sometimes none at all on my machine.

Kent, let’s continue the discussion on the original thread.

Thanks,

Jon

*From:* Koenig, Christian <[email protected]> <mailto:[email protected]>
*Sent:* Tuesday, April 14, 2020 10:47 AM
*To:* Deucher, Alexander <[email protected]> <mailto:[email protected]>
*Cc:* Russell, Kent <[email protected]> <mailto:[email protected]>; [email protected] <mailto:[email protected]>; Kuehling, Felix <[email protected]> <mailto:[email protected]>; Kim, Jonathan <[email protected]> <mailto:[email protected]>
*Subject:* Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

That's exactly my concern as well.

This looks a bit like the test creates erroneous data somehow, but there doesn't seems to be a RAS check in the MM data path.

And now that we use the BAR path it goes up in flames.

I just don't see how we can create erroneous data in a test case?

Christian.

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected] <mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

------------------------------------------------------------------------

*From:* amd-gfx <[email protected] <mailto:[email protected]>> on behalf of Russell, Kent <[email protected] <mailto:[email protected]>>
*Sent:* Tuesday, April 14, 2020 10:19 AM *To:* Koenig, Christian <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]> <[email protected] <mailto:[email protected]>> *Cc:* Kuehling, Felix <[email protected] <mailto:[email protected]>>; Kim, Jonathan <[email protected] <mailto:[email protected]>> *Subject:* RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

 Kent

> -----Original Message-----
> From: Christian König <[email protected] <mailto:[email protected]>>
> Sent: Tuesday, April 14, 2020 9:52 AM
> To: Russell, Kent <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]>
> Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
> amdgpu_device_vram_access v2"
>
> Am 13.04.20 um 20:20 schrieb Kent Russell:
> > This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e.
> > The original patch causes a RAS event and subsequent kernel hard-hang
> > when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and > > Arcturus
> >
> > dmesg output at hang time:
> > [drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected!
> > amdgpu 0000:67:00.0: GPU reset begin!
> > Evicting PASID 0x8000 queues
> > Started evicting pasid 0x8000
> > qcm fence wait loop timeout expired
> > The cp might be in an unrecoverable state due to an unsuccessful
> > queues preemption Failed to evict process queues Failed to suspend
> > process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid
> > 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost
> > due to RAS ERREVENT_ATHUB_INTERRUPT
> > amdgpu: [powerplay] Failed to send message 0x26, response 0x0
> > amdgpu: [powerplay] Failed to set soft min gfxclk ! > > amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
> > amdgpu: [powerplay] Failed to send message 0x7, response 0x0
> > amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu
> features!
> > amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
> > amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM!
> > [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP
> > block <powerplay> failed -5
>
> Do you have more information on what's going wrong here since this is a really
> important patch for KFD debugging.
>
> >
> > Signed-off-by: Kent Russell <[email protected] <mailto:[email protected]>>
>
> Reviewed-by: Christian König <[email protected] <mailto:[email protected]>>
>
> > ---
> > drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26
----------------------
> >   1 file changed, 26 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > index cf5d6e585634..a3f997f84020 100644 > > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > @@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct
> amdgpu_device *adev, loff_t pos,
> >      uint32_t hi = ~0;
> >      uint64_t last;
> >
> > -
> > -#ifdef CONFIG_64BIT
> > -   last = min(pos + size, adev->gmc.visible_vram_size);
> > -   if (last > pos) {
> > -           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
> > -           size_t count = last - pos;
> > -
> > -           if (write) {
> > - memcpy_toio(addr, buf, count);
> > -                   mb();
> > - amdgpu_asic_flush_hdp(adev, NULL);
> > -           } else {
> > - amdgpu_asic_invalidate_hdp(adev, NULL); > > -                   mb();
> > - memcpy_fromio(buf, addr, count);
> > -           }
> > -
> > -           if (count == size)
> > -                   return;
> > -
> > -           pos += count;
> > -           buf += count / 4;
> > -           size -= count;
> > -   }
> > -#endif
> > -
> > spin_lock_irqsave(&adev->mmio_idx_lock, flags); > >      for (last = pos + size; pos < last; pos += 4) { > > uint32_t tmp = pos >> 31;
_______________________________________________
amd-gfx mailing list [email protected] <mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected] <mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

------------------------------------------------------------------------

*From:* amd-gfx <[email protected] <mailto:[email protected]>> on behalf of Russell, Kent <[email protected] <mailto:[email protected]>>
*Sent:* Tuesday, April 14, 2020 10:19 AM *To:* Koenig, Christian <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]> <[email protected] <mailto:[email protected]>> *Cc:* Kuehling, Felix <[email protected] <mailto:[email protected]>>; Kim, Jonathan <[email protected] <mailto:[email protected]>> *Subject:* RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

 Kent

> -----Original Message-----
> From: Christian König <[email protected] <mailto:[email protected]>>
> Sent: Tuesday, April 14, 2020 9:52 AM
> To: Russell, Kent <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]>
> Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
> amdgpu_device_vram_access v2"
>
> Am 13.04.20 um 20:20 schrieb Kent Russell:
> > This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e.
> > The original patch causes a RAS event and subsequent kernel hard-hang
> > when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and > > Arcturus
> >
> > dmesg output at hang time:
> > [drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected!
> > amdgpu 0000:67:00.0: GPU reset begin!
> > Evicting PASID 0x8000 queues
> > Started evicting pasid 0x8000
> > qcm fence wait loop timeout expired
> > The cp might be in an unrecoverable state due to an unsuccessful
> > queues preemption Failed to evict process queues Failed to suspend
> > process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid
> > 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost
> > due to RAS ERREVENT_ATHUB_INTERRUPT
> > amdgpu: [powerplay] Failed to send message 0x26, response 0x0
> > amdgpu: [powerplay] Failed to set soft min gfxclk ! > > amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
> > amdgpu: [powerplay] Failed to send message 0x7, response 0x0
> > amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu
> features!
> > amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
> > amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM!
> > [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP
> > block <powerplay> failed -5
>
> Do you have more information on what's going wrong here since this is a really
> important patch for KFD debugging.
>
> >
> > Signed-off-by: Kent Russell <[email protected] <mailto:[email protected]>>
>
> Reviewed-by: Christian König <[email protected] <mailto:[email protected]>>
>
> > ---
> > drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26
----------------------
> >   1 file changed, 26 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > index cf5d6e585634..a3f997f84020 100644 > > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > @@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct
> amdgpu_device *adev, loff_t pos,
> >      uint32_t hi = ~0;
> >      uint64_t last;
> >
> > -
> > -#ifdef CONFIG_64BIT
> > -   last = min(pos + size, adev->gmc.visible_vram_size);
> > -   if (last > pos) {
> > -           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
> > -           size_t count = last - pos;
> > -
> > -           if (write) {
> > - memcpy_toio(addr, buf, count);
> > - mb();
> > - amdgpu_asic_flush_hdp(adev, NULL);
> > -           } else {
> > - amdgpu_asic_invalidate_hdp(adev, NULL); > > - mb();
> > - memcpy_fromio(buf, addr, count);
> > -           }
> > -
> > -           if (count == size)
> > - return;
> > -
> > -           pos += count;
> > -           buf += count / 4;
> > -           size -= count;
> > -   }
> > -#endif
> > -
> > spin_lock_irqsave(&adev->mmio_idx_lock, flags); > >      for (last = pos + size; pos < last; pos += 4) { > > uint32_t tmp = pos >> 31;
_______________________________________________
amd-gfx mailing list [email protected] <mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected] <mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

------------------------------------------------------------------------

*From:* amd-gfx <[email protected] <mailto:[email protected]>> on behalf of Russell, Kent <[email protected] <mailto:[email protected]>>
*Sent:* Tuesday, April 14, 2020 10:19 AM *To:* Koenig, Christian <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]> <[email protected] <mailto:[email protected]>> *Cc:* Kuehling, Felix <[email protected] <mailto:[email protected]>>; Kim, Jonathan <[email protected] <mailto:[email protected]>> *Subject:* RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

 Kent

> -----Original Message-----
> From: Christian König <[email protected] <mailto:[email protected]>>
> Sent: Tuesday, April 14, 2020 9:52 AM
> To: Russell, Kent <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]>
> Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
> amdgpu_device_vram_access v2"
>
> Am 13.04.20 um 20:20 schrieb Kent Russell:
> > This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e.
> > The original patch causes a RAS event and subsequent kernel hard-hang
> > when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and > > Arcturus
> >
> > dmesg output at hang time:
> > [drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected!
> > amdgpu 0000:67:00.0: GPU reset begin!
> > Evicting PASID 0x8000 queues
> > Started evicting pasid 0x8000
> > qcm fence wait loop timeout expired
> > The cp might be in an unrecoverable state due to an unsuccessful
> > queues preemption Failed to evict process queues Failed to suspend
> > process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid
> > 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost
> > due to RAS ERREVENT_ATHUB_INTERRUPT
> > amdgpu: [powerplay] Failed to send message 0x26, response 0x0
> > amdgpu: [powerplay] Failed to set soft min gfxclk ! > > amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
> > amdgpu: [powerplay] Failed to send message 0x7, response 0x0
> > amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu
> features!
> > amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
> > amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM!
> > [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP
> > block <powerplay> failed -5
>
> Do you have more information on what's going wrong here since this is a really
> important patch for KFD debugging.
>
> >
> > Signed-off-by: Kent Russell <[email protected] <mailto:[email protected]>>
>
> Reviewed-by: Christian König <[email protected] <mailto:[email protected]>>
>
> > ---
> > drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26
----------------------
> >   1 file changed, 26 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > index cf5d6e585634..a3f997f84020 100644 > > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > @@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct
> amdgpu_device *adev, loff_t pos,
> >      uint32_t hi = ~0;
> >      uint64_t last;
> >
> > -
> > -#ifdef CONFIG_64BIT
> > -   last = min(pos + size, adev->gmc.visible_vram_size);
> > -   if (last > pos) {
> > -           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
> > -           size_t count = last - pos;
> > -
> > -           if (write) {
> > - memcpy_toio(addr, buf, count);
> > - mb();
> > - amdgpu_asic_flush_hdp(adev, NULL);
> > -           } else {
> > - amdgpu_asic_invalidate_hdp(adev, NULL); > > - mb();
> > - memcpy_fromio(buf, addr, count);
> > -           }
> > -
> > -           if (count == size)
> > - return;
> > -
> > -           pos += count;
> > -           buf += count / 4;
> > -           size -= count;
> > -   }
> > -#endif
> > -
> > spin_lock_irqsave(&adev->mmio_idx_lock, flags); > >      for (last = pos + size; pos < last; pos += 4) { > >              uint32_t tmp = pos >> 31;
_______________________________________________
amd-gfx mailing list [email protected] <mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected] <mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

------------------------------------------------------------------------

*From:* amd-gfx <[email protected] <mailto:[email protected]>> on behalf of Russell, Kent <[email protected] <mailto:[email protected]>>
*Sent:* Tuesday, April 14, 2020 10:19 AM *To:* Koenig, Christian <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]> <[email protected] <mailto:[email protected]>> *Cc:* Kuehling, Felix <[email protected] <mailto:[email protected]>>; Kim, Jonathan <[email protected] <mailto:[email protected]>> *Subject:* RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

 Kent

> -----Original Message-----
> From: Christian König <[email protected] <mailto:[email protected]>>
> Sent: Tuesday, April 14, 2020 9:52 AM
> To: Russell, Kent <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]>
> Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
> amdgpu_device_vram_access v2"
>
> Am 13.04.20 um 20:20 schrieb Kent Russell:
> > This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e.
> > The original patch causes a RAS event and subsequent kernel hard-hang
> > when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and > > Arcturus
> >
> > dmesg output at hang time:
> > [drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected!
> > amdgpu 0000:67:00.0: GPU reset begin!
> > Evicting PASID 0x8000 queues
> > Started evicting pasid 0x8000
> > qcm fence wait loop timeout expired
> > The cp might be in an unrecoverable state due to an unsuccessful
> > queues preemption Failed to evict process queues Failed to suspend
> > process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid
> > 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost
> > due to RAS ERREVENT_ATHUB_INTERRUPT
> > amdgpu: [powerplay] Failed to send message 0x26, response 0x0
> > amdgpu: [powerplay] Failed to set soft min gfxclk ! > > amdgpu: [powerplay] Failed to upload DPM Bootup Levels!
> > amdgpu: [powerplay] Failed to send message 0x7, response 0x0
> > amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu
> features!
> > amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
> > amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM!
> > [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP
> > block <powerplay> failed -5
>
> Do you have more information on what's going wrong here since this is a really
> important patch for KFD debugging.
>
> >
> > Signed-off-by: Kent Russell <[email protected] <mailto:[email protected]>>
>
> Reviewed-by: Christian König <[email protected] <mailto:[email protected]>>
>
> > ---
> > drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26
----------------------
> >   1 file changed, 26 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > index cf5d6e585634..a3f997f84020 100644 > > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > @@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct
> amdgpu_device *adev, loff_t pos,
> >      uint32_t hi = ~0;
> >      uint64_t last;
> >
> > -
> > -#ifdef CONFIG_64BIT
> > -   last = min(pos + size, adev->gmc.visible_vram_size);
> > -   if (last > pos) {
> > -           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
> > -           size_t count = last - pos;
> > -
> > -           if (write) {
> > - memcpy_toio(addr, buf, count);
> > - mb();
> > - amdgpu_asic_flush_hdp(adev, NULL);
> > -           } else {
> > - amdgpu_asic_invalidate_hdp(adev, NULL); > > - mb();
> > - memcpy_fromio(buf, addr, count);
> > -           }
> > -
> > -           if (count == size)
> > - return;
> > -
> > -           pos += count;
> > -           buf += count / 4;
> > -           size -= count;
> > -   }
> > -#endif
> > -
> > spin_lock_irqsave(&adev->mmio_idx_lock, flags); > >      for (last = pos + size; pos < last; pos += 4) { > >              uint32_t tmp = pos >> 31;
_______________________________________________
amd-gfx mailing list [email protected] <mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

Am 14.04.2020 16:35 schrieb "Deucher, Alexander" <[email protected] <mailto:[email protected]>>:

[AMD Public Use]

If this causes an issue, any access to vram via the BAR could cause an issue.

Alex

------------------------------------------------------------------------

*From:* amd-gfx <[email protected] <mailto:[email protected]>> on behalf of Russell, Kent <[email protected] <mailto:[email protected]>>
*Sent:* Tuesday, April 14, 2020 10:19 AM
*To:* Koenig, Christian <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]> <[email protected] <mailto:[email protected]>>
*Cc:* Kuehling, Felix <[email protected] <mailto:[email protected]>>; Kim, Jonathan <[email protected] <mailto:[email protected]>> *Subject:* RE: [PATCH] Revert "drm/amdgpu: use the BAR if possible in amdgpu_device_vram_access v2"

[AMD Official Use Only - Internal Distribution Only]

On VG20 or MI100, as soon as we run the subtest, we get the dmesg output below, and then the kernel ends up hanging. I don't know enough about the test itself to know why this is occurring, but Jon Kim and Felix were discussing it on a separate thread when the issue was first reported, so they can hopefully provide some additional information.

 Kent

> -----Original Message-----
> From: Christian König <[email protected] <mailto:[email protected]>>
> Sent: Tuesday, April 14, 2020 9:52 AM
> To: Russell, Kent <[email protected] <mailto:[email protected]>>; [email protected] <mailto:[email protected]>
> Subject: Re: [PATCH] Revert "drm/amdgpu: use the BAR if possible in
> amdgpu_device_vram_access v2"
>
> Am 13.04.20 um 20:20 schrieb Kent Russell:
> > This reverts commit c12b84d6e0d70f1185e6daddfd12afb671791b6e.
> > The original patch causes a RAS event and subsequent kernel hard-hang
> > when running the KFDMemoryTest.PtraceAccessInvisibleVram on VG20 and > > Arcturus
> >
> > dmesg output at hang time:
> > [drm] RAS event of type ERREVENT_ATHUB_INTERRUPT detected! > > amdgpu 0000:67:00.0: GPU reset begin!
> > Evicting PASID 0x8000 queues
> > Started evicting pasid 0x8000
> > qcm fence wait loop timeout expired
> > The cp might be in an unrecoverable state due to an unsuccessful
> > queues preemption Failed to evict process queues Failed to suspend
> > process 0x8000 Finished evicting pasid 0x8000 Started restoring pasid
> > 0x8000 Finished restoring pasid 0x8000 [drm] UVD VCPU state may lost
> > due to RAS ERREVENT_ATHUB_INTERRUPT
> > amdgpu: [powerplay] Failed to send message 0x26, response 0x0
> > amdgpu: [powerplay] Failed to set soft min gfxclk ! > > amdgpu: [powerplay] Failed to upload DPM Bootup Levels! > > amdgpu: [powerplay] Failed to send message 0x7, response 0x0
> > amdgpu: [powerplay] [DisableAllSMUFeatures] Failed to disable all smu
> features!
> > amdgpu: [powerplay] [DisableDpmTasks] Failed to disable all smu features!
> > amdgpu: [powerplay] [PowerOffAsic] Failed to disable DPM! > > [drm:amdgpu_device_ip_suspend_phase2 [amdgpu]] *ERROR* suspend of IP
> > block <powerplay> failed -5
>
> Do you have more information on what's going wrong here since this is a really
> important patch for KFD debugging.
>
> >
> > Signed-off-by: Kent Russell <[email protected] <mailto:[email protected]>>
>
> Reviewed-by: Christian König <[email protected] <mailto:[email protected]>>
>
> > ---
> > drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 26
----------------------
> >   1 file changed, 26 deletions(-)
> >
> > diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c
> > index cf5d6e585634..a3f997f84020 100644
> > --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_device.c > > @@ -254,32 +254,6 @@ void amdgpu_device_vram_access(struct > amdgpu_device *adev, loff_t pos,
> >      uint32_t hi = ~0;
> >      uint64_t last;
> >
> > -
> > -#ifdef CONFIG_64BIT
> > -   last = min(pos + size, adev->gmc.visible_vram_size); > > -   if (last > pos) {
> > -           void __iomem *addr = adev->mman.aper_base_kaddr + pos;
> > -           size_t count = last - pos;
> > -
> > -           if (write) {
> > - memcpy_toio(addr, buf, count);
> > -                   mb();
> > - amdgpu_asic_flush_hdp(adev, NULL);
> > -           } else {
> > - amdgpu_asic_invalidate_hdp(adev, NULL);
> > -                   mb();
> > - memcpy_fromio(buf, addr, count);
> > -           }
> > -
> > -           if (count == size)
> > -                   return;
> > -
> > -           pos += count;
> > -           buf += count / 4;
> > -           size -= count;
> > -   }
> > -#endif
> > -

Quoted 9 lines

> > spin_lock_irqsave(&adev->mmio_idx_lock, flags); > >      for (last = pos + size; pos < last; pos += 4) { > >              uint32_t tmp = pos >> 31;
_______________________________________________
amd-gfx mailing list [email protected] <mailto:[email protected]>
https://nam11.safelinks.protection.outlook.com/?url=https%3A%2F%2Flists.freedesktop.org%2Fmailman%2Flistinfo%2Famd-gfx&amp;data=02%7C01%7Calexander.deucher%40amd.com%7C68e0bfea2a5f4a909ab108d7e07ed164%7C3dd8961fe4884e608e11a82d994e183d%7C0%7C0%7C637224707637289768&amp;sdata=ttNOHJt0IwywpOIWahKjjuC6OkT1jxduc6iMzYzndpg%3D&amp;reserved=0

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