Skip to contentSkip to conversation

amd/amdkfd: sync all devices to wait all processes being evicted

PATCH In review Zhigang Luo 3 participants 4 messages Apr 1, 2024 – Apr 3, 2024

Subsystems (3)
lore
Share & tools
lore
#0 lore raw

PATCH 1/2 amd/amdkfd: sync all devices to wait all processes being evicted

If there are more than one device doing reset in parallel, the first device will call kfd_suspend_all_processes() to evict all processes on all devices, this call takes time to finish. other device will start reset and recover without waiting. if the process has not been evicted before doing recover, it will be restored, then caused page fault.

Change-Id Ib1eddb56b69ecd41fe703abd169944154f48b0cd
---
 drivers/gpu/drm/amd/amdkfd/kfd_device.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/amd/amdkfd/kfd_device.c +1 -1
index 041ec3de55e7..55f89c858c7a 100644
@@ -969,11 +969,11 @@ void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm)
 	if (!run_pm) {
 		mutex_lock(&kfd_processes_mutex);
 		count = ++kfd_locked;
-		mutex_unlock(&kfd_processes_mutex);
 
 		/* For first KFD device suspend all the KFD processes */
 		if (count == 1)
 			kfd_suspend_all_processes();
+		mutex_unlock(&kfd_processes_mutex);
 	}
 
 	for (i = 0; i < kfd->num_nodes; i++) {
Signature · 1 line
2.25.1
#1 lore raw

PATCH 2/2 amd/amdgpu: improve VF recover time

1. change AMDGPU_VF2PF_UPDATE_MAX_RETRY_LIMIT from 30 to 5.
2. set fatel error detected flag.

Change-Id If1e0357deffa4549d4e83e925c8d764f7f8c9f42
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_device.c | 1 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c   | 1 +
 drivers/gpu/drm/amd/amdgpu/amdgpu_virt.h   | 2 +-
 3 files changed, 3 insertions(+), 1 deletion(-)
drivers/gpu/drm/amd/amdgpu/amdgpu_device.c +1 -0
index 70261eb9b0bb..5e88f6f6a481 100644
@@ -4982,6 +4982,7 @@ static int amdgpu_device_reset_sriov(struct amdgpu_device *adev,
 		r = amdgpu_virt_reset_gpu(adev);
 	if (r)
 		return r;
+	amdgpu_ras_set_fed(adev, false);
 	amdgpu_irq_gpu_reset_resume_helper(adev);
 
 	/* some sw clean up VF needs to do before recover */
drivers/gpu/drm/amd/amdgpu/amdgpu_virt.c +1 -0
index aed60aaf1a55..4b6e50cfba08 100644
@@ -593,6 +593,7 @@ static void amdgpu_virt_update_vf2pf_work_item(struct work_struct *work)
 		adev->virt.vf2pf_update_retry_cnt++;
 		if ((adev->virt.vf2pf_update_retry_cnt >= AMDGPU_VF2PF_UPDATE_MAX_RETRY_LIMIT) &&
 		    amdgpu_sriov_runtime(adev) && !amdgpu_in_reset(adev)) {
+			amdgpu_ras_set_fed(adev, true);
 			if (amdgpu_reset_domain_schedule(adev->reset_domain,
 							  &adev->virt.flr_work))
 				return;
drivers/gpu/drm/amd/amdgpu/amdgpu_virt.h +1 -1
index a858bc98cad4..0a57b637c4bd 100644
@@ -52,7 +52,7 @@
 /* tonga/fiji use this offset */
 #define mmBIF_IOV_FUNC_IDENTIFIER 0x1503
 
-#define AMDGPU_VF2PF_UPDATE_MAX_RETRY_LIMIT 30
+#define AMDGPU_VF2PF_UPDATE_MAX_RETRY_LIMIT 5
 
 enum amdgpu_sriov_vf_mode {
 	SRIOV_VF_MODE_BARE_METAL = 0,
Signature · 1 line
2.25.1
maintainer
#2 lore raw
In reply to Zhigang Luo
Quoted 9 lines · Zhigang Luo

If there are more than one device doing reset in parallel, the first device will call kfd_suspend_all_processes() to evict all processes on all devices, this call takes time to finish. other device will start reset and recover without waiting. if the process has not been evicted before doing recover, it will be restored, then caused page fault.

Change-Id Ib1eddb56b69ecd41fe703abd169944154f48b0cd

Please remove the Change-Id: before you push. Other than that, this patch is

Quoted 9 lines
---
  drivers/gpu/drm/amd/amdkfd/kfd_device.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/amd/amdkfd/kfd_device.c +0 -0
index 041ec3de55e7..55f89c858c7a 100644

drivers/gpu/drm/amd/amdkfd/kfd_device.c void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm) · line 975

Quoted 8 lines
drivers/gpu/drm/amd/amdkfd/kfd_device.c +0 -1
@@ -969,11 +969,11 @@ void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm)
  	if (!run_pm) {
  		mutex_lock(&kfd_processes_mutex);
  		count = ++kfd_locked;
-		mutex_unlock(&kfd_processes_mutex);
 
  		/* For first KFD device suspend all the KFD processes */
  		if (count == 1)
  			kfd_suspend_all_processes();

This could be simplified now. The variable "count" was only needed for the broken attempt to do call suspend outside the lock. Now you can just do:

	mutex_lock(&kfd_processes_mutex);
	if (++kfd_locked == 1)
		kfd_suspend_all_processes();
	mutex_unlock(&kfd_processes_mutex);

To be consistent, we probably need to make a similar change in kgd2kfd_resume and run kfd_resume_all_processes under the lock as well. Otherwise there could be a race condition between suspend and resume.

Regards,
  Felix

Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_device.c +1 -0
+		mutex_unlock(&kfd_processes_mutex);
  	}
 
  	for (i = 0; i < kfd->num_nodes; i++) {
#3 lore raw
In reply to Zhigang Luo
Quoted 21 lines · Zhigang Luo

Caution: This message originated from an External Source. Use proper caution when opening attachments, clicking links, or responding.

If there are more than one device doing reset in parallel, the first device will call kfd_suspend_all_processes() to evict all processes on all devices, this call takes time to finish. other device will start reset and recover without waiting. if the process has not been evicted before doing recover, it will be restored, then caused page fault.

Change-Id Ib1eddb56b69ecd41fe703abd169944154f48b0cd
---
  drivers/gpu/drm/amd/amdkfd/kfd_device.c | 2 +-
  1 file changed, 1 insertion(+), 1 deletion(-)
drivers/gpu/drm/amd/amdkfd/kfd_device.c +0 -0
index 041ec3de55e7..55f89c858c7a 100644

drivers/gpu/drm/amd/amdkfd/kfd_device.c void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm) · line 977

Quoted 10 lines
drivers/gpu/drm/amd/amdkfd/kfd_device.c +1 -1
@@ -969,11 +969,11 @@ void kgd2kfd_suspend(struct kfd_dev *kfd, bool run_pm)
         if (!run_pm) {
                 mutex_lock(&kfd_processes_mutex);
                 count = ++kfd_locked;
-               mutex_unlock(&kfd_processes_mutex);
 
                 /* For first KFD device suspend all the KFD processes */
                 if (count == 1)
                         kfd_suspend_all_processes();
+               mutex_unlock(&kfd_processes_mutex);
         }

I do not understand why use kfd_lock here. You want evict all processes when first device got suspended. The kfd_lock indicates if all kfd driver functions got locked. It is not same meaning as device suspend. That is not your patch issue, but I think using different flag to record device suspend is better. ex, if kfd_lock got set for some other reasons, we would skip evicting processes here.

Regards

Xiaogang

Quoted 4 lines
drivers/gpu/drm/amd/amdkfd/kfd_device.c +0 -1
         for (i = 0; i < kfd->num_nodes; i++) {
--

2.25.1