[PATCH 3/3] drm/amdkfd: add RAS poison consumption support for utcl2

[Zhang, Hawking]

[AMD Official Use Only]

Copy Felix

@@ -119,10 +121,14 @@ static void event_interrupt_poison_consumption(struct kfd_dev *dev,
 		break;
 	case SOC15_INTSRC_SDMA_ECC:
 	default:
+		if (client_id == SOC15_IH_CLIENTID_UTCL2)
+			ret = kfd_dqm_evict_pasid(dev->dqm, pasid);
 		break;
 	}

This will break SDMA - We haven't enabled optimized poison consumption handling for sdma yet.

I'd suggest we explicitly call out the interrupt source id UTCL2_FAULT as a case, even it is the same as VM_FAULT. And it should be fine to start evict_queue directly after that because in ISR it already guarantee this is from UTCL2 client, right?

+		if (client_id == SOC15_IH_CLIENTID_UTCL2 &&
+		    dev->kfd2kgd->is_ras_utcl2_poison &&
+		    dev->kfd2kgd->is_ras_utcl2_poison(dev->adev, client_id)) {
+			event_interrupt_poison_consumption(dev, ih_ring_entry);


In addition, is_ras_utcl2_poison can be renamed to query_utcl2_ras_status or poison_status, while utcl2_fault_clear to reset_utlc2_poison_status to align with naming style of ras hw op.

Thinking about this more, it's better we add this in gfx ras op, and expose to KFD. Thoughts?

Regards,
Hawking

-----Original Message-----
From: Zhou1, Tao <Tao.Zhou1 at amd.com> 
Sent: Monday, March 14, 2022 15:03
To: amd-gfx at lists.freedesktop.org; Zhang, Hawking <Hawking.Zhang at amd.com>; Yang, Stanley <Stanley.Yang at amd.com>; Chai, Thomas <YiPeng.Chai at amd.com>
Cc: Zhou1, Tao <Tao.Zhou1 at amd.com>
Subject: [PATCH 3/3] drm/amdkfd: add RAS poison consumption support for utcl2

Do RAS page retirement and use gpu reset as fallback in utcl2 fault handler.

Signed-off-by: Tao Zhou <tao.zhou1 at amd.com>
---
 .../gpu/drm/amd/amdkfd/kfd_int_process_v9.c   | 23 ++++++++++++++++---
 1 file changed, 20 insertions(+), 3 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c
index f7def0bf0730..3991f71d865b 100644
--- a/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c
+++ b/drivers/gpu/drm/amd/amdkfd/kfd_int_process_v9.c
@@ -93,11 +93,12 @@ enum SQ_INTERRUPT_ERROR_TYPE {  static void event_interrupt_poison_consumption(struct kfd_dev *dev,
 				const uint32_t *ih_ring_entry)
 {
-	uint16_t source_id, pasid;
+	uint16_t source_id, client_id, pasid;
 	int ret = -EINVAL;
 	struct kfd_process *p;
 
 	source_id = SOC15_SOURCE_ID_FROM_IH_ENTRY(ih_ring_entry);
+	client_id = SOC15_CLIENT_ID_FROM_IH_ENTRY(ih_ring_entry);
 	pasid = SOC15_PASID_FROM_IH_ENTRY(ih_ring_entry);
 
 	p = kfd_lookup_process_by_pasid(pasid);
@@ -110,6 +111,7 @@ static void event_interrupt_poison_consumption(struct kfd_dev *dev,
 		return;
 	}
 
+	pr_debug("RAS poison consumption handling\n");
 	atomic_set(&p->poison, 1);
 	kfd_unref_process(p);
 
@@ -119,10 +121,14 @@ static void event_interrupt_poison_consumption(struct kfd_dev *dev,
 		break;
 	case SOC15_INTSRC_SDMA_ECC:
 	default:
+		if (client_id == SOC15_IH_CLIENTID_UTCL2)
+			ret = kfd_dqm_evict_pasid(dev->dqm, pasid);
 		break;
 	}
 
-	kfd_signal_poison_consumed_event(dev, pasid);
+	/* utcl2 page fault has its own vm fault event */
+	if (client_id != SOC15_IH_CLIENTID_UTCL2)
+		kfd_signal_poison_consumed_event(dev, pasid);
 
 	/* resetting queue passes, do page retirement without gpu reset
 	 * resetting queue fails, fallback to gpu reset solution @@ -314,7 +320,18 @@ static void event_interrupt_wq_v9(struct kfd_dev *dev,
 		info.prot_write = ring_id & 0x20;
 
 		kfd_smi_event_update_vmfault(dev, pasid);
-		kfd_dqm_evict_pasid(dev->dqm, pasid);
+
+		if (client_id == SOC15_IH_CLIENTID_UTCL2 &&
+		    dev->kfd2kgd->is_ras_utcl2_poison &&
+		    dev->kfd2kgd->is_ras_utcl2_poison(dev->adev, client_id)) {
+			event_interrupt_poison_consumption(dev, ih_ring_entry);
+
+			if (dev->kfd2kgd->utcl2_fault_clear)
+				dev->kfd2kgd->utcl2_fault_clear(dev->adev);
+		}
+		else
+			kfd_dqm_evict_pasid(dev->dqm, pasid);
+
 		kfd_signal_vm_fault_event(dev, pasid, &info);
 	}
 }
--
2.35.1