kernel_5.10 no rt

hc

2023-12-11 6778948f9de86c3cfaf36725a7c87dcff9ba247f

 kernel/drivers/gpu/drm/amd/amdkfd/kfd_events.c
..
..
@@ -187,7 +187,7 @@
187
187
 	if (p->signal_mapped_size &&
188
188
 	    p->signal_event_count == p->signal_mapped_size / 8) {
189
189
 		if (!p->signal_event_limit_reached) {
190
-			pr_warn("Signal event wasn't created because limit was reached\n");
190
+			pr_debug("Signal event wasn't created because limit was reached\n");
191
191
 			p->signal_event_limit_reached = true;
192
192
 		}
193
193
 		return -ENOSPC;
..
..
@@ -346,7 +346,6 @@
346
346
 		ret = create_signal_event(devkfd, p, ev);
347
347
 		if (!ret) {
348
348
 			*event_page_offset = KFD_MMAP_TYPE_EVENTS;
349
-			*event_page_offset <<= PAGE_SHIFT;
350
349
 			*event_slot_index = ev->event_id;
351
350
 		}
352
351
 		break;
..
..
@@ -461,7 +460,7 @@
461
460
 	}
462
461
 }
463
462
 
464
-void kfd_signal_event_interrupt(unsigned int pasid, uint32_t partial_id,
463
+void kfd_signal_event_interrupt(u32 pasid, uint32_t partial_id,
465
464
 				uint32_t valid_id_bits)
466
465
 {
467
466
 	struct kfd_event *ev = NULL;
..
..
@@ -532,6 +531,8 @@
532
531
 	event_waiters = kmalloc_array(num_events,
533
532
 					sizeof(struct kfd_event_waiter),
534
533
 					GFP_KERNEL);
534
+	if (!event_waiters)
535
+		return NULL;
535
536
 
536
537
 	for (i = 0; (event_waiters) && (i < num_events) ; i++) {
537
538
 		init_wait(&event_waiters[i].wait);
..
..
@@ -852,8 +853,8 @@
852
853
 
853
854
 	if (type == KFD_EVENT_TYPE_MEMORY) {
854
855
 		dev_warn(kfd_device,
855
-			"Sending SIGSEGV to HSA Process with PID %d ",
856
-				p->lead_thread->pid);
856
+			"Sending SIGSEGV to process %d (pasid 0x%x)",
857
+				p->lead_thread->pid, p->pasid);
857
858
 		send_sig(SIGSEGV, p->lead_thread, 0);
858
859
 	}
859
860
 
..
..
@@ -861,19 +862,19 @@
861
862
 	if (send_signal) {
862
863
 		if (send_sigterm) {
863
864
 			dev_warn(kfd_device,
864
-				"Sending SIGTERM to HSA Process with PID %d ",
865
-					p->lead_thread->pid);
865
+				"Sending SIGTERM to process %d (pasid 0x%x)",
866
+					p->lead_thread->pid, p->pasid);
866
867
 			send_sig(SIGTERM, p->lead_thread, 0);
867
868
 		} else {
868
869
 			dev_err(kfd_device,
869
-				"HSA Process (PID %d) got unhandled exception",
870
-				p->lead_thread->pid);
870
+				"Process %d (pasid 0x%x) got unhandled exception",
871
+				p->lead_thread->pid, p->pasid);
871
872
 		}
872
873
 	}
873
874
 }
874
875
 
875
876
 #ifdef KFD_SUPPORT_IOMMU_V2
876
-void kfd_signal_iommu_event(struct kfd_dev *dev, unsigned int pasid,
877
+void kfd_signal_iommu_event(struct kfd_dev *dev, u32 pasid,
877
878
 		unsigned long address, bool is_write_requested,
878
879
 		bool is_execute_requested)
879
880
 {
..
..
@@ -902,7 +903,7 @@
902
903
 
903
904
 	memset(&memory_exception_data, 0, sizeof(memory_exception_data));
904
905
 
905
-	down_read(&mm->mmap_sem);
906
+	mmap_read_lock(mm);
906
907
 	vma = find_vma(mm, address);
907
908
 
908
909
 	memory_exception_data.gpu_id = dev->id;
..
..
@@ -925,7 +926,7 @@
925
926
 			memory_exception_data.failure.NoExecute = 0;
926
927
 	}
927
928
 
928
-	up_read(&mm->mmap_sem);
929
+	mmap_read_unlock(mm);
929
930
 	mmput(mm);
930
931
 
931
932
 	pr_debug("notpresent %d, noexecute %d, readonly %d\n",
..
..
@@ -936,7 +937,8 @@
936
937
 	/* Workaround on Raven to not kill the process when memory is freed
937
938
 	 * before IOMMU is able to finish processing all the excessive PPRs
938
939
 	 */
939
-	if (dev->device_info->asic_family != CHIP_RAVEN) {
940
+	if (dev->device_info->asic_family != CHIP_RAVEN &&
941
+	    dev->device_info->asic_family != CHIP_RENOIR) {
940
942
 		mutex_lock(&p->event_mutex);
941
943
 
942
944
 		/* Lookup events by type and signal them */
..
..
@@ -950,7 +952,7 @@
950
952
 }
951
953
 #endif /* KFD_SUPPORT_IOMMU_V2 */
952
954
 
953
-void kfd_signal_hw_exception_event(unsigned int pasid)
955
+void kfd_signal_hw_exception_event(u32 pasid)
954
956
 {
955
957
 	/*
956
958
 	 * Because we are called from arbitrary context (workqueue) as opposed
..
..
@@ -971,7 +973,7 @@
971
973
 	kfd_unref_process(p);
972
974
 }
973
975
 
974
-void kfd_signal_vm_fault_event(struct kfd_dev *dev, unsigned int pasid,
976
+void kfd_signal_vm_fault_event(struct kfd_dev *dev, u32 pasid,
975
977
 				struct kfd_vm_fault_info *info)
976
978
 {
977
979
 	struct kfd_event *ev;
..
..
@@ -983,7 +985,7 @@
983
985
 		return; /* Presumably process exited. */
984
986
 	memset(&memory_exception_data, 0, sizeof(memory_exception_data));
985
987
 	memory_exception_data.gpu_id = dev->id;
986
-	memory_exception_data.failure.imprecise = 1;
988
+	memory_exception_data.failure.imprecise = true;
987
989
 	/* Set failure reason */
988
990
 	if (info) {
989
991
 		memory_exception_data.va = (info->page_addr) << PAGE_SHIFT;
..
..
@@ -1011,25 +1013,41 @@
1011
1013
 void kfd_signal_reset_event(struct kfd_dev *dev)
1012
1014
 {
1013
1015
 	struct kfd_hsa_hw_exception_data hw_exception_data;
1016
+	struct kfd_hsa_memory_exception_data memory_exception_data;
1014
1017
 	struct kfd_process *p;
1015
1018
 	struct kfd_event *ev;
1016
1019
 	unsigned int temp;
1017
1020
 	uint32_t id, idx;
1021
+	int reset_cause = atomic_read(&dev->sram_ecc_flag) ?
1022
+			KFD_HW_EXCEPTION_ECC :
1023
+			KFD_HW_EXCEPTION_GPU_HANG;
1018
1024
 
1019
1025
 	/* Whole gpu reset caused by GPU hang and memory is lost */
1020
1026
 	memset(&hw_exception_data, 0, sizeof(hw_exception_data));
1021
1027
 	hw_exception_data.gpu_id = dev->id;
1022
1028
 	hw_exception_data.memory_lost = 1;
1029
+	hw_exception_data.reset_cause = reset_cause;
1030
+
1031
+	memset(&memory_exception_data, 0, sizeof(memory_exception_data));
1032
+	memory_exception_data.ErrorType = KFD_MEM_ERR_SRAM_ECC;
1033
+	memory_exception_data.gpu_id = dev->id;
1034
+	memory_exception_data.failure.imprecise = true;
1023
1035
 
1024
1036
 	idx = srcu_read_lock(&kfd_processes_srcu);
1025
1037
 	hash_for_each_rcu(kfd_processes_table, temp, p, kfd_processes) {
1026
1038
 		mutex_lock(&p->event_mutex);
1027
1039
 		id = KFD_FIRST_NONSIGNAL_EVENT_ID;
1028
-		idr_for_each_entry_continue(&p->event_idr, ev, id)
1040
+		idr_for_each_entry_continue(&p->event_idr, ev, id) {
1029
1041
 			if (ev->type == KFD_EVENT_TYPE_HW_EXCEPTION) {
1030
1042
 				ev->hw_exception_data = hw_exception_data;
1031
1043
 				set_event(ev);
1032
1044
 			}
1045
+			if (ev->type == KFD_EVENT_TYPE_MEMORY &&
1046
+			    reset_cause == KFD_HW_EXCEPTION_ECC) {
1047
+				ev->memory_exception_data = memory_exception_data;
1048
+				set_event(ev);
1049
+			}
1050
+		}
1033
1051
 		mutex_unlock(&p->event_mutex);
1034
1052
 	}
1035
1053
 	srcu_read_unlock(&kfd_processes_srcu, idx);