drm/amdgpu: Fix the looply call svm_range_restore_pages issue
Subsystems (2)
PATCH 1/2 drm/amdgpu: Fix the looply call svm_range_restore_pages issue
As the delayed free pt, the wanted freed bo has been reused which will cause unexpected page fault, and then call svm_range_restore_pages.
Detail as below:
1.It wants to free the pt in follow code, but it is not freed immediately and used “schedule_work(&vm->pt_free_work);”.
[ 92.276838] Call Trace: [ 92.276841] dump_stack+0x63/0xa0 [ 92.276887] amdgpu_vm_pt_free_list+0xfb/0x120 [amdgpu] [ 92.276932] amdgpu_vm_update_range+0x69c/0x8e0 [amdgpu] [ 92.276990] svm_range_unmap_from_gpus+0x112/0x310 [amdgpu] [ 92.277046] svm_range_cpu_invalidate_pagetables+0x725/0x780 [amdgpu] [ 92.277050] ? __alloc_pages_nodemask+0x19f/0x3e0 [ 92.277051] mn_itree_invalidate+0x72/0xc0 [ 92.277052] __mmu_notifier_invalidate_range_start+0x48/0x60 [ 92.277054] migrate_vma_collect+0xf6/0x100 [ 92.277055] migrate_vma_setup+0xcf/0x120 [ 92.277109] svm_migrate_ram_to_vram+0x256/0x6b0 [amdgpu]
2.Call svm_range_map_to_gpu->amdgpu_vm_update_range to update the page table, at this time it will use the same entry bo which is the want free bo in step1.
3.Then it executes the pt_free_work to free the bo. At this time, the GPU access memory will cause page fault as pt bo has been freed. And then it will call svm_range_restore_pages again.
How to fix?
Add a workqueue, and flush the workqueue each time before updating page table.
--- drivers/gpu/drm/amd/amdgpu/amdgpu.h | 1 + drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 8 ++++++-- drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c | 2 +- 3 files changed, 8 insertions(+), 3 deletions(-)
drivers/gpu/drm/amd/amdgpu/amdgpu.h +1 -0
@@ -1209,6 +1209,7 @@ struct amdgpu_device {struct mutex enforce_isolation_mutex;struct amdgpu_init_level *init_lvl; +struct workqueue_struct *wq; }; static inline uint32_t amdgpu_ip_version(const struct amdgpu_device *adev,
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +6 -2
@@ -983,6 +983,7 @@ int amdgpu_vm_update_range(struct amdgpu_device *adev, struct amdgpu_vm *vm,*/ flush_tlb |=amdgpu_ip_version (adev,GC_HWIP ,0 ) <IP_VERSION (9 ,0 ,0 ); +flush_workqueue (adev->wq);memset (¶ms,0 ,sizeof (params)); params.adev = adev; params.vm = vm;
@@ -2614,7 +2615,7 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)amdgpu_amdkfd_gpuvm_destroy_cb (adev, vm);flush_work (&vm->pt_free_work); - +cancel_work_sync (&vm->pt_free_work); root =amdgpu_bo_ref (vm->root.bo);amdgpu_bo_reserve (root, true);amdgpu_vm_put_task_info (vm->task_info);
@@ -2716,6 +2717,8 @@ void amdgpu_vm_manager_init(struct amdgpu_device *adev)#endif xa_init_flags (&adev->vm_manager.pasids,XA_FLAGS_LOCK_IRQ ); + adev->wq =alloc_workqueue ("amdgpu_recycle" , +WQ_MEM_RECLAIM |WQ_HIGHPRI |WQ_UNBOUND ,16 ); }/**
@@ -2729,7 +2732,8 @@ void amdgpu_vm_manager_fini(struct amdgpu_device *adev) {WARN_ON (!xa_empty (&adev->vm_manager.pasids));xa_destroy (&adev->vm_manager.pasids); - +flush_workqueue (adev->wq); +destroy_workqueue (adev->wq);amdgpu_vmid_mgr_fini (adev); }
drivers/gpu/drm/amd/amdgpu/amdgpu_vm_pt.c +1 -1
@@ -589,7 +589,7 @@ void amdgpu_vm_pt_free_list(struct amdgpu_device *adev,spin_lock (&vm->status_lock);list_splice_init (¶ms->tlb_flush_waitlist, &vm->pt_freed);spin_unlock (&vm->status_lock); -schedule_work (&vm->pt_free_work); +queue_work (adev->wq, &vm->pt_free_work);return ; }
Signature · 1 line
2.36.1