[PATCH 0/4] Backported amdgpu ttm deadlock fixes for 4.14

[PATCH 0/4] Backported amdgpu ttm deadlock fixes for 4.14

[Lyude Paul]

I haven't gone to see where it started, but as of late a good number of
pretty nasty deadlock issues have appeared with the kernel. Easy
reproduction recipe on a laptop with i915/amdgpu prime with lockdep enabled:

DRI_PRIME=1 glxinfo

Additionally, some more race conditions exist that I've managed to
trigger with piglit and lockdep enabled after applying these patches:

    =============================
    WARNING: suspicious RCU usage
    4.14.3Lyude-Test+ #2 Not tainted
    -----------------------------
    ./include/linux/reservation.h:216 suspicious rcu_dereference_protected() usage!

    other info that might help us debug this:

    rcu_scheduler_active = 2, debug_locks = 1
    1 lock held by ext_image_dma_b/27451:
     #0:  (reservation_ww_class_mutex){+.+.}, at: [<ffffffffa034f2ff>] ttm_bo_unref+0x9f/0x3c0 [ttm]

    stack backtrace:
    CPU: 0 PID: 27451 Comm: ext_image_dma_b Not tainted 4.14.3Lyude-Test+ #2
    Hardware name: HP HP ZBook 15 G4/8275, BIOS P70 Ver. 01.02 06/09/2017
    Call Trace:
     dump_stack+0x8e/0xce
     lockdep_rcu_suspicious+0xc5/0x100
     reservation_object_copy_fences+0x292/0x2b0
     ? ttm_bo_unref+0x9f/0x3c0 [ttm]
     ttm_bo_unref+0xbd/0x3c0 [ttm]
     amdgpu_bo_unref+0x2a/0x50 [amdgpu]
     amdgpu_gem_object_free+0x4b/0x50 [amdgpu]
     drm_gem_object_free+0x1f/0x40 [drm]
     drm_gem_object_put_unlocked+0x40/0xb0 [drm]
     drm_gem_object_handle_put_unlocked+0x6c/0xb0 [drm]
     drm_gem_object_release_handle+0x51/0x90 [drm]
     drm_gem_handle_delete+0x5e/0x90 [drm]
     ? drm_gem_handle_create+0x40/0x40 [drm]
     drm_gem_close_ioctl+0x20/0x30 [drm]
     drm_ioctl_kernel+0x5d/0xb0 [drm]
     drm_ioctl+0x2f7/0x3b0 [drm]
     ? drm_gem_handle_create+0x40/0x40 [drm]
     ? trace_hardirqs_on_caller+0xf4/0x190
     ? trace_hardirqs_on+0xd/0x10
     amdgpu_drm_ioctl+0x4f/0x90 [amdgpu]
     do_vfs_ioctl+0x93/0x670
     ? __fget+0x108/0x1f0
     SyS_ioctl+0x79/0x90
     entry_SYSCALL_64_fastpath+0x23/0xc2

I've also added the relevant fixes for the issue mentioned above.

Christian König (3):
  drm/ttm: fix ttm_bo_cleanup_refs_or_queue once more
  dma-buf: make reservation_object_copy_fences rcu save
  drm/amdgpu: reserve root PD while releasing it

Michel Dänzer (1):
  drm/ttm: Always and only destroy bo->ttm_resv in ttm_bo_release_list

 drivers/dma-buf/reservation.c          | 56 +++++++++++++++++++++++++---------
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 13 ++++++--
 drivers/gpu/drm/ttm/ttm_bo.c           | 43 +++++++++++++-------------
 3 files changed, 74 insertions(+), 38 deletions(-)

--
2.14.3

_______________________________________________
amd-gfx mailing list
amd-gfx@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/amd-gfx

[PATCH 3/4] drm/amdgpu: reserve root PD while releasing it

[Lyude Paul]

From: Christian König <christian.koenig@amd.com>

Otherwise somebody could try to evict it at the same time and try to use
half torn down structures.

commit 2642cf110d08a403f585a051e4cbf45a90b3adea upstream

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-and-Tested-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Lyude Paul <lyude@redhat.com>
---
 drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 13 +++++++++++--
 1 file changed, 11 insertions(+), 2 deletions(-)

diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
index bd20ff018512..591c6dc2c1f7 100644
--- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
+++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c
@@ -2586,7 +2586,8 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
 {
 	struct amdgpu_bo_va_mapping *mapping, *tmp;
 	bool prt_fini_needed = !!adev->gart.gart_funcs->set_prt;
-	int i;
+	struct amdgpu_bo *root;
+	int i, r;
 
 	amd_sched_entity_fini(vm->entity.sched, &vm->entity);
 
@@ -2609,7 +2610,15 @@ void amdgpu_vm_fini(struct amdgpu_device *adev, struct amdgpu_vm *vm)
 		amdgpu_vm_free_mapping(adev, vm, mapping, NULL);
 	}
 
-	amdgpu_vm_free_levels(&vm->root);
+	root = amdgpu_bo_ref(vm->root.bo);
+	r = amdgpu_bo_reserve(root, true);
+	if (r) {
+		dev_err(adev->dev, "Leaking page tables because BO reservation failed\n");
+	} else {
+		amdgpu_vm_free_levels(&vm->root);
+		amdgpu_bo_unreserve(root);
+	}
+	amdgpu_bo_unref(&root);
 	dma_fence_put(vm->last_dir_update);
 	for (i = 0; i < AMDGPU_MAX_VMHUBS; i++)
 		amdgpu_vm_free_reserved_vmid(adev, vm, i);
-- 
2.14.3

_______________________________________________
dri-devel mailing list
dri-devel@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/dri-devel

Re: [PATCH 0/4] Backported amdgpu ttm deadlock fixes for 4.14

[Christian König]

Am 01.12.2017 um 01:23 schrieb Lyude Paul:
> I haven't gone to see where it started, but as of late a good number of
> pretty nasty deadlock issues have appeared with the kernel. Easy
> reproduction recipe on a laptop with i915/amdgpu prime with lockdep enabled:
>
> DRI_PRIME=1 glxinfo

Acked-by: Christian König <christian.koenig@amd.com>

Thanks for taking care of this,
Christian.

>
> Additionally, some more race conditions exist that I've managed to
> trigger with piglit and lockdep enabled after applying these patches:
>
>      =============================
>      WARNING: suspicious RCU usage
>      4.14.3Lyude-Test+ #2 Not tainted
>      -----------------------------
>      ./include/linux/reservation.h:216 suspicious rcu_dereference_protected() usage!
>
>      other info that might help us debug this:
>
>      rcu_scheduler_active = 2, debug_locks = 1
>      1 lock held by ext_image_dma_b/27451:
>       #0:  (reservation_ww_class_mutex){+.+.}, at: [<ffffffffa034f2ff>] ttm_bo_unref+0x9f/0x3c0 [ttm]
>
>      stack backtrace:
>      CPU: 0 PID: 27451 Comm: ext_image_dma_b Not tainted 4.14.3Lyude-Test+ #2
>      Hardware name: HP HP ZBook 15 G4/8275, BIOS P70 Ver. 01.02 06/09/2017
>      Call Trace:
>       dump_stack+0x8e/0xce
>       lockdep_rcu_suspicious+0xc5/0x100
>       reservation_object_copy_fences+0x292/0x2b0
>       ? ttm_bo_unref+0x9f/0x3c0 [ttm]
>       ttm_bo_unref+0xbd/0x3c0 [ttm]
>       amdgpu_bo_unref+0x2a/0x50 [amdgpu]
>       amdgpu_gem_object_free+0x4b/0x50 [amdgpu]
>       drm_gem_object_free+0x1f/0x40 [drm]
>       drm_gem_object_put_unlocked+0x40/0xb0 [drm]
>       drm_gem_object_handle_put_unlocked+0x6c/0xb0 [drm]
>       drm_gem_object_release_handle+0x51/0x90 [drm]
>       drm_gem_handle_delete+0x5e/0x90 [drm]
>       ? drm_gem_handle_create+0x40/0x40 [drm]
>       drm_gem_close_ioctl+0x20/0x30 [drm]
>       drm_ioctl_kernel+0x5d/0xb0 [drm]
>       drm_ioctl+0x2f7/0x3b0 [drm]
>       ? drm_gem_handle_create+0x40/0x40 [drm]
>       ? trace_hardirqs_on_caller+0xf4/0x190
>       ? trace_hardirqs_on+0xd/0x10
>       amdgpu_drm_ioctl+0x4f/0x90 [amdgpu]
>       do_vfs_ioctl+0x93/0x670
>       ? __fget+0x108/0x1f0
>       SyS_ioctl+0x79/0x90
>       entry_SYSCALL_64_fastpath+0x23/0xc2
>
> I've also added the relevant fixes for the issue mentioned above.
>
> Christian König (3):
>    drm/ttm: fix ttm_bo_cleanup_refs_or_queue once more
>    dma-buf: make reservation_object_copy_fences rcu save
>    drm/amdgpu: reserve root PD while releasing it
>
> Michel Dänzer (1):
>    drm/ttm: Always and only destroy bo->ttm_resv in ttm_bo_release_list
>
>   drivers/dma-buf/reservation.c          | 56 +++++++++++++++++++++++++---------
>   drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 13 ++++++--
>   drivers/gpu/drm/ttm/ttm_bo.c           | 43 +++++++++++++-------------
>   3 files changed, 74 insertions(+), 38 deletions(-)
>
> --
> 2.14.3
>
> _______________________________________________
> amd-gfx mailing list
> amd-gfx@lists.freedesktop.org
> https://lists.freedesktop.org/mailman/listinfo/amd-gfx

Re: [PATCH 0/4] Backported amdgpu ttm deadlock fixes for 4.14

[Greg KH]

On Thu, Nov 30, 2017 at 07:23:02PM -0500, Lyude Paul wrote:
> I haven't gone to see where it started, but as of late a good number of
> pretty nasty deadlock issues have appeared with the kernel. Easy
> reproduction recipe on a laptop with i915/amdgpu prime with lockdep enabled:
> 
> DRI_PRIME=1 glxinfo
> 
> Additionally, some more race conditions exist that I've managed to
> trigger with piglit and lockdep enabled after applying these patches:
> 
>     =============================
>     WARNING: suspicious RCU usage
>     4.14.3Lyude-Test+ #2 Not tainted
>     -----------------------------
>     ./include/linux/reservation.h:216 suspicious rcu_dereference_protected() usage!
> 
>     other info that might help us debug this:
> 
>     rcu_scheduler_active = 2, debug_locks = 1
>     1 lock held by ext_image_dma_b/27451:
>      #0:  (reservation_ww_class_mutex){+.+.}, at: [<ffffffffa034f2ff>] ttm_bo_unref+0x9f/0x3c0 [ttm]
> 
>     stack backtrace:
>     CPU: 0 PID: 27451 Comm: ext_image_dma_b Not tainted 4.14.3Lyude-Test+ #2
>     Hardware name: HP HP ZBook 15 G4/8275, BIOS P70 Ver. 01.02 06/09/2017
>     Call Trace:
>      dump_stack+0x8e/0xce
>      lockdep_rcu_suspicious+0xc5/0x100
>      reservation_object_copy_fences+0x292/0x2b0
>      ? ttm_bo_unref+0x9f/0x3c0 [ttm]
>      ttm_bo_unref+0xbd/0x3c0 [ttm]
>      amdgpu_bo_unref+0x2a/0x50 [amdgpu]
>      amdgpu_gem_object_free+0x4b/0x50 [amdgpu]
>      drm_gem_object_free+0x1f/0x40 [drm]
>      drm_gem_object_put_unlocked+0x40/0xb0 [drm]
>      drm_gem_object_handle_put_unlocked+0x6c/0xb0 [drm]
>      drm_gem_object_release_handle+0x51/0x90 [drm]
>      drm_gem_handle_delete+0x5e/0x90 [drm]
>      ? drm_gem_handle_create+0x40/0x40 [drm]
>      drm_gem_close_ioctl+0x20/0x30 [drm]
>      drm_ioctl_kernel+0x5d/0xb0 [drm]
>      drm_ioctl+0x2f7/0x3b0 [drm]
>      ? drm_gem_handle_create+0x40/0x40 [drm]
>      ? trace_hardirqs_on_caller+0xf4/0x190
>      ? trace_hardirqs_on+0xd/0x10
>      amdgpu_drm_ioctl+0x4f/0x90 [amdgpu]
>      do_vfs_ioctl+0x93/0x670
>      ? __fget+0x108/0x1f0
>      SyS_ioctl+0x79/0x90
>      entry_SYSCALL_64_fastpath+0x23/0xc2
> 
> I've also added the relevant fixes for the issue mentioned above.
> 
> Christian König (3):
>   drm/ttm: fix ttm_bo_cleanup_refs_or_queue once more
>   dma-buf: make reservation_object_copy_fences rcu save
>   drm/amdgpu: reserve root PD while releasing it
> 
> Michel Dänzer (1):
>   drm/ttm: Always and only destroy bo->ttm_resv in ttm_bo_release_list

All now queued up, thanks.

greg k-h