Re: GPU hang trying to run OpenCL kernels on x86_64
Hi Slava, The two x86_64 systems I tried are: - System One Tyan S7025 with dual Xeon X5675 and 48GB registered ECC memory, with a NVIDIA GTX 1050Ti 4GB(also used for display) and an AMD RX 550 4GB Running standard Ubuntu 16.04.4 with kernels linux-image-4.13.0-38-generic and linux-image-4.4.0-122-generic, mesa-17.2.8-0ubuntu0, libdrm-2.4.83-1 and amdgpu-pro 17.50/amdgpu-pro 18.10 lsb_release -a Description: Ubuntu 16.04.4 LTS BIOS configuration: ACPI enabled v3.0 ACPI APIC support Enabled ACPI SRAT table Enabled SR-IOV Enabled Intel VT-d Disabled PCI MMIO 64 Bits support Disabled - System Two Tyan S7002 with dual Xeon X5670 and 12GB registered ECC memory, with an AMD RX 480 Running Ubuntu 18.04 with kernels vanilla 4.16.7 and linux-image-4.15.0-20-generic, mesa-18.0.0~rc5-1ubuntu1, libdrm-2.4.91-2 and mesa-opencl-icd, libclc-0.2.0+git20180312-1 BIOS configuration: ACPI enabled v2.0 ACPI APIC support Enabled ACPI SRAT table Enabled SR-IOV Enabled Intel VT-d Disabled PCI MMIO 64 Bits support Disabled amdgpu-pro-install --headless --opencl=legacy When I try to run the attached openCL code (which computes a cross-correlation between two square matrices directly by cross correlation function definition), the GPU hangs, but there are also other kernels where this also happens. As soon as I try to run the kernel the system hangs at the first kernel computation on all the two systems, and after a couple of seconds dmesg shows: [drm:amdgpu_job_timedout [amdgpu]] *ERROR* ring gfx timeout, last signaled seq=2, last emitted seq=3 [drm] IP block:gmc_v8_0 is hung! [drm] IP block:tonga_ih is hung! [drm] IP block:gfx_v8_0 is hung! [drm] IP block:sdma_v3_0 is hung! [drm] IP block:uvd_v6_0 is hung! [drm] IP block:vce_v3_0 is hung! [drm] GPU recovery disabled. - On another system with armhf 32 bits, 1GB ram, 512GB SSD, AMD RX 480 or AMD RX 550 with Ubuntu 17.10, vanilla kernel 4.16.7, mesa-18.0.2, libdrm-2.4.92-git, libclc-git at commit 3d994f2ff2cbb4531223fe2657144cb19f0c5328 (15/Nov/2017) The kernels work properly on the same AMD cards. On Fri, May 4, 2018 at 7:18 PM, Abramov, Slavawrote: > Luis, > > > Can you please provide more details on your system environment and steps on > configuring the software and reproducing the issue? > > > > Slava A > > > From: amd-gfx on behalf of Luís > Mendes > Sent: Friday, May 4, 2018 12:27:47 PM > To: amd-gfx list; Koenig, Christian; Michel Dänzer > Subject: GPU hang trying to run OpenCL kernels on x86_64 > > Hi, > > I am a collaborator with Syncleus/aparapi project on github and I've > been testing OpenCL on AMD and NVIDIA cards. > > Currently I have a set of kernels that hang the GPU (AMD RX 460 and > AMD RX 550) across all compute units on x86_64 running vanilla kernel > 4.16.7 on Ubuntu 18.04, also on Ubuntu 16.04.4 with AMDGPU PRO 17.50 > and 18.10 show the same problems, in fact, AMDGPU-PRO 18.10 is even > worse. > > However the same set of kernels run happily on armhf with vanilla > Linux 4.16.7 and mesa 18.0 (mesa-opencl-icd and libclc for amdgcn), > Ubuntu 17.10, on an AMD RX460 and an AMD RX 550. > > Luís Mendes > ___ > amd-gfx mailing list > amd-gfx@lists.freedesktop.org > https://lists.freedesktop.org/mailman/listinfo/amd-gfx May 04, 2018 10:37:41 PM com.aparapi.internal.kernel.KernelRunner executeInternalInner INFO: typedef struct This_s{ __global int *tilesGeometry; __global int *inputGeometry; __global int *threadOutputStart; __global int *outputGeometry; __global int *threadOffsetI; __global int *threadOffsetJ; __global float *matrixInF; __global float *matrixInG; __global float *matrixOut; int passid; }This; int get_pass_id(This *this){ return this->passid; } short pt_ist_ceris_vipivist_jobs_CrossCorrelationKernel__signX(This *this, short x){ short value = (short)((x + x) + 1); return((short)(value / abs(value))); } short pt_ist_ceris_vipivist_jobs_CrossCorrelationKernel__relocateX(This *this, short x, short dimX){ short result = (short)(((pt_ist_ceris_vipivist_jobs_CrossCorrelationKernel__signX(this, x) + 1) * (x + 1)) / 2); result = (short)(((pt_ist_ceris_vipivist_jobs_CrossCorrelationKernel__signX(this, (short)(dimX - result)) + 1) * result) / 2); return(result); } __kernel void run( __global int *tilesGeometry, __global int *inputGeometry, __global int *threadOutputStart, __global int *outputGeometry, __global int *threadOffsetI, __global int *threadOffsetJ, __global float *matrixInF, __global float *matrixInG, __global float *matrixOut, int passid ){ This thisStruct; This* this= this->tilesGeometry = tilesGeometry; this->inputGeometry = inputGeometry; this->threadOutputStart = threadOutputStart; this->outputGeometry =
Re: GPU hang trying to run OpenCL kernels on x86_64
Luis, Can you please provide more details on your system environment and steps on configuring the software and reproducing the issue? Slava A From: amd-gfxon behalf of Luís Mendes Sent: Friday, May 4, 2018 12:27:47 PM To: amd-gfx list; Koenig, Christian; Michel Dänzer Subject: GPU hang trying to run OpenCL kernels on x86_64 Hi, I am a collaborator with Syncleus/aparapi project on github and I've been testing OpenCL on AMD and NVIDIA cards. Currently I have a set of kernels that hang the GPU (AMD RX 460 and AMD RX 550) across all compute units on x86_64 running vanilla kernel 4.16.7 on Ubuntu 18.04, also on Ubuntu 16.04.4 with AMDGPU PRO 17.50 and 18.10 show the same problems, in fact, AMDGPU-PRO 18.10 is even worse. However the same set of kernels run happily on armhf with vanilla Linux 4.16.7 and mesa 18.0 (mesa-opencl-icd and libclc for amdgcn), Ubuntu 17.10, on an AMD RX460 and an AMD RX 550. Luís Mendes ___ amd-gfx mailing list amd-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/amd-gfx ___ amd-gfx mailing list amd-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/amd-gfx
GPU hang trying to run OpenCL kernels on x86_64
Hi, I am a collaborator with Syncleus/aparapi project on github and I've been testing OpenCL on AMD and NVIDIA cards. Currently I have a set of kernels that hang the GPU (AMD RX 460 and AMD RX 550) across all compute units on x86_64 running vanilla kernel 4.16.7 on Ubuntu 18.04, also on Ubuntu 16.04.4 with AMDGPU PRO 17.50 and 18.10 show the same problems, in fact, AMDGPU-PRO 18.10 is even worse. However the same set of kernels run happily on armhf with vanilla Linux 4.16.7 and mesa 18.0 (mesa-opencl-icd and libclc for amdgcn), Ubuntu 17.10, on an AMD RX460 and an AMD RX 550. Luís Mendes ___ amd-gfx mailing list amd-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: amdgpu 0000:01:00.0: IH ring buffer overflow (0x00000010, 0x00000000, 0x00000020)
Am 27.04.2018 um 14:02 schrieb Paul Menzel: Dear Linux AMD folks, I get the overrun message below. Sorry for the delayed response, I was on vacation and then on sick leave :( $ more /proc/version Linux version 4.14.30.mx64.211 (r...@holidayincambodia.molgen.mpg.de) (gcc version 7.3.0 (GCC)) #1 SMP Tue Mar 27 12:40:07 CEST 2018 That one is ancient, please try an up to date kernel as well. [90612.637194] amdgpu :01:00.0: IH ring buffer overflow (0x0010, 0x, 0x0020) Should this be reported and fixed? How can this be debugged? (`drm.debug=0xe`?). We do not know how to reproduce it yet. Yes, that should certainly be fixed. The issue is that your hardware is producing so many interrupts that the CPU can't catch up any more processing them. That either sounds like a hardware problem or something is blocking the CPU for quite a long time with interrupts disabled. Regards, Christian. Kind regards, Paul ___ amd-gfx mailing list amd-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/amd-gfx ___ amd-gfx mailing list amd-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH] drm/amdgpu: invalidate parent bo when shadow bo was invalidated
Am 04.05.2018 um 08:44 schrieb Chunming Zhou: Shadow BO is located on GTT and its parent (PT and PD) BO could located on VRAM. In some case, the BO on GTT could be evicted but the parent did not. This may cause the shadow BO not be put in the evict list and could not be invalidate correctly. v2: suggested by Christian Change-Id: Iad10d9a3031fa2b243879b9e58ee4d8c527eb433 Signed-off-by: Chunming ZhouReported-by: Shaoyun Liu Reviewed-by: Christian König --- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 4 1 file changed, 4 insertions(+) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c index 71dcdefce255..8e71d3984016 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c @@ -2252,6 +2252,10 @@ void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev, { struct amdgpu_vm_bo_base *bo_base; + /* shadow bo doesn't have bo base, its validation needs its parent */ + if (bo->parent && bo->parent->shadow == bo) + bo = bo->parent; + list_for_each_entry(bo_base, >va, bo_list) { struct amdgpu_vm *vm = bo_base->vm; ___ amd-gfx mailing list amd-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH] drm/amd/display: remove need of modeset flag for overlay planes
On 05/03/2018 02:11 PM, Harry Wentland wrote: On 2018-05-03 04:00 AM, S, Shirish wrote: On 5/2/2018 7:21 PM, Harry Wentland wrote: On 2018-04-27 06:27 AM, Shirish S wrote: This patch is in continuation to the "843e3c7 drm/amd/display: defer modeset check in dm_update_planes_state" where we started to eliminate the dependency on DRM_MODE_ATOMIC_ALLOW_MODESET to be set by the user space, which as such is not mandatory. After deferring, this patch eliminates the dependency on the flag for overlay planes. Apologies for the late response. I had to think about this patch for a long time since I'm not quite comfortable with it. This has to be done in stages as its a pretty complex and requires thorough testing before we free primary planes as well from dependency on modeset flag. Signed-off-by: Shirish S--- drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c | 8 +--- 1 file changed, 5 insertions(+), 3 deletions(-) diff --git a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c index 1a63c04..87b661d 100644 --- a/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c +++ b/drivers/gpu/drm/amd/display/amdgpu_dm/amdgpu_dm.c @@ -4174,7 +4174,7 @@ static void amdgpu_dm_commit_planes(struct drm_atomic_state *state, } spin_unlock_irqrestore(>dev->event_lock, flags); - if (!pflip_needed) { + if (!pflip_needed || plane->type == DRM_PLANE_TYPE_OVERLAY) { Does this mean that whenever we have an overlay plane we won't do amdgpu_dm_do_flip but commit_planes_to_stream instead? Is this really the behavior we want? commit_planes_to_stream was intended to program a new surface on a modeset whereas amdgpu_dm_do_flip was intended for pageflips. Need of "modeset" flag to program new surface is what we want to fix in this patch for underlay plane and in next stages, fix manifestations caused by this approach as and when seen. Since the user space doesn't send modeset flag for new surface, hence to program it, this patch checks the plane type to construct planes_count before calling commit_planes_to_stream(). Looking at the allow_modeset flag was never quite right and we anticipated having to rework this when having to deal with things like multiple planes. What really has to happen is that we determine the surface_update_type in atomic_check and then use that in atomic_commit to either program the surface only (UPDATE_TYPE_FAST) without having to lock all pipes or to lock all pipes (see lock_and_validation_needed in amdgpu_dm_atomic_check) if we need to reprogram mode (UPDATE_TYPE_FULL). I don't remember exactly what UPDATE_TYPE_MED is used for. I don't feel comfortable taking a shortcut for DRM_PLANE_TYPE_OVERLAY without first having a plan and patches for how to deal with the above-mentioned. Bhawan and Andrey had a look at this before but it was never quite ready. The work was non-trivial and potentially impacts lots of configurations and scenarios if we don't get it right. If you're curious you can look at this change (apologies to everyone else for posting AMD-internal link): http://git.amd.com:8080/#/c/103931/11 If we use commit_planes_to_stream we end up losing things like the immediate_flip flag, as well as the wait for the right moment to program the flip that amdgpu_dm_do_flip does. From the code, amdgpu_dm_do_flip does what you mentioned only for primary plane and hence either way its not set for underlay. The code wasn't designed with underlay in mind, so it will need work. Harry I support Harry's comments, we definitely need to strive to remove dependency on page_fleep needed flag, AFAIK we are the only ATOMIC KMS driver which makes a distinction between page fleep and other surface updates, but it's better to sit and create a general plan of how to address it for all type of planes instead of patching for overlay only. Andrey Regards, Shirish S Even more importantly we won't wait for fences (reservation_object_wait_timeout_rcu). Harry WARN_ON(!dm_new_plane_state->dc_state); plane_states_constructed[planes_count] = dm_new_plane_state->dc_state; @@ -4884,7 +4884,8 @@ static int dm_update_planes_state(struct dc *dc, /* Remove any changed/removed planes */ if (!enable) { - if (pflip_needed) + if (pflip_needed && + plane && plane->type != DRM_PLANE_TYPE_OVERLAY) continue; if (!old_plane_crtc) @@ -4931,7 +4932,8 @@ static int dm_update_planes_state(struct dc *dc, if (!dm_new_crtc_state->stream) continue; - if (pflip_needed) + if (pflip_needed && + plane && plane->type != DRM_PLANE_TYPE_OVERLAY) continue; WARN_ON(dm_new_plane_state->dc_state); ___ amd-gfx
Re: [Linaro-mm-sig] [PATCH 4/8] dma-buf: add peer2peer flag
Am Mittwoch, den 25.04.2018, 13:44 -0400 schrieb Alex Deucher: > On Wed, Apr 25, 2018 at 2:41 AM, Christoph Hellwig> wrote: > > On Wed, Apr 25, 2018 at 02:24:36AM -0400, Alex Deucher wrote: > > > > It has a non-coherent transaction mode (which the chipset can opt to > > > > not implement and still flush), to make sure the AGP horror show > > > > doesn't happen again and GPU folks are happy with PCIe. That's at > > > > least my understanding from digging around in amd the last time we had > > > > coherency issues between intel and amd gpus. GPUs have some bits > > > > somewhere (in the pagetables, or in the buffer object description > > > > table created by userspace) to control that stuff. > > > > > > Right. We have a bit in the GPU page table entries that determines > > > whether we snoop the CPU's cache or not. > > > > I can see how that works with the GPU on the same SOC or SOC set as the > > CPU. But how is that going to work for a GPU that is a plain old PCIe > > card? The cache snooping in that case is happening in the PCIe root > > complex. > > I'm not a pci expert, but as far as I know, the device sends either a > snooped or non-snooped transaction on the bus. I think the > transaction descriptor supports a no snoop attribute. Our GPUs have > supported this feature for probably 20 years if not more, going back > to PCI. Using non-snooped transactions have lower latency and faster > throughput compared to snooped transactions. PCI-X (as in the thing which as all the rage before PCIe) added a attribute phase to each transaction where the requestor can enable relaxed ordering and/or no-snoop on a transaction basis. As those are strictly performance optimizations the root-complex isn't required to honor those attributes, but implementations that care about performance usually will. Regards, Lucas ___ amd-gfx mailing list amd-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH] drm/amdgpu: invalidate parent bo when shadow bo was invalidated
On 05/04/2018 02:44 PM, Chunming Zhou wrote: Shadow BO is located on GTT and its parent (PT and PD) BO could located on VRAM. In some case, the BO on GTT could be evicted but the parent did not. This may cause the shadow BO not be put in the evict list and could not be invalidate correctly. v2: suggested by Christian Change-Id: Iad10d9a3031fa2b243879b9e58ee4d8c527eb433 Signed-off-by: Chunming ZhouReported-by: Shaoyun Liu Reviewed-by: Junwei Zhang --- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 4 1 file changed, 4 insertions(+) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c index 71dcdefce255..8e71d3984016 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c @@ -2252,6 +2252,10 @@ void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev, { struct amdgpu_vm_bo_base *bo_base; + /* shadow bo doesn't have bo base, its validation needs its parent */ + if (bo->parent && bo->parent->shadow == bo) + bo = bo->parent; + list_for_each_entry(bo_base, >va, bo_list) { struct amdgpu_vm *vm = bo_base->vm; ___ amd-gfx mailing list amd-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/amd-gfx
[PATCH] drm/amdgpu: invalidate parent bo when shadow bo was invalidated
Shadow BO is located on GTT and its parent (PT and PD) BO could located on VRAM. In some case, the BO on GTT could be evicted but the parent did not. This may cause the shadow BO not be put in the evict list and could not be invalidate correctly. v2: suggested by Christian Change-Id: Iad10d9a3031fa2b243879b9e58ee4d8c527eb433 Signed-off-by: Chunming ZhouReported-by: Shaoyun Liu --- drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 4 1 file changed, 4 insertions(+) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c index 71dcdefce255..8e71d3984016 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c @@ -2252,6 +2252,10 @@ void amdgpu_vm_bo_invalidate(struct amdgpu_device *adev, { struct amdgpu_vm_bo_base *bo_base; + /* shadow bo doesn't have bo base, its validation needs its parent */ + if (bo->parent && bo->parent->shadow == bo) + bo = bo->parent; + list_for_each_entry(bo_base, >va, bo_list) { struct amdgpu_vm *vm = bo_base->vm; -- 2.14.1 ___ amd-gfx mailing list amd-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/amd-gfx
Re: [PATCH 2/2] drm/amdgpu: link shadow bo as well
On 2018年05月03日 18:14, Christian König wrote: Am 24.04.2018 um 09:35 schrieb Chunming Zhou: Shadow BO is located on GTT and its parent (PT and PD) BO could located on VRAM. In some case, the BO on GTT could be evicted but the parent did not. This may cause the shadow BO not be put in the evict list and could not be invalidate correctly. Change-Id: Iad10d9a3031fa2b243879b9e58ee4d8c527eb433 Signed-off-by: Chunming ZhouReported-by: Shaoyun Liu Way to much memory usage and to complicated. What we need to do is just to handle the real BO as evicted when the shadow BO is evicted. Something like the following at the start of amdgpu_vm_bo_invalidate() should be sufficient: if (bo->parent && bo->parent->shadow == bo) bo = bo->parent; clever, will send a patch soon. Regards, David Zhou Regards, Christian. --- drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c | 5 - drivers/gpu/drm/amd/amdgpu/amdgpu_object.h | 1 + drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c | 11 +++ drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h | 1 + 4 files changed, 13 insertions(+), 5 deletions(-) diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c index e1756b68a17b..9c9f6fc5c994 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_cs.c @@ -462,11 +462,6 @@ static int amdgpu_cs_validate(void *param, struct amdgpu_bo *bo) do { r = amdgpu_cs_bo_validate(p, bo); } while (r == -ENOMEM && amdgpu_cs_try_evict(p, bo)); - if (r) - return r; - - if (bo->shadow) - r = amdgpu_cs_bo_validate(p, bo->shadow); return r; } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h index 540e03fa159f..8078da36aec7 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_object.h @@ -58,6 +58,7 @@ struct amdgpu_bo_va_mapping { /* User space allocated BO in a VM */ struct amdgpu_bo_va { struct amdgpu_vm_bo_base base; + struct amdgpu_vm_bo_base shadow_base; /* protected by bo being reserved */ unsigned ref_count; diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c index c75b96433ee7..026147cc5104 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c @@ -450,8 +450,13 @@ static int amdgpu_vm_alloc_levels(struct amdgpu_device *adev, pt->parent = amdgpu_bo_ref(parent->base.bo); amdgpu_vm_bo_base_init(>base, vm, pt); + amdgpu_vm_bo_base_init(>shadow_base, vm, + pt->shadow); spin_lock(>status_lock); list_move(>base.vm_status, >relocated); + if (pt->shadow) + list_move(>shadow_base.vm_status, + >relocated); spin_unlock(>status_lock); } @@ -1875,6 +1880,7 @@ struct amdgpu_bo_va *amdgpu_vm_bo_add(struct amdgpu_device *adev, return NULL; } amdgpu_vm_bo_base_init(_va->base, vm, bo); + amdgpu_vm_bo_base_init(_va->shadow_base, vm, bo ? bo->shadow : NULL); bo_va->ref_count = 1; INIT_LIST_HEAD(_va->valids); @@ -2220,9 +2226,11 @@ void amdgpu_vm_bo_rmv(struct amdgpu_device *adev, struct amdgpu_vm *vm = bo_va->base.vm; list_del(_va->base.bo_list); + list_del(_va->shadow_base.bo_list); spin_lock(>status_lock); list_del(_va->base.vm_status); + list_del(_va->shadow_base.vm_status); spin_unlock(>status_lock); list_for_each_entry_safe(mapping, next, _va->valids, list) { @@ -2456,6 +2464,7 @@ int amdgpu_vm_init(struct amdgpu_device *adev, struct amdgpu_vm *vm, goto error_unreserve; amdgpu_vm_bo_base_init(>root.base, vm, root); + amdgpu_vm_bo_base_init(>root.shadow_base, vm, root->shadow); amdgpu_bo_unreserve(vm->root.base.bo); if (pasid) { @@ -2575,6 +2584,8 @@ static void amdgpu_vm_free_levels(struct amdgpu_device *adev, if (parent->base.bo) { list_del(>base.bo_list); list_del(>base.vm_status); + list_del(>shadow_base.bo_list); + list_del(>shadow_base.vm_status); amdgpu_bo_unref(>base.bo->shadow); amdgpu_bo_unref(>base.bo); } diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h index 30f080364c97..f4ae6c6b28b8 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_vm.h @@ -153,6 +153,7 @@ struct amdgpu_vm_bo_base { struct amdgpu_vm_pt { struct amdgpu_vm_bo_base base; + struct amdgpu_vm_bo_base shadow_base; bool huge; /* array of page tables, one for each directory entry */ ___ amd-gfx mailing list amd-gfx@lists.freedesktop.org