From 22a36e660d014925114feb09a2680bb3c2d1e279 Mon Sep 17 00:00:00 2001 From: Vitaly Prosyak Date: Thu, 6 Nov 2025 12:35:53 -0500 Subject: drm/amdgpu: disable peer-to-peer access for DCC-enabled GC12 VRAM surfaces MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Certain multi-GPU configurations (especially GFX12) may hit data corruption when a DCC-compressed VRAM surface is shared across GPUs using peer-to-peer (P2P) DMA transfers. Such surfaces rely on device-local metadata and cannot be safely accessed through a remote GPU’s page tables. Attempting to import a DCC-enabled surface through P2P leads to incorrect rendering or GPU faults. This change disables P2P for DCC-enabled VRAM buffers that are contiguous and allocated on GFX12+ hardware. In these cases, the importer falls back to the standard system-memory path, avoiding invalid access to compressed surfaces. Future work could consider optional migration (VRAM→System→VRAM) if a performance regression is observed when `attach->peer2peer = false`. Tested on: - Dual RX 9700 XT (Navi4x) setup - GNOME and Wayland compositor scenarios - Confirmed no corruption after disabling P2P under these conditions v2: Remove check TTM_PL_VRAM & TTM_PL_FLAG_CONTIGUOUS. v3: simplify for upsteam and fix ip version check (Alex) Suggested-by: Christian König Signed-off-by: Vitaly Prosyak Reviewed-by: Christian König Signed-off-by: Alex Deucher (cherry picked from commit 9dff2bb709e6fbd97e263fd12bf12802d2b5a0cf) Cc: stable@vger.kernel.org --- drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c | 12 ++++++++++++ 1 file changed, 12 insertions(+) (limited to 'drivers/gpu') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c index 8561ad7f6180..ed3bef1edfe4 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_dma_buf.c @@ -82,6 +82,18 @@ static int amdgpu_dma_buf_attach(struct dma_buf *dmabuf, struct amdgpu_bo *bo = gem_to_amdgpu_bo(obj); struct amdgpu_device *adev = amdgpu_ttm_adev(bo->tbo.bdev); + /* + * Disable peer-to-peer access for DCC-enabled VRAM surfaces on GFX12+. + * Such buffers cannot be safely accessed over P2P due to device-local + * compression metadata. Fallback to system-memory path instead. + * Device supports GFX12 (GC 12.x or newer) + * BO was created with the AMDGPU_GEM_CREATE_GFX12_DCC flag + * + */ + if (amdgpu_ip_version(adev, GC_HWIP, 0) >= IP_VERSION(12, 0, 0) && + bo->flags & AMDGPU_GEM_CREATE_GFX12_DCC) + attach->peer2peer = false; + if (!amdgpu_dmabuf_is_xgmi_accessible(attach_adev, bo) && pci_p2pdma_distance(adev->pdev, attach->dev, false) < 0) attach->peer2peer = false; -- cgit v1.2.3 From 9f8fd538e244b87e4556833da51ddd986f50cc81 Mon Sep 17 00:00:00 2001 From: Pierre-Eric Pelloux-Prayer Date: Tue, 4 Nov 2025 10:42:45 +0100 Subject: drm/amdgpu: jump to the correct label on failure MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit drm_sched_entity_init wasn't called yet, so the only thing to do is to release allocated memory. This doesn't fix any bug since entity is zero allocated and drm_sched_entity_fini does nothing in this case. Signed-off-by: Pierre-Eric Pelloux-Prayer Reviewed-by: Tvrtko Ursulin Acked-by: Christian König Signed-off-by: Alex Deucher (cherry picked from commit ec49374ccb8da86b465beaf09c367f3dfd648d8f) --- drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) (limited to 'drivers/gpu') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c index f5d5c45ddc0d..afedea02188d 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_ctx.c @@ -236,7 +236,7 @@ static int amdgpu_ctx_init_entity(struct amdgpu_ctx *ctx, u32 hw_ip, r = amdgpu_xcp_select_scheds(adev, hw_ip, hw_prio, fpriv, &num_scheds, &scheds); if (r) - goto cleanup_entity; + goto error_free_entity; } /* disable load balance if the hw engine retains context among dependent jobs */ -- cgit v1.2.3 From 6623c5f9fd877868fba133b4ae4dab0052e82dad Mon Sep 17 00:00:00 2001 From: "Jesse.Zhang" Date: Fri, 24 Oct 2025 16:09:25 +0800 Subject: drm/amdgpu: fix lock warning in amdgpu_userq_fence_driver_process MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Fix a potential deadlock caused by inconsistent spinlock usage between interrupt and process contexts in the userq fence driver. The issue occurs when amdgpu_userq_fence_driver_process() is called from both: - Interrupt context: gfx_v11_0_eop_irq() -> amdgpu_userq_fence_driver_process() - Process context: amdgpu_eviction_fence_suspend_worker() -> amdgpu_userq_fence_driver_force_completion() -> amdgpu_userq_fence_driver_process() In interrupt context, the spinlock was acquired without disabling interrupts, leaving it in {IN-HARDIRQ-W} state. When the same lock is acquired in process context, the kernel detects inconsistent locking since the process context acquisition would enable interrupts while holding a lock previously acquired in interrupt context. Kernel log shows: [ 4039.310790] inconsistent {IN-HARDIRQ-W} -> {HARDIRQ-ON-W} usage. [ 4039.310804] kworker/7:2/409 [HC0[0]:SC0[0]:HE1:SE1] takes: [ 4039.310818] ffff9284e1bed000 (&fence_drv->fence_list_lock){?...}-{3:3}, [ 4039.310993] {IN-HARDIRQ-W} state was registered at: [ 4039.311004] lock_acquire+0xc6/0x300 [ 4039.311018] _raw_spin_lock+0x39/0x80 [ 4039.311031] amdgpu_userq_fence_driver_process.part.0+0x30/0x180 [amdgpu] [ 4039.311146] amdgpu_userq_fence_driver_process+0x17/0x30 [amdgpu] [ 4039.311257] gfx_v11_0_eop_irq+0x132/0x170 [amdgpu] Fix by using spin_lock_irqsave()/spin_unlock_irqrestore() to properly manage interrupt state regardless of calling context. Reviewed-by: Christian König Signed-off-by: Jesse Zhang Signed-off-by: Alex Deucher (cherry picked from commit ded3ad780cf97a04927773c4600823b84f7f3cc2) Cc: stable@vger.kernel.org --- drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) (limited to 'drivers/gpu') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c index 761bad98da3e..4d0096d0baa9 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_userq_fence.c @@ -151,15 +151,16 @@ void amdgpu_userq_fence_driver_process(struct amdgpu_userq_fence_driver *fence_d { struct amdgpu_userq_fence *userq_fence, *tmp; struct dma_fence *fence; + unsigned long flags; u64 rptr; int i; if (!fence_drv) return; + spin_lock_irqsave(&fence_drv->fence_list_lock, flags); rptr = amdgpu_userq_fence_read(fence_drv); - spin_lock(&fence_drv->fence_list_lock); list_for_each_entry_safe(userq_fence, tmp, &fence_drv->fences, link) { fence = &userq_fence->base; @@ -174,7 +175,7 @@ void amdgpu_userq_fence_driver_process(struct amdgpu_userq_fence_driver *fence_d list_del(&userq_fence->link); dma_fence_put(fence); } - spin_unlock(&fence_drv->fence_list_lock); + spin_unlock_irqrestore(&fence_drv->fence_list_lock, flags); } void amdgpu_userq_fence_driver_destroy(struct kref *ref) -- cgit v1.2.3 From 33c995709121a3a29d4567a08c943bf7a5b24b78 Mon Sep 17 00:00:00 2001 From: Ivan Lipski Date: Thu, 23 Oct 2025 10:03:59 -0400 Subject: drm/amd/display: Allow VRR params change if unsynced with the stream MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit [Why] When changing resolution (e.g., 4K → FHD) in mirror/clone mode with certain monitors, the monitor blanks and loses connection due to an early exit in vrr_settings_require_update(). The function only checks if VRR state, fixed refresh target, or min/max refresh rate range has changed. During mode changes, if the calculated min/max refresh values remain the same even though the stream's v_total changed, the function returns early without updating vrr_params.adjust.v_total_min/max, leaving the monitor's VRR timing parameters unsynced with the new mode, causing it to blank out. [How] Explicitly adjust VRR parameters to the stream's nominal v_total when VRR is supported, but inactive. Fixes: 6d31602a9f57 ("drm/amd/display: more liberal vmin/vmax update for freesync") Reviewed-by: Aurabindo Pillai Signed-off-by: Ivan Lipski Signed-off-by: Fangzhi Zuo Tested-by: Dan Wheeler Signed-off-by: Alex Deucher (cherry picked from commit 607df8248a011524211ee34850345305a1913f9e) --- drivers/gpu/drm/amd/display/modules/freesync/freesync.c | 11 +++++++++++ 1 file changed, 11 insertions(+) (limited to 'drivers/gpu') diff --git a/drivers/gpu/drm/amd/display/modules/freesync/freesync.c b/drivers/gpu/drm/amd/display/modules/freesync/freesync.c index ce421bcddcb0..1aae46d703ba 100644 --- a/drivers/gpu/drm/amd/display/modules/freesync/freesync.c +++ b/drivers/gpu/drm/amd/display/modules/freesync/freesync.c @@ -1260,6 +1260,17 @@ void mod_freesync_handle_v_update(struct mod_freesync *mod_freesync, update_v_total_for_static_ramp( core_freesync, stream, in_out_vrr); } + + /* + * If VRR is inactive, set vtotal min and max to nominal vtotal + */ + if (in_out_vrr->state == VRR_STATE_INACTIVE) { + in_out_vrr->adjust.v_total_min = + mod_freesync_calc_v_total_from_refresh(stream, + in_out_vrr->max_refresh_in_uhz); + in_out_vrr->adjust.v_total_max = in_out_vrr->adjust.v_total_min; + return; + } } unsigned long long mod_freesync_calc_nominal_field_rate( -- cgit v1.2.3 From 7132f7e025f9382157543dd86a62d161335b48b9 Mon Sep 17 00:00:00 2001 From: Sultan Alsawaf Date: Fri, 7 Nov 2025 13:07:13 -0500 Subject: drm/amd/amdgpu: Ensure isp_kernel_buffer_alloc() creates a new BO When the BO pointer provided to amdgpu_bo_create_kernel() points to non-NULL, amdgpu_bo_create_kernel() takes it as a hint to pin that address rather than allocate a new BO. This functionality is never desired for allocating ISP buffers. A new BO should always be created when isp_kernel_buffer_alloc() is called, per the description for isp_kernel_buffer_alloc(). Ensure this by zeroing *bo right before the amdgpu_bo_create_kernel() call. Fixes: 55d42f616976 ("drm/amd/amdgpu: Add helper functions for isp buffers") Reviewed-by: Mario Limonciello (AMD) Reviewed-by: Pratap Nirujogi Signed-off-by: Sultan Alsawaf Signed-off-by: Alex Deucher (cherry picked from commit 73c8c29baac7f0c7e703d92eba009008cbb5228e) --- drivers/gpu/drm/amd/amdgpu/amdgpu_isp.c | 2 ++ 1 file changed, 2 insertions(+) (limited to 'drivers/gpu') diff --git a/drivers/gpu/drm/amd/amdgpu/amdgpu_isp.c b/drivers/gpu/drm/amd/amdgpu/amdgpu_isp.c index 9cddbf50442a..37270c4dab8d 100644 --- a/drivers/gpu/drm/amd/amdgpu/amdgpu_isp.c +++ b/drivers/gpu/drm/amd/amdgpu/amdgpu_isp.c @@ -280,6 +280,8 @@ int isp_kernel_buffer_alloc(struct device *dev, u64 size, if (ret) return ret; + /* Ensure *bo is NULL so a new BO will be created */ + *bo = NULL; ret = amdgpu_bo_create_kernel(adev, size, ISP_MC_ADDR_ALIGN, -- cgit v1.2.3 From bbe3c115030da431c9ec843c18d5583e59482dd2 Mon Sep 17 00:00:00 2001 From: Sathishkumar S Date: Tue, 7 Oct 2025 13:17:51 +0530 Subject: drm/amdgpu/jpeg: Add parse_cs for JPEG5_0_1 enable parse_cs callback for JPEG5_0_1. Signed-off-by: Sathishkumar S Reviewed-by: Leo Liu Signed-off-by: Alex Deucher (cherry picked from commit 547985579932c1de13f57f8bcf62cd9361b9d3d3) Cc: stable@vger.kernel.org --- drivers/gpu/drm/amd/amdgpu/jpeg_v5_0_1.c | 1 + 1 file changed, 1 insertion(+) (limited to 'drivers/gpu') diff --git a/drivers/gpu/drm/amd/amdgpu/jpeg_v5_0_1.c b/drivers/gpu/drm/amd/amdgpu/jpeg_v5_0_1.c index baf097d2e1ac..ab0bf880d3d8 100644 --- a/drivers/gpu/drm/amd/amdgpu/jpeg_v5_0_1.c +++ b/drivers/gpu/drm/amd/amdgpu/jpeg_v5_0_1.c @@ -878,6 +878,7 @@ static const struct amdgpu_ring_funcs jpeg_v5_0_1_dec_ring_vm_funcs = { .get_rptr = jpeg_v5_0_1_dec_ring_get_rptr, .get_wptr = jpeg_v5_0_1_dec_ring_get_wptr, .set_wptr = jpeg_v5_0_1_dec_ring_set_wptr, + .parse_cs = amdgpu_jpeg_dec_parse_cs, .emit_frame_size = SOC15_FLUSH_GPU_TLB_NUM_WREG * 6 + SOC15_FLUSH_GPU_TLB_NUM_REG_WAIT * 8 + -- cgit v1.2.3 From d15deafab5d722afb9e2f83c5edcdef9d9d98bd1 Mon Sep 17 00:00:00 2001 From: Jonathan Kim Date: Thu, 6 Nov 2025 10:17:06 -0500 Subject: drm/amdkfd: relax checks for over allocation of save area Over allocation of save area is not fatal, only under allocation is. ROCm has various components that independently claim authority over save area size. Unless KFD decides to claim single authority, relax size checks. Signed-off-by: Jonathan Kim Reviewed-by: Philip Yang Signed-off-by: Alex Deucher (cherry picked from commit 15bd4958fe38e763bc17b607ba55155254a01f55) Cc: stable@vger.kernel.org --- drivers/gpu/drm/amd/amdkfd/kfd_queue.c | 12 ++++++------ 1 file changed, 6 insertions(+), 6 deletions(-) (limited to 'drivers/gpu') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_queue.c b/drivers/gpu/drm/amd/amdkfd/kfd_queue.c index a65c67cf56ff..f1e7583650c4 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_queue.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_queue.c @@ -297,16 +297,16 @@ int kfd_queue_acquire_buffers(struct kfd_process_device *pdd, struct queue_prope goto out_err_unreserve; } - if (properties->ctx_save_restore_area_size != topo_dev->node_props.cwsr_size) { - pr_debug("queue cwsr size 0x%x not equal to node cwsr size 0x%x\n", + if (properties->ctx_save_restore_area_size < topo_dev->node_props.cwsr_size) { + pr_debug("queue cwsr size 0x%x not sufficient for node cwsr size 0x%x\n", properties->ctx_save_restore_area_size, topo_dev->node_props.cwsr_size); err = -EINVAL; goto out_err_unreserve; } - total_cwsr_size = (topo_dev->node_props.cwsr_size + topo_dev->node_props.debug_memory_size) - * NUM_XCC(pdd->dev->xcc_mask); + total_cwsr_size = (properties->ctx_save_restore_area_size + + topo_dev->node_props.debug_memory_size) * NUM_XCC(pdd->dev->xcc_mask); total_cwsr_size = ALIGN(total_cwsr_size, PAGE_SIZE); err = kfd_queue_buffer_get(vm, (void *)properties->ctx_save_restore_area_address, @@ -352,8 +352,8 @@ int kfd_queue_release_buffers(struct kfd_process_device *pdd, struct queue_prope topo_dev = kfd_topology_device_by_id(pdd->dev->id); if (!topo_dev) return -EINVAL; - total_cwsr_size = (topo_dev->node_props.cwsr_size + topo_dev->node_props.debug_memory_size) - * NUM_XCC(pdd->dev->xcc_mask); + total_cwsr_size = (properties->ctx_save_restore_area_size + + topo_dev->node_props.debug_memory_size) * NUM_XCC(pdd->dev->xcc_mask); total_cwsr_size = ALIGN(total_cwsr_size, PAGE_SIZE); kfd_queue_buffer_svm_put(pdd, properties->ctx_save_restore_area_address, total_cwsr_size); -- cgit v1.2.3 From eac32ff42393efa6657efc821231b8d802c1d485 Mon Sep 17 00:00:00 2001 From: Harish Kasiviswanathan Date: Tue, 28 Oct 2025 14:37:07 -0400 Subject: drm/amdkfd: Fix GPU mappings for APU after prefetch Fix the following corner case:- Consider a 2M huge page SVM allocation, followed by prefetch call for the first 4K page. The whole range is initially mapped with single PTE. After the prefetch, this range gets split to first page + rest of the pages. Currently, the first page mapping is not updated on MI300A (APU) since page hasn't migrated. However, after range split PTE mapping it not valid. Fix this by forcing page table update for the whole range when prefetch is called. Calling prefetch on APU doesn't improve performance. If all it deteriotes. However, functionality has to be supported. v2: Use apu_prefer_gtt as this issue doesn't apply to APUs with carveout VRAM v3: Simplify by setting the flag for all ASICs as it doesn't affect dGPU v4: Remove v2 and v3 changes. Force update_mapping when range is split at a size that is not aligned to prange granularity Suggested-by: Philip Yang Signed-off-by: Harish Kasiviswanathan Reviewed-by: Philip Yang Reviewed-by: Felix Kuehling Signed-off-by: Alex Deucher (cherry picked from commit 076470b9f6f8d9c7c8ca73a9f054942a686f9ba7) --- drivers/gpu/drm/amd/amdkfd/kfd_svm.c | 2 ++ 1 file changed, 2 insertions(+) (limited to 'drivers/gpu') diff --git a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c index 9d72411c3379..74a1d3e1d52b 100644 --- a/drivers/gpu/drm/amd/amdkfd/kfd_svm.c +++ b/drivers/gpu/drm/amd/amdkfd/kfd_svm.c @@ -3687,6 +3687,8 @@ svm_range_set_attr(struct kfd_process *p, struct mm_struct *mm, svm_range_apply_attrs(p, prange, nattr, attrs, &update_mapping); /* TODO: unmap ranges from GPU that lost access */ } + update_mapping |= !p->xnack_enabled && !list_empty(&remap_list); + list_for_each_entry_safe(prange, next, &remove_list, update_list) { pr_debug("unlink old 0x%p prange 0x%p [0x%lx 0x%lx]\n", prange->svms, prange, prange->start, -- cgit v1.2.3