Allow using multiple SDMA schedulers only on GPUs where
we are allowed to do concurrent VM flushes.
This consideration is necessary because all GART windows
are mapped in VMID 0 (the kernel VMID) so each buffer
entity would flush VMID 0 concurrently.
Practically this means that we can't use multiple SDMA
engines for TTM on GFX6-8 and Navi 1x.
Fixes: 01c836788b ("drm/amdgpu: pass all the sdma scheds to amdgpu_mman")
Fixes: e4029f7a94 ("drm/amdgpu: only use working sdma schedulers for ttm")
Cc: Pierre-Eric Pelloux-Prayer <pierre-eric.pelloux-prayer@amd.com>
Signed-off-by: Timur Kristóf <timur.kristof@gmail.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit a8171229bc)
amdgpu_acpi_vfct_bios() fetches the VFCT table with acpi_get_table()
but never releases it. acpi_get_table() takes a reference on the
table (incrementing its validation_count and mapping it on the 0->1
transition); without a paired acpi_put_table() the mapping is leaked
on every call, whether or not a matching VBIOS image is found.
Route all exit paths after the table is acquired through a common
acpi_put_table(). The VBIOS image is copied out with kmemdup() before
the table is released, so it remains valid for the caller.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Link: https://patch.msgid.link/20260708193518.702584-3-mario.limonciello@amd.com
Signed-off-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit ca5988682b)
Cc: stable@vger.kernel.org
amdgpu_bo_create_reserved() only allocates a new BO when
*bo_ptr (struct amdgpu_bo **bo_ptr as input parameter) is
NULL, it simply skips creation when *bo_ptr is non-NULL.
But it unconditionally reserves, pins, gart allocates
and maps the BO afterwards.
When the same non-NULL BO pointer is passed in again,
for example firmware buffers that live in adev and are
re-loaded on every resume / cp_resume / start
under AMDGPU_FW_LOAD_DIRECT, amdgpu_bo_pin() just increases
pin_count unconditionally, however the matching teardown only unpins
once, so pin_count never drops to zero, so TTM is not able
to move, swap or evict a BO, causing BO leaks.
This commit fixes this issue by only pinning the bo
once at creation, and repeated calls no longer
take additional pin references.
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit 3ddc0ae762)
Cc: stable@vger.kernel.org
pre_reset only force-completes fences of MAPPED queues. A queue in any
other state (e.g. mid-eviction) keeps its last_fence pending; after a
GPU reset that fence never signals, so the eviction/suspend worker and
process teardown (amdgpu_evf_mgr_flush_suspend) wait on it forever and
wedge the machine:
INFO: task kworker/6:28 blocked for more than 120 seconds.
Workqueue: events amdgpu_eviction_fence_suspend_worker [amdgpu]
Call Trace:
dma_fence_wait_timeout+0x7e/0x130
amdgpu_userq_evict+0x67/0x140 [amdgpu]
amdgpu_eviction_fence_suspend_worker+0xd8/0x160 [amdgpu]
process_scheduled_works+0xa6/0x420
Force-complete every queue's fence regardless of state. The unmap and
mark-hung step stays gated on MAPPED, since unmapping a queue that is
not mapped is invalid.
Fixes: 290f46cf57 ("drm/amdgpu: Implement user queue reset functionality")
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Jesse Zhang <Jesse.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit 9102b39fa9)
Cc: stable@vger.kernel.org
GC 11.7.0 and 11.7.1 should map to AMDGPU_FAMILY_GC_11_5_4 for DCN42.
Fixes: cf591e67c0 ("drm/amdgpu: add support for GC IP version 11.7.0")
Fixes: a928d8d81e ("drm/amdgpu: add support for GC IP version 11.7.1")
Signed-off-by: Roman Li <Roman.Li@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit f8ee6447e7)
drm-misc-fixes for v7.2-rc3:
- Fix uaf in amdxdna mmap failure path.
- A lot of deadlocks, access races and return value fixes in amdxdna.
- Fix analogix_dp bitshifts during link training.
- Use direct label in drm_exec.
- Fix absent indirect bo handling in v3d.
- Sync on first active crtc in fb_dirty, rather than first crtc.
- Rework try_harder in the buddy allocator.
- Make imagination function static to solve compiler warning.
- Fix imagination error checking.
Signed-off-by: Dave Airlie <airlied@redhat.com>
From: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Link: https://patch.msgid.link/71e5b48b-307f-47f5-8fd5-b60ea43e4196@linux.intel.com
The vm pointer returned from amdgpu_vm_get_vm_from_pasid() is only
valid while the lock is still being held. Once xa_unlock_irqrestore is
called and returned, the pointer is no longer under lock and is subject
to modification. Since, the caller still dereferences vm->task_info in
amdgpu_vm_get_task_info_vm() after the lock is removed, this causes a
use after unlock problem.
Remove the lifetime issue present in amdgpu_vm_get_task_info_pasid()
through removing the amdgpu_vm_get_vm_from_pasid() function from
amdgpu_vm.c and making the relevant code inline to hold the lock while
it is still in use.
Signed-off-by: Shahyan Soltani <shahyan.soltani@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit 9d01579f3f)
Cc: stable@vger.kernel.org
Destroying a hung user queue issues a MES REMOVE_QUEUE that times out,
The destroy path only logged the error and freed the queue, so the
next userq submission failed and forced a GPU reset attributed to an innocent workload.
Kick the userq reset work when unmap fails so the GPU is recovered at
destroy time.
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Jesse Zhang <Jesse.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit 8396b9de41)
Cc: stable@vger.kernel.org
JDPG does not support on VCN5
This patch will disable JDPG, because DPG is not correctly
copying the JRBC Read/Write Pointers (R/WPTR) from the PG
(Power Gating) block to JRBC.
Signed-off-by: Suresh Guttula <suresh.guttula@amd.com>
Reviewed-by: Mario Limonciello <mario.limonciello@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit ea3fdd1eda)
jpeg_v4_0_3_is_idle() initializes ret to false and then accumulates ring
idle status using &=. Since false & condition always remains false, the
function can never report the JPEG block as idle.
Initialize ret to true so the function returns true only when all JPEG
rings report RB_JOB_DONE.
Signed-off-by: Boyuan Zhang <boyuan.zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit e9df8e9d04)
Cc: stable@vger.kernel.org
Running RCCL unit tests on a system with a 64K PAGE_SIZE triggers
the following warning and causes the test to terminate on latest
upstream kernel:
WARNING: drivers/gpu/drm/amd/amdgpu/amdgpu_object.c:1335 at
amdgpu_bo_release_notify+0x1bc/0x280 [amdgpu],
CPU#18: rccl-UnitTests/33151
Call trace:
amdgpu_bo_release_notify
ttm_bo_release
amdgpu_gem_object_free
drm_gem_object_free
amdgpu_bo_unref
amdgpu_bo_create
amdgpu_bo_create_user
amdgpu_gem_object_create
amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu
kfd_ioctl_alloc_memory_of_gpu
kfd_ioctl
sys_ioctl
The warning is triggered because
amdgpu_ttm_next_clear_entity() returns NULL when a clear buffer
operation is requested. This happens because the GART window
allocation for the default_entity, clear_entity and move_entity
fails during initialization.
Commit [1] introduced separate GART windows for the
default_entity, clear_entity and move_entity of each SDMA
instance. Their sizes are derived from
AMDGPU_GTT_MAX_TRANSFER_SIZE, which is currently defined as 1024
pages. This implicitly assumes a 4K PAGE_SIZE, where 1024 pages
correspond to a 4MB transfer. On a 64K PAGE_SIZE system, however,
the same value expands to 64MB.
The default_entity and clear_entity each allocate one
AMDGPU_GTT_MAX_TRANSFER_SIZE GART window, while the move_entity
allocates two such windows. This results in 16MB of GART space
per SDMA instance on a 4K PAGE_SIZE system, but 256MB per SDMA
instance on a 64K PAGE_SIZE system.
On an MI210 system with five SDMA instances and a 512MB GART
aperture, the total GART space required becomes 1.25GB,
exceeding the available GART aperture. Consequently, GART window
allocation fails, amdgpu_ttm_next_clear_entity() returns NULL,
and the above warning is triggered.
Redefine AMDGPU_GTT_MAX_TRANSFER_SIZE in bytes instead of page
units. Where a page count is required, convert it using
PAGE_SHIFT. This preserves the existing 4MB transfer size across
all PAGE_SIZE configurations while keeping GART window
allocations within the available GART aperture.
[1] https://lore.kernel.org/all/20260408100327.1372-3-pierre-eric.pelloux-prayer@amd.com/#t
Closes: https://gitlab.freedesktop.org/drm/amd/-/work_items/5435
Fixes: 897ee11ec0 ("drm/amdgpu: create multiple clear/move ttm entities")
Signed-off-by: Donet Tom <donettom@linux.ibm.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit 27213b776a)
Cc: stable@vger.kernel.org
jpeg_v5_0_1_is_idle() initializes ret to false and then accumulates ring
idle status using &=. Since false & condition always remains false, the
function can never report the JPEG block as idle.
Initialize ret to true so the function returns true only when all JPEG
rings report RB_JOB_DONE.
Signed-off-by: Boyuan Zhang <boyuan.zhang@amd.com>
Reviewed-by: David (Ming Qiang) Wu <David.Wu3@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit 680adf5fae)
Cc: stable@vger.kernel.org
This state can be reached via other means than physical moves, like PRT
bindings. Make the name match the actual purpose of the state.
Signed-off-by: Natalie Vock <natalie.vock@gmx.de>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit 1f7a795fb9)
The "moved" VM state is a bit unfortunately named, because BOs can end
up in this state without being physically moved. While we need to
invalidate every mapping when BOs are physically moved, in some other
cases like PRT binds/unbinds there is no need to refresh mappings except
those affected by the bind.
Full invalidation of all BO mappings manifested as severe regressions in
PRT bind performance, which this patch fixes. The offending patch is
4cdbba5a16 ("drm/amdgpu: restructure VM state machine v4") in the
amd-staging-drm-next tree, although it has not yet propagated anywhere
else.
Fixes: 4cdbba5a16 ("drm/amdgpu: restructure VM state machine v4")
Closes: https://gitlab.freedesktop.org/drm/amd/-/work_items/5437
Signed-off-by: Natalie Vock <natalie.vock@gmx.de>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit 0b2fa33b42)
Cc: stable@vger.kernel.org
When ACP soft reset poll times out, original code returns early without cleanup,
leaking MFD child devices, genpd links and all ACP heap allocations.
Replace direct early return with goto out to force run all cleanup logic
regardless of reset success, preserve timeout error code for caller.
Signed-off-by: Ce Sun <cesun102@amd.com>
Reviewed-by: Tao Zhou <tao.zhou1@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit 98073e4328)
Cc: stable@vger.kernel.org
When creating an user-queue, the user space
provides a doorbell BO handle and an offset within
the bo to obtain a doorbell.
However current implementation using xa_store_irq()
to store a doorbell, which allows a later queue created
with the same BO and offset parameters to overwrite an
existing queue and doorbell mapping.
This can cause problems like misrouting fence IRQ
processing to a wrong queue, and mislead the cleanup
process of one queue erasing the mapping of another queue.
This commit fixes this issue by replacing xa_store_irq with
xa_insert_irq, which rejects mapping a reserved
doorbell to a newly created queue
Signed-off-by: Zhu Lingshan <lingshan.zhu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit 6244eae229)
Cc: stable@vger.kernel.org
Try KQ by ring_id first (KCQ and UQ never share a HW slot); fall back
to amdgpu_userq_process_fence_irq() on miss, since KCQ EOPs were
misrouted into the userq fence path when enable_mes is true.
Require a strict (me,pipe,queue) match in the gfx case, then userq gfx
EOPs fall through to amdgpu_userq_process_fence_irq().
Suggested-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Jesse Zhang <jesse.zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit 6c1f4f7ff0)
Cc: stable@vger.kernel.org
Try KQ by ring_id first (KCQ and UQ never share a HW slot); fall back
to amdgpu_userq_process_fence_irq() on miss, since KQ EOPs were
misrouted into the userq fence path when enable_mes is true.
Require a strict (me,pipe,queue) match in the gfx case, then userq gfx
EOPs fall through to amdgpu_userq_process_fence_irq().
Suggested-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Jesse Zhang <jesse.zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit 88e589cc81)
Cc: stable@vger.kernel.org
amdgpu_pci_remove() calls drm_dev_unplug() before invoking the driver
fini routines. This causes drm_dev_enter() in amdgpu_ttm_fini() to
always return false, so iounmap(aper_base_kaddr) never runs on normal
driver unload, leaving an orphaned entry in the x86 PAT interval tree.
On connected_to_cpu hardware, the aperture is mapped write-back (WB) via
ioremap_cache(). On reload, IP discovery calls memremap(..., MEMREMAP_WC)
over the same range. The WC vs WB conflict causes:
ioremap error for 0x..., requested 0x1, got 0x0
amdgpu: discovery failed: -2
Fix by switching to devres-managed mappings so cleanup is guaranteed
regardless of drm_dev_enter() state:
- connected_to_cpu path: devm_memremap(MEMREMAP_WB). For
IORESOURCE_SYSTEM_RAM ranges this takes the try_ram_remap() shortcut,
returning __va(offset) from the existing kernel direct map. No new
ioremap VA or PAT entry is created, so there is nothing to orphan.
- dGPU path: devm_ioremap_wc() registers iounmap() as a devres action,
guaranteeing cleanup at device_del() time.
Also remove iounmap(aper_base_kaddr) from amdgpu_device_unmap_mmio()
since the mapping is now devres-owned.
v2: Remove redundant x86_64 guard (Lijo)
Fixes: 9d0af8b4de ("drm/amdgpu: pre-map device buffer as cached for A+A config")
Signed-off-by: Asad Kamal <asad.kamal@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Lijo Lazar <lijo.lazar@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit d871e99879)
Cc: stable@vger.kernel.org
Fix a security vulnerability where malicious VCE command streams
with oversized dimensions (e.g. 65536×65536) cause 32-bit integer
overflow, wrapping the calculated buffer size to 0. This bypasses
validation and allows GPU firmware to perform out-of-bound memory
access.
The fix uses 64-bit arithmetic to detect overflow and rejects
invalid dimensions before they reach the hardware.
V2: remove redundant check
V3: modify max height value
V4: remove size64
Signed-off-by: Boyuan Zhang <boyuan.zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit cbe408dba5)
Cc: stable@vger.kernel.org
Reuse the parameter length returned by
vcn_v4_0_enc_find_ib_param() instead of rereading it from
the IB.
This avoids a potential TOCTOU issue if the IB contents
change between reads.
Signed-off-by: Boyuan Zhang <boyuan.zhang@amd.com>
Reviewed-by: David Rosca <david.rosca@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit dbb02b4755)
Cc: stable@vger.kernel.org
When width or height is less than 16, width_in_mb or height_in_mb
becomes 0, leading to fs_in_mb being 0. This causes a division by
zero when calculating num_dpb_buffer in H264 and H264 Perf decode
paths.
Add validation to reject frames with width < 16 or height < 16
before performing any calculations that depend on these values.
V2: Format change - move up all vaiable definitions.
V3: Use warn_once to avoid spam.
Signed-off-by: Boyuan Zhang <boyuan.zhang@amd.com>
Reviewed-by: Leo Liu <leo.liu@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
(cherry picked from commit 3e41d26c70)
Cc: stable@vger.kernel.org
