Use sdma linear copy to migrate data between ram and vram. The sdma
linear copy command uses kernel buffer function queue to access system
memory through gart table.
Use reserved gart table window 0 to map system page address, and vram
page address is direct mapping. Use the same kernel buffer function to
fill in gart table mapping, so this is serialized with memory copy by
sdma job submit. We only need wait for the last memory copy sdma fence
for larger buffer migration.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
amdgpu_gmc_get_vm_pte use bo_va->is_xgmi same hive information to set
pte flags to update GPU mapping. Add local structure variable bo_va, and
update bo_va.is_xgmi, pass it to mapping->bo_va while mapping to GPU.
Assuming xgmi pstate is hi after boot.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
If svm range perfetch location is not zero, use TTM to alloc
amdgpu_bo vram nodes to validate svm range, then map vram nodes to GPUs.
Use offset to sub allocate from the same amdgpu_bo to handle overlap
vram range while adding new range or unmapping range.
svm_bo has ref count to trace the shared ranges. If all ranges of shared
amdgpu_bo are migrated to ram, ref count becomes 0, then amdgpu_bo is
released, all ranges svm_bo is set to NULL.
To migrate range from ram back to vram, allocate the same amdgpu_bo
with previous offset if the range has svm_bo.
If prange migrate to VRAM, no CPU mapping exist, then process exit will
not have unmap callback for this prange to free prange and svm bo. Free
outstanding pranges from svms list before process is freed in
svm_range_list_fini.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
HMM migration alloc sizeof(struct page) on system memory for each VRAM
page, it is 1GB system memory reserved for 64GB VRAM. To avoid
application OOM, increase system memory used size based on VRAM size of
all GPUs, then application alloc memory will fail if system memory usage
reach the limit.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Oak Zeng <Oak.Zeng@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Xnack retries are used for page fault recovery. Some AMD chip
families support continuously retry while page table entries are invalid.
The driver must handle the page fault interrupt and fill in a valid entry
for the GPU to continue.
This ioctl allows to enable/disable XNACK retries per KFD process.
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
XNACK mode controls the SQ RETRY_DISABLE setting that determines,
whether recoverable page faults can be supported on GFXv9 hardware.
Only on Aldebaran we can support different processes running with
different XNACK modes. On older chips all processes must use the same
RETRY_DISABLE setting. However, processes not relying on recoverable
page faults can work with RETRY enabled. This means XNACK off is always
available as a fallback so we can use the same mode on all GPUs in a
process.
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
HMM interval notifier callback notify CPU page table will be updated,
stop process queues if the updated address belongs to svm range
registered in process svms objects tree. Scheduled restore work to
update GPU page table using new pages address in the updated svm range.
The restore worker flushes any deferred work to make sure it restores
an up-to-date svm_range_list.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
When application explicitly call unmap or unmap from mmput when
application exit, driver will receive MMU_NOTIFY_UNMAP event to remove
svm range from process svms object tree and list first, unmap from GPUs
(in the following patch).
Split the svm ranges to handle partial unmapping of svm ranges. To
avoid deadlocks, updating MMU notifiers, range lists and interval trees
is done in a deferred worker. New child ranges are attached to their
parent range's child_list until the worker can update the
svm_range_list. svm_range_set_attr flushes deferred work and takes the
mmap_write_lock to guarantee that it has an up-to-date svm_range_list.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
svm range structure stores the range start address, size, attributes,
flags, prefetch location and gpu bitmap which indicates which GPU this
range maps to. Same virtual address is shared by CPU and GPUs.
Process has svm range list which uses both interval tree and list to
store all svm ranges registered by the process. Interval tree is used by
GPU vm fault handler and CPU page fault handler to get svm range
structure from the specific address. List is used to scan all ranges in
eviction restore work.
No overlap range interval [start, last] exist in svms object interval
tree. If process registers new range which has overlap with old range,
the old range split into 2 ranges depending on the overlap happens at
head or tail part of old range.
Apply attributes preferred location, prefetch location, mapping flags,
migration granularity to svm range, store mapping gpu index into bitmap.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Add svm (shared virtual memory) ioctl data structure and API definition.
The svm ioctl API is designed to be extensible in the future. All
operations are provided by a single IOCTL to preserve ioctl number
space. The arguments structure ends with a variable size array of
attributes that can be used to set or get one or multiple attributes.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
svm range uses gpu bitmap to store which GPU svm range maps to.
Application pass driver gpu id to specify GPU, the helper is needed to
convert gpu id to gpu bitmap idx.
Access through kfd_process_device pointers array from kfd_process.
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
DRM render node file handles are used for CPU mapping of BOs using mmap
by the Thunk. It uses the DRM render node of the GPU where the BO was
allocated.
DRM allows mmap access automatically when it creates a GEM handle for a
BO. KFD BOs don't have GEM handles, so KFD needs to manage access
manually. Use drm_vma_node_allow to allow user mode to mmap BOs allocated
with kfd_ioctl_alloc_memory_of_gpu through the DRM render node that was
used in the kfd_ioctl_acquire_vm call for the same GPU.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Philip Yang <philip.yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
amdgpu_amdkfd_gpuvm_alloc_memory_of_gpu needs the drm_priv to allow mmap
to access the BO through the corresponding file descriptor. The VM can
also be extracted from drm_priv, so drm_priv can replace the vm parameter
in the kfd2kgd interface.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Philip Yang <philip.yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
In poison progogate mode, when driver receive the edc error interrupt
from SQ, driver should kill the process by pasid which is using the
poison data, and then trigger GPU reset.
Signed-off-by: Dennis Li <Dennis.Li@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
ROCm user mode has acquired VMs from DRM file descriptors for as long
as it supported the upstream KFD. Legacy code to support older versions
of ROCm is not needed any more.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Power Management IP is initialized/enabled before KFD init. When a
thermal throttling happens before kfd_smi_init is done, calling the KFD
SMI update function causes a stack dump by referring a NULL pointer (
smi_clients list). Check if kfd_init is completed before calling the
function.
Signed-off-by: Amber Lin <Amber.Lin@amd.com>
Reviewed-by: Mukul Joshi <mukul.joshi@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Amdgpu driver uses 4-byte data type as DQM fence memory,
and transmits GPU address of fence memory to microcode
through query status PM4 message. However, query status
PM4 message definition and microcode processing are all
processed according to 8 bytes. Fence memory only allocates
4 bytes of memory, but microcode does write 8 bytes of memory,
so there is a memory corruption.
Changes since v1:
* Change dqm->fence_addr as a u64 pointer to fix this issue,
also fix up query_status and amdkfd_fence_wait_timeout function
uses 64 bit fence value to make them consistent.
Signed-off-by: Qu Huang <jinsdb@126.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Remove per_device_list from kfd_process and replace it with a
kfd_process_device pointers array of MAX_GPU_INSTANCES size. This helps
to manage the kfd_process_devices binded to a specific kfd_process.
Also, functions used by kfd_chardev to iterate over the list were
removed, since they are not valid anymore. Instead, it was replaced by a
local loop iterating the array.
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Jonathan Kim <jonathan.kim@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
A few functions that were intentended for the perf events support are
currently declared in arch/x86/events/amd/iommu.h, which mens they are
not in scope for the actual function definition. Also amdkfd has started
using a few of them using externs in a .c file. End that misery by
moving the prototypes to the proper header.
Signed-off-by: Christoph Hellwig <hch@lst.de>
Link: https://lore.kernel.org/r/20210402143312.372386-5-hch@lst.de
Signed-off-by: Joerg Roedel <jroedel@suse.de>
Amdgpu driver uses 4-byte data type as DQM fence memory,
and transmits GPU address of fence memory to microcode
through query status PM4 message. However, query status
PM4 message definition and microcode processing are all
processed according to 8 bytes. Fence memory only allocates
4 bytes of memory, but microcode does write 8 bytes of memory,
so there is a memory corruption.
Changes since v1:
* Change dqm->fence_addr as a u64 pointer to fix this issue,
also fix up query_status and amdkfd_fence_wait_timeout function
uses 64 bit fence value to make them consistent.
Signed-off-by: Qu Huang <jinsdb@126.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org
Using 'imply AMD_IOMMU_V2' does not guarantee that the driver can link
against the exported functions. If the GPU driver is built-in but the
IOMMU driver is a loadable module, the kfd_iommu.c file is indeed
built but does not work:
x86_64-linux-ld: drivers/gpu/drm/amd/amdkfd/kfd_iommu.o: in function `kfd_iommu_bind_process_to_device':
kfd_iommu.c:(.text+0x516): undefined reference to `amd_iommu_bind_pasid'
x86_64-linux-ld: drivers/gpu/drm/amd/amdkfd/kfd_iommu.o: in function `kfd_iommu_unbind_process':
kfd_iommu.c:(.text+0x691): undefined reference to `amd_iommu_unbind_pasid'
x86_64-linux-ld: drivers/gpu/drm/amd/amdkfd/kfd_iommu.o: in function `kfd_iommu_suspend':
kfd_iommu.c:(.text+0x966): undefined reference to `amd_iommu_set_invalidate_ctx_cb'
x86_64-linux-ld: kfd_iommu.c:(.text+0x97f): undefined reference to `amd_iommu_set_invalid_ppr_cb'
x86_64-linux-ld: kfd_iommu.c:(.text+0x9a4): undefined reference to `amd_iommu_free_device'
x86_64-linux-ld: drivers/gpu/drm/amd/amdkfd/kfd_iommu.o: in function `kfd_iommu_resume':
kfd_iommu.c:(.text+0xa9a): undefined reference to `amd_iommu_init_device'
x86_64-linux-ld: kfd_iommu.c:(.text+0xadc): undefined reference to `amd_iommu_set_invalidate_ctx_cb'
x86_64-linux-ld: kfd_iommu.c:(.text+0xaff): undefined reference to `amd_iommu_set_invalid_ppr_cb'
x86_64-linux-ld: kfd_iommu.c:(.text+0xc72): undefined reference to `amd_iommu_bind_pasid'
x86_64-linux-ld: kfd_iommu.c:(.text+0xe08): undefined reference to `amd_iommu_set_invalidate_ctx_cb'
x86_64-linux-ld: kfd_iommu.c:(.text+0xe26): undefined reference to `amd_iommu_set_invalid_ppr_cb'
x86_64-linux-ld: kfd_iommu.c:(.text+0xe42): undefined reference to `amd_iommu_free_device'
Use IS_REACHABLE to only build IOMMU-V2 support if the amd_iommu symbols
are reachable by the amdkfd driver. Output a warning if they are not,
because that may not be what the user was expecting.
Fixes: 64d1c3a43a ("drm/amdkfd: Centralize IOMMUv2 code and make it conditional")
Reported-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Create dedicated Aldebaran kfd2kgd callbacks to prepare
for new per-vmid register instructions for debug trap
setting functions and sending host traps.
v2: rebase (Alex)
Signed-off-by: Jonathan Kim <Jonathan.Kim@amd.com>
Reviewed-by: Oak Zeng <Oak.Zeng@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
MEC firmware can silently fail the queue preemption request
without time out. In this case, HIQ's MQD's queue_doorbell_id
will be set. Check this field to see whether last queue preemption
was successful or not.
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Suggested-by: Jay Cornwall <Jay.Cornwall@amd.com>
Acked-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
get_num_acc_vgprs does not set status.scc if the number of acc vgprs
is 0, so use an and instruction to set the condition code.
The Aldebaran handler binary was not based on the latest version of
the sources, so this update to the binary is the minimal change only
adding two instructions to set the condition code.
A newer version of the handler should be generated and tested in
another commit.
Signed-off-by: Laurent Morichetti <laurent.morichetti@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Currently host-gpu io link is always reported as PCIe however, on some
A+A systems, there could be one xgmi link available. This change exposes
xgmi link via sysfs when it is present.
v2: fix includes (Alex)
Reviewed-by: Oak Zeng <oak.zeng@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Rajneesh Bhardwaj <rajneesh.bhardwaj@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Similar to arcturus, but ARCH/ACC VGPRs may now be split unevenly.
A new field in SQ_WAVE_GPR_ALLOC tracks the boundary between the two
sets of VGPRs.
Squash below patches:
drm/amdkfd: Use preprocessor for IP-specific trap handler code
drm/amdkfd: Fix VGPR restore race in gfx8/gfx9 trap handler
drm/amdkfd: Remove duplicated code in gfx9 trap handler
drm/amdkfd: Separate ARCH/ACC VGPR restore in trap handler
drm/amdkfd: Reverse order of ARCH/ACC VGPR restore in trap handler
Signed-off-by: Jay Cornwall <jay.cornwall@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Trap handler is set per-process per-device and is unrelated
to queue management.
Move implementation closer to TMA setup code.
Signed-off-by: Jay Cornwall <jay.cornwall@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Existing, buggy user mode breaks when SRAM ECC is correctly reported as
"enabled". To avoid breaking existing user mode, deprecate that bit and
leave it as 0. Define a new bit to report the actual SRAM ECC mode that
new, correct user mode can use in the future.
Fixes: 7ec177bdcfc1 ("drm/amdkfd: fix set kfd node ras properties value")
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Kent Russell <kent.russell@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
If init_cwsr_apu fails, we currently leave the kfd_process structure in
place anyway. The next kfd_open will then succeed, using the existing
kfd_process structure. Fix that by cleaning up the kfd_process after a
failure in init_cwsr_apu.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Philip Yang <philip.yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
We use mmu_notifier_put to free the MMU notifier. That needs to be
paired with mmu_notifier_get to work correctly. Othewrise the next patch
would cause a kernel oops.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Philip Yang <philip.yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
memalloc_nofs_save/restore are no longer sufficient to prevent recursive
lock warnings when holding locks that can be taken in MMU notifiers. Use
memalloc_noreclaim_save/restore instead.
Fixes: f920e413ff ("mm: track mmu notifiers in fs_reclaim_acquire/release")
CC: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org # 5.10.x
Instead of caching the value during amdgpu_device_init, just call the
function directly. This avoids issues where the unique_id hasn't been
saved by the time that KFD's topology snapshot is done (e.g. Arcturus).
KFD's topology information from the amdgpu_device was initially cached
at KFD initialization due to amdkfd and amdgpu being separate modules.
Now that they are combined together, we can directly call the functions
that we need and avoid this unnecessary duplication and complexity.
As a side-effect of this change, we also remove unique_id=0 for CPUs,
which is obviously not unique.
v2: Drop previous patch printing unique_id in hex
Signed-off-by: Kent Russell <kent.russell@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
ROCm user mode depends on userptr support. Without it, KFD is basically
useless.
Make sure HSA_AMD selects the same options as DRM_AMDGPU_USERPTR to avoid
broken configurations where userptr gets enabled but its dependencies are
disabled.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Philip Yang <philip.yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
KASAN reported a slab-out-of-bounds read of size 1 in
kdf_create_vcrat_image_cpu().
This occurs when, for example, when on an x86_64 with a single NUMA node
because kfd_fill_iolink_info_for_cpu() is a no-op, but afterwards the
sub_type_hdr->length, which is out-of-bounds, is read and multiplied by
entries. Fortunately, entries is 0 in this case so the overall
crat_table->length is still correct.
Check if there were any entries before de-referencing sub_type_hdr which
may be pointing to out-of-bounds memory.
Fixes: b7b6c38529 ("drm/amdkfd: Calculate CPU VCRAT size dynamically (v2)")
Suggested-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Jeremy Cline <jcline@redhat.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
We can have the same src ids for different client ids so make sure to
check both the client id and the source id when handling interrupts.
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
GFX10 CP firmware expects PCIe atomics support. Don't enumerate GFX10
devices on platforms (PCIe v2) that don't support PCIe atomics.
Currently, some of the applications like clinfo soft hangs on platforms
without PCIe atomics support.
Signed-off-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Release dmabuf reference before returning from kfd_ioctl_import_dmabuf.
amdgpu_amdkfd_gpuvm_import_dmabuf takes a reference to the underlying
GEM BO and doesn't keep the reference to the dmabuf wrapper.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Kent Russell <kent.russell@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
