[Why]
Avoid conflict with address ranges mapped by SVM
mechanism that try to be allocated again through
ioctl_alloc in the same process. And viceversa.
[How]
For ioctl_alloc_memory_of_gpu allocations
Check if the address range passed into ioctl memory
alloc does not exist already in the kfd_process
svms->objects interval tree.
For SVM allocations
Look for the address range into the interval tree VA from
the VM inside of each pdds used in a kfd_process.
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This fix the deadlock with the BO reservations during SVM_BO evictions
while allocations in VRAM are concurrently performed. More specific,
while the ttm waits for the fence to be signaled (ttm_bo_wait), it
already has the BO reserved. In parallel, the restore worker might be
running, prefetching memory to VRAM. This also requires to reserve the
BO, but blocks the mmap semaphore first. The deadlock happens when the
SVM_BO eviction worker kicks in and waits for the mmap semaphore held
in restore worker. Preventing signal the fence back, causing the
deadlock until the ttm times out.
We don't need to hold the BO reservation anymore during validation
and mapping. Now the physical addresses are taken from hmm_range_fault.
We also take migrate_mutex to prevent range migration while
validate_and_map update GPU page table.
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
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 can get the pdev and asic type from the adev. No need
to pass them explicitly.
v2: squash in build fix for !CONFIG_HSA_AMD from Anson
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
In ras poison mode, page retirement will be handled by the irq handler of the
module which consumes corrupted data.
v2: rename ras_process_cb to ras_poison_consumption_handler.
move the handler's implementation from ASIC specific file to common
file.
v3: call gpu reset for xGMI connected mode.
Signed-off-by: Tao Zhou <tao.zhou1@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Use resource_size function on resource object
instead of explicit computation.
Clean up coccicheck warning:
./drivers/gpu/drm/amd/amdkfd/kfd_migrate.c:905:10-13: ERROR: Missing
resource_size with res
Reported-by: Abaci Robot <abaci@linux.alibaba.com>
Reviewed-by: Amos Kong <kongjianjun@gmail.com>
Signed-off-by: Yang Li <yang.lee@linux.alibaba.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Device manager releases device-specific resources when a driver
disconnects from a device, devm_memunmap_pages and
devm_release_mem_region calls in svm_migrate_fini are redundant.
It causes below warning trace after patch "drm/amdgpu: Split
amdgpu_device_fini into early and late", so remove function
svm_migrate_fini.
BUG: https://gitlab.freedesktop.org/drm/amd/-/issues/1718
WARNING: CPU: 1 PID: 3646 at drivers/base/devres.c:795
devm_release_action+0x51/0x60
Call Trace:
? memunmap_pages+0x360/0x360
svm_migrate_fini+0x2d/0x60 [amdgpu]
kgd2kfd_device_exit+0x23/0xa0 [amdgpu]
amdgpu_amdkfd_device_fini_sw+0x1d/0x30 [amdgpu]
amdgpu_device_fini_sw+0x45/0x290 [amdgpu]
amdgpu_driver_release_kms+0x12/0x30 [amdgpu]
drm_dev_release+0x20/0x40 [drm]
release_nodes+0x196/0x1e0
device_release_driver_internal+0x104/0x1d0
driver_detach+0x47/0x90
bus_remove_driver+0x7a/0xd0
pci_unregister_driver+0x3d/0x90
amdgpu_exit+0x11/0x20 [amdgpu]
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
If svm migration init failed to create pgmap for device memory, set
pgmap type to 0 to disable device SVM support capability.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
For xnack off, restore work dma unmap previous system memory page, and
dma map the updated system memory page to update GPU mapping, this is
not dma mapping leaking, remove the WARN_ONCE for dma mapping leaking.
prange->dma_addr store the VRAM page pfn after the range migrated to
VRAM, should not dma unmap VRAM page when updating GPU mapping or
remove prange. Add helper svm_is_valid_dma_mapping_addr to check VRAM
page and error cases.
Mask out SVM_RANGE_VRAM_DOMAIN flag in dma_addr before calling amdgpu vm
update to avoid BUG_ON(*addr & 0xFFFF00000000003FULL), and set it again
immediately after. This flag is used to know the type of page later to
dma unmapping system memory page.
Fixes: 1d5dbfe6c0 ("drm/amdkfd: classify and map mixed svm range pages in GPU")
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
SVM range may includes multiple VMAs with different vm_flags, if prange
page index is the last page of the VMA offset + npages, update GPU
mapping to create GPU page table with same VMA access permission.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
On some GPUs the PCIe atomic requirement for KFD depends on the MEC
firmware version. Add a firmware version check for this. The minimum
firmware version that works without atomics can be updated in the
device_info structure for each GPU type.
Move PCIe atomic detection from kgd2kfd_probe into kgd2kfd_device_init
because the MEC firmware is not loaded yet at the probe stage.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Guchun Chen <guchun.chen@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
On systems with multiple SH per SE compute_static_thread_mgmt_se#
is split into independent masks, one for each SH, in the upper and
lower 16 bits. We need to detect this and apply cu masking to each
SH. The cu mask bits are assigned first to each SE, then to
alternate SHs, then finally to higher CU id. This ensures that
the maximum number of SPIs are engaged as early as possible while
balancing CU assignment to each SH.
v2: Use max SH/SE rather than max SH in cu_per_sh.
v3: Fix comment blocks, ensure se_mask is initially zero filled,
and correctly assign se.sh.cu positions to unset bits in cu_mask.
Signed-off-by: Sean Keely <Sean.Keely@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Restore retry fault or prefetch range, or restore svm range after
eviction to map range to GPU with correct read or write access
permission.
Range may includes multiple VMAs, update GPU page table with offset of
prange, number of pages for each VMA according VMA access permission.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Check range access permission to restore GPU retry fault, if GPU retry
fault on address which belongs to VMA, and VMA has no read or write
permission requested by GPU, failed to restore the address. The vm fault
event will pass back to user space.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
KFDSVMRangeTest.SetGetAttributesTest randomly fails in stress test.
Note: Google Test filter = KFDSVMRangeTest.*
[==========] Running 18 tests from 1 test case.
[----------] Global test environment set-up.
[----------] 18 tests from KFDSVMRangeTest
[ RUN ] KFDSVMRangeTest.BasicSystemMemTest
[ OK ] KFDSVMRangeTest.BasicSystemMemTest (30 ms)
[ RUN ] KFDSVMRangeTest.SetGetAttributesTest
[ ] Get default atrributes
/home/yifan/brahma/libhsakmt/tests/kfdtest/src/KFDSVMRangeTest.cpp:154: Failure
Value of: expectedDefaultResults[i]
Actual: 4294967295
Expected: outputAttributes[i].value
Which is: 0
/home/yifan/brahma/libhsakmt/tests/kfdtest/src/KFDSVMRangeTest.cpp:154: Failure
Value of: expectedDefaultResults[i]
Actual: 4294967295
Expected: outputAttributes[i].value
Which is: 0
/home/yifan/brahma/libhsakmt/tests/kfdtest/src/KFDSVMRangeTest.cpp:152: Failure
Value of: expectedDefaultResults[i]
Actual: 4
Expected: outputAttributes[i].type
Which is: 2
[ ] Setting/Getting atrributes
[ FAILED ]
the root cause is that svm work queue has not finished when svm_range_get_attr is called, thus
some garbage svm interval tree data make svm_range_get_attr get wrong result. Flush work queue before
iterate svm interval tree.
Signed-off-by: Yifan Zhang <yifan1.zhang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
For xnack on, if range ACCESS or ACCESS_IN_PLACE (AIP) by single GPU, or
range is ACCESS_IN_PLACE by mGPUs and all mGPUs connection on XGMI same
hive, the best prefetch location is prefetch_loc GPU. Otherwise, the best
prefetch location is always CPU because GPU does not have coherent
mapping VRAM of other GPUs even with large-BAR PCIe connection.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Currently the SVM get_attr call allows querying, which flags are set
in the entire address range. Add the opposite query, which flags are
clear in the entire address range. Both queries can be combined in a
single get_attr call, which allows answering questions such as, "is
this address range coherent, non-coherent, or a mix of both"?
Proposed userspace for UAPI:
https://github.com/RadeonOpenCompute/ROCR-Runtime/tree/memory_model_queries
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Philip Yand <philip.yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Add u32 gfx_target_version field to kfd_node_properties and
kfd_device_info. Populate <asic>_device_info structs accordingly and
expose to sysfs.
This allows eliminating device-ID-based lookup tables in user mode for
future ASICs.
Signed-off-by: Graham Sider <Graham.Sider@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
It is to workaround HW bug on other Asics and based on
reverting two commits back:
drm/amdkfd: Add heavy-weight TLB flush after unmapping
drm/amdkfd: Add memory sync before TLB flush on unmap
Signed-off-by: Eric Huang <jinhuieric.huang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
start_cpsch and stop_cpsch can be called during kfd device
initialization or during gpu reset/recovery. So they can
run concurrently. Currently in start_cpsch and stop_cpsch,
pm_init and pm_uninit is not protected by the dpm lock.
Imagine such a case that user use packet manager's function
to submit a pm4 packet to hang hws (ie through command
cat /sys/class/kfd/kfd/topology/nodes/1/gpu_id | sudo tee
/sys/kernel/debug/kfd/hang_hws), while kfd device is under
device reset/recovery so packet manager can be not initialized.
There will be unpredictable protection fault in such case.
This patch moves pm_init/uninit inside the dpm lock and check
packet manager is initialized before using packet manager
function.
Signed-off-by: Oak Zeng <Oak.Zeng@amd.com>
Acked-by: Christian Konig <christian.koenig@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Similar to xGMI reporting the min/max bandwidth between direct peers, PCIe
will report the min/max bandwidth to the KFD.
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Report the min/max bandwidth in megabytes to the kfd for direct
xgmi connections only. Indirect peers will report 0 since
indirect route is unknown.
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
prange is NULL if vm fault retry on invalid address, for this case, can
not use prange to get pdd, use adev to get gpuidx and then get pdd
instead, then increase pdd vm fault counter.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
prange is NULL if vm fault retry on invalid address, for this case, can
not use prange to get pdd, use adev to get gpuidx and then get pdd
instead, then increase pdd vm fault counter.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Each zone-device page holds a reference to the SVM BO that manages its
backing storage. This is necessary to correctly hold on to the BO in
case zone_device pages are shared with a child-process.
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This is for debug purposes only.
It conditionally generates partial migrations to test mixed
CPU/GPU memory domain pages in a prange easily.
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
