The function kfd_lookup_process_by_pasid will increase the reference
count of kfd_process object, its caller should call kfd_unref_process to
decrease the reference count. Otherwise resource leakage will happen.
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>
Add the function pointer.
Signed-off-by: Chengming Gui <Jack.Gui@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: Tao Zhou <tao.zhou1@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Need do a heavy-weight TLB flush to make sure we have no more dirty data
in the cache for the unmapped pages.
Define enum TLB_FLUSH_TYPE, add flush_type parameter to
amdgpu_amdkfd_flush_gpu_tlb_pasid.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
This reverts commit 1704ac8e43.
After "drm/amdgpu: flush TLB if valid PDE turns into PTE" is checked
in, this workaround is not needed.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Mapping huge page, 2MB aligned address with 2MB size, uses PDE0 as PTE.
If previously valid PDE0, PDE0.V=1 and PDE0.P=0 turns into PTE, this
requires TLB flush, otherwise page table walker will not read updated
PDE0.
Change page table update mapping to return table_freed flag to indicate
the previously valid PDE may have turned into a PTE if page table is
freed.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Now that we found the underlying problem we can re-apply this patch.
This reverts commit 6b44b667e2.
v2: rebase on KFD changes
Signed-off-by: Christian König <christian.koenig@amd.com>
Acked-by: Felix Kuehling <Felix.Kuehling@amd.com>
Tested-by: Nirmoy Das <nirmoy.das@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
MTYPE UC was used for a specific use case that ended up not being
implemented. Use NC for better performance for coarse-grained memory where
cache coherence during shader execution is not required.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Oak Zeng <Oak.Zeng@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
MTYPE UC was used for a specific use case that ended up not being
implemented. Use NC for better performance for coarse-grained memory where
cache coherence during shader execution is not required.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Oak Zeng <Oak.Zeng@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The user applications maybe register the KFD_EVENT_TYPE_HW_EXCEPTION and
KFD_EVENT_TYPE_MEMORY events, driver could notify them when poison data
consumed. Beside that, some applications maybe register SIGBUS signal
hander. These applications will handle poison data by themselves, exit
or re-create context to re-dispatch works.
Signed-off-by: Dennis Li <Dennis.Li@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
If xnack is on, new range is created to recover retry vm fault or
created by SVM API calls, set all GPUs have access to the range.
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 migration copy failed because process is killed, or out of VRAM or
system memory, pass error code back to caller to handle error
gracefully.
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>
Rename,
ras_hw_supported --> ras_hw_enabled, and
ras_features --> ras_enabled,
to show that ras_enabled is a subset of
ras_hw_enabled, which itself is a subset
of the ASIC capability.
Cc: Alexander Deucher <Alexander.Deucher@amd.com>
Cc: John Clements <john.clements@amd.com>
Cc: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Luben Tuikov <luben.tuikov@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Reviewed-by: John Clements <John.Clements@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
In NPS4 BIOS we need to find the closest numa node when creating
topology io link between cpu and gpu, if PCI driver doesn't set
it.
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>
The L1 cache information has been updated and the L2/L3
information has been added. The changes have been made
for Vega10 and newer ASICs. There are no changes
for the older ASICs before Vega10.
Signed-off-by: Mike Li <Tianxinmike.Li@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
To account for various PCIe and xGMI setups, check the no atomics settings
for a device in relation to every direct peer.
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>
This function is only used in this source file.
Reported-by: kernel test robot <lkp@intel.com>
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Rodrigo Siqueira <Rodrigo.Siqueira@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
To workaround the situation that vm retry fault keep coming after page
table update. We are investigating the root cause, but once this issue
happens, application will stuck and sometimes have to reboot to recover.
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>
update kfd_supported_devices to enable Aldebaran virtualization support
Signed-off-by: Zhigang Luo <zhigang.luo@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
GPUs connected to CPUs over xGMI are bidirectional so set weight by a
single hop both ways.
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Tested-by: Ramesh Errabolu <ramesh.errabolu@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Link atomics support over xGMI should be reported independently of PCIe.
Do not set NO_ATOMICS flags on devices that support xGMI but that do not
have atomics support over PCIe.
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Tested-by: Ramesh Errabolu <ramesh.errabolu@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Including:
- Big cleanup of almost unsused parts of the IOMMU API by
Christoph Hellwig. This mostly affects the Freescale PAMU
driver.
- New IOMMU driver for Unisoc SOCs
- ARM SMMU Updates from Will:
- SMMUv3: Drop vestigial PREFETCH_ADDR support
- SMMUv3: Elide TLB sync logic for empty gather
- SMMUv3: Fix "Service Failure Mode" handling
- SMMUv2: New Qualcomm compatible string
- Removal of the AMD IOMMU performance counter writeable check
on AMD. It caused long boot delays on some machines and is
only needed to work around an errata on some older (possibly
pre-production) chips. If someone is still hit by this
hardware issue anyway the performance counters will just
return 0.
- Support for targeted invalidations in the AMD IOMMU driver.
Before that the driver only invalidated a single 4k page or the
whole IO/TLB for an address space. This has been extended now
and is mostly useful for emulated AMD IOMMUs.
- Several fixes for the Shared Virtual Memory support in the
Intel VT-d driver
- Mediatek drivers can now be built as modules
- Re-introduction of the forcedac boot option which got lost
when converting the Intel VT-d driver to the common dma-iommu
implementation.
- Extension of the IOMMU device registration interface and
support iommu_ops to be const again when drivers are built as
modules.
-----BEGIN PGP SIGNATURE-----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=qF2C
-----END PGP SIGNATURE-----
Merge tag 'iommu-updates-v5.13' of git://git.kernel.org/pub/scm/linux/kernel/git/joro/iommu
Pull iommu updates from Joerg Roedel:
- Big cleanup of almost unsused parts of the IOMMU API by Christoph
Hellwig. This mostly affects the Freescale PAMU driver.
- New IOMMU driver for Unisoc SOCs
- ARM SMMU Updates from Will:
- Drop vestigial PREFETCH_ADDR support (SMMUv3)
- Elide TLB sync logic for empty gather (SMMUv3)
- Fix "Service Failure Mode" handling (SMMUv3)
- New Qualcomm compatible string (SMMUv2)
- Removal of the AMD IOMMU performance counter writeable check on AMD.
It caused long boot delays on some machines and is only needed to
work around an errata on some older (possibly pre-production) chips.
If someone is still hit by this hardware issue anyway the performance
counters will just return 0.
- Support for targeted invalidations in the AMD IOMMU driver. Before
that the driver only invalidated a single 4k page or the whole IO/TLB
for an address space. This has been extended now and is mostly useful
for emulated AMD IOMMUs.
- Several fixes for the Shared Virtual Memory support in the Intel VT-d
driver
- Mediatek drivers can now be built as modules
- Re-introduction of the forcedac boot option which got lost when
converting the Intel VT-d driver to the common dma-iommu
implementation.
- Extension of the IOMMU device registration interface and support
iommu_ops to be const again when drivers are built as modules.
* tag 'iommu-updates-v5.13' of git://git.kernel.org/pub/scm/linux/kernel/git/joro/iommu: (84 commits)
iommu: Streamline registration interface
iommu: Statically set module owner
iommu/mediatek-v1: Add error handle for mtk_iommu_probe
iommu/mediatek-v1: Avoid build fail when build as module
iommu/mediatek: Always enable the clk on resume
iommu/fsl-pamu: Fix uninitialized variable warning
iommu/vt-d: Force to flush iotlb before creating superpage
iommu/amd: Put newline after closing bracket in warning
iommu/vt-d: Fix an error handling path in 'intel_prepare_irq_remapping()'
iommu/vt-d: Fix build error of pasid_enable_wpe() with !X86
iommu/amd: Remove performance counter pre-initialization test
Revert "iommu/amd: Fix performance counter initialization"
iommu/amd: Remove duplicate check of devid
iommu/exynos: Remove unneeded local variable initialization
iommu/amd: Page-specific invalidations for more than one page
iommu/arm-smmu-v3: Remove the unused fields for PREFETCH_CONFIG command
iommu/vt-d: Avoid unnecessary cache flush in pasid entry teardown
iommu/vt-d: Invalidate PASID cache when root/context entry changed
iommu/vt-d: Remove WO permissions on second-level paging entries
iommu/vt-d: Report the right page fault address
...
v2: updated MEC FW version after validating gws with debugger
Signed-off-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Reviewed-by: Joseph Greathouse <Joseph.Greathouse@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
After draining the stale retry fault, or failed to validate the range
to recover, have to remove the fault address from fault filter ring, to
be able to handle subsequent retry interrupt on same address. Otherwise
the retry fault will not be processed to recover until timeout passed.
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>
Retry fault interrupt maybe pending in IH ring after GPU page table
is updated to recover the vm fault, because each page of the range
generate retry fault interrupt. There is race if application unmap
range to remove and free the range first and then retry fault work
restore_pages handle the retry fault interrupt, because range can not be
found, this vm fault can not be recovered and report incorrect GPU vm
fault to application.
Before unmap to remove and free range, drain retry fault interrupt
from IH ring1 to ensure no retry fault comes after the range is removed.
Drain retry fault interrupt skip the range which is on deferred list
to remove, or the range is child range, which is split by unmap, does
not add to svms and have interval notifier.
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>
GPU vm retry fault recover range need retry validation if
1. range is split in parallel by unmap while recover
2. range migrate to system memory and range is updated in system
memory while recover
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>
The plural of 'process' should be 'processes'.
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>
Use devm_memunmap_pages instead of memunmap_pages to release pgmap
and remove pgmap from device action, to avoid double free pgmap when
unloading driver module.
Release device memory region if failed to create device memory pages
structure.
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>
There is a spelling mistake in a pr_debug message. Fix it.
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Reviewed-by: Nirmoy Das <nirmoy.das@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.h is included in kfd_priv.h
Signed-off-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: John Clements <John.Clements@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Fixed a kernel-doc error in the documentation of a function.
Signed-off-by: Fabio M. De Francesco <fmdefrancesco@gmail.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>
The variable r is being initialized with a value that is never read
and it is being updated later with a new value. The initialization is
redundant and can be removed.
Addresses-Coverity: ("Unused value")
Signed-off-by: Colin Ian King <colin.king@canonical.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>
Currently the call to kfd_process_gpuidx_from_gpuid is returning an
int value and this is being assigned to a uint32_t variable gpuidx
and this is being checked for a negative error return which is always
going to be false. Fix this by making gpuidx an int32_t. This makes
gpuidx also type consistent with the use of gpuidx from the callers.
Addresses-Coverity: ("Unsigned compared against 0")
Fixes: cda0f85bfa ("drm/amdkfd: refine migration policy with xnack on")
Signed-off-by: Colin Ian King <colin.king@canonical.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>
Attribute access value for default ranges is set, based on
process xnack on/off.
XNACK ON has GPU access attribute for unregistered ranges through page
fault. While XNACK OFF has no access attribute for unregistered ranges.
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>
SVM ranges are created for unregistered memory, triggered
by page faults. These ranges are migrated/mapped to
GPU VRAM memory.
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>
In order to support multi-process debugging, HWS PM4 packet
MAP_PROCESS requires an extension of 5 DWORDS to support targeting of
per-vmid SPI debug control registers as well as watch points per process.
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>
Control whether to build SVM support into amdgpu with a Kconfig option.
This makes it easier to disable it in production kernels if this new
feature causes problems in production environments.
Use "depends on" instead of "select" for DEVICE_PRIVATE, as is
recommended for visible options.
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>
SVMAPISupported property added to HSA_CAPABILITY, the value match
HSA_CAPABILITY defined in Thunk spec:
SVMAPISupported: it will not be supported on older kernels that don't
have HMM or on systems with GFXv8 or older GPUs without support for
48-bit virtual addresses.
CoherentHostAccess property added to HSA_MEMORYPROPERTY, the value match
HSA_MEMORYPROPERTY defined in Thunk spec:
CoherentHostAccess: whether or not device memory can be coherently
accessed by the host CPU.
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 prefetch range to gpu with acutal location is another gpu, or GPU
retry fault restore pages to migrate the range with acutal location is
gpu, then migrate from one gpu to another gpu.
Use system memory as bridge because sdma engine may not able to access
another gpu vram, use sdma of source gpu to migrate to system memory,
then use sdma of destination gpu to migrate from system memory to gpu.
Print out gpuid or gpuidx in debug messages.
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>
With xnack on, add validate timestamp in order to handle GPU vm fault
from multiple GPUs.
If GPU retry fault need migrate the range to the best restore location,
use range validate timestamp to record system timestamp after range is
restored to update GPU page table.
Because multiple pages of same range have multiple retry fault, define
AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING to the long time period that
pending retry fault may still comes after page table update, to skip
duplicate retry fault of same range.
If difference between system timestamp and range last validate timestamp
is bigger than AMDGPU_SVM_RANGE_RETRY_FAULT_PENDING, that means the
retry fault is from another GPU, then continue to handle retry fault
recover.
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>
With xnack on, GPU vm fault handler decide the best restore location,
then migrate range to the best restore location and update GPU mapping
to recover the GPU vm fault.
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_bo eviction mechanism is different from regular BOs.
Every SVM_BO created contains one eviction fence and one
worker item for eviction process.
SVM_BOs can be attached to one or more pranges.
For SVM_BO eviction mechanism, TTM will start to call
enable_signal callback for every SVM_BO until VRAM space
is available.
Here, all the ttm_evict calls are synchronous, this guarantees
that each eviction has completed and the fence has signaled before
it returns.
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
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>
Page table restore implementation in SVM API. This is called from
the fault handler at amdgpu_vm. To update page tables through
the page fault retry IH.
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
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>
GPU page tables are invalidated by unmapping prange directly at
the mmu notifier, when page fault retry is enabled through
amdgpu_noretry global parameter. The restore page table is
performed at the page fault handler.
If xnack is on, we update GPU mappings after migration to avoid
unnecessary GPUVM faults.
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
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 CPU page fault happens, HMM pgmap_ops callback migrate_to_ram start
migrate memory from vram to ram in steps:
1. migrate_vma_pages get vram pages, and notify HMM to invalidate the
pages, HMM interval notifier callback evict process queues
2. Allocate system memory pages
3. Use svm copy memory to migrate data from vram to ram
4. migrate_vma_pages copy pages structure from vram pages to ram pages
5. Return VM_FAULT_SIGBUS if migration failed, to notify application
6. migrate_vma_finalize put vram pages, page_free callback free vram
pages and vram nodes
7. Restore work wait for migration is finished, then update GPU page
table mapping to system memory, and resume process queues
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>
Register svm range with same address and size but perferred_location
is changed from CPU to GPU or from GPU to CPU, trigger migration the svm
range from ram to vram or from vram to ram.
If svm range prefetch location is GPU with flags
KFD_IOCTL_SVM_FLAG_HOST_ACCESS, validate the svm range on ram first,
then migrate it from ram to vram.
After migrating to vram is done, CPU access will have cpu page fault,
page fault handler migrate it back to ram and resume cpu access.
Migration steps:
1. migrate_vma_pages get svm range ram pages, notify the
interval is invalidated and unmap from CPU page table, HMM interval
notifier callback evict process queues
2. Allocate new pages in vram using TTM
3. Use svm copy memory to sdma copy data from ram to vram
4. migrate_vma_pages copy ram pages structure to vram pages structure
5. migrate_vma_finalize put ram pages to free ram pages and memory
6. Restore work wait for migration is finished, then update GPUs page
table mapping to new vram pages, resume process queues
If migrate_vma_setup failed to collect all ram pages of range, retry 3
times until success to start 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>
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>
