This is a temporary work around for GPU hang issues observed on GFX11.
Change-Id: I98fbedbbd1c51fe402c2116b35ca548931a390c9
Signed-off-by: David Belanger <david.belanger@amd.com>
This reverts commit 7787a039bd.
It causes a regression in pytorch benchmark.
Change-Id: I96173dbd061cf38d6f451c02cb181ae51b7f625e
Signed-off-by: Eric Huang <jinhuieric.huang@amd.com>
This reverts commit 178a619b80.
There are some openMP issues that were introduced after SVM userptr
feature was added.
Signed-off-by: Alex Sierra <Alex.Sierra@amd.com>
Change-Id: I7ef87c5232a3bcbe594c743fa4b4958601845ba5
This reverts commit 45fad29752.
There are some openMP issues that were introduced after SVM userptr
feature was added.
Signed-off-by: Alex Sierra <Alex.Sierra@amd.com>
Change-Id: I6566c9f0d39d05ecb92f38159880763f432939a5
This reverts commit 8a746bdaed.
There are some openMP issues that were introduced after SVM userptr
feature was added.
Signed-off-by: Alex Sierra <Alex.Sierra@amd.com>
Change-Id: Ib01046571d2c84fa0fd228ecba0dee0eae3f994d
This is handled by __fmm_release calling aperture_release_area.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Change-Id: Ib8ed300e1734f03aeb9dfc8074897ece310b8af9
Use a common helper for CPU mappings to reduce duplicate code.
Consistently use MAP_SHARED for all render_fd mappings.
Remove double-mapping for AQL queue buffers on the CPU. This workaround
is only needed on the GPU.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Change-Id: Iff86c8cc9f1e5c982614b3f11129bc2cf8cbba02
The NULL pointer check was the only way for that function to fail. And it
was done after the pointer was accessed. Simplify this by just returning
the result as a return value instead of using a pointer as output
parameter. This way the function can never fail and the caller doesn't
need to do any error handling.
Declare the function in libhsakmt.h instead of duplicating the
declaration in fmm.c.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Change-Id: I91b90d66166fd3b5cdc47c73a9bbc369c45b51fe
Setting this variable to '0' will force to disable memory
registration/allocation through SVM API mechanism.
Not setting this or setting to '1', SVM API will be used only if all
GPUs support it.
Signed-off-by: Alex Sierra <Alex.Sierra@amd.com>
Change-Id: Icdf7656de09aa9988b567ec6c024953398e9bb48
When is hsaKmtCreateQueue called first time for node
doorbells[NodeId].size is initialized to zero in init_process_doorbells
but used to calculate the doorbell offset. It works just by accident
because doorbells[NodeId].size is uint32_t so -1 will be 0xFFFFFFFF which
is zero extended into 0x00000000FFFFFFFF and it will work as long as mmap
offset bits are not within lower 32 bits.
Bug: https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface/issues/78
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Change-Id: Ia791adfc51363d4704cb50fa4f01137b7dd48a75
When hsa is closed, it would close open fds for /dev/kfd but
not for /dev/dri/renderD*. This caused issues with CRIU
checkpoint, which expects that /dev/kfd will be open if
/dev/dri/renderD* is.
As a workaround for the CRIU behaviour, leave /dev/kfd open
when closing hsa.
Signed-off-by: David Francis <David.Francis@amd.com>
Change-Id: Ie1b2d5b1d8986750b0e560ae2934b7c73cff942e
Get more debug information about user pointers that were registered
through SVM API, and triggered by memory exception events.
A new kfdtest with this use case was also included inside
KFDExceptionTest.
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Change-Id: I8e9df3c1c6c3f42d7b9235d12406d80d31746443
Register and map userptrs through Shared Virtual Memory(SVM) API at
the Kernel level when available. Using this approach, performance
will be improve as register/unregister memory will not trigger any
system call to KFD driver.
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Change-Id: I20723cbeb340bf48b95e1115f0102c031397bc14
Prior to launch some ASICs may re-use PCI DIDs from older generations.
This can cause issues during topology initialization as hsa_gfxip_table
lookups will override sysfs-provided gfx versions, causing incorrect
gfxip selection. Since no new entries will be added to hsa_gfxip_table,
limit its search only to pre-GFX11 ASICs.
Signed-off-by: Graham Sider <Graham.Sider@amd.com>
Change-Id: I53eaefac5db2650a36a6ce9f21daf750f50cfd26
fork process copy-on-write MMU nitifier on CWSR range will evict user
queues, and then update GPU mapping and resume queues, use MADV_DONTFORK
to avoid COW MMU notifier callback on CWSR SVM range.
Use mmap to alloc SVM range for CWSR because posix_memalign don't alloc
new range in child process, this fails to register svm range as range is
invalid address in forked child process.
Change-Id: Ibaea56a691dd6f577ed2e1f2d43f4a3500b8316f
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
To remove duplicate mmap aligned allocation code.
Change-Id: Ibc05cc4aaf6d190bd2382e33bdeca1496960c5f2
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
mmap alloc larger address range with align padding page plus guard
pages, then unmap the padding and guard pages at beginning and end
of the range, return aligned address range.
Change-Id: Iaf3c711a079c744289efbafee9b5e63aaf724765
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
GFX1036(ISA version) is not included in the previous range.
This patch can really include all gfx10 series ASICs.
Signed-off-by: Aaron Liu <aaron.liu@amd.com>
Change-Id: I0e28dbfc031c216166b306b9fb39f644f75a330f
Avoiding the segfault, runtime debugger enable is not supported
if the firmware of gpu doesn't support debug exceptions.
Signed-off-by: jie1zhan <jesse.zhang@amd.com>
Change-Id: Ifad57a6e78cb1c92b1f8927355ece8c64e89c51b
Remove potential double free condition when free_queue() is called
after hsaKmtDestroyQueue() if mapping doorbell fails during queue
creation.
Signed-off-by: Mukul Joshi <mukul.joshi@amd.com>
Change-Id: If2aa19c455b30d2940b232dbafb9cc1eaad721a5
Queue ctx_save_restore memory is allocated with size
ctx_save_restore_size + debug_memory_size, use the same size
in free_queue to free ctx_save_restore memory.
Change-Id: I4902ff15fb82ddea64b8342b89776a1bf5c38d13
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Avoiding segfault when an invalid SharedMemoryHandle is passed in
when calling fmm_register_shared_memory.
Change-Id: I0e0bbed01487fc10afcbb170eb9330e70b209d14
Signed-off-by: David Yat Sin <David.YatSin@amd.com>
Now that HsaNodeProperties is passed in to
topology_get_node_props_from_drm, check that pointer instead of the
pointer for MarketingName (which throws a compiler warning)
Signed-off-by: kent.russell@amd.com <kent.russell@amd.com>
Change-Id: If76b24e1bab5a62e514ab440b6316c7b7cd264c1
Query family id info from drm render node, then
ROCr can query this info directly from Thunk
instead of parsing the info by itself.
Signed-off-by: Lang Yu <Lang.Yu@amd.com>
Change-Id: I030bd27ab2379fbf87f3d787302c3b8613456278
The debugger can now request snapshot copies with entry size and
set/clear watchpoints by device.
v3: drop min version check to v10.0
v2: check runtime allowance from v10.3 to 13.x
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Change-Id: I9befefb596201a11591de218db29a9317b41e69b
The debugger depends on the CWSR area being executable. Set the right
flag when registering SVM memory.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Tested-by: Laurent Morichetti <Laurent.Morichetti@amd.com>
Change-Id: I7441e214d1a4da8324d775e777976fabd1c81a6f
The CMA feature is deprecated and about to be removed from the DKMS
branch. It was never supported upstream. Leave dummy functions in
place for now.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Change-Id: I9e51403d753cb91630553aff4f19e931af509740
hsaKmtMapMemoryToGPU should not try to map VRAM on peer GPUs that don't
have an IO-Link to the memory. The new P2P mapping code in KFD will
fail otherwise.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Change-Id: I6d59b55651b98756865a0f69eafef3e386372cf3
This allows init_process_apertures to use the whole consistent topoology
instead of taking its own partial snapshot.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Change-Id: Ia13e7aa7fcd090ea8d6cacd4babb29a27c20207f
Kernel debug IOCTL got version bumped to v11.
Updated runtime enable but missed runtime disable check update.
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Change-Id: I71a8970ccfe7dc517abe7b4ad962369aea6a0496
To improve performance on queue preemption, allocate ctx s/r
area in VRAM instead of system memory, and migrate it back
to system memory when VRAM is full.
Signed-off-by: Eric Huang <jinhuieric.huang@amd.com>
Change-Id: If775782027188dbe84b6868260e429373675434c
The KFD no longer allow debug ops that modify HW state prior to
trap activation so permit bump in major version.
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Change-Id: I072d3998b7b043df9a67f0f6762b0afdfa9382c6
Non-paged allocation for queue memory necessary for binding wptr to
GART. Required to support usermode queue oversubscription with MES on
GFX11.
Change ensures queue memory does not specify ATS.
Signed-off-by: Graham Sider <Graham.Sider@amd.com>
Change-Id: I10b23b0205c90dad902c711a88cfb5e9b4979617
System Management Interface event is read from anonymous file handle,
this helper wrap the ioctl interface to get anonymous file handle for
GPU nodeid.
Define SMI event IDs, event triggers, copy the same value from
kfd_ioctl.h to avoid translation.
Change-Id: I5c8ba5301473bb3b80bb4e2aa33a9f675bedb001
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Add KFDGpuID to HsaNodeProperties to return gpu_id to upper layer,
gpu_id is hash ID generated by KFD to distinguish GPUs on the system.
ROCr and ROCProfiler will use gpu_id to analyze SMI event message.
Change-Id: I6eabe6849230e04120674f5bc55e6ea254a532d6
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
This env variable sets the max VA alignment order size as
"PAGE_SIZE * 2^alignment order" during mapping. By default the order
size is set to 9(2MB).
Signed-off-by: Alex Sierra <alex.sierra@amd.com>
Change-Id: I01ae4e0963f4d21c7c367464e60f865bc58d7fac
New API function to report available memory per GPU
Signed-off-by: Daniel Phillips <Daniel.Phillips@amd.com>
Change-Id: I63c1e4ca0020c657977ab3635947ab0ed0a81440
Tested on Talos II with Vega 64
POWER systems allocate NUMA nodes on multiples of 8 to allow CPU
onlining / offlining
Set the correct NUMA mask bits when requesting node-bound memory
allocations
This is a cleanup/squash/rebase of:
https://github.com/RadeonOpenCompute/ROCT-Thunk-Interface/pull/47
Change-Id: Id4af6dff7e66e9d464d6b17a1e99087eb3ac8e51
Signed-off-by: Jeremy Newton <Jeremy.Newton@amd.com>
Import the latest version from the kernel tree.
Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Change-Id: If5f998ad55085ebd5020adaa382181204d834e3e
Currently, context save area size passed to KFD includes the
size of the debug area. Change this to report the actual size
of the context save area to KFD.
Signed-off-by: Mukul Joshi <mukul.joshi@amd.com>
Change-Id: I5d440ae802255a97ade046775f6a000bae79d5d5
Mapped memory areas become invalid after fork, and the child process is
required to remap the memory areas after a fork. So we mark these device
memory mappings with MADV_DONTFORK so that they are removed from the
child process after fork.
This was causing some issues when doing CRIU checkpoint/restore because
CRIU and amdgpu_plugin were not able to handle these mappings.
Change-Id: I50eb334aecea6dab7522d94da0273adcf4fb1ce0
Signed-off-by: David Yat Sin <david.yatsin@amd.com>
The AMD compiler team has confirmed that they expect gfx90c
to be gfx90c, with a major/minor/stepping of 9, 0, and 12
respectively. It appears that there is a typo in the libhsakmt
topology information that lists this part as gfx902. This patch
fixes the issue.
Signed-off-by: Joseph Greathouse <Joseph.Greathouse@amd.com>
Change-Id: I6f907a7aa6f190b12aba8bb4210c7b341b3c720b
The return code is just -1 if any error occurs. To detect debugger
unavailable we need to check the actual ioctl error code.
Change-Id: I8a294c754196aec916809497ec8e810da2f072b8
Signed-off-by: Sean Keely <Sean.Keely@amd.com>
For userptr, after taking aperture lock, decrease registration_count and
ensure object registration_count equal to 0 to release KFD and thunk
object.
Move decrementing of registration count from fmm_deregister_memory into
__fmm_release to avoid a race condition when dropping the aperture lock
in fmm_deregister_memory.
Change-Id: I5381fa6b8a77a1516af2554e5174e91969c338c4
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
to faciliate debugging, print errno when queue destroy fail
current log give very little information when fail:
[ RUN ] KFDQMTest.AllSdmaQueues
[ ] Regular SDMA engines number: 1 SDMA queues per engine: 2
[ OK ] KFDQMTest.AllSdmaQueues (11 ms)
[ RUN ] KFDQMTest.AllXgmiSdmaQueues
[ ] XGMI SDMA engines number: 0 SDMA queues per engine: 2
[ OK ] KFDQMTest.AllXgmiSdmaQueues (6 ms)
[ RUN ] KFDQMTest.AllQueues
/home/foreman/build/hsakmt-roct-amdgpu-1.0.9.40500/sources/libhsakmt/tests/kfdtest/src/KFDQMTest.cpp:381: Failure
Value of: (cpQueues[i].Destroy())
Actual: 1
Expected: HSAKMT_STATUS_SUCCESS
Which is: 0
Signed-off-by: Yifan Zhang <yifan1.zhang@amd.com>
Change-Id: I5b1b5b616a5fd7ff198360c893a7aeed685022bd
After calling ioctl to create userptr obj, take aperture lock, check if
there is same userptr obj created after finding object, to catch the
race that multiple threads register same userptr to multiple GPUs.
If same userptr obj exist, then increase userptr registeration_count,
and free the newly create obj.
Change-Id: I63ae3a4f54da8aedd11c124d8d53ebe727b8203a
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
David Belanger
Eric Huang
Alex Sierra
Felix Kuehling
David Yat Sin
David Francis
Graham Sider
Philip Yang
Aaron Liu
jie1zhan
Mukul Joshi
kent.russell@amd.com
Lang Yu
Jonathan Kim
Daniel Phillips
Timothy Pearson
Joseph Greathouse
Sean Keely
Yifan Zhang