In a consecutive packet submission, for example unmap and query status,
when CP is reading wptr caused by unmap packet doorbell ring, if in some
case CP operates slower (e.g. doorbell_mode=1) and wptr has been updated
to next packet (query status), but the query status packet content has
not been flushed to memory yet, it will cause CP fetched stalled data.
Adding mb to ensure ring buffer has been updated before updating wptr.
Also adding a mb to ensure wptr updated before doorbell ring.
Signed-off-by: Victor Zhao <Victor.Zhao@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
kfd process kref count(process->ref) is initialized to 1 by kref_init. After
it is created not need to increase its kref. Instad add kfd process kref at kfd
process mmu notifier allocation since we already decrease the kref at
free_notifier of mmu_notifier_ops, so pair them.
When user process opens kfd node multiple times the kfd process kref is
increased each time to balance with kfd node close operation.
Signed-off-by: Xiaogang Chen <xiaogang.chen@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
In kfd_procfs_show(), the sdma_activity_work_handler is a local variable
and the sdma_activity_work_handler.sdma_activity_work should initialize
with INIT_WORK_ONSTACK() instead of INIT_WORK().
Fixes: 32cb59f313 ("drm/amdkfd: Track SDMA utilization per process")
Signed-off-by: Yuan Can <yuancan@huawei.com>
Signed-off-by: Felix Kuehling <felix.kuehling@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
GFX 12 does not require a page size cap for the trap handler because
it does not require a CWSR work around like GFX 11 did.
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Reviewed-by: David Belanger <david.belanger@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The `kfd_get_cu_occupancy` function previously declared a large
`cu_occupancy` array as a local variable, which could lead to stack
overflows due to excessive stack usage. This commit replaces the static
array allocation with dynamic memory allocation using `kcalloc`,
thereby reducing the stack size.
This change avoids the risk of stack overflows in kernel space, in
scenarios where `AMDGPU_MAX_QUEUES` is large. The allocated memory is
freed using `kfree` before the function returns to prevent memory
leaks.
Fixes the below with gcc W=1:
drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_process.c: In function ‘kfd_get_cu_occupancy’:
drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_process.c:322:1: warning: the frame size of 1056 bytes is larger than 1024 bytes [-Wframe-larger-than=]
322 | }
| ^
Fixes: 6ae9e1aba9 ("drm/amdkfd: Update logic for CU occupancy calculations")
Cc: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Cc: Felix Kuehling <felix.kuehling@amd.com>
Cc: Christian König <christian.koenig@amd.com>
Cc: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Srinivasan Shanmugam <srinivasan.shanmugam@amd.com>
Suggested-by: Mukul Joshi <mukul.joshi@amd.com>
Reviewed-by: Mukul Joshi <mukul.joshi@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Host drivers can create partial hives per guest by disabling xgmi sharing
between certain peers in the main hive.
Typically, these partial hives are fully connected per guest session.
In the event that the host makes a mistake by adding a non-shared node
to a guest session, have the KFD reflect sharing disabled by severing
the IO link.
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Tested-by: James Yao <yiqing.yao@amd.com>
Reviewed-by: Harish Kasiviswanathan <harish.kasiviswanathan@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
as the adding of mb() should be sufficient in function unmap_queues_cpsch,
remove the add of volatile type as recommended
Signed-off-by: Victor Zhao <Victor.Zhao@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
make sure KFD_FENCE_INIT write to fence_addr before pm_send_query_status
called, to avoid qcm fence timeout caused by incorrect ordering.
Signed-off-by: Victor Zhao <Victor.Zhao@amd.com>
Reviewed-by: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Process device data pdd->vram_usage is read by rocm-smi via sysfs, this
is currently missing the svm_bo usage accounting, so "rocm-smi
--showpids" per process VRAM usage report is incorrect.
Add pdd->vram_usage accounting when svm_bo allocation and release,
change to atomic64_t type because it is updated outside process mutex
now.
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>
Add new SMI event to report the dropped event count.
When the event kfifo is full, drop count is not zero, or no enough space
left to store the event message, increase drop count.
After reading event out from kfifo, if event was dropped, drop_count is
not zero, generate a dropped event record and reset drop count to zero.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: James Zhu <James.Zhu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
SMI event fifo size 1KB was enough to report GPU vm fault or reset
event, but could drop the more frequent SVM migration events. Increase
kfifo size to 8KB to store about 100 migrate events, less chance to drop
the migrate events if lots of migration happened in the short period of
time. Add KFD prefix to the macro name.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: James Zhu <James.Zhu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
If page migration failed, also output migrate end event to match with
migrate start event, with failure error_code added to the end of the
migrate message macro. This will not break uAPI because application uses
old message macro sscanf drop and ignore the error_code.
Output GPU page fault restore end event if migration failed.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: James Zhu <James.Zhu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Currently, the code uses the IH_VMID_X_LUT register to map
a queue's vmid to the corresponding PASID. This logic is racy
since CP can update the VMID-PASID mapping anytime especially
when there are more processes than number of vmids. Update the
logic to calculate CU occupancy by matching doorbell offset of
the queue with valid wave counts against the process's queues.
Signed-off-by: Mukul Joshi <mukul.joshi@amd.com>
Reviewed-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
We were removing the kernfs entry for queue info before checking if the
queue could be destroyed. If it failed to get destroyed (e.g. during
some GPU resets), then we would try to delete it later during pqm
teardown, but the file was already removed. This led to a kernel WARN
trying to remove size, gpuid and type. Move the remove to after the
destroy check.
Signed-off-by: Kent Russell <kent.russell@amd.com>
Reviewed-by: Jonathan Kim <jonathan.kim@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Instead of trying to use close_fd() on failure exits, just have
criu_get_prime_handle() store the file reference without inserting
it into descriptor table.
Then, once the callers are past the last failure exit, they can go
and either insert all those file references into the corresponding
slots of descriptor table, or drop all those file references and
free the unused descriptors.
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
To avoid memory leaks, release q_extra_data when exiting the restore queue.
v2: Correct the proto (Alex)
Signed-off-by: Jesse Zhang <jesse.zhang@amd.com>
Reviewed-by: Tim Huang <tim.huang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Document how to use SMI system management interface to enable and
receive SVM events. Document SVM event triggers.
Define SVM events message string format macro that could be used by user
mode for sscanf to parse the event. Add it to uAPI header file to make
it obvious that is changing uAPI in future.
No functional changes.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: James Zhu <James.Zhu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
If a queue is being destroyed but causes a HWS hang on removal, the KFD
may issue an unnecessary gpu reset if the destroyed queue can be fixed
by a queue reset.
This is because the queue has been removed from the KFD's queue list
prior to the preemption action on destroy so the reset call will fail to
match the HQD PQ reset information against the KFD's queue record to do
the actual reset.
To fix this, deactivate the queue prior to preemption since it's being
destroyed anyways and remove the queue from the KFD's queue list after
preemption.
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>
Enables users to update SVM's default granularity, used in
buffer migration and handling of recoverable page faults.
Param value is set in terms of log(numPages(buffer)),
e.g. 9 for a 2 MIB buffer
Signed-off-by: Ramesh Errabolu <Ramesh.Errabolu@amd.com>
Reviewed-by: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
ih1 is not initialized for APUs. Don't drain it or NULL pointer
error will be triggered.
Fixes: 6ef29715ac ("drm/amdkfd: Change kfd/svm page fault drain handling")
Signed-off-by: Yifan Zhang <yifan1.zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
When app unmap vm ranges(munmap) kfd/svm starts drain pending page fault and
not handle any incoming pages fault of this process until a deferred work item
got executed by default system wq. The time period of "not handle page fault"
can be long and is unpredicable. That is advese to kfd performance on page
faults recovery.
This patch uses time stamp of incoming page fault to decide to drop or recover
page fault. When app unmap vm ranges kfd records each gpu device's ih ring
current time stamp. These time stamps are used at kfd page fault recovery
routine.
Any page fault happened on unmapped ranges after unmap events is application
bug that accesses vm range after unmap. It is not driver work to cover that.
By using time stamp of page fault do not need drain page faults at deferred
work. So, the time period that kfd does not handle page faults is reduced
and can be controlled.
Signed-off-by: Xiaogang.Chen <Xiaogang.Chen@amd.com>
Reviewed-by: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Traditional utcl2 fault_status polling does not
work in SRIOV environment. The polling of fault
status register from guest side will be dropped
by hardware.
Driver should switch to check utcl2 interrupt
source id to identify utcl2 poison event. It is
set to 1 when poisoned data interrupts are
signaled.
v2: drop the unused local variable (Tao)
Signed-off-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: Tao Zhou <tao.zhou1@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Based on the recommendation of MEC FW, update BadOpcode interrupt
handling by unmapping all queues, removing the queue that got the
interrupt from scheduling and remapping rest of the queues back when
using MES scheduler. This is done to prevent the case where unmapping
of the bad queue can fail thereby causing a GPU reset.
Signed-off-by: Mukul Joshi <mukul.joshi@amd.com>
Acked-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
When amdgpu enable enforce_isolation, KFD enables single-process mode in
HWS and sets exec_cleaner_shader bit in MAP_PROCESS.
Signed-off-by: Amber Lin <Amber.Lin@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Provide amdgpu_amdkfd_stop_sched() for amdgpu to stop KFD scheduling
compute work on HIQ. amdgpu_amdkfd_start_sched() resumes the scheduling.
When amdgpu_amdkfd_stop_sched is called, KFD will unmap queues from
runlist. If users send ioctls to KFD to create queues, they'll be added
but those queues won't be mapped to runlist (so not scheduled) until
amdgpu_amdkfd_start_sched is called.
v2: fix build (Alex)
Signed-off-by: Amber Lin <Amber.Lin@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Add helper function kfd_queue_unreference_buffers to reduce queue buffer
refcount, separate it from release queue buffers.
Because it is circular locking to hold dqm_lock to take vm lock,
kfd_ioctl_destroy_queue should take vm lock, unreference queue buffers
first, but not release queue buffers, to handle error in case failed to
hold vm lock. Then hold dqm_lock to remove queue from queue list and
then release queue buffers.
Restore process worker restore queue hold dqm_lock, will always find
the queue with valid queue buffers.
v2 (Felix):
- renamed kfd_queue_unreference_buffer(s) to kfd_queue_unref_bo_va(s)
- added two FIXME comments for follow up
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Signed-off-by: Felix Kuehling <felix.kuehling@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
When users dynamically set the partition mode through sysfs writes,
this can lead to a double lock situation where the KFD is trying to take
the partition lock when updating the recommended SDMA engines.
Have the KFD reference its saved socket device number count instead.
Also ensure we have enough SDMA xGMI engines to report the recommended
engines in the first place.
Fixes: e06b71b231 ("drm/amdkfd: allow users to target recommended SDMA engines")
Signed-off-by: Jonathan Kim <jonathan.kim@amd.com>
Reviewed-by: Lijo Lazar <lijo.lazar@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Support per-queue reset for GFX9. The recommendation is for the driver
to target reset the HW queue via a SPI MMIO register write.
Since this requires pipe and HW queue info and MEC FW is limited to
doorbell reports of hung queues after an unmap failure, scan the HW
queue slots defined by SET_RESOURCES first to identify the user queue
candidates to reset.
Only signal reset events to processes that have had a queue reset.
If queue reset fails, fall back to GPU reset.
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>
Fixes the below if kernel config not enable HMM support
>> drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_queue.c:107:26: error:
implicit declaration of function 'svm_range_from_addr'
>> drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_queue.c:107:24: error:
assignment to 'struct svm_range *' from 'int' makes pointer from integer
without a cast [-Wint-conversion]
>> drivers/gpu/drm/amd/amdgpu/../amdkfd/kfd_queue.c:111:28: error:
invalid use of undefined type 'struct svm_range'
Fixes: b049504e21 ("drm/amdkfd: Validate user queue svm memory residency")
Reported-by: kernel test robot <lkp@intel.com>
Closes: https://lore.kernel.org/oe-kbuild-all/202407252127.zvnxaKRA-lkp@intel.com/
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>
Certain GPUs have better copy performance over xGMI on specific
SDMA engines depending on the source and destination GPU.
Allow users to create SDMA queues on these recommended engines.
Close to 2x overall performance has been observed with this
optimization.
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>
When creating KFD user compute queue, check if queue eop buffer size,
cwsr area size, ctl stack size equal to the size of KFD node
properities.
Check the entire cwsr area which may split into multiple svm ranges
aligned to granularity boundary.
Signed-off-by: Philip Yang <Philip.Yang@amd.com>
Reviewed-by: Felix Kuehling <felix.kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
