As more and more new asics start to reuse the old device IDs before
launch, there is a need to quickly override the existing asic type
corresponding to the reused device ID through a kernel parameter. With
this, engineers no longer need to rely on local hack patches,
facilitating cooperation across teams.
Signed-off-by: Yong Zhao <Yong.Zhao@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
ras recovery_init should be called after ttm init,
bad page reserve should be put in front of gpu reset since i2c
may be unstable during gpu reset.
add cleanup for recovery_init and recovery_fini
v2: add more comment and print.
remove cancel_work_sync in recovery_init.
Signed-off-by: Tao Zhou <tao.zhou1@amd.com>
Reviewed-by: Guchun Chen <guchun.chen@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
In case of RAS error allow user configure auto system
reboot through ras_ctrl.
This is also part of the temproray work around for the RAS
hang problem.
v4: Use latest kernel API for disk sync.
Signed-off-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Problem:
Under certain conditions, when some IP bocks take a RAS error,
we can get into a situation where a GPU reset is not possible
due to issues in RAS in SMU/PSP.
Temporary fix until proper solution in PSP/SMU is ready:
When uncorrectable error happens the DF will unconditionally
broadcast error event packets to all its clients/slave upon
receiving fatal error event and freeze all its outbound queues,
err_event_athub interrupt will be triggered.
In such case and we use this interrupt
to issue GPU reset. THe GPU reset code is modified for such case to avoid HW
reset, only stops schedulers, deatches all in progress and not yet scheduled
job's fences, set error code on them and signals.
Also reject any new incoming job submissions from user space.
All this is done to notify the applications of the problem.
v2:
Extract amdgpu_amdkfd_pre/post_reset from amdgpu_device_lock/unlock_adev
Move amdgpu_job_stop_all_jobs_on_sched to amdgpu_job.c
Remove print param from amdgpu_ras_query_error_count
v3:
Update based on prevoius bug fixing patch to properly call amdgpu_amdkfd_pre_reset
for other XGMI hive memebers.
Signed-off-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
Acked-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Issue 1:
In XGMI case amdgpu_device_lock_adev for other devices in hive
was called to late, after access to their repsective schedulers.
So relocate the lock to the begining of accessing the other devs.
Issue 2:
Using amdgpu_device_ip_need_full_reset to switch the device list from
all devices in hive to the single 'master' device who owns this reset
call is wrong because when stopping schedulers we iterate all the devices
in hive but when restarting we will only reactivate the 'master' device.
Also, in case amdgpu_device_pre_asic_reset conlcudes that full reset IS
needed we then have to stop schedulers for all devices in hive and not
only the 'master' but with amdgpu_device_ip_need_full_reset we
already missed the opprotunity do to so. So just remove this logic and
always stop and start all schedulers for all devices in hive.
Also minor cleanup and print fix.
v4: Minor coding style fix.
Signed-off-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
Acked-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
for SOC15/vega10 the BACO reset & mode1 would introduce vram lost
in high end address range, current kmd's vram lost checking cannot
catch it since it only check very ahead visible frame buffer
v2:
cover NV as well
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
what we really want is a read or write that is guaranteed to be 64 bits
at a time, atomic64 operations are supported on all architectures
Signed-off-by: Tao Zhou <tao.zhou1@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
amdgpu_ip_block.status.hw for GMC wasn't set to
false on suspend during GPU reset and so on resume gmc_v9_0_resume
wasn't called.
Caused by 'drm/amdgpu: fix double ucode load by PSP(v3)'
Signed-off-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
previously the ucode loading of PSP was repreated, one executed in
phase_1 init/re-init/resume and the other in fw_loading routine
Avoid this double loading by clearing ip_blocks.status.hw in suspend or reset
prior to the FW loading and any block's hw_init/resume
v2:
still do the smu fw loading since it is needed by bare-metal
v3:
drop the change in reinit_early_sriov, just clear all block's status.hw
in the head place and set the status.hw after hw_init done is enough
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Emily Deng <Emily.Deng@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
we can simplify all those unnecessary function under
SRIOV for vega10 since:
1) PSP L1 policy is by force enabled in SRIOV
2) original logic always set all flags which make itself
a dummy step
besides,
1) the ih_doorbell_range set should also be skipped
for VEGA10 SRIOV.
2) the gfx_common registers should also be skipped
for VEGA10 SRIOV.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Emily Deng <Emily.Deng@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
When doing a GPU reset or unloading the driver, we need to
put the SMU into the apprpriate state for the re-init after
the reset or unload to reliably work.
I don't think this is necessary for BACO because the SMU actually
controls the BACO state to it needs to be active.
For suspend (S3), the asic is put into D3 so the SMU would be
powered down so I don't think we need to put the SMU into
any special state.
Reviewed-by: Evan Quan <evan.quan@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
We need to serialize access to the psp ring if there are multiple
callers at runtime.
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
When CONFIG_PERF_EVENTS is disabled, we cannot compile the pmu
portion of the amdgpu driver:
drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c:48:38: error: no member named 'hw' in 'struct perf_event'
struct hw_perf_event *hwc = &event->hw;
~~~~~ ^
drivers/gpu/drm/amd/amdgpu/amdgpu_pmu.c:51:13: error: no member named 'attr' in 'struct perf_event'
if (event->attr.type != event->pmu->type)
~~~~~ ^
...
Use conditional compilation for this file.
Fixes: 9c7c85f7ea ("drm/amdgpu: add pmu counters")
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
GPU atomics operation depends on PCIE atomics support.
Always enable PCIE atomics ops support in case that
it hasn't been enabled.
Signed-off-by: Jack Xiao <Jack.Xiao@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[WHY]
We don't want to expose sensitive ASIC information before ASIC release.
[HOW]
Encode the soc_bounding_box in the gpu_info FW (for Linux) and read it
at driver load.
v2: fix warning when CONFIG_DRM_AMD_DC_DCN2_0 is not set (Alex)
Signed-off-by: Harry Wentland <harry.wentland@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The MCBP unit test is used to test the functionality of MCBP.
It emualtes to send preemption request and resubmit the unfinished
jobs.
v2: squash in fixes (Alex)
v3: squash in memory leak fix (Jack)
Acked-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Jack Xiao <Jack.Xiao@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
