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>
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>
for SRIOV the SOS fw of PSP is loaded in hypervisor thus
guest won't tell the version of it, and judging feature by
reading the sos fw version in guest side is completely wrong
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>
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>
It's incorrect to do soft reset for SRIOV, when GFX
hang the WREG would stuck there becuase it goes KIQ way.
the GPU reset counter is incorrect: always increase twice
for each timedout
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>
this patch actually refactor mailbox implmentations, and
all below changes are needed together to fix all those mailbox
handshake issues exposured by heavey TDR test.
1)refactor all mailbox functions based on byte accessing for mb_control
reason is to avoid touching non-related bits when writing trn/rcv part of
mailbox_control, this way some incorrect INTR sent to hypervisor
side could be avoided, and it fixes couple handshake bug.
2)trans_msg function re-impled: put a invalid
logic before transmitting message to make sure the ACK bit is in
a clear status, otherwise there is chance that ACK asserted already
before transmitting message and lead to fake ACK polling.
(hypervisor side have some tricks to workaround ACK bit being corrupted
by VF FLR which hase an side effects that may make guest side ACK bit
asserted wrongly), and clear TRANS_MSG words after message transferred.
3)for mailbox_flr_work, it is also re-worked: it takes the mutex lock
first if invoked, to block gpu recover's participate too early while
hypervisor side is doing VF FLR. (hypervisor sends FLR_NOTIFY to guest
before doing VF FLR and sentds FLR_COMPLETE after VF FLR done, and
the FLR_NOTIFY will trigger interrupt to guest which lead to
mailbox_flr_work being invoked)
This can avoid the issue that mailbox trans msg being cleared by its VF FLR.
4)for mailbox_rcv_irq IRQ routine, it should only peek msg and schedule
mailbox_flr_work, instead of ACK to hypervisor itself, because FLR_NOTIFY
msg sent from hypervisor side doesn't need VF's ACK (this is because
VF's ACK would lead to hypervisor clear its trans_valid/msg, and this
would cause handshake bug if trans_valid/msg is cleared not due to
correct VF ACK but from a wrong VF ACK like this "FLR_NOTIFY" one)
This fixed handshake bug that sometimes GUEST always couldn't receive
"READY_TO_ACCESS_GPU" msg from hypervisor.
5)seperate polling time limite accordingly:
POLL ACK cost no more than 500ms
POLL MSG cost no more than 12000ms
POLL FLR finish cost no more than 500ms
6) we still need to set adev into in_gpu_reset mode after we received
FLR_NOTIFY from host side, this can prevent innocent app wrongly succesed
to open amdgpu dri device.
FLR_NOFITY is received due to an IDLE hang detected from hypervisor side
which indicating GPU is already die in this VF.
v2:
use MACRO as the offset of mailbox_control register
don't test if NOTIFY_CMPL event in rcv_msg since it won't
recieve that message anymore
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Pixel Ding <Pixel.Ding@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
mailbox registers can be accessed with a byte boundry according
to BIF team, so this patch prepares register byte access
and will be used by following patches.
Actually, for mailbox registers once the byte field is touched even not changed,
the mailbox behaves, so we need the byte width accessing to those sort of regs.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Pixel Ding <Pixel.Ding@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
sometimes GPU is switched to other VFs and won't swich
back soon, so the kiq reg access will not signal within
a short period, instead of busy waiting a long time(MAX_KEQ_REG_WAIT)
and returning TMO we can istead sleep 5ms and try again
later (non irq context)
And since the waiting in kiq_r/weg is busy wait, so MAX_KIQ_REG_WAIT
shouldn't set to a long time, set it to 10ms is more appropriate.
if gpu already in reset state, don't retry the KIQ reg access
otherwise it would always hang because KIQ was already die usually.
v2:
replace schedule() with msleep() for the wait
v3:
use while loop for the wait repeating
use macros for the sleep period
more description for it
v4:
drop unused variable
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com
Reviewed-by: Pixel Ding <Pixel.Ding@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
1)create a routine "handle_vram_lost" to do the vram
recovery, and put it into amdgpu_device_reset/reset_sriov,
this way no need of the extra paramter to hold the
VRAM LOST information and the related macros can be removed.
3)show vram_recover failure if time out, and set TMO equal to
lockup_timeout if vram_recover is under SRIOV runtime mode.
4)report error if any ip reset failed for SR-IOV
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
fix:
should do right shift on wb before clearing
cleanups:
1,should memset all wb buffer
2,set max wb number to 128 (total 4KB) is big enough
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
WB_FREE should be put after all engines's hw_fini
done, otherwise the invalid wptr/rptr_addr would still
be used by engines which trigger abnormal bugs.
This fixes couple DMAR reading error in host side for SRIOV
after guest kmd is unloaded.
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>
otherwise there will be DMAR reading error comes out from CP since
GFX is still alive and CPC's WPTR_POLL is still enabled, which would
lead to DMAR read error.
fix:
we can hault CPG after hw_fini, but cannot halt CPC becaues KIQ
stil need to be alive to let RLCV invoke, but its WPTR_POLL could
be disabled.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
should use bo_create_kernel instead of split to two
function that create and pin the SA bo
issue:
before this patch, there are DMAR read error in host
side when running SRIOV test, the DMAR address dropped
in the range of SA bo.
fix:
after this cleanups of SA init and fini, above DMAR
eror gone.
v2:
keep sa_bo's fini instead of suspend, to keep
reporting error
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
issue:
sometime GFX/MM ib test hit timeout under SRIOV env, root cause
is that engine doesn't come back soon enough so the current
IB test considered as timed out.
fix:
for SRIOV GFX IB test wait time need to be expanded a lot during
SRIOV runtimei mode since it couldn't really begin before GFX engine
come back.
for SRIOV MM IB test it always need more time since MM scheduling
is not go together with GFX engine, it is controled by h/w MM
scheduler so no matter runtime or exclusive mode MM IB test
always need more time.
v2:
use ring type instead of idx to judge
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
SRIOV doesn't give VF cg/pg feature so the MM's idle_work
is skipped for SR-IOV
v2:
remove superfluous changes
since idle_work is not scheduled for SR-IOV so the condition
check for SR-IOV inside idle_work also can be dropped
v3:
drop the SRIOV check in amdgpu_vce/uvd_suspend
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
WB_FREE should be put after all engines's hw_fini
done, otherwise the invalid wptr/rptr_addr would still
be used by engines which trigger abnormal bugs.
This fixes couple DMAR reading error in host side for SRIOV
after guest kmd is unloaded.
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>
two reasons to switch SCRATCH reg method to WB method:
1)Because when doing IB test we don't want to involve KIQ health
status affect, and since SCRATCH register access is go through
KIQ that way GFX IB test would failed due to KIQ fail.
2)acccessing SCRATCH register cost much more time than WB method
because SCRATCH register access runs through KIQ which at least could
begin after GPU world switch back to current Guest VF
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
issue:
under SR-IOV sometimes the iB test will fail on
gfx ring
fix:
with cond_exec inserted in RB the gfx engine would
skip part packets if RLCV issue PREEMPT on gfx engine
if gfx engine is prior to COND_EXEC packet, this is
okay for regular command from UMD, but for the ib test
since the whole dma format doesn't support PREEMPT
so must remove the COND_EXEC from it.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
fix:
should do right shift on wb before clearing
cleanups:
1,should memset all wb buffer
2,set max wb number to 128 (total 4KB) is big enough
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
shrink_pages is in unit of Order after ttm_page_pool_free,
but it is used by nr_free in next round so need change
it into native page unit
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Roger He <Hongbo.He@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
this VRAM evict is not needed and also cost 2seconds
to finish because the IRQ is software side disabled
before it.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
shrink_pages is in unit of Order after ttm_page_pool_free,
but it is used by nr_free in next round so need change
it into native page unit
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Roger He <Hongbo.He@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
found RING0 test fail after S3 resume regression, which is
introduced by 1cfd8e237f0318e330190ac21d63c58ae6a1f66c
Because after suspend VRAM will be cleared, so driver must
unpin the GART table(resident in VRAM) during suspend so it
can be evicted to system ram and must correspondingly pin it
during resume so the GART table could be restored to VRAM.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
virtual DCE Timer structure is already released
after its sw_fini(), so we need to cancel the
its Timer in hw_fini() otherwise the Timer canceling
is missed.
v2:
use for loop and num_crtc to replace original code
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>
NULL pointer is because original logic will step into
set_pde_pte() even after the gart.ptr is freed due to
there are twice gart_unbind() on all gart area.
also, there are other minor fixes:
1,since gart_init only create dummy page, the corresponding
gart_fini shouldn't do more like unbinding all GART, this is
unnecessary because in driver fini stage all GART unbinding
had already been done during each IP's SW_FINI (GMC's
SW_FINI is the last one called), so remove the step
for the GART unbinding in gart_fini().
2,gart_fini() is already invoked during each GMC IP's gart_fini
routine,e.g. gmc_vx_0_gart_fini(), so no need to manually
call it during ttm_fini().
3,amdgpu_gem_force_release() should be put ahead of
amdgpu_vm_manager_fini()
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
for gart_ram_alloc/free, they are never used in driver thus
ripe them out totally.
for gart_vram_pin/unpin, they are not needed becuase we can
use bo_creat_kernel/free to replace the original manual way
in the gart_vram_alloc/free, thus gart_vram_pin/unpin can
also be riped out.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
if app close CTX right after IB submit, gpu recover
will fail to find out the entity behind this guilty
job thus lead to no job skipping for this guilty job.
to fix this corner case just move the increasement of
job->karma out of the entity iteration.
v2:
only do karma increasment if bad->s_priority != KERNEL
because we always consider KERNEL job be correct and always
want to recover an unfinished kernel job (sometimes kernel
job is interrupted by VF FLR or other GPU hang event)
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-By: Xiangliang Yu <Xiangliang.Yu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
1,no sriov check since gpu recover is unified
2,need CPU_ACCESS_REQUIRED flag for VRAM if SRIOV
because otherwise after following PIN the first allocated
VRAM bo is wasted due to some TTM mgr reason.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
since now gpu reset is unified with gpu_recover
for both bare-metal and SR-IOV:
1)rename in_sriov_reset to in_gpu_reset
2)move lock_reset from adev->virt to adev
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
1,new imple names amdgpu_gpu_recover which gives more hint
on what it does compared with gpu_reset
2,gpu_recover unify bare-metal and SR-IOV, only the asic reset
part is implemented differently
3,gpu_recover will increase hang job karma and mark its entity/context
as guilty if exceeds limit
V2:
4,in scheduler main routine the job from guilty context will be immedialy
fake signaled after it poped from queue and its fence be set with
"-ECANCELED" error
5,in scheduler recovery routine all jobs from the guilty entity would be
dropped
6,in run_job() routine the real IB submission would be skipped if @skip parameter
equales true or there was VRAM lost occured.
V3:
7,replace deprecated gpu reset, use new gpu recover
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
jobs are skipped under two cases
1)when the entity behind this job marked guilty, the job
poped from this entity's queue will be dropped in sched_main loop.
2)in job_recovery(), skip the scheduling job if its karma detected
above limit, and also skipped as well for other jobs sharing the
same fence context. this approach is becuase job_recovery() cannot
access job->entity due to entity may already dead.
v2:
some logic fix
v3:
when entity detected guilty, don't drop the job in the poping
stage, instead set its fence error as -ECANCELED
in run_job(), skip the scheduling either:1) fence->error < 0
or 2) there was a VRAM LOST occurred on this job.
this way we can unify the job skipping logic.
with this feature we can introduce new gpu recover feature.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
reset_counter marks the reset counter number once the context
is created, shouldn't be changed due to query.
To keep U/K interface on the ctx_query and keep ctx's reset_counter
logic compatible with GPU RESET feature, now use another var named
"reset_counter_query" to replace the original checked & updated in
amdgpu_ctx_query.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
merge the setting guilty on context into this function
to avoid implement extra routine.
v2:
go through entity list and compare the fence_ctx
before operate on the entity, otherwise the entity
may be just a wild pointer
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Chunming Zhou <David1.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
this way the real interested guilty is connected to entity->guilty
pointer, and we can use entity->pointer later in gpu recovery procedure
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Chunming Zhou <David1.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
this member will be used later, it will points to
the real var inside of context and CS_SUBMIT & gpu schdduler
can decide if skip a job depends on context->guilty or *entity->guilty
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Chunming Zhou <David1.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
since gpu_scheduler source domain cannot access amdgpu variable
so need create the hang_limit membewr for sched, and it can
refer it for the upcoming GPU RESET patches
v2:
make hang_limit a parameter of sched_init()
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Chunming Zhou <David1.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
this can fix the memory leak under the case that not all
BO are freed during "takedown" stage, because originally
it blocks following kfree on mgr.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
1,it should not work on non-SR-IOV case
2,the NO_VBIOS error is incorrect, should
handle it under detect_sriov_bios.
3,wrap the whole detect_sriov_bios with sriov check
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>
with current WB usage we only use 57 slots, so 512
is extreamly sufficient, and reduce to 512 can
make WB fit into one page.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
this fix memory leak due to request_firmware after driver
unloaded
v2:
release gmc firmware for gmc6/7/8 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>
this way after KIQ MQD released in drv unloading, CPC
can still let KIQ access this MQD thus RLCV SAVE_VF
will not fail
v2:
always use VRAM domain for KIQ MQD no matter BM or SRIOV
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
GPU reset will require all hw doing hw_init thus
ucode_init_bo will be invoked again, which lead to
memory leak
skip the fw_buf allocation during sriov gpu reset to avoid
memory leak.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
otherwise a gpu hang will make application couldn't be killed
under timedout=0 mode
v2:
Fix memoryleak job/job->s_fence issue
unlock mn
remove the ERROR msg after waiting being interrupted
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
RLC need CSB registers initiated under SRIOV during world switch
otherwise the clear state buffer behav will not be recovered to
current VF scheme after switch back
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
bo_free on csa is too late to put in amdgpu_fini because that
time ttm is already finished,
Move it earlier to avoid the page fault.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Signed-off-by: Horace Chen <horace.chen@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
FRAME_CONTROL(begin) is needed for vega10 due to ucode logic change,
it can fix some CTS random fail under gfx preemption enabled mode.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
currently in_reset is only used in sriov gpu reset, and it
will be used for other non-gfx hw component later, like
PSP, so move it from gfx to adev and rename to in_sriov_reset
make more sense.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
we observe abnormal number from:
/sys/devices/virtual/drm/amdttm/buffer_objects/bo_count
bo_count is atomic_inc which is "int" type,
shouldn't explicitly turn it to unsigned long.
Signed-off-by: Monk Liu <monk.liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
we observe abnormal number from:
/sys/devices/virtual/drm/amdttm/buffer_objects/bo_count
bo_count is atomic_inc which is "int" type,
shouldn't explicitly turn it to unsigned long.
Signed-off-by: Monk Liu <monk.liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
for SR-IOV, we must keep the pipeline-sync in the protection
of COND_EXEC, otherwise the command consumed by CPG is not
consistent when world switch triggerd, e.g.:
world switch hit and the IB frame is skipped so the fence
won't signal, thus CP will jump to the next DMAframe's pipeline-sync
command, and it will make CP hang foever.
after pipelin-sync moved into COND_EXEC the consistency can be
guaranteed
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
because if the fence is really signaled, it could already
released so the fence pointer is a wild pointer, but if
we use job->base.node we are safe because job will not
be released untill amdgpu_job_timedout finished.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
1,TDR will kickout guilty job if it hang exceed the threshold
of the given one from kernel paramter "job_hang_limit", that
way a bad command stream will not infinitly cause GPU hang.
by default this threshold is 1 so a job will be kicked out
after it hang.
2,if a job timeout TDR routine will not reset all sched/ring,
instead if will only reset on the givn one which is indicated
by @job of amdgpu_sriov_gpu_reset, that way we don't need to
reset and recover each sched/ring if we already know which job
cause GPU hang.
3,unblock sriov_gpu_reset for AI family.
V2:
1:put kickout guilty job after sched parked.
2:since parking scheduler prior to kickout already occupies a
while, we can do last check on the in question job before
doing hw_reset.
TODO:
1:when a job is considered as guilty, we should mark some flag
in its fence status flag, and let UMD side aware that this
fence signaling is not due to job complete but job hang.
2:if gpu reset cause all video memory lost, we need introduce
a new policy to implement TDR, like drop all jobs not yet
signaled, and all IOCTL on this device will return ERROR
DEVICE_LOST.
this will be implemented later.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
because we don't want to do sriov-gpu-reset under certain
cases, so just split those two funtion and don't invoke
sr-iov one from bare-metal one.
V2:
remove debugfs_gpu_reset routine on SRIOV case.
Signed-off-by: Monk Liu <Monk.Liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Monk Liu