Since the introduction of A6xx support, we've been enabling the CX GMU
power counter 0 in a bit of a weird spot. Move it to hw_init so that
GMU wrapper GPUs can reuse the same code paths. As a bonus, this order
makes it easier to compare mainline and downstream register access traces.
Signed-off-by: Konrad Dybcio <konrad.dybcio@linaro.org>
Patchwork: https://patchwork.freedesktop.org/patch/542765/
Signed-off-by: Rob Clark <robdclark@chromium.org>
Currently we're only deasserting REG_A6XX_RBBM_GBIF_HALT, but we also
need REG_A6XX_GBIF_HALT to be set to 0.
This is typically done automatically on successful GX collapse, but in
case that fails, we should take care of it.
Also, add a memory barrier to ensure it's gone through before jumping
to further initialization.
Reviewed-by: Dmitry Baryshkov <dmitry.baryshkov@linaro.org>
Signed-off-by: Konrad Dybcio <konrad.dybcio@linaro.org>
Patchwork: https://patchwork.freedesktop.org/patch/542760/
Signed-off-by: Rob Clark <robdclark@chromium.org>
Introduce a6xx_gpu_sw_reset() in preparation for adding GMU wrapper
GPUs and reuse it in a6xx_gmu_force_off().
This helper, contrary to the original usage in GMU code paths, adds
a readback+delay sequence to ensure that the reset is never deasserted
too quickly due to e.g. OoO execution going crazy.
Signed-off-by: Konrad Dybcio <konrad.dybcio@linaro.org>
Patchwork: https://patchwork.freedesktop.org/patch/542758/
Signed-off-by: Rob Clark <robdclark@chromium.org>
This function is responsible for telling the GPU to halt transactions
on all of its relevant buses, drain them and leave them in a predictable
state, so that the GPU can be e.g. reset cleanly.
Move the function to a6xx_gpu.c, remove the static keyword and add a
prototype in a6xx_gpu.h to accomodate for the move.
Reviewed-by: Akhil P Oommen <quic_akhilpo@quicinc.com>
Signed-off-by: Konrad Dybcio <konrad.dybcio@linaro.org>
Patchwork: https://patchwork.freedesktop.org/patch/542762/
Signed-off-by: Rob Clark <robdclark@chromium.org>
This seems to have existed for ever but is now more apparant after
commit 9bff18d134 ("drm/ttm: use per BO cleanup workers")
My analysis: two threads are running, one in the irq signalling the
fence, in dma_fence_signal_timestamp_locked, it has done the
DMA_FENCE_FLAG_SIGNALLED_BIT setting, but hasn't yet reached the
callbacks.
The second thread in nouveau_cli_work_ready, where it sees the fence is
signalled, so then puts the fence, cleanups the object and frees the
work item, which contains the callback.
Thread one goes again and tries to call the callback and causes the
use-after-free.
Proposed fix: lock the fence signalled check in nouveau_cli_work_ready,
so either the callbacks are done or the memory is freed.
Reviewed-by: Karol Herbst <kherbst@redhat.com>
Fixes: 11e451e740 ("drm/nouveau: remove fence wait code from deferred client work handler")
Cc: stable@vger.kernel.org
Signed-off-by: Dave Airlie <airlied@redhat.com>
Link: https://lore.kernel.org/dri-devel/20230615024008.1600281-1-airlied@gmail.com/
drm/i915 feature pull #2 for v6.5:
Features and functionality:
- Meteorlake PM demand (Vinod, Mika)
- Switch to dedicated workqueues to stop using flush_scheduled_work() (Luca)
Refactoring and cleanups:
- Move display runtime init under display/ (Matt)
- Async flip error message clarifications (Arun)
Fixes:
- Remove 10bit gamma on desktop gen3 parts, they don't support it (Ville)
- Fix driver probe error handling if driver creation fails (Matt)
- Fix all -Wunused-but-set-variable warnings, and enable it for i915 (Jani)
- Stop using edid_blob_ptr (Jani)
- Fix log level for "CDS interlane align done" (Khaled)
- Fix an unnecessary include prefix (Matt)
Merges:
- Backmerge drm-next to sync with drm-intel-gt-next (Jani)
Signed-off-by: Dave Airlie <airlied@redhat.com>
From: Jani Nikula <jani.nikula@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/87o7lnpxz2.fsf@intel.com
[Why]
The sequence for collecting down_reply from source perspective should
be:
Request_n->repeat (get partial reply of Request_n->clear message ready
flag to ack DPRX that the message is received) till all partial
replies for Request_n are received->new Request_n+1.
Now there is chance that drm_dp_mst_hpd_irq() will fire new down
request in the tx queue when the down reply is incomplete. Source is
restricted to generate interveleaved message transactions so we should
avoid it.
Also, while assembling partial reply packets, reading out DPCD DOWN_REP
Sideband MSG buffer + clearing DOWN_REP_MSG_RDY flag should be
wrapped up as a complete operation for reading out a reply packet.
Kicking off a new request before clearing DOWN_REP_MSG_RDY flag might
be risky. e.g. If the reply of the new request has overwritten the
DPRX DOWN_REP Sideband MSG buffer before source writing one to clear
DOWN_REP_MSG_RDY flag, source then unintentionally flushes the reply
for the new request. Should handle the up request in the same way.
[How]
Separete drm_dp_mst_hpd_irq() into 2 steps. After acking the MST IRQ
event, driver calls drm_dp_mst_hpd_irq_send_new_request() and might
trigger drm_dp_mst_kick_tx() only when there is no on going message
transaction.
Changes since v1:
* Reworked on review comments received
-> Adjust the fix to let driver explicitly kick off new down request
when mst irq event is handled and acked
-> Adjust the commit message
Changes since v2:
* Adjust the commit message
* Adjust the naming of the divided 2 functions and add a new input
parameter "ack".
* Adjust code flow as per review comments.
Changes since v3:
* Update the function description of drm_dp_mst_hpd_irq_handle_event
Changes since v4:
* Change ack of drm_dp_mst_hpd_irq_handle_event() to be an array align
the size of esi[]
Signed-off-by: Wayne Lin <Wayne.Lin@amd.com>
Reviewed-by: Lyude Paul <lyude@redhat.com>
Acked-by: Jani Nikula <jani.nikula@intel.com>
Cc: stable@vger.kernel.org
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
If hmm_range_fault returns -EBUSY, we should call hmm_range_fault again
to validate the remaining pages. On one system with NUMA auto balancing
enabled, hmm_range_fault takes 6 seconds for 1GB range because CPU
migrate the range one page at a time. To be safe, increase timeout value
to 1 second for 128MB 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>
Update user space last_event_age when event age is enabled.
It is only for KFD_EVENT_TYPE_SIGNAL which is checked by user space.
Signed-off-by: James Zhu <James.Zhu@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
drm_sched_entity_add_dependency_cb ignores the scheduled fence and return false.
If entity's dependency is a scheduler error fence and drm_sched_stop is called
due to TDR, drm_sched_entity_pop_job will wait for the dependency infinitely.
[How]
Do not wait or ignore the scheduled error fence, add drm_sched_entity_wakeup
callback for the dependency with scheduled error fence.
Signed-off-by: ZhenGuo Yin <zhenguo.yin@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
UMD is not aware of entity error, and will keep submitting jobs
into the error entity.
[How]
Add entity error check when getting entity from ctx.
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: ZhenGuo Yin <zhenguo.yin@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Instead of using the VRAM lost counter add a 64bit token which indicates
if a context or job is still valid to use.
Should the VRAM be lost or the page tables need re-creation the token will
change indicating that userspace needs to act and re-create the contexts
and re-submit the work.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Luben Tuikov <luben.tuikov@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
When some problem with the updates of page tables is detected reset the
state machine of the VM and re-create all page tables from scratch.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Luben Tuikov <luben.tuikov@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Mark as experimental for now until we get closer to production
to avoid possible undesireable behavior when mixing newer
boards with older kernels.
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
PARTITION_MODE field in PARTITION_COMPUTE_STATUS register is defined as
below by firmware.
SPX = 0, DPX = 1, TPX = 2, QPX = 3, CPX = 4
Change driver definition accordingly.
Signed-off-by: Lijo Lazar <lijo.lazar@amd.com>
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: Le Ma <le.ma@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Update eccinfo table structure according to smu v13_0_0 interface.
v2: Calculate array size instead of using macro definition.
Signed-off-by: Candice Li <candice.li@amd.com>
Reviewed-by: Lijo Lazar <lijo.lazar@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The fixed commit listed in the Fixes tag below, introduced a bug in
amdgpu_ras.c::amdgpu_reserve_page_direct(), in that when introducing the new
amdgpu_umc_fill_error_record() and internally in that new function the physical
address (argument "uint64_t retired_page"--wrong name) is right-shifted by
AMDGPU_GPU_PAGE_SHIFT. Thus, in amdgpu_reserve_page_direct() when we pass
"address" to that new function, we should NOT right-shift it, since this
results, erroneously, in the page address to be 0 for first
2^(2*AMDGPU_GPU_PAGE_SHIFT) memory addresses.
This commit fixes this bug.
Cc: Tao Zhou <tao.zhou1@amd.com>
Cc: Hawking Zhang <Hawking.Zhang@amd.com>
Cc: Alex Deucher <Alexander.Deucher@amd.com>
Fixes: 400013b268 ("drm/amdgpu: add umc_fill_error_record to make code more simple")
Signed-off-by: Luben Tuikov <luben.tuikov@amd.com>
Link: https://lore.kernel.org/r/20230610113536.10621-1-luben.tuikov@amd.com
Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
The wptr needs to be incremented at at least 64 dword intervals,
use 256 to align with windows. This should fix potential hangs
with unaligned updates.
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Aaron Liu <aaron.liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
