drm/amdgpu/gfx: move more common KIQ code to amdgpu_gfx.c

Lots more common stuff.

Reviewed-by: Alex Xie <AlexBin.Xie@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: move mec queue helpers to amdgpu_gfx.h

They are gfx related, not general helpers.

Reviewed-by: Alex Xie <AlexBin.Xie@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu/gfx9: remove spurious line in kiq setup

This overrode what queue was actually assigned for kiq.

Reviewed-by: Alex Xie <AlexBin.Xie@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu/gfx8: whitespace change

Make it consistent.

Reviewed-by: Alex Xie <AlexBin.Xie@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu/gfx9: Raven has two MECs

This was missed when Andres' queue patches were rebased.

Fixes: 42794b27 (drm/amdgpu: take ownership of per-pipe configuration v3)
Reviewed-by: Alex Xie <AlexBin.Xie@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: move gfx_v*_0_compute_queue_acquire to common code

Same function was duplicated in all gfx IP files.

Reviewed-by: Alex Xie <AlexBin.Xie@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: fix mec queue policy on single MEC asics

Fixes hangs on single MEC asics.

Fixes: 2ed286fb434 (drm/amdgpu: new queue policy, take first 2 queues of each pipe v2)
Reviewed-by: Alex Xie <AlexBin.Xie@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu/gfx: create a common bitmask function (v2)

The same function was duplicated in all the gfx IPs. Use
a single implementation for all.

v2: use static inline (Alex Xie)

Reviewed-by: Alex Xie <AlexBin.Xie@amd.com>
Suggested-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu/gfx8: drop per-APU CU limits

Always use the max for the family rather than the per sku limits.
This makes sure the mask is always the max size to avoid reporting
the wrong number of CUs.

Reviewed-by: Alex Xie <AlexBin.Xie@amd.com>
Reviewed-by: Andres Rodriguez <andresx7@gmail.com>
Cc: stable@vger.kernel.org
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu/gfx6: properly cache mc_arb_ramcfg

This was missing for gfx6.

Acked-by: Huang Rui <ray.huang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org

drm/amdgpu/gfx9: new queue policy, take first 2 queues of each pipe

Instead of taking the first pipe and giving the rest to kfd, take the
first 2 queues of each pipe.

Effectively, amdgpu and amdkfd own the same number of queues. But
because the queues are spread over multiple pipes the hardware will be
able to better handle concurrent compute workloads.

amdgpu goes from 1 pipe to 4 pipes, i.e. from 1 compute threads to 4
amdkfd goes from 3 pipe to 4 pipes, i.e. from 3 compute threads to 4

gfx9 was missed when this patch set was rebased to include gfx9.

Acked-by: Tom St Denis <tom.stdenis@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu/gfx9: allocate queues horizontally across pipes

Pipes provide better concurrency than queues, therefore we want to make
sure that apps use queues from different pipes whenever possible.

Optimize for the trivial case where an app will consume rings in order,
therefore we don't want adjacent rings to belong to the same pipe.

gfx9 was missed when these patches were rebased.

Reviewed-by: Tom St Denis <tom.stdenis@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/powerplay: fix memory leak in cz_hwmgr backend

vddc_dep_on_dal_pwrl is allocated and initialized in cz_hwmgr_backend_init
Thus free the memory in cz_hwmgr_backend_fini

Signed-off-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org

drm/amd/powerplay: fix memory leak in rv_hwmgr backend

vddc_dep_on_dal_pwrl and vq_budgeting_table are allocated and initialized
in rv_hwmgr_backend_init. Thus free the memory in rv_hwmgr_backend_fini

Signed-off-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/powerplay: add sclk and mclk overdrive for vega10

For overclocking sclk and mclk.

Signed-off-by: Eric Huang <JinHuiEric.Huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/powerplay: fix populate dpm level failed when s3 on vega10.

As the min clk may be  large than boot level can support.
in this case, just ignore the min clk.

Signed-off-by: Rex Zhu <Rex.Zhu@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: update to use RREG32_SOC15/WREG32_SOC15 for gmc9

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: update to use RREG32_SOC15/WREG32_SOC15 for mmhub

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: update to use RREG32_SOC15/WREG32_SOC15 for gfxhub

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: fix the gart table cleared issue for S3

Something writes over the first 8 MB so reserve this
on vega10 until we root cause it.

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: add ip block number prints

User is able to follow the ip block number to write the ip_block_mask for
selecting the one which user would like to enable.

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: add ip name print for selecting ips with ip_block_mask

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: remove mmhub ip

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: remove gfxhub ip

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: export mmhub get clockgating into gmc

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: export mmhub set clockgating into gmc

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: export mmhub sw_init into gmc

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: export gfxhub sw_init into gmc

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: fix to miss program invalidation at resume

This patch moves invalidation into gart enable function from hw_init.
Because we would like align the sequence calling between init and resume.

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: abstract setup vmid config for gfxhub/mmhub

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: abstract disable identity aperture for gfxhub/mmhub

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: abstract system domain enablement for gfxhub/mmhub

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: abstract cache initialization for gfxhub/mmhub

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: abstract TLB initialization for gfxhub/mmhub

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: abstract system aperture initialization for gfxhub/mmhub

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: abstract gart aperture initialization for gfxhub/mmhub

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: abstract gart table initialization for gfxhub/mmhub

Signed-off-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: add saved_bo to save vce 4.0 context when suspend

We are using PSP to resume firmware after suspend, and it is
resumed at where it got suspended, so we'd better save the
the context.

Signed-off-by: Leo Liu <leo.liu@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: use existing function amdgpu_bo_create_kernel

To simplify vce bo create

Signed-off-by: Leo Liu <leo.liu@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: add vcpu_bo cpu address for vce

Signed-off-by: Leo Liu <leo.liu@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: Move compute vm bug logic to amdgpu_vm.c

  In review, Christian would like to keep the logic
  inside amdgpu_vm.c with a cost of slightly slower.
  The loop is still optimized out with this patch.

v2: remove the if statement. Now it is not slower.

Signed-off-by: Alex Xie <AlexBin.Xie@amd.com>
Reviewed-by: Christian König <christian.koeng@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/powerplay: enable CKS by default on vega10.

Signed-off-by: Rex Zhu <Rex.Zhu@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/powerplay: Align with VBIOS to support AVFS parameters.

Signed-off-by: Rex Zhu <Rex.Zhu@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/powerplay: Add floor DCEF for DS on boot.

Use the vbios to look up the default frequencies
for socclk and dcefclk.

Signed-off-by: Rex Zhu <Rex.Zhu@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: use LRU mapping policy for SDMA engines

Spreading the load across multiple SDMA engines can increase memory
transfer performance.

Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: guarantee bijective mapping of ring ids for LRU v3

Depending on usage patterns, the current LRU policy may create a
non-injective mapping between userspace ring ids and kernel rings.

This behaviour is undesired as apps that attempt to fill all HW blocks
would be unable to reach some of them.

This change forces the LRU policy to create bijective mappings only.

v2: compress ring_blacklist
v3: simplify amdgpu_ring_is_blacklisted() logic

Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: implement lru amdgpu_queue_mgr policy for compute v4

Use an LRU policy to map usermode rings to HW compute queues.

Most compute clients use one queue, and usually the first queue
available. This results in poor pipe/queue work distribution when
multiple compute apps are running. In most cases pipe 0 queue 0 is
the only queue that gets used.

In order to better distribute work across multiple HW queues, we adopt
a policy to map the usermode ring ids to the LRU HW queue.

This fixes a large majority of multi-app compute workloads sharing the
same HW queue, even though 7 other queues are available.

v2: use ring->funcs->type instead of ring->hw_ip
v3: remove amdgpu_queue_mapper_funcs
v4: change ring_lru_list_lock to spinlock, grab only once in lru_get()

Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: untie user ring ids from kernel ring ids v6

Add amdgpu_queue_mgr, a mechanism that allows disjointing usermode's
ring ids from the kernel's ring ids.

The queue manager maintains a per-file descriptor map of user ring ids
to amdgpu_ring pointers. Once a map is created it is permanent (this is
required to maintain FIFO execution guarantees for a context's ring).

Different queue map policies can be configured for each HW IP.
Currently all HW IPs use the identity mapper, i.e. kernel ring id is
equal to the user ring id.

The purpose of this mechanism is to distribute the load across multiple
queues more effectively for HW IPs that support multiple rings.
Userspace clients are unable to check whether a specific resource is in
use by a different client. Therefore, it is up to the kernel driver to
make the optimal choice.

v2: remove amdgpu_queue_mapper_funcs
v3: made amdgpu_queue_mgr per context instead of per-fd
v4: add context_put on error paths
v5: rebase and include new IPs UVD_ENC & VCN_*
v6: drop unused amdgpu_ring_is_valid_index (Alex)

Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: workaround tonga HW bug in HQD programming sequence

Tonga based asics may experience hangs when an HQD's EOP parameters
are modified.

Workaround this HW issue by avoiding writes to these registers for
tonga asics.

Based on the following ROCm commit:
2a0fb8 - drm/amdgpu: Synchronize KFD HQD load protocol with CP scheduler

From the ROCm git repository:
https://github.com/RadeonOpenCompute/ROCK-Kernel-Driver.git

CC: Jay Cornwall <Jay.Cornwall@amd.com>
Suggested-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: condense mqd programming sequence

The MQD structure matches the reg layout. Take advantage of this to
simplify HQD programming.

Note that the ACTIVE field still needs to be programmed last.

Suggested-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: new queue policy, take first 2 queues of each pipe v2

Instead of taking the first pipe and giving the rest to kfd, take the
first 2 queues of each pipe.

Effectively, amdgpu and amdkfd own the same number of queues. But
because the queues are spread over multiple pipes the hardware will be
able to better handle concurrent compute workloads.

amdgpu goes from 1 pipe to 4 pipes, i.e. from 1 compute threads to 4
amdkfd goes from 3 pipe to 4 pipes, i.e. from 3 compute threads to 4

v2: fix policy comment

Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: avoid KIQ clashing with compute or KFD queues v2

Instead of picking an arbitrary queue for KIQ, search for one according
to policy. The queue must be unused.

Also report the KIQ as an unavailable resource to KFD.

In testing I ran into KCQ initialization issues when using pipes 2/3 of
MEC2 for the KIQ. Therefore the policy disallows grabbing one of these.

v2: fix (ring.me + 1) to (ring.me -1) in amdgpu_amdkfd_device_init

Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: remove hardcoded queue_mask in PACKET3_SET_RESOURCES

The assumption that we are only using the first pipe no longer holds.
Instead, calculate the queue_mask from the queue_bitmap.

Acked-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: allocate queues horizontally across pipes

Pipes provide better concurrency than queues, therefore we want to make
sure that apps use queues from different pipes whenever possible.

Optimize for the trivial case where an app will consume rings in order,
therefore we don't want adjacent rings to belong to the same pipe.

Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Acked-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: remove duplicate magic constants from amdgpu_amdkfd_gfx*.c

This information is already available in adev.

Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdkfd: allow split HQD on per-queue granularity v5

Update the KGD to KFD interface to allow sharing pipes with queue
granularity instead of pipe granularity.

This allows for more interesting pipe/queue splits.

v2: fix overflow check for res.queue_mask
v3: fix shift overflow when setting res.queue_mask
v4: fix comment in is_pipeline_enabled()
v5: clamp res.queue_mask to the first MEC only

Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: teach amdgpu how to enable interrupts for any pipe v3

The current implementation is hardcoded to enable ME1/PIPE0 interrupts
only.

This patch allows amdgpu to enable interrupts for any pipe of ME1.

v2: added gfx9 support
v3: use soc15_grbm_select for gfx9

Acked-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: allow split of queues with kfd at queue granularity v4

Previously the queue/pipe split with kfd operated with pipe
granularity. This patch allows amdgpu to take ownership of an arbitrary
set of queues.

It also consolidates the last few magic numbers in the compute
initialization process into mec_init.

v2: support for gfx9
v3: renamed AMDGPU_MAX_QUEUES to AMDGPU_MAX_COMPUTE_QUEUES
v4: fix off-by-one in num_mec checks in *_compute_queue_acquire

Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: take ownership of per-pipe configuration v3

Make amdgpu the owner of all per-pipe state of the HQDs.

This change will allow us to split the queues between kfd and amdgpu
with a queue granularity instead of pipe granularity.

This patch fixes kfd allocating an HDP_EOP region for its 3 pipes which
goes unused.

v2: support for gfx9
v3: fix gfx7 HPD intitialization

Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/radeon: take ownership of pipe initialization

Take ownership of pipe initialization away from KFD.

Note that hpd_eop_gpu_addr was already large enough to accomodate all
pipes.

Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: rename rdev to adev

Rename straggler instances of r(adeon)dev to a(mdgpu)dev

Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: fix kgd_hqd_load failing to update shadow_wptr

The return value from copy_form_user is 0 for the success case.

Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: unify MQD programming sequence for kfd and amdgpu v2

Use the same gfx_*_mqd_commit function for kfd and amdgpu codepaths.

This removes the last duplicates of this programming sequence.

v2: fix cp_hqd_pq_wptr value

Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Acked-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: remove duplicate definition of cik_mqd

The gfxv7 contains a slightly different version of cik_mqd called
bonaire_mqd. This can introduce subtle bugs if fixes are not applied in
both places.

Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: detect timeout error when deactivating hqd

Handle HQD deactivation timeouts instead of ignoring them.

Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: refactor MQD/HQD initialization v3

The MQD programming sequence currently exists in 3 different places.
Refactor it to absorb all the duplicates.

The success path remains mostly identical except for a slightly
different order in the non-kiq case. This shouldn't matter if the HQD
is disabled.

The error handling paths have been updated to deal with the new code
structure.

v2: the non-kiq path for gfxv8 was dropped in the rebase
v3: split MEC_HPD_SIZE rename, dropped doorbell changes

Reviewed-by: Edward O'Callaghan <funfunctor@folklore1984.net>
Acked-by: Christian König <christian.koenig@amd.com>
Acked-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: clarify MEC_HPD_SIZE is specific to a gfx generation

Rename MEC_HPD_SIZE to GFXN_MEC_HPD_SIZE to clarify it is specific to a
gfx generation.

Signed-off-by: Andres Rodriguez <andresx7@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/powerplay: code clean up in vega10_hwmgr.c

Signed-off-by: Rex Zhu <Rex.Zhu@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

Revert "drm/amd/powerplay: disable engine spread spectrum feature on Vega10."

This reverts commit f8fdaa0e7b81698ba2ad8c2d20c7f9a44c75e0c6.
firmware add support for this feature, so still ctrl by vbios.

Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/powerplay: enable deep sleep by default for vega10

Signed-off-by: Rex Zhu <Rex.Zhu@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/powerplay: enable ulv feature by default for vega10.

Signed-off-by: Rex Zhu <Rex.Zhu@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: stop joining VM PTE updates

This isn't beneficial any more since VRAM allocations are now split
so that they fits into a single page table.

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Junwei Zhang <Jerry.Zhang@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: cache the complete pde

Makes it easier to update the PDE with huge pages.

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Junwei Zhang <Jerry.Zhang@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: Remove two ! operations in an if condition

 Make the code easier to understand.

Signed-off-by: Alex Xie <AlexBin.Xie@amd.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: Optimize a function called by every IB sheduling

  Move several if statements and a loop statment from
  run time to initialization time.

Signed-off-by: Alex Xie <AlexBin.Xie@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: Program ring for vce instance 1 at its register space

We need program ring buffer on instance 1 register space domain,
when only if instance 1 available, with two instances or instance 0,
and we need only program instance 0 regsiter space domain for ring.

Signed-off-by: Leo Liu <leo.liu@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: Return EINVAL if no PT BO

This change is also useful for the upcoming changes where page tables
can be updated by CPU.

Signed-off-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Junwei Zhang <Jerry.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: add some extra VM error handling

If updating the PDs fails we now invalidate all entries to try again later.

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Junwei Zhang <Jerry.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: cleanup adjust_mc_addr handling v4

Rename adjust_mc_addr to get_vm_pde and check the address bits in one place.

v2: handle vcn as well, keep setting the valid bit manually,
    add a BUG_ON() for GMC v6, v7 and v8 as well.
v3: handle vcn_v1_0_enc_ring_emit_vm_flush as well.
v4: fix the BUG_ON mask for GFX6-8

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: enable lbpw on raven

Signed-off-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: init lbpw on raven

Load Balancing Per Watt (LBPW) allows dynamically disable CUs
when they are idle

Signed-off-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: update sdma 4.1 raven specific golden settings

Signed-off-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: update GC 9.1 raven specific golden settings

Signed-off-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: update GC 9.1 golden settings

Signed-off-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: remove unnecessary debug message

remnants from bring-up.

Signed-off-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/powerplay: fix set tools address for Vega10

Tools fb address was failed to send to smu when smu
was not running. Changing sequence will fix it.

Signed-off-by: Eric Huang <JinHuiEric.Huang@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/powerplay: fix soft pptable size for Vega10

It is to fix bug of sysfs entry pp_table which had size 0 of output before.

Signed-off-by: Eric Huang <JinHuiEric.Huang@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: optimize amdgpu driver load & resume time

amdgpu_device_resume() & amdgpu_device_init() have a high
time consuming call of amdgpu_late_init() which sets the
clock_gating state of all IP blocks and is blocking.
This patch defers only this setting of clock gating state
operation to post resume of amdgpu driver but ideally before
the UI comes up or in some cases post ui as well.

With this change the resume time of amdgpu_device comes down
from 1.299s to 0.199s which further helps in reducing the overall
system resume time.

V1: made the optimization applicable during driver load as well.

TEST:(For ChromiumOS on STONEY only)
* UI comes up
* amdgpu_late_init() call gets called consistently and no errors reported.

Signed-off-by: Shirish S <shirish.s@amd.com>
Reviewed-by: Huang Rui <ray.huang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: fix null point error when rmmod amdgpu.

this bug happened when amdgpu load failed.

[   75.740951] BUG: unable to handle kernel paging request at 00000000000031c0
[   75.748167] IP: [<ffffffffa064a0e0>] amdgpu_fbdev_restore_mode+0x20/0x60 [amdgpu]
[   75.755774] PGD 0

[   75.759185] Oops: 0000 [#1] SMP
[   75.762408] Modules linked in: amdgpu(OE-) ttm(OE) drm_kms_helper(OE) drm(OE) i2c_algo_bit(E) fb_sys_fops(E) syscopyarea(E) sysfillrect(E) sysimgblt(E) rpcsec_gss_krb5(E) nfsv4(E) nfs(E) fscache(E) eeepc_wmi(E) asus_wmi(E) sparse_keymap(E) intel_rapl(E) snd_hda_codec_hdmi(E) snd_hda_codec_realtek(E) snd_hda_codec_generic(E) snd_hda_intel(E) snd_hda_codec(E) snd_hda_core(E) x86_pkg_temp_thermal(E) intel_powerclamp(E) snd_hwdep(E) snd_pcm(E) snd_seq_midi(E) coretemp(E) kvm_intel(E) snd_seq_midi_event(E) snd_rawmidi(E) kvm(E) snd_seq(E) joydev(E) snd_seq_device(E) snd_timer(E) irqbypass(E) crct10dif_pclmul(E) crc32_pclmul(E) mei_me(E) ghash_clmulni_intel(E) snd(E) aesni_intel(E) mei(E) soundcore(E) aes_x86_64(E) shpchp(E) serio_raw(E) lrw(E) acpi_pad(E) gf128mul(E) glue_helper(E) ablk_helper(E) mac_hid(E)
[   75.835574]  cryptd(E) parport_pc(E) ppdev(E) lp(E) nfsd(E) parport(E) auth_rpcgss(E) nfs_acl(E) lockd(E) grace(E) sunrpc(E) autofs4(E) hid_generic(E) usbhid(E) mxm_wmi(E) psmouse(E) e1000e(E) ptp(E) pps_core(E) ahci(E) libahci(E) wmi(E) video(E) i2c_hid(E) hid(E)
[   75.858489] CPU: 5 PID: 1603 Comm: rmmod Tainted: G           OE   4.9.0-custom #2
[   75.866183] Hardware name: System manufacturer System Product Name/Z170-A, BIOS 0901 08/31/2015
[   75.875050] task: ffff88045d1bbb80 task.stack: ffffc90002de4000
[   75.881094] RIP: 0010:[<ffffffffa064a0e0>]  [<ffffffffa064a0e0>] amdgpu_fbdev_restore_mode+0x20/0x60 [amdgpu]
[   75.891238] RSP: 0018:ffffc90002de7d48  EFLAGS: 00010286
[   75.896648] RAX: 0000000000000000 RBX: 0000000000000000 RCX: 0000000000000001
[   75.903933] RDX: 0000000000000000 RSI: ffff88045d1bbb80 RDI: 0000000000000286
[   75.911183] RBP: ffffc90002de7d50 R08: 0000000000000502 R09: 0000000000000004
[   75.918449] R10: 0000000000000000 R11: 0000000000000001 R12: ffff880464bf0000
[   75.925675] R13: ffffffffa0853000 R14: 0000000000000000 R15: 0000564e44f88210
[   75.932980] FS:  00007f13d5400700(0000) GS:ffff880476540000(0000) knlGS:0000000000000000
[   75.941238] CS:  0010 DS: 0000 ES: 0000 CR0: 0000000080050033
[   75.947088] CR2: 00000000000031c0 CR3: 000000045fd0b000 CR4: 00000000003406e0
[   75.954332] DR0: 0000000000000000 DR1: 0000000000000000 DR2: 0000000000000000
[   75.961566] DR3: 0000000000000000 DR6: 00000000fffe0ff0 DR7: 0000000000000400
[   75.968834] Stack:
[   75.970881]  ffff880464bf0000 ffffc90002de7d60 ffffffffa0636592 ffffc90002de7d80
[   75.978454]  ffffffffa059015f ffff880464bf0000 ffff880464bf0000 ffffc90002de7da8
[   75.986076]  ffffffffa0595216 ffff880464bf0000 ffff880460f4d000 ffffffffa0853000
[   75.993692] Call Trace:
[   75.996177]  [<ffffffffa0636592>] amdgpu_driver_lastclose_kms+0x12/0x20 [amdgpu]
[   76.003700]  [<ffffffffa059015f>] drm_lastclose+0x2f/0xd0 [drm]
[   76.009777]  [<ffffffffa0595216>] drm_dev_unregister+0x16/0xd0 [drm]
[   76.016255]  [<ffffffffa0595944>] drm_put_dev+0x34/0x70 [drm]
[   76.022139]  [<ffffffffa062f365>] amdgpu_pci_remove+0x15/0x20 [amdgpu]
[   76.028800]  [<ffffffff81416499>] pci_device_remove+0x39/0xc0
[   76.034661]  [<ffffffff81531caa>] __device_release_driver+0x9a/0x140
[   76.041121]  [<ffffffff81531e58>] driver_detach+0xb8/0xc0
[   76.046575]  [<ffffffff81530c95>] bus_remove_driver+0x55/0xd0
[   76.052401]  [<ffffffff815325fc>] driver_unregister+0x2c/0x50
[   76.058244]  [<ffffffff81416289>] pci_unregister_driver+0x29/0x90
[   76.064466]  [<ffffffffa0596c5e>] drm_pci_exit+0x9e/0xb0 [drm]
[   76.070507]  [<ffffffffa0796d71>] amdgpu_exit+0x1c/0x32 [amdgpu]
[   76.076609]  [<ffffffff81104810>] SyS_delete_module+0x1a0/0x200
[   76.082627]  [<ffffffff810e2b1a>] ? rcu_eqs_enter.isra.36+0x4a/0x50
[   76.089001]  [<ffffffff8100392e>] do_syscall_64+0x6e/0x180
[   76.094583]  [<ffffffff817e1d2f>] entry_SYSCALL64_slow_path+0x25/0x25
[   76.101114] Code: 94 c0 c3 31 c0 5d c3 0f 1f 40 00 0f 1f 44 00 00 55 31 c0 48 89 e5 53 48 89 fb 48 c7 c7 1d 21 84 a0 e8 ab 77 b3 e0 e8 fc 8b d7 e0 <48> 8b bb c0 31 00 00 48 85 ff 74 09 e8 ff eb fc ff 85 c0 75 03
[   76.121432] RIP  [<ffffffffa064a0e0>] amdgpu_fbdev_restore_mode+0x20/0x60 [amdgpu]

Signed-off-by: Rex Zhu <Rex.Zhu@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/powerplay: set powerplay support cap on raven

Signed-off-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/powerplay: bypass pptable process on raven

Signed-off-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/powerplay: add get_temperature interface for raven

Signed-off-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/powerplay: update magic number for rv hw backend

Signed-off-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/powerplay: using MinFClock/MaxFclock to report Min/Max memory clock limits

Signed-off-by: Hawking Zhang <Hawking.Zhang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/radeon: Cleanup pageflipping IRQ handling for evergreen, si

Same as the previous patch, but for pageflipping now. This also lets us
clear up the copy paste for vblank/vline IRQs.

Changes since v1:
- Preserve the order all registers are written back

Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Lyude <lyude@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/radeon: Cleanup HDMI audio interrupt handling for evergreen

Same as the previous patch, but now for handling HDMI audio interrupts.

Changes since v1:
- Preserve the order we write back all registers

Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Lyude <lyude@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/radeon: Cleanup display interrupt handling for evergreen, si

The current code here is really, really bad. A huge amount of it looks
to be copy pasted, it has some weird hatred of arrays and code sharing,
switch cases everywhere for things that really don't need them, and it
makes the file seem immensely more complex then it actually is. This is
a pain for maintanence, and is vulnerable to more weird irq handling
bugs.

So, let's start cleaning this up a bit. Modify all of the IRQ handlers
for evergreen/si so that they just use for loops. As well, we add a
helper function radeon_irq_kms_set_irq_n_enabled(), whose purpose is
just to update the state of registers that enable/disable interrupts
while printing any changes to the set of enabled interrupts to the
kernel log.

Note in this commit, since vblank/vline irq acking is intertwined with
page flip irq acking, we can't cut out all of the copy paste in
evergreen/si_irq_ack() just yet.

Changes since v1:
- Preserve order we write back all registers

Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Lyude <lyude@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu/gfx9: Switch baremetal to use KIQ for compute ring management. (v4)

KIQ is the Kernel Interface Queue for managing the MEC.  Rather than setting
up rings via direct MMIO of ring registers, the rings are configured via
special packets sent to the KIQ.  The allows the MEC to better manage shared
resources and certain power events. It also reduces the code paths in the
driver to support and is required for MEC powergating.

v2: drop gfx_v9_0_cp_compute_fini() as well
v3: rebase on latest changes derived from gfx8, add unmap queues on
hw_fini
v4: fix copy/paste typo in error message (Rex)

Acked-by: Tom St Denis <tom.stdenis@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu/gfx9: fix resume of KIQ and KCQs (v2)

No need to reset the wptr and clear the rings.  The UNMAP_QUEUES
packet writes the current MQD state back the MQD on suspend,
so there is no need to reset it as well.

v2: fix from gfx8 (Rex)

Acked-by: Tom St Denis <tom.stdenis@amd.com>
Acked-by: monk liu <monk.liu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu/gfx9: move SET_RESOURCES into the same command stream

As the KCQ setup.  This way we only have to wait once for the
entire MEC.

Acked-by: Tom St Denis <tom.stdenis@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>