drm/amdgpu: check flush fence context instead of same ring v2

Otherwise we can run into the following situation:

1. Process A grabs ID 1 for ring 0.
2. Process B grabs ID 1 for ring 0.
3. Process A grabs ID 1 for ring 1.
4. Process A tries to reuse ID1 for ring 0 but things he doesn't need to flush.

v2: check the context of the flush fence instead of messing with the owner field.

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

drm/radeon: support backlight control for UNIPHY3

Same interface as other UNIPHY blocks

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

drm/amdgpu: support backlight control for UNIPHY3

Same interface as other UNIPHY blocks

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

drm/amdgpu: remove usec timeout loop from IB tests

We already waited for the fence, so waiting for the registers
is completely pointless and just copy & pasted from the ring test.

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

drm/amdgpu: cleanup hw reference handling in the IB tests

Reference should be taken when we make the assignment, not anywhere else.

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

drm/amdgpu: cleanup UVD coding style

Cleanup 80 chars limit.

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Alexandre Demers <alexandre.f.demers@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: allow multiple sessions in the same VCE IB

We always used updated firmware for amdgpu, so this actually should work fine.

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

drm/amdgpu: cleanup VCE coding style

Fix 80 chars issues and remove some dead code as well.

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

drm/amdgpu: sanitize fence numbers

Looks like the VCE block sometimes still sends nonsense
fence numbers on startup.

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

drm/amdgpu/powerplay: endian fixes for ppatomctrl.c

Atom tables are in LE format.

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

drm/amd/powerplay: don't add invalid voltage.

if atomctrl_get_voltage_evv_on_sclk_ai returns
non zero (fail) in the expansion of the
PP_ASSERT_WITH_CODE macro the continue will
actually do nothing, So invalid voltage will be
added to ppbable.

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: add read/write function for GC CAC programming

Create a GC_CAC_IND_INDEX/DATA pair of funcitons to program
all the CAC registers

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 definitions related to di/dt feature for fiji and polaris.

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 shared definitions for di/dt feature.

v1: delete some comflict definitions between polaris and fiji.

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: remove gfx8 registers that vary between asics

those register mask definitions are different in polaris compare to
former gfx 8 gpus, so remove them from misusing.

Signed-off-by: Ken Wang <Qingqing.Wang@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/powerplay: add mvdd dpm support.

SMC requires master switch bit to be set.

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: get number of shade engine by cgs interface.

the num of shade engine was needed to
measure the activity of the graphics core
and to enable di/dt feature.

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: remove more of the ring backup code

Not used anymore.

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

drm/amd/powerplay:  Unify family defines

s/AMD_FAMILY_/AMDGPU_FAMILY_/

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

drm/amdgpu: clean up ring_backup code, no need more

Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: ib test first after gpu reset

Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: recovery hw jobs when gpu reset V3

V3: directly use pd_addr.

Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: abstract amdgpu_vm_is_gpu_reset

Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: add a bool to specify if needing vm flush V2

which avoids job->vm_pd_addr be changed.

V2: pass job structure to amdgpu_vm_grab_id and amdgpu_vm_flush directly.

Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: add amd_sched_job_recovery

Which is to recover hw jobs when gpu reset.

Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: force completion for gpu reset

After all hw jobs are reset, hw fence is meaningless, so force_completion

Cc: William Lewis <minutemaidpark@hotmail.com>
Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: block ttm first before parking scheduler

Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd: add amd_sched_hw_job_reset

amd_sched_hw_job_reset will remove callback from hw fence.

Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd: add parent for sched fence

Parent of sched fence is hw fence which is to signal sched fence.

Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: remove evict vram

Previous vm fault is since page talbe losts connection with vmid after gpu reset.
Now the issue is fixed by recovery. No need more.

If we want to save vram for some EDC card, we will need to consider a complete solution.

Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: put old hw fence of job if gpu reset

Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: remove fence parameter from amd_sched_job_init

We return the fence as part of the job structur anyway,
no need to do this twice.

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

drm/amdgpu: earlier free SA resources

Keep the time we don't have a fence associated with the resource smaller.

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

drm/amdgpu: shorten amdgpu_job_free_resources

The fence and the sync object are not hardware resources.

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

drm/amdgpu: fix user fence handling once more

Same problem as with the VM page tables. The user fence address must be
determined before the job is scheduled, not when the IB is executed.

This fixes a security problem where user fences could be used to overwrite
any part of VRAM.

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

drm/amd/amdgpu: Add bank selection for MMIO debugfs (v3)

(v2) Added INSTANCE selector
(v3) Changed order of bank selectors

Signed-off-by: Tom St Denis <tom.stdenis@amd.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/amdgpu: Add gca config debug entry (v4)

A binary entry that lists GCA configuration data (and can be
read by umr).

(v2) Use kmalloc instead of vmalloc
(v3) Minor indentation correction
(v4) agd: Squash in kmalloc fix

Signed-off-by: Tom St Denis <tom.stdenis@amd.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/gfx: add instance field to select_se_sh (v3)

Add ability to specify instance in select_se_sh callback.
Defaults to 0xffffffff all over the driver.

(v2) Don't enable INSTANCE_BROADCAST by default
(v3) Style changes

Signed-off-by: Tom St Denis <tom.stdenis@amd.com>
Reviewed-by: Nicolai Hähnle <nicolai.haehnle@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: Poll for both connect/disconnect on analog connectors

DRM_CONNECTOR_POLL_CONNECT only enables polling for connections, not
disconnections. Because of this, we end up losing hotplug polling for
analog connectors once they get connected.

Easy way to reproduce:
 - Grab a machine with an AMD GPU and a VGA port
 - Plug a monitor into the VGA port, wait for it to update the connector
   from disconnected to connected
 - Disconnect the monitor on VGA, a hotplug event is never sent for the
   removal of the connector.

Originally, only using DRM_CONNECTOR_POLL_CONNECT might have been a good
idea since doing VGA polling can sometimes result in having to mess with
the DAC voltages to figure out whether or not there's actually something
there since VGA doesn't have HPD. Doing this would have the potential of
showing visible artifacts on the screen every time we ran a poll while a
VGA display was connected. Luckily, amdgpu_vga_detect() only resorts to
this sort of polling if the poll is forced, and DRM's polling helper
doesn't force it's polls.

Additionally, this removes some assignments to connector->polled that
weren't actually doing anything.

Cc: stable@vger.kernel.org
Signed-off-by: Lyude <cpaul@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/radeon: Poll for both connect/disconnect on analog connectors

DRM_CONNECTOR_POLL_CONNECT only enables polling for connections, not
disconnections. Because of this, we end up losing hotplug polling for
analog connectors once they get connected.

Easy way to reproduce:
 - Grab a machine with a radeon GPU and a VGA port
 - Plug a monitor into the VGA port, wait for it to update the connector
   from disconnected to connected
 - Disconnect the monitor on VGA, a hotplug event is never sent for the
   removal of the connector.

Originally, only using DRM_CONNECTOR_POLL_CONNECT might have been a good
idea since doing VGA polling can sometimes result in having to mess with
the DAC voltages to figure out whether or not there's actually something
there since VGA doesn't have HPD. Doing this would have the potential of
showing visible artifacts on the screen every time we ran a poll while a
VGA display was connected. Luckily, radeon_vga_detect() only resorts to
this sort of polling if the poll is forced, and DRM's polling helper
doesn't force it's polls.

Additionally, this removes some assignments to connector->polled that
weren't actually doing anything.

Cc: stable@vger.kernel.org
Signed-off-by: Lyude <cpaul@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: stop/resume fb access when gpu reset V3

V2: Add wait_for_mc_idle after stopping fb access
V3:
1. Remove resume fb access since there is no need to
    do that for gpu reset.
2. Move stop fb access to amdgpu_gpu_reset function,
since it's the same for all asics.

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com> (V1)
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/powerplay: change condition judgment as function's return value changed.

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

drm/amd/powerplay: functions's return state was reversed

In commit 195567e99bdf6491a370b71a1dcf6b4c891495d7,
use true/false instead of 1/0 to fix build warning.
But the original logic: '0' means true and '1' means false.

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

drm/amdgpu: stop disabling irqs when it isn't neccessary

A regular spin_lock/unlock should do here as well.

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

drm/amdgpu: must update page table after gpu reset

Record the gpu reset count in vmid to identify if gpu reset happened.

Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: save/restore bios scratch when gpu reset

Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: must update page table after gpu reset

Record the gpu reset count in vmid to identify if gpu reset happened.

Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/ttm: fix stupid parameter inversion in the pipeline code

We want to keep the newest fence, not the oldest one.

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

drm/amd/powerplay: fix typos in comment in polaris' hwmgr

Signed-off-by: Alexandre Demers <alexandre.f.demers@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/powerplay: fix trivial typo and tidy comment

Signed-off-by: Alexandre Demers <alexandre.f.demers@gmail.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/amdgpu: Set DRIVER_MODESET feature flag at build time

This flag was being set unconditionally at runtime so just set it at
compile time instead.

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Frank Binns <frank.binns@imgtec.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: Unpin BO if we can't get fences in amdgpu_crtc_page_flip

If reservation_object_get_fences_rcu failed, we'd previously go directly
to the cleanup label, so we'd leave the BO pinned.

While we're at it, remove two amdgpu_bo_unreserve calls in favour of two
new labels.

Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu/gmc7: remove duplicate wait_for_idle functions

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

drm/amdgpu/gmc8: remove duplicate wait_for_idle functions

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

drm/amdgpu/dce8: fix flash with white screen on monitor

Fixed mc stop and resume hardware programming sequence.

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

drm/amdgpu: skip invalid ip blocks in ip helpers

Skip IP blocks that are not enabled.

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

drm/amdgpu: add missing breaks

Saves a few CPU cycles.

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

drm/amdgpu: add IP helpers for wait_for_idle and is_idle

Helpers to to call the IP functions for the selected IP.

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

drm/amdgpu/powerplay/cz: add missing call to powergate VCE

Not enabled yet, but missing the call.

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

drm/amdgpu/dce11: update async flip update time

Use UPDATE_IMMEDIATE (update on next data request boundary) rather
than UPDATE_H_RETRACE (update on next line boundary).  The data
request boundary is less than a scanline, so it update will happen
sooner.

Cc: Vitaly Prosyak <vitaly.prosyak@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/ttm: wait for eviction in ttm_bo_force_list_clean

Now that we can pipeline evictions we need to wait for
them to finish when we cleanup a memory domain.

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

drm/amdgpu: don't update page tables for VM emulation

It's just overhead to do so and allocating a VMID
when we don't need one is actually a bit dangerous.

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

drm/amdgpu: implement HDP functions for UVD v2

Flush and invalidate the HDP caches.

v2: fix typo in comment

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

drm/amdgpu: change pcie_gen_cap magic code to macro

This patch changes pcie_gen_cap magic code to macro to make it more
readable.

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Rex Zhu <Rex.Zhu@amd.com>
Signed-off-by: Huang Rui <ray.huang@amd.com>
Cc: Eric Huang <JinHuiEric.Huang@amd.com>
Cc: Ken Wang <Qingqing.Wang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu/gfx7: make gfx_v7_0_rlc_stop static

Not used outside of gfx7.

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

drm/amdgpu/gfx7: switch to using the existing rlc callbacks

gfx8 already uses them.  Remove the direct exports and
use the callbacks fpr gfx7.

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

drm/amdgpu: move select_se_sh into the gfx struct

It's gfx IP specific, not asic specific, so move to a
gfx callback.

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

drm/amdgpu: move get_gpu_clock_counter into the gfx struct

It's gfx IP specific, not asic specific, so move to a
gfx callback.

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

drm/amdgpu: drop wait_for_mc_idle asic callback

Only used in the gmc IP modules so just call the local
function directly.

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

drm/amdgpu/gmc: make some functions static

These are not used outside of the respective gmc ip modules.

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

drm/amdgpu: validate VM PTs only on eviction

We don't need to validate them again if the eviction counter didn't changed.

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

drm/amdgpu: add eviction counter

Keep track of the number of evictions since boot.

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

drm/amdgpu: add amdgpu_irq_gpu_reset_resume_helper

irq need to update when gpu reset happens.

Signed-off-by: Chunming Zhou <David1.Zhou@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>

drm/amd/powerplay: make sure VCE is disabled by default

This patch is a port of similar patch for amdgpu
when PP is disabled. Since the code flow is little
different when PP is enabled, we need to make sure
the patch is applied for PP enabled path as well.

With the current code, when we boot with the amdgpu
driver enabled and loaded, the VCE also automatically
remains enabled since bootup. This can be verified from
the output of amdgpu_pm_info. It does not matter whether
we boot into command line directly or into X, the VCE
stays enabled the entire time.

This patch addresses the issue and makes sure that
VCE is turned on only during playback, and remains
disabled otherwise.

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Arindam Nath <arindam.nath@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amd/amdgpu: make sure VCE is disabled by default

With the current code, when we boot with the amdgpu
driver enabled and loaded, the VCE also automatically
remains enabled since bootup. This can be verified from
the output of amdgpu_pm_info. It does not matter whether
we boot into command line directly or into X, the VCE
stays enabled the entire time.

This patch addresses the issue and makes sure that
VCE is turned on only during playback, and remains
disaled otherwise.

Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Arindam Nath <arindam.nath@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: disable compute pipeline sync workaround when using fixed fw

No need to stall the pipe when we are using firmware with the
fix.

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

drm/amdgpu/gfx8: set USER_SHADER_ARRAY_CONFIG based on disable_cu parameter

v2: do not overwrite register when bitmap is zero

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

drm/amdgpu/gfx7: set USER_SHADER_ARRAY_CONFIG based on disable_cu parameter

v2: do not overwrite register when bitmap is zero

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

drm/amdgpu: add disable_cu parameter

This parameter will allow disabling individual CUs on module load, e.g.
amdgpu.disable_cu=2.0.3,2.0.4 to disable CUs 3 and 4 of SE2.

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

drm/amdgpu: evict vram when gpu reset

On workstation cards with ECC vram, the entirety of vram is cleared to 0
on asic init to set the ECC status correctly.  On non ECC boards, I don't
think they do any explicit clearing, but the vram controller is reset
which may cause issues with the data there.

Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: block scheduler when gpu reset

Signed-off-by: Chunming Zhou <David1.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

amdgpu: use NULL instead of 0 for pointer

In the AMD powerplay driver, a pointer is checked for validity by
comparing against an integer '0', which causes a harmless warning
when building with "make W=1":

drivers/gpu/drm/amd/amdgpu/../powerplay/hwmgr/processpptables.c:1502:16: error: ordered comparison of pointer with integer zero [-Werror=extra]

This changes the code to the more conventional "if (pointer)" check.

Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Arnd Bergmann <arnd@arndb.de>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: pipeline evictions as well

This boosts Xonotic from 38fps to 47fps when artificially limiting VRAM to
256MB for testing. It should improve all CPU bound rendering situations
where we have a lot of swapping to/from VRAM.

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

drm/amdgpu: save the PD addr before scheduling the job

When we pipeline evictions the page directory could already be
moving somewhere else when grab_id is called.

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

drm/ttm: add the infrastructure for pipelined evictions

Free up the memory immediately, remember the last eviction for each domain and
make new allocations depend on the last eviction to be completed.

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

drm/ttm: simplify ttm_bo_wait

As far as I can see no need for a custom implementation any more.

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

drm/ttm: remove TTM_BO_PRIV_FLAG_MOVING

Instead of using the flag just remember the fence of the last move operation.

This avoids waiting for command submissions pipelined after the move, but
before accessing the BO with the CPU again.

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

drm/ttm: remove no_gpu_wait param from ttm_bo_move_accel_cleanup

It isn't used and not waiting for the GPU after scheduling a move is
actually quite dangerous.

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

drm/amdgpu: remove pre move wait

Not needed any more.

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

drm/amdgpu: sync to buffer moves before VM updates

Otherwise we could update the VM page tables while the move is only scheduled.

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

drm/ttm: wait for BO idle after the move in ttm_bo_swapout

Final part to avoid pre move waits.

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

drm/ttm: drop waiting for idle in ttm_bo_evict.

That is unnecessary now.

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

drm/ttm: drop wait for idle in ttm_bo_move_buffer

That is unnecessary now.

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

drm/ttm: wait for BO idle in ttm_bo_move_memcpy

When we want to pipeline accelerated moves we need to wait in the fallback path.

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

drm/ttm: add wait for idle in all drivers bo_move functions

Wait for idle before moving the BO in all drivers implementing
an accelerated move function.

This should keep the current behavior when removing the pre move wait.

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

drm/ttm: remove dummy bo_move implementations

It's pointless to only call the default implementation.

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

drm/ttm: remove NULL checks when calling ttm_tt_destroy

The function is a no-op with a NULL pointer.

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

drm/ttm: cleanup ttm_tt_(unbind|destroy)

ttm_tt_destroy should be the only one unbinding the object.

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

drm/amdgpu: some improvement in parsing inputs

It changes the way to skip newline character and also avoids
warning message from some compiler.

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

drm/amd/powerplay: remove useless soft pptable in Asic related backend

The soft pptable was used for re-uploading pptable as cache, but since
previous commits, the generic codes for uploading pptable are used and
backend is released during resetting powerplay. So it becomes redundance.

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