drm/amdgpu: improve sa_bo->fence by kernel fence

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>

drm/amdgpu: modify amdgpu_fence_wait_any() to amdgpu_fence_wait_multiple()

Rename the function and update the related code with this modified function.
Add the new parameter of bool wait_all.

If wait_all is true, it will return when all fences are signaled or timeout.
If wait_all is false, it will return when any fence is signaled or timeout.

Signed-off-by: Junwei Zhang <Jerry.Zhang@amd.com>
Reviewed-by: Monk Liu <monk.liu@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>

drm/amdgpu: fix IH ring allocation for bus addresses (v2)

Use pci_alloc_consistent rather than kzalloc since we
need 256 byte aligned memory for the ring buffer.

v2: fix copy paste typo in free function noticed
by Jammy.

bug:
https://bugs.freedesktop.org/show_bug.cgi?id=91749

Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org

drm/amdgpu: Don't link train DisplayPort on HPD until we get the dpcd

This is a port of:
DRM - radeon: Don't link train DisplayPort on HPD until we get the dpcd
to amdgpu.

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

DRM - radeon: Don't link train DisplayPort on HPD until we get the dpcd

Most of the time this isn't an issue since hotplugging an adaptor will
trigger a crtc mode change which in turn, causes the driver to probe
every DisplayPort for a dpcd. However, in cases where hotplugging
doesn't cause a mode change (specifically when one unplugs a monitor
from a DisplayPort connector, then plugs that same monitor back in
seconds later on the same port without any other monitors connected), we
never probe for the dpcd before starting the initial link training. What
happens from there looks like this:

- GPU has only one monitor connected. It's connected via
  DisplayPort, and does not go through an adaptor of any sort.

- User unplugs DisplayPort connector from GPU.

- Change in HPD is detected by the driver, we probe every
  DisplayPort for a possible connection.

- Probe the port the user originally had the monitor connected
  on for it's dpcd. This fails, and we clear the first (and only
  the first) byte of the dpcd to indicate we no longer have a
  dpcd for this port.

- User plugs the previously disconnected monitor back into the
  same DisplayPort.

- radeon_connector_hotplug() is called before everyone else,
  and tries to handle the link training. Since only the first
  byte of the dpcd is zeroed, the driver is able to complete
  link training but does so against the wrong dpcd, causing it
  to initialize the link with the wrong settings.

- Display stays blank (usually), dpcd is probed after the
  initial link training, and the driver prints no obvious
  messages to the log.

In theory, since only one byte of the dpcd is chopped off (specifically,
the byte that contains the revision information for DisplayPort), it's
not entirely impossible that this bug may not show on certain monitors.
For instance, the only reason this bug was visible on my ASUS PB238
monitor was due to the fact that this monitor using the enhanced framing
symbol sequence, the flag for which is ignored if the radeon driver
thinks that the DisplayPort version is below 1.1.

Signed-off-by: Stephen Chandler Paul <cpaul@redhat.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org

drm/amdgpu: fix CS error handling v2

Stop double freeing the the BO list by pulling the content
of amdgpu_cs_parser_prepare_job() into the IOCTL function again.

v2: better commit message

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com> (v1)
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: fix zeroing all IB fields manually v2

The problem now is that we don't necessarily call amdgpu_ib_get()
in some error paths and so work with uninitialized data.

Better require that the memory is already zeroed.

v2: better commit message

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com> (v1)
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: free the job immediately after dispatching it

Fixes a whole bunch of lockdep warnings.

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

drm/amdgpu: use a spinlock instead of a mutex for the rq

More appropriate and fixes some nasty lockdep warnings.

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

drm/amdgpu: don't need to use bo_list_clone any more

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>

drm/amdgpu: some clean up for cs_ioctl

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>

drm/amdgpu: prepare job should be common code path

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>

drm/amdgpu: abstract amdgpu_job for scheduler

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>

drm/amdgpu: bump the DRM version for new allowed mem-mapped registers

Used by mesa, etc. for profiling.

Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Marek Olšák <marek.olsak@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

Merge branch 'drm-next-fsl-dcu' of https://github.com/Jianwei-Wang/linux-drm-fsl-dcu into drm-next

Merge Freescale DCU FRM driver.

* 'drm-next-fsl-dcu' of https://github.com/Jianwei-Wang/linux-drm-fsl-dcu:
  MAINTAINERS: Add Freescale DCU DRM driver maintainer
  devicetree: Add NEC to the vendor-prefix list
  drm/layerscape: Add Freescale DCU DRM driver

MAINTAINERS: Add Freescale DCU DRM driver maintainer

Add Alison and myself as maintainers of the Freescale DCU DRM driver.

Signed-off-by: Alison Wang <b18965@freescale.com>
Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com>
Signed-off-by: Jianwei Wang <jianwei.wang.chn@gmail.com>

devicetree: Add NEC to the vendor-prefix list

NEC represent NEC LCD Technologies, Ltd.

Signed-off-by: Alison Wang <b18965@freescale.com>
Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com>
Signed-off-by: Jianwei Wang <jianwei.wang.chn@gmail.com>

drm/layerscape: Add Freescale DCU DRM driver

This patch add support for Two Dimensional Animation and Compositing
Engine (2D-ACE) on the Freescale SoCs.

2D-ACE is a Freescale display controller. 2D-ACE describes
the functionality of the module extremely well its name is a value
that cannot be used as a token in programming languages.
Instead the valid token "DCU" is used to tag the register names and
function names.

The Display Controller Unit (DCU) module is a system master that
fetches graphics stored in internal or external memory and displays
them on a TFT LCD panel. A wide range of panel sizes is supported
and the timing of the interface signals is highly configurable.
Graphics are read directly from memory and then blended in real-time,
which allows for dynamic content creation with minimal CPU
intervention.

The features:
(1) Full RGB888 output to TFT LCD panel.
(2) Blending of each pixel using up to 4 source layers
dependent
on size of panel.
(3) Each graphic layer can be placed with one pixel resolution
in either axis.
(4) Each graphic layer support RGB565 and RGB888 direct colors
without alpha channel and BGRA8888 BGRA4444 ARGB1555 direct
colors
with an alpha channel and YUV422 format.
(5) Each graphic layer support alpha blending with 8-bit
resolution.
This is a simplified version, only one primary plane, one
framebuffer, one crtc, one connector and one encoder for TFT
LCD panel.

Signed-off-by: Alison Wang <b18965@freescale.com>
Signed-off-by: Xiubo Li <lixiubo@cmss.chinamobile.com>
Signed-off-by: Jianwei Wang <jianwei.wang.chn@gmail.com>
Acked-by: Daniel Vetter <daniel.vetter@ffwll.ch>

Merge branch 'drm-atmel-hlcdc-devel' of https://github.com/bbrezillon/linux-at91 into drm-next

The following PR add support for 3 more atmel SoCs and for some missing
features (new input formats and PRIME support).

* 'drm-atmel-hlcdc-devel' of https://github.com/bbrezillon/linux-at91:
  drm: atmel-hlcdc: add support for sama5d4 SoCs
  drm: atmel-hlcdc: add support for at91sam9n12 SoC
  drm: atmel-hlcdc: add support for at91sam9x5 SoCs
  drm: atmel-hlcdc: add RGB565 and RGB444 output support
  drm: atmel-hlcdc: add the missing DRM_ATOMIC flag
  drm: atmel-hlcdc: add PRIME support

Merge branch 'drm-next-4.3' of git://people.freedesktop.org/~agd5f/linux into drm-next

amdgpu and radeon changes for 4.3. Highlights:
- Fiji support for amdgpu.
- CGS support for amdgpu.  This is a new driver
  internal cross-component API.
- Initial GPU scheduler for amdgpu.  Still disabled
  by default.
- Lots of bug fixes and optimizations

* 'drm-next-4.3' of git://people.freedesktop.org/~agd5f/linux: (130 commits)
  drm/amdgpu: wait on page directory changes. v2
  drm/amdgpu: Select BACKLIGHT_LCD_SUPPORT
  drm/radeon: Select BACKLIGHT_LCD_SUPPORT
  drm/amdgpu: cleanup sheduler rq handling v2
  drm/amdgpu: move prepare work out of scheduler to cs_ioctl
  drm/amdgpu: fix unnecessary wake up
  drm/amdgpu: fix duplicated mapping invoke bug
  drm/amdgpu: drop bo_list_clone when no scheduler
  drm/amdgpu: disable GPU reset by default
  drm/amdgpu: fix type mismatch error
  drm/amdgpu: add reference for **fence
  drm/amdgpu: fix waiting for all fences before flipping
  drm/amdgpu: fix UVD return code checking
  drm/amdgpu: remove scheduler fence list v2
  drm/amdgpu: remove amd_sched_wait_emit v2
  drm/amdgpu: remove unecessary scheduler fence callbacks
  drm/amdgpu: fix scheduler fence implementation
  drm/amdgpu: don't grab dev->struct_mutex in pm functions
  drm/amdgpu: Don't take dev->struct_mutex in bo_force_delete
  drm/radeon: Don't take dev->struct_mutex in pm functions
  ...

drm: atmel-hlcdc: add support for sama5d4 SoCs

Describe capabilities of the HLCDC IP found on sama5d4 SoCs and add a
new entry to the atmel_hlcdc_of_match table.

Signed-off-by: Boris Brezillon <boris.brezillon@free-electrons.com>

drm: atmel-hlcdc: add support for at91sam9n12 SoC

Describe capabilities of the HLCDC IP found on at91sam9n12 SoC and add a
new entry to the atmel_hlcdc_of_match table.

Signed-off-by: Boris Brezillon <boris.brezillon@free-electrons.com>

drm: atmel-hlcdc: add support for at91sam9x5 SoCs

Describe capabilities of the HLCDC IP found on at91sam9x5 SoCs and add a
new entry to the atmel_hlcdc_of_match table.

Signed-off-by: Boris Brezillon <boris.brezillon@free-electrons.com>

drm: atmel-hlcdc: add RGB565 and RGB444 output support

The HLCDC IP supports RGB565 and RGB444 output formats.

Signed-off-by: Boris Brezillon <boris.brezillon@free-electrons.com>

drm: atmel-hlcdc: add the missing DRM_ATOMIC flag

The atmel-hlcdc driver already supports atomic operations, add the
missing DRM_ATOMIC flag to expose the atomic features to userspace.

Signed-off-by: Boris Brezillon <boris.brezillon@free-electrons.com>

drm: atmel-hlcdc: add PRIME support

Signed-off-by: Boris Brezillon <boris.brezillon@free-electrons.com>

drm/amdgpu: wait on page directory changes. v2

Pagetables can be moved and therefore the page directory update can be necessary
for the current cs even if none of the the bo's are moved. In that scenario
there is no fence between the sdma0 and gfx ring, so we add one.

v2 (chk): rebased

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

drm/amdgpu: Select BACKLIGHT_LCD_SUPPORT

Explicitly select BACKLIGHT_LCD_SUPPORT to satisfy the direct dependency
of BACKLIGHT_CLASS_DEVICE.

Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: Christian König <christian.koenig@amd.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/radeon: Select BACKLIGHT_LCD_SUPPORT

Explicitly select BACKLIGHT_LCD_SUPPORT to satisfy the direct dependency
of BACKLIGHT_CLASS_DEVICE.

Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: Christian König <christian.koenig@amd.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: cleanup sheduler rq handling v2

Rework run queue implementation, especially remove the odd list handling.

v2: cleanup the code only, no algorithem change.

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

drm/amdgpu: move prepare work out of scheduler to cs_ioctl

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>

drm/amdgpu: fix unnecessary wake up

decrease CPU extra overhead.

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>

drm/amdgpu: fix duplicated mapping invoke bug

fix the bug that there is duplicated bo_update_mapping issued

Signed-off-by: monk.liu <monk.liu@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>

drm/amdgpu: drop bo_list_clone when no scheduler

bo_list_clone() will take a lot of time when bo_list hold too much
elements, like above 7000

Signed-off-by: Monk.Liu <monk.liu@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Jammy Zhou <jammy.zhou@amd.com>

drm/amdgpu: disable GPU reset by default

It's not validated yet and causes more harm than good.
Avoids spurious resets.

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

drm/amdgpu: fix type mismatch error

remaining timeout returned by amdgpu_fence_wait_any can be larger than
max int value, thus the truncated 32 bit value in r ends up being
negative while its original long value is positive.

Signed-off-by: monk.liu <monk.liu@amd.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <jammy.zhou@amd.com>

drm/amdgpu: add reference for **fence

fix fence is released when pass to **fence sometimes.
add reference for it.

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>

drm/amdgpu: fix waiting for all fences before flipping

Otherwise we might see corruption.

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

drm/amdgpu: fix UVD return code checking

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Tested-and-Reviewed-by: Leo Liu <leo.liu@amd.com>

drm/amdgpu: remove scheduler fence list v2

Unused and missing proper locking.

v2: add locking comment to commit message.

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com> (v1)

drm/amdgpu: remove amd_sched_wait_emit v2

Not used any more.

v2: remove amd_sched_emit as well.

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

drm/amdgpu: remove unecessary scheduler fence callbacks

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

drm/amdgpu: fix scheduler fence implementation

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

drm/amdgpu: don't grab dev->struct_mutex in pm functions

Similar to radeon, except that amdgpu doesn't even use struct_mutex to
protect anything like the shared z buffer (sane gpu architecture,
yay!). And the code already grabs the globa adev->ring_lock, so this
code can't race with itself. Which makes struct_mutex completely
redundnant. Remove it.

Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: "Christian König" <christian.koenig@amd.com>
Reviewed-by: Thierry Reding <treding@nvidia.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: Don't take dev->struct_mutex in bo_force_delete

It really doesn't protect anything which doesn't have other locks
already. Also this is run from driver unload code so not much need for
locks anyway.

Same changes as for radeon really.

Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: "Christian König" <christian.koenig@amd.com>
Reviewed-by: Thierry Reding <treding@nvidia.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/radeon: Don't take dev->struct_mutex in pm functions

We already grab 2 device-global locks (write-sema rdev->pm.mclk_lock
and rdev->ring_lock), adding another global mutex won't serialize this
code more. And since there's really nothing interesting that gets
protected in radeon by dev->struct mutex (we only have the global z
buffer owners and it's still serializing gem bo destruction in the drm
core - which is irrelevant since radeon uses ttm anyway internally)
this doesn't add protection. Remove it.

Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: "Christian König" <christian.koenig@amd.com>
Reviewed-by: Thierry Reding <treding@nvidia.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/radeon: Don't take dev->struct_mutex in bo_force_delete

It really doesn't protect anything which doesn't have other locks
already. Also this is run from driver unload code so not much need for
locks anyway.

Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: "Christian König" <christian.koenig@amd.com>
Reviewed-by: Thierry Reding <treding@nvidia.com>
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>

drm/amdgpu: remove VI hw bug workaround v3

The workaround simply doesn't work because VM mappings
are controlled by userspace not the kernel.

Additional to that this is just a performance problem
which happens if you have holes in your VM mapping.

v2: adjust virtual addr alignment as well.
v3: fix trivial warning

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Monk Liu <monk.liu@amd.com> (v1)
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com> (v2)

drm/amdgpu: cleanup amdgpu_fence_ring_wait_seq

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

drm/amdgpu: remove duplicate amdgpu_fence_process implementation

Looks like that somehow got missed while during porting the radeon changes.

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

drm/amdgpu: remove amdgpu_fence_wait

It was just a wrapper for fence_wait anyway.

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

drm/amdgpu: use the reservation obj wait for the UVD msg

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

drm/amdgpu: remove amdgpu_fence_signaled

The common kernel function does the same thing.

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

drm/amd: add scheduler fence implementation (v2)

scheduler fence is based on kernel fence framework.

v2: squash in Christian's build fix

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>

drm/amdgpu: use kernel submit helper in vm

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>

drm/amdgpu: use amd_sched_job in its backend ops

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>

drm/amdgpu: cleanup and fix scheduler fence handling v2

v2: rebased

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com> (v1)
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>

drm/amdgpu: merge amd_sched_entity and amd_context_entity v2

Avoiding a couple of casts.

v2: rename c_entity to entity as well

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

drm/amdgpu: fix coding style in a couple of places

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>

drm/amdgpu: remove unused parent entity

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>

drm/amdgpu: process sched job exactly triggered by fence signal

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>

Revert "drm/amdgpu: return new seq_no for amd_sched_push_job"

This reverts commit d1d33da8eb86b8ca41dd9ed95738030df5267b95.

Reviewed-by: Christian K?nig <christian.koenig@amd.com>
Conflicts:
drivers/gpu/drm/amd/amdgpu/amdgpu_sched.c
drivers/gpu/drm/amd/amdgpu/amdgpu_vm.c

drm/amdgpu: cleanup amdgpu_ctx inti/fini v2

Cleanup the kernel context handling.

v2: rebased

Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com> (v1)

drm/amdgpu: stop leaking the ctx id into the scheduler v2

Id's are for the IOCTL ABI only.

v2: remove tgid as well

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

drm/amdgpu: cleanup ctx_mgr init/fini

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

drm/amdgpu: fix bo list handling in CS

We didn't initialized the mutex in the cloned bo list resulting in nice
warnings from lockdep. Also fixes error handling in this function.

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

drm/amdgpu: reorder the code to avoid forward declerations

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

drm/amdgpu: move sched job process from isr to fence callback

This way can avoid interrupt lost, and can process sched job exactly.

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Jammy Zhou <Jammy.Zhou@amd.com>

drm/amdgpu: add amd_sched_next_queued_seq function

This function is used to get the next queued sequence number

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

drm/amdgpu: make last_handled_seq atomic

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

drm/amdgpu: add amd_sched_commit

This function is to update last_emitted_v_seq and wake up the waiters.

It should be called by driver in the run_job backend function

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

drm/amdgpu: return new seq_no for amd_sched_push_job

It is clean to update last_queued_v_seq in the scheduler module

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

drm/amdgpu: some code refinement v2

Fix the code alignment, etc.

v2: rebase the code

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

drm/amdgpu: fix null pointer by previous cleanup

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>

drm/amdgpu: add kernel fence in ib_submit_kernel_helper

every sbumission should be able to get a fence.

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <jammy.zhou@amd.com>

drm/amdgpu: use kernel fence for sdma ib test

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <jammy.zhou@amd.com>

drm/amdgpu: use kernel fence for gfx ib test

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <jammy.zhou@amd.com>

drm/amdgpu: use kernel fence in amdgpu_test

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <jammy.zhou@amd.com>

drm/amdgpu: use kernel fence for vce ib test

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <jammy.zhou@amd.com>

drm/amdgpu: change uvd ib test to use kernel fence directly

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <jammy.zhou@amd.com>

drm/amdgpu: use kernel fence for last_pt_update

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <jammy.zhou@amd.com>

drm/amdgpu: use kernel fence diretly in amdgpu_bo_fence

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>
Reviewed-by: Jammy Zhou <jammy.zhou@amd.com>

drm/amdgpu: clean up amd sched wait_ts and wait_signal

Remove code not used at the moment.

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

drm/amdgpu: stop using addr to check for BO move v3

It is theoretically possible that a swapped out BO gets the
same GTT address, but different backing pages while being swapped in.

Instead just use another VA state to note updated areas.
Ported from not upstream yet radeon commit with the same name.

v2: fix some bugs in the original implementation found in the radeon code.
v3: squash in VCE/UVD fix

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

drm/amdgpu: move wait_queue_head from adev to ring (v2)

thus unnecessary wake up could be avoid between rings
v2:
move wait_queue_head to fence_drv from ring

Signed-off-by: monk.liu <monk.liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>

drm/amdgpu: re-implement fence_default_wait

use fence_wait_any to implement fence_default_wait

Signed-off-by: monk.liu <monk.liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>

drm/amdgpu: new implement for fence_wait_any (v2)

origninal method will sleep/schedule at the granurarity of HZ/2 and
based on seq signal method, the new implement is based on kernel fance
interface, no unnecessary schedule at all

v2: replace logic of original amdgpu_fence_wait_any

Signed-off-by: monk.liu <monk.liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>

drm/amdgpu: use kernel fence interface when possible

Signed-off-by: monk.liu <monk.liu@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>

drm/amdgpu: use scheduler for VCE ib test

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>

drm/amdgpu: use scheduler for UVD ib test

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>

drm/amdgpu: use gpu scheduler for sdma ib test

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>

drm/amdgpu: Use gpu scheduler for gfx ring ib test

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>

drm/amdgpu: add helper function for kernel submission

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>

drm/amdgpu: fix seq in ctx_add_fence

if enabling scheduler, then the queued seq is assigned
when pushing job before emitting job.

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>

drm/amdgpu: wait forever for wait emit

the job must be emitted by scheduler, otherwise scheduler is abnormal.

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>

drm/amdgpu: add amdgpu.sched_hw_submission option

This option can be used to specify the max number of submissions in the
active HW queue. The default value is 2 now.

Signed-off-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>

drm/amdgpu: add amdgpu.sched_jobs option

This option can be used to specify the max job number in the job queue,
and it is 16 by default.

Signed-off-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>

drm/amdgpu: silent the message for GPU scheduler creation

Signed-off-by: Jammy Zhou <Jammy.Zhou@amd.com>
Reviewed-by: Chunming Zhou <david1.zhou@amd.com>

drm/amdgpu: fix syncing to VM updates

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

drm/amdgpu: add check for callback

it is possible that the callback isn't defined sometimes.

Signed-off-by: Chunming Zhou <david1.zhou@amd.com>
Reviewed-by: Christian K?nig <christian.koenig@amd.com>