OSDN Git Service

android-x86/kernel.git
5 years agodrm/amdgpu/powerplay: fix typo in BACO header guards
Alex Deucher [Fri, 15 Feb 2019 22:20:04 +0000 (17:20 -0500)]
drm/amdgpu/powerplay: fix typo in BACO header guards

s/BOCO/BACO/g

Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu/powerplay: fix return codes in BACO code
Alex Deucher [Fri, 15 Feb 2019 22:14:36 +0000 (17:14 -0500)]
drm/amdgpu/powerplay: fix return codes in BACO code

Use a proper return code rather than -1.

Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu: add missing license on baco files
Alex Deucher [Mon, 11 Feb 2019 02:05:31 +0000 (21:05 -0500)]
drm/amdgpu: add missing license on baco files

Trivial.

Reviewed-by: Hawking Zhang <Hawking.Zhang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu: partial revert cleanup setting bulk_movable v2
Christian König [Wed, 30 Jan 2019 12:41:05 +0000 (13:41 +0100)]
drm/amdgpu: partial revert cleanup setting bulk_movable v2

We still need to set bulk_movable to false when new BOs are added or removed.

v2: also set it to false on removal

Signed-off-by: Christian König <christian.koenig@amd.com>
Tested-by: StDenis, Tom <Tom.StDenis@amd.com>
Tested-by: Przemek Socha <soprwa@gmail.com>
Reviewed-by: Zhou, David(ChunMing) <David1.Zhou@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu: cleanup setting bulk_movable
Christian König [Mon, 28 Jan 2019 12:41:58 +0000 (13:41 +0100)]
drm/amdgpu: cleanup setting bulk_movable

We only need to set this to false now when BOs are removed from the LRU.

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>
5 years agodrm/amd/powerplay/smu10_hwmgr: use struct_size() in kzalloc()
Gustavo A. R. Silva [Tue, 19 Feb 2019 18:55:09 +0000 (12:55 -0600)]
drm/amd/powerplay/smu10_hwmgr: use struct_size() in kzalloc()

One of the more common cases of allocation size calculations is finding
the size of a structure that has a zero-sized array at the end, along
with memory for some number of elements for that array. For example:

struct foo {
    int stuff;
    struct boo entry[];
};

size = sizeof(struct foo) + count * sizeof(struct boo);
instance = kzalloc(size, GFP_KERNEL);

Instead of leaving these open-coded and prone to type mistakes, we can
now use the new struct_size() helper:

instance = kzalloc(struct_size(instance, entry, count), GFP_KERNEL);

Notice that, in this case, variable table_size is not necessary, hence
it is removed.

This code was detected with the help of Coccinelle.

Signed-off-by: Gustavo A. R. Silva <gustavo@embeddedor.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amd/powerplay/smu8_hwmgr: use struct_size() in kzalloc()
Gustavo A. R. Silva [Tue, 19 Feb 2019 18:42:32 +0000 (12:42 -0600)]
drm/amd/powerplay/smu8_hwmgr: use struct_size() in kzalloc()

One of the more common cases of allocation size calculations is finding
the size of a structure that has a zero-sized array at the end, along
with memory for some number of elements for that array. For example:

struct foo {
    int stuff;
    struct boo entry[];
};

size = sizeof(struct foo) + count * sizeof(struct boo);
instance = kzalloc(size, GFP_KERNEL);

Instead of leaving these open-coded and prone to type mistakes, we can
now use the new struct_size() helper:

instance = kzalloc(struct_size(instance, entry, count), GFP_KERNEL);

Notice that, in this case, variable table_size is not necessary, hence
it is removed.

This code was detected with the help of Coccinelle.

Signed-off-by: Gustavo A. R. Silva <gustavo@embeddedor.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amd/display: Refactor for setup periodic interrupt.
Yongqiang Sun [Fri, 25 Jan 2019 19:40:14 +0000 (14:40 -0500)]
drm/amd/display: Refactor for setup periodic interrupt.

[Why]
Current periodic interrupt start point calc in optc
is not clear.

[How]
1. DM convert delta time to lines number and dc will calculate the
   start position as per lines number and interrupt type.
2. hwss calculates the start point as per line offset.
3. optc programs vertical interrupts register as per start point
   and interrupt source.

Signed-off-by: Yongqiang Sun <yongqiang.sun@amd.com>
Reviewed-by: Tony Cheng <Tony.Cheng@amd.com>
Acked-by: Leo Li <sunpeng.li@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amd/display: Clear stream->mode_changed after commit
Nicholas Kazlauskas [Fri, 25 Jan 2019 20:23:09 +0000 (15:23 -0500)]
drm/amd/display: Clear stream->mode_changed after commit

[Why]
The stream->mode_changed flag can persist in the following sequence
of atomic commits:

Commit 1:
Enable CRTC0 (mode_changed = true), Enable CRTC1 (mode_changed = true)

Commit 2:
Disable CRTC1 (mode_changed = false)

In this sequence we want to keep the exiting CRTC0 but it's not in the
atomic state for the commit since it hasn't been modified. In this case
the stream->mode_changed flag persists as true and we don't re-program
the planes for the existing stream.

[How]
The flag needs to be cleared and it makes the most sense to do it within
DC after the state has been committed. Nothing following dc_commit_state
should think that the stream's mode has changed.

Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: Leo Li <sunpeng.li@amd.com>
Acked-by: Tony Cheng <Tony.Cheng@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amd/display: Do cursor updates after stream updates
Nicholas Kazlauskas [Fri, 25 Jan 2019 20:30:24 +0000 (15:30 -0500)]
drm/amd/display: Do cursor updates after stream updates

[Why]
Cursor updates used to happen after vblank/flip/stream updates before
the stream update refactor. They now happen before stream updates
which means that they're not going to be synced with fb changes
and that they're going to programmed for pipes that we're disabling
within the same commit.

[How]
Move them after stream updates.

Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: David Francis <David.Francis@amd.com>
Acked-by: Leo Li <sunpeng.li@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amd/display: Fix update type mismatches in atomic check
Nicholas Kazlauskas [Tue, 22 Jan 2019 19:09:34 +0000 (14:09 -0500)]
drm/amd/display: Fix update type mismatches in atomic check

[Why]
Whenever a stream or plane is added or removed from the context the
pointer will change from old to new. We set lock and validation
needed in these cases. But not all of these cases match update_type
from dm_determine_update_type_for_commit - an example being overlay
plane updates.

There are warnings for a few of these cases that should be fixed.

[How]
We can closer align to DC (and lock_and_validation_needed) by
comparing stream and plane pointers.

Since the old stream/old plane state is never freed until sometime
after the commit tail work finishes we are guaranteed to never get
back the same block of memory when we remove and create a stream or
plane state in the same commit.

Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: Leo Li <sunpeng.li@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amd/display: Don't expose support for DRM_FORMAT_RGB888
Nicholas Kazlauskas [Mon, 21 Jan 2019 14:44:47 +0000 (09:44 -0500)]
drm/amd/display: Don't expose support for DRM_FORMAT_RGB888

[Why]
This format isn't supported in DC and some IGT tests fail since we
expose support for it.

[How]
Remove it.

Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: Harry Wentland <Harry.Wentland@amd.com>
Reviewed-by: Leo Li <sunpeng.li@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amd/display: Fix wrong z-order when updating overlay planes
Nicholas Kazlauskas [Fri, 18 Jan 2019 18:17:55 +0000 (13:17 -0500)]
drm/amd/display: Fix wrong z-order when updating overlay planes

[Why]
If a commit updates an overlay plane via the legacy plane IOCTL
then the only plane in the state will be the overlay plane.

Overlay planes need to be added first to the DC context, but in the
scenario above the plane will be added last. This will result in wrong
z-order during rendering.

[How]
If any non-cursor plane has been updated then the rest of the
non-cursor planes should be added to the CRTC state.

The cursor plane doesn't need to be included for stream updates and
locking it will cause performance issues. It should be ignored.

DC requires that the surface count passed during stream updates
be the number of surfaces currently on the stream to enable fast
updates. This previously wasn't the case without this patch, so this
also allows this optimization to occur.

Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: Leo Li <sunpeng.li@amd.com>
Acked-by: Tony Cheng <Tony.Cheng@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amd/display: send pipe set command to dmcu when backlight is set
Josip Pavic [Wed, 30 Jan 2019 21:23:47 +0000 (16:23 -0500)]
drm/amd/display: send pipe set command to dmcu when backlight is set

[Why]
Previously, a change removed code that would send a pipe set command
to dmcu each time the backlight was set, as it was thought to be
superfluous. However, it is possible for the backlight to be set
before a valid pipe has been set, which causes DMCU to hang after a
DPMS restore on some systems.

[How]
Send a pipe set command to DMCU prior to setting the backlight.

Signed-off-by: Josip Pavic <Josip.Pavic@amd.com>
Reviewed-by: Anthony Koo <Anthony.Koo@amd.com>
Acked-by: Leo Li <sunpeng.li@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amd/display: Ungate stream before programming registers
Gary Kattan [Fri, 25 Jan 2019 23:04:14 +0000 (15:04 -0800)]
drm/amd/display: Ungate stream before programming registers

[Why]
Certain tests fail after a fresh reboot. This is caused by writing to
registers prior to ungating the stream we're trying to program.

[How]
Make sure the stream is ungated before writing to its registers.
This also enables power-gating plane resources before init_hw
initializes them.
Additionally, this does some refactoring to move gating/ungating
from enable/disable_plane functions to where stream resources are
enabled/disabled.

Signed-off-by: Gary Kattan <gary.kattan@amd.com>
Reviewed-by: Dmytro Laktyushkin <Dmytro.Laktyushkin@amd.com>
Acked-by: Leo Li <sunpeng.li@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amd/display: Increase precision for backlight curve
Anthony Koo [Wed, 30 Jan 2019 16:05:38 +0000 (11:05 -0500)]
drm/amd/display: Increase precision for backlight curve

[Why]
We are currently losing precision when we convert from
16 bit --> 8 bit --> 16 bit.

[How]
We shouldn't down convert unnecessarily and lose precision.
Keep values at 16 bit and use directly.

Signed-off-by: Anthony Koo <Anthony.Koo@amd.com>
Reviewed-by: Aric Cyr <Aric.Cyr@amd.com>
Acked-by: Leo Li <sunpeng.li@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amd/display: remove screen flashes on seamless boot
Anthony Koo [Sun, 20 Jan 2019 06:54:01 +0000 (01:54 -0500)]
drm/amd/display: remove screen flashes on seamless boot

[Why]
We want boot to desktop to be seamless

[How]
During init pipes, avoid touching the pipes where GOP has already
enabled the HW to the state we want.

Signed-off-by: Anthony Koo <Anthony.Koo@amd.com>
Reviewed-by: Aric Cyr <Aric.Cyr@amd.com>
Acked-by: Leo Li <sunpeng.li@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amd/display: send pipe set command to dmcu when stream unblanks
Josip Pavic [Tue, 29 Jan 2019 19:15:03 +0000 (14:15 -0500)]
drm/amd/display: send pipe set command to dmcu when stream unblanks

[Why]
When stream is blanked, pipe set command is sent to dmcu to notify it
that the abm pipe is disabled. When stream is unblanked, no notification is
made to dmcu that the abm pipe has been enabled, resulting in abm not
being enabled in the firmware.

[How]
When stream is unblanked, send a pipe set command to dmcu.

Signed-off-by: Josip Pavic <Josip.Pavic@amd.com>
Reviewed-by: Anthony Koo <Anthony.Koo@amd.com>
Acked-by: Leo Li <sunpeng.li@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agoRevert "drm/amdgpu: Fix bugs in setting CP RB/MEC DOORBELL_RANGE registers"
Yong Zhao [Tue, 19 Feb 2019 16:21:51 +0000 (11:21 -0500)]
Revert "drm/amdgpu: Fix bugs in setting CP RB/MEC DOORBELL_RANGE registers"

The original change caused a regression, so revert it until the new fix
is ready.

BUG: https://bugs.freedesktop.org/show_bug.cgi?id=109650

This reverts commit 764c85fef41722db0f21558c6c2fb38bee172d19.

Signed-off-by: Yong Zhao <Yong.Zhao@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agoRevert "drm/amdgpu: Delete user queue doorbell variables"
Yong Zhao [Tue, 19 Feb 2019 16:21:41 +0000 (11:21 -0500)]
Revert "drm/amdgpu: Delete user queue doorbell variables"

This reverts commit 9006c6bd9059cb9807fa863bafc1d776222cb61b.

Signed-off-by: Yong Zhao <Yong.Zhao@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/radeon/evergreen_cs: fix missing break in switch statement
Gustavo A. R. Silva [Fri, 15 Feb 2019 20:29:26 +0000 (14:29 -0600)]
drm/radeon/evergreen_cs: fix missing break in switch statement

Add missing break statement in order to prevent the code from falling
through to case CB_TARGET_MASK.

This bug was found thanks to the ongoing efforts to enable
-Wimplicit-fallthrough.

Fixes: dd220a00e8bd ("drm/radeon/kms: add support for streamout v7")
Cc: stable@vger.kernel.org
Signed-off-by: Gustavo A. R. Silva <gustavo@embeddedor.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu/si_dpm: Mark expected switch fall-throughs
Gustavo A. R. Silva [Fri, 15 Feb 2019 16:46:43 +0000 (10:46 -0600)]
drm/amdgpu/si_dpm: Mark expected switch fall-throughs

In preparation to enabling -Wimplicit-fallthrough, mark switch
cases where we are expecting to fall through.

Warning level 3 was used: -Wimplicit-fallthrough=3

This patch is part of the ongoing efforts to enable
-Wimplicit-fallthrough.

Signed-off-by: Gustavo A. R. Silva <gustavo@embeddedor.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/radeon/ci_dpm: Mark expected switch fall-throughs
Gustavo A. R. Silva [Fri, 15 Feb 2019 17:08:18 +0000 (11:08 -0600)]
drm/radeon/ci_dpm: Mark expected switch fall-throughs

In preparation to enabling -Wimplicit-fallthrough, mark switch
cases where we are expecting to fall through.

Warning level 3 was used: -Wimplicit-fallthrough=3

This patch is part of the ongoing efforts to enable
-Wimplicit-fallthrough.

Signed-off-by: Gustavo A. R. Silva <gustavo@embeddedor.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amd/powerplay/smu7_hwmgr: Mark expected switch fall-throughs
Gustavo A. R. Silva [Fri, 15 Feb 2019 16:57:17 +0000 (10:57 -0600)]
drm/amd/powerplay/smu7_hwmgr: Mark expected switch fall-throughs

In preparation to enabling -Wimplicit-fallthrough, mark switch
cases where we are expecting to fall through.

Warning level 3 was used: -Wimplicit-fallthrough=3

This patch is part of the ongoing efforts to enable
-Wimplicit-fallthrough.

Signed-off-by: Gustavo A. R. Silva <gustavo@embeddedor.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amd/display/dce_mem_input: Mark expected switch fall-through
Gustavo A. R. Silva [Fri, 15 Feb 2019 16:54:42 +0000 (10:54 -0600)]
drm/amd/display/dce_mem_input: Mark expected switch fall-through

In preparation to enabling -Wimplicit-fallthrough, mark switch
cases where we are expecting to fall through.

Warning level 3 was used: -Wimplicit-fallthrough=3

Notice that, in this particular case, the code comment is modified
in accordance with what GCC is expecting to find.

This patch is part of the ongoing efforts to enable
-Wimplicit-fallthrough.

Signed-off-by: Gustavo A. R. Silva <gustavo@embeddedor.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/radeon/si_dpm: Mark expected switch fall-throughs
Gustavo A. R. Silva [Fri, 15 Feb 2019 17:10:31 +0000 (11:10 -0600)]
drm/radeon/si_dpm: Mark expected switch fall-throughs

In preparation to enabling -Wimplicit-fallthrough, mark switch
cases where we are expecting to fall through.

Warning level 3 was used: -Wimplicit-fallthrough=3

This patch is part of the ongoing efforts to enable
-Wimplicit-fallthrough.

Signed-off-by: Gustavo A. R. Silva <gustavo@embeddedor.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amd/display/dc/bios_parser2: Mark expected switch fall-throughs
Gustavo A. R. Silva [Fri, 15 Feb 2019 16:50:23 +0000 (10:50 -0600)]
drm/amd/display/dc/bios_parser2: Mark expected switch fall-throughs

In preparation to enabling -Wimplicit-fallthrough, mark switch
cases where we are expecting to fall through.

Warning level 3 was used: -Wimplicit-fallthrough=3

This patch is part of the ongoing efforts to enable
-Wimplicit-fallthrough.

Signed-off-by: Gustavo A. R. Silva <gustavo@embeddedor.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdkfd: Optimize out sdma doorbell array in kgd2kfd_shared_resources
Yong Zhao [Thu, 10 Jan 2019 04:31:14 +0000 (23:31 -0500)]
drm/amdkfd: Optimize out sdma doorbell array in kgd2kfd_shared_resources

We can directly calculate sdma doorbell indexes in the process doorbell
pages through the doorbell_index structure in amdgpu_device, so no need
to cache them in kgd2kfd_shared_resources any more. This alleviates the
adaptation needs when new SDMA configurations are introduced.

Signed-off-by: Yong Zhao <Yong.Zhao@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdkfd: Fix bugs regarding CP queue doorbell mask on SOC15
Yong Zhao [Wed, 13 Feb 2019 18:15:05 +0000 (13:15 -0500)]
drm/amdkfd: Fix bugs regarding CP queue doorbell mask on SOC15

Reserved doorbells for SDMA IH and VCN were not properly masked out
when allocating doorbells for CP user queues. This patch fixed that.

Signed-off-by: Yong Zhao <Yong.Zhao@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu: Add first_non_cp and last_non_cp in amdgpu_doorbell_index
Yong Zhao [Wed, 13 Feb 2019 18:13:50 +0000 (13:13 -0500)]
drm/amdgpu: Add first_non_cp and last_non_cp in amdgpu_doorbell_index

They will be used to inform KFD the doorbell range not usable for CP.

Signed-off-by: Yong Zhao <Yong.Zhao@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdkfd: Move a constant definition around
Yong Zhao [Tue, 15 Jan 2019 23:51:27 +0000 (18:51 -0500)]
drm/amdkfd: Move a constant definition around

The similar definitions should be consecutive.

Signed-off-by: Yong Zhao <Yong.Zhao@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu: remove some old unused dpm helpers
Alex Deucher [Thu, 14 Feb 2019 20:54:57 +0000 (15:54 -0500)]
drm/amdgpu: remove some old unused dpm helpers

Carried over from radeon, but no longer used.

Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu: Add command to override the context priority.
Bas Nieuwenhuizen [Wed, 30 Jan 2019 01:53:22 +0000 (02:53 +0100)]
drm/amdgpu: Add command to override the context priority.

Given a master fd we can then override the priority of the context
in another fd.

Using these overrides was recommended by Christian instead of trying
to submit from a master fd, and I am adding a way to override a
single context instead of the entire process so we can only upgrade
a single Vulkan queue and not effectively the entire process.

Reused the flags field as it was checked to be 0 anyways, so nothing
used it. This is source-incompatible (due to the name change), but
ABI compatible.

Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu: Check if fd really is an amdgpu fd.
Bas Nieuwenhuizen [Wed, 30 Jan 2019 01:53:21 +0000 (02:53 +0100)]
drm/amdgpu: Check if fd really is an amdgpu fd.

Otherwise we interpret the file private data as drm & amdgpu data
while it might not be, possibly allowing one to get memory corruption.

Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu: Only add rqs for initialized rings.
Bas Nieuwenhuizen [Wed, 30 Jan 2019 01:53:20 +0000 (02:53 +0100)]
drm/amdgpu: Only add rqs for initialized rings.

I don't see another way to figure out if a ring is initialized if
the hardware block might not be initialized.

Entities have been fixed up to handle num_rqs = 0.

Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/sched: Fix entities with 0 rqs.
Bas Nieuwenhuizen [Wed, 30 Jan 2019 01:53:19 +0000 (02:53 +0100)]
drm/sched: Fix entities with 0 rqs.

Some blocks in amdgpu can have 0 rqs.

Job creation already fails with -ENOENT when entity->rq is NULL,
so jobs cannot be pushed. Without a rq there is no scheduler to
pop jobs, and rq selection already does the right thing with a
list of length 0.

So the operations we need to fix are:
  - Creation, do not set rq to rq_list[0] if the list can have length 0.
  - Do not flush any jobs when there is no rq.
  - On entity destruction handle the rq = NULL case.
  - on set_priority, do not try to change the rq if it is NULL.

Signed-off-by: Bas Nieuwenhuizen <bas@basnieuwenhuizen.nl>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu: don't clamp debugfs register access to the BAR size
Alex Deucher [Mon, 11 Feb 2019 21:49:47 +0000 (16:49 -0500)]
drm/amdgpu: don't clamp debugfs register access to the BAR size

This prevents us from accessing extended registers in tools like
umr.  The register access functions already check if the offset
is beyond the BAR size and use the indirect accessors with locking
so this is safe.

Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu: fix several indentation issues
Colin Ian King [Tue, 12 Feb 2019 14:05:08 +0000 (14:05 +0000)]
drm/amdgpu: fix several indentation issues

There are several statements that are incorrectly indented. Fix these.

Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Colin Ian King <colin.king@canonical.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu/powerplay: declare firmware for CI cards
Alex Deucher [Wed, 13 Feb 2019 20:07:06 +0000 (15:07 -0500)]
drm/amdgpu/powerplay: declare firmware for CI cards

Missing firmware declaration caused firmware requirement to
not be noted by the module and may cause firmware to not
be available in initrd.

Fixes: bc4b539e385088 "drm/amdgpu: remove old CI DPM implementation"
Reviewed-by: James Zhu <James.Zhu@amd.com>
Tested-by: James Zhu <James.Zhu@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amd/display: Fix deadlock with display during hanged ring recovery.
Andrey Grodzovsky [Wed, 13 Feb 2019 18:53:45 +0000 (13:53 -0500)]
drm/amd/display: Fix deadlock with display during hanged ring recovery.

When ring hang happens amdgpu_dm_commit_planes during flip is holding
the BO reserved and then stack waiting for fences to signal in
reservation_object_wait_timeout_rcu (which won't signal because there
was a hnag). Then when we try to shutdown display block during reset
recovery from drm_atomic_helper_suspend we also try to reserve the BO
from dm_plane_helper_cleanup_fb ending in deadlock.
Also remove useless WARN_ON

Signed-off-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu: Delete user queue doorbell variables
Yong Zhao [Fri, 1 Feb 2019 23:36:21 +0000 (18:36 -0500)]
drm/amdgpu: Delete user queue doorbell variables

They are no longer used, so delete them to avoid confusion.

Signed-off-by: Yong Zhao <Yong.Zhao@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu: Fix bugs in setting CP RB/MEC DOORBELL_RANGE registers
Yong Zhao [Tue, 5 Feb 2019 20:17:40 +0000 (15:17 -0500)]
drm/amdgpu: Fix bugs in setting CP RB/MEC DOORBELL_RANGE registers

CP_RB_DOORBELL_RANGE_LOWER/UPPER and CP_MEC_DOORBELL_RANGE_LOWER/UPPER
are used for waking up an idle scheduler and for power gating support.
Usually the first few doorbells in pci doorbell bar are used for RB
and all leftover for MEC. This patch fixes the incorrect settings.

Theoretically, gfx ring doorbells should come before all MEC doorbells
to be consistent with the design. However, since the doorbell
allocations are agreed by all and we are not free to change them, also
considering the kernel MEC ring doorbells which are before gfx ring
doorbells are not used often, we compromise by leaving the doorbell
allocations unchanged.

Signed-off-by: Yong Zhao <Yong.Zhao@amd.com>
Reviewed-by: Felix Kuehling <Felix.Kuehling@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu: Simplify eviction fence handling
Felix Kuehling [Mon, 4 Feb 2019 23:17:26 +0000 (18:17 -0500)]
drm/amdgpu: Simplify eviction fence handling

Temporarily removing eviction fences to avoid triggering them by
accident is no longer necessary due to the fence_owner logic in
amdgpu_sync_resv.

As a result the ef_list usage of amdgpu_amdkfd_remove_eviction_fence
and amdgpu_amdkfd_add_eviction_fence are no longer needed.

Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu: Avoid setting off KFD eviction fences in amdgpu_vm
Felix Kuehling [Thu, 7 Feb 2019 17:08:14 +0000 (12:08 -0500)]
drm/amdgpu: Avoid setting off KFD eviction fences in amdgpu_vm

Use FENCE_OWNER_KFD to synchronize PT/PD initialization and clearing
of page table entries. This avoids triggering KFD eviction fences on
the PD reservation objects of compute VMs.

Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu: Replace ttm_bo_wait with amdgpu_bo_sync_wait
Felix Kuehling [Mon, 4 Feb 2019 22:53:05 +0000 (17:53 -0500)]
drm/amdgpu: Replace ttm_bo_wait with amdgpu_bo_sync_wait

The fence_owner logic in amdgpu_sync_wait will allow waiting without
having to temporarily remove eviction fences.

Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu: Add helper to wait for BO fences using a sync object
Felix Kuehling [Mon, 4 Feb 2019 22:46:35 +0000 (17:46 -0500)]
drm/amdgpu: Add helper to wait for BO fences using a sync object

Creates a temporary sync object to wait for the BO reservation. This
generalizes amdgpu_vm_wait_pd.

Signed-off-by: Felix Kuehling <Felix.Kuehling@amd.com>
Reviewed-by: Christian König <christian.koenig@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu: tighten gpu_recover in mailbox_flr to avoid duplicate recover in sriov
wentalou [Wed, 30 Jan 2019 03:13:01 +0000 (11:13 +0800)]
drm/amdgpu: tighten gpu_recover in mailbox_flr to avoid duplicate recover in sriov

sriov's gpu_recover inside xgpu_ai_mailbox_flr_work would cause duplicate recover in TDR.
TDR's gpu_recover would be triggered by amdgpu_job_timedout,
that could avoid vk-cts failure by unexpected recover.

Signed-off-by: Wentao Lou <Wentao.Lou@amd.com>
Acked-by: Andrey Grodzovsky <andrey.grodzovsky@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu: cleanup amdgpu_ih_process a bit more
Christian König [Wed, 9 Jan 2019 14:36:29 +0000 (15:36 +0100)]
drm/amdgpu: cleanup amdgpu_ih_process a bit more

Remove the callback and call the dispatcher directly.

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>
5 years agodrm/virtio: do NOT reuse resource ids
Gerd Hoffmann [Fri, 8 Feb 2019 14:04:09 +0000 (15:04 +0100)]
drm/virtio: do NOT reuse resource ids

Bisected guest kernel changes crashing qemu.  Landed at
"6c1cd97bda drm/virtio: fix resource id handling".  Looked again, and
noticed we where not only leaking *some* ids, but *all* ids.  The old
code never ever called virtio_gpu_resource_id_put().

So, commit 6c1cd97bda effectively makes the linux kernel starting
re-using IDs after releasing them, and apparently virglrenderer can't
deal with that.  Oops.

This patch puts a temporary stopgap into place for the 5.0 release.

Signed-off-by: Gerd Hoffmann <kraxel@redhat.com>
Reviewed-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190208140409.15280-1-kraxel@redhat.com
5 years agoMerge branch 'drm-next-5.1' of git://people.freedesktop.org/~agd5f/linux into drm...
Dave Airlie [Mon, 11 Feb 2019 04:04:05 +0000 (14:04 +1000)]
Merge branch 'drm-next-5.1' of git://people.freedesktop.org/~agd5f/linux into drm-next

Updates for 5.1:
- GDS fixes
- Add AMDGPU_CHUNK_ID_SCHEDULED_DEPENDENCIES interface
- GPUVM fixes
- PCIE DPM switching fixes for vega20
- Vega10 uclk DPM regression fix
- DC Freesync fixes
- DC ABM fixes
- Various DC cleanups

Signed-off-by: Dave Airlie <airlied@redhat.com>
From: Alex Deucher <alexdeucher@gmail.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190208210214.27666-1-alexander.deucher@amd.com
5 years agoMerge tag 'drm-intel-next-2019-02-07' of git://anongit.freedesktop.org/drm/drm-intel...
Dave Airlie [Mon, 11 Feb 2019 03:41:53 +0000 (13:41 +1000)]
Merge tag 'drm-intel-next-2019-02-07' of git://anongit.freedesktop.org/drm/drm-intel into drm-next

UAPI Changes:

- Expose RPCS (SSEU) configuration to userspace for Ice Lake
in order to allow userspace to reconfigure the subslice config
per context basis. (Tvrtko, Lionel)

Driver Changes:

- Execbuf and preemption improvements including selftests (Chris)
- Rename HAS_GMCH_DISPLAY/HAS_GMCH (Rodrigo)
- Debugfs error handling fix for robustness (Greg)
- Improve reg_rw traces (Ville)
- Push clear_intel_crtc_state onto the heap (Chris)
- Watermark fixes for Ice Lake (Ville)
- Fix enable count array size and bounds checking (Tvrtko)
- MST Fixes (Lyude)
- Prevent race and handle error on I915_GEM_MMAP (Joonas)
- Initial rework for an full atomic gamma mode (Ville)

Signed-off-by: Dave Airlie <airlied@redhat.com>
From: Rodrigo Vivi <rodrigo.vivi@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190208165000.GA30314@intel.com
5 years agoMerge tag 'drm/tegra/for-5.1-rc1' of git://anongit.freedesktop.org/tegra/linux into...
Dave Airlie [Mon, 11 Feb 2019 03:32:38 +0000 (13:32 +1000)]
Merge tag 'drm/tegra/for-5.1-rc1' of git://anongit.freedesktop.org/tegra/linux into drm-next

drm/tegra: Changes for v5.1-rc1

This set of changes starts of with some refactoring of the CEC support
to make it reusable on Tegra210 and later. Following are a couple of
fixes for HDMI audio support (via HDA).

The bulk here is a set of preparatory patches working towards enabling
Tegra186 support for host1x and VIC. Additional patches will be needed
to fully enable this, but they're not quite ready yet.

To round things off, this also adds support for configuring the SOR
crossbar using device tree, and fixes a couple of job-related issues in
the host1x code.

Signed-off-by: Dave Airlie <airlied@redhat.com>
From: Thierry Reding <thierry.reding@gmail.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190208144721.25830-1-thierry.reding@gmail.com
5 years agoMerge tag 'du-next-20190208' of git://linuxtv.org/pinchartl/media into drm-next
Dave Airlie [Mon, 11 Feb 2019 03:17:00 +0000 (13:17 +1000)]
Merge tag 'du-next-20190208' of git://linuxtv.org/pinchartl/media into drm-next

Renesas display drivers changes for v5.1 (2nd part):

- R8A7744 LVDS support
- DPAD0 output support on D3/E3

Signed-off-by: Dave Airlie <airlied@redhat.com>
From: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190208003355.GG10386@pendragon.ideasonboard.com
5 years agoMerge tag 'exynos-drm-next-for-v5.1' of git://git.kernel.org/pub/scm/linux/kernel...
Dave Airlie [Fri, 8 Feb 2019 00:35:26 +0000 (10:35 +1000)]
Merge tag 'exynos-drm-next-for-v5.1' of git://git./linux/kernel/git/daeinki/drm-exynos into drm-next

- Add rotator support for s5pv210
  . With this patch series, s5pv210 SoC can use rotator module but
    only NV12 and XRGB8888 formats are supported.
- Modify e-mail address
  . It changes email address of scaler module author.

Signed-off-by: Dave Airlie <airlied@redhat.com>
From: Inki Dae <inki.dae@samsung.com>
Link: https://patchwork.freedesktop.org/patch/msgid/079a9586-9d85-7d38-2658-ce312b6d71e8@samsung.com
5 years agodrm: rcar-du: Turn LVDS clock output on/off for DPAD0 output on D3/E3
Laurent Pinchart [Wed, 16 Jan 2019 23:59:34 +0000 (01:59 +0200)]
drm: rcar-du: Turn LVDS clock output on/off for DPAD0 output on D3/E3

On the D3 and E3 SoCs the LVDS PLL clock output provides the dot clock
to the DU channels, even when the LVDS outputs are not in use. Enable
and disable the LVDS clock output when enabling or disabling a CRTC
connected to the DPAD0 output.

Signed-off-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
5 years agodrm: rcar-du: lvds: Add API to enable/disable clock output
Laurent Pinchart [Wed, 16 Jan 2019 23:11:37 +0000 (01:11 +0200)]
drm: rcar-du: lvds: Add API to enable/disable clock output

On the D3 and E3 platforms, the LVDS internal PLL supplies the pixel
clock to the DU. This works automatically for LVDS outputs as the LVDS
encoder is enabled through the bridge API, enabling the internal PLL and
clock output. However, when using the DU DPAD output with the LVDS
outputs turned off, the LVDS PLL needs to be controlled manually. Add an
API to do so, to be called by the DU driver.

The drivers/gpu/drm/rcar-du/ directory has to be treated as obj-y
unconditionally, as the LVDS driver could be built-in while the DU
driver is compiled as a module.

Signed-off-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
5 years agodrm: rcar-du: lvds: Don't fail probe if output is not connected on D3/E3
Laurent Pinchart [Wed, 16 Jan 2019 22:43:08 +0000 (00:43 +0200)]
drm: rcar-du: lvds: Don't fail probe if output is not connected on D3/E3

On the D3 and E3 SoCs the LVDS encoder has an extended internal PLL and
supplies a clock to the DU. That clock is used not only for the LVDS
outputs but also for the DPAD output. The LVDS encoder thus needs to be
available to the DU even when its output is disabled. Don't fail probe
in that case on D3 and E3.

Signed-off-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
5 years agodrm: rcar-du: Simplify encoder registration
Laurent Pinchart [Wed, 16 Jan 2019 22:40:49 +0000 (00:40 +0200)]
drm: rcar-du: Simplify encoder registration

Before the driver fully moved to drm_bridge and drm_panel, it was
necessary to parse DT and locate encoder and connector nodes. The
connector node is now unused and can be removed as a parameter to
rcar_du_encoder_init(). As a consequence rcar_du_encoders_init_one() can
be greatly simplified, removing most of the DT parsing.

Signed-off-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
5 years agodrm: rcar-du: lvds: Add r8a7744 support
Biju Das [Tue, 22 Jan 2019 15:25:47 +0000 (15:25 +0000)]
drm: rcar-du: lvds: Add r8a7744 support

The LVDS encoders on RZ/G1N SoC is similar to RZ/G1M. Add support for
RZ/G1N (R8A7744) SoC to the LVDS encoder driver.

Signed-off-by: Biju Das <biju.das@bp.renesas.com>
Reviewed-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Signed-off-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Reviewed-by: Simon Horman <horms+renesas@verge.net.au>
5 years agodt-bindings: display: renesas: lvds: Document r8a7744 bindings
Biju Das [Tue, 22 Jan 2019 15:25:46 +0000 (15:25 +0000)]
dt-bindings: display: renesas: lvds: Document r8a7744 bindings

Document the RZ/G1N (R8A7744) LVDS bindings.

Signed-off-by: Biju Das <biju.das@bp.renesas.com>
Reviewed-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Signed-off-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Reviewed-by: Simon Horman <horms+renesas@verge.net.au>
Reviewed-by: Rob Herring <robh@kernel.org>
5 years agodrm: rcar-du: add missing of_node_put
Julia Lawall [Mon, 14 Jan 2019 16:44:56 +0000 (17:44 +0100)]
drm: rcar-du: add missing of_node_put

Add an of_node_put when the result of of_graph_get_remote_port_parent is
not available.

Add a second of_node_put if no encoder is selected (encoder remains NULL).

The semantic match that finds the first problem is as follows
(http://coccinelle.lip6.fr):

// <smpl>
@r exists@
local idexpression e;
expression x;
@@
e = of_graph_get_remote_port_parent(...);
... when != x = e
    when != true e == NULL
    when != of_node_put(e)
    when != of_fwnode_handle(e)
(
return e;
|
*return ...;
)
// </smpl>

Signed-off-by: Julia Lawall <Julia.Lawall@lip6.fr>
Reviewed-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
Reviewed-by: Kieran Bingham <kieran.bingham+renesas@ideasonboard.com>
Signed-off-by: Laurent Pinchart <laurent.pinchart+renesas@ideasonboard.com>
5 years agodrm/amd/display: Check hpd_gpio for NULL before accessing it
Harry Wentland [Thu, 7 Feb 2019 20:12:35 +0000 (15:12 -0500)]
drm/amd/display: Check hpd_gpio for NULL before accessing it

dal_gpio_open and dal_gpio_unlock_pin dereference hpd_gpio.
Check for NULL before calling those functions.

Fixes: ac627caf6b9275a ("drm/amd/display: add gpio lock/unlock")
Reported-by: Przemek Socha <soprwa@gmail.com>
CC: Chiawen Huang <chiawen.huang@amd.com>
CC: Tony Cheng <Tony.Cheng@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Tested-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Harry Wentland <harry.wentland@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/i915: Update DRIVER_DATE to 20190207
Rodrigo Vivi [Thu, 7 Feb 2019 20:45:32 +0000 (12:45 -0800)]
drm/i915: Update DRIVER_DATE to 20190207

Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
5 years agodrm/i915: Move LUT programming to happen after vblank waits
Ville Syrjälä [Tue, 5 Feb 2019 16:08:41 +0000 (18:08 +0200)]
drm/i915: Move LUT programming to happen after vblank waits

The LUTs are single buffered so we should program them after
the double buffered pipe updates have been latched by the
hardware.

We'll also fix up the IPS vs. split gamma w/a to do the IPS
disable like everyone else. Note that this is currently dead
code as we don't use the split gamma mode on HSW, but that
will be fixed up shortly.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190205160848.24662-7-ville.syrjala@linux.intel.com
Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
5 years agodrm/i915: Split color mgmt based on single vs. double buffered registers
Ville Syrjälä [Tue, 5 Feb 2019 16:08:40 +0000 (18:08 +0200)]
drm/i915: Split color mgmt based on single vs. double buffered registers

Split the color management hooks along the single vs. double
buffered registers line. Of the currently programmed registers
GAMMA_MODE and the ilk+ pipe CSC are double buffered, the
LUTS and CHV CGM block are single buffered.

The double buffered register will be programmed during the
normal pipe update with evasion, and also during pipe enable
so that the settings will already be correct when the pipe
starts up before the planes are enabled.

The single buffered registers are currently programmed before
the vblank evade. Which is totally wrong, but we'll correct
that later.

v2: Add some docs to explain the two vfuncs (Matt,Uma)
    Rebase

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190205160848.24662-6-ville.syrjala@linux.intel.com
Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
5 years agodrm/i915: Pull GAMMA_MODE write out from haswell_load_luts()
Ville Syrjälä [Tue, 5 Feb 2019 16:08:39 +0000 (18:08 +0200)]
drm/i915: Pull GAMMA_MODE write out from haswell_load_luts()

For bdw+ let's move the GAMMA_MODE write for the legacy LUT
mode into the .load_luts() funciton directly, rather than
relying on haswell_load_luts(). We'll be getting rid of
haswell_load_luts() entirely soon, and it's anyway cleaner
to have the GAMMA_MODE write in a single place.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Reviewed-by: Uma Shankar <uma.shankar@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190205160848.24662-5-ville.syrjala@linux.intel.com
Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
5 years agodrm/i915: Constify the state arguments to the color management stuff
Ville Syrjälä [Tue, 5 Feb 2019 16:08:38 +0000 (18:08 +0200)]
drm/i915: Constify the state arguments to the color management stuff

Pass the crtc state etc. as const to the color management commit
functions. And while at it polish some of the local variables.

v2: Rebase

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Reviewed-by: Uma Shankar <uma.shankar@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190205160848.24662-4-ville.syrjala@linux.intel.com
Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
5 years agodrm/i915: Precompute gamma_mode
Ville Syrjälä [Tue, 5 Feb 2019 16:08:37 +0000 (18:08 +0200)]
drm/i915: Precompute gamma_mode

We shouldn't be computing gamma mode during the commit phase.
Move it to the check phase.

v2: Reword comments a bit (Matt)
    Rebase

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Matt Roper <matthew.d.roper@intel.com>
Reviewed-by: Uma Shankar <uma.shankar@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190205160848.24662-3-ville.syrjala@linux.intel.com
Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
5 years agodrm/i915: Split the gamma/csc enable bits from the plane_ctl() function
Ville Syrjälä [Tue, 5 Feb 2019 16:08:36 +0000 (18:08 +0200)]
drm/i915: Split the gamma/csc enable bits from the plane_ctl() function

On g4x+ the pipe gamma enable bit for the primary plane affects
the pipe bottom color as well. The same for the pipe csc enable
bit on ilk+. Thus we must configure those bits correctly even
when the primary plane is disabled.

To make the feasible let's split those settings from the
plane_ctl() function into a seprate funciton that we can
call from the ->disable_plane() hook as well.

For consistency we'll do that on all the plane types. While
that has no real benefits at this time, it'll become useful
when we start to control the pipe gamma/csc enable bits
dynamically when we overhaul the color management code.

On pre-g4x there doesn't appear to be any way to gamma
correct the pipe bottom color, but sticking to the same
pattern doesn't hurt. And it'll still help us to do
crtc state readout correctly for the pipe gamma enable
bit for the color management overhaul.

An alternative apporach would be to still precompute these
bits into plane_state->ctl, but that would require that we
run through the plane check even when the plane isn't logically
enabled on any crtc. Currently that condition causes us to
short circuit the entire thing and not call ->check_plane().
There would also be some chicken and egg problems with
->check_plane() vs. crtc color state check that would
requite splitting certain things into multiple steps.
So all in all this seems like the easier route.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190205160848.24662-2-ville.syrjala@linux.intel.com
Reviewed-by: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
5 years agodrm/amdgpu: fix NULL ptr dref in the VM code
Christian König [Thu, 7 Feb 2019 10:41:59 +0000 (11:41 +0100)]
drm/amdgpu: fix NULL ptr dref in the VM code

The exclusive fence is of course perfectly optional here.

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>
5 years agodrm/amd/powerplay: add override pcie parameters for Vega20 (v2)
Harish Kasiviswanathan [Tue, 5 Feb 2019 19:05:11 +0000 (14:05 -0500)]
drm/amd/powerplay: add override pcie parameters for Vega20 (v2)

v2: Fix SMU message format
    Send override message after SMU enable features

Signed-off-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Reviewed-by: Eric Huang <JinhuiEric.Huang@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agodrm/amdgpu: Fix pci platform speed and width
Harish Kasiviswanathan [Fri, 1 Feb 2019 22:57:48 +0000 (17:57 -0500)]
drm/amdgpu: Fix pci platform speed and width

The new Vega series GPU cards have in-built bridges. To get the pcie
speed and width supported by the platform walk the hierarchy and get the
slowest link.

Signed-off-by: Harish Kasiviswanathan <Harish.Kasiviswanathan@amd.com>
Acked-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
5 years agogpu: host1x: Continue CDMA execution starting with a next job
Dmitry Osipenko [Tue, 7 Aug 2018 13:07:13 +0000 (16:07 +0300)]
gpu: host1x: Continue CDMA execution starting with a next job

Currently gathers of a hung job are getting NOP'ed and a restarted CDMA
executes the NOP'ed gathers. There shouldn't be a reason to not restart
CDMA execution starting with a next job, avoiding the unnecessary churning
with gathers NOP'ing.

Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Reviewed-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
5 years agogpu: host1x: Don't complete a completed job
Dmitry Osipenko [Tue, 7 Aug 2018 13:07:12 +0000 (16:07 +0300)]
gpu: host1x: Don't complete a completed job

There is a chance that the last job has been completed at the time of
CDMA timeout handler invocation. In this case there is no need to complete
the completed job.

Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Reviewed-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
5 years agogpu: host1x: Cancel only job that actually got stuck
Dmitry Osipenko [Tue, 7 Aug 2018 13:07:11 +0000 (16:07 +0300)]
gpu: host1x: Cancel only job that actually got stuck

Host1x doesn't have information about jobs inter-dependency, that is
something that will become available once host1x will get a proper
jobs scheduler implementation. Currently a hang job causes other unrelated
jobs to be canceled, that is a relic from downstream driver which is
irrelevant to upstream. Let's cancel only the hanging job and not to touch
other jobs in queue.

Signed-off-by: Dmitry Osipenko <digetx@gmail.com>
Reviewed-by: Mikko Perttunen <mperttunen@nvidia.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
5 years agodrm/tegra: sor: Support device tree crossbar configuration
Thierry Reding [Fri, 25 Jan 2019 10:00:58 +0000 (11:00 +0100)]
drm/tegra: sor: Support device tree crossbar configuration

The crossbar configuration is usually the same across all designs for a
given SoC generation. But sometimes there are designs that require some
other configuration.

Implement support for parsing the crossbar configuration from a device
tree. If the crossbar configuration is not present in the device tree,
fall back to the default crossbar configuration.

Signed-off-by: Thierry Reding <treding@nvidia.com>
5 years agodt-bindings: display: tegra: Support SOR crossbar configuration
Thierry Reding [Fri, 25 Jan 2019 10:00:57 +0000 (11:00 +0100)]
dt-bindings: display: tegra: Support SOR crossbar configuration

The SOR has a crossbar that can map each lane of the SOR to each of the
SOR pads. The mapping is usually the same across designs for a specific
SoC generation, but every now and then there's a design that doesn't.

Allow the crossbar configuration to be specified in device tree to make
it possible to support these designs.

Signed-off-by: Thierry Reding <treding@nvidia.com>
5 years agodrm/tegra: vic: Support stream ID register programming
Thierry Reding [Fri, 1 Feb 2019 13:28:36 +0000 (14:28 +0100)]
drm/tegra: vic: Support stream ID register programming

The version of VIC found in Tegra186 and later incorporates improvements
with regards to context isolation. As part of those improvements, stream
ID registers were added that allow to specify separate stream IDs for
the Falcon microcontroller and the VIC memory interface.

While it is possible to also set the stream ID dynamically at runtime to
allow userspace contexts to be completely separated, this commit doesn't
implement that yet. Instead, the static VIC stream ID is programmed when
the Falcon is booted. This ensures that memory accesses by the Falcon or
the VIC are properly translated via the SMMU.

Signed-off-by: Thierry Reding <treding@nvidia.com>
5 years agodrm/tegra: vic: Do not clear driver data
Thierry Reding [Fri, 1 Feb 2019 13:28:35 +0000 (14:28 +0100)]
drm/tegra: vic: Do not clear driver data

Upon driver failure, the driver core will take care of clearing the
driver data, so there's no need to do so explicitly in the driver.

Reviewed-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
5 years agodrm/tegra: Restrict IOVA space to DMA mask
Thierry Reding [Fri, 1 Feb 2019 13:28:34 +0000 (14:28 +0100)]
drm/tegra: Restrict IOVA space to DMA mask

On Tegra186 and later, the ARM SMMU provides an input address space that
is 48 bits wide. However, memory clients can only address up to 40 bits.
If the geometry is used as-is, allocations of IOVA space can end up in a
region that cannot be addressed by the memory clients.

To fix this, restrict the IOVA space to the DMA mask of the host1x
device. Note that, technically, the IOVA space needs to be restricted to
the intersection of the DMA masks for all clients that are attached to
the IOMMU domain. In practice using the DMA mask of the host1x device is
sufficient because all host1x clients share the same DMA mask.

Signed-off-by: Thierry Reding <treding@nvidia.com>
5 years agodrm/tegra: Setup shared IOMMU domain after initialization
Thierry Reding [Fri, 1 Feb 2019 13:28:33 +0000 (14:28 +0100)]
drm/tegra: Setup shared IOMMU domain after initialization

Move initialization of the shared IOMMU domain after the host1x device
has been initialized. At this point all the Tegra DRM clients have been
attached to the shared IOMMU domain.

This is important because Tegra186 and later use an ARM SMMU, for which
the driver defers setting up the geometry for a domain until a device is
attached to it. This is to ensure that the domain is properly set up for
a specific ARM SMMU instance, which is unknown at allocation time.

Reviewed-by: Dmitry Osipenko <digetx@gmail.com>
Signed-off-by: Thierry Reding <treding@nvidia.com>
5 years agodrm/tegra: vic: Load firmware on demand
Thierry Reding [Fri, 1 Feb 2019 13:28:32 +0000 (14:28 +0100)]
drm/tegra: vic: Load firmware on demand

Loading the firmware requires an allocation of IOVA space to make sure
that the VIC's Falcon microcontroller can read the firmware if address
translation via the SMMU is enabled.

However, the allocation currently happens at a time where the geometry
of an IOMMU domain may not have been initialized yet. This happens for
example on Tegra186 and later where an ARM SMMU is used. Domains which
are created by the ARM SMMU driver postpone the geometry setup until a
device is attached to the domain. This is because IOMMU domains aren't
attached to a specific IOMMU instance at allocation time and hence the
input address space, which defines the geometry, is not known yet.

Work around this by postponing the firmware load until it is needed at
the time where a channel is opened to the VIC. At this time the shared
IOMMU domain's geometry has been properly initialized.

As a byproduct this allows the Tegra DRM to be created in the absence
of VIC firmware, since the VIC initialization no longer fails if the
firmware can't be found.

Based on an earlier patch by Dmitry Osipenko <digetx@gmail.com>.

Signed-off-by: Thierry Reding <treding@nvidia.com>
Reviewed-by: Dmitry Osipenko <digetx@gmail.com>
5 years agodrm/tegra: Store parent pointer in Tegra DRM clients
Thierry Reding [Fri, 1 Feb 2019 13:28:31 +0000 (14:28 +0100)]
drm/tegra: Store parent pointer in Tegra DRM clients

Tegra DRM clients need access to their parent, so store a pointer to it
upon registration. It's technically possible to get at this by going via
the host1x client's parent and getting the driver data, but that's quite
complicated and not very transparent. It's much more straightforward and
natural to let the children know about their parent.

Signed-off-by: Thierry Reding <treding@nvidia.com>
Reviewed-by: Dmitry Osipenko <digetx@gmail.com>
5 years agogpu: host1x: Optimize CDMA push buffer memory usage
Thierry Reding [Fri, 1 Feb 2019 13:28:30 +0000 (14:28 +0100)]
gpu: host1x: Optimize CDMA push buffer memory usage

The host1x CDMA push buffer is terminated by a special opcode (RESTART)
that tells the CDMA to wrap around to the beginning of the push buffer.
To accomodate the RESTART opcode, an extra 4 bytes are allocated on top
of the 512 * 8 = 4096 bytes needed for the 512 slots (1 slot = 2 words)
that are used for other commands passed to CDMA. This requires that two
memory pages are allocated, but most of the second page (4092 bytes) is
never used.

Decrease the number of slots to 511 so that the RESTART opcode fits
within the page. Adjust the push buffer wraparound code to take into
account push buffer sizes that are not a power of two.

Signed-off-by: Thierry Reding <treding@nvidia.com>
5 years agogpu: host1x: Use correct semantics for HOST1X_CHANNEL_DMAEND
Thierry Reding [Fri, 1 Feb 2019 13:28:29 +0000 (14:28 +0100)]
gpu: host1x: Use correct semantics for HOST1X_CHANNEL_DMAEND

The HOST1X_CHANNEL_DMAEND is an offset relative to the value written to
the HOST1X_CHANNEL_DMASTART register, but it is currently treated as an
absolute address. This can cause SMMU faults if the CDMA fetches past a
pushbuffer's IOMMU mapping.

Properly setting the DMAEND prevents the CDMA from fetching beyond that
address and avoid such issues. This is currently not observed because a
whole (almost) page of essentially scratch space absorbs any excessive
prefetching by CDMA. However, changing the number of slots in the push
buffer can trigger these SMMU faults.

Signed-off-by: Thierry Reding <treding@nvidia.com>
5 years agogpu: host1x: Support 40-bit addressing on Tegra186
Thierry Reding [Fri, 1 Feb 2019 13:28:28 +0000 (14:28 +0100)]
gpu: host1x: Support 40-bit addressing on Tegra186

The host1x and clients instantiated on Tegra186 support addressing 40
bits of memory.

Signed-off-by: Thierry Reding <treding@nvidia.com>
5 years agogpu: host1x: Restrict IOVA space to DMA mask
Thierry Reding [Fri, 1 Feb 2019 13:28:27 +0000 (14:28 +0100)]
gpu: host1x: Restrict IOVA space to DMA mask

On Tegra186 and later, the ARM SMMU provides an input address space that
is 48 bits wide. However, memory clients can only address up to 40 bits.
If the geometry is used as-is, allocations of IOVA space can end up in a
region that is not addressable by the memory clients.

To fix this, restrict the IOVA space to the DMA mask of the host1x
device.

Signed-off-by: Thierry Reding <treding@nvidia.com>
5 years agogpu: host1x: Support 40-bit addressing
Thierry Reding [Fri, 1 Feb 2019 13:28:25 +0000 (14:28 +0100)]
gpu: host1x: Support 40-bit addressing

Tegra186 and later support 40 bits of address space. Additional
registers need to be programmed to store the full 40 bits of push
buffer addresses.

Since command stream gathers can also reside in buffers in a 40-bit
address space, a new variant of the GATHER opcode is also introduced.
It takes two parameters: the first parameter contains the lower 32
bits of the address and the second parameter contains bits 32 to 39.

Signed-off-by: Thierry Reding <treding@nvidia.com>
5 years agogpu: host1x: Introduce support for wide opcodes
Thierry Reding [Fri, 1 Feb 2019 13:28:24 +0000 (14:28 +0100)]
gpu: host1x: Introduce support for wide opcodes

The CDMA push buffer can currently only handle opcodes that take a
single word parameter. However, the host1x implementation on Tegra186
and later supports opcodes that require multiple words as parameters.

Unfortunately the way the push buffer is structured, these wide opcodes
cannot simply be composed of two regular opcodes because that could
result in the wide opcode being split across the end of the push buffer
and the final RESTART opcode required to wrap the push buffer around
would break the wide opcode.

One way to fix this would be to remove the concept of slots to simplify
push buffer operations. However, that's not entirely trivial and should
be done in a separate patch. For now, simply use a different function
to push four-word opcodes into the push buffer. Technically only three
words are pushed, with the fourth word used as padding to preserve the
2-word alignment required by the slots abstraction. The fourth word is
always a NOP opcode.

Additional care must be taken when the end of the push buffer is
reached. If a four-word opcode doesn't fit into the push buffer without
being split by the boundary, NOP opcodes will be introduced and the new
wide opcode placed at the beginning of the push buffer.

Signed-off-by: Thierry Reding <treding@nvidia.com>
5 years agogpu: host1x: Program the channel stream ID
Thierry Reding [Fri, 1 Feb 2019 13:28:23 +0000 (14:28 +0100)]
gpu: host1x: Program the channel stream ID

When processing command streams, make sure the host1x's stream ID is
programmed for the channel so that addresses are properly translated
through the SMMU.

Signed-off-by: Thierry Reding <treding@nvidia.com>
5 years agodrm/i915: Don't set update_wm_post on g4x+
Ville Syrjälä [Wed, 6 Feb 2019 18:54:33 +0000 (20:54 +0200)]
drm/i915: Don't set update_wm_post on g4x+

update_wm_post is meant for pre-g4x only. Don't ever set
it on g4x+.

The only effect of a bogus update_wm_post on g4x+ could
be that we clear the legacy_cursor_update flag in
intel_atomic_commit(). Since legacy_cursor_update is
only set for legacy cursor updates (as the name suggests)
and we only set update_wm_post for a modeset the two
cases should never occur at the same time. But let's
be consistent in setting update_wm_post so we don't
end up confusing so many people.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190206185433.8116-1-ville.syrjala@linux.intel.com
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
5 years agodrm/i915: Hack and slash, throttle execbuffer hogs
Chris Wilson [Thu, 7 Feb 2019 07:18:22 +0000 (07:18 +0000)]
drm/i915: Hack and slash, throttle execbuffer hogs

Apply backpressure to hogs that emit requests faster than the GPU can
process them by waiting for their ring to be less than half-full before
proceeding with taking the struct_mutex.

This is a gross hack to apply throttling backpressure, the long term
goal is to remove the struct_mutex contention so that each client
naturally waits, preferably in an asynchronous, nonblocking fashion
(pipelined operations for the win), for their own resources and never
blocks another client within the driver at least. (Realtime priority
goals would extend to ensuring that resource contention favours high
priority clients as well.)

This patch only limits excessive request production and does not attempt
to throttle clients that block waiting for eviction (either global GTT or
system memory) or any other global resources, see above for the long term
goal.

No microbenchmarks are harmed (to the best of my knowledge).

Testcase: igt/gem_exec_schedule/pi-ringfull-*
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: John Harrison <John.C.Harrison@Intel.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190207071829.5574-1-chris@chris-wilson.co.uk
5 years agodrm/i915: Handle vm_mmap error during I915_GEM_MMAP ioctl with WC set
Joonas Lahtinen [Thu, 7 Feb 2019 08:54:54 +0000 (10:54 +0200)]
drm/i915: Handle vm_mmap error during I915_GEM_MMAP ioctl with WC set

Add err goto label and use it when VMA can't be established or changes
underneath.

v2:
- Dropping Fixes: as it's indeed impossible to race an object to the
  error address. (Chris)
v3:
- Use IS_ERR_VALUE (Chris)

Reported-by: Adam Zabrocki <adamza@microsoft.com>
Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Cc: Adam Zabrocki <adamza@microsoft.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> #v2
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20190207085454.10598-2-joonas.lahtinen@linux.intel.com
5 years agodrm/i915: Prevent a race during I915_GEM_MMAP ioctl with WC set
Joonas Lahtinen [Thu, 7 Feb 2019 08:54:53 +0000 (10:54 +0200)]
drm/i915: Prevent a race during I915_GEM_MMAP ioctl with WC set

Make sure the underlying VMA in the process address space is the
same as it was during vm_mmap to avoid applying WC to wrong VMA.

A more long-term solution would be to have vm_mmap_locked variant
in linux/mmap.h for when caller wants to hold mmap_sem for an
extended duration.

v2:
- Refactor the compare function

Fixes: 1816f9236303 ("drm/i915: Support creation of unbound wc user mappings for objects")
Reported-by: Adam Zabrocki <adamza@microsoft.com>
Suggested-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: <stable@vger.kernel.org> # v4.0+
Cc: Akash Goel <akash.goel@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Cc: Adam Zabrocki <adamza@microsoft.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com> #v1
Link: https://patchwork.freedesktop.org/patch/msgid/20190207085454.10598-1-joonas.lahtinen@linux.intel.com
5 years agodrm/exynos: Change Andrzej Pietrasiewicz's e-mail address
Andrzej Pietrasiewicz [Thu, 13 Dec 2018 13:19:46 +0000 (14:19 +0100)]
drm/exynos: Change Andrzej Pietrasiewicz's e-mail address

My @samusung.com address is going to cease existing soon, so change it to
an address which can actually be used to contact me.

Signed-off-by: Andrzej Pietrasiewicz <andrzej.p@samsung.com>
Signed-off-by: Inki Dae <inki.dae@samsung.com>
5 years agodt-bindings: gpu: samsung-rotator: Document s5pv210 support
Paweł Chmiel [Fri, 28 Dec 2018 15:19:03 +0000 (16:19 +0100)]
dt-bindings: gpu: samsung-rotator: Document s5pv210 support

This commit documents new compatible for s5pv210 soc,
which will be also supported by this driver.

Signed-off-by: Paweł Chmiel <pawel.mikolaj.chmiel@gmail.com>
Reviewed-by: Rob Herring <robh@kernel.org>
Acked-by: Krzysztof Kozlowski <krzk@kernel.org>
Signed-off-by: Inki Dae <inki.dae@samsung.com>
5 years agodrm/exynos: rotator: Add support for s5pv210
Paweł Chmiel [Wed, 19 Dec 2018 15:57:02 +0000 (16:57 +0100)]
drm/exynos: rotator: Add support for s5pv210

This commit adds support for s5pv210.
Currently only NV12 and XRGB8888 formats are supported.
It was tested by using tool from
https://www.spinics.net/lists/linux-samsung-soc/msg60498.html

Signed-off-by: Paweł Chmiel <pawel.mikolaj.chmiel@gmail.com>
Signed-off-by: Inki Dae <inki.dae@samsung.com>
5 years agodrm/i915: Don't send hotplug in intel_dp_check_mst_status()
Lyude Paul [Tue, 29 Jan 2019 19:10:01 +0000 (14:10 -0500)]
drm/i915: Don't send hotplug in intel_dp_check_mst_status()

This hotplug also isn't needed: drm_dp_mst_topology_mgr_set_mst()
already sends a hotplug on its own from drm_dp_destroy_connector_work()
after destroying connectors in the MST topology.

Signed-off-by: Lyude Paul <lyude@redhat.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Daniel Vetter <daniel@ffwll.ch>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190129191001.442-4-lyude@redhat.com
5 years agodrm/i915: Don't send MST hotplugs during resume
Lyude Paul [Tue, 29 Jan 2019 19:10:00 +0000 (14:10 -0500)]
drm/i915: Don't send MST hotplugs during resume

We have a bad habit of calling drm_fb_helper_hotplug_event() far more
then we actually need to. MST appears to be one of these cases, where we
call drm_fb_helper_hotplug_event() if we fail to resume a connected MST
topology in intel_dp_mst_resume(). We don't actually need to do this at
all though since hotplug events are already sent from
drm_dp_connector_destroy_work() every time connectors are unregistered
from userspace's PoV. Additionally, extra calls to
drm_fb_helper_hotplug_event() also just mean more of a chance of doing a
connector probe somewhere we shouldn't.

So, don't send any hotplug events during resume if the MST topology
fails to come up. Just rely on the DP MST helpers to send them for us.

Signed-off-by: Lyude Paul <lyude@redhat.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: Daniel Vetter <daniel@ffwll.ch>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190129191001.442-3-lyude@redhat.com
5 years agodrm/i915: Block fbdev HPD processing during suspend
Lyude Paul [Tue, 29 Jan 2019 19:09:59 +0000 (14:09 -0500)]
drm/i915: Block fbdev HPD processing during suspend

When resuming, we check whether or not any previously connected
MST topologies are still present and if so, attempt to resume them. If
this fails, we disable said MST topologies and fire off a hotplug event
so that userspace knows to reprobe.

However, sending a hotplug event involves calling
drm_fb_helper_hotplug_event(), which in turn results in fbcon doing a
connector reprobe in the caller's thread - something we can't do at the
point in which i915 calls drm_dp_mst_topology_mgr_resume() since
hotplugging hasn't been fully initialized yet.

This currently causes some rather subtle but fatal issues. For example,
on my T480s the laptop dock connected to it usually disappears during a
suspend cycle, and comes back up a short while after the system has been
resumed. This guarantees pretty much every suspend and resume cycle,
drm_dp_mst_topology_mgr_set_mst(mgr, false); will be caused and in turn,
a connector hotplug will occur. Now it's Rute Goldberg time: when the
connector hotplug occurs, i915 reprobes /all/ of the connectors,
including eDP. However, eDP probing requires that we power on the panel
VDD which in turn, grabs a wakeref to the appropriate power domain on
the GPU (on my T480s, this is the PORT_DDI_A_IO domain). This is where
things start breaking, since this all happens before
intel_power_domains_enable() is called we end up leaking the wakeref
that was acquired and never releasing it later. Come next suspend/resume
cycle, this causes us to fail to shut down the GPU properly, which
causes it not to resume properly and die a horrible complicated death.

(as a note: this only happens when there's both an eDP panel and MST
topology connected which is removed mid-suspend. One or the other seems
to always be OK).

We could try to fix the VDD wakeref leak, but this doesn't seem like
it's worth it at all since we aren't able to handle hotplug detection
while resuming anyway. So, let's go with a more robust solution inspired
by nouveau: block fbdev from handling hotplug events until we resume
fbdev. This allows us to still send sysfs hotplug events to be handled
later by user space while we're resuming, while also preventing us from
actually processing any hotplug events we receive until it's safe.

This fixes the wakeref leak observed on the T480s and as such, also
fixes suspend/resume with MST topologies connected on this machine.

Changes since v2:
* Don't call drm_fb_helper_hotplug_event() under lock, do it after lock
  (Chris Wilson)
* Don't call drm_fb_helper_hotplug_event() in
  intel_fbdev_output_poll_changed() under lock (Chris Wilson)
* Always set ifbdev->hpd_waiting (Chris Wilson)

Signed-off-by: Lyude Paul <lyude@redhat.com>
Fixes: 0e32b39ceed6 ("drm/i915: add DP 1.2 MST support (v0.7)")
Cc: Todd Previte <tprevite@gmail.com>
Cc: Dave Airlie <airlied@redhat.com>
Cc: Jani Nikula <jani.nikula@linux.intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Rodrigo Vivi <rodrigo.vivi@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Cc: intel-gfx@lists.freedesktop.org
Cc: <stable@vger.kernel.org> # v3.17+
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20190129191001.442-2-lyude@redhat.com