Don't block others while waiting for the fences to finish, concurrent
submission is perfectly valid in this case and holding the lock can
prevent killed applications from terminating.
Signed-off-by: Christian König <ckoenig.leichtzumerken@gmail.com>
Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
DC and DM already support DRM_FORMAT_RGB565, it's just missing from the
list of valid formats.
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Leo Li <sunpeng.li@amd.com>
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Acked-by: Alex Deucher <alexander.deucher@amd.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Add a fast_validate parameter in dc_validate_global_state for future use
Signed-off-by: Joshua Aberback <joshua.aberback@amd.com>
Reviewed-by: Tony Cheng <Tony.Cheng@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
As long as we have at least one non-cursor plane enabled on a CRTC then
the CRTC itself can remain enabled.
This will allow for commits where there's an overlay plane enabled but
no primary plane enabled.
[How]
Remove existing primary plane fb != NULL checks and replace them with
the new does_crtc_have_active_plane helper.
This will be called from atomic check when validating the CRTC.
Since the primary plane state can now potentially be NULL we'll need
to guard for that when accessing it in some of the cursor logic.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: David Francis <David.Francis@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
Surface scaling info updates can affect bandwidth and blocks. We need
to be checking these with global validation to avoid underflow or
corruption.
[How]
Drop the state->allow_modeset early exit in
dm_determine_update_type_for_commit. Most of those should be considered
fast now anyway.
Fill in scaling info and it to the surface update in atomic
check.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: David Francis <David.Francis@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
DM thinks that the update type should be full whenever a stream or
plane is added or removed (including recreations).
This won't match in the case where DC thinks what looks like a fast
update to DM is actually a medium or full - like scaling changes that
affect bandwidth and clocks.
[How]
Drop this warning. DC knows better than the DM does for determining
cases like this.
The other warning can be kept for now since it would warn on a pretty
serious DC or DM bug.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: David Francis <David.Francis@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
DC expects the surface memory address to identify the surface.
This doesn't work with what we're doing with the temporary surfaces,
it will always assume this is a full update because the surface
isn't in the current context.
[How]
Use the surface directly. This doesn't give us much improvement yet,
since we always create a new dc_plane_state when state->allow_modeset
is true.
The call into dc_check_update_surfaces_for_stream also needs to be
locked, for two reasons:
1. It checks the current DC state
2. It modifies the surface update flags
Both of which could be currently in the middle of commit work from
commit tail.
A TODO here is to pass the context explicitly into this function and
find a way to get the surface update flags out of it without modifying
the surface in place.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: David Francis <David.Francis@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
Planes have downscaling limits and upscaling limits per format and DM
is expected to validate these using DC caps. We should fail atomic
check validation if we aren't capable of doing the scaling.
[How]
We don't currently create store which DC plane maps to which DRM plane
so we can't easily check the caps directly. For now add basic
constraints that cover the absolute min and max downscale / upscale
limits for most RGB and YUV formats across ASICs.
Leave a TODO indicating that these should really be done with DC caps.
We'll probably need to subclass DRM planes again in order to correctly
identify which DC plane maps to it.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: Harry Wentland <Harry.Wentland@amd.com>
Reviewed-by: Tony Cheng <Tony.Cheng@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
We currently don't do DC validation for medium or full updates where
the plane state isn't created. There are some medium and full updates
that can cause bandwidth or clock changes to occur resulting in
underflow or corruption.
We need to be able to fill surface and plane info updates during
atomic commit for dm_determine_update_type for commit. Since we already
do this during atomic commit tail it would be good if we had the same
logic in both places for creating these structures.
[How]
Introduce fill_dc_scaling_info and fill_dc_plane_info_and_addr.
These two functions cover the following three update structures:
- struct dc_scaling_info
- struct dc_plane_info
- struct dc_plane_address
Cleanup and adapter the existing fill_plane_* helpers to work with
these functions.
Update call sites that used most of these sub helpers directly to work
with the new functions. The exception being prepare_fb - we just want
the new buffer attributes specifically in the case where we're
creating the plane. This is needed for dc_commit_state in the case
where the FB hasn't been previously been used.
This isn't quite a refactor, but functionally driver behavior should
be mostly the smae as before. The one exception is that we now check
the return code for fill_plane_buffer_attributes which means that
commits will be rejected that try to enable DCC with erroneous
parameters.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: David Francis <David.Francis@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
Pitch was only calculated based on format whenever the plane state
was recreated. This could result in surface corruption due to the
incorrect pitch being programmed when the surface pitch changed during
commits where state->allow_modeset = false.
[How]
Recalculate pitch at the same time we update the buffer address and
other buffer attributes. This function was previously called
fill_plane_tiling_attributes but I've also renamed it to
fill_plane_buffer_attributes to clarify the actual intent of the
function now that it's handling most buffer related attributes.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: David Francis <David.Francis@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
The overlay will be incorrectly placed *below* the primary plane for
commits with state->allow_modeset = true because the primary plane
won't be removed and recreated in the same commit.
[How]
Add the should_reset_plane helper to determine if the plane should be
reset or not. If we need to add or force reset any plane in the context
then we'll need to do the same for every plane on the stream.
Unfortunately we need to do the remove / recreate routine for removing
planes as well since DC currently isn't well equipped to handle the
plane with the top pipe being removed with other planes still active.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: Sun peng Li <Sunpeng.Li@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
Plane scaling parameters are not correctly filled or updated when
performing fast updates.
They're filled when creating the dc plane state and during atomic check.
While the atomic check code path happens for the plane even during fast
updates, the issue is that they're done in place on the dc_plane_state
directly. This dc_plane_state may be the current state plane state
being used by the hardware, so these parameters won't be correctly
programmed.
The new scaling parameters should instead be passed as an update
to the plane.
[How]
Update fill_rects_from_plane_state to not modify dc_plane_state
directly. Update the call sites that use this to fill in the appropriate
values.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: Sun peng Li <Sunpeng.Li@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[why]
there are some scaling capabilities such as fp16 which are known to be unsupported
on a given ASIC. exposing these static capabilities allows much simpler implementation
for OS interfaces which require to report such static capabilities to reduce the
number of dynamic validation calls
[how]
refactor the existing plane caps to be more extensible, and add fp16 and scaling
capabilities
Signed-off-by: Jun Lei <Jun.Lei@amd.com>
Reviewed-by: Tony Cheng <Tony.Cheng@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com>
Acked-by: Harry Wentland <Harry.Wentland@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
We need DC's color space to match the color encoding and color space
specified by userspace to correctly render YUV surfaces.
[How]
Add the DRM color properties when the DC plane supports NV12.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: Sun peng Li <Sunpeng.Li@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
Hardware can support video surfaces and DC tells us which planes are
suitable via DC plane caps.
[How]
The supported formats array will now vary based on what DC tells us,
so create an array and fill it dynamically based on plane types and
caps.
Ideally we'd query support for every format via DC plane caps, but for
the framework is in place to do so later with this.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: Sun peng Li <Sunpeng.Li@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
When deciding to add properties or expose formats on DRM planes we
should be querying the caps for the DC plane it's supposed to represent.
[How]
Pass plane caps down into plane initialization, refactoring overlay
plane initialization to have the overlay plane be represented by
the first overlay capable DC plane.
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>
[Why]
When entering S3, amdgpu first calls DRM to cache the current atomic
state, then commit the 'all-disabled' state. This sets dc->current_state
to point at the private atomic object's dm_atomic_state->context, as
any regular atomic commit would.
Afterwards, amdgpu_dm calls dc_set_power_state() with S3 power state.
This invalidates dc->current_state by wiping it to 0, consequently
wiping dm_atomic_state->context.
During resume, the cached atomic state is restored. When getting the
private object however, the dm_atomic_state - containing the wiped
context - is duplicated into the atomic state. This causes DC validation
to fail during atomic check, as necessary function pointers in dc_state
are now NULL.
[How]
Recreate the private object's dm_atomic_state->context during resume,
restoring any static values such as function pointers.
A TODO item is added to move static read-only values out of dc_state -
they shouldn't be there anyways.
Signed-off-by: Leo Li <sunpeng.li@amd.com>
Reviewed-by: Nicholas Kazlauskas <Nicholas.Kazlauskas@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
We need DC's color space to match the color encoding and color space
specified by userspace to correctly render YUV surfaces.
[How]
Convert the DRM color encoding and color range properties to the
appropriate DC colorspace option and update the color space when
performing surface updates.
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>
The brace initialization used here generates warnings on some
compilers. For example, on GCC 4.9:
[...] In function ‘dm_determine_update_type_for_commit’:
[...] error: missing braces around initializer [-Werror=missing-braces]
struct dc_stream_update stream_update = { 0 };
^
Use memset to make this more portable.
v2: Specify the compiler / diagnostic in the commit message (Paul)
Cc: Sun peng Li <Sunpeng.Li@amd.com>
Cc: Harry Wentland <Harry.Wentland@amd.com>
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>
[Why]
Multiple threads were writing back to one global VBA in DC resulting
in multiple threads overwriting eachother's data
[How]
Add an instance of DML (which contains VBA) to each context and
change all calls that used dc->dml to use context->dml. Created a
seperate copy constructor for linux in a case where there is no
access to DC.
Signed-off-by: Aidan Wood <Aidan.Wood@amd.com>
Reviewed-by: Jun Lei <Jun.Lei@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet.Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
We want vblank counts and timestamps of flip completion as sent
in pageflip completion events to be consistent with the vblank
count and timestamp of the vblank of flip completion, like in non
VRR mode.
In VRR mode, drm_update_vblank_count() - and thereby vblank
count and timestamp updates - must be delayed until after the
end of front-porch of each vblank, as it is only safe to
calculate vblank timestamps outside of the front-porch, when
we actually know when the vblank will end or has ended.
The function drm_update_vblank_count() which updates timestamps
and counts gets called by drm_crtc_accurate_vblank_count() or by
drm_crtc_handle_vblank().
Therefore we must make sure that pageflip events for a completed
flip are only sent out after drm_crtc_accurate_vblank_count() or
drm_crtc_handle_vblank() is executed, after end of front-porch
for the vblank of flip completion.
Two cases:
a) Pageflip irq handler executes inside front-porch:
In this case we must defer sending pageflip events until
drm_crtc_handle_vblank() executes after end of front-porch,
and thereby calculates proper vblank count and timestamp.
Iow. the pflip irq handler must just arm a pageflip event
to be sent out by drm_crtc_handle_vblank() later on.
b) Pageflip irq handler executes after end of front-porch, e.g.,
after flip completion in back-porch or due to a massively
delayed handler invocation into the active scanout of the new
frame. In this case we can call drm_crtc_accurate_vblank_count()
to safely force calculation of a proper vblank count and
timestamp, and must send the pageflip completion event
ourselves from the pageflip irq handler.
This is the same behaviour as needed for standard fixed refresh
rate mode.
To decide from within pageflip handler if we are in case a) or b),
we check the current scanout position against the boundary of
front-porch. In non-VRR mode we just do what we did in the past.
Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com>
Reviewed-by: Nicholas Kazlauskas <Nicholas.Kazlauskas@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
In VRR mode, proper vblank/pageflip timestamps can only be computed
after the display scanout position has left front-porch. Therefore
delay calls to drm_crtc_handle_vblank(), and thereby calls to
drm_update_vblank_count() and pageflip event delivery, to after the
end of front-porch when in VRR mode.
We add a new vupdate irq, which triggers at the end of the vupdate
interval, ie. at the end of vblank, and calls the core vblank handler
function. The new irq handler is not executed in standard non-VRR
mode, so vblank handling for fixed refresh rate mode is identical
to the past implementation.
v2: Implement feedback by Nicholas and Paul Menzel.
Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com>
Acked-by: Harry Wentland <harry.wentland@amd.com>
Reviewed-by: Nicholas Kazlauskas <Nicholas.Kazlauskas@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
For throttling to work correctly, we always need a baseline vblank
count last_flip_vblank that increments at start of front-porch.
This is the case for drm_crtc_vblank_count() in non-VRR mode, where
the vblank irq fires at start of front-porch and triggers DRM core
vblank handling, but it is no longer the case in VRR mode, where
core vblank handling is done later, after end of front-porch.
Therefore drm_crtc_vblank_count() is no longer useful for this.
We also can't use drm_crtc_accurate_vblank_count(), as that would
screw up vblank timestamps in VRR mode when called in front-porch.
To solve this, use the cooked hardware vblank counter returned by
amdgpu_get_vblank_counter_kms() instead, as that one is cooked to
always increment at start of front-porch, independent of when
vblank related irq's fire.
This patch allows vblank irq handling to happen anywhere within
vblank of even after it, without a negative impact on flip
throttling, so followup patches can shift the vblank core
handling trigger point wherever they need it.
Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com>
Reviewed-by: Nicholas Kazlauskas <Nicholas.Kazlauskas@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
During VRR mode we can not allow vblank irq dis-/enable
transitions, as an enable after a disable can happen at
an arbitrary time during the video refresh cycle, e.g.,
with a high likelyhood inside vblank front-porch. An
enable during front-porch would cause vblank timestamp
updates/calculations which are completely bogus, given
the code can't know when the vblank will end as long
as we are in front-porch with no page flip completed.
Hold a permanent vblank reference on the crtc while
in active VRR mode to prevent a vblank disable, and
drop the reference again when switching back to fixed
refresh rate non-VRR mode.
v2: Make sure transition is also handled if vrr is
disabled and stream gets disabled in the same
atomic commit by moving the call to the transition
function outside of plane commit.
Suggested by Nicholas.
v3: Trivial rebase onto previous patch.
Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com>
Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
We need the VRR active/inactive state info earlier in
the commit sequence, so VRR related setup functions like
amdgpu_dm_handle_vrr_transition() know the final VRR state
when they need to do their hw setup work.
Split update_freesync_state_on_stream() into an early part,
that can run at the beginning of commit tail before the
vrr transition handling, and a late part that must run after
vrr transition handling inside the commit planes code for
enabled crtc's.
Suggested by Nicholas Kazlauskas.
Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com>
Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
amdgpu:
- Switch to HMM for userptr (reverted until HMM fixes land)
- New experimental SMU 11 replacement for powerplay for vega20 (not enabled by default)
- Initial RAS support for vega20
- BACO support for vega12
- BACO fixes for vega20
- Rework IH handling for page fault and retry interrupts
- Cleanly split CPU and GPU paths for GPUVM updates
- Powerplay fixes
- XGMI fixes
- Rework how DC interacts with atomic for planes
- Clean up and simplify DC/Powerplay interfaces
- Misc cleanups and bug fixes
amdkfd:
- Switch to HMM for userptr (reverted until HMM fixes land)
- Add initial RAS support
- MQD fixes
ttm:
- Unify DRM_FILE_PAGE_OFFSET handling
- Account for kernel allocations in kernel zone only
- Misc cleanups
Signed-off-by: Dave Airlie <airlied@redhat.com>
From: Alex Deucher <alexdeucher@gmail.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190402170820.22197-1-alexander.deucher@amd.com
We want DRM planes to be initialized in the following order:
- primary planes
- overlay planes
- cursor planes
to support existing userspace expectations for plane z-ordering. This
means that we also need to register CRTCs after all planes have been
initialized since overlay planes can be placed on any CRTC.
So the only reason why we have the mode_info->planes list is to
remember the primary planes for use later when we need to register
the CRTC.
Overlay planes have no purpose being in this list. DRM will cleanup
any planes that we've registered for us, so the only planes that need to
be explicitly cleaned up are the ones that have failed to register.
By dropping the explicit free on every plane in the mode_info->planes
list this patch also fixes a double-free in the case where we fail to
initialize only some of the planes.
Cc: Leo Li <sunpeng.li@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
We need arbitrary read/write over DP AUX DPCD
for debugging
[How]
Three debugfs entries
Set the target address by writing to
"aux_dpcd_address"
(The first four bytes written are used)
Set the transaction size in bytes by writing to
"aux_dpcd_size"
(The first four bytes written are used)
Start a transaction by reading/writing
"aux_dpcd_data"
Do note: there is no concerrency protection at all
Accessing these entries in quick succession can lead
to strange behaviour
Signed-off-by: David Francis <David.Francis@amd.com>
Reviewed-by: Harry Wentland <Harry.Wentland@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet.Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
Somewhere in the atomic check reshuffle ABM got lost.
ABM is a crtc property (copied from a connector property).
It can change without a modeset, just like underscan.
[How]
In the skip_modeset branch of atomic check crtc updates,
copy over the abm property.
Signed-off-by: David Francis <David.Francis@amd.com>
Reviewed-by: Nicholas Kazlauskas <Nicholas.Kazlauskas@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet.Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
To help prevent plane state not being set to the correct default
value if any new properties are added in the future.
[How]
Use the drm helper - which seems to be the common solution among other
DRM drivers.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: Sun peng Li <Sunpeng.Li@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet.Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
The plane_reset callback is subclassed but hasn't been updated since
the drm helper got updated to include resetting alpha related state
(state->alpha and state->pixel_blend_mode). The overlay planes
exposed by amdgpu_dm were therefore being rendered as invisible by
default ever since supported was exposed for alpha blending properties
on overlays.
This caused regressions in igt@kms_plane_multiple@atomic-tiling-none
and igt@kms_plane@plane-position-covered-pipe tests.
[How]
Reset the plane state values to their correct values as defined in
the drm helper.
This fixes the IGT test regression.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: Harry Wentland <Harry.Wentland@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet.Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
Incorrect hardcoded assumptions are made regarding luma and chroma
alignment. The actual values set for the DRM framebuffer should be used
when programming the address.
[How]
Respect the given pitch for both luma and chroma planes - it's not like
we can force the alignment to anything else at this point anyway.
Use the FB offset for the chroma planes directly. DRM already
provides this to us so there's no need to calculate it manually.
While we don't actually use the chroma surface size parameters on Raven,
these should have technically been fb->width / 2 and fb->height / 2
since the chroma plane is half size of the luma plane for NV12.
Leave a TODO indicating that those should be set based on the actual
surface format instead since this is only correct for YUV420 formats.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: Harry Wentland <harry.wentland@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
For new DC planes the correct plane address fields are filled based
on whether the plane had a graphics or video format.
However, when we perform stream and plane updates using DC we only ever
fill in the graphics format fields. This causing corruption and hangs
when using video surface formats like NV12 for planes.
[How]
Use the same logic everywhere we update dc_plane_address - always
fill in the correct fields based on the surface format type.
There are 3 places this is done:
- Atomic check, during DC plane creation
- Atomic commit, during plane prepare_fb
- Atomic commit tail, during amdgpu_dm_commit_planes
We use the fill_plane_tiling_attributes in all 3 locations and it
already needs the address to update DCC attributes, so the surface
address update logic can be moved into this helper.
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>
[Why]
The DRM overlay planes DM exposes support RGBA formats but are currently
forced as fully opaque over whatever they overlay.
[How]
Expose DRM blending mode and alpha properties to userspace.
The overlays exposed support per-pixel pre-multiplied alpha along with
global plane opacity.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: Harry Wentland <Harry.Wentland@amd.com>
Acked-by: Leo Li <sunpeng.li@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
Can happen on ASICs with 6 planes, but this isn't a bug since we haven't
written outside the array.
[How]
Use <= instead of <.
Cc: Leo Li <sunpeng.li@amd.com>
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reported-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Michel Dänzer <michel.daenzer@amd.com>
Tested-by: Michel Dänzer <michel.daenzer@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
New DRM versions manage locking for private objects for us, so this
is no longer needed.
This also prevents a WARN_ON from occurring when the private object is
duplicated during the forced atomic commit that occurs from the HPD
handler.
The HPD handler calls drm_modeset_lock_all before the forced commit
and if the private object is duplicated then the
DEBUG_LOCKS_WARN_ON(ww_ctx->done_acquire) warning will be triggered
since we're trying to lock something when everything should have
already been locked.
[How]
Drop the lock and let DRM manage this.
Cc: Harry Wentland <harry.wentland@amd.com>
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>
[Why]
Tiling and DCC attributes can change when swapping framebuffers but
these will only ever get updated on full commits where
state->allow_modeset is true. But for the page-flip IOCTL
state->allow_modeset = false so these aren't updated and DCC changes
aren't being programmed.
[How]
Split out updating the tiling and DCC attributes into its own function
that's called when creating the plane for the first time but also
called when switching framebuffers before commit stream updates.
These functions also don't modify the plane state directly since the
plane state being used as reference is often the current one
being used by the hardware.
DC needs to be able to compare plane_state to the stream update
structures to determine if the hardware programming should happen. If
it's modified in place then nothing will be programmed and we can run
into concurrency issues.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: Harry Wentland <Harry.Wentland@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet.Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
commit_planes is indented quite far
[How]
Move the pageflip code from an if statement to after a
continue
Signed-off-by: David Francis <David.Francis@amd.com>
Reviewed-by: Nicholas Kazlauskas <Nicholas.Kazlauskas@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet.Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
IGT expects that pageflips can be triggered with the same
framebuffer before and after the commit
[How]
Expand the definition of pageflip to include any change
with an old framebuffer and a new framebuffer, even if they're
the same
Signed-off-by: David Francis <David.Francis@amd.com>
Reviewed-by: Nicholas Kazlauskas <Nicholas.Kazlauskas@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet.Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
Raven has support for combining pipes for DRM_PLANE_TYPE_OVERLAY use
but no overlays are exposed to userspace.
[How]
Expose overlay planes based on DC plane caps.
If all the pipes are in use then the atomic commits can fail, but this
is expected behavior for userspace.
Only support RGB on overlays for now.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: Tony Cheng <Tony.Cheng@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet.Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
Primary and underlay planes were previously exposed to DRM by using
max_planes and max_slave_planes.
The value for max_planes was always pipe_count + has_underlay.
If there was an underlay pipe, then max_slave_planes = 1.
Raven has pipe_count = 4, max_planes = 4, and max_slave_planes = 1.
So during plane initialziation it was actually "creating"
1 overlay plane and 3 primary planes... or it would be, had its
plane_type array not been dm_plane_type_default, which will only create
DRM_PLANE_TYPE_PRIMARY planes.
We can expose primary planes as supporting more than one CRTC at a time
to more closely resemble plane behavior on DCN but userspace doesn't
really expect planes to be used in this manner and will either
ignore the planes or crash.
Planes with index greater than max_streams are marked as supporting
all CRTCs. No ASIC currently has primary plane count greater than the
stream count but we shouldn't expose more than necessary.
[How]
Drop support for underlay planes. They aren't well tested and don't
fully work right at the moment.
Only create one primary plane per CRTC so we're not creating overlays.
Initialize plane types directly instead of referencing a misleading
array of plane types.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: Tony Cheng <Tony.Cheng@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet.Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
The kms_plane@plane-position-covered-pipe-*-planes subtests can produce
a sequence of atomic commits such that neither active_changed nor
mode_changed but connectors_changed.
When this happens we remove the old stream from the context and add
a new stream but the new stream doesn't have mode_changed=true set.
This incorrect programming sequence causes CRC mismatches to occur in
the test.
The stream->mode_changed value should be set whenever a new stream
is created.
[How]
A new stream is created whenever drm_atomic_crtc_needs_modeset is true.
We previously covered the active_changed and mode_changed conditions
for the CRTC but connectors_changed is also checked within
drm_atomic_crtc_needs_modeset.
So just use drm_atomic_crtc_needs_modeset directly to determine the
mode_changed flag.
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Reviewed-by: Sun peng Li <Sunpeng.Li@amd.com>
Acked-by: Bhawanpreet Lakha <Bhawanpreet.Lakha@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
dc_commit_updates_for_stream is called twice per stream: once
with the flip data and once will all other data. This causes
problems when these DC calls have different numbers of planes
For example, a commit with a pageflip on plane A and a
non-pageflip change on plane B will first call
into DC with just plane A, causing plane B to be
disabled. Then it will call into DC with both planes,
re-enabling plane B
[How]
Merge flip and full into a single bundle
Apart from the single DC call, the logic should not be
changed by this patch
Signed-off-by: David Francis <David.Francis@amd.com>
Reviewed-by: Nicholas Kazlauskas <Nicholas.Kazlauskas@amd.com>
Acked-by: Leo Li <sunpeng.li@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
[Why]
The wait_for_vblank boolean in commit_tail was passed by reference
into each stream commit, and if that commit was an asynchronous
flip, it would disable vblank waits on all subsequent flips.
This made the behaviour depend on crtc order in a non-intuitive way,
although since the asynchronous pageflip flag is only used by the
legacy IOCTLs at the moment it is never an issue
[How]
Find wait_for_vblank before doing any stream commits
Signed-off-by: David Francis <David.Francis@amd.com>
Reviewed-by: Nicholas Kazlauskas <Nicholas.Kazlauskas@amd.com>
Acked-by: Leo Li <sunpeng.li@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
In VRR mode, keep track of the vblank count of the last
completed pageflip in amdgpu_crtc->last_flip_vblank, as
recorded in the pageflip completion handler after each
completed flip.
Use that count to prevent mmio programming a new pageflip
within the same vblank in which the last pageflip completed,
iow. to throttle pageflips to at most one flip per video
frame, while at the same time allowing to request a flip
not only before start of vblank, but also anywhere within
vblank.
The old logic did the same, and made sense for regular fixed
refresh rate flipping, but in vrr mode it prevents requesting
a flip anywhere inside the possibly huge vblank, thereby
reducing framerate in vrr mode instead of improving it, by
delaying a slightly delayed flip requests up to a maximum
vblank duration + 1 scanout duration. This would limit VRR
usefulness to only help applications with a very high GPU
demand, which can submit the flip request before start of
vblank, but then have to wait long for fences to complete.
With this method a flip can be both requested and - after
fences have completed - executed, ie. it doesn't matter if
the request (amdgpu_dm_do_flip()) gets delayed until deep
into the extended vblank due to cpu execution delays. This
also allows clients which want to regulate framerate within
the vrr range a much more fine-grained control of flip timing,
a feature that might be useful for video playback, and is
very useful for neuroscience/vision research applications.
In regular non-VRR mode, retain the old flip submission
behavior. This to keep flip scheduling for fullscreen X11/GLX
OpenGL clients intact, if they use the GLX_OML_sync_control
extensions glXSwapBufferMscOML(, ..., target_msc,...) function
with a specific target_msc target vblank count.
glXSwapBuffersMscOML() or DRI3/Present PresentPixmap() will
not flip at the proper target_msc for a non-zero target_msc
if VRR mode is active with this patch. They'd often flip one
frame too early. However, this limitation should not matter
much in VRR mode, as scheduling based on vblank counts is
pretty futile/unusable under variable refresh duration
anyway, so no real extra harm is done.
According to some testing already done with this patch by
Nicholas on top of my tests, IGT tests didn't report any
problems. If fixes stuttering and flickering when flipping
at rates below the minimum vrr refresh rate.
Fixes: bb47de7366 ("drm/amdgpu: Set FreeSync state using drm VRR
properties")
Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com>
Cc: <stable@vger.kernel.org>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: Michel Dänzer <michel@daenzer.net>
Tested-by: Bruno Filipe <bmilreu@gmail.com>
Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
In VRR mode, keep track of the vblank count of the last
completed pageflip in amdgpu_crtc->last_flip_vblank, as
recorded in the pageflip completion handler after each
completed flip.
Use that count to prevent mmio programming a new pageflip
within the same vblank in which the last pageflip completed,
iow. to throttle pageflips to at most one flip per video
frame, while at the same time allowing to request a flip
not only before start of vblank, but also anywhere within
vblank.
The old logic did the same, and made sense for regular fixed
refresh rate flipping, but in vrr mode it prevents requesting
a flip anywhere inside the possibly huge vblank, thereby
reducing framerate in vrr mode instead of improving it, by
delaying a slightly delayed flip requests up to a maximum
vblank duration + 1 scanout duration. This would limit VRR
usefulness to only help applications with a very high GPU
demand, which can submit the flip request before start of
vblank, but then have to wait long for fences to complete.
With this method a flip can be both requested and - after
fences have completed - executed, ie. it doesn't matter if
the request (amdgpu_dm_do_flip()) gets delayed until deep
into the extended vblank due to cpu execution delays. This
also allows clients which want to regulate framerate within
the vrr range a much more fine-grained control of flip timing,
a feature that might be useful for video playback, and is
very useful for neuroscience/vision research applications.
In regular non-VRR mode, retain the old flip submission
behavior. This to keep flip scheduling for fullscreen X11/GLX
OpenGL clients intact, if they use the GLX_OML_sync_control
extensions glXSwapBufferMscOML(, ..., target_msc,...) function
with a specific target_msc target vblank count.
glXSwapBuffersMscOML() or DRI3/Present PresentPixmap() will
not flip at the proper target_msc for a non-zero target_msc
if VRR mode is active with this patch. They'd often flip one
frame too early. However, this limitation should not matter
much in VRR mode, as scheduling based on vblank counts is
pretty futile/unusable under variable refresh duration
anyway, so no real extra harm is done.
According to some testing already done with this patch by
Nicholas on top of my tests, IGT tests didn't report any
problems. If fixes stuttering and flickering when flipping
at rates below the minimum vrr refresh rate.
Fixes: bb47de7366 ("drm/amdgpu: Set FreeSync state using drm VRR
properties")
Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com>
Cc: <stable@vger.kernel.org>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: Michel Dänzer <michel@daenzer.net>
Tested-by: Bruno Filipe <bmilreu@gmail.com>
Reviewed-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Signed-off-by: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
-----BEGIN PGP SIGNATURE-----
iQFSBAABCAA8FiEEq68RxlopcLEwq+PEeb4+QwBBGIYFAlx8YYIeHHRvcnZhbGRz
QGxpbnV4LWZvdW5kYXRpb24ub3JnAAoJEHm+PkMAQRiGs5MIAIiVVIH+C0St60vf
nzfGpVe+BETe199UveM4Ic2FWMk97ZhSk5Oj6HwYY9vnw4iwoRCZIO5B8Dna4nxY
8XjiwxpJRVLq+7Y1d61O6NHo6UjFHF0GMzyeJeNNUq+mCISxZdLsqzsszt9X09mA
GoJjZ0UMw2Tkz/s3Ie4MumKASc+y2CjJc0ZVEZlJsMaqMJLIfUn/CrTzHBivmuqJ
sV6ZkP4as6h87bI9mi79p8pzvVooCRJ10cg4A/DHG4t2bEAIlB4t5dfZRFzVMhVo
cCPRk9tiA9y4I3zBjcuAZMODcBpfdWoQK8TqYw2cDS3LEDMgnEdIH6snMYHr69z0
kZJjA2A=
=Qs0l
-----END PGP SIGNATURE-----
Merge v5.0 into drm-next
There is a really hairy resolution involving amdgpu fixes, that I'd rather confirm here.
Also some misc fixes are landed by me, but the pr has them as well.
Signed-off-by: Dave Airlie <airlied@redhat.com>
In VRR mode, keep track of the vblank count of the last
completed pageflip in amdgpu_crtc->last_flip_vblank, as
recorded in the pageflip completion handler after each
completed flip.
Use that count to prevent mmio programming a new pageflip
within the same vblank in which the last pageflip completed,
iow. to throttle pageflips to at most one flip per video
frame, while at the same time allowing to request a flip
not only before start of vblank, but also anywhere within
vblank.
The old logic did the same, and made sense for regular fixed
refresh rate flipping, but in vrr mode it prevents requesting
a flip anywhere inside the possibly huge vblank, thereby
reducing framerate in vrr mode instead of improving it, by
delaying a slightly delayed flip requests up to a maximum
vblank duration + 1 scanout duration. This would limit VRR
usefulness to only help applications with a very high GPU
demand, which can submit the flip request before start of
vblank, but then have to wait long for fences to complete.
With this method a flip can be both requested and - after
fences have completed - executed, ie. it doesn't matter if
the request (amdgpu_dm_do_flip()) gets delayed until deep
into the extended vblank due to cpu execution delays. This
also allows clients which want to regulate framerate within
the vrr range a much more fine-grained control of flip timing,
a feature that might be useful for video playback, and is
very useful for neuroscience/vision research applications.
In regular non-VRR mode, retain the old flip submission
behavior. This to keep flip scheduling for fullscreen X11/GLX
OpenGL clients intact, if they use the GLX_OML_sync_control
extensions glXSwapBufferMscOML(, ..., target_msc,...) function
with a specific target_msc target vblank count.
glXSwapBuffersMscOML() or DRI3/Present PresentPixmap() will
not flip at the proper target_msc for a non-zero target_msc
if VRR mode is active with this patch. They'd often flip one
frame too early. However, this limitation should not matter
much in VRR mode, as scheduling based on vblank counts is
pretty futile/unusable under variable refresh duration
anyway, so no real extra harm is done.
Signed-off-by: Mario Kleiner <mario.kleiner.de@gmail.com>
Cc: Nicholas Kazlauskas <nicholas.kazlauskas@amd.com>
Cc: Harry Wentland <harry.wentland@amd.com>
Cc: Alex Deucher <alexander.deucher@amd.com>
Cc: Michel Dänzer <michel@daenzer.net>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
