If we have to force the hardware to go through a full modeset
due to eg. cdclk reprogramming, we need to preserve
crtc_state->inherited for all crtcs that have not otherwise
gone through the whole compute_config() stuff after connectors
have been detected.
Otherwise eg. cdclk induced modeset glk_force_audio_cdclk()
will clear the inherited flag, and thus the first real commit
coming from userspace later on will not be forced through
the full .compute_config() path and so eg. audio state may
not get properly recomputed.
But instead of adding all kinds of ad-hoc crtc_state->inherited
preservation hacks all over, let's change things so that we
only clear it for the crtcs directly included in userspace/client
initiated commits.
Should be far less fragile since now we just need to remember
to flag the internal commits, and not worry about where new
crtcs might get pulled in.
Closes: https://gitlab.freedesktop.org/drm/intel/-/issues/5260
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230328122357.1697-1-ville.syrjala@linux.intel.com
Reviewed-by: Luca Coelho <luciano.coelho@intel.com>
Move VRR enabling/disabling into a place where it also works
for fastsets.
With this we always start the transcoder up in non-VRR mode.
Granted we already did that but for a very short period of
time. But now that we might end up doing a bit more with the
transcoder in non-VRR mode it seems prudent to also update
the active timings as the transcoder changes its operating
mode.
crtc_state->vrr.enable still tracks whether VRR is actually
enabled or not, but now we configure all the other VRR timing
registers whenever VRR is possible (whether we actually enable
it or not). crtc_state->vrr.flipline can now serve as our
"is VRR possible" bit of state.
I decided to leave the MSA timing ignore bit set all the time
whether VRR is actually enabled or not. If the sink can figure
out the timings with that information when VRR is active then
surely it can also do it when VRR is inactive.
v2: Protect intel_vrr_set_transcoder_timings() with HAS_VRR()
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230321135615.27338-1-ville.syrjala@linux.intel.com
Reviewed-by: Mitul Golani <mitulkumar.ajitkumar.golani@intel.com>
Read out the pipe/output csc matrices on ilk+ and stash the results
(in the hardware specific format) into the appropriate place
in the crtc state.
Note that on skl/glk/icl the pipe csc unit suffers from an issue
where *reads* of the coefficient/offset registers also disarm
the double buffer update (if currently armed via CSC_MODE write).
So it's rather important that the readout only happens after the
csc registers have been latched. Fortunately the state checker
only runs after the start of vblank where the latching happens.
And on skl/glk the DMC + CSC register read has the potential to
corrupt the latched CSC register values, so let's add a comment
reminding us that the DC states should remain off until the
readout has been completed.
TODO: maybe we could somehow check to make sure PSR has in fact
latched the new register values already, and that DC states
have been off all along?
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230329135002.3096-9-ville.syrjala@linux.intel.com
Reviewed-by: Ankit Nautiyal <ankit.k.nautiyal@intel.com>
The encoder update_prepare()/complete() hooks were added to hold a
TC port link reference for all outputs in the atomic state around the
whole modeset enable sequence - thus locking the ports' TC mode - and
set the TBT/DP-alt PLL type corresponding to the current TC mode.
Since nothing depends on the PLL selection before/after then encoder's
pre_pll_enable/post_pll_disable hooks are called, the above steps can be
moved to these hooks, so do that and remove the
update_prepare()/complete() hooks.
Reviewed-by: Mika Kahola <mika.kahola@intel.com>
Signed-off-by: Imre Deak <imre.deak@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230323142035.1432621-30-imre.deak@intel.com
Keeping DC states enabled is incompatible with the _noarm()/_arm()
split we use for writing pipe/plane registers. When DC5 and PSR
are enabled, all pipe/plane registers effectively become self-arming
on account of DC5 exit arming the update, and PSR exit latching it.
What probably saves us most of the time is that (with PIPE_MISC[21]=0)
all pipe register writes themselves trigger PSR exit, and then
we don't re-enter PSR until the idle frame count has elapsed.
So it may be that the PSR exit happens already before we've
updated the state too much.
Also the PSR1 panel (at least on this KBL) seems to discard the first
frame we trasmit, presumably still scanning out from its internal
framebuffer at that point. So only the second frame we transmit is
actually visible. But I suppose that could also be panel specific
behaviour. I haven't checked out how other PSR panels behave, nor
did I bother to check what the eDP spec has to say about this.
And since this really is all about DC states, let's switch from
the MODESET domain to the DC_OFF domain. Functionally they are
100% identical. We should probably remove the MODESET domain...
And for good measure let's toss in an assert to the place where
we do the _noarm() register writes to make sure DC states are
in fact off.
v2: Just use intel_display_power_is_enabled() (Imre)
Cc: <stable@vger.kernel.org> #v5.17+
Cc: Manasi Navare <navaremanasi@google.com>
Cc: Drew Davenport <ddavenport@chromium.org>
Cc: Jouni Högander <jouni.hogander@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Fixes: d13dde4495 ("drm/i915: Split pipe+output CSC programming to noarm+arm pair")
Fixes: f8a005eb89 ("drm/i915: Optimize icl+ universal plane programming")
Fixes: 890b6ec4a5 ("drm/i915: Split skl+ plane update into noarm+arm pair")
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230320183532.17727-1-ville.syrjala@linux.intel.com
Add some (probably overkill) locking to protect the vblank
timestamping constants updates during seamless M/N fastsets.
As everything should be naturally aligned I think the individual
pieces should probably end up updating atomically enough. So this
is only really meant to guarantee everyone sees a consistent whole.
All the drm_vblank.c usage is covered by vblank_time_lock,
and uncore.lock will take care of __intel_get_crtc_scanline()
that can also be called from outside the core vblank functionality.
Currently only crtc_clock and framedur_ns can change, but in
the future might fastset also across eg. vtotal/vblank_end
changes, so let's just grab the locks across the whole thing.
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230310235828.17439-2-ville.syrjala@linux.intel.com
intel_crtc_prepare_cleared_state() is unintentionally losing
the "inherited" flag. This will happen if intel_initial_commit()
is forced to go through the full modeset calculations for
whatever reason.
Afterwards the first real commit from userspace will not get
forced to the full modeset path, and thus eg. audio state may
not get recomputed properly. So if the monitor was already
enabled during boot audio will not work until userspace itself
does an explicit full modeset.
Cc: stable@vger.kernel.org
Tested-by: Lee Shawn C <shawn.c.lee@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230223152048.20878-1-ville.syrjala@linux.intel.com
Reviewed-by: Uma Shankar <uma.shankar@intel.com>
The DSI code has some local hacks to program TRANS_VBLANK on
TGL+ (ICL DSI transcoders didn't have this register). That
will not work when we need to start using the delayed vblank
(for DSB purposes). Too lazy to figure out what the is going
on there, so just sprinkle FIXMEs in the hopes someone else
will spot them eventually.
v2: Only TRANS_{HBLANK,SET_CONTEXT_LATENCY} still no not
exist for DSI transcoders, only TRANS_VBLANK
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230213225258.2127-12-ville.syrjala@linux.intel.com
Acked-by: Jani Nikula <jani.nikula@intel.com>
On TGL VBLANK.VBLANK_START was the mechanism by which we can
delay the pipe's internal vblank in relation to the transcoder's
vblank. On ADL+ that no longer does anything. Instead we must
now use the new TRANS_SET_CONTEXT_LATENCY register. Program it
accordingly.
And since VBLANK.VBLANK_START is no longer used by the hardware
on ADL+ let's just zero it out to make it stand out in register
dumps. Seeing the zeroed value should hopefully remind people
to check the other register instead.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230213225258.2127-11-ville.syrjala@linux.intel.com
Acked-by: Jani Nikula <jani.nikula@intel.com>
Define the contents of the transcoder timing registers using
REG_GENMASK() & co. For ease of maintenance let's just define
the bitmasks with the full 16bit width (also used by the
current hand rolled stuff) even though not all bits are actually
used. None of the unsued bits have ever contained anything.
Jani spotted that the CRT load detection code did use narrower
bitmasks, so that is now going to change. But that is fine
since any garbage in the high bits would have been caught by
the state checker that always used the full 16bit masks.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230213225258.2127-10-ville.syrjala@linux.intel.com
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Rename PIPECONF to TRANSCONF to make it clear what it actually
applies to.
While the usual convention is to pick the earliers name I think
in this case it's more clear to use the later name. Especially
as even the register offset is in the wrong range (0x70000 vs.
0x60000) and thus makes it look like this is per-pipe.
There is one place in gvt that's doing something with TRANSCONF
while iterating with for_each_pipe(). So that might not be doing
the right thing for TRANSCODER_EDP, dunno. Not knowing what it
does I left it as is to avoid breakage.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230213225258.2127-5-ville.syrjala@linux.intel.com
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Name the CPU transcoder timing registers TRANS_FOO rather than
just FOO. This is the modern name, after the pipe/transcoder split
happened. Makes it a bit more obvious whether you pass in a pipe or
a transcoder.
PIPESRC is a bit special as it's a pipe register, even though it
lives in the transcoder registers range (0x60000 instead of 0x70000).
And BCLRPAT I suppose is a transcoder register (since it has something
to do with the timing generator), but it doesn't even exist after gen4
so I left it to use the only name it ever had in bspec.
And while at it let's pass in the correct enum in few more
places why don't we. Although in all those places the distinction
doesn't matter.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230213225258.2127-4-ville.syrjala@linux.intel.com
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Add new files intel_wm.[ch] and i9xx_wm.[ch] under display/ to hold
generic and pre-SKL watermark code, respectively. SKL+ watermark code
has already been split out to skl_watermark.[ch].
Use the _wm.[ch] naming for brevity; we may want to rename
skl_watermark.[ch] later accordingly.
Add new intel_wm_init() to call either skl_wm_init() or
i9xx_wm_init(i915) depending on the platform, the latter comprising of
the remains of intel_init_pm().
Sprinkle in some minor checkpatch fixes while moving the code.
v2:
- Rebase
- Fix copyright year
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/ddf04a07a37f0368b3fef85d4ebb924082fec6cd.1676317696.git.jani.nikula@intel.com
Determining whether the display engine is present on a platform happens
only in intel_device_info_runtime_init(). Initializing the display power
functionality depends on this condition, so move
intel_power_domains_init() later after the runtime init function has
been called.
The next patch fixing platforms without display, depends on this patch.
Acked-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Imre Deak <imre.deak@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230208114300.3123934-3-imre.deak@intel.com
Current implementation of async flip w/a relies on assumption that
previous atomic commit contains valid information if async_flip is still
enabled on the plane. It is incorrect. If previous commit did not modify
the plane its state->uapi.async_flip can be false. As a result DMAR/PIPE
errors can be observed:
i915 0000:00:02.0: [drm] *ERROR* Fault errors on pipe A: 0x00000080
i915 0000:00:02.0: [drm] *ERROR* Fault errors on pipe A: 0x00000080
DMAR: DRHD: handling fault status reg 2
DMAR: [DMA Read NO_PASID] Request device [00:02.0] fault addr 0x0 [fault reason 0x06] PTE Read access is not set
v2: update async_flip_planes in more reliable places (Ville)
v3: reset async_flip_planes and do_async_flip in more scenarios (Ville)
v4: move all resets to plane loops (Ville)
Signed-off-by: Andrzej Hajda <andrzej.hajda@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230127153003.2225111-1-andrzej.hajda@intel.com
