Bspec lists different VT-d guard numbers (the number of dummy
padding PTEs) for different platforms and plane types. Use those
instead of just assuming the max glk+ number for everything.
This could avoid a bit of overhead on older platforms due to
reduced padding, and it makes it easier to cross check with the
spec.
Note that VLV/CHV do not document this w/a at all, so not sure
if it's actually needed or not. Nor do we actually know how much
padding is required if it is needed. For now use the same 128
PTEs that we use for snb-bdw primary planes.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20250122151755.6928-5-ville.syrjala@linux.intel.com
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Async flips often require bigger alignment that sync flips.
Currently we have HAS_ASYNC_FLIPS() checks strewn about to
inidcate that async flips are generally supported and thus
we want more alignment. Switch that over to using
intel_plane_can_async_flip() so that we can handle these
in a slightly less messy way. Currently we don't have cases
where async flips would require different alignment for
different modifiers on the same plane.
We'll also move the HAS_ASYNC_FLIPS() check to the plane init
code so that we can still use that as a quick way to disable
the async flips workarounds for testing purposes.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20241009182207.22900-5-ville.syrjala@linux.intel.com
Reviewed-by: Jouni Högander <jouni.hogander@intel.com>
We need to be able to do both MMIO and DSB based pipe/plane
programming. To that end plumb the 'dsb' all way from the top
into the plane commit hooks.
The compiler appears smart enough to combine the branches from
all the back-to-back register writes into a single branch.
So the generated asm ends up looking more or less like this:
plane_hook()
{
if (dsb) {
intel_dsb_reg_write();
intel_dsb_reg_write();
...
} else {
intel_de_write_fw();
intel_de_write_fw();
...
}
}
which seems like a reasonably efficient way to do this.
An alternative I was also considering is some kind of closure
(register write function + display vs. dsb pointer passed to it).
That does result is smaller code as there are no branches anymore,
but having each register access go via function pointer sounds
less efficient.
Not that I actually measured the overhead of either approach yet.
Also the reg_rw tracepoint seems to be making a huge mess of the
generated code for the mmio path. And additionally there's some
kind of IS_GSI_REG() hack in __raw_uncore_read() which ends up
generating a pointless branch for every mmio register access.
So looks like there might be quite a bit of room for improvement
in the mmio path still.
Reviewed-by: Animesh Manna <animesh.manna@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240930170415.23841-12-ville.syrjala@linux.intel.com
Different hardware generations have different scanout alignment
requirements. Introduce a new vfunc that will allow us to
make that distinction without horrible if-ladders.
For now we directly plug in the existing intel_surf_alignment()
and intel_cursor_alignment() functions.
For fbdev we (temporarily) introduce intel_fbdev_min_alignment()
that simply queries the alignment from the primary plane of
the first crtc.
TODO: someone will need to fix xe's alignment handling
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240612204712.31404-4-ville.syrjala@linux.intel.com
Acked-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
On bdw-glk the sync->async flip change takes an extra frame due to
the double buffering behaviour of the async flip plane control bit.
Since on skl+ we are now explicitly converting the first async flip
to a sync flip (in order to allow changing the modifier and/or
ddb/watermarks) we are now taking two extra frames until async flips
are actually active. We can drop that back down to one frame by
setting the async flip bit already during the sync flip.
Note that on bdw we don't currently do the extra sync flip (see
intel_plane_do_async_flip()) so technically we wouldn't have
to deal with this in i9xx_plane_update_arm(). But I added the
relevant snippet of code there as well, just in case we ever
decide to go for the extra sync flip on pre-skl platforms as
well (we might, for example, want to change the fb stride).
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240430095639.26390-5-ville.syrjala@linux.intel.com
Reviewed-by: Arun R Murthy <arun.r.murthy@intel.com>
On MTL the GOP (for whatever reason) likes to bind its framebuffer
high up in the ggtt address space. This can conflict with whatever
ggtt_reserve_guc_top() is trying to do, and the result is that
ggtt_reserve_guc_top() fails and then we proceed to explode when
trying to tear down the driver. Thus far I haven't analyzed what
causes the actual fireworks, but it's not super important as even
if it didn't explode we'd still fail the driver load and the user
would be left with an unusable GPU.
To remedy this (without having to figure out exactly what
ggtt_reserve_guc_top() is trying to achieve) we can attempt to
relocate the BIOS framebuffer to a lower ggtt address. We can do
this at this early point in driver init because nothing else is
supposed to be clobbering the ggtt yet. So we simply change where
in the ggtt we pin the vma, the original PTEs will be left as is,
and the new PTEs will get written with the same dma addresses.
The plane will keep on scanning out from the original PTEs until
we are done with the whole process, and at that point we rewrite
the plane's surface address register to point at the new ggtt
address.
Since we don't need a specific ggtt address for the plane
(apart from needing it to land in the mappable region for
normal stolen objects) we'll just try to pin it without a fixed
offset first. It should end up at the lowest available address
(which really should be 0 at this point in the driver init).
If that fails we'll fall back to just pinning it exactly to the
origianal address.
To make sure we don't accidentlally pin it partially over the
original ggtt range (as that would corrupt the original PTEs)
we reserve the original range temporarily during this process.
v2: Try to pin explicitly to ggtt offset 0 as otherwise DG2 puts it
even higher (atm we have no PIN_LOW flag to force it low)
v3: "fix" xe
Reviewed-by: Andrzej Hajda <andrzej.hajda@intel.com>
Tested-by: Paz Zcharya <pazz@chromium.org>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20240202224340.30647-16-ville.syrjala@linux.intel.com
Acked-by: Lucas De Marchi <lucas.demarchi@intel.com>
Split out frontbuffer related declarations and static inlines from
gem/i915_gem_object.h into new gem/i915_gem_object_frontbuffer.h.
The main goal is to reduce header interdependencies. With
gem/i915_gem_object.h including display/intel_frontbuffer.h,
modification of the latter causes a whopping 300+ objects to be rebuilt,
while many of the source files actually needing it aren't explicitly
including it at all.
After the change, only 21 objects depend on display/intel_frontbuffer.h,
directly or indirectly.
Cc: Jouni Högander <jouni.hogander@intel.com>
Reviewed-by: Jouni Högander <jouni.hogander@intel.com>
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20230830085127.2416842-1-jani.nikula@intel.com
Turns out many of the files that need i915_reg.h get it implicitly via
{display/intel_de.h, gt/intel_context.h} -> i915_trace.h -> i915_irq.h
-> i915_reg.h. Since i915_trace.h doesn't actually need i915_irq.h,
makes sense to drop it, but that requires adding quite a few new
includes all over the place.
Prefer including i915_reg.h where needed instead of adding another
implicit include, because eventually we'll want to split up i915_reg.h
and only include the specific registers at each place.
Also some places actually needed i915_irq.h too.
Cc: Lucas De Marchi <lucas.demarchi@intel.com>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/6e78a2e0ac1bffaf5af3b5ccc21dff05e6518cef.1668008071.git.jani.nikula@intel.com
Drop the locks around most primary plane register writes.
The lock isn't needed since each plane's register are neatly
contained on their own cachelines.
The one exception we have to make is DSPADDR/DSPSURF which is
(ab)used to also trigger FBC nukes on pre-snb (since the
hardware doesn't seem to have any dedicated mechanism to
trigger nukes). So we need to keep the lock around it to
protect against the rmw performed by the fbc code.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20220210062403.18690-5-ville.syrjala@linux.intel.com
Reviewed-by: Stanislav Lisovskiy <stanislav.lisovskiy@intel.com>
Use REG_GENMASK() & co. when dealing with PIPESRC.
Note that i9xx_get_initial_plane_config() will now use the
full 16 bit mask whereas previously it used 12 bits only.
But intel_get_pipe_src_size() already used the full 16 bits
on all platforms anyway, so at least we're consistent now.
The high bits beyond the max supported pipe source size
should not be set in any case so this seems fine.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20211112193813.8224-7-ville.syrjala@linux.intel.com
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
