drm/i915: use i915 and Ironlake CRTC enable/disable functions in prepare/commit

This will allow us to optimize our prepare/commit paths a bit better.

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
[ickle: minor tweak to handle the cursor across pipe resizing]
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: don't unlock panel regs

This was just a workaround for some broken Ironlake CRTC code.

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: split i9xx CRTC enable/disable code

So we can use it for CRTC prepare/commit.

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: split Ironlake CRTC enable/disable code

This way we can also use it in CRTC prepare/commit.  Also makes it
easier to split out FDI and other code.

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: Adapt workqueue to new alloc_workqueue interface

create_singlethreaded_workqueue() is being phased out for a new
concurrency managed task infrastructure.

Adapt our workqueue constructor to explicitly create a domain that only
allows the execution of a single task at any time. All the tasks are
expected to require the dev->struct_mutex, so would block concurrency of
other tasks if we allow more than a single i915 task to be run at once.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915/debug: Include Ironlake in self-refresh status

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: Make the connector->encoder relationship explicit

Currently we have a exact mapping of a connector onto an encoder for its
whole lifetime. Make this an explicit property of the structure and so
simplify the code.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>

drm/i915: Use the direct mapping of pipe->crtc

Why iterate all the crtcs to find the pipe, when we already know which
crtc is attached to which pipe?

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>

drm/i915: Rename intel_encoder->enc to base for consistency

[Patch is slightly larger than is strictly necessary to fixup
surrounding checkpatch.pl errors.]

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>

drm/i915/dp: Flush the PLL register write before sleeping

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: make sure panel is sequenced off when starting a mode set

Otherwise we may not be able to train the DP link.

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: make sure VDD AUX power has time to settle

When turning on or off the VDD AUX bit, we need to give the panel time
to start or stop or AUX transactions may fail.

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: don't change VDD AUX status in panel power functions

Mode set sequence outlines when the AUX VDD bit should be set and
cleared, and it's separate from the panel power sequence.

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: split DP link training across panel power sequencing

Mode set sequence requires that we start training, then enable the
panel, then complete training.  So split the DP training function into
two parts; the first enables the DP port and sets training pattern 1 and
the second completes the training.

As part of this, remove some redundant function args from the various DP
handling functions and use the intel_dp fields everywhere we can.

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
[ickle: removed first ironlake_edp_backlight_on() on advice of jbarnes]
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: use VDD AUX for panel power around detection and in prepare

Mode setting sequence specifies that we use VDD AUX for configuration
and detection, and early in the mode set sequence.  Only later (after
DP_A has started training) should we actually enable panel power.

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
[ickle: checkpatch.pl complaining about whitespace]
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: use 125MHz reference clock for PCH attached eDP

Fix the test so we don't try to use the 450MHz refclk on PCH attached
eDP.

References:
  https://bugs.freedesktop.org/show_bug.cgi?id=29141

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: fix eDP detection

Panel needs to be powered up.

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

i915: snprintf returns large values

snprintf() returns the number of bytes which would have been used if
there was enough space.  It can be larger than the size of the buffer.
Obviously in this case the buffer is large enough but everyone just
copy and pastes this code so it's better to limit it and set a good
example.

Signed-off-by: Dan Carpenter <error27@gmail.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: die, i915_probe_agp, die

Use the detection from intel-gtt.ko instead. Hooray!

Also move the stolen mem allocator to the other gtt stuff in dev_prv->mem.

v2: Chris Wilson noted that my error handling was crap. Fix it. He also
said that this fixes a problem on his i845. Indeed, i915_probe_agp
misses a special case for i830/i845 stolen mem detection.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=25476
Cc: stable@kernel.org
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: drop prealloc_start from i915_dma gtt init

Not used and simply confusing.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

intel-gtt: use chipset generation number some more

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

intel-gtt: call init_gtt_init in probe function

This way create_gatt_table become dummy glue functions for the fake
agp driver - rename them accordingly (and kill the now unnecessary
i9xx copy).

With this change, the gtt initialization code is almost independant
from the agp stuff. Two things are still missing:
- the scratch page is created by the generic agp code.
- filling the whole gtt with scratch_page ptes is not yet consolidated -
  this needs abstracted pte handling, first.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

intel-gtt: consolidate i9xx setup

The only difference between i915 and i965 was the calculation of the
gtt address. So merge these two paths into one. Otherwise the same
changes as in the i830 setup consolidation.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

intel-gtt: consolidate i830 setup

Slighlty reordered sequence was necessary. Also don't set
agp_bridge->gatt_bus_addr anymore. Only used by generic agp helper
functions, hence unnecessary for the intel fake agp driver.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

intel-gtt: consolidate the gtt ioremap calls

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

intel-gtt: i830: adjust ioremap of regs and gtt to i9xx

This way around this can be extracted into common code.

Also use a common cleanup function (and give it a generic name).

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

intel-gtt: i965: use detected gtt size for mapping

Also move the Sandybdridge size detection into gtt_total_entries, like
the rest.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

intel-gtt: i915: use detected gtt size for mapping

Slight reordering of the init sequence required.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

intel-gtt: introduce intel_gtt_driver

Same idea as INTEL_INFO from drm/i915. This
- reduces the dependancy on agp_driver
- stops the what-does-IS_I965G-mean confusion (here it's just gen4, in
  drm/i915 it's gen >=4)
- further prepares the separation of the fake agp driver from the rest.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

intel-gtt: fix gtt_total_entries detection

In commit f1befe71 Chris Wilson added some code to clear the full gtt
on g33/pineview instead of just the mappable part. The code looks like
it was copy-pasted from agp/intel-gtt.c, at least an identical piece
of code is still there (in intel_i830_init_gtt_entries). This lead to
a regression in 2.6.35 which was supposedly fixed in commit e7b96f28

Now this commit makes absolutely no sense to me. It seems to be
slightly confused about chipset generations - it references docs for
4th gen but the regression concerns 3rd gen g33. Luckily the the g33
gmch docs are available with the GMCH Graphics Control pci config
register definitions. The other (bigger problem) is that the new
check in there uses the i830 stolen mem bits (.5M, 1M or 8M of stolen
mem). They are different since the i855GM.

The most likely case is that it hits the 512M fallback, which was
probably the right thing for the boxes this was tested on.

So the original approach by Chris Wilson seems to be wrong and the
current code is definitely wrong. There is a third approach by Jesse
Barnes from his RFC patch "Who wants a bigger GTT mapping range?"
where he simply shoves g33 in the same clause like later chipset
generations.

I've asked him and Jesse confirmed that this should work. So implement
it.

Bugzilla: https://bugzilla.kernel.org/show_bug.cgi?id=16891$
Tested-by: Anisse Astier <anisse@astier.eu>
Cc: stable@kernel.org
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

intel-gtt: s/i8[13]0/fake_agp for generic functions

Start to separate the fake agp driver from the rest of intel-gtt.c

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

intel-gtt: adjust overhead entries in intel_gtt_stolen_entries

agp/intel_gtt.c and drm/i915/i915_dma.c don't calculate this the same
way: The intel-gtt code seems to use the actual gtt size, the drm
module just the mappable. Go with the logic from the drm module because
that's the more conservative choice.

But conserve the original code in intel_gtt_total_size for later use.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

intel-gtt: drop unnecessary conditions in intel_gtt_stolen_entries

The dedection function in drm/i915/i915_dma.c works without it, so
drop it here, too. All the values are disdinct, anyway.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

intel-gtt: sane variable names for intel_gtt_stolen_entries

This somewhat aligns it with the version in drm/i915/i915_dma.c.
Changes:
- s/gtt_entries/stolen_size
- track overhead entries in a seperate var (the effective gtt size
  calculation will be extracted later on).
- subtract the overhead at the end instead of in each clause.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

intel-gtt: generic intel_fake_agp_fetch_size

This uses the new mappable gtt size detection from the previous patch.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

intel-gtt: new function intel_gtt_mappable_entries

This implementation is stolen from drm/i915, but is equivalent to
the code sprinkled over intel-gtt.c in the various fetch_size functions.
It's not yet used anywhere, though.

Also introduce intel_gtt_init which only calls intel_gtt_stolen_entries.
Over the course of the next patches, this will grow untill it contains
the complete init sequence starting from the call to gtt_mappable_entries.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

intel-gtt: s/intel_i830_init_gtt_entries/intel_gtt_stolen_entries

First simple step towards a more generic initialization. This
is needed to disentangle the agp stuff from the stuff that is
actually needed by drm/i915.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

intel-gtt: store a local pointer to the bridge pci dev

When the intel-gtt code now longer depends on agp, we cannot rely
on this. So store a local reference in intel-gtt.c.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

intel-gtt: introduce drm/intel-gtt.h

Add a few definitions to it that are already shared and that will
be shared in the future (like the number of stolen entries).
No functional changes in here.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

agp/intel: make intel-gtt.c into a real source file

Now that the disentangling is complete, stop including intel-gtt.c
from intel-agp.c.

The linux build system _really_ doesn't allow .c source files with the
same name as the module. It fails with the following message when trying
to build such a bugger:

make[3]: Circular drivers/char/agp/intel-agp.o <- drivers/char/agp/intel-agp.o dependency dropped.

Instead of renameing intel-agp.c I've simply created a new module out
of intel-gtt.c. Renaming intel-agp.ko to something else is not an option
for it will surely kill someones boot process.

This also paves the way to use the gtt code without loading the agp
driver.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

agp/intel: split out gmch/gtt probe, part 2

This just splits the device list into two and moves the gtt related stuff
to intel-gtt.c. The two new devices lists also lose the not longer needed
fields. There where only about 5 cases anyway with both a gmch and a
possible agp port, so the duplication of entries is rather small.
Additionally kill 2 out of the three Ironlake mobile entries that
only differed in host bridge pci id.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: Don't disable panel for modesetting if pfit hasn't changed

It seems to be possible to program a new mode without disabling the panel
if the panel fitter setup doesn't change. Add support for that.

Signed-off-by: Matthew Garrett <mjg@redhat.com>
Acked-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: set dither bits on eDP panels too

We really need a macro to test whether a given connector has a panel
attached rather than sprinkling HAS_PCH_SPLIT/IS_eDP/has_edp_encoder
etc all over. In the meantime, fix the bug...

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
[ickle: tidy up the duplicity in the conditionals]
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: fix pipeconf dither bit definitions

Make them match the others and add BPP definitions.

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: Remove impossible error handling from bit17 swizzling

Our usage of kmap() cannot return NULL here, so remove the unnecessary
error handling.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: Add ringbuffer wait reset to hangcheck

The GPU records whether it is currently waiting for a completion of a
WAIT_FOR_EVENT in the RB_WAIT bit in the ringbuffer control registers.
On third generation chipsets and later, a write of 1 to this bit breaks
the hang and returns the GPU to arbitration, i.e. the GPU should
continue executing the reminder of the batchbuffer and return to normal
operations.

By adding this to hangcheck we can avoid a full GPU reset under these
conditions.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: Clear scanline waits after disabling the pipe.

If we disable the pipe and the GPU is currently waiting on a scanline
WAIT_FOR_EVENT, the GPU will hang. Fortunately, there is a magic bit
which we can write on i915+ to break this wait after disabling the
pipe.

References:

  Bug 29252 - [Arrandale] Hung WAIT_FOR_EVENT when running rss-glx-skyrocket
  https://bugs.freedesktop.org/show_bug.cgi?id=29252

  Bug 28964 - [i965gm] GPU infinite MI_WAIT_FOR_EVENT while watching video in Totem
  https://bugs.freedesktop.org/show_bug.cgi?id=28964

and many others.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Jesse Barnes <jbarnes@virtuousgeek.org>

drm/i915: Show framebuffer info in debugfs

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: Show device capabilities in debugfs

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915/tv: After disabling the pipe, use wait_for_vblank_off()

Hopefully this is a contributing factor to the spurious TV detection
repoted by Ivan Bulatovic and others.

References:

  Bug 16871 - "TV1 connected" with no tv
  https://bugzilla.kernel.org/show_bug.cgi?id=16871

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reported-by: Ivan Bulatovic <combuster@gmx.com>
Cc: Jesse Barnes <jbarnes@virtuousgeek.org>

drm/i915: Refactor panel backlight controls

There were two instances of code to control the panel backlight and
neither handled the complete set of device variations.

Fixes:

  Bug 29716 - [GM965] Regression: Backlight resets to minimum when changing resolution
  https://bugs.freedesktop.org/show_bug.cgi?id=29716

And a bug on one of my PineView boxes which overflowed the backlight
value.

Incorporates part of a similar patch by Matthew Garrett that exposes a
native Intel backlight controller.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: Remove redundant initialisation of fb_base

We do it whilst configuring dev->mode_config, so remove the out-of-place
earlier initialisation.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: Quieten sparse warnings for missing prototypes.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: Kill the active list spinlock

This spinlock only served debugging purposes in a time when we could not
be sure of the mutex ever being released upon a GPU hang. As we now
should be able rely on hangcheck to do the job for us (and that error
reporting should not itself require the struct mutex) we can kill the
incomplete attempt at protection.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: Remove the random SyncFlush during initialisation

We have no idea why we request a SyncFlush via INSTPM at that point in
time -- we certainly never check for its completion...

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: Compile out error state without DEBUG_FS

Alexander reported that the compilation of intel_overlay.c was failing
due to an inclusion that was only valid with CONFIG_DEBUG_FS. As the
whole error reporting is only useful with debugfs enabled, remove all
the redundant error state collection code when compiling without
CONFIG_DEBUG_FS.

Reported-by: Alexander Lam <lambchop468@gmail.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915/overlay: Use a continuation hook to finish work after a flip.

Slightly easier to follow than the state machine and now possible as the
control structure is opaque and hw_wedged is no longer interferred with.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/overlay: Make the overlay control struct opaque.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/overlay: Pass interruptible to switch_off()

During DPMS we currently do not want the overlay code to be
interruptible, so pass that information down and only take the
uninterrruptible paths.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/overlay: Workaround i830 overlay activation bug.

On i830, there exists a bug where an overlay on pipe B requires the mode
clock on pipe A in order to activate. So workaround this by activating
pipe A when trying to enable the overlay on pipe B.

References:

  [Bug 29007] GPU hang on video playback with overlay
  https://bugs.freedesktop.org/show_bug.cgi?id=29007

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915/overlay: Make do_put_image() as static

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: Preallocate requests

By allocating the request prior to writing to the ringbuffer, we can
abort the operation without leaving the GPU in an inconsistent state.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/overlay: Tidy release_old_vid()

Inline the call to wait_flip() and simplify the resulting code.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/overlay: Combine SWITCH_OFF into a single step

We can program the h/w to first wait on the flip and then switch off
without relying on s/w intervention. This removes the need for a double
step switch off, bringing much rejoicing.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/overlay: Explicitly pass regs from map to unmap

The scoping of the validity of the mapping is thus clarified.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/overlay: Refactor do_wait_request()

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/overlay: Tidy check_overlay_dst()

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/overlay: Tidy update_pfit_vscale_ratio()

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/overlay: Remove duplicated definition of OFC_UPDATE

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/overlay: Destroy reg_bo on shutdown.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/overlay: Use the recommended page alignment for physical regs

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/overlay: Tidy attribute checking.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/overlay: Use non-atomic mappings for the common case.

The only time where an atomic mapping is required is during
error-capture and there we cannot use the default slot, but need to
specifically use one of the IRQ slots. So separate out the two
conditions and use the atomic mapping only when appropriate.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/overlay: Move capabilities bits to common info block.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/overlay: Ensure that the reg_bo is in the GTT prior to writing.

Just makes sure that writes are not being aliased by the CPU cache and
do make it out to main memory.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=24977
Cc: stable@kernel.org

drm/i915/overlay: Missing breaks between case statements for color depth

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/overlay: Whitespace

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: drop i915_add_request right in front of i915_wait_request

... take advantage of the new implicit request issuing of
i915_wait_request.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: move the wait_rendering call into flush_gpu_write_domain

One caller (for the pageflip support) wants a purely pipelined flush.
Distinguish this case by a new parameter. This will also be useful
later on for pipelined fencing.

v2: Simplify the code by depending upon the implicit request emitting
of i915_wait_request.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
[ickle: And drop the non-interruptible support in the process.]
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: drop seqno argument from i915_gem_object_move_to_active

By moving one i915_add_request we can solely depend on the new
auto-seqno-numbering behaviour.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: kill a no longer necessary BUG_ON

i915_gem_object_move_to_active can handle zero seqno for us now.
And not emitting a request is not fatal here - we'll try to emit
a new one if we have to wait for some rendering to complete.

In case this assumption ever gets accidentally broken, there's already
a BUG_ON to catch it in i915_do_wait_request.

So just silently ignore ENOMEM here instead of screwing up the whole
drm.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: move flushing list processing to i915_retire_commands

... instead of threading flush_domains through the execbuf code to
i915_add_request.

With this change 2 small cleanups are possible (likewise the majority
of the patch):

- The flush_domains parameter of i915_add_request is always 0. Drop it
  and the corresponding logic.
- Ditto for the seqno param of i915_gem_process_flushing_list.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: only one interrupt per batchbuffer is not enough!

Previously I thought that one interrupt per batchbuffer should be
enough. Now tedious benchmarking showed this to be wrong.

Therefore track whether any commands have been isssued with a future
seqno (like pipelined fencing changes or flushes). If this is the case
emit a request before issueing the batchbuffer.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: move flushing list processing to i915_gem_flush

Now that we can move objects to the active list without already having
emitted a request, move the flushing list handling into i915_gem_flush.
This makes more sense and allows to drop a few i915_add_request calls
that are not strictly necessary.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: allow lazy emitting of requests

Sometimes (like when flushing in preparation of batchbuffer execution)
we know that we'll emit a request but haven't yet done so. Allow this
case by simply taking the next seqno by default. Ensure that a request
is eventually emitted before waiting for an request by issuing it
in i915_wait_request iff this is not yet done.

Also replace one open-coded version of i915_gem_object_wait_rendering,
to prevent future code-diversion.

Chris Wilson asked me to explain and clarify what this patch does and why.
Here it goes:

Old way of moving objects onto the active list and associating them with a
reques:

1. i915_add_request + store the returned seqno somewhere
2. i915_gem_object_move_to_active (with the stored seqno as parameter)

For the current users, this is all fine. But I'd like to associate objects
(and fence regs) with the batchbuffer request deep down in the execbuf
call-chain. I thought about three ways of implementing this.

a) Don't care, just emit request when we need a new seqno. When heavily
pipelining fence reg changes, this would have caused tons of superflous
request (and corresponding irqs).

b) Thread all changed fences, objects, whatever through the execbuf-maze,
so that when we emit a request, we can store the new seqno at all the right
places.

c) Kill that seqno-threading-around business by simply storing the next
seqno, i.e. allow 2. to be done before 1. in the above sequence.

I've decided to implement c) (in this patch). The following patches are
just fall-out that resulted from this small conceptual change.

* We can handle the flushing list processing where we actually emit a flush
  (i915_gem_flush and i915_retire_commands) instead of in i915_add_request.
  The code makes IMHO more sense this way (and i915_add_request looses the
  flush_domains parameter, obviously).

* We can avoid emitting unnecessary requests. IMHO there's no point in
  emitting more than one request per batchbuffer (with or without an
  corresponding irq).

* By enforcing 2. before 1. ordering in the above sequence the seqno
  argument of i915_gem_object_move_to_active is redundant and can be
  dropped.

v2: Now i915_wait_request issues request if it is not yet emitted.
Also introduce i915_gem_next_request_seqno(dev) just in case we ever
need to do some prep work before using a new seqno.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
[ickle: Keep i915_gem_object_set_to_display_plane() uninterruptible.]
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: add MMIO debug output

Useful for capturing register read/write traces to send to the hw guys.

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915/sdvo: Preserve pixel-multiplier

Store the pixel-multiplier on the adjusted mode and avoid modifying the
requested mode.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: Sanity check user framebuffer parameters on creation

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915/tv: Mark the format names as constant and so avoid the memleak

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915/tv: Poll for DAC state change

Instead of sleeping for an arbitrary length of time (the documentation
fails to specify how long to wait for) wait until the load detection has
changed state (or at most the 20ms as before).

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915/tv: Preserve reserved DAC bits during mode-setting

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: Revert extra intel_wait_for_vblank to prevent stalls.

With the extra intel_wait_for_vblank added in commit
9d0498a2bf7455159b317f19531a3e5db2ecc9c4 periodic stalls were being
triggered (which were detected by i915_hangcheck_elapsed). Partially
revert this change for now.

Signed-off-by: Sitsofe Wheeler <sitsofe@yahoo.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: Ironlake page-flipping is per-plane not per-pipe

Fix a minor confusion between intel_page_flip_finish(pipe) and
intel_page_flip_finish_plane(plane) -- should have no effect as
currently we map pipe 0 to plane 0 (and pipe 1 to plane 1).

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: Addin-offset is an unreliable indicator of LVDS presence (v2)

My Samsung N210 has a VBT with DEVICE_TYPE_INT_LFP with a zero
addin-offset. With the check in place, the panel was declared absent.

v2: Only trust BIOS writers that have graduated to writing OpRegions.
(We are all doomed.)

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Zhao Yakui <yakui.zhao@intel.com>
Cc: Adam Jackson <ajax@redhat.com>
Reviewed-by: Adam Jackson <ajax@redhat.com>

drm/i915: Use the VBT from OpRegion when available (v3)

It is recommended that we use the Video BIOS tables that were copied
into the OpRegion during POST when initialising the driver. This saves
us from having to furtle around inside the ROM ourselves and possibly
allows the vBIOS to adjust the tables prior to initialisation.

On some systems, such as the Samsung N210, there is no accessible VBIOS
and the only means of finding the VBT is through the OpRegion.

v2: Rearrange the code so that ASLE is enabled along with ACPI
v3: Enable OpRegion parsing even without ACPI

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Matthew Garrett <mjg@redhat.com>

drm/i915: Rename i915_opregion.c to intel_opregion.c

It's part of the generic Intel driver infrastructure so rename it in
prepreparation for using it for VBT.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915/crt: Flush register prior to waiting for vblank.

If we don't flush the write then we can not be sure that the border
colour will have taken effect by the time we try to read it back.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: Avoid using msleep under kdb and wait_for()

wait_for() uses msleep() to yield the cpu whilst spinning waiting for a
register to change. kdb asserts that mode changes are atomic and so
prohibits msleep. The alternative would be to use mdelay or to simply
probe the register more often instead of busy waiting.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>

drm/i915: Drop the msleep parameter to wait_for()

Jesse's feedback from using the wait_for() macro was that the msleep
argument was that it was superfluous and made the macro more difficult
to use and to read. As the actually amount of time to sleep is not
critical, the crucial part is to sleep and let the processor schedule
something else whilst we wait for the event, replace the argument with a
hardcoded value.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Jesse Barnes <jbarnes@virtuousgeek.org>

drm/i915: unload: fix retire_work races

ums-gem code correctly cancels the retire work (at lastclose time),
kms does not do so. Fix this by canceling the work right after ideling
the gpu.

While staring at the code I noticed that the work function is not
static. Fix this, too.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>