openwrt/staging/blogic.git
12 years agodrm/i915: don't grab dev->struct_mutex for userspace forcewak
Daniel Vetter [Thu, 9 Aug 2012 20:33:58 +0000 (22:33 +0200)]
drm/i915: don't grab dev->struct_mutex for userspace forcewak

Since forcewake is now protected by a spinlock, we don't need to grab
dev->struct_mutex any more. This way we can also get rid of a stale
comment, noticed by Ben Widawsky while reviewing some locking changes.

v2: Kill the unused variable ret, noticed by Fengguang Wu.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: try harder to find WR PLL clock settings
Paulo Zanoni [Fri, 10 Aug 2012 13:03:03 +0000 (10:03 -0300)]
drm/i915: try harder to find WR PLL clock settings

If we don't find the exact refresh rate, go with the next one. This
makes some modes work for me. They won't have the best settings, but
will at least have something. Just returning from this function when
we don't find the perfect settings does not help us at all.

Version 2:
  - Remove duplicate lines on the clock table.
  - Add back debug message with refresh, p, n2 and r2.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: use the correct encoder type when comparing
Paulo Zanoni [Fri, 10 Aug 2012 13:05:10 +0000 (10:05 -0300)]
drm/i915: use the correct encoder type when comparing

  - intel_encoder->type is INTEL_OUTPUT_SOMETHING
  - drm_encoder->encoder_type is DRM_MODE_ENCODER_SOMETHING

Here we're using intel_encoder, so compare the oranges against
oranges. While at it, rename the variable to "intel_encoder" so we
keep our naming standards used everywhere.

Luckily this was not a bug because both DRM_MODE_ENCODER_DAC and
INTEL_OUTPUT_ANALOG are defined as 1. This is the only case where the
drm definition matches the intel definition.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Lazily apply the SNB+ seqno w/a
Chris Wilson [Thu, 9 Aug 2012 09:58:30 +0000 (10:58 +0100)]
drm/i915: Lazily apply the SNB+ seqno w/a

Avoid the forcewake overhead when simply retiring requests, as often the
last seen seqno is good enough to satisfy the retirment process and will
be promptly re-run in any case. Only ensure that we force the coherent
seqno read when we are explicitly waiting upon a completion event to be
sure that none go missing, and also for when we are reporting seqno
values in case of error or debugging.

This greatly reduces the load for userspace using the busy-ioctl to
track active buffers, for instance halving the CPU used by X in pushing
the pixels from a software render (flash). The effect will be even more
magnified with userptr and so providing a zero-copy upload path in that
instance, or in similar instances where X is simply compositing DRI
buffers.

v2: Reverse the polarity of the tachyon stream. Daniel suggested that
'force' was too generic for the parameter name and that 'lazy_coherency'
better encapsulated the semantics of it being an optimization and its
purpose. Also notice that gen6_get_seqno() is only used by gen6/7
chipsets and so the test for IS_GEN6 || IS_GEN7 is redundant in that
function.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: enable rc6 on ilk again
Daniel Vetter [Wed, 8 Aug 2012 21:35:40 +0000 (23:35 +0200)]
drm/i915: enable rc6 on ilk again

I have the faint hope that the total absence of any locking for the
rps code wasn't too good an idea and could very well have caused some
rc6 related regressions.

Unfortunately we've never managed to reproduce these issues on any of
our own machines, so the only way to go about this is to enable it and
see what happens.

While at it, kill some stale comments and improve the logging.

Acked-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: fix up ilk drps/ips locking
Daniel Vetter [Thu, 9 Aug 2012 14:46:01 +0000 (16:46 +0200)]
drm/i915: fix up ilk drps/ips locking

We change the drps/ips sw/hw state from different callers: Our own irq
handler, the external intel-ips module and from process context. Most
of these callers don't take any lock at all.

Protect everything by making the mchdev_lock irqsave and grabbing it in
all relevant callsites. Note that we have to convert a few sleeps in the
drps enable/disable code to delays, but alas, I'm not volunteering to
restructure the code around a few work items.

For paranoia add a spin_locked assert to ironlake_set_drps, too.

v2: Move one access inside the lock protection. Caught by the
dev_priv->ips mass-rename ...

v3: Resolve rebase conflict.

Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: DE_PCU_EVENT irq is ilk-only
Daniel Vetter [Wed, 8 Aug 2012 21:35:37 +0000 (23:35 +0200)]
drm/i915: DE_PCU_EVENT irq is ilk-only

Like all the other drps/ips stuff. Hence add the corresponding check,
give the function a preciser prefix and move the single reg clearing into
the rps handling function, too.

Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: kill dev_priv->mchdev_lock
Daniel Vetter [Wed, 8 Aug 2012 21:35:36 +0000 (23:35 +0200)]
drm/i915: kill dev_priv->mchdev_lock

It's only ever a pointer to the global mchdev_lock, and we don't use
it at all.

Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: move all rps state into dev_priv->rps
Daniel Vetter [Wed, 8 Aug 2012 21:35:35 +0000 (23:35 +0200)]
drm/i915: move all rps state into dev_priv->rps

This way it's easier so see what belongs together, and what is used
by the ilk ips code. Also add some comments that explain the locking.

Note that (cur|min|max)_delay need to be duplicated, because
they're also used by the ips code.

v2: Missed one place that the dev_priv->ips change caught ...

Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: use mutex_lock_interruptible for debugfs files
Daniel Vetter [Thu, 9 Aug 2012 13:07:02 +0000 (15:07 +0200)]
drm/i915: use mutex_lock_interruptible for debugfs files

It's no fun if your shell hangs when the driver has gone on vacation
and you want to know why ...

Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: fixup up debugfs rps state handling
Daniel Vetter [Thu, 9 Aug 2012 13:07:01 +0000 (15:07 +0200)]
drm/i915: fixup up debugfs rps state handling

- Take the dev->struct_mutex around access the corresponding state
  (and adjusting the rps hw state).
- Add an assert to gen6_set_rps to ensure we don't forget about this
  in the future.
- Don't set up the min/max_freq files if it doesn't apply to the hw.
  And do the same for the gen6+ cache sharing file while at it.

v2: Move the gen6+ checks into the read/write callbacks. Thanks to the
awesome drm midlayer we can't check that when registering the debugfs
files, because the driver is not yet fully set up, specifically the
->load callback hasn't run yet.

Oh how I despise this disaster ...

v3: Also add a WARN_ON(mutex_is_locked) in set_rps to check the
locking.

v4: Use mutex_lock_interruptible, suggested by Chris Wilson.

Reviewed-by: Ben Widawsky <ben@bwidawsk.net> (for v2)
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: properly guard ilk ips state
Daniel Vetter [Thu, 9 Aug 2012 14:44:54 +0000 (16:44 +0200)]
drm/i915: properly guard ilk ips state

The update_gfx_val function called from mark_busy wasn't taking the
mchdev_lock, as it should have. Also sprinkle a few spinlock asserts
over the code to document things better.

Things are still rather confusing, especially since a few variables
in dev_priv are used by both the gen6+ rps code and the ilk ips code.
But protected by totally different locks. Follow-on patches will clean
that up.

v2: Don't add a deadlock ... hence split up update_gfx_val into a
wrapper that grabs the lock and an internal __ variant for callsites
within intel_pm.c that already have taken the lock.

v3: Mark the internal helper as static, noticed by Ben Widawsky.

v4: Damien Lespiau had questions about the safety of the ips setup
sequence, explain in a comment why it works.

Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: add parentheses around PIXCLK_GATE definitions
Paulo Zanoni [Wed, 8 Aug 2012 17:15:32 +0000 (14:15 -0300)]
drm/i915: add parentheses around PIXCLK_GATE definitions

By looking at the current way we're using these definitions I don't
think this commit will fix any bug, but programmers from the future
are evil and will certainly find ways to combine macro expansion with
operator precedence to introduce bugs that are hard to find.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: reindent Haswell register definitions
Paulo Zanoni [Wed, 8 Aug 2012 17:15:31 +0000 (14:15 -0300)]
drm/i915: reindent Haswell register definitions

It's the only part of the i915_reg.h file that looks totally wrongly
indented, so I assume my editor config is the correct one.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: completely reset the value of DDI_FUNC_CTL
Paulo Zanoni [Wed, 8 Aug 2012 17:15:30 +0000 (14:15 -0300)]
drm/i915: completely reset the value of DDI_FUNC_CTL

Don't rely on previous values already set on the register. Everything
we're not explicitly setting should be zero for now.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: correctly set the DDI_FUNC_CTL bpc field
Paulo Zanoni [Wed, 8 Aug 2012 17:15:29 +0000 (14:15 -0300)]
drm/i915: correctly set the DDI_FUNC_CTL bpc field

Correctly erase the values previously set and also check for 6bpc and
10bpc.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: set the DDI sync polarity bits
Paulo Zanoni [Wed, 8 Aug 2012 17:15:28 +0000 (14:15 -0300)]
drm/i915: set the DDI sync polarity bits

During my tests, everything worked even if the wrong polarity was set.
Still, we should try to set the correct values.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: fix pipe DDI mode select
Paulo Zanoni [Wed, 8 Aug 2012 17:15:27 +0000 (14:15 -0300)]
drm/i915: fix pipe DDI mode select

Mask the value before changing it and also select DVI when needed.

DVI was working in cases where the BIOS was setting the correct value
because we were not masking the value before changing it.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: dump the device info
Daniel Vetter [Wed, 8 Aug 2012 20:01:51 +0000 (22:01 +0200)]
drm/i915: dump the device info

Handy for lazy people like me, or when people forget to add the output
of lspci -nn.

v2: Chris Wilson noticed that we have this duplicated already in the
i915_capabilites debugfs file. But there \n as separator looks better,
which would be a bit verbose in dmesg. Abuse the preprocessor to
extract this all.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: fixup desired rps frequency computation
Daniel Vetter [Wed, 8 Aug 2012 15:42:52 +0000 (17:42 +0200)]
drm/i915: fixup desired rps frequency computation

In commit

commit 20b46e59dd102665ce7168baa215e5b1ee66b69b
Author: Daniel Vetter <daniel.vetter@ffwll.ch>
Date:   Thu Jul 26 11:16:14 2012 +0200

    drm/i915: Only set the down rps limit when at the loweset frequency

The computation for the new desired frequency was extracted, but since
the desired frequency was passed-by value, the adjustments didn't
propgate back. Fix this.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Add I915_GEM_PARAM_HAS_SEMAPHORES
Chris Wilson [Wed, 8 Aug 2012 09:23:22 +0000 (10:23 +0100)]
drm/i915: Add I915_GEM_PARAM_HAS_SEMAPHORES

Userspace tries to estimate the cost of ring switching based on whether
the GPU and GEM supports semaphores. (If we have multiple rings and no
semaphores, userspace assumes that the cost of switching rings between
batches is exorbitant and will endeavour to keep the next batch on the
active ring - as a coarse approximation to tracking both destination and
source surfaces.) Currently userspace has to guess whether semaphores
exist based on the chipset generation and the module parameter,
i915.semaphores. This is a crude and inaccurate guess as the defaults
internally depend upon other chipset features being enabled or disabled,
nor does it extend well into the future. By exporting a HAS_SEMAPHORES
parameter, we can easily query the driver and obtain an accurate answer.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Only apply the SNB pipe control w/a to gen6
Chris Wilson [Fri, 20 Jul 2012 17:02:28 +0000 (18:02 +0100)]
drm/i915: Only apply the SNB pipe control w/a to gen6

The requirements for the sync flush to be emitted prior to the render
cache flush is only true for SandyBridge. On IvyBridge and friends we
can just emit the flushes with an inline CS stall.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: prevent possible pin leak on error path
Eugeni Dodonov [Mon, 18 Jun 2012 22:03:38 +0000 (19:03 -0300)]
drm/i915: prevent possible pin leak on error path

We should not hit this under any sane conditions, but still, this does not
looks right.

CC: Chris Wilson <chris@chris-wilson.co.uk>
CC: Daniel Vetter <daniel.vetter@ffwll.ch>
CC: stable@vger.kernel.org
Reported-by: Herton Ronaldo Krzesinski <herton.krzesinski@canonical.com>
Reviewed-by: Chris Wlison <chris@chris-wilson.co.uk>
Signed-off-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: rip out sanitize_pm again
Daniel Vetter [Thu, 26 Jul 2012 09:50:05 +0000 (11:50 +0200)]
drm/i915: rip out sanitize_pm again

We believe to have squashed all issues around the gen6+ rps interrupt
generation and why the gpu sometimes got stuck. With that cleared up,
there's no user left for the sanitize_pm infrastructure, so let's just
rip it out.

Note that 'intel_reg_write 0xa014 0x13070000' is the w/a if we find
ourselves stuck again.

Acked-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Only set the down rps limit when at the loweset frequency
Daniel Vetter [Thu, 26 Jul 2012 09:16:14 +0000 (11:16 +0200)]
drm/i915: Only set the down rps limit when at the loweset frequency

The power docs say that when the gt leaves rc6, it is in the lowest
frequency and only about 25 usec later will switch to the frequency
selected in GEN6_RPNSWREQ. If the downclock limit expires in that
window and the down limit is set to the lowest possible frequency, the
hw will not send the down interrupt. Which leads to a too high gpu
clock and wasted power.

Chris Wilson already worked on this with

commit 7b9e0ae6da0a7eaf2680a1a788f08df123724f3b
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Sat Apr 28 08:56:39 2012 +0100

    drm/i915: Always update RPS interrupts thresholds along with
    frequency

but got the logic inverted: The current code set the down limit as
long as we haven't reached it. Instead of only once with reached the
lowest frequency.

Note that we can't always set the downclock limit to 0, because
otherwise the hw will keep on bugging us with downclock request irqs
once the lowest level is reached.

For similar reasons also always set the upclock limit, otherwise the
hw might poke us again with interrupts.

v2: Chris Wilson noticed that the limit reg is also computed in
sanitize_pm. To avoid duplication, extract the code into a common
function.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Export ability of changing cache levels to userspace
Chris Wilson [Tue, 10 Jul 2012 09:27:08 +0000 (10:27 +0100)]
drm/i915: Export ability of changing cache levels to userspace

By selecting the cache level (essentially whether or not the CPU snoops
any updates to the bo, and on more recent machines whether it resides
inside the CPU's last-level-cache) a userspace driver is able to then
manage all of its memory within buffer objects, if it so desires. This
enables the userspace driver to accelerate uploads and more importantly
downloads from the GPU and to able to mix CPU and GPU rendering/activity
efficiently.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
[danvet: Added code comment about where we plan to stuff platform
specific cacheing control bits in the ioctl struct.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Segregate memory domains in the GTT using coloring
Chris Wilson [Thu, 26 Jul 2012 10:49:32 +0000 (11:49 +0100)]
drm/i915: Segregate memory domains in the GTT using coloring

Several functions of the GPU have the restriction that differing memory
domains cannot be placed next to each other (as the GPU may prefetch
beyond the end of one domain and hang as it crosses into the other
domain). We use the facility of the drm_mm to mark ranges with a
particular color that corresponds to the cache attributes of those pages
in order to prevent allocating adjacent blocks of differing memory
types.

v2: Rebase ontop of drm_mm coloring v2.
v3: Fix rebinding existing gtt_space and add a verification routine.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Expand DPF support to Haswell
Ben Widawsky [Wed, 25 Jul 2012 03:47:32 +0000 (20:47 -0700)]
drm/i915: Expand DPF support to Haswell

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Macro to determine DPF support
Ben Widawsky [Wed, 25 Jul 2012 03:47:31 +0000 (20:47 -0700)]
drm/i915: Macro to determine DPF support

Originally I had a macro specifically for DPF support, and Daniel, with
good reason asked me to change it to this. It's not the way I would have
gone (and indeed I didn't), but for now there is no distinction as all
platforms with L3 also have DPF.

Note: The good reasons are that dpf is a l3$ feature (at least on
currrent hw), hence I don't expect one to go without the other.

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
[danvet: added note]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Add contexts for HSW
Ben Widawsky [Wed, 25 Jul 2012 03:47:30 +0000 (20:47 -0700)]
drm/i915: Add contexts for HSW

Basic context support on HSW is no different than previous generations.
The size of the context object changes, but that's about it.

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Avoid concurrent access when marking the device as idle/busy
Chris Wilson [Sat, 21 Jul 2012 11:31:41 +0000 (12:31 +0100)]
drm/i915: Avoid concurrent access when marking the device as idle/busy

As suggested by Daniel, rip out the independent timers for device and
crtc busyness and integrate the manual powermanagement of the display
engine into the GEM core and its request tracking. The benefits are that
the code is a lot smaller, fewer moving parts and should fit more neatly
into the overall activity tracking of the driver.

v2: Complete overhaul and removal of the racy timers and workers.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Split i915_gem_flush_ring() into seperate invalidate/flush funcs
Chris Wilson [Fri, 20 Jul 2012 11:41:08 +0000 (12:41 +0100)]
drm/i915: Split i915_gem_flush_ring() into seperate invalidate/flush funcs

By moving the function to intel_ringbuffer and currying the appropriate
parameter, hopefully we make the callsites easier to read and
understand.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Clear the pending_gpu_fenced_access flag at the start of execbuffer
Chris Wilson [Fri, 20 Jul 2012 11:41:07 +0000 (12:41 +0100)]
drm/i915: Clear the pending_gpu_fenced_access flag at the start of execbuffer

Otherwise once we use the buffer with a BLT command on gen2/3, we will
always regard future command submissions as continuing the fenced
access. However, now that we flush/invalidate between every batch we can
drop this pessimism.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Replace the complex flushing logic with simple invalidate/flush all
Daniel Vetter [Sat, 21 Jul 2012 10:25:01 +0000 (12:25 +0200)]
drm/i915: Replace the complex flushing logic with simple invalidate/flush all

Now that we unconditionally flush and invalidate between every batch
buffer, we no longer need the complex logic to decide which domains
require flushing. Remove it and rejoice.

v2 (danvet): Keep around the flip waiting logic. It's gross and
broken, I know, but we can't just kill that thing ... even if we just
keep it around as a reminder that things are broken.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Remove the explicit flush of the GPU write domain
Chris Wilson [Fri, 20 Jul 2012 11:41:05 +0000 (12:41 +0100)]
drm/i915: Remove the explicit flush of the GPU write domain

Rely instead on the insertion of the implicit flush before the seqno
breadcrumb.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Remove explicit flush from i915_gem_object_flush_fence()
Chris Wilson [Fri, 20 Jul 2012 11:41:04 +0000 (12:41 +0100)]
drm/i915: Remove explicit flush from i915_gem_object_flush_fence()

As the flush is either performed explictly immediately after the
execbuffer dispatch, or before the serialisation of last_fenced_seqno we
can forgo the explict i915_gem_flush_ring().

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Remove the per-ring write list
Chris Wilson [Fri, 20 Jul 2012 11:41:03 +0000 (12:41 +0100)]
drm/i915: Remove the per-ring write list

This is now handled by a global flag to ensure we emit a flush before
the next serialisation point (if we failed to queue one previously).

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Remove the defunct flushing list
Chris Wilson [Fri, 20 Jul 2012 11:41:02 +0000 (12:41 +0100)]
drm/i915: Remove the defunct flushing list

As we guarantee to emit a flush before emitting the breadcrumb or
the next batchbuffer, there is no further need for the flushing list.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Replace the pending_gpu_write flag with an explicit seqno
Chris Wilson [Fri, 20 Jul 2012 11:41:01 +0000 (12:41 +0100)]
drm/i915: Replace the pending_gpu_write flag with an explicit seqno

As we always flush the GPU cache prior to emitting the breadcrumb, we no
longer have to worry about the deferred flush causing the
pending_gpu_write to be delayed. So we can instead utilize the known
last_write_seqno to hopefully minimise the wait times.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Remove assertion over write domain after i915_gem_object_sync()
Chris Wilson [Fri, 20 Jul 2012 11:41:00 +0000 (12:41 +0100)]
drm/i915: Remove assertion over write domain after i915_gem_object_sync()

As we move to lazily clearing the GPU write domain only when the buffer
becomes inactive, this leaves a window of opportunity for
i915_gem_object_pin_to_display_plane() to detect a seemingly
inconsistent value. This function is special as it tries to pipeline the
operation to avoid the stall and so may not retires the buffer and we
may not get the opportunity to clear the write domain. However, we know
all is good, so drop the assertion.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Allow late allocation of request for i915_add_request()
Chris Wilson [Fri, 20 Jul 2012 11:40:59 +0000 (12:40 +0100)]
drm/i915: Allow late allocation of request for i915_add_request()

Request preallocation was added to i915_add_request() in order to
support the overlay. However, not all users care and can quite happily
ignore the failure to allocate the request as they will simply repeat
the request in the future.

By pushing the allocation down into i915_add_request(), we can then
remove some rather ugly error handling in the callers.

v2: Nullify request->file_priv otherwise we chase a garbage pointer
when retiring requests.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: add inte_crt->adpa_reg
Daniel Vetter [Wed, 11 Jul 2012 14:27:57 +0000 (16:27 +0200)]
drm/i915: add inte_crt->adpa_reg

With the base addresses shifting around, this is easier to handle.
Also move to the real reg offset on vlv.

Acked-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: create VLV_DSIPLAY_BASE #define
Daniel Vetter [Wed, 11 Jul 2012 14:27:55 +0000 (16:27 +0200)]
drm/i915: create VLV_DSIPLAY_BASE #define

Will be used more in the next patch.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Return a mask of the active rings in the high word of busy_ioctl
Chris Wilson [Wed, 4 Jul 2012 11:25:08 +0000 (12:25 +0100)]
drm/i915: Return a mask of the active rings in the high word of busy_ioctl

The intention is to help select which engine to use for copies with
interoperating clients - such as a GL client making a request to the X
server to perform a SwapBuffers, which may require copying from the
active GL back buffer to the X front buffer.

We choose to report a mask of the active rings to future proof the
interface against any changes which may allow for the object to reside
upon multiple rings.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
[danvet: bikeshed away the write ring mask and add the explanation
Chris sent in a follow-up mail why we decided to use masks.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: add register read IOCTL
Ben Widawsky [Thu, 12 Jul 2012 18:01:05 +0000 (11:01 -0700)]
drm/i915: add register read IOCTL

The interface's immediate purpose is to do synchronous timestamp queries
as required by GL_TIMESTAMP. The GPU has a register for reading the
timestamp but because that would normally require root access through
libpciaccess, the IOCTL can provide this service instead.

Currently the implementation whitelists only the render ring timestamp
register, because that is the only thing we need to expose at this time.

v2: make size implicit based on the register offset
Add a generation check

Reviewed-by: Eric Anholt <eric@anholt.net>
Cc: Jacek Lawrynowicz <jacek.lawrynowicz@intel.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
[danvet: fixup the ioctl numerb:]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Reserve ioctl numbers for set/get_caching
Daniel Vetter [Wed, 18 Jul 2012 18:03:05 +0000 (20:03 +0200)]
drm/i915: Reserve ioctl numbers for set/get_caching

I'm planing to merge this next week for 3.7, but I'd like to avoid
stupid conflicts with the exsting userspace when merging the new
reg_read ioctl (which doesn't have userspace yet, but this caching
interface has).

Header extracted from Chris Wilson's patch, but fix up the copy&pasted
comment in the interface struct.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Support for ns2501-DVO
Thomas Richter [Wed, 18 Jul 2012 17:22:30 +0000 (19:22 +0200)]
drm/i915: Support for ns2501-DVO

This patch adds support for the ns2501 DVO, found in some older Fujitsu/Siemens Labtops.
It is in the state of "works for me".
Includes now proper DPMS support. Includes switching between resolutions -
from 640x480 to 1024x768.
Currently assumes that the native display resolution is 1024x768.

The ns2501 seems to be rather critical - if the output PLL is not
running, the chip doesn't seem to be clocked and then doesn't react
on i2c messages. Thus, a quick'n-dirty trick ensures that the DVO
is active before submitting any i2c messages to it. This is
probably to be reviewed.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=17902
Signed-off-by: Thomas Richter <thor@math.tu-berlin.de>
[danvet: fixup whitespace fail.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: add port field to struct intel_dp and use it
Paulo Zanoni [Tue, 17 Jul 2012 20:53:45 +0000 (17:53 -0300)]
drm/i915: add port field to struct intel_dp and use it

This will be needed for Haswell, but already has its uses here.

This patch started as a small patch written patch by Shobhit Kumar,
but it has changed so much that none of its original lines remain.

Credits-to: Shobhit Kumar <shobhit.kumar@intel.com>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: move common code to intel_dp_set_link_train
Paulo Zanoni [Tue, 17 Jul 2012 19:55:16 +0000 (16:55 -0300)]
drm/i915: move common code to intel_dp_set_link_train

We have some common code that we always run before calling
intel_dp_set_link_train. This common code sets the correct training
patterns to the DP variable. If we add more calls to
intel_dp_set_link_train, we'll also have to duplicate this common
code. So instead of repeating this code whenever we call
intel_dp_set_link_train, we move the code to inside the function: now
we check which training pattern we're going to set and then we set the
DP register according to it.

One of the side-effects of this change is that now we never forget to
mask the training pattern bits before changing them. It looks like
this was working before because we were first masking the bits, then
writing 00, 01 and then 11.

This patch also enables us to use the intel_dp_set_link_train function
when disabling link training: in this case we need to avoid writing
the DP_TRAINING_LANE*_SET AUX commands.

As a bonus, the big intel_dp_{start,complete}_link_train functions
will get smaller and a little bit easier to read.

Version 2 changes:
 - Rewrite commit message.
 - Also clear the training pattern bits before changing them.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: add port parameter to intel_hdmi_init
Daniel Vetter [Thu, 12 Jul 2012 18:19:59 +0000 (20:19 +0200)]
drm/i915: add port parameter to intel_hdmi_init

Instead of having a giant if cascade to figure this out according to
the passed-in register. We could do quite a bit more cleaning up and
all by using the port at more places, but I think this should be part
of a bigger rework to introduce a struct intel_digital_port which
would keep track of all these things. I guess this will be part of
some haswell-DP-induced refactoring.

For now this rips out the big cascade, which is what annoyed me so
much.

v2: Add port variable name back for the func decl (I've tried to trick
myself below the 80 char limit).

Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: simplify possible_clones computation
Daniel Vetter [Thu, 12 Jul 2012 18:08:18 +0000 (20:08 +0200)]
drm/i915: simplify possible_clones computation

Intel hw only has one MUX for encoders, so outputs are either not
cloneable or all in the same group of cloneable outputs. This neatly
simplifies the code and allows us to ditch some ugly if cascades in
the dp and hdmi init code (well, we need these if cascades for other
stuff still, but that can be taken care of in follow-up patches).

Note that this changes two things:
- dvo can now be cloned with sdvo, but dvo is gen2 whereas sdvo is
  gen3+, so no problem. Note that the old code had a bug and didn't
  allow cloning crt with dvo (but only the other way round).
- sdvo-lvds can now be cloned with sdvo-non-tv. Spec says this won't
  work, but the only reason I've found is that you can't use the
  panel-fitter (used for lvds upscaling) with anything else. But we
  don't use the panel fitter for sdvo-lvds. Imo this part of Bspec is
  a) rather confusing b) mostly as a guideline to implementors (i.e.
  explicitly stating what is already implicit from the spec, without
  always going into the details of why). So I think we can ignore this
  - worst case we'll get a bug report from a user with with sdvo-lvds
  and sdvo-tmds and have to add that special case back in.

Because sdvo lvds is a bit special explain in comments why sdvo LVDS
outputs can be cloned, but native LVDS and eDP can't be cloned - we
use the panel fitter for the later, but not for sdvo.

Note that this also uncoditionally initializes the panel_vdd work used
by eDP. Trying to be clever doesn't buy us anything (but strange bugs)
and this way we can kill the is_edp check.

v2: Incorporate review from Paulo
- Add in a missing space.
- Pimp comment message to address his concerns.

Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: group ADPA #defines together
Daniel Vetter [Wed, 11 Jul 2012 14:27:56 +0000 (16:27 +0200)]
drm/i915: group ADPA #defines together

Splitting them up between pch and gmch variants just makes it harder
to find things. Especially since the hotplug bits are actually valid
on earlier chips, too.

v2: Fixed the comment as pointed out by Paulo Zanoni.

Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Cleanup context switching through do_switch()
Chris Wilson [Sun, 15 Jul 2012 11:34:24 +0000 (12:34 +0100)]
drm/i915: Cleanup context switching through do_switch()

When bug hunting, I found the interface to do_switch() overly
complicated and I believe festered the earlier bug. This aims to make
the code a little clearer.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Move DP structs to shared location
Shobhit Kumar [Fri, 29 Jun 2012 19:03:35 +0000 (16:03 -0300)]
drm/i915: Move DP structs to shared location

Move the DP structure to shared location so that it can be used from
within the ddi module.

Changes from Paulo:
- Move less code to intel_drv.h
- Remove #include statement
- Replace a tab with a space in train_set

Signed-off-by: Shobhit Kumar <shobhit.kumar@intel.com>
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: unbreak lastclose for failed driver init
Daniel Vetter [Sat, 21 Jul 2012 14:47:09 +0000 (16:47 +0200)]
drm/i915: unbreak lastclose for failed driver init

We now refuse to load on gen6+ if kms is not enabled:

commit 26394d9251879231b85e6c8cf899fa43e75c68f1
Author: Daniel Vetter <daniel.vetter@ffwll.ch>
Date:   Mon Mar 26 21:33:18 2012 +0200

    drm/i915: refuse to load on gen6+ without kms

Which results in the drm core calling our lastclose function to clean
up the mess, but that one is neatly broken for such failure cases
since kms has been introduced in

commit 79e539453b34e35f39299a899d263b0a1f1670bd
Author: Jesse Barnes <jbarnes@virtuousgeek.org>
Date:   Fri Nov 7 14:24:08 2008 -0800

    DRM: i915: add mode setting support

Reported-and-tested-by: Paulo Zanoni <przanoni@gmail.com>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Set the context before setting up regs for the context.
Eric Anholt [Mon, 23 Jul 2012 19:33:55 +0000 (12:33 -0700)]
drm/i915: Set the context before setting up regs for the context.

Fixes failures in transform feedback on gen7 because our SOL_RESET
flag was setting the transform feedback offsets in the old context
(occasionally happened to be ours) instead of the new context.

Signed-off-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: constify mode in crtc_mode_fixup
Daniel Vetter [Fri, 20 Jul 2012 08:30:45 +0000 (10:30 +0200)]
drm/i915: constify mode in crtc_mode_fixup

Laurent Pinchart missed this when sending in is giant constify patch:

commit e811f5ae19043b2ac2c28e147a4274038e655598
Author: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Date:   Tue Jul 17 17:56:50 2012 +0200

    drm: Make the .mode_fixup() operations mode argument a const pointer

Acked-by; Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915/lvds: ditch ->prepare special case
Daniel Vetter [Wed, 11 Jul 2012 14:27:52 +0000 (16:27 +0200)]
drm/i915/lvds: ditch ->prepare special case

LVDS is the first output where dpms on/off and prepare/commit don't
perfectly match. Now the idea behind this special case seems to be
that for simple resolution changes on the LVDS we don't need to stop
the pipe, because (at least on newer chips) we can adjust the panel
fitter on the fly.

There are a few problems with the current code though:
- We still stop and restart the pipe unconditionally, because the crtc
  helper code isn't flexible enough.
- We show some ugly flickering, especially when changing crtcs (this
  the crtc helper would actually take into account, but we don't
  implement the encoder->get_crtc callback required to make this work
  properly).

So it doesn't even work as advertised. I agree that it would be nice
to do resolution changes on LVDS (and also eDP) whithout blacking the
screen where the panel fitter allows to do that. But imo we should
implement this as a special case a few layers up in the mode set code,
akin to how we already detect simple framebuffer changes (and only
update the required registers with ->mode_set_base).

Until this is all in place, make our lives easier and just rip it out.

Also note that this seems to fix actual bugs with enabling the lvds
output, see:

http://lists.freedesktop.org/archives/intel-gfx/2012-July/018614.html

Cc: Takashi Iwai <tiwai@suse.de>
Cc: Giacomo Comes <comes@naic.edu>
Acked-by: Chris Wilson <chris@chris-wilson.co.uk>
Tested-by: Takashi Iwai <tiwai@suse.de>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: dereferencing an error pointer
Dan Carpenter [Tue, 17 Jul 2012 06:44:49 +0000 (09:44 +0300)]
drm/i915: dereferencing an error pointer

We need to check that "ctx" is a valid pointer before dereferencing it.

Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: fix invalid reference handling of the default ctx obj
Chris Wilson [Sun, 15 Jul 2012 11:34:23 +0000 (12:34 +0100)]
drm/i915: fix invalid reference handling of the default ctx obj

Otherwise we end up trying to unpin a freed object and BUG.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Add -EIO to the list of known errors for __wait_seqno
Chris Wilson [Mon, 16 Jul 2012 12:05:34 +0000 (13:05 +0100)]
drm/i915: Add -EIO to the list of known errors for __wait_seqno

This prevents a WARN introduced with

  commit de2b998552c1534e87bfbc51ec5734b02bc89020
  Author: Daniel Vetter <daniel.vetter@ffwll.ch>
  Date:   Wed Jul 4 22:52:50 2012 +0200

      drm/i915: don't return a spurious -EIO from intel_ring_begin

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Flush the context object from the CPU caches upon switching
Chris Wilson [Sun, 15 Jul 2012 11:34:22 +0000 (12:34 +0100)]
drm/i915: Flush the context object from the CPU caches upon switching

The issue is that we stale data in the CPU caches, when we come to
swap-out the object, the CPU may short-circuit the reads from those
cacheline and so corrupt the context object.

Secondary, leaving the context object as being marked in the CPU write
domain whilst on the GPU active list is a bad idea and will throw
warnings later.

Note: Thanks to calling set_to_gtt_domain with write = false and not
setting any gpu write domain when putting a context object onto the
active list (when we switch away from it) the set_to_gtt_domain call
won't block.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
[danvet: Added a note to the commit message and a comment in the code
to explain the clever non-blocking trick.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Make the lock for pageflips interruptible
Chris Wilson [Wed, 23 May 2012 10:13:58 +0000 (11:13 +0100)]
drm/i915: Make the lock for pageflips interruptible

As we take the struct_mutex lock to access the command-stream, there is
a possibility that we may need to wait for a GPU hang and so should make
the lock both interruptible and error-checking.

References: https://bugs.freedesktop.org/show_bug.cgi?id=50069
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: don't forget the PCH backlight registers
Paulo Zanoni [Sat, 14 Jul 2012 14:57:12 +0000 (11:57 -0300)]
drm/i915: don't forget the PCH backlight registers

When we enable/disable the CPU backlight registers we can't forget to
enable/disable the PCH backlight registers. Since we're using the CPU
registers we should also unset the override bit.

Fixes a regression on the following commit:
  drm/i915: properly enable the blc controller on the right pipe

The commit just deleted the code that sets the PCH registers, so it
was relying on the values set by the BIOS. I told my BIOS to boot on
the DVI monitor instead of the LVDS panel, so I noticed the bug.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Insert a flush between batches if the breadcrumb was dropped
Chris Wilson [Fri, 13 Jul 2012 13:14:08 +0000 (14:14 +0100)]
drm/i915: Insert a flush between batches if the breadcrumb was dropped

If we drop the breadcrumb request after a batch due to a signal for
example we aim to fix it up at the next opportunity. In this case we
emit a second batchbuffer with no waits upon the first and so no
opportunity to insert the missing request, so we need to emit the
missing flush for coherency. (Note that that invalidating the render
cache is the same as flushing it, so there should have been no
observable corruption.)

Note that beside simply adding the missing flush, avoiding potential
render corruption, this will also fix at least parts of the problem
introduced by some funny interaction of these two commits:

commit de2b998552c1534e87bfbc51ec5734b02bc89020
Author: Daniel Vetter <daniel.vetter@ffwll.ch>
Date:   Wed Jul 4 22:52:50 2012 +0200

    drm/i915: don't return a spurious -EIO from intel_ring_begin

which allowed intel_ring_begin to return -ERESTARTSYS and

commit cc889e0f6ce6a63c62db17d702ecfed86d58083f
Author: Daniel Vetter <daniel.vetter@ffwll.ch>
Date:   Wed Jun 13 20:45:19 2012 +0200

    drm/i915: disable flushing_list/gpu_write_list

which essentially disabled the flushing list.

The issue happens when we submit a batch & emit it, but get
interrupted (thanks to the first patch) while trying to emit the
flush. On the next batch we still assume that the full gpu domain
handling is in effect and hence compute the invalidate&flushing
domains. But thanks to the 2nd patch we totally ignore these and only
invalidate all gpu domains, presuming that any required flushes have
been issued already.  Which is wrong and eventually results in us
updating the new write_domain values with the computed
pending_write_domain values, which leaves an object with write_domain
== 0 on the gpu_write_list.

As soon as we try to unbind that object, things blow up.

Fix this by emitting the missing flush according to the new
ring->gpu_caches_dirty flag.

Note that this does _not_ fix all the current cases where we end up
with an object on the flushing_list that can't be flushed.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=52040
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
[danvet: Add bug explanation to commit message.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: missing error case in init status page
Ben Widawsky [Fri, 13 Jul 2012 06:16:12 +0000 (23:16 -0700)]
drm/i915: missing error case in init status page

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: mask tiled bit when updating ILK sprites
Ander Conselvan de Oliveira [Fri, 13 Jul 2012 12:50:33 +0000 (15:50 +0300)]
drm/i915: mask tiled bit when updating ILK sprites

Or going from tiled to untiled may break.

Signed-off-by: Ander Conselvan de Oliveira <ander.conselvan.de.oliveira@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: try to train DP even harder
Paulo Zanoni [Fri, 29 Jun 2012 19:03:34 +0000 (16:03 -0300)]
drm/i915: try to train DP even harder

While debugging Haswell link train failures I observed that we never
try the maximum voltage configuration more than once consecutively. We
start the training, the monitor keeps telling us to increase the
voltage, then when we reach the maximum we just go back to the start
(because of the "memset" above "voltage_tries = 0"). When we reach
this point, we keep alternating between the maximum and the minimum
voltages until we give up.

The DP spec suggests that we should try the same voltage 5 times
before giving up. This patch makes us try the maximum voltage at
least 5 times before going back to the minimum voltages.

This patch does not fix any particular bug I'm aware of.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: kill intel_ddc_probe
Daniel Vetter [Wed, 11 Jul 2012 10:31:53 +0000 (12:31 +0200)]
drm/i915: kill intel_ddc_probe

We have way too much lying hardware to rely on a simple "does someone
answer on the ddc i2c address?" check. And now it's unused, so just
kill it.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: check whether we actually received an edid in detect_ddc
Daniel Vetter [Wed, 11 Jul 2012 10:31:52 +0000 (12:31 +0200)]
drm/i915: check whether we actually received an edid in detect_ddc

Somehow detect_ddc manages to fall through all checks when we think
that something responds on the ddc i2c address, but the edid read
failed. Fix this up by explicitly checking for this case.

This fixes a regression on newer chips because since

commit aaa377302b2994fcc2c66741b47da33feb489dca
Author: Daniel Vetter <daniel.vetter@ffwll.ch>
Date:   Sat Jun 16 15:30:32 2012 +0200

    drm/i915/crt: Do not rely upon the HPD presence pin

we use ddc detection also on hotplug capable platforms. And one of
these reads all 0s for any i2c transaction if nothing is connected to
the vga port.

v2: Implement Chris Wilson's review:
- simplify logic, default to "nothing detected"
- kill stale comment
- BUG_ON(!crt->type != ANALOG)

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=51900
Tested-by: Yang Guang <guang.a.yang@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: fix up PCH backlight #define mixup
Daniel Vetter [Tue, 10 Jul 2012 22:31:06 +0000 (00:31 +0200)]
drm/i915: fix up PCH backlight #define mixup

I so totally suck.

This can cause a black screen if (for whatever reason) the bios
hasn't set this bit itself.

This regression has been introduced in

commit 7cf4160148136deb31ee5f2802857dd935a38529
Author: Daniel Vetter <daniel.vetter@ffwll.ch>
Date:   Tue Jun 5 10:07:09 2012 +0200

    drm/i915: clear up backlight #define confusion on gen4+

Tested-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Add comments to explain the BSD tail write workaround
Chris Wilson [Thu, 5 Jul 2012 16:14:01 +0000 (17:14 +0100)]
drm/i915: Add comments to explain the BSD tail write workaround

Having had to dive into the bspec to understand what each stage of the
workaround meant, and how that the ring broadcasting IDLE corresponded
with the GT powering down the ring (i.e. rc6) add comments to aide
the next reader.

And since the register "is used to control all aspects of PSMI and power
saving functions" that makes it quite interesting to inspect with
regards to RC6 hangs, so add it to the error-state.

v2: Rediscover the piece of magic, set the RNCID to 0 before waiting for
the ring to wake up.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Disable the BLT on pre-production SNB hardware
Chris Wilson [Thu, 5 Jul 2012 22:49:40 +0000 (23:49 +0100)]
drm/i915: Disable the BLT on pre-production SNB hardware

It never quite worked despite the numerous workarounds, yet I still see
people trying to use this hardware and filing bug reports. As we no
longer even try to implement the workarounds, since 6a233c78878
(drm/i915/ringbuffer: kill snb blt workaround), simply disable the ring.

v2: Add a message to inform the user about the limited capabilities of
their pre-production hardware.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: initialize power wells in modeset_init_hw
Eugeni Dodonov [Fri, 6 Jul 2012 18:42:36 +0000 (15:42 -0300)]
drm/i915: initialize power wells in modeset_init_hw

This initializes power wells within the modeset_init_hw routine.
Testing has shown that this works for both driver load time and for
suspend-resume code paths.

Signed-off-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915: Only request PM interrupts for the events we handled
Chris Wilson [Thu, 5 Jul 2012 14:02:17 +0000 (15:02 +0100)]
drm/i915: Only request PM interrupts for the events we handled

There is little point waking up every 10ms to service an interrupt which
we then promptly ignore. So only program the the PMIER to enable
interrupts for those events which we do handle, not all of them!

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Jesse Barnes <jbarnes@virtuousgeek.org>
Cc: Eugeni Dodonov <eugeni.dodonov@intel.com>
Cc: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915/context: Add missing IVB context sizes
Ben Widawsky [Wed, 18 Jul 2012 17:10:10 +0000 (10:10 -0700)]
drm/i915/context: Add missing IVB context sizes

There were some fields missed. Daniel pointed this out in review, and I
know I fixed it, but something happened somehow and some time.

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/i915/context/: s/CTX/CXT
Ben Widawsky [Wed, 18 Jul 2012 17:10:09 +0000 (10:10 -0700)]
drm/i915/context/: s/CTX/CXT

*sigh* the docs had it spelled wrong, corrected it, and then proceeded
to re-do the original error. The original code preserved this history,
and this patch attempts to keep in sync with the current docs.

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
12 years agodrm/sis: fixup sis_mm ioctl structs
Daniel Vetter [Sun, 24 Jun 2012 17:57:24 +0000 (19:57 +0200)]
drm/sis: fixup sis_mm ioctl structs

Userspace uses long in quite a few places more than the kernel. Which
gives me neat proof that I'm the only guy on this side of the galaxy
who ever tried to run glxgears on a 64bit machine with sis graphics on
linux.

Note that the longs in drm_sis_mem_t aren't aligned properly, so this
won't even work with 32bit userspace on 64bit kernel as-is. Hence the
patch can't break that, either.

Nope, I'm not nuts enough to write the 32bit ioctl compat layer for
this and test it with some wine app. Even though hunting the ebay
dungeons for a sis card actually supported by the mesa drivers casts
some doubts on this ...

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agodrm: kill i915/i830 ids from drm_pciids.h
Daniel Vetter [Tue, 25 Oct 2011 23:03:05 +0000 (01:03 +0200)]
drm: kill i915/i830 ids from drm_pciids.h

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agodrm: unconditionally clean up dma buffers of closing clients
Daniel Vetter [Tue, 25 Oct 2011 22:31:26 +0000 (00:31 +0200)]
drm: unconditionally clean up dma buffers of closing clients

With the last patch to ditch DMA_QUEUE support, we should be able
to call the dma cleanup uncoditionally, even when the master has
disappeared.

Do so because it just makes more sense.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agodrm: kill dma queue support
Daniel Vetter [Tue, 25 Oct 2011 22:54:41 +0000 (00:54 +0200)]
drm: kill dma queue support

Absolutely unused. All the values are only ever initialized and
then used at most in some debug printout functions.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agodrm: ditch strange DRIVER_DMA_QUEUE only error bail-out
Daniel Vetter [Tue, 25 Oct 2011 22:53:57 +0000 (00:53 +0200)]
drm: ditch strange DRIVER_DMA_QUEUE only error bail-out

Only one driver (i810) even sets that flag. Now the actual locking
code uncoditionally promotes lock->context to an unsigned int.

Closer inspection of the userspace reveals that the drm lock context
is defined as an unsigned int (at least on linux). I suspect we just
have a strange case of signedness confusion going on.

Tested on my i815, doesn't seem to break anything.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agodrm: kill reclaim_buffers callback
Daniel Vetter [Tue, 25 Oct 2011 22:20:57 +0000 (00:20 +0200)]
drm: kill reclaim_buffers callback

All leftover users either haven't set DRIVER_HAVE_DMA, in which
case this will never be called, or use the drm_core implementation.

Call that directly in the only callsite.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agodrm/savage: clean up reclaim_buffers
Daniel Vetter [Tue, 25 Oct 2011 22:14:15 +0000 (00:14 +0200)]
drm/savage: clean up reclaim_buffers

The reclaim_buffers function of the savage driver actually wants to run
with the hw_lock held - at least there are printks in the call-chain
to that effect. But the drm core only calls reclaim_buffers as used
by savage _after_ forcefully dropping the hwlock (in case it's still
hold by the closing fd).

So do the same idlelock dance as for the other dma drivers and hope
that papers over any issues.

v2: Don't let the idlelock linger around.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Tested-by: Tormod Volden <debian.tormod@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agodrm: kill reclaim_buffers_locked
Daniel Vetter [Tue, 25 Oct 2011 21:57:28 +0000 (23:57 +0200)]
drm: kill reclaim_buffers_locked

i810 was the last user of this code, with that gone, kill it.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agoRevert "Revert "drm/i810: cleanup reclaim_buffers""
Daniel Vetter [Mon, 11 Jun 2012 19:50:23 +0000 (21:50 +0200)]
Revert "Revert "drm/i810: cleanup reclaim_buffers""

This reverts commit 6e877b576ddf7cde5db2e9a6dcb56fef0ea77e64,
reinstating the original commit:

commit 87499ffdcb1c70f66988cd8febc4ead0ba2f9118
Author: Daniel Vetter <daniel.vetter@ffwll.ch>
Date:   Tue Oct 25 23:51:24 2011 +0200

    drm/i810: cleanup reclaim_buffers

    My dear old i815 always hits the deadlocked on reclaim_buffers
    warning. Switch over to the idlelock duct-tape on hope that
    works better. I've fired up my i815 and now closing glxgears doesn't
    take 5 seconds anymore. \o/

The original problem with that was that I've moved it ahead in the
series so that it could be included despite some patches not being
ready quite yet. The little problem is that this patch required some
of the previous rework to work correctly.

Now that everything is in the right order again, this actually works
on my i810 and does speed up closing gl apps as the original commit
claimed. Without hanging the machine, as the revert says.

Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agodrm: kill reclaim_buffers_idlelocked functions
Daniel Vetter [Tue, 25 Oct 2011 21:42:29 +0000 (23:42 +0200)]
drm: kill reclaim_buffers_idlelocked functions

The only two users are now folded into the drivers preclose functions,
so this is unused.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agodrm/sis: clean up reclaim_buffers
Daniel Vetter [Tue, 25 Oct 2011 21:42:59 +0000 (23:42 +0200)]
drm/sis: clean up reclaim_buffers

Like for via.

v2: Actually drop the idlelock again if taken.

v3: Fixup.

v4: Fixup the "has master" vs. "is master" confusion the refactor
introduced.

v5: Drop the idlelock in the early return path.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agodrm/via: clean up reclaim_buffers
Daniel Vetter [Tue, 25 Oct 2011 21:37:09 +0000 (23:37 +0200)]
drm/via: clean up reclaim_buffers

A few things
- kill reclaim_buffers, it's never ever called because via does not set
  DRIVER_HAVE_DMA
- inline the idlelock dance into the buffer reclaim logic and make it
  a simple preclose cleanup function
- directly call the the dma_quiescent function and kill the needless
  if check.

v2: Actually drop the idlelock when we take it. Reported by James
Simmons.

v3: Rebased onto latest drm-next.

v4: Fixup the refactor.

v5: More fixup the refactor - I've accidentally changed the check for
any master to checking whether the closing fd is the master.

v6: Don't forget to drop the idlelock in the early return path, too.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agodrm/udl: port over blanking code from udlfb.
Dave Airlie [Tue, 26 Jun 2012 13:53:07 +0000 (14:53 +0100)]
drm/udl: port over blanking code from udlfb.

This ports over the dpms code from udlfb, and should mean
a better chance of turning on some udl devices.

Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agodrm/radeon/kms: auto detect pcie link speed from root port
Dave Airlie [Wed, 27 Jun 2012 07:35:54 +0000 (08:35 +0100)]
drm/radeon/kms: auto detect pcie link speed from root port

This check the root ports supported link speeds and enables
GEN2 mode if the 5.0 GT link speed is available.

The first 3.0 cards are SI so they will probably need more investigation.

Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agodrm/pci: add support for getting the supported link bw.
Dave Airlie [Wed, 27 Jun 2012 07:35:53 +0000 (08:35 +0100)]
drm/pci: add support for getting the supported link bw.

This should work for PCIE3.0 as well.

Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agopci_regs: define LNKSTA2 pcie cap + bits.
Dave Airlie [Wed, 27 Jun 2012 07:35:52 +0000 (08:35 +0100)]
pci_regs: define LNKSTA2 pcie cap + bits.

We need these for detecting the max link speed for drm drivers.

Acked-by: Bjorn Helgaas <bhelgass@google.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agodrm/radeon: improve GPU lockup debugging info on r6xx/r7xx/r8xx/r9xx
Jerome Glisse [Wed, 27 Jun 2012 16:25:01 +0000 (12:25 -0400)]
drm/radeon: improve GPU lockup debugging info on r6xx/r7xx/r8xx/r9xx

Print various CP register that have valuable informations regarding
GPU lockup.

Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agodrm/mgag200: fix null pointer dereference
Devendra Naga [Sat, 7 Jul 2012 09:22:15 +0000 (09:22 +0000)]
drm/mgag200: fix null pointer dereference

we are referencing the pointer after doing alloc_apertures,
as alloc_apertures kzallocs, the kzalloc may fail and we get a NULL.

so we need to check for NULL before we dereference this pointer

Signed-off-by: Devendra Naga <devendra.aaru@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agodrm/radeon: Try harder to avoid HW cursor ending on a multiple of 128 columns.
Michel Dänzer [Tue, 17 Jul 2012 17:02:09 +0000 (19:02 +0200)]
drm/radeon: Try harder to avoid HW cursor ending on a multiple of 128 columns.

This could previously fail if either of the enabled displays was using a
horizontal resolution that is a multiple of 128, and only the leftmost column
of the cursor was (supposed to be) visible at the right edge of that display.

The solution is to move the cursor one pixel to the left in that case.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=33183

Cc: stable@vger.kernel.org
Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agodrm: Make the .mode_fixup() operations mode argument a const pointer
Laurent Pinchart [Tue, 17 Jul 2012 15:56:50 +0000 (17:56 +0200)]
drm: Make the .mode_fixup() operations mode argument a const pointer

The passed mode must not be modified by the operation, make it const.

Signed-off-by: Laurent Pinchart <laurent.pinchart@ideasonboard.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agodrm: remove the list_head from drm_mode_set
Daniel Vetter [Wed, 11 Jul 2012 14:28:08 +0000 (16:28 +0200)]
drm: remove the list_head from drm_mode_set

It's unused. At it confused me quite a bit until I've discovered that.

Cc: dri-devel@lists.freedesktop.org
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agodrm/fb helper: don't call drm_crtc_helper_set_config
Daniel Vetter [Wed, 11 Jul 2012 14:28:07 +0000 (16:28 +0200)]
drm/fb helper: don't call drm_crtc_helper_set_config

Go through the interface vtable instead, because not everyone might be
using the crtc helper code.

Cc: dri-devel@lists.freedesktop.org
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>
12 years agodrm/fb-helper: delay hotplug handling when partially bound
Daniel Vetter [Fri, 15 Jun 2012 09:01:22 +0000 (11:01 +0200)]
drm/fb-helper: delay hotplug handling when partially bound

Ok, this requires quite a dance to actually hit:
1) We plug in a 2nd screen, enable it in both X and (by vt-switching)
in the fbcon.
2) We disable that screen again in with xrandr.
3) We vt-switch again, so that fbcon displays on the 2nd screen, but X
on the first screen. This obviously needs a driver that doesn't switch
off unused functions when regaining the VT.
3) When X controls the vt, we unplug that screen.

Now drm_fb_helper_hotplug_event we noticed that that some crtcs are
bound, but because we still have the fbcon on the 2nd screeen we also
have bound set. Which means the fbcon wrongly assumes it's in control
of everything an happily disables the output on the 2nd screen, but
enables its fb on the first screen.

Work around this issue by counting how many crtcs are bound and how
many are bound to fbcon and assuming that when fbcon isn't bound to
all of them, it better not touch the output configuration.

Conceptually this is the same as only restoring the fbcon output
configuration on the driver's ->lastclose, when we're sure that no one
else is using kms. So this should be consistent with existing kms
drivers.

Chris has created a separate patch for the intel ddx, but I think we
should fix this issue here regardless - the fbcon messing with the
output config while it's not fully in control simply isn't a too
polite behaviour.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=50772
Tested-by: Maxim A. Nikulin <M.A.Nikulin@gmail.com>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Dave Airlie <airlied@redhat.com>