Currently hsw_write_vm_values() may write to certain watermark
registers needlessly. For instance if only, say, LP3 changes,
the current code will again disable all LP1+ watermarks even
though only LP3 needs to be reconfigured.

Add an easy to read function that will compute the dirtyness of the
watermarks, and use that information to further optimize the watermark
programming.

v2: Disable LP1+ watermarks around changing LP0 watermarks for Paulo
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NPaulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

To make it easier to check what watermark updates are actually
necessary, keep copies of the relevant bits that match the current
hardware state.

Also add DDB partitioning into hsw_wm_values as that's another piece
of state we want to track.

We don't read out the hardware state on init yet, so we can't really
start using this yet, but it will be used later.
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NPaulo Zanoni <paulo.r.zanoni@intel.com>
[danvet: Paulo asked for a comment around the memcmp to say that we
depend upon zero-initializing the entire structures due to padding.
But a later patch in this series removes the memcmp again. So this is
ok as-is.]
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

The fbc_wm_enabled member in intel_wm_config is useless for the time
being. The original idea for it was that we'd pre-compute it and so
that the WM merging process could know whether it needs to worry
about FBC watermarks at all.

But we don't have a convenient way to pre-check for the possibility
of FBC being used. intel_update_fbc() should be split up for that.
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NPaulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

On HSW the LP1,LP2,LP3 levels are either 1,2,3 or 1,3,4. We make the
conversion from LPn to to the level at one point current. Later we're
going to do it in a few places, so move it to a separate function.
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NPaulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Using the 5/6 DDB split make sense only when sprites are enabled.
So check that before we waste any cycles computing the merged
watermarks with the 5/6 DDB split.
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NPaulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Move the watermark max computations into haswell_update_wm(). This
allows keeping the 1/2 vs. 5/6 split code in one place, and avoid having
to pass around so many things. We also save a bit of stack space by only
requiring one copy of struct hsw_wm_maximums.

Also move the intel_wm_config out from hsw_compute_wm_parameters() and
pass it it. We'll have some need for it in haswell_update_wm() later.
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NPaulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Let's try to keep using the intermediate intel_pipe_wm representation
for as long as possible. It avoids subtle knowledge about the
internals of the hardware registers when trying to choose the
best watermark configuration.

While at it replace the memset() w/ zero initialization.
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NPaulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

I want to convert hsw_find_best_result() to use intel_pipe_wm, so we
need to move the merging to happen outside hsw_compute_wm_results().
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NPaulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

No point in re-computing the watermarks for all pipes, when only one
pipe has changed. The watermarks stored under intel_crtc.wm.active are
still valid for the other pipes. We just need to redo the merging.

We can also skip the merge/update procedure completely if the new
watermarks for the affected pipe come out unchanged.
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NPaulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Introduce a new struct intel_pipe_wm which contains all the
watermarks for a single pipe. Use it to unify the LP0 and LP1+
watermark computations so that we can just iterate through the
watermark levels neatly and call ilk_compute_wm_level() for each.

Also add another tool ilk_wm_merge() that merges the LP1+ watermarks
from all pipes. For that, embed one intel_pipe_wm inside intel_crtc that
contains the currently valid watermarks for each pipe.

This is mainly preparatory work for pre-computing the watermarks for
each pipe and merging them at a later time. For now the merging still
happens immediately.

v2: Add some comments about level 0 DDB split and intel_wm_config
    Add WARN_ON for level 0 being disabled
    s/lp_wm/merged
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NPaulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

As we delay the initial RPS enabling (upon boot and after resume), there
is a chance that we may start to render and trigger RPS boosts before we
set up the punit. Any changes we make could result in inconsistent
hardware state, with a danger of causing undefined behaviour. However,
as the boosting is a optional tweak to RPS, we can simply ignore it
whilst RPS is not yet enabled.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: Jesse Barnes <jbarnes@virtuousgeek.org>
Reviewed-by: NJesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Let's try to avoid these confusing negated booleans.
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NJani Nikula <jani.nikula@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

The policy's max frequency is not equal to the CPU's max frequency. The
ring frequency is derived from the CPU frequency, and not the policy
frequency.

One example of how this may differ through sysfs. If the sysfs max
frequency is modified, that will be used for the max ring frequency
calculation.
(/sys/devices/system/cpu/cpu0/cpufreq/scaling_max_freq). As far as I
know, no current governor uses anything but max as the default, but in
theory, they could. Similarly distributions might set policy as part of
their init process.

It's ideal to use the real frequency because when we're currently scaled
up on the GPU. In this case we likely want to race to idle, and using a
less than max ring frequency is non-optimal for this situation.

AFAIK, this patch should have no impact on a majority of people.

This behavior hasn't been changed since it was first introduced:
commit 23b2f8bb
Author: Jesse Barnes <jbarnes@virtuousgeek.org>
Date:   Tue Jun 28 13:04:16 2011 -0700

    drm/i915: load a ring frequency scaling table v3

CC: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: NBen Widawsky <ben@bwidawsk.net>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NRodrigo Vivi <rodrigo.vivi@gmail.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

The intel_flush_primary_plane name actually tells us which plane
we're talking about.

Also reorganize the internals a bit and add a missing POSTING_READ()
to make sure the hardware has seen the changes by the time we
return from the function.
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

drm_vblank_init() is too ugly. Make it a bit easier on the eye by
collecting all the per-crtc vblank counters, timestamps etc. to
a structure and just allocate an array of those.
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NDaniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: NAlex Deucher <alexander.deucher@amd.com>
Signed-off-by: NDave Airlie <airlied@redhat.com>

After applying wait-boost we often find ourselves stuck at higher clocks
than required. The current threshold value requires the GPU to be
continuously and completely idle for 313ms before it is dropped by one
bin. Conversely, we require the GPU to be busy for an average of 90% over
a 84ms period before we upclock. So the current thresholds almost never
downclock the GPU, and respond very slowly to sudden demands for more
power. It is easy to observe that we currently lock into the wrong bin
and both underperform in benchmarks and consume more power than optimal
(just by repeating the task and measuring the different results).

An alternative approach, as discussed in the bspec, is to use a
continuous threshold for upclocking, and an average value for downclocking.
This is good for quickly detecting and reacting to state changes within a
frame, however it fails with the common throttling method of waiting
upon the outstanding frame - at least it is difficult to choose a
threshold that works well at 15,000fps and at 60fps. So continue to use
average busy/idle loads to determine frequency change.

v2: Use 3 power zones to keep frequencies low in steady-state mostly
idle (e.g. scrolling, interactive 2D drawing), and frequencies high
for demanding games. In between those end-states, we use a
fast-reclocking algorithm to converge more quickly on the desired bin.

v3: Bug fixes - make sure we reset adj after switching power zones.

v4: Tune - drop the continuous busy thresholds as it prevents us from
choosing the right frequency for glxgears style swap benchmarks. Instead
the goal is to be able to find the right clocks irrespective of the
wait-boost.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Kenneth Graunke <kenneth@whitecape.org>
Cc: Stéphane Marchesin <stephane.marchesin@gmail.com>
Cc: Owen Taylor <otaylor@redhat.com>
Cc: "Meng, Mengmeng" <mengmeng.meng@intel.com>
Cc: "Zhuang, Lena" <lena.zhuang@intel.com>
Reviewed-by: NJesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

If we encounter a situation where the CPU blocks waiting for results
from the GPU, give the GPU a kick to boost its the frequency.

This should work to reduce user interface stalls and to quickly promote
mesa to high frequencies - but the cost is that our requested frequency
stalls high (as we do not idle for long enough before rc6 to start
reducing frequencies, nor are we aggressive at down clocking an
underused GPU). However, this should be mitigated by rc6 itself powering
off the GPU when idle, and that energy use is dependent upon the workload
of the GPU in addition to its frequency (e.g. the math or sampler
functions only consume power when used). Still, this is likely to
adversely affect light workloads.

In particular, this nearly eliminates the highly noticeable wake-up lag
in animations from idle. For example, expose or workspace transitions.
(However, given the situation where we fail to downclock, our requested
frequency is almost always the maximum, except for Baytrail where we
manually downclock upon idling. This often masks the latency of
upclocking after being idle, so animations are typically smooth - at the
cost of increased power consumption.)

Stéphane raised the concern that this will punish good applications and
reward bad applications - but due to the nature of how mesa performs its
client throttling, I believe all mesa applications will be roughly
equally affected. To address this concern, and to prevent applications
like compositors from permanently boosting the RPS state, we ratelimit the
frequency of the wait-boosts each client recieves.

Unfortunately, this techinique is ineffective with Ironlake - which also
has dynamic render power states and suffers just as dramatically. For
Ironlake, the thermal/power headroom is shared with the CPU through
Intelligent Power Sharing and the intel-ips module. This leaves us with
no GPU boost frequencies available when coming out of idle, and due to
hardware limitations we cannot change the arbitration between the CPU and
GPU quickly enough to be effective.

v2: Limit each client to receiving a single boost for each active period.
    Tested by QA to only marginally increase power, and to demonstrably
    increase throughput in games. No latency measurements yet.

v3: Cater for front-buffer rendering with manual throttling.

v4: Tidy up.

v5: Sadly the compositor needs frequent boosts as it may never idle, but
due to its picking mechanism (using ReadPixels) may require frequent
waits. Those waits, along with the waits for the vrefresh swap, conspire
to keep the GPU at low frequencies despite the interactive latency. To
overcome this we ditch the one-boost-per-active-period and just ratelimit
the number of wait-boosts each client can receive.
Reported-and-tested-by: NPaul Neumann <paul104x@yahoo.de>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=68716Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Kenneth Graunke <kenneth@whitecape.org>
Cc: Stéphane Marchesin <stephane.marchesin@gmail.com>
Cc: Owen Taylor <otaylor@redhat.com>
Cc: "Meng, Mengmeng" <mengmeng.meng@intel.com>
Cc: "Zhuang, Lena" <lena.zhuang@intel.com>
Reviewed-by: NJesse Barnes <jbarnes@virtuousgeek.org>
[danvet: No extern for function prototypes in headers.]
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

This patch attempts to clean up the ring/IA scaling programming in the
following ways.
1. Fix the comment about the DDR frequency. The math is 266MHz, not
133MHz. Formula was right, docs are wrong.

2. Mask the DCLK register since I don't know how it is defined on future
platforms.

3. use mult_frac instead of magic math.

This helps for future platform enabling.

v2: Actually use the right patch. The v1 was a mix of things, none of
which was right. Note that due to rounding, we actually get different
values (slightly higher) for the effective ring frequency.

v3: Use 1.25 instead of 1.33 as the original code did. (Jesse)

CC: Jesse Barnes <jbarnes@virtuousgeek.org>
CC: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NBen Widawsky <ben@bwidawsk.net>
Reviewed-by: NJesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Signed-off-by: NDamien Lespiau <damien.lespiau@intel.com>
Reviewed-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

It indicates a probable BIOS bug, but it appears to be harmless, and
there's nothing the user can do about it anyway, so reduce to a debug
msg.  I've filed a bug with the BIOS folks about it anyway, so hopefully
they'll fix whatever GT SB read they were doing when the GT was off.

References: https://bugs.freedesktop.org/show_bug.cgi?id=69396Signed-off-by: NJesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

And add some reg defines while we're at it.  Since the units of the RC6
residency counter are actually in CZ clocks, we want to just use the
high bits or we'll overflow too frequently.
Signed-off-by: NJesse Barnes <jbarnes@virtuousgeek.org>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

struct drm_mode_display now has a separate crtc_ version of the clock to
be used when we're talking about the timings given to the harwadre (was
far as the mode is concerned).

This commit is really the result of a git grep adjusted_mode.*clock and
replacing those by adjusted_mode.crtc_clock. No functional change.

v2: Rebased on drm-intel-queued-next
Reviewed-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: NDamien Lespiau <damien.lespiau@intel.com>
Acked-by: NDave Airlie <airlied@gmail.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Done while reviewing all our allocations for fubar. Also a few errant
cases of lacking () for the sizeof operator - just a bit of OCD.

I've left out all the conversions that also should use kcalloc from
this patch  (it's only 2).
Reviewed-by: NJani Nikula <jani.nikula@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

VGA registers/memory live inside the the display power well. Add a power
domain for VGA.
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NPaulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

We increase/decrease the power well refcount in several places now, and
all of those places need to do the same thing, so pull that code into
a few small helper functions.

v2: Rename the funcs to __intel_power_well_{get,put}
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NPaulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Add APIs to get/put power well references for specific purposes.

v2: Split the i915_request change to another patch
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NPaulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Reorganize the internal i915_request power well handling to use the
reference count just like everyone else. This way all we need to do is
check the reference count and we know whether the power well needs to be
enabled of disabled.

v2: Split he intel_display_power_{get,put} change to another patch.
    Add intel_resume_power_well() to make sure we enable the power
    well on resume
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NPaulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Disabling it isn't really an option on these platforms, but having it
available for power comparisons is useful.
Signed-off-by: NJesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Rather that mess about with hdisplay/vdisplay from requested_mode, add
explicit pipe src size information to pipe config.

Now requested_mode is only really relevant for dvo/sdvo output timings.
For everything else either adjusted_mode or pipe src size should be
used.

In many places where we end up using pipe source size, we should
actually use the primary plane size, but we don't currently store
that information explicitly. As long as we treat primaries as full
screen only, we can get away with this. Eventually when we move
primaries over to drm_plane, we need to fix it all up.

v2: Add a comment to explain what pipe_src_{w,h} are
    Add a note about primary planes to commit message
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NDamien Lespiau <damien.lespiau@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Move intel_crtc_active() to intel_display.c and make it available
elsewhere as well.

intel_edp_psr_match_conditions() already has one open coded copy,
so replace that one with a call to intel_crtc_active().

v2: Copy paste a big comment from danvet's mail explaining
    when we can ditch the extra checks
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NDamien Lespiau <damien.lespiau@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

The clock in crtc->mode doesn't necessarily mean anything. Let's look
at the clock in adjusted_mode instead.
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NDamien Lespiau <damien.lespiau@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Currently most of the watermark code looks at crtc->mode which is the
user requested mode. The only piece of information there that is
relevant is hdisplay, the rest must come from adjusted_mode. Convert
all of the code to use requested_mode and adjusted_mode from
pipe config appropriately.
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NDamien Lespiau <damien.lespiau@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Check the mode flags from the adjusted_mode, not user requested mode.
The hdisplay/vdisplay check actually checkes the primary plane size,
so those still need to come from the user requested mode.

Extract both modes from pipe config instead of the drm_crtc.
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NDamien Lespiau <damien.lespiau@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Pull the expected max WM level determinations out to a separate
function. Will have another user soon.
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NPaulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Unify the code a bit to use ilk_compute_wm_level for all watermark
levels.
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NPaulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

hsw_pipe_wm_parameters and hsw_wm_maximums typically are read only. Make
them const.
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NPaulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Passing the appropriate crtc to intel_update_watermarks() should help
in avoiding needless work in the future.

v2: Avoid clash with internal 'crtc' variable in some wm functions
Reviewed-by: NPaulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Detangle the additional state of whether or not the hw has the pfit
enabled from whether it has zero size. This allows us to cleanly
distinguish in the code when we expect the pfit to be enabled (for
Haswell pc8), and when the BIOS is confused and needs sanitizing.
Reported-by: Nshui yanwei <yangweix.shui@intel.com>
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=68251Tested-by: Nshui yanwei <yangweix.shui@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

We shouldn't disable the trickle feed bits on Haswell. Our
documentation explicitly says the trickle feed bits of PRI_CTL and
CUR_CTL should not be programmed to 1, and the hardware engineer also
asked us to not program the SPR_CTL field to 1. Leaving the bits as 1
could cause underflows.
Reported-by: NArthur Runyan <arthur.j.runyan@intel.com>
Signed-off-by: NPaulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Submitting a batchbuffer which simulates a gpu
hang by doing MI_BATCH_BUFFER_START into itself,
to test hangcheck, started to hard hang the whole box
(IVB). Bisecting lead to this commit:

commit 664b422c2966cd39b8f67e8d53a566ea8c877cd6
Author: Vinit Azad <vinit.azad@intel.com>
Date:   Wed Aug 14 13:34:33 2013 -0700

    drm/i915: Only unmask required PM interrupts

Experimenting with the mask register showed that
unmasking EI UP will prevent the hard hang in IVB and SNB.
HSW doesn't hang with EI UP masked.

Considering we are just disabling interrupts that aren't even
delivered to driver, this change is more likely to paper over some
weirdness in gpu's internal state machine. But until better
explanation can be found, let's trade little bit of power
for stability on these architectures.

v2: - Unmask EI_EXPIRED directly in I915_WRITE (Vinit)
v3: - Only unmask on SNB and IVB

Cc: Vinit Azad <vinit.azad@intel.com>
Signed-off-by: NMika Kuoppala <mika.kuoppala@intel.com>
Acked-by: NVinit Azad <vinit.azad@intel.com>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>