GitHub/LineageOS/android_kernel_motorola_exynos9610.git
10 years agodrm/i915: sanitize PUNIT register macro definitions
Imre Deak [Tue, 4 Mar 2014 17:23:02 +0000 (19:23 +0200)]
drm/i915: sanitize PUNIT register macro definitions

In the upcoming patches we'll need to access the rest of the fields in
the punit power gating register, so prepare for that.

v2:
- add doc reference for the power well subsystem IDs (Jesse)
- remove IDs for non-existant DPIO_RX[23] subsystems (Jesse)

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: vlv: keep first level vblank IRQs masked
Imre Deak [Tue, 4 Mar 2014 17:23:01 +0000 (19:23 +0200)]
drm/i915: vlv: keep first level vblank IRQs masked

This is a left-over from

commit b7e634cc8dcd320123199a18bae0937b40dc28b8
Author: Imre Deak <imre.deak@intel.com>
Date:   Tue Feb 4 21:35:45 2014 +0200

drm/i915: vlv: don't unmask IIR[DISPLAY_PIPE_A/B_VBLANK] interrupt

where we stopped unmasking the vblank IRQs, but left them enabled in the
IER register. Disable them in IER too.

v2:
- remove comment becoming stale after this change (Ville)

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: check pipe power domain when reading its hw state
Imre Deak [Wed, 5 Mar 2014 14:20:55 +0000 (16:20 +0200)]
drm/i915: check pipe power domain when reading its hw state

We can read out the pipe HW state only if the required power domain is
on. If not we consider the pipe to be off.

v2:
- no change
v3:
- push down the power domain checks into the specific crtc
  get_pipe_config handlers (Daniel)

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
[danvet: Appease checkpatch.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: check port power domain when reading the encoder hw state
Imre Deak [Wed, 5 Mar 2014 14:20:54 +0000 (16:20 +0200)]
drm/i915: check port power domain when reading the encoder hw state

Since the encoder is tied to its port, we need to make sure the power
domain for that port is on before reading out the encoder HW state.

Note that this also covers also all connector get_hw_state handlers,
since all those just call the corresponding encoder get_hw_state
handler, which checks - after this change - for all power domains
the connector needs.

v2:
- no change
v3:
- push down the power domain checks into the specific encoder
  get_hw_state handlers (Daniel)

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: get port power domain in connector detect handlers
Imre Deak [Wed, 5 Mar 2014 14:20:53 +0000 (16:20 +0200)]
drm/i915: get port power domain in connector detect handlers

The connector detect and get_mode handlers need to access the port
specific HW blocks to read the EDID etc. Get/put the port power domains
around these handlers.

v2:
- get port power domain for HDMI too (Ville)
- get port power domain for the DP,HDMI audio detect handlers (Jesse)
- Leave the intel_runtime_pm_get/put in the DP detect function in place.
  Instead of just removing them, these should be moved to the appropriate
  power_well enable/disable handlers. We can do this after Paulo's
  'Merge PC8 with runtime PM, v2' patchset.
v3:
- rebased on latest -nightly

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: add port power domains
Imre Deak [Tue, 4 Mar 2014 17:22:57 +0000 (19:22 +0200)]
drm/i915: add port power domains

Parts that poke port specific HW blocks like the encoder HW state
readout or connector hotplug detect code need a way to check whether
required power domains are on or enable/disable these. For this purpose
add a set of power domains that refer to the port HW blocks. Get the
proper port power domains during modeset.

For now when requesting the power domain for a DDI port get it for a 4
lane configuration. This can be optimized later to request only the 2
lane power domain, when proper support is added on the VLV PHY side for
this. Atm, the PHY setup code assumes a 4 lane config in all cases.

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: add noop power well handlers instead of NULL checking them
Imre Deak [Tue, 4 Mar 2014 17:22:56 +0000 (19:22 +0200)]
drm/i915: add noop power well handlers instead of NULL checking them

Reading code free of special cases wins over the small overhead of
calling a noop handler. Suggested by Jesse.

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: split power well 'set' handler to separate enable/disable/sync_hw
Imre Deak [Tue, 4 Mar 2014 17:22:55 +0000 (19:22 +0200)]
drm/i915: split power well 'set' handler to separate enable/disable/sync_hw

Split the 'set' power well handler into an 'enable', 'disable' and
'sync_hw' handler. This maps more conveniently to higher level
operations, for example it allows us to push the hsw package c8 handling
into the corresponding hsw/bdw enable/disable handlers and the hsw BIOS
hand-over setting into the hsw/bdw sync_hw handler.

No functional change.

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
[danvet: Appease checkpatch's whitespace complaints.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: add init power domain to always-on power wells
Imre Deak [Tue, 4 Mar 2014 17:22:54 +0000 (19:22 +0200)]
drm/i915: add init power domain to always-on power wells

Whenever we request a power domain it has to guarantee that all HW
resources are enabled that are needed to access a HW register associated
with that power domain. In case a register is on an always-on power well
this won't result in turning on a power well, but it may require
enabling some other HW resource. One such resource is the HSW/BDW device
D0 state that is required for all register accesses and thus for all
power wells/power domains.

So far the init power domain (guaranteeing access to all HW registers)
was part of the default i9xx always-on power well, but not the HSW/BDW
always-on power wells. Add the domain to the latter power wells too.

Atm, all the always-on power wells have noop handlers, so this doesn't
change the functionality.

v2:
- clarify semantics of always-on power wells (Paulo)

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: move power domain macros to intel_pm.c
Imre Deak [Tue, 4 Mar 2014 17:22:53 +0000 (19:22 +0200)]
drm/i915: move power domain macros to intel_pm.c

These macros are used only locally, so move them to the .c file.

No functional change.

v2:
- add init power domain to always-on power wells in the following
  - separate - patch (Paulo)

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Disable full ppgtt by default
Daniel Vetter [Thu, 6 Mar 2014 08:40:43 +0000 (09:40 +0100)]
drm/i915: Disable full ppgtt by default

There are too many oustanding issues:

- Fence handling in the current code is broken. There's a patch series
  from me, but it's blocked on and extended review (which includes
  writing the testcases).

- IOMMU mapping handling is broken, we need to properly refcount it -
  currently it gets destroyed when the first vma is unbound, so way
  too early.

- There's a pending reset issue on snb. Since Mika's reset work and
  full ppgtt have been pulled in in separate branches and ended up
  intermittingly breaking each another it's unclear who's the exact
  culprit here.

- We still have persistent evidince of crazy recursion bugs through
  vma_unbind and ppgtt_relase, e.g.

  https://bugs.freedesktop.org/show_bug.cgi?id=73383

  This issue (and a few others meanwhile resolved) have blocked our
  performance measuring/tuning group since 3 months.

- Secure batch dispatching is broken. This is blocking Brad Volkin's
  command checker work since 3 months.

All these issues are confirmed to only happen when full ppgtt is
enabled, falling back to aliasing ppgtt resolves them. But even
aliasing ppgtt itself still has a regression:

- We currently unconditionally bind objects into the aliasing ppgtt,
  which means all priviledged objects like ringbuffers are visible to
  unpriviledged access again. On top of that this also breaks the
  command checker for aliasing ppgtt, since it can't hide the
  validated batch any more.

Furthermore topic/full-ppgtt has never been reviewed:

- Lifetime rules around vma unbinding/release are unclear, resulting
  into this awesome hack called ppgtt_release. Which seems to take the
  blame for most of the recursion fallout.

- Context/ring init works different on gpu reset than anywhere else.
  Such differeneces have in the past always lead to really hard to
  track down bugs.

- Aliasing ppgtt is treated in a bunch of places as a real address
  space, but it isn't - the real address space is always the global
  gtt in that case. This results in a bit a mess between contexts and
  ppgtt object, further complication the context/ppgtt/vma lifetime
  rules.

- We don't have any docs describing the overall concepts introduced
  with full ppgtt. A short, concise overview describing vmas and some
  of the strange bits around them (like the unbound vmas used by
  execbuf, or the new binding rules) really is needed.

Note that a lot of the post topic/full-ppgtt merge fallout has already
been addressed, this entire list here of 10 issues really only contains
the still outstanding issues.

Finally the 3.15 merge window is approaching and I think we need to
use the remaining time to ensure that our fallback option of using
aliasing ppgtt is in solid shape. Hence I think it's time to throw the
switch. While at it demote the helper from static inline status
because really.

Cc: Ben Widawsky <ben@bwidawsk.net>
Cc: Dave Airlie <airlied@gmail.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: move modeset_update_power_wells earlier
Imre Deak [Wed, 5 Mar 2014 14:20:52 +0000 (16:20 +0200)]
drm/i915: move modeset_update_power_wells earlier

These functions will be needed by the valleyview specific power well
update functionality added in an upcoming patch, so move them earlier.

No functional change.

v2:
- no change
v3:
- rebase on latest -nightly

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org> (v2)
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: fold in __intel_power_well_get/put functions
Imre Deak [Tue, 4 Mar 2014 17:22:51 +0000 (19:22 +0200)]
drm/i915: fold in __intel_power_well_get/put functions

These functions are used only by a single call site and are simple
enough to just fold them in.

Note that in later patches the parts folded in here are further
simplified as we'll remove hsw_{disable,enable}_package_c8 and the NULL
check of the power well enable/disable handlers. All this means that at
the end intel_display_power_get/put() becomes more understandable as we
don't need to jump between two functions when reading the code.

No functional change.

v2:
- clarify the rational for the change (Chris)

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Unify CHICKEN_PIPESL_1 register definitions
Ville Syrjälä [Wed, 5 Mar 2014 11:05:47 +0000 (13:05 +0200)]
drm/i915: Unify CHICKEN_PIPESL_1 register definitions

We have two names for the same register CHICKEN_PIPESL_1 and
HSW_PIPE_SLICE_CHICKEN_1. Unify it to just one.

Also rename the FBCQ disable bit to resemble the name we've
given to a similar bit on earlier platforms.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Use RMW to update chicken bits in gen7_enable_fbc()
Ville Syrjälä [Wed, 5 Mar 2014 11:05:46 +0000 (13:05 +0200)]
drm/i915: Use RMW to update chicken bits in gen7_enable_fbc()

gen7_enable_fbc() may write to some registers which we've already
touched, so use RMW so that we don't undo any previous updates.

Also note that we implemnt WaFbcAsynchFlipDisableFbcQueue:bdw.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Don't clobber CHICKEN_PIPESL_1 on BDW
Ville Syrjälä [Wed, 5 Mar 2014 11:05:45 +0000 (13:05 +0200)]
drm/i915: Don't clobber CHICKEN_PIPESL_1 on BDW

Misplaced parens cause us to totally clobber the CHICKEN_PIPESL_1
registers with 0xffffffff. Move the parens to the correct place
to avoid this.

In particular this caused bit 30 of said registers to be set, which
caused the sprite CSC to produce incorrect results.

Cc: stable@vger.kernel.org
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=72220
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: reverse dp link param selection, prefer fast over wide again
Daniel Vetter [Mon, 3 Mar 2014 10:18:10 +0000 (11:18 +0100)]
drm/i915: reverse dp link param selection, prefer fast over wide again

... it's this time of the year again. Originally we've frobbed this to
fix up some regressions, but maybe our DP code improved sufficiently
now that we can dare to do again what the spec recommends.

This reverts

commit 2514bc510d0c3aadcc5204056bb440fa36845147
Author: Jesse Barnes <jbarnes@virtuousgeek.org>
Date:   Thu Jun 21 15:13:50 2012 -0700

    drm/i915: prefer wide & slow to fast & narrow in DP configs

I'm pretty sure I'll regret this patch, but otoh I expect we won't
make progress here without poking the devil occasionally.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=73694
Cc: peter@colberg.org
Cc: Jesse Barnes <jbarnes@virtuousgeek.org>
Tested-by: Itai BEN YAACOV <candeb@free.fr>
Tested-by: David En <d.engraf@arcor.de>
Reported-and-Tested-by: Marcus Bergner <marcusbergner@gmail.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Acked-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: No need to put forcewake after a reset
Mika Kuoppala [Wed, 5 Mar 2014 16:08:18 +0000 (18:08 +0200)]
drm/i915: No need to put forcewake after a reset

As we now have intel_uncore_forcewake_reset() no need
to do explicit put after reset.

v2: rebase

Signed-off-by: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Fix i915_switch_context() argument name in kerneldoc
Damien Lespiau [Mon, 3 Mar 2014 23:57:24 +0000 (23:57 +0000)]
drm/i915: Fix i915_switch_context() argument name in kerneldoc

While reading some code, out of boredom, stumbled on a tiny tiny fix.

Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Remove unused to_gem_object() macro
Damien Lespiau [Mon, 3 Mar 2014 17:42:37 +0000 (17:42 +0000)]
drm/i915: Remove unused to_gem_object() macro

That macro was only ever used to convert ring->private into a gem object
(hence the forceful cast). ring->private doesn't even exist anymore as
it was transmogrified by Chris in:

  commit 0d1aacac36530fce058d7a0db3da7befd5765417
  Author: Chris Wilson <chris@chris-wilson.co.uk>
  Date:   Mon Aug 26 20:58:11 2013 +0100

      drm/i915: Embed the ring->private within the struct intel_ring_buffer

Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Make i915_gem_retire_requests_ring() static
Damien Lespiau [Mon, 3 Mar 2014 17:42:36 +0000 (17:42 +0000)]
drm/i915: Make i915_gem_retire_requests_ring() static

Its last usage outside of i915_gem.c was removed in:

  commit 1f70999f9052f5a1b0ce1a55aff3808f2ec9fe42
  Author: Chris Wilson <chris@chris-wilson.co.uk>
  Date:   Mon Jan 27 22:43:07 2014 +0000

     drm/i915: Prevent recursion by retiring requests when the ring is full

Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Don't just say it, actually force edp vdd
Patrik Jakobsson [Mon, 3 Mar 2014 23:42:44 +0000 (00:42 +0100)]
drm/i915: Don't just say it, actually force edp vdd

This patch fixes the blank screen bug introduced in 3.14-rc1 on the
MacBook Air 6,2. The comments state that we need to force edp vdd so
lets put it back.

The regression was introduced by the following commit:

commit dff392dbd258381a6c3164f38420593f2d291e3b
Author: Paulo Zanoni <paulo.r.zanoni@intel.com>
Date:   Fri Dec 6 17:32:41 2013 -0200

    drm/i915: don't touch the VDD when disabling the panel

v2: Wrap intel_disable_dp() with _vdd_on and _vdd_off

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=74628
Cc: Paulo Zanoni <paulo.r.zanoni@intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Patrik Jakobsson <patrik.r.jakobsson@gmail.com>
Acked-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Make num_sprites a per-pipe value
Damien Lespiau [Mon, 3 Mar 2014 17:31:48 +0000 (17:31 +0000)]
drm/i915: Make num_sprites a per-pipe value

In the future, we need to be able to specify per-pipe number of
planes/sprites. Let's start today!

Signed-off-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>
10 years agodrm/i915: Add a for_each_sprite() macro
Damien Lespiau [Mon, 3 Mar 2014 17:31:47 +0000 (17:31 +0000)]
drm/i915: Add a for_each_sprite() macro

This macro is similar to for_each_pipe() we already have. Convert the
two call sites we have at the same time.

Signed-off-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>
10 years agodrm/i915: Replace a few for_each_pipe(i) by for_each_pipe(pipe)
Damien Lespiau [Mon, 3 Mar 2014 17:31:46 +0000 (17:31 +0000)]
drm/i915: Replace a few for_each_pipe(i) by for_each_pipe(pipe)

Consistency throughout the code base is good and remove some room for
mistakes (as explained in the "drm/i915: Use a pipe variable to cycle
through the pipes" commit)

So, let's replace the for_each_pipe(i) occurences by for_each_pipe(pipe)
when it's reasonable and practical to do so (eg. when there isn't another
pipe variable already).

Suggested-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-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>
10 years agodrm/i915: Don't declare unnecessary shadowing variable
Damien Lespiau [Mon, 3 Mar 2014 17:31:45 +0000 (17:31 +0000)]
drm/i915: Don't declare unnecessary shadowing variable

'i' is already defined in the function scope and used elsewhere. Let's
use it instead.

Signed-off-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>
10 years agodrm/i915: Use a pipe variable to cycle through the pipes
Damien Lespiau [Mon, 3 Mar 2014 17:31:44 +0000 (17:31 +0000)]
drm/i915: Use a pipe variable to cycle through the pipes

I recently fumbled a patch because I wrote twice num_sprites[i], and it
was the right thing to do in only 50% of the cases.

This patch ensures I need to write num_sprites[pipe], ie it should be
self-documented that it's per-pipe number of sprites without having to
look at what is 'i' this time around.

It's all a lame excuse, but it does make it harder to redo the same
mistake.

Signed-off-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>
10 years agodrm/i915: We implement WaDisableAsyncFlipPerfMode:bdw
Ville Syrjälä [Thu, 27 Feb 2014 19:59:03 +0000 (21:59 +0200)]
drm/i915: We implement WaDisableAsyncFlipPerfMode:bdw

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Implement WaDisableSDEUnitClockGating:bdw
Ville Syrjälä [Thu, 27 Feb 2014 19:59:02 +0000 (21:59 +0200)]
drm/i915: Implement WaDisableSDEUnitClockGating:bdw

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Disable semaphore wait event idle message on BDW
Ville Syrjälä [Thu, 27 Feb 2014 19:59:01 +0000 (21:59 +0200)]
drm/i915: Disable semaphore wait event idle message on BDW

According to BSpec we need to always set this magic bit in ring buffer
mode.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Use DIV_ROUND_UP() when calculating number of required FDI lanes
Ville Syrjälä [Thu, 27 Feb 2014 12:23:14 +0000 (14:23 +0200)]
drm/i915: Use DIV_ROUND_UP() when calculating number of required FDI lanes

If we need precisely N lanes to satisfy the FDI bandwidth requirement,
the code would still claim that we need N+1 lanes. Use DIV_ROUND_UP()
to get a more accurate answer.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Fix DDI port_clock for VGA output
Ville Syrjälä [Thu, 27 Feb 2014 12:23:12 +0000 (14:23 +0200)]
drm/i915: Fix DDI port_clock for VGA output

On DDI there's no PLL as such to generate the pixel clock for VGA.
Instead we derive the pixel clock from the FDI link frequency. So
to make .compute_config match what .get_config does, we need to
set the port_clock based on the FDI link frequency.

Note that we don't even check the port_clock when selecting the
PLL for VGA output. We just assume SPLL at 1.35GHz is what we want,
and that does match with the asumption of FDI frequency of 2.7Ghz
we have in intel_fdi_link_freq().

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=74955
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Don't access fifodbg registers on gen8
Mika Kuoppala [Fri, 21 Feb 2014 16:47:36 +0000 (18:47 +0200)]
drm/i915: Don't access fifodbg registers on gen8

as they don't exists.

v2: rename gen6_*_mt_* to gen7_*_mt_* as they never get called
    with gen6 (Chris)

Signed-off-by: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net> (v1)
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Do forcewake reset on gen8
Mika Kuoppala [Fri, 21 Feb 2014 15:32:00 +0000 (17:32 +0200)]
drm/i915: Do forcewake reset on gen8

When we get control from BIOS there might be mt forcewake
bits already set. This causes us to do double mt get
without proper clear/ack sequence.

Fix this by clearing mt forcewake register on init,
like we do with older gens.

Signed-off-by: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: don't flood the logs about bdw semaphores
Jani Nikula [Wed, 5 Mar 2014 12:17:28 +0000 (14:17 +0200)]
drm/i915: don't flood the logs about bdw semaphores

BDW is no longer flagged as preliminary hw, but without
i915.preliminary_hw_support module param set the logs are filled with
WARNs about it.

Just make semaphores off the BDW per-chip default for now.

CC: Ben Widawsky <ben@bwidawsk.net>
Reported-by: Sebastien Dufour <sebastien.dufour@intel.com>
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Add thread stall DOP clock gating workaround on Broadwell.
Kenneth Graunke [Thu, 27 Feb 2014 07:59:31 +0000 (23:59 -0800)]
drm/i915: Add thread stall DOP clock gating workaround on Broadwell.

Ben and I believe this will be necessary on production hardware.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
[danvet: Shuffle lines to group all ROW_CHICKEN writes and add a
cautious comment that this might not be needed on production hw.]
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Add a partial instruction shootdown workaround on Broadwell.
Kenneth Graunke [Thu, 27 Feb 2014 07:59:30 +0000 (23:59 -0800)]
drm/i915: Add a partial instruction shootdown workaround on Broadwell.

I believe this will be necessary on production hardware.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
[danvet: Fix whitespace fail spotted by checkpatch. Also add missing
:bdw w/a tag that Ville spotted.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Add suspend count to error state
Mika Kuoppala [Tue, 25 Feb 2014 15:11:28 +0000 (17:11 +0200)]
drm/i915: Add suspend count to error state

For example if we get bug reports with similar error states and
suspend count is always 1, that might lead the Sherlocks to
right general direction.

Suggested-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Add reset count to error state
Mika Kuoppala [Tue, 25 Feb 2014 15:11:27 +0000 (17:11 +0200)]
drm/i915: Add reset count to error state

By default we keep only the error state from first hang. However
some sneaky user might have cleared the first error state and we
assume mistakenly that it is from first hang. As sometimes this
matters, it is better to explicitly store the reset count.

Signed-off-by: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Add reason for capture in error state
Mika Kuoppala [Tue, 25 Feb 2014 15:11:26 +0000 (17:11 +0200)]
drm/i915: Add reason for capture in error state

We capture error state not only when the GPU hangs but also on
other situations as in interrupt errors and in situations where
we can kick things forward without GPU reset. There will be log
entry on most of these cases. But as error state capture might be
only thing we have, if dmesg was not captured. Or as in GEN4 case,
interrupt error can trigger error state capture without log entry,
the exact reason why capture was made is hard to decipher.

v2: Split out the the error code stuff to separate patch (Ben)

References: https://bugs.freedesktop.org/show_bug.cgi?id=74193
Signed-off-by: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Add error code into error state
Mika Kuoppala [Tue, 25 Feb 2014 15:11:25 +0000 (17:11 +0200)]
drm/i915: Add error code into error state

commit 011cf577b2531dfbd2254bd9ec147ad71471abaf
Author: Ben Widawsky <benjamin.widawsky@intel.com>
Date:   Tue Feb 4 12:18:55 2014 +0000

    drm/i915: Generate a hang error code

added error code debug into dmesg. Store this also
with error state to make matching dmesg logs and error
states easier.

As we need to have full ring state for error code generation,
do full capture always, print hang message into log and then
decide if we need to keep the error state.

Signed-off-by: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Record pid/comm of hanging task
Chris Wilson [Tue, 25 Feb 2014 15:11:24 +0000 (17:11 +0200)]
drm/i915: Record pid/comm of hanging task

After finding the guilty batch and request, we can use it to find the
process that submitted the batch and then add the culprit into the error
state.

This is a slightly different approach from Ben's in that instead of
adding the extra information into the struct i915_hw_context, we use the
information already captured in struct drm_file which is then referenced
from the request.

v2: Also capture the workaround buffer for gen2, so that we can compare
    its contents against the intended batch for the active request.

v3: Rebase (Mika)
v4: Check for null context (Chris)
    checkpatch warnings fixed

Link: http://lists.freedesktop.org/archives/intel-gfx/2013-August/032280.html
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> (v2)
Signed-off-by: Mika Kuoppala <mika.kuoppala@intel.com> (v4)
Acked-by: Ben Widawsky <ben@bwidawsk.net>
Cc: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Rely on accurate request tracking for finding hung batches
Chris Wilson [Tue, 25 Feb 2014 15:11:23 +0000 (17:11 +0200)]
drm/i915: Rely on accurate request tracking for finding hung batches

In the past, it was possible to have multiple batches per request due to
a stray signal or ENOMEM. As a result we had to scan each active object
(filtered by those having the COMMAND domain) for the one that contained
the ACTHD pointer. This was then made more complicated by the
introduction of ppgtt, whereby ACTHD then pointed into the address space
of the context and so also needed to be taken into account.

This is a fairly robust approach (though the implementation is a little
fragile and depends upon the per-generation setup, registers and
parameters). However, due to the requirements for hangstats, we needed a
robust method for associating batches with a particular request and
having that we can rely upon it for finding the associated batch object
for error capture.

If the batch buffer tracking is not robust enough, that should become
apparent quite quickly through an erroneous error capture. That should
also help to make sure that the runtime reporting to userspace is
robust. It also means that we then report the oldest incomplete batch on
each ring, which can be useful for determining the state of userspace at
the time of a hang.

v2: Use i915_gem_find_active_request (Mika)

v3: remove check for ring->get_seqno, split long lines (Ben)

v4: check that context is available (Chris)
    checkpatch warnings fixed

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> (v1)
Signed-off-by: Mika Kuoppala <mika.kuoppala@intel.com> (v3)
Cc: Ben Widawsky <benjamin.widawsky@intel.com>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net> (v3)
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Reset vma->mm_list after unbinding
Chris Wilson [Tue, 25 Feb 2014 14:23:28 +0000 (14:23 +0000)]
drm/i915: Reset vma->mm_list after unbinding

In place of true activity counting, we walk the list of vma associated
with an object managing each on the vm's active/inactive list everytime
we call move-to-inactive. This depends upon the vma->mm_list being
cleared after unbinding, or else we run into difficulty when tracking
the object in multiple vm's - we see a use-after free and corruption of
the mm_list.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Streamline VLV forcewake handling
Ville Syrjälä [Thu, 27 Feb 2014 20:07:21 +0000 (22:07 +0200)]
drm/i915: Streamline VLV forcewake handling

It occured to me that when we're trying to wake up both render
and media wells on VLV, we might end up calling the low level
force_wake_get/put two times even though one call would be
enough. Make that happen by figuring out which wells really
need to be woken up based on the forcewake counts.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by:Deepak S <deepak.s@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Drop the forcewake count inc/dec around register read on VLV
Ville Syrjälä [Mon, 24 Feb 2014 15:02:09 +0000 (17:02 +0200)]
drm/i915: Drop the forcewake count inc/dec around register read on VLV

VLV is the only platform where we increment/decrement the forcewake
count around register access. Drop the inc/dec on VLV to make the
forcewake code a bit more unified.

The inc/dec are not necessary since we hold the uncore lock around
the whole operation.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Deepak S <deepak.s@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Fix VLV forcewake after reset
Ville Syrjälä [Mon, 24 Feb 2014 15:02:08 +0000 (17:02 +0200)]
drm/i915: Fix VLV forcewake after reset

Use the render/media specific forcewake counts to properly restore the
forcewake status after a GPU reset on VLV.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Deepak S <deepak.s@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Perform pageflip using mmio if the GPU is terminally wedged
Chris Wilson [Thu, 20 Feb 2014 09:26:13 +0000 (09:26 +0000)]
drm/i915: Perform pageflip using mmio if the GPU is terminally wedged

After a hang and failed reset, we cannot use the GPU to execute the page
flip instructions. Instead we can force a synchronous mmio flip. (Later,
we can reduce the synchronicity of the mmio flip by moving some of the
delays off to a worker, like the current page flip code; see vblank
tasks.)

References: https://bugs.freedesktop.org/show_bug.cgi?id=72631
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: assert we're not runtime suspended when accessing registers
Paulo Zanoni [Fri, 21 Feb 2014 16:52:28 +0000 (13:52 -0300)]
drm/i915: assert we're not runtime suspended when accessing registers

I could swear this was already happening in the current code...

Also, put the reads and writes in a generic place, so we don't forget
it again when we add runtime PM support to new platforms.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: assert force wake is disabled when we runtime suspend
Paulo Zanoni [Fri, 21 Feb 2014 16:52:26 +0000 (13:52 -0300)]
drm/i915: assert force wake is disabled when we runtime suspend

Just to be sure...

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: call assert_device_not_suspended at gen6_force_wake_work
Paulo Zanoni [Fri, 21 Feb 2014 16:52:25 +0000 (13:52 -0300)]
drm/i915: call assert_device_not_suspended at gen6_force_wake_work

Because we shouldn't be runtime suspended when forcewake is supposed
to be enabled.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
[danvet: Update commit message - no WARN expected since the bugfix for
issues hit with this assert is already in. And resolve conflicts with
the change from worker to timer for the delayed fw release.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: kill dev_priv->pc8.gpu_idle
Paulo Zanoni [Fri, 21 Feb 2014 16:52:24 +0000 (13:52 -0300)]
drm/i915: kill dev_priv->pc8.gpu_idle

Since the addition of dev_priv->mm.busy, there's no more need for
dev_priv->pc8.gpu_idle, so kill it.

Notice that when you remove gpu_idle, hsw_package_c8_gpu_idle and
hsw_package_c8_gpu_busy become identical to hsw_enable_package_c8 and
hsw_disable_package_c8, so just use them.

Also, when we boot the machine, dev_priv->mm.busy initially considers
the machine as idle. This is opposed to dev_priv->pc8.gpu_idle, which
considered it busy. So dev_priv->pc8.disable_count has to be
initalized to 1 now.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: get/put runtime PM in more places at i915_debugfs.c
Paulo Zanoni [Fri, 21 Feb 2014 16:52:23 +0000 (13:52 -0300)]
drm/i915: get/put runtime PM in more places at i915_debugfs.c

These are places where we read (not write) registers while we're
runtime suspended.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: get runtime PM while trying to detect CRT
Paulo Zanoni [Fri, 21 Feb 2014 16:52:22 +0000 (13:52 -0300)]
drm/i915: get runtime PM while trying to detect CRT

Otherwise we'll read registers that return 0xffffffff, trigger some
WARNs, think CRT is actually connected (because certain bits are 1),
and fail the drm-resources-equal testcase!

Tested on a SNB machine with runtime PM support (which is not upstream
yet, but is already on my public tree at freedesktop.org, and will
hopefully eventually become upstream).

Testcase: igt/pm_pc8/drm-resources-equal
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: put runtime PM only when we actually release force_wake
Paulo Zanoni [Fri, 21 Feb 2014 20:58:29 +0000 (17:58 -0300)]
drm/i915: put runtime PM only when we actually release force_wake

When we call gen6_gt_force_wake_put we don't actually put force_wake,
we just schedule gen6_force_wake_work through mod_delayed_work, and
that will eventually release force_wake.

The problem is that we call intel_runtime_pm_put directly at
gen6_gt_force_wake_put, so most of the times we put our runtime PM
reference before the delayed work happens, so we may runtime suspend
while force_wake is still supposed to be enabled if the graphics
autosuspend_delay_ms is too small.

Now the nice thing about the current code is that after it triggers
the delayed work function it gets a refcount, and it only triggers the
delayed work function if refcount is zero. This guarantees that when
we schedule the funciton, it will run before we try to schedule it
again, which simplifies the problem and allows for the current
solution to work properly (hopefully!).

v2: - Keep the VLV refcounts balanced (Jesse)

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: put runtime PM only at the end of intel_mark_idle
Paulo Zanoni [Fri, 21 Feb 2014 16:52:19 +0000 (13:52 -0300)]
drm/i915: put runtime PM only at the end of intel_mark_idle

Because intel_mark_idle still touches some registers: it needs the
machine to be awake. If you set both the autosuspend and PC8 delays to
zero, you can get a "Device suspended" WARN when gen6_rps_idle touches
registers.

This is not easy to reproduce, but happens once in a while when
running pm_pc8.

Testcase: igt/pm_pc8
Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Don't ban default context when stop_rings!=0
Ville Syrjälä [Fri, 21 Feb 2014 14:26:47 +0000 (16:26 +0200)]
drm/i915: Don't ban default context when stop_rings!=0

If we've explicitly stopped the rings for testing purposes, don't ban
the default context. Fixes kms_flip hang tests.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Acked-by: Mika Kuoppala <mika.kuoppala@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: print connector mode list in display_info
Jesse Barnes [Thu, 20 Feb 2014 20:39:57 +0000 (12:39 -0800)]
drm/i915: print connector mode list in display_info

Useful for bug reports.

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Update VBT data structures to have MIPI block enhancements
Shobhit Kumar [Fri, 28 Feb 2014 05:48:46 +0000 (11:18 +0530)]
drm/i915: Update VBT data structures to have MIPI block enhancements

MIPI Block #52 which provides configuration details for the MIPI panel
including dphy settings as per panel and tcon specs

Block #53 gives information on panel enable sequences

v2: Address review comemnts from Jani
    - Move panel ids from intel_dsi.h to intel_bios.h
    - bdb_mipi_config structure improvements for cleaner code
    - Adding units for the pps delays, all in ms
    - change data structure to be more cleaner and simple

v3: Corrected the unit for pps delays as 100us

Signed-off-by: Shobhit Kumar <shobhit.kumar@intel.com>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Convert the forcewake worker into a timer func
Chris Wilson [Wed, 5 Mar 2014 12:00:39 +0000 (12:00 +0000)]
drm/i915: Convert the forcewake worker into a timer func

We don't want to suffer scheduling delay when turning off the GPU after
waking it up to touch registers. Ideally, we only want to keep the GPU
awake for the register access sequence, with a single forcewake dance on
the first access and release immediately after the last. We set a timer
on the first access so that we only dance once and on the next scheduler
tick, we drop the forcewake again.

This moves the cleanup routine from the common i915 workqueue to a timer
func so that we don't anger powertop, and drop the forcewake again
quicker.

v2: Enable the deferred force_wake_put for regular register reads as
    well.
v3: Beautification and make sure we disable forcewake when shutting
    down.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Ben Widawsky <ben@bwidawsk.net>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915/bdw: Add FBC support
Ben Widawsky [Fri, 21 Feb 2014 00:01:20 +0000 (16:01 -0800)]
drm/i915/bdw: Add FBC support

This got lost when we shuffled around our internal branch and
GEN7_FEATURES macro. There were no HW changes to support FBC, so we just
need to set the flag.

v2: Don't allow FBC for any pipe but A on platforms with DDI. (Paulo)

Cc: Daisy Sun <daisy.sun@intel.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Accurately track when we mark the hardware as idle/busy
Chris Wilson [Fri, 21 Feb 2014 17:55:39 +0000 (17:55 +0000)]
drm/i915: Accurately track when we mark the hardware as idle/busy

We currently call intel_mark_idle() too often, as we do so as a
side-effect of processing the request queue. However, we the calls to
intel_mark_idle() are expected to be paired with a call to
intel_mark_busy() (or else we try to idle the hardware by accessing
registers that are already disabled). Make the idle/busy tracking
explicit to prevent the multiple calls.

v2: We can drop some of the complexity in __i915_add_request() as
queue_delayed_work() already behaves as we want (not requeuing the item
if it is already in the queue) and mark_busy/mark_idle imply that the
idle task is inactive.

v3: We do still need to cancel the pending idle task so that it is sent
again after the current busy load completes (not in the middle of it).

Reported-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Tested-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Fix forcewake counts for gen8
Mika Kuoppala [Tue, 18 Feb 2014 17:10:24 +0000 (19:10 +0200)]
drm/i915: Fix forcewake counts for gen8

Sometimes generic driver code gets forcewake explicitly by
gen6_gt_force_wake_get(), which check forcewake_count before accessing
hardware. However the register access with gen8_write function access
low level hw accessors directly, ignoring the forcewake_count. This
leads to nested forcewake get from hardware, in ring init and possibly
elsewhere, causing forcewake ack clear errors and/or hangs.

Fix this by checking the forcewake count also in gen8_write

v2: Read side doesn't care about shadowed registers,
    Remove __needs_put funkiness from gen8_write. (Ville)
    Improved commit message.

References: https://bugs.freedesktop.org/show_bug.cgi?id=74007
Signed-off-by: Mika Kuoppala <mika.kuoppala@intel.com>
Cc: Ben Widawsky <benjamin.widawsky@intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Mika Kuoppala <mika.kuoppala@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: move hsw power domain comment to its right place
Imre Deak [Mon, 17 Feb 2014 22:02:19 +0000 (00:02 +0200)]
drm/i915: move hsw power domain comment to its right place

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: use power domain api to check vga power state
Imre Deak [Mon, 17 Feb 2014 22:02:16 +0000 (00:02 +0200)]
drm/i915: use power domain api to check vga power state

This way we can reuse the check on other platforms too. Also factor out
a version of the function that doesn't check if the power is on, we'll
need to call this from within the power domain framework.

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: switch order of power domain init wrt. irq install
Imre Deak [Mon, 17 Feb 2014 22:02:15 +0000 (00:02 +0200)]
drm/i915: switch order of power domain init wrt. irq install

On VLV at least the display IRQ register access and functionality
depends on its power well to be on, so move the power domain HW init
before we install the IRQs.

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: use drm_i915_private everywhere in the power domain api
Imre Deak [Mon, 17 Feb 2014 22:02:02 +0000 (00:02 +0200)]
drm/i915: use drm_i915_private everywhere in the power domain api

The power domains framework is internal to the i915 driver, so pass
drm_i915_private instead of drm_device to its functions.

Also remove a dangling intel_set_power_well() declaration.

No functional change.

Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: ignore bios output config if not all outputs are on
Daniel Vetter [Tue, 4 Mar 2014 20:08:42 +0000 (21:08 +0100)]
drm/i915: ignore bios output config if not all outputs are on

Both Ville and QA rather immediately complained that with the new
initial_config logic from Jesse not all outputs get enabled. Since the
fbdev emulation pretty much tries to always enable as many outputs as
possible (it even has hotplug handling and all that) fall back if more
outputs could have been enabled.

v2: Fix up my confusion about what enabled means - it's passed from
the fbdev helper, we need to check for a non-zero connector->encoder
link. Spotted by Ville.

v3: Add some debug output as requested by Jesse for debugging fallback
issues.

Cc: Jesse Barnes <jbarnes@virtuousgeek.org>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=75552
Tested-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: s/any_enabled/!fallback/ in fbdev_initial_config
Daniel Vetter [Tue, 4 Mar 2014 20:08:41 +0000 (21:08 +0100)]
drm/i915: s/any_enabled/!fallback/ in fbdev_initial_config

It started as a simple check whether anything is lit up, but now is't
used to driver the general fallback logic to the default output
configuration selector in the helper library. So rename it for more
clarity.

Cc: Jesse Barnes <jbarnes@virtuousgeek.org>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Reject changes of fb base when we have a flip pending
Chris Wilson [Tue, 4 Mar 2014 13:15:08 +0000 (13:15 +0000)]
drm/i915: Reject changes of fb base when we have a flip pending

This should be impossible due to the wait for outstanding flips that the
caller is meant to perform prior to updating the scanout base. Paranoia
tells me to check anyway.

References: https://bugs.freedesktop.org/show_bug.cgi?id=75502
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agoRevert "drm/i915: enable HiZ Raw Stall Optimization on IVB"
Chris Wilson [Tue, 4 Mar 2014 09:41:43 +0000 (09:41 +0000)]
Revert "drm/i915: enable HiZ Raw Stall Optimization on IVB"

This reverts commit 116f2b6da868dec7539103574d0421cd6221e931.

This optimization causes widespread corruption in games, and even in
glxgears, on my ivb:gt1. The corruption appears like z-fighting of
overlapping polygons in the HiZ buffer.

The observation ties in very closely with the description of the
optimization disabled by default on IVB:

"The Hierarchical Z RAW Stall Optimization allows non-overlapping
polygons in the same 8x4 pixel/sample area to be processed without
stalling waiting for the earlier ones to write to Hierarchical Z
buffer."

No reason is given for why it is disabled by default, usually for such
optimizations it is that it is incomplete. However, there is no
indication whether this a gt1 only issue either. Before considering
reenabling this optimization, I would first suggest reproducing the
corruption in piglit.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=75623
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Chia-I Wu <olv@lunarg.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: re-add locking around hw state readout
Jesse Barnes [Fri, 21 Feb 2014 21:13:39 +0000 (13:13 -0800)]
drm/i915: re-add locking around hw state readout

To silence locking complaints.  This was a rebase failure on my part in

commit fa9fa083d0606cb323f6105c17702460ea0a6780
Author: Jesse Barnes <jbarnes@virtuousgeek.org>
Date:   Tue Feb 11 15:28:56 2014 -0800

    drm/i915: read out hw state earlier v2

Reported-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: honor forced connector modes v2
Jesse Barnes [Thu, 20 Feb 2014 20:28:07 +0000 (12:28 -0800)]
drm/i915: honor forced connector modes v2

In the move over to use BIOS connector configs, we lost the ability to
force a specific set of connectors on or off.  Try to remedy that by
dropping back to the old behavior if we detect a hard coded connector
config.

v2: don't deref connector state for disabled connectors (Jesse)

Reported-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Revert workaround for disabling L3 cache aging on IVB
Chris Wilson [Fri, 14 Feb 2014 22:34:43 +0000 (22:34 +0000)]
drm/i915: Revert workaround for disabling L3 cache aging on IVB

In commit e4e0c058a19c41150d12ad2d3023b3cf09c5de67
Author: Eugeni Dodonov <eugeni.dodonov@intel.com>
Date:   Wed Feb 8 12:53:50 2012 -0800

    drm/i915: gen7: Implement an L3 caching workaround.

the L3 cache aging was disabled. This was part of a shotgun response
to a number of GPU hang bugs, but there appears to be no documentation
to suggest that disabling the L3 cache age was ever required (to prevent
the GPU hangs).

Restoring the L3 cache age is a minor performance win of around 2%
on IVB:GT2. (Note that this value seems to be consistent across a number
of tests and so appears to be above the usual noise.)

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Revert workaround for disabling L3 cache aging on BYT
Sinclair Yeh [Wed, 19 Feb 2014 21:09:31 +0000 (13:09 -0800)]
drm/i915: Revert workaround for disabling L3 cache aging on BYT

V2:  edit the commit message to contain more info
The W/A spreadsheet says this is still required, but the b-spec says
it's not for BYT-T.  So the documentation is not clear.  However,
our experience with the other SKUs of BYT-I/M on Android and Linux
suggests that setting this bit actually causes GPU hang for certain
OGL benchmark applications.

Removing this bit completely resolves the GPU hangs.

Signed-off-by: Sinclair Yeh <sinclair.yeh@intel.com>
Acked-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915/bdw: Kill ppgtt->num_pt_pages
Ben Widawsky [Fri, 21 Feb 2014 21:06:34 +0000 (13:06 -0800)]
drm/i915/bdw: Kill ppgtt->num_pt_pages

With the original PPGTT implementation if the number of PDPs was not a
power of two, the number of pages for the page tables would end up being
rounded up. The code actually had a bug here afaict, but this is a
theoretical bug as I don't believe this can actually occur with the
current code/HW..

With the rework of the page table allocations, there is no longer a
distinction between number of page table pages, and number of page
directory entries. To avoid confusion, kill the redundant (and newer)
struct member.

Cc: Imre Deak <imre.deak@intel.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Split GEN6 PPGTT initialization up
Ben Widawsky [Thu, 20 Feb 2014 06:05:49 +0000 (22:05 -0800)]
drm/i915: Split GEN6 PPGTT initialization up

Simply to match the GEN8 style of PPGTT initialization, split up the
allocations and mappings. Unlike GEN8, we skip a separate dma_addr_t
allocation function, as it is much simpler pre-gen8.

With this code it would be easy to make a more general PPGTT
initialization function with per GEN alloc/map/etc. or use a common
helper, similar to the ringbuffer code. I don't see a benefit to doing
this just yet, but who knows...

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Split GEN6 PPGTT cleanup
Ben Widawsky [Thu, 20 Feb 2014 06:05:48 +0000 (22:05 -0800)]
drm/i915: Split GEN6 PPGTT cleanup

This cleanup is similar to the GEN8 cleanup (though less necessary).
Having everything split will make cleaning the initialization path error
paths easier to understand.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Update i915_gem_gtt.c copyright
Ben Widawsky [Thu, 20 Feb 2014 06:05:47 +0000 (22:05 -0800)]
drm/i915: Update i915_gem_gtt.c copyright

I keep meaning to do this... by now almost the entire file has been
written by an Intel employee (including Daniel post-2010).

Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agoRevert "drm/i915/bdw: Limit GTT to 2GB"
Ben Widawsky [Thu, 20 Feb 2014 06:05:46 +0000 (22:05 -0800)]
Revert "drm/i915/bdw: Limit GTT to 2GB"

This reverts commit 3a2ffb65eec6dbda2fd8151894f51c18b42c8d41.

Now that the code is fixed to use smaller allocations, it should be safe
to let the full GGTT be used on BDW.

The testcase for this is anything which uses more than half of the GTT,
thus eclipsing the old limit.

Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915/bdw: Reorganize PT allocations
Ben Widawsky [Thu, 20 Feb 2014 19:51:21 +0000 (11:51 -0800)]
drm/i915/bdw: Reorganize PT allocations

The previous allocation mechanism would get 2 contiguous allocations,
one for the page directories, and one for the page tables. As each page
table is 1 page, and there are 512 of these per page directory, this
goes to 2MB. An unfriendly request at best. Worse still, our HW now
supports 4 page directories, and a 2MB allocation is not allowed.

In order to fix this, this patch attempts to split up each page table
allocation into a single, discrete allocation. There is nothing really
fancy about the patch itself, it just has to manage an extra pointer
indirection, and have a fancier bit of logic to free up the pages.

To accommodate some of the added complexity, two new helpers are
introduced to allocate, and free the page table pages.

NOTE: I really wanted to split the way we do allocations, and the way in
which we identify the page table/page directory being used. I found
splitting this functionality up to be too unwieldy. I apologize in
advance to the reviewer. I'd recommend looking at the result, rather
than the diff.

v2/NOTE2: This patch predated commit:
6f1cc993518462ccf039e195fabd47e7aa5bfd13
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Tue Dec 31 15:50:31 2013 +0000

    drm/i915: Avoid dereference past end of page arr

It fixed the same issue as that patch, but because of the limbo state of
PPGTT, Chris patch was merged instead. The excess churn is a result of
my using my original patch, which has my preferred naming. Primarily
act_* is changed to which_*, but it's mostly the same otherwise. I've
kept the convention Chris used for the pte wrap (I had something
slightly different, and broken - but fixable)

v3: Rename which_p[..]e to drop which_ (Chris)
Remove BUG_ON in inner loop (Chris)
Redo the pde/pdpe wrap logic (Chris)

v4: s/1MB/2MB in commit message (Imre)
Plug leaking gen8_pt_pages in both the error path, as well as general
free case (Imre)

v5: Rename leftover "which_" variables (Imre)
Add the pde = 0 wrap that was missed from v3 (Imre)

Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
[danvet: Squash in fixup from Ben.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Make clear/insert vfuncs args absolute
Ben Widawsky [Thu, 20 Feb 2014 19:50:33 +0000 (11:50 -0800)]
drm/i915: Make clear/insert vfuncs args absolute

This patch converts insert_entries and clear_range, both functions which
are specific to the VM. These functions tend to encapsulate the gen
specific PTE writes. Passing absolute addresses to the insert_entries,
and clear_range will help make the logic clearer within the functions as
to what's going on. Currently, all callers simply do the appropriate
page shift, which IMO, ends up looking weird with an upcoming change for
the gen8 page table allocations.

Up until now, the PPGTT was a funky 2 level page table. GEN8 changes
this to look more like a 3 level page table, and to that extent we need
a significant amount more memory simply for the page tables. To address
this, the allocations will be split up in finer amounts.

v2: Replace size_t with uint64_t (Chris, Imre)

v3: Fix size in gen8_ppgtt_init (Ben)
Fix Size in i915_gem_suspend_gtt_mappings/restore (Imre)

Reviewed-by: Imre Deak <imre.deak@intel.com> (v2)
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915/bdw: Split ppgtt initialization up
Ben Widawsky [Thu, 20 Feb 2014 06:05:43 +0000 (22:05 -0800)]
drm/i915/bdw: Split ppgtt initialization up

Like cleanup in an earlier patch, the code becomes much more readable,
and easier to extend if we extract out helper functions for the various
stages of init.

Note that with this patch it becomes really simple, and tempting to begin
using the 'goto out' idiom with explicit free/fini semantics. I've
kept the error path as similar as possible to the cleanup() function to
make sure cleanup is as robust as possible

v2: Remove comment "NB:From here on, ppgtt->base.cleanup() should
function properly"
Update commit message to reflect above

v3: Rebased on top of bugfixes found in the previous patch by Imre
Moved number of pd pages assertion to the proper place (Imre)

v4:
Allocate dma address space for num_pd_pages, not num_pd_entries (Ben)
Don't use gen8_pt_dma_addr after free on error path (Imre)
With new fix from v4 of the previous patch.

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915/bdw: Reorganize PPGTT init
Ben Widawsky [Thu, 20 Feb 2014 06:05:42 +0000 (22:05 -0800)]
drm/i915/bdw: Reorganize PPGTT init

Create 3 clear stages in PPGTT init. This will help with upcoming
changes be more readable. The 3 stages are, allocation, dma mapping, and
writing the P[DT]Es

One nice benefit to the patches is that it makes 2 very clear error
points, allocation, and mapping, and avoids having to do any handling
after writing PTEs (something which was likely buggy before). This
simplified error handling I suspect will be helpful when we move to
deferred/dynamic page table allocation and mapping.

The patches also attempts to break up some of the steps into more
logical reviewable chunks, particularly when we free.

v2: Don't call cleanup on the error path since that takes down the
drm_mm and list entry, which aren't setup at this point.

v3: Fixes addressing Imre's comments from:
<1392821989.19792.13.camel@intelbox>

Don't do dynamic allocation for the page table DMA addresses. I can't
remember why I did it in the first place. This addresses one of Imre's
other issues.

Fix error path leak of page tables.

v4: Fix the fix of the error path leak. Original fix still leaked page
tables. (Imre)

Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915/bdw: Free PPGTT struct
Ben Widawsky [Thu, 20 Feb 2014 19:47:07 +0000 (11:47 -0800)]
drm/i915/bdw: Free PPGTT struct

GEN8 never freed the PPGTT struct. As GEN8 doesn't use full PPGTT, the
leak is small and only found on a module reload. ie. I don't think this
needs to go to stable.

v2: The very naive, kfree in gen8 ppgtt cleanup, is subject to a double
free on PPGTT initialization failure. (Spotted by Imre). Instead this
patch pulls the ppgtt struct freeing out of the cleanup and leaves it to
the allocators/callers or the one doing the last kref_put as in standard
convention

Reported-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Move ppgtt_release out of the header
Ben Widawsky [Thu, 20 Feb 2014 19:47:06 +0000 (11:47 -0800)]
drm/i915: Move ppgtt_release out of the header

At one time it was expected to be called in multiple places by kref_put.
At the current time however, it is all contained within
i915_gem_context.c.

This patch makes an upcoming required addition a bit nicer since it too
doesn't need to be defined in a header file.

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Add a comment about WIZ hashing vs. thread counts
Ville Syrjälä [Wed, 5 Feb 2014 10:43:47 +0000 (12:43 +0200)]
drm/i915: Add a comment about WIZ hashing vs. thread counts

Add a comment next to our WIZ hashing setup to remind people about the
link between WIZ hashing disable bit and PS/WM thread counts.

Suggested-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Change BDW WIZ hashing mode to 16x4
Ville Syrjälä [Tue, 4 Feb 2014 19:59:21 +0000 (21:59 +0200)]
drm/i915: Change BDW WIZ hashing mode to 16x4

BSpec recommends using 8x4 hashing mode when MSAA is used. But in
practice 16x4 seems to have a slight edge in performance (on IVB and
HSW at least). So just use 16x4.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Antti Koskipää <antti.koskipaa@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Change HSW WIZ hashing mode to 16x4
Ville Syrjälä [Tue, 4 Feb 2014 19:59:20 +0000 (21:59 +0200)]
drm/i915: Change HSW WIZ hashing mode to 16x4

BSpec recommends using 8x4 hashing mode when MSAA is used. But in
practice 16x4 seems to have a slight edge in performance (on IVB and
HSW at least). So just use 16x4.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Antti Koskipää <antti.koskipaa@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Change IVB WIZ hashing mode to 16x4
Ville Syrjälä [Tue, 4 Feb 2014 19:59:19 +0000 (21:59 +0200)]
drm/i915: Change IVB WIZ hashing mode to 16x4

BSpec recommends using 8x4 hashing mode when MSAA is used. But in
practice 16x4 seems to have a slight edge in performance (on IVB and
HSW at least). So just use 16x4.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Antti Koskipää <antti.koskipaa@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: There's no need to mask all 3D_CHICKEN bits on SNB
Ville Syrjälä [Tue, 4 Feb 2014 19:59:17 +0000 (21:59 +0200)]
drm/i915: There's no need to mask all 3D_CHICKEN bits on SNB

The need to set all of the mask bits for 3D_CHICKEN3 was required
only for pre-production hardware.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Antti Koskipää <antti.koskipaa@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Assume we implement WaStripsFansDisableFastClipPerformanceFix:snb
Ville Syrjälä [Tue, 4 Feb 2014 19:59:16 +0000 (21:59 +0200)]
drm/i915: Assume we implement WaStripsFansDisableFastClipPerformanceFix:snb

Based on the name, the workaround we implement is
WaStripsFansDisableFastClipPerformanceFix. Unfortunately there's no
description in the w/a database, so this is just a guess.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Antti Koskipää <antti.koskipaa@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Fix SNB GT_MODE register setup
Ville Syrjälä [Tue, 4 Feb 2014 19:59:15 +0000 (21:59 +0200)]
drm/i915: Fix SNB GT_MODE register setup

On SNB we set up WaSetupGtModeTdRowDispatch:snb early in
gen6_init_clock_gating(). That sets a bit in the GEN6_GT_MODE register.
However later we go and disable all the bits in the same register. And
then we go on to set some other bit. So apparently we never actually
implemented this workaround since the "disable all bits" part was there
already before the w/a got supposedly implemented.

These are the relevant commits:

 commit 6547fbdbfff62c99e4f7b4f985ff8b3454f33b0f
 Author: Daniel Vetter <daniel.vetter@ffwll.ch>
 Date:   Fri Dec 14 23:38:29 2012 +0100

    drm/i915: Implement WaSetupGtModeTdRowDispatch

 commit f8f2ac9a76b0f80a6763ca316116a7bab8486997
 Author: Ben Widawsky <ben@bwidawsk.net>
 Date:   Wed Oct 3 19:34:24 2012 -0700

    drm/i915: Fix GT_MODE default value

So, let's drop the "disable all bits" part, move both writes to
closer proxomity to each other, and name the WIZ hashing bits
appropriately. BSpec is still a bit confused how the bits should
actually be interpreted, but I took the the description for the
high bit since the low bit part only lists values for a single bit.

Also add a comment about our choice of WIZ hashing mode.

Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Antti Koskipää <antti.koskipaa@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: get/put runtime PM without holding rps.hw_lock
Paulo Zanoni [Thu, 19 Dec 2013 13:54:52 +0000 (11:54 -0200)]
drm/i915: get/put runtime PM without holding rps.hw_lock

We'll need this when we merge PC8 and Runtime PM: the PC8
enable/disable functions need that lock.

Also, it's good practice to not hold a lock for longer than strictly
needed.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: rename modeset_update_power_wells
Paulo Zanoni [Thu, 19 Dec 2013 13:54:51 +0000 (11:54 -0200)]
drm/i915: rename modeset_update_power_wells

To modeset_update_crtc_power_domains, since this function is
responsible for updating all the power domains of all CRTCs after a
modeset. In the future we should also run this function on all
platforms, not just Haswell.

Signed-off-by: Paulo Zanoni <paulo.r.zanoni@intel.com>
Reviewed-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Remove dead code
Thierry Reding [Fri, 21 Feb 2014 07:55:27 +0000 (08:55 +0100)]
drm/i915: Remove dead code

The i915 driver sets DRIVER_GEM unconditionally, so testing for the
feature will always fail.

Signed-off-by: Thierry Reding <treding@nvidia.com>
[danvet: Fix up conflicts.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: sprinkle static
Daniel Vetter [Sun, 2 Mar 2014 20:18:00 +0000 (21:18 +0100)]
drm/i915: sprinkle static

Apparently we've missed a few more than what Fengguang's 0-day tester
recently reported in i915_irq.c ... Makes sparse happy again (ignore
some spurious stuff about ksyms of exported functions).

Cc: kbuild test robot <fengguang.wu@intel.com>
Cc: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: tune down user-triggerable dmesg noise in the cursor/overlay code
Daniel Vetter [Fri, 14 Feb 2014 13:06:06 +0000 (14:06 +0100)]
drm/i915: tune down user-triggerable dmesg noise in the cursor/overlay code

Spotted while auditing the code for fencing issues.

Cc: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: fix NULL deref in the load detect code
Daniel Vetter [Fri, 14 Feb 2014 15:35:54 +0000 (16:35 +0100)]
drm/i915: fix NULL deref in the load detect code

Looks like I've missed one of the potential NULL deref bugs in Jesse's
fbdev->fb embedded struct to pointer conversions. Fix it up.

This regression has been introduced in

commit 8bcd45534ddf68ab71aeed709dacd9cf65dc0f75
Author: Jesse Barnes <jbarnes@virtuousgeek.org>
Date:   Fri Feb 7 12:10:38 2014 -0800

    drm/i915: alloc intel_fb in the intel_fbdev struct

Cc: Jesse Barnes <jbarnes@virtuousgeek.org>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
10 years agodrm/i915: Only bind each object rather than for every execbuffer
Daniel Vetter [Fri, 14 Feb 2014 13:01:21 +0000 (14:01 +0100)]
drm/i915: Only bind each object rather than for every execbuffer

One side-effect of the introduction of ppgtt was that we needed to
rebind the object into the appropriate vm (and global gtt in some
peculiar cases). For simplicity this was done twice for every object on
every call to execbuffer. However, that adds a tremendous amount of CPU
overhead (rewriting all the PTE for all objects into WC memory) per
draw. The fix is to push all the decision about which vm to bind into
and when down into the low-level bind routines through hints rather than
performing the bind unconditionally in the execbuffer routine.

Note that this is a regression introduced in the full ppgtt feature
branch, before this we've only done re-bound objects when the relevant
has_(aliasing_ppgtt|global_gtt)_mapping flag was clear. But since
that's per-object and not per-vma that optimization broke.

v2: Split out prep work and unrelated changes.

v3: Bring back functional change around PIN_GLOBAL that I've
accidentally split out.

v4: Remove the temporary hack for the old binding logic to avoid
bisection issues.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=72906
Tested-by: jianx.zhou@intel.com
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk> (v1)
Cc: Ben Widawsky <benjamin.widawsky@intel.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Acked-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>