agp/intel-gtt: export the gtt pagetable iomapping

We need this because ppgtt page directory entries need to be in the
global gtt pagetable.

Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Tested-by: Chris Wilson <chris@chris-wilson.co.uk>
Tested-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

agp/intel-gtt: export the scratch page dma address

To implement a PPGTT for drm/i915 that fully aliases the GTT, we also
need to properly alias the scratch page.

Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Tested-by: Chris Wilson <chris@chris-wilson.co.uk>
Tested-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: dump even more into the error_state

Chris Wilson and me have again stared at funny error states and it's
been pretty clear from the start that something was seriously amiss.
The seqnos last seen by the cpu were a few hundred behind those that
the gpu could have possibly emitted last before it died ...

Chris now tracked it down (hopefully, definit verdict's still out),
but in hindsight we'd have found the bug by simply dumping the cpu
side tracking of the ring head and tail registers.

Fix this and prevent an identical time-waster in the future.

Because the hangs always involved semaphores in one way or another,
we've tried to dump the mbox registers, but couldn't find any
inconsistencies. Still, dump them too.

Reviewed-and-wanted-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: s/DRM_ERROR/DRM_DEBUG in i915_gem_execbuffer.c

These are all user-trigerable, so tune down their loudness a notch.
For some of these we have i-g-t tests (because they prevent
newly-discovered bugs), without this patches running the test suite
leaves behind a dirty dmesg.

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

drm/i915: add gen6+ registers to i915_swizzle_info

Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: consolidate swizzling control bit frobbing

On gen5 we also need to correctly set up swizzling in the display
scanout engine, but only there. Consolidate this into the same
function.

This has a small effect on ums setups - the kernel now also sets this
bit in addition to userspace setting it. Given that this code only
runs when userspace either can't (resume, gpu reset) or explicitly
won't(gem_init) touch the hw this shouldn't have an adverse effect.

Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: swizzling support for snb/ivb

We have to do this manually. Somebody had a Great Idea.

I've measured speed-ups just a few percent above the noise level
(below 5% for the best case), but no slowdows. Chris Wilson measured
quite a bit more (10-20% above the usual snb variance) on a more
recent and better tuned version of sna, but also recorded a few
slow-downs on benchmarks know for uglier amounts of snb-induced
variance.

v2: Incorporate Ben Widawsky's preliminary review comments and
elaborate a bit about the performance impact in the changelog.

v3: Add a comment as to why we don't need to check the 3rd memory
channel.

v4: Fixup whitespace.

Acked-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: Handle unmappable buffers during error state capture

As the buffer is not necessarily accessible through the GTT at the time
of a GPU hang, and capturing some of its contents is far more valuable
than skipping it, provide a clflushed fallback read path. We still
prefer to read through the GTT as that is more consistent with the GPU
access of the same buffer. So example it will demonstrate any errorneous
tiling or swizzling of the command buffer as seen by the GPU.

This becomes necessary with use of CPU relocations and lazy GTT binding,
but could potentially happen anyway as a result of a pathological error.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: rewrite shmem_pread_slow to use copy_to_user

Like for shmem_pwrite_slow. The only difference is that because we
read data, we can leave the fetched cachelines in the cpu: In the case
that the object isn't in the cpu read domain anymore, the clflush for
the next cpu read domain invalidation will simply drop these
cachelines.

slow_shmem_bit17_copy is now ununsed, so kill it.

With this patch tests/gem_mmap_gtt now actually works.

v2: add __ to copy_to_user_swizzled as suggested by Chris Wilson.

v3: Fixup the swizzling logic, it swizzled the wrong pages.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=38115
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: rewrite shmem_pwrite_slow to use copy_from_user

... instead of get_user_pages, because that fails on non page-backed
user addresses like e.g. a gtt mapping of a bo.

To get there essentially copy the vfs read path into pagecache. We
can't call that right away because we have to take care of bit17
swizzling. To not deadlock with our own pagefault handler we need
to completely drop struct_mutex, reducing the atomicty-guarantees
of our userspace abi. Implications for racing with other gem ioctl:

- execbuf, pwrite, pread: Due to -EFAULT fallback to slow paths there's
  already the risk of the pwrite call not being atomic, no degration.
- read/write access to mmaps: already fully racy, no degration.
- set_tiling: Calling set_tiling while reading/writing is already
  pretty much undefined, now it just got a bit worse. set_tiling is
  only called by libdrm on unused/new bos, so no problem.
- set_domain: When changing to the gtt domain while copying (without any
  read/write access, e.g. for synchronization), we might leave unflushed
  data in the cpu caches. The clflush_object at the end of pwrite_slow
  takes care of this problem.
- truncating of purgeable objects: the shmem_read_mapping_page call could
  reinstate backing storage for truncated objects. The check at the end
  of pwrite_slow takes care of this.

v2:
- add missing intel_gtt_chipset_flush
- add __ to copy_from_user_swizzled as suggest by Chris Wilson.

v3: Fixup bit17 swizzling, it swizzled the wrong pages.

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

drm/i915: fall through pwrite_gtt_slow to the shmem slow path

The gtt_pwrite slowpath grabs the userspace memory with
get_user_pages. This will not work for non-page backed memory, like a
gtt mmapped gem object. Hence fall throuh to the shmem paths if we hit
-EFAULT in the gtt paths.

Now the shmem paths have exactly the same problem, but this way we
only need to rearrange the code in one write path.

v2: v1 accidentaly falls back to shmem pwrite for phys objects. Fixed.

v3: Make the codeflow around phys_pwrite cleara as suggested by Chris
Wilson.

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

drm/i915: add debugfs file for swizzling information

This will also come handy for the gen6+ swizzling support, where the
driver is supposed to control swizzling depending upon dram
configuration.

v2: CxDRB3 are 16 bit regs! Noticed by Chris Wilson.

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

drm/i915: fix swizzle detection for gen3

It looks like the desktop variants of i915 and i945 also have the DCC
register to control dram channel interleave and cpu side bit6
swizzling.

Unfortunately internal Cspec/ConfigDB documentation for these ancient chips
have already been dropped and there seem to be no archives. Also
somebody thought the swizzling behaviour is surely a worthy secret to
keep and redacted any mention of these fields from the published Intel
datasheets.

I suspect the hw engineers were really proud of the page coloring
they've achieved in their first dual channel dram controller with
bit17 - after all Bspec explains in great length the optimal layout of
page frame numbers modulo 4 for the color and depth buffers, too.
Later on when they've started to work on VT-d they shamefully
discoverd their stupidity and tried to cover the tracks ...

Tested-by: Daniel Vetter <daniel.vetter@ffwll.ch> (i915g)
Tested-by: Pavel Ondračka <pavel.ondracka@email.cz> (i945g)
Tested-by: Chris Wilson <chris@chris-wilson.co.uk>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=42625
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: Remove the upper limit on the bo size for mapping into the CPU domain

The original intention of comparing the bo against the mappable GTT
limits was to prevent a subsequent faulting of the bo into the GTT from
clearing the entire GTT in vain. However, that was clearly a cut'n'paste
mistake as a CPU mapping never binds the bo into the aperture. Whilst
there may be some merit to limiting the maximum size of the bo to
something that can be utilized by the GPU, that limit itself does not
belong as a safeguard to mmapping the bo, so remove the check entirely.

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

drm/i915: add per-ring fault reg to error_state

This was pretty handy when figuring out what exactly went wrong with
ppgtt and it might also be useful when we stop filling the entire gart
with scratch page entries.

Also add the gen6+ DONE reg while at it.

v2: Chris Wilson suggested to allocate the error_state with kzalloc
for better paranoia. Also kill existing spurious clears of the
error_state while at it.

Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: reject GTT domain in relocations

This confuses our domain tracking and can (for gtt write domains) lead
to a subsequent oops.

Tested by tests/gem_exec_bad_domains from i-g-t.

Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: remove the i915_batchbuffer_info debugfs file

With the error_state facility in place, this has outlived it's
usefulness. It also oopses with the lates llc-reloc patches because
it directly access objects through the gtt without any checks.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: capture error_state also for stuck rings

Since quite a while we also the basic output configuration in the
error_state, so it should contain enough information to diagnose
these MI_WAIT hangs.

Reviewed-and-tested-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: refactor debugfs create functions

All r/w debugfs files are created equal.

v2: Add some newlines to make the code easier on the eyes as requested
by Ben Widawsky.

Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: refactor debugfs open function

Only forcewake has an open with special semantics, the other r/w
debugfs only assign the file private pointer.

Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: don't trash the gtt when running out of fences

With the fence accounting fixed up in the previous commit not finding
enough fences is a fatal error and userspace bug. Trashing the entire
gtt is not gonna turn up that missing fence, so don't to this by
returning another error thatn ENOSPC.

This has the added benefit that it's easier to distinguish fence
accounting errors from gtt space accounting issues.

TTM serves as precendence for the EDEADLK error code - it returns it
when the reservation code needs resources already blocked by the
current reservation.

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

drm/i915: Separate fence pin counting from normal bind pin counting

In order to correctly account for reserving space in the GTT and fences
for a batch buffer, we need to independently track whether the fence is
pinned due to a fenced GPU access in the batch or whether the buffer is
pinned in the aperture. Currently we count the fenced as pinned if the
buffer has already been seen in the execbuffer. This leads to a false
accounting of available fence registers, causing frequent mass evictions.
Worse, if coupled with the change to make i915_gem_object_get_fence()
report EDADLK upon fence starvation, the batchbuffer can fail with only
one fence required...

Fixes intel-gpu-tools/tests/gem_fenced_exec_thrash

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=38735
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Tested-by: Paul Neumann <paul104x@yahoo.de>
[danvet: Resolve the functional conflict with Jesse Barnes sprite
patches, acked by Chris Wilson on irc.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/ringbuffer: kill snb blt workaround

This was just to facilitate product enablement with pre-production hw.
Allows us to kill quite a bit of cruft.

Signed-off-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: collect more per ring error state

Based on a patch by Ben Widawsky, but with different colors
for the bikeshed.

In contrast to Ben's patch this one doesn't add the fault regs.
Afaics they're for the optional page fault support which
- we're not enabling
- and which seems to be unsupported by the hw team. Recent bspec
  lacks tons of information about this that the public docs released
  half a year back still contain.

Also dump ring HEAD/TAIL registers - I've recently seen a few
error_state where just guessing these is not good enough.

v2: Also dump INSTPM for every ring.

v3: Fix a few really silly goof-ups spotted by Chris Wilson.

Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
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>

drm/i915: refactor ring error state capture to use arrays

The code already got unwieldy and we want to dump more per-ring
registers.

Only functional change is that we now also capture the video
ring registers on ilk.

v2: fixup a refactor fumble spotted by Chris Wilson.

Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
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>

drm/i915: switch ring->id to be a real id

... and add a helpr function for the places where we want a flag.

This way we can use ring->id to index into arrays.

v2: Resurrect the missing beautification-space Chris Wilson noted.
I'm moving this space around because I'll reuse ring_str in the next
patch.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: set AUD_CONFIG N_value_index for DisplayPort

It should be programmed to "0" for HDMI or "1" for DisplayPort.

This enables DisplayPort audio for

- HP EliteBook 8460p
  (whose BIOS does not set the N_value_index bit for us)

- DisplayPort monitor hot plugged after boot
  (otherwise most BIOS will fill the N_value_index bit for us)

Tested-by: Robert Lemaire <rlemaire@suse.com>
Reviewed-by: Keith Packard <keithp@keithp.com>
Signed-off-by: Wu Fengguang <fengguang.wu@intel.com>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: correct lock type in destroy

This is only relevant when using module unloading, and really only helps
get rid of a probably benign warning.

I can't remember if I sent this out already, but it's not turning up in
any of my searches.

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Keith Packard <keithp@keithp.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: drm/i915: Fix recursive calls to unmap

After the ILK vt-d workaround patches it became clear that we had
introduced a bug.  Chris Wilson tracked down the issue to recursive
calls to unmap. This happens because we try to optimize waiting on
requests by calling retire requests after the wait, which may drop the
last reference on an object and end up freeing the object (and then
unmap the object from the gtt).

After the last patch we can now choose to defer processing the retire
list.

Kudos to Chris Wilson for tracking this one down.

This patch fixes gem_unref_active_buffers from i-g-t. It was tested by
forcing do_idle_maps to true.

This also fixes tests/gem_linear_blits in intel-gpu-tools.

Reported-by: guang.a.yang@intel.com
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=42180
Reviewed-by: Keith Packard <keithp@keithp.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: argument to control retiring behavior

Sometimes it may be the case when we idle the gpu or wait on something
we don't actually want to process the retiring list. This patch allows
callers to choose the behavior.

Reviewed-by: Keith Packard <keithp@keithp.com>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: Correct the bit number for the MI_FLUSH_ENABLE.

Older specs claimed this was bit 11, but newer specs and the actual
simulator code say it was bit 12.  Regardless, we don't use MI_FLUSH,
or try to enable it any more.

Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
[danvet: Anyone trying to use this bit, please read all the relevant
discussions, it's epic.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: Remove the MI_FLUSH_ENABLE setting.

We have always been using the wrong bit -- it's bit 12.  However, the
bit also doesn't do anything -- hardware has always accepted the
MI_FLUSH command even when it was specced not to.

Given that there is only one MI_FLUSH emitted in all of the driver
stack on gen6+ (in i965_video.c of the 2d driver, and it should be
using other code to do its flush instead), just remove the MI_FLUSH
enable instead of trying to fix it.

Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: fixup assert_pipe to take the pipe A quirk into account

This was completely spamming dmesg on my i855gm. This issue was just
shortly introduced with:

commit 931872fceabacf2d4f8b6fbd51611c167e83164c
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Mon Jan 16 23:01:13 2012 +0000

    drm/i915: Check that plane/pipe is disabled before removing the fb

Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: print out which pixel format we do not support

Otherwise, we are left with pretty bogus message saying that the pixel
format is not supported while leaving the details to the telepatic powers.

v2: use DRM_DEBUG_KMS instead of DRM_ERROR

Signed-off-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: properly mask and or watermark values for sprites

Now that we're using the sprite WM fields, we need to take care not to
clobber them in the main update_wm functions.  While we're at it, make
sure we mask out the old sprite wm value before or'ing in the new one
when the sprite wm is updated.

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Reviewed-by: Keith Packard <keithp@keithp.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: clarify gen2 pageflip cmd

I've reviewed gen2 pageflip code to hunt down multiple prepare pageflip
issues. The only thing I've found is a slight but functionally
meaningless confusion about the length of the mi cmd.

Fix it up and add a comment about what this dword should be (according
to docs at least).

Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: add a LLC feature flag in device description

LLC is not SNB/IVB-specific, so we should check for it in a more generic
way.

Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: kill i915_mem.c

Some decent history digging indicates that this was to be used for the
GLX_MESA_allocate_memory extension but never actually implemented for
any released i915 userspace code.

So just rip it out.

v2: Fixup the Makefile.

Acked-by: Dave Airlie <airlied@gmail.com>
Cc: Keith Whitwell <keithw@vmware.com>
Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-Off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: Use kcalloc instead of kzalloc to allocate array

The advantage of kcalloc is, that will prevent integer overflows which could
result from the multiplication of number of elements and size and it is also
a bit nicer to read.

The semantic patch that makes this change is available
in https://lkml.org/lkml/2011/11/25/107

Signed-off-by: Thomas Meyer <thomas@m3y3r.de>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/dp: Check for AUXCH error before checking for success

This is paranoid, but I am entirely willing to believe the hardware
could come up with a condition where I get a status with both the 'done'
and 'receive error' bits set.

Signed-off-by: Adam Jackson <ajax@redhat.com>
Acked-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/dp: Use auxch precharge value of 5 everywhere

The default in the Sandybridge docs is 5, as on Ironlake, and I have no
reason to believe 3 would work any better.

Signed-off-by: Adam Jackson <ajax@redhat.com>
Acked-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915/dp: Tweak auxch clock divider for PCH

Matches the advice in the Sandybridge documentation.

Signed-off-by: Adam Jackson <ajax@redhat.com>
Acked-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: Remove a comment about PCH from the non-PCH path

Signed-off-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: Fix assert_pch_hdmi_disabled to mention HDMI (not DP)

Signed-off-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: Implement plane-disabled assertion for PCH too

Signed-off-by: Adam Jackson <ajax@redhat.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drivers: i915: Fix BLC PWM register setup

There is an error in i915_read_blc_pwm_ctl, where the register values
are not being copied correctly.  BLC_PWM_CTL and BLC_PWM_CTL2 are
getting mixed up.  This patch fixes that so that saveBLC_PWM_CTL2 and
not saveBLC_PWM_CTL is copied to the BLC_PWM_CTL2 register.

Signed-off-by: Simon Que <sque@chromium.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: Check that plane/pipe is disabled before removing the fb

Staring at an error state such as:

PGTBL_ER: 0x00000400
    Display B: Invalid tiling
fence[0] = 05001001
    valid, x-tiled, pitch: 512, start: 0x05000000, size: 1048576
Pinned [2]:
  00000000   131072 0001 0001 00000000 P uncached
  00020000  4096000 0041 0000 00000000 P uncached (name: 1)
Plane [1]:
  CNTR: c0000000 # enabled | gamma
  STRIDE: 00001400
  SIZE: 03ff04ff
  POS: 00000000
  ADDR: 05000000

Suggests that we did not clear the DSPBCNTR prior to unpinning the
framebuffer and reusing the GTT space. Impossible! Unless our DPMS
bookkeeping ran afoul again...

In the meantime add an assertion that the plane is decoupled from the
framebuffer prior to release.

Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: fix typo in function name

Fix function name in comments, a left-over from when i965_reset was
renamed to i915_reset.

Signed-off-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: split out pll divider code

This cleans up the mode set path a little further, making it easier to
extend for future platforms.

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
[danvet: shut up stupid gcc warning about potential use of
un-initlized fp2]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: split 9xx refclk & sdvo tv code out

Makes the mode set routine a little cleaner and easier to extend.

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

agp/intel: Add pci id for hostbridge from has/qemu

This is needed to run the simulator.

Cc: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
[danvet: added a comment in case people wonder what it's for.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: there is no pipe CxSR on ironlake

After checking the specs and discussing with Jesse, turns out CxSR is not
available on Ironlake and gen5, and its advertisement on the device
description is misleading.

Acked-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: Only look for matching clocks for LVDS downclock

This patch enforces that the downclock clock source is the same as the preferred
clock source for LVDS. This fixes a bug where the driver chooses a downclock
clock source with a different P than the preferred mode clock source. This
happened even if the preferred clock source implemented an acceptable rate for
the downclock. The result of this bug is that downclock is disabled.

Signed-off-by: Sean Paul <seanpaul@chromium.org>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/i915: Silence _DSM errors

<@ajax> mjg59: how concerned should i be about [drm:intel_dsm_pci_probe]
        *ERROR* failed to get supported _DSM functions ?
<@mjg59> ajax: Entirely unconcerned

Signed-off-by: Adam Jackson <ajax@redhat.com>
Acked-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>

drm/nouveau/pm: fix build with HWMON off

Reported-by: Randy Dunlap <rdunlap@xenotime.net>
Signed-off-by: Dave Airlie <airlied@redhat.com>

gma500: silence gcc warnings in mid_get_vbt_data()

Add a cast here to silence a Gcc warning.
drivers/gpu/drm/gma500/mid_bios.c:214:28: warning:
  cast from pointer to integer of different size [-Wpointer-to-int-cast]

Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Acked-by: Alan Cox <alan@linux.intel.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm/ttm: fix condition (and vs or)

The "if (!p && !p->dev)" condition isn't right because || was intended
instead of &&.  But actually, "p" is the list cursor and so it's always
non-NULL and we can just remove that bit.  We can remove the another
similar check as well.

Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Acked-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm/radeon: double lock typo in radeon_vm_bo_rmv()

The second lock should be an unlock or it causes a deadlock.

Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm/radeon: use after free in radeon_vm_bo_add()

"bo_va" is dereferenced in the error message.

Signed-off-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm/sis|via: don't return stack garbage from free_mem ioctl

Fallout from my "kill drm_sman" refactor. Unfortunately gcc seems to
have failed me and not warned about this.

Tested-by: Jakob Bornecrantz <wallbraker@gmail.com> (on via)
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm/radeon/kms: remove pointless CS flags priority struct

Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: Christian König <deathsimple@vodafone.de>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm/radeon/kms: check if vm is supported in VA ioctl

Add a VM manager enabled field and use it to check if
vm is enabled.

Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: jglisse@redhat.com
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm: introduce drm_can_sleep and use in intel/radeon drivers. (v2)

So we have a few places where the drm drivers would like to sleep to
be nice to the system, mainly in the modesetting paths, but we also
have two cases were atomic modesetting must take place, panic writing
and kernel debugger. So provide a central inline to determine if a
sleep or delay should be used and use this in the intel and radeon drivers.

v2: drop intel_drv.h MSLEEP macro, nobody uses it.

Based on patch from Michel Dänzer <michel.daenzer@amd.com>

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=43941

Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>

radeon: Fix disabling PCI bus mastering on big endian hosts.

It would previously write basically random bits to PCI configuration space...
Not very surprising that the GPU tended to stop responding completely. The
resulting MCE even froze the whole machine sometimes.

Now resetting the GPU after a lockup has at least a fighting chance of
succeeding.

Cc: stable@vger.kernel.org
Signed-off-by: Michel Dänzer <michel.daenzer@amd.com>
Reviewed-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

ttm: fix agp since ttm tt rework

ttm tt rework modified the way we allocate and populate the
ttm_tt structure, the AGP side was missing some bit to properly
work. Fix those and fix radeon and nouveau AGP support.

Tested on radeon only so far.

Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Reviewed-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

agp: Fix multi-line warning message whitespace

Signed-off-by: Tormod Volden <debian.tormod@gmail.com>
Reviewed-by: Corbin Simpson <MostAwesomeDude@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm/ttm/dma: Fix accounting error when calling ttm_mem_global_free_page and don't try to free freed pages.

The code to figure out how many pages to shrink the pool
ends up capping the 'count' at _manager->options.max_size - which is OK.
Except that the 'count' is also used when accounting for how many pages
are recycled - which we end up with the invalid values. This fixes
it by using a different value for the amount of pages to shrink.

On top of that we would free the cached page pool - which is nonsense
as they are deleted from the pool - so there are no free pages in that
pool..

Also we also missed the opportunity to batch the amount of pages
to free (similar to how ttm_page_alloc.c does it). This reintroduces
the code that was lost during rebasing.

Reviewed-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm/ttm/dma: Only call set_pages_array_wb when the page is not in WB pool.

Otherwise we are doing redundant work. Especially since the 'unbind'
and 'unpopulate' have been merged and nouveau driver ends up calling
it quite excessivly. On a GeForce 8600 GT with Gnome Shell (GNOME 3)
we end up spending about 54% CPU time in __change_page_attr_set_clr
checking the page flags.

The callgraph (annotated) looks as so before this patch:

    53.29%  gnome-shell  [kernel.kallsyms]                   [k] static_protections
            |
            --- static_protections
               |
               |--91.80%-- __change_page_attr_set_clr
               |          change_page_attr_set_clr
               |          set_pages_array_wb
               |          |
               |          |--96.55%-- ttm_dma_unpopulate
               |          |          nouveau_ttm_tt_unpopulate
               |          |          ttm_tt_destroy
               |          |          ttm_bo_cleanup_memtype_use
               |          |          ttm_bo_release
               |          |          kref_put
               |          |          ttm_bo_unref
               |          |          nouveau_gem_object_del
               |          |          drm_gem_object_free
               |          |          kref_put
               |          |          drm_gem_object_unreference_unlocked
               |          |          drm_gem_object_handle_unreference_unlocked.part.1
               |          |          drm_gem_handle_delete
               |          |          drm_gem_close_ioctl
               |          |          drm_ioctl
               |          |          do_vfs_ioctl
               |          |          sys_ioctl
               |          |          system_call_fastpath
               |          |          __GI___ioctl
               |          |
               |           --3.45%-- ttm_dma_pages_put
               |                     ttm_dma_page_pool_free
               |                     ttm_dma_unpopulate
               |                     nouveau_ttm_tt_unpopulate
               |                     ttm_tt_destroy
               |                     ttm_bo_cleanup_memtype_use
               |                     ttm_bo_release
               |                     kref_put
               |                     ttm_bo_unref
               |                     nouveau_gem_object_del
               |                     drm_gem_object_free
               |                     kref_put
               |                     drm_gem_object_unreference_unlocked
               |                     drm_gem_object_handle_unreference_unlocked.part.1
               |                     drm_gem_handle_delete
               |                     drm_gem_close_ioctl
               |                     drm_ioctl
               |                     do_vfs_ioctl
               |                     sys_ioctl
               |                     system_call_fastpath
               |                     __GI___ioctl
               |
                --8.20%-- change_page_attr_set_clr
                          set_pages_array_wb
                          |
                          |--93.76%-- ttm_dma_unpopulate
                          |          nouveau_ttm_tt_unpopulate
                          |          ttm_tt_destroy
                          |          ttm_bo_cleanup_memtype_use
                          |          ttm_bo_release
                          |          kref_put
                          |          ttm_bo_unref
                          |          nouveau_gem_object_del
                          |          drm_gem_object_free
                          |          kref_put
                          |          drm_gem_object_unreference_unlocked
                          |          drm_gem_object_handle_unreference_unlocked.part.1
                          |          drm_gem_handle_delete
                          |          drm_gem_close_ioctl
                          |          drm_ioctl
                          |          do_vfs_ioctl
                          |          sys_ioctl
                          |          system_call_fastpath
                          |          __GI___ioctl
                          |
                           --6.24%-- ttm_dma_pages_put
                                     ttm_dma_page_pool_free
                                     ttm_dma_unpopulate
                                     nouveau_ttm_tt_unpopulate
                                     ttm_tt_destroy
                                     ttm_bo_cleanup_memtype_use
                                     ttm_bo_release
                                     kref_put
                                     ttm_bo_unref
                                     nouveau_gem_object_del
                                     drm_gem_object_free
                                     kref_put
                                     drm_gem_object_unreference_unlocked
                                     drm_gem_object_handle_unreference_unlocked.part.1
                                     drm_gem_handle_delete
                                     drm_gem_close_ioctl
                                     drm_ioctl
                                     do_vfs_ioctl
                                     sys_ioctl
                                     system_call_fastpath
                                     __GI___ioctl

and after this patch all of that disappears.

Signed-off-by: Konrad Rzeszutek Wilk <konrad.wilk@oracle.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm/radeon/kms: sync across multiple rings when doing bo moves v3

We need to synchronize across rings when doing a bo move to make
sure we the buffer is idle if it's in use by a different ring than
the ring doing the move.

v2: fix fence setup for bo moves

v3: add missing ring lock/unlock

Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm/radeon/kms: Add support for multi-ring sync in CS ioctl (v2)

Use semaphores to sync buffers across rings in the CS
ioctl.  Add a reloc flag to allow userspace to skip
sync for buffers.

agd5f: port to latest CS ioctl changes.

v2: add ring lock/unlock to make sure changes hit the ring.

Signed-off-by: Christian König <deathsimple@vodafone.de>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm/radeon: GPU virtual memory support v22

Virtual address space are per drm client (opener of /dev/drm).
Client are in charge of virtual address space, they need to
map bo into it by calling DRM_RADEON_GEM_VA ioctl.

First 16M of virtual address space is reserved by the kernel.

Once using 2 level page table we should be able to have a small
vram memory footprint for each pt (there would be one pt for all
gart, one for all vram and then one first level for each virtual
address space).

Plan include using the sub allocator for a common vm page table
area and using memcpy to copy vm page table in & out. Or use
a gart object and copy things in & out using dma.

v2: agd5f fixes:
- Add vram base offset for vram pages.  The GPU physical address of a
vram page is FB_OFFSET + page offset.  FB_OFFSET is 0 on discrete
cards and the physical bus address of the stolen memory on
integrated chips.
- VM_CONTEXT1_PROTECTION_FAULT_DEFAULT_ADDR covers all vmid's >= 1

v3: agd5f:
- integrate with the semaphore/multi-ring stuff

v4:
- rebase on top ttm dma & multi-ring stuff
- userspace is now in charge of the address space
- no more specific cs vm ioctl, instead cs ioctl has a new
  chunk

v5:
- properly handle mem == NULL case from move_notify callback
- fix the vm cleanup path

v6:
- fix update of page table to only happen on valid mem placement

v7:
- add tlb flush for each vm context
- add flags to define mapping property (readable, writeable, snooped)
- make ring id implicit from ib->fence->ring, up to each asic callback
  to then do ring specific scheduling if vm ib scheduling function

v8:
- add query for ib limit and kernel reserved virtual space
- rename vm->size to max_pfn (maximum number of page)
- update gem_va ioctl to also allow unmap operation
- bump kernel version to allow userspace to query for vm support

v9:
- rebuild page table only when bind and incrementaly depending
  on bo referenced by cs and that have been moved
- allow virtual address space to grow
- use sa allocator for vram page table
- return invalid when querying vm limit on non cayman GPU
- dump vm fault register on lockup

v10: agd5f:
- Move the vm schedule_ib callback to a standalone function, remove
  the callback and use the existing ib_execute callback for VM IBs.

v11:
- rebase on top of lastest Linus

v12: agd5f:
- remove spurious backslash
- set IB vm_id to 0 in radeon_ib_get()

v13: agd5f:
- fix handling of RADEON_CHUNK_ID_FLAGS

v14:
- fix va destruction
- fix suspend resume
- forbid bo to have several different va in same vm

v15:
- rebase

v16:
- cleanup left over of vm init/fini

v17: agd5f:
- cs checker

v18: agd5f:
- reworks the CS ioctl to better support multiple rings and
VM.  Rather than adding a new chunk id for VM, just re-use the
IB chunk id and add a new flags for VM mode.  Also define additional
dwords for the flags chunk id to define the what ring we want to use
(gfx, compute, uvd, etc.) and the priority.

v19:
- fix cs fini in weird case of no ib
- semi working flush fix for ni
- rebase on top of sa allocator changes

v20: agd5f:
- further CS ioctl cleanups from Christian's comments

v21: agd5f:
- integrate CS checker improvements

v22: agd5f:
- final cleanups for release, only allow VM CS on cayman

Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm: make DRM_UNLOCKED ioctls with their own mutex

drm_getclient, drm_getstats and drm_getmap (with a few minor
adjustments) do not need global mutex, so fix that and
make the said ioctls DRM_UNLOCKED. Details:

  drm_getclient: the only thing that should be protected here
  is dev->filelist and that is already protected everywhere with
  dev->struct_mutex.

  drm_getstats: there is no need for any mutex here because the
  loop runs through quasi-static (set at load time only)
  data, and the actual count access is done with atomic_read()

  drm_getmap already uses dev->struct_mutex to protect
  dev->maplist, which also used to protect the same structure
  everywhere else except at three places:
  * drm_getsarea, which doesn't grab *any* mutex before
    touching dev->maplist (so no drm_global_mutex doesn't help
    here either; different issue for a different patch).
    However, drivers seem to call it only at
    initialization time so it probably doesn't matter
  * drm_master_destroy, which is called from drm_master_put,
    which in turn is protected with dev->struct_mutex
    everywhere else in drm module, so we are good here too.
  * drm_getsareactx, which releases the dev->struct_mutex
    too early, but this patch includes the fix for that.

v2: * incorporate comments received from Daniel Vetter
    * include the (long) explanation above to make it clear what
      we are doing (and why), also at Daniel Vetter's request
    * tighten up mutex grab/release locations to only
      encompass real critical sections, rather than some
      random code around them

Signed-off-by: Ilija Hadzic <ihadzic@research.bell-labs.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm: no need to hold global mutex for static data

drm_getcap and drm_version ioctls only reads static data,
there is no need to protect them with drm_global_mutex,
so make them DRM_UNLOCKED

Signed-off-by: Ilija Hadzic <ihadzic@research.bell-labs.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm/radeon/benchmark: common modes sweep ignores 640x480@32

Sweep common_modes array should start with index 0.

Signed-off-by: Chen Jie <chenj@lemote.com>
Reviewed-by: Ilija Hadzic <ihadzic@research.bell-labs.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm/radeon/kms: disable writeback on pre-R300 asics

We often end up missing fences on older asics with
writeback enabled which leads to delays in the userspace
accel code, so just disable it by default on those asics.

Reported-by: Helge Deller <deller@gmx.de>
Reported-by: Dave Airlie <airlied@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Cc: stable@vger.kernel.org
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm/radeon: allocate semaphore from the ib pool

This allow to share the ib pool with semaphore and avoid
having more bo around.

Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm/radeon: make ib size variable

This avoid to waste ib pool size and avoid a bunch of wait for
previous ib to finish.

Signed-off-by: Jerome Glisse <jglisse@redhat.com>
Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm/radeon/kms: add missing ring ready check in sync tests

Signed-off-by: Alex Deucher <alexander.deucher@amd.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm: add support for private planes

In cases where the scanout hw is sufficiently similar between "overlay"
and traditional crtc layers, it might be convenient to allow the driver
to create internal drm_plane helper objects used by the drm_crtc
implementation, rather than duplicate code between the plane and crtc.
A private plane is not exposed to userspace.

Signed-off-by: Rob Clark <rob@ti.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm: disconnect plane from fb/crtc when disabled

Since plane->fb and plane->crtc are set in drm_mode_setplane()
after update_plane(), They should be cleared after disable().

Signed-off-by: Rob Clark <rob@ti.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm/i915: only set the intel_crtc DPMS mode to on if the mode set succeeded

If it failed, leave it in the "off" state.

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Keith Packard <keithp@keithp.com>

drm/i915: don't disable a PCH DPLL that's in use

If a PCH pipe PLL is being used by transcoder C, don't disable it.

Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Keith Packard <keithp@keithp.com>

drm/i915: don't bail out of intel_wait_ring_buffer too early

In the pre-gem days with non-existing hangcheck and gpu reset code,
this timeout of 3 seconds was pretty important to avoid stuck
processes.

But now we have the hangcheck code in gem that goes to great length
to ensure that the gpu is really dead before declaring it wedged.

So there's no need for this timeout anymore. Actually it's even harmful
because we can bail out too early (e.g. with xscreensaver slip)
when running giant batchbuffers. And our code isn't robust enough
to properly unroll any state-changes, we pretty much rely on the gpu
reset code cleaning up the mess (like cache tracking, fencing state,
active list/request tracking, ...).

With this change intel_begin_ring can only fail when the gpu is
wedged, and it will return -EAGAIN (like wait_request in case the
gpu reset is still outstanding).

v2: Chris Wilson noted that on resume timers aren't running and hence
we won't ever get kicked out of this loop by the hangcheck code. Use
an insanely large timeout instead for the HAS_GEM case to prevent
resume bugs from totally hanging the machine.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Acked-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Signed-off-by: Keith Packard <keithp@keithp.com>

drm/i915: kicking rings stuck on semaphores considered harmful

If our semaphore logic gets confused and we have a ring stuck waiting
for one, there's a decent chance it'll just execute garbage when being
kicked. Also, kicking the ring obscures the place where the error
first occured, making error_state decoding much harder.

So drop this an let gpu reset handle this mess in a clean fashion.

In contrast, kicking rings stuck on MI_WAIT is rather harmless, at
worst there'll be a bit of screen-flickering. There's also old
broken userspace out there which needs this as a  work-around.

Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Chris Wilson <chris@hchris-wilson.co.uk>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Keith Packard <keithp@keithp.com>

drm/i915: Add support for resetting the SO write pointers on gen7.

These registers are automatically incremented by the hardware during
transform feedback to track where the next streamed vertex output
should go.  Unlike the previous generation, which had a packet for
setting the corresponding registers to a defined value, gen7 only has
MI_LOAD_REGISTER_IMM to do so.  That's a secure packet (since it loads
an arbitrary register), so we need to do it from the kernel, and it
needs to be settable atomically with the batchbuffer execution so that
two clients doing transform feedback don't stomp on each others'
state.

Instead of building a more complicated interface involcing setting the
registers to a specific value, just set them to 0 when asked and
userland can tweak its pointers accordingly.

Signed-off-by: Eric Anholt <eric@anholt.net>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Reviewed-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Keith Packard <keithp@keithp.com>

drm/i915: Make the fallback IRQ wait not sleep.

The waits we do here are generally so short that sleeping is a bad
idea unless we have an IRQ to wake us up.  Improves regression test
performance from 18 minutes to 3.5 minutes on gen7, which is now
consistent with the previous generation.

Signed-off-by: Eric Anholt <eric@anholt.net>
Tested-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Acked-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Keith Packard <keithp@keithp.com>

drm/i915: Work around gen7 BLT ring synchronization issues.

Previous to this commit, testing easily reproduced a failure where the
seqno would apparently arrive after the IRQ associated with it, with test programs as simple as:

for (;;) {
    glCopyPixels(0, 0, 1, 1);
    glFinish();
}

Various workarounds we've seen for previous generations didn't work to
fix this issue, so until new information comes in, replace the IRQ
waits on the BLT ring with polling.

Signed-off-by: Eric Anholt <eric@anholt.net>
Tested-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Acked-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Keith Packard <keithp@keithp.com>

drm/i915: Do the fallback non-IRQ wait in ring throttle, too.

As a workaround for IRQ synchronization issues in the gen7 BLT ring,
we want to turn the two wait functions into polling loops.

Signed-off-by: Eric Anholt <eric@anholt.net>
Tested-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Acked-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Keith Packard <keithp@keithp.com>

drm/i915: Set two chicken bits implicated in missed IRQs on Ivybridge.

They don't fix our problems alone, but we're told to set them.

Signed-off-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Keith Packard <keithp@keithp.com>

drm/i915: add color key support v4

Add new ioctls for getting and setting the current destination color
key.  This allows for simple overlay display control by matching a color
key value in the primary plane before blending the overlay on top.

v2: remove unnecessary mutex acquire/release around reg accesses
v3: add support for full color key management
v4: fix copy & paste bug in snb_get_colorkey
    don't bother checking min/max values against docs as the docs are likely
    wrong (how could we handle 10bpc surface formats?)

Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>

drm/i915: track sprite coverage and disable primary plane if possible

To save power when the sprite is full screen, we can disable the primary
plane on the same pipe.  Track the sprite status and enable/disable the
primary opportunistically.

v2: remove primary plane enable/disable hooks; they're identical

Reviewed-by: Daniel Vetter <daniel@ffwll.ch>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Keith Packard <keithp@keithp.com>

drm/i915: add SNB and IVB video sprite support v6

The video sprites support various video surface formats natively and can
handle scaling as well.  So add support for them using the new DRM core
sprite support functions.

v2: use drm specific fourcc header and defines
v3: address Daniel's comments:
  - don't take struct mutex around register access (only needed for
    regs in the GT power well)
  - don't hold struct mutex across vblank waits
  - fix up update_plane API (pass obj instead of GTT offset)
  - add interlaced defines for sprite regs
  - drop unnecessary 'reg' variables
  - comment double buffered reg flushing
  Also fix w/h confusion when writing the scaling reg.
v4: more fixes, address more comments from Daniel, and include Hai's fix
  - prevent divide by zero in scaling calculation (Hai Lan)
  - update to Ville's new DRM_FORMAT_* types
  - fix sprite watermark handling (calc based on CRTC size, separate
    from normal display wm)
  - remove private refcounts now that the fb cleanups handles things
v5: add linear surface support
v6: remove color key clearing & setting from update_plane

For this version, I tested DPMS since it came up in the last review;
DPMS off/on works ok when a video player is working under X, but for
power saving we'll probably want to do something smarter.  I'll leave
that for a separate patch on top.  Likewise with the refcounting/fb
layer handling, which are really separate cleanups.

Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Keith Packard <keithp@keithp.com>

drm/i915: Clean up multi-threaded forcewake patch

We learned that the ECOBUS register was inside the GT power well, and
so *did* need force wake to be read, so it gets removed from the list
of 'doesn't need force wake' registers.

That means the code reading ECOBUS after forcing the mt_force_wake
function to be called needs to use I915_READ_NOTRACE; it doesn't need
to do more force wake fun as it's already done it manually.

This also adds a comment explaining why the MT forcewake testing code
only needs to call mt_forcewake_get/put and not disable RC6 manually
-- the ECOBUS read will return 0 if the device is in RC6 and isn't
using MT forcewake, causing the test to work correctly.

Signed-off-by: Keith Packard <keithp@keithp.com>
Cc: Jesse Barnes <jbarnes@virtuousgeek.org>

drm/i915: drpc debugfs update for gen6

Many of the old fields from Ironlake have gone away. Strip all those
fields, and try to update to fields people care about. RC information
isn't exactly ideal anymore. All we can guarantee when we read the
register is that we're not using forcewake, ie. the software isn't
forcing the hardware to stay awake. The downside is that in doing this
we may wait a while and that causes an unnaturally idle state on the
GPU.

Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=42578
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Keith Packard <keithp@keithp.com>

drm/i915: Update GEN6_RP_CONTROL definitions

This matches the modern specs more accurately.

This will be used by the following patch to fix the way we display RC
status.

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Reviewed-by: Eugeni Dodonov <eugeni.dodonov@intel.com>
Signed-off-by: Keith Packard <keithp@keithp.com>

drm/i915: Force sync command ordering (Gen6+)

The docs say this is required for Gen7, and since the bit was added for
Gen6, we are also setting it there pit pf paranoia. Particularly as
Chris points out, if PIPE_CONTROL counts as a 3d state packet.

This was found through doc inspection by Ken and applies to Gen6+;

Reported-by: Kenneth Graunke <kenneth@whitecape.org>
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Eric Anholt <eric@anholt.net>
Signed-off-by: Keith Packard <keithp@keithp.com>

drm/i915: relative_constants_mode race fix

dev_priv keeps track of the current addressing mode that gets set at
execbuffer time. Unfortunately the existing code was doing this before
acquiring struct_mutex which leaves a race with another thread also
doing an execbuffer. If that wasn't bad enough, relocate_slow drops
struct_mutex which opens a much more likely error where another thread
comes in and modifies the state while relocate_slow is being slow.

The solution here is to just defer setting this state until we
absolutely need it, and we know we'll have struct_mutex for the
remainder of our code path.

v2: Keith noticed a bug in the original patch.

Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Keith Packard <keithp@keithp.com>

Merge branch 'drm-radeon-testing' of ../drm-radeon-next into drm-core-next

This merges the evergreen HDMI audio support.

* 'drm-radeon-testing' of ../drm-radeon-next:
  drm/radeon/kms: define TMDS/LVTM HDMI enabling bits
  drm/radeon/kms: workaround invalid AVI infoframe checksum issue
  drm/radeon/kms: setup HDMI mode on Evergreen encoders
  drm/radeon/kms: support for audio on Evergreen
  drm/radeon/kms: minor HDMI audio cleanups
  drm/radeon/kms: do not force DVI mode on DCE4 if audio is on
ridge

Conflicts:
drivers/gpu/drm/radeon/evergreen.c

drm/radeon/kms: define TMDS/LVTM HDMI enabling bits

The names has been taken from free M76 specs.

Signed-off-by: Rafał Miłecki <zajec5@gmail.com>
Reviewed-by: Alex Deucher <alexdeucher@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>

drm/radeon/kms: workaround invalid AVI infoframe checksum issue

This change was verified to fix both issues with no video I've
investigated. I've also checked checksum calculation with fglrx on:
RV620, HD54xx, HD5450, HD6310, HD6320.

Cc: stable@vger.kernel.org
Signed-off-by: Rafał Miłecki <zajec5@gmail.com>
Signed-off-by: Dave Airlie <airlied@redhat.com>