... and put them to so good use.

Note that there's functional change in vlv clock gating code, we now
no longer spuriously read back the current value of the bit. According
to Bspec the high bits should always read zero, so ORing this in
should have no effect.
Reviewed-by: NJesse Barnes <jbarnes@virtuousgeek.org>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NEugeni Dodonov <eugeni.dodonov@intel.com>
Signed-Off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Rename obj->tiling_changed to obj->fence_dirty so that it is clear that
it flags when the parameters for an active fence (including the
no-fence) register are changed.

Also, do not set this flag when the object does not have a fence
register allocated currently and the gpu does not depend upon the
unfence. This case works exactly like when a tiled object lost its
fence and hence does not need additional handling for the tiling
change in the code.

v2: Use fence_dirty to better express what the flag tracks and add a few
more details to the comments to serve as a reminder of how the GPU also
uses the unfenced register slot.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
[danvet: Add some bikeshed to the commit message about the stricter
use of fence_dirty.]
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

As with one of the earlier patches in the series, we're forced to cast
for copy_[to|from]_user. Again because of the nature of the GEN x86
exclusivity, this should be safe.
Signed-off-by: NBen Widawsky <benjamin.widawsky@intel.com>
[danvet: Added some bikeshed.]
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

We can also take advantage of the new 'no retire' mode for seqno waiting
to avoid having to take a reference on the old fence object whilst
flushing an existing fence.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Now that we have a routine that is able to clear the fences as well as
setup up the register for a tiled object, remove the surplus routines to
clear the fences.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

One clarification that we make is to the existing semantics of
obj->tiling_changed to only mean that we need to update an associated
fence register (including the NO_FENCE when executing an untiled but
fenced GPU command). If we do not have a fence register or pending
fenced GPU access for the object (after put_fence() for example), then
we can clear the tiling_changed flag as any fence will necessarily be
rewritten upon acquisition.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Update the existing architecture specific fence writing routines to
either update the fence to point to a tiled object or to clear them in
preparation to remove the other fence writing routes.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

As i915_wait_request() will first check for an already passed seqno,
doing it also in the caller is a waste of space for a cold path.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

As the fences are stored in LRU order, we can simply reuse the oldest if
we do not have an unused register.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

As we now never pipeline a fence update, obj->last_fenced_ring is always
the same as the obj->ring whenever obj->last_fenced_seqno is active, so
remove it.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

As we now no longer track a pipelined fence change, we never use
ring->setup_seqno and can kill it.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Step 2 is then to replace the pipelined parameter with NULL and perform
constant folding to remove dead code.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

We never succeeded in getting pipelined fencing to work (unresolved
spurious GPU hangs), so begin the process of dismantling and removal
the broken code.

Step 1 is the removal of the pipeline parameter to get_fence().
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

For some reason snb has 2 fields to set ppgtt cacheability. This one
here does not exist on gen7.

This might explain why ppgtt wasn't a win on snb like on ivb - not
enough pte caching.

v2: Fixup rebase fail.
Reviewed-by: NBen Widawsky <ben@bwidawsk.net>
Signed-Off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Bspec says that we need to set this: vol1c.3 "Blitter Command
Streamer", Section 1.1.2.1 "GAB_CTL_REG - GAB Unit Control Register".

We don't really rely on pagefaults, but who knows what this all
affects.
Reviewed-by: NBen Widawsky <ben@bwidawsk.net>
Signed-Off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

... we will botch up the bit17 swizzling. Furthermore tiled pwrite is
a (now) unused slowpath, so no one really cares.

This fixes the last swizzling issues I have with i-g-t on my bit17
swizzling i915G. No regression, it's been broken since the dawn of
gem, but it's nice for regression tracking when really _all_ i-g-t
tests work.

Actually this is not true, Chris Wilson noticed while reviewing this
patch that the commit

commit d9e86c0e
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Wed Nov 10 16:40:20 2010 +0000

    drm/i915: Pipelined fencing [infrastructure]

contained a functional change that broke things.
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-Off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Waiting for seqno-1 in our object synchronization code is an
implementation detail given how we've decided to do the waits within the
rest of our code.
Requested-by: NDaniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: NBen Widawsky <benjamin.widawsky@intel.com>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

This fixes a long standing issue where emitting the semaphore updates
may have failed, but we've already updated our internal data structure.
Reported-by: NDaniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: NBen Widawsky <benjamin.widawsky@intel.com>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

When I extracted the synchronization code for implementing semaphorified
pageflips (74f5f6e0), I neglected the non pipelined case which also
calls this code. The modesetting code wants to make sure the object has
finished rendering to the frame before configuring the scanout (ie.
non-pipelined case).

As a result of a follow on discussion on IRC, I've decided to add a
comment about the function itself which received much inspiration from
Chris as well. So really, this patch was ghost-written by Chris :).
Reported-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NBen Widawsky <benjamin.widawsky@intel.com>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Tested-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Similar to allowing a buffer to be simultaneously read by the GPU and
through the GTT, we wish to allow readback of the pages through the CPU
domain whilst they are also being read by the GPU. Domain coherency
is managed by allowing multiple readers, but only a single writer.

This is used by mesa for its program cache which it may search for every
new program every frame and then renews should it need to add. During
renewal, mesa copies the program bo currently executing through a CPU
mapping onto the new bo. This patch allows the search and that copy to
proceed without causing a stall on the current batch.

Testcase: i-g-t/tests/gem_cpu_concurrent_blit
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NDaniel Vetter <daniel.vetter@ffwll.ch>
Signed-Off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

In theory this will have performance and power improvements. Performance
because we don't need to stall when the scanout BO is busy, and power
because we don't have to stall when the BO is busy (and the ring can
even go to sleep if the HW supports it).

v2:
squash 2 patches into 1 (me)
un-inline the enable_semaphores function (Daniel)
remove comment about SNB hangs from i915_gem_object_sync (Chris)
rename intel_enable_semaphores to i915_semaphore_is_enabled (me)
removed page flip comment; "no why" (Chris)

To address other comments from Daniel (irc):
update the comment to say 'vt-d is crap, don't enable semaphores'
  - I think you misinterpreted Chris' comment, it already exists.
checking out whether we can pageflip on the render ring on ivb (didn't
work on early silicon)
  - We don't want to enable workarounds for early silicon unless we have
    to.
  - I can't find any references in the docs about this.
optionally use it if the fb is already busy on the render ring
  - This should be how the code already worked, unless I am
    misunderstanding your meaning.
Signed-off-by: NBen Widawsky <ben@bwidawsk.net>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

By simplifying the rules to calling get_fence when writing to the
through the GTT in a tiled manner, and calling put_fence before writing
to the object through the GTT in a linear manner, the code becomes
clearer and there is less chance of making a mistake.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NDaniel Vetter <daniel.vetter@ffwll.ch>
[danvet: fixed up conflict with ppgtt code and spelling in a new
comment.]
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

This fixes a resume regression introduced in

commit 7dd49065
Author: Chris Wilson <chris@chris-wilson.co.uk>
Date:   Wed Mar 21 10:48:18 2012 +0000

    drm/i915: Mark untiled BLT commands as fenced on gen2/3

which fixed fencing tracking for untiled blt commands.

A side effect of that patch was that now also untiled objects have a
non-zero obj->last_fenced_seqno to track when a fence can be set up
after a pipelined tiling change. Unfortunately this was only cleared
by the fence setup and teardown code, resulting in tons of untiled but
inactive objects with non-zero last_fenced_seqno.

Now after resume we completely reset the seqno tracking, both on the
driver side (by setting dev_priv->next_seqno = 1) and on the hw side
(by allocating a new hws page, which contains the seqnos). Hilarity
and indefinite waits ensued from the stale seqnos in
obj->last_fenced_seqno from before the suspend.

The fix is to properly clear the fencing tracking state like we
already do for the normal gpu rendering while moving objects off the
active list.
Reported-and-tested-by: N"Rafael J. Wysocki" <rjw@sisk.pl>
Cc: Jiri Slaby <jslaby@suse.cz>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-Off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Ums is already disabled, but on ilk we can additionally disable gem
initialization when using user mode setting. Upstream never support
ilk without kernel modesetting and not even the RHEL ilk ums backport
needs gem - that driver is based on xf86-video-intel version 2.2,
which is pre-gem.
Reviewed-by: NAdam Jackson <ajax@redhat.com>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-Off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

The BLT commands on gen2/3 utilize the fence registers and so we cannot
modify any fences for the object whilst those commands are in flight.
Currently we marked tiled commands as occupying a fence, but forgot to
restrict the untiled commands from preventing a fence being assigned
before they were completed.

One side-effect is that we ten have to double check that a fence was
allocated for a fenced buffer during move-to-active.
Reported-by: NJiri Slaby <jirislaby@gmail.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=43427
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=47990Reviewed-by: NDaniel Vetter <daniel.vetter@ffwll.ch>
Testcase: i-g-t/tests/gem_tiled_after_untiled_blt
Tested-by: NDaniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: stable@kernel.org
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

The ppgtt page directory lives in a snatched part of the gtt pte
range. Which naturally gets cleared on hibernate when we pull the
power. Suspend to ram (which is what I've tested) works because
despite the fact that this is a mmio region, it is actually back by
system ram.

Fix this by moving the page directory setup code to the ppgtt init
code (which gets called on resume).

This fixes hibernate on my ivb and snb.
Reviewed-by: NBen Widawsky <ben@bwidawsk.net>
Signed-Off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Haven't seen this yet, but it doesn't hurt.
Signed-off-by: NJesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

Beside helping the compiler untangle this maze they double-up as
documentation for which parts of the code aren't performance-critical
but just around to keep old (but already dead-slow) userspace from
breaking.
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-Off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

The issue is that with inline clflushing the clflushing isn't properly
swizzled. Fix this by
- always clflushing entire 128 byte chunks and
- unconditionally flush before writes when swizzling a given page.
  We could be clever and check whether we pwrite a partial 128 byte
  chunk instead of a partial cacheline, but I've figured that's not
  worth it.

Now the usual approach is to fold this into the original patch series, but
I've opted against this because
- this fixes a corner case only very old userspace relies on and
- I'd like to not invalidate all the testing the pwrite rewrite has gotten.

This fixes the regression notice by tests/gem_tiled_partial_prite_pread
from i-g-t. Unfortunately it doesn't fix the issues with partial pwrites to
tiled buffers on bit17 swizzling machines. But that is also broken without
the pwrite patches, so likely a different issue (or a problem with the
testcase).

v2: Simplify the patch by dropping the overly clever partial write
logic for swizzled pages.
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-Off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

drm/i915 wants to read/write more than one page in its fastpath
and hence needs to prefault more than PAGE_SIZE bytes.

Add new functions in filemap.h to make that possible.

Also kill a copy&pasted spurious space in both functions while at it.

v2: As suggested by Andrew Morton, add a multipage parameter to both
functions to avoid the additional branch for the pagemap.c hotpath.
My gcc 4.6 here seems to dtrt and indeed reap these branches where not
needed.

v3: Becaus I couldn't find a way around adding a uaddr += PAGE_SIZE to
the filemap.c hotpaths (that the compiler couldn't remove again),
let's go with separate new functions for the multipage use-case.

v4: Adjust comment to CodingStlye and fix spelling.
Acked-by: NAndrew Morton <akpm@linux-foundation.org>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

While moving around things, this two functions slowly grew out of any
sane bounds. So extract a few lines that do the copying and
clflushing. Also add a few comments to explain what's going on.

v2: Again do s/needs_clflush/needs_clflush_after/ in the write paths
as suggested by Chris Wilson.
Tested-by: NChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-Off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

It's around 20% faster.
Tested-by: NChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-Off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

It's too expensive to move it around just for that pwrite, especially
when we're trashing on the mappable gtt part like crazy.
Tested-by: NChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-Off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

In micro-benchmarking of the usual pwrite use-pattern of alternating
pwrites with gtt domain reads from the gpu, this yields around 30%
improvement of pwrite throughput across all buffers size. The trick is
that we can avoid clflush cachelines that we will overwrite completely
anyway.

Furthermore for partial pwrites it gives a proportional speedup on top
of the 30% percent because we only clflush back the part of the buffer
we're actually writing.

v2: Simplify the clflush-before-write logic, as suggested by Chris
Wilson.

v3: Finishing touches suggested by Chris Wilson:
- add comment to needs_clflush_before and only set this if the bo is
  uncached.
- s/needs_clflush/needs_clflush_after/ in the write paths to clearly
  differentiate it from needs_clflush_before.
Tested-by: NChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-Off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

The pagemap.h prefault helpers do the prefaulting by simply writing
some data into every page. Hence we should not prefault when we're not
yet commited to to actually writing data to userspace. The problem is
now that
- we can't prefault while holding dev->struct_mutex for we could
  deadlock with our own pagefault handler
- we need to grab dev->struct_mutex before copying to sync up with any
  outsanding gpu writes.

Therefore only prefault when we're dropping the lock the first time in
the pread slowpath - at that point we're committed to the write, don't
wait on the gpu anymore and hence won't return early (with e.g.
-EINTR).
Tested-by: NChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-Off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

With the proper prefault, it's extremely unlikely that we fall back
to the gtt slowpath.

So just kill it and use the shmem_pwrite path as fallback.

To further clean up the code, move the preparatory gem calls into the
respective pwrite functions. This way the gtt_fast->shmem fallback
is much more obvious.
Tested-by: NChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

This speeds up pwrite and pread from ~120 µs ro ~100 µs for
reading/writing 1mb on my snb (if the backing storage pages
are already pinned, of course).

v2: Chris Wilson pointed out a glaring page reference bug - I've
unconditionally dropped the reference. With that fixed (and the
associated reduction of dirt in dmesg) it's now even a notch faster.

v3: Unconditionaly grab a page reference when dropping
dev->struct_mutex to simplify the code-flow.
Tested-by: NChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

~120 µs instead fo ~210 µs to write 1mb on my snb. I like this.
Tested-by: NChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

No longer needed.
Tested-by: NChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>

This is obviously gonna slow down pread. But for a half-way realistic
micro-benchmark, it doesn't matter: Non-broken userspace reads back
data from the gpu once before the gpu again dirties it.

So all this ranged clflush tracking is just a waste of time.

No pread performance change (neglecting the dumb benchmark of
constantly reading the same data) measured.

As an added bonus, this avoids clflush on read on coherent objects.
Which means that partial preads on snb are now roughly 4x as fast.
This will be usefull for e.g. the libva encoder - when I finally get
around to fix that up.

v2: Properly sync with the gpu on LLC machines.
Tested-by: NChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NChris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: NDaniel Vetter <daniel.vetter@ffwll.ch>