Previously, it was possible for the GPU memory accesses to be swizzled
to try to optimize the fetches for tiled buffers. This swizzling was on
top of what the memory controller in the uncore already does.
With broadwell, we drop that GPU side swizzling, and the corresponding
initialization in 3 units (GAM, GT, DE). All those bits are reserved, as
specs put it:
Before Gen8, there was a historical configuration control field to
swizzle address bit[6] for in X/Y tiling modes. This was set in three
different places: TILECTL[1:0], ARB_MODE[5:4], and
DISP_ARB_CTL[14:13]"
For Gen8 the swizzle fields are all reserved, and the CPU's memory
controller performs all address swizzling modifications.
This also means that user space doesn't have to manually swizzle when
accessing tiled buffers from the CPU, and so we always return
I915_BIT_6_SWIZZLE_NONE from i915_gem_detect_bit_6_swizzle(), which
short-circuits the initialization of the registers mentionned above in
i915_gem_init_swizzling().
v2: Refine the explanation a bit more (Daniel)
v3: Make it BDW+ specific (Steve)
Cc: Steve Aarnio <steve.j.aarnio@linux.intel.com>
Signed-off-by: Damien Lespiau <damien.lespiau@intel.com>
[danvet: Keep the actual code to set the tiling bits for now, in case
some bios escaped to the wild that uses this - we'd need it for
fastboot.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
This migrates the fence tracking onto the existing seqno
infrastructure so that the later conversion to tracking via requests is
simplified.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Remove the rest of the references to drm_i915_private_t. No functional
changes.
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
[danvet: Drop hunk in i915_cmd_parser.c]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Assuming that all framebuffer related metadata is invariant simplifies
our userspace input data checking. And current userspace always first
updates the tiling of an object before creating a framebuffer with it.
This allows us to upconvert a check in pin_and_fence to a WARN.
In the future it should also be helpful to know which buffer objects
are potential scanout targets for e.g. frontbuffer rendering tracking
and similar things.
Note that SNA shipped for one prerelease with code which will be
broken through this patch. But users shouldn't notice since it's
purely an optimization and will transparently fall back to allocating
a new fb. i-g-t also had offending code (now fixed), but we don't
really care about breaking the test-suite.
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Grumpily-reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
No buffer overflows here, but better safe than sorry.
v2:
- Fixup the sizeof conversion, I've missed the pointer deref (Jani).
- Drop the redundant GFP_ZERO, kcalloc alreads memsets (Jani).
- Use kmalloc_array for the execbuf fastpath to avoid the memset
(Chris). I've opted to leave all other conversions as-is since they
aren't in a fastpath and dealing with cleared memory instead of
random garbage is just generally nicer.
Cc: Jani Nikula <jani.nikula@linux.intel.com>
Cc: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Jani Nikula <jani.nikula@intel.com>
[danvet: Drop the contentious kmalloc_array hunk in execbuf.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
As alluded to in several patches, and it will be reiterated later... A
VMA is an abstraction for a GEM BO bound into an address space.
Therefore it stands to reason, that the existing bind, and unbind are
the ones which will be the most impacted. This patch implements this,
and updates all callers which weren't already updated in the series
(because it was too messy).
This patch represents the bulk of an earlier, larger patch. I've pulled
out a bunch of things by the request of Daniel. The history is preserved
for posterity with the email convention of ">" One big change from the
original patch aside from a bunch of cropping is I've created an
i915_vma_unbind() function. That is because we always have the VMA
anyway, and doing an extra lookup is useful. There is a caveat, we
retain an i915_gem_object_ggtt_unbind, for the global cases which might
not talk in VMAs.
> drm/i915: plumb VM into object operations
>
> This patch was formerly known as:
> "drm/i915: Create VMAs (part 3) - plumbing"
>
> This patch adds a VM argument, bind/unbind, and the object
> offset/size/color getters/setters. It preserves the old ggtt helper
> functions because things still need, and will continue to need them.
>
> Some code will still need to be ported over after this.
>
> v2: Fix purge to pick an object and unbind all vmas
> This was doable because of the global bound list change.
>
> v3: With the commit to actually pin/unpin pages in place, there is no
> longer a need to check if unbind succeeded before calling put_pages().
> Make put_pages only BUG() after checking pin count.
>
> v4: Rebased on top of the new hangcheck work by Mika
> plumbed eb_destroy also
> Many checkpatch related fixes
>
> v5: Very large rebase
>
> v6:
> Change BUG_ON to WARN_ON (Daniel)
> Rename vm to ggtt in preallocate stolen, since it is always ggtt when
> dealing with stolen memory. (Daniel)
> list_for_each will short-circuit already (Daniel)
> remove superflous space (Daniel)
> Use per object list of vmas (Daniel)
> Make obj_bound_any() use obj_bound for each vm (Ben)
> s/bind_to_gtt/bind_to_vm/ (Ben)
>
> Fixed up the inactive shrinker. As Daniel noticed the code could
> potentially count the same object multiple times. While it's not
> possible in the current case, since 1 object can only ever be bound into
> 1 address space thus far - we may as well try to get something more
> future proof in place now. With a prep patch before this to switch over
> to using the bound list + inactive check, we're now able to carry that
> forward for every address space an object is bound into.
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
[danvet: Rebase on top of the loss of "drm/i915: Cleanup more of VMA
in destroy".]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Soon we want to gut a lot of our existing assumptions how many address
spaces an object can live in, and in doing so, embed the drm_mm_node in
the object (and later the VMA).
It's possible in the future we'll want to add more getter/setter
methods, but for now this is enough to enable the VMAs.
v2: Reworked commit message (Ben)
Added comments to the main functions (Ben)
sed -i "s/i915_gem_obj_set_color/i915_gem_obj_ggtt_set_color/" drivers/gpu/drm/i915/*.[ch]
sed -i "s/i915_gem_obj_bound/i915_gem_obj_ggtt_bound/" drivers/gpu/drm/i915/*.[ch]
sed -i "s/i915_gem_obj_size/i915_gem_obj_ggtt_size/" drivers/gpu/drm/i915/*.[ch]
sed -i "s/i915_gem_obj_offset/i915_gem_obj_ggtt_offset/" drivers/gpu/drm/i915/*.[ch]
(Daniel)
v3: Rebased on new reserve_node patch
Changed DRM_DEBUG_KMS to actually work (will need fixing later)
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
BSpec contains several scattered notes which state that the maximum
fence stride was increased to 256KB on IVB.
Testing on real hardware agrees.
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Our checks for an invalid fence stride forgot to guard against
zero stride on gen4+. Fix it.
v2: Avoid duplicated code (danvet)
Signed-off-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
The i915 driver uses sg lists for memory without backing 'struct page'
pages, similarly to other IO memory regions, setting only the DMA
address for these. It does this, so that it can program the HW MMU
tables in a uniform way both for sg lists with and without backing pages.
Without a valid page pointer we can't call nth_page to get the current
page in __sg_page_iter_next, so add a helper that relevant users can
call separately. Also add a helper to get the DMA address of the current
page (idea from Daniel).
Convert all places in i915, to use the new API.
Signed-off-by: Imre Deak <imre.deak@intel.com>
Reviewed-by: Damien Lespiau <damien.lespiau@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
So far the assumption was that each dma scatter list entry contains only
a single page. This might not hold in the future, when we'll introduce
compact scatter lists, so prepare for this everywhere in the i915 code
where we walk such a list.
We'll fix the place _creating_ these lists separately in the next patch
to help the reviewing/bisectability.
Reference: http://www.spinics.net/lists/dri-devel/msg33917.html
Signed-off-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
The purpose of the gtt structure is to help isolate our gtt specific
properties from the rest of the code (in doing so it help us finish the
isolation from the AGP connection).
The following members are pulled out (and renamed):
gtt_start
gtt_total
gtt_mappable_end
gtt_mappable
gtt_base_addr
gsm
The gtt structure will serve as a nice place to put gen specific gtt
routines in upcoming patches. As far as what else I feel belongs in this
structure: it is meant to encapsulate the GTT's physical properties.
This is why I've not added fields which track various drm_mm properties,
or things like gtt_mtrr (which is itself a pretty transient field).
Reviewed-by: Rodrigo Vivi <rodrigo.vivi@gmail.com>
[Ben modified commit messages]
Signed-off-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
The two functions are rather similar, so merge them.
Signed-off-by: Imre Deak <imre.deak@intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
On a machine with bit17 swizzling, we need to store the bit17 of the
physical page address in put-pages. This requires a memory allocation,
on average less than a page, which may be difficult to satisfy is the
request to put-pages is on behalf of the shrinker. We could allow that
allocation to pull from the reserved memory pools, but it seems much
safer to preallocate the array for tiled objects on affected machines.
v2: Export i915_gem_object_needs_bit17_swizzle() for reuse.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Daniel writes:
Bigger -fixes pile, mostly because I've included Ajax' DP dongle stuff,
as discussed on irc. Otherwise just small things:
- regression fix to finally make 6bpc auto-dither on dp work (Jani)
- reinstate an snb ctx w/a that accidentally got lost in a rework (Chris)
- fixup the DP train sequence, logic-goof-up uncovered by Coverty (Chris)
- fix set_caching locking (Ben)
- fix spurious segfault on con-current gtt mmap faulting (Dimitry and Mika)
- some pageflip correctness fixes (still hunting down some issues, but
these are the worst offenders of confused code that we've tracked down
thus far) from Chris and me
- fixup swizzling settings on vlv (Jesse)
- gt_mode w/a from Ben added, fixes snb gt1 rc6+hw ctx hangs.
* 'drm-intel-fixes' of git://people.freedesktop.org/~danvet/drm-intel:
drm/i915: Fix GT_MODE default value
drm/i915: don't frob the vblank ts in finish_page_flip
drm/i915: call drm_handle_vblank before finish_page_flip
drm/i915: print warning if vmi915_gem_fault error is not handled
drm/i915: EBUSY status handling added to i915_gem_fault().
drm/i915: Try harder to complete DP training pattern 1
drm/i915: set swizzling to none on VLV
drm/dp: Make sink count DP 1.2 aware
drm/dp: Document DP spec versions for various DPCD registers
drm/i915/dp: Be smarter about connection sense for branch devices
drm/i915/dp: Fetch downstream port info if needed during DPCD fetch
drm/dp: Update DPCD defines
drm: Export drm_probe_ddc()
drm/i915: Flush the pending flips on the CRTC before modification
drm/i915: Actually invalidate the TLB for the SandyBridge HW contexts w/a
drm/i915: Fix set_caching locking
drm/i915: use adjusted_mode instead of mode for checking the 6bpc force flag
Pull drm merge (part 1) from Dave Airlie:
"So first of all my tree and uapi stuff has a conflict mess, its my
fault as the nouveau stuff didn't hit -next as were trying to rebase
regressions out of it before we merged.
Highlights:
- SH mobile modesetting driver and associated helpers
- some DRM core documentation
- i915 modesetting rework, haswell hdmi, haswell and vlv fixes, write
combined pte writing, ilk rc6 support,
- nouveau: major driver rework into a hw core driver, makes features
like SLI a lot saner to implement,
- psb: add eDP/DP support for Cedarview
- radeon: 2 layer page tables, async VM pte updates, better PLL
selection for > 2 screens, better ACPI interactions
The rest is general grab bag of fixes.
So why part 1? well I have the exynos pull req which came in a bit
late but was waiting for me to do something they shouldn't have and it
looks fairly safe, and David Howells has some more header cleanups
he'd like me to pull, that seem like a good idea, but I'd like to get
this merge out of the way so -next dosen't get blocked."
Tons of conflicts mostly due to silly include line changes, but mostly
mindless. A few other small semantic conflicts too, noted from Dave's
pre-merged branch.
* 'drm-next' of git://people.freedesktop.org/~airlied/linux: (447 commits)
drm/nv98/crypt: fix fuc build with latest envyas
drm/nouveau/devinit: fixup various issues with subdev ctor/init ordering
drm/nv41/vm: fix and enable use of "real" pciegart
drm/nv44/vm: fix and enable use of "real" pciegart
drm/nv04/dmaobj: fixup vm target handling in preparation for nv4x pcie
drm/nouveau: store supported dma mask in vmmgr
drm/nvc0/ibus: initial implementation of subdev
drm/nouveau/therm: add support for fan-control modes
drm/nouveau/hwmon: rename pwm0* to pmw1* to follow hwmon's rules
drm/nouveau/therm: calculate the pwm divisor on nv50+
drm/nouveau/fan: rewrite the fan tachometer driver to get more precision, faster
drm/nouveau/therm: move thermal-related functions to the therm subdev
drm/nouveau/bios: parse the pwm divisor from the perf table
drm/nouveau/therm: use the EXTDEV table to detect i2c monitoring devices
drm/nouveau/therm: rework thermal table parsing
drm/nouveau/gpio: expose the PWM/TOGGLE parameter found in the gpio vbios table
drm/nouveau: fix pm initialization order
drm/nouveau/bios: check that fixed tvdac gpio data is valid before using it
drm/nouveau: log channel debug/error messages from client object rather than drm client
drm/nouveau: have drm debugging macros build on top of core macros
...
We don't have bit 6 swizzling on VLV, so this function is easy.
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Convert #include "..." to #include <path/...> in drivers/gpu/.
Signed-off-by: David Howells <dhowells@redhat.com>
Acked-by: Dave Airlie <airlied@redhat.com>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Acked-by: Dave Jones <davej@redhat.com>
Remove redundant DRM UAPI header #inclusions from drivers/gpu/.
Remove redundant #inclusions of core DRM UAPI headers (drm.h, drm_mode.h and
drm_sarea.h). They are now #included via drmP.h and drm_crtc.h via a preceding
patch.
Without this patch and the patch to make include the UAPI headers from the core
headers, after the UAPI split, the DRM C sources cannot find these UAPI headers
because the DRM code relies on specific -I flags to make #include "..." work
on headers in include/drm/ - but that does not work after the UAPI split without
adding more -I flags.
Signed-off-by: David Howells <dhowells@redhat.com>
Acked-by: Dave Airlie <airlied@redhat.com>
Acked-by: Arnd Bergmann <arnd@arndb.de>
Acked-by: Thomas Gleixner <tglx@linutronix.de>
Acked-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Acked-by: Dave Jones <davej@redhat.com>
Rather than have multiple data structures for describing our page layout
in conjunction with the array of pages, we can migrate all users over to
a scatterlist.
One major advantage, other than unifying the page tracking structures,
this offers is that we replace the vmalloc'ed array (which can be up to
a megabyte in size) with a chain of individual pages which helps reduce
memory pressure.
The disadvantage is that we then do not have a simple array to iterate,
or to access randomly. The common case for this is in the relocation
processing, which will typically fit within a single scatterlist page
and so be almost the same cost as the simple array. For iterating over
the array, the extra function call could be optimised away, but in
reality is an insignificant cost of either binding the pages, or
performing the pwrite/pread.
v2: Fix drm_clflush_sg() to not invoke wbinvd as well! And fix the
trivial compile error from rebasing.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
If we fail to unbind and so abort the change in tiling, we will have
removed the VMA for the object for no reason. The likelihood of unbind
failing is slim (other than ERESTARTSYS which will cause userspace to
try again), so the change is mostly for the principle.
Also improve the slightly stale comment.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel 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: Chris Wilson <chris@chris-wilson.co.uk>
[danvet: Add some bikeshed to the commit message about the stricter
use of fence_dirty.]
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
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>
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>
Use the helper function already employed by the pwrite/pread
functions.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Keith Packard <keithp@keithp.com>
Fixes tests/gem_tiled_pread on my snb. I know, mesa doesn't use this
on gen6+, but I also hate failing testcases.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Ben Widawsky <ben@bwidawsk.net>
Signed-off-by: Keith Packard <keithp@keithp.com>
Align unfenced buffers on older hardware to the power-of-two object
size. The docs suggest that it should be possible to align only to a
power-of-two tile height, but using the already computed fence size is
easier and always correct. We also have to make sure that we unbind
misaligned buffers upon tiling changes.
In order to prevent a repetition of this bug, we change the interface
to the alignment computation routines to force the caller to provide
the requested alignment and size of the GTT binding rather than assume
the current values on the object.
Reported-and-tested-by: Sitosfe Wheeler <sitsofe@yahoo.com>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=36326
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: stable@kernel.org
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Keith Packard <keithp@keithp.com>
Treat it like Ironlake and Sandy Bridge.
Signed-off-by: Jesse Barnes <jbarnes@virtuousgeek.org>
Reviewed-by: Keith Packard <keithp@keithp.com>
Signed-off-by: Keith Packard <keithp@keithp.com>
Early gen3 and gen2 chipset do not have the relaxed per-surface tiling
constraints of the later chipsets, so we need to check that the GTT
alignment is correct for the new tiling. If it is not, we need to
rebind.
Reported-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
This reverts commit c2e0eb1670.
As it turns out, userspace already depends upon being able to enable
tiling on existing bo which it promises to be large enough for its
purposes i.e. it will not access beyond the end of the last full-tile
row.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=35016
Reported-and-tested-by: Kamal Mostafa <kamal@canonical.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
It looks like gen2 has a peculiar interleaved 2-row inter-tile
layout. Probably inherited from i81x which had 2kb tiles (which
naturally fit an even-number-of-tile-rows scheme to fit onto 4kb
pages). There is no other mention of this in any docs (also not
in the Intel internal documention according to Chris Wilson).
Problem manifests itself in corruptions in the second half of the
last tile row (if the bo has an odd number of tiles). Which can
only happen with relaxed tiling (introduced in a00b10c360).
So reject set_tiling calls that don't satisfy this constrain to
prevent broken userspace from causing havoc. While at it, also
check the size for newer chipsets.
LKML: https://lkml.org/lkml/2011/2/19/5
Reported-by: Indan Zupancic <indan@nul.nu>
Tested-by: Indan Zupancic <indan@nul.nu>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Dave Airlie spotted that we had a potential bug should we ever rearrange
the drm_i915_gem_object so not the base drm_gem_object was not its first
member. He noticed that we often convert the return of
drm_gem_object_lookup() immediately into drm_i915_gem_object and then
check the result for nullity. This is only valid when the base object is
the first member and so the superobject has the same address. Play safe
instead and use the compiler to convert back to the original return
address for sanity testing.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Instead of reporting EIO upfront in the entrance of an ioctl that may or
may not attempt to use the GPU, defer the actual detection of an invalid
ioctl to when we issue a GPU instruction. This allows us to continue to
use bo in video memory (via pread/pwrite and mmap) after the GPU has hung.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
With this change, every batchbuffer can use all available fences (save
pinned and scanout, of course) without ever stalling the gpu!
In theory. Currently the actual pipelined update of the register is
disabled due to some stability issues. However, just the deferred update
is a significant win.
Based on a series of patches by Daniel Vetter.
The premise is that before every access to a buffer through the GTT we
have to declare whether we need a register or not. If the access is by
the GPU, a pipelined update to the register is made via the ringbuffer,
and we track the last seqno of the batches that access it. If by the
CPU we wait for the last GPU access and update the register (either
to clear or to set it for the current buffer).
One advantage of being able to pipeline changes is that we can defer the
actual updating of the fence register until we first need to access the
object through the GTT, i.e. we can eliminate the stall on set_tiling.
This is important as the userspace bo cache does not track the tiling
status of active buffers which generate frequent stalls on gen3 when
enabling tiling for an already bound buffer.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
As we may bind an object with the correct alignment, but with an invalid
size, it may pass the current checks on whether the object may be reused
with a fence.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
So long as we adhere to the fence registers rules for alignment and no
overlaps (including with unfenced accesses to linear memory) and account
for the tiled access in our size allocation, we do not have to allocate
the full fenced region for the object. This allows us to fight the bloat
tiling imposed on pre-i965 chipsets and frees up RAM for real use. [Inside
the GTT we still suffer the additional alignment constraints, so it doesn't
magic allow us to render larger scenes without stalls -- we need the
expanded GTT and fence pipelining to overcome those...]
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Owain Ainsworth reported an issue between the interaction of the
hangcheck and userspace immediately (and permanently) falling back to
s/w rasterisation. In order to break the mutex and begin resetting the
GPU, we must abort the current operation (usually within the wait) and
climb sufficiently far back up the call chain to drop the mutex. In his
implementation, Owain has a loop within the ioctl handler to detect the
hang and then sleep until the error handler has run. I've chosen to
return to userspace and report an EAGAIN which should trigger the
userspace ioctl handler to repeat the call (simply because it felt less
invasive...). Before hitting a wedged GPU, we then wait upon completion
of the error handler.
Reported-by: Owain G. Ainsworth <zerooa@googlemail.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
As we currently may need to acquire a fence register during a modeset,
we need to be able to do so in an uninterruptible manner. So expose that
parameter to the callers of the fence management code.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
This is consistent with trying to access a filename that not exist
within a directory which is a good analogy here. The main reason for the
change is that it is easy to confuse the error code of EBADF as an
performing an ioctl on an invalid file descriptor (rather than an
unknown object).
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Dave Airlie <airlied@redhat.com>
We generally issue an error message at the point of failure, and so this
warning with a fairly pointless stacktrace is superfluous and ugly.
Needless to say, the common trigger for this WARN happens to be EIO
where this is pure noise.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Eric Anholt <eric@anholt.net>
Makes no sense and complicates matters for pipelined tiling changes.
So don't allow it and return -EBUSY.
v2: Fix reference leak. Thanks to Owain Ainsworth for spotting this.
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Eric Anholt <eric@anholt.net>
Current code is definitely crap: Largest pitch allowed spills into
the TILING_Y bit of the fence registers ... :(
I've rewritten the limits check under the assumption that 3rd gen hw
has a 3d pitch limit of 8kb (like 2nd gen). This is supported by an
otherwise totally misleading XXX comment.
This bug mostly resulted in tiling-corrupted pixmaps because the kernel
allowed too wide buffers to be tiled. Bug brought to the light by the
xf86-video-intel 2.11 release because that unconditionally enabled
tiling for pixmaps, relying on the kernel to check things. Tiling for
the framebuffer was not affected because the ddx does some additional
checks there ensure the buffer is within hw-limits.
v2: Instead of computing the value that would be written into the
hw fence registers and then checking the limits simply check whether
the stride is above the 8kb limit. To better document the hw, add
some WARN_ONs in i915_write_fence_reg like I've done for the i830
case (using the right limits).
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=27449
Tested-by: Alexander Lam <lambchop468@gmail.com>
Cc: stable@kernel.org
Signed-off-by: Eric Anholt <eric@anholt.net>
Damien Lespiau