When transitioning to the GTT or CPU domain we wait on all rendering
from i915 to complete (with the optimisation of allowing concurrent read
access by both the GPU and client). We don't yet ensure all rendering
from third parties (tracked by implicit fences on the dma-buf) is
complete. Since implicitly tracked rendering by third parties will
ignore our cache-domain tracking, we have to always wait upon rendering
from third-parties when transitioning to direct access to the backing
store. We still rely on clients notifying us of cache domain changes
(i.e. they need to move to the GTT read or write domain after doing a CPU
access before letting the third party render again).
v2:
This introduces a potential WARN_ON into i915_gem_object_free() as the
current i915_vma_unbind() calls i915_gem_object_wait_rendering(). To
hit this path we first need to render with the GPU, have a dma-buf
attached with an unsignaled fence and then interrupt the wait. It does
get fixed later in the series (when i915_vma_unbind() only waits on the
active VMA and not all, including third-party, rendering.
To offset that risk, use the __i915_vma_unbind_no_wait hack.
Testcase: igt/prime_vgem/basic-fence-read
Testcase: igt/prime_vgem/basic-fence-mmap
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1469002875-2335-8-git-send-email-chris@chris-wilson.co.uk
Since commit a6f766f397 ("drm/i915: Limit ring synchronisation (sw
sempahores) RPS boosts") and commit bcafc4e38b ("drm/i915: Limit mmio
flip RPS boosts") we have limited the waitboosting for semaphores and
flips. Ideally we do not want to boost in either of these instances as no
userspace consumer is waiting upon the results (though a userspace producer
may be stalled trying to submit an execbuf - but in this case the
producer is being throttled due to the engine being saturated with
work). With the introduction of NO_WAITBOOST in the previous patch, we
can finally disable these needless boosts.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1469002875-2335-6-git-send-email-chris@chris-wilson.co.uk
Setting a write-back cache policy in the MOCS entry definition also
implies snooping, which has a considerable overhead. This is
unexpected for a few reasons:
- From user-space's point of view since it didn't want a coherent
surface (it didn't set the buffer as such via the set caching IOCTL).
- There is a separate MOCS entry field for snooping (which we never
set).
- This MOCS table is about caching in (e)LLC and there is no (e)LLC on
BXT. There is a separate table for L3 cache control.
Considering the above the current behavior of snooping looks like an
unintentional side-effect of the WB setting. Changing it to be LLC-UC
gets rid of the snooping without any ill-effects. For a coherent
surface the application would use a separate MOCS entry at index 1 and
call the set caching IOCTL to setup the PTE entries for the
corresponding buffer to be snooped. In the future we could also add a
new MOCS entry for coherent surfaces.
This resulted in 70% improvement in synthetic texturing benchmarks.
Kudos to Valtteri Rantala, Eero Tamminen and Michael T Frederick and
Ville who helped to narrow the source of problem to the kernel and to
the snooping behaviour in particular.
With a follow-up change to adjust the 3rd entry value
igt/gem_mocs_settings is passing after this change.
v2:
- Rebase on v2 of patch 1/2.
v3:
- Set the entry as LLC uncached instead of PTE-passthrough. This way
we also keep snooping disabled, but we also make the cacheability/
coherency setting indepent of the PTE which is managed by the
kernel. (Chris)
CC: Rong R Yang <rong.r.yang@intel.com>
CC: Yakui Zhao <yakui.zhao@intel.com>
CC: Valtteri Rantala <valtteri.rantala@intel.com>
CC: Eero Tamminen <eero.t.tamminen@intel.com>
CC: Michael T Frederick <michael.t.frederick@intel.com>
CC: Ville Syrjälä <ville.syrjala@linux.intel.com>
CC: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Imre Deak <imre.deak@intel.com>
Acked-by: Zhao Yakui <yakui.zhao@intel.com>
Tested-by: Rong R Yang <rong.r.yang@intel.com>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1467380406-11954-3-git-send-email-imre.deak@intel.com
Fairly minimal, there's still lots of functions without any docs, and
which aren't static. But probably we want to first clean this up some more.
- Drop the bogus const. Marking argument pointers themselves (instead of
what they point at) as const provides roughly 0 value. And it's confusing,
since the data the pointer points at _is_ being changed.
- Remove kerneldoc for static functions. Keep comments where they seem valuable.
- Indent and whitespace fixes.
- Blockquote the bit field definitions of the descriptor for correct layouting.
Acked-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/1468612088-9721-9-git-send-email-daniel.vetter@ffwll.ch
Precursor for fix to secure batch execution. We will need to be able to
retrieve the batch VMA (as well as the batch itself) from the eb list,
so this patch extracts that part of eb_get_batch() into a separate
function, and moves both parts to a more logical place in the file, near
where the eb list is created.
Also, it may not be obvious, but the current execbuffer2 ioctl interface
requires that the buffer object containing the batch-to-be-executed be
the LAST entry in the exec2_list[] array (I expected it to be the
first!).
To clarify this, we can replace the rather obscure construct
"list_entry(eb->vmas.prev, ...)"
in the old version of eb_get_batch() with the equivalent but more
explicit
"list_last_entry(&eb->vmas,...)"
in the new eb_get_batch_vma() and of course add an explanatory comment.
Signed-off-by: Dave Gordon <david.s.gordon@intel.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/1468504324-12690-2-git-send-email-david.s.gordon@intel.com
Two different sets of flag bits are stored in the 'flags' member of a
'struct drm_i915_gem_exec_object2', and they're defined in two different
source files, increasing the risk of an accidental clash.
Some flags in this field are supplied by the user; these are defined in
i915_drm.h, and they start from the LSB and work up.
Other flags are defined in i915_gem_execbuffer, for internal use within
that file only; they start from the MSB and work down.
So here we add a compile-time check that the two sets of flags do not
overlap, which would cause all sorts of confusion.
Signed-off-by: Dave Gordon <david.s.gordon@intel.com>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/1468504324-12690-1-git-send-email-david.s.gordon@intel.com
Currently execlists is exempt from emitting a request to switch each
ring away from the current context over to the dev_priv->kernel_context
(for whatever reason, just under execlists the GGTT is unlikely to be as
fragmented, however the switch may help in some extreme cases). Extract
the switcher and enable it for execlsts as well, as we need to do so in
a later patch to force the context switch before suspend. (And since for
that switch we explicitly require the disposable kernel context, rename
the extracted function.)
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1468590980-6186-1-git-send-email-chris@chris-wilson.co.uk
Chris Wilson
Rodrigo Vivi
Imre Deak
Daniel Vetter
Dave Gordon
Bob Paauwe
Ville Syrjälä