When userspace is passing around swapbuffers using DRI, we frequently
have to open and close the same object in the foreign address space.
This shows itself as the same object being rebound at roughly 30fps
(with a second object also being rebound at 30fps), which involves us
having to rewrite the page tables and maintain the drm_mm range manager
every time.
However, since the object still exists and it is only the local handle
that disappears, if we are lazy and do not unbind the VMA immediately
when the local user closes the object but defer it until the GPU is
idle, then we can reuse the same VMA binding. We still have to be
careful to mark the handle and lookup tables as closed to maintain the
uABI, just allowing the underlying VMA to be resurrected if the user is
able to access the same object from the same context again.
If the object itself is destroyed (neither userspace keeping a handle to
it), the VMA will be reaped immediately as usual.
In the future, this will be even more useful as instantiating a new VMA
for use on the GPU will become heavier. A nuisance indeed, so nip it in
the bud.
v2: s/__i915_vma_final_close/i915_vma_destroy/ etc.
v3: Leave a hint as to why we deferred the unbind on close.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180503195115.22309-1-chris@chris-wilson.co.uk
i915 is the only driver using those fields in the drm_gem_object
structure, so they only waste memory for all other drivers.
Move the fields into drm_i915_gem_object instead and patch the i915 code
with the following sed commands:
sed -i "s/obj->base.read_domains/obj->read_domains/g" drivers/gpu/drm/i915/*.c drivers/gpu/drm/i915/*/*.c
sed -i "s/obj->base.write_domain/obj->write_domain/g" drivers/gpu/drm/i915/*.c drivers/gpu/drm/i915/*/*.c
Change is only compile tested.
v2: move fields around as suggested by Chris.
Signed-off-by: Christian König <christian.koenig@amd.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20180216124338.9087-1-christian.koenig@amd.com
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
As the request will, in the following patch, implicitly invoke a
context-switch on construction, we should precede that with a GPU TLB
invalidation. Also, even before using GGTT, we always want to invalidate
the TLBs for any updates (as well as the ppgtt invalidates that are
unconditionally applied by execbuf). Since we almost always require the
TLB invalidate, do it unconditionally on request allocation and so we can
remove it from all other paths.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20171120102002.22254-1-chris@chris-wilson.co.uk
Reviewed-by: Mika Kuoppala <mika.kuoppala@linux.intel.com>
We check whether the multiplies will overflow prior to calling
kmalloc_array so that we can respond with -EINVAL for the invalid user
arguments rather than treating it as an -ENOMEM that would otherwise
occur. However, as Dan Carpenter pointed out, we did an addition on the
unsigned int prior to passing to kmalloc_array where it would be
promoted to size_t for the calculation, thereby allowing it to overflow
and underallocate.
v2: buffer_count is currently limited to INT_MAX because we treat it as
signaled variable for LUT_HANDLE in eb_lookup_vma
v3: Move common checks for eb1/eb2 into the same function
v4: Put the check back for nfence*sizeof(user_fence) overflow
v5: access_ok uses ULONG_MAX but kvmalloc_array uses SIZE_MAX
v6: size_t and unsigned long are not type-equivalent on 32b
Reported-by: Dan Carpenter <dan.carpenter@oracle.com>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20171116105059.25142-1-chris@chris-wilson.co.uk
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
In case the object has changed tiling between calls to execbuf, we need
to check if the existing offset inside the GTT matches the new tiling
constraint. We even need to do this for "unfenced" tiled objects, where
the 3D commands use an implied fence and so the object still needs to
match the physical fence restrictions on alignment (only required for
gen2 and early gen3).
In commit 2889caa923 ("drm/i915: Eliminate lots of iterations over
the execobjects array"), the idea was to remove the second guessing and
only set the NEEDS_MAP flag when required. However, the entire check
for an unusable offset for fencing was removed and not just the
secondary check. I.e.
/* avoid costly ping-pong once a batch bo ended up non-mappable */
if (entry->flags & __EXEC_OBJECT_NEEDS_MAP &&
!i915_vma_is_map_and_fenceable(vma))
return !only_mappable_for_reloc(entry->flags);
was entirely removed as the ping-pong between execbuf passes was fixed,
but its primary purpose in forcing unaligned unfenced access to be
rebound was forgotten.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=103502
Fixes: 2889caa923 ("drm/i915: Eliminate lots of iterations over the execobjects array")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20171031103607.17836-1-chris@chris-wilson.co.uk
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Getting started with v4.15 features:
- Cannonlake workarounds (Rodrigo, Oscar)
- Infoframe refactoring and fixes to enable infoframes for DP (Ville)
- VBT definition updates (Jani)
- Sparse warning fixes (Ville, Chris)
- Crtc state usage fixes and cleanups (Ville)
- DP vswing, pre-emph and buffer translation refactoring and fixes (Rodrigo)
- Prevent IPS from interfering with CRC capture (Ville, Marta)
- Enable Mesa to advertise ARB_timer_query (Nanley)
- Refactor GT number into intel_device_info (Lionel)
- Avoid eDP DP AUX CH timeouts harder (Manasi)
- CDCLK check improvements (Ville)
- Restore GPU clock boost on missed pageflip vblanks (Chris)
- Fence register reservation API for vGPU (Changbin)
- First batch of CCS fixes (Ville)
- Finally, numerous GEM fixes, cleanups and improvements (Chris)
* tag 'drm-intel-next-2017-09-07' of git://anongit.freedesktop.org/git/drm-intel: (100 commits)
drm/i915: Update DRIVER_DATE to 20170907
drm/i915/cnl: WaThrottleEUPerfToAvoidTDBackPressure:cnl(pre-prod)
drm/i915: Lift has-pinned-pages assert to caller of ____i915_gem_object_get_pages
drm/i915: Display WA #1133 WaFbcSkipSegments:cnl, glk
drm/i915/cnl: Allow the reg_read ioctl to read the RCS TIMESTAMP register
drm/i915: Move device_info.has_snoop into the static tables
drm/i915: Disable MI_STORE_DATA_IMM for i915g/i915gm
drm/i915: Re-enable GTT following a device reset
drm/i915/cnp: Wa 1181: Fix Backlight issue
drm/i915: Annotate user relocs with __user
drm/i915: Constify load detect mode
drm/i915/perf: Remove __user from u64 in drm_i915_perf_oa_config
drm/i915: Silence sparse by using gfp_t
drm/i915: io unmap functions want __iomem
drm/i915: Add __rcu to radix tree slot pointer
drm/i915: Wake up the device for the fbdev setup
drm/i915: Add interface to reserve fence registers for vGPU
drm/i915: Use correct path to trace include
drm/i915: Fix the missing PPAT cache attributes on CNL
drm/i915: Fix enum pipe vs. enum transcoder for the PCH transcoder
...
GFP_TEMPORARY was introduced by commit e12ba74d8f ("Group short-lived
and reclaimable kernel allocations") along with __GFP_RECLAIMABLE. It's
primary motivation was to allow users to tell that an allocation is
short lived and so the allocator can try to place such allocations close
together and prevent long term fragmentation. As much as this sounds
like a reasonable semantic it becomes much less clear when to use the
highlevel GFP_TEMPORARY allocation flag. How long is temporary? Can the
context holding that memory sleep? Can it take locks? It seems there is
no good answer for those questions.
The current implementation of GFP_TEMPORARY is basically GFP_KERNEL |
__GFP_RECLAIMABLE which in itself is tricky because basically none of
the existing caller provide a way to reclaim the allocated memory. So
this is rather misleading and hard to evaluate for any benefits.
I have checked some random users and none of them has added the flag
with a specific justification. I suspect most of them just copied from
other existing users and others just thought it might be a good idea to
use without any measuring. This suggests that GFP_TEMPORARY just
motivates for cargo cult usage without any reasoning.
I believe that our gfp flags are quite complex already and especially
those with highlevel semantic should be clearly defined to prevent from
confusion and abuse. Therefore I propose dropping GFP_TEMPORARY and
replace all existing users to simply use GFP_KERNEL. Please note that
SLAB users with shrinkers will still get __GFP_RECLAIMABLE heuristic and
so they will be placed properly for memory fragmentation prevention.
I can see reasons we might want some gfp flag to reflect shorterm
allocations but I propose starting from a clear semantic definition and
only then add users with proper justification.
This was been brought up before LSF this year by Matthew [1] and it
turned out that GFP_TEMPORARY really doesn't have a clear semantic. It
seems to be a heuristic without any measured advantage for most (if not
all) its current users. The follow up discussion has revealed that
opinions on what might be temporary allocation differ a lot between
developers. So rather than trying to tweak existing users into a
semantic which they haven't expected I propose to simply remove the flag
and start from scratch if we really need a semantic for short term
allocations.
[1] http://lkml.kernel.org/r/20170118054945.GD18349@bombadil.infradead.org
[akpm@linux-foundation.org: fix typo]
[akpm@linux-foundation.org: coding-style fixes]
[sfr@canb.auug.org.au: drm/i915: fix up]
Link: http://lkml.kernel.org/r/20170816144703.378d4f4d@canb.auug.org.au
Link: http://lkml.kernel.org/r/20170728091904.14627-1-mhocko@kernel.org
Signed-off-by: Michal Hocko <mhocko@suse.com>
Signed-off-by: Stephen Rothwell <sfr@canb.auug.org.au>
Acked-by: Mel Gorman <mgorman@suse.de>
Acked-by: Vlastimil Babka <vbabka@suse.cz>
Cc: Matthew Wilcox <willy@infradead.org>
Cc: Neil Brown <neilb@suse.de>
Cc: "Theodore Ts'o" <tytso@mit.edu>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
The early gen3 machines (i915g/Grantsdale and i915gm/Alviso) share a lot
of characteristics in their MI/GTT blocks with gen2, and in particular
can only use physical addresses in MI_STORE_DATA_IMM. This makes it
incompatible with our usage, so include those two machines in the
blacklist to prevent usage.
v2: Make it easy for gcc and rewrite it as a switch to save some space.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Ville Syrjälä <ville.syrjala@linux.intel.com> #v1
Link: https://patchwork.freedesktop.org/patch/msgid/20170906152859.5304-1-chris@chris-wilson.co.uk
Sometimes we know we are the only user of the bo, but since we take a
protective pin_pages early on, an attempt to change the vmap on the
object is denied because it is busy. i915_gem_object_pin_map() cannot
tell from our single pin_count if the operation is safe. Instead we must
pass that information down from the caller in the manner of
I915_MAP_OVERRIDE.
This issue has existed from the introduction of the mapping, but was
never noticed as the only place where this conflict might happen is for
cached kernel buffers (such as allocated by i915_gem_batch_pool_get()).
Until recently there was only a single user (the cmdparser) so no
conflicts ever occurred. However, we now use it to allocate batches for
different operations (using MAP_WC on !llc for writes) in addition to the
existing shadow batch (using MAP_WB for reads).
We could either keep both mappings cached, or use a different write
mechanism if we detect a MAP_WB already exists (i.e. clflush
afterwards), but as we haven't seen this issue in the wild (it requires
hitting the GPU reloc path in addition to the cmdparser) for simplicity
just allow the mappings to be recreated.
v2: Include the i915_MAP_OVERRIDE bit in the enum so the compiler knows
about all the valid values.
Fixes: 7dd4f6729f ("drm/i915: Async GPU relocation processing")
Testcase: igt/gem_lut_handle # byt, completely by accident
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20170828104631.8606-1-chris@chris-wilson.co.uk
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
(cherry picked from commit a575c67617)
Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
Sometimes we know we are the only user of the bo, but since we take a
protective pin_pages early on, an attempt to change the vmap on the
object is denied because it is busy. i915_gem_object_pin_map() cannot
tell from our single pin_count if the operation is safe. Instead we must
pass that information down from the caller in the manner of
I915_MAP_OVERRIDE.
This issue has existed from the introduction of the mapping, but was
never noticed as the only place where this conflict might happen is for
cached kernel buffers (such as allocated by i915_gem_batch_pool_get()).
Until recently there was only a single user (the cmdparser) so no
conflicts ever occurred. However, we now use it to allocate batches for
different operations (using MAP_WC on !llc for writes) in addition to the
existing shadow batch (using MAP_WB for reads).
We could either keep both mappings cached, or use a different write
mechanism if we detect a MAP_WB already exists (i.e. clflush
afterwards), but as we haven't seen this issue in the wild (it requires
hitting the GPU reloc path in addition to the cmdparser) for simplicity
just allow the mappings to be recreated.
v2: Include the i915_MAP_OVERRIDE bit in the enum so the compiler knows
about all the valid values.
Fixes: 7dd4f6729f ("drm/i915: Async GPU relocation processing")
Testcase: igt/gem_lut_handle # byt, completely by accident
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20170828104631.8606-1-chris@chris-wilson.co.uk
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
This was the competing idea long ago, but it was only with the rewrite
of the idr as an radixtree and using the radixtree directly ourselves,
along with the realisation that we can store the vma directly in the
radixtree and only need a list for the reverse mapping, that made the
patch performant enough to displace using a hashtable. Though the vma ht
is fast and doesn't require any extra allocation (as we can embed the node
inside the vma), it does require a thread for resizing and serialization
and will have the occasional slow lookup. That is hairy enough to
investigate alternatives and favour them if equivalent in peak performance.
One advantage of allocating an indirection entry is that we can support a
single shared bo between many clients, something that was done on a
first-come first-serve basis for shared GGTT vma previously. To offset
the extra allocations, we create yet another kmem_cache for them.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20170816085210.4199-5-chris@chris-wilson.co.uk
When userspace is doing most of the work, avoiding relocs (using
NO_RELOC) and opting out of implicit synchronisation (using ASYNC), we
still spend a lot of time processing the arrays in execbuf, even though
we now should have nothing to do most of the time. One issue that
becomes readily apparent in profiling anv is that iterating over the
large execobj[] is unfriendly to the loop prefetchers of the CPU and it
much prefers iterating over a pair of arrays rather than one big array.
v2: Clear vma[] on construction to handle errors during vma lookup
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20170816085210.4199-3-chris@chris-wilson.co.uk
This commit adds support for waiting on or signaling DRM syncobjs as
part of execbuf. It does so by hijacking the currently unused cliprects
pointer to instead point to an array of i915_gem_exec_fence structs
which containe a DRM syncobj and a flags parameter which specifies
whether to wait on it or to signal it. This implementation
theoretically allows for both flags to be set in which case it waits on
the dma_fence that was in the syncobj and then immediately replaces it
with the dma_fence from the current execbuf.
v2:
- Rebase on new syncobj API
v3:
- Pull everything out into helpers
- Do all allocation in gem_execbuffer2
- Pack the flags in the bottom 2 bits of the drm_syncobj*
v4:
- Prevent a potential race on syncobj->fence
Testcase: igt/gem_exec_fence/syncobj*
Signed-off-by: Jason Ekstrand <jason@jlekstrand.net>
Link: https://patchwork.freedesktop.org/patch/msgid/1499289202-25441-1-git-send-email-jason.ekstrand@intel.com
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Link: https://patchwork.freedesktop.org/patch/msgid/20170815145733.4562-1-chris@chris-wilson.co.uk
The goal here was to minimise doing any thing or any check inside the
kernel that was not strictly required. For a userspace that assumes
complete control over the cache domains, the kernel is usually using
outdated information and may trigger clflushes where none were
required.
However, swapping is a situation where userspace has no knowledge of the
domain transfer, and will leave the object in the CPU cache. The kernel
must flush this out to the backing storage prior to use with the GPU. As
we use an asynchronous task tracked by an implicit fence for this, we
also need to cancel the ASYNC flag on the object so that the object will
wait for the clflush to complete before being executed. This also absolves
userspace of the responsibility imposed by commit 77ae995789 ("drm/i915:
Enable userspace to opt-out of implicit fencing") that its needed to ensure
that the object was out of the CPU cache prior to use on the GPU.
Fixes: 77ae995789 ("drm/i915: Enable userspace to opt-out of implicit fencing")
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=101571
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Jason Ekstrand <jason@jlekstrand.net>
Reviewed-by: Jason Ekstrand <jason@jlekstrand.net>
Link: https://patchwork.freedesktop.org/patch/msgid/20170721145037.25105-5-chris@chris-wilson.co.uk
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
I was being overly paranoid in not updating the execobject.offset after
performing the fallback copy where we set reloc.presumed_offset to -1.
The thinking was to ensure that a subsequent NORELOC execbuf would be
forced to process the invalid relocations. However this is overkill so
long as we *only* update the execobject.offset following a successful
update of the relocation value witin the batch. If we have to repeat the
execbuf due to a later interruption, then we may skip the relocations on
the second pass (honouring NORELOC) since the execobject.offset match
the actual offsets (even though reloc.presumed_offset is garbage).
Subsequent calls to execbuf with NORELOC should themselves ensure that
the reloc.presumed_offset have been corrected in case of future
migration.
Reporting back the actual execobject.offset, even when
reloc.presumed_offset is garbage, ensures that reuse of those objects
use the latest information to avoid relocations.
Fixes: 2889caa923 ("drm/i915: Eliminate lots of iterations over the execobjects array")
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=101635
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20170721145037.25105-4-chris@chris-wilson.co.uk
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Signed-off-by: Daniel Vetter <daniel.vetter@ffwll.ch>
When computing a hash for looking up relocation target handles in an
execbuf, we start with a large size for the hashtable and proceed to
halve it until the allocation succeeds. The final attempt is with an
order of 0 (i.e. a single element). This means that we then pass bits=0
to hash_32() which then computes "hash >> (32 - 0)" to lookup the single
element. Right shifting a value by the width of the operand is
undefined, so limit the smallest hash table we use to order 1.
v2: Keep the retry allocation flag for the final pass
Fixes: 4ff4b44cbb ("drm/i915: Store a direct lookup from object handle to vma")
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/20170629150425.27508-1-chris@chris-wilson.co.uk
