There is an improbable, but not impossible, case that if we leave the
pages unpin as we operate on the object, then somebody via the shrinker
may steal the lock (which lock? right now, it is struct_mutex, THE lock)
and change the cache domains after we have already inspected them.
(Whilst here, avail ourselves of the opportunity to take a couple of
steps to make the two functions look more similar.)
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/20160818161718.27187-11-chris@chris-wilson.co.uk
If we quickly switch from writing through the GTT to a read of the
physical page directly with the CPU (e.g. performing relocations through
the GTT and then running the command parser), we can observe that the
writes are not visible to the CPU. It is not a coherency problem, as
extensive investigations with clflush have demonstrated, but a mere
timing issue - we have to wait for the GTT to complete it's write before
we start our read from the CPU.
The issue can be illustrated in userspace with:
gtt = gem_mmap__gtt(fd, handle, 0, OBJECT_SIZE, PROT_READ | PROT_WRITE);
cpu = gem_mmap__cpu(fd, handle, 0, OBJECT_SIZE, PROT_READ | PROT_WRITE);
gem_set_domain(fd, handle, I915_GEM_DOMAIN_GTT, I915_GEM_DOMAIN_GTT);
for (i = 0; i < OBJECT_SIZE / 64; i++) {
int x = 16*i + (i%16);
gtt[x] = i;
clflush(&cpu[x], sizeof(cpu[x]));
assert(cpu[x] == i);
}
Experimenting with that shows that this behaviour is indeed limited to
recent Atom-class hardware.
Testcase: igt/gem_exec_flush/basic-batch-default-cmd #byt
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/20160818161718.27187-10-chris@chris-wilson.co.uk
This is a companion to i915_gem_obj_prepare_shmem_read() that prepares
the backing storage for direct writes. It first serialises with the GPU,
pins the backing storage and then indicates what clfushes are required in
order for the writes to be coherent.
Whilst here, fix support for ancient CPUs without clflush for which we
cannot do the GTT+clflush tricks.
v2: Add i915_gem_obj_finish_shmem_access() for symmetry
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/20160818161718.27187-8-chris@chris-wilson.co.uk
If we cannot release the fence (for example if someone is inexplicably
trying to write into a tiled framebuffer that is currently pinned to the
display! *cough* kms_frontbuffer_tracking *cough*) fallback to using the
page-by-page pwrite/pread interface, rather than fail the syscall
entirely.
Since this is triggerable by the user (along pwrite) we have to remove
the WARN_ON(fence->pin_count).
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/20160818161718.27187-6-chris@chris-wilson.co.uk
Daniel Vetter proposed a new challenge to the serialisation inside the
busy-ioctl that exposed a flaw that could result in us reporting the
wrong engine as being busy. If the request is reallocated as we test
its busyness and then reassigned to this object by another thread, we
would not notice that the test itself was incorrect.
We are faced with a choice of using __i915_gem_active_get_request_rcu()
to first acquire a reference to the request preventing the race, or to
acknowledge the race and accept the limitations upon the accuracy of the
busy flags. Note that we guarantee that we never falsely report the
object as idle (providing userspace itself doesn't race), and so the
most important use of the busy-ioctl and its guarantees are fulfilled.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Daniel Vetter <daniel.vetter@intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Daniel Vetter <daniel.vetter@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1471337440-16777-1-git-send-email-chris@chris-wilson.co.uk
Backmerge because too many conflicts, and also we need to get at the
latest struct fence patches from Gustavo. Requested by Chris Wilson.
Signed-off-by: Daniel Vetter <daniel.vetter@intel.com>
- refactor ddi buffer programming a bit (Ville)
- large-scale renaming to untangle naming in the gem code (Chris)
- rework vma/active tracking for accurately reaping idle mappings of shared
objects (Chris)
- misc dp sst/mst probing corner case fixes (Ville)
- tons of cleanup&tunings all around in gem
- lockless (rcu-protected) request lookup, plus use it everywhere for
non(b)locking waits (Chris)
- pipe crc debugfs fixes (Rodrigo)
- random fixes all over
* tag 'drm-intel-next-2016-08-08' of git://anongit.freedesktop.org/drm-intel: (222 commits)
drm/i915: Update DRIVER_DATE to 20160808
drm/i915: fix aliasing_ppgtt leak
drm/i915: Update comment before i915_spin_request
drm/i915: Use drm official vblank_no_hw_counter callback.
drm/i915: Fix copy_to_user usage for pipe_crc
Revert "drm/i915: Track active streams also for DP SST"
drm/i915: fix WaInsertDummyPushConstPs
drm/i915: Assert that the request hasn't been retired
drm/i915: Repack fence tiling mode and stride into a single integer
drm/i915: Document and reject invalid tiling modes
drm/i915: Remove locking for get_tiling
drm/i915: Remove pinned check from madvise ioctl
drm/i915: Reduce locking inside swfinish ioctl
drm/i915: Remove (struct_mutex) locking for busy-ioctl
drm/i915: Remove (struct_mutex) locking for wait-ioctl
drm/i915: Do a nonblocking wait first in pread/pwrite
drm/i915: Remove unused no-shrinker-steal
drm/i915: Tidy generation of the GTT mmap offset
drm/i915/shrinker: Wait before acquiring struct_mutex under oom
drm/i915: Simplify do_idling() (Ironlake vt-d w/a)
...
vmaps has a provision for controlling the page protection bits, with which
we can use to control the mapping type, e.g. WB, WC, UC or even WT.
To allow the caller to choose their mapping type, we add a parameter to
i915_gem_object_pin_map - but we still only allow one vmap to be cached
per object. If the object is currently not pinned, then we recreate the
previous vmap with the new access type, but if it was pinned we report an
error. This effectively limits the access via i915_gem_object_pin_map to a
single mapping type for the lifetime of the object. Not usually a problem,
but something to be aware of when setting up the object's vmap.
We will want to vary the access type to enable WC mappings of ringbuffer
and context objects on !llc platforms, as well as other objects where we
need coherent access to the GPU's pages without going through the GTT
v2: Remove the redundant braces around pin count check and fix the marker
in documentation (Chris)
v3:
- Add a new enum for the vmalloc mapping type & pass that as an argument to
i915_object_pin_map. (Tvrtko)
- Use PAGE_MASK to extract or filter the mapping type info and remove a
superfluous BUG_ON.(Tvrtko)
v4:
- Rename the enums and clean up the pin_map function. (Chris)
v5: Drop the VM_NO_GUARD, minor cosmetics.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Signed-off-by: Akash Goel <akash.goel@intel.com>
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1471001999-17787-1-git-send-email-chris@chris-wilson.co.uk
In commit 2529d57050 ("drm/i915: Drop racy markup of missed-irqs from
idle-worker") the racy detection of missed interrupts was removed when
we went idle. This however opened up the issue that the stuck waiters
were not being reported, causing a test case failure. If we move the
stuck waiter detection out of hangcheck and into the breadcrumb
mechanims (i.e. the waiter) itself, we can avoid this issue entirely.
This leaves hangcheck looking for a stuck GPU (inspecting for request
advancement and HEAD motion), and breadcrumbs looking for a stuck
waiter - hopefully make both easier to understand by their segregation.
v2: Reduce the error message as we now run independently of hangcheck,
and the hanging batch used by igt also counts as a stuck waiter causing
extra warnings in dmesg.
v3: Move the breadcrumb's hangcheck kickstart to the first missed wait.
Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=97104
Fixes: 2529d57050 (waiter"drm/i915: Drop racy markup of missed-irqs...")
Testcase: igt/drv_missed_irq
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@linux.intel.com>
Cc: Mika Kuoppala <mika.kuoppala@intel.com>
Reviewed-by: Mika Kuoppala <mika.kuoppala@intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1470761272-1245-2-git-send-email-chris@chris-wilson.co.uk
We don't need to incur the overhead of checking whether the object is
pinned prior to changing its madvise. If the object is pinned, the
madvise will not take effect until it is unpinned and so we cannot free
the pages being pointed at by hardware. Marking a pinned object with
allocated pages as DONTNEED will not trigger any undue warnings. The check
is therefore superfluous, and by removing it we can remove a linear walk
over all the vma the object has.
Still despite it being an overzealous check, that error code is part of
the current ABI and so we must proceed with caution.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: http://patchwork.freedesktop.org/patch/msgid/1470388464-28458-15-git-send-email-chris@chris-wilson.co.uk
With a bit of care (and leniency) we can iterate over the object and
wait for previous rendering to complete with judicial use of atomic
reference counting. The ABI requires us to ensure that an active object
is eventually flushed (like the busy-ioctl) which is guaranteed by our
management of requests (i.e. everything that is submitted to hardware is
flushed in the same request). All we have to do is ensure that we can
detect when the requests are complete for reporting when the object is
idle (without triggering ETIME), locklessly - this is handled by
i915_gem_active_wait_unlocked().
The impact of this is actually quite small - the return to userspace
following the wait was already lockless and so we don't see much gain in
latency improvement upon completing the wait. What we do achieve here is
completing an already finished wait without hitting the struct_mutex,
our hold is quite short and so we are typically just a victim of
contention rather than a cause - but it is still one less contention
point!
v2: Break up a long line.
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/1470388464-28458-12-git-send-email-chris@chris-wilson.co.uk
If we try and read or write to an active request, we first must wait
upon the GPU completing that request. Let's do that without holding the
mutex (and so allow someone else to access the GPU whilst we wait). Upon
completion, we will acquire the mutex and only then start the operation
(i.e. we do not rely on state from before the initial wait).
v2: Repaint the goto labels
v3: Move the tracepoints back to the start of the ioctls
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/1470388464-28458-11-git-send-email-chris@chris-wilson.co.uk
If we make the observation that mmap-offsets are only released when we
free an object, we can then deduce that the shrinker only creates free
space in the mmap arena indirectly by flushing the request list and
freeing expired objects. If we combine this with the lockless
vma-manager and lockless idling, we can avoid taking our big struct_mutex
until we need to actually free the requests.
One side-effect is that we defer the madvise checking until we need the
pages (i.e. the fault handler). This brings us into line with the other
delayed checks (and madvise in general).
v2: s/ret/err/ and use if (!err) rather than if (ret == 0)
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/1470388464-28458-9-git-send-email-chris@chris-wilson.co.uk
The principal motivation for this was to try and eliminate the
struct_mutex from i915_gem_suspend - but we still need to hold the mutex
current for the i915_gem_context_lost(). (The issue there is that there
may be an indirect lockdep cycle between cpu_hotplug (i.e. suspend) and
struct_mutex via the stop_machine().) For the moment, enabling last
request tracking for the engine, allows us to do busyness checking and
waiting without requiring the struct_mutex - which is useful in its own
right.
As a side-effect of having a robust means for tracking engine busyness,
we can replace our other busyness heuristic, that of comparing against
the last submitted seqno. For paranoid reasons, we have a semi-ordered
check of that seqno inside the hangchecker, which we can now improve to
an ordered check of the engine's busyness (removing a locked xchg in the
process).
v2: Pass along "bool interruptible" as being unlocked we cannot rely on
i915->mm.interruptible being stable or even under our control.
v3: Replace check Ironlake i915_gpu_busy() with the common precalculated value
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/1470388464-28458-6-git-send-email-chris@chris-wilson.co.uk
Before suspending (or unloading), we would first wait upon all rendering
to be completed and then disable the rings. This later step is a remanent
from DRI1 days when we did not use request tracking for all operations
upon the ring. Now that we are sure we are waiting upon the very last
operation by the engine, we can forgo clobbering the ring registers,
though we do keep the assert that the engine is indeed idle before
sleeping.
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/1470388464-28458-5-git-send-email-chris@chris-wilson.co.uk
We can completely avoid taking the struct_mutex around the non-blocking
waits by switching over to the RCU request management (trading the mutex
for a RCU read lock and some complex atomic operations). The improvement
is that we gain further contention reduction, and overall the code
become simpler due to the reduced mutex dancing.
v2: Move i915_gem_fault tracepoint back to the start of the function,
before the unlocked wait.
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/1470388464-28458-2-git-send-email-chris@chris-wilson.co.uk
If we enable RCU for the requests (providing a grace period where we can
inspect a "dead" request before it is freed), we can allow callers to
carefully perform lockless lookup of an active request.
However, by enabling deferred freeing of requests, we can potentially
hog a lot of memory when dealing with tens of thousands of requests per
second - with a quick insertion of a synchronize_rcu() inside our
shrinker callback, that issue disappears.
v2: Currently, it is our responsibility to handle reclaim i.e. to avoid
hogging memory with the delayed slab frees. At the moment, we wait for a
grace period in the shrinker, and block for all RCU callbacks on oom.
Suggested alternatives focus on flushing our RCU callback when we have a
certain number of outstanding request frees, and blocking on that flush
after a second high watermark. (So rather than wait for the system to
run out of memory, we stop issuing requests - both are nondeterministic.)
Paul E. McKenney wrote:
Another approach is synchronize_rcu() after some largish number of
requests. The advantage of this approach is that it throttles the
production of callbacks at the source. The corresponding disadvantage
is that it slows things up.
Another approach is to use call_rcu(), but if the previous call_rcu()
is still in flight, block waiting for it. Yet another approach is
the get_state_synchronize_rcu() / cond_synchronize_rcu() pair. The
idea is to do something like this:
cond_synchronize_rcu(cookie);
cookie = get_state_synchronize_rcu();
You would of course do an initial get_state_synchronize_rcu() to
get things going. This would not block unless there was less than
one grace period's worth of time between invocations. But this
assumes a busy system, where there is almost always a grace period
in flight. But you can make that happen as follows:
cond_synchronize_rcu(cookie);
cookie = get_state_synchronize_rcu();
call_rcu(&my_rcu_head, noop_function);
Note that you need additional code to make sure that the old callback
has completed before doing a new one. Setting and clearing a flag
with appropriate memory ordering control suffices (e.g,. smp_load_acquire()
and smp_store_release()).
v3: More comments on compiler and processor order of operations within
the RCU lookup and discover we can use rcu_access_pointer() here instead.
v4: Wrap i915_gem_active_get_rcu() to take the rcu_read_lock itself.
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Maarten Lankhorst <maarten.lankhorst@linux.intel.com>
Cc: "Goel, Akash" <akash.goel@intel.com>
Cc: Josh Triplett <josh@joshtriplett.org>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Reviewed-by: Daniel Vetter <daniel.vetter@ffwll.ch>
Link: http://patchwork.freedesktop.org/patch/msgid/1470324762-2545-25-git-send-email-chris@chris-wilson.co.uk
We are motivated to avoid using a bitfield for obj->active for a couple
of reasons. Firstly, we wish to document our lockless read of obj->active
using READ_ONCE inside i915_gem_busy_ioctl() and that requires an
integral type (i.e. not a bitfield). Secondly, gcc produces abysmal code
when presented with a bitfield and that shows up high on the profiles of
request tracking (mainly due to excess memory traffic as it converts
the bitfield to a register and back and generates frequent AGI in the
process).
v2: BIT, break up a long line in compute the other engines, new paint
for i915_gem_object_is_active (now i915_gem_object_get_active).
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/1470324762-2545-23-git-send-email-chris@chris-wilson.co.uk
Not only is i915_vma_pin() called for every single object on every single
execbuf, it is usually a simple increment as the VMA is already bound for
execution by the GPU. Rearrange the tests for unbound and pin_count
overflow so that we can do the increment and test very cheaply and
compact enough to inline the operation into execbuf. The trick used is
to note that we can check for an overflow bit (keeping space available
for it inside the flags) at the same time as checking the binding bits.
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/1470324762-2545-17-git-send-email-chris@chris-wilson.co.uk
