Skip to content

K
Kernel

openeuler / Kernel
大约 1 年 前同步成功

通知 6
Star 0
Fork 0
K
Kernel

体验新版 GitCode，发现更多精彩内容 >>

切换分支/标签
  1. 04 10月, 2019 1 次提交
    • C
      drm/i915: Pull i915_vma_pin under the vm->mutex · 2850748e
      由 Chris Wilson 提交于 
      Replace the struct_mutex requirement for pinning the i915_vma with the
local vm->mutex instead. Note that the vm->mutex is tainted by the
shrinker (we require unbinding from inside fs-reclaim) and so we cannot
allocate while holding that mutex. Instead we have to preallocate
workers to do allocate and apply the PTE updates after we have we
reserved their slot in the drm_mm (using fences to order the PTE writes
with the GPU work and with later unbind).

In adding the asynchronous vma binding, one subtle requirement is to
avoid coupling the binding fence into the backing object->resv. That is
the asynchronous binding only applies to the vma timeline itself and not
to the pages as that is a more global timeline (the binding of one vma
does not need to be ordered with another vma, nor does the implicit GEM
fencing depend on a vma, only on writes to the backing store). Keeping
the vma binding distinct from the backing store timelines is verified by
a number of async gem_exec_fence and gem_exec_schedule tests. The way we
do this is quite simple, we keep the fence for the vma binding separate
and only wait on it as required, and never add it to the obj->resv
itself.

Another consequence in reducing the locking around the vma is the
destruction of the vma is no longer globally serialised by struct_mutex.
A natural solution would be to add a kref to i915_vma, but that requires
decoupling the reference cycles, possibly by introducing a new
i915_mm_pages object that is own by both obj->mm and vma->pages.
However, we have not taken that route due to the overshadowing lmem/ttm
discussions, and instead play a series of complicated games with
trylocks to (hopefully) ensure that only one destruction path is called!

v2: Add some commentary, and some helpers to reduce patch churn.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: NTvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20191004134015.13204-4-chris@chris-wilson.co.uk
      2850748e
  3. 11 9月, 2019 1 次提交
  5. 03 9月, 2019 1 次提交
    • C
      drm/i915: Replace obj->pin_global with obj->frontbuffer · 5a90606d
      由 Chris Wilson 提交于 
      obj->pin_global was originally used as a means to keep the shrinker off
the active scanout, but we use the vma->pin_count itself for that and
the obj->frontbuffer to delay shrinking active framebuffers. The other
role that obj->pin_global gained was for spotting display objects inside
GEM and working harder to keep those coherent; for which we can again
simply inspect obj->frontbuffer directly.

Coming up next, we will want to manipulate the pin_global counter
outside of the principle locks, so would need to make pin_global atomic.
However, since obj->frontbuffer is already managed atomically, it makes
sense to use that the primary key for display objects instead of having
pin_global.

Ville pointed out the principle difference is that obj->frontbuffer is
set for as long as an intel_framebuffer is attached to an object, but
obj->pin_global was only raised for as long as the object was active. In
practice, this means that we consider the object as being on the scanout
for longer than is strictly required, causing us to be more proactive in
flushing -- though it should be true that we would have flushed
eventually when the back became the front, except that on the flip path
that flush is async but when hit from another ioctl it will be
synchronous.

v2: i915_gem_object_is_framebuffer()
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: Ville Syrjälä <ville.syrjala@linux.intel.com>
Reviewed-by: NVille Syrjälä <ville.syrjala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190902040303.14195-5-chris@chris-wilson.co.uk
      5a90606d
  7. 06 8月, 2019 1 次提交
  9. 03 8月, 2019 1 次提交
    • C
      drm/i915: Hide unshrinkable context objects from the shrinker · 1aff1903
      由 Chris Wilson 提交于 
      The shrinker cannot touch objects used by the contexts (logical state
and ring). Currently we mark those as "pin_global" to let the shrinker
skip over them, however, if we remove them from the shrinker lists
entirely, we don't event have to include them in our shrink accounting.

By keeping the unshrinkable objects in our shrinker tracking, we report
a large number of objects available to be shrunk, and leave the shrinker
deeply unsatisfied when we fail to reclaim those. The shrinker will
persist in trying to reclaim the unavailable objects, forcing the system
into a livelock (not even hitting the dread oomkiller).

v2: Extend unshrinkable protection for perma-pinned scratch and guc
allocations (Tvrtko)
v3: Notice that we should be pinned when marking unshrinkable and so the
link cannot be empty; merge duplicate paths.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: NMatthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190802212137.22207-1-chris@chris-wilson.co.uk
      1aff1903
  11. 03 7月, 2019 1 次提交
  13. 22 6月, 2019 2 次提交
  15. 19 6月, 2019 1 次提交
  17. 15 6月, 2019 1 次提交
    • C
      drm/i915: Keep contexts pinned until after the next kernel context switch · ce476c80
      由 Chris Wilson 提交于 
      We need to keep the context image pinned in memory until after the GPU
has finished writing into it. Since it continues to write as we signal
the final breadcrumb, we need to keep it pinned until the request after
it is complete. Currently we know the order in which requests execute on
each engine, and so to remove that presumption we need to identify a
request/context-switch we know must occur after our completion. Any
request queued after the signal must imply a context switch, for
simplicity we use a fresh request from the kernel context.

The sequence of operations for keeping the context pinned until saved is:

 - On context activation, we preallocate a node for each physical engine
   the context may operate on. This is to avoid allocations during
   unpinning, which may be from inside FS_RECLAIM context (aka the
   shrinker)

 - On context deactivation on retirement of the last active request (which
   is before we know the context has been saved), we add the
   preallocated node onto a barrier list on each engine

 - On engine idling, we emit a switch to kernel context. When this
   switch completes, we know that all previous contexts must have been
   saved, and so on retiring this request we can finally unpin all the
   contexts that were marked as deactivated prior to the switch.

We can enhance this in future by flushing all the idle contexts on a
regular heartbeat pulse of a switch to kernel context, which will also
be used to check for hung engines.

v2: intel_context_active_acquire/_release
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Reviewed-by: NMika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190614164606.15633-1-chris@chris-wilson.co.uk
      ce476c80
  19. 14 6月, 2019 2 次提交
  21. 12 6月, 2019 2 次提交
  23. 11 6月, 2019 1 次提交
  25. 01 6月, 2019 2 次提交
    • C
      drm/i915: Report all objects with allocated pages to the shrinker · d82b4b26
      由 Chris Wilson 提交于 
      Currently, we try to report to the shrinker the precise number of
objects (pages) that are available to be reaped at this moment. This
requires searching all objects with allocated pages to see if they
fulfill the search criteria, and this count is performed quite
frequently. (The shrinker tries to free ~128 pages on each invocation,
before which we count all the objects; counting takes longer than
unbinding the objects!) If we take the pragmatic view that with
sufficient desire, all objects are eventually reapable (they become
inactive, or no longer used as framebuffer etc), we can simply return
the count of pinned pages maintained during get_pages/put_pages rather
than walk the lists every time.

The downside is that we may (slightly) over-report the number of
objects/pages we could shrink and so penalize ourselves by shrinking
more than required. This is mitigated by keeping the order in which we
shrink objects such that we avoid penalizing active and frequently used
objects, and if memory is so tight that we need to free them we would
need to anyway.

v2: Only expose shrinkable objects to the shrinker; a small reduction in
not considering stolen and foreign objects.
v3: Restore the tracking from a "backup" copy from before the gem/ split
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Matthew Auld <matthew.auld@intel.com>
Reviewed-by: NMatthew Auld <matthew.auld@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190530203500.26272-2-chris@chris-wilson.co.uk
      d82b4b26
    • C
      drm/i915: Track the purgeable objects on a separate eviction list · 3b4fa964
      由 Chris Wilson 提交于 
      Currently the purgeable objects, I915_MADV_DONTNEED, are mixed in the
normal bound/unbound lists. Every shrinker pass starts with an attempt
to purge from this set of unneeded objects, which entails us doing a
walk over both lists looking for any candidates. If there are none, and
since we are shrinking we can reasonably assume that the lists are
full!, this becomes a very slow futile walk.

If we separate out the purgeable objects into own list, this search then
becomes its own phase that is preferentially handled during shrinking.
Instead the cost becomes that we then need to filter the purgeable list
if we want to distinguish between bound and unbound objects.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Matthew Auld <matthew.william.auld@gmail.com>
Reviewed-by: NMatthew Auld <matthew.william.auld@gmail.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190530203500.26272-1-chris@chris-wilson.co.uk
      3b4fa964
  27. 28 5月, 2019 2 次提交
  29. 20 4月, 2019 1 次提交
    • C
      drm/i915: Start writeback from the shrinker · 2d6692e6
      由 Chris Wilson 提交于 
      When we are called to relieve mempressue via the shrinker, the only way
we can make progress is either by discarding unwanted pages (those
objects that userspace has marked MADV_DONTNEED) or by reclaiming the
dirty objects via swap. As we know that is the only way to make further
progress, we can initiate the writeback as we invalidate the objects.
This means the objects we put onto the inactive anon lru list are
already marked for reclaim+writeback and so will trigger a wait upon the
writeback inside direct reclaim, greatly improving the success rate of
direct reclaim on i915 objects.

The corollary is that we may start a slow swap on opportunistic
mempressure from the likes of the compaction + migration kthreads. This
is limited by those threads only being allowed to shrink idle pages, but
also that if we reactivate the page before it is swapped out by gpu
activity, we only page the cost of repinning the page. The cost is most
felt when an object is reused after mempressure, which hopefully
excludes the latency sensitive tasks (as we are just extending the
impact of swap thrashing to them).

Apparently this is not the first time we've had this idea. Back in
commit 5537252b ("drm/i915: Invalidate our pages under memory
pressure") we wanted to start writeback but settled on invalidate after
Hugh Dickins warned us about a possibility of a deadlock within shmemfs
if we started writeback from shrink_slab. Looking at the callchain,
using writeback from i915_gem_shrink should be equivalent to the pageout
also employed by shrink_slab, i.e. it should not be any riskier afaict.

v2: Leave mmapings intact. At this point, the only mmapings of our
objects will be via CPU mmaps on the shmemfs filp, which are
out-of-scope for our LRU tracking. Instead leave those pages to the
inactive anon LRU page list for aging and pageout as normal.

v3: Be selective on which paths trigger writeback, in particular
excluding paths shrinking just to reclaim vm space (e.g. mmap, vmap
reapers) and avoid starting writeback on the entire process space from
within the pm freezer.

References: https://bugs.freedesktop.org/show_bug.cgi?id=108686Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: Matthew Auld <matthew.auld@intel.com>
Cc: Daniel Vetter <daniel.vetter@ffwll.ch>
Cc: Michal Hocko <mhocko@suse.com>
Reviewed-by: Joonas Lahtinen <joonas.lahtinen@linux.intel.com> #v1
Link: https://patchwork.freedesktop.org/patch/msgid/20190420115539.29081-1-chris@chris-wilson.co.uk
      2d6692e6
  31. 29 1月, 2019 2 次提交
    • C
      drm/i915: Pull VM lists under the VM mutex. · 09d7e46b
      由 Chris Wilson 提交于 
      A starting point to counter the pervasive struct_mutex. For the goal of
avoiding (or at least blocking under them!) global locks during user
request submission, a simple but important step is being able to manage
each clients GTT separately. For which, we want to replace using the
struct_mutex as the guard for all things GTT/VM and switch instead to a
specific mutex inside i915_address_space.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Reviewed-by: NTvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190128102356.15037-2-chris@chris-wilson.co.uk
      09d7e46b
    • C
      drm/i915: Stop tracking MRU activity on VMA · 499197dc
      由 Chris Wilson 提交于 
      Our goal is to remove struct_mutex and replace it with fine grained
locking. One of the thorny issues is our eviction logic for reclaiming
space for an execbuffer (or GTT mmaping, among a few other examples).
While eviction itself is easy to move under a per-VM mutex, performing
the activity tracking is less agreeable. One solution is not to do any
MRU tracking and do a simple coarse evaluation during eviction of
active/inactive, with a loose temporal ordering of last
insertion/evaluation. That keeps all the locking constrained to when we
are manipulating the VM itself, neatly avoiding the tricky handling of
possible recursive locking during execbuf and elsewhere.

Note that discarding the MRU (currently implemented as a pair of lists,
to avoid scanning the active list for a NONBLOCKING search) is unlikely
to impact upon our efficiency to reclaim VM space (where we think a LRU
model is best) as our current strategy is to use random idle replacement
first before doing a search, and over time the use of softpinned 48b
per-ppGTT is growing (thereby eliminating any need to perform any eviction
searches, in theory at least) with the remaining users being found on
much older devices (gen2-gen6).

v2: Changelog and commentary rewritten to elaborate on the duality of a
single list being both an inactive and active list.
v3: Consolidate bool parameters into a single set of flags; don't
comment on the duality of a single variable being a multiplicity of
bits.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: NTvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190128102356.15037-1-chris@chris-wilson.co.uk
      499197dc
  33. 15 1月, 2019 3 次提交
  35. 10 1月, 2019 2 次提交
  37. 09 1月, 2019 1 次提交
  39. 08 1月, 2019 1 次提交
    • C
      drm/i915: Return immediately if trylock fails for direct-reclaim · d25f71a1
      由 Chris Wilson 提交于 
      Ignore trying to shrink from i915 if we fail to acquire the struct_mutex
in the shrinker while performing direct-reclaim. The trade-off being
(much) lower latency for non-i915 clients at an increased risk of being
unable to obtain a page from direct-reclaim without hitting the
oom-notifier. The proviso being that we still keep trying to hard
obtain the lock for kswapd so that we can reap under heavy memory
pressure.

v2: Taint all mutexes taken within the shrinker with the struct_mutex
subclass as an early warning system, and drop I915_SHRINK_ACTIVE from
vmap to reduce the number of dangerous paths. We also have to drop
I915_SHRINK_ACTIVE from oom-notifier to be able to make the same claim
that ACTIVE is only used from outside context, which fits in with a
longer strategy of avoiding stalls due to scanning active during
shrinking.

The danger in using the subclass struct_mutex is that we declare
ourselves more knowledgable than lockdep and deprive ourselves of
automatic coverage. Instead, we require ourselves to mark up any mutex
taken inside the shrinker in order to detect lock-inversion, and if we
miss any we are doomed to a deadlock at the worst possible moment.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: NTvrtko Ursulin <tvrtko.ursulin@intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20190107115509.12523-1-chris@chris-wilson.co.uk
      d25f71a1
  41. 11 7月, 2018 1 次提交
  43. 09 7月, 2018 1 次提交
    • C
      drm/i915: Provide a timeout to i915_gem_wait_for_idle() · ec625fb9
      由 Chris Wilson 提交于 
      Usually we have no idea about the upper bound we need to wait to catch
up with userspace when idling the device, but in a few situations we
know the system was idle beforehand and can provide a short timeout in
order to very quickly catch a failure, long before hangcheck kicks in.

In the following patches, we will use the timeout to curtain two overly
long waits, where we know we can expect the GPU to complete within a
reasonable time or declare it broken.

In particular, with a broken GPU we expect it to fail during the initial
GPU setup where do a couple of context switches to record the defaults.
This is a task that takes a few milliseconds even on the slowest of
devices, but we may have to wait 60s for hangcheck to give in and
declare the machine inoperable. In this a case where any gpu hang is
unacceptable, both from a timeliness and practical standpoint.

The other improvement is that in selftests, we do not need to arm an
independent timer to inject a wedge, as we can just limit the timeout on
the wait directly.

v2: Include the timeout parameter in the trace.
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Cc: Mika Kuoppala <mika.kuoppala@linux.intel.com>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Reviewed-by: NMika Kuoppala <mika.kuoppala@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20180709122044.7028-1-chris@chris-wilson.co.uk
      ec625fb9
  45. 06 6月, 2018 1 次提交
  47. 22 2月, 2018 1 次提交
  49. 01 2月, 2018 1 次提交
  51. 18 1月, 2018 1 次提交
  53. 28 11月, 2017 2 次提交
  55. 09 11月, 2017 1 次提交
    • C
      drm/i915: Idle the GPU before shinking everything · 0676e794
      由 Chris Wilson 提交于 
      The handling of contexts are peculiar. Instead of tieing their vma to
activity, we pin the context. This means that we cannot simply unbind
the context object itself at will (which would normally cause us to wait
for the vma to be idle), but must manually idle the GPU and retire
requests first.

A consequence of this peculiarity is when doing a last desperate attempt
to recover memory. If the memory is tied up inside active context
objects, we will fail to recover any memory simply by trying to unbind
the objects without first doing a wait-for-idle.

A side-effect of removing the call to shrinker_lock_uninterruptible()
from i915_gem_shrinker_oom() was that we removed an unlocked
wait-for-idle, and so lost the "natural" shrinkage of context objects.
By replacing that with a locked wait from inside i915_gem_shrink(), we
not only replace it with the ability to recover all context objects, but
do so for all i915_gem_shrink_all() callers.

v2: Switching requires request allocation, which is not permitted from
inside the shrinker as it only uses ordinary allocations.

References: https://bugs.freedesktop.org/show_bug.cgi?id=102936
Fixes: f2123818 ("drm/i915: Move dev_priv->mm.[un]bound_list to its own lock")
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20171108094400.1386-1-chris@chris-wilson.co.ukReviewed-by: NJoonas Lahtinen <joonas.lahtinen@linux.intel.com>
(cherry picked from commit 2f6a3783)
Signed-off-by: NJani Nikula <jani.nikula@intel.com>
      0676e794
  57. 08 11月, 2017 1 次提交
    • C
      drm/i915: Idle the GPU before shinking everything · 2f6a3783
      由 Chris Wilson 提交于 
      The handling of contexts are peculiar. Instead of tieing their vma to
activity, we pin the context. This means that we cannot simply unbind
the context object itself at will (which would normally cause us to wait
for the vma to be idle), but must manually idle the GPU and retire
requests first.

A consequence of this peculiarity is when doing a last desperate attempt
to recover memory. If the memory is tied up inside active context
objects, we will fail to recover any memory simply by trying to unbind
the objects without first doing a wait-for-idle.

A side-effect of removing the call to shrinker_lock_uninterruptible()
from i915_gem_shrinker_oom() was that we removed an unlocked
wait-for-idle, and so lost the "natural" shrinkage of context objects.
By replacing that with a locked wait from inside i915_gem_shrink(), we
not only replace it with the ability to recover all context objects, but
do so for all i915_gem_shrink_all() callers.

v2: Switching requires request allocation, which is not permitted from
inside the shrinker as it only uses ordinary allocations.

References: https://bugs.freedesktop.org/show_bug.cgi?id=102936
Fixes: f2123818 ("drm/i915: Move dev_priv->mm.[un]bound_list to its own lock")
Signed-off-by: NChris Wilson <chris@chris-wilson.co.uk>
Cc: Joonas Lahtinen <joonas.lahtinen@linux.intel.com>
Link: https://patchwork.freedesktop.org/patch/msgid/20171108094400.1386-1-chris@chris-wilson.co.ukReviewed-by: NJoonas Lahtinen <joonas.lahtinen@linux.intel.com>
      2f6a3783
  59. 17 10月, 2017 1 次提交