On 10/16/20 12:44 PM, Maarten Lankhorst wrote:
Instead of force unbinding and rebinding every time, we try to check if our notifier seqcount is still correct when pages are bound. This way we only rebind userptr when we need to, and prevent stalls. Changes since v1: - Missing mutex_unlock, reported by kbuild. Reported-by: kernel test robot <l...@intel.com> Reported-by: Dan Carpenter <dan.carpen...@oracle.com> Signed-off-by: Maarten Lankhorst <maarten.lankho...@linux.intel.com> --- drivers/gpu/drm/i915/gem/i915_gem_userptr.c | 27 ++++++++++++++++++--- 1 file changed, 24 insertions(+), 3 deletions(-) diff --git a/drivers/gpu/drm/i915/gem/i915_gem_userptr.c b/drivers/gpu/drm/i915/gem/i915_gem_userptr.c index 8f05b6d90d54..b3fd5eecf0a1 100644 --- a/drivers/gpu/drm/i915/gem/i915_gem_userptr.c +++ b/drivers/gpu/drm/i915/gem/i915_gem_userptr.c @@ -268,12 +268,33 @@ int i915_gem_object_userptr_submit_init(struct drm_i915_gem_object *obj) if (ret) return ret;- /* Make sure userptr is unbound for next attempt, so we don't use stale pages. */- ret = i915_gem_object_userptr_unbind(obj, false); + /* optimistically try to preserve current pages while unlocked */ + if (i915_gem_object_has_pages(obj) && + !mmu_interval_check_retry(&obj->userptr.notifier, + obj->userptr.notifier_seq)) { + spin_lock(&i915->mm.notifier_lock); + if (obj->userptr.pvec && + !mmu_interval_read_retry(&obj->userptr.notifier, + obj->userptr.notifier_seq)) {
In theory, although extremely unlikely, on 32-bit the internal seqno may wrap giving a false negative above, That could be remedied by clearing obj->userptr.notifier_seq in the notifier. (Even seqnos will always be considered invalid).
+ obj->userptr.page_ref++; + + /* We can keep using the current binding, this is the fastpath */ + ret = 1; + } + spin_unlock(&i915->mm.notifier_lock); + } + + if (!ret) { + /* Make sure userptr is unbound for next attempt, so we don't use stale pages. */ + ret = i915_gem_object_userptr_unbind(obj, false); + } i915_gem_object_unlock(obj); - if (ret) + if (ret < 0) return ret;+ if (ret > 0)+ return 0; + notifier_seq = mmu_interval_read_begin(&obj->userptr.notifier);
This function will sleep until the core is done with the last invalidation, during which time all validation checks above will return false. We should place this call before any other checks in the function.
Furthermore the sleep is uninterruptible. We probably need a core change to get this right.
pvec = kvmalloc_array(num_pages, sizeof(struct page *), GFP_KERNEL);
_______________________________________________ Intel-gfx mailing list Intel-gfx@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/intel-gfx
