drm/i915: Avoid lock dropping between rescheduling
authorChris Wilson <chris@chris-wilson.co.uk>
Mon, 27 Mar 2017 20:21:43 +0000 (21:21 +0100)
committerJani Nikula <jani.nikula@intel.com>
Fri, 31 Mar 2017 10:39:47 +0000 (13:39 +0300)
Unlocking is dangerous. In this case we combine an early update to the
out-of-queue request, because we know that it will be inserted into the
correct FIFO priority-ordered slot when it becomes ready in the future.
However, given sufficient enthusiasm, it may become ready as we are
continuing to reschedule, and so may gazump the FIFO if we have since
dropped its spinlock. The result is that it may be executed too early,
before its dependencies.

v2: Move all work into the second phase over the topological sort. This
removes the shortcut on the out-of-rbtree request to ensure that we only
adjust its priority after adjusting all of its dependencies.

Fixes: 20311bd35060 ("drm/i915/scheduler: Execute requests in order of priorities")
Testcase: igt/gem_exec_whisper
Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
Cc: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
Cc: <stable@vger.kernel.org> # v4.10+
Link: http://patchwork.freedesktop.org/patch/msgid/20170327202143.7972-1-chris@chris-wilson.co.uk
Reviewed-by: Tvrtko Ursulin <tvrtko.ursulin@intel.com>
(cherry picked from commit a79a524e9260d4ffaff88348615e70fb3d393692)
Signed-off-by: Jani Nikula <jani.nikula@intel.com>
drivers/gpu/drm/i915/intel_lrc.c

index 91555d4e912991690c06149fddd4a1b8aaddb2f9..47517a02f0a439125b3b3a769e6848a4c4928ca2 100644 (file)
@@ -670,15 +670,14 @@ static void execlists_submit_request(struct drm_i915_gem_request *request)
 static struct intel_engine_cs *
 pt_lock_engine(struct i915_priotree *pt, struct intel_engine_cs *locked)
 {
-       struct intel_engine_cs *engine;
+       struct intel_engine_cs *engine =
+               container_of(pt, struct drm_i915_gem_request, priotree)->engine;
+
+       GEM_BUG_ON(!locked);
 
-       engine = container_of(pt,
-                             struct drm_i915_gem_request,
-                             priotree)->engine;
        if (engine != locked) {
-               if (locked)
-                       spin_unlock_irq(&locked->timeline->lock);
-               spin_lock_irq(&engine->timeline->lock);
+               spin_unlock(&locked->timeline->lock);
+               spin_lock(&engine->timeline->lock);
        }
 
        return engine;
@@ -686,7 +685,7 @@ pt_lock_engine(struct i915_priotree *pt, struct intel_engine_cs *locked)
 
 static void execlists_schedule(struct drm_i915_gem_request *request, int prio)
 {
-       struct intel_engine_cs *engine = NULL;
+       struct intel_engine_cs *engine;
        struct i915_dependency *dep, *p;
        struct i915_dependency stack;
        LIST_HEAD(dfs);
@@ -720,26 +719,23 @@ static void execlists_schedule(struct drm_i915_gem_request *request, int prio)
        list_for_each_entry_safe(dep, p, &dfs, dfs_link) {
                struct i915_priotree *pt = dep->signaler;
 
-               list_for_each_entry(p, &pt->signalers_list, signal_link)
+               /* Within an engine, there can be no cycle, but we may
+                * refer to the same dependency chain multiple times
+                * (redundant dependencies are not eliminated) and across
+                * engines.
+                */
+               list_for_each_entry(p, &pt->signalers_list, signal_link) {
+                       GEM_BUG_ON(p->signaler->priority < pt->priority);
                        if (prio > READ_ONCE(p->signaler->priority))
                                list_move_tail(&p->dfs_link, &dfs);
+               }
 
                list_safe_reset_next(dep, p, dfs_link);
-               if (!RB_EMPTY_NODE(&pt->node))
-                       continue;
-
-               engine = pt_lock_engine(pt, engine);
-
-               /* If it is not already in the rbtree, we can update the
-                * priority inplace and skip over it (and its dependencies)
-                * if it is referenced *again* as we descend the dfs.
-                */
-               if (prio > pt->priority && RB_EMPTY_NODE(&pt->node)) {
-                       pt->priority = prio;
-                       list_del_init(&dep->dfs_link);
-               }
        }
 
+       engine = request->engine;
+       spin_lock_irq(&engine->timeline->lock);
+
        /* Fifo and depth-first replacement ensure our deps execute before us */
        list_for_each_entry_safe_reverse(dep, p, &dfs, dfs_link) {
                struct i915_priotree *pt = dep->signaler;
@@ -751,16 +747,15 @@ static void execlists_schedule(struct drm_i915_gem_request *request, int prio)
                if (prio <= pt->priority)
                        continue;
 
-               GEM_BUG_ON(RB_EMPTY_NODE(&pt->node));
-
                pt->priority = prio;
-               rb_erase(&pt->node, &engine->execlist_queue);
-               if (insert_request(pt, &engine->execlist_queue))
-                       engine->execlist_first = &pt->node;
+               if (!RB_EMPTY_NODE(&pt->node)) {
+                       rb_erase(&pt->node, &engine->execlist_queue);
+                       if (insert_request(pt, &engine->execlist_queue))
+                               engine->execlist_first = &pt->node;
+               }
        }
 
-       if (engine)
-               spin_unlock_irq(&engine->timeline->lock);
+       spin_unlock_irq(&engine->timeline->lock);
 
        /* XXX Do we need to preempt to make room for us and our deps? */
 }