/* * Active refs memory management * * To be more economical with memory, we reap all the i915_active trees as * they idle (when we know the active requests are inactive) and allocate the * nodes from a local slab cache to hopefully reduce the fragmentation.
*/ staticstruct kmem_cache *slab_cache;
staticvoid debug_active_deactivate(struct i915_active *ref)
{
lockdep_assert_held(&ref->tree_lock); if (!atomic_read(&ref->count)) /* after the last dec */
debug_object_deactivate(ref, &active_debug_desc);
}
/* return the unused nodes to our slabcache -- flushing the allocator */ if (!atomic_dec_and_lock_irqsave(&ref->count, &ref->tree_lock, flags)) return;
/* Even if we have not used the cache, we may still have a barrier */ if (!ref->cache)
ref->cache = fetch_node(ref->tree.rb_node);
/* Keep the MRU cached node for reuse */ if (ref->cache) { /* Discard all other nodes in the tree */
rb_erase(&ref->cache->node, &ref->tree);
root = ref->tree;
/* Rebuild the tree with only the cached node */
rb_link_node(&ref->cache->node, NULL, &ref->tree.rb_node);
rb_insert_color(&ref->cache->node, &ref->tree);
GEM_BUG_ON(ref->tree.rb_node != &ref->cache->node);
/* Make the cached node available for reuse with any timeline */
ref->cache->timeline = 0; /* needs cmpxchg(u64) */
}
spin_unlock_irqrestore(&ref->tree_lock, flags);
/* After the final retire, the entire struct may be freed */ if (ref->retire)
ref->retire(ref);
/* ... except if you wait on it, you must manage your own references! */
wake_up_var(ref);
/* Finally free the discarded timeline tree */
rbtree_postorder_for_each_entry_safe(it, n, &root, node) {
GEM_BUG_ON(i915_active_fence_isset(&it->base));
kmem_cache_free(slab_cache, it);
}
}
GEM_BUG_ON(idx == 0); /* 0 is the unordered timeline, rsvd for cache */
/* * We track the most recently used timeline to skip a rbtree search * for the common case, under typical loads we never need the rbtree * at all. We can reuse the last slot if it is empty, that is * after the previous activity has been retired, or if it matches the * current timeline.
*/
it = READ_ONCE(ref->cache); if (it) {
u64 cached = READ_ONCE(it->timeline);
/* Once claimed, this slot will only belong to this idx */ if (cached == idx) return it;
/* * An unclaimed cache [.timeline=0] can only be claimed once. * * If the value is already non-zero, some other thread has * claimed the cache and we know that is does not match our * idx. If, and only if, the timeline is currently zero is it * worth competing to claim it atomically for ourselves (for * only the winner of that race will cmpxchg return the old * value of 0).
*/ if (!cached && !cmpxchg64(&it->timeline, 0, idx)) return it;
}
BUILD_BUG_ON(offsetof(typeof(*it), node));
/* While active, the tree can only be built; not destroyed */
GEM_BUG_ON(i915_active_is_idle(ref));
it = fetch_node(ref->tree.rb_node); while (it) { if (it->timeline < idx) {
it = fetch_node(it->node.rb_right);
} elseif (it->timeline > idx) {
it = fetch_node(it->node.rb_left);
} else {
WRITE_ONCE(ref->cache, it); break;
}
}
/* NB: If the tree rotated beneath us, we may miss our target. */ return it;
}
if (node->timeline < idx)
p = &parent->rb_right; else
p = &parent->rb_left;
}
/* * XXX: We should preallocate this before i915_active_ref() is ever * called, but we cannot call into fs_reclaim() anyway, so use GFP_ATOMIC.
*/
node = kmem_cache_alloc(slab_cache, GFP_ATOMIC); if (!node) goto out;
/* * Rebuild the llist excluding our node. We may perform this * outside of the kernel_context timeline mutex and so someone * else may be manipulating the engine->barrier_tasks, in * which case either we or they will be upset :) * * A second __active_del_barrier() will report failure to claim * the active_node and the caller will just shrug and know not to * claim ownership of its node. * * A concurrent i915_request_add_active_barriers() will miss adding * any of the tasks, but we will try again on the next -- and since * we are actively using the barrier, we know that there will be * at least another opportunity when we idle.
*/
llist_for_each_safe(pos, next, llist_del_all(&engine->barrier_tasks)) { if (node == barrier_from_ll(pos)) {
node = NULL; continue;
}
pos->next = head;
head = pos; if (!tail)
tail = pos;
} if (head)
llist_add_batch(head, tail, &engine->barrier_tasks);
staticbool
replace_barrier(struct i915_active *ref, struct i915_active_fence *active)
{ if (!is_barrier(active)) /* proto-node used by our idle barrier? */ returnfalse;
/* * This request is on the kernel_context timeline, and so * we can use it to substitute for the pending idle-barrer * request that we want to emit on the kernel_context.
*/ return __active_del_barrier(ref, node_from_active(active));
}
enable_signaling(&ref->excl);
rbtree_postorder_for_each_entry_safe(it, n, &ref->tree, node) {
err = flush_barrier(it); /* unconnected idle barrier? */ if (err) break;
enable_signaling(&it->base);
}
return err;
}
int __i915_active_wait(struct i915_active *ref, int state)
{
might_sleep();
/* Any fence added after the wait begins will not be auto-signaled */ if (i915_active_acquire_if_busy(ref)) { int err;
err = flush_lazy_signals(ref);
i915_active_release(ref); if (err) return err;
if (___wait_var_event(ref, i915_active_is_idle(ref),
state, 0, 0, schedule())) return -EINTR;
}
/* * After the wait is complete, the caller may free the active. * We have to flush any concurrent retirement before returning.
*/
flush_work(&ref->work); return 0;
}
/* * Try to reuse any existing barrier nodes already allocated for this * i915_active, due to overlapping active phases there is likely a * node kept alive (as we reuse before parking). We prefer to reuse * completely idle barriers (less hassle in manipulating the llists), * but otherwise any will do.
*/ if (ref->cache && is_idle_barrier(ref->cache, idx)) {
p = &ref->cache->node; goto match;
}
prev = NULL;
p = ref->tree.rb_node; while (p) { struct active_node *node =
rb_entry(p, struct active_node, node);
if (is_idle_barrier(node, idx)) goto match;
prev = p; if (node->timeline < idx)
p = READ_ONCE(p->rb_right); else
p = READ_ONCE(p->rb_left);
}
/* * No quick match, but we did find the leftmost rb_node for the * kernel_context. Walk the rb_tree in-order to see if there were * any idle-barriers on this timeline that we missed, or just use * the first pending barrier.
*/ for (p = prev; p; p = rb_next(p)) { struct active_node *node =
rb_entry(p, struct active_node, node); struct intel_engine_cs *engine;
if (node->timeline > idx) break;
if (node->timeline < idx) continue;
if (is_idle_barrier(node, idx)) goto match;
/* * The list of pending barriers is protected by the * kernel_context timeline, which notably we do not hold * here. i915_request_add_active_barriers() may consume * the barrier before we claim it, so we have to check * for success.
*/
engine = __barrier_to_engine(node);
smp_rmb(); /* serialise with add_active_barriers */ if (is_barrier(&node->base) &&
____active_del_barrier(ref, node, engine)) goto match;
}
return NULL;
match:
spin_lock_irq(&ref->tree_lock);
rb_erase(p, &ref->tree); /* Hide from waits and sibling allocations */ if (p == &ref->cache->node)
WRITE_ONCE(ref->cache, NULL);
spin_unlock_irq(&ref->tree_lock);
/* Wait until the previous preallocation is completed */ while (!llist_empty(&ref->preallocated_barriers))
cond_resched();
/* * Preallocate a node for each physical engine supporting the target * engine (remember virtual engines have more than one sibling). * We can then use the preallocated nodes in * i915_active_acquire_barrier()
*/
GEM_BUG_ON(!mask);
for_each_engine_masked(engine, gt, mask, tmp) {
u64 idx = engine->kernel_context->timeline->fence_context; struct llist_node *prev = first; struct active_node *node;
rcu_read_lock();
node = reuse_idle_barrier(ref, idx);
rcu_read_unlock(); if (!node) {
node = kmem_cache_alloc(slab_cache, GFP_KERNEL); if (!node) goto unwind;
if (!i915_active_fence_isset(&node->base)) { /* * Mark this as being *our* unconnected proto-node. * * Since this node is not in any list, and we have * decoupled it from the rbtree, we can reuse the * request to indicate this is an idle-barrier node * and then we can use the rb_node and list pointers * for our tracking of the pending barrier.
*/
RCU_INIT_POINTER(node->base.fence, ERR_PTR(-EAGAIN));
node->base.cb.node.prev = (void *)engine;
__i915_active_acquire(ref);
}
GEM_BUG_ON(rcu_access_pointer(node->base.fence) != ERR_PTR(-EAGAIN));
GEM_BUG_ON(barrier_to_engine(node) != engine);
first = barrier_to_ll(node);
first->next = prev; if (!last)
last = first;
intel_engine_pm_get(engine);
}
GEM_BUG_ON(!llist_empty(&ref->preallocated_barriers));
llist_add_batch(first, last, &ref->preallocated_barriers);
return 0;
unwind: while (first) { struct active_node *node = barrier_from_ll(first);
/* * Transfer the list of preallocated barriers into the * i915_active rbtree, but only as proto-nodes. They will be * populated by i915_request_add_active_barriers() to point to the * request that will eventually release them.
*/
llist_for_each_safe(pos, next, take_preallocated_barriers(ref)) { struct active_node *node = barrier_from_ll(pos); struct intel_engine_cs *engine = barrier_to_engine(node); struct rb_node **p, *parent;
spin_lock_irqsave_nested(&ref->tree_lock, flags,
SINGLE_DEPTH_NESTING);
parent = NULL;
p = &ref->tree.rb_node; while (*p) { struct active_node *it;
parent = *p;
it = rb_entry(parent, struct active_node, node); if (it->timeline < node->timeline)
p = &parent->rb_right; else
p = &parent->rb_left;
}
rb_link_node(&node->node, parent, p);
rb_insert_color(&node->node, &ref->tree);
spin_unlock_irqrestore(&ref->tree_lock, flags);
node = llist_del_all(&engine->barrier_tasks); if (!node) return; /* * Attach the list of proto-fences to the in-flight request such * that the parent i915_active will be released when this request * is retired.
*/
spin_lock_irqsave(&rq->lock, flags);
llist_for_each_safe(node, next, node) { /* serialise with reuse_idle_barrier */
smp_store_mb(*ll_to_fence_slot(node), &rq->fence);
list_add_tail((struct list_head *)node, &rq->fence.cb_list);
}
spin_unlock_irqrestore(&rq->lock, flags);
}
/* * __i915_active_fence_set: Update the last active fence along its timeline * @active: the active tracker * @fence: the new fence (under construction) * * Records the new @fence as the last active fence along its timeline in * this active tracker, moving the tracking callbacks from the previous * fence onto this one. Gets and returns a reference to the previous fence * (if not already completed), which the caller must put after making sure * that it is executed before the new fence. To ensure that the order of * fences within the timeline of the i915_active_fence is understood, it * should be locked by the caller.
*/ struct dma_fence *
__i915_active_fence_set(struct i915_active_fence *active, struct dma_fence *fence)
{ struct dma_fence *prev; unsignedlong flags;
/* * In case of fences embedded in i915_requests, their memory is * SLAB_FAILSAFE_BY_RCU, then it can be reused right after release * by new requests. Then, there is a risk of passing back a pointer * to a new, completely unrelated fence that reuses the same memory * while tracked under a different active tracker. Combined with i915 * perf open/close operations that build await dependencies between * engine kernel context requests and user requests from different * timelines, this can lead to dependency loops and infinite waits. * * As a countermeasure, we try to get a reference to the active->fence * first, so if we succeed and pass it back to our user then it is not * released and potentially reused by an unrelated request before the * user has a chance to set up an await dependency on it.
*/
prev = i915_active_fence_get(active); if (fence == prev) return fence;
/* * Consider that we have two threads arriving (A and B), with * C already resident as the active->fence. * * Both A and B have got a reference to C or NULL, depending on the * timing of the interrupt handler. Let's assume that if A has got C * then it has locked C first (before B). * * Note the strong ordering of the timeline also provides consistent * nesting rules for the fence->lock; the inner lock is always the * older lock.
*/
spin_lock_irqsave(fence->lock, flags); if (prev)
spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
/* * A does the cmpxchg first, and so it sees C or NULL, as before, or * something else, depending on the timing of other threads and/or * interrupt handler. If not the same as before then A unlocks C if * applicable and retries, starting from an attempt to get a new * active->fence. Meanwhile, B follows the same path as A. * Once A succeeds with cmpxch, B fails again, retires, gets A from * active->fence, locks it as soon as A completes, and possibly * succeeds with cmpxchg.
*/ while (cmpxchg(__active_fence_slot(active), prev, fence) != prev) { if (prev) {
spin_unlock(prev->lock);
dma_fence_put(prev);
}
spin_unlock_irqrestore(fence->lock, flags);
spin_lock_irqsave(fence->lock, flags); if (prev)
spin_lock_nested(prev->lock, SINGLE_DEPTH_NESTING);
}
/* * If prev is NULL then the previous fence must have been signaled * and we know that we are first on the timeline. If it is still * present then, having the lock on that fence already acquired, we * serialise with the interrupt handler, in the process of removing it * from any future interrupt callback. A will then wait on C before * executing (if present). * * As B is second, it sees A as the previous fence and so waits for * it to complete its transition and takes over the occupancy for * itself -- remembering that it needs to wait on A before executing.
*/ if (prev) {
__list_del_entry(&active->cb.node);
spin_unlock(prev->lock); /* serialise with prev->cb_list */
}
list_add_tail(&active->cb.node, &fence->cb_list);
spin_unlock_irqrestore(fence->lock, flags);
return prev;
}
int i915_active_fence_set(struct i915_active_fence *active, struct i915_request *rq)
{ struct dma_fence *fence; int err = 0;
/* Must maintain timeline ordering wrt previous active requests */
fence = __i915_active_fence_set(active, &rq->fence); if (fence) {
err = i915_request_await_dma_fence(rq, fence);
dma_fence_put(fence);
}
Die Informationen auf dieser Webseite wurden
nach bestem Wissen sorgfältig zusammengestellt. Es wird jedoch weder Vollständigkeit, noch Richtigkeit,
noch Qualität der bereit gestellten Informationen zugesichert.
Bemerkung:
Die farbliche Syntaxdarstellung und die Messung sind noch experimentell.
