History

rcu: Support reclaim for head-less object

Update the kvfree_call_rcu() function with head-less support. This allows RCU to reclaim objects without an embedded rcu_head. tree-RCU: We introduce two chains of arrays to store SLAB-backed and vmalloc pointers, each. Storage in either of these arrays does not require embedding an rcu_head within the object. Maintaining the arrays may become impossible due to high memory pressure. For such cases there is an emergency path. Objects with rcu_head inside are just queued on a backup rcu_head list. Later on that list is drained. As for the head-less variant, as the current context can sleep, the following emergency measures are applied: a) Synchronously wait until a grace period has elapsed. b) Call kvfree(). tiny-RCU: For double argument calls, there are no new changes in behavior. For single argument call, kvfree() is directly inlined on the current stack after a synchronize_rcu() call. Note that for tiny-RCU, any call to synchronize_rcu() is actually a quiescent state, therefore it does nothing.

Uladzislau Rezki (Sony) committed 6 years ago

Commit 3042f831 parent ce4dce12 files · +60 −3Browse files

Unified · Split

2 files
  1. include/linux/rcutiny.h +17 −1
  2. kernel/rcu/tree.c +43 −2
include/linux/rcutiny.h+17−1
@@ -34,9 +34,25 @@ static inline void synchronize_rcu_expedited(void) · 33 unmodified lines
3434 synchronize_rcu();
3535 }
3636
37+/*
38+ * Add one more declaration of kvfree() here. It is
39+ * not so straight forward to just include <linux/mm.h>
40+ * where it is defined due to getting many compile
41+ * errors caused by that include.
42+ */
43+extern void kvfree(const void *addr);
44+
3745 static inline void kvfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
3846 {
39− call_rcu(head, func);
47+ if (head) {
48+ call_rcu(head, func);
49+ return;
50+ }
51+
52+ // kvfree_rcu(one_arg) call.
53+ might_sleep();
54+ synchronize_rcu();
55+ kvfree((void *) func);
4056 }
4157
4258 void rcu_qs(void);
kernel/rcu/tree.c+43−2
@@ -3314,6 +3314,13 @@ kvfree_call_rcu_add_ptr_to_bulk(struct kfree_rcu_cpu *krcp, void *ptr) · 3313 unmodified lines
33143314 if (IS_ENABLED(CONFIG_PREEMPT_RT))
33153315 return false;
33163316
3317+ /*
3318+ * NOTE: For one argument of kvfree_rcu() we can
3319+ * drop the lock and get the page in sleepable
3320+ * context. That would allow to maintain an array
3321+ * for the CONFIG_PREEMPT_RT as well if no cached
3322+ * pages are available.
3323+ */
33173324 bnode = (struct kvfree_rcu_bulk_data *)
33183325 __get_free_page(GFP_NOWAIT | __GFP_NOWARN);
33193326 }
@@ -3353,16 +3360,33 @@ void kvfree_call_rcu(struct rcu_head *head, rcu_callback_t func) · 33 unmodified lines
33533360 {
33543361 unsigned long flags;
33553362 struct kfree_rcu_cpu *krcp;
3363+ bool success;
33563364 void *ptr;
33573365
3366+ if (head) {
3367+ ptr = (void *) head - (unsigned long) func;
3368+ } else {
3369+ /*
3370+ * Please note there is a limitation for the head-less
3371+ * variant, that is why there is a clear rule for such
3372+ * objects: it can be used from might_sleep() context
3373+ * only. For other places please embed an rcu_head to
3374+ * your data.
3375+ */
3376+ might_sleep();
3377+ ptr = (unsigned long *) func;
3378+ }
3379+
33583380 krcp = krc_this_cpu_lock(&flags);
3359− ptr = (void *)head - (unsigned long)func;
33603381
33613382 // Queue the object but don't yet schedule the batch.
33623383 if (debug_rcu_head_queue(ptr)) {
33633384 // Probable double kfree_rcu(), just leak.
33643385 WARN_ONCE(1, "%s(): Double-freed call. rcu_head %p\n",
33653386 __func__, head);
3387+
3388+ // Mark as success and leave.
3389+ success = true;
33663390 goto unlock_return;
33673391 }
33683392
@@ -3370,10 +3394,16 @@ void kvfree_call_rcu(struct rcu_head *head, rcu_callback_t func) · 1 unmodified line
33703394 * Under high memory pressure GFP_NOWAIT can fail,
33713395 * in that case the emergency path is maintained.
33723396 */
3373− if (unlikely(!kvfree_call_rcu_add_ptr_to_bulk(krcp, ptr))) {
3397+ success = kvfree_call_rcu_add_ptr_to_bulk(krcp, ptr);
3398+ if (!success) {
3399+ if (head == NULL)
3400+ // Inline if kvfree_rcu(one_arg) call.
3401+ goto unlock_return;
3402+
33743403 head->func = func;
33753404 head->next = krcp->head;
33763405 krcp->head = head;
3406+ success = true;
33773407 }
33783408
33793409 WRITE_ONCE(krcp->count, krcp->count + 1);
@@ -3387,6 +3417,17 @@ void kvfree_call_rcu(struct rcu_head *head, rcu_callback_t func) · 7 unmodified lines
33873417
33883418 unlock_return:
33893419 krc_this_cpu_unlock(krcp, flags);
3420+
3421+ /*
3422+ * Inline kvfree() after synchronize_rcu(). We can do
3423+ * it from might_sleep() context only, so the current
3424+ * CPU can pass the QS state.
3425+ */
3426+ if (!success) {
3427+ debug_rcu_head_unqueue((struct rcu_head *) ptr);
3428+ synchronize_rcu();
3429+ kvfree(ptr);
3430+ }
33903431 }
33913432 EXPORT_SYMBOL_GPL(kvfree_call_rcu);
33923433