bpf: Change bpf_mem_cache draining process.

The next patch will introduce cross-cpu llist access and existing
irq_work_sync() + drain_mem_cache() + rcu_barrier_tasks_trace() mechanism will
not be enough, since irq_work_sync() + drain_mem_cache() on cpu A won't
guarantee that llist on cpu A are empty. The free_bulk() on cpu B might add
objects back to llist of cpu A. Add 'bool draining' flag.
The modified sequence looks like:
for_each_cpu:
  WRITE_ONCE(c->draining, true); // do_call_rcu_ttrace() won't be doing call_rcu() any more
  irq_work_sync(); // wait for irq_work callback (free_bulk) to finish
  drain_mem_cache(); // free all objects
rcu_barrier_tasks_trace(); // wait for RCU callbacks to execute

Signed-off-by: Alexei Starovoitov <ast@kernel.org>
Signed-off-by: Daniel Borkmann <daniel@iogearbox.net>
Acked-by: Hou Tao <houtao1@huawei.com>
Link: https://lore.kernel.org/bpf/20230706033447.54696-8-alexei.starovoitov@gmail.com

Author: Alexei Starovoitov

kernel/bpf/memalloc.c

@@ -98,6 +98,7 @@ struct bpf_mem_cache {
 	int free_cnt;
 	int low_watermark, high_watermark, batch;
 	int percpu_size;
+	bool draining;
 
 	/* list of objects to be freed after RCU tasks trace GP */
 	struct llist_head free_by_rcu_ttrace;
@@ -301,6 +302,12 @@ static void do_call_rcu_ttrace(struct bpf_mem_cache *c)
 		 * from __free_rcu() and from drain_mem_cache().
 		 */
 		__llist_add(llnode, &c->waiting_for_gp_ttrace);
+
+	if (unlikely(READ_ONCE(c->draining))) {
+		__free_rcu(&c->rcu_ttrace);
+		return;
+	}
+
 	/* Use call_rcu_tasks_trace() to wait for sleepable progs to finish.
 	 * If RCU Tasks Trace grace period implies RCU grace period, free
 	 * these elements directly, else use call_rcu() to wait for normal
@@ -544,15 +551,7 @@ void bpf_mem_alloc_destroy(struct bpf_mem_alloc *ma)
 		rcu_in_progress = 0;
 		for_each_possible_cpu(cpu) {
 			c = per_cpu_ptr(ma->cache, cpu);
-			/*
-			 * refill_work may be unfinished for PREEMPT_RT kernel
-			 * in which irq work is invoked in a per-CPU RT thread.
-			 * It is also possible for kernel with
-			 * arch_irq_work_has_interrupt() being false and irq
-			 * work is invoked in timer interrupt. So waiting for
-			 * the completion of irq work to ease the handling of
-			 * concurrency.
-			 */
+			WRITE_ONCE(c->draining, true);
 			irq_work_sync(&c->refill_work);
 			drain_mem_cache(c);
 			rcu_in_progress += atomic_read(&c->call_rcu_ttrace_in_progress);
@@ -568,6 +567,7 @@ void bpf_mem_alloc_destroy(struct bpf_mem_alloc *ma)
 			cc = per_cpu_ptr(ma->caches, cpu);
 			for (i = 0; i < NUM_CACHES; i++) {
 				c = &cc->cache[i];
+				WRITE_ONCE(c->draining, true);
 				irq_work_sync(&c->refill_work);
 				drain_mem_cache(c);
 				rcu_in_progress += atomic_read(&c->call_rcu_ttrace_in_progress);