aio: fix serial draining in exit_aio()
exit_aio() currently serializes killing io contexts. Each context
killing ends up having to do percpu_ref_kill(), which in turns has
to wait for an RCU grace period. This can take a long time, depending
on the number of contexts. And there's no point in doing them serially,
when we could be waiting for all of them in one fell swoop.
This patches makes my fio thread offload test case exit 0.2s instead
of almost 6s.
Reviewed-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Jens Axboe <axboe@fb.com>
diff --git a/fs/aio.c b/fs/aio.c
index f8e52a1..cabb5ed 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -77,6 +77,11 @@
unsigned reqs_available;
};
+struct ctx_rq_wait {
+ struct completion comp;
+ atomic_t count;
+};
+
struct kioctx {
struct percpu_ref users;
atomic_t dead;
@@ -115,7 +120,7 @@
/*
* signals when all in-flight requests are done
*/
- struct completion *requests_done;
+ struct ctx_rq_wait *rq_wait;
struct {
/*
@@ -535,8 +540,8 @@
struct kioctx *ctx = container_of(ref, struct kioctx, reqs);
/* At this point we know that there are no any in-flight requests */
- if (ctx->requests_done)
- complete(ctx->requests_done);
+ if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
+ complete(&ctx->rq_wait->comp);
INIT_WORK(&ctx->free_work, free_ioctx);
schedule_work(&ctx->free_work);
@@ -744,7 +749,7 @@
* the rapid destruction of the kioctx.
*/
static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
- struct completion *requests_done)
+ struct ctx_rq_wait *wait)
{
struct kioctx_table *table;
@@ -773,7 +778,7 @@
if (ctx->mmap_size)
vm_munmap(ctx->mmap_base, ctx->mmap_size);
- ctx->requests_done = requests_done;
+ ctx->rq_wait = wait;
percpu_ref_kill(&ctx->users);
return 0;
}
@@ -805,18 +810,24 @@
void exit_aio(struct mm_struct *mm)
{
struct kioctx_table *table = rcu_dereference_raw(mm->ioctx_table);
- int i;
+ struct ctx_rq_wait wait;
+ int i, skipped;
if (!table)
return;
+ atomic_set(&wait.count, table->nr);
+ init_completion(&wait.comp);
+
+ skipped = 0;
for (i = 0; i < table->nr; ++i) {
struct kioctx *ctx = table->table[i];
- struct completion requests_done =
- COMPLETION_INITIALIZER_ONSTACK(requests_done);
- if (!ctx)
+ if (!ctx) {
+ skipped++;
continue;
+ }
+
/*
* We don't need to bother with munmap() here - exit_mmap(mm)
* is coming and it'll unmap everything. And we simply can't,
@@ -825,10 +836,12 @@
* that it needs to unmap the area, just set it to 0.
*/
ctx->mmap_size = 0;
- kill_ioctx(mm, ctx, &requests_done);
+ kill_ioctx(mm, ctx, &wait);
+ }
+ if (!atomic_sub_and_test(skipped, &wait.count)) {
/* Wait until all IO for the context are done. */
- wait_for_completion(&requests_done);
+ wait_for_completion(&wait.comp);
}
RCU_INIT_POINTER(mm->ioctx_table, NULL);
@@ -1313,15 +1326,17 @@
{
struct kioctx *ioctx = lookup_ioctx(ctx);
if (likely(NULL != ioctx)) {
- struct completion requests_done =
- COMPLETION_INITIALIZER_ONSTACK(requests_done);
+ struct ctx_rq_wait wait;
int ret;
+ init_completion(&wait.comp);
+ atomic_set(&wait.count, 1);
+
/* Pass requests_done to kill_ioctx() where it can be set
* in a thread-safe way. If we try to set it here then we have
* a race condition if two io_destroy() called simultaneously.
*/
- ret = kill_ioctx(current->mm, ioctx, &requests_done);
+ ret = kill_ioctx(current->mm, ioctx, &wait);
percpu_ref_put(&ioctx->users);
/* Wait until all IO for the context are done. Otherwise kernel
@@ -1329,7 +1344,7 @@
* is destroyed.
*/
if (!ret)
- wait_for_completion(&requests_done);
+ wait_for_completion(&wait.comp);
return ret;
}