Commit dc48e56d authored by Jens Axboe's avatar Jens Axboe
Browse files

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: default avatarJeff Moyer <>
Signed-off-by: default avatarJens Axboe <>
parent 2963e3f7
......@@ -77,6 +77,11 @@ struct kioctx_cpu {
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 @@ struct kioctx {
* signals when all in-flight requests are done
struct completion *requests_done;
struct ctx_rq_wait *rq_wait;
struct {
......@@ -535,8 +540,8 @@ static void free_ioctx_reqs(struct percpu_ref *ref)
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)
if (ctx->rq_wait && atomic_dec_and_test(&ctx->rq_wait->count))
INIT_WORK(&ctx->free_work, free_ioctx);
......@@ -744,7 +749,7 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
* 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 @@ static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
if (ctx->mmap_size)
vm_munmap(ctx->mmap_base, ctx->mmap_size);
ctx->requests_done = requests_done;
ctx->rq_wait = wait;
return 0;
......@@ -805,18 +810,24 @@ EXPORT_SYMBOL(wait_on_sync_kiocb);
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)
atomic_set(&wait.count, table->nr);
skipped = 0;
for (i = 0; i < table->nr; ++i) {
struct kioctx *ctx = table->table[i];
struct completion requests_done =
if (!ctx)
if (!ctx) {
* 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 @@ void exit_aio(struct mm_struct *mm)
* 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. */
RCU_INIT_POINTER(mm->ioctx_table, NULL);
......@@ -1313,15 +1326,17 @@ SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
struct kioctx *ioctx = lookup_ioctx(ctx);
if (likely(NULL != ioctx)) {
struct completion requests_done =
struct ctx_rq_wait wait;
int ret;
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);
/* Wait until all IO for the context are done. Otherwise kernel
......@@ -1329,7 +1344,7 @@ SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
* is destroyed.
if (!ret)
return ret;
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment