aio: make the lookup_ioctx() lockless

The mm->ioctx_list is currently protected by a reader-writer lock,
so we always grab that lock on the read side for doing ioctx
lookups. As the workload is extremely reader biased, turn this into
an rcu hlist so we can make lookup_ioctx() lockless. Get rid of
the rwlock and use a spinlock for providing update side exclusion.

There's usually only 1 entry on this list, so it doesn't make sense
to look into fancier data structures.

Reviewed-by: Jeff Moyer <jmoyer@redhat.com>
Signed-off-by: Jens Axboe <jens.axboe@oracle.com>

1 files changed, 56 insertions, 44 deletions

diff --git a/fs/aio.c b/fs/aio.c
index f658441..d6f89d3 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -191,6 +191,20 @@ static int aio_setup_ring(struct kioctx *ctx)
 	kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK), km); \
 } while(0)
 
+static void ctx_rcu_free(struct rcu_head *head)
+{
+	struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
+	unsigned nr_events = ctx->max_reqs;
+
+	kmem_cache_free(kioctx_cachep, ctx);
+
+	if (nr_events) {
+		spin_lock(&aio_nr_lock);
+		BUG_ON(aio_nr - nr_events > aio_nr);
+		aio_nr -= nr_events;
+		spin_unlock(&aio_nr_lock);
+	}
+}
 
 /* __put_ioctx
  *	Called when the last user of an aio context has gone away,
@@ -198,8 +212,6 @@ static int aio_setup_ring(struct kioctx *ctx)
  */
 static void __put_ioctx(struct kioctx *ctx)
 {
-	unsigned nr_events = ctx->max_reqs;
-
 	BUG_ON(ctx->reqs_active);
 
 	cancel_delayed_work(&ctx->wq);
@@ -208,14 +220,7 @@ static void __put_ioctx(struct kioctx *ctx)
 	mmdrop(ctx->mm);
 	ctx->mm = NULL;
 	pr_debug("__put_ioctx: freeing %p\n", ctx);
-	kmem_cache_free(kioctx_cachep, ctx);
-
-	if (nr_events) {
-		spin_lock(&aio_nr_lock);
-		BUG_ON(aio_nr - nr_events > aio_nr);
-		aio_nr -= nr_events;
-		spin_unlock(&aio_nr_lock);
-	}
+	call_rcu(&ctx->rcu_head, ctx_rcu_free);
 }
 
 #define get_ioctx(kioctx) do {						\
@@ -235,6 +240,7 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
 {
 	struct mm_struct *mm;
 	struct kioctx *ctx;
+	int did_sync = 0;
 
 	/* Prevent overflows */
 	if ((nr_events > (0x10000000U / sizeof(struct io_event))) ||
@@ -267,21 +273,30 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
 		goto out_freectx;
 
 	/* limit the number of system wide aios */
-	spin_lock(&aio_nr_lock);
-	if (aio_nr + ctx->max_reqs > aio_max_nr ||
-	    aio_nr + ctx->max_reqs < aio_nr)
-		ctx->max_reqs = 0;
-	else
-		aio_nr += ctx->max_reqs;
-	spin_unlock(&aio_nr_lock);
+	do {
+		spin_lock_bh(&aio_nr_lock);
+		if (aio_nr + nr_events > aio_max_nr ||
+		    aio_nr + nr_events < aio_nr)
+			ctx->max_reqs = 0;
+		else
+			aio_nr += ctx->max_reqs;
+		spin_unlock_bh(&aio_nr_lock);
+		if (ctx->max_reqs || did_sync)
+			break;
+
+		/* wait for rcu callbacks to have completed before giving up */
+		synchronize_rcu();
+		did_sync = 1;
+		ctx->max_reqs = nr_events;
+	} while (1);
+
 	if (ctx->max_reqs == 0)
 		goto out_cleanup;
 
 	/* now link into global list. */
-	write_lock(&mm->ioctx_list_lock);
-	ctx->next = mm->ioctx_list;
-	mm->ioctx_list = ctx;
-	write_unlock(&mm->ioctx_list_lock);
+	spin_lock(&mm->ioctx_lock);
+	hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
+	spin_unlock(&mm->ioctx_lock);
 
 	dprintk("aio: allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
 		ctx, ctx->user_id, current->mm, ctx->ring_info.nr);
@@ -375,11 +390,12 @@ ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
  */
 void exit_aio(struct mm_struct *mm)
 {
-	struct kioctx *ctx = mm->ioctx_list;
-	mm->ioctx_list = NULL;
-	while (ctx) {
-		struct kioctx *next = ctx->next;
-		ctx->next = NULL;
+	struct kioctx *ctx;
+
+	while (!hlist_empty(&mm->ioctx_list)) {
+		ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
+		hlist_del_rcu(&ctx->list);
+
 		aio_cancel_all(ctx);
 
 		wait_for_all_aios(ctx);
@@ -394,7 +410,6 @@ void exit_aio(struct mm_struct *mm)
 				atomic_read(&ctx->users), ctx->dead,
 				ctx->reqs_active);
 		put_ioctx(ctx);
-		ctx = next;
 	}
 }
 
@@ -555,19 +570,21 @@ int aio_put_req(struct kiocb *req)
 
 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
 {
-	struct kioctx *ioctx;
-	struct mm_struct *mm;
+	struct mm_struct *mm = current->mm;
+	struct kioctx *ctx = NULL;
+	struct hlist_node *n;
 
-	mm = current->mm;
-	read_lock(&mm->ioctx_list_lock);
-	for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next)
-		if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) {
-			get_ioctx(ioctx);
+	rcu_read_lock();
+
+	hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
+		if (ctx->user_id == ctx_id && !ctx->dead) {
+			get_ioctx(ctx);
 			break;
 		}
-	read_unlock(&mm->ioctx_list_lock);
+	}
 
-	return ioctx;
+	rcu_read_unlock();
+	return ctx;
 }
 
 /*
@@ -1215,19 +1232,14 @@ out:
 static void io_destroy(struct kioctx *ioctx)
 {
 	struct mm_struct *mm = current->mm;
-	struct kioctx **tmp;
 	int was_dead;
 
 	/* delete the entry from the list is someone else hasn't already */
-	write_lock(&mm->ioctx_list_lock);
+	spin_lock(&mm->ioctx_lock);
 	was_dead = ioctx->dead;
 	ioctx->dead = 1;
-	for (tmp = &mm->ioctx_list; *tmp && *tmp != ioctx;
-	     tmp = &(*tmp)->next)
-		;
-	if (*tmp)
-		*tmp = ioctx->next;
-	write_unlock(&mm->ioctx_list_lock);
+	hlist_del_rcu(&ioctx->list);
+	spin_unlock(&mm->ioctx_lock);
 
 	dprintk("aio_release(%p)\n", ioctx);
 	if (likely(!was_dead))