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>

arch/s390/mm/pgtable.c

@@ -263,7 +263,7 @@ int s390_enable_sie(void)
 	/* lets check if we are allowed to replace the mm */
 	task_lock(tsk);
 	if (!tsk->mm || atomic_read(&tsk->mm->mm_users) > 1 ||
-	    tsk->mm != tsk->active_mm || tsk->mm->ioctx_list) {
+	    tsk->mm != tsk->active_mm || !hlist_empty(&tsk->mm->ioctx_list)) {
 		task_unlock(tsk);
 		return -EINVAL;
 	}
@@ -279,7 +279,7 @@ int s390_enable_sie(void)
 	/* Now lets check again if something happened */
 	task_lock(tsk);
 	if (!tsk->mm || atomic_read(&tsk->mm->mm_users) > 1 ||
-	    tsk->mm != tsk->active_mm || tsk->mm->ioctx_list) {
+	    tsk->mm != tsk->active_mm || !hlist_empty(&tsk->mm->ioctx_list)) {
 		mmput(mm);
 		task_unlock(tsk);
 		return -EINVAL;

fs/aio.c

@@ -191,23 +191,11 @@ 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)
-/* __put_ioctx
- *	Called when the last user of an aio context has gone away,
- *	and the struct needs to be freed.
- */
-static void __put_ioctx(struct kioctx *ctx)
 {
+	struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
 	unsigned nr_events = ctx->max_reqs;
 
-	BUG_ON(ctx->reqs_active);
-
-	cancel_delayed_work(&ctx->wq);
-	cancel_work_sync(&ctx->wq.work);
-	aio_free_ring(ctx);
-	mmdrop(ctx->mm);
-	ctx->mm = NULL;
-	pr_debug("__put_ioctx: freeing %p\n", ctx);
 	kmem_cache_free(kioctx_cachep, ctx);
 
 	if (nr_events) {
@@ -218,6 +206,23 @@ static void __put_ioctx(struct kioctx *ctx)
 	}
 }
 
+/* __put_ioctx
+ *	Called when the last user of an aio context has gone away,
+ *	and the struct needs to be freed.
+ */
+static void __put_ioctx(struct kioctx *ctx)
+{
+	BUG_ON(ctx->reqs_active);
+
+	cancel_delayed_work(&ctx->wq);
+	cancel_work_sync(&ctx->wq.work);
+	aio_free_ring(ctx);
+	mmdrop(ctx->mm);
+	ctx->mm = NULL;
+	pr_debug("__put_ioctx: freeing %p\n", ctx);
+	call_rcu(&ctx->rcu_head, ctx_rcu_free);
+}
+
 #define get_ioctx(kioctx) do {						\
 	BUG_ON(atomic_read(&(kioctx)->users) <= 0);			\
 	atomic_inc(&(kioctx)->users);					\
@@ -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);
+	do {
-	if (aio_nr + ctx->max_reqs > aio_max_nr ||
+		spin_lock_bh(&aio_nr_lock);
-	    aio_nr + ctx->max_reqs < aio_nr)
+		if (aio_nr + nr_events > aio_max_nr ||
-		ctx->max_reqs = 0;
+		    aio_nr + nr_events < aio_nr)
-	else
+			ctx->max_reqs = 0;
-		aio_nr += ctx->max_reqs;
+		else
-	spin_unlock(&aio_nr_lock);
+			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);
+	spin_lock(&mm->ioctx_lock);
-	ctx->next = mm->ioctx_list;
+	hlist_add_head_rcu(&ctx->list, &mm->ioctx_list);
-	mm->ioctx_list = ctx;
+	spin_unlock(&mm->ioctx_lock);
-	write_unlock(&mm->ioctx_list_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;
+	struct kioctx *ctx;
-	mm->ioctx_list = NULL;
+
-	while (ctx) {
+	while (!hlist_empty(&mm->ioctx_list)) {
-		struct kioctx *next = ctx->next;
+		ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
-		ctx->next = NULL;
+		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 = current->mm;
-	struct mm_struct *mm;
+	struct kioctx *ctx = NULL;
+	struct hlist_node *n;
 
-	mm = current->mm;
+	rcu_read_lock();
-	read_lock(&mm->ioctx_list_lock);
+
-	for (ioctx = mm->ioctx_list; ioctx; ioctx = ioctx->next)
+	hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
-		if (likely(ioctx->user_id == ctx_id && !ioctx->dead)) {
+		if (ctx->user_id == ctx_id && !ctx->dead) {
-			get_ioctx(ioctx);
+			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;
+	hlist_del_rcu(&ioctx->list);
-	     tmp = &(*tmp)->next)
+	spin_unlock(&mm->ioctx_lock);
-		;
-	if (*tmp)
-		*tmp = ioctx->next;
-	write_unlock(&mm->ioctx_list_lock);
 
 	dprintk("aio_release(%p)\n", ioctx);
 	if (likely(!was_dead))

include/linux/aio.h

@@ -5,6 +5,7 @@
 #include <linux/workqueue.h>
 #include <linux/aio_abi.h>
 #include <linux/uio.h>
+#include <linux/rcupdate.h>
 
 #include <asm/atomic.h>
 
@@ -183,7 +184,7 @@ struct kioctx {
 
 	/* This needs improving */
 	unsigned long		user_id;
-	struct kioctx		*next;
+	struct hlist_node	list;
 
 	wait_queue_head_t	wait;
 
@@ -199,6 +200,8 @@ struct kioctx {
 	struct aio_ring_info	ring_info;
 
 	struct delayed_work	wq;
+
+	struct rcu_head		rcu_head;
 };
 
 /* prototypes */

include/linux/mm_types.h

@@ -232,8 +232,9 @@ struct mm_struct {
 	struct core_state *core_state; /* coredumping support */
 
 	/* aio bits */
-	rwlock_t		ioctx_list_lock;	/* aio lock */
+	spinlock_t		ioctx_lock;
-	struct kioctx		*ioctx_list;
+	struct hlist_head	ioctx_list;
+
 #ifdef CONFIG_MM_OWNER
 	/*
 	 * "owner" points to a task that is regarded as the canonical

kernel/fork.c

@@ -415,8 +415,8 @@ static struct mm_struct * mm_init(struct mm_struct * mm, struct task_struct *p)
 	set_mm_counter(mm, file_rss, 0);
 	set_mm_counter(mm, anon_rss, 0);
 	spin_lock_init(&mm->page_table_lock);
-	rwlock_init(&mm->ioctx_list_lock);
+	spin_lock_init(&mm->ioctx_lock);
-	mm->ioctx_list = NULL;
+	INIT_HLIST_HEAD(&mm->ioctx_list);
 	mm->free_area_cache = TASK_UNMAPPED_BASE;
 	mm->cached_hole_size = ~0UL;
 	mm_init_owner(mm, p);