AIO performance improvements/cleanups, v3

index 3dcfaf4..8a8cd02 100644
@@ -14,9 +14,6 @@
  * use_mm
  *	Makes the calling kernel thread take on the specified
  *	mm context.
- *	Called by the retry thread execute retries within the
- *	iocb issuer's mm context, so that copy_from/to_user
- *	operations work seamlessly for aio.
  *	(Note: this routine is intended to be called only
  *	from a kernel thread context)
  */

index 31ff6db..b46a09f 100644
@@ -9,44 +9,22 @@
 
 #include <linux/atomic.h>
 
-#define AIO_MAXSEGS		4
-#define AIO_KIOGRP_NR_ATOMIC	8
-
 struct kioctx;
 
-/* Notes on cancelling a kiocb:
- *	If a kiocb is cancelled, aio_complete may return 0 to indicate 
- *	that cancel has not yet disposed of the kiocb.  All cancel 
- *	operations *must* call aio_put_req to dispose of the kiocb 
- *	to guard against races with the completion code.
- */
-#define KIOCB_C_CANCELLED	0x01
-#define KIOCB_C_COMPLETE	0x02
-
 #define KIOCB_SYNC_KEY		(~0U)
 
 /* ki_flags bits */
-/*
- * This may be used for cancel/retry serialization in the future, but
- * for now it's unused and we probably don't want modules to even
- * think they can use it.
- */
-/* #define KIF_LOCKED		0 */
 #define KIF_KICKED		1
 #define KIF_CANCELLED		2
 
-#define kiocbTryLock(iocb)	test_and_set_bit(KIF_LOCKED, &(iocb)->ki_flags)
 #define kiocbTryKick(iocb)	test_and_set_bit(KIF_KICKED, &(iocb)->ki_flags)
 
-#define kiocbSetLocked(iocb)	set_bit(KIF_LOCKED, &(iocb)->ki_flags)
 #define kiocbSetKicked(iocb)	set_bit(KIF_KICKED, &(iocb)->ki_flags)
 #define kiocbSetCancelled(iocb)	set_bit(KIF_CANCELLED, &(iocb)->ki_flags)
 
-#define kiocbClearLocked(iocb)	clear_bit(KIF_LOCKED, &(iocb)->ki_flags)
 #define kiocbClearKicked(iocb)	clear_bit(KIF_KICKED, &(iocb)->ki_flags)
 #define kiocbClearCancelled(iocb)	clear_bit(KIF_CANCELLED, &(iocb)->ki_flags)
 
-#define kiocbIsLocked(iocb)	test_bit(KIF_LOCKED, &(iocb)->ki_flags)
 #define kiocbIsKicked(iocb)	test_bit(KIF_KICKED, &(iocb)->ki_flags)
 #define kiocbIsCancelled(iocb)	test_bit(KIF_CANCELLED, &(iocb)->ki_flags)
 

@@ -207,8 +185,6 @@ struct kioctx {
 };
 
 /* prototypes */
-extern unsigned aio_max_size;
-
 #ifdef CONFIG_AIO
 extern ssize_t wait_on_sync_kiocb(struct kiocb *iocb);
 extern int aio_put_req(struct kiocb *iocb);

index 76494ca..2a3f001 100644
@@ -24,6 +24,7 @@
 #include <linux/sched.h>
 #include <linux/slab.h>
 #include <linux/poll.h>
+#include <linux/mmu_context.h>
 
 #include <linux/device.h>
 #include <linux/moduleparam.h>

@@ -514,6 +515,9 @@ static long ep_ioctl(struct file *fd, unsigned code, unsigned long value)
 struct kiocb_priv {
 	struct usb_request	*req;
 	struct ep_data		*epdata;
+	struct kiocb 		*iocb;
+	struct mm_struct 	*mm;
+	struct work_struct	work;
 	void			*buf;
 	const struct iovec	*iv;
 	unsigned long		nr_segs;

@@ -541,15 +545,12 @@ static int ep_aio_cancel(struct kiocb *iocb, struct io_event *e)
 	return value;
 }
 
-static ssize_t ep_aio_read_retry(struct kiocb *iocb)
+static ssize_t ep_copy_to_user(struct kiocb_priv *priv)
 {
-	struct kiocb_priv	*priv = iocb->private;
 	ssize_t			len, total;
 	void			*to_copy;
 	int			i;
 
-	/* we "retry" to get the right mm context for this: */
-
 	/* copy stuff into user buffers */
 	total = priv->actual;
 	len = 0;

@@ -569,9 +570,26 @@ static ssize_t ep_aio_read_retry(struct kiocb *iocb)
 		if (total == 0)
 			break;
 	}
+
+	return len;
+}
+
+static void ep_user_copy_worker(struct work_struct *work)
+{
+	struct kiocb_priv *priv = container_of(work, struct kiocb_priv, work);
+	struct mm_struct *mm = priv->mm;
+	struct kiocb *iocb = priv->iocb;
+	size_t ret;
+
+	use_mm(mm);
+	ret = ep_copy_to_user(priv);
+	unuse_mm(mm);
+
+	/* completing the iocb can drop the ctx and mm, don't touch mm after */
+	aio_complete(iocb, ret, ret);
+
 	kfree(priv->buf);
 	kfree(priv);
-	return len;
 }
 
 static void ep_aio_complete(struct usb_ep *ep, struct usb_request *req)

@@ -597,14 +615,14 @@ static void ep_aio_complete(struct usb_ep *ep, struct usb_request *req)
 		aio_complete(iocb, req->actual ? req->actual : req->status,
 				req->status);
 	} else {
-		/* retry() won't report both; so we hide some faults */
+		/* ep_copy_to_user() won't report both; we hide some faults */
 		if (unlikely(0 != req->status))
 			DBG(epdata->dev, "%s fault %d len %d\n",
 				ep->name, req->status, req->actual);
 
 		priv->buf = req->buf;
 		priv->actual = req->actual;
-		kick_iocb(iocb);
+		schedule_work(&priv->work);
 	}
 	spin_unlock(&epdata->dev->lock);
 

@@ -634,8 +652,10 @@ fail:
 		return value;
 	}
 	iocb->private = priv;
+	priv->iocb = iocb;
 	priv->iv = iv;
 	priv->nr_segs = nr_segs;
+	INIT_WORK(&priv->work, ep_user_copy_worker);
 
 	value = get_ready_ep(iocb->ki_filp->f_flags, epdata);
 	if (unlikely(value < 0)) {

@@ -647,6 +667,7 @@ fail:
 	get_ep(epdata);
 	priv->epdata = epdata;
 	priv->actual = 0;
+	priv->mm = current->mm; /* mm teardown waits for iocbs in exit_aio() */
 
 	/* each kiocb is coupled to one usb_request, but we can't
 	 * allocate or submit those if the host disconnected.

@@ -675,7 +696,7 @@ fail:
 		kfree(priv);
 		put_ep(epdata);
 	} else
-		value = (iv ? -EIOCBRETRY : -EIOCBQUEUED);
+		value = -EIOCBQUEUED;
 	return value;
 }
 

@@ -693,7 +714,6 @@ ep_aio_read(struct kiocb *iocb, const struct iovec *iov,
 	if (unlikely(!buf))
 		return -ENOMEM;
 
-	iocb->ki_retry = ep_aio_read_retry;
 	return ep_aio_rwtail(iocb, buf, iocb->ki_left, epdata, iov, nr_segs);
 }
 

index 71f613c..1de4f78 100644
@@ -54,11 +54,6 @@ unsigned long aio_max_nr = 0x10000; /* system wide maximum number of aio request
 static struct kmem_cache	*kiocb_cachep;
 static struct kmem_cache	*kioctx_cachep;
 
-static struct workqueue_struct *aio_wq;
-
-static void aio_kick_handler(struct work_struct *);
-static void aio_queue_work(struct kioctx *);
-
 /* aio_setup
  *	Creates the slab caches used by the aio routines, panic on
  *	failure as this is done early during the boot sequence.

@@ -68,9 +63,6 @@ static int __init aio_setup(void)
 	kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
 	kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
 
-	aio_wq = alloc_workqueue("aio", 0, 1);	/* used to limit concurrency */
-	BUG_ON(!aio_wq);
-
 	pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
 
 	return 0;

@@ -86,7 +78,6 @@ static void aio_free_ring(struct kioctx *ctx)
 		put_page(info->ring_pages[i]);
 
 	if (info->mmap_size) {
-		BUG_ON(ctx->mm != current->mm);
 		vm_munmap(info->mmap_base, info->mmap_size);
 	}
 

@@ -101,6 +92,7 @@ static int aio_setup_ring(struct kioctx *ctx)
 	struct aio_ring *ring;
 	struct aio_ring_info *info = &ctx->ring_info;
 	unsigned nr_events = ctx->max_reqs;
+	struct mm_struct *mm = current->mm;
 	unsigned long size;
 	int nr_pages;
 

@@ -126,22 +118,21 @@ static int aio_setup_ring(struct kioctx *ctx)
 
 	info->mmap_size = nr_pages * PAGE_SIZE;
 	dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
-	down_write(&ctx->mm->mmap_sem);
+	down_write(&mm->mmap_sem);
 	info->mmap_base = do_mmap_pgoff(NULL, 0, info->mmap_size, 
 					PROT_READ|PROT_WRITE,
 					MAP_ANONYMOUS|MAP_PRIVATE, 0);
 	if (IS_ERR((void *)info->mmap_base)) {
-		up_write(&ctx->mm->mmap_sem);
+		up_write(&mm->mmap_sem);
 		info->mmap_size = 0;
 		aio_free_ring(ctx);
 		return -EAGAIN;
 	}
 
 	dprintk("mmap address: 0x%08lx\n", info->mmap_base);
-	info->nr_pages = get_user_pages(current, ctx->mm,
-					info->mmap_base, nr_pages, 
+	info->nr_pages = get_user_pages(current, mm, info->mmap_base, nr_pages, 
 					1, 0, info->ring_pages, NULL);
-	up_write(&ctx->mm->mmap_sem);
+	up_write(&mm->mmap_sem);
 
 	if (unlikely(info->nr_pages != nr_pages)) {
 		aio_free_ring(ctx);

@@ -203,10 +194,7 @@ static void __put_ioctx(struct kioctx *ctx)
 	unsigned nr_events = ctx->max_reqs;
 	BUG_ON(ctx->reqs_active);
 
-	cancel_delayed_work_sync(&ctx->wq);
 	aio_free_ring(ctx);
-	mmdrop(ctx->mm);
-	ctx->mm = NULL;
 	if (nr_events) {
 		spin_lock(&aio_nr_lock);
 		BUG_ON(aio_nr - nr_events > aio_nr);

@@ -234,7 +222,7 @@ static inline void put_ioctx(struct kioctx *kioctx)
  */
 static struct kioctx *ioctx_alloc(unsigned nr_events)
 {
-	struct mm_struct *mm;
+	struct mm_struct *mm = current->mm;
 	struct kioctx *ctx;
 	int err = -ENOMEM;
 

@@ -253,8 +241,6 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
 		return ERR_PTR(-ENOMEM);
 
 	ctx->max_reqs = nr_events;
-	mm = ctx->mm = current->mm;
-	atomic_inc(&mm->mm_count);
 
 	atomic_set(&ctx->users, 2);
 	spin_lock_init(&ctx->ctx_lock);

@@ -262,8 +248,6 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
 	init_waitqueue_head(&ctx->wait);
 
 	INIT_LIST_HEAD(&ctx->active_reqs);
-	INIT_LIST_HEAD(&ctx->run_list);
-	INIT_DELAYED_WORK(&ctx->wq, aio_kick_handler);
 
 	if (aio_setup_ring(ctx) < 0)
 		goto out_freectx;

@@ -284,14 +268,13 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
 	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);
+		ctx, ctx->user_id, mm, ctx->ring_info.nr);
 	return ctx;
 
 out_cleanup:
 	err = -EAGAIN;
 	aio_free_ring(ctx);
 out_freectx:
-	mmdrop(mm);
 	kmem_cache_free(kioctx_cachep, ctx);
 	dprintk("aio: error allocating ioctx %d\n", err);
 	return ERR_PTR(err);

@@ -388,8 +371,6 @@ void exit_aio(struct mm_struct *mm)
 		 * as indicator that it needs to unmap the area,
 		 * just set it to 0; aio_free_ring() is the only
 		 * place that uses ->mmap_size, so it's safe.
-		 * That way we get all munmap done to current->mm -
-		 * all other callers have ctx->mm == current->mm.
 		 */
 		ctx->ring_info.mmap_size = 0;
 		put_ioctx(ctx);

@@ -423,7 +404,6 @@ static struct kiocb *__aio_get_req(struct kioctx *ctx)
 	req->ki_dtor = NULL;
 	req->private = NULL;
 	req->ki_iovec = NULL;
-	INIT_LIST_HEAD(&req->ki_run_list);
 	req->ki_eventfd = NULL;
 
 	return req;

@@ -609,281 +589,6 @@ static struct kioctx *lookup_ioctx(unsigned long ctx_id)
 	return ret;
 }
 
-/*
- * Queue up a kiocb to be retried. Assumes that the kiocb
- * has already been marked as kicked, and places it on
- * the retry run list for the corresponding ioctx, if it
- * isn't already queued. Returns 1 if it actually queued
- * the kiocb (to tell the caller to activate the work
- * queue to process it), or 0, if it found that it was
- * already queued.
- */
-static inline int __queue_kicked_iocb(struct kiocb *iocb)
-{
-	struct kioctx *ctx = iocb->ki_ctx;
-
-	assert_spin_locked(&ctx->ctx_lock);
-
-	if (list_empty(&iocb->ki_run_list)) {
-		list_add_tail(&iocb->ki_run_list,
-			&ctx->run_list);
-		return 1;
-	}
-	return 0;
-}
-
-/* aio_run_iocb
- *	This is the core aio execution routine. It is
- *	invoked both for initial i/o submission and
- *	subsequent retries via the aio_kick_handler.
- *	Expects to be invoked with iocb->ki_ctx->lock
- *	already held. The lock is released and reacquired
- *	as needed during processing.
- *
- * Calls the iocb retry method (already setup for the
- * iocb on initial submission) for operation specific
- * handling, but takes care of most of common retry
- * execution details for a given iocb. The retry method
- * needs to be non-blocking as far as possible, to avoid
- * holding up other iocbs waiting to be serviced by the
- * retry kernel thread.
- *
- * The trickier parts in this code have to do with
- * ensuring that only one retry instance is in progress
- * for a given iocb at any time. Providing that guarantee
- * simplifies the coding of individual aio operations as
- * it avoids various potential races.
- */
-static ssize_t aio_run_iocb(struct kiocb *iocb)
-{
-	struct kioctx	*ctx = iocb->ki_ctx;
-	ssize_t (*retry)(struct kiocb *);
-	ssize_t ret;
-
-	if (!(retry = iocb->ki_retry)) {
-		printk("aio_run_iocb: iocb->ki_retry = NULL\n");
-		return 0;
-	}
-
-	/*
-	 * We don't want the next retry iteration for this
-	 * operation to start until this one has returned and
-	 * updated the iocb state. However, wait_queue functions
-	 * can trigger a kick_iocb from interrupt context in the
-	 * meantime, indicating that data is available for the next
-	 * iteration. We want to remember that and enable the
-	 * next retry iteration _after_ we are through with
-	 * this one.
-	 *
-	 * So, in order to be able to register a "kick", but
-	 * prevent it from being queued now, we clear the kick
-	 * flag, but make the kick code *think* that the iocb is
-	 * still on the run list until we are actually done.
-	 * When we are done with this iteration, we check if
-	 * the iocb was kicked in the meantime and if so, queue
-	 * it up afresh.
-	 */
-
-	kiocbClearKicked(iocb);
-
-	/*
-	 * This is so that aio_complete knows it doesn't need to
-	 * pull the iocb off the run list (We can't just call
-	 * INIT_LIST_HEAD because we don't want a kick_iocb to
-	 * queue this on the run list yet)
-	 */
-	iocb->ki_run_list.next = iocb->ki_run_list.prev = NULL;
-	spin_unlock_irq(&ctx->ctx_lock);
-
-	/* Quit retrying if the i/o has been cancelled */
-	if (kiocbIsCancelled(iocb)) {
-		ret = -EINTR;
-		aio_complete(iocb, ret, 0);
-		/* must not access the iocb after this */
-		goto out;
-	}
-
-	/*
-	 * Now we are all set to call the retry method in async
-	 * context.
-	 */
-	ret = retry(iocb);
-
-	if (ret != -EIOCBRETRY && ret != -EIOCBQUEUED) {
-		/*
-		 * There's no easy way to restart the syscall since other AIO's
-		 * may be already running. Just fail this IO with EINTR.
-		 */
-		if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
-			     ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
-			ret = -EINTR;
-		aio_complete(iocb, ret, 0);
-	}
-out:
-	spin_lock_irq(&ctx->ctx_lock);
-
-	if (-EIOCBRETRY == ret) {
-		/*
-		 * OK, now that we are done with this iteration
-		 * and know that there is more left to go,
-		 * this is where we let go so that a subsequent
-		 * "kick" can start the next iteration
-		 */
-
-		/* will make __queue_kicked_iocb succeed from here on */
-		INIT_LIST_HEAD(&iocb->ki_run_list);
-		/* we must queue the next iteration ourselves, if it
-		 * has already been kicked */
-		if (kiocbIsKicked(iocb)) {
-			__queue_kicked_iocb(iocb);
-
-			/*
-			 * __queue_kicked_iocb will always return 1 here, because
-			 * iocb->ki_run_list is empty at this point so it should
-			 * be safe to unconditionally queue the context into the
-			 * work queue.
-			 */
-			aio_queue_work(ctx);
-		}
-	}
-	return ret;
-}
-
-/*
- * __aio_run_iocbs:
- * 	Process all pending retries queued on the ioctx
- * 	run list.
- * Assumes it is operating within the aio issuer's mm
- * context.
- */
-static int __aio_run_iocbs(struct kioctx *ctx)
-{
-	struct kiocb *iocb;
-	struct list_head run_list;
-
-	assert_spin_locked(&ctx->ctx_lock);
-
-	list_replace_init(&ctx->run_list, &run_list);
-	while (!list_empty(&run_list)) {
-		iocb = list_entry(run_list.next, struct kiocb,
-			ki_run_list);
-		list_del(&iocb->ki_run_list);
-		/*
-		 * Hold an extra reference while retrying i/o.
-		 */
-		iocb->ki_users++;       /* grab extra reference */
-		aio_run_iocb(iocb);
-		__aio_put_req(ctx, iocb);
- 	}
-	if (!list_empty(&ctx->run_list))
-		return 1;
-	return 0;
-}
-
-static void aio_queue_work(struct kioctx * ctx)
-{
-	unsigned long timeout;
-	/*
-	 * if someone is waiting, get the work started right
-	 * away, otherwise, use a longer delay
-	 */
-	smp_mb();
-	if (waitqueue_active(&ctx->wait))
-		timeout = 1;
-	else
-		timeout = HZ/10;
-	queue_delayed_work(aio_wq, &ctx->wq, timeout);
-}
-
-/*
- * aio_run_all_iocbs:
- *	Process all pending retries queued on the ioctx
- *	run list, and keep running them until the list
- *	stays empty.
- * Assumes it is operating within the aio issuer's mm context.
- */
-static inline void aio_run_all_iocbs(struct kioctx *ctx)
-{
-	spin_lock_irq(&ctx->ctx_lock);
-	while (__aio_run_iocbs(ctx))
-		;
-	spin_unlock_irq(&ctx->ctx_lock);
-}
-
-/*
- * aio_kick_handler:
- * 	Work queue handler triggered to process pending
- * 	retries on an ioctx. Takes on the aio issuer's
- *	mm context before running the iocbs, so that
- *	copy_xxx_user operates on the issuer's address
- *      space.
- * Run on aiod's context.
- */
-static void aio_kick_handler(struct work_struct *work)
-{
-	struct kioctx *ctx = container_of(work, struct kioctx, wq.work);
-	mm_segment_t oldfs = get_fs();
-	struct mm_struct *mm;
-	int requeue;
-
-	set_fs(USER_DS);
-	use_mm(ctx->mm);
-	spin_lock_irq(&ctx->ctx_lock);
-	requeue =__aio_run_iocbs(ctx);
-	mm = ctx->mm;
-	spin_unlock_irq(&ctx->ctx_lock);
- 	unuse_mm(mm);
-	set_fs(oldfs);
-	/*
-	 * we're in a worker thread already; no point using non-zero delay
-	 */
-	if (requeue)
-		queue_delayed_work(aio_wq, &ctx->wq, 0);
-}
-
-
-/*
- * Called by kick_iocb to queue the kiocb for retry
- * and if required activate the aio work queue to process
- * it
- */
-static void try_queue_kicked_iocb(struct kiocb *iocb)
-{
- 	struct kioctx	*ctx = iocb->ki_ctx;
-	unsigned long flags;
-	int run = 0;
-
-	spin_lock_irqsave(&ctx->ctx_lock, flags);
-	/* set this inside the lock so that we can't race with aio_run_iocb()
-	 * testing it and putting the iocb on the run list under the lock */
-	if (!kiocbTryKick(iocb))
-		run = __queue_kicked_iocb(iocb);
-	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
-	if (run)
-		aio_queue_work(ctx);
-}
-
-/*
- * kick_iocb:
- *      Called typically from a wait queue callback context
- *      to trigger a retry of the iocb.
- *      The retry is usually executed by aio workqueue
- *      threads (See aio_kick_handler).
- */
-void kick_iocb(struct kiocb *iocb)
-{
-	/* sync iocbs are easy: they can only ever be executing from a 
-	 * single context. */
-	if (is_sync_kiocb(iocb)) {
-		kiocbSetKicked(iocb);
-	        wake_up_process(iocb->ki_obj.tsk);
-		return;
-	}
-
-	try_queue_kicked_iocb(iocb);
-}
-EXPORT_SYMBOL(kick_iocb);
-
 /* aio_complete
  *	Called when the io request on the given iocb is complete.
  *	Returns true if this is the last user of the request.  The 

@@ -924,9 +629,6 @@ int aio_complete(struct kiocb *iocb, long res, long res2)
 	 */
 	spin_lock_irqsave(&ctx->ctx_lock, flags);
 
-	if (iocb->ki_run_list.prev && !list_empty(&iocb->ki_run_list))
-		list_del_init(&iocb->ki_run_list);
-
 	/*
 	 * cancelled requests don't get events, userland was given one
 	 * when the event got cancelled.

@@ -1081,13 +783,11 @@ static int read_events(struct kioctx *ctx,
 	int			i = 0;
 	struct io_event		ent;
 	struct aio_timeout	to;
-	int			retry = 0;
 
 	/* needed to zero any padding within an entry (there shouldn't be 
 	 * any, but C is fun!
 	 */
 	memset(&ent, 0, sizeof(ent));
-retry:
 	ret = 0;
 	while (likely(i < nr)) {
 		ret = aio_read_evt(ctx, &ent);

@@ -1117,13 +817,6 @@ retry:
 
 	/* End fast path */
 
-	/* racey check, but it gets redone */
-	if (!retry && unlikely(!list_empty(&ctx->run_list))) {
-		retry = 1;
-		aio_run_all_iocbs(ctx);
-		goto retry;
-	}
-
 	init_timeout(&to);
 	if (timeout) {
 		struct timespec	ts;

@@ -1343,7 +1036,7 @@ static ssize_t aio_rw_vect_retry(struct kiocb *iocb)
 	/* If we managed to write some out we return that, rather than
 	 * the eventual error. */
 	if (opcode == IOCB_CMD_PWRITEV
-	    && ret < 0 && ret != -EIOCBQUEUED && ret != -EIOCBRETRY
+	    && ret < 0 && ret != -EIOCBQUEUED
 	    && iocb->ki_nbytes - iocb->ki_left)
 		ret = iocb->ki_nbytes - iocb->ki_left;
 

@@ -1585,18 +1278,27 @@ static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
 	 * don't see ctx->dead set here, io_destroy() waits for our IO to
 	 * finish.
 	 */
-	if (ctx->dead) {
-		spin_unlock_irq(&ctx->ctx_lock);
+	if (ctx->dead)
 		ret = -EINVAL;
+	spin_unlock_irq(&ctx->ctx_lock);
+	if (ret)
 		goto out_put_req;
+
+	if (unlikely(kiocbIsCancelled(req))) {
+		ret = -EINTR;
+	} else {
+		ret = req->ki_retry(req);
 	}
-	aio_run_iocb(req);
-	if (!list_empty(&ctx->run_list)) {
-		/* drain the run list */
-		while (__aio_run_iocbs(ctx))
-			;
+	if (ret != -EIOCBQUEUED) {
+		/*
+		 * There's no easy way to restart the syscall since other AIO's
+		 * may be already running. Just fail this IO with EINTR.
+		 */
+		if (unlikely(ret == -ERESTARTSYS || ret == -ERESTARTNOINTR ||
+			     ret == -ERESTARTNOHAND || ret == -ERESTART_RESTARTBLOCK))
+			ret = -EINTR;
+		aio_complete(req, ret, 0);
 	}
-	spin_unlock_irq(&ctx->ctx_lock);
 
 	aio_put_req(req);	/* drop extra ref to req */
 	return 0;

index 4f7795f..bad2583 100644
@@ -2322,7 +2322,7 @@ int ocfs2_inode_lock_full_nested(struct inode *inode,
 	status = __ocfs2_cluster_lock(osb, lockres, level, dlm_flags,
 				      arg_flags, subclass, _RET_IP_);
 	if (status < 0) {
-		if (status != -EAGAIN && status != -EIOCBRETRY)
+		if (status != -EAGAIN)
 			mlog_errno(status);
 		goto bail;
 	}

index d065348..7347732 100644
@@ -318,16 +318,6 @@ int rw_verify_area(int read_write, struct file *file, loff_t *ppos, size_t count
 	return count > MAX_RW_COUNT ? MAX_RW_COUNT : count;
 }
 
-static void wait_on_retry_sync_kiocb(struct kiocb *iocb)
-{
-	set_current_state(TASK_UNINTERRUPTIBLE);
-	if (!kiocbIsKicked(iocb))
-		schedule();
-	else
-		kiocbClearKicked(iocb);
-	__set_current_state(TASK_RUNNING);
-}
-
 ssize_t do_sync_read(struct file *filp, char __user *buf, size_t len, loff_t *ppos)
 {
 	struct iovec iov = { .iov_base = buf, .iov_len = len };

@@ -339,13 +329,7 @@ ssize_t do_sync_read(struct file *filp, char __user *buf, size_t len, loff_t *pp
 	kiocb.ki_left = len;
 	kiocb.ki_nbytes = len;
 
-	for (;;) {
-		ret = filp->f_op->aio_read(&kiocb, &iov, 1, kiocb.ki_pos);
-		if (ret != -EIOCBRETRY)
-			break;
-		wait_on_retry_sync_kiocb(&kiocb);
-	}
-
+	ret = filp->f_op->aio_read(&kiocb, &iov, 1, kiocb.ki_pos);
 	if (-EIOCBQUEUED == ret)
 		ret = wait_on_sync_kiocb(&kiocb);
 	*ppos = kiocb.ki_pos;

@@ -395,13 +379,7 @@ ssize_t do_sync_write(struct file *filp, const char __user *buf, size_t len, lof
 	kiocb.ki_left = len;
 	kiocb.ki_nbytes = len;
 
-	for (;;) {
-		ret = filp->f_op->aio_write(&kiocb, &iov, 1, kiocb.ki_pos);
-		if (ret != -EIOCBRETRY)
-			break;
-		wait_on_retry_sync_kiocb(&kiocb);
-	}
-
+	ret = filp->f_op->aio_write(&kiocb, &iov, 1, kiocb.ki_pos);
 	if (-EIOCBQUEUED == ret)
 		ret = wait_on_sync_kiocb(&kiocb);
 	*ppos = kiocb.ki_pos;

@@ -568,13 +546,7 @@ ssize_t do_sync_readv_writev(struct file *filp, const struct iovec *iov,
 	kiocb.ki_left = len;
 	kiocb.ki_nbytes = len;
 
-	for (;;) {
-		ret = fn(&kiocb, iov, nr_segs, kiocb.ki_pos);
-		if (ret != -EIOCBRETRY)
-			break;
-		wait_on_retry_sync_kiocb(&kiocb);
-	}
-
+	ret = fn(&kiocb, iov, nr_segs, kiocb.ki_pos);
 	if (ret == -EIOCBQUEUED)
 		ret = wait_on_sync_kiocb(&kiocb);
 	*ppos = kiocb.ki_pos;

index b46a09f..019204e 100644
@@ -14,18 +14,12 @@ struct kioctx;
 #define KIOCB_SYNC_KEY		(~0U)
 
 /* ki_flags bits */
-#define KIF_KICKED		1
 #define KIF_CANCELLED		2
 
-#define kiocbTryKick(iocb)	test_and_set_bit(KIF_KICKED, &(iocb)->ki_flags)
-
-#define kiocbSetKicked(iocb)	set_bit(KIF_KICKED, &(iocb)->ki_flags)
 #define kiocbSetCancelled(iocb)	set_bit(KIF_CANCELLED, &(iocb)->ki_flags)
 
-#define kiocbClearKicked(iocb)	clear_bit(KIF_KICKED, &(iocb)->ki_flags)
 #define kiocbClearCancelled(iocb)	clear_bit(KIF_CANCELLED, &(iocb)->ki_flags)
 
-#define kiocbIsKicked(iocb)	test_bit(KIF_KICKED, &(iocb)->ki_flags)
 #define kiocbIsCancelled(iocb)	test_bit(KIF_CANCELLED, &(iocb)->ki_flags)
 
 /* is there a better place to document function pointer methods? */

@@ -52,18 +46,8 @@ struct kioctx;
  * not ask the method again -- ki_retry must ensure forward progress.
  * aio_complete() must be called once and only once in the future, multiple
  * calls may result in undefined behaviour.
- *
- * If ki_retry returns -EIOCBRETRY it has made a promise that kick_iocb()
- * will be called on the kiocb pointer in the future.  This may happen
- * through generic helpers that associate kiocb->ki_wait with a wait
- * queue head that ki_retry uses via current->io_wait.  It can also happen
- * with custom tracking and manual calls to kick_iocb(), though that is
- * discouraged.  In either case, kick_iocb() must be called once and only
- * once.  ki_retry must ensure forward progress, the AIO core will wait
- * indefinitely for kick_iocb() to be called.
  */
 struct kiocb {
-	struct list_head	ki_run_list;
 	unsigned long		ki_flags;
 	int			ki_users;
 	unsigned		ki_key;		/* id of this request */

@@ -160,7 +144,6 @@ static inline unsigned aio_ring_avail(struct aio_ring_info *info,
 struct kioctx {
 	atomic_t		users;
 	int			dead;
-	struct mm_struct	*mm;
 
 	/* This needs improving */
 	unsigned long		user_id;

@@ -172,15 +155,12 @@ struct kioctx {
 
 	int			reqs_active;
 	struct list_head	active_reqs;	/* used for cancellation */
-	struct list_head	run_list;	/* used for kicked reqs */
 
 	/* sys_io_setup currently limits this to an unsigned int */
 	unsigned		max_reqs;
 
 	struct aio_ring_info	ring_info;
 
-	struct delayed_work	wq;
-
 	struct rcu_head		rcu_head;
 };
 

@@ -188,7 +168,6 @@ struct kioctx {
 #ifdef CONFIG_AIO
 extern ssize_t wait_on_sync_kiocb(struct kiocb *iocb);
 extern int aio_put_req(struct kiocb *iocb);
-extern void kick_iocb(struct kiocb *iocb);
 extern int aio_complete(struct kiocb *iocb, long res, long res2);
 struct mm_struct;
 extern void exit_aio(struct mm_struct *mm);

@@ -197,7 +176,6 @@ extern long do_io_submit(aio_context_t ctx_id, long nr,
 #else
 static inline ssize_t wait_on_sync_kiocb(struct kiocb *iocb) { return 0; }
 static inline int aio_put_req(struct kiocb *iocb) { return 0; }
-static inline void kick_iocb(struct kiocb *iocb) { }
 static inline int aio_complete(struct kiocb *iocb, long res, long res2) { return 0; }
 struct mm_struct;
 static inline void exit_aio(struct mm_struct *mm) { }

index f6bf082..89627b9 100644
@@ -28,6 +28,5 @@
 #define EBADTYPE	527	/* Type not supported by server */
 #define EJUKEBOX	528	/* Request initiated, but will not complete before timeout */
 #define EIOCBQUEUED	529	/* iocb queued, will get completion event */
-#define EIOCBRETRY	530	/* iocb queued, will trigger a retry */
 
 #endif

index 0537903..968ae6e 100644
@@ -627,6 +627,18 @@ static ssize_t write_null(struct file *file, const char __user *buf,
 	return count;
 }
 
+static ssize_t aio_read_null(struct kiocb *iocb, const struct iovec *iov,
+			     unsigned long nr_segs, loff_t pos)
+{
+	return 0;
+}
+
+static ssize_t aio_write_null(struct kiocb *iocb, const struct iovec *iov,
+			      unsigned long nr_segs, loff_t pos)
+{
+	return iov_length(iov, nr_segs);
+}
+
 static int pipe_to_null(struct pipe_inode_info *info, struct pipe_buffer *buf,
 			struct splice_desc *sd)
 {

@@ -670,6 +682,24 @@ static ssize_t read_zero(struct file *file, char __user *buf,
 	return written ? written : -EFAULT;
 }
 
+static ssize_t aio_read_zero(struct kiocb *iocb, const struct iovec *iov,
+			     unsigned long nr_segs, loff_t pos)
+{
+	size_t written = 0;
+	unsigned long i;
+	ssize_t ret;
+
+	for (i = 0; i < nr_segs; i++) {
+		ret = read_zero(iocb->ki_filp, iov[i].iov_base, iov[i].iov_len,
+				&pos);
+		if (ret < 0)
+			break;
+		written += ret;
+	}
+
+	return written ? written : -EFAULT;
+}
+
 static int mmap_zero(struct file *file, struct vm_area_struct *vma)
 {
 #ifndef CONFIG_MMU

@@ -738,6 +768,7 @@ static int open_port(struct inode * inode, struct file * filp)
 #define full_lseek      null_lseek
 #define write_zero	write_null
 #define read_full       read_zero
+#define aio_write_zero	aio_write_null
 #define open_mem	open_port
 #define open_kmem	open_mem
 #define open_oldmem	open_mem

@@ -766,6 +797,8 @@ static const struct file_operations null_fops = {
 	.llseek		= null_lseek,
 	.read		= read_null,
 	.write		= write_null,
+	.aio_read	= aio_read_null,
+	.aio_write	= aio_write_null,
 	.splice_write	= splice_write_null,
 };
 

@@ -782,6 +815,8 @@ static const struct file_operations zero_fops = {
 	.llseek		= zero_lseek,
 	.read		= read_zero,
 	.write		= write_zero,
+	.aio_read	= aio_read_zero,
+	.aio_write	= aio_write_zero,
 	.mmap		= mmap_zero,
 };
 

index 1de4f78..0b85822 100644
@@ -528,7 +528,7 @@ static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
 /* __aio_put_req
  *	Returns true if this put was the last user of the request.
  */
-static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
+static void __aio_put_req(struct kioctx *ctx, struct kiocb *req)
 {
 	dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
 		req, atomic_long_read(&req->ki_filp->f_count));

@@ -538,7 +538,7 @@ static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
 	req->ki_users--;
 	BUG_ON(req->ki_users < 0);
 	if (likely(req->ki_users))
-		return 0;
+		return;
 	list_del(&req->ki_list);		/* remove from active_reqs */
 	req->ki_cancel = NULL;
 	req->ki_retry = NULL;

@@ -546,21 +546,18 @@ static int __aio_put_req(struct kioctx *ctx, struct kiocb *req)
 	fput(req->ki_filp);
 	req->ki_filp = NULL;
 	really_put_req(ctx, req);
-	return 1;
 }
 
 /* aio_put_req
  *	Returns true if this put was the last user of the kiocb,
  *	false if the request is still in use.
  */
-int aio_put_req(struct kiocb *req)
+void aio_put_req(struct kiocb *req)
 {
 	struct kioctx *ctx = req->ki_ctx;
-	int ret;
 	spin_lock_irq(&ctx->ctx_lock);
-	ret = __aio_put_req(ctx, req);
+	__aio_put_req(ctx, req);
 	spin_unlock_irq(&ctx->ctx_lock);
-	return ret;
 }
 EXPORT_SYMBOL(aio_put_req);
 

@@ -591,10 +588,8 @@ static struct kioctx *lookup_ioctx(unsigned long ctx_id)
 
 /* aio_complete
  *	Called when the io request on the given iocb is complete.
- *	Returns true if this is the last user of the request.  The 
- *	only other user of the request can be the cancellation code.
  */
-int aio_complete(struct kiocb *iocb, long res, long res2)
+void aio_complete(struct kiocb *iocb, long res, long res2)
 {
 	struct kioctx	*ctx = iocb->ki_ctx;
 	struct aio_ring_info	*info;

@@ -602,7 +597,6 @@ int aio_complete(struct kiocb *iocb, long res, long res2)
 	struct io_event	*event;
 	unsigned long	flags;
 	unsigned long	tail;
-	int		ret;
 
 	/*
 	 * Special case handling for sync iocbs:

@@ -616,7 +610,7 @@ int aio_complete(struct kiocb *iocb, long res, long res2)
 		iocb->ki_user_data = res;
 		iocb->ki_users = 0;
 		wake_up_process(iocb->ki_obj.tsk);
-		return 1;
+		return;
 	}
 
 	info = &ctx->ring_info;

@@ -675,7 +669,7 @@ int aio_complete(struct kiocb *iocb, long res, long res2)
 
 put_rq:
 	/* everything turned out well, dispose of the aiocb. */
-	ret = __aio_put_req(ctx, iocb);
+	__aio_put_req(ctx, iocb);
 
 	/*
 	 * We have to order our ring_info tail store above and test

@@ -689,7 +683,6 @@ put_rq:
 		wake_up(&ctx->wait);
 
 	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
-	return ret;
 }
 EXPORT_SYMBOL(aio_complete);
 

index 019204e..615d55a 100644
@@ -167,16 +167,16 @@ struct kioctx {
 /* prototypes */
 #ifdef CONFIG_AIO
 extern ssize_t wait_on_sync_kiocb(struct kiocb *iocb);
-extern int aio_put_req(struct kiocb *iocb);
-extern int aio_complete(struct kiocb *iocb, long res, long res2);
+extern void aio_put_req(struct kiocb *iocb);
+extern void aio_complete(struct kiocb *iocb, long res, long res2);
 struct mm_struct;
 extern void exit_aio(struct mm_struct *mm);
 extern long do_io_submit(aio_context_t ctx_id, long nr,
 			 struct iocb __user *__user *iocbpp, bool compat);
 #else
 static inline ssize_t wait_on_sync_kiocb(struct kiocb *iocb) { return 0; }
-static inline int aio_put_req(struct kiocb *iocb) { return 0; }
-static inline int aio_complete(struct kiocb *iocb, long res, long res2) { return 0; }
+static inline void aio_put_req(struct kiocb *iocb) { }
+static inline void aio_complete(struct kiocb *iocb, long res, long res2) { }
 struct mm_struct;
 static inline void exit_aio(struct mm_struct *mm) { }
 static inline long do_io_submit(aio_context_t ctx_id, long nr,

index 0b85822..e1d4084 100644
@@ -217,6 +217,29 @@ static inline void put_ioctx(struct kioctx *kioctx)
 		__put_ioctx(kioctx);
 }
 
+static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb,
+			struct io_event *res)
+{
+	int (*cancel)(struct kiocb *, struct io_event *);
+	int ret = -EINVAL;
+
+	cancel = kiocb->ki_cancel;
+	kiocbSetCancelled(kiocb);
+	if (cancel) {
+		kiocb->ki_users++;
+		spin_unlock_irq(&ctx->ctx_lock);
+
+		memset(res, 0, sizeof(*res));
+		res->obj = (u64) kiocb->ki_obj.user;
+		res->data = kiocb->ki_user_data;
+		ret = cancel(kiocb, res);
+
+		spin_lock_irq(&ctx->ctx_lock);
+	}
+
+	return ret;
+}
+
 /* ioctx_alloc
  *	Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
  */

@@ -287,25 +310,19 @@ out_freectx:
  */
 static void kill_ctx(struct kioctx *ctx)
 {
-	int (*cancel)(struct kiocb *, struct io_event *);
 	struct task_struct *tsk = current;
 	DECLARE_WAITQUEUE(wait, tsk);
 	struct io_event res;
+	struct kiocb *req;
 
 	spin_lock_irq(&ctx->ctx_lock);
 	ctx->dead = 1;
 	while (!list_empty(&ctx->active_reqs)) {
-		struct list_head *pos = ctx->active_reqs.next;
-		struct kiocb *iocb = list_kiocb(pos);
-		list_del_init(&iocb->ki_list);
-		cancel = iocb->ki_cancel;
-		kiocbSetCancelled(iocb);
-		if (cancel) {
-			iocb->ki_users++;
-			spin_unlock_irq(&ctx->ctx_lock);
-			cancel(iocb, &res);
-			spin_lock_irq(&ctx->ctx_lock);
-		}
+		req = list_first_entry(&ctx->active_reqs,
+					struct kiocb, ki_list);
+
+		list_del_init(&req->ki_list);
+		kiocb_cancel(ctx, req, &res);
 	}
 
 	if (!ctx->reqs_active)

@@ -1409,7 +1426,7 @@ static struct kiocb *lookup_kiocb(struct kioctx *ctx, struct iocb __user *iocb,
 SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
 		struct io_event __user *, result)
 {
-	int (*cancel)(struct kiocb *iocb, struct io_event *res);
+	struct io_event res;
 	struct kioctx *ctx;
 	struct kiocb *kiocb;
 	u32 key;

@@ -1424,32 +1441,22 @@ SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
 		return -EINVAL;
 
 	spin_lock_irq(&ctx->ctx_lock);
-	ret = -EAGAIN;
+
 	kiocb = lookup_kiocb(ctx, iocb, key);
-	if (kiocb && kiocb->ki_cancel) {
-		cancel = kiocb->ki_cancel;
-		kiocb->ki_users ++;
-		kiocbSetCancelled(kiocb);
-	} else
-		cancel = NULL;
+	if (kiocb)
+		ret = kiocb_cancel(ctx, kiocb, &res);
+	else
+		ret = -EINVAL;
+
 	spin_unlock_irq(&ctx->ctx_lock);
 
-	if (NULL != cancel) {
-		struct io_event tmp;
-		pr_debug("calling cancel\n");
-		memset(&tmp, 0, sizeof(tmp));
-		tmp.obj = (u64)(unsigned long)kiocb->ki_obj.user;
-		tmp.data = kiocb->ki_user_data;
-		ret = cancel(kiocb, &tmp);
-		if (!ret) {
-			/* Cancellation succeeded -- copy the result
-			 * into the user's buffer.
-			 */
-			if (copy_to_user(result, &tmp, sizeof(tmp)))
-				ret = -EFAULT;
-		}
-	} else
-		ret = -EINVAL;
+	if (!ret) {
+		/* Cancellation succeeded -- copy the result
+		 * into the user's buffer.
+		 */
+		if (copy_to_user(result, &res, sizeof(res)))
+			ret = -EFAULT;
+	}
 
 	put_ioctx(ctx);
 

index 2a3f001..7640e01 100644
@@ -25,6 +25,7 @@
 #include <linux/slab.h>
 #include <linux/poll.h>
 #include <linux/mmu_context.h>
+#include <linux/aio.h>
 
 #include <linux/device.h>
 #include <linux/moduleparam.h>

67 lines not on this page · raw message

67 lines not on this page · raw message

index 8fcea98..868ac0a 100644
@@ -8,6 +8,8 @@
  *
  *	See ../COPYING for licensing terms.
  */
+#define pr_fmt(fmt) "%s: " fmt, __func__
+
 #include <linux/kernel.h>
 #include <linux/init.h>
 #include <linux/errno.h>

@@ -18,8 +20,6 @@
 #include <linux/backing-dev.h>
 #include <linux/uio.h>
 
-#define DEBUG 0
-
 #include <linux/sched.h>
 #include <linux/fs.h>
 #include <linux/file.h>

@@ -39,12 +39,6 @@
 #include <asm/kmap_types.h>
 #include <asm/uaccess.h>
 
-#if DEBUG > 1
-#define dprintk		printk
-#else
-#define dprintk(x...)	do { ; } while (0)
-#endif
-
 #define AIO_RING_MAGIC			0xa10a10a1
 #define AIO_RING_COMPAT_FEATURES	1
 #define AIO_RING_INCOMPAT_FEATURES	0

@@ -124,7 +118,7 @@ static int __init aio_setup(void)
 	kiocb_cachep = KMEM_CACHE(kiocb, SLAB_HWCACHE_ALIGN|SLAB_PANIC);
 	kioctx_cachep = KMEM_CACHE(kioctx,SLAB_HWCACHE_ALIGN|SLAB_PANIC);
 
-	pr_debug("aio_setup: sizeof(struct page) = %d\n", (int)sizeof(struct page));
+	pr_debug("sizeof(struct page) = %zu\n", sizeof(struct page));
 
 	return 0;
 }

@@ -178,7 +172,7 @@ static int aio_setup_ring(struct kioctx *ctx)
 	}
 
 	info->mmap_size = nr_pages * PAGE_SIZE;
-	dprintk("attempting mmap of %lu bytes\n", info->mmap_size);
+	pr_debug("attempting mmap of %lu bytes\n", info->mmap_size);
 	down_write(&mm->mmap_sem);
 	info->mmap_base = do_mmap_pgoff(NULL, 0, info->mmap_size, 
 					PROT_READ|PROT_WRITE,

@@ -190,7 +184,7 @@ static int aio_setup_ring(struct kioctx *ctx)
 		return -EAGAIN;
 	}
 
-	dprintk("mmap address: 0x%08lx\n", info->mmap_base);
+	pr_debug("mmap address: 0x%08lx\n", info->mmap_base);
 	info->nr_pages = get_user_pages(current, mm, info->mmap_base, nr_pages, 
 					1, 0, info->ring_pages, NULL);
 	up_write(&mm->mmap_sem);

@@ -262,7 +256,7 @@ static void __put_ioctx(struct kioctx *ctx)
 		aio_nr -= nr_events;
 		spin_unlock(&aio_nr_lock);
 	}
-	pr_debug("__put_ioctx: freeing %p\n", ctx);
+	pr_debug("freeing %p\n", ctx);
 	call_rcu(&ctx->rcu_head, ctx_rcu_free);
 }
 

@@ -351,7 +345,7 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
 	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",
+	pr_debug("allocated ioctx %p[%ld]: mm=%p mask=0x%x\n",
 		ctx, ctx->user_id, mm, ctx->ring_info.nr);
 	return ctx;
 

@@ -360,7 +354,7 @@ out_cleanup:
 	aio_free_ring(ctx);
 out_freectx:
 	kmem_cache_free(kioctx_cachep, ctx);
-	dprintk("aio: error allocating ioctx %d\n", err);
+	pr_debug("error allocating ioctx %d\n", err);
 	return ERR_PTR(err);
 }
 

@@ -608,8 +602,8 @@ static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
  */
 static void __aio_put_req(struct kioctx *ctx, struct kiocb *req)
 {
-	dprintk(KERN_DEBUG "aio_put(%p): f_count=%ld\n",
-		req, atomic_long_read(&req->ki_filp->f_count));
+	pr_debug("(%p): f_count=%ld\n",
+		 req, atomic_long_read(&req->ki_filp->f_count));
 
 	assert_spin_locked(&ctx->ctx_lock);
 

@@ -720,9 +714,9 @@ void aio_complete(struct kiocb *iocb, long res, long res2)
 	event->res = res;
 	event->res2 = res2;
 
-	dprintk("aio_complete: %p[%lu]: %p: %p %Lx %lx %lx\n",
-		ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
-		res, res2);
+	pr_debug("%p[%lu]: %p: %p %Lx %lx %lx\n",
+		 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
+		 res, res2);
 
 	/* after flagging the request as done, we
 	 * must never even look at it again

@@ -778,9 +772,7 @@ static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
 	int ret = 0;
 
 	ring = kmap_atomic(info->ring_pages[0]);
-	dprintk("in aio_read_evt h%lu t%lu m%lu\n",
-		 (unsigned long)ring->head, (unsigned long)ring->tail,
-		 (unsigned long)ring->nr);
+	pr_debug("h%u t%u m%u\n", ring->head, ring->tail, ring->nr);
 
 	if (ring->head == ring->tail)
 		goto out;

@@ -801,8 +793,7 @@ static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
 
 out:
 	kunmap_atomic(ring);
-	dprintk("leaving aio_read_evt: %d  h%lu t%lu\n", ret,
-		 (unsigned long)ring->head, (unsigned long)ring->tail);
+	pr_debug("%d  h%u t%u\n", ret, ring->head, ring->tail);
 	return ret;
 }
 

@@ -865,13 +856,13 @@ static int read_events(struct kioctx *ctx,
 		if (unlikely(ret <= 0))
 			break;
 
-		dprintk("read event: %Lx %Lx %Lx %Lx\n",
-			ent.data, ent.obj, ent.res, ent.res2);
+		pr_debug("%Lx %Lx %Lx %Lx\n",
+			 ent.data, ent.obj, ent.res, ent.res2);
 
 		/* Could we split the check in two? */
 		ret = -EFAULT;
 		if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
-			dprintk("aio: lost an event due to EFAULT.\n");
+			pr_debug("lost an event due to EFAULT.\n");
 			break;
 		}
 		ret = 0;

@@ -934,7 +925,7 @@ static int read_events(struct kioctx *ctx,
 
 		ret = -EFAULT;
 		if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
-			dprintk("aio: lost an event due to EFAULT.\n");
+			pr_debug("lost an event due to EFAULT.\n");
 			break;
 		}
 

@@ -965,7 +956,7 @@ static void io_destroy(struct kioctx *ioctx)
 	hlist_del_rcu(&ioctx->list);
 	spin_unlock(&mm->ioctx_lock);
 
-	dprintk("aio_release(%p)\n", ioctx);
+	pr_debug("(%p)\n", ioctx);
 	if (likely(!was_dead))
 		put_ioctx(ioctx);	/* twice for the list */
 

@@ -1258,7 +1249,7 @@ static ssize_t aio_setup_iocb(struct kiocb *kiocb, bool compat)
 			kiocb->ki_retry = aio_fsync;
 		break;
 	default:
-		dprintk("EINVAL: io_submit: no operation provided\n");
+		pr_debug("EINVAL: no operation provided\n");
 		ret = -EINVAL;
 	}
 

@@ -1278,7 +1269,7 @@ static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
 
 	/* enforce forwards compatibility on users */
 	if (unlikely(iocb->aio_reserved1 || iocb->aio_reserved2)) {
-		pr_debug("EINVAL: io_submit: reserve field set\n");
+		pr_debug("EINVAL: reserve field set\n");
 		return -EINVAL;
 	}
 

@@ -1319,7 +1310,7 @@ static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
 
 	ret = put_user(req->ki_key, &user_iocb->aio_key);
 	if (unlikely(ret)) {
-		dprintk("EFAULT: aio_key\n");
+		pr_debug("EFAULT: aio_key\n");
 		goto out_put_req;
 	}
 

@@ -1400,7 +1391,7 @@ long do_io_submit(aio_context_t ctx_id, long nr,
 
 	ctx = lookup_ioctx(ctx_id);
 	if (unlikely(!ctx)) {
-		pr_debug("EINVAL: io_submit: invalid context id\n");
+		pr_debug("EINVAL: invalid context id\n");
 		return -EINVAL;
 	}
 

index 868ac0a..db6cb02 100644
@@ -584,6 +584,8 @@ static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
 {
 	assert_spin_locked(&ctx->ctx_lock);
 
+	if (req->ki_filp)
+		fput(req->ki_filp);
 	if (req->ki_eventfd != NULL)
 		eventfd_ctx_put(req->ki_eventfd);
 	if (req->ki_dtor)

@@ -602,9 +604,6 @@ static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
  */
 static void __aio_put_req(struct kioctx *ctx, struct kiocb *req)
 {
-	pr_debug("(%p): f_count=%ld\n",
-		 req, atomic_long_read(&req->ki_filp->f_count));
-
 	assert_spin_locked(&ctx->ctx_lock);
 
 	req->ki_users--;

@@ -615,8 +614,6 @@ static void __aio_put_req(struct kioctx *ctx, struct kiocb *req)
 	req->ki_cancel = NULL;
 	req->ki_retry = NULL;
 
-	fput(req->ki_filp);
-	req->ki_filp = NULL;
 	really_put_req(ctx, req);
 }
 

@@ -1264,7 +1261,6 @@ static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
 			 bool compat)
 {
 	struct kiocb *req;
-	struct file *file;
 	ssize_t ret;
 
 	/* enforce forwards compatibility on users */

@@ -1283,16 +1279,16 @@ static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
 		return -EINVAL;
 	}
 
-	file = fget(iocb->aio_fildes);
-	if (unlikely(!file))
-		return -EBADF;
-
 	req = aio_get_req(ctx, batch);  /* returns with 2 references to req */
-	if (unlikely(!req)) {
-		fput(file);
+	if (unlikely(!req))
 		return -EAGAIN;
+
+	req->ki_filp = fget(iocb->aio_fildes);
+	if (unlikely(!req->ki_filp)) {
+		ret = -EBADF;
+		goto out_put_req;
 	}
-	req->ki_filp = file;
+
 	if (iocb->aio_flags & IOCB_FLAG_RESFD) {
 		/*
 		 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an

index db6cb02..37eac67 100644
@@ -89,7 +89,7 @@ struct kioctx {
 
 	spinlock_t		ctx_lock;
 
-	int			reqs_active;
+	atomic_t		reqs_active;
 	struct list_head	active_reqs;	/* used for cancellation */
 
 	/* sys_io_setup currently limits this to an unsigned int */

@@ -247,7 +247,7 @@ static void ctx_rcu_free(struct rcu_head *head)
 static void __put_ioctx(struct kioctx *ctx)
 {
 	unsigned nr_events = ctx->max_reqs;
-	BUG_ON(ctx->reqs_active);
+	BUG_ON(atomic_read(&ctx->reqs_active));
 
 	aio_free_ring(ctx);
 	if (nr_events) {

@@ -281,7 +281,7 @@ static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb,
 	cancel = kiocb->ki_cancel;
 	kiocbSetCancelled(kiocb);
 	if (cancel) {
-		kiocb->ki_users++;
+		atomic_inc(&kiocb->ki_users);
 		spin_unlock_irq(&ctx->ctx_lock);
 
 		memset(res, 0, sizeof(*res));

@@ -380,12 +380,12 @@ static void kill_ctx(struct kioctx *ctx)
 		kiocb_cancel(ctx, req, &res);
 	}
 
-	if (!ctx->reqs_active)
+	if (!atomic_read(&ctx->reqs_active))
 		goto out;
 
 	add_wait_queue(&ctx->wait, &wait);
 	set_task_state(tsk, TASK_UNINTERRUPTIBLE);
-	while (ctx->reqs_active) {
+	while (atomic_read(&ctx->reqs_active)) {
 		spin_unlock_irq(&ctx->ctx_lock);
 		io_schedule();
 		set_task_state(tsk, TASK_UNINTERRUPTIBLE);

@@ -403,9 +403,9 @@ out:
  */
 ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
 {
-	while (iocb->ki_users) {
+	while (atomic_read(&iocb->ki_users)) {
 		set_current_state(TASK_UNINTERRUPTIBLE);
-		if (!iocb->ki_users)
+		if (!atomic_read(&iocb->ki_users))
 			break;
 		io_schedule();
 	}

@@ -435,7 +435,7 @@ void exit_aio(struct mm_struct *mm)
 			printk(KERN_DEBUG
 				"exit_aio:ioctx still alive: %d %d %d\n",
 				atomic_read(&ctx->users), ctx->dead,
-				ctx->reqs_active);
+				atomic_read(&ctx->reqs_active));
 		/*
 		 * We don't need to bother with munmap() here -
 		 * exit_mmap(mm) is coming and it'll unmap everything.

@@ -450,11 +450,11 @@ void exit_aio(struct mm_struct *mm)
 }
 
 /* aio_get_req
- *	Allocate a slot for an aio request.  Increments the users count
+ *	Allocate a slot for an aio request.  Increments the ki_users count
  * of the kioctx so that the kioctx stays around until all requests are
  * complete.  Returns NULL if no requests are free.
  *
- * Returns with kiocb->users set to 2.  The io submit code path holds
+ * Returns with kiocb->ki_users set to 2.  The io submit code path holds
  * an extra reference while submitting the i/o.
  * This prevents races between the aio code path referencing the
  * req (after submitting it) and aio_complete() freeing the req.

@@ -468,7 +468,7 @@ static struct kiocb *__aio_get_req(struct kioctx *ctx)
 		return NULL;
 
 	req->ki_flags = 0;
-	req->ki_users = 2;
+	atomic_set(&req->ki_users, 2);
 	req->ki_key = 0;
 	req->ki_ctx = ctx;
 	req->ki_cancel = NULL;

@@ -509,9 +509,9 @@ static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch)
 		list_del(&req->ki_batch);
 		list_del(&req->ki_list);
 		kmem_cache_free(kiocb_cachep, req);
-		ctx->reqs_active--;
+		atomic_dec(&ctx->reqs_active);
 	}
-	if (unlikely(!ctx->reqs_active && ctx->dead))
+	if (unlikely(!atomic_read(&ctx->reqs_active) && ctx->dead))
 		wake_up_all(&ctx->wait);
 	spin_unlock_irq(&ctx->ctx_lock);
 }

@@ -542,7 +542,7 @@ static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
 	spin_lock_irq(&ctx->ctx_lock);
 	ring = kmap_atomic(ctx->ring_info.ring_pages[0]);
 
-	avail = aio_ring_avail(&ctx->ring_info, ring) - ctx->reqs_active;
+	avail = aio_ring_avail(&ctx->ring_info, ring) - atomic_read(&ctx->reqs_active);
 	BUG_ON(avail < 0);
 	if (avail < allocated) {
 		/* Trim back the number of requests. */

@@ -557,7 +557,7 @@ static int kiocb_batch_refill(struct kioctx *ctx, struct kiocb_batch *batch)
 	batch->count -= allocated;
 	list_for_each_entry(req, &batch->head, ki_batch) {
 		list_add(&req->ki_list, &ctx->active_reqs);
-		ctx->reqs_active++;
+		atomic_inc(&ctx->reqs_active);
 	}
 
 	kunmap_atomic(ring);

@@ -580,10 +580,8 @@ static inline struct kiocb *aio_get_req(struct kioctx *ctx,
 	return req;
 }
 
-static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
+static void kiocb_free(struct kiocb *req)
 {
-	assert_spin_locked(&ctx->ctx_lock);
-
 	if (req->ki_filp)
 		fput(req->ki_filp);
 	if (req->ki_eventfd != NULL)

@@ -593,40 +591,12 @@ static inline void really_put_req(struct kioctx *ctx, struct kiocb *req)
 	if (req->ki_iovec != &req->ki_inline_vec)
 		kfree(req->ki_iovec);
 	kmem_cache_free(kiocb_cachep, req);
-	ctx->reqs_active--;
-
-	if (unlikely(!ctx->reqs_active && ctx->dead))
-		wake_up_all(&ctx->wait);
 }
 
-/* __aio_put_req
- *	Returns true if this put was the last user of the request.
- */
-static void __aio_put_req(struct kioctx *ctx, struct kiocb *req)
-{
-	assert_spin_locked(&ctx->ctx_lock);
-
-	req->ki_users--;
-	BUG_ON(req->ki_users < 0);
-	if (likely(req->ki_users))
-		return;
-	list_del(&req->ki_list);		/* remove from active_reqs */
-	req->ki_cancel = NULL;
-	req->ki_retry = NULL;
-
-	really_put_req(ctx, req);
-}
-
-/* aio_put_req
- *	Returns true if this put was the last user of the kiocb,
- *	false if the request is still in use.
- */
 void aio_put_req(struct kiocb *req)
 {
-	struct kioctx *ctx = req->ki_ctx;
-	spin_lock_irq(&ctx->ctx_lock);
-	__aio_put_req(ctx, req);
-	spin_unlock_irq(&ctx->ctx_lock);
+	if (atomic_dec_and_test(&req->ki_users))
+		kiocb_free(req);
 }
 EXPORT_SYMBOL(aio_put_req);
 

@@ -675,9 +645,9 @@ void aio_complete(struct kiocb *iocb, long res, long res2)
 	 *  - the sync task helpfully left a reference to itself in the iocb
 	 */
 	if (is_sync_kiocb(iocb)) {
-		BUG_ON(iocb->ki_users != 1);
+		BUG_ON(atomic_read(&iocb->ki_users) != 1);
 		iocb->ki_user_data = res;
-		iocb->ki_users = 0;
+		atomic_set(&iocb->ki_users, 0);
 		wake_up_process(iocb->ki_obj.tsk);
 		return;
 	}

@@ -692,6 +662,8 @@ void aio_complete(struct kiocb *iocb, long res, long res2)
 	 */
 	spin_lock_irqsave(&ctx->ctx_lock, flags);
 
+	list_del(&iocb->ki_list); /* remove from active_reqs */
+
 	/*
 	 * cancelled requests don't get events, userland was given one
 	 * when the event got cancelled.

@@ -738,7 +710,8 @@ void aio_complete(struct kiocb *iocb, long res, long res2)
 
 put_rq:
 	/* everything turned out well, dispose of the aiocb. */
-	__aio_put_req(ctx, iocb);
+	aio_put_req(iocb);
+	atomic_dec(&ctx->reqs_active);
 
 	/*
 	 * We have to order our ring_info tail store above and test

@@ -903,7 +876,7 @@ static int read_events(struct kioctx *ctx,
 				break;
 			/* Try to only show up in io wait if there are ops
 			 *  in flight */
-			if (ctx->reqs_active)
+			if (atomic_read(&ctx->reqs_active))
 				io_schedule();
 			else
 				schedule();

@@ -1362,6 +1335,14 @@ static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
 	return 0;
 
 out_put_req:
+	spin_lock_irq(&ctx->ctx_lock);
+	list_del(&req->ki_list);
+	spin_unlock_irq(&ctx->ctx_lock);
+
+	atomic_dec(&ctx->reqs_active);
+	if (unlikely(!atomic_read(&ctx->reqs_active) && ctx->dead))
+		wake_up_all(&ctx->wait);
+
 	aio_put_req(req);	/* drop extra ref to req */
 	aio_put_req(req);	/* drop i/o ref to req */
 	return ret;

index 7b1eb23..1e728f0 100644
@@ -49,7 +49,7 @@ struct kioctx;
  */
 struct kiocb {
 	unsigned long		ki_flags;
-	int			ki_users;
+	atomic_t		ki_users;
 	unsigned		ki_key;		/* id of this request */
 
 	struct file		*ki_filp;

@@ -96,7 +96,7 @@ static inline bool is_sync_kiocb(struct kiocb *kiocb)
 static inline void init_sync_kiocb(struct kiocb *kiocb, struct file *filp)
 {
 	*kiocb = (struct kiocb) {
-			.ki_users = 1,
+			.ki_users = ATOMIC_INIT(1),
 			.ki_key = KIOCB_SYNC_KEY,
 			.ki_filp = filp,
 			.ki_obj.tsk = current,

index 37eac67..e0eb23d 100644
@@ -79,7 +79,7 @@ static inline unsigned aio_ring_avail(struct aio_ring_info *info,
 
 struct kioctx {
 	atomic_t		users;
-	int			dead;
+	atomic_t		dead;
 
 	/* This needs improving */
 	unsigned long		user_id;

@@ -98,6 +98,7 @@ struct kioctx {
 	struct aio_ring_info	ring_info;
 
 	struct rcu_head		rcu_head;
+	struct work_struct	rcu_work;
 };
 
 /*------ sysctl variables----*/

@@ -234,44 +235,6 @@ static int aio_setup_ring(struct kioctx *ctx)
 	kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK)); \
 } while(0)
 
-static void ctx_rcu_free(struct rcu_head *head)
-{
-	struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
-	kmem_cache_free(kioctx_cachep, 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)
-{
-	unsigned nr_events = ctx->max_reqs;
-	BUG_ON(atomic_read(&ctx->reqs_active));
-
-	aio_free_ring(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);
-	}
-	pr_debug("freeing %p\n", ctx);
-	call_rcu(&ctx->rcu_head, ctx_rcu_free);
-}
-
-static inline int try_get_ioctx(struct kioctx *kioctx)
-{
-	return atomic_inc_not_zero(&kioctx->users);
-}
-
-static inline void put_ioctx(struct kioctx *kioctx)
-{
-	BUG_ON(atomic_read(&kioctx->users) <= 0);
-	if (unlikely(atomic_dec_and_test(&kioctx->users)))
-		__put_ioctx(kioctx);
-}
-
 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb,
 			struct io_event *res)
 {

@@ -295,6 +258,61 @@ static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb,
 	return ret;
 }
 
+static void free_ioctx_rcu(struct rcu_head *head)
+{
+	struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
+	kmem_cache_free(kioctx_cachep, ctx);
+}
+
+/*
+ * When this function runs, the kioctx has been removed from the "hash table"
+ * and ctx->users has dropped to 0, so we know no more kiocbs can be submitted -
+ * now it's safe to cancel any that need to be.
+ */
+static void free_ioctx(struct kioctx *ctx)
+{
+	struct io_event res;
+	struct kiocb *req;
+
+	spin_lock_irq(&ctx->ctx_lock);
+
+	while (!list_empty(&ctx->active_reqs)) {
+		req = list_first_entry(&ctx->active_reqs,
+				       struct kiocb, ki_list);
+
+		list_del_init(&req->ki_list);
+		kiocb_cancel(ctx, req, &res);
+	}
+
+	spin_unlock_irq(&ctx->ctx_lock);
+
+	wait_event(ctx->wait, !atomic_read(&ctx->reqs_active));
+
+	aio_free_ring(ctx);
+
+	spin_lock(&aio_nr_lock);
+	BUG_ON(aio_nr - ctx->max_reqs > aio_nr);
+	aio_nr -= ctx->max_reqs;
+	spin_unlock(&aio_nr_lock);
+
+	pr_debug("freeing %p\n", ctx);
+
+	/*
+	 * Here the call_rcu() is between the wait_event() for reqs_active to
+	 * hit 0, and freeing the ioctx.
+	 *
+	 * aio_complete() decrements reqs_active, but it has to touch the ioctx
+	 * after to issue a wakeup so we use rcu.
+	 */
+	call_rcu(&ctx->rcu_head, free_ioctx_rcu);
+}
+
+static void put_ioctx(struct kioctx *ctx)
+{
+	if (unlikely(atomic_dec_and_test(&ctx->users)))
+		free_ioctx(ctx);
+}
+
 /* ioctx_alloc
  *	Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
  */

@@ -321,6 +339,7 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
 	ctx->max_reqs = nr_events;
 
 	atomic_set(&ctx->users, 2);
+	atomic_set(&ctx->dead, 0);
 	spin_lock_init(&ctx->ctx_lock);
 	spin_lock_init(&ctx->ring_info.ring_lock);
 	init_waitqueue_head(&ctx->wait);

@@ -358,44 +377,43 @@ out_freectx:
 	return ERR_PTR(err);
 }
 
-/* kill_ctx
- *	Cancels all outstanding aio requests on an aio context.  Used 
- *	when the processes owning a context have all exited to encourage 
- *	the rapid destruction of the kioctx.
- */
-static void kill_ctx(struct kioctx *ctx)
+static void kill_ioctx_work(struct work_struct *work)
 {
-	struct task_struct *tsk = current;
-	DECLARE_WAITQUEUE(wait, tsk);
-	struct io_event res;
-	struct kiocb *req;
+	struct kioctx *ctx = container_of(work, struct kioctx, rcu_work);
 
-	spin_lock_irq(&ctx->ctx_lock);
-	ctx->dead = 1;
-	while (!list_empty(&ctx->active_reqs)) {
-		req = list_first_entry(&ctx->active_reqs,
-					struct kiocb, ki_list);
+	wake_up_all(&ctx->wait);
+	put_ioctx(ctx);
+}
 
-		list_del_init(&req->ki_list);
-		kiocb_cancel(ctx, req, &res);
-	}
+static void kill_ioctx_rcu(struct rcu_head *head)
+{
+	struct kioctx *ctx = container_of(head, struct kioctx, rcu_head);
 
-	if (!atomic_read(&ctx->reqs_active))
-		goto out;
+	INIT_WORK(&ctx->rcu_work, kill_ioctx_work);
+	schedule_work(&ctx->rcu_work);
+}
 
-	add_wait_queue(&ctx->wait, &wait);
-	set_task_state(tsk, TASK_UNINTERRUPTIBLE);
-	while (atomic_read(&ctx->reqs_active)) {
-		spin_unlock_irq(&ctx->ctx_lock);
-		io_schedule();
-		set_task_state(tsk, TASK_UNINTERRUPTIBLE);
-		spin_lock_irq(&ctx->ctx_lock);
-	}
-	__set_task_state(tsk, TASK_RUNNING);
-	remove_wait_queue(&ctx->wait, &wait);
+/* kill_ioctx
+ *	Cancels all outstanding aio requests on an aio context.  Used
+ *	when the processes owning a context have all exited to encourage
+ *	the rapid destruction of the kioctx.
+ */
+static void kill_ioctx(struct kioctx *ctx)
+{
+	if (!atomic_xchg(&ctx->dead, 1)) {
+		hlist_del_rcu(&ctx->list);
+		/* Between hlist_del_rcu() and dropping the initial ref */
+		synchronize_rcu();
 
-out:
-	spin_unlock_irq(&ctx->ctx_lock);
+		/*
+		 * We can't punt to workqueue here because put_ioctx() ->
+		 * free_ioctx() will unmap the ringbuffer, and that has to be
+		 * done in the original process's context. kill_ioctx_rcu/work()
+		 * exist for exit_aio(), as in that path free_ioctx() won't do
+		 * the unmap.
+		 */
+		kill_ioctx_work(&ctx->rcu_work);
+	}
 }
 
 /* wait_on_sync_kiocb:

@@ -414,27 +432,25 @@ ssize_t wait_on_sync_kiocb(struct kiocb *iocb)
 }
 EXPORT_SYMBOL(wait_on_sync_kiocb);
 
-/* exit_aio: called when the last user of mm goes away.  At this point, 
- * there is no way for any new requests to be submited or any of the 
- * io_* syscalls to be called on the context.  However, there may be 
- * outstanding requests which hold references to the context; as they 
- * go away, they will call put_ioctx and release any pinned memory
- * associated with the request (held via struct page * references).
+/*
+ * exit_aio: called when the last user of mm goes away.  At this point, there is
+ * no way for any new requests to be submited or any of the io_* syscalls to be
+ * called on the context.
+ *
+ * There may be outstanding kiocbs, but free_ioctx() will explicitly wait on
+ * them.
  */
 void exit_aio(struct mm_struct *mm)
 {
 	struct kioctx *ctx;
+	struct hlist_node *p, *n;
 
-	while (!hlist_empty(&mm->ioctx_list)) {
-		ctx = hlist_entry(mm->ioctx_list.first, struct kioctx, list);
-		hlist_del_rcu(&ctx->list);
-
-		kill_ctx(ctx);
-
+	hlist_for_each_entry_safe(ctx, p, n, &mm->ioctx_list, list) {
 		if (1 != atomic_read(&ctx->users))
 			printk(KERN_DEBUG
 				"exit_aio:ioctx still alive: %d %d %d\n",
-				atomic_read(&ctx->users), ctx->dead,
+				atomic_read(&ctx->users),
+				atomic_read(&ctx->dead),
 				atomic_read(&ctx->reqs_active));
 		/*
 		 * We don't need to bother with munmap() here -

@@ -445,7 +461,11 @@ void exit_aio(struct mm_struct *mm)
 		 * place that uses ->mmap_size, so it's safe.
 		 */
 		ctx->ring_info.mmap_size = 0;
-		put_ioctx(ctx);
+
+		if (!atomic_xchg(&ctx->dead, 1)) {
+			hlist_del_rcu(&ctx->list);
+			call_rcu(&ctx->rcu_head, kill_ioctx_rcu);
+		}
 	}
 }
 

@@ -511,8 +531,6 @@ static void kiocb_batch_free(struct kioctx *ctx, struct kiocb_batch *batch)
 		kmem_cache_free(kiocb_cachep, req);
 		atomic_dec(&ctx->reqs_active);
 	}
-	if (unlikely(!atomic_read(&ctx->reqs_active) && ctx->dead))
-		wake_up_all(&ctx->wait);
 	spin_unlock_irq(&ctx->ctx_lock);
 }
 

@@ -608,18 +626,12 @@ static struct kioctx *lookup_ioctx(unsigned long ctx_id)
 
 	rcu_read_lock();
 
-	hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list) {
-		/*
-		 * RCU protects us against accessing freed memory but
-		 * we have to be careful not to get a reference when the
-		 * reference count already dropped to 0 (ctx->dead test
-		 * is unreliable because of races).
-		 */
-		if (ctx->user_id == ctx_id && !ctx->dead && try_get_ioctx(ctx)){
+	hlist_for_each_entry_rcu(ctx, n, &mm->ioctx_list, list)
+		if (ctx->user_id == ctx_id){
+			atomic_inc(&ctx->users);
 			ret = ctx;
 			break;
 		}
-	}
 
 	rcu_read_unlock();
 	return ret;

@@ -654,12 +666,15 @@ void aio_complete(struct kiocb *iocb, long res, long res2)
 
 	info = &ctx->ring_info;
 
-	/* add a completion event to the ring buffer.
-	 * must be done holding ctx->ctx_lock to prevent
-	 * other code from messing with the tail
-	 * pointer since we might be called from irq
-	 * context.
+	/*
+	 * Add a completion event to the ring buffer. Must be done holding
+	 * ctx->ctx_lock to prevent other code from messing with the tail
+	 * pointer since we might be called from irq context.
+	 *
+	 * Take rcu_read_lock() in case the kioctx is being destroyed, as we
+	 * need to issue a wakeup after decrementing reqs_active.
 	 */
+	rcu_read_lock();
 	spin_lock_irqsave(&ctx->ctx_lock, flags);
 
 	list_del(&iocb->ki_list); /* remove from active_reqs */

@@ -725,6 +740,7 @@ put_rq:
 		wake_up(&ctx->wait);
 
 	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
+	rcu_read_unlock();
 }
 EXPORT_SYMBOL(aio_complete);
 

@@ -868,7 +884,7 @@ static int read_events(struct kioctx *ctx,
 				break;
 			if (min_nr <= i)
 				break;
-			if (unlikely(ctx->dead)) {
+			if (unlikely(atomic_read(&ctx->dead))) {
 				ret = -EINVAL;
 				break;
 			}

@@ -911,35 +927,6 @@ out:
 	return i ? i : ret;
 }
 
-/* Take an ioctx and remove it from the list of ioctx's.  Protects 
- * against races with itself via ->dead.
- */
-static void io_destroy(struct kioctx *ioctx)
-{
-	struct mm_struct *mm = current->mm;
-	int was_dead;
-
-	/* delete the entry from the list is someone else hasn't already */
-	spin_lock(&mm->ioctx_lock);
-	was_dead = ioctx->dead;
-	ioctx->dead = 1;
-	hlist_del_rcu(&ioctx->list);
-	spin_unlock(&mm->ioctx_lock);
-
-	pr_debug("(%p)\n", ioctx);
-	if (likely(!was_dead))
-		put_ioctx(ioctx);	/* twice for the list */
-
-	kill_ctx(ioctx);
-
-	/*
-	 * Wake up any waiters.  The setting of ctx->dead must be seen
-	 * by other CPUs at this point.  Right now, we rely on the
-	 * locking done by the above calls to ensure this consistency.
-	 */
-	wake_up_all(&ioctx->wait);
-}
-
 /* sys_io_setup:
  *	Create an aio_context capable of receiving at least nr_events.
  *	ctxp must not point to an aio_context that already exists, and

@@ -975,7 +962,7 @@ SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
 	if (!IS_ERR(ioctx)) {
 		ret = put_user(ioctx->user_id, ctxp);
 		if (ret)
-			io_destroy(ioctx);
+			kill_ioctx(ioctx);
 		put_ioctx(ioctx);
 	}
 

@@ -993,7 +980,7 @@ SYSCALL_DEFINE1(io_destroy, aio_context_t, ctx)
 {
 	struct kioctx *ioctx = lookup_ioctx(ctx);
 	if (likely(NULL != ioctx)) {
-		io_destroy(ioctx);
+		kill_ioctx(ioctx);
 		put_ioctx(ioctx);
 		return 0;
 	}

@@ -1296,25 +1283,6 @@ static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
 	if (ret)
 		goto out_put_req;
 
-	spin_lock_irq(&ctx->ctx_lock);
-	/*
-	 * We could have raced with io_destroy() and are currently holding a
-	 * reference to ctx which should be destroyed. We cannot submit IO
-	 * since ctx gets freed as soon as io_submit() puts its reference.  The
-	 * check here is reliable: io_destroy() sets ctx->dead before waiting
-	 * for outstanding IO and the barrier between these two is realized by
-	 * unlock of mm->ioctx_lock and lock of ctx->ctx_lock.  Analogously we
-	 * increment ctx->reqs_active before checking for ctx->dead and the
-	 * barrier is realized by unlock and lock of ctx->ctx_lock. Thus if we
-	 * don't see ctx->dead set here, io_destroy() waits for our IO to
-	 * finish.
-	 */
-	if (ctx->dead)
-		ret = -EINVAL;
-	spin_unlock_irq(&ctx->ctx_lock);
-	if (ret)
-		goto out_put_req;
-
 	if (unlikely(kiocbIsCancelled(req))) {
 		ret = -EINTR;
 	} else {

@@ -1340,9 +1308,6 @@ out_put_req:
 	spin_unlock_irq(&ctx->ctx_lock);
 
 	atomic_dec(&ctx->reqs_active);
-	if (unlikely(!atomic_read(&ctx->reqs_active) && ctx->dead))
-		wake_up_all(&ctx->wait);
-
 	aio_put_req(req);	/* drop extra ref to req */
 	aio_put_req(req);	/* drop i/o ref to req */
 	return ret;

index 168dfe1..3088723 100644
@@ -330,6 +330,92 @@ do {									\
 	__ret;								\
 })
 
+#define __wait_event_hrtimeout(wq, condition, timeout, state)		\
+({									\
+	int __ret = 0;							\
+	DEFINE_WAIT(__wait);						\
+	struct hrtimer_sleeper __t;					\
+									\
+	hrtimer_init_on_stack(&__t.timer, CLOCK_MONOTONIC,		\
+			      HRTIMER_MODE_REL);			\
+	hrtimer_init_sleeper(&__t, current);				\
+	if ((timeout).tv64 != KTIME_MAX)				\
+		hrtimer_start_range_ns(&__t.timer, timeout,		\
+				       current->timer_slack_ns,		\
+				       HRTIMER_MODE_REL);		\
+									\
+	for (;;) {							\
+		prepare_to_wait(&wq, &__wait, state);			\
+		if (condition)						\
+			break;						\
+		if (state == TASK_INTERRUPTIBLE &&			\
+		    signal_pending(current)) {				\
+			__ret = -ERESTARTSYS;				\
+			break;						\
+		}							\
+		if (!__t.task) {					\
+			__ret = -ETIME;					\
+			break;						\
+		}							\
+		schedule();						\
+	}								\
+									\
+	hrtimer_cancel(&__t.timer);					\
+	destroy_hrtimer_on_stack(&__t.timer);				\
+	finish_wait(&wq, &__wait);					\
+	__ret;								\
+})
+
+/**
+ * wait_event_hrtimeout - sleep until a condition gets true or a timeout elapses
+ * @wq: the waitqueue to wait on
+ * @condition: a C expression for the event to wait for
+ * @timeout: timeout, in jiffies
+ *
+ * The process is put to sleep (TASK_UNINTERRUPTIBLE) until the
+ * @condition evaluates to true or a signal is received.
+ * The @condition is checked each time the waitqueue @wq is woken up.
+ *
+ * wake_up() has to be called after changing any variable that could
+ * change the result of the wait condition.
+ *
+ * The function returns 0 if @condition became true, or -ETIME if the timeout
+ * elapsed.
+ */
+#define wait_event_hrtimeout(wq, condition, timeout)			\
+({									\
+	int __ret = 0;							\
+	if (!(condition))						\
+		__ret = __wait_event_hrtimeout(wq, condition, timeout,	\
+					       TASK_UNINTERRUPTIBLE);	\
+	__ret;								\
+})
+
+/**
+ * wait_event_interruptible_hrtimeout - sleep until a condition gets true or a timeout elapses
+ * @wq: the waitqueue to wait on
+ * @condition: a C expression for the event to wait for
+ * @timeout: timeout, in jiffies
+ *
+ * The process is put to sleep (TASK_INTERRUPTIBLE) until the
+ * @condition evaluates to true or a signal is received.
+ * The @condition is checked each time the waitqueue @wq is woken up.
+ *
+ * wake_up() has to be called after changing any variable that could
+ * change the result of the wait condition.
+ *
+ * The function returns 0 if @condition became true, -ERESTARTSYS if it was
+ * interrupted by a signal, or -ETIME if the timeout elapsed.
+ */
+#define wait_event_interruptible_hrtimeout(wq, condition, timeout)	\
+({									\
+	long __ret = 0;							\
+	if (!(condition))						\
+		__ret = __wait_event_hrtimeout(wq, condition, timeout,	\
+					       TASK_INTERRUPTIBLE);	\
+	__ret;								\
+})
+
 #define __wait_event_interruptible_exclusive(wq, condition, ret)	\
 do {									\
 	DEFINE_WAIT(__wait);						\

@@ -370,7 +370,7 @@ do {									\
  * wait_event_hrtimeout - sleep until a condition gets true or a timeout elapses
  * @wq: the waitqueue to wait on
  * @condition: a C expression for the event to wait for
- * @timeout: timeout, in jiffies
+ * @timeout: timeout, as a ktime_t
  *
  * The process is put to sleep (TASK_UNINTERRUPTIBLE) until the
  * @condition evaluates to true or a signal is received.

@@ -395,7 +395,7 @@ do {									\
  * wait_event_interruptible_hrtimeout - sleep until a condition gets true or a timeout elapses
  * @wq: the waitqueue to wait on
  * @condition: a C expression for the event to wait for
- * @timeout: timeout, in jiffies
+ * @timeout: timeout, as a ktime_t
  *
  * The process is put to sleep (TASK_INTERRUPTIBLE) until the
  * @condition evaluates to true or a signal is received.

index e0eb23d..06e1dd0 100644
@@ -63,7 +63,7 @@ struct aio_ring_info {
 	unsigned long		mmap_size;
 
 	struct page		**ring_pages;
-	spinlock_t		ring_lock;
+	struct mutex		ring_lock;
 	long			nr_pages;
 
 	unsigned		nr, tail;

@@ -341,7 +341,7 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
 	atomic_set(&ctx->users, 2);
 	atomic_set(&ctx->dead, 0);
 	spin_lock_init(&ctx->ctx_lock);
-	spin_lock_init(&ctx->ring_info.ring_lock);
+	mutex_init(&ctx->ring_info.ring_lock);
 	init_waitqueue_head(&ctx->wait);
 
 	INIT_LIST_HEAD(&ctx->active_reqs);

@@ -744,187 +744,115 @@ put_rq:
 }
 EXPORT_SYMBOL(aio_complete);
 
-/* aio_read_evt
- *	Pull an event off of the ioctx's event ring.  Returns the number of 
- *	events fetched (0 or 1 ;-)
- *	FIXME: make this use cmpxchg.
- *	TODO: make the ringbuffer user mmap()able (requires FIXME).
+/* aio_read_events
+ *	Pull an event off of the ioctx's event ring.  Returns the number of
+ *	events fetched
  */
-static int aio_read_evt(struct kioctx *ioctx, struct io_event *ent)
+static int aio_read_events_ring(struct kioctx *ctx,
+				struct io_event __user *event, long nr)
 {
-	struct aio_ring_info *info = &ioctx->ring_info;
+	struct aio_ring_info *info = &ctx->ring_info;
 	struct aio_ring *ring;
-	unsigned long head;
-	int ret = 0;
+	unsigned head, pos;
+	int ret = 0, copy_ret;
+
+	if (!mutex_trylock(&info->ring_lock)) {
+		__set_current_state(TASK_RUNNING);
+		mutex_lock(&info->ring_lock);
+	}
 
 	ring = kmap_atomic(info->ring_pages[0]);
-	pr_debug("h%u t%u m%u\n", ring->head, ring->tail, ring->nr);
+	head = ring->head;
+	kunmap_atomic(ring);
+
+	pr_debug("h%u t%u m%u\n", head, info->tail, info->nr);
 
-	if (ring->head == ring->tail)
+	if (head == info->tail)
 		goto out;
 
-	spin_lock(&info->ring_lock);
-
-	head = ring->head % info->nr;
-	if (head != ring->tail) {
-		struct io_event *evp = aio_ring_event(info, head);
-		*ent = *evp;
-		head = (head + 1) % info->nr;
-		smp_mb(); /* finish reading the event before updatng the head */
-		ring->head = head;
-		ret = 1;
-		put_aio_ring_event(evp);
+	__set_current_state(TASK_RUNNING);
+
+	while (ret < nr) {
+		unsigned i = (head < info->tail ? info->tail : info->nr) - head;
+		struct io_event *ev;
+		struct page *page;
+
+		if (head == info->tail)
+			break;
+
+		i = min_t(int, i, nr - ret);
+		i = min_t(int, i, AIO_EVENTS_PER_PAGE -
+			  ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
+
+		pos = head + AIO_EVENTS_OFFSET;
+		page = info->ring_pages[pos / AIO_EVENTS_PER_PAGE];
+		pos %= AIO_EVENTS_PER_PAGE;
+
+		ev = kmap(page);
+		copy_ret = copy_to_user(event + ret, ev + pos, sizeof(*ev) * i);
+		kunmap(page);
+
+		if (unlikely(copy_ret)) {
+			ret = -EFAULT;
+			goto out;
+		}
+
+		ret += i;
+		head += i;
+		head %= info->nr;
 	}
-	spin_unlock(&info->ring_lock);
 
-out:
+	ring = kmap_atomic(info->ring_pages[0]);
+	ring->head = head;
 	kunmap_atomic(ring);
-	pr_debug("%d  h%u t%u\n", ret, ring->head, ring->tail);
+
+	pr_debug("%d  h%u t%u\n", ret, head, info->tail);
+out:
+	mutex_unlock(&info->ring_lock);
+
 	return ret;
 }
 
-struct aio_timeout {
-	struct timer_list	timer;
-	int			timed_out;
-	struct task_struct	*p;
-};
-
-static void timeout_func(unsigned long data)
+static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
+			    struct io_event __user *event, long *i)
 {
-	struct aio_timeout *to = (struct aio_timeout *)data;
+	long ret = aio_read_events_ring(ctx, event + *i, nr - *i);
 
-	to->timed_out = 1;
-	wake_up_process(to->p);
-}
+	if (ret > 0)
+		*i += ret;
 
-static inline void init_timeout(struct aio_timeout *to)
-{
-	setup_timer_on_stack(&to->timer, timeout_func, (unsigned long) to);
-	to->timed_out = 0;
-	to->p = current;
-}
+	if (unlikely(atomic_read(&ctx->dead)))
+		ret = -EINVAL;
 
-static inline void set_timeout(long start_jiffies, struct aio_timeout *to,
-			       const struct timespec *ts)
-{
-	to->timer.expires = start_jiffies + timespec_to_jiffies(ts);
-	if (time_after(to->timer.expires, jiffies))
-		add_timer(&to->timer);
-	else
-		to->timed_out = 1;
-}
+	if (!*i)
+		*i = ret;
 
-static inline void clear_timeout(struct aio_timeout *to)
-{
-	del_singleshot_timer_sync(&to->timer);
+	return ret < 0 || *i >= min_nr;
 }
 
-static int read_events(struct kioctx *ctx,
-			long min_nr, long nr,
+static long read_events(struct kioctx *ctx, long min_nr, long nr,
 			struct io_event __user *event,
 			struct timespec __user *timeout)
 {
-	long			start_jiffies = jiffies;
-	struct task_struct	*tsk = current;
-	DECLARE_WAITQUEUE(wait, tsk);
-	int			ret;
-	int			i = 0;
-	struct io_event		ent;
-	struct aio_timeout	to;
-
-	/* needed to zero any padding within an entry (there shouldn't be 
-	 * any, but C is fun!
-	 */
-	memset(&ent, 0, sizeof(ent));
-	ret = 0;
-	while (likely(i < nr)) {
-		ret = aio_read_evt(ctx, &ent);
-		if (unlikely(ret <= 0))
-			break;
-
-		pr_debug("%Lx %Lx %Lx %Lx\n",
-			 ent.data, ent.obj, ent.res, ent.res2);
-
-		/* Could we split the check in two? */
-		ret = -EFAULT;
-		if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
-			pr_debug("lost an event due to EFAULT.\n");
-			break;
-		}
-		ret = 0;
-
-		/* Good, event copied to userland, update counts. */
-		event ++;
-		i ++;
-	}
-
-	if (min_nr <= i)
-		return i;
-	if (ret)
-		return ret;
-
-	/* End fast path */
+	ktime_t until = { .tv64 = KTIME_MAX };
+	long ret = 0;
 
-	init_timeout(&to);
 	if (timeout) {
 		struct timespec	ts;
-		ret = -EFAULT;
+
 		if (unlikely(copy_from_user(&ts, timeout, sizeof(ts))))
-			goto out;
+			return -EFAULT;
 
-		set_timeout(start_jiffies, &to, &ts);
+		until = timespec_to_ktime(ts);
 	}
 
-	while (likely(i < nr)) {
-		add_wait_queue_exclusive(&ctx->wait, &wait);
-		do {
-			set_task_state(tsk, TASK_INTERRUPTIBLE);
-			ret = aio_read_evt(ctx, &ent);
-			if (ret)
-				break;
-			if (min_nr <= i)
-				break;
-			if (unlikely(atomic_read(&ctx->dead))) {
-				ret = -EINVAL;
-				break;
-			}
-			if (to.timed_out)	/* Only check after read evt */
-				break;
-			/* Try to only show up in io wait if there are ops
-			 *  in flight */
-			if (atomic_read(&ctx->reqs_active))
-				io_schedule();
-			else
-				schedule();
-			if (signal_pending(tsk)) {
-				ret = -EINTR;
-				break;
-			}
-			/*ret = aio_read_evt(ctx, &ent);*/
-		} while (1) ;
-
-		set_task_state(tsk, TASK_RUNNING);
-		remove_wait_queue(&ctx->wait, &wait);
-
-		if (unlikely(ret <= 0))
-			break;
-
-		ret = -EFAULT;
-		if (unlikely(copy_to_user(event, &ent, sizeof(ent)))) {
-			pr_debug("lost an event due to EFAULT.\n");
-			break;
-		}
+	wait_event_interruptible_hrtimeout(ctx->wait,
+			aio_read_events(ctx, min_nr, nr, event, &ret), until);
 
-		/* Good, event copied to userland, update counts. */
-		event ++;
-		i ++;
-	}
+	if (!ret && signal_pending(current))
+		ret = -EINTR;
 
-	if (timeout)
-		clear_timeout(&to);
-out:
-	destroy_timer_on_stack(&to.timer);
-	return i ? i : ret;
+	return ret;
 }
 
 /* sys_io_setup:

index 4033ebb..21b2c27 100644
@@ -837,12 +837,13 @@ EXPORT_SYMBOL(aio_complete_batch);
  *	Pull an event off of the ioctx's event ring.  Returns the number of
  *	events fetched
  */
-static int aio_read_events_ring(struct kioctx *ctx,
-				struct io_event __user *event, long nr)
+static long aio_read_events_ring(struct kioctx *ctx,
+				 struct io_event __user *event, long nr)
 {
 	struct aio_ring *ring;
 	unsigned head, pos;
-	int ret = 0, copy_ret;
+	long ret = 0;
+	int copy_ret;
 
 	if (!mutex_trylock(&ctx->ring_lock)) {
 		__set_current_state(TASK_RUNNING);

@@ -861,23 +862,24 @@ static int aio_read_events_ring(struct kioctx *ctx,
 	__set_current_state(TASK_RUNNING);
 
 	while (ret < nr) {
-		unsigned i = (head < ctx->shadow_tail ? ctx->shadow_tail : ctx->nr) - head;
+		long avail = (head < ctx->shadow_tail
+			      ? ctx->shadow_tail : ctx->nr) - head;
 		struct io_event *ev;
 		struct page *page;
 
 		if (head == ctx->shadow_tail)
 			break;
 
-		i = min_t(int, i, nr - ret);
-		i = min_t(int, i, AIO_EVENTS_PER_PAGE -
-			  ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
+		avail = min(avail, nr - ret);
+		avail = min_t(long, avail, AIO_EVENTS_PER_PAGE -
+			      ((head + AIO_EVENTS_OFFSET) % AIO_EVENTS_PER_PAGE));
 
 		pos = head + AIO_EVENTS_OFFSET;
 		page = ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE];
 		pos %= AIO_EVENTS_PER_PAGE;
 
 		ev = kmap(page);
-		copy_ret = copy_to_user(event + ret, ev + pos, sizeof(*ev) * i);
+		copy_ret = copy_to_user(event + ret, ev + pos, sizeof(*ev) * avail);
 		kunmap(page);
 
 		if (unlikely(copy_ret)) {

@@ -885,8 +887,8 @@ static int aio_read_events_ring(struct kioctx *ctx,
 			goto out;
 		}
 
-		ret += i;
-		head += i;
+		ret += avail;
+		head += avail;
 		head %= ctx->nr;
 	}
 

@@ -895,7 +897,7 @@ static int aio_read_events_ring(struct kioctx *ctx,
 	kunmap_atomic(ring);
 	flush_dcache_page(ctx->ring_pages[0]);
 
-	pr_debug("%d  h%u t%u\n", ret, head, ctx->shadow_tail);
+	pr_debug("%li  h%u t%u\n", ret, head, ctx->shadow_tail);
 
 	put_reqs_available(ctx, ret);
 out:

index 21b2c27..932170a 100644
@@ -102,6 +102,19 @@ struct kioctx {
 	struct {
 		struct mutex	ring_lock;
 		wait_queue_head_t wait;
+
+		/*
+		 * Copy of the real tail, that aio_complete uses - to reduce
+		 * cacheline bouncing. The real tail will tend to be much more
+		 * contended - since typically events are delivered one at a
+		 * time, and then aio_read_events() slurps them up a bunch at a
+		 * time - so it's helpful if aio_read_events() isn't also
+		 * contending for the tail. So, aio_complete() updates
+		 * shadow_tail whenever it updates tail.
+		 *
+		 * Also needed because tail is used as a hacky lock and isn't
+		 * always the real tail.
+		 */
 		unsigned	shadow_tail;
 	} ____cacheline_aligned_in_smp;
 

@@ -845,10 +858,7 @@ static long aio_read_events_ring(struct kioctx *ctx,
 	long ret = 0;
 	int copy_ret;
 
-	if (!mutex_trylock(&ctx->ring_lock)) {
-		__set_current_state(TASK_RUNNING);
-		mutex_lock(&ctx->ring_lock);
-	}
+	mutex_lock(&ctx->ring_lock);
 
 	ring = kmap_atomic(ctx->ring_pages[0]);
 	head = ring->head;

@@ -859,8 +869,6 @@ static long aio_read_events_ring(struct kioctx *ctx,
 	if (head == ctx->shadow_tail)
 		goto out;
 
-	__set_current_state(TASK_RUNNING);
-
 	while (ret < nr) {
 		long avail = (head < ctx->shadow_tail
 			      ? ctx->shadow_tail : ctx->nr) - head;

@@ -939,6 +947,20 @@ static long read_events(struct kioctx *ctx, long min_nr, long nr,
 		until = timespec_to_ktime(ts);
 	}
 
+	/*
+	 * Note that aio_read_events() is being called as the conditional - i.e.
+	 * we're calling it after prepare_to_wait() has set task state to
+	 * TASK_INTERRUPTIBLE.
+	 *
+	 * But aio_read_events() can block, and if it blocks it's going to flip
+	 * the task state back to TASK_RUNNING.
+	 *
+	 * This should be ok, provided it doesn't flip the state back to
+	 * TASK_RUNNING and return 0 too much - that causes us to spin. That
+	 * will only happen if the mutex_lock() call blocks, and we then find
+	 * the ringbuffer empty. So in practice we should be ok, but it's
+	 * something to be aware of when touching this code.
+	 */
 	wait_event_interruptible_hrtimeout(ctx->wait,
 			aio_read_events(ctx, min_nr, nr, event, &ret), until);
 

index 06e1dd0..c1047c8 100644
@@ -208,33 +208,15 @@ static int aio_setup_ring(struct kioctx *ctx)
 	ring->incompat_features = AIO_RING_INCOMPAT_FEATURES;
 	ring->header_length = sizeof(struct aio_ring);
 	kunmap_atomic(ring);
+	flush_dcache_page(info->ring_pages[0]);
 
 	return 0;
 }
 
-
-/* aio_ring_event: returns a pointer to the event at the given index from
- * kmap_atomic().  Release the pointer with put_aio_ring_event();
- */
 #define AIO_EVENTS_PER_PAGE	(PAGE_SIZE / sizeof(struct io_event))
 #define AIO_EVENTS_FIRST_PAGE	((PAGE_SIZE - sizeof(struct aio_ring)) / sizeof(struct io_event))
 #define AIO_EVENTS_OFFSET	(AIO_EVENTS_PER_PAGE - AIO_EVENTS_FIRST_PAGE)
 
-#define aio_ring_event(info, nr) ({					\
-	unsigned pos = (nr) + AIO_EVENTS_OFFSET;			\
-	struct io_event *__event;					\
-	__event = kmap_atomic(						\
-			(info)->ring_pages[pos / AIO_EVENTS_PER_PAGE]); \
-	__event += pos % AIO_EVENTS_PER_PAGE;				\
-	__event;							\
-})
-
-#define put_aio_ring_event(event) do {		\
-	struct io_event *__event = (event);	\
-	(void)__event;				\
-	kunmap_atomic((void *)((unsigned long)__event & PAGE_MASK)); \
-} while(0)
-
 static int kiocb_cancel(struct kioctx *ctx, struct kiocb *kiocb,
 			struct io_event *res)
 {

@@ -645,9 +627,9 @@ void aio_complete(struct kiocb *iocb, long res, long res2)
 	struct kioctx	*ctx = iocb->ki_ctx;
 	struct aio_ring_info	*info;
 	struct aio_ring	*ring;
-	struct io_event	*event;
+	struct io_event	*ev_page, *event;
 	unsigned long	flags;
-	unsigned long	tail;
+	unsigned tail, pos;
 
 	/*
 	 * Special case handling for sync iocbs:

@@ -686,19 +668,24 @@ void aio_complete(struct kiocb *iocb, long res, long res2)
 	if (kiocbIsCancelled(iocb))
 		goto put_rq;
 
-	ring = kmap_atomic(info->ring_pages[0]);
-
 	tail = info->tail;
-	event = aio_ring_event(info, tail);
+	pos = tail + AIO_EVENTS_OFFSET;
+
 	if (++tail >= info->nr)
 		tail = 0;
 
+	ev_page = kmap_atomic(info->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
+	event = ev_page + pos % AIO_EVENTS_PER_PAGE;
+
 	event->obj = (u64)(unsigned long)iocb->ki_obj.user;
 	event->data = iocb->ki_user_data;
 	event->res = res;
 	event->res2 = res2;
 
-	pr_debug("%p[%lu]: %p: %p %Lx %lx %lx\n",
+	kunmap_atomic(ev_page);
+	flush_dcache_page(info->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
+
+	pr_debug("%p[%u]: %p: %p %Lx %lx %lx\n",
 		 ctx, tail, iocb, iocb->ki_obj.user, iocb->ki_user_data,
 		 res, res2);
 

@@ -708,12 +695,13 @@ void aio_complete(struct kiocb *iocb, long res, long res2)
 	smp_wmb();	/* make event visible before updating tail */
 
 	info->tail = tail;
-	ring->tail = tail;
 
-	put_aio_ring_event(event);
+	ring = kmap_atomic(info->ring_pages[0]);
+	ring->tail = tail;
 	kunmap_atomic(ring);
+	flush_dcache_page(info->ring_pages[0]);
 
-	pr_debug("added to ring %p at [%lu]\n", iocb, tail);
+	pr_debug("added to ring %p at [%u]\n", iocb, tail);
 
 	/*
 	 * Check if the user asked us to deliver the result through an

@@ -805,6 +793,7 @@ static int aio_read_events_ring(struct kioctx *ctx,
 	ring = kmap_atomic(info->ring_pages[0]);
 	ring->head = head;
 	kunmap_atomic(ring);
+	flush_dcache_page(info->ring_pages[0]);
 
 	pr_debug("%d  h%u t%u\n", ret, head, info->tail);
 out: