[PATCH] block: avoid extra bio reference for async O_DIRECT

[PATCH] block: avoid extra bio reference for async O_DIRECT

[Jens Axboe]

The bio referencing has a trick that doesn't do any actual atomic
inc/dec on the reference count until we have to elevator to > 1. For the
async IO O_DIRECT case, we can't use the simple DIO variants, so we use
__blkdev_direct_IO(). It always grabs an extra reference to the bio
after allocation, which means we then enter the slower path of actually
having to do atomic_inc/dec on the count.

We don't need to do that for the async case, unless we end up going
multi-bio, in which case we're already doing huge amounts of IO. For the
smaller IO case (< BIO_MAX_PAGES), we can do without the extra ref.

Signed-off-by: Jens Axboe <axboe@kernel.dk>

---

diff --git a/fs/block_dev.c b/fs/block_dev.c
index f3bd41f80421..44fc9da13f68 100644
--- a/fs/block_dev.c
+++ b/fs/block_dev.c
@@ -292,10 +292,26 @@ static int blkdev_iopoll(struct kiocb *kiocb, bool wait)
 	return blk_poll(q, READ_ONCE(kiocb->ki_cookie), wait);
 }
 
+static void blkdev_bio_dirty_release(struct bio *bio, bool should_dirty)
+{
+	if (should_dirty) {
+		bio_check_pages_dirty(bio);
+	} else {
+		struct bio_vec *bvec;
+		int i;
+
+		if (!bio_flagged(bio, BIO_HOLD_PAGES))
+			bio_for_each_segment_all(bvec, bio, i)
+				put_page(bvec->bv_page);
+		bio_put(bio);
+	}
+}
+
 static void blkdev_bio_end_io(struct bio *bio)
 {
 	struct blkdev_dio *dio = bio->bi_private;
 	bool should_dirty = dio->should_dirty;
+	bool should_free = false;
 
 	if (dio->multi_bio && !atomic_dec_and_test(&dio->ref)) {
 		if (bio->bi_status && !dio->bio.bi_status)
@@ -313,26 +329,20 @@ static void blkdev_bio_end_io(struct bio *bio)
 			}
 
 			dio->iocb->ki_complete(iocb, ret, 0);
-			bio_put(&dio->bio);
 		} else {
 			struct task_struct *waiter = dio->waiter;
 
 			WRITE_ONCE(dio->waiter, NULL);
 			blk_wake_io_task(waiter);
 		}
+		/* last bio is done, now we can free the parent holding dio */
+		should_free = dio->multi_bio;
 	}
 
-	if (should_dirty) {
-		bio_check_pages_dirty(bio);
-	} else {
-		struct bio_vec *bvec;
-		int i;
-
-		if (!bio_flagged(bio, BIO_HOLD_PAGES))
-			bio_for_each_segment_all(bvec, bio, i)
-				put_page(bvec->bv_page);
-		bio_put(bio);
-	}
+	if (!dio->is_sync)
+		blkdev_bio_dirty_release(bio, should_dirty);
+	if (should_free)
+		bio_put(&dio->bio);
 }
 
 static ssize_t
@@ -355,14 +365,15 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, int nr_pages)
 		return -EINVAL;
 
 	bio = bio_alloc_bioset(GFP_KERNEL, nr_pages, &blkdev_dio_pool);
-	bio_get(bio); /* extra ref for the completion handler */
 
 	dio = container_of(bio, struct blkdev_dio, bio);
 	dio->is_sync = is_sync = is_sync_kiocb(iocb);
-	if (dio->is_sync)
+	if (dio->is_sync) {
 		dio->waiter = current;
-	else
+		bio_get(bio);
+	} else {
 		dio->iocb = iocb;
+	}
 
 	dio->size = 0;
 	dio->multi_bio = false;
@@ -425,6 +436,13 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, int nr_pages)
 		}
 
 		if (!dio->multi_bio) {
+			/*
+			 * async needs an extra reference if we are goign
+			 * multi-bio only, so we can ensure that 'dio'
+			 * sticks around.
+			 */
+			if (!is_sync)
+				bio_get(bio);
 			dio->multi_bio = true;
 			atomic_set(&dio->ref, 2);
 		} else {
@@ -458,6 +476,7 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, int nr_pages)
 	if (likely(!ret))
 		ret = dio->size;
 
+	blkdev_bio_dirty_release(bio, dio->should_dirty);
 	bio_put(&dio->bio);
 	return ret;
 }

-- 
Jens Axboe

Re: [PATCH] block: avoid extra bio reference for async O_DIRECT

[Jens Axboe]

On 11/29/18 3:55 PM, Jens Axboe wrote:
> The bio referencing has a trick that doesn't do any actual atomic
> inc/dec on the reference count until we have to elevator to > 1. For the
> async IO O_DIRECT case, we can't use the simple DIO variants, so we use
> __blkdev_direct_IO(). It always grabs an extra reference to the bio
> after allocation, which means we then enter the slower path of actually
> having to do atomic_inc/dec on the count.
> 
> We don't need to do that for the async case, unless we end up going
> multi-bio, in which case we're already doing huge amounts of IO. For the
> smaller IO case (< BIO_MAX_PAGES), we can do without the extra ref.

I accidentally generated this against the aio-poll branch, here's a
version that is not.


diff --git a/fs/block_dev.c b/fs/block_dev.c
index d233a59ea364..5c0a224bf9cb 100644
--- a/fs/block_dev.c
+++ b/fs/block_dev.c
@@ -281,10 +281,25 @@ struct blkdev_dio {
 
 static struct bio_set blkdev_dio_pool;
 
+static void blkdev_bio_dirty_release(struct bio *bio, bool should_dirty)
+{
+	if (should_dirty) {
+		bio_check_pages_dirty(bio);
+	} else {
+		struct bio_vec *bvec;
+		int i;
+
+		bio_for_each_segment_all(bvec, bio, i)
+			put_page(bvec->bv_page);
+		bio_put(bio);
+	}
+}
+
 static void blkdev_bio_end_io(struct bio *bio)
 {
 	struct blkdev_dio *dio = bio->bi_private;
 	bool should_dirty = dio->should_dirty;
+	bool should_free = false;
 
 	if (dio->multi_bio && !atomic_dec_and_test(&dio->ref)) {
 		if (bio->bi_status && !dio->bio.bi_status)
@@ -302,25 +317,20 @@ static void blkdev_bio_end_io(struct bio *bio)
 			}
 
 			dio->iocb->ki_complete(iocb, ret, 0);
-			bio_put(&dio->bio);
 		} else {
 			struct task_struct *waiter = dio->waiter;
 
 			WRITE_ONCE(dio->waiter, NULL);
 			blk_wake_io_task(waiter);
 		}
+		/* last bio is done, now we can free the parent holding dio */
+		should_free = dio->multi_bio;
 	}
 
-	if (should_dirty) {
-		bio_check_pages_dirty(bio);
-	} else {
-		struct bio_vec *bvec;
-		int i;
-
-		bio_for_each_segment_all(bvec, bio, i)
-			put_page(bvec->bv_page);
-		bio_put(bio);
-	}
+	if (!dio->is_sync)
+		blkdev_bio_dirty_release(bio, should_dirty);
+	if (should_free)
+		bio_put(&dio->bio);
 }
 
 static ssize_t
@@ -343,14 +353,15 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, int nr_pages)
 		return -EINVAL;
 
 	bio = bio_alloc_bioset(GFP_KERNEL, nr_pages, &blkdev_dio_pool);
-	bio_get(bio); /* extra ref for the completion handler */
 
 	dio = container_of(bio, struct blkdev_dio, bio);
 	dio->is_sync = is_sync = is_sync_kiocb(iocb);
-	if (dio->is_sync)
+	if (dio->is_sync) {
 		dio->waiter = current;
-	else
+		bio_get(bio);
+	} else {
 		dio->iocb = iocb;
+	}
 
 	dio->size = 0;
 	dio->multi_bio = false;
@@ -400,6 +411,13 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, int nr_pages)
 		}
 
 		if (!dio->multi_bio) {
+			/*
+			 * async needs an extra reference if we are goign
+			 * multi-bio only, so we can ensure that 'dio'
+			 * sticks around.
+			 */
+			if (!is_sync)
+				bio_get(bio);
 			dio->multi_bio = true;
 			atomic_set(&dio->ref, 2);
 		} else {
@@ -433,6 +451,7 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, int nr_pages)
 	if (likely(!ret))
 		ret = dio->size;
 
+	blkdev_bio_dirty_release(bio, dio->should_dirty);
 	bio_put(&dio->bio);
 	return ret;
 }

-- 
Jens Axboe

Re: [PATCH] block: avoid extra bio reference for async O_DIRECT

[Christoph Hellwig]

I like the idea, but I don't think it is correct as-is.  We can't just
move the blkdev_bio_dirty_release for the last bio into the caller, as
the completion order that decrements the refcount might have different
ordering.

Something like the simpler patch below might archive your goal, though:

---
From 923ee8e2358de04037ba6f7269aaf321f7b2e173 Mon Sep 17 00:00:00 2001
From: Christoph Hellwig <hch@lst.de>
Date: Fri, 30 Nov 2018 09:23:48 +0100
Subject: block: avoid extra bio reference for async O_DIRECT

The bio referencing has a trick that doesn't do any actual atomic
inc/dec on the reference count until we have to elevator to > 1. For the
async IO O_DIRECT case, we can't use the simple DIO variants, so we use
__blkdev_direct_IO(). It always grabs an extra reference to the bio
after allocation, which means we then enter the slower path of actually
having to do atomic_inc/dec on the count.

We don't need to do that for the async case, unless we end up going
multi-bio, in which case we're already doing huge amounts of IO. For the
smaller IO case (< BIO_MAX_PAGES), we can do without the extra ref.

Based on an earlier patch (and commit log) from Jens Axboe.

Signed-off-by: Christoph Hellwig <hch@lst.de>
---
 fs/block_dev.c | 17 +++++++++++++----
 1 file changed, 13 insertions(+), 4 deletions(-)

diff --git a/fs/block_dev.c b/fs/block_dev.c
index d233a59ea364..e1886cc7048f 100644
--- a/fs/block_dev.c
+++ b/fs/block_dev.c
@@ -302,7 +302,8 @@ static void blkdev_bio_end_io(struct bio *bio)
 			}
 
 			dio->iocb->ki_complete(iocb, ret, 0);
-			bio_put(&dio->bio);
+			if (dio->multi_bio)
+				bio_put(&dio->bio);
 		} else {
 			struct task_struct *waiter = dio->waiter;
 
@@ -343,14 +344,15 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, int nr_pages)
 		return -EINVAL;
 
 	bio = bio_alloc_bioset(GFP_KERNEL, nr_pages, &blkdev_dio_pool);
-	bio_get(bio); /* extra ref for the completion handler */
 
 	dio = container_of(bio, struct blkdev_dio, bio);
 	dio->is_sync = is_sync = is_sync_kiocb(iocb);
-	if (dio->is_sync)
+	if (dio->is_sync) {
 		dio->waiter = current;
-	else
+		bio_get(bio);
+	} else {
 		dio->iocb = iocb;
+	}
 
 	dio->size = 0;
 	dio->multi_bio = false;
@@ -400,6 +402,13 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, int nr_pages)
 		}
 
 		if (!dio->multi_bio) {
+			/*
+			 * AIO needs an extra reference to ensure the dio
+			 * structure which is embedded into the first bio
+			 * stays around.
+			 */
+			if (!is_sync)
+				bio_get(bio);
 			dio->multi_bio = true;
 			atomic_set(&dio->ref, 2);
 		} else {
-- 
2.19.1

Re: [PATCH] block: avoid extra bio reference for async O_DIRECT

[Jens Axboe]

On 11/30/18 1:38 AM, Christoph Hellwig wrote:
> I like the idea, but I don't think it is correct as-is.  We can't just
> move the blkdev_bio_dirty_release for the last bio into the caller, as
> the completion order that decrements the refcount might have different
> ordering.

Ugh yeah, not sure why I forgot about the multi case for that one.

> Something like the simpler patch below might archive your goal, though:

That does look cleaner. Let me run it through some testing just to
ensure we don't leak or use-after-free.

-- 
Jens Axboe