[PATCHSET v7] Support for polled and buffed aio (and more)

[PATCHSET v7] Support for polled and buffed aio (and more)

[Jens Axboe]

For the grand introduction to this feature, see my original posting
here:

https://lore.kernel.org/linux-block/20181117235317.7366-1-axboe@kernel.dk/

and refer to the previous postings of this patchset for whatever
features were added there. Particularly v4 has some performance results:

https://lore.kernel.org/linux-block/20181130165646.27341-1-axboe@kernel.dk/

See changelog for other changes, but this specifically adds support for
kernel side backing for the SQ/CQ ring interface. There are two modes:

1) IOCTX_FLAG_SQTHREAD. This adds a single thread backing. Right now
   this needs to be woken up through io_ring_enter(2), but I'm also
   planning a polled mode where an application can submit IO without
   ever entering the kernel.

2) IOCTX_FLAG_SQWQ. Like SQTHREAD, this adds a kernel side backing for
   IO submission. This one uses a workqueue, so we can have higher
   levels of concurrency. The main use case here is buffered AIO. With
   this, we can support buffered aio with QD > 1. For cached data, we
   avoid doing a workqueue punt for reads. So those should return
   without any extra context switches.


Find it here for your cgit browsing:

http://git.kernel.dk/cgit/linux-block/log/?h=aio-poll

or clone/pull from this branch:

git://git.kernel.dk/linux-block aio-poll

Since v6:

- Add support for async buffered IO, through kthread/workqueue backing
- Rename aio_kiocb flags
- Comment new syscalls
- Bug fixes

 Documentation/filesystems/vfs.txt      |    3 +
 Documentation/sysctl/fs.txt            |    8 +-
 arch/x86/entry/syscalls/syscall_64.tbl |    2 +
 block/bio.c                            |   33 +-
 fs/aio.c                               | 1914 ++++++++++++++++++++++--
 fs/block_dev.c                         |   34 +-
 fs/file.c                              |   15 +-
 fs/file_table.c                        |   10 +-
 fs/gfs2/file.c                         |    2 +
 fs/iomap.c                             |   57 +-
 fs/xfs/xfs_file.c                      |    1 +
 include/linux/bio.h                    |    1 +
 include/linux/blk_types.h              |    2 +
 include/linux/file.h                   |    2 +
 include/linux/fs.h                     |    5 +-
 include/linux/iomap.h                  |    1 +
 include/linux/syscalls.h               |    5 +
 include/uapi/asm-generic/unistd.h      |    4 +-
 include/uapi/linux/aio_abi.h           |   35 +
 kernel/sys_ni.c                        |    2 +
 20 files changed, 1939 insertions(+), 197 deletions(-)

-- 
Jens Axboe

[PATCH 01/27] fs: add an iopoll method to struct file_operations

[Jens Axboe]

From: Christoph Hellwig <hch@lst.de>

This new methods is used to explicitly poll for I/O completion for an
iocb.  It must be called for any iocb submitted asynchronously (that
is with a non-null ki_complete) which has the IOCB_HIPRI flag set.

The method is assisted by a new ki_cookie field in struct iocb to store
the polling cookie.

TODO: we can probably union ki_cookie with the existing hint and I/O
priority fields to avoid struct kiocb growth.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 Documentation/filesystems/vfs.txt | 3 +++
 include/linux/fs.h                | 2 ++
 2 files changed, 5 insertions(+)

diff --git a/Documentation/filesystems/vfs.txt b/Documentation/filesystems/vfs.txt
index 5f71a252e2e0..d9dc5e4d82b9 100644
--- a/Documentation/filesystems/vfs.txt
+++ b/Documentation/filesystems/vfs.txt
@@ -857,6 +857,7 @@ struct file_operations {
 	ssize_t (*write) (struct file *, const char __user *, size_t, loff_t *);
 	ssize_t (*read_iter) (struct kiocb *, struct iov_iter *);
 	ssize_t (*write_iter) (struct kiocb *, struct iov_iter *);
+	int (*iopoll)(struct kiocb *kiocb, bool spin);
 	int (*iterate) (struct file *, struct dir_context *);
 	int (*iterate_shared) (struct file *, struct dir_context *);
 	__poll_t (*poll) (struct file *, struct poll_table_struct *);
@@ -902,6 +903,8 @@ otherwise noted.
 
   write_iter: possibly asynchronous write with iov_iter as source
 
+  iopoll: called when aio wants to poll for completions on HIPRI iocbs
+
   iterate: called when the VFS needs to read the directory contents
 
   iterate_shared: called when the VFS needs to read the directory contents
diff --git a/include/linux/fs.h b/include/linux/fs.h
index a1ab233e6469..6a5f71f8ae06 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -310,6 +310,7 @@ struct kiocb {
 	int			ki_flags;
 	u16			ki_hint;
 	u16			ki_ioprio; /* See linux/ioprio.h */
+	unsigned int		ki_cookie; /* for ->iopoll */
 } __randomize_layout;
 
 static inline bool is_sync_kiocb(struct kiocb *kiocb)
@@ -1781,6 +1782,7 @@ struct file_operations {
 	ssize_t (*write) (struct file *, const char __user *, size_t, loff_t *);
 	ssize_t (*read_iter) (struct kiocb *, struct iov_iter *);
 	ssize_t (*write_iter) (struct kiocb *, struct iov_iter *);
+	int (*iopoll)(struct kiocb *kiocb, bool spin);
 	int (*iterate) (struct file *, struct dir_context *);
 	int (*iterate_shared) (struct file *, struct dir_context *);
 	__poll_t (*poll) (struct file *, struct poll_table_struct *);
-- 
2.17.1

[PATCH 02/27] block: add REQ_HIPRI_ASYNC

[Jens Axboe]

For the upcoming async polled IO, we can't sleep allocating requests.
If we do, then we introduce a deadlock where the submitter already
has async polled IO in-flight, but can't wait for them to complete
since polled requests must be active found and reaped.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 include/linux/blk_types.h | 1 +
 1 file changed, 1 insertion(+)

diff --git a/include/linux/blk_types.h b/include/linux/blk_types.h
index 46c005d601ac..921d734d6b5d 100644
--- a/include/linux/blk_types.h
+++ b/include/linux/blk_types.h
@@ -347,6 +347,7 @@ enum req_flag_bits {
 #define REQ_NOWAIT		(1ULL << __REQ_NOWAIT)
 #define REQ_NOUNMAP		(1ULL << __REQ_NOUNMAP)
 #define REQ_HIPRI		(1ULL << __REQ_HIPRI)
+#define REQ_HIPRI_ASYNC		(REQ_HIPRI | REQ_NOWAIT)
 
 #define REQ_DRV			(1ULL << __REQ_DRV)
 #define REQ_SWAP		(1ULL << __REQ_SWAP)
-- 
2.17.1

[PATCH 03/27] block: wire up block device iopoll method

[Jens Axboe]

From: Christoph Hellwig <hch@lst.de>

Just call blk_poll on the iocb cookie, we can derive the block device
from the inode trivially.

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/block_dev.c | 10 ++++++++++
 1 file changed, 10 insertions(+)

diff --git a/fs/block_dev.c b/fs/block_dev.c
index e1886cc7048f..6de8d35f6e41 100644
--- a/fs/block_dev.c
+++ b/fs/block_dev.c
@@ -281,6 +281,14 @@ struct blkdev_dio {
 
 static struct bio_set blkdev_dio_pool;
 
+static int blkdev_iopoll(struct kiocb *kiocb, bool wait)
+{
+	struct block_device *bdev = I_BDEV(kiocb->ki_filp->f_mapping->host);
+	struct request_queue *q = bdev_get_queue(bdev);
+
+	return blk_poll(q, READ_ONCE(kiocb->ki_cookie), wait);
+}
+
 static void blkdev_bio_end_io(struct bio *bio)
 {
 	struct blkdev_dio *dio = bio->bi_private;
@@ -398,6 +406,7 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, int nr_pages)
 				bio->bi_opf |= REQ_HIPRI;
 
 			qc = submit_bio(bio);
+			WRITE_ONCE(iocb->ki_cookie, qc);
 			break;
 		}
 
@@ -2070,6 +2079,7 @@ const struct file_operations def_blk_fops = {
 	.llseek		= block_llseek,
 	.read_iter	= blkdev_read_iter,
 	.write_iter	= blkdev_write_iter,
+	.iopoll		= blkdev_iopoll,
 	.mmap		= generic_file_mmap,
 	.fsync		= blkdev_fsync,
 	.unlocked_ioctl	= block_ioctl,
-- 
2.17.1

[PATCH 04/27] block: use REQ_HIPRI_ASYNC for non-sync polled IO

[Jens Axboe]

Tell the block layer if it's a sync or async polled request, so it can
do the right thing.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/block_dev.c | 8 ++++++--
 1 file changed, 6 insertions(+), 2 deletions(-)

diff --git a/fs/block_dev.c b/fs/block_dev.c
index 6de8d35f6e41..b8f574615792 100644
--- a/fs/block_dev.c
+++ b/fs/block_dev.c
@@ -402,8 +402,12 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, int nr_pages)
 
 		nr_pages = iov_iter_npages(iter, BIO_MAX_PAGES);
 		if (!nr_pages) {
-			if (iocb->ki_flags & IOCB_HIPRI)
-				bio->bi_opf |= REQ_HIPRI;
+			if (iocb->ki_flags & IOCB_HIPRI) {
+				if (!is_sync)
+					bio->bi_opf |= REQ_HIPRI_ASYNC;
+				else
+					bio->bi_opf |= REQ_HIPRI;
+			}
 
 			qc = submit_bio(bio);
 			WRITE_ONCE(iocb->ki_cookie, qc);
-- 
2.17.1

[PATCH 05/27] iomap: wire up the iopoll method

[Jens Axboe]

From: Christoph Hellwig <hch@lst.de>

Store the request queue the last bio was submitted to in the iocb
private data in addition to the cookie so that we find the right block
device.  Also refactor the common direct I/O bio submission code into a
nice little helper.

Signed-off-by: Christoph Hellwig <hch@lst.de>

Modified to use REQ_HIPRI_ASYNC for async polled IO.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/gfs2/file.c        |  2 ++
 fs/iomap.c            | 47 +++++++++++++++++++++++++++++--------------
 fs/xfs/xfs_file.c     |  1 +
 include/linux/iomap.h |  1 +
 4 files changed, 36 insertions(+), 15 deletions(-)

diff --git a/fs/gfs2/file.c b/fs/gfs2/file.c
index 45a17b770d97..358157efc5b7 100644
--- a/fs/gfs2/file.c
+++ b/fs/gfs2/file.c
@@ -1280,6 +1280,7 @@ const struct file_operations gfs2_file_fops = {
 	.llseek		= gfs2_llseek,
 	.read_iter	= gfs2_file_read_iter,
 	.write_iter	= gfs2_file_write_iter,
+	.iopoll		= iomap_dio_iopoll,
 	.unlocked_ioctl	= gfs2_ioctl,
 	.mmap		= gfs2_mmap,
 	.open		= gfs2_open,
@@ -1310,6 +1311,7 @@ const struct file_operations gfs2_file_fops_nolock = {
 	.llseek		= gfs2_llseek,
 	.read_iter	= gfs2_file_read_iter,
 	.write_iter	= gfs2_file_write_iter,
+	.iopoll		= iomap_dio_iopoll,
 	.unlocked_ioctl	= gfs2_ioctl,
 	.mmap		= gfs2_mmap,
 	.open		= gfs2_open,
diff --git a/fs/iomap.c b/fs/iomap.c
index 9a5bf1e8925b..f3039989de73 100644
--- a/fs/iomap.c
+++ b/fs/iomap.c
@@ -1441,6 +1441,32 @@ struct iomap_dio {
 	};
 };
 
+int iomap_dio_iopoll(struct kiocb *kiocb, bool spin)
+{
+	struct request_queue *q = READ_ONCE(kiocb->private);
+
+	if (!q)
+		return 0;
+	return blk_poll(q, READ_ONCE(kiocb->ki_cookie), spin);
+}
+EXPORT_SYMBOL_GPL(iomap_dio_iopoll);
+
+static void iomap_dio_submit_bio(struct iomap_dio *dio, struct iomap *iomap,
+		struct bio *bio)
+{
+	atomic_inc(&dio->ref);
+
+	if (dio->iocb->ki_flags & IOCB_HIPRI) {
+		if (!dio->wait_for_completion)
+			bio->bi_opf |= REQ_HIPRI_ASYNC;
+		else
+			bio->bi_opf |= REQ_HIPRI;
+	}
+
+	dio->submit.last_queue = bdev_get_queue(iomap->bdev);
+	dio->submit.cookie = submit_bio(bio);
+}
+
 static ssize_t iomap_dio_complete(struct iomap_dio *dio)
 {
 	struct kiocb *iocb = dio->iocb;
@@ -1553,7 +1579,7 @@ static void iomap_dio_bio_end_io(struct bio *bio)
 	}
 }
 
-static blk_qc_t
+static void
 iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
 		unsigned len)
 {
@@ -1567,15 +1593,10 @@ iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
 	bio->bi_private = dio;
 	bio->bi_end_io = iomap_dio_bio_end_io;
 
-	if (dio->iocb->ki_flags & IOCB_HIPRI)
-		flags |= REQ_HIPRI;
-
 	get_page(page);
 	__bio_add_page(bio, page, len, 0);
 	bio_set_op_attrs(bio, REQ_OP_WRITE, flags);
-
-	atomic_inc(&dio->ref);
-	return submit_bio(bio);
+	iomap_dio_submit_bio(dio, iomap, bio);
 }
 
 static loff_t
@@ -1678,9 +1699,6 @@ iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
 				bio_set_pages_dirty(bio);
 		}
 
-		if (dio->iocb->ki_flags & IOCB_HIPRI)
-			bio->bi_opf |= REQ_HIPRI;
-
 		iov_iter_advance(dio->submit.iter, n);
 
 		dio->size += n;
@@ -1688,11 +1706,7 @@ iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
 		copied += n;
 
 		nr_pages = iov_iter_npages(&iter, BIO_MAX_PAGES);
-
-		atomic_inc(&dio->ref);
-
-		dio->submit.last_queue = bdev_get_queue(iomap->bdev);
-		dio->submit.cookie = submit_bio(bio);
+		iomap_dio_submit_bio(dio, iomap, bio);
 	} while (nr_pages);
 
 	/*
@@ -1903,6 +1917,9 @@ iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 	if (dio->flags & IOMAP_DIO_WRITE_FUA)
 		dio->flags &= ~IOMAP_DIO_NEED_SYNC;
 
+	WRITE_ONCE(iocb->ki_cookie, dio->submit.cookie);
+	WRITE_ONCE(iocb->private, dio->submit.last_queue);
+
 	if (!atomic_dec_and_test(&dio->ref)) {
 		if (!dio->wait_for_completion)
 			return -EIOCBQUEUED;
diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index e47425071e65..60c2da41f0fc 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -1203,6 +1203,7 @@ const struct file_operations xfs_file_operations = {
 	.write_iter	= xfs_file_write_iter,
 	.splice_read	= generic_file_splice_read,
 	.splice_write	= iter_file_splice_write,
+	.iopoll		= iomap_dio_iopoll,
 	.unlocked_ioctl	= xfs_file_ioctl,
 #ifdef CONFIG_COMPAT
 	.compat_ioctl	= xfs_file_compat_ioctl,
diff --git a/include/linux/iomap.h b/include/linux/iomap.h
index 9a4258154b25..0fefb5455bda 100644
--- a/include/linux/iomap.h
+++ b/include/linux/iomap.h
@@ -162,6 +162,7 @@ typedef int (iomap_dio_end_io_t)(struct kiocb *iocb, ssize_t ret,
 		unsigned flags);
 ssize_t iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 		const struct iomap_ops *ops, iomap_dio_end_io_t end_io);
+int iomap_dio_iopoll(struct kiocb *kiocb, bool spin);
 
 #ifdef CONFIG_SWAP
 struct file;
-- 
2.17.1

[PATCH 06/27] aio: use assigned completion handler

[Jens Axboe]

We know this is a read/write request, but in preparation for
having different kinds of those, ensure that we call the assigned
handler instead of assuming it's aio_complete_rq().

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/aio.c | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/fs/aio.c b/fs/aio.c
index 05647d352bf3..cf0de61743e8 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -1490,7 +1490,7 @@ static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
 		ret = -EINTR;
 		/*FALLTHRU*/
 	default:
-		aio_complete_rw(req, ret, 0);
+		req->ki_complete(req, ret, 0);
 	}
 }
 
-- 
2.17.1

[PATCH 07/27] aio: separate out ring reservation from req allocation

[Jens Axboe]

From: Christoph Hellwig <hch@lst.de>

This is in preparation for certain types of IO not needing a ring
reserveration.

Signed-off-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/aio.c | 30 +++++++++++++++++-------------
 1 file changed, 17 insertions(+), 13 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index cf0de61743e8..eaceb40e6cf5 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -901,7 +901,7 @@ static void put_reqs_available(struct kioctx *ctx, unsigned nr)
 	local_irq_restore(flags);
 }
 
-static bool get_reqs_available(struct kioctx *ctx)
+static bool __get_reqs_available(struct kioctx *ctx)
 {
 	struct kioctx_cpu *kcpu;
 	bool ret = false;
@@ -993,6 +993,14 @@ static void user_refill_reqs_available(struct kioctx *ctx)
 	spin_unlock_irq(&ctx->completion_lock);
 }
 
+static bool get_reqs_available(struct kioctx *ctx)
+{
+	if (__get_reqs_available(ctx))
+		return true;
+	user_refill_reqs_available(ctx);
+	return __get_reqs_available(ctx);
+}
+
 /* aio_get_req
  *	Allocate a slot for an aio request.
  * Returns NULL if no requests are free.
@@ -1001,24 +1009,15 @@ static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
 {
 	struct aio_kiocb *req;
 
-	if (!get_reqs_available(ctx)) {
-		user_refill_reqs_available(ctx);
-		if (!get_reqs_available(ctx))
-			return NULL;
-	}
-
 	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
 	if (unlikely(!req))
-		goto out_put;
+		return NULL;
 
 	percpu_ref_get(&ctx->reqs);
 	INIT_LIST_HEAD(&req->ki_list);
 	refcount_set(&req->ki_refcnt, 0);
 	req->ki_ctx = ctx;
 	return req;
-out_put:
-	put_reqs_available(ctx, 1);
-	return NULL;
 }
 
 static struct kioctx *lookup_ioctx(unsigned long ctx_id)
@@ -1805,9 +1804,13 @@ static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
 		return -EINVAL;
 	}
 
+	if (!get_reqs_available(ctx))
+		return -EAGAIN;
+
+	ret = -EAGAIN;
 	req = aio_get_req(ctx);
 	if (unlikely(!req))
-		return -EAGAIN;
+		goto out_put_reqs_available;
 
 	if (iocb.aio_flags & IOCB_FLAG_RESFD) {
 		/*
@@ -1870,11 +1873,12 @@ static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
 		goto out_put_req;
 	return 0;
 out_put_req:
-	put_reqs_available(ctx, 1);
 	percpu_ref_put(&ctx->reqs);
 	if (req->ki_eventfd)
 		eventfd_ctx_put(req->ki_eventfd);
 	kmem_cache_free(kiocb_cachep, req);
+out_put_reqs_available:
+	put_reqs_available(ctx, 1);
 	return ret;
 }
 
-- 
2.17.1

[PATCH 08/27] aio: don't zero entire aio_kiocb aio_get_req()

[Jens Axboe]

It's 192 bytes, fairly substantial. Most items don't need to be cleared,
especially not upfront. Clear the ones we do need to clear, and leave
the other ones for setup when the iocb is prepared and submitted.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/aio.c | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index eaceb40e6cf5..522c04864d82 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -1009,14 +1009,15 @@ static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
 {
 	struct aio_kiocb *req;
 
-	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL|__GFP_ZERO);
+	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
 	if (unlikely(!req))
 		return NULL;
 
 	percpu_ref_get(&ctx->reqs);
+	req->ki_ctx = ctx;
 	INIT_LIST_HEAD(&req->ki_list);
 	refcount_set(&req->ki_refcnt, 0);
-	req->ki_ctx = ctx;
+	req->ki_eventfd = NULL;
 	return req;
 }
 
@@ -1730,6 +1731,10 @@ static ssize_t aio_poll(struct aio_kiocb *aiocb, struct iocb *iocb)
 	if (unlikely(!req->file))
 		return -EBADF;
 
+	req->head = NULL;
+	req->woken = false;
+	req->cancelled = false;
+
 	apt.pt._qproc = aio_poll_queue_proc;
 	apt.pt._key = req->events;
 	apt.iocb = aiocb;
-- 
2.17.1

[PATCH 09/27] aio: only use blk plugs for > 2 depth submissions

[Jens Axboe]

Plugging is meant to optimize submission of a string of IOs, if we don't
have more than 2 being submitted, don't bother setting up a plug.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/aio.c | 18 ++++++++++++++----
 1 file changed, 14 insertions(+), 4 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index 522c04864d82..ed6c3914477a 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -69,6 +69,12 @@ struct aio_ring {
 	struct io_event		io_events[0];
 }; /* 128 bytes + ring size */
 
+/*
+ * Plugging is meant to work with larger batches of IOs. If we don't
+ * have more than the below, then don't bother setting up a plug.
+ */
+#define AIO_PLUG_THRESHOLD	2
+
 #define AIO_RING_PAGES	8
 
 struct kioctx_table {
@@ -1919,7 +1925,8 @@ SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
 	if (nr > ctx->nr_events)
 		nr = ctx->nr_events;
 
-	blk_start_plug(&plug);
+	if (nr > AIO_PLUG_THRESHOLD)
+		blk_start_plug(&plug);
 	for (i = 0; i < nr; i++) {
 		struct iocb __user *user_iocb;
 
@@ -1932,7 +1939,8 @@ SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
 		if (ret)
 			break;
 	}
-	blk_finish_plug(&plug);
+	if (nr > AIO_PLUG_THRESHOLD)
+		blk_finish_plug(&plug);
 
 	percpu_ref_put(&ctx->users);
 	return i ? i : ret;
@@ -1959,7 +1967,8 @@ COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
 	if (nr > ctx->nr_events)
 		nr = ctx->nr_events;
 
-	blk_start_plug(&plug);
+	if (nr > AIO_PLUG_THRESHOLD)
+		blk_start_plug(&plug);
 	for (i = 0; i < nr; i++) {
 		compat_uptr_t user_iocb;
 
@@ -1972,7 +1981,8 @@ COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
 		if (ret)
 			break;
 	}
-	blk_finish_plug(&plug);
+	if (nr > AIO_PLUG_THRESHOLD)
+		blk_finish_plug(&plug);
 
 	percpu_ref_put(&ctx->users);
 	return i ? i : ret;
-- 
2.17.1

[PATCH 10/27] aio: use iocb_put() instead of open coding it

[Jens Axboe]

Replace the percpu_ref_put() + kmem_cache_free() with a call to
iocb_put() instead.

Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/aio.c | 3 +--
 1 file changed, 1 insertion(+), 2 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index ed6c3914477a..cf93b92bfb1e 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -1884,10 +1884,9 @@ static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
 		goto out_put_req;
 	return 0;
 out_put_req:
-	percpu_ref_put(&ctx->reqs);
 	if (req->ki_eventfd)
 		eventfd_ctx_put(req->ki_eventfd);
-	kmem_cache_free(kiocb_cachep, req);
+	iocb_put(req);
 out_put_reqs_available:
 	put_reqs_available(ctx, 1);
 	return ret;
-- 
2.17.1

[PATCH 11/27] aio: split out iocb copy from io_submit_one()

[Jens Axboe]

In preparation of handing in iocbs in a different fashion as well. Also
make it clear that the iocb being passed in isn't modified, by marking
it const throughout.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/aio.c | 68 +++++++++++++++++++++++++++++++-------------------------
 1 file changed, 38 insertions(+), 30 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index cf93b92bfb1e..06c8bcc72496 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -1420,7 +1420,7 @@ static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
 	aio_complete(iocb, res, res2);
 }
 
-static int aio_prep_rw(struct kiocb *req, struct iocb *iocb)
+static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
 {
 	int ret;
 
@@ -1461,7 +1461,7 @@ static int aio_prep_rw(struct kiocb *req, struct iocb *iocb)
 	return ret;
 }
 
-static int aio_setup_rw(int rw, struct iocb *iocb, struct iovec **iovec,
+static int aio_setup_rw(int rw, const struct iocb *iocb, struct iovec **iovec,
 		bool vectored, bool compat, struct iov_iter *iter)
 {
 	void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
@@ -1500,8 +1500,8 @@ static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
 	}
 }
 
-static ssize_t aio_read(struct kiocb *req, struct iocb *iocb, bool vectored,
-		bool compat)
+static ssize_t aio_read(struct kiocb *req, const struct iocb *iocb,
+			bool vectored, bool compat)
 {
 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
 	struct iov_iter iter;
@@ -1533,8 +1533,8 @@ static ssize_t aio_read(struct kiocb *req, struct iocb *iocb, bool vectored,
 	return ret;
 }
 
-static ssize_t aio_write(struct kiocb *req, struct iocb *iocb, bool vectored,
-		bool compat)
+static ssize_t aio_write(struct kiocb *req, const struct iocb *iocb,
+			 bool vectored, bool compat)
 {
 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
 	struct iov_iter iter;
@@ -1589,7 +1589,8 @@ static void aio_fsync_work(struct work_struct *work)
 	aio_complete(container_of(req, struct aio_kiocb, fsync), ret, 0);
 }
 
-static int aio_fsync(struct fsync_iocb *req, struct iocb *iocb, bool datasync)
+static int aio_fsync(struct fsync_iocb *req, const struct iocb *iocb,
+		     bool datasync)
 {
 	if (unlikely(iocb->aio_buf || iocb->aio_offset || iocb->aio_nbytes ||
 			iocb->aio_rw_flags))
@@ -1717,7 +1718,7 @@ aio_poll_queue_proc(struct file *file, struct wait_queue_head *head,
 	add_wait_queue(head, &pt->iocb->poll.wait);
 }
 
-static ssize_t aio_poll(struct aio_kiocb *aiocb, struct iocb *iocb)
+static ssize_t aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
 {
 	struct kioctx *ctx = aiocb->ki_ctx;
 	struct poll_iocb *req = &aiocb->poll;
@@ -1789,27 +1790,23 @@ static ssize_t aio_poll(struct aio_kiocb *aiocb, struct iocb *iocb)
 	return 0;
 }
 
-static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
-			 bool compat)
+static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
+			   struct iocb __user *user_iocb, bool compat)
 {
 	struct aio_kiocb *req;
-	struct iocb iocb;
 	ssize_t ret;
 
-	if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
-		return -EFAULT;
-
 	/* enforce forwards compatibility on users */
-	if (unlikely(iocb.aio_reserved2)) {
+	if (unlikely(iocb->aio_reserved2)) {
 		pr_debug("EINVAL: reserve field set\n");
 		return -EINVAL;
 	}
 
 	/* prevent overflows */
 	if (unlikely(
-	    (iocb.aio_buf != (unsigned long)iocb.aio_buf) ||
-	    (iocb.aio_nbytes != (size_t)iocb.aio_nbytes) ||
-	    ((ssize_t)iocb.aio_nbytes < 0)
+	    (iocb->aio_buf != (unsigned long)iocb->aio_buf) ||
+	    (iocb->aio_nbytes != (size_t)iocb->aio_nbytes) ||
+	    ((ssize_t)iocb->aio_nbytes < 0)
 	   )) {
 		pr_debug("EINVAL: overflow check\n");
 		return -EINVAL;
@@ -1823,14 +1820,14 @@ static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
 	if (unlikely(!req))
 		goto out_put_reqs_available;
 
-	if (iocb.aio_flags & IOCB_FLAG_RESFD) {
+	if (iocb->aio_flags & IOCB_FLAG_RESFD) {
 		/*
 		 * If the IOCB_FLAG_RESFD flag of aio_flags is set, get an
 		 * instance of the file* now. The file descriptor must be
 		 * an eventfd() fd, and will be signaled for each completed
 		 * event using the eventfd_signal() function.
 		 */
-		req->ki_eventfd = eventfd_ctx_fdget((int) iocb.aio_resfd);
+		req->ki_eventfd = eventfd_ctx_fdget((int) iocb->aio_resfd);
 		if (IS_ERR(req->ki_eventfd)) {
 			ret = PTR_ERR(req->ki_eventfd);
 			req->ki_eventfd = NULL;
@@ -1845,32 +1842,32 @@ static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
 	}
 
 	req->ki_user_iocb = user_iocb;
-	req->ki_user_data = iocb.aio_data;
+	req->ki_user_data = iocb->aio_data;
 
-	switch (iocb.aio_lio_opcode) {
+	switch (iocb->aio_lio_opcode) {
 	case IOCB_CMD_PREAD:
-		ret = aio_read(&req->rw, &iocb, false, compat);
+		ret = aio_read(&req->rw, iocb, false, compat);
 		break;
 	case IOCB_CMD_PWRITE:
-		ret = aio_write(&req->rw, &iocb, false, compat);
+		ret = aio_write(&req->rw, iocb, false, compat);
 		break;
 	case IOCB_CMD_PREADV:
-		ret = aio_read(&req->rw, &iocb, true, compat);
+		ret = aio_read(&req->rw, iocb, true, compat);
 		break;
 	case IOCB_CMD_PWRITEV:
-		ret = aio_write(&req->rw, &iocb, true, compat);
+		ret = aio_write(&req->rw, iocb, true, compat);
 		break;
 	case IOCB_CMD_FSYNC:
-		ret = aio_fsync(&req->fsync, &iocb, false);
+		ret = aio_fsync(&req->fsync, iocb, false);
 		break;
 	case IOCB_CMD_FDSYNC:
-		ret = aio_fsync(&req->fsync, &iocb, true);
+		ret = aio_fsync(&req->fsync, iocb, true);
 		break;
 	case IOCB_CMD_POLL:
-		ret = aio_poll(req, &iocb);
+		ret = aio_poll(req, iocb);
 		break;
 	default:
-		pr_debug("invalid aio operation %d\n", iocb.aio_lio_opcode);
+		pr_debug("invalid aio operation %d\n", iocb->aio_lio_opcode);
 		ret = -EINVAL;
 		break;
 	}
@@ -1892,6 +1889,17 @@ static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
 	return ret;
 }
 
+static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
+			 bool compat)
+{
+	struct iocb iocb;
+
+	if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
+		return -EFAULT;
+
+	return __io_submit_one(ctx, &iocb, user_iocb, compat);
+}
+
 /* sys_io_submit:
  *	Queue the nr iocbs pointed to by iocbpp for processing.  Returns
  *	the number of iocbs queued.  May return -EINVAL if the aio_context
-- 
2.17.1

[PATCH 12/27] aio: abstract out io_event filler helper

[Jens Axboe]

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/aio.c | 14 ++++++++++----
 1 file changed, 10 insertions(+), 4 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index 06c8bcc72496..173f1f79dc8f 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -1063,6 +1063,15 @@ static inline void iocb_put(struct aio_kiocb *iocb)
 	}
 }
 
+static void aio_fill_event(struct io_event *ev, struct aio_kiocb *iocb,
+			   long res, long res2)
+{
+	ev->obj = (u64)(unsigned long)iocb->ki_user_iocb;
+	ev->data = iocb->ki_user_data;
+	ev->res = res;
+	ev->res2 = res2;
+}
+
 /* aio_complete
  *	Called when the io request on the given iocb is complete.
  */
@@ -1090,10 +1099,7 @@ static void aio_complete(struct aio_kiocb *iocb, long res, long res2)
 	ev_page = kmap_atomic(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
 	event = ev_page + pos % AIO_EVENTS_PER_PAGE;
 
-	event->obj = (u64)(unsigned long)iocb->ki_user_iocb;
-	event->data = iocb->ki_user_data;
-	event->res = res;
-	event->res2 = res2;
+	aio_fill_event(event, iocb, res, res2);
 
 	kunmap_atomic(ev_page);
 	flush_dcache_page(ctx->ring_pages[pos / AIO_EVENTS_PER_PAGE]);
-- 
2.17.1

[PATCH 13/27] aio: add io_setup2() system call

[Jens Axboe]

This is just like io_setup(), except add a flags argument to let the
caller control/define some of the io_context behavior.

Outside of the flags, we add an iocb array and two user pointers for
future use.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 Documentation/sysctl/fs.txt            |  8 +--
 arch/x86/entry/syscalls/syscall_64.tbl |  1 +
 fs/aio.c                               | 80 ++++++++++++++++++--------
 include/linux/syscalls.h               |  3 +
 include/uapi/asm-generic/unistd.h      |  4 +-
 kernel/sys_ni.c                        |  1 +
 6 files changed, 67 insertions(+), 30 deletions(-)

diff --git a/Documentation/sysctl/fs.txt b/Documentation/sysctl/fs.txt
index 819caf8ca05f..5e484eb7a25f 100644
--- a/Documentation/sysctl/fs.txt
+++ b/Documentation/sysctl/fs.txt
@@ -47,10 +47,10 @@ Currently, these files are in /proc/sys/fs:
 aio-nr & aio-max-nr:
 
 aio-nr is the running total of the number of events specified on the
-io_setup system call for all currently active aio contexts.  If aio-nr
-reaches aio-max-nr then io_setup will fail with EAGAIN.  Note that
-raising aio-max-nr does not result in the pre-allocation or re-sizing
-of any kernel data structures.
+io_setup/io_setup2 system call for all currently active aio contexts.
+If aio-nr reaches aio-max-nr then io_setup will fail with EAGAIN.
+Note that raising aio-max-nr does not result in the pre-allocation or
+re-sizing of any kernel data structures.
 
 ==============================================================
 
diff --git a/arch/x86/entry/syscalls/syscall_64.tbl b/arch/x86/entry/syscalls/syscall_64.tbl
index f0b1709a5ffb..67c357225fb0 100644
--- a/arch/x86/entry/syscalls/syscall_64.tbl
+++ b/arch/x86/entry/syscalls/syscall_64.tbl
@@ -343,6 +343,7 @@
 332	common	statx			__x64_sys_statx
 333	common	io_pgetevents		__x64_sys_io_pgetevents
 334	common	rseq			__x64_sys_rseq
+335	common	io_setup2		__x64_sys_io_setup2
 
 #
 # x32-specific system call numbers start at 512 to avoid cache impact
diff --git a/fs/aio.c b/fs/aio.c
index 173f1f79dc8f..0bad70eab553 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -100,6 +100,8 @@ struct kioctx {
 
 	unsigned long		user_id;
 
+	unsigned int		flags;
+
 	struct __percpu kioctx_cpu *cpu;
 
 	/*
@@ -686,10 +688,8 @@ static void aio_nr_sub(unsigned nr)
 	spin_unlock(&aio_nr_lock);
 }
 
-/* ioctx_alloc
- *	Allocates and initializes an ioctx.  Returns an ERR_PTR if it failed.
- */
-static struct kioctx *ioctx_alloc(unsigned nr_events)
+static struct kioctx *io_setup_flags(unsigned long ctxid,
+				     unsigned int nr_events, unsigned int flags)
 {
 	struct mm_struct *mm = current->mm;
 	struct kioctx *ctx;
@@ -701,6 +701,12 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
 	 */
 	unsigned int max_reqs = nr_events;
 
+	if (unlikely(ctxid || nr_events == 0)) {
+		pr_debug("EINVAL: ctx %lu nr_events %u\n",
+		         ctxid, nr_events);
+		return ERR_PTR(-EINVAL);
+	}
+
 	/*
 	 * We keep track of the number of available ringbuffer slots, to prevent
 	 * overflow (reqs_available), and we also use percpu counters for this.
@@ -726,6 +732,7 @@ static struct kioctx *ioctx_alloc(unsigned nr_events)
 	if (!ctx)
 		return ERR_PTR(-ENOMEM);
 
+	ctx->flags = flags;
 	ctx->max_reqs = max_reqs;
 
 	spin_lock_init(&ctx->ctx_lock);
@@ -1281,6 +1288,45 @@ static long read_events(struct kioctx *ctx, long min_nr, long nr,
 	return ret;
 }
 
+/* sys_io_setup2:
+ *	Like sys_io_setup(), except that it takes a set of flags
+ *	(IOCTX_FLAG_*), and some pointers to user structures:
+ *
+ *	*iocbs - pointer to array of struct iocb, for when
+ *	IOCTX_FLAG_USERIOCB is set in flags.
+ *
+ *	*user1 - reserved for future use
+ *
+ *	*user2 - reserved for future use.
+ */
+SYSCALL_DEFINE6(io_setup2, u32, nr_events, u32, flags, struct iocb __user *,
+		iocbs, void __user *, user1, void __user *, user2,
+		aio_context_t __user *, ctxp)
+{
+	struct kioctx *ioctx;
+	unsigned long ctx;
+	long ret;
+
+	if (flags || user1 || user2)
+		return -EINVAL;
+
+	ret = get_user(ctx, ctxp);
+	if (unlikely(ret))
+		goto out;
+
+	ioctx = io_setup_flags(ctx, nr_events, flags);
+	ret = PTR_ERR(ioctx);
+	if (IS_ERR(ioctx))
+		goto out;
+
+	ret = put_user(ioctx->user_id, ctxp);
+	if (ret)
+		kill_ioctx(current->mm, ioctx, NULL);
+	percpu_ref_put(&ioctx->users);
+out:
+	return ret;
+}
+
 /* 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
@@ -1296,7 +1342,7 @@ static long read_events(struct kioctx *ctx, long min_nr, long nr,
  */
 SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
 {
-	struct kioctx *ioctx = NULL;
+	struct kioctx *ioctx;
 	unsigned long ctx;
 	long ret;
 
@@ -1304,14 +1350,7 @@ SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
 	if (unlikely(ret))
 		goto out;
 
-	ret = -EINVAL;
-	if (unlikely(ctx || nr_events == 0)) {
-		pr_debug("EINVAL: ctx %lu nr_events %u\n",
-		         ctx, nr_events);
-		goto out;
-	}
-
-	ioctx = ioctx_alloc(nr_events);
+	ioctx = io_setup_flags(ctx, nr_events, 0);
 	ret = PTR_ERR(ioctx);
 	if (!IS_ERR(ioctx)) {
 		ret = put_user(ioctx->user_id, ctxp);
@@ -1327,7 +1366,7 @@ SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
 #ifdef CONFIG_COMPAT
 COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
 {
-	struct kioctx *ioctx = NULL;
+	struct kioctx *ioctx;
 	unsigned long ctx;
 	long ret;
 
@@ -1335,23 +1374,14 @@ COMPAT_SYSCALL_DEFINE2(io_setup, unsigned, nr_events, u32 __user *, ctx32p)
 	if (unlikely(ret))
 		goto out;
 
-	ret = -EINVAL;
-	if (unlikely(ctx || nr_events == 0)) {
-		pr_debug("EINVAL: ctx %lu nr_events %u\n",
-		         ctx, nr_events);
-		goto out;
-	}
-
-	ioctx = ioctx_alloc(nr_events);
+	ioctx = io_setup_flags(ctx, nr_events, 0);
 	ret = PTR_ERR(ioctx);
 	if (!IS_ERR(ioctx)) {
-		/* truncating is ok because it's a user address */
-		ret = put_user((u32)ioctx->user_id, ctx32p);
+		ret = put_user(ioctx->user_id, ctx32p);
 		if (ret)
 			kill_ioctx(current->mm, ioctx, NULL);
 		percpu_ref_put(&ioctx->users);
 	}
-
 out:
 	return ret;
 }
diff --git a/include/linux/syscalls.h b/include/linux/syscalls.h
index 2ac3d13a915b..a20a663d583f 100644
--- a/include/linux/syscalls.h
+++ b/include/linux/syscalls.h
@@ -287,6 +287,9 @@ static inline void addr_limit_user_check(void)
  */
 #ifndef CONFIG_ARCH_HAS_SYSCALL_WRAPPER
 asmlinkage long sys_io_setup(unsigned nr_reqs, aio_context_t __user *ctx);
+asmlinkage long sys_io_setup2(unsigned, unsigned, struct iocb __user *,
+				void __user *, void __user *,
+				aio_context_t __user *);
 asmlinkage long sys_io_destroy(aio_context_t ctx);
 asmlinkage long sys_io_submit(aio_context_t, long,
 			struct iocb __user * __user *);
diff --git a/include/uapi/asm-generic/unistd.h b/include/uapi/asm-generic/unistd.h
index c7f3321fbe43..1bbaa4c59f20 100644
--- a/include/uapi/asm-generic/unistd.h
+++ b/include/uapi/asm-generic/unistd.h
@@ -738,9 +738,11 @@ __SYSCALL(__NR_statx,     sys_statx)
 __SC_COMP(__NR_io_pgetevents, sys_io_pgetevents, compat_sys_io_pgetevents)
 #define __NR_rseq 293
 __SYSCALL(__NR_rseq, sys_rseq)
+#define __NR_io_setup2 294
+__SYSCALL(__NR_io_setup2, sys_io_setup2)
 
 #undef __NR_syscalls
-#define __NR_syscalls 294
+#define __NR_syscalls 295
 
 /*
  * 32 bit systems traditionally used different
diff --git a/kernel/sys_ni.c b/kernel/sys_ni.c
index df556175be50..17c8b4393669 100644
--- a/kernel/sys_ni.c
+++ b/kernel/sys_ni.c
@@ -37,6 +37,7 @@ asmlinkage long sys_ni_syscall(void)
  */
 
 COND_SYSCALL(io_setup);
+COND_SYSCALL(io_setup2);
 COND_SYSCALL_COMPAT(io_setup);
 COND_SYSCALL(io_destroy);
 COND_SYSCALL(io_submit);
-- 
2.17.1

[PATCH 14/27] aio: add support for having user mapped iocbs

[Jens Axboe]

For io_submit(), we have to first copy each pointer to an iocb, then
copy the iocb. The latter is 64 bytes in size, and that's a lot of
copying for a single IO.

Add support for setting IOCTX_FLAG_USERIOCB through the new io_setup2()
system call, which allows the iocbs to reside in userspace. If this flag
is used, then io_submit() doesn't take pointers to iocbs anymore, it
takes an index value into the array of iocbs instead. Similary, for
io_getevents(), the iocb ->obj will be the index, not the pointer to the
iocb.

See the change made to fio to support this feature, it's pretty
trivialy to adapt to. For applications, like fio, that previously
embedded the iocb inside a application private structure, some sort
of lookup table/structure is needed to find the private IO structure
from the index at io_getevents() time.

http://git.kernel.dk/cgit/fio/commit/?id=3c3168e91329c83880c91e5abc28b9d6b940fd95

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/aio.c                     | 126 +++++++++++++++++++++++++++++++----
 include/uapi/linux/aio_abi.h |   2 +
 2 files changed, 116 insertions(+), 12 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index 0bad70eab553..4e8c471a2598 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -92,6 +92,11 @@ struct ctx_rq_wait {
 	atomic_t count;
 };
 
+struct aio_mapped_range {
+	struct page **pages;
+	long nr_pages;
+};
+
 struct kioctx {
 	struct percpu_ref	users;
 	atomic_t		dead;
@@ -127,6 +132,8 @@ struct kioctx {
 	struct page		**ring_pages;
 	long			nr_pages;
 
+	struct aio_mapped_range	iocb_range;
+
 	struct rcu_work		free_rwork;	/* see free_ioctx() */
 
 	/*
@@ -222,6 +229,11 @@ static struct vfsmount *aio_mnt;
 static const struct file_operations aio_ring_fops;
 static const struct address_space_operations aio_ctx_aops;
 
+static const unsigned int iocb_page_shift =
+				ilog2(PAGE_SIZE / sizeof(struct iocb));
+
+static void aio_useriocb_unmap(struct kioctx *);
+
 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
 {
 	struct file *file;
@@ -578,6 +590,7 @@ static void free_ioctx(struct work_struct *work)
 					  free_rwork);
 	pr_debug("freeing %p\n", ctx);
 
+	aio_useriocb_unmap(ctx);
 	aio_free_ring(ctx);
 	free_percpu(ctx->cpu);
 	percpu_ref_exit(&ctx->reqs);
@@ -1288,6 +1301,70 @@ static long read_events(struct kioctx *ctx, long min_nr, long nr,
 	return ret;
 }
 
+static struct iocb *aio_iocb_from_index(struct kioctx *ctx, int index)
+{
+	struct iocb *iocb;
+
+	iocb = page_address(ctx->iocb_range.pages[index >> iocb_page_shift]);
+	index &= ((1 << iocb_page_shift) - 1);
+	return iocb + index;
+}
+
+static void aio_unmap_range(struct aio_mapped_range *range)
+{
+	int i;
+
+	if (!range->nr_pages)
+		return;
+
+	for (i = 0; i < range->nr_pages; i++)
+		put_page(range->pages[i]);
+
+	kfree(range->pages);
+	range->pages = NULL;
+	range->nr_pages = 0;
+}
+
+static int aio_map_range(struct aio_mapped_range *range, void __user *uaddr,
+			 size_t size, int gup_flags)
+{
+	int nr_pages, ret;
+
+	if ((unsigned long) uaddr & ~PAGE_MASK)
+		return -EINVAL;
+
+	nr_pages = (size + PAGE_SIZE - 1) >> PAGE_SHIFT;
+
+	range->pages = kzalloc(nr_pages * sizeof(struct page *), GFP_KERNEL);
+	if (!range->pages)
+		return -ENOMEM;
+
+	down_write(&current->mm->mmap_sem);
+	ret = get_user_pages((unsigned long) uaddr, nr_pages, gup_flags,
+				range->pages, NULL);
+	up_write(&current->mm->mmap_sem);
+
+	if (ret < nr_pages) {
+		kfree(range->pages);
+		return -ENOMEM;
+	}
+
+	range->nr_pages = nr_pages;
+	return 0;
+}
+
+static void aio_useriocb_unmap(struct kioctx *ctx)
+{
+	aio_unmap_range(&ctx->iocb_range);
+}
+
+static int aio_useriocb_map(struct kioctx *ctx, struct iocb __user *iocbs)
+{
+	size_t size = sizeof(struct iocb) * ctx->max_reqs;
+
+	return aio_map_range(&ctx->iocb_range, iocbs, size, 0);
+}
+
 /* sys_io_setup2:
  *	Like sys_io_setup(), except that it takes a set of flags
  *	(IOCTX_FLAG_*), and some pointers to user structures:
@@ -1307,7 +1384,9 @@ SYSCALL_DEFINE6(io_setup2, u32, nr_events, u32, flags, struct iocb __user *,
 	unsigned long ctx;
 	long ret;
 
-	if (flags || user1 || user2)
+	if (user1 || user2)
+		return -EINVAL;
+	if (flags & ~IOCTX_FLAG_USERIOCB)
 		return -EINVAL;
 
 	ret = get_user(ctx, ctxp);
@@ -1319,9 +1398,17 @@ SYSCALL_DEFINE6(io_setup2, u32, nr_events, u32, flags, struct iocb __user *,
 	if (IS_ERR(ioctx))
 		goto out;
 
+	if (flags & IOCTX_FLAG_USERIOCB) {
+		ret = aio_useriocb_map(ioctx, iocbs);
+		if (ret)
+			goto err;
+	}
+
 	ret = put_user(ioctx->user_id, ctxp);
-	if (ret)
+	if (ret) {
+err:
 		kill_ioctx(current->mm, ioctx, NULL);
+	}
 	percpu_ref_put(&ioctx->users);
 out:
 	return ret;
@@ -1871,10 +1958,13 @@ static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
 		}
 	}
 
-	ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
-	if (unlikely(ret)) {
-		pr_debug("EFAULT: aio_key\n");
-		goto out_put_req;
+	/* Don't support cancel on user mapped iocbs */
+	if (!(ctx->flags & IOCTX_FLAG_USERIOCB)) {
+		ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
+		if (unlikely(ret)) {
+			pr_debug("EFAULT: aio_key\n");
+			goto out_put_req;
+		}
 	}
 
 	req->ki_user_iocb = user_iocb;
@@ -1928,12 +2018,22 @@ static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
 			 bool compat)
 {
-	struct iocb iocb;
+	struct iocb iocb, *iocbp;
 
-	if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
-		return -EFAULT;
+	if (ctx->flags & IOCTX_FLAG_USERIOCB) {
+		unsigned long iocb_index = (unsigned long) user_iocb;
 
-	return __io_submit_one(ctx, &iocb, user_iocb, compat);
+		if (iocb_index >= ctx->max_reqs)
+			return -EINVAL;
+
+		iocbp = aio_iocb_from_index(ctx, iocb_index);
+	} else {
+		if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
+			return -EFAULT;
+		iocbp = &iocb;
+	}
+
+	return __io_submit_one(ctx, iocbp, user_iocb, compat);
 }
 
 /* sys_io_submit:
@@ -2077,6 +2177,9 @@ SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
 	if (unlikely(!ctx))
 		return -EINVAL;
 
+	if (ctx->flags & IOCTX_FLAG_USERIOCB)
+		goto err;
+
 	spin_lock_irq(&ctx->ctx_lock);
 	kiocb = lookup_kiocb(ctx, iocb);
 	if (kiocb) {
@@ -2093,9 +2196,8 @@ SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
 		 */
 		ret = -EINPROGRESS;
 	}
-
+err:
 	percpu_ref_put(&ctx->users);
-
 	return ret;
 }
 
diff --git a/include/uapi/linux/aio_abi.h b/include/uapi/linux/aio_abi.h
index 8387e0af0f76..814e6606c413 100644
--- a/include/uapi/linux/aio_abi.h
+++ b/include/uapi/linux/aio_abi.h
@@ -106,6 +106,8 @@ struct iocb {
 	__u32	aio_resfd;
 }; /* 64 bytes */
 
+#define IOCTX_FLAG_USERIOCB	(1 << 0)	/* iocbs are user mapped */
+
 #undef IFBIG
 #undef IFLITTLE
 
-- 
2.17.1

[PATCH 15/27] aio: support for IO polling

[Jens Axboe]

Add polled variants of PREAD/PREADV and PWRITE/PWRITEV. These act
like their non-polled counterparts, except we expect to poll for
completion of them. The polling happens at io_getevent() time, and
works just like non-polled IO.

To setup an io_context for polled IO, the application must call
io_setup2() with IOCTX_FLAG_IOPOLL as one of the flags. It is illegal
to mix and match polled and non-polled IO on an io_context.

Polled IO doesn't support the user mapped completion ring. Events
must be reaped through the io_getevents() system call. For non-irq
driven poll devices, there's no way to support completion reaping
from userspace by just looking at the ring. The application itself
is the one that pulls completion entries.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/aio.c                     | 396 +++++++++++++++++++++++++++++++----
 include/uapi/linux/aio_abi.h |   3 +
 2 files changed, 363 insertions(+), 36 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index 4e8c471a2598..c4dbd5e1c350 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -153,6 +153,18 @@ struct kioctx {
 		atomic_t	reqs_available;
 	} ____cacheline_aligned_in_smp;
 
+	/* iopoll submission state */
+	struct {
+		spinlock_t poll_lock;
+		struct list_head poll_submitted;
+	} ____cacheline_aligned_in_smp;
+
+	/* iopoll completion state */
+	struct {
+		struct list_head poll_completing;
+		struct mutex getevents_lock;
+	} ____cacheline_aligned_in_smp;
+
 	struct {
 		spinlock_t	ctx_lock;
 		struct list_head active_reqs;	/* used for cancellation */
@@ -205,14 +217,27 @@ struct aio_kiocb {
 	__u64			ki_user_data;	/* user's data for completion */
 
 	struct list_head	ki_list;	/* the aio core uses this
-						 * for cancellation */
+						 * for cancellation, or for
+						 * polled IO */
+
+	unsigned long		ki_flags;
+#define KIOCB_F_POLL_COMPLETED	0	/* polled IO has completed */
+#define KIOCB_F_POLL_EAGAIN	1	/* polled submission got EAGAIN */
+
 	refcount_t		ki_refcnt;
 
-	/*
-	 * If the aio_resfd field of the userspace iocb is not zero,
-	 * this is the underlying eventfd context to deliver events to.
-	 */
-	struct eventfd_ctx	*ki_eventfd;
+	union {
+		/*
+		 * If the aio_resfd field of the userspace iocb is not zero,
+		 * this is the underlying eventfd context to deliver events to.
+		 */
+		struct eventfd_ctx	*ki_eventfd;
+
+		/*
+		 * For polled IO, stash completion info here
+		 */
+		struct io_event		ki_ev;
+	};
 };
 
 /*------ sysctl variables----*/
@@ -233,6 +258,7 @@ static const unsigned int iocb_page_shift =
 				ilog2(PAGE_SIZE / sizeof(struct iocb));
 
 static void aio_useriocb_unmap(struct kioctx *);
+static void aio_iopoll_reap_events(struct kioctx *);
 
 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
 {
@@ -471,11 +497,15 @@ static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
 	int i;
 	struct file *file;
 
-	/* Compensate for the ring buffer's head/tail overlap entry */
-	nr_events += 2;	/* 1 is required, 2 for good luck */
-
+	/*
+	 * Compensate for the ring buffer's head/tail overlap entry.
+	 * IO polling doesn't require any io event entries
+	 */
 	size = sizeof(struct aio_ring);
-	size += sizeof(struct io_event) * nr_events;
+	if (!(ctx->flags & IOCTX_FLAG_IOPOLL)) {
+		nr_events += 2;	/* 1 is required, 2 for good luck */
+		size += sizeof(struct io_event) * nr_events;
+	}
 
 	nr_pages = PFN_UP(size);
 	if (nr_pages < 0)
@@ -758,6 +788,11 @@ static struct kioctx *io_setup_flags(unsigned long ctxid,
 
 	INIT_LIST_HEAD(&ctx->active_reqs);
 
+	spin_lock_init(&ctx->poll_lock);
+	INIT_LIST_HEAD(&ctx->poll_submitted);
+	INIT_LIST_HEAD(&ctx->poll_completing);
+	mutex_init(&ctx->getevents_lock);
+
 	if (percpu_ref_init(&ctx->users, free_ioctx_users, 0, GFP_KERNEL))
 		goto err;
 
@@ -829,11 +864,15 @@ static int kill_ioctx(struct mm_struct *mm, struct kioctx *ctx,
 {
 	struct kioctx_table *table;
 
+	mutex_lock(&ctx->getevents_lock);
 	spin_lock(&mm->ioctx_lock);
 	if (atomic_xchg(&ctx->dead, 1)) {
 		spin_unlock(&mm->ioctx_lock);
+		mutex_unlock(&ctx->getevents_lock);
 		return -EINVAL;
 	}
+	aio_iopoll_reap_events(ctx);
+	mutex_unlock(&ctx->getevents_lock);
 
 	table = rcu_dereference_raw(mm->ioctx_table);
 	WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
@@ -1042,6 +1081,7 @@ static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
 	percpu_ref_get(&ctx->reqs);
 	req->ki_ctx = ctx;
 	INIT_LIST_HEAD(&req->ki_list);
+	req->ki_flags = 0;
 	refcount_set(&req->ki_refcnt, 0);
 	req->ki_eventfd = NULL;
 	return req;
@@ -1083,6 +1123,15 @@ static inline void iocb_put(struct aio_kiocb *iocb)
 	}
 }
 
+static void iocb_put_many(struct kioctx *ctx, void **iocbs, int *nr)
+{
+	if (*nr) {
+		percpu_ref_put_many(&ctx->reqs, *nr);
+		kmem_cache_free_bulk(kiocb_cachep, *nr, iocbs);
+		*nr = 0;
+	}
+}
+
 static void aio_fill_event(struct io_event *ev, struct aio_kiocb *iocb,
 			   long res, long res2)
 {
@@ -1272,6 +1321,185 @@ static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
 	return ret < 0 || *i >= min_nr;
 }
 
+#define AIO_IOPOLL_BATCH	8
+
+/*
+ * Process completed iocb iopoll entries, copying the result to userspace.
+ */
+static long aio_iopoll_reap(struct kioctx *ctx, struct io_event __user *evs,
+			    unsigned int *nr_events, long max)
+{
+	void *iocbs[AIO_IOPOLL_BATCH];
+	struct aio_kiocb *iocb, *n;
+	int to_free = 0, ret = 0;
+
+	/* Shouldn't happen... */
+	if (*nr_events >= max)
+		return 0;
+
+	list_for_each_entry_safe(iocb, n, &ctx->poll_completing, ki_list) {
+		if (*nr_events == max)
+			break;
+		if (!test_bit(KIOCB_F_POLL_COMPLETED, &iocb->ki_flags))
+			continue;
+		if (to_free == AIO_IOPOLL_BATCH)
+			iocb_put_many(ctx, iocbs, &to_free);
+
+		list_del(&iocb->ki_list);
+		iocbs[to_free++] = iocb;
+
+		fput(iocb->rw.ki_filp);
+
+		if (evs && copy_to_user(evs + *nr_events, &iocb->ki_ev,
+		    sizeof(iocb->ki_ev))) {
+			ret = -EFAULT;
+			break;
+		}
+		(*nr_events)++;
+	}
+
+	if (to_free)
+		iocb_put_many(ctx, iocbs, &to_free);
+
+	return ret;
+}
+
+/*
+ * Poll for a mininum of 'min' events, and a maximum of 'max'. Note that if
+ * min == 0 we consider that a non-spinning poll check - we'll still enter
+ * the driver poll loop, but only as a non-spinning completion check.
+ */
+static int aio_iopoll_getevents(struct kioctx *ctx,
+				struct io_event __user *event,
+				unsigned int *nr_events, long min, long max)
+{
+	struct aio_kiocb *iocb;
+	int to_poll, polled, ret;
+
+	/*
+	 * Check if we already have done events that satisfy what we need
+	 */
+	if (!list_empty(&ctx->poll_completing)) {
+		ret = aio_iopoll_reap(ctx, event, nr_events, max);
+		if (ret < 0)
+			return ret;
+		if ((min && *nr_events >= min) || *nr_events >= max)
+			return 0;
+	}
+
+	/*
+	 * Take in a new working set from the submitted list, if possible.
+	 */
+	if (!list_empty_careful(&ctx->poll_submitted)) {
+		spin_lock(&ctx->poll_lock);
+		list_splice_init(&ctx->poll_submitted, &ctx->poll_completing);
+		spin_unlock(&ctx->poll_lock);
+	}
+
+	if (list_empty(&ctx->poll_completing))
+		return 0;
+
+	/*
+	 * Check again now that we have a new batch.
+	 */
+	ret = aio_iopoll_reap(ctx, event, nr_events, max);
+	if (ret < 0)
+		return ret;
+	if ((min && *nr_events >= min) || *nr_events >= max)
+		return 0;
+
+	/*
+	 * Find up to 'max' worth of events to poll for, including the
+	 * events we already successfully polled
+	 */
+	polled = to_poll = 0;
+	list_for_each_entry(iocb, &ctx->poll_completing, ki_list) {
+		/*
+		 * Poll for needed events with spin == true, anything after
+		 * that we just check if we have more, up to max.
+		 */
+		bool spin = !polled || *nr_events < min;
+		struct kiocb *kiocb = &iocb->rw;
+
+		if (test_bit(KIOCB_F_POLL_COMPLETED, &iocb->ki_flags))
+			break;
+		if (++to_poll + *nr_events > max)
+			break;
+
+		ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
+		if (ret < 0)
+			return ret;
+
+		polled += ret;
+		if (polled + *nr_events >= max)
+			break;
+	}
+
+	ret = aio_iopoll_reap(ctx, event, nr_events, max);
+	if (ret < 0)
+		return ret;
+	if (*nr_events >= min)
+		return 0;
+	return to_poll;
+}
+
+/*
+ * We can't just wait for polled events to come to us, we have to actively
+ * find and complete them.
+ */
+static void aio_iopoll_reap_events(struct kioctx *ctx)
+{
+	if (!(ctx->flags & IOCTX_FLAG_IOPOLL))
+		return;
+
+	while (!list_empty_careful(&ctx->poll_submitted) ||
+	       !list_empty(&ctx->poll_completing)) {
+		unsigned int nr_events = 0;
+
+		aio_iopoll_getevents(ctx, NULL, &nr_events, 1, UINT_MAX);
+	}
+}
+
+static int __aio_iopoll_check(struct kioctx *ctx, struct io_event __user *event,
+			      unsigned int *nr_events, long min_nr, long max_nr)
+{
+	int ret = 0;
+
+	while (!*nr_events || !need_resched()) {
+		int tmin = 0;
+
+		if (*nr_events < min_nr)
+			tmin = min_nr - *nr_events;
+
+		ret = aio_iopoll_getevents(ctx, event, nr_events, tmin, max_nr);
+		if (ret <= 0)
+			break;
+		ret = 0;
+	}
+
+	return ret;
+}
+
+static int aio_iopoll_check(struct kioctx *ctx, long min_nr, long nr,
+			    struct io_event __user *event)
+{
+	unsigned int nr_events = 0;
+	int ret;
+
+	/* Only allow one thread polling at a time */
+	if (!mutex_trylock(&ctx->getevents_lock))
+		return -EBUSY;
+	if (unlikely(atomic_read(&ctx->dead))) {
+		ret = -EINVAL;
+		goto err;
+	}
+
+	ret = __aio_iopoll_check(ctx, event, &nr_events, min_nr, nr);
+err:
+	mutex_unlock(&ctx->getevents_lock);
+	return nr_events ? nr_events : ret;
+}
+
 static long read_events(struct kioctx *ctx, long min_nr, long nr,
 			struct io_event __user *event,
 			ktime_t until)
@@ -1386,7 +1614,7 @@ SYSCALL_DEFINE6(io_setup2, u32, nr_events, u32, flags, struct iocb __user *,
 
 	if (user1 || user2)
 		return -EINVAL;
-	if (flags & ~IOCTX_FLAG_USERIOCB)
+	if (flags & ~(IOCTX_FLAG_USERIOCB | IOCTX_FLAG_IOPOLL))
 		return -EINVAL;
 
 	ret = get_user(ctx, ctxp);
@@ -1520,13 +1748,8 @@ static void aio_remove_iocb(struct aio_kiocb *iocb)
 	spin_unlock_irqrestore(&ctx->ctx_lock, flags);
 }
 
-static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
+static void kiocb_end_write(struct kiocb *kiocb)
 {
-	struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
-
-	if (!list_empty_careful(&iocb->ki_list))
-		aio_remove_iocb(iocb);
-
 	if (kiocb->ki_flags & IOCB_WRITE) {
 		struct inode *inode = file_inode(kiocb->ki_filp);
 
@@ -1538,19 +1761,48 @@ static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
 			__sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
 		file_end_write(kiocb->ki_filp);
 	}
+}
+
+static void aio_complete_rw(struct kiocb *kiocb, long res, long res2)
+{
+	struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
+
+	if (!list_empty_careful(&iocb->ki_list))
+		aio_remove_iocb(iocb);
+
+	kiocb_end_write(kiocb);
 
 	fput(kiocb->ki_filp);
 	aio_complete(iocb, res, res2);
 }
 
-static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
+static void aio_complete_rw_poll(struct kiocb *kiocb, long res, long res2)
 {
+	struct aio_kiocb *iocb = container_of(kiocb, struct aio_kiocb, rw);
+
+	kiocb_end_write(kiocb);
+
+	/*
+	 * Handle EAGAIN from resource limits with polled IO inline, don't
+	 * pass the event back to userspace.
+	 */
+	if (unlikely(res == -EAGAIN))
+		set_bit(KIOCB_F_POLL_EAGAIN, &iocb->ki_flags);
+	else {
+		aio_fill_event(&iocb->ki_ev, iocb, res, res2);
+		set_bit(KIOCB_F_POLL_COMPLETED, &iocb->ki_flags);
+	}
+}
+
+static int aio_prep_rw(struct aio_kiocb *kiocb, const struct iocb *iocb)
+{
+	struct kioctx *ctx = kiocb->ki_ctx;
+	struct kiocb *req = &kiocb->rw;
 	int ret;
 
 	req->ki_filp = fget(iocb->aio_fildes);
 	if (unlikely(!req->ki_filp))
 		return -EBADF;
-	req->ki_complete = aio_complete_rw;
 	req->ki_pos = iocb->aio_offset;
 	req->ki_flags = iocb_flags(req->ki_filp);
 	if (iocb->aio_flags & IOCB_FLAG_RESFD)
@@ -1576,9 +1828,35 @@ static int aio_prep_rw(struct kiocb *req, const struct iocb *iocb)
 	if (unlikely(ret))
 		goto out_fput;
 
-	req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
-	return 0;
+	if (iocb->aio_flags & IOCB_FLAG_HIPRI) {
+		/* shares space in the union, and is rather pointless.. */
+		ret = -EINVAL;
+		if (iocb->aio_flags & IOCB_FLAG_RESFD)
+			goto out_fput;
+
+		/* can't submit polled IO to a non-polled ctx */
+		if (!(ctx->flags & IOCTX_FLAG_IOPOLL))
+			goto out_fput;
+
+		ret = -EOPNOTSUPP;
+		if (!(req->ki_flags & IOCB_DIRECT) ||
+		    !req->ki_filp->f_op->iopoll)
+			goto out_fput;
+
+		req->ki_flags |= IOCB_HIPRI;
+		req->ki_complete = aio_complete_rw_poll;
+	} else {
+		/* can't submit non-polled IO to a polled ctx */
+		ret = -EINVAL;
+		if (ctx->flags & IOCTX_FLAG_IOPOLL)
+			goto out_fput;
 
+		/* no one is going to poll for this I/O */
+		req->ki_flags &= ~IOCB_HIPRI;
+		req->ki_complete = aio_complete_rw;
+	}
+
+	return 0;
 out_fput:
 	fput(req->ki_filp);
 	return ret;
@@ -1623,15 +1901,40 @@ static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
 	}
 }
 
-static ssize_t aio_read(struct kiocb *req, const struct iocb *iocb,
+/*
+ * After the iocb has been issued, it's safe to be found on the poll list.
+ * Adding the kiocb to the list AFTER submission ensures that we don't
+ * find it from a io_getevents() thread before the issuer is done accessing
+ * the kiocb cookie.
+ */
+static void aio_iopoll_iocb_issued(struct aio_kiocb *kiocb)
+{
+	/*
+	 * For fast devices, IO may have already completed. If it has, add
+	 * it to the front so we find it first. We can't add to the poll_done
+	 * list as that's unlocked from the completion side.
+	 */
+	const int front = test_bit(KIOCB_F_POLL_COMPLETED, &kiocb->ki_flags);
+	struct kioctx *ctx = kiocb->ki_ctx;
+
+	spin_lock(&ctx->poll_lock);
+	if (front)
+		list_add(&kiocb->ki_list, &ctx->poll_submitted);
+	else
+		list_add_tail(&kiocb->ki_list, &ctx->poll_submitted);
+	spin_unlock(&ctx->poll_lock);
+}
+
+static ssize_t aio_read(struct aio_kiocb *kiocb, const struct iocb *iocb,
 			bool vectored, bool compat)
 {
 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
+	struct kiocb *req = &kiocb->rw;
 	struct iov_iter iter;
 	struct file *file;
 	ssize_t ret;
 
-	ret = aio_prep_rw(req, iocb);
+	ret = aio_prep_rw(kiocb, iocb);
 	if (ret)
 		return ret;
 	file = req->ki_filp;
@@ -1656,15 +1959,16 @@ static ssize_t aio_read(struct kiocb *req, const struct iocb *iocb,
 	return ret;
 }
 
-static ssize_t aio_write(struct kiocb *req, const struct iocb *iocb,
+static ssize_t aio_write(struct aio_kiocb *kiocb, const struct iocb *iocb,
 			 bool vectored, bool compat)
 {
 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
+	struct kiocb *req = &kiocb->rw;
 	struct iov_iter iter;
 	struct file *file;
 	ssize_t ret;
 
-	ret = aio_prep_rw(req, iocb);
+	ret = aio_prep_rw(kiocb, iocb);
 	if (ret)
 		return ret;
 	file = req->ki_filp;
@@ -1935,7 +2239,8 @@ static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
 		return -EINVAL;
 	}
 
-	if (!get_reqs_available(ctx))
+	/* Poll IO doesn't need ring reservations */
+	if (!(ctx->flags & IOCTX_FLAG_IOPOLL) && !get_reqs_available(ctx))
 		return -EAGAIN;
 
 	ret = -EAGAIN;
@@ -1958,8 +2263,8 @@ static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
 		}
 	}
 
-	/* Don't support cancel on user mapped iocbs */
-	if (!(ctx->flags & IOCTX_FLAG_USERIOCB)) {
+	/* Don't support cancel on user mapped iocbs or polled context */
+	if (!(ctx->flags & (IOCTX_FLAG_USERIOCB | IOCTX_FLAG_IOPOLL))) {
 		ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
 		if (unlikely(ret)) {
 			pr_debug("EFAULT: aio_key\n");
@@ -1970,26 +2275,33 @@ static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
 	req->ki_user_iocb = user_iocb;
 	req->ki_user_data = iocb->aio_data;
 
+	ret = -EINVAL;
 	switch (iocb->aio_lio_opcode) {
 	case IOCB_CMD_PREAD:
-		ret = aio_read(&req->rw, iocb, false, compat);
+		ret = aio_read(req, iocb, false, compat);
 		break;
 	case IOCB_CMD_PWRITE:
-		ret = aio_write(&req->rw, iocb, false, compat);
+		ret = aio_write(req, iocb, false, compat);
 		break;
 	case IOCB_CMD_PREADV:
-		ret = aio_read(&req->rw, iocb, true, compat);
+		ret = aio_read(req, iocb, true, compat);
 		break;
 	case IOCB_CMD_PWRITEV:
-		ret = aio_write(&req->rw, iocb, true, compat);
+		ret = aio_write(req, iocb, true, compat);
 		break;
 	case IOCB_CMD_FSYNC:
+		if (ctx->flags & IOCTX_FLAG_IOPOLL)
+			break;
 		ret = aio_fsync(&req->fsync, iocb, false);
 		break;
 	case IOCB_CMD_FDSYNC:
+		if (ctx->flags & IOCTX_FLAG_IOPOLL)
+			break;
 		ret = aio_fsync(&req->fsync, iocb, true);
 		break;
 	case IOCB_CMD_POLL:
+		if (ctx->flags & IOCTX_FLAG_IOPOLL)
+			break;
 		ret = aio_poll(req, iocb);
 		break;
 	default:
@@ -2005,13 +2317,21 @@ static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
 	 */
 	if (ret)
 		goto out_put_req;
+	if (ctx->flags & IOCTX_FLAG_IOPOLL) {
+		if (test_bit(KIOCB_F_POLL_EAGAIN, &req->ki_flags)) {
+			ret = -EAGAIN;
+			goto out_put_req;
+		}
+		aio_iopoll_iocb_issued(req);
+	}
 	return 0;
 out_put_req:
 	if (req->ki_eventfd)
 		eventfd_ctx_put(req->ki_eventfd);
 	iocb_put(req);
 out_put_reqs_available:
-	put_reqs_available(ctx, 1);
+	if (!(ctx->flags & IOCTX_FLAG_IOPOLL))
+		put_reqs_available(ctx, 1);
 	return ret;
 }
 
@@ -2177,7 +2497,7 @@ SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
 	if (unlikely(!ctx))
 		return -EINVAL;
 
-	if (ctx->flags & IOCTX_FLAG_USERIOCB)
+	if (ctx->flags & (IOCTX_FLAG_USERIOCB | IOCTX_FLAG_IOPOLL))
 		goto err;
 
 	spin_lock_irq(&ctx->ctx_lock);
@@ -2212,8 +2532,12 @@ static long do_io_getevents(aio_context_t ctx_id,
 	long ret = -EINVAL;
 
 	if (likely(ioctx)) {
-		if (likely(min_nr <= nr && min_nr >= 0))
-			ret = read_events(ioctx, min_nr, nr, events, until);
+		if (likely(min_nr <= nr && min_nr >= 0)) {
+			if (ioctx->flags & IOCTX_FLAG_IOPOLL)
+				ret = aio_iopoll_check(ioctx, min_nr, nr, events);
+			else
+				ret = read_events(ioctx, min_nr, nr, events, until);
+		}
 		percpu_ref_put(&ioctx->users);
 	}
 
diff --git a/include/uapi/linux/aio_abi.h b/include/uapi/linux/aio_abi.h
index 814e6606c413..ea0b9a19f4df 100644
--- a/include/uapi/linux/aio_abi.h
+++ b/include/uapi/linux/aio_abi.h
@@ -52,9 +52,11 @@ enum {
  *                   is valid.
  * IOCB_FLAG_IOPRIO - Set if the "aio_reqprio" member of the "struct iocb"
  *                    is valid.
+ * IOCB_FLAG_HIPRI - Use IO completion polling
  */
 #define IOCB_FLAG_RESFD		(1 << 0)
 #define IOCB_FLAG_IOPRIO	(1 << 1)
+#define IOCB_FLAG_HIPRI		(1 << 2)
 
 /* read() from /dev/aio returns these structures. */
 struct io_event {
@@ -107,6 +109,7 @@ struct iocb {
 }; /* 64 bytes */
 
 #define IOCTX_FLAG_USERIOCB	(1 << 0)	/* iocbs are user mapped */
+#define IOCTX_FLAG_IOPOLL	(1 << 1)	/* io_context is polled */
 
 #undef IFBIG
 #undef IFLITTLE
-- 
2.17.1

[PATCH 16/27] aio: add submission side request cache

[Jens Axboe]

We have to add each submitted polled request to the io_context
poll_submitted list, which means we have to grab the poll_lock. We
already use the block plug to batch submissions if we're doing a batch
of IO submissions, extend that to cover the poll requests internally as
well.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/aio.c | 136 +++++++++++++++++++++++++++++++++++++++++++++----------
 1 file changed, 113 insertions(+), 23 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index c4dbd5e1c350..2e8cde976cb4 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -240,6 +240,21 @@ struct aio_kiocb {
 	};
 };
 
+struct aio_submit_state {
+	struct kioctx *ctx;
+
+	struct blk_plug plug;
+#ifdef CONFIG_BLOCK
+	struct blk_plug_cb plug_cb;
+#endif
+
+	/*
+	 * Polled iocbs that have been submitted, but not added to the ctx yet
+	 */
+	struct list_head req_list;
+	unsigned int req_count;
+};
+
 /*------ sysctl variables----*/
 static DEFINE_SPINLOCK(aio_nr_lock);
 unsigned long aio_nr;		/* current system wide number of aio requests */
@@ -257,6 +272,15 @@ static const struct address_space_operations aio_ctx_aops;
 static const unsigned int iocb_page_shift =
 				ilog2(PAGE_SIZE / sizeof(struct iocb));
 
+/*
+ * We rely on block level unplugs to flush pending requests, if we schedule
+ */
+#ifdef CONFIG_BLOCK
+static const bool aio_use_state_req_list = true;
+#else
+static const bool aio_use_state_req_list = false;
+#endif
+
 static void aio_useriocb_unmap(struct kioctx *);
 static void aio_iopoll_reap_events(struct kioctx *);
 
@@ -1901,13 +1925,28 @@ static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
 	}
 }
 
+/*
+ * Called either at the end of IO submission, or through a plug callback
+ * because we're going to schedule. Moves out local batch of requests to
+ * the ctx poll list, so they can be found for polling + reaping.
+ */
+static void aio_flush_state_reqs(struct kioctx *ctx,
+				 struct aio_submit_state *state)
+{
+	spin_lock(&ctx->poll_lock);
+	list_splice_tail_init(&state->req_list, &ctx->poll_submitted);
+	spin_unlock(&ctx->poll_lock);
+	state->req_count = 0;
+}
+
 /*
  * After the iocb has been issued, it's safe to be found on the poll list.
  * Adding the kiocb to the list AFTER submission ensures that we don't
  * find it from a io_getevents() thread before the issuer is done accessing
  * the kiocb cookie.
  */
-static void aio_iopoll_iocb_issued(struct aio_kiocb *kiocb)
+static void aio_iopoll_iocb_issued(struct aio_submit_state *state,
+				   struct aio_kiocb *kiocb)
 {
 	/*
 	 * For fast devices, IO may have already completed. If it has, add
@@ -1917,12 +1956,21 @@ static void aio_iopoll_iocb_issued(struct aio_kiocb *kiocb)
 	const int front = test_bit(KIOCB_F_POLL_COMPLETED, &kiocb->ki_flags);
 	struct kioctx *ctx = kiocb->ki_ctx;
 
-	spin_lock(&ctx->poll_lock);
-	if (front)
-		list_add(&kiocb->ki_list, &ctx->poll_submitted);
-	else
-		list_add_tail(&kiocb->ki_list, &ctx->poll_submitted);
-	spin_unlock(&ctx->poll_lock);
+	if (!state || !aio_use_state_req_list) {
+		spin_lock(&ctx->poll_lock);
+		if (front)
+			list_add(&kiocb->ki_list, &ctx->poll_submitted);
+		else
+			list_add_tail(&kiocb->ki_list, &ctx->poll_submitted);
+		spin_unlock(&ctx->poll_lock);
+	} else {
+		if (front)
+			list_add(&kiocb->ki_list, &state->req_list);
+		else
+			list_add_tail(&kiocb->ki_list, &state->req_list);
+		if (++state->req_count >= AIO_IOPOLL_BATCH)
+			aio_flush_state_reqs(ctx, state);
+	}
 }
 
 static ssize_t aio_read(struct aio_kiocb *kiocb, const struct iocb *iocb,
@@ -2218,7 +2266,8 @@ static ssize_t aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
 }
 
 static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
-			   struct iocb __user *user_iocb, bool compat)
+			   struct iocb __user *user_iocb,
+			   struct aio_submit_state *state, bool compat)
 {
 	struct aio_kiocb *req;
 	ssize_t ret;
@@ -2322,7 +2371,7 @@ static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
 			ret = -EAGAIN;
 			goto out_put_req;
 		}
-		aio_iopoll_iocb_issued(req);
+		aio_iopoll_iocb_issued(state, req);
 	}
 	return 0;
 out_put_req:
@@ -2336,7 +2385,7 @@ static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
 }
 
 static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
-			 bool compat)
+			 struct aio_submit_state *state, bool compat)
 {
 	struct iocb iocb, *iocbp;
 
@@ -2353,7 +2402,44 @@ static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
 		iocbp = &iocb;
 	}
 
-	return __io_submit_one(ctx, iocbp, user_iocb, compat);
+	return __io_submit_one(ctx, iocbp, user_iocb, state, compat);
+}
+
+#ifdef CONFIG_BLOCK
+static void aio_state_unplug(struct blk_plug_cb *cb, bool from_schedule)
+{
+	struct aio_submit_state *state;
+
+	state = container_of(cb, struct aio_submit_state, plug_cb);
+	if (!list_empty(&state->req_list))
+		aio_flush_state_reqs(state->ctx, state);
+}
+#endif
+
+/*
+ * Batched submission is done, ensure local IO is flushed out.
+ */
+static void aio_submit_state_end(struct aio_submit_state *state)
+{
+	blk_finish_plug(&state->plug);
+	if (!list_empty(&state->req_list))
+		aio_flush_state_reqs(state->ctx, state);
+}
+
+/*
+ * Start submission side cache.
+ */
+static void aio_submit_state_start(struct aio_submit_state *state,
+				   struct kioctx *ctx)
+{
+	state->ctx = ctx;
+	INIT_LIST_HEAD(&state->req_list);
+	state->req_count = 0;
+#ifdef CONFIG_BLOCK
+	state->plug_cb.callback = aio_state_unplug;
+	blk_start_plug(&state->plug);
+	list_add(&state->plug_cb.list, &state->plug.cb_list);
+#endif
 }
 
 /* sys_io_submit:
@@ -2371,10 +2457,10 @@ static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
 SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
 		struct iocb __user * __user *, iocbpp)
 {
+	struct aio_submit_state state, *statep = NULL;
 	struct kioctx *ctx;
 	long ret = 0;
 	int i = 0;
-	struct blk_plug plug;
 
 	if (unlikely(nr < 0))
 		return -EINVAL;
@@ -2388,8 +2474,10 @@ SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
 	if (nr > ctx->nr_events)
 		nr = ctx->nr_events;
 
-	if (nr > AIO_PLUG_THRESHOLD)
-		blk_start_plug(&plug);
+	if (nr > AIO_PLUG_THRESHOLD) {
+		aio_submit_state_start(&state, ctx);
+		statep = &state;
+	}
 	for (i = 0; i < nr; i++) {
 		struct iocb __user *user_iocb;
 
@@ -2398,12 +2486,12 @@ SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
 			break;
 		}
 
-		ret = io_submit_one(ctx, user_iocb, false);
+		ret = io_submit_one(ctx, user_iocb, statep, false);
 		if (ret)
 			break;
 	}
-	if (nr > AIO_PLUG_THRESHOLD)
-		blk_finish_plug(&plug);
+	if (statep)
+		aio_submit_state_end(statep);
 
 	percpu_ref_put(&ctx->users);
 	return i ? i : ret;
@@ -2413,10 +2501,10 @@ SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
 COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
 		       int, nr, compat_uptr_t __user *, iocbpp)
 {
+	struct aio_submit_state state, *statep = NULL;
 	struct kioctx *ctx;
 	long ret = 0;
 	int i = 0;
-	struct blk_plug plug;
 
 	if (unlikely(nr < 0))
 		return -EINVAL;
@@ -2430,8 +2518,10 @@ COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
 	if (nr > ctx->nr_events)
 		nr = ctx->nr_events;
 
-	if (nr > AIO_PLUG_THRESHOLD)
-		blk_start_plug(&plug);
+	if (nr > AIO_PLUG_THRESHOLD) {
+		aio_submit_state_start(&state, ctx);
+		statep = &state;
+	}
 	for (i = 0; i < nr; i++) {
 		compat_uptr_t user_iocb;
 
@@ -2440,12 +2530,12 @@ COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
 			break;
 		}
 
-		ret = io_submit_one(ctx, compat_ptr(user_iocb), true);
+		ret = io_submit_one(ctx, compat_ptr(user_iocb), statep, true);
 		if (ret)
 			break;
 	}
-	if (nr > AIO_PLUG_THRESHOLD)
-		blk_finish_plug(&plug);
+	if (statep)
+		aio_submit_state_end(statep);
 
 	percpu_ref_put(&ctx->users);
 	return i ? i : ret;
-- 
2.17.1

[PATCH 17/27] fs: add fget_many() and fput_many()

[Jens Axboe]

Some uses cases repeatedly get and put references to the same file, but
the only exposed interface is doing these one at the time. As each of
these entail an atomic inc or dec on a shared structure, that cost can
add up.

Add fget_many(), which works just like fget(), except it takes an
argument for how many references to get on the file. Ditto fput_many(),
which can drop an arbitrary number of references to a file.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/file.c            | 15 ++++++++++-----
 fs/file_table.c      | 10 ++++++++--
 include/linux/file.h |  2 ++
 include/linux/fs.h   |  3 ++-
 4 files changed, 22 insertions(+), 8 deletions(-)

diff --git a/fs/file.c b/fs/file.c
index 7ffd6e9d103d..ad9870edfd51 100644
--- a/fs/file.c
+++ b/fs/file.c
@@ -676,7 +676,7 @@ void do_close_on_exec(struct files_struct *files)
 	spin_unlock(&files->file_lock);
 }
 
-static struct file *__fget(unsigned int fd, fmode_t mask)
+static struct file *__fget(unsigned int fd, fmode_t mask, unsigned int refs)
 {
 	struct files_struct *files = current->files;
 	struct file *file;
@@ -691,7 +691,7 @@ static struct file *__fget(unsigned int fd, fmode_t mask)
 		 */
 		if (file->f_mode & mask)
 			file = NULL;
-		else if (!get_file_rcu(file))
+		else if (!get_file_rcu_many(file, refs))
 			goto loop;
 	}
 	rcu_read_unlock();
@@ -699,15 +699,20 @@ static struct file *__fget(unsigned int fd, fmode_t mask)
 	return file;
 }
 
+struct file *fget_many(unsigned int fd, unsigned int refs)
+{
+	return __fget(fd, FMODE_PATH, refs);
+}
+
 struct file *fget(unsigned int fd)
 {
-	return __fget(fd, FMODE_PATH);
+	return fget_many(fd, 1);
 }
 EXPORT_SYMBOL(fget);
 
 struct file *fget_raw(unsigned int fd)
 {
-	return __fget(fd, 0);
+	return __fget(fd, 0, 1);
 }
 EXPORT_SYMBOL(fget_raw);
 
@@ -738,7 +743,7 @@ static unsigned long __fget_light(unsigned int fd, fmode_t mask)
 			return 0;
 		return (unsigned long)file;
 	} else {
-		file = __fget(fd, mask);
+		file = __fget(fd, mask, 1);
 		if (!file)
 			return 0;
 		return FDPUT_FPUT | (unsigned long)file;
diff --git a/fs/file_table.c b/fs/file_table.c
index e49af4caf15d..6a3964df33e4 100644
--- a/fs/file_table.c
+++ b/fs/file_table.c
@@ -326,9 +326,9 @@ void flush_delayed_fput(void)
 
 static DECLARE_DELAYED_WORK(delayed_fput_work, delayed_fput);
 
-void fput(struct file *file)
+void fput_many(struct file *file, unsigned int refs)
 {
-	if (atomic_long_dec_and_test(&file->f_count)) {
+	if (atomic_long_sub_and_test(refs, &file->f_count)) {
 		struct task_struct *task = current;
 
 		if (likely(!in_interrupt() && !(task->flags & PF_KTHREAD))) {
@@ -347,6 +347,12 @@ void fput(struct file *file)
 	}
 }
 
+void fput(struct file *file)
+{
+	fput_many(file, 1);
+}
+
+
 /*
  * synchronous analog of fput(); for kernel threads that might be needed
  * in some umount() (and thus can't use flush_delayed_fput() without
diff --git a/include/linux/file.h b/include/linux/file.h
index 6b2fb032416c..3fcddff56bc4 100644
--- a/include/linux/file.h
+++ b/include/linux/file.h
@@ -13,6 +13,7 @@
 struct file;
 
 extern void fput(struct file *);
+extern void fput_many(struct file *, unsigned int);
 
 struct file_operations;
 struct vfsmount;
@@ -44,6 +45,7 @@ static inline void fdput(struct fd fd)
 }
 
 extern struct file *fget(unsigned int fd);
+extern struct file *fget_many(unsigned int fd, unsigned int refs);
 extern struct file *fget_raw(unsigned int fd);
 extern unsigned long __fdget(unsigned int fd);
 extern unsigned long __fdget_raw(unsigned int fd);
diff --git a/include/linux/fs.h b/include/linux/fs.h
index 6a5f71f8ae06..dc54a65c401a 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -952,7 +952,8 @@ static inline struct file *get_file(struct file *f)
 	atomic_long_inc(&f->f_count);
 	return f;
 }
-#define get_file_rcu(x) atomic_long_inc_not_zero(&(x)->f_count)
+#define get_file_rcu_many(x, cnt) atomic_long_add_unless(&(x)->f_count, (cnt), 0)
+#define get_file_rcu(x) get_file_rcu_many((x), 1)
 #define fput_atomic(x)	atomic_long_add_unless(&(x)->f_count, -1, 1)
 #define file_count(x)	atomic_long_read(&(x)->f_count)
 
-- 
2.17.1

[PATCH 18/27] aio: use fget/fput_many() for file references

[Jens Axboe]

On the submission side, add file reference batching to the
aio_submit_state. We get as many references as the number of iocbs we
are submitting, and drop unused ones if we end up switching files. The
assumption here is that we're usually only dealing with one fd, and if
there are multiple, hopefuly they are at least somewhat ordered. Could
trivially be extended to cover multiple fds, if needed.

On the completion side we do the same thing, except this is trivially
done just locally in aio_iopoll_reap().

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/aio.c | 106 +++++++++++++++++++++++++++++++++++++++++++++++--------
 1 file changed, 91 insertions(+), 15 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index 2e8cde976cb4..6cbfe9905637 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -253,6 +253,15 @@ struct aio_submit_state {
 	 */
 	struct list_head req_list;
 	unsigned int req_count;
+
+	/*
+	 * File reference cache
+	 */
+	struct file *file;
+	unsigned int fd;
+	unsigned int has_refs;
+	unsigned int used_refs;
+	unsigned int ios_left;
 };
 
 /*------ sysctl variables----*/
@@ -1355,7 +1364,8 @@ static long aio_iopoll_reap(struct kioctx *ctx, struct io_event __user *evs,
 {
 	void *iocbs[AIO_IOPOLL_BATCH];
 	struct aio_kiocb *iocb, *n;
-	int to_free = 0, ret = 0;
+	int file_count, to_free = 0, ret = 0;
+	struct file *file = NULL;
 
 	/* Shouldn't happen... */
 	if (*nr_events >= max)
@@ -1372,7 +1382,20 @@ static long aio_iopoll_reap(struct kioctx *ctx, struct io_event __user *evs,
 		list_del(&iocb->ki_list);
 		iocbs[to_free++] = iocb;
 
-		fput(iocb->rw.ki_filp);
+		/*
+		 * Batched puts of the same file, to avoid dirtying the
+		 * file usage count multiple times, if avoidable.
+		 */
+		if (!file) {
+			file = iocb->rw.ki_filp;
+			file_count = 1;
+		} else if (file == iocb->rw.ki_filp) {
+			file_count++;
+		} else {
+			fput_many(file, file_count);
+			file = iocb->rw.ki_filp;
+			file_count = 1;
+		}
 
 		if (evs && copy_to_user(evs + *nr_events, &iocb->ki_ev,
 		    sizeof(iocb->ki_ev))) {
@@ -1382,6 +1405,9 @@ static long aio_iopoll_reap(struct kioctx *ctx, struct io_event __user *evs,
 		(*nr_events)++;
 	}
 
+	if (file)
+		fput_many(file, file_count);
+
 	if (to_free)
 		iocb_put_many(ctx, iocbs, &to_free);
 
@@ -1818,13 +1844,58 @@ static void aio_complete_rw_poll(struct kiocb *kiocb, long res, long res2)
 	}
 }
 
-static int aio_prep_rw(struct aio_kiocb *kiocb, const struct iocb *iocb)
+static void aio_file_put(struct aio_submit_state *state)
+{
+	if (state->file) {
+		int diff = state->has_refs - state->used_refs;
+
+		if (diff)
+			fput_many(state->file, diff);
+		state->file = NULL;
+	}
+}
+
+/*
+ * Get as many references to a file as we have IOs left in this submission,
+ * assuming most submissions are for one file, or at least that each file
+ * has more than one submission.
+ */
+static struct file *aio_file_get(struct aio_submit_state *state, int fd)
+{
+	if (!state)
+		return fget(fd);
+
+	if (!state->file) {
+get_file:
+		state->file = fget_many(fd, state->ios_left);
+		if (!state->file)
+			return NULL;
+
+		state->fd = fd;
+		state->has_refs = state->ios_left;
+		state->used_refs = 1;
+		state->ios_left--;
+		return state->file;
+	}
+
+	if (state->fd == fd) {
+		state->used_refs++;
+		state->ios_left--;
+		return state->file;
+	}
+
+	aio_file_put(state);
+	goto get_file;
+}
+
+static int aio_prep_rw(struct aio_kiocb *kiocb, const struct iocb *iocb,
+		       struct aio_submit_state *state)
 {
 	struct kioctx *ctx = kiocb->ki_ctx;
 	struct kiocb *req = &kiocb->rw;
 	int ret;
 
-	req->ki_filp = fget(iocb->aio_fildes);
+	req->ki_filp = aio_file_get(state, iocb->aio_fildes);
 	if (unlikely(!req->ki_filp))
 		return -EBADF;
 	req->ki_pos = iocb->aio_offset;
@@ -1974,7 +2045,8 @@ static void aio_iopoll_iocb_issued(struct aio_submit_state *state,
 }
 
 static ssize_t aio_read(struct aio_kiocb *kiocb, const struct iocb *iocb,
-			bool vectored, bool compat)
+			struct aio_submit_state *state, bool vectored,
+			bool compat)
 {
 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
 	struct kiocb *req = &kiocb->rw;
@@ -1982,7 +2054,7 @@ static ssize_t aio_read(struct aio_kiocb *kiocb, const struct iocb *iocb,
 	struct file *file;
 	ssize_t ret;
 
-	ret = aio_prep_rw(kiocb, iocb);
+	ret = aio_prep_rw(kiocb, iocb, state);
 	if (ret)
 		return ret;
 	file = req->ki_filp;
@@ -2008,7 +2080,8 @@ static ssize_t aio_read(struct aio_kiocb *kiocb, const struct iocb *iocb,
 }
 
 static ssize_t aio_write(struct aio_kiocb *kiocb, const struct iocb *iocb,
-			 bool vectored, bool compat)
+			 struct aio_submit_state *state, bool vectored,
+			 bool compat)
 {
 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
 	struct kiocb *req = &kiocb->rw;
@@ -2016,7 +2089,7 @@ static ssize_t aio_write(struct aio_kiocb *kiocb, const struct iocb *iocb,
 	struct file *file;
 	ssize_t ret;
 
-	ret = aio_prep_rw(kiocb, iocb);
+	ret = aio_prep_rw(kiocb, iocb, state);
 	if (ret)
 		return ret;
 	file = req->ki_filp;
@@ -2327,16 +2400,16 @@ static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
 	ret = -EINVAL;
 	switch (iocb->aio_lio_opcode) {
 	case IOCB_CMD_PREAD:
-		ret = aio_read(req, iocb, false, compat);
+		ret = aio_read(req, iocb, state, false, compat);
 		break;
 	case IOCB_CMD_PWRITE:
-		ret = aio_write(req, iocb, false, compat);
+		ret = aio_write(req, iocb, state, false, compat);
 		break;
 	case IOCB_CMD_PREADV:
-		ret = aio_read(req, iocb, true, compat);
+		ret = aio_read(req, iocb, state, true, compat);
 		break;
 	case IOCB_CMD_PWRITEV:
-		ret = aio_write(req, iocb, true, compat);
+		ret = aio_write(req, iocb, state, true, compat);
 		break;
 	case IOCB_CMD_FSYNC:
 		if (ctx->flags & IOCTX_FLAG_IOPOLL)
@@ -2424,17 +2497,20 @@ static void aio_submit_state_end(struct aio_submit_state *state)
 	blk_finish_plug(&state->plug);
 	if (!list_empty(&state->req_list))
 		aio_flush_state_reqs(state->ctx, state);
+	aio_file_put(state);
 }
 
 /*
  * Start submission side cache.
  */
 static void aio_submit_state_start(struct aio_submit_state *state,
-				   struct kioctx *ctx)
+				   struct kioctx *ctx, int max_ios)
 {
 	state->ctx = ctx;
 	INIT_LIST_HEAD(&state->req_list);
 	state->req_count = 0;
+	state->file = NULL;
+	state->ios_left = max_ios;
 #ifdef CONFIG_BLOCK
 	state->plug_cb.callback = aio_state_unplug;
 	blk_start_plug(&state->plug);
@@ -2475,7 +2551,7 @@ SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
 		nr = ctx->nr_events;
 
 	if (nr > AIO_PLUG_THRESHOLD) {
-		aio_submit_state_start(&state, ctx);
+		aio_submit_state_start(&state, ctx, nr);
 		statep = &state;
 	}
 	for (i = 0; i < nr; i++) {
@@ -2519,7 +2595,7 @@ COMPAT_SYSCALL_DEFINE3(io_submit, compat_aio_context_t, ctx_id,
 		nr = ctx->nr_events;
 
 	if (nr > AIO_PLUG_THRESHOLD) {
-		aio_submit_state_start(&state, ctx);
+		aio_submit_state_start(&state, ctx, nr);
 		statep = &state;
 	}
 	for (i = 0; i < nr; i++) {
-- 
2.17.1

[PATCH 19/27] aio: split iocb init from allocation

[Jens Axboe]

In preparation from having pre-allocated requests, that we then just
need to initialize before use.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/aio.c | 17 +++++++++++------
 1 file changed, 11 insertions(+), 6 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index 6cbfe9905637..e9dbaedda7ae 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -1099,6 +1099,16 @@ static bool get_reqs_available(struct kioctx *ctx)
 	return __get_reqs_available(ctx);
 }
 
+static void aio_iocb_init(struct kioctx *ctx, struct aio_kiocb *req)
+{
+	percpu_ref_get(&ctx->reqs);
+	req->ki_ctx = ctx;
+	INIT_LIST_HEAD(&req->ki_list);
+	req->ki_flags = 0;
+	refcount_set(&req->ki_refcnt, 0);
+	req->ki_eventfd = NULL;
+}
+
 /* aio_get_req
  *	Allocate a slot for an aio request.
  * Returns NULL if no requests are free.
@@ -1111,12 +1121,7 @@ static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
 	if (unlikely(!req))
 		return NULL;
 
-	percpu_ref_get(&ctx->reqs);
-	req->ki_ctx = ctx;
-	INIT_LIST_HEAD(&req->ki_list);
-	req->ki_flags = 0;
-	refcount_set(&req->ki_refcnt, 0);
-	req->ki_eventfd = NULL;
+	aio_iocb_init(ctx, req);
 	return req;
 }
 
-- 
2.17.1

[PATCH 20/27] aio: batch aio_kiocb allocation

[Jens Axboe]

Similarly to how we use the state->ios_left to know how many references
to get to a file, we can use it to allocate the aio_kiocb's we need in
bulk.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/aio.c | 47 +++++++++++++++++++++++++++++++++++++++--------
 1 file changed, 39 insertions(+), 8 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index e9dbaedda7ae..a385e7c06bfa 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -240,6 +240,8 @@ struct aio_kiocb {
 	};
 };
 
+#define AIO_IOPOLL_BATCH	8
+
 struct aio_submit_state {
 	struct kioctx *ctx;
 
@@ -254,6 +256,13 @@ struct aio_submit_state {
 	struct list_head req_list;
 	unsigned int req_count;
 
+	/*
+	 * aio_kiocb alloc cache
+	 */
+	void *iocbs[AIO_IOPOLL_BATCH];
+	unsigned int free_iocbs;
+	unsigned int cur_iocb;
+
 	/*
 	 * File reference cache
 	 */
@@ -1113,15 +1122,35 @@ static void aio_iocb_init(struct kioctx *ctx, struct aio_kiocb *req)
  *	Allocate a slot for an aio request.
  * Returns NULL if no requests are free.
  */
-static inline struct aio_kiocb *aio_get_req(struct kioctx *ctx)
+static struct aio_kiocb *aio_get_req(struct kioctx *ctx,
+				     struct aio_submit_state *state)
 {
 	struct aio_kiocb *req;
 
-	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
-	if (unlikely(!req))
-		return NULL;
+	if (!state)
+		req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
+	else if (!state->free_iocbs) {
+		size_t size;
+
+		size = min_t(size_t, state->ios_left, ARRAY_SIZE(state->iocbs));
+		size = kmem_cache_alloc_bulk(kiocb_cachep, GFP_KERNEL, size,
+						state->iocbs);
+		if (size < 0)
+			return ERR_PTR(size);
+		else if (!size)
+			return ERR_PTR(-ENOMEM);
+		state->free_iocbs = size - 1;
+		state->cur_iocb = 1;
+		req = state->iocbs[0];
+	} else {
+		req = state->iocbs[state->cur_iocb];
+		state->free_iocbs--;
+		state->cur_iocb++;
+	}
+
+	if (req)
+		aio_iocb_init(ctx, req);
 
-	aio_iocb_init(ctx, req);
 	return req;
 }
 
@@ -1359,8 +1388,6 @@ static bool aio_read_events(struct kioctx *ctx, long min_nr, long nr,
 	return ret < 0 || *i >= min_nr;
 }
 
-#define AIO_IOPOLL_BATCH	8
-
 /*
  * Process completed iocb iopoll entries, copying the result to userspace.
  */
@@ -2371,7 +2398,7 @@ static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
 		return -EAGAIN;
 
 	ret = -EAGAIN;
-	req = aio_get_req(ctx);
+	req = aio_get_req(ctx, state);
 	if (unlikely(!req))
 		goto out_put_reqs_available;
 
@@ -2503,6 +2530,9 @@ static void aio_submit_state_end(struct aio_submit_state *state)
 	if (!list_empty(&state->req_list))
 		aio_flush_state_reqs(state->ctx, state);
 	aio_file_put(state);
+	if (state->free_iocbs)
+		kmem_cache_free_bulk(kiocb_cachep, state->free_iocbs,
+					&state->iocbs[state->cur_iocb]);
 }
 
 /*
@@ -2514,6 +2544,7 @@ static void aio_submit_state_start(struct aio_submit_state *state,
 	state->ctx = ctx;
 	INIT_LIST_HEAD(&state->req_list);
 	state->req_count = 0;
+	state->free_iocbs = 0;
 	state->file = NULL;
 	state->ios_left = max_ios;
 #ifdef CONFIG_BLOCK
-- 
2.17.1

[PATCH 21/27] block: add BIO_HOLD_PAGES flag

[Jens Axboe]

For user mapped IO, we do get_user_pages() upfront, and then do a
put_page() on each page at end_io time to release the page reference. In
preparation for having permanently mapped pages, add a BIO_HOLD_PAGES
flag that tells us not to release the pages, the caller will do that.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 block/bio.c               | 6 ++++--
 include/linux/blk_types.h | 1 +
 2 files changed, 5 insertions(+), 2 deletions(-)

diff --git a/block/bio.c b/block/bio.c
index 036e3f0cc736..03dde1c03ae6 100644
--- a/block/bio.c
+++ b/block/bio.c
@@ -1636,7 +1636,8 @@ static void bio_dirty_fn(struct work_struct *work)
 		next = bio->bi_private;
 
 		bio_set_pages_dirty(bio);
-		bio_release_pages(bio);
+		if (!bio_flagged(bio, BIO_HOLD_PAGES))
+			bio_release_pages(bio);
 		bio_put(bio);
 	}
 }
@@ -1652,7 +1653,8 @@ void bio_check_pages_dirty(struct bio *bio)
 			goto defer;
 	}
 
-	bio_release_pages(bio);
+	if (!bio_flagged(bio, BIO_HOLD_PAGES))
+		bio_release_pages(bio);
 	bio_put(bio);
 	return;
 defer:
diff --git a/include/linux/blk_types.h b/include/linux/blk_types.h
index 921d734d6b5d..356a4c89b0d9 100644
--- a/include/linux/blk_types.h
+++ b/include/linux/blk_types.h
@@ -228,6 +228,7 @@ struct bio {
 #define BIO_TRACE_COMPLETION 10	/* bio_endio() should trace the final completion
 				 * of this bio. */
 #define BIO_QUEUE_ENTERED 11	/* can use blk_queue_enter_live() */
+#define BIO_HOLD_PAGES	12	/* don't put O_DIRECT pages */
 
 /* See BVEC_POOL_OFFSET below before adding new flags */
 
-- 
2.17.1

[PATCH 22/27] block: implement bio helper to add iter bvec pages to bio

[Jens Axboe]

For an ITER_BVEC, we can just iterate the iov and add the pages
to the bio directly.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 block/bio.c         | 27 +++++++++++++++++++++++++++
 include/linux/bio.h |  1 +
 2 files changed, 28 insertions(+)

diff --git a/block/bio.c b/block/bio.c
index 03dde1c03ae6..1da1391e8b1d 100644
--- a/block/bio.c
+++ b/block/bio.c
@@ -904,6 +904,33 @@ int bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter)
 }
 EXPORT_SYMBOL_GPL(bio_iov_iter_get_pages);
 
+/**
+ * bio_iov_bvec_add_pages - add pages from an ITER_BVEC to a bio
+ * @bio: bio to add pages to
+ * @iter: iov iterator describing the region to be added
+ *
+ * Iterate pages in the @iter and add them to the bio. We flag the
+ * @bio with BIO_HOLD_PAGES, telling IO completion not to free them.
+ */
+int bio_iov_bvec_add_pages(struct bio *bio, struct iov_iter *iter)
+{
+	unsigned short orig_vcnt = bio->bi_vcnt;
+	const struct bio_vec *bv;
+
+	do {
+		size_t size;
+
+		bv = iter->bvec + iter->iov_offset;
+		size = bio_add_page(bio, bv->bv_page, bv->bv_len, bv->bv_offset);
+		if (size != bv->bv_len)
+			break;
+		iov_iter_advance(iter, size);
+	} while (iov_iter_count(iter) && !bio_full(bio));
+
+	bio_set_flag(bio, BIO_HOLD_PAGES);
+	return bio->bi_vcnt > orig_vcnt ? 0 : -EINVAL;
+}
+
 static void submit_bio_wait_endio(struct bio *bio)
 {
 	complete(bio->bi_private);
diff --git a/include/linux/bio.h b/include/linux/bio.h
index 7380b094dcca..ca25ea890192 100644
--- a/include/linux/bio.h
+++ b/include/linux/bio.h
@@ -434,6 +434,7 @@ bool __bio_try_merge_page(struct bio *bio, struct page *page,
 void __bio_add_page(struct bio *bio, struct page *page,
 		unsigned int len, unsigned int off);
 int bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter);
+int bio_iov_bvec_add_pages(struct bio *bio, struct iov_iter *iter);
 struct rq_map_data;
 extern struct bio *bio_map_user_iov(struct request_queue *,
 				    struct iov_iter *, gfp_t);
-- 
2.17.1

[PATCH 23/27] fs: add support for mapping an ITER_BVEC for O_DIRECT

[Jens Axboe]

This adds support for sync/async O_DIRECT to make a bvec type iter
for bdev access, as well as iomap.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/block_dev.c | 16 ++++++++++++----
 fs/iomap.c     | 10 +++++++---
 2 files changed, 19 insertions(+), 7 deletions(-)

diff --git a/fs/block_dev.c b/fs/block_dev.c
index b8f574615792..236c6abe649d 100644
--- a/fs/block_dev.c
+++ b/fs/block_dev.c
@@ -219,7 +219,10 @@ __blkdev_direct_IO_simple(struct kiocb *iocb, struct iov_iter *iter,
 	bio.bi_end_io = blkdev_bio_end_io_simple;
 	bio.bi_ioprio = iocb->ki_ioprio;
 
-	ret = bio_iov_iter_get_pages(&bio, iter);
+	if (iov_iter_is_bvec(iter))
+		ret = bio_iov_bvec_add_pages(&bio, iter);
+	else
+		ret = bio_iov_iter_get_pages(&bio, iter);
 	if (unlikely(ret))
 		goto out;
 	ret = bio.bi_iter.bi_size;
@@ -326,8 +329,9 @@ static void blkdev_bio_end_io(struct bio *bio)
 		struct bio_vec *bvec;
 		int i;
 
-		bio_for_each_segment_all(bvec, bio, i)
-			put_page(bvec->bv_page);
+		if (!bio_flagged(bio, BIO_HOLD_PAGES))
+			bio_for_each_segment_all(bvec, bio, i)
+				put_page(bvec->bv_page);
 		bio_put(bio);
 	}
 }
@@ -381,7 +385,11 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, int nr_pages)
 		bio->bi_end_io = blkdev_bio_end_io;
 		bio->bi_ioprio = iocb->ki_ioprio;
 
-		ret = bio_iov_iter_get_pages(bio, iter);
+		if (iov_iter_is_bvec(iter))
+			ret = bio_iov_bvec_add_pages(bio, iter);
+		else
+			ret = bio_iov_iter_get_pages(bio, iter);
+
 		if (unlikely(ret)) {
 			bio->bi_status = BLK_STS_IOERR;
 			bio_endio(bio);
diff --git a/fs/iomap.c b/fs/iomap.c
index f3039989de73..2bb309a320a3 100644
--- a/fs/iomap.c
+++ b/fs/iomap.c
@@ -1573,8 +1573,9 @@ static void iomap_dio_bio_end_io(struct bio *bio)
 		struct bio_vec *bvec;
 		int i;
 
-		bio_for_each_segment_all(bvec, bio, i)
-			put_page(bvec->bv_page);
+		if (!bio_flagged(bio, BIO_HOLD_PAGES))
+			bio_for_each_segment_all(bvec, bio, i)
+				put_page(bvec->bv_page);
 		bio_put(bio);
 	}
 }
@@ -1673,7 +1674,10 @@ iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
 		bio->bi_private = dio;
 		bio->bi_end_io = iomap_dio_bio_end_io;
 
-		ret = bio_iov_iter_get_pages(bio, &iter);
+		if (iov_iter_is_bvec(&iter))
+			ret = bio_iov_bvec_add_pages(bio, &iter);
+		else
+			ret = bio_iov_iter_get_pages(bio, &iter);
 		if (unlikely(ret)) {
 			/*
 			 * We have to stop part way through an IO. We must fall
-- 
2.17.1

[PATCH 24/27] aio: add support for pre-mapped user IO buffers

[Jens Axboe]

If we have fixed user buffers, we can map them into the kernel when we
setup the io_context. That avoids the need to do get_user_pages() for
each and every IO.

To utilize this feature, the application must set both
IOCTX_FLAG_USERIOCB, to provide iocb's in userspace, and then
IOCTX_FLAG_FIXEDBUFS. The latter tells aio that the iocbs that are
mapped already contain valid destination and sizes. These buffers can
then be mapped into the kernel for the life time of the io_context, as
opposed to just the duration of the each single IO.

Only works with non-vectored read/write commands for now, not with
PREADV/PWRITEV.

A limit of 4M is imposed as the largest buffer we currently support.
There's nothing preventing us from going larger, but we need some cap,
and 4M seemed like it would definitely be big enough. RLIMIT_MEMLOCK
is used to cap the total amount of memory pinned.

See the fio change for how to utilize this feature:

http://git.kernel.dk/cgit/fio/commit/?id=2041bd343da1c1e955253f62374588718c64f0f3

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/aio.c                     | 193 ++++++++++++++++++++++++++++++++---
 include/uapi/linux/aio_abi.h |   1 +
 2 files changed, 177 insertions(+), 17 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index a385e7c06bfa..7bd5975a83e6 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -42,6 +42,7 @@
 #include <linux/ramfs.h>
 #include <linux/percpu-refcount.h>
 #include <linux/mount.h>
+#include <linux/sizes.h>
 
 #include <asm/kmap_types.h>
 #include <linux/uaccess.h>
@@ -97,6 +98,11 @@ struct aio_mapped_range {
 	long nr_pages;
 };
 
+struct aio_mapped_ubuf {
+	struct bio_vec *bvec;
+	unsigned int nr_bvecs;
+};
+
 struct kioctx {
 	struct percpu_ref	users;
 	atomic_t		dead;
@@ -132,6 +138,8 @@ struct kioctx {
 	struct page		**ring_pages;
 	long			nr_pages;
 
+	struct aio_mapped_ubuf	*user_bufs;
+
 	struct aio_mapped_range	iocb_range;
 
 	struct rcu_work		free_rwork;	/* see free_ioctx() */
@@ -301,6 +309,7 @@ static const bool aio_use_state_req_list = false;
 
 static void aio_useriocb_unmap(struct kioctx *);
 static void aio_iopoll_reap_events(struct kioctx *);
+static void aio_iocb_buffer_unmap(struct kioctx *);
 
 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
 {
@@ -662,6 +671,7 @@ static void free_ioctx(struct work_struct *work)
 					  free_rwork);
 	pr_debug("freeing %p\n", ctx);
 
+	aio_iocb_buffer_unmap(ctx);
 	aio_useriocb_unmap(ctx);
 	aio_free_ring(ctx);
 	free_percpu(ctx->cpu);
@@ -1675,6 +1685,122 @@ static int aio_useriocb_map(struct kioctx *ctx, struct iocb __user *iocbs)
 	return aio_map_range(&ctx->iocb_range, iocbs, size, 0);
 }
 
+static void aio_iocb_buffer_unmap(struct kioctx *ctx)
+{
+	int i, j;
+
+	if (!ctx->user_bufs)
+		return;
+
+	for (i = 0; i < ctx->max_reqs; i++) {
+		struct aio_mapped_ubuf *amu = &ctx->user_bufs[i];
+
+		for (j = 0; j < amu->nr_bvecs; j++)
+			put_page(amu->bvec[j].bv_page);
+
+		kfree(amu->bvec);
+		amu->nr_bvecs = 0;
+	}
+
+	kfree(ctx->user_bufs);
+	ctx->user_bufs = NULL;
+}
+
+static int aio_iocb_buffer_map(struct kioctx *ctx)
+{
+	unsigned long total_pages, page_limit;
+	struct page **pages = NULL;
+	int i, j, got_pages = 0;
+	struct iocb *iocb;
+	int ret = -EINVAL;
+
+	ctx->user_bufs = kzalloc(ctx->max_reqs * sizeof(struct aio_mapped_ubuf),
+					GFP_KERNEL);
+	if (!ctx->user_bufs)
+		return -ENOMEM;
+
+	/* Don't allow more pages than we can safely lock */
+	total_pages = 0;
+	page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
+
+	for (i = 0; i < ctx->max_reqs; i++) {
+		struct aio_mapped_ubuf *amu = &ctx->user_bufs[i];
+		unsigned long off, start, end, ubuf;
+		int pret, nr_pages;
+		size_t size;
+
+		iocb = aio_iocb_from_index(ctx, i);
+
+		/*
+		 * Don't impose further limits on the size and buffer
+		 * constraints here, we'll -EINVAL later when IO is
+		 * submitted if they are wrong.
+		 */
+		ret = -EFAULT;
+		if (!iocb->aio_buf)
+			goto err;
+
+		/* arbitrary limit, but we need something */
+		if (iocb->aio_nbytes > SZ_4M)
+			goto err;
+
+		ubuf = iocb->aio_buf;
+		end = (ubuf + iocb->aio_nbytes + PAGE_SIZE - 1) >> PAGE_SHIFT;
+		start = ubuf >> PAGE_SHIFT;
+		nr_pages = end - start;
+
+		ret = -ENOMEM;
+		if (total_pages + nr_pages > page_limit)
+			goto err;
+
+		if (!pages || nr_pages > got_pages) {
+			kfree(pages);
+			pages = kmalloc(nr_pages * sizeof(struct page *),
+					GFP_KERNEL);
+			if (!pages)
+				goto err;
+			got_pages = nr_pages;
+		}
+
+		amu->bvec = kmalloc(nr_pages * sizeof(struct bio_vec),
+					GFP_KERNEL);
+		if (!amu->bvec)
+			goto err;
+
+		down_write(&current->mm->mmap_sem);
+		pret = get_user_pages((unsigned long) iocb->aio_buf, nr_pages,
+					1, pages, NULL);
+		up_write(&current->mm->mmap_sem);
+
+		if (pret < nr_pages) {
+			if (pret < 0)
+				ret = pret;
+			goto err;
+		}
+
+		off = ubuf & ~PAGE_MASK;
+		size = iocb->aio_nbytes;
+		for (j = 0; j < nr_pages; j++) {
+			size_t vec_len;
+
+			vec_len = min_t(size_t, size, PAGE_SIZE - off);
+			amu->bvec[j].bv_page = pages[j];
+			amu->bvec[j].bv_len = vec_len;
+			amu->bvec[j].bv_offset = off;
+			off = 0;
+			size -= vec_len;
+		}
+		amu->nr_bvecs = nr_pages;
+		total_pages += nr_pages;
+	}
+	kfree(pages);
+	return 0;
+err:
+	kfree(pages);
+	aio_iocb_buffer_unmap(ctx);
+	return ret;
+}
+
 /* sys_io_setup2:
  *	Like sys_io_setup(), except that it takes a set of flags
  *	(IOCTX_FLAG_*), and some pointers to user structures:
@@ -1696,7 +1822,8 @@ SYSCALL_DEFINE6(io_setup2, u32, nr_events, u32, flags, struct iocb __user *,
 
 	if (user1 || user2)
 		return -EINVAL;
-	if (flags & ~(IOCTX_FLAG_USERIOCB | IOCTX_FLAG_IOPOLL))
+	if (flags & ~(IOCTX_FLAG_USERIOCB | IOCTX_FLAG_IOPOLL |
+		      IOCTX_FLAG_FIXEDBUFS))
 		return -EINVAL;
 
 	ret = get_user(ctx, ctxp);
@@ -1712,6 +1839,15 @@ SYSCALL_DEFINE6(io_setup2, u32, nr_events, u32, flags, struct iocb __user *,
 		ret = aio_useriocb_map(ioctx, iocbs);
 		if (ret)
 			goto err;
+		if (flags & IOCTX_FLAG_FIXEDBUFS) {
+			ret = aio_iocb_buffer_map(ioctx);
+			if (ret)
+				goto err;
+		}
+	} else if (flags & IOCTX_FLAG_FIXEDBUFS) {
+		/* can only support fixed bufs with user mapped iocbs */
+		ret = -EINVAL;
+		goto err;
 	}
 
 	ret = put_user(ioctx->user_id, ctxp);
@@ -1989,23 +2125,39 @@ static int aio_prep_rw(struct aio_kiocb *kiocb, const struct iocb *iocb,
 	return ret;
 }
 
-static int aio_setup_rw(int rw, const struct iocb *iocb, struct iovec **iovec,
-		bool vectored, bool compat, struct iov_iter *iter)
+static int aio_setup_rw(int rw, struct aio_kiocb *kiocb,
+		const struct iocb *iocb, struct iovec **iovec, bool vectored,
+		bool compat, bool kaddr, struct iov_iter *iter)
 {
-	void __user *buf = (void __user *)(uintptr_t)iocb->aio_buf;
+	void __user *ubuf = (void __user *)(uintptr_t)iocb->aio_buf;
 	size_t len = iocb->aio_nbytes;
 
 	if (!vectored) {
-		ssize_t ret = import_single_range(rw, buf, len, *iovec, iter);
+		ssize_t ret;
+
+		if (!kaddr) {
+			ret = import_single_range(rw, ubuf, len, *iovec, iter);
+		} else {
+			long index = (long) kiocb->ki_user_iocb;
+			struct aio_mapped_ubuf *amu;
+
+			/* __io_submit_one() already validated the index */
+			amu = &kiocb->ki_ctx->user_bufs[index];
+			iov_iter_bvec(iter, rw, amu->bvec, amu->nr_bvecs, len);
+			ret = 0;
+
+		}
 		*iovec = NULL;
 		return ret;
 	}
+	if (kaddr)
+		return -EINVAL;
 #ifdef CONFIG_COMPAT
 	if (compat)
-		return compat_import_iovec(rw, buf, len, UIO_FASTIOV, iovec,
+		return compat_import_iovec(rw, ubuf, len, UIO_FASTIOV, iovec,
 				iter);
 #endif
-	return import_iovec(rw, buf, len, UIO_FASTIOV, iovec, iter);
+	return import_iovec(rw, ubuf, len, UIO_FASTIOV, iovec, iter);
 }
 
 static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
@@ -2078,7 +2230,7 @@ static void aio_iopoll_iocb_issued(struct aio_submit_state *state,
 
 static ssize_t aio_read(struct aio_kiocb *kiocb, const struct iocb *iocb,
 			struct aio_submit_state *state, bool vectored,
-			bool compat)
+			bool compat, bool kaddr)
 {
 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
 	struct kiocb *req = &kiocb->rw;
@@ -2098,9 +2250,11 @@ static ssize_t aio_read(struct aio_kiocb *kiocb, const struct iocb *iocb,
 	if (unlikely(!file->f_op->read_iter))
 		goto out_fput;
 
-	ret = aio_setup_rw(READ, iocb, &iovec, vectored, compat, &iter);
+	ret = aio_setup_rw(READ, kiocb, iocb, &iovec, vectored, compat, kaddr,
+				&iter);
 	if (ret)
 		goto out_fput;
+
 	ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
 	if (!ret)
 		aio_rw_done(req, call_read_iter(file, req, &iter));
@@ -2113,7 +2267,7 @@ static ssize_t aio_read(struct aio_kiocb *kiocb, const struct iocb *iocb,
 
 static ssize_t aio_write(struct aio_kiocb *kiocb, const struct iocb *iocb,
 			 struct aio_submit_state *state, bool vectored,
-			 bool compat)
+			 bool compat, bool kaddr)
 {
 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
 	struct kiocb *req = &kiocb->rw;
@@ -2133,7 +2287,8 @@ static ssize_t aio_write(struct aio_kiocb *kiocb, const struct iocb *iocb,
 	if (unlikely(!file->f_op->write_iter))
 		goto out_fput;
 
-	ret = aio_setup_rw(WRITE, iocb, &iovec, vectored, compat, &iter);
+	ret = aio_setup_rw(WRITE, kiocb, iocb, &iovec, vectored, compat, kaddr,
+				&iter);
 	if (ret)
 		goto out_fput;
 	ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
@@ -2372,7 +2527,8 @@ static ssize_t aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
 
 static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
 			   struct iocb __user *user_iocb,
-			   struct aio_submit_state *state, bool compat)
+			   struct aio_submit_state *state, bool compat,
+			   bool kaddr)
 {
 	struct aio_kiocb *req;
 	ssize_t ret;
@@ -2432,16 +2588,16 @@ static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
 	ret = -EINVAL;
 	switch (iocb->aio_lio_opcode) {
 	case IOCB_CMD_PREAD:
-		ret = aio_read(req, iocb, state, false, compat);
+		ret = aio_read(req, iocb, state, false, compat, kaddr);
 		break;
 	case IOCB_CMD_PWRITE:
-		ret = aio_write(req, iocb, state, false, compat);
+		ret = aio_write(req, iocb, state, false, compat, kaddr);
 		break;
 	case IOCB_CMD_PREADV:
-		ret = aio_read(req, iocb, state, true, compat);
+		ret = aio_read(req, iocb, state, true, compat, kaddr);
 		break;
 	case IOCB_CMD_PWRITEV:
-		ret = aio_write(req, iocb, state, true, compat);
+		ret = aio_write(req, iocb, state, true, compat, kaddr);
 		break;
 	case IOCB_CMD_FSYNC:
 		if (ctx->flags & IOCTX_FLAG_IOPOLL)
@@ -2493,6 +2649,7 @@ static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
 			 struct aio_submit_state *state, bool compat)
 {
 	struct iocb iocb, *iocbp;
+	bool kaddr;
 
 	if (ctx->flags & IOCTX_FLAG_USERIOCB) {
 		unsigned long iocb_index = (unsigned long) user_iocb;
@@ -2500,14 +2657,16 @@ static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
 		if (iocb_index >= ctx->max_reqs)
 			return -EINVAL;
 
+		kaddr = (ctx->flags & IOCTX_FLAG_FIXEDBUFS) != 0;
 		iocbp = aio_iocb_from_index(ctx, iocb_index);
 	} else {
 		if (unlikely(copy_from_user(&iocb, user_iocb, sizeof(iocb))))
 			return -EFAULT;
+		kaddr = false;
 		iocbp = &iocb;
 	}
 
-	return __io_submit_one(ctx, iocbp, user_iocb, state, compat);
+	return __io_submit_one(ctx, iocbp, user_iocb, state, compat, kaddr);
 }
 
 #ifdef CONFIG_BLOCK
diff --git a/include/uapi/linux/aio_abi.h b/include/uapi/linux/aio_abi.h
index ea0b9a19f4df..05d72cf86bd3 100644
--- a/include/uapi/linux/aio_abi.h
+++ b/include/uapi/linux/aio_abi.h
@@ -110,6 +110,7 @@ struct iocb {
 
 #define IOCTX_FLAG_USERIOCB	(1 << 0)	/* iocbs are user mapped */
 #define IOCTX_FLAG_IOPOLL	(1 << 1)	/* io_context is polled */
+#define IOCTX_FLAG_FIXEDBUFS	(1 << 2)	/* IO buffers are fixed */
 
 #undef IFBIG
 #undef IFLITTLE
-- 
2.17.1

[PATCH 25/27] aio: split old ring complete out from aio_complete()

[Jens Axboe]

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/aio.c | 17 ++++++++++++-----
 1 file changed, 12 insertions(+), 5 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index 7bd5975a83e6..a69e6228fe62 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -1218,12 +1218,9 @@ static void aio_fill_event(struct io_event *ev, struct aio_kiocb *iocb,
 	ev->res2 = res2;
 }
 
-/* aio_complete
- *	Called when the io request on the given iocb is complete.
- */
-static void aio_complete(struct aio_kiocb *iocb, long res, long res2)
+static void aio_ring_complete(struct kioctx *ctx, struct aio_kiocb *iocb,
+			      long res, long res2)
 {
-	struct kioctx	*ctx = iocb->ki_ctx;
 	struct aio_ring	*ring;
 	struct io_event	*ev_page, *event;
 	unsigned tail, pos, head;
@@ -1273,6 +1270,16 @@ static void aio_complete(struct aio_kiocb *iocb, long res, long res2)
 	spin_unlock_irqrestore(&ctx->completion_lock, flags);
 
 	pr_debug("added to ring %p at [%u]\n", iocb, tail);
+}
+
+/* aio_complete
+ *	Called when the io request on the given iocb is complete.
+ */
+static void aio_complete(struct aio_kiocb *iocb, long res, long res2)
+{
+	struct kioctx *ctx = iocb->ki_ctx;
+
+	aio_ring_complete(ctx, iocb, res, res2);
 
 	/*
 	 * Check if the user asked us to deliver the result through an
-- 
2.17.1

[PATCH 26/27] aio: add support for submission/completion rings

[Jens Axboe]

Experimental support for submitting and completing IO through rings
shared between the application and kernel.

The submission rings are struct iocb, like we would submit through
io_submit(), and the completion rings are struct io_event, like we
would pass in (and copy back) from io_getevents().

A new system call is added for this, io_ring_enter(). This system
call submits IO that is queued in the SQ ring, and/or completes IO
and stores the results in the CQ ring.

This could be augmented with a kernel thread that does the submission
and polling, then the application would never have to enter the
kernel to do IO.

Sample application: http://git.kernel.dk/cgit/fio/plain/t/aio-ring.c

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 arch/x86/entry/syscalls/syscall_64.tbl |   1 +
 fs/aio.c                               | 456 +++++++++++++++++++++++--
 include/linux/syscalls.h               |   4 +-
 include/uapi/linux/aio_abi.h           |  26 ++
 kernel/sys_ni.c                        |   1 +
 5 files changed, 452 insertions(+), 36 deletions(-)

diff --git a/arch/x86/entry/syscalls/syscall_64.tbl b/arch/x86/entry/syscalls/syscall_64.tbl
index 67c357225fb0..55a26700a637 100644
--- a/arch/x86/entry/syscalls/syscall_64.tbl
+++ b/arch/x86/entry/syscalls/syscall_64.tbl
@@ -344,6 +344,7 @@
 333	common	io_pgetevents		__x64_sys_io_pgetevents
 334	common	rseq			__x64_sys_rseq
 335	common	io_setup2		__x64_sys_io_setup2
+336	common	io_ring_enter		__x64_sys_io_ring_enter
 
 #
 # x32-specific system call numbers start at 512 to avoid cache impact
diff --git a/fs/aio.c b/fs/aio.c
index a69e6228fe62..f4f39a7f8f94 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -142,6 +142,11 @@ struct kioctx {
 
 	struct aio_mapped_range	iocb_range;
 
+	/* if used, completion and submission rings */
+	struct aio_mapped_range	sq_ring;
+	struct aio_mapped_range	cq_ring;
+	int			cq_ring_overflow;
+
 	struct rcu_work		free_rwork;	/* see free_ioctx() */
 
 	/*
@@ -297,6 +302,8 @@ static const struct address_space_operations aio_ctx_aops;
 
 static const unsigned int iocb_page_shift =
 				ilog2(PAGE_SIZE / sizeof(struct iocb));
+static const unsigned int event_page_shift =
+				ilog2(PAGE_SIZE / sizeof(struct io_event));
 
 /*
  * We rely on block level unplugs to flush pending requests, if we schedule
@@ -307,6 +314,7 @@ static const bool aio_use_state_req_list = true;
 static const bool aio_use_state_req_list = false;
 #endif
 
+static void aio_scqring_unmap(struct kioctx *);
 static void aio_useriocb_unmap(struct kioctx *);
 static void aio_iopoll_reap_events(struct kioctx *);
 static void aio_iocb_buffer_unmap(struct kioctx *);
@@ -539,6 +547,12 @@ static const struct address_space_operations aio_ctx_aops = {
 #endif
 };
 
+/* Polled IO or SQ/CQ rings don't use the old ring */
+static bool aio_ctx_old_ring(struct kioctx *ctx)
+{
+	return !(ctx->flags & (IOCTX_FLAG_IOPOLL | IOCTX_FLAG_SCQRING));
+}
+
 static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
 {
 	struct aio_ring *ring;
@@ -553,7 +567,7 @@ static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
 	 * IO polling doesn't require any io event entries
 	 */
 	size = sizeof(struct aio_ring);
-	if (!(ctx->flags & IOCTX_FLAG_IOPOLL)) {
+	if (aio_ctx_old_ring(ctx)) {
 		nr_events += 2;	/* 1 is required, 2 for good luck */
 		size += sizeof(struct io_event) * nr_events;
 	}
@@ -640,6 +654,17 @@ static int aio_setup_ring(struct kioctx *ctx, unsigned int nr_events)
 	return 0;
 }
 
+/*
+ * Don't support cancel on anything that isn't regular aio
+ */
+static bool aio_ctx_supports_cancel(struct kioctx *ctx)
+{
+	int noflags = IOCTX_FLAG_USERIOCB | IOCTX_FLAG_IOPOLL |
+		      IOCTX_FLAG_SCQRING;
+
+	return (ctx->flags & noflags) == 0;
+}
+
 #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)
@@ -650,6 +675,8 @@ void kiocb_set_cancel_fn(struct kiocb *iocb, kiocb_cancel_fn *cancel)
 	struct kioctx *ctx = req->ki_ctx;
 	unsigned long flags;
 
+	if (WARN_ON_ONCE(!aio_ctx_supports_cancel(ctx)))
+		return;
 	if (WARN_ON_ONCE(!list_empty(&req->ki_list)))
 		return;
 
@@ -673,6 +700,7 @@ static void free_ioctx(struct work_struct *work)
 
 	aio_iocb_buffer_unmap(ctx);
 	aio_useriocb_unmap(ctx);
+	aio_scqring_unmap(ctx);
 	aio_free_ring(ctx);
 	free_percpu(ctx->cpu);
 	percpu_ref_exit(&ctx->reqs);
@@ -1218,6 +1246,47 @@ static void aio_fill_event(struct io_event *ev, struct aio_kiocb *iocb,
 	ev->res2 = res2;
 }
 
+static struct io_event *__aio_get_cqring_ev(struct aio_io_event_ring *ring,
+					    struct aio_mapped_range *range,
+					    unsigned *next_tail)
+{
+	struct io_event *ev;
+	unsigned tail;
+
+	smp_rmb();
+	tail = READ_ONCE(ring->tail);
+	*next_tail = tail + 1;
+	if (*next_tail == ring->nr_events)
+		*next_tail = 0;
+	if (*next_tail == READ_ONCE(ring->head))
+		return NULL;
+
+	/* io_event array starts offset one into the mapped range */
+	tail++;
+	ev = page_address(range->pages[tail >> event_page_shift]);
+	tail &= ((1 << event_page_shift) - 1);
+	return ev + tail;
+}
+
+static void aio_commit_cqring(struct kioctx *ctx, unsigned next_tail)
+{
+	struct aio_io_event_ring *ring;
+
+	ring = page_address(ctx->cq_ring.pages[0]);
+	if (next_tail != ring->tail) {
+		ring->tail = next_tail;
+		smp_wmb();
+	}
+}
+
+static struct io_event *aio_peek_cqring(struct kioctx *ctx, unsigned *ntail)
+{
+	struct aio_io_event_ring *ring;
+
+	ring = page_address(ctx->cq_ring.pages[0]);
+	return __aio_get_cqring_ev(ring, &ctx->cq_ring, ntail);
+}
+
 static void aio_ring_complete(struct kioctx *ctx, struct aio_kiocb *iocb,
 			      long res, long res2)
 {
@@ -1279,7 +1348,36 @@ static void aio_complete(struct aio_kiocb *iocb, long res, long res2)
 {
 	struct kioctx *ctx = iocb->ki_ctx;
 
-	aio_ring_complete(ctx, iocb, res, res2);
+	if (ctx->flags & IOCTX_FLAG_SCQRING) {
+		unsigned long flags;
+		struct io_event *ev;
+		unsigned int tail;
+
+		/*
+		 * If we can't get a cq entry, userspace overflowed the
+		 * submission (by quite a lot). Flag it as an overflow
+		 * condition, and next io_ring_enter(2) call will return
+		 * -EOVERFLOW.
+		 */
+		spin_lock_irqsave(&ctx->completion_lock, flags);
+		ev = aio_peek_cqring(ctx, &tail);
+		if (ev) {
+			aio_fill_event(ev, iocb, res, res2);
+			aio_commit_cqring(ctx, tail);
+		} else
+			ctx->cq_ring_overflow = 1;
+		spin_unlock_irqrestore(&ctx->completion_lock, flags);
+	} else {
+		aio_ring_complete(ctx, iocb, res, res2);
+
+		/*
+		 * We have to order our ring_info tail store above and test
+		 * of the wait list below outside the wait lock.  This is
+		 * like in wake_up_bit() where clearing a bit has to be
+		 * ordered with the unlocked test.
+		 */
+		smp_mb();
+	}
 
 	/*
 	 * Check if the user asked us to deliver the result through an
@@ -1291,14 +1389,6 @@ static void aio_complete(struct aio_kiocb *iocb, long res, long res2)
 		eventfd_ctx_put(iocb->ki_eventfd);
 	}
 
-	/*
-	 * We have to order our ring_info tail store above and test
-	 * of the wait list below outside the wait lock.  This is
-	 * like in wake_up_bit() where clearing a bit has to be
-	 * ordered with the unlocked test.
-	 */
-	smp_mb();
-
 	if (waitqueue_active(&ctx->wait))
 		wake_up(&ctx->wait);
 	iocb_put(iocb);
@@ -1421,6 +1511,9 @@ static long aio_iopoll_reap(struct kioctx *ctx, struct io_event __user *evs,
 		return 0;
 
 	list_for_each_entry_safe(iocb, n, &ctx->poll_completing, ki_list) {
+		struct io_event *ev = NULL;
+		unsigned int next_tail;
+
 		if (*nr_events == max)
 			break;
 		if (!test_bit(KIOCB_F_POLL_COMPLETED, &iocb->ki_flags))
@@ -1428,6 +1521,14 @@ static long aio_iopoll_reap(struct kioctx *ctx, struct io_event __user *evs,
 		if (to_free == AIO_IOPOLL_BATCH)
 			iocb_put_many(ctx, iocbs, &to_free);
 
+		/* Will only happen if the application over-commits */
+		ret = -EAGAIN;
+		if (ctx->flags & IOCTX_FLAG_SCQRING) {
+			ev = aio_peek_cqring(ctx, &next_tail);
+			if (!ev)
+				break;
+		}
+
 		list_del(&iocb->ki_list);
 		iocbs[to_free++] = iocb;
 
@@ -1446,8 +1547,11 @@ static long aio_iopoll_reap(struct kioctx *ctx, struct io_event __user *evs,
 			file_count = 1;
 		}
 
-		if (evs && copy_to_user(evs + *nr_events, &iocb->ki_ev,
-		    sizeof(iocb->ki_ev))) {
+		if (ev) {
+			memcpy(ev, &iocb->ki_ev, sizeof(*ev));
+			aio_commit_cqring(ctx, next_tail);
+		} else if (evs && copy_to_user(evs + *nr_events, &iocb->ki_ev,
+				sizeof(iocb->ki_ev))) {
 			ret = -EFAULT;
 			break;
 		}
@@ -1628,15 +1732,42 @@ static long read_events(struct kioctx *ctx, long min_nr, long nr,
 	return ret;
 }
 
-static struct iocb *aio_iocb_from_index(struct kioctx *ctx, int index)
+static struct iocb *__aio_sqring_from_index(struct aio_iocb_ring *ring,
+					    struct aio_mapped_range *range,
+					    int index)
 {
 	struct iocb *iocb;
 
-	iocb = page_address(ctx->iocb_range.pages[index >> iocb_page_shift]);
+	/* iocb array starts offset one into the mapped range */
+	index++;
+	iocb = page_address(range->pages[index >> iocb_page_shift]);
 	index &= ((1 << iocb_page_shift) - 1);
 	return iocb + index;
 }
 
+static struct iocb *aio_sqring_from_index(struct kioctx *ctx, int index)
+{
+	struct aio_iocb_ring *ring;
+
+	ring = page_address(ctx->sq_ring.pages[0]);
+	return __aio_sqring_from_index(ring, &ctx->sq_ring, index);
+}
+
+static struct iocb *aio_iocb_from_index(struct kioctx *ctx, int index)
+{
+	struct iocb *iocb;
+
+	if (ctx->flags & IOCTX_FLAG_SCQRING) {
+		iocb = aio_sqring_from_index(ctx, index);
+	} else {
+		iocb = page_address(ctx->iocb_range.pages[index >> iocb_page_shift]);
+		index &= ((1 << iocb_page_shift) - 1);
+		iocb += index;
+	}
+
+	return iocb;
+}
+
 static void aio_unmap_range(struct aio_mapped_range *range)
 {
 	int i;
@@ -1692,6 +1823,52 @@ static int aio_useriocb_map(struct kioctx *ctx, struct iocb __user *iocbs)
 	return aio_map_range(&ctx->iocb_range, iocbs, size, 0);
 }
 
+static void aio_scqring_unmap(struct kioctx *ctx)
+{
+	aio_unmap_range(&ctx->sq_ring);
+	aio_unmap_range(&ctx->cq_ring);
+}
+
+static int aio_scqring_map(struct kioctx *ctx,
+			   struct aio_iocb_ring __user *sq_ring,
+			   struct aio_io_event_ring __user *cq_ring)
+{
+	struct aio_iocb_ring *ksq_ring;
+	struct aio_io_event_ring *kcq_ring;
+	int ret, sq_ring_size, cq_ring_size;
+	size_t size;
+
+	/*
+	 * The CQ ring size is QD + 1, so we don't have to track full condition
+	 * for head == tail. The SQ ring we make twice that in size, to make
+	 * room for having more inflight than the QD.
+	 */
+	sq_ring_size = ctx->max_reqs;
+	cq_ring_size = 2 * ctx->max_reqs;
+
+	size = sq_ring_size * sizeof(struct iocb);
+	ret = aio_map_range(&ctx->sq_ring, sq_ring,
+			    sq_ring_size * sizeof(struct iocb), 0);
+	if (ret)
+		return ret;
+
+	ret = aio_map_range(&ctx->cq_ring, cq_ring,
+			    cq_ring_size * sizeof(struct io_event), FOLL_WRITE);
+	if (ret) {
+		aio_unmap_range(&ctx->sq_ring);
+		return ret;
+	}
+
+	ksq_ring = page_address(ctx->sq_ring.pages[0]);
+	ksq_ring->nr_events = sq_ring_size;
+	ksq_ring->head = ksq_ring->tail = 0;
+
+	kcq_ring = page_address(ctx->cq_ring.pages[0]);
+	kcq_ring->nr_events = cq_ring_size;
+	kcq_ring->head = kcq_ring->tail = 0;
+	return 0;
+}
+
 static void aio_iocb_buffer_unmap(struct kioctx *ctx)
 {
 	int i, j;
@@ -1815,22 +1992,22 @@ static int aio_iocb_buffer_map(struct kioctx *ctx)
  *	*iocbs - pointer to array of struct iocb, for when
  *	IOCTX_FLAG_USERIOCB is set in flags.
  *
- *	*user1 - reserved for future use
+ *	*sq_ring - pointer to the userspace SQ ring, if used.
  *
- *	*user2 - reserved for future use.
+ *	*cq_ring - pointer to the userspace CQ ring, if used.
  */
-SYSCALL_DEFINE6(io_setup2, u32, nr_events, u32, flags, struct iocb __user *,
-		iocbs, void __user *, user1, void __user *, user2,
+SYSCALL_DEFINE6(io_setup2, u32, nr_events, u32, flags,
+		struct iocb __user *, iocbs,
+		struct aio_iocb_ring __user *, sq_ring,
+		struct aio_io_event_ring __user *, cq_ring,
 		aio_context_t __user *, ctxp)
 {
 	struct kioctx *ioctx;
 	unsigned long ctx;
 	long ret;
 
-	if (user1 || user2)
-		return -EINVAL;
 	if (flags & ~(IOCTX_FLAG_USERIOCB | IOCTX_FLAG_IOPOLL |
-		      IOCTX_FLAG_FIXEDBUFS))
+		      IOCTX_FLAG_FIXEDBUFS | IOCTX_FLAG_SCQRING))
 		return -EINVAL;
 
 	ret = get_user(ctx, ctxp);
@@ -1843,18 +2020,26 @@ SYSCALL_DEFINE6(io_setup2, u32, nr_events, u32, flags, struct iocb __user *,
 		goto out;
 
 	if (flags & IOCTX_FLAG_USERIOCB) {
+		ret = -EINVAL;
+		if (flags & IOCTX_FLAG_SCQRING)
+			goto err;
+
 		ret = aio_useriocb_map(ioctx, iocbs);
 		if (ret)
 			goto err;
-		if (flags & IOCTX_FLAG_FIXEDBUFS) {
-			ret = aio_iocb_buffer_map(ioctx);
-			if (ret)
-				goto err;
-		}
-	} else if (flags & IOCTX_FLAG_FIXEDBUFS) {
-		/* can only support fixed bufs with user mapped iocbs */
+	}
+	if (flags & IOCTX_FLAG_SCQRING) {
+		ret = aio_scqring_map(ioctx, sq_ring, cq_ring);
+		if (ret)
+			goto err;
+	}
+	if (flags & IOCTX_FLAG_FIXEDBUFS) {
 		ret = -EINVAL;
-		goto err;
+		if (!(flags & (IOCTX_FLAG_USERIOCB | IOCTX_FLAG_SCQRING)))
+			goto err;
+		ret = aio_iocb_buffer_map(ioctx);
+		if (ret)
+			goto err;
 	}
 
 	ret = put_user(ioctx->user_id, ctxp);
@@ -2556,8 +2741,7 @@ static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
 		return -EINVAL;
 	}
 
-	/* Poll IO doesn't need ring reservations */
-	if (!(ctx->flags & IOCTX_FLAG_IOPOLL) && !get_reqs_available(ctx))
+	if (aio_ctx_old_ring(ctx) && !get_reqs_available(ctx))
 		return -EAGAIN;
 
 	ret = -EAGAIN;
@@ -2581,7 +2765,7 @@ static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
 	}
 
 	/* Don't support cancel on user mapped iocbs or polled context */
-	if (!(ctx->flags & (IOCTX_FLAG_USERIOCB | IOCTX_FLAG_IOPOLL))) {
+	if (aio_ctx_supports_cancel(ctx)) {
 		ret = put_user(KIOCB_KEY, &user_iocb->aio_key);
 		if (unlikely(ret)) {
 			pr_debug("EFAULT: aio_key\n");
@@ -2647,7 +2831,7 @@ static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
 		eventfd_ctx_put(req->ki_eventfd);
 	iocb_put(req);
 out_put_reqs_available:
-	if (!(ctx->flags & IOCTX_FLAG_IOPOLL))
+	if (aio_ctx_old_ring(ctx))
 		put_reqs_available(ctx, 1);
 	return ret;
 }
@@ -2720,6 +2904,201 @@ static void aio_submit_state_start(struct aio_submit_state *state,
 #endif
 }
 
+static struct iocb *__aio_get_sqring(struct aio_iocb_ring *ring,
+				     struct aio_mapped_range *range,
+				     unsigned *next_head)
+{
+	unsigned head;
+
+	smp_rmb();
+	head = READ_ONCE(ring->head);
+	if (head == READ_ONCE(ring->tail))
+		return NULL;
+
+	*next_head = head + 1;
+	if (*next_head == ring->nr_events)
+		*next_head = 0;
+
+	return __aio_sqring_from_index(ring, range, head);
+}
+
+static void aio_commit_sqring(struct kioctx *ctx, unsigned next_head)
+{
+	struct aio_iocb_ring *ring;
+
+	ring = page_address(ctx->sq_ring.pages[0]);
+	if (ring->head != next_head) {
+		ring->head = next_head;
+		smp_wmb();
+	}
+}
+
+static const struct iocb *aio_peek_sqring(struct kioctx *ctx, unsigned *nhead)
+{
+	struct aio_iocb_ring *ring;
+
+	ring = page_address(ctx->sq_ring.pages[0]);
+	return __aio_get_sqring(ring, &ctx->sq_ring, nhead);
+}
+
+static int aio_ring_submit(struct kioctx *ctx, unsigned int to_submit)
+{
+	bool kaddr = (ctx->flags & IOCTX_FLAG_FIXEDBUFS) != 0;
+	struct aio_submit_state state, *statep = NULL;
+	int i, ret = 0, submit = 0;
+
+	if (to_submit > AIO_PLUG_THRESHOLD) {
+		aio_submit_state_start(&state, ctx, to_submit);
+		statep = &state;
+	}
+
+	for (i = 0; i < to_submit; i++) {
+		const struct iocb *iocb;
+		unsigned int next_head;
+
+		iocb = aio_peek_sqring(ctx, &next_head);
+		if (!iocb)
+			break;
+
+		ret = __io_submit_one(ctx, iocb, NULL, NULL, false, kaddr);
+		if (ret)
+			break;
+
+		submit++;
+		aio_commit_sqring(ctx, next_head);
+	}
+
+	if (statep)
+		aio_submit_state_end(statep);
+
+	return submit ? submit : ret;
+}
+
+/*
+ * Wait until events become available, if we don't already have some. The
+ * application must reap them itself, as they reside on the shared cq ring.
+ */
+static int aio_cqring_wait(struct kioctx *ctx, int min_events)
+{
+	struct aio_io_event_ring *ring;
+	DEFINE_WAIT(wait);
+	int ret = 0;
+
+	ring = page_address(ctx->cq_ring.pages[0]);
+	smp_rmb();
+	if (ring->head != ring->tail)
+		return 0;
+
+	do {
+		prepare_to_wait(&ctx->wait, &wait, TASK_INTERRUPTIBLE);
+
+		ret = 0;
+		smp_rmb();
+		if (ring->head != ring->tail)
+			break;
+		if (!min_events)
+			break;
+
+		schedule();
+
+		ret = -EINVAL;
+		if (atomic_read(&ctx->dead))
+			break;
+		ret = -EINTR;
+		if (signal_pending(current))
+			break;
+	} while (1);
+
+	finish_wait(&ctx->wait, &wait);
+	return ret;
+}
+
+static int __io_ring_enter(struct kioctx *ctx, unsigned int to_submit,
+			   unsigned int min_complete, unsigned int flags)
+{
+	int ret = 0;
+
+	if (flags & IORING_FLAG_SUBMIT) {
+		ret = aio_ring_submit(ctx, to_submit);
+		if (ret < 0)
+			return ret;
+	}
+	if (flags & IORING_FLAG_GETEVENTS) {
+		unsigned int nr_events = 0;
+		int get_ret;
+
+		if (!ret && to_submit)
+			min_complete = 0;
+
+		if (ctx->flags & IOCTX_FLAG_IOPOLL)
+			get_ret = __aio_iopoll_check(ctx, NULL, &nr_events,
+							min_complete, -1U);
+		else
+			get_ret = aio_cqring_wait(ctx, min_complete);
+
+		if (get_ret < 0 && !ret)
+			ret = get_ret;
+	}
+
+	return ret;
+}
+
+/* sys_io_ring_enter:
+ *	Alternative way to both submit and complete IO, instead of using
+ *	io_submit(2) and io_getevents(2). Requires the use of the SQ/CQ
+ *	ring interface, hence the io_context must be setup with
+ *	io_setup2() and IOCTX_FLAG_SCQRING must be specified (and the
+ *	sq_ring/cq_ring passed in).
+ *
+ *	Returns the number of IOs submitted, if IORING_FLAG_SUBMIT
+ *	is used, otherwise returns 0 for IORING_FLAG_GETEVENTS success,
+ *	but not the number of events, as those will have to be found
+ *	by the application by reading the CQ ring anyway.
+ *
+ *	Apart from that, the error returns are much like io_submit()
+ *	and io_getevents(), since a lot of the same error conditions
+ *	are shared.
+ */
+SYSCALL_DEFINE4(io_ring_enter, aio_context_t, ctx_id, u32, to_submit,
+		u32, min_complete, u32, flags)
+{
+	struct kioctx *ctx;
+	long ret;
+
+	BUILD_BUG_ON(sizeof(struct aio_iocb_ring) != sizeof(struct iocb));
+	BUILD_BUG_ON(sizeof(struct aio_io_event_ring) !=
+			sizeof(struct io_event));
+
+	ctx = lookup_ioctx(ctx_id);
+	if (!ctx) {
+		pr_debug("EINVAL: invalid context id\n");
+		return -EINVAL;
+	}
+
+	ret = -EBUSY;
+	if (!mutex_trylock(&ctx->getevents_lock))
+		goto err;
+
+	ret = -EOVERFLOW;
+	if (ctx->cq_ring_overflow) {
+		ctx->cq_ring_overflow = 0;
+		goto err_unlock;
+	}
+
+	ret = -EINVAL;
+	if (unlikely(atomic_read(&ctx->dead)))
+		goto err_unlock;
+
+	if (ctx->flags & IOCTX_FLAG_SCQRING)
+		ret = __io_ring_enter(ctx, to_submit, min_complete, flags);
+
+err_unlock:
+	mutex_unlock(&ctx->getevents_lock);
+err:
+	percpu_ref_put(&ctx->users);
+	return ret;
+}
+
 /* sys_io_submit:
  *	Queue the nr iocbs pointed to by iocbpp for processing.  Returns
  *	the number of iocbs queued.  May return -EINVAL if the aio_context
@@ -2749,6 +3128,10 @@ SYSCALL_DEFINE3(io_submit, aio_context_t, ctx_id, long, nr,
 		return -EINVAL;
 	}
 
+	/* SCQRING must use io_ring_enter() */
+	if (ctx->flags & IOCTX_FLAG_SCQRING)
+		return -EINVAL;
+
 	if (nr > ctx->nr_events)
 		nr = ctx->nr_events;
 
@@ -2865,7 +3248,7 @@ SYSCALL_DEFINE3(io_cancel, aio_context_t, ctx_id, struct iocb __user *, iocb,
 	if (unlikely(!ctx))
 		return -EINVAL;
 
-	if (ctx->flags & (IOCTX_FLAG_USERIOCB | IOCTX_FLAG_IOPOLL))
+	if (!aio_ctx_supports_cancel(ctx))
 		goto err;
 
 	spin_lock_irq(&ctx->ctx_lock);
@@ -2900,7 +3283,10 @@ static long do_io_getevents(aio_context_t ctx_id,
 	long ret = -EINVAL;
 
 	if (likely(ioctx)) {
-		if (likely(min_nr <= nr && min_nr >= 0)) {
+		/* SCQRING must use io_ring_enter() */
+		if (ioctx->flags & IOCTX_FLAG_SCQRING)
+			ret = -EINVAL;
+		else if (min_nr <= nr && min_nr >= 0) {
 			if (ioctx->flags & IOCTX_FLAG_IOPOLL)
 				ret = aio_iopoll_check(ioctx, min_nr, nr, events);
 			else
diff --git a/include/linux/syscalls.h b/include/linux/syscalls.h
index a20a663d583f..576725d00020 100644
--- a/include/linux/syscalls.h
+++ b/include/linux/syscalls.h
@@ -288,8 +288,10 @@ static inline void addr_limit_user_check(void)
 #ifndef CONFIG_ARCH_HAS_SYSCALL_WRAPPER
 asmlinkage long sys_io_setup(unsigned nr_reqs, aio_context_t __user *ctx);
 asmlinkage long sys_io_setup2(unsigned, unsigned, struct iocb __user *,
-				void __user *, void __user *,
+				struct aio_iocb_ring __user *,
+				struct aio_io_event_ring __user *,
 				aio_context_t __user *);
+asmlinkage long sys_io_ring_enter(aio_context_t, unsigned, unsigned, unsigned);
 asmlinkage long sys_io_destroy(aio_context_t ctx);
 asmlinkage long sys_io_submit(aio_context_t, long,
 			struct iocb __user * __user *);
diff --git a/include/uapi/linux/aio_abi.h b/include/uapi/linux/aio_abi.h
index 05d72cf86bd3..9fb7d0ec868f 100644
--- a/include/uapi/linux/aio_abi.h
+++ b/include/uapi/linux/aio_abi.h
@@ -111,6 +111,32 @@ struct iocb {
 #define IOCTX_FLAG_USERIOCB	(1 << 0)	/* iocbs are user mapped */
 #define IOCTX_FLAG_IOPOLL	(1 << 1)	/* io_context is polled */
 #define IOCTX_FLAG_FIXEDBUFS	(1 << 2)	/* IO buffers are fixed */
+#define IOCTX_FLAG_SCQRING	(1 << 3)	/* Use SQ/CQ rings */
+
+struct aio_iocb_ring {
+	union {
+		struct {
+			u32 head, tail;
+			u32 nr_events;
+		};
+		struct iocb pad_iocb;
+	};
+	struct iocb iocbs[0];
+};
+
+struct aio_io_event_ring {
+	union {
+		struct {
+			u32 head, tail;
+			u32 nr_events;
+		};
+		struct io_event pad_event;
+	};
+	struct io_event events[0];
+};
+
+#define IORING_FLAG_SUBMIT	(1 << 0)
+#define IORING_FLAG_GETEVENTS	(1 << 1)
 
 #undef IFBIG
 #undef IFLITTLE
diff --git a/kernel/sys_ni.c b/kernel/sys_ni.c
index 17c8b4393669..a32b7ea93838 100644
--- a/kernel/sys_ni.c
+++ b/kernel/sys_ni.c
@@ -38,6 +38,7 @@ asmlinkage long sys_ni_syscall(void)
 
 COND_SYSCALL(io_setup);
 COND_SYSCALL(io_setup2);
+COND_SYSCALL(io_ring_enter);
 COND_SYSCALL_COMPAT(io_setup);
 COND_SYSCALL(io_destroy);
 COND_SYSCALL(io_submit);
-- 
2.17.1

[PATCH 27/27] aio: support kernel side submission for aio with SCQRING

[Jens Axboe]

Add support for backing the io_context with either a thread, or a
workqueue and letting those handle the submission for us. This can
be used to reduce overhead for submission, or to always make submission
async. The latter is particularly useful for buffered aio, which is
now fully async with this feature.

For polled IO, we could have the kernel side thread hammer on the SQ
ring and submit when it finds IO. This would mean that an application
would NEVER have to enter the kernel to do IO! Didn't add this yet,
but it would be trivial to add.

If an application sets IOCTX_FLAG_SCQTHREAD, the io_context gets a
single thread backing. If used with buffered IO, this will limit
the device queue depth to 1, but it will be async, IOs will simply
be serialized.

Or an application can set IOCTX_FLAG_SQWQ, in which case the io_context
gets a work queue backing. The concurrency level is the mininum of
twice the available CPUs, or the queue depth specific for the context.
For this mode, we attempt to do buffered reads inline, in case they are
cached. So we should only punt to a workqueue, if we would have to block
to get our data.

Tested with polling, no polling, fixedbufs, no fixedbufs, buffered,
O_DIRECT.

See the sample application for how to use it:

http://git.kernel.dk/cgit/fio/plain/t/aio-ring.c

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/aio.c                     | 416 ++++++++++++++++++++++++++++++++---
 include/uapi/linux/aio_abi.h |   3 +
 2 files changed, 389 insertions(+), 30 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index f4f39a7f8f94..44284b1f4ec9 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -25,6 +25,7 @@
 #include <linux/sched/signal.h>
 #include <linux/fs.h>
 #include <linux/file.h>
+#include <linux/fdtable.h>
 #include <linux/mm.h>
 #include <linux/mman.h>
 #include <linux/mmu_context.h>
@@ -43,6 +44,7 @@
 #include <linux/percpu-refcount.h>
 #include <linux/mount.h>
 #include <linux/sizes.h>
+#include <linux/sched/mm.h>
 
 #include <asm/kmap_types.h>
 #include <linux/uaccess.h>
@@ -103,6 +105,14 @@ struct aio_mapped_ubuf {
 	unsigned int nr_bvecs;
 };
 
+struct aio_sq_offload {
+	struct task_struct *thread;	/* if using a thread */
+	struct workqueue_struct *wq;	/* wq offload */
+	struct mm_struct *mm;
+	struct files_struct *files;
+	wait_queue_head_t wait;
+};
+
 struct kioctx {
 	struct percpu_ref	users;
 	atomic_t		dead;
@@ -146,6 +156,10 @@ struct kioctx {
 	struct aio_mapped_range	sq_ring;
 	struct aio_mapped_range	cq_ring;
 	int			cq_ring_overflow;
+	int			submit_eagain;
+
+	/* sq ring submitter thread, if used */
+	struct aio_sq_offload	sq_offload;
 
 	struct rcu_work		free_rwork;	/* see free_ioctx() */
 
@@ -236,6 +250,7 @@ struct aio_kiocb {
 	unsigned long		ki_flags;
 #define KIOCB_F_POLL_COMPLETED	0	/* polled IO has completed */
 #define KIOCB_F_POLL_EAGAIN	1	/* polled submission got EAGAIN */
+#define KIOCB_F_FORCE_NONBLOCK	2	/* inline submission attempt */
 
 	refcount_t		ki_refcnt;
 
@@ -1354,19 +1369,31 @@ static void aio_complete(struct aio_kiocb *iocb, long res, long res2)
 		unsigned int tail;
 
 		/*
-		 * If we can't get a cq entry, userspace overflowed the
-		 * submission (by quite a lot). Flag it as an overflow
-		 * condition, and next io_ring_enter(2) call will return
-		 * -EOVERFLOW.
+		 * Catch EAGAIN early if we've forced a nonblock attempt, as
+		 * we don't want to pass that back down to userspace through
+		 * the CQ ring. Just mark the ctx as such, so the caller will
+		 * see it and punt to workqueue. This is just for buffered
+		 * aio reads.
 		 */
-		spin_lock_irqsave(&ctx->completion_lock, flags);
-		ev = aio_peek_cqring(ctx, &tail);
-		if (ev) {
-			aio_fill_event(ev, iocb, res, res2);
-			aio_commit_cqring(ctx, tail);
-		} else
-			ctx->cq_ring_overflow = 1;
-		spin_unlock_irqrestore(&ctx->completion_lock, flags);
+		if (res == -EAGAIN &&
+		    test_bit(KIOCB_F_FORCE_NONBLOCK, &iocb->ki_flags)) {
+			ctx->submit_eagain = 1;
+		} else {
+			/*
+			 * If we can't get a cq entry, userspace overflowed the
+			 * submission (by quite a lot). Flag it as an overflow
+			 * condition, and next io_ring_enter(2) call will return
+			 * -EOVERFLOW.
+			 */
+			spin_lock_irqsave(&ctx->completion_lock, flags);
+			ev = aio_peek_cqring(ctx, &tail);
+			if (ev) {
+				aio_fill_event(ev, iocb, res, res2);
+				aio_commit_cqring(ctx, tail);
+			} else
+				ctx->cq_ring_overflow = 1;
+			spin_unlock_irqrestore(&ctx->completion_lock, flags);
+		}
 	} else {
 		aio_ring_complete(ctx, iocb, res, res2);
 
@@ -1768,6 +1795,63 @@ static struct iocb *aio_iocb_from_index(struct kioctx *ctx, int index)
 	return iocb;
 }
 
+static int aio_sq_thread(void *);
+
+static int aio_sq_thread_start(struct kioctx *ctx, struct aio_iocb_ring *ring)
+{
+	struct aio_sq_offload *aso = &ctx->sq_offload;
+	int ret;
+
+	memset(aso, 0, sizeof(*aso));
+	init_waitqueue_head(&aso->wait);
+
+	if (!(ctx->flags & IOCTX_FLAG_FIXEDBUFS))
+		aso->mm = current->mm;
+
+	ret = -EBADF;
+	aso->files = get_files_struct(current);
+	if (!aso->files)
+		goto err;
+
+	if (ctx->flags & IOCTX_FLAG_SQTHREAD) {
+		char name[32];
+
+		snprintf(name, sizeof(name), "aio-sq-%lu/%d", ctx->user_id,
+					ring->sq_thread_cpu);
+		aso->thread = kthread_create_on_cpu(aio_sq_thread, ctx,
+						ring->sq_thread_cpu, name);
+		if (IS_ERR(aso->thread)) {
+			ret = PTR_ERR(aso->thread);
+			aso->thread = NULL;
+			goto err;
+		}
+		wake_up_process(aso->thread);
+	} else if (ctx->flags & IOCTX_FLAG_SQWQ) {
+		int concurrency;
+
+		/* Do QD, or 2 * CPUS, whatever is smallest */
+		concurrency = min(ring->nr_events - 1, 2 * num_online_cpus());
+		aso->wq = alloc_workqueue("aio-sq-%lu",
+						WQ_UNBOUND | WQ_FREEZABLE | WQ_SYSFS,
+						concurrency,
+						ctx->user_id);
+		if (!aso->wq) {
+			ret = -ENOMEM;
+			goto err;
+		}
+	}
+
+	return 0;
+err:
+	if (aso->files) {
+		put_files_struct(aso->files);
+		aso->files = NULL;
+	}
+	if (aso->mm)
+		aso->mm = NULL;
+	return ret;
+}
+
 static void aio_unmap_range(struct aio_mapped_range *range)
 {
 	int i;
@@ -1825,6 +1909,19 @@ static int aio_useriocb_map(struct kioctx *ctx, struct iocb __user *iocbs)
 
 static void aio_scqring_unmap(struct kioctx *ctx)
 {
+	struct aio_sq_offload *aso = &ctx->sq_offload;
+
+	if (aso->thread) {
+		kthread_stop(aso->thread);
+		aso->thread = NULL;
+	} else if (aso->wq) {
+		destroy_workqueue(aso->wq);
+		aso->wq = NULL;
+	}
+	if (aso->files) {
+		put_files_struct(aso->files);
+		aso->files = NULL;
+	}
 	aio_unmap_range(&ctx->sq_ring);
 	aio_unmap_range(&ctx->cq_ring);
 }
@@ -1854,10 +1951,8 @@ static int aio_scqring_map(struct kioctx *ctx,
 
 	ret = aio_map_range(&ctx->cq_ring, cq_ring,
 			    cq_ring_size * sizeof(struct io_event), FOLL_WRITE);
-	if (ret) {
-		aio_unmap_range(&ctx->sq_ring);
-		return ret;
-	}
+	if (ret)
+		goto err;
 
 	ksq_ring = page_address(ctx->sq_ring.pages[0]);
 	ksq_ring->nr_events = sq_ring_size;
@@ -1866,7 +1961,16 @@ static int aio_scqring_map(struct kioctx *ctx,
 	kcq_ring = page_address(ctx->cq_ring.pages[0]);
 	kcq_ring->nr_events = cq_ring_size;
 	kcq_ring->head = kcq_ring->tail = 0;
-	return 0;
+
+	if (ctx->flags & (IOCTX_FLAG_SQTHREAD | IOCTX_FLAG_SQWQ))
+		ret = aio_sq_thread_start(ctx, ksq_ring);
+
+err:
+	if (ret) {
+		aio_unmap_range(&ctx->sq_ring);
+		aio_unmap_range(&ctx->cq_ring);
+	}
+	return ret;
 }
 
 static void aio_iocb_buffer_unmap(struct kioctx *ctx)
@@ -2007,7 +2111,8 @@ SYSCALL_DEFINE6(io_setup2, u32, nr_events, u32, flags,
 	long ret;
 
 	if (flags & ~(IOCTX_FLAG_USERIOCB | IOCTX_FLAG_IOPOLL |
-		      IOCTX_FLAG_FIXEDBUFS | IOCTX_FLAG_SCQRING))
+		      IOCTX_FLAG_FIXEDBUFS | IOCTX_FLAG_SCQRING |
+		      IOCTX_FLAG_SQTHREAD | IOCTX_FLAG_SQWQ))
 		return -EINVAL;
 
 	ret = get_user(ctx, ctxp);
@@ -2249,7 +2354,7 @@ static struct file *aio_file_get(struct aio_submit_state *state, int fd)
 }
 
 static int aio_prep_rw(struct aio_kiocb *kiocb, const struct iocb *iocb,
-		       struct aio_submit_state *state)
+		       struct aio_submit_state *state, bool force_nonblock)
 {
 	struct kioctx *ctx = kiocb->ki_ctx;
 	struct kiocb *req = &kiocb->rw;
@@ -2282,6 +2387,10 @@ static int aio_prep_rw(struct aio_kiocb *kiocb, const struct iocb *iocb,
 	ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
 	if (unlikely(ret))
 		goto out_fput;
+	if (force_nonblock) {
+		req->ki_flags |= IOCB_NOWAIT;
+		set_bit(KIOCB_F_FORCE_NONBLOCK, &kiocb->ki_flags);
+	}
 
 	if (iocb->aio_flags & IOCB_FLAG_HIPRI) {
 		/* shares space in the union, and is rather pointless.. */
@@ -2422,7 +2531,7 @@ static void aio_iopoll_iocb_issued(struct aio_submit_state *state,
 
 static ssize_t aio_read(struct aio_kiocb *kiocb, const struct iocb *iocb,
 			struct aio_submit_state *state, bool vectored,
-			bool compat, bool kaddr)
+			bool compat, bool kaddr, bool force_nonblock)
 {
 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
 	struct kiocb *req = &kiocb->rw;
@@ -2430,7 +2539,7 @@ static ssize_t aio_read(struct aio_kiocb *kiocb, const struct iocb *iocb,
 	struct file *file;
 	ssize_t ret;
 
-	ret = aio_prep_rw(kiocb, iocb, state);
+	ret = aio_prep_rw(kiocb, iocb, state, force_nonblock);
 	if (ret)
 		return ret;
 	file = req->ki_filp;
@@ -2467,7 +2576,7 @@ static ssize_t aio_write(struct aio_kiocb *kiocb, const struct iocb *iocb,
 	struct file *file;
 	ssize_t ret;
 
-	ret = aio_prep_rw(kiocb, iocb, state);
+	ret = aio_prep_rw(kiocb, iocb, state, false);
 	if (ret)
 		return ret;
 	file = req->ki_filp;
@@ -2720,7 +2829,7 @@ static ssize_t aio_poll(struct aio_kiocb *aiocb, const struct iocb *iocb)
 static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
 			   struct iocb __user *user_iocb,
 			   struct aio_submit_state *state, bool compat,
-			   bool kaddr)
+			   bool kaddr, bool force_nonblock)
 {
 	struct aio_kiocb *req;
 	ssize_t ret;
@@ -2779,13 +2888,15 @@ static int __io_submit_one(struct kioctx *ctx, const struct iocb *iocb,
 	ret = -EINVAL;
 	switch (iocb->aio_lio_opcode) {
 	case IOCB_CMD_PREAD:
-		ret = aio_read(req, iocb, state, false, compat, kaddr);
+		ret = aio_read(req, iocb, state, false, compat, kaddr,
+				force_nonblock);
 		break;
 	case IOCB_CMD_PWRITE:
 		ret = aio_write(req, iocb, state, false, compat, kaddr);
 		break;
 	case IOCB_CMD_PREADV:
-		ret = aio_read(req, iocb, state, true, compat, kaddr);
+		ret = aio_read(req, iocb, state, true, compat, kaddr,
+				force_nonblock);
 		break;
 	case IOCB_CMD_PWRITEV:
 		ret = aio_write(req, iocb, state, true, compat, kaddr);
@@ -2857,7 +2968,8 @@ static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
 		iocbp = &iocb;
 	}
 
-	return __io_submit_one(ctx, iocbp, user_iocb, state, compat, kaddr);
+	return __io_submit_one(ctx, iocbp, user_iocb, state, compat, kaddr,
+				false);
 }
 
 #ifdef CONFIG_BLOCK
@@ -2960,7 +3072,8 @@ static int aio_ring_submit(struct kioctx *ctx, unsigned int to_submit)
 		if (!iocb)
 			break;
 
-		ret = __io_submit_one(ctx, iocb, NULL, NULL, false, kaddr);
+		ret = __io_submit_one(ctx, iocb, NULL, NULL, false, kaddr,
+					false);
 		if (ret)
 			break;
 
@@ -3013,15 +3126,258 @@ static int aio_cqring_wait(struct kioctx *ctx, int min_events)
 	return ret;
 }
 
+static void aio_fill_cq_error(struct kioctx *ctx, const struct iocb *iocb,
+			      long ret)
+{
+	struct io_event *ev;
+	unsigned tail;
+
+	/*
+	 * Only really need the lock for non-polled IO, but this is an error
+	 * so not worth checking. Just lock it so we know kernel access to
+	 * the CQ ring is serialized.
+	 */
+	spin_lock_irq(&ctx->completion_lock);
+	ev = aio_peek_cqring(ctx, &tail);
+	ev->obj = iocb->aio_data;
+	ev->data = 0;
+	ev->res = ret;
+	ev->res2 = 0;
+	aio_commit_cqring(ctx, tail);
+	spin_unlock_irq(&ctx->completion_lock);
+
+	/*
+	 * for thread offload, app could already be sleeping in io_ring_enter()
+	 * before we get to flag the error. wake them up, if needed.
+	 */
+	if (ctx->flags & (IOCTX_FLAG_SQTHREAD | IOCTX_FLAG_SQWQ))
+		if (waitqueue_active(&ctx->wait))
+			wake_up(&ctx->wait);
+}
+
+struct aio_io_work {
+	struct work_struct work;
+	struct kioctx *ctx;
+	struct iocb iocb;
+};
+
+/*
+ * sq thread only supports O_DIRECT or FIXEDBUFS IO
+ */
+static int aio_sq_thread(void *data)
+{
+	const struct iocb *iocbs[AIO_IOPOLL_BATCH];
+	struct aio_submit_state state;
+	struct kioctx *ctx = data;
+	struct aio_sq_offload *aso = &ctx->sq_offload;
+	struct mm_struct *cur_mm = NULL;
+	struct files_struct *old_files;
+	mm_segment_t old_fs;
+	DEFINE_WAIT(wait);
+
+	old_files = current->files;
+	current->files = aso->files;
+
+	old_fs = get_fs();
+	set_fs(USER_DS);
+
+	while (!kthread_should_stop()) {
+		struct aio_submit_state *statep = NULL;
+		const struct iocb *iocb;
+		bool mm_fault = false;
+		unsigned int nhead;
+		int ret, i, j;
+
+		iocb = aio_peek_sqring(ctx, &nhead);
+		if (!iocb) {
+			prepare_to_wait(&aso->wait, &wait, TASK_INTERRUPTIBLE);
+			iocb = aio_peek_sqring(ctx, &nhead);
+			if (!iocb) {
+				/*
+				 * Drop cur_mm before scheduler. We can't hold
+				 * it for long periods, and it would also
+				 * introduce a deadlock with kill_ioctx().
+				 */
+				if (cur_mm) {
+					unuse_mm(cur_mm);
+					mmput(cur_mm);
+					cur_mm = NULL;
+				}
+				schedule();
+			}
+			finish_wait(&aso->wait, &wait);
+			if (!iocb)
+				continue;
+		}
+
+		/* If ->mm is set, we're not doing FIXEDBUFS */
+		if (aso->mm) {
+			mm_fault = !mmget_not_zero(aso->mm);
+			if (!mm_fault) {
+				use_mm(aso->mm);
+				cur_mm = aso->mm;
+			}
+		}
+
+		i = 0;
+		do {
+			if (i == ARRAY_SIZE(iocbs))
+				break;
+			iocbs[i++] = iocb;
+			aio_commit_sqring(ctx, nhead);
+		} while ((iocb = aio_peek_sqring(ctx, &nhead)) != NULL);
+
+		if (i > AIO_PLUG_THRESHOLD) {
+			aio_submit_state_start(&state, ctx, i);
+			statep = &state;
+		}
+
+		for (j = 0; j < i; j++) {
+			if (unlikely(mm_fault))
+				ret = -EFAULT;
+			else
+				ret = __io_submit_one(ctx, iocbs[j], NULL, NULL,
+							false, !cur_mm, false);
+			if (!ret)
+				continue;
+
+			aio_fill_cq_error(ctx, iocbs[j], ret);
+		}
+
+		if (statep)
+			aio_submit_state_end(&state);
+	}
+	current->files = old_files;
+	set_fs(old_fs);
+	if (cur_mm) {
+		unuse_mm(cur_mm);
+		mmput(cur_mm);
+	}
+	return 0;
+}
+
+static void aio_sq_wq_submit_work(struct work_struct *work)
+{
+	struct aio_io_work *aiw = container_of(work, struct aio_io_work, work);
+	struct kioctx *ctx = aiw->ctx;
+	struct aio_sq_offload *aso = &ctx->sq_offload;
+	mm_segment_t old_fs = get_fs();
+	struct files_struct *old_files;
+	int ret;
+
+	old_files = current->files;
+	current->files = aso->files;
+
+	if (aso->mm) {
+		if (!mmget_not_zero(aso->mm)) {
+			ret = -EFAULT;
+			goto err;
+		}
+		use_mm(aso->mm);
+	}
+
+	set_fs(USER_DS);
+
+	ret = __io_submit_one(ctx, &aiw->iocb, NULL, NULL, false, !aso->mm,
+				false);
+
+	set_fs(old_fs);
+	if (aso->mm) {
+		unuse_mm(aso->mm);
+		mmput(aso->mm);
+	}
+
+err:
+	if (ret)
+		aio_fill_cq_error(ctx, &aiw->iocb, ret);
+	current->files = old_files;
+	kfree(aiw);
+}
+
+/*
+ * If this is a read, try a cached inline read first. If the IO is in the
+ * page cache, we can satisfy it without blocking and without having to
+ * punt to a threaded execution. This is much faster, particularly for
+ * lower queue depth IO, and it's always a lot more efficient.
+ */
+static int aio_sq_try_inline(struct kioctx *ctx, struct aio_io_work *aiw)
+{
+	struct aio_sq_offload *aso = &ctx->sq_offload;
+	int ret;
+
+	if (aiw->iocb.aio_lio_opcode != IOCB_CMD_PREAD &&
+	    aiw->iocb.aio_lio_opcode != IOCB_CMD_PREADV)
+		return -EAGAIN;
+
+	ret = __io_submit_one(ctx, &aiw->iocb, NULL, NULL, false, !aso->mm,
+				true);
+
+	if (ret == -EAGAIN || ctx->submit_eagain) {
+		ctx->submit_eagain = 0;
+		return -EAGAIN;
+	}
+
+	/*
+	 * We're done - even if this was an error, return 0. The error will
+	 * be in the CQ ring for the application.
+	 */
+	kfree(aiw);
+	return 0;
+}
+
+static int aio_sq_wq_submit(struct kioctx *ctx, unsigned int to_submit)
+{
+	struct aio_io_work *work;
+	const struct iocb *iocb;
+	unsigned nhead;
+	int ret, queued;
+
+	ret = queued = 0;
+	while ((iocb = aio_peek_sqring(ctx, &nhead)) != NULL) {
+		work = kmalloc(sizeof(*work), GFP_KERNEL);
+		if (!work) {
+			ret = -ENOMEM;
+			break;
+		}
+		memcpy(&work->iocb, iocb, sizeof(*iocb));
+		aio_commit_sqring(ctx, nhead);
+		ret = aio_sq_try_inline(ctx, work);
+		if (ret == -EAGAIN) {
+			INIT_WORK(&work->work, aio_sq_wq_submit_work);
+			work->ctx = ctx;
+			queue_work(ctx->sq_offload.wq, &work->work);
+			ret = 0;
+		}
+		queued++;
+		if (queued == to_submit)
+			break;
+	}
+
+	return queued ? queued : ret;
+}
+
 static int __io_ring_enter(struct kioctx *ctx, unsigned int to_submit,
 			   unsigned int min_complete, unsigned int flags)
 {
 	int ret = 0;
 
 	if (flags & IORING_FLAG_SUBMIT) {
-		ret = aio_ring_submit(ctx, to_submit);
-		if (ret < 0)
-			return ret;
+		/*
+		 * Three options here:
+		 * 1) We have an sq thread, just wake it up to do submissions
+		 * 2) We have an sq wq, queue a work item for each iocb
+		 * 3) Submit directly
+		 */
+		if (to_submit && (ctx->flags & IOCTX_FLAG_SQTHREAD)) {
+			wake_up(&ctx->sq_offload.wait);
+			ret = to_submit;
+		} else if (to_submit && (ctx->flags & IOCTX_FLAG_SQWQ)) {
+			ret = aio_sq_wq_submit(ctx, to_submit);
+		} else {
+			ret = aio_ring_submit(ctx, to_submit);
+			if (ret < 0)
+				return ret;
+		}
 	}
 	if (flags & IORING_FLAG_GETEVENTS) {
 		unsigned int nr_events = 0;
diff --git a/include/uapi/linux/aio_abi.h b/include/uapi/linux/aio_abi.h
index 9fb7d0ec868f..500b37feeaa8 100644
--- a/include/uapi/linux/aio_abi.h
+++ b/include/uapi/linux/aio_abi.h
@@ -112,12 +112,15 @@ struct iocb {
 #define IOCTX_FLAG_IOPOLL	(1 << 1)	/* io_context is polled */
 #define IOCTX_FLAG_FIXEDBUFS	(1 << 2)	/* IO buffers are fixed */
 #define IOCTX_FLAG_SCQRING	(1 << 3)	/* Use SQ/CQ rings */
+#define IOCTX_FLAG_SQTHREAD	(1 << 4)	/* Use SQ thread */
+#define IOCTX_FLAG_SQWQ		(1 << 5)	/* Use SQ workqueue */
 
 struct aio_iocb_ring {
 	union {
 		struct {
 			u32 head, tail;
 			u32 nr_events;
+			u32 sq_thread_cpu;
 		};
 		struct iocb pad_iocb;
 	};
-- 
2.17.1