[PATCHSET v3 0/18] Support for polled aio

[PATCHSET v3 0/18] Support for polled aio

[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/

The patchset continues to evolve, and has grown some optimizations that
benefit non-polled aio as well. One such example is the user mapped
iocbs, which avoid copying a full iocb for each IO. If you look at
memory performance, we copy 72b for each IO at submission time (8 byte
pointer, 64 byte iocb), then copy back 16 bytes at completion for the
io_event. That's 88 bytes of memory copies, for something that's
potentially a 512b IO.

In terms of observed performance, I've been able to do 1070K 4k IOPS
with this on a single device, using a single thread for both IO
submissions and completions. For actual real life use cases, on a
flash storage box permance is up 25% in terms of IOPS/throughput.

As before, patches are on top of my mq-perf branch

Changes since v2:

- Fix an off-by-one causing QD=1 aio polls to sometimes stall.
- Add fget_many/fput_many optimization.
- Ensure that io_cancel() is never called on a polled iocb.
- Add support for user mapped iocbs.
- IOCTX_FLAG_IOPOLL is now (1 << 1), since it was reordered after
  the user iocb support.
- Kill full aio_kiocb zero on alloc.
- Move fops->iopoll() storage back to the iocb (->ki_cookie). The usage
  in the dio private storage was inherently racy, making the store of
  the submission cookie open to use-after-free.
- Fix narrow use-before-submit race in fops->iopoll() access.
- Batch move submitted iocbs to poll_submitted list.
- Pass back EAGAIN at submission time, don't punt to getevents() time.
- Split some patches up.
- Address review concerns.
- Various little fixes and optimizations.


 Documentation/filesystems/vfs.txt      |   3 +
 arch/x86/entry/syscalls/syscall_64.tbl |   1 +
 fs/aio.c                               | 841 +++++++++++++++++++++----
 fs/block_dev.c                         |  20 +-
 fs/file.c                              |  15 +-
 fs/file_table.c                        |  10 +-
 fs/gfs2/file.c                         |   2 +
 fs/iomap.c                             |  51 +-
 fs/xfs/xfs_file.c                      |   1 +
 include/linux/file.h                   |   2 +
 include/linux/fs.h                     |   5 +-
 include/linux/iomap.h                  |   1 +
 include/linux/syscalls.h               |   2 +
 include/uapi/asm-generic/unistd.h      |   4 +-
 include/uapi/linux/aio_abi.h           |   5 +
 kernel/sys_ni.c                        |   1 +
 16 files changed, 819 insertions(+), 145 deletions(-)

-- 
Jens Axboe

[PATCH 01/20] aio: fix failure to put the file pointer

[Jens Axboe]

If the ioprio capability check fails, we return without putting
the file pointer.

Fixes: d9a08a9e616b ("fs: Add aio iopriority support")
Reviewed-by: Christoph Hellwig <hch@lst.de>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/aio.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/fs/aio.c b/fs/aio.c
index b984918be4b7..205390c0c1bb 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -1436,6 +1436,7 @@ static int aio_prep_rw(struct kiocb *req, struct iocb *iocb)
 		ret = ioprio_check_cap(iocb->aio_reqprio);
 		if (ret) {
 			pr_debug("aio ioprio check cap error: %d\n", ret);
+			fput(req->ki_filp);
 			return ret;
 		}
 
-- 
2.17.1

[PATCH 02/20] aio: clear IOCB_HIPRI

[Jens Axboe]

From: Christoph Hellwig <hch@lst.de>

No one is going to poll for aio (yet), so we must clear the HIPRI
flag, as we would otherwise send it down the poll queues, where no
one will be polling for completions.

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

IOCB_HIPRI, not RWF_HIPRI.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/aio.c | 11 ++++++++---
 1 file changed, 8 insertions(+), 3 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index 205390c0c1bb..05647d352bf3 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -1436,8 +1436,7 @@ static int aio_prep_rw(struct kiocb *req, struct iocb *iocb)
 		ret = ioprio_check_cap(iocb->aio_reqprio);
 		if (ret) {
 			pr_debug("aio ioprio check cap error: %d\n", ret);
-			fput(req->ki_filp);
-			return ret;
+			goto out_fput;
 		}
 
 		req->ki_ioprio = iocb->aio_reqprio;
@@ -1446,7 +1445,13 @@ static int aio_prep_rw(struct kiocb *req, struct iocb *iocb)
 
 	ret = kiocb_set_rw_flags(req, iocb->aio_rw_flags);
 	if (unlikely(ret))
-		fput(req->ki_filp);
+		goto out_fput;
+
+	req->ki_flags &= ~IOCB_HIPRI; /* no one is going to poll for this I/O */
+	return 0;
+
+out_fput:
+	fput(req->ki_filp);
 	return ret;
 }
 
-- 
2.17.1

[PATCH 03/20] 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.

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 04/20] 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.

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 d233a59ea364..9acb26f7c9ec 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;
@@ -396,6 +404,7 @@ __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;
 		}
 
@@ -2061,6 +2070,7 @@ 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 05/20] block: ensure that async polled IO is marked REQ_NOWAIT

[Jens Axboe]

We can't wait for polled events to complete, as they may require active
polling from whoever submitted it. If that is the same task that is
submitting new IO, we could deadlock waiting for IO to complete that
this task is supposed to be completing itself.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/block_dev.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/fs/block_dev.c b/fs/block_dev.c
index 9acb26f7c9ec..9580ffe0cf24 100644
--- a/fs/block_dev.c
+++ b/fs/block_dev.c
@@ -400,8 +400,16 @@ __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)
+			if (iocb->ki_flags & IOCB_HIPRI) {
 				bio->bi_opf |= REQ_HIPRI;
+				/*
+				 * For async polled IO, we can't wait for
+				 * requests to complete, as they may also be
+				 * polled and require active reaping.
+				 */
+				if (!is_sync)
+					bio->bi_opf |= REQ_NOWAIT;
+			}
 
 			qc = submit_bio(bio);
 			WRITE_ONCE(iocb->ki_cookie, qc);
-- 
2.17.1

[PATCH 06/20] 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>
Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/gfs2/file.c        |  2 ++
 fs/iomap.c            | 43 ++++++++++++++++++++++++++++---------------
 fs/xfs/xfs_file.c     |  1 +
 include/linux/iomap.h |  1 +
 4 files changed, 32 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 74c1f37f0fd6..16787b3b09fd 100644
--- a/fs/iomap.c
+++ b/fs/iomap.c
@@ -1436,6 +1436,28 @@ 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)
+		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;
@@ -1548,7 +1570,7 @@ 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)
 {
@@ -1562,15 +1584,10 @@ 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
@@ -1666,9 +1683,6 @@ 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;
@@ -1676,11 +1690,7 @@ 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);
 
 	if (need_zeroout) {
@@ -1883,6 +1893,9 @@ 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 53c9ab8fb777..603e705781a4 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 07/20] iomap: ensure that async polled IO is marked REQ_NOWAIT

[Jens Axboe]

We can't wait for polled events to complete, as they may require active
polling from whoever submitted it. If that is the same task that is
submitting new IO, we could deadlock waiting for IO to complete that
this task is supposed to be completing itself.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/iomap.c | 10 +++++++++-
 1 file changed, 9 insertions(+), 1 deletion(-)

diff --git a/fs/iomap.c b/fs/iomap.c
index 16787b3b09fd..96d60b9b2bea 100644
--- a/fs/iomap.c
+++ b/fs/iomap.c
@@ -1451,8 +1451,16 @@ static void iomap_dio_submit_bio(struct iomap_dio *dio, struct iomap *iomap,
 {
 	atomic_inc(&dio->ref);
 
-	if (dio->iocb->ki_flags & IOCB_HIPRI)
+	if (dio->iocb->ki_flags & IOCB_HIPRI) {
 		bio->bi_opf |= REQ_HIPRI;
+		/*
+		 * For async polled IO, we can't wait for requests to
+		 * complete, as they may also be polled and require active
+		 * reaping.
+		 */
+		if (!dio->wait_for_completion)
+			bio->bi_opf |= REQ_NOWAIT;
+	}
 
 	dio->submit.last_queue = bdev_get_queue(iomap->bdev);
 	dio->submit.cookie = submit_bio(bio);
-- 
2.17.1

[PATCH 08/20] 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 09/20] aio: separate out ring reservation from req allocation

[Jens Axboe]

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

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 10/20] 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.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/aio.c | 19 ++++++++++++-------
 1 file changed, 12 insertions(+), 7 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index eaceb40e6cf5..681f2072f81b 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);
-	if (unlikely(!req))
-		return NULL;
+	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
+	if (req) {
+		percpu_ref_get(&ctx->reqs);
+		req->ki_ctx = ctx;
+		INIT_LIST_HEAD(&req->ki_list);
+		refcount_set(&req->ki_refcnt, 0);
+		req->ki_eventfd = NULL;
+	}
 
-	percpu_ref_get(&ctx->reqs);
-	INIT_LIST_HEAD(&req->ki_list);
-	refcount_set(&req->ki_refcnt, 0);
-	req->ki_ctx = ctx;
 	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 11/20] 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.

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 681f2072f81b..533cb7b1112f 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -1887,6 +1887,12 @@ static int io_submit_one(struct kioctx *ctx, struct iocb __user *user_iocb,
 	return ret;
 }
 
+/*
+ * 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
+
 /* 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
@@ -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 12/20] 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.

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 533cb7b1112f..e8457f9486e3 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -1878,10 +1878,9 @@ 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 13/20] 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 e8457f9486e3..ba5758c854e8 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -1414,7 +1414,7 @@ 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;
 
@@ -1455,7 +1455,7 @@ 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;
@@ -1494,8 +1494,8 @@ 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;
@@ -1527,8 +1527,8 @@ 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;
@@ -1583,7 +1583,8 @@ 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))
@@ -1711,7 +1712,7 @@ 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;
@@ -1783,27 +1784,23 @@ 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;
@@ -1817,14 +1814,14 @@ 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;
@@ -1839,32 +1836,32 @@ 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;
 	}
@@ -1886,6 +1883,17 @@ 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);
+}
+
 /*
  * 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.
-- 
2.17.1

[PATCH 14/20] 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 ba5758c854e8..12859ea1cb64 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -1057,6 +1057,15 @@ 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.
  */
@@ -1084,10 +1093,7 @@ 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 15/20] 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 that,
we pass in an iocb array for future use.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 arch/x86/entry/syscalls/syscall_64.tbl |  1 +
 fs/aio.c                               | 70 ++++++++++++++++----------
 include/linux/syscalls.h               |  2 +
 include/uapi/asm-generic/unistd.h      |  4 +-
 kernel/sys_ni.c                        |  1 +
 5 files changed, 50 insertions(+), 28 deletions(-)

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 12859ea1cb64..74831ce2185e 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -94,6 +94,8 @@ struct kioctx {
 
 	unsigned long		user_id;
 
+	unsigned int		flags;
+
 	struct __percpu kioctx_cpu *cpu;
 
 	/*
@@ -680,21 +682,24 @@ 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;
 	int err = -ENOMEM;
-
 	/*
 	 * Store the original nr_events -- what userspace passed to io_setup(),
 	 * for counting against the global limit -- before it changes.
 	 */
 	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.
@@ -720,6 +725,7 @@ 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);
@@ -1275,6 +1281,33 @@ static long read_events(struct kioctx *ctx, long min_nr, long nr,
 	return ret;
 }
 
+SYSCALL_DEFINE4(io_setup2, u32, nr_events, u32, flags, struct iocb * __user,
+		iocbs, aio_context_t __user *, ctxp)
+{
+	struct kioctx *ioctx;
+	unsigned long ctx;
+	long ret;
+
+	if (flags)
+		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
@@ -1290,7 +1323,7 @@ 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;
 
@@ -1298,14 +1331,7 @@ 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);
@@ -1313,7 +1339,6 @@ SYSCALL_DEFINE2(io_setup, unsigned, nr_events, aio_context_t __user *, ctxp)
 			kill_ioctx(current->mm, ioctx, NULL);
 		percpu_ref_put(&ioctx->users);
 	}
-
 out:
 	return ret;
 }
@@ -1321,7 +1346,7 @@ 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;
 
@@ -1329,23 +1354,14 @@ 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..b661e78717e6 100644
--- a/include/linux/syscalls.h
+++ b/include/linux/syscalls.h
@@ -287,6 +287,8 @@ 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 *,
+				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 538546edbfbd..b4527ed373b0 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 16/20] 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                     | 99 +++++++++++++++++++++++++++++++-----
 include/uapi/linux/aio_abi.h |  2 +
 2 files changed, 89 insertions(+), 12 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index 74831ce2185e..e98121df92f6 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -121,6 +121,9 @@ struct kioctx {
 	struct page		**ring_pages;
 	long			nr_pages;
 
+	struct page		**iocb_pages;
+	long			iocb_nr_pages;
+
 	struct rcu_work		free_rwork;	/* see free_ioctx() */
 
 	/*
@@ -216,6 +219,11 @@ 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_free(struct kioctx *);
+
 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
 {
 	struct file *file;
@@ -572,6 +580,7 @@ static void free_ioctx(struct work_struct *work)
 					  free_rwork);
 	pr_debug("freeing %p\n", ctx);
 
+	aio_useriocb_free(ctx);
 	aio_free_ring(ctx);
 	free_percpu(ctx->cpu);
 	percpu_ref_exit(&ctx->reqs);
@@ -1281,6 +1290,45 @@ static long read_events(struct kioctx *ctx, long min_nr, long nr,
 	return ret;
 }
 
+static void aio_useriocb_free(struct kioctx *ctx)
+{
+	int i;
+
+	if (!ctx->iocb_nr_pages)
+		return;
+
+	for (i = 0; i < ctx->iocb_nr_pages; i++)
+		put_page(ctx->iocb_pages[i]);
+
+	kfree(ctx->iocb_pages);
+}
+
+static int aio_useriocb_map(struct kioctx *ctx, struct iocb __user *iocbs,
+			    unsigned int nr_events)
+{
+	int nr_pages, ret;
+
+	if ((unsigned long) iocbs & ~PAGE_MASK)
+		return -EINVAL;
+
+	nr_pages = sizeof(struct iocb) * nr_events;
+	nr_pages = (nr_pages + PAGE_SIZE - 1) >> PAGE_SHIFT;
+
+	ctx->iocb_pages = kzalloc(nr_pages * sizeof(struct page *), GFP_KERNEL);
+	if (!ctx->iocb_pages)
+		return -ENOMEM;
+
+	ret = get_user_pages_fast((unsigned long) iocbs, nr_pages, 0,
+					ctx->iocb_pages);
+	if (ret < nr_pages) {
+		kfree(ctx->iocb_pages);
+		return -ENOMEM;
+	}
+
+	ctx->iocb_nr_pages = nr_pages;
+	return 0;
+}
+
 SYSCALL_DEFINE4(io_setup2, u32, nr_events, u32, flags, struct iocb * __user,
 		iocbs, aio_context_t __user *, ctxp)
 {
@@ -1288,7 +1336,7 @@ SYSCALL_DEFINE4(io_setup2, u32, nr_events, u32, flags, struct iocb * __user,
 	unsigned long ctx;
 	long ret;
 
-	if (flags)
+	if (flags & ~IOCTX_FLAG_USERIOCB)
 		return -EINVAL;
 
 	ret = get_user(ctx, ctxp);
@@ -1300,9 +1348,17 @@ SYSCALL_DEFINE4(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, nr_events);
+		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;
@@ -1851,10 +1907,13 @@ 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;
@@ -1908,12 +1967,26 @@ 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;
+		unsigned int page_index;
 
-	return __io_submit_one(ctx, &iocb, user_iocb, compat);
+		if (iocb_index >= ctx->nr_events)
+			return -EINVAL;
+
+		page_index = iocb_index >> iocb_page_shift;
+		iocb_index &= ((1 << iocb_page_shift) - 1);
+		iocbp = page_address(ctx->iocb_pages[page_index]);
+		iocbp += 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);
 }
 
 /*
@@ -2063,6 +2136,9 @@ 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) {
@@ -2079,9 +2155,8 @@ 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 17/20] 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                     | 378 +++++++++++++++++++++++++++++++----
 include/uapi/linux/aio_abi.h |   3 +
 2 files changed, 345 insertions(+), 36 deletions(-)

diff --git a/fs/aio.c b/fs/aio.c
index e98121df92f6..db73c8af1a0a 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -143,6 +143,18 @@ 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 */
@@ -195,14 +207,27 @@ 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 IOCB_POLL_COMPLETED	0
+#define IOCB_POLL_BUSY		1
+
 	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----*/
@@ -223,6 +248,7 @@ static const unsigned int iocb_page_shift =
 				ilog2(PAGE_SIZE / sizeof(struct iocb));
 
 static void aio_useriocb_free(struct kioctx *);
+static void aio_iopoll_reap_events(struct kioctx *);
 
 static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
 {
@@ -461,11 +487,15 @@ 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)
@@ -747,6 +777,11 @@ 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;
 
@@ -818,11 +853,15 @@ 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]));
@@ -1029,6 +1068,7 @@ 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;
 	}
@@ -1072,6 +1112,15 @@ 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)
 {
@@ -1261,6 +1310,166 @@ 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;
+
+	list_for_each_entry_safe(iocb, n, &ctx->poll_completing, ki_list) {
+		if (*nr_events == max)
+			break;
+		if (!test_bit(IOCB_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;
+}
+
+static int __aio_iopoll_check(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 (*nr_events >= min)
+			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 (*nr_events >= min)
+		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(IOCB_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_check(ctx, NULL, &nr_events, 1, UINT_MAX);
+	}
+}
+
+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 = 0;
+
+	/* 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;
+	}
+
+	while (!nr_events || !need_resched()) {
+		int tmin = 0;
+
+		if (nr_events < min_nr)
+			tmin = min_nr - nr_events;
+
+		ret = __aio_iopoll_check(ctx, event, &nr_events, tmin, nr);
+		if (ret <= 0)
+			break;
+		ret = 0;
+	}
+
+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)
@@ -1336,7 +1545,7 @@ SYSCALL_DEFINE4(io_setup2, u32, nr_events, u32, flags, struct iocb * __user,
 	unsigned long ctx;
 	long ret;
 
-	if (flags & ~IOCTX_FLAG_USERIOCB)
+	if (flags & ~(IOCTX_FLAG_USERIOCB | IOCTX_FLAG_IOPOLL))
 		return -EINVAL;
 
 	ret = get_user(ctx, ctxp);
@@ -1469,13 +1678,8 @@ 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);
 
@@ -1487,19 +1691,48 @@ 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(IOCB_POLL_BUSY, &iocb->ki_flags);
+	else {
+		aio_fill_event(&iocb->ki_ev, iocb, res, res2);
+		set_bit(IOCB_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)
@@ -1525,9 +1758,35 @@ 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;
@@ -1570,17 +1829,43 @@ static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
 	default:
 		req->ki_complete(req, ret, 0);
 	}
+
 }
 
-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_add = test_bit(IOCB_POLL_COMPLETED, &kiocb->ki_flags);
+	struct kioctx *ctx = kiocb->ki_ctx;
+
+	spin_lock(&ctx->poll_lock);
+	if (front_add)
+		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;
@@ -1605,15 +1890,16 @@ 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;
@@ -1884,7 +2170,8 @@ 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;
@@ -1907,8 +2194,8 @@ 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");
@@ -1919,26 +2206,33 @@ 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:
@@ -1954,13 +2248,21 @@ 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(IOCB_POLL_BUSY, &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;
 }
 
@@ -2136,7 +2438,7 @@ 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);
@@ -2171,8 +2473,12 @@ 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 18/20] 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 db73c8af1a0a..74afd1ff7fc9 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -247,6 +247,15 @@ 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_free(struct kioctx *);
 static void aio_iopoll_reap_events(struct kioctx *);
 
@@ -1724,6 +1733,21 @@ static void aio_complete_rw_poll(struct kiocb *kiocb, long res, long res2)
 	}
 }
 
+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;
+};
+
 static int aio_prep_rw(struct aio_kiocb *kiocb, const struct iocb *iocb)
 {
 	struct kioctx *ctx = kiocb->ki_ctx;
@@ -1832,13 +1856,28 @@ 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
@@ -1848,12 +1887,21 @@ static void aio_iopoll_iocb_issued(struct aio_kiocb *kiocb)
 	const int front_add = test_bit(IOCB_POLL_COMPLETED, &kiocb->ki_flags);
 	struct kioctx *ctx = kiocb->ki_ctx;
 
-	spin_lock(&ctx->poll_lock);
-	if (front_add)
-		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_add)
+			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_add)
+			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,
@@ -2149,7 +2197,8 @@ 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;
@@ -2253,7 +2302,7 @@ 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:
@@ -2267,7 +2316,7 @@ 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;
 
@@ -2288,7 +2337,44 @@ 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
 }
 
 /*
@@ -2312,10 +2398,10 @@ 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;
@@ -2329,8 +2415,10 @@ 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;
 
@@ -2339,12 +2427,12 @@ 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;
@@ -2354,10 +2442,10 @@ 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;
@@ -2371,8 +2459,10 @@ 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;
 
@@ -2381,12 +2471,12 @@ 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 19/20] 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 20/20] 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 74afd1ff7fc9..70d96c2e0579 100644
--- a/fs/aio.c
+++ b/fs/aio.c
@@ -1329,7 +1329,8 @@ 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;
 
 	list_for_each_entry_safe(iocb, n, &ctx->poll_completing, ki_list) {
 		if (*nr_events == max)
@@ -1342,7 +1343,20 @@ 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))) {
@@ -1352,6 +1366,9 @@ 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);
 
@@ -1746,15 +1763,69 @@ 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;
 };
 
-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;
@@ -1905,7 +1976,8 @@ 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;
@@ -1913,7 +1985,7 @@ 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;
@@ -1939,7 +2011,8 @@ 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;
@@ -1947,7 +2020,7 @@ 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;
@@ -2258,16 +2331,16 @@ 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)
@@ -2359,17 +2432,20 @@ 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);
@@ -2416,7 +2492,7 @@ 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++) {
@@ -2460,7 +2536,7 @@ 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

Re: [PATCH 01/20] aio: fix failure to put the file pointer

[Johannes Thumshirn]

Reivewed-by: Johannes Thumshirn <jthumshirn@suse.de>
-- 
Johannes Thumshirn                                        SUSE Labs
jthumshirn@suse.de                                +49 911 74053 689
SUSE LINUX GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: Felix Imendörffer, Jane Smithard, Graham Norton
HRB 21284 (AG Nürnberg)
Key fingerprint = EC38 9CAB C2C4 F25D 8600 D0D0 0393 969D 2D76 0850

Re: [PATCH 02/20] aio: clear IOCB_HIPRI

[Johannes Thumshirn]

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
-- 
Johannes Thumshirn                                        SUSE Labs
jthumshirn@suse.de                                +49 911 74053 689
SUSE LINUX GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: Felix Imendörffer, Jane Smithard, Graham Norton
HRB 21284 (AG Nürnberg)
Key fingerprint = EC38 9CAB C2C4 F25D 8600 D0D0 0393 969D 2D76 0850

Re: [PATCH 03/20] fs: add an iopoll method to struct file_operations

[Johannes Thumshirn]

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
-- 
Johannes Thumshirn                                        SUSE Labs
jthumshirn@suse.de                                +49 911 74053 689
SUSE LINUX GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: Felix Imendörffer, Jane Smithard, Graham Norton
HRB 21284 (AG Nürnberg)
Key fingerprint = EC38 9CAB C2C4 F25D 8600 D0D0 0393 969D 2D76 0850

Re: [PATCH 04/20] block: wire up block device iopoll method

[Johannes Thumshirn]

Reviewed-by: Johannes Thumshirn <jthumshirn@suse.de>
-- 
Johannes Thumshirn                                        SUSE Labs
jthumshirn@suse.de                                +49 911 74053 689
SUSE LINUX GmbH, Maxfeldstr. 5, 90409 Nürnberg
GF: Felix Imendörffer, Jane Smithard, Graham Norton
HRB 21284 (AG Nürnberg)
Key fingerprint = EC38 9CAB C2C4 F25D 8600 D0D0 0393 969D 2D76 0850

Re: [PATCH 17/20] aio: support for IO polling

[Benny Halevy]

On Mon, 2018-11-26 at 09:45 -0700, Jens Axboe wrote:
> 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                     | 378 +++++++++++++++++++++++++++++++----
>  include/uapi/linux/aio_abi.h |   3 +
>  2 files changed, 345 insertions(+), 36 deletions(-)
> 
> diff --git a/fs/aio.c b/fs/aio.c
> index e98121df92f6..db73c8af1a0a 100644
> --- a/fs/aio.c
> +++ b/fs/aio.c
> @@ -143,6 +143,18 @@ 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 */
> @@ -195,14 +207,27 @@ 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 IOCB_POLL_COMPLETED	0
> +#define IOCB_POLL_BUSY		1
> +
>  	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----*/
> @@ -223,6 +248,7 @@ static const unsigned int iocb_page_shift =
>  				ilog2(PAGE_SIZE / sizeof(struct iocb));
>  
>  static void aio_useriocb_free(struct kioctx *);
> +static void aio_iopoll_reap_events(struct kioctx *);
>  
>  static struct file *aio_private_file(struct kioctx *ctx, loff_t nr_pages)
>  {
> @@ -461,11 +487,15 @@ 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)
> @@ -747,6 +777,11 @@ 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;
>  
> @@ -818,11 +853,15 @@ 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);

Is it worth handling the mutex lock and calling aio_iopoll_reap_events
only if (ctx->flags & IOCTX_FLAG_IOPOLL)?  If so, testing it can be
removed from aio_iopoll_reap_events() (and maybe it could even
be open coded
here since this is its only call site apparently)

>  
>  	table = rcu_dereference_raw(mm->ioctx_table);
>  	WARN_ON(ctx != rcu_access_pointer(table->table[ctx->id]));
> @@ -1029,6 +1068,7 @@ 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;
>  	}
> @@ -1072,6 +1112,15 @@ static inline void iocb_put(struct aio_kiocb *iocb)
>  	}
>  }
>  
> +static void iocb_put_many(struct kioctx *ctx, void **iocbs, int *nr)
> +{
> +	if (nr) {

How can nr by NULL?
And what's the point of supporting this case?
Did you mean: if (*nr)?
(In this case, if safe to call the functions below with *nr==0,
I'm not sure it's worth optimizing... especially since this is a static
function and its callers make sure to call it only when *nr > 0)

> +		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)
>  {
> @@ -1261,6 +1310,166 @@ 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;
> +
> +	list_for_each_entry_safe(iocb, n, &ctx->poll_completing, ki_list) {
> +		if (*nr_events == max)

*nr_events >= max would be safer.

> +			break;
> +		if (!test_bit(IOCB_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;
> +}
> +
> +static int __aio_iopoll_check(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 (*nr_events >= min)
> +			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 (*nr_events >= min)
> +		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(IOCB_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_check(ctx, NULL, &nr_events, 1, UINT_MAX);
> +	}
> +}
> +
> +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 = 0;
> +
> +	/* 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;
> +	}
> +
> +	while (!nr_events || !need_resched()) {
> +		int tmin = 0;
> +
> +		if (nr_events < min_nr)
> +			tmin = min_nr - nr_events;
> +
> +		ret = __aio_iopoll_check(ctx, event, &nr_events, tmin, nr);
> +		if (ret <= 0)
> +			break;
> +		ret = 0;
> +	}
> +
> +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)
> @@ -1336,7 +1545,7 @@ SYSCALL_DEFINE4(io_setup2, u32, nr_events, u32, flags, struct iocb * __user,
>  	unsigned long ctx;
>  	long ret;
>  
> -	if (flags & ~IOCTX_FLAG_USERIOCB)
> +	if (flags & ~(IOCTX_FLAG_USERIOCB | IOCTX_FLAG_IOPOLL))
>  		return -EINVAL;
>  
>  	ret = get_user(ctx, ctxp);
> @@ -1469,13 +1678,8 @@ 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);
>  
> @@ -1487,19 +1691,48 @@ 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(IOCB_POLL_BUSY, &iocb->ki_flags);
> +	else {
> +		aio_fill_event(&iocb->ki_ev, iocb, res, res2);
> +		set_bit(IOCB_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)
> @@ -1525,9 +1758,35 @@ 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;
> @@ -1570,17 +1829,43 @@ static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
>  	default:
>  		req->ki_complete(req, ret, 0);
>  	}
> +

nit: this hunk is probably unintentional

>  }
>  
> -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_add = test_bit(IOCB_POLL_COMPLETED, &kiocb->ki_flags);
> +	struct kioctx *ctx = kiocb->ki_ctx;
> +
> +	spin_lock(&ctx->poll_lock);
> +	if (front_add)
> +		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;
> @@ -1605,15 +1890,16 @@ 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;
> @@ -1884,7 +2170,8 @@ 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;
> @@ -1907,8 +2194,8 @@ 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");
> @@ -1919,26 +2206,33 @@ 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:
> @@ -1954,13 +2248,21 @@ 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(IOCB_POLL_BUSY, &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;
>  }
>  
> @@ -2136,7 +2438,7 @@ 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);
> @@ -2171,8 +2473,12 @@ 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

Re: [PATCH 17/20] aio: support for IO polling

[Jens Axboe]

On 11/27/18 2:53 AM, Benny Halevy wrote:
>> @@ -818,11 +853,15 @@ 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);
> 
> Is it worth handling the mutex lock and calling aio_iopoll_reap_events
> only if (ctx->flags & IOCTX_FLAG_IOPOLL)?  If so, testing it can be
> removed from aio_iopoll_reap_events() (and maybe it could even
> be open coded
> here since this is its only call site apparently)

I don't think it really matters, this only happens when you tear down an
io_context. FWIW, I think it's cleaner to retain the test in the
function, not outside it.

>> @@ -1072,6 +1112,15 @@ static inline void iocb_put(struct aio_kiocb *iocb)
>>  	}
>>  }
>>  
>> +static void iocb_put_many(struct kioctx *ctx, void **iocbs, int *nr)
>> +{
>> +	if (nr) {
> 
> How can nr by NULL?
> And what's the point of supporting this case?
> Did you mean: if (*nr)?
> (In this case, if safe to call the functions below with *nr==0,
> I'm not sure it's worth optimizing... especially since this is a static
> function and its callers make sure to call it only when *nr > 0)

Indeed, that should be if (*nr), thanks! The slub implementation of the
bulk free complains if you pass in nr == 0. Outside of that, a single
check should be better than checking in multiple spots.

>> @@ -1261,6 +1310,166 @@ 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;
>> +
>> +	list_for_each_entry_safe(iocb, n, &ctx->poll_completing, ki_list) {
>> +		if (*nr_events == max)
> 
> *nr_events >= max would be safer.

I don't see how we can get there with it being larger than already, that
would be a big bug if we fill more events than userspace asked for.

>> @@ -1570,17 +1829,43 @@ static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
>>  	default:
>>  		req->ki_complete(req, ret, 0);
>>  	}
>> +
> 
> nit: this hunk is probably unintentional

Looks like it, I'll kill it.


-- 
Jens Axboe

Re: [PATCH 17/20] aio: support for IO polling

[Benny Halevy]

On Tue, 2018-11-27 at 08:24 -0700, Jens Axboe wrote:
> On 11/27/18 2:53 AM, Benny Halevy wrote:
> > > @@ -818,11 +853,15 @@ 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);
> > 
> > Is it worth handling the mutex lock and calling aio_iopoll_reap_events
> > only if (ctx->flags & IOCTX_FLAG_IOPOLL)?  If so, testing it can be
> > removed from aio_iopoll_reap_events() (and maybe it could even
> > be open coded
> > here since this is its only call site apparently)
> 
> I don't think it really matters, this only happens when you tear down an
> io_context. FWIW, I think it's cleaner to retain the test in the
> function, not outside it.
> 
> > > @@ -1072,6 +1112,15 @@ static inline void iocb_put(struct aio_kiocb *iocb)
> > >  	}
> > >  }
> > >  
> > > +static void iocb_put_many(struct kioctx *ctx, void **iocbs, int *nr)
> > > +{
> > > +	if (nr) {
> > 
> > How can nr by NULL?
> > And what's the point of supporting this case?
> > Did you mean: if (*nr)?
> > (In this case, if safe to call the functions below with *nr==0,
> > I'm not sure it's worth optimizing... especially since this is a static
> > function and its callers make sure to call it only when *nr > 0)
> 
> Indeed, that should be if (*nr), thanks! The slub implementation of the
> bulk free complains if you pass in nr == 0. Outside of that, a single
> check should be better than checking in multiple spots.
> 

Cool. The compiler might also optimize it away when inlining this function
if the caller tests *nr for being non-zero too.

> > > @@ -1261,6 +1310,166 @@ 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;
> > > +
> > > +	list_for_each_entry_safe(iocb, n, &ctx->poll_completing, ki_list) {
> > > +		if (*nr_events == max)
> > 
> > *nr_events >= max would be safer.
> 
> I don't see how we can get there with it being larger than already, that
> would be a big bug if we fill more events than userspace asked for.
> 

Currently we indeed can't, but if the code changes in the future
and we do, this will reduce the damage - hence being safer (and it costs nothing
in terms of performance).

> > > @@ -1570,17 +1829,43 @@ static inline void aio_rw_done(struct kiocb *req, ssize_t ret)
> > >  	default:
> > >  		req->ki_complete(req, ret, 0);
> > >  	}
> > > +
> > 
> > nit: this hunk is probably unintentional
> 
> Looks like it, I'll kill it.
> 
> 

Re: [PATCH 17/20] aio: support for IO polling

[Jens Axboe]

On 11/28/18 2:33 AM, Benny Halevy wrote:
>> I don't see how we can get there with it being larger than already,
>> that would be a big bug if we fill more events than userspace asked
>> for.
>>
> 
> Currently we indeed can't, but if the code changes in the future and
> we do, this will reduce the damage - hence being safer (and it costs
> nothing in terms of performance).

The thing is, if we're ever over max, we have potentially corrupted user
space memory by copying back too many events. So if anything, it should
be a BUG() condition, not just a check.

-- 
Jens Axboe

Re: [PATCH 17/20] aio: support for IO polling

[Benny Halevy]

On Wed, 2018-11-28 at 11:50 -0700, Jens Axboe wrote:
> On 11/28/18 2:33 AM, Benny Halevy wrote:
> > > I don't see how we can get there with it being larger than already,
> > > that would be a big bug if we fill more events than userspace asked
> > > for.
> > > 
> > 
> > Currently we indeed can't, but if the code changes in the future and
> > we do, this will reduce the damage - hence being safer (and it costs
> > nothing in terms of performance).
> 
> The thing is, if we're ever over max, we have potentially corrupted user
> space memory by copying back too many events. So if anything, it should
> be a BUG() condition, not just a check.
> 

Agreed.