[PATCHSET v2] io_uring IO interface

[PATCHSET v2] io_uring IO interface

[Jens Axboe]

Here's v2 of the io_uring interface. See the v1 posting for some more info:

https://lore.kernel.org/linux-block/20190108165645.19311-1-axboe@kernel.dk/

The data structures changed, to improve the symmetry of the submission
and completion side. The io_uring_iocb is now io_uring_sqe, but it
otherwise remains the same as before. Ditto on the completion side,
where io_uring_event is now io_uring_cqe.

I've updated the fio io_uring test app, and the io_uring engine. The
liburing git repo has also been adapted to the various changes since the
v1 posting. As a reminder, the liburing git repo contains some helpers
for doing IO without having to muck with the ring directly, setting up
an io_uring context, etc. Clone that here:

git://git.kernel.dk/liburing

In terms of usage, there's also a small test app here:

http://git.kernel.dk/cgit/fio/plain/t/io_uring.c

and the liburing repo has a few test apps in test/ as well.

Patches are aginst 5.0-rc1, but can also be found here:

git://git.kernel.dk/linux-block io_uring

Changes since v1:

- Fail IORING_OP_{READ,WRITE}_FIXED if not configured
- Fix ctx drop ref issue on failure to close ring_fd when sq thread/wq
  are in use
- Move to separate Kconfig entry (CONFIG_IO_URING)
- Add SPDX headers
- Drop gcc ism of zero sized arrays
- Rename io_uring_iocb -> io_uring_sqe
- Rename io_uring_event -> io_uring_cqe
- Drop needless io_event_ring and io_iocb_ring structures
- Drop ctx->max_reqs, use ->sq_entries
- Drop unused ->ring_lock
- Drop io_ring_ctx slab cache
- Fix state batched kiocb alloc failure to put ctx
- Fix missing write ordering barrier when filling in the cqe
- Drop io_req_init()
- Various renames
- Fix a few lines that were too long
- Address other minor review comments
- Fix IORING_SETUP_SQPOLL being set without IORING_SETUP_SQTHREAD
- Drop IORING_SETUP_FIXEDBUFS, iovecs being non-NULL is enough
- Fix error handling free of ctx in setup path
- Change standard read/write commands to be iov based READV/WRITEV
- Pass in struct sqe_submit instead of separate sqe/index everywhere
- Fix reap of polled events on fops->release()
- Lock uring for sq thread polling
- Don't grab ->completion_lock for polled IO cqe filling
- Fix ev_flags vs flags typo
- Consolidate parts of the io_ring_ctx alignment

 Documentation/filesystems/vfs.txt      |    3 +
 arch/x86/entry/syscalls/syscall_64.tbl |    2 +
 block/bio.c                            |   59 +-
 fs/Makefile                            |    1 +
 fs/block_dev.c                         |   19 +-
 fs/file.c                              |   15 +-
 fs/file_table.c                        |    9 +-
 fs/gfs2/file.c                         |    2 +
 fs/io_uring.c                          | 1890 ++++++++++++++++++++++++
 fs/iomap.c                             |   48 +-
 fs/xfs/xfs_file.c                      |    1 +
 include/linux/bio.h                    |   14 +
 include/linux/blk_types.h              |    1 +
 include/linux/file.h                   |    2 +
 include/linux/fs.h                     |    6 +-
 include/linux/iomap.h                  |    1 +
 include/linux/syscalls.h               |    5 +
 include/uapi/linux/io_uring.h          |  114 ++
 init/Kconfig                           |    8 +
 kernel/sys_ni.c                        |    2 +
 20 files changed, 2163 insertions(+), 39 deletions(-)

-- 
Jens Axboe

[PATCH 01/15] fs: add an iopoll method to struct file_operations

[Jens Axboe]

From: Christoph Hellwig <hch@lst.de>

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

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

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

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

diff --git a/Documentation/filesystems/vfs.txt b/Documentation/filesystems/vfs.txt
index 8dc8e9c2913f..761c6fd24a53 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 811c77743dad..ccb0b7a63aa5 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)
@@ -1786,6 +1787,7 @@ struct file_operations {
 	ssize_t (*write) (struct file *, const char __user *, size_t, loff_t *);
 	ssize_t (*read_iter) (struct kiocb *, struct iov_iter *);
 	ssize_t (*write_iter) (struct kiocb *, struct iov_iter *);
+	int (*iopoll)(struct kiocb *kiocb, bool spin);
 	int (*iterate) (struct file *, struct dir_context *);
 	int (*iterate_shared) (struct file *, struct dir_context *);
 	__poll_t (*poll) (struct file *, struct poll_table_struct *);
-- 
2.17.1

[PATCH 02/15] block: wire up block device iopoll method

[Jens Axboe]

From: Christoph Hellwig <hch@lst.de>

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

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

diff --git a/fs/block_dev.c b/fs/block_dev.c
index c546cdce77e6..5415579f3e14 100644
--- a/fs/block_dev.c
+++ b/fs/block_dev.c
@@ -279,6 +279,14 @@ 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;
 		}
 
@@ -2068,6 +2077,7 @@ 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 03/15] block: add bio_set_polled() helper

[Jens Axboe]

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

Utilize the helper in the blockdev DIRECT_IO code.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/block_dev.c      |  4 ++--
 include/linux/bio.h | 14 ++++++++++++++
 2 files changed, 16 insertions(+), 2 deletions(-)

diff --git a/fs/block_dev.c b/fs/block_dev.c
index 5415579f3e14..2ebd2a0d7789 100644
--- a/fs/block_dev.c
+++ b/fs/block_dev.c
@@ -233,7 +233,7 @@ __blkdev_direct_IO_simple(struct kiocb *iocb, struct iov_iter *iter,
 		task_io_account_write(ret);
 	}
 	if (iocb->ki_flags & IOCB_HIPRI)
-		bio.bi_opf |= REQ_HIPRI;
+		bio_set_polled(&bio, iocb);
 
 	qc = submit_bio(&bio);
 	for (;;) {
@@ -401,7 +401,7 @@ __blkdev_direct_IO(struct kiocb *iocb, struct iov_iter *iter, int nr_pages)
 		nr_pages = iov_iter_npages(iter, BIO_MAX_PAGES);
 		if (!nr_pages) {
 			if (iocb->ki_flags & IOCB_HIPRI)
-				bio->bi_opf |= REQ_HIPRI;
+				bio_set_polled(bio, iocb);
 
 			qc = submit_bio(bio);
 			WRITE_ONCE(iocb->ki_cookie, qc);
diff --git a/include/linux/bio.h b/include/linux/bio.h
index 7380b094dcca..f6f0a2b3cbc8 100644
--- a/include/linux/bio.h
+++ b/include/linux/bio.h
@@ -823,5 +823,19 @@ static inline int bio_integrity_add_page(struct bio *bio, struct page *page,
 
 #endif /* CONFIG_BLK_DEV_INTEGRITY */
 
+/*
+ * Mark a bio as polled. Note that for async polled IO, the caller must
+ * expect -EWOULDBLOCK if we cannot allocate a request (or other resources).
+ * We cannot block waiting for requests on polled IO, as those completions
+ * must be found by the caller. This is different than IRQ driven IO, where
+ * it's safe to wait for IO to complete.
+ */
+static inline void bio_set_polled(struct bio *bio, struct kiocb *kiocb)
+{
+	bio->bi_opf |= REQ_HIPRI;
+	if (!is_sync_kiocb(kiocb))
+		bio->bi_opf |= REQ_NOWAIT;
+}
+
 #endif /* CONFIG_BLOCK */
 #endif /* __LINUX_BIO_H */
-- 
2.17.1

[PATCH 04/15] iomap: wire up the iopoll method

[Jens Axboe]

From: Christoph Hellwig <hch@lst.de>

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

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

Modified to use bio_set_polled().

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 a2dea5bc0427..58a768e59712 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 a3088fae567b..4ee50b76b4a1 100644
--- a/fs/iomap.c
+++ b/fs/iomap.c
@@ -1454,6 +1454,28 @@ 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_set_polled(bio, dio->iocb);
+
+	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;
@@ -1566,7 +1588,7 @@ 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)
 {
@@ -1580,15 +1602,10 @@ 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
@@ -1691,9 +1708,6 @@ 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;
@@ -1701,11 +1715,7 @@ 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);
 
 	/*
@@ -1916,6 +1926,9 @@ iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 	if (dio->flags & IOMAP_DIO_WRITE_FUA)
 		dio->flags &= ~IOMAP_DIO_NEED_SYNC;
 
+	WRITE_ONCE(iocb->ki_cookie, dio->submit.cookie);
+	WRITE_ONCE(iocb->private, dio->submit.last_queue);
+
 	if (!atomic_dec_and_test(&dio->ref)) {
 		if (!dio->wait_for_completion)
 			return -EIOCBQUEUED;
diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index e47425071e65..60c2da41f0fc 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -1203,6 +1203,7 @@ const struct file_operations xfs_file_operations = {
 	.write_iter	= xfs_file_write_iter,
 	.splice_read	= generic_file_splice_read,
 	.splice_write	= iter_file_splice_write,
+	.iopoll		= iomap_dio_iopoll,
 	.unlocked_ioctl	= xfs_file_ioctl,
 #ifdef CONFIG_COMPAT
 	.compat_ioctl	= xfs_file_compat_ioctl,
diff --git a/include/linux/iomap.h b/include/linux/iomap.h
index 9a4258154b25..0fefb5455bda 100644
--- a/include/linux/iomap.h
+++ b/include/linux/iomap.h
@@ -162,6 +162,7 @@ typedef int (iomap_dio_end_io_t)(struct kiocb *iocb, ssize_t ret,
 		unsigned flags);
 ssize_t iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 		const struct iomap_ops *ops, iomap_dio_end_io_t end_io);
+int iomap_dio_iopoll(struct kiocb *kiocb, bool spin);
 
 #ifdef CONFIG_SWAP
 struct file;
-- 
2.17.1

[PATCH 05/15] Add io_uring IO interface

[Jens Axboe]

The submission queue (SQ) and completion queue (CQ) rings are shared
between the application and the kernel. This eliminates the need to
copy data back and forth to submit and complete IO.

IO submissions use the io_uring_sqe data structure, and completions
are generated in the form of io_uring_sqe data structures. The SQ
ring is an index into the io_uring_sqe array, which makes it possible
to submit a batch of IOs without them being contiguous in the ring.
The CQ ring is always contiguous, as completion events are inherently
unordered and can point to any io_uring_iocb.

Two new system calls are added for this:

io_uring_setup(entries, iovecs, params)
	Sets up a context for doing async IO. On success, returns a file
	descriptor that the application can mmap to gain access to the
	SQ ring, CQ ring, and io_uring_iocbs.

io_uring_enter(fd, to_submit, min_complete, flags)
	Initiates IO against the rings mapped to this fd, or waits for
	them to complete, or both The behavior is controlled by the
	parameters passed in. If 'min_complete' is non-zero, then we'll
	try and submit new IO. If IORING_ENTER_GETEVENTS is set, the
	kernel will wait for 'min_complete' events, if they aren't
	already available.

With this setup, it's possible to do async IO with a single system
call. Future developments will enable polled IO with this interface,
and polled submission as well. The latter will enable an application
to do IO without doing ANY system calls at all.

For IRQ driven IO, an application only needs to enter the kernel for
completions if it wants to wait for them to occur.

Sample application: http://git.kernel.dk/cgit/fio/plain/t/io_uring.c

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 arch/x86/entry/syscalls/syscall_64.tbl |   2 +
 fs/Makefile                            |   1 +
 fs/io_uring.c                          | 838 +++++++++++++++++++++++++
 include/linux/syscalls.h               |   5 +
 include/uapi/linux/io_uring.h          |  96 +++
 init/Kconfig                           |   8 +
 kernel/sys_ni.c                        |   2 +
 7 files changed, 952 insertions(+)
 create mode 100644 fs/io_uring.c
 create mode 100644 include/uapi/linux/io_uring.h

diff --git a/arch/x86/entry/syscalls/syscall_64.tbl b/arch/x86/entry/syscalls/syscall_64.tbl
index f0b1709a5ffb..453ff7a79002 100644
--- a/arch/x86/entry/syscalls/syscall_64.tbl
+++ b/arch/x86/entry/syscalls/syscall_64.tbl
@@ -343,6 +343,8 @@
 332	common	statx			__x64_sys_statx
 333	common	io_pgetevents		__x64_sys_io_pgetevents
 334	common	rseq			__x64_sys_rseq
+335	common	io_uring_setup		__x64_sys_io_uring_setup
+336	common	io_uring_enter		__x64_sys_io_uring_enter
 
 #
 # x32-specific system call numbers start at 512 to avoid cache impact
diff --git a/fs/Makefile b/fs/Makefile
index 293733f61594..8e15d6fc4340 100644
--- a/fs/Makefile
+++ b/fs/Makefile
@@ -30,6 +30,7 @@ obj-$(CONFIG_TIMERFD)		+= timerfd.o
 obj-$(CONFIG_EVENTFD)		+= eventfd.o
 obj-$(CONFIG_USERFAULTFD)	+= userfaultfd.o
 obj-$(CONFIG_AIO)               += aio.o
+obj-$(CONFIG_IO_URING)		+= io_uring.o
 obj-$(CONFIG_FS_DAX)		+= dax.o
 obj-$(CONFIG_FS_ENCRYPTION)	+= crypto/
 obj-$(CONFIG_FILE_LOCKING)      += locks.o
diff --git a/fs/io_uring.c b/fs/io_uring.c
new file mode 100644
index 000000000000..0bad563f3486
--- /dev/null
+++ b/fs/io_uring.c
@@ -0,0 +1,838 @@
+// SPDX-License-Identifier: GPL-2.0
+/*
+ * Shared application/kernel submission and completion ring pairs, for
+ * supporting fast/efficient IO.
+ *
+ * Copyright (C) 2019 Jens Axboe
+ */
+#include <linux/kernel.h>
+#include <linux/init.h>
+#include <linux/errno.h>
+#include <linux/syscalls.h>
+#include <linux/refcount.h>
+#include <linux/uio.h>
+
+#include <linux/sched/signal.h>
+#include <linux/fs.h>
+#include <linux/file.h>
+#include <linux/mm.h>
+#include <linux/mman.h>
+#include <linux/mmu_context.h>
+#include <linux/percpu.h>
+#include <linux/slab.h>
+#include <linux/workqueue.h>
+#include <linux/blkdev.h>
+#include <linux/anon_inodes.h>
+
+#include <linux/uaccess.h>
+#include <linux/nospec.h>
+
+#include <uapi/linux/io_uring.h>
+
+#include "internal.h"
+
+struct io_uring {
+	u32 head ____cacheline_aligned_in_smp;
+	u32 tail ____cacheline_aligned_in_smp;
+};
+
+struct io_sq_ring {
+	struct io_uring		r;
+	u32			ring_mask;
+	u32			ring_entries;
+	u32			dropped;
+	u32			flags;
+	u32			array[];
+};
+
+struct io_cq_ring {
+	struct io_uring		r;
+	u32			ring_mask;
+	u32			ring_entries;
+	u32			overflow;
+	struct io_uring_cqe	cqes[];
+};
+
+struct io_ring_ctx {
+	struct percpu_ref	refs;
+
+	unsigned int		flags;
+
+	/* SQ ring */
+	struct io_sq_ring	*sq_ring;
+	unsigned		sq_entries;
+	unsigned		sq_mask;
+	struct io_uring_sqe	*sq_sqes;
+
+	/* CQ ring */
+	struct io_cq_ring	*cq_ring;
+	unsigned		cq_entries;
+	unsigned		cq_mask;
+
+	struct completion	ctx_done;
+
+	struct {
+		struct mutex		uring_lock;
+		wait_queue_head_t	wait;
+	} ____cacheline_aligned_in_smp;
+
+	struct {
+		spinlock_t		completion_lock;
+	} ____cacheline_aligned_in_smp;
+};
+
+struct fsync_iocb {
+	struct work_struct	work;
+	struct file		*file;
+	bool			datasync;
+};
+
+struct io_kiocb {
+	union {
+		struct kiocb		rw;
+		struct fsync_iocb	fsync;
+	};
+
+	struct io_ring_ctx	*ki_ctx;
+	unsigned long		ki_index;
+	struct list_head	ki_list;
+	unsigned long		ki_flags;
+};
+
+#define IO_PLUG_THRESHOLD		2
+
+struct sqe_submit {
+	const struct io_uring_sqe *sqe;
+	unsigned index;
+};
+
+static struct kmem_cache *kiocb_cachep;
+
+static const struct file_operations io_scqring_fops;
+
+static void io_ring_ctx_ref_free(struct percpu_ref *ref)
+{
+	struct io_ring_ctx *ctx = container_of(ref, struct io_ring_ctx, refs);
+
+	complete(&ctx->ctx_done);
+}
+
+static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
+{
+	struct io_ring_ctx *ctx;
+
+	ctx = kzalloc(sizeof(*ctx), GFP_KERNEL);
+	if (!ctx)
+		return NULL;
+
+	if (percpu_ref_init(&ctx->refs, io_ring_ctx_ref_free, 0, GFP_KERNEL)) {
+		kfree(ctx);
+		return NULL;
+	}
+
+	ctx->flags = p->flags;
+
+	init_completion(&ctx->ctx_done);
+
+	spin_lock_init(&ctx->completion_lock);
+	init_waitqueue_head(&ctx->wait);
+	mutex_init(&ctx->uring_lock);
+
+	return ctx;
+}
+
+static void io_inc_cqring(struct io_ring_ctx *ctx)
+{
+	struct io_cq_ring *ring = ctx->cq_ring;
+
+	ring->r.tail++;
+	smp_wmb();
+}
+
+static struct io_uring_cqe *io_peek_cqring(struct io_ring_ctx *ctx)
+{
+	struct io_cq_ring *ring = ctx->cq_ring;
+	unsigned tail;
+
+	smp_rmb();
+	tail = READ_ONCE(ring->r.tail);
+	if (tail + 1 == READ_ONCE(ring->r.head))
+		return NULL;
+
+	return &ring->cqes[tail & ctx->cq_mask];
+}
+
+static struct io_kiocb *io_get_kiocb(struct io_ring_ctx *ctx)
+{
+	struct io_kiocb *req;
+
+	if (!percpu_ref_tryget(&ctx->refs))
+		return NULL;
+
+	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
+	if (!req)
+		return NULL;
+
+	req->ki_ctx = ctx;
+	INIT_LIST_HEAD(&req->ki_list);
+	req->ki_flags = 0;
+	return req;
+}
+
+static void io_ring_drop_ctx_ref(struct io_ring_ctx *ctx, unsigned refs)
+{
+	percpu_ref_put_many(&ctx->refs, refs);
+
+	if (waitqueue_active(&ctx->wait))
+		wake_up(&ctx->wait);
+}
+
+static void io_free_kiocb(struct io_kiocb *iocb)
+{
+	kmem_cache_free(kiocb_cachep, iocb);
+	io_ring_drop_ctx_ref(iocb->ki_ctx, 1);
+}
+
+static void kiocb_end_write(struct kiocb *kiocb)
+{
+	if (kiocb->ki_flags & IOCB_WRITE) {
+		struct inode *inode = file_inode(kiocb->ki_filp);
+
+		/*
+		 * Tell lockdep we inherited freeze protection from submission
+		 * thread.
+		 */
+		if (S_ISREG(inode->i_mode))
+			__sb_writers_acquired(inode->i_sb, SB_FREEZE_WRITE);
+		file_end_write(kiocb->ki_filp);
+	}
+}
+
+static void io_cqring_fill_event(struct io_ring_ctx *ctx, unsigned ki_index,
+				 long res, unsigned ev_flags)
+{
+	struct io_uring_cqe *cqe;
+	unsigned long flags;
+
+	/*
+	 * If we can't get a cq entry, userspace overflowed the
+	 * submission (by quite a lot). Increment the overflow count in
+	 * the ring.
+	 */
+	spin_lock_irqsave(&ctx->completion_lock, flags);
+	cqe = io_peek_cqring(ctx);
+	if (cqe) {
+		cqe->index = ki_index;
+		cqe->res = res;
+		cqe->flags = ev_flags;
+		smp_wmb();
+		io_inc_cqring(ctx);
+	} else
+		ctx->cq_ring->overflow++;
+	spin_unlock_irqrestore(&ctx->completion_lock, flags);
+}
+
+static void io_complete_scqring_rw(struct kiocb *kiocb, long res, long res2)
+{
+	struct io_kiocb *iocb = container_of(kiocb, struct io_kiocb, rw);
+
+	kiocb_end_write(kiocb);
+
+	fput(kiocb->ki_filp);
+	io_cqring_fill_event(iocb->ki_ctx, iocb->ki_index, res, 0);
+	io_free_kiocb(iocb);
+}
+
+static int io_prep_rw(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe)
+{
+	struct kiocb *req = &kiocb->rw;
+	int ret;
+
+	req->ki_filp = fget(sqe->fd);
+	if (unlikely(!req->ki_filp))
+		return -EBADF;
+	req->ki_pos = sqe->off;
+	req->ki_flags = iocb_flags(req->ki_filp);
+	req->ki_hint = ki_hint_validate(file_write_hint(req->ki_filp));
+	if (sqe->ioprio) {
+		ret = ioprio_check_cap(sqe->ioprio);
+		if (ret)
+			goto out_fput;
+
+		req->ki_ioprio = sqe->ioprio;
+	} else
+		req->ki_ioprio = get_current_ioprio();
+
+	ret = kiocb_set_rw_flags(req, sqe->rw_flags);
+	if (unlikely(ret))
+		goto out_fput;
+	if (req->ki_flags & IOCB_HIPRI) {
+		ret = -EINVAL;
+		goto out_fput;
+	}
+
+	req->ki_complete = io_complete_scqring_rw;
+	return 0;
+out_fput:
+	fput(req->ki_filp);
+	return ret;
+}
+
+static inline void io_rw_done(struct kiocb *req, ssize_t ret)
+{
+	switch (ret) {
+	case -EIOCBQUEUED:
+		break;
+	case -ERESTARTSYS:
+	case -ERESTARTNOINTR:
+	case -ERESTARTNOHAND:
+	case -ERESTART_RESTARTBLOCK:
+		/*
+		 * There's no easy way to restart the syscall since other AIO's
+		 * may be already running. Just fail this IO with EINTR.
+		 */
+		ret = -EINTR;
+		/*FALLTHRU*/
+	default:
+		req->ki_complete(req, ret, 0);
+	}
+}
+
+static ssize_t io_read(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe)
+{
+	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
+	void __user *buf = (void __user *) (uintptr_t) sqe->addr;
+	struct kiocb *req = &kiocb->rw;
+	struct iov_iter iter;
+	struct file *file;
+	ssize_t ret;
+
+	ret = io_prep_rw(kiocb, sqe);
+	if (ret)
+		return ret;
+	file = req->ki_filp;
+
+	ret = -EBADF;
+	if (unlikely(!(file->f_mode & FMODE_READ)))
+		goto out_fput;
+	ret = -EINVAL;
+	if (unlikely(!file->f_op->read_iter))
+		goto out_fput;
+
+	ret = import_iovec(READ, buf, sqe->len, UIO_FASTIOV, &iovec, &iter);
+	if (ret)
+		goto out_fput;
+
+	ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
+	if (!ret)
+		io_rw_done(req, call_read_iter(file, req, &iter));
+	kfree(iovec);
+out_fput:
+	if (unlikely(ret))
+		fput(file);
+	return ret;
+}
+
+static ssize_t io_write(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe)
+{
+	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
+	void __user *buf = (void __user *) (uintptr_t) sqe->addr;
+	struct kiocb *req = &kiocb->rw;
+	struct iov_iter iter;
+	struct file *file;
+	ssize_t ret;
+
+	ret = io_prep_rw(kiocb, sqe);
+	if (ret)
+		return ret;
+	file = req->ki_filp;
+
+	ret = -EBADF;
+	if (unlikely(!(file->f_mode & FMODE_WRITE)))
+		goto out_fput;
+	ret = -EINVAL;
+	if (unlikely(!file->f_op->write_iter))
+		goto out_fput;
+
+	ret = import_iovec(WRITE, buf, sqe->len, UIO_FASTIOV, &iovec, &iter);
+	if (ret)
+		goto out_fput;
+
+	ret = rw_verify_area(WRITE, file, &req->ki_pos, iov_iter_count(&iter));
+	if (!ret) {
+		/*
+		 * Open-code file_start_write here to grab freeze protection,
+		 * which will be released by another thread in
+		 * io_complete_rw().  Fool lockdep by telling it the lock got
+		 * released so that it doesn't complain about the held lock when
+		 * we return to userspace.
+		 */
+		if (S_ISREG(file_inode(file)->i_mode)) {
+			__sb_start_write(file_inode(file)->i_sb,
+						SB_FREEZE_WRITE, true);
+			__sb_writers_release(file_inode(file)->i_sb,
+						SB_FREEZE_WRITE);
+		}
+		req->ki_flags |= IOCB_WRITE;
+		io_rw_done(req, call_write_iter(file, req, &iter));
+	}
+out_fput:
+	if (unlikely(ret))
+		fput(file);
+	return ret;
+}
+
+static void io_fsync_work(struct work_struct *work)
+{
+	struct fsync_iocb *req = container_of(work, struct fsync_iocb, work);
+	struct io_kiocb *iocb = container_of(req, struct io_kiocb, fsync);
+	int ret;
+
+	ret = vfs_fsync(req->file, req->datasync);
+	fput(req->file);
+
+	io_cqring_fill_event(iocb->ki_ctx, iocb->ki_index, ret, 0);
+	io_free_kiocb(iocb);
+}
+
+static int io_fsync(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe,
+		    bool datasync)
+{
+	struct fsync_iocb *req = &kiocb->fsync;
+
+	if (unlikely(sqe->addr || sqe->off || sqe->len || sqe->__resv))
+		return -EINVAL;
+
+	req->file = fget(sqe->fd);
+	if (unlikely(!req->file))
+		return -EBADF;
+	if (unlikely(!req->file->f_op->fsync)) {
+		fput(req->file);
+		return -EINVAL;
+	}
+
+	req->datasync = datasync;
+	INIT_WORK(&req->work, io_fsync_work);
+	schedule_work(&req->work);
+	return 0;
+}
+
+static int io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s)
+{
+	const struct io_uring_sqe *sqe = s->sqe;
+	struct io_kiocb *req;
+	ssize_t ret;
+
+	/* enforce forwards compatibility on users */
+	if (unlikely(sqe->flags))
+		return -EINVAL;
+
+	req = io_get_kiocb(ctx);
+	if (unlikely(!req))
+		return -EAGAIN;
+
+	ret = -EINVAL;
+	if (s->index >= ctx->sq_entries)
+		goto out_put_req;
+	req->ki_index = s->index;
+
+	ret = -EINVAL;
+	switch (sqe->opcode) {
+	case IORING_OP_READV:
+		ret = io_read(req, sqe);
+		break;
+	case IORING_OP_WRITEV:
+		ret = io_write(req, sqe);
+		break;
+	case IORING_OP_FSYNC:
+		ret = io_fsync(req, sqe, false);
+		break;
+	case IORING_OP_FDSYNC:
+		ret = io_fsync(req, sqe, true);
+		break;
+	default:
+		ret = -EINVAL;
+		break;
+	}
+
+	/*
+	 * If ret is 0, ->ki_complete() has either been called, or will get
+	 * called later on. Anything else, we need to free the req.
+	 */
+	if (ret)
+		goto out_put_req;
+	return 0;
+out_put_req:
+	io_free_kiocb(req);
+	return ret;
+}
+
+static void io_inc_sqring(struct io_ring_ctx *ctx)
+{
+	struct io_sq_ring *ring = ctx->sq_ring;
+
+	ring->r.head++;
+	smp_wmb();
+}
+
+static bool io_peek_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s)
+{
+	struct io_sq_ring *ring = ctx->sq_ring;
+	unsigned head;
+
+	smp_rmb();
+	head = READ_ONCE(ring->r.head);
+	if (head == READ_ONCE(ring->r.tail))
+		return false;
+
+	head = ring->array[head & ctx->sq_mask];
+	if (head < ctx->sq_entries) {
+		s->index = head;
+		s->sqe = &ctx->sq_sqes[head];
+		return true;
+	}
+
+	/* drop invalid entries */
+	ring->r.head++;
+	ring->dropped++;
+	smp_wmb();
+	return false;
+}
+
+static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit)
+{
+	int i, ret = 0, submit = 0;
+	struct blk_plug plug;
+
+	if (to_submit > IO_PLUG_THRESHOLD)
+		blk_start_plug(&plug);
+
+	for (i = 0; i < to_submit; i++) {
+		struct sqe_submit s;
+
+		if (!io_peek_sqring(ctx, &s))
+			break;
+
+		ret = io_submit_sqe(ctx, &s);
+		if (ret)
+			break;
+
+		submit++;
+		io_inc_sqring(ctx);
+	}
+
+	if (to_submit > IO_PLUG_THRESHOLD)
+		blk_finish_plug(&plug);
+
+	return submit ? submit : ret;
+}
+
+/*
+ * Wait until events become available, if we don't already have some. The
+ * application must reap them itself, as they reside on the shared cq ring.
+ */
+static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events)
+{
+	struct io_cq_ring *ring = ctx->cq_ring;
+	DEFINE_WAIT(wait);
+	int ret = 0;
+
+	smp_rmb();
+	if (ring->r.head != ring->r.tail)
+		return 0;
+	if (!min_events)
+		return 0;
+
+	do {
+		prepare_to_wait(&ctx->wait, &wait, TASK_INTERRUPTIBLE);
+
+		ret = 0;
+		smp_rmb();
+		if (ring->r.head != ring->r.tail)
+			break;
+
+		schedule();
+
+		ret = -EINTR;
+		if (signal_pending(current))
+			break;
+	} while (1);
+
+	finish_wait(&ctx->wait, &wait);
+	return ring->r.head == ring->r.tail ? ret : 0;
+}
+
+static int __io_uring_enter(struct io_ring_ctx *ctx, unsigned to_submit,
+			    unsigned min_complete, unsigned flags)
+{
+	int ret = 0;
+
+	if (to_submit) {
+		ret = io_ring_submit(ctx, to_submit);
+		if (ret < 0)
+			return ret;
+	}
+	if (flags & IORING_ENTER_GETEVENTS) {
+		int get_ret;
+
+		if (!ret && to_submit)
+			min_complete = 0;
+
+		get_ret = io_cqring_wait(ctx, min_complete);
+		if (get_ret < 0 && !ret)
+			ret = get_ret;
+	}
+
+	return ret;
+}
+
+static void io_free_scq_urings(struct io_ring_ctx *ctx)
+{
+	if (ctx->sq_ring) {
+		page_frag_free(ctx->sq_ring);
+		ctx->sq_ring = NULL;
+	}
+	if (ctx->sq_sqes) {
+		page_frag_free(ctx->sq_sqes);
+		ctx->sq_sqes = NULL;
+	}
+	if (ctx->cq_ring) {
+		page_frag_free(ctx->cq_ring);
+		ctx->cq_ring = NULL;
+	}
+}
+
+static void io_ring_ctx_free(struct io_ring_ctx *ctx)
+{
+	io_free_scq_urings(ctx);
+	percpu_ref_exit(&ctx->refs);
+	kfree(ctx);
+}
+
+static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
+{
+	percpu_ref_kill(&ctx->refs);
+	wait_for_completion(&ctx->ctx_done);
+	io_ring_ctx_free(ctx);
+}
+
+static int io_scqring_release(struct inode *inode, struct file *file)
+{
+	struct io_ring_ctx *ctx = file->private_data;
+
+	file->private_data = NULL;
+	io_ring_ctx_wait_and_kill(ctx);
+	return 0;
+}
+
+static int io_scqring_mmap(struct file *file, struct vm_area_struct *vma)
+{
+	loff_t offset = (loff_t) vma->vm_pgoff << PAGE_SHIFT;
+	unsigned long sz = vma->vm_end - vma->vm_start;
+	struct io_ring_ctx *ctx = file->private_data;
+	unsigned long pfn;
+	struct page *page;
+	void *ptr;
+
+	switch (offset) {
+	case IORING_OFF_SQ_RING:
+		ptr = ctx->sq_ring;
+		break;
+	case IORING_OFF_SQES:
+		ptr = ctx->sq_sqes;
+		break;
+	case IORING_OFF_CQ_RING:
+		ptr = ctx->cq_ring;
+		break;
+	default:
+		return -EINVAL;
+	}
+
+	page = virt_to_head_page(ptr);
+	if (sz > (PAGE_SIZE << compound_order(page)))
+		return -EINVAL;
+
+	pfn = virt_to_phys(ptr) >> PAGE_SHIFT;
+	return remap_pfn_range(vma, vma->vm_start, pfn, sz, vma->vm_page_prot);
+}
+
+SYSCALL_DEFINE4(io_uring_enter, unsigned int, fd, u32, to_submit,
+		u32, min_complete, u32, flags)
+{
+	struct io_ring_ctx *ctx;
+	long ret = -EBADF;
+	struct fd f;
+
+	f = fdget(fd);
+	if (!f.file)
+		return -EBADF;
+
+	ret = -EOPNOTSUPP;
+	if (f.file->f_op != &io_scqring_fops)
+		goto out_fput;
+
+	ret = -EINVAL;
+	ctx = f.file->private_data;
+	if (!percpu_ref_tryget(&ctx->refs))
+		goto out_fput;
+
+	ret = -EBUSY;
+	if (mutex_trylock(&ctx->uring_lock)) {
+		ret = __io_uring_enter(ctx, to_submit, min_complete, flags);
+		mutex_unlock(&ctx->uring_lock);
+	}
+	io_ring_drop_ctx_ref(ctx, 1);
+out_fput:
+	fdput(f);
+	return ret;
+}
+
+static const struct file_operations io_scqring_fops = {
+	.release	= io_scqring_release,
+	.mmap		= io_scqring_mmap,
+};
+
+static void *io_mem_alloc(size_t size)
+{
+	gfp_t gfp_flags = GFP_KERNEL | __GFP_ZERO | __GFP_NOWARN | __GFP_COMP |
+				__GFP_NORETRY;
+
+	return (void *) __get_free_pages(gfp_flags, get_order(size));
+}
+
+static int io_allocate_scq_urings(struct io_ring_ctx *ctx,
+				  struct io_uring_params *p)
+{
+	struct io_sq_ring *sq_ring;
+	struct io_cq_ring *cq_ring;
+	size_t size;
+	int ret;
+
+	sq_ring = io_mem_alloc(struct_size(sq_ring, array, p->sq_entries));
+	if (!sq_ring)
+		return -ENOMEM;
+
+	ctx->sq_ring = sq_ring;
+	sq_ring->ring_mask = p->sq_entries - 1;
+	sq_ring->ring_entries = p->sq_entries;
+	ctx->sq_mask = sq_ring->ring_mask;
+	ctx->sq_entries = sq_ring->ring_entries;
+
+	ret = -EOVERFLOW;
+	size = array_size(sizeof(struct io_uring_sqe), p->sq_entries);
+	if (size == SIZE_MAX)
+		goto err;
+	ret = -ENOMEM;
+	ctx->sq_sqes = io_mem_alloc(size);
+	if (!ctx->sq_sqes)
+		goto err;
+
+	cq_ring = io_mem_alloc(struct_size(cq_ring, cqes, p->cq_entries));
+	if (!cq_ring)
+		goto err;
+
+	ctx->cq_ring = cq_ring;
+	cq_ring->ring_mask = p->cq_entries - 1;
+	cq_ring->ring_entries = p->cq_entries;
+	ctx->cq_mask = cq_ring->ring_mask;
+	ctx->cq_entries = cq_ring->ring_entries;
+	return 0;
+err:
+	io_free_scq_urings(ctx);
+	return ret;
+}
+
+static void io_fill_offsets(struct io_uring_params *p)
+{
+	memset(&p->sq_off, 0, sizeof(p->sq_off));
+	p->sq_off.head = offsetof(struct io_sq_ring, r.head);
+	p->sq_off.tail = offsetof(struct io_sq_ring, r.tail);
+	p->sq_off.ring_mask = offsetof(struct io_sq_ring, ring_mask);
+	p->sq_off.ring_entries = offsetof(struct io_sq_ring, ring_entries);
+	p->sq_off.flags = offsetof(struct io_sq_ring, flags);
+	p->sq_off.dropped = offsetof(struct io_sq_ring, dropped);
+	p->sq_off.array = offsetof(struct io_sq_ring, array);
+
+	memset(&p->cq_off, 0, sizeof(p->cq_off));
+	p->cq_off.head = offsetof(struct io_cq_ring, r.head);
+	p->cq_off.tail = offsetof(struct io_cq_ring, r.tail);
+	p->cq_off.ring_mask = offsetof(struct io_cq_ring, ring_mask);
+	p->cq_off.ring_entries = offsetof(struct io_cq_ring, ring_entries);
+	p->cq_off.overflow = offsetof(struct io_cq_ring, overflow);
+	p->cq_off.cqes = offsetof(struct io_cq_ring, cqes);
+}
+
+static int io_uring_create(unsigned entries, struct io_uring_params *p)
+{
+	struct io_ring_ctx *ctx;
+	int ret;
+
+	/*
+	 * Use twice as many entries for the CQ ring. It's possible for the
+	 * application to drive a higher depth than the size of the SQ ring,
+	 * since the sqes are only used at submission time. This allows for
+	 * some flexibility in overcommitting a bit.
+	 */
+	p->sq_entries = roundup_pow_of_two(entries);
+	p->cq_entries = 2 * p->sq_entries;
+
+	ctx = io_ring_ctx_alloc(p);
+	if (!ctx)
+		return -ENOMEM;
+
+	ret = io_allocate_scq_urings(ctx, p);
+	if (ret)
+		goto err;
+
+	ret = anon_inode_getfd("[io_uring]", &io_scqring_fops, ctx,
+				O_RDWR | O_CLOEXEC);
+	if (ret < 0)
+		goto err;
+
+	io_fill_offsets(p);
+	return ret;
+err:
+	io_ring_ctx_wait_and_kill(ctx);
+	return ret;
+}
+
+/*
+ * Sets up an aio uring context, and returns the fd. Applications asks for a
+ * ring size, we return the actual sq/cq ring sizes (among other things) in the
+ * params structure passed in.
+ */
+SYSCALL_DEFINE3(io_uring_setup, u32, entries, struct iovec __user *, iovecs,
+		struct io_uring_params __user *, params)
+{
+	struct io_uring_params p;
+	long ret;
+	int i;
+
+	if (copy_from_user(&p, params, sizeof(p)))
+		return -EFAULT;
+	for (i = 0; i < ARRAY_SIZE(p.resv); i++) {
+		if (p.resv[i])
+			return -EINVAL;
+	}
+
+	if (p.flags)
+		return -EINVAL;
+	if (iovecs)
+		return -EINVAL;
+
+	ret = io_uring_create(entries, &p);
+	if (ret < 0)
+		return ret;
+
+	if (copy_to_user(params, &p, sizeof(p)))
+		return -EFAULT;
+
+	return ret;
+}
+
+static int __init io_uring_setup(void)
+{
+	kiocb_cachep = KMEM_CACHE(io_kiocb, SLAB_HWCACHE_ALIGN | SLAB_PANIC);
+	return 0;
+};
+__initcall(io_uring_setup);
diff --git a/include/linux/syscalls.h b/include/linux/syscalls.h
index 257cccba3062..6d40939f65cd 100644
--- a/include/linux/syscalls.h
+++ b/include/linux/syscalls.h
@@ -69,6 +69,7 @@ struct file_handle;
 struct sigaltstack;
 struct rseq;
 union bpf_attr;
+struct io_uring_params;
 
 #include <linux/types.h>
 #include <linux/aio_abi.h>
@@ -309,6 +310,10 @@ asmlinkage long sys_io_pgetevents_time32(aio_context_t ctx_id,
 				struct io_event __user *events,
 				struct old_timespec32 __user *timeout,
 				const struct __aio_sigset *sig);
+asmlinkage long sys_io_uring_setup(u32 entries, struct iovec __user *iov,
+				struct io_uring_params __user *p);
+asmlinkage long sys_io_uring_enter(unsigned int fd, u32 to_submit,
+				u32 min_complete, u32 flags);
 
 /* fs/xattr.c */
 asmlinkage long sys_setxattr(const char __user *path, const char __user *name,
diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
new file mode 100644
index 000000000000..ae30ed41965f
--- /dev/null
+++ b/include/uapi/linux/io_uring.h
@@ -0,0 +1,96 @@
+/* SPDX-License-Identifier: GPL-2.0 WITH Linux-syscall-note */
+/*
+ * Header file for the io_uring interface.
+ *
+ * Copyright (C) 2019 Jens Axboe
+ * Copyright (C) 2019 Christoph Hellwig
+ */
+#ifndef LINUX_IO_URING_H
+#define LINUX_IO_URING_H
+
+#include <linux/fs.h>
+#include <linux/types.h>
+
+/*
+ * IO submission data structure (Submission Queue Entry)
+ */
+struct io_uring_sqe {
+	__u8	opcode;
+	__u8	flags;
+	__u16	ioprio;
+	__s32	fd;
+	__u64	off;
+	union {
+		void	*addr;
+		__u64	__pad;
+	};
+	__u32	len;
+	union {
+		__kernel_rwf_t	rw_flags;
+		__u32		__resv;
+	};
+};
+
+#define IORING_OP_READV		1
+#define IORING_OP_WRITEV	2
+#define IORING_OP_FSYNC		3
+#define IORING_OP_FDSYNC	4
+
+/*
+ * IO completion data structure (Completion Queue Entry)
+ */
+struct io_uring_cqe {
+	__u64	index;		/* what sqe this event came from */
+	__s32	res;		/* result code for this event */
+	__u32	flags;
+};
+
+/*
+ * Magic offsets for the application to mmap the data it needs
+ */
+#define IORING_OFF_SQ_RING		0ULL
+#define IORING_OFF_CQ_RING		0x8000000ULL
+#define IORING_OFF_SQES			0x10000000ULL
+
+/*
+ * Filled with the offset for mmap(2)
+ */
+struct io_sqring_offsets {
+	__u32 head;
+	__u32 tail;
+	__u32 ring_mask;
+	__u32 ring_entries;
+	__u32 flags;
+	__u32 dropped;
+	__u32 array;
+	__u32 resv[3];
+};
+
+struct io_cqring_offsets {
+	__u32 head;
+	__u32 tail;
+	__u32 ring_mask;
+	__u32 ring_entries;
+	__u32 overflow;
+	__u32 cqes;
+	__u32 resv[4];
+};
+
+/*
+ * io_uring_enter(2) flags
+ */
+#define IORING_ENTER_GETEVENTS	(1 << 0)
+
+/*
+ * Passed in for io_uring_setup(2). Copied back with updated info on success
+ */
+struct io_uring_params {
+	__u32 sq_entries;
+	__u32 cq_entries;
+	__u32 flags;
+	__u16 resv[10];
+	struct io_sqring_offsets sq_off;
+	struct io_cqring_offsets cq_off;
+};
+
+#endif
diff --git a/init/Kconfig b/init/Kconfig
index d47cb77a220e..6fbb2f40e912 100644
--- a/init/Kconfig
+++ b/init/Kconfig
@@ -1402,6 +1402,14 @@ config AIO
 	  by some high performance threaded applications. Disabling
 	  this option saves about 7k.
 
+config IO_URING
+	bool "Enable IO uring support" if EXPERT
+	default y
+	help
+	  This option enables support for the io_uring interface, enabling
+	  applications to submit and completion IO through submission and
+	  completion rings that are shared between the kernel and application.
+
 config ADVISE_SYSCALLS
 	bool "Enable madvise/fadvise syscalls" if EXPERT
 	default y
diff --git a/kernel/sys_ni.c b/kernel/sys_ni.c
index ab9d0e3c6d50..ee5e523564bb 100644
--- a/kernel/sys_ni.c
+++ b/kernel/sys_ni.c
@@ -46,6 +46,8 @@ COND_SYSCALL(io_getevents);
 COND_SYSCALL(io_pgetevents);
 COND_SYSCALL_COMPAT(io_getevents);
 COND_SYSCALL_COMPAT(io_pgetevents);
+COND_SYSCALL(io_uring_setup);
+COND_SYSCALL(io_uring_enter);
 
 /* fs/xattr.c */
 
-- 
2.17.1

Re: [PATCH 05/15] Add io_uring IO interface

[Martin K. Petersen]

Jens,

> +struct io_uring_sqe {
> +	__u8	opcode;
> +	__u8	flags;
> +	__u16	ioprio;
> +	__s32	fd;
> +	__u64	off;
> +	union {
> +		void	*addr;
> +		__u64	__pad;
> +	};
> +	__u32	len;
> +	union {
> +		__kernel_rwf_t	rw_flags;
> +		__u32		__resv;
> +	};
> +};

A bit tongue in cheek and yet somewhat serious: While I'm super excited
about the 4 x 64 bitness of the sqe, where does the integrity buffer go?
Or the 128-bit KV store key. How do we extend this interface beyond the
flags?

-- 
Martin K. Petersen	Oracle Linux Engineering

Re: [PATCH 05/15] Add io_uring IO interface

[Jens Axboe]

On 1/11/19 11:19 AM, Martin K. Petersen wrote:
> 
> Jens,
> 
>> +struct io_uring_sqe {
>> +	__u8	opcode;
>> +	__u8	flags;
>> +	__u16	ioprio;
>> +	__s32	fd;
>> +	__u64	off;
>> +	union {
>> +		void	*addr;
>> +		__u64	__pad;
>> +	};
>> +	__u32	len;
>> +	union {
>> +		__kernel_rwf_t	rw_flags;
>> +		__u32		__resv;
>> +	};
>> +};
> 
> A bit tongue in cheek and yet somewhat serious: While I'm super excited
> about the 4 x 64 bitness of the sqe, where does the integrity buffer go?
> Or the 128-bit KV store key. How do we extend this interface beyond the
> flags?

For integrity buffers, how about we stash them on the side? The newer
series has an extra system call, io_uring_register(), which is currently
used for registering files and buffers for IO on the side. You could
trivially tie an integrity buffer to an sqe through that.

For KV, I thint that's an actually interesting use case (sorry,
integrity), and we might just want to bite the bullet and extend the sqe
to full 64 bytes.  We're currently at 48 bytes, which leaves us with 16
bytes of space for KV, and other use cases.

-- 
Jens Axboe

Re: [PATCH 05/15] Add io_uring IO interface

[Jens Axboe]

On 1/11/19 11:34 AM, Jens Axboe wrote:
> On 1/11/19 11:19 AM, Martin K. Petersen wrote:
>>
>> Jens,
>>
>>> +struct io_uring_sqe {
>>> +	__u8	opcode;
>>> +	__u8	flags;
>>> +	__u16	ioprio;
>>> +	__s32	fd;
>>> +	__u64	off;
>>> +	union {
>>> +		void	*addr;
>>> +		__u64	__pad;
>>> +	};
>>> +	__u32	len;
>>> +	union {
>>> +		__kernel_rwf_t	rw_flags;
>>> +		__u32		__resv;
>>> +	};
>>> +};
>>
>> A bit tongue in cheek and yet somewhat serious: While I'm super excited
>> about the 4 x 64 bitness of the sqe, where does the integrity buffer go?
>> Or the 128-bit KV store key. How do we extend this interface beyond the
>> flags?
> 
> For integrity buffers, how about we stash them on the side? The newer
> series has an extra system call, io_uring_register(), which is currently
> used for registering files and buffers for IO on the side. You could
> trivially tie an integrity buffer to an sqe through that.
> 
> For KV, I thint that's an actually interesting use case (sorry,
> integrity), and we might just want to bite the bullet and extend the sqe
> to full 64 bytes.  We're currently at 48 bytes, which leaves us with 16
> bytes of space for KV, and other use cases.

I bit the bullet and bumped the size. 64 bytes is a nicer size in terms
of cachelines anyway, and I really doubt that 48 vs 64 bytes makes a
size consumption problem for anyone.

The buf_index is only used for the fixed buffers, which means that we have
16 bytes / 128 bits that we can grab for things like KV.

-- 
Jens Axboe

Re: [PATCH 05/15] Add io_uring IO interface

[Martin K. Petersen]

Jens,

> I bit the bullet and bumped the size. 64 bytes is a nicer size in terms
> of cachelines anyway, and I really doubt that 48 vs 64 bytes makes a
> size consumption problem for anyone.
>
> The buf_index is only used for the fixed buffers, which means that we have
> 16 bytes / 128 bits that we can grab for things like KV.

Great!

-- 
Martin K. Petersen	Oracle Linux Engineering

[PATCH 06/15] io_uring: support for IO polling

[Jens Axboe]

Add support for polled read and write commands. These act like their
non-polled counterparts, except we expect to poll for completion of
them.

To use polling, io_uring_setup() must be used with the
IORING_SETUP_IOPOLL flag being set. It is illegal to mix and match
polled and non-polled IO on an io_uring.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/io_uring.c                 | 247 ++++++++++++++++++++++++++++++++--
 include/uapi/linux/io_uring.h |   5 +
 2 files changed, 239 insertions(+), 13 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 0bad563f3486..c872bfb32a03 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -71,13 +71,17 @@ struct io_ring_ctx {
 
 	struct completion	ctx_done;
 
+	/* iopoll submission state */
 	struct {
-		struct mutex		uring_lock;
-		wait_queue_head_t	wait;
+		spinlock_t		poll_lock;
+		struct list_head	poll_submitted;
 	} ____cacheline_aligned_in_smp;
 
 	struct {
+		struct list_head	poll_completing;
 		spinlock_t		completion_lock;
+		struct mutex		uring_lock;
+		wait_queue_head_t	wait;
 	} ____cacheline_aligned_in_smp;
 };
 
@@ -97,9 +101,12 @@ struct io_kiocb {
 	unsigned long		ki_index;
 	struct list_head	ki_list;
 	unsigned long		ki_flags;
+#define KIOCB_F_IOPOLL_COMPLETED	0	/* polled IO has completed */
+#define KIOCB_F_IOPOLL_EAGAIN		1	/* submission got EAGAIN */
 };
 
 #define IO_PLUG_THRESHOLD		2
+#define IO_IOPOLL_BATCH			8
 
 struct sqe_submit {
 	const struct io_uring_sqe *sqe;
@@ -136,6 +143,9 @@ static struct io_ring_ctx *io_ring_ctx_alloc(struct io_uring_params *p)
 
 	spin_lock_init(&ctx->completion_lock);
 	init_waitqueue_head(&ctx->wait);
+	spin_lock_init(&ctx->poll_lock);
+	INIT_LIST_HEAD(&ctx->poll_submitted);
+	INIT_LIST_HEAD(&ctx->poll_completing);
 	mutex_init(&ctx->uring_lock);
 
 	return ctx;
@@ -187,12 +197,151 @@ static void io_ring_drop_ctx_ref(struct io_ring_ctx *ctx, unsigned refs)
 		wake_up(&ctx->wait);
 }
 
+static void io_free_kiocb_many(struct io_ring_ctx *ctx, void **iocbs, int *nr)
+{
+	if (*nr) {
+		kmem_cache_free_bulk(kiocb_cachep, *nr, iocbs);
+		io_ring_drop_ctx_ref(ctx, *nr);
+		*nr = 0;
+	}
+}
+
 static void io_free_kiocb(struct io_kiocb *iocb)
 {
 	kmem_cache_free(kiocb_cachep, iocb);
 	io_ring_drop_ctx_ref(iocb->ki_ctx, 1);
 }
 
+/*
+ * Find and free completed poll iocbs
+ */
+static void io_iopoll_reap(struct io_ring_ctx *ctx, unsigned int *nr_events)
+{
+	void *iocbs[IO_IOPOLL_BATCH];
+	struct io_kiocb *iocb, *n;
+	int to_free = 0;
+
+	list_for_each_entry_safe(iocb, n, &ctx->poll_completing, ki_list) {
+		if (!test_bit(KIOCB_F_IOPOLL_COMPLETED, &iocb->ki_flags))
+			continue;
+		if (to_free == ARRAY_SIZE(iocbs))
+			io_free_kiocb_many(ctx, iocbs, &to_free);
+
+		list_del(&iocb->ki_list);
+		iocbs[to_free++] = iocb;
+
+		fput(iocb->rw.ki_filp);
+		(*nr_events)++;
+	}
+
+	if (to_free)
+		io_free_kiocb_many(ctx, iocbs, &to_free);
+}
+
+/*
+ * Poll for a mininum of 'min' events, and a maximum of 'max'. Note that if
+ * min == 0 we consider that a non-spinning poll check - we'll still enter
+ * the driver poll loop, but only as a non-spinning completion check.
+ */
+static int io_iopoll_getevents(struct io_ring_ctx *ctx, unsigned int *nr_events,
+				long min)
+{
+	struct io_kiocb *iocb;
+	int found, polled, ret;
+
+	/*
+	 * Check if we already have done events that satisfy what we need
+	 */
+	if (!list_empty(&ctx->poll_completing)) {
+		io_iopoll_reap(ctx, nr_events);
+		if (min && *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.
+	 */
+	io_iopoll_reap(ctx, nr_events);
+	if (min && *nr_events >= min)
+		return 0;
+
+	polled = found = 0;
+	list_for_each_entry(iocb, &ctx->poll_completing, ki_list) {
+		/*
+		 * Poll for needed events with spin == true, anything after
+		 * that we just check if we have more, up to max.
+		 */
+		bool spin = !polled || *nr_events < min;
+		struct kiocb *kiocb = &iocb->rw;
+
+		if (test_bit(KIOCB_F_IOPOLL_COMPLETED, &iocb->ki_flags))
+			break;
+
+		found++;
+		ret = kiocb->ki_filp->f_op->iopoll(kiocb, spin);
+		if (ret < 0)
+			return ret;
+
+		polled += ret;
+	}
+
+	io_iopoll_reap(ctx, nr_events);
+	if (*nr_events >= min)
+		return 0;
+	return found;
+}
+
+/*
+ * We can't just wait for polled events to come to us, we have to actively
+ * find and complete them.
+ */
+static void io_iopoll_reap_events(struct io_ring_ctx *ctx)
+{
+	if (!(ctx->flags & IORING_SETUP_IOPOLL))
+		return;
+
+	mutex_lock(&ctx->uring_lock);
+	while (!list_empty_careful(&ctx->poll_submitted) ||
+	       !list_empty(&ctx->poll_completing)) {
+		unsigned int nr_events = 0;
+
+		io_iopoll_getevents(ctx, &nr_events, 1);
+	}
+	mutex_unlock(&ctx->uring_lock);
+}
+
+static int io_iopoll_check(struct io_ring_ctx *ctx, unsigned *nr_events,
+			   long min)
+{
+	int ret = 0;
+
+	while (!*nr_events || !need_resched()) {
+		int tmin = 0;
+
+		if (*nr_events < min)
+			tmin = min - *nr_events;
+
+		ret = io_iopoll_getevents(ctx, nr_events, tmin);
+		if (ret <= 0)
+			break;
+		ret = 0;
+	}
+
+	return ret;
+}
+
 static void kiocb_end_write(struct kiocb *kiocb)
 {
 	if (kiocb->ki_flags & IOCB_WRITE) {
@@ -208,18 +357,16 @@ static void kiocb_end_write(struct kiocb *kiocb)
 	}
 }
 
-static void io_cqring_fill_event(struct io_ring_ctx *ctx, unsigned ki_index,
-				 long res, unsigned ev_flags)
+static void __io_cqring_fill_event(struct io_ring_ctx *ctx, unsigned ki_index,
+				   long res, unsigned ev_flags)
 {
 	struct io_uring_cqe *cqe;
-	unsigned long flags;
 
 	/*
 	 * If we can't get a cq entry, userspace overflowed the
 	 * submission (by quite a lot). Increment the overflow count in
 	 * the ring.
 	 */
-	spin_lock_irqsave(&ctx->completion_lock, flags);
 	cqe = io_peek_cqring(ctx);
 	if (cqe) {
 		cqe->index = ki_index;
@@ -229,6 +376,15 @@ static void io_cqring_fill_event(struct io_ring_ctx *ctx, unsigned ki_index,
 		io_inc_cqring(ctx);
 	} else
 		ctx->cq_ring->overflow++;
+}
+
+static void io_cqring_fill_event(struct io_ring_ctx *ctx, unsigned ki_index,
+				 long res, unsigned ev_flags)
+{
+	unsigned long flags;
+
+	spin_lock_irqsave(&ctx->completion_lock, flags);
+	__io_cqring_fill_event(ctx, ki_index, res, ev_flags);
 	spin_unlock_irqrestore(&ctx->completion_lock, flags);
 }
 
@@ -243,8 +399,23 @@ static void io_complete_scqring_rw(struct kiocb *kiocb, long res, long res2)
 	io_free_kiocb(iocb);
 }
 
+static void io_complete_scqring_iopoll(struct kiocb *kiocb, long res, long res2)
+{
+	struct io_kiocb *iocb = container_of(kiocb, struct io_kiocb, rw);
+
+	kiocb_end_write(kiocb);
+
+	if (unlikely(res == -EAGAIN)) {
+		set_bit(KIOCB_F_IOPOLL_EAGAIN, &iocb->ki_flags);
+	} else {
+		__io_cqring_fill_event(iocb->ki_ctx, iocb->ki_index, res, 0);
+		set_bit(KIOCB_F_IOPOLL_COMPLETED, &iocb->ki_flags);
+	}
+}
+
 static int io_prep_rw(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe)
 {
+	struct io_ring_ctx *ctx = kiocb->ki_ctx;
 	struct kiocb *req = &kiocb->rw;
 	int ret;
 
@@ -266,12 +437,22 @@ static int io_prep_rw(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe)
 	ret = kiocb_set_rw_flags(req, sqe->rw_flags);
 	if (unlikely(ret))
 		goto out_fput;
-	if (req->ki_flags & IOCB_HIPRI) {
-		ret = -EINVAL;
-		goto out_fput;
-	}
 
-	req->ki_complete = io_complete_scqring_rw;
+	if (ctx->flags & IORING_SETUP_IOPOLL) {
+		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 = io_complete_scqring_iopoll;
+	} else {
+		if (req->ki_flags & IOCB_HIPRI) {
+			ret = -EINVAL;
+			goto out_fput;
+		}
+		req->ki_complete = io_complete_scqring_rw;
+	}
 	return 0;
 out_fput:
 	fput(req->ki_filp);
@@ -298,6 +479,30 @@ static inline void io_rw_done(struct kiocb *req, ssize_t ret)
 	}
 }
 
+/*
+ * 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 io_iopoll_kiocb_issued(struct io_kiocb *kiocb)
+{
+	/*
+	 * For fast devices, IO may have already completed. If it has, add
+	 * it to the front so we find it first. We can't add to the poll_done
+	 * list as that's unlocked from the completion side.
+	 */
+	const int front = test_bit(KIOCB_F_IOPOLL_COMPLETED, &kiocb->ki_flags);
+	struct io_ring_ctx *ctx = kiocb->ki_ctx;
+
+	spin_lock(&ctx->poll_lock);
+	if (front)
+		list_add(&kiocb->ki_list, &ctx->poll_submitted);
+	else
+		list_add_tail(&kiocb->ki_list, &ctx->poll_submitted);
+	spin_unlock(&ctx->poll_lock);
+}
+
 static ssize_t io_read(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe)
 {
 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
@@ -400,6 +605,8 @@ static int io_fsync(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe,
 {
 	struct fsync_iocb *req = &kiocb->fsync;
 
+	if (kiocb->ki_ctx->flags & IORING_SETUP_IOPOLL)
+		return -EINVAL;
 	if (unlikely(sqe->addr || sqe->off || sqe->len || sqe->__resv))
 		return -EINVAL;
 
@@ -461,6 +668,13 @@ static int io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s)
 	 */
 	if (ret)
 		goto out_put_req;
+	if (ctx->flags & IORING_SETUP_IOPOLL) {
+		if (test_bit(KIOCB_F_IOPOLL_EAGAIN, &req->ki_flags)) {
+			ret = -EAGAIN;
+			goto out_put_req;
+		}
+		io_iopoll_kiocb_issued(req);
+	}
 	return 0;
 out_put_req:
 	io_free_kiocb(req);
@@ -573,12 +787,17 @@ static int __io_uring_enter(struct io_ring_ctx *ctx, unsigned to_submit,
 			return ret;
 	}
 	if (flags & IORING_ENTER_GETEVENTS) {
+		unsigned nr_events = 0;
 		int get_ret;
 
 		if (!ret && to_submit)
 			min_complete = 0;
 
-		get_ret = io_cqring_wait(ctx, min_complete);
+		if (ctx->flags & IORING_SETUP_IOPOLL)
+			get_ret = io_iopoll_check(ctx, &nr_events,
+							min_complete);
+		else
+			get_ret = io_cqring_wait(ctx, min_complete);
 		if (get_ret < 0 && !ret)
 			ret = get_ret;
 	}
@@ -604,6 +823,7 @@ static void io_free_scq_urings(struct io_ring_ctx *ctx)
 
 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
 {
+	io_iopoll_reap_events(ctx);
 	io_free_scq_urings(ctx);
 	percpu_ref_exit(&ctx->refs);
 	kfree(ctx);
@@ -612,6 +832,7 @@ static void io_ring_ctx_free(struct io_ring_ctx *ctx)
 static void io_ring_ctx_wait_and_kill(struct io_ring_ctx *ctx)
 {
 	percpu_ref_kill(&ctx->refs);
+	io_iopoll_reap_events(ctx);
 	wait_for_completion(&ctx->ctx_done);
 	io_ring_ctx_free(ctx);
 }
@@ -815,7 +1036,7 @@ SYSCALL_DEFINE3(io_uring_setup, u32, entries, struct iovec __user *, iovecs,
 			return -EINVAL;
 	}
 
-	if (p.flags)
+	if (p.flags & ~IORING_SETUP_IOPOLL)
 		return -EINVAL;
 	if (iovecs)
 		return -EINVAL;
diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
index ae30ed41965f..ba9e5b851f73 100644
--- a/include/uapi/linux/io_uring.h
+++ b/include/uapi/linux/io_uring.h
@@ -31,6 +31,11 @@ struct io_uring_sqe {
 	};
 };
 
+/*
+ * io_uring_setup() flags
+ */
+#define IORING_SETUP_IOPOLL	(1 << 0)	/* io_context is polled */
+
 #define IORING_OP_READV		1
 #define IORING_OP_WRITEV	2
 #define IORING_OP_FSYNC		3
-- 
2.17.1

[PATCH 07/15] io_uring: add submission side request cache

[Jens Axboe]

We have to add each submitted polled request to the io_ring_ctx
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/io_uring.c | 122 +++++++++++++++++++++++++++++++++++++++++++-------
 1 file changed, 106 insertions(+), 16 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index c872bfb32a03..f7938156552f 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -113,6 +113,21 @@ struct sqe_submit {
 	unsigned index;
 };
 
+struct io_submit_state {
+	struct io_ring_ctx *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 struct kmem_cache *kiocb_cachep;
 
 static const struct file_operations io_scqring_fops;
@@ -480,21 +495,29 @@ static inline void io_rw_done(struct kiocb *req, ssize_t ret)
 }
 
 /*
- * 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.
+ * 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 io_iopoll_kiocb_issued(struct io_kiocb *kiocb)
+static void io_flush_state_reqs(struct io_ring_ctx *ctx,
+				 struct io_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;
+}
+
+static void io_iopoll_iocb_add_list(struct io_kiocb *kiocb)
+{
+	const int front = test_bit(KIOCB_F_IOPOLL_COMPLETED, &kiocb->ki_flags);
+	struct io_ring_ctx *ctx = kiocb->ki_ctx;
+
 	/*
 	 * For fast devices, IO may have already completed. If it has, add
 	 * it to the front so we find it first. We can't add to the poll_done
 	 * list as that's unlocked from the completion side.
 	 */
-	const int front = test_bit(KIOCB_F_IOPOLL_COMPLETED, &kiocb->ki_flags);
-	struct io_ring_ctx *ctx = kiocb->ki_ctx;
-
 	spin_lock(&ctx->poll_lock);
 	if (front)
 		list_add(&kiocb->ki_list, &ctx->poll_submitted);
@@ -503,6 +526,33 @@ static void io_iopoll_kiocb_issued(struct io_kiocb *kiocb)
 	spin_unlock(&ctx->poll_lock);
 }
 
+static void io_iopoll_iocb_add_state(struct io_submit_state *state,
+				     struct io_kiocb *kiocb)
+{
+	if (test_bit(KIOCB_F_IOPOLL_COMPLETED, &kiocb->ki_flags))
+		list_add(&kiocb->ki_list, &state->req_list);
+	else
+		list_add_tail(&kiocb->ki_list, &state->req_list);
+
+	if (++state->req_count >= IO_IOPOLL_BATCH)
+		io_flush_state_reqs(state->ctx, state);
+}
+
+/*
+ * 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 io_iopoll_kiocb_issued(struct io_submit_state *state,
+				   struct io_kiocb *kiocb)
+{
+	if (!state || !IS_ENABLED(CONFIG_BLOCK))
+		io_iopoll_iocb_add_list(kiocb);
+	else
+		io_iopoll_iocb_add_state(state, kiocb);
+}
+
 static ssize_t io_read(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe)
 {
 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
@@ -624,7 +674,8 @@ static int io_fsync(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe,
 	return 0;
 }
 
-static int io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s)
+static int io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s,
+			 struct io_submit_state *state)
 {
 	const struct io_uring_sqe *sqe = s->sqe;
 	struct io_kiocb *req;
@@ -673,7 +724,7 @@ static int io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s)
 			ret = -EAGAIN;
 			goto out_put_req;
 		}
-		io_iopoll_kiocb_issued(req);
+		io_iopoll_kiocb_issued(state, req);
 	}
 	return 0;
 out_put_req:
@@ -681,6 +732,43 @@ static int io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s)
 	return ret;
 }
 
+#ifdef CONFIG_BLOCK
+static void io_state_unplug(struct blk_plug_cb *cb, bool from_schedule)
+{
+	struct io_submit_state *state;
+
+	state = container_of(cb, struct io_submit_state, plug_cb);
+	if (!list_empty(&state->req_list))
+		io_flush_state_reqs(state->ctx, state);
+}
+#endif
+
+/*
+ * Batched submission is done, ensure local IO is flushed out.
+ */
+static void io_submit_state_end(struct io_submit_state *state)
+{
+	blk_finish_plug(&state->plug);
+	if (!list_empty(&state->req_list))
+		io_flush_state_reqs(state->ctx, state);
+}
+
+/*
+ * Start submission side cache.
+ */
+static void io_submit_state_start(struct io_submit_state *state,
+				  struct io_ring_ctx *ctx)
+{
+	state->ctx = ctx;
+	INIT_LIST_HEAD(&state->req_list);
+	state->req_count = 0;
+#ifdef CONFIG_BLOCK
+	state->plug_cb.callback = io_state_unplug;
+	blk_start_plug(&state->plug);
+	list_add(&state->plug_cb.list, &state->plug.cb_list);
+#endif
+}
+
 static void io_inc_sqring(struct io_ring_ctx *ctx)
 {
 	struct io_sq_ring *ring = ctx->sq_ring;
@@ -715,11 +803,13 @@ static bool io_peek_sqring(struct io_ring_ctx *ctx, struct sqe_submit *s)
 
 static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit)
 {
+	struct io_submit_state state, *statep = NULL;
 	int i, ret = 0, submit = 0;
-	struct blk_plug plug;
 
-	if (to_submit > IO_PLUG_THRESHOLD)
-		blk_start_plug(&plug);
+	if (to_submit > IO_PLUG_THRESHOLD) {
+		io_submit_state_start(&state, ctx);
+		statep = &state;
+	}
 
 	for (i = 0; i < to_submit; i++) {
 		struct sqe_submit s;
@@ -727,7 +817,7 @@ static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit)
 		if (!io_peek_sqring(ctx, &s))
 			break;
 
-		ret = io_submit_sqe(ctx, &s);
+		ret = io_submit_sqe(ctx, &s, statep);
 		if (ret)
 			break;
 
@@ -735,8 +825,8 @@ static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit)
 		io_inc_sqring(ctx);
 	}
 
-	if (to_submit > IO_PLUG_THRESHOLD)
-		blk_finish_plug(&plug);
+	if (statep)
+		io_submit_state_end(statep);
 
 	return submit ? submit : ret;
 }
-- 
2.17.1

[PATCH 08/15] 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      |  9 +++++++--
 include/linux/file.h |  2 ++
 include/linux/fs.h   |  4 +++-
 4 files changed, 22 insertions(+), 8 deletions(-)

diff --git a/fs/file.c b/fs/file.c
index 3209ee271c41..e0d7ce70e860 100644
--- a/fs/file.c
+++ b/fs/file.c
@@ -705,7 +705,7 @@ 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;
@@ -720,7 +720,7 @@ 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();
@@ -728,15 +728,20 @@ 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);
 
@@ -767,7 +772,7 @@ 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 5679e7fcb6b0..155d7514a094 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 +347,11 @@ 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 ccb0b7a63aa5..acaad78b6781 100644
--- a/include/linux/fs.h
+++ b/include/linux/fs.h
@@ -952,7 +952,9 @@ 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 09/15] io_uring: use fget/fput_many() for file references

[Jens Axboe]

On the submission side, add file reference batching to the
io_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 io_iopoll_reap().

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/io_uring.c | 105 +++++++++++++++++++++++++++++++++++++++++++-------
 1 file changed, 92 insertions(+), 13 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index f7938156552f..cd2dfc153338 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -126,6 +126,15 @@ struct io_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 struct kmem_cache *kiocb_cachep;
@@ -234,7 +243,8 @@ static void io_iopoll_reap(struct io_ring_ctx *ctx, unsigned int *nr_events)
 {
 	void *iocbs[IO_IOPOLL_BATCH];
 	struct io_kiocb *iocb, *n;
-	int to_free = 0;
+	int file_count, to_free = 0;
+	struct file *file = NULL;
 
 	list_for_each_entry_safe(iocb, n, &ctx->poll_completing, ki_list) {
 		if (!test_bit(KIOCB_F_IOPOLL_COMPLETED, &iocb->ki_flags))
@@ -245,10 +255,27 @@ static void io_iopoll_reap(struct io_ring_ctx *ctx, unsigned int *nr_events)
 		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;
+		}
+
 		(*nr_events)++;
 	}
 
+	if (file)
+		fput_many(file, file_count);
+
 	if (to_free)
 		io_free_kiocb_many(ctx, iocbs, &to_free);
 }
@@ -428,13 +455,60 @@ static void io_complete_scqring_iopoll(struct kiocb *kiocb, long res, long res2)
 	}
 }
 
-static int io_prep_rw(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe)
+static void io_file_put(struct io_submit_state *state, struct file *file)
+{
+	if (!state) {
+		fput(file);
+	} else 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 *io_file_get(struct io_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;
+	}
+
+	io_file_put(state, NULL);
+	goto get_file;
+}
+
+static int io_prep_rw(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe,
+		      struct io_submit_state *state)
 {
 	struct io_ring_ctx *ctx = kiocb->ki_ctx;
 	struct kiocb *req = &kiocb->rw;
 	int ret;
 
-	req->ki_filp = fget(sqe->fd);
+	req->ki_filp = io_file_get(state, sqe->fd);
 	if (unlikely(!req->ki_filp))
 		return -EBADF;
 	req->ki_pos = sqe->off;
@@ -470,7 +544,7 @@ static int io_prep_rw(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe)
 	}
 	return 0;
 out_fput:
-	fput(req->ki_filp);
+	io_file_put(state, req->ki_filp);
 	return ret;
 }
 
@@ -553,7 +627,8 @@ static void io_iopoll_kiocb_issued(struct io_submit_state *state,
 		io_iopoll_iocb_add_state(state, kiocb);
 }
 
-static ssize_t io_read(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe)
+static ssize_t io_read(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe,
+		       struct io_submit_state *state)
 {
 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
 	void __user *buf = (void __user *) (uintptr_t) sqe->addr;
@@ -562,7 +637,7 @@ static ssize_t io_read(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe)
 	struct file *file;
 	ssize_t ret;
 
-	ret = io_prep_rw(kiocb, sqe);
+	ret = io_prep_rw(kiocb, sqe, state);
 	if (ret)
 		return ret;
 	file = req->ki_filp;
@@ -588,7 +663,8 @@ static ssize_t io_read(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe)
 	return ret;
 }
 
-static ssize_t io_write(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe)
+static ssize_t io_write(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe,
+			struct io_submit_state *state)
 {
 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
 	void __user *buf = (void __user *) (uintptr_t) sqe->addr;
@@ -597,7 +673,7 @@ static ssize_t io_write(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe)
 	struct file *file;
 	ssize_t ret;
 
-	ret = io_prep_rw(kiocb, sqe);
+	ret = io_prep_rw(kiocb, sqe, state);
 	if (ret)
 		return ret;
 	file = req->ki_filp;
@@ -697,10 +773,10 @@ static int io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s,
 	ret = -EINVAL;
 	switch (sqe->opcode) {
 	case IORING_OP_READV:
-		ret = io_read(req, sqe);
+		ret = io_read(req, sqe, state);
 		break;
 	case IORING_OP_WRITEV:
-		ret = io_write(req, sqe);
+		ret = io_write(req, sqe, state);
 		break;
 	case IORING_OP_FSYNC:
 		ret = io_fsync(req, sqe, false);
@@ -751,17 +827,20 @@ static void io_submit_state_end(struct io_submit_state *state)
 	blk_finish_plug(&state->plug);
 	if (!list_empty(&state->req_list))
 		io_flush_state_reqs(state->ctx, state);
+	io_file_put(state, NULL);
 }
 
 /*
  * Start submission side cache.
  */
 static void io_submit_state_start(struct io_submit_state *state,
-				  struct io_ring_ctx *ctx)
+				  struct io_ring_ctx *ctx, unsigned 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 = io_state_unplug;
 	blk_start_plug(&state->plug);
@@ -807,7 +886,7 @@ static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit)
 	int i, ret = 0, submit = 0;
 
 	if (to_submit > IO_PLUG_THRESHOLD) {
-		io_submit_state_start(&state, ctx);
+		io_submit_state_start(&state, ctx, to_submit);
 		statep = &state;
 	}
 
-- 
2.17.1

[PATCH 10/15] io_uring: batch io_kiocb allocation

[Jens Axboe]

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

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/io_uring.c | 71 +++++++++++++++++++++++++++++++++++++--------------
 1 file changed, 52 insertions(+), 19 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index cd2dfc153338..b5233786b5a8 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -127,6 +127,13 @@ struct io_submit_state {
 	struct list_head req_list;
 	unsigned int req_count;
 
+	/*
+	 * io_kiocb alloc cache
+	 */
+	void *kiocbs[IO_IOPOLL_BATCH];
+	unsigned int free_kiocbs;
+	unsigned int cur_kiocb;
+
 	/*
 	 * File reference cache
 	 */
@@ -196,36 +203,58 @@ static struct io_uring_cqe *io_peek_cqring(struct io_ring_ctx *ctx)
 	return &ring->cqes[tail & ctx->cq_mask];
 }
 
-static struct io_kiocb *io_get_kiocb(struct io_ring_ctx *ctx)
+static void io_ring_drop_ctx_refs(struct io_ring_ctx *ctx, unsigned refs)
+{
+	percpu_ref_put_many(&ctx->refs, refs);
+
+	if (waitqueue_active(&ctx->wait))
+		wake_up(&ctx->wait);
+}
+
+static struct io_kiocb *io_get_kiocb(struct io_ring_ctx *ctx,
+				   struct io_submit_state *state)
 {
 	struct io_kiocb *req;
 
 	if (!percpu_ref_tryget(&ctx->refs))
 		return NULL;
 
-	req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
-	if (!req)
-		return NULL;
-
-	req->ki_ctx = ctx;
-	INIT_LIST_HEAD(&req->ki_list);
-	req->ki_flags = 0;
-	return req;
-}
+	if (!state)
+		req = kmem_cache_alloc(kiocb_cachep, GFP_KERNEL);
+	else if (!state->free_kiocbs) {
+		size_t sz;
+		int ret;
+
+		sz = min_t(size_t, state->ios_left, ARRAY_SIZE(state->kiocbs));
+		ret = kmem_cache_alloc_bulk(kiocb_cachep, GFP_KERNEL, sz,
+						state->kiocbs);
+		if (ret <= 0)
+			goto out;
+		state->free_kiocbs = ret - 1;
+		state->cur_kiocb = 1;
+		req = state->kiocbs[0];
+	} else {
+		req = state->kiocbs[state->cur_kiocb];
+		state->free_kiocbs--;
+		state->cur_kiocb++;
+	}
 
-static void io_ring_drop_ctx_ref(struct io_ring_ctx *ctx, unsigned refs)
-{
-	percpu_ref_put_many(&ctx->refs, refs);
+	if (req) {
+		req->ki_ctx = ctx;
+		req->ki_flags = 0;
+		return req;
+	}
 
-	if (waitqueue_active(&ctx->wait))
-		wake_up(&ctx->wait);
+out:
+	io_ring_drop_ctx_refs(ctx, 1);
+	return NULL;
 }
 
 static void io_free_kiocb_many(struct io_ring_ctx *ctx, void **iocbs, int *nr)
 {
 	if (*nr) {
 		kmem_cache_free_bulk(kiocb_cachep, *nr, iocbs);
-		io_ring_drop_ctx_ref(ctx, *nr);
+		io_ring_drop_ctx_refs(ctx, *nr);
 		*nr = 0;
 	}
 }
@@ -233,7 +262,7 @@ static void io_free_kiocb_many(struct io_ring_ctx *ctx, void **iocbs, int *nr)
 static void io_free_kiocb(struct io_kiocb *iocb)
 {
 	kmem_cache_free(kiocb_cachep, iocb);
-	io_ring_drop_ctx_ref(iocb->ki_ctx, 1);
+	io_ring_drop_ctx_refs(iocb->ki_ctx, 1);
 }
 
 /*
@@ -761,7 +790,7 @@ static int io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s,
 	if (unlikely(sqe->flags))
 		return -EINVAL;
 
-	req = io_get_kiocb(ctx);
+	req = io_get_kiocb(ctx, state);
 	if (unlikely(!req))
 		return -EAGAIN;
 
@@ -828,6 +857,9 @@ static void io_submit_state_end(struct io_submit_state *state)
 	if (!list_empty(&state->req_list))
 		io_flush_state_reqs(state->ctx, state);
 	io_file_put(state, NULL);
+	if (state->free_kiocbs)
+		kmem_cache_free_bulk(kiocb_cachep, state->free_kiocbs,
+					&state->kiocbs[state->cur_kiocb]);
 }
 
 /*
@@ -839,6 +871,7 @@ static void io_submit_state_start(struct io_submit_state *state,
 	state->ctx = ctx;
 	INIT_LIST_HEAD(&state->req_list);
 	state->req_count = 0;
+	state->free_kiocbs = 0;
 	state->file = NULL;
 	state->ios_left = max_ios;
 #ifdef CONFIG_BLOCK
@@ -1071,7 +1104,7 @@ SYSCALL_DEFINE4(io_uring_enter, unsigned int, fd, u32, to_submit,
 		ret = __io_uring_enter(ctx, to_submit, min_complete, flags);
 		mutex_unlock(&ctx->uring_lock);
 	}
-	io_ring_drop_ctx_ref(ctx, 1);
+	io_ring_drop_ctx_refs(ctx, 1);
 out_fput:
 	fdput(f);
 	return ret;
-- 
2.17.1

[PATCH 11/15] block: implement bio helper to add iter bvec pages to bio

[Jens Axboe]

For an ITER_BVEC, we can just iterate the iov and add the pages
to the bio directly. This requires that the caller doesn't releases
the pages on IO completion, we add a BIO_HOLD_PAGES flag for that.

The current two callers of bio_iov_iter_get_pages() are updated to
check if they need to release pages on completion. This makes them
work with bvecs that contain kernel mapped pages already.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 block/bio.c               | 59 ++++++++++++++++++++++++++++++++-------
 fs/block_dev.c            |  5 ++--
 fs/iomap.c                |  5 ++--
 include/linux/blk_types.h |  1 +
 4 files changed, 56 insertions(+), 14 deletions(-)

diff --git a/block/bio.c b/block/bio.c
index 4db1008309ed..7af4f45d2ed6 100644
--- a/block/bio.c
+++ b/block/bio.c
@@ -828,6 +828,23 @@ int bio_add_page(struct bio *bio, struct page *page,
 }
 EXPORT_SYMBOL(bio_add_page);
 
+static int __bio_iov_bvec_add_pages(struct bio *bio, struct iov_iter *iter)
+{
+	const struct bio_vec *bv = iter->bvec;
+	unsigned int len;
+	size_t size;
+
+	len = min_t(size_t, bv->bv_len, iter->count);
+	size = bio_add_page(bio, bv->bv_page, len,
+				bv->bv_offset + iter->iov_offset);
+	if (size == len) {
+		iov_iter_advance(iter, size);
+		return 0;
+	}
+
+	return -EINVAL;
+}
+
 #define PAGE_PTRS_PER_BVEC     (sizeof(struct bio_vec) / sizeof(struct page *))
 
 /**
@@ -876,23 +893,43 @@ static int __bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter)
 }
 
 /**
- * bio_iov_iter_get_pages - pin user or kernel pages and add them to a bio
+ * bio_iov_iter_get_pages - add user or kernel pages to a bio
  * @bio: bio to add pages to
- * @iter: iov iterator describing the region to be mapped
+ * @iter: iov iterator describing the region to be added
+ *
+ * This takes either an iterator pointing to user memory, or one pointing to
+ * kernel pages (BVEC iterator). If we're adding user pages, we pin them and
+ * map them into the kernel. On IO completion, the caller should put those
+ * pages. If we're adding kernel pages, we just have to add the pages to the
+ * bio directly. We don't grab an extra reference to those pages (the user
+ * should already have that), and we don't put the page on IO completion.
+ * The caller needs to check if the bio is flagged BIO_HOLD_PAGES on IO
+ * completion. If it isn't, then pages should be released.
  *
- * Pins pages from *iter and appends them to @bio's bvec array. The
- * pages will have to be released using put_page() when done.
  * The function tries, but does not guarantee, to pin as many pages as
- * fit into the bio, or are requested in *iter, whatever is smaller.
- * If MM encounters an error pinning the requested pages, it stops.
- * Error is returned only if 0 pages could be pinned.
+ * fit into the bio, or are requested in *iter, whatever is smaller. If
+ * MM encounters an error pinning the requested pages, it stops. Error
+ * is returned only if 0 pages could be pinned.
  */
 int bio_iov_iter_get_pages(struct bio *bio, struct iov_iter *iter)
 {
+	const bool is_bvec = iov_iter_is_bvec(iter);
 	unsigned short orig_vcnt = bio->bi_vcnt;
 
+	/*
+	 * If this is a BVEC iter, then the pages are kernel pages. Don't
+	 * release them on IO completion.
+	 */
+	if (is_bvec)
+		bio_set_flag(bio, BIO_HOLD_PAGES);
+
 	do {
-		int ret = __bio_iov_iter_get_pages(bio, iter);
+		int ret;
+
+		if (is_bvec)
+			ret = __bio_iov_bvec_add_pages(bio, iter);
+		else
+			ret = __bio_iov_iter_get_pages(bio, iter);
 
 		if (unlikely(ret))
 			return bio->bi_vcnt > orig_vcnt ? 0 : ret;
@@ -1634,7 +1671,8 @@ static void bio_dirty_fn(struct work_struct *work)
 		next = bio->bi_private;
 
 		bio_set_pages_dirty(bio);
-		bio_release_pages(bio);
+		if (!bio_flagged(bio, BIO_HOLD_PAGES))
+			bio_release_pages(bio);
 		bio_put(bio);
 	}
 }
@@ -1650,7 +1688,8 @@ void bio_check_pages_dirty(struct bio *bio)
 			goto defer;
 	}
 
-	bio_release_pages(bio);
+	if (!bio_flagged(bio, BIO_HOLD_PAGES))
+		bio_release_pages(bio);
 	bio_put(bio);
 	return;
 defer:
diff --git a/fs/block_dev.c b/fs/block_dev.c
index 2ebd2a0d7789..b7742014c9de 100644
--- a/fs/block_dev.c
+++ b/fs/block_dev.c
@@ -324,8 +324,9 @@ static void blkdev_bio_end_io(struct bio *bio)
 		struct bio_vec *bvec;
 		int i;
 
-		bio_for_each_segment_all(bvec, bio, i)
-			put_page(bvec->bv_page);
+		if (!bio_flagged(bio, BIO_HOLD_PAGES))
+			bio_for_each_segment_all(bvec, bio, i)
+				put_page(bvec->bv_page);
 		bio_put(bio);
 	}
 }
diff --git a/fs/iomap.c b/fs/iomap.c
index 4ee50b76b4a1..0a64c9c51203 100644
--- a/fs/iomap.c
+++ b/fs/iomap.c
@@ -1582,8 +1582,9 @@ static void iomap_dio_bio_end_io(struct bio *bio)
 		struct bio_vec *bvec;
 		int i;
 
-		bio_for_each_segment_all(bvec, bio, i)
-			put_page(bvec->bv_page);
+		if (!bio_flagged(bio, BIO_HOLD_PAGES))
+			bio_for_each_segment_all(bvec, bio, i)
+				put_page(bvec->bv_page);
 		bio_put(bio);
 	}
 }
diff --git a/include/linux/blk_types.h b/include/linux/blk_types.h
index 5c7e7f859a24..97e206855cd3 100644
--- a/include/linux/blk_types.h
+++ b/include/linux/blk_types.h
@@ -215,6 +215,7 @@ struct bio {
 /*
  * bio flags
  */
+#define BIO_HOLD_PAGES	0	/* don't put O_DIRECT pages */
 #define BIO_SEG_VALID	1	/* bi_phys_segments valid */
 #define BIO_CLONED	2	/* doesn't own data */
 #define BIO_BOUNCED	3	/* bio is a bounce bio */
-- 
2.17.1

[PATCH 12/15] io_uring: add support for pre-mapped user IO buffers

[Jens Axboe]

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

To utilize this feature, the application must pass in an array of iovecs
that contain the desired buffer addresses and lengths. These buffers can
then be mapped into the kernel for the life time of the io_uring, as
opposed to just the duration of the each single IO. The application can
then use the IORING_OP_{READ,WRITE}_FIXED to perform IO to these fixed
locations.

It's perfectly valid to setup a larger buffer, and then sometimes only
use parts of it for an IO. As long as the range is within the originally
mapped region, it will work just fine.

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

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/io_uring.c                 | 202 ++++++++++++++++++++++++++++++++--
 include/uapi/linux/io_uring.h |   2 +
 2 files changed, 196 insertions(+), 8 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index b5233786b5a8..7ab20258e39b 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -23,6 +23,8 @@
 #include <linux/workqueue.h>
 #include <linux/blkdev.h>
 #include <linux/anon_inodes.h>
+#include <linux/sizes.h>
+#include <linux/nospec.h>
 
 #include <linux/uaccess.h>
 #include <linux/nospec.h>
@@ -53,6 +55,13 @@ struct io_cq_ring {
 	struct io_uring_cqe	cqes[];
 };
 
+struct io_mapped_ubuf {
+	u64		ubuf;
+	size_t		len;
+	struct		bio_vec *bvec;
+	unsigned int	nr_bvecs;
+};
+
 struct io_ring_ctx {
 	struct percpu_ref	refs;
 
@@ -69,6 +78,9 @@ struct io_ring_ctx {
 	unsigned		cq_entries;
 	unsigned		cq_mask;
 
+	/* if used, fixed mapped user buffers */
+	struct io_mapped_ubuf	*user_bufs;
+
 	struct completion	ctx_done;
 
 	/* iopoll submission state */
@@ -656,11 +668,42 @@ static void io_iopoll_kiocb_issued(struct io_submit_state *state,
 		io_iopoll_iocb_add_state(state, kiocb);
 }
 
+static int io_import_fixed(int rw, struct io_kiocb *kiocb,
+			   const struct io_uring_sqe *sqe,
+			   struct iov_iter *iter)
+{
+	struct io_ring_ctx *ctx = kiocb->ki_ctx;
+	struct io_mapped_ubuf *imu;
+	size_t len = sqe->len;
+	size_t offset;
+	int index;
+
+	/* attempt to use fixed buffers without having provided iovecs */
+	if (!ctx->user_bufs)
+		return -EFAULT;
+
+	/* io_submit_sqe() already validated the index */
+	index = array_index_nospec(kiocb->ki_index, ctx->sq_entries);
+	imu = &ctx->user_bufs[index];
+	if ((unsigned long) sqe->addr < imu->ubuf ||
+	    (unsigned long) sqe->addr + len > imu->ubuf + imu->len)
+		return -EFAULT;
+
+	/*
+	 * May not be a start of buffer, set size appropriately
+	 * and advance us to the beginning.
+	 */
+	offset = (unsigned long) sqe->addr - imu->ubuf;
+	iov_iter_bvec(iter, rw, imu->bvec, imu->nr_bvecs, offset + len);
+	if (offset)
+		iov_iter_advance(iter, offset);
+	return 0;
+}
+
 static ssize_t io_read(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe,
 		       struct io_submit_state *state)
 {
 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
-	void __user *buf = (void __user *) (uintptr_t) sqe->addr;
 	struct kiocb *req = &kiocb->rw;
 	struct iov_iter iter;
 	struct file *file;
@@ -678,7 +721,15 @@ static ssize_t io_read(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe,
 	if (unlikely(!file->f_op->read_iter))
 		goto out_fput;
 
-	ret = import_iovec(READ, buf, sqe->len, UIO_FASTIOV, &iovec, &iter);
+	if (sqe->opcode == IORING_OP_READV) {
+		void __user *buf = (void __user *) (uintptr_t) sqe->addr;
+
+		ret = import_iovec(READ, buf, sqe->len, UIO_FASTIOV, &iovec,
+					&iter);
+	} else {
+		ret = io_import_fixed(READ, kiocb, sqe, &iter);
+		iovec = NULL;
+	}
 	if (ret)
 		goto out_fput;
 
@@ -696,7 +747,6 @@ static ssize_t io_write(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe,
 			struct io_submit_state *state)
 {
 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
-	void __user *buf = (void __user *) (uintptr_t) sqe->addr;
 	struct kiocb *req = &kiocb->rw;
 	struct iov_iter iter;
 	struct file *file;
@@ -714,7 +764,14 @@ static ssize_t io_write(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe,
 	if (unlikely(!file->f_op->write_iter))
 		goto out_fput;
 
-	ret = import_iovec(WRITE, buf, sqe->len, UIO_FASTIOV, &iovec, &iter);
+	if (sqe->opcode == IORING_OP_WRITEV) {
+		void __user *buf = (void __user *) (uintptr_t) sqe->addr;
+
+		ret = import_iovec(WRITE, buf, sqe->len, UIO_FASTIOV, &iovec, &iter);
+	} else {
+		ret = io_import_fixed(WRITE, kiocb, sqe, &iter);
+		iovec = NULL;
+	}
 	if (ret)
 		goto out_fput;
 
@@ -802,9 +859,11 @@ static int io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s,
 	ret = -EINVAL;
 	switch (sqe->opcode) {
 	case IORING_OP_READV:
+	case IORING_OP_READ_FIXED:
 		ret = io_read(req, sqe, state);
 		break;
 	case IORING_OP_WRITEV:
+	case IORING_OP_WRITE_FIXED:
 		ret = io_write(req, sqe, state);
 		break;
 	case IORING_OP_FSYNC:
@@ -1007,6 +1066,127 @@ static int __io_uring_enter(struct io_ring_ctx *ctx, unsigned to_submit,
 	return ret;
 }
 
+static void io_sqe_buffer_unmap(struct io_ring_ctx *ctx)
+{
+	int i, j;
+
+	if (!ctx->user_bufs)
+		return;
+
+	for (i = 0; i < ctx->sq_entries; i++) {
+		struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
+
+		for (j = 0; j < imu->nr_bvecs; j++)
+			put_page(imu->bvec[j].bv_page);
+
+		kfree(imu->bvec);
+		imu->nr_bvecs = 0;
+	}
+
+	kfree(ctx->user_bufs);
+	ctx->user_bufs = NULL;
+}
+
+static int io_sqe_buffer_map(struct io_ring_ctx *ctx,
+			     struct iovec __user *iovecs)
+{
+	unsigned long total_pages, page_limit;
+	struct page **pages = NULL;
+	int i, j, got_pages = 0;
+	int ret = -EINVAL;
+
+	ctx->user_bufs = kcalloc(ctx->sq_entries, sizeof(struct io_mapped_ubuf),
+					GFP_KERNEL);
+	if (!ctx->user_bufs)
+		return -ENOMEM;
+
+	/* Don't allow more pages than we can safely lock */
+	total_pages = 0;
+	page_limit = rlimit(RLIMIT_MEMLOCK) >> PAGE_SHIFT;
+
+	for (i = 0; i < ctx->sq_entries; i++) {
+		struct io_mapped_ubuf *imu = &ctx->user_bufs[i];
+		unsigned long off, start, end, ubuf;
+		int pret, nr_pages;
+		struct iovec iov;
+		size_t size;
+
+		ret = -EFAULT;
+		if (copy_from_user(&iov, &iovecs[i], sizeof(iov)))
+			goto err;
+
+		/*
+		 * Don't impose further limits on the size and buffer
+		 * constraints here, we'll -EINVAL later when IO is
+		 * submitted if they are wrong.
+		 */
+		ret = -EFAULT;
+		if (!iov.iov_base)
+			goto err;
+
+		/* arbitrary limit, but we need something */
+		if (iov.iov_len > SZ_4M)
+			goto err;
+
+		ubuf = (unsigned long) iov.iov_base;
+		end = (ubuf + iov.iov_len + PAGE_SIZE - 1) >> PAGE_SHIFT;
+		start = ubuf >> PAGE_SHIFT;
+		nr_pages = end - start;
+
+		ret = -ENOMEM;
+		if (total_pages + nr_pages > page_limit)
+			goto err;
+
+		if (!pages || nr_pages > got_pages) {
+			kfree(pages);
+			pages = kmalloc(nr_pages * sizeof(struct page *),
+					GFP_KERNEL);
+			if (!pages)
+				goto err;
+			got_pages = nr_pages;
+		}
+
+		imu->bvec = kmalloc(nr_pages * sizeof(struct bio_vec),
+					GFP_KERNEL);
+		if (!imu->bvec)
+			goto err;
+
+		down_write(&current->mm->mmap_sem);
+		pret = get_user_pages_longterm(ubuf, nr_pages, 1, pages, NULL);
+		up_write(&current->mm->mmap_sem);
+
+		if (pret < nr_pages) {
+			if (pret < 0)
+				ret = pret;
+			goto err;
+		}
+
+		off = ubuf & ~PAGE_MASK;
+		size = iov.iov_len;
+		for (j = 0; j < nr_pages; j++) {
+			size_t vec_len;
+
+			vec_len = min_t(size_t, size, PAGE_SIZE - off);
+			imu->bvec[j].bv_page = pages[j];
+			imu->bvec[j].bv_len = vec_len;
+			imu->bvec[j].bv_offset = off;
+			off = 0;
+			size -= vec_len;
+		}
+		/* store original address for later verification */
+		imu->ubuf = ubuf;
+		imu->len = iov.iov_len;
+		imu->nr_bvecs = nr_pages;
+		total_pages += nr_pages;
+	}
+	kfree(pages);
+	return 0;
+err:
+	kfree(pages);
+	io_sqe_buffer_unmap(ctx);
+	return ret;
+}
+
 static void io_free_scq_urings(struct io_ring_ctx *ctx)
 {
 	if (ctx->sq_ring) {
@@ -1027,6 +1207,7 @@ static void io_ring_ctx_free(struct io_ring_ctx *ctx)
 {
 	io_iopoll_reap_events(ctx);
 	io_free_scq_urings(ctx);
+	io_sqe_buffer_unmap(ctx);
 	percpu_ref_exit(&ctx->refs);
 	kfree(ctx);
 }
@@ -1185,7 +1366,8 @@ static void io_fill_offsets(struct io_uring_params *p)
 	p->cq_off.cqes = offsetof(struct io_cq_ring, cqes);
 }
 
-static int io_uring_create(unsigned entries, struct io_uring_params *p)
+static int io_uring_create(unsigned entries, struct io_uring_params *p,
+			   struct iovec __user *iovecs)
 {
 	struct io_ring_ctx *ctx;
 	int ret;
@@ -1207,6 +1389,12 @@ static int io_uring_create(unsigned entries, struct io_uring_params *p)
 	if (ret)
 		goto err;
 
+	if (iovecs) {
+		ret = io_sqe_buffer_map(ctx, iovecs);
+		if (ret)
+			goto err;
+	}
+
 	ret = anon_inode_getfd("[io_uring]", &io_scqring_fops, ctx,
 				O_RDWR | O_CLOEXEC);
 	if (ret < 0)
@@ -1240,10 +1428,8 @@ SYSCALL_DEFINE3(io_uring_setup, u32, entries, struct iovec __user *, iovecs,
 
 	if (p.flags & ~IORING_SETUP_IOPOLL)
 		return -EINVAL;
-	if (iovecs)
-		return -EINVAL;
 
-	ret = io_uring_create(entries, &p);
+	ret = io_uring_create(entries, &p, iovecs);
 	if (ret < 0)
 		return ret;
 
diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
index ba9e5b851f73..80d1a8224b9c 100644
--- a/include/uapi/linux/io_uring.h
+++ b/include/uapi/linux/io_uring.h
@@ -40,6 +40,8 @@ struct io_uring_sqe {
 #define IORING_OP_WRITEV	2
 #define IORING_OP_FSYNC		3
 #define IORING_OP_FDSYNC	4
+#define IORING_OP_READ_FIXED	5
+#define IORING_OP_WRITE_FIXED	6
 
 /*
  * IO completion data structure (Completion Queue Entry)
-- 
2.17.1

[PATCH 13/15] io_uring: support kernel side submission

[Jens Axboe]

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

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

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

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

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

See this sample application for how to use it:

http://git.kernel.dk/cgit/fio/plain/t/io_uring.c

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/io_uring.c                 | 378 ++++++++++++++++++++++++++++++++--
 include/uapi/linux/io_uring.h |   5 +-
 2 files changed, 369 insertions(+), 14 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 7ab20258e39b..da46872ecd67 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -15,6 +15,7 @@
 #include <linux/sched/signal.h>
 #include <linux/fs.h>
 #include <linux/file.h>
+#include <linux/fdtable.h>
 #include <linux/mm.h>
 #include <linux/mman.h>
 #include <linux/mmu_context.h>
@@ -25,6 +26,8 @@
 #include <linux/anon_inodes.h>
 #include <linux/sizes.h>
 #include <linux/nospec.h>
+#include <linux/kthread.h>
+#include <linux/sched/mm.h>
 
 #include <linux/uaccess.h>
 #include <linux/nospec.h>
@@ -62,6 +65,14 @@ struct io_mapped_ubuf {
 	unsigned int	nr_bvecs;
 };
 
+struct io_sq_offload {
+	struct task_struct	*thread;	/* if using a thread */
+	struct workqueue_struct	*wq;		/* wq offload */
+	struct mm_struct	*mm;
+	struct files_struct	*files;
+	wait_queue_head_t	wait;
+};
+
 struct io_ring_ctx {
 	struct percpu_ref	refs;
 
@@ -71,6 +82,7 @@ struct io_ring_ctx {
 	struct io_sq_ring	*sq_ring;
 	unsigned		sq_entries;
 	unsigned		sq_mask;
+	unsigned		sq_thread_cpu;
 	struct io_uring_sqe	*sq_sqes;
 
 	/* CQ ring */
@@ -81,6 +93,9 @@ struct io_ring_ctx {
 	/* if used, fixed mapped user buffers */
 	struct io_mapped_ubuf	*user_bufs;
 
+	/* sq ring submitter thread, if used */
+	struct io_sq_offload	sq_offload;
+
 	struct completion	ctx_done;
 
 	/* iopoll submission state */
@@ -115,6 +130,7 @@ struct io_kiocb {
 	unsigned long		ki_flags;
 #define KIOCB_F_IOPOLL_COMPLETED	0	/* polled IO has completed */
 #define KIOCB_F_IOPOLL_EAGAIN		1	/* submission got EAGAIN */
+#define KIOCB_F_FORCE_NONBLOCK		2	/* inline submission attempt */
 };
 
 #define IO_PLUG_THRESHOLD		2
@@ -125,6 +141,12 @@ struct sqe_submit {
 	unsigned index;
 };
 
+struct io_work {
+	struct work_struct work;
+	struct io_ring_ctx *ctx;
+	struct sqe_submit submit;
+};
+
 struct io_submit_state {
 	struct io_ring_ctx *ctx;
 
@@ -471,6 +493,20 @@ static void io_cqring_fill_event(struct io_ring_ctx *ctx, unsigned ki_index,
 	spin_unlock_irqrestore(&ctx->completion_lock, flags);
 }
 
+static void io_fill_cq_error(struct io_ring_ctx *ctx, struct sqe_submit *s,
+			     long error)
+{
+	io_cqring_fill_event(ctx, s->index, error, 0);
+
+	/*
+	 * for thread offload, app could already be sleeping in io_ring_enter()
+	 * before we get to flag the error. wake them up, if needed.
+	 */
+	if (ctx->flags & (IORING_SETUP_SQTHREAD | IORING_SETUP_SQWQ))
+		if (waitqueue_active(&ctx->wait))
+			wake_up(&ctx->wait);
+}
+
 static void io_complete_scqring_rw(struct kiocb *kiocb, long res, long res2)
 {
 	struct io_kiocb *iocb = container_of(kiocb, struct io_kiocb, rw);
@@ -543,7 +579,7 @@ static struct file *io_file_get(struct io_submit_state *state, int fd)
 }
 
 static int io_prep_rw(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe,
-		      struct io_submit_state *state)
+		      struct io_submit_state *state, bool force_nonblock)
 {
 	struct io_ring_ctx *ctx = kiocb->ki_ctx;
 	struct kiocb *req = &kiocb->rw;
@@ -567,6 +603,10 @@ static int io_prep_rw(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe,
 	ret = kiocb_set_rw_flags(req, sqe->rw_flags);
 	if (unlikely(ret))
 		goto out_fput;
+	if (force_nonblock) {
+		req->ki_flags |= IOCB_NOWAIT;
+		set_bit(KIOCB_F_FORCE_NONBLOCK, &kiocb->ki_flags);
+	}
 
 	if (ctx->flags & IORING_SETUP_IOPOLL) {
 		ret = -EOPNOTSUPP;
@@ -701,7 +741,7 @@ static int io_import_fixed(int rw, struct io_kiocb *kiocb,
 }
 
 static ssize_t io_read(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe,
-		       struct io_submit_state *state)
+		       struct io_submit_state *state, bool nonblock)
 {
 	struct iovec inline_vecs[UIO_FASTIOV], *iovec = inline_vecs;
 	struct kiocb *req = &kiocb->rw;
@@ -709,7 +749,7 @@ static ssize_t io_read(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe,
 	struct file *file;
 	ssize_t ret;
 
-	ret = io_prep_rw(kiocb, sqe, state);
+	ret = io_prep_rw(kiocb, sqe, state, nonblock);
 	if (ret)
 		return ret;
 	file = req->ki_filp;
@@ -734,8 +774,18 @@ static ssize_t io_read(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe,
 		goto out_fput;
 
 	ret = rw_verify_area(READ, file, &req->ki_pos, iov_iter_count(&iter));
-	if (!ret)
-		io_rw_done(req, call_read_iter(file, req, &iter));
+	if (!ret) {
+		ssize_t ret2;
+
+		/*
+		 * Catch -EAGAIN return for forced non-blocking submission
+		 */
+		ret2 = call_read_iter(file, req, &iter);
+		if (!nonblock || ret2 != -EAGAIN)
+			io_rw_done(req, ret2);
+		else
+			ret = -EAGAIN;
+	}
 	kfree(iovec);
 out_fput:
 	if (unlikely(ret))
@@ -752,7 +802,7 @@ static ssize_t io_write(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe,
 	struct file *file;
 	ssize_t ret;
 
-	ret = io_prep_rw(kiocb, sqe, state);
+	ret = io_prep_rw(kiocb, sqe, state, false);
 	if (ret)
 		return ret;
 	file = req->ki_filp;
@@ -837,7 +887,7 @@ static int io_fsync(struct io_kiocb *kiocb, const struct io_uring_sqe *sqe,
 }
 
 static int io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s,
-			 struct io_submit_state *state)
+			 struct io_submit_state *state, bool force_nonblock)
 {
 	const struct io_uring_sqe *sqe = s->sqe;
 	struct io_kiocb *req;
@@ -860,7 +910,7 @@ static int io_submit_sqe(struct io_ring_ctx *ctx, struct sqe_submit *s,
 	switch (sqe->opcode) {
 	case IORING_OP_READV:
 	case IORING_OP_READ_FIXED:
-		ret = io_read(req, sqe, state);
+		ret = io_read(req, sqe, state, force_nonblock);
 		break;
 	case IORING_OP_WRITEV:
 	case IORING_OP_WRITE_FIXED:
@@ -988,7 +1038,7 @@ static int io_ring_submit(struct io_ring_ctx *ctx, unsigned int to_submit)
 		if (!io_peek_sqring(ctx, &s))
 			break;
 
-		ret = io_submit_sqe(ctx, &s, statep);
+		ret = io_submit_sqe(ctx, &s, statep, false);
 		if (ret)
 			break;
 
@@ -1037,15 +1087,237 @@ static int io_cqring_wait(struct io_ring_ctx *ctx, int min_events)
 	return ring->r.head == ring->r.tail ? ret : 0;
 }
 
+static void io_sq_wq_submit_work(struct work_struct *work)
+{
+	struct io_work *iw = container_of(work, struct io_work, work);
+	struct io_ring_ctx *ctx = iw->ctx;
+	struct io_sq_offload *sqo = &ctx->sq_offload;
+	mm_segment_t old_fs = get_fs();
+	struct files_struct *old_files;
+	int ret;
+
+	old_files = current->files;
+	current->files = sqo->files;
+
+	if (sqo->mm) {
+		if (!mmget_not_zero(sqo->mm)) {
+			ret = -EFAULT;
+			goto err;
+		}
+		use_mm(sqo->mm);
+	}
+
+	set_fs(USER_DS);
+
+	ret = io_submit_sqe(ctx, &iw->submit, NULL, false);
+
+	set_fs(old_fs);
+	if (sqo->mm) {
+		unuse_mm(sqo->mm);
+		mmput(sqo->mm);
+	}
+
+err:
+	if (ret)
+		io_fill_cq_error(ctx, &iw->submit, ret);
+	current->files = old_files;
+	kfree(iw);
+}
+
+static int io_queue_async_work(struct io_ring_ctx *ctx, struct sqe_submit *s)
+{
+	struct io_work *work;
+
+	work = kmalloc(sizeof(*work), GFP_KERNEL);
+	if (!work)
+		return -ENOMEM;
+
+	memcpy(&work->submit, s, sizeof(*s));
+	INIT_WORK(&work->work, io_sq_wq_submit_work);
+	work->ctx = ctx;
+	queue_work(ctx->sq_offload.wq, &work->work);
+	return 0;
+}
+
+static int io_submit_sqes(struct io_ring_ctx *ctx, struct sqe_submit *sqes,
+			  unsigned int nr, struct mm_struct *cur_mm,
+			  bool mm_fault)
+{
+	struct io_submit_state state, *statep = NULL;
+	int ret, i, submitted = 0;
+
+	if (nr > IO_PLUG_THRESHOLD) {
+		io_submit_state_start(&state, ctx, nr);
+		statep = &state;
+	}
+
+	for (i = 0; i < nr; i++) {
+		if (unlikely(mm_fault))
+			ret = -EFAULT;
+		else
+			ret = io_submit_sqe(ctx, &sqes[i], statep, false);
+		if (!ret) {
+			submitted++;
+			continue;
+		}
+
+		io_fill_cq_error(ctx, &sqes[i], ret);
+	}
+
+	if (statep)
+		io_submit_state_end(&state);
+
+	return submitted;
+}
+
+/*
+ * sq thread only supports O_DIRECT or FIXEDBUFS IO
+ */
+static int io_sq_thread(void *data)
+{
+	struct sqe_submit sqes[IO_IOPOLL_BATCH];
+	struct io_ring_ctx *ctx = data;
+	struct io_sq_offload *sqo = &ctx->sq_offload;
+	struct mm_struct *cur_mm = NULL;
+	struct files_struct *old_files;
+	mm_segment_t old_fs;
+	DEFINE_WAIT(wait);
+
+	old_files = current->files;
+	current->files = sqo->files;
+
+	old_fs = get_fs();
+	set_fs(USER_DS);
+
+	while (!kthread_should_stop()) {
+		bool mm_fault = false;
+		int i;
+
+		if (!io_peek_sqring(ctx, &sqes[0])) {
+			/*
+			 * Drop cur_mm before scheduling, we can't hold it for
+			 * long periods (or over schedule()). Do this before
+			 * adding ourselves to the waitqueue, as the unuse/drop
+			 * may sleep.
+			 */
+			if (cur_mm) {
+				unuse_mm(cur_mm);
+				mmput(cur_mm);
+				cur_mm = NULL;
+			}
+
+			prepare_to_wait(&sqo->wait, &wait, TASK_INTERRUPTIBLE);
+			if (!io_peek_sqring(ctx, &sqes[0])) {
+				if (kthread_should_park())
+					kthread_parkme();
+				if (kthread_should_stop()) {
+					finish_wait(&sqo->wait, &wait);
+					break;
+				}
+				if (signal_pending(current))
+					flush_signals(current);
+				schedule();
+				finish_wait(&sqo->wait, &wait);
+				continue;
+			}
+			finish_wait(&sqo->wait, &wait);
+		}
+
+		/* If ->mm is set, we're not doing FIXEDBUFS */
+		if (sqo->mm && !cur_mm) {
+			mm_fault = !mmget_not_zero(sqo->mm);
+			if (!mm_fault) {
+				use_mm(sqo->mm);
+				cur_mm = sqo->mm;
+			}
+		}
+
+		i = 0;
+		do {
+			if (i == ARRAY_SIZE(sqes))
+				break;
+			i++;
+			io_inc_sqring(ctx);
+		} while (io_peek_sqring(ctx, &sqes[i]));
+
+		io_submit_sqes(ctx, sqes, i, cur_mm, mm_fault);
+	}
+	current->files = old_files;
+	set_fs(old_fs);
+	if (cur_mm) {
+		unuse_mm(cur_mm);
+		mmput(cur_mm);
+	}
+	return 0;
+}
+
+/*
+ * If this is a read, try a cached inline read first. If the IO is in the
+ * page cache, we can satisfy it without blocking and without having to
+ * punt to a threaded execution. This is much faster, particularly for
+ * lower queue depth IO, and it's always a lot more efficient.
+ */
+static bool io_sq_try_inline(struct io_ring_ctx *ctx, struct sqe_submit *s)
+{
+	int ret;
+
+	if (s->sqe->opcode != IORING_OP_READV &&
+	    s->sqe->opcode != IORING_OP_READ_FIXED)
+		return false;
+
+	ret = io_submit_sqe(ctx, s, NULL, true);
+
+	/*
+	 * If we get -EAGAIN, return false to submit out-of-line. Any other
+	 * result and we're done, caller will fill in CQ ring event.
+	 */
+	return ret != -EAGAIN;
+}
+
+static int io_sq_wq_submit(struct io_ring_ctx *ctx, unsigned int to_submit)
+{
+	struct sqe_submit s;
+	int ret, queued;
+
+	ret = queued = 0;
+	while (io_peek_sqring(ctx, &s)) {
+		ret = io_sq_try_inline(ctx, &s);
+		if (!ret) {
+			ret = io_queue_async_work(ctx, &s);
+			if (ret)
+				break;
+		}
+		io_inc_sqring(ctx);
+		queued++;
+		if (queued == to_submit)
+			break;
+	}
+
+	return queued ? queued : ret;
+}
+
 static int __io_uring_enter(struct io_ring_ctx *ctx, unsigned to_submit,
 			    unsigned min_complete, unsigned flags)
 {
 	int ret = 0;
 
 	if (to_submit) {
-		ret = io_ring_submit(ctx, to_submit);
-		if (ret < 0)
-			return ret;
+		/*
+		 * Three options here:
+		 * 1) We have an sq thread, just wake it up to do submissions
+		 * 2) We have an sq wq, queue a work item for each sqe
+		 * 3) Submit directly
+		 */
+		if (ctx->flags & IORING_SETUP_SQTHREAD) {
+			wake_up(&ctx->sq_offload.wait);
+			ret = to_submit;
+		} else if (ctx->flags & IORING_SETUP_SQWQ) {
+			ret = io_sq_wq_submit(ctx, to_submit);
+		} else {
+			ret = io_ring_submit(ctx, to_submit);
+			if (ret < 0)
+				return ret;
+		}
 	}
 	if (flags & IORING_ENTER_GETEVENTS) {
 		unsigned nr_events = 0;
@@ -1187,6 +1459,77 @@ static int io_sqe_buffer_map(struct io_ring_ctx *ctx,
 	return ret;
 }
 
+static int io_sq_thread(void *);
+
+static int io_sq_thread_start(struct io_ring_ctx *ctx)
+{
+	struct io_sq_offload *sqo = &ctx->sq_offload;
+	int ret;
+
+	memset(sqo, 0, sizeof(*sqo));
+	init_waitqueue_head(&sqo->wait);
+
+	if (!ctx->user_bufs)
+		sqo->mm = current->mm;
+
+	/*
+	 * This is safe since 'current' has the fd installed, and if that
+	 * gets closed on exit, then fops->release() is invoked which
+	 * waits for the sq thread and sq workqueue to flush and exit
+	 * before exiting.
+	 */
+	ret = -EBADF;
+	sqo->files = current->files;
+	if (!sqo->files)
+		goto err;
+
+	if (ctx->flags & IORING_SETUP_SQTHREAD) {
+		sqo->thread = kthread_create_on_cpu(io_sq_thread, ctx,
+							ctx->sq_thread_cpu,
+							"io_uring-sq");
+		if (IS_ERR(sqo->thread)) {
+			ret = PTR_ERR(sqo->thread);
+			sqo->thread = NULL;
+			goto err;
+		}
+		wake_up_process(sqo->thread);
+	} else if (ctx->flags & IORING_SETUP_SQWQ) {
+		int concurrency;
+
+		/* Do QD, or 2 * CPUS, whatever is smallest */
+		concurrency = min(ctx->sq_entries - 1, 2 * num_online_cpus());
+		sqo->wq = alloc_workqueue("io_ring-wq",
+						WQ_UNBOUND | WQ_FREEZABLE,
+						concurrency);
+		if (!sqo->wq) {
+			ret = -ENOMEM;
+			goto err;
+		}
+	}
+
+	return 0;
+err:
+	if (sqo->files)
+		sqo->files = NULL;
+	if (sqo->mm)
+		sqo->mm = NULL;
+	return ret;
+}
+
+static void io_sq_thread_stop(struct io_ring_ctx *ctx)
+{
+	struct io_sq_offload *sqo = &ctx->sq_offload;
+
+	if (sqo->thread) {
+		kthread_park(sqo->thread);
+		kthread_stop(sqo->thread);
+		sqo->thread = NULL;
+	} else if (sqo->wq) {
+		destroy_workqueue(sqo->wq);
+		sqo->wq = NULL;
+	}
+}
+
 static void io_free_scq_urings(struct io_ring_ctx *ctx)
 {
 	if (ctx->sq_ring) {
@@ -1205,6 +1548,7 @@ static void io_free_scq_urings(struct io_ring_ctx *ctx)
 
 static void io_ring_ctx_free(struct io_ring_ctx *ctx)
 {
+	io_sq_thread_stop(ctx);
 	io_iopoll_reap_events(ctx);
 	io_free_scq_urings(ctx);
 	io_sqe_buffer_unmap(ctx);
@@ -1394,6 +1738,13 @@ static int io_uring_create(unsigned entries, struct io_uring_params *p,
 		if (ret)
 			goto err;
 	}
+	if (p->flags & (IORING_SETUP_SQTHREAD | IORING_SETUP_SQWQ)) {
+		ctx->sq_thread_cpu = p->sq_thread_cpu;
+
+		ret = io_sq_thread_start(ctx);
+		if (ret)
+			goto err;
+	}
 
 	ret = anon_inode_getfd("[io_uring]", &io_scqring_fops, ctx,
 				O_RDWR | O_CLOEXEC);
@@ -1426,7 +1777,8 @@ SYSCALL_DEFINE3(io_uring_setup, u32, entries, struct iovec __user *, iovecs,
 			return -EINVAL;
 	}
 
-	if (p.flags & ~IORING_SETUP_IOPOLL)
+	if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQTHREAD |
+			IORING_SETUP_SQWQ))
 		return -EINVAL;
 
 	ret = io_uring_create(entries, &p, iovecs);
diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
index 80d1a8224b9c..79004940f7da 100644
--- a/include/uapi/linux/io_uring.h
+++ b/include/uapi/linux/io_uring.h
@@ -35,6 +35,8 @@ struct io_uring_sqe {
  * io_uring_setup() flags
  */
 #define IORING_SETUP_IOPOLL	(1 << 0)	/* io_context is polled */
+#define	IORING_SETUP_SQTHREAD	(1 << 1)	/* Use SQ thread */
+#define IORING_SETUP_SQWQ	(1 << 2)	/* Use SQ workqueue */
 
 #define IORING_OP_READV		1
 #define IORING_OP_WRITEV	2
@@ -95,7 +97,8 @@ struct io_uring_params {
 	__u32 sq_entries;
 	__u32 cq_entries;
 	__u32 flags;
-	__u16 resv[10];
+	__u16 sq_thread_cpu;
+	__u16 resv[9];
 	struct io_sqring_offsets sq_off;
 	struct io_cqring_offsets cq_off;
 };
-- 
2.17.1

[PATCH 14/15] io_uring: add submission polling

[Jens Axboe]

This enables an application to do IO, without ever entering the kernel.
By using the SQ ring to fill in new events and watching for completions
on the CQ ring, we can submit and reap IOs without doing a single system
call. The kernel side thread will poll for new submissions, and in case
of HIPRI/polled IO, it'll also poll for completions.

For O_DIRECT, we can do this with just SQTHREAD being enabled. For
buffered aio, we need the workqueue as well. If we can satisfy the
buffered inline from the SQTHREAD, we do that. If not, we punt to the
workqueue. This is just like buffered aio off the io_uring_enter(2)
system call.

Proof of concept. If the thread has been idle for 1 second, it will set
sq_ring->flags |= IORING_SQ_NEED_WAKEUP. The application will have to
call io_uring_enter() to start things back up again. If IO is kept busy,
that will never be needed. Basically an application that has this
feature enabled will guard it's io_uring_enter(2) call with:

barrier();
if (*sq_ring->flags & IORING_SQ_NEED_WAKEUP)
	io_uring_enter(fd, to_submit, 0, 0);

instead of calling it unconditionally.

Improvements:

1) Maybe have smarter backoff. Busy loop for X time, then go to
   monitor/mwait, finally the schedule we have now after an idle
   second. Might not be worth the complexity.

2) Probably want the application to pass in the appropriate grace
   period, not hard code it at 1 second.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/io_uring.c                 | 102 +++++++++++++++++++++++++++++++---
 include/uapi/linux/io_uring.h |   3 +
 2 files changed, 97 insertions(+), 8 deletions(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index da46872ecd67..6c62329b00ec 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -67,6 +67,7 @@ struct io_mapped_ubuf {
 
 struct io_sq_offload {
 	struct task_struct	*thread;	/* if using a thread */
+	bool			thread_poll;
 	struct workqueue_struct	*wq;		/* wq offload */
 	struct mm_struct	*mm;
 	struct files_struct	*files;
@@ -1145,17 +1146,35 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, struct sqe_submit *sqes,
 {
 	struct io_submit_state state, *statep = NULL;
 	int ret, i, submitted = 0;
+	bool nonblock;
 
 	if (nr > IO_PLUG_THRESHOLD) {
 		io_submit_state_start(&state, ctx, nr);
 		statep = &state;
 	}
 
+	/*
+	 * Having both a thread and a workqueue only makes sense for buffered
+	 * IO, where we can't submit in an async fashion. Use the NOWAIT
+	 * trick from the SQ thread, and punt to the workqueue if we can't
+	 * satisfy this iocb without blocking. This is only necessary
+	 * for buffered IO with sqthread polled submission.
+	 */
+	nonblock = (ctx->flags & IORING_SETUP_SQWQ) != 0;
+
 	for (i = 0; i < nr; i++) {
-		if (unlikely(mm_fault))
+		if (unlikely(mm_fault)) {
 			ret = -EFAULT;
-		else
-			ret = io_submit_sqe(ctx, &sqes[i], statep, false);
+		} else {
+			ret = io_submit_sqe(ctx, &sqes[i], statep, nonblock);
+			/* nogo, submit to workqueue */
+			if (nonblock && ret == -EAGAIN)
+				ret = io_queue_async_work(ctx, &sqes[i]);
+			if (!ret) {
+				submitted++;
+				continue;
+			}
+		}
 		if (!ret) {
 			submitted++;
 			continue;
@@ -1171,7 +1190,10 @@ static int io_submit_sqes(struct io_ring_ctx *ctx, struct sqe_submit *sqes,
 }
 
 /*
- * sq thread only supports O_DIRECT or FIXEDBUFS IO
+ * SQ thread is woken if the app asked for offloaded submission. This can
+ * be either O_DIRECT, in which case we do submissions directly, or it can
+ * be buffered IO, in which case we do them inline if we can do so without
+ * blocking. If we can't, then we punt to a workqueue.
  */
 static int io_sq_thread(void *data)
 {
@@ -1182,6 +1204,8 @@ static int io_sq_thread(void *data)
 	struct files_struct *old_files;
 	mm_segment_t old_fs;
 	DEFINE_WAIT(wait);
+	unsigned inflight;
+	unsigned long timeout;
 
 	old_files = current->files;
 	current->files = sqo->files;
@@ -1189,11 +1213,49 @@ static int io_sq_thread(void *data)
 	old_fs = get_fs();
 	set_fs(USER_DS);
 
+	timeout = inflight = 0;
 	while (!kthread_should_stop()) {
 		bool mm_fault = false;
 		int i;
 
+		if (sqo->thread_poll && inflight) {
+			unsigned int nr_events = 0;
+
+			/*
+			 * Normal IO, just pretend everything completed.
+			 * We don't have to poll completions for that.
+			 */
+			if (ctx->flags & IORING_SETUP_IOPOLL) {
+				/*
+				 * App should not use IORING_ENTER_GETEVENTS
+				 * with thread polling, but if it does, then
+				 * ensure we are mutually exclusive.
+				 */
+				if (mutex_trylock(&ctx->uring_lock)) {
+					io_iopoll_check(ctx, &nr_events, 0);
+					mutex_unlock(&ctx->uring_lock);
+				}
+			} else {
+				nr_events = inflight;
+			}
+
+			inflight -= nr_events;
+			if (!inflight)
+				timeout = jiffies + HZ;
+		}
+
 		if (!io_peek_sqring(ctx, &sqes[0])) {
+			/*
+			 * If we're polling, let us spin for a second without
+			 * work before going to sleep.
+			 */
+			if (sqo->thread_poll) {
+				if (inflight || !time_after(jiffies, timeout)) {
+					cpu_relax();
+					continue;
+				}
+			}
+
 			/*
 			 * Drop cur_mm before scheduling, we can't hold it for
 			 * long periods (or over schedule()). Do this before
@@ -1207,6 +1269,13 @@ static int io_sq_thread(void *data)
 			}
 
 			prepare_to_wait(&sqo->wait, &wait, TASK_INTERRUPTIBLE);
+
+			/* Tell userspace we may need a wakeup call */
+			if (sqo->thread_poll) {
+				ctx->sq_ring->flags |= IORING_SQ_NEED_WAKEUP;
+				smp_wmb();
+			}
+
 			if (!io_peek_sqring(ctx, &sqes[0])) {
 				if (kthread_should_park())
 					kthread_parkme();
@@ -1218,6 +1287,13 @@ static int io_sq_thread(void *data)
 					flush_signals(current);
 				schedule();
 				finish_wait(&sqo->wait, &wait);
+
+				if (sqo->thread_poll) {
+					struct io_sq_ring *ring;
+
+					ring = ctx->sq_ring;
+					ring->flags &= ~IORING_SQ_NEED_WAKEUP;
+				}
 				continue;
 			}
 			finish_wait(&sqo->wait, &wait);
@@ -1240,7 +1316,7 @@ static int io_sq_thread(void *data)
 			io_inc_sqring(ctx);
 		} while (io_peek_sqring(ctx, &sqes[i]));
 
-		io_submit_sqes(ctx, sqes, i, cur_mm, mm_fault);
+		inflight += io_submit_sqes(ctx, sqes, i, cur_mm, mm_fault);
 	}
 	current->files = old_files;
 	set_fs(old_fs);
@@ -1483,6 +1559,9 @@ static int io_sq_thread_start(struct io_ring_ctx *ctx)
 	if (!sqo->files)
 		goto err;
 
+	if (ctx->flags & IORING_SETUP_SQPOLL)
+		sqo->thread_poll = true;
+
 	if (ctx->flags & IORING_SETUP_SQTHREAD) {
 		sqo->thread = kthread_create_on_cpu(io_sq_thread, ctx,
 							ctx->sq_thread_cpu,
@@ -1493,7 +1572,8 @@ static int io_sq_thread_start(struct io_ring_ctx *ctx)
 			goto err;
 		}
 		wake_up_process(sqo->thread);
-	} else if (ctx->flags & IORING_SETUP_SQWQ) {
+	}
+	if (ctx->flags & IORING_SETUP_SQWQ) {
 		int concurrency;
 
 		/* Do QD, or 2 * CPUS, whatever is smallest */
@@ -1524,7 +1604,8 @@ static void io_sq_thread_stop(struct io_ring_ctx *ctx)
 		kthread_park(sqo->thread);
 		kthread_stop(sqo->thread);
 		sqo->thread = NULL;
-	} else if (sqo->wq) {
+	}
+	if (sqo->wq) {
 		destroy_workqueue(sqo->wq);
 		sqo->wq = NULL;
 	}
@@ -1738,6 +1819,11 @@ static int io_uring_create(unsigned entries, struct io_uring_params *p,
 		if (ret)
 			goto err;
 	}
+	if ((p->flags & IORING_SETUP_SQPOLL) &&
+	   !(p->flags & IORING_SETUP_SQTHREAD)) {
+		ret = -EINVAL;
+		goto err;
+	}
 	if (p->flags & (IORING_SETUP_SQTHREAD | IORING_SETUP_SQWQ)) {
 		ctx->sq_thread_cpu = p->sq_thread_cpu;
 
@@ -1778,7 +1864,7 @@ SYSCALL_DEFINE3(io_uring_setup, u32, entries, struct iovec __user *, iovecs,
 	}
 
 	if (p.flags & ~(IORING_SETUP_IOPOLL | IORING_SETUP_SQTHREAD |
-			IORING_SETUP_SQWQ))
+			IORING_SETUP_SQWQ | IORING_SETUP_SQPOLL))
 		return -EINVAL;
 
 	ret = io_uring_create(entries, &p, iovecs);
diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
index 79004940f7da..9321eb97479d 100644
--- a/include/uapi/linux/io_uring.h
+++ b/include/uapi/linux/io_uring.h
@@ -37,6 +37,7 @@ struct io_uring_sqe {
 #define IORING_SETUP_IOPOLL	(1 << 0)	/* io_context is polled */
 #define	IORING_SETUP_SQTHREAD	(1 << 1)	/* Use SQ thread */
 #define IORING_SETUP_SQWQ	(1 << 2)	/* Use SQ workqueue */
+#define IORING_SETUP_SQPOLL	(1 << 3)	/* SQ thread polls */
 
 #define IORING_OP_READV		1
 #define IORING_OP_WRITEV	2
@@ -75,6 +76,8 @@ struct io_sqring_offsets {
 	__u32 resv[3];
 };
 
+#define IORING_SQ_NEED_WAKEUP	(1 << 0) /* needs io_uring_enter wakeup */
+
 struct io_cqring_offsets {
 	__u32 head;
 	__u32 tail;
-- 
2.17.1

[PATCH 15/15] io_uring: add io_uring_event cache hit information

[Jens Axboe]

Add hint on whether a read was served out of the page cache, or if it
hit media. This is useful for buffered async IO, O_DIRECT reads would
never have this set (for obvious reasons).

If the read hit page cache, cqe->flags will have IOCQE_FLAG_CACHEHIT
set.

Signed-off-by: Jens Axboe <axboe@kernel.dk>
---
 fs/io_uring.c                 | 7 ++++++-
 include/uapi/linux/io_uring.h | 5 +++++
 2 files changed, 11 insertions(+), 1 deletion(-)

diff --git a/fs/io_uring.c b/fs/io_uring.c
index 6c62329b00ec..f2a603c447ba 100644
--- a/fs/io_uring.c
+++ b/fs/io_uring.c
@@ -511,11 +511,16 @@ static void io_fill_cq_error(struct io_ring_ctx *ctx, struct sqe_submit *s,
 static void io_complete_scqring_rw(struct kiocb *kiocb, long res, long res2)
 {
 	struct io_kiocb *iocb = container_of(kiocb, struct io_kiocb, rw);
+	unsigned ev_flags = 0;
 
 	kiocb_end_write(kiocb);
 
 	fput(kiocb->ki_filp);
-	io_cqring_fill_event(iocb->ki_ctx, iocb->ki_index, res, 0);
+
+	if (res > 0 && test_bit(KIOCB_F_FORCE_NONBLOCK, &iocb->ki_flags))
+		ev_flags = IOCQE_FLAG_CACHEHIT;
+
+	io_cqring_fill_event(iocb->ki_ctx, iocb->ki_index, res, ev_flags);
 	io_free_kiocb(iocb);
 }
 
diff --git a/include/uapi/linux/io_uring.h b/include/uapi/linux/io_uring.h
index 9321eb97479d..20e4c22e040d 100644
--- a/include/uapi/linux/io_uring.h
+++ b/include/uapi/linux/io_uring.h
@@ -55,6 +55,11 @@ struct io_uring_cqe {
 	__u32	flags;
 };
 
+/*
+ * io_uring_event->flags
+ */
+#define IOCQE_FLAG_CACHEHIT	(1 << 0)	/* IO did not hit media */
+
 /*
  * Magic offsets for the application to mmap the data it needs
  */
-- 
2.17.1

Re: [PATCH 15/15] io_uring: add io_uring_event cache hit information

[Jeff Moyer]

Jens Axboe <axboe@kernel.dk> writes:

> Add hint on whether a read was served out of the page cache, or if it
> hit media. This is useful for buffered async IO, O_DIRECT reads would
> never have this set (for obvious reasons).
>
> If the read hit page cache, cqe->flags will have IOCQE_FLAG_CACHEHIT
> set.

We may want to hold off on this one until the whole mincore/RWF_NOWAIT
debate is sorted. [1]

Cheers,
Jeff

[1] https://lore.kernel.org/lkml/20190109022430.GE27534@dastard/

Re: [PATCH 15/15] io_uring: add io_uring_event cache hit information

[Jens Axboe]

On 1/10/19 4:12 PM, Jeff Moyer wrote:
> Jens Axboe <axboe@kernel.dk> writes:
> 
>> Add hint on whether a read was served out of the page cache, or if it
>> hit media. This is useful for buffered async IO, O_DIRECT reads would
>> never have this set (for obvious reasons).
>>
>> If the read hit page cache, cqe->flags will have IOCQE_FLAG_CACHEHIT
>> set.
> 
> We may want to hold off on this one until the whole mincore/RWF_NOWAIT
> debate is sorted. [1]

Definitely, it's why it's separate and at the end of the series. But
in reality, this doesn't leak anything that timing doesn't already
tell you. So it's kind of a moot point.

-- 
Jens Axboe

Re: [PATCHSET v2] io_uring IO interface

[Roman Penyaev]

Hi Jens,

That is interesting.  Recently I sent an rfc related to epoll uring:

https://lore.kernel.org/lkml/20190109164025.24554-1-rpenyaev@suse.de

which can be mapped to userspace and all ready events can be consumed
from it directly.  I am wondering, is it possible to make some common
API for all kind of ready events / urings, or it doesn't make any
sense?

--
Roman


On 2019-01-10 03:43, Jens Axboe wrote:
> Here's v2 of the io_uring interface. See the v1 posting for some more 
> info:
> 
> https://lore.kernel.org/linux-block/20190108165645.19311-1-axboe@kernel.dk/
> 
> The data structures changed, to improve the symmetry of the submission
> and completion side. The io_uring_iocb is now io_uring_sqe, but it
> otherwise remains the same as before. Ditto on the completion side,
> where io_uring_event is now io_uring_cqe.
> 
> I've updated the fio io_uring test app, and the io_uring engine. The
> liburing git repo has also been adapted to the various changes since 
> the
> v1 posting. As a reminder, the liburing git repo contains some helpers
> for doing IO without having to muck with the ring directly, setting up
> an io_uring context, etc. Clone that here:
> 
> git://git.kernel.dk/liburing
> 
> In terms of usage, there's also a small test app here:
> 
> http://git.kernel.dk/cgit/fio/plain/t/io_uring.c
> 
> and the liburing repo has a few test apps in test/ as well.
> 
> Patches are aginst 5.0-rc1, but can also be found here:
> 
> git://git.kernel.dk/linux-block io_uring
> 
> Changes since v1:
> 
> - Fail IORING_OP_{READ,WRITE}_FIXED if not configured
> - Fix ctx drop ref issue on failure to close ring_fd when sq thread/wq
>   are in use
> - Move to separate Kconfig entry (CONFIG_IO_URING)
> - Add SPDX headers
> - Drop gcc ism of zero sized arrays
> - Rename io_uring_iocb -> io_uring_sqe
> - Rename io_uring_event -> io_uring_cqe
> - Drop needless io_event_ring and io_iocb_ring structures
> - Drop ctx->max_reqs, use ->sq_entries
> - Drop unused ->ring_lock
> - Drop io_ring_ctx slab cache
> - Fix state batched kiocb alloc failure to put ctx
> - Fix missing write ordering barrier when filling in the cqe
> - Drop io_req_init()
> - Various renames
> - Fix a few lines that were too long
> - Address other minor review comments
> - Fix IORING_SETUP_SQPOLL being set without IORING_SETUP_SQTHREAD
> - Drop IORING_SETUP_FIXEDBUFS, iovecs being non-NULL is enough
> - Fix error handling free of ctx in setup path
> - Change standard read/write commands to be iov based READV/WRITEV
> - Pass in struct sqe_submit instead of separate sqe/index everywhere
> - Fix reap of polled events on fops->release()
> - Lock uring for sq thread polling
> - Don't grab ->completion_lock for polled IO cqe filling
> - Fix ev_flags vs flags typo
> - Consolidate parts of the io_ring_ctx alignment
> 
>  Documentation/filesystems/vfs.txt      |    3 +
>  arch/x86/entry/syscalls/syscall_64.tbl |    2 +
>  block/bio.c                            |   59 +-
>  fs/Makefile                            |    1 +
>  fs/block_dev.c                         |   19 +-
>  fs/file.c                              |   15 +-
>  fs/file_table.c                        |    9 +-
>  fs/gfs2/file.c                         |    2 +
>  fs/io_uring.c                          | 1890 ++++++++++++++++++++++++
>  fs/iomap.c                             |   48 +-
>  fs/xfs/xfs_file.c                      |    1 +
>  include/linux/bio.h                    |   14 +
>  include/linux/blk_types.h              |    1 +
>  include/linux/file.h                   |    2 +
>  include/linux/fs.h                     |    6 +-
>  include/linux/iomap.h                  |    1 +
>  include/linux/syscalls.h               |    5 +
>  include/uapi/linux/io_uring.h          |  114 ++
>  init/Kconfig                           |    8 +
>  kernel/sys_ni.c                        |    2 +
>  20 files changed, 2163 insertions(+), 39 deletions(-)

Re: [PATCHSET v2] io_uring IO interface

[Ilya Dryomov]

On Fri, Jan 11, 2019 at 10:51 AM Roman Penyaev <rpenyaev@suse.de> wrote:
>
> Hi Jens,
>
> That is interesting.  Recently I sent an rfc related to epoll uring:
>
> https://lore.kernel.org/lkml/20190109164025.24554-1-rpenyaev@suse.de
>
> which can be mapped to userspace and all ready events can be consumed
> from it directly.  I am wondering, is it possible to make some common
> API for all kind of ready events / urings, or it doesn't make any
> sense?

I think you can use the new IOCB_CMD_POLL from Christoph and avoid
epoll_wait() in favor of aio/io_uring interface, at least in new high
performance applications.  Reaping events entirely in userspace (i.e.
performing io_getevents() without entering the kernel) has been
possible for a long time even with the existing aio interface.

Thanks,

                Ilya

Re: [PATCHSET v2] io_uring IO interface

[Christoph Hellwig]

On Fri, Jan 11, 2019 at 05:11:57PM +0100, Ilya Dryomov wrote:
> I think you can use the new IOCB_CMD_POLL from Christoph and avoid
> epoll_wait() in favor of aio/io_uring interface, at least in new high
> performance applications.  Reaping events entirely in userspace (i.e.
> performing io_getevents() without entering the kernel) has been
> possible for a long time even with the existing aio interface.

For io_uring we can reuse the IOCB_CMD_POLL concept, but we'd have to
add a new cancel command, as the uring right now doesn't support
cancelation.  But I'd rather make that command a new opcode instead
of a separate syscall, which would lead to a nicer design.

A prototype for this should be fairly easy, I'd just want someone
to actually use it for real life testing, like ScyllaDB does for
IOCB_CMD_POLL.

Re: [PATCHSET v2] io_uring IO interface

[Roman Penyaev]

On 2019-01-11 17:11, Ilya Dryomov wrote:
> On Fri, Jan 11, 2019 at 10:51 AM Roman Penyaev <rpenyaev@suse.de> 
> wrote:
>> 
>> Hi Jens,
>> 
>> That is interesting.  Recently I sent an rfc related to epoll uring:
>> 
>> https://lore.kernel.org/lkml/20190109164025.24554-1-rpenyaev@suse.de
>> 
>> which can be mapped to userspace and all ready events can be consumed
>> from it directly.  I am wondering, is it possible to make some common
>> API for all kind of ready events / urings, or it doesn't make any
>> sense?
> 
> I think you can use the new IOCB_CMD_POLL from Christoph and avoid
> epoll_wait() in favor of aio/io_uring interface, at least in new high
> performance applications.

Yeah, I saw this extension for aio from Christoph.  I was motivated to
extend epoll with uring to avoid constant recharging of file 
descriptors,
i.e. once you inserted descriptor to epoll you just consume events from
uring (of course in that particular case only edge triggered events are
supported).

Also recently for epoll I fixed contention on event callback making hot
path completely lockless, i.e. with uring epoll can become a nice thingy
in terms of performance.

But can any descriptor (on vfs layer) be extended to have a uring? To 
have
some common API? Then if event source (say socket) has a uring (do not
know how, just thoughts) and event destination (aio, epoll) has a uring,
then reading on userside can be a matter of traversing urings.

> Reaping events entirely in userspace (i.e.
> performing io_getevents() without entering the kernel) has been
> possible for a long time even with the existing aio interface.

True.

--
Roman

Re: [PATCHSET v2] io_uring IO interface

[Jens Axboe]

On 1/11/19 2:46 AM, Roman Penyaev wrote:
> Hi Jens,
> 
> That is interesting.  Recently I sent an rfc related to epoll uring:
> 
> https://lore.kernel.org/lkml/20190109164025.24554-1-rpenyaev@suse.de
> 
> which can be mapped to userspace and all ready events can be consumed
> from it directly.  I am wondering, is it possible to make some common
> API for all kind of ready events / urings, or it doesn't make any
> sense?

Not sure that's easily possible, even out of the two rings in io_uring,
the sq and cq rings aren't the same. The latter is sequentially
written, as completions come in. The former ring is actually indexes
into the array, so you can submit things out of order when needed.

-- 
Jens Axboe