On 2020/03/25 7:46, Dave Chinner wrote:

On Wed, Mar 25, 2020 at 12:24:53AM +0900, Johannes Thumshirn wrote:

quoted

Use REQ_OP_ZONE_APPEND for direct I/O write BIOs, instead of REQ_OP_WRITE
if the file-system requests it. The file system can make this request
by setting the new flag IOCB_ZONE_APPEND for a direct I/O kiocb before
calling iompa_dio_rw(). Using this information, this function propagates
the zone append flag using IOMAP_ZONE_APPEND to the file system
iomap_begin() method. The BIOs submitted for the zone append DIO will be
set to use the REQ_OP_ZONE_APPEND operation.

Since zone append operations cannot be split, the iomap_apply() and
iomap_dio_rw() internal loops are executed only once, which may result
in short writes.

Signed-off-by: Johannes Thumshirn <redacted>
---
 fs/iomap/direct-io.c  | 80 ++++++++++++++++++++++++++++++++++++-------
 include/linux/fs.h    |  1 +
 include/linux/iomap.h | 22 ++++++------
 3 files changed, 79 insertions(+), 24 deletions(-)

diff --git a/fs/iomap/direct-io.c b/fs/iomap/direct-io.c
index 23837926c0c5..b3e2aadce72f 100644
--- a/fs/iomap/direct-io.c
+++ b/fs/iomap/direct-io.c

@@ -17,6 +17,7 @@
  * Private flags for iomap_dio, must not overlap with the public ones in
  * iomap.h:
  */
+#define IOMAP_DIO_ZONE_APPEND	(1 << 27)
 #define IOMAP_DIO_WRITE_FUA	(1 << 28)
 #define IOMAP_DIO_NEED_SYNC	(1 << 29)
 #define IOMAP_DIO_WRITE		(1 << 30)

@@ -39,6 +40,7 @@ struct iomap_dio {
 			struct task_struct	*waiter;
 			struct request_queue	*last_queue;
 			blk_qc_t		cookie;
+			sector_t		sector;
 		} submit;
 
 		/* used for aio completion: */

@@ -151,6 +153,9 @@ static void iomap_dio_bio_end_io(struct bio *bio)
 	if (bio->bi_status)
 		iomap_dio_set_error(dio, blk_status_to_errno(bio->bi_status));
 
+	if (dio->flags & IOMAP_DIO_ZONE_APPEND)
+		dio->submit.sector = bio->bi_iter.bi_sector;
+
 	if (atomic_dec_and_test(&dio->ref)) {
 		if (dio->wait_for_completion) {
 			struct task_struct *waiter = dio->submit.waiter;

@@ -194,6 +199,21 @@ iomap_dio_zero(struct iomap_dio *dio, struct iomap *iomap, loff_t pos,
 	iomap_dio_submit_bio(dio, iomap, bio);
 }
 
+static sector_t
+iomap_dio_bio_sector(struct iomap_dio *dio, struct iomap *iomap, loff_t pos)
+{
+	sector_t sector = iomap_sector(iomap, pos);
+
+	/*
+	 * For zone append writes, the BIO needs to point at the start of the
+	 * zone to append to.
+	 */
+	if (dio->flags & IOMAP_DIO_ZONE_APPEND)
+		sector = ALIGN_DOWN(sector, bdev_zone_sectors(iomap->bdev));
+
+	return sector;
+}

This seems to me like it should be done by the ->iomap_begin
implementation when mapping the IO. I don't see why this needs to be
specially handled by the iomap dio code.

Fair point. However, iomap_sector() does not simply return iomap->addr >> 9. So
that means that in iomap_begin, the mapping address and offset returned needs to
account for the iomap_sector() calculation, i.e.

(iomap->addr + pos - iomap->offset) >> SECTOR_SHIFT

which does not result in a very obvious values for iomap address and offset.
Well I guess for zone append we simply need to set
iomap->offset = pos;
and
iomap->addr = zone start sector;
and that would then work. This however look fragile to me.

quoted

+
 static loff_t
 iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
 		struct iomap_dio *dio, struct iomap *iomap)

@@ -204,6 +224,7 @@ iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
 	struct bio *bio;
 	bool need_zeroout = false;
 	bool use_fua = false;
+	bool zone_append = false;
 	int nr_pages, ret = 0;
 	size_t copied = 0;
 	size_t orig_count;

@@ -235,6 +256,9 @@ iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
 			use_fua = true;
 	}
 
+	if (dio->flags & IOMAP_DIO_ZONE_APPEND)
+		zone_append = true;
+
 	/*
 	 * Save the original count and trim the iter to just the extent we
 	 * are operating on right now.  The iter will be re-expanded once

@@ -266,12 +290,28 @@ iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
 
 		bio = bio_alloc(GFP_KERNEL, nr_pages);
 		bio_set_dev(bio, iomap->bdev);
-		bio->bi_iter.bi_sector = iomap_sector(iomap, pos);
+		bio->bi_iter.bi_sector = iomap_dio_bio_sector(dio, iomap, pos);
 		bio->bi_write_hint = dio->iocb->ki_hint;
 		bio->bi_ioprio = dio->iocb->ki_ioprio;
 		bio->bi_private = dio;
 		bio->bi_end_io = iomap_dio_bio_end_io;
 
+		if (dio->flags & IOMAP_DIO_WRITE) {
+			bio->bi_opf = REQ_SYNC | REQ_IDLE;
+			if (zone_append)
+				bio->bi_opf |= REQ_OP_ZONE_APPEND;
+			else
+				bio->bi_opf |= REQ_OP_WRITE;
+			if (use_fua)
+				bio->bi_opf |= REQ_FUA;
+			else
+				dio->flags &= ~IOMAP_DIO_WRITE_FUA;
+		} else {
+			bio->bi_opf = REQ_OP_READ;
+			if (dio->flags & IOMAP_DIO_DIRTY)
+				bio_set_pages_dirty(bio);
+		}

Why move all this code? If it's needed, please split it into a
separate patchi to separate it from the new functionality...

The BIO add page in bio_iov_iter_get_pages() needs to know that the BIO is a
zone append to stop adding pages before the BIO grows too large and result in a
split when it is submitted. So bio->bi_opf needs to be set before calling
bio_iov_iter_get_pages(). So this change is related to the new functionality.
But I guess it is OK to do it regardless of the BIO operation.

quoted

+
 		ret = bio_iov_iter_get_pages(bio, dio->submit.iter);
 		if (unlikely(ret)) {
 			/*

@@ -284,19 +324,10 @@ iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
 			goto zero_tail;
 		}
 
-		n = bio->bi_iter.bi_size;
-		if (dio->flags & IOMAP_DIO_WRITE) {
-			bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
-			if (use_fua)
-				bio->bi_opf |= REQ_FUA;
-			else
-				dio->flags &= ~IOMAP_DIO_WRITE_FUA;
+		if (dio->flags & IOMAP_DIO_WRITE)
 			task_io_account_write(n);
-		} else {
-			bio->bi_opf = REQ_OP_READ;
-			if (dio->flags & IOMAP_DIO_DIRTY)
-				bio_set_pages_dirty(bio);
-		}
+
+		n = bio->bi_iter.bi_size;
 
 		dio->size += n;
 		pos += n;

@@ -304,6 +335,15 @@ iomap_dio_bio_actor(struct inode *inode, loff_t pos, loff_t length,
 
 		nr_pages = iov_iter_npages(dio->submit.iter, BIO_MAX_PAGES);
 		iomap_dio_submit_bio(dio, iomap, bio);
+
+		/*
+		 * Issuing multiple BIOs for a large zone append write can
+		 * result in reordering of the write fragments and to data
+		 * corruption. So always stop after the first BIO is issued.
+		 */
+		if (zone_append)
+			break;

I don't think this sort of functionality should be tied to "zone
append". If there is a need for "issue a single (short) bio only" it
should be a flag to iomap_dio_rw() set by the filesystem, which can
then handle the short read/write that is returned.

Yes, that would be cleaner.

quoted

+		/*
+		 * Zone append writes cannot be split and be shorted. Break
+		 * here to let the user know instead of sending more IOs which
+		 * could get reordered and corrupt the written data.
+		 */
+		if (flags & IOMAP_ZONE_APPEND)
+			break;

ditto.

Cheers,

Dave.

-- 
Damien Le Moal
Western Digital Research