On Fri, Jun 21, 2024 at 01:02:34PM +0100, John Garry wrote:

On 21/06/2024 07:13, Hannes Reinecke wrote:

quoted

On 6/20/24 14:53, John Garry wrote:

quoted

Support atomic writes by submitting a single BIO with the REQ_ATOMIC set.

It must be ensured that the atomic write adheres to its rules, like
naturally aligned offset, so call blkdev_dio_invalid() ->
blkdev_atomic_write_valid() [with renaming blkdev_dio_unaligned() to
blkdev_dio_invalid()] for this purpose. The BIO submission path currently
checks for atomic writes which are too large, so no need to check here.

In blkdev_direct_IO(), if the nr_pages exceeds BIO_MAX_VECS, then we
cannot
produce a single BIO, so error in this case.

Finally set FMODE_CAN_ATOMIC_WRITE when the bdev can support atomic
writes
and the associated file flag is for O_DIRECT.

Reviewed-by: Martin K. Petersen <martin.petersen@oracle.com>
Signed-off-by: John Garry <john.g.garry@oracle.com>
---
  block/fops.c | 20 +++++++++++++++++---
  1 file changed, 17 insertions(+), 3 deletions(-)

diff --git a/block/fops.c b/block/fops.c
index 376265935714..be36c9fbd500 100644
--- a/block/fops.c
+++ b/block/fops.c

@@ -34,9 +34,12 @@ static blk_opf_t dio_bio_write_op(struct kiocb *iocb)

      return opf;
  }
-static bool blkdev_dio_unaligned(struct block_device *bdev, loff_t pos,
-                  struct iov_iter *iter)
+static bool blkdev_dio_invalid(struct block_device *bdev, loff_t pos,
+                struct iov_iter *iter, bool is_atomic)
  {
+    if (is_atomic && !generic_atomic_write_valid(iter, pos))
+        return true;
+
      return pos & (bdev_logical_block_size(bdev) - 1) ||
          !bdev_iter_is_aligned(bdev, iter);
  }

@@ -72,6 +75,8 @@ static ssize_t __blkdev_direct_IO_simple(struct

kiocb *iocb,
      bio.bi_iter.bi_sector = pos >> SECTOR_SHIFT;
      bio.bi_write_hint = file_inode(iocb->ki_filp)->i_write_hint;
      bio.bi_ioprio = iocb->ki_ioprio;
+    if (iocb->ki_flags & IOCB_ATOMIC)
+        bio.bi_opf |= REQ_ATOMIC;
      ret = bio_iov_iter_get_pages(&bio, iter);
      if (unlikely(ret))

@@ -343,6 +348,9 @@ static ssize_t __blkdev_direct_IO_async(struct

kiocb *iocb,
          task_io_account_write(bio->bi_iter.bi_size);
      }
+    if (iocb->ki_flags & IOCB_ATOMIC)
+        bio->bi_opf |= REQ_ATOMIC;
+
      if (iocb->ki_flags & IOCB_NOWAIT)
          bio->bi_opf |= REQ_NOWAIT;

@@ -359,12 +367,13 @@ static ssize_t __blkdev_direct_IO_async(struct

kiocb *iocb,
  static ssize_t blkdev_direct_IO(struct kiocb *iocb, struct
iov_iter *iter)
  {
      struct block_device *bdev = I_BDEV(iocb->ki_filp->f_mapping->host);
+    bool is_atomic = iocb->ki_flags & IOCB_ATOMIC;
      unsigned int nr_pages;
      if (!iov_iter_count(iter))
          return 0;
-    if (blkdev_dio_unaligned(bdev, iocb->ki_pos, iter))
+    if (blkdev_dio_invalid(bdev, iocb->ki_pos, iter, is_atomic))

Why not passing in iocb->ki_flags here?
Or, indeed, the entire iocb?

We could (pass the iocb), but we only need to look up one thing - ki_pos. We
already have is_atomic local. I am just trying to make things as efficient
as possible. If you really think it's better (to pass iocb), then it can be
changed.

I certainly do. ;)

https://lore.kernel.org/linux-xfs/20240620212401.GA3058325@frogsfrogsfrogs/ (local)

--D

Thanks,
John