Merge tag 'vfs-6.7.super' of gitolite.kernel.org:pub/scm/linux/kernel/git/vfs/vfs

Pull vfs superblock updates from Christian Brauner:
 "This contains the work to make block device opening functions return a
  struct bdev_handle instead of just a struct block_device. The same
  struct bdev_handle is then also passed to block device closing
  functions.

  This allows us to propagate context from opening to closing a block
  device without having to modify all users everytime.

  Sidenote, in the future we might even want to try and have block
  device opening functions return a struct file directly but that's a
  series on top of this.

  These are further preparatory changes to be able to count writable
  opens and blocking writes to mounted block devices. That's a separate
  piece of work for next cycle and for that we absolutely need the
  changes to btrfs that have been quietly dropped somehow.

  Originally the series contained a patch that removed the old
  blkdev_*() helpers. But since this would've caused needles churn in
  -next for bcachefs we ended up delaying it.

  The second piece of work addresses one of the major annoyances about
  the work last cycle, namely that we required dropping s_umount
  whenever we used the superblock and fs_holder_ops for a block device.

  The reason for that requirement had been that in some codepaths
  s_umount could've been taken under disk->open_mutex (that's always
  been the case, at least theoretically). For example, on surprise block
  device removal or media change. And opening and closing block devices
  required grabbing disk->open_mutex as well.

  So we did the work and went through the block layer and fixed all
  those places so that s_umount is never taken under disk->open_mutex.
  This means no more brittle games where we yield and reacquire s_umount
  during block device opening and closing and no more requirements where
  block devices need to be closed. Filesystems don't need to care about
  this.

  There's a bunch of other follow-up work such as moving block device
  freezing and thawing to holder operations which makes it work for all
  block devices and not just the main block device just as we did for
  surprise removal. But that is for next cycle.

  Tested with fstests for all major fses, blktests, LTP"

* tag 'vfs-6.7.super' of gitolite.kernel.org:pub/scm/linux/kernel/git/vfs/vfs: (37 commits)
  porting: update locking requirements
  fs: assert that open_mutex isn't held over holder ops
  block: assert that we're not holding open_mutex over blk_report_disk_dead
  block: move bdev_mark_dead out of disk_check_media_change
  block: WARN_ON_ONCE() when we remove active partitions
  block: simplify bdev_del_partition()
  fs: Avoid grabbing sb->s_umount under bdev->bd_holder_lock
  jfs: fix log->bdev_handle null ptr deref in lbmStartIO
  bcache: Fixup error handling in register_cache()
  xfs: Convert to bdev_open_by_path()
  reiserfs: Convert to bdev_open_by_dev/path()
  ocfs2: Convert to use bdev_open_by_dev()
  nfs/blocklayout: Convert to use bdev_open_by_dev/path()
  jfs: Convert to bdev_open_by_dev()
  f2fs: Convert to bdev_open_by_dev/path()
  ext4: Convert to bdev_open_by_dev()
  erofs: Convert to use bdev_open_by_path()
  btrfs: Convert to bdev_open_by_path()
  fs: Convert to bdev_open_by_dev()
  mm/swap: Convert to use bdev_open_by_dev()
  ...

1 files changed, 28 insertions, 16 deletions

diff --git a/block/fops.c b/block/fops.c
index 73e42742543f..0abaac705daf 100644
--- a/block/fops.c
+++ b/block/fops.c
@@ -542,15 +542,31 @@ static int blkdev_fsync(struct file *filp, loff_t start, loff_t end,
 	return error;
 }
 
+/**
+ * file_to_blk_mode - get block open flags from file flags
+ * @file: file whose open flags should be converted
+ *
+ * Look at file open flags and generate corresponding block open flags from
+ * them. The function works both for file just being open (e.g. during ->open
+ * callback) and for file that is already open. This is actually non-trivial
+ * (see comment in the function).
+ */
 blk_mode_t file_to_blk_mode(struct file *file)
 {
 	blk_mode_t mode = 0;
+	struct bdev_handle *handle = file->private_data;
 
 	if (file->f_mode & FMODE_READ)
 		mode |= BLK_OPEN_READ;
 	if (file->f_mode & FMODE_WRITE)
 		mode |= BLK_OPEN_WRITE;
-	if (file->private_data)
+	/*
+	 * do_dentry_open() clears O_EXCL from f_flags, use handle->mode to
+	 * determine whether the open was exclusive for already open files.
+	 */
+	if (handle)
+		mode |= handle->mode & BLK_OPEN_EXCL;
+	else if (file->f_flags & O_EXCL)
 		mode |= BLK_OPEN_EXCL;
 	if (file->f_flags & O_NDELAY)
 		mode |= BLK_OPEN_NDELAY;
@@ -568,7 +584,8 @@ blk_mode_t file_to_blk_mode(struct file *file)
 
 static int blkdev_open(struct inode *inode, struct file *filp)
 {
-	struct block_device *bdev;
+	struct bdev_handle *handle;
+	blk_mode_t mode;
 
 	/*
 	 * Preserve backwards compatibility and allow large file access
@@ -579,29 +596,24 @@ static int blkdev_open(struct inode *inode, struct file *filp)
 	filp->f_flags |= O_LARGEFILE;
 	filp->f_mode |= FMODE_BUF_RASYNC | FMODE_CAN_ODIRECT;
 
-	/*
-	 * Use the file private data to store the holder for exclusive openes.
-	 * file_to_blk_mode relies on it being present to set BLK_OPEN_EXCL.
-	 */
-	if (filp->f_flags & O_EXCL)
-		filp->private_data = filp;
-
-	bdev = blkdev_get_by_dev(inode->i_rdev, file_to_blk_mode(filp),
-				 filp->private_data, NULL);
-	if (IS_ERR(bdev))
-		return PTR_ERR(bdev);
+	mode = file_to_blk_mode(filp);
+	handle = bdev_open_by_dev(inode->i_rdev, mode,
+			mode & BLK_OPEN_EXCL ? filp : NULL, NULL);
+	if (IS_ERR(handle))
+		return PTR_ERR(handle);
 
-	if (bdev_nowait(bdev))
+	if (bdev_nowait(handle->bdev))
 		filp->f_mode |= FMODE_NOWAIT;
 
-	filp->f_mapping = bdev->bd_inode->i_mapping;
+	filp->f_mapping = handle->bdev->bd_inode->i_mapping;
 	filp->f_wb_err = filemap_sample_wb_err(filp->f_mapping);
+	filp->private_data = handle;
 	return 0;
 }
 
 static int blkdev_release(struct inode *inode, struct file *filp)
 {
-	blkdev_put(I_BDEV(filp->f_mapping->host), filp->private_data);
+	bdev_release(filp->private_data);
 	return 0;
 }