From: Christoph Hellwig <hch@lst.de>
Date: Tue, 18 Jan 2022 08:47:22 +0000 (+0100)
Subject: block: cleanup the bd_holder locking
X-Git-Url: https://www.infradead.org/git/?a=commitdiff_plain;h=9cbaaa9a2298e5bc367379c47a561570c47a01eb;p=users%2Fhch%2Fblock.git

block: cleanup the bd_holder locking

Consistently use bdev_lock for all holder related fields and factor all
holder related work into helpers holding it instead of sometimes using
open_mutex.

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

diff --git a/block/bdev.c b/block/bdev.c
index 2e1d76ceabca..a301a4264419 100644
--- a/block/bdev.c
+++ b/block/bdev.c
@@ -611,17 +611,16 @@ static void bd_clear_claiming(struct block_device *whole, void *holder)
 	wake_up_bit(&whole->bd_claiming, 0);
 }
 
-/**
- * bd_finish_claiming - finish claiming of a block device
- * @bdev: block device of interest
- * @holder: holder that has claimed @bdev
- *
+/*
  * Finish exclusive open of a block device. Mark the device as exlusively
  * open by the holder and wake up all waiters for exclusive open to finish.
  */
-static void bd_finish_claiming(struct block_device *bdev, void *holder)
+static bool bd_finish_claiming(struct block_device *bdev, void *holder,
+		fmode_t mode)
 {
 	struct block_device *whole = bdev_whole(bdev);
+	struct gendisk *disk = bdev->bd_disk;
+	bool unblock_events = false;
 
 	spin_lock(&bdev_lock);
 	BUG_ON(!bd_may_claim(bdev, whole, holder));
@@ -634,6 +633,42 @@ static void bd_finish_claiming(struct block_device *bdev, void *holder)
 	bdev->bd_holders++;
 	bdev->bd_holder = holder;
 	bd_clear_claiming(whole, holder);
+
+	/*
+	 * Block event polling for write claims if requested.  Any write holder
+	 * makes the write_holder state stick until all are released.  This is
+	 * good enough and tracking individual writeable reference is too
+	 * fragile given the way @mode is used in blkdev_get/put().
+	 */
+	if (!bdev->bd_write_holder && (mode & FMODE_WRITE) &&
+	    (disk->event_flags & DISK_EVENT_FLAG_BLOCK_ON_EXCL_WRITE)) {
+		bdev->bd_write_holder = true;
+		unblock_events = true;
+	}
+
+	spin_unlock(&bdev_lock);
+	return unblock_events;
+}
+
+/*
+ * Release a claim on the device.
+ */
+static void bd_unclaim(struct block_device *bdev)
+{
+	struct block_device *whole = bdev_whole(bdev);
+
+	spin_lock(&bdev_lock);
+	WARN_ON_ONCE(bdev->bd_holders < 1);
+	WARN_ON_ONCE(whole->bd_holders < 1);
+	if (!--whole->bd_holders)
+		whole->bd_holder = NULL;
+	if (!--bdev->bd_holders) {
+		bdev->bd_holder = NULL;
+		if (bdev->bd_write_holder) {
+			disk_unblock_events(bdev->bd_disk);
+			bdev->bd_write_holder = false;
+		}
+	}
 	spin_unlock(&bdev_lock);
 }
 
@@ -781,7 +816,6 @@ void blkdev_put_no_open(struct block_device *bdev)
  */
 struct block_device *blkdev_get_by_dev(dev_t dev, fmode_t mode, void *holder)
 {
-	bool unblock_events = true;
 	struct block_device *bdev;
 	struct gendisk *disk;
 	int ret;
@@ -818,33 +852,17 @@ struct block_device *blkdev_get_by_dev(dev_t dev, fmode_t mode, void *holder)
 		ret = blkdev_get_whole(bdev, mode);
 	if (ret)
 		goto put_module;
-	if (mode & FMODE_EXCL) {
-		bd_finish_claiming(bdev, holder);
-
-		/*
-		 * Block event polling for write claims if requested.  Any write
-		 * holder makes the write_holder state stick until all are
-		 * released.  This is good enough and tracking individual
-		 * writeable reference is too fragile given the way @mode is
-		 * used in blkdev_get/put().
-		 */
-		if ((mode & FMODE_WRITE) && !bdev->bd_write_holder &&
-		    (disk->event_flags & DISK_EVENT_FLAG_BLOCK_ON_EXCL_WRITE)) {
-			bdev->bd_write_holder = true;
-			unblock_events = false;
-		}
-	}
 	mutex_unlock(&disk->open_mutex);
 
-	if (unblock_events)
+	if (!(mode & FMODE_EXCL) || bd_finish_claiming(bdev, holder, mode))
 		disk_unblock_events(disk);
 	return bdev;
 put_module:
 	module_put(disk->fops->owner);
 abort_claiming:
+	mutex_unlock(&disk->open_mutex);
 	if (mode & FMODE_EXCL)
 		bd_abort_claiming(bdev, holder);
-	mutex_unlock(&disk->open_mutex);
 	disk_unblock_events(disk);
 put_blkdev:
 	blkdev_put_no_open(bdev);
@@ -904,43 +922,15 @@ void blkdev_put(struct block_device *bdev, fmode_t mode)
 	if (bdev->bd_openers == 1)
 		sync_blockdev(bdev);
 
-	mutex_lock(&disk->open_mutex);
-	if (mode & FMODE_EXCL) {
-		struct block_device *whole = bdev_whole(bdev);
-		bool bdev_free;
-
-		/*
-		 * Release a claim on the device.  The holder fields
-		 * are protected with bdev_lock.  open_mutex is to
-		 * synchronize disk_holder unlinking.
-		 */
-		spin_lock(&bdev_lock);
-
-		WARN_ON_ONCE(--bdev->bd_holders < 0);
-		WARN_ON_ONCE(--whole->bd_holders < 0);
-
-		if ((bdev_free = !bdev->bd_holders))
-			bdev->bd_holder = NULL;
-		if (!whole->bd_holders)
-			whole->bd_holder = NULL;
-
-		spin_unlock(&bdev_lock);
-
-		/*
-		 * If this was the last claim, remove holder link and
-		 * unblock evpoll if it was a write holder.
-		 */
-		if (bdev_free && bdev->bd_write_holder) {
-			disk_unblock_events(disk);
-			bdev->bd_write_holder = false;
-		}
-	}
+	if (mode & FMODE_EXCL)
+		bd_unclaim(bdev);
 
 	/*
 	 * Trigger event checking and tell drivers to flush MEDIA_CHANGE
 	 * event.  This is to ensure detection of media removal commanded
 	 * from userland - e.g. eject(1).
 	 */
+	mutex_lock(&disk->open_mutex);
 	disk_flush_events(disk, DISK_EVENT_MEDIA_CHANGE);
 
 	if (bdev_is_partition(bdev))