Preallocate flush bio, sysfs tunable

For devices that support flushing, we allocate a bio, submit, wait for it and then free it. The bio allocation does not fail so ENOMEM is not a problem but we still may unnecessarily stress the allocation subsystem.

Instead, we can allocate the device at the same time we allocate the device and reuse it each time we need to flush the barriers. The bio is reset before each use. Reference counting is simplified to just device allocation (get) and freeing (put).

Note for write_dev_flush: we check the queue flush status again as we can't use the existence of bio as before.

Signed-off-by David Sterba <[email protected]>

---
  fs/btrfs/disk-io.c | 24 ++++++------------------
  fs/btrfs/volumes.c | 12 ++++++++++++
  2 files changed, 18 insertions(+), 18 deletions(-)

fs/btrfs/disk-io.c +3 -10

index 2b00ebff13f8..27d44d6ab775 100644
@@ -3482,9 +3482,7 @@ static int write_dev_supers(struct btrfs_device *device,
   */
  static void btrfs_end_empty_barrier(struct bio *bio)
  {
-	if (bio->bi_private)
-		complete(bio->bi_private);
-	bio_put(bio);
+	complete(bio->bi_private);
  }
 
  /*

@@ -3494,26 +3492,19 @@ static void btrfs_end_empty_barrier(struct bio *bio)
  static void write_dev_flush(struct btrfs_device *device)
  {
  	struct request_queue *q = bdev_get_queue(device->bdev);
-	struct bio *bio;
+	struct bio *bio = device->flush_bio;
 
  	if (!test_bit(QUEUE_FLAG_WC, &q->queue_flags))
  		return;
 
-	/*
-	 * one reference for us, and we leave it for the
-	 * caller
-	 */
-	device->flush_bio = NULL;
-	bio = btrfs_io_bio_alloc(0);
+	bio_reset(bio);

fs/btrfs/disk-io.c +0 -4

  	wait_for_completion(&device->flush_wait);

@@ -3535,10 +3527,6 @@ static int wait_dev_flush(struct btrfs_device *device)
  				BTRFS_DEV_STAT_FLUSH_ERRS);
  	}
 
-	/* drop the reference from the wait == 0 run */
-	bio_put(bio);
-	device->flush_bio = NULL;
-
  	return ret;
  }
 

fs/btrfs/volumes.c +0 -0

index 8bb1f4e5905a..251ae81e4363 100644

fs/btrfs/volumes.c +7 -0

+	if (!dev->flush_bio) {
+		kfree(dev);
+		return ERR_PTR(-ENOMEM);
+	}
+	bio_get(dev->flush_bio);
+
  	INIT_LIST_HEAD(&dev->dev_list);
  	INIT_LIST_HEAD(&dev->dev_alloc_list);
  	INIT_LIST_HEAD(&dev->resized_list);

@@ -838,6 +849,7 @@ static void __free_device(struct work_struct *work)
 
  	device = container_of(work, struct btrfs_device, rcu_work);
  	rcu_string_free(device->name);
+	bio_put(device->flush_bio);
  	kfree(device);
  }
 

We should really just wait in wait_dev_flush and let the caller decide what to do with the error value.

Signed-off-by David Sterba <[email protected]>

---
  fs/btrfs/disk-io.c | 11 +++--------
  1 file changed, 3 insertions(+), 8 deletions(-)

fs/btrfs/disk-io.c +3 -8

index 05ff81ecb887..59a732a13370 100644
@@ -3512,7 +3512,6 @@ static void write_dev_flush(struct btrfs_device *device)
   */
  static int wait_dev_flush(struct btrfs_device *device)
  {
-	int ret = 0;
  	struct request_queue *q = bdev_get_queue(device->bdev);
  	struct bio *bio = device->flush_bio;
 

@@ -3521,13 +3520,7 @@ static int wait_dev_flush(struct btrfs_device *device)
 
  	wait_for_completion_io(&device->flush_wait);
 
-	if (bio->bi_error) {
-		ret = bio->bi_error;
-		btrfs_dev_stat_inc_and_print(device,
-				BTRFS_DEV_STAT_FLUSH_ERRS);
-	}
-
-	return ret;
+	return bio->bi_error;
  }
 
  static int check_barrier_error(struct btrfs_fs_devices *fsdevs)

@@ -3586,6 +3579,8 @@ static int barrier_all_devices(struct btrfs_fs_info *info)
  		ret = wait_dev_flush(dev);
  		if (ret) {
  			dev->last_flush_error = ret;
+			btrfs_dev_stat_inc_and_print(dev,
+					BTRFS_DEV_STAT_FLUSH_ERRS);
  			errors_wait++;
  		}
  	}

We need a device capable of device barriers so we can test the flush code. To aid testing, add a per-filesystem status flag that affects the barriers regerdless of the device capabilities and obviously does not give the same guarantees.

It's off by default, sysfs tunable will follow.

Signed-off-by David Sterba <[email protected]>

---
  fs/btrfs/ctree.h   | 1 +
  fs/btrfs/disk-io.c | 8 +++++---
  fs/btrfs/volumes.h | 1 +
  3 files changed, 7 insertions(+), 3 deletions(-)

fs/btrfs/ctree.h +1 -0

index f0f5f28784b6..dcf4404f7d61 100644
@@ -716,6 +716,7 @@ struct btrfs_delayed_root;
  #define BTRFS_FS_LOG1_ERR			12
  #define BTRFS_FS_LOG2_ERR			13
  #define BTRFS_FS_QUOTA_OVERRIDE			14
+#define BTRFS_FS_FORCE_DEV_FLUSH		15
 
  /*
   * Indicate that a whole-filesystem exclusive operation is running

fs/btrfs/disk-io.c +0 -0

index 59a732a13370..659a3b4645d2 100644

fs/btrfs/disk-io.c +3 -2

  	bio_reset(bio);

@@ -3505,6 +3506,7 @@ static void write_dev_flush(struct btrfs_device *device)
  	bio->bi_private = &device->flush_wait;
 
  	submit_bio(bio);
+	device->flush_bio_sent = 1;
  }
 
  /*

@@ -3512,12 +3514,12 @@ static void write_dev_flush(struct btrfs_device *device)
   */
  static int wait_dev_flush(struct btrfs_device *device)
  {
-	struct request_queue *q = bdev_get_queue(device->bdev);
  	struct bio *bio = device->flush_bio;
 
-	if (!test_bit(QUEUE_FLAG_WC, &q->queue_flags))
+	if (!device->flush_bio_sent)
  		return 0;
 
+	device->flush_bio_sent = 0;
  	wait_for_completion_io(&device->flush_wait);
 
  	return bio->bi_error;

fs/btrfs/volumes.h +1 -0

index 35327efecdbb..6f45fd60d15a 100644
@@ -75,6 +75,7 @@ struct btrfs_device {
  	int can_discard;
  	int is_tgtdev_for_dev_replace;
  	int last_flush_error;
+	int flush_bio_sent;
 
  #ifdef __BTRFS_NEED_DEVICE_DATA_ORDERED
  	seqcount_t data_seqcount;

Add per-filesystem tunable to switch on/off the device barriers, regardless of the actual device support. This is a debugging feature.

The path to the sysfs file is /sys/fs/btrfs/UUID/force_dev_flush, allowed values are 0 and 1. Default is 0.

Signed-off-by David Sterba <[email protected]>

---
  fs/btrfs/sysfs.c | 47 +++++++++++++++++++++++++++++++++++++++++++++++
  1 file changed, 47 insertions(+)

fs/btrfs/sysfs.c +47 -0

index c2d5f3580b4c..197d43911936 100644
@@ -487,12 +487,59 @@ static ssize_t quota_override_store(struct kobject *kobj,
 
  BTRFS_ATTR_RW(quota_override, quota_override_show, quota_override_store);
 
+static ssize_t force_dev_flush_show(struct kobject *kobj,
+				   struct kobj_attribute *a, char *buf)
+{
+	struct btrfs_fs_info *fs_info = to_fs_info(kobj);
+	int force;
+
+	if (!fs_info)
+		return -EPERM;
+
+	force = !!(test_bit(BTRFS_FS_FORCE_DEV_FLUSH, &fs_info->flags));
+	return snprintf(buf, PAGE_SIZE, "%d\n", force);
+}
+
+static ssize_t force_dev_flush_store(struct kobject *kobj,
+				    struct kobj_attribute *a,
+				    const char *buf, size_t len)
+{
+	struct btrfs_fs_info *fs_info = to_fs_info(kobj);
+	unsigned long force;
+	int err;
+
+	if (!fs_info)
+		return -EPERM;
+
+	if (!capable(CAP_SYS_ADMIN))
+		return -EPERM;
+
+	err = kstrtoul(buf, 10, &force);
+	if (err)
+		return err;
+	if (force > 1)
+		return -EINVAL;
+
+	if (force) {
+		set_bit(BTRFS_FS_FORCE_DEV_FLUSH, &fs_info->flags);
+		btrfs_info(fs_info, "Forced device flushes enabled");
+	} else {
+		clear_bit(BTRFS_FS_FORCE_DEV_FLUSH, &fs_info->flags);
+		btrfs_info(fs_info, "Forced device flushes disabled");
+	}
+
+	return len;
+}
+
+BTRFS_ATTR_RW(force_dev_flush, force_dev_flush_show, force_dev_flush_store);
+
  static const struct attribute *btrfs_attrs[] = {
  	BTRFS_ATTR_PTR(label),
  	BTRFS_ATTR_PTR(nodesize),
  	BTRFS_ATTR_PTR(sectorsize),
  	BTRFS_ATTR_PTR(clone_alignment),
  	BTRFS_ATTR_PTR(quota_override),
+	BTRFS_ATTR_PTR(force_dev_flush),
  	NULL,
  };
 

Quoted 9 lines

fs/btrfs/disk-io.c +0 -0

index 59a732a13370..659a3b4645d2 100644

fs/btrfs/disk-io.c static void write_dev_flush(struct btrfs_device · line 3499

Quoted 7 lines

fs/btrfs/disk-io.c +2 -1

@@ -3494,7 +3494,8 @@ static void write_dev_flush(struct btrfs_device 
 *device)
      struct request_queue *q = bdev_get_queue(device->bdev);
      struct bio *bio = device->flush_bio;
-    if (!test_bit(QUEUE_FLAG_WC, &q->queue_flags))
+    if (!test_bit(BTRFS_FS_FORCE_DEV_FLUSH, &device->fs_info->flags)
+        && !test_bit(QUEUE_FLAG_WC, &q->queue_flags))
          return;

Now I understand what you meant. But the most common case in our test set up is a device with write cache. So BTRFS_FS_FORCE_DEV_FLUSH does not bring any additional force. IMO.

Thanks, Anand

Quoted 11 lines · David Sterba

For devices that support flushing, we allocate a bio, submit, wait for it and then free it. The bio allocation does not fail so ENOMEM is not a problem but we still may unnecessarily stress the allocation subsystem.

Instead, we can allocate the device at the same time we allocate the device and reuse it each time we need to flush the barriers. The bio is reset before each use. Reference counting is simplified to just device allocation (get) and freeing (put).

Note for write_dev_flush: we check the queue flush status again as we can't use the existence of bio as before.

Looks good few items as below..

Quoted 45 lines

Signed-off-by David Sterba <[email protected]>

---
  fs/btrfs/disk-io.c | 24 ++++++------------------
  fs/btrfs/volumes.c | 12 ++++++++++++
  2 files changed, 18 insertions(+), 18 deletions(-)

fs/btrfs/disk-io.c +3 -11

index 2b00ebff13f8..27d44d6ab775 100644
@@ -3482,9 +3482,7 @@ static int write_dev_supers(struct btrfs_device *device,
   */
  static void btrfs_end_empty_barrier(struct bio *bio)
  {
-	if (bio->bi_private)
-		complete(bio->bi_private);
-	bio_put(bio);
+	complete(bio->bi_private);
  }
 
  /*

@@ -3494,26 +3492,19 @@ static void btrfs_end_empty_barrier(struct bio *bio)
  static void write_dev_flush(struct btrfs_device *device)
  {
  	struct request_queue *q = bdev_get_queue(device->bdev);
-	struct bio *bio;
+	struct bio *bio = device->flush_bio;
 
  	if (!test_bit(QUEUE_FLAG_WC, &q->queue_flags))
  		return;
 
-	/*
-	 * one reference for us, and we leave it for the
-	 * caller
-	 */
-	device->flush_bio = NULL;
-	bio = btrfs_io_bio_alloc(0);
+	bio_reset(bio);
  	bio->bi_end_io = btrfs_end_empty_barrier;
  	bio->bi_bdev = device->bdev;
  	bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_PREFLUSH;
  	init_completion(&device->flush_wait);
  	bio->bi_private = &device->flush_wait;
-	device->flush_bio = bio;
 

fs/btrfs/disk-io.c static void btrfs_end_empty_barrier(struct bio *bio) · line 3507

Quoted 3 lines

fs/btrfs/disk-io.c +1 -2

-	bio_get(bio);
-	btrfsic_submit_bio(bio);
+	submit_bio(bio);

Originally it went through the btrfsic. There is no mention of this change if its not an oversight.

fs/btrfs/disk-io.c static void write_dev_flush(struct btrfs_device *device) · line 3520

Quoted 11 lines

fs/btrfs/disk-io.c +2 -1

  /*

@@ -3522,9 +3513,10 @@ static void write_dev_flush(struct btrfs_device *device)
  static int wait_dev_flush(struct btrfs_device *device)
  {
  	int ret = 0;
+	struct request_queue *q = bdev_get_queue(device->bdev);
  	struct bio *bio = device->flush_bio;
 
-	if (!bio)
+	if (!test_bit(QUEUE_FLAG_WC, &q->queue_flags))
  		return 0;

It returns here if its write through. Which can be toggled after write_dev_flush() has been called such as..

echo "write back" > /sys/block/sdd/queue/write_cache write_dev_flush(sdd)
echo "write through" > /sys/block/sdd/queue/write_cache wait_dev_flush(sdd)

So it would fails to check error.

Quoted 22 lines

Quoted 9 lines

fs/btrfs/disk-io.c +0 -0

index 59a732a13370..659a3b4645d2 100644

fs/btrfs/disk-io.c static void write_dev_flush(struct btrfs_device · line 3499

Quoted 7 lines

fs/btrfs/disk-io.c +2 -1

@@ -3494,7 +3494,8 @@ static void write_dev_flush(struct btrfs_device 
 *device)
      struct request_queue *q = bdev_get_queue(device->bdev);
      struct bio *bio = device->flush_bio;
-    if (!test_bit(QUEUE_FLAG_WC, &q->queue_flags))
+    if (!test_bit(BTRFS_FS_FORCE_DEV_FLUSH, &device->fs_info->flags)
+        && !test_bit(QUEUE_FLAG_WC, &q->queue_flags))
          return;

Now I understand what you meant. But the most common case in our test set up is a device with write cache. So BTRFS_FS_FORCE_DEV_FLUSH does not bring any additional force. IMO.

Quoted 1 line

Thanks, Anand

Or one another idea is we could remove

    !test_bit(QUEUE_FLAG_WC, &q->queue_flags)

which purpose is to only fail early.

If we remove it there is consistency in our code with or with out the write cache.