loop: properly observe rotational flag of underlying device

PATCH In review Holger Hoffstätte 4 participants 6 messages Nov 11, 2015 – May 12, 2016 Block layer

Share & tools

Holger Hoffstätte

author

2015-11-11 15:21 UTC #0 lore raw

PATCH 1/1 loop: properly observe rotational flag of underlying device

The loop driver always declares the rotational flag of its device as rotational, even when the device of the mapped file is nonrotational, as is the case with SSDs or on tmpfs. This can confuse filesystem tools which are SSD-aware; in my case I frequently forget to tell mkfs.btrfs that my loop device on tmpfs is nonrotational, and that I really don't need any automatic metadata redundancy.

The attached patch fixes this by introspecting the rotational flag of the mapped file's underlying block device, if it exists. If the mapped file's filesystem has no associated block device - as is the case on e.g. tmpfs - we assume nonrotational storage. If there is a better way to identify such non-devices I'd love to hear them.

Signed-off-by Holger Hoffstätte <[email protected]>

---
 drivers/block/loop.c | 19 +++++++++++++++++++
 1 file changed, 19 insertions(+)

drivers/block/loop.c +19 -0

index 423f4ca..2984aca 100644
@@ -843,6 +843,24 @@ static void loop_config_discard(struct loop_device *lo)
 	queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, q);
 }
 
+static void loop_update_rotational(struct loop_device *lo)
+{
+	struct file *file = lo->lo_backing_file;
+	struct inode *file_inode = file->f_mapping->host;
+	struct block_device *file_bdev = file_inode->i_sb->s_bdev;
+	struct request_queue *q = lo->lo_queue;
+	bool nonrot = true;
+
+	/* not all filesystems (e.g. tmpfs) have a sb->s_bdev */
+	if (file_bdev)
+		nonrot = blk_queue_nonrot(bdev_get_queue(file_bdev));
+
+	if (nonrot)
+		queue_flag_set_unlocked(QUEUE_FLAG_NONROT, q);
+	else
+		queue_flag_clear_unlocked(QUEUE_FLAG_NONROT, q);
+}
+
 static void loop_unprepare_queue(struct loop_device *lo)
 {
 	flush_kthread_worker(&lo->worker);

@@ -939,6 +957,7 @@ static int loop_set_fd(struct loop_device *lo, fmode_t mode,
 	if (!(lo_flags & LO_FLAGS_READ_ONLY) && file->f_op->fsync)
 		blk_queue_flush(lo->lo_queue, REQ_FLUSH);
 
+	loop_update_rotational(lo);
 	loop_update_dio(lo);
 	set_capacity(lo->lo_disk, size);
 	bd_set_size(bdev, size << 9);

Signature · 1 line

2.6.3

Jens Axboe

2015-11-11 21:29 UTC #1 lore raw

Quoted 34 lines · Holger Hoffstätte

The loop driver always declares the rotational flag of its device as rotational, even when the device of the mapped file is nonrotational, as is the case with SSDs or on tmpfs. This can confuse filesystem tools which are SSD-aware; in my case I frequently forget to tell mkfs.btrfs that my loop device on tmpfs is nonrotational, and that I really don't need any automatic metadata redundancy.
The attached patch fixes this by introspecting the rotational flag of the mapped file's underlying block device, if it exists. If the mapped file's filesystem has no associated block device - as is the case on e.g. tmpfs - we assume nonrotational storage. If there is a better way to identify such non-devices I'd love to hear them.
Signed-off-by Holger Hoffstätte <[email protected]>
---
  drivers/block/loop.c | 19 +++++++++++++++++++
  1 file changed, 19 insertions(+)
drivers/block/loop.c +7 -0
index 423f4ca..2984aca 100644
@@ -843,6 +843,24 @@ static void loop_config_discard(struct loop_device *lo)
  	queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, q);
  }
 
+static void loop_update_rotational(struct loop_device *lo)
+{
+	struct file *file = lo->lo_backing_file;
+	struct inode *file_inode = file->f_mapping->host;
+	struct block_device *file_bdev = file_inode->i_sb->s_bdev;
+	struct request_queue *q = lo->lo_queue;
+	bool nonrot = true;

drivers/block/loop.c static void loop_config_discard(struct loop_device *lo) · line 862

Quoted 10 lines

drivers/block/loop.c +10 -0

+
+	/* not all filesystems (e.g. tmpfs) have a sb->s_bdev */
+	if (file_bdev)
+		nonrot = blk_queue_nonrot(bdev_get_queue(file_bdev));
+
+	if (nonrot)
+		queue_flag_set_unlocked(QUEUE_FLAG_NONROT, q);
+	else
+		queue_flag_clear_unlocked(QUEUE_FLAG_NONROT, q);
+}

Are we sure we want to change the default from rot to nonrot?

Apart from that, looks good.

Signature · 1 line

Jens Axboe

Holger Hoffstätte

author

2015-11-11 22:08 UTC #2 lore raw

Quoted 39 lines · Jens Axboe

Quoted 38 lines · Holger Hoffstätte
The loop driver always declares the rotational flag of its device as rotational, even when the device of the mapped file is nonrotational, as is the case with SSDs or on tmpfs. This can confuse filesystem tools which are SSD-aware; in my case I frequently forget to tell mkfs.btrfs that my loop device on tmpfs is nonrotational, and that I really don't need any automatic metadata redundancy.
The attached patch fixes this by introspecting the rotational flag of the mapped file's underlying block device, if it exists. If the mapped file's filesystem has no associated block device - as is the case on e.g. tmpfs - we assume nonrotational storage. If there is a better way to identify such non-devices I'd love to hear them.
Signed-off-by Holger Hoffstätte <[email protected]>
---
  drivers/block/loop.c | 19 +++++++++++++++++++
  1 file changed, 19 insertions(+)
drivers/block/loop.c +11 -0
index 423f4ca..2984aca 100644
@@ -843,6 +843,24 @@ static void loop_config_discard(struct loop_device *lo)
      queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, q);
  }
 
+static void loop_update_rotational(struct loop_device *lo)
+{
+    struct file *file = lo->lo_backing_file;
+    struct inode *file_inode = file->f_mapping->host;
+    struct block_device *file_bdev = file_inode->i_sb->s_bdev;
+    struct request_queue *q = lo->lo_queue;
+    bool nonrot = true;
+
+    /* not all filesystems (e.g. tmpfs) have a sb->s_bdev */
+    if (file_bdev)
+        nonrot = blk_queue_nonrot(bdev_get_queue(file_bdev));

Quoted 8 lines

drivers/block/loop.c static void loop_config_discard(struct loop_device *lo) · line 862
Quoted 6 lines
drivers/block/loop.c +6 -0
+
+    if (nonrot)
+        queue_flag_set_unlocked(QUEUE_FLAG_NONROT, q);
+    else
+        queue_flag_clear_unlocked(QUEUE_FLAG_NONROT, q);
+}
Are we sure we want to change the default from rot to nonrot?

Well, that's why I asked for a better way to identify tmpfs. It took me several hours to figure out that tmpfs doesn't have an s_bdev, and could not find a better way than to assume that a superblock without backing device is probably something virtual/nonrotational/nvm etc.

Alternatively I could look at sb->s_type and set nonrot for known fs types, but that seemed too ugly - not to mention conceptually weird.

Quoted 1 line

Apart from that, looks good.

phew :)

thanks,
Holger

Holger Hoffstätte

author

2016-01-11 23:28 UTC #3 lore raw

Quoted 49 lines · Holger Hoffstätte

Quoted 39 lines · Jens Axboe
Quoted 38 lines · Holger Hoffstätte
The loop driver always declares the rotational flag of its device as rotational, even when the device of the mapped file is nonrotational, as is the case with SSDs or on tmpfs. This can confuse filesystem tools which are SSD-aware; in my case I frequently forget to tell mkfs.btrfs that my loop device on tmpfs is nonrotational, and that I really don't need any automatic metadata redundancy.
The attached patch fixes this by introspecting the rotational flag of the mapped file's underlying block device, if it exists. If the mapped file's filesystem has no associated block device - as is the case on e.g. tmpfs - we assume nonrotational storage. If there is a better way to identify such non-devices I'd love to hear them.
Signed-off-by Holger Hoffstätte <[email protected]>
---
  drivers/block/loop.c | 19 +++++++++++++++++++
  1 file changed, 19 insertions(+)
drivers/block/loop.c +11 -0
index 423f4ca..2984aca 100644
@@ -843,6 +843,24 @@ static void loop_config_discard(struct loop_device *lo)
      queue_flag_set_unlocked(QUEUE_FLAG_DISCARD, q);
  }
 
+static void loop_update_rotational(struct loop_device *lo)
+{
+    struct file *file = lo->lo_backing_file;
+    struct inode *file_inode = file->f_mapping->host;
+    struct block_device *file_bdev = file_inode->i_sb->s_bdev;
+    struct request_queue *q = lo->lo_queue;
+    bool nonrot = true;
+
+    /* not all filesystems (e.g. tmpfs) have a sb->s_bdev */
+    if (file_bdev)
+        nonrot = blk_queue_nonrot(bdev_get_queue(file_bdev));
Quoted 8 lines
drivers/block/loop.c static void loop_config_discard(struct loop_device *lo) · line 862
Quoted 6 lines
drivers/block/loop.c +6 -0
+
+    if (nonrot)
+        queue_flag_set_unlocked(QUEUE_FLAG_NONROT, q);
+    else
+        queue_flag_clear_unlocked(QUEUE_FLAG_NONROT, q);
+}
Are we sure we want to change the default from rot to nonrot?

Quoted 9 lines

Well, that's why I asked for a better way to identify tmpfs. It took me several hours to figure out that tmpfs doesn't have an s_bdev, and could not find a better way than to assume that a superblock without backing device is probably something virtual/nonrotational/nvm etc.
Alternatively I could look at sb->s_type and set nonrot for known fs types, but that seemed too ugly - not to mention conceptually weird.
Quoted 1 line
Apart from that, looks good.

Jens,

I haven't seen this merged in any trees yet and was wondering if there's any chance to get this into 4.5? If there's something left to fix up please let me know.

Thanks,
Holger

gwendal grignou

2016-05-12 20:30 UTC #4 lore raw

Quoted 8 lines · Holger Hoffstätte <holger.hoffstaette <at> googlemail.com> writes:

Quoted 6 lines · Holger Hoffstätte
Quoted 5 lines · Jens Axboe
Quoted 4 lines · Holger Hoffstätte
The loop driver always declares the rotational flag of its device as rotational, even when the device of the mapped file is nonrotational, as is the case with SSDs or on tmpfs. This can confuse filesystem

tools

Quoted 1 line

Quoted 1 line
Quoted 1 line
Quoted 1 line
which are SSD-aware; in my case I frequently forget to tell

mkfs.btrfs

Quoted 1 line

Quoted 1 line
Quoted 1 line
Quoted 1 line
that my loop device on tmpfs is nonrotational, and that I really

don't

Quoted 3 lines

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
need any automatic metadata redundancy.
The attached patch fixes this by introspecting the rotational flag of

the

Quoted 1 line

Quoted 1 line
Quoted 1 line
Quoted 1 line
mapped file's underlying block device, if it exists. If the mapped

file's

Quoted 1 line

Quoted 1 line
Quoted 1 line
Quoted 1 line
filesystem has no associated block device - as is the case on e.g.

tmpfs -

Quoted 1 line

Quoted 1 line
Quoted 1 line
Quoted 1 line
we assume nonrotational storage. If there is a better way to identify

such

Quoted 3 lines

Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
non-devices I'd love to hear them.
Signed-off-by Holger Hoffstätte <holger.hoffstaette <at>

googlemail.com>

Quoted 5 lines

Jens,
I haven't seen this merged in any trees yet and was wondering if there's any chance to get this into 4.5? If there's something left to fix up

please

Quoted 6 lines

let me know.
Thanks,
Holger

This patch proved useful for ureadahead: when we use it on a loop device, it would use the HDD method to place the data in cache using the pack information instead of the SSD method.

Signed-off-by Gwendal Grignou <[email protected]>

Holger Hoffstätte

2016-05-12 22:30 UTC #5 lore raw

[cc: linux-block]

Quoted 42 lines · gwendal grignou

Quoted 8 lines · Holger Hoffstätte <holger.hoffstaette <at> googlemail.com> writes:
Quoted 6 lines · Holger Hoffstätte
Quoted 5 lines · Jens Axboe
Quoted 4 lines · Holger Hoffstätte
The loop driver always declares the rotational flag of its device as rotational, even when the device of the mapped file is nonrotational, as is the case with SSDs or on tmpfs. This can confuse filesystem
tools
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
which are SSD-aware; in my case I frequently forget to tell
mkfs.btrfs
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
that my loop device on tmpfs is nonrotational, and that I really
don't
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
need any automatic metadata redundancy.
The attached patch fixes this by introspecting the rotational flag of
the
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
mapped file's underlying block device, if it exists. If the mapped
file's
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
filesystem has no associated block device - as is the case on e.g.
tmpfs -
Quoted 1 line
Quoted 1 line
Quoted 1 line
Quoted 1 line
we assume nonrotational storage. If there is a better way to identify
such
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
Quoted 3 lines
non-devices I'd love to hear them.
Signed-off-by Holger Hoffstätte <holger.hoffstaette <at>
googlemail.com>
Quoted 5 lines
Jens,
I haven't seen this merged in any trees yet and was wondering if there's any chance to get this into 4.5? If there's something left to fix up
please
Quoted 6 lines
let me know.
Thanks,
Holger

Quoted 5 lines

This patch proved useful for ureadahead: when we use it on a loop device, it would use the HDD method to place the data in cache using the pack information instead of the SSD method.
Signed-off-by Gwendal Grignou <[email protected]>

I had completely forgotten about this, and apparently so had Jens. ;) Thanks for the feedback, glad to hear it is useful.

Jens, any objections to merge this for 4.7? It should still apply cleanly. The original patch was at: https://lkml.org/lkml/2015/11/11/288

Thanks,
Holger