Reporting free space to userspace programs

Reporting free space to userspace programs

[Leszek Ciesielski]

Hi,

the results of running 'df' against a btrfs volume are somewhat
unintuitive from a user point of view. On a single drive btrfs volume,
created with 'mkfs.btrfs -m raid1 -d raid1 /dev/sda6', I am getting
the following result:

/dev/sda6             1.4T  594G  804G  43% /mnt

while 'btrfs-show' displays much more expected result:

Label: none  uuid: 46e2f2b6-e3a6-4b02-8fdc-f9d0fb0882e0
        Total devices 1 FS bytes used 593.15GB
        devid    1 size 1.36TB used 1.26TB path /dev/sda6

IMHO it would be more intuitive for df in this case to show 699GB
total capacity (based on the fact that data is mirrored, and users
probably are not concerned with metadata handling during normal
usage), the 'used space' probably should include the space taken up by
metadata in addition to data usage (after all, this space is not
available for user data) and free space should report only data space
available (because this is what the user is usually expecting). Or, in
other words: the result of 'df' should not concern the user with the
details of raid0/raid1/raid10 used either for data or metadata.

Anyone care to comment?

kudos :-)


Leszek 'skolima' Ciesielski
---
I am running kernel 2.6.31-gentoo and using btrfs-progs 0.19, please
excuse me if my comments are no longer relevant (however, I did check
commit messages from the version I use up until
linux/kernel/git/mason/btrfs-unstable.git master-HEAD and did not find
anything related to the topic).

Re: Reporting free space to userspace programs

[jim owens]

Leszek Ciesielski wrote:
> Hi,
> 
> the results of running 'df' against a btrfs volume are somewhat
> unintuitive from a user point of view. On a single drive btrfs volume,
> created with 'mkfs.btrfs -m raid1 -d raid1 /dev/sda6', I am getting
> the following result:
> 
> /dev/sda6             1.4T  594G  804G  43% /mnt
> 
> while 'btrfs-show' displays much more expected result:
> 
> Label: none  uuid: 46e2f2b6-e3a6-4b02-8fdc-f9d0fb0882e0
>         Total devices 1 FS bytes used 593.15GB
>         devid    1 size 1.36TB used 1.26TB path /dev/sda6
> 
> IMHO it would be more intuitive for df in this case to show 699GB
> total capacity (based on the fact that data is mirrored, and users
> probably are not concerned with metadata handling during normal
> usage), the 'used space' probably should include the space taken up by
> metadata in addition to data usage (after all, this space is not
> available for user data) and free space should report only data space
> available (because this is what the user is usually expecting). Or, in
> other words: the result of 'df' should not concern the user with the
> details of raid0/raid1/raid10 used either for data or metadata.

I agree that df output sucks... but I've been there before with
another filesystem on another OS.  The sad fact is df output is
too simplistic for the features of modern (last 20 years) systems.

There is no way to make df report a value other than "raw space"
(which is what btrfs reports today) that will be accurate under
all possible raid conditions.  The problem is each file can be
stored in a different raid (OK not done now, but permitted) and
different COW state.  That means space_used_per_user_file_block
is not constant.

So btrfs can only report "best case" or "worst case", but neither
will be true.

jim

Re: Reporting free space to userspace programs

[Leszek Ciesielski]

Hi,

in a long overdue followup to my previous email, I am sending a patch
that modifies the result of running 'df' against a btrfs volume. I
understand that, give the simplicity of 'df', there is not 'correct'
solution - I do think however, that the changed output is more
intuitive. Most importantly - the free/used space percentage are
reported correctly, which should decrease the frequency of 'my 50%
filled btrfs volume is failing with ENOSPC' emails.

Would anyone like to comment?

diff --git a/fs/btrfs/super.c b/fs/btrfs/super.c
index 8a1ea6e..893c154 100644
--- a/fs/btrfs/super.c
+++ b/fs/btrfs/super.c
@@ -623,13 +623,18 @@ static int btrfs_statfs(struct dentry *dentry,
struct kstatfs *buf)
 {
 	struct btrfs_root *root =3D btrfs_sb(dentry->d_sb);
 	struct btrfs_super_block *disk_super =3D &root->fs_info->super_copy;
+	struct btrfs_device *device;
 	int bits =3D dentry->d_sb->s_blocksize_bits;
 	__be32 *fsid =3D (__be32 *)root->fs_info->fsid;

 	buf->f_namelen =3D BTRFS_NAME_LEN;
 	buf->f_blocks =3D btrfs_super_total_bytes(disk_super) >> bits;
-	buf->f_bfree =3D buf->f_blocks -
-		(btrfs_super_bytes_used(disk_super) >> bits);
+	buf->f_bfree =3D buf->f_blocks;
+	mutex_lock(&root->fs_info->fs_devices->device_list_mutex);
+	list_for_each_entry(device, &root->fs_info->fs_devices->devices, dev_=
list) {
+	    buf->f_bfree -=3D (device->bytes_used >> bits);
+	}
+	mutex_unlock(&root->fs_info->fs_devices->device_list_mutex);
 	buf->f_bavail =3D buf->f_bfree;
 	buf->f_bsize =3D dentry->d_sb->s_blocksize;
 	buf->f_type =3D BTRFS_SUPER_MAGIC;


On Fri, Oct 30, 2009 at 2:21 PM, jim owens <jowens@hp.com> wrote:
> Leszek Ciesielski wrote:
>>
>> Hi,
>>
>> the results of running 'df' against a btrfs volume are somewhat
>> unintuitive from a user point of view. On a single drive btrfs volum=
e,
>> created with 'mkfs.btrfs -m raid1 -d raid1 /dev/sda6', I am getting
>> the following result:
>>
>> /dev/sda6 =A0 =A0 =A0 =A0 =A0 =A0 1.4T =A0594G =A0804G =A043% /mnt
>>
>> while 'btrfs-show' displays much more expected result:
>>
>> Label: none =A0uuid: 46e2f2b6-e3a6-4b02-8fdc-f9d0fb0882e0
>> =A0 =A0 =A0 =A0Total devices 1 FS bytes used 593.15GB
>> =A0 =A0 =A0 =A0devid =A0 =A01 size 1.36TB used 1.26TB path /dev/sda6
>>
>> IMHO it would be more intuitive for df in this case to show 699GB
>> total capacity (based on the fact that data is mirrored, and users
>> probably are not concerned with metadata handling during normal
>> usage), the 'used space' probably should include the space taken up =
by
>> metadata in addition to data usage (after all, this space is not
>> available for user data) and free space should report only data spac=
e
>> available (because this is what the user is usually expecting). Or, =
in
>> other words: the result of 'df' should not concern the user with the
>> details of raid0/raid1/raid10 used either for data or metadata.
>
> I agree that df output sucks... but I've been there before with
> another filesystem on another OS. =A0The sad fact is df output is
> too simplistic for the features of modern (last 20 years) systems.
>
> There is no way to make df report a value other than "raw space"
> (which is what btrfs reports today) that will be accurate under
> all possible raid conditions. =A0The problem is each file can be
> stored in a different raid (OK not done now, but permitted) and
> different COW state. =A0That means space_used_per_user_file_block
> is not constant.
>
> So btrfs can only report "best case" or "worst case", but neither
> will be true.
>
> jim
>
>
--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" =
in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html

Re: Reporting free space to userspace programs

[Mitch Harder]

On Tue, Feb 23, 2010 at 4:27 PM, Leszek Ciesielski <skolima@gmail.com> =
wrote:
> Hi,
>
> in a long overdue followup to my previous email, I am sending a patch
> that modifies the result of running 'df' against a btrfs volume. I
> understand that, give the simplicity of 'df', there is not 'correct'
> solution - I do think however, that the changed output is more
> intuitive. Most importantly - the free/used space percentage are
> reported correctly, which should decrease the frequency of 'my 50%
> filled btrfs volume is failing with ENOSPC' emails.
>
> Would anyone like to comment?
>
> diff --git a/fs/btrfs/super.c b/fs/btrfs/super.c
> index 8a1ea6e..893c154 100644
> --- a/fs/btrfs/super.c
> +++ b/fs/btrfs/super.c
> @@ -623,13 +623,18 @@ static int btrfs_statfs(struct dentry *dentry,
> struct kstatfs *buf)
> =A0{
> =A0 =A0 =A0 =A0struct btrfs_root *root =3D btrfs_sb(dentry->d_sb);
> =A0 =A0 =A0 =A0struct btrfs_super_block *disk_super =3D &root->fs_inf=
o->super_copy;
> + =A0 =A0 =A0 struct btrfs_device *device;
> =A0 =A0 =A0 =A0int bits =3D dentry->d_sb->s_blocksize_bits;
> =A0 =A0 =A0 =A0__be32 *fsid =3D (__be32 *)root->fs_info->fsid;
>
> =A0 =A0 =A0 =A0buf->f_namelen =3D BTRFS_NAME_LEN;
> =A0 =A0 =A0 =A0buf->f_blocks =3D btrfs_super_total_bytes(disk_super) =
>> bits;
> - =A0 =A0 =A0 buf->f_bfree =3D buf->f_blocks -
> - =A0 =A0 =A0 =A0 =A0 =A0 =A0 (btrfs_super_bytes_used(disk_super) >> =
bits);
> + =A0 =A0 =A0 buf->f_bfree =3D buf->f_blocks;
> + =A0 =A0 =A0 mutex_lock(&root->fs_info->fs_devices->device_list_mute=
x);
> + =A0 =A0 =A0 list_for_each_entry(device, &root->fs_info->fs_devices-=
>devices, dev_list) {
> + =A0 =A0 =A0 =A0 =A0 buf->f_bfree -=3D (device->bytes_used >> bits);
> + =A0 =A0 =A0 }
> + =A0 =A0 =A0 mutex_unlock(&root->fs_info->fs_devices->device_list_mu=
tex);
> =A0 =A0 =A0 =A0buf->f_bavail =3D buf->f_bfree;
> =A0 =A0 =A0 =A0buf->f_bsize =3D dentry->d_sb->s_blocksize;
> =A0 =A0 =A0 =A0buf->f_type =3D BTRFS_SUPER_MAGIC;
>
>
> On Fri, Oct 30, 2009 at 2:21 PM, jim owens <jowens@hp.com> wrote:
>> Leszek Ciesielski wrote:
>>>
>>> Hi,
>>>
>>> the results of running 'df' against a btrfs volume are somewhat
>>> unintuitive from a user point of view. On a single drive btrfs volu=
me,
>>> created with 'mkfs.btrfs -m raid1 -d raid1 /dev/sda6', I am getting
>>> the following result:
>>>
>>> /dev/sda6 =A0 =A0 =A0 =A0 =A0 =A0 1.4T =A0594G =A0804G =A043% /mnt
>>>
>>> while 'btrfs-show' displays much more expected result:
>>>
>>> Label: none =A0uuid: 46e2f2b6-e3a6-4b02-8fdc-f9d0fb0882e0
>>> =A0 =A0 =A0 =A0Total devices 1 FS bytes used 593.15GB
>>> =A0 =A0 =A0 =A0devid =A0 =A01 size 1.36TB used 1.26TB path /dev/sda=
6
>>>
>>> IMHO it would be more intuitive for df in this case to show 699GB
>>> total capacity (based on the fact that data is mirrored, and users
>>> probably are not concerned with metadata handling during normal
>>> usage), the 'used space' probably should include the space taken up=
 by
>>> metadata in addition to data usage (after all, this space is not
>>> available for user data) and free space should report only data spa=
ce
>>> available (because this is what the user is usually expecting). Or,=
 in
>>> other words: the result of 'df' should not concern the user with th=
e
>>> details of raid0/raid1/raid10 used either for data or metadata.
>>
>> I agree that df output sucks... but I've been there before with
>> another filesystem on another OS. =A0The sad fact is df output is
>> too simplistic for the features of modern (last 20 years) systems.
>>
>> There is no way to make df report a value other than "raw space"
>> (which is what btrfs reports today) that will be accurate under
>> all possible raid conditions. =A0The problem is each file can be
>> stored in a different raid (OK not done now, but permitted) and
>> different COW state. =A0That means space_used_per_user_file_block
>> is not constant.
>>
>> So btrfs can only report "best case" or "worst case", but neither
>> will be true.
>>
>> jim
>>
>>
> --
> To unsubscribe from this list: send the line "unsubscribe linux-btrfs=
" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at =A0http://vger.kernel.org/majordomo-info.html
>

I felt like the 'df' value that was returned after applying this patch
was misleading.

Btrfs pre-allocates an amount of space for data, metadata, and system
data, but it doesn't actually use all that space (until the volume
begins to get full).

The 'df' value returned with this patch applied is the amount of space
that is 'allocated' (which is labeled as 'used' in the btrfs-show
output, but really isn't all used yet).

=46or example, when I tested this patch on my volume, the result was
based on 10.00GB of allocated data space.  But only 7.22GB of the
allocated data space had been used.

It highlights an odd problem.  The volume runs out of space based on
the allocated areas.  So in that sense this patch gives a more
accurate representation of when the disk is full.  But from what I've
seen so far, btrfs operates most of the time with a fair amount of fat
between allocated space and actually used space.  So this patch
doesn't give you a good idea of how much space is actually left for
you to use.
--
To unsubscribe from this list: send the line "unsubscribe linux-btrfs" =
in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html