Re: [PATCH RFC] btrfs: introduce procfs interface for the device list

[Chris Mason <clm@fb.com>]

On Wed, Oct 1, 2014 at 3:41 AM, Anand Jain <Anand.Jain@oracle.com> 
wrote:
> 
> Hi Chris,
> 
>  Thanks for commenting. Some clarifying comments as below.
> 
> 
> On 30/09/2014 22:23, Chris Mason wrote:
>> 
>> 
>> On Mon, Sep 29, 2014 at 1:09 AM, Anand Jain <anand.jain@oracle.com> 
>> wrote:
>>> From: Anand Jain <Anand.Jain@oracle.com>
>>> 
>>> (added RFC prefix to the patch header)
>>> (as of now just an experimental interface)
>>> 
>>> This patch introduces profs interface /proc/fs/btrfs/devlist,
>>> which as of now exports all the members of kernel fs_devices.
>>> 
>>> The current /sys/fs/btrfs interface works when the fs is
>>> mounted, and is on the file directory hierarchy and also has
>>> the sysfs limitation max output of U64 per file.
>>> 
>>> Here btrfs procfs uses seq_file to export all the members of
>>> fs_devices. Also shows the contents when device is not mounted,
>>> but have registered with btrfs kernel (useful as an alternative
>>> to buggy ready ioctl)
>>> 
>>> An attempt is made to follow the some standard file format
>>> output such as ini. So that a simple warper python script will
>>> provide end user useful interfaces.
>>> 
>>> Further planning to add few more members to the interface such as
>>> group profile info. The long term idea is to make procfs
>>> interface a onestop btrfs application interface for the device and
>>> fs info from the kernel, where a simple python script can make
>>> use of it.
>> 
>> Hi Anand,
>> 
>> We're going to have a really hard time getting a new proc interface
>> merged in, and after we've recently fixed up all (most?) of our sysfs
>> races, I'd rather not have to do it all over again with /proc.
> 
>  This does not use fsid/devid based file-directory. So races as were
>  in sysfs implementation does not apply here. (But there are 
> opportunity
>  to optimize the code at the place mentioned in the code as todo).

Right, proc has different races ;)  Again the bar for new interfaces in 
proc is really very high.  It's not the direction the rest of the 
kernel is using.

> 
> 
>> I know
>> the lack of a seq interface is a difficult compromise to make in 
>> sysfs,
>> but at this point I think we're stuck with it.  Which specific part 
>> do
>> you hope to improve by dumping more information out in a single file?
> 
>  Since its a single file and dumping most of the members of fs_devices
>  we would ensure the interface will remain unchanged for a long time
>  and helps debugging. This is hard to do when we layout files per
>  parameter value.
> 
>  Less clutter. But needs python script abstraction to provide what
>  user want. Better than using ioctls.
> 
>  file-parameter-layout might introduce races. So here there is no file
>  parameter layout, its just one file /proc/fs/btrfs/devlist, provides
>  an interface which is compatible with parser such as python
>  configparser, with which application can organize it using a simple
>  script.
> 
> 
> Further,
>  This also exports all registered devices which may not be mounted.
>  (sysfs implementation does not).

For these features, we need to work within the sysfs and udev 
frameworks.  It will integrate better with the direction the distros 
are using for management in general.  I really understand that in some 
ways the proc interface would be easier to write and easier to use, but 
this is one of those times that consistency with the rest of the kernel 
comes first.

Thanks again for the time you've spent improving the device management 
side of things.  For now, sysfs and udev are the best choices overall.

-chris