Re: [RFC v3 11/13] vfs: add 3 new ioctl interfaces

[Dave Chinner <david@fromorbit.com>]

On Wed, Oct 10, 2012 at 06:07:33PM +0800, zwu.kernel@gmail.com wrote:
> From: Zhi Yong Wu <wuzhy@linux.vnet.ibm.com>
> 
>   FS_IOC_GET_HEAT_INFO: return a struct containing the various
> metrics collected in btrfs_freq_data structs, and also return a

I think you mean hot_freq_data :P

> calculated data temperature based on those metrics. Optionally, retrieve
> the temperature from the hot data hash list instead of recalculating it.

To get the heat info for a specific file you have to know what file
you want to get that info for, right?  I can see the usefulness of
asking for the heat data on a specific file, but how do you find the
hot files in the first place? i.e. the big question the user
interface needs to answer is "what files are hot?".

Once userspace knows what the hottest files are, it can open them
and query the data via the above ioctl, but expecting userspace to
iterate millions of inodes in a filesystem to find hot files is very
inefficient.

FWIW, if you were to return file handles to the hottest files, then
the application could open and query them without even needing to
know the path name to them. This woul dbe exceedingly useful for
defragmentation programs, especially as that is the way xfs_fsr
already operates on candidate files.(*)

IOWs, sometimes the pathname is irrelevant to the operations that
applications want to perform - all they care about having an
efficient method of finding the inode they want and getting a file
descriptor that points to the file. Given the heat map info fits
right in to the sort of operations defrag and data mover tools
already do, it kind of makes sense to optimise the interface towards
those uses....

(*) i.e. finds them via bulkstat which returns handle information
along with all the other inode data, then opens the file by handle
to do the defrag work....

>   FS_IOC_GET_HEAT_OPTS: return an integer representing the current
> state of hot data tracking and migration:
> 
> 0 = do nothing
> 1 = track frequency of access
> 
>   FS_IOC_SET_HEAT_OPTS: change the state of hot data tracking and
> migration, as described above.

I can't see how this is a manageable interface. It is not
persistent, so after every filesystem mount you'd have to set the
flag on all your inodes again. Hence, for the moment, I'd suggest
that dropping per-inode tracking control until all the core issues
are sorted out....

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com