Re: [PATCH] vfs: Add a trace point in the mark_inode_dirty function

[Ingo Molnar <mingo@elte.hu>]

* Theodore Tso <tytso@mit.edu> wrote:

> On Wed, Nov 11, 2009 at 08:45:42AM +0100, Ingo Molnar wrote:
> > Without an inode->vfs-name lookup/matching service it's of limited 
> > utility though to developers and users. So inode numbers are fine (as 
> > nicely unique physical identifiers)- as long as corresponding vfs name 
> > string is available too.
> 
> Inode numbers are quite usable for me; but I'm not afraid to do
> 
>       debugfs /dev/sdb -R "ncheck 12345"
> 
> :-)
> 
> If you really want to avoid that, one relatively lightweight thing we 
> could do, which would avoid needing to dump the entire pathname out, 
> would be to print out the triple (devno, dir_ino, file_ino), and then 
> provide a privileged syscall which translates this to a user-visible 
> pathname.  It won't be necessarily the pathname which the user used to 
> open the file (since there might be links, and bind mounts, et. al), 
> but if the goal is to give one of the user-friendly names of the inode 
> (as opposed to _the_ pathname used to open the file), it's quite 
> sufficient.

Hm, why add a new syscall to retrieve the name we already had when the 
event happened?

Also, why add a new syscall to retrieve something that might not exist 
anymore? (the VFS namespace is quite dynamic - post-processing to 
retrieve names is fundamentally racy)

What matters most for analysis is the 'name of the moment' - the thing 
that the app used at that point.

Arjan isnt doing this just randomly, he's one of the few people trying 
to speed up Linux booting - and this is the info he finds useful. We 
should give that information in a reasonable way, and the tracepoint he 
proposed looks pretty reasonable.

	Ingo