Re: [take2] Inotify: nested attributes support.

[Pavel Machek <pavel@suse.cz>]

On Wed 2008-11-26 11:29:36, Evgeniy Polyakov wrote:
> On Wed, Nov 26, 2008 at 12:15:38AM -0800, Andrew Morton (akpm@linux-foundation.org) wrote:
> > OK, so we have a super-duper framework which will allow us to add pids
> > (and other things) to inotify messages.
> 
> Yup :)
> 
> > This still doesn't provide a reason for anyone to be interested in the
> > code!  Why do we want pids in inotify messages?
>  
> I actually cared only about myself :)
> I started the thread and implementation, because my application has to
> differentiate IO made by itself and any IO made by system (another
> users, crons, whatever else), inotify did not give me that info, so I
> extended it. As of others: PID/TID may be used by watching applications
> to reduce own load to not process own IO, things like beagle may show
> who actually made changes into the file.

Actually, does the kernel even know who initiated the i/o?

Take two threads, both mapping /etc/something , both of them writing
through the mmap. Kernel sees dirty pages so it writes them back, but
which thread is repsonsible for the write?
 
> > And how does this work give that pids are (no longer) system-wide unique?
> 
> It gets pids from the caller's task_struct (via current), so its data is
> as unique as process calling getpid() or syscall(__NR_gettid).

What happens on mmap()?
									Pavel
-- 
(english) http://www.livejournal.com/~pavelmachek
(cesky, pictures) http://atrey.karlin.mff.cuni.cz/~pavel/picture/horses/blog.html