Re: udevd is killing file write performance.

[John McCutchan <john@johnmccutchan.com>]

On Mon, 2006-27-02 at 21:11 +1100, Nick Piggin wrote:
> John McCutchan wrote:
> > On Fri, 2006-24-02 at 18:07 +1100, Nick Piggin wrote:
> 
> >>I saw this problem when testing my lockless pagecache a while back.
> >>
> >>Attached is a first implementation of what was my idea then of how
> >>to solve it... note it is pretty rough and I never got around to doing
> >>much testing of it.
> >>
> >>Basically: moves work out of inotify event time and to inotify attach
> >>/detach time while staying out of the core VFS.
> > 
> > 
> > 
> > This looks really good. There might be some corner cases but it looks
> > like it will solve this problem nicely.
> > 
> 
> Thanks. You should see I sent a new version which fixes several bugs
> and cleans up the code a bit.
> 

Yeah, it looks good. I haven't had time to test it myself but nothing
jumps out at as being wrong. I can only say that about the code that
touches inotify -- the rest of the VFS someone else will need to comment
on.

> There might be some areas of potential problems:
> - creating and deleting watches on directories with many entries will
>    take a long time. Is anyone likely to be creating and destroying
>    these things at a very high frequency? Probably nobody cares except
>    it might twist some real-time knickers.
> 

That's not a typical inotify usage pattern. Typically a watch is created
and left until the directory is deleted, or the application closes.

> - concurrent operations in the same watched directory will incur the
>    same scalability penalty. I think this is basically a non-issue since
>    the sheer number of events coming out will likely be a bigger problem.
>    Doctor, it hurts when I do this.
> 

Again, Yeah, I don't think we need to worry.

-- 
John McCutchan <john@johnmccutchan.com>