[lkml]   [2009]   [Mar]   [25]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
Messages in this thread
SubjectRe: [RFC PATCH 0/3] sysfs: allow suicide
* Alan Stern <>:
> On Tue, 24 Mar 2009, Alex Chiang wrote:
> > Hi all,
> >
> > This is a refreshed version of the patch series Tejun posted quite a while
> > ago that allowed sysfs attributes to commit suicide directly:
> >
> >
> > The most contentious part is patch 1/3, wherein sysfs abuses the
> > module notifier call chain, and basically prevents all module unloads
> > until suicidal sysfs attributes have completed.
> >
> > This is poison of a different flavor from last time. The earlier version
> > of this series modified the module API and created an interface that
> > allowed anyone to inhibit module unload.
> >
> > This time, only sysfs is allowed to be so... special. Which is a slight
> > improvement, but the question as to whether sysfs should be allowed to
> > do something like this is unresolved.
> I tend to agree with Eric that this feels a little like a band-aid, and
> a more general solution would be preferable. But I don't have one to
> offer, and getting the immediate problems fixed is also important.

Well, getting the sysfs callback off the global workqueue is an
immediate fix that:

- introduces no conceptual change
- fixes the lockdep false positive
- doesn't try to be clever with references

If the consensus here is that this suicide patch series is simply
a band-aid, then I think my other patch will have solved the
problem as much as possible without getting mired in a
conversation about truth and beauty.

> Why change the inhibit-module-unload interface? This new approach
> seems a lot more complicated than needed; a simple rwsem should work
> okay. Exposing it to the entire kernel when only sysfs uses it doesn't
> matter -- there must be plenty of EXPORTed symbols with only one user.

My concern was more the other way around, that exposing a
sledgehammer interface to anyone who wants to inhibit module
unload might not seem like such a wise choice.

I felt that going through the blocking notifier call chain was a
little more proper, in the sense of, "ok well we're going to
allow this inhibit-unload but we know exactly who's doing it".

But that seems irrelevant now.

> Which reminds me... What happens if two different processes write to
> the same suicidal sysfs attribute at the same time?

Good question; I didn't test that with Tejun's patches.

Using the callback mechanism, and a recent patch I wrote that
Greg accepted for 2.6.30, we only allow one in-flight callback
per sysfs attribute/kobject at a time. The loser of the race gets
-EAGAIN while the remove is occurring, and then when the
attribute goes away, gets "file not found" (or something



 \ /
  Last update: 2009-03-26 00:07    [W:0.103 / U:0.924 seconds]
©2003-2018 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site