Re: [PATCH -next v2 0/4] mm: replace mmap_sem for mm->exe_file serialization

From: Oleg Nesterov
Date: Sun Mar 15 2015 - 11:29:07 EST


On 03/15, Davidlohr Bueso wrote:
>
> On Sun, 2015-03-15 at 15:21 +0100, Oleg Nesterov wrote:
> > I didn't even read this version, but honestly I don't like it anyway.
> >
> > I leave the review to Cyrill and Konstantin though, If they like these
> > changes I won't argue.
> >
> > But I simply can't understand why are you doing this.
> >
> >
> >
> > Yes, this code needs cleanups, I agree. Does this series makes it better?
> > To me it doesn't, and the diffstat below shows that it blows the code.
>
> Looking at some of the caller paths now, I have to disagree.

And I believe you are wrong. But let me repeat, I leave this to Cyrill
and Konstantin. Cleanups are always subjective.

> > In fact, to me it complicates this code. For example. Personally I think
> > that MMF_EXE_FILE_CHANGED should die. And currently we can just remove it.
>
> How could you remove this?

Just remove this flag and the test_and_set_bit(MMF_EXE_FILE_CHANGED) check.
Again, this is subjective, but to me it looks ugly. Why do we allow to
change ->exe_file but only once?

> > Not after your patch which adds another dependency.
>
> I don't add another dependency, I just replace the current one.

But you did. If we remove test_and_set_bit(MMF_EXE_FILE_CHANGED)
set_mm_exe_file() becomes racy with your patch. Sure, this is fixable too.

> > Or do you think this is performance improvement? I don't think so. Yes,
> > prctl() abuses mmap_sem, but this not a hot path and the task can only
> > abuse its own ->mm.
>
> I've tried to make it as clear as possible this is a not performance
> patch. I guess I've failed. Let me repeat it again: this is *not*
> performance motivated ;)

OK.

> This kind of things under mmap_sem prevents
> lock breakup.

Could you spell?

> > Hmm. And this series is simply wrong without more changes in audit paths.
> > Unfortunately this is fixable, but let me NACK at least this version ;)
>
> Could you explain this? Are you referring to the audit.c user? If so
> that caller has already been updated.

I do not see these changes in Linus's tree. OK, if those caller's were
already changed somewhere else then unfortunately I can't nack this patch
by technical reasons ;)

But perhaps you should mention that this change depends on other patches
and name them.

> > Speaking of cleanups... IIRC Konstantin suggested to rcuify this pointer
> > and I agree, this looks better than the new lock.
>
> Yes, I can do that in patch 1, but as mentioned, rcu is not really the
> question to me, it's the lock for when we change the exe file, so if
> it's not mmap_sem we'd still need another lock.

Not if we keep MMF_EXE_FILE_CHANGED. See above, we can change it lockless.
And even without MMF_EXE_FILE_CHANGED, we can use xchg().

Oleg.

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/