svn commit: r238907 - projects/calloutng/sys/kern

[Attilio Rao]

On Sat, Sep 8, 2012 at 4:38 PM, Konstantin Belousov <kostikbel at gmail.com> wrote:
> On Fri, Sep 07, 2012 at 11:35:15PM +0100, Attilio Rao wrote:
>> On Mon, Jul 30, 2012 at 9:39 PM, Attilio Rao <attilio at freebsd.org> wrote:
>> > On Mon, Jul 30, 2012 at 4:31 PM, Andriy Gapon <avg at freebsd.org> wrote:
>> >> on 30/07/2012 18:04 Attilio Rao said the following:
>> >>> On 7/30/12, Andriy Gapon <avg at freebsd.org> wrote:
>> >>>> on 30/07/2012 17:56 Attilio Rao said the following:
>> >>>>> More explicitly, I think such combination TDP_NOSLEEPING +
>> >>>>> TDP_NOBLOCKING (name invented) should be set on entering the interrupt
>> >>>>> context, not only related to this part of callouts. This would be a
>> >>>>> very good help for catching buggy situations.
>> >>>>
>> >>>> Something very tangential.  I think it would also be nice to check if a
>> >>>> thread has
>> >>>> any(?) locks held when returning to userland.
>> >>>
>> >>> This happens already for INVARIANTS case, with td_locks counters.
>> >>> In the !INVARIANTS case, this doesn't happen because you don't want to
>> >>> add the burden to bump td_locks for the fast case and I think it is a
>> >>> good approach.
>> >>
>> >> Ah, I missed that, thank you.
>> >> BTW, it seems that td_locks is checked twice in normal syscallret() path: once in
>> >> syscallret() itself and then in userret().  On this note, would it make sense to
>> >> move the whole nine yards of asserts from syscallret() to userret()?
>> >> I mean it might make sense to have those checks (td_critnest, td_pflags) in other
>> >> paths to userland.
>> >
>> > Nice catch.
>> > The checks were added to syscallret() in r208453. While this is fine,
>> > I think that putting them in userret() may give them more exposure and
>> > cover also cases like traps which are not covered right now.
>> > If you want to make a patch that moves these conditions in userret()
>> > I'd be in favor of it.
>>
>> More specifically, what do you think about this patch?:
>> http://www.freebsd.org/~attilio/userret_diag.patch
>>
>> Of course I moved the XEN par too before the checks.
>> The patch survived to few consecutive and parallel buildworlds, FWIW.
>
> At least in fork_return(), the last assert which checks that Giant is not
> held, is no longer needed.

Actually, this is unnecessary also with -CURRENT stock code today,
because userret() already checks for td_locks. And it seems
fork_return() is not the only function where this happens, as trap()
did this too on x86 and possibly all the other architectures grew it
with cut&paste. Possibly we need this further, separate, patch before
userred_diag:
http://www.freebsd.org/~attilio/userret_nogiant.patch

> I had similar thought from the time when I added TDP_NOFAULTING check to the
> syscallret(), but the loss of the syscall name in the panic output always
> stopped me.

I think in case of a lock/td_pinned/td_critnest/etc. leak, having the
syscall number in the panic message won't change anything as you will
likely need a coredump and possibly instrument your kernel with
further debugging, etc. to see what's going on.

Attilio


-- 
Peace can only be achieved by understanding - A. Einstein