Re: ipv4/tcp.c:4234:1: error: the frame size of 1152 bytes is larger than 1024 bytes [-Werror=frame-larger-than=]

[Brendan Higgins]

On Wed, Sep 8, 2021 at 10:16 AM Shuah Khan <skhan@xxxxxxxxxxxxxxxxxxx> wrote:
>
> On 9/8/21 11:05 AM, Arnd Bergmann wrote:
> > On Wed, Sep 8, 2021 at 4:12 PM Shuah Khan <skhan@xxxxxxxxxxxxxxxxxxx> wrote:
> >> On 9/7/21 5:14 PM, Linus Torvalds wrote:
> >>> The KUNIT macros create all these individually reasonably small
> >>> initialized structures on stack, and when you have more than a small
> >>> handful of them the KUNIT infrastructure just makes the stack space
> >>> explode. Sometimes the compiler will be able to re-use the stack
> >>> slots, but it seems to be an iffy proposition to depend on it - it
> >>> seems to be a combination of luck and various config options.
> >>>
> >>
> >> I have been concerned about these macros creeping in for a while.
> >> I will take a closer look and work with Brendan to come with a plan
> >> to address it.
> >
> > I've previously sent patches to turn off the structleak plugin for
> > any kunit test file to work around this, but only a few of those patches
> > got merged and new files have been added since. It would
> > definitely help to come up with a proper fix, but my structleak-disable
> > hack should be sufficient as a quick fix.
> >
>
> Looks like these are RFC patches and the discussion went cold. Let's pick
> this back up and we can make progress.
>
> https://lore.kernel.org/lkml/CAFd5g45+JqKDqewqz2oZtnphA-_0w62FdSTkRs43K_NJUgnLBg@xxxxxxxxxxxxxx/

I can try to get the patch reapplying and send it out (I just figured
that Arnd or Kees would want to send it out :-)  since it was your
idea).

I definitely agree that in the cases where KUnit is not actually
contributing to blowing the stack - struct leak just thinks it is,
this is fine; however, it sounds like Linus' concerns with KUnit's
macros go deeper than this. Arnd, I think you sketched out a way to
make the KUNIT_* macros take up less space, but after some
investigation we found that it was pretty inflexible.

Ideally test cases should never get big enough for KUNIT_* macros to
be a problem (when they do it is usually an indication that your test
case is trying to do too many things); nevertheless, we are still in
this situation.

I think I will need to dust off some cobwebs out of my brain to
remember why I didn't like the idea of making the KUNIT_* macros take
up less stack space.