[PATCH] sandbox: Support signal handling only when requested

[Simon Glass]

Hi Heinrich,

On Sun, 6 Jun 2021 at 15:35, Heinrich Schuchardt <xypron.glpk at gmx.de> wrote:
>
> Am 6. Juni 2021 20:07:31 MESZ schrieb Sean Anderson <seanga2 at gmail.com>:
> >On 6/6/21 1:57 PM, Heinrich Schuchardt wrote:
> >> On 6/6/21 7:52 PM, Sean Anderson wrote:
> >>> On 6/6/21 1:28 PM, Heinrich Schuchardt wrote:
> >>>> On 6/6/21 6:44 PM, Simon Glass wrote:
> >>>>> Hi Heinrich,
> >>>>>
> >>>>> On Mon, 22 Mar 2021 at 18:56, Simon Glass <sjg at chromium.org>
> >wrote:
> >>>>>>
> >>>>>> Hi Heinrich,
> >>>>>>
> >>>>>> On Mon, 22 Mar 2021 at 23:02, Heinrich Schuchardt
> >>>>>> <xypron.glpk at gmx.de> wrote:
> >>>>>>>
> >>>>>>> On 22.03.21 06:21, Simon Glass wrote:
> >>>>>>>> At present if sandbox crashes it prints a message and tries to
> >>>>>>>> exit. But
> >>>>>>>> with the recently introduced signal handler, it often seems to
> >get
> >>>>>>>> stuck
> >>>>>>>> in a loop until the stack overflows:
> >>>>>>>>
> >>>>>>>> Segmentation violation
> >>>>>>>>
> >>>>>>>> Segmentation violation
> >>>>>>>>
> >>>>>>>> Segmentation violation
> >>>>>>>>
> >>>>>>>> Segmentation violation
> >>>>>>>>
> >>>>>>>> Segmentation violation
> >>>>>>>>
> >>>>>>>> Segmentation violation
> >>>>>>>>
> >>>>>>>> Segmentation violation
> >>>>>>>> ...
> >>>>>>>
> >>>>>>> Hello Simon,
> >>>>>>>
> >>>>>>> do you have a reproducible example? I never have seen this.
> >>>>>>
> >>>>>> https://source.denx.de/u-boot/custodians/u-boot-dm/-/jobs/242433
> >>>>>>
> >>>>>> You need to run that commit with pytest though...it does not
> >happen
> >>>>>> when run directly.
> >>>>>>
> >>>>>> BTW this sems to expose some rather nasty bug in dlmalloc or how
> >it is
> >>>>>> used. I notice that as soon as the first test is run, the 'top'
> >value
> >>>>>> in dlmalloc is outside the range of the malloc pool, which seems
> >>>>>> wrong. I wonder if there is something broken with how
> >>>>>> dm_test_pre_run() and dm_test_post_run() work.
> >>>>>>
> >>>>>>>
> >>>>>>> Corrupting gd could cause an endless recursive loop, as these
> >lines
> >>>>>>> follow printing the observed string:
> >>>>>>>
> >>>>>>>          printf("pc = 0x%lx, ", pc);
> >>>>>>>          printf("pc_reloc =0x%lx\n\n", pc - gd->reloc_off);
> >>>>>>
> >>>>>> Yes I suspect printf() is dead.
> >>>>>>
> >>>>>>>
> >>>>>>> If we remove SA_NODEFER from the signal mask in
> >arch/sandbox/cpu/os.c,
> >>>>>>> recursion cannot occur anymore. If a segmentation violation
> >occurs
> >>>>>>> inside the handler it will be delegated to the default handler.
> >>>>>>>
> >>>>>>> Furthermore we could consider removing the signal handler at the
> >start
> >>>>>>> of os_signal_action().
> >>>>>>
> >>>>>> The issue is that if you get a segfault you really don't know if
> >you
> >>>>>> can continue and do anything else.
> >>>>>>
> >>>>>> What is the goal with the signal handler? I don't think the user
> >can
> >>>>>> do anything about it.
> >>>>
> >>>> Hello Simon,
> >>>>
> >>>> the signal handler prints out the crash location and this makes
> >>>> analyzing problems much easier. It proved valuable to me several
> >times.
> >>>
> >>> Can't you just rerun with gdb?
> >>
> >> This would require that the problem is easily reproducible which may
> >not
> >> be the case.
> >
> >Hm, perhaps you could keep track of how many times we've tried to catch
> >a signal, and bail if this is the second time around. E.g. something
> >like
> >
>
> Removing SA_NODEFER from the signal mask will let the OS kick in if an exception occurs in a signal handler.
>
> No counter is needed.

Yes that is correct.

However I am still going to apply this patch, since I would prefer
that the default behaviour is to exit with a signal, rather than
continue. It indicates some sort of error (except if we are running a
strange test), so it seems wrong to try to continue. It may produce
other issues and sandbox is in an unknown state.

I am happy to discuss another way / patch for doing what you need.

Regards,
Simon

[..]