Crash with bus error on calling vips_rank

[akash-akya]

Program cashes when vips_rank called successively on same image.

Program exit with error:

[2]    60540 bus error  ./a.out

To Reproduce

Sample image: https://raw.githubusercontent.com/elixir-image/image/main/test/support/images/tulips.png

#include <vips/vips.h>

int
main (int argc, char **argv)
{
  if (VIPS_INIT (argv[0]))
    return (-1);

  VipsImage *img;
  size_t size;

  img = vips_image_new_from_file ("tulips.png", NULL);

  for (uint i = 0; i < 500; i++) {
    if (vips_rank (img, &img, 3, 3, 7, NULL) != 0) {
      vips_error_exit (NULL);
    }
  }

  (void)vips_image_write_to_memory (img, &size);

  return (0);
}

• gcc -g -Wall seg_fault.c `pkg-config vips --cflags --libs`
• run ./a.out

Unfortunately, I couldn't manage to get core-dump as I updated my machine recently, and I have to configure it to produce one.

Environment
* OS: MacOS 14.2.1
* Vips: vips-8.15.0

[jcupitt]

Hi @akash-akya,

Yes, you'll see crashes with very long pipelines, this is expected behaviour caused by a C stack overflow.

[akash-akya]

I see. I guess there is nothing we can do to improve the visibility, like showing warning?

Closing the issue the then. Thanks for the quick response!

[jcupitt]

It's hard to do well.

You need to get the stack pointer on program startup, then on each recursion, libvips would need to get the stack pointer again, compare it to the stack size (which it would have to measure in some platform-independent way) and fail if it was close to the limit. This means this kind of thing is best done at the application level.

nip2 tries to address C stack limits with a recursion checker, but I think (from memory!) it just assumes the C stack is 2MB. You'd probably need something better.

Maybe vix could track pipeline depth and have a parameter somewhere the user can adjust with a maximum safe value?

[akash-akya]

Ah ok, understood.

Maybe vix could track pipeline depth and have a parameter somewhere the user can adjust with a maximum safe value?

Yes, that seems possible, I'll look into it. Thanks for the pointer!

[jcupitt]

We might need to add something like int vips_image_get_pipeline_depth(VipsImage*) -- I think this might be hard to track reliably in vix.

Another possible issue (thinking about it again) is that the pipeline depth that you can use depends on how full the C stack is when you trigger evaluation, which in turn depends on the vix code that's executing.

Maybe these checks would not be accurate enough to be useful :(