Re: Bug: Fatal errors result in infinite stream of error messages

[rui wang <ruiv.wang@gmail.com>]

On 12/4/14, Borislav Petkov <bp@alien8.de> wrote:
> On Wed, Dec 03, 2014 at 05:11:49PM +0800, rui wang wrote:
>> The problem is because kdump fails to load a new kernel, and we're
>> executing past crash_kexec() in panic(). And it calls
>> bust_spinlocks(0) which calls into the GPU driver trying to unblank
>> the screen, which eventually calls __schedule() while waiting for a
>> mutex to be released. But we're still in the machine check context.
>> The infinite stream of errors is because there's a for(;;) loop in
>> __mutex_lock_common(), so we enter __schedule() again and again.
>
> Hmm, there's a bust_spinlocks(1) call in mce_panic() for which I have no
> idea what it is for? To stop us from scheduling?
>
> If so, why doesn't it stop us...?
>
> There's also this:
>
> void console_unblank(void)
> {
>         struct console *c;
>
>         /*
>          * console_unblank can no longer be called in interrupt context
> unless
> ====>    * oops_in_progress is set to 1..
>          */
>         if (oops_in_progress) {
>                 if (down_trylock_console_sem() != 0)
>                         return;
>         } else
>                 console_lock();
>

That points to the direction of a potential fix. There are places
under drivers/gpu/drm/ which should do this kind of check in order to
make it work in panic. An existing example is here :

void drm_warn_on_modeset_not_all_locked(struct drm_device *dev)
{
        struct drm_crtc *crtc;

        /* Locking is currently fubar in the panic handler. */
        if (oops_in_progress) <======
                return;

        list_for_each_entry(crtc, &dev->mode_config.crtc_list, head)
                WARN_ON(!drm_modeset_is_locked(&crtc->mutex));

        WARN_ON(!drm_modeset_is_locked(&dev->mode_config.connection_mutex));
        WARN_ON(!mutex_is_locked(&dev->mode_config.mutex));
}

I'll send a patch later.

Thanks
Rui