On Mon, Jun 20, 2011 at 4:03 PM, David Rientjes [off-list ref] wrote:
On Mon, 20 Jun 2011, Mandeep Singh Baines wrote:
quoted
Hi Dave,
I think this change is causing a regression I'm seeing in panic.
Before this change, I'd get a
reboot on panic (we've configured as such).
With this change, my machine gets wedged if the machine is running in
X when the panic occurs.
I traced the code flow to this:
bust_spinlocks(0);
->unblank_screen();
->do_unblank_screen(0);
->vc->vc_sw->con_blank(vc, 0, 0);
->fbcon_blank(vc, 0, 0);
->update_screen(vc);
->redraw_screen(vc, 0);
->vc->vc_sw->con_switch(vc);
->fbcon_switch(vc);
->ops->update_start(info);
->bit_update_start(info);
->fb_pan_display(info, &ops->var);
->info->fbops->fb_pan_display(var, info);
->drm_fb_helper_pan_display(var, info);
->mutex_lock(&dev->mode_config.mutex); *this blocks*
With this change, there is now a lot going on in the panic path. Stuff
that I'm not sure is safe when panicking. In addition to the
mutex_lock, there is also a del_timer_sync()
now happening in the context of panic().
I see this bug with a 2.6.38 kernel but did a quick scan of a newer
kernels and did not see anything that changed in this path so I
suspect its still there.
Reverting this change fixes the regression.
Chris Fowler reports something similar when running 2.6.38 by inducing a
kernel panic via the oom killer -- see
http://marc.info/?l=linux-kernel&m0805985022791. I've added him to the
cc so he can participate in the thread and cherry-pick any fixes (last
status update was that he was going to be trying 2.6.38.8).
One potential fix might be to convert the mutex_lock to a try if
oops_in_progress but
I suspect oops_in_progress checks may be needed in a bunch of other places in
the screen_unblank code path.