Re: [PATCH] mm: use VM_BUG_ON*() helpers to dump more debugging info

[Michal Hocko <mhocko@suse.com>]

On Mon 06-04-20 09:40:05, David Hildenbrand wrote:
> On 06.04.20 05:55, qiwuchen55@gmail.com wrote:
> > From: chenqiwu <chenqiwu@xiaomi.com>
> > 
> > This patch use VM_BUG_ON*() helpers instead of simple BUG_ON()
> > in some of the main mm codes. If CONFIG_DEBUG_VM is set, we can
> > get more debugging information when the bug is hit.
> > 
> 
> The "issue" in this context of VM_BUG_ON*() is that, without
> CONFIG_DEBUG_VM,  there won't really be any runtime checks anymore,
> meaning a production system would not stop and BUG_ON() (which would be
> disruptive, but there is a chance to debug this), instead it would
> happily continue to run, eventually messing up something else.
> 
> This is a clear change introduced in this series.
> 
> My gut feeling is that we want to convert this on a per-case basis instead.

This definitely should be done on per-case basis. Many of those bug ons
are historical and they wouldn't be allowed these days. So I would
definitely recommend going through each of them and re-evaluate whether
they are really needed and whether they serve any purpose. A rule of
thumb is that if we can handle the situation more gracefully then the
BUG_ON should be simply dropped. If a verbatim output in a debugging
mode would serve a good purpose then WARN{_ONECE} would serve a good
purpose and if a crash is would help debugging in a DEBUG_VM mode then
use the VM_BUG_ON instead.
-- 
Michal Hocko
SUSE Labs