On Wed, 7 Mar 2012, KOSAKI Motohiro wrote:
> > It's unnecessary to BUG() in situations when a mempolicy has an
> > unsupported mode, it just means that a mode doesn't have complete coverage
> > in all mempolicy functions -- which is an error, but not a fatal error --
> > or that a bit has flipped. Regardless, it's sufficient to warn the user
> > in the kernel log of the situation once and then proceed without crashing
> > the system.
> >
> > This patch converts nearly all the BUG()'s in mm/mempolicy.c to
> > WARN_ON_ONCE(1) and provides the necessary code to return successfully.
>
> I'm sorry. I simple don't understand the purpose of this patch. every
> mem policy syscalls have input check then we can't hit BUG()s in
> mempolicy.c. To me, BUG() is obvious notation than WARN_ON_ONCE().
>
Right, this patch doesn't functionally change anything except it will (1)
continue to warn users when there's a legitimate mempolicy code error by
way of WARN_ON_ONCE() (which is good), just without crashing the machine
unnecessarily and (2) allow the system to stay alive since no mempolicy
error changed by this bug is fatal. We should only be using BUG() when
the side-effects of continuing are fatal; doing WARN_ON_ONCE(1) is
sufficient annotation, I think, that this code should never be reached --
BUG() has no advantage here.
> We usually use WARN_ON_ONCE() for hw drivers code. Because of, the
> warn-on mean "we believe this route never reach, but we afraid there
> is crazy buggy hardware".
>
> And, now BUG() has renreachable() annotation. why don't it work?
>
>
> #define BUG() \
> do { \
> asm volatile("ud2"); \
> unreachable(); \
> } while (0)
>
That's not compiled for CONFIG_BUG=n; such a config fallsback to
include/asm-generic/bug.h which just does
#define BUG() do {} while (0)
because CONFIG_BUG specifically _wants_ to bypass BUG()s and is reasonably
protected by CONFIG_EXPERT.