From: eeh@netbsd.org
Subject: Re: kernel stack overflow detection
Date: 26 Jun 2002 17:14:16 -0000

> | there're 2 things. (sorry for confusing)
> |
> | KSTACK_CHECK_MAGIC is (intended to be) MI.
> | KSTACK_CHECK_DR0 is i386-specific.
> 
> So, does KSTACK_CHECK_MAGIC work without any MD changes?
> If not, what MD changes are needed?

if kernel stack range on the arch can be calculated by
macros on sys/proc.h (KSTACK_{END,SIZE} in my patch),
no MD changes are needed.
if it can't, you need to override these macros.

> Why do you also need KSTACK_CHECK_DR0 if KSTACK_CHECK_MAGIC
> should do the job in a MI manner?
> 
> Or, if you need MD code and it can all be done in MD code,
> why is there a need for an MI version, which you have admitted
> does have noticeable performance impact?  Do you expect anyone
> to actually use the MI code?

MI version produce messages like following:
warning: kernel stack left %d bytes(pid %u)

MD version traps exactly when overflow occurs,
so trace on ddb is more useful.

> sparc already has kernel stack redzone code.
> sparc64 already has kernel stack redzone code.
> vax already has kernel stack redzone code.
> hppa already has stack redzine code.

i made a patch for vax and ok'ed by ragge.
i have no idea about sparc, sparc64 and hppa.
i think HPPA_REDZONE isn't compilable..
(lack of ",")
        /* Create the kernel stack red zone. */
        pmap_redzone((vaddr_t)p->p_addr + HPPA_REDZONE
                (vaddr_t)p->p_addr + USPACE, 1);

> Do we really need a MI version?

at least, i want.

---
YAMAMOTO Takashi<yamt@mwd.biglobe.ne.jp>