Question about Kernel Reporting Sigfaults in <arch>/mm/fault.c

Question about Kernel Reporting Sigfaults in <arch>/mm/fault.c

[Sean Kamath]

Hi.

[All lines are from 2.6.16's latest git repository from kernel.org.]

We recently started noticing error messages showing up in the messages file
every time a user process segfaulted.  Doing some investigation, it turns
out it's only on x86_64 boxen (the other boxes are i386).  We traced this
down to arch/x86_64/mm/fault.c:do_page_fault() (under bad_area_nosemaphore):

478 if (exception_trace && unhandled_signal(tsk, SIGSEGV)) {
479 	printk(
480 	"%s%s[%d]: segfault at %016lx rip %016lx rsp %016lx error %lx\n",
481 		tsk->pid > 1 ? KERN_INFO : KERN_EMERG,
482 		tsk->comm, tsk->pid, address, regs->rip,
483 		regs->rsp, error_code);
484 }

exception_trace is set to 1 (line 296) and not used anywhere else in the file.

arch/i386/mm/fault.c does not have this.
arch/sparc64/mm/fault.c does not have this.

Oddly, sparc/mm/fault.c *does* have (a version of) this, only ifdef'ed out:

319 #if 0
320 	printk("Fault whee %s [%d]: segfaults at %08lx pc=%08lx\n",
321 	tsk->comm, tsk->pid, address, regs->pc);
322 #endif

This leads me to suspect that the segfault reporting in in x86_64 is
vestigial from a time when it was helpful.

The question(s): Is this intentional (to have segfaults reported on x86_64,
and (possibly) nothing else)?  Or was "exception_trace" supposed to be a
flag but never fleshed out?

I admit, it would be nice to be able to toggle on and off segfault
reporting in the kernel (from a system administration point of view, this
is helpful to be able to go back to developers and say 'your program is
crashing a lot' -- something necessary if you are protecting end-users from
a lot of core files . . .) on all platforms.

Sean

Re: Question about Kernel Reporting Sigfaults in <arch>/mm/fault.c

[Andi Kleen]

Sean Kamath <skamath@anim.dreamworks.com> writes:
> 
> exception_trace is set to 1 (line 296) and not used anywhere else in the file.
> 

It can be set using /proc/sys/debug/exception-trace

> The question(s): Is this intentional (to have segfaults reported on x86_64,

Yes.

> I admit, it would be nice to be able to toggle on and off segfault
> reporting in the kernel (from a system administration point of view, this
> is helpful to be able to go back to developers and say 'your program is
> crashing a lot' -- something necessary if you are protecting end-users from
> a lot of core files . . .) on all platforms.

It was originally was added when the port was still young to debug user
land faster, but it keeps people honest and is still very useful.

In fact it found some long standing bugs that were never noticed on i386 with
programs silently segfaulting.

-Andi