Re: [BUG 2.6.20-rc2-mm1] init segfaults whenCONFIG_PROFILE_LIKELY=y

[Randy Dunlap]

On Sun, 31 Dec 2006 13:11:26 -0800 Daniel Walker wrote:

> On Sun, 2006-12-31 at 12:43 -0800, Randy Dunlap wrote:
> > On Sun, 31 Dec 2006 11:45:09 -0800 Daniel Walker wrote:
> > 
> > > On Sun, 2006-12-31 at 23:04 +0800, Fengguang Wu wrote:
> > > > Hi,
> > > > 
> > > > The following messages keeps popping up when CONFIG_PROFILE_LIKELY=y:
> > > > 
> > > > init[1]: segfault at ffffffff8118c110 rip ffffffff8118c110 rsp 00007fff9a9d14d8 error 15
> > > > init[1]: segfault at ffffffff8118c110 rip ffffffff8118c110 rsp 00007fff9a9d14d8 error 15
> > > > init[1]: segfault at ffffffff8118c110 rip ffffffff8118c110 rsp 00007fff9a9d14d8 error 15
> > > > init[1]: segfault at ffffffff8118c110 rip ffffffff8118c110 rsp 00007fff9a9d14d8 error 15
> > > > init[1]: segfault at ffffffff8118c110 rip ffffffff8118c110 rsp 00007fff9a9d14d8 error 15
> > > > init[1]: segfault at ffffffff8118c110 rip ffffffff8118c110 rsp 00007fff9a9d14d8 error 15
> > > > init[1]: segfault at ffffffff8118c110 rip ffffffff8118c110 rsp 00007fff9a9d14d8 error 15
> > > > init[1]: segfault at ffffffff8118c110 rip ffffffff8118c110 rsp 00007fff9a9d14d8 error 15
> > > > 
> > > 
> > > 
> > > Does this seem like an appropriate solution? This just reconstitutes
> > > Ingo's patch by removing the unlikely calls that got added recently. 
> > 
> > How does this fix the problem?  (if it does)
> > What is the real cause of the problem?
> 
> Well I tested it so I sure hope it fixes it (unless I've gone mad). I
> guess we can wait for Fengguang to test it tho.

Yes, it fixed it in my testing also.

> > > Maybe a comment into vsyscall.c that says to stay away from all macro's
> > > and possible debug code that could be added might be helpful ?
> > 
> > Why?
> 
> I don't know very much about vsyscalls, but from what I've read they
> actually reside in userspace. So with and "unlikely" added into that
> code, and profiling on, you will end up calling do_check_likely() which
> is in kernel space that's how the segfault happens. 
> 
> I imagine this goes for all debugging in kernel space, you can't add it
> into a vsyscall. That's my reasoning behind adding a comment.

OK, thanks.  The first explanation was a bit lacking IMO,
but this one begins to help.

---
~Randy
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/