2005-02-09 20:09:03

by Chris Friesen

[permalink] [raw]
Subject: [BUG] in copy_siginfo_to_user32 on ppc64 (and others?) in 2.6.9/2.6.10

I found a bug which has since been fixed, but I'm hoping to save others
the problems that I had tracking it down.

It was fairly confusing--the information in the siginfo_t struct was
different based on whether I used a signal handler in the regular way,
or blocked the signal and retrieved the information using sigtimedwait().

After much instrumentation of the kernel, I tracked it down.

Until recently (Jan 5), ppc64 had its own version of
compat_sys_rt_sigtimedwait, which simply called sys_rt_sigtimedwait()
then copied the results to the userspace struct using
copy_siginfo_to_user32().

Unfortunately, sys_rt_sigtimedwait() only copies the lower 16 bits of
si_code, and the ppc64 version of copy_siginfo_to_user32() keyed on the
upper 16 bits to decide what information to copy. Thus, it always ended
up in the default case of the switch statement, and only ever copied
si_pid and si_uid.

Oops.

Chris