[linuxsh-dev] Signal/Restart changes

[Jeremy S. <jas...@pa...>]

Hi Niibe,

I saw the recent signal.c and entry.S changes; if I understand
correctly, the idea is to make sure that r0 (arg4) is restored correctly
when a syscall is restarted (better than providing random arbitrary
arguments!) and that strace always sees a new task return 0 from
fork().  But after picking them up, I had some problems; here's what
seems to be going on.

In signal.c:restore_sigcontext(), r0 was never copied back to the
exception frame.  This was fine when it was copied to a local variable
returned by sys_sigreturn() to entry.S:syscall_ret(), since it was put
into the exception frame there -- but now that sys_sigreturn() goes
directly to ret_from_syscall, the r0 value is not restored; the
exception frame retains whatever arbitrary r0 value was left by the
signal handler, and when the original syscall resumes (if it was *not*
restarted) then the return value is bogus.  (Similarly, I suspect that
if a process was simply scheduled during a timeslice interrupt and is
rescheduled with a pending signal, the r0 value may be stomped on as
well.)

Simply putting the COPY(regs[0]) in restore_sigcontext() seemed to help
with the some of the problems I saw, but maybe not all.  Does that seem
a reasonable thing to do?  Or are all these changes still considered
experimental, and I just shouldn't be using them yet?

Thanks,
  --Jeremy Siegel
  (js...@mv...)