Re: [RFC] Attempt to clean up sigsuspend et al

[David Woodhouse <dwmw2@infradead.org>]

On Sun, 2005-11-13 at 21:09 +1100, Paul Mackerras wrote:
> > We don't need to do it on entry; we can do it on exit. If do_signal()
> > sees that the nvgprs weren't saved in the pt_regs it's playing with,
> > then it can set a new TIF_SAVENVPGRS_TO_SIG_FRAME bit. I don't think we
> > even need to store a pointer saying where to put them -- the signal
> > frame will always be at the top of the userspace stack.
> 
> Two problems with that: (1) for *sigsuspend, if do_signal returns 0,
> then we need to go back and keep waiting, and (2) it doesn't handle
> the other syscalls that need the full register set, such as fork,
> clone etc.  Also, for the *sigsuspend case, we need to pass the saved
> signal set to do_signal, not NULL (as the call from the exception exit
> path does).

It doesn't address fork/vfork/clone; certainly. I haven't really looked
at that code path at all yet. It was the *suspend() functions which
offended me, mostly because I need similar behaviour for pselect() and
ppoll() and I didn't want to add to the mess.

But I think other things you mention are OK -- we'd still call
do_signal() from *sigsuspend() just as we do at the moment, with the
same loop. It's just that do_signal() would notice that the nvgprs
aren't saved, and would set the TIF_SAVENVPGRS_TO_SIG_FRAME bit. ..

	if (!regs->trap & 1)
		set_thread_flag(TIF_SAVENVGPRS_SIG);

We end up returning to userspace via entry.S, and the code there notices
the SAVENVGPRS bit and saves _just_ r14-r31 into the signal frame. It
doesn't have to call do_signal() again, so doesn't have to get the right
saved signal mask when doing so.

(Actually in the case of swapcontext() we don't know that the nvgprs
should be saved in a signal frame at the top of the userspace stack; the
context will be elsewhere. We probably do need to store a pointer in the
thread_info, but that's OK too)

-- 
dwmw2