Re: [RFC] TIF_NOTIFY_RESUME, arch/*/*/*signal*.c and all such

From: Oleg Nesterov <oleg@redhat.com>
To: Al Viro <viro@ZenIV.linux.org.uk>
Cc: Linus Torvalds <torvalds@linux-foundation.org>,
	linux-arch@vger.kernel.org, linux-kernel@vger.kernel.org,
	Russell King <linux@arm.linux.org.uk>, Tejun Heo <tj@kernel.org>,
	Arnd Bergmann <arnd@arndb.de>,
	Roland McGrath <roland@hack.frob.com>
Subject: Re: [RFC] TIF_NOTIFY_RESUME, arch/*/*/*signal*.c and all such
Date: Wed, 25 Apr 2012 19:02:30 +0200	[thread overview]
Message-ID: <20120425170230.GA24977@redhat.com> (raw)
In-Reply-To: <20120425161002.GH6871@ZenIV.linux.org.uk>

On 04/25, Al Viro wrote:
>
> On Wed, Apr 25, 2012 at 05:46:11PM +0200, Oleg Nesterov wrote:
> >
> > 	--- x/arch/x86/kernel/signal.c
> > 	+++ x/arch/x86/kernel/signal.c
> > 	@@ -711,6 +711,13 @@ handle_signal(unsigned long sig, siginfo
> > 				regs->ax = regs->orig_ax;
> > 				regs->ip -= 2;
> > 				break;
> > 	+
> > 	+		case -EINTR:
> > 	+			break;
> > 	+
> > 	+		default:
> > 	+			if (regs->orig_ax == NR_eintr)
> > 	+				regs->ax = NR_eintr;
> > 			}
> > 		}
> > 	
> > 	@@ -791,6 +798,7 @@ static void do_signal(struct pt_regs *re
> > 			case -ERESTARTSYS:
> > 			case -ERESTARTNOINTR:
> > 				regs->ax = regs->orig_ax;
> > 	+			regs->orig_ax = NR_eintr;
> > 				regs->ip -= 2;
> > 				break;
> > 	
> >
> > this ignores ERESTART_RESTARTBLOCK for simplicity.
>
> Ehh...  You do realize that it's that simple only on architectures that
> have syscall number in a register?

No, I do not ;) I simply know nothing about other architectures.

> Take a look at e.g. insert_restart_trampoline()
> in arch/parisc/kernel/signal.c

Will do. Not sure I will be able to undestand it though. And btw I still
didn't look at the signal patches in your tree, I've just returned from
vacation today.

> I'm not sure that SA_RESTART case is actually worth bothering with -

Neither me. But,

> AFAICS, your mechanism won't cover SA_RESTART/SA_RESTART/!SA_RESTART
> combinations anyway,

Confused. The hack above doesn't even try to cover this case (and it is
obviously wrong in ERESTARTSYS/NOINTR case), I only tried to illustrate
the idea.

> 	syscall interrupted
> 		SA_RESTART signal arrives, handler1 entered
> 			!SA_RESTART signal arrives, handler2 entered
> 			handler2 returns
> 		hander1 returns
> 	syscall restarts

And there is another case:

	syscall interrupted, it returns ERESTARTSYS.

	do_signal() sees no signal, restarts the syscall

	!SA_RESTART comes before we return to user-mode

This doesn't really differ from sigsuspend/ERESTARTNOHAND we discussed
before.

Oleg.