Martin Pelikan <martin.peli...@gmail.com> writes:
> Hi!
>
> Recently, I had to unleash my crappy Atom and decided to run OpenBSD on it.
> Then I became irritated with the performance of most X11 applications, and
> started poking around:
> - libfreetype.so is quite common
> - it does setjmp (disguised as ft_setjmp) quite a lot
> - setjmp needs to do sigprocmask
> - sigprocmask (I presume) needs the kernel lock.
If you're right that atomic_{clear,set}bits_int is correct and
sufficient and actually faster, then all dynamic executables would
benefit from this speedup (sigprocmask is used in ld.so(1)).
> sigprocmask(2) is quite simple actually, but wasn't marked as NOLOCK,
> so I made it so. And then I made a pointless test to demonstrate my point,
> and ran it on my pointless Atom N270 (two threads, one core from 1980s)
> as well as in VMware, which is the only 4-core amd64 machine I can get
> on right now. (did I say anything about using virtual machines to measure
> things?)
>
> #include <setjmp.h>
>
> int
> main(void)
> {
> jmp_buf env;
> unsigned i;
>
> /* on amd64 this was too slow, so 20M was used instead */
> for (i = 0; i < 4000000; ++i) {
> setjmp(env);
> }
> return (0);
> }
>
> Before:
> Atom, i386, single process, 0m3.80s real 0m1.15s user 0m2.50s system
> Atom, i386, two processes, 0m4.81s real 0m1.44s user 0m3.09s system
>
> amd64, single process, i<20M, 0m3.35s real 0m0.65s user 0m2.23s system
> amd64, two processes, i<20M, 0m4.48s real 0m0.66s user 0m3.37s system
>
> After:
> Atom, i386, single process, 0m2.93s real 0m0.95s user 0m1.35s system
> Atom, i386, two processes, 0m3.67s real 0m1.36s user 0m2.15s system
>
> amd64, single process, i<20M, 0m2.14s real 0m0.63s user 0m1.12s system
> amd64, two processes, i<20M, 0m2.57s real 0m0.57s user 0m1.55s system
>
>
> If anyone has a real test with some real-world application using lots
> of setjmps, I'd like to see it. What'd I like to see most, is if does
> it work on other architectures and if it is indeed correct. Although
> the numbers suggest quite an improvement, bear in mind this kind of
> load does not at all resemble any software actually doing useful stuff.
>
> Any comments?
> --
> Martin Pelikan
>
>
> Index: kern/init_sysent.c
> ===================================================================
> RCS file: /cvs/src/sys/kern/init_sysent.c,v
> retrieving revision 1.145
> diff -u -p -r1.145 init_sysent.c
> --- kern/init_sysent.c 9 Jun 2013 13:10:27 -0000 1.145
> +++ kern/init_sysent.c 19 Jun 2013 09:49:57 -0000
> @@ -136,7 +136,7 @@ struct sysent sysent[] = {
> sys_sigaction }, /* 46 = sigaction */
> { 0, 0, 0,
> sys_getgid }, /* 47 = getgid */
> - { 2, s(struct sys_sigprocmask_args), 0,
> + { 2, s(struct sys_sigprocmask_args), SY_NOLOCK | 0,
> sys_sigprocmask }, /* 48 = sigprocmask */
> { 2, s(struct sys_getlogin_args), 0,
> sys_getlogin }, /* 49 = getlogin */
> Index: kern/kern_sig.c
> ===================================================================
> RCS file: /cvs/src/sys/kern/kern_sig.c,v
> retrieving revision 1.152
> diff -u -p -r1.152 kern_sig.c
> --- kern/kern_sig.c 1 Jun 2013 04:05:26 -0000 1.152
> +++ kern/kern_sig.c 19 Jun 2013 09:49:57 -0000
> @@ -429,28 +429,25 @@ sys_sigprocmask(struct proc *p, void *v,
> syscallarg(sigset_t) mask;
> } */ *uap = v;
> int error = 0;
> - int s;
> sigset_t mask;
>
> *retval = p->p_sigmask;
> - mask = SCARG(uap, mask);
> - s = splhigh();
> + mask = SCARG(uap, mask) &~ sigcantmask;
>
> switch (SCARG(uap, how)) {
> case SIG_BLOCK:
> - p->p_sigmask |= mask &~ sigcantmask;
> + atomic_setbits_int(&p->p_sigmask, mask);
> break;
> case SIG_UNBLOCK:
> - p->p_sigmask &= ~mask;
> + atomic_clearbits_int(&p->p_sigmask, mask);
> break;
> case SIG_SETMASK:
> - p->p_sigmask = mask &~ sigcantmask;
> + p->p_sigmask = mask;
> break;
> default:
> error = EINVAL;
> break;
> }
> - splx(s);
> return (error);
> }
>
> Index: kern/syscalls.master
> ===================================================================
> RCS file: /cvs/src/sys/kern/syscalls.master,v
> retrieving revision 1.133
> diff -u -p -r1.133 syscalls.master
> --- kern/syscalls.master 9 Jun 2013 13:10:19 -0000 1.133
> +++ kern/syscalls.master 19 Jun 2013 09:49:57 -0000
> @@ -123,7 +123,7 @@
> const struct sigaction *nsa, \
> struct sigaction *osa); }
> 47 STD { gid_t sys_getgid(void); }
> -48 STD { int sys_sigprocmask(int how, sigset_t mask); }
> +48 STD NOLOCK { int sys_sigprocmask(int how, sigset_t mask); }
> 49 STD { int sys_getlogin(char *namebuf, u_int namelen); }
> 50 STD { int sys_setlogin(const char *namebuf); }
> #ifdef ACCOUNTING
>
--
Jérémie Courrèges-Anglas
PGP Key fingerprint: 61DB D9A0 00A4 67CF 2A90 8961 6191 8FBF 06A1 1494
