Date: Mon, 12 May 2003 14:04:50 +1000 (EST) From: Bruce Evans <bde@zeta.org.au> To: Don Bowman <don@sandvine.com> Cc: freebsd-net@freebsd.org Subject: RE: polling(4) and idle time/cpu usage percentages Message-ID: <20030512133324.C77949@gamplex.bde.org> In-Reply-To: <FE045D4D9F7AED4CBFF1B3B813C8533701B3660D@mail.sandvine.com> References: <FE045D4D9F7AED4CBFF1B3B813C8533701B3660D@mail.sandvine.com>
next in thread | previous in thread | raw e-mail | index | archive | help
On Sun, 11 May 2003, Don Bowman wrote:
> From: Bruce Evans [mailto:bde@zeta.org.au]
> > Did you try my hack?
>
> I tried the hack, as below. The other thing that makes idle
> wildly inaccurate is the symmetric multi-threading on the xeon
> (aka hyperthreading).
The variable needs to be per-cpu for the SMP case. Perhaps there
are other complications for SMP (from having to forward clock interrupts).
Hyperthreading might increase them. Anyway, get the !SMP case working
first.
> Index: kern_clock.c
> ===================================================================
> RCS file: /usr/cvs/src/sys/kern/kern_clock.c,v
> retrieving revision 1.105.2.9.1000.2
> diff -U3 -r1.105.2.9.1000.2 kern_clock.c
> --- kern_clock.c 13 Feb 2003 23:05:58 -0000 1.105.2.9.1000.2
> +++ kern_clock.c 10 May 2003 23:41:47 -0000
> @@ -68,6 +68,7 @@
> #endif
>
> #ifdef DEVICE_POLLING
> +extern int in_polling;
Per-cpu variables are complicated to initialized in RELENG_4. I think
an array with index cpuid can be used with little cost here (cpuid is
a per-cpu global giving the cpu number).
> @@ -550,6 +551,11 @@
> } else if (p != NULL) {
> p->p_sticks++;
> cp_time[CP_SYS]++;
> +#if defined(DEVICE_POLLING)
> + } else if (in_polling) {
Maybe in_polling[cpuid].
> + p->p_sticks++;
Don't incrememnt this. p should always be NULL here.
> + cp_time[CP_SYS]++;
> +#endif
> } else
> cp_time[CP_IDLE]++;
> }
> Index: kern_poll.c
> ===================================================================
> RCS file: /usr/cvs/src/sys/kern/kern_poll.c,v
> retrieving revision 1.2.2.4.1000.1
> diff -U3 -r1.2.2.4.1000.1 kern_poll.c
> --- kern_poll.c 10 Feb 2003 16:49:19 -0000 1.2.2.4.1000.1
> +++ kern_poll.c 10 May 2003 23:37:11 -0000
> @@ -54,6 +54,8 @@
> void ether_poll(int); /* polling while in trap */
> int idle_poll(void); /* poll while in idle loop */
>
> +int in_polling;
> +
> /*
> * Polling support for [network] device drivers.
> *
> @@ -268,11 +270,13 @@
> {
> if (poll_in_idle_loop && poll_handlers > 0) {
> int s = splimp();
> + in_polling = 1;
> enable_intr();
> ether_poll(poll_each_burst);
> disable_intr();
> splx(s);
> vm_page_zero_idle();
> + in_polling = 0;
> return 1;
> } else
> return vm_page_zero_idle();
>
Meybe set the variable for the whole function and name the variable
without using "polling" so that it counts work done by vm_page_zero_idle()
too. The above is better if you just want to count network overhead.
Since the null pointer is apparently never followed in statclock(),
the above apparently doesn't work. I think you aren't actually
calling it for the SMP case. swtch.s has a separate idle loop for
the SMP case in RELENG_4. Only the !SMP case calls the above unless
you have changed it. So the idle time may actually be idle.
Bruce
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20030512133324.C77949>