1999-07-22 22:13:36 +04:00
|
|
|
/* $NetBSD: kern_synch.c,v 1.60 1999/07/22 18:13:37 thorpej Exp $ */
|
1994-06-29 10:29:24 +04:00
|
|
|
|
|
|
|
/*-
|
|
|
|
* Copyright (c) 1982, 1986, 1990, 1991, 1993
|
|
|
|
* The Regents of the University of California. All rights reserved.
|
|
|
|
* (c) UNIX System Laboratories, Inc.
|
|
|
|
* All or some portions of this file are derived from material licensed
|
|
|
|
* to the University of California by American Telephone and Telegraph
|
|
|
|
* Co. or Unix System Laboratories, Inc. and are reproduced herein with
|
|
|
|
* the permission of UNIX System Laboratories, Inc.
|
|
|
|
*
|
|
|
|
* Redistribution and use in source and binary forms, with or without
|
|
|
|
* modification, are permitted provided that the following conditions
|
|
|
|
* are met:
|
|
|
|
* 1. Redistributions of source code must retain the above copyright
|
|
|
|
* notice, this list of conditions and the following disclaimer.
|
|
|
|
* 2. Redistributions in binary form must reproduce the above copyright
|
|
|
|
* notice, this list of conditions and the following disclaimer in the
|
|
|
|
* documentation and/or other materials provided with the distribution.
|
|
|
|
* 3. All advertising materials mentioning features or use of this software
|
|
|
|
* must display the following acknowledgement:
|
|
|
|
* This product includes software developed by the University of
|
|
|
|
* California, Berkeley and its contributors.
|
|
|
|
* 4. Neither the name of the University nor the names of its contributors
|
|
|
|
* may be used to endorse or promote products derived from this software
|
|
|
|
* without specific prior written permission.
|
|
|
|
*
|
|
|
|
* THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
|
|
|
|
* ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
|
|
|
|
* IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
|
|
|
|
* ARE DISCLAIMED. IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
|
|
|
|
* FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
|
|
|
|
* DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
|
|
|
|
* OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
|
|
|
|
* HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
|
|
|
|
* LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
|
|
|
|
* OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
|
|
|
|
* SUCH DAMAGE.
|
|
|
|
*
|
1998-03-01 05:20:01 +03:00
|
|
|
* @(#)kern_synch.c 8.9 (Berkeley) 5/19/95
|
1994-06-29 10:29:24 +04:00
|
|
|
*/
|
|
|
|
|
1998-07-05 02:18:13 +04:00
|
|
|
#include "opt_ddb.h"
|
1998-06-26 01:17:15 +04:00
|
|
|
#include "opt_ktrace.h"
|
1998-02-10 17:08:44 +03:00
|
|
|
|
1994-06-29 10:29:24 +04:00
|
|
|
#include <sys/param.h>
|
|
|
|
#include <sys/systm.h>
|
|
|
|
#include <sys/proc.h>
|
|
|
|
#include <sys/kernel.h>
|
|
|
|
#include <sys/buf.h>
|
|
|
|
#include <sys/signalvar.h>
|
|
|
|
#include <sys/resourcevar.h>
|
1996-02-04 05:15:01 +03:00
|
|
|
#include <vm/vm.h>
|
Scheduler bug fixes and reorganization
* fix the ancient nice(1) bug, where nice +20 processes incorrectly
steal 10 - 20% of the CPU, (or even more depending on load average)
* provide a new schedclk() mechanism at a new clock at schedhz, so high
platform hz values don't cause nice +0 processes to look like they are
niced
* change the algorithm slightly, and reorganize the code a lot
* fix percent-CPU calculation bugs, and eliminate some no-op code
=== nice bug === Correctly divide the scheduler queues between niced and
compute-bound processes. The current nice weight of two (sort of, see
`algorithm change' below) neatly divides the USRPRI queues in half; this
should have been used to clip p_estcpu, instead of UCHAR_MAX. Besides
being the wrong amount, clipping an unsigned char to UCHAR_MAX is a no-op,
and it was done after decay_cpu() which can only _reduce_ the value. It
has to be kept <= NICE_WEIGHT * PRIO_MAX - PPQ or processes can
scheduler-penalize themselves onto the same queue as nice +20 processes.
(Or even a higher one.)
=== New schedclk() mechansism === Some platforms should be cutting down
stathz before hitting the scheduler, since the scheduler algorithm only
works right in the vicinity of 64 Hz. Rather than prescale hz, then scale
back and forth by 4 every time p_estcpu is touched (each occurance an
abstraction violation), use p_estcpu without scaling and require schedhz
to be generated directly at the right frequency. Use a default stathz (well,
actually, profhz) / 4, so nothing changes unless a platform defines schedhz
and a new clock. Define these for alpha, where hz==1024, and nice was
totally broke.
=== Algorithm change === The nice value used to be added to the
exponentially-decayed scheduler history value p_estcpu, in _addition_ to
be incorporated directly (with greater wieght) into the priority calculation.
At first glance, it appears to be a pointless increase of 1/8 the nice
effect (pri = p_estcpu/4 + nice*2), but it's actually at least 3x that
because it will ramp up linearly but be decayed only exponentially, thus
converging to an additional .75 nice for a loadaverage of one. I killed
this, it makes the behavior hard to control, almost impossible to analyze,
and the effect (~~nothing at for the first second, then somewhat increased
niceness after three seconds or more, depending on load average) pointless.
=== Other bugs === hz -> profhz in the p_pctcpu = f(p_cpticks) calcuation.
Collect scheduler functionality. Try to put each abstraction in just one
place.
1999-02-23 05:56:03 +03:00
|
|
|
#include <sys/sched.h>
|
1998-02-05 10:59:28 +03:00
|
|
|
|
|
|
|
#include <uvm/uvm_extern.h>
|
|
|
|
|
1994-06-29 10:29:24 +04:00
|
|
|
#ifdef KTRACE
|
|
|
|
#include <sys/ktrace.h>
|
|
|
|
#endif
|
|
|
|
|
Scheduler bug fixes and reorganization
* fix the ancient nice(1) bug, where nice +20 processes incorrectly
steal 10 - 20% of the CPU, (or even more depending on load average)
* provide a new schedclk() mechanism at a new clock at schedhz, so high
platform hz values don't cause nice +0 processes to look like they are
niced
* change the algorithm slightly, and reorganize the code a lot
* fix percent-CPU calculation bugs, and eliminate some no-op code
=== nice bug === Correctly divide the scheduler queues between niced and
compute-bound processes. The current nice weight of two (sort of, see
`algorithm change' below) neatly divides the USRPRI queues in half; this
should have been used to clip p_estcpu, instead of UCHAR_MAX. Besides
being the wrong amount, clipping an unsigned char to UCHAR_MAX is a no-op,
and it was done after decay_cpu() which can only _reduce_ the value. It
has to be kept <= NICE_WEIGHT * PRIO_MAX - PPQ or processes can
scheduler-penalize themselves onto the same queue as nice +20 processes.
(Or even a higher one.)
=== New schedclk() mechansism === Some platforms should be cutting down
stathz before hitting the scheduler, since the scheduler algorithm only
works right in the vicinity of 64 Hz. Rather than prescale hz, then scale
back and forth by 4 every time p_estcpu is touched (each occurance an
abstraction violation), use p_estcpu without scaling and require schedhz
to be generated directly at the right frequency. Use a default stathz (well,
actually, profhz) / 4, so nothing changes unless a platform defines schedhz
and a new clock. Define these for alpha, where hz==1024, and nice was
totally broke.
=== Algorithm change === The nice value used to be added to the
exponentially-decayed scheduler history value p_estcpu, in _addition_ to
be incorporated directly (with greater wieght) into the priority calculation.
At first glance, it appears to be a pointless increase of 1/8 the nice
effect (pri = p_estcpu/4 + nice*2), but it's actually at least 3x that
because it will ramp up linearly but be decayed only exponentially, thus
converging to an additional .75 nice for a loadaverage of one. I killed
this, it makes the behavior hard to control, almost impossible to analyze,
and the effect (~~nothing at for the first second, then somewhat increased
niceness after three seconds or more, depending on load average) pointless.
=== Other bugs === hz -> profhz in the p_pctcpu = f(p_cpticks) calcuation.
Collect scheduler functionality. Try to put each abstraction in just one
place.
1999-02-23 05:56:03 +03:00
|
|
|
#define NICE_WEIGHT 2 /* priorities per nice level */
|
|
|
|
#define PPQ (128 / NQS) /* priorities per queue */
|
|
|
|
|
|
|
|
#define ESTCPULIM(e) min((e), NICE_WEIGHT * PRIO_MAX - PPQ)
|
|
|
|
|
1994-06-29 10:29:24 +04:00
|
|
|
#include <machine/cpu.h>
|
|
|
|
|
|
|
|
u_char curpriority; /* usrpri of curproc */
|
|
|
|
int lbolt; /* once a second sleep address */
|
|
|
|
|
1996-02-04 05:15:01 +03:00
|
|
|
void roundrobin __P((void *));
|
|
|
|
void schedcpu __P((void *));
|
|
|
|
void updatepri __P((struct proc *));
|
|
|
|
void endtsleep __P((void *));
|
|
|
|
|
1994-06-29 10:29:24 +04:00
|
|
|
/*
|
|
|
|
* Force switch among equal priority processes every 100ms.
|
|
|
|
*/
|
|
|
|
/* ARGSUSED */
|
|
|
|
void
|
|
|
|
roundrobin(arg)
|
|
|
|
void *arg;
|
|
|
|
{
|
|
|
|
|
|
|
|
need_resched();
|
|
|
|
timeout(roundrobin, NULL, hz / 10);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Constants for digital decay and forget:
|
|
|
|
* 90% of (p_estcpu) usage in 5 * loadav time
|
|
|
|
* 95% of (p_pctcpu) usage in 60 seconds (load insensitive)
|
|
|
|
* Note that, as ps(1) mentions, this can let percentages
|
|
|
|
* total over 100% (I've seen 137.9% for 3 processes).
|
|
|
|
*
|
|
|
|
* Note that hardclock updates p_estcpu and p_cpticks independently.
|
|
|
|
*
|
|
|
|
* We wish to decay away 90% of p_estcpu in (5 * loadavg) seconds.
|
|
|
|
* That is, the system wants to compute a value of decay such
|
|
|
|
* that the following for loop:
|
|
|
|
* for (i = 0; i < (5 * loadavg); i++)
|
|
|
|
* p_estcpu *= decay;
|
|
|
|
* will compute
|
|
|
|
* p_estcpu *= 0.1;
|
|
|
|
* for all values of loadavg:
|
|
|
|
*
|
|
|
|
* Mathematically this loop can be expressed by saying:
|
|
|
|
* decay ** (5 * loadavg) ~= .1
|
|
|
|
*
|
|
|
|
* The system computes decay as:
|
|
|
|
* decay = (2 * loadavg) / (2 * loadavg + 1)
|
|
|
|
*
|
|
|
|
* We wish to prove that the system's computation of decay
|
|
|
|
* will always fulfill the equation:
|
|
|
|
* decay ** (5 * loadavg) ~= .1
|
|
|
|
*
|
|
|
|
* If we compute b as:
|
|
|
|
* b = 2 * loadavg
|
|
|
|
* then
|
|
|
|
* decay = b / (b + 1)
|
|
|
|
*
|
|
|
|
* We now need to prove two things:
|
|
|
|
* 1) Given factor ** (5 * loadavg) ~= .1, prove factor == b/(b+1)
|
|
|
|
* 2) Given b/(b+1) ** power ~= .1, prove power == (5 * loadavg)
|
|
|
|
*
|
|
|
|
* Facts:
|
|
|
|
* For x close to zero, exp(x) =~ 1 + x, since
|
|
|
|
* exp(x) = 0! + x**1/1! + x**2/2! + ... .
|
|
|
|
* therefore exp(-1/b) =~ 1 - (1/b) = (b-1)/b.
|
|
|
|
* For x close to zero, ln(1+x) =~ x, since
|
|
|
|
* ln(1+x) = x - x**2/2 + x**3/3 - ... -1 < x < 1
|
|
|
|
* therefore ln(b/(b+1)) = ln(1 - 1/(b+1)) =~ -1/(b+1).
|
|
|
|
* ln(.1) =~ -2.30
|
|
|
|
*
|
|
|
|
* Proof of (1):
|
|
|
|
* Solve (factor)**(power) =~ .1 given power (5*loadav):
|
|
|
|
* solving for factor,
|
|
|
|
* ln(factor) =~ (-2.30/5*loadav), or
|
|
|
|
* factor =~ exp(-1/((5/2.30)*loadav)) =~ exp(-1/(2*loadav)) =
|
|
|
|
* exp(-1/b) =~ (b-1)/b =~ b/(b+1). QED
|
|
|
|
*
|
|
|
|
* Proof of (2):
|
|
|
|
* Solve (factor)**(power) =~ .1 given factor == (b/(b+1)):
|
|
|
|
* solving for power,
|
|
|
|
* power*ln(b/(b+1)) =~ -2.30, or
|
|
|
|
* power =~ 2.3 * (b + 1) = 4.6*loadav + 2.3 =~ 5*loadav. QED
|
|
|
|
*
|
|
|
|
* Actual power values for the implemented algorithm are as follows:
|
|
|
|
* loadav: 1 2 3 4
|
|
|
|
* power: 5.68 10.32 14.94 19.55
|
|
|
|
*/
|
|
|
|
|
|
|
|
/* calculations for digital decay to forget 90% of usage in 5*loadav sec */
|
|
|
|
#define loadfactor(loadav) (2 * (loadav))
|
|
|
|
#define decay_cpu(loadfac, cpu) (((loadfac) * (cpu)) / ((loadfac) + FSCALE))
|
|
|
|
|
|
|
|
/* decay 95% of `p_pctcpu' in 60 seconds; see CCPU_SHIFT before changing */
|
|
|
|
fixpt_t ccpu = 0.95122942450071400909 * FSCALE; /* exp(-1/20) */
|
|
|
|
|
|
|
|
/*
|
|
|
|
* If `ccpu' is not equal to `exp(-1/20)' and you still want to use the
|
|
|
|
* faster/more-accurate formula, you'll have to estimate CCPU_SHIFT below
|
|
|
|
* and possibly adjust FSHIFT in "param.h" so that (FSHIFT >= CCPU_SHIFT).
|
|
|
|
*
|
|
|
|
* To estimate CCPU_SHIFT for exp(-1/20), the following formula was used:
|
|
|
|
* 1 - exp(-1/20) ~= 0.0487 ~= 0.0488 == 1 (fixed pt, *11* bits).
|
|
|
|
*
|
|
|
|
* If you dont want to bother with the faster/more-accurate formula, you
|
|
|
|
* can set CCPU_SHIFT to (FSHIFT + 1) which will use a slower/less-accurate
|
|
|
|
* (more general) method of calculating the %age of CPU used by a process.
|
|
|
|
*/
|
|
|
|
#define CCPU_SHIFT 11
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Recompute process priorities, every hz ticks.
|
|
|
|
*/
|
|
|
|
/* ARGSUSED */
|
|
|
|
void
|
|
|
|
schedcpu(arg)
|
|
|
|
void *arg;
|
|
|
|
{
|
|
|
|
register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
|
|
|
|
register struct proc *p;
|
|
|
|
register int s;
|
|
|
|
register unsigned int newcpu;
|
|
|
|
|
|
|
|
wakeup((caddr_t)&lbolt);
|
1994-08-30 07:04:28 +04:00
|
|
|
for (p = allproc.lh_first; p != 0; p = p->p_list.le_next) {
|
1994-06-29 10:29:24 +04:00
|
|
|
/*
|
|
|
|
* Increment time in/out of memory and sleep time
|
|
|
|
* (if sleeping). We ignore overflow; with 16-bit int's
|
|
|
|
* (remember them?) overflow takes 45 days.
|
|
|
|
*/
|
|
|
|
p->p_swtime++;
|
|
|
|
if (p->p_stat == SSLEEP || p->p_stat == SSTOP)
|
|
|
|
p->p_slptime++;
|
|
|
|
p->p_pctcpu = (p->p_pctcpu * ccpu) >> FSHIFT;
|
|
|
|
/*
|
|
|
|
* If the process has slept the entire second,
|
|
|
|
* stop recalculating its priority until it wakes up.
|
|
|
|
*/
|
|
|
|
if (p->p_slptime > 1)
|
|
|
|
continue;
|
|
|
|
s = splstatclock(); /* prevent state changes */
|
|
|
|
/*
|
|
|
|
* p_pctcpu is only for ps.
|
|
|
|
*/
|
Scheduler bug fixes and reorganization
* fix the ancient nice(1) bug, where nice +20 processes incorrectly
steal 10 - 20% of the CPU, (or even more depending on load average)
* provide a new schedclk() mechanism at a new clock at schedhz, so high
platform hz values don't cause nice +0 processes to look like they are
niced
* change the algorithm slightly, and reorganize the code a lot
* fix percent-CPU calculation bugs, and eliminate some no-op code
=== nice bug === Correctly divide the scheduler queues between niced and
compute-bound processes. The current nice weight of two (sort of, see
`algorithm change' below) neatly divides the USRPRI queues in half; this
should have been used to clip p_estcpu, instead of UCHAR_MAX. Besides
being the wrong amount, clipping an unsigned char to UCHAR_MAX is a no-op,
and it was done after decay_cpu() which can only _reduce_ the value. It
has to be kept <= NICE_WEIGHT * PRIO_MAX - PPQ or processes can
scheduler-penalize themselves onto the same queue as nice +20 processes.
(Or even a higher one.)
=== New schedclk() mechansism === Some platforms should be cutting down
stathz before hitting the scheduler, since the scheduler algorithm only
works right in the vicinity of 64 Hz. Rather than prescale hz, then scale
back and forth by 4 every time p_estcpu is touched (each occurance an
abstraction violation), use p_estcpu without scaling and require schedhz
to be generated directly at the right frequency. Use a default stathz (well,
actually, profhz) / 4, so nothing changes unless a platform defines schedhz
and a new clock. Define these for alpha, where hz==1024, and nice was
totally broke.
=== Algorithm change === The nice value used to be added to the
exponentially-decayed scheduler history value p_estcpu, in _addition_ to
be incorporated directly (with greater wieght) into the priority calculation.
At first glance, it appears to be a pointless increase of 1/8 the nice
effect (pri = p_estcpu/4 + nice*2), but it's actually at least 3x that
because it will ramp up linearly but be decayed only exponentially, thus
converging to an additional .75 nice for a loadaverage of one. I killed
this, it makes the behavior hard to control, almost impossible to analyze,
and the effect (~~nothing at for the first second, then somewhat increased
niceness after three seconds or more, depending on load average) pointless.
=== Other bugs === hz -> profhz in the p_pctcpu = f(p_cpticks) calcuation.
Collect scheduler functionality. Try to put each abstraction in just one
place.
1999-02-23 05:56:03 +03:00
|
|
|
KASSERT(profhz);
|
1994-06-29 10:29:24 +04:00
|
|
|
#if (FSHIFT >= CCPU_SHIFT)
|
Scheduler bug fixes and reorganization
* fix the ancient nice(1) bug, where nice +20 processes incorrectly
steal 10 - 20% of the CPU, (or even more depending on load average)
* provide a new schedclk() mechanism at a new clock at schedhz, so high
platform hz values don't cause nice +0 processes to look like they are
niced
* change the algorithm slightly, and reorganize the code a lot
* fix percent-CPU calculation bugs, and eliminate some no-op code
=== nice bug === Correctly divide the scheduler queues between niced and
compute-bound processes. The current nice weight of two (sort of, see
`algorithm change' below) neatly divides the USRPRI queues in half; this
should have been used to clip p_estcpu, instead of UCHAR_MAX. Besides
being the wrong amount, clipping an unsigned char to UCHAR_MAX is a no-op,
and it was done after decay_cpu() which can only _reduce_ the value. It
has to be kept <= NICE_WEIGHT * PRIO_MAX - PPQ or processes can
scheduler-penalize themselves onto the same queue as nice +20 processes.
(Or even a higher one.)
=== New schedclk() mechansism === Some platforms should be cutting down
stathz before hitting the scheduler, since the scheduler algorithm only
works right in the vicinity of 64 Hz. Rather than prescale hz, then scale
back and forth by 4 every time p_estcpu is touched (each occurance an
abstraction violation), use p_estcpu without scaling and require schedhz
to be generated directly at the right frequency. Use a default stathz (well,
actually, profhz) / 4, so nothing changes unless a platform defines schedhz
and a new clock. Define these for alpha, where hz==1024, and nice was
totally broke.
=== Algorithm change === The nice value used to be added to the
exponentially-decayed scheduler history value p_estcpu, in _addition_ to
be incorporated directly (with greater wieght) into the priority calculation.
At first glance, it appears to be a pointless increase of 1/8 the nice
effect (pri = p_estcpu/4 + nice*2), but it's actually at least 3x that
because it will ramp up linearly but be decayed only exponentially, thus
converging to an additional .75 nice for a loadaverage of one. I killed
this, it makes the behavior hard to control, almost impossible to analyze,
and the effect (~~nothing at for the first second, then somewhat increased
niceness after three seconds or more, depending on load average) pointless.
=== Other bugs === hz -> profhz in the p_pctcpu = f(p_cpticks) calcuation.
Collect scheduler functionality. Try to put each abstraction in just one
place.
1999-02-23 05:56:03 +03:00
|
|
|
p->p_pctcpu += (profhz == 100)?
|
1994-06-29 10:29:24 +04:00
|
|
|
((fixpt_t) p->p_cpticks) << (FSHIFT - CCPU_SHIFT):
|
|
|
|
100 * (((fixpt_t) p->p_cpticks)
|
Scheduler bug fixes and reorganization
* fix the ancient nice(1) bug, where nice +20 processes incorrectly
steal 10 - 20% of the CPU, (or even more depending on load average)
* provide a new schedclk() mechanism at a new clock at schedhz, so high
platform hz values don't cause nice +0 processes to look like they are
niced
* change the algorithm slightly, and reorganize the code a lot
* fix percent-CPU calculation bugs, and eliminate some no-op code
=== nice bug === Correctly divide the scheduler queues between niced and
compute-bound processes. The current nice weight of two (sort of, see
`algorithm change' below) neatly divides the USRPRI queues in half; this
should have been used to clip p_estcpu, instead of UCHAR_MAX. Besides
being the wrong amount, clipping an unsigned char to UCHAR_MAX is a no-op,
and it was done after decay_cpu() which can only _reduce_ the value. It
has to be kept <= NICE_WEIGHT * PRIO_MAX - PPQ or processes can
scheduler-penalize themselves onto the same queue as nice +20 processes.
(Or even a higher one.)
=== New schedclk() mechansism === Some platforms should be cutting down
stathz before hitting the scheduler, since the scheduler algorithm only
works right in the vicinity of 64 Hz. Rather than prescale hz, then scale
back and forth by 4 every time p_estcpu is touched (each occurance an
abstraction violation), use p_estcpu without scaling and require schedhz
to be generated directly at the right frequency. Use a default stathz (well,
actually, profhz) / 4, so nothing changes unless a platform defines schedhz
and a new clock. Define these for alpha, where hz==1024, and nice was
totally broke.
=== Algorithm change === The nice value used to be added to the
exponentially-decayed scheduler history value p_estcpu, in _addition_ to
be incorporated directly (with greater wieght) into the priority calculation.
At first glance, it appears to be a pointless increase of 1/8 the nice
effect (pri = p_estcpu/4 + nice*2), but it's actually at least 3x that
because it will ramp up linearly but be decayed only exponentially, thus
converging to an additional .75 nice for a loadaverage of one. I killed
this, it makes the behavior hard to control, almost impossible to analyze,
and the effect (~~nothing at for the first second, then somewhat increased
niceness after three seconds or more, depending on load average) pointless.
=== Other bugs === hz -> profhz in the p_pctcpu = f(p_cpticks) calcuation.
Collect scheduler functionality. Try to put each abstraction in just one
place.
1999-02-23 05:56:03 +03:00
|
|
|
<< (FSHIFT - CCPU_SHIFT)) / profhz;
|
1994-06-29 10:29:24 +04:00
|
|
|
#else
|
|
|
|
p->p_pctcpu += ((FSCALE - ccpu) *
|
Scheduler bug fixes and reorganization
* fix the ancient nice(1) bug, where nice +20 processes incorrectly
steal 10 - 20% of the CPU, (or even more depending on load average)
* provide a new schedclk() mechanism at a new clock at schedhz, so high
platform hz values don't cause nice +0 processes to look like they are
niced
* change the algorithm slightly, and reorganize the code a lot
* fix percent-CPU calculation bugs, and eliminate some no-op code
=== nice bug === Correctly divide the scheduler queues between niced and
compute-bound processes. The current nice weight of two (sort of, see
`algorithm change' below) neatly divides the USRPRI queues in half; this
should have been used to clip p_estcpu, instead of UCHAR_MAX. Besides
being the wrong amount, clipping an unsigned char to UCHAR_MAX is a no-op,
and it was done after decay_cpu() which can only _reduce_ the value. It
has to be kept <= NICE_WEIGHT * PRIO_MAX - PPQ or processes can
scheduler-penalize themselves onto the same queue as nice +20 processes.
(Or even a higher one.)
=== New schedclk() mechansism === Some platforms should be cutting down
stathz before hitting the scheduler, since the scheduler algorithm only
works right in the vicinity of 64 Hz. Rather than prescale hz, then scale
back and forth by 4 every time p_estcpu is touched (each occurance an
abstraction violation), use p_estcpu without scaling and require schedhz
to be generated directly at the right frequency. Use a default stathz (well,
actually, profhz) / 4, so nothing changes unless a platform defines schedhz
and a new clock. Define these for alpha, where hz==1024, and nice was
totally broke.
=== Algorithm change === The nice value used to be added to the
exponentially-decayed scheduler history value p_estcpu, in _addition_ to
be incorporated directly (with greater wieght) into the priority calculation.
At first glance, it appears to be a pointless increase of 1/8 the nice
effect (pri = p_estcpu/4 + nice*2), but it's actually at least 3x that
because it will ramp up linearly but be decayed only exponentially, thus
converging to an additional .75 nice for a loadaverage of one. I killed
this, it makes the behavior hard to control, almost impossible to analyze,
and the effect (~~nothing at for the first second, then somewhat increased
niceness after three seconds or more, depending on load average) pointless.
=== Other bugs === hz -> profhz in the p_pctcpu = f(p_cpticks) calcuation.
Collect scheduler functionality. Try to put each abstraction in just one
place.
1999-02-23 05:56:03 +03:00
|
|
|
(p->p_cpticks * FSCALE / profhz)) >> FSHIFT;
|
1994-06-29 10:29:24 +04:00
|
|
|
#endif
|
|
|
|
p->p_cpticks = 0;
|
Scheduler bug fixes and reorganization
* fix the ancient nice(1) bug, where nice +20 processes incorrectly
steal 10 - 20% of the CPU, (or even more depending on load average)
* provide a new schedclk() mechanism at a new clock at schedhz, so high
platform hz values don't cause nice +0 processes to look like they are
niced
* change the algorithm slightly, and reorganize the code a lot
* fix percent-CPU calculation bugs, and eliminate some no-op code
=== nice bug === Correctly divide the scheduler queues between niced and
compute-bound processes. The current nice weight of two (sort of, see
`algorithm change' below) neatly divides the USRPRI queues in half; this
should have been used to clip p_estcpu, instead of UCHAR_MAX. Besides
being the wrong amount, clipping an unsigned char to UCHAR_MAX is a no-op,
and it was done after decay_cpu() which can only _reduce_ the value. It
has to be kept <= NICE_WEIGHT * PRIO_MAX - PPQ or processes can
scheduler-penalize themselves onto the same queue as nice +20 processes.
(Or even a higher one.)
=== New schedclk() mechansism === Some platforms should be cutting down
stathz before hitting the scheduler, since the scheduler algorithm only
works right in the vicinity of 64 Hz. Rather than prescale hz, then scale
back and forth by 4 every time p_estcpu is touched (each occurance an
abstraction violation), use p_estcpu without scaling and require schedhz
to be generated directly at the right frequency. Use a default stathz (well,
actually, profhz) / 4, so nothing changes unless a platform defines schedhz
and a new clock. Define these for alpha, where hz==1024, and nice was
totally broke.
=== Algorithm change === The nice value used to be added to the
exponentially-decayed scheduler history value p_estcpu, in _addition_ to
be incorporated directly (with greater wieght) into the priority calculation.
At first glance, it appears to be a pointless increase of 1/8 the nice
effect (pri = p_estcpu/4 + nice*2), but it's actually at least 3x that
because it will ramp up linearly but be decayed only exponentially, thus
converging to an additional .75 nice for a loadaverage of one. I killed
this, it makes the behavior hard to control, almost impossible to analyze,
and the effect (~~nothing at for the first second, then somewhat increased
niceness after three seconds or more, depending on load average) pointless.
=== Other bugs === hz -> profhz in the p_pctcpu = f(p_cpticks) calcuation.
Collect scheduler functionality. Try to put each abstraction in just one
place.
1999-02-23 05:56:03 +03:00
|
|
|
newcpu = (u_int)decay_cpu(loadfac, p->p_estcpu);
|
|
|
|
p->p_estcpu = newcpu;
|
1994-06-29 10:29:24 +04:00
|
|
|
resetpriority(p);
|
|
|
|
if (p->p_priority >= PUSER) {
|
|
|
|
if ((p != curproc) &&
|
|
|
|
p->p_stat == SRUN &&
|
|
|
|
(p->p_flag & P_INMEM) &&
|
|
|
|
(p->p_priority / PPQ) != (p->p_usrpri / PPQ)) {
|
1996-11-06 23:19:19 +03:00
|
|
|
remrunqueue(p);
|
1994-06-29 10:29:24 +04:00
|
|
|
p->p_priority = p->p_usrpri;
|
|
|
|
setrunqueue(p);
|
|
|
|
} else
|
|
|
|
p->p_priority = p->p_usrpri;
|
|
|
|
}
|
|
|
|
splx(s);
|
|
|
|
}
|
1998-02-05 10:59:28 +03:00
|
|
|
uvm_meter();
|
1994-06-29 10:29:24 +04:00
|
|
|
timeout(schedcpu, (void *)0, hz);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Recalculate the priority of a process after it has slept for a while.
|
|
|
|
* For all load averages >= 1 and max p_estcpu of 255, sleeping for at
|
|
|
|
* least six times the loadfactor will decay p_estcpu to zero.
|
|
|
|
*/
|
|
|
|
void
|
|
|
|
updatepri(p)
|
|
|
|
register struct proc *p;
|
|
|
|
{
|
|
|
|
register unsigned int newcpu = p->p_estcpu;
|
|
|
|
register fixpt_t loadfac = loadfactor(averunnable.ldavg[0]);
|
|
|
|
|
|
|
|
if (p->p_slptime > 5 * loadfac)
|
|
|
|
p->p_estcpu = 0;
|
|
|
|
else {
|
|
|
|
p->p_slptime--; /* the first time was done in schedcpu */
|
|
|
|
while (newcpu && --p->p_slptime)
|
|
|
|
newcpu = (int) decay_cpu(loadfac, newcpu);
|
Scheduler bug fixes and reorganization
* fix the ancient nice(1) bug, where nice +20 processes incorrectly
steal 10 - 20% of the CPU, (or even more depending on load average)
* provide a new schedclk() mechanism at a new clock at schedhz, so high
platform hz values don't cause nice +0 processes to look like they are
niced
* change the algorithm slightly, and reorganize the code a lot
* fix percent-CPU calculation bugs, and eliminate some no-op code
=== nice bug === Correctly divide the scheduler queues between niced and
compute-bound processes. The current nice weight of two (sort of, see
`algorithm change' below) neatly divides the USRPRI queues in half; this
should have been used to clip p_estcpu, instead of UCHAR_MAX. Besides
being the wrong amount, clipping an unsigned char to UCHAR_MAX is a no-op,
and it was done after decay_cpu() which can only _reduce_ the value. It
has to be kept <= NICE_WEIGHT * PRIO_MAX - PPQ or processes can
scheduler-penalize themselves onto the same queue as nice +20 processes.
(Or even a higher one.)
=== New schedclk() mechansism === Some platforms should be cutting down
stathz before hitting the scheduler, since the scheduler algorithm only
works right in the vicinity of 64 Hz. Rather than prescale hz, then scale
back and forth by 4 every time p_estcpu is touched (each occurance an
abstraction violation), use p_estcpu without scaling and require schedhz
to be generated directly at the right frequency. Use a default stathz (well,
actually, profhz) / 4, so nothing changes unless a platform defines schedhz
and a new clock. Define these for alpha, where hz==1024, and nice was
totally broke.
=== Algorithm change === The nice value used to be added to the
exponentially-decayed scheduler history value p_estcpu, in _addition_ to
be incorporated directly (with greater wieght) into the priority calculation.
At first glance, it appears to be a pointless increase of 1/8 the nice
effect (pri = p_estcpu/4 + nice*2), but it's actually at least 3x that
because it will ramp up linearly but be decayed only exponentially, thus
converging to an additional .75 nice for a loadaverage of one. I killed
this, it makes the behavior hard to control, almost impossible to analyze,
and the effect (~~nothing at for the first second, then somewhat increased
niceness after three seconds or more, depending on load average) pointless.
=== Other bugs === hz -> profhz in the p_pctcpu = f(p_cpticks) calcuation.
Collect scheduler functionality. Try to put each abstraction in just one
place.
1999-02-23 05:56:03 +03:00
|
|
|
p->p_estcpu = newcpu;
|
1994-06-29 10:29:24 +04:00
|
|
|
}
|
|
|
|
resetpriority(p);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* We're only looking at 7 bits of the address; everything is
|
|
|
|
* aligned to 4, lots of things are aligned to greater powers
|
|
|
|
* of 2. Shift right by 8, i.e. drop the bottom 256 worth.
|
|
|
|
*/
|
|
|
|
#define TABLESIZE 128
|
1994-10-31 00:43:03 +03:00
|
|
|
#define LOOKUP(x) (((long)(x) >> 8) & (TABLESIZE - 1))
|
1994-06-29 10:29:24 +04:00
|
|
|
struct slpque {
|
|
|
|
struct proc *sq_head;
|
|
|
|
struct proc **sq_tailp;
|
|
|
|
} slpque[TABLESIZE];
|
|
|
|
|
|
|
|
/*
|
|
|
|
* During autoconfiguration or after a panic, a sleep will simply
|
|
|
|
* lower the priority briefly to allow interrupts, then return.
|
|
|
|
* The priority to be used (safepri) is machine-dependent, thus this
|
|
|
|
* value is initialized and maintained in the machine-dependent layers.
|
|
|
|
* This priority will typically be 0, or the lowest priority
|
|
|
|
* that is safe for use on the interrupt stack; it can be made
|
|
|
|
* higher to block network software interrupts after panics.
|
|
|
|
*/
|
|
|
|
int safepri;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* General sleep call. Suspends the current process until a wakeup is
|
|
|
|
* performed on the specified identifier. The process will then be made
|
|
|
|
* runnable with the specified priority. Sleeps at most timo/hz seconds
|
|
|
|
* (0 means no timeout). If pri includes PCATCH flag, signals are checked
|
|
|
|
* before and after sleeping, else signals are not checked. Returns 0 if
|
|
|
|
* awakened, EWOULDBLOCK if the timeout expires. If PCATCH is set and a
|
|
|
|
* signal needs to be delivered, ERESTART is returned if the current system
|
|
|
|
* call should be restarted if possible, and EINTR is returned if the system
|
|
|
|
* call should be interrupted by the signal (return EINTR).
|
|
|
|
*/
|
|
|
|
int
|
|
|
|
tsleep(ident, priority, wmesg, timo)
|
|
|
|
void *ident;
|
|
|
|
int priority, timo;
|
1997-10-09 16:49:44 +04:00
|
|
|
const char *wmesg;
|
1994-06-29 10:29:24 +04:00
|
|
|
{
|
|
|
|
register struct proc *p = curproc;
|
|
|
|
register struct slpque *qp;
|
1998-02-12 23:39:41 +03:00
|
|
|
register int s;
|
1994-06-29 10:29:24 +04:00
|
|
|
int sig, catch = priority & PCATCH;
|
|
|
|
extern int cold;
|
|
|
|
void endtsleep __P((void *));
|
|
|
|
|
|
|
|
if (cold || panicstr) {
|
|
|
|
/*
|
|
|
|
* After a panic, or during autoconfiguration,
|
|
|
|
* just give interrupts a chance, then just return;
|
|
|
|
* don't run any other procs or panic below,
|
|
|
|
* in case this is the idle process and already asleep.
|
|
|
|
*/
|
1996-10-16 03:00:10 +04:00
|
|
|
s = splhigh();
|
1994-06-29 10:29:24 +04:00
|
|
|
splx(safepri);
|
|
|
|
splx(s);
|
|
|
|
return (0);
|
|
|
|
}
|
1996-10-16 03:00:10 +04:00
|
|
|
|
|
|
|
#ifdef KTRACE
|
|
|
|
if (KTRPOINT(p, KTR_CSW))
|
|
|
|
ktrcsw(p->p_tracep, 1, 0);
|
|
|
|
#endif
|
|
|
|
s = splhigh();
|
|
|
|
|
1994-06-29 10:29:24 +04:00
|
|
|
#ifdef DIAGNOSTIC
|
|
|
|
if (ident == NULL || p->p_stat != SRUN || p->p_back)
|
|
|
|
panic("tsleep");
|
|
|
|
#endif
|
|
|
|
p->p_wchan = ident;
|
|
|
|
p->p_wmesg = wmesg;
|
|
|
|
p->p_slptime = 0;
|
|
|
|
p->p_priority = priority & PRIMASK;
|
|
|
|
qp = &slpque[LOOKUP(ident)];
|
|
|
|
if (qp->sq_head == 0)
|
|
|
|
qp->sq_head = p;
|
|
|
|
else
|
|
|
|
*qp->sq_tailp = p;
|
|
|
|
*(qp->sq_tailp = &p->p_forw) = 0;
|
|
|
|
if (timo)
|
|
|
|
timeout(endtsleep, (void *)p, timo);
|
|
|
|
/*
|
|
|
|
* We put ourselves on the sleep queue and start our timeout
|
|
|
|
* before calling CURSIG, as we could stop there, and a wakeup
|
|
|
|
* or a SIGCONT (or both) could occur while we were stopped.
|
|
|
|
* A SIGCONT would cause us to be marked as SSLEEP
|
|
|
|
* without resuming us, thus we must be ready for sleep
|
|
|
|
* when CURSIG is called. If the wakeup happens while we're
|
|
|
|
* stopped, p->p_wchan will be 0 upon return from CURSIG.
|
|
|
|
*/
|
|
|
|
if (catch) {
|
|
|
|
p->p_flag |= P_SINTR;
|
1996-02-04 05:15:01 +03:00
|
|
|
if ((sig = CURSIG(p)) != 0) {
|
1994-06-29 10:29:24 +04:00
|
|
|
if (p->p_wchan)
|
|
|
|
unsleep(p);
|
|
|
|
p->p_stat = SRUN;
|
|
|
|
goto resume;
|
|
|
|
}
|
|
|
|
if (p->p_wchan == 0) {
|
|
|
|
catch = 0;
|
|
|
|
goto resume;
|
|
|
|
}
|
|
|
|
} else
|
|
|
|
sig = 0;
|
|
|
|
p->p_stat = SSLEEP;
|
|
|
|
p->p_stats->p_ru.ru_nvcsw++;
|
|
|
|
mi_switch();
|
|
|
|
#ifdef DDB
|
|
|
|
/* handy breakpoint location after process "wakes" */
|
|
|
|
asm(".globl bpendtsleep ; bpendtsleep:");
|
|
|
|
#endif
|
|
|
|
resume:
|
|
|
|
curpriority = p->p_usrpri;
|
|
|
|
splx(s);
|
|
|
|
p->p_flag &= ~P_SINTR;
|
|
|
|
if (p->p_flag & P_TIMEOUT) {
|
|
|
|
p->p_flag &= ~P_TIMEOUT;
|
|
|
|
if (sig == 0) {
|
|
|
|
#ifdef KTRACE
|
|
|
|
if (KTRPOINT(p, KTR_CSW))
|
|
|
|
ktrcsw(p->p_tracep, 0, 0);
|
|
|
|
#endif
|
|
|
|
return (EWOULDBLOCK);
|
|
|
|
}
|
|
|
|
} else if (timo)
|
|
|
|
untimeout(endtsleep, (void *)p);
|
1996-02-04 05:15:01 +03:00
|
|
|
if (catch && (sig != 0 || (sig = CURSIG(p)) != 0)) {
|
1994-06-29 10:29:24 +04:00
|
|
|
#ifdef KTRACE
|
|
|
|
if (KTRPOINT(p, KTR_CSW))
|
|
|
|
ktrcsw(p->p_tracep, 0, 0);
|
|
|
|
#endif
|
1998-09-11 16:50:05 +04:00
|
|
|
if ((p->p_sigacts->ps_sigact[sig].sa_flags & SA_RESTART) == 0)
|
1994-06-29 10:29:24 +04:00
|
|
|
return (EINTR);
|
|
|
|
return (ERESTART);
|
|
|
|
}
|
|
|
|
#ifdef KTRACE
|
|
|
|
if (KTRPOINT(p, KTR_CSW))
|
|
|
|
ktrcsw(p->p_tracep, 0, 0);
|
|
|
|
#endif
|
|
|
|
return (0);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Implement timeout for tsleep.
|
|
|
|
* If process hasn't been awakened (wchan non-zero),
|
|
|
|
* set timeout flag and undo the sleep. If proc
|
|
|
|
* is stopped, just unsleep so it will remain stopped.
|
|
|
|
*/
|
|
|
|
void
|
|
|
|
endtsleep(arg)
|
|
|
|
void *arg;
|
|
|
|
{
|
|
|
|
register struct proc *p;
|
|
|
|
int s;
|
|
|
|
|
|
|
|
p = (struct proc *)arg;
|
|
|
|
s = splhigh();
|
|
|
|
if (p->p_wchan) {
|
|
|
|
if (p->p_stat == SSLEEP)
|
|
|
|
setrunnable(p);
|
|
|
|
else
|
|
|
|
unsleep(p);
|
|
|
|
p->p_flag |= P_TIMEOUT;
|
|
|
|
}
|
|
|
|
splx(s);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Short-term, non-interruptable sleep.
|
|
|
|
*/
|
|
|
|
void
|
|
|
|
sleep(ident, priority)
|
|
|
|
void *ident;
|
|
|
|
int priority;
|
|
|
|
{
|
|
|
|
register struct proc *p = curproc;
|
|
|
|
register struct slpque *qp;
|
1998-02-12 23:39:41 +03:00
|
|
|
register int s;
|
1994-06-29 10:29:24 +04:00
|
|
|
extern int cold;
|
|
|
|
|
|
|
|
#ifdef DIAGNOSTIC
|
|
|
|
if (priority > PZERO) {
|
1996-10-13 06:32:29 +04:00
|
|
|
printf("sleep called with priority %d > PZERO, wchan: %p\n",
|
1994-06-29 10:29:24 +04:00
|
|
|
priority, ident);
|
|
|
|
panic("old sleep");
|
|
|
|
}
|
|
|
|
#endif
|
|
|
|
s = splhigh();
|
|
|
|
if (cold || panicstr) {
|
|
|
|
/*
|
|
|
|
* After a panic, or during autoconfiguration,
|
|
|
|
* just give interrupts a chance, then just return;
|
|
|
|
* don't run any other procs or panic below,
|
|
|
|
* in case this is the idle process and already asleep.
|
|
|
|
*/
|
|
|
|
splx(safepri);
|
|
|
|
splx(s);
|
|
|
|
return;
|
|
|
|
}
|
|
|
|
#ifdef DIAGNOSTIC
|
|
|
|
if (ident == NULL || p->p_stat != SRUN || p->p_back)
|
|
|
|
panic("sleep");
|
|
|
|
#endif
|
|
|
|
p->p_wchan = ident;
|
|
|
|
p->p_wmesg = NULL;
|
|
|
|
p->p_slptime = 0;
|
|
|
|
p->p_priority = priority;
|
|
|
|
qp = &slpque[LOOKUP(ident)];
|
|
|
|
if (qp->sq_head == 0)
|
|
|
|
qp->sq_head = p;
|
|
|
|
else
|
|
|
|
*qp->sq_tailp = p;
|
|
|
|
*(qp->sq_tailp = &p->p_forw) = 0;
|
|
|
|
p->p_stat = SSLEEP;
|
|
|
|
p->p_stats->p_ru.ru_nvcsw++;
|
|
|
|
#ifdef KTRACE
|
|
|
|
if (KTRPOINT(p, KTR_CSW))
|
|
|
|
ktrcsw(p->p_tracep, 1, 0);
|
|
|
|
#endif
|
|
|
|
mi_switch();
|
|
|
|
#ifdef DDB
|
|
|
|
/* handy breakpoint location after process "wakes" */
|
|
|
|
asm(".globl bpendsleep ; bpendsleep:");
|
|
|
|
#endif
|
|
|
|
#ifdef KTRACE
|
|
|
|
if (KTRPOINT(p, KTR_CSW))
|
|
|
|
ktrcsw(p->p_tracep, 0, 0);
|
|
|
|
#endif
|
|
|
|
curpriority = p->p_usrpri;
|
|
|
|
splx(s);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Remove a process from its wait queue
|
|
|
|
*/
|
|
|
|
void
|
|
|
|
unsleep(p)
|
|
|
|
register struct proc *p;
|
|
|
|
{
|
|
|
|
register struct slpque *qp;
|
|
|
|
register struct proc **hp;
|
|
|
|
int s;
|
|
|
|
|
|
|
|
s = splhigh();
|
|
|
|
if (p->p_wchan) {
|
|
|
|
hp = &(qp = &slpque[LOOKUP(p->p_wchan)])->sq_head;
|
|
|
|
while (*hp != p)
|
|
|
|
hp = &(*hp)->p_forw;
|
|
|
|
*hp = p->p_forw;
|
|
|
|
if (qp->sq_tailp == &p->p_forw)
|
|
|
|
qp->sq_tailp = hp;
|
|
|
|
p->p_wchan = 0;
|
|
|
|
}
|
|
|
|
splx(s);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Make all processes sleeping on the specified identifier runnable.
|
|
|
|
*/
|
|
|
|
void
|
|
|
|
wakeup(ident)
|
|
|
|
register void *ident;
|
|
|
|
{
|
|
|
|
register struct slpque *qp;
|
|
|
|
register struct proc *p, **q;
|
|
|
|
int s;
|
|
|
|
|
|
|
|
s = splhigh();
|
|
|
|
qp = &slpque[LOOKUP(ident)];
|
|
|
|
restart:
|
1996-02-04 05:15:01 +03:00
|
|
|
for (q = &qp->sq_head; (p = *q) != NULL; ) {
|
1994-06-29 10:29:24 +04:00
|
|
|
#ifdef DIAGNOSTIC
|
1996-02-04 05:15:01 +03:00
|
|
|
if (p->p_back || (p->p_stat != SSLEEP && p->p_stat != SSTOP))
|
1994-06-29 10:29:24 +04:00
|
|
|
panic("wakeup");
|
|
|
|
#endif
|
|
|
|
if (p->p_wchan == ident) {
|
|
|
|
p->p_wchan = 0;
|
|
|
|
*q = p->p_forw;
|
|
|
|
if (qp->sq_tailp == &p->p_forw)
|
|
|
|
qp->sq_tailp = q;
|
|
|
|
if (p->p_stat == SSLEEP) {
|
|
|
|
/* OPTIMIZED EXPANSION OF setrunnable(p); */
|
|
|
|
if (p->p_slptime > 1)
|
|
|
|
updatepri(p);
|
|
|
|
p->p_slptime = 0;
|
|
|
|
p->p_stat = SRUN;
|
|
|
|
if (p->p_flag & P_INMEM)
|
|
|
|
setrunqueue(p);
|
|
|
|
/*
|
|
|
|
* Since curpriority is a user priority,
|
|
|
|
* p->p_priority is always better than
|
|
|
|
* curpriority.
|
|
|
|
*/
|
|
|
|
if ((p->p_flag & P_INMEM) == 0)
|
|
|
|
wakeup((caddr_t)&proc0);
|
|
|
|
else
|
|
|
|
need_resched();
|
|
|
|
/* END INLINE EXPANSION */
|
|
|
|
goto restart;
|
|
|
|
}
|
|
|
|
} else
|
|
|
|
q = &p->p_forw;
|
|
|
|
}
|
|
|
|
splx(s);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* The machine independent parts of mi_switch().
|
|
|
|
* Must be called at splstatclock() or higher.
|
|
|
|
*/
|
|
|
|
void
|
|
|
|
mi_switch()
|
|
|
|
{
|
|
|
|
register struct proc *p = curproc; /* XXX */
|
|
|
|
register struct rlimit *rlim;
|
|
|
|
register long s, u;
|
|
|
|
struct timeval tv;
|
|
|
|
|
1998-03-01 05:20:01 +03:00
|
|
|
#ifdef DEBUG
|
1998-11-04 09:19:55 +03:00
|
|
|
if (p->p_simple_locks) {
|
|
|
|
printf("p->p_simple_locks %d\n", p->p_simple_locks);
|
|
|
|
#ifdef LOCKDEBUG
|
|
|
|
simple_lock_dump();
|
|
|
|
#endif
|
1998-03-01 05:20:01 +03:00
|
|
|
panic("sleep: holding simple lock");
|
1998-11-04 09:19:55 +03:00
|
|
|
}
|
1998-03-01 05:20:01 +03:00
|
|
|
#endif
|
1994-06-29 10:29:24 +04:00
|
|
|
/*
|
|
|
|
* Compute the amount of time during which the current
|
|
|
|
* process was running, and add that to its total so far.
|
|
|
|
*/
|
|
|
|
microtime(&tv);
|
|
|
|
u = p->p_rtime.tv_usec + (tv.tv_usec - runtime.tv_usec);
|
|
|
|
s = p->p_rtime.tv_sec + (tv.tv_sec - runtime.tv_sec);
|
|
|
|
if (u < 0) {
|
|
|
|
u += 1000000;
|
|
|
|
s--;
|
|
|
|
} else if (u >= 1000000) {
|
|
|
|
u -= 1000000;
|
|
|
|
s++;
|
|
|
|
}
|
|
|
|
p->p_rtime.tv_usec = u;
|
|
|
|
p->p_rtime.tv_sec = s;
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Check if the process exceeds its cpu resource allocation.
|
|
|
|
* If over max, kill it. In any case, if it has run for more
|
|
|
|
* than 10 minutes, reduce priority to give others a chance.
|
|
|
|
*/
|
|
|
|
rlim = &p->p_rlimit[RLIMIT_CPU];
|
|
|
|
if (s >= rlim->rlim_cur) {
|
|
|
|
if (s >= rlim->rlim_max)
|
|
|
|
psignal(p, SIGKILL);
|
|
|
|
else {
|
|
|
|
psignal(p, SIGXCPU);
|
|
|
|
if (rlim->rlim_cur < rlim->rlim_max)
|
|
|
|
rlim->rlim_cur += 5;
|
|
|
|
}
|
|
|
|
}
|
1996-07-18 01:54:04 +04:00
|
|
|
if (autonicetime && s > autonicetime && p->p_ucred->cr_uid && p->p_nice == NZERO) {
|
1996-10-02 22:04:56 +04:00
|
|
|
p->p_nice = autoniceval + NZERO;
|
1994-06-29 10:29:24 +04:00
|
|
|
resetpriority(p);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Pick a new current process and record its start time.
|
|
|
|
*/
|
1998-02-05 10:59:28 +03:00
|
|
|
uvmexp.swtch++;
|
1994-06-29 10:29:24 +04:00
|
|
|
cpu_switch(p);
|
|
|
|
microtime(&runtime);
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Initialize the (doubly-linked) run queues
|
|
|
|
* to be empty.
|
|
|
|
*/
|
|
|
|
void
|
|
|
|
rqinit()
|
|
|
|
{
|
|
|
|
register int i;
|
|
|
|
|
|
|
|
for (i = 0; i < NQS; i++)
|
|
|
|
qs[i].ph_link = qs[i].ph_rlink = (struct proc *)&qs[i];
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Change process state to be runnable,
|
|
|
|
* placing it on the run queue if it is in memory,
|
|
|
|
* and awakening the swapper if it isn't in memory.
|
|
|
|
*/
|
|
|
|
void
|
|
|
|
setrunnable(p)
|
|
|
|
register struct proc *p;
|
|
|
|
{
|
|
|
|
register int s;
|
|
|
|
|
|
|
|
s = splhigh();
|
|
|
|
switch (p->p_stat) {
|
|
|
|
case 0:
|
|
|
|
case SRUN:
|
|
|
|
case SZOMB:
|
1999-07-22 22:13:36 +04:00
|
|
|
case SDEAD:
|
1994-06-29 10:29:24 +04:00
|
|
|
default:
|
|
|
|
panic("setrunnable");
|
|
|
|
case SSTOP:
|
1995-06-09 03:51:01 +04:00
|
|
|
/*
|
|
|
|
* If we're being traced (possibly because someone attached us
|
|
|
|
* while we were stopped), check for a signal from the debugger.
|
|
|
|
*/
|
1998-09-11 16:50:05 +04:00
|
|
|
if ((p->p_flag & P_TRACED) != 0 && p->p_xstat != 0) {
|
|
|
|
sigaddset(&p->p_siglist, p->p_xstat);
|
|
|
|
p->p_sigcheck = 1;
|
|
|
|
}
|
1994-06-29 10:29:24 +04:00
|
|
|
case SSLEEP:
|
|
|
|
unsleep(p); /* e.g. when sending signals */
|
|
|
|
break;
|
|
|
|
|
|
|
|
case SIDL:
|
|
|
|
break;
|
|
|
|
}
|
|
|
|
p->p_stat = SRUN;
|
|
|
|
if (p->p_flag & P_INMEM)
|
|
|
|
setrunqueue(p);
|
|
|
|
splx(s);
|
|
|
|
if (p->p_slptime > 1)
|
|
|
|
updatepri(p);
|
|
|
|
p->p_slptime = 0;
|
|
|
|
if ((p->p_flag & P_INMEM) == 0)
|
|
|
|
wakeup((caddr_t)&proc0);
|
|
|
|
else if (p->p_priority < curpriority)
|
|
|
|
need_resched();
|
|
|
|
}
|
|
|
|
|
|
|
|
/*
|
|
|
|
* Compute the priority of a process when running in user mode.
|
|
|
|
* Arrange to reschedule if the resulting priority is better
|
|
|
|
* than that of the current process.
|
|
|
|
*/
|
|
|
|
void
|
|
|
|
resetpriority(p)
|
|
|
|
register struct proc *p;
|
|
|
|
{
|
|
|
|
register unsigned int newpriority;
|
|
|
|
|
Scheduler bug fixes and reorganization
* fix the ancient nice(1) bug, where nice +20 processes incorrectly
steal 10 - 20% of the CPU, (or even more depending on load average)
* provide a new schedclk() mechanism at a new clock at schedhz, so high
platform hz values don't cause nice +0 processes to look like they are
niced
* change the algorithm slightly, and reorganize the code a lot
* fix percent-CPU calculation bugs, and eliminate some no-op code
=== nice bug === Correctly divide the scheduler queues between niced and
compute-bound processes. The current nice weight of two (sort of, see
`algorithm change' below) neatly divides the USRPRI queues in half; this
should have been used to clip p_estcpu, instead of UCHAR_MAX. Besides
being the wrong amount, clipping an unsigned char to UCHAR_MAX is a no-op,
and it was done after decay_cpu() which can only _reduce_ the value. It
has to be kept <= NICE_WEIGHT * PRIO_MAX - PPQ or processes can
scheduler-penalize themselves onto the same queue as nice +20 processes.
(Or even a higher one.)
=== New schedclk() mechansism === Some platforms should be cutting down
stathz before hitting the scheduler, since the scheduler algorithm only
works right in the vicinity of 64 Hz. Rather than prescale hz, then scale
back and forth by 4 every time p_estcpu is touched (each occurance an
abstraction violation), use p_estcpu without scaling and require schedhz
to be generated directly at the right frequency. Use a default stathz (well,
actually, profhz) / 4, so nothing changes unless a platform defines schedhz
and a new clock. Define these for alpha, where hz==1024, and nice was
totally broke.
=== Algorithm change === The nice value used to be added to the
exponentially-decayed scheduler history value p_estcpu, in _addition_ to
be incorporated directly (with greater wieght) into the priority calculation.
At first glance, it appears to be a pointless increase of 1/8 the nice
effect (pri = p_estcpu/4 + nice*2), but it's actually at least 3x that
because it will ramp up linearly but be decayed only exponentially, thus
converging to an additional .75 nice for a loadaverage of one. I killed
this, it makes the behavior hard to control, almost impossible to analyze,
and the effect (~~nothing at for the first second, then somewhat increased
niceness after three seconds or more, depending on load average) pointless.
=== Other bugs === hz -> profhz in the p_pctcpu = f(p_cpticks) calcuation.
Collect scheduler functionality. Try to put each abstraction in just one
place.
1999-02-23 05:56:03 +03:00
|
|
|
newpriority = PUSER + p->p_estcpu + NICE_WEIGHT * (p->p_nice - NZERO);
|
1994-06-29 10:29:24 +04:00
|
|
|
newpriority = min(newpriority, MAXPRI);
|
|
|
|
p->p_usrpri = newpriority;
|
|
|
|
if (newpriority < curpriority)
|
|
|
|
need_resched();
|
|
|
|
}
|
Scheduler bug fixes and reorganization
* fix the ancient nice(1) bug, where nice +20 processes incorrectly
steal 10 - 20% of the CPU, (or even more depending on load average)
* provide a new schedclk() mechanism at a new clock at schedhz, so high
platform hz values don't cause nice +0 processes to look like they are
niced
* change the algorithm slightly, and reorganize the code a lot
* fix percent-CPU calculation bugs, and eliminate some no-op code
=== nice bug === Correctly divide the scheduler queues between niced and
compute-bound processes. The current nice weight of two (sort of, see
`algorithm change' below) neatly divides the USRPRI queues in half; this
should have been used to clip p_estcpu, instead of UCHAR_MAX. Besides
being the wrong amount, clipping an unsigned char to UCHAR_MAX is a no-op,
and it was done after decay_cpu() which can only _reduce_ the value. It
has to be kept <= NICE_WEIGHT * PRIO_MAX - PPQ or processes can
scheduler-penalize themselves onto the same queue as nice +20 processes.
(Or even a higher one.)
=== New schedclk() mechansism === Some platforms should be cutting down
stathz before hitting the scheduler, since the scheduler algorithm only
works right in the vicinity of 64 Hz. Rather than prescale hz, then scale
back and forth by 4 every time p_estcpu is touched (each occurance an
abstraction violation), use p_estcpu without scaling and require schedhz
to be generated directly at the right frequency. Use a default stathz (well,
actually, profhz) / 4, so nothing changes unless a platform defines schedhz
and a new clock. Define these for alpha, where hz==1024, and nice was
totally broke.
=== Algorithm change === The nice value used to be added to the
exponentially-decayed scheduler history value p_estcpu, in _addition_ to
be incorporated directly (with greater wieght) into the priority calculation.
At first glance, it appears to be a pointless increase of 1/8 the nice
effect (pri = p_estcpu/4 + nice*2), but it's actually at least 3x that
because it will ramp up linearly but be decayed only exponentially, thus
converging to an additional .75 nice for a loadaverage of one. I killed
this, it makes the behavior hard to control, almost impossible to analyze,
and the effect (~~nothing at for the first second, then somewhat increased
niceness after three seconds or more, depending on load average) pointless.
=== Other bugs === hz -> profhz in the p_pctcpu = f(p_cpticks) calcuation.
Collect scheduler functionality. Try to put each abstraction in just one
place.
1999-02-23 05:56:03 +03:00
|
|
|
|
|
|
|
/*
|
1999-02-28 21:14:57 +03:00
|
|
|
* We adjust the priority of the current process. The priority of a process
|
|
|
|
* gets worse as it accumulates CPU time. The cpu usage estimator (p_estcpu)
|
|
|
|
* is increased here. The formula for computing priorities (in kern_synch.c)
|
|
|
|
* will compute a different value each time p_estcpu increases. This can
|
|
|
|
* cause a switch, but unless the priority crosses a PPQ boundary the actual
|
|
|
|
* queue will not change. The cpu usage estimator ramps up quite quickly
|
|
|
|
* when the process is running (linearly), and decays away exponentially, at
|
|
|
|
* a rate which is proportionally slower when the system is busy. The basic
|
|
|
|
* principal is that the system will 90% forget that the process used a lot
|
|
|
|
* of CPU time in 5 * loadav seconds. This causes the system to favor
|
|
|
|
* processes which haven't run much recently, and to round-robin among other
|
|
|
|
* processes.
|
Scheduler bug fixes and reorganization
* fix the ancient nice(1) bug, where nice +20 processes incorrectly
steal 10 - 20% of the CPU, (or even more depending on load average)
* provide a new schedclk() mechanism at a new clock at schedhz, so high
platform hz values don't cause nice +0 processes to look like they are
niced
* change the algorithm slightly, and reorganize the code a lot
* fix percent-CPU calculation bugs, and eliminate some no-op code
=== nice bug === Correctly divide the scheduler queues between niced and
compute-bound processes. The current nice weight of two (sort of, see
`algorithm change' below) neatly divides the USRPRI queues in half; this
should have been used to clip p_estcpu, instead of UCHAR_MAX. Besides
being the wrong amount, clipping an unsigned char to UCHAR_MAX is a no-op,
and it was done after decay_cpu() which can only _reduce_ the value. It
has to be kept <= NICE_WEIGHT * PRIO_MAX - PPQ or processes can
scheduler-penalize themselves onto the same queue as nice +20 processes.
(Or even a higher one.)
=== New schedclk() mechansism === Some platforms should be cutting down
stathz before hitting the scheduler, since the scheduler algorithm only
works right in the vicinity of 64 Hz. Rather than prescale hz, then scale
back and forth by 4 every time p_estcpu is touched (each occurance an
abstraction violation), use p_estcpu without scaling and require schedhz
to be generated directly at the right frequency. Use a default stathz (well,
actually, profhz) / 4, so nothing changes unless a platform defines schedhz
and a new clock. Define these for alpha, where hz==1024, and nice was
totally broke.
=== Algorithm change === The nice value used to be added to the
exponentially-decayed scheduler history value p_estcpu, in _addition_ to
be incorporated directly (with greater wieght) into the priority calculation.
At first glance, it appears to be a pointless increase of 1/8 the nice
effect (pri = p_estcpu/4 + nice*2), but it's actually at least 3x that
because it will ramp up linearly but be decayed only exponentially, thus
converging to an additional .75 nice for a loadaverage of one. I killed
this, it makes the behavior hard to control, almost impossible to analyze,
and the effect (~~nothing at for the first second, then somewhat increased
niceness after three seconds or more, depending on load average) pointless.
=== Other bugs === hz -> profhz in the p_pctcpu = f(p_cpticks) calcuation.
Collect scheduler functionality. Try to put each abstraction in just one
place.
1999-02-23 05:56:03 +03:00
|
|
|
*/
|
|
|
|
|
|
|
|
void
|
1999-02-28 21:14:57 +03:00
|
|
|
schedclock(p)
|
Scheduler bug fixes and reorganization
* fix the ancient nice(1) bug, where nice +20 processes incorrectly
steal 10 - 20% of the CPU, (or even more depending on load average)
* provide a new schedclk() mechanism at a new clock at schedhz, so high
platform hz values don't cause nice +0 processes to look like they are
niced
* change the algorithm slightly, and reorganize the code a lot
* fix percent-CPU calculation bugs, and eliminate some no-op code
=== nice bug === Correctly divide the scheduler queues between niced and
compute-bound processes. The current nice weight of two (sort of, see
`algorithm change' below) neatly divides the USRPRI queues in half; this
should have been used to clip p_estcpu, instead of UCHAR_MAX. Besides
being the wrong amount, clipping an unsigned char to UCHAR_MAX is a no-op,
and it was done after decay_cpu() which can only _reduce_ the value. It
has to be kept <= NICE_WEIGHT * PRIO_MAX - PPQ or processes can
scheduler-penalize themselves onto the same queue as nice +20 processes.
(Or even a higher one.)
=== New schedclk() mechansism === Some platforms should be cutting down
stathz before hitting the scheduler, since the scheduler algorithm only
works right in the vicinity of 64 Hz. Rather than prescale hz, then scale
back and forth by 4 every time p_estcpu is touched (each occurance an
abstraction violation), use p_estcpu without scaling and require schedhz
to be generated directly at the right frequency. Use a default stathz (well,
actually, profhz) / 4, so nothing changes unless a platform defines schedhz
and a new clock. Define these for alpha, where hz==1024, and nice was
totally broke.
=== Algorithm change === The nice value used to be added to the
exponentially-decayed scheduler history value p_estcpu, in _addition_ to
be incorporated directly (with greater wieght) into the priority calculation.
At first glance, it appears to be a pointless increase of 1/8 the nice
effect (pri = p_estcpu/4 + nice*2), but it's actually at least 3x that
because it will ramp up linearly but be decayed only exponentially, thus
converging to an additional .75 nice for a loadaverage of one. I killed
this, it makes the behavior hard to control, almost impossible to analyze,
and the effect (~~nothing at for the first second, then somewhat increased
niceness after three seconds or more, depending on load average) pointless.
=== Other bugs === hz -> profhz in the p_pctcpu = f(p_cpticks) calcuation.
Collect scheduler functionality. Try to put each abstraction in just one
place.
1999-02-23 05:56:03 +03:00
|
|
|
struct proc *p;
|
|
|
|
{
|
|
|
|
p->p_estcpu = ESTCPULIM(p->p_estcpu + 1);
|
|
|
|
resetpriority(p);
|
|
|
|
if (p->p_priority >= PUSER)
|
|
|
|
p->p_priority = p->p_usrpri;
|
|
|
|
}
|