wrong for many years that negative niceness would lower the priority
of a process below PUSER, and once below PUSER, there were conditionals
in the code that are required to test for whether a process was in
the kernel which would break.
The breakage could (and did) cause lock-ups, basically nothing else
but the least nice program being able to run in some conditions. The
algorithm which adjusts the priority now subtracts PRIO_MIN to do
things properly, and the ESTCPULIM() algorithm was updated to use
PRIO_TOTAL (PRIO_MAX - PRIO_MIN) to calculate the estcpu.
NICE_WEIGHT is now 1 to accomodate the full range of priorities better
(a -20 process with full CPU time has the priority of a +0 process with
no CPU time). There are now 20 queues (exactly; 80 priorities) for
use in user processes' scheduling, and PUSER has been lowered to 48
to accomplish this.
This means, to the user, that things will be scheduled more correctly
(noticeable), there is no lock-up anymore WRT a niced -20 process
never releasing the CPU time for other processes. In this fair system,
tsleep()ed < PUSER processes now will get the proper higher priority
than priority >= PUSER user processes.
The detective work of this was done by me, along with part of the
solution. Luoqi Chen has provided most of the solution, and really
helped me understand what was happening better, to boot :)
Submitted by: luoqi
Concept reviewed by: bde
syscall path inward. A system call may select whether it needs the MP
lock or not (the default being that it does need it).
A great deal of conditional SMP code for various deadended experiments
has been removed. 'cil' and 'cml' have been removed entirely, and the
locking around the cpl has been removed. The conditional
separately-locked fast-interrupt code has been removed, meaning that
interrupts must hold the CPL now (but they pretty much had to anyway).
Another reason for doing this is that the original separate-lock for
interrupts just doesn't apply to the interrupt thread mechanism being
contemplated.
Modifications to the cpl may now ONLY occur while holding the MP
lock. For example, if an otherwise MP safe syscall needs to mess with
the cpl, it must hold the MP lock for the duration and must (as usual)
save/restore the cpl in a nested fashion.
This is precursor work for the real meat coming later: avoiding having
to hold the MP lock for common syscalls and I/O's and interrupt threads.
It is expected that the spl mechanisms and new interrupt threading
mechanisms will be able to run in tandem, allowing a slow piecemeal
transition to occur.
This patch should result in a moderate performance improvement due to
the considerable amount of code that has been removed from the critical
path, especially the simplification of the spl*() calls. The real
performance gains will come later.
Approved by: jkh
Reviewed by: current, bde (exception.s)
Some work taken from: luoqi's patch
with the known bogus currtpriority. This undoes the previous changes to
sys/i386/i386/trap.c, sys/alpha/alpha/trap.c, sys/sys/systm.h
Now we have the patch set approved by bde.
Approved by: bde
commit to kern_synch.c:
----------------------------
revision 1.55
date: 1999/02/23 02:56:03; author: ross; state: Exp; lines: +39 -10
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.
----------------------------
The details are a little different in FreeBSD:
=== nice bug === Fixing this is the main point of this commit. We use
essentially the same clipping rule as NetBSD (our limit on p_estcpu
differs by a scale factor). However, clipping at all is fundamentally
bad. It gives free CPU the hoggiest hogs once they reach the limit, and
reaching the limit is normal for long-running hogs. This will be fixed
later.
=== New schedclk() mechanism === We don't use the NetBSD schedclk()
(now schedclock()) mechanism. We require (real)stathz to be about 128
and scale by an extra factor of 2 compared with NetBSD's statclock().
We scale p_estcpu instead of scaling the clock. This is more accurate
and flexible.
=== Algorithm change === Same change.
=== Other bugs === The p_pctcpu bug was fixed long ago. We don't try as
hard to abstract functionality yet.
Related changes: the new limit on p_estcpu must be exported to kern_exit.c
for clipping in wait1().
Agreed with by: dufault
and extend. The new function containing the code is named schedclock()
as in NetBSD, but it has slightly different semantics (it already handles
incrementation of p->p_cpticks, and it should handle any calling frequency).
Agreed with in principle by: dufault
Alot of the code in sys/kern directly accesses the *Q_HEAD and *Q_ENTRY
structures for list operations. This patch makes all list operations
in sys/kern use the queue(3) macros, rather than directly accessing the
*Q_{HEAD,ENTRY} structures.
This batch of changes compile to the same object files.
Reviewed by: phk
Submitted by: Jake Burkholder <jake@checker.org>
PR: 14914
-----------------------------
The core of the signalling code has been rewritten to operate
on the new sigset_t. No methodological changes have been made.
Most references to a sigset_t object are through macros (see
signalvar.h) to create a level of abstraction and to provide
a basis for further improvements.
The NSIG constant has not been changed to reflect the maximum
number of signals possible. The reason is that it breaks
programs (especially shells) which assume that all signals
have a non-null name in sys_signame. See src/bin/sh/trap.c
for an example. Instead _SIG_MAXSIG has been introduced to
hold the maximum signal possible with the new sigset_t.
struct sigprop has been moved from signalvar.h to kern_sig.c
because a) it is only used there, and b) access must be done
though function sigprop(). The latter because the table doesn't
holds properties for all signals, but only for the first NSIG
signals.
signal.h has been reorganized to make reading easier and to
add the new and/or modified structures. The "old" structures
are moved to signalvar.h to prevent namespace polution.
Especially the coda filesystem suffers from the change, because
it contained lines like (p->p_sigmask == SIGIO), which is easy
to do for integral types, but not for compound types.
NOTE: kdump (and port linux_kdump) must be recompiled.
Thanks to Garrett Wollman and Daniel Eischen for pressing the
importance of changing sigreturn as well.
numbers as chars or use bogus casts in an attempt to unmisrepresnt
them. In top, don't assume that 0xff is the only negative cpu
number when cpu numbers are (mis)represented.
Higher numbers led to smaller quanta.
In discussion with BDE, change this parameter to be in uSecs
to make it machine independent,
and limit it to non zero multiples of 'tick' (rounding down).
Also make the variabel globally available so that the present function that
returns its value (used for posix scheduling I believe) can go away.
Submitted by: Bruce Evans <bde@freebsd.org>
not per-process. Keep it in `switchtime' consistently.
It is now clear that the timestamp is always valid in fork_trampoline()
except when the child is running on a previously idle cpu, which
can only happen if there are multiple cpus, so don't check or set
the timestamp in fork_trampoline except in the (i386) SMP case.
Just remove the alpha code for setting it unconditionally, since
there is no SMP case for alpha and the code had rotted.
Parts reviewed by: dfr, phk
often for it to be a good criterion for switching kernel cpu hogs --
it is true after most wakeups. Use the criterion "has been running
for >= 2 quanta" instead.
runtime. p_runtime is unsigned while p_cpulimit is not, so this avoids the
nasty side effect of the process getting killed when the runtime comes up
"negative" due to other bugs.
respectively. Most of the longs should probably have been
u_longs, but this changes is just to prevent warnings about
casts between pointers and integers of different sizes, not
to fix poorly chosen types.
Clean up (or if antipodic: down) some of the msgbuf stuff.
Use an inline function rather than a macro for timecounter delta.
Maintain process "on-cpu" time as 64 bits of microseconds to avoid
needless second rollover overhead.
Avoid calling microuptime the second time in mi_switch() if we do
not pass through _idle in cpu_switch()
This should reduce our context-switch overhead a bit, in particular
on pre-P5 and SMP systems.
WARNING: Programs which muck about with struct proc in userland
will have to be fixed.
Reviewed, but found imperfect by: bde
update of cpu usage as shown by top when one process is cpu bound
(no system calls) while the system is otherwise idle (except for top).
Don't attempt to switch to the BSP in boot(). If the system was idle when
an interrupt caused a panic, this won't work. Instead, switch to the BSP
in cpu_reset.
Remove some spurious forward_statclock/forward_hardclock warnings.
* Figure out UTC relative to boottime. Four new functions provide
time relative to boottime.
* move "runtime" into struct proc. This helps fix the calcru()
problem in SMP.
* kill mono_time.
* add timespec{add|sub|cmp} macros to time.h. (XXX: These may change!)
* nanosleep, select & poll takes long sleeps one day at a time
Reviewed by: bde
Tested by: ache and others
_KPOSIX_PRIORITY_SCHEDULING options to work. Changes:
Change all "posix4" to "p1003_1b". Misnamed files are left
as "posix4" until I'm told if I can simply delete them and add
new ones;
Add _POSIX_PRIORITY_SCHEDULING system calls for FreeBSD and Linux;
Add man pages for _POSIX_PRIORITY_SCHEDULING system calls;
Add options to LINT;
Minor fixes to P1003_1B code during testing.
Fix for RTPRIO scheduler to eliminate invalid context switches.
POSIX.4 headers and sysctl variables. Nothing should change
unless POSIX4 is defined or _POSIX_VERSION is set to 199309.
