xref: /illumos-gate/usr/src/cmd/intrd/intrd.pl (revision 657a8c206b913d1ee578fd725f0b25eca5b77253)
1#!/usr/perl5/bin/perl
2#
3# CDDL HEADER START
4#
5# The contents of this file are subject to the terms of the
6# Common Development and Distribution License (the "License").
7# You may not use this file except in compliance with the License.
8#
9# You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10# or http://www.opensolaris.org/os/licensing.
11# See the License for the specific language governing permissions
12# and limitations under the License.
13#
14# When distributing Covered Code, include this CDDL HEADER in each
15# file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16# If applicable, add the following below this CDDL HEADER, with the
17# fields enclosed by brackets "[]" replaced with your own identifying
18# information: Portions Copyright [yyyy] [name of copyright owner]
19#
20# CDDL HEADER END
21#
22
23#
24# Copyright 2008 Sun Microsystems, Inc.  All rights reserved.
25# Use is subject to license terms.
26#
27
28require 5.8.4;
29use strict;
30use warnings;
31use POSIX;
32use File::Basename("basename");
33
34my $cmdname = basename($0);
35
36my $using_scengen = 0;	# 1 if using scenario simulator
37my $debug = 0;
38
39my $normal_sleeptime = 10;		# time to sleep between samples
40my $idle_sleeptime = 45;		# time to sleep when idle
41my $onecpu_sleeptime = (60 * 15);	# used if only 1 CPU on system
42my $sleeptime = $normal_sleeptime;	# either normal_ or idle_ or onecpu_
43
44my $idle_intrload = .1;			# idle if interrupt load < 10%
45
46my $timerange_toohi    = .01;
47my $statslen = 60;	# time period (in secs) to keep in @deltas
48
49
50# Parse arguments. intrd does not accept any public arguments; the two
51# arguments below are meant for testing purposes. -D generates a significant
52# amount of syslog output. -S <filename> loads the filename as a perl
53# script. That file is expected to implement a kstat "simulator" which
54# can be used to feed information to intrd and verify intrd's responses.
55
56while ($_ = shift @ARGV) {
57	if ($_ eq "-S" && $#ARGV != -1) {
58		$using_scengen = 1;
59		do $ARGV[0];	# load simulator
60		shift @ARGV;
61	} elsif ($_ eq "-D") {
62		$debug = 1;
63	}
64}
65
66if ($using_scengen == 0) {
67	require Sun::Solaris::Kstat;
68	require Sun::Solaris::Intrs;
69	import Sun::Solaris::Intrs(qw(intrmove is_pcplusmp));
70	require Sys::Syslog;
71	import Sys::Syslog;
72	openlog($cmdname, 'pid', 'daemon');
73	setlogmask(Sys::Syslog::LOG_UPTO($debug > 0 ? &Sys::Syslog::LOG_DEBUG :
74	    &Sys::Syslog::LOG_INFO));
75}
76
77my $asserted = 0;
78my $assert_level = 'debug';	# syslog level for assertion failures
79sub VERIFY($@)
80{
81	my $bad = (shift() == 0);	# $_[0] == 0 means assert failed
82	if ($bad) {
83		my $msg = shift();
84		syslog($assert_level, "VERIFY: $msg", @_);
85		$asserted++;
86	}
87	return ($bad);
88}
89
90
91
92
93sub getstat($$);
94sub generate_delta($$);
95sub compress_deltas($);
96sub dumpdelta($);
97
98sub goodness($);
99sub imbalanced($$);
100sub do_reconfig($);
101
102sub goodness_cpu($$);		# private function
103sub move_intr($$$$);		# private function
104sub ivecs_to_string(@);		# private function
105sub do_find_goal($$$$);		# private function
106sub find_goal($$);		# private function
107sub do_reconfig_cpu2cpu($$$$);	# private function
108sub do_reconfig_cpu($$$);	# private function
109
110
111#
112# What follow are the basic data structures routines of intrd.
113#
114# getstat() is responsible for reading the kstats and generating a "stat" hash.
115#
116# generate_delta() is responsible for taking two "stat" hashes and creating
117# a new "delta" hash that represents what has changed over time.
118#
119# compress_deltas() is responsible for taking a list of deltas and generating
120# a single delta hash that encompasses all the time periods described by the
121# deltas.
122
123
124#
125# getstat() is handed a reference to a kstat and generates a hash, returned
126# by reference, containing all the fields from the kstats which we need.
127# If it returns the scalar 0, it failed to gather the kstats, and the caller
128# should react accordingly.
129#
130# getstat() is also responsible for maintaining a reasonable $sleeptime.
131#
132# {"snaptime"}          kstat's snaptime
133# {<cpuid>}             one hash reference per online cpu
134#  ->{"tot"}            == cpu:<cpuid>:sys:cpu_nsec_{user + kernel + idle}
135#  ->{"crtime"}         == cpu:<cpuid>:sys:crtime
136#  ->{"ivecs"}
137#     ->{<cookie#>}     iterates over pci_intrs::<nexus>:cookie
138#        ->{"time"}     == pci_intrs:<ivec#>:<nexus>:time (in nsec)
139#        ->{"pil"}      == pci_intrs:<ivec#>:<nexus>:pil
140#        ->{"crtime"}   == pci_intrs:<ivec#>:<nexus>:crtime
141#        ->{"ino"}      == pci_intrs:<ivec#>:<nexus>:ino
142#        ->{"num_ino"}  == num inos of single device instance sharing this entry
143#				Will be > 1 on pcplusmp X86 systems for devices
144#				with multiple MSI interrupts.
145#        ->{"buspath"}  == pci_intrs:<ivec#>:<nexus>:buspath
146#        ->{"name"}     == pci_intrs:<ivec#>:<nexus>:name
147#        ->{"ihs"}      == pci_intrs:<ivec#>:<nexus>:ihs
148#
149
150sub getstat($$)
151{
152	my ($ks, $pcplusmp_sys) = @_;
153
154	my $cpucnt = 0;
155	my %stat = ();
156	my ($minsnap, $maxsnap);
157
158	# Hash of hash which matches (MSI device, ino) combos to kstats.
159	my %msidevs = ();
160
161	# kstats are not generated atomically. Each kstat hierarchy will
162	# have been generated within the kernel at a different time. On a
163	# thrashing system, we may not run quickly enough in order to get
164	# coherent kstat timing information across all the kstats. To
165	# determine if this is occurring, $minsnap/$maxsnap are used to
166	# find the breadth between the first and last snaptime of all the
167	# kstats we access. $maxsnap - $minsnap roughly represents the
168	# total time taken up in getstat(). If this time approaches the
169	# time between snapshots, our results may not be useful.
170
171	$minsnap = -1;		# snaptime is always a positive number
172	$maxsnap = $minsnap;
173
174	# Iterate over the cpus in cpu:<cpuid>::. Check
175	# cpu_info:<cpuid>:cpu_info<cpuid>:state to make sure the
176	# processor is "on-line". If not, it isn't accepting interrupts
177	# and doesn't concern us.
178	#
179	# Record cpu:<cpuid>:sys:snaptime, and check $minsnap/$maxsnap.
180
181	while (my ($cpu, $cpst) = each %{$ks->{cpu}}) {
182		next if !exists($ks->{cpu_info}{$cpu}{"cpu_info$cpu"}{state});
183		#"state" fld of kstat w/
184		#		  modname    inst name-"cpuinfo0"
185		my $state = $ks->{cpu_info}{$cpu}{"cpu_info$cpu"}{state};
186		next if ($state !~ /^on-line\0/);
187		my $cpu_sys = $cpst->{sys};
188
189		$stat{$cpu}{tot} = ($cpu_sys->{cpu_nsec_idle} +
190				    $cpu_sys->{cpu_nsec_user} +
191				    $cpu_sys->{cpu_nsec_kernel});
192		$stat{$cpu}{crtime} = $cpu_sys->{crtime};
193		$stat{$cpu}{ivecs} = {};
194
195		if ($minsnap == -1 || $cpu_sys->{snaptime} < $minsnap) {
196			$minsnap = $cpu_sys->{snaptime};
197		}
198		if ($cpu_sys->{snaptime} > $maxsnap) {
199			$maxsnap = $cpu_sys->{snaptime};
200		}
201		$cpucnt++;
202	}
203
204	if ($cpucnt <= 1) {
205		$sleeptime = $onecpu_sleeptime;
206		return (0);	# nothing to do with 1 CPU
207	}
208
209	# Iterate over the ivecs. If the cpu is not on-line, ignore the
210	# ivecs mapped to it, if any.
211	#
212	# Record pci_intrs:{inum}:<nexus>:time, snaptime, crtime, pil,
213	# ino, name, and buspath. Check $minsnap/$maxsnap.
214
215	foreach my $inst (values(%{$ks->{pci_intrs}})) {
216		my $intrcfg = (values(%$inst))[0];
217		my $cpu = $intrcfg->{cpu};
218
219		next unless exists $stat{$cpu};
220		next if ($intrcfg->{type} =~ /^disabled\0/);
221
222		# Perl looks beyond NULL chars in pattern matching.
223		# Truncate name field at the first NULL
224		$intrcfg->{name} =~ s/\0.*$//;
225
226		if ($intrcfg->{snaptime} < $minsnap) {
227			$minsnap = $intrcfg->{snaptime};
228		} elsif ($intrcfg->{snaptime} > $maxsnap) {
229			$maxsnap = $intrcfg->{snaptime};
230		}
231
232		my $cookie = "$intrcfg->{buspath} $intrcfg->{ino}";
233		if (exists $stat{$cpu}{ivecs}{$cookie}) {
234			my $cookiestats = $stat{$cpu}{ivecs}{$cookie};
235
236			$cookiestats->{time} += $intrcfg->{time};
237			$cookiestats->{name} .= "/$intrcfg->{name}";
238
239			# If this new interrupt sharing $cookie represents a
240			# change from an earlier getstat, make sure that
241			# generate_delta will see the change by setting
242			# crtime to the most recent crtime of its components.
243
244			if ($intrcfg->{crtime} > $cookiestats->{crtime}) {
245				$cookiestats->{crtime} = $intrcfg->{crtime};
246			}
247			$cookiestats->{ihs}++;
248			next;
249		}
250		$stat{$cpu}{ivecs}{$cookie}{time} = $intrcfg->{time};
251		$stat{$cpu}{ivecs}{$cookie}{crtime} = $intrcfg->{crtime};
252		$stat{$cpu}{ivecs}{$cookie}{pil} = $intrcfg->{pil};
253		$stat{$cpu}{ivecs}{$cookie}{ino} = $intrcfg->{ino};
254		$stat{$cpu}{ivecs}{$cookie}{num_ino} = 1;
255		$stat{$cpu}{ivecs}{$cookie}{buspath} = $intrcfg->{buspath};
256		$stat{$cpu}{ivecs}{$cookie}{name} = $intrcfg->{name};
257		$stat{$cpu}{ivecs}{$cookie}{ihs} = 1;
258
259		if ($pcplusmp_sys && ($intrcfg->{type} =~ /^msi\0/)) {
260			if (!(exists($msidevs{$intrcfg->{name}}))) {
261				$msidevs{$intrcfg->{name}} = {};
262			}
263			$msidevs{$intrcfg->{name}}{$intrcfg->{ino}} =
264			    \$stat{$cpu}{ivecs}{$cookie};
265		}
266	}
267
268	# All MSI interrupts of a device instance share a single MSI address.
269	# On X86 systems with an APIC, this MSI address is interpreted as CPU
270	# routing info by the APIC.  For this reason, on these platforms, all
271	# interrupts for MSI devices must be moved to the same CPU at the same
272	# time.
273	#
274	# Since all interrupts will be on the same CPU on these platforms, all
275	# interrupts can be consolidated into one ivec entry.  For such devices,
276	# num_ino will be > 1 to denote that a group move is needed.
277
278	# Loop thru all MSI devices on X86 pcplusmp systems.
279	# Nop on other systems.
280	foreach my $msidevkey (sort keys %msidevs) {
281
282		# Loop thru inos of the device, sorted by lowest value first
283		# For each cookie found for a device, incr num_ino for the
284		# lowest cookie and remove other cookies.
285
286		# Assumes PIL is the same for first and current cookies
287
288		my $first_ino = -1;
289		my $first_cookiep;
290		my $curr_cookiep;
291		foreach my $inokey (sort keys %{$msidevs{$msidevkey}}) {
292			$curr_cookiep = $msidevs{$msidevkey}{$inokey};
293			if ($first_ino == -1) {
294				$first_ino = $inokey;
295				$first_cookiep = $curr_cookiep;
296			} else {
297				$$first_cookiep->{num_ino}++;
298				$$first_cookiep->{time} +=
299				    $$curr_cookiep->{time};
300				if ($$curr_cookiep->{crtime} >
301				    $$first_cookiep->{crtime}) {
302					$$first_cookiep->{crtime} =
303					    $$curr_cookiep->{crtime};
304				}
305				# Invalidate this cookie, less complicated and
306				# more efficient than deleting it.
307				$$curr_cookiep->{num_ino} = 0;
308			}
309		}
310	}
311
312	# We define the timerange as the amount of time spent gathering the
313	# various kstats, divided by our sleeptime. If we take a lot of time
314	# to access the kstats, and then we create a delta comparing these
315	# kstats with a prior set of kstats, that delta will cover
316	# substaintially different amount of time depending upon which
317	# interrupt or CPU is being examined.
318	#
319	# By checking the timerange here, we guarantee that any deltas
320	# created from these kstats will contain self-consistent data,
321	# in that all CPUs and interrupts cover a similar span of time.
322	#
323	# $timerange_toohi is the upper bound. Any timerange above
324	# this is thrown out as garbage. If the stat is safely within this
325	# bound, we treat the stat as representing an instant in time, rather
326	# than the time range it actually spans. We arbitrarily choose minsnap
327	# as the snaptime of the stat.
328
329	$stat{snaptime} = $minsnap;
330	my $timerange = ($maxsnap - $minsnap) / $sleeptime;
331	return (0) if ($timerange > $timerange_toohi);	# i.e. failure
332	return (\%stat);
333}
334
335#
336# dumpdelta takes a reference to our "delta" structure:
337# {"missing"}           "1" if the delta's component stats had inconsistencies
338# {"minsnap"}           time of the first kstat snaptime used in this delta
339# {"maxsnap"}           time of the last kstat snaptime used in this delta
340# {"goodness"}          cost function applied to this delta
341# {"avgintrload"}       avg of interrupt load across cpus, as a percentage
342# {"avgintrnsec"}       avg number of nsec spent in interrupts, per cpu
343# {<cpuid>}             iterates over on-line cpus
344#  ->{"intrs"}          cpu's movable intr time (sum of "time" for each ivec)
345#  ->{"tot"}            CPU load from all sources in nsec
346#  ->{"bigintr"}        largest value of {ivecs}{<ivec#>}{time} from below
347#  ->{"intrload"}       intrs / tot
348#  ->{"ivecs"}
349#     ->{<ivec#>}       iterates over ivecs for this cpu
350#        ->{"time"}     time used by this interrupt (in nsec)
351#        ->{"pil"}      pil level of this interrupt
352#        ->{"ino"}      interrupt number (or base vector if MSI group)
353#        ->{"buspath"}  filename of the directory of the device's bus
354#        ->{"name"}     device name
355#        ->{"ihs"}      number of different handlers sharing this ino
356#        ->{"num_ino"}  number of interrupt vectors in MSI group
357#
358# It prints out the delta structure in a nice, human readable display.
359#
360
361sub dumpdelta($)
362{
363	my ($delta) = @_;
364
365	# print global info
366
367	syslog('debug', "dumpdelta:");
368	syslog('debug', " RECONFIGURATION IN DELTA") if $delta->{missing} > 0;
369	syslog('debug', " avgintrload: %5.2f%%  avgintrnsec: %d",
370	       $delta->{avgintrload} * 100, $delta->{avgintrnsec});
371	syslog('debug', "    goodness: %5.2f%%", $delta->{goodness} * 100)
372	    if exists($delta->{goodness});
373
374	# iterate over cpus
375
376	while (my ($cpu, $cpst) = each %$delta) {
377		next if !ref($cpst);		# skip non-cpuid entries
378		my $tot = $cpst->{tot};
379		syslog('debug', "    cpu %3d intr %7.3f%%  (bigintr %7.3f%%)",
380		       $cpu, $cpst->{intrload}*100, $cpst->{bigintr}*100/$tot);
381		syslog('debug', "        intrs %d, bigintr %d",
382		       $cpst->{intrs}, $cpst->{bigintr});
383
384		# iterate over ivecs on this cpu
385
386		while (my ($ivec, $ivst) = each %{$cpst->{ivecs}}) {
387			syslog('debug', "    %15s:\"%s\": %7.3f%%  %d",
388			    ($ivst->{ihs} > 1 ? "$ivst->{name}($ivst->{ihs})" :
389			    $ivst->{name}), $ivec,
390			    $ivst->{time}*100 / $tot, $ivst->{time});
391		}
392	}
393}
394
395#
396# generate_delta($stat, $newstat) takes two stat references, returned from
397# getstat(), and creates a %delta. %delta (not surprisingly) contains the
398# same basic info as stat and newstat, but with the timestamps as deltas
399# instead of absolute times. We return a reference to the delta.
400#
401
402sub generate_delta($$)
403{
404	my ($stat, $newstat) = @_;
405
406	my %delta = ();
407	my $intrload;
408	my $intrnsec;
409	my $cpus;
410
411	# Take the worstcase timerange
412	$delta{minsnap} = $stat->{snaptime};
413	$delta{maxsnap} = $newstat->{snaptime};
414	if (VERIFY($delta{maxsnap} > $delta{minsnap},
415	    "generate_delta: stats aren't ascending")) {
416		$delta{missing} = 1;
417		return (\%delta);
418	}
419
420	# if there are a different number of cpus in the stats, set missing
421
422	$delta{missing} = (keys(%$stat) != keys(%$newstat));
423	if (VERIFY($delta{missing} == 0,
424	    "generate_delta: number of CPUs changed")) {
425		return (\%delta);
426	}
427
428	# scan through every cpu in %newstat and compare against %stat
429
430	while (my ($cpu, $newcpst) = each %$newstat) {
431		next if !ref($newcpst);		# skip non-cpuid fields
432
433		# If %stat is missing a cpu from %newstat, then it was just
434		# onlined. Mark missing.
435
436		if (VERIFY(exists $stat->{$cpu} &&
437		    $stat->{$cpu}{crtime} == $newcpst->{crtime},
438		    "generate_delta: cpu $cpu changed")) {
439			$delta{missing} = 1;
440			return (\%delta);
441		}
442		my $cpst = $stat->{$cpu};
443		$delta{$cpu}{tot} = $newcpst->{tot} - $cpst->{tot};
444		if (VERIFY($delta{$cpu}{tot} >= 0,
445		    "generate_delta: deltas are not ascending?")) {
446			$delta{missing} = 1;
447			delete($delta{$cpu});
448			return (\%delta);
449		}
450		# Avoid remote chance of division by zero
451		$delta{$cpu}{tot} = 1 if $delta{$cpu}{tot} == 0;
452		$delta{$cpu}{intrs} = 0;
453		$delta{$cpu}{bigintr} = 0;
454
455		my %ivecs = ();
456		$delta{$cpu}{ivecs} = \%ivecs;
457
458		# if the number of ivecs differs, set missing
459
460		if (VERIFY(keys(%{$cpst->{ivecs}}) ==
461			   keys(%{$newcpst->{ivecs}}),
462			   "generate_delta: cpu $cpu has more/less".
463			   " interrupts")) {
464			$delta{missing} = 1;
465			return (\%delta);
466		}
467
468		while (my ($inum, $newivec) = each %{$newcpst->{ivecs}}) {
469
470			# Unused cookie, corresponding to an MSI vector which
471			# is part of a group.  The whole group is accounted for
472			# by a different cookie.
473			next if ($newivec->{num_ino} == 0);
474
475			# If this ivec doesn't exist in $stat, or if $stat
476			# shows a different crtime, set missing.
477			if (VERIFY(exists $cpst->{ivecs}{$inum} &&
478				   $cpst->{ivecs}{$inum}{crtime} ==
479				   $newivec->{crtime},
480				   "generate_delta: cpu $cpu inum $inum".
481				   " has changed")) {
482				$delta{missing} = 1;
483				return (\%delta);
484			}
485			my $ivec = $cpst->{ivecs}{$inum};
486
487			# Create $delta{$cpu}{ivecs}{$inum}.
488
489			my %dltivec = ();
490			$delta{$cpu}{ivecs}{$inum} = \%dltivec;
491
492			# calculate time used by this interrupt
493
494			my $time = $newivec->{time} - $ivec->{time};
495			if (VERIFY($time >= 0,
496				   "generate_delta: ivec went backwards?")) {
497				$delta{missing} = 1;
498				delete($delta{$cpu}{ivecs}{$inum});
499				return (\%delta);
500			}
501			$delta{$cpu}{intrs} += $time;
502			$dltivec{time} = $time;
503			if ($time > $delta{$cpu}{bigintr}) {
504				$delta{$cpu}{bigintr} = $time;
505			}
506
507			# Transfer over basic info about the kstat. We
508			# don't have to worry about discrepancies between
509			# ivec and newivec because we verified that both
510			# have the same crtime.
511
512			$dltivec{pil} = $newivec->{pil};
513			$dltivec{ino} = $newivec->{ino};
514			$dltivec{buspath} = $newivec->{buspath};
515			$dltivec{name} = $newivec->{name};
516			$dltivec{ihs} = $newivec->{ihs};
517			$dltivec{num_ino} = $newivec->{num_ino};
518		}
519		if ($delta{$cpu}{tot} < $delta{$cpu}{intrs}) {
520			# Ewww! Hopefully just a rounding error.
521			# Make something up.
522			$delta{$cpu}{tot} = $delta{$cpu}{intrs};
523		}
524		$delta{$cpu}{intrload} =
525		       $delta{$cpu}{intrs} / $delta{$cpu}{tot};
526		$intrload += $delta{$cpu}{intrload};
527		$intrnsec += $delta{$cpu}{intrs};
528		$cpus++;
529	}
530	if ($cpus > 0) {
531		$delta{avgintrload} = $intrload / $cpus;
532		$delta{avgintrnsec} = $intrnsec / $cpus;
533	} else {
534		$delta{avgintrload} = 0;
535		$delta{avgintrnsec} = 0;
536	}
537	return (\%delta);
538}
539
540
541# compress_delta takes a list of deltas, and returns a single new delta
542# which represents the combined information from all the deltas. The deltas
543# provided are assumed to be sequential in time. The resulting compressed
544# delta looks just like any other delta. This new delta is also more accurate
545# since its statistics are averaged over a longer period than any of the
546# original deltas.
547
548sub compress_deltas ($)
549{
550	my ($deltas) = @_;
551
552	my %newdelta = ();
553	my ($intrs, $tot);
554	my $cpus = 0;
555	my ($high_intrload) = 0;
556
557	if (VERIFY($#$deltas != -1,
558		   "compress_deltas: list of delta is empty?")) {
559		return (0);
560	}
561	$newdelta{minsnap} = $deltas->[0]{minsnap};
562	$newdelta{maxsnap} = $deltas->[$#$deltas]{maxsnap};
563	$newdelta{missing} = 0;
564
565	foreach my $delta (@$deltas) {
566		if (VERIFY($delta->{missing} == 0,
567		    "compressing bad deltas?")) {
568			return (0);
569		}
570		while (my ($cpuid, $cpu) = each %$delta) {
571			next if !ref($cpu);
572
573			$intrs += $cpu->{intrs};
574			$tot += $cpu->{tot};
575			$newdelta{$cpuid}{intrs} += $cpu->{intrs};
576			$newdelta{$cpuid}{tot} += $cpu->{tot};
577			if (!exists $newdelta{$cpuid}{ivecs}) {
578				my %ivecs = ();
579				$newdelta{$cpuid}{ivecs} = \%ivecs;
580			}
581			while (my ($inum, $ivec) = each %{$cpu->{ivecs}}) {
582				my $newivecs = $newdelta{$cpuid}{ivecs};
583				$newivecs->{$inum}{time} += $ivec->{time};
584				$newivecs->{$inum}{pil} = $ivec->{pil};
585				$newivecs->{$inum}{ino} = $ivec->{ino};
586				$newivecs->{$inum}{buspath} = $ivec->{buspath};
587				$newivecs->{$inum}{name} = $ivec->{name};
588				$newivecs->{$inum}{ihs} = $ivec->{ihs};
589				$newivecs->{$inum}{num_ino} = $ivec->{num_ino};
590			}
591		}
592	}
593	foreach my $cpu (values(%newdelta)) {
594		next if !ref($cpu); # ignore non-cpu fields
595		$cpus++;
596
597		my $bigintr = 0;
598		foreach my $ivec (values(%{$cpu->{ivecs}})) {
599			if ($ivec->{time} > $bigintr) {
600				$bigintr = $ivec->{time};
601			}
602		}
603		$cpu->{bigintr} = $bigintr;
604		$cpu->{intrload} = $cpu->{intrs} / $cpu->{tot};
605		if ($high_intrload < $cpu->{intrload}) {
606			$high_intrload = $cpu->{intrload};
607		}
608		$cpu->{tot} = 1 if $cpu->{tot} <= 0;
609	}
610	if ($cpus == 0) {
611		$newdelta{avgintrnsec} = 0;
612		$newdelta{avgintrload} = 0;
613	} else {
614		$newdelta{avgintrnsec} = $intrs / $cpus;
615		$newdelta{avgintrload} = $intrs / $tot;
616	}
617	$sleeptime = ($high_intrload < $idle_intrload) ? $idle_sleeptime :
618	    $normal_sleeptime;
619	return (\%newdelta);
620}
621
622
623
624
625
626# What follow are the core functions responsible for examining the deltas
627# generated above and deciding what to do about them.
628#
629# goodness() and its helper goodness_cpu() return a heuristic which describe
630# how good (or bad) the current interrupt balance is. The value returned will
631# be between 0 and 1, with 0 representing maximum goodness, and 1 representing
632# maximum badness.
633#
634# imbalanced() compares a current and historical value of goodness, and
635# determines if there has been enough change to warrant evaluating a
636# reconfiguration of the interrupts
637#
638# do_reconfig(), and its helpers, do_reconfig_cpu(), do_reconfig_cpu2cpu(),
639# find_goal(), do_find_goal(), and move_intr(), are responsible for examining
640# a delta and determining the best possible assignment of interrupts to CPUs.
641#
642# It is important that do_reconfig() be in alignment with goodness(). If
643# do_reconfig were to generate a new interrupt distribution that worsened
644# goodness, we could get into a pathological loop with intrd fighting itself,
645# constantly deciding that things are imbalanced, and then changing things
646# only to make them worse.
647
648
649
650# any goodness over $goodness_unsafe_load is considered really bad
651# goodness must drop by at least $goodness_mindelta for a reconfig
652
653my $goodness_unsafe_load = .9;
654my $goodness_mindelta = .1;
655
656# goodness(%delta) examines a delta and return its "goodness". goodness will
657# be between 0 (best) and 1 (major bad). goodness is determined by evaluating
658# the goodness of each individual cpu, and returning the worst case. This
659# helps on systems with many CPUs, where otherwise a single pathological CPU
660# might otherwise be ignored because the average was OK.
661#
662# To calculate the goodness of an individual CPU, we start by looking at its
663# load due to interrupts. If the load is above a certain high threshold and
664# there is more than one interrupt assigned to this CPU, we set goodness
665# to worst-case. If the load is below the average interrupt load of all CPUs,
666# then we return best-case, since what's to complain about?
667#
668# Otherwise we look at how much the load is above the average, and return
669# that as the goodness, with one caveat: we never return more than the CPU's
670# interrupt load ignoring its largest single interrupt source. This is
671# because a CPU with one high-load interrupt, and no other interrupts, is
672# perfectly balanced. Nothing can be done to improve the situation, and thus
673# it is perfectly balanced even if the interrupt's load is 100%.
674
675sub goodness($)
676{
677	my ($delta) = @_;
678
679	return (1) if $delta->{missing} > 0;
680
681	my $high_goodness = 0;
682	my $goodness;
683
684	foreach my $cpu (values(%$delta)) {
685		next if !ref($cpu);		# skip non-cpuid fields
686
687		$goodness = goodness_cpu($cpu, $delta->{avgintrload});
688		if (VERIFY($goodness >= 0 && $goodness <= 1,
689			   "goodness: cpu goodness out of range?")) {
690			dumpdelta($delta);
691			return (1);
692		}
693		if ($goodness == 1) {
694			return (1);	# worst case, no need to continue
695		}
696		if ($goodness > $high_goodness) {
697			$high_goodness = $goodness;
698		}
699	}
700	return ($high_goodness);
701}
702
703sub goodness_cpu($$)		# private function
704{
705	my ($cpu, $avgintrload) = @_;
706
707	my $goodness;
708	my $load = $cpu->{intrs} / $cpu->{tot};
709
710	return (0) if ($load < $avgintrload);	# low loads are perfectly good
711
712	# Calculate $load_no_bigintr, which represents the load
713	# due to interrupts, excluding the one biggest interrupt.
714	# This is the most gain we can get on this CPU from
715	# offloading interrupts.
716
717	my $load_no_bigintr = ($cpu->{intrs} - $cpu->{bigintr}) / $cpu->{tot};
718
719	# A major imbalance is indicated if a CPU is saturated
720	# with interrupt handling, and it has more than one
721	# source of interrupts. Those other interrupts could be
722	# starved if of a lower pil. Return a goodness of 1,
723	# which is the worst possible return value,
724	# which will effectively contaminate this entire delta.
725
726	my $cnt = keys(%{$cpu->{ivecs}});
727
728	if ($load > $goodness_unsafe_load && $cnt > 1) {
729		return (1);
730	}
731	$goodness = $load - $avgintrload;
732	if ($goodness > $load_no_bigintr) {
733		$goodness = $load_no_bigintr;
734	}
735	return ($goodness);
736}
737
738
739# imbalanced() is used by the main routine to determine if the goodness
740# has shifted far enough from our last baseline to warrant a reassignment
741# of interrupts. A very high goodness indicates that a CPU is way out of
742# whack. If the goodness has varied too much since the baseline, then
743# perhaps a reconfiguration is worth considering.
744
745sub imbalanced ($$)
746{
747	my ($goodness, $baseline) = @_;
748
749	# Return 1 if we are pathological, or creeping away from the baseline
750
751	return (1) if $goodness > .50;
752	return (1) if abs($goodness - $baseline) > $goodness_mindelta;
753	return (0);
754}
755
756# do_reconfig(), do_reconfig_cpu(), and do_reconfig_cpu2cpu(), are the
757# decision-making functions responsible for generating a new interrupt
758# distribution. They are designed with the definition of goodness() in
759# mind, i.e. they use the same definition of "good distribution" as does
760# goodness().
761#
762# do_reconfig() is responsible for deciding whether a redistribution is
763# actually warranted. If the goodness is already pretty good, it doesn't
764# waste the CPU time to generate a new distribution. If it
765# calculates a new distribution and finds that it is not sufficiently
766# improved from the prior distirbution, it will not do the redistribution,
767# mainly to avoid the disruption to system performance caused by
768# rejuggling interrupts.
769#
770# Its main loop works by going through a list of cpus sorted from
771# highest to lowest interrupt load. It removes the highest-load cpus
772# one at a time and hands them off to do_reconfig_cpu(). This function
773# then re-sorts the remaining CPUs from lowest to highest interrupt load,
774# and one at a time attempts to rejuggle interrupts between the original
775# high-load CPU and the low-load CPU. Rejuggling on a high-load CPU is
776# considered finished as soon as its interrupt load is within
777# $goodness_mindelta of the average interrupt load. Such a CPU will have
778# a goodness of below the $goodness_mindelta threshold.
779
780#
781# move_intr(\%delta, $inum, $oldcpu, $newcpu)
782# used by reconfiguration code to move an interrupt between cpus within
783# a delta. This manipulates data structures, and does not actually move
784# the interrupt on the running system.
785#
786sub move_intr($$$$)		# private function
787{
788	my ($delta, $inum, $oldcpuid, $newcpuid) = @_;
789
790	my $ivec = $delta->{$oldcpuid}{ivecs}{$inum};
791
792	# Remove ivec from old cpu
793
794	my $oldcpu = $delta->{$oldcpuid};
795	$oldcpu->{intrs} -= $ivec->{time};
796	$oldcpu->{intrload} = $oldcpu->{intrs} / $oldcpu->{tot};
797	delete($oldcpu->{ivecs}{$inum});
798
799	VERIFY($oldcpu->{intrs} >= 0, "move_intr: intr's time > total time?");
800	VERIFY($ivec->{time} <= $oldcpu->{bigintr},
801	       "move_intr: intr's time > bigintr?");
802
803	if ($ivec->{time} >= $oldcpu->{bigintr}) {
804		my $bigtime = 0;
805
806		foreach my $ivec (values(%{$oldcpu->{ivecs}})) {
807			$bigtime = $ivec->{time} if $ivec->{time} > $bigtime;
808		}
809		$oldcpu->{bigintr} = $bigtime;
810	}
811
812	# Add ivec onto new cpu
813
814	my $newcpu = $delta->{$newcpuid};
815
816	$ivec->{nowcpu} = $newcpuid;
817	$newcpu->{intrs} += $ivec->{time};
818	$newcpu->{intrload} = $newcpu->{intrs} / $newcpu->{tot};
819	$newcpu->{ivecs}{$inum} = $ivec;
820
821	$newcpu->{bigintr} = $ivec->{time}
822		if $ivec->{time} > $newcpu->{bigintr};
823}
824
825sub move_intr_check($$$)	# private function
826{
827	my ($delta, $oldcpuid, $newcpuid) = @_;
828
829	VERIFY($delta->{$oldcpuid}{tot} >= $delta->{$oldcpuid}{intrs},
830	       "Moved interrupts left 100+%% load on src cpu");
831	VERIFY($delta->{$newcpuid}{tot} >= $delta->{$newcpuid}{intrs},
832	       "Moved interrupts left 100+%% load on tgt cpu");
833}
834
835sub ivecs_to_string(@)		# private function
836{
837	my $str = "";
838	foreach my $ivec (@_) {
839		$str = "$str $ivec->{inum}";
840	}
841	return ($str);
842}
843
844
845sub do_reconfig($)
846{
847	my ($delta) = @_;
848
849	my $goodness = $delta->{goodness};
850
851	# We can't improve goodness to better than 0. We should stop here
852	# if, even if we achieve a goodness of 0, the improvement is still
853	# too small to merit the action.
854
855	if ($goodness - 0 < $goodness_mindelta) {
856		syslog('debug', "goodness good enough, don't reconfig");
857		return (0);
858	}
859
860	syslog('notice', "Optimizing interrupt assignments");
861
862	if (VERIFY ($delta->{missing} == 0, "RECONFIG Aborted: should not ".
863	    "have a delta with missing")) {
864		return (-1);
865	}
866
867	# Make a list of all cpuids, and also add some extra information
868	# to the ivec structures.
869
870	my @cpusortlist = ();
871
872	while (my ($cpuid, $cpu) = each %$delta) {
873		next if !ref($cpu);	# skip non-cpu entries
874
875		push(@cpusortlist, $cpuid);
876		while (my ($inum, $ivec) = each %{$cpu->{ivecs}}) {
877			$ivec->{origcpu} = $cpuid;
878			$ivec->{nowcpu} = $cpuid;
879			$ivec->{inum} = $inum;
880		}
881	}
882
883	# Sort the list of CPUs from highest to lowest interrupt load.
884	# Remove the top CPU from that list and attempt to redistribute
885	# its interrupts. If the CPU has a goodness below a threshold,
886	# just ignore the CPU and move to the next one. If the CPU's
887	# load falls below the average load plus that same threshold,
888	# then there are no CPUs left worth reconfiguring, and we're done.
889
890	while (@cpusortlist) {
891		# Re-sort cpusortlist each time, since do_reconfig_cpu can
892		# move interrupts around.
893
894		@cpusortlist =
895		    sort({$delta->{$b}{intrload} <=> $delta->{$a}{intrload}}
896		    @cpusortlist);
897
898		my $cpu = shift(@cpusortlist);
899		if (($delta->{$cpu}{intrload} <= $goodness_unsafe_load) &&
900		    ($delta->{$cpu}{intrload} <=
901		    $delta->{avgintrload} + $goodness_mindelta)) {
902			syslog('debug', "finished reconfig: cpu $cpu load ".
903			    "$delta->{$cpu}{intrload} avgload ".
904			    "$delta->{avgintrload}");
905			last;
906		}
907		if (goodness_cpu($delta->{$cpu}, $delta->{avgintrload}) <
908		    $goodness_mindelta) {
909			next;
910		}
911		do_reconfig_cpu($delta, \@cpusortlist, $cpu);
912	}
913
914	# How good a job did we do? If the improvement was minimal, and
915	# our goodness wasn't pathological (and thus needing any help it
916	# can get), then don't bother moving the interrupts.
917
918	my $newgoodness = goodness($delta);
919	VERIFY($newgoodness <= $goodness,
920	       "reconfig: result has worse goodness?");
921
922	if (($goodness != 1 || $newgoodness == 1) &&
923	    $goodness - $newgoodness < $goodness_mindelta) {
924		syslog('debug', "goodness already near optimum, ".
925		       "don't reconfig");
926		return (0);
927	}
928	syslog('debug', "goodness %5.2f%% --> %5.2f%%", $goodness*100,
929	       $newgoodness*100);
930
931	# Time to move those interrupts!
932
933	my $ret = 1;
934	my $warned = 0;
935	while (my ($cpuid, $cpu) = each %$delta) {
936		next if $cpuid =~ /\D/;
937		while (my ($inum, $ivec) = each %{$cpu->{ivecs}}) {
938			next if ($ivec->{origcpu} == $cpuid);
939
940			if (!intrmove($ivec->{buspath}, $ivec->{ino},
941			    $cpuid, $ivec->{num_ino})) {
942				syslog('warning', "Unable to move interrupts")
943				    if $warned++ == 0;
944				syslog('debug', "Unable to move buspath ".
945				    "$ivec->{buspath} ino $ivec->{ino} to ".
946				    "cpu $cpuid");
947				$ret = -1;
948			}
949		}
950	}
951
952	syslog('notice', "Interrupt assignments optimized");
953	return ($ret);
954}
955
956sub do_reconfig_cpu($$$)	# private function
957{
958	my ($delta, $cpusortlist, $oldcpuid) = @_;
959
960	# We have been asked to rejuggle interrupts between $oldcpuid and
961	# other CPUs found on $cpusortlist so as to improve the load on
962	# $oldcpuid. We reverse $cpusortlist to get our own copy of the
963	# list, sorted from lowest to highest interrupt load. One at a
964	# time, shift a CPU off of this list of CPUs, and attempt to
965	# rejuggle interrupts between the two CPUs. Don't do this if the
966	# other CPU has a higher load than oldcpuid. We're done rejuggling
967	# once $oldcpuid's goodness falls below a threshold.
968
969	syslog('debug', "reconfiguring $oldcpuid");
970
971	my $cpu = $delta->{$oldcpuid};
972	my $avgintrload = $delta->{avgintrload};
973
974	my @cputargetlist = reverse(@$cpusortlist); # make a copy of the list
975	while ($#cputargetlist != -1) {
976 		last if goodness_cpu($cpu, $avgintrload) < $goodness_mindelta;
977
978		my $tgtcpuid = shift(@cputargetlist);
979		my $tgt = $delta->{$tgtcpuid};
980		my $load = $cpu->{intrload};
981		my $tgtload = $tgt->{intrload};
982		last if $tgtload > $load;
983		do_reconfig_cpu2cpu($delta, $oldcpuid, $tgtcpuid, $load);
984	}
985}
986
987sub do_reconfig_cpu2cpu($$$$)	# private function
988{
989	my ($delta, $srccpuid, $tgtcpuid, $srcload) = @_;
990
991	# We've been asked to consider interrupt juggling between srccpuid
992	# (with a high interrupt load) and tgtcpuid (with a lower interrupt
993	# load). First, make a single list with all of the ivecs from both
994	# CPUs, and sort the list from highest to lowest load.
995
996	syslog('debug', "exchanging intrs between $srccpuid and $tgtcpuid");
997
998	# Gather together all the ivecs and sort by load
999
1000	my @ivecs = (values(%{$delta->{$srccpuid}{ivecs}}),
1001	    values(%{$delta->{$tgtcpuid}{ivecs}}));
1002	return if $#ivecs == -1;
1003
1004	@ivecs = sort({$b->{time} <=> $a->{time}} @ivecs);
1005
1006	# Our "goal" load for srccpuid is the average load across all CPUs.
1007	# find_goal() will find determine the optimum selection of the
1008	# available interrupts which comes closest to this goal without
1009	# falling below the goal.
1010
1011	my $goal = $delta->{avgintrnsec};
1012
1013	# We know that the interrupt load on tgtcpuid is less than that on
1014	# srccpuid, but its load could still be above avgintrnsec. Don't
1015	# choose a goal which would bring srccpuid below the load on tgtcpuid.
1016
1017	my $avgnsec =
1018	    ($delta->{$srccpuid}{intrs} + $delta->{$tgtcpuid}{intrs}) / 2;
1019	if ($goal < $avgnsec) {
1020		$goal = $avgnsec;
1021	}
1022
1023	# If the largest of the interrupts is on srccpuid, leave it there.
1024	# This can help minimize the disruption caused by moving interrupts.
1025
1026	if ($ivecs[0]->{origcpu} == $srccpuid) {
1027		syslog('debug', "Keeping $ivecs[0]->{inum} on $srccpuid");
1028		$goal -= $ivecs[0]->{time};
1029		shift(@ivecs);
1030	}
1031
1032	syslog('debug', "GOAL: inums should total $goal");
1033	find_goal(\@ivecs, $goal);
1034
1035	# find_goal() returned its results to us by setting $ivec->{goal} if
1036	# the ivec should be on srccpuid, or clearing it for tgtcpuid.
1037	# Call move_intr() to update our $delta with the new results.
1038
1039	foreach my $ivec (@ivecs) {
1040		syslog('debug', "ivec $ivec->{inum} goal $ivec->{goal}");
1041		VERIFY($ivec->{nowcpu} == $srccpuid ||
1042		    $ivec->{nowcpu} == $tgtcpuid, "cpu2cpu found an ".
1043		    "interrupt not currently on src or tgt cpu");
1044
1045		if ($ivec->{goal} && $ivec->{nowcpu} != $srccpuid) {
1046			move_intr($delta, $ivec->{inum}, $ivec->{nowcpu},
1047			    $srccpuid);
1048		} elsif ($ivec->{goal} == 0 && $ivec->{nowcpu} != $tgtcpuid) {
1049			move_intr($delta, $ivec->{inum}, $ivec->{nowcpu},
1050			    $tgtcpuid);
1051		}
1052	}
1053	move_intr_check($delta, $srccpuid, $tgtcpuid); # asserts
1054
1055	my $newload = $delta->{$srccpuid}{intrs} / $delta->{$srccpuid}{tot};
1056	VERIFY($newload <= $srcload && $newload > $delta->{avgintrload},
1057	    "cpu2cpu: new load didn't end up in expected range");
1058}
1059
1060
1061# find_goal() and its helper do_find_goal() are used to find the best
1062# combination of interrupts in order to generate a load that is as close
1063# as possible to a goal load without falling below that goal. Before returning
1064# to its caller, find_goal() sets a new value in the hash of each interrupt,
1065# {goal}, which if set signifies that this interrupt is one of the interrupts
1066# identified as part of the set of interrupts which best meet the goal.
1067#
1068# The arguments to find_goal are a list of ivecs (hash references), sorted
1069# by descending {time}, and the goal load. The goal is relative to {time}.
1070# The best fit is determined by performing a depth-first search. do_find_goal
1071# is the recursive subroutine which carries out the search.
1072#
1073# It is passed an index as an argument, originally 0. On a given invocation,
1074# it is only to consider interrupts in the ivecs array starting at that index.
1075# It then considers two possibilities:
1076#   1) What is the best goal-fit if I include ivecs[index]?
1077#   2) What is the best goal-fit if I exclude ivecs[index]?
1078# To determine case 1, it subtracts the load of ivecs[index] from the goal,
1079# and calls itself recursively with that new goal and index++.
1080# To determine case 2, it calls itself recursively with the same goal and
1081# index++.
1082#
1083# It then compares the two results, decide which one best meets the goals,
1084# and returns the result. The return value is the best-fit's interrupt load,
1085# followed by a list of all the interrupts which make up that best-fit.
1086#
1087# As an optimization, a second array loads[] is created which mirrors ivecs[].
1088# loads[i] will equal the total loads of all ivecs[i..$#ivecs]. This is used
1089# by do_find_goal to avoid recursing all the way to the end of the ivecs
1090# array if including all remaining interrupts will still leave the best-fit
1091# at below goal load. If so, it then includes all remaining interrupts on
1092# the goal list and returns.
1093#
1094sub find_goal($$)		# private function
1095{
1096	my ($ivecs, $goal) = @_;
1097
1098	my @goals;
1099	my $load;
1100	my $ivec;
1101
1102	if ($goal <= 0) {
1103		@goals = ();	# the empty set will best meet the goal
1104	} else {
1105		syslog('debug', "finding goal from intrs %s",
1106		    ivecs_to_string(@$ivecs));
1107
1108		# Generate @loads array
1109
1110		my $tot = 0;
1111		foreach $ivec (@$ivecs) {
1112			$tot += $ivec->{time};
1113		}
1114		my @loads = ();
1115		foreach $ivec (@$ivecs) {
1116			push(@loads, $tot);
1117			$tot -= $ivec->{time};
1118		}
1119		($load, @goals) = do_find_goal($ivecs, \@loads, $goal, 0);
1120		VERIFY($load >= $goal, "find_goal didn't meet goals");
1121	}
1122	syslog('debug', "goals found: %s", ivecs_to_string(@goals));
1123
1124	# Set or clear $ivec->{goal} for each ivec, based on returned @goals
1125
1126	foreach $ivec (@$ivecs) {
1127		if ($#goals > -1 && $ivec == $goals[0]) {
1128			syslog('debug', "inum $ivec->{inum} on source cpu");
1129			$ivec->{goal} = 1;
1130			shift(@goals);
1131		} else {
1132			syslog('debug', "inum $ivec->{inum} on target cpu");
1133			$ivec->{goal} = 0;
1134		}
1135	}
1136}
1137
1138
1139sub do_find_goal($$$$)		# private function
1140{
1141	my ($ivecs, $loads, $goal, $idx) = @_;
1142
1143	if ($idx > $#{$ivecs}) {
1144		return (0);
1145	}
1146	syslog('debug', "$idx: finding goal $goal inum $ivecs->[$idx]{inum}");
1147
1148	my $load = $ivecs->[$idx]{time};
1149	my @goals_with = ();
1150	my @goals_without = ();
1151	my ($with, $without);
1152
1153	# If we include all remaining items and we're still below goal,
1154	# stop here. We can just return a result that includes $idx and all
1155	# subsequent ivecs. Since this will still be below goal, there's
1156	# nothing better to be done.
1157
1158	if ($loads->[$idx] <= $goal) {
1159		syslog('debug',
1160		    "$idx: including all remaining intrs %s with load %d",
1161		    ivecs_to_string(@$ivecs[$idx .. $#{$ivecs}]),
1162		    $loads->[$idx]);
1163		return ($loads->[$idx], @$ivecs[$idx .. $#{$ivecs}]);
1164	}
1165
1166	# Evaluate the "with" option, i.e. the best matching goal which
1167	# includes $ivecs->[$idx]. If idx's load is more than our goal load,
1168	# stop here. Once we're above the goal, there is no need to consider
1169	# further interrupts since they'll only take us further from the goal.
1170
1171	if ($goal <= $load) {
1172		$with = $load;	# stop here
1173	} else {
1174		($with, @goals_with) =
1175		    do_find_goal($ivecs, $loads, $goal - $load, $idx + 1);
1176		$with += $load;
1177	}
1178	syslog('debug', "$idx: with-load $with intrs %s",
1179	       ivecs_to_string($ivecs->[$idx], @goals_with));
1180
1181	# Evaluate the "without" option, i.e. the best matching goal which
1182	# excludes $ivecs->[$idx].
1183
1184	($without, @goals_without) =
1185	    &do_find_goal($ivecs, $loads, $goal, $idx + 1);
1186	syslog('debug', "$idx: without-load $without intrs %s",
1187	       ivecs_to_string(@goals_without));
1188
1189	# We now have our "with" and "without" options, and we choose which
1190	# best fits the goal. If one is greater than goal and the other is
1191	# below goal, we choose the one that is greater. If they are both
1192	# below goal, then we choose the one that is greater. If they are
1193	# both above goal, then we choose the smaller.
1194
1195	my $which;		# 0 == with, 1 == without
1196	if ($with >= $goal && $without < $goal) {
1197		$which = 0;
1198	} elsif ($with < $goal && $without >= $goal) {
1199		$which = 1;
1200	} elsif ($with >= $goal && $without >= $goal) {
1201		$which = ($without < $with);
1202	} else {
1203		$which = ($without > $with);
1204	}
1205
1206	# Return the load of our best case scenario, followed by all the ivecs
1207	# which compose that goal.
1208
1209	if ($which == 1) {	# without
1210		syslog('debug', "$idx: going without");
1211		return ($without, @goals_without);
1212	} else {
1213		syslog('debug', "$idx: going with");
1214		return ($with, $ivecs->[$idx], @goals_with);
1215	}
1216	# Not reached
1217}
1218
1219
1220
1221
1222syslog('debug', "intrd is starting".($debug ? " (debug)" : ""));
1223
1224my @deltas = ();
1225my $deltas_tottime = 0;		# sum of maxsnap-minsnap across @deltas
1226my $avggoodness;
1227my $baseline_goodness = 0;
1228my $compdelta;
1229
1230my $do_reconfig;
1231
1232# temp variables
1233my $goodness;
1234my $deltatime;
1235my $olddelta;
1236my $olddeltatime;
1237my $delta;
1238my $newstat;
1239my $below_statslen;
1240my $newtime;
1241my $ret;
1242
1243
1244my $gotsig = 0;
1245$SIG{INT} = sub { $gotsig = 1; };     # don't die in the middle of retargeting
1246$SIG{HUP} = $SIG{INT};
1247$SIG{TERM} = $SIG{INT};
1248
1249my $ks;
1250if ($using_scengen == 0) {
1251	$ks = Sun::Solaris::Kstat->new();
1252} else {
1253	$ks = myks_update();	# supplied by the simulator
1254}
1255
1256# If no pci_intrs kstats were found, we need to exit, but we can't because
1257# SMF will restart us and/or report an error to the administrator. But
1258# there's nothing an administrator can do. So print out a message for SMF
1259# logs and silently pause forever.
1260
1261if (!exists($ks->{pci_intrs})) {
1262	print STDERR "$cmdname: no interrupts were found; ".
1263	    "your PCI bus may not yet be supported\n";
1264	pause() while $gotsig == 0;
1265	exit 0;
1266}
1267
1268# See if this is a system with a pcplusmp APIC.
1269# Such systems will get special handling.
1270# Assume that if one bus has a pcplusmp APIC that they all do.
1271
1272# Get a list of pci_intrs kstats.
1273my @elem = values(%{$ks->{pci_intrs}});
1274my $elem0 = $elem[0];
1275my $elemval = (values(%$elem0))[0];
1276
1277# Use its buspath to query the system.  It is assumed that either all or none
1278# of the busses on a system are hosted by the pcplusmp APIC.
1279my $pcplusmp_sys = is_pcplusmp($elemval->{buspath});
1280
1281my $stat = getstat($ks, $pcplusmp_sys);
1282
1283for (;;) {
1284	sub clear_deltas {
1285		@deltas = ();
1286		$deltas_tottime = 0;
1287		$stat = 0;   # prevent next gen_delta() from setting {missing}
1288	}
1289
1290	# 1. Sleep, update the kstats, and save the new stats in $newstat.
1291
1292	exit 0 if $gotsig;		# if we got ^C / SIGTERM, exit
1293	if ($using_scengen == 0) {
1294		sleep($sleeptime);
1295		exit 0 if $gotsig;	# if we got ^C / SIGTERM, exit
1296		$ks->update();
1297	} else {
1298		$ks = myks_update();
1299	}
1300	$newstat = getstat($ks, $pcplusmp_sys);
1301
1302	# $stat or $newstat could be zero if they're uninitialized, or if
1303	# getstat() failed. If $stat is zero, move $newstat to $stat, sleep
1304	# and try again. If $newstat is zero, then we also sleep and try
1305	# again, hoping the problem will clear up.
1306
1307	next if (!ref $newstat);
1308	if (!ref $stat) {
1309		$stat = $newstat;
1310		next;
1311	}
1312
1313	# 2. Compare $newstat with the prior set of values, result in %$delta.
1314
1315	$delta = generate_delta($stat, $newstat);
1316	dumpdelta($delta) if $debug;	# Dump most recent stats to stdout.
1317	$stat = $newstat;	# The new stats now become the old stats.
1318
1319
1320	# 3. If $delta->{missing}, then there has been a reconfiguration of
1321	# either cpus or interrupts (probably both). We need to toss out our
1322	# old set of statistics and start from scratch.
1323	#
1324	# Also, if the delta covers a very long range of time, then we've
1325	# been experiencing a system overload that has resulted in intrd
1326	# not being allowed to run effectively for a while now. As above,
1327	# toss our old statistics and start from scratch.
1328
1329	$deltatime = $delta->{maxsnap} - $delta->{minsnap};
1330	if ($delta->{missing} > 0 || $deltatime > $statslen) {
1331		clear_deltas();
1332		syslog('debug', "evaluating interrupt assignments");
1333		next;
1334	}
1335
1336
1337	# 4. Incorporate new delta into the list of deltas, and associated
1338	# statistics. If we've just now received $statslen deltas, then it's
1339	# time to evaluate a reconfiguration.
1340
1341	$below_statslen = ($deltas_tottime < $statslen);
1342	$deltas_tottime += $deltatime;
1343	$do_reconfig = ($below_statslen && $deltas_tottime >= $statslen);
1344	push(@deltas, $delta);
1345
1346	# 5. Remove old deltas if total time is more than $statslen. We use
1347	# @deltas as a moving average of the last $statslen seconds. Shift
1348	# off the olders deltas, but only if that doesn't cause us to fall
1349	# below $statslen seconds.
1350
1351	while (@deltas > 1) {
1352		$olddelta = $deltas[0];
1353		$olddeltatime = $olddelta->{maxsnap} - $olddelta->{minsnap};
1354		$newtime = $deltas_tottime - $olddeltatime;
1355		last if ($newtime < $statslen);
1356
1357		shift(@deltas);
1358		$deltas_tottime = $newtime;
1359	}
1360
1361	# 6. The brains of the operation are here. First, check if we're
1362	# imbalanced, and if so set $do_reconfig. If $do_reconfig is set,
1363	# either because of imbalance or above in step 4, we evaluate a
1364	# new configuration.
1365	#
1366	# First, take @deltas and generate a single "compressed" delta
1367	# which summarizes them all. Pass that to do_reconfig and see
1368	# what it does with it:
1369	#
1370	# $ret == -1 : failure
1371	# $ret ==  0 : current config is optimal (or close enough)
1372	# $ret ==  1 : reconfiguration has occurred
1373	#
1374	# If $ret is -1 or 1, dump all our deltas and start from scratch.
1375	# Step 4 above will set do_reconfig soon thereafter.
1376	#
1377	# If $ret is 0, then nothing has happened because we're already
1378	# good enough. Set baseline_goodness to current goodness.
1379
1380	$compdelta = compress_deltas(\@deltas);
1381	if (VERIFY(ref($compdelta) eq "HASH", "couldn't compress deltas")) {
1382		clear_deltas();
1383		next;
1384	}
1385	$compdelta->{goodness} = goodness($compdelta);
1386	dumpdelta($compdelta) if $debug;
1387
1388	$goodness = $compdelta->{goodness};
1389	syslog('debug', "GOODNESS: %5.2f%%", $goodness * 100);
1390
1391	if ($deltas_tottime >= $statslen &&
1392	    imbalanced($goodness, $baseline_goodness)) {
1393		$do_reconfig = 1;
1394	}
1395
1396	if ($do_reconfig) {
1397		$ret = do_reconfig($compdelta);
1398
1399		if ($ret != 0) {
1400			clear_deltas();
1401			syslog('debug', "do_reconfig FAILED!") if $ret == -1;
1402		} else {
1403			syslog('debug', "setting new baseline of $goodness");
1404			$baseline_goodness = $goodness;
1405		}
1406	}
1407	syslog('debug', "---------------------------------------");
1408}
1409