xref: /illumos-gate/usr/src/cmd/intrd/intrd.pl (revision 2917a9c9)
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 2007 Sun Microsystems, Inc.  All rights reserved.
25# Use is subject to license terms.
26#
27#ident	"%Z%%M%	%I%	%E% SMI"
28#
29
30require 5.6.1;
31use strict;
32use warnings;
33use POSIX;
34use File::Basename("basename");
35
36my $cmdname = basename($0);
37
38my $using_scengen = 0;	# 1 if using scenario simulator
39my $debug = 0;
40
41my $normal_sleeptime = 10;		# time to sleep between samples
42my $idle_sleeptime = 45;		# time to sleep when idle
43my $onecpu_sleeptime = (60 * 15);	# used if only 1 CPU on system
44my $sleeptime = $normal_sleeptime;	# either normal_ or idle_ or onecpu_
45
46my $idle_intrload = .1;			# idle if interrupt load < 10%
47
48my $timerange_toohi    = .01;
49my $statslen = 60;	# time period (in secs) to keep in @deltas
50
51
52# Parse arguments. intrd does not accept any public arguments; the two
53# arguments below are meant for testing purposes. -D generates a significant
54# amount of syslog output. -S <filename> loads the filename as a perl
55# script. That file is expected to implement a kstat "simulator" which
56# can be used to feed information to intrd and verify intrd's responses.
57
58while ($_ = shift @ARGV) {
59	if ($_ eq "-S" && $#ARGV != -1) {
60		$using_scengen = 1;
61		do $ARGV[0];	# load simulator
62		shift @ARGV;
63	} elsif ($_ eq "-D") {
64		$debug = 1;
65	}
66}
67
68if ($using_scengen == 0) {
69	require Sun::Solaris::Kstat;
70	require Sun::Solaris::Intrs;
71	import Sun::Solaris::Intrs(qw(intrmove is_pcplusmp));
72	require Sys::Syslog;
73	import Sys::Syslog;
74	openlog($cmdname, 'pid', 'daemon');
75	setlogmask(Sys::Syslog::LOG_UPTO($debug > 0 ? &Sys::Syslog::LOG_DEBUG :
76	    &Sys::Syslog::LOG_INFO));
77}
78
79my $asserted = 0;
80my $assert_level = 'debug';	# syslog level for assertion failures
81sub VERIFY($@)
82{
83	my $bad = (shift() == 0);	# $_[0] == 0 means assert failed
84	if ($bad) {
85		my $msg = shift();
86		syslog($assert_level, "VERIFY: $msg", @_);
87		$asserted++;
88	}
89	return ($bad);
90}
91
92
93
94
95sub getstat($$);
96sub generate_delta($$);
97sub compress_deltas($);
98sub dumpdelta($);
99
100sub goodness($);
101sub imbalanced($$);
102sub do_reconfig($);
103
104sub goodness_cpu($$);		# private function
105sub move_intr($$$$);		# private function
106sub ivecs_to_string(@);		# private function
107sub do_find_goal($$$$);		# private function
108sub find_goal($$);		# private function
109sub do_reconfig_cpu2cpu($$$$);	# private function
110sub do_reconfig_cpu($$$);	# private function
111
112
113#
114# What follow are the basic data structures routines of intrd.
115#
116# getstat() is responsible for reading the kstats and generating a "stat" hash.
117#
118# generate_delta() is responsible for taking two "stat" hashes and creating
119# a new "delta" hash that represents what has changed over time.
120#
121# compress_deltas() is responsible for taking a list of deltas and generating
122# a single delta hash that encompasses all the time periods described by the
123# deltas.
124
125
126#
127# getstat() is handed a reference to a kstat and generates a hash, returned
128# by reference, containing all the fields from the kstats which we need.
129# If it returns the scalar 0, it failed to gather the kstats, and the caller
130# should react accordingly.
131#
132# getstat() is also responsible for maintaining a reasonable $sleeptime.
133#
134# {"snaptime"}          kstat's snaptime
135# {<cpuid>}             one hash reference per online cpu
136#  ->{"tot"}            == cpu:<cpuid>:sys:cpu_nsec_{user + kernel + idle}
137#  ->{"crtime"}         == cpu:<cpuid>:sys:crtime
138#  ->{"ivecs"}
139#     ->{<cookie#>}     iterates over pci_intrs::<nexus>:cookie
140#        ->{"time"}     == pci_intrs:<ivec#>:<nexus>:time (in nsec)
141#        ->{"pil"}      == pci_intrs:<ivec#>:<nexus>:pil
142#        ->{"crtime"}   == pci_intrs:<ivec#>:<nexus>:crtime
143#        ->{"ino"}      == pci_intrs:<ivec#>:<nexus>:ino
144#        ->{"num_ino"}  == num inos of single device instance sharing this entry
145#				Will be > 1 on pcplusmp X86 systems for devices
146#				with multiple MSI interrupts.
147#        ->{"buspath"}  == pci_intrs:<ivec#>:<nexus>:buspath
148#        ->{"name"}     == pci_intrs:<ivec#>:<nexus>:name
149#        ->{"ihs"}      == pci_intrs:<ivec#>:<nexus>:ihs
150#
151
152sub getstat($$)
153{
154	my ($ks, $pcplusmp_sys) = @_;
155
156	my $cpucnt = 0;
157	my %stat = ();
158	my ($minsnap, $maxsnap);
159
160	# Hash of hash which matches (MSI device, ino) combos to kstats.
161	my %msidevs = ();
162
163	# kstats are not generated atomically. Each kstat hierarchy will
164	# have been generated within the kernel at a different time. On a
165	# thrashing system, we may not run quickly enough in order to get
166	# coherent kstat timing information across all the kstats. To
167	# determine if this is occurring, $minsnap/$maxsnap are used to
168	# find the breadth between the first and last snaptime of all the
169	# kstats we access. $maxsnap - $minsnap roughly represents the
170	# total time taken up in getstat(). If this time approaches the
171	# time between snapshots, our results may not be useful.
172
173	$minsnap = -1;		# snaptime is always a positive number
174	$maxsnap = $minsnap;
175
176	# Iterate over the cpus in cpu:<cpuid>::. Check
177	# cpu_info:<cpuid>:cpu_info<cpuid>:state to make sure the
178	# processor is "on-line". If not, it isn't accepting interrupts
179	# and doesn't concern us.
180	#
181	# Record cpu:<cpuid>:sys:snaptime, and check $minsnap/$maxsnap.
182
183	while (my ($cpu, $cpst) = each %{$ks->{cpu}}) {
184		next if !exists($ks->{cpu_info}{$cpu}{"cpu_info$cpu"}{state});
185		#"state" fld of kstat w/
186		#		  modname    inst name-"cpuinfo0"
187		my $state = $ks->{cpu_info}{$cpu}{"cpu_info$cpu"}{state};
188		next if ($state !~ /^on-line\0/);
189		my $cpu_sys = $cpst->{sys};
190
191		$stat{$cpu}{tot} = ($cpu_sys->{cpu_nsec_idle} +
192				    $cpu_sys->{cpu_nsec_user} +
193				    $cpu_sys->{cpu_nsec_kernel});
194		$stat{$cpu}{crtime} = $cpu_sys->{crtime};
195		$stat{$cpu}{ivecs} = {};
196
197		if ($minsnap == -1 || $cpu_sys->{snaptime} < $minsnap) {
198			$minsnap = $cpu_sys->{snaptime};
199		}
200		if ($cpu_sys->{snaptime} > $maxsnap) {
201			$maxsnap = $cpu_sys->{snaptime};
202		}
203		$cpucnt++;
204	}
205
206	if ($cpucnt <= 1) {
207		$sleeptime = $onecpu_sleeptime;
208		return (0);	# nothing to do with 1 CPU
209	}
210
211	# Iterate over the ivecs. If the cpu is not on-line, ignore the
212	# ivecs mapped to it, if any.
213	#
214	# Record pci_intrs:{inum}:<nexus>:time, snaptime, crtime, pil,
215	# ino, name, and buspath. Check $minsnap/$maxsnap.
216
217	foreach my $inst (values(%{$ks->{pci_intrs}})) {
218		my $intrcfg = (values(%$inst))[0];
219		my $cpu = $intrcfg->{cpu};
220
221		next unless exists $stat{$cpu};
222		next if ($intrcfg->{type} =~ /^disabled\0/);
223
224		# Perl looks beyond NULL chars in pattern matching.
225		# Truncate name field at the first NULL
226		$intrcfg->{name} =~ s/\0.*$//;
227
228		if ($intrcfg->{snaptime} < $minsnap) {
229			$minsnap = $intrcfg->{snaptime};
230		} elsif ($intrcfg->{snaptime} > $maxsnap) {
231			$maxsnap = $intrcfg->{snaptime};
232		}
233
234		my $cookie = "$intrcfg->{buspath} $intrcfg->{ino}";
235		if (exists $stat{$cpu}{ivecs}{$cookie}) {
236			my $cookiestats = $stat{$cpu}{ivecs}{$cookie};
237
238			$cookiestats->{time} += $intrcfg->{time};
239			$cookiestats->{name} .= "/$intrcfg->{name}";
240
241			# If this new interrupt sharing $cookie represents a
242			# change from an earlier getstat, make sure that
243			# generate_delta will see the change by setting
244			# crtime to the most recent crtime of its components.
245
246			if ($intrcfg->{crtime} > $cookiestats->{crtime}) {
247				$cookiestats->{crtime} = $intrcfg->{crtime};
248			}
249			$cookiestats->{ihs}++;
250			next;
251		}
252		$stat{$cpu}{ivecs}{$cookie}{time} = $intrcfg->{time};
253		$stat{$cpu}{ivecs}{$cookie}{crtime} = $intrcfg->{crtime};
254		$stat{$cpu}{ivecs}{$cookie}{pil} = $intrcfg->{pil};
255		$stat{$cpu}{ivecs}{$cookie}{ino} = $intrcfg->{ino};
256		$stat{$cpu}{ivecs}{$cookie}{num_ino} = 1;
257		$stat{$cpu}{ivecs}{$cookie}{buspath} = $intrcfg->{buspath};
258		$stat{$cpu}{ivecs}{$cookie}{name} = $intrcfg->{name};
259		$stat{$cpu}{ivecs}{$cookie}{ihs} = 1;
260
261		if ($pcplusmp_sys && ($intrcfg->{type} =~ /^msi\0/)) {
262			if (!(exists($msidevs{$intrcfg->{name}}))) {
263				$msidevs{$intrcfg->{name}} = {};
264			}
265			$msidevs{$intrcfg->{name}}{$intrcfg->{ino}} =
266			    \$stat{$cpu}{ivecs}{$cookie};
267		}
268	}
269
270	# All MSI interrupts of a device instance share a single MSI address.
271	# On X86 systems with an APIC, this MSI address is interpreted as CPU
272	# routing info by the APIC.  For this reason, on these platforms, all
273	# interrupts for MSI devices must be moved to the same CPU at the same
274	# time.
275	#
276	# Since all interrupts will be on the same CPU on these platforms, all
277	# interrupts can be consolidated into one ivec entry.  For such devices,
278	# num_ino will be > 1 to denote that a group move is needed.
279
280	# Loop thru all MSI devices on X86 pcplusmp systems.
281	# Nop on other systems.
282	foreach my $msidevkey (sort keys %msidevs) {
283
284		# Loop thru inos of the device, sorted by lowest value first
285		# For each cookie found for a device, incr num_ino for the
286		# lowest cookie and remove other cookies.
287
288		# Assumes PIL is the same for first and current cookies
289
290		my $first_ino = -1;
291		my $first_cookiep;
292		my $curr_cookiep;
293		foreach my $inokey (sort keys %{$msidevs{$msidevkey}}) {
294			$curr_cookiep = $msidevs{$msidevkey}{$inokey};
295			if ($first_ino == -1) {
296				$first_ino = $inokey;
297				$first_cookiep = $curr_cookiep;
298			} else {
299				$$first_cookiep->{num_ino}++;
300				$$first_cookiep->{time} +=
301				    $$curr_cookiep->{time};
302				if ($$curr_cookiep->{crtime} >
303				    $$first_cookiep->{crtime}) {
304					$$first_cookiep->{crtime} =
305					    $$curr_cookiep->{crtime};
306				}
307				# Invalidate this cookie, less complicated and
308				# more efficient than deleting it.
309				$$curr_cookiep->{num_ino} = 0;
310			}
311		}
312	}
313
314	# We define the timerange as the amount of time spent gathering the
315	# various kstats, divided by our sleeptime. If we take a lot of time
316	# to access the kstats, and then we create a delta comparing these
317	# kstats with a prior set of kstats, that delta will cover
318	# substaintially different amount of time depending upon which
319	# interrupt or CPU is being examined.
320	#
321	# By checking the timerange here, we guarantee that any deltas
322	# created from these kstats will contain self-consistent data,
323	# in that all CPUs and interrupts cover a similar span of time.
324	#
325	# $timerange_toohi is the upper bound. Any timerange above
326	# this is thrown out as garbage. If the stat is safely within this
327	# bound, we treat the stat as representing an instant in time, rather
328	# than the time range it actually spans. We arbitrarily choose minsnap
329	# as the snaptime of the stat.
330
331	$stat{snaptime} = $minsnap;
332	my $timerange = ($maxsnap - $minsnap) / $sleeptime;
333	return (0) if ($timerange > $timerange_toohi);	# i.e. failure
334	return (\%stat);
335}
336
337#
338# dumpdelta takes a reference to our "delta" structure:
339# {"missing"}           "1" if the delta's component stats had inconsistencies
340# {"minsnap"}           time of the first kstat snaptime used in this delta
341# {"maxsnap"}           time of the last kstat snaptime used in this delta
342# {"goodness"}          cost function applied to this delta
343# {"avgintrload"}       avg of interrupt load across cpus, as a percentage
344# {"avgintrnsec"}       avg number of nsec spent in interrupts, per cpu
345# {<cpuid>}             iterates over on-line cpus
346#  ->{"intrs"}          cpu's movable intr time (sum of "time" for each ivec)
347#  ->{"tot"}            CPU load from all sources in nsec
348#  ->{"bigintr"}        largest value of {ivecs}{<ivec#>}{time} from below
349#  ->{"intrload"}       intrs / tot
350#  ->{"ivecs"}
351#     ->{<ivec#>}       iterates over ivecs for this cpu
352#        ->{"time"}     time used by this interrupt (in nsec)
353#        ->{"pil"}      pil level of this interrupt
354#        ->{"ino"}      interrupt number (or base vector if MSI group)
355#        ->{"buspath"}  filename of the directory of the device's bus
356#        ->{"name"}     device name
357#        ->{"ihs"}      number of different handlers sharing this ino
358#        ->{"num_ino"}  number of interrupt vectors in MSI group
359#
360# It prints out the delta structure in a nice, human readable display.
361#
362
363sub dumpdelta($)
364{
365	my ($delta) = @_;
366
367	# print global info
368
369	syslog('debug', "dumpdelta:");
370	syslog('debug', " RECONFIGURATION IN DELTA") if $delta->{missing} > 0;
371	syslog('debug', " avgintrload: %5.2f%%  avgintrnsec: %d",
372	       $delta->{avgintrload} * 100, $delta->{avgintrnsec});
373	syslog('debug', "    goodness: %5.2f%%", $delta->{goodness} * 100)
374	    if exists($delta->{goodness});
375
376	# iterate over cpus
377
378	while (my ($cpu, $cpst) = each %$delta) {
379		next if !ref($cpst);		# skip non-cpuid entries
380		my $tot = $cpst->{tot};
381		syslog('debug', "    cpu %3d intr %7.3f%%  (bigintr %7.3f%%)",
382		       $cpu, $cpst->{intrload}*100, $cpst->{bigintr}*100/$tot);
383		syslog('debug', "        intrs %d, bigintr %d",
384		       $cpst->{intrs}, $cpst->{bigintr});
385
386		# iterate over ivecs on this cpu
387
388		while (my ($ivec, $ivst) = each %{$cpst->{ivecs}}) {
389			syslog('debug', "    %15s:\"%s\": %7.3f%%  %d",
390			    ($ivst->{ihs} > 1 ? "$ivst->{name}($ivst->{ihs})" :
391			    $ivst->{name}), $ivec,
392			    $ivst->{time}*100 / $tot, $ivst->{time});
393		}
394	}
395}
396
397#
398# generate_delta($stat, $newstat) takes two stat references, returned from
399# getstat(), and creates a %delta. %delta (not surprisingly) contains the
400# same basic info as stat and newstat, but with the timestamps as deltas
401# instead of absolute times. We return a reference to the delta.
402#
403
404sub generate_delta($$)
405{
406	my ($stat, $newstat) = @_;
407
408	my %delta = ();
409	my $intrload;
410	my $intrnsec;
411	my $cpus;
412
413	# Take the worstcase timerange
414	$delta{minsnap} = $stat->{snaptime};
415	$delta{maxsnap} = $newstat->{snaptime};
416	if (VERIFY($delta{maxsnap} > $delta{minsnap},
417	    "generate_delta: stats aren't ascending")) {
418		$delta{missing} = 1;
419		return (\%delta);
420	}
421
422	# if there are a different number of cpus in the stats, set missing
423
424	$delta{missing} = (keys(%$stat) != keys(%$newstat));
425	if (VERIFY($delta{missing} == 0,
426	    "generate_delta: number of CPUs changed")) {
427		return (\%delta);
428	}
429
430	# scan through every cpu in %newstat and compare against %stat
431
432	while (my ($cpu, $newcpst) = each %$newstat) {
433		next if !ref($newcpst);		# skip non-cpuid fields
434
435		# If %stat is missing a cpu from %newstat, then it was just
436		# onlined. Mark missing.
437
438		if (VERIFY(exists $stat->{$cpu} &&
439		    $stat->{$cpu}{crtime} == $newcpst->{crtime},
440		    "generate_delta: cpu $cpu changed")) {
441			$delta{missing} = 1;
442			return (\%delta);
443		}
444		my $cpst = $stat->{$cpu};
445		$delta{$cpu}{tot} = $newcpst->{tot} - $cpst->{tot};
446		if (VERIFY($delta{$cpu}{tot} >= 0,
447		    "generate_delta: deltas are not ascending?")) {
448			$delta{missing} = 1;
449			delete($delta{$cpu});
450			return (\%delta);
451		}
452		# Avoid remote chance of division by zero
453		$delta{$cpu}{tot} = 1 if $delta{$cpu}{tot} == 0;
454		$delta{$cpu}{intrs} = 0;
455		$delta{$cpu}{bigintr} = 0;
456
457		my %ivecs = ();
458		$delta{$cpu}{ivecs} = \%ivecs;
459
460		# if the number of ivecs differs, set missing
461
462		if (VERIFY(keys(%{$cpst->{ivecs}}) ==
463			   keys(%{$newcpst->{ivecs}}),
464			   "generate_delta: cpu $cpu has more/less".
465			   " interrupts")) {
466			$delta{missing} = 1;
467			return (\%delta);
468		}
469
470		while (my ($inum, $newivec) = each %{$newcpst->{ivecs}}) {
471
472			# Unused cookie, corresponding to an MSI vector which
473			# is part of a group.  The whole group is accounted for
474			# by a different cookie.
475			next if ($newivec->{num_ino} == 0);
476
477			# If this ivec doesn't exist in $stat, or if $stat
478			# shows a different crtime, set missing.
479			if (VERIFY(exists $cpst->{ivecs}{$inum} &&
480				   $cpst->{ivecs}{$inum}{crtime} ==
481				   $newivec->{crtime},
482				   "generate_delta: cpu $cpu inum $inum".
483				   " has changed")) {
484				$delta{missing} = 1;
485				return (\%delta);
486			}
487			my $ivec = $cpst->{ivecs}{$inum};
488
489			# Create $delta{$cpu}{ivecs}{$inum}.
490
491			my %dltivec = ();
492			$delta{$cpu}{ivecs}{$inum} = \%dltivec;
493
494			# calculate time used by this interrupt
495
496			my $time = $newivec->{time} - $ivec->{time};
497			if (VERIFY($time >= 0,
498				   "generate_delta: ivec went backwards?")) {
499				$delta{missing} = 1;
500				delete($delta{$cpu}{ivecs}{$inum});
501				return (\%delta);
502			}
503			$delta{$cpu}{intrs} += $time;
504			$dltivec{time} = $time;
505			if ($time > $delta{$cpu}{bigintr}) {
506				$delta{$cpu}{bigintr} = $time;
507			}
508
509			# Transfer over basic info about the kstat. We
510			# don't have to worry about discrepancies between
511			# ivec and newivec because we verified that both
512			# have the same crtime.
513
514			$dltivec{pil} = $newivec->{pil};
515			$dltivec{ino} = $newivec->{ino};
516			$dltivec{buspath} = $newivec->{buspath};
517			$dltivec{name} = $newivec->{name};
518			$dltivec{ihs} = $newivec->{ihs};
519			$dltivec{num_ino} = $newivec->{num_ino};
520		}
521		if ($delta{$cpu}{tot} < $delta{$cpu}{intrs}) {
522			# Ewww! Hopefully just a rounding error.
523			# Make something up.
524			$delta{$cpu}{tot} = $delta{$cpu}{intrs};
525		}
526		$delta{$cpu}{intrload} =
527		       $delta{$cpu}{intrs} / $delta{$cpu}{tot};
528		$intrload += $delta{$cpu}{intrload};
529		$intrnsec += $delta{$cpu}{intrs};
530		$cpus++;
531	}
532	if ($cpus > 0) {
533		$delta{avgintrload} = $intrload / $cpus;
534		$delta{avgintrnsec} = $intrnsec / $cpus;
535	} else {
536		$delta{avgintrload} = 0;
537		$delta{avgintrnsec} = 0;
538	}
539	return (\%delta);
540}
541
542
543# compress_delta takes a list of deltas, and returns a single new delta
544# which represents the combined information from all the deltas. The deltas
545# provided are assumed to be sequential in time. The resulting compressed
546# delta looks just like any other delta. This new delta is also more accurate
547# since its statistics are averaged over a longer period than any of the
548# original deltas.
549
550sub compress_deltas ($)
551{
552	my ($deltas) = @_;
553
554	my %newdelta = ();
555	my ($intrs, $tot);
556	my $cpus = 0;
557	my ($high_intrload) = 0;
558
559	if (VERIFY($#$deltas != -1,
560		   "compress_deltas: list of delta is empty?")) {
561		return (0);
562	}
563	$newdelta{minsnap} = $deltas->[0]{minsnap};
564	$newdelta{maxsnap} = $deltas->[$#$deltas]{maxsnap};
565	$newdelta{missing} = 0;
566
567	foreach my $delta (@$deltas) {
568		if (VERIFY($delta->{missing} == 0,
569		    "compressing bad deltas?")) {
570			return (0);
571		}
572		while (my ($cpuid, $cpu) = each %$delta) {
573			next if !ref($cpu);
574
575			$intrs += $cpu->{intrs};
576			$tot += $cpu->{tot};
577			$newdelta{$cpuid}{intrs} += $cpu->{intrs};
578			$newdelta{$cpuid}{tot} += $cpu->{tot};
579			if (!exists $newdelta{$cpuid}{ivecs}) {
580				my %ivecs = ();
581				$newdelta{$cpuid}{ivecs} = \%ivecs;
582			}
583			while (my ($inum, $ivec) = each %{$cpu->{ivecs}}) {
584				my $newivecs = $newdelta{$cpuid}{ivecs};
585				$newivecs->{$inum}{time} += $ivec->{time};
586				$newivecs->{$inum}{pil} = $ivec->{pil};
587				$newivecs->{$inum}{ino} = $ivec->{ino};
588				$newivecs->{$inum}{buspath} = $ivec->{buspath};
589				$newivecs->{$inum}{name} = $ivec->{name};
590				$newivecs->{$inum}{ihs} = $ivec->{ihs};
591				$newivecs->{$inum}{num_ino} = $ivec->{num_ino};
592			}
593		}
594	}
595	foreach my $cpu (values(%newdelta)) {
596		next if !ref($cpu); # ignore non-cpu fields
597		$cpus++;
598
599		my $bigintr = 0;
600		foreach my $ivec (values(%{$cpu->{ivecs}})) {
601			if ($ivec->{time} > $bigintr) {
602				$bigintr = $ivec->{time};
603			}
604		}
605		$cpu->{bigintr} = $bigintr;
606		$cpu->{intrload} = $cpu->{intrs} / $cpu->{tot};
607		if ($high_intrload < $cpu->{intrload}) {
608			$high_intrload = $cpu->{intrload};
609		}
610		$cpu->{tot} = 1 if $cpu->{tot} <= 0;
611	}
612	if ($cpus == 0) {
613		$newdelta{avgintrnsec} = 0;
614		$newdelta{avgintrload} = 0;
615	} else {
616		$newdelta{avgintrnsec} = $intrs / $cpus;
617		$newdelta{avgintrload} = $intrs / $tot;
618	}
619	$sleeptime = ($high_intrload < $idle_intrload) ? $idle_sleeptime :
620	    $normal_sleeptime;
621	return (\%newdelta);
622}
623
624
625
626
627
628# What follow are the core functions responsible for examining the deltas
629# generated above and deciding what to do about them.
630#
631# goodness() and its helper goodness_cpu() return a heuristic which describe
632# how good (or bad) the current interrupt balance is. The value returned will
633# be between 0 and 1, with 0 representing maximum goodness, and 1 representing
634# maximum badness.
635#
636# imbalanced() compares a current and historical value of goodness, and
637# determines if there has been enough change to warrant evaluating a
638# reconfiguration of the interrupts
639#
640# do_reconfig(), and its helpers, do_reconfig_cpu(), do_reconfig_cpu2cpu(),
641# find_goal(), do_find_goal(), and move_intr(), are responsible for examining
642# a delta and determining the best possible assignment of interrupts to CPUs.
643#
644# It is important that do_reconfig() be in alignment with goodness(). If
645# do_reconfig were to generate a new interrupt distribution that worsened
646# goodness, we could get into a pathological loop with intrd fighting itself,
647# constantly deciding that things are imbalanced, and then changing things
648# only to make them worse.
649
650
651
652# any goodness over $goodness_unsafe_load is considered really bad
653# goodness must drop by at least $goodness_mindelta for a reconfig
654
655my $goodness_unsafe_load = .9;
656my $goodness_mindelta = .1;
657
658# goodness(%delta) examines a delta and return its "goodness". goodness will
659# be between 0 (best) and 1 (major bad). goodness is determined by evaluating
660# the goodness of each individual cpu, and returning the worst case. This
661# helps on systems with many CPUs, where otherwise a single pathological CPU
662# might otherwise be ignored because the average was OK.
663#
664# To calculate the goodness of an individual CPU, we start by looking at its
665# load due to interrupts. If the load is above a certain high threshold and
666# there is more than one interrupt assigned to this CPU, we set goodness
667# to worst-case. If the load is below the average interrupt load of all CPUs,
668# then we return best-case, since what's to complain about?
669#
670# Otherwise we look at how much the load is above the average, and return
671# that as the goodness, with one caveat: we never return more than the CPU's
672# interrupt load ignoring its largest single interrupt source. This is
673# because a CPU with one high-load interrupt, and no other interrupts, is
674# perfectly balanced. Nothing can be done to improve the situation, and thus
675# it is perfectly balanced even if the interrupt's load is 100%.
676
677sub goodness($)
678{
679	my ($delta) = @_;
680
681	return (1) if $delta->{missing} > 0;
682
683	my $high_goodness = 0;
684	my $goodness;
685
686	foreach my $cpu (values(%$delta)) {
687		next if !ref($cpu);		# skip non-cpuid fields
688
689		$goodness = goodness_cpu($cpu, $delta->{avgintrload});
690		if (VERIFY($goodness >= 0 && $goodness <= 1,
691			   "goodness: cpu goodness out of range?")) {
692			dumpdelta($delta);
693			return (1);
694		}
695		if ($goodness == 1) {
696			return (1);	# worst case, no need to continue
697		}
698		if ($goodness > $high_goodness) {
699			$high_goodness = $goodness;
700		}
701	}
702	return ($high_goodness);
703}
704
705sub goodness_cpu($$)		# private function
706{
707	my ($cpu, $avgintrload) = @_;
708
709	my $goodness;
710	my $load = $cpu->{intrs} / $cpu->{tot};
711
712	return (0) if ($load < $avgintrload);	# low loads are perfectly good
713
714	# Calculate $load_no_bigintr, which represents the load
715	# due to interrupts, excluding the one biggest interrupt.
716	# This is the most gain we can get on this CPU from
717	# offloading interrupts.
718
719	my $load_no_bigintr = ($cpu->{intrs} - $cpu->{bigintr}) / $cpu->{tot};
720
721	# A major imbalance is indicated if a CPU is saturated
722	# with interrupt handling, and it has more than one
723	# source of interrupts. Those other interrupts could be
724	# starved if of a lower pil. Return a goodness of 1,
725	# which is the worst possible return value,
726	# which will effectively contaminate this entire delta.
727
728	my $cnt = keys(%{$cpu->{ivecs}});
729
730	if ($load > $goodness_unsafe_load && $cnt > 1) {
731		return (1);
732	}
733	$goodness = $load - $avgintrload;
734	if ($goodness > $load_no_bigintr) {
735		$goodness = $load_no_bigintr;
736	}
737	return ($goodness);
738}
739
740
741# imbalanced() is used by the main routine to determine if the goodness
742# has shifted far enough from our last baseline to warrant a reassignment
743# of interrupts. A very high goodness indicates that a CPU is way out of
744# whack. If the goodness has varied too much since the baseline, then
745# perhaps a reconfiguration is worth considering.
746
747sub imbalanced ($$)
748{
749	my ($goodness, $baseline) = @_;
750
751	# Return 1 if we are pathological, or creeping away from the baseline
752
753	return (1) if $goodness > .50;
754	return (1) if abs($goodness - $baseline) > $goodness_mindelta;
755	return (0);
756}
757
758# do_reconfig(), do_reconfig_cpu(), and do_reconfig_cpu2cpu(), are the
759# decision-making functions responsible for generating a new interrupt
760# distribution. They are designed with the definition of goodness() in
761# mind, i.e. they use the same definition of "good distribution" as does
762# goodness().
763#
764# do_reconfig() is responsible for deciding whether a redistribution is
765# actually warranted. If the goodness is already pretty good, it doesn't
766# waste the CPU time to generate a new distribution. If it
767# calculates a new distribution and finds that it is not sufficiently
768# improved from the prior distirbution, it will not do the redistribution,
769# mainly to avoid the disruption to system performance caused by
770# rejuggling interrupts.
771#
772# Its main loop works by going through a list of cpus sorted from
773# highest to lowest interrupt load. It removes the highest-load cpus
774# one at a time and hands them off to do_reconfig_cpu(). This function
775# then re-sorts the remaining CPUs from lowest to highest interrupt load,
776# and one at a time attempts to rejuggle interrupts between the original
777# high-load CPU and the low-load CPU. Rejuggling on a high-load CPU is
778# considered finished as soon as its interrupt load is within
779# $goodness_mindelta of the average interrupt load. Such a CPU will have
780# a goodness of below the $goodness_mindelta threshold.
781
782#
783# move_intr(\%delta, $inum, $oldcpu, $newcpu)
784# used by reconfiguration code to move an interrupt between cpus within
785# a delta. This manipulates data structures, and does not actually move
786# the interrupt on the running system.
787#
788sub move_intr($$$$)		# private function
789{
790	my ($delta, $inum, $oldcpuid, $newcpuid) = @_;
791
792	my $ivec = $delta->{$oldcpuid}{ivecs}{$inum};
793
794	# Remove ivec from old cpu
795
796	my $oldcpu = $delta->{$oldcpuid};
797	$oldcpu->{intrs} -= $ivec->{time};
798	$oldcpu->{intrload} = $oldcpu->{intrs} / $oldcpu->{tot};
799	delete($oldcpu->{ivecs}{$inum});
800
801	VERIFY($oldcpu->{intrs} >= 0, "move_intr: intr's time > total time?");
802	VERIFY($ivec->{time} <= $oldcpu->{bigintr},
803	       "move_intr: intr's time > bigintr?");
804
805	if ($ivec->{time} >= $oldcpu->{bigintr}) {
806		my $bigtime = 0;
807
808		foreach my $ivec (values(%{$oldcpu->{ivecs}})) {
809			$bigtime = $ivec->{time} if $ivec->{time} > $bigtime;
810		}
811		$oldcpu->{bigintr} = $bigtime;
812	}
813
814	# Add ivec onto new cpu
815
816	my $newcpu = $delta->{$newcpuid};
817
818	$ivec->{nowcpu} = $newcpuid;
819	$newcpu->{intrs} += $ivec->{time};
820	$newcpu->{intrload} = $newcpu->{intrs} / $newcpu->{tot};
821	$newcpu->{ivecs}{$inum} = $ivec;
822
823	$newcpu->{bigintr} = $ivec->{time}
824		if $ivec->{time} > $newcpu->{bigintr};
825}
826
827sub move_intr_check($$$)	# private function
828{
829	my ($delta, $oldcpuid, $newcpuid) = @_;
830
831	VERIFY($delta->{$oldcpuid}{tot} >= $delta->{$oldcpuid}{intrs},
832	       "Moved interrupts left 100+%% load on src cpu");
833	VERIFY($delta->{$newcpuid}{tot} >= $delta->{$newcpuid}{intrs},
834	       "Moved interrupts left 100+%% load on tgt cpu");
835}
836
837sub ivecs_to_string(@)		# private function
838{
839	my $str = "";
840	foreach my $ivec (@_) {
841		$str = "$str $ivec->{inum}";
842	}
843	return ($str);
844}
845
846
847sub do_reconfig($)
848{
849	my ($delta) = @_;
850
851	my $goodness = $delta->{goodness};
852
853	# We can't improve goodness to better than 0. We should stop here
854	# if, even if we achieve a goodness of 0, the improvement is still
855	# too small to merit the action.
856
857	if ($goodness - 0 < $goodness_mindelta) {
858		syslog('debug', "goodness good enough, don't reconfig");
859		return (0);
860	}
861
862	syslog('notice', "Optimizing interrupt assignments");
863
864	if (VERIFY ($delta->{missing} == 0, "RECONFIG Aborted: should not ".
865	    "have a delta with missing")) {
866		return (-1);
867	}
868
869	# Make a list of all cpuids, and also add some extra information
870	# to the ivec structures.
871
872	my @cpusortlist = ();
873
874	while (my ($cpuid, $cpu) = each %$delta) {
875		next if !ref($cpu);	# skip non-cpu entries
876
877		push(@cpusortlist, $cpuid);
878		while (my ($inum, $ivec) = each %{$cpu->{ivecs}}) {
879			$ivec->{origcpu} = $cpuid;
880			$ivec->{nowcpu} = $cpuid;
881			$ivec->{inum} = $inum;
882		}
883	}
884
885	# Sort the list of CPUs from highest to lowest interrupt load.
886	# Remove the top CPU from that list and attempt to redistribute
887	# its interrupts. If the CPU has a goodness below a threshold,
888	# just ignore the CPU and move to the next one. If the CPU's
889	# load falls below the average load plus that same threshold,
890	# then there are no CPUs left worth reconfiguring, and we're done.
891
892	while (@cpusortlist) {
893		# Re-sort cpusortlist each time, since do_reconfig_cpu can
894		# move interrupts around.
895
896		@cpusortlist =
897		    sort({$delta->{$b}{intrload} <=> $delta->{$a}{intrload}}
898		    @cpusortlist);
899
900		my $cpu = shift(@cpusortlist);
901		if (($delta->{$cpu}{intrload} <= $goodness_unsafe_load) &&
902		    ($delta->{$cpu}{intrload} <=
903		    $delta->{avgintrload} + $goodness_mindelta)) {
904			syslog('debug', "finished reconfig: cpu $cpu load ".
905			    "$delta->{$cpu}{intrload} avgload ".
906			    "$delta->{avgintrload}");
907			last;
908		}
909		if (goodness_cpu($delta->{$cpu}, $delta->{avgintrload}) <
910		    $goodness_mindelta) {
911			next;
912		}
913		do_reconfig_cpu($delta, \@cpusortlist, $cpu);
914	}
915
916	# How good a job did we do? If the improvement was minimal, and
917	# our goodness wasn't pathological (and thus needing any help it
918	# can get), then don't bother moving the interrupts.
919
920	my $newgoodness = goodness($delta);
921	VERIFY($newgoodness <= $goodness,
922	       "reconfig: result has worse goodness?");
923
924	if (($goodness != 1 || $newgoodness == 1) &&
925	    $goodness - $newgoodness < $goodness_mindelta) {
926		syslog('debug', "goodness already near optimum, ".
927		       "don't reconfig");
928		return (0);
929	}
930	syslog('debug', "goodness %5.2f%% --> %5.2f%%", $goodness*100,
931	       $newgoodness*100);
932
933	# Time to move those interrupts!
934
935	my $ret = 1;
936	my $warned = 0;
937	while (my ($cpuid, $cpu) = each %$delta) {
938		next if $cpuid =~ /\D/;
939		while (my ($inum, $ivec) = each %{$cpu->{ivecs}}) {
940			next if ($ivec->{origcpu} == $cpuid);
941
942			if (!intrmove($ivec->{buspath}, $ivec->{ino},
943			    $cpuid, $ivec->{num_ino})) {
944				syslog('warning', "Unable to move interrupts")
945				    if $warned++ == 0;
946				syslog('debug', "Unable to move buspath ".
947				    "$ivec->{buspath} ino $ivec->{ino} to ".
948				    "cpu $cpuid");
949				$ret = -1;
950			}
951		}
952	}
953
954	syslog('notice', "Interrupt assignments optimized");
955	return ($ret);
956}
957
958sub do_reconfig_cpu($$$)	# private function
959{
960	my ($delta, $cpusortlist, $oldcpuid) = @_;
961
962	# We have been asked to rejuggle interrupts between $oldcpuid and
963	# other CPUs found on $cpusortlist so as to improve the load on
964	# $oldcpuid. We reverse $cpusortlist to get our own copy of the
965	# list, sorted from lowest to highest interrupt load. One at a
966	# time, shift a CPU off of this list of CPUs, and attempt to
967	# rejuggle interrupts between the two CPUs. Don't do this if the
968	# other CPU has a higher load than oldcpuid. We're done rejuggling
969	# once $oldcpuid's goodness falls below a threshold.
970
971	syslog('debug', "reconfiguring $oldcpuid");
972
973	my $cpu = $delta->{$oldcpuid};
974	my $avgintrload = $delta->{avgintrload};
975
976	my @cputargetlist = reverse(@$cpusortlist); # make a copy of the list
977	while ($#cputargetlist != -1) {
978 		last if goodness_cpu($cpu, $avgintrload) < $goodness_mindelta;
979
980		my $tgtcpuid = shift(@cputargetlist);
981		my $tgt = $delta->{$tgtcpuid};
982		my $load = $cpu->{intrload};
983		my $tgtload = $tgt->{intrload};
984		last if $tgtload > $load;
985		do_reconfig_cpu2cpu($delta, $oldcpuid, $tgtcpuid, $load);
986	}
987}
988
989sub do_reconfig_cpu2cpu($$$$)	# private function
990{
991	my ($delta, $srccpuid, $tgtcpuid, $srcload) = @_;
992
993	# We've been asked to consider interrupt juggling between srccpuid
994	# (with a high interrupt load) and tgtcpuid (with a lower interrupt
995	# load). First, make a single list with all of the ivecs from both
996	# CPUs, and sort the list from highest to lowest load.
997
998	syslog('debug', "exchanging intrs between $srccpuid and $tgtcpuid");
999
1000	# Gather together all the ivecs and sort by load
1001
1002	my @ivecs = (values(%{$delta->{$srccpuid}{ivecs}}),
1003	    values(%{$delta->{$tgtcpuid}{ivecs}}));
1004	return if $#ivecs == -1;
1005
1006	@ivecs = sort({$b->{time} <=> $a->{time}} @ivecs);
1007
1008	# Our "goal" load for srccpuid is the average load across all CPUs.
1009	# find_goal() will find determine the optimum selection of the
1010	# available interrupts which comes closest to this goal without
1011	# falling below the goal.
1012
1013	my $goal = $delta->{avgintrnsec};
1014
1015	# We know that the interrupt load on tgtcpuid is less than that on
1016	# srccpuid, but its load could still be above avgintrnsec. Don't
1017	# choose a goal which would bring srccpuid below the load on tgtcpuid.
1018
1019	my $avgnsec =
1020	    ($delta->{$srccpuid}{intrs} + $delta->{$tgtcpuid}{intrs}) / 2;
1021	if ($goal < $avgnsec) {
1022		$goal = $avgnsec;
1023	}
1024
1025	# If the largest of the interrupts is on srccpuid, leave it there.
1026	# This can help minimize the disruption caused by moving interrupts.
1027
1028	if ($ivecs[0]->{origcpu} == $srccpuid) {
1029		syslog('debug', "Keeping $ivecs[0]->{inum} on $srccpuid");
1030		$goal -= $ivecs[0]->{time};
1031		shift(@ivecs);
1032	}
1033
1034	syslog('debug', "GOAL: inums should total $goal");
1035	find_goal(\@ivecs, $goal);
1036
1037	# find_goal() returned its results to us by setting $ivec->{goal} if
1038	# the ivec should be on srccpuid, or clearing it for tgtcpuid.
1039	# Call move_intr() to update our $delta with the new results.
1040
1041	foreach my $ivec (@ivecs) {
1042		syslog('debug', "ivec $ivec->{inum} goal $ivec->{goal}");
1043		VERIFY($ivec->{nowcpu} == $srccpuid ||
1044		    $ivec->{nowcpu} == $tgtcpuid, "cpu2cpu found an ".
1045		    "interrupt not currently on src or tgt cpu");
1046
1047		if ($ivec->{goal} && $ivec->{nowcpu} != $srccpuid) {
1048			move_intr($delta, $ivec->{inum}, $ivec->{nowcpu},
1049			    $srccpuid);
1050		} elsif ($ivec->{goal} == 0 && $ivec->{nowcpu} != $tgtcpuid) {
1051			move_intr($delta, $ivec->{inum}, $ivec->{nowcpu},
1052			    $tgtcpuid);
1053		}
1054	}
1055	move_intr_check($delta, $srccpuid, $tgtcpuid); # asserts
1056
1057	my $newload = $delta->{$srccpuid}{intrs} / $delta->{$srccpuid}{tot};
1058	VERIFY($newload <= $srcload && $newload > $delta->{avgintrload},
1059	    "cpu2cpu: new load didn't end up in expected range");
1060}
1061
1062
1063# find_goal() and its helper do_find_goal() are used to find the best
1064# combination of interrupts in order to generate a load that is as close
1065# as possible to a goal load without falling below that goal. Before returning
1066# to its caller, find_goal() sets a new value in the hash of each interrupt,
1067# {goal}, which if set signifies that this interrupt is one of the interrupts
1068# identified as part of the set of interrupts which best meet the goal.
1069#
1070# The arguments to find_goal are a list of ivecs (hash references), sorted
1071# by descending {time}, and the goal load. The goal is relative to {time}.
1072# The best fit is determined by performing a depth-first search. do_find_goal
1073# is the recursive subroutine which carries out the search.
1074#
1075# It is passed an index as an argument, originally 0. On a given invocation,
1076# it is only to consider interrupts in the ivecs array starting at that index.
1077# It then considers two possibilities:
1078#   1) What is the best goal-fit if I include ivecs[index]?
1079#   2) What is the best goal-fit if I exclude ivecs[index]?
1080# To determine case 1, it subtracts the load of ivecs[index] from the goal,
1081# and calls itself recursively with that new goal and index++.
1082# To determine case 2, it calls itself recursively with the same goal and
1083# index++.
1084#
1085# It then compares the two results, decide which one best meets the goals,
1086# and returns the result. The return value is the best-fit's interrupt load,
1087# followed by a list of all the interrupts which make up that best-fit.
1088#
1089# As an optimization, a second array loads[] is created which mirrors ivecs[].
1090# loads[i] will equal the total loads of all ivecs[i..$#ivecs]. This is used
1091# by do_find_goal to avoid recursing all the way to the end of the ivecs
1092# array if including all remaining interrupts will still leave the best-fit
1093# at below goal load. If so, it then includes all remaining interrupts on
1094# the goal list and returns.
1095#
1096sub find_goal($$)		# private function
1097{
1098	my ($ivecs, $goal) = @_;
1099
1100	my @goals;
1101	my $load;
1102	my $ivec;
1103
1104	if ($goal <= 0) {
1105		@goals = ();	# the empty set will best meet the goal
1106	} else {
1107		syslog('debug', "finding goal from intrs %s",
1108		    ivecs_to_string(@$ivecs));
1109
1110		# Generate @loads array
1111
1112		my $tot = 0;
1113		foreach $ivec (@$ivecs) {
1114			$tot += $ivec->{time};
1115		}
1116		my @loads = ();
1117		foreach $ivec (@$ivecs) {
1118			push(@loads, $tot);
1119			$tot -= $ivec->{time};
1120		}
1121		($load, @goals) = do_find_goal($ivecs, \@loads, $goal, 0);
1122		VERIFY($load >= $goal, "find_goal didn't meet goals");
1123	}
1124	syslog('debug', "goals found: %s", ivecs_to_string(@goals));
1125
1126	# Set or clear $ivec->{goal} for each ivec, based on returned @goals
1127
1128	foreach $ivec (@$ivecs) {
1129		if ($#goals > -1 && $ivec == $goals[0]) {
1130			syslog('debug', "inum $ivec->{inum} on source cpu");
1131			$ivec->{goal} = 1;
1132			shift(@goals);
1133		} else {
1134			syslog('debug', "inum $ivec->{inum} on target cpu");
1135			$ivec->{goal} = 0;
1136		}
1137	}
1138}
1139
1140
1141sub do_find_goal($$$$)		# private function
1142{
1143	my ($ivecs, $loads, $goal, $idx) = @_;
1144
1145	if ($idx > $#{$ivecs}) {
1146		return (0);
1147	}
1148	syslog('debug', "$idx: finding goal $goal inum $ivecs->[$idx]{inum}");
1149
1150	my $load = $ivecs->[$idx]{time};
1151	my @goals_with = ();
1152	my @goals_without = ();
1153	my ($with, $without);
1154
1155	# If we include all remaining items and we're still below goal,
1156	# stop here. We can just return a result that includes $idx and all
1157	# subsequent ivecs. Since this will still be below goal, there's
1158	# nothing better to be done.
1159
1160	if ($loads->[$idx] <= $goal) {
1161		syslog('debug',
1162		    "$idx: including all remaining intrs %s with load %d",
1163		    ivecs_to_string(@$ivecs[$idx .. $#{$ivecs}]),
1164		    $loads->[$idx]);
1165		return ($loads->[$idx], @$ivecs[$idx .. $#{$ivecs}]);
1166	}
1167
1168	# Evaluate the "with" option, i.e. the best matching goal which
1169	# includes $ivecs->[$idx]. If idx's load is more than our goal load,
1170	# stop here. Once we're above the goal, there is no need to consider
1171	# further interrupts since they'll only take us further from the goal.
1172
1173	if ($goal <= $load) {
1174		$with = $load;	# stop here
1175	} else {
1176		($with, @goals_with) =
1177		    do_find_goal($ivecs, $loads, $goal - $load, $idx + 1);
1178		$with += $load;
1179	}
1180	syslog('debug', "$idx: with-load $with intrs %s",
1181	       ivecs_to_string($ivecs->[$idx], @goals_with));
1182
1183	# Evaluate the "without" option, i.e. the best matching goal which
1184	# excludes $ivecs->[$idx].
1185
1186	($without, @goals_without) =
1187	    &do_find_goal($ivecs, $loads, $goal, $idx + 1);
1188	syslog('debug', "$idx: without-load $without intrs %s",
1189	       ivecs_to_string(@goals_without));
1190
1191	# We now have our "with" and "without" options, and we choose which
1192	# best fits the goal. If one is greater than goal and the other is
1193	# below goal, we choose the one that is greater. If they are both
1194	# below goal, then we choose the one that is greater. If they are
1195	# both above goal, then we choose the smaller.
1196
1197	my $which;		# 0 == with, 1 == without
1198	if ($with >= $goal && $without < $goal) {
1199		$which = 0;
1200	} elsif ($with < $goal && $without >= $goal) {
1201		$which = 1;
1202	} elsif ($with >= $goal && $without >= $goal) {
1203		$which = ($without < $with);
1204	} else {
1205		$which = ($without > $with);
1206	}
1207
1208	# Return the load of our best case scenario, followed by all the ivecs
1209	# which compose that goal.
1210
1211	if ($which == 1) {	# without
1212		syslog('debug', "$idx: going without");
1213		return ($without, @goals_without);
1214	} else {
1215		syslog('debug', "$idx: going with");
1216		return ($with, $ivecs->[$idx], @goals_with);
1217	}
1218	# Not reached
1219}
1220
1221
1222
1223
1224syslog('debug', "intrd is starting".($debug ? " (debug)" : ""));
1225
1226my @deltas = ();
1227my $deltas_tottime = 0;		# sum of maxsnap-minsnap across @deltas
1228my $avggoodness;
1229my $baseline_goodness = 0;
1230my $compdelta;
1231
1232my $do_reconfig;
1233
1234# temp variables
1235my $goodness;
1236my $deltatime;
1237my $olddelta;
1238my $olddeltatime;
1239my $delta;
1240my $newstat;
1241my $below_statslen;
1242my $newtime;
1243my $ret;
1244
1245
1246my $gotsig = 0;
1247$SIG{INT} = sub { $gotsig = 1; };     # don't die in the middle of retargeting
1248$SIG{HUP} = $SIG{INT};
1249$SIG{TERM} = $SIG{INT};
1250
1251my $ks;
1252if ($using_scengen == 0) {
1253	$ks = Sun::Solaris::Kstat->new();
1254} else {
1255	$ks = myks_update();	# supplied by the simulator
1256}
1257
1258# If no pci_intrs kstats were found, we need to exit, but we can't because
1259# SMF will restart us and/or report an error to the administrator. But
1260# there's nothing an administrator can do. So print out a message for SMF
1261# logs and silently pause forever.
1262
1263if (!exists($ks->{pci_intrs})) {
1264	print STDERR "$cmdname: no interrupts were found; ".
1265	    "your PCI bus may not yet be supported\n";
1266	pause() while $gotsig == 0;
1267	exit 0;
1268}
1269
1270# See if this is a system with a pcplusmp APIC.
1271# Such systems will get special handling.
1272# Assume that if one bus has a pcplusmp APIC that they all do.
1273
1274# Get a list of pci_intrs kstats.
1275my @elem = values(%{$ks->{pci_intrs}});
1276my $elem0 = $elem[0];
1277my $elemval = (values(%$elem0))[0];
1278
1279# Use its buspath to query the system.  It is assumed that either all or none
1280# of the busses on a system are hosted by the pcplusmp APIC.
1281my $pcplusmp_sys = is_pcplusmp($elemval->{buspath});
1282
1283my $stat = getstat($ks, $pcplusmp_sys);
1284
1285for (;;) {
1286	sub clear_deltas {
1287		@deltas = ();
1288		$deltas_tottime = 0;
1289		$stat = 0;   # prevent next gen_delta() from setting {missing}
1290	}
1291
1292	# 1. Sleep, update the kstats, and save the new stats in $newstat.
1293
1294	exit 0 if $gotsig;		# if we got ^C / SIGTERM, exit
1295	if ($using_scengen == 0) {
1296		sleep($sleeptime);
1297		exit 0 if $gotsig;	# if we got ^C / SIGTERM, exit
1298		$ks->update();
1299	} else {
1300		$ks = myks_update();
1301	}
1302	$newstat = getstat($ks, $pcplusmp_sys);
1303
1304	# $stat or $newstat could be zero if they're uninitialized, or if
1305	# getstat() failed. If $stat is zero, move $newstat to $stat, sleep
1306	# and try again. If $newstat is zero, then we also sleep and try
1307	# again, hoping the problem will clear up.
1308
1309	next if (!ref $newstat);
1310	if (!ref $stat) {
1311		$stat = $newstat;
1312		next;
1313	}
1314
1315	# 2. Compare $newstat with the prior set of values, result in %$delta.
1316
1317	$delta = generate_delta($stat, $newstat);
1318	dumpdelta($delta) if $debug;	# Dump most recent stats to stdout.
1319	$stat = $newstat;	# The new stats now become the old stats.
1320
1321
1322	# 3. If $delta->{missing}, then there has been a reconfiguration of
1323	# either cpus or interrupts (probably both). We need to toss out our
1324	# old set of statistics and start from scratch.
1325	#
1326	# Also, if the delta covers a very long range of time, then we've
1327	# been experiencing a system overload that has resulted in intrd
1328	# not being allowed to run effectively for a while now. As above,
1329	# toss our old statistics and start from scratch.
1330
1331	$deltatime = $delta->{maxsnap} - $delta->{minsnap};
1332	if ($delta->{missing} > 0 || $deltatime > $statslen) {
1333		clear_deltas();
1334		syslog('debug', "evaluating interrupt assignments");
1335		next;
1336	}
1337
1338
1339	# 4. Incorporate new delta into the list of deltas, and associated
1340	# statistics. If we've just now received $statslen deltas, then it's
1341	# time to evaluate a reconfiguration.
1342
1343	$below_statslen = ($deltas_tottime < $statslen);
1344	$deltas_tottime += $deltatime;
1345	$do_reconfig = ($below_statslen && $deltas_tottime >= $statslen);
1346	push(@deltas, $delta);
1347
1348	# 5. Remove old deltas if total time is more than $statslen. We use
1349	# @deltas as a moving average of the last $statslen seconds. Shift
1350	# off the olders deltas, but only if that doesn't cause us to fall
1351	# below $statslen seconds.
1352
1353	while (@deltas > 1) {
1354		$olddelta = $deltas[0];
1355		$olddeltatime = $olddelta->{maxsnap} - $olddelta->{minsnap};
1356		$newtime = $deltas_tottime - $olddeltatime;
1357		last if ($newtime < $statslen);
1358
1359		shift(@deltas);
1360		$deltas_tottime = $newtime;
1361	}
1362
1363	# 6. The brains of the operation are here. First, check if we're
1364	# imbalanced, and if so set $do_reconfig. If $do_reconfig is set,
1365	# either because of imbalance or above in step 4, we evaluate a
1366	# new configuration.
1367	#
1368	# First, take @deltas and generate a single "compressed" delta
1369	# which summarizes them all. Pass that to do_reconfig and see
1370	# what it does with it:
1371	#
1372	# $ret == -1 : failure
1373	# $ret ==  0 : current config is optimal (or close enough)
1374	# $ret ==  1 : reconfiguration has occurred
1375	#
1376	# If $ret is -1 or 1, dump all our deltas and start from scratch.
1377	# Step 4 above will set do_reconfig soon thereafter.
1378	#
1379	# If $ret is 0, then nothing has happened because we're already
1380	# good enough. Set baseline_goodness to current goodness.
1381
1382	$compdelta = compress_deltas(\@deltas);
1383	if (VERIFY(ref($compdelta) eq "HASH", "couldn't compress deltas")) {
1384		clear_deltas();
1385		next;
1386	}
1387	$compdelta->{goodness} = goodness($compdelta);
1388	dumpdelta($compdelta) if $debug;
1389
1390	$goodness = $compdelta->{goodness};
1391	syslog('debug', "GOODNESS: %5.2f%%", $goodness * 100);
1392
1393	if ($deltas_tottime >= $statslen &&
1394	    imbalanced($goodness, $baseline_goodness)) {
1395		$do_reconfig = 1;
1396	}
1397
1398	if ($do_reconfig) {
1399		$ret = do_reconfig($compdelta);
1400
1401		if ($ret != 0) {
1402			clear_deltas();
1403			syslog('debug', "do_reconfig FAILED!") if $ret == -1;
1404		} else {
1405			syslog('debug', "setting new baseline of $goodness");
1406			$baseline_goodness = $goodness;
1407		}
1408	}
1409	syslog('debug', "---------------------------------------");
1410}
1411