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, Version 1.0 only 7# (the "License"). You may not use this file except in compliance 8# with the License. 9# 10# You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE 11# or http://www.opensolaris.org/os/licensing. 12# See the License for the specific language governing permissions 13# and limitations under the License. 14# 15# When distributing Covered Code, include this CDDL HEADER in each 16# file and include the License file at usr/src/OPENSOLARIS.LICENSE. 17# If applicable, add the following below this CDDL HEADER, with the 18# fields enclosed by brackets "[]" replaced with your own identifying 19# information: Portions Copyright [yyyy] [name of copyright owner] 20# 21# CDDL HEADER END 22# 23 24# 25# Copyright 2005 Sun Microsystems, Inc. All rights reserved. 26# Use is subject to license terms. 27# 28#ident "%Z%%M% %I% %E% SMI" 29# 30 31require 5.6.1; 32use strict; 33use warnings; 34use POSIX; 35use File::Basename("basename"); 36 37my $cmdname = basename($0); 38 39my $using_scengen = 0; # 1 if using scenario simulator 40my $debug = 0; 41 42my $min_sleeptime = 1; 43my $max_sleeptime = 15; 44my $onecpu_sleeptime = (60 * 15); # used if only 1 CPU on system 45my $sleeptime = $min_sleeptime; # time to sleep between kstat updates 46 47# For timerange_foo variables, see comments at tail of &getstat() 48 49my $timerange_toohi = .01; 50my $timerange_hithresh = .0003; 51my $timerange_lothresh = $timerange_hithresh / 2; 52my $unsafe_timerange = .02; 53 54my $statslen = 60; # time period (in secs) to keep in @deltas 55 56 57# Parse arguments. intrd does not accept any public arguments; the two 58# arguments below are meant for testing purposes. -D generates a significant 59# amount of syslog output. -S <filename> loads the filename as a perl 60# script. That file is expected to implement a kstat "simulator" which 61# can be used to feed information to intrd and verify intrd's responses. 62 63while ($_ = shift @ARGV) { 64 if ($_ eq "-S" && $#ARGV != -1) { 65 $using_scengen = 1; 66 do $ARGV[0]; # load simulator 67 shift @ARGV; 68 } elsif ($_ eq "-D") { 69 $debug = 1; 70 } 71} 72 73if ($using_scengen == 0) { 74 require Sun::Solaris::Kstat; 75 require Sun::Solaris::Intrs; 76 import Sun::Solaris::Intrs(qw(intrmove)); 77 require Sys::Syslog; 78 import Sys::Syslog; 79 openlog($cmdname, 'pid', 'daemon'); 80 setlogmask(Sys::Syslog::LOG_UPTO($debug > 0 ? &Sys::Syslog::LOG_DEBUG : 81 &Sys::Syslog::LOG_INFO)); 82} 83 84 85my $asserted = 0; 86my $assert_level = 'debug'; # syslog level for assertion failures 87sub VERIFY($@) 88{ 89 my $bad = (shift() == 0); # $_[0] == 0 means assert failed 90 if ($bad) { 91 my $msg = shift(); 92 syslog($assert_level, "VERIFY: $msg", @_); 93 $asserted++; 94 } 95 return ($bad); 96} 97 98 99 100 101sub getstat($); 102sub generate_delta($$); 103sub compress_deltas($); 104sub dumpdelta($); 105 106sub goodness($); 107sub imbalanced($$); 108sub do_reconfig($); 109 110sub goodness_cpu($$); # private function 111sub move_intr($$$$); # private function 112sub ivecs_to_string(@); # private function 113sub do_find_goal($$$$); # private function 114sub find_goal($$); # private function 115sub do_reconfig_cpu2cpu($$$$); # private function 116sub do_reconfig_cpu($$$); # private function 117 118 119# 120# What follow are the basic data structures routines of intrd. 121# 122# getstat() is responsible for reading the kstats and generating a "stat" hash. 123# 124# generate_delta() is responsible for taking two "stat" hashes and creating 125# a new "delta" hash that represents what has changed over time. 126# 127# compress_deltas() is responsible for taking a list of deltas and generating 128# a single delta hash that encompasses all the time periods described by the 129# deltas. 130 131 132# 133# getstat() is handed a reference to a kstat and generates a hash, returned 134# by reference, containing all the fields from the kstats which we need. 135# If it returns the scalar 0, it failed to gather the kstats, and the caller 136# should react accordingly. 137# 138# getstat() is also responsible for maintaining a reasonable $sleeptime. 139# 140# {"snaptime"} kstat's snaptime 141# {<cpuid>} one hash reference per online cpu 142# ->{"tot"} == cpu:<cpuid>:sys:cpu_nsec_{user + kernel + idle} 143# ->{"crtime"} == cpu:<cpuid>:sys:crtime 144# ->{"ivecs"} 145# ->{<cookie#>} iterates over pci_intrs::config:cookie 146# ->{"time"} == pci_intrs:<ivec#>:config:time (in nsec) 147# ->{"pil"} == pci_intrs:<ivec#>:config:pil 148# ->{"crtime"} == pci_intrs:<ivec#>:config:crtime 149# ->{"ino"} == pci_intrs:<ivec#>:config:ino 150# ->{"buspath"} == pci_intrs:<ivec#>:config:buspath 151# ->{"name"} == pci_intrs:<ivec#>:config:name 152# ->{"ihs"} == pci_intrs:<ivec#>:config:ihs 153# 154 155sub getstat($) 156{ 157 my ($ks) = @_; 158 159 my $cpucnt = 0; 160 my %stat = (); 161 my ($minsnap, $maxsnap); 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 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}:config:time, snaptime, crtime, pil, 213 # ino, name, and buspath. Check $minsnap/$maxsnap. 214 215 foreach my $inst (values(%{$ks->{pci_intrs}})) { 216 my $intrcfg = $inst->{config}; 217 my $cpu = $intrcfg->{cpu}; 218 219 next unless exists $stat{$cpu}; 220 221 if ($intrcfg->{snaptime} < $minsnap) { 222 $minsnap = $intrcfg->{snaptime}; 223 } elsif ($intrcfg->{snaptime} > $maxsnap) { 224 $maxsnap = $intrcfg->{snaptime}; 225 } 226 227 my $cookie = "$intrcfg->{buspath} $intrcfg->{ino}"; 228 if (exists $stat{$cpu}{ivecs}{$cookie}) { 229 my $cookiestats = $stat{$cpu}{ivecs}{$cookie}; 230 231 $cookiestats->{time} += $intrcfg->{time}; 232 $cookiestats->{name} .= "/$intrcfg->{name}"; 233 234 # If this new interrupt sharing $cookie represents a 235 # change from an earlier getstat, make sure that 236 # generate_delta will see the change by setting 237 # crtime to the most recent crtime of its components. 238 239 if ($intrcfg->{crtime} > $cookiestats->{crtime}) { 240 $cookiestats->{crtime} = $intrcfg->{crtime}; 241 } 242 $cookiestats->{ihs}++; 243 next; 244 } 245 $stat{$cpu}{ivecs}{$cookie}{time} = $intrcfg->{time}; 246 $stat{$cpu}{ivecs}{$cookie}{crtime} = $intrcfg->{crtime}; 247 $stat{$cpu}{ivecs}{$cookie}{pil} = $intrcfg->{pil}; 248 $stat{$cpu}{ivecs}{$cookie}{ino} = $intrcfg->{ino}; 249 $stat{$cpu}{ivecs}{$cookie}{buspath} = $intrcfg->{buspath}; 250 $stat{$cpu}{ivecs}{$cookie}{name} = $intrcfg->{name}; 251 $stat{$cpu}{ivecs}{$cookie}{ihs} = 1; 252 } 253 254 # We define the timerange as the amount of time spent gathering the 255 # various kstats, divided by our sleeptime. If we take a lot of time 256 # to access the kstats, and then we create a delta comparing these 257 # kstats with a prior set of kstats, that delta will cover 258 # substaintially different amount of time depending upon which 259 # interrupt or CPU is being examined. 260 # 261 # By checking the timerange here, we guarantee that any deltas 262 # created from these kstats will contain self-consistent data, 263 # in that all CPUs and interrupts cover a similar span of time. 264 # 265 # We attempt to keep this timerange between $timerange_lothresh and 266 # $timerange_hithresh. If the timerange gets too large, not only are 267 # there the accuracy concerns above, but it means that intrd is using 268 # a lot of CPU time. If the timerange gets too small, that means our 269 # sleep time is large, and we could fail to react quickly enough to a 270 # sudden change. 271 # 272 # Finally, $timerange_toohi is the upper bound. Any timerange above 273 # this is thrown out as garbage. If the stat is safely within this 274 # bound, we treat the stat as representing an instant in time, rather 275 # than the time range it actually spans. We arbitrarily choose minsnap 276 # as the snaptime of the stat. 277 278 $stat{snaptime} = $minsnap; 279 my $timerange = ($maxsnap - $minsnap) / $sleeptime; 280 if ($sleeptime == $onecpu_sleeptime) { 281 $sleeptime = $min_sleeptime; # time to come out of idling 282 } elsif ($timerange > $timerange_hithresh && 283 $sleeptime < $max_sleeptime) { 284 $sleeptime++; 285 } elsif ($timerange < $timerange_lothresh && 286 $sleeptime > $min_sleeptime) { 287 $sleeptime--; 288 } 289 return (0) if ($timerange > $timerange_toohi); # i.e. failure 290 return (\%stat); 291} 292 293# 294# dumpdelta takes a reference to our "delta" structure: 295# {"missing"} "1" if the delta's component stats had inconsistencies 296# {"minsnap"} time of the first kstat snaptime used in this delta 297# {"maxsnap"} time of the last kstat snaptime used in this delta 298# {"goodness"} cost function applied to this delta 299# {"avgintrload"} avg of interrupt load across cpus, as a percentage 300# {"avgintrnsec"} avg number of nsec spent in interrupts, per cpu 301# {<cpuid>} iterates over on-line cpus 302# ->{"intrs"} cpu's movable intr time (sum of "time" for each ivec) 303# ->{"tot"} CPU load from all sources 304# ->{"bigintr"} largest value of {ivecs}{<ivec#>}{time} from below 305# ->{"intrload"} intrs / tot 306# ->{"ivecs"} 307# ->{<ivec#>} iterates over ivecs for this cpu 308# ->{"time"} time used by this interrupt (in nsec) 309# ->{"pil"} pil level of this interrupt 310# ->{"ino"} interrupt number 311# ->{"buspath"} filename of the directory of the device's bus 312# ->{"name"} device name 313# ->{"ihs"} number of different handlers sharing this ino 314# 315# It prints out the delta structure in a nice, human readable display. 316# 317 318sub dumpdelta($) 319{ 320 my ($delta) = @_; 321 322 # print global info 323 324 syslog('debug', "dumpdelta:"); 325 syslog('debug', " RECONFIGURATION IN DELTA") if $delta->{missing} > 0; 326 syslog('debug', " avgintrload: %5.2f%% avgintrnsec: %d", 327 $delta->{avgintrload} * 100, $delta->{avgintrnsec}); 328 syslog('debug', " goodness: %5.2f%%", $delta->{goodness} * 100) 329 if exists($delta->{goodness}); 330 331 # iterate over cpus 332 333 while (my ($cpu, $cpst) = each %$delta) { 334 next if !ref($cpst); # skip non-cpuid entries 335 my $tot = $cpst->{tot}; 336 syslog('debug', " cpu %3d intr %7.3f%% (bigintr %7.3f%%)", 337 $cpu, $cpst->{intrload}*100, $cpst->{bigintr}*100/$tot); 338 syslog('debug', " intrs %d, bigintr %d", 339 $cpst->{intrs}, $cpst->{bigintr}); 340 341 # iterate over ivecs on this cpu 342 343 while (my ($ivec, $ivst) = each %{$cpst->{ivecs}}) { 344 syslog('debug', " %15s:\"%s\": %7.3f%% %d", 345 ($ivst->{ihs} > 1 ? "$ivst->{name}($ivst->{ihs})" : 346 $ivst->{name}), $ivec, 347 $ivst->{time}*100 / $tot, $ivst->{time}); 348 } 349 } 350} 351 352# 353# generate_delta($stat, $newstat) takes two stat references, returned from 354# getstat(), and creates a %delta. %delta (not surprisingly) contains the 355# same basic info as stat and newstat, but with the timestamps as deltas 356# instead of absolute times. We return a reference to the delta. 357# 358 359sub generate_delta($$) 360{ 361 my ($stat, $newstat) = @_; 362 363 my %delta = (); 364 my $intrload; 365 my $intrnsec; 366 my $cpus; 367 368 # Take the worstcase timerange 369 $delta{minsnap} = $stat->{snaptime}; 370 $delta{maxsnap} = $newstat->{snaptime}; 371 if (VERIFY($delta{maxsnap} > $delta{minsnap}, 372 "generate_delta: stats aren't ascending")) { 373 $delta{missing} = 1; 374 return (\%delta); 375 } 376 377 # if there are a different number of cpus in the stats, set missing 378 379 $delta{missing} = (keys(%$stat) != keys(%$newstat)); 380 if (VERIFY($delta{missing} == 0, 381 "generate_delta: number of CPUs changed")) { 382 return (\%delta); 383 } 384 385 # scan through every cpu in %newstat and compare against %stat 386 387 while (my ($cpu, $newcpst) = each %$newstat) { 388 next if !ref($newcpst); # skip non-cpuid fields 389 390 # If %stat is missing a cpu from %newstat, then it was just 391 # onlined. Mark missing. 392 393 if (VERIFY(exists $stat->{$cpu} && 394 $stat->{$cpu}{crtime} == $newcpst->{crtime}, 395 "generate_delta: cpu $cpu changed")) { 396 $delta{missing} = 1; 397 return (\%delta); 398 } 399 my $cpst = $stat->{$cpu}; 400 $delta{$cpu}{tot} = $newcpst->{tot} - $cpst->{tot}; 401 if (VERIFY($delta{$cpu}{tot} >= 0, 402 "generate_delta: deltas are not ascending?")) { 403 $delta{missing} = 1; 404 delete($delta{$cpu}); 405 return (\%delta); 406 } 407 # Avoid remote chance of division by zero 408 $delta{$cpu}{tot} = 1 if $delta{$cpu}{tot} == 0; 409 $delta{$cpu}{intrs} = 0; 410 $delta{$cpu}{bigintr} = 0; 411 412 my %ivecs = (); 413 $delta{$cpu}{ivecs} = \%ivecs; 414 415 # if the number of ivecs differs, set missing 416 417 if (VERIFY(keys(%{$cpst->{ivecs}}) == 418 keys(%{$newcpst->{ivecs}}), 419 "generate_delta: cpu $cpu has more/less". 420 " interrupts")) { 421 $delta{missing} = 1; 422 return (\%delta); 423 } 424 425 while (my ($inum, $newivec) = each %{$newcpst->{ivecs}}) { 426 # If this ivec doesn't exist in $stat, or if $stat 427 # shows a different crtime, set missing. 428 429 if (VERIFY(exists $cpst->{ivecs}{$inum} && 430 $cpst->{ivecs}{$inum}{crtime} == 431 $newivec->{crtime}, 432 "generate_delta: cpu $cpu inum $inum". 433 " has changed")) { 434 $delta{missing} = 1; 435 return (\%delta); 436 } 437 my $ivec = $cpst->{ivecs}{$inum}; 438 439 # Create $delta{$cpu}{ivecs}{$inum}. 440 441 my %dltivec = (); 442 $delta{$cpu}{ivecs}{$inum} = \%dltivec; 443 444 # calculate time used by this interrupt 445 446 my $time = $newivec->{time} - $ivec->{time}; 447 if (VERIFY($time >= 0, 448 "generate_delta: ivec went backwards?")) { 449 $delta{missing} = 1; 450 delete($delta{$cpu}{ivecs}{$inum}); 451 return (\%delta); 452 } 453 $delta{$cpu}{intrs} += $time; 454 $dltivec{time} = $time; 455 if ($time > $delta{$cpu}{bigintr}) { 456 $delta{$cpu}{bigintr} = $time; 457 } 458 459 # Transfer over basic info about the kstat. We 460 # don't have to worry about discrepancies between 461 # ivec and newivec because we verified that both 462 # have the same crtime. 463 464 $dltivec{pil} = $newivec->{pil}; 465 $dltivec{ino} = $newivec->{ino}; 466 $dltivec{buspath} = $newivec->{buspath}; 467 $dltivec{name} = $newivec->{name}; 468 $dltivec{ihs} = $newivec->{ihs}; 469 } 470 if ($delta{$cpu}{tot} < $delta{$cpu}{intrs}) { 471 # Ewww! Hopefully just a rounding error. 472 # Make something up. 473 $delta{$cpu}{tot} = $delta{$cpu}{intrs}; 474 } 475 $delta{$cpu}{intrload} = 476 $delta{$cpu}{intrs} / $delta{$cpu}{tot}; 477 $intrload += $delta{$cpu}{intrload}; 478 $intrnsec += $delta{$cpu}{intrs}; 479 $cpus++; 480 } 481 if ($cpus > 0) { 482 $delta{avgintrload} = $intrload / $cpus; 483 $delta{avgintrnsec} = $intrnsec / $cpus; 484 } else { 485 $delta{avgintrload} = 0; 486 $delta{avgintrnsec} = 0; 487 } 488 return (\%delta); 489} 490 491 492# compress_delta takes a list of deltas, and returns a single new delta 493# which represents the combined information from all the deltas. The deltas 494# provided are assumed to be sequential in time. The resulting compressed 495# delta looks just like any other delta. This new delta is also more accurate 496# since its statistics are averaged over a longer period than any of the 497# original deltas. 498 499sub compress_deltas ($) 500{ 501 my ($deltas) = @_; 502 503 my %newdelta = (); 504 my ($intrs, $tot); 505 my $cpus = 0; 506 507 if (VERIFY($#$deltas != -1, 508 "compress_deltas: list of delta is empty?")) { 509 return (0); 510 } 511 $newdelta{minsnap} = $deltas->[0]{minsnap}; 512 $newdelta{maxsnap} = $deltas->[$#$deltas]{maxsnap}; 513 $newdelta{missing} = 0; 514 515 foreach my $delta (@$deltas) { 516 if (VERIFY($delta->{missing} == 0, 517 "compressing bad deltas?")) { 518 return (0); 519 } 520 while (my ($cpuid, $cpu) = each %$delta) { 521 next if !ref($cpu); 522 523 $intrs += $cpu->{intrs}; 524 $tot += $cpu->{tot}; 525 $newdelta{$cpuid}{intrs} += $cpu->{intrs}; 526 $newdelta{$cpuid}{tot} += $cpu->{tot}; 527 if (!exists $newdelta{$cpuid}{ivecs}) { 528 my %ivecs = (); 529 $newdelta{$cpuid}{ivecs} = \%ivecs; 530 } 531 while (my ($inum, $ivec) = each %{$cpu->{ivecs}}) { 532 my $newivecs = $newdelta{$cpuid}{ivecs}; 533 $newivecs->{$inum}{time} += $ivec->{time}; 534 $newivecs->{$inum}{pil} = $ivec->{pil}; 535 $newivecs->{$inum}{ino} = $ivec->{ino}; 536 $newivecs->{$inum}{buspath} = $ivec->{buspath}; 537 $newivecs->{$inum}{name} = $ivec->{name}; 538 $newivecs->{$inum}{ihs} = $ivec->{ihs}; 539 } 540 } 541 } 542 foreach my $cpu (values(%newdelta)) { 543 next if !ref($cpu); # ignore non-cpu fields 544 $cpus++; 545 546 my $bigintr = 0; 547 foreach my $ivec (values(%{$cpu->{ivecs}})) { 548 if ($ivec->{time} > $bigintr) { 549 $bigintr = $ivec->{time}; 550 } 551 } 552 $cpu->{bigintr} = $bigintr; 553 $cpu->{intrload} = $cpu->{intrs} / $cpu->{tot}; 554 $cpu->{tot} = 1 if $cpu->{tot} <= 0; 555 } 556 if ($cpus == 0) { 557 $newdelta{avgintrnsec} = 0; 558 $newdelta{avgintrload} = 0; 559 } else { 560 $newdelta{avgintrnsec} = $intrs / $cpus; 561 $newdelta{avgintrload} = $intrs / $tot; 562 } 563 return (\%newdelta); 564} 565 566 567 568 569 570# What follow are the core functions responsible for examining the deltas 571# generated above and deciding what to do about them. 572# 573# goodness() and its helper goodness_cpu() return a heuristic which describe 574# how good (or bad) the current interrupt balance is. The value returned will 575# be between 0 and 1, with 0 representing maximum goodness, and 1 representing 576# maximum badness. 577# 578# imbalanced() compares a current and historical value of goodness, and 579# determines if there has been enough change to warrant evaluating a 580# reconfiguration of the interrupts 581# 582# do_reconfig(), and its helpers, do_reconfig_cpu(), do_reconfig_cpu2cpu(), 583# find_goal(), do_find_goal(), and move_intr(), are responsible for examining 584# a delta and determining the best possible assignment of interrupts to CPUs. 585# 586# It is important that do_reconfig() be in alignment with goodness(). If 587# do_reconfig were to generate a new interrupt distribution that worsened 588# goodness, we could get into a pathological loop with intrd fighting itself, 589# constantly deciding that things are imbalanced, and then changing things 590# only to make them worse. 591 592 593 594# any goodness over $goodness_unsafe_load is considered really bad 595# goodness must drop by at least $goodness_mindelta for a reconfig 596 597my $goodness_unsafe_load = .9; 598my $goodness_mindelta = .1; 599 600# goodness(%delta) examines a delta and return its "goodness". goodness will 601# be between 0 (best) and 1 (major bad). goodness is determined by evaluating 602# the goodness of each individual cpu, and returning the worst case. This 603# helps on systems with many CPUs, where otherwise a single pathological CPU 604# might otherwise be ignored because the average was OK. 605# 606# To calculate the goodness of an individual CPU, we start by looking at its 607# load due to interrupts. If the load is above a certain high threshold and 608# there is more than one interrupt assigned to this CPU, we set goodness 609# to worst-case. If the load is below the average interrupt load of all CPUs, 610# then we return best-case, since what's to complain about? 611# 612# Otherwise we look at how much the load is above the average, and return 613# that as the goodness, with one caveat: we never return more than the CPU's 614# interrupt load ignoring its largest single interrupt source. This is 615# because a CPU with one high-load interrupt, and no other interrupts, is 616# perfectly balanced. Nothing can be done to improve the situation, and thus 617# it is perfectly balanced even if the interrupt's load is 100%. 618 619sub goodness($) 620{ 621 my ($delta) = @_; 622 623 return (1) if $delta->{missing} > 0; 624 625 my $high_goodness = 0; 626 my $goodness; 627 628 foreach my $cpu (values(%$delta)) { 629 next if !ref($cpu); # skip non-cpuid fields 630 631 $goodness = goodness_cpu($cpu, $delta->{avgintrload}); 632 if (VERIFY($goodness >= 0 && $goodness <= 1, 633 "goodness: cpu goodness out of range?")) { 634 dumpdelta($delta); 635 return (1); 636 } 637 if ($goodness == 1) { 638 return (1); # worst case, no need to continue 639 } 640 if ($goodness > $high_goodness) { 641 $high_goodness = $goodness; 642 } 643 } 644 return ($high_goodness); 645} 646 647sub goodness_cpu($$) # private function 648{ 649 my ($cpu, $avgintrload) = @_; 650 651 my $goodness; 652 my $load = $cpu->{intrs} / $cpu->{tot}; 653 654 return (0) if ($load < $avgintrload); # low loads are perfectly good 655 656 # Calculate $load_no_bigintr, which represents the load 657 # due to interrupts, excluding the one biggest interrupt. 658 # This is the most gain we can get on this CPU from 659 # offloading interrupts. 660 661 my $load_no_bigintr = ($cpu->{intrs} - $cpu->{bigintr}) / $cpu->{tot}; 662 663 # A major imbalance is indicated if a CPU is saturated 664 # with interrupt handling, and it has more than one 665 # source of interrupts. Those other interrupts could be 666 # starved if of a lower pil. Return a goodness of 1, 667 # which is the worst possible return value, 668 # which will effectively contaminate this entire delta. 669 670 my $cnt = keys(%{$cpu->{ivecs}}); 671 672 if ($load > $goodness_unsafe_load && $cnt > 1) { 673 return (1); 674 } 675 $goodness = $load - $avgintrload; 676 if ($goodness > $load_no_bigintr) { 677 $goodness = $load_no_bigintr; 678 } 679 return ($goodness); 680} 681 682 683# imbalanced() is used by the main routine to determine if the goodness 684# has shifted far enough from our last baseline to warrant a reassignment 685# of interrupts. A very high goodness indicates that a CPU is way out of 686# whack. If the goodness has varied too much since the baseline, then 687# perhaps a reconfiguration is worth considering. 688 689sub imbalanced ($$) 690{ 691 my ($goodness, $baseline) = @_; 692 693 # Return 1 if we are pathological, or creeping away from the baseline 694 695 return (1) if $goodness > .50; 696 return (1) if abs($goodness - $baseline) > $goodness_mindelta; 697 return (0); 698} 699 700# do_reconfig(), do_reconfig_cpu(), and do_reconfig_cpu2cpu(), are the 701# decision-making functions responsible for generating a new interrupt 702# distribution. They are designed with the definition of goodness() in 703# mind, i.e. they use the same definition of "good distribution" as does 704# goodness(). 705# 706# do_reconfig() is responsible for deciding whether a redistribution is 707# actually warranted. If the goodness is already pretty good, it doesn't 708# waste the CPU time to generate a new distribution. If it 709# calculates a new distribution and finds that it is not sufficiently 710# improved from the prior distirbution, it will not do the redistribution, 711# mainly to avoid the disruption to system performance caused by 712# rejuggling interrupts. 713# 714# Its main loop works by going through a list of cpus sorted from 715# highest to lowest interrupt load. It removes the highest-load cpus 716# one at a time and hands them off to do_reconfig_cpu(). This function 717# then re-sorts the remaining CPUs from lowest to highest interrupt load, 718# and one at a time attempts to rejuggle interrupts between the original 719# high-load CPU and the low-load CPU. Rejuggling on a high-load CPU is 720# considered finished as soon as its interrupt load is within 721# $goodness_mindelta of the average interrupt load. Such a CPU will have 722# a goodness of below the $goodness_mindelta threshold. 723 724# 725# move_intr(\%delta, $inum, $oldcpu, $newcpu) 726# used by reconfiguration code to move an interrupt between cpus within 727# a delta. This manipulates data structures, and does not actually move 728# the interrupt on the running system. 729# 730sub move_intr($$$$) # private function 731{ 732 my ($delta, $inum, $oldcpuid, $newcpuid) = @_; 733 734 my $ivec = $delta->{$oldcpuid}{ivecs}{$inum}; 735 736 # Remove ivec from old cpu 737 738 my $oldcpu = $delta->{$oldcpuid}; 739 $oldcpu->{intrs} -= $ivec->{time}; 740 $oldcpu->{intrload} = $oldcpu->{intrs} / $oldcpu->{tot}; 741 delete($oldcpu->{ivecs}{$inum}); 742 743 VERIFY($oldcpu->{intrs} >= 0, "move_intr: intr's time > total time?"); 744 VERIFY($ivec->{time} <= $oldcpu->{bigintr}, 745 "move_intr: intr's time > bigintr?"); 746 747 if ($ivec->{time} >= $oldcpu->{bigintr}) { 748 my $bigtime = 0; 749 750 foreach my $ivec (values(%{$oldcpu->{ivecs}})) { 751 $bigtime = $ivec->{time} if $ivec->{time} > $bigtime; 752 } 753 $oldcpu->{bigintr} = $bigtime; 754 } 755 756 # Add ivec onto new cpu 757 758 my $newcpu = $delta->{$newcpuid}; 759 760 $ivec->{nowcpu} = $newcpuid; 761 $newcpu->{intrs} += $ivec->{time}; 762 $newcpu->{intrload} = $newcpu->{intrs} / $newcpu->{tot}; 763 $newcpu->{ivecs}{$inum} = $ivec; 764 765 $newcpu->{bigintr} = $ivec->{time} 766 if $ivec->{time} > $newcpu->{bigintr}; 767} 768 769sub move_intr_check($$$) # private function 770{ 771 my ($delta, $oldcpuid, $newcpuid) = @_; 772 773 VERIFY($delta->{$oldcpuid}{tot} >= $delta->{$oldcpuid}{intrs}, 774 "Moved interrupts left 100+%% load on src cpu"); 775 VERIFY($delta->{$newcpuid}{tot} >= $delta->{$newcpuid}{intrs}, 776 "Moved interrupts left 100+%% load on tgt cpu"); 777} 778 779sub ivecs_to_string(@) # private function 780{ 781 my $str = ""; 782 foreach my $ivec (@_) { 783 $str = "$str $ivec->{inum}"; 784 } 785 return ($str); 786} 787 788 789sub do_reconfig($) 790{ 791 my ($delta) = @_; 792 793 my $goodness = $delta->{goodness}; 794 795 # We can't improve goodness to better than 0. We should stop here 796 # if, even if we achieve a goodness of 0, the improvement is still 797 # too small to merit the action. 798 799 if ($goodness - 0 < $goodness_mindelta) { 800 syslog('debug', "goodness good enough, don't reconfig"); 801 return (0); 802 } 803 804 syslog('notice', "Optimizing interrupt assignments"); 805 806 if (VERIFY ($delta->{missing} == 0, "RECONFIG Aborted: should not ". 807 "have a delta with missing")) { 808 return (-1); 809 } 810 811 # Make a list of all cpuids, and also add some extra information 812 # to the ivec structures. 813 814 my @cpusortlist = (); 815 816 while (my ($cpuid, $cpu) = each %$delta) { 817 next if !ref($cpu); # skip non-cpu entries 818 819 push(@cpusortlist, $cpuid); 820 while (my ($inum, $ivec) = each %{$cpu->{ivecs}}) { 821 $ivec->{origcpu} = $cpuid; 822 $ivec->{nowcpu} = $cpuid; 823 $ivec->{inum} = $inum; 824 } 825 } 826 827 # Sort the list of CPUs from highest to lowest interrupt load. 828 # Remove the top CPU from that list and attempt to redistribute 829 # its interrupts. If the CPU has a goodness below a threshold, 830 # just ignore the CPU and move to the next one. If the CPU's 831 # load falls below the average load plus that same threshold, 832 # then there are no CPUs left worth reconfiguring, and we're done. 833 834 while (@cpusortlist) { 835 # Re-sort cpusortlist each time, since do_reconfig_cpu can 836 # move interrupts around. 837 838 @cpusortlist = 839 sort({$delta->{$b}{intrload} <=> $delta->{$a}{intrload}} 840 @cpusortlist); 841 842 my $cpu = shift(@cpusortlist); 843 if (($delta->{$cpu}{intrload} <= $goodness_unsafe_load) && 844 ($delta->{$cpu}{intrload} <= 845 $delta->{avgintrload} + $goodness_mindelta)) { 846 syslog('debug', "finished reconfig: cpu $cpu load ". 847 "$delta->{$cpu}{intrload} avgload ". 848 "$delta->{avgintrload}"); 849 last; 850 } 851 if (goodness_cpu($delta->{$cpu}, $delta->{avgintrload}) < 852 $goodness_mindelta) { 853 next; 854 } 855 do_reconfig_cpu($delta, \@cpusortlist, $cpu); 856 } 857 858 # How good a job did we do? If the improvement was minimal, and 859 # our goodness wasn't pathological (and thus needing any help it 860 # can get), then don't bother moving the interrupts. 861 862 my $newgoodness = goodness($delta); 863 VERIFY($newgoodness <= $goodness, 864 "reconfig: result has worse goodness?"); 865 866 if (($goodness != 1 || $newgoodness == 1) && 867 $goodness - $newgoodness < $goodness_mindelta) { 868 syslog('debug', "goodness already near optimum, ". 869 "don't reconfig"); 870 return (0); 871 } 872 syslog('debug', "goodness %5.2f%% --> %5.2f%%", $goodness*100, 873 $newgoodness*100); 874 875 # Time to move those interrupts! 876 877 my $ret = 1; 878 my $warned = 0; 879 while (my ($cpuid, $cpu) = each %$delta) { 880 next if $cpuid =~ /\D/; 881 while (my ($inum, $ivec) = each %{$cpu->{ivecs}}) { 882 next if ($ivec->{origcpu} == $cpuid); 883 884 if (!intrmove($ivec->{buspath}, $ivec->{ino}, 885 $cpuid)) { 886 syslog('warning', "Unable to move interrupts") 887 if $warned++ == 0; 888 syslog('debug', "Unable to move buspath ". 889 "$ivec->{buspath} ino $ivec->{ino} to ". 890 "cpu $cpuid"); 891 $ret = -1; 892 } 893 } 894 } 895 896 syslog('notice', "Interrupt assignments optimized"); 897 return ($ret); 898} 899 900sub do_reconfig_cpu($$$) # private function 901{ 902 my ($delta, $cpusortlist, $oldcpuid) = @_; 903 904 # We have been asked to rejuggle interrupts between $oldcpuid and 905 # other CPUs found on $cpusortlist so as to improve the load on 906 # $oldcpuid. We reverse $cpusortlist to get our own copy of the 907 # list, sorted from lowest to highest interrupt load. One at a 908 # time, shift a CPU off of this list of CPUs, and attempt to 909 # rejuggle interrupts between the two CPUs. Don't do this if the 910 # other CPU has a higher load than oldcpuid. We're done rejuggling 911 # once $oldcpuid's goodness falls below a threshold. 912 913 syslog('debug', "reconfiguring $oldcpuid"); 914 915 my $cpu = $delta->{$oldcpuid}; 916 my $avgintrload = $delta->{avgintrload}; 917 918 my @cputargetlist = reverse(@$cpusortlist); # make a copy of the list 919 while ($#cputargetlist != -1) { 920 last if goodness_cpu($cpu, $avgintrload) < $goodness_mindelta; 921 922 my $tgtcpuid = shift(@cputargetlist); 923 my $tgt = $delta->{$tgtcpuid}; 924 my $load = $cpu->{intrload}; 925 my $tgtload = $tgt->{intrload}; 926 last if $tgtload > $load; 927 do_reconfig_cpu2cpu($delta, $oldcpuid, $tgtcpuid, $load); 928 } 929} 930 931sub do_reconfig_cpu2cpu($$$$) # private function 932{ 933 my ($delta, $srccpuid, $tgtcpuid, $srcload) = @_; 934 935 # We've been asked to consider interrupt juggling between srccpuid 936 # (with a high interrupt load) and tgtcpuid (with a lower interrupt 937 # load). First, make a single list with all of the ivecs from both 938 # CPUs, and sort the list from highest to lowest load. 939 940 syslog('debug', "exchanging intrs between $srccpuid and $tgtcpuid"); 941 942 # Gather together all the ivecs and sort by load 943 944 my @ivecs = (values(%{$delta->{$srccpuid}{ivecs}}), 945 values(%{$delta->{$tgtcpuid}{ivecs}})); 946 return if $#ivecs == -1; 947 948 @ivecs = sort({$b->{time} <=> $a->{time}} @ivecs); 949 950 # Our "goal" load for srccpuid is the average load across all CPUs. 951 # find_goal() will find determine the optimum selection of the 952 # available interrupts which comes closest to this goal without 953 # falling below the goal. 954 955 my $goal = $delta->{avgintrnsec}; 956 957 # We know that the interrupt load on tgtcpuid is less than that on 958 # srccpuid, but its load could still be above avgintrnsec. Don't 959 # choose a goal which would bring srccpuid below the load on tgtcpuid. 960 961 my $avgnsec = 962 ($delta->{$srccpuid}{intrs} + $delta->{$tgtcpuid}{intrs}) / 2; 963 if ($goal < $avgnsec) { 964 $goal = $avgnsec; 965 } 966 967 # If the largest of the interrupts is on srccpuid, leave it there. 968 # This can help minimize the disruption caused by moving interrupts. 969 970 if ($ivecs[0]->{origcpu} == $srccpuid) { 971 syslog('debug', "Keeping $ivecs[0]->{inum} on $srccpuid"); 972 $goal -= $ivecs[0]->{time}; 973 shift(@ivecs); 974 } 975 976 syslog('debug', "GOAL: inums should total $goal"); 977 find_goal(\@ivecs, $goal); 978 979 # find_goal() returned its results to us by setting $ivec->{goal} if 980 # the ivec should be on srccpuid, or clearing it for tgtcpuid. 981 # Call move_intr() to update our $delta with the new results. 982 983 foreach my $ivec (@ivecs) { 984 syslog('debug', "ivec $ivec->{inum} goal $ivec->{goal}"); 985 VERIFY($ivec->{nowcpu} == $srccpuid || 986 $ivec->{nowcpu} == $tgtcpuid, "cpu2cpu found an ". 987 "interrupt not currently on src or tgt cpu"); 988 989 if ($ivec->{goal} && $ivec->{nowcpu} != $srccpuid) { 990 move_intr($delta, $ivec->{inum}, $ivec->{nowcpu}, 991 $srccpuid); 992 } elsif ($ivec->{goal} == 0 && $ivec->{nowcpu} != $tgtcpuid) { 993 move_intr($delta, $ivec->{inum}, $ivec->{nowcpu}, 994 $tgtcpuid); 995 } 996 } 997 move_intr_check($delta, $srccpuid, $tgtcpuid); # asserts 998 999 my $newload = $delta->{$srccpuid}{intrs} / $delta->{$srccpuid}{tot}; 1000 VERIFY($newload <= $srcload && $newload > $delta->{avgintrload}, 1001 "cpu2cpu: new load didn't end up in expected range"); 1002} 1003 1004 1005# find_goal() and its helper do_find_goal() are used to find the best 1006# combination of interrupts in order to generate a load that is as close 1007# as possible to a goal load without falling below that goal. Before returning 1008# to its caller, find_goal() sets a new value in the hash of each interrupt, 1009# {goal}, which if set signifies that this interrupt is one of the interrupts 1010# identified as part of the set of interrupts which best meet the goal. 1011# 1012# The arguments to find_goal are a list of ivecs (hash references), sorted 1013# by descending {time}, and the goal load. The goal is relative to {time}. 1014# The best fit is determined by performing a depth-first search. do_find_goal 1015# is the recursive subroutine which carries out the search. 1016# 1017# It is passed an index as an argument, originally 0. On a given invocation, 1018# it is only to consider interrupts in the ivecs array starting at that index. 1019# It then considers two possibilities: 1020# 1) What is the best goal-fit if I include ivecs[index]? 1021# 2) What is the best goal-fit if I exclude ivecs[index]? 1022# To determine case 1, it subtracts the load of ivecs[index] from the goal, 1023# and calls itself recursively with that new goal and index++. 1024# To determine case 2, it calls itself recursively with the same goal and 1025# index++. 1026# 1027# It then compares the two results, decide which one best meets the goals, 1028# and returns the result. The return value is the best-fit's interrupt load, 1029# followed by a list of all the interrupts which make up that best-fit. 1030# 1031# As an optimization, a second array loads[] is created which mirrors ivecs[]. 1032# loads[i] will equal the total loads of all ivecs[i..$#ivecs]. This is used 1033# by do_find_goal to avoid recursing all the way to the end of the ivecs 1034# array if including all remaining interrupts will still leave the best-fit 1035# at below goal load. If so, it then includes all remaining interrupts on 1036# the goal list and returns. 1037# 1038sub find_goal($$) # private function 1039{ 1040 my ($ivecs, $goal) = @_; 1041 1042 my @goals; 1043 my $load; 1044 my $ivec; 1045 1046 if ($goal <= 0) { 1047 @goals = (); # the empty set will best meet the goal 1048 } else { 1049 syslog('debug', "finding goal from intrs %s", 1050 ivecs_to_string(@$ivecs)); 1051 1052 # Generate @loads array 1053 1054 my $tot = 0; 1055 foreach $ivec (@$ivecs) { 1056 $tot += $ivec->{time}; 1057 } 1058 my @loads = (); 1059 foreach $ivec (@$ivecs) { 1060 push(@loads, $tot); 1061 $tot -= $ivec->{time}; 1062 } 1063 ($load, @goals) = do_find_goal($ivecs, \@loads, $goal, 0); 1064 VERIFY($load >= $goal, "find_goal didn't meet goals"); 1065 } 1066 syslog('debug', "goals found: %s", ivecs_to_string(@goals)); 1067 1068 # Set or clear $ivec->{goal} for each ivec, based on returned @goals 1069 1070 foreach $ivec (@$ivecs) { 1071 if ($#goals > -1 && $ivec == $goals[0]) { 1072 syslog('debug', "inum $ivec->{inum} on source cpu"); 1073 $ivec->{goal} = 1; 1074 shift(@goals); 1075 } else { 1076 syslog('debug', "inum $ivec->{inum} on target cpu"); 1077 $ivec->{goal} = 0; 1078 } 1079 } 1080} 1081 1082 1083sub do_find_goal($$$$) # private function 1084{ 1085 my ($ivecs, $loads, $goal, $idx) = @_; 1086 1087 if ($idx > $#{$ivecs}) { 1088 return (0); 1089 } 1090 syslog('debug', "$idx: finding goal $goal inum $ivecs->[$idx]{inum}"); 1091 1092 my $load = $ivecs->[$idx]{time}; 1093 my @goals_with = (); 1094 my @goals_without = (); 1095 my ($with, $without); 1096 1097 # If we include all remaining items and we're still below goal, 1098 # stop here. We can just return a result that includes $idx and all 1099 # subsequent ivecs. Since this will still be below goal, there's 1100 # nothing better to be done. 1101 1102 if ($loads->[$idx] <= $goal) { 1103 syslog('debug', 1104 "$idx: including all remaining intrs %s with load %d", 1105 ivecs_to_string(@$ivecs[$idx .. $#{$ivecs}]), 1106 $loads->[$idx]); 1107 return ($loads->[$idx], @$ivecs[$idx .. $#{$ivecs}]); 1108 } 1109 1110 # Evaluate the "with" option, i.e. the best matching goal which 1111 # includes $ivecs->[$idx]. If idx's load is more than our goal load, 1112 # stop here. Once we're above the goal, there is no need to consider 1113 # further interrupts since they'll only take us further from the goal. 1114 1115 if ($goal <= $load) { 1116 $with = $load; # stop here 1117 } else { 1118 ($with, @goals_with) = 1119 do_find_goal($ivecs, $loads, $goal - $load, $idx + 1); 1120 $with += $load; 1121 } 1122 syslog('debug', "$idx: with-load $with intrs %s", 1123 ivecs_to_string($ivecs->[$idx], @goals_with)); 1124 1125 # Evaluate the "without" option, i.e. the best matching goal which 1126 # excludes $ivecs->[$idx]. 1127 1128 ($without, @goals_without) = 1129 &do_find_goal($ivecs, $loads, $goal, $idx + 1); 1130 syslog('debug', "$idx: without-load $without intrs %s", 1131 ivecs_to_string(@goals_without)); 1132 1133 # We now have our "with" and "without" options, and we choose which 1134 # best fits the goal. If one is greater than goal and the other is 1135 # below goal, we choose the one that is greater. If they are both 1136 # below goal, then we choose the one that is greater. If they are 1137 # both above goal, then we choose the smaller. 1138 1139 my $which; # 0 == with, 1 == without 1140 if ($with >= $goal && $without < $goal) { 1141 $which = 0; 1142 } elsif ($with < $goal && $without >= $goal) { 1143 $which = 1; 1144 } elsif ($with >= $goal && $without >= $goal) { 1145 $which = ($without < $with); 1146 } else { 1147 $which = ($without > $with); 1148 } 1149 1150 # Return the load of our best case scenario, followed by all the ivecs 1151 # which compose that goal. 1152 1153 if ($which == 1) { # without 1154 syslog('debug', "$idx: going without"); 1155 return ($without, @goals_without); 1156 } else { 1157 syslog('debug', "$idx: going with"); 1158 return ($with, $ivecs->[$idx], @goals_with); 1159 } 1160 # Not reached 1161} 1162 1163 1164 1165 1166syslog('debug', "intrd is starting".($debug ? " (debug)" : "")); 1167 1168my @deltas = (); 1169my $deltas_tottime = 0; # sum of maxsnap-minsnap across @deltas 1170my $avggoodness; 1171my $baseline_goodness = 0; 1172my $compdelta; 1173 1174my $do_reconfig; 1175 1176# temp variables 1177my $goodness; 1178my $deltatime; 1179my $olddelta; 1180my $olddeltatime; 1181my $delta; 1182my $newstat; 1183my $below_statslen; 1184my $newtime; 1185my $ret; 1186 1187 1188my $gotsig = 0; 1189$SIG{INT} = sub { $gotsig = 1; }; # don't die in the middle of retargeting 1190$SIG{HUP} = $SIG{INT}; 1191$SIG{TERM} = $SIG{INT}; 1192 1193my $ks; 1194if ($using_scengen == 0) { 1195 $ks = Sun::Solaris::Kstat->new(); 1196} else { 1197 $ks = myks_update(); # supplied by the simulator 1198} 1199 1200# If no pci_intrs kstats were found, we need to exit, but we can't because 1201# SMF will restart us and/or report an error to the administrator. But 1202# there's nothing an administrator can do. So print out a message for SMF 1203# logs and silently pause forever. 1204 1205if (!exists($ks->{pci_intrs})) { 1206 print STDERR "$cmdname: no interrupts were found; ". 1207 "your PCI bus may not yet be supported\n"; 1208 pause() while $gotsig == 0; 1209 exit 0; 1210} 1211 1212my $stat = getstat($ks); 1213 1214 1215 1216for (;;) { 1217 sub clear_deltas { 1218 @deltas = (); 1219 $deltas_tottime = 0; 1220 $stat = 0; # prevent next gen_delta() from setting {missing} 1221 } 1222 1223 # 1. Sleep, update the kstats, and save the new stats in $newstat. 1224 1225 exit 0 if $gotsig; # if we got ^C / SIGTERM, exit 1226 if ($using_scengen == 0) { 1227 sleep($sleeptime); 1228 exit 0 if $gotsig; # if we got ^C / SIGTERM, exit 1229 $ks->update(); 1230 } else { 1231 $ks = myks_update(); 1232 } 1233 $newstat = getstat($ks); 1234 1235 # $stat or $newstat could be zero if they're uninitialized, or if 1236 # getstat() failed. If $stat is zero, move $newstat to $stat, sleep 1237 # and try again. If $newstat is zero, then we also sleep and try 1238 # again, hoping the problem will clear up. 1239 1240 next if (!ref $newstat); 1241 if (!ref $stat) { 1242 $stat = $newstat; 1243 next; 1244 } 1245 1246 1247 # 2. Compare $newstat with the prior set of values, result in %$delta. 1248 1249 $delta = generate_delta($stat, $newstat); 1250 dumpdelta($delta) if $debug; # Dump most recent stats to stdout. 1251 $stat = $newstat; # The new stats now become the old stats. 1252 1253 1254 # 3. If $delta->{missing}, then there has been a reconfiguration of 1255 # either cpus or interrupts (probably both). We need to toss out our 1256 # old set of statistics and start from scratch. 1257 # 1258 # Also, if the delta covers a very long range of time, then we've 1259 # been experiencing a system overload that has resulted in intrd 1260 # not being allowed to run effectively for a while now. As above, 1261 # toss our old statistics and start from scratch. 1262 1263 $deltatime = $delta->{maxsnap} - $delta->{minsnap}; 1264 if ($delta->{missing} > 0 || $deltatime > $statslen) { 1265 clear_deltas(); 1266 syslog('debug', "evaluating interrupt assignments"); 1267 next; 1268 } 1269 1270 1271 # 4. Incorporate new delta into the list of deltas, and associated 1272 # statistics. If we've just now received $statslen deltas, then it's 1273 # time to evaluate a reconfiguration. 1274 1275 $below_statslen = ($deltas_tottime < $statslen); 1276 $deltas_tottime += $deltatime; 1277 $do_reconfig = ($below_statslen && $deltas_tottime >= $statslen); 1278 push(@deltas, $delta); 1279 1280 # 5. Remove old deltas if total time is more than $statslen. We use 1281 # @deltas as a moving average of the last $statslen seconds. Shift 1282 # off the olders deltas, but only if that doesn't cause us to fall 1283 # below $statslen seconds. 1284 1285 while (@deltas > 1) { 1286 $olddelta = $deltas[0]; 1287 $olddeltatime = $olddelta->{maxsnap} - $olddelta->{minsnap}; 1288 $newtime = $deltas_tottime - $olddeltatime; 1289 last if ($newtime < $statslen); 1290 1291 shift(@deltas); 1292 $deltas_tottime = $newtime; 1293 } 1294 1295 # 6. The brains of the operation are here. First, check if we're 1296 # imbalanced, and if so set $do_reconfig. If $do_reconfig is set, 1297 # either because of imbalance or above in step 4, we evaluate a 1298 # new configuration. 1299 # 1300 # First, take @deltas and generate a single "compressed" delta 1301 # which summarizes them all. Pass that to do_reconfig and see 1302 # what it does with it: 1303 # 1304 # $ret == -1 : failure 1305 # $ret == 0 : current config is optimal (or close enough) 1306 # $ret == 1 : reconfiguration has occurred 1307 # 1308 # If $ret is -1 or 1, dump all our deltas and start from scratch. 1309 # Step 4 above will set do_reconfig soon thereafter. 1310 # 1311 # If $ret is 0, then nothing has happened because we're already 1312 # good enough. Set baseline_goodness to current goodness. 1313 1314 $compdelta = compress_deltas(\@deltas); 1315 if (VERIFY(ref($compdelta) eq "HASH", "couldn't compress deltas")) { 1316 clear_deltas(); 1317 next; 1318 } 1319 $compdelta->{goodness} = goodness($compdelta); 1320 dumpdelta($compdelta) if $debug; 1321 1322 $goodness = $compdelta->{goodness}; 1323 syslog('debug', "GOODNESS: %5.2f%%", $goodness * 100); 1324 1325 if ($deltas_tottime >= $statslen && 1326 imbalanced($goodness, $baseline_goodness)) { 1327 $do_reconfig = 1; 1328 } 1329 1330 if ($do_reconfig) { 1331 $ret = do_reconfig($compdelta); 1332 1333 if ($ret != 0) { 1334 clear_deltas(); 1335 syslog('debug', "do_reconfig FAILED!") if $ret == -1; 1336 } else { 1337 syslog('debug', "setting new baseline of $goodness"); 1338 $baseline_goodness = $goodness; 1339 } 1340 } 1341 syslog('debug', "---------------------------------------"); 1342} 1343