1#!/usr/bin/perl -w 2# (c) 2008, Steven Rostedt <srostedt@redhat.com> 3# Licensed under the terms of the GNU GPL License version 2 4# 5# recordmcount.pl - makes a section called __mcount_loc that holds 6# all the offsets to the calls to mcount. 7# 8# 9# What we want to end up with this is that each object file will have a 10# section called __mcount_loc that will hold the list of pointers to mcount 11# callers. After final linking, the vmlinux will have within .init.data the 12# list of all callers to mcount between __start_mcount_loc and __stop_mcount_loc. 13# Later on boot up, the kernel will read this list, save the locations and turn 14# them into nops. When tracing or profiling is later enabled, these locations 15# will then be converted back to pointers to some function. 16# 17# This is no easy feat. This script is called just after the original 18# object is compiled and before it is linked. 19# 20# When parse this object file using 'objdump', the references to the call 21# sites are offsets from the section that the call site is in. Hence, all 22# functions in a section that has a call site to mcount, will have the 23# offset from the beginning of the section and not the beginning of the 24# function. 25# 26# But where this section will reside finally in vmlinx is undetermined at 27# this point. So we can't use this kind of offsets to record the final 28# address of this call site. 29# 30# The trick is to change the call offset referring the start of a section to 31# referring a function symbol in this section. During the link step, 'ld' will 32# compute the final address according to the information we record. 33# 34# e.g. 35# 36# .section ".sched.text", "ax" 37# [...] 38# func1: 39# [...] 40# call mcount (offset: 0x10) 41# [...] 42# ret 43# .globl fun2 44# func2: (offset: 0x20) 45# [...] 46# [...] 47# ret 48# func3: 49# [...] 50# call mcount (offset: 0x30) 51# [...] 52# 53# Both relocation offsets for the mcounts in the above example will be 54# offset from .sched.text. If we choose global symbol func2 as a reference and 55# make another file called tmp.s with the new offsets: 56# 57# .section __mcount_loc 58# .quad func2 - 0x10 59# .quad func2 + 0x10 60# 61# We can then compile this tmp.s into tmp.o, and link it back to the original 62# object. 63# 64# In our algorithm, we will choose the first global function we meet in this 65# section as the reference. But this gets hard if there is no global functions 66# in this section. In such a case we have to select a local one. E.g. func1: 67# 68# .section ".sched.text", "ax" 69# func1: 70# [...] 71# call mcount (offset: 0x10) 72# [...] 73# ret 74# func2: 75# [...] 76# call mcount (offset: 0x20) 77# [...] 78# .section "other.section" 79# 80# If we make the tmp.s the same as above, when we link together with 81# the original object, we will end up with two symbols for func1: 82# one local, one global. After final compile, we will end up with 83# an undefined reference to func1 or a wrong reference to another global 84# func1 in other files. 85# 86# Since local objects can reference local variables, we need to find 87# a way to make tmp.o reference the local objects of the original object 88# file after it is linked together. To do this, we convert func1 89# into a global symbol before linking tmp.o. Then after we link tmp.o 90# we will only have a single symbol for func1 that is global. 91# We can convert func1 back into a local symbol and we are done. 92# 93# Here are the steps we take: 94# 95# 1) Record all the local and weak symbols by using 'nm' 96# 2) Use objdump to find all the call site offsets and sections for 97# mcount. 98# 3) Compile the list into its own object. 99# 4) Do we have to deal with local functions? If not, go to step 8. 100# 5) Make an object that converts these local functions to global symbols 101# with objcopy. 102# 6) Link together this new object with the list object. 103# 7) Convert the local functions back to local symbols and rename 104# the result as the original object. 105# 8) Link the object with the list object. 106# 9) Move the result back to the original object. 107# 108 109use strict; 110 111my $P = $0; 112$P =~ s@.*/@@g; 113 114my $V = '0.1'; 115 116if ($#ARGV != 11) { 117 print "usage: $P arch endian bits objdump objcopy cc ld nm rm mv is_module inputfile\n"; 118 print "version: $V\n"; 119 exit(1); 120} 121 122my ($arch, $endian, $bits, $objdump, $objcopy, $cc, 123 $ld, $nm, $rm, $mv, $is_module, $inputfile) = @ARGV; 124 125# This file refers to mcount and shouldn't be ftraced, so lets' ignore it 126if ($inputfile =~ m,kernel/trace/ftrace\.o$,) { 127 exit(0); 128} 129 130# Acceptable sections to record. 131my %text_sections = ( 132 ".text" => 1, 133 ".ref.text" => 1, 134 ".sched.text" => 1, 135 ".spinlock.text" => 1, 136 ".irqentry.text" => 1, 137 ".softirqentry.text" => 1, 138 ".kprobes.text" => 1, 139 ".cpuidle.text" => 1, 140 ".text.unlikely" => 1, 141); 142 143# Note: we are nice to C-programmers here, thus we skip the '||='-idiom. 144$objdump = 'objdump' if (!$objdump); 145$objcopy = 'objcopy' if (!$objcopy); 146$cc = 'gcc' if (!$cc); 147$ld = 'ld' if (!$ld); 148$nm = 'nm' if (!$nm); 149$rm = 'rm' if (!$rm); 150$mv = 'mv' if (!$mv); 151 152#print STDERR "running: $P '$arch' '$objdump' '$objcopy' '$cc' '$ld' " . 153# "'$nm' '$rm' '$mv' '$inputfile'\n"; 154 155my %locals; # List of local (static) functions 156my %weak; # List of weak functions 157my %convert; # List of local functions used that needs conversion 158 159my $type; 160my $local_regex; # Match a local function (return function) 161my $weak_regex; # Match a weak function (return function) 162my $section_regex; # Find the start of a section 163my $function_regex; # Find the name of a function 164 # (return offset and func name) 165my $mcount_regex; # Find the call site to mcount (return offset) 166my $mcount_adjust; # Address adjustment to mcount offset 167my $alignment; # The .align value to use for $mcount_section 168my $section_type; # Section header plus possible alignment command 169my $can_use_local = 0; # If we can use local function references 170 171# Shut up recordmcount if user has older objcopy 172my $quiet_recordmcount = ".tmp_quiet_recordmcount"; 173my $print_warning = 1; 174$print_warning = 0 if ( -f $quiet_recordmcount); 175 176## 177# check_objcopy - whether objcopy supports --globalize-symbols 178# 179# --globalize-symbols came out in 2.17, we must test the version 180# of objcopy, and if it is less than 2.17, then we can not 181# record local functions. 182sub check_objcopy 183{ 184 open (IN, "$objcopy --version |") or die "error running $objcopy"; 185 while (<IN>) { 186 if (/objcopy.*\s(\d+)\.(\d+)/) { 187 $can_use_local = 1 if ($1 > 2 || ($1 == 2 && $2 >= 17)); 188 last; 189 } 190 } 191 close (IN); 192 193 if (!$can_use_local && $print_warning) { 194 print STDERR "WARNING: could not find objcopy version or version " . 195 "is less than 2.17.\n" . 196 "\tLocal function references are disabled.\n"; 197 open (QUIET, ">$quiet_recordmcount"); 198 printf QUIET "Disables the warning from recordmcount.pl\n"; 199 close QUIET; 200 } 201} 202 203if ($arch =~ /(x86(_64)?)|(i386)/) { 204 if ($bits == 64) { 205 $arch = "x86_64"; 206 } else { 207 $arch = "i386"; 208 } 209} 210 211# 212# We base the defaults off of i386, the other archs may 213# feel free to change them in the below if statements. 214# 215$local_regex = "^[0-9a-fA-F]+\\s+t\\s+(\\S+)"; 216$weak_regex = "^[0-9a-fA-F]+\\s+([wW])\\s+(\\S+)"; 217$section_regex = "Disassembly of section\\s+(\\S+):"; 218$function_regex = "^([0-9a-fA-F]+)\\s+<(.*?)>:"; 219$mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s(mcount|__fentry__)\$"; 220$section_type = '@progbits'; 221$mcount_adjust = 0; 222$type = ".long"; 223 224if ($arch eq "x86_64") { 225 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s(mcount|__fentry__)([+-]0x[0-9a-zA-Z]+)?\$"; 226 $type = ".quad"; 227 $alignment = 8; 228 $mcount_adjust = -1; 229 230 # force flags for this arch 231 $ld .= " -m elf_x86_64"; 232 $objdump .= " -M x86-64"; 233 $objcopy .= " -O elf64-x86-64"; 234 $cc .= " -m64"; 235 236} elsif ($arch eq "i386") { 237 $alignment = 4; 238 $mcount_adjust = -1; 239 240 # force flags for this arch 241 $ld .= " -m elf_i386"; 242 $objdump .= " -M i386"; 243 $objcopy .= " -O elf32-i386"; 244 $cc .= " -m32"; 245 246} elsif ($arch eq "s390" && $bits == 64) { 247 if ($cc =~ /-DCC_USING_HOTPATCH/) { 248 $mcount_regex = "^\\s*([0-9a-fA-F]+):\\s*c0 04 00 00 00 00\\s*brcl\\s*0,[0-9a-f]+ <([^\+]*)>\$"; 249 $mcount_adjust = 0; 250 } else { 251 $mcount_regex = "^\\s*([0-9a-fA-F]+):\\s*R_390_(PC|PLT)32DBL\\s+_mcount\\+0x2\$"; 252 $mcount_adjust = -14; 253 } 254 $alignment = 8; 255 $type = ".quad"; 256 $ld .= " -m elf64_s390"; 257 $cc .= " -m64"; 258 259} elsif ($arch eq "sh") { 260 $alignment = 2; 261 262 # force flags for this arch 263 $ld .= " -m shlelf_linux"; 264 $objcopy .= " -O elf32-sh-linux"; 265 266} elsif ($arch eq "powerpc") { 267 $local_regex = "^[0-9a-fA-F]+\\s+t\\s+(\\.?\\S+)"; 268 # See comment in the sparc64 section for why we use '\w'. 269 $function_regex = "^([0-9a-fA-F]+)\\s+<(\\.?\\w*?)>:"; 270 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s\\.?_mcount\$"; 271 272 if ($bits == 64) { 273 $type = ".quad"; 274 } 275 276} elsif ($arch eq "arm") { 277 $alignment = 2; 278 $section_type = '%progbits'; 279 $mcount_regex = "^\\s*([0-9a-fA-F]+):\\s*R_ARM_(CALL|PC24|THM_CALL)" . 280 "\\s+(__gnu_mcount_nc|mcount)\$"; 281 282} elsif ($arch eq "arm64") { 283 $alignment = 3; 284 $section_type = '%progbits'; 285 $mcount_regex = "^\\s*([0-9a-fA-F]+):\\s*R_AARCH64_CALL26\\s+_mcount\$"; 286 $type = ".quad"; 287} elsif ($arch eq "ia64") { 288 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s_mcount\$"; 289 $type = "data8"; 290 291 if ($is_module eq "0") { 292 $cc .= " -mconstant-gp"; 293 } 294} elsif ($arch eq "sparc64") { 295 # In the objdump output there are giblets like: 296 # 0000000000000000 <igmp_net_exit-0x18>: 297 # As there's some data blobs that get emitted into the 298 # text section before the first instructions and the first 299 # real symbols. We don't want to match that, so to combat 300 # this we use '\w' so we'll match just plain symbol names, 301 # and not those that also include hex offsets inside of the 302 # '<>' brackets. Actually the generic function_regex setting 303 # could safely use this too. 304 $function_regex = "^([0-9a-fA-F]+)\\s+<(\\w*?)>:"; 305 306 # Sparc64 calls '_mcount' instead of plain 'mcount'. 307 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s_mcount\$"; 308 309 $alignment = 8; 310 $type = ".xword"; 311 $ld .= " -m elf64_sparc"; 312 $cc .= " -m64"; 313 $objcopy .= " -O elf64-sparc"; 314} elsif ($arch eq "mips") { 315 # To enable module support, we need to enable the -mlong-calls option 316 # of gcc for module, after using this option, we can not get the real 317 # offset of the calling to _mcount, but the offset of the lui 318 # instruction or the addiu one. herein, we record the address of the 319 # first one, and then we can replace this instruction by a branch 320 # instruction to jump over the profiling function to filter the 321 # indicated functions, or swith back to the lui instruction to trace 322 # them, which means dynamic tracing. 323 # 324 # c: 3c030000 lui v1,0x0 325 # c: R_MIPS_HI16 _mcount 326 # c: R_MIPS_NONE *ABS* 327 # c: R_MIPS_NONE *ABS* 328 # 10: 64630000 daddiu v1,v1,0 329 # 10: R_MIPS_LO16 _mcount 330 # 10: R_MIPS_NONE *ABS* 331 # 10: R_MIPS_NONE *ABS* 332 # 14: 03e0082d move at,ra 333 # 18: 0060f809 jalr v1 334 # 335 # for the kernel: 336 # 337 # 10: 03e0082d move at,ra 338 # 14: 0c000000 jal 0 <loongson_halt> 339 # 14: R_MIPS_26 _mcount 340 # 14: R_MIPS_NONE *ABS* 341 # 14: R_MIPS_NONE *ABS* 342 # 18: 00020021 nop 343 if ($is_module eq "0") { 344 $mcount_regex = "^\\s*([0-9a-fA-F]+): R_MIPS_26\\s+_mcount\$"; 345 } else { 346 $mcount_regex = "^\\s*([0-9a-fA-F]+): R_MIPS_HI16\\s+_mcount\$"; 347 } 348 $objdump .= " -Melf-trad".$endian."mips "; 349 350 if ($endian eq "big") { 351 $endian = " -EB "; 352 $ld .= " -melf".$bits."btsmip"; 353 } else { 354 $endian = " -EL "; 355 $ld .= " -melf".$bits."ltsmip"; 356 } 357 358 $cc .= " -mno-abicalls -fno-pic -mabi=" . $bits . $endian; 359 $ld .= $endian; 360 361 if ($bits == 64) { 362 $function_regex = 363 "^([0-9a-fA-F]+)\\s+<(.|[^\$]L.*?|\$[^L].*?|[^\$][^L].*?)>:"; 364 $type = ".dword"; 365 } 366} elsif ($arch eq "microblaze") { 367 # Microblaze calls '_mcount' instead of plain 'mcount'. 368 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s_mcount\$"; 369} elsif ($arch eq "blackfin") { 370 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s__mcount\$"; 371 $mcount_adjust = -4; 372} elsif ($arch eq "tilegx" || $arch eq "tile") { 373 # Default to the newer TILE-Gx architecture if only "tile" is given. 374 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s__mcount\$"; 375 $type = ".quad"; 376 $alignment = 8; 377} else { 378 die "Arch $arch is not supported with CONFIG_FTRACE_MCOUNT_RECORD"; 379} 380 381my $text_found = 0; 382my $read_function = 0; 383my $opened = 0; 384my $mcount_section = "__mcount_loc"; 385 386my $dirname; 387my $filename; 388my $prefix; 389my $ext; 390 391if ($inputfile =~ m,^(.*)/([^/]*)$,) { 392 $dirname = $1; 393 $filename = $2; 394} else { 395 $dirname = "."; 396 $filename = $inputfile; 397} 398 399if ($filename =~ m,^(.*)(\.\S),) { 400 $prefix = $1; 401 $ext = $2; 402} else { 403 $prefix = $filename; 404 $ext = ""; 405} 406 407my $mcount_s = $dirname . "/.tmp_mc_" . $prefix . ".s"; 408my $mcount_o = $dirname . "/.tmp_mc_" . $prefix . ".o"; 409 410check_objcopy(); 411 412# 413# Step 1: find all the local (static functions) and weak symbols. 414# 't' is local, 'w/W' is weak 415# 416open (IN, "$nm $inputfile|") || die "error running $nm"; 417while (<IN>) { 418 if (/$local_regex/) { 419 $locals{$1} = 1; 420 } elsif (/$weak_regex/) { 421 $weak{$2} = $1; 422 } 423} 424close(IN); 425 426my @offsets; # Array of offsets of mcount callers 427my $ref_func; # reference function to use for offsets 428my $offset = 0; # offset of ref_func to section beginning 429 430## 431# update_funcs - print out the current mcount callers 432# 433# Go through the list of offsets to callers and write them to 434# the output file in a format that can be read by an assembler. 435# 436sub update_funcs 437{ 438 return unless ($ref_func and @offsets); 439 440 # Sanity check on weak function. A weak function may be overwritten by 441 # another function of the same name, making all these offsets incorrect. 442 if (defined $weak{$ref_func}) { 443 die "$inputfile: ERROR: referencing weak function" . 444 " $ref_func for mcount\n"; 445 } 446 447 # is this function static? If so, note this fact. 448 if (defined $locals{$ref_func}) { 449 450 # only use locals if objcopy supports globalize-symbols 451 if (!$can_use_local) { 452 return; 453 } 454 $convert{$ref_func} = 1; 455 } 456 457 # Loop through all the mcount caller offsets and print a reference 458 # to the caller based from the ref_func. 459 if (!$opened) { 460 open(FILE, ">$mcount_s") || die "can't create $mcount_s\n"; 461 $opened = 1; 462 print FILE "\t.section $mcount_section,\"a\",$section_type\n"; 463 print FILE "\t.align $alignment\n" if (defined($alignment)); 464 } 465 foreach my $cur_offset (@offsets) { 466 printf FILE "\t%s %s + %d\n", $type, $ref_func, $cur_offset - $offset; 467 } 468} 469 470# 471# Step 2: find the sections and mcount call sites 472# 473open(IN, "$objdump -hdr $inputfile|") || die "error running $objdump"; 474 475my $text; 476 477 478# read headers first 479my $read_headers = 1; 480 481while (<IN>) { 482 483 if ($read_headers && /$mcount_section/) { 484 # 485 # Somehow the make process can execute this script on an 486 # object twice. If it does, we would duplicate the mcount 487 # section and it will cause the function tracer self test 488 # to fail. Check if the mcount section exists, and if it does, 489 # warn and exit. 490 # 491 print STDERR "ERROR: $mcount_section already in $inputfile\n" . 492 "\tThis may be an indication that your build is corrupted.\n" . 493 "\tDelete $inputfile and try again. If the same object file\n" . 494 "\tstill causes an issue, then disable CONFIG_DYNAMIC_FTRACE.\n"; 495 exit(-1); 496 } 497 498 # is it a section? 499 if (/$section_regex/) { 500 $read_headers = 0; 501 502 # Only record text sections that we know are safe 503 $read_function = defined($text_sections{$1}); 504 # print out any recorded offsets 505 update_funcs(); 506 507 # reset all markers and arrays 508 $text_found = 0; 509 undef($ref_func); 510 undef(@offsets); 511 512 # section found, now is this a start of a function? 513 } elsif ($read_function && /$function_regex/) { 514 $text_found = 1; 515 $text = $2; 516 517 # if this is either a local function or a weak function 518 # keep looking for functions that are global that 519 # we can use safely. 520 if (!defined($locals{$text}) && !defined($weak{$text})) { 521 $ref_func = $text; 522 $read_function = 0; 523 $offset = hex $1; 524 } else { 525 # if we already have a function, and this is weak, skip it 526 if (!defined($ref_func) && !defined($weak{$text}) && 527 # PPC64 can have symbols that start with .L and 528 # gcc considers these special. Don't use them! 529 $text !~ /^\.L/) { 530 $ref_func = $text; 531 $offset = hex $1; 532 } 533 } 534 } 535 # is this a call site to mcount? If so, record it to print later 536 if ($text_found && /$mcount_regex/) { 537 push(@offsets, (hex $1) + $mcount_adjust); 538 } 539} 540 541# dump out anymore offsets that may have been found 542update_funcs(); 543 544# If we did not find any mcount callers, we are done (do nothing). 545if (!$opened) { 546 exit(0); 547} 548 549close(FILE); 550 551# 552# Step 3: Compile the file that holds the list of call sites to mcount. 553# 554`$cc -o $mcount_o -c $mcount_s`; 555 556my @converts = keys %convert; 557 558# 559# Step 4: Do we have sections that started with local functions? 560# 561if ($#converts >= 0) { 562 my $globallist = ""; 563 my $locallist = ""; 564 565 foreach my $con (@converts) { 566 $globallist .= " --globalize-symbol $con"; 567 $locallist .= " --localize-symbol $con"; 568 } 569 570 my $globalobj = $dirname . "/.tmp_gl_" . $filename; 571 my $globalmix = $dirname . "/.tmp_mx_" . $filename; 572 573 # 574 # Step 5: set up each local function as a global 575 # 576 `$objcopy $globallist $inputfile $globalobj`; 577 578 # 579 # Step 6: Link the global version to our list. 580 # 581 `$ld -r $globalobj $mcount_o -o $globalmix`; 582 583 # 584 # Step 7: Convert the local functions back into local symbols 585 # 586 `$objcopy $locallist $globalmix $inputfile`; 587 588 # Remove the temp files 589 `$rm $globalobj $globalmix`; 590 591} else { 592 593 my $mix = $dirname . "/.tmp_mx_" . $filename; 594 595 # 596 # Step 8: Link the object with our list of call sites object. 597 # 598 `$ld -r $inputfile $mcount_o -o $mix`; 599 600 # 601 # Step 9: Move the result back to the original object. 602 # 603 `$mv $mix $inputfile`; 604} 605 606# Clean up the temp files 607`$rm $mcount_o $mcount_s`; 608 609exit(0); 610