1#!/usr/bin/env perl 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 warnings; 110use strict; 111 112my $P = $0; 113$P =~ s@.*/@@g; 114 115my $V = '0.1'; 116 117if ($#ARGV != 11) { 118 print "usage: $P arch endian bits objdump objcopy cc ld nm rm mv is_module inputfile\n"; 119 print "version: $V\n"; 120 exit(1); 121} 122 123my ($arch, $endian, $bits, $objdump, $objcopy, $cc, 124 $ld, $nm, $rm, $mv, $is_module, $inputfile) = @ARGV; 125 126# This file refers to mcount and shouldn't be ftraced, so lets' ignore it 127if ($inputfile =~ m,kernel/trace/ftrace\.o$,) { 128 exit(0); 129} 130 131# Acceptable sections to record. 132my %text_sections = ( 133 ".text" => 1, 134 ".init.text" => 1, 135 ".ref.text" => 1, 136 ".sched.text" => 1, 137 ".spinlock.text" => 1, 138 ".irqentry.text" => 1, 139 ".softirqentry.text" => 1, 140 ".kprobes.text" => 1, 141 ".cpuidle.text" => 1, 142 ".text.unlikely" => 1, 143); 144 145# Acceptable section-prefixes to record. 146my %text_section_prefixes = ( 147 ".text." => 1, 148); 149 150# Note: we are nice to C-programmers here, thus we skip the '||='-idiom. 151$objdump = 'objdump' if (!$objdump); 152$objcopy = 'objcopy' if (!$objcopy); 153$cc = 'gcc' if (!$cc); 154$ld = 'ld' if (!$ld); 155$nm = 'nm' if (!$nm); 156$rm = 'rm' if (!$rm); 157$mv = 'mv' if (!$mv); 158 159#print STDERR "running: $P '$arch' '$objdump' '$objcopy' '$cc' '$ld' " . 160# "'$nm' '$rm' '$mv' '$inputfile'\n"; 161 162my %locals; # List of local (static) functions 163my %weak; # List of weak functions 164my %convert; # List of local functions used that needs conversion 165 166my $type; 167my $local_regex; # Match a local function (return function) 168my $weak_regex; # Match a weak function (return function) 169my $section_regex; # Find the start of a section 170my $function_regex; # Find the name of a function 171 # (return offset and func name) 172my $mcount_regex; # Find the call site to mcount (return offset) 173my $mcount_adjust; # Address adjustment to mcount offset 174my $alignment; # The .align value to use for $mcount_section 175my $section_type; # Section header plus possible alignment command 176my $can_use_local = 0; # If we can use local function references 177 178# Shut up recordmcount if user has older objcopy 179my $quiet_recordmcount = ".tmp_quiet_recordmcount"; 180my $print_warning = 1; 181$print_warning = 0 if ( -f $quiet_recordmcount); 182 183## 184# check_objcopy - whether objcopy supports --globalize-symbols 185# 186# --globalize-symbols came out in 2.17, we must test the version 187# of objcopy, and if it is less than 2.17, then we can not 188# record local functions. 189sub check_objcopy 190{ 191 open (IN, "$objcopy --version |") or die "error running $objcopy"; 192 while (<IN>) { 193 if (/objcopy.*\s(\d+)\.(\d+)/) { 194 $can_use_local = 1 if ($1 > 2 || ($1 == 2 && $2 >= 17)); 195 last; 196 } 197 } 198 close (IN); 199 200 if (!$can_use_local && $print_warning) { 201 print STDERR "WARNING: could not find objcopy version or version " . 202 "is less than 2.17.\n" . 203 "\tLocal function references are disabled.\n"; 204 open (QUIET, ">$quiet_recordmcount"); 205 printf QUIET "Disables the warning from recordmcount.pl\n"; 206 close QUIET; 207 } 208} 209 210if ($arch =~ /(x86(_64)?)|(i386)/) { 211 if ($bits == 64) { 212 $arch = "x86_64"; 213 } else { 214 $arch = "i386"; 215 } 216} 217 218# 219# We base the defaults off of i386, the other archs may 220# feel free to change them in the below if statements. 221# 222$local_regex = "^[0-9a-fA-F]+\\s+t\\s+(\\S+)"; 223$weak_regex = "^[0-9a-fA-F]+\\s+([wW])\\s+(\\S+)"; 224$section_regex = "Disassembly of section\\s+(\\S+):"; 225$function_regex = "^([0-9a-fA-F]+)\\s+<(.*?)>:"; 226$mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s(mcount|__fentry__)\$"; 227$section_type = '@progbits'; 228$mcount_adjust = 0; 229$type = ".long"; 230 231if ($arch eq "x86_64") { 232 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s(mcount|__fentry__)([+-]0x[0-9a-zA-Z]+)?\$"; 233 $type = ".quad"; 234 $alignment = 8; 235 $mcount_adjust = -1; 236 237 # force flags for this arch 238 $ld .= " -m elf_x86_64"; 239 $objdump .= " -M x86-64"; 240 $objcopy .= " -O elf64-x86-64"; 241 $cc .= " -m64"; 242 243} elsif ($arch eq "i386") { 244 $alignment = 4; 245 $mcount_adjust = -1; 246 247 # force flags for this arch 248 $ld .= " -m elf_i386"; 249 $objdump .= " -M i386"; 250 $objcopy .= " -O elf32-i386"; 251 $cc .= " -m32"; 252 253} elsif ($arch eq "s390" && $bits == 64) { 254 if ($cc =~ /-DCC_USING_HOTPATCH/) { 255 $mcount_regex = "^\\s*([0-9a-fA-F]+):\\s*c0 04 00 00 00 00\\s*brcl\\s*0,[0-9a-f]+ <([^\+]*)>\$"; 256 $mcount_adjust = 0; 257 } else { 258 $mcount_regex = "^\\s*([0-9a-fA-F]+):\\s*R_390_(PC|PLT)32DBL\\s+_mcount\\+0x2\$"; 259 $mcount_adjust = -14; 260 } 261 $alignment = 8; 262 $type = ".quad"; 263 $ld .= " -m elf64_s390"; 264 $cc .= " -m64"; 265 266} elsif ($arch eq "sh") { 267 $alignment = 2; 268 269 # force flags for this arch 270 $ld .= " -m shlelf_linux"; 271 $objcopy .= " -O elf32-sh-linux"; 272 273} elsif ($arch eq "powerpc") { 274 my $ldemulation; 275 276 $local_regex = "^[0-9a-fA-F]+\\s+t\\s+(\\.?\\S+)"; 277 # See comment in the sparc64 section for why we use '\w'. 278 $function_regex = "^([0-9a-fA-F]+)\\s+<(\\.?\\w*?)>:"; 279 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s\\.?_mcount\$"; 280 281 if ($endian eq "big") { 282 $cc .= " -mbig-endian "; 283 $ld .= " -EB "; 284 $ldemulation = "ppc" 285 } else { 286 $cc .= " -mlittle-endian "; 287 $ld .= " -EL "; 288 $ldemulation = "lppc" 289 } 290 if ($bits == 64) { 291 $type = ".quad"; 292 $cc .= " -m64 "; 293 $ld .= " -m elf64".$ldemulation." "; 294 } else { 295 $cc .= " -m32 "; 296 $ld .= " -m elf32".$ldemulation." "; 297 } 298 299} elsif ($arch eq "arm") { 300 $alignment = 2; 301 $section_type = '%progbits'; 302 $mcount_regex = "^\\s*([0-9a-fA-F]+):\\s*R_ARM_(CALL|PC24|THM_CALL)" . 303 "\\s+(__gnu_mcount_nc|mcount)\$"; 304 305} elsif ($arch eq "arm64") { 306 $alignment = 3; 307 $section_type = '%progbits'; 308 $mcount_regex = "^\\s*([0-9a-fA-F]+):\\s*R_AARCH64_CALL26\\s+_mcount\$"; 309 $type = ".quad"; 310} elsif ($arch eq "ia64") { 311 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s_mcount\$"; 312 $type = "data8"; 313 314 if ($is_module eq "0") { 315 $cc .= " -mconstant-gp"; 316 } 317} elsif ($arch eq "sparc64") { 318 # In the objdump output there are giblets like: 319 # 0000000000000000 <igmp_net_exit-0x18>: 320 # As there's some data blobs that get emitted into the 321 # text section before the first instructions and the first 322 # real symbols. We don't want to match that, so to combat 323 # this we use '\w' so we'll match just plain symbol names, 324 # and not those that also include hex offsets inside of the 325 # '<>' brackets. Actually the generic function_regex setting 326 # could safely use this too. 327 $function_regex = "^([0-9a-fA-F]+)\\s+<(\\w*?)>:"; 328 329 # Sparc64 calls '_mcount' instead of plain 'mcount'. 330 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s_mcount\$"; 331 332 $alignment = 8; 333 $type = ".xword"; 334 $ld .= " -m elf64_sparc"; 335 $cc .= " -m64"; 336 $objcopy .= " -O elf64-sparc"; 337} elsif ($arch eq "mips") { 338 # To enable module support, we need to enable the -mlong-calls option 339 # of gcc for module, after using this option, we can not get the real 340 # offset of the calling to _mcount, but the offset of the lui 341 # instruction or the addiu one. herein, we record the address of the 342 # first one, and then we can replace this instruction by a branch 343 # instruction to jump over the profiling function to filter the 344 # indicated functions, or switch back to the lui instruction to trace 345 # them, which means dynamic tracing. 346 # 347 # c: 3c030000 lui v1,0x0 348 # c: R_MIPS_HI16 _mcount 349 # c: R_MIPS_NONE *ABS* 350 # c: R_MIPS_NONE *ABS* 351 # 10: 64630000 daddiu v1,v1,0 352 # 10: R_MIPS_LO16 _mcount 353 # 10: R_MIPS_NONE *ABS* 354 # 10: R_MIPS_NONE *ABS* 355 # 14: 03e0082d move at,ra 356 # 18: 0060f809 jalr v1 357 # 358 # for the kernel: 359 # 360 # 10: 03e0082d move at,ra 361 # 14: 0c000000 jal 0 <loongson_halt> 362 # 14: R_MIPS_26 _mcount 363 # 14: R_MIPS_NONE *ABS* 364 # 14: R_MIPS_NONE *ABS* 365 # 18: 00020021 nop 366 if ($is_module eq "0") { 367 $mcount_regex = "^\\s*([0-9a-fA-F]+): R_MIPS_26\\s+_mcount\$"; 368 } else { 369 $mcount_regex = "^\\s*([0-9a-fA-F]+): R_MIPS_HI16\\s+_mcount\$"; 370 } 371 $objdump .= " -Melf-trad".$endian."mips "; 372 373 if ($endian eq "big") { 374 $endian = " -EB "; 375 $ld .= " -melf".$bits."btsmip"; 376 } else { 377 $endian = " -EL "; 378 $ld .= " -melf".$bits."ltsmip"; 379 } 380 381 $cc .= " -mno-abicalls -fno-pic -mabi=" . $bits . $endian; 382 $ld .= $endian; 383 384 if ($bits == 64) { 385 $function_regex = 386 "^([0-9a-fA-F]+)\\s+<(.|[^\$]L.*?|\$[^L].*?|[^\$][^L].*?)>:"; 387 $type = ".dword"; 388 } 389} elsif ($arch eq "microblaze") { 390 # Microblaze calls '_mcount' instead of plain 'mcount'. 391 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s_mcount\$"; 392} elsif ($arch eq "riscv") { 393 $function_regex = "^([0-9a-fA-F]+)\\s+<([^.0-9][0-9a-zA-Z_\\.]+)>:"; 394 $mcount_regex = "^\\s*([0-9a-fA-F]+):\\sR_RISCV_CALL\\s_mcount\$"; 395 $type = ".quad"; 396 $alignment = 2; 397} elsif ($arch eq "nds32") { 398 $mcount_regex = "^\\s*([0-9a-fA-F]+):\\s*R_NDS32_HI20_RELA\\s+_mcount\$"; 399 $alignment = 2; 400} elsif ($arch eq "csky") { 401 $mcount_regex = "^\\s*([0-9a-fA-F]+):\\s*R_CKCORE_PCREL_JSR_IMM26BY2\\s+_mcount\$"; 402 $alignment = 2; 403} else { 404 die "Arch $arch is not supported with CONFIG_FTRACE_MCOUNT_RECORD"; 405} 406 407my $text_found = 0; 408my $read_function = 0; 409my $opened = 0; 410my $mcount_section = "__mcount_loc"; 411 412my $dirname; 413my $filename; 414my $prefix; 415my $ext; 416 417if ($inputfile =~ m,^(.*)/([^/]*)$,) { 418 $dirname = $1; 419 $filename = $2; 420} else { 421 $dirname = "."; 422 $filename = $inputfile; 423} 424 425if ($filename =~ m,^(.*)(\.\S),) { 426 $prefix = $1; 427 $ext = $2; 428} else { 429 $prefix = $filename; 430 $ext = ""; 431} 432 433my $mcount_s = $dirname . "/.tmp_mc_" . $prefix . ".s"; 434my $mcount_o = $dirname . "/.tmp_mc_" . $prefix . ".o"; 435 436check_objcopy(); 437 438# 439# Step 1: find all the local (static functions) and weak symbols. 440# 't' is local, 'w/W' is weak 441# 442open (IN, "$nm $inputfile|") || die "error running $nm"; 443while (<IN>) { 444 if (/$local_regex/) { 445 $locals{$1} = 1; 446 } elsif (/$weak_regex/) { 447 $weak{$2} = $1; 448 } 449} 450close(IN); 451 452my @offsets; # Array of offsets of mcount callers 453my $ref_func; # reference function to use for offsets 454my $offset = 0; # offset of ref_func to section beginning 455 456## 457# update_funcs - print out the current mcount callers 458# 459# Go through the list of offsets to callers and write them to 460# the output file in a format that can be read by an assembler. 461# 462sub update_funcs 463{ 464 return unless ($ref_func and @offsets); 465 466 # Sanity check on weak function. A weak function may be overwritten by 467 # another function of the same name, making all these offsets incorrect. 468 if (defined $weak{$ref_func}) { 469 die "$inputfile: ERROR: referencing weak function" . 470 " $ref_func for mcount\n"; 471 } 472 473 # is this function static? If so, note this fact. 474 if (defined $locals{$ref_func}) { 475 476 # only use locals if objcopy supports globalize-symbols 477 if (!$can_use_local) { 478 return; 479 } 480 $convert{$ref_func} = 1; 481 } 482 483 # Loop through all the mcount caller offsets and print a reference 484 # to the caller based from the ref_func. 485 if (!$opened) { 486 open(FILE, ">$mcount_s") || die "can't create $mcount_s\n"; 487 $opened = 1; 488 print FILE "\t.section $mcount_section,\"a\",$section_type\n"; 489 print FILE "\t.align $alignment\n" if (defined($alignment)); 490 } 491 foreach my $cur_offset (@offsets) { 492 printf FILE "\t%s %s + %d\n", $type, $ref_func, $cur_offset - $offset; 493 } 494} 495 496# 497# Step 2: find the sections and mcount call sites 498# 499open(IN, "LANG=C $objdump -hdr $inputfile|") || die "error running $objdump"; 500 501my $text; 502 503 504# read headers first 505my $read_headers = 1; 506 507while (<IN>) { 508 509 if ($read_headers && /$mcount_section/) { 510 # 511 # Somehow the make process can execute this script on an 512 # object twice. If it does, we would duplicate the mcount 513 # section and it will cause the function tracer self test 514 # to fail. Check if the mcount section exists, and if it does, 515 # warn and exit. 516 # 517 print STDERR "ERROR: $mcount_section already in $inputfile\n" . 518 "\tThis may be an indication that your build is corrupted.\n" . 519 "\tDelete $inputfile and try again. If the same object file\n" . 520 "\tstill causes an issue, then disable CONFIG_DYNAMIC_FTRACE.\n"; 521 exit(-1); 522 } 523 524 # is it a section? 525 if (/$section_regex/) { 526 $read_headers = 0; 527 528 # Only record text sections that we know are safe 529 $read_function = defined($text_sections{$1}); 530 if (!$read_function) { 531 foreach my $prefix (keys %text_section_prefixes) { 532 if (substr($1, 0, length $prefix) eq $prefix) { 533 $read_function = 1; 534 last; 535 } 536 } 537 } 538 # print out any recorded offsets 539 update_funcs(); 540 541 # reset all markers and arrays 542 $text_found = 0; 543 undef($ref_func); 544 undef(@offsets); 545 546 # section found, now is this a start of a function? 547 } elsif ($read_function && /$function_regex/) { 548 $text_found = 1; 549 $text = $2; 550 551 # if this is either a local function or a weak function 552 # keep looking for functions that are global that 553 # we can use safely. 554 if (!defined($locals{$text}) && !defined($weak{$text})) { 555 $ref_func = $text; 556 $read_function = 0; 557 $offset = hex $1; 558 } else { 559 # if we already have a function, and this is weak, skip it 560 if (!defined($ref_func) && !defined($weak{$text}) && 561 # PPC64 can have symbols that start with .L and 562 # gcc considers these special. Don't use them! 563 $text !~ /^\.L/) { 564 $ref_func = $text; 565 $offset = hex $1; 566 } 567 } 568 } 569 # is this a call site to mcount? If so, record it to print later 570 if ($text_found && /$mcount_regex/) { 571 push(@offsets, (hex $1) + $mcount_adjust); 572 } 573} 574 575# dump out anymore offsets that may have been found 576update_funcs(); 577 578# If we did not find any mcount callers, we are done (do nothing). 579if (!$opened) { 580 exit(0); 581} 582 583close(FILE); 584 585# 586# Step 3: Compile the file that holds the list of call sites to mcount. 587# 588`$cc -o $mcount_o -c $mcount_s`; 589 590my @converts = keys %convert; 591 592# 593# Step 4: Do we have sections that started with local functions? 594# 595if ($#converts >= 0) { 596 my $globallist = ""; 597 my $locallist = ""; 598 599 foreach my $con (@converts) { 600 $globallist .= " --globalize-symbol $con"; 601 $locallist .= " --localize-symbol $con"; 602 } 603 604 my $globalobj = $dirname . "/.tmp_gl_" . $filename; 605 my $globalmix = $dirname . "/.tmp_mx_" . $filename; 606 607 # 608 # Step 5: set up each local function as a global 609 # 610 `$objcopy $globallist $inputfile $globalobj`; 611 612 # 613 # Step 6: Link the global version to our list. 614 # 615 `$ld -r $globalobj $mcount_o -o $globalmix`; 616 617 # 618 # Step 7: Convert the local functions back into local symbols 619 # 620 `$objcopy $locallist $globalmix $inputfile`; 621 622 # Remove the temp files 623 `$rm $globalobj $globalmix`; 624 625} else { 626 627 my $mix = $dirname . "/.tmp_mx_" . $filename; 628 629 # 630 # Step 8: Link the object with our list of call sites object. 631 # 632 `$ld -r $inputfile $mcount_o -o $mix`; 633 634 # 635 # Step 9: Move the result back to the original object. 636 # 637 `$mv $mix $inputfile`; 638} 639 640# Clean up the temp files 641`$rm $mcount_o $mcount_s`; 642 643exit(0); 644