1#!/usr/bin/env perl 2# SPDX-License-Identifier: GPL-2.0-only 3# (c) 2008, Steven Rostedt <srostedt@redhat.com> 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 } 258 $alignment = 8; 259 $type = ".quad"; 260 $ld .= " -m elf64_s390"; 261 $cc .= " -m64"; 262 263} elsif ($arch eq "sh") { 264 $alignment = 2; 265 266 # force flags for this arch 267 $ld .= " -m shlelf_linux"; 268 if ($endian eq "big") { 269 $objcopy .= " -O elf32-shbig-linux"; 270 } else { 271 $objcopy .= " -O elf32-sh-linux"; 272 } 273 274} elsif ($arch eq "powerpc") { 275 my $ldemulation; 276 277 $local_regex = "^[0-9a-fA-F]+\\s+t\\s+(\\.?\\S+)"; 278 # See comment in the sparc64 section for why we use '\w'. 279 $function_regex = "^([0-9a-fA-F]+)\\s+<(\\.?\\w*?)>:"; 280 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s\\.?_mcount\$"; 281 282 if ($endian eq "big") { 283 $cc .= " -mbig-endian "; 284 $ld .= " -EB "; 285 $ldemulation = "ppc" 286 } else { 287 $cc .= " -mlittle-endian "; 288 $ld .= " -EL "; 289 $ldemulation = "lppc" 290 } 291 if ($bits == 64) { 292 $type = ".quad"; 293 $cc .= " -m64 "; 294 $ld .= " -m elf64".$ldemulation." "; 295 } else { 296 $cc .= " -m32 "; 297 $ld .= " -m elf32".$ldemulation." "; 298 } 299 300} elsif ($arch eq "arm") { 301 $alignment = 2; 302 $section_type = '%progbits'; 303 $mcount_regex = "^\\s*([0-9a-fA-F]+):\\s*R_ARM_(CALL|PC24|THM_CALL)" . 304 "\\s+(__gnu_mcount_nc|mcount)\$"; 305 306} elsif ($arch eq "arm64") { 307 $alignment = 3; 308 $section_type = '%progbits'; 309 $mcount_regex = "^\\s*([0-9a-fA-F]+):\\s*R_AARCH64_CALL26\\s+_mcount\$"; 310 $type = ".quad"; 311} elsif ($arch eq "ia64") { 312 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s_mcount\$"; 313 $type = "data8"; 314 315 if ($is_module eq "0") { 316 $cc .= " -mconstant-gp"; 317 } 318} elsif ($arch eq "sparc64") { 319 # In the objdump output there are giblets like: 320 # 0000000000000000 <igmp_net_exit-0x18>: 321 # As there's some data blobs that get emitted into the 322 # text section before the first instructions and the first 323 # real symbols. We don't want to match that, so to combat 324 # this we use '\w' so we'll match just plain symbol names, 325 # and not those that also include hex offsets inside of the 326 # '<>' brackets. Actually the generic function_regex setting 327 # could safely use this too. 328 $function_regex = "^([0-9a-fA-F]+)\\s+<(\\w*?)>:"; 329 330 # Sparc64 calls '_mcount' instead of plain 'mcount'. 331 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s_mcount\$"; 332 333 $alignment = 8; 334 $type = ".xword"; 335 $ld .= " -m elf64_sparc"; 336 $cc .= " -m64"; 337 $objcopy .= " -O elf64-sparc"; 338} elsif ($arch eq "mips") { 339 # To enable module support, we need to enable the -mlong-calls option 340 # of gcc for module, after using this option, we can not get the real 341 # offset of the calling to _mcount, but the offset of the lui 342 # instruction or the addiu one. herein, we record the address of the 343 # first one, and then we can replace this instruction by a branch 344 # instruction to jump over the profiling function to filter the 345 # indicated functions, or switch back to the lui instruction to trace 346 # them, which means dynamic tracing. 347 # 348 # c: 3c030000 lui v1,0x0 349 # c: R_MIPS_HI16 _mcount 350 # c: R_MIPS_NONE *ABS* 351 # c: R_MIPS_NONE *ABS* 352 # 10: 64630000 daddiu v1,v1,0 353 # 10: R_MIPS_LO16 _mcount 354 # 10: R_MIPS_NONE *ABS* 355 # 10: R_MIPS_NONE *ABS* 356 # 14: 03e0082d move at,ra 357 # 18: 0060f809 jalr v1 358 # 359 # for the kernel: 360 # 361 # 10: 03e0082d move at,ra 362 # 14: 0c000000 jal 0 <loongson_halt> 363 # 14: R_MIPS_26 _mcount 364 # 14: R_MIPS_NONE *ABS* 365 # 14: R_MIPS_NONE *ABS* 366 # 18: 00020021 nop 367 if ($is_module eq "0") { 368 $mcount_regex = "^\\s*([0-9a-fA-F]+): R_MIPS_26\\s+_mcount\$"; 369 } else { 370 $mcount_regex = "^\\s*([0-9a-fA-F]+): R_MIPS_HI16\\s+_mcount\$"; 371 } 372 $objdump .= " -Melf-trad".$endian."mips "; 373 374 if ($endian eq "big") { 375 $endian = " -EB "; 376 $ld .= " -melf".$bits."btsmip"; 377 } else { 378 $endian = " -EL "; 379 $ld .= " -melf".$bits."ltsmip"; 380 } 381 382 $cc .= " -mno-abicalls -fno-pic -mabi=" . $bits . $endian; 383 $ld .= $endian; 384 385 if ($bits == 64) { 386 $function_regex = 387 "^([0-9a-fA-F]+)\\s+<(.|[^\$]L.*?|\$[^L].*?|[^\$][^L].*?)>:"; 388 $type = ".dword"; 389 } 390} elsif ($arch eq "microblaze") { 391 # Microblaze calls '_mcount' instead of plain 'mcount'. 392 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s_mcount\$"; 393} elsif ($arch eq "riscv") { 394 $function_regex = "^([0-9a-fA-F]+)\\s+<([^.0-9][0-9a-zA-Z_\\.]+)>:"; 395 $mcount_regex = "^\\s*([0-9a-fA-F]+):\\sR_RISCV_CALL(_PLT)?\\s_?mcount\$"; 396 $type = ".quad"; 397 $alignment = 2; 398} elsif ($arch eq "nds32") { 399 $mcount_regex = "^\\s*([0-9a-fA-F]+):\\s*R_NDS32_HI20_RELA\\s+_mcount\$"; 400 $alignment = 2; 401} elsif ($arch eq "csky") { 402 $mcount_regex = "^\\s*([0-9a-fA-F]+):\\s*R_CKCORE_PCREL_JSR_IMM26BY2\\s+_mcount\$"; 403 $alignment = 2; 404} else { 405 die "Arch $arch is not supported with CONFIG_FTRACE_MCOUNT_RECORD"; 406} 407 408my $text_found = 0; 409my $read_function = 0; 410my $opened = 0; 411my $mcount_section = "__mcount_loc"; 412 413my $dirname; 414my $filename; 415my $prefix; 416my $ext; 417 418if ($inputfile =~ m,^(.*)/([^/]*)$,) { 419 $dirname = $1; 420 $filename = $2; 421} else { 422 $dirname = "."; 423 $filename = $inputfile; 424} 425 426if ($filename =~ m,^(.*)(\.\S),) { 427 $prefix = $1; 428 $ext = $2; 429} else { 430 $prefix = $filename; 431 $ext = ""; 432} 433 434my $mcount_s = $dirname . "/.tmp_mc_" . $prefix . ".s"; 435my $mcount_o = $dirname . "/.tmp_mc_" . $prefix . ".o"; 436 437check_objcopy(); 438 439# 440# Step 1: find all the local (static functions) and weak symbols. 441# 't' is local, 'w/W' is weak 442# 443open (IN, "$nm $inputfile|") || die "error running $nm"; 444while (<IN>) { 445 if (/$local_regex/) { 446 $locals{$1} = 1; 447 } elsif (/$weak_regex/) { 448 $weak{$2} = $1; 449 } 450} 451close(IN); 452 453my @offsets; # Array of offsets of mcount callers 454my $ref_func; # reference function to use for offsets 455my $offset = 0; # offset of ref_func to section beginning 456 457## 458# update_funcs - print out the current mcount callers 459# 460# Go through the list of offsets to callers and write them to 461# the output file in a format that can be read by an assembler. 462# 463sub update_funcs 464{ 465 return unless ($ref_func and @offsets); 466 467 # Sanity check on weak function. A weak function may be overwritten by 468 # another function of the same name, making all these offsets incorrect. 469 if (defined $weak{$ref_func}) { 470 die "$inputfile: ERROR: referencing weak function" . 471 " $ref_func for mcount\n"; 472 } 473 474 # is this function static? If so, note this fact. 475 if (defined $locals{$ref_func}) { 476 477 # only use locals if objcopy supports globalize-symbols 478 if (!$can_use_local) { 479 return; 480 } 481 $convert{$ref_func} = 1; 482 } 483 484 # Loop through all the mcount caller offsets and print a reference 485 # to the caller based from the ref_func. 486 if (!$opened) { 487 open(FILE, ">$mcount_s") || die "can't create $mcount_s\n"; 488 $opened = 1; 489 print FILE "\t.section $mcount_section,\"a\",$section_type\n"; 490 print FILE "\t.align $alignment\n" if (defined($alignment)); 491 } 492 foreach my $cur_offset (@offsets) { 493 printf FILE "\t%s %s + %d\n", $type, $ref_func, $cur_offset - $offset; 494 } 495} 496 497# 498# Step 2: find the sections and mcount call sites 499# 500open(IN, "LANG=C $objdump -hdr $inputfile|") || die "error running $objdump"; 501 502my $text; 503 504 505# read headers first 506my $read_headers = 1; 507 508while (<IN>) { 509 510 if ($read_headers && /$mcount_section/) { 511 # 512 # Somehow the make process can execute this script on an 513 # object twice. If it does, we would duplicate the mcount 514 # section and it will cause the function tracer self test 515 # to fail. Check if the mcount section exists, and if it does, 516 # warn and exit. 517 # 518 print STDERR "ERROR: $mcount_section already in $inputfile\n" . 519 "\tThis may be an indication that your build is corrupted.\n" . 520 "\tDelete $inputfile and try again. If the same object file\n" . 521 "\tstill causes an issue, then disable CONFIG_DYNAMIC_FTRACE.\n"; 522 exit(-1); 523 } 524 525 # is it a section? 526 if (/$section_regex/) { 527 $read_headers = 0; 528 529 # Only record text sections that we know are safe 530 $read_function = defined($text_sections{$1}); 531 if (!$read_function) { 532 foreach my $prefix (keys %text_section_prefixes) { 533 if (substr($1, 0, length $prefix) eq $prefix) { 534 $read_function = 1; 535 last; 536 } 537 } 538 } 539 # print out any recorded offsets 540 update_funcs(); 541 542 # reset all markers and arrays 543 $text_found = 0; 544 undef($ref_func); 545 undef(@offsets); 546 547 # section found, now is this a start of a function? 548 } elsif ($read_function && /$function_regex/) { 549 $text_found = 1; 550 $text = $2; 551 552 # if this is either a local function or a weak function 553 # keep looking for functions that are global that 554 # we can use safely. 555 if (!defined($locals{$text}) && !defined($weak{$text})) { 556 $ref_func = $text; 557 $read_function = 0; 558 $offset = hex $1; 559 } else { 560 # if we already have a function, and this is weak, skip it 561 if (!defined($ref_func) && !defined($weak{$text}) && 562 # PPC64 can have symbols that start with .L and 563 # gcc considers these special. Don't use them! 564 $text !~ /^\.L/) { 565 $ref_func = $text; 566 $offset = hex $1; 567 } 568 } 569 } 570 # is this a call site to mcount? If so, record it to print later 571 if ($text_found && /$mcount_regex/) { 572 push(@offsets, (hex $1) + $mcount_adjust); 573 } 574} 575 576# dump out anymore offsets that may have been found 577update_funcs(); 578 579# If we did not find any mcount callers, we are done (do nothing). 580if (!$opened) { 581 exit(0); 582} 583 584close(FILE); 585 586# 587# Step 3: Compile the file that holds the list of call sites to mcount. 588# 589`$cc -o $mcount_o -c $mcount_s`; 590 591my @converts = keys %convert; 592 593# 594# Step 4: Do we have sections that started with local functions? 595# 596if ($#converts >= 0) { 597 my $globallist = ""; 598 my $locallist = ""; 599 600 foreach my $con (@converts) { 601 $globallist .= " --globalize-symbol $con"; 602 $locallist .= " --localize-symbol $con"; 603 } 604 605 my $globalobj = $dirname . "/.tmp_gl_" . $filename; 606 my $globalmix = $dirname . "/.tmp_mx_" . $filename; 607 608 # 609 # Step 5: set up each local function as a global 610 # 611 `$objcopy $globallist $inputfile $globalobj`; 612 613 # 614 # Step 6: Link the global version to our list. 615 # 616 `$ld -r $globalobj $mcount_o -o $globalmix`; 617 618 # 619 # Step 7: Convert the local functions back into local symbols 620 # 621 `$objcopy $locallist $globalmix $inputfile`; 622 623 # Remove the temp files 624 `$rm $globalobj $globalmix`; 625 626} else { 627 628 my $mix = $dirname . "/.tmp_mx_" . $filename; 629 630 # 631 # Step 8: Link the object with our list of call sites object. 632 # 633 `$ld -r $inputfile $mcount_o -o $mix`; 634 635 # 636 # Step 9: Move the result back to the original object. 637 # 638 `$mv $mix $inputfile`; 639} 640 641# Clean up the temp files 642`$rm $mcount_o $mcount_s`; 643 644exit(0); 645 646# vim: softtabstop=4 647