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