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