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