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