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