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$,) { 127 exit(0); 128} 129 130# Acceptable sections to record. 131my %text_sections = ( 132 ".text" => 1, 133 ".ref.text" => 1, 134 ".sched.text" => 1, 135 ".spinlock.text" => 1, 136 ".irqentry.text" => 1, 137 ".text.unlikely" => 1, 138); 139 140# Note: we are nice to C-programmers here, thus we skip the '||='-idiom. 141$objdump = 'objdump' if (!$objdump); 142$objcopy = 'objcopy' if (!$objcopy); 143$cc = 'gcc' if (!$cc); 144$ld = 'ld' if (!$ld); 145$nm = 'nm' if (!$nm); 146$rm = 'rm' if (!$rm); 147$mv = 'mv' if (!$mv); 148 149#print STDERR "running: $P '$arch' '$objdump' '$objcopy' '$cc' '$ld' " . 150# "'$nm' '$rm' '$mv' '$inputfile'\n"; 151 152my %locals; # List of local (static) functions 153my %weak; # List of weak functions 154my %convert; # List of local functions used that needs conversion 155 156my $type; 157my $local_regex; # Match a local function (return function) 158my $weak_regex; # Match a weak function (return function) 159my $section_regex; # Find the start of a section 160my $function_regex; # Find the name of a function 161 # (return offset and func name) 162my $mcount_regex; # Find the call site to mcount (return offset) 163my $mcount_adjust; # Address adjustment to mcount offset 164my $alignment; # The .align value to use for $mcount_section 165my $section_type; # Section header plus possible alignment command 166my $can_use_local = 0; # If we can use local function references 167 168# Shut up recordmcount if user has older objcopy 169my $quiet_recordmcount = ".tmp_quiet_recordmcount"; 170my $print_warning = 1; 171$print_warning = 0 if ( -f $quiet_recordmcount); 172 173## 174# check_objcopy - whether objcopy supports --globalize-symbols 175# 176# --globalize-symbols came out in 2.17, we must test the version 177# of objcopy, and if it is less than 2.17, then we can not 178# record local functions. 179sub check_objcopy 180{ 181 open (IN, "$objcopy --version |") or die "error running $objcopy"; 182 while (<IN>) { 183 if (/objcopy.*\s(\d+)\.(\d+)/) { 184 $can_use_local = 1 if ($1 > 2 || ($1 == 2 && $2 >= 17)); 185 last; 186 } 187 } 188 close (IN); 189 190 if (!$can_use_local && $print_warning) { 191 print STDERR "WARNING: could not find objcopy version or version " . 192 "is less than 2.17.\n" . 193 "\tLocal function references are disabled.\n"; 194 open (QUIET, ">$quiet_recordmcount"); 195 printf QUIET "Disables the warning from recordmcount.pl\n"; 196 close QUIET; 197 } 198} 199 200if ($arch =~ /(x86(_64)?)|(i386)/) { 201 if ($bits == 64) { 202 $arch = "x86_64"; 203 } else { 204 $arch = "i386"; 205 } 206} 207 208# 209# We base the defaults off of i386, the other archs may 210# feel free to change them in the below if statements. 211# 212$local_regex = "^[0-9a-fA-F]+\\s+t\\s+(\\S+)"; 213$weak_regex = "^[0-9a-fA-F]+\\s+([wW])\\s+(\\S+)"; 214$section_regex = "Disassembly of section\\s+(\\S+):"; 215$function_regex = "^([0-9a-fA-F]+)\\s+<(.*?)>:"; 216$mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\smcount\$"; 217$section_type = '@progbits'; 218$mcount_adjust = 0; 219$type = ".long"; 220 221if ($arch eq "x86_64") { 222 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\smcount([+-]0x[0-9a-zA-Z]+)?\$"; 223 $type = ".quad"; 224 $alignment = 8; 225 226 # force flags for this arch 227 $ld .= " -m elf_x86_64"; 228 $objdump .= " -M x86-64"; 229 $objcopy .= " -O elf64-x86-64"; 230 $cc .= " -m64"; 231 232} elsif ($arch eq "i386") { 233 $alignment = 4; 234 235 # force flags for this arch 236 $ld .= " -m elf_i386"; 237 $objdump .= " -M i386"; 238 $objcopy .= " -O elf32-i386"; 239 $cc .= " -m32"; 240 241} elsif ($arch eq "s390" && $bits == 32) { 242 $mcount_regex = "^\\s*([0-9a-fA-F]+):\\s*R_390_32\\s+_mcount\$"; 243 $alignment = 4; 244 $ld .= " -m elf_s390"; 245 $cc .= " -m31"; 246 247} elsif ($arch eq "s390" && $bits == 64) { 248 $mcount_regex = "^\\s*([0-9a-fA-F]+):\\s*R_390_(PC|PLT)32DBL\\s+_mcount\\+0x2\$"; 249 $alignment = 8; 250 $type = ".quad"; 251 $ld .= " -m elf64_s390"; 252 $cc .= " -m64"; 253 254} elsif ($arch eq "sh") { 255 $alignment = 2; 256 257 # force flags for this arch 258 $ld .= " -m shlelf_linux"; 259 $objcopy .= " -O elf32-sh-linux"; 260 $cc .= " -m32"; 261 262} elsif ($arch eq "powerpc") { 263 $local_regex = "^[0-9a-fA-F]+\\s+t\\s+(\\.?\\S+)"; 264 $function_regex = "^([0-9a-fA-F]+)\\s+<(\\.?.*?)>:"; 265 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s\\.?_mcount\$"; 266 267 if ($bits == 64) { 268 $type = ".quad"; 269 } 270 271} elsif ($arch eq "arm") { 272 $alignment = 2; 273 $section_type = '%progbits'; 274 $mcount_regex = "^\\s*([0-9a-fA-F]+):\\s*R_ARM_(CALL|PC24|THM_CALL)" . 275 "\\s+(__gnu_mcount_nc|mcount)\$"; 276 277} elsif ($arch eq "ia64") { 278 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s_mcount\$"; 279 $type = "data8"; 280 281 if ($is_module eq "0") { 282 $cc .= " -mconstant-gp"; 283 } 284} elsif ($arch eq "sparc64") { 285 # In the objdump output there are giblets like: 286 # 0000000000000000 <igmp_net_exit-0x18>: 287 # As there's some data blobs that get emitted into the 288 # text section before the first instructions and the first 289 # real symbols. We don't want to match that, so to combat 290 # this we use '\w' so we'll match just plain symbol names, 291 # and not those that also include hex offsets inside of the 292 # '<>' brackets. Actually the generic function_regex setting 293 # could safely use this too. 294 $function_regex = "^([0-9a-fA-F]+)\\s+<(\\w*?)>:"; 295 296 # Sparc64 calls '_mcount' instead of plain 'mcount'. 297 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s_mcount\$"; 298 299 $alignment = 8; 300 $type = ".xword"; 301 $ld .= " -m elf64_sparc"; 302 $cc .= " -m64"; 303 $objcopy .= " -O elf64-sparc"; 304} elsif ($arch eq "mips") { 305 # To enable module support, we need to enable the -mlong-calls option 306 # of gcc for module, after using this option, we can not get the real 307 # offset of the calling to _mcount, but the offset of the lui 308 # instruction or the addiu one. herein, we record the address of the 309 # first one, and then we can replace this instruction by a branch 310 # instruction to jump over the profiling function to filter the 311 # indicated functions, or swith back to the lui instruction to trace 312 # them, which means dynamic tracing. 313 # 314 # c: 3c030000 lui v1,0x0 315 # c: R_MIPS_HI16 _mcount 316 # c: R_MIPS_NONE *ABS* 317 # c: R_MIPS_NONE *ABS* 318 # 10: 64630000 daddiu v1,v1,0 319 # 10: R_MIPS_LO16 _mcount 320 # 10: R_MIPS_NONE *ABS* 321 # 10: R_MIPS_NONE *ABS* 322 # 14: 03e0082d move at,ra 323 # 18: 0060f809 jalr v1 324 # 325 # for the kernel: 326 # 327 # 10: 03e0082d move at,ra 328 # 14: 0c000000 jal 0 <loongson_halt> 329 # 14: R_MIPS_26 _mcount 330 # 14: R_MIPS_NONE *ABS* 331 # 14: R_MIPS_NONE *ABS* 332 # 18: 00020021 nop 333 if ($is_module eq "0") { 334 $mcount_regex = "^\\s*([0-9a-fA-F]+): R_MIPS_26\\s+_mcount\$"; 335 } else { 336 $mcount_regex = "^\\s*([0-9a-fA-F]+): R_MIPS_HI16\\s+_mcount\$"; 337 } 338 $objdump .= " -Melf-trad".$endian."mips "; 339 340 if ($endian eq "big") { 341 $endian = " -EB "; 342 $ld .= " -melf".$bits."btsmip"; 343 } else { 344 $endian = " -EL "; 345 $ld .= " -melf".$bits."ltsmip"; 346 } 347 348 $cc .= " -mno-abicalls -fno-pic -mabi=" . $bits . $endian; 349 $ld .= $endian; 350 351 if ($bits == 64) { 352 $function_regex = 353 "^([0-9a-fA-F]+)\\s+<(.|[^\$]L.*?|\$[^L].*?|[^\$][^L].*?)>:"; 354 $type = ".dword"; 355 } 356} elsif ($arch eq "microblaze") { 357 # Microblaze calls '_mcount' instead of plain 'mcount'. 358 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s_mcount\$"; 359} elsif ($arch eq "blackfin") { 360 $mcount_regex = "^\\s*([0-9a-fA-F]+):.*\\s__mcount\$"; 361 $mcount_adjust = -4; 362} else { 363 die "Arch $arch is not supported with CONFIG_FTRACE_MCOUNT_RECORD"; 364} 365 366my $text_found = 0; 367my $read_function = 0; 368my $opened = 0; 369my $mcount_section = "__mcount_loc"; 370 371my $dirname; 372my $filename; 373my $prefix; 374my $ext; 375 376if ($inputfile =~ m,^(.*)/([^/]*)$,) { 377 $dirname = $1; 378 $filename = $2; 379} else { 380 $dirname = "."; 381 $filename = $inputfile; 382} 383 384if ($filename =~ m,^(.*)(\.\S),) { 385 $prefix = $1; 386 $ext = $2; 387} else { 388 $prefix = $filename; 389 $ext = ""; 390} 391 392my $mcount_s = $dirname . "/.tmp_mc_" . $prefix . ".s"; 393my $mcount_o = $dirname . "/.tmp_mc_" . $prefix . ".o"; 394 395check_objcopy(); 396 397# 398# Step 1: find all the local (static functions) and weak symbols. 399# 't' is local, 'w/W' is weak 400# 401open (IN, "$nm $inputfile|") || die "error running $nm"; 402while (<IN>) { 403 if (/$local_regex/) { 404 $locals{$1} = 1; 405 } elsif (/$weak_regex/) { 406 $weak{$2} = $1; 407 } 408} 409close(IN); 410 411my @offsets; # Array of offsets of mcount callers 412my $ref_func; # reference function to use for offsets 413my $offset = 0; # offset of ref_func to section beginning 414 415## 416# update_funcs - print out the current mcount callers 417# 418# Go through the list of offsets to callers and write them to 419# the output file in a format that can be read by an assembler. 420# 421sub update_funcs 422{ 423 return unless ($ref_func and @offsets); 424 425 # Sanity check on weak function. A weak function may be overwritten by 426 # another function of the same name, making all these offsets incorrect. 427 if (defined $weak{$ref_func}) { 428 die "$inputfile: ERROR: referencing weak function" . 429 " $ref_func for mcount\n"; 430 } 431 432 # is this function static? If so, note this fact. 433 if (defined $locals{$ref_func}) { 434 435 # only use locals if objcopy supports globalize-symbols 436 if (!$can_use_local) { 437 return; 438 } 439 $convert{$ref_func} = 1; 440 } 441 442 # Loop through all the mcount caller offsets and print a reference 443 # to the caller based from the ref_func. 444 if (!$opened) { 445 open(FILE, ">$mcount_s") || die "can't create $mcount_s\n"; 446 $opened = 1; 447 print FILE "\t.section $mcount_section,\"a\",$section_type\n"; 448 print FILE "\t.align $alignment\n" if (defined($alignment)); 449 } 450 foreach my $cur_offset (@offsets) { 451 printf FILE "\t%s %s + %d\n", $type, $ref_func, $cur_offset - $offset; 452 } 453} 454 455# 456# Step 2: find the sections and mcount call sites 457# 458open(IN, "$objdump -hdr $inputfile|") || die "error running $objdump"; 459 460my $text; 461 462 463# read headers first 464my $read_headers = 1; 465 466while (<IN>) { 467 468 if ($read_headers && /$mcount_section/) { 469 # 470 # Somehow the make process can execute this script on an 471 # object twice. If it does, we would duplicate the mcount 472 # section and it will cause the function tracer self test 473 # to fail. Check if the mcount section exists, and if it does, 474 # warn and exit. 475 # 476 print STDERR "ERROR: $mcount_section already in $inputfile\n" . 477 "\tThis may be an indication that your build is corrupted.\n" . 478 "\tDelete $inputfile and try again. If the same object file\n" . 479 "\tstill causes an issue, then disable CONFIG_DYNAMIC_FTRACE.\n"; 480 exit(-1); 481 } 482 483 # is it a section? 484 if (/$section_regex/) { 485 $read_headers = 0; 486 487 # Only record text sections that we know are safe 488 $read_function = defined($text_sections{$1}); 489 # print out any recorded offsets 490 update_funcs(); 491 492 # reset all markers and arrays 493 $text_found = 0; 494 undef($ref_func); 495 undef(@offsets); 496 497 # section found, now is this a start of a function? 498 } elsif ($read_function && /$function_regex/) { 499 $text_found = 1; 500 $text = $2; 501 502 # if this is either a local function or a weak function 503 # keep looking for functions that are global that 504 # we can use safely. 505 if (!defined($locals{$text}) && !defined($weak{$text})) { 506 $ref_func = $text; 507 $read_function = 0; 508 $offset = hex $1; 509 } else { 510 # if we already have a function, and this is weak, skip it 511 if (!defined($ref_func) && !defined($weak{$text}) && 512 # PPC64 can have symbols that start with .L and 513 # gcc considers these special. Don't use them! 514 $text !~ /^\.L/) { 515 $ref_func = $text; 516 $offset = hex $1; 517 } 518 } 519 } 520 # is this a call site to mcount? If so, record it to print later 521 if ($text_found && /$mcount_regex/) { 522 push(@offsets, (hex $1) + $mcount_adjust); 523 } 524} 525 526# dump out anymore offsets that may have been found 527update_funcs(); 528 529# If we did not find any mcount callers, we are done (do nothing). 530if (!$opened) { 531 exit(0); 532} 533 534close(FILE); 535 536# 537# Step 3: Compile the file that holds the list of call sites to mcount. 538# 539`$cc -o $mcount_o -c $mcount_s`; 540 541my @converts = keys %convert; 542 543# 544# Step 4: Do we have sections that started with local functions? 545# 546if ($#converts >= 0) { 547 my $globallist = ""; 548 my $locallist = ""; 549 550 foreach my $con (@converts) { 551 $globallist .= " --globalize-symbol $con"; 552 $locallist .= " --localize-symbol $con"; 553 } 554 555 my $globalobj = $dirname . "/.tmp_gl_" . $filename; 556 my $globalmix = $dirname . "/.tmp_mx_" . $filename; 557 558 # 559 # Step 5: set up each local function as a global 560 # 561 `$objcopy $globallist $inputfile $globalobj`; 562 563 # 564 # Step 6: Link the global version to our list. 565 # 566 `$ld -r $globalobj $mcount_o -o $globalmix`; 567 568 # 569 # Step 7: Convert the local functions back into local symbols 570 # 571 `$objcopy $locallist $globalmix $inputfile`; 572 573 # Remove the temp files 574 `$rm $globalobj $globalmix`; 575 576} else { 577 578 my $mix = $dirname . "/.tmp_mx_" . $filename; 579 580 # 581 # Step 8: Link the object with our list of call sites object. 582 # 583 `$ld -r $inputfile $mcount_o -o $mix`; 584 585 # 586 # Step 9: Move the result back to the original object. 587 # 588 `$mv $mix $inputfile`; 589} 590 591# Clean up the temp files 592`$rm $mcount_o $mcount_s`; 593 594exit(0); 595