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