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