1<?xml version="1.0"?> <!-- -*- sgml -*- --> 2<!DOCTYPE chapter PUBLIC "-//OASIS//DTD DocBook XML V4.2//EN" 3 "http://www.oasis-open.org/docbook/xml/4.2/docbookx.dtd" 4[ <!ENTITY % vg-entities SYSTEM "../../docs/xml/vg-entities.xml"> %vg-entities; ]> 5 6 7<chapter id="drd-manual" xreflabel="DRD: a thread error detector"> 8 <title>DRD: a thread error detector</title> 9 10<para>To use this tool, you must specify 11<option>--tool=drd</option> 12on the Valgrind command line.</para> 13 14 15<sect1 id="drd-manual.overview" xreflabel="Overview"> 16<title>Overview</title> 17 18<para> 19DRD is a Valgrind tool for detecting errors in multithreaded C and C++ 20programs. The tool works for any program that uses the POSIX threading 21primitives or that uses threading concepts built on top of the POSIX threading 22primitives. 23</para> 24 25<sect2 id="drd-manual.mt-progr-models" xreflabel="MT-progr-models"> 26<title>Multithreaded Programming Paradigms</title> 27 28<para> 29There are two possible reasons for using multithreading in a program: 30<itemizedlist> 31 <listitem> 32 <para> 33 To model concurrent activities. Assigning one thread to each activity 34 can be a great simplification compared to multiplexing the states of 35 multiple activities in a single thread. This is why most server software 36 and embedded software is multithreaded. 37 </para> 38 </listitem> 39 <listitem> 40 <para> 41 To use multiple CPU cores simultaneously for speeding up 42 computations. This is why many High Performance Computing (HPC) 43 applications are multithreaded. 44 </para> 45 </listitem> 46</itemizedlist> 47</para> 48 49<para> 50Multithreaded programs can use one or more of the following programming 51paradigms. Which paradigm is appropriate depends e.g. on the application type. 52Some examples of multithreaded programming paradigms are: 53<itemizedlist> 54 <listitem> 55 <para> 56 Locking. Data that is shared over threads is protected from concurrent 57 accesses via locking. E.g. the POSIX threads library, the Qt library 58 and the Boost.Thread library support this paradigm directly. 59 </para> 60 </listitem> 61 <listitem> 62 <para> 63 Message passing. No data is shared between threads, but threads exchange 64 data by passing messages to each other. Examples of implementations of 65 the message passing paradigm are MPI and CORBA. 66 </para> 67 </listitem> 68 <listitem> 69 <para> 70 Automatic parallelization. A compiler converts a sequential program into 71 a multithreaded program. The original program may or may not contain 72 parallelization hints. One example of such parallelization hints is the 73 OpenMP standard. In this standard a set of directives are defined which 74 tell a compiler how to parallelize a C, C++ or Fortran program. OpenMP 75 is well suited for computational intensive applications. As an example, 76 an open source image processing software package is using OpenMP to 77 maximize performance on systems with multiple CPU 78 cores. GCC supports the 79 OpenMP standard from version 4.2.0 on. 80 </para> 81 </listitem> 82 <listitem> 83 <para> 84 Software Transactional Memory (STM). Any data that is shared between 85 threads is updated via transactions. After each transaction it is 86 verified whether there were any conflicting transactions. If there were 87 conflicts, the transaction is aborted, otherwise it is committed. This 88 is a so-called optimistic approach. There is a prototype of the Intel C++ 89 Compiler available that supports STM. Research about the addition of 90 STM support to GCC is ongoing. 91 </para> 92 </listitem> 93</itemizedlist> 94</para> 95 96<para> 97DRD supports any combination of multithreaded programming paradigms as 98long as the implementation of these paradigms is based on the POSIX 99threads primitives. DRD however does not support programs that use 100e.g. Linux' futexes directly. Attempts to analyze such programs with 101DRD will cause DRD to report many false positives. 102</para> 103 104</sect2> 105 106 107<sect2 id="drd-manual.pthreads-model" xreflabel="Pthreads-model"> 108<title>POSIX Threads Programming Model</title> 109 110<para> 111POSIX threads, also known as Pthreads, is the most widely available 112threading library on Unix systems. 113</para> 114 115<para> 116The POSIX threads programming model is based on the following abstractions: 117<itemizedlist> 118 <listitem> 119 <para> 120 A shared address space. All threads running within the same 121 process share the same address space. All data, whether shared or 122 not, is identified by its address. 123 </para> 124 </listitem> 125 <listitem> 126 <para> 127 Regular load and store operations, which allow to read values 128 from or to write values to the memory shared by all threads 129 running in the same process. 130 </para> 131 </listitem> 132 <listitem> 133 <para> 134 Atomic store and load-modify-store operations. While these are 135 not mentioned in the POSIX threads standard, most 136 microprocessors support atomic memory operations. 137 </para> 138 </listitem> 139 <listitem> 140 <para> 141 Threads. Each thread represents a concurrent activity. 142 </para> 143 </listitem> 144 <listitem> 145 <para> 146 Synchronization objects and operations on these synchronization 147 objects. The following types of synchronization objects have been 148 defined in the POSIX threads standard: mutexes, condition variables, 149 semaphores, reader-writer synchronization objects, barriers and 150 spinlocks. 151 </para> 152 </listitem> 153</itemizedlist> 154</para> 155 156<para> 157Which source code statements generate which memory accesses depends on 158the <emphasis>memory model</emphasis> of the programming language being 159used. There is not yet a definitive memory model for the C and C++ 160languages. For a draft memory model, see also the document 161<ulink url="http://www.open-std.org/jtc1/sc22/wg21/docs/papers/2007/n2338.html"> 162WG21/N2338: Concurrency memory model compiler consequences</ulink>. 163</para> 164 165<para> 166For more information about POSIX threads, see also the Single UNIX 167Specification version 3, also known as 168<ulink url="http://www.opengroup.org/onlinepubs/000095399/idx/threads.html"> 169IEEE Std 1003.1</ulink>. 170</para> 171 172</sect2> 173 174 175<sect2 id="drd-manual.mt-problems" xreflabel="MT-Problems"> 176<title>Multithreaded Programming Problems</title> 177 178<para> 179Depending on which multithreading paradigm is being used in a program, 180one or more of the following problems can occur: 181<itemizedlist> 182 <listitem> 183 <para> 184 Data races. One or more threads access the same memory location without 185 sufficient locking. Most but not all data races are programming errors 186 and are the cause of subtle and hard-to-find bugs. 187 </para> 188 </listitem> 189 <listitem> 190 <para> 191 Lock contention. One thread blocks the progress of one or more other 192 threads by holding a lock too long. 193 </para> 194 </listitem> 195 <listitem> 196 <para> 197 Improper use of the POSIX threads API. Most implementations of the POSIX 198 threads API have been optimized for runtime speed. Such implementations 199 will not complain on certain errors, e.g. when a mutex is being unlocked 200 by another thread than the thread that obtained a lock on the mutex. 201 </para> 202 </listitem> 203 <listitem> 204 <para> 205 Deadlock. A deadlock occurs when two or more threads wait for 206 each other indefinitely. 207 </para> 208 </listitem> 209 <listitem> 210 <para> 211 False sharing. If threads that run on different processor cores 212 access different variables located in the same cache line 213 frequently, this will slow down the involved threads a lot due 214 to frequent exchange of cache lines. 215 </para> 216 </listitem> 217</itemizedlist> 218</para> 219 220<para> 221Although the likelihood of the occurrence of data races can be reduced 222through a disciplined programming style, a tool for automatic 223detection of data races is a necessity when developing multithreaded 224software. DRD can detect these, as well as lock contention and 225improper use of the POSIX threads API. 226</para> 227 228</sect2> 229 230 231<sect2 id="drd-manual.data-race-detection" xreflabel="data-race-detection"> 232<title>Data Race Detection</title> 233 234<para> 235The result of load and store operations performed by a multithreaded program 236depends on the order in which memory operations are performed. This order is 237determined by: 238<orderedlist> 239 <listitem> 240 <para> 241 All memory operations performed by the same thread are performed in 242 <emphasis>program order</emphasis>, that is, the order determined by the 243 program source code and the results of previous load operations. 244 </para> 245 </listitem> 246 <listitem> 247 <para> 248 Synchronization operations determine certain ordering constraints on 249 memory operations performed by different threads. These ordering 250 constraints are called the <emphasis>synchronization order</emphasis>. 251 </para> 252 </listitem> 253</orderedlist> 254The combination of program order and synchronization order is called the 255<emphasis>happens-before relationship</emphasis>. This concept was first 256defined by S. Adve et al in the paper <emphasis>Detecting data races on weak 257memory systems</emphasis>, ACM SIGARCH Computer Architecture News, v.19 n.3, 258p.234-243, May 1991. 259</para> 260 261<para> 262Two memory operations <emphasis>conflict</emphasis> if both operations are 263performed by different threads, refer to the same memory location and at least 264one of them is a store operation. 265</para> 266 267<para> 268A multithreaded program is <emphasis>data-race free</emphasis> if all 269conflicting memory accesses are ordered by synchronization 270operations. 271</para> 272 273<para> 274A well known way to ensure that a multithreaded program is data-race 275free is to ensure that a locking discipline is followed. It is e.g. 276possible to associate a mutex with each shared data item, and to hold 277a lock on the associated mutex while the shared data is accessed. 278</para> 279 280<para> 281All programs that follow a locking discipline are data-race free, but not all 282data-race free programs follow a locking discipline. There exist multithreaded 283programs where access to shared data is arbitrated via condition variables, 284semaphores or barriers. As an example, a certain class of HPC applications 285consists of a sequence of computation steps separated in time by barriers, and 286where these barriers are the only means of synchronization. Although there are 287many conflicting memory accesses in such applications and although such 288applications do not make use mutexes, most of these applications do not 289contain data races. 290</para> 291 292<para> 293There exist two different approaches for verifying the correctness of 294multithreaded programs at runtime. The approach of the so-called Eraser 295algorithm is to verify whether all shared memory accesses follow a consistent 296locking strategy. And the happens-before data race detectors verify directly 297whether all interthread memory accesses are ordered by synchronization 298operations. While the last approach is more complex to implement, and while it 299is more sensitive to OS scheduling, it is a general approach that works for 300all classes of multithreaded programs. An important advantage of 301happens-before data race detectors is that these do not report any false 302positives. 303</para> 304 305<para> 306DRD is based on the happens-before algorithm. 307</para> 308 309</sect2> 310 311 312</sect1> 313 314 315<sect1 id="drd-manual.using-drd" xreflabel="Using DRD"> 316<title>Using DRD</title> 317 318<sect2 id="drd-manual.options" xreflabel="DRD Command-line Options"> 319<title>DRD Command-line Options</title> 320 321<para>The following command-line options are available for controlling the 322behavior of the DRD tool itself:</para> 323 324<!-- start of xi:include in the manpage --> 325<variablelist id="drd.opts.list"> 326 <varlistentry> 327 <term> 328 <option><![CDATA[--check-stack-var=<yes|no> [default: no]]]></option> 329 </term> 330 <listitem> 331 <para> 332 Controls whether DRD detects data races on stack 333 variables. Verifying stack variables is disabled by default because 334 most programs do not share stack variables over threads. 335 </para> 336 </listitem> 337 </varlistentry> 338 <varlistentry> 339 <term> 340 <option><![CDATA[--exclusive-threshold=<n> [default: off]]]></option> 341 </term> 342 <listitem> 343 <para> 344 Print an error message if any mutex or writer lock has been 345 held longer than the time specified in milliseconds. This 346 option enables the detection of lock contention. 347 </para> 348 </listitem> 349 </varlistentry> 350 <varlistentry> 351 <term> 352 <option><![CDATA[--join-list-vol=<n> [default: 10]]]></option> 353 </term> 354 <listitem> 355 <para> 356 Data races that occur between a statement at the end of one thread 357 and another thread can be missed if memory access information is 358 discarded immediately after a thread has been joined. This option 359 allows to specify for how many joined threads memory access information 360 should be retained. 361 </para> 362 </listitem> 363 </varlistentry> 364 <varlistentry> 365 <term> 366 <option> 367 <![CDATA[--first-race-only=<yes|no> [default: no]]]> 368 </option> 369 </term> 370 <listitem> 371 <para> 372 Whether to report only the first data race that has been detected on a 373 memory location or all data races that have been detected on a memory 374 location. 375 </para> 376 </listitem> 377 </varlistentry> 378 <varlistentry> 379 <term> 380 <option> 381 <![CDATA[--free-is-write=<yes|no> [default: no]]]> 382 </option> 383 </term> 384 <listitem> 385 <para> 386 Whether to report races between accessing memory and freeing 387 memory. Enabling this option may cause DRD to run slightly 388 slower. Notes:</para> 389 <itemizedlist> 390 <listitem> 391 <para> 392 Don't enable this option when using custom memory allocators 393 that use 394 the <computeroutput>VG_USERREQ__MALLOCLIKE_BLOCK</computeroutput> 395 and <computeroutput>VG_USERREQ__FREELIKE_BLOCK</computeroutput> 396 because that would result in false positives. 397 </para> 398 </listitem> 399 <listitem> 400 <para>Don't enable this option when using reference-counted 401 objects because that will result in false positives, even when 402 that code has been annotated properly with 403 <computeroutput>ANNOTATE_HAPPENS_BEFORE</computeroutput> 404 and <computeroutput>ANNOTATE_HAPPENS_AFTER</computeroutput>. See 405 e.g. the output of the following command for an example: 406 <computeroutput>valgrind --tool=drd --free-is-write=yes 407 drd/tests/annotate_smart_pointer</computeroutput>. 408 </para> 409 </listitem> 410 </itemizedlist> 411 </listitem> 412 </varlistentry> 413 <varlistentry> 414 <term> 415 <option> 416 <![CDATA[--report-signal-unlocked=<yes|no> [default: yes]]]> 417 </option> 418 </term> 419 <listitem> 420 <para> 421 Whether to report calls to 422 <function>pthread_cond_signal</function> and 423 <function>pthread_cond_broadcast</function> where the mutex 424 associated with the signal through 425 <function>pthread_cond_wait</function> or 426 <function>pthread_cond_timed_wait</function>is not locked at 427 the time the signal is sent. Sending a signal without holding 428 a lock on the associated mutex is a common programming error 429 which can cause subtle race conditions and unpredictable 430 behavior. There exist some uncommon synchronization patterns 431 however where it is safe to send a signal without holding a 432 lock on the associated mutex. 433 </para> 434 </listitem> 435 </varlistentry> 436 <varlistentry> 437 <term> 438 <option><![CDATA[--segment-merging=<yes|no> [default: yes]]]></option> 439 </term> 440 <listitem> 441 <para> 442 Controls segment merging. Segment merging is an algorithm to 443 limit memory usage of the data race detection 444 algorithm. Disabling segment merging may improve the accuracy 445 of the so-called 'other segments' displayed in race reports 446 but can also trigger an out of memory error. 447 </para> 448 </listitem> 449 </varlistentry> 450 <varlistentry> 451 <term> 452 <option><![CDATA[--segment-merging-interval=<n> [default: 10]]]></option> 453 </term> 454 <listitem> 455 <para> 456 Perform segment merging only after the specified number of new 457 segments have been created. This is an advanced configuration option 458 that allows to choose whether to minimize DRD's memory usage by 459 choosing a low value or to let DRD run faster by choosing a slightly 460 higher value. The optimal value for this parameter depends on the 461 program being analyzed. The default value works well for most programs. 462 </para> 463 </listitem> 464 </varlistentry> 465 <varlistentry> 466 <term> 467 <option><![CDATA[--shared-threshold=<n> [default: off]]]></option> 468 </term> 469 <listitem> 470 <para> 471 Print an error message if a reader lock has been held longer 472 than the specified time (in milliseconds). This option enables 473 the detection of lock contention. 474 </para> 475 </listitem> 476 </varlistentry> 477 <varlistentry> 478 <term> 479 <option><![CDATA[--show-confl-seg=<yes|no> [default: yes]]]></option> 480 </term> 481 <listitem> 482 <para> 483 Show conflicting segments in race reports. Since this 484 information can help to find the cause of a data race, this 485 option is enabled by default. Disabling this option makes the 486 output of DRD more compact. 487 </para> 488 </listitem> 489 </varlistentry> 490 <varlistentry> 491 <term> 492 <option><![CDATA[--show-stack-usage=<yes|no> [default: no]]]></option> 493 </term> 494 <listitem> 495 <para> 496 Print stack usage at thread exit time. When a program creates a large 497 number of threads it becomes important to limit the amount of virtual 498 memory allocated for thread stacks. This option makes it possible to 499 observe how much stack memory has been used by each thread of the 500 client program. Note: the DRD tool itself allocates some temporary 501 data on the client thread stack. The space necessary for this 502 temporary data must be allocated by the client program when it 503 allocates stack memory, but is not included in stack usage reported by 504 DRD. 505 </para> 506 </listitem> 507 </varlistentry> 508 <varlistentry> 509 <term> 510 <option><![CDATA[--ignore-thread-creation=<yes|no> [default: no]]]></option> 511 </term> 512 <listitem> 513 <para> 514 Controls whether all activities during thread creation should be 515 ignored. By default enabled only on Solaris. 516 Solaris provides higher throughput, parallelism and scalability than 517 other operating systems, at the cost of more fine-grained locking 518 activity. This means for example that when a thread is created under 519 glibc, just one big lock is used for all thread setup. Solaris libc 520 uses several fine-grained locks and the creator thread resumes its 521 activities as soon as possible, leaving for example stack and TLS setup 522 sequence to the created thread. 523 This situation confuses DRD as it assumes there is some false ordering 524 in place between creator and created thread; and therefore many types 525 of race conditions in the application would not be reported. To prevent 526 such false ordering, this command line option is set to 527 <computeroutput>yes</computeroutput> by default on Solaris. 528 All activity (loads, stores, client requests) is therefore ignored 529 during:</para> 530 <itemizedlist> 531 <listitem> 532 <para> 533 pthread_create() call in the creator thread 534 </para> 535 </listitem> 536 <listitem> 537 <para> 538 thread creation phase (stack and TLS setup) in the created thread 539 </para> 540 </listitem> 541 </itemizedlist> 542 </listitem> 543 </varlistentry> 544</variablelist> 545<!-- end of xi:include in the manpage --> 546 547<!-- start of xi:include in the manpage --> 548<para> 549The following options are available for monitoring the behavior of the 550client program: 551</para> 552 553<variablelist id="drd.debugopts.list"> 554 <varlistentry> 555 <term> 556 <option><![CDATA[--trace-addr=<address> [default: none]]]></option> 557 </term> 558 <listitem> 559 <para> 560 Trace all load and store activity for the specified 561 address. This option may be specified more than once. 562 </para> 563 </listitem> 564 </varlistentry> 565 <varlistentry> 566 <term> 567 <option><![CDATA[--ptrace-addr=<address> [default: none]]]></option> 568 </term> 569 <listitem> 570 <para> 571 Trace all load and store activity for the specified address and keep 572 doing that even after the memory at that address has been freed and 573 reallocated. 574 </para> 575 </listitem> 576 </varlistentry> 577 <varlistentry> 578 <term> 579 <option><![CDATA[--trace-alloc=<yes|no> [default: no]]]></option> 580 </term> 581 <listitem> 582 <para> 583 Trace all memory allocations and deallocations. May produce a huge 584 amount of output. 585 </para> 586 </listitem> 587 </varlistentry> 588 <varlistentry> 589 <term> 590 <option><![CDATA[--trace-barrier=<yes|no> [default: no]]]></option> 591 </term> 592 <listitem> 593 <para> 594 Trace all barrier activity. 595 </para> 596 </listitem> 597 </varlistentry> 598 <varlistentry> 599 <term> 600 <option><![CDATA[--trace-cond=<yes|no> [default: no]]]></option> 601 </term> 602 <listitem> 603 <para> 604 Trace all condition variable activity. 605 </para> 606 </listitem> 607 </varlistentry> 608 <varlistentry> 609 <term> 610 <option><![CDATA[--trace-fork-join=<yes|no> [default: no]]]></option> 611 </term> 612 <listitem> 613 <para> 614 Trace all thread creation and all thread termination events. 615 </para> 616 </listitem> 617 </varlistentry> 618 <varlistentry> 619 <term> 620 <option><![CDATA[--trace-hb=<yes|no> [default: no]]]></option> 621 </term> 622 <listitem> 623 <para> 624 Trace execution of the <literal>ANNOTATE_HAPPENS_BEFORE()</literal>, 625 <literal>ANNOTATE_HAPPENS_AFTER()</literal> and 626 <literal>ANNOTATE_HAPPENS_DONE()</literal> client requests. 627 </para> 628 </listitem> 629 </varlistentry> 630 <varlistentry> 631 <term> 632 <option><![CDATA[--trace-mutex=<yes|no> [default: no]]]></option> 633 </term> 634 <listitem> 635 <para> 636 Trace all mutex activity. 637 </para> 638 </listitem> 639 </varlistentry> 640 <varlistentry> 641 <term> 642 <option><![CDATA[--trace-rwlock=<yes|no> [default: no]]]></option> 643 </term> 644 <listitem> 645 <para> 646 Trace all reader-writer lock activity. 647 </para> 648 </listitem> 649 </varlistentry> 650 <varlistentry> 651 <term> 652 <option><![CDATA[--trace-semaphore=<yes|no> [default: no]]]></option> 653 </term> 654 <listitem> 655 <para> 656 Trace all semaphore activity. 657 </para> 658 </listitem> 659 </varlistentry> 660</variablelist> 661<!-- end of xi:include in the manpage --> 662 663</sect2> 664 665 666<sect2 id="drd-manual.data-races" xreflabel="Data Races"> 667<title>Detected Errors: Data Races</title> 668 669<para> 670DRD prints a message every time it detects a data race. Please keep 671the following in mind when interpreting DRD's output: 672<itemizedlist> 673 <listitem> 674 <para> 675 Every thread is assigned a <emphasis>thread ID</emphasis> by the DRD 676 tool. A thread ID is a number. Thread ID's start at one and are never 677 recycled. 678 </para> 679 </listitem> 680 <listitem> 681 <para> 682 The term <emphasis>segment</emphasis> refers to a consecutive 683 sequence of load, store and synchronization operations, all 684 issued by the same thread. A segment always starts and ends at a 685 synchronization operation. Data race analysis is performed 686 between segments instead of between individual load and store 687 operations because of performance reasons. 688 </para> 689 </listitem> 690 <listitem> 691 <para> 692 There are always at least two memory accesses involved in a data 693 race. Memory accesses involved in a data race are called 694 <emphasis>conflicting memory accesses</emphasis>. DRD prints a 695 report for each memory access that conflicts with a past memory 696 access. 697 </para> 698 </listitem> 699</itemizedlist> 700</para> 701 702<para> 703Below you can find an example of a message printed by DRD when it 704detects a data race: 705</para> 706<programlisting><![CDATA[ 707$ valgrind --tool=drd --read-var-info=yes drd/tests/rwlock_race 708... 709==9466== Thread 3: 710==9466== Conflicting load by thread 3 at 0x006020b8 size 4 711==9466== at 0x400B6C: thread_func (rwlock_race.c:29) 712==9466== by 0x4C291DF: vg_thread_wrapper (drd_pthread_intercepts.c:186) 713==9466== by 0x4E3403F: start_thread (in /lib64/libpthread-2.8.so) 714==9466== by 0x53250CC: clone (in /lib64/libc-2.8.so) 715==9466== Location 0x6020b8 is 0 bytes inside local var "s_racy" 716==9466== declared at rwlock_race.c:18, in frame #0 of thread 3 717==9466== Other segment start (thread 2) 718==9466== at 0x4C2847D: pthread_rwlock_rdlock* (drd_pthread_intercepts.c:813) 719==9466== by 0x400B6B: thread_func (rwlock_race.c:28) 720==9466== by 0x4C291DF: vg_thread_wrapper (drd_pthread_intercepts.c:186) 721==9466== by 0x4E3403F: start_thread (in /lib64/libpthread-2.8.so) 722==9466== by 0x53250CC: clone (in /lib64/libc-2.8.so) 723==9466== Other segment end (thread 2) 724==9466== at 0x4C28B54: pthread_rwlock_unlock* (drd_pthread_intercepts.c:912) 725==9466== by 0x400B84: thread_func (rwlock_race.c:30) 726==9466== by 0x4C291DF: vg_thread_wrapper (drd_pthread_intercepts.c:186) 727==9466== by 0x4E3403F: start_thread (in /lib64/libpthread-2.8.so) 728==9466== by 0x53250CC: clone (in /lib64/libc-2.8.so) 729... 730]]></programlisting> 731 732<para> 733The above report has the following meaning: 734<itemizedlist> 735 <listitem> 736 <para> 737 The number in the column on the left is the process ID of the 738 process being analyzed by DRD. 739 </para> 740 </listitem> 741 <listitem> 742 <para> 743 The first line ("Thread 3") tells you the thread ID for 744 the thread in which context the data race has been detected. 745 </para> 746 </listitem> 747 <listitem> 748 <para> 749 The next line tells which kind of operation was performed (load or 750 store) and by which thread. On the same line the start address and the 751 number of bytes involved in the conflicting access are also displayed. 752 </para> 753 </listitem> 754 <listitem> 755 <para> 756 Next, the call stack of the conflicting access is displayed. If 757 your program has been compiled with debug information 758 (<option>-g</option>), this call stack will include file names and 759 line numbers. The two 760 bottommost frames in this call stack (<function>clone</function> 761 and <function>start_thread</function>) show how the NPTL starts 762 a thread. The third frame 763 (<function>vg_thread_wrapper</function>) is added by DRD. The 764 fourth frame (<function>thread_func</function>) is the first 765 interesting line because it shows the thread entry point, that 766 is the function that has been passed as the third argument to 767 <function>pthread_create</function>. 768 </para> 769 </listitem> 770 <listitem> 771 <para> 772 Next, the allocation context for the conflicting address is 773 displayed. For dynamically allocated data the allocation call 774 stack is shown. For static variables and stack variables the 775 allocation context is only shown when the option 776 <option>--read-var-info=yes</option> has been 777 specified. Otherwise DRD will print <computeroutput>Allocation 778 context: unknown</computeroutput>. 779 </para> 780 </listitem> 781 <listitem> 782 <para> 783 A conflicting access involves at least two memory accesses. For 784 one of these accesses an exact call stack is displayed, and for 785 the other accesses an approximate call stack is displayed, 786 namely the start and the end of the segments of the other 787 accesses. This information can be interpreted as follows: 788 <orderedlist> 789 <listitem> 790 <para> 791 Start at the bottom of both call stacks, and count the 792 number stack frames with identical function name, file 793 name and line number. In the above example the three 794 bottommost frames are identical 795 (<function>clone</function>, 796 <function>start_thread</function> and 797 <function>vg_thread_wrapper</function>). 798 </para> 799 </listitem> 800 <listitem> 801 <para> 802 The next higher stack frame in both call stacks now tells 803 you between in which source code region the other memory 804 access happened. The above output tells that the other 805 memory access involved in the data race happened between 806 source code lines 28 and 30 in file 807 <computeroutput>rwlock_race.c</computeroutput>. 808 </para> 809 </listitem> 810 </orderedlist> 811 </para> 812 </listitem> 813</itemizedlist> 814</para> 815 816</sect2> 817 818 819<sect2 id="drd-manual.lock-contention" xreflabel="Lock Contention"> 820<title>Detected Errors: Lock Contention</title> 821 822<para> 823Threads must be able to make progress without being blocked for too long by 824other threads. Sometimes a thread has to wait until a mutex or reader-writer 825synchronization object is unlocked by another thread. This is called 826<emphasis>lock contention</emphasis>. 827</para> 828 829<para> 830Lock contention causes delays. Such delays should be as short as 831possible. The two command line options 832<literal>--exclusive-threshold=<n></literal> and 833<literal>--shared-threshold=<n></literal> make it possible to 834detect excessive lock contention by making DRD report any lock that 835has been held longer than the specified threshold. An example: 836</para> 837<programlisting><![CDATA[ 838$ valgrind --tool=drd --exclusive-threshold=10 drd/tests/hold_lock -i 500 839... 840==10668== Acquired at: 841==10668== at 0x4C267C8: pthread_mutex_lock (drd_pthread_intercepts.c:395) 842==10668== by 0x400D92: main (hold_lock.c:51) 843==10668== Lock on mutex 0x7fefffd50 was held during 503 ms (threshold: 10 ms). 844==10668== at 0x4C26ADA: pthread_mutex_unlock (drd_pthread_intercepts.c:441) 845==10668== by 0x400DB5: main (hold_lock.c:55) 846... 847]]></programlisting> 848 849<para> 850The <literal>hold_lock</literal> test program holds a lock as long as 851specified by the <literal>-i</literal> (interval) argument. The DRD 852output reports that the lock acquired at line 51 in source file 853<literal>hold_lock.c</literal> and released at line 55 was held during 854503 ms, while a threshold of 10 ms was specified to DRD. 855</para> 856 857</sect2> 858 859 860<sect2 id="drd-manual.api-checks" xreflabel="API Checks"> 861<title>Detected Errors: Misuse of the POSIX threads API</title> 862 863<para> 864 DRD is able to detect and report the following misuses of the POSIX 865 threads API: 866 <itemizedlist> 867 <listitem> 868 <para> 869 Passing the address of one type of synchronization object 870 (e.g. a mutex) to a POSIX API call that expects a pointer to 871 another type of synchronization object (e.g. a condition 872 variable). 873 </para> 874 </listitem> 875 <listitem> 876 <para> 877 Attempts to unlock a mutex that has not been locked. 878 </para> 879 </listitem> 880 <listitem> 881 <para> 882 Attempts to unlock a mutex that was locked by another thread. 883 </para> 884 </listitem> 885 <listitem> 886 <para> 887 Attempts to lock a mutex of type 888 <literal>PTHREAD_MUTEX_NORMAL</literal> or a spinlock 889 recursively. 890 </para> 891 </listitem> 892 <listitem> 893 <para> 894 Destruction or deallocation of a locked mutex. 895 </para> 896 </listitem> 897 <listitem> 898 <para> 899 Sending a signal to a condition variable while no lock is held 900 on the mutex associated with the condition variable. 901 </para> 902 </listitem> 903 <listitem> 904 <para> 905 Calling <function>pthread_cond_wait</function> on a mutex 906 that is not locked, that is locked by another thread or that 907 has been locked recursively. 908 </para> 909 </listitem> 910 <listitem> 911 <para> 912 Associating two different mutexes with a condition variable 913 through <function>pthread_cond_wait</function>. 914 </para> 915 </listitem> 916 <listitem> 917 <para> 918 Destruction or deallocation of a condition variable that is 919 being waited upon. 920 </para> 921 </listitem> 922 <listitem> 923 <para> 924 Destruction or deallocation of a locked reader-writer synchronization 925 object. 926 </para> 927 </listitem> 928 <listitem> 929 <para> 930 Attempts to unlock a reader-writer synchronization object that was not 931 locked by the calling thread. 932 </para> 933 </listitem> 934 <listitem> 935 <para> 936 Attempts to recursively lock a reader-writer synchronization object 937 exclusively. 938 </para> 939 </listitem> 940 <listitem> 941 <para> 942 Attempts to pass the address of a user-defined reader-writer 943 synchronization object to a POSIX threads function. 944 </para> 945 </listitem> 946 <listitem> 947 <para> 948 Attempts to pass the address of a POSIX reader-writer synchronization 949 object to one of the annotations for user-defined reader-writer 950 synchronization objects. 951 </para> 952 </listitem> 953 <listitem> 954 <para> 955 Reinitialization of a mutex, condition variable, reader-writer 956 lock, semaphore or barrier. 957 </para> 958 </listitem> 959 <listitem> 960 <para> 961 Destruction or deallocation of a semaphore or barrier that is 962 being waited upon. 963 </para> 964 </listitem> 965 <listitem> 966 <para> 967 Missing synchronization between barrier wait and barrier destruction. 968 </para> 969 </listitem> 970 <listitem> 971 <para> 972 Exiting a thread without first unlocking the spinlocks, mutexes or 973 reader-writer synchronization objects that were locked by that thread. 974 </para> 975 </listitem> 976 <listitem> 977 <para> 978 Passing an invalid thread ID to <function>pthread_join</function> 979 or <function>pthread_cancel</function>. 980 </para> 981 </listitem> 982 </itemizedlist> 983</para> 984 985</sect2> 986 987 988<sect2 id="drd-manual.clientreqs" xreflabel="Client requests"> 989<title>Client Requests</title> 990 991<para> 992Just as for other Valgrind tools it is possible to let a client program 993interact with the DRD tool through client requests. In addition to the 994client requests several macros have been defined that allow to use the 995client requests in a convenient way. 996</para> 997 998<para> 999The interface between client programs and the DRD tool is defined in 1000the header file <literal><valgrind/drd.h></literal>. The 1001available macros and client requests are: 1002<itemizedlist> 1003 <listitem> 1004 <para> 1005 The macro <literal>DRD_GET_VALGRIND_THREADID</literal> and the 1006 corresponding client 1007 request <varname>VG_USERREQ__DRD_GET_VALGRIND_THREAD_ID</varname>. 1008 Query the thread ID that has been assigned by the Valgrind core to the 1009 thread executing this client request. Valgrind's thread ID's start at 1010 one and are recycled in case a thread stops. 1011 </para> 1012 </listitem> 1013 <listitem> 1014 <para> 1015 The macro <literal>DRD_GET_DRD_THREADID</literal> and the corresponding 1016 client request <varname>VG_USERREQ__DRD_GET_DRD_THREAD_ID</varname>. 1017 Query the thread ID that has been assigned by DRD to the thread 1018 executing this client request. These are the thread ID's reported by DRD 1019 in data race reports and in trace messages. DRD's thread ID's start at 1020 one and are never recycled. 1021 </para> 1022 </listitem> 1023 <listitem> 1024 <para> 1025 The macros <literal>DRD_IGNORE_VAR(x)</literal>, 1026 <literal>ANNOTATE_TRACE_MEMORY(&x)</literal> and the corresponding 1027 client request <varname>VG_USERREQ__DRD_START_SUPPRESSION</varname>. Some 1028 applications contain intentional races. There exist e.g. applications 1029 where the same value is assigned to a shared variable from two different 1030 threads. It may be more convenient to suppress such races than to solve 1031 these. This client request allows to suppress such races. 1032 </para> 1033 </listitem> 1034 <listitem> 1035 <para> 1036 The macro <literal>DRD_STOP_IGNORING_VAR(x)</literal> and the 1037 corresponding client request 1038 <varname>VG_USERREQ__DRD_FINISH_SUPPRESSION</varname>. Tell DRD 1039 to no longer ignore data races for the address range that was suppressed 1040 either via the macro <literal>DRD_IGNORE_VAR(x)</literal> or via the 1041 client request <varname>VG_USERREQ__DRD_START_SUPPRESSION</varname>. 1042 </para> 1043 </listitem> 1044 <listitem> 1045 <para> 1046 The macro <literal>DRD_TRACE_VAR(x)</literal>. Trace all load and store 1047 activity for the address range starting at <literal>&x</literal> and 1048 occupying <literal>sizeof(x)</literal> bytes. When DRD reports a data 1049 race on a specified variable, and it's not immediately clear which 1050 source code statements triggered the conflicting accesses, it can be 1051 very helpful to trace all activity on the offending memory location. 1052 </para> 1053 </listitem> 1054 <listitem> 1055 <para> 1056 The macro <literal>DRD_STOP_TRACING_VAR(x)</literal>. Stop tracing load 1057 and store activity for the address range starting 1058 at <literal>&x</literal> and occupying <literal>sizeof(x)</literal> 1059 bytes. 1060 </para> 1061 </listitem> 1062 <listitem> 1063 <para> 1064 The macro <literal>ANNOTATE_TRACE_MEMORY(&x)</literal>. Trace all 1065 load and store activity that touches at least the single byte at the 1066 address <literal>&x</literal>. 1067 </para> 1068 </listitem> 1069 <listitem> 1070 <para> 1071 The client request <varname>VG_USERREQ__DRD_START_TRACE_ADDR</varname>, 1072 which allows to trace all load and store activity for the specified 1073 address range. 1074 </para> 1075 </listitem> 1076 <listitem> 1077 <para> 1078 The client 1079 request <varname>VG_USERREQ__DRD_STOP_TRACE_ADDR</varname>. Do no longer 1080 trace load and store activity for the specified address range. 1081 </para> 1082 </listitem> 1083 <listitem> 1084 <para> 1085 The macro <literal>ANNOTATE_HAPPENS_BEFORE(addr)</literal> tells DRD to 1086 insert a mark. Insert this macro just after an access to the variable at 1087 the specified address has been performed. 1088 </para> 1089 </listitem> 1090 <listitem> 1091 <para> 1092 The macro <literal>ANNOTATE_HAPPENS_AFTER(addr)</literal> tells DRD that 1093 the next access to the variable at the specified address should be 1094 considered to have happened after the access just before the latest 1095 <literal>ANNOTATE_HAPPENS_BEFORE(addr)</literal> annotation that 1096 references the same variable. The purpose of these two macros is to tell 1097 DRD about the order of inter-thread memory accesses implemented via 1098 atomic memory operations. See 1099 also <literal>drd/tests/annotate_smart_pointer.cpp</literal> for an 1100 example. 1101 </para> 1102 </listitem> 1103 <listitem> 1104 <para> 1105 The macro <literal>ANNOTATE_RWLOCK_CREATE(rwlock)</literal> tells DRD 1106 that the object at address <literal>rwlock</literal> is a 1107 reader-writer synchronization object that is not a 1108 <literal>pthread_rwlock_t</literal> synchronization object. See 1109 also <literal>drd/tests/annotate_rwlock.c</literal> for an example. 1110 </para> 1111 </listitem> 1112 <listitem> 1113 <para> 1114 The macro <literal>ANNOTATE_RWLOCK_DESTROY(rwlock)</literal> tells DRD 1115 that the reader-writer synchronization object at 1116 address <literal>rwlock</literal> has been destroyed. 1117 </para> 1118 </listitem> 1119 <listitem> 1120 <para> 1121 The macro <literal>ANNOTATE_WRITERLOCK_ACQUIRED(rwlock)</literal> tells 1122 DRD that a writer lock has been acquired on the reader-writer 1123 synchronization object at address <literal>rwlock</literal>. 1124 </para> 1125 </listitem> 1126 <listitem> 1127 <para> 1128 The macro <literal>ANNOTATE_READERLOCK_ACQUIRED(rwlock)</literal> tells 1129 DRD that a reader lock has been acquired on the reader-writer 1130 synchronization object at address <literal>rwlock</literal>. 1131 </para> 1132 </listitem> 1133 <listitem> 1134 <para> 1135 The macro <literal>ANNOTATE_RWLOCK_ACQUIRED(rwlock, is_w)</literal> 1136 tells DRD that a writer lock (when <literal>is_w != 0</literal>) or that 1137 a reader lock (when <literal>is_w == 0</literal>) has been acquired on 1138 the reader-writer synchronization object at 1139 address <literal>rwlock</literal>. 1140 </para> 1141 </listitem> 1142 <listitem> 1143 <para> 1144 The macro <literal>ANNOTATE_WRITERLOCK_RELEASED(rwlock)</literal> tells 1145 DRD that a writer lock has been released on the reader-writer 1146 synchronization object at address <literal>rwlock</literal>. 1147 </para> 1148 </listitem> 1149 <listitem> 1150 <para> 1151 The macro <literal>ANNOTATE_READERLOCK_RELEASED(rwlock)</literal> tells 1152 DRD that a reader lock has been released on the reader-writer 1153 synchronization object at address <literal>rwlock</literal>. 1154 </para> 1155 </listitem> 1156 <listitem> 1157 <para> 1158 The macro <literal>ANNOTATE_RWLOCK_RELEASED(rwlock, is_w)</literal> 1159 tells DRD that a writer lock (when <literal>is_w != 0</literal>) or that 1160 a reader lock (when <literal>is_w == 0</literal>) has been released on 1161 the reader-writer synchronization object at 1162 address <literal>rwlock</literal>. 1163 </para> 1164 </listitem> 1165 <listitem> 1166 <para> 1167 The macro <literal>ANNOTATE_BARRIER_INIT(barrier, count, 1168 reinitialization_allowed)</literal> tells DRD that a new barrier object 1169 at the address <literal>barrier</literal> has been initialized, 1170 that <literal>count</literal> threads participate in each barrier and 1171 also whether or not barrier reinitialization without intervening 1172 destruction should be reported as an error. See 1173 also <literal>drd/tests/annotate_barrier.c</literal> for an example. 1174 </para> 1175 </listitem> 1176 <listitem> 1177 <para> 1178 The macro <literal>ANNOTATE_BARRIER_DESTROY(barrier)</literal> 1179 tells DRD that a barrier object is about to be destroyed. 1180 </para> 1181 </listitem> 1182 <listitem> 1183 <para> 1184 The macro <literal>ANNOTATE_BARRIER_WAIT_BEFORE(barrier)</literal> 1185 tells DRD that waiting for a barrier will start. 1186 </para> 1187 </listitem> 1188 <listitem> 1189 <para> 1190 The macro <literal>ANNOTATE_BARRIER_WAIT_AFTER(barrier)</literal> 1191 tells DRD that waiting for a barrier has finished. 1192 </para> 1193 </listitem> 1194 <listitem> 1195 <para> 1196 The macro <literal>ANNOTATE_BENIGN_RACE_SIZED(addr, size, 1197 descr)</literal> tells DRD that any races detected on the specified 1198 address are benign and hence should not be 1199 reported. The <literal>descr</literal> argument is ignored but can be 1200 used to document why data races on <literal>addr</literal> are benign. 1201 </para> 1202 </listitem> 1203 <listitem> 1204 <para> 1205 The macro <literal>ANNOTATE_BENIGN_RACE_STATIC(var, descr)</literal> 1206 tells DRD that any races detected on the specified static variable are 1207 benign and hence should not be reported. The <literal>descr</literal> 1208 argument is ignored but can be used to document why data races 1209 on <literal>var</literal> are benign. Note: this macro can only be 1210 used in C++ programs and not in C programs. 1211 </para> 1212 </listitem> 1213 <listitem> 1214 <para> 1215 The macro <literal>ANNOTATE_IGNORE_READS_BEGIN</literal> tells 1216 DRD to ignore all memory loads performed by the current thread. 1217 </para> 1218 </listitem> 1219 <listitem> 1220 <para> 1221 The macro <literal>ANNOTATE_IGNORE_READS_END</literal> tells 1222 DRD to stop ignoring the memory loads performed by the current thread. 1223 </para> 1224 </listitem> 1225 <listitem> 1226 <para> 1227 The macro <literal>ANNOTATE_IGNORE_WRITES_BEGIN</literal> tells 1228 DRD to ignore all memory stores performed by the current thread. 1229 </para> 1230 </listitem> 1231 <listitem> 1232 <para> 1233 The macro <literal>ANNOTATE_IGNORE_WRITES_END</literal> tells 1234 DRD to stop ignoring the memory stores performed by the current thread. 1235 </para> 1236 </listitem> 1237 <listitem> 1238 <para> 1239 The macro <literal>ANNOTATE_IGNORE_READS_AND_WRITES_BEGIN</literal> tells 1240 DRD to ignore all memory accesses performed by the current thread. 1241 </para> 1242 </listitem> 1243 <listitem> 1244 <para> 1245 The macro <literal>ANNOTATE_IGNORE_READS_AND_WRITES_END</literal> tells 1246 DRD to stop ignoring the memory accesses performed by the current thread. 1247 </para> 1248 </listitem> 1249 <listitem> 1250 <para> 1251 The macro <literal>ANNOTATE_NEW_MEMORY(addr, size)</literal> tells 1252 DRD that the specified memory range has been allocated by a custom 1253 memory allocator in the client program and that the client program 1254 will start using this memory range. 1255 </para> 1256 </listitem> 1257 <listitem> 1258 <para> 1259 The macro <literal>ANNOTATE_THREAD_NAME(name)</literal> tells DRD to 1260 associate the specified name with the current thread and to include this 1261 name in the error messages printed by DRD. 1262 </para> 1263 </listitem> 1264 <listitem> 1265 <para> 1266 The macros <literal>VALGRIND_MALLOCLIKE_BLOCK</literal> and 1267 <literal>VALGRIND_FREELIKE_BLOCK</literal> from the Valgrind core are 1268 implemented; they are described in 1269 <xref linkend="manual-core-adv.clientreq"/>. 1270 </para> 1271 </listitem> 1272</itemizedlist> 1273</para> 1274 1275<para> 1276Note: if you compiled Valgrind yourself, the header file 1277<literal><valgrind/drd.h></literal> will have been installed in 1278the directory <literal>/usr/include</literal> by the command 1279<literal>make install</literal>. If you obtained Valgrind by 1280installing it as a package however, you will probably have to install 1281another package with a name like <literal>valgrind-devel</literal> 1282before Valgrind's header files are available. 1283</para> 1284 1285</sect2> 1286 1287 1288<sect2 id="drd-manual.C++11" xreflabel="C++11"> 1289<title>Debugging C++11 Programs</title> 1290 1291<para>If you want to use the C++11 class std::thread you will need to do the 1292 following to annotate the std::shared_ptr<> objects used in the 1293 implementation of that class: 1294<itemizedlist> 1295 <listitem> 1296 <para>Add the following code at the start of a common header or at the 1297 start of each source file, before any C++ header files are included:</para> 1298 <programlisting> 1299#include <valgrind/drd.h> 1300#define _GLIBCXX_SYNCHRONIZATION_HAPPENS_BEFORE(addr) ANNOTATE_HAPPENS_BEFORE(addr) 1301#define _GLIBCXX_SYNCHRONIZATION_HAPPENS_AFTER(addr) ANNOTATE_HAPPENS_AFTER(addr) 1302</programlisting> 1303 </listitem> 1304 <listitem> 1305 <para>Download the gcc source code and from source file 1306 libstdc++-v3/src/c++11/thread.cc copy the implementation of the 1307 <computeroutput>execute_native_thread_routine()</computeroutput> 1308 and <computeroutput>std::thread::_M_start_thread()</computeroutput> 1309 functions into a source file that is linked with your application. Make 1310 sure that also in this source file the 1311 _GLIBCXX_SYNCHRONIZATION_HAPPENS_*() macros are defined properly.</para> 1312 </listitem> 1313</itemizedlist> 1314</para> 1315<para>For more information, see also <emphasis>The 1316GNU C++ Library Manual, Debugging Support</emphasis> 1317(<ulink url="http://gcc.gnu.org/onlinedocs/libstdc++/manual/debug.html">http://gcc.gnu.org/onlinedocs/libstdc++/manual/debug.html</ulink>).</para> 1318 1319</sect2> 1320 1321 1322<sect2 id="drd-manual.gnome" xreflabel="GNOME"> 1323<title>Debugging GNOME Programs</title> 1324 1325<para> 1326GNOME applications use the threading primitives provided by the 1327<computeroutput>glib</computeroutput> and 1328<computeroutput>gthread</computeroutput> libraries. These libraries 1329are built on top of POSIX threads, and hence are directly supported by 1330DRD. Please keep in mind that you have to call 1331<function>g_thread_init</function> before creating any threads, or 1332DRD will report several data races on glib functions. See also the 1333<ulink 1334url="http://library.gnome.org/devel/glib/stable/glib-Threads.html">GLib 1335Reference Manual</ulink> for more information about 1336<function>g_thread_init</function>. 1337</para> 1338 1339<para> 1340One of the many facilities provided by the <literal>glib</literal> 1341library is a block allocator, called <literal>g_slice</literal>. You 1342have to disable this block allocator when using DRD by adding the 1343following to the shell environment variables: 1344<literal>G_SLICE=always-malloc</literal>. See also the <ulink 1345url="http://library.gnome.org/devel/glib/stable/glib-Memory-Slices.html">GLib 1346Reference Manual</ulink> for more information. 1347</para> 1348 1349</sect2> 1350 1351 1352<sect2 id="drd-manual.boost.thread" xreflabel="Boost.Thread"> 1353<title>Debugging Boost.Thread Programs</title> 1354 1355<para> 1356The Boost.Thread library is the threading library included with the 1357cross-platform Boost Libraries. This threading library is an early 1358implementation of the upcoming C++0x threading library. 1359</para> 1360 1361<para> 1362Applications that use the Boost.Thread library should run fine under DRD. 1363</para> 1364 1365<para> 1366More information about Boost.Thread can be found here: 1367<itemizedlist> 1368 <listitem> 1369 <para> 1370 Anthony Williams, <ulink 1371 url="http://www.boost.org/doc/libs/1_37_0/doc/html/thread.html">Boost.Thread</ulink> 1372 Library Documentation, Boost website, 2007. 1373 </para> 1374 </listitem> 1375 <listitem> 1376 <para> 1377 Anthony Williams, <ulink 1378 url="http://www.ddj.com/cpp/211600441">What's New in Boost 1379 Threads?</ulink>, Recent changes to the Boost Thread library, 1380 Dr. Dobbs Magazine, October 2008. 1381 </para> 1382 </listitem> 1383</itemizedlist> 1384</para> 1385 1386</sect2> 1387 1388 1389<sect2 id="drd-manual.openmp" xreflabel="OpenMP"> 1390<title>Debugging OpenMP Programs</title> 1391 1392<para> 1393OpenMP stands for <emphasis>Open Multi-Processing</emphasis>. The OpenMP 1394standard consists of a set of compiler directives for C, C++ and Fortran 1395programs that allows a compiler to transform a sequential program into a 1396parallel program. OpenMP is well suited for HPC applications and allows to 1397work at a higher level compared to direct use of the POSIX threads API. While 1398OpenMP ensures that the POSIX API is used correctly, OpenMP programs can still 1399contain data races. So it definitely makes sense to verify OpenMP programs 1400with a thread checking tool. 1401</para> 1402 1403<para> 1404DRD supports OpenMP shared-memory programs generated by GCC. GCC 1405supports OpenMP since version 4.2.0. GCC's runtime support 1406for OpenMP programs is provided by a library called 1407<literal>libgomp</literal>. The synchronization primitives implemented 1408in this library use Linux' futex system call directly, unless the 1409library has been configured with the 1410<literal>--disable-linux-futex</literal> option. DRD only supports 1411libgomp libraries that have been configured with this option and in 1412which symbol information is present. For most Linux distributions this 1413means that you will have to recompile GCC. See also the script 1414<literal>drd/scripts/download-and-build-gcc</literal> in the 1415Valgrind source tree for an example of how to compile GCC. You will 1416also have to make sure that the newly compiled 1417<literal>libgomp.so</literal> library is loaded when OpenMP programs 1418are started. This is possible by adding a line similar to the 1419following to your shell startup script: 1420</para> 1421<programlisting><![CDATA[ 1422export LD_LIBRARY_PATH=~/gcc-4.4.0/lib64:~/gcc-4.4.0/lib: 1423]]></programlisting> 1424 1425<para> 1426As an example, the test OpenMP test program 1427<literal>drd/tests/omp_matinv</literal> triggers a data race 1428when the option -r has been specified on the command line. The data 1429race is triggered by the following code: 1430</para> 1431<programlisting><![CDATA[ 1432#pragma omp parallel for private(j) 1433for (j = 0; j < rows; j++) 1434{ 1435 if (i != j) 1436 { 1437 const elem_t factor = a[j * cols + i]; 1438 for (k = 0; k < cols; k++) 1439 { 1440 a[j * cols + k] -= a[i * cols + k] * factor; 1441 } 1442 } 1443} 1444]]></programlisting> 1445 1446<para> 1447The above code is racy because the variable <literal>k</literal> has 1448not been declared private. DRD will print the following error message 1449for the above code: 1450</para> 1451<programlisting><![CDATA[ 1452$ valgrind --tool=drd --check-stack-var=yes --read-var-info=yes drd/tests/omp_matinv 3 -t 2 -r 1453... 1454Conflicting store by thread 1/1 at 0x7fefffbc4 size 4 1455 at 0x4014A0: gj.omp_fn.0 (omp_matinv.c:203) 1456 by 0x401211: gj (omp_matinv.c:159) 1457 by 0x40166A: invert_matrix (omp_matinv.c:238) 1458 by 0x4019B4: main (omp_matinv.c:316) 1459Location 0x7fefffbc4 is 0 bytes inside local var "k" 1460declared at omp_matinv.c:160, in frame #0 of thread 1 1461... 1462]]></programlisting> 1463<para> 1464In the above output the function name <function>gj.omp_fn.0</function> 1465has been generated by GCC from the function name 1466<function>gj</function>. The allocation context information shows that the 1467data race has been caused by modifying the variable <literal>k</literal>. 1468</para> 1469 1470<para> 1471Note: for GCC versions before 4.4.0, no allocation context information is 1472shown. With these GCC versions the most usable information in the above output 1473is the source file name and the line number where the data race has been 1474detected (<literal>omp_matinv.c:203</literal>). 1475</para> 1476 1477<para> 1478For more information about OpenMP, see also 1479<ulink url="http://openmp.org/">openmp.org</ulink>. 1480</para> 1481 1482</sect2> 1483 1484 1485<sect2 id="drd-manual.cust-mem-alloc" xreflabel="Custom Memory Allocators"> 1486<title>DRD and Custom Memory Allocators</title> 1487 1488<para> 1489DRD tracks all memory allocation events that happen via the 1490standard memory allocation and deallocation functions 1491(<function>malloc</function>, <function>free</function>, 1492<function>new</function> and <function>delete</function>), via entry 1493and exit of stack frames or that have been annotated with Valgrind's 1494memory pool client requests. DRD uses memory allocation and deallocation 1495information for two purposes: 1496<itemizedlist> 1497 <listitem> 1498 <para> 1499 To know where the scope ends of POSIX objects that have not been 1500 destroyed explicitly. It is e.g. not required by the POSIX 1501 threads standard to call 1502 <function>pthread_mutex_destroy</function> before freeing the 1503 memory in which a mutex object resides. 1504 </para> 1505 </listitem> 1506 <listitem> 1507 <para> 1508 To know where the scope of variables ends. If e.g. heap memory 1509 has been used by one thread, that thread frees that memory, and 1510 another thread allocates and starts using that memory, no data 1511 races must be reported for that memory. 1512 </para> 1513 </listitem> 1514</itemizedlist> 1515</para> 1516 1517<para> 1518It is essential for correct operation of DRD that the tool knows about 1519memory allocation and deallocation events. When analyzing a client program 1520with DRD that uses a custom memory allocator, either instrument the custom 1521memory allocator with the <literal>VALGRIND_MALLOCLIKE_BLOCK</literal> 1522and <literal>VALGRIND_FREELIKE_BLOCK</literal> macros or disable the 1523custom memory allocator. 1524</para> 1525 1526<para> 1527As an example, the GNU libstdc++ library can be configured 1528to use standard memory allocation functions instead of memory pools by 1529setting the environment variable 1530<literal>GLIBCXX_FORCE_NEW</literal>. For more information, see also 1531the <ulink 1532url="http://gcc.gnu.org/onlinedocs/libstdc++/manual/bk01pt04ch11.html">libstdc++ 1533manual</ulink>. 1534</para> 1535 1536</sect2> 1537 1538 1539<sect2 id="drd-manual.drd-versus-memcheck" xreflabel="DRD Versus Memcheck"> 1540<title>DRD Versus Memcheck</title> 1541 1542<para> 1543It is essential for correct operation of DRD that there are no memory 1544errors such as dangling pointers in the client program. Which means that 1545it is a good idea to make sure that your program is Memcheck-clean 1546before you analyze it with DRD. It is possible however that some of 1547the Memcheck reports are caused by data races. In this case it makes 1548sense to run DRD before Memcheck. 1549</para> 1550 1551<para> 1552So which tool should be run first? In case both DRD and Memcheck 1553complain about a program, a possible approach is to run both tools 1554alternatingly and to fix as many errors as possible after each run of 1555each tool until none of the two tools prints any more error messages. 1556</para> 1557 1558</sect2> 1559 1560 1561<sect2 id="drd-manual.resource-requirements" xreflabel="Resource Requirements"> 1562<title>Resource Requirements</title> 1563 1564<para> 1565The requirements of DRD with regard to heap and stack memory and the 1566effect on the execution time of client programs are as follows: 1567<itemizedlist> 1568 <listitem> 1569 <para> 1570 When running a program under DRD with default DRD options, 1571 between 1.1 and 3.6 times more memory will be needed compared to 1572 a native run of the client program. More memory will be needed 1573 if loading debug information has been enabled 1574 (<literal>--read-var-info=yes</literal>). 1575 </para> 1576 </listitem> 1577 <listitem> 1578 <para> 1579 DRD allocates some of its temporary data structures on the stack 1580 of the client program threads. This amount of data is limited to 1581 1 - 2 KB. Make sure that thread stacks are sufficiently large. 1582 </para> 1583 </listitem> 1584 <listitem> 1585 <para> 1586 Most applications will run between 20 and 50 times slower under 1587 DRD than a native single-threaded run. The slowdown will be most 1588 noticeable for applications which perform frequent mutex lock / 1589 unlock operations. 1590 </para> 1591 </listitem> 1592</itemizedlist> 1593</para> 1594 1595</sect2> 1596 1597 1598<sect2 id="drd-manual.effective-use" xreflabel="Effective Use"> 1599<title>Hints and Tips for Effective Use of DRD</title> 1600 1601<para> 1602The following information may be helpful when using DRD: 1603<itemizedlist> 1604 <listitem> 1605 <para> 1606 Make sure that debug information is present in the executable 1607 being analyzed, such that DRD can print function name and line 1608 number information in stack traces. Most compilers can be told 1609 to include debug information via compiler option 1610 <option>-g</option>. 1611 </para> 1612 </listitem> 1613 <listitem> 1614 <para> 1615 Compile with option <option>-O1</option> instead of 1616 <option>-O0</option>. This will reduce the amount of generated 1617 code, may reduce the amount of debug info and will speed up 1618 DRD's processing of the client program. For more information, 1619 see also <xref linkend="manual-core.started"/>. 1620 </para> 1621 </listitem> 1622 <listitem> 1623 <para> 1624 If DRD reports any errors on libraries that are part of your 1625 Linux distribution like e.g. <literal>libc.so</literal> or 1626 <literal>libstdc++.so</literal>, installing the debug packages 1627 for these libraries will make the output of DRD a lot more 1628 detailed. 1629 </para> 1630 </listitem> 1631 <listitem> 1632 <para> 1633 When using C++, do not send output from more than one thread to 1634 <literal>std::cout</literal>. Doing so would not only 1635 generate multiple data race reports, it could also result in 1636 output from several threads getting mixed up. Either use 1637 <function>printf</function> or do the following: 1638 <orderedlist> 1639 <listitem> 1640 <para>Derive a class from <literal>std::ostreambuf</literal> 1641 and let that class send output line by line to 1642 <literal>stdout</literal>. This will avoid that individual 1643 lines of text produced by different threads get mixed 1644 up.</para> 1645 </listitem> 1646 <listitem> 1647 <para>Create one instance of <literal>std::ostream</literal> 1648 for each thread. This makes stream formatting settings 1649 thread-local. Pass a per-thread instance of the class 1650 derived from <literal>std::ostreambuf</literal> to the 1651 constructor of each instance. </para> 1652 </listitem> 1653 <listitem> 1654 <para>Let each thread send its output to its own instance of 1655 <literal>std::ostream</literal> instead of 1656 <literal>std::cout</literal>.</para> 1657 </listitem> 1658 </orderedlist> 1659 </para> 1660 </listitem> 1661</itemizedlist> 1662</para> 1663 1664</sect2> 1665 1666 1667</sect1> 1668 1669 1670<sect1 id="drd-manual.Pthreads" xreflabel="Pthreads"> 1671<title>Using the POSIX Threads API Effectively</title> 1672 1673<sect2 id="drd-manual.mutex-types" xreflabel="mutex-types"> 1674<title>Mutex types</title> 1675 1676<para> 1677The Single UNIX Specification version two defines the following four 1678mutex types (see also the documentation of <ulink 1679url="http://www.opengroup.org/onlinepubs/007908799/xsh/pthread_mutexattr_settype.html"><function>pthread_mutexattr_settype</function></ulink>): 1680<itemizedlist> 1681 <listitem> 1682 <para> 1683 <emphasis>normal</emphasis>, which means that no error checking 1684 is performed, and that the mutex is non-recursive. 1685 </para> 1686 </listitem> 1687 <listitem> 1688 <para> 1689 <emphasis>error checking</emphasis>, which means that the mutex 1690 is non-recursive and that error checking is performed. 1691 </para> 1692 </listitem> 1693 <listitem> 1694 <para> 1695 <emphasis>recursive</emphasis>, which means that a mutex may be 1696 locked recursively. 1697 </para> 1698 </listitem> 1699 <listitem> 1700 <para> 1701 <emphasis>default</emphasis>, which means that error checking 1702 behavior is undefined, and that the behavior for recursive 1703 locking is also undefined. Or: portable code must neither 1704 trigger error conditions through the Pthreads API nor attempt to 1705 lock a mutex of default type recursively. 1706 </para> 1707 </listitem> 1708</itemizedlist> 1709</para> 1710 1711<para> 1712In complex applications it is not always clear from beforehand which 1713mutex will be locked recursively and which mutex will not be locked 1714recursively. Attempts lock a non-recursive mutex recursively will 1715result in race conditions that are very hard to find without a thread 1716checking tool. So either use the error checking mutex type and 1717consistently check the return value of Pthread API mutex calls, or use 1718the recursive mutex type. 1719</para> 1720 1721</sect2> 1722 1723<sect2 id="drd-manual.condvar" xreflabel="condition-variables"> 1724<title>Condition variables</title> 1725 1726<para> 1727A condition variable allows one thread to wake up one or more other 1728threads. Condition variables are often used to notify one or more 1729threads about state changes of shared data. Unfortunately it is very 1730easy to introduce race conditions by using condition variables as the 1731only means of state information propagation. A better approach is to 1732let threads poll for changes of a state variable that is protected by 1733a mutex, and to use condition variables only as a thread wakeup 1734mechanism. See also the source file 1735<computeroutput>drd/tests/monitor_example.cpp</computeroutput> for an 1736example of how to implement this concept in C++. The monitor concept 1737used in this example is a well known and very useful concept -- see 1738also Wikipedia for more information about the <ulink 1739url="http://en.wikipedia.org/wiki/Monitor_(synchronization)">monitor</ulink> 1740concept. 1741</para> 1742 1743</sect2> 1744 1745<sect2 id="drd-manual.pctw" xreflabel="pthread_cond_timedwait"> 1746<title>pthread_cond_timedwait and timeouts</title> 1747 1748<para> 1749Historically the function 1750<function>pthread_cond_timedwait</function> only allowed the 1751specification of an absolute timeout, that is a timeout independent of 1752the time when this function was called. However, almost every call to 1753this function expresses a relative timeout. This typically happens by 1754passing the sum of 1755<computeroutput>clock_gettime(CLOCK_REALTIME)</computeroutput> and a 1756relative timeout as the third argument. This approach is incorrect 1757since forward or backward clock adjustments by e.g. ntpd will affect 1758the timeout. A more reliable approach is as follows: 1759<itemizedlist> 1760 <listitem> 1761 <para> 1762 When initializing a condition variable through 1763 <function>pthread_cond_init</function>, specify that the timeout of 1764 <function>pthread_cond_timedwait</function> will use the clock 1765 <literal>CLOCK_MONOTONIC</literal> instead of 1766 <literal>CLOCK_REALTIME</literal>. You can do this via 1767 <computeroutput>pthread_condattr_setclock(..., 1768 CLOCK_MONOTONIC)</computeroutput>. 1769 </para> 1770 </listitem> 1771 <listitem> 1772 <para> 1773 When calling <function>pthread_cond_timedwait</function>, pass 1774 the sum of 1775 <computeroutput>clock_gettime(CLOCK_MONOTONIC)</computeroutput> 1776 and a relative timeout as the third argument. 1777 </para> 1778 </listitem> 1779</itemizedlist> 1780See also 1781<computeroutput>drd/tests/monitor_example.cpp</computeroutput> for an 1782example. 1783</para> 1784 1785</sect2> 1786 1787</sect1> 1788 1789 1790<sect1 id="drd-manual.limitations" xreflabel="Limitations"> 1791<title>Limitations</title> 1792 1793<para>DRD currently has the following limitations:</para> 1794 1795<itemizedlist> 1796 <listitem> 1797 <para> 1798 DRD, just like Memcheck, will refuse to start on Linux 1799 distributions where all symbol information has been removed from 1800 <filename>ld.so</filename>. This is e.g. the case for the PPC editions 1801 of openSUSE and Gentoo. You will have to install the glibc debuginfo 1802 package on these platforms before you can use DRD. See also openSUSE 1803 bug <ulink url="http://bugzilla.novell.com/show_bug.cgi?id=396197"> 1804 396197</ulink> and Gentoo bug <ulink 1805 url="http://bugs.gentoo.org/214065">214065</ulink>. 1806 </para> 1807 </listitem> 1808 <listitem> 1809 <para> 1810 With gcc 4.4.3 and before, DRD may report data races on the C++ 1811 class <literal>std::string</literal> in a multithreaded program. This is 1812 a know <literal>libstdc++</literal> issue -- see also GCC bug 1813 <ulink url="http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40518">40518</ulink> 1814 for more information. 1815 </para> 1816 </listitem> 1817 <listitem> 1818 <para> 1819 If you compile the DRD source code yourself, you need GCC 3.0 or 1820 later. GCC 2.95 is not supported. 1821 </para> 1822 </listitem> 1823 <listitem> 1824 <para> 1825 Of the two POSIX threads implementations for Linux, only the 1826 NPTL (Native POSIX Thread Library) is supported. The older 1827 LinuxThreads library is not supported. 1828 </para> 1829 </listitem> 1830</itemizedlist> 1831 1832</sect1> 1833 1834 1835<sect1 id="drd-manual.feedback" xreflabel="Feedback"> 1836<title>Feedback</title> 1837 1838<para> 1839If you have any comments, suggestions, feedback or bug reports about 1840DRD, feel free to either post a message on the Valgrind users mailing 1841list or to file a bug report. See also <ulink 1842url="&vg-url;">&vg-url;</ulink> for more information. 1843</para> 1844 1845</sect1> 1846 1847 1848</chapter> 1849