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; ]>
7 <chapter id="drd-manual" xreflabel="DRD: a thread error detector">
8 <title>DRD: a thread error detector</title>
10 <para>To use this tool, you must specify
11 <option>--tool=drd</option>
12 on the Valgrind command line.</para>
15 <sect1 id="drd-manual.overview" xreflabel="Overview">
16 <title>Overview</title>
19 DRD is a Valgrind tool for detecting errors in multithreaded C and C++
20 programs. The tool works for any program that uses the POSIX threading
21 primitives or that uses threading concepts built on top of the POSIX threading
25 <sect2 id="drd-manual.mt-progr-models" xreflabel="MT-progr-models">
26 <title>Multithreaded Programming Paradigms</title>
29 There are two possible reasons for using multithreading in a program:
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.
41 To use multiple CPU cores simultaneously for speeding up
42 computations. This is why many High Performance Computing (HPC)
43 applications are multithreaded.
50 Multithreaded programs can use one or more of the following programming
51 paradigms. Which paradigm is appropriate depends e.g. on the application type.
52 Some examples of multithreaded programming paradigms are:
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.
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.
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.
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.
97 DRD supports any combination of multithreaded programming paradigms as
98 long as the implementation of these paradigms is based on the POSIX
99 threads primitives. DRD however does not support programs that use
100 e.g. Linux' futexes directly. Attempts to analyze such programs with
101 DRD will cause DRD to report many false positives.
107 <sect2 id="drd-manual.pthreads-model" xreflabel="Pthreads-model">
108 <title>POSIX Threads Programming Model</title>
111 POSIX threads, also known as Pthreads, is the most widely available
112 threading library on Unix systems.
116 The POSIX threads programming model is based on the following abstractions:
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.
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.
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.
141 Threads. Each thread represents a concurrent activity.
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
157 Which source code statements generate which memory accesses depends on
158 the <emphasis>memory model</emphasis> of the programming language being
159 used. There is not yet a definitive memory model for the C and C++
160 languages. 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">
162 WG21/N2338: Concurrency memory model compiler consequences</ulink>.
166 For more information about POSIX threads, see also the Single UNIX
167 Specification version 3, also known as
168 <ulink url="http://www.opengroup.org/onlinepubs/000095399/idx/threads.html">
169 IEEE Std 1003.1</ulink>.
175 <sect2 id="drd-manual.mt-problems" xreflabel="MT-Problems">
176 <title>Multithreaded Programming Problems</title>
179 Depending on which multithreading paradigm is being used in a program,
180 one or more of the following problems can occur:
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.
191 Lock contention. One thread blocks the progress of one or more other
192 threads by holding a lock too long.
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.
205 Deadlock. A deadlock occurs when two or more threads wait for
206 each other indefinitely.
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.
221 Although the likelihood of the occurrence of data races can be reduced
222 through a disciplined programming style, a tool for automatic
223 detection of data races is a necessity when developing multithreaded
224 software. DRD can detect these, as well as lock contention and
225 improper use of the POSIX threads API.
231 <sect2 id="drd-manual.data-race-detection" xreflabel="data-race-detection">
232 <title>Data Race Detection</title>
235 The result of load and store operations performed by a multithreaded program
236 depends on the order in which memory operations are performed. This order is
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.
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>.
254 The combination of program order and synchronization order is called the
255 <emphasis>happens-before relationship</emphasis>. This concept was first
256 defined by S. Adve et al in the paper <emphasis>Detecting data races on weak
257 memory systems</emphasis>, ACM SIGARCH Computer Architecture News, v.19 n.3,
262 Two memory operations <emphasis>conflict</emphasis> if both operations are
263 performed by different threads, refer to the same memory location and at least
264 one of them is a store operation.
268 A multithreaded program is <emphasis>data-race free</emphasis> if all
269 conflicting memory accesses are ordered by synchronization
274 A well known way to ensure that a multithreaded program is data-race
275 free is to ensure that a locking discipline is followed. It is e.g.
276 possible to associate a mutex with each shared data item, and to hold
277 a lock on the associated mutex while the shared data is accessed.
281 All programs that follow a locking discipline are data-race free, but not all
282 data-race free programs follow a locking discipline. There exist multithreaded
283 programs where access to shared data is arbitrated via condition variables,
284 semaphores or barriers. As an example, a certain class of HPC applications
285 consists of a sequence of computation steps separated in time by barriers, and
286 where these barriers are the only means of synchronization. Although there are
287 many conflicting memory accesses in such applications and although such
288 applications do not make use mutexes, most of these applications do not
293 There exist two different approaches for verifying the correctness of
294 multithreaded programs at runtime. The approach of the so-called Eraser
295 algorithm is to verify whether all shared memory accesses follow a consistent
296 locking strategy. And the happens-before data race detectors verify directly
297 whether all interthread memory accesses are ordered by synchronization
298 operations. While the last approach is more complex to implement, and while it
299 is more sensitive to OS scheduling, it is a general approach that works for
300 all classes of multithreaded programs. An important advantage of
301 happens-before data race detectors is that these do not report any false
306 DRD is based on the happens-before algorithm.
315 <sect1 id="drd-manual.using-drd" xreflabel="Using DRD">
316 <title>Using DRD</title>
318 <sect2 id="drd-manual.options" xreflabel="DRD Command-line Options">
319 <title>DRD Command-line Options</title>
321 <para>The following command-line options are available for controlling the
322 behavior of the DRD tool itself:</para>
324 <!-- start of xi:include in the manpage -->
325 <variablelist id="drd.opts.list">
328 <option><![CDATA[--check-stack-var=<yes|no> [default: no]]]></option>
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.
340 <option><![CDATA[--exclusive-threshold=<n> [default: off]]]></option>
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.
353 <![CDATA[--first-race-only=<yes|no> [default: no]]]>
358 Whether to report only the first data race that has been detected on a
359 memory location or all data races that have been detected on a memory
367 <![CDATA[--free-is-write=<yes|no> [default: no]]]>
372 Whether to report races between accessing memory and freeing
373 memory. Enabling this option may cause DRD to run slightly
378 Don't enable this option when using custom memory allocators
380 the <computeroutput>VG_USERREQ__MALLOCLIKE_BLOCK</computeroutput>
381 and <computeroutput>VG_USERREQ__FREELIKE_BLOCK</computeroutput>
382 because that would result in false positives.
386 <para>Don't enable this option when using reference-counted
387 objects because that will result in false positives, even when
388 that code has been annotated properly with
389 <computeroutput>ANNOTATE_HAPPENS_BEFORE</computeroutput>
390 and <computeroutput>ANNOTATE_HAPPENS_AFTER</computeroutput>. See
391 e.g. the output of the following command for an example:
392 <computeroutput>valgrind --tool=drd --free-is-write=yes
393 drd/tests/annotate_smart_pointer</computeroutput>.
403 <![CDATA[--report-signal-unlocked=<yes|no> [default: yes]]]>
408 Whether to report calls to
409 <function>pthread_cond_signal</function> and
410 <function>pthread_cond_broadcast</function> where the mutex
411 associated with the signal through
412 <function>pthread_cond_wait</function> or
413 <function>pthread_cond_timed_wait</function>is not locked at
414 the time the signal is sent. Sending a signal without holding
415 a lock on the associated mutex is a common programming error
416 which can cause subtle race conditions and unpredictable
417 behavior. There exist some uncommon synchronization patterns
418 however where it is safe to send a signal without holding a
419 lock on the associated mutex.
425 <option><![CDATA[--segment-merging=<yes|no> [default: yes]]]></option>
429 Controls segment merging. Segment merging is an algorithm to
430 limit memory usage of the data race detection
431 algorithm. Disabling segment merging may improve the accuracy
432 of the so-called 'other segments' displayed in race reports
433 but can also trigger an out of memory error.
439 <option><![CDATA[--segment-merging-interval=<n> [default: 10]]]></option>
443 Perform segment merging only after the specified number of new
444 segments have been created. This is an advanced configuration option
445 that allows to choose whether to minimize DRD's memory usage by
446 choosing a low value or to let DRD run faster by choosing a slightly
447 higher value. The optimal value for this parameter depends on the
448 program being analyzed. The default value works well for most programs.
454 <option><![CDATA[--shared-threshold=<n> [default: off]]]></option>
458 Print an error message if a reader lock has been held longer
459 than the specified time (in milliseconds). This option enables
460 the detection of lock contention.
466 <option><![CDATA[--show-confl-seg=<yes|no> [default: yes]]]></option>
470 Show conflicting segments in race reports. Since this
471 information can help to find the cause of a data race, this
472 option is enabled by default. Disabling this option makes the
473 output of DRD more compact.
479 <option><![CDATA[--show-stack-usage=<yes|no> [default: no]]]></option>
483 Print stack usage at thread exit time. When a program creates a large
484 number of threads it becomes important to limit the amount of virtual
485 memory allocated for thread stacks. This option makes it possible to
486 observe how much stack memory has been used by each thread of the the
487 client program. Note: the DRD tool itself allocates some temporary
488 data on the client thread stack. The space necessary for this
489 temporary data must be allocated by the client program when it
490 allocates stack memory, but is not included in stack usage reported by
496 <!-- end of xi:include in the manpage -->
498 <!-- start of xi:include in the manpage -->
500 The following options are available for monitoring the behavior of the
504 <variablelist id="drd.debugopts.list">
507 <option><![CDATA[--trace-addr=<address> [default: none]]]></option>
511 Trace all load and store activity for the specified
512 address. This option may be specified more than once.
518 <option><![CDATA[--trace-alloc=<yes|no> [default: no]]]></option>
522 Trace all memory allocations and deallocations. May produce a huge
529 <option><![CDATA[--trace-barrier=<yes|no> [default: no]]]></option>
533 Trace all barrier activity.
539 <option><![CDATA[--trace-cond=<yes|no> [default: no]]]></option>
543 Trace all condition variable activity.
549 <option><![CDATA[--trace-fork-join=<yes|no> [default: no]]]></option>
553 Trace all thread creation and all thread termination events.
559 <option><![CDATA[--trace-mutex=<yes|no> [default: no]]]></option>
563 Trace all mutex activity.
569 <option><![CDATA[--trace-rwlock=<yes|no> [default: no]]]></option>
573 Trace all reader-writer lock activity.
579 <option><![CDATA[--trace-semaphore=<yes|no> [default: no]]]></option>
583 Trace all semaphore activity.
588 <!-- end of xi:include in the manpage -->
593 <sect2 id="drd-manual.data-races" xreflabel="Data Races">
594 <title>Detected Errors: Data Races</title>
597 DRD prints a message every time it detects a data race. Please keep
598 the following in mind when interpreting DRD's output:
602 Every thread is assigned a <emphasis>thread ID</emphasis> by the DRD
603 tool. A thread ID is a number. Thread ID's start at one and are never
609 The term <emphasis>segment</emphasis> refers to a consecutive
610 sequence of load, store and synchronization operations, all
611 issued by the same thread. A segment always starts and ends at a
612 synchronization operation. Data race analysis is performed
613 between segments instead of between individual load and store
614 operations because of performance reasons.
619 There are always at least two memory accesses involved in a data
620 race. Memory accesses involved in a data race are called
621 <emphasis>conflicting memory accesses</emphasis>. DRD prints a
622 report for each memory access that conflicts with a past memory
630 Below you can find an example of a message printed by DRD when it
633 <programlisting><![CDATA[
634 $ valgrind --tool=drd --read-var-info=yes drd/tests/rwlock_race
637 ==9466== Conflicting load by thread 3 at 0x006020b8 size 4
638 ==9466== at 0x400B6C: thread_func (rwlock_race.c:29)
639 ==9466== by 0x4C291DF: vg_thread_wrapper (drd_pthread_intercepts.c:186)
640 ==9466== by 0x4E3403F: start_thread (in /lib64/libpthread-2.8.so)
641 ==9466== by 0x53250CC: clone (in /lib64/libc-2.8.so)
642 ==9466== Location 0x6020b8 is 0 bytes inside local var "s_racy"
643 ==9466== declared at rwlock_race.c:18, in frame #0 of thread 3
644 ==9466== Other segment start (thread 2)
645 ==9466== at 0x4C2847D: pthread_rwlock_rdlock* (drd_pthread_intercepts.c:813)
646 ==9466== by 0x400B6B: thread_func (rwlock_race.c:28)
647 ==9466== by 0x4C291DF: vg_thread_wrapper (drd_pthread_intercepts.c:186)
648 ==9466== by 0x4E3403F: start_thread (in /lib64/libpthread-2.8.so)
649 ==9466== by 0x53250CC: clone (in /lib64/libc-2.8.so)
650 ==9466== Other segment end (thread 2)
651 ==9466== at 0x4C28B54: pthread_rwlock_unlock* (drd_pthread_intercepts.c:912)
652 ==9466== by 0x400B84: thread_func (rwlock_race.c:30)
653 ==9466== by 0x4C291DF: vg_thread_wrapper (drd_pthread_intercepts.c:186)
654 ==9466== by 0x4E3403F: start_thread (in /lib64/libpthread-2.8.so)
655 ==9466== by 0x53250CC: clone (in /lib64/libc-2.8.so)
660 The above report has the following meaning:
664 The number in the column on the left is the process ID of the
665 process being analyzed by DRD.
670 The first line ("Thread 3") tells you the thread ID for
671 the thread in which context the data race has been detected.
676 The next line tells which kind of operation was performed (load or
677 store) and by which thread. On the same line the start address and the
678 number of bytes involved in the conflicting access are also displayed.
683 Next, the call stack of the conflicting access is displayed. If
684 your program has been compiled with debug information
685 (<option>-g</option>), this call stack will include file names and
686 line numbers. The two
687 bottommost frames in this call stack (<function>clone</function>
688 and <function>start_thread</function>) show how the NPTL starts
689 a thread. The third frame
690 (<function>vg_thread_wrapper</function>) is added by DRD. The
691 fourth frame (<function>thread_func</function>) is the first
692 interesting line because it shows the thread entry point, that
693 is the function that has been passed as the third argument to
694 <function>pthread_create</function>.
699 Next, the allocation context for the conflicting address is
700 displayed. For dynamically allocated data the allocation call
701 stack is shown. For static variables and stack variables the
702 allocation context is only shown when the option
703 <option>--read-var-info=yes</option> has been
704 specified. Otherwise DRD will print <computeroutput>Allocation
705 context: unknown</computeroutput>.
710 A conflicting access involves at least two memory accesses. For
711 one of these accesses an exact call stack is displayed, and for
712 the other accesses an approximate call stack is displayed,
713 namely the start and the end of the segments of the other
714 accesses. This information can be interpreted as follows:
718 Start at the bottom of both call stacks, and count the
719 number stack frames with identical function name, file
720 name and line number. In the above example the three
721 bottommost frames are identical
722 (<function>clone</function>,
723 <function>start_thread</function> and
724 <function>vg_thread_wrapper</function>).
729 The next higher stack frame in both call stacks now tells
730 you between in which source code region the other memory
731 access happened. The above output tells that the other
732 memory access involved in the data race happened between
733 source code lines 28 and 30 in file
734 <computeroutput>rwlock_race.c</computeroutput>.
746 <sect2 id="drd-manual.lock-contention" xreflabel="Lock Contention">
747 <title>Detected Errors: Lock Contention</title>
750 Threads must be able to make progress without being blocked for too long by
751 other threads. Sometimes a thread has to wait until a mutex or reader-writer
752 synchronization object is unlocked by another thread. This is called
753 <emphasis>lock contention</emphasis>.
757 Lock contention causes delays. Such delays should be as short as
758 possible. The two command line options
759 <literal>--exclusive-threshold=<n></literal> and
760 <literal>--shared-threshold=<n></literal> make it possible to
761 detect excessive lock contention by making DRD report any lock that
762 has been held longer than the specified threshold. An example:
764 <programlisting><![CDATA[
765 $ valgrind --tool=drd --exclusive-threshold=10 drd/tests/hold_lock -i 500
767 ==10668== Acquired at:
768 ==10668== at 0x4C267C8: pthread_mutex_lock (drd_pthread_intercepts.c:395)
769 ==10668== by 0x400D92: main (hold_lock.c:51)
770 ==10668== Lock on mutex 0x7fefffd50 was held during 503 ms (threshold: 10 ms).
771 ==10668== at 0x4C26ADA: pthread_mutex_unlock (drd_pthread_intercepts.c:441)
772 ==10668== by 0x400DB5: main (hold_lock.c:55)
777 The <literal>hold_lock</literal> test program holds a lock as long as
778 specified by the <literal>-i</literal> (interval) argument. The DRD
779 output reports that the lock acquired at line 51 in source file
780 <literal>hold_lock.c</literal> and released at line 55 was held during
781 503 ms, while a threshold of 10 ms was specified to DRD.
787 <sect2 id="drd-manual.api-checks" xreflabel="API Checks">
788 <title>Detected Errors: Misuse of the POSIX threads API</title>
791 DRD is able to detect and report the following misuses of the POSIX
796 Passing the address of one type of synchronization object
797 (e.g. a mutex) to a POSIX API call that expects a pointer to
798 another type of synchronization object (e.g. a condition
804 Attempts to unlock a mutex that has not been locked.
809 Attempts to unlock a mutex that was locked by another thread.
814 Attempts to lock a mutex of type
815 <literal>PTHREAD_MUTEX_NORMAL</literal> or a spinlock
821 Destruction or deallocation of a locked mutex.
826 Sending a signal to a condition variable while no lock is held
827 on the mutex associated with the condition variable.
832 Calling <function>pthread_cond_wait</function> on a mutex
833 that is not locked, that is locked by another thread or that
834 has been locked recursively.
839 Associating two different mutexes with a condition variable
840 through <function>pthread_cond_wait</function>.
845 Destruction or deallocation of a condition variable that is
851 Destruction or deallocation of a locked reader-writer synchronization
857 Attempts to unlock a reader-writer synchronization object that was not
858 locked by the calling thread.
863 Attempts to recursively lock a reader-writer synchronization object
869 Attempts to pass the address of a user-defined reader-writer
870 synchronization object to a POSIX threads function.
875 Attempts to pass the address of a POSIX reader-writer synchronization
876 object to one of the annotations for user-defined reader-writer
877 synchronization objects.
882 Reinitialization of a mutex, condition variable, reader-writer
883 lock, semaphore or barrier.
888 Destruction or deallocation of a semaphore or barrier that is
894 Missing synchronization between barrier wait and barrier destruction.
899 Exiting a thread without first unlocking the spinlocks, mutexes or
900 reader-writer synchronization objects that were locked by that thread.
905 Passing an invalid thread ID to <function>pthread_join</function>
906 or <function>pthread_cancel</function>.
915 <sect2 id="drd-manual.clientreqs" xreflabel="Client requests">
916 <title>Client Requests</title>
919 Just as for other Valgrind tools it is possible to let a client program
920 interact with the DRD tool through client requests. In addition to the
921 client requests several macros have been defined that allow to use the
922 client requests in a convenient way.
926 The interface between client programs and the DRD tool is defined in
927 the header file <literal><valgrind/drd.h></literal>. The
928 available macros and client requests are:
932 The macro <literal>DRD_GET_VALGRIND_THREADID</literal> and the
934 request <varname>VG_USERREQ__DRD_GET_VALGRIND_THREAD_ID</varname>.
935 Query the thread ID that has been assigned by the Valgrind core to the
936 thread executing this client request. Valgrind's thread ID's start at
937 one and are recycled in case a thread stops.
942 The macro <literal>DRD_GET_DRD_THREADID</literal> and the corresponding
943 client request <varname>VG_USERREQ__DRD_GET_DRD_THREAD_ID</varname>.
944 Query the thread ID that has been assigned by DRD to the thread
945 executing this client request. These are the thread ID's reported by DRD
946 in data race reports and in trace messages. DRD's thread ID's start at
947 one and are never recycled.
952 The macros <literal>DRD_IGNORE_VAR(x)</literal>,
953 <literal>ANNOTATE_TRACE_MEMORY(&x)</literal> and the corresponding
954 client request <varname>VG_USERREQ__DRD_START_SUPPRESSION</varname>. Some
955 applications contain intentional races. There exist e.g. applications
956 where the same value is assigned to a shared variable from two different
957 threads. It may be more convenient to suppress such races than to solve
958 these. This client request allows to suppress such races.
963 The macro <literal>DRD_STOP_IGNORING_VAR(x)</literal> and the
964 corresponding client request
965 <varname>VG_USERREQ__DRD_FINISH_SUPPRESSION</varname>. Tell DRD
966 to no longer ignore data races for the address range that was suppressed
967 either via the macro <literal>DRD_IGNORE_VAR(x)</literal> or via the
968 client request <varname>VG_USERREQ__DRD_START_SUPPRESSION</varname>.
973 The macro <literal>DRD_TRACE_VAR(x)</literal>. Trace all load and store
974 activity for the address range starting at <literal>&x</literal> and
975 occupying <literal>sizeof(x)</literal> bytes. When DRD reports a data
976 race on a specified variable, and it's not immediately clear which
977 source code statements triggered the conflicting accesses, it can be
978 very helpful to trace all activity on the offending memory location.
983 The macro <literal>ANNOTATE_TRACE_MEMORY(&x)</literal>. Trace all
984 load and store activity that touches at least the single byte at the
985 address <literal>&x</literal>.
990 The client request <varname>VG_USERREQ__DRD_START_TRACE_ADDR</varname>,
991 which allows to trace all load and store activity for the specified
998 request <varname>VG_USERREQ__DRD_STOP_TRACE_ADDR</varname>. Do no longer
999 trace load and store activity for the specified address range.
1004 The macro <literal>ANNOTATE_HAPPENS_BEFORE(addr)</literal> tells DRD to
1005 insert a mark. Insert this macro just after an access to the variable at
1006 the specified address has been performed.
1011 The macro <literal>ANNOTATE_HAPPENS_AFTER(addr)</literal> tells DRD that
1012 the next access to the variable at the specified address should be
1013 considered to have happened after the access just before the latest
1014 <literal>ANNOTATE_HAPPENS_BEFORE(addr)</literal> annotation that
1015 references the same variable. The purpose of these two macros is to tell
1016 DRD about the order of inter-thread memory accesses implemented via
1017 atomic memory operations. See
1018 also <literal>drd/tests/annotate_smart_pointer.cpp</literal> for an
1024 The macro <literal>ANNOTATE_RWLOCK_CREATE(rwlock)</literal> tells DRD
1025 that the object at address <literal>rwlock</literal> is a
1026 reader-writer synchronization object that is not a
1027 <literal>pthread_rwlock_t</literal> synchronization object. See
1028 also <literal>drd/tests/annotate_rwlock.c</literal> for an example.
1033 The macro <literal>ANNOTATE_RWLOCK_DESTROY(rwlock)</literal> tells DRD
1034 that the reader-writer synchronization object at
1035 address <literal>rwlock</literal> has been destroyed.
1040 The macro <literal>ANNOTATE_WRITERLOCK_ACQUIRED(rwlock)</literal> tells
1041 DRD that a writer lock has been acquired on the reader-writer
1042 synchronization object at address <literal>rwlock</literal>.
1047 The macro <literal>ANNOTATE_READERLOCK_ACQUIRED(rwlock)</literal> tells
1048 DRD that a reader lock has been acquired on the reader-writer
1049 synchronization object at address <literal>rwlock</literal>.
1054 The macro <literal>ANNOTATE_RWLOCK_ACQUIRED(rwlock, is_w)</literal>
1055 tells DRD that a writer lock (when <literal>is_w != 0</literal>) or that
1056 a reader lock (when <literal>is_w == 0</literal>) has been acquired on
1057 the reader-writer synchronization object at
1058 address <literal>rwlock</literal>.
1063 The macro <literal>ANNOTATE_WRITERLOCK_RELEASED(rwlock)</literal> tells
1064 DRD that a writer lock has been released on the reader-writer
1065 synchronization object at address <literal>rwlock</literal>.
1070 The macro <literal>ANNOTATE_READERLOCK_RELEASED(rwlock)</literal> tells
1071 DRD that a reader lock has been released on the reader-writer
1072 synchronization object at address <literal>rwlock</literal>.
1077 The macro <literal>ANNOTATE_RWLOCK_RELEASED(rwlock, is_w)</literal>
1078 tells DRD that a writer lock (when <literal>is_w != 0</literal>) or that
1079 a reader lock (when <literal>is_w == 0</literal>) has been released on
1080 the reader-writer synchronization object at
1081 address <literal>rwlock</literal>.
1086 The macro <literal>ANNOTATE_BARRIER_INIT(barrier, count,
1087 reinitialization_allowed)</literal> tells DRD that a new barrier object
1088 at the address <literal>barrier</literal> has been initialized,
1089 that <literal>count</literal> threads participate in each barrier and
1090 also whether or not barrier reinitialization without intervening
1091 destruction should be reported as an error. See
1092 also <literal>drd/tests/annotate_barrier.c</literal> for an example.
1097 The macro <literal>ANNOTATE_BARRIER_DESTROY(barrier)</literal>
1098 tells DRD that a barrier object is about to be destroyed.
1103 The macro <literal>ANNOTATE_BARRIER_WAIT_BEFORE(barrier)</literal>
1104 tells DRD that waiting for a barrier will start.
1109 The macro <literal>ANNOTATE_BARRIER_WAIT_AFTER(barrier)</literal>
1110 tells DRD that waiting for a barrier has finished.
1115 The macro <literal>ANNOTATE_BENIGN_RACE_SIZED(addr, size,
1116 descr)</literal> tells DRD that any races detected on the specified
1117 address are benign and hence should not be
1118 reported. The <literal>descr</literal> argument is ignored but can be
1119 used to document why data races on <literal>addr</literal> are benign.
1124 The macro <literal>ANNOTATE_BENIGN_RACE_STATIC(var, descr)</literal>
1125 tells DRD that any races detected on the specified static variable are
1126 benign and hence should not be reported. The <literal>descr</literal>
1127 argument is ignored but can be used to document why data races
1128 on <literal>var</literal> are benign. Note: this macro can only be
1129 used in C++ programs and not in C programs.
1134 The macro <literal>ANNOTATE_IGNORE_READS_BEGIN</literal> tells
1135 DRD to ignore all memory loads performed by the current thread.
1140 The macro <literal>ANNOTATE_IGNORE_READS_END</literal> tells
1141 DRD to stop ignoring the memory loads performed by the current thread.
1146 The macro <literal>ANNOTATE_IGNORE_WRITES_BEGIN</literal> tells
1147 DRD to ignore all memory stores performed by the current thread.
1152 The macro <literal>ANNOTATE_IGNORE_WRITES_END</literal> tells
1153 DRD to stop ignoring the memory stores performed by the current thread.
1158 The macro <literal>ANNOTATE_IGNORE_READS_AND_WRITES_BEGIN</literal> tells
1159 DRD to ignore all memory accesses performed by the current thread.
1164 The macro <literal>ANNOTATE_IGNORE_READS_AND_WRITES_END</literal> tells
1165 DRD to stop ignoring the memory accesses performed by the current thread.
1170 The macro <literal>ANNOTATE_NEW_MEMORY(addr, size)</literal> tells
1171 DRD that the specified memory range has been allocated by a custom
1172 memory allocator in the client program and that the client program
1173 will start using this memory range.
1178 The macro <literal>ANNOTATE_THREAD_NAME(name)</literal> tells DRD to
1179 associate the specified name with the current thread and to include this
1180 name in the error messages printed by DRD.
1185 The macros <literal>VALGRIND_MALLOCLIKE_BLOCK</literal> and
1186 <literal>VALGRIND_FREELIKE_BLOCK</literal> from the Valgrind core are
1187 implemented; they are described in
1188 <xref linkend="manual-core-adv.clientreq"/>.
1195 Note: if you compiled Valgrind yourself, the header file
1196 <literal><valgrind/drd.h></literal> will have been installed in
1197 the directory <literal>/usr/include</literal> by the command
1198 <literal>make install</literal>. If you obtained Valgrind by
1199 installing it as a package however, you will probably have to install
1200 another package with a name like <literal>valgrind-devel</literal>
1201 before Valgrind's header files are available.
1207 <sect2 id="drd-manual.gnome" xreflabel="GNOME">
1208 <title>Debugging GNOME Programs</title>
1211 GNOME applications use the threading primitives provided by the
1212 <computeroutput>glib</computeroutput> and
1213 <computeroutput>gthread</computeroutput> libraries. These libraries
1214 are built on top of POSIX threads, and hence are directly supported by
1215 DRD. Please keep in mind that you have to call
1216 <function>g_thread_init</function> before creating any threads, or
1217 DRD will report several data races on glib functions. See also the
1219 url="http://library.gnome.org/devel/glib/stable/glib-Threads.html">GLib
1220 Reference Manual</ulink> for more information about
1221 <function>g_thread_init</function>.
1225 One of the many facilities provided by the <literal>glib</literal>
1226 library is a block allocator, called <literal>g_slice</literal>. You
1227 have to disable this block allocator when using DRD by adding the
1228 following to the shell environment variables:
1229 <literal>G_SLICE=always-malloc</literal>. See also the <ulink
1230 url="http://library.gnome.org/devel/glib/stable/glib-Memory-Slices.html">GLib
1231 Reference Manual</ulink> for more information.
1237 <sect2 id="drd-manual.boost.thread" xreflabel="Boost.Thread">
1238 <title>Debugging Boost.Thread Programs</title>
1241 The Boost.Thread library is the threading library included with the
1242 cross-platform Boost Libraries. This threading library is an early
1243 implementation of the upcoming C++0x threading library.
1247 Applications that use the Boost.Thread library should run fine under DRD.
1251 More information about Boost.Thread can be found here:
1255 Anthony Williams, <ulink
1256 url="http://www.boost.org/doc/libs/1_37_0/doc/html/thread.html">Boost.Thread</ulink>
1257 Library Documentation, Boost website, 2007.
1262 Anthony Williams, <ulink
1263 url="http://www.ddj.com/cpp/211600441">What's New in Boost
1264 Threads?</ulink>, Recent changes to the Boost Thread library,
1265 Dr. Dobbs Magazine, October 2008.
1274 <sect2 id="drd-manual.openmp" xreflabel="OpenMP">
1275 <title>Debugging OpenMP Programs</title>
1278 OpenMP stands for <emphasis>Open Multi-Processing</emphasis>. The OpenMP
1279 standard consists of a set of compiler directives for C, C++ and Fortran
1280 programs that allows a compiler to transform a sequential program into a
1281 parallel program. OpenMP is well suited for HPC applications and allows to
1282 work at a higher level compared to direct use of the POSIX threads API. While
1283 OpenMP ensures that the POSIX API is used correctly, OpenMP programs can still
1284 contain data races. So it definitely makes sense to verify OpenMP programs
1285 with a thread checking tool.
1289 DRD supports OpenMP shared-memory programs generated by GCC. GCC
1290 supports OpenMP since version 4.2.0. GCC's runtime support
1291 for OpenMP programs is provided by a library called
1292 <literal>libgomp</literal>. The synchronization primitives implemented
1293 in this library use Linux' futex system call directly, unless the
1294 library has been configured with the
1295 <literal>--disable-linux-futex</literal> option. DRD only supports
1296 libgomp libraries that have been configured with this option and in
1297 which symbol information is present. For most Linux distributions this
1298 means that you will have to recompile GCC. See also the script
1299 <literal>drd/scripts/download-and-build-gcc</literal> in the
1300 Valgrind source tree for an example of how to compile GCC. You will
1301 also have to make sure that the newly compiled
1302 <literal>libgomp.so</literal> library is loaded when OpenMP programs
1303 are started. This is possible by adding a line similar to the
1304 following to your shell startup script:
1306 <programlisting><![CDATA[
1307 export LD_LIBRARY_PATH=~/gcc-4.4.0/lib64:~/gcc-4.4.0/lib:
1308 ]]></programlisting>
1311 As an example, the test OpenMP test program
1312 <literal>drd/tests/omp_matinv</literal> triggers a data race
1313 when the option -r has been specified on the command line. The data
1314 race is triggered by the following code:
1316 <programlisting><![CDATA[
1317 #pragma omp parallel for private(j)
1318 for (j = 0; j < rows; j++)
1322 const elem_t factor = a[j * cols + i];
1323 for (k = 0; k < cols; k++)
1325 a[j * cols + k] -= a[i * cols + k] * factor;
1329 ]]></programlisting>
1332 The above code is racy because the variable <literal>k</literal> has
1333 not been declared private. DRD will print the following error message
1336 <programlisting><![CDATA[
1337 $ valgrind --tool=drd --check-stack-var=yes --read-var-info=yes drd/tests/omp_matinv 3 -t 2 -r
1339 Conflicting store by thread 1/1 at 0x7fefffbc4 size 4
1340 at 0x4014A0: gj.omp_fn.0 (omp_matinv.c:203)
1341 by 0x401211: gj (omp_matinv.c:159)
1342 by 0x40166A: invert_matrix (omp_matinv.c:238)
1343 by 0x4019B4: main (omp_matinv.c:316)
1344 Location 0x7fefffbc4 is 0 bytes inside local var "k"
1345 declared at omp_matinv.c:160, in frame #0 of thread 1
1347 ]]></programlisting>
1349 In the above output the function name <function>gj.omp_fn.0</function>
1350 has been generated by GCC from the function name
1351 <function>gj</function>. The allocation context information shows that the
1352 data race has been caused by modifying the variable <literal>k</literal>.
1356 Note: for GCC versions before 4.4.0, no allocation context information is
1357 shown. With these GCC versions the most usable information in the above output
1358 is the source file name and the line number where the data race has been
1359 detected (<literal>omp_matinv.c:203</literal>).
1363 For more information about OpenMP, see also
1364 <ulink url="http://openmp.org/">openmp.org</ulink>.
1370 <sect2 id="drd-manual.cust-mem-alloc" xreflabel="Custom Memory Allocators">
1371 <title>DRD and Custom Memory Allocators</title>
1374 DRD tracks all memory allocation events that happen via the
1375 standard memory allocation and deallocation functions
1376 (<function>malloc</function>, <function>free</function>,
1377 <function>new</function> and <function>delete</function>), via entry
1378 and exit of stack frames or that have been annotated with Valgrind's
1379 memory pool client requests. DRD uses memory allocation and deallocation
1380 information for two purposes:
1384 To know where the scope ends of POSIX objects that have not been
1385 destroyed explicitly. It is e.g. not required by the POSIX
1386 threads standard to call
1387 <function>pthread_mutex_destroy</function> before freeing the
1388 memory in which a mutex object resides.
1393 To know where the scope of variables ends. If e.g. heap memory
1394 has been used by one thread, that thread frees that memory, and
1395 another thread allocates and starts using that memory, no data
1396 races must be reported for that memory.
1403 It is essential for correct operation of DRD that the tool knows about
1404 memory allocation and deallocation events. When analyzing a client program
1405 with DRD that uses a custom memory allocator, either instrument the custom
1406 memory allocator with the <literal>VALGRIND_MALLOCLIKE_BLOCK</literal>
1407 and <literal>VALGRIND_FREELIKE_BLOCK</literal> macros or disable the
1408 custom memory allocator.
1412 As an example, the GNU libstdc++ library can be configured
1413 to use standard memory allocation functions instead of memory pools by
1414 setting the environment variable
1415 <literal>GLIBCXX_FORCE_NEW</literal>. For more information, see also
1417 url="http://gcc.gnu.org/onlinedocs/libstdc++/manual/bk01pt04ch11.html">libstdc++
1424 <sect2 id="drd-manual.drd-versus-memcheck" xreflabel="DRD Versus Memcheck">
1425 <title>DRD Versus Memcheck</title>
1428 It is essential for correct operation of DRD that there are no memory
1429 errors such as dangling pointers in the client program. Which means that
1430 it is a good idea to make sure that your program is Memcheck-clean
1431 before you analyze it with DRD. It is possible however that some of
1432 the Memcheck reports are caused by data races. In this case it makes
1433 sense to run DRD before Memcheck.
1437 So which tool should be run first? In case both DRD and Memcheck
1438 complain about a program, a possible approach is to run both tools
1439 alternatingly and to fix as many errors as possible after each run of
1440 each tool until none of the two tools prints any more error messages.
1446 <sect2 id="drd-manual.resource-requirements" xreflabel="Resource Requirements">
1447 <title>Resource Requirements</title>
1450 The requirements of DRD with regard to heap and stack memory and the
1451 effect on the execution time of client programs are as follows:
1455 When running a program under DRD with default DRD options,
1456 between 1.1 and 3.6 times more memory will be needed compared to
1457 a native run of the client program. More memory will be needed
1458 if loading debug information has been enabled
1459 (<literal>--read-var-info=yes</literal>).
1464 DRD allocates some of its temporary data structures on the stack
1465 of the client program threads. This amount of data is limited to
1466 1 - 2 KB. Make sure that thread stacks are sufficiently large.
1471 Most applications will run between 20 and 50 times slower under
1472 DRD than a native single-threaded run. The slowdown will be most
1473 noticeable for applications which perform frequent mutex lock /
1483 <sect2 id="drd-manual.effective-use" xreflabel="Effective Use">
1484 <title>Hints and Tips for Effective Use of DRD</title>
1487 The following information may be helpful when using DRD:
1491 Make sure that debug information is present in the executable
1492 being analyzed, such that DRD can print function name and line
1493 number information in stack traces. Most compilers can be told
1494 to include debug information via compiler option
1495 <option>-g</option>.
1500 Compile with option <option>-O1</option> instead of
1501 <option>-O0</option>. This will reduce the amount of generated
1502 code, may reduce the amount of debug info and will speed up
1503 DRD's processing of the client program. For more information,
1504 see also <xref linkend="manual-core.started"/>.
1509 If DRD reports any errors on libraries that are part of your
1510 Linux distribution like e.g. <literal>libc.so</literal> or
1511 <literal>libstdc++.so</literal>, installing the debug packages
1512 for these libraries will make the output of DRD a lot more
1518 When using C++, do not send output from more than one thread to
1519 <literal>std::cout</literal>. Doing so would not only
1520 generate multiple data race reports, it could also result in
1521 output from several threads getting mixed up. Either use
1522 <function>printf</function> or do the following:
1525 <para>Derive a class from <literal>std::ostreambuf</literal>
1526 and let that class send output line by line to
1527 <literal>stdout</literal>. This will avoid that individual
1528 lines of text produced by different threads get mixed
1532 <para>Create one instance of <literal>std::ostream</literal>
1533 for each thread. This makes stream formatting settings
1534 thread-local. Pass a per-thread instance of the class
1535 derived from <literal>std::ostreambuf</literal> to the
1536 constructor of each instance. </para>
1539 <para>Let each thread send its output to its own instance of
1540 <literal>std::ostream</literal> instead of
1541 <literal>std::cout</literal>.</para>
1555 <sect1 id="drd-manual.Pthreads" xreflabel="Pthreads">
1556 <title>Using the POSIX Threads API Effectively</title>
1558 <sect2 id="drd-manual.mutex-types" xreflabel="mutex-types">
1559 <title>Mutex types</title>
1562 The Single UNIX Specification version two defines the following four
1563 mutex types (see also the documentation of <ulink
1564 url="http://www.opengroup.org/onlinepubs/007908799/xsh/pthread_mutexattr_settype.html"><function>pthread_mutexattr_settype</function></ulink>):
1568 <emphasis>normal</emphasis>, which means that no error checking
1569 is performed, and that the mutex is non-recursive.
1574 <emphasis>error checking</emphasis>, which means that the mutex
1575 is non-recursive and that error checking is performed.
1580 <emphasis>recursive</emphasis>, which means that a mutex may be
1586 <emphasis>default</emphasis>, which means that error checking
1587 behavior is undefined, and that the behavior for recursive
1588 locking is also undefined. Or: portable code must neither
1589 trigger error conditions through the Pthreads API nor attempt to
1590 lock a mutex of default type recursively.
1597 In complex applications it is not always clear from beforehand which
1598 mutex will be locked recursively and which mutex will not be locked
1599 recursively. Attempts lock a non-recursive mutex recursively will
1600 result in race conditions that are very hard to find without a thread
1601 checking tool. So either use the error checking mutex type and
1602 consistently check the return value of Pthread API mutex calls, or use
1603 the recursive mutex type.
1608 <sect2 id="drd-manual.condvar" xreflabel="condition-variables">
1609 <title>Condition variables</title>
1612 A condition variable allows one thread to wake up one or more other
1613 threads. Condition variables are often used to notify one or more
1614 threads about state changes of shared data. Unfortunately it is very
1615 easy to introduce race conditions by using condition variables as the
1616 only means of state information propagation. A better approach is to
1617 let threads poll for changes of a state variable that is protected by
1618 a mutex, and to use condition variables only as a thread wakeup
1619 mechanism. See also the source file
1620 <computeroutput>drd/tests/monitor_example.cpp</computeroutput> for an
1621 example of how to implement this concept in C++. The monitor concept
1622 used in this example is a well known and very useful concept -- see
1623 also Wikipedia for more information about the <ulink
1624 url="http://en.wikipedia.org/wiki/Monitor_(synchronization)">monitor</ulink>
1630 <sect2 id="drd-manual.pctw" xreflabel="pthread_cond_timedwait">
1631 <title>pthread_cond_timedwait and timeouts</title>
1634 Historically the function
1635 <function>pthread_cond_timedwait</function> only allowed the
1636 specification of an absolute timeout, that is a timeout independent of
1637 the time when this function was called. However, almost every call to
1638 this function expresses a relative timeout. This typically happens by
1640 <computeroutput>clock_gettime(CLOCK_REALTIME)</computeroutput> and a
1641 relative timeout as the third argument. This approach is incorrect
1642 since forward or backward clock adjustments by e.g. ntpd will affect
1643 the timeout. A more reliable approach is as follows:
1647 When initializing a condition variable through
1648 <function>pthread_cond_init</function>, specify that the timeout of
1649 <function>pthread_cond_timedwait</function> will use the clock
1650 <literal>CLOCK_MONOTONIC</literal> instead of
1651 <literal>CLOCK_REALTIME</literal>. You can do this via
1652 <computeroutput>pthread_condattr_setclock(...,
1653 CLOCK_MONOTONIC)</computeroutput>.
1658 When calling <function>pthread_cond_timedwait</function>, pass
1660 <computeroutput>clock_gettime(CLOCK_MONOTONIC)</computeroutput>
1661 and a relative timeout as the third argument.
1666 <computeroutput>drd/tests/monitor_example.cpp</computeroutput> for an
1675 <sect1 id="drd-manual.limitations" xreflabel="Limitations">
1676 <title>Limitations</title>
1678 <para>DRD currently has the following limitations:</para>
1683 DRD, just like Memcheck, will refuse to start on Linux
1684 distributions where all symbol information has been removed from
1685 <filename>ld.so</filename>. This is e.g. the case for the PPC editions
1686 of openSUSE and Gentoo. You will have to install the glibc debuginfo
1687 package on these platforms before you can use DRD. See also openSUSE
1688 bug <ulink url="http://bugzilla.novell.com/show_bug.cgi?id=396197">
1689 396197</ulink> and Gentoo bug <ulink
1690 url="http://bugs.gentoo.org/214065">214065</ulink>.
1695 With gcc 4.4.3 and before, DRD may report data races on the C++
1696 class <literal>std::string</literal> in a multithreaded program. This is
1697 a know <literal>libstdc++</literal> issue -- see also GCC bug
1698 <ulink url="http://gcc.gnu.org/bugzilla/show_bug.cgi?id=40518">40518</ulink>
1699 for more information.
1704 When address tracing is enabled, no information on atomic stores
1710 If you compile the DRD source code yourself, you need GCC 3.0 or
1711 later. GCC 2.95 is not supported.
1716 Of the two POSIX threads implementations for Linux, only the
1717 NPTL (Native POSIX Thread Library) is supported. The older
1718 LinuxThreads library is not supported.
1726 <sect1 id="drd-manual.feedback" xreflabel="Feedback">
1727 <title>Feedback</title>
1730 If you have any comments, suggestions, feedback or bug reports about
1731 DRD, feel free to either post a message on the Valgrind users mailing
1732 list or to file a bug report. See also <ulink
1733 url="&vg-url;">&vg-url;</ulink> for more information.