8. DRD: a thread error detector

Table of Contents

8.1. Overview
8.1.1. Multithreaded Programming Paradigms
8.1.2. POSIX Threads Programming Model
8.1.3. Multithreaded Programming Problems
8.1.4. Data Race Detection
8.2. Using DRD
8.2.1. DRD Command-line Options
8.2.2. Detected Errors: Data Races
8.2.3. Detected Errors: Lock Contention
8.2.4. Detected Errors: Misuse of the POSIX threads API
8.2.5. Client Requests
8.2.6. Debugging C++11 Programs
8.2.7. Debugging GNOME Programs
8.2.8. Debugging Boost.Thread Programs
8.2.9. Debugging OpenMP Programs
8.2.10. DRD and Custom Memory Allocators
8.2.11. DRD Versus Memcheck
8.2.12. Resource Requirements
8.2.13. Hints and Tips for Effective Use of DRD
8.3. Using the POSIX Threads API Effectively
8.3.1. Mutex types
8.3.2. Condition variables
8.3.3. pthread_cond_timedwait and timeouts
8.4. Limitations
8.5. Feedback

To use this tool, you must specify --tool=drd on the Valgrind command line.

8.1. Overview

DRD is a Valgrind tool for detecting errors in multithreaded C and C++ programs. The tool works for any program that uses the POSIX threading primitives or that uses threading concepts built on top of the POSIX threading primitives.

8.1.1. Multithreaded Programming Paradigms

There are two possible reasons for using multithreading in a program:

  • To model concurrent activities. Assigning one thread to each activity can be a great simplification compared to multiplexing the states of multiple activities in a single thread. This is why most server software and embedded software is multithreaded.

  • To use multiple CPU cores simultaneously for speeding up computations. This is why many High Performance Computing (HPC) applications are multithreaded.

Multithreaded programs can use one or more of the following programming paradigms. Which paradigm is appropriate depends e.g. on the application type. Some examples of multithreaded programming paradigms are:

  • Locking. Data that is shared over threads is protected from concurrent accesses via locking. E.g. the POSIX threads library, the Qt library and the Boost.Thread library support this paradigm directly.

  • Message passing. No data is shared between threads, but threads exchange data by passing messages to each other. Examples of implementations of the message passing paradigm are MPI and CORBA.

  • Automatic parallelization. A compiler converts a sequential program into a multithreaded program. The original program may or may not contain parallelization hints. One example of such parallelization hints is the OpenMP standard. In this standard a set of directives are defined which tell a compiler how to parallelize a C, C++ or Fortran program. OpenMP is well suited for computational intensive applications. As an example, an open source image processing software package is using OpenMP to maximize performance on systems with multiple CPU cores. GCC supports the OpenMP standard from version 4.2.0 on.

  • Software Transactional Memory (STM). Any data that is shared between threads is updated via transactions. After each transaction it is verified whether there were any conflicting transactions. If there were conflicts, the transaction is aborted, otherwise it is committed. This is a so-called optimistic approach. There is a prototype of the Intel C++ Compiler available that supports STM. Research about the addition of STM support to GCC is ongoing.

DRD supports any combination of multithreaded programming paradigms as long as the implementation of these paradigms is based on the POSIX threads primitives. DRD however does not support programs that use e.g. Linux' futexes directly. Attempts to analyze such programs with DRD will cause DRD to report many false positives.

8.1.2. POSIX Threads Programming Model

POSIX threads, also known as Pthreads, is the most widely available threading library on Unix systems.

The POSIX threads programming model is based on the following abstractions:

  • A shared address space. All threads running within the same process share the same address space. All data, whether shared or not, is identified by its address.

  • Regular load and store operations, which allow to read values from or to write values to the memory shared by all threads running in the same process.

  • Atomic store and load-modify-store operations. While these are not mentioned in the POSIX threads standard, most microprocessors support atomic memory operations.

  • Threads. Each thread represents a concurrent activity.

  • Synchronization objects and operations on these synchronization objects. The following types of synchronization objects have been defined in the POSIX threads standard: mutexes, condition variables, semaphores, reader-writer synchronization objects, barriers and spinlocks.

Which source code statements generate which memory accesses depends on the memory model of the programming language being used. There is not yet a definitive memory model for the C and C++ languages. For a draft memory model, see also the document WG21/N2338: Concurrency memory model compiler consequences.

For more information about POSIX threads, see also the Single UNIX Specification version 3, also known as IEEE Std 1003.1.

8.1.3. Multithreaded Programming Problems

Depending on which multithreading paradigm is being used in a program, one or more of the following problems can occur:

  • Data races. One or more threads access the same memory location without sufficient locking. Most but not all data races are programming errors and are the cause of subtle and hard-to-find bugs.

  • Lock contention. One thread blocks the progress of one or more other threads by holding a lock too long.

  • Improper use of the POSIX threads API. Most implementations of the POSIX threads API have been optimized for runtime speed. Such implementations will not complain on certain errors, e.g. when a mutex is being unlocked by another thread than the thread that obtained a lock on the mutex.

  • Deadlock. A deadlock occurs when two or more threads wait for each other indefinitely.

  • False sharing. If threads that run on different processor cores access different variables located in the same cache line frequently, this will slow down the involved threads a lot due to frequent exchange of cache lines.

Although the likelihood of the occurrence of data races can be reduced through a disciplined programming style, a tool for automatic detection of data races is a necessity when developing multithreaded software. DRD can detect these, as well as lock contention and improper use of the POSIX threads API.

8.1.4. Data Race Detection

The result of load and store operations performed by a multithreaded program depends on the order in which memory operations are performed. This order is determined by:

  1. All memory operations performed by the same thread are performed in program order, that is, the order determined by the program source code and the results of previous load operations.

  2. Synchronization operations determine certain ordering constraints on memory operations performed by different threads. These ordering constraints are called the synchronization order.

The combination of program order and synchronization order is called the happens-before relationship. This concept was first defined by S. Adve et al in the paper Detecting data races on weak memory systems, ACM SIGARCH Computer Architecture News, v.19 n.3, p.234-243, May 1991.

Two memory operations conflict if both operations are performed by different threads, refer to the same memory location and at least one of them is a store operation.

A multithreaded program is data-race free if all conflicting memory accesses are ordered by synchronization operations.

A well known way to ensure that a multithreaded program is data-race free is to ensure that a locking discipline is followed. It is e.g. possible to associate a mutex with each shared data item, and to hold a lock on the associated mutex while the shared data is accessed.

All programs that follow a locking discipline are data-race free, but not all data-race free programs follow a locking discipline. There exist multithreaded programs where access to shared data is arbitrated via condition variables, semaphores or barriers. As an example, a certain class of HPC applications consists of a sequence of computation steps separated in time by barriers, and where these barriers are the only means of synchronization. Although there are many conflicting memory accesses in such applications and although such applications do not make use mutexes, most of these applications do not contain data races.

There exist two different approaches for verifying the correctness of multithreaded programs at runtime. The approach of the so-called Eraser algorithm is to verify whether all shared memory accesses follow a consistent locking strategy. And the happens-before data race detectors verify directly whether all interthread memory accesses are ordered by synchronization operations. While the last approach is more complex to implement, and while it is more sensitive to OS scheduling, it is a general approach that works for all classes of multithreaded programs. An important advantage of happens-before data race detectors is that these do not report any false positives.

DRD is based on the happens-before algorithm.

8.2. Using DRD

8.2.1. DRD Command-line Options

The following command-line options are available for controlling the behavior of the DRD tool itself:

--check-stack-var=<yes|no> [default: no]

Controls whether DRD detects data races on stack variables. Verifying stack variables is disabled by default because most programs do not share stack variables over threads.

--exclusive-threshold=<n> [default: off]

Print an error message if any mutex or writer lock has been held longer than the time specified in milliseconds. This option enables the detection of lock contention.

--join-list-vol=<n> [default: 10]

Data races that occur between a statement at the end of one thread and another thread can be missed if memory access information is discarded immediately after a thread has been joined. This option allows to specify for how many joined threads memory access information should be retained.

--first-race-only=<yes|no> [default: no]

Whether to report only the first data race that has been detected on a memory location or all data races that have been detected on a memory location.

--free-is-write=<yes|no> [default: no]

Whether to report races between accessing memory and freeing memory. Enabling this option may cause DRD to run slightly slower. Notes:

  • Don't enable this option when using custom memory allocators that use the VG_USERREQ__MALLOCLIKE_BLOCK and VG_USERREQ__FREELIKE_BLOCK because that would result in false positives.

  • Don't enable this option when using reference-counted objects because that will result in false positives, even when that code has been annotated properly with ANNOTATE_HAPPENS_BEFORE and ANNOTATE_HAPPENS_AFTER. See e.g. the output of the following command for an example: valgrind --tool=drd --free-is-write=yes drd/tests/annotate_smart_pointer.

--report-signal-unlocked=<yes|no> [default: yes]

Whether to report calls to pthread_cond_signal and pthread_cond_broadcast where the mutex associated with the signal through pthread_cond_wait or pthread_cond_timed_waitis not locked at the time the signal is sent. Sending a signal without holding a lock on the associated mutex is a common programming error which can cause subtle race conditions and unpredictable behavior. There exist some uncommon synchronization patterns however where it is safe to send a signal without holding a lock on the associated mutex.

--segment-merging=<yes|no> [default: yes]

Controls segment merging. Segment merging is an algorithm to limit memory usage of the data race detection algorithm. Disabling segment merging may improve the accuracy of the so-called 'other segments' displayed in race reports but can also trigger an out of memory error.

--segment-merging-interval=<n> [default: 10]

Perform segment merging only after the specified number of new segments have been created. This is an advanced configuration option that allows to choose whether to minimize DRD's memory usage by choosing a low value or to let DRD run faster by choosing a slightly higher value. The optimal value for this parameter depends on the program being analyzed. The default value works well for most programs.

--shared-threshold=<n> [default: off]

Print an error message if a reader lock has been held longer than the specified time (in milliseconds). This option enables the detection of lock contention.

--show-confl-seg=<yes|no> [default: yes]

Show conflicting segments in race reports. Since this information can help to find the cause of a data race, this option is enabled by default. Disabling this option makes the output of DRD more compact.

--show-stack-usage=<yes|no> [default: no]

Print stack usage at thread exit time. When a program creates a large number of threads it becomes important to limit the amount of virtual memory allocated for thread stacks. This option makes it possible to observe how much stack memory has been used by each thread of the client program. Note: the DRD tool itself allocates some temporary data on the client thread stack. The space necessary for this temporary data must be allocated by the client program when it allocates stack memory, but is not included in stack usage reported by DRD.

--ignore-thread-creation=<yes|no> [default: no]

Controls whether all activities during thread creation should be ignored. By default enabled only on Solaris. Solaris provides higher throughput, parallelism and scalability than other operating systems, at the cost of more fine-grained locking activity. This means for example that when a thread is created under glibc, just one big lock is used for all thread setup. Solaris libc uses several fine-grained locks and the creator thread resumes its activities as soon as possible, leaving for example stack and TLS setup sequence to the created thread. This situation confuses DRD as it assumes there is some false ordering in place between creator and created thread; and therefore many types of race conditions in the application would not be reported. To prevent such false ordering, this command line option is set to yes by default on Solaris. All activity (loads, stores, client requests) is therefore ignored during:

  • pthread_create() call in the creator thread

  • thread creation phase (stack and TLS setup) in the created thread

The following options are available for monitoring the behavior of the client program:

--trace-addr=<address> [default: none]

Trace all load and store activity for the specified address. This option may be specified more than once.

--ptrace-addr=<address> [default: none]

Trace all load and store activity for the specified address and keep doing that even after the memory at that address has been freed and reallocated.

--trace-alloc=<yes|no> [default: no]

Trace all memory allocations and deallocations. May produce a huge amount of output.

--trace-barrier=<yes|no> [default: no]

Trace all barrier activity.

--trace-cond=<yes|no> [default: no]

Trace all condition variable activity.

--trace-fork-join=<yes|no> [default: no]

Trace all thread creation and all thread termination events.

--trace-hb=<yes|no> [default: no]

Trace execution of the ANNOTATE_HAPPENS_BEFORE(), ANNOTATE_HAPPENS_AFTER() and ANNOTATE_HAPPENS_DONE() client requests.

--trace-mutex=<yes|no> [default: no]

Trace all mutex activity.

--trace-rwlock=<yes|no> [default: no]

Trace all reader-writer lock activity.

--trace-semaphore=<yes|no> [default: no]

Trace all semaphore activity.

8.2.2. Detected Errors: Data Races

DRD prints a message every time it detects a data race. Please keep the following in mind when interpreting DRD's output:

  • Every thread is assigned a thread ID by the DRD tool. A thread ID is a number. Thread ID's start at one and are never recycled.

  • The term segment refers to a consecutive sequence of load, store and synchronization operations, all issued by the same thread. A segment always starts and ends at a synchronization operation. Data race analysis is performed between segments instead of between individual load and store operations because of performance reasons.

  • There are always at least two memory accesses involved in a data race. Memory accesses involved in a data race are called conflicting memory accesses. DRD prints a report for each memory access that conflicts with a past memory access.

Below you can find an example of a message printed by DRD when it detects a data race:

$ valgrind --tool=drd --read-var-info=yes drd/tests/rwlock_race
...
==9466== Thread 3:
==9466== Conflicting load by thread 3 at 0x006020b8 size 4
==9466==    at 0x400B6C: thread_func (rwlock_race.c:29)
==9466==    by 0x4C291DF: vg_thread_wrapper (drd_pthread_intercepts.c:186)
==9466==    by 0x4E3403F: start_thread (in /lib64/libpthread-2.8.so)
==9466==    by 0x53250CC: clone (in /lib64/libc-2.8.so)
==9466== Location 0x6020b8 is 0 bytes inside local var "s_racy"
==9466== declared at rwlock_race.c:18, in frame #0 of thread 3
==9466== Other segment start (thread 2)
==9466==    at 0x4C2847D: pthread_rwlock_rdlock* (drd_pthread_intercepts.c:813)
==9466==    by 0x400B6B: thread_func (rwlock_race.c:28)
==9466==    by 0x4C291DF: vg_thread_wrapper (drd_pthread_intercepts.c:186)
==9466==    by 0x4E3403F: start_thread (in /lib64/libpthread-2.8.so)
==9466==    by 0x53250CC: clone (in /lib64/libc-2.8.so)
==9466== Other segment end (thread 2)
==9466==    at 0x4C28B54: pthread_rwlock_unlock* (drd_pthread_intercepts.c:912)
==9466==    by 0x400B84: thread_func (rwlock_race.c:30)
==9466==    by 0x4C291DF: vg_thread_wrapper (drd_pthread_intercepts.c:186)
==9466==    by 0x4E3403F: start_thread (in /lib64/libpthread-2.8.so)
==9466==    by 0x53250CC: clone (in /lib64/libc-2.8.so)
...

The above report has the following meaning:

  • The number in the column on the left is the process ID of the process being analyzed by DRD.

  • The first line ("Thread 3") tells you the thread ID for the thread in which context the data race has been detected.

  • The next line tells which kind of operation was performed (load or store) and by which thread. On the same line the start address and the number of bytes involved in the conflicting access are also displayed.

  • Next, the call stack of the conflicting access is displayed. If your program has been compiled with debug information (-g), this call stack will include file names and line numbers. The two bottommost frames in this call stack (clone and start_thread) show how the NPTL starts a thread. The third frame (vg_thread_wrapper) is added by DRD. The fourth frame (thread_func) is the first interesting line because it shows the thread entry point, that is the function that has been passed as the third argument to pthread_create.

  • Next, the allocation context for the conflicting address is displayed. For dynamically allocated data the allocation call stack is shown. For static variables and stack variables the allocation context is only shown when the option --read-var-info=yes has been specified. Otherwise DRD will print Allocation context: unknown.

  • A conflicting access involves at least two memory accesses. For one of these accesses an exact call stack is displayed, and for the other accesses an approximate call stack is displayed, namely the start and the end of the segments of the other accesses. This information can be interpreted as follows:

    1. Start at the bottom of both call stacks, and count the number stack frames with identical function name, file name and line number. In the above example the three bottommost frames are identical (clone, start_thread and vg_thread_wrapper).

    2. The next higher stack frame in both call stacks now tells you between in which source code region the other memory access happened. The above output tells that the other memory access involved in the data race happened between source code lines 28 and 30 in file rwlock_race.c.

8.2.3. Detected Errors: Lock Contention

Threads must be able to make progress without being blocked for too long by other threads. Sometimes a thread has to wait until a mutex or reader-writer synchronization object is unlocked by another thread. This is called lock contention.

Lock contention causes delays. Such delays should be as short as possible. The two command line options --exclusive-threshold=<n> and --shared-threshold=<n> make it possible to detect excessive lock contention by making DRD report any lock that has been held longer than the specified threshold. An example:

$ valgrind --tool=drd --exclusive-threshold=10 drd/tests/hold_lock -i 500
...
==10668== Acquired at:
==10668==    at 0x4C267C8: pthread_mutex_lock (drd_pthread_intercepts.c:395)
==10668==    by 0x400D92: main (hold_lock.c:51)
==10668== Lock on mutex 0x7fefffd50 was held during 503 ms (threshold: 10 ms).
==10668==    at 0x4C26ADA: pthread_mutex_unlock (drd_pthread_intercepts.c:441)
==10668==    by 0x400DB5: main (hold_lock.c:55)
...

The hold_lock test program holds a lock as long as specified by the -i (interval) argument. The DRD output reports that the lock acquired at line 51 in source file hold_lock.c and released at line 55 was held during 503 ms, while a threshold of 10 ms was specified to DRD.

8.2.4. Detected Errors: Misuse of the POSIX threads API

DRD is able to detect and report the following misuses of the POSIX threads API:

  • Passing the address of one type of synchronization object (e.g. a mutex) to a POSIX API call that expects a pointer to another type of synchronization object (e.g. a condition variable).

  • Attempts to unlock a mutex that has not been locked.

  • Attempts to unlock a mutex that was locked by another thread.

  • Attempts to lock a mutex of type PTHREAD_MUTEX_NORMAL or a spinlock recursively.

  • Destruction or deallocation of a locked mutex.

  • Sending a signal to a condition variable while no lock is held on the mutex associated with the condition variable.

  • Calling pthread_cond_wait on a mutex that is not locked, that is locked by another thread or that has been locked recursively.

  • Associating two different mutexes with a condition variable through pthread_cond_wait.

  • Destruction or deallocation of a condition variable that is being waited upon.

  • Destruction or deallocation of a locked reader-writer synchronization object.

  • Attempts to unlock a reader-writer synchronization object that was not locked by the calling thread.

  • Attempts to recursively lock a reader-writer synchronization object exclusively.

  • Attempts to pass the address of a user-defined reader-writer synchronization object to a POSIX threads function.

  • Attempts to pass the address of a POSIX reader-writer synchronization object to one of the annotations for user-defined reader-writer synchronization objects.

  • Reinitialization of a mutex, condition variable, reader-writer lock, semaphore or barrier.

  • Destruction or deallocation of a semaphore or barrier that is being waited upon.

  • Missing synchronization between barrier wait and barrier destruction.

  • Exiting a thread without first unlocking the spinlocks, mutexes or reader-writer synchronization objects that were locked by that thread.

  • Passing an invalid thread ID to pthread_join or pthread_cancel.

8.2.5. Client Requests

Just as for other Valgrind tools it is possible to let a client program interact with the DRD tool through client requests. In addition to the client requests several macros have been defined that allow to use the client requests in a convenient way.

The interface between client programs and the DRD tool is defined in the header file <valgrind/drd.h>. The available macros and client requests are:

  • The macro DRD_GET_VALGRIND_THREADID and the corresponding client request VG_USERREQ__DRD_GET_VALGRIND_THREAD_ID. Query the thread ID that has been assigned by the Valgrind core to the thread executing this client request. Valgrind's thread ID's start at one and are recycled in case a thread stops.

  • The macro DRD_GET_DRD_THREADID and the corresponding client request VG_USERREQ__DRD_GET_DRD_THREAD_ID. Query the thread ID that has been assigned by DRD to the thread executing this client request. These are the thread ID's reported by DRD in data race reports and in trace messages. DRD's thread ID's start at one and are never recycled.

  • The macros DRD_IGNORE_VAR(x), ANNOTATE_TRACE_MEMORY(&x) and the corresponding client request VG_USERREQ__DRD_START_SUPPRESSION. Some applications contain intentional races. There exist e.g. applications where the same value is assigned to a shared variable from two different threads. It may be more convenient to suppress such races than to solve these. This client request allows to suppress such races.

  • The macro DRD_STOP_IGNORING_VAR(x) and the corresponding client request VG_USERREQ__DRD_FINISH_SUPPRESSION. Tell DRD to no longer ignore data races for the address range that was suppressed either via the macro DRD_IGNORE_VAR(x) or via the client request VG_USERREQ__DRD_START_SUPPRESSION.

  • The macro DRD_TRACE_VAR(x). Trace all load and store activity for the address range starting at &x and occupying sizeof(x) bytes. When DRD reports a data race on a specified variable, and it's not immediately clear which source code statements triggered the conflicting accesses, it can be very helpful to trace all activity on the offending memory location.

  • The macro DRD_STOP_TRACING_VAR(x). Stop tracing load and store activity for the address range starting at &x and occupying sizeof(x) bytes.

  • The macro ANNOTATE_TRACE_MEMORY(&x). Trace all load and store activity that touches at least the single byte at the address &x.

  • The client request VG_USERREQ__DRD_START_TRACE_ADDR, which allows to trace all load and store activity for the specified address range.

  • The client request VG_USERREQ__DRD_STOP_TRACE_ADDR. Do no longer trace load and store activity for the specified address range.

  • The macro ANNOTATE_HAPPENS_BEFORE(addr) tells DRD to insert a mark. Insert this macro just after an access to the variable at the specified address has been performed.

  • The macro ANNOTATE_HAPPENS_AFTER(addr) tells DRD that the next access to the variable at the specified address should be considered to have happened after the access just before the latest ANNOTATE_HAPPENS_BEFORE(addr) annotation that references the same variable. The purpose of these two macros is to tell DRD about the order of inter-thread memory accesses implemented via atomic memory operations. See also drd/tests/annotate_smart_pointer.cpp for an example.

  • The macro ANNOTATE_RWLOCK_CREATE(rwlock) tells DRD that the object at address rwlock is a reader-writer synchronization object that is not a pthread_rwlock_t synchronization object. See also drd/tests/annotate_rwlock.c for an example.

  • The macro ANNOTATE_RWLOCK_DESTROY(rwlock) tells DRD that the reader-writer synchronization object at address rwlock has been destroyed.

  • The macro ANNOTATE_WRITERLOCK_ACQUIRED(rwlock) tells DRD that a writer lock has been acquired on the reader-writer synchronization object at address rwlock.

  • The macro ANNOTATE_READERLOCK_ACQUIRED(rwlock) tells DRD that a reader lock has been acquired on the reader-writer synchronization object at address rwlock.

  • The macro ANNOTATE_RWLOCK_ACQUIRED(rwlock, is_w) tells DRD that a writer lock (when is_w != 0) or that a reader lock (when is_w == 0) has been acquired on the reader-writer synchronization object at address rwlock.

  • The macro ANNOTATE_WRITERLOCK_RELEASED(rwlock) tells DRD that a writer lock has been released on the reader-writer synchronization object at address rwlock.

  • The macro ANNOTATE_READERLOCK_RELEASED(rwlock) tells DRD that a reader lock has been released on the reader-writer synchronization object at address rwlock.

  • The macro ANNOTATE_RWLOCK_RELEASED(rwlock, is_w) tells DRD that a writer lock (when is_w != 0) or that a reader lock (when is_w == 0) has been released on the reader-writer synchronization object at address rwlock.

  • The macro ANNOTATE_BARRIER_INIT(barrier, count, reinitialization_allowed) tells DRD that a new barrier object at the address barrier has been initialized, that count threads participate in each barrier and also whether or not barrier reinitialization without intervening destruction should be reported as an error. See also drd/tests/annotate_barrier.c for an example.

  • The macro ANNOTATE_BARRIER_DESTROY(barrier) tells DRD that a barrier object is about to be destroyed.

  • The macro ANNOTATE_BARRIER_WAIT_BEFORE(barrier) tells DRD that waiting for a barrier will start.

  • The macro ANNOTATE_BARRIER_WAIT_AFTER(barrier) tells DRD that waiting for a barrier has finished.

  • The macro ANNOTATE_BENIGN_RACE_SIZED(addr, size, descr) tells DRD that any races detected on the specified address are benign and hence should not be reported. The descr argument is ignored but can be used to document why data races on addr are benign.

  • The macro ANNOTATE_BENIGN_RACE_STATIC(var, descr) tells DRD that any races detected on the specified static variable are benign and hence should not be reported. The descr argument is ignored but can be used to document why data races on var are benign. Note: this macro can only be used in C++ programs and not in C programs.

  • The macro ANNOTATE_IGNORE_READS_BEGIN tells DRD to ignore all memory loads performed by the current thread.

  • The macro ANNOTATE_IGNORE_READS_END tells DRD to stop ignoring the memory loads performed by the current thread.

  • The macro ANNOTATE_IGNORE_WRITES_BEGIN tells DRD to ignore all memory stores performed by the current thread.

  • The macro ANNOTATE_IGNORE_WRITES_END tells DRD to stop ignoring the memory stores performed by the current thread.

  • The macro ANNOTATE_IGNORE_READS_AND_WRITES_BEGIN tells DRD to ignore all memory accesses performed by the current thread.

  • The macro ANNOTATE_IGNORE_READS_AND_WRITES_END tells DRD to stop ignoring the memory accesses performed by the current thread.

  • The macro ANNOTATE_NEW_MEMORY(addr, size) tells DRD that the specified memory range has been allocated by a custom memory allocator in the client program and that the client program will start using this memory range.

  • The macro ANNOTATE_THREAD_NAME(name) tells DRD to associate the specified name with the current thread and to include this name in the error messages printed by DRD.

  • The macros VALGRIND_MALLOCLIKE_BLOCK and VALGRIND_FREELIKE_BLOCK from the Valgrind core are implemented; they are described in The Client Request mechanism.

Note: if you compiled Valgrind yourself, the header file <valgrind/drd.h> will have been installed in the directory /usr/include by the command make install. If you obtained Valgrind by installing it as a package however, you will probably have to install another package with a name like valgrind-devel before Valgrind's header files are available.

8.2.6. Debugging C++11 Programs

If you want to use the C++11 class std::thread you will need to do the following to annotate the std::shared_ptr<> objects used in the implementation of that class:

  • Add the following code at the start of a common header or at the start of each source file, before any C++ header files are included:

    #include <valgrind/drd.h>
    #define _GLIBCXX_SYNCHRONIZATION_HAPPENS_BEFORE(addr) ANNOTATE_HAPPENS_BEFORE(addr)
    #define _GLIBCXX_SYNCHRONIZATION_HAPPENS_AFTER(addr)  ANNOTATE_HAPPENS_AFTER(addr)
    
  • Download the gcc source code and from source file libstdc++-v3/src/c++11/thread.cc copy the implementation of the execute_native_thread_routine() and std::thread::_M_start_thread() functions into a source file that is linked with your application. Make sure that also in this source file the _GLIBCXX_SYNCHRONIZATION_HAPPENS_*() macros are defined properly.

For more information, see also The GNU C++ Library Manual, Debugging Support (http://gcc.gnu.org/onlinedocs/libstdc++/manual/debug.html).

8.2.7. Debugging GNOME Programs

GNOME applications use the threading primitives provided by the glib and gthread libraries. These libraries are built on top of POSIX threads, and hence are directly supported by DRD. Please keep in mind that you have to call g_thread_init before creating any threads, or DRD will report several data races on glib functions. See also the GLib Reference Manual for more information about g_thread_init.

One of the many facilities provided by the glib library is a block allocator, called g_slice. You have to disable this block allocator when using DRD by adding the following to the shell environment variables: G_SLICE=always-malloc. See also the GLib Reference Manual for more information.

8.2.8. Debugging Boost.Thread Programs

The Boost.Thread library is the threading library included with the cross-platform Boost Libraries. This threading library is an early implementation of the upcoming C++0x threading library.

Applications that use the Boost.Thread library should run fine under DRD.

More information about Boost.Thread can be found here:

  • Anthony Williams, Boost.Thread Library Documentation, Boost website, 2007.

  • Anthony Williams, What's New in Boost Threads?, Recent changes to the Boost Thread library, Dr. Dobbs Magazine, October 2008.

8.2.9. Debugging OpenMP Programs

OpenMP stands for Open Multi-Processing. The OpenMP standard consists of a set of compiler directives for C, C++ and Fortran programs that allows a compiler to transform a sequential program into a parallel program. OpenMP is well suited for HPC applications and allows to work at a higher level compared to direct use of the POSIX threads API. While OpenMP ensures that the POSIX API is used correctly, OpenMP programs can still contain data races. So it definitely makes sense to verify OpenMP programs with a thread checking tool.

DRD supports OpenMP shared-memory programs generated by GCC. GCC supports OpenMP since version 4.2.0. GCC's runtime support for OpenMP programs is provided by a library called libgomp. The synchronization primitives implemented in this library use Linux' futex system call directly, unless the library has been configured with the --disable-linux-futex option. DRD only supports libgomp libraries that have been configured with this option and in which symbol information is present. For most Linux distributions this means that you will have to recompile GCC. See also the script drd/scripts/download-and-build-gcc in the Valgrind source tree for an example of how to compile GCC. You will also have to make sure that the newly compiled libgomp.so library is loaded when OpenMP programs are started. This is possible by adding a line similar to the following to your shell startup script:

export LD_LIBRARY_PATH=~/gcc-4.4.0/lib64:~/gcc-4.4.0/lib:

As an example, the test OpenMP test program drd/tests/omp_matinv triggers a data race when the option -r has been specified on the command line. The data race is triggered by the following code:

#pragma omp parallel for private(j)
for (j = 0; j < rows; j++)
{
  if (i != j)
  {
    const elem_t factor = a[j * cols + i];
    for (k = 0; k < cols; k++)
    {
      a[j * cols + k] -= a[i * cols + k] * factor;
    }
  }
}

The above code is racy because the variable k has not been declared private. DRD will print the following error message for the above code:

$ valgrind --tool=drd --check-stack-var=yes --read-var-info=yes drd/tests/omp_matinv 3 -t 2 -r
...
Conflicting store by thread 1/1 at 0x7fefffbc4 size 4
   at 0x4014A0: gj.omp_fn.0 (omp_matinv.c:203)
   by 0x401211: gj (omp_matinv.c:159)
   by 0x40166A: invert_matrix (omp_matinv.c:238)
   by 0x4019B4: main (omp_matinv.c:316)
Location 0x7fefffbc4 is 0 bytes inside local var "k"
declared at omp_matinv.c:160, in frame #0 of thread 1
...

In the above output the function name gj.omp_fn.0 has been generated by GCC from the function name gj. The allocation context information shows that the data race has been caused by modifying the variable k.

Note: for GCC versions before 4.4.0, no allocation context information is shown. With these GCC versions the most usable information in the above output is the source file name and the line number where the data race has been detected (omp_matinv.c:203).

For more information about OpenMP, see also openmp.org.

8.2.10. DRD and Custom Memory Allocators

DRD tracks all memory allocation events that happen via the standard memory allocation and deallocation functions (malloc, free, new and delete), via entry and exit of stack frames or that have been annotated with Valgrind's memory pool client requests. DRD uses memory allocation and deallocation information for two purposes:

  • To know where the scope ends of POSIX objects that have not been destroyed explicitly. It is e.g. not required by the POSIX threads standard to call pthread_mutex_destroy before freeing the memory in which a mutex object resides.

  • To know where the scope of variables ends. If e.g. heap memory has been used by one thread, that thread frees that memory, and another thread allocates and starts using that memory, no data races must be reported for that memory.

It is essential for correct operation of DRD that the tool knows about memory allocation and deallocation events. When analyzing a client program with DRD that uses a custom memory allocator, either instrument the custom memory allocator with the VALGRIND_MALLOCLIKE_BLOCK and VALGRIND_FREELIKE_BLOCK macros or disable the custom memory allocator.

As an example, the GNU libstdc++ library can be configured to use standard memory allocation functions instead of memory pools by setting the environment variable GLIBCXX_FORCE_NEW. For more information, see also the libstdc++ manual.

8.2.11. DRD Versus Memcheck

It is essential for correct operation of DRD that there are no memory errors such as dangling pointers in the client program. Which means that it is a good idea to make sure that your program is Memcheck-clean before you analyze it with DRD. It is possible however that some of the Memcheck reports are caused by data races. In this case it makes sense to run DRD before Memcheck.

So which tool should be run first? In case both DRD and Memcheck complain about a program, a possible approach is to run both tools alternatingly and to fix as many errors as possible after each run of each tool until none of the two tools prints any more error messages.

8.2.12. Resource Requirements

The requirements of DRD with regard to heap and stack memory and the effect on the execution time of client programs are as follows:

  • When running a program under DRD with default DRD options, between 1.1 and 3.6 times more memory will be needed compared to a native run of the client program. More memory will be needed if loading debug information has been enabled (--read-var-info=yes).

  • DRD allocates some of its temporary data structures on the stack of the client program threads. This amount of data is limited to 1 - 2 KB. Make sure that thread stacks are sufficiently large.

  • Most applications will run between 20 and 50 times slower under DRD than a native single-threaded run. The slowdown will be most noticeable for applications which perform frequent mutex lock / unlock operations.

8.2.13. Hints and Tips for Effective Use of DRD

The following information may be helpful when using DRD:

  • Make sure that debug information is present in the executable being analyzed, such that DRD can print function name and line number information in stack traces. Most compilers can be told to include debug information via compiler option -g.

  • Compile with option -O1 instead of -O0. This will reduce the amount of generated code, may reduce the amount of debug info and will speed up DRD's processing of the client program. For more information, see also Getting started.

  • If DRD reports any errors on libraries that are part of your Linux distribution like e.g. libc.so or libstdc++.so, installing the debug packages for these libraries will make the output of DRD a lot more detailed.

  • When using C++, do not send output from more than one thread to std::cout. Doing so would not only generate multiple data race reports, it could also result in output from several threads getting mixed up. Either use printf or do the following:

    1. Derive a class from std::ostreambuf and let that class send output line by line to stdout. This will avoid that individual lines of text produced by different threads get mixed up.

    2. Create one instance of std::ostream for each thread. This makes stream formatting settings thread-local. Pass a per-thread instance of the class derived from std::ostreambuf to the constructor of each instance.

    3. Let each thread send its output to its own instance of std::ostream instead of std::cout.

8.3. Using the POSIX Threads API Effectively

8.3.1. Mutex types

The Single UNIX Specification version two defines the following four mutex types (see also the documentation of pthread_mutexattr_settype):

  • normal, which means that no error checking is performed, and that the mutex is non-recursive.

  • error checking, which means that the mutex is non-recursive and that error checking is performed.

  • recursive, which means that a mutex may be locked recursively.

  • default, which means that error checking behavior is undefined, and that the behavior for recursive locking is also undefined. Or: portable code must neither trigger error conditions through the Pthreads API nor attempt to lock a mutex of default type recursively.

In complex applications it is not always clear from beforehand which mutex will be locked recursively and which mutex will not be locked recursively. Attempts lock a non-recursive mutex recursively will result in race conditions that are very hard to find without a thread checking tool. So either use the error checking mutex type and consistently check the return value of Pthread API mutex calls, or use the recursive mutex type.

8.3.2. Condition variables

A condition variable allows one thread to wake up one or more other threads. Condition variables are often used to notify one or more threads about state changes of shared data. Unfortunately it is very easy to introduce race conditions by using condition variables as the only means of state information propagation. A better approach is to let threads poll for changes of a state variable that is protected by a mutex, and to use condition variables only as a thread wakeup mechanism. See also the source file drd/tests/monitor_example.cpp for an example of how to implement this concept in C++. The monitor concept used in this example is a well known and very useful concept -- see also Wikipedia for more information about the monitor concept.

8.3.3. pthread_cond_timedwait and timeouts

Historically the function pthread_cond_timedwait only allowed the specification of an absolute timeout, that is a timeout independent of the time when this function was called. However, almost every call to this function expresses a relative timeout. This typically happens by passing the sum of clock_gettime(CLOCK_REALTIME) and a relative timeout as the third argument. This approach is incorrect since forward or backward clock adjustments by e.g. ntpd will affect the timeout. A more reliable approach is as follows:

  • When initializing a condition variable through pthread_cond_init, specify that the timeout of pthread_cond_timedwait will use the clock CLOCK_MONOTONIC instead of CLOCK_REALTIME. You can do this via pthread_condattr_setclock(..., CLOCK_MONOTONIC).

  • When calling pthread_cond_timedwait, pass the sum of clock_gettime(CLOCK_MONOTONIC) and a relative timeout as the third argument.

See also drd/tests/monitor_example.cpp for an example.

8.4. Limitations

DRD currently has the following limitations:

  • DRD, just like Memcheck, will refuse to start on Linux distributions where all symbol information has been removed from ld.so. This is e.g. the case for the PPC editions of openSUSE and Gentoo. You will have to install the glibc debuginfo package on these platforms before you can use DRD. See also openSUSE bug 396197 and Gentoo bug 214065.

  • With gcc 4.4.3 and before, DRD may report data races on the C++ class std::string in a multithreaded program. This is a know libstdc++ issue -- see also GCC bug 40518 for more information.

  • If you compile the DRD source code yourself, you need GCC 3.0 or later. GCC 2.95 is not supported.

  • Of the two POSIX threads implementations for Linux, only the NPTL (Native POSIX Thread Library) is supported. The older LinuxThreads library is not supported.

8.5. Feedback

If you have any comments, suggestions, feedback or bug reports about DRD, feel free to either post a message on the Valgrind users mailing list or to file a bug report. See also http://www.valgrind.org/ for more information.