			Valgrind User Manual

2. Using and understanding the Valgrind core

This chapter describes the Valgrind core services, command-line options and behaviours. That means it is relevant regardless of what particular tool you are using. The information should be sufficient for you to make effective day-to-day use of Valgrind. Advanced topics related to the Valgrind core are described in Valgrind's core: advanced topics.

A point of terminology: most references to "Valgrind" in this chapter refer to the Valgrind core services.

2.1. What Valgrind does with your program

Valgrind is designed to be as non-intrusive as possible. It works directly with existing executables. You don't need to recompile, relink, or otherwise modify the program to be checked.

You invoke Valgrind like this:

valgrind [valgrind-options] your-prog [your-prog-options]

The most important option is --tool which dictates which Valgrind tool to run. For example, if want to run the command ls -l using the memory-checking tool Memcheck, issue this command:

valgrind --tool=memcheck ls -l

However, Memcheck is the default, so if you want to use it you can omit the --tool option.

Regardless of which tool is in use, Valgrind takes control of your program before it starts. Debugging information is read from the executable and associated libraries, so that error messages and other outputs can be phrased in terms of source code locations, when appropriate.

Your program is then run on a synthetic CPU provided by the Valgrind core. As new code is executed for the first time, the core hands the code to the selected tool. The tool adds its own instrumentation code to this and hands the result back to the core, which coordinates the continued execution of this instrumented code.

The amount of instrumentation code added varies widely between tools. At one end of the scale, Memcheck adds code to check every memory access and every value computed, making it run 10-50 times slower than natively. At the other end of the spectrum, the minimal tool, called Nulgrind, adds no instrumentation at all and causes in total "only" about a 4 times slowdown.

Valgrind simulates every single instruction your program executes. Because of this, the active tool checks, or profiles, not only the code in your application but also in all supporting dynamically-linked libraries, including the C library, graphical libraries, and so on.

If you're using an error-detection tool, Valgrind may detect errors in system libraries, for example the GNU C or X11 libraries, which you have to use. You might not be interested in these errors, since you probably have no control over that code. Therefore, Valgrind allows you to selectively suppress errors, by recording them in a suppressions file which is read when Valgrind starts up. The build mechanism selects default suppressions which give reasonable behaviour for the OS and libraries detected on your machine. To make it easier to write suppressions, you can use the --gen-suppressions=yes option. This tells Valgrind to print out a suppression for each reported error, which you can then copy into a suppressions file.

Different error-checking tools report different kinds of errors. The suppression mechanism therefore allows you to say which tool or tool(s) each suppression applies to.

2.2. Getting started

First off, consider whether it might be beneficial to recompile your application and supporting libraries with debugging info enabled (the -g option). Without debugging info, the best Valgrind tools will be able to do is guess which function a particular piece of code belongs to, which makes both error messages and profiling output nearly useless. With -g, you'll get messages which point directly to the relevant source code lines.

Another option you might like to consider, if you are working with C++, is -fno-inline. That makes it easier to see the function-call chain, which can help reduce confusion when navigating around large C++ apps. For example, debugging OpenOffice.org with Memcheck is a bit easier when using this option. You don't have to do this, but doing so helps Valgrind produce more accurate and less confusing error reports. Chances are you're set up like this already, if you intended to debug your program with GNU GDB, or some other debugger. Alternatively, the Valgrind option --read-inline-info=yes instructs Valgrind to read the debug information describing inlining information. With this, function call chain will be properly shown, even when your application is compiled with inlining.

If you are planning to use Memcheck: On rare occasions, compiler optimisations (at -O2 and above, and sometimes -O1) have been observed to generate code which fools Memcheck into wrongly reporting uninitialised value errors, or missing uninitialised value errors. We have looked in detail into fixing this, and unfortunately the result is that doing so would give a further significant slowdown in what is already a slow tool. So the best solution is to turn off optimisation altogether. Since this often makes things unmanageably slow, a reasonable compromise is to use -O. This gets you the majority of the benefits of higher optimisation levels whilst keeping relatively small the chances of false positives or false negatives from Memcheck. Also, you should compile your code with -Wall because it can identify some or all of the problems that Valgrind can miss at the higher optimisation levels. (Using -Wall is also a good idea in general.) All other tools (as far as we know) are unaffected by optimisation level, and for profiling tools like Cachegrind it is better to compile your program at its normal optimisation level.

Valgrind understands the DWARF2/3/4 formats used by GCC 3.1 and later. The reader for "stabs" debugging format (used by GCC versions prior to 3.1) has been disabled in Valgrind 3.9.0.

When you're ready to roll, run Valgrind as described above. Note that you should run the real (machine-code) executable here. If your application is started by, for example, a shell or Perl script, you'll need to modify it to invoke Valgrind on the real executables. Running such scripts directly under Valgrind will result in you getting error reports pertaining to /bin/sh, /usr/bin/perl, or whatever interpreter you're using. This may not be what you want and can be confusing. You can force the issue by giving the option --trace-children=yes, but confusion is still likely.

2.3. The Commentary

Valgrind tools write a commentary, a stream of text, detailing error reports and other significant events. All lines in the commentary have following form:

==12345== some-message-from-Valgrind

The 12345 is the process ID. This scheme makes it easy to distinguish program output from Valgrind commentary, and also easy to differentiate commentaries from different processes which have become merged together, for whatever reason.

By default, Valgrind tools write only essential messages to the commentary, so as to avoid flooding you with information of secondary importance. If you want more information about what is happening, re-run, passing the -v option to Valgrind. A second -v gives yet more detail.

You can direct the commentary to three different places:

The default: send it to a file descriptor, which is by default 2 (stderr). So, if you give the core no options, it will write commentary to the standard error stream. If you want to send it to some other file descriptor, for example number 9, you can specify --log-fd=9.

This is the simplest and most common arrangement, but can cause problems when Valgrinding entire trees of processes which expect specific file descriptors, particularly stdin/stdout/stderr, to be available for their own use.
A less intrusive option is to write the commentary to a file, which you specify by --log-file=filename. There are special format specifiers that can be used to use a process ID or an environment variable name in the log file name. These are useful/necessary if your program invokes multiple processes (especially for MPI programs). See the basic options section for more details.
The least intrusive option is to send the commentary to a network socket. The socket is specified as an IP address and port number pair, like this: --log-socket=192.168.0.1:12345 if you want to send the output to host IP 192.168.0.1 port 12345 (note: we have no idea if 12345 is a port of pre-existing significance). You can also omit the port number: --log-socket=192.168.0.1, in which case a default port of 1500 is used. This default is defined by the constant VG_CLO_DEFAULT_LOGPORT in the sources.

Note, unfortunately, that you have to use an IP address here, rather than a hostname.

Writing to a network socket is pointless if you don't have something listening at the other end. We provide a simple listener program, valgrind-listener, which accepts connections on the specified port and copies whatever it is sent to stdout. Probably someone will tell us this is a horrible security risk. It seems likely that people will write more sophisticated listeners in the fullness of time.

valgrind-listener can accept simultaneous connections from up to 50 Valgrinded processes. In front of each line of output it prints the current number of active connections in round brackets.

valgrind-listener accepts three command-line options:

-e --exit-at-zero

When the number of connected processes falls back to zero, exit. Without this, it will run forever, that is, until you send it Control-C.

--max-connect=INTEGER

By default, the listener can connect to up to 50 processes. Occasionally, that number is too small. Use this option to provide a different limit. E.g. --max-connect=100.

portnumber

Changes the port it listens on from the default (1500). The specified port must be in the range 1024 to 65535. The same restriction applies to port numbers specified by a --log-socket to Valgrind itself.

If a Valgrinded process fails to connect to a listener, for whatever reason (the listener isn't running, invalid or unreachable host or port, etc), Valgrind switches back to writing the commentary to stderr. The same goes for any process which loses an established connection to a listener. In other words, killing the listener doesn't kill the processes sending data to it.

Here is an important point about the relationship between the commentary and profiling output from tools. The commentary contains a mix of messages from the Valgrind core and the selected tool. If the tool reports errors, it will report them to the commentary. However, if the tool does profiling, the profile data will be written to a file of some kind, depending on the tool, and independent of what --log-* options are in force. The commentary is intended to be a low-bandwidth, human-readable channel. Profiling data, on the other hand, is usually voluminous and not meaningful without further processing, which is why we have chosen this arrangement.

2.4. Reporting of errors

When an error-checking tool detects something bad happening in the program, an error message is written to the commentary. Here's an example from Memcheck:

==25832== Invalid read of size 4
==25832==    at 0x8048724: BandMatrix::ReSize(int, int, int) (bogon.cpp:45)
==25832==    by 0x80487AF: main (bogon.cpp:66)
==25832==  Address 0xBFFFF74C is not stack'd, malloc'd or free'd

This message says that the program did an illegal 4-byte read of address 0xBFFFF74C, which, as far as Memcheck can tell, is not a valid stack address, nor corresponds to any current heap blocks or recently freed heap blocks. The read is happening at line 45 of bogon.cpp, called from line 66 of the same file, etc. For errors associated with an identified (current or freed) heap block, for example reading freed memory, Valgrind reports not only the location where the error happened, but also where the associated heap block was allocated/freed.

Valgrind remembers all error reports. When an error is detected, it is compared against old reports, to see if it is a duplicate. If so, the error is noted, but no further commentary is emitted. This avoids you being swamped with bazillions of duplicate error reports.

If you want to know how many times each error occurred, run with the -v option. When execution finishes, all the reports are printed out, along with, and sorted by, their occurrence counts. This makes it easy to see which errors have occurred most frequently.

Errors are reported before the associated operation actually happens. For example, if you're using Memcheck and your program attempts to read from address zero, Memcheck will emit a message to this effect, and your program will then likely die with a segmentation fault.

In general, you should try and fix errors in the order that they are reported. Not doing so can be confusing. For example, a program which copies uninitialised values to several memory locations, and later uses them, will generate several error messages, when run on Memcheck. The first such error message may well give the most direct clue to the root cause of the problem.

The process of detecting duplicate errors is quite an expensive one and can become a significant performance overhead if your program generates huge quantities of errors. To avoid serious problems, Valgrind will simply stop collecting errors after 1,000 different errors have been seen, or 10,000,000 errors in total have been seen. In this situation you might as well stop your program and fix it, because Valgrind won't tell you anything else useful after this. Note that the 1,000/10,000,000 limits apply after suppressed errors are removed. These limits are defined in m_errormgr.c and can be increased if necessary.

To avoid this cutoff you can use the --error-limit=no option. Then Valgrind will always show errors, regardless of how many there are. Use this option carefully, since it may have a bad effect on performance.

2.5. Suppressing errors

The error-checking tools detect numerous problems in the system libraries, such as the C library, which come pre-installed with your OS. You can't easily fix these, but you don't want to see these errors (and yes, there are many!) So Valgrind reads a list of errors to suppress at startup. A default suppression file is created by the ./configure script when the system is built.

You can modify and add to the suppressions file at your leisure, or, better, write your own. Multiple suppression files are allowed. This is useful if part of your project contains errors you can't or don't want to fix, yet you don't want to continuously be reminded of them.

Note: By far the easiest way to add suppressions is to use the --gen-suppressions=yes option described in Core Command-line Options. This generates suppressions automatically. For best results, though, you may want to edit the output of --gen-suppressions=yes by hand, in which case it would be advisable to read through this section.

Each error to be suppressed is described very specifically, to minimise the possibility that a suppression-directive inadvertently suppresses a bunch of similar errors which you did want to see. The suppression mechanism is designed to allow precise yet flexible specification of errors to suppress.

If you use the -v option, at the end of execution, Valgrind prints out one line for each used suppression, giving the number of times it got used, its name and the filename and line number where the suppression is defined. Depending on the suppression kind, the filename and line number are optionally followed by additional information (such as the number of blocks and bytes suppressed by a Memcheck leak suppression). Here's the suppressions used by a run of valgrind -v --tool=memcheck ls -l:

--1610-- used_suppression:      2 dl-hack3-cond-1 /usr/lib/valgrind/default.supp:1234
--1610-- used_suppression:      2 glibc-2.5.x-on-SUSE-10.2-(PPC)-2a /usr/lib/valgrind/default.supp:1234

Multiple suppressions files are allowed. Valgrind loads suppression patterns from $PREFIX/lib/valgrind/default.supp unless --default-suppressions=no has been specified. You can ask to add suppressions from additional files by specifying --suppressions=/path/to/file.supp one or more times.

If you want to understand more about suppressions, look at an existing suppressions file whilst reading the following documentation. The file glibc-2.3.supp, in the source distribution, provides some good examples.

Blank and comment lines in a suppression file are ignored. Comment lines are made of 0 or more blanks followed by a # character followed by some text.

Each suppression has the following components:

First line: its name. This merely gives a handy name to the suppression, by which it is referred to in the summary of used suppressions printed out when a program finishes. It's not important what the name is; any identifying string will do.
Second line: name of the tool(s) that the suppression is for (if more than one, comma-separated), and the name of the suppression itself, separated by a colon (n.b.: no spaces are allowed), eg:
```
tool_name1,tool_name2:suppression_name
```
Recall that Valgrind is a modular system, in which different instrumentation tools can observe your program whilst it is running. Since different tools detect different kinds of errors, it is necessary to say which tool(s) the suppression is meaningful to.

Tools will complain, at startup, if a tool does not understand any suppression directed to it. Tools ignore suppressions which are not directed to them. As a result, it is quite practical to put suppressions for all tools into the same suppression file.
Next line: a small number of suppression types have extra information after the second line (eg. the Param suppression for Memcheck)
Remaining lines: This is the calling context for the error -- the chain of function calls that led to it. There can be up to 24 of these lines.

Locations may be names of either shared objects, functions, or source lines. They begin with obj:, fun:, or src: respectively. Function, object, and file names to match against may use the wildcard characters * and ?. Source lines are specified using the form filename[:lineNumber].

Important note: C++ function names must be mangled. If you are writing suppressions by hand, use the --demangle=no option to get the mangled names in your error messages. An example of a mangled C++ name is _ZN9QListView4showEv. This is the form that the GNU C++ compiler uses internally, and the form that must be used in suppression files. The equivalent demangled name, QListView::show(), is what you see at the C++ source code level.

A location line may also be simply "..." (three dots). This is a frame-level wildcard, which matches zero or more frames. Frame level wildcards are useful because they make it easy to ignore varying numbers of uninteresting frames in between frames of interest. That is often important when writing suppressions which are intended to be robust against variations in the amount of function inlining done by compilers.
Finally, the entire suppression must be between curly braces. Each brace must be the first character on its own line.

A suppression only suppresses an error when the error matches all the details in the suppression. Here's an example:

{
  __gconv_transform_ascii_internal/__mbrtowc/mbtowc
  Memcheck:Value4
  fun:__gconv_transform_ascii_internal
  fun:__mbr*toc
  fun:mbtowc
}

What it means is: for Memcheck only, suppress a use-of-uninitialised-value error, when the data size is 4, when it occurs in the function __gconv_transform_ascii_internal, when that is called from any function of name matching __mbr*toc, when that is called from mbtowc. It doesn't apply under any other circumstances. The string by which this suppression is identified to the user is __gconv_transform_ascii_internal/__mbrtowc/mbtowc.

(See Writing suppression files for more details on the specifics of Memcheck's suppression kinds.)

Another example, again for the Memcheck tool:

{
  libX11.so.6.2/libX11.so.6.2/libXaw.so.7.0
  Memcheck:Value4
  obj:/usr/X11R6/lib/libX11.so.6.2
  obj:/usr/X11R6/lib/libX11.so.6.2
  obj:/usr/X11R6/lib/libXaw.so.7.0
}

This suppresses any size 4 uninitialised-value error which occurs anywhere in libX11.so.6.2, when called from anywhere in the same library, when called from anywhere in libXaw.so.7.0. The inexact specification of locations is regrettable, but is about all you can hope for, given that the X11 libraries shipped on the Linux distro on which this example was made have had their symbol tables removed.

An example of the src: specification, again for the Memcheck tool:

{
  libX11.so.6.2/libX11.so.6.2/libXaw.so.7.0
  Memcheck:Value4
  src:valid.c:321
}

This suppresses any size-4 uninitialised-value error which occurs at line 321 in valid.c.

Although the above two examples do not make this clear, you can freely mix obj:, fun:, and src: lines in a suppression.

Finally, here's an example using three frame-level wildcards:

{
   a-contrived-example
   Memcheck:Leak
   fun:malloc
   ...
   fun:ddd
   ...
   fun:ccc
   ...
   fun:main
}

This suppresses Memcheck memory-leak errors, in the case where the allocation was done by main calling (though any number of intermediaries, including zero) ccc, calling onwards via ddd and eventually to malloc..

2.6. Debuginfod

Valgrind supports the downloading of debuginfo files via debuginfod, an HTTP server for distributing ELF/DWARF debugging information. When a debuginfo file cannot be found locally, Valgrind is able to query debuginfod servers for the file using its build-id.

In order to use this feature debuginfod-find must be installed and $DEBUGINFOD_URLS must contain URLs of debuginfod servers. Valgrind does not support debuginfod-find verbose output that is normally enabled with $DEBUGINFOD_PROGRESS and $DEBUGINFOD_VERBOSE. These environment variables will be ignored.

For more information regarding debuginfod, see https://sourceware.org/elfutils/Debuginfod.html

2.7. Core Command-line Options

As mentioned above, Valgrind's core accepts a common set of options. The tools also accept tool-specific options, which are documented separately for each tool.

Valgrind's default settings succeed in giving reasonable behaviour in most cases. We group the available options by rough categories.

2.7.1. Tool-selection Option

The single most important option.

--tool=<toolname> [default: memcheck]: Run the Valgrind tool called toolname, e.g. memcheck, cachegrind, callgrind, helgrind, drd, massif, dhat, lackey, none, exp-bbv, etc.

2.7.2. Basic Options

These options work with all tools.

-h --help

Show help for all options, both for the core and for the selected tool. If the option is repeated it is equivalent to giving --help-debug.

--help-debug

Same as --help, but also lists debugging options which usually are only of use to Valgrind's developers.

--version

Show the version number of the Valgrind core. Tools can have their own version numbers. There is a scheme in place to ensure that tools only execute when the core version is one they are known to work with. This was done to minimise the chances of strange problems arising from tool-vs-core version incompatibilities.

-q, --quiet

Run silently, and only print error messages. Useful if you are running regression tests or have some other automated test machinery.

-v, --verbose

Be more verbose. Gives extra information on various aspects of your program, such as: the shared objects loaded, the suppressions used, the progress of the instrumentation and execution engines, and warnings about unusual behaviour. Repeating the option increases the verbosity level.

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

When enabled, Valgrind will trace into sub-processes initiated via the exec system call. This is necessary for multi-process programs.

Note that Valgrind does trace into the child of a fork (it would be difficult not to, since fork makes an identical copy of a process), so this option is arguably badly named. However, most children of fork calls immediately call exec anyway.

--trace-children-skip=patt1,patt2,...

This option only has an effect when --trace-children=yes is specified. It allows for some children to be skipped. The option takes a comma separated list of patterns for the names of child executables that Valgrind should not trace into. Patterns may include the metacharacters ? and *, which have the usual meaning.

This can be useful for pruning uninteresting branches from a tree of processes being run on Valgrind. But you should be careful when using it. When Valgrind skips tracing into an executable, it doesn't just skip tracing that executable, it also skips tracing any of that executable's child processes. In other words, the flag doesn't merely cause tracing to stop at the specified executables -- it skips tracing of entire process subtrees rooted at any of the specified executables.

--trace-children-skip-by-arg=patt1,patt2,...

This is the same as --trace-children-skip, with one difference: the decision as to whether to trace into a child process is made by examining the arguments to the child process, rather than the name of its executable.

--child-silent-after-fork=<yes|no> [default: no]

When enabled, Valgrind will not show any debugging or logging output for the child process resulting from a fork call. This can make the output less confusing (although more misleading) when dealing with processes that create children. It is particularly useful in conjunction with --trace-children=. Use of this option is also strongly recommended if you are requesting XML output (--xml=yes), since otherwise the XML from child and parent may become mixed up, which usually makes it useless.

--vgdb=<no|yes|full> [default: yes]

Valgrind will provide "gdbserver" functionality when --vgdb=yes or --vgdb=full is specified. This allows an external GNU GDB debugger to control and debug your program when it runs on Valgrind. --vgdb=full incurs significant performance overheads, but provides more precise breakpoints and watchpoints. See Debugging your program using Valgrind's gdbserver and GDB for a detailed description.

If the embedded gdbserver is enabled but no gdb is currently being used, the vgdb command line utility can send "monitor commands" to Valgrind from a shell. The Valgrind core provides a set of Valgrind monitor commands. A tool can optionally provide tool specific monitor commands, which are documented in the tool specific chapter.

--vgdb-error=<number> [default: 999999999]

Use this option when the Valgrind gdbserver is enabled with --vgdb=yes or --vgdb=full. Tools that report errors will wait for "number" errors to be reported before freezing the program and waiting for you to connect with GDB. It follows that a value of zero will cause the gdbserver to be started before your program is executed. This is typically used to insert GDB breakpoints before execution, and also works with tools that do not report errors, such as Massif.

--vgdb-stop-at=<set> [default: none]

Use this option when the Valgrind gdbserver is enabled with --vgdb=yes or --vgdb=full. The Valgrind gdbserver will be invoked for each error after --vgdb-error have been reported. You can additionally ask the Valgrind gdbserver to be invoked for other events, specified in one of the following ways:

a comma separated list of one or more of startup exit valgrindabexit.

The values startup exit valgrindabexit respectively indicate to invoke gdbserver before your program is executed, after the last instruction of your program, on Valgrind abnormal exit (e.g. internal error, out of memory, ...).

Note: startup and --vgdb-error=0 will both cause Valgrind gdbserver to be invoked before your program is executed. The --vgdb-error=0 will in addition cause your program to stop on all subsequent errors.
all to specify the complete set. It is equivalent to --vgdb-stop-at=startup,exit,valgrindabexit.
none for the empty set.

--track-fds=<yes|no|all> [default: no]

When enabled, Valgrind will print out a list of open file descriptors on exit or on request, via the gdbserver monitor command v.info open_fds. Along with each file descriptor is printed a stack backtrace of where the file was opened and any details relating to the file descriptor such as the file name or socket details. Use all to include reporting on stdin, stdout and stderr.

--time-stamp=<yes|no> [default: no]

When enabled, each message is preceded with an indication of the elapsed wallclock time since startup, expressed as days, hours, minutes, seconds and milliseconds.

--log-fd=<number> [default: 2, stderr]

Specifies that Valgrind should send all of its messages to the specified file descriptor. The default, 2, is the standard error channel (stderr). Note that this may interfere with the client's own use of stderr, as Valgrind's output will be interleaved with any output that the client sends to stderr.

--log-file=<filename>

Specifies that Valgrind should send all of its messages to the specified file. If the file name is empty, it causes an abort. There are three special format specifiers that can be used in the file name.

%p is replaced with the current process ID. This is very useful for program that invoke multiple processes. WARNING: If you use --trace-children=yes and your program invokes multiple processes OR your program forks without calling exec afterwards, and you don't use this specifier (or the %q specifier below), the Valgrind output from all those processes will go into one file, possibly jumbled up, and possibly incomplete. Note: If the program forks and calls exec afterwards, Valgrind output of the child from the period between fork and exec will be lost. Fortunately this gap is really tiny for most programs; and modern programs use posix_spawn anyway.

%n is replaced with a file sequence number unique for this process. This is useful for processes that produces several files from the same filename template.

%q{FOO} is replaced with the contents of the environment variable FOO. If the {FOO} part is malformed, it causes an abort. This specifier is rarely needed, but very useful in certain circumstances (eg. when running MPI programs). The idea is that you specify a variable which will be set differently for each process in the job, for example BPROC_RANK or whatever is applicable in your MPI setup. If the named environment variable is not set, it causes an abort. Note that in some shells, the { and } characters may need to be escaped with a backslash.

%% is replaced with %.

If an % is followed by any other character, it causes an abort.

If the file name specifies a relative file name, it is put in the program's initial working directory: this is the current directory when the program started its execution after the fork or after the exec. If it specifies an absolute file name (ie. starts with '/') then it is put there.

--log-socket=<ip-address:port-number>

Specifies that Valgrind should send all of its messages to the specified port at the specified IP address. The port may be omitted, in which case port 1500 is used. If a connection cannot be made to the specified socket, Valgrind falls back to writing output to the standard error (stderr). This option is intended to be used in conjunction with the valgrind-listener program. For further details, see the commentary in the manual.

2.7.3. Error-related Options

These options are used by all tools that can report errors, e.g. Memcheck, but not Cachegrind.

--xml=<yes|no> [default: no]

When enabled, the important parts of the output (e.g. tool error messages) will be in XML format rather than plain text. Furthermore, the XML output will be sent to a different output channel than the plain text output. Therefore, you also must use one of --xml-fd, --xml-file or --xml-socket to specify where the XML is to be sent.

Less important messages will still be printed in plain text, but because the XML output and plain text output are sent to different output channels (the destination of the plain text output is still controlled by --log-fd, --log-file and --log-socket) this should not cause problems.

This option is aimed at making life easier for tools that consume Valgrind's output as input, such as GUI front ends. Currently this option works with Memcheck, Helgrind and DRD. The output format is specified in the file docs/internals/xml-output-protocol4.txt in the source tree for Valgrind 3.5.0 or later.

The recommended options for a GUI to pass, when requesting XML output, are: --xml=yes to enable XML output, --xml-file to send the XML output to a (presumably GUI-selected) file, --log-file to send the plain text output to a second GUI-selected file, --child-silent-after-fork=yes, and -q to restrict the plain text output to critical error messages created by Valgrind itself. For example, failure to read a specified suppressions file counts as a critical error message. In this way, for a successful run the text output file will be empty. But if it isn't empty, then it will contain important information which the GUI user should be made aware of.

--xml-fd=<number> [default: -1, disabled]

Specifies that Valgrind should send its XML output to the specified file descriptor. It must be used in conjunction with --xml=yes.

--xml-file=<filename>

Specifies that Valgrind should send its XML output to the specified file. It must be used in conjunction with --xml=yes. Any %p or %q sequences appearing in the filename are expanded in exactly the same way as they are for --log-file. See the description of --log-file for details.

--xml-socket=<ip-address:port-number>

Specifies that Valgrind should send its XML output the specified port at the specified IP address. It must be used in conjunction with --xml=yes. The form of the argument is the same as that used by --log-socket. See the description of --log-socket for further details.

--xml-user-comment=<string>

Embeds an extra user comment string at the start of the XML output. Only works when --xml=yes is specified; ignored otherwise.

--demangle=<yes|no> [default: yes]

Enable/disable automatic demangling (decoding) of C++ names. Enabled by default. When enabled, Valgrind will attempt to translate encoded C++ names back to something approaching the original. The demangler handles symbols mangled by g++ versions 2.X, 3.X and 4.X.

An important fact about demangling is that function names mentioned in suppressions files should be in their mangled form. Valgrind does not demangle function names when searching for applicable suppressions, because to do otherwise would make suppression file contents dependent on the state of Valgrind's demangling machinery, and also slow down suppression matching.

--num-callers=<number> [default: 12]

Specifies the maximum number of entries shown in stack traces that identify program locations. Note that errors are commoned up using only the top four function locations (the place in the current function, and that of its three immediate callers). So this doesn't affect the total number of errors reported.

The maximum value for this is 500. Note that higher settings will make Valgrind run a bit more slowly and take a bit more memory, but can be useful when working with programs with deeply-nested call chains.

--unw-stack-scan-thresh=<number> [default: 0] , --unw-stack-scan-frames=<number> [default: 5]

Stack-scanning support is available only on ARM targets.

These flags enable and control stack unwinding by stack scanning. When the normal stack unwinding mechanisms -- usage of Dwarf CFI records, and frame-pointer following -- fail, stack scanning may be able to recover a stack trace.

Note that stack scanning is an imprecise, heuristic mechanism that may give very misleading results, or none at all. It should be used only in emergencies, when normal unwinding fails, and it is important to nevertheless have stack traces.

Stack scanning is a simple technique: the unwinder reads words from the stack, and tries to guess which of them might be return addresses, by checking to see if they point just after ARM or Thumb call instructions. If so, the word is added to the backtrace.

The main danger occurs when a function call returns, leaving its return address exposed, and a new function is called, but the new function does not overwrite the old address. The result of this is that the backtrace may contain entries for functions which have already returned, and so be very confusing.

A second limitation of this implementation is that it will scan only the page (4KB, normally) containing the starting stack pointer. If the stack frames are large, this may result in only a few (or not even any) being present in the trace. Also, if you are unlucky and have an initial stack pointer near the end of its containing page, the scan may miss all interesting frames.

By default stack scanning is disabled. The normal use case is to ask for it when a stack trace would otherwise be very short. So, to enable it, use --unw-stack-scan-thresh=number. This requests Valgrind to try using stack scanning to "extend" stack traces which contain fewer than number frames.

If stack scanning does take place, it will only generate at most the number of frames specified by --unw-stack-scan-frames. Typically, stack scanning generates so many garbage entries that this value is set to a low value (5) by default. In no case will a stack trace larger than the value specified by --num-callers be created.

--error-limit=<yes|no> [default: yes]

When enabled, Valgrind stops reporting errors after 10,000,000 in total, or 1,000 different ones, have been seen. This is to stop the error tracking machinery from becoming a huge performance overhead in programs with many errors.

--error-exitcode=<number> [default: 0]

Specifies an alternative exit code to return if Valgrind reported any errors in the run. When set to the default value (zero), the return value from Valgrind will always be the return value of the process being simulated. When set to a nonzero value, that value is returned instead, if Valgrind detects any errors. This is useful for using Valgrind as part of an automated test suite, since it makes it easy to detect test cases for which Valgrind has reported errors, just by inspecting return codes.

--exit-on-first-error=<yes|no> [default: no]

If this option is enabled, Valgrind exits on the first error. A nonzero exit value must be defined using --error-exitcode option. Useful if you are running regression tests or have some other automated test machinery.

--error-markers=<begin>,<end> [default: none]

When errors are output as plain text (i.e. XML not used), --error-markers instructs to output a line containing the begin (end) string before (after) each error.

Such marker lines facilitate searching for errors and/or extracting errors in an output file that contain valgrind errors mixed with the program output.

Note that empty markers are accepted. So, only using a begin (or an end) marker is possible.

--show-error-list=no|yes [default: no]

If this option is enabled, for tools that report errors, valgrind will show the list of detected errors and the list of used suppressions at exit.

Note that at verbosity 2 and above, valgrind automatically shows the list of detected errors and the list of used suppressions at exit, unless --show-error-list=no is selected.

-s

Specifying -s is equivalent to --show-error-list=yes.

--sigill-diagnostics=<yes|no> [default: yes]

Enable/disable printing of illegal instruction diagnostics. Enabled by default, but defaults to disabled when --quiet is given. The default can always be explicitly overridden by giving this option.

When enabled, a warning message will be printed, along with some diagnostics, whenever an instruction is encountered that Valgrind cannot decode or translate, before the program is given a SIGILL signal. Often an illegal instruction indicates a bug in the program or missing support for the particular instruction in Valgrind. But some programs do deliberately try to execute an instruction that might be missing and trap the SIGILL signal to detect processor features. Using this flag makes it possible to avoid the diagnostic output that you would otherwise get in such cases.

--keep-debuginfo=<yes|no> [default: no]

When enabled, keep ("archive") symbols and all other debuginfo for unloaded code. This allows saved stack traces to include file/line info for code that has been dlclose'd (or similar). Be careful with this, since it can lead to unbounded memory use for programs which repeatedly load and unload shared objects.

Some tools and some functionalities have only limited support for archived debug info. Memcheck fully supports it. Generally, tools that report errors can use archived debug info to show the error stack traces. The known limitations are: Helgrind's past access stack trace of a race condition is does not use archived debug info. Massif (and more generally the xtree Massif output format) does not make use of archived debug info. Only Memcheck has been (somewhat) tested with --keep-debuginfo=yes, so other tools may have unknown limitations.

--show-below-main=<yes|no> [default: no]

By default, stack traces for errors do not show any functions that appear beneath main because most of the time it's uninteresting C library stuff and/or gobbledygook. Alternatively, if main is not present in the stack trace, stack traces will not show any functions below main-like functions such as glibc's __libc_start_main. Furthermore, if main-like functions are present in the trace, they are normalised as (below main), in order to make the output more deterministic.

If this option is enabled, all stack trace entries will be shown and main-like functions will not be normalised.

--fullpath-after=<string> [default: don't show source paths]

By default Valgrind only shows the filenames in stack traces, but not full paths to source files. When using Valgrind in large projects where the sources reside in multiple different directories, this can be inconvenient. --fullpath-after provides a flexible solution to this problem. When this option is present, the path to each source file is shown, with the following all-important caveat: if string is found in the path, then the path up to and including string is omitted, else the path is shown unmodified. Note that string is not required to be a prefix of the path.

For example, consider a file named /home/janedoe/blah/src/foo/bar/xyzzy.c. Specifying --fullpath-after=/home/janedoe/blah/src/ will cause Valgrind to show the name as foo/bar/xyzzy.c.

Because the string is not required to be a prefix, --fullpath-after=src/ will produce the same output. This is useful when the path contains arbitrary machine-generated characters. For example, the path /my/build/dir/C32A1B47/blah/src/foo/xyzzy can be pruned to foo/xyzzy using --fullpath-after=/blah/src/.

If you simply want to see the full path, just specify an empty string: --fullpath-after=. This isn't a special case, merely a logical consequence of the above rules.

Finally, you can use --fullpath-after multiple times. Any appearance of it causes Valgrind to switch to producing full paths and applying the above filtering rule. Each produced path is compared against all the --fullpath-after-specified strings, in the order specified. The first string to match causes the path to be truncated as described above. If none match, the full path is shown. This facilitates chopping off prefixes when the sources are drawn from a number of unrelated directories.

--extra-debuginfo-path=<path> [default: undefined and unused]

By default Valgrind searches in several well-known paths for debug objects, such as /usr/lib/debug/.

However, there may be scenarios where you may wish to put debug objects at an arbitrary location, such as external storage when running Valgrind on a mobile device with limited local storage. Another example might be a situation where you do not have permission to install debug object packages on the system where you are running Valgrind.

In these scenarios, you may provide an absolute path as an extra, final place for Valgrind to search for debug objects by specifying --extra-debuginfo-path=/path/to/debug/objects. The given path will be prepended to the absolute path name of the searched-for object. For example, if Valgrind is looking for the debuginfo for /w/x/y/zz.so and --extra-debuginfo-path=/a/b/c is specified, it will look for a debug object at /a/b/c/w/x/y/zz.so.

This flag should only be specified once. If it is specified multiple times, only the last instance is honoured.

--debuginfo-server=ipaddr:port [default: undefined and unused]

This is a new, experimental, feature introduced in version 3.9.0.

In some scenarios it may be convenient to read debuginfo from objects stored on a different machine. With this flag, Valgrind will query a debuginfo server running on ipaddr and listening on port port, if it cannot find the debuginfo object in the local filesystem.

The debuginfo server must accept TCP connections on port port. The debuginfo server is contained in the source file auxprogs/valgrind-di-server.c. It will only serve from the directory it is started in. port defaults to 1500 in both client and server if not specified.

If Valgrind looks for the debuginfo for /w/x/y/zz.so by using the debuginfo server, it will strip the pathname components and merely request zz.so on the server. That in turn will look only in its current working directory for a matching debuginfo object.

The debuginfo data is transmitted in small fragments (8 KB) as requested by Valgrind. Each block is compressed using LZO to reduce transmission time. The implementation has been tuned for best performance over a single-stage 802.11g (WiFi) network link.

Note that checks for matching primary vs debug objects, using GNU debuglink CRC scheme, are performed even when using the debuginfo server. To disable such checking, you need to also specify --allow-mismatched-debuginfo=yes.

By default the Valgrind build system will build valgrind-di-server for the target platform, which is almost certainly not what you want. So far we have been unable to find out how to get automake/autoconf to build it for the build platform. If you want to use it, you will have to recompile it by hand using the command shown at the top of auxprogs/valgrind-di-server.c.

Valgrind can also download debuginfo via debuginfod. See the DEBUGINFOD section for more information.

--allow-mismatched-debuginfo=no|yes [no]

When reading debuginfo from separate debuginfo objects, Valgrind will by default check that the main and debuginfo objects match, using the GNU debuglink mechanism. This guarantees that it does not read debuginfo from out of date debuginfo objects, and also ensures that Valgrind can't crash as a result of mismatches.

This check can be overridden using --allow-mismatched-debuginfo=yes. This may be useful when the debuginfo and main objects have not been split in the proper way. Be careful when using this, though: it disables all consistency checking, and Valgrind has been observed to crash when the main and debuginfo objects don't match.

--suppressions=<filename> [default: $PREFIX/lib/valgrind/default.supp]

Specifies an extra file from which to read descriptions of errors to suppress. You may use up to 100 extra suppression files.

--gen-suppressions=<yes|no|all> [default: no]

When set to yes, Valgrind will pause after every error shown and print the line:

---- Print suppression ? --- [Return/N/n/Y/y/C/c] ----

Pressing Ret, or N Ret or n Ret, causes Valgrind continue execution without printing a suppression for this error.

Pressing Y Ret or y Ret causes Valgrind to write a suppression for this error. You can then cut and paste it into a suppression file if you don't want to hear about the error in the future.

When set to all, Valgrind will print a suppression for every reported error, without querying the user.

This option is particularly useful with C++ programs, as it prints out the suppressions with mangled names, as required.

Note that the suppressions printed are as specific as possible. You may want to common up similar ones, by adding wildcards to function names, and by using frame-level wildcards. The wildcarding facilities are powerful yet flexible, and with a bit of careful editing, you may be able to suppress a whole family of related errors with only a few suppressions.

Sometimes two different errors are suppressed by the same suppression, in which case Valgrind will output the suppression more than once, but you only need to have one copy in your suppression file (but having more than one won't cause problems). Also, the suppression name is given as <insert a suppression name here>; the name doesn't really matter, it's only used with the -v option which prints out all used suppression records.

--input-fd=<number> [default: 0, stdin]

When using --gen-suppressions=yes, Valgrind will stop so as to read keyboard input from you when each error occurs. By default it reads from the standard input (stdin), which is problematic for programs which close stdin. This option allows you to specify an alternative file descriptor from which to read input.

--dsymutil=no|yes [yes]

This option is only relevant when running Valgrind on Mac OS X.

Mac OS X uses a deferred debug information (debuginfo) linking scheme. When object files containing debuginfo are linked into a .dylib or an executable, the debuginfo is not copied into the final file. Instead, the debuginfo must be linked manually by running dsymutil, a system-provided utility, on the executable or .dylib. The resulting combined debuginfo is placed in a directory alongside the executable or .dylib, but with the extension .dSYM.

With --dsymutil=no, Valgrind will detect cases where the .dSYM directory is either missing, or is present but does not appear to match the associated executable or .dylib, most likely because it is out of date. In these cases, Valgrind will print a warning message but take no further action.

With --dsymutil=yes, Valgrind will, in such cases, automatically run dsymutil as necessary to bring the debuginfo up to date. For all practical purposes, if you always use --dsymutil=yes, then there is never any need to run dsymutil manually or as part of your applications's build system, since Valgrind will run it as necessary.

Valgrind will not attempt to run dsymutil on any executable or library in /usr/, /bin/, /sbin/, /opt/, /sw/, /System/, /Library/ or /Applications/ since dsymutil will always fail in such situations. It fails both because the debuginfo for such pre-installed system components is not available anywhere, and also because it would require write privileges in those directories.

Be careful when using --dsymutil=yes, since it will cause pre-existing .dSYM directories to be silently deleted and re-created. Also note that dsymutil is quite slow, sometimes excessively so.

--max-stackframe=<number> [default: 2000000]

The maximum size of a stack frame. If the stack pointer moves by more than this amount then Valgrind will assume that the program is switching to a different stack.

You may need to use this option if your program has large stack-allocated arrays. Valgrind keeps track of your program's stack pointer. If it changes by more than the threshold amount, Valgrind assumes your program is switching to a different stack, and Memcheck behaves differently than it would for a stack pointer change smaller than the threshold. Usually this heuristic works well. However, if your program allocates large structures on the stack, this heuristic will be fooled, and Memcheck will subsequently report large numbers of invalid stack accesses. This option allows you to change the threshold to a different value.

You should only consider use of this option if Valgrind's debug output directs you to do so. In that case it will tell you the new threshold you should specify.

In general, allocating large structures on the stack is a bad idea, because you can easily run out of stack space, especially on systems with limited memory or which expect to support large numbers of threads each with a small stack, and also because the error checking performed by Memcheck is more effective for heap-allocated data than for stack-allocated data. If you have to use this option, you may wish to consider rewriting your code to allocate on the heap rather than on the stack.

--main-stacksize=<number> [default: use current 'ulimit' value]

Specifies the size of the main thread's stack.

To simplify its memory management, Valgrind reserves all required space for the main thread's stack at startup. That means it needs to know the required stack size at startup.

By default, Valgrind uses the current "ulimit" value for the stack size, or 16 MB, whichever is lower. In many cases this gives a stack size in the range 8 to 16 MB, which almost never overflows for most applications.

If you need a larger total stack size, use --main-stacksize to specify it. Only set it as high as you need, since reserving far more space than you need (that is, hundreds of megabytes more than you need) constrains Valgrind's memory allocators and may reduce the total amount of memory that Valgrind can use. This is only really of significance on 32-bit machines.

On Linux, you may request a stack of size up to 2GB. Valgrind will stop with a diagnostic message if the stack cannot be allocated.

--main-stacksize only affects the stack size for the program's initial thread. It has no bearing on the size of thread stacks, as Valgrind does not allocate those.

You may need to use both --main-stacksize and --max-stackframe together. It is important to understand that --main-stacksize sets the maximum total stack size, whilst --max-stackframe specifies the largest size of any one stack frame. You will have to work out the --main-stacksize value for yourself (usually, if your applications segfaults). But Valgrind will tell you the needed --max-stackframe size, if necessary.

As discussed further in the description of --max-stackframe, a requirement for a large stack is a sign of potential portability problems. You are best advised to place all large data in heap-allocated memory.

--max-threads=<number> [default: 500]

By default, Valgrind can handle to up to 500 threads. Occasionally, that number is too small. Use this option to provide a different limit. E.g. --max-threads=3000.

2.7.4. malloc-related Options

For tools that use their own version of malloc (e.g. Memcheck, Massif, Helgrind, DRD), the following options apply.

--alignment=<number> [default: 8 or 16, depending on the platform]

By default Valgrind's malloc, realloc, etc, return a block whose starting address is 8-byte aligned or 16-byte aligned (the value depends on the platform and matches the platform default). This option allows you to specify a different alignment. The supplied value must be greater than or equal to the default, less than or equal to 4096, and must be a power of two.

--redzone-size=<number> [default: depends on the tool]

Valgrind's malloc, realloc, etc, add padding blocks before and after each heap block allocated by the program being run. Such padding blocks are called redzones. The default value for the redzone size depends on the tool. For example, Memcheck adds and protects a minimum of 16 bytes before and after each block allocated by the client. This allows it to detect block underruns or overruns of up to 16 bytes.

Increasing the redzone size makes it possible to detect overruns of larger distances, but increases the amount of memory used by Valgrind. Decreasing the redzone size will reduce the memory needed by Valgrind but also reduces the chances of detecting over/underruns, so is not recommended.

--xtree-memory=none|allocs|full [none]

Tools replacing Valgrind's malloc, realloc, etc, can optionally produce an execution tree detailing which piece of code is responsible for heap memory usage. See Execution Trees for a detailed explanation about execution trees.

When set to none, no memory execution tree is produced.

When set to allocs, the memory execution tree gives the current number of allocated bytes and the current number of allocated blocks.

When set to full, the memory execution tree gives 6 different measurements : the current number of allocated bytes and blocks (same values as for allocs), the total number of allocated bytes and blocks, the total number of freed bytes and blocks.

Note that the overhead in cpu and memory to produce an xtree depends on the tool. The overhead in cpu is small for the value allocs, as the information needed to produce this report is maintained in any case by the tool. For massif and helgrind, specifying full implies to capture a stack trace for each free operation, while normally these tools only capture an allocation stack trace. For Memcheck, the cpu overhead for the value full is small, as this can only be used in combination with --keep-stacktraces=alloc-and-free or --keep-stacktraces=alloc-then-free, which already records a stack trace for each free operation. The memory overhead varies between 5 and 10 words per unique stacktrace in the xtree, plus the memory needed to record the stack trace for the free operations, if needed specifically for the xtree.

--xtree-memory-file=<filename> [default: xtmemory.kcg.%p]

Specifies that Valgrind should produce the xtree memory report in the specified file. Any %p or %q sequences appearing in the filename are expanded in exactly the same way as they are for --log-file. See the description of --log-file for details.

If the filename contains the extension .ms, then the produced file format will be a massif output file format. If the filename contains the extension .kcg or no extension is provided or recognised, then the produced file format will be a callgrind output format.

See Execution Trees for a detailed explanation about execution trees formats.

2.7.5. Uncommon Options

These options apply to all tools, as they affect certain obscure workings of the Valgrind core. Most people won't need to use them.

--smc-check=<none|stack|all|all-non-file> [default: all-non-file for x86/amd64/s390x, stack for other archs]

This option controls Valgrind's detection of self-modifying code. If no checking is done, when a program executes some code, then overwrites it with new code, and executes the new code, Valgrind will continue to execute the translations it made for the old code. This will likely lead to incorrect behaviour and/or crashes.

For "modern" architectures -- anything that's not x86, amd64 or s390x -- the default is stack. This is because a correct program must take explicit action to reestablish D-I cache coherence following code modification. Valgrind observes and honours such actions, with the result that self-modifying code is transparently handled with zero extra cost.

For x86, amd64 and s390x, the program is not required to notify the hardware of required D-I coherence syncing. Hence the default is all-non-file, which covers the normal case of generating code into an anonymous (non-file-backed) mmap'd area.

The meanings of the four available settings are as follows. No detection (none), detect self-modifying code on the stack (which is used by GCC to implement nested functions) (stack), detect self-modifying code everywhere (all), and detect self-modifying code everywhere except in file-backed mappings (all-non-file).

Running with all will slow Valgrind down noticeably. Running with none will rarely speed things up, since very little code gets dynamically generated in most programs. The VALGRIND_DISCARD_TRANSLATIONS client request is an alternative to --smc-check=all and --smc-check=all-non-file that requires more programmer effort but allows Valgrind to run your program faster, by telling it precisely when translations need to be re-made.

--smc-check=all-non-file provides a cheaper but more limited version of --smc-check=all. It adds checks to any translations that do not originate from file-backed memory mappings. Typical applications that generate code, for example JITs in web browsers, generate code into anonymous mmaped areas, whereas the "fixed" code of the browser always lives in file-backed mappings. --smc-check=all-non-file takes advantage of this observation, limiting the overhead of checking to code which is likely to be JIT generated.

--read-inline-info=<yes|no> [default: see below]

When enabled, Valgrind will read information about inlined function calls from DWARF3 debug info. This slows Valgrind startup and makes it use more memory (typically for each inlined piece of code, 6 words and space for the function name), but it results in more descriptive stacktraces. Currently, this functionality is enabled by default only for Linux, Android and Solaris targets and only for the tools Memcheck, Massif, Helgrind and DRD. Here is an example of some stacktraces with --read-inline-info=no:

==15380== Conditional jump or move depends on uninitialised value(s)
==15380==    at 0x80484EA: main (inlinfo.c:6)
==15380== 
==15380== Conditional jump or move depends on uninitialised value(s)
==15380==    at 0x8048550: fun_noninline (inlinfo.c:6)
==15380==    by 0x804850E: main (inlinfo.c:34)
==15380== 
==15380== Conditional jump or move depends on uninitialised value(s)
==15380==    at 0x8048520: main (inlinfo.c:6)

And here are the same errors with --read-inline-info=yes:

==15377== Conditional jump or move depends on uninitialised value(s)
==15377==    at 0x80484EA: fun_d (inlinfo.c:6)
==15377==    by 0x80484EA: fun_c (inlinfo.c:14)
==15377==    by 0x80484EA: fun_b (inlinfo.c:20)
==15377==    by 0x80484EA: fun_a (inlinfo.c:26)
==15377==    by 0x80484EA: main (inlinfo.c:33)
==15377== 
==15377== Conditional jump or move depends on uninitialised value(s)
==15377==    at 0x8048550: fun_d (inlinfo.c:6)
==15377==    by 0x8048550: fun_noninline (inlinfo.c:41)
==15377==    by 0x804850E: main (inlinfo.c:34)
==15377== 
==15377== Conditional jump or move depends on uninitialised value(s)
==15377==    at 0x8048520: fun_d (inlinfo.c:6)
==15377==    by 0x8048520: main (inlinfo.c:35)

--read-var-info=<yes|no> [default: no]

When enabled, Valgrind will read information about variable types and locations from DWARF3 debug info. This slows Valgrind startup significantly and makes it use significantly more memory, but for the tools that can take advantage of it (Memcheck, Helgrind, DRD) it can result in more precise error messages. For example, here are some standard errors issued by Memcheck:

==15363== Uninitialised byte(s) found during client check request
==15363==    at 0x80484A9: croak (varinfo1.c:28)
==15363==    by 0x8048544: main (varinfo1.c:55)
==15363==  Address 0x80497f7 is 7 bytes inside data symbol "global_i2"
==15363== 
==15363== Uninitialised byte(s) found during client check request
==15363==    at 0x80484A9: croak (varinfo1.c:28)
==15363==    by 0x8048550: main (varinfo1.c:56)
==15363==  Address 0xbea0d0cc is on thread 1's stack
==15363==  in frame #1, created by main (varinfo1.c:45)

And here are the same errors with --read-var-info=yes:

==15370== Uninitialised byte(s) found during client check request
==15370==    at 0x80484A9: croak (varinfo1.c:28)
==15370==    by 0x8048544: main (varinfo1.c:55)
==15370==  Location 0x80497f7 is 0 bytes inside global_i2[7],
==15370==  a global variable declared at varinfo1.c:41
==15370== 
==15370== Uninitialised byte(s) found during client check request
==15370==    at 0x80484A9: croak (varinfo1.c:28)
==15370==    by 0x8048550: main (varinfo1.c:56)
==15370==  Location 0xbeb4a0cc is 0 bytes inside local var "local"
==15370==  declared at varinfo1.c:46, in frame #1 of thread 1

--vgdb-poll=<number> [default: 5000]

As part of its main loop, the Valgrind scheduler will poll to check if some activity (such as an external command or some input from a gdb) has to be handled by gdbserver. This activity poll will be done after having run the given number of basic blocks (or slightly more than the given number of basic blocks). This poll is quite cheap so the default value is set relatively low. You might further decrease this value if vgdb cannot use ptrace system call to interrupt Valgrind if all threads are (most of the time) blocked in a system call.

--vgdb-shadow-registers=no|yes [default: no]

When activated, gdbserver will expose the Valgrind shadow registers to GDB. With this, the value of the Valgrind shadow registers can be examined or changed using GDB. Exposing shadow registers only works with GDB version 7.1 or later.

--vgdb-prefix=<prefix> [default: /tmp/vgdb-pipe]

To communicate with gdb/vgdb, the Valgrind gdbserver creates 3 files (2 named FIFOs and a mmap shared memory file). The prefix option controls the directory and prefix for the creation of these files.

--run-libc-freeres=<yes|no> [default: yes]

This option is only relevant when running Valgrind on Linux.

The GNU C library (libc.so), which is used by all programs, may allocate memory for its own uses. Usually it doesn't bother to free that memory when the program ends—there would be no point, since the Linux kernel reclaims all process resources when a process exits anyway, so it would just slow things down.

The glibc authors realised that this behaviour causes leak checkers, such as Valgrind, to falsely report leaks in glibc, when a leak check is done at exit. In order to avoid this, they provided a routine called __libc_freeres specifically to make glibc release all memory it has allocated. Memcheck therefore tries to run __libc_freeres at exit.

Unfortunately, in some very old versions of glibc, __libc_freeres is sufficiently buggy to cause segmentation faults. This was particularly noticeable on Red Hat 7.1. So this option is provided in order to inhibit the run of __libc_freeres. If your program seems to run fine on Valgrind, but segfaults at exit, you may find that --run-libc-freeres=no fixes that, although at the cost of possibly falsely reporting space leaks in libc.so.

--run-cxx-freeres=<yes|no> [default: yes]

This option is only relevant when running Valgrind on Linux or Solaris C++ programs.

The GNU Standard C++ library (libstdc++.so), which is used by all C++ programs compiled with g++, may allocate memory for its own uses. Usually it doesn't bother to free that memory when the program ends—there would be no point, since the kernel reclaims all process resources when a process exits anyway, so it would just slow things down.

The gcc authors realised that this behaviour causes leak checkers, such as Valgrind, to falsely report leaks in libstdc++, when a leak check is done at exit. In order to avoid this, they provided a routine called __gnu_cxx::__freeres specifically to make libstdc++ release all memory it has allocated. Memcheck therefore tries to run __gnu_cxx::__freeres at exit.

For the sake of flexibility and unforeseen problems with __gnu_cxx::__freeres, option --run-cxx-freeres=no exists, although at the cost of possibly falsely reporting space leaks in libstdc++.so.

--sim-hints=hint1,hint2,...

Pass miscellaneous hints to Valgrind which slightly modify the simulated behaviour in nonstandard or dangerous ways, possibly to help the simulation of strange features. By default no hints are enabled. Use with caution! Currently known hints are:

lax-ioctls: Be very lax about ioctl handling; the only assumption is that the size is correct. Doesn't require the full buffer to be initialised when writing. Without this, using some device drivers with a large number of strange ioctl commands becomes very tiresome.
fuse-compatible: Enable special handling for certain system calls that may block in a FUSE file-system. This may be necessary when running Valgrind on a multi-threaded program that uses one thread to manage a FUSE file-system and another thread to access that file-system.
enable-outer: Enable some special magic needed when the program being run is itself Valgrind.
no-inner-prefix: Disable printing a prefix > in front of each stdout or stderr output line in an inner Valgrind being run by an outer Valgrind. This is useful when running Valgrind regression tests in an outer/inner setup. Note that the prefix > will always be printed in front of the inner debug logging lines.
no-nptl-pthread-stackcache: This hint is only relevant when running Valgrind on Linux; it is ignored on Solaris and Mac OS X.

The GNU glibc pthread library (libpthread.so), which is used by pthread programs, maintains a cache of pthread stacks. When a pthread terminates, the memory used for the pthread stack and some thread local storage related data structure are not always directly released. This memory is kept in a cache (up to a certain size), and is re-used if a new thread is started.

This cache causes the helgrind tool to report some false positive race condition errors on this cached memory, as helgrind does not understand the internal glibc cache synchronisation primitives. So, when using helgrind, disabling the cache helps to avoid false positive race conditions, in particular when using thread local storage variables (e.g. variables using the __thread qualifier).

When using the memcheck tool, disabling the cache ensures the memory used by glibc to handle __thread variables is directly released when a thread terminates.

Note: Valgrind disables the cache using some internal knowledge of the glibc stack cache implementation and by examining the debug information of the pthread library. This technique is thus somewhat fragile and might not work for all glibc versions. This has been successfully tested with various glibc versions (e.g. 2.11, 2.16, 2.18) on various platforms.
lax-doors: (Solaris only) Be very lax about door syscall handling over unrecognised door file descriptors. Does not require that full buffer is initialised when writing. Without this, programs using libdoor(3LIB) functionality with completely proprietary semantics may report large number of false positives.
fallback-llsc: (MIPS and ARM64 only): Enables an alternative implementation of Load-Linked (LL) and Store-Conditional (SC) instructions. The standard implementation gives more correct behaviour, but can cause indefinite looping on certain processor implementations that are intolerant of extra memory references between LL and SC. So far this is known only to happen on Cavium 3 cores. You should not need to use this flag, since the relevant cores are detected at startup and the alternative implementation is automatically enabled if necessary. There is no equivalent anti-flag: you cannot force-disable the alternative implementation, if it is automatically enabled. The underlying problem exists because the "standard" implementation of LL and SC is done by copying through LL and SC instructions into the instrumented code. However, tools may insert extra instrumentation memory references in between the LL and SC instructions. These memory references are not present in the original uninstrumented code, and their presence in the instrumented code can cause the SC instructions to persistently fail, leading to indefinite looping in LL-SC blocks. The alternative implementation gives correct behaviour of LL and SC instructions between threads in a process, up to and including the ABA scenario. It also gives correct behaviour between a Valgrinded thread and a non-Valgrinded thread running in a different process, that communicate via shared memory, but only up to and including correct CAS behaviour -- in this case the ABA scenario may not be correctly handled.

--fair-sched=<no|yes|try> [default: no]

The --fair-sched option controls the locking mechanism used by Valgrind to serialise thread execution. The locking mechanism controls the way the threads are scheduled, and different settings give different trade-offs between fairness and performance. For more details about the Valgrind thread serialisation scheme and its impact on performance and thread scheduling, see Scheduling and Multi-Thread Performance.

The value --fair-sched=yes activates a fair scheduler. In short, if multiple threads are ready to run, the threads will be scheduled in a round robin fashion. This mechanism is not available on all platforms or Linux versions. If not available, using --fair-sched=yes will cause Valgrind to terminate with an error.

You may find this setting improves overall responsiveness if you are running an interactive multithreaded program, for example a web browser, on Valgrind.
The value --fair-sched=try activates fair scheduling if available on the platform. Otherwise, it will automatically fall back to --fair-sched=no.
The value --fair-sched=no activates a scheduler which does not guarantee fairness between threads ready to run, but which in general gives the highest performance.

--kernel-variant=variant1,variant2,...

Handle system calls and ioctls arising from minor variants of the default kernel for this platform. This is useful for running on hacked kernels or with kernel modules which support nonstandard ioctls, for example. Use with caution. If you don't understand what this option does then you almost certainly don't need it. Currently known variants are:

bproc: support the sys_broc system call on x86. This is for running on BProc, which is a minor variant of standard Linux which is sometimes used for building clusters.
android-no-hw-tls: some versions of the Android emulator for ARM do not provide a hardware TLS (thread-local state) register, and Valgrind crashes at startup. Use this variant to select software support for TLS.
android-gpu-sgx5xx: use this to support handling of proprietary ioctls for the PowerVR SGX 5XX series of GPUs on Android devices. Failure to select this does not cause stability problems, but may cause Memcheck to report false errors after the program performs GPU-specific ioctls.
android-gpu-adreno3xx: similarly, use this to support handling of proprietary ioctls for the Qualcomm Adreno 3XX series of GPUs on Android devices.

--merge-recursive-frames=<number> [default: 0]

Some recursive algorithms, for example balanced binary tree implementations, create many different stack traces, each containing cycles of calls. A cycle is defined as two identical program counter values separated by zero or more other program counter values. Valgrind may then use a lot of memory to store all these stack traces. This is a poor use of memory considering that such stack traces contain repeated uninteresting recursive calls instead of more interesting information such as the function that has initiated the recursive call.

The option --merge-recursive-frames=<number> instructs Valgrind to detect and merge recursive call cycles having a size of up to <number> frames. When such a cycle is detected, Valgrind records the cycle in the stack trace as a unique program counter.

The value 0 (the default) causes no recursive call merging. A value of 1 will cause stack traces of simple recursive algorithms (for example, a factorial implementation) to be collapsed. A value of 2 will usually be needed to collapse stack traces produced by recursive algorithms such as binary trees, quick sort, etc. Higher values might be needed for more complex recursive algorithms.

Note: recursive calls are detected by analysis of program counter values. They are not detected by looking at function names.

--num-transtab-sectors=<number> [default: 6 for Android platforms, 16 for all others]

Valgrind translates and instruments your program's machine code in small fragments (basic blocks). The translations are stored in a translation cache that is divided into a number of sections (sectors). If the cache is full, the sector containing the oldest translations is emptied and reused. If these old translations are needed again, Valgrind must re-translate and re-instrument the corresponding machine code, which is expensive. If the "executed instructions" working set of a program is big, increasing the number of sectors may improve performance by reducing the number of re-translations needed. Sectors are allocated on demand. Once allocated, a sector can never be freed, and occupies considerable space, depending on the tool and the value of --avg-transtab-entry-size (about 40 MB per sector for Memcheck). Use the option --stats=yes to obtain precise information about the memory used by a sector and the allocation and recycling of sectors.

--avg-transtab-entry-size=<number> [default: 0, meaning use tool provided default]

Average size of translated basic block. This average size is used to dimension the size of a sector. Each tool provides a default value to be used. If this default value is too small, the translation sectors will become full too quickly. If this default value is too big, a significant part of the translation sector memory will be unused. Note that the average size of a basic block translation depends on the tool, and might depend on tool options. For example, the memcheck option --track-origins=yes increases the size of the basic block translations. Use --avg-transtab-entry-size to tune the size of the sectors, either to gain memory or to avoid too many retranslations.

--aspace-minaddr=<address> [default: depends on the platform]

To avoid potential conflicts with some system libraries, Valgrind does not use the address space below --aspace-minaddr value, keeping it reserved in case a library specifically requests memory in this region. So, some "pessimistic" value is guessed by Valgrind depending on the platform. On linux, by default, Valgrind avoids using the first 64MB even if typically there is no conflict in this complete zone. You can use the option --aspace-minaddr to have your memory hungry application benefitting from more of this lower memory. On the other hand, if you encounter a conflict, increasing aspace-minaddr value might solve it. Conflicts will typically manifest themselves with mmap failures in the low range of the address space. The provided address must be page aligned and must be equal or bigger to 0x1000 (4KB). To find the default value on your platform, do something such as valgrind -d -d date 2>&1 | grep -i minaddr. Values lower than 0x10000 (64KB) are known to create problems on some distributions.

--valgrind-stacksize=<number> [default: 1MB]

For each thread, Valgrind needs its own 'private' stack. The default size for these stacks is largely dimensioned, and so should be sufficient in most cases. In case the size is too small, Valgrind will segfault. Before segfaulting, a warning might be produced by Valgrind when approaching the limit.

Use the option --valgrind-stacksize if such an (unlikely) warning is produced, or Valgrind dies due to a segmentation violation. Such segmentation violations have been seen when demangling huge C++ symbols.

If your application uses many threads and needs a lot of memory, you can gain some memory by reducing the size of these Valgrind stacks using the option --valgrind-stacksize.

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

When enabled, Valgrind will emit warnings about its CPU emulation in certain cases. These are usually not interesting.

--require-text-symbol=:sonamepatt:fnnamepatt

When a shared object whose soname matches sonamepatt is loaded into the process, examine all the text symbols it exports. If none of those match fnnamepatt, print an error message and abandon the run. This makes it possible to ensure that the run does not continue unless a given shared object contains a particular function name.

Both sonamepatt and fnnamepatt can be written using the usual ? and * wildcards. For example: ":*libc.so*:foo?bar". You may use characters other than a colon to separate the two patterns. It is only important that the first character and the separator character are the same. For example, the above example could also be written "Q*libc.so*Qfoo?bar". Multiple --require-text-symbol flags are allowed, in which case shared objects that are loaded into the process will be checked against all of them.

The purpose of this is to support reliable usage of marked-up libraries. For example, suppose we have a version of GCC's libgomp.so which has been marked up with annotations to support Helgrind. It is only too easy and confusing to load the wrong, un-annotated libgomp.so into the application. So the idea is: add a text symbol in the marked-up library, for example annotated_for_helgrind_3_6, and then give the flag --require-text-symbol=:*libgomp*so*:annotated_for_helgrind_3_6 so that when libgomp.so is loaded, Valgrind scans its symbol table, and if the symbol isn't present the run is aborted, rather than continuing silently with the un-marked-up library. Note that you should put the entire flag in quotes to stop shells expanding up the * and ? wildcards.

--soname-synonyms=syn1=pattern1,syn2=pattern2,...

When a shared library is loaded, Valgrind checks for functions in the library that must be replaced or wrapped. For example, Memcheck replaces some string and memory functions (strchr, strlen, strcpy, memchr, memcpy, memmove, etc.) with its own versions. Such replacements are normally done only in shared libraries whose soname matches a predefined soname pattern (e.g. libc.so* on linux). By default, no replacement is done for a statically linked binary or for alternative libraries, except for the allocation functions (malloc, free, calloc, memalign, realloc, operator new, operator delete, etc.) Such allocation functions are intercepted by default in any shared library or in the executable if they are exported as global symbols. This means that if a replacement allocation library such as tcmalloc is found, its functions are also intercepted by default. In some cases, the replacements allow --soname-synonyms to specify one additional synonym pattern, giving flexibility in the replacement. Or to prevent interception of all public allocation symbols.

Currently, this flexibility is only allowed for the malloc related functions, using the synonym somalloc. This synonym is usable for all tools doing standard replacement of malloc related functions (e.g. memcheck, helgrind, drd, massif, dhat).

Alternate malloc library: to replace the malloc related functions in a specific alternate library with soname mymalloclib.so (and not in any others), give the option --soname-synonyms=somalloc=mymalloclib.so. A pattern can be used to match multiple libraries sonames. For example, --soname-synonyms=somalloc=*tcmalloc* will match the soname of all variants of the tcmalloc library (native, debug, profiled, ... tcmalloc variants).

Note: the soname of a elf shared library can be retrieved using the readelf utility.
Replacements in a statically linked library are done by using the NONE pattern. For example, if you link with libtcmalloc.a, and only want to intercept the malloc related functions in the executable (and standard libraries) themselves, but not any other shared libraries, you can give the option --soname-synonyms=somalloc=NONE. Note that a NONE pattern will match the main executable and any shared library having no soname.
To run a "default" Firefox build for Linux, in which JEMalloc is linked in to the main executable, use --soname-synonyms=somalloc=NONE.
To only intercept allocation symbols in the default system libraries, but not in any other shared library or the executable defining public malloc or operator new related functions use a non-existing library name like --soname-synonyms=somalloc=nouserintercepts (where nouserintercepts can be any non-existing library name).
Shared library of the dynamic (runtime) linker is excluded from searching for global public symbols, such as those for the malloc related functions (identified by somalloc synonym).

--progress-interval=<number> [default: 0, meaning 'disabled']

This is an enhancement to Valgrind's debugging output. It is unlikely to be of interest to end users.

When number is set to a non-zero value, Valgrind will print a one-line progress summary every number seconds. Valid settings for number are between 0 and 3600 inclusive. Here's some example output with number set to 10:

PROGRESS: U 110s, W 113s, 97.3% CPU, EvC 414.79M, TIn 616.7k, TOut 0.5k, #thr 67
PROGRESS: U 120s, W 124s, 96.8% CPU, EvC 505.27M, TIn 636.6k, TOut 3.0k, #thr 64
PROGRESS: U 130s, W 134s, 97.0% CPU, EvC 574.90M, TIn 657.5k, TOut 3.0k, #thr 63

Each line shows:

U: total user time
W: total wallclock time
CPU: overall average cpu use
EvC: number of event checks. An event check is a backwards branch in the simulated program, so this is a measure of forward progress of the program
TIn: number of code blocks instrumented by the JIT
TOut: number of instrumented code blocks that have been thrown away
#thr: number of threads in the program

From the progress of these, it is possible to observe:

when the program is compute bound (TIn rises slowly, EvC rises rapidly)
when the program is in a spinloop (TIn/TOut fixed, EvC rises rapidly)
when the program is JIT-bound (TIn rises rapidly)
when the program is rapidly discarding code (TOut rises rapidly)
when the program is about to achieve some expected state (EvC arrives at some value you expect)
when the program is idling (U rises more slowly than W)

2.7.6. Debugging Options

There are also some options for debugging Valgrind itself. You shouldn't need to use them in the normal run of things. If you wish to see the list, use the --help-debug option.

If you wish to debug your program rather than debugging Valgrind itself, then you should use the options --vgdb=yes or --vgdb=full.

2.7.7. Setting Default Options

Note that Valgrind also reads options from three places:

The file ~/.valgrindrc
The environment variable $VALGRIND_OPTS
The file ./.valgrindrc

These are processed in the given order, before the command-line options. Options processed later override those processed earlier; for example, options in ./.valgrindrc will take precedence over those in ~/.valgrindrc.

Please note that the ./.valgrindrc file is ignored if it is not a regular file, or is marked as world writeable, or is not owned by the current user. This is because the ./.valgrindrc can contain options that are potentially harmful or can be used by a local attacker to execute code under your user account.

Any tool-specific options put in $VALGRIND_OPTS or the .valgrindrc files should be prefixed with the tool name and a colon. For example, if you want Memcheck to always do leak checking, you can put the following entry in ~/.valgrindrc:

--memcheck:leak-check=yes

This will be ignored if any tool other than Memcheck is run. Without the memcheck: part, this will cause problems if you select other tools that don't understand --leak-check=yes.

2.7.8. Dynamically Changing Options

The value of some command line options can be changed dynamically while your program is running under Valgrind.

The dynamically changeable options of the valgrind core and a given tool can be listed using option --help-dyn-options, for example:

$ valgrind --tool=memcheck --help-dyn-options
  dynamically changeable options:
    -v -q -d --stats --vgdb=no --vgdb=yes --vgdb=full --vgdb-poll --vgdb-error
    --vgdb-stop-at --error-markers --show-error-list -s --show-below-main
    --time-stamp --trace-children --child-silent-after-fork --trace-sched
    --trace-signals --trace-symtab --trace-cfi --debug-dump=syms
    --debug-dump=line --debug-dump=frames --trace-redir --trace-syscalls
    --sym-offsets --progress-interval --merge-recursive-frames
    --vex-iropt-verbosity --suppressions --trace-flags --trace-notbelow
    --trace-notabove --profile-flags --gen-suppressions=no
    --gen-suppressions=yes --gen-suppressions=all --errors-for-leak-kinds
    --show-leak-kinds --leak-check-heuristics --show-reachable
    --show-possibly-lost --freelist-vol --freelist-big-blocks --leak-check=no
    --leak-check=summary --leak-check=yes --leak-check=full --ignore-ranges
    --ignore-range-below-sp --show-mismatched-frees
valgrind: Use --help for more information.
$

The dynamic options can be changed the following ways:

From the shell, using vgdb and the monitor command v.clo:
```
$ vgdb "v.clo --trace-children=yes --child-silent-after-fork=no"
sending command v.clo --trace-children=yes --child-silent-after-fork=no to pid 4404
$
```
Note: you must use double quotes around the monitor command to avoid vgdb interpreting the valgrind options as its own options.

From gdb, using the the monitor command v.clo:

(gdb) monitor v.clo --trace-children=yes --child-silent-after-fork=no
(gdb)

From your program, using the client request VALGRIND_CLO_CHANGE(option):

      VALGRIND_CLO_CHANGE ("--trace-children=yes");
      VALGRIND_CLO_CHANGE ("--child-silent-after-fork=no");

Dynamically changeable options can be used in various circumstances, such as changing the valgrind behaviour during execution, loading suppression files as part of shared library initialisation, change or set valgrind options in child processes, ...

2.8. Support for Threads

Threaded programs are fully supported.

The main thing to point out with respect to threaded programs is that your program will use the native threading library, but Valgrind serialises execution so that only one (kernel) thread is running at a time. This approach avoids the horrible implementation problems of implementing a truly multithreaded version of Valgrind, but it does mean that threaded apps never use more than one CPU simultaneously, even if you have a multiprocessor or multicore machine.

Valgrind doesn't schedule the threads itself. It merely ensures that only one thread runs at once, using a simple locking scheme. The actual thread scheduling remains under control of the OS kernel. What this does mean, though, is that your program will see very different scheduling when run on Valgrind than it does when running normally. This is both because Valgrind is serialising the threads, and because the code runs so much slower than normal.

This difference in scheduling may cause your program to behave differently, if you have some kind of concurrency, critical race, locking, or similar, bugs. In that case you might consider using the tools Helgrind and/or DRD to track them down.

On Linux, Valgrind also supports direct use of the clone system call, futex and so on. clone is supported where either everything is shared (a thread) or nothing is shared (fork-like); partial sharing will fail.

2.8.1. Scheduling and Multi-Thread Performance

A thread executes code only when it holds the abovementioned lock. After executing some number of instructions, the running thread will release the lock. All threads ready to run will then compete to acquire the lock.

The --fair-sched option controls the locking mechanism used to serialise thread execution.

The default pipe based locking mechanism (--fair-sched=no) is available on all platforms. Pipe based locking does not guarantee fairness between threads: it is quite likely that a thread that has just released the lock reacquires it immediately, even though other threads are ready to run. When using pipe based locking, different runs of the same multithreaded application might give very different thread scheduling.

An alternative locking mechanism, based on futexes, is available on some platforms. If available, it is activated by --fair-sched=yes or --fair-sched=try. Futex based locking ensures fairness (round-robin scheduling) between threads: if multiple threads are ready to run, the lock will be given to the thread which first requested the lock. Note that a thread which is blocked in a system call (e.g. in a blocking read system call) has not (yet) requested the lock: such a thread requests the lock only after the system call is finished.

The fairness of the futex based locking produces better reproducibility of thread scheduling for different executions of a multithreaded application. This better reproducibility is particularly helpful when using Helgrind or DRD.

Valgrind's use of thread serialisation implies that only one thread at a time may run. On a multiprocessor/multicore system, the running thread is assigned to one of the CPUs by the OS kernel scheduler. When a thread acquires the lock, sometimes the thread will be assigned to the same CPU as the thread that just released the lock. Sometimes, the thread will be assigned to another CPU. When using pipe based locking, the thread that just acquired the lock will usually be scheduled on the same CPU as the thread that just released the lock. With the futex based mechanism, the thread that just acquired the lock will more often be scheduled on another CPU.

Valgrind's thread serialisation and CPU assignment by the OS kernel scheduler can interact badly with the CPU frequency scaling available on many modern CPUs. To decrease power consumption, the frequency of a CPU or core is automatically decreased if the CPU/core has not been used recently. If the OS kernel often assigns the thread which just acquired the lock to another CPU/core, it is quite likely that this CPU/core is currently at a low frequency. The frequency of this CPU will be increased after some time. However, during this time, the (only) running thread will have run at the low frequency. Once this thread has run for some time, it will release the lock. Another thread will acquire this lock, and might be scheduled again on another CPU whose clock frequency was decreased in the meantime.

The futex based locking causes threads to change CPUs/cores more often. So, if CPU frequency scaling is activated, the futex based locking might decrease significantly the performance of a multithreaded app running under Valgrind. Performance losses of up to 50% degradation have been observed, as compared to running on a machine for which CPU frequency scaling has been disabled. The pipe based locking locking scheme also interacts badly with CPU frequency scaling, with performance losses in the range 10..20% having been observed.

To avoid such performance degradation, you should indicate to the kernel that all CPUs/cores should always run at maximum clock speed. Depending on your Linux distribution, CPU frequency scaling may be controlled using a graphical interface or using command line such as cpufreq-selector or cpufreq-set.

An alternative way to avoid these problems is to tell the OS scheduler to tie a Valgrind process to a specific (fixed) CPU using the taskset command. This should ensure that the selected CPU does not fall below its maximum frequency setting so long as any thread of the program has work to do.

2.9. Handling of Signals

Valgrind has a fairly complete signal implementation. It should be able to cope with any POSIX-compliant use of signals.

If you're using signals in clever ways (for example, catching SIGSEGV, modifying page state and restarting the instruction), you're probably relying on precise exceptions. In this case, you will need to use --vex-iropt-register-updates=allregs-at-mem-access or --vex-iropt-register-updates=allregs-at-each-insn.

If your program dies as a result of a fatal core-dumping signal, Valgrind will generate its own core file (vgcore.NNNNN) containing your program's state. You may use this core file for post-mortem debugging with GDB or similar. (Note: it will not generate a core if your core dump size limit is 0.) At the time of writing the core dumps do not include all the floating point register information.

In the unlikely event that Valgrind itself crashes, the operating system will create a core dump in the usual way.

2.10. Execution Trees

An execution tree (xtree) is made of a set of stack traces, each stack trace is associated with some resource consumptions or event counts. Depending on the xtree, different event counts/resource consumptions can be recorded in the xtree. Multiple tools can produce memory use xtree. Memcheck can output the leak search results in an xtree.

A typical usage for an xtree is to show a graphical or textual representation of the heap usage of a program. The below figure is a heap usage xtree graphical representation produced by kcachegrind. In the kcachegrind output, you can see that main current heap usage (allocated indirectly) is 528 bytes : 388 bytes allocated indirectly via a call to function f1 and 140 bytes indirectly allocated via a call to function f2. f2 has allocated memory by calling g2, while f1 has allocated memory by calling g11 and g12. g11, g12 and g1 have directly called a memory allocation function (malloc), and so have a non zero 'Self' value. Note that when kcachegrind shows an xtree, the 'Called' column and call nr indications in the Call Graph are not significant (always set to 0 or 1, independently of the real nr of calls. The kcachegrind versions >= 0.8.0 do not show anymore such irrelevant xtree call number information.

An xtree heap memory report is produced at the end of the execution when required using the option --xtree-memory. It can also be produced on demand using the xtmemory monitor command (see Valgrind monitor commands). Currently, an xtree heap memory report can be produced by the memcheck, helgrind and massif tools.

The xtrees produced by the option --xtree-memory or the xtmemory monitor command are showing the following events/resource consumption describing heap usage:

curB current number of Bytes allocated. The number of allocated bytes is added to the curB value of a stack trace for each allocation. It is decreased when a block allocated by this stack trace is released (by another "freeing" stack trace)
curBk current number of Blocks allocated, maintained similary to curB : +1 for each allocation, -1 when the block is freed.
totB total allocated Bytes. This is increased for each allocation with the number of allocated bytes.
totBk total allocated Blocks, maintained similary to totB : +1 for each allocation.
totFdB total Freed Bytes, increased each time a block is released by this ("freeing") stack trace : + nr freed bytes for each free operation.
totFdBk total Freed Blocks, maintained similarly to totFdB : +1 for each free operation.

Note that the last 4 counts are produced only when the --xtree-memory=full was given at startup.

Xtrees can be saved in 2 file formats, the "Callgrind Format" and the "Massif Format".

Callgrind Format

An xtree file in the Callgrind Format contains a single callgraph, associating each stack trace with the values recorded in the xtree.

Different Callgrind Format file visualisers are available:

Valgrind distribution includes the callgrind_annotate command line utility that reads in the xtree data, and prints a sorted lists of functions, optionally with source annotation. Note that due to xtree specificities, you must give the option --inclusive=yes to callgrind_annotate.

For graphical visualization of the data, you can use KCachegrind, which is a KDE/Qt based GUI that makes it easy to navigate the large amount of data that an xtree can contain.

Note that xtree Callgrind Format does not make use of the inline information even when specifying --read-inline-info=yes.
Massif Format

An xtree file in the Massif Format contains one detailed tree callgraph data for each type of event recorded in the xtree. So, for --xtree-memory=alloc, the output file will contain 2 detailed trees (for the counts curB and curBk), while --xtree-memory=full will give a file with 6 detailed trees.

Different Massif Format file visualisers are available. Valgrind distribution includes the ms_print command line utility that produces an easy to read reprentation of a massif output file. See Using Massif and ms_print and Using massif-visualizer for more details about visualising Massif Format output files.

Note that xtree Massif Format makes use of the inline information when specifying --read-inline-info=yes.

Note that for equivalent information, the Callgrind Format is more compact than the Massif Format. However, the Callgrind Format always contains the full data: there is no filtering done during file production, filtering is done by visualisers such as kcachegrind. kcachegrind is particularly easy to use to analyse big xtree data containing multiple events counts or resources consumption. The Massif Format (optionally) only contains a part of the data. For example, the Massif tool might filter some of the data, according to the --threshold option.

To clarify the xtree concept, the below gives several extracts of the output produced by the following commands:

valgrind --xtree-memory=full --xtree-memory-file=xtmemory.kcg mfg
callgrind_annotate --auto=yes --inclusive=yes --sort=curB:100,curBk:100,totB:100,totBk:100,totFdB:100,totFdBk:100  xtmemory.kcg

The below extract shows that the program mfg has allocated in total 770 bytes in 60 different blocks. Of these 60 blocks, 19 were freed, releasing a total of 242 bytes. The heap currently contains 528 bytes in 41 blocks.

--------------------------------------------------------------------------------
curB curBk totB totBk totFdB totFdBk 
--------------------------------------------------------------------------------
 528    41  770    60    242      19  PROGRAM TOTALS

The below gives more details about which functions have allocated or released memory. As an example, we see that main has (directly or indirectly) allocated 770 bytes of memory and freed (directly or indirectly) 242 bytes of memory. The function f1 has (directly or indirectly) allocated 570 bytes of memory, and has not (directly or indirectly) freed memory. Of the 570 bytes allocated by function f1, 388 bytes (34 blocks) have not been released.

--------------------------------------------------------------------------------
curB curBk totB totBk totFdB totFdBk  file:function
--------------------------------------------------------------------------------
 528    41  770    60    242      19  mfg.c:main
 388    34  570    50      0       0  mfg.c:f1
 220    20  330    30      0       0  mfg.c:g11
 168    14  240    20      0       0  mfg.c:g12
 140     7  200    10      0       0  mfg.c:g2
 140     7  200    10      0       0  mfg.c:f2
   0     0    0     0    131      10  mfg.c:freeY
   0     0    0     0    111       9  mfg.c:freeX

The below gives a more detailed information about the callgraph and which source lines/calls have (directly or indirectly) allocated or released memory. The below shows that the 770 bytes allocated by main have been indirectly allocated by calls to f1 and f2. Similarly, we see that the 570 bytes allocated by f1 have been indirectly allocated by calls to g11 and g12. Of the 330 bytes allocated by the 30 calls to g11, 168 bytes have not been freed. The function freeY (called once by main) has released in total 10 blocks and 131 bytes.

--------------------------------------------------------------------------------
-- Auto-annotated source: /home/philippe/valgrind/littleprogs/ + mfg.c
--------------------------------------------------------------------------------
curB curBk totB totBk totFdB totFdBk 
....
   .     .    .     .      .       .  static void freeY(void)
   .     .    .     .      .       .  {
   .     .    .     .      .       .     int i;
   .     .    .     .      .       .     for (i = 0; i < next_ptr; i++)
   .     .    .     .      .       .        if(i % 5 == 0 && ptrs[i] != NULL)
   0     0    0     0    131      10           free(ptrs[i]);
   .     .    .     .      .       .  }
   .     .    .     .      .       .  static void f1(void)
   .     .    .     .      .       .  {
   .     .    .     .      .       .     int i;
   .     .    .     .      .       .     for (i = 0; i < 30; i++)
 220    20  330    30      0       0        g11();
   .     .    .     .      .       .     for (i = 0; i < 20; i++)
 168    14  240    20      0       0        g12();
   .     .    .     .      .       .  }
   .     .    .     .      .       .  int main()
   .     .    .     .      .       .  {
 388    34  570    50      0       0     f1();
 140     7  200    10      0       0     f2();
   0     0    0     0    111       9     freeX();
   0     0    0     0    131      10     freeY();
   .     .    .     .      .       .     return 0;
   .     .    .     .      .       .  }

Heap memory xtrees are helping to understand how your (big) program is using the heap. A full heap memory xtree helps to pin point some code that allocates a lot of small objects : allocating such small objects might be replaced by more efficient technique, such as allocating a big block using malloc, and then diviving this block into smaller blocks in order to decrease the cpu and/or memory overhead of allocating a lot of small blocks. Such full xtree information complements e.g. what callgrind can show: callgrind can show the number of calls to a function (such as malloc) but does not indicate the volume of memory allocated (or freed).

A full heap memory xtree also can identify the code that allocates and frees a lot of blocks : the total foot print of the program might not reflect the fact that the same memory was over and over allocated then released.

Finally, Xtree visualisers such as kcachegrind are helping to identify big memory consumers, in order to possibly optimise the amount of memory needed by your program.

2.11. Building and Installing Valgrind

We use the standard Unix ./configure, make, make install mechanism. Once you have completed make install you may then want to run the regression tests with make regtest.

In addition to the usual --prefix=/path/to/install/tree, there are three options which affect how Valgrind is built:

--enable-inner

This builds Valgrind with some special magic hacks which make it possible to run it on a standard build of Valgrind (what the developers call "self-hosting"). Ordinarily you should not use this option as various kinds of safety checks are disabled.
--enable-only64bit

--enable-only32bit

On 64-bit platforms (amd64-linux, ppc64-linux, amd64-darwin), Valgrind is by default built in such a way that both 32-bit and 64-bit executables can be run. Sometimes this cleverness is a problem for a variety of reasons. These two options allow for single-target builds in this situation. If you issue both, the configure script will complain. Note they are ignored on 32-bit-only platforms (x86-linux, ppc32-linux, arm-linux, x86-darwin).

The configure script tests the version of the X server currently indicated by the current $DISPLAY. This is a known bug. The intention was to detect the version of the current X client libraries, so that correct suppressions could be selected for them, but instead the test checks the server version. This is just plain wrong.

If you are building a binary package of Valgrind for distribution, please read README_PACKAGERS Readme Packagers. It contains some important information.

Apart from that, there's not much excitement here. Let us know if you have build problems.

2.12. If You Have Problems

See Limitations for the known limitations of Valgrind, and for a list of programs which are known not to work on it.

All parts of the system make heavy use of assertions and internal self-checks. They are permanently enabled, and we have no plans to disable them. If one of them breaks, please mail us!

If you get an assertion failure in m_mallocfree.c, this may have happened because your program wrote off the end of a heap block, or before its beginning, thus corrupting heap metadata. Valgrind hopefully will have emitted a message to that effect before dying in this way.

Read the Valgrind FAQ for more advice about common problems, crashes, etc.

2.13. Limitations

The following list of limitations seems long. However, most programs actually work fine.

Valgrind will run programs on the supported platforms subject to the following constraints:

On Linux, Valgrind determines at startup the size of the 'brk segment' using the RLIMIT_DATA rlim_cur, with a minimum of 1 MB and a maximum of 8 MB. Valgrind outputs a message each time a program tries to extend the brk segment beyond the size determined at startup. Most programs will work properly with this limit, typically by switching to the use of mmap to get more memory. If your program really needs a big brk segment, you must change the 8 MB hardcoded limit and recompile Valgrind.
On x86 and amd64, there is no support for 3DNow! instructions. If the translator encounters these, Valgrind will generate a SIGILL when the instruction is executed. Apart from that, on x86 and amd64, essentially all instructions are supported, up to and including AVX and AES in 64-bit mode and SSSE3 in 32-bit mode. 32-bit mode does in fact support the bare minimum SSE4 instructions needed to run programs on MacOSX 10.6 on 32-bit targets.
On ppc32 and ppc64, almost all integer, floating point and Altivec instructions are supported. Specifically: integer and FP insns that are mandatory for PowerPC, the "General-purpose optional" group (fsqrt, fsqrts, stfiwx), the "Graphics optional" group (fre, fres, frsqrte, frsqrtes), and the Altivec (also known as VMX) SIMD instruction set, are supported. Also, instructions from the Power ISA 2.05 specification, as present in POWER6 CPUs, are supported.
On ARM, essentially the entire ARMv7-A instruction set is supported, in both ARM and Thumb mode. ThumbEE and Jazelle are not supported. NEON, VFPv3 and ARMv6 media support is fairly complete.
If your program does its own memory management, rather than using malloc/new/free/delete, it should still work, but Memcheck's error checking won't be so effective. If you describe your program's memory management scheme using "client requests" (see The Client Request mechanism), Memcheck can do better. Nevertheless, using malloc/new and free/delete is still the best approach.
Valgrind's signal simulation is not as robust as it could be. Basic POSIX-compliant sigaction and sigprocmask functionality is supplied, but it's conceivable that things could go badly awry if you do weird things with signals. Workaround: don't. Programs that do non-POSIX signal tricks are in any case inherently unportable, so should be avoided if possible.
Machine instructions, and system calls, have been implemented on demand. So it's possible, although unlikely, that a program will fall over with a message to that effect. If this happens, please report all the details printed out, so we can try and implement the missing feature.
Memory consumption of your program is majorly increased whilst running under Valgrind's Memcheck tool. This is due to the large amount of administrative information maintained behind the scenes. Another cause is that Valgrind dynamically translates the original executable. Translated, instrumented code is 12-18 times larger than the original so you can easily end up with 150+ MB of translations when running (eg) a web browser.
Valgrind can handle dynamically-generated code just fine. If you regenerate code over the top of old code (ie. at the same memory addresses), if the code is on the stack Valgrind will realise the code has changed, and work correctly. This is necessary to handle the trampolines GCC uses to implemented nested functions. If you regenerate code somewhere other than the stack, and you are running on an 32- or 64-bit x86 CPU, you will need to use the --smc-check=all option, and Valgrind will run more slowly than normal. Or you can add client requests that tell Valgrind when your program has overwritten code.

On other platforms (ARM, PowerPC) Valgrind observes and honours the cache invalidation hints that programs are obliged to emit to notify new code, and so self-modifying-code support should work automatically, without the need for --smc-check=all.
Valgrind has the following limitations in its implementation of x86/AMD64 floating point relative to IEEE754.

Precision: There is no support for 80 bit arithmetic. Internally, Valgrind represents all such "long double" numbers in 64 bits, and so there may be some differences in results. Whether or not this is critical remains to be seen. Note, the x86/amd64 fldt/fstpt instructions (read/write 80-bit numbers) are correctly simulated, using conversions to/from 64 bits, so that in-memory images of 80-bit numbers look correct if anyone wants to see.

The impression observed from many FP regression tests is that the accuracy differences aren't significant. Generally speaking, if a program relies on 80-bit precision, there may be difficulties porting it to non x86/amd64 platforms which only support 64-bit FP precision. Even on x86/amd64, the program may get different results depending on whether it is compiled to use SSE2 instructions (64-bits only), or x87 instructions (80-bit). The net effect is to make FP programs behave as if they had been run on a machine with 64-bit IEEE floats, for example PowerPC. On amd64 FP arithmetic is done by default on SSE2, so amd64 looks more like PowerPC than x86 from an FP perspective, and there are far fewer noticeable accuracy differences than with x86.

Rounding: Valgrind does observe the 4 IEEE-mandated rounding modes (to nearest, to +infinity, to -infinity, to zero) for the following conversions: float to integer, integer to float where there is a possibility of loss of precision, and float-to-float rounding. For all other FP operations, only the IEEE default mode (round to nearest) is supported.

Numeric exceptions in FP code: IEEE754 defines five types of numeric exception that can happen: invalid operation (sqrt of negative number, etc), division by zero, overflow, underflow, inexact (loss of precision).

For each exception, two courses of action are defined by IEEE754: either (1) a user-defined exception handler may be called, or (2) a default action is defined, which "fixes things up" and allows the computation to proceed without throwing an exception.

Currently Valgrind only supports the default fixup actions. Again, feedback on the importance of exception support would be appreciated.

When Valgrind detects that the program is trying to exceed any of these limitations (setting exception handlers, rounding mode, or precision control), it can print a message giving a traceback of where this has happened, and continue execution. This behaviour used to be the default, but the messages are annoying and so showing them is now disabled by default. Use --show-emwarns=yes to see them.

The above limitations define precisely the IEEE754 'default' behaviour: default fixup on all exceptions, round-to-nearest operations, and 64-bit precision.
Valgrind has the following limitations in its implementation of x86/AMD64 SSE2 FP arithmetic, relative to IEEE754.

Essentially the same: no exceptions, and limited observance of rounding mode. Also, SSE2 has control bits which make it treat denormalised numbers as zero (DAZ) and a related action, flush denormals to zero (FTZ). Both of these cause SSE2 arithmetic to be less accurate than IEEE requires. Valgrind detects, ignores, and can warn about, attempts to enable either mode.
Valgrind has the following limitations in its implementation of ARM VFPv3 arithmetic, relative to IEEE754.

Essentially the same: no exceptions, and limited observance of rounding mode. Also, switching the VFP unit into vector mode will cause Valgrind to abort the program -- it has no way to emulate vector uses of VFP at a reasonable performance level. This is no big deal given that non-scalar uses of VFP instructions are in any case deprecated.
Valgrind has the following limitations in its implementation of PPC32 and PPC64 floating point arithmetic, relative to IEEE754.

Scalar (non-Altivec): Valgrind provides a bit-exact emulation of all floating point instructions, except for "fre" and "fres", which are done more precisely than required by the PowerPC architecture specification. All floating point operations observe the current rounding mode.

However, fpscr[FPRF] is not set after each operation. That could be done but would give measurable performance overheads, and so far no need for it has been found.

As on x86/AMD64, IEEE754 exceptions are not supported: all floating point exceptions are handled using the default IEEE fixup actions. Valgrind detects, ignores, and can warn about, attempts to unmask the 5 IEEE FP exception kinds by writing to the floating-point status and control register (fpscr).

Vector (Altivec, VMX): essentially as with x86/AMD64 SSE/SSE2: no exceptions, and limited observance of rounding mode. For Altivec, FP arithmetic is done in IEEE/Java mode, which is more accurate than the Linux default setting. "More accurate" means that denormals are handled properly, rather than simply being flushed to zero.

Programs which are known not to work are:

emacs starts up but immediately concludes it is out of memory and aborts. It may be that Memcheck does not provide a good enough emulation of the mallinfo function. Emacs works fine if you build it to use the standard malloc/free routines.

2.14. An Example Run

This is the log for a run of a small program using Memcheck. The program is in fact correct, and the reported error is as the result of a potentially serious code generation bug in GNU g++ (snapshot 20010527).

sewardj@phoenix:~/newmat10$ ~/Valgrind-6/valgrind -v ./bogon 
==25832== Valgrind 0.10, a memory error detector for x86 RedHat 7.1.
==25832== Copyright (C) 2000-2001, and GNU GPL'd, by Julian Seward.
==25832== Startup, with flags:
==25832== --suppressions=/home/sewardj/Valgrind/redhat71.supp
==25832== reading syms from /lib/ld-linux.so.2
==25832== reading syms from /lib/libc.so.6
==25832== reading syms from /mnt/pima/jrs/Inst/lib/libgcc_s.so.0
==25832== reading syms from /lib/libm.so.6
==25832== reading syms from /mnt/pima/jrs/Inst/lib/libstdc++.so.3
==25832== reading syms from /home/sewardj/Valgrind/valgrind.so
==25832== reading syms from /proc/self/exe
==25832== 
==25832== Invalid read of size 4
==25832==    at 0x8048724: BandMatrix::ReSize(int,int,int) (bogon.cpp:45)
==25832==    by 0x80487AF: main (bogon.cpp:66)
==25832==  Address 0xBFFFF74C is not stack'd, malloc'd or free'd
==25832==
==25832== ERROR SUMMARY: 1 errors from 1 contexts (suppressed: 0 from 0)
==25832== malloc/free: in use at exit: 0 bytes in 0 blocks.
==25832== malloc/free: 0 allocs, 0 frees, 0 bytes allocated.
==25832== For a detailed leak analysis, rerun with: --leak-check=yes

The GCC folks fixed this about a week before GCC 3.0 shipped.

2.15. Warning Messages You Might See

Some of these only appear if you run in verbose mode (enabled by -v):

More than 100 errors detected. Subsequent errors will still be recorded, but in less detail than before.

After 100 different errors have been shown, Valgrind becomes more conservative about collecting them. It then requires only the program counters in the top two stack frames to match when deciding whether or not two errors are really the same one. Prior to this point, the PCs in the top four frames are required to match. This hack has the effect of slowing down the appearance of new errors after the first 100. The 100 constant can be changed by recompiling Valgrind.
More than 1000 errors detected. I'm not reporting any more. Final error counts may be inaccurate. Go fix your program!

After 1000 different errors have been detected, Valgrind ignores any more. It seems unlikely that collecting even more different ones would be of practical help to anybody, and it avoids the danger that Valgrind spends more and more of its time comparing new errors against an ever-growing collection. As above, the 1000 number is a compile-time constant.
Warning: client switching stacks?

Valgrind spotted such a large change in the stack pointer that it guesses the client is switching to a different stack. At this point it makes a kludgey guess where the base of the new stack is, and sets memory permissions accordingly. At the moment "large change" is defined as a change of more that 2000000 in the value of the stack pointer register. If Valgrind guesses wrong, you may get many bogus error messages following this and/or have crashes in the stack trace recording code. You might avoid these problems by informing Valgrind about the stack bounds using VALGRIND_STACK_REGISTER client request.
Warning: client attempted to close Valgrind's logfile fd <number>

Valgrind doesn't allow the client to close the logfile, because you'd never see any diagnostic information after that point. If you see this message, you may want to use the --log-fd=<number> option to specify a different logfile file-descriptor number.
Warning: noted but unhandled ioctl <number>

Valgrind observed a call to one of the vast family of ioctl system calls, but did not modify its memory status info (because nobody has yet written a suitable wrapper). The call will still have gone through, but you may get spurious errors after this as a result of the non-update of the memory info.
Warning: set address range perms: large range <number>

Diagnostic message, mostly for benefit of the Valgrind developers, to do with memory permissions.