• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1<?xml version="1.0"?> <!-- -*- sgml -*- -->
2<!DOCTYPE chapter PUBLIC "-//OASIS//DTD DocBook XML V4.2//EN"
3          "http://www.oasis-open.org/docbook/xml/4.2/docbookx.dtd">
4
5
6<chapter id="mc-manual" xreflabel="Memcheck: a memory error detector">
7<title>Memcheck: a memory error detector</title>
8
9<para>To use this tool, you may specify <option>--tool=memcheck</option>
10on the Valgrind command line.  You don't have to, though, since Memcheck
11is the default tool.</para>
12
13
14<sect1 id="mc-manual.overview" xreflabel="Overview">
15<title>Overview</title>
16
17<para>Memcheck is a memory error detector.  It can detect the following
18problems that are common in C and C++ programs.</para>
19
20<itemizedlist>
21  <listitem>
22    <para>Accessing memory you shouldn't, e.g. overrunning and underrunning
23    heap blocks, overrunning the top of the stack, and accessing memory after
24    it has been freed.</para>
25  </listitem>
26
27  <listitem>
28    <para>Using undefined values, i.e. values that have not been initialised,
29    or that have been derived from other undefined values.</para>
30  </listitem>
31
32  <listitem>
33    <para>Incorrect freeing of heap memory, such as double-freeing heap
34    blocks, or mismatched use of
35    <function>malloc</function>/<computeroutput>new</computeroutput>/<computeroutput>new[]</computeroutput>
36    versus
37    <function>free</function>/<computeroutput>delete</computeroutput>/<computeroutput>delete[]</computeroutput></para>
38  </listitem>
39
40  <listitem>
41    <para>Overlapping <computeroutput>src</computeroutput> and
42    <computeroutput>dst</computeroutput> pointers in
43    <computeroutput>memcpy</computeroutput> and related
44    functions.</para>
45  </listitem>
46
47  <listitem>
48    <para>Passing a fishy (presumably negative) value to the
49    <computeroutput>size</computeroutput> parameter of a memory
50    allocation function.</para>
51  </listitem>
52
53  <listitem>
54    <para>Memory leaks.</para>
55  </listitem>
56</itemizedlist>
57
58<para>Problems like these can be difficult to find by other means,
59often remaining undetected for long periods, then causing occasional,
60  difficult-to-diagnose crashes.</para>
61
62<para>Memcheck also provides <xref linkend="manual-core.xtree"/> memory
63  profiling using the command line
64  option <computeroutput>--xtree-memory</computeroutput> and the monitor command
65   <computeroutput>xtmemory</computeroutput>.</para>
66</sect1>
67
68
69
70<sect1 id="mc-manual.errormsgs"
71       xreflabel="Explanation of error messages from Memcheck">
72<title>Explanation of error messages from Memcheck</title>
73
74<para>Memcheck issues a range of error messages.  This section presents a
75quick summary of what error messages mean.  The precise behaviour of the
76error-checking machinery is described in <xref
77linkend="mc-manual.machine"/>.</para>
78
79
80<sect2 id="mc-manual.badrw"
81       xreflabel="Illegal read / Illegal write errors">
82<title>Illegal read / Illegal write errors</title>
83
84<para>For example:</para>
85<programlisting><![CDATA[
86Invalid read of size 4
87   at 0x40F6BBCC: (within /usr/lib/libpng.so.2.1.0.9)
88   by 0x40F6B804: (within /usr/lib/libpng.so.2.1.0.9)
89   by 0x40B07FF4: read_png_image(QImageIO *) (kernel/qpngio.cpp:326)
90   by 0x40AC751B: QImageIO::read() (kernel/qimage.cpp:3621)
91 Address 0xBFFFF0E0 is not stack'd, malloc'd or free'd
92]]></programlisting>
93
94<para>This happens when your program reads or writes memory at a place
95which Memcheck reckons it shouldn't.  In this example, the program did a
964-byte read at address 0xBFFFF0E0, somewhere within the system-supplied
97library libpng.so.2.1.0.9, which was called from somewhere else in the
98same library, called from line 326 of <filename>qpngio.cpp</filename>,
99and so on.</para>
100
101<para>Memcheck tries to establish what the illegal address might relate
102to, since that's often useful.  So, if it points into a block of memory
103which has already been freed, you'll be informed of this, and also where
104the block was freed.  Likewise, if it should turn out to be just off
105the end of a heap block, a common result of off-by-one-errors in
106array subscripting, you'll be informed of this fact, and also where the
107block was allocated.  If you use the <option><xref
108linkend="opt.read-var-info"/></option> option Memcheck will run more slowly
109but may give a more detailed description of any illegal address.</para>
110
111<para>In this example, Memcheck can't identify the address.  Actually
112the address is on the stack, but, for some reason, this is not a valid
113stack address -- it is below the stack pointer and that isn't allowed.
114In this particular case it's probably caused by GCC generating invalid
115code, a known bug in some ancient versions of GCC.</para>
116
117<para>Note that Memcheck only tells you that your program is about to
118access memory at an illegal address.  It can't stop the access from
119happening.  So, if your program makes an access which normally would
120result in a segmentation fault, you program will still suffer the same
121fate -- but you will get a message from Memcheck immediately prior to
122this.  In this particular example, reading junk on the stack is
123non-fatal, and the program stays alive.</para>
124
125</sect2>
126
127
128
129<sect2 id="mc-manual.uninitvals"
130       xreflabel="Use of uninitialised values">
131<title>Use of uninitialised values</title>
132
133<para>For example:</para>
134<programlisting><![CDATA[
135Conditional jump or move depends on uninitialised value(s)
136   at 0x402DFA94: _IO_vfprintf (_itoa.h:49)
137   by 0x402E8476: _IO_printf (printf.c:36)
138   by 0x8048472: main (tests/manuel1.c:8)
139]]></programlisting>
140
141<para>An uninitialised-value use error is reported when your program
142uses a value which hasn't been initialised -- in other words, is
143undefined.  Here, the undefined value is used somewhere inside the
144<function>printf</function> machinery of the C library.  This error was
145reported when running the following small program:</para>
146<programlisting><![CDATA[
147int main()
148{
149  int x;
150  printf ("x = %d\n", x);
151}]]></programlisting>
152
153<para>It is important to understand that your program can copy around
154junk (uninitialised) data as much as it likes.  Memcheck observes this
155and keeps track of the data, but does not complain.  A complaint is
156issued only when your program attempts to make use of uninitialised
157data in a way that might affect your program's externally-visible behaviour.
158In this example, <varname>x</varname> is uninitialised.  Memcheck observes
159the value being passed to <function>_IO_printf</function> and thence to
160<function>_IO_vfprintf</function>, but makes no comment.  However,
161<function>_IO_vfprintf</function> has to examine the value of
162<varname>x</varname> so it can turn it into the corresponding ASCII string,
163and it is at this point that Memcheck complains.</para>
164
165<para>Sources of uninitialised data tend to be:</para>
166<itemizedlist>
167  <listitem>
168    <para>Local variables in procedures which have not been initialised,
169    as in the example above.</para>
170  </listitem>
171  <listitem>
172    <para>The contents of heap blocks (allocated with
173    <function>malloc</function>, <function>new</function>, or a similar
174    function) before you (or a constructor) write something there.
175    </para>
176  </listitem>
177</itemizedlist>
178
179<para>To see information on the sources of uninitialised data in your
180program, use the <option>--track-origins=yes</option> option.  This
181makes Memcheck run more slowly, but can make it much easier to track down
182the root causes of uninitialised value errors.</para>
183
184</sect2>
185
186
187
188<sect2 id="mc-manual.bad-syscall-args"
189       xreflabel="Use of uninitialised or unaddressable values in system
190       calls">
191<title>Use of uninitialised or unaddressable values in system
192       calls</title>
193
194<para>Memcheck checks all parameters to system calls:
195<itemizedlist>
196  <listitem>
197    <para>It checks all the direct parameters themselves, whether they are
198    initialised.</para>
199  </listitem>
200  <listitem>
201    <para>Also, if a system call needs to read from a buffer provided by
202    your program, Memcheck checks that the entire buffer is addressable
203    and its contents are initialised.</para>
204  </listitem>
205  <listitem>
206    <para>Also, if the system call needs to write to a user-supplied
207    buffer, Memcheck checks that the buffer is addressable.</para>
208  </listitem>
209</itemizedlist>
210</para>
211
212<para>After the system call, Memcheck updates its tracked information to
213precisely reflect any changes in memory state caused by the system
214call.</para>
215
216<para>Here's an example of two system calls with invalid parameters:</para>
217<programlisting><![CDATA[
218  #include <stdlib.h>
219  #include <unistd.h>
220  int main( void )
221  {
222    char* arr  = malloc(10);
223    int*  arr2 = malloc(sizeof(int));
224    write( 1 /* stdout */, arr, 10 );
225    exit(arr2[0]);
226  }
227]]></programlisting>
228
229<para>You get these complaints ...</para>
230<programlisting><![CDATA[
231  Syscall param write(buf) points to uninitialised byte(s)
232     at 0x25A48723: __write_nocancel (in /lib/tls/libc-2.3.3.so)
233     by 0x259AFAD3: __libc_start_main (in /lib/tls/libc-2.3.3.so)
234     by 0x8048348: (within /auto/homes/njn25/grind/head4/a.out)
235   Address 0x25AB8028 is 0 bytes inside a block of size 10 alloc'd
236     at 0x259852B0: malloc (vg_replace_malloc.c:130)
237     by 0x80483F1: main (a.c:5)
238
239  Syscall param exit(error_code) contains uninitialised byte(s)
240     at 0x25A21B44: __GI__exit (in /lib/tls/libc-2.3.3.so)
241     by 0x8048426: main (a.c:8)
242]]></programlisting>
243
244<para>... because the program has (a) written uninitialised junk
245from the heap block to the standard output, and (b) passed an
246uninitialised value to <function>exit</function>.  Note that the first
247error refers to the memory pointed to by
248<computeroutput>buf</computeroutput> (not
249<computeroutput>buf</computeroutput> itself), but the second error
250refers directly to <computeroutput>exit</computeroutput>'s argument
251<computeroutput>arr2[0]</computeroutput>.</para>
252
253</sect2>
254
255
256<sect2 id="mc-manual.badfrees" xreflabel="Illegal frees">
257<title>Illegal frees</title>
258
259<para>For example:</para>
260<programlisting><![CDATA[
261Invalid free()
262   at 0x4004FFDF: free (vg_clientmalloc.c:577)
263   by 0x80484C7: main (tests/doublefree.c:10)
264 Address 0x3807F7B4 is 0 bytes inside a block of size 177 free'd
265   at 0x4004FFDF: free (vg_clientmalloc.c:577)
266   by 0x80484C7: main (tests/doublefree.c:10)
267]]></programlisting>
268
269<para>Memcheck keeps track of the blocks allocated by your program
270with <function>malloc</function>/<computeroutput>new</computeroutput>,
271so it can know exactly whether or not the argument to
272<function>free</function>/<computeroutput>delete</computeroutput> is
273legitimate or not.  Here, this test program has freed the same block
274twice.  As with the illegal read/write errors, Memcheck attempts to
275make sense of the address freed.  If, as here, the address is one
276which has previously been freed, you wil be told that -- making
277duplicate frees of the same block easy to spot.  You will also get this
278message if you try to free a pointer that doesn't point to the start of a
279heap block.</para>
280
281</sect2>
282
283
284<sect2 id="mc-manual.rudefn"
285       xreflabel="When a heap block is freed with an inappropriate deallocation
286function">
287<title>When a heap block is freed with an inappropriate deallocation
288function</title>
289
290<para>In the following example, a block allocated with
291<function>new[]</function> has wrongly been deallocated with
292<function>free</function>:</para>
293<programlisting><![CDATA[
294Mismatched free() / delete / delete []
295   at 0x40043249: free (vg_clientfuncs.c:171)
296   by 0x4102BB4E: QGArray::~QGArray(void) (tools/qgarray.cpp:149)
297   by 0x4C261C41: PptDoc::~PptDoc(void) (include/qmemarray.h:60)
298   by 0x4C261F0E: PptXml::~PptXml(void) (pptxml.cc:44)
299 Address 0x4BB292A8 is 0 bytes inside a block of size 64 alloc'd
300   at 0x4004318C: operator new[](unsigned int) (vg_clientfuncs.c:152)
301   by 0x4C21BC15: KLaola::readSBStream(int) const (klaola.cc:314)
302   by 0x4C21C155: KLaola::stream(KLaola::OLENode const *) (klaola.cc:416)
303   by 0x4C21788F: OLEFilter::convert(QCString const &) (olefilter.cc:272)
304]]></programlisting>
305
306<para>In <literal>C++</literal> it's important to deallocate memory in a
307way compatible with how it was allocated.  The deal is:</para>
308<itemizedlist>
309  <listitem>
310    <para>If allocated with
311    <function>malloc</function>,
312    <function>calloc</function>,
313    <function>realloc</function>,
314    <function>valloc</function> or
315    <function>memalign</function>, you must
316    deallocate with <function>free</function>.</para>
317  </listitem>
318  <listitem>
319   <para>If allocated with <function>new</function>, you must deallocate
320   with <function>delete</function>.</para>
321  </listitem>
322  <listitem>
323    <para>If allocated with <function>new[]</function>, you must
324    deallocate with <function>delete[]</function>.</para>
325  </listitem>
326</itemizedlist>
327
328<para>The worst thing is that on Linux apparently it doesn't matter if
329you do mix these up, but the same program may then crash on a
330different platform, Solaris for example.  So it's best to fix it
331properly.  According to the KDE folks "it's amazing how many C++
332programmers don't know this".</para>
333
334<para>The reason behind the requirement is as follows.  In some C++
335implementations, <function>delete[]</function> must be used for
336objects allocated by <function>new[]</function> because the compiler
337stores the size of the array and the pointer-to-member to the
338destructor of the array's content just before the pointer actually
339returned.  <function>delete</function> doesn't account for this and will get
340confused, possibly corrupting the heap.</para>
341
342</sect2>
343
344
345
346<sect2 id="mc-manual.overlap"
347       xreflabel="Overlapping source and destination blocks">
348<title>Overlapping source and destination blocks</title>
349
350<para>The following C library functions copy some data from one
351memory block to another (or something similar):
352<function>memcpy</function>,
353<function>strcpy</function>,
354<function>strncpy</function>,
355<function>strcat</function>,
356<function>strncat</function>.
357The blocks pointed to by their <computeroutput>src</computeroutput> and
358<computeroutput>dst</computeroutput> pointers aren't allowed to overlap.
359The POSIX standards have wording along the lines "If copying takes place
360between objects that overlap, the behavior is undefined." Therefore,
361Memcheck checks for this.
362</para>
363
364<para>For example:</para>
365<programlisting><![CDATA[
366==27492== Source and destination overlap in memcpy(0xbffff294, 0xbffff280, 21)
367==27492==    at 0x40026CDC: memcpy (mc_replace_strmem.c:71)
368==27492==    by 0x804865A: main (overlap.c:40)
369]]></programlisting>
370
371<para>You don't want the two blocks to overlap because one of them could
372get partially overwritten by the copying.</para>
373
374<para>You might think that Memcheck is being overly pedantic reporting
375this in the case where <computeroutput>dst</computeroutput> is less than
376<computeroutput>src</computeroutput>.  For example, the obvious way to
377implement <function>memcpy</function> is by copying from the first
378byte to the last.  However, the optimisation guides of some
379architectures recommend copying from the last byte down to the first.
380Also, some implementations of <function>memcpy</function> zero
381<computeroutput>dst</computeroutput> before copying, because zeroing the
382destination's cache line(s) can improve performance.</para>
383
384<para>The moral of the story is: if you want to write truly portable
385code, don't make any assumptions about the language
386implementation.</para>
387
388</sect2>
389
390
391<sect2 id="mc-manual.fishyvalue"
392       xreflabel="Fishy argument values">
393<title>Fishy argument values</title>
394
395<para>All memory allocation functions take an argument specifying the
396size of the memory block that should be allocated. Clearly, the requested
397size should be a non-negative value and is typically not excessively large.
398For instance, it is extremely unlikly that the size of an allocation
399request exceeds 2**63 bytes on a 64-bit machine. It is much more likely that
400such a value is the result of an erroneous size calculation and is in effect
401a negative value (that just happens to appear excessively large because
402the bit pattern is interpreted as an unsigned integer).
403Such a value is called a "fishy value".
404
405The <varname>size</varname> argument of the following allocation functions
406is checked for being fishy:
407<function>malloc</function>,
408<function>calloc</function>,
409<function>realloc</function>,
410<function>memalign</function>,
411<function>new</function>,
412<function>new []</function>.
413<function>__builtin_new</function>,
414<function>__builtin_vec_new</function>,
415For <function>calloc</function> both arguments are being checked.
416</para>
417
418<para>For example:</para>
419<programlisting><![CDATA[
420==32233== Argument 'size' of function malloc has a fishy (possibly negative) value: -3
421==32233==    at 0x4C2CFA7: malloc (vg_replace_malloc.c:298)
422==32233==    by 0x400555: foo (fishy.c:15)
423==32233==    by 0x400583: main (fishy.c:23)
424]]></programlisting>
425
426<para>In earlier Valgrind versions those values were being referred to
427as "silly arguments" and no back-trace was included.
428</para>
429
430</sect2>
431
432
433<sect2 id="mc-manual.leaks" xreflabel="Memory leak detection">
434<title>Memory leak detection</title>
435
436<para>Memcheck keeps track of all heap blocks issued in response to
437calls to
438<function>malloc</function>/<function>new</function> et al.
439So when the program exits, it knows which blocks have not been freed.
440</para>
441
442<para>If <option>--leak-check</option> is set appropriately, for each
443remaining block, Memcheck determines if the block is reachable from pointers
444within the root-set.  The root-set consists of (a) general purpose registers
445of all threads, and (b) initialised, aligned, pointer-sized data words in
446accessible client memory, including stacks.</para>
447
448<para>There are two ways a block can be reached.  The first is with a
449"start-pointer", i.e. a pointer to the start of the block.  The second is with
450an "interior-pointer", i.e. a pointer to the middle of the block.  There are
451several ways we know of that an interior-pointer can occur:</para>
452
453<itemizedlist>
454  <listitem>
455    <para>The pointer might have originally been a start-pointer and have been
456    moved along deliberately (or not deliberately) by the program.  In
457    particular, this can happen if your program uses tagged pointers, i.e.
458    if it uses the bottom one, two or three bits of a pointer, which are
459    normally always zero due to alignment, in order to store extra
460    information.</para>
461  </listitem>
462
463  <listitem>
464    <para>It might be a random junk value in memory, entirely unrelated, just
465    a coincidence.</para>
466  </listitem>
467
468  <listitem>
469    <para>It might be a pointer to the inner char array of a C++
470    <computeroutput>std::string</computeroutput>.  For example, some
471    compilers add 3 words at the beginning of the std::string to
472    store the length, the capacity and a reference count before the
473    memory containing the array of characters. They return a pointer
474    just after these 3 words, pointing at the char array.</para>
475  </listitem>
476
477  <listitem>
478    <para>Some code might allocate a block of memory, and use the first 8
479    bytes to store (block size - 8) as a 64bit number.
480    <computeroutput>sqlite3MemMalloc</computeroutput> does this.</para>
481  </listitem>
482
483  <listitem>
484    <para>It might be a pointer to an array of C++ objects (which possess
485    destructors) allocated with <computeroutput>new[]</computeroutput>.  In
486    this case, some compilers store a "magic cookie" containing the array
487    length at the start of the allocated block, and return a pointer to just
488    past that magic cookie, i.e. an interior-pointer.
489    See <ulink url="http://theory.uwinnipeg.ca/gnu/gcc/gxxint_14.html">this
490    page</ulink> for more information.</para>
491  </listitem>
492
493  <listitem>
494    <para>It might be a pointer to an inner part of a C++ object using
495    multiple inheritance. </para>
496  </listitem>
497</itemizedlist>
498
499<para>You can optionally activate heuristics to use during the leak
500search to detect the interior pointers corresponding to
501the <computeroutput>stdstring</computeroutput>,
502<computeroutput>length64</computeroutput>,
503<computeroutput>newarray</computeroutput>
504and <computeroutput>multipleinheritance</computeroutput> cases. If the
505heuristic detects that an interior pointer corresponds to such a case,
506the block will be considered as reachable by the interior
507pointer. In other words, the interior pointer will be treated
508as if it were a start pointer.</para>
509
510
511<para>With that in mind, consider the nine possible cases described by the
512following figure.</para>
513
514<programlisting><![CDATA[
515     Pointer chain            AAA Leak Case   BBB Leak Case
516     -------------            -------------   -------------
517(1)  RRR ------------> BBB                    DR
518(2)  RRR ---> AAA ---> BBB    DR              IR
519(3)  RRR               BBB                    DL
520(4)  RRR      AAA ---> BBB    DL              IL
521(5)  RRR ------?-----> BBB                    (y)DR, (n)DL
522(6)  RRR ---> AAA -?-> BBB    DR              (y)IR, (n)DL
523(7)  RRR -?-> AAA ---> BBB    (y)DR, (n)DL    (y)IR, (n)IL
524(8)  RRR -?-> AAA -?-> BBB    (y)DR, (n)DL    (y,y)IR, (n,y)IL, (_,n)DL
525(9)  RRR      AAA -?-> BBB    DL              (y)IL, (n)DL
526
527Pointer chain legend:
528- RRR: a root set node or DR block
529- AAA, BBB: heap blocks
530- --->: a start-pointer
531- -?->: an interior-pointer
532
533Leak Case legend:
534- DR: Directly reachable
535- IR: Indirectly reachable
536- DL: Directly lost
537- IL: Indirectly lost
538- (y)XY: it's XY if the interior-pointer is a real pointer
539- (n)XY: it's XY if the interior-pointer is not a real pointer
540- (_)XY: it's XY in either case
541]]></programlisting>
542
543<para>Every possible case can be reduced to one of the above nine.  Memcheck
544merges some of these cases in its output, resulting in the following four
545leak kinds.</para>
546
547
548<itemizedlist>
549
550  <listitem>
551    <para>"Still reachable". This covers cases 1 and 2 (for the BBB blocks)
552    above.  A start-pointer or chain of start-pointers to the block is
553    found.  Since the block is still pointed at, the programmer could, at
554    least in principle, have freed it before program exit.  "Still reachable"
555    blocks are very common and arguably not a problem. So, by default,
556    Memcheck won't report such blocks individually.</para>
557  </listitem>
558
559  <listitem>
560    <para>"Definitely lost".  This covers case 3 (for the BBB blocks) above.
561    This means that no pointer to the block can be found.  The block is
562    classified as "lost", because the programmer could not possibly have
563    freed it at program exit, since no pointer to it exists.  This is likely
564    a symptom of having lost the pointer at some earlier point in the
565    program.  Such cases should be fixed by the programmer.</para>
566    </listitem>
567
568  <listitem>
569    <para>"Indirectly lost".  This covers cases 4 and 9 (for the BBB blocks)
570    above.  This means that the block is lost, not because there are no
571    pointers to it, but rather because all the blocks that point to it are
572    themselves lost.  For example, if you have a binary tree and the root
573    node is lost, all its children nodes will be indirectly lost.  Because
574    the problem will disappear if the definitely lost block that caused the
575    indirect leak is fixed, Memcheck won't report such blocks individually
576    by default.</para>
577  </listitem>
578
579  <listitem>
580    <para>"Possibly lost".  This covers cases 5--8 (for the BBB blocks)
581    above.  This means that a chain of one or more pointers to the block has
582    been found, but at least one of the pointers is an interior-pointer.
583    This could just be a random value in memory that happens to point into a
584    block, and so you shouldn't consider this ok unless you know you have
585    interior-pointers.</para>
586  </listitem>
587
588</itemizedlist>
589
590<para>(Note: This mapping of the nine possible cases onto four leak kinds is
591not necessarily the best way that leaks could be reported;  in particular,
592interior-pointers are treated inconsistently.  It is possible the
593categorisation may be improved in the future.)</para>
594
595<para>Furthermore, if suppressions exists for a block, it will be reported
596as "suppressed" no matter what which of the above four kinds it belongs
597to.</para>
598
599
600<para>The following is an example leak summary.</para>
601
602<programlisting><![CDATA[
603LEAK SUMMARY:
604   definitely lost: 48 bytes in 3 blocks.
605   indirectly lost: 32 bytes in 2 blocks.
606     possibly lost: 96 bytes in 6 blocks.
607   still reachable: 64 bytes in 4 blocks.
608        suppressed: 0 bytes in 0 blocks.
609]]></programlisting>
610
611<para>If heuristics have been used to consider some blocks as
612reachable, the leak summary details the heuristically reachable subset
613of 'still reachable:' per heuristic. In the below example, of the 95
614bytes still reachable, 87 bytes (56+7+8+16) have been considered
615heuristically reachable.
616</para>
617
618<programlisting><![CDATA[
619LEAK SUMMARY:
620   definitely lost: 4 bytes in 1 blocks
621   indirectly lost: 0 bytes in 0 blocks
622     possibly lost: 0 bytes in 0 blocks
623   still reachable: 95 bytes in 6 blocks
624                      of which reachable via heuristic:
625                        stdstring          : 56 bytes in 2 blocks
626                        length64           : 16 bytes in 1 blocks
627                        newarray           : 7 bytes in 1 blocks
628                        multipleinheritance: 8 bytes in 1 blocks
629        suppressed: 0 bytes in 0 blocks
630]]></programlisting>
631
632<para>If <option>--leak-check=full</option> is specified,
633Memcheck will give details for each definitely lost or possibly lost block,
634including where it was allocated.  (Actually, it merges results for all
635blocks that have the same leak kind and sufficiently similar stack traces
636into a single "loss record".  The
637<option>--leak-resolution</option> lets you control the
638meaning of "sufficiently similar".)  It cannot tell you when or how or why
639the pointer to a leaked block was lost; you have to work that out for
640yourself.  In general, you should attempt to ensure your programs do not
641have any definitely lost or possibly lost blocks at exit.</para>
642
643<para>For example:</para>
644<programlisting><![CDATA[
6458 bytes in 1 blocks are definitely lost in loss record 1 of 14
646   at 0x........: malloc (vg_replace_malloc.c:...)
647   by 0x........: mk (leak-tree.c:11)
648   by 0x........: main (leak-tree.c:39)
649
65088 (8 direct, 80 indirect) bytes in 1 blocks are definitely lost in loss record 13 of 14
651   at 0x........: malloc (vg_replace_malloc.c:...)
652   by 0x........: mk (leak-tree.c:11)
653   by 0x........: main (leak-tree.c:25)
654]]></programlisting>
655
656<para>The first message describes a simple case of a single 8 byte block
657that has been definitely lost.  The second case mentions another 8 byte
658block that has been definitely lost;  the difference is that a further 80
659bytes in other blocks are indirectly lost because of this lost block.
660The loss records are not presented in any notable order, so the loss record
661numbers aren't particularly meaningful. The loss record numbers can be used
662in the Valgrind gdbserver to list the addresses of the leaked blocks and/or give
663more details about how a block is still reachable.</para>
664
665<para>The option <option>--show-leak-kinds=&lt;set&gt;</option>
666controls the set of leak kinds to show
667when <option>--leak-check=full</option> is specified. </para>
668
669<para>The <option>&lt;set&gt;</option> of leak kinds is specified
670in one of the following ways:
671
672<itemizedlist>
673  <listitem><para>a comma separated list of one or more of
674    <option>definite indirect possible reachable</option>.</para>
675  </listitem>
676
677  <listitem><para><option>all</option> to specify the complete set (all leak kinds).</para>
678  </listitem>
679
680  <listitem><para><option>none</option> for the empty set.</para>
681  </listitem>
682</itemizedlist>
683
684</para>
685
686<para> The default value for the leak kinds to show is
687  <option>--show-leak-kinds=definite,possible</option>.
688</para>
689
690<para>To also show the reachable and indirectly lost blocks in
691addition to the definitely and possibly lost blocks, you can
692use <option>--show-leak-kinds=all</option>.  To only show the
693reachable and indirectly lost blocks, use
694<option>--show-leak-kinds=indirect,reachable</option>.  The reachable
695and indirectly lost blocks will then be presented as shown in
696the following two examples.</para>
697
698<programlisting><![CDATA[
69964 bytes in 4 blocks are still reachable in loss record 2 of 4
700   at 0x........: malloc (vg_replace_malloc.c:177)
701   by 0x........: mk (leak-cases.c:52)
702   by 0x........: main (leak-cases.c:74)
703
70432 bytes in 2 blocks are indirectly lost in loss record 1 of 4
705   at 0x........: malloc (vg_replace_malloc.c:177)
706   by 0x........: mk (leak-cases.c:52)
707   by 0x........: main (leak-cases.c:80)
708]]></programlisting>
709
710<para>Because there are different kinds of leaks with different
711severities, an interesting question is: which leaks should be
712counted as true "errors" and which should not?
713</para>
714
715<para> The answer to this question affects the numbers printed in
716the <computeroutput>ERROR SUMMARY</computeroutput> line, and also the
717effect of the <option>--error-exitcode</option> option.  First, a leak
718is only counted as a true "error"
719if <option>--leak-check=full</option> is specified.  Then, the
720option <option>--errors-for-leak-kinds=&lt;set&gt;</option> controls
721the set of leak kinds to consider as errors.  The default value
722is <option>--errors-for-leak-kinds=definite,possible</option>
723</para>
724
725</sect2>
726
727</sect1>
728
729
730
731<sect1 id="mc-manual.options"
732       xreflabel="Memcheck Command-Line Options">
733<title>Memcheck Command-Line Options</title>
734
735<!-- start of xi:include in the manpage -->
736<variablelist id="mc.opts.list">
737
738  <varlistentry id="opt.leak-check" xreflabel="--leak-check">
739    <term>
740      <option><![CDATA[--leak-check=<no|summary|yes|full> [default: summary] ]]></option>
741    </term>
742    <listitem>
743      <para>When enabled, search for memory leaks when the client
744      program finishes.  If set to <varname>summary</varname>, it says how
745      many leaks occurred.  If set to <varname>full</varname> or
746      <varname>yes</varname>, each individual leak will be shown
747      in detail and/or counted as an error, as specified by the options
748      <option>--show-leak-kinds</option> and
749      <option>--errors-for-leak-kinds</option>. </para>
750    </listitem>
751  </varlistentry>
752
753  <varlistentry id="opt.leak-resolution" xreflabel="--leak-resolution">
754    <term>
755      <option><![CDATA[--leak-resolution=<low|med|high> [default: high] ]]></option>
756    </term>
757    <listitem>
758      <para>When doing leak checking, determines how willing
759      Memcheck is to consider different backtraces to
760      be the same for the purposes of merging multiple leaks into a single
761      leak report.  When set to <varname>low</varname>, only the first
762      two entries need match.  When <varname>med</varname>, four entries
763      have to match.  When <varname>high</varname>, all entries need to
764      match.</para>
765
766      <para>For hardcore leak debugging, you probably want to use
767      <option>--leak-resolution=high</option> together with
768      <option>--num-callers=40</option> or some such large number.
769      </para>
770
771      <para>Note that the <option>--leak-resolution</option> setting
772      does not affect Memcheck's ability to find
773      leaks.  It only changes how the results are presented.</para>
774    </listitem>
775  </varlistentry>
776
777  <varlistentry id="opt.show-leak-kinds" xreflabel="--show-leak-kinds">
778    <term>
779      <option><![CDATA[--show-leak-kinds=<set> [default: definite,possible] ]]></option>
780    </term>
781    <listitem>
782      <para>Specifies the leak kinds to show in a <varname>full</varname>
783      leak search, in one of the following ways: </para>
784
785      <itemizedlist>
786        <listitem><para>a comma separated list of one or more of
787            <option>definite indirect possible reachable</option>.</para>
788        </listitem>
789
790        <listitem><para><option>all</option> to specify the complete set (all leak kinds).
791            It is equivalent to
792            <option>--show-leak-kinds=definite,indirect,possible,reachable</option>.</para>
793        </listitem>
794
795        <listitem><para><option>none</option> for the empty set.</para>
796        </listitem>
797      </itemizedlist>
798    </listitem>
799  </varlistentry>
800
801
802  <varlistentry id="opt.errors-for-leak-kinds" xreflabel="--errors-for-leak-kinds">
803    <term>
804      <option><![CDATA[--errors-for-leak-kinds=<set> [default: definite,possible] ]]></option>
805    </term>
806    <listitem>
807      <para>Specifies the leak kinds to count as errors in a
808        <varname>full</varname> leak search. The
809        <option><![CDATA[<set>]]></option> is specified similarly to
810        <option>--show-leak-kinds</option>
811      </para>
812    </listitem>
813  </varlistentry>
814
815
816  <varlistentry id="opt.leak-check-heuristics" xreflabel="--leak-check-heuristics">
817    <term>
818      <option><![CDATA[--leak-check-heuristics=<set> [default: all] ]]></option>
819    </term>
820    <listitem>
821      <para>Specifies the set of leak check heuristics to be used
822        during leak searches.  The heuristics control which interior pointers
823        to a block cause it to be considered as reachable.
824        The heuristic set is specified in one of the following ways:</para>
825
826      <itemizedlist>
827        <listitem><para>a comma separated list of one or more of
828            <option>stdstring length64 newarray multipleinheritance</option>.</para>
829        </listitem>
830
831        <listitem><para><option>all</option> to activate the complete set of
832            heuristics.
833            It is equivalent to
834            <option>--leak-check-heuristics=stdstring,length64,newarray,multipleinheritance</option>.</para>
835        </listitem>
836
837        <listitem><para><option>none</option> for the empty set.</para>
838        </listitem>
839      </itemizedlist>
840    </listitem>
841
842    <para>Note that these heuristics are dependent on the layout of the objects
843      produced by the C++ compiler. They have been tested with some gcc versions
844      (e.g. 4.4 and 4.7). They might not work properly with other C++ compilers.
845    </para>
846  </varlistentry>
847
848
849  <varlistentry id="opt.show-reachable" xreflabel="--show-reachable">
850    <term>
851      <option><![CDATA[--show-reachable=<yes|no> ]]></option>
852    </term>
853    <term>
854      <option><![CDATA[--show-possibly-lost=<yes|no> ]]></option>
855    </term>
856    <listitem>
857      <para>These options provide an alternative way to specify the leak kinds to show:
858      </para>
859      <itemizedlist>
860        <listitem>
861	  <para>
862            <option>--show-reachable=no --show-possibly-lost=yes</option> is equivalent to
863            <option>--show-leak-kinds=definite,possible</option>.
864	  </para>
865        </listitem>
866        <listitem>
867	  <para>
868            <option>--show-reachable=no --show-possibly-lost=no</option> is equivalent to
869            <option>--show-leak-kinds=definite</option>.
870	  </para>
871        </listitem>
872        <listitem>
873	  <para>
874            <option>--show-reachable=yes</option> is equivalent to
875            <option>--show-leak-kinds=all</option>.
876	  </para>
877        </listitem>
878      </itemizedlist>
879    </listitem>
880    <para> Note that  <option>--show-possibly-lost=no</option> has no effect
881      if <option>--show-reachable=yes</option> is specified.</para>
882  </varlistentry>
883
884  <varlistentry id="opt.xtree-leak" xreflabel="--xtree-leak">
885    <term>
886      <option><![CDATA[--xtree-leak=<no|yes> [no] ]]></option>
887    </term>
888    <listitem>
889      <para>If set to yes, the results for the leak search done at exit will be
890        output in a 'Callgrind Format' execution tree file. Note that this
891        automatically sets the option <option>--leak-check=full</option>.
892        The produced file
893       will contain the following events:</para>
894      <itemizedlist>
895        <listitem><para><option>RB</option> : Reachable Bytes</para></listitem>
896         <listitem><para><option>PB</option> : Possibly lost Bytes</para></listitem>
897         <listitem><para><option>IB</option> : Indirectly lost Bytes</para></listitem>
898         <listitem><para><option>DB</option> : Definitely lost Bytes (direct plus indirect)</para></listitem>
899         <listitem><para><option>DIB</option> : Definitely Indirectly lost Bytes (subset of DB)</para></listitem>
900         <listitem><para><option>RBk</option> : reachable Blocks</para></listitem>
901         <listitem><para><option>PBk</option> : Possibly lost Blocks</para></listitem>
902         <listitem><para><option>IBk</option> : Indirectly lost Blocks</para></listitem>
903         <listitem><para><option>DBk</option> : Definitely lost Blocks</para></listitem>
904      </itemizedlist>
905
906      <para>The increase or decrease for all events above will also be output in
907        the file to provide the delta (increase or decreaseà between 2
908        successive leak searches. For example, <option>iRB</option> is the
909        increase of the <option>RB</option> event, <option>dPBk</option> is the
910        decrease of <option>PBk</option> event. The values for the increase and
911        decrease events will be zero for the first leak search done.</para>
912
913      <para>See <xref linkend="manual-core.xtree"/> for a detailed explanation
914        about execution trees.</para>
915    </listitem>
916  </varlistentry>
917
918  <varlistentry id="opt.xtree-leak-file" xreflabel="--xtree-leak-file">
919    <term>
920      <option><![CDATA[--xtree-leak-file=<filename> [default:
921      xtleak.kcg.%p] ]]></option>
922    </term>
923    <listitem>
924      <para>Specifies that Valgrind should produce the xtree leak
925        report in the specified file.  Any <option>%p</option>,
926        <option>%q</option> or  <option>%n</option> sequences appearing in
927        the filename are expanded
928        in exactly the same way as they are for <option>--log-file</option>.
929        See the description of <xref linkend="opt.log-file"/>
930        for details. </para>
931      <para>See <xref linkend="manual-core.xtree"/>
932      for a detailed explanation about execution trees formats. </para>
933    </listitem>
934  </varlistentry>
935
936  <varlistentry id="opt.undef-value-errors" xreflabel="--undef-value-errors">
937    <term>
938      <option><![CDATA[--undef-value-errors=<yes|no> [default: yes] ]]></option>
939    </term>
940    <listitem>
941      <para>Controls whether Memcheck reports
942      uses of undefined value errors.  Set this to
943      <varname>no</varname> if you don't want to see undefined value
944      errors.  It also has the side effect of speeding up
945      Memcheck somewhat.
946      </para>
947    </listitem>
948  </varlistentry>
949
950  <varlistentry id="opt.track-origins" xreflabel="--track-origins">
951    <term>
952      <option><![CDATA[--track-origins=<yes|no> [default: no] ]]></option>
953    </term>
954      <listitem>
955        <para>Controls whether Memcheck tracks
956        the origin of uninitialised values.  By default, it does not,
957        which means that although it can tell you that an
958        uninitialised value is being used in a dangerous way, it
959        cannot tell you where the uninitialised value came from.  This
960        often makes it difficult to track down the root problem.
961        </para>
962        <para>When set
963        to <varname>yes</varname>, Memcheck keeps
964        track of the origins of all uninitialised values.  Then, when
965        an uninitialised value error is
966        reported, Memcheck will try to show the
967        origin of the value.  An origin can be one of the following
968        four places: a heap block, a stack allocation, a client
969        request, or miscellaneous other sources (eg, a call
970        to <varname>brk</varname>).
971        </para>
972        <para>For uninitialised values originating from a heap
973        block, Memcheck shows where the block was
974        allocated.  For uninitialised values originating from a stack
975        allocation, Memcheck can tell you which
976        function allocated the value, but no more than that -- typically
977        it shows you the source location of the opening brace of the
978        function.  So you should carefully check that all of the
979        function's local variables are initialised properly.
980        </para>
981        <para>Performance overhead: origin tracking is expensive.  It
982        halves Memcheck's speed and increases
983        memory use by a minimum of 100MB, and possibly more.
984        Nevertheless it can drastically reduce the effort required to
985        identify the root cause of uninitialised value errors, and so
986        is often a programmer productivity win, despite running
987        more slowly.
988        </para>
989        <para>Accuracy: Memcheck tracks origins
990        quite accurately.  To avoid very large space and time
991        overheads, some approximations are made.  It is possible,
992        although unlikely, that Memcheck will report an incorrect origin, or
993        not be able to identify any origin.
994        </para>
995        <para>Note that the combination
996        <option>--track-origins=yes</option>
997        and <option>--undef-value-errors=no</option> is
998        nonsensical.  Memcheck checks for and
999        rejects this combination at startup.
1000        </para>
1001      </listitem>
1002  </varlistentry>
1003
1004  <varlistentry id="opt.partial-loads-ok" xreflabel="--partial-loads-ok">
1005    <term>
1006      <option><![CDATA[--partial-loads-ok=<yes|no> [default: yes] ]]></option>
1007    </term>
1008    <listitem>
1009      <para>Controls how Memcheck handles 32-, 64-, 128- and 256-bit
1010      naturally aligned loads from addresses for which some bytes are
1011      addressable and others are not.  When <varname>yes</varname>, such
1012      loads do not produce an address error.  Instead, loaded bytes
1013      originating from illegal addresses are marked as uninitialised, and
1014      those corresponding to legal addresses are handled in the normal
1015      way.</para>
1016
1017      <para>When <varname>no</varname>, loads from partially invalid
1018      addresses are treated the same as loads from completely invalid
1019      addresses: an illegal-address error is issued, and the resulting
1020      bytes are marked as initialised.</para>
1021
1022      <para>Note that code that behaves in this way is in violation of
1023      the ISO C/C++ standards, and should be considered broken.  If
1024      at all possible, such code should be fixed.</para>
1025    </listitem>
1026  </varlistentry>
1027
1028  <varlistentry id="opt.expensive-definedness-checks" xreflabel="--expensive-definedness-checks">
1029    <term>
1030      <option><![CDATA[--expensive-definedness-checks=<yes|no> [default: no] ]]></option>
1031    </term>
1032    <listitem>
1033      <para>Controls whether Memcheck should employ more precise but also more
1034      expensive (time consuming) algorithms when checking the definedness of a
1035      value. The default setting is not to do that and it is usually
1036      sufficient. However, for highly optimised code valgrind may sometimes
1037      incorrectly complain.
1038      Invoking valgrind with <option>--expensive-definedness-checks=yes</option>
1039      helps but comes at a performance cost. Runtime degradation of
1040      25% have been observed but the extra cost depends a lot on the
1041      application at hand.
1042      </para>
1043    </listitem>
1044  </varlistentry>
1045
1046  <varlistentry id="opt.keep-stacktraces" xreflabel="--keep-stacktraces">
1047    <term>
1048      <option><![CDATA[--keep-stacktraces=alloc|free|alloc-and-free|alloc-then-free|none [default: alloc-and-free] ]]></option>
1049    </term>
1050    <listitem>
1051      <para>Controls which stack trace(s) to keep for malloc'd and/or
1052      free'd blocks.
1053      </para>
1054
1055      <para>With <varname>alloc-then-free</varname>, a stack trace is
1056      recorded at allocation time, and is associated with the block.
1057      When the block is freed, a second stack trace is recorded, and
1058      this replaces the allocation stack trace.  As a result, any "use
1059      after free" errors relating to this block can only show a stack
1060      trace for where the block was freed.
1061      </para>
1062
1063      <para>With <varname>alloc-and-free</varname>, both allocation
1064      and the deallocation stack traces for the block are stored.
1065      Hence a "use after free" error will
1066      show both, which may make the error easier to diagnose.
1067      Compared to <varname>alloc-then-free</varname>, this setting
1068      slightly increases Valgrind's memory use as the block contains two
1069      references instead of one.
1070      </para>
1071
1072      <para>With <varname>alloc</varname>, only the allocation stack
1073      trace is recorded (and reported).  With <varname>free</varname>,
1074      only the deallocation stack trace is recorded (and reported).
1075      These values somewhat decrease Valgrind's memory and cpu usage.
1076      They can be useful depending on the error types you are
1077      searching for and the level of detail you need to analyse
1078      them.  For example, if you are only interested in memory leak
1079      errors, it is sufficient to record the allocation stack traces.
1080      </para>
1081
1082      <para>With <varname>none</varname>, no stack traces are recorded
1083      for malloc and free operations. If your program allocates a lot
1084      of blocks and/or allocates/frees from many different stack
1085      traces, this can significantly decrease cpu and/or memory
1086      required.  Of course, few details will be reported for errors
1087      related to heap blocks.
1088      </para>
1089
1090      <para>Note that once a stack trace is recorded, Valgrind keeps
1091      the stack trace in memory even if it is not referenced by any
1092      block.  Some programs (for example, recursive algorithms) can
1093      generate a huge number of stack traces. If Valgrind uses too
1094      much memory in such circumstances, you can reduce the memory
1095      required with the options <varname>--keep-stacktraces</varname>
1096      and/or by using a smaller value for the
1097      option <varname>--num-callers</varname>.
1098      </para>
1099
1100      <para>If you want to use
1101        <computeroutput>--xtree-memory=full</computeroutput> memory profiling
1102        (see <xref linkend="manual-core.xtree"/> ), then you cannot
1103        specify <varname>--keep-stacktraces=free</varname>
1104        or <varname>--keep-stacktraces=none</varname>.</para>
1105
1106    </listitem>
1107  </varlistentry>
1108
1109  <varlistentry id="opt.freelist-vol" xreflabel="--freelist-vol">
1110    <term>
1111      <option><![CDATA[--freelist-vol=<number> [default: 20000000] ]]></option>
1112    </term>
1113    <listitem>
1114      <para>When the client program releases memory using
1115      <function>free</function> (in <literal>C</literal>) or
1116      <computeroutput>delete</computeroutput>
1117      (<literal>C++</literal>), that memory is not immediately made
1118      available for re-allocation.  Instead, it is marked inaccessible
1119      and placed in a queue of freed blocks.  The purpose is to defer as
1120      long as possible the point at which freed-up memory comes back
1121      into circulation.  This increases the chance that
1122      Memcheck will be able to detect invalid
1123      accesses to blocks for some significant period of time after they
1124      have been freed.</para>
1125
1126      <para>This option specifies the maximum total size, in bytes, of the
1127      blocks in the queue.  The default value is twenty million bytes.
1128      Increasing this increases the total amount of memory used by
1129      Memcheck but may detect invalid uses of freed
1130      blocks which would otherwise go undetected.</para>
1131    </listitem>
1132  </varlistentry>
1133
1134  <varlistentry id="opt.freelist-big-blocks" xreflabel="--freelist-big-blocks">
1135    <term>
1136      <option><![CDATA[--freelist-big-blocks=<number> [default: 1000000] ]]></option>
1137    </term>
1138    <listitem>
1139      <para>When making blocks from the queue of freed blocks available
1140      for re-allocation, Memcheck will in priority re-circulate the blocks
1141      with a size greater or equal to <option>--freelist-big-blocks</option>.
1142      This ensures that freeing big blocks (in particular freeing blocks bigger than
1143      <option>--freelist-vol</option>) does not immediately lead to a re-circulation
1144      of all (or a lot of) the small blocks in the free list. In other words,
1145      this option increases the likelihood to discover dangling pointers
1146      for the "small" blocks, even when big blocks are freed.</para>
1147      <para>Setting a value of 0 means that all the blocks are re-circulated
1148      in a FIFO order. </para>
1149    </listitem>
1150  </varlistentry>
1151
1152  <varlistentry id="opt.workaround-gcc296-bugs" xreflabel="--workaround-gcc296-bugs">
1153    <term>
1154      <option><![CDATA[--workaround-gcc296-bugs=<yes|no> [default: no] ]]></option>
1155    </term>
1156    <listitem>
1157      <para>When enabled, assume that reads and writes some small
1158      distance below the stack pointer are due to bugs in GCC 2.96, and
1159      does not report them.  The "small distance" is 256 bytes by
1160      default.  Note that GCC 2.96 is the default compiler on some ancient
1161      Linux distributions (RedHat 7.X) and so you may need to use this
1162      option.  Do not use it if you do not have to, as it can cause real
1163      errors to be overlooked.  A better alternative is to use a more
1164      recent GCC in which this bug is fixed.</para>
1165
1166      <para>You may also need to use this option when working with
1167      GCC 3.X or 4.X on 32-bit PowerPC Linux.  This is because
1168      GCC generates code which occasionally accesses below the
1169      stack pointer, particularly for floating-point to/from integer
1170      conversions.  This is in violation of the 32-bit PowerPC ELF
1171      specification, which makes no provision for locations below the
1172      stack pointer to be accessible.</para>
1173
1174      <para>This option is deprecated as of version 3.12 and may be
1175      removed from future versions.  You should instead use
1176      <option>--ignore-range-below-sp</option> to specify the exact
1177      range of offsets below the stack pointer that should be ignored.
1178      A suitable equivalent
1179      is <option>--ignore-range-below-sp=1024-1</option>.
1180      </para>
1181    </listitem>
1182  </varlistentry>
1183
1184  <varlistentry id="opt.ignore-range-below-sp"
1185                xreflabel="--ignore-range-below-sp">
1186    <term>
1187      <option><![CDATA[--ignore-range-below-sp=<number>-<number> ]]></option>
1188    </term>
1189    <listitem>
1190      <para>This is a more general replacement for the deprecated
1191      <option>--workaround-gcc296-bugs</option> option.  When
1192       specified, it causes Memcheck not to report errors for accesses
1193       at the specified offsets below the stack pointer.  The two
1194       offsets must be positive decimal numbers and -- somewhat
1195       counterintuitively -- the first one must be larger, in order to
1196       imply a non-wraparound address range to ignore.  For example,
1197       to ignore 4 byte accesses at 8192 bytes below the stack
1198       pointer,
1199       use <option>--ignore-range-below-sp=8192-8189</option>.  Only
1200       one range may be specified.
1201      </para>
1202    </listitem>
1203  </varlistentry>
1204
1205  <varlistentry id="opt.show-mismatched-frees"
1206                xreflabel="--show-mismatched-frees">
1207    <term>
1208      <option><![CDATA[--show-mismatched-frees=<yes|no> [default: yes] ]]></option>
1209    </term>
1210    <listitem>
1211      <para>When enabled, Memcheck checks that heap blocks are
1212      deallocated using a function that matches the allocating
1213      function.  That is, it expects <varname>free</varname> to be
1214      used to deallocate blocks allocated
1215      by <varname>malloc</varname>, <varname>delete</varname> for
1216      blocks allocated by <varname>new</varname>,
1217      and <varname>delete[]</varname> for blocks allocated
1218      by <varname>new[]</varname>.  If a mismatch is detected, an
1219      error is reported.  This is in general important because in some
1220      environments, freeing with a non-matching function can cause
1221      crashes.</para>
1222
1223      <para>There is however a scenario where such mismatches cannot
1224      be avoided.  That is when the user provides implementations of
1225      <varname>new</varname>/<varname>new[]</varname> that
1226      call <varname>malloc</varname> and
1227      of <varname>delete</varname>/<varname>delete[]</varname> that
1228      call <varname>free</varname>, and these functions are
1229      asymmetrically inlined.  For example, imagine
1230      that <varname>delete[]</varname> is inlined
1231      but <varname>new[]</varname> is not.  The result is that
1232      Memcheck "sees" all <varname>delete[]</varname> calls as direct
1233      calls to <varname>free</varname>, even when the program source
1234      contains no mismatched calls.</para>
1235
1236      <para>This causes a lot of confusing and irrelevant error
1237      reports.  <varname>--show-mismatched-frees=no</varname> disables
1238      these checks.  It is not generally advisable to disable them,
1239      though, because you may miss real errors as a result.</para>
1240    </listitem>
1241  </varlistentry>
1242
1243  <varlistentry id="opt.ignore-ranges" xreflabel="--ignore-ranges">
1244    <term>
1245      <option><![CDATA[--ignore-ranges=0xPP-0xQQ[,0xRR-0xSS] ]]></option>
1246    </term>
1247    <listitem>
1248    <para>Any ranges listed in this option (and multiple ranges can be
1249    specified, separated by commas) will be ignored by Memcheck's
1250    addressability checking.</para>
1251    </listitem>
1252  </varlistentry>
1253
1254  <varlistentry id="opt.malloc-fill" xreflabel="--malloc-fill">
1255    <term>
1256      <option><![CDATA[--malloc-fill=<hexnumber> ]]></option>
1257    </term>
1258    <listitem>
1259      <para>Fills blocks allocated
1260      by <computeroutput>malloc</computeroutput>,
1261         <computeroutput>new</computeroutput>, etc, but not
1262      by <computeroutput>calloc</computeroutput>, with the specified
1263      byte.  This can be useful when trying to shake out obscure
1264      memory corruption problems.  The allocated area is still
1265      regarded by Memcheck as undefined -- this option only affects its
1266      contents. Note that <option>--malloc-fill</option> does not
1267      affect a block of memory when it is used as argument
1268      to client requests VALGRIND_MEMPOOL_ALLOC or
1269      VALGRIND_MALLOCLIKE_BLOCK.
1270      </para>
1271    </listitem>
1272  </varlistentry>
1273
1274  <varlistentry id="opt.free-fill" xreflabel="--free-fill">
1275    <term>
1276      <option><![CDATA[--free-fill=<hexnumber> ]]></option>
1277    </term>
1278    <listitem>
1279      <para>Fills blocks freed
1280      by <computeroutput>free</computeroutput>,
1281         <computeroutput>delete</computeroutput>, etc, with the
1282      specified byte value.  This can be useful when trying to shake out
1283      obscure memory corruption problems.  The freed area is still
1284      regarded by Memcheck as not valid for access -- this option only
1285      affects its contents. Note that <option>--free-fill</option> does not
1286      affect a block of memory when it is used as argument to
1287      client requests VALGRIND_MEMPOOL_FREE or VALGRIND_FREELIKE_BLOCK.
1288      </para>
1289    </listitem>
1290  </varlistentry>
1291
1292</variablelist>
1293<!-- end of xi:include in the manpage -->
1294
1295</sect1>
1296
1297
1298<sect1 id="mc-manual.suppfiles" xreflabel="Writing suppression files">
1299<title>Writing suppression files</title>
1300
1301<para>The basic suppression format is described in
1302<xref linkend="manual-core.suppress"/>.</para>
1303
1304<para>The suppression-type (second) line should have the form:</para>
1305<programlisting><![CDATA[
1306Memcheck:suppression_type]]></programlisting>
1307
1308<para>The Memcheck suppression types are as follows:</para>
1309
1310<itemizedlist>
1311  <listitem>
1312    <para><varname>Value1</varname>,
1313    <varname>Value2</varname>,
1314    <varname>Value4</varname>,
1315    <varname>Value8</varname>,
1316    <varname>Value16</varname>,
1317    meaning an uninitialised-value error when
1318    using a value of 1, 2, 4, 8 or 16 bytes.</para>
1319  </listitem>
1320
1321  <listitem>
1322    <para><varname>Cond</varname> (or its old
1323    name, <varname>Value0</varname>), meaning use
1324    of an uninitialised CPU condition code.</para>
1325  </listitem>
1326
1327  <listitem>
1328    <para><varname>Addr1</varname>,
1329    <varname>Addr2</varname>,
1330    <varname>Addr4</varname>,
1331    <varname>Addr8</varname>,
1332    <varname>Addr16</varname>,
1333    meaning an invalid address during a
1334    memory access of 1, 2, 4, 8 or 16 bytes respectively.</para>
1335  </listitem>
1336
1337  <listitem>
1338    <para><varname>Jump</varname>, meaning an
1339    jump to an unaddressable location error.</para>
1340  </listitem>
1341
1342  <listitem>
1343    <para><varname>Param</varname>, meaning an
1344    invalid system call parameter error.</para>
1345  </listitem>
1346
1347  <listitem>
1348    <para><varname>Free</varname>, meaning an
1349    invalid or mismatching free.</para>
1350  </listitem>
1351
1352  <listitem>
1353    <para><varname>Overlap</varname>, meaning a
1354    <computeroutput>src</computeroutput> /
1355    <computeroutput>dst</computeroutput> overlap in
1356    <function>memcpy</function> or a similar function.</para>
1357  </listitem>
1358
1359  <listitem>
1360    <para><varname>Leak</varname>, meaning
1361    a memory leak.</para>
1362  </listitem>
1363
1364</itemizedlist>
1365
1366<para><computeroutput>Param</computeroutput> errors have a mandatory extra
1367information line at this point, which is the name of the offending
1368system call parameter. </para>
1369
1370<para><computeroutput>Leak</computeroutput> errors have an optional
1371extra information line, with the following format:</para>
1372<programlisting><![CDATA[
1373match-leak-kinds:<set>]]></programlisting>
1374<para>where <computeroutput>&lt;set&gt;</computeroutput> specifies which
1375leak kinds are matched by this suppression entry.
1376<computeroutput>&lt;set&gt;</computeroutput> is specified in the
1377same way as with the option <option>--show-leak-kinds</option>, that is,
1378one of the following:</para>
1379<itemizedlist>
1380  <listitem>a comma separated list of one or more of
1381    <option>definite indirect possible reachable</option>.
1382  </listitem>
1383
1384  <listitem><option>all</option> to specify the complete set (all leak kinds).
1385  </listitem>
1386
1387  <listitem><option>none</option> for the empty set.
1388  </listitem>
1389</itemizedlist>
1390<para>If this optional extra line is not present, the suppression
1391entry will match all leak kinds.</para>
1392
1393<para>Be aware that leak suppressions that are created using
1394<option>--gen-suppressions</option> will contain this optional extra
1395line, and therefore may match fewer leaks than you expect.  You may
1396want to remove the line before using the generated
1397suppressions.</para>
1398
1399<para>The other Memcheck error kinds do not have extra lines.</para>
1400
1401<para>
1402If you give the <option>-v</option> option, Valgrind will print
1403the list of used suppressions at the end of execution.
1404For a leak suppression, this output gives the number of different
1405loss records that match the suppression, and the number of bytes
1406and blocks suppressed by the suppression.
1407If the run contains multiple leak checks, the number of bytes and blocks
1408are reset to zero before each new leak check. Note that the number of different
1409loss records is not reset to zero.</para>
1410<para>In the example below, in the last leak search, 7 blocks and 96 bytes have
1411been suppressed by a suppression with the name
1412<option>some_leak_suppression</option>:</para>
1413<programlisting><![CDATA[
1414--21041-- used_suppression:     10 some_other_leak_suppression s.supp:14 suppressed: 12,400 bytes in 1 blocks
1415--21041-- used_suppression:     39 some_leak_suppression s.supp:2 suppressed: 96 bytes in 7 blocks
1416]]></programlisting>
1417
1418<para>For <varname>ValueN</varname> and <varname>AddrN</varname>
1419errors, the first line of the calling context is either the name of
1420the function in which the error occurred, or, failing that, the full
1421path of the <filename>.so</filename> file or executable containing the
1422error location.  For <varname>Free</varname> errors, the first line is
1423the name of the function doing the freeing (eg,
1424<function>free</function>, <function>__builtin_vec_delete</function>,
1425etc).  For <varname>Overlap</varname> errors, the first line is the name of the
1426function with the overlapping arguments (eg.
1427<function>memcpy</function>, <function>strcpy</function>, etc).</para>
1428
1429<para>The last part of any suppression specifies the rest of the
1430calling context that needs to be matched.</para>
1431
1432</sect1>
1433
1434
1435
1436<sect1 id="mc-manual.machine"
1437       xreflabel="Details of Memcheck's checking machinery">
1438<title>Details of Memcheck's checking machinery</title>
1439
1440<para>Read this section if you want to know, in detail, exactly
1441what and how Memcheck is checking.</para>
1442
1443
1444<sect2 id="mc-manual.value" xreflabel="Valid-value (V) bit">
1445<title>Valid-value (V) bits</title>
1446
1447<para>It is simplest to think of Memcheck implementing a synthetic CPU
1448which is identical to a real CPU, except for one crucial detail.  Every
1449bit (literally) of data processed, stored and handled by the real CPU
1450has, in the synthetic CPU, an associated "valid-value" bit, which says
1451whether or not the accompanying bit has a legitimate value.  In the
1452discussions which follow, this bit is referred to as the V (valid-value)
1453bit.</para>
1454
1455<para>Each byte in the system therefore has a 8 V bits which follow it
1456wherever it goes.  For example, when the CPU loads a word-size item (4
1457bytes) from memory, it also loads the corresponding 32 V bits from a
1458bitmap which stores the V bits for the process' entire address space.
1459If the CPU should later write the whole or some part of that value to
1460memory at a different address, the relevant V bits will be stored back
1461in the V-bit bitmap.</para>
1462
1463<para>In short, each bit in the system has (conceptually) an associated V
1464bit, which follows it around everywhere, even inside the CPU.  Yes, all the
1465CPU's registers (integer, floating point, vector and condition registers)
1466have their own V bit vectors.  For this to work, Memcheck uses a great deal
1467of compression to represent the V bits compactly.</para>
1468
1469<para>Copying values around does not cause Memcheck to check for, or
1470report on, errors.  However, when a value is used in a way which might
1471conceivably affect your program's externally-visible behaviour,
1472the associated V bits are immediately checked.  If any of these indicate
1473that the value is undefined (even partially), an error is reported.</para>
1474
1475<para>Here's an (admittedly nonsensical) example:</para>
1476<programlisting><![CDATA[
1477int i, j;
1478int a[10], b[10];
1479for ( i = 0; i < 10; i++ ) {
1480  j = a[i];
1481  b[i] = j;
1482}]]></programlisting>
1483
1484<para>Memcheck emits no complaints about this, since it merely copies
1485uninitialised values from <varname>a[]</varname> into
1486<varname>b[]</varname>, and doesn't use them in a way which could
1487affect the behaviour of the program.  However, if
1488the loop is changed to:</para>
1489<programlisting><![CDATA[
1490for ( i = 0; i < 10; i++ ) {
1491  j += a[i];
1492}
1493if ( j == 77 )
1494  printf("hello there\n");
1495]]></programlisting>
1496
1497<para>then Memcheck will complain, at the
1498<computeroutput>if</computeroutput>, that the condition depends on
1499uninitialised values.  Note that it <command>doesn't</command> complain
1500at the <varname>j += a[i];</varname>, since at that point the
1501undefinedness is not "observable".  It's only when a decision has to be
1502made as to whether or not to do the <function>printf</function> -- an
1503observable action of your program -- that Memcheck complains.</para>
1504
1505<para>Most low level operations, such as adds, cause Memcheck to use the
1506V bits for the operands to calculate the V bits for the result.  Even if
1507the result is partially or wholly undefined, it does not
1508complain.</para>
1509
1510<para>Checks on definedness only occur in three places: when a value is
1511used to generate a memory address, when control flow decision needs to
1512be made, and when a system call is detected, Memcheck checks definedness
1513of parameters as required.</para>
1514
1515<para>If a check should detect undefinedness, an error message is
1516issued.  The resulting value is subsequently regarded as well-defined.
1517To do otherwise would give long chains of error messages.  In other
1518words, once Memcheck reports an undefined value error, it tries to
1519avoid reporting further errors derived from that same undefined
1520value.</para>
1521
1522<para>This sounds overcomplicated.  Why not just check all reads from
1523memory, and complain if an undefined value is loaded into a CPU
1524register?  Well, that doesn't work well, because perfectly legitimate C
1525programs routinely copy uninitialised values around in memory, and we
1526don't want endless complaints about that.  Here's the canonical example.
1527Consider a struct like this:</para>
1528<programlisting><![CDATA[
1529struct S { int x; char c; };
1530struct S s1, s2;
1531s1.x = 42;
1532s1.c = 'z';
1533s2 = s1;
1534]]></programlisting>
1535
1536<para>The question to ask is: how large is <varname>struct S</varname>,
1537in bytes?  An <varname>int</varname> is 4 bytes and a
1538<varname>char</varname> one byte, so perhaps a <varname>struct
1539S</varname> occupies 5 bytes?  Wrong.  All non-toy compilers we know
1540of will round the size of <varname>struct S</varname> up to a whole
1541number of words, in this case 8 bytes.  Not doing this forces compilers
1542to generate truly appalling code for accessing arrays of
1543<varname>struct S</varname>'s on some architectures.</para>
1544
1545<para>So <varname>s1</varname> occupies 8 bytes, yet only 5 of them will
1546be initialised.  For the assignment <varname>s2 = s1</varname>, GCC
1547generates code to copy all 8 bytes wholesale into <varname>s2</varname>
1548without regard for their meaning.  If Memcheck simply checked values as
1549they came out of memory, it would yelp every time a structure assignment
1550like this happened.  So the more complicated behaviour described above
1551is necessary.  This allows GCC to copy
1552<varname>s1</varname> into <varname>s2</varname> any way it likes, and a
1553warning will only be emitted if the uninitialised values are later
1554used.</para>
1555
1556</sect2>
1557
1558
1559<sect2 id="mc-manual.vaddress" xreflabel=" Valid-address (A) bits">
1560<title>Valid-address (A) bits</title>
1561
1562<para>Notice that the previous subsection describes how the validity of
1563values is established and maintained without having to say whether the
1564program does or does not have the right to access any particular memory
1565location.  We now consider the latter question.</para>
1566
1567<para>As described above, every bit in memory or in the CPU has an
1568associated valid-value (V) bit.  In addition, all bytes in memory, but
1569not in the CPU, have an associated valid-address (A) bit.  This
1570indicates whether or not the program can legitimately read or write that
1571location.  It does not give any indication of the validity of the data
1572at that location -- that's the job of the V bits -- only whether or not
1573the location may be accessed.</para>
1574
1575<para>Every time your program reads or writes memory, Memcheck checks
1576the A bits associated with the address.  If any of them indicate an
1577invalid address, an error is emitted.  Note that the reads and writes
1578themselves do not change the A bits, only consult them.</para>
1579
1580<para>So how do the A bits get set/cleared?  Like this:</para>
1581
1582<itemizedlist>
1583  <listitem>
1584    <para>When the program starts, all the global data areas are
1585    marked as accessible.</para>
1586  </listitem>
1587
1588  <listitem>
1589    <para>When the program does
1590    <function>malloc</function>/<computeroutput>new</computeroutput>,
1591    the A bits for exactly the area allocated, and not a byte more,
1592    are marked as accessible.  Upon freeing the area the A bits are
1593    changed to indicate inaccessibility.</para>
1594  </listitem>
1595
1596  <listitem>
1597    <para>When the stack pointer register (<literal>SP</literal>) moves
1598    up or down, A bits are set.  The rule is that the area from
1599    <literal>SP</literal> up to the base of the stack is marked as
1600    accessible, and below <literal>SP</literal> is inaccessible.  (If
1601    that sounds illogical, bear in mind that the stack grows down, not
1602    up, on almost all Unix systems, including GNU/Linux.)  Tracking
1603    <literal>SP</literal> like this has the useful side-effect that the
1604    section of stack used by a function for local variables etc is
1605    automatically marked accessible on function entry and inaccessible
1606    on exit.</para>
1607  </listitem>
1608
1609  <listitem>
1610    <para>When doing system calls, A bits are changed appropriately.
1611    For example, <literal>mmap</literal>
1612    magically makes files appear in the process'
1613    address space, so the A bits must be updated if <literal>mmap</literal>
1614    succeeds.</para>
1615  </listitem>
1616
1617  <listitem>
1618    <para>Optionally, your program can tell Memcheck about such changes
1619    explicitly, using the client request mechanism described
1620    above.</para>
1621  </listitem>
1622
1623</itemizedlist>
1624
1625</sect2>
1626
1627
1628<sect2 id="mc-manual.together" xreflabel="Putting it all together">
1629<title>Putting it all together</title>
1630
1631<para>Memcheck's checking machinery can be summarised as
1632follows:</para>
1633
1634<itemizedlist>
1635  <listitem>
1636    <para>Each byte in memory has 8 associated V (valid-value) bits,
1637    saying whether or not the byte has a defined value, and a single A
1638    (valid-address) bit, saying whether or not the program currently has
1639    the right to read/write that address.  As mentioned above, heavy
1640    use of compression means the overhead is typically around 25%.</para>
1641  </listitem>
1642
1643  <listitem>
1644    <para>When memory is read or written, the relevant A bits are
1645    consulted.  If they indicate an invalid address, Memcheck emits an
1646    Invalid read or Invalid write error.</para>
1647  </listitem>
1648
1649  <listitem>
1650    <para>When memory is read into the CPU's registers, the relevant V
1651    bits are fetched from memory and stored in the simulated CPU.  They
1652    are not consulted.</para>
1653  </listitem>
1654
1655  <listitem>
1656    <para>When a register is written out to memory, the V bits for that
1657    register are written back to memory too.</para>
1658  </listitem>
1659
1660  <listitem>
1661    <para>When values in CPU registers are used to generate a memory
1662    address, or to determine the outcome of a conditional branch, the V
1663    bits for those values are checked, and an error emitted if any of
1664    them are undefined.</para>
1665  </listitem>
1666
1667  <listitem>
1668    <para>When values in CPU registers are used for any other purpose,
1669    Memcheck computes the V bits for the result, but does not check
1670    them.</para>
1671  </listitem>
1672
1673  <listitem>
1674    <para>Once the V bits for a value in the CPU have been checked, they
1675    are then set to indicate validity.  This avoids long chains of
1676    errors.</para>
1677  </listitem>
1678
1679  <listitem>
1680    <para>When values are loaded from memory, Memcheck checks the A bits
1681    for that location and issues an illegal-address warning if needed.
1682    In that case, the V bits loaded are forced to indicate Valid,
1683    despite the location being invalid.</para>
1684
1685    <para>This apparently strange choice reduces the amount of confusing
1686    information presented to the user.  It avoids the unpleasant
1687    phenomenon in which memory is read from a place which is both
1688    unaddressable and contains invalid values, and, as a result, you get
1689    not only an invalid-address (read/write) error, but also a
1690    potentially large set of uninitialised-value errors, one for every
1691    time the value is used.</para>
1692
1693    <para>There is a hazy boundary case to do with multi-byte loads from
1694    addresses which are partially valid and partially invalid.  See
1695    details of the option <option>--partial-loads-ok</option> for details.
1696    </para>
1697  </listitem>
1698
1699</itemizedlist>
1700
1701
1702<para>Memcheck intercepts calls to <function>malloc</function>,
1703<function>calloc</function>, <function>realloc</function>,
1704<function>valloc</function>, <function>memalign</function>,
1705<function>free</function>, <computeroutput>new</computeroutput>,
1706<computeroutput>new[]</computeroutput>,
1707<computeroutput>delete</computeroutput> and
1708<computeroutput>delete[]</computeroutput>.  The behaviour you get
1709is:</para>
1710
1711<itemizedlist>
1712
1713  <listitem>
1714    <para><function>malloc</function>/<function>new</function>/<computeroutput>new[]</computeroutput>:
1715    the returned memory is marked as addressable but not having valid
1716    values.  This means you have to write to it before you can read
1717    it.</para>
1718  </listitem>
1719
1720  <listitem>
1721    <para><function>calloc</function>: returned memory is marked both
1722    addressable and valid, since <function>calloc</function> clears
1723    the area to zero.</para>
1724  </listitem>
1725
1726  <listitem>
1727    <para><function>realloc</function>: if the new size is larger than
1728    the old, the new section is addressable but invalid, as with
1729    <function>malloc</function>.  If the new size is smaller, the
1730    dropped-off section is marked as unaddressable.  You may only pass to
1731    <function>realloc</function> a pointer previously issued to you by
1732    <function>malloc</function>/<function>calloc</function>/<function>realloc</function>.</para>
1733  </listitem>
1734
1735  <listitem>
1736    <para><function>free</function>/<computeroutput>delete</computeroutput>/<computeroutput>delete[]</computeroutput>:
1737    you may only pass to these functions a pointer previously issued
1738    to you by the corresponding allocation function.  Otherwise,
1739    Memcheck complains.  If the pointer is indeed valid, Memcheck
1740    marks the entire area it points at as unaddressable, and places
1741    the block in the freed-blocks-queue.  The aim is to defer as long
1742    as possible reallocation of this block.  Until that happens, all
1743    attempts to access it will elicit an invalid-address error, as you
1744    would hope.</para>
1745  </listitem>
1746
1747</itemizedlist>
1748
1749</sect2>
1750</sect1>
1751
1752<sect1 id="mc-manual.monitor-commands" xreflabel="Memcheck Monitor Commands">
1753<title>Memcheck Monitor Commands</title>
1754<para>The Memcheck tool provides monitor commands handled by Valgrind's
1755built-in gdbserver (see <xref linkend="manual-core-adv.gdbserver-commandhandling"/>).
1756</para>
1757
1758<itemizedlist>
1759  <listitem>
1760    <para><varname>xb &lt;addr&gt; [&lt;len&gt;]</varname>
1761      shows the definedness (V) bits and values for &lt;len&gt; (default 1)
1762      bytes starting at &lt;addr&gt;.
1763      For each 8 bytes, two lines are output.
1764    </para>
1765    <para>
1766      The first line shows the validity bits for 8 bytes.
1767      The definedness of each byte in the range is given using two hexadecimal
1768      digits.  These hexadecimal digits encode the validity of each bit of the
1769      corresponding byte,
1770      using 0 if the bit is defined and 1 if the bit is undefined.
1771      If a byte is not addressable, its validity bits are replaced
1772      by <varname>__</varname> (a double underscore).
1773    </para>
1774    <para>
1775      The second line shows the values of the bytes below the corresponding
1776      validity bits. The format used to show the bytes data is similar to the
1777      GDB command 'x /&lt;len&gt;xb &lt;addr&gt;'. The value for a non
1778      addressable bytes is shown as ?? (two question marks).
1779    </para>
1780    <para>
1781      In the following example, <varname>string10</varname> is an array
1782      of 10 characters, in which the even numbered bytes are
1783      undefined. In the below example, the byte corresponding
1784      to <varname>string10[5]</varname> is not addressable.
1785    </para>
1786<programlisting><![CDATA[
1787(gdb) p &string10
1788$4 = (char (*)[10]) 0x804a2f0
1789(gdb) mo xb 0x804a2f0 10
1790                  ff      00      ff      00      ff      __      ff      00
17910x804A2F0:      0x3f    0x6e    0x3f    0x65    0x3f    0x??     0x3f    0x65
1792                  ff      00
17930x804A2F8:      0x3f    0x00
1794Address 0x804A2F0 len 10 has 1 bytes unaddressable
1795(gdb)
1796]]></programlisting>
1797
1798    <para> The command xb cannot be used with registers. To get
1799      the validity bits of a register, you must start Valgrind with the
1800      option <option>--vgdb-shadow-registers=yes</option>. The validity
1801      bits of a register can then be obtained by printing the 'shadow 1'
1802      corresponding register.  In the below x86 example, the register
1803      eax has all its bits undefined, while the register ebx is fully
1804      defined.
1805    </para>
1806<programlisting><![CDATA[
1807(gdb) p /x $eaxs1
1808$9 = 0xffffffff
1809(gdb) p /x $ebxs1
1810$10 = 0x0
1811(gdb)
1812]]></programlisting>
1813
1814  </listitem>
1815
1816  <listitem>
1817    <para><varname>get_vbits &lt;addr&gt; [&lt;len&gt;]</varname>
1818    shows the definedness (V) bits for &lt;len&gt; (default 1) bytes
1819    starting at &lt;addr&gt; using the same convention as the
1820    <varname>xb</varname> command. <varname>get_vbits</varname> only
1821    shows the V bits (grouped by 4 bytes). It does not show the values.
1822    If you want to associate V bits with the corresponding byte values, the
1823    <varname>xb</varname> command will be easier to use, in particular
1824    on little endian computers when associating undefined parts of an integer
1825    with their V bits values.
1826    </para>
1827    <para>
1828    The following example shows the result of <varname>get_vibts</varname>
1829    on the <varname>string10</varname> used in the  <varname>xb</varname>
1830    command explanation.
1831    </para>
1832<programlisting><![CDATA[
1833(gdb) monitor get_vbits 0x804a2f0 10
1834ff00ff00 ff__ff00 ff00
1835Address 0x804A2F0 len 10 has 1 bytes unaddressable
1836(gdb)
1837]]></programlisting>
1838
1839  </listitem>
1840
1841  <listitem>
1842    <para><varname>make_memory
1843    [noaccess|undefined|defined|Definedifaddressable] &lt;addr&gt;
1844    [&lt;len&gt;]</varname> marks the range of &lt;len&gt; (default 1)
1845    bytes at &lt;addr&gt; as having the given status. Parameter
1846    <varname>noaccess</varname> marks the range as non-accessible, so
1847    Memcheck will report an error on any access to it.
1848    <varname>undefined</varname> or <varname>defined</varname> mark
1849    the area as accessible, but Memcheck regards the bytes in it
1850    respectively as having undefined or defined values.
1851    <varname>Definedifaddressable</varname> marks as defined, bytes in
1852    the range which are already addressible, but makes no change to
1853    the status of bytes in the range which are not addressible. Note
1854    that the first letter of <varname>Definedifaddressable</varname>
1855    is an uppercase D to avoid confusion with <varname>defined</varname>.
1856    </para>
1857
1858    <para>
1859    In the following example, the first byte of the
1860    <varname>string10</varname> is marked as defined:
1861    </para>
1862<programlisting><![CDATA[
1863(gdb) monitor make_memory defined 0x8049e28  1
1864(gdb) monitor get_vbits 0x8049e28 10
18650000ff00 ff00ff00 ff00
1866(gdb)
1867]]></programlisting>
1868  </listitem>
1869
1870  <listitem>
1871    <para><varname>check_memory [addressable|defined] &lt;addr&gt;
1872    [&lt;len&gt;]</varname> checks that the range of &lt;len&gt;
1873    (default 1) bytes at &lt;addr&gt; has the specified accessibility.
1874    It then outputs a description of &lt;addr&gt;. In the following
1875    example, a detailed description is available because the
1876    option <option>--read-var-info=yes</option> was given at Valgrind
1877    startup:
1878    </para>
1879<programlisting><![CDATA[
1880(gdb) monitor check_memory defined 0x8049e28  1
1881Address 0x8049E28 len 1 defined
1882==14698==  Location 0x8049e28 is 0 bytes inside string10[0],
1883==14698==  declared at prog.c:10, in frame #0 of thread 1
1884(gdb)
1885]]></programlisting>
1886  </listitem>
1887
1888  <listitem>
1889    <para><varname>leak_check [full*|summary|xtleak]
1890                              [kinds &lt;set&gt;|reachable|possibleleak*|definiteleak]
1891                              [heuristics heur1,heur2,...]
1892                              [increased*|changed|any]
1893                              [unlimited*|limited &lt;max_loss_records_output&gt;]
1894          </varname>
1895    performs a leak check. The <varname>*</varname> in the arguments
1896    indicates the default values. </para>
1897
1898    <para> If the <varname>[full*|summary|xtleak]</varname> argument is
1899    <varname>summary</varname>, only a summary of the leak search is given;
1900    otherwise a full leak report is produced.  A full leak report gives
1901    detailed information for each leak: the stack trace where the leaked blocks
1902    were allocated, the number of blocks leaked and their total size.  When a
1903    full report is requested, the next two arguments further specify what
1904    kind of leaks to report.  A leak's details are shown if they match
1905    both the second and third argument. A full leak report might
1906    output detailed information for many leaks. The nr of leaks for
1907    which information is output can be controlled using
1908    the <varname>limited</varname> argument followed by the maximum nr
1909    of leak records to output. If this maximum is reached, the leak
1910    search  outputs the records with the biggest number of bytes.
1911    </para>
1912    <para>The value <varname>xtleak</varname> also produces a full leak report,
1913      but output it as an xtree in a file xtleak.kcg.%p.%n (see <xref linkend="opt.log-file"/>).
1914      See <xref linkend="manual-core.xtree"/>
1915      for a detailed explanation about execution trees formats.
1916      See <xref linkend="opt.xtree-leak"/> for the description of the events
1917      in a xtree leak file.
1918      </para>
1919
1920    <para>The <varname>kinds</varname> argument controls what kind of blocks
1921    are shown for a <varname>full</varname> leak search.  The set of leak kinds
1922    to show can be specified using a <varname>&lt;set&gt;</varname> similarly
1923    to the command line option <option>--show-leak-kinds</option>.
1924    Alternatively, the  value <varname>definiteleak</varname>
1925    is equivalent to <varname>kinds definite</varname>, the
1926    value <varname>possibleleak</varname> is equivalent to
1927    <varname>kinds definite,possible</varname> : it will also show
1928    possibly leaked blocks, .i.e those for which only an interior
1929    pointer was found.  The value <varname>reachable</varname> will
1930    show all block categories (i.e. is equivalent to <varname>kinds
1931    all</varname>).
1932    </para>
1933
1934    <para>The <varname>heuristics</varname> argument controls the heuristics
1935    used during the leak search. The set of heuristics to use can be specified
1936    using a <varname>&lt;set&gt;</varname> similarly
1937    to the command line option <option>--leak-check-heuristics</option>.
1938    The default value for the <varname>heuristics</varname> argument is
1939    <varname>heuristics none</varname>.
1940    </para>
1941
1942    <para>The <varname>[increased*|changed|any]</varname> argument controls what
1943    kinds of changes are shown for a <varname>full</varname> leak search. The
1944    value <varname>increased</varname> specifies that only block
1945    allocation stacks with an increased number of leaked bytes or
1946    blocks since the previous leak check should be shown.  The
1947    value <varname>changed</varname> specifies that allocation stacks
1948    with any change since the previous leak check should be shown.
1949    The value <varname>any</varname> specifies that all leak entries
1950    should be shown, regardless of any increase or decrease.  When
1951    If <varname>increased</varname> or <varname>changed</varname> are
1952    specified, the leak report entries will show the delta relative to
1953    the previous leak report.
1954    </para>
1955
1956    <para>The following example shows usage of the
1957    <varname>leak_check</varname> monitor command on
1958    the <varname>memcheck/tests/leak-cases.c</varname> regression
1959    test. The first command outputs one entry having an increase in
1960    the leaked bytes.  The second command is the same as the first
1961    command, but uses the abbreviated forms accepted by GDB and the
1962    Valgrind gdbserver. It only outputs the summary information, as
1963    there was no increase since the previous leak search.</para>
1964<programlisting><![CDATA[
1965(gdb) monitor leak_check full possibleleak increased
1966==19520== 16 (+16) bytes in 1 (+1) blocks are possibly lost in loss record 9 of 12
1967==19520==    at 0x40070B4: malloc (vg_replace_malloc.c:263)
1968==19520==    by 0x80484D5: mk (leak-cases.c:52)
1969==19520==    by 0x804855F: f (leak-cases.c:81)
1970==19520==    by 0x80488E0: main (leak-cases.c:107)
1971==19520==
1972==19520== LEAK SUMMARY:
1973==19520==    definitely lost: 32 (+0) bytes in 2 (+0) blocks
1974==19520==    indirectly lost: 16 (+0) bytes in 1 (+0) blocks
1975==19520==      possibly lost: 32 (+16) bytes in 2 (+1) blocks
1976==19520==    still reachable: 96 (+16) bytes in 6 (+1) blocks
1977==19520==         suppressed: 0 (+0) bytes in 0 (+0) blocks
1978==19520== Reachable blocks (those to which a pointer was found) are not shown.
1979==19520== To see them, add 'reachable any' args to leak_check
1980==19520==
1981(gdb) mo l
1982==19520== LEAK SUMMARY:
1983==19520==    definitely lost: 32 (+0) bytes in 2 (+0) blocks
1984==19520==    indirectly lost: 16 (+0) bytes in 1 (+0) blocks
1985==19520==      possibly lost: 32 (+0) bytes in 2 (+0) blocks
1986==19520==    still reachable: 96 (+0) bytes in 6 (+0) blocks
1987==19520==         suppressed: 0 (+0) bytes in 0 (+0) blocks
1988==19520== Reachable blocks (those to which a pointer was found) are not shown.
1989==19520== To see them, add 'reachable any' args to leak_check
1990==19520==
1991(gdb)
1992]]></programlisting>
1993    <para>Note that when using Valgrind's gdbserver, it is not
1994    necessary to rerun
1995    with <option>--leak-check=full</option>
1996    <option>--show-reachable=yes</option> to see the reachable
1997    blocks. You can obtain the same information without rerunning by
1998    using the GDB command <computeroutput>monitor leak_check full
1999    reachable any</computeroutput> (or, using
2000    abbreviation: <computeroutput>mo l f r a</computeroutput>).
2001    </para>
2002  </listitem>
2003
2004  <listitem>
2005    <para><varname>block_list &lt;loss_record_nr&gt;|&lt;loss_record_nr_from&gt;..&lt;loss_record_nr_to&gt;
2006        [unlimited*|limited &lt;max_blocks&gt;]
2007        [heuristics heur1,heur2,...]
2008      </varname>
2009      shows the list of blocks belonging to
2010      <varname>&lt;loss_record_nr&gt;</varname> (or to the loss records range
2011      <varname>&lt;loss_record_nr_from&gt;..&lt;loss_record_nr_to&gt;</varname>).
2012      The nr of blocks to print can be controlled using the
2013      <varname>limited</varname> argument followed by the maximum nr
2014      of blocks to output.
2015      If one or more heuristics are given, only prints the loss records
2016      and blocks found via one of the given <varname>heur1,heur2,...</varname>
2017      heuristics.
2018    </para>
2019
2020    <para> A leak search merges the allocated blocks in loss records :
2021    a loss record re-groups all blocks having the same state (for
2022    example, Definitely Lost) and the same allocation backtrace.
2023    Each loss record is identified in the leak search result
2024    by a loss record number.
2025    The <varname>block_list</varname> command shows the loss record information
2026    followed by the addresses and sizes of the blocks which have been
2027    merged in the loss record. If a block was found using an heuristic, the block size
2028    is followed by the heuristic.
2029    </para>
2030
2031    <para> If a directly lost block causes some other blocks to be indirectly
2032    lost, the block_list command will also show these indirectly lost blocks.
2033    The indirectly lost blocks will be indented according to the level of indirection
2034    between the directly lost block and the indirectly lost block(s).
2035    Each indirectly lost block is followed by the reference of its loss record.
2036    </para>
2037
2038    <para> The block_list command can be used on the results of a leak search as long
2039    as no block has been freed after this leak search: as soon as the program frees
2040    a block, a new leak search is needed before block_list can be used again.
2041    </para>
2042
2043    <para>
2044    In the below example, the program leaks a tree structure by losing the pointer to
2045    the block A (top of the tree).
2046    So, the block A is directly lost, causing an indirect
2047    loss of blocks B to G. The first block_list command shows the loss record of A
2048    (a definitely lost block with address 0x4028028, size 16). The addresses and sizes
2049    of the indirectly lost blocks due to block A are shown below the block A.
2050    The second command shows the details of one of the indirect loss records output
2051    by the first command.
2052    </para>
2053<programlisting><![CDATA[
2054           A
2055         /   \
2056        B     C
2057       / \   / \
2058      D   E F   G
2059]]></programlisting>
2060
2061<programlisting><![CDATA[
2062(gdb) bt
2063#0  main () at leak-tree.c:69
2064(gdb) monitor leak_check full any
2065==19552== 112 (16 direct, 96 indirect) bytes in 1 blocks are definitely lost in loss record 7 of 7
2066==19552==    at 0x40070B4: malloc (vg_replace_malloc.c:263)
2067==19552==    by 0x80484D5: mk (leak-tree.c:28)
2068==19552==    by 0x80484FC: f (leak-tree.c:41)
2069==19552==    by 0x8048856: main (leak-tree.c:63)
2070==19552==
2071==19552== LEAK SUMMARY:
2072==19552==    definitely lost: 16 bytes in 1 blocks
2073==19552==    indirectly lost: 96 bytes in 6 blocks
2074==19552==      possibly lost: 0 bytes in 0 blocks
2075==19552==    still reachable: 0 bytes in 0 blocks
2076==19552==         suppressed: 0 bytes in 0 blocks
2077==19552==
2078(gdb) monitor block_list 7
2079==19552== 112 (16 direct, 96 indirect) bytes in 1 blocks are definitely lost in loss record 7 of 7
2080==19552==    at 0x40070B4: malloc (vg_replace_malloc.c:263)
2081==19552==    by 0x80484D5: mk (leak-tree.c:28)
2082==19552==    by 0x80484FC: f (leak-tree.c:41)
2083==19552==    by 0x8048856: main (leak-tree.c:63)
2084==19552== 0x4028028[16]
2085==19552==   0x4028068[16] indirect loss record 1
2086==19552==      0x40280E8[16] indirect loss record 3
2087==19552==      0x4028128[16] indirect loss record 4
2088==19552==   0x40280A8[16] indirect loss record 2
2089==19552==      0x4028168[16] indirect loss record 5
2090==19552==      0x40281A8[16] indirect loss record 6
2091(gdb) mo b 2
2092==19552== 16 bytes in 1 blocks are indirectly lost in loss record 2 of 7
2093==19552==    at 0x40070B4: malloc (vg_replace_malloc.c:263)
2094==19552==    by 0x80484D5: mk (leak-tree.c:28)
2095==19552==    by 0x8048519: f (leak-tree.c:43)
2096==19552==    by 0x8048856: main (leak-tree.c:63)
2097==19552== 0x40280A8[16]
2098==19552==   0x4028168[16] indirect loss record 5
2099==19552==   0x40281A8[16] indirect loss record 6
2100(gdb)
2101
2102]]></programlisting>
2103
2104  </listitem>
2105
2106  <listitem>
2107    <para><varname>who_points_at &lt;addr&gt; [&lt;len&gt;]</varname>
2108    shows all the locations where a pointer to addr is found.
2109    If len is equal to 1, the command only shows the locations pointing
2110    exactly at addr (i.e. the "start pointers" to addr).
2111    If len is &gt; 1, "interior pointers" pointing at the len first bytes
2112    will also be shown.
2113    </para>
2114
2115    <para>The locations searched for are the same as the locations
2116    used in the leak search. So, <varname>who_points_at</varname> can a.o.
2117    be used to show why the leak search still can reach a block, or can
2118    search for dangling pointers to a freed block.
2119    Each location pointing at addr (or pointing inside addr if interior pointers
2120    are being searched for) will be described.
2121    </para>
2122
2123    <para>In the below example, the pointers to the 'tree block A' (see example
2124    in command <varname>block_list</varname>) is shown before the tree was leaked.
2125    The descriptions are detailed as the option <option>--read-var-info=yes</option>
2126    was given at Valgrind startup. The second call shows the pointers (start and interior
2127    pointers) to block G. The block G (0x40281A8) is reachable via block C (0x40280a8)
2128    and register ECX of tid 1 (tid is the Valgrind thread id).
2129    It is "interior reachable" via the register EBX.
2130    </para>
2131
2132<programlisting><![CDATA[
2133(gdb) monitor who_points_at 0x4028028
2134==20852== Searching for pointers to 0x4028028
2135==20852== *0x8049e20 points at 0x4028028
2136==20852==  Location 0x8049e20 is 0 bytes inside global var "t"
2137==20852==  declared at leak-tree.c:35
2138(gdb) monitor who_points_at 0x40281A8 16
2139==20852== Searching for pointers pointing in 16 bytes from 0x40281a8
2140==20852== *0x40280ac points at 0x40281a8
2141==20852==  Address 0x40280ac is 4 bytes inside a block of size 16 alloc'd
2142==20852==    at 0x40070B4: malloc (vg_replace_malloc.c:263)
2143==20852==    by 0x80484D5: mk (leak-tree.c:28)
2144==20852==    by 0x8048519: f (leak-tree.c:43)
2145==20852==    by 0x8048856: main (leak-tree.c:63)
2146==20852== tid 1 register ECX points at 0x40281a8
2147==20852== tid 1 register EBX interior points at 2 bytes inside 0x40281a8
2148(gdb)
2149]]></programlisting>
2150
2151  <para> When <varname>who_points_at</varname> finds an interior pointer,
2152  it will report the heuristic(s) with which this interior pointer
2153  will be considered as reachable. Note that this is done independently
2154  of the value of the option <option>--leak-check-heuristics</option>.
2155  In the below example, the loss record 6 indicates a possibly lost
2156  block. <varname>who_points_at</varname> reports that there is an interior
2157  pointer pointing in this block, and that the block can be considered
2158  reachable using the heuristic
2159  <computeroutput>multipleinheritance</computeroutput>.
2160  </para>
2161
2162<programlisting><![CDATA[
2163(gdb) monitor block_list 6
2164==3748== 8 bytes in 1 blocks are possibly lost in loss record 6 of 7
2165==3748==    at 0x4007D77: operator new(unsigned int) (vg_replace_malloc.c:313)
2166==3748==    by 0x8048954: main (leak_cpp_interior.cpp:43)
2167==3748== 0x402A0E0[8]
2168(gdb) monitor who_points_at 0x402A0E0 8
2169==3748== Searching for pointers pointing in 8 bytes from 0x402a0e0
2170==3748== *0xbe8ee078 interior points at 4 bytes inside 0x402a0e0
2171==3748==  Address 0xbe8ee078 is on thread 1's stack
2172==3748== block at 0x402a0e0 considered reachable by ptr 0x402a0e4 using multipleinheritance heuristic
2173(gdb)
2174]]></programlisting>
2175
2176  </listitem>
2177
2178</itemizedlist>
2179
2180</sect1>
2181
2182<sect1 id="mc-manual.clientreqs" xreflabel="Client requests">
2183<title>Client Requests</title>
2184
2185<para>The following client requests are defined in
2186<filename>memcheck.h</filename>.
2187See <filename>memcheck.h</filename> for exact details of their
2188arguments.</para>
2189
2190<itemizedlist>
2191
2192  <listitem>
2193    <para><varname>VALGRIND_MAKE_MEM_NOACCESS</varname>,
2194    <varname>VALGRIND_MAKE_MEM_UNDEFINED</varname> and
2195    <varname>VALGRIND_MAKE_MEM_DEFINED</varname>.
2196    These mark address ranges as completely inaccessible,
2197    accessible but containing undefined data, and accessible and
2198    containing defined data, respectively. They return -1, when
2199    run on Valgrind and 0 otherwise.</para>
2200  </listitem>
2201
2202  <listitem>
2203    <para><varname>VALGRIND_MAKE_MEM_DEFINED_IF_ADDRESSABLE</varname>.
2204    This is just like <varname>VALGRIND_MAKE_MEM_DEFINED</varname> but only
2205    affects those bytes that are already addressable.</para>
2206  </listitem>
2207
2208  <listitem>
2209    <para><varname>VALGRIND_CHECK_MEM_IS_ADDRESSABLE</varname> and
2210    <varname>VALGRIND_CHECK_MEM_IS_DEFINED</varname>: check immediately
2211    whether or not the given address range has the relevant property,
2212    and if not, print an error message.  Also, for the convenience of
2213    the client, returns zero if the relevant property holds; otherwise,
2214    the returned value is the address of the first byte for which the
2215    property is not true.  Always returns 0 when not run on
2216    Valgrind.</para>
2217  </listitem>
2218
2219  <listitem>
2220    <para><varname>VALGRIND_CHECK_VALUE_IS_DEFINED</varname>: a quick and easy
2221    way to find out whether Valgrind thinks a particular value
2222    (lvalue, to be precise) is addressable and defined.  Prints an error
2223    message if not.  It has no return value.</para>
2224  </listitem>
2225
2226  <listitem>
2227    <para><varname>VALGRIND_DO_LEAK_CHECK</varname>: does a full memory leak
2228    check (like <option>--leak-check=full</option>) right now.
2229    This is useful for incrementally checking for leaks between arbitrary
2230    places in the program's execution.  It has no return value.</para>
2231  </listitem>
2232
2233  <listitem>
2234    <para><varname>VALGRIND_DO_ADDED_LEAK_CHECK</varname>: same as
2235   <varname> VALGRIND_DO_LEAK_CHECK</varname> but only shows the
2236    entries for which there was an increase in leaked bytes or leaked
2237    number of blocks since the previous leak search.  It has no return
2238    value.</para>
2239  </listitem>
2240
2241  <listitem>
2242    <para><varname>VALGRIND_DO_CHANGED_LEAK_CHECK</varname>: same as
2243    <varname>VALGRIND_DO_LEAK_CHECK</varname> but only shows the
2244    entries for which there was an increase or decrease in leaked
2245    bytes or leaked number of blocks since the previous leak search. It
2246    has no return value.</para>
2247  </listitem>
2248
2249  <listitem>
2250    <para><varname>VALGRIND_DO_QUICK_LEAK_CHECK</varname>: like
2251    <varname>VALGRIND_DO_LEAK_CHECK</varname>, except it produces only a leak
2252    summary (like <option>--leak-check=summary</option>).
2253    It has no return value.</para>
2254  </listitem>
2255
2256  <listitem>
2257    <para><varname>VALGRIND_COUNT_LEAKS</varname>: fills in the four
2258    arguments with the number of bytes of memory found by the previous
2259    leak check to be leaked (i.e. the sum of direct leaks and indirect leaks),
2260    dubious, reachable and suppressed.  This is useful in test harness code,
2261    after calling <varname>VALGRIND_DO_LEAK_CHECK</varname> or
2262    <varname>VALGRIND_DO_QUICK_LEAK_CHECK</varname>.</para>
2263  </listitem>
2264
2265  <listitem>
2266    <para><varname>VALGRIND_COUNT_LEAK_BLOCKS</varname>: identical to
2267    <varname>VALGRIND_COUNT_LEAKS</varname> except that it returns the
2268    number of blocks rather than the number of bytes in each
2269    category.</para>
2270  </listitem>
2271
2272  <listitem>
2273    <para><varname>VALGRIND_GET_VBITS</varname> and
2274    <varname>VALGRIND_SET_VBITS</varname>: allow you to get and set the
2275    V (validity) bits for an address range.  You should probably only
2276    set V bits that you have got with
2277    <varname>VALGRIND_GET_VBITS</varname>.  Only for those who really
2278    know what they are doing.</para>
2279  </listitem>
2280
2281  <listitem>
2282    <para><varname>VALGRIND_CREATE_BLOCK</varname> and
2283    <varname>VALGRIND_DISCARD</varname>.  <varname>VALGRIND_CREATE_BLOCK</varname>
2284    takes an address, a number of bytes and a character string.  The
2285    specified address range is then associated with that string.  When
2286    Memcheck reports an invalid access to an address in the range, it
2287    will describe it in terms of this block rather than in terms of
2288    any other block it knows about.  Note that the use of this macro
2289    does not actually change the state of memory in any way -- it
2290    merely gives a name for the range.
2291    </para>
2292
2293    <para>At some point you may want Memcheck to stop reporting errors
2294    in terms of the block named
2295    by <varname>VALGRIND_CREATE_BLOCK</varname>.  To make this
2296    possible, <varname>VALGRIND_CREATE_BLOCK</varname> returns a
2297    "block handle", which is a C <varname>int</varname> value.  You
2298    can pass this block handle to <varname>VALGRIND_DISCARD</varname>.
2299    After doing so, Valgrind will no longer relate addressing errors
2300    in the specified range to the block.  Passing invalid handles to
2301    <varname>VALGRIND_DISCARD</varname> is harmless.
2302   </para>
2303  </listitem>
2304
2305</itemizedlist>
2306
2307</sect1>
2308
2309
2310
2311
2312<sect1 id="mc-manual.mempools" xreflabel="Memory Pools">
2313<title>Memory Pools: describing and working with custom allocators</title>
2314
2315<para>Some programs use custom memory allocators, often for performance
2316reasons.  Left to itself, Memcheck is unable to understand the
2317behaviour of custom allocation schemes as well as it understands the
2318standard allocators, and so may miss errors and leaks in your program.  What
2319this section describes is a way to give Memcheck enough of a description of
2320your custom allocator that it can make at least some sense of what is
2321happening.</para>
2322
2323<para>There are many different sorts of custom allocator, so Memcheck
2324attempts to reason about them using a loose, abstract model.  We
2325use the following terminology when describing custom allocation
2326systems:</para>
2327
2328<itemizedlist>
2329  <listitem>
2330    <para>Custom allocation involves a set of independent "memory pools".
2331    </para>
2332  </listitem>
2333  <listitem>
2334    <para>Memcheck's notion of a a memory pool consists of a single "anchor
2335    address" and a set of non-overlapping "chunks" associated with the
2336    anchor address.</para>
2337  </listitem>
2338  <listitem>
2339    <para>Typically a pool's anchor address is the address of a
2340    book-keeping "header" structure.</para>
2341  </listitem>
2342  <listitem>
2343    <para>Typically the pool's chunks are drawn from a contiguous
2344    "superblock" acquired through the system
2345    <function>malloc</function> or
2346    <function>mmap</function>.</para>
2347  </listitem>
2348
2349</itemizedlist>
2350
2351<para>Keep in mind that the last two points above say "typically": the
2352Valgrind mempool client request API is intentionally vague about the
2353exact structure of a mempool. There is no specific mention made of
2354headers or superblocks. Nevertheless, the following picture may help
2355elucidate the intention of the terms in the API:</para>
2356
2357<programlisting><![CDATA[
2358   "pool"
2359   (anchor address)
2360   |
2361   v
2362   +--------+---+
2363   | header | o |
2364   +--------+-|-+
2365              |
2366              v                  superblock
2367              +------+---+--------------+---+------------------+
2368              |      |rzB|  allocation  |rzB|                  |
2369              +------+---+--------------+---+------------------+
2370                         ^              ^
2371                         |              |
2372                       "addr"     "addr"+"size"
2373]]></programlisting>
2374
2375<para>
2376Note that the header and the superblock may be contiguous or
2377discontiguous, and there may be multiple superblocks associated with a
2378single header; such variations are opaque to Memcheck. The API
2379only requires that your allocation scheme can present sensible values
2380of "pool", "addr" and "size".</para>
2381
2382<para>
2383Typically, before making client requests related to mempools, a client
2384program will have allocated such a header and superblock for their
2385mempool, and marked the superblock NOACCESS using the
2386<varname>VALGRIND_MAKE_MEM_NOACCESS</varname> client request.</para>
2387
2388<para>
2389When dealing with mempools, the goal is to maintain a particular
2390invariant condition: that Memcheck believes the unallocated portions
2391of the pool's superblock (including redzones) are NOACCESS. To
2392maintain this invariant, the client program must ensure that the
2393superblock starts out in that state; Memcheck cannot make it so, since
2394Memcheck never explicitly learns about the superblock of a pool, only
2395the allocated chunks within the pool.</para>
2396
2397<para>
2398Once the header and superblock for a pool are established and properly
2399marked, there are a number of client requests programs can use to
2400inform Memcheck about changes to the state of a mempool:</para>
2401
2402<itemizedlist>
2403
2404  <listitem>
2405    <para>
2406    <varname>VALGRIND_CREATE_MEMPOOL(pool, rzB, is_zeroed)</varname>:
2407    This request registers the address <varname>pool</varname> as the anchor
2408    address for a memory pool. It also provides a size
2409    <varname>rzB</varname>, specifying how large the redzones placed around
2410    chunks allocated from the pool should be. Finally, it provides an
2411    <varname>is_zeroed</varname> argument that specifies whether the pool's
2412    chunks are zeroed (more precisely: defined) when allocated.
2413    </para>
2414    <para>
2415    Upon completion of this request, no chunks are associated with the
2416    pool.  The request simply tells Memcheck that the pool exists, so that
2417    subsequent calls can refer to it as a pool.
2418    </para>
2419  </listitem>
2420
2421  <listitem>
2422    <!-- Note: the below is mostly a copy of valgrind.h. Keep in sync! -->
2423    <para>
2424      <varname>VALGRIND_CREATE_MEMPOOL_EXT(pool, rzB, is_zeroed, flags)</varname>:
2425      Create a memory pool with some flags (that can
2426      be OR-ed together) specifying extended behaviour.  When flags is
2427      zero, the behaviour is identical to
2428    <varname>VALGRIND_CREATE_MEMPOOL</varname>.</para>
2429    <itemizedlist>
2430      <listitem>
2431	<para> The flag <varname>VALGRIND_MEMPOOL_METAPOOL</varname>
2432          specifies that the pieces of memory associated with the pool
2433          using <varname>VALGRIND_MEMPOOL_ALLOC</varname> will be used
2434          by the application as superblocks to dole out MALLOC_LIKE
2435          blocks using <varname>VALGRIND_MALLOCLIKE_BLOCK</varname>.
2436          In other words, a meta pool is a "2 levels" pool : first
2437          level is the blocks described
2438          by <varname>VALGRIND_MEMPOOL_ALLOC</varname>.  The second
2439          level blocks are described
2440          using <varname>VALGRIND_MALLOCLIKE_BLOCK</varname>.  Note
2441          that the association between the pool and the second level
2442          blocks is implicit : second level blocks will be located
2443          inside first level blocks. It is necessary to use
2444          the <varname>VALGRIND_MEMPOOL_METAPOOL</varname> flag for
2445          such 2 levels pools, as otherwise valgrind will detect
2446          overlapping memory blocks, and will abort execution
2447          (e.g. during leak search).
2448	</para>
2449      </listitem>
2450      <listitem>
2451	<para>
2452	  <varname>VALGRIND_MEMPOOL_AUTO_FREE</varname>.  Such a meta
2453          pool can also be marked as an 'auto free' pool using the
2454          flag <varname>VALGRIND_MEMPOOL_AUTO_FREE</varname>, which
2455          must be OR-ed together with
2456          the <varname>VALGRIND_MEMPOOL_METAPOOL</varname>. For an
2457          'auto free' pool, <varname>VALGRIND_MEMPOOL_FREE</varname>
2458          will automatically free the second level blocks that are
2459          contained inside the first level block freed
2460          with <varname>VALGRIND_MEMPOOL_FREE</varname>.  In other
2461          words, calling <varname>VALGRIND_MEMPOOL_FREE</varname> will
2462          cause implicit calls
2463          to <varname>VALGRIND_FREELIKE_BLOCK</varname> for all the
2464          second level blocks included in the first level block.
2465          Note: it is an error to use
2466          the <varname>VALGRIND_MEMPOOL_AUTO_FREE</varname> flag
2467          without the
2468         <varname>VALGRIND_MEMPOOL_METAPOOL</varname> flag.
2469	</para>
2470      </listitem>
2471    </itemizedlist>
2472  </listitem>
2473
2474  <listitem>
2475    <para><varname>VALGRIND_DESTROY_MEMPOOL(pool)</varname>:
2476    This request tells Memcheck that a pool is being torn down. Memcheck
2477    then removes all records of chunks associated with the pool, as well
2478    as its record of the pool's existence. While destroying its records of
2479    a mempool, Memcheck resets the redzones of any live chunks in the pool
2480    to NOACCESS.
2481    </para>
2482  </listitem>
2483
2484  <listitem>
2485    <para><varname>VALGRIND_MEMPOOL_ALLOC(pool, addr, size)</varname>:
2486    This request informs Memcheck that a <varname>size</varname>-byte chunk
2487    has been allocated at <varname>addr</varname>, and associates the chunk with the
2488    specified
2489    <varname>pool</varname>. If the pool was created with nonzero
2490    <varname>rzB</varname> redzones, Memcheck will mark the
2491    <varname>rzB</varname> bytes before and after the chunk as NOACCESS. If
2492    the pool was created with the <varname>is_zeroed</varname> argument set,
2493    Memcheck will mark the chunk as DEFINED, otherwise Memcheck will mark
2494    the chunk as UNDEFINED.
2495    </para>
2496  </listitem>
2497
2498  <listitem>
2499    <para><varname>VALGRIND_MEMPOOL_FREE(pool, addr)</varname>:
2500    This request informs Memcheck that the chunk at <varname>addr</varname>
2501    should no longer be considered allocated. Memcheck will mark the chunk
2502    associated with <varname>addr</varname> as NOACCESS, and delete its
2503    record of the chunk's existence.
2504    </para>
2505  </listitem>
2506
2507  <listitem>
2508    <para><varname>VALGRIND_MEMPOOL_TRIM(pool, addr, size)</varname>:
2509    This request trims the chunks associated with <varname>pool</varname>.
2510    The request only operates on chunks associated with
2511    <varname>pool</varname>. Trimming is formally defined as:</para>
2512    <itemizedlist>
2513      <listitem>
2514        <para> All chunks entirely inside the range
2515        <varname>addr..(addr+size-1)</varname> are preserved.</para>
2516      </listitem>
2517      <listitem>
2518        <para>All chunks entirely outside the range
2519        <varname>addr..(addr+size-1)</varname> are discarded, as though
2520        <varname>VALGRIND_MEMPOOL_FREE</varname> was called on them. </para>
2521      </listitem>
2522      <listitem>
2523        <para>All other chunks must intersect with the range
2524        <varname>addr..(addr+size-1)</varname>; areas outside the
2525        intersection are marked as NOACCESS, as though they had been
2526        independently freed with
2527        <varname>VALGRIND_MEMPOOL_FREE</varname>.</para>
2528      </listitem>
2529    </itemizedlist>
2530    <para>This is a somewhat rare request, but can be useful in
2531    implementing the type of mass-free operations common in custom
2532    LIFO allocators.</para>
2533  </listitem>
2534
2535  <listitem>
2536    <para><varname>VALGRIND_MOVE_MEMPOOL(poolA, poolB)</varname>: This
2537    request informs Memcheck that the pool previously anchored at
2538    address <varname>poolA</varname> has moved to anchor address
2539    <varname>poolB</varname>.  This is a rare request, typically only needed
2540    if you <function>realloc</function> the header of a mempool.</para>
2541    <para>No memory-status bits are altered by this request.</para>
2542  </listitem>
2543
2544  <listitem>
2545    <para>
2546    <varname>VALGRIND_MEMPOOL_CHANGE(pool, addrA, addrB,
2547    size)</varname>: This request informs Memcheck that the chunk
2548    previously allocated at address <varname>addrA</varname> within
2549    <varname>pool</varname> has been moved and/or resized, and should be
2550    changed to cover the region <varname>addrB..(addrB+size-1)</varname>. This
2551    is a rare request, typically only needed if you
2552    <function>realloc</function> a superblock or wish to extend a chunk
2553    without changing its memory-status bits.
2554    </para>
2555    <para>No memory-status bits are altered by this request.
2556    </para>
2557  </listitem>
2558
2559  <listitem>
2560    <para><varname>VALGRIND_MEMPOOL_EXISTS(pool)</varname>:
2561    This request informs the caller whether or not Memcheck is currently
2562    tracking a mempool at anchor address <varname>pool</varname>. It
2563    evaluates to 1 when there is a mempool associated with that address, 0
2564    otherwise. This is a rare request, only useful in circumstances when
2565    client code might have lost track of the set of active mempools.
2566    </para>
2567  </listitem>
2568
2569</itemizedlist>
2570
2571</sect1>
2572
2573
2574
2575
2576
2577
2578
2579<sect1 id="mc-manual.mpiwrap" xreflabel="MPI Wrappers">
2580<title>Debugging MPI Parallel Programs with Valgrind</title>
2581
2582<para>Memcheck supports debugging of distributed-memory applications
2583which use the MPI message passing standard.  This support consists of a
2584library of wrapper functions for the
2585<computeroutput>PMPI_*</computeroutput> interface.  When incorporated
2586into the application's address space, either by direct linking or by
2587<computeroutput>LD_PRELOAD</computeroutput>, the wrappers intercept
2588calls to <computeroutput>PMPI_Send</computeroutput>,
2589<computeroutput>PMPI_Recv</computeroutput>, etc.  They then
2590use client requests to inform Memcheck of memory state changes caused
2591by the function being wrapped.  This reduces the number of false
2592positives that Memcheck otherwise typically reports for MPI
2593applications.</para>
2594
2595<para>The wrappers also take the opportunity to carefully check
2596size and definedness of buffers passed as arguments to MPI functions, hence
2597detecting errors such as passing undefined data to
2598<computeroutput>PMPI_Send</computeroutput>, or receiving data into a
2599buffer which is too small.</para>
2600
2601<para>Unlike most of the rest of Valgrind, the wrapper library is subject to a
2602BSD-style license, so you can link it into any code base you like.
2603See the top of <computeroutput>mpi/libmpiwrap.c</computeroutput>
2604for license details.</para>
2605
2606
2607<sect2 id="mc-manual.mpiwrap.build" xreflabel="Building MPI Wrappers">
2608<title>Building and installing the wrappers</title>
2609
2610<para> The wrapper library will be built automatically if possible.
2611Valgrind's configure script will look for a suitable
2612<computeroutput>mpicc</computeroutput> to build it with.  This must be
2613the same <computeroutput>mpicc</computeroutput> you use to build the
2614MPI application you want to debug.  By default, Valgrind tries
2615<computeroutput>mpicc</computeroutput>, but you can specify a
2616different one by using the configure-time option
2617<option>--with-mpicc</option>.  Currently the
2618wrappers are only buildable with
2619<computeroutput>mpicc</computeroutput>s which are based on GNU
2620GCC or Intel's C++ Compiler.</para>
2621
2622<para>Check that the configure script prints a line like this:</para>
2623
2624<programlisting><![CDATA[
2625checking for usable MPI2-compliant mpicc and mpi.h... yes, mpicc
2626]]></programlisting>
2627
2628<para>If it says <computeroutput>... no</computeroutput>, your
2629<computeroutput>mpicc</computeroutput> has failed to compile and link
2630a test MPI2 program.</para>
2631
2632<para>If the configure test succeeds, continue in the usual way with
2633<computeroutput>make</computeroutput> and <computeroutput>make
2634install</computeroutput>.  The final install tree should then contain
2635<computeroutput>libmpiwrap-&lt;platform&gt;.so</computeroutput>.
2636</para>
2637
2638<para>Compile up a test MPI program (eg, MPI hello-world) and try
2639this:</para>
2640
2641<programlisting><![CDATA[
2642LD_PRELOAD=$prefix/lib/valgrind/libmpiwrap-<platform>.so   \
2643           mpirun [args] $prefix/bin/valgrind ./hello
2644]]></programlisting>
2645
2646<para>You should see something similar to the following</para>
2647
2648<programlisting><![CDATA[
2649valgrind MPI wrappers 31901: Active for pid 31901
2650valgrind MPI wrappers 31901: Try MPIWRAP_DEBUG=help for possible options
2651]]></programlisting>
2652
2653<para>repeated for every process in the group.  If you do not see
2654these, there is an build/installation problem of some kind.</para>
2655
2656<para> The MPI functions to be wrapped are assumed to be in an ELF
2657shared object with soname matching
2658<computeroutput>libmpi.so*</computeroutput>.  This is known to be
2659correct at least for Open MPI and Quadrics MPI, and can easily be
2660changed if required.</para>
2661</sect2>
2662
2663
2664<sect2 id="mc-manual.mpiwrap.gettingstarted"
2665       xreflabel="Getting started with MPI Wrappers">
2666<title>Getting started</title>
2667
2668<para>Compile your MPI application as usual, taking care to link it
2669using the same <computeroutput>mpicc</computeroutput> that your
2670Valgrind build was configured with.</para>
2671
2672<para>
2673Use the following basic scheme to run your application on Valgrind with
2674the wrappers engaged:</para>
2675
2676<programlisting><![CDATA[
2677MPIWRAP_DEBUG=[wrapper-args]                                  \
2678   LD_PRELOAD=$prefix/lib/valgrind/libmpiwrap-<platform>.so   \
2679   mpirun [mpirun-args]                                       \
2680   $prefix/bin/valgrind [valgrind-args]                       \
2681   [application] [app-args]
2682]]></programlisting>
2683
2684<para>As an alternative to
2685<computeroutput>LD_PRELOAD</computeroutput>ing
2686<computeroutput>libmpiwrap-&lt;platform&gt;.so</computeroutput>, you can
2687simply link it to your application if desired.  This should not disturb
2688native behaviour of your application in any way.</para>
2689</sect2>
2690
2691
2692<sect2 id="mc-manual.mpiwrap.controlling"
2693       xreflabel="Controlling the MPI Wrappers">
2694<title>Controlling the wrapper library</title>
2695
2696<para>Environment variable
2697<computeroutput>MPIWRAP_DEBUG</computeroutput> is consulted at
2698startup.  The default behaviour is to print a starting banner</para>
2699
2700<programlisting><![CDATA[
2701valgrind MPI wrappers 16386: Active for pid 16386
2702valgrind MPI wrappers 16386: Try MPIWRAP_DEBUG=help for possible options
2703]]></programlisting>
2704
2705<para> and then be relatively quiet.</para>
2706
2707<para>You can give a list of comma-separated options in
2708<computeroutput>MPIWRAP_DEBUG</computeroutput>.  These are</para>
2709
2710<itemizedlist>
2711  <listitem>
2712    <para><computeroutput>verbose</computeroutput>:
2713    show entries/exits of all wrappers.  Also show extra
2714    debugging info, such as the status of outstanding
2715    <computeroutput>MPI_Request</computeroutput>s resulting
2716    from uncompleted <computeroutput>MPI_Irecv</computeroutput>s.</para>
2717  </listitem>
2718  <listitem>
2719    <para><computeroutput>quiet</computeroutput>:
2720    opposite of <computeroutput>verbose</computeroutput>, only print
2721    anything when the wrappers want
2722    to report a detected programming error, or in case of catastrophic
2723    failure of the wrappers.</para>
2724  </listitem>
2725  <listitem>
2726    <para><computeroutput>warn</computeroutput>:
2727    by default, functions which lack proper wrappers
2728    are not commented on, just silently
2729    ignored.  This causes a warning to be printed for each unwrapped
2730    function used, up to a maximum of three warnings per function.</para>
2731  </listitem>
2732  <listitem>
2733    <para><computeroutput>strict</computeroutput>:
2734    print an error message and abort the program if
2735    a function lacking a wrapper is used.</para>
2736  </listitem>
2737</itemizedlist>
2738
2739<para> If you want to use Valgrind's XML output facility
2740(<option>--xml=yes</option>), you should pass
2741<computeroutput>quiet</computeroutput> in
2742<computeroutput>MPIWRAP_DEBUG</computeroutput> so as to get rid of any
2743extraneous printing from the wrappers.</para>
2744
2745</sect2>
2746
2747
2748<sect2 id="mc-manual.mpiwrap.limitations.functions"
2749       xreflabel="Functions: Abilities and Limitations">
2750<title>Functions</title>
2751
2752<para>All MPI2 functions except
2753<computeroutput>MPI_Wtick</computeroutput>,
2754<computeroutput>MPI_Wtime</computeroutput> and
2755<computeroutput>MPI_Pcontrol</computeroutput> have wrappers.  The
2756first two are not wrapped because they return a
2757<computeroutput>double</computeroutput>, which Valgrind's
2758function-wrap mechanism cannot handle (but it could easily be
2759extended to do so).  <computeroutput>MPI_Pcontrol</computeroutput> cannot be
2760wrapped as it has variable arity:
2761<computeroutput>int MPI_Pcontrol(const int level, ...)</computeroutput></para>
2762
2763<para>Most functions are wrapped with a default wrapper which does
2764nothing except complain or abort if it is called, depending on
2765settings in <computeroutput>MPIWRAP_DEBUG</computeroutput> listed
2766above.  The following functions have "real", do-something-useful
2767wrappers:</para>
2768
2769<programlisting><![CDATA[
2770PMPI_Send PMPI_Bsend PMPI_Ssend PMPI_Rsend
2771
2772PMPI_Recv PMPI_Get_count
2773
2774PMPI_Isend PMPI_Ibsend PMPI_Issend PMPI_Irsend
2775
2776PMPI_Irecv
2777PMPI_Wait PMPI_Waitall
2778PMPI_Test PMPI_Testall
2779
2780PMPI_Iprobe PMPI_Probe
2781
2782PMPI_Cancel
2783
2784PMPI_Sendrecv
2785
2786PMPI_Type_commit PMPI_Type_free
2787
2788PMPI_Pack PMPI_Unpack
2789
2790PMPI_Bcast PMPI_Gather PMPI_Scatter PMPI_Alltoall
2791PMPI_Reduce PMPI_Allreduce PMPI_Op_create
2792
2793PMPI_Comm_create PMPI_Comm_dup PMPI_Comm_free PMPI_Comm_rank PMPI_Comm_size
2794
2795PMPI_Error_string
2796PMPI_Init PMPI_Initialized PMPI_Finalize
2797]]></programlisting>
2798
2799<para> A few functions such as
2800<computeroutput>PMPI_Address</computeroutput> are listed as
2801<computeroutput>HAS_NO_WRAPPER</computeroutput>.  They have no wrapper
2802at all as there is nothing worth checking, and giving a no-op wrapper
2803would reduce performance for no reason.</para>
2804
2805<para> Note that the wrapper library itself can itself generate large
2806numbers of calls to the MPI implementation, especially when walking
2807complex types.  The most common functions called are
2808<computeroutput>PMPI_Extent</computeroutput>,
2809<computeroutput>PMPI_Type_get_envelope</computeroutput>,
2810<computeroutput>PMPI_Type_get_contents</computeroutput>, and
2811<computeroutput>PMPI_Type_free</computeroutput>.  </para>
2812</sect2>
2813
2814<sect2 id="mc-manual.mpiwrap.limitations.types"
2815       xreflabel="Types: Abilities and Limitations">
2816<title>Types</title>
2817
2818<para> MPI-1.1 structured types are supported, and walked exactly.
2819The currently supported combiners are
2820<computeroutput>MPI_COMBINER_NAMED</computeroutput>,
2821<computeroutput>MPI_COMBINER_CONTIGUOUS</computeroutput>,
2822<computeroutput>MPI_COMBINER_VECTOR</computeroutput>,
2823<computeroutput>MPI_COMBINER_HVECTOR</computeroutput>
2824<computeroutput>MPI_COMBINER_INDEXED</computeroutput>,
2825<computeroutput>MPI_COMBINER_HINDEXED</computeroutput> and
2826<computeroutput>MPI_COMBINER_STRUCT</computeroutput>.  This should
2827cover all MPI-1.1 types.  The mechanism (function
2828<computeroutput>walk_type</computeroutput>) should extend easily to
2829cover MPI2 combiners.</para>
2830
2831<para>MPI defines some named structured types
2832(<computeroutput>MPI_FLOAT_INT</computeroutput>,
2833<computeroutput>MPI_DOUBLE_INT</computeroutput>,
2834<computeroutput>MPI_LONG_INT</computeroutput>,
2835<computeroutput>MPI_2INT</computeroutput>,
2836<computeroutput>MPI_SHORT_INT</computeroutput>,
2837<computeroutput>MPI_LONG_DOUBLE_INT</computeroutput>) which are pairs
2838of some basic type and a C <computeroutput>int</computeroutput>.
2839Unfortunately the MPI specification makes it impossible to look inside
2840these types and see where the fields are.  Therefore these wrappers
2841assume the types are laid out as <computeroutput>struct { float val;
2842int loc; }</computeroutput> (for
2843<computeroutput>MPI_FLOAT_INT</computeroutput>), etc, and act
2844accordingly.  This appears to be correct at least for Open MPI 1.0.2
2845and for Quadrics MPI.</para>
2846
2847<para>If <computeroutput>strict</computeroutput> is an option specified
2848in <computeroutput>MPIWRAP_DEBUG</computeroutput>, the application
2849will abort if an unhandled type is encountered.  Otherwise, the
2850application will print a warning message and continue.</para>
2851
2852<para>Some effort is made to mark/check memory ranges corresponding to
2853arrays of values in a single pass.  This is important for performance
2854since asking Valgrind to mark/check any range, no matter how small,
2855carries quite a large constant cost.  This optimisation is applied to
2856arrays of primitive types (<computeroutput>double</computeroutput>,
2857<computeroutput>float</computeroutput>,
2858<computeroutput>int</computeroutput>,
2859<computeroutput>long</computeroutput>, <computeroutput>long
2860long</computeroutput>, <computeroutput>short</computeroutput>,
2861<computeroutput>char</computeroutput>, and <computeroutput>long
2862double</computeroutput> on platforms where <computeroutput>sizeof(long
2863double) == 8</computeroutput>).  For arrays of all other types, the
2864wrappers handle each element individually and so there can be a very
2865large performance cost.</para>
2866
2867</sect2>
2868
2869
2870<sect2 id="mc-manual.mpiwrap.writingwrappers"
2871       xreflabel="Writing new MPI Wrappers">
2872<title>Writing new wrappers</title>
2873
2874<para>
2875For the most part the wrappers are straightforward.  The only
2876significant complexity arises with nonblocking receives.</para>
2877
2878<para>The issue is that <computeroutput>MPI_Irecv</computeroutput>
2879states the recv buffer and returns immediately, giving a handle
2880(<computeroutput>MPI_Request</computeroutput>) for the transaction.
2881Later the user will have to poll for completion with
2882<computeroutput>MPI_Wait</computeroutput> etc, and when the
2883transaction completes successfully, the wrappers have to paint the
2884recv buffer.  But the recv buffer details are not presented to
2885<computeroutput>MPI_Wait</computeroutput> -- only the handle is.  The
2886library therefore maintains a shadow table which associates
2887uncompleted <computeroutput>MPI_Request</computeroutput>s with the
2888corresponding buffer address/count/type.  When an operation completes,
2889the table is searched for the associated address/count/type info, and
2890memory is marked accordingly.</para>
2891
2892<para>Access to the table is guarded by a (POSIX pthreads) lock, so as
2893to make the library thread-safe.</para>
2894
2895<para>The table is allocated with
2896<computeroutput>malloc</computeroutput> and never
2897<computeroutput>free</computeroutput>d, so it will show up in leak
2898checks.</para>
2899
2900<para>Writing new wrappers should be fairly easy.  The source file is
2901<computeroutput>mpi/libmpiwrap.c</computeroutput>.  If possible,
2902find an existing wrapper for a function of similar behaviour to the
2903one you want to wrap, and use it as a starting point.  The wrappers
2904are organised in sections in the same order as the MPI 1.1 spec, to
2905aid navigation.  When adding a wrapper, remember to comment out the
2906definition of the default wrapper in the long list of defaults at the
2907bottom of the file (do not remove it, just comment it out).</para>
2908</sect2>
2909
2910<sect2 id="mc-manual.mpiwrap.whattoexpect"
2911       xreflabel="What to expect with MPI Wrappers">
2912<title>What to expect when using the wrappers</title>
2913
2914<para>The wrappers should reduce Memcheck's false-error rate on MPI
2915applications.  Because the wrapping is done at the MPI interface,
2916there will still potentially be a large number of errors reported in
2917the MPI implementation below the interface.  The best you can do is
2918try to suppress them.</para>
2919
2920<para>You may also find that the input-side (buffer
2921length/definedness) checks find errors in your MPI use, for example
2922passing too short a buffer to
2923<computeroutput>MPI_Recv</computeroutput>.</para>
2924
2925<para>Functions which are not wrapped may increase the false
2926error rate.  A possible approach is to run with
2927<computeroutput>MPI_DEBUG</computeroutput> containing
2928<computeroutput>warn</computeroutput>.  This will show you functions
2929which lack proper wrappers but which are nevertheless used.  You can
2930then write wrappers for them.
2931</para>
2932
2933<para>A known source of potential false errors are the
2934<computeroutput>PMPI_Reduce</computeroutput> family of functions, when
2935using a custom (user-defined) reduction function.  In a reduction
2936operation, each node notionally sends data to a "central point" which
2937uses the specified reduction function to merge the data items into a
2938single item.  Hence, in general, data is passed between nodes and fed
2939to the reduction function, but the wrapper library cannot mark the
2940transferred data as initialised before it is handed to the reduction
2941function, because all that happens "inside" the
2942<computeroutput>PMPI_Reduce</computeroutput> call.  As a result you
2943may see false positives reported in your reduction function.</para>
2944
2945</sect2>
2946
2947</sect1>
2948
2949
2950
2951
2952
2953</chapter>
2954