1Dark Matter Detector (DMD)
2==========================
3
4.. note::
5
6    This section was copied with minimal modification from
7    https://developer.mozilla.org/en-US/docs/Mozilla/Performance/DMD. Some
8    parts of the documentation may be out-of-date.
9
10DMD (short for "dark matter detector") is a heap profiler within Firefox. It
11has four modes.
12
13* "Dark Matter" mode. In this mode, DMD tracks the contents of the heap,
14  including which heap blocks have been reported by memory reporters. It helps
15  us reduce the "heap-unclassified" value in Firefox's about:memory page, and
16  also detects if any heap blocks are reported twice. Originally, this was the
17  only mode that DMD had, which explains DMD's name. This is the default mode.
18
19* "Live" mode. In this mode, DMD tracks the current contents of the heap. You
20  can dump that information to file, giving a profile of the live heap blocks
21  at that point in time. This is good for understanding how memory is used at
22  an interesting point in time, such as peak memory usage.
23
24* "Cumulative" mode. In this mode, DMD tracks both the past and current
25  contents of the heap. You can dump that information to file, giving a profile
26  of the heap usage for the entire session. This is good for finding parts of
27  the code that cause high heap churn, e.g. by allocating many short-lived
28  allocations.
29
30* "Heap scanning" mode. This mode is like live mode, but it also records the
31  contents of every live block in the log. This can be used to investigate
32  leaks by figuring out which objects might be holding references to other
33  objects.
34
35
36Building and Running
37--------------------
38
39Nightly Firefox
40~~~~~~~~~~~~~~~
41
42The easiest way to use DMD is with the normal Nightly Firefox build, which
43has DMD already enabled in the build. To have DMD active while running it,
44you just need to set the environment variable ``DMD=1`` when running. For
45instance, on OSX, you can run something like:
46
47.. code-block: bash
48
49    DMD=1 /Applications/Firefox\ Nightly.app/Contents/MacOS/firefox
50
51You can tell it is working by going to ``about:memory`` and looking for "Save
52DMD Output". If DMD has been properly enabled, the "Save" button won't be
53grayed out. Look at the "Trigger" section below to see the full list of ways
54to get a DMD report once you have it activated. Note that stack information
55you get will likely be less detailed, due to being unable to symbolicate. You
56will be able to get function names, but not line numbers.
57
58Processing the output
59---------------------
60
61DMD outputs one gzipped JSON file per process that contains a description of
62that process's heap. You can analyze these files (either gzipped or not)
63using ``dmd.py``. On Nightly Firefox, ``dmd.py`` is included in the
64distribution. For instance on OS X, it is located in the directory
65``/Applications/Firefox Nightly.app/Contents/Resources/``. For Nightly,
66symbolication will fail, but you can at least get some information. In a
67local build, ``dmd.py`` will be located in the directory
68``$OBJDIR/dist/bin/``.
69
70Some platforms (Linux, Mac, Android) require stack fixing, which adds missing
71filename, function name and line number information. This will occur
72automatically the first time you run ``dmd.py`` on the output file. This can
73take 10s of seconds or more to complete. (This will fail if your build does
74not contain symbols. However, if you have crash reporter symbols for your
75build (as tryserver builds do) you can use this script instead: clone the
76whole repo, edit the paths at the top of ``resymbolicate_dmd.py`` and run
77it.) The simplest way to do this is to just run the ``dmd.py`` script on your
78DMD report while your working directory is ``$OBJDIR/dist/bin``. This will
79allow the local libraries to be found and used.
80
81If you invoke ``dmd.py`` without arguments you will get output appropriate
82for the mode in which DMD was invoked.
83
84"Dark matter" mode output
85~~~~~~~~~~~~~~~~~~~~~~~~~
86
87For "dark matter" mode, ``dmd.py``'s output describes how the live heap blocks
88are covered by memory reports. This output is broken into multiple sections.
89
901. "Invocation". This tells you how DMD was invoked, i.e. what options were used.
91
922. "Twice-reported stack trace records". This tells you which heap blocks
93   were reported twice or more. The presence of any such records indicates bugs
94   in one or more memory reporters.
95
963. "Unreported stack trace records". This tells you which heap blocks were
97   not reported, which indicate where additional memory reporters would be most
98   helpful.
99
1004. "Once-reported stack trace records": like the "Unreported stack trace
101   records" section, but for blocks reported once.
102
1035. "Summary": gives measurements of the total heap, and the
104   unreported/once-reported/twice-reported portions of it.
105
106The "Twice-reported stack trace records" and "Unreported stack trace records"
107sections are the most important, because they indicate ways in which the
108memory reporters can be improved.
109
110Here's an example stack trace record from the "Unreported stack trace
111records" section.
112
113.. code-block::
114
115    Unreported {
116      150 blocks in heap block record 283 of 5,495
117      21,600 bytes (20,400 requested / 1,200 slop)
118      Individual block sizes: 144 x 150
119      0.00% of the heap (16.85% cumulative)
120      0.02% of unreported (94.68% cumulative)
121      Allocated at {
122        #01: replace_malloc (/home/njn/moz/mi5/go64dmd/memory/replace/dmd/../../../../memory/replace/dmd/DMD.cpp:1286)
123        #02: malloc (/home/njn/moz/mi5/go64dmd/memory/build/../../../memory/build/replace_malloc.c:153)
124        #03: moz_xmalloc (/home/njn/moz/mi5/memory/mozalloc/mozalloc.cpp:84)
125        #04: nsCycleCollectingAutoRefCnt::incr(void*, nsCycleCollectionParticipant*) (/home/njn/moz/mi5/go64dmd/dom/xul/../../dist/include/nsISupportsImpl.h:250)
126        #05: nsXULElement::Create(nsXULPrototypeElement*, nsIDocument*, bool, bool,mozilla::dom::Element**) (/home/njn/moz/mi5/dom/xul/nsXULElement.cpp:287)
127        #06: nsXBLContentSink::CreateElement(char16_t const**, unsigned int, mozilla::dom::NodeInfo*, unsigned int, nsIContent**, bool*, mozilla::dom::FromParser) (/home/njn/moz/mi5/dom/xbl/nsXBLContentSink.cpp:874)
128        #07: nsCOMPtr<nsIContent>::StartAssignment() (/home/njn/moz/mi5/go64dmd/dom/xml/../../dist/include/nsCOMPtr.h:753)
129        #08: nsXMLContentSink::HandleStartElement(char16_t const*, char16_t const**, unsigned int, unsigned int, bool) (/home/njn/moz/mi5/dom/xml/nsXMLContentSink.cpp:1007)
130      }
131    }
132
133It tells you that there were 150 heap blocks that were allocated from the
134program point indicated by the "Allocated at" stack trace, that these blocks
135took up 21,600 bytes, that all 150 blocks had a size of 144 bytes, and that
1361,200 of those bytes were "slop" (wasted space caused by the heap allocator
137rounding up request sizes). It also indicates what percentage of the total
138heap size and the unreported portion of the heap these blocks represent.
139
140Within each section, records are listed from largest to smallest.
141
142Once-reported and twice-reported stack trace records also have stack traces for the report point(s). For example:
143
144.. code-block::
145
146    Reported at {
147      #01: mozilla::dmd::Report(void const*) (/home/njn/moz/mi2/memory/replace/dmd/DMD.cpp:1740) 0x7f68652581ca
148      #02: CycleCollectorMallocSizeOf(void const*) (/home/njn/moz/mi2/xpcom/base/nsCycleCollector.cpp:3008) 0x7f6860fdfe02
149      #03: nsPurpleBuffer::SizeOfExcludingThis(unsigned long (*)(void const*)) const (/home/njn/moz/mi2/xpcom/base/nsCycleCollector.cpp:933) 0x7f6860fdb7af
150      #04: nsCycleCollector::SizeOfIncludingThis(unsigned long (*)(void const*), unsigned long*, unsigned long*, unsigned long*, unsigned long*, unsigned long*) const (/home/njn/moz/mi2/xpcom/base/nsCycleCollector.cpp:3029) 0x7f6860fdb6b1
151      #05: CycleCollectorMultiReporter::CollectReports(nsIMemoryMultiReporterCallback*, nsISupports*) (/home/njn/moz/mi2/xpcom/base/nsCycleCollector.cpp:3075) 0x7f6860fde432
152      #06: nsMemoryInfoDumper::DumpMemoryReportsToFileImpl(nsAString_internal const&) (/home/njn/moz/mi2/xpcom/base/nsMemoryInfoDumper.cpp:626) 0x7f6860fece79
153      #07: nsMemoryInfoDumper::DumpMemoryReportsToFile(nsAString_internal const&, bool, bool) (/home/njn/moz/mi2/xpcom/base/nsMemoryInfoDumper.cpp:344) 0x7f6860febaf9
154      #08: mozilla::(anonymous namespace)::DumpMemoryReportsRunnable::Run() (/home/njn/moz/mi2/xpcom/base/nsMemoryInfoDumper.cpp:58) 0x7f6860fefe03
155    }
156
157You can tell which memory reporter made the report by the name of the
158``MallocSizeOf`` function near the top of the stack trace. In this case it
159was the cycle collector's reporter.
160
161By default, DMD does not record an allocation stack trace for most blocks, to
162make it run faster. The decision on whether to record is done
163probabilistically, and larger blocks are more likely to have an allocation
164stack trace recorded. All unreported blocks that lack an allocation stack
165trace will end up in a single record. For example:
166
167.. code-block::
168
169    Unreported {
170      420,010 blocks in heap block record 2 of 5,495
171      29,203,408 bytes (27,777,288 requested / 1,426,120 slop)
172      Individual block sizes: 2,048 x 3; 1,024 x 103; 512 x 147; 496 x 7; 480 x 31; 464 x 6; 448 x 50; 432 x 41; 416 x 28; 400 x 53; 384 x 43; 368 x 216; 352 x 141; 336 x 58; 320 x 104; 304 x 5,130; 288 x 150; 272 x 591; 256 x 6,017; 240 x 1,372; 224 x 93; 208 x 488; 192 x 1,919; 176 x 18,903; 160 x 1,754; 144 x 5,041; 128 x 36,709; 112 x 5,571; 96 x 6,280; 80 x 40,738; 64 x 37,925; 48 x 78,392; 32 x 136,199; 16 x 31,001; 8 x 4,706
173      3.78% of the heap (10.24% cumulative)
174      21.24% of unreported (57.53% cumulative)
175      Allocated at {
176        #01: (no stack trace recorded due to --stacks=partial)
177      }
178    }
179
180In contrast, stack traces are always recorded when a block is reported, which
181means you can end up with records like this where the allocation point is
182unknown but the reporting point is known:
183
184.. code-block::
185
186    Once-reported {
187      104,491 blocks in heap block record 13 of 4,689
188      10,392,000 bytes (10,392,000 requested / 0 slop)
189      Individual block sizes: 512 x 124; 256 x 242; 192 x 813; 128 x 54,664; 64 x 48,648
190      1.35% of the heap (48.65% cumulative)
191      1.64% of once-reported (59.18% cumulative)
192      Allocated at {
193        #01: (no stack trace recorded due to --stacks=partial)
194      }
195      Reported at {
196        #01: mozilla::dmd::DMDFuncs::Report(void const*) (/home/njn/moz/mi5/go64dmd/memory/replace/dmd/../../../../memory/replace/dmd/DMD.cpp:1646)
197        #02: WindowsMallocSizeOf(void const*) (/home/njn/moz/mi5/dom/base/nsWindowMemoryReporter.cpp:189)
198        #03: nsAttrAndChildArray::SizeOfExcludingThis(unsigned long (*)(void const*)) const (/home/njn/moz/mi5/dom/base/nsAttrAndChildArray.cpp:880)
199        #04: mozilla::dom::FragmentOrElement::SizeOfExcludingThis(unsigned long (*)(void const*)) const (/home/njn/moz/mi5/dom/base/FragmentOrElement.cpp:2337)
200        #05: nsINode::SizeOfIncludingThis(unsigned long (*)(void const*)) const (/home/njn/moz/mi5/go64dmd/parser/html/../../../dom/base/nsINode.h:307)
201        #06: mozilla::dom::NodeInfo::NodeType() const (/home/njn/moz/mi5/go64dmd/dom/base/../../dist/include/mozilla/dom/NodeInfo.h:127)
202        #07: nsHTMLDocument::DocAddSizeOfExcludingThis(nsWindowSizes*) const (/home/njn/moz/mi5/dom/html/nsHTMLDocument.cpp:3710)
203        #08: nsIDocument::DocAddSizeOfIncludingThis(nsWindowSizes*) const (/home/njn/moz/mi5/dom/base/nsDocument.cpp:12820)
204      }
205    }
206
207The choice of whether to record an allocation stack trace for all blocks is controlled by an option (see below).
208
209"Live" mode output
210~~~~~~~~~~~~~~~~~~
211
212For "live" mode, dmd.py's output describes what live heap blocks are present.
213This output is broken into multiple sections.
214
2151. "Invocation". This tells you how DMD was invoked, i.e. what options were used.
2162. "Live stack trace records". This tells you which heap blocks were present.
2173. "Summary": gives measurements of the total heap.
218
219The individual records are similar to those output in "dark matter" mode.
220
221
222"Cumulative" mode output
223~~~~~~~~~~~~~~~~~~~~~~~~
224
225For "cumulative" mode, dmd.py's output describes how the live heap blocks are
226covered by memory reports. This output is broken into multiple sections.
227
2281. "Invocation". This tells you how DMD was invoked, i.e. what options were used.
2292. "Cumulative stack trace records". This tells you which heap blocks were allocated during the session.
2303. "Summary": gives measurements of the total (cumulative) heap.
231
232The individual records are similar to those output in "dark matter" mode.
233
234
235"Scan" mode output
236~~~~~~~~~~~~~~~~~~
237
238For "scan" mode, the output of ``dmd.py`` is the same as "live" mode. A
239separate script, ``block_analyzer.py``, can be used to find out information
240about which blocks refer to a particular block. ``dmd.py --clamp-contents``
241needs to be run on the log first. See this other page for an overview of how
242to use heap scan mode to fix a leak involving refcounted objects.
243
244Options
245-------
246
247Runtime
248~~~~~~~
249
250When you run ``mach run --dmd`` you can specify additional options to control
251how DMD runs. Run ``mach help run`` for documentation on these.
252
253The most interesting one is ``--mode``. Acceptable values are ``dark-matter``
254(the default), ``live``, ``cumulative``, and ``scan``.
255
256Another interesting one is ``--stacks``. Acceptable values are ``partial``
257(the default) and ``full``. In the former case most blocks will not have an
258allocation stack trace recorded. However, because larger blocks are more
259likely to have one recorded, most allocated bytes should have an allocation
260stack trace even though most allocated blocks do not. Use ``--stacks=full``
261if you want complete information, but note that DMD will run substantially
262slower in that case.
263
264The options may also be put in the environment variable DMD, or set DMD to 1
265to enable DMD with default options (dark-matter and partial stacks).
266
267The ``MOZ_DMD_SHUTDOWN_LOG`` environment variable, if set, triggers a DMD run
268at shutdown; its value must be a directory where the logs will be placed.
269Which processes get logged is controlled by the ``MOZ_DMD_LOG_PROCESS``
270environment variable, which can take the following values.
271
272* Unset: log all processes.
273* "default": log the parent process only.
274* "tab": log content processes only.
275
276For example, if you have
277
278.. code-block::
279
280    MOZ_DMD_SHUTDOWN_LOG=~/dmdlogs/ MOZ_DMD_LOG_PROCESS=tab
281
282then DMD logs for content processes will be saved to ~/dmdlogs/.
283
284.. note::
285
286    To dump DMD data from Content processes, you'll need to disable the sandbox
287
288.. note::
289
290    ``MOZ_DMD_SHUTDOWN_LOG`` must (currently) include the trailing separator (''/")
291
292
293Post-processing
294~~~~~~~~~~~~~~~
295
296``dmd.py`` also takes options that control how it works. Run ``dmd.py -h``
297for documentation. The following options are the most interesting ones.
298
299* ``-f`` / ``--max-frames``. By default, records show up to 8 stack frames. You
300  can choose a smaller number, in which case more allocations will be
301  aggregated into each record, but you'll have less context. Or you can choose
302  a larger number, in which cases allocations will be split across more
303  records, but you will have more context. There is no single best value, but
304  values in the range 2..10 are often good. The maximum is 24.
305
306* ``-a`` / ``--ignore-alloc-frames``. Many allocation stack traces start with
307  multiple frames that mention allocation wrapper functions, e.g.
308  ``js_calloc()`` calls replace_calloc(). This option filters these out. It
309  often helps improve the quality of the output when using a small
310  ``--max-frames`` value.
311
312* ``-s`` / ``--sort-by``. This controls how records are sorted. Acceptable
313  values are usable (the default), ``req``, ``slop`` and ``num-blocks``.
314
315* ``--clamp-contents``. For a heap scan log, this performs a conservative
316  pointer analysis on the contents of each block, changing any value that is a
317  pointer into the middle of a live block into a pointer to the start of that
318  block. All other values are changes to null. In addition, all trailing nulls
319  are removed from the block contents.
320
321As an example that combines multiple options, if you apply the following
322command to a profile obtained in "live" mode:
323
324.. code-block::
325
326    dmd.py -r -f 2 -a -s slop
327
328it will give you a good idea of where the major sources of slop are.
329
330``dmd.py`` can also compute the difference between two DMD output files, so
331long as those files were produced in the same mode. Simply pass it two
332filenames instead of one to get the difference.
333
334
335Which heap blocks are reported?
336-------------------------------
337
338At this stage you might wonder how DMD knows, in "dark matter" mode, which
339allocations have been reported and which haven't. DMD only knows about heap
340blocks that are measured via a function created with one of the following two
341macros:
342
343.. code-block::
344
345    MOZ_DEFINE_MALLOC_SIZE_OF
346    MOZ_DEFINE_MALLOC_SIZE_OF_ON_ALLOC
347
348Fortunately, most of the existing memory reporters do this.
349