DMD

DMD (short for "dark matter detector") is a heap profiler within Firefox. It has four modes.

  • "Dark Matter" mode. In this mode, DMD tracks the contents of the heap, including which heap blocks have been reported by memory reporters. It helps us reduce the "heap-unclassified" value in Firefox's about:memory page, and also detects if any heap blocks are reported twice. Originally, this was the only mode that DMD had, which explains DMD's name. This is the default mode.
  • "Live" mode. In this mode, DMD tracks the current contents of the heap. You can dump that information to file, giving a profile of the live heap blocks at that point in time. This is good for understanding how memory is used at an interesting point in time, such as peak memory usage.
  • "Cumulative" mode. In this mode, DMD tracks both the past and current contents of the heap. You can dump that information to file, giving a profile of the heap usage for the entire session. This is good for finding parts of the code that cause high heap churn, e.g. by allocating many short-lived allocations.
  • "Heap scanning" mode. This mode is like live mode, but it also records the contents of every live block in the log. This can be used to investigate leaks by figuring out which objects might be holding references to other objects.

Building and Running

Desktop Firefox (Linux)

Build

Build Firefox with these options:

ac_add_options --enable-dmd

If building via try server, modify browser/config/mozconfigs/linux64/common-opt or a similar file before pushing.

Launch

Use mach run --dmd; use --mode to choose the mode. Add --debug to run under gdb.

Trigger

There are three ways to trigger a DMD snapshot.

  1. Visit about:memory and click the DMD button (depending on how old your build is, it might be labelled "Save" or "Analyze reports" or "DMD"). The button won't be present in non-DMD builds, and will be grayed out in DMD builds if DMD isn't enabled at start-up.
  2. If you wish to drigger DMD dumps from within C++ or JavaScript code, you can use nsIMemoryInfoDumper.dumpMemoryToTempDir. For example, from JavaScript code you can do the following.
    const Cc = Components.classes;
    let mydumper = Cc["@mozilla.org/memory-info-dumper;1"]
                     .getService(Ci.nsIMemoryInfoDumper);
    mydumper.dumpMemoryInfoToTempDir(identifier, anonymize, minimize);
    
    This will dump memory reports and DMD output to the temporary directory. identifier is a string that will be used for part of the filename (or a timestamp will be used if it is an empty string); anonymize is a boolean that indicates if the memory reports should be anonymized; and minimize is a boolean that indicates if memory usage should be minimized first.
  3. (Linux only) You can send signal 34 to the firefox process, e.g. with the following command.
    $ killall -34 firefox

Each one of these steps triggers all the memory reporters and then DMD analyzes the reports, printing commentary like this:

DMD[5222] opened /tmp/dmd-1414556492-5222.json.gz for writing
DMD[5222] Dump 1 {
DMD[5222]   Constructing the heap block list...
DMD[5222]   Constructing the stack trace table...
DMD[5222]   Constructing the stack frame table...
DMD[5222] }

In an e10s-enabled build, you'll see separate output for each process. This step can take 10 or more seconds and may make Firefox freeze temporarily.

If you see the "opened" line, it tells you where the file was saved. It's always in a temp directory, and the filenames are always of the form dmd-<pid>.

Desktop Firefox (Mac)

Build

Build with these options:

ac_add_options --enable-dmd

If building via try server, modify browser/config/mozconfigs/macosx64/common-opt or a similar file before pushing.

Launch

Use mach run --dmd; use --mode to choose the mode. Add --debug to run under lldb.

Trigger

Follow the Trigger instructions for Linux. Note that on Mac this step can take 30+ seconds.

Desktop Firefox (Windows)

Build

Build with these options:

ac_add_options --enable-dmd

If building via try server, modify browser/config/mozconfigs/win32/common-opt. Also, add this line to build/mozconfig.common:

MOZ_CRASHREPORTER_UPLOAD_FULL_SYMBOLS=1

Launch

On a local build, use mach run --dmd; use --mode to choose the mode.

On a build done by the try server, follow these instructions instead.

Trigger

Follow the Trigger instructions for Linux.

Fennec

In order to use DMD on Fennec you will need root access on the Android device. Instructions on how to root your device is outside the scope of this document.

Build

Build with these options:

ac_add_options --enable-dmd

Prep

In order to prepare your device for running Fennec with DMD enabled, you will need to do a few things. First, you will need to push the libdmd.so library to the device so that it can by dynamically loaded by Fennec. You can do this by running:

adb push $OBJDIR/dist/bin/libdmd.so /sdcard/

Second, you will need to make an executable wrapper for Fennec which sets an environment variable before launching it. (If you are familiar with the recommended "--es env0" method for setting environment variables when launching Fennec, note that you cannot use this method here because those are processed too late in the startup process. If you are not familiar with that method, you can ignore this parenthetical note.) First make the executable wrapper on your host machine using the editor of your choice. Name the file dmd_fennec and enter this as the contents:

#!/system/bin/sh
export MOZ_REPLACE_MALLOC_LIB=/sdcard/libdmd.so
exec "$@"

If you want to use other DMD options, you can enter additional environment variables above. You will need to push this to the device and make it executable. Since you cannot mark files in /sdcard/ as executable, we will use /data/local/tmp for this purpose:

adb push dmd_fennec /data/local/tmp
adb shell
cd /data/local/tmp
chmod 755 dmd_fennec

The final step is to make Android use the above wrapper script while launching Fennec, so that the environment variable is present when Fennec starts up. Assuming you have done a local build, the app identifier will be org.mozilla.fennec_$USERNAME ($USERNAME is your username on the host machine) and so we do this as shown below. If you are using a DMD-enabled try build, or build from other source, adjust the app identifier as necessary.

adb shell
su    # You need root access for the setprop command to take effect
setprop wrap.org.mozilla.fennec_$USERNAME "/data/local/tmp/dmd_fennec"

Once this is set up, starting the org.mozilla.fennec_$USERNAME app will use the wrapper script.

Launch

Launch Fennec either by tapping on the icon as usual, or from the command line (as before, be sure to replace org.mozilla.fennec_$USERNAME with the app identifier as appropriate).

adb shell am start -n org.mozilla.fennec_$USERNAME/.App

Trigger

Use the existing memory-report dumping hook:

adb shell am broadcast -a org.mozilla.gecko.MEMORY_DUMP

In logcat, you should see output similar to this:

I/DMD     (20731): opened /storage/emulated/0/Download/memory-reports/dmd-default-20731.json.gz for writing
...
I/GeckoConsole(20731): nsIMemoryInfoDumper dumped reports to /storage/emulated/0/Download/memory-reports/unified-memory-report-default-20731.json.gz

The path is where the memory reports and DMD reports get dumped to. You can pull them like so:

adb pull /sdcard/Download/memory-reports/dmd-default-20731.json.gz
adb pull /sdcard/Download/memory-reports/unified-memory-report-default-20731.json.gz

Processing the output

DMD outputs one gzipped JSON file per process that contains a description of that process's heap. You can analyze these files (either gzipped or not) using $OBJDIR/dist/bin/dmd.py.

Some platforms (Linux, Mac, Android) require stack fixing, which adds missing filename, function name and line number information. This will occur automatically the first time you run dmd.py on the output file. This can take several minutes to complete. (This will fail if your build does not contain symbols. However, if you have crash reporter symbols for your build -- as tryserver builds do -- you can use this script instead: clone the whole repo, edit the paths at the top of resymbolicate_dmd.py and run it.) The simplest way to do this is to just run the dmd.py script on your DMD report while your working directory is $OBJDIR/dist/bin. This will allow the local libraries to be found and used.

If you invoke dmd.py without arguments you will get output appropriate for the mode in which DMD was invoked.

"Dark matter" mode output

For "dark matter" mode, dmd.py's output describes how the live heap blocks are covered by memory reports. This output is broken into multiple sections.

  1. "Invocation". This tells you how DMD was invoked, i.e. what options were used.
  2. "Twice-reported stack trace records". This tells you which heap blocks were reported twice or more. The presence of any such records indicates bugs in one or more memory reporters.
  3. "Unreported stack trace records". This tells you which heap blocks were not reported, which indicate where additional memory reporters would be most helpful.
  4. "Once-reported stack trace records": like the "Unreported stack trace records" section, but for blocks reported once.
  5. "Summary": gives measurements of the total heap, and the unreported/once-reported/twice-reported portions of it.

The "Twice-reported stack trace records" and "Unreported stack trace records" sections are the most important, because they indicate ways in which the memory reporters can be improved.

Here's an example stack trace record from the "Unreported stack trace records" section.

Unreported {
  150 blocks in heap block record 283 of 5,495
  21,600 bytes (20,400 requested / 1,200 slop)
  Individual block sizes: 144 x 150
  0.00% of the heap (16.85% cumulative)
  0.02% of unreported (94.68% cumulative)
  Allocated at {
    #01: replace_malloc (/home/njn/moz/mi5/go64dmd/memory/replace/dmd/../../../../memory/replace/dmd/DMD.cpp:1286)
    #02: malloc (/home/njn/moz/mi5/go64dmd/memory/build/../../../memory/build/replace_malloc.c:153)
    #03: moz_xmalloc (/home/njn/moz/mi5/memory/mozalloc/mozalloc.cpp:84)
    #04: nsCycleCollectingAutoRefCnt::incr(void*, nsCycleCollectionParticipant*) (/home/njn/moz/mi5/go64dmd/dom/xul/../../dist/include/nsISupportsImpl.h:250)
    #05: nsXULElement::Create(nsXULPrototypeElement*, nsIDocument*, bool, bool,mozilla::dom::Element**) (/home/njn/moz/mi5/dom/xul/nsXULElement.cpp:287)
    #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)
    #07: nsCOMPtr<nsIContent>::StartAssignment() (/home/njn/moz/mi5/go64dmd/dom/xml/../../dist/include/nsCOMPtr.h:753)
    #08: nsXMLContentSink::HandleStartElement(char16_t const*, char16_t const**, unsigned int, unsigned int, bool) (/home/njn/moz/mi5/dom/xml/nsXMLContentSink.cpp:1007)
  }
}

It tells you that there were 150 heap blocks that were allocated from the program point indicated by the "Allocated at" stack trace, that these blocks took up 21,600 bytes, that all 150 blocks had a size of 144 bytes, and that 1,200 of those bytes were "slop" (wasted space caused by the heap allocator rounding up request sizes). It also indicates what percentage of the total heap size and the unreported portion of the heap these blocks represent.

Within each section, records are listed from largest to smallest.

Once-reported and twice-reported stack trace records also have stack traces for the report point(s). For example:

Reported at {
  #01: mozilla::dmd::Report(void const*) (/home/njn/moz/mi2/memory/replace/dmd/DMD.cpp:1740) 0x7f68652581ca
  #02: CycleCollectorMallocSizeOf(void const*) (/home/njn/moz/mi2/xpcom/base/nsCycleCollector.cpp:3008) 0x7f6860fdfe02
  #03: nsPurpleBuffer::SizeOfExcludingThis(unsigned long (*)(void const*)) const (/home/njn/moz/mi2/xpcom/base/nsCycleCollector.cpp:933) 0x7f6860fdb7af
  #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
  #05: CycleCollectorMultiReporter::CollectReports(nsIMemoryMultiReporterCallback*, nsISupports*) (/home/njn/moz/mi2/xpcom/base/nsCycleCollector.cpp:3075) 0x7f6860fde432
  #06: nsMemoryInfoDumper::DumpMemoryReportsToFileImpl(nsAString_internal const&) (/home/njn/moz/mi2/xpcom/base/nsMemoryInfoDumper.cpp:626) 0x7f6860fece79
  #07: nsMemoryInfoDumper::DumpMemoryReportsToFile(nsAString_internal const&, bool, bool) (/home/njn/moz/mi2/xpcom/base/nsMemoryInfoDumper.cpp:344) 0x7f6860febaf9
  #08: mozilla::(anonymous namespace)::DumpMemoryReportsRunnable::Run() (/home/njn/moz/mi2/xpcom/base/nsMemoryInfoDumper.cpp:58) 0x7f6860fefe03
}

You can tell which memory reporter made the report by the name of the MallocSizeOf function near the top of the stack trace. In this case it was the cycle collector's reporter.

By default, DMD does not record an allocation stack trace for most blocks, to make it run faster. The decision on whether to record is done probabilistically, and larger blocks are more likely to have an allocation stack trace recorded. All unreported blocks that lack an allocation stack trace will end up in a single record. For example:

Unreported {
  420,010 blocks in heap block record 2 of 5,495
  29,203,408 bytes (27,777,288 requested / 1,426,120 slop)
  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
  3.78% of the heap (10.24% cumulative)
  21.24% of unreported (57.53% cumulative)
  Allocated at {
    #01: (no stack trace recorded due to --stacks=partial)
  }
}

In contrast, stack traces are always recorded when a block is reported, which means you can end up with records like this where the allocation point is unknown but the reporting point is known:

Once-reported {
  104,491 blocks in heap block record 13 of 4,689
  10,392,000 bytes (10,392,000 requested / 0 slop)
  Individual block sizes: 512 x 124; 256 x 242; 192 x 813; 128 x 54,664; 64 x 48,648
  1.35% of the heap (48.65% cumulative)
  1.64% of once-reported (59.18% cumulative)
  Allocated at {
    #01: (no stack trace recorded due to --stacks=partial)
  }
  Reported at {
    #01: mozilla::dmd::DMDFuncs::Report(void const*) (/home/njn/moz/mi5/go64dmd/memory/replace/dmd/../../../../memory/replace/dmd/DMD.cpp:1646)
    #02: WindowsMallocSizeOf(void const*) (/home/njn/moz/mi5/dom/base/nsWindowMemoryReporter.cpp:189)
    #03: nsAttrAndChildArray::SizeOfExcludingThis(unsigned long (*)(void const*)) const (/home/njn/moz/mi5/dom/base/nsAttrAndChildArray.cpp:880)
    #04: mozilla::dom::FragmentOrElement::SizeOfExcludingThis(unsigned long (*)(void const*)) const (/home/njn/moz/mi5/dom/base/FragmentOrElement.cpp:2337)
    #05: nsINode::SizeOfIncludingThis(unsigned long (*)(void const*)) const (/home/njn/moz/mi5/go64dmd/parser/html/../../../dom/base/nsINode.h:307)
    #06: mozilla::dom::NodeInfo::NodeType() const (/home/njn/moz/mi5/go64dmd/dom/base/../../dist/include/mozilla/dom/NodeInfo.h:127)
    #07: nsHTMLDocument::DocAddSizeOfExcludingThis(nsWindowSizes*) const (/home/njn/moz/mi5/dom/html/nsHTMLDocument.cpp:3710)
    #08: nsIDocument::DocAddSizeOfIncludingThis(nsWindowSizes*) const (/home/njn/moz/mi5/dom/base/nsDocument.cpp:12820)
  }
}

The choice of whether to record an allocation stack trace for all blocks is controlled by an option (see below).

"Live" mode output


For "live" mode, dmd.py's output describes what live heap blocks are present. This output is broken into multiple sections.

  1. "Invocation". This tells you how DMD was invoked, i.e. what options were used.
  2. "Live stack trace records". This tells you which heap blocks were present.
  3. "Summary": gives measurements of the total heap.

The individual records are similar to those output in "dark matter" mode.

"Cumulative" mode output

For "cumulative" mode, dmd.py's output describes how the live heap blocks are covered by memory reports. This output is broken into multiple sections.

  1. "Invocation". This tells you how DMD was invoked, i.e. what options were used.
  2. "Cumulative stack trace records". This tells you which heap blocks were allocated during the session.
  3. "Summary": gives measurements of the total (cumulative) heap.

The individual records are similar to those output in "dark matter" mode.

"Scan" mode output

For "scan" mode, the output of dmd.py is the same as "live" mode. A separate script, block_analyzer.py, can be used to find out information about which blocks refer to a particular block. dmd.py --clamp-contents needs to be run on the log first.

Options

Runtime

When you run mach run --dmd you can specify additional options to control how DMD runs. Run mach help run for documentation on these.

The most interesting one is --mode. Acceptable values are dark-matter (the default), live, cumulative, and scan.

Another interesting one is --stacks. Acceptable values are partial (the default) and full. In the former case most blocks will not have an allocation stack trace recorded. However, because larger blocks are more likely to have one recorded, most allocated bytes should have an allocation stack trace even though most allocated blocks do not. Use --stacks=full if you want complete information, but note that DMD will run substantially slower in that case.

The MOZ_DMD_SHUTDOWN_LOG environment variable, if set, triggers a DMD run at shutdown; its value must be a directory where the logs will be placed. Which processes get logged is controlled by the MOZ_DMD_LOG_PROCESS environment variable, which can take the following values.

  • Unset: log all processes.
  • "default": log the parent process only.
  • "tab": log content processes only.

For example, if you have

MOZ_DMD_SHUTDOWN_LOG=~/dmdlogs/ MOZ_DMD_LOG_PROCESS=tab

then DMD logs for content processes will be saved to ~/dmdlogs/.

Post-processing

dmd.py also takes options that control how it works. Run dmd.py -h for documentation. The following options are the most interesting ones.

  • -f / --max-frames. By default, records show up to 8 stack frames. You can choose a smaller number, in which case more allocations will be aggregated into each record, but you'll have less context. Or you can choose a larger number, in which cases allocations will be split across more records, but you will have more context. There is no single best value, but values in the range 2..10 are often good. The maximum is 24.

  • -a / --ignore-alloc-frames. Many allocation stack traces start with multiple frames that mention allocation wrapper functions, e.g. js_calloc() calls replace_calloc(). This option filters these out. It often helps improve the quality of the output when using a small --max-frames value.

  • -s / --sort-by. This controls how records are sorted. Acceptable values are usable (the default), req, slop and num-blocks.

  • --clamp-contents. For a heap scan log, this performs a conservative pointer analysis on the contents of each block, changing any value that is a pointer into the middle of a live block into a pointer to the start of that block. All other values are changes to null. In addition, all trailing nulls are removed from the block contents.

As an example that combines multiple options, if you apply the following command to a profile obtained in "live" mode:

dmd.py -r -f 2 -a -s slop

it will give you a good idea of where the major sources of slop are.

dmd.py can also compute the difference between two DMD output files, so long as those files were produced in the same mode. Simply pass it two filenames instead of one to get the difference.

Which heap blocks are reported?

At this stage you might wonder how DMD knows, in "dark matter" mode, which allocations have been reported and which haven't. DMD only knows about heap blocks that are measured via a function created with one of the following two macros:

MOZ_DEFINE_MALLOC_SIZE_OF
MOZ_DEFINE_MALLOC_SIZE_OF_ON_ALLOC

Fortunately, most of the existing memory reporters do this. See Performance/Memory_Reporting for more details about how memory reporters are written.

Troubleshooting

Contact Nick Nethercote ("njn" on IRC), Eric Rahm ("erahm" on IRC) or Nathan Froyd ("froydnj" on IRC).

Document Tags and Contributors

 Contributors to this page: kmaglione, nnethercote, kanru, Rovanion, tnikkel, amccreight, glandium, staktrace
 Last updated by: kmaglione,