heaphook

Replace all the dynamic heap allocation functions by LD_PRELOAD.

Build and Install

This library heaphook is prepared as a ament_cmake package.

$ mkdir -p /path/to/heaphook_ws && cd /path/to/heaphook_ws
$ mkdir src & git clone git@github.com:tier4/heaphook.git src
$ colcon build

The shared libraries that will be specified in LD_PRELOAD are generated under install/heaphook/lib. This path is added to LD_LIBRARY_PATH by the setup script.

$ source install/setup.bash

How to use

For now, We provide two kinds of the heaphook libraries.

• libpreloaded_heaptrack.so: Records all the heap allocation/deallocation function calls and generate a log file for visualizing the history of heap consumption.
• libpreloaded_tlsf.so: Replaces all the heap allocation/deallocation with TLSF (Tow-Level Segregated Fit) memory allocator.
• libpreloaded_backtrace.so: Records all malloc/new function calls with their backtraces where the memory allocations take place.

A typical use case is to utilize libpreloaded_heaptrack to grasp the transition and maximum value of heap consumtion of the target process and to determine the initial allocated memory pool size for libpreloaded_tlsf. Of cource, it is also useful to utilize libpreloaded_heaptrack just to know the heap consumption of the target process.

libpreloaded_heaptrack

You can track the heap consumption of the specified process.

$ LD_PRELOAD=libpreloaded_heaptrack.so executable

A log file is generated under the current working directory in the format heaplog.{%pid}.log. You can visualize heap consumption transitions in PDF format based on the generated log file. The parser depends on progressbar python library, so install it before.

$ pip install progressbar
$ python3 heaplog_parser.py heaplog.{%pid}.log // Generates heaplog.{%pid}.pdf

libpreloaded_tlsf

You need to specify the initial allocaton size for the memory pool and the size of additional memory to be allocated when the initial memory pool size is not sufficient. The initial size of the memory pool should be set to a value with a margin greater than the peak heap usage of the target process. Extending the memory pool size during the runtime should be avoided, as it leads to overhead in the allocation functions.

$ LD_PRELOAD=libpreloaded_tlsf.so INITIAL_MEMPOOL_SIZE=1000000 ADDITIONAL_MEMPOOL_SIZE=1000000 executable

When the initial size of the memory pool is insufficient, it first allocates the additional size specified by ADDITIONAL_MEMPOOL_SIZE, and if it is still insufficient, it allocates double the value of the previous allocation until it is sufficient.

As an example, here is what happens when malloc(1000) is called when the existing memory pool is exhausted and ADDITIONAL_MEMPOOL_SIZE=100.

1. 100 bytes added to the memory pool (still not sufficient)
2. 200 bytes added to the memory pool (still not sufficient)
3. 400 bytes added to the memory pool (still not sufficient)
4. 800 bytes added to the memory pool (still not sufficient)
5. 1600 bytes added to the memory pool (finally sufficient)

The added memory pool areas are not contiguous with each other in the virual address space, so it is not necessarily enough even if the total size of the added memory pools exceeds the size of the memory allocation request.

libpreloaded_backtrace.so

Use the following command to trace the callers:

$ LD_PRELOAD=libpreloaded_backtrace.so executable

Two log files are genearted under the working directory in the format of top_alloc_bytes_bt.{%pid}.{%tid}.log and top_num_calls_bt.{%pid}.{%tid}.log. By default, top 10 callers are logged. The environment variable NUM_TOPS can control the number of callers. In addition, callers that just make one malloc/new are not logged as they are not the major source of page faults. To disable it, just set environment variable SHOW_NON_RECURRENT_CALLERS=1.

To show the source code file name and the line numbers, run the following command:

$ python3 backtrace_analyzer.py -i top_alloc_bytes_bt.{%pid}.{%tid}.log
or
$ python3 backtrace_analyzer.py -i top_alloc_bytes_bt.{%pid}.{%tid}.log | c++filt  // demangle C++ function names

The result will be more user-friendly if the executables are linked with -rdynamic -no-pie -fno-pie options. Or even aggressive, build your code with CMAKE_BUILD_TYPE=RelWithDebInfo.

Integrate with ROS2 launch

You can easily integrate heaphook with ROS2 launch systems. From the launch file, you can replace all heap allocations of the process corresponding to the targeted Node and ComposableNodeContainer.

<node pkg="..." exec="..." name="...">
  <env name="LD_PRELOAD" value="libpreloaded_heaptrack.so" />
</node>

<node pkg="..." exec="..." name="...">
  <env name="LD_PRELOAD" value="libpreloaded_tlsf.so" />
  <env name="INITIAL_MEMPOOL_SIZE" value="100000000" />
  <env name="ADDITIONAL_MEMPOOL_SIZE" value="100000000" />
</node>

```python container = ComposableNodeContainer( …, additional_env={“LD_PRELOAD”: “libpreloaded_heaptrack.so”},

File truncated at 100 lines see the full file

Changelog for package heaphook

0.1.1 (2023-08-12)

• fix: add workaround to build on ubuntu 20.04 (#5)
• Contributors: Daisuke Nishimatsu

0.1.0 (2023-05-23)

• ci: add build-and-test workflow (#2)

  • add build-and-test workflow

  * Update .github/workflows/build-and-test.yaml Co-authored-by: Kenji Miyake <<31987104+kenji-miyake@users.noreply.github.com>>

  • Update build-and-test.yaml
  • fix linter error
  • fix
  • fix

  * fix

  Co-authored-by: Kenji Miyake <<31987104+kenji-miyake@users.noreply.github.com>>

• style: apply lint (#1)

  • apply lint

  * refactor package.xml ---------

• Update CMakeLists.txt

• Update package.xml

• Update package.xml metadata

• Create LICENSE

• Update README.md

• Create README.md

• Prevent infinit loop when additional mempool is not enough

• Enable to configure mempool size

• Delete unnecessary code

• Hanble exhaustion of memory pool

• Aquire lock for tlsf library

• Fix bug

• Add hooks for specialized malloc functions

• Deal with alignment related functions

• Fix

• Introduce tlsf allocator

• Create ament_cmake package

• Add malloc_usable_size hook

• Add aligned allocation hooks

• Add realloc hook

• Add calloc hook

• Safe figure as pdf in heaplog parser

• Add progress bar to heaplog parser

• Add python heaplog parser

• Change log format

• Complete heaplog parser

• Add mmaped log parser base

• Fix bug (prevent logging in logging thread)

• Make logging incremental and async

• Log malloc/free history

• Hook free

• First malloc hook

• Contributors: Daisuke Nishimatsu, Takahiro Ishikawa