memleax debugs memory leak of a running process by attaching it,
without recompiling or restarting.
Because the debugging work depends on CPU architecture and OS heavily,
and I test
memleax only on several programs, and it is not used widely
by now. So there must be bugs.
Besides, I write a new tool
which works by hooking memory functions by LD_PRELOAD.
It's much simpler and has much less impact on performance.
So I am not going to improve
memleax debugs memory leak of a running process by attaching it.
It hooks the target process's invocation of memory allocation and free,
and reports the memory blocks which live long enough as memory leak, in real time.
The default expire threshold is 10 seconds, however you should always
set it by
-e option according to your scenarios.
It is very convenient to use, and suitable for production environment.
There is no need to recompile the program or restart the target process.
memleax to monitor the target process, wait for the real-time memory
leak report, and then kill it (e.g. by Ctrl-C) to stop monitoring.
memleax follows new threads, but not forked processes.
If you want to debug multiple processes, just run multiple
Because target process's each invocation of memory allocation and free makes a TRAP, the performance impact depends on the frequency of memory invocation in target process.
For example, it impacts lightly to
nginx with HTTP, while heavily with HTTPS,
Valgrind starts target process, while
memleax attaches to a running process;
Valgrind gives memory leak report after target process quits, while
reports in real time;
Valgrind reports all unfreed memory include program init, while
reports only after attaching, skipping the init phase;
Valgrind runs target process on its virtual CPU, which makes it slow.
memleax hooks memory APIs, which may be less slow if the target process
call memory APIs not often.
Valgrind debugs kinds of memory bugs, while
memleax is lightweight and
only detects memory leak.
In summary, I think
Valgrind is more powerful, while
memleax is more
convenient and suitable for production environment.
x86_64, tested on CentOS 7.2 and Ubuntu 16.04
aarch64, tested on Raspbian and pi64.
amd64, tested on FreeBSD 10.3
memleax can not show function backtrace on
you could try to compile the target program with
-funwind-tables flag of GCC.
There are DEB and RPM packages for releases.
For Arch Linux users,
memleax is available in AUR. Thanks to jelly.
For FreeBSD users,
memleax is available in FreeBSD Ports Collection.
Thanks to tabrarg.
I tried to submit
memleax to Fedora EPEL,
but failed. Any help is welcomed.
The development packages of the following libraries are required:
libdwis preferred. They are used to read dwarf debug-line information. If you do not have them neither, set
configureto disable it. As a result you will not see file name and line number in backtrace.
These packages may have different names in different distributions, such as
libelf may names
NOTE: On FreeBSD 10.3, there are built-in
libdwarf still can be installed by
memleax works with
built-in libelf and
pkg libdwarf. So you should
pkg, and must not install
After all required libraries are installed, run
$ mkdir build $ cd build $ cmake .. $ make $ sudo make install
To debug a running process, run:
$ memleax [options] <target-pid>
memleax begins to monitor the target process, and report memory leak in real time.
You should always set expire time by
-e options according to your scenarios.
For example, if you are debugging an HTTP server with keepalive, and there are
connections last for more than 5 minutes, you should set
-e 360 to cover it.
If your program is expected to free every memory in 1 second, you should set
to get report in time.
The memory blocks live longer than the threshold, are showed as:
CallStack: memory expires with 101 bytes, backtrace: 0x00007fd322bd8220 libc-2.17.so malloc()+0 0x000000000040084e test foo()+14 foo.c:12 0x0000000000400875 test bar()+37 bar.c:20 0x0000000000400acb test main()+364 test.c:80
CallStack is the ID of CallStack where memory leak happens.
The backtrace is showed only on the first time, while it only shows the ID and counter if expiring again:
CallStack: memory expires with 101 bytes, 2 times again
If the expired memory block is freed later, it shows:
CallStack: expired-memory frees after 10 seconds
If there are too many expired-memory-blocks are freed on one CallStack, this CallStack will not be showed again:
Warning: too many expired-free at CallStack. will not show this CallStack later
When you think you have found the answer, stop the debug.
memleax quits on:
After quitting, it also gives statistics for the CallStacks with memory leak:
CallStack: may-leak=20 (2020 bytes) expired=20 (2020 bytes), free_expired=0 (0 bytes) alloc=20 (2020 bytes), free=0 (0 bytes) freed memory live time: min=0 max=0 average=0 un-freed memory live time: max=20 0x00007fd322bd8220 libc-2.17.so malloc()+0 0x000000000040084e test foo()+14 foo.c:12 0x0000000000400875 test bar()+37 bar.c:20 0x0000000000400acb test main()+364 test.c:80