leak, kmem, umem – help find memory leaks

leak [ –abcds ] [ –f binary ] [ –r res ] [ –x width ] pid ...

kmem [ kernel ]

umem pid [ textfile ]

Leak examines the named processes, which should be sharing their data and bss segments, for memory leaks. It uses a mark and sweep–style algorithm to determine which allocated blocks are no longer reachable from the set of root pointers. The set of root pointers is created by looking through the shared bss segment as well as each process's registers.

Unless directed otherwise, leak prints, for each block, a line with seven space–separated fields: the string block, the address of the block, the size of the block, the first two words of the block, and the function names represented by the first two words of the block. Usually, the first two words of the block contain the malloc and realloc tags (see malloc(2)), useful for finding who allocated the leaked blocks.

If the –s or the –c option is given, leak will instead present a sequence of acid(1) commands that show each leaky allocation site. With –s a comment appears next to each command to indicate how many lost blocks were allocated at that point in the program. With –c the comments are extended to indicate also the total number of bytes lost at that point in the program, and an additional comment line gives the overall total number of bytes.

If the –a option is given, leak will print information as decribed above, but for all allocated blocks, not only leaked ones. If the –d option is given, leak will print information as described above, but for all free blocks, i.e. those freed, or those that are not yet in use (fragmentation?). The –a and –d options can be combined.

If the –b option is given, leak will print a Plan 9 image file graphically summarizing the memory arenas. In the image, each pixel represents res (default 8) bytes. The color code is:
dark blue   Completely allocated.
bright blueContains malloc headers.
bright red   Contains malloc headers for leaked memory.
dark red    Contains leaked memory.
yellow      Completely free
white      Padding to fill out the image. The bright pixels representing headers help in counting the number of blocks. Magnifying the images with lens(1) is often useful.

If given a name rather than a list of process ids, leak echoes back a command–line with process ids of every process with that name.

The –f option specifies a binary to go on the acid(1) command–line used to inspect the processes, and is only necessary when inspecting processes started from stripped binaries.

Umem prints a summary of all allocated blocks in the process with id pid. Each line of the summary gives the count and total size of blocks allocated at an allocation point. The list is sorted by count in decreasing order. Umem prints summarizes all allocations, not just memory leaks, but it is faster and requires less memory than leak.

Kmem is like umem but prints a summary for the running kernel.

List lost blocks in 8.out. This depends on the fact that there is only one instance of 8.out running; if there were more, the output of leak –s 8.out would need editing before sending to the shell.
% leak –s 8.out
leak –s 229 230
% leak –s 8.out | rc
src(0x0000bf1b); // 64
src(0x000016f5); // 7
src(0x0000a988); // 7

View the memory usage graphic for the window system.
% leak –b rio | rc | page

List the top allocation points in the kernel, first by count and then by total size:
% kmem | sed 10q
% kmem | sort –nr +1 | sed 10q


getcallerpc(2), setmalloctag in malloc(2)

Leak and kmem depend on the internal structure of the libc pool memory allocator (see pool(2)). Since the ANSI/POSIX environment uses a different allocator, leak will not work on APE programs.

Leak is not speedy, and acidleak can consume more memory than the process(es) being examined.

These commands require /sys/src/libc/port/pool.acid to be present and generated from pool.c.