watchmalloc

- debugging memory allocator

Synopsis

#include <stdlib.h>

void *malloc(size_t size);

void free(void *ptr);

void *realloc(void *ptr, size_t size);

void *memalign(size_t alignment, size_t size);

void *valloc(size_t size);

void *calloc(size_t nelem, size_t elsize);

#include <malloc.h>

int mallopt(int cmd, int value);

struct mallinfo mallinfo(void);

Description

The collection of malloc() functions in this shared object are an optional replacement for the standard versions of the same functions in the system C library. See malloc(3C). They provide a more strict interface than the standard versions and enable enforcement of the interface through the watchpoint facility of /proc. See proc(4).

Any dynamically linked application can be run with these functions in place of the standard functions if the following string is present in the environment (see ld.so.1(1)):

 LD_PRELOAD=watchmalloc.so.1

The individual function interfaces are identical to the standard ones as described in malloc(3C). However, laxities provided in the standard versions are not permitted when the watchpoint facility is enabled (see WATCHPOINTS below):

• Memory may not be freed more than once.

• A pointer to freed memory may not be used in a call to realloc().

• A call to malloc() immediately following a call to free() will not return the same space.

• Any reference to memory that has been freed yields undefined results.

To enforce these restrictions partially, without great loss in speed as compared to the watchpoint facility described below, a freed block of memory is overwritten with the pattern 0xdeadbeef before returning from free(). The malloc() function returns with the allocated memory filled with the pattern 0xbaddcafe as a precaution against applications incorrectly expecting to receive back unmodified memory from the last free(). The calloc() function always returns with the memory zero-filled.

Entry points for mallopt() and mallinfo() are provided as empty routines, and are present only because some malloc() implementations provide them.

WATCHPOINTS

The watchpoint facility of /proc can be applied by a process to itself. The functions in watchmalloc.so.1 use this feature if the following string is present in the environment:

MALLOC_DEBUG=WATCH

This causes every block of freed memory to be covered with WA_WRITE watched areas. If the application attempts to write any part of freed memory, it will trigger a watchpoint trap, resulting in a SIGTRAP signal, which normally produces an application core dump.

A header is maintained before each block of allocated memory. Each header is covered with a watched area, thereby providing a red zone before and after each block of allocated memory (the header for the subsequent memory block serves as the trailing red zone for its preceding memory block). Writing just before or just after a memory block returned by malloc() will trigger a watchpoint trap.

Watchpoints incur a large performance penalty. Requesting MALLOC_DEBUG=WATCH can cause the application to run 10 to 100 times slower, depending on the use made of allocated memory.

Further options are enabled by specifying a comma-separated string of options:

MALLOC_DEBUG=WATCH,RW,STOP

WATCH

Enables WA_WRITE watched areas as described above.

RW

Enables both WA_READ and WA_WRITE watched areas. An attempt either to read or write freed memory or the red zones will trigger a watchpoint trap. This incurs even more overhead and can cause the application to run up to 1000 times slower.

STOP

The process will stop showing a FLTWATCH machine fault if it triggers a watchpoint trap, rather than dumping core with a SIGTRAP signal. This allows a debugger to be attached to the live process at the point where it underwent the watchpoint trap. Also, the various /proc tools described in proc(1) can be used to examine the stopped process.

One of WATCH or RW must be specified, else the watchpoint facility is not engaged. RW overrides WATCH. Unrecognized options are silently ignored.

LIMITATIONS

Sizes of memory blocks allocated by malloc() are rounded up to the worst-case alignment size, 8 bytes for 32-bit processes and 16 bytes for 64-bit processes. Accessing the extra space allocated for a memory block is technically a memory violation but is in fact innocuous. Such accesses are not detected by the watchpoint facility of watchmalloc.

Interposition of watchmalloc.so.1 fails innocuously if the target application is statically linked with respect to its malloc() functions.

Attributes

See attributes(5) for descriptions of the following attributes:

See Also

proc(1), bsdmalloc(3MALLOC), calloc(3C), free(3C), malloc(3C), malloc(3MALLOC), mapmalloc(3MALLOC), memalign(3C), realloc(3C), valloc(3C), libmapmalloc(3LIB), proc(4), attributes(5)