UMEM_ALLOC(3MALLOC) Memory Allocation Library Functions UMEM_ALLOC(3MALLOC)

umem_alloc, umem_zalloc, umem_free, umem_nofail_callback - fast, scalable memory allocation

cc [ flag ... ] file... -lumem [ library ... ]
#include <umem.h>
void *umem_alloc(size_t size, int  flags);

void *umem_zalloc(size_t size, int  flags);

void umem_free(void *buf, size_t size);

void umem_nofail_callback((int (*callback)(void));

void *malloc(size_t size);

void *calloc(size_t nelem, size_t elsize);

void free(void *ptr);

void *memalign(size_t alignment, size_t size);

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

void *valloc(size_t size);

The umem_alloc() function returns a pointer to a block of size bytes suitably aligned for any variable type. The initial contents of memory allocated using umem_alloc() is undefined. The flags argument determines the behavior of umem_alloc() if it is unable to fulfill the request. The flags argument can take the following values:

UMEM_DEFAULT

Return NULL on failure.

UMEM_NOFAIL

Call an optional callback (set with umem_nofail_callback()) on failure. The callback takes no arguments and can finish by:
returning UMEM_CALLBACK_RETRY, in which case the allocation will be retried. If the allocation fails, the callback will be invoked again.
returning UMEM_CALLBACK_EXIT(status), in which case exit(2) is invoked with status as its argument. The exit() function is called only once. If multiple threads return from the UMEM_NOFAIL callback with UMEM_CALLBACK_EXIT(status), one will call exit() while the other blocks until exit() terminates the program.
invoking a context-changing function (setcontext(2)) or a non-local jump (longjmp(3C) or siglongjmp(3C)), or ending the current thread of control (thr_exit(3C) or pthread_exit(3C)). The application is responsible for any necessary cleanup. The state of libumem remains consistent.
If no callback has been set or the callback has been set to NULL, umem_alloc(..., UMEM_NOFAIL) behaves as though the callback returned UMEM_CALLBACK_EXIT(255).

The libumem library can call callbacks from any place that a UMEM_NOFAIL allocation is issued. In multithreaded applications, callbacks are expected to perform their own concurrency management.

The function call umem_alloc(0, flag) always returns NULL. The function call umem_free(NULL, 0) is allowed.

The umem_zalloc() function has the same semantics as umem_alloc(), but the block of memory is initialized to zeros before it is returned.

The umem_free() function frees blocks previously allocated using umem_alloc() and umem_zalloc(). The buffer address and size must exactly match the original allocation. Memory must not be returned piecemeal.

The umem_nofail_callback() function sets the process-wide UMEM_NOFAIL callback. See the description of UMEM_NOFAIL for more information.

The malloc(), calloc(), free(), memalign(), realloc(), and valloc() functions are as described in malloc(3C). The libumem library provides these functions for backwards-compatibility with the standard functions.

See umem_debug(3MALLOC) for environment variables that effect the debugging features of the libumem library.

UMEM_OPTIONS

Contains a list of comma-separated options. Unrecognized options are ignored. The options that are supported are:

backend=sbrk
backend=mmap

Set the underlying function used to allocate memory. This option can be set to sbrk (the default) for an sbrk(2)-based source or mmap for an mmap(2)-based source. If set to a value that is not supported, sbrk will be used.

perthread_cache=size

libumem allows for each thread to cache recently freed small allocations for future allocations. The size argument, which accepts k, m, g, and t, suffixes denotes the maximum amount of memory each thread can use for this purpose. The default amount used is 1 MB. Any buffers in the per-thread cache are freed when the thread exits. The efficacy of the per-thread cache can be determined with the ::umastat mdb(1) dcmd debugger command.

allocator=best
allocator=first
allocator=instant
allocator=next

Set the underlying allocation strategy. The best fit strategy tells libumem to use the smallest free segment possible. The instant fit strategy approximates the best fit strategy in constant cpu time. The first fit strategy takes the first free segment that can honor the allocation. The next fit strategy uses the next free segment after the previously allocated one.

Example 1 Using the umem_alloc() function.


#include <stdio.h>
#include <umem.h>
...
char *buf = umem_alloc(1024, UMEM_DEFAULT);
if (buf == NULL) {

fprintf(stderr, "out of memory\n");
return (1); } /* cannot assume anything about buf's contents */ ... umem_free(buf, 1024); ...

Example 2 Using the umem_zalloc() function


#include <stdio.h>
#include <umem.h>
...
char *buf = umem_zalloc(1024, UMEM_DEFAULT);
if (buf == NULL) {

fprintf(stderr, "out of memory\n");
return (1); } /* buf contains zeros */ ... umem_free(buf, 1024); ...

Example 3 Using UMEM_NOFAIL


#include <stdlib.h>
#include <stdio.h>
#include <umem.h>
/*

* Note that the allocation code below does not have to
* check for umem_alloc() returning NULL
*/ int my_failure_handler(void) {
(void) fprintf(stderr, "out of memory\n");
return (UMEM_CALLBACK_EXIT(255)); } ... umem_nofail_callback(my_failure_handler); ... int i; char *buf[100]; for (i = 0; i < 100; i++)
buf[i] = umem_alloc(1024 * 1024, UMEM_NOFAIL); ... for (i = 0; i < 100; i++)
umem_free(buf[i], 1024 * 1024); ...

Example 4 Using UMEM_NOFAIL in a multithreaded application


#define _REENTRANT
#include <thread.h>
#include <stdio.h>
#include <umem.h>
void *
start_func(void *the_arg)
{

int *info = (int *)the_arg;
char *buf = umem_alloc(1024 * 1024, UMEM_NOFAIL);
/* does not need to check for buf == NULL */
buf[0] = 0;
...
/*
* if there were other UMEM_NOFAIL allocations,
* we would need to arrange for buf to be
* umem_free()ed upon failure.
*/
...
umem_free(buf, 1024 * 1024);
return (the_arg); } ... int my_failure_handler(void) {
/* terminate the current thread with status NULL */
thr_exit(NULL); } ... umem_nofail_callback(my_failure_handler); ... int my_arg; thread_t tid; void *status; (void) thr_create(NULL, NULL, start_func, &my_arg, 0,
NULL); ... while (thr_join(0, &tid, &status) != 0)
; if (status == NULL) {
(void) fprintf(stderr, "thread %d ran out of memory\n",
tid); } ...

See attributes(7) for descriptions of the following attributes:

ATTRIBUTE TYPE ATTRIBUTE VALUE
Interface Stability Committed
MT-Level MT-Safe
Standard See below.

For malloc(), calloc(), free(), realloc(), and valloc(), see standards(7).

exit(2), mmap(2), sbrk(2), longjmp(3C), malloc(3C), pthread_exit(3C), thr_exit(3C), libumem(3LIB), bsdmalloc(3MALLOC), malloc(3MALLOC), mapmalloc(3MALLOC), umem_cache_create(3MALLOC), umem_debug(3MALLOC), watchmalloc(3MALLOC), attributes(7), standards(7)

Modular Debugger Guide:

https://illumos.org/books/mdb/

Any of the following can cause undefined results:

Passing a pointer returned from umem_alloc() or umem_zalloc() to free() or realloc().
Passing a pointer returned from malloc(), calloc(), valloc(), memalign(), or realloc() to umem_free().
Writing past the end of a buffer allocated using umem_alloc() or umem_zalloc()
Performing UMEM_NOFAIL allocations from an atexit(3C) handler.

If the UMEM_NOFAIL callback performs UMEM_NOFAIL allocations, infinite recursion can occur.

The following list compares the features of the malloc(3C), bsdmalloc(3MALLOC), malloc(3MALLOC), mtmalloc(3MALLOC) , and the libumem functions.

The malloc(3C), bsdmalloc(3MALLOC), and malloc(3MALLOC) functions have no support for concurrency. The libumem and mtmalloc(3MALLOC) functions support concurrent allocations.
The bsdmalloc(3MALLOC) functions afford better performance but are space-inefficient.
The malloc(3MALLOC) functions are space-efficient but have slower performance.
The standard, fully SCD-compliant malloc(3C) functions are a trade-off between performance and space-efficiency.
The mtmalloc(3MALLOC) functions provide fast, concurrent malloc() implementations that are not space-efficient.
The libumem functions provide a fast, concurrent allocation implementation that in most cases is more space-efficient than mtmalloc(3MALLOC).
December 9, 2017 OmniOS