man.bsd.lv manual page server

Manual Page Search Parameters
KMEM(9) Kernel Developer's Manual KMEM(9)

kmem — kernel wired memory allocator

#include <sys/kmem.h>

void *
kmem_alloc(size_t size, km_flag_t kmflags);

void *
kmem_zalloc(size_t size, km_flag_t kmflags);

void
kmem_free(void *p, size_t size);

void *
kmem_intr_alloc(size_t size, km_flag_t kmflags);

void *
kmem_intr_zalloc(size_t size, km_flag_t kmflags);

void
kmem_intr_free(void *p, size_t size);

char *
kmem_asprintf(const char *fmt, ...);

char *
kmem_strdupsize(const char *str, size_t *size, km_flag_t kmflags);

void
kmem_strfree(char *str);


options KMEM_SIZE
options KMEM_REDZONE
options KMEM_GUARD

() allocates kernel wired memory. It takes the following arguments.

size
Specify the size of allocation in bytes.
kmflags
Either of the following:
If the allocation cannot be satisfied immediately, sleep until enough memory is available. If KM_SLEEP is specified, then the allocation cannot fail.
Don't sleep. Immediately return NULL if there is not enough memory available. It should only be used when failure to allocate will not have harmful, user-visible effects.

Use of KM_NOSLEEP is strongly discouraged as it can create transient, hard to debug failures that occur when the system is under memory pressure.

In situations where it is not possible to sleep, for example because locks are held by the caller, the code path should be restructured to allow the allocation to be made in another place.

The contents of allocated memory are uninitialized.

Unlike Solaris, kmem_alloc(0, flags) is illegal.

() is the equivalent of kmem_alloc(), except that it initializes the memory to zero.

() functions as the well known () function, but allocates memory using kmem_alloc(). This routine can sleep during allocation. The size of the allocated area is the length of the returned character string, plus one (for the NUL terminator). This must be taken into consideration when freeing the returned area with kmem_free().

() frees kernel wired memory allocated by kmem_alloc() or kmem_zalloc() so that it can be used for other purposes. It takes the following arguments.

p
The pointer to the memory being freed. It must be the one returned by kmem_alloc() or kmem_zalloc().
size
The size of the memory being freed, in bytes. It must be the same as the size argument used for kmem_alloc() or kmem_zalloc() when the memory was allocated.

Freeing NULL is illegal.

(), () and () are the equivalents of the above kmem routines which can be called from the interrupt context. These routines are for the special cases. Normally, pool_cache(9) should be used for memory allocation from interrupt context.

The () function is a utility function that can be used to copy the string in the str argument to a new buffer allocated using kmem_alloc() and optionally return the size of the allocation (the length of the string plus the trailing NUL) in the size argument if that is not NULL.

The () function can be used to free a NUL terminated string computing the length of the string using strlen(3) and adding one for the NUL and then using kmem_free().

Making KM_SLEEP allocations while holding mutexes or reader/writer locks is discouraged, as the caller can sleep for an unbounded amount of time in order to satisfy the allocation. This can in turn block other threads that wish to acquire locks held by the caller. It should be noted that kmem_free() may also block.

For some locks this is permissible or even unavoidable. For others, particularly locks that may be taken from soft interrupt context, it is a serious problem. As a general rule it is better not to allow this type of situation to develop. One way to circumvent the problem is to make allocations speculative and part of a retryable sequence. For example:

  retry:
        /* speculative unlocked check */
        if (need to allocate) {
                new_item = kmem_alloc(sizeof(*new_item), KM_SLEEP);
        } else {
                new_item = NULL;
        }
        mutex_enter(lock);
        /* check while holding lock for true status */
        if (need to allocate) {
                if (new_item == NULL) {
                        mutex_exit(lock);
                        goto retry;
                }
                consume(new_item);
                new_item = NULL;
        }
        mutex_exit(lock);
        if (new_item != NULL) {
                /* did not use it after all */
                kmem_free(new_item, sizeof(*new_item));
        }

Kernels compiled with the KMEM_SIZE option ensure the size given in () matches the actual allocated size. On kmem_alloc(), the kernel will allocate an additional contiguous kmem page of eight bytes in the buffer, will register the allocated size in the first kmem page of that buffer, and will return a pointer to the second kmem page in that same buffer. When freeing, the kernel reads the first page, and compares the size registered with the one given in kmem_free(). Any mismatch triggers a panic.

KMEM_SIZE is enabled by default on DIAGNOSTIC and DEBUG.

Kernels compiled with the KMEM_REDZONE option add a dynamic pattern of two bytes at the end of each allocated buffer, and check this pattern when freeing to ensure the caller hasn't written outside the requested area. This option does not introduce a significant performance impact, but has two drawbacks: it only catches write overflows, and catches them only on ().

KMEM_REDZONE is enabled by default on DIAGNOSTIC.

Kernels compiled with the KMEM_GUARD option perform CPU intensive sanity checks on kmem operations. It adds additional, very high overhead runtime verification to kmem operations. It must be enabled with KMEM_SIZE.

KMEM_GUARD tries to catch the following types of bugs:

  • Overflow at time of occurrence, by means of a guard page. An unmapped guard page sits immediately after the requested area; a read/write overflow therefore triggers a page fault.
  • Underflow at (), by using KMEM_SIZE's registered size. If an underflow occurs, the size stored by KMEM_SIZE will be overwritten, which means that when freeing, the kernel will spot the mismatch.
  • Use-after-free at time of occurrence. When freeing, the memory is unmapped, and depending on the value of kmem_guard_depth, the kernel will more or less delay the recycling of that memory. Which means that any ulterior read/write access to the memory will trigger a page fault, given it hasn't been recycled yet.

To enable it, boot the system with the -d option, which causes the debugger to be entered early during the kernel boot process. Issue commands such as the following:

db> w kmem_guard_depth 0t30000
db> c

This instructs kmem_guard to queue up to 60000 (30000*2) pages of unmapped KVA to catch use-after-free type errors. When () is called, memory backing a freed item is unmapped and the kernel VA space pushed onto a FIFO. The VA space will not be reused until another 30k items have been freed. Until reused the kernel will catch invalid accesses and panic with a page fault. Limitations:

  • It has a severe impact on performance.
  • It is best used on a 64-bit machine with lots of RAM.

KMEM_GUARD is enabled by default on DEBUG.

On success, kmem_alloc(), kmem_asprintf(), kmem_intr_alloc(), kmem_intr_zalloc(), kmem_strdupsize(), and kmem_zalloc() return a pointer to allocated memory. Otherwise, NULL is returned.

The kmem subsystem is implemented within the file sys/kern/subr_kmem.c.

intro(9), memoryallocators(9), percpu(9), pool_cache(9), uvm_km(9)

The kmem_alloc(), kmem_asprintf(), kmem_free(), kmem_strdupsize(), kmem_strfree(), and kmem_zalloc() functions cannot be used from interrupt context, from a soft interrupt, or from a callout. Use pool_cache(9) in these situations.

As the memory allocated by kmem_alloc() is uninitialized, it can contain security-sensitive data left by its previous user. It is the caller's responsibility not to expose it to the world.

November 7, 2017 NetBSD-9.2