NAME

dpcpu — Kernel Dynamic Per-CPU Memory Allocator

SYNOPSIS

#include <sys/pcpu.h>

Per-CPU Variable Definition and Declaration

DPCPU_DEFINE(type, name);

DPCPU_DEFINE_STATIC(type, name);

DPCPU_DECLARE(type, name);

Current CPU Accessor Functions

DPCPU_PTR(name);

DPCPU_GET(name);

DPCPU_SET(name, value);

Named CPU Accessor Functions

DPCPU_ID_PTR(cpu, name);

DPCPU_ID_GET(cpu, name);

DPCPU_ID_SET(cpu, name, value);

DESCRIPTION

dpcpu instantiates one instance of a global variable with each CPU in the system. Dynamically allocated per-CPU variables are defined using DPCPU_DEFINE(), which defines a variable of name name and type type. Arbitrary C types may be used, including structures and arrays. If no initialization is provided, then each per-CPU instance of the variable will be zero-filled (i.e., as though allocated in BSS):

DPCPU_DEFINE(int, foo_int);

Values may also be initialized statically with the definition, causing each per-CPU instance to be initialized with the value:

DPCPU_DEFINE(int, foo_int) = 1;

Values that can be defined as static must use DPCPU_DEFINE_STATIC():

DPCPU_DEFINE_STATIC(int, foo_int);

DPCPU_DECLARE() produces a declaration of the per-CPU variable suitable for use in header files.

The current CPU's variable instance can be accessed via DPCPU_PTR (which returns a pointer to the per-CPU instance), DPCPU_GET (which retrieves the value of the per-CPU instance), and DPCPU_SET (which sets the value of the per-CPU instance).

Instances of variables associated with specific CPUs can be accessed via the DPCPU_ID_PTR, DPCPU_ID_GET, and DPGPU_ID_SET accessor functions, which accept an additional CPU ID argument, cpu.

Synchronization

In addition to the ordinary synchronization concerns associated with global variables, which may imply the use of atomic(9), mutex(9), or other kernel synchronization primitives, it is further the case that thread migration could dynamically change the instance of a variable being accessed by a thread between operations. This requires additional care when reasoning about and protecting per-CPU variables.

For example, it may be desirable to protect access using critical_section(9) to prevent both preemption and migration during use. Alternatively, it may be desirable to cache the CPU ID at the start of a sequence of accesses, using suitable synchronization to make non-atomic sequences safe in the presence of migration.

DPCPU_DEFINE_STATIC(int, foo_int);
DPCPU_DEFINE_STATIC(struct mutex, foo_lock);

void
foo_int_increment(void)
{
    int cpu, value;

    /* Safe as atomic access. */
    atomic_add_int(DPCPU_PTR(foo_int), 1);

    /*
     * Protect with a critical section, which prevents preemption
     * and migration.  However, access to instances from remote CPUs
     * is not safe, as critical sections prevent concurrent access
     * only from the current CPU.
     */
    critical_enter();
    value = DPCPU_GET(foo_int);
    value++;
    DPCPU_SET(foo_int, value);
    critical_exit();

    /*
     * Protect with a per-CPU mutex, tolerating migration, but
     * potentially accessing the variable from multiple CPUs if
     * migration occurs after reading curcpu.  Remote access to a
     * per-CPU variable is safe as long as the correct mutex is
     * acquired.
     */
    cpu = curcpu;
    mtx_lock(DPCPU_ID_PTR(cpu, foo_lock));
    value = DPCPU_ID_GET(cpu, foo_int);
    value++;
    DPCPU_ID_SET(cpu, foo_int);
    mtx_unlock(DPCPU_ID_PTR(cpu, foo_lock));
}

HISTORY

dpcpu was first introduced by Jeff Roberson in FreeBSD 8.0. This manual page was written by Robert N. M. Watson.