Updated: 2022/Sep/29
Please read Privacy Policy. It's for your privacy.
XCALL(9) Kernel Developer's Manual XCALL(9) NAME xcall, xc_broadcast, xc_unicast, xc_wait, xc_barrier - cross-call interface SYNOPSIS #include <sys/xcall.h> typedef void (*xcfunc_t)(void *, void *); uint64_t xc_broadcast(u_int flags, xcfunc_t func, void *arg1, void *arg2); uint64_t xc_unicast(u_int flags, xcfunc_t func, void *arg1, void *arg2, struct cpu_info *ci); void xc_wait(uint64_t where); void xc_barrier(u_int flags); DESCRIPTION The machine-independent xcall interface allows any CPU in the system to request that an arbitrary function be executed on any other CPU. Sometimes it is necessary to modify hardware state that is tied directly to individual CPUs (such as a CPU's local timer), and these updates can not be done remotely by another CPU. The LWP requesting the update may be unable to guarantee that it will be running on the CPU where the update must occur, when the update occurs. Additionally, it is sometimes necessary to modify per-CPU software state from a remote CPU. Where these update operations are so rare or the access to the per-CPU data so frequent that the cost of using locking or atomic operations to provide coherency is prohibitive, another way must be found. Cross calls help to solve these types of problem. However, since this facility is heavyweight, it is expected that it will not be used often. xcall provides a mechanism for making "low priority" cross calls. The function to be executed runs on the remote CPU within a thread context, and not from a software interrupt, so it can ensure that it is not interrupting other code running on the CPU, and so has exclusive access to the CPU. Keep in mind that unless disabled, it may cause a kernel preemption. xcall also provides a mechanism for making "high priority" cross calls. The function to be executed runs on the remote CPU within a software interrupt context, possibly interrupting other lower-priority code running on the CPU. NOTES Functions being called should be relatively lightweight. They may block on locks, but carefully and minimally, to not interfere with other cross calls in the system. FUNCTIONS xc_broadcast(flags, func, arg1, arg2) Call (*func)(arg1, arg2) on all CPUs in the system. Return a uint64_t "ticket" to xc_wait() on for the cross-call to complete. flags should be XC_HIGHPRI or XC_HIGHPRI_IPL(ipl) for a "high priority" call, and 0 for a "low priority" call. XC_HIGHPRI uses an IPL_SOFTSERIAL software interrupt while XC_HIGHPRI_IPL uses a software interrupt with an IPL specified by ipl. xc_broadcast() should not be called from interrupt context. xc_unicast(flags, func, arg1, arg2, ci) Like xc_broadcast(), but call func on only the CPU indicated by ci. xc_unicast() also returns a "ticket". xc_wait(where) Wait on the "ticket" returned by a prior xc_broadcast() or xc_unicast() for the corresponding cross-call to complete. xc_wait() should be called from a thread context. xc_barrier(flags) Issue a broadcast cross-call that does nothing, and wait for it to complete. This functions like a memory barrier that forces all prior operations in program order to globally happen before all subsequent operations in program order, as witnessed by every CPU. This additionally waits for all higher-priority activity on the CPU to complete, according to flags: - xc_barrier(0) waits for any pending kpreempt_disable(9) sections or activity at interrupt priority level above IPL_NONE to finish on all CPUs. - xc_barrier(XC_HIGHPRI_IPL(ipl)) waits for any pending activity at the software interrupt priority level ipl or higher to finish on all CPUs. xc_barrier() is much more expensive than membar_ops(3), so it should be used sparingly, only to publish information infrequently -- for example, during module load and unload -- when the cost of a memory barrier on the consumer side would be prohibitive. CODE REFERENCES The xcall interface is implemented within the file sys/kern/subr_xcall.c. SEE ALSO membar_ops(3), kpreempt(9), percpu(9), softint(9) HISTORY The xcall interface first appeared in NetBSD 5.0. AUTHORS Andrew Doran <ad@NetBSD.org> NetBSD 10.99 February 1, 2020 NetBSD 10.99