xref: /netbsd-src/sys/kern/subr_ipi.c (revision 76c7fc5f6b13ed0b1508e6b313e88e59977ed78e)
1 /*	$NetBSD: subr_ipi.c,v 1.7 2019/10/16 18:29:49 christos Exp $	*/
2 
3 /*-
4  * Copyright (c) 2014 The NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Mindaugas Rasiukevicius.
9  *
10  * Redistribution and use in source and binary forms, with or without
11  * modification, are permitted provided that the following conditions
12  * are met:
13  * 1. Redistributions of source code must retain the above copyright
14  *    notice, this list of conditions and the following disclaimer.
15  * 2. Redistributions in binary form must reproduce the above copyright
16  *    notice, this list of conditions and the following disclaimer in the
17  *    documentation and/or other materials provided with the distribution.
18  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 /*
33  * Inter-processor interrupt (IPI) interface: asynchronous IPIs to
34  * invoke functions with a constant argument and synchronous IPIs
35  * with the cross-call support.
36  */
37 
38 #include <sys/cdefs.h>
39 __KERNEL_RCSID(0, "$NetBSD: subr_ipi.c,v 1.7 2019/10/16 18:29:49 christos Exp $");
40 
41 #include <sys/param.h>
42 #include <sys/types.h>
43 
44 #include <sys/atomic.h>
45 #include <sys/evcnt.h>
46 #include <sys/cpu.h>
47 #include <sys/ipi.h>
48 #include <sys/intr.h>
49 #include <sys/kcpuset.h>
50 #include <sys/kmem.h>
51 #include <sys/lock.h>
52 #include <sys/mutex.h>
53 
54 /*
55  * An array of the IPI handlers used for asynchronous invocation.
56  * The lock protects the slot allocation.
57  */
58 
59 typedef struct {
60 	ipi_func_t	func;
61 	void *		arg;
62 } ipi_intr_t;
63 
64 static kmutex_t		ipi_mngmt_lock;
65 static ipi_intr_t	ipi_intrs[IPI_MAXREG]	__cacheline_aligned;
66 
67 /*
68  * Per-CPU mailbox for IPI messages: it is a single cache line storing
69  * up to IPI_MSG_MAX messages.  This interface is built on top of the
70  * synchronous IPIs.
71  */
72 
73 #define	IPI_MSG_SLOTS	(CACHE_LINE_SIZE / sizeof(ipi_msg_t *))
74 #define	IPI_MSG_MAX	IPI_MSG_SLOTS
75 
76 typedef struct {
77 	ipi_msg_t *	msg[IPI_MSG_SLOTS];
78 } ipi_mbox_t;
79 
80 
81 /* Mailboxes for the synchronous IPIs. */
82 static ipi_mbox_t *	ipi_mboxes	__read_mostly;
83 static struct evcnt	ipi_mboxfull_ev	__cacheline_aligned;
84 static void		ipi_msg_cpu_handler(void *);
85 
86 /* Handler for the synchronous IPIs - it must be zero. */
87 #define	IPI_SYNCH_ID	0
88 
89 #ifndef MULTIPROCESSOR
90 #define	cpu_ipi(ci)	KASSERT(ci == NULL)
91 #endif
92 
93 void
94 ipi_sysinit(void)
95 {
96 	const size_t len = ncpu * sizeof(ipi_mbox_t);
97 
98 	/* Initialise the per-CPU bit fields. */
99 	for (u_int i = 0; i < ncpu; i++) {
100 		struct cpu_info *ci = cpu_lookup(i);
101 		memset(&ci->ci_ipipend, 0, sizeof(ci->ci_ipipend));
102 	}
103 	mutex_init(&ipi_mngmt_lock, MUTEX_DEFAULT, IPL_NONE);
104 	memset(ipi_intrs, 0, sizeof(ipi_intrs));
105 
106 	/* Allocate per-CPU IPI mailboxes. */
107 	ipi_mboxes = kmem_zalloc(len, KM_SLEEP);
108 	KASSERT(ipi_mboxes != NULL);
109 
110 	/*
111 	 * Register the handler for synchronous IPIs.  This mechanism
112 	 * is built on top of the asynchronous interface.  Slot zero is
113 	 * reserved permanently; it is also handy to use zero as a failure
114 	 * for other registers (as it is potentially less error-prone).
115 	 */
116 	ipi_intrs[IPI_SYNCH_ID].func = ipi_msg_cpu_handler;
117 
118 	evcnt_attach_dynamic(&ipi_mboxfull_ev, EVCNT_TYPE_MISC, NULL,
119 	   "ipi", "full");
120 }
121 
122 /*
123  * ipi_register: register an asynchronous IPI handler.
124  *
125  * => Returns IPI ID which is greater than zero; on failure - zero.
126  */
127 u_int
128 ipi_register(ipi_func_t func, void *arg)
129 {
130 	mutex_enter(&ipi_mngmt_lock);
131 	for (u_int i = 0; i < IPI_MAXREG; i++) {
132 		if (ipi_intrs[i].func == NULL) {
133 			/* Register the function. */
134 			ipi_intrs[i].func = func;
135 			ipi_intrs[i].arg = arg;
136 			mutex_exit(&ipi_mngmt_lock);
137 
138 			KASSERT(i != IPI_SYNCH_ID);
139 			return i;
140 		}
141 	}
142 	mutex_exit(&ipi_mngmt_lock);
143 	printf("WARNING: ipi_register: table full, increase IPI_MAXREG\n");
144 	return 0;
145 }
146 
147 /*
148  * ipi_unregister: release the IPI handler given the ID.
149  */
150 void
151 ipi_unregister(u_int ipi_id)
152 {
153 	ipi_msg_t ipimsg = { .func = __FPTRCAST(ipi_func_t, nullop) };
154 
155 	KASSERT(ipi_id != IPI_SYNCH_ID);
156 	KASSERT(ipi_id < IPI_MAXREG);
157 
158 	/* Release the slot. */
159 	mutex_enter(&ipi_mngmt_lock);
160 	KASSERT(ipi_intrs[ipi_id].func != NULL);
161 	ipi_intrs[ipi_id].func = NULL;
162 
163 	/* Ensure that there are no IPIs in flight. */
164 	kpreempt_disable();
165 	ipi_broadcast(&ipimsg, false);
166 	ipi_wait(&ipimsg);
167 	kpreempt_enable();
168 	mutex_exit(&ipi_mngmt_lock);
169 }
170 
171 /*
172  * ipi_mark_pending: internal routine to mark an IPI pending on the
173  * specified CPU (which might be curcpu()).
174  */
175 static bool
176 ipi_mark_pending(u_int ipi_id, struct cpu_info *ci)
177 {
178 	const u_int i = ipi_id >> IPI_BITW_SHIFT;
179 	const uint32_t bitm = 1U << (ipi_id & IPI_BITW_MASK);
180 
181 	KASSERT(ipi_id < IPI_MAXREG);
182 	KASSERT(kpreempt_disabled());
183 
184 	/* Mark as pending and send an IPI. */
185 	if (membar_consumer(), (ci->ci_ipipend[i] & bitm) == 0) {
186 		atomic_or_32(&ci->ci_ipipend[i], bitm);
187 		return true;
188 	}
189 	return false;
190 }
191 
192 /*
193  * ipi_trigger: asynchronously send an IPI to the specified CPU.
194  */
195 void
196 ipi_trigger(u_int ipi_id, struct cpu_info *ci)
197 {
198 
199 	KASSERT(curcpu() != ci);
200 	if (ipi_mark_pending(ipi_id, ci)) {
201 		cpu_ipi(ci);
202 	}
203 }
204 
205 /*
206  * ipi_trigger_multi_internal: the guts of ipi_trigger_multi() and
207  * ipi_trigger_broadcast().
208  */
209 static void
210 ipi_trigger_multi_internal(u_int ipi_id, const kcpuset_t *target,
211     bool skip_self)
212 {
213 	const cpuid_t selfid = cpu_index(curcpu());
214 	CPU_INFO_ITERATOR cii;
215 	struct cpu_info *ci;
216 
217 	KASSERT(kpreempt_disabled());
218 	KASSERT(target != NULL);
219 
220 	for (CPU_INFO_FOREACH(cii, ci)) {
221 		const cpuid_t cpuid = cpu_index(ci);
222 
223 		if (!kcpuset_isset(target, cpuid) || cpuid == selfid) {
224 			continue;
225 		}
226 		ipi_trigger(ipi_id, ci);
227 	}
228 	if (!skip_self && kcpuset_isset(target, selfid)) {
229 		ipi_mark_pending(ipi_id, curcpu());
230 		int s = splhigh();
231 		ipi_cpu_handler();
232 		splx(s);
233 	}
234 }
235 
236 /*
237  * ipi_trigger_multi: same as ipi_trigger() but sends to the multiple
238  * CPUs given the target CPU set.
239  */
240 void
241 ipi_trigger_multi(u_int ipi_id, const kcpuset_t *target)
242 {
243 	ipi_trigger_multi_internal(ipi_id, target, false);
244 }
245 
246 /*
247  * ipi_trigger_broadcast: same as ipi_trigger_multi() to kcpuset_attached,
248  * optionally skipping the sending CPU.
249  */
250 void
251 ipi_trigger_broadcast(u_int ipi_id, bool skip_self)
252 {
253 	ipi_trigger_multi_internal(ipi_id, kcpuset_attached, skip_self);
254 }
255 
256 /*
257  * put_msg: insert message into the mailbox.
258  */
259 static inline void
260 put_msg(ipi_mbox_t *mbox, ipi_msg_t *msg)
261 {
262 	int count = SPINLOCK_BACKOFF_MIN;
263 again:
264 	for (u_int i = 0; i < IPI_MSG_MAX; i++) {
265 		if (__predict_true(mbox->msg[i] == NULL) &&
266 		    atomic_cas_ptr(&mbox->msg[i], NULL, msg) == NULL) {
267 			return;
268 		}
269 	}
270 
271 	/* All slots are full: we have to spin-wait. */
272 	ipi_mboxfull_ev.ev_count++;
273 	SPINLOCK_BACKOFF(count);
274 	goto again;
275 }
276 
277 /*
278  * ipi_cpu_handler: the IPI handler.
279  */
280 void
281 ipi_cpu_handler(void)
282 {
283 	struct cpu_info * const ci = curcpu();
284 
285 	/*
286 	 * Handle asynchronous IPIs: inspect per-CPU bit field, extract
287 	 * IPI ID numbers and execute functions in those slots.
288 	 */
289 	for (u_int i = 0; i < IPI_BITWORDS; i++) {
290 		uint32_t pending, bit;
291 
292 		if (ci->ci_ipipend[i] == 0) {
293 			continue;
294 		}
295 		pending = atomic_swap_32(&ci->ci_ipipend[i], 0);
296 #ifndef __HAVE_ATOMIC_AS_MEMBAR
297 		membar_producer();
298 #endif
299 		while ((bit = ffs(pending)) != 0) {
300 			const u_int ipi_id = (i << IPI_BITW_SHIFT) | --bit;
301 			ipi_intr_t *ipi_hdl = &ipi_intrs[ipi_id];
302 
303 			pending &= ~(1U << bit);
304 			KASSERT(ipi_hdl->func != NULL);
305 			ipi_hdl->func(ipi_hdl->arg);
306 		}
307 	}
308 }
309 
310 /*
311  * ipi_msg_cpu_handler: handle synchronous IPIs - iterate mailbox,
312  * execute the passed functions and acknowledge the messages.
313  */
314 static void
315 ipi_msg_cpu_handler(void *arg __unused)
316 {
317 	const struct cpu_info * const ci = curcpu();
318 	ipi_mbox_t *mbox = &ipi_mboxes[cpu_index(ci)];
319 
320 	for (u_int i = 0; i < IPI_MSG_MAX; i++) {
321 		ipi_msg_t *msg;
322 
323 		/* Get the message. */
324 		if ((msg = mbox->msg[i]) == NULL) {
325 			continue;
326 		}
327 		mbox->msg[i] = NULL;
328 
329 		/* Execute the handler. */
330 		KASSERT(msg->func);
331 		msg->func(msg->arg);
332 
333 		/* Ack the request. */
334 #ifndef __HAVE_ATOMIC_AS_MEMBAR
335 		membar_producer();
336 #endif
337 		atomic_dec_uint(&msg->_pending);
338 	}
339 }
340 
341 /*
342  * ipi_unicast: send an IPI to a single CPU.
343  *
344  * => The CPU must be remote; must not be local.
345  * => The caller must ipi_wait() on the message for completion.
346  */
347 void
348 ipi_unicast(ipi_msg_t *msg, struct cpu_info *ci)
349 {
350 	const cpuid_t id = cpu_index(ci);
351 
352 	KASSERT(msg->func != NULL);
353 	KASSERT(kpreempt_disabled());
354 	KASSERT(curcpu() != ci);
355 
356 	msg->_pending = 1;
357 	membar_producer();
358 
359 	put_msg(&ipi_mboxes[id], msg);
360 	ipi_trigger(IPI_SYNCH_ID, ci);
361 }
362 
363 /*
364  * ipi_multicast: send an IPI to each CPU in the specified set.
365  *
366  * => The caller must ipi_wait() on the message for completion.
367  */
368 void
369 ipi_multicast(ipi_msg_t *msg, const kcpuset_t *target)
370 {
371 	const struct cpu_info * const self = curcpu();
372 	CPU_INFO_ITERATOR cii;
373 	struct cpu_info *ci;
374 	u_int local;
375 
376 	KASSERT(msg->func != NULL);
377 	KASSERT(kpreempt_disabled());
378 
379 	local = !!kcpuset_isset(target, cpu_index(self));
380 	msg->_pending = kcpuset_countset(target) - local;
381 	membar_producer();
382 
383 	for (CPU_INFO_FOREACH(cii, ci)) {
384 		cpuid_t id;
385 
386 		if (__predict_false(ci == self)) {
387 			continue;
388 		}
389 		id = cpu_index(ci);
390 		if (!kcpuset_isset(target, id)) {
391 			continue;
392 		}
393 		put_msg(&ipi_mboxes[id], msg);
394 		ipi_trigger(IPI_SYNCH_ID, ci);
395 	}
396 	if (local) {
397 		msg->func(msg->arg);
398 	}
399 }
400 
401 /*
402  * ipi_broadcast: send an IPI to all CPUs.
403  *
404  * => The caller must ipi_wait() on the message for completion.
405  */
406 void
407 ipi_broadcast(ipi_msg_t *msg, bool skip_self)
408 {
409 	const struct cpu_info * const self = curcpu();
410 	CPU_INFO_ITERATOR cii;
411 	struct cpu_info *ci;
412 
413 	KASSERT(msg->func != NULL);
414 	KASSERT(kpreempt_disabled());
415 
416 	msg->_pending = ncpu - 1;
417 	membar_producer();
418 
419 	/* Broadcast IPIs for remote CPUs. */
420 	for (CPU_INFO_FOREACH(cii, ci)) {
421 		cpuid_t id;
422 
423 		if (__predict_false(ci == self)) {
424 			continue;
425 		}
426 		id = cpu_index(ci);
427 		put_msg(&ipi_mboxes[id], msg);
428 		ipi_trigger(IPI_SYNCH_ID, ci);
429 	}
430 
431 	if (!skip_self) {
432 		/* Finally, execute locally. */
433 		msg->func(msg->arg);
434 	}
435 }
436 
437 /*
438  * ipi_wait: spin-wait until the message is processed.
439  */
440 void
441 ipi_wait(ipi_msg_t *msg)
442 {
443 	int count = SPINLOCK_BACKOFF_MIN;
444 
445 	while (msg->_pending) {
446 		KASSERT(msg->_pending < ncpu);
447 		SPINLOCK_BACKOFF(count);
448 	}
449 }
450