subr_xcall.c revision 1.21 1 /* $NetBSD: subr_xcall.c,v 1.21 2018/02/01 03:15:29 ozaki-r Exp $ */
2
3 /*-
4 * Copyright (c) 2007-2010 The NetBSD Foundation, Inc.
5 * All rights reserved.
6 *
7 * This code is derived from software contributed to The NetBSD Foundation
8 * by Andrew Doran and 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 * Cross call support
34 *
35 * Background
36 *
37 * Sometimes it is necessary to modify hardware state that is tied
38 * directly to individual CPUs (such as a CPU's local timer), and
39 * these updates can not be done remotely by another CPU. The LWP
40 * requesting the update may be unable to guarantee that it will be
41 * running on the CPU where the update must occur, when the update
42 * occurs.
43 *
44 * Additionally, it's sometimes necessary to modify per-CPU software
45 * state from a remote CPU. Where these update operations are so
46 * rare or the access to the per-CPU data so frequent that the cost
47 * of using locking or atomic operations to provide coherency is
48 * prohibitive, another way must be found.
49 *
50 * Cross calls help to solve these types of problem by allowing
51 * any CPU in the system to request that an arbitrary function be
52 * executed on any other CPU.
53 *
54 * Implementation
55 *
56 * A slow mechanism for making 'low priority' cross calls is
57 * provided. The function to be executed runs on the remote CPU
58 * within a bound kthread. No queueing is provided, and the
59 * implementation uses global state. The function being called may
60 * block briefly on locks, but in doing so must be careful to not
61 * interfere with other cross calls in the system. The function is
62 * called with thread context and not from a soft interrupt, so it
63 * can ensure that it is not interrupting other code running on the
64 * CPU, and so has exclusive access to the CPU. Since this facility
65 * is heavyweight, it's expected that it will not be used often.
66 *
67 * Cross calls must not allocate memory, as the pagedaemon uses
68 * them (and memory allocation may need to wait on the pagedaemon).
69 *
70 * A low-overhead mechanism for high priority calls (XC_HIGHPRI) is
71 * also provided. The function to be executed runs on a software
72 * interrupt context, at IPL_SOFTSERIAL level, and is expected to
73 * be very lightweight, e.g. avoid blocking.
74 */
75
76 #include <sys/cdefs.h>
77 __KERNEL_RCSID(0, "$NetBSD: subr_xcall.c,v 1.21 2018/02/01 03:15:29 ozaki-r Exp $");
78
79 #include <sys/types.h>
80 #include <sys/param.h>
81 #include <sys/xcall.h>
82 #include <sys/mutex.h>
83 #include <sys/condvar.h>
84 #include <sys/evcnt.h>
85 #include <sys/kthread.h>
86 #include <sys/cpu.h>
87
88 #ifdef _RUMPKERNEL
89 #include "rump_private.h"
90 #endif
91
92 /* Cross-call state box. */
93 typedef struct {
94 kmutex_t xc_lock;
95 kcondvar_t xc_busy;
96 xcfunc_t xc_func;
97 void * xc_arg1;
98 void * xc_arg2;
99 uint64_t xc_headp;
100 uint64_t xc_donep;
101 unsigned int xc_ipl;
102 } xc_state_t;
103
104 /* Bit indicating high (1) or low (0) priority. */
105 #define XC_PRI_BIT (1ULL << 63)
106
107 /* Low priority xcall structures. */
108 static xc_state_t xc_low_pri __cacheline_aligned;
109
110 /* High priority xcall structures. */
111 static xc_state_t xc_high_pri __cacheline_aligned;
112 static void * xc_sihs[4] __cacheline_aligned;
113
114 /* Event counters. */
115 static struct evcnt xc_unicast_ev __cacheline_aligned;
116 static struct evcnt xc_broadcast_ev __cacheline_aligned;
117
118 static void xc_init(void);
119 static void xc_thread(void *);
120
121 static inline uint64_t xc_highpri(xcfunc_t, void *, void *, struct cpu_info *,
122 unsigned int);
123 static inline uint64_t xc_lowpri(xcfunc_t, void *, void *, struct cpu_info *);
124
125 /* The internal form of IPL */
126 #define XC_IPL_MASK 0xff00
127 /*
128 * Assign 0 to XC_IPL_SOFTSERIAL to treat IPL_SOFTSERIAL as the default value
129 * (just XC_HIGHPRI).
130 */
131 #define XC_IPL_SOFTSERIAL 0
132 #define XC_IPL_SOFTBIO 1
133 #define XC_IPL_SOFTCLOCK 2
134 #define XC_IPL_SOFTNET 3
135 #define XC_IPL_MAX XC_IPL_SOFTNET
136
137 CTASSERT(XC_IPL_MAX <= __arraycount(xc_sihs));
138
139 /*
140 * xc_init:
141 *
142 * Initialize low and high priority cross-call structures.
143 */
144 static void
145 xc_init(void)
146 {
147 xc_state_t *xclo = &xc_low_pri, *xchi = &xc_high_pri;
148
149 memset(xclo, 0, sizeof(xc_state_t));
150 mutex_init(&xclo->xc_lock, MUTEX_DEFAULT, IPL_NONE);
151 cv_init(&xclo->xc_busy, "xclocv");
152
153 memset(xchi, 0, sizeof(xc_state_t));
154 mutex_init(&xchi->xc_lock, MUTEX_DEFAULT, IPL_SOFTSERIAL);
155 cv_init(&xchi->xc_busy, "xchicv");
156
157 #define SETUP_SOFTINT(xipl, sipl) do { \
158 xc_sihs[(xipl)] = softint_establish( (sipl) | SOFTINT_MPSAFE,\
159 xc__highpri_intr, NULL); \
160 KASSERT(xc_sihs[(xipl)] != NULL); \
161 } while (0)
162
163 SETUP_SOFTINT(XC_IPL_SOFTSERIAL, SOFTINT_SERIAL);
164 SETUP_SOFTINT(XC_IPL_SOFTBIO, SOFTINT_BIO);
165 SETUP_SOFTINT(XC_IPL_SOFTCLOCK, SOFTINT_CLOCK);
166 SETUP_SOFTINT(XC_IPL_SOFTNET, SOFTINT_NET);
167
168 #undef SETUP_SOFTINT
169
170 evcnt_attach_dynamic(&xc_unicast_ev, EVCNT_TYPE_MISC, NULL,
171 "crosscall", "unicast");
172 evcnt_attach_dynamic(&xc_broadcast_ev, EVCNT_TYPE_MISC, NULL,
173 "crosscall", "broadcast");
174 }
175
176 /*
177 * Encode an IPL to a form that can be embedded into flags of xc_broadcast
178 * or xc_unicast.
179 */
180 unsigned int
181 xc_encode_ipl(int ipl)
182 {
183
184 switch (ipl) {
185 case IPL_SOFTSERIAL:
186 return __SHIFTIN(XC_IPL_SOFTSERIAL, XC_IPL_MASK);
187 case IPL_SOFTBIO:
188 return __SHIFTIN(XC_IPL_SOFTBIO, XC_IPL_MASK);
189 case IPL_SOFTCLOCK:
190 return __SHIFTIN(XC_IPL_SOFTCLOCK, XC_IPL_MASK);
191 case IPL_SOFTNET:
192 return __SHIFTIN(XC_IPL_SOFTNET, XC_IPL_MASK);
193 }
194
195 panic("Invalid IPL: %d", ipl);
196 }
197
198 /*
199 * Extract an XC_IPL from flags of xc_broadcast or xc_unicast.
200 */
201 static inline unsigned int
202 xc_extract_ipl(unsigned int flags)
203 {
204
205 return __SHIFTOUT(flags, XC_IPL_MASK);
206 }
207
208 /*
209 * xc_init_cpu:
210 *
211 * Initialize the cross-call subsystem. Called once for each CPU
212 * in the system as they are attached.
213 */
214 void
215 xc_init_cpu(struct cpu_info *ci)
216 {
217 static bool again = false;
218 int error __diagused;
219
220 if (!again) {
221 /* Autoconfiguration will prevent re-entry. */
222 xc_init();
223 again = true;
224 }
225 cv_init(&ci->ci_data.cpu_xcall, "xcall");
226 error = kthread_create(PRI_XCALL, KTHREAD_MPSAFE, ci, xc_thread,
227 NULL, NULL, "xcall/%u", ci->ci_index);
228 KASSERT(error == 0);
229 }
230
231 /*
232 * xc_broadcast:
233 *
234 * Trigger a call on all CPUs in the system.
235 */
236 uint64_t
237 xc_broadcast(unsigned int flags, xcfunc_t func, void *arg1, void *arg2)
238 {
239
240 KASSERT(!cpu_intr_p() && !cpu_softintr_p());
241
242 if ((flags & XC_HIGHPRI) != 0) {
243 int ipl = xc_extract_ipl(flags);
244 return xc_highpri(func, arg1, arg2, NULL, ipl);
245 } else {
246 return xc_lowpri(func, arg1, arg2, NULL);
247 }
248 }
249
250 /*
251 * xc_unicast:
252 *
253 * Trigger a call on one CPU.
254 */
255 uint64_t
256 xc_unicast(unsigned int flags, xcfunc_t func, void *arg1, void *arg2,
257 struct cpu_info *ci)
258 {
259
260 KASSERT(ci != NULL);
261 KASSERT(!cpu_intr_p() && !cpu_softintr_p());
262
263 if ((flags & XC_HIGHPRI) != 0) {
264 int ipl = xc_extract_ipl(flags);
265 return xc_highpri(func, arg1, arg2, ci, ipl);
266 } else {
267 return xc_lowpri(func, arg1, arg2, ci);
268 }
269 }
270
271 /*
272 * xc_wait:
273 *
274 * Wait for a cross call to complete.
275 */
276 void
277 xc_wait(uint64_t where)
278 {
279 xc_state_t *xc;
280
281 KASSERT(!cpu_intr_p() && !cpu_softintr_p());
282
283 /* Determine whether it is high or low priority cross-call. */
284 if ((where & XC_PRI_BIT) != 0) {
285 xc = &xc_high_pri;
286 where &= ~XC_PRI_BIT;
287 } else {
288 xc = &xc_low_pri;
289 }
290
291 /* Fast path, if already done. */
292 if (xc->xc_donep >= where) {
293 return;
294 }
295
296 /* Slow path: block until awoken. */
297 mutex_enter(&xc->xc_lock);
298 while (xc->xc_donep < where) {
299 cv_wait(&xc->xc_busy, &xc->xc_lock);
300 }
301 mutex_exit(&xc->xc_lock);
302 }
303
304 /*
305 * xc_lowpri:
306 *
307 * Trigger a low priority call on one or more CPUs.
308 */
309 static inline uint64_t
310 xc_lowpri(xcfunc_t func, void *arg1, void *arg2, struct cpu_info *ci)
311 {
312 xc_state_t *xc = &xc_low_pri;
313 CPU_INFO_ITERATOR cii;
314 uint64_t where;
315
316 mutex_enter(&xc->xc_lock);
317 while (xc->xc_headp != xc->xc_donep) {
318 cv_wait(&xc->xc_busy, &xc->xc_lock);
319 }
320 xc->xc_arg1 = arg1;
321 xc->xc_arg2 = arg2;
322 xc->xc_func = func;
323 if (ci == NULL) {
324 xc_broadcast_ev.ev_count++;
325 for (CPU_INFO_FOREACH(cii, ci)) {
326 if ((ci->ci_schedstate.spc_flags & SPCF_RUNNING) == 0)
327 continue;
328 xc->xc_headp += 1;
329 ci->ci_data.cpu_xcall_pending = true;
330 cv_signal(&ci->ci_data.cpu_xcall);
331 }
332 } else {
333 xc_unicast_ev.ev_count++;
334 xc->xc_headp += 1;
335 ci->ci_data.cpu_xcall_pending = true;
336 cv_signal(&ci->ci_data.cpu_xcall);
337 }
338 KASSERT(xc->xc_donep < xc->xc_headp);
339 where = xc->xc_headp;
340 mutex_exit(&xc->xc_lock);
341
342 /* Return a low priority ticket. */
343 KASSERT((where & XC_PRI_BIT) == 0);
344 return where;
345 }
346
347 /*
348 * xc_thread:
349 *
350 * One thread per-CPU to dispatch low priority calls.
351 */
352 static void
353 xc_thread(void *cookie)
354 {
355 struct cpu_info *ci = curcpu();
356 xc_state_t *xc = &xc_low_pri;
357 void *arg1, *arg2;
358 xcfunc_t func;
359
360 mutex_enter(&xc->xc_lock);
361 for (;;) {
362 while (!ci->ci_data.cpu_xcall_pending) {
363 if (xc->xc_headp == xc->xc_donep) {
364 cv_broadcast(&xc->xc_busy);
365 }
366 cv_wait(&ci->ci_data.cpu_xcall, &xc->xc_lock);
367 KASSERT(ci == curcpu());
368 }
369 ci->ci_data.cpu_xcall_pending = false;
370 func = xc->xc_func;
371 arg1 = xc->xc_arg1;
372 arg2 = xc->xc_arg2;
373 mutex_exit(&xc->xc_lock);
374
375 KASSERT(func != NULL);
376 (*func)(arg1, arg2);
377
378 mutex_enter(&xc->xc_lock);
379 xc->xc_donep++;
380 }
381 /* NOTREACHED */
382 }
383
384 /*
385 * xc_ipi_handler:
386 *
387 * Handler of cross-call IPI.
388 */
389 void
390 xc_ipi_handler(void)
391 {
392 xc_state_t *xc = & xc_high_pri;
393
394 KASSERT(xc->xc_ipl < __arraycount(xc_sihs));
395
396 /* Executes xc__highpri_intr() via software interrupt. */
397 softint_schedule(xc_sihs[xc->xc_ipl]);
398 }
399
400 /*
401 * xc__highpri_intr:
402 *
403 * A software interrupt handler for high priority calls.
404 */
405 void
406 xc__highpri_intr(void *dummy)
407 {
408 xc_state_t *xc = &xc_high_pri;
409 void *arg1, *arg2;
410 xcfunc_t func;
411
412 KASSERTMSG(!cpu_intr_p(), "high priority xcall for function %p",
413 xc->xc_func);
414 /*
415 * Lock-less fetch of function and its arguments.
416 * Safe since it cannot change at this point.
417 */
418 KASSERT(xc->xc_donep < xc->xc_headp);
419 func = xc->xc_func;
420 arg1 = xc->xc_arg1;
421 arg2 = xc->xc_arg2;
422
423 KASSERT(func != NULL);
424 (*func)(arg1, arg2);
425
426 /*
427 * Note the request as done, and if we have reached the head,
428 * cross-call has been processed - notify waiters, if any.
429 */
430 mutex_enter(&xc->xc_lock);
431 if (++xc->xc_donep == xc->xc_headp) {
432 cv_broadcast(&xc->xc_busy);
433 }
434 mutex_exit(&xc->xc_lock);
435 }
436
437 /*
438 * xc_highpri:
439 *
440 * Trigger a high priority call on one or more CPUs.
441 */
442 static inline uint64_t
443 xc_highpri(xcfunc_t func, void *arg1, void *arg2, struct cpu_info *ci,
444 unsigned int ipl)
445 {
446 xc_state_t *xc = &xc_high_pri;
447 uint64_t where;
448
449 mutex_enter(&xc->xc_lock);
450 while (xc->xc_headp != xc->xc_donep) {
451 cv_wait(&xc->xc_busy, &xc->xc_lock);
452 }
453 xc->xc_func = func;
454 xc->xc_arg1 = arg1;
455 xc->xc_arg2 = arg2;
456 xc->xc_headp += (ci ? 1 : ncpu);
457 xc->xc_ipl = ipl;
458 where = xc->xc_headp;
459 mutex_exit(&xc->xc_lock);
460
461 /*
462 * Send the IPI once lock is released.
463 * Note: it will handle the local CPU case.
464 */
465
466 #ifdef _RUMPKERNEL
467 rump_xc_highpri(ci);
468 #else
469 #ifdef MULTIPROCESSOR
470 kpreempt_disable();
471 if (curcpu() == ci) {
472 /* Unicast: local CPU. */
473 xc_ipi_handler();
474 } else if (ci) {
475 /* Unicast: remote CPU. */
476 xc_send_ipi(ci);
477 } else {
478 /* Broadcast: all, including local. */
479 xc_send_ipi(NULL);
480 xc_ipi_handler();
481 }
482 kpreempt_enable();
483 #else
484 KASSERT(ci == NULL || curcpu() == ci);
485 xc_ipi_handler();
486 #endif
487 #endif
488
489 /* Indicate a high priority ticket. */
490 return (where | XC_PRI_BIT);
491 }
492