subr_xcall.c revision 1.22 1 /* $NetBSD: subr_xcall.c,v 1.22 2018/02/03 11:30:01 martin 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 1.22 2018/02/03 11:30:01 martin 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 #if IPL_SOFTCLOCK != IPL_SOFTBIO
190 case IPL_SOFTCLOCK:
191 return __SHIFTIN(XC_IPL_SOFTCLOCK, XC_IPL_MASK);
192 #endif
193 #if IPL_SOFTNET != IPL_SOFTBIO
194 case IPL_SOFTNET:
195 return __SHIFTIN(XC_IPL_SOFTNET, XC_IPL_MASK);
196 #endif
197 }
198
199 panic("Invalid IPL: %d", ipl);
200 }
201
202 /*
203 * Extract an XC_IPL from flags of xc_broadcast or xc_unicast.
204 */
205 static inline unsigned int
206 xc_extract_ipl(unsigned int flags)
207 {
208
209 return __SHIFTOUT(flags, XC_IPL_MASK);
210 }
211
212 /*
213 * xc_init_cpu:
214 *
215 * Initialize the cross-call subsystem. Called once for each CPU
216 * in the system as they are attached.
217 */
218 void
219 xc_init_cpu(struct cpu_info *ci)
220 {
221 static bool again = false;
222 int error __diagused;
223
224 if (!again) {
225 /* Autoconfiguration will prevent re-entry. */
226 xc_init();
227 again = true;
228 }
229 cv_init(&ci->ci_data.cpu_xcall, "xcall");
230 error = kthread_create(PRI_XCALL, KTHREAD_MPSAFE, ci, xc_thread,
231 NULL, NULL, "xcall/%u", ci->ci_index);
232 KASSERT(error == 0);
233 }
234
235 /*
236 * xc_broadcast:
237 *
238 * Trigger a call on all CPUs in the system.
239 */
240 uint64_t
241 xc_broadcast(unsigned int flags, xcfunc_t func, void *arg1, void *arg2)
242 {
243
244 KASSERT(!cpu_intr_p() && !cpu_softintr_p());
245
246 if ((flags & XC_HIGHPRI) != 0) {
247 int ipl = xc_extract_ipl(flags);
248 return xc_highpri(func, arg1, arg2, NULL, ipl);
249 } else {
250 return xc_lowpri(func, arg1, arg2, NULL);
251 }
252 }
253
254 /*
255 * xc_unicast:
256 *
257 * Trigger a call on one CPU.
258 */
259 uint64_t
260 xc_unicast(unsigned int flags, xcfunc_t func, void *arg1, void *arg2,
261 struct cpu_info *ci)
262 {
263
264 KASSERT(ci != NULL);
265 KASSERT(!cpu_intr_p() && !cpu_softintr_p());
266
267 if ((flags & XC_HIGHPRI) != 0) {
268 int ipl = xc_extract_ipl(flags);
269 return xc_highpri(func, arg1, arg2, ci, ipl);
270 } else {
271 return xc_lowpri(func, arg1, arg2, ci);
272 }
273 }
274
275 /*
276 * xc_wait:
277 *
278 * Wait for a cross call to complete.
279 */
280 void
281 xc_wait(uint64_t where)
282 {
283 xc_state_t *xc;
284
285 KASSERT(!cpu_intr_p() && !cpu_softintr_p());
286
287 /* Determine whether it is high or low priority cross-call. */
288 if ((where & XC_PRI_BIT) != 0) {
289 xc = &xc_high_pri;
290 where &= ~XC_PRI_BIT;
291 } else {
292 xc = &xc_low_pri;
293 }
294
295 /* Fast path, if already done. */
296 if (xc->xc_donep >= where) {
297 return;
298 }
299
300 /* Slow path: block until awoken. */
301 mutex_enter(&xc->xc_lock);
302 while (xc->xc_donep < where) {
303 cv_wait(&xc->xc_busy, &xc->xc_lock);
304 }
305 mutex_exit(&xc->xc_lock);
306 }
307
308 /*
309 * xc_lowpri:
310 *
311 * Trigger a low priority call on one or more CPUs.
312 */
313 static inline uint64_t
314 xc_lowpri(xcfunc_t func, void *arg1, void *arg2, struct cpu_info *ci)
315 {
316 xc_state_t *xc = &xc_low_pri;
317 CPU_INFO_ITERATOR cii;
318 uint64_t where;
319
320 mutex_enter(&xc->xc_lock);
321 while (xc->xc_headp != xc->xc_donep) {
322 cv_wait(&xc->xc_busy, &xc->xc_lock);
323 }
324 xc->xc_arg1 = arg1;
325 xc->xc_arg2 = arg2;
326 xc->xc_func = func;
327 if (ci == NULL) {
328 xc_broadcast_ev.ev_count++;
329 for (CPU_INFO_FOREACH(cii, ci)) {
330 if ((ci->ci_schedstate.spc_flags & SPCF_RUNNING) == 0)
331 continue;
332 xc->xc_headp += 1;
333 ci->ci_data.cpu_xcall_pending = true;
334 cv_signal(&ci->ci_data.cpu_xcall);
335 }
336 } else {
337 xc_unicast_ev.ev_count++;
338 xc->xc_headp += 1;
339 ci->ci_data.cpu_xcall_pending = true;
340 cv_signal(&ci->ci_data.cpu_xcall);
341 }
342 KASSERT(xc->xc_donep < xc->xc_headp);
343 where = xc->xc_headp;
344 mutex_exit(&xc->xc_lock);
345
346 /* Return a low priority ticket. */
347 KASSERT((where & XC_PRI_BIT) == 0);
348 return where;
349 }
350
351 /*
352 * xc_thread:
353 *
354 * One thread per-CPU to dispatch low priority calls.
355 */
356 static void
357 xc_thread(void *cookie)
358 {
359 struct cpu_info *ci = curcpu();
360 xc_state_t *xc = &xc_low_pri;
361 void *arg1, *arg2;
362 xcfunc_t func;
363
364 mutex_enter(&xc->xc_lock);
365 for (;;) {
366 while (!ci->ci_data.cpu_xcall_pending) {
367 if (xc->xc_headp == xc->xc_donep) {
368 cv_broadcast(&xc->xc_busy);
369 }
370 cv_wait(&ci->ci_data.cpu_xcall, &xc->xc_lock);
371 KASSERT(ci == curcpu());
372 }
373 ci->ci_data.cpu_xcall_pending = false;
374 func = xc->xc_func;
375 arg1 = xc->xc_arg1;
376 arg2 = xc->xc_arg2;
377 mutex_exit(&xc->xc_lock);
378
379 KASSERT(func != NULL);
380 (*func)(arg1, arg2);
381
382 mutex_enter(&xc->xc_lock);
383 xc->xc_donep++;
384 }
385 /* NOTREACHED */
386 }
387
388 /*
389 * xc_ipi_handler:
390 *
391 * Handler of cross-call IPI.
392 */
393 void
394 xc_ipi_handler(void)
395 {
396 xc_state_t *xc = & xc_high_pri;
397
398 KASSERT(xc->xc_ipl < __arraycount(xc_sihs));
399
400 /* Executes xc__highpri_intr() via software interrupt. */
401 softint_schedule(xc_sihs[xc->xc_ipl]);
402 }
403
404 /*
405 * xc__highpri_intr:
406 *
407 * A software interrupt handler for high priority calls.
408 */
409 void
410 xc__highpri_intr(void *dummy)
411 {
412 xc_state_t *xc = &xc_high_pri;
413 void *arg1, *arg2;
414 xcfunc_t func;
415
416 KASSERTMSG(!cpu_intr_p(), "high priority xcall for function %p",
417 xc->xc_func);
418 /*
419 * Lock-less fetch of function and its arguments.
420 * Safe since it cannot change at this point.
421 */
422 KASSERT(xc->xc_donep < xc->xc_headp);
423 func = xc->xc_func;
424 arg1 = xc->xc_arg1;
425 arg2 = xc->xc_arg2;
426
427 KASSERT(func != NULL);
428 (*func)(arg1, arg2);
429
430 /*
431 * Note the request as done, and if we have reached the head,
432 * cross-call has been processed - notify waiters, if any.
433 */
434 mutex_enter(&xc->xc_lock);
435 if (++xc->xc_donep == xc->xc_headp) {
436 cv_broadcast(&xc->xc_busy);
437 }
438 mutex_exit(&xc->xc_lock);
439 }
440
441 /*
442 * xc_highpri:
443 *
444 * Trigger a high priority call on one or more CPUs.
445 */
446 static inline uint64_t
447 xc_highpri(xcfunc_t func, void *arg1, void *arg2, struct cpu_info *ci,
448 unsigned int ipl)
449 {
450 xc_state_t *xc = &xc_high_pri;
451 uint64_t where;
452
453 mutex_enter(&xc->xc_lock);
454 while (xc->xc_headp != xc->xc_donep) {
455 cv_wait(&xc->xc_busy, &xc->xc_lock);
456 }
457 xc->xc_func = func;
458 xc->xc_arg1 = arg1;
459 xc->xc_arg2 = arg2;
460 xc->xc_headp += (ci ? 1 : ncpu);
461 xc->xc_ipl = ipl;
462 where = xc->xc_headp;
463 mutex_exit(&xc->xc_lock);
464
465 /*
466 * Send the IPI once lock is released.
467 * Note: it will handle the local CPU case.
468 */
469
470 #ifdef _RUMPKERNEL
471 rump_xc_highpri(ci);
472 #else
473 #ifdef MULTIPROCESSOR
474 kpreempt_disable();
475 if (curcpu() == ci) {
476 /* Unicast: local CPU. */
477 xc_ipi_handler();
478 } else if (ci) {
479 /* Unicast: remote CPU. */
480 xc_send_ipi(ci);
481 } else {
482 /* Broadcast: all, including local. */
483 xc_send_ipi(NULL);
484 xc_ipi_handler();
485 }
486 kpreempt_enable();
487 #else
488 KASSERT(ci == NULL || curcpu() == ci);
489 xc_ipi_handler();
490 #endif
491 #endif
492
493 /* Indicate a high priority ticket. */
494 return (where | XC_PRI_BIT);
495 }
496