xref: /netbsd-src/sys/kern/kern_cpu.c (revision 288bb96063654ec504ca8732afc683d3ebc514b5)
1 /*	$NetBSD: kern_cpu.c,v 1.47 2011/06/29 06:22:21 matt Exp $	*/
2 
3 /*-
4  * Copyright (c) 2007, 2008, 2009, 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.
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  * Copyright (c)2007 YAMAMOTO Takashi,
34  * All rights reserved.
35  *
36  * Redistribution and use in source and binary forms, with or without
37  * modification, are permitted provided that the following conditions
38  * are met:
39  * 1. Redistributions of source code must retain the above copyright
40  *    notice, this list of conditions and the following disclaimer.
41  * 2. Redistributions in binary form must reproduce the above copyright
42  *    notice, this list of conditions and the following disclaimer in the
43  *    documentation and/or other materials provided with the distribution.
44  *
45  * THIS SOFTWARE IS PROVIDED BY THE AUTHOR AND CONTRIBUTORS ``AS IS'' AND
46  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
47  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
48  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR OR CONTRIBUTORS BE LIABLE
49  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
50  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
51  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
52  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
53  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
54  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
55  * SUCH DAMAGE.
56  */
57 
58 #include <sys/cdefs.h>
59 __KERNEL_RCSID(0, "$NetBSD: kern_cpu.c,v 1.47 2011/06/29 06:22:21 matt Exp $");
60 
61 #include <sys/param.h>
62 #include <sys/systm.h>
63 #include <sys/idle.h>
64 #include <sys/sched.h>
65 #include <sys/intr.h>
66 #include <sys/conf.h>
67 #include <sys/cpu.h>
68 #include <sys/cpuio.h>
69 #include <sys/proc.h>
70 #include <sys/percpu.h>
71 #include <sys/kernel.h>
72 #include <sys/kauth.h>
73 #include <sys/xcall.h>
74 #include <sys/pool.h>
75 #include <sys/kmem.h>
76 #include <sys/select.h>
77 #include <sys/namei.h>
78 #include <sys/callout.h>
79 
80 #include <uvm/uvm_extern.h>
81 
82 /*
83  * If the port has state that cpu_data is the first thing in cpu_info,
84  * verify the claim is true.  This will prevent the from getting out
85  * of sync.
86  */
87 #ifdef __HAVE_CPU_DATA_FIRST
88 CTASSERT(offsetof(struct cpu_info, ci_data) == 0);
89 #else
90 CTASSERT(offsetof(struct cpu_info, ci_data) != 0);
91 #endif
92 
93 void	cpuctlattach(int);
94 
95 static void	cpu_xc_online(struct cpu_info *);
96 static void	cpu_xc_offline(struct cpu_info *);
97 
98 dev_type_ioctl(cpuctl_ioctl);
99 
100 const struct cdevsw cpuctl_cdevsw = {
101 	nullopen, nullclose, nullread, nullwrite, cpuctl_ioctl,
102 	nullstop, notty, nopoll, nommap, nokqfilter,
103 	D_OTHER | D_MPSAFE
104 };
105 
106 kmutex_t	cpu_lock		__cacheline_aligned;
107 int		ncpu			__read_mostly;
108 int		ncpuonline		__read_mostly;
109 bool		mp_online		__read_mostly;
110 struct cpuqueue	cpu_queue		__cacheline_aligned
111     = CIRCLEQ_HEAD_INITIALIZER(cpu_queue);
112 
113 static struct cpu_info **cpu_infos	__read_mostly;
114 
115 int
116 mi_cpu_attach(struct cpu_info *ci)
117 {
118 	int error;
119 
120 	KASSERT(maxcpus > 0);
121 
122 	ci->ci_index = ncpu;
123 	CIRCLEQ_INSERT_TAIL(&cpu_queue, ci, ci_data.cpu_qchain);
124 	TAILQ_INIT(&ci->ci_data.cpu_ld_locks);
125 	__cpu_simple_lock_init(&ci->ci_data.cpu_ld_lock);
126 
127 	/* This is useful for eg, per-cpu evcnt */
128 	snprintf(ci->ci_data.cpu_name, sizeof(ci->ci_data.cpu_name), "cpu%d",
129 	    cpu_index(ci));
130 
131 	if (__predict_false(cpu_infos == NULL)) {
132 		cpu_infos =
133 		    kmem_zalloc(sizeof(cpu_infos[0]) * maxcpus, KM_SLEEP);
134 	}
135 	cpu_infos[cpu_index(ci)] = ci;
136 
137 	sched_cpuattach(ci);
138 
139 	error = create_idle_lwp(ci);
140 	if (error != 0) {
141 		/* XXX revert sched_cpuattach */
142 		return error;
143 	}
144 
145 	if (ci == curcpu())
146 		ci->ci_data.cpu_onproc = curlwp;
147 	else
148 		ci->ci_data.cpu_onproc = ci->ci_data.cpu_idlelwp;
149 
150 	percpu_init_cpu(ci);
151 	softint_init(ci);
152 	callout_init_cpu(ci);
153 	xc_init_cpu(ci);
154 	pool_cache_cpu_init(ci);
155 	selsysinit(ci);
156 	cache_cpu_init(ci);
157 	TAILQ_INIT(&ci->ci_data.cpu_biodone);
158 	ncpu++;
159 	ncpuonline++;
160 
161 	return 0;
162 }
163 
164 void
165 cpuctlattach(int dummy)
166 {
167 
168 	KASSERT(cpu_infos != NULL);
169 }
170 
171 int
172 cpuctl_ioctl(dev_t dev, u_long cmd, void *data, int flag, lwp_t *l)
173 {
174 	CPU_INFO_ITERATOR cii;
175 	cpustate_t *cs;
176 	struct cpu_info *ci;
177 	int error, i;
178 	u_int id;
179 
180 	error = 0;
181 
182 	mutex_enter(&cpu_lock);
183 	switch (cmd) {
184 	case IOC_CPU_SETSTATE:
185 		if (error == 0)
186 			cs = data;
187 		error = kauth_authorize_system(l->l_cred,
188 		    KAUTH_SYSTEM_CPU, KAUTH_REQ_SYSTEM_CPU_SETSTATE, cs, NULL,
189 		    NULL);
190 		if (error != 0)
191 			break;
192 		if (cs->cs_id >= maxcpus ||
193 		    (ci = cpu_lookup(cs->cs_id)) == NULL) {
194 			error = ESRCH;
195 			break;
196 		}
197 		error = cpu_setintr(ci, cs->cs_intr);
198 		error = cpu_setstate(ci, cs->cs_online);
199 		break;
200 
201 	case IOC_CPU_GETSTATE:
202 		if (error == 0)
203 			cs = data;
204 		id = cs->cs_id;
205 		memset(cs, 0, sizeof(*cs));
206 		cs->cs_id = id;
207 		if (cs->cs_id >= maxcpus ||
208 		    (ci = cpu_lookup(id)) == NULL) {
209 			error = ESRCH;
210 			break;
211 		}
212 		if ((ci->ci_schedstate.spc_flags & SPCF_OFFLINE) != 0)
213 			cs->cs_online = false;
214 		else
215 			cs->cs_online = true;
216 		if ((ci->ci_schedstate.spc_flags & SPCF_NOINTR) != 0)
217 			cs->cs_intr = false;
218 		else
219 			cs->cs_intr = true;
220 		cs->cs_lastmod = (int32_t)ci->ci_schedstate.spc_lastmod;
221 		cs->cs_lastmodhi = (int32_t)
222 		    (ci->ci_schedstate.spc_lastmod >> 32);
223 		cs->cs_intrcnt = cpu_intr_count(ci) + 1;
224 		break;
225 
226 	case IOC_CPU_MAPID:
227 		i = 0;
228 		for (CPU_INFO_FOREACH(cii, ci)) {
229 			if (i++ == *(int *)data)
230 				break;
231 		}
232 		if (ci == NULL)
233 			error = ESRCH;
234 		else
235 			*(int *)data = cpu_index(ci);
236 		break;
237 
238 	case IOC_CPU_GETCOUNT:
239 		*(int *)data = ncpu;
240 		break;
241 
242 	default:
243 		error = ENOTTY;
244 		break;
245 	}
246 	mutex_exit(&cpu_lock);
247 
248 	return error;
249 }
250 
251 struct cpu_info *
252 cpu_lookup(u_int idx)
253 {
254 	struct cpu_info *ci;
255 
256 	KASSERT(idx < maxcpus);
257 
258 	if (__predict_false(cpu_infos == NULL)) {
259 		KASSERT(idx == 0);
260 		return curcpu();
261 	}
262 
263 	ci = cpu_infos[idx];
264 	KASSERT(ci == NULL || cpu_index(ci) == idx);
265 
266 	return ci;
267 }
268 
269 static void
270 cpu_xc_offline(struct cpu_info *ci)
271 {
272 	struct schedstate_percpu *spc, *mspc = NULL;
273 	struct cpu_info *target_ci;
274 	struct lwp *l;
275 	CPU_INFO_ITERATOR cii;
276 	int s;
277 
278 	/*
279 	 * Thread that made the cross call (separate context) holds
280 	 * cpu_lock on our behalf.
281 	 */
282 	spc = &ci->ci_schedstate;
283 	s = splsched();
284 	spc->spc_flags |= SPCF_OFFLINE;
285 	splx(s);
286 
287 	/* Take the first available CPU for the migration. */
288 	for (CPU_INFO_FOREACH(cii, target_ci)) {
289 		mspc = &target_ci->ci_schedstate;
290 		if ((mspc->spc_flags & SPCF_OFFLINE) == 0)
291 			break;
292 	}
293 	KASSERT(target_ci != NULL);
294 
295 	/*
296 	 * Migrate all non-bound threads to the other CPU.  Note that this
297 	 * runs from the xcall thread, thus handling of LSONPROC is not needed.
298 	 */
299 	mutex_enter(proc_lock);
300 	LIST_FOREACH(l, &alllwp, l_list) {
301 		struct cpu_info *mci;
302 
303 		lwp_lock(l);
304 		if (l->l_cpu != ci || (l->l_pflag & (LP_BOUND | LP_INTR))) {
305 			lwp_unlock(l);
306 			continue;
307 		}
308 		/* Normal case - no affinity */
309 		if ((l->l_flag & LW_AFFINITY) == 0) {
310 			lwp_migrate(l, target_ci);
311 			continue;
312 		}
313 		/* Affinity is set, find an online CPU in the set */
314 		KASSERT(l->l_affinity != NULL);
315 		for (CPU_INFO_FOREACH(cii, mci)) {
316 			mspc = &mci->ci_schedstate;
317 			if ((mspc->spc_flags & SPCF_OFFLINE) == 0 &&
318 			    kcpuset_isset(cpu_index(mci), l->l_affinity))
319 				break;
320 		}
321 		if (mci == NULL) {
322 			lwp_unlock(l);
323 			mutex_exit(proc_lock);
324 			goto fail;
325 		}
326 		lwp_migrate(l, mci);
327 	}
328 	mutex_exit(proc_lock);
329 
330 #ifdef __HAVE_MD_CPU_OFFLINE
331 	cpu_offline_md();
332 #endif
333 	return;
334 fail:
335 	/* Just unset the SPCF_OFFLINE flag, caller will check */
336 	s = splsched();
337 	spc->spc_flags &= ~SPCF_OFFLINE;
338 	splx(s);
339 }
340 
341 static void
342 cpu_xc_online(struct cpu_info *ci)
343 {
344 	struct schedstate_percpu *spc;
345 	int s;
346 
347 	spc = &ci->ci_schedstate;
348 	s = splsched();
349 	spc->spc_flags &= ~SPCF_OFFLINE;
350 	splx(s);
351 }
352 
353 int
354 cpu_setstate(struct cpu_info *ci, bool online)
355 {
356 	struct schedstate_percpu *spc;
357 	CPU_INFO_ITERATOR cii;
358 	struct cpu_info *ci2;
359 	uint64_t where;
360 	xcfunc_t func;
361 	int nonline;
362 
363 	spc = &ci->ci_schedstate;
364 
365 	KASSERT(mutex_owned(&cpu_lock));
366 
367 	if (online) {
368 		if ((spc->spc_flags & SPCF_OFFLINE) == 0)
369 			return 0;
370 		func = (xcfunc_t)cpu_xc_online;
371 		ncpuonline++;
372 	} else {
373 		if ((spc->spc_flags & SPCF_OFFLINE) != 0)
374 			return 0;
375 		nonline = 0;
376 		/*
377 		 * Ensure that at least one CPU within the processor set
378 		 * stays online.  Revisit this later.
379 		 */
380 		for (CPU_INFO_FOREACH(cii, ci2)) {
381 			if ((ci2->ci_schedstate.spc_flags & SPCF_OFFLINE) != 0)
382 				continue;
383 			if (ci2->ci_schedstate.spc_psid != spc->spc_psid)
384 				continue;
385 			nonline++;
386 		}
387 		if (nonline == 1)
388 			return EBUSY;
389 		func = (xcfunc_t)cpu_xc_offline;
390 		ncpuonline--;
391 	}
392 
393 	where = xc_unicast(0, func, ci, NULL, ci);
394 	xc_wait(where);
395 	if (online) {
396 		KASSERT((spc->spc_flags & SPCF_OFFLINE) == 0);
397 	} else if ((spc->spc_flags & SPCF_OFFLINE) == 0) {
398 		/* If was not set offline, then it is busy */
399 		return EBUSY;
400 	}
401 
402 	spc->spc_lastmod = time_second;
403 	return 0;
404 }
405 
406 #ifdef __HAVE_INTR_CONTROL
407 static void
408 cpu_xc_intr(struct cpu_info *ci)
409 {
410 	struct schedstate_percpu *spc;
411 	int s;
412 
413 	spc = &ci->ci_schedstate;
414 	s = splsched();
415 	spc->spc_flags &= ~SPCF_NOINTR;
416 	splx(s);
417 }
418 
419 static void
420 cpu_xc_nointr(struct cpu_info *ci)
421 {
422 	struct schedstate_percpu *spc;
423 	int s;
424 
425 	spc = &ci->ci_schedstate;
426 	s = splsched();
427 	spc->spc_flags |= SPCF_NOINTR;
428 	splx(s);
429 }
430 
431 int
432 cpu_setintr(struct cpu_info *ci, bool intr)
433 {
434 	struct schedstate_percpu *spc;
435 	CPU_INFO_ITERATOR cii;
436 	struct cpu_info *ci2;
437 	uint64_t where;
438 	xcfunc_t func;
439 	int nintr;
440 
441 	spc = &ci->ci_schedstate;
442 
443 	KASSERT(mutex_owned(&cpu_lock));
444 
445 	if (intr) {
446 		if ((spc->spc_flags & SPCF_NOINTR) == 0)
447 			return 0;
448 		func = (xcfunc_t)cpu_xc_intr;
449 	} else {
450 		if ((spc->spc_flags & SPCF_NOINTR) != 0)
451 			return 0;
452 		/*
453 		 * Ensure that at least one CPU within the system
454 		 * is handing device interrupts.
455 		 */
456 		nintr = 0;
457 		for (CPU_INFO_FOREACH(cii, ci2)) {
458 			if ((ci2->ci_schedstate.spc_flags & SPCF_NOINTR) != 0)
459 				continue;
460 			if (ci2 == ci)
461 				continue;
462 			nintr++;
463 		}
464 		if (nintr == 0)
465 			return EBUSY;
466 		func = (xcfunc_t)cpu_xc_nointr;
467 	}
468 
469 	where = xc_unicast(0, func, ci, NULL, ci);
470 	xc_wait(where);
471 	if (intr) {
472 		KASSERT((spc->spc_flags & SPCF_NOINTR) == 0);
473 	} else if ((spc->spc_flags & SPCF_NOINTR) == 0) {
474 		/* If was not set offline, then it is busy */
475 		return EBUSY;
476 	}
477 
478 	/* Direct interrupts away from the CPU and record the change. */
479 	cpu_intr_redistribute();
480 	spc->spc_lastmod = time_second;
481 	return 0;
482 }
483 #else	/* __HAVE_INTR_CONTROL */
484 int
485 cpu_setintr(struct cpu_info *ci, bool intr)
486 {
487 
488 	return EOPNOTSUPP;
489 }
490 
491 u_int
492 cpu_intr_count(struct cpu_info *ci)
493 {
494 
495 	return 0;	/* 0 == "don't know" */
496 }
497 #endif	/* __HAVE_INTR_CONTROL */
498 
499 bool
500 cpu_softintr_p(void)
501 {
502 
503 	return (curlwp->l_pflag & LP_INTR) != 0;
504 }
505