xref: /netbsd-src/sys/kern/kern_cpu.c (revision 9aa0541bdf64142d9a27c2cf274394d60182818f)
1 /*	$NetBSD: kern_cpu.c,v 1.51 2011/09/11 14:54:49 jdc 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 1.51 2011/09/11 14:54:49 jdc 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 
111 kcpuset_t *	kcpuset_attached	__read_mostly;
112 
113 struct cpuqueue	cpu_queue		__cacheline_aligned
114     = CIRCLEQ_HEAD_INITIALIZER(cpu_queue);
115 
116 static struct cpu_info **cpu_infos	__read_mostly;
117 
118 int
119 mi_cpu_attach(struct cpu_info *ci)
120 {
121 	int error;
122 
123 	KASSERT(maxcpus > 0);
124 
125 	ci->ci_index = ncpu;
126 	CIRCLEQ_INSERT_TAIL(&cpu_queue, ci, ci_data.cpu_qchain);
127 	TAILQ_INIT(&ci->ci_data.cpu_ld_locks);
128 	__cpu_simple_lock_init(&ci->ci_data.cpu_ld_lock);
129 
130 	/* This is useful for eg, per-cpu evcnt */
131 	snprintf(ci->ci_data.cpu_name, sizeof(ci->ci_data.cpu_name), "cpu%d",
132 	    cpu_index(ci));
133 
134 	if (__predict_false(cpu_infos == NULL)) {
135 		cpu_infos =
136 		    kmem_zalloc(sizeof(cpu_infos[0]) * maxcpus, KM_SLEEP);
137 		kcpuset_create(&kcpuset_attached, true);
138 	}
139 	cpu_infos[cpu_index(ci)] = ci;
140 	kcpuset_set(kcpuset_attached, ci->ci_index);
141 
142 	sched_cpuattach(ci);
143 
144 	error = create_idle_lwp(ci);
145 	if (error != 0) {
146 		/* XXX revert sched_cpuattach */
147 		return error;
148 	}
149 
150 	if (ci == curcpu())
151 		ci->ci_data.cpu_onproc = curlwp;
152 	else
153 		ci->ci_data.cpu_onproc = ci->ci_data.cpu_idlelwp;
154 
155 	percpu_init_cpu(ci);
156 	softint_init(ci);
157 	callout_init_cpu(ci);
158 	xc_init_cpu(ci);
159 	pool_cache_cpu_init(ci);
160 	selsysinit(ci);
161 	cache_cpu_init(ci);
162 	TAILQ_INIT(&ci->ci_data.cpu_biodone);
163 	ncpu++;
164 	ncpuonline++;
165 
166 	return 0;
167 }
168 
169 void
170 cpuctlattach(int dummy)
171 {
172 
173 	KASSERT(cpu_infos != NULL);
174 }
175 
176 int
177 cpuctl_ioctl(dev_t dev, u_long cmd, void *data, int flag, lwp_t *l)
178 {
179 	CPU_INFO_ITERATOR cii;
180 	cpustate_t *cs;
181 	struct cpu_info *ci;
182 	int error, i;
183 	u_int id;
184 
185 	error = 0;
186 
187 	mutex_enter(&cpu_lock);
188 	switch (cmd) {
189 	case IOC_CPU_SETSTATE:
190 		if (error == 0)
191 			cs = data;
192 		error = kauth_authorize_system(l->l_cred,
193 		    KAUTH_SYSTEM_CPU, KAUTH_REQ_SYSTEM_CPU_SETSTATE, cs, NULL,
194 		    NULL);
195 		if (error != 0)
196 			break;
197 		if (cs->cs_id >= maxcpus ||
198 		    (ci = cpu_lookup(cs->cs_id)) == NULL) {
199 			error = ESRCH;
200 			break;
201 		}
202 		error = cpu_setintr(ci, cs->cs_intr);
203 		error = cpu_setstate(ci, cs->cs_online);
204 		break;
205 
206 	case IOC_CPU_GETSTATE:
207 		if (error == 0)
208 			cs = data;
209 		id = cs->cs_id;
210 		memset(cs, 0, sizeof(*cs));
211 		cs->cs_id = id;
212 		if (cs->cs_id >= maxcpus ||
213 		    (ci = cpu_lookup(id)) == NULL) {
214 			error = ESRCH;
215 			break;
216 		}
217 		if ((ci->ci_schedstate.spc_flags & SPCF_OFFLINE) != 0)
218 			cs->cs_online = false;
219 		else
220 			cs->cs_online = true;
221 		if ((ci->ci_schedstate.spc_flags & SPCF_NOINTR) != 0)
222 			cs->cs_intr = false;
223 		else
224 			cs->cs_intr = true;
225 		cs->cs_lastmod = (int32_t)ci->ci_schedstate.spc_lastmod;
226 		cs->cs_lastmodhi = (int32_t)
227 		    (ci->ci_schedstate.spc_lastmod >> 32);
228 		cs->cs_intrcnt = cpu_intr_count(ci) + 1;
229 		cs->cs_hwid = ci->ci_cpuid;
230 		break;
231 
232 	case IOC_CPU_MAPID:
233 		i = 0;
234 		for (CPU_INFO_FOREACH(cii, ci)) {
235 			if (i++ == *(int *)data)
236 				break;
237 		}
238 		if (ci == NULL)
239 			error = ESRCH;
240 		else
241 			*(int *)data = cpu_index(ci);
242 		break;
243 
244 	case IOC_CPU_GETCOUNT:
245 		*(int *)data = ncpu;
246 		break;
247 
248 	default:
249 		error = ENOTTY;
250 		break;
251 	}
252 	mutex_exit(&cpu_lock);
253 
254 	return error;
255 }
256 
257 struct cpu_info *
258 cpu_lookup(u_int idx)
259 {
260 	struct cpu_info *ci;
261 
262 	KASSERT(idx < maxcpus);
263 
264 	if (__predict_false(cpu_infos == NULL)) {
265 		KASSERT(idx == 0);
266 		return curcpu();
267 	}
268 
269 	ci = cpu_infos[idx];
270 	KASSERT(ci == NULL || cpu_index(ci) == idx);
271 
272 	return ci;
273 }
274 
275 static void
276 cpu_xc_offline(struct cpu_info *ci)
277 {
278 	struct schedstate_percpu *spc, *mspc = NULL;
279 	struct cpu_info *target_ci;
280 	struct lwp *l;
281 	CPU_INFO_ITERATOR cii;
282 	int s;
283 
284 	/*
285 	 * Thread that made the cross call (separate context) holds
286 	 * cpu_lock on our behalf.
287 	 */
288 	spc = &ci->ci_schedstate;
289 	s = splsched();
290 	spc->spc_flags |= SPCF_OFFLINE;
291 	splx(s);
292 
293 	/* Take the first available CPU for the migration. */
294 	for (CPU_INFO_FOREACH(cii, target_ci)) {
295 		mspc = &target_ci->ci_schedstate;
296 		if ((mspc->spc_flags & SPCF_OFFLINE) == 0)
297 			break;
298 	}
299 	KASSERT(target_ci != NULL);
300 
301 	/*
302 	 * Migrate all non-bound threads to the other CPU.  Note that this
303 	 * runs from the xcall thread, thus handling of LSONPROC is not needed.
304 	 */
305 	mutex_enter(proc_lock);
306 	LIST_FOREACH(l, &alllwp, l_list) {
307 		struct cpu_info *mci;
308 
309 		lwp_lock(l);
310 		if (l->l_cpu != ci || (l->l_pflag & (LP_BOUND | LP_INTR))) {
311 			lwp_unlock(l);
312 			continue;
313 		}
314 		/* Regular case - no affinity. */
315 		if (l->l_affinity == NULL) {
316 			lwp_migrate(l, target_ci);
317 			continue;
318 		}
319 		/* Affinity is set, find an online CPU in the set. */
320 		for (CPU_INFO_FOREACH(cii, mci)) {
321 			mspc = &mci->ci_schedstate;
322 			if ((mspc->spc_flags & SPCF_OFFLINE) == 0 &&
323 			    kcpuset_isset(l->l_affinity, cpu_index(mci)))
324 				break;
325 		}
326 		if (mci == NULL) {
327 			lwp_unlock(l);
328 			mutex_exit(proc_lock);
329 			goto fail;
330 		}
331 		lwp_migrate(l, mci);
332 	}
333 	mutex_exit(proc_lock);
334 
335 #ifdef __HAVE_MD_CPU_OFFLINE
336 	cpu_offline_md();
337 #endif
338 	return;
339 fail:
340 	/* Just unset the SPCF_OFFLINE flag, caller will check */
341 	s = splsched();
342 	spc->spc_flags &= ~SPCF_OFFLINE;
343 	splx(s);
344 }
345 
346 static void
347 cpu_xc_online(struct cpu_info *ci)
348 {
349 	struct schedstate_percpu *spc;
350 	int s;
351 
352 	spc = &ci->ci_schedstate;
353 	s = splsched();
354 	spc->spc_flags &= ~SPCF_OFFLINE;
355 	splx(s);
356 }
357 
358 int
359 cpu_setstate(struct cpu_info *ci, bool online)
360 {
361 	struct schedstate_percpu *spc;
362 	CPU_INFO_ITERATOR cii;
363 	struct cpu_info *ci2;
364 	uint64_t where;
365 	xcfunc_t func;
366 	int nonline;
367 
368 	spc = &ci->ci_schedstate;
369 
370 	KASSERT(mutex_owned(&cpu_lock));
371 
372 	if (online) {
373 		if ((spc->spc_flags & SPCF_OFFLINE) == 0)
374 			return 0;
375 		func = (xcfunc_t)cpu_xc_online;
376 		ncpuonline++;
377 	} else {
378 		if ((spc->spc_flags & SPCF_OFFLINE) != 0)
379 			return 0;
380 		nonline = 0;
381 		/*
382 		 * Ensure that at least one CPU within the processor set
383 		 * stays online.  Revisit this later.
384 		 */
385 		for (CPU_INFO_FOREACH(cii, ci2)) {
386 			if ((ci2->ci_schedstate.spc_flags & SPCF_OFFLINE) != 0)
387 				continue;
388 			if (ci2->ci_schedstate.spc_psid != spc->spc_psid)
389 				continue;
390 			nonline++;
391 		}
392 		if (nonline == 1)
393 			return EBUSY;
394 		func = (xcfunc_t)cpu_xc_offline;
395 		ncpuonline--;
396 	}
397 
398 	where = xc_unicast(0, func, ci, NULL, ci);
399 	xc_wait(where);
400 	if (online) {
401 		KASSERT((spc->spc_flags & SPCF_OFFLINE) == 0);
402 	} else if ((spc->spc_flags & SPCF_OFFLINE) == 0) {
403 		/* If was not set offline, then it is busy */
404 		return EBUSY;
405 	}
406 
407 	spc->spc_lastmod = time_second;
408 	return 0;
409 }
410 
411 #ifdef __HAVE_INTR_CONTROL
412 static void
413 cpu_xc_intr(struct cpu_info *ci)
414 {
415 	struct schedstate_percpu *spc;
416 	int s;
417 
418 	spc = &ci->ci_schedstate;
419 	s = splsched();
420 	spc->spc_flags &= ~SPCF_NOINTR;
421 	splx(s);
422 }
423 
424 static void
425 cpu_xc_nointr(struct cpu_info *ci)
426 {
427 	struct schedstate_percpu *spc;
428 	int s;
429 
430 	spc = &ci->ci_schedstate;
431 	s = splsched();
432 	spc->spc_flags |= SPCF_NOINTR;
433 	splx(s);
434 }
435 
436 int
437 cpu_setintr(struct cpu_info *ci, bool intr)
438 {
439 	struct schedstate_percpu *spc;
440 	CPU_INFO_ITERATOR cii;
441 	struct cpu_info *ci2;
442 	uint64_t where;
443 	xcfunc_t func;
444 	int nintr;
445 
446 	spc = &ci->ci_schedstate;
447 
448 	KASSERT(mutex_owned(&cpu_lock));
449 
450 	if (intr) {
451 		if ((spc->spc_flags & SPCF_NOINTR) == 0)
452 			return 0;
453 		func = (xcfunc_t)cpu_xc_intr;
454 	} else {
455 		if ((spc->spc_flags & SPCF_NOINTR) != 0)
456 			return 0;
457 		/*
458 		 * Ensure that at least one CPU within the system
459 		 * is handing device interrupts.
460 		 */
461 		nintr = 0;
462 		for (CPU_INFO_FOREACH(cii, ci2)) {
463 			if ((ci2->ci_schedstate.spc_flags & SPCF_NOINTR) != 0)
464 				continue;
465 			if (ci2 == ci)
466 				continue;
467 			nintr++;
468 		}
469 		if (nintr == 0)
470 			return EBUSY;
471 		func = (xcfunc_t)cpu_xc_nointr;
472 	}
473 
474 	where = xc_unicast(0, func, ci, NULL, ci);
475 	xc_wait(where);
476 	if (intr) {
477 		KASSERT((spc->spc_flags & SPCF_NOINTR) == 0);
478 	} else if ((spc->spc_flags & SPCF_NOINTR) == 0) {
479 		/* If was not set offline, then it is busy */
480 		return EBUSY;
481 	}
482 
483 	/* Direct interrupts away from the CPU and record the change. */
484 	cpu_intr_redistribute();
485 	spc->spc_lastmod = time_second;
486 	return 0;
487 }
488 #else	/* __HAVE_INTR_CONTROL */
489 int
490 cpu_setintr(struct cpu_info *ci, bool intr)
491 {
492 
493 	return EOPNOTSUPP;
494 }
495 
496 u_int
497 cpu_intr_count(struct cpu_info *ci)
498 {
499 
500 	return 0;	/* 0 == "don't know" */
501 }
502 #endif	/* __HAVE_INTR_CONTROL */
503 
504 bool
505 cpu_softintr_p(void)
506 {
507 
508 	return (curlwp->l_pflag & LP_INTR) != 0;
509 }
510