xref: /freebsd/sys/x86/x86/cpu_machdep.c (revision 5dae51da3da0cc94d17bd67b308fad304ebec7e0)
1 /*-
2  * Copyright (c) 2003 Peter Wemm.
3  * Copyright (c) 1992 Terrence R. Lambert.
4  * Copyright (c) 1982, 1987, 1990 The Regents of the University of California.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to Berkeley by
8  * William Jolitz.
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  * 3. All advertising materials mentioning features or use of this software
19  *    must display the following acknowledgement:
20  *	This product includes software developed by the University of
21  *	California, Berkeley and its contributors.
22  * 4. Neither the name of the University nor the names of its contributors
23  *    may be used to endorse or promote products derived from this software
24  *    without specific prior written permission.
25  *
26  * THIS SOFTWARE IS PROVIDED BY THE REGENTS AND CONTRIBUTORS ``AS IS'' AND
27  * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
28  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
29  * ARE DISCLAIMED.  IN NO EVENT SHALL THE REGENTS OR CONTRIBUTORS BE LIABLE
30  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
31  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
32  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
33  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
34  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
35  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
36  * SUCH DAMAGE.
37  *
38  *	from: @(#)machdep.c	7.4 (Berkeley) 6/3/91
39  */
40 
41 #include <sys/cdefs.h>
42 __FBSDID("$FreeBSD$");
43 
44 #include "opt_atpic.h"
45 #include "opt_compat.h"
46 #include "opt_cpu.h"
47 #include "opt_ddb.h"
48 #include "opt_inet.h"
49 #include "opt_isa.h"
50 #include "opt_kdb.h"
51 #include "opt_kstack_pages.h"
52 #include "opt_maxmem.h"
53 #include "opt_mp_watchdog.h"
54 #include "opt_platform.h"
55 #ifdef __i386__
56 #include "opt_npx.h"
57 #include "opt_apic.h"
58 #include "opt_xbox.h"
59 #endif
60 
61 #include <sys/param.h>
62 #include <sys/proc.h>
63 #include <sys/systm.h>
64 #include <sys/bus.h>
65 #include <sys/cpu.h>
66 #include <sys/kdb.h>
67 #include <sys/kernel.h>
68 #include <sys/ktr.h>
69 #include <sys/lock.h>
70 #include <sys/malloc.h>
71 #include <sys/mutex.h>
72 #include <sys/pcpu.h>
73 #include <sys/rwlock.h>
74 #include <sys/sched.h>
75 #ifdef SMP
76 #include <sys/smp.h>
77 #endif
78 #include <sys/sysctl.h>
79 
80 #include <machine/clock.h>
81 #include <machine/cpu.h>
82 #include <machine/cputypes.h>
83 #include <machine/specialreg.h>
84 #include <machine/md_var.h>
85 #include <machine/mp_watchdog.h>
86 #include <machine/tss.h>
87 #ifdef SMP
88 #include <machine/smp.h>
89 #endif
90 #include <x86/acpica_machdep.h>
91 
92 #include <vm/vm.h>
93 #include <vm/vm_extern.h>
94 #include <vm/vm_kern.h>
95 #include <vm/vm_page.h>
96 #include <vm/vm_map.h>
97 #include <vm/vm_object.h>
98 #include <vm/vm_pager.h>
99 #include <vm/vm_param.h>
100 
101 #define	STATE_RUNNING	0x0
102 #define	STATE_MWAIT	0x1
103 #define	STATE_SLEEPING	0x2
104 
105 /*
106  * Machine dependent boot() routine
107  *
108  * I haven't seen anything to put here yet
109  * Possibly some stuff might be grafted back here from boot()
110  */
111 void
112 cpu_boot(int howto)
113 {
114 }
115 
116 /*
117  * Flush the D-cache for non-DMA I/O so that the I-cache can
118  * be made coherent later.
119  */
120 void
121 cpu_flush_dcache(void *ptr, size_t len)
122 {
123 	/* Not applicable */
124 }
125 
126 void
127 acpi_cpu_c1(void)
128 {
129 
130 	__asm __volatile("sti; hlt");
131 }
132 
133 void
134 acpi_cpu_idle_mwait(uint32_t mwait_hint)
135 {
136 	int *state;
137 
138 	/*
139 	 * XXXKIB.  Software coordination mode should be supported,
140 	 * but all Intel CPUs provide hardware coordination.
141 	 */
142 
143 	state = (int *)PCPU_PTR(monitorbuf);
144 	KASSERT(*state == STATE_SLEEPING,
145 		("cpu_mwait_cx: wrong monitorbuf state"));
146 	*state = STATE_MWAIT;
147 	cpu_monitor(state, 0, 0);
148 	if (*state == STATE_MWAIT)
149 		cpu_mwait(MWAIT_INTRBREAK, mwait_hint);
150 
151 	/*
152 	 * We should exit on any event that interrupts mwait, because
153 	 * that event might be a wanted interrupt.
154 	 */
155 	*state = STATE_RUNNING;
156 }
157 
158 /* Get current clock frequency for the given cpu id. */
159 int
160 cpu_est_clockrate(int cpu_id, uint64_t *rate)
161 {
162 	uint64_t tsc1, tsc2;
163 	uint64_t acnt, mcnt, perf;
164 	register_t reg;
165 
166 	if (pcpu_find(cpu_id) == NULL || rate == NULL)
167 		return (EINVAL);
168 #ifdef __i386__
169 	if ((cpu_feature & CPUID_TSC) == 0)
170 		return (EOPNOTSUPP);
171 #endif
172 
173 	/*
174 	 * If TSC is P-state invariant and APERF/MPERF MSRs do not exist,
175 	 * DELAY(9) based logic fails.
176 	 */
177 	if (tsc_is_invariant && !tsc_perf_stat)
178 		return (EOPNOTSUPP);
179 
180 #ifdef SMP
181 	if (smp_cpus > 1) {
182 		/* Schedule ourselves on the indicated cpu. */
183 		thread_lock(curthread);
184 		sched_bind(curthread, cpu_id);
185 		thread_unlock(curthread);
186 	}
187 #endif
188 
189 	/* Calibrate by measuring a short delay. */
190 	reg = intr_disable();
191 	if (tsc_is_invariant) {
192 		wrmsr(MSR_MPERF, 0);
193 		wrmsr(MSR_APERF, 0);
194 		tsc1 = rdtsc();
195 		DELAY(1000);
196 		mcnt = rdmsr(MSR_MPERF);
197 		acnt = rdmsr(MSR_APERF);
198 		tsc2 = rdtsc();
199 		intr_restore(reg);
200 		perf = 1000 * acnt / mcnt;
201 		*rate = (tsc2 - tsc1) * perf;
202 	} else {
203 		tsc1 = rdtsc();
204 		DELAY(1000);
205 		tsc2 = rdtsc();
206 		intr_restore(reg);
207 		*rate = (tsc2 - tsc1) * 1000;
208 	}
209 
210 #ifdef SMP
211 	if (smp_cpus > 1) {
212 		thread_lock(curthread);
213 		sched_unbind(curthread);
214 		thread_unlock(curthread);
215 	}
216 #endif
217 
218 	return (0);
219 }
220 
221 /*
222  * Shutdown the CPU as much as possible
223  */
224 void
225 cpu_halt(void)
226 {
227 	for (;;)
228 		halt();
229 }
230 
231 bool
232 cpu_mwait_usable(void)
233 {
234 
235 	return ((cpu_feature2 & CPUID2_MON) != 0 && ((cpu_mon_mwait_flags &
236 	    (CPUID5_MON_MWAIT_EXT | CPUID5_MWAIT_INTRBREAK)) ==
237 	    (CPUID5_MON_MWAIT_EXT | CPUID5_MWAIT_INTRBREAK)));
238 }
239 
240 void (*cpu_idle_hook)(sbintime_t) = NULL;	/* ACPI idle hook. */
241 static int	cpu_ident_amdc1e = 0;	/* AMD C1E supported. */
242 static int	idle_mwait = 1;		/* Use MONITOR/MWAIT for short idle. */
243 SYSCTL_INT(_machdep, OID_AUTO, idle_mwait, CTLFLAG_RWTUN, &idle_mwait,
244     0, "Use MONITOR/MWAIT for short idle");
245 
246 #ifndef PC98
247 static void
248 cpu_idle_acpi(sbintime_t sbt)
249 {
250 	int *state;
251 
252 	state = (int *)PCPU_PTR(monitorbuf);
253 	*state = STATE_SLEEPING;
254 
255 	/* See comments in cpu_idle_hlt(). */
256 	disable_intr();
257 	if (sched_runnable())
258 		enable_intr();
259 	else if (cpu_idle_hook)
260 		cpu_idle_hook(sbt);
261 	else
262 		acpi_cpu_c1();
263 	*state = STATE_RUNNING;
264 }
265 #endif /* !PC98 */
266 
267 static void
268 cpu_idle_hlt(sbintime_t sbt)
269 {
270 	int *state;
271 
272 	state = (int *)PCPU_PTR(monitorbuf);
273 	*state = STATE_SLEEPING;
274 
275 	/*
276 	 * Since we may be in a critical section from cpu_idle(), if
277 	 * an interrupt fires during that critical section we may have
278 	 * a pending preemption.  If the CPU halts, then that thread
279 	 * may not execute until a later interrupt awakens the CPU.
280 	 * To handle this race, check for a runnable thread after
281 	 * disabling interrupts and immediately return if one is
282 	 * found.  Also, we must absolutely guarentee that hlt is
283 	 * the next instruction after sti.  This ensures that any
284 	 * interrupt that fires after the call to disable_intr() will
285 	 * immediately awaken the CPU from hlt.  Finally, please note
286 	 * that on x86 this works fine because of interrupts enabled only
287 	 * after the instruction following sti takes place, while IF is set
288 	 * to 1 immediately, allowing hlt instruction to acknowledge the
289 	 * interrupt.
290 	 */
291 	disable_intr();
292 	if (sched_runnable())
293 		enable_intr();
294 	else
295 		acpi_cpu_c1();
296 	*state = STATE_RUNNING;
297 }
298 
299 static void
300 cpu_idle_mwait(sbintime_t sbt)
301 {
302 	int *state;
303 
304 	state = (int *)PCPU_PTR(monitorbuf);
305 	*state = STATE_MWAIT;
306 
307 	/* See comments in cpu_idle_hlt(). */
308 	disable_intr();
309 	if (sched_runnable()) {
310 		enable_intr();
311 		*state = STATE_RUNNING;
312 		return;
313 	}
314 	cpu_monitor(state, 0, 0);
315 	if (*state == STATE_MWAIT)
316 		__asm __volatile("sti; mwait" : : "a" (MWAIT_C1), "c" (0));
317 	else
318 		enable_intr();
319 	*state = STATE_RUNNING;
320 }
321 
322 static void
323 cpu_idle_spin(sbintime_t sbt)
324 {
325 	int *state;
326 	int i;
327 
328 	state = (int *)PCPU_PTR(monitorbuf);
329 	*state = STATE_RUNNING;
330 
331 	/*
332 	 * The sched_runnable() call is racy but as long as there is
333 	 * a loop missing it one time will have just a little impact if any
334 	 * (and it is much better than missing the check at all).
335 	 */
336 	for (i = 0; i < 1000; i++) {
337 		if (sched_runnable())
338 			return;
339 		cpu_spinwait();
340 	}
341 }
342 
343 /*
344  * C1E renders the local APIC timer dead, so we disable it by
345  * reading the Interrupt Pending Message register and clearing
346  * both C1eOnCmpHalt (bit 28) and SmiOnCmpHalt (bit 27).
347  *
348  * Reference:
349  *   "BIOS and Kernel Developer's Guide for AMD NPT Family 0Fh Processors"
350  *   #32559 revision 3.00+
351  */
352 #define	MSR_AMDK8_IPM		0xc0010055
353 #define	AMDK8_SMIONCMPHALT	(1ULL << 27)
354 #define	AMDK8_C1EONCMPHALT	(1ULL << 28)
355 #define	AMDK8_CMPHALT		(AMDK8_SMIONCMPHALT | AMDK8_C1EONCMPHALT)
356 
357 void
358 cpu_probe_amdc1e(void)
359 {
360 
361 	/*
362 	 * Detect the presence of C1E capability mostly on latest
363 	 * dual-cores (or future) k8 family.
364 	 */
365 	if (cpu_vendor_id == CPU_VENDOR_AMD &&
366 	    (cpu_id & 0x00000f00) == 0x00000f00 &&
367 	    (cpu_id & 0x0fff0000) >=  0x00040000) {
368 		cpu_ident_amdc1e = 1;
369 	}
370 }
371 
372 #if defined(__i386__) && defined(PC98)
373 void (*cpu_idle_fn)(sbintime_t) = cpu_idle_hlt;
374 #else
375 void (*cpu_idle_fn)(sbintime_t) = cpu_idle_acpi;
376 #endif
377 
378 void
379 cpu_idle(int busy)
380 {
381 	uint64_t msr;
382 	sbintime_t sbt = -1;
383 
384 	CTR2(KTR_SPARE2, "cpu_idle(%d) at %d",
385 	    busy, curcpu);
386 #ifdef MP_WATCHDOG
387 	ap_watchdog(PCPU_GET(cpuid));
388 #endif
389 
390 	/* If we are busy - try to use fast methods. */
391 	if (busy) {
392 		if ((cpu_feature2 & CPUID2_MON) && idle_mwait) {
393 			cpu_idle_mwait(busy);
394 			goto out;
395 		}
396 	}
397 
398 	/* If we have time - switch timers into idle mode. */
399 	if (!busy) {
400 		critical_enter();
401 		sbt = cpu_idleclock();
402 	}
403 
404 	/* Apply AMD APIC timer C1E workaround. */
405 	if (cpu_ident_amdc1e && cpu_disable_c3_sleep) {
406 		msr = rdmsr(MSR_AMDK8_IPM);
407 		if (msr & AMDK8_CMPHALT)
408 			wrmsr(MSR_AMDK8_IPM, msr & ~AMDK8_CMPHALT);
409 	}
410 
411 	/* Call main idle method. */
412 	cpu_idle_fn(sbt);
413 
414 	/* Switch timers back into active mode. */
415 	if (!busy) {
416 		cpu_activeclock();
417 		critical_exit();
418 	}
419 out:
420 	CTR2(KTR_SPARE2, "cpu_idle(%d) at %d done",
421 	    busy, curcpu);
422 }
423 
424 int
425 cpu_idle_wakeup(int cpu)
426 {
427 	struct pcpu *pcpu;
428 	int *state;
429 
430 	pcpu = pcpu_find(cpu);
431 	state = (int *)pcpu->pc_monitorbuf;
432 	/*
433 	 * This doesn't need to be atomic since missing the race will
434 	 * simply result in unnecessary IPIs.
435 	 */
436 	if (*state == STATE_SLEEPING)
437 		return (0);
438 	if (*state == STATE_MWAIT)
439 		*state = STATE_RUNNING;
440 	return (1);
441 }
442 
443 /*
444  * Ordered by speed/power consumption.
445  */
446 struct {
447 	void	*id_fn;
448 	char	*id_name;
449 } idle_tbl[] = {
450 	{ cpu_idle_spin, "spin" },
451 	{ cpu_idle_mwait, "mwait" },
452 	{ cpu_idle_hlt, "hlt" },
453 #if !defined(__i386__) || !defined(PC98)
454 	{ cpu_idle_acpi, "acpi" },
455 #endif
456 	{ NULL, NULL }
457 };
458 
459 static int
460 idle_sysctl_available(SYSCTL_HANDLER_ARGS)
461 {
462 	char *avail, *p;
463 	int error;
464 	int i;
465 
466 	avail = malloc(256, M_TEMP, M_WAITOK);
467 	p = avail;
468 	for (i = 0; idle_tbl[i].id_name != NULL; i++) {
469 		if (strstr(idle_tbl[i].id_name, "mwait") &&
470 		    (cpu_feature2 & CPUID2_MON) == 0)
471 			continue;
472 #if !defined(__i386__) || !defined(PC98)
473 		if (strcmp(idle_tbl[i].id_name, "acpi") == 0 &&
474 		    cpu_idle_hook == NULL)
475 			continue;
476 #endif
477 		p += sprintf(p, "%s%s", p != avail ? ", " : "",
478 		    idle_tbl[i].id_name);
479 	}
480 	error = sysctl_handle_string(oidp, avail, 0, req);
481 	free(avail, M_TEMP);
482 	return (error);
483 }
484 
485 SYSCTL_PROC(_machdep, OID_AUTO, idle_available, CTLTYPE_STRING | CTLFLAG_RD,
486     0, 0, idle_sysctl_available, "A", "list of available idle functions");
487 
488 static int
489 idle_sysctl(SYSCTL_HANDLER_ARGS)
490 {
491 	char buf[16];
492 	int error;
493 	char *p;
494 	int i;
495 
496 	p = "unknown";
497 	for (i = 0; idle_tbl[i].id_name != NULL; i++) {
498 		if (idle_tbl[i].id_fn == cpu_idle_fn) {
499 			p = idle_tbl[i].id_name;
500 			break;
501 		}
502 	}
503 	strncpy(buf, p, sizeof(buf));
504 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
505 	if (error != 0 || req->newptr == NULL)
506 		return (error);
507 	for (i = 0; idle_tbl[i].id_name != NULL; i++) {
508 		if (strstr(idle_tbl[i].id_name, "mwait") &&
509 		    (cpu_feature2 & CPUID2_MON) == 0)
510 			continue;
511 #if !defined(__i386__) || !defined(PC98)
512 		if (strcmp(idle_tbl[i].id_name, "acpi") == 0 &&
513 		    cpu_idle_hook == NULL)
514 			continue;
515 #endif
516 		if (strcmp(idle_tbl[i].id_name, buf))
517 			continue;
518 		cpu_idle_fn = idle_tbl[i].id_fn;
519 		return (0);
520 	}
521 	return (EINVAL);
522 }
523 
524 SYSCTL_PROC(_machdep, OID_AUTO, idle, CTLTYPE_STRING | CTLFLAG_RW, 0, 0,
525     idle_sysctl, "A", "currently selected idle function");
526 
527 static int panic_on_nmi = 1;
528 SYSCTL_INT(_machdep, OID_AUTO, panic_on_nmi, CTLFLAG_RWTUN,
529     &panic_on_nmi, 0,
530     "Panic on NMI");
531 int nmi_is_broadcast = 1;
532 SYSCTL_INT(_machdep, OID_AUTO, nmi_is_broadcast, CTLFLAG_RWTUN,
533     &nmi_is_broadcast, 0,
534     "Chipset NMI is broadcast");
535 #ifdef KDB
536 int kdb_on_nmi = 1;
537 SYSCTL_INT(_machdep, OID_AUTO, kdb_on_nmi, CTLFLAG_RWTUN,
538     &kdb_on_nmi, 0,
539     "Go to KDB on NMI");
540 #endif
541 
542 #ifdef DEV_ISA
543 void
544 nmi_call_kdb(u_int cpu, u_int type, struct trapframe *frame)
545 {
546 
547 	/* machine/parity/power fail/"kitchen sink" faults */
548 	if (isa_nmi(frame->tf_err) == 0) {
549 #ifdef KDB
550 		/*
551 		 * NMI can be hooked up to a pushbutton for debugging.
552 		 */
553 		if (kdb_on_nmi) {
554 			printf("NMI/cpu%d ... going to debugger\n", cpu);
555 			kdb_trap(type, 0, frame);
556 		}
557 #endif /* KDB */
558 	} else if (panic_on_nmi) {
559 		panic("NMI indicates hardware failure");
560 	}
561 }
562 #endif
563 
564 void
565 nmi_handle_intr(u_int type, struct trapframe *frame)
566 {
567 
568 #ifdef DEV_ISA
569 #ifdef SMP
570 	if (nmi_is_broadcast) {
571 		nmi_call_kdb_smp(type, frame);
572 		return;
573 	}
574 #endif
575 	nmi_call_kdb(PCPU_GET(cpuid), type, frame);
576 #endif
577 }
578