xref: /freebsd/sys/x86/x86/cpu_machdep.c (revision 5bf5ca772c6de2d53344a78cf461447cc322ccea)
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_apic.h"
57 #endif
58 
59 #include <sys/param.h>
60 #include <sys/proc.h>
61 #include <sys/systm.h>
62 #include <sys/bus.h>
63 #include <sys/cpu.h>
64 #include <sys/kdb.h>
65 #include <sys/kernel.h>
66 #include <sys/ktr.h>
67 #include <sys/lock.h>
68 #include <sys/malloc.h>
69 #include <sys/mutex.h>
70 #include <sys/pcpu.h>
71 #include <sys/rwlock.h>
72 #include <sys/sched.h>
73 #ifdef SMP
74 #include <sys/smp.h>
75 #endif
76 #include <sys/sysctl.h>
77 
78 #include <machine/clock.h>
79 #include <machine/cpu.h>
80 #include <machine/cputypes.h>
81 #include <machine/specialreg.h>
82 #include <machine/md_var.h>
83 #include <machine/mp_watchdog.h>
84 #include <machine/tss.h>
85 #ifdef SMP
86 #include <machine/smp.h>
87 #endif
88 #include <x86/acpica_machdep.h>
89 
90 #include <vm/vm.h>
91 #include <vm/vm_extern.h>
92 #include <vm/vm_kern.h>
93 #include <vm/vm_page.h>
94 #include <vm/vm_map.h>
95 #include <vm/vm_object.h>
96 #include <vm/vm_pager.h>
97 #include <vm/vm_param.h>
98 
99 #define	STATE_RUNNING	0x0
100 #define	STATE_MWAIT	0x1
101 #define	STATE_SLEEPING	0x2
102 
103 /*
104  * Machine dependent boot() routine
105  *
106  * I haven't seen anything to put here yet
107  * Possibly some stuff might be grafted back here from boot()
108  */
109 void
110 cpu_boot(int howto)
111 {
112 }
113 
114 /*
115  * Flush the D-cache for non-DMA I/O so that the I-cache can
116  * be made coherent later.
117  */
118 void
119 cpu_flush_dcache(void *ptr, size_t len)
120 {
121 	/* Not applicable */
122 }
123 
124 void
125 acpi_cpu_c1(void)
126 {
127 
128 	__asm __volatile("sti; hlt");
129 }
130 
131 /*
132  * Use mwait to pause execution while waiting for an interrupt or
133  * another thread to signal that there is more work.
134  *
135  * NOTE: Interrupts will cause a wakeup; however, this function does
136  * not enable interrupt handling. The caller is responsible to enable
137  * interrupts.
138  */
139 void
140 acpi_cpu_idle_mwait(uint32_t mwait_hint)
141 {
142 	int *state;
143 
144 	/*
145 	 * A comment in Linux patch claims that 'CPUs run faster with
146 	 * speculation protection disabled. All CPU threads in a core
147 	 * must disable speculation protection for it to be
148 	 * disabled. Disable it while we are idle so the other
149 	 * hyperthread can run fast.'
150 	 *
151 	 * XXXKIB.  Software coordination mode should be supported,
152 	 * but all Intel CPUs provide hardware coordination.
153 	 */
154 
155 	state = (int *)PCPU_PTR(monitorbuf);
156 	KASSERT(*state == STATE_SLEEPING,
157 		("cpu_mwait_cx: wrong monitorbuf state"));
158 	*state = STATE_MWAIT;
159 	handle_ibrs_entry();
160 	cpu_monitor(state, 0, 0);
161 	if (*state == STATE_MWAIT)
162 		cpu_mwait(MWAIT_INTRBREAK, mwait_hint);
163 	handle_ibrs_exit();
164 
165 	/*
166 	 * We should exit on any event that interrupts mwait, because
167 	 * that event might be a wanted interrupt.
168 	 */
169 	*state = STATE_RUNNING;
170 }
171 
172 /* Get current clock frequency for the given cpu id. */
173 int
174 cpu_est_clockrate(int cpu_id, uint64_t *rate)
175 {
176 	uint64_t tsc1, tsc2;
177 	uint64_t acnt, mcnt, perf;
178 	register_t reg;
179 
180 	if (pcpu_find(cpu_id) == NULL || rate == NULL)
181 		return (EINVAL);
182 #ifdef __i386__
183 	if ((cpu_feature & CPUID_TSC) == 0)
184 		return (EOPNOTSUPP);
185 #endif
186 
187 	/*
188 	 * If TSC is P-state invariant and APERF/MPERF MSRs do not exist,
189 	 * DELAY(9) based logic fails.
190 	 */
191 	if (tsc_is_invariant && !tsc_perf_stat)
192 		return (EOPNOTSUPP);
193 
194 #ifdef SMP
195 	if (smp_cpus > 1) {
196 		/* Schedule ourselves on the indicated cpu. */
197 		thread_lock(curthread);
198 		sched_bind(curthread, cpu_id);
199 		thread_unlock(curthread);
200 	}
201 #endif
202 
203 	/* Calibrate by measuring a short delay. */
204 	reg = intr_disable();
205 	if (tsc_is_invariant) {
206 		wrmsr(MSR_MPERF, 0);
207 		wrmsr(MSR_APERF, 0);
208 		tsc1 = rdtsc();
209 		DELAY(1000);
210 		mcnt = rdmsr(MSR_MPERF);
211 		acnt = rdmsr(MSR_APERF);
212 		tsc2 = rdtsc();
213 		intr_restore(reg);
214 		perf = 1000 * acnt / mcnt;
215 		*rate = (tsc2 - tsc1) * perf;
216 	} else {
217 		tsc1 = rdtsc();
218 		DELAY(1000);
219 		tsc2 = rdtsc();
220 		intr_restore(reg);
221 		*rate = (tsc2 - tsc1) * 1000;
222 	}
223 
224 #ifdef SMP
225 	if (smp_cpus > 1) {
226 		thread_lock(curthread);
227 		sched_unbind(curthread);
228 		thread_unlock(curthread);
229 	}
230 #endif
231 
232 	return (0);
233 }
234 
235 /*
236  * Shutdown the CPU as much as possible
237  */
238 void
239 cpu_halt(void)
240 {
241 	for (;;)
242 		halt();
243 }
244 
245 bool
246 cpu_mwait_usable(void)
247 {
248 
249 	return ((cpu_feature2 & CPUID2_MON) != 0 && ((cpu_mon_mwait_flags &
250 	    (CPUID5_MON_MWAIT_EXT | CPUID5_MWAIT_INTRBREAK)) ==
251 	    (CPUID5_MON_MWAIT_EXT | CPUID5_MWAIT_INTRBREAK)));
252 }
253 
254 void (*cpu_idle_hook)(sbintime_t) = NULL;	/* ACPI idle hook. */
255 static int	cpu_ident_amdc1e = 0;	/* AMD C1E supported. */
256 static int	idle_mwait = 1;		/* Use MONITOR/MWAIT for short idle. */
257 SYSCTL_INT(_machdep, OID_AUTO, idle_mwait, CTLFLAG_RWTUN, &idle_mwait,
258     0, "Use MONITOR/MWAIT for short idle");
259 
260 static void
261 cpu_idle_acpi(sbintime_t sbt)
262 {
263 	int *state;
264 
265 	state = (int *)PCPU_PTR(monitorbuf);
266 	*state = STATE_SLEEPING;
267 
268 	/* See comments in cpu_idle_hlt(). */
269 	disable_intr();
270 	if (sched_runnable())
271 		enable_intr();
272 	else if (cpu_idle_hook)
273 		cpu_idle_hook(sbt);
274 	else
275 		acpi_cpu_c1();
276 	*state = STATE_RUNNING;
277 }
278 
279 static void
280 cpu_idle_hlt(sbintime_t sbt)
281 {
282 	int *state;
283 
284 	state = (int *)PCPU_PTR(monitorbuf);
285 	*state = STATE_SLEEPING;
286 
287 	/*
288 	 * Since we may be in a critical section from cpu_idle(), if
289 	 * an interrupt fires during that critical section we may have
290 	 * a pending preemption.  If the CPU halts, then that thread
291 	 * may not execute until a later interrupt awakens the CPU.
292 	 * To handle this race, check for a runnable thread after
293 	 * disabling interrupts and immediately return if one is
294 	 * found.  Also, we must absolutely guarentee that hlt is
295 	 * the next instruction after sti.  This ensures that any
296 	 * interrupt that fires after the call to disable_intr() will
297 	 * immediately awaken the CPU from hlt.  Finally, please note
298 	 * that on x86 this works fine because of interrupts enabled only
299 	 * after the instruction following sti takes place, while IF is set
300 	 * to 1 immediately, allowing hlt instruction to acknowledge the
301 	 * interrupt.
302 	 */
303 	disable_intr();
304 	if (sched_runnable())
305 		enable_intr();
306 	else
307 		acpi_cpu_c1();
308 	*state = STATE_RUNNING;
309 }
310 
311 static void
312 cpu_idle_mwait(sbintime_t sbt)
313 {
314 	int *state;
315 
316 	state = (int *)PCPU_PTR(monitorbuf);
317 	*state = STATE_MWAIT;
318 
319 	/* See comments in cpu_idle_hlt(). */
320 	disable_intr();
321 	if (sched_runnable()) {
322 		enable_intr();
323 		*state = STATE_RUNNING;
324 		return;
325 	}
326 	cpu_monitor(state, 0, 0);
327 	if (*state == STATE_MWAIT)
328 		__asm __volatile("sti; mwait" : : "a" (MWAIT_C1), "c" (0));
329 	else
330 		enable_intr();
331 	*state = STATE_RUNNING;
332 }
333 
334 static void
335 cpu_idle_spin(sbintime_t sbt)
336 {
337 	int *state;
338 	int i;
339 
340 	state = (int *)PCPU_PTR(monitorbuf);
341 	*state = STATE_RUNNING;
342 
343 	/*
344 	 * The sched_runnable() call is racy but as long as there is
345 	 * a loop missing it one time will have just a little impact if any
346 	 * (and it is much better than missing the check at all).
347 	 */
348 	for (i = 0; i < 1000; i++) {
349 		if (sched_runnable())
350 			return;
351 		cpu_spinwait();
352 	}
353 }
354 
355 /*
356  * C1E renders the local APIC timer dead, so we disable it by
357  * reading the Interrupt Pending Message register and clearing
358  * both C1eOnCmpHalt (bit 28) and SmiOnCmpHalt (bit 27).
359  *
360  * Reference:
361  *   "BIOS and Kernel Developer's Guide for AMD NPT Family 0Fh Processors"
362  *   #32559 revision 3.00+
363  */
364 #define	MSR_AMDK8_IPM		0xc0010055
365 #define	AMDK8_SMIONCMPHALT	(1ULL << 27)
366 #define	AMDK8_C1EONCMPHALT	(1ULL << 28)
367 #define	AMDK8_CMPHALT		(AMDK8_SMIONCMPHALT | AMDK8_C1EONCMPHALT)
368 
369 void
370 cpu_probe_amdc1e(void)
371 {
372 
373 	/*
374 	 * Detect the presence of C1E capability mostly on latest
375 	 * dual-cores (or future) k8 family.
376 	 */
377 	if (cpu_vendor_id == CPU_VENDOR_AMD &&
378 	    (cpu_id & 0x00000f00) == 0x00000f00 &&
379 	    (cpu_id & 0x0fff0000) >=  0x00040000) {
380 		cpu_ident_amdc1e = 1;
381 	}
382 }
383 
384 void (*cpu_idle_fn)(sbintime_t) = cpu_idle_acpi;
385 
386 void
387 cpu_idle(int busy)
388 {
389 	uint64_t msr;
390 	sbintime_t sbt = -1;
391 
392 	CTR2(KTR_SPARE2, "cpu_idle(%d) at %d",
393 	    busy, curcpu);
394 #ifdef MP_WATCHDOG
395 	ap_watchdog(PCPU_GET(cpuid));
396 #endif
397 
398 	/* If we are busy - try to use fast methods. */
399 	if (busy) {
400 		if ((cpu_feature2 & CPUID2_MON) && idle_mwait) {
401 			cpu_idle_mwait(busy);
402 			goto out;
403 		}
404 	}
405 
406 	/* If we have time - switch timers into idle mode. */
407 	if (!busy) {
408 		critical_enter();
409 		sbt = cpu_idleclock();
410 	}
411 
412 	/* Apply AMD APIC timer C1E workaround. */
413 	if (cpu_ident_amdc1e && cpu_disable_c3_sleep) {
414 		msr = rdmsr(MSR_AMDK8_IPM);
415 		if (msr & AMDK8_CMPHALT)
416 			wrmsr(MSR_AMDK8_IPM, msr & ~AMDK8_CMPHALT);
417 	}
418 
419 	/* Call main idle method. */
420 	cpu_idle_fn(sbt);
421 
422 	/* Switch timers back into active mode. */
423 	if (!busy) {
424 		cpu_activeclock();
425 		critical_exit();
426 	}
427 out:
428 	CTR2(KTR_SPARE2, "cpu_idle(%d) at %d done",
429 	    busy, curcpu);
430 }
431 
432 int
433 cpu_idle_wakeup(int cpu)
434 {
435 	struct pcpu *pcpu;
436 	int *state;
437 
438 	pcpu = pcpu_find(cpu);
439 	state = (int *)pcpu->pc_monitorbuf;
440 	/*
441 	 * This doesn't need to be atomic since missing the race will
442 	 * simply result in unnecessary IPIs.
443 	 */
444 	if (*state == STATE_SLEEPING)
445 		return (0);
446 	if (*state == STATE_MWAIT)
447 		*state = STATE_RUNNING;
448 	return (1);
449 }
450 
451 /*
452  * Ordered by speed/power consumption.
453  */
454 struct {
455 	void	*id_fn;
456 	char	*id_name;
457 } idle_tbl[] = {
458 	{ cpu_idle_spin, "spin" },
459 	{ cpu_idle_mwait, "mwait" },
460 	{ cpu_idle_hlt, "hlt" },
461 	{ cpu_idle_acpi, "acpi" },
462 	{ NULL, NULL }
463 };
464 
465 static int
466 idle_sysctl_available(SYSCTL_HANDLER_ARGS)
467 {
468 	char *avail, *p;
469 	int error;
470 	int i;
471 
472 	avail = malloc(256, M_TEMP, M_WAITOK);
473 	p = avail;
474 	for (i = 0; idle_tbl[i].id_name != NULL; i++) {
475 		if (strstr(idle_tbl[i].id_name, "mwait") &&
476 		    (cpu_feature2 & CPUID2_MON) == 0)
477 			continue;
478 		if (strcmp(idle_tbl[i].id_name, "acpi") == 0 &&
479 		    cpu_idle_hook == NULL)
480 			continue;
481 		p += sprintf(p, "%s%s", p != avail ? ", " : "",
482 		    idle_tbl[i].id_name);
483 	}
484 	error = sysctl_handle_string(oidp, avail, 0, req);
485 	free(avail, M_TEMP);
486 	return (error);
487 }
488 
489 SYSCTL_PROC(_machdep, OID_AUTO, idle_available, CTLTYPE_STRING | CTLFLAG_RD,
490     0, 0, idle_sysctl_available, "A", "list of available idle functions");
491 
492 static int
493 idle_sysctl(SYSCTL_HANDLER_ARGS)
494 {
495 	char buf[16];
496 	int error;
497 	char *p;
498 	int i;
499 
500 	p = "unknown";
501 	for (i = 0; idle_tbl[i].id_name != NULL; i++) {
502 		if (idle_tbl[i].id_fn == cpu_idle_fn) {
503 			p = idle_tbl[i].id_name;
504 			break;
505 		}
506 	}
507 	strncpy(buf, p, sizeof(buf));
508 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
509 	if (error != 0 || req->newptr == NULL)
510 		return (error);
511 	for (i = 0; idle_tbl[i].id_name != NULL; i++) {
512 		if (strstr(idle_tbl[i].id_name, "mwait") &&
513 		    (cpu_feature2 & CPUID2_MON) == 0)
514 			continue;
515 		if (strcmp(idle_tbl[i].id_name, "acpi") == 0 &&
516 		    cpu_idle_hook == NULL)
517 			continue;
518 		if (strcmp(idle_tbl[i].id_name, buf))
519 			continue;
520 		cpu_idle_fn = idle_tbl[i].id_fn;
521 		return (0);
522 	}
523 	return (EINVAL);
524 }
525 
526 SYSCTL_PROC(_machdep, OID_AUTO, idle, CTLTYPE_STRING | CTLFLAG_RW, 0, 0,
527     idle_sysctl, "A", "currently selected idle function");
528 
529 static int panic_on_nmi = 1;
530 SYSCTL_INT(_machdep, OID_AUTO, panic_on_nmi, CTLFLAG_RWTUN,
531     &panic_on_nmi, 0,
532     "Panic on NMI");
533 int nmi_is_broadcast = 1;
534 SYSCTL_INT(_machdep, OID_AUTO, nmi_is_broadcast, CTLFLAG_RWTUN,
535     &nmi_is_broadcast, 0,
536     "Chipset NMI is broadcast");
537 #ifdef KDB
538 int kdb_on_nmi = 1;
539 SYSCTL_INT(_machdep, OID_AUTO, kdb_on_nmi, CTLFLAG_RWTUN,
540     &kdb_on_nmi, 0,
541     "Go to KDB on NMI");
542 #endif
543 
544 #ifdef DEV_ISA
545 void
546 nmi_call_kdb(u_int cpu, u_int type, struct trapframe *frame)
547 {
548 
549 	/* machine/parity/power fail/"kitchen sink" faults */
550 	if (isa_nmi(frame->tf_err) == 0) {
551 #ifdef KDB
552 		/*
553 		 * NMI can be hooked up to a pushbutton for debugging.
554 		 */
555 		if (kdb_on_nmi) {
556 			printf("NMI/cpu%d ... going to debugger\n", cpu);
557 			kdb_trap(type, 0, frame);
558 		}
559 #endif /* KDB */
560 	} else if (panic_on_nmi) {
561 		panic("NMI indicates hardware failure");
562 	}
563 }
564 #endif
565 
566 void
567 nmi_handle_intr(u_int type, struct trapframe *frame)
568 {
569 
570 #ifdef DEV_ISA
571 #ifdef SMP
572 	if (nmi_is_broadcast) {
573 		nmi_call_kdb_smp(type, frame);
574 		return;
575 	}
576 #endif
577 	nmi_call_kdb(PCPU_GET(cpuid), type, frame);
578 #endif
579 }
580 
581 int hw_ibrs_active;
582 int hw_ibrs_disable = 1;
583 
584 SYSCTL_INT(_hw, OID_AUTO, ibrs_active, CTLFLAG_RD, &hw_ibrs_active, 0,
585     "Indirect Branch Restricted Speculation active");
586 
587 void
588 hw_ibrs_recalculate(void)
589 {
590 	uint64_t v;
591 
592 	if ((cpu_ia32_arch_caps & IA32_ARCH_CAP_IBRS_ALL) != 0) {
593 		if (hw_ibrs_disable) {
594 			v= rdmsr(MSR_IA32_SPEC_CTRL);
595 			v &= ~(uint64_t)IA32_SPEC_CTRL_IBRS;
596 			wrmsr(MSR_IA32_SPEC_CTRL, v);
597 		} else {
598 			v= rdmsr(MSR_IA32_SPEC_CTRL);
599 			v |= IA32_SPEC_CTRL_IBRS;
600 			wrmsr(MSR_IA32_SPEC_CTRL, v);
601 		}
602 		return;
603 	}
604 	hw_ibrs_active = (cpu_stdext_feature3 & CPUID_STDEXT3_IBPB) != 0 &&
605 	    !hw_ibrs_disable;
606 }
607 
608 static int
609 hw_ibrs_disable_handler(SYSCTL_HANDLER_ARGS)
610 {
611 	int error, val;
612 
613 	val = hw_ibrs_disable;
614 	error = sysctl_handle_int(oidp, &val, 0, req);
615 	if (error != 0 || req->newptr == NULL)
616 		return (error);
617 	hw_ibrs_disable = val != 0;
618 	hw_ibrs_recalculate();
619 	return (0);
620 }
621 SYSCTL_PROC(_hw, OID_AUTO, ibrs_disable, CTLTYPE_INT | CTLFLAG_RWTUN |
622     CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0, hw_ibrs_disable_handler, "I",
623     "Disable Indirect Branch Restricted Speculation");
624