xref: /freebsd/sys/x86/x86/cpu_machdep.c (revision 48c779cdecb5f803e5fe5d761987e976ca9609db)
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_acpi.h"
45 #include "opt_atpic.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/domainset.h>
65 #include <sys/kdb.h>
66 #include <sys/kernel.h>
67 #include <sys/ktr.h>
68 #include <sys/lock.h>
69 #include <sys/malloc.h>
70 #include <sys/mutex.h>
71 #include <sys/pcpu.h>
72 #include <sys/rwlock.h>
73 #include <sys/sched.h>
74 #include <sys/smp.h>
75 #include <sys/sysctl.h>
76 
77 #include <machine/clock.h>
78 #include <machine/cpu.h>
79 #include <machine/cputypes.h>
80 #include <machine/specialreg.h>
81 #include <machine/md_var.h>
82 #include <machine/mp_watchdog.h>
83 #include <machine/tss.h>
84 #ifdef SMP
85 #include <machine/smp.h>
86 #endif
87 #ifdef CPU_ELAN
88 #include <machine/elan_mmcr.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 #include <isa/isareg.h>
102 
103 #include <contrib/dev/acpica/include/acpi.h>
104 
105 #define	STATE_RUNNING	0x0
106 #define	STATE_MWAIT	0x1
107 #define	STATE_SLEEPING	0x2
108 
109 #ifdef SMP
110 static u_int	cpu_reset_proxyid;
111 static volatile u_int	cpu_reset_proxy_active;
112 #endif
113 
114 /*
115  * Automatically initialized per CPU errata in cpu_idle_tun below.
116  */
117 bool mwait_cpustop_broken = false;
118 SYSCTL_BOOL(_machdep, OID_AUTO, mwait_cpustop_broken, CTLFLAG_RDTUN,
119     &mwait_cpustop_broken, 0,
120     "Can not reliably wake MONITOR/MWAIT cpus without interrupts");
121 
122 /*
123  * Machine dependent boot() routine
124  *
125  * I haven't seen anything to put here yet
126  * Possibly some stuff might be grafted back here from boot()
127  */
128 void
129 cpu_boot(int howto)
130 {
131 }
132 
133 /*
134  * Flush the D-cache for non-DMA I/O so that the I-cache can
135  * be made coherent later.
136  */
137 void
138 cpu_flush_dcache(void *ptr, size_t len)
139 {
140 	/* Not applicable */
141 }
142 
143 void
144 acpi_cpu_c1(void)
145 {
146 
147 	__asm __volatile("sti; hlt");
148 }
149 
150 /*
151  * Use mwait to pause execution while waiting for an interrupt or
152  * another thread to signal that there is more work.
153  *
154  * NOTE: Interrupts will cause a wakeup; however, this function does
155  * not enable interrupt handling. The caller is responsible to enable
156  * interrupts.
157  */
158 void
159 acpi_cpu_idle_mwait(uint32_t mwait_hint)
160 {
161 	int *state;
162 	uint64_t v;
163 
164 	/*
165 	 * A comment in Linux patch claims that 'CPUs run faster with
166 	 * speculation protection disabled. All CPU threads in a core
167 	 * must disable speculation protection for it to be
168 	 * disabled. Disable it while we are idle so the other
169 	 * hyperthread can run fast.'
170 	 *
171 	 * XXXKIB.  Software coordination mode should be supported,
172 	 * but all Intel CPUs provide hardware coordination.
173 	 */
174 
175 	state = &PCPU_PTR(monitorbuf)->idle_state;
176 	KASSERT(atomic_load_int(state) == STATE_SLEEPING,
177 	    ("cpu_mwait_cx: wrong monitorbuf state"));
178 	atomic_store_int(state, STATE_MWAIT);
179 	if (PCPU_GET(ibpb_set) || hw_ssb_active) {
180 		v = rdmsr(MSR_IA32_SPEC_CTRL);
181 		wrmsr(MSR_IA32_SPEC_CTRL, v & ~(IA32_SPEC_CTRL_IBRS |
182 		    IA32_SPEC_CTRL_STIBP | IA32_SPEC_CTRL_SSBD));
183 	} else {
184 		v = 0;
185 	}
186 	cpu_monitor(state, 0, 0);
187 	if (atomic_load_int(state) == STATE_MWAIT)
188 		cpu_mwait(MWAIT_INTRBREAK, mwait_hint);
189 
190 	/*
191 	 * SSB cannot be disabled while we sleep, or rather, if it was
192 	 * disabled, the sysctl thread will bind to our cpu to tweak
193 	 * MSR.
194 	 */
195 	if (v != 0)
196 		wrmsr(MSR_IA32_SPEC_CTRL, v);
197 
198 	/*
199 	 * We should exit on any event that interrupts mwait, because
200 	 * that event might be a wanted interrupt.
201 	 */
202 	atomic_store_int(state, STATE_RUNNING);
203 }
204 
205 /* Get current clock frequency for the given cpu id. */
206 int
207 cpu_est_clockrate(int cpu_id, uint64_t *rate)
208 {
209 	uint64_t tsc1, tsc2;
210 	uint64_t acnt, mcnt, perf;
211 	register_t reg;
212 
213 	if (pcpu_find(cpu_id) == NULL || rate == NULL)
214 		return (EINVAL);
215 #ifdef __i386__
216 	if ((cpu_feature & CPUID_TSC) == 0)
217 		return (EOPNOTSUPP);
218 #endif
219 
220 	/*
221 	 * If TSC is P-state invariant and APERF/MPERF MSRs do not exist,
222 	 * DELAY(9) based logic fails.
223 	 */
224 	if (tsc_is_invariant && !tsc_perf_stat)
225 		return (EOPNOTSUPP);
226 
227 #ifdef SMP
228 	if (smp_cpus > 1) {
229 		/* Schedule ourselves on the indicated cpu. */
230 		thread_lock(curthread);
231 		sched_bind(curthread, cpu_id);
232 		thread_unlock(curthread);
233 	}
234 #endif
235 
236 	/* Calibrate by measuring a short delay. */
237 	reg = intr_disable();
238 	if (tsc_is_invariant) {
239 		wrmsr(MSR_MPERF, 0);
240 		wrmsr(MSR_APERF, 0);
241 		tsc1 = rdtsc();
242 		DELAY(1000);
243 		mcnt = rdmsr(MSR_MPERF);
244 		acnt = rdmsr(MSR_APERF);
245 		tsc2 = rdtsc();
246 		intr_restore(reg);
247 		perf = 1000 * acnt / mcnt;
248 		*rate = (tsc2 - tsc1) * perf;
249 	} else {
250 		tsc1 = rdtsc();
251 		DELAY(1000);
252 		tsc2 = rdtsc();
253 		intr_restore(reg);
254 		*rate = (tsc2 - tsc1) * 1000;
255 	}
256 
257 #ifdef SMP
258 	if (smp_cpus > 1) {
259 		thread_lock(curthread);
260 		sched_unbind(curthread);
261 		thread_unlock(curthread);
262 	}
263 #endif
264 
265 	return (0);
266 }
267 
268 /*
269  * Shutdown the CPU as much as possible
270  */
271 void
272 cpu_halt(void)
273 {
274 	for (;;)
275 		halt();
276 }
277 
278 static void
279 cpu_reset_real(void)
280 {
281 	struct region_descriptor null_idt;
282 	int b;
283 
284 	disable_intr();
285 #ifdef CPU_ELAN
286 	if (elan_mmcr != NULL)
287 		elan_mmcr->RESCFG = 1;
288 #endif
289 #ifdef __i386__
290 	if (cpu == CPU_GEODE1100) {
291 		/* Attempt Geode's own reset */
292 		outl(0xcf8, 0x80009044ul);
293 		outl(0xcfc, 0xf);
294 	}
295 #endif
296 #if !defined(BROKEN_KEYBOARD_RESET)
297 	/*
298 	 * Attempt to do a CPU reset via the keyboard controller,
299 	 * do not turn off GateA20, as any machine that fails
300 	 * to do the reset here would then end up in no man's land.
301 	 */
302 	outb(IO_KBD + 4, 0xFE);
303 	DELAY(500000);	/* wait 0.5 sec to see if that did it */
304 #endif
305 
306 	/*
307 	 * Attempt to force a reset via the Reset Control register at
308 	 * I/O port 0xcf9.  Bit 2 forces a system reset when it
309 	 * transitions from 0 to 1.  Bit 1 selects the type of reset
310 	 * to attempt: 0 selects a "soft" reset, and 1 selects a
311 	 * "hard" reset.  We try a "hard" reset.  The first write sets
312 	 * bit 1 to select a "hard" reset and clears bit 2.  The
313 	 * second write forces a 0 -> 1 transition in bit 2 to trigger
314 	 * a reset.
315 	 */
316 	outb(0xcf9, 0x2);
317 	outb(0xcf9, 0x6);
318 	DELAY(500000);  /* wait 0.5 sec to see if that did it */
319 
320 	/*
321 	 * Attempt to force a reset via the Fast A20 and Init register
322 	 * at I/O port 0x92.  Bit 1 serves as an alternate A20 gate.
323 	 * Bit 0 asserts INIT# when set to 1.  We are careful to only
324 	 * preserve bit 1 while setting bit 0.  We also must clear bit
325 	 * 0 before setting it if it isn't already clear.
326 	 */
327 	b = inb(0x92);
328 	if (b != 0xff) {
329 		if ((b & 0x1) != 0)
330 			outb(0x92, b & 0xfe);
331 		outb(0x92, b | 0x1);
332 		DELAY(500000);  /* wait 0.5 sec to see if that did it */
333 	}
334 
335 	printf("No known reset method worked, attempting CPU shutdown\n");
336 	DELAY(1000000); /* wait 1 sec for printf to complete */
337 
338 	/* Wipe the IDT. */
339 	null_idt.rd_limit = 0;
340 	null_idt.rd_base = 0;
341 	lidt(&null_idt);
342 
343 	/* "good night, sweet prince .... <THUNK!>" */
344 	breakpoint();
345 
346 	/* NOTREACHED */
347 	while(1);
348 }
349 
350 #ifdef SMP
351 static void
352 cpu_reset_proxy(void)
353 {
354 
355 	cpu_reset_proxy_active = 1;
356 	while (cpu_reset_proxy_active == 1)
357 		ia32_pause(); /* Wait for other cpu to see that we've started */
358 
359 	printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid);
360 	DELAY(1000000);
361 	cpu_reset_real();
362 }
363 #endif
364 
365 void
366 cpu_reset(void)
367 {
368 #ifdef SMP
369 	struct monitorbuf *mb;
370 	cpuset_t map;
371 	u_int cnt;
372 
373 	if (smp_started) {
374 		map = all_cpus;
375 		CPU_CLR(PCPU_GET(cpuid), &map);
376 		CPU_NAND(&map, &stopped_cpus);
377 		if (!CPU_EMPTY(&map)) {
378 			printf("cpu_reset: Stopping other CPUs\n");
379 			stop_cpus(map);
380 		}
381 
382 		if (PCPU_GET(cpuid) != 0) {
383 			cpu_reset_proxyid = PCPU_GET(cpuid);
384 			cpustop_restartfunc = cpu_reset_proxy;
385 			cpu_reset_proxy_active = 0;
386 			printf("cpu_reset: Restarting BSP\n");
387 
388 			/* Restart CPU #0. */
389 			CPU_SETOF(0, &started_cpus);
390 			mb = &pcpu_find(0)->pc_monitorbuf;
391 			atomic_store_int(&mb->stop_state,
392 			    MONITOR_STOPSTATE_RUNNING);
393 
394 			cnt = 0;
395 			while (cpu_reset_proxy_active == 0 && cnt < 10000000) {
396 				ia32_pause();
397 				cnt++;	/* Wait for BSP to announce restart */
398 			}
399 			if (cpu_reset_proxy_active == 0) {
400 				printf("cpu_reset: Failed to restart BSP\n");
401 			} else {
402 				cpu_reset_proxy_active = 2;
403 				while (1)
404 					ia32_pause();
405 				/* NOTREACHED */
406 			}
407 		}
408 
409 		DELAY(1000000);
410 	}
411 #endif
412 	cpu_reset_real();
413 	/* NOTREACHED */
414 }
415 
416 bool
417 cpu_mwait_usable(void)
418 {
419 
420 	return ((cpu_feature2 & CPUID2_MON) != 0 && ((cpu_mon_mwait_flags &
421 	    (CPUID5_MON_MWAIT_EXT | CPUID5_MWAIT_INTRBREAK)) ==
422 	    (CPUID5_MON_MWAIT_EXT | CPUID5_MWAIT_INTRBREAK)));
423 }
424 
425 void (*cpu_idle_hook)(sbintime_t) = NULL;	/* ACPI idle hook. */
426 static int	cpu_ident_amdc1e = 0;	/* AMD C1E supported. */
427 static int	idle_mwait = 1;		/* Use MONITOR/MWAIT for short idle. */
428 SYSCTL_INT(_machdep, OID_AUTO, idle_mwait, CTLFLAG_RWTUN, &idle_mwait,
429     0, "Use MONITOR/MWAIT for short idle");
430 
431 static void
432 cpu_idle_acpi(sbintime_t sbt)
433 {
434 	int *state;
435 
436 	state = &PCPU_PTR(monitorbuf)->idle_state;
437 	atomic_store_int(state, STATE_SLEEPING);
438 
439 	/* See comments in cpu_idle_hlt(). */
440 	disable_intr();
441 	if (sched_runnable())
442 		enable_intr();
443 	else if (cpu_idle_hook)
444 		cpu_idle_hook(sbt);
445 	else
446 		acpi_cpu_c1();
447 	atomic_store_int(state, STATE_RUNNING);
448 }
449 
450 static void
451 cpu_idle_hlt(sbintime_t sbt)
452 {
453 	int *state;
454 
455 	state = &PCPU_PTR(monitorbuf)->idle_state;
456 	atomic_store_int(state, STATE_SLEEPING);
457 
458 	/*
459 	 * Since we may be in a critical section from cpu_idle(), if
460 	 * an interrupt fires during that critical section we may have
461 	 * a pending preemption.  If the CPU halts, then that thread
462 	 * may not execute until a later interrupt awakens the CPU.
463 	 * To handle this race, check for a runnable thread after
464 	 * disabling interrupts and immediately return if one is
465 	 * found.  Also, we must absolutely guarentee that hlt is
466 	 * the next instruction after sti.  This ensures that any
467 	 * interrupt that fires after the call to disable_intr() will
468 	 * immediately awaken the CPU from hlt.  Finally, please note
469 	 * that on x86 this works fine because of interrupts enabled only
470 	 * after the instruction following sti takes place, while IF is set
471 	 * to 1 immediately, allowing hlt instruction to acknowledge the
472 	 * interrupt.
473 	 */
474 	disable_intr();
475 	if (sched_runnable())
476 		enable_intr();
477 	else
478 		acpi_cpu_c1();
479 	atomic_store_int(state, STATE_RUNNING);
480 }
481 
482 static void
483 cpu_idle_mwait(sbintime_t sbt)
484 {
485 	int *state;
486 
487 	state = &PCPU_PTR(monitorbuf)->idle_state;
488 	atomic_store_int(state, STATE_MWAIT);
489 
490 	/* See comments in cpu_idle_hlt(). */
491 	disable_intr();
492 	if (sched_runnable()) {
493 		atomic_store_int(state, STATE_RUNNING);
494 		enable_intr();
495 		return;
496 	}
497 
498 	cpu_monitor(state, 0, 0);
499 	if (atomic_load_int(state) == STATE_MWAIT)
500 		__asm __volatile("sti; mwait" : : "a" (MWAIT_C1), "c" (0));
501 	else
502 		enable_intr();
503 	atomic_store_int(state, STATE_RUNNING);
504 }
505 
506 static void
507 cpu_idle_spin(sbintime_t sbt)
508 {
509 	int *state;
510 	int i;
511 
512 	state = &PCPU_PTR(monitorbuf)->idle_state;
513 	atomic_store_int(state, STATE_RUNNING);
514 
515 	/*
516 	 * The sched_runnable() call is racy but as long as there is
517 	 * a loop missing it one time will have just a little impact if any
518 	 * (and it is much better than missing the check at all).
519 	 */
520 	for (i = 0; i < 1000; i++) {
521 		if (sched_runnable())
522 			return;
523 		cpu_spinwait();
524 	}
525 }
526 
527 /*
528  * C1E renders the local APIC timer dead, so we disable it by
529  * reading the Interrupt Pending Message register and clearing
530  * both C1eOnCmpHalt (bit 28) and SmiOnCmpHalt (bit 27).
531  *
532  * Reference:
533  *   "BIOS and Kernel Developer's Guide for AMD NPT Family 0Fh Processors"
534  *   #32559 revision 3.00+
535  */
536 #define	MSR_AMDK8_IPM		0xc0010055
537 #define	AMDK8_SMIONCMPHALT	(1ULL << 27)
538 #define	AMDK8_C1EONCMPHALT	(1ULL << 28)
539 #define	AMDK8_CMPHALT		(AMDK8_SMIONCMPHALT | AMDK8_C1EONCMPHALT)
540 
541 void
542 cpu_probe_amdc1e(void)
543 {
544 
545 	/*
546 	 * Detect the presence of C1E capability mostly on latest
547 	 * dual-cores (or future) k8 family.
548 	 */
549 	if (cpu_vendor_id == CPU_VENDOR_AMD &&
550 	    (cpu_id & 0x00000f00) == 0x00000f00 &&
551 	    (cpu_id & 0x0fff0000) >=  0x00040000) {
552 		cpu_ident_amdc1e = 1;
553 	}
554 }
555 
556 void (*cpu_idle_fn)(sbintime_t) = cpu_idle_acpi;
557 
558 void
559 cpu_idle(int busy)
560 {
561 	uint64_t msr;
562 	sbintime_t sbt = -1;
563 
564 	CTR2(KTR_SPARE2, "cpu_idle(%d) at %d",
565 	    busy, curcpu);
566 #ifdef MP_WATCHDOG
567 	ap_watchdog(PCPU_GET(cpuid));
568 #endif
569 
570 	/* If we are busy - try to use fast methods. */
571 	if (busy) {
572 		if ((cpu_feature2 & CPUID2_MON) && idle_mwait) {
573 			cpu_idle_mwait(busy);
574 			goto out;
575 		}
576 	}
577 
578 	/* If we have time - switch timers into idle mode. */
579 	if (!busy) {
580 		critical_enter();
581 		sbt = cpu_idleclock();
582 	}
583 
584 	/* Apply AMD APIC timer C1E workaround. */
585 	if (cpu_ident_amdc1e && cpu_disable_c3_sleep) {
586 		msr = rdmsr(MSR_AMDK8_IPM);
587 		if (msr & AMDK8_CMPHALT)
588 			wrmsr(MSR_AMDK8_IPM, msr & ~AMDK8_CMPHALT);
589 	}
590 
591 	/* Call main idle method. */
592 	cpu_idle_fn(sbt);
593 
594 	/* Switch timers back into active mode. */
595 	if (!busy) {
596 		cpu_activeclock();
597 		critical_exit();
598 	}
599 out:
600 	CTR2(KTR_SPARE2, "cpu_idle(%d) at %d done",
601 	    busy, curcpu);
602 }
603 
604 static int cpu_idle_apl31_workaround;
605 SYSCTL_INT(_machdep, OID_AUTO, idle_apl31, CTLFLAG_RW,
606     &cpu_idle_apl31_workaround, 0,
607     "Apollo Lake APL31 MWAIT bug workaround");
608 
609 int
610 cpu_idle_wakeup(int cpu)
611 {
612 	struct monitorbuf *mb;
613 	int *state;
614 
615 	mb = &pcpu_find(cpu)->pc_monitorbuf;
616 	state = &mb->idle_state;
617 	switch (atomic_load_int(state)) {
618 	case STATE_SLEEPING:
619 		return (0);
620 	case STATE_MWAIT:
621 		atomic_store_int(state, STATE_RUNNING);
622 		return (cpu_idle_apl31_workaround ? 0 : 1);
623 	case STATE_RUNNING:
624 		return (1);
625 	default:
626 		panic("bad monitor state");
627 		return (1);
628 	}
629 }
630 
631 /*
632  * Ordered by speed/power consumption.
633  */
634 static struct {
635 	void	*id_fn;
636 	char	*id_name;
637 	int	id_cpuid2_flag;
638 } idle_tbl[] = {
639 	{ .id_fn = cpu_idle_spin, .id_name = "spin" },
640 	{ .id_fn = cpu_idle_mwait, .id_name = "mwait",
641 	    .id_cpuid2_flag = CPUID2_MON },
642 	{ .id_fn = cpu_idle_hlt, .id_name = "hlt" },
643 	{ .id_fn = cpu_idle_acpi, .id_name = "acpi" },
644 };
645 
646 static int
647 idle_sysctl_available(SYSCTL_HANDLER_ARGS)
648 {
649 	char *avail, *p;
650 	int error;
651 	int i;
652 
653 	avail = malloc(256, M_TEMP, M_WAITOK);
654 	p = avail;
655 	for (i = 0; i < nitems(idle_tbl); i++) {
656 		if (idle_tbl[i].id_cpuid2_flag != 0 &&
657 		    (cpu_feature2 & idle_tbl[i].id_cpuid2_flag) == 0)
658 			continue;
659 		if (strcmp(idle_tbl[i].id_name, "acpi") == 0 &&
660 		    cpu_idle_hook == NULL)
661 			continue;
662 		p += sprintf(p, "%s%s", p != avail ? ", " : "",
663 		    idle_tbl[i].id_name);
664 	}
665 	error = sysctl_handle_string(oidp, avail, 0, req);
666 	free(avail, M_TEMP);
667 	return (error);
668 }
669 
670 SYSCTL_PROC(_machdep, OID_AUTO, idle_available, CTLTYPE_STRING | CTLFLAG_RD,
671     0, 0, idle_sysctl_available, "A", "list of available idle functions");
672 
673 static bool
674 cpu_idle_selector(const char *new_idle_name)
675 {
676 	int i;
677 
678 	for (i = 0; i < nitems(idle_tbl); i++) {
679 		if (idle_tbl[i].id_cpuid2_flag != 0 &&
680 		    (cpu_feature2 & idle_tbl[i].id_cpuid2_flag) == 0)
681 			continue;
682 		if (strcmp(idle_tbl[i].id_name, "acpi") == 0 &&
683 		    cpu_idle_hook == NULL)
684 			continue;
685 		if (strcmp(idle_tbl[i].id_name, new_idle_name))
686 			continue;
687 		cpu_idle_fn = idle_tbl[i].id_fn;
688 		if (bootverbose)
689 			printf("CPU idle set to %s\n", idle_tbl[i].id_name);
690 		return (true);
691 	}
692 	return (false);
693 }
694 
695 static int
696 cpu_idle_sysctl(SYSCTL_HANDLER_ARGS)
697 {
698 	char buf[16], *p;
699 	int error, i;
700 
701 	p = "unknown";
702 	for (i = 0; i < nitems(idle_tbl); i++) {
703 		if (idle_tbl[i].id_fn == cpu_idle_fn) {
704 			p = idle_tbl[i].id_name;
705 			break;
706 		}
707 	}
708 	strncpy(buf, p, sizeof(buf));
709 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
710 	if (error != 0 || req->newptr == NULL)
711 		return (error);
712 	return (cpu_idle_selector(buf) ? 0 : EINVAL);
713 }
714 
715 SYSCTL_PROC(_machdep, OID_AUTO, idle, CTLTYPE_STRING | CTLFLAG_RW, 0, 0,
716     cpu_idle_sysctl, "A", "currently selected idle function");
717 
718 static void
719 cpu_idle_tun(void *unused __unused)
720 {
721 	char tunvar[16];
722 
723 	if (TUNABLE_STR_FETCH("machdep.idle", tunvar, sizeof(tunvar)))
724 		cpu_idle_selector(tunvar);
725 	else if (cpu_vendor_id == CPU_VENDOR_AMD &&
726 	    CPUID_TO_FAMILY(cpu_id) == 0x17 && CPUID_TO_MODEL(cpu_id) == 0x1) {
727 		/* Ryzen erratas 1057, 1109. */
728 		cpu_idle_selector("hlt");
729 		idle_mwait = 0;
730 		mwait_cpustop_broken = true;
731 	}
732 
733 	if (cpu_vendor_id == CPU_VENDOR_INTEL && cpu_id == 0x506c9) {
734 		/*
735 		 * Apollo Lake errata APL31 (public errata APL30).
736 		 * Stores to the armed address range may not trigger
737 		 * MWAIT to resume execution.  OS needs to use
738 		 * interrupts to wake processors from MWAIT-induced
739 		 * sleep states.
740 		 */
741 		cpu_idle_apl31_workaround = 1;
742 		mwait_cpustop_broken = true;
743 	}
744 	TUNABLE_INT_FETCH("machdep.idle_apl31", &cpu_idle_apl31_workaround);
745 }
746 SYSINIT(cpu_idle_tun, SI_SUB_CPU, SI_ORDER_MIDDLE, cpu_idle_tun, NULL);
747 
748 static int panic_on_nmi = 1;
749 SYSCTL_INT(_machdep, OID_AUTO, panic_on_nmi, CTLFLAG_RWTUN,
750     &panic_on_nmi, 0,
751     "Panic on NMI raised by hardware failure");
752 int nmi_is_broadcast = 1;
753 SYSCTL_INT(_machdep, OID_AUTO, nmi_is_broadcast, CTLFLAG_RWTUN,
754     &nmi_is_broadcast, 0,
755     "Chipset NMI is broadcast");
756 #ifdef KDB
757 int kdb_on_nmi = 1;
758 SYSCTL_INT(_machdep, OID_AUTO, kdb_on_nmi, CTLFLAG_RWTUN,
759     &kdb_on_nmi, 0,
760     "Go to KDB on NMI with unknown source");
761 #endif
762 
763 void
764 nmi_call_kdb(u_int cpu, u_int type, struct trapframe *frame)
765 {
766 	bool claimed = false;
767 
768 #ifdef DEV_ISA
769 	/* machine/parity/power fail/"kitchen sink" faults */
770 	if (isa_nmi(frame->tf_err)) {
771 		claimed = true;
772 		if (panic_on_nmi)
773 			panic("NMI indicates hardware failure");
774 	}
775 #endif /* DEV_ISA */
776 #ifdef KDB
777 	if (!claimed && kdb_on_nmi) {
778 		/*
779 		 * NMI can be hooked up to a pushbutton for debugging.
780 		 */
781 		printf("NMI/cpu%d ... going to debugger\n", cpu);
782 		kdb_trap(type, 0, frame);
783 	}
784 #endif /* KDB */
785 }
786 
787 void
788 nmi_handle_intr(u_int type, struct trapframe *frame)
789 {
790 
791 #ifdef SMP
792 	if (nmi_is_broadcast) {
793 		nmi_call_kdb_smp(type, frame);
794 		return;
795 	}
796 #endif
797 	nmi_call_kdb(PCPU_GET(cpuid), type, frame);
798 }
799 
800 int hw_ibrs_active;
801 int hw_ibrs_disable = 1;
802 
803 SYSCTL_INT(_hw, OID_AUTO, ibrs_active, CTLFLAG_RD, &hw_ibrs_active, 0,
804     "Indirect Branch Restricted Speculation active");
805 
806 void
807 hw_ibrs_recalculate(void)
808 {
809 	uint64_t v;
810 
811 	if ((cpu_ia32_arch_caps & IA32_ARCH_CAP_IBRS_ALL) != 0) {
812 		if (hw_ibrs_disable) {
813 			v = rdmsr(MSR_IA32_SPEC_CTRL);
814 			v &= ~(uint64_t)IA32_SPEC_CTRL_IBRS;
815 			wrmsr(MSR_IA32_SPEC_CTRL, v);
816 		} else {
817 			v = rdmsr(MSR_IA32_SPEC_CTRL);
818 			v |= IA32_SPEC_CTRL_IBRS;
819 			wrmsr(MSR_IA32_SPEC_CTRL, v);
820 		}
821 		return;
822 	}
823 	hw_ibrs_active = (cpu_stdext_feature3 & CPUID_STDEXT3_IBPB) != 0 &&
824 	    !hw_ibrs_disable;
825 }
826 
827 static int
828 hw_ibrs_disable_handler(SYSCTL_HANDLER_ARGS)
829 {
830 	int error, val;
831 
832 	val = hw_ibrs_disable;
833 	error = sysctl_handle_int(oidp, &val, 0, req);
834 	if (error != 0 || req->newptr == NULL)
835 		return (error);
836 	hw_ibrs_disable = val != 0;
837 	hw_ibrs_recalculate();
838 	return (0);
839 }
840 SYSCTL_PROC(_hw, OID_AUTO, ibrs_disable, CTLTYPE_INT | CTLFLAG_RWTUN |
841     CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0, hw_ibrs_disable_handler, "I",
842     "Disable Indirect Branch Restricted Speculation");
843 
844 int hw_ssb_active;
845 int hw_ssb_disable;
846 
847 SYSCTL_INT(_hw, OID_AUTO, spec_store_bypass_disable_active, CTLFLAG_RD,
848     &hw_ssb_active, 0,
849     "Speculative Store Bypass Disable active");
850 
851 static void
852 hw_ssb_set_one(bool enable)
853 {
854 	uint64_t v;
855 
856 	v = rdmsr(MSR_IA32_SPEC_CTRL);
857 	if (enable)
858 		v |= (uint64_t)IA32_SPEC_CTRL_SSBD;
859 	else
860 		v &= ~(uint64_t)IA32_SPEC_CTRL_SSBD;
861 	wrmsr(MSR_IA32_SPEC_CTRL, v);
862 }
863 
864 static void
865 hw_ssb_set(bool enable, bool for_all_cpus)
866 {
867 	struct thread *td;
868 	int bound_cpu, i, is_bound;
869 
870 	if ((cpu_stdext_feature3 & CPUID_STDEXT3_SSBD) == 0) {
871 		hw_ssb_active = 0;
872 		return;
873 	}
874 	hw_ssb_active = enable;
875 	if (for_all_cpus) {
876 		td = curthread;
877 		thread_lock(td);
878 		is_bound = sched_is_bound(td);
879 		bound_cpu = td->td_oncpu;
880 		CPU_FOREACH(i) {
881 			sched_bind(td, i);
882 			hw_ssb_set_one(enable);
883 		}
884 		if (is_bound)
885 			sched_bind(td, bound_cpu);
886 		else
887 			sched_unbind(td);
888 		thread_unlock(td);
889 	} else {
890 		hw_ssb_set_one(enable);
891 	}
892 }
893 
894 void
895 hw_ssb_recalculate(bool all_cpus)
896 {
897 
898 	switch (hw_ssb_disable) {
899 	default:
900 		hw_ssb_disable = 0;
901 		/* FALLTHROUGH */
902 	case 0: /* off */
903 		hw_ssb_set(false, all_cpus);
904 		break;
905 	case 1: /* on */
906 		hw_ssb_set(true, all_cpus);
907 		break;
908 	case 2: /* auto */
909 		hw_ssb_set((cpu_ia32_arch_caps & IA32_ARCH_CAP_SSB_NO) != 0 ?
910 		    false : true, all_cpus);
911 		break;
912 	}
913 }
914 
915 static int
916 hw_ssb_disable_handler(SYSCTL_HANDLER_ARGS)
917 {
918 	int error, val;
919 
920 	val = hw_ssb_disable;
921 	error = sysctl_handle_int(oidp, &val, 0, req);
922 	if (error != 0 || req->newptr == NULL)
923 		return (error);
924 	hw_ssb_disable = val;
925 	hw_ssb_recalculate(true);
926 	return (0);
927 }
928 SYSCTL_PROC(_hw, OID_AUTO, spec_store_bypass_disable, CTLTYPE_INT |
929     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
930     hw_ssb_disable_handler, "I",
931     "Speculative Store Bypass Disable (0 - off, 1 - on, 2 - auto");
932 
933 int hw_mds_disable;
934 
935 /*
936  * Handler for Microarchitectural Data Sampling issues.  Really not a
937  * pointer to C function: on amd64 the code must not change any CPU
938  * architectural state except possibly %rflags. Also, it is always
939  * called with interrupts disabled.
940  */
941 void mds_handler_void(void);
942 void mds_handler_verw(void);
943 void mds_handler_ivb(void);
944 void mds_handler_bdw(void);
945 void mds_handler_skl_sse(void);
946 void mds_handler_skl_avx(void);
947 void mds_handler_skl_avx512(void);
948 void mds_handler_silvermont(void);
949 void (*mds_handler)(void) = mds_handler_void;
950 
951 static int
952 sysctl_hw_mds_disable_state_handler(SYSCTL_HANDLER_ARGS)
953 {
954 	const char *state;
955 
956 	if (mds_handler == mds_handler_void)
957 		state = "inactive";
958 	else if (mds_handler == mds_handler_verw)
959 		state = "VERW";
960 	else if (mds_handler == mds_handler_ivb)
961 		state = "software IvyBridge";
962 	else if (mds_handler == mds_handler_bdw)
963 		state = "software Broadwell";
964 	else if (mds_handler == mds_handler_skl_sse)
965 		state = "software Skylake SSE";
966 	else if (mds_handler == mds_handler_skl_avx)
967 		state = "software Skylake AVX";
968 	else if (mds_handler == mds_handler_skl_avx512)
969 		state = "software Skylake AVX512";
970 	else if (mds_handler == mds_handler_silvermont)
971 		state = "software Silvermont";
972 	else
973 		state = "unknown";
974 	return (SYSCTL_OUT(req, state, strlen(state)));
975 }
976 
977 SYSCTL_PROC(_hw, OID_AUTO, mds_disable_state,
978     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
979     sysctl_hw_mds_disable_state_handler, "A",
980     "Microarchitectural Data Sampling Mitigation state");
981 
982 _Static_assert(__offsetof(struct pcpu, pc_mds_tmp) % 64 == 0, "MDS AVX512");
983 
984 void
985 hw_mds_recalculate(void)
986 {
987 	struct pcpu *pc;
988 	vm_offset_t b64;
989 	u_long xcr0;
990 	int i;
991 
992 	/*
993 	 * Allow user to force VERW variant even if MD_CLEAR is not
994 	 * reported.  For instance, hypervisor might unknowingly
995 	 * filter the cap out.
996 	 * For the similar reasons, and for testing, allow to enable
997 	 * mitigation even for RDCL_NO or MDS_NO caps.
998 	 */
999 	if (cpu_vendor_id != CPU_VENDOR_INTEL || hw_mds_disable == 0 ||
1000 	    ((cpu_ia32_arch_caps & (IA32_ARCH_CAP_RDCL_NO |
1001 	    IA32_ARCH_CAP_MDS_NO)) != 0 && hw_mds_disable == 3)) {
1002 		mds_handler = mds_handler_void;
1003 	} else if (((cpu_stdext_feature3 & CPUID_STDEXT3_MD_CLEAR) != 0 &&
1004 	    hw_mds_disable == 3) || hw_mds_disable == 1) {
1005 		mds_handler = mds_handler_verw;
1006 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1007 	    (CPUID_TO_MODEL(cpu_id) == 0x2e || CPUID_TO_MODEL(cpu_id) == 0x1e ||
1008 	    CPUID_TO_MODEL(cpu_id) == 0x1f || CPUID_TO_MODEL(cpu_id) == 0x1a ||
1009 	    CPUID_TO_MODEL(cpu_id) == 0x2f || CPUID_TO_MODEL(cpu_id) == 0x25 ||
1010 	    CPUID_TO_MODEL(cpu_id) == 0x2c || CPUID_TO_MODEL(cpu_id) == 0x2d ||
1011 	    CPUID_TO_MODEL(cpu_id) == 0x2a || CPUID_TO_MODEL(cpu_id) == 0x3e ||
1012 	    CPUID_TO_MODEL(cpu_id) == 0x3a) &&
1013 	    (hw_mds_disable == 2 || hw_mds_disable == 3)) {
1014 		/*
1015 		 * Nehalem, SandyBridge, IvyBridge
1016 		 */
1017 		CPU_FOREACH(i) {
1018 			pc = pcpu_find(i);
1019 			if (pc->pc_mds_buf == NULL) {
1020 				pc->pc_mds_buf = malloc_domainset(672, M_TEMP,
1021 				    DOMAINSET_PREF(pc->pc_domain), M_WAITOK);
1022 				bzero(pc->pc_mds_buf, 16);
1023 			}
1024 		}
1025 		mds_handler = mds_handler_ivb;
1026 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1027 	    (CPUID_TO_MODEL(cpu_id) == 0x3f || CPUID_TO_MODEL(cpu_id) == 0x3c ||
1028 	    CPUID_TO_MODEL(cpu_id) == 0x45 || CPUID_TO_MODEL(cpu_id) == 0x46 ||
1029 	    CPUID_TO_MODEL(cpu_id) == 0x56 || CPUID_TO_MODEL(cpu_id) == 0x4f ||
1030 	    CPUID_TO_MODEL(cpu_id) == 0x47 || CPUID_TO_MODEL(cpu_id) == 0x3d) &&
1031 	    (hw_mds_disable == 2 || hw_mds_disable == 3)) {
1032 		/*
1033 		 * Haswell, Broadwell
1034 		 */
1035 		CPU_FOREACH(i) {
1036 			pc = pcpu_find(i);
1037 			if (pc->pc_mds_buf == NULL) {
1038 				pc->pc_mds_buf = malloc_domainset(1536, M_TEMP,
1039 				    DOMAINSET_PREF(pc->pc_domain), M_WAITOK);
1040 				bzero(pc->pc_mds_buf, 16);
1041 			}
1042 		}
1043 		mds_handler = mds_handler_bdw;
1044 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1045 	    ((CPUID_TO_MODEL(cpu_id) == 0x55 && (cpu_id &
1046 	    CPUID_STEPPING) <= 5) ||
1047 	    CPUID_TO_MODEL(cpu_id) == 0x4e || CPUID_TO_MODEL(cpu_id) == 0x5e ||
1048 	    (CPUID_TO_MODEL(cpu_id) == 0x8e && (cpu_id &
1049 	    CPUID_STEPPING) <= 0xb) ||
1050 	    (CPUID_TO_MODEL(cpu_id) == 0x9e && (cpu_id &
1051 	    CPUID_STEPPING) <= 0xc)) &&
1052 	    (hw_mds_disable == 2 || hw_mds_disable == 3)) {
1053 		/*
1054 		 * Skylake, KabyLake, CoffeeLake, WhiskeyLake,
1055 		 * CascadeLake
1056 		 */
1057 		CPU_FOREACH(i) {
1058 			pc = pcpu_find(i);
1059 			if (pc->pc_mds_buf == NULL) {
1060 				pc->pc_mds_buf = malloc_domainset(6 * 1024,
1061 				    M_TEMP, DOMAINSET_PREF(pc->pc_domain),
1062 				    M_WAITOK);
1063 				b64 = (vm_offset_t)malloc_domainset(64 + 63,
1064 				    M_TEMP, DOMAINSET_PREF(pc->pc_domain),
1065 				    M_WAITOK);
1066 				pc->pc_mds_buf64 = (void *)roundup2(b64, 64);
1067 				bzero(pc->pc_mds_buf64, 64);
1068 			}
1069 		}
1070 		xcr0 = rxcr(0);
1071 		if ((xcr0 & XFEATURE_ENABLED_ZMM_HI256) != 0 &&
1072 		    (cpu_stdext_feature2 & CPUID_STDEXT_AVX512DQ) != 0)
1073 			mds_handler = mds_handler_skl_avx512;
1074 		else if ((xcr0 & XFEATURE_ENABLED_AVX) != 0 &&
1075 		    (cpu_feature2 & CPUID2_AVX) != 0)
1076 			mds_handler = mds_handler_skl_avx;
1077 		else
1078 			mds_handler = mds_handler_skl_sse;
1079 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1080 	    ((CPUID_TO_MODEL(cpu_id) == 0x37 ||
1081 	    CPUID_TO_MODEL(cpu_id) == 0x4a ||
1082 	    CPUID_TO_MODEL(cpu_id) == 0x4c ||
1083 	    CPUID_TO_MODEL(cpu_id) == 0x4d ||
1084 	    CPUID_TO_MODEL(cpu_id) == 0x5a ||
1085 	    CPUID_TO_MODEL(cpu_id) == 0x5d ||
1086 	    CPUID_TO_MODEL(cpu_id) == 0x6e ||
1087 	    CPUID_TO_MODEL(cpu_id) == 0x65 ||
1088 	    CPUID_TO_MODEL(cpu_id) == 0x75 ||
1089 	    CPUID_TO_MODEL(cpu_id) == 0x1c ||
1090 	    CPUID_TO_MODEL(cpu_id) == 0x26 ||
1091 	    CPUID_TO_MODEL(cpu_id) == 0x27 ||
1092 	    CPUID_TO_MODEL(cpu_id) == 0x35 ||
1093 	    CPUID_TO_MODEL(cpu_id) == 0x36 ||
1094 	    CPUID_TO_MODEL(cpu_id) == 0x7a))) {
1095 		/* Silvermont, Airmont */
1096 		CPU_FOREACH(i) {
1097 			pc = pcpu_find(i);
1098 			if (pc->pc_mds_buf == NULL)
1099 				pc->pc_mds_buf = malloc(256, M_TEMP, M_WAITOK);
1100 		}
1101 		mds_handler = mds_handler_silvermont;
1102 	} else {
1103 		hw_mds_disable = 0;
1104 		mds_handler = mds_handler_void;
1105 	}
1106 }
1107 
1108 static void
1109 hw_mds_recalculate_boot(void *arg __unused)
1110 {
1111 
1112 	hw_mds_recalculate();
1113 }
1114 SYSINIT(mds_recalc, SI_SUB_SMP, SI_ORDER_ANY, hw_mds_recalculate_boot, NULL);
1115 
1116 static int
1117 sysctl_mds_disable_handler(SYSCTL_HANDLER_ARGS)
1118 {
1119 	int error, val;
1120 
1121 	val = hw_mds_disable;
1122 	error = sysctl_handle_int(oidp, &val, 0, req);
1123 	if (error != 0 || req->newptr == NULL)
1124 		return (error);
1125 	if (val < 0 || val > 3)
1126 		return (EINVAL);
1127 	hw_mds_disable = val;
1128 	hw_mds_recalculate();
1129 	return (0);
1130 }
1131 
1132 SYSCTL_PROC(_hw, OID_AUTO, mds_disable, CTLTYPE_INT |
1133     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1134     sysctl_mds_disable_handler, "I",
1135     "Microarchitectural Data Sampling Mitigation "
1136     "(0 - off, 1 - on VERW, 2 - on SW, 3 - on AUTO");
1137 
1138 /*
1139  * Enable and restore kernel text write permissions.
1140  * Callers must ensure that disable_wp()/restore_wp() are executed
1141  * without rescheduling on the same core.
1142  */
1143 bool
1144 disable_wp(void)
1145 {
1146 	u_int cr0;
1147 
1148 	cr0 = rcr0();
1149 	if ((cr0 & CR0_WP) == 0)
1150 		return (false);
1151 	load_cr0(cr0 & ~CR0_WP);
1152 	return (true);
1153 }
1154 
1155 void
1156 restore_wp(bool old_wp)
1157 {
1158 
1159 	if (old_wp)
1160 		load_cr0(rcr0() | CR0_WP);
1161 }
1162 
1163 bool
1164 acpi_get_fadt_bootflags(uint16_t *flagsp)
1165 {
1166 #ifdef DEV_ACPI
1167 	ACPI_TABLE_FADT *fadt;
1168 	vm_paddr_t physaddr;
1169 
1170 	physaddr = acpi_find_table(ACPI_SIG_FADT);
1171 	if (physaddr == 0)
1172 		return (false);
1173 	fadt = acpi_map_table(physaddr, ACPI_SIG_FADT);
1174 	if (fadt == NULL)
1175 		return (false);
1176 	*flagsp = fadt->BootFlags;
1177 	acpi_unmap_table(fadt);
1178 	return (true);
1179 #else
1180 	return (false);
1181 #endif
1182 }
1183