xref: /freebsd/sys/x86/x86/cpu_machdep.c (revision a0b9e2e854027e6ff61fb075a1309dbc71c42b54)
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 struct msr_op_arg {
115 	u_int msr;
116 	int op;
117 	uint64_t arg1;
118 };
119 
120 static void
121 x86_msr_op_one(void *argp)
122 {
123 	struct msr_op_arg *a;
124 	uint64_t v;
125 
126 	a = argp;
127 	switch (a->op) {
128 	case MSR_OP_ANDNOT:
129 		v = rdmsr(a->msr);
130 		v &= ~a->arg1;
131 		wrmsr(a->msr, v);
132 		break;
133 	case MSR_OP_OR:
134 		v = rdmsr(a->msr);
135 		v |= a->arg1;
136 		wrmsr(a->msr, v);
137 		break;
138 	case MSR_OP_WRITE:
139 		wrmsr(a->msr, a->arg1);
140 		break;
141 	}
142 }
143 
144 #define	MSR_OP_EXMODE_MASK	0xf0000000
145 #define	MSR_OP_OP_MASK		0x000000ff
146 
147 void
148 x86_msr_op(u_int msr, u_int op, uint64_t arg1)
149 {
150 	struct thread *td;
151 	struct msr_op_arg a;
152 	u_int exmode;
153 	int bound_cpu, i, is_bound;
154 
155 	a.op = op & MSR_OP_OP_MASK;
156 	MPASS(a.op == MSR_OP_ANDNOT || a.op == MSR_OP_OR ||
157 	    a.op == MSR_OP_WRITE);
158 	exmode = op & MSR_OP_EXMODE_MASK;
159 	MPASS(exmode == MSR_OP_LOCAL || exmode == MSR_OP_SCHED ||
160 	    exmode == MSR_OP_RENDEZVOUS);
161 	a.msr = msr;
162 	a.arg1 = arg1;
163 	switch (exmode) {
164 	case MSR_OP_LOCAL:
165 		x86_msr_op_one(&a);
166 		break;
167 	case MSR_OP_SCHED:
168 		td = curthread;
169 		thread_lock(td);
170 		is_bound = sched_is_bound(td);
171 		bound_cpu = td->td_oncpu;
172 		CPU_FOREACH(i) {
173 			sched_bind(td, i);
174 			x86_msr_op_one(&a);
175 		}
176 		if (is_bound)
177 			sched_bind(td, bound_cpu);
178 		else
179 			sched_unbind(td);
180 		thread_unlock(td);
181 		break;
182 	case MSR_OP_RENDEZVOUS:
183 		smp_rendezvous(NULL, x86_msr_op_one, NULL, &a);
184 		break;
185 	}
186 }
187 
188 /*
189  * Automatically initialized per CPU errata in cpu_idle_tun below.
190  */
191 bool mwait_cpustop_broken = false;
192 SYSCTL_BOOL(_machdep, OID_AUTO, mwait_cpustop_broken, CTLFLAG_RDTUN,
193     &mwait_cpustop_broken, 0,
194     "Can not reliably wake MONITOR/MWAIT cpus without interrupts");
195 
196 /*
197  * Flush the D-cache for non-DMA I/O so that the I-cache can
198  * be made coherent later.
199  */
200 void
201 cpu_flush_dcache(void *ptr, size_t len)
202 {
203 	/* Not applicable */
204 }
205 
206 void
207 acpi_cpu_c1(void)
208 {
209 
210 	__asm __volatile("sti; hlt");
211 }
212 
213 /*
214  * Use mwait to pause execution while waiting for an interrupt or
215  * another thread to signal that there is more work.
216  *
217  * NOTE: Interrupts will cause a wakeup; however, this function does
218  * not enable interrupt handling. The caller is responsible to enable
219  * interrupts.
220  */
221 void
222 acpi_cpu_idle_mwait(uint32_t mwait_hint)
223 {
224 	int *state;
225 	uint64_t v;
226 
227 	/*
228 	 * A comment in Linux patch claims that 'CPUs run faster with
229 	 * speculation protection disabled. All CPU threads in a core
230 	 * must disable speculation protection for it to be
231 	 * disabled. Disable it while we are idle so the other
232 	 * hyperthread can run fast.'
233 	 *
234 	 * XXXKIB.  Software coordination mode should be supported,
235 	 * but all Intel CPUs provide hardware coordination.
236 	 */
237 
238 	state = &PCPU_PTR(monitorbuf)->idle_state;
239 	KASSERT(atomic_load_int(state) == STATE_SLEEPING,
240 	    ("cpu_mwait_cx: wrong monitorbuf state"));
241 	atomic_store_int(state, STATE_MWAIT);
242 	if (PCPU_GET(ibpb_set) || hw_ssb_active) {
243 		v = rdmsr(MSR_IA32_SPEC_CTRL);
244 		wrmsr(MSR_IA32_SPEC_CTRL, v & ~(IA32_SPEC_CTRL_IBRS |
245 		    IA32_SPEC_CTRL_STIBP | IA32_SPEC_CTRL_SSBD));
246 	} else {
247 		v = 0;
248 	}
249 	cpu_monitor(state, 0, 0);
250 	if (atomic_load_int(state) == STATE_MWAIT)
251 		cpu_mwait(MWAIT_INTRBREAK, mwait_hint);
252 
253 	/*
254 	 * SSB cannot be disabled while we sleep, or rather, if it was
255 	 * disabled, the sysctl thread will bind to our cpu to tweak
256 	 * MSR.
257 	 */
258 	if (v != 0)
259 		wrmsr(MSR_IA32_SPEC_CTRL, v);
260 
261 	/*
262 	 * We should exit on any event that interrupts mwait, because
263 	 * that event might be a wanted interrupt.
264 	 */
265 	atomic_store_int(state, STATE_RUNNING);
266 }
267 
268 /* Get current clock frequency for the given cpu id. */
269 int
270 cpu_est_clockrate(int cpu_id, uint64_t *rate)
271 {
272 	uint64_t tsc1, tsc2;
273 	uint64_t acnt, mcnt, perf;
274 	register_t reg;
275 
276 	if (pcpu_find(cpu_id) == NULL || rate == NULL)
277 		return (EINVAL);
278 #ifdef __i386__
279 	if ((cpu_feature & CPUID_TSC) == 0)
280 		return (EOPNOTSUPP);
281 #endif
282 
283 	/*
284 	 * If TSC is P-state invariant and APERF/MPERF MSRs do not exist,
285 	 * DELAY(9) based logic fails.
286 	 */
287 	if (tsc_is_invariant && !tsc_perf_stat)
288 		return (EOPNOTSUPP);
289 
290 #ifdef SMP
291 	if (smp_cpus > 1) {
292 		/* Schedule ourselves on the indicated cpu. */
293 		thread_lock(curthread);
294 		sched_bind(curthread, cpu_id);
295 		thread_unlock(curthread);
296 	}
297 #endif
298 
299 	/* Calibrate by measuring a short delay. */
300 	reg = intr_disable();
301 	if (tsc_is_invariant) {
302 		wrmsr(MSR_MPERF, 0);
303 		wrmsr(MSR_APERF, 0);
304 		tsc1 = rdtsc();
305 		DELAY(1000);
306 		mcnt = rdmsr(MSR_MPERF);
307 		acnt = rdmsr(MSR_APERF);
308 		tsc2 = rdtsc();
309 		intr_restore(reg);
310 		perf = 1000 * acnt / mcnt;
311 		*rate = (tsc2 - tsc1) * perf;
312 	} else {
313 		tsc1 = rdtsc();
314 		DELAY(1000);
315 		tsc2 = rdtsc();
316 		intr_restore(reg);
317 		*rate = (tsc2 - tsc1) * 1000;
318 	}
319 
320 #ifdef SMP
321 	if (smp_cpus > 1) {
322 		thread_lock(curthread);
323 		sched_unbind(curthread);
324 		thread_unlock(curthread);
325 	}
326 #endif
327 
328 	return (0);
329 }
330 
331 /*
332  * Shutdown the CPU as much as possible
333  */
334 void
335 cpu_halt(void)
336 {
337 	for (;;)
338 		halt();
339 }
340 
341 static void
342 cpu_reset_real(void)
343 {
344 	struct region_descriptor null_idt;
345 	int b;
346 
347 	disable_intr();
348 #ifdef CPU_ELAN
349 	if (elan_mmcr != NULL)
350 		elan_mmcr->RESCFG = 1;
351 #endif
352 #ifdef __i386__
353 	if (cpu == CPU_GEODE1100) {
354 		/* Attempt Geode's own reset */
355 		outl(0xcf8, 0x80009044ul);
356 		outl(0xcfc, 0xf);
357 	}
358 #endif
359 #if !defined(BROKEN_KEYBOARD_RESET)
360 	/*
361 	 * Attempt to do a CPU reset via the keyboard controller,
362 	 * do not turn off GateA20, as any machine that fails
363 	 * to do the reset here would then end up in no man's land.
364 	 */
365 	outb(IO_KBD + 4, 0xFE);
366 	DELAY(500000);	/* wait 0.5 sec to see if that did it */
367 #endif
368 
369 	/*
370 	 * Attempt to force a reset via the Reset Control register at
371 	 * I/O port 0xcf9.  Bit 2 forces a system reset when it
372 	 * transitions from 0 to 1.  Bit 1 selects the type of reset
373 	 * to attempt: 0 selects a "soft" reset, and 1 selects a
374 	 * "hard" reset.  We try a "hard" reset.  The first write sets
375 	 * bit 1 to select a "hard" reset and clears bit 2.  The
376 	 * second write forces a 0 -> 1 transition in bit 2 to trigger
377 	 * a reset.
378 	 */
379 	outb(0xcf9, 0x2);
380 	outb(0xcf9, 0x6);
381 	DELAY(500000);  /* wait 0.5 sec to see if that did it */
382 
383 	/*
384 	 * Attempt to force a reset via the Fast A20 and Init register
385 	 * at I/O port 0x92.  Bit 1 serves as an alternate A20 gate.
386 	 * Bit 0 asserts INIT# when set to 1.  We are careful to only
387 	 * preserve bit 1 while setting bit 0.  We also must clear bit
388 	 * 0 before setting it if it isn't already clear.
389 	 */
390 	b = inb(0x92);
391 	if (b != 0xff) {
392 		if ((b & 0x1) != 0)
393 			outb(0x92, b & 0xfe);
394 		outb(0x92, b | 0x1);
395 		DELAY(500000);  /* wait 0.5 sec to see if that did it */
396 	}
397 
398 	printf("No known reset method worked, attempting CPU shutdown\n");
399 	DELAY(1000000); /* wait 1 sec for printf to complete */
400 
401 	/* Wipe the IDT. */
402 	null_idt.rd_limit = 0;
403 	null_idt.rd_base = 0;
404 	lidt(&null_idt);
405 
406 	/* "good night, sweet prince .... <THUNK!>" */
407 	breakpoint();
408 
409 	/* NOTREACHED */
410 	while(1);
411 }
412 
413 #ifdef SMP
414 static void
415 cpu_reset_proxy(void)
416 {
417 
418 	cpu_reset_proxy_active = 1;
419 	while (cpu_reset_proxy_active == 1)
420 		ia32_pause(); /* Wait for other cpu to see that we've started */
421 
422 	printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid);
423 	DELAY(1000000);
424 	cpu_reset_real();
425 }
426 #endif
427 
428 void
429 cpu_reset(void)
430 {
431 #ifdef SMP
432 	struct monitorbuf *mb;
433 	cpuset_t map;
434 	u_int cnt;
435 
436 	if (smp_started) {
437 		map = all_cpus;
438 		CPU_CLR(PCPU_GET(cpuid), &map);
439 		CPU_ANDNOT(&map, &stopped_cpus);
440 		if (!CPU_EMPTY(&map)) {
441 			printf("cpu_reset: Stopping other CPUs\n");
442 			stop_cpus(map);
443 		}
444 
445 		if (PCPU_GET(cpuid) != 0) {
446 			cpu_reset_proxyid = PCPU_GET(cpuid);
447 			cpustop_restartfunc = cpu_reset_proxy;
448 			cpu_reset_proxy_active = 0;
449 			printf("cpu_reset: Restarting BSP\n");
450 
451 			/* Restart CPU #0. */
452 			CPU_SETOF(0, &started_cpus);
453 			mb = &pcpu_find(0)->pc_monitorbuf;
454 			atomic_store_int(&mb->stop_state,
455 			    MONITOR_STOPSTATE_RUNNING);
456 
457 			cnt = 0;
458 			while (cpu_reset_proxy_active == 0 && cnt < 10000000) {
459 				ia32_pause();
460 				cnt++;	/* Wait for BSP to announce restart */
461 			}
462 			if (cpu_reset_proxy_active == 0) {
463 				printf("cpu_reset: Failed to restart BSP\n");
464 			} else {
465 				cpu_reset_proxy_active = 2;
466 				while (1)
467 					ia32_pause();
468 				/* NOTREACHED */
469 			}
470 		}
471 
472 		DELAY(1000000);
473 	}
474 #endif
475 	cpu_reset_real();
476 	/* NOTREACHED */
477 }
478 
479 bool
480 cpu_mwait_usable(void)
481 {
482 
483 	return ((cpu_feature2 & CPUID2_MON) != 0 && ((cpu_mon_mwait_flags &
484 	    (CPUID5_MON_MWAIT_EXT | CPUID5_MWAIT_INTRBREAK)) ==
485 	    (CPUID5_MON_MWAIT_EXT | CPUID5_MWAIT_INTRBREAK)));
486 }
487 
488 void (*cpu_idle_hook)(sbintime_t) = NULL;	/* ACPI idle hook. */
489 
490 int cpu_amdc1e_bug = 0;			/* AMD C1E APIC workaround required. */
491 
492 static int	idle_mwait = 1;		/* Use MONITOR/MWAIT for short idle. */
493 SYSCTL_INT(_machdep, OID_AUTO, idle_mwait, CTLFLAG_RWTUN, &idle_mwait,
494     0, "Use MONITOR/MWAIT for short idle");
495 
496 static void
497 cpu_idle_acpi(sbintime_t sbt)
498 {
499 	int *state;
500 
501 	state = &PCPU_PTR(monitorbuf)->idle_state;
502 	atomic_store_int(state, STATE_SLEEPING);
503 
504 	/* See comments in cpu_idle_hlt(). */
505 	disable_intr();
506 	if (sched_runnable())
507 		enable_intr();
508 	else if (cpu_idle_hook)
509 		cpu_idle_hook(sbt);
510 	else
511 		acpi_cpu_c1();
512 	atomic_store_int(state, STATE_RUNNING);
513 }
514 
515 static void
516 cpu_idle_hlt(sbintime_t sbt)
517 {
518 	int *state;
519 
520 	state = &PCPU_PTR(monitorbuf)->idle_state;
521 	atomic_store_int(state, STATE_SLEEPING);
522 
523 	/*
524 	 * Since we may be in a critical section from cpu_idle(), if
525 	 * an interrupt fires during that critical section we may have
526 	 * a pending preemption.  If the CPU halts, then that thread
527 	 * may not execute until a later interrupt awakens the CPU.
528 	 * To handle this race, check for a runnable thread after
529 	 * disabling interrupts and immediately return if one is
530 	 * found.  Also, we must absolutely guarentee that hlt is
531 	 * the next instruction after sti.  This ensures that any
532 	 * interrupt that fires after the call to disable_intr() will
533 	 * immediately awaken the CPU from hlt.  Finally, please note
534 	 * that on x86 this works fine because of interrupts enabled only
535 	 * after the instruction following sti takes place, while IF is set
536 	 * to 1 immediately, allowing hlt instruction to acknowledge the
537 	 * interrupt.
538 	 */
539 	disable_intr();
540 	if (sched_runnable())
541 		enable_intr();
542 	else
543 		acpi_cpu_c1();
544 	atomic_store_int(state, STATE_RUNNING);
545 }
546 
547 static void
548 cpu_idle_mwait(sbintime_t sbt)
549 {
550 	int *state;
551 
552 	state = &PCPU_PTR(monitorbuf)->idle_state;
553 	atomic_store_int(state, STATE_MWAIT);
554 
555 	/* See comments in cpu_idle_hlt(). */
556 	disable_intr();
557 	if (sched_runnable()) {
558 		atomic_store_int(state, STATE_RUNNING);
559 		enable_intr();
560 		return;
561 	}
562 
563 	cpu_monitor(state, 0, 0);
564 	if (atomic_load_int(state) == STATE_MWAIT)
565 		__asm __volatile("sti; mwait" : : "a" (MWAIT_C1), "c" (0));
566 	else
567 		enable_intr();
568 	atomic_store_int(state, STATE_RUNNING);
569 }
570 
571 static void
572 cpu_idle_spin(sbintime_t sbt)
573 {
574 	int *state;
575 	int i;
576 
577 	state = &PCPU_PTR(monitorbuf)->idle_state;
578 	atomic_store_int(state, STATE_RUNNING);
579 
580 	/*
581 	 * The sched_runnable() call is racy but as long as there is
582 	 * a loop missing it one time will have just a little impact if any
583 	 * (and it is much better than missing the check at all).
584 	 */
585 	for (i = 0; i < 1000; i++) {
586 		if (sched_runnable())
587 			return;
588 		cpu_spinwait();
589 	}
590 }
591 
592 void (*cpu_idle_fn)(sbintime_t) = cpu_idle_acpi;
593 
594 void
595 cpu_idle(int busy)
596 {
597 	uint64_t msr;
598 	sbintime_t sbt = -1;
599 
600 	CTR2(KTR_SPARE2, "cpu_idle(%d) at %d",
601 	    busy, curcpu);
602 #ifdef MP_WATCHDOG
603 	ap_watchdog(PCPU_GET(cpuid));
604 #endif
605 
606 	/* If we are busy - try to use fast methods. */
607 	if (busy) {
608 		if ((cpu_feature2 & CPUID2_MON) && idle_mwait) {
609 			cpu_idle_mwait(busy);
610 			goto out;
611 		}
612 	}
613 
614 	/* If we have time - switch timers into idle mode. */
615 	if (!busy) {
616 		critical_enter();
617 		sbt = cpu_idleclock();
618 	}
619 
620 	/* Apply AMD APIC timer C1E workaround. */
621 	if (cpu_amdc1e_bug && cpu_disable_c3_sleep) {
622 		msr = rdmsr(MSR_AMDK8_IPM);
623 		if ((msr & (AMDK8_SMIONCMPHALT | AMDK8_C1EONCMPHALT)) != 0)
624 			wrmsr(MSR_AMDK8_IPM, msr & ~(AMDK8_SMIONCMPHALT |
625 			    AMDK8_C1EONCMPHALT));
626 	}
627 
628 	/* Call main idle method. */
629 	cpu_idle_fn(sbt);
630 
631 	/* Switch timers back into active mode. */
632 	if (!busy) {
633 		cpu_activeclock();
634 		critical_exit();
635 	}
636 out:
637 	CTR2(KTR_SPARE2, "cpu_idle(%d) at %d done",
638 	    busy, curcpu);
639 }
640 
641 static int cpu_idle_apl31_workaround;
642 SYSCTL_INT(_machdep, OID_AUTO, idle_apl31, CTLFLAG_RW,
643     &cpu_idle_apl31_workaround, 0,
644     "Apollo Lake APL31 MWAIT bug workaround");
645 
646 int
647 cpu_idle_wakeup(int cpu)
648 {
649 	struct monitorbuf *mb;
650 	int *state;
651 
652 	mb = &pcpu_find(cpu)->pc_monitorbuf;
653 	state = &mb->idle_state;
654 	switch (atomic_load_int(state)) {
655 	case STATE_SLEEPING:
656 		return (0);
657 	case STATE_MWAIT:
658 		atomic_store_int(state, STATE_RUNNING);
659 		return (cpu_idle_apl31_workaround ? 0 : 1);
660 	case STATE_RUNNING:
661 		return (1);
662 	default:
663 		panic("bad monitor state");
664 		return (1);
665 	}
666 }
667 
668 /*
669  * Ordered by speed/power consumption.
670  */
671 static struct {
672 	void	*id_fn;
673 	char	*id_name;
674 	int	id_cpuid2_flag;
675 } idle_tbl[] = {
676 	{ .id_fn = cpu_idle_spin, .id_name = "spin" },
677 	{ .id_fn = cpu_idle_mwait, .id_name = "mwait",
678 	    .id_cpuid2_flag = CPUID2_MON },
679 	{ .id_fn = cpu_idle_hlt, .id_name = "hlt" },
680 	{ .id_fn = cpu_idle_acpi, .id_name = "acpi" },
681 };
682 
683 static int
684 idle_sysctl_available(SYSCTL_HANDLER_ARGS)
685 {
686 	char *avail, *p;
687 	int error;
688 	int i;
689 
690 	avail = malloc(256, M_TEMP, M_WAITOK);
691 	p = avail;
692 	for (i = 0; i < nitems(idle_tbl); i++) {
693 		if (idle_tbl[i].id_cpuid2_flag != 0 &&
694 		    (cpu_feature2 & idle_tbl[i].id_cpuid2_flag) == 0)
695 			continue;
696 		if (strcmp(idle_tbl[i].id_name, "acpi") == 0 &&
697 		    cpu_idle_hook == NULL)
698 			continue;
699 		p += sprintf(p, "%s%s", p != avail ? ", " : "",
700 		    idle_tbl[i].id_name);
701 	}
702 	error = sysctl_handle_string(oidp, avail, 0, req);
703 	free(avail, M_TEMP);
704 	return (error);
705 }
706 
707 SYSCTL_PROC(_machdep, OID_AUTO, idle_available,
708     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
709     0, 0, idle_sysctl_available, "A",
710     "list of available idle functions");
711 
712 static bool
713 cpu_idle_selector(const char *new_idle_name)
714 {
715 	int i;
716 
717 	for (i = 0; i < nitems(idle_tbl); i++) {
718 		if (idle_tbl[i].id_cpuid2_flag != 0 &&
719 		    (cpu_feature2 & idle_tbl[i].id_cpuid2_flag) == 0)
720 			continue;
721 		if (strcmp(idle_tbl[i].id_name, "acpi") == 0 &&
722 		    cpu_idle_hook == NULL)
723 			continue;
724 		if (strcmp(idle_tbl[i].id_name, new_idle_name))
725 			continue;
726 		cpu_idle_fn = idle_tbl[i].id_fn;
727 		if (bootverbose)
728 			printf("CPU idle set to %s\n", idle_tbl[i].id_name);
729 		return (true);
730 	}
731 	return (false);
732 }
733 
734 static int
735 cpu_idle_sysctl(SYSCTL_HANDLER_ARGS)
736 {
737 	char buf[16], *p;
738 	int error, i;
739 
740 	p = "unknown";
741 	for (i = 0; i < nitems(idle_tbl); i++) {
742 		if (idle_tbl[i].id_fn == cpu_idle_fn) {
743 			p = idle_tbl[i].id_name;
744 			break;
745 		}
746 	}
747 	strncpy(buf, p, sizeof(buf));
748 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
749 	if (error != 0 || req->newptr == NULL)
750 		return (error);
751 	return (cpu_idle_selector(buf) ? 0 : EINVAL);
752 }
753 
754 SYSCTL_PROC(_machdep, OID_AUTO, idle,
755     CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
756     0, 0, cpu_idle_sysctl, "A",
757     "currently selected idle function");
758 
759 static void
760 cpu_idle_tun(void *unused __unused)
761 {
762 	char tunvar[16];
763 
764 	if (TUNABLE_STR_FETCH("machdep.idle", tunvar, sizeof(tunvar)))
765 		cpu_idle_selector(tunvar);
766 	else if (cpu_vendor_id == CPU_VENDOR_AMD &&
767 	    CPUID_TO_FAMILY(cpu_id) == 0x17 && CPUID_TO_MODEL(cpu_id) == 0x1) {
768 		/* Ryzen erratas 1057, 1109. */
769 		cpu_idle_selector("hlt");
770 		idle_mwait = 0;
771 		mwait_cpustop_broken = true;
772 	}
773 
774 	if (cpu_vendor_id == CPU_VENDOR_INTEL && cpu_id == 0x506c9) {
775 		/*
776 		 * Apollo Lake errata APL31 (public errata APL30).
777 		 * Stores to the armed address range may not trigger
778 		 * MWAIT to resume execution.  OS needs to use
779 		 * interrupts to wake processors from MWAIT-induced
780 		 * sleep states.
781 		 */
782 		cpu_idle_apl31_workaround = 1;
783 		mwait_cpustop_broken = true;
784 	}
785 	TUNABLE_INT_FETCH("machdep.idle_apl31", &cpu_idle_apl31_workaround);
786 }
787 SYSINIT(cpu_idle_tun, SI_SUB_CPU, SI_ORDER_MIDDLE, cpu_idle_tun, NULL);
788 
789 static int panic_on_nmi = 0xff;
790 SYSCTL_INT(_machdep, OID_AUTO, panic_on_nmi, CTLFLAG_RWTUN,
791     &panic_on_nmi, 0,
792     "Panic on NMI: 1 = H/W failure; 2 = unknown; 0xff = all");
793 int nmi_is_broadcast = 1;
794 SYSCTL_INT(_machdep, OID_AUTO, nmi_is_broadcast, CTLFLAG_RWTUN,
795     &nmi_is_broadcast, 0,
796     "Chipset NMI is broadcast");
797 int (*apei_nmi)(void);
798 
799 void
800 nmi_call_kdb(u_int cpu, u_int type, struct trapframe *frame)
801 {
802 	bool claimed = false;
803 
804 #ifdef DEV_ISA
805 	/* machine/parity/power fail/"kitchen sink" faults */
806 	if (isa_nmi(frame->tf_err)) {
807 		claimed = true;
808 		if ((panic_on_nmi & 1) != 0)
809 			panic("NMI indicates hardware failure");
810 	}
811 #endif /* DEV_ISA */
812 
813 	/* ACPI Platform Error Interfaces callback. */
814 	if (apei_nmi != NULL && (*apei_nmi)())
815 		claimed = true;
816 
817 	/*
818 	 * NMIs can be useful for debugging.  They can be hooked up to a
819 	 * pushbutton, usually on an ISA, PCI, or PCIe card.  They can also be
820 	 * generated by an IPMI BMC, either manually or in response to a
821 	 * watchdog timeout.  For example, see the "power diag" command in
822 	 * ports/sysutils/ipmitool.  They can also be generated by a
823 	 * hypervisor; see "bhyvectl --inject-nmi".
824 	 */
825 
826 #ifdef KDB
827 	if (!claimed && (panic_on_nmi & 2) != 0) {
828 		if (debugger_on_panic) {
829 			printf("NMI/cpu%d ... going to debugger\n", cpu);
830 			claimed = kdb_trap(type, 0, frame);
831 		}
832 	}
833 #endif /* KDB */
834 
835 	if (!claimed && panic_on_nmi != 0)
836 		panic("NMI");
837 }
838 
839 void
840 nmi_handle_intr(u_int type, struct trapframe *frame)
841 {
842 
843 #ifdef SMP
844 	if (nmi_is_broadcast) {
845 		nmi_call_kdb_smp(type, frame);
846 		return;
847 	}
848 #endif
849 	nmi_call_kdb(PCPU_GET(cpuid), type, frame);
850 }
851 
852 static int hw_ibrs_active;
853 int hw_ibrs_ibpb_active;
854 int hw_ibrs_disable = 1;
855 
856 SYSCTL_INT(_hw, OID_AUTO, ibrs_active, CTLFLAG_RD, &hw_ibrs_active, 0,
857     "Indirect Branch Restricted Speculation active");
858 
859 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, ibrs,
860     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
861     "Indirect Branch Restricted Speculation active");
862 
863 SYSCTL_INT(_machdep_mitigations_ibrs, OID_AUTO, active, CTLFLAG_RD,
864     &hw_ibrs_active, 0, "Indirect Branch Restricted Speculation active");
865 
866 void
867 hw_ibrs_recalculate(bool for_all_cpus)
868 {
869 	if ((cpu_ia32_arch_caps & IA32_ARCH_CAP_IBRS_ALL) != 0) {
870 		x86_msr_op(MSR_IA32_SPEC_CTRL, (for_all_cpus ?
871 		    MSR_OP_RENDEZVOUS : MSR_OP_LOCAL) |
872 		    (hw_ibrs_disable != 0 ? MSR_OP_ANDNOT : MSR_OP_OR),
873 		    IA32_SPEC_CTRL_IBRS);
874 		hw_ibrs_active = hw_ibrs_disable == 0;
875 		hw_ibrs_ibpb_active = 0;
876 	} else {
877 		hw_ibrs_active = hw_ibrs_ibpb_active = (cpu_stdext_feature3 &
878 		    CPUID_STDEXT3_IBPB) != 0 && !hw_ibrs_disable;
879 	}
880 }
881 
882 static int
883 hw_ibrs_disable_handler(SYSCTL_HANDLER_ARGS)
884 {
885 	int error, val;
886 
887 	val = hw_ibrs_disable;
888 	error = sysctl_handle_int(oidp, &val, 0, req);
889 	if (error != 0 || req->newptr == NULL)
890 		return (error);
891 	hw_ibrs_disable = val != 0;
892 	hw_ibrs_recalculate(true);
893 	return (0);
894 }
895 SYSCTL_PROC(_hw, OID_AUTO, ibrs_disable, CTLTYPE_INT | CTLFLAG_RWTUN |
896     CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0, hw_ibrs_disable_handler, "I",
897     "Disable Indirect Branch Restricted Speculation");
898 
899 SYSCTL_PROC(_machdep_mitigations_ibrs, OID_AUTO, disable, CTLTYPE_INT |
900     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
901     hw_ibrs_disable_handler, "I",
902     "Disable Indirect Branch Restricted Speculation");
903 
904 int hw_ssb_active;
905 int hw_ssb_disable;
906 
907 SYSCTL_INT(_hw, OID_AUTO, spec_store_bypass_disable_active, CTLFLAG_RD,
908     &hw_ssb_active, 0,
909     "Speculative Store Bypass Disable active");
910 
911 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, ssb,
912     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
913     "Speculative Store Bypass Disable active");
914 
915 SYSCTL_INT(_machdep_mitigations_ssb, OID_AUTO, active, CTLFLAG_RD,
916     &hw_ssb_active, 0, "Speculative Store Bypass Disable active");
917 
918 static void
919 hw_ssb_set(bool enable, bool for_all_cpus)
920 {
921 
922 	if ((cpu_stdext_feature3 & CPUID_STDEXT3_SSBD) == 0) {
923 		hw_ssb_active = 0;
924 		return;
925 	}
926 	hw_ssb_active = enable;
927 	x86_msr_op(MSR_IA32_SPEC_CTRL,
928 	    (enable ? MSR_OP_OR : MSR_OP_ANDNOT) |
929 	    (for_all_cpus ? MSR_OP_SCHED : MSR_OP_LOCAL), IA32_SPEC_CTRL_SSBD);
930 }
931 
932 void
933 hw_ssb_recalculate(bool all_cpus)
934 {
935 
936 	switch (hw_ssb_disable) {
937 	default:
938 		hw_ssb_disable = 0;
939 		/* FALLTHROUGH */
940 	case 0: /* off */
941 		hw_ssb_set(false, all_cpus);
942 		break;
943 	case 1: /* on */
944 		hw_ssb_set(true, all_cpus);
945 		break;
946 	case 2: /* auto */
947 		hw_ssb_set((cpu_ia32_arch_caps & IA32_ARCH_CAP_SSB_NO) != 0 ?
948 		    false : true, all_cpus);
949 		break;
950 	}
951 }
952 
953 static int
954 hw_ssb_disable_handler(SYSCTL_HANDLER_ARGS)
955 {
956 	int error, val;
957 
958 	val = hw_ssb_disable;
959 	error = sysctl_handle_int(oidp, &val, 0, req);
960 	if (error != 0 || req->newptr == NULL)
961 		return (error);
962 	hw_ssb_disable = val;
963 	hw_ssb_recalculate(true);
964 	return (0);
965 }
966 SYSCTL_PROC(_hw, OID_AUTO, spec_store_bypass_disable, CTLTYPE_INT |
967     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
968     hw_ssb_disable_handler, "I",
969     "Speculative Store Bypass Disable (0 - off, 1 - on, 2 - auto");
970 
971 SYSCTL_PROC(_machdep_mitigations_ssb, OID_AUTO, disable, CTLTYPE_INT |
972     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
973     hw_ssb_disable_handler, "I",
974     "Speculative Store Bypass Disable (0 - off, 1 - on, 2 - auto");
975 
976 int hw_mds_disable;
977 
978 /*
979  * Handler for Microarchitectural Data Sampling issues.  Really not a
980  * pointer to C function: on amd64 the code must not change any CPU
981  * architectural state except possibly %rflags. Also, it is always
982  * called with interrupts disabled.
983  */
984 void mds_handler_void(void);
985 void mds_handler_verw(void);
986 void mds_handler_ivb(void);
987 void mds_handler_bdw(void);
988 void mds_handler_skl_sse(void);
989 void mds_handler_skl_avx(void);
990 void mds_handler_skl_avx512(void);
991 void mds_handler_silvermont(void);
992 void (*mds_handler)(void) = mds_handler_void;
993 
994 static int
995 sysctl_hw_mds_disable_state_handler(SYSCTL_HANDLER_ARGS)
996 {
997 	const char *state;
998 
999 	if (mds_handler == mds_handler_void)
1000 		state = "inactive";
1001 	else if (mds_handler == mds_handler_verw)
1002 		state = "VERW";
1003 	else if (mds_handler == mds_handler_ivb)
1004 		state = "software IvyBridge";
1005 	else if (mds_handler == mds_handler_bdw)
1006 		state = "software Broadwell";
1007 	else if (mds_handler == mds_handler_skl_sse)
1008 		state = "software Skylake SSE";
1009 	else if (mds_handler == mds_handler_skl_avx)
1010 		state = "software Skylake AVX";
1011 	else if (mds_handler == mds_handler_skl_avx512)
1012 		state = "software Skylake AVX512";
1013 	else if (mds_handler == mds_handler_silvermont)
1014 		state = "software Silvermont";
1015 	else
1016 		state = "unknown";
1017 	return (SYSCTL_OUT(req, state, strlen(state)));
1018 }
1019 
1020 SYSCTL_PROC(_hw, OID_AUTO, mds_disable_state,
1021     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1022     sysctl_hw_mds_disable_state_handler, "A",
1023     "Microarchitectural Data Sampling Mitigation state");
1024 
1025 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, mds,
1026     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1027     "Microarchitectural Data Sampling Mitigation state");
1028 
1029 SYSCTL_PROC(_machdep_mitigations_mds, OID_AUTO, state,
1030     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1031     sysctl_hw_mds_disable_state_handler, "A",
1032     "Microarchitectural Data Sampling Mitigation state");
1033 
1034 _Static_assert(__offsetof(struct pcpu, pc_mds_tmp) % 64 == 0, "MDS AVX512");
1035 
1036 void
1037 hw_mds_recalculate(void)
1038 {
1039 	struct pcpu *pc;
1040 	vm_offset_t b64;
1041 	u_long xcr0;
1042 	int i;
1043 
1044 	/*
1045 	 * Allow user to force VERW variant even if MD_CLEAR is not
1046 	 * reported.  For instance, hypervisor might unknowingly
1047 	 * filter the cap out.
1048 	 * For the similar reasons, and for testing, allow to enable
1049 	 * mitigation even when MDS_NO cap is set.
1050 	 */
1051 	if (cpu_vendor_id != CPU_VENDOR_INTEL || hw_mds_disable == 0 ||
1052 	    ((cpu_ia32_arch_caps & IA32_ARCH_CAP_MDS_NO) != 0 &&
1053 	    hw_mds_disable == 3)) {
1054 		mds_handler = mds_handler_void;
1055 	} else if (((cpu_stdext_feature3 & CPUID_STDEXT3_MD_CLEAR) != 0 &&
1056 	    hw_mds_disable == 3) || hw_mds_disable == 1) {
1057 		mds_handler = mds_handler_verw;
1058 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1059 	    (CPUID_TO_MODEL(cpu_id) == 0x2e || CPUID_TO_MODEL(cpu_id) == 0x1e ||
1060 	    CPUID_TO_MODEL(cpu_id) == 0x1f || CPUID_TO_MODEL(cpu_id) == 0x1a ||
1061 	    CPUID_TO_MODEL(cpu_id) == 0x2f || CPUID_TO_MODEL(cpu_id) == 0x25 ||
1062 	    CPUID_TO_MODEL(cpu_id) == 0x2c || CPUID_TO_MODEL(cpu_id) == 0x2d ||
1063 	    CPUID_TO_MODEL(cpu_id) == 0x2a || CPUID_TO_MODEL(cpu_id) == 0x3e ||
1064 	    CPUID_TO_MODEL(cpu_id) == 0x3a) &&
1065 	    (hw_mds_disable == 2 || hw_mds_disable == 3)) {
1066 		/*
1067 		 * Nehalem, SandyBridge, IvyBridge
1068 		 */
1069 		CPU_FOREACH(i) {
1070 			pc = pcpu_find(i);
1071 			if (pc->pc_mds_buf == NULL) {
1072 				pc->pc_mds_buf = malloc_domainset(672, M_TEMP,
1073 				    DOMAINSET_PREF(pc->pc_domain), M_WAITOK);
1074 				bzero(pc->pc_mds_buf, 16);
1075 			}
1076 		}
1077 		mds_handler = mds_handler_ivb;
1078 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1079 	    (CPUID_TO_MODEL(cpu_id) == 0x3f || CPUID_TO_MODEL(cpu_id) == 0x3c ||
1080 	    CPUID_TO_MODEL(cpu_id) == 0x45 || CPUID_TO_MODEL(cpu_id) == 0x46 ||
1081 	    CPUID_TO_MODEL(cpu_id) == 0x56 || CPUID_TO_MODEL(cpu_id) == 0x4f ||
1082 	    CPUID_TO_MODEL(cpu_id) == 0x47 || CPUID_TO_MODEL(cpu_id) == 0x3d) &&
1083 	    (hw_mds_disable == 2 || hw_mds_disable == 3)) {
1084 		/*
1085 		 * Haswell, Broadwell
1086 		 */
1087 		CPU_FOREACH(i) {
1088 			pc = pcpu_find(i);
1089 			if (pc->pc_mds_buf == NULL) {
1090 				pc->pc_mds_buf = malloc_domainset(1536, M_TEMP,
1091 				    DOMAINSET_PREF(pc->pc_domain), M_WAITOK);
1092 				bzero(pc->pc_mds_buf, 16);
1093 			}
1094 		}
1095 		mds_handler = mds_handler_bdw;
1096 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1097 	    ((CPUID_TO_MODEL(cpu_id) == 0x55 && (cpu_id &
1098 	    CPUID_STEPPING) <= 5) ||
1099 	    CPUID_TO_MODEL(cpu_id) == 0x4e || CPUID_TO_MODEL(cpu_id) == 0x5e ||
1100 	    (CPUID_TO_MODEL(cpu_id) == 0x8e && (cpu_id &
1101 	    CPUID_STEPPING) <= 0xb) ||
1102 	    (CPUID_TO_MODEL(cpu_id) == 0x9e && (cpu_id &
1103 	    CPUID_STEPPING) <= 0xc)) &&
1104 	    (hw_mds_disable == 2 || hw_mds_disable == 3)) {
1105 		/*
1106 		 * Skylake, KabyLake, CoffeeLake, WhiskeyLake,
1107 		 * CascadeLake
1108 		 */
1109 		CPU_FOREACH(i) {
1110 			pc = pcpu_find(i);
1111 			if (pc->pc_mds_buf == NULL) {
1112 				pc->pc_mds_buf = malloc_domainset(6 * 1024,
1113 				    M_TEMP, DOMAINSET_PREF(pc->pc_domain),
1114 				    M_WAITOK);
1115 				b64 = (vm_offset_t)malloc_domainset(64 + 63,
1116 				    M_TEMP, DOMAINSET_PREF(pc->pc_domain),
1117 				    M_WAITOK);
1118 				pc->pc_mds_buf64 = (void *)roundup2(b64, 64);
1119 				bzero(pc->pc_mds_buf64, 64);
1120 			}
1121 		}
1122 		xcr0 = rxcr(0);
1123 		if ((xcr0 & XFEATURE_ENABLED_ZMM_HI256) != 0 &&
1124 		    (cpu_stdext_feature & CPUID_STDEXT_AVX512DQ) != 0)
1125 			mds_handler = mds_handler_skl_avx512;
1126 		else if ((xcr0 & XFEATURE_ENABLED_AVX) != 0 &&
1127 		    (cpu_feature2 & CPUID2_AVX) != 0)
1128 			mds_handler = mds_handler_skl_avx;
1129 		else
1130 			mds_handler = mds_handler_skl_sse;
1131 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1132 	    ((CPUID_TO_MODEL(cpu_id) == 0x37 ||
1133 	    CPUID_TO_MODEL(cpu_id) == 0x4a ||
1134 	    CPUID_TO_MODEL(cpu_id) == 0x4c ||
1135 	    CPUID_TO_MODEL(cpu_id) == 0x4d ||
1136 	    CPUID_TO_MODEL(cpu_id) == 0x5a ||
1137 	    CPUID_TO_MODEL(cpu_id) == 0x5d ||
1138 	    CPUID_TO_MODEL(cpu_id) == 0x6e ||
1139 	    CPUID_TO_MODEL(cpu_id) == 0x65 ||
1140 	    CPUID_TO_MODEL(cpu_id) == 0x75 ||
1141 	    CPUID_TO_MODEL(cpu_id) == 0x1c ||
1142 	    CPUID_TO_MODEL(cpu_id) == 0x26 ||
1143 	    CPUID_TO_MODEL(cpu_id) == 0x27 ||
1144 	    CPUID_TO_MODEL(cpu_id) == 0x35 ||
1145 	    CPUID_TO_MODEL(cpu_id) == 0x36 ||
1146 	    CPUID_TO_MODEL(cpu_id) == 0x7a))) {
1147 		/* Silvermont, Airmont */
1148 		CPU_FOREACH(i) {
1149 			pc = pcpu_find(i);
1150 			if (pc->pc_mds_buf == NULL)
1151 				pc->pc_mds_buf = malloc(256, M_TEMP, M_WAITOK);
1152 		}
1153 		mds_handler = mds_handler_silvermont;
1154 	} else {
1155 		hw_mds_disable = 0;
1156 		mds_handler = mds_handler_void;
1157 	}
1158 }
1159 
1160 static void
1161 hw_mds_recalculate_boot(void *arg __unused)
1162 {
1163 
1164 	hw_mds_recalculate();
1165 }
1166 SYSINIT(mds_recalc, SI_SUB_SMP, SI_ORDER_ANY, hw_mds_recalculate_boot, NULL);
1167 
1168 static int
1169 sysctl_mds_disable_handler(SYSCTL_HANDLER_ARGS)
1170 {
1171 	int error, val;
1172 
1173 	val = hw_mds_disable;
1174 	error = sysctl_handle_int(oidp, &val, 0, req);
1175 	if (error != 0 || req->newptr == NULL)
1176 		return (error);
1177 	if (val < 0 || val > 3)
1178 		return (EINVAL);
1179 	hw_mds_disable = val;
1180 	hw_mds_recalculate();
1181 	return (0);
1182 }
1183 
1184 SYSCTL_PROC(_hw, OID_AUTO, mds_disable, CTLTYPE_INT |
1185     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1186     sysctl_mds_disable_handler, "I",
1187     "Microarchitectural Data Sampling Mitigation "
1188     "(0 - off, 1 - on VERW, 2 - on SW, 3 - on AUTO");
1189 
1190 SYSCTL_PROC(_machdep_mitigations_mds, OID_AUTO, disable, CTLTYPE_INT |
1191     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1192     sysctl_mds_disable_handler, "I",
1193     "Microarchitectural Data Sampling Mitigation "
1194     "(0 - off, 1 - on VERW, 2 - on SW, 3 - on AUTO");
1195 
1196 /*
1197  * Intel Transactional Memory Asynchronous Abort Mitigation
1198  * CVE-2019-11135
1199  */
1200 int x86_taa_enable;
1201 int x86_taa_state;
1202 enum {
1203 	TAA_NONE	= 0,	/* No mitigation enabled */
1204 	TAA_TSX_DISABLE	= 1,	/* Disable TSX via MSR */
1205 	TAA_VERW	= 2,	/* Use VERW mitigation */
1206 	TAA_AUTO	= 3,	/* Automatically select the mitigation */
1207 
1208 	/* The states below are not selectable by the operator */
1209 
1210 	TAA_TAA_UC	= 4,	/* Mitigation present in microcode */
1211 	TAA_NOT_PRESENT	= 5	/* TSX is not present */
1212 };
1213 
1214 static void
1215 taa_set(bool enable, bool all)
1216 {
1217 
1218 	x86_msr_op(MSR_IA32_TSX_CTRL,
1219 	    (enable ? MSR_OP_OR : MSR_OP_ANDNOT) |
1220 	    (all ? MSR_OP_RENDEZVOUS : MSR_OP_LOCAL),
1221 	    IA32_TSX_CTRL_RTM_DISABLE | IA32_TSX_CTRL_TSX_CPUID_CLEAR);
1222 }
1223 
1224 void
1225 x86_taa_recalculate(void)
1226 {
1227 	static int taa_saved_mds_disable = 0;
1228 	int taa_need = 0, taa_state = 0;
1229 	int mds_disable = 0, need_mds_recalc = 0;
1230 
1231 	/* Check CPUID.07h.EBX.HLE and RTM for the presence of TSX */
1232 	if ((cpu_stdext_feature & CPUID_STDEXT_HLE) == 0 ||
1233 	    (cpu_stdext_feature & CPUID_STDEXT_RTM) == 0) {
1234 		/* TSX is not present */
1235 		x86_taa_state = TAA_NOT_PRESENT;
1236 		return;
1237 	}
1238 
1239 	/* Check to see what mitigation options the CPU gives us */
1240 	if (cpu_ia32_arch_caps & IA32_ARCH_CAP_TAA_NO) {
1241 		/* CPU is not suseptible to TAA */
1242 		taa_need = TAA_TAA_UC;
1243 	} else if (cpu_ia32_arch_caps & IA32_ARCH_CAP_TSX_CTRL) {
1244 		/*
1245 		 * CPU can turn off TSX.  This is the next best option
1246 		 * if TAA_NO hardware mitigation isn't present
1247 		 */
1248 		taa_need = TAA_TSX_DISABLE;
1249 	} else {
1250 		/* No TSX/TAA specific remedies are available. */
1251 		if (x86_taa_enable == TAA_TSX_DISABLE) {
1252 			if (bootverbose)
1253 				printf("TSX control not available\n");
1254 			return;
1255 		} else
1256 			taa_need = TAA_VERW;
1257 	}
1258 
1259 	/* Can we automatically take action, or are we being forced? */
1260 	if (x86_taa_enable == TAA_AUTO)
1261 		taa_state = taa_need;
1262 	else
1263 		taa_state = x86_taa_enable;
1264 
1265 	/* No state change, nothing to do */
1266 	if (taa_state == x86_taa_state) {
1267 		if (bootverbose)
1268 			printf("No TSX change made\n");
1269 		return;
1270 	}
1271 
1272 	/* Does the MSR need to be turned on or off? */
1273 	if (taa_state == TAA_TSX_DISABLE)
1274 		taa_set(true, true);
1275 	else if (x86_taa_state == TAA_TSX_DISABLE)
1276 		taa_set(false, true);
1277 
1278 	/* Does MDS need to be set to turn on VERW? */
1279 	if (taa_state == TAA_VERW) {
1280 		taa_saved_mds_disable = hw_mds_disable;
1281 		mds_disable = hw_mds_disable = 1;
1282 		need_mds_recalc = 1;
1283 	} else if (x86_taa_state == TAA_VERW) {
1284 		mds_disable = hw_mds_disable = taa_saved_mds_disable;
1285 		need_mds_recalc = 1;
1286 	}
1287 	if (need_mds_recalc) {
1288 		hw_mds_recalculate();
1289 		if (mds_disable != hw_mds_disable) {
1290 			if (bootverbose)
1291 				printf("Cannot change MDS state for TAA\n");
1292 			/* Don't update our state */
1293 			return;
1294 		}
1295 	}
1296 
1297 	x86_taa_state = taa_state;
1298 	return;
1299 }
1300 
1301 static void
1302 taa_recalculate_boot(void * arg __unused)
1303 {
1304 
1305 	x86_taa_recalculate();
1306 }
1307 SYSINIT(taa_recalc, SI_SUB_SMP, SI_ORDER_ANY, taa_recalculate_boot, NULL);
1308 
1309 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, taa,
1310     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1311     "TSX Asynchronous Abort Mitigation");
1312 
1313 static int
1314 sysctl_taa_handler(SYSCTL_HANDLER_ARGS)
1315 {
1316 	int error, val;
1317 
1318 	val = x86_taa_enable;
1319 	error = sysctl_handle_int(oidp, &val, 0, req);
1320 	if (error != 0 || req->newptr == NULL)
1321 		return (error);
1322 	if (val < TAA_NONE || val > TAA_AUTO)
1323 		return (EINVAL);
1324 	x86_taa_enable = val;
1325 	x86_taa_recalculate();
1326 	return (0);
1327 }
1328 
1329 SYSCTL_PROC(_machdep_mitigations_taa, OID_AUTO, enable, CTLTYPE_INT |
1330     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1331     sysctl_taa_handler, "I",
1332     "TAA Mitigation enablement control "
1333     "(0 - off, 1 - disable TSX, 2 - VERW, 3 - on AUTO");
1334 
1335 static int
1336 sysctl_taa_state_handler(SYSCTL_HANDLER_ARGS)
1337 {
1338 	const char *state;
1339 
1340 	switch (x86_taa_state) {
1341 	case TAA_NONE:
1342 		state = "inactive";
1343 		break;
1344 	case TAA_TSX_DISABLE:
1345 		state = "TSX disabled";
1346 		break;
1347 	case TAA_VERW:
1348 		state = "VERW";
1349 		break;
1350 	case TAA_TAA_UC:
1351 		state = "Mitigated in microcode";
1352 		break;
1353 	case TAA_NOT_PRESENT:
1354 		state = "TSX not present";
1355 		break;
1356 	default:
1357 		state = "unknown";
1358 	}
1359 
1360 	return (SYSCTL_OUT(req, state, strlen(state)));
1361 }
1362 
1363 SYSCTL_PROC(_machdep_mitigations_taa, OID_AUTO, state,
1364     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1365     sysctl_taa_state_handler, "A",
1366     "TAA Mitigation state");
1367 
1368 int __read_frequently cpu_flush_rsb_ctxsw;
1369 SYSCTL_INT(_machdep_mitigations, OID_AUTO, flush_rsb_ctxsw,
1370     CTLFLAG_RW | CTLFLAG_NOFETCH, &cpu_flush_rsb_ctxsw, 0,
1371     "Flush Return Stack Buffer on context switch");
1372 
1373 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, rngds,
1374     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1375     "MCU Optimization, disable RDSEED mitigation");
1376 
1377 int x86_rngds_mitg_enable = 1;
1378 void
1379 x86_rngds_mitg_recalculate(bool all_cpus)
1380 {
1381 	if ((cpu_stdext_feature3 & CPUID_STDEXT3_MCUOPT) == 0)
1382 		return;
1383 	x86_msr_op(MSR_IA32_MCU_OPT_CTRL,
1384 	    (x86_rngds_mitg_enable ? MSR_OP_OR : MSR_OP_ANDNOT) |
1385 	    (all_cpus ? MSR_OP_RENDEZVOUS : MSR_OP_LOCAL),
1386 	    IA32_RNGDS_MITG_DIS);
1387 }
1388 
1389 static int
1390 sysctl_rngds_mitg_enable_handler(SYSCTL_HANDLER_ARGS)
1391 {
1392 	int error, val;
1393 
1394 	val = x86_rngds_mitg_enable;
1395 	error = sysctl_handle_int(oidp, &val, 0, req);
1396 	if (error != 0 || req->newptr == NULL)
1397 		return (error);
1398 	x86_rngds_mitg_enable = val;
1399 	x86_rngds_mitg_recalculate(true);
1400 	return (0);
1401 }
1402 SYSCTL_PROC(_machdep_mitigations_rngds, OID_AUTO, enable, CTLTYPE_INT |
1403     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1404     sysctl_rngds_mitg_enable_handler, "I",
1405     "MCU Optimization, disabling RDSEED mitigation control "
1406     "(0 - mitigation disabled (RDSEED optimized), 1 - mitigation enabled");
1407 
1408 static int
1409 sysctl_rngds_state_handler(SYSCTL_HANDLER_ARGS)
1410 {
1411 	const char *state;
1412 
1413 	if ((cpu_stdext_feature3 & CPUID_STDEXT3_MCUOPT) == 0) {
1414 		state = "Not applicable";
1415 	} else if (x86_rngds_mitg_enable == 0) {
1416 		state = "RDSEED not serialized";
1417 	} else {
1418 		state = "Mitigated";
1419 	}
1420 	return (SYSCTL_OUT(req, state, strlen(state)));
1421 }
1422 SYSCTL_PROC(_machdep_mitigations_rngds, OID_AUTO, state,
1423     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1424     sysctl_rngds_state_handler, "A",
1425     "MCU Optimization state");
1426 
1427 /*
1428  * Enable and restore kernel text write permissions.
1429  * Callers must ensure that disable_wp()/restore_wp() are executed
1430  * without rescheduling on the same core.
1431  */
1432 bool
1433 disable_wp(void)
1434 {
1435 	u_int cr0;
1436 
1437 	cr0 = rcr0();
1438 	if ((cr0 & CR0_WP) == 0)
1439 		return (false);
1440 	load_cr0(cr0 & ~CR0_WP);
1441 	return (true);
1442 }
1443 
1444 void
1445 restore_wp(bool old_wp)
1446 {
1447 
1448 	if (old_wp)
1449 		load_cr0(rcr0() | CR0_WP);
1450 }
1451 
1452 bool
1453 acpi_get_fadt_bootflags(uint16_t *flagsp)
1454 {
1455 #ifdef DEV_ACPI
1456 	ACPI_TABLE_FADT *fadt;
1457 	vm_paddr_t physaddr;
1458 
1459 	physaddr = acpi_find_table(ACPI_SIG_FADT);
1460 	if (physaddr == 0)
1461 		return (false);
1462 	fadt = acpi_map_table(physaddr, ACPI_SIG_FADT);
1463 	if (fadt == NULL)
1464 		return (false);
1465 	*flagsp = fadt->BootFlags;
1466 	acpi_unmap_table(fadt);
1467 	return (true);
1468 #else
1469 	return (false);
1470 #endif
1471 }
1472