xref: /freebsd/sys/x86/x86/cpu_machdep.c (revision 9eb0e8326d0fe73ae947959c1df327238d3b2d53)
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_platform.h"
54 #include "opt_sched.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/cpufunc.h>
80 #include <machine/cputypes.h>
81 #include <machine/specialreg.h>
82 #include <machine/md_var.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 #include <x86/ifunc.h>
92 
93 #include <vm/vm.h>
94 #include <vm/vm_extern.h>
95 #include <vm/vm_kern.h>
96 #include <vm/vm_page.h>
97 #include <vm/vm_map.h>
98 #include <vm/vm_object.h>
99 #include <vm/vm_pager.h>
100 #include <vm/vm_param.h>
101 
102 #include <isa/isareg.h>
103 
104 #include <contrib/dev/acpica/include/acpi.h>
105 
106 #define	STATE_RUNNING	0x0
107 #define	STATE_MWAIT	0x1
108 #define	STATE_SLEEPING	0x2
109 
110 #ifdef SMP
111 static u_int	cpu_reset_proxyid;
112 static volatile u_int	cpu_reset_proxy_active;
113 #endif
114 
115 char bootmethod[16];
116 SYSCTL_STRING(_machdep, OID_AUTO, bootmethod, CTLFLAG_RD, bootmethod, 0,
117     "System firmware boot method");
118 
119 struct msr_op_arg {
120 	u_int msr;
121 	int op;
122 	uint64_t arg1;
123 	uint64_t *res;
124 };
125 
126 static void
127 x86_msr_op_one(void *argp)
128 {
129 	struct msr_op_arg *a;
130 	uint64_t v;
131 
132 	a = argp;
133 	switch (a->op) {
134 	case MSR_OP_ANDNOT:
135 		v = rdmsr(a->msr);
136 		v &= ~a->arg1;
137 		wrmsr(a->msr, v);
138 		break;
139 	case MSR_OP_OR:
140 		v = rdmsr(a->msr);
141 		v |= a->arg1;
142 		wrmsr(a->msr, v);
143 		break;
144 	case MSR_OP_WRITE:
145 		wrmsr(a->msr, a->arg1);
146 		break;
147 	case MSR_OP_READ:
148 		v = rdmsr(a->msr);
149 		*a->res = v;
150 		break;
151 	}
152 }
153 
154 #define	MSR_OP_EXMODE_MASK	0xf0000000
155 #define	MSR_OP_OP_MASK		0x000000ff
156 #define	MSR_OP_GET_CPUID(x)	(((x) & ~MSR_OP_EXMODE_MASK) >> 8)
157 
158 void
159 x86_msr_op(u_int msr, u_int op, uint64_t arg1, uint64_t *res)
160 {
161 	struct thread *td;
162 	struct msr_op_arg a;
163 	cpuset_t set;
164 	u_int exmode;
165 	int bound_cpu, cpu, i, is_bound;
166 
167 	a.op = op & MSR_OP_OP_MASK;
168 	MPASS(a.op == MSR_OP_ANDNOT || a.op == MSR_OP_OR ||
169 	    a.op == MSR_OP_WRITE || a.op == MSR_OP_READ);
170 	exmode = op & MSR_OP_EXMODE_MASK;
171 	MPASS(exmode == MSR_OP_LOCAL || exmode == MSR_OP_SCHED_ALL ||
172 	    exmode == MSR_OP_SCHED_ONE || exmode == MSR_OP_RENDEZVOUS_ALL ||
173 	    exmode == MSR_OP_RENDEZVOUS_ONE);
174 	a.msr = msr;
175 	a.arg1 = arg1;
176 	a.res = res;
177 	switch (exmode) {
178 	case MSR_OP_LOCAL:
179 		x86_msr_op_one(&a);
180 		break;
181 	case MSR_OP_SCHED_ALL:
182 		td = curthread;
183 		thread_lock(td);
184 		is_bound = sched_is_bound(td);
185 		bound_cpu = td->td_oncpu;
186 		CPU_FOREACH(i) {
187 			sched_bind(td, i);
188 			x86_msr_op_one(&a);
189 		}
190 		if (is_bound)
191 			sched_bind(td, bound_cpu);
192 		else
193 			sched_unbind(td);
194 		thread_unlock(td);
195 		break;
196 	case MSR_OP_SCHED_ONE:
197 		td = curthread;
198 		cpu = MSR_OP_GET_CPUID(op);
199 		thread_lock(td);
200 		is_bound = sched_is_bound(td);
201 		bound_cpu = td->td_oncpu;
202 		if (!is_bound || bound_cpu != cpu)
203 			sched_bind(td, cpu);
204 		x86_msr_op_one(&a);
205 		if (is_bound) {
206 			if (bound_cpu != cpu)
207 				sched_bind(td, bound_cpu);
208 		} else {
209 			sched_unbind(td);
210 		}
211 		thread_unlock(td);
212 		break;
213 	case MSR_OP_RENDEZVOUS_ALL:
214 		smp_rendezvous(smp_no_rendezvous_barrier, x86_msr_op_one,
215 		    smp_no_rendezvous_barrier, &a);
216 		break;
217 	case MSR_OP_RENDEZVOUS_ONE:
218 		cpu = MSR_OP_GET_CPUID(op);
219 		CPU_SETOF(cpu, &set);
220 		smp_rendezvous_cpus(set, smp_no_rendezvous_barrier,
221 		    x86_msr_op_one, smp_no_rendezvous_barrier, &a);
222 		break;
223 	}
224 }
225 
226 /*
227  * Automatically initialized per CPU errata in cpu_idle_tun below.
228  */
229 bool mwait_cpustop_broken = false;
230 SYSCTL_BOOL(_machdep, OID_AUTO, mwait_cpustop_broken, CTLFLAG_RDTUN,
231     &mwait_cpustop_broken, 0,
232     "Can not reliably wake MONITOR/MWAIT cpus without interrupts");
233 
234 /*
235  * Flush the D-cache for non-DMA I/O so that the I-cache can
236  * be made coherent later.
237  */
238 void
239 cpu_flush_dcache(void *ptr, size_t len)
240 {
241 	/* Not applicable */
242 }
243 
244 void
245 acpi_cpu_c1(void)
246 {
247 
248 	__asm __volatile("sti; hlt");
249 }
250 
251 /*
252  * Use mwait to pause execution while waiting for an interrupt or
253  * another thread to signal that there is more work.
254  *
255  * NOTE: Interrupts will cause a wakeup; however, this function does
256  * not enable interrupt handling. The caller is responsible to enable
257  * interrupts.
258  */
259 void
260 acpi_cpu_idle_mwait(uint32_t mwait_hint)
261 {
262 	int *state;
263 	uint64_t v;
264 
265 	/*
266 	 * A comment in Linux patch claims that 'CPUs run faster with
267 	 * speculation protection disabled. All CPU threads in a core
268 	 * must disable speculation protection for it to be
269 	 * disabled. Disable it while we are idle so the other
270 	 * hyperthread can run fast.'
271 	 *
272 	 * XXXKIB.  Software coordination mode should be supported,
273 	 * but all Intel CPUs provide hardware coordination.
274 	 */
275 
276 	state = &PCPU_PTR(monitorbuf)->idle_state;
277 	KASSERT(atomic_load_int(state) == STATE_SLEEPING,
278 	    ("cpu_mwait_cx: wrong monitorbuf state"));
279 	atomic_store_int(state, STATE_MWAIT);
280 	if (PCPU_GET(ibpb_set) || hw_ssb_active) {
281 		v = rdmsr(MSR_IA32_SPEC_CTRL);
282 		wrmsr(MSR_IA32_SPEC_CTRL, v & ~(IA32_SPEC_CTRL_IBRS |
283 		    IA32_SPEC_CTRL_STIBP | IA32_SPEC_CTRL_SSBD));
284 	} else {
285 		v = 0;
286 	}
287 	cpu_monitor(state, 0, 0);
288 	if (atomic_load_int(state) == STATE_MWAIT)
289 		cpu_mwait(MWAIT_INTRBREAK, mwait_hint);
290 
291 	/*
292 	 * SSB cannot be disabled while we sleep, or rather, if it was
293 	 * disabled, the sysctl thread will bind to our cpu to tweak
294 	 * MSR.
295 	 */
296 	if (v != 0)
297 		wrmsr(MSR_IA32_SPEC_CTRL, v);
298 
299 	/*
300 	 * We should exit on any event that interrupts mwait, because
301 	 * that event might be a wanted interrupt.
302 	 */
303 	atomic_store_int(state, STATE_RUNNING);
304 }
305 
306 /* Get current clock frequency for the given cpu id. */
307 int
308 cpu_est_clockrate(int cpu_id, uint64_t *rate)
309 {
310 	uint64_t tsc1, tsc2;
311 	uint64_t acnt, mcnt, perf;
312 	register_t reg;
313 
314 	if (pcpu_find(cpu_id) == NULL || rate == NULL)
315 		return (EINVAL);
316 #ifdef __i386__
317 	if ((cpu_feature & CPUID_TSC) == 0)
318 		return (EOPNOTSUPP);
319 #endif
320 
321 	/*
322 	 * If TSC is P-state invariant and APERF/MPERF MSRs do not exist,
323 	 * DELAY(9) based logic fails.
324 	 */
325 	if (tsc_is_invariant && !tsc_perf_stat)
326 		return (EOPNOTSUPP);
327 
328 #ifdef SMP
329 	if (smp_cpus > 1) {
330 		/* Schedule ourselves on the indicated cpu. */
331 		thread_lock(curthread);
332 		sched_bind(curthread, cpu_id);
333 		thread_unlock(curthread);
334 	}
335 #endif
336 
337 	/* Calibrate by measuring a short delay. */
338 	reg = intr_disable();
339 	if (tsc_is_invariant) {
340 		wrmsr(MSR_MPERF, 0);
341 		wrmsr(MSR_APERF, 0);
342 		tsc1 = rdtsc();
343 		DELAY(1000);
344 		mcnt = rdmsr(MSR_MPERF);
345 		acnt = rdmsr(MSR_APERF);
346 		tsc2 = rdtsc();
347 		intr_restore(reg);
348 		perf = 1000 * acnt / mcnt;
349 		*rate = (tsc2 - tsc1) * perf;
350 	} else {
351 		tsc1 = rdtsc();
352 		DELAY(1000);
353 		tsc2 = rdtsc();
354 		intr_restore(reg);
355 		*rate = (tsc2 - tsc1) * 1000;
356 	}
357 
358 #ifdef SMP
359 	if (smp_cpus > 1) {
360 		thread_lock(curthread);
361 		sched_unbind(curthread);
362 		thread_unlock(curthread);
363 	}
364 #endif
365 
366 	return (0);
367 }
368 
369 /*
370  * Shutdown the CPU as much as possible
371  */
372 void
373 cpu_halt(void)
374 {
375 	for (;;)
376 		halt();
377 }
378 
379 static void
380 cpu_reset_real(void)
381 {
382 	struct region_descriptor null_idt;
383 	int b;
384 
385 	disable_intr();
386 #ifdef CPU_ELAN
387 	if (elan_mmcr != NULL)
388 		elan_mmcr->RESCFG = 1;
389 #endif
390 #ifdef __i386__
391 	if (cpu == CPU_GEODE1100) {
392 		/* Attempt Geode's own reset */
393 		outl(0xcf8, 0x80009044ul);
394 		outl(0xcfc, 0xf);
395 	}
396 #endif
397 #if !defined(BROKEN_KEYBOARD_RESET)
398 	/*
399 	 * Attempt to do a CPU reset via the keyboard controller,
400 	 * do not turn off GateA20, as any machine that fails
401 	 * to do the reset here would then end up in no man's land.
402 	 */
403 	outb(IO_KBD + 4, 0xFE);
404 	DELAY(500000);	/* wait 0.5 sec to see if that did it */
405 #endif
406 
407 	/*
408 	 * Attempt to force a reset via the Reset Control register at
409 	 * I/O port 0xcf9.  Bit 2 forces a system reset when it
410 	 * transitions from 0 to 1.  Bit 1 selects the type of reset
411 	 * to attempt: 0 selects a "soft" reset, and 1 selects a
412 	 * "hard" reset.  We try a "hard" reset.  The first write sets
413 	 * bit 1 to select a "hard" reset and clears bit 2.  The
414 	 * second write forces a 0 -> 1 transition in bit 2 to trigger
415 	 * a reset.
416 	 */
417 	outb(0xcf9, 0x2);
418 	outb(0xcf9, 0x6);
419 	DELAY(500000);  /* wait 0.5 sec to see if that did it */
420 
421 	/*
422 	 * Attempt to force a reset via the Fast A20 and Init register
423 	 * at I/O port 0x92.  Bit 1 serves as an alternate A20 gate.
424 	 * Bit 0 asserts INIT# when set to 1.  We are careful to only
425 	 * preserve bit 1 while setting bit 0.  We also must clear bit
426 	 * 0 before setting it if it isn't already clear.
427 	 */
428 	b = inb(0x92);
429 	if (b != 0xff) {
430 		if ((b & 0x1) != 0)
431 			outb(0x92, b & 0xfe);
432 		outb(0x92, b | 0x1);
433 		DELAY(500000);  /* wait 0.5 sec to see if that did it */
434 	}
435 
436 	printf("No known reset method worked, attempting CPU shutdown\n");
437 	DELAY(1000000); /* wait 1 sec for printf to complete */
438 
439 	/* Wipe the IDT. */
440 	null_idt.rd_limit = 0;
441 	null_idt.rd_base = 0;
442 	lidt(&null_idt);
443 
444 	/* "good night, sweet prince .... <THUNK!>" */
445 	breakpoint();
446 
447 	/* NOTREACHED */
448 	while(1);
449 }
450 
451 #ifdef SMP
452 static void
453 cpu_reset_proxy(void)
454 {
455 
456 	cpu_reset_proxy_active = 1;
457 	while (cpu_reset_proxy_active == 1)
458 		ia32_pause(); /* Wait for other cpu to see that we've started */
459 
460 	printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid);
461 	DELAY(1000000);
462 	cpu_reset_real();
463 }
464 #endif
465 
466 void
467 cpu_reset(void)
468 {
469 #ifdef SMP
470 	struct monitorbuf *mb;
471 	cpuset_t map;
472 	u_int cnt;
473 
474 	if (smp_started) {
475 		map = all_cpus;
476 		CPU_CLR(PCPU_GET(cpuid), &map);
477 		CPU_ANDNOT(&map, &map, &stopped_cpus);
478 		if (!CPU_EMPTY(&map)) {
479 			printf("cpu_reset: Stopping other CPUs\n");
480 			stop_cpus(map);
481 		}
482 
483 		if (PCPU_GET(cpuid) != 0) {
484 			cpu_reset_proxyid = PCPU_GET(cpuid);
485 			cpustop_restartfunc = cpu_reset_proxy;
486 			cpu_reset_proxy_active = 0;
487 			printf("cpu_reset: Restarting BSP\n");
488 
489 			/* Restart CPU #0. */
490 			CPU_SETOF(0, &started_cpus);
491 			mb = &pcpu_find(0)->pc_monitorbuf;
492 			atomic_store_int(&mb->stop_state,
493 			    MONITOR_STOPSTATE_RUNNING);
494 
495 			cnt = 0;
496 			while (cpu_reset_proxy_active == 0 && cnt < 10000000) {
497 				ia32_pause();
498 				cnt++;	/* Wait for BSP to announce restart */
499 			}
500 			if (cpu_reset_proxy_active == 0) {
501 				printf("cpu_reset: Failed to restart BSP\n");
502 			} else {
503 				cpu_reset_proxy_active = 2;
504 				while (1)
505 					ia32_pause();
506 				/* NOTREACHED */
507 			}
508 		}
509 	}
510 #endif
511 	cpu_reset_real();
512 	/* NOTREACHED */
513 }
514 
515 bool
516 cpu_mwait_usable(void)
517 {
518 
519 	return ((cpu_feature2 & CPUID2_MON) != 0 && ((cpu_mon_mwait_flags &
520 	    (CPUID5_MON_MWAIT_EXT | CPUID5_MWAIT_INTRBREAK)) ==
521 	    (CPUID5_MON_MWAIT_EXT | CPUID5_MWAIT_INTRBREAK)));
522 }
523 
524 void (*cpu_idle_hook)(sbintime_t) = NULL;	/* ACPI idle hook. */
525 
526 int cpu_amdc1e_bug = 0;			/* AMD C1E APIC workaround required. */
527 
528 static int	idle_mwait = 1;		/* Use MONITOR/MWAIT for short idle. */
529 SYSCTL_INT(_machdep, OID_AUTO, idle_mwait, CTLFLAG_RWTUN, &idle_mwait,
530     0, "Use MONITOR/MWAIT for short idle");
531 
532 static bool
533 cpu_idle_enter(int *statep, int newstate)
534 {
535 	KASSERT(atomic_load_int(statep) == STATE_RUNNING,
536 	    ("%s: state %d", __func__, atomic_load_int(statep)));
537 
538 	/*
539 	 * A fence is needed to prevent reordering of the load in
540 	 * sched_runnable() with this store to the idle state word.  Without it,
541 	 * cpu_idle_wakeup() can observe the state as STATE_RUNNING after having
542 	 * added load to the queue, and elide an IPI.  Then, sched_runnable()
543 	 * can observe tdq_load == 0, so the CPU ends up idling with pending
544 	 * work.  tdq_notify() similarly ensures that a prior update to tdq_load
545 	 * is visible before calling cpu_idle_wakeup().
546 	 */
547 	atomic_store_int(statep, newstate);
548 #if defined(SCHED_ULE) && defined(SMP)
549 	atomic_thread_fence_seq_cst();
550 #endif
551 
552 	/*
553 	 * Since we may be in a critical section from cpu_idle(), if
554 	 * an interrupt fires during that critical section we may have
555 	 * a pending preemption.  If the CPU halts, then that thread
556 	 * may not execute until a later interrupt awakens the CPU.
557 	 * To handle this race, check for a runnable thread after
558 	 * disabling interrupts and immediately return if one is
559 	 * found.  Also, we must absolutely guarentee that hlt is
560 	 * the next instruction after sti.  This ensures that any
561 	 * interrupt that fires after the call to disable_intr() will
562 	 * immediately awaken the CPU from hlt.  Finally, please note
563 	 * that on x86 this works fine because of interrupts enabled only
564 	 * after the instruction following sti takes place, while IF is set
565 	 * to 1 immediately, allowing hlt instruction to acknowledge the
566 	 * interrupt.
567 	 */
568 	disable_intr();
569 	if (sched_runnable()) {
570 		enable_intr();
571 		atomic_store_int(statep, STATE_RUNNING);
572 		return (false);
573 	} else {
574 		return (true);
575 	}
576 }
577 
578 static void
579 cpu_idle_exit(int *statep)
580 {
581 	atomic_store_int(statep, STATE_RUNNING);
582 }
583 
584 static void
585 cpu_idle_acpi(sbintime_t sbt)
586 {
587 	int *state;
588 
589 	state = &PCPU_PTR(monitorbuf)->idle_state;
590 	if (cpu_idle_enter(state, STATE_SLEEPING)) {
591 		if (cpu_idle_hook)
592 			cpu_idle_hook(sbt);
593 		else
594 			acpi_cpu_c1();
595 		cpu_idle_exit(state);
596 	}
597 }
598 
599 static void
600 cpu_idle_hlt(sbintime_t sbt)
601 {
602 	int *state;
603 
604 	state = &PCPU_PTR(monitorbuf)->idle_state;
605 	if (cpu_idle_enter(state, STATE_SLEEPING)) {
606 		acpi_cpu_c1();
607 		atomic_store_int(state, STATE_RUNNING);
608 	}
609 }
610 
611 static void
612 cpu_idle_mwait(sbintime_t sbt)
613 {
614 	int *state;
615 
616 	state = &PCPU_PTR(monitorbuf)->idle_state;
617 	if (cpu_idle_enter(state, STATE_MWAIT)) {
618 		cpu_monitor(state, 0, 0);
619 		if (atomic_load_int(state) == STATE_MWAIT)
620 			__asm __volatile("sti; mwait" : : "a" (MWAIT_C1), "c" (0));
621 		else
622 			enable_intr();
623 		cpu_idle_exit(state);
624 	}
625 }
626 
627 static void
628 cpu_idle_spin(sbintime_t sbt)
629 {
630 	int *state;
631 	int i;
632 
633 	state = &PCPU_PTR(monitorbuf)->idle_state;
634 	atomic_store_int(state, STATE_RUNNING);
635 
636 	/*
637 	 * The sched_runnable() call is racy but as long as there is
638 	 * a loop missing it one time will have just a little impact if any
639 	 * (and it is much better than missing the check at all).
640 	 */
641 	for (i = 0; i < 1000; i++) {
642 		if (sched_runnable())
643 			return;
644 		cpu_spinwait();
645 	}
646 }
647 
648 void (*cpu_idle_fn)(sbintime_t) = cpu_idle_acpi;
649 
650 void
651 cpu_idle(int busy)
652 {
653 	uint64_t msr;
654 	sbintime_t sbt = -1;
655 
656 	CTR1(KTR_SPARE2, "cpu_idle(%d)", busy);
657 
658 	/* If we are busy - try to use fast methods. */
659 	if (busy) {
660 		if ((cpu_feature2 & CPUID2_MON) && idle_mwait) {
661 			cpu_idle_mwait(busy);
662 			goto out;
663 		}
664 	}
665 
666 	/* If we have time - switch timers into idle mode. */
667 	if (!busy) {
668 		critical_enter();
669 		sbt = cpu_idleclock();
670 	}
671 
672 	/* Apply AMD APIC timer C1E workaround. */
673 	if (cpu_amdc1e_bug && cpu_disable_c3_sleep) {
674 		msr = rdmsr(MSR_AMDK8_IPM);
675 		if ((msr & (AMDK8_SMIONCMPHALT | AMDK8_C1EONCMPHALT)) != 0)
676 			wrmsr(MSR_AMDK8_IPM, msr & ~(AMDK8_SMIONCMPHALT |
677 			    AMDK8_C1EONCMPHALT));
678 	}
679 
680 	/* Call main idle method. */
681 	cpu_idle_fn(sbt);
682 
683 	/* Switch timers back into active mode. */
684 	if (!busy) {
685 		cpu_activeclock();
686 		critical_exit();
687 	}
688 out:
689 	CTR1(KTR_SPARE2, "cpu_idle(%d) done", busy);
690 }
691 
692 static int cpu_idle_apl31_workaround;
693 SYSCTL_INT(_machdep, OID_AUTO, idle_apl31, CTLFLAG_RW,
694     &cpu_idle_apl31_workaround, 0,
695     "Apollo Lake APL31 MWAIT bug workaround");
696 
697 int
698 cpu_idle_wakeup(int cpu)
699 {
700 	struct monitorbuf *mb;
701 	int *state;
702 
703 	mb = &pcpu_find(cpu)->pc_monitorbuf;
704 	state = &mb->idle_state;
705 	switch (atomic_load_int(state)) {
706 	case STATE_SLEEPING:
707 		return (0);
708 	case STATE_MWAIT:
709 		atomic_store_int(state, STATE_RUNNING);
710 		return (cpu_idle_apl31_workaround ? 0 : 1);
711 	case STATE_RUNNING:
712 		return (1);
713 	default:
714 		panic("bad monitor state");
715 		return (1);
716 	}
717 }
718 
719 /*
720  * Ordered by speed/power consumption.
721  */
722 static struct {
723 	void	*id_fn;
724 	char	*id_name;
725 	int	id_cpuid2_flag;
726 } idle_tbl[] = {
727 	{ .id_fn = cpu_idle_spin, .id_name = "spin" },
728 	{ .id_fn = cpu_idle_mwait, .id_name = "mwait",
729 	    .id_cpuid2_flag = CPUID2_MON },
730 	{ .id_fn = cpu_idle_hlt, .id_name = "hlt" },
731 	{ .id_fn = cpu_idle_acpi, .id_name = "acpi" },
732 };
733 
734 static int
735 idle_sysctl_available(SYSCTL_HANDLER_ARGS)
736 {
737 	char *avail, *p;
738 	int error;
739 	int i;
740 
741 	avail = malloc(256, M_TEMP, M_WAITOK);
742 	p = avail;
743 	for (i = 0; i < nitems(idle_tbl); i++) {
744 		if (idle_tbl[i].id_cpuid2_flag != 0 &&
745 		    (cpu_feature2 & idle_tbl[i].id_cpuid2_flag) == 0)
746 			continue;
747 		if (strcmp(idle_tbl[i].id_name, "acpi") == 0 &&
748 		    cpu_idle_hook == NULL)
749 			continue;
750 		p += sprintf(p, "%s%s", p != avail ? ", " : "",
751 		    idle_tbl[i].id_name);
752 	}
753 	error = sysctl_handle_string(oidp, avail, 0, req);
754 	free(avail, M_TEMP);
755 	return (error);
756 }
757 
758 SYSCTL_PROC(_machdep, OID_AUTO, idle_available,
759     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE,
760     0, 0, idle_sysctl_available, "A",
761     "list of available idle functions");
762 
763 static bool
764 cpu_idle_selector(const char *new_idle_name)
765 {
766 	int i;
767 
768 	for (i = 0; i < nitems(idle_tbl); i++) {
769 		if (idle_tbl[i].id_cpuid2_flag != 0 &&
770 		    (cpu_feature2 & idle_tbl[i].id_cpuid2_flag) == 0)
771 			continue;
772 		if (strcmp(idle_tbl[i].id_name, "acpi") == 0 &&
773 		    cpu_idle_hook == NULL)
774 			continue;
775 		if (strcmp(idle_tbl[i].id_name, new_idle_name))
776 			continue;
777 		cpu_idle_fn = idle_tbl[i].id_fn;
778 		if (bootverbose)
779 			printf("CPU idle set to %s\n", idle_tbl[i].id_name);
780 		return (true);
781 	}
782 	return (false);
783 }
784 
785 static int
786 cpu_idle_sysctl(SYSCTL_HANDLER_ARGS)
787 {
788 	char buf[16], *p;
789 	int error, i;
790 
791 	p = "unknown";
792 	for (i = 0; i < nitems(idle_tbl); i++) {
793 		if (idle_tbl[i].id_fn == cpu_idle_fn) {
794 			p = idle_tbl[i].id_name;
795 			break;
796 		}
797 	}
798 	strncpy(buf, p, sizeof(buf));
799 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
800 	if (error != 0 || req->newptr == NULL)
801 		return (error);
802 	return (cpu_idle_selector(buf) ? 0 : EINVAL);
803 }
804 
805 SYSCTL_PROC(_machdep, OID_AUTO, idle,
806     CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
807     0, 0, cpu_idle_sysctl, "A",
808     "currently selected idle function");
809 
810 static void
811 cpu_idle_tun(void *unused __unused)
812 {
813 	char tunvar[16];
814 
815 	if (TUNABLE_STR_FETCH("machdep.idle", tunvar, sizeof(tunvar)))
816 		cpu_idle_selector(tunvar);
817 	else if (cpu_vendor_id == CPU_VENDOR_AMD &&
818 	    CPUID_TO_FAMILY(cpu_id) == 0x17 && CPUID_TO_MODEL(cpu_id) == 0x1) {
819 		/* Ryzen erratas 1057, 1109. */
820 		cpu_idle_selector("hlt");
821 		idle_mwait = 0;
822 		mwait_cpustop_broken = true;
823 	}
824 
825 	if (cpu_vendor_id == CPU_VENDOR_INTEL &&
826 	    (cpu_id == 0x506c9 || cpu_id == 0x506ca)) {
827 		/*
828 		 * Apollo Lake errata APL31 (public errata APL30).
829 		 * Stores to the armed address range may not trigger
830 		 * MWAIT to resume execution.  OS needs to use
831 		 * interrupts to wake processors from MWAIT-induced
832 		 * sleep states.
833 		 */
834 		cpu_idle_apl31_workaround = 1;
835 		mwait_cpustop_broken = true;
836 	}
837 	TUNABLE_INT_FETCH("machdep.idle_apl31", &cpu_idle_apl31_workaround);
838 }
839 SYSINIT(cpu_idle_tun, SI_SUB_CPU, SI_ORDER_MIDDLE, cpu_idle_tun, NULL);
840 
841 static int panic_on_nmi = 0xff;
842 SYSCTL_INT(_machdep, OID_AUTO, panic_on_nmi, CTLFLAG_RWTUN,
843     &panic_on_nmi, 0,
844     "Panic on NMI: 1 = H/W failure; 2 = unknown; 0xff = all");
845 int nmi_is_broadcast = 1;
846 SYSCTL_INT(_machdep, OID_AUTO, nmi_is_broadcast, CTLFLAG_RWTUN,
847     &nmi_is_broadcast, 0,
848     "Chipset NMI is broadcast");
849 int (*apei_nmi)(void);
850 
851 void
852 nmi_call_kdb(u_int cpu, u_int type, struct trapframe *frame)
853 {
854 	bool claimed = false;
855 
856 #ifdef DEV_ISA
857 	/* machine/parity/power fail/"kitchen sink" faults */
858 	if (isa_nmi(frame->tf_err)) {
859 		claimed = true;
860 		if ((panic_on_nmi & 1) != 0)
861 			panic("NMI indicates hardware failure");
862 	}
863 #endif /* DEV_ISA */
864 
865 	/* ACPI Platform Error Interfaces callback. */
866 	if (apei_nmi != NULL && (*apei_nmi)())
867 		claimed = true;
868 
869 	/*
870 	 * NMIs can be useful for debugging.  They can be hooked up to a
871 	 * pushbutton, usually on an ISA, PCI, or PCIe card.  They can also be
872 	 * generated by an IPMI BMC, either manually or in response to a
873 	 * watchdog timeout.  For example, see the "power diag" command in
874 	 * ports/sysutils/ipmitool.  They can also be generated by a
875 	 * hypervisor; see "bhyvectl --inject-nmi".
876 	 */
877 
878 #ifdef KDB
879 	if (!claimed && (panic_on_nmi & 2) != 0) {
880 		if (debugger_on_panic) {
881 			printf("NMI/cpu%d ... going to debugger\n", cpu);
882 			claimed = kdb_trap(type, 0, frame);
883 		}
884 	}
885 #endif /* KDB */
886 
887 	if (!claimed && panic_on_nmi != 0)
888 		panic("NMI");
889 }
890 
891 void
892 nmi_handle_intr(u_int type, struct trapframe *frame)
893 {
894 
895 #ifdef SMP
896 	if (nmi_is_broadcast) {
897 		nmi_call_kdb_smp(type, frame);
898 		return;
899 	}
900 #endif
901 	nmi_call_kdb(PCPU_GET(cpuid), type, frame);
902 }
903 
904 static int hw_ibrs_active;
905 int hw_ibrs_ibpb_active;
906 int hw_ibrs_disable = 1;
907 
908 SYSCTL_INT(_hw, OID_AUTO, ibrs_active, CTLFLAG_RD, &hw_ibrs_active, 0,
909     "Indirect Branch Restricted Speculation active");
910 
911 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, ibrs,
912     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
913     "Indirect Branch Restricted Speculation active");
914 
915 SYSCTL_INT(_machdep_mitigations_ibrs, OID_AUTO, active, CTLFLAG_RD,
916     &hw_ibrs_active, 0, "Indirect Branch Restricted Speculation active");
917 
918 void
919 hw_ibrs_recalculate(bool for_all_cpus)
920 {
921 	if ((cpu_ia32_arch_caps & IA32_ARCH_CAP_IBRS_ALL) != 0) {
922 		x86_msr_op(MSR_IA32_SPEC_CTRL, (for_all_cpus ?
923 		    MSR_OP_RENDEZVOUS_ALL : MSR_OP_LOCAL) |
924 		    (hw_ibrs_disable != 0 ? MSR_OP_ANDNOT : MSR_OP_OR),
925 		    IA32_SPEC_CTRL_IBRS, NULL);
926 		hw_ibrs_active = hw_ibrs_disable == 0;
927 		hw_ibrs_ibpb_active = 0;
928 	} else {
929 		hw_ibrs_active = hw_ibrs_ibpb_active = (cpu_stdext_feature3 &
930 		    CPUID_STDEXT3_IBPB) != 0 && !hw_ibrs_disable;
931 	}
932 }
933 
934 static int
935 hw_ibrs_disable_handler(SYSCTL_HANDLER_ARGS)
936 {
937 	int error, val;
938 
939 	val = hw_ibrs_disable;
940 	error = sysctl_handle_int(oidp, &val, 0, req);
941 	if (error != 0 || req->newptr == NULL)
942 		return (error);
943 	hw_ibrs_disable = val != 0;
944 	hw_ibrs_recalculate(true);
945 	return (0);
946 }
947 SYSCTL_PROC(_hw, OID_AUTO, ibrs_disable, CTLTYPE_INT | CTLFLAG_RWTUN |
948     CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0, hw_ibrs_disable_handler, "I",
949     "Disable Indirect Branch Restricted Speculation");
950 
951 SYSCTL_PROC(_machdep_mitigations_ibrs, OID_AUTO, disable, CTLTYPE_INT |
952     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
953     hw_ibrs_disable_handler, "I",
954     "Disable Indirect Branch Restricted Speculation");
955 
956 int hw_ssb_active;
957 int hw_ssb_disable;
958 
959 SYSCTL_INT(_hw, OID_AUTO, spec_store_bypass_disable_active, CTLFLAG_RD,
960     &hw_ssb_active, 0,
961     "Speculative Store Bypass Disable active");
962 
963 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, ssb,
964     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
965     "Speculative Store Bypass Disable active");
966 
967 SYSCTL_INT(_machdep_mitigations_ssb, OID_AUTO, active, CTLFLAG_RD,
968     &hw_ssb_active, 0, "Speculative Store Bypass Disable active");
969 
970 static void
971 hw_ssb_set(bool enable, bool for_all_cpus)
972 {
973 
974 	if ((cpu_stdext_feature3 & CPUID_STDEXT3_SSBD) == 0) {
975 		hw_ssb_active = 0;
976 		return;
977 	}
978 	hw_ssb_active = enable;
979 	x86_msr_op(MSR_IA32_SPEC_CTRL,
980 	    (enable ? MSR_OP_OR : MSR_OP_ANDNOT) |
981 	    (for_all_cpus ? MSR_OP_SCHED_ALL : MSR_OP_LOCAL),
982 	    IA32_SPEC_CTRL_SSBD, NULL);
983 }
984 
985 void
986 hw_ssb_recalculate(bool all_cpus)
987 {
988 
989 	switch (hw_ssb_disable) {
990 	default:
991 		hw_ssb_disable = 0;
992 		/* FALLTHROUGH */
993 	case 0: /* off */
994 		hw_ssb_set(false, all_cpus);
995 		break;
996 	case 1: /* on */
997 		hw_ssb_set(true, all_cpus);
998 		break;
999 	case 2: /* auto */
1000 		hw_ssb_set((cpu_ia32_arch_caps & IA32_ARCH_CAP_SSB_NO) != 0 ?
1001 		    false : true, all_cpus);
1002 		break;
1003 	}
1004 }
1005 
1006 static int
1007 hw_ssb_disable_handler(SYSCTL_HANDLER_ARGS)
1008 {
1009 	int error, val;
1010 
1011 	val = hw_ssb_disable;
1012 	error = sysctl_handle_int(oidp, &val, 0, req);
1013 	if (error != 0 || req->newptr == NULL)
1014 		return (error);
1015 	hw_ssb_disable = val;
1016 	hw_ssb_recalculate(true);
1017 	return (0);
1018 }
1019 SYSCTL_PROC(_hw, OID_AUTO, spec_store_bypass_disable, CTLTYPE_INT |
1020     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1021     hw_ssb_disable_handler, "I",
1022     "Speculative Store Bypass Disable (0 - off, 1 - on, 2 - auto)");
1023 
1024 SYSCTL_PROC(_machdep_mitigations_ssb, OID_AUTO, disable, CTLTYPE_INT |
1025     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1026     hw_ssb_disable_handler, "I",
1027     "Speculative Store Bypass Disable (0 - off, 1 - on, 2 - auto)");
1028 
1029 int hw_mds_disable;
1030 
1031 /*
1032  * Handler for Microarchitectural Data Sampling issues.  Really not a
1033  * pointer to C function: on amd64 the code must not change any CPU
1034  * architectural state except possibly %rflags. Also, it is always
1035  * called with interrupts disabled.
1036  */
1037 void mds_handler_void(void);
1038 void mds_handler_verw(void);
1039 void mds_handler_ivb(void);
1040 void mds_handler_bdw(void);
1041 void mds_handler_skl_sse(void);
1042 void mds_handler_skl_avx(void);
1043 void mds_handler_skl_avx512(void);
1044 void mds_handler_silvermont(void);
1045 void (*mds_handler)(void) = mds_handler_void;
1046 
1047 static int
1048 sysctl_hw_mds_disable_state_handler(SYSCTL_HANDLER_ARGS)
1049 {
1050 	const char *state;
1051 
1052 	if (mds_handler == mds_handler_void)
1053 		state = "inactive";
1054 	else if (mds_handler == mds_handler_verw)
1055 		state = "VERW";
1056 	else if (mds_handler == mds_handler_ivb)
1057 		state = "software IvyBridge";
1058 	else if (mds_handler == mds_handler_bdw)
1059 		state = "software Broadwell";
1060 	else if (mds_handler == mds_handler_skl_sse)
1061 		state = "software Skylake SSE";
1062 	else if (mds_handler == mds_handler_skl_avx)
1063 		state = "software Skylake AVX";
1064 	else if (mds_handler == mds_handler_skl_avx512)
1065 		state = "software Skylake AVX512";
1066 	else if (mds_handler == mds_handler_silvermont)
1067 		state = "software Silvermont";
1068 	else
1069 		state = "unknown";
1070 	return (SYSCTL_OUT(req, state, strlen(state)));
1071 }
1072 
1073 SYSCTL_PROC(_hw, OID_AUTO, mds_disable_state,
1074     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1075     sysctl_hw_mds_disable_state_handler, "A",
1076     "Microarchitectural Data Sampling Mitigation state");
1077 
1078 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, mds,
1079     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1080     "Microarchitectural Data Sampling Mitigation state");
1081 
1082 SYSCTL_PROC(_machdep_mitigations_mds, OID_AUTO, state,
1083     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1084     sysctl_hw_mds_disable_state_handler, "A",
1085     "Microarchitectural Data Sampling Mitigation state");
1086 
1087 _Static_assert(__offsetof(struct pcpu, pc_mds_tmp) % 64 == 0, "MDS AVX512");
1088 
1089 void
1090 hw_mds_recalculate(void)
1091 {
1092 	struct pcpu *pc;
1093 	vm_offset_t b64;
1094 	u_long xcr0;
1095 	int i;
1096 
1097 	/*
1098 	 * Allow user to force VERW variant even if MD_CLEAR is not
1099 	 * reported.  For instance, hypervisor might unknowingly
1100 	 * filter the cap out.
1101 	 * For the similar reasons, and for testing, allow to enable
1102 	 * mitigation even when MDS_NO cap is set.
1103 	 */
1104 	if (cpu_vendor_id != CPU_VENDOR_INTEL || hw_mds_disable == 0 ||
1105 	    ((cpu_ia32_arch_caps & IA32_ARCH_CAP_MDS_NO) != 0 &&
1106 	    hw_mds_disable == 3)) {
1107 		mds_handler = mds_handler_void;
1108 	} else if (((cpu_stdext_feature3 & CPUID_STDEXT3_MD_CLEAR) != 0 &&
1109 	    hw_mds_disable == 3) || hw_mds_disable == 1) {
1110 		mds_handler = mds_handler_verw;
1111 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1112 	    (CPUID_TO_MODEL(cpu_id) == 0x2e || CPUID_TO_MODEL(cpu_id) == 0x1e ||
1113 	    CPUID_TO_MODEL(cpu_id) == 0x1f || CPUID_TO_MODEL(cpu_id) == 0x1a ||
1114 	    CPUID_TO_MODEL(cpu_id) == 0x2f || CPUID_TO_MODEL(cpu_id) == 0x25 ||
1115 	    CPUID_TO_MODEL(cpu_id) == 0x2c || CPUID_TO_MODEL(cpu_id) == 0x2d ||
1116 	    CPUID_TO_MODEL(cpu_id) == 0x2a || CPUID_TO_MODEL(cpu_id) == 0x3e ||
1117 	    CPUID_TO_MODEL(cpu_id) == 0x3a) &&
1118 	    (hw_mds_disable == 2 || hw_mds_disable == 3)) {
1119 		/*
1120 		 * Nehalem, SandyBridge, IvyBridge
1121 		 */
1122 		CPU_FOREACH(i) {
1123 			pc = pcpu_find(i);
1124 			if (pc->pc_mds_buf == NULL) {
1125 				pc->pc_mds_buf = malloc_domainset(672, M_TEMP,
1126 				    DOMAINSET_PREF(pc->pc_domain), M_WAITOK);
1127 				bzero(pc->pc_mds_buf, 16);
1128 			}
1129 		}
1130 		mds_handler = mds_handler_ivb;
1131 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1132 	    (CPUID_TO_MODEL(cpu_id) == 0x3f || CPUID_TO_MODEL(cpu_id) == 0x3c ||
1133 	    CPUID_TO_MODEL(cpu_id) == 0x45 || CPUID_TO_MODEL(cpu_id) == 0x46 ||
1134 	    CPUID_TO_MODEL(cpu_id) == 0x56 || CPUID_TO_MODEL(cpu_id) == 0x4f ||
1135 	    CPUID_TO_MODEL(cpu_id) == 0x47 || CPUID_TO_MODEL(cpu_id) == 0x3d) &&
1136 	    (hw_mds_disable == 2 || hw_mds_disable == 3)) {
1137 		/*
1138 		 * Haswell, Broadwell
1139 		 */
1140 		CPU_FOREACH(i) {
1141 			pc = pcpu_find(i);
1142 			if (pc->pc_mds_buf == NULL) {
1143 				pc->pc_mds_buf = malloc_domainset(1536, M_TEMP,
1144 				    DOMAINSET_PREF(pc->pc_domain), M_WAITOK);
1145 				bzero(pc->pc_mds_buf, 16);
1146 			}
1147 		}
1148 		mds_handler = mds_handler_bdw;
1149 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1150 	    ((CPUID_TO_MODEL(cpu_id) == 0x55 && (cpu_id &
1151 	    CPUID_STEPPING) <= 5) ||
1152 	    CPUID_TO_MODEL(cpu_id) == 0x4e || CPUID_TO_MODEL(cpu_id) == 0x5e ||
1153 	    (CPUID_TO_MODEL(cpu_id) == 0x8e && (cpu_id &
1154 	    CPUID_STEPPING) <= 0xb) ||
1155 	    (CPUID_TO_MODEL(cpu_id) == 0x9e && (cpu_id &
1156 	    CPUID_STEPPING) <= 0xc)) &&
1157 	    (hw_mds_disable == 2 || hw_mds_disable == 3)) {
1158 		/*
1159 		 * Skylake, KabyLake, CoffeeLake, WhiskeyLake,
1160 		 * CascadeLake
1161 		 */
1162 		CPU_FOREACH(i) {
1163 			pc = pcpu_find(i);
1164 			if (pc->pc_mds_buf == NULL) {
1165 				pc->pc_mds_buf = malloc_domainset(6 * 1024,
1166 				    M_TEMP, DOMAINSET_PREF(pc->pc_domain),
1167 				    M_WAITOK);
1168 				b64 = (vm_offset_t)malloc_domainset(64 + 63,
1169 				    M_TEMP, DOMAINSET_PREF(pc->pc_domain),
1170 				    M_WAITOK);
1171 				pc->pc_mds_buf64 = (void *)roundup2(b64, 64);
1172 				bzero(pc->pc_mds_buf64, 64);
1173 			}
1174 		}
1175 		xcr0 = rxcr(0);
1176 		if ((xcr0 & XFEATURE_ENABLED_ZMM_HI256) != 0 &&
1177 		    (cpu_stdext_feature & CPUID_STDEXT_AVX512DQ) != 0)
1178 			mds_handler = mds_handler_skl_avx512;
1179 		else if ((xcr0 & XFEATURE_ENABLED_AVX) != 0 &&
1180 		    (cpu_feature2 & CPUID2_AVX) != 0)
1181 			mds_handler = mds_handler_skl_avx;
1182 		else
1183 			mds_handler = mds_handler_skl_sse;
1184 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1185 	    ((CPUID_TO_MODEL(cpu_id) == 0x37 ||
1186 	    CPUID_TO_MODEL(cpu_id) == 0x4a ||
1187 	    CPUID_TO_MODEL(cpu_id) == 0x4c ||
1188 	    CPUID_TO_MODEL(cpu_id) == 0x4d ||
1189 	    CPUID_TO_MODEL(cpu_id) == 0x5a ||
1190 	    CPUID_TO_MODEL(cpu_id) == 0x5d ||
1191 	    CPUID_TO_MODEL(cpu_id) == 0x6e ||
1192 	    CPUID_TO_MODEL(cpu_id) == 0x65 ||
1193 	    CPUID_TO_MODEL(cpu_id) == 0x75 ||
1194 	    CPUID_TO_MODEL(cpu_id) == 0x1c ||
1195 	    CPUID_TO_MODEL(cpu_id) == 0x26 ||
1196 	    CPUID_TO_MODEL(cpu_id) == 0x27 ||
1197 	    CPUID_TO_MODEL(cpu_id) == 0x35 ||
1198 	    CPUID_TO_MODEL(cpu_id) == 0x36 ||
1199 	    CPUID_TO_MODEL(cpu_id) == 0x7a))) {
1200 		/* Silvermont, Airmont */
1201 		CPU_FOREACH(i) {
1202 			pc = pcpu_find(i);
1203 			if (pc->pc_mds_buf == NULL)
1204 				pc->pc_mds_buf = malloc(256, M_TEMP, M_WAITOK);
1205 		}
1206 		mds_handler = mds_handler_silvermont;
1207 	} else {
1208 		hw_mds_disable = 0;
1209 		mds_handler = mds_handler_void;
1210 	}
1211 }
1212 
1213 static void
1214 hw_mds_recalculate_boot(void *arg __unused)
1215 {
1216 
1217 	hw_mds_recalculate();
1218 }
1219 SYSINIT(mds_recalc, SI_SUB_SMP, SI_ORDER_ANY, hw_mds_recalculate_boot, NULL);
1220 
1221 static int
1222 sysctl_mds_disable_handler(SYSCTL_HANDLER_ARGS)
1223 {
1224 	int error, val;
1225 
1226 	val = hw_mds_disable;
1227 	error = sysctl_handle_int(oidp, &val, 0, req);
1228 	if (error != 0 || req->newptr == NULL)
1229 		return (error);
1230 	if (val < 0 || val > 3)
1231 		return (EINVAL);
1232 	hw_mds_disable = val;
1233 	hw_mds_recalculate();
1234 	return (0);
1235 }
1236 
1237 SYSCTL_PROC(_hw, OID_AUTO, mds_disable, CTLTYPE_INT |
1238     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1239     sysctl_mds_disable_handler, "I",
1240     "Microarchitectural Data Sampling Mitigation "
1241     "(0 - off, 1 - on VERW, 2 - on SW, 3 - on AUTO)");
1242 
1243 SYSCTL_PROC(_machdep_mitigations_mds, OID_AUTO, disable, CTLTYPE_INT |
1244     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1245     sysctl_mds_disable_handler, "I",
1246     "Microarchitectural Data Sampling Mitigation "
1247     "(0 - off, 1 - on VERW, 2 - on SW, 3 - on AUTO)");
1248 
1249 /*
1250  * Intel Transactional Memory Asynchronous Abort Mitigation
1251  * CVE-2019-11135
1252  */
1253 int x86_taa_enable;
1254 int x86_taa_state;
1255 enum {
1256 	TAA_NONE	= 0,	/* No mitigation enabled */
1257 	TAA_TSX_DISABLE	= 1,	/* Disable TSX via MSR */
1258 	TAA_VERW	= 2,	/* Use VERW mitigation */
1259 	TAA_AUTO	= 3,	/* Automatically select the mitigation */
1260 
1261 	/* The states below are not selectable by the operator */
1262 
1263 	TAA_TAA_UC	= 4,	/* Mitigation present in microcode */
1264 	TAA_NOT_PRESENT	= 5	/* TSX is not present */
1265 };
1266 
1267 static void
1268 taa_set(bool enable, bool all)
1269 {
1270 
1271 	x86_msr_op(MSR_IA32_TSX_CTRL,
1272 	    (enable ? MSR_OP_OR : MSR_OP_ANDNOT) |
1273 	    (all ? MSR_OP_RENDEZVOUS_ALL : MSR_OP_LOCAL),
1274 	    IA32_TSX_CTRL_RTM_DISABLE | IA32_TSX_CTRL_TSX_CPUID_CLEAR,
1275 	    NULL);
1276 }
1277 
1278 void
1279 x86_taa_recalculate(void)
1280 {
1281 	static int taa_saved_mds_disable = 0;
1282 	int taa_need = 0, taa_state = 0;
1283 	int mds_disable = 0, need_mds_recalc = 0;
1284 
1285 	/* Check CPUID.07h.EBX.HLE and RTM for the presence of TSX */
1286 	if ((cpu_stdext_feature & CPUID_STDEXT_HLE) == 0 ||
1287 	    (cpu_stdext_feature & CPUID_STDEXT_RTM) == 0) {
1288 		/* TSX is not present */
1289 		x86_taa_state = TAA_NOT_PRESENT;
1290 		return;
1291 	}
1292 
1293 	/* Check to see what mitigation options the CPU gives us */
1294 	if (cpu_ia32_arch_caps & IA32_ARCH_CAP_TAA_NO) {
1295 		/* CPU is not suseptible to TAA */
1296 		taa_need = TAA_TAA_UC;
1297 	} else if (cpu_ia32_arch_caps & IA32_ARCH_CAP_TSX_CTRL) {
1298 		/*
1299 		 * CPU can turn off TSX.  This is the next best option
1300 		 * if TAA_NO hardware mitigation isn't present
1301 		 */
1302 		taa_need = TAA_TSX_DISABLE;
1303 	} else {
1304 		/* No TSX/TAA specific remedies are available. */
1305 		if (x86_taa_enable == TAA_TSX_DISABLE) {
1306 			if (bootverbose)
1307 				printf("TSX control not available\n");
1308 			return;
1309 		} else
1310 			taa_need = TAA_VERW;
1311 	}
1312 
1313 	/* Can we automatically take action, or are we being forced? */
1314 	if (x86_taa_enable == TAA_AUTO)
1315 		taa_state = taa_need;
1316 	else
1317 		taa_state = x86_taa_enable;
1318 
1319 	/* No state change, nothing to do */
1320 	if (taa_state == x86_taa_state) {
1321 		if (bootverbose)
1322 			printf("No TSX change made\n");
1323 		return;
1324 	}
1325 
1326 	/* Does the MSR need to be turned on or off? */
1327 	if (taa_state == TAA_TSX_DISABLE)
1328 		taa_set(true, true);
1329 	else if (x86_taa_state == TAA_TSX_DISABLE)
1330 		taa_set(false, true);
1331 
1332 	/* Does MDS need to be set to turn on VERW? */
1333 	if (taa_state == TAA_VERW) {
1334 		taa_saved_mds_disable = hw_mds_disable;
1335 		mds_disable = hw_mds_disable = 1;
1336 		need_mds_recalc = 1;
1337 	} else if (x86_taa_state == TAA_VERW) {
1338 		mds_disable = hw_mds_disable = taa_saved_mds_disable;
1339 		need_mds_recalc = 1;
1340 	}
1341 	if (need_mds_recalc) {
1342 		hw_mds_recalculate();
1343 		if (mds_disable != hw_mds_disable) {
1344 			if (bootverbose)
1345 				printf("Cannot change MDS state for TAA\n");
1346 			/* Don't update our state */
1347 			return;
1348 		}
1349 	}
1350 
1351 	x86_taa_state = taa_state;
1352 	return;
1353 }
1354 
1355 static void
1356 taa_recalculate_boot(void * arg __unused)
1357 {
1358 
1359 	x86_taa_recalculate();
1360 }
1361 SYSINIT(taa_recalc, SI_SUB_SMP, SI_ORDER_ANY, taa_recalculate_boot, NULL);
1362 
1363 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, taa,
1364     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1365     "TSX Asynchronous Abort Mitigation");
1366 
1367 static int
1368 sysctl_taa_handler(SYSCTL_HANDLER_ARGS)
1369 {
1370 	int error, val;
1371 
1372 	val = x86_taa_enable;
1373 	error = sysctl_handle_int(oidp, &val, 0, req);
1374 	if (error != 0 || req->newptr == NULL)
1375 		return (error);
1376 	if (val < TAA_NONE || val > TAA_AUTO)
1377 		return (EINVAL);
1378 	x86_taa_enable = val;
1379 	x86_taa_recalculate();
1380 	return (0);
1381 }
1382 
1383 SYSCTL_PROC(_machdep_mitigations_taa, OID_AUTO, enable, CTLTYPE_INT |
1384     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1385     sysctl_taa_handler, "I",
1386     "TAA Mitigation enablement control "
1387     "(0 - off, 1 - disable TSX, 2 - VERW, 3 - on AUTO)");
1388 
1389 static int
1390 sysctl_taa_state_handler(SYSCTL_HANDLER_ARGS)
1391 {
1392 	const char *state;
1393 
1394 	switch (x86_taa_state) {
1395 	case TAA_NONE:
1396 		state = "inactive";
1397 		break;
1398 	case TAA_TSX_DISABLE:
1399 		state = "TSX disabled";
1400 		break;
1401 	case TAA_VERW:
1402 		state = "VERW";
1403 		break;
1404 	case TAA_TAA_UC:
1405 		state = "Mitigated in microcode";
1406 		break;
1407 	case TAA_NOT_PRESENT:
1408 		state = "TSX not present";
1409 		break;
1410 	default:
1411 		state = "unknown";
1412 	}
1413 
1414 	return (SYSCTL_OUT(req, state, strlen(state)));
1415 }
1416 
1417 SYSCTL_PROC(_machdep_mitigations_taa, OID_AUTO, state,
1418     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1419     sysctl_taa_state_handler, "A",
1420     "TAA Mitigation state");
1421 
1422 int __read_frequently cpu_flush_rsb_ctxsw;
1423 SYSCTL_INT(_machdep_mitigations, OID_AUTO, flush_rsb_ctxsw,
1424     CTLFLAG_RW | CTLFLAG_NOFETCH, &cpu_flush_rsb_ctxsw, 0,
1425     "Flush Return Stack Buffer on context switch");
1426 
1427 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, rngds,
1428     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1429     "MCU Optimization, disable RDSEED mitigation");
1430 
1431 int x86_rngds_mitg_enable = 1;
1432 void
1433 x86_rngds_mitg_recalculate(bool all_cpus)
1434 {
1435 	if ((cpu_stdext_feature3 & CPUID_STDEXT3_MCUOPT) == 0)
1436 		return;
1437 	x86_msr_op(MSR_IA32_MCU_OPT_CTRL,
1438 	    (x86_rngds_mitg_enable ? MSR_OP_OR : MSR_OP_ANDNOT) |
1439 	    (all_cpus ? MSR_OP_RENDEZVOUS_ALL : MSR_OP_LOCAL),
1440 	    IA32_RNGDS_MITG_DIS, NULL);
1441 }
1442 
1443 static int
1444 sysctl_rngds_mitg_enable_handler(SYSCTL_HANDLER_ARGS)
1445 {
1446 	int error, val;
1447 
1448 	val = x86_rngds_mitg_enable;
1449 	error = sysctl_handle_int(oidp, &val, 0, req);
1450 	if (error != 0 || req->newptr == NULL)
1451 		return (error);
1452 	x86_rngds_mitg_enable = val;
1453 	x86_rngds_mitg_recalculate(true);
1454 	return (0);
1455 }
1456 SYSCTL_PROC(_machdep_mitigations_rngds, OID_AUTO, enable, CTLTYPE_INT |
1457     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1458     sysctl_rngds_mitg_enable_handler, "I",
1459     "MCU Optimization, disabling RDSEED mitigation control "
1460     "(0 - mitigation disabled (RDSEED optimized), 1 - mitigation enabled)");
1461 
1462 static int
1463 sysctl_rngds_state_handler(SYSCTL_HANDLER_ARGS)
1464 {
1465 	const char *state;
1466 
1467 	if ((cpu_stdext_feature3 & CPUID_STDEXT3_MCUOPT) == 0) {
1468 		state = "Not applicable";
1469 	} else if (x86_rngds_mitg_enable == 0) {
1470 		state = "RDSEED not serialized";
1471 	} else {
1472 		state = "Mitigated";
1473 	}
1474 	return (SYSCTL_OUT(req, state, strlen(state)));
1475 }
1476 SYSCTL_PROC(_machdep_mitigations_rngds, OID_AUTO, state,
1477     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1478     sysctl_rngds_state_handler, "A",
1479     "MCU Optimization state");
1480 
1481 /*
1482  * Enable and restore kernel text write permissions.
1483  * Callers must ensure that disable_wp()/restore_wp() are executed
1484  * without rescheduling on the same core.
1485  */
1486 bool
1487 disable_wp(void)
1488 {
1489 	u_int cr0;
1490 
1491 	cr0 = rcr0();
1492 	if ((cr0 & CR0_WP) == 0)
1493 		return (false);
1494 	load_cr0(cr0 & ~CR0_WP);
1495 	return (true);
1496 }
1497 
1498 void
1499 restore_wp(bool old_wp)
1500 {
1501 
1502 	if (old_wp)
1503 		load_cr0(rcr0() | CR0_WP);
1504 }
1505 
1506 bool
1507 acpi_get_fadt_bootflags(uint16_t *flagsp)
1508 {
1509 #ifdef DEV_ACPI
1510 	ACPI_TABLE_FADT *fadt;
1511 	vm_paddr_t physaddr;
1512 
1513 	physaddr = acpi_find_table(ACPI_SIG_FADT);
1514 	if (physaddr == 0)
1515 		return (false);
1516 	fadt = acpi_map_table(physaddr, ACPI_SIG_FADT);
1517 	if (fadt == NULL)
1518 		return (false);
1519 	*flagsp = fadt->BootFlags;
1520 	acpi_unmap_table(fadt);
1521 	return (true);
1522 #else
1523 	return (false);
1524 #endif
1525 }
1526 
1527 DEFINE_IFUNC(, uint64_t, rdtsc_ordered, (void))
1528 {
1529 	bool cpu_is_amd = cpu_vendor_id == CPU_VENDOR_AMD ||
1530 	    cpu_vendor_id == CPU_VENDOR_HYGON;
1531 
1532 	if ((amd_feature & AMDID_RDTSCP) != 0)
1533 		return (rdtscp);
1534 	else if ((cpu_feature & CPUID_SSE2) != 0)
1535 		return (cpu_is_amd ? rdtsc_ordered_mfence :
1536 		    rdtsc_ordered_lfence);
1537 	else
1538 		return (rdtsc);
1539 }
1540