xref: /freebsd/sys/x86/x86/cpu_machdep.c (revision a134ebd6e63f658f2d3d04ac0c60d23bcaa86dd7)
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  * Machine dependent boot() routine
198  *
199  * I haven't seen anything to put here yet
200  * Possibly some stuff might be grafted back here from boot()
201  */
202 void
203 cpu_boot(int howto)
204 {
205 }
206 
207 /*
208  * Flush the D-cache for non-DMA I/O so that the I-cache can
209  * be made coherent later.
210  */
211 void
212 cpu_flush_dcache(void *ptr, size_t len)
213 {
214 	/* Not applicable */
215 }
216 
217 void
218 acpi_cpu_c1(void)
219 {
220 
221 	__asm __volatile("sti; hlt");
222 }
223 
224 /*
225  * Use mwait to pause execution while waiting for an interrupt or
226  * another thread to signal that there is more work.
227  *
228  * NOTE: Interrupts will cause a wakeup; however, this function does
229  * not enable interrupt handling. The caller is responsible to enable
230  * interrupts.
231  */
232 void
233 acpi_cpu_idle_mwait(uint32_t mwait_hint)
234 {
235 	int *state;
236 	uint64_t v;
237 
238 	/*
239 	 * A comment in Linux patch claims that 'CPUs run faster with
240 	 * speculation protection disabled. All CPU threads in a core
241 	 * must disable speculation protection for it to be
242 	 * disabled. Disable it while we are idle so the other
243 	 * hyperthread can run fast.'
244 	 *
245 	 * XXXKIB.  Software coordination mode should be supported,
246 	 * but all Intel CPUs provide hardware coordination.
247 	 */
248 
249 	state = &PCPU_PTR(monitorbuf)->idle_state;
250 	KASSERT(atomic_load_int(state) == STATE_SLEEPING,
251 	    ("cpu_mwait_cx: wrong monitorbuf state"));
252 	atomic_store_int(state, STATE_MWAIT);
253 	if (PCPU_GET(ibpb_set) || hw_ssb_active) {
254 		v = rdmsr(MSR_IA32_SPEC_CTRL);
255 		wrmsr(MSR_IA32_SPEC_CTRL, v & ~(IA32_SPEC_CTRL_IBRS |
256 		    IA32_SPEC_CTRL_STIBP | IA32_SPEC_CTRL_SSBD));
257 	} else {
258 		v = 0;
259 	}
260 	cpu_monitor(state, 0, 0);
261 	if (atomic_load_int(state) == STATE_MWAIT)
262 		cpu_mwait(MWAIT_INTRBREAK, mwait_hint);
263 
264 	/*
265 	 * SSB cannot be disabled while we sleep, or rather, if it was
266 	 * disabled, the sysctl thread will bind to our cpu to tweak
267 	 * MSR.
268 	 */
269 	if (v != 0)
270 		wrmsr(MSR_IA32_SPEC_CTRL, v);
271 
272 	/*
273 	 * We should exit on any event that interrupts mwait, because
274 	 * that event might be a wanted interrupt.
275 	 */
276 	atomic_store_int(state, STATE_RUNNING);
277 }
278 
279 /* Get current clock frequency for the given cpu id. */
280 int
281 cpu_est_clockrate(int cpu_id, uint64_t *rate)
282 {
283 	uint64_t tsc1, tsc2;
284 	uint64_t acnt, mcnt, perf;
285 	register_t reg;
286 
287 	if (pcpu_find(cpu_id) == NULL || rate == NULL)
288 		return (EINVAL);
289 #ifdef __i386__
290 	if ((cpu_feature & CPUID_TSC) == 0)
291 		return (EOPNOTSUPP);
292 #endif
293 
294 	/*
295 	 * If TSC is P-state invariant and APERF/MPERF MSRs do not exist,
296 	 * DELAY(9) based logic fails.
297 	 */
298 	if (tsc_is_invariant && !tsc_perf_stat)
299 		return (EOPNOTSUPP);
300 
301 #ifdef SMP
302 	if (smp_cpus > 1) {
303 		/* Schedule ourselves on the indicated cpu. */
304 		thread_lock(curthread);
305 		sched_bind(curthread, cpu_id);
306 		thread_unlock(curthread);
307 	}
308 #endif
309 
310 	/* Calibrate by measuring a short delay. */
311 	reg = intr_disable();
312 	if (tsc_is_invariant) {
313 		wrmsr(MSR_MPERF, 0);
314 		wrmsr(MSR_APERF, 0);
315 		tsc1 = rdtsc();
316 		DELAY(1000);
317 		mcnt = rdmsr(MSR_MPERF);
318 		acnt = rdmsr(MSR_APERF);
319 		tsc2 = rdtsc();
320 		intr_restore(reg);
321 		perf = 1000 * acnt / mcnt;
322 		*rate = (tsc2 - tsc1) * perf;
323 	} else {
324 		tsc1 = rdtsc();
325 		DELAY(1000);
326 		tsc2 = rdtsc();
327 		intr_restore(reg);
328 		*rate = (tsc2 - tsc1) * 1000;
329 	}
330 
331 #ifdef SMP
332 	if (smp_cpus > 1) {
333 		thread_lock(curthread);
334 		sched_unbind(curthread);
335 		thread_unlock(curthread);
336 	}
337 #endif
338 
339 	return (0);
340 }
341 
342 /*
343  * Shutdown the CPU as much as possible
344  */
345 void
346 cpu_halt(void)
347 {
348 	for (;;)
349 		halt();
350 }
351 
352 static void
353 cpu_reset_real(void)
354 {
355 	struct region_descriptor null_idt;
356 	int b;
357 
358 	disable_intr();
359 #ifdef CPU_ELAN
360 	if (elan_mmcr != NULL)
361 		elan_mmcr->RESCFG = 1;
362 #endif
363 #ifdef __i386__
364 	if (cpu == CPU_GEODE1100) {
365 		/* Attempt Geode's own reset */
366 		outl(0xcf8, 0x80009044ul);
367 		outl(0xcfc, 0xf);
368 	}
369 #endif
370 #if !defined(BROKEN_KEYBOARD_RESET)
371 	/*
372 	 * Attempt to do a CPU reset via the keyboard controller,
373 	 * do not turn off GateA20, as any machine that fails
374 	 * to do the reset here would then end up in no man's land.
375 	 */
376 	outb(IO_KBD + 4, 0xFE);
377 	DELAY(500000);	/* wait 0.5 sec to see if that did it */
378 #endif
379 
380 	/*
381 	 * Attempt to force a reset via the Reset Control register at
382 	 * I/O port 0xcf9.  Bit 2 forces a system reset when it
383 	 * transitions from 0 to 1.  Bit 1 selects the type of reset
384 	 * to attempt: 0 selects a "soft" reset, and 1 selects a
385 	 * "hard" reset.  We try a "hard" reset.  The first write sets
386 	 * bit 1 to select a "hard" reset and clears bit 2.  The
387 	 * second write forces a 0 -> 1 transition in bit 2 to trigger
388 	 * a reset.
389 	 */
390 	outb(0xcf9, 0x2);
391 	outb(0xcf9, 0x6);
392 	DELAY(500000);  /* wait 0.5 sec to see if that did it */
393 
394 	/*
395 	 * Attempt to force a reset via the Fast A20 and Init register
396 	 * at I/O port 0x92.  Bit 1 serves as an alternate A20 gate.
397 	 * Bit 0 asserts INIT# when set to 1.  We are careful to only
398 	 * preserve bit 1 while setting bit 0.  We also must clear bit
399 	 * 0 before setting it if it isn't already clear.
400 	 */
401 	b = inb(0x92);
402 	if (b != 0xff) {
403 		if ((b & 0x1) != 0)
404 			outb(0x92, b & 0xfe);
405 		outb(0x92, b | 0x1);
406 		DELAY(500000);  /* wait 0.5 sec to see if that did it */
407 	}
408 
409 	printf("No known reset method worked, attempting CPU shutdown\n");
410 	DELAY(1000000); /* wait 1 sec for printf to complete */
411 
412 	/* Wipe the IDT. */
413 	null_idt.rd_limit = 0;
414 	null_idt.rd_base = 0;
415 	lidt(&null_idt);
416 
417 	/* "good night, sweet prince .... <THUNK!>" */
418 	breakpoint();
419 
420 	/* NOTREACHED */
421 	while(1);
422 }
423 
424 #ifdef SMP
425 static void
426 cpu_reset_proxy(void)
427 {
428 
429 	cpu_reset_proxy_active = 1;
430 	while (cpu_reset_proxy_active == 1)
431 		ia32_pause(); /* Wait for other cpu to see that we've started */
432 
433 	printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid);
434 	DELAY(1000000);
435 	cpu_reset_real();
436 }
437 #endif
438 
439 void
440 cpu_reset(void)
441 {
442 #ifdef SMP
443 	struct monitorbuf *mb;
444 	cpuset_t map;
445 	u_int cnt;
446 
447 	if (smp_started) {
448 		map = all_cpus;
449 		CPU_CLR(PCPU_GET(cpuid), &map);
450 		CPU_ANDNOT(&map, &stopped_cpus);
451 		if (!CPU_EMPTY(&map)) {
452 			printf("cpu_reset: Stopping other CPUs\n");
453 			stop_cpus(map);
454 		}
455 
456 		if (PCPU_GET(cpuid) != 0) {
457 			cpu_reset_proxyid = PCPU_GET(cpuid);
458 			cpustop_restartfunc = cpu_reset_proxy;
459 			cpu_reset_proxy_active = 0;
460 			printf("cpu_reset: Restarting BSP\n");
461 
462 			/* Restart CPU #0. */
463 			CPU_SETOF(0, &started_cpus);
464 			mb = &pcpu_find(0)->pc_monitorbuf;
465 			atomic_store_int(&mb->stop_state,
466 			    MONITOR_STOPSTATE_RUNNING);
467 
468 			cnt = 0;
469 			while (cpu_reset_proxy_active == 0 && cnt < 10000000) {
470 				ia32_pause();
471 				cnt++;	/* Wait for BSP to announce restart */
472 			}
473 			if (cpu_reset_proxy_active == 0) {
474 				printf("cpu_reset: Failed to restart BSP\n");
475 			} else {
476 				cpu_reset_proxy_active = 2;
477 				while (1)
478 					ia32_pause();
479 				/* NOTREACHED */
480 			}
481 		}
482 
483 		DELAY(1000000);
484 	}
485 #endif
486 	cpu_reset_real();
487 	/* NOTREACHED */
488 }
489 
490 bool
491 cpu_mwait_usable(void)
492 {
493 
494 	return ((cpu_feature2 & CPUID2_MON) != 0 && ((cpu_mon_mwait_flags &
495 	    (CPUID5_MON_MWAIT_EXT | CPUID5_MWAIT_INTRBREAK)) ==
496 	    (CPUID5_MON_MWAIT_EXT | CPUID5_MWAIT_INTRBREAK)));
497 }
498 
499 void (*cpu_idle_hook)(sbintime_t) = NULL;	/* ACPI idle hook. */
500 static int	cpu_ident_amdc1e = 0;	/* AMD C1E supported. */
501 static int	idle_mwait = 1;		/* Use MONITOR/MWAIT for short idle. */
502 SYSCTL_INT(_machdep, OID_AUTO, idle_mwait, CTLFLAG_RWTUN, &idle_mwait,
503     0, "Use MONITOR/MWAIT for short idle");
504 
505 static void
506 cpu_idle_acpi(sbintime_t sbt)
507 {
508 	int *state;
509 
510 	state = &PCPU_PTR(monitorbuf)->idle_state;
511 	atomic_store_int(state, STATE_SLEEPING);
512 
513 	/* See comments in cpu_idle_hlt(). */
514 	disable_intr();
515 	if (sched_runnable())
516 		enable_intr();
517 	else if (cpu_idle_hook)
518 		cpu_idle_hook(sbt);
519 	else
520 		acpi_cpu_c1();
521 	atomic_store_int(state, STATE_RUNNING);
522 }
523 
524 static void
525 cpu_idle_hlt(sbintime_t sbt)
526 {
527 	int *state;
528 
529 	state = &PCPU_PTR(monitorbuf)->idle_state;
530 	atomic_store_int(state, STATE_SLEEPING);
531 
532 	/*
533 	 * Since we may be in a critical section from cpu_idle(), if
534 	 * an interrupt fires during that critical section we may have
535 	 * a pending preemption.  If the CPU halts, then that thread
536 	 * may not execute until a later interrupt awakens the CPU.
537 	 * To handle this race, check for a runnable thread after
538 	 * disabling interrupts and immediately return if one is
539 	 * found.  Also, we must absolutely guarentee that hlt is
540 	 * the next instruction after sti.  This ensures that any
541 	 * interrupt that fires after the call to disable_intr() will
542 	 * immediately awaken the CPU from hlt.  Finally, please note
543 	 * that on x86 this works fine because of interrupts enabled only
544 	 * after the instruction following sti takes place, while IF is set
545 	 * to 1 immediately, allowing hlt instruction to acknowledge the
546 	 * interrupt.
547 	 */
548 	disable_intr();
549 	if (sched_runnable())
550 		enable_intr();
551 	else
552 		acpi_cpu_c1();
553 	atomic_store_int(state, STATE_RUNNING);
554 }
555 
556 static void
557 cpu_idle_mwait(sbintime_t sbt)
558 {
559 	int *state;
560 
561 	state = &PCPU_PTR(monitorbuf)->idle_state;
562 	atomic_store_int(state, STATE_MWAIT);
563 
564 	/* See comments in cpu_idle_hlt(). */
565 	disable_intr();
566 	if (sched_runnable()) {
567 		atomic_store_int(state, STATE_RUNNING);
568 		enable_intr();
569 		return;
570 	}
571 
572 	cpu_monitor(state, 0, 0);
573 	if (atomic_load_int(state) == STATE_MWAIT)
574 		__asm __volatile("sti; mwait" : : "a" (MWAIT_C1), "c" (0));
575 	else
576 		enable_intr();
577 	atomic_store_int(state, STATE_RUNNING);
578 }
579 
580 static void
581 cpu_idle_spin(sbintime_t sbt)
582 {
583 	int *state;
584 	int i;
585 
586 	state = &PCPU_PTR(monitorbuf)->idle_state;
587 	atomic_store_int(state, STATE_RUNNING);
588 
589 	/*
590 	 * The sched_runnable() call is racy but as long as there is
591 	 * a loop missing it one time will have just a little impact if any
592 	 * (and it is much better than missing the check at all).
593 	 */
594 	for (i = 0; i < 1000; i++) {
595 		if (sched_runnable())
596 			return;
597 		cpu_spinwait();
598 	}
599 }
600 
601 /*
602  * C1E renders the local APIC timer dead, so we disable it by
603  * reading the Interrupt Pending Message register and clearing
604  * both C1eOnCmpHalt (bit 28) and SmiOnCmpHalt (bit 27).
605  *
606  * Reference:
607  *   "BIOS and Kernel Developer's Guide for AMD NPT Family 0Fh Processors"
608  *   #32559 revision 3.00+
609  */
610 #define	MSR_AMDK8_IPM		0xc0010055
611 #define	AMDK8_SMIONCMPHALT	(1ULL << 27)
612 #define	AMDK8_C1EONCMPHALT	(1ULL << 28)
613 #define	AMDK8_CMPHALT		(AMDK8_SMIONCMPHALT | AMDK8_C1EONCMPHALT)
614 
615 void
616 cpu_probe_amdc1e(void)
617 {
618 
619 	/*
620 	 * Detect the presence of C1E capability mostly on latest
621 	 * dual-cores (or future) k8 family.
622 	 */
623 	if (cpu_vendor_id == CPU_VENDOR_AMD &&
624 	    (cpu_id & 0x00000f00) == 0x00000f00 &&
625 	    (cpu_id & 0x0fff0000) >=  0x00040000) {
626 		cpu_ident_amdc1e = 1;
627 	}
628 }
629 
630 void (*cpu_idle_fn)(sbintime_t) = cpu_idle_acpi;
631 
632 void
633 cpu_idle(int busy)
634 {
635 	uint64_t msr;
636 	sbintime_t sbt = -1;
637 
638 	CTR2(KTR_SPARE2, "cpu_idle(%d) at %d",
639 	    busy, curcpu);
640 #ifdef MP_WATCHDOG
641 	ap_watchdog(PCPU_GET(cpuid));
642 #endif
643 
644 	/* If we are busy - try to use fast methods. */
645 	if (busy) {
646 		if ((cpu_feature2 & CPUID2_MON) && idle_mwait) {
647 			cpu_idle_mwait(busy);
648 			goto out;
649 		}
650 	}
651 
652 	/* If we have time - switch timers into idle mode. */
653 	if (!busy) {
654 		critical_enter();
655 		sbt = cpu_idleclock();
656 	}
657 
658 	/* Apply AMD APIC timer C1E workaround. */
659 	if (cpu_ident_amdc1e && cpu_disable_c3_sleep) {
660 		msr = rdmsr(MSR_AMDK8_IPM);
661 		if (msr & AMDK8_CMPHALT)
662 			wrmsr(MSR_AMDK8_IPM, msr & ~AMDK8_CMPHALT);
663 	}
664 
665 	/* Call main idle method. */
666 	cpu_idle_fn(sbt);
667 
668 	/* Switch timers back into active mode. */
669 	if (!busy) {
670 		cpu_activeclock();
671 		critical_exit();
672 	}
673 out:
674 	CTR2(KTR_SPARE2, "cpu_idle(%d) at %d done",
675 	    busy, curcpu);
676 }
677 
678 static int cpu_idle_apl31_workaround;
679 SYSCTL_INT(_machdep, OID_AUTO, idle_apl31, CTLFLAG_RW,
680     &cpu_idle_apl31_workaround, 0,
681     "Apollo Lake APL31 MWAIT bug workaround");
682 
683 int
684 cpu_idle_wakeup(int cpu)
685 {
686 	struct monitorbuf *mb;
687 	int *state;
688 
689 	mb = &pcpu_find(cpu)->pc_monitorbuf;
690 	state = &mb->idle_state;
691 	switch (atomic_load_int(state)) {
692 	case STATE_SLEEPING:
693 		return (0);
694 	case STATE_MWAIT:
695 		atomic_store_int(state, STATE_RUNNING);
696 		return (cpu_idle_apl31_workaround ? 0 : 1);
697 	case STATE_RUNNING:
698 		return (1);
699 	default:
700 		panic("bad monitor state");
701 		return (1);
702 	}
703 }
704 
705 /*
706  * Ordered by speed/power consumption.
707  */
708 static struct {
709 	void	*id_fn;
710 	char	*id_name;
711 	int	id_cpuid2_flag;
712 } idle_tbl[] = {
713 	{ .id_fn = cpu_idle_spin, .id_name = "spin" },
714 	{ .id_fn = cpu_idle_mwait, .id_name = "mwait",
715 	    .id_cpuid2_flag = CPUID2_MON },
716 	{ .id_fn = cpu_idle_hlt, .id_name = "hlt" },
717 	{ .id_fn = cpu_idle_acpi, .id_name = "acpi" },
718 };
719 
720 static int
721 idle_sysctl_available(SYSCTL_HANDLER_ARGS)
722 {
723 	char *avail, *p;
724 	int error;
725 	int i;
726 
727 	avail = malloc(256, M_TEMP, M_WAITOK);
728 	p = avail;
729 	for (i = 0; i < nitems(idle_tbl); i++) {
730 		if (idle_tbl[i].id_cpuid2_flag != 0 &&
731 		    (cpu_feature2 & idle_tbl[i].id_cpuid2_flag) == 0)
732 			continue;
733 		if (strcmp(idle_tbl[i].id_name, "acpi") == 0 &&
734 		    cpu_idle_hook == NULL)
735 			continue;
736 		p += sprintf(p, "%s%s", p != avail ? ", " : "",
737 		    idle_tbl[i].id_name);
738 	}
739 	error = sysctl_handle_string(oidp, avail, 0, req);
740 	free(avail, M_TEMP);
741 	return (error);
742 }
743 
744 SYSCTL_PROC(_machdep, OID_AUTO, idle_available,
745     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_NEEDGIANT,
746     0, 0, idle_sysctl_available, "A",
747     "list of available idle functions");
748 
749 static bool
750 cpu_idle_selector(const char *new_idle_name)
751 {
752 	int i;
753 
754 	for (i = 0; i < nitems(idle_tbl); i++) {
755 		if (idle_tbl[i].id_cpuid2_flag != 0 &&
756 		    (cpu_feature2 & idle_tbl[i].id_cpuid2_flag) == 0)
757 			continue;
758 		if (strcmp(idle_tbl[i].id_name, "acpi") == 0 &&
759 		    cpu_idle_hook == NULL)
760 			continue;
761 		if (strcmp(idle_tbl[i].id_name, new_idle_name))
762 			continue;
763 		cpu_idle_fn = idle_tbl[i].id_fn;
764 		if (bootverbose)
765 			printf("CPU idle set to %s\n", idle_tbl[i].id_name);
766 		return (true);
767 	}
768 	return (false);
769 }
770 
771 static int
772 cpu_idle_sysctl(SYSCTL_HANDLER_ARGS)
773 {
774 	char buf[16], *p;
775 	int error, i;
776 
777 	p = "unknown";
778 	for (i = 0; i < nitems(idle_tbl); i++) {
779 		if (idle_tbl[i].id_fn == cpu_idle_fn) {
780 			p = idle_tbl[i].id_name;
781 			break;
782 		}
783 	}
784 	strncpy(buf, p, sizeof(buf));
785 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
786 	if (error != 0 || req->newptr == NULL)
787 		return (error);
788 	return (cpu_idle_selector(buf) ? 0 : EINVAL);
789 }
790 
791 SYSCTL_PROC(_machdep, OID_AUTO, idle,
792     CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_NEEDGIANT,
793     0, 0, cpu_idle_sysctl, "A",
794     "currently selected idle function");
795 
796 static void
797 cpu_idle_tun(void *unused __unused)
798 {
799 	char tunvar[16];
800 
801 	if (TUNABLE_STR_FETCH("machdep.idle", tunvar, sizeof(tunvar)))
802 		cpu_idle_selector(tunvar);
803 	else if (cpu_vendor_id == CPU_VENDOR_AMD &&
804 	    CPUID_TO_FAMILY(cpu_id) == 0x17 && CPUID_TO_MODEL(cpu_id) == 0x1) {
805 		/* Ryzen erratas 1057, 1109. */
806 		cpu_idle_selector("hlt");
807 		idle_mwait = 0;
808 		mwait_cpustop_broken = true;
809 	}
810 
811 	if (cpu_vendor_id == CPU_VENDOR_INTEL && cpu_id == 0x506c9) {
812 		/*
813 		 * Apollo Lake errata APL31 (public errata APL30).
814 		 * Stores to the armed address range may not trigger
815 		 * MWAIT to resume execution.  OS needs to use
816 		 * interrupts to wake processors from MWAIT-induced
817 		 * sleep states.
818 		 */
819 		cpu_idle_apl31_workaround = 1;
820 		mwait_cpustop_broken = true;
821 	}
822 	TUNABLE_INT_FETCH("machdep.idle_apl31", &cpu_idle_apl31_workaround);
823 }
824 SYSINIT(cpu_idle_tun, SI_SUB_CPU, SI_ORDER_MIDDLE, cpu_idle_tun, NULL);
825 
826 static int panic_on_nmi = 0xff;
827 SYSCTL_INT(_machdep, OID_AUTO, panic_on_nmi, CTLFLAG_RWTUN,
828     &panic_on_nmi, 0,
829     "Panic on NMI: 1 = H/W failure; 2 = unknown; 0xff = all");
830 int nmi_is_broadcast = 1;
831 SYSCTL_INT(_machdep, OID_AUTO, nmi_is_broadcast, CTLFLAG_RWTUN,
832     &nmi_is_broadcast, 0,
833     "Chipset NMI is broadcast");
834 int (*apei_nmi)(void);
835 
836 void
837 nmi_call_kdb(u_int cpu, u_int type, struct trapframe *frame)
838 {
839 	bool claimed = false;
840 
841 #ifdef DEV_ISA
842 	/* machine/parity/power fail/"kitchen sink" faults */
843 	if (isa_nmi(frame->tf_err)) {
844 		claimed = true;
845 		if ((panic_on_nmi & 1) != 0)
846 			panic("NMI indicates hardware failure");
847 	}
848 #endif /* DEV_ISA */
849 
850 	/* ACPI Platform Error Interfaces callback. */
851 	if (apei_nmi != NULL && (*apei_nmi)())
852 		claimed = true;
853 
854 	/*
855 	 * NMIs can be useful for debugging.  They can be hooked up to a
856 	 * pushbutton, usually on an ISA, PCI, or PCIe card.  They can also be
857 	 * generated by an IPMI BMC, either manually or in response to a
858 	 * watchdog timeout.  For example, see the "power diag" command in
859 	 * ports/sysutils/ipmitool.  They can also be generated by a
860 	 * hypervisor; see "bhyvectl --inject-nmi".
861 	 */
862 
863 #ifdef KDB
864 	if (!claimed && (panic_on_nmi & 2) != 0) {
865 		if (debugger_on_panic) {
866 			printf("NMI/cpu%d ... going to debugger\n", cpu);
867 			claimed = kdb_trap(type, 0, frame);
868 		}
869 	}
870 #endif /* KDB */
871 
872 	if (!claimed && panic_on_nmi != 0)
873 		panic("NMI");
874 }
875 
876 void
877 nmi_handle_intr(u_int type, struct trapframe *frame)
878 {
879 
880 #ifdef SMP
881 	if (nmi_is_broadcast) {
882 		nmi_call_kdb_smp(type, frame);
883 		return;
884 	}
885 #endif
886 	nmi_call_kdb(PCPU_GET(cpuid), type, frame);
887 }
888 
889 static int hw_ibrs_active;
890 int hw_ibrs_ibpb_active;
891 int hw_ibrs_disable = 1;
892 
893 SYSCTL_INT(_hw, OID_AUTO, ibrs_active, CTLFLAG_RD, &hw_ibrs_active, 0,
894     "Indirect Branch Restricted Speculation active");
895 
896 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, ibrs,
897     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
898     "Indirect Branch Restricted Speculation active");
899 
900 SYSCTL_INT(_machdep_mitigations_ibrs, OID_AUTO, active, CTLFLAG_RD,
901     &hw_ibrs_active, 0, "Indirect Branch Restricted Speculation active");
902 
903 void
904 hw_ibrs_recalculate(bool for_all_cpus)
905 {
906 	if ((cpu_ia32_arch_caps & IA32_ARCH_CAP_IBRS_ALL) != 0) {
907 		x86_msr_op(MSR_IA32_SPEC_CTRL, (for_all_cpus ?
908 		    MSR_OP_RENDEZVOUS : MSR_OP_LOCAL) |
909 		    (hw_ibrs_disable != 0 ? MSR_OP_ANDNOT : MSR_OP_OR),
910 		    IA32_SPEC_CTRL_IBRS);
911 		hw_ibrs_active = hw_ibrs_disable == 0;
912 		hw_ibrs_ibpb_active = 0;
913 	} else {
914 		hw_ibrs_active = hw_ibrs_ibpb_active = (cpu_stdext_feature3 &
915 		    CPUID_STDEXT3_IBPB) != 0 && !hw_ibrs_disable;
916 	}
917 }
918 
919 static int
920 hw_ibrs_disable_handler(SYSCTL_HANDLER_ARGS)
921 {
922 	int error, val;
923 
924 	val = hw_ibrs_disable;
925 	error = sysctl_handle_int(oidp, &val, 0, req);
926 	if (error != 0 || req->newptr == NULL)
927 		return (error);
928 	hw_ibrs_disable = val != 0;
929 	hw_ibrs_recalculate(true);
930 	return (0);
931 }
932 SYSCTL_PROC(_hw, OID_AUTO, ibrs_disable, CTLTYPE_INT | CTLFLAG_RWTUN |
933     CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0, hw_ibrs_disable_handler, "I",
934     "Disable Indirect Branch Restricted Speculation");
935 
936 SYSCTL_PROC(_machdep_mitigations_ibrs, OID_AUTO, disable, CTLTYPE_INT |
937     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
938     hw_ibrs_disable_handler, "I",
939     "Disable Indirect Branch Restricted Speculation");
940 
941 int hw_ssb_active;
942 int hw_ssb_disable;
943 
944 SYSCTL_INT(_hw, OID_AUTO, spec_store_bypass_disable_active, CTLFLAG_RD,
945     &hw_ssb_active, 0,
946     "Speculative Store Bypass Disable active");
947 
948 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, ssb,
949     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
950     "Speculative Store Bypass Disable active");
951 
952 SYSCTL_INT(_machdep_mitigations_ssb, OID_AUTO, active, CTLFLAG_RD,
953     &hw_ssb_active, 0, "Speculative Store Bypass Disable active");
954 
955 static void
956 hw_ssb_set(bool enable, bool for_all_cpus)
957 {
958 
959 	if ((cpu_stdext_feature3 & CPUID_STDEXT3_SSBD) == 0) {
960 		hw_ssb_active = 0;
961 		return;
962 	}
963 	hw_ssb_active = enable;
964 	x86_msr_op(MSR_IA32_SPEC_CTRL,
965 	    (enable ? MSR_OP_OR : MSR_OP_ANDNOT) |
966 	    (for_all_cpus ? MSR_OP_SCHED : MSR_OP_LOCAL), IA32_SPEC_CTRL_SSBD);
967 }
968 
969 void
970 hw_ssb_recalculate(bool all_cpus)
971 {
972 
973 	switch (hw_ssb_disable) {
974 	default:
975 		hw_ssb_disable = 0;
976 		/* FALLTHROUGH */
977 	case 0: /* off */
978 		hw_ssb_set(false, all_cpus);
979 		break;
980 	case 1: /* on */
981 		hw_ssb_set(true, all_cpus);
982 		break;
983 	case 2: /* auto */
984 		hw_ssb_set((cpu_ia32_arch_caps & IA32_ARCH_CAP_SSB_NO) != 0 ?
985 		    false : true, all_cpus);
986 		break;
987 	}
988 }
989 
990 static int
991 hw_ssb_disable_handler(SYSCTL_HANDLER_ARGS)
992 {
993 	int error, val;
994 
995 	val = hw_ssb_disable;
996 	error = sysctl_handle_int(oidp, &val, 0, req);
997 	if (error != 0 || req->newptr == NULL)
998 		return (error);
999 	hw_ssb_disable = val;
1000 	hw_ssb_recalculate(true);
1001 	return (0);
1002 }
1003 SYSCTL_PROC(_hw, OID_AUTO, spec_store_bypass_disable, CTLTYPE_INT |
1004     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1005     hw_ssb_disable_handler, "I",
1006     "Speculative Store Bypass Disable (0 - off, 1 - on, 2 - auto");
1007 
1008 SYSCTL_PROC(_machdep_mitigations_ssb, OID_AUTO, disable, CTLTYPE_INT |
1009     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1010     hw_ssb_disable_handler, "I",
1011     "Speculative Store Bypass Disable (0 - off, 1 - on, 2 - auto");
1012 
1013 int hw_mds_disable;
1014 
1015 /*
1016  * Handler for Microarchitectural Data Sampling issues.  Really not a
1017  * pointer to C function: on amd64 the code must not change any CPU
1018  * architectural state except possibly %rflags. Also, it is always
1019  * called with interrupts disabled.
1020  */
1021 void mds_handler_void(void);
1022 void mds_handler_verw(void);
1023 void mds_handler_ivb(void);
1024 void mds_handler_bdw(void);
1025 void mds_handler_skl_sse(void);
1026 void mds_handler_skl_avx(void);
1027 void mds_handler_skl_avx512(void);
1028 void mds_handler_silvermont(void);
1029 void (*mds_handler)(void) = mds_handler_void;
1030 
1031 static int
1032 sysctl_hw_mds_disable_state_handler(SYSCTL_HANDLER_ARGS)
1033 {
1034 	const char *state;
1035 
1036 	if (mds_handler == mds_handler_void)
1037 		state = "inactive";
1038 	else if (mds_handler == mds_handler_verw)
1039 		state = "VERW";
1040 	else if (mds_handler == mds_handler_ivb)
1041 		state = "software IvyBridge";
1042 	else if (mds_handler == mds_handler_bdw)
1043 		state = "software Broadwell";
1044 	else if (mds_handler == mds_handler_skl_sse)
1045 		state = "software Skylake SSE";
1046 	else if (mds_handler == mds_handler_skl_avx)
1047 		state = "software Skylake AVX";
1048 	else if (mds_handler == mds_handler_skl_avx512)
1049 		state = "software Skylake AVX512";
1050 	else if (mds_handler == mds_handler_silvermont)
1051 		state = "software Silvermont";
1052 	else
1053 		state = "unknown";
1054 	return (SYSCTL_OUT(req, state, strlen(state)));
1055 }
1056 
1057 SYSCTL_PROC(_hw, OID_AUTO, mds_disable_state,
1058     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1059     sysctl_hw_mds_disable_state_handler, "A",
1060     "Microarchitectural Data Sampling Mitigation state");
1061 
1062 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, mds,
1063     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1064     "Microarchitectural Data Sampling Mitigation state");
1065 
1066 SYSCTL_PROC(_machdep_mitigations_mds, OID_AUTO, state,
1067     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1068     sysctl_hw_mds_disable_state_handler, "A",
1069     "Microarchitectural Data Sampling Mitigation state");
1070 
1071 _Static_assert(__offsetof(struct pcpu, pc_mds_tmp) % 64 == 0, "MDS AVX512");
1072 
1073 void
1074 hw_mds_recalculate(void)
1075 {
1076 	struct pcpu *pc;
1077 	vm_offset_t b64;
1078 	u_long xcr0;
1079 	int i;
1080 
1081 	/*
1082 	 * Allow user to force VERW variant even if MD_CLEAR is not
1083 	 * reported.  For instance, hypervisor might unknowingly
1084 	 * filter the cap out.
1085 	 * For the similar reasons, and for testing, allow to enable
1086 	 * mitigation even when MDS_NO cap is set.
1087 	 */
1088 	if (cpu_vendor_id != CPU_VENDOR_INTEL || hw_mds_disable == 0 ||
1089 	    ((cpu_ia32_arch_caps & IA32_ARCH_CAP_MDS_NO) != 0 &&
1090 	    hw_mds_disable == 3)) {
1091 		mds_handler = mds_handler_void;
1092 	} else if (((cpu_stdext_feature3 & CPUID_STDEXT3_MD_CLEAR) != 0 &&
1093 	    hw_mds_disable == 3) || hw_mds_disable == 1) {
1094 		mds_handler = mds_handler_verw;
1095 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1096 	    (CPUID_TO_MODEL(cpu_id) == 0x2e || CPUID_TO_MODEL(cpu_id) == 0x1e ||
1097 	    CPUID_TO_MODEL(cpu_id) == 0x1f || CPUID_TO_MODEL(cpu_id) == 0x1a ||
1098 	    CPUID_TO_MODEL(cpu_id) == 0x2f || CPUID_TO_MODEL(cpu_id) == 0x25 ||
1099 	    CPUID_TO_MODEL(cpu_id) == 0x2c || CPUID_TO_MODEL(cpu_id) == 0x2d ||
1100 	    CPUID_TO_MODEL(cpu_id) == 0x2a || CPUID_TO_MODEL(cpu_id) == 0x3e ||
1101 	    CPUID_TO_MODEL(cpu_id) == 0x3a) &&
1102 	    (hw_mds_disable == 2 || hw_mds_disable == 3)) {
1103 		/*
1104 		 * Nehalem, SandyBridge, IvyBridge
1105 		 */
1106 		CPU_FOREACH(i) {
1107 			pc = pcpu_find(i);
1108 			if (pc->pc_mds_buf == NULL) {
1109 				pc->pc_mds_buf = malloc_domainset(672, M_TEMP,
1110 				    DOMAINSET_PREF(pc->pc_domain), M_WAITOK);
1111 				bzero(pc->pc_mds_buf, 16);
1112 			}
1113 		}
1114 		mds_handler = mds_handler_ivb;
1115 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1116 	    (CPUID_TO_MODEL(cpu_id) == 0x3f || CPUID_TO_MODEL(cpu_id) == 0x3c ||
1117 	    CPUID_TO_MODEL(cpu_id) == 0x45 || CPUID_TO_MODEL(cpu_id) == 0x46 ||
1118 	    CPUID_TO_MODEL(cpu_id) == 0x56 || CPUID_TO_MODEL(cpu_id) == 0x4f ||
1119 	    CPUID_TO_MODEL(cpu_id) == 0x47 || CPUID_TO_MODEL(cpu_id) == 0x3d) &&
1120 	    (hw_mds_disable == 2 || hw_mds_disable == 3)) {
1121 		/*
1122 		 * Haswell, Broadwell
1123 		 */
1124 		CPU_FOREACH(i) {
1125 			pc = pcpu_find(i);
1126 			if (pc->pc_mds_buf == NULL) {
1127 				pc->pc_mds_buf = malloc_domainset(1536, M_TEMP,
1128 				    DOMAINSET_PREF(pc->pc_domain), M_WAITOK);
1129 				bzero(pc->pc_mds_buf, 16);
1130 			}
1131 		}
1132 		mds_handler = mds_handler_bdw;
1133 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1134 	    ((CPUID_TO_MODEL(cpu_id) == 0x55 && (cpu_id &
1135 	    CPUID_STEPPING) <= 5) ||
1136 	    CPUID_TO_MODEL(cpu_id) == 0x4e || CPUID_TO_MODEL(cpu_id) == 0x5e ||
1137 	    (CPUID_TO_MODEL(cpu_id) == 0x8e && (cpu_id &
1138 	    CPUID_STEPPING) <= 0xb) ||
1139 	    (CPUID_TO_MODEL(cpu_id) == 0x9e && (cpu_id &
1140 	    CPUID_STEPPING) <= 0xc)) &&
1141 	    (hw_mds_disable == 2 || hw_mds_disable == 3)) {
1142 		/*
1143 		 * Skylake, KabyLake, CoffeeLake, WhiskeyLake,
1144 		 * CascadeLake
1145 		 */
1146 		CPU_FOREACH(i) {
1147 			pc = pcpu_find(i);
1148 			if (pc->pc_mds_buf == NULL) {
1149 				pc->pc_mds_buf = malloc_domainset(6 * 1024,
1150 				    M_TEMP, DOMAINSET_PREF(pc->pc_domain),
1151 				    M_WAITOK);
1152 				b64 = (vm_offset_t)malloc_domainset(64 + 63,
1153 				    M_TEMP, DOMAINSET_PREF(pc->pc_domain),
1154 				    M_WAITOK);
1155 				pc->pc_mds_buf64 = (void *)roundup2(b64, 64);
1156 				bzero(pc->pc_mds_buf64, 64);
1157 			}
1158 		}
1159 		xcr0 = rxcr(0);
1160 		if ((xcr0 & XFEATURE_ENABLED_ZMM_HI256) != 0 &&
1161 		    (cpu_stdext_feature & CPUID_STDEXT_AVX512DQ) != 0)
1162 			mds_handler = mds_handler_skl_avx512;
1163 		else if ((xcr0 & XFEATURE_ENABLED_AVX) != 0 &&
1164 		    (cpu_feature2 & CPUID2_AVX) != 0)
1165 			mds_handler = mds_handler_skl_avx;
1166 		else
1167 			mds_handler = mds_handler_skl_sse;
1168 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1169 	    ((CPUID_TO_MODEL(cpu_id) == 0x37 ||
1170 	    CPUID_TO_MODEL(cpu_id) == 0x4a ||
1171 	    CPUID_TO_MODEL(cpu_id) == 0x4c ||
1172 	    CPUID_TO_MODEL(cpu_id) == 0x4d ||
1173 	    CPUID_TO_MODEL(cpu_id) == 0x5a ||
1174 	    CPUID_TO_MODEL(cpu_id) == 0x5d ||
1175 	    CPUID_TO_MODEL(cpu_id) == 0x6e ||
1176 	    CPUID_TO_MODEL(cpu_id) == 0x65 ||
1177 	    CPUID_TO_MODEL(cpu_id) == 0x75 ||
1178 	    CPUID_TO_MODEL(cpu_id) == 0x1c ||
1179 	    CPUID_TO_MODEL(cpu_id) == 0x26 ||
1180 	    CPUID_TO_MODEL(cpu_id) == 0x27 ||
1181 	    CPUID_TO_MODEL(cpu_id) == 0x35 ||
1182 	    CPUID_TO_MODEL(cpu_id) == 0x36 ||
1183 	    CPUID_TO_MODEL(cpu_id) == 0x7a))) {
1184 		/* Silvermont, Airmont */
1185 		CPU_FOREACH(i) {
1186 			pc = pcpu_find(i);
1187 			if (pc->pc_mds_buf == NULL)
1188 				pc->pc_mds_buf = malloc(256, M_TEMP, M_WAITOK);
1189 		}
1190 		mds_handler = mds_handler_silvermont;
1191 	} else {
1192 		hw_mds_disable = 0;
1193 		mds_handler = mds_handler_void;
1194 	}
1195 }
1196 
1197 static void
1198 hw_mds_recalculate_boot(void *arg __unused)
1199 {
1200 
1201 	hw_mds_recalculate();
1202 }
1203 SYSINIT(mds_recalc, SI_SUB_SMP, SI_ORDER_ANY, hw_mds_recalculate_boot, NULL);
1204 
1205 static int
1206 sysctl_mds_disable_handler(SYSCTL_HANDLER_ARGS)
1207 {
1208 	int error, val;
1209 
1210 	val = hw_mds_disable;
1211 	error = sysctl_handle_int(oidp, &val, 0, req);
1212 	if (error != 0 || req->newptr == NULL)
1213 		return (error);
1214 	if (val < 0 || val > 3)
1215 		return (EINVAL);
1216 	hw_mds_disable = val;
1217 	hw_mds_recalculate();
1218 	return (0);
1219 }
1220 
1221 SYSCTL_PROC(_hw, OID_AUTO, mds_disable, CTLTYPE_INT |
1222     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1223     sysctl_mds_disable_handler, "I",
1224     "Microarchitectural Data Sampling Mitigation "
1225     "(0 - off, 1 - on VERW, 2 - on SW, 3 - on AUTO");
1226 
1227 SYSCTL_PROC(_machdep_mitigations_mds, OID_AUTO, disable, CTLTYPE_INT |
1228     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1229     sysctl_mds_disable_handler, "I",
1230     "Microarchitectural Data Sampling Mitigation "
1231     "(0 - off, 1 - on VERW, 2 - on SW, 3 - on AUTO");
1232 
1233 /*
1234  * Intel Transactional Memory Asynchronous Abort Mitigation
1235  * CVE-2019-11135
1236  */
1237 int x86_taa_enable;
1238 int x86_taa_state;
1239 enum {
1240 	TAA_NONE	= 0,	/* No mitigation enabled */
1241 	TAA_TSX_DISABLE	= 1,	/* Disable TSX via MSR */
1242 	TAA_VERW	= 2,	/* Use VERW mitigation */
1243 	TAA_AUTO	= 3,	/* Automatically select the mitigation */
1244 
1245 	/* The states below are not selectable by the operator */
1246 
1247 	TAA_TAA_UC	= 4,	/* Mitigation present in microcode */
1248 	TAA_NOT_PRESENT	= 5	/* TSX is not present */
1249 };
1250 
1251 static void
1252 taa_set(bool enable, bool all)
1253 {
1254 
1255 	x86_msr_op(MSR_IA32_TSX_CTRL,
1256 	    (enable ? MSR_OP_OR : MSR_OP_ANDNOT) |
1257 	    (all ? MSR_OP_RENDEZVOUS : MSR_OP_LOCAL),
1258 	    IA32_TSX_CTRL_RTM_DISABLE | IA32_TSX_CTRL_TSX_CPUID_CLEAR);
1259 }
1260 
1261 void
1262 x86_taa_recalculate(void)
1263 {
1264 	static int taa_saved_mds_disable = 0;
1265 	int taa_need = 0, taa_state = 0;
1266 	int mds_disable = 0, need_mds_recalc = 0;
1267 
1268 	/* Check CPUID.07h.EBX.HLE and RTM for the presence of TSX */
1269 	if ((cpu_stdext_feature & CPUID_STDEXT_HLE) == 0 ||
1270 	    (cpu_stdext_feature & CPUID_STDEXT_RTM) == 0) {
1271 		/* TSX is not present */
1272 		x86_taa_state = TAA_NOT_PRESENT;
1273 		return;
1274 	}
1275 
1276 	/* Check to see what mitigation options the CPU gives us */
1277 	if (cpu_ia32_arch_caps & IA32_ARCH_CAP_TAA_NO) {
1278 		/* CPU is not suseptible to TAA */
1279 		taa_need = TAA_TAA_UC;
1280 	} else if (cpu_ia32_arch_caps & IA32_ARCH_CAP_TSX_CTRL) {
1281 		/*
1282 		 * CPU can turn off TSX.  This is the next best option
1283 		 * if TAA_NO hardware mitigation isn't present
1284 		 */
1285 		taa_need = TAA_TSX_DISABLE;
1286 	} else {
1287 		/* No TSX/TAA specific remedies are available. */
1288 		if (x86_taa_enable == TAA_TSX_DISABLE) {
1289 			if (bootverbose)
1290 				printf("TSX control not available\n");
1291 			return;
1292 		} else
1293 			taa_need = TAA_VERW;
1294 	}
1295 
1296 	/* Can we automatically take action, or are we being forced? */
1297 	if (x86_taa_enable == TAA_AUTO)
1298 		taa_state = taa_need;
1299 	else
1300 		taa_state = x86_taa_enable;
1301 
1302 	/* No state change, nothing to do */
1303 	if (taa_state == x86_taa_state) {
1304 		if (bootverbose)
1305 			printf("No TSX change made\n");
1306 		return;
1307 	}
1308 
1309 	/* Does the MSR need to be turned on or off? */
1310 	if (taa_state == TAA_TSX_DISABLE)
1311 		taa_set(true, true);
1312 	else if (x86_taa_state == TAA_TSX_DISABLE)
1313 		taa_set(false, true);
1314 
1315 	/* Does MDS need to be set to turn on VERW? */
1316 	if (taa_state == TAA_VERW) {
1317 		taa_saved_mds_disable = hw_mds_disable;
1318 		mds_disable = hw_mds_disable = 1;
1319 		need_mds_recalc = 1;
1320 	} else if (x86_taa_state == TAA_VERW) {
1321 		mds_disable = hw_mds_disable = taa_saved_mds_disable;
1322 		need_mds_recalc = 1;
1323 	}
1324 	if (need_mds_recalc) {
1325 		hw_mds_recalculate();
1326 		if (mds_disable != hw_mds_disable) {
1327 			if (bootverbose)
1328 				printf("Cannot change MDS state for TAA\n");
1329 			/* Don't update our state */
1330 			return;
1331 		}
1332 	}
1333 
1334 	x86_taa_state = taa_state;
1335 	return;
1336 }
1337 
1338 static void
1339 taa_recalculate_boot(void * arg __unused)
1340 {
1341 
1342 	x86_taa_recalculate();
1343 }
1344 SYSINIT(taa_recalc, SI_SUB_SMP, SI_ORDER_ANY, taa_recalculate_boot, NULL);
1345 
1346 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, taa,
1347     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1348     "TSX Asynchronous Abort Mitigation");
1349 
1350 static int
1351 sysctl_taa_handler(SYSCTL_HANDLER_ARGS)
1352 {
1353 	int error, val;
1354 
1355 	val = x86_taa_enable;
1356 	error = sysctl_handle_int(oidp, &val, 0, req);
1357 	if (error != 0 || req->newptr == NULL)
1358 		return (error);
1359 	if (val < TAA_NONE || val > TAA_AUTO)
1360 		return (EINVAL);
1361 	x86_taa_enable = val;
1362 	x86_taa_recalculate();
1363 	return (0);
1364 }
1365 
1366 SYSCTL_PROC(_machdep_mitigations_taa, OID_AUTO, enable, CTLTYPE_INT |
1367     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1368     sysctl_taa_handler, "I",
1369     "TAA Mitigation enablement control "
1370     "(0 - off, 1 - disable TSX, 2 - VERW, 3 - on AUTO");
1371 
1372 static int
1373 sysctl_taa_state_handler(SYSCTL_HANDLER_ARGS)
1374 {
1375 	const char *state;
1376 
1377 	switch (x86_taa_state) {
1378 	case TAA_NONE:
1379 		state = "inactive";
1380 		break;
1381 	case TAA_TSX_DISABLE:
1382 		state = "TSX disabled";
1383 		break;
1384 	case TAA_VERW:
1385 		state = "VERW";
1386 		break;
1387 	case TAA_TAA_UC:
1388 		state = "Mitigated in microcode";
1389 		break;
1390 	case TAA_NOT_PRESENT:
1391 		state = "TSX not present";
1392 		break;
1393 	default:
1394 		state = "unknown";
1395 	}
1396 
1397 	return (SYSCTL_OUT(req, state, strlen(state)));
1398 }
1399 
1400 SYSCTL_PROC(_machdep_mitigations_taa, OID_AUTO, state,
1401     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1402     sysctl_taa_state_handler, "A",
1403     "TAA Mitigation state");
1404 
1405 int __read_frequently cpu_flush_rsb_ctxsw;
1406 SYSCTL_INT(_machdep_mitigations, OID_AUTO, flush_rsb_ctxsw,
1407     CTLFLAG_RW | CTLFLAG_NOFETCH, &cpu_flush_rsb_ctxsw, 0,
1408     "Flush Return Stack Buffer on context switch");
1409 
1410 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, rngds,
1411     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1412     "MCU Optimization, disable RDSEED mitigation");
1413 
1414 int x86_rngds_mitg_enable = 1;
1415 void
1416 x86_rngds_mitg_recalculate(bool all_cpus)
1417 {
1418 	if ((cpu_stdext_feature3 & CPUID_STDEXT3_MCUOPT) == 0)
1419 		return;
1420 	x86_msr_op(MSR_IA32_MCU_OPT_CTRL,
1421 	    (x86_rngds_mitg_enable ? MSR_OP_OR : MSR_OP_ANDNOT) |
1422 	    (all_cpus ? MSR_OP_RENDEZVOUS : MSR_OP_LOCAL),
1423 	    IA32_RNGDS_MITG_DIS);
1424 }
1425 
1426 static int
1427 sysctl_rngds_mitg_enable_handler(SYSCTL_HANDLER_ARGS)
1428 {
1429 	int error, val;
1430 
1431 	val = x86_rngds_mitg_enable;
1432 	error = sysctl_handle_int(oidp, &val, 0, req);
1433 	if (error != 0 || req->newptr == NULL)
1434 		return (error);
1435 	x86_rngds_mitg_enable = val;
1436 	x86_rngds_mitg_recalculate(true);
1437 	return (0);
1438 }
1439 SYSCTL_PROC(_machdep_mitigations_rngds, OID_AUTO, enable, CTLTYPE_INT |
1440     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1441     sysctl_rngds_mitg_enable_handler, "I",
1442     "MCU Optimization, disabling RDSEED mitigation control "
1443     "(0 - mitigation disabled (RDSEED optimized), 1 - mitigation enabled");
1444 
1445 static int
1446 sysctl_rngds_state_handler(SYSCTL_HANDLER_ARGS)
1447 {
1448 	const char *state;
1449 
1450 	if ((cpu_stdext_feature3 & CPUID_STDEXT3_MCUOPT) == 0) {
1451 		state = "Not applicable";
1452 	} else if (x86_rngds_mitg_enable == 0) {
1453 		state = "RDSEED not serialized";
1454 	} else {
1455 		state = "Mitigated";
1456 	}
1457 	return (SYSCTL_OUT(req, state, strlen(state)));
1458 }
1459 SYSCTL_PROC(_machdep_mitigations_rngds, OID_AUTO, state,
1460     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1461     sysctl_rngds_state_handler, "A",
1462     "MCU Optimization state");
1463 
1464 /*
1465  * Enable and restore kernel text write permissions.
1466  * Callers must ensure that disable_wp()/restore_wp() are executed
1467  * without rescheduling on the same core.
1468  */
1469 bool
1470 disable_wp(void)
1471 {
1472 	u_int cr0;
1473 
1474 	cr0 = rcr0();
1475 	if ((cr0 & CR0_WP) == 0)
1476 		return (false);
1477 	load_cr0(cr0 & ~CR0_WP);
1478 	return (true);
1479 }
1480 
1481 void
1482 restore_wp(bool old_wp)
1483 {
1484 
1485 	if (old_wp)
1486 		load_cr0(rcr0() | CR0_WP);
1487 }
1488 
1489 bool
1490 acpi_get_fadt_bootflags(uint16_t *flagsp)
1491 {
1492 #ifdef DEV_ACPI
1493 	ACPI_TABLE_FADT *fadt;
1494 	vm_paddr_t physaddr;
1495 
1496 	physaddr = acpi_find_table(ACPI_SIG_FADT);
1497 	if (physaddr == 0)
1498 		return (false);
1499 	fadt = acpi_map_table(physaddr, ACPI_SIG_FADT);
1500 	if (fadt == NULL)
1501 		return (false);
1502 	*flagsp = fadt->BootFlags;
1503 	acpi_unmap_table(fadt);
1504 	return (true);
1505 #else
1506 	return (false);
1507 #endif
1508 }
1509