xref: /freebsd/sys/x86/x86/cpu_machdep.c (revision e9ac41698b2f322d55ccf9da50a3596edb2c1800)
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 
39 #include <sys/cdefs.h>
40 #include "opt_acpi.h"
41 #include "opt_atpic.h"
42 #include "opt_cpu.h"
43 #include "opt_ddb.h"
44 #include "opt_inet.h"
45 #include "opt_isa.h"
46 #include "opt_kdb.h"
47 #include "opt_kstack_pages.h"
48 #include "opt_maxmem.h"
49 #include "opt_platform.h"
50 #include "opt_sched.h"
51 #ifdef __i386__
52 #include "opt_apic.h"
53 #endif
54 
55 #include <sys/param.h>
56 #include <sys/proc.h>
57 #include <sys/systm.h>
58 #include <sys/bus.h>
59 #include <sys/cpu.h>
60 #include <sys/domainset.h>
61 #include <sys/kdb.h>
62 #include <sys/kernel.h>
63 #include <sys/ktr.h>
64 #include <sys/lock.h>
65 #include <sys/malloc.h>
66 #include <sys/mutex.h>
67 #include <sys/pcpu.h>
68 #include <sys/rwlock.h>
69 #include <sys/sched.h>
70 #include <sys/smp.h>
71 #include <sys/sysctl.h>
72 
73 #include <machine/clock.h>
74 #include <machine/cpu.h>
75 #include <machine/cpufunc.h>
76 #include <machine/cputypes.h>
77 #include <machine/specialreg.h>
78 #include <machine/md_var.h>
79 #include <machine/tss.h>
80 #ifdef SMP
81 #include <machine/smp.h>
82 #endif
83 #ifdef CPU_ELAN
84 #include <machine/elan_mmcr.h>
85 #endif
86 #include <x86/acpica_machdep.h>
87 #include <x86/ifunc.h>
88 
89 #include <vm/vm.h>
90 #include <vm/vm_extern.h>
91 #include <vm/vm_kern.h>
92 #include <vm/vm_page.h>
93 #include <vm/vm_map.h>
94 #include <vm/vm_object.h>
95 #include <vm/vm_pager.h>
96 #include <vm/vm_param.h>
97 
98 #include <isa/isareg.h>
99 
100 #include <contrib/dev/acpica/include/acpi.h>
101 
102 #define	STATE_RUNNING	0x0
103 #define	STATE_MWAIT	0x1
104 #define	STATE_SLEEPING	0x2
105 
106 #ifdef SMP
107 static u_int	cpu_reset_proxyid;
108 static volatile u_int	cpu_reset_proxy_active;
109 #endif
110 
111 char bootmethod[16];
112 SYSCTL_STRING(_machdep, OID_AUTO, bootmethod, CTLFLAG_RD, bootmethod, 0,
113     "System firmware boot method");
114 
115 struct msr_op_arg {
116 	u_int msr;
117 	int op;
118 	uint64_t arg1;
119 	uint64_t *res;
120 };
121 
122 static void
123 x86_msr_op_one(void *argp)
124 {
125 	struct msr_op_arg *a;
126 	uint64_t v;
127 
128 	a = argp;
129 	switch (a->op) {
130 	case MSR_OP_ANDNOT:
131 		v = rdmsr(a->msr);
132 		v &= ~a->arg1;
133 		wrmsr(a->msr, v);
134 		break;
135 	case MSR_OP_OR:
136 		v = rdmsr(a->msr);
137 		v |= a->arg1;
138 		wrmsr(a->msr, v);
139 		break;
140 	case MSR_OP_WRITE:
141 		wrmsr(a->msr, a->arg1);
142 		break;
143 	case MSR_OP_READ:
144 		v = rdmsr(a->msr);
145 		*a->res = v;
146 		break;
147 	}
148 }
149 
150 #define	MSR_OP_EXMODE_MASK	0xf0000000
151 #define	MSR_OP_OP_MASK		0x000000ff
152 #define	MSR_OP_GET_CPUID(x)	(((x) & ~MSR_OP_EXMODE_MASK) >> 8)
153 
154 void
155 x86_msr_op(u_int msr, u_int op, uint64_t arg1, uint64_t *res)
156 {
157 	struct thread *td;
158 	struct msr_op_arg a;
159 	cpuset_t set;
160 	u_int exmode;
161 	int bound_cpu, cpu, i, is_bound;
162 
163 	a.op = op & MSR_OP_OP_MASK;
164 	MPASS(a.op == MSR_OP_ANDNOT || a.op == MSR_OP_OR ||
165 	    a.op == MSR_OP_WRITE || a.op == MSR_OP_READ);
166 	exmode = op & MSR_OP_EXMODE_MASK;
167 	MPASS(exmode == MSR_OP_LOCAL || exmode == MSR_OP_SCHED_ALL ||
168 	    exmode == MSR_OP_SCHED_ONE || exmode == MSR_OP_RENDEZVOUS_ALL ||
169 	    exmode == MSR_OP_RENDEZVOUS_ONE);
170 	a.msr = msr;
171 	a.arg1 = arg1;
172 	a.res = res;
173 	switch (exmode) {
174 	case MSR_OP_LOCAL:
175 		x86_msr_op_one(&a);
176 		break;
177 	case MSR_OP_SCHED_ALL:
178 		td = curthread;
179 		thread_lock(td);
180 		is_bound = sched_is_bound(td);
181 		bound_cpu = td->td_oncpu;
182 		CPU_FOREACH(i) {
183 			sched_bind(td, i);
184 			x86_msr_op_one(&a);
185 		}
186 		if (is_bound)
187 			sched_bind(td, bound_cpu);
188 		else
189 			sched_unbind(td);
190 		thread_unlock(td);
191 		break;
192 	case MSR_OP_SCHED_ONE:
193 		td = curthread;
194 		cpu = MSR_OP_GET_CPUID(op);
195 		thread_lock(td);
196 		is_bound = sched_is_bound(td);
197 		bound_cpu = td->td_oncpu;
198 		if (!is_bound || bound_cpu != cpu)
199 			sched_bind(td, cpu);
200 		x86_msr_op_one(&a);
201 		if (is_bound) {
202 			if (bound_cpu != cpu)
203 				sched_bind(td, bound_cpu);
204 		} else {
205 			sched_unbind(td);
206 		}
207 		thread_unlock(td);
208 		break;
209 	case MSR_OP_RENDEZVOUS_ALL:
210 		smp_rendezvous(smp_no_rendezvous_barrier, x86_msr_op_one,
211 		    smp_no_rendezvous_barrier, &a);
212 		break;
213 	case MSR_OP_RENDEZVOUS_ONE:
214 		cpu = MSR_OP_GET_CPUID(op);
215 		CPU_SETOF(cpu, &set);
216 		smp_rendezvous_cpus(set, smp_no_rendezvous_barrier,
217 		    x86_msr_op_one, smp_no_rendezvous_barrier, &a);
218 		break;
219 	}
220 }
221 
222 /*
223  * Automatically initialized per CPU errata in cpu_idle_tun below.
224  */
225 bool mwait_cpustop_broken = false;
226 SYSCTL_BOOL(_machdep, OID_AUTO, mwait_cpustop_broken, CTLFLAG_RDTUN,
227     &mwait_cpustop_broken, 0,
228     "Can not reliably wake MONITOR/MWAIT cpus without interrupts");
229 
230 /*
231  * Flush the D-cache for non-DMA I/O so that the I-cache can
232  * be made coherent later.
233  */
234 void
235 cpu_flush_dcache(void *ptr, size_t len)
236 {
237 	/* Not applicable */
238 }
239 
240 void
241 acpi_cpu_c1(void)
242 {
243 
244 	__asm __volatile("sti; hlt");
245 }
246 
247 /*
248  * Use mwait to pause execution while waiting for an interrupt or
249  * another thread to signal that there is more work.
250  *
251  * NOTE: Interrupts will cause a wakeup; however, this function does
252  * not enable interrupt handling. The caller is responsible to enable
253  * interrupts.
254  */
255 void
256 acpi_cpu_idle_mwait(uint32_t mwait_hint)
257 {
258 	int *state;
259 	uint64_t v;
260 
261 	/*
262 	 * A comment in Linux patch claims that 'CPUs run faster with
263 	 * speculation protection disabled. All CPU threads in a core
264 	 * must disable speculation protection for it to be
265 	 * disabled. Disable it while we are idle so the other
266 	 * hyperthread can run fast.'
267 	 *
268 	 * XXXKIB.  Software coordination mode should be supported,
269 	 * but all Intel CPUs provide hardware coordination.
270 	 */
271 
272 	state = &PCPU_PTR(monitorbuf)->idle_state;
273 	KASSERT(atomic_load_int(state) == STATE_SLEEPING,
274 	    ("cpu_mwait_cx: wrong monitorbuf state"));
275 	atomic_store_int(state, STATE_MWAIT);
276 	if (PCPU_GET(ibpb_set) || hw_ssb_active) {
277 		v = rdmsr(MSR_IA32_SPEC_CTRL);
278 		wrmsr(MSR_IA32_SPEC_CTRL, v & ~(IA32_SPEC_CTRL_IBRS |
279 		    IA32_SPEC_CTRL_STIBP | IA32_SPEC_CTRL_SSBD));
280 	} else {
281 		v = 0;
282 	}
283 	cpu_monitor(state, 0, 0);
284 	if (atomic_load_int(state) == STATE_MWAIT)
285 		cpu_mwait(MWAIT_INTRBREAK, mwait_hint);
286 
287 	/*
288 	 * SSB cannot be disabled while we sleep, or rather, if it was
289 	 * disabled, the sysctl thread will bind to our cpu to tweak
290 	 * MSR.
291 	 */
292 	if (v != 0)
293 		wrmsr(MSR_IA32_SPEC_CTRL, v);
294 
295 	/*
296 	 * We should exit on any event that interrupts mwait, because
297 	 * that event might be a wanted interrupt.
298 	 */
299 	atomic_store_int(state, STATE_RUNNING);
300 }
301 
302 /* Get current clock frequency for the given cpu id. */
303 int
304 cpu_est_clockrate(int cpu_id, uint64_t *rate)
305 {
306 	uint64_t tsc1, tsc2;
307 	uint64_t acnt, mcnt, perf;
308 	register_t reg;
309 
310 	if (pcpu_find(cpu_id) == NULL || rate == NULL)
311 		return (EINVAL);
312 #ifdef __i386__
313 	if ((cpu_feature & CPUID_TSC) == 0)
314 		return (EOPNOTSUPP);
315 #endif
316 
317 	/*
318 	 * If TSC is P-state invariant and APERF/MPERF MSRs do not exist,
319 	 * DELAY(9) based logic fails.
320 	 */
321 	if (tsc_is_invariant && !tsc_perf_stat)
322 		return (EOPNOTSUPP);
323 
324 #ifdef SMP
325 	if (smp_cpus > 1) {
326 		/* Schedule ourselves on the indicated cpu. */
327 		thread_lock(curthread);
328 		sched_bind(curthread, cpu_id);
329 		thread_unlock(curthread);
330 	}
331 #endif
332 
333 	/* Calibrate by measuring a short delay. */
334 	reg = intr_disable();
335 	if (tsc_is_invariant) {
336 		wrmsr(MSR_MPERF, 0);
337 		wrmsr(MSR_APERF, 0);
338 		tsc1 = rdtsc();
339 		DELAY(1000);
340 		mcnt = rdmsr(MSR_MPERF);
341 		acnt = rdmsr(MSR_APERF);
342 		tsc2 = rdtsc();
343 		intr_restore(reg);
344 		perf = 1000 * acnt / mcnt;
345 		*rate = (tsc2 - tsc1) * perf;
346 	} else {
347 		tsc1 = rdtsc();
348 		DELAY(1000);
349 		tsc2 = rdtsc();
350 		intr_restore(reg);
351 		*rate = (tsc2 - tsc1) * 1000;
352 	}
353 
354 #ifdef SMP
355 	if (smp_cpus > 1) {
356 		thread_lock(curthread);
357 		sched_unbind(curthread);
358 		thread_unlock(curthread);
359 	}
360 #endif
361 
362 	return (0);
363 }
364 
365 /*
366  * Shutdown the CPU as much as possible
367  */
368 void
369 cpu_halt(void)
370 {
371 	for (;;)
372 		halt();
373 }
374 
375 static void
376 cpu_reset_real(void)
377 {
378 	struct region_descriptor null_idt;
379 	int b;
380 
381 	disable_intr();
382 #ifdef CPU_ELAN
383 	if (elan_mmcr != NULL)
384 		elan_mmcr->RESCFG = 1;
385 #endif
386 #ifdef __i386__
387 	if (cpu == CPU_GEODE1100) {
388 		/* Attempt Geode's own reset */
389 		outl(0xcf8, 0x80009044ul);
390 		outl(0xcfc, 0xf);
391 	}
392 #endif
393 #if !defined(BROKEN_KEYBOARD_RESET)
394 	/*
395 	 * Attempt to do a CPU reset via the keyboard controller,
396 	 * do not turn off GateA20, as any machine that fails
397 	 * to do the reset here would then end up in no man's land.
398 	 */
399 	outb(IO_KBD + 4, 0xFE);
400 	DELAY(500000);	/* wait 0.5 sec to see if that did it */
401 #endif
402 
403 	/*
404 	 * Attempt to force a reset via the Reset Control register at
405 	 * I/O port 0xcf9.  Bit 2 forces a system reset when it
406 	 * transitions from 0 to 1.  Bit 1 selects the type of reset
407 	 * to attempt: 0 selects a "soft" reset, and 1 selects a
408 	 * "hard" reset.  We try a "hard" reset.  The first write sets
409 	 * bit 1 to select a "hard" reset and clears bit 2.  The
410 	 * second write forces a 0 -> 1 transition in bit 2 to trigger
411 	 * a reset.
412 	 */
413 	outb(0xcf9, 0x2);
414 	outb(0xcf9, 0x6);
415 	DELAY(500000);  /* wait 0.5 sec to see if that did it */
416 
417 	/*
418 	 * Attempt to force a reset via the Fast A20 and Init register
419 	 * at I/O port 0x92.  Bit 1 serves as an alternate A20 gate.
420 	 * Bit 0 asserts INIT# when set to 1.  We are careful to only
421 	 * preserve bit 1 while setting bit 0.  We also must clear bit
422 	 * 0 before setting it if it isn't already clear.
423 	 */
424 	b = inb(0x92);
425 	if (b != 0xff) {
426 		if ((b & 0x1) != 0)
427 			outb(0x92, b & 0xfe);
428 		outb(0x92, b | 0x1);
429 		DELAY(500000);  /* wait 0.5 sec to see if that did it */
430 	}
431 
432 	printf("No known reset method worked, attempting CPU shutdown\n");
433 	DELAY(1000000); /* wait 1 sec for printf to complete */
434 
435 	/* Wipe the IDT. */
436 	null_idt.rd_limit = 0;
437 	null_idt.rd_base = 0;
438 	lidt(&null_idt);
439 
440 	/* "good night, sweet prince .... <THUNK!>" */
441 	breakpoint();
442 
443 	/* NOTREACHED */
444 	while(1);
445 }
446 
447 #ifdef SMP
448 static void
449 cpu_reset_proxy(void)
450 {
451 
452 	cpu_reset_proxy_active = 1;
453 	while (cpu_reset_proxy_active == 1)
454 		ia32_pause(); /* Wait for other cpu to see that we've started */
455 
456 	printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid);
457 	DELAY(1000000);
458 	cpu_reset_real();
459 }
460 #endif
461 
462 void
463 cpu_reset(void)
464 {
465 #ifdef SMP
466 	struct monitorbuf *mb;
467 	cpuset_t map;
468 	u_int cnt;
469 
470 	if (smp_started) {
471 		map = all_cpus;
472 		CPU_CLR(PCPU_GET(cpuid), &map);
473 		CPU_ANDNOT(&map, &map, &stopped_cpus);
474 		if (!CPU_EMPTY(&map)) {
475 			printf("cpu_reset: Stopping other CPUs\n");
476 			stop_cpus(map);
477 		}
478 
479 		if (PCPU_GET(cpuid) != 0) {
480 			cpu_reset_proxyid = PCPU_GET(cpuid);
481 			cpustop_restartfunc = cpu_reset_proxy;
482 			cpu_reset_proxy_active = 0;
483 			printf("cpu_reset: Restarting BSP\n");
484 
485 			/* Restart CPU #0. */
486 			CPU_SETOF(0, &started_cpus);
487 			mb = &pcpu_find(0)->pc_monitorbuf;
488 			atomic_store_int(&mb->stop_state,
489 			    MONITOR_STOPSTATE_RUNNING);
490 
491 			cnt = 0;
492 			while (cpu_reset_proxy_active == 0 && cnt < 10000000) {
493 				ia32_pause();
494 				cnt++;	/* Wait for BSP to announce restart */
495 			}
496 			if (cpu_reset_proxy_active == 0) {
497 				printf("cpu_reset: Failed to restart BSP\n");
498 			} else {
499 				cpu_reset_proxy_active = 2;
500 				while (1)
501 					ia32_pause();
502 				/* NOTREACHED */
503 			}
504 		}
505 	}
506 #endif
507 	cpu_reset_real();
508 	/* NOTREACHED */
509 }
510 
511 bool
512 cpu_mwait_usable(void)
513 {
514 
515 	return ((cpu_feature2 & CPUID2_MON) != 0 && ((cpu_mon_mwait_flags &
516 	    (CPUID5_MON_MWAIT_EXT | CPUID5_MWAIT_INTRBREAK)) ==
517 	    (CPUID5_MON_MWAIT_EXT | CPUID5_MWAIT_INTRBREAK)));
518 }
519 
520 void (*cpu_idle_hook)(sbintime_t) = NULL;	/* ACPI idle hook. */
521 
522 int cpu_amdc1e_bug = 0;			/* AMD C1E APIC workaround required. */
523 
524 static int	idle_mwait = 1;		/* Use MONITOR/MWAIT for short idle. */
525 SYSCTL_INT(_machdep, OID_AUTO, idle_mwait, CTLFLAG_RWTUN, &idle_mwait,
526     0, "Use MONITOR/MWAIT for short idle");
527 
528 static bool
529 cpu_idle_enter(int *statep, int newstate)
530 {
531 	KASSERT(atomic_load_int(statep) == STATE_RUNNING,
532 	    ("%s: state %d", __func__, atomic_load_int(statep)));
533 
534 	/*
535 	 * A fence is needed to prevent reordering of the load in
536 	 * sched_runnable() with this store to the idle state word.  Without it,
537 	 * cpu_idle_wakeup() can observe the state as STATE_RUNNING after having
538 	 * added load to the queue, and elide an IPI.  Then, sched_runnable()
539 	 * can observe tdq_load == 0, so the CPU ends up idling with pending
540 	 * work.  tdq_notify() similarly ensures that a prior update to tdq_load
541 	 * is visible before calling cpu_idle_wakeup().
542 	 */
543 	atomic_store_int(statep, newstate);
544 #if defined(SCHED_ULE) && defined(SMP)
545 	atomic_thread_fence_seq_cst();
546 #endif
547 
548 	/*
549 	 * Since we may be in a critical section from cpu_idle(), if
550 	 * an interrupt fires during that critical section we may have
551 	 * a pending preemption.  If the CPU halts, then that thread
552 	 * may not execute until a later interrupt awakens the CPU.
553 	 * To handle this race, check for a runnable thread after
554 	 * disabling interrupts and immediately return if one is
555 	 * found.  Also, we must absolutely guarentee that hlt is
556 	 * the next instruction after sti.  This ensures that any
557 	 * interrupt that fires after the call to disable_intr() will
558 	 * immediately awaken the CPU from hlt.  Finally, please note
559 	 * that on x86 this works fine because of interrupts enabled only
560 	 * after the instruction following sti takes place, while IF is set
561 	 * to 1 immediately, allowing hlt instruction to acknowledge the
562 	 * interrupt.
563 	 */
564 	disable_intr();
565 	if (sched_runnable()) {
566 		enable_intr();
567 		atomic_store_int(statep, STATE_RUNNING);
568 		return (false);
569 	} else {
570 		return (true);
571 	}
572 }
573 
574 static void
575 cpu_idle_exit(int *statep)
576 {
577 	atomic_store_int(statep, STATE_RUNNING);
578 }
579 
580 static void
581 cpu_idle_acpi(sbintime_t sbt)
582 {
583 	int *state;
584 
585 	state = &PCPU_PTR(monitorbuf)->idle_state;
586 	if (cpu_idle_enter(state, STATE_SLEEPING)) {
587 		if (cpu_idle_hook)
588 			cpu_idle_hook(sbt);
589 		else
590 			acpi_cpu_c1();
591 		cpu_idle_exit(state);
592 	}
593 }
594 
595 static void
596 cpu_idle_hlt(sbintime_t sbt)
597 {
598 	int *state;
599 
600 	state = &PCPU_PTR(monitorbuf)->idle_state;
601 	if (cpu_idle_enter(state, STATE_SLEEPING)) {
602 		acpi_cpu_c1();
603 		atomic_store_int(state, STATE_RUNNING);
604 	}
605 }
606 
607 static void
608 cpu_idle_mwait(sbintime_t sbt)
609 {
610 	int *state;
611 
612 	state = &PCPU_PTR(monitorbuf)->idle_state;
613 	if (cpu_idle_enter(state, STATE_MWAIT)) {
614 		cpu_monitor(state, 0, 0);
615 		if (atomic_load_int(state) == STATE_MWAIT)
616 			__asm __volatile("sti; mwait" : : "a" (MWAIT_C1), "c" (0));
617 		else
618 			enable_intr();
619 		cpu_idle_exit(state);
620 	}
621 }
622 
623 static void
624 cpu_idle_spin(sbintime_t sbt)
625 {
626 	int *state;
627 	int i;
628 
629 	state = &PCPU_PTR(monitorbuf)->idle_state;
630 	atomic_store_int(state, STATE_RUNNING);
631 
632 	/*
633 	 * The sched_runnable() call is racy but as long as there is
634 	 * a loop missing it one time will have just a little impact if any
635 	 * (and it is much better than missing the check at all).
636 	 */
637 	for (i = 0; i < 1000; i++) {
638 		if (sched_runnable())
639 			return;
640 		cpu_spinwait();
641 	}
642 }
643 
644 void (*cpu_idle_fn)(sbintime_t) = cpu_idle_acpi;
645 
646 void
647 cpu_idle(int busy)
648 {
649 	uint64_t msr;
650 	sbintime_t sbt = -1;
651 
652 	CTR1(KTR_SPARE2, "cpu_idle(%d)", busy);
653 
654 	/* If we are busy - try to use fast methods. */
655 	if (busy) {
656 		if ((cpu_feature2 & CPUID2_MON) && idle_mwait) {
657 			cpu_idle_mwait(busy);
658 			goto out;
659 		}
660 	}
661 
662 	/* If we have time - switch timers into idle mode. */
663 	if (!busy) {
664 		critical_enter();
665 		sbt = cpu_idleclock();
666 	}
667 
668 	/* Apply AMD APIC timer C1E workaround. */
669 	if (cpu_amdc1e_bug && cpu_disable_c3_sleep) {
670 		msr = rdmsr(MSR_AMDK8_IPM);
671 		if ((msr & (AMDK8_SMIONCMPHALT | AMDK8_C1EONCMPHALT)) != 0)
672 			wrmsr(MSR_AMDK8_IPM, msr & ~(AMDK8_SMIONCMPHALT |
673 			    AMDK8_C1EONCMPHALT));
674 	}
675 
676 	/* Call main idle method. */
677 	cpu_idle_fn(sbt);
678 
679 	/* Switch timers back into active mode. */
680 	if (!busy) {
681 		cpu_activeclock();
682 		critical_exit();
683 	}
684 out:
685 	CTR1(KTR_SPARE2, "cpu_idle(%d) done", busy);
686 }
687 
688 static int cpu_idle_apl31_workaround;
689 SYSCTL_INT(_machdep, OID_AUTO, idle_apl31, CTLFLAG_RWTUN | CTLFLAG_NOFETCH,
690     &cpu_idle_apl31_workaround, 0,
691     "Apollo Lake APL31 MWAIT bug workaround");
692 
693 int
694 cpu_idle_wakeup(int cpu)
695 {
696 	struct monitorbuf *mb;
697 	int *state;
698 
699 	mb = &pcpu_find(cpu)->pc_monitorbuf;
700 	state = &mb->idle_state;
701 	switch (atomic_load_int(state)) {
702 	case STATE_SLEEPING:
703 		return (0);
704 	case STATE_MWAIT:
705 		atomic_store_int(state, STATE_RUNNING);
706 		return (cpu_idle_apl31_workaround ? 0 : 1);
707 	case STATE_RUNNING:
708 		return (1);
709 	default:
710 		panic("bad monitor state");
711 		return (1);
712 	}
713 }
714 
715 /*
716  * Ordered by speed/power consumption.
717  */
718 static const struct {
719 	void	*id_fn;
720 	const char *id_name;
721 	int	id_cpuid2_flag;
722 } idle_tbl[] = {
723 	{ .id_fn = cpu_idle_spin, .id_name = "spin" },
724 	{ .id_fn = cpu_idle_mwait, .id_name = "mwait",
725 	    .id_cpuid2_flag = CPUID2_MON },
726 	{ .id_fn = cpu_idle_hlt, .id_name = "hlt" },
727 	{ .id_fn = cpu_idle_acpi, .id_name = "acpi" },
728 };
729 
730 static int
731 idle_sysctl_available(SYSCTL_HANDLER_ARGS)
732 {
733 	char *avail, *p;
734 	int error;
735 	int i;
736 
737 	avail = malloc(256, M_TEMP, M_WAITOK);
738 	p = avail;
739 	for (i = 0; i < nitems(idle_tbl); i++) {
740 		if (idle_tbl[i].id_cpuid2_flag != 0 &&
741 		    (cpu_feature2 & idle_tbl[i].id_cpuid2_flag) == 0)
742 			continue;
743 		if (strcmp(idle_tbl[i].id_name, "acpi") == 0 &&
744 		    cpu_idle_hook == NULL)
745 			continue;
746 		p += sprintf(p, "%s%s", p != avail ? ", " : "",
747 		    idle_tbl[i].id_name);
748 	}
749 	error = sysctl_handle_string(oidp, avail, 0, req);
750 	free(avail, M_TEMP);
751 	return (error);
752 }
753 
754 SYSCTL_PROC(_machdep, OID_AUTO, idle_available,
755     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE,
756     0, 0, idle_sysctl_available, "A",
757     "list of available idle functions");
758 
759 static bool
760 cpu_idle_selector(const char *new_idle_name)
761 {
762 	int i;
763 
764 	for (i = 0; i < nitems(idle_tbl); i++) {
765 		if (idle_tbl[i].id_cpuid2_flag != 0 &&
766 		    (cpu_feature2 & idle_tbl[i].id_cpuid2_flag) == 0)
767 			continue;
768 		if (strcmp(idle_tbl[i].id_name, "acpi") == 0 &&
769 		    cpu_idle_hook == NULL)
770 			continue;
771 		if (strcmp(idle_tbl[i].id_name, new_idle_name))
772 			continue;
773 		cpu_idle_fn = idle_tbl[i].id_fn;
774 		if (bootverbose)
775 			printf("CPU idle set to %s\n", idle_tbl[i].id_name);
776 		return (true);
777 	}
778 	return (false);
779 }
780 
781 static int
782 cpu_idle_sysctl(SYSCTL_HANDLER_ARGS)
783 {
784 	char buf[16];
785 	const char *p;
786 	int error, i;
787 
788 	p = "unknown";
789 	for (i = 0; i < nitems(idle_tbl); i++) {
790 		if (idle_tbl[i].id_fn == cpu_idle_fn) {
791 			p = idle_tbl[i].id_name;
792 			break;
793 		}
794 	}
795 	strncpy(buf, p, sizeof(buf));
796 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
797 	if (error != 0 || req->newptr == NULL)
798 		return (error);
799 	return (cpu_idle_selector(buf) ? 0 : EINVAL);
800 }
801 
802 SYSCTL_PROC(_machdep, OID_AUTO, idle,
803     CTLTYPE_STRING | CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE,
804     0, 0, cpu_idle_sysctl, "A",
805     "currently selected idle function");
806 
807 static void
808 cpu_idle_tun(void *unused __unused)
809 {
810 	char tunvar[16];
811 
812 	if (TUNABLE_STR_FETCH("machdep.idle", tunvar, sizeof(tunvar)))
813 		cpu_idle_selector(tunvar);
814 	else if (cpu_vendor_id == CPU_VENDOR_AMD &&
815 	    CPUID_TO_FAMILY(cpu_id) == 0x17 && CPUID_TO_MODEL(cpu_id) == 0x1) {
816 		/* Ryzen erratas 1057, 1109. */
817 		cpu_idle_selector("hlt");
818 		idle_mwait = 0;
819 		mwait_cpustop_broken = true;
820 	}
821 
822 	if (cpu_vendor_id == CPU_VENDOR_INTEL &&
823 	    CPUID_TO_FAMILY(cpu_id) == 0x6 && CPUID_TO_MODEL(cpu_id) == 0x5c) {
824 		/*
825 		 * Apollo Lake errata APL31 (public errata APL30).
826 		 * Stores to the armed address range may not trigger
827 		 * MWAIT to resume execution.  OS needs to use
828 		 * interrupts to wake processors from MWAIT-induced
829 		 * sleep states.
830 		 */
831 		cpu_idle_apl31_workaround = 1;
832 		mwait_cpustop_broken = true;
833 	}
834 	TUNABLE_INT_FETCH("machdep.idle_apl31", &cpu_idle_apl31_workaround);
835 }
836 SYSINIT(cpu_idle_tun, SI_SUB_CPU, SI_ORDER_MIDDLE, cpu_idle_tun, NULL);
837 
838 static int panic_on_nmi = 0xff;
839 SYSCTL_INT(_machdep, OID_AUTO, panic_on_nmi, CTLFLAG_RWTUN,
840     &panic_on_nmi, 0,
841     "Panic on NMI: 1 = H/W failure; 2 = unknown; 0xff = all");
842 int nmi_is_broadcast = 1;
843 SYSCTL_INT(_machdep, OID_AUTO, nmi_is_broadcast, CTLFLAG_RWTUN,
844     &nmi_is_broadcast, 0,
845     "Chipset NMI is broadcast");
846 int (*apei_nmi)(void);
847 
848 void
849 nmi_call_kdb(u_int cpu, u_int type, struct trapframe *frame)
850 {
851 	bool claimed = false;
852 
853 #ifdef DEV_ISA
854 	/* machine/parity/power fail/"kitchen sink" faults */
855 	if (isa_nmi(frame->tf_err)) {
856 		claimed = true;
857 		if ((panic_on_nmi & 1) != 0)
858 			panic("NMI indicates hardware failure");
859 	}
860 #endif /* DEV_ISA */
861 
862 	/* ACPI Platform Error Interfaces callback. */
863 	if (apei_nmi != NULL && (*apei_nmi)())
864 		claimed = true;
865 
866 	/*
867 	 * NMIs can be useful for debugging.  They can be hooked up to a
868 	 * pushbutton, usually on an ISA, PCI, or PCIe card.  They can also be
869 	 * generated by an IPMI BMC, either manually or in response to a
870 	 * watchdog timeout.  For example, see the "power diag" command in
871 	 * ports/sysutils/ipmitool.  They can also be generated by a
872 	 * hypervisor; see "bhyvectl --inject-nmi".
873 	 */
874 
875 #ifdef KDB
876 	if (!claimed && (panic_on_nmi & 2) != 0) {
877 		if (debugger_on_panic) {
878 			printf("NMI/cpu%d ... going to debugger\n", cpu);
879 			claimed = kdb_trap(type, 0, frame);
880 		}
881 	}
882 #endif /* KDB */
883 
884 	if (!claimed && panic_on_nmi != 0)
885 		panic("NMI");
886 }
887 
888 void
889 nmi_handle_intr(u_int type, struct trapframe *frame)
890 {
891 
892 #ifdef SMP
893 	if (nmi_is_broadcast) {
894 		nmi_call_kdb_smp(type, frame);
895 		return;
896 	}
897 #endif
898 	nmi_call_kdb(PCPU_GET(cpuid), type, frame);
899 }
900 
901 static int hw_ibrs_active;
902 int hw_ibrs_ibpb_active;
903 int hw_ibrs_disable = 1;
904 
905 SYSCTL_INT(_hw, OID_AUTO, ibrs_active, CTLFLAG_RD, &hw_ibrs_active, 0,
906     "Indirect Branch Restricted Speculation active");
907 
908 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, ibrs,
909     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
910     "Indirect Branch Restricted Speculation active");
911 
912 SYSCTL_INT(_machdep_mitigations_ibrs, OID_AUTO, active, CTLFLAG_RD,
913     &hw_ibrs_active, 0, "Indirect Branch Restricted Speculation active");
914 
915 void
916 hw_ibrs_recalculate(bool for_all_cpus)
917 {
918 	if ((cpu_ia32_arch_caps & IA32_ARCH_CAP_IBRS_ALL) != 0) {
919 		x86_msr_op(MSR_IA32_SPEC_CTRL, (for_all_cpus ?
920 		    MSR_OP_RENDEZVOUS_ALL : MSR_OP_LOCAL) |
921 		    (hw_ibrs_disable != 0 ? MSR_OP_ANDNOT : MSR_OP_OR),
922 		    IA32_SPEC_CTRL_IBRS, NULL);
923 		hw_ibrs_active = hw_ibrs_disable == 0;
924 		hw_ibrs_ibpb_active = 0;
925 	} else {
926 		hw_ibrs_active = hw_ibrs_ibpb_active = (cpu_stdext_feature3 &
927 		    CPUID_STDEXT3_IBPB) != 0 && !hw_ibrs_disable;
928 	}
929 }
930 
931 static int
932 hw_ibrs_disable_handler(SYSCTL_HANDLER_ARGS)
933 {
934 	int error, val;
935 
936 	val = hw_ibrs_disable;
937 	error = sysctl_handle_int(oidp, &val, 0, req);
938 	if (error != 0 || req->newptr == NULL)
939 		return (error);
940 	hw_ibrs_disable = val != 0;
941 	hw_ibrs_recalculate(true);
942 	return (0);
943 }
944 SYSCTL_PROC(_hw, OID_AUTO, ibrs_disable, CTLTYPE_INT | CTLFLAG_RWTUN |
945     CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0, hw_ibrs_disable_handler, "I",
946     "Disable Indirect Branch Restricted Speculation");
947 
948 SYSCTL_PROC(_machdep_mitigations_ibrs, OID_AUTO, disable, CTLTYPE_INT |
949     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
950     hw_ibrs_disable_handler, "I",
951     "Disable Indirect Branch Restricted Speculation");
952 
953 int hw_ssb_active;
954 int hw_ssb_disable;
955 
956 SYSCTL_INT(_hw, OID_AUTO, spec_store_bypass_disable_active, CTLFLAG_RD,
957     &hw_ssb_active, 0,
958     "Speculative Store Bypass Disable active");
959 
960 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, ssb,
961     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
962     "Speculative Store Bypass Disable active");
963 
964 SYSCTL_INT(_machdep_mitigations_ssb, OID_AUTO, active, CTLFLAG_RD,
965     &hw_ssb_active, 0, "Speculative Store Bypass Disable active");
966 
967 static void
968 hw_ssb_set(bool enable, bool for_all_cpus)
969 {
970 
971 	if ((cpu_stdext_feature3 & CPUID_STDEXT3_SSBD) == 0) {
972 		hw_ssb_active = 0;
973 		return;
974 	}
975 	hw_ssb_active = enable;
976 	x86_msr_op(MSR_IA32_SPEC_CTRL,
977 	    (enable ? MSR_OP_OR : MSR_OP_ANDNOT) |
978 	    (for_all_cpus ? MSR_OP_SCHED_ALL : MSR_OP_LOCAL),
979 	    IA32_SPEC_CTRL_SSBD, NULL);
980 }
981 
982 void
983 hw_ssb_recalculate(bool all_cpus)
984 {
985 
986 	switch (hw_ssb_disable) {
987 	default:
988 		hw_ssb_disable = 0;
989 		/* FALLTHROUGH */
990 	case 0: /* off */
991 		hw_ssb_set(false, all_cpus);
992 		break;
993 	case 1: /* on */
994 		hw_ssb_set(true, all_cpus);
995 		break;
996 	case 2: /* auto */
997 		hw_ssb_set((cpu_ia32_arch_caps & IA32_ARCH_CAP_SSB_NO) != 0 ?
998 		    false : true, all_cpus);
999 		break;
1000 	}
1001 }
1002 
1003 static int
1004 hw_ssb_disable_handler(SYSCTL_HANDLER_ARGS)
1005 {
1006 	int error, val;
1007 
1008 	val = hw_ssb_disable;
1009 	error = sysctl_handle_int(oidp, &val, 0, req);
1010 	if (error != 0 || req->newptr == NULL)
1011 		return (error);
1012 	hw_ssb_disable = val;
1013 	hw_ssb_recalculate(true);
1014 	return (0);
1015 }
1016 SYSCTL_PROC(_hw, OID_AUTO, spec_store_bypass_disable, CTLTYPE_INT |
1017     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1018     hw_ssb_disable_handler, "I",
1019     "Speculative Store Bypass Disable (0 - off, 1 - on, 2 - auto)");
1020 
1021 SYSCTL_PROC(_machdep_mitigations_ssb, OID_AUTO, disable, CTLTYPE_INT |
1022     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1023     hw_ssb_disable_handler, "I",
1024     "Speculative Store Bypass Disable (0 - off, 1 - on, 2 - auto)");
1025 
1026 int hw_mds_disable;
1027 
1028 /*
1029  * Handler for Microarchitectural Data Sampling issues.  Really not a
1030  * pointer to C function: on amd64 the code must not change any CPU
1031  * architectural state except possibly %rflags. Also, it is always
1032  * called with interrupts disabled.
1033  */
1034 void mds_handler_void(void);
1035 void mds_handler_verw(void);
1036 void mds_handler_ivb(void);
1037 void mds_handler_bdw(void);
1038 void mds_handler_skl_sse(void);
1039 void mds_handler_skl_avx(void);
1040 void mds_handler_skl_avx512(void);
1041 void mds_handler_silvermont(void);
1042 void (*mds_handler)(void) = mds_handler_void;
1043 
1044 static int
1045 sysctl_hw_mds_disable_state_handler(SYSCTL_HANDLER_ARGS)
1046 {
1047 	const char *state;
1048 
1049 	if (mds_handler == mds_handler_void)
1050 		state = "inactive";
1051 	else if (mds_handler == mds_handler_verw)
1052 		state = "VERW";
1053 	else if (mds_handler == mds_handler_ivb)
1054 		state = "software IvyBridge";
1055 	else if (mds_handler == mds_handler_bdw)
1056 		state = "software Broadwell";
1057 	else if (mds_handler == mds_handler_skl_sse)
1058 		state = "software Skylake SSE";
1059 	else if (mds_handler == mds_handler_skl_avx)
1060 		state = "software Skylake AVX";
1061 	else if (mds_handler == mds_handler_skl_avx512)
1062 		state = "software Skylake AVX512";
1063 	else if (mds_handler == mds_handler_silvermont)
1064 		state = "software Silvermont";
1065 	else
1066 		state = "unknown";
1067 	return (SYSCTL_OUT(req, state, strlen(state)));
1068 }
1069 
1070 SYSCTL_PROC(_hw, OID_AUTO, mds_disable_state,
1071     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1072     sysctl_hw_mds_disable_state_handler, "A",
1073     "Microarchitectural Data Sampling Mitigation state");
1074 
1075 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, mds,
1076     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1077     "Microarchitectural Data Sampling Mitigation state");
1078 
1079 SYSCTL_PROC(_machdep_mitigations_mds, OID_AUTO, state,
1080     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1081     sysctl_hw_mds_disable_state_handler, "A",
1082     "Microarchitectural Data Sampling Mitigation state");
1083 
1084 _Static_assert(__offsetof(struct pcpu, pc_mds_tmp) % 64 == 0, "MDS AVX512");
1085 
1086 void
1087 hw_mds_recalculate(void)
1088 {
1089 	struct pcpu *pc;
1090 	vm_offset_t b64;
1091 	u_long xcr0;
1092 	int i;
1093 
1094 	/*
1095 	 * Allow user to force VERW variant even if MD_CLEAR is not
1096 	 * reported.  For instance, hypervisor might unknowingly
1097 	 * filter the cap out.
1098 	 * For the similar reasons, and for testing, allow to enable
1099 	 * mitigation even when MDS_NO cap is set.
1100 	 */
1101 	if (cpu_vendor_id != CPU_VENDOR_INTEL || hw_mds_disable == 0 ||
1102 	    ((cpu_ia32_arch_caps & IA32_ARCH_CAP_MDS_NO) != 0 &&
1103 	    hw_mds_disable == 3)) {
1104 		mds_handler = mds_handler_void;
1105 	} else if (((cpu_stdext_feature3 & CPUID_STDEXT3_MD_CLEAR) != 0 &&
1106 	    hw_mds_disable == 3) || hw_mds_disable == 1) {
1107 		mds_handler = mds_handler_verw;
1108 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1109 	    (CPUID_TO_MODEL(cpu_id) == 0x2e || CPUID_TO_MODEL(cpu_id) == 0x1e ||
1110 	    CPUID_TO_MODEL(cpu_id) == 0x1f || CPUID_TO_MODEL(cpu_id) == 0x1a ||
1111 	    CPUID_TO_MODEL(cpu_id) == 0x2f || CPUID_TO_MODEL(cpu_id) == 0x25 ||
1112 	    CPUID_TO_MODEL(cpu_id) == 0x2c || CPUID_TO_MODEL(cpu_id) == 0x2d ||
1113 	    CPUID_TO_MODEL(cpu_id) == 0x2a || CPUID_TO_MODEL(cpu_id) == 0x3e ||
1114 	    CPUID_TO_MODEL(cpu_id) == 0x3a) &&
1115 	    (hw_mds_disable == 2 || hw_mds_disable == 3)) {
1116 		/*
1117 		 * Nehalem, SandyBridge, IvyBridge
1118 		 */
1119 		CPU_FOREACH(i) {
1120 			pc = pcpu_find(i);
1121 			if (pc->pc_mds_buf == NULL) {
1122 				pc->pc_mds_buf = malloc_domainset(672, M_TEMP,
1123 				    DOMAINSET_PREF(pc->pc_domain), M_WAITOK);
1124 				bzero(pc->pc_mds_buf, 16);
1125 			}
1126 		}
1127 		mds_handler = mds_handler_ivb;
1128 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1129 	    (CPUID_TO_MODEL(cpu_id) == 0x3f || CPUID_TO_MODEL(cpu_id) == 0x3c ||
1130 	    CPUID_TO_MODEL(cpu_id) == 0x45 || CPUID_TO_MODEL(cpu_id) == 0x46 ||
1131 	    CPUID_TO_MODEL(cpu_id) == 0x56 || CPUID_TO_MODEL(cpu_id) == 0x4f ||
1132 	    CPUID_TO_MODEL(cpu_id) == 0x47 || CPUID_TO_MODEL(cpu_id) == 0x3d) &&
1133 	    (hw_mds_disable == 2 || hw_mds_disable == 3)) {
1134 		/*
1135 		 * Haswell, Broadwell
1136 		 */
1137 		CPU_FOREACH(i) {
1138 			pc = pcpu_find(i);
1139 			if (pc->pc_mds_buf == NULL) {
1140 				pc->pc_mds_buf = malloc_domainset(1536, M_TEMP,
1141 				    DOMAINSET_PREF(pc->pc_domain), M_WAITOK);
1142 				bzero(pc->pc_mds_buf, 16);
1143 			}
1144 		}
1145 		mds_handler = mds_handler_bdw;
1146 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1147 	    ((CPUID_TO_MODEL(cpu_id) == 0x55 && (cpu_id &
1148 	    CPUID_STEPPING) <= 5) ||
1149 	    CPUID_TO_MODEL(cpu_id) == 0x4e || CPUID_TO_MODEL(cpu_id) == 0x5e ||
1150 	    (CPUID_TO_MODEL(cpu_id) == 0x8e && (cpu_id &
1151 	    CPUID_STEPPING) <= 0xb) ||
1152 	    (CPUID_TO_MODEL(cpu_id) == 0x9e && (cpu_id &
1153 	    CPUID_STEPPING) <= 0xc)) &&
1154 	    (hw_mds_disable == 2 || hw_mds_disable == 3)) {
1155 		/*
1156 		 * Skylake, KabyLake, CoffeeLake, WhiskeyLake,
1157 		 * CascadeLake
1158 		 */
1159 		CPU_FOREACH(i) {
1160 			pc = pcpu_find(i);
1161 			if (pc->pc_mds_buf == NULL) {
1162 				pc->pc_mds_buf = malloc_domainset(6 * 1024,
1163 				    M_TEMP, DOMAINSET_PREF(pc->pc_domain),
1164 				    M_WAITOK);
1165 				b64 = (vm_offset_t)malloc_domainset(64 + 63,
1166 				    M_TEMP, DOMAINSET_PREF(pc->pc_domain),
1167 				    M_WAITOK);
1168 				pc->pc_mds_buf64 = (void *)roundup2(b64, 64);
1169 				bzero(pc->pc_mds_buf64, 64);
1170 			}
1171 		}
1172 		xcr0 = rxcr(0);
1173 		if ((xcr0 & XFEATURE_ENABLED_ZMM_HI256) != 0 &&
1174 		    (cpu_stdext_feature & CPUID_STDEXT_AVX512DQ) != 0)
1175 			mds_handler = mds_handler_skl_avx512;
1176 		else if ((xcr0 & XFEATURE_ENABLED_AVX) != 0 &&
1177 		    (cpu_feature2 & CPUID2_AVX) != 0)
1178 			mds_handler = mds_handler_skl_avx;
1179 		else
1180 			mds_handler = mds_handler_skl_sse;
1181 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1182 	    ((CPUID_TO_MODEL(cpu_id) == 0x37 ||
1183 	    CPUID_TO_MODEL(cpu_id) == 0x4a ||
1184 	    CPUID_TO_MODEL(cpu_id) == 0x4c ||
1185 	    CPUID_TO_MODEL(cpu_id) == 0x4d ||
1186 	    CPUID_TO_MODEL(cpu_id) == 0x5a ||
1187 	    CPUID_TO_MODEL(cpu_id) == 0x5d ||
1188 	    CPUID_TO_MODEL(cpu_id) == 0x6e ||
1189 	    CPUID_TO_MODEL(cpu_id) == 0x65 ||
1190 	    CPUID_TO_MODEL(cpu_id) == 0x75 ||
1191 	    CPUID_TO_MODEL(cpu_id) == 0x1c ||
1192 	    CPUID_TO_MODEL(cpu_id) == 0x26 ||
1193 	    CPUID_TO_MODEL(cpu_id) == 0x27 ||
1194 	    CPUID_TO_MODEL(cpu_id) == 0x35 ||
1195 	    CPUID_TO_MODEL(cpu_id) == 0x36 ||
1196 	    CPUID_TO_MODEL(cpu_id) == 0x7a))) {
1197 		/* Silvermont, Airmont */
1198 		CPU_FOREACH(i) {
1199 			pc = pcpu_find(i);
1200 			if (pc->pc_mds_buf == NULL)
1201 				pc->pc_mds_buf = malloc(256, M_TEMP, M_WAITOK);
1202 		}
1203 		mds_handler = mds_handler_silvermont;
1204 	} else {
1205 		hw_mds_disable = 0;
1206 		mds_handler = mds_handler_void;
1207 	}
1208 }
1209 
1210 static void
1211 hw_mds_recalculate_boot(void *arg __unused)
1212 {
1213 
1214 	hw_mds_recalculate();
1215 }
1216 SYSINIT(mds_recalc, SI_SUB_SMP, SI_ORDER_ANY, hw_mds_recalculate_boot, NULL);
1217 
1218 static int
1219 sysctl_mds_disable_handler(SYSCTL_HANDLER_ARGS)
1220 {
1221 	int error, val;
1222 
1223 	val = hw_mds_disable;
1224 	error = sysctl_handle_int(oidp, &val, 0, req);
1225 	if (error != 0 || req->newptr == NULL)
1226 		return (error);
1227 	if (val < 0 || val > 3)
1228 		return (EINVAL);
1229 	hw_mds_disable = val;
1230 	hw_mds_recalculate();
1231 	return (0);
1232 }
1233 
1234 SYSCTL_PROC(_hw, OID_AUTO, mds_disable, CTLTYPE_INT |
1235     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1236     sysctl_mds_disable_handler, "I",
1237     "Microarchitectural Data Sampling Mitigation "
1238     "(0 - off, 1 - on VERW, 2 - on SW, 3 - on AUTO)");
1239 
1240 SYSCTL_PROC(_machdep_mitigations_mds, OID_AUTO, disable, CTLTYPE_INT |
1241     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1242     sysctl_mds_disable_handler, "I",
1243     "Microarchitectural Data Sampling Mitigation "
1244     "(0 - off, 1 - on VERW, 2 - on SW, 3 - on AUTO)");
1245 
1246 /*
1247  * Intel Transactional Memory Asynchronous Abort Mitigation
1248  * CVE-2019-11135
1249  */
1250 int x86_taa_enable;
1251 int x86_taa_state;
1252 enum {
1253 	TAA_NONE	= 0,	/* No mitigation enabled */
1254 	TAA_TSX_DISABLE	= 1,	/* Disable TSX via MSR */
1255 	TAA_VERW	= 2,	/* Use VERW mitigation */
1256 	TAA_AUTO	= 3,	/* Automatically select the mitigation */
1257 
1258 	/* The states below are not selectable by the operator */
1259 
1260 	TAA_TAA_UC	= 4,	/* Mitigation present in microcode */
1261 	TAA_NOT_PRESENT	= 5	/* TSX is not present */
1262 };
1263 
1264 static void
1265 taa_set(bool enable, bool all)
1266 {
1267 
1268 	x86_msr_op(MSR_IA32_TSX_CTRL,
1269 	    (enable ? MSR_OP_OR : MSR_OP_ANDNOT) |
1270 	    (all ? MSR_OP_RENDEZVOUS_ALL : MSR_OP_LOCAL),
1271 	    IA32_TSX_CTRL_RTM_DISABLE | IA32_TSX_CTRL_TSX_CPUID_CLEAR,
1272 	    NULL);
1273 }
1274 
1275 void
1276 x86_taa_recalculate(void)
1277 {
1278 	static int taa_saved_mds_disable = 0;
1279 	int taa_need = 0, taa_state = 0;
1280 	int mds_disable = 0, need_mds_recalc = 0;
1281 
1282 	/* Check CPUID.07h.EBX.HLE and RTM for the presence of TSX */
1283 	if ((cpu_stdext_feature & CPUID_STDEXT_HLE) == 0 ||
1284 	    (cpu_stdext_feature & CPUID_STDEXT_RTM) == 0) {
1285 		/* TSX is not present */
1286 		x86_taa_state = TAA_NOT_PRESENT;
1287 		return;
1288 	}
1289 
1290 	/* Check to see what mitigation options the CPU gives us */
1291 	if (cpu_ia32_arch_caps & IA32_ARCH_CAP_TAA_NO) {
1292 		/* CPU is not suseptible to TAA */
1293 		taa_need = TAA_TAA_UC;
1294 	} else if (cpu_ia32_arch_caps & IA32_ARCH_CAP_TSX_CTRL) {
1295 		/*
1296 		 * CPU can turn off TSX.  This is the next best option
1297 		 * if TAA_NO hardware mitigation isn't present
1298 		 */
1299 		taa_need = TAA_TSX_DISABLE;
1300 	} else {
1301 		/* No TSX/TAA specific remedies are available. */
1302 		if (x86_taa_enable == TAA_TSX_DISABLE) {
1303 			if (bootverbose)
1304 				printf("TSX control not available\n");
1305 			return;
1306 		} else
1307 			taa_need = TAA_VERW;
1308 	}
1309 
1310 	/* Can we automatically take action, or are we being forced? */
1311 	if (x86_taa_enable == TAA_AUTO)
1312 		taa_state = taa_need;
1313 	else
1314 		taa_state = x86_taa_enable;
1315 
1316 	/* No state change, nothing to do */
1317 	if (taa_state == x86_taa_state) {
1318 		if (bootverbose)
1319 			printf("No TSX change made\n");
1320 		return;
1321 	}
1322 
1323 	/* Does the MSR need to be turned on or off? */
1324 	if (taa_state == TAA_TSX_DISABLE)
1325 		taa_set(true, true);
1326 	else if (x86_taa_state == TAA_TSX_DISABLE)
1327 		taa_set(false, true);
1328 
1329 	/* Does MDS need to be set to turn on VERW? */
1330 	if (taa_state == TAA_VERW) {
1331 		taa_saved_mds_disable = hw_mds_disable;
1332 		mds_disable = hw_mds_disable = 1;
1333 		need_mds_recalc = 1;
1334 	} else if (x86_taa_state == TAA_VERW) {
1335 		mds_disable = hw_mds_disable = taa_saved_mds_disable;
1336 		need_mds_recalc = 1;
1337 	}
1338 	if (need_mds_recalc) {
1339 		hw_mds_recalculate();
1340 		if (mds_disable != hw_mds_disable) {
1341 			if (bootverbose)
1342 				printf("Cannot change MDS state for TAA\n");
1343 			/* Don't update our state */
1344 			return;
1345 		}
1346 	}
1347 
1348 	x86_taa_state = taa_state;
1349 	return;
1350 }
1351 
1352 static void
1353 taa_recalculate_boot(void * arg __unused)
1354 {
1355 
1356 	x86_taa_recalculate();
1357 }
1358 SYSINIT(taa_recalc, SI_SUB_SMP, SI_ORDER_ANY, taa_recalculate_boot, NULL);
1359 
1360 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, taa,
1361     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1362     "TSX Asynchronous Abort Mitigation");
1363 
1364 static int
1365 sysctl_taa_handler(SYSCTL_HANDLER_ARGS)
1366 {
1367 	int error, val;
1368 
1369 	val = x86_taa_enable;
1370 	error = sysctl_handle_int(oidp, &val, 0, req);
1371 	if (error != 0 || req->newptr == NULL)
1372 		return (error);
1373 	if (val < TAA_NONE || val > TAA_AUTO)
1374 		return (EINVAL);
1375 	x86_taa_enable = val;
1376 	x86_taa_recalculate();
1377 	return (0);
1378 }
1379 
1380 SYSCTL_PROC(_machdep_mitigations_taa, OID_AUTO, enable, CTLTYPE_INT |
1381     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1382     sysctl_taa_handler, "I",
1383     "TAA Mitigation enablement control "
1384     "(0 - off, 1 - disable TSX, 2 - VERW, 3 - on AUTO)");
1385 
1386 static int
1387 sysctl_taa_state_handler(SYSCTL_HANDLER_ARGS)
1388 {
1389 	const char *state;
1390 
1391 	switch (x86_taa_state) {
1392 	case TAA_NONE:
1393 		state = "inactive";
1394 		break;
1395 	case TAA_TSX_DISABLE:
1396 		state = "TSX disabled";
1397 		break;
1398 	case TAA_VERW:
1399 		state = "VERW";
1400 		break;
1401 	case TAA_TAA_UC:
1402 		state = "Mitigated in microcode";
1403 		break;
1404 	case TAA_NOT_PRESENT:
1405 		state = "TSX not present";
1406 		break;
1407 	default:
1408 		state = "unknown";
1409 	}
1410 
1411 	return (SYSCTL_OUT(req, state, strlen(state)));
1412 }
1413 
1414 SYSCTL_PROC(_machdep_mitigations_taa, OID_AUTO, state,
1415     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1416     sysctl_taa_state_handler, "A",
1417     "TAA Mitigation state");
1418 
1419 int __read_frequently cpu_flush_rsb_ctxsw;
1420 SYSCTL_INT(_machdep_mitigations, OID_AUTO, flush_rsb_ctxsw,
1421     CTLFLAG_RW | CTLFLAG_NOFETCH, &cpu_flush_rsb_ctxsw, 0,
1422     "Flush Return Stack Buffer on context switch");
1423 
1424 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, rngds,
1425     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1426     "MCU Optimization, disable RDSEED mitigation");
1427 
1428 int x86_rngds_mitg_enable = 1;
1429 void
1430 x86_rngds_mitg_recalculate(bool all_cpus)
1431 {
1432 	if ((cpu_stdext_feature3 & CPUID_STDEXT3_MCUOPT) == 0)
1433 		return;
1434 	x86_msr_op(MSR_IA32_MCU_OPT_CTRL,
1435 	    (x86_rngds_mitg_enable ? MSR_OP_OR : MSR_OP_ANDNOT) |
1436 	    (all_cpus ? MSR_OP_RENDEZVOUS_ALL : MSR_OP_LOCAL),
1437 	    IA32_RNGDS_MITG_DIS, NULL);
1438 }
1439 
1440 static int
1441 sysctl_rngds_mitg_enable_handler(SYSCTL_HANDLER_ARGS)
1442 {
1443 	int error, val;
1444 
1445 	val = x86_rngds_mitg_enable;
1446 	error = sysctl_handle_int(oidp, &val, 0, req);
1447 	if (error != 0 || req->newptr == NULL)
1448 		return (error);
1449 	x86_rngds_mitg_enable = val;
1450 	x86_rngds_mitg_recalculate(true);
1451 	return (0);
1452 }
1453 SYSCTL_PROC(_machdep_mitigations_rngds, OID_AUTO, enable, CTLTYPE_INT |
1454     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1455     sysctl_rngds_mitg_enable_handler, "I",
1456     "MCU Optimization, disabling RDSEED mitigation control "
1457     "(0 - mitigation disabled (RDSEED optimized), 1 - mitigation enabled)");
1458 
1459 static int
1460 sysctl_rngds_state_handler(SYSCTL_HANDLER_ARGS)
1461 {
1462 	const char *state;
1463 
1464 	if ((cpu_stdext_feature3 & CPUID_STDEXT3_MCUOPT) == 0) {
1465 		state = "Not applicable";
1466 	} else if (x86_rngds_mitg_enable == 0) {
1467 		state = "RDSEED not serialized";
1468 	} else {
1469 		state = "Mitigated";
1470 	}
1471 	return (SYSCTL_OUT(req, state, strlen(state)));
1472 }
1473 SYSCTL_PROC(_machdep_mitigations_rngds, OID_AUTO, state,
1474     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1475     sysctl_rngds_state_handler, "A",
1476     "MCU Optimization state");
1477 
1478 
1479 /*
1480  * Zenbleed.
1481  *
1482  * No corresponding errata is publicly listed.  AMD has issued a security
1483  * bulletin (AMD-SB-7008), entitled "Cross-Process Information Leak".  This
1484  * document lists (as of August 2023) platform firmware's availability target
1485  * dates, with most being November/December 2023.  It will then be up to
1486  * motherboard manufacturers to produce corresponding BIOS updates, which will
1487  * happen with an inevitable lag.  Additionally, for a variety of reasons,
1488  * operators might not be able to apply them everywhere due.  On the side of
1489  * standalone CPU microcodes, no plans for availability have been published so
1490  * far.  However, a developer appearing to be an AMD employee has hardcoded in
1491  * Linux revision numbers of future microcodes that are presumed to fix the
1492  * vulnerability.
1493  *
1494  * Given the stability issues encountered with early microcode releases for Rome
1495  * (the only microcode publicly released so far) and the absence of official
1496  * communication on standalone CPU microcodes, we have opted instead for
1497  * matching by default all AMD Zen2 processors which, according to the
1498  * vulnerability's discoverer, are all affected (see
1499  * https://lock.cmpxchg8b.com/zenbleed.html).  This policy, also adopted by
1500  * OpenBSD, may be overriden using the tunable/sysctl
1501  * 'machdep.mitigations.zenbleed.enable'.  We might revise it later depending on
1502  * official statements, microcode updates' public availability and community
1503  * assessment that they actually fix the vulnerability without any instability
1504  * side effects.
1505  */
1506 
1507 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, zenbleed,
1508     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1509     "Zenbleed OS-triggered prevention (via chicken bit)");
1510 
1511 /* 2 is auto, see below. */
1512 int zenbleed_enable = 2;
1513 
1514 void
1515 zenbleed_sanitize_enable(void)
1516 {
1517 	/* Default to auto (2). */
1518 	if (zenbleed_enable < 0 || zenbleed_enable > 2)
1519 		zenbleed_enable = 2;
1520 }
1521 
1522 static bool
1523 zenbleed_chicken_bit_applicable(void)
1524 {
1525 	/* Concerns only bare-metal AMD Zen2 processors. */
1526 	return (cpu_vendor_id == CPU_VENDOR_AMD &&
1527 	    CPUID_TO_FAMILY(cpu_id) == 0x17 &&
1528 	    CPUID_TO_MODEL(cpu_id) >= 0x30 &&
1529 	    vm_guest == VM_GUEST_NO);
1530 }
1531 
1532 static bool
1533 zenbleed_chicken_bit_should_enable(void)
1534 {
1535 	/*
1536 	 * Obey tunable/sysctl.
1537 	 *
1538 	 * As explained above, currently, the automatic setting (2) and the "on"
1539 	 * one (1) have the same effect.  In the future, we might additionally
1540 	 * check for specific microcode revisions as part of the automatic
1541 	 * determination.
1542 	 */
1543 	return (zenbleed_enable != 0);
1544 }
1545 
1546 void
1547 zenbleed_check_and_apply(bool all_cpus)
1548 {
1549 	bool set;
1550 
1551 	if (!zenbleed_chicken_bit_applicable())
1552 		return;
1553 
1554 	set = zenbleed_chicken_bit_should_enable();
1555 
1556 	x86_msr_op(MSR_DE_CFG,
1557 	    (set ? MSR_OP_OR : MSR_OP_ANDNOT) |
1558 	    (all_cpus ? MSR_OP_RENDEZVOUS_ALL : MSR_OP_LOCAL),
1559 	    DE_CFG_ZEN2_FP_BACKUP_FIX_BIT, NULL);
1560 }
1561 
1562 static int
1563 sysctl_zenbleed_enable_handler(SYSCTL_HANDLER_ARGS)
1564 {
1565 	int error, val;
1566 
1567 	val = zenbleed_enable;
1568 	error = sysctl_handle_int(oidp, &val, 0, req);
1569 	if (error != 0 || req->newptr == NULL)
1570 		return (error);
1571 	zenbleed_enable = val;
1572 	zenbleed_sanitize_enable();
1573 	zenbleed_check_and_apply(true);
1574 	return (0);
1575 }
1576 SYSCTL_PROC(_machdep_mitigations_zenbleed, OID_AUTO, enable, CTLTYPE_INT |
1577     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1578     sysctl_zenbleed_enable_handler, "I",
1579     "Enable Zenbleed OS-triggered mitigation (chicken bit) "
1580     "(0: Force disable, 1: Force enable, 2: Automatic determination)");
1581 
1582 static int
1583 sysctl_zenbleed_state_handler(SYSCTL_HANDLER_ARGS)
1584 {
1585 	const char *state;
1586 
1587 	if (!zenbleed_chicken_bit_applicable())
1588 		state = "Not applicable";
1589 	else if (zenbleed_chicken_bit_should_enable())
1590 		state = "Mitigation enabled";
1591 	else
1592 		state = "Mitigation disabled";
1593 	return (SYSCTL_OUT(req, state, strlen(state)));
1594 }
1595 SYSCTL_PROC(_machdep_mitigations_zenbleed, OID_AUTO, state,
1596     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1597     sysctl_zenbleed_state_handler, "A",
1598     "Zenbleed OS-triggered mitigation (chicken bit) state");
1599 
1600 
1601 /*
1602  * Enable and restore kernel text write permissions.
1603  * Callers must ensure that disable_wp()/restore_wp() are executed
1604  * without rescheduling on the same core.
1605  */
1606 bool
1607 disable_wp(void)
1608 {
1609 	u_int cr0;
1610 
1611 	cr0 = rcr0();
1612 	if ((cr0 & CR0_WP) == 0)
1613 		return (false);
1614 	load_cr0(cr0 & ~CR0_WP);
1615 	return (true);
1616 }
1617 
1618 void
1619 restore_wp(bool old_wp)
1620 {
1621 
1622 	if (old_wp)
1623 		load_cr0(rcr0() | CR0_WP);
1624 }
1625 
1626 bool
1627 acpi_get_fadt_bootflags(uint16_t *flagsp)
1628 {
1629 #ifdef DEV_ACPI
1630 	ACPI_TABLE_FADT *fadt;
1631 	vm_paddr_t physaddr;
1632 
1633 	physaddr = acpi_find_table(ACPI_SIG_FADT);
1634 	if (physaddr == 0)
1635 		return (false);
1636 	fadt = acpi_map_table(physaddr, ACPI_SIG_FADT);
1637 	if (fadt == NULL)
1638 		return (false);
1639 	*flagsp = fadt->BootFlags;
1640 	acpi_unmap_table(fadt);
1641 	return (true);
1642 #else
1643 	return (false);
1644 #endif
1645 }
1646 
1647 DEFINE_IFUNC(, uint64_t, rdtsc_ordered, (void))
1648 {
1649 	bool cpu_is_amd = cpu_vendor_id == CPU_VENDOR_AMD ||
1650 	    cpu_vendor_id == CPU_VENDOR_HYGON;
1651 
1652 	if ((amd_feature & AMDID_RDTSCP) != 0)
1653 		return (rdtscp);
1654 	else if ((cpu_feature & CPUID_SSE2) != 0)
1655 		return (cpu_is_amd ? rdtsc_ordered_mfence :
1656 		    rdtsc_ordered_lfence);
1657 	else
1658 		return (rdtsc);
1659 }
1660