xref: /freebsd-13.1/sys/x86/x86/cpu_machdep.c (revision dc411487)
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/cpufunc.h>
80 #include <machine/cputypes.h>
81 #include <machine/specialreg.h>
82 #include <machine/md_var.h>
83 #include <machine/mp_watchdog.h>
84 #include <machine/tss.h>
85 #ifdef SMP
86 #include <machine/smp.h>
87 #endif
88 #ifdef CPU_ELAN
89 #include <machine/elan_mmcr.h>
90 #endif
91 #include <x86/acpica_machdep.h>
92 #include <x86/ifunc.h>
93 
94 #include <vm/vm.h>
95 #include <vm/vm_extern.h>
96 #include <vm/vm_kern.h>
97 #include <vm/vm_page.h>
98 #include <vm/vm_map.h>
99 #include <vm/vm_object.h>
100 #include <vm/vm_pager.h>
101 #include <vm/vm_param.h>
102 
103 #include <isa/isareg.h>
104 
105 #include <contrib/dev/acpica/include/acpi.h>
106 
107 #define	STATE_RUNNING	0x0
108 #define	STATE_MWAIT	0x1
109 #define	STATE_SLEEPING	0x2
110 
111 #ifdef SMP
112 static u_int	cpu_reset_proxyid;
113 static volatile u_int	cpu_reset_proxy_active;
114 #endif
115 
116 struct msr_op_arg {
117 	u_int msr;
118 	int op;
119 	uint64_t arg1;
120 	uint64_t *res;
121 };
122 
123 static void
x86_msr_op_one(void * argp)124 x86_msr_op_one(void *argp)
125 {
126 	struct msr_op_arg *a;
127 	uint64_t v;
128 
129 	a = argp;
130 	switch (a->op) {
131 	case MSR_OP_ANDNOT:
132 		v = rdmsr(a->msr);
133 		v &= ~a->arg1;
134 		wrmsr(a->msr, v);
135 		break;
136 	case MSR_OP_OR:
137 		v = rdmsr(a->msr);
138 		v |= a->arg1;
139 		wrmsr(a->msr, v);
140 		break;
141 	case MSR_OP_WRITE:
142 		wrmsr(a->msr, a->arg1);
143 		break;
144 	case MSR_OP_READ:
145 		v = rdmsr(a->msr);
146 		*a->res = v;
147 		break;
148 	}
149 }
150 
151 #define	MSR_OP_EXMODE_MASK	0xf0000000
152 #define	MSR_OP_OP_MASK		0x000000ff
153 #define	MSR_OP_GET_CPUID(x)	(((x) & ~MSR_OP_EXMODE_MASK) >> 8)
154 
155 void
x86_msr_op(u_int msr,u_int op,uint64_t arg1,uint64_t * res)156 x86_msr_op(u_int msr, u_int op, uint64_t arg1, uint64_t *res)
157 {
158 	struct thread *td;
159 	struct msr_op_arg a;
160 	cpuset_t set;
161 	u_int exmode;
162 	int bound_cpu, cpu, i, is_bound;
163 
164 	a.op = op & MSR_OP_OP_MASK;
165 	MPASS(a.op == MSR_OP_ANDNOT || a.op == MSR_OP_OR ||
166 	    a.op == MSR_OP_WRITE || a.op == MSR_OP_READ);
167 	exmode = op & MSR_OP_EXMODE_MASK;
168 	MPASS(exmode == MSR_OP_LOCAL || exmode == MSR_OP_SCHED_ALL ||
169 	    exmode == MSR_OP_SCHED_ONE || exmode == MSR_OP_RENDEZVOUS_ALL ||
170 	    exmode == MSR_OP_RENDEZVOUS_ONE);
171 	a.msr = msr;
172 	a.arg1 = arg1;
173 	a.res = res;
174 	switch (exmode) {
175 	case MSR_OP_LOCAL:
176 		x86_msr_op_one(&a);
177 		break;
178 	case MSR_OP_SCHED_ALL:
179 		td = curthread;
180 		thread_lock(td);
181 		is_bound = sched_is_bound(td);
182 		bound_cpu = td->td_oncpu;
183 		CPU_FOREACH(i) {
184 			sched_bind(td, i);
185 			x86_msr_op_one(&a);
186 		}
187 		if (is_bound)
188 			sched_bind(td, bound_cpu);
189 		else
190 			sched_unbind(td);
191 		thread_unlock(td);
192 		break;
193 	case MSR_OP_SCHED_ONE:
194 		td = curthread;
195 		cpu = MSR_OP_GET_CPUID(op);
196 		thread_lock(td);
197 		is_bound = sched_is_bound(td);
198 		bound_cpu = td->td_oncpu;
199 		if (!is_bound || bound_cpu != cpu)
200 			sched_bind(td, cpu);
201 		x86_msr_op_one(&a);
202 		if (is_bound) {
203 			if (bound_cpu != cpu)
204 				sched_bind(td, bound_cpu);
205 		} else {
206 			sched_unbind(td);
207 		}
208 		thread_unlock(td);
209 		break;
210 	case MSR_OP_RENDEZVOUS_ALL:
211 		smp_rendezvous(smp_no_rendezvous_barrier, x86_msr_op_one,
212 		    smp_no_rendezvous_barrier, &a);
213 		break;
214 	case MSR_OP_RENDEZVOUS_ONE:
215 		cpu = MSR_OP_GET_CPUID(op);
216 		CPU_SETOF(cpu, &set);
217 		smp_rendezvous_cpus(set, smp_no_rendezvous_barrier,
218 		    x86_msr_op_one, smp_no_rendezvous_barrier, &a);
219 		break;
220 	}
221 }
222 
223 /*
224  * Automatically initialized per CPU errata in cpu_idle_tun below.
225  */
226 bool mwait_cpustop_broken = false;
227 SYSCTL_BOOL(_machdep, OID_AUTO, mwait_cpustop_broken, CTLFLAG_RDTUN,
228     &mwait_cpustop_broken, 0,
229     "Can not reliably wake MONITOR/MWAIT cpus without interrupts");
230 
231 /*
232  * Flush the D-cache for non-DMA I/O so that the I-cache can
233  * be made coherent later.
234  */
235 void
cpu_flush_dcache(void * ptr,size_t len)236 cpu_flush_dcache(void *ptr, size_t len)
237 {
238 	/* Not applicable */
239 }
240 
241 void
acpi_cpu_c1(void)242 acpi_cpu_c1(void)
243 {
244 
245 	__asm __volatile("sti; hlt");
246 }
247 
248 /*
249  * Use mwait to pause execution while waiting for an interrupt or
250  * another thread to signal that there is more work.
251  *
252  * NOTE: Interrupts will cause a wakeup; however, this function does
253  * not enable interrupt handling. The caller is responsible to enable
254  * interrupts.
255  */
256 void
acpi_cpu_idle_mwait(uint32_t mwait_hint)257 acpi_cpu_idle_mwait(uint32_t mwait_hint)
258 {
259 	int *state;
260 	uint64_t v;
261 
262 	/*
263 	 * A comment in Linux patch claims that 'CPUs run faster with
264 	 * speculation protection disabled. All CPU threads in a core
265 	 * must disable speculation protection for it to be
266 	 * disabled. Disable it while we are idle so the other
267 	 * hyperthread can run fast.'
268 	 *
269 	 * XXXKIB.  Software coordination mode should be supported,
270 	 * but all Intel CPUs provide hardware coordination.
271 	 */
272 
273 	state = &PCPU_PTR(monitorbuf)->idle_state;
274 	KASSERT(atomic_load_int(state) == STATE_SLEEPING,
275 	    ("cpu_mwait_cx: wrong monitorbuf state"));
276 	atomic_store_int(state, STATE_MWAIT);
277 	if (PCPU_GET(ibpb_set) || hw_ssb_active) {
278 		v = rdmsr(MSR_IA32_SPEC_CTRL);
279 		wrmsr(MSR_IA32_SPEC_CTRL, v & ~(IA32_SPEC_CTRL_IBRS |
280 		    IA32_SPEC_CTRL_STIBP | IA32_SPEC_CTRL_SSBD));
281 	} else {
282 		v = 0;
283 	}
284 	cpu_monitor(state, 0, 0);
285 	if (atomic_load_int(state) == STATE_MWAIT)
286 		cpu_mwait(MWAIT_INTRBREAK, mwait_hint);
287 
288 	/*
289 	 * SSB cannot be disabled while we sleep, or rather, if it was
290 	 * disabled, the sysctl thread will bind to our cpu to tweak
291 	 * MSR.
292 	 */
293 	if (v != 0)
294 		wrmsr(MSR_IA32_SPEC_CTRL, v);
295 
296 	/*
297 	 * We should exit on any event that interrupts mwait, because
298 	 * that event might be a wanted interrupt.
299 	 */
300 	atomic_store_int(state, STATE_RUNNING);
301 }
302 
303 /* Get current clock frequency for the given cpu id. */
304 int
cpu_est_clockrate(int cpu_id,uint64_t * rate)305 cpu_est_clockrate(int cpu_id, uint64_t *rate)
306 {
307 	uint64_t tsc1, tsc2;
308 	uint64_t acnt, mcnt, perf;
309 	register_t reg;
310 
311 	if (pcpu_find(cpu_id) == NULL || rate == NULL)
312 		return (EINVAL);
313 #ifdef __i386__
314 	if ((cpu_feature & CPUID_TSC) == 0)
315 		return (EOPNOTSUPP);
316 #endif
317 
318 	/*
319 	 * If TSC is P-state invariant and APERF/MPERF MSRs do not exist,
320 	 * DELAY(9) based logic fails.
321 	 */
322 	if (tsc_is_invariant && !tsc_perf_stat)
323 		return (EOPNOTSUPP);
324 
325 #ifdef SMP
326 	if (smp_cpus > 1) {
327 		/* Schedule ourselves on the indicated cpu. */
328 		thread_lock(curthread);
329 		sched_bind(curthread, cpu_id);
330 		thread_unlock(curthread);
331 	}
332 #endif
333 
334 	/* Calibrate by measuring a short delay. */
335 	reg = intr_disable();
336 	if (tsc_is_invariant) {
337 		wrmsr(MSR_MPERF, 0);
338 		wrmsr(MSR_APERF, 0);
339 		tsc1 = rdtsc();
340 		DELAY(1000);
341 		mcnt = rdmsr(MSR_MPERF);
342 		acnt = rdmsr(MSR_APERF);
343 		tsc2 = rdtsc();
344 		intr_restore(reg);
345 		perf = 1000 * acnt / mcnt;
346 		*rate = (tsc2 - tsc1) * perf;
347 	} else {
348 		tsc1 = rdtsc();
349 		DELAY(1000);
350 		tsc2 = rdtsc();
351 		intr_restore(reg);
352 		*rate = (tsc2 - tsc1) * 1000;
353 	}
354 
355 #ifdef SMP
356 	if (smp_cpus > 1) {
357 		thread_lock(curthread);
358 		sched_unbind(curthread);
359 		thread_unlock(curthread);
360 	}
361 #endif
362 
363 	return (0);
364 }
365 
366 /*
367  * Shutdown the CPU as much as possible
368  */
369 void
cpu_halt(void)370 cpu_halt(void)
371 {
372 	for (;;)
373 		halt();
374 }
375 
376 static void
cpu_reset_real(void)377 cpu_reset_real(void)
378 {
379 	struct region_descriptor null_idt;
380 	int b;
381 
382 	disable_intr();
383 #ifdef CPU_ELAN
384 	if (elan_mmcr != NULL)
385 		elan_mmcr->RESCFG = 1;
386 #endif
387 #ifdef __i386__
388 	if (cpu == CPU_GEODE1100) {
389 		/* Attempt Geode's own reset */
390 		outl(0xcf8, 0x80009044ul);
391 		outl(0xcfc, 0xf);
392 	}
393 #endif
394 #if !defined(BROKEN_KEYBOARD_RESET)
395 	/*
396 	 * Attempt to do a CPU reset via the keyboard controller,
397 	 * do not turn off GateA20, as any machine that fails
398 	 * to do the reset here would then end up in no man's land.
399 	 */
400 	outb(IO_KBD + 4, 0xFE);
401 	DELAY(500000);	/* wait 0.5 sec to see if that did it */
402 #endif
403 
404 	/*
405 	 * Attempt to force a reset via the Reset Control register at
406 	 * I/O port 0xcf9.  Bit 2 forces a system reset when it
407 	 * transitions from 0 to 1.  Bit 1 selects the type of reset
408 	 * to attempt: 0 selects a "soft" reset, and 1 selects a
409 	 * "hard" reset.  We try a "hard" reset.  The first write sets
410 	 * bit 1 to select a "hard" reset and clears bit 2.  The
411 	 * second write forces a 0 -> 1 transition in bit 2 to trigger
412 	 * a reset.
413 	 */
414 	outb(0xcf9, 0x2);
415 	outb(0xcf9, 0x6);
416 	DELAY(500000);  /* wait 0.5 sec to see if that did it */
417 
418 	/*
419 	 * Attempt to force a reset via the Fast A20 and Init register
420 	 * at I/O port 0x92.  Bit 1 serves as an alternate A20 gate.
421 	 * Bit 0 asserts INIT# when set to 1.  We are careful to only
422 	 * preserve bit 1 while setting bit 0.  We also must clear bit
423 	 * 0 before setting it if it isn't already clear.
424 	 */
425 	b = inb(0x92);
426 	if (b != 0xff) {
427 		if ((b & 0x1) != 0)
428 			outb(0x92, b & 0xfe);
429 		outb(0x92, b | 0x1);
430 		DELAY(500000);  /* wait 0.5 sec to see if that did it */
431 	}
432 
433 	printf("No known reset method worked, attempting CPU shutdown\n");
434 	DELAY(1000000); /* wait 1 sec for printf to complete */
435 
436 	/* Wipe the IDT. */
437 	null_idt.rd_limit = 0;
438 	null_idt.rd_base = 0;
439 	lidt(&null_idt);
440 
441 	/* "good night, sweet prince .... <THUNK!>" */
442 	breakpoint();
443 
444 	/* NOTREACHED */
445 	while(1);
446 }
447 
448 #ifdef SMP
449 static void
cpu_reset_proxy(void)450 cpu_reset_proxy(void)
451 {
452 
453 	cpu_reset_proxy_active = 1;
454 	while (cpu_reset_proxy_active == 1)
455 		ia32_pause(); /* Wait for other cpu to see that we've started */
456 
457 	printf("cpu_reset_proxy: Stopped CPU %d\n", cpu_reset_proxyid);
458 	DELAY(1000000);
459 	cpu_reset_real();
460 }
461 #endif
462 
463 void
cpu_reset(void)464 cpu_reset(void)
465 {
466 #ifdef SMP
467 	struct monitorbuf *mb;
468 	cpuset_t map;
469 	u_int cnt;
470 
471 	if (smp_started) {
472 		map = all_cpus;
473 		CPU_CLR(PCPU_GET(cpuid), &map);
474 		CPU_ANDNOT(&map, &map, &stopped_cpus);
475 		if (!CPU_EMPTY(&map)) {
476 			printf("cpu_reset: Stopping other CPUs\n");
477 			stop_cpus(map);
478 		}
479 
480 		if (PCPU_GET(cpuid) != 0) {
481 			cpu_reset_proxyid = PCPU_GET(cpuid);
482 			cpustop_restartfunc = cpu_reset_proxy;
483 			cpu_reset_proxy_active = 0;
484 			printf("cpu_reset: Restarting BSP\n");
485 
486 			/* Restart CPU #0. */
487 			CPU_SETOF(0, &started_cpus);
488 			mb = &pcpu_find(0)->pc_monitorbuf;
489 			atomic_store_int(&mb->stop_state,
490 			    MONITOR_STOPSTATE_RUNNING);
491 
492 			cnt = 0;
493 			while (cpu_reset_proxy_active == 0 && cnt < 10000000) {
494 				ia32_pause();
495 				cnt++;	/* Wait for BSP to announce restart */
496 			}
497 			if (cpu_reset_proxy_active == 0) {
498 				printf("cpu_reset: Failed to restart BSP\n");
499 			} else {
500 				cpu_reset_proxy_active = 2;
501 				while (1)
502 					ia32_pause();
503 				/* NOTREACHED */
504 			}
505 		}
506 
507 		DELAY(1000000);
508 	}
509 #endif
510 	cpu_reset_real();
511 	/* NOTREACHED */
512 }
513 
514 bool
cpu_mwait_usable(void)515 cpu_mwait_usable(void)
516 {
517 
518 	return ((cpu_feature2 & CPUID2_MON) != 0 && ((cpu_mon_mwait_flags &
519 	    (CPUID5_MON_MWAIT_EXT | CPUID5_MWAIT_INTRBREAK)) ==
520 	    (CPUID5_MON_MWAIT_EXT | CPUID5_MWAIT_INTRBREAK)));
521 }
522 
523 void (*cpu_idle_hook)(sbintime_t) = NULL;	/* ACPI idle hook. */
524 
525 int cpu_amdc1e_bug = 0;			/* AMD C1E APIC workaround required. */
526 
527 static int	idle_mwait = 1;		/* Use MONITOR/MWAIT for short idle. */
528 SYSCTL_INT(_machdep, OID_AUTO, idle_mwait, CTLFLAG_RWTUN, &idle_mwait,
529     0, "Use MONITOR/MWAIT for short idle");
530 
531 static void
cpu_idle_acpi(sbintime_t sbt)532 cpu_idle_acpi(sbintime_t sbt)
533 {
534 	int *state;
535 
536 	state = &PCPU_PTR(monitorbuf)->idle_state;
537 	atomic_store_int(state, STATE_SLEEPING);
538 
539 	/* See comments in cpu_idle_hlt(). */
540 	disable_intr();
541 	if (sched_runnable())
542 		enable_intr();
543 	else if (cpu_idle_hook)
544 		cpu_idle_hook(sbt);
545 	else
546 		acpi_cpu_c1();
547 	atomic_store_int(state, STATE_RUNNING);
548 }
549 
550 static void
cpu_idle_hlt(sbintime_t sbt)551 cpu_idle_hlt(sbintime_t sbt)
552 {
553 	int *state;
554 
555 	state = &PCPU_PTR(monitorbuf)->idle_state;
556 	atomic_store_int(state, STATE_SLEEPING);
557 
558 	/*
559 	 * Since we may be in a critical section from cpu_idle(), if
560 	 * an interrupt fires during that critical section we may have
561 	 * a pending preemption.  If the CPU halts, then that thread
562 	 * may not execute until a later interrupt awakens the CPU.
563 	 * To handle this race, check for a runnable thread after
564 	 * disabling interrupts and immediately return if one is
565 	 * found.  Also, we must absolutely guarentee that hlt is
566 	 * the next instruction after sti.  This ensures that any
567 	 * interrupt that fires after the call to disable_intr() will
568 	 * immediately awaken the CPU from hlt.  Finally, please note
569 	 * that on x86 this works fine because of interrupts enabled only
570 	 * after the instruction following sti takes place, while IF is set
571 	 * to 1 immediately, allowing hlt instruction to acknowledge the
572 	 * interrupt.
573 	 */
574 	disable_intr();
575 	if (sched_runnable())
576 		enable_intr();
577 	else
578 		acpi_cpu_c1();
579 	atomic_store_int(state, STATE_RUNNING);
580 }
581 
582 static void
cpu_idle_mwait(sbintime_t sbt)583 cpu_idle_mwait(sbintime_t sbt)
584 {
585 	int *state;
586 
587 	state = &PCPU_PTR(monitorbuf)->idle_state;
588 	atomic_store_int(state, STATE_MWAIT);
589 
590 	/* See comments in cpu_idle_hlt(). */
591 	disable_intr();
592 	if (sched_runnable()) {
593 		atomic_store_int(state, STATE_RUNNING);
594 		enable_intr();
595 		return;
596 	}
597 
598 	cpu_monitor(state, 0, 0);
599 	if (atomic_load_int(state) == STATE_MWAIT)
600 		__asm __volatile("sti; mwait" : : "a" (MWAIT_C1), "c" (0));
601 	else
602 		enable_intr();
603 	atomic_store_int(state, STATE_RUNNING);
604 }
605 
606 static void
cpu_idle_spin(sbintime_t sbt)607 cpu_idle_spin(sbintime_t sbt)
608 {
609 	int *state;
610 	int i;
611 
612 	state = &PCPU_PTR(monitorbuf)->idle_state;
613 	atomic_store_int(state, STATE_RUNNING);
614 
615 	/*
616 	 * The sched_runnable() call is racy but as long as there is
617 	 * a loop missing it one time will have just a little impact if any
618 	 * (and it is much better than missing the check at all).
619 	 */
620 	for (i = 0; i < 1000; i++) {
621 		if (sched_runnable())
622 			return;
623 		cpu_spinwait();
624 	}
625 }
626 
627 void (*cpu_idle_fn)(sbintime_t) = cpu_idle_acpi;
628 
629 void
cpu_idle(int busy)630 cpu_idle(int busy)
631 {
632 	uint64_t msr;
633 	sbintime_t sbt = -1;
634 
635 	CTR2(KTR_SPARE2, "cpu_idle(%d) at %d",
636 	    busy, curcpu);
637 #ifdef MP_WATCHDOG
638 	ap_watchdog(PCPU_GET(cpuid));
639 #endif
640 
641 	/* If we are busy - try to use fast methods. */
642 	if (busy) {
643 		if ((cpu_feature2 & CPUID2_MON) && idle_mwait) {
644 			cpu_idle_mwait(busy);
645 			goto out;
646 		}
647 	}
648 
649 	/* If we have time - switch timers into idle mode. */
650 	if (!busy) {
651 		critical_enter();
652 		sbt = cpu_idleclock();
653 	}
654 
655 	/* Apply AMD APIC timer C1E workaround. */
656 	if (cpu_amdc1e_bug && cpu_disable_c3_sleep) {
657 		msr = rdmsr(MSR_AMDK8_IPM);
658 		if ((msr & (AMDK8_SMIONCMPHALT | AMDK8_C1EONCMPHALT)) != 0)
659 			wrmsr(MSR_AMDK8_IPM, msr & ~(AMDK8_SMIONCMPHALT |
660 			    AMDK8_C1EONCMPHALT));
661 	}
662 
663 	/* Call main idle method. */
664 	cpu_idle_fn(sbt);
665 
666 	/* Switch timers back into active mode. */
667 	if (!busy) {
668 		cpu_activeclock();
669 		critical_exit();
670 	}
671 out:
672 	CTR2(KTR_SPARE2, "cpu_idle(%d) at %d done",
673 	    busy, curcpu);
674 }
675 
676 static int cpu_idle_apl31_workaround;
677 SYSCTL_INT(_machdep, OID_AUTO, idle_apl31, CTLFLAG_RW,
678     &cpu_idle_apl31_workaround, 0,
679     "Apollo Lake APL31 MWAIT bug workaround");
680 
681 int
cpu_idle_wakeup(int cpu)682 cpu_idle_wakeup(int cpu)
683 {
684 	struct monitorbuf *mb;
685 	int *state;
686 
687 	mb = &pcpu_find(cpu)->pc_monitorbuf;
688 	state = &mb->idle_state;
689 	switch (atomic_load_int(state)) {
690 	case STATE_SLEEPING:
691 		return (0);
692 	case STATE_MWAIT:
693 		atomic_store_int(state, STATE_RUNNING);
694 		return (cpu_idle_apl31_workaround ? 0 : 1);
695 	case STATE_RUNNING:
696 		return (1);
697 	default:
698 		panic("bad monitor state");
699 		return (1);
700 	}
701 }
702 
703 /*
704  * Ordered by speed/power consumption.
705  */
706 static struct {
707 	void	*id_fn;
708 	char	*id_name;
709 	int	id_cpuid2_flag;
710 } idle_tbl[] = {
711 	{ .id_fn = cpu_idle_spin, .id_name = "spin" },
712 	{ .id_fn = cpu_idle_mwait, .id_name = "mwait",
713 	    .id_cpuid2_flag = CPUID2_MON },
714 	{ .id_fn = cpu_idle_hlt, .id_name = "hlt" },
715 	{ .id_fn = cpu_idle_acpi, .id_name = "acpi" },
716 };
717 
718 static int
idle_sysctl_available(SYSCTL_HANDLER_ARGS)719 idle_sysctl_available(SYSCTL_HANDLER_ARGS)
720 {
721 	char *avail, *p;
722 	int error;
723 	int i;
724 
725 	avail = malloc(256, M_TEMP, M_WAITOK);
726 	p = avail;
727 	for (i = 0; i < nitems(idle_tbl); i++) {
728 		if (idle_tbl[i].id_cpuid2_flag != 0 &&
729 		    (cpu_feature2 & idle_tbl[i].id_cpuid2_flag) == 0)
730 			continue;
731 		if (strcmp(idle_tbl[i].id_name, "acpi") == 0 &&
732 		    cpu_idle_hook == NULL)
733 			continue;
734 		p += sprintf(p, "%s%s", p != avail ? ", " : "",
735 		    idle_tbl[i].id_name);
736 	}
737 	error = sysctl_handle_string(oidp, avail, 0, req);
738 	free(avail, M_TEMP);
739 	return (error);
740 }
741 
742 SYSCTL_PROC(_machdep, OID_AUTO, idle_available,
743     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE,
744     0, 0, idle_sysctl_available, "A",
745     "list of available idle functions");
746 
747 static bool
cpu_idle_selector(const char * new_idle_name)748 cpu_idle_selector(const char *new_idle_name)
749 {
750 	int i;
751 
752 	for (i = 0; i < nitems(idle_tbl); i++) {
753 		if (idle_tbl[i].id_cpuid2_flag != 0 &&
754 		    (cpu_feature2 & idle_tbl[i].id_cpuid2_flag) == 0)
755 			continue;
756 		if (strcmp(idle_tbl[i].id_name, "acpi") == 0 &&
757 		    cpu_idle_hook == NULL)
758 			continue;
759 		if (strcmp(idle_tbl[i].id_name, new_idle_name))
760 			continue;
761 		cpu_idle_fn = idle_tbl[i].id_fn;
762 		if (bootverbose)
763 			printf("CPU idle set to %s\n", idle_tbl[i].id_name);
764 		return (true);
765 	}
766 	return (false);
767 }
768 
769 static int
cpu_idle_sysctl(SYSCTL_HANDLER_ARGS)770 cpu_idle_sysctl(SYSCTL_HANDLER_ARGS)
771 {
772 	char buf[16], *p;
773 	int error, i;
774 
775 	p = "unknown";
776 	for (i = 0; i < nitems(idle_tbl); i++) {
777 		if (idle_tbl[i].id_fn == cpu_idle_fn) {
778 			p = idle_tbl[i].id_name;
779 			break;
780 		}
781 	}
782 	strncpy(buf, p, sizeof(buf));
783 	error = sysctl_handle_string(oidp, buf, sizeof(buf), req);
784 	if (error != 0 || req->newptr == NULL)
785 		return (error);
786 	return (cpu_idle_selector(buf) ? 0 : EINVAL);
787 }
788 
789 SYSCTL_PROC(_machdep, OID_AUTO, idle,
790     CTLTYPE_STRING | CTLFLAG_RW | CTLFLAG_MPSAFE,
791     0, 0, cpu_idle_sysctl, "A",
792     "currently selected idle function");
793 
794 static void
cpu_idle_tun(void * unused __unused)795 cpu_idle_tun(void *unused __unused)
796 {
797 	char tunvar[16];
798 
799 	if (TUNABLE_STR_FETCH("machdep.idle", tunvar, sizeof(tunvar)))
800 		cpu_idle_selector(tunvar);
801 	else if (cpu_vendor_id == CPU_VENDOR_AMD &&
802 	    CPUID_TO_FAMILY(cpu_id) == 0x17 && CPUID_TO_MODEL(cpu_id) == 0x1) {
803 		/* Ryzen erratas 1057, 1109. */
804 		cpu_idle_selector("hlt");
805 		idle_mwait = 0;
806 		mwait_cpustop_broken = true;
807 	}
808 
809 	if (cpu_vendor_id == CPU_VENDOR_INTEL && cpu_id == 0x506c9) {
810 		/*
811 		 * Apollo Lake errata APL31 (public errata APL30).
812 		 * Stores to the armed address range may not trigger
813 		 * MWAIT to resume execution.  OS needs to use
814 		 * interrupts to wake processors from MWAIT-induced
815 		 * sleep states.
816 		 */
817 		cpu_idle_apl31_workaround = 1;
818 		mwait_cpustop_broken = true;
819 	}
820 	TUNABLE_INT_FETCH("machdep.idle_apl31", &cpu_idle_apl31_workaround);
821 }
822 SYSINIT(cpu_idle_tun, SI_SUB_CPU, SI_ORDER_MIDDLE, cpu_idle_tun, NULL);
823 
824 static int panic_on_nmi = 0xff;
825 SYSCTL_INT(_machdep, OID_AUTO, panic_on_nmi, CTLFLAG_RWTUN,
826     &panic_on_nmi, 0,
827     "Panic on NMI: 1 = H/W failure; 2 = unknown; 0xff = all");
828 int nmi_is_broadcast = 1;
829 SYSCTL_INT(_machdep, OID_AUTO, nmi_is_broadcast, CTLFLAG_RWTUN,
830     &nmi_is_broadcast, 0,
831     "Chipset NMI is broadcast");
832 int (*apei_nmi)(void);
833 
834 void
nmi_call_kdb(u_int cpu,u_int type,struct trapframe * frame)835 nmi_call_kdb(u_int cpu, u_int type, struct trapframe *frame)
836 {
837 	bool claimed = false;
838 
839 #ifdef DEV_ISA
840 	/* machine/parity/power fail/"kitchen sink" faults */
841 	if (isa_nmi(frame->tf_err)) {
842 		claimed = true;
843 		if ((panic_on_nmi & 1) != 0)
844 			panic("NMI indicates hardware failure");
845 	}
846 #endif /* DEV_ISA */
847 
848 	/* ACPI Platform Error Interfaces callback. */
849 	if (apei_nmi != NULL && (*apei_nmi)())
850 		claimed = true;
851 
852 	/*
853 	 * NMIs can be useful for debugging.  They can be hooked up to a
854 	 * pushbutton, usually on an ISA, PCI, or PCIe card.  They can also be
855 	 * generated by an IPMI BMC, either manually or in response to a
856 	 * watchdog timeout.  For example, see the "power diag" command in
857 	 * ports/sysutils/ipmitool.  They can also be generated by a
858 	 * hypervisor; see "bhyvectl --inject-nmi".
859 	 */
860 
861 #ifdef KDB
862 	if (!claimed && (panic_on_nmi & 2) != 0) {
863 		if (debugger_on_panic) {
864 			printf("NMI/cpu%d ... going to debugger\n", cpu);
865 			claimed = kdb_trap(type, 0, frame);
866 		}
867 	}
868 #endif /* KDB */
869 
870 	if (!claimed && panic_on_nmi != 0)
871 		panic("NMI");
872 }
873 
874 void
nmi_handle_intr(u_int type,struct trapframe * frame)875 nmi_handle_intr(u_int type, struct trapframe *frame)
876 {
877 
878 #ifdef SMP
879 	if (nmi_is_broadcast) {
880 		nmi_call_kdb_smp(type, frame);
881 		return;
882 	}
883 #endif
884 	nmi_call_kdb(PCPU_GET(cpuid), type, frame);
885 }
886 
887 static int hw_ibrs_active;
888 int hw_ibrs_ibpb_active;
889 int hw_ibrs_disable = 1;
890 
891 SYSCTL_INT(_hw, OID_AUTO, ibrs_active, CTLFLAG_RD, &hw_ibrs_active, 0,
892     "Indirect Branch Restricted Speculation active");
893 
894 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, ibrs,
895     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
896     "Indirect Branch Restricted Speculation active");
897 
898 SYSCTL_INT(_machdep_mitigations_ibrs, OID_AUTO, active, CTLFLAG_RD,
899     &hw_ibrs_active, 0, "Indirect Branch Restricted Speculation active");
900 
901 void
hw_ibrs_recalculate(bool for_all_cpus)902 hw_ibrs_recalculate(bool for_all_cpus)
903 {
904 	if ((cpu_ia32_arch_caps & IA32_ARCH_CAP_IBRS_ALL) != 0) {
905 		x86_msr_op(MSR_IA32_SPEC_CTRL, (for_all_cpus ?
906 		    MSR_OP_RENDEZVOUS_ALL : MSR_OP_LOCAL) |
907 		    (hw_ibrs_disable != 0 ? MSR_OP_ANDNOT : MSR_OP_OR),
908 		    IA32_SPEC_CTRL_IBRS, NULL);
909 		hw_ibrs_active = hw_ibrs_disable == 0;
910 		hw_ibrs_ibpb_active = 0;
911 	} else {
912 		hw_ibrs_active = hw_ibrs_ibpb_active = (cpu_stdext_feature3 &
913 		    CPUID_STDEXT3_IBPB) != 0 && !hw_ibrs_disable;
914 	}
915 }
916 
917 static int
hw_ibrs_disable_handler(SYSCTL_HANDLER_ARGS)918 hw_ibrs_disable_handler(SYSCTL_HANDLER_ARGS)
919 {
920 	int error, val;
921 
922 	val = hw_ibrs_disable;
923 	error = sysctl_handle_int(oidp, &val, 0, req);
924 	if (error != 0 || req->newptr == NULL)
925 		return (error);
926 	hw_ibrs_disable = val != 0;
927 	hw_ibrs_recalculate(true);
928 	return (0);
929 }
930 SYSCTL_PROC(_hw, OID_AUTO, ibrs_disable, CTLTYPE_INT | CTLFLAG_RWTUN |
931     CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0, hw_ibrs_disable_handler, "I",
932     "Disable Indirect Branch Restricted Speculation");
933 
934 SYSCTL_PROC(_machdep_mitigations_ibrs, OID_AUTO, disable, CTLTYPE_INT |
935     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
936     hw_ibrs_disable_handler, "I",
937     "Disable Indirect Branch Restricted Speculation");
938 
939 int hw_ssb_active;
940 int hw_ssb_disable;
941 
942 SYSCTL_INT(_hw, OID_AUTO, spec_store_bypass_disable_active, CTLFLAG_RD,
943     &hw_ssb_active, 0,
944     "Speculative Store Bypass Disable active");
945 
946 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, ssb,
947     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
948     "Speculative Store Bypass Disable active");
949 
950 SYSCTL_INT(_machdep_mitigations_ssb, OID_AUTO, active, CTLFLAG_RD,
951     &hw_ssb_active, 0, "Speculative Store Bypass Disable active");
952 
953 static void
hw_ssb_set(bool enable,bool for_all_cpus)954 hw_ssb_set(bool enable, bool for_all_cpus)
955 {
956 
957 	if ((cpu_stdext_feature3 & CPUID_STDEXT3_SSBD) == 0) {
958 		hw_ssb_active = 0;
959 		return;
960 	}
961 	hw_ssb_active = enable;
962 	x86_msr_op(MSR_IA32_SPEC_CTRL,
963 	    (enable ? MSR_OP_OR : MSR_OP_ANDNOT) |
964 	    (for_all_cpus ? MSR_OP_SCHED_ALL : MSR_OP_LOCAL),
965 	    IA32_SPEC_CTRL_SSBD, NULL);
966 }
967 
968 void
hw_ssb_recalculate(bool all_cpus)969 hw_ssb_recalculate(bool all_cpus)
970 {
971 
972 	switch (hw_ssb_disable) {
973 	default:
974 		hw_ssb_disable = 0;
975 		/* FALLTHROUGH */
976 	case 0: /* off */
977 		hw_ssb_set(false, all_cpus);
978 		break;
979 	case 1: /* on */
980 		hw_ssb_set(true, all_cpus);
981 		break;
982 	case 2: /* auto */
983 		hw_ssb_set((cpu_ia32_arch_caps & IA32_ARCH_CAP_SSB_NO) != 0 ?
984 		    false : true, all_cpus);
985 		break;
986 	}
987 }
988 
989 static int
hw_ssb_disable_handler(SYSCTL_HANDLER_ARGS)990 hw_ssb_disable_handler(SYSCTL_HANDLER_ARGS)
991 {
992 	int error, val;
993 
994 	val = hw_ssb_disable;
995 	error = sysctl_handle_int(oidp, &val, 0, req);
996 	if (error != 0 || req->newptr == NULL)
997 		return (error);
998 	hw_ssb_disable = val;
999 	hw_ssb_recalculate(true);
1000 	return (0);
1001 }
1002 SYSCTL_PROC(_hw, OID_AUTO, spec_store_bypass_disable, CTLTYPE_INT |
1003     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1004     hw_ssb_disable_handler, "I",
1005     "Speculative Store Bypass Disable (0 - off, 1 - on, 2 - auto)");
1006 
1007 SYSCTL_PROC(_machdep_mitigations_ssb, OID_AUTO, disable, CTLTYPE_INT |
1008     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1009     hw_ssb_disable_handler, "I",
1010     "Speculative Store Bypass Disable (0 - off, 1 - on, 2 - auto)");
1011 
1012 int hw_mds_disable;
1013 
1014 /*
1015  * Handler for Microarchitectural Data Sampling issues.  Really not a
1016  * pointer to C function: on amd64 the code must not change any CPU
1017  * architectural state except possibly %rflags. Also, it is always
1018  * called with interrupts disabled.
1019  */
1020 void mds_handler_void(void);
1021 void mds_handler_verw(void);
1022 void mds_handler_ivb(void);
1023 void mds_handler_bdw(void);
1024 void mds_handler_skl_sse(void);
1025 void mds_handler_skl_avx(void);
1026 void mds_handler_skl_avx512(void);
1027 void mds_handler_silvermont(void);
1028 void (*mds_handler)(void) = mds_handler_void;
1029 
1030 static int
sysctl_hw_mds_disable_state_handler(SYSCTL_HANDLER_ARGS)1031 sysctl_hw_mds_disable_state_handler(SYSCTL_HANDLER_ARGS)
1032 {
1033 	const char *state;
1034 
1035 	if (mds_handler == mds_handler_void)
1036 		state = "inactive";
1037 	else if (mds_handler == mds_handler_verw)
1038 		state = "VERW";
1039 	else if (mds_handler == mds_handler_ivb)
1040 		state = "software IvyBridge";
1041 	else if (mds_handler == mds_handler_bdw)
1042 		state = "software Broadwell";
1043 	else if (mds_handler == mds_handler_skl_sse)
1044 		state = "software Skylake SSE";
1045 	else if (mds_handler == mds_handler_skl_avx)
1046 		state = "software Skylake AVX";
1047 	else if (mds_handler == mds_handler_skl_avx512)
1048 		state = "software Skylake AVX512";
1049 	else if (mds_handler == mds_handler_silvermont)
1050 		state = "software Silvermont";
1051 	else
1052 		state = "unknown";
1053 	return (SYSCTL_OUT(req, state, strlen(state)));
1054 }
1055 
1056 SYSCTL_PROC(_hw, OID_AUTO, mds_disable_state,
1057     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1058     sysctl_hw_mds_disable_state_handler, "A",
1059     "Microarchitectural Data Sampling Mitigation state");
1060 
1061 SYSCTL_NODE(_machdep_mitigations, OID_AUTO, mds,
1062     CTLFLAG_RW | CTLFLAG_MPSAFE, 0,
1063     "Microarchitectural Data Sampling Mitigation state");
1064 
1065 SYSCTL_PROC(_machdep_mitigations_mds, OID_AUTO, state,
1066     CTLTYPE_STRING | CTLFLAG_RD | CTLFLAG_MPSAFE, NULL, 0,
1067     sysctl_hw_mds_disable_state_handler, "A",
1068     "Microarchitectural Data Sampling Mitigation state");
1069 
1070 _Static_assert(__offsetof(struct pcpu, pc_mds_tmp) % 64 == 0, "MDS AVX512");
1071 
1072 void
hw_mds_recalculate(void)1073 hw_mds_recalculate(void)
1074 {
1075 	struct pcpu *pc;
1076 	vm_offset_t b64;
1077 	u_long xcr0;
1078 	int i;
1079 
1080 	/*
1081 	 * Allow user to force VERW variant even if MD_CLEAR is not
1082 	 * reported.  For instance, hypervisor might unknowingly
1083 	 * filter the cap out.
1084 	 * For the similar reasons, and for testing, allow to enable
1085 	 * mitigation even when MDS_NO cap is set.
1086 	 */
1087 	if (cpu_vendor_id != CPU_VENDOR_INTEL || hw_mds_disable == 0 ||
1088 	    ((cpu_ia32_arch_caps & IA32_ARCH_CAP_MDS_NO) != 0 &&
1089 	    hw_mds_disable == 3)) {
1090 		mds_handler = mds_handler_void;
1091 	} else if (((cpu_stdext_feature3 & CPUID_STDEXT3_MD_CLEAR) != 0 &&
1092 	    hw_mds_disable == 3) || hw_mds_disable == 1) {
1093 		mds_handler = mds_handler_verw;
1094 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1095 	    (CPUID_TO_MODEL(cpu_id) == 0x2e || CPUID_TO_MODEL(cpu_id) == 0x1e ||
1096 	    CPUID_TO_MODEL(cpu_id) == 0x1f || CPUID_TO_MODEL(cpu_id) == 0x1a ||
1097 	    CPUID_TO_MODEL(cpu_id) == 0x2f || CPUID_TO_MODEL(cpu_id) == 0x25 ||
1098 	    CPUID_TO_MODEL(cpu_id) == 0x2c || CPUID_TO_MODEL(cpu_id) == 0x2d ||
1099 	    CPUID_TO_MODEL(cpu_id) == 0x2a || CPUID_TO_MODEL(cpu_id) == 0x3e ||
1100 	    CPUID_TO_MODEL(cpu_id) == 0x3a) &&
1101 	    (hw_mds_disable == 2 || hw_mds_disable == 3)) {
1102 		/*
1103 		 * Nehalem, SandyBridge, IvyBridge
1104 		 */
1105 		CPU_FOREACH(i) {
1106 			pc = pcpu_find(i);
1107 			if (pc->pc_mds_buf == NULL) {
1108 				pc->pc_mds_buf = malloc_domainset(672, M_TEMP,
1109 				    DOMAINSET_PREF(pc->pc_domain), M_WAITOK);
1110 				bzero(pc->pc_mds_buf, 16);
1111 			}
1112 		}
1113 		mds_handler = mds_handler_ivb;
1114 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1115 	    (CPUID_TO_MODEL(cpu_id) == 0x3f || CPUID_TO_MODEL(cpu_id) == 0x3c ||
1116 	    CPUID_TO_MODEL(cpu_id) == 0x45 || CPUID_TO_MODEL(cpu_id) == 0x46 ||
1117 	    CPUID_TO_MODEL(cpu_id) == 0x56 || CPUID_TO_MODEL(cpu_id) == 0x4f ||
1118 	    CPUID_TO_MODEL(cpu_id) == 0x47 || CPUID_TO_MODEL(cpu_id) == 0x3d) &&
1119 	    (hw_mds_disable == 2 || hw_mds_disable == 3)) {
1120 		/*
1121 		 * Haswell, Broadwell
1122 		 */
1123 		CPU_FOREACH(i) {
1124 			pc = pcpu_find(i);
1125 			if (pc->pc_mds_buf == NULL) {
1126 				pc->pc_mds_buf = malloc_domainset(1536, M_TEMP,
1127 				    DOMAINSET_PREF(pc->pc_domain), M_WAITOK);
1128 				bzero(pc->pc_mds_buf, 16);
1129 			}
1130 		}
1131 		mds_handler = mds_handler_bdw;
1132 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1133 	    ((CPUID_TO_MODEL(cpu_id) == 0x55 && (cpu_id &
1134 	    CPUID_STEPPING) <= 5) ||
1135 	    CPUID_TO_MODEL(cpu_id) == 0x4e || CPUID_TO_MODEL(cpu_id) == 0x5e ||
1136 	    (CPUID_TO_MODEL(cpu_id) == 0x8e && (cpu_id &
1137 	    CPUID_STEPPING) <= 0xb) ||
1138 	    (CPUID_TO_MODEL(cpu_id) == 0x9e && (cpu_id &
1139 	    CPUID_STEPPING) <= 0xc)) &&
1140 	    (hw_mds_disable == 2 || hw_mds_disable == 3)) {
1141 		/*
1142 		 * Skylake, KabyLake, CoffeeLake, WhiskeyLake,
1143 		 * CascadeLake
1144 		 */
1145 		CPU_FOREACH(i) {
1146 			pc = pcpu_find(i);
1147 			if (pc->pc_mds_buf == NULL) {
1148 				pc->pc_mds_buf = malloc_domainset(6 * 1024,
1149 				    M_TEMP, DOMAINSET_PREF(pc->pc_domain),
1150 				    M_WAITOK);
1151 				b64 = (vm_offset_t)malloc_domainset(64 + 63,
1152 				    M_TEMP, DOMAINSET_PREF(pc->pc_domain),
1153 				    M_WAITOK);
1154 				pc->pc_mds_buf64 = (void *)roundup2(b64, 64);
1155 				bzero(pc->pc_mds_buf64, 64);
1156 			}
1157 		}
1158 		xcr0 = rxcr(0);
1159 		if ((xcr0 & XFEATURE_ENABLED_ZMM_HI256) != 0 &&
1160 		    (cpu_stdext_feature & CPUID_STDEXT_AVX512DQ) != 0)
1161 			mds_handler = mds_handler_skl_avx512;
1162 		else if ((xcr0 & XFEATURE_ENABLED_AVX) != 0 &&
1163 		    (cpu_feature2 & CPUID2_AVX) != 0)
1164 			mds_handler = mds_handler_skl_avx;
1165 		else
1166 			mds_handler = mds_handler_skl_sse;
1167 	} else if (CPUID_TO_FAMILY(cpu_id) == 0x6 &&
1168 	    ((CPUID_TO_MODEL(cpu_id) == 0x37 ||
1169 	    CPUID_TO_MODEL(cpu_id) == 0x4a ||
1170 	    CPUID_TO_MODEL(cpu_id) == 0x4c ||
1171 	    CPUID_TO_MODEL(cpu_id) == 0x4d ||
1172 	    CPUID_TO_MODEL(cpu_id) == 0x5a ||
1173 	    CPUID_TO_MODEL(cpu_id) == 0x5d ||
1174 	    CPUID_TO_MODEL(cpu_id) == 0x6e ||
1175 	    CPUID_TO_MODEL(cpu_id) == 0x65 ||
1176 	    CPUID_TO_MODEL(cpu_id) == 0x75 ||
1177 	    CPUID_TO_MODEL(cpu_id) == 0x1c ||
1178 	    CPUID_TO_MODEL(cpu_id) == 0x26 ||
1179 	    CPUID_TO_MODEL(cpu_id) == 0x27 ||
1180 	    CPUID_TO_MODEL(cpu_id) == 0x35 ||
1181 	    CPUID_TO_MODEL(cpu_id) == 0x36 ||
1182 	    CPUID_TO_MODEL(cpu_id) == 0x7a))) {
1183 		/* Silvermont, Airmont */
1184 		CPU_FOREACH(i) {
1185 			pc = pcpu_find(i);
1186 			if (pc->pc_mds_buf == NULL)
1187 				pc->pc_mds_buf = malloc(256, M_TEMP, M_WAITOK);
1188 		}
1189 		mds_handler = mds_handler_silvermont;
1190 	} else {
1191 		hw_mds_disable = 0;
1192 		mds_handler = mds_handler_void;
1193 	}
1194 }
1195 
1196 static void
hw_mds_recalculate_boot(void * arg __unused)1197 hw_mds_recalculate_boot(void *arg __unused)
1198 {
1199 
1200 	hw_mds_recalculate();
1201 }
1202 SYSINIT(mds_recalc, SI_SUB_SMP, SI_ORDER_ANY, hw_mds_recalculate_boot, NULL);
1203 
1204 static int
sysctl_mds_disable_handler(SYSCTL_HANDLER_ARGS)1205 sysctl_mds_disable_handler(SYSCTL_HANDLER_ARGS)
1206 {
1207 	int error, val;
1208 
1209 	val = hw_mds_disable;
1210 	error = sysctl_handle_int(oidp, &val, 0, req);
1211 	if (error != 0 || req->newptr == NULL)
1212 		return (error);
1213 	if (val < 0 || val > 3)
1214 		return (EINVAL);
1215 	hw_mds_disable = val;
1216 	hw_mds_recalculate();
1217 	return (0);
1218 }
1219 
1220 SYSCTL_PROC(_hw, OID_AUTO, mds_disable, CTLTYPE_INT |
1221     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1222     sysctl_mds_disable_handler, "I",
1223     "Microarchitectural Data Sampling Mitigation "
1224     "(0 - off, 1 - on VERW, 2 - on SW, 3 - on AUTO)");
1225 
1226 SYSCTL_PROC(_machdep_mitigations_mds, OID_AUTO, disable, CTLTYPE_INT |
1227     CTLFLAG_RWTUN | CTLFLAG_NOFETCH | CTLFLAG_MPSAFE, NULL, 0,
1228     sysctl_mds_disable_handler, "I",
1229     "Microarchitectural Data Sampling Mitigation "
1230     "(0 - off, 1 - on VERW, 2 - on SW, 3 - on AUTO)");
1231 
1232 /*
1233  * Intel Transactional Memory Asynchronous Abort Mitigation
1234  * CVE-2019-11135
1235  */
1236 int x86_taa_enable;
1237 int x86_taa_state;
1238 enum {
1239 	TAA_NONE	= 0,	/* No mitigation enabled */
1240 	TAA_TSX_DISABLE	= 1,	/* Disable TSX via MSR */
1241 	TAA_VERW	= 2,	/* Use VERW mitigation */
1242 	TAA_AUTO	= 3,	/* Automatically select the mitigation */
1243 
1244 	/* The states below are not selectable by the operator */
1245 
1246 	TAA_TAA_UC	= 4,	/* Mitigation present in microcode */
1247 	TAA_NOT_PRESENT	= 5	/* TSX is not present */
1248 };
1249 
1250 static void
taa_set(bool enable,bool all)1251 taa_set(bool enable, bool all)
1252 {
1253 
1254 	x86_msr_op(MSR_IA32_TSX_CTRL,
1255 	    (enable ? MSR_OP_OR : MSR_OP_ANDNOT) |
1256 	    (all ? MSR_OP_RENDEZVOUS_ALL : MSR_OP_LOCAL),
1257 	    IA32_TSX_CTRL_RTM_DISABLE | IA32_TSX_CTRL_TSX_CPUID_CLEAR,
1258 	    NULL);
1259 }
1260 
1261 void
x86_taa_recalculate(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
taa_recalculate_boot(void * arg __unused)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
sysctl_taa_handler(SYSCTL_HANDLER_ARGS)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
sysctl_taa_state_handler(SYSCTL_HANDLER_ARGS)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
x86_rngds_mitg_recalculate(bool all_cpus)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_ALL : MSR_OP_LOCAL),
1423 	    IA32_RNGDS_MITG_DIS, NULL);
1424 }
1425 
1426 static int
sysctl_rngds_mitg_enable_handler(SYSCTL_HANDLER_ARGS)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
sysctl_rngds_state_handler(SYSCTL_HANDLER_ARGS)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
disable_wp(void)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
restore_wp(bool old_wp)1482 restore_wp(bool old_wp)
1483 {
1484 
1485 	if (old_wp)
1486 		load_cr0(rcr0() | CR0_WP);
1487 }
1488 
1489 bool
acpi_get_fadt_bootflags(uint16_t * flagsp)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 
1510 DEFINE_IFUNC(, uint64_t, rdtsc_ordered, (void))
1511 {
1512 	bool cpu_is_amd = cpu_vendor_id == CPU_VENDOR_AMD ||
1513 	    cpu_vendor_id == CPU_VENDOR_HYGON;
1514 
1515 	if ((amd_feature & AMDID_RDTSCP) != 0)
1516 		return (rdtscp);
1517 	else if ((cpu_feature & CPUID_SSE2) != 0)
1518 		return (cpu_is_amd ? rdtsc_ordered_mfence :
1519 		    rdtsc_ordered_lfence);
1520 	else
1521 		return (rdtsc);
1522 }
1523