xref: /netbsd-src/sys/arch/x86/x86/cpu.c (revision 1f05e87cf6355a797166eee0e58e6c51b599f251)
1 /*	$NetBSD: cpu.c,v 1.210 2024/04/22 23:07:47 andvar Exp $	*/
2 
3 /*
4  * Copyright (c) 2000-2020 NetBSD Foundation, Inc.
5  * All rights reserved.
6  *
7  * This code is derived from software contributed to The NetBSD Foundation
8  * by Bill Sommerfeld of RedBack Networks Inc, and by Andrew Doran.
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  *
19  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
20  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
21  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
22  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
23  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
24  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
25  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
26  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
27  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
28  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
29  * POSSIBILITY OF SUCH DAMAGE.
30  */
31 
32 /*
33  * Copyright (c) 1999 Stefan Grefen
34  *
35  * Redistribution and use in source and binary forms, with or without
36  * modification, are permitted provided that the following conditions
37  * are met:
38  * 1. Redistributions of source code must retain the above copyright
39  *    notice, this list of conditions and the following disclaimer.
40  * 2. Redistributions in binary form must reproduce the above copyright
41  *    notice, this list of conditions and the following disclaimer in the
42  *    documentation and/or other materials provided with the distribution.
43  * 3. All advertising materials mentioning features or use of this software
44  *    must display the following acknowledgement:
45  *      This product includes software developed by the NetBSD
46  *      Foundation, Inc. and its contributors.
47  * 4. Neither the name of The NetBSD Foundation nor the names of its
48  *    contributors may be used to endorse or promote products derived
49  *    from this software without specific prior written permission.
50  *
51  * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS ``AS IS'' AND ANY
52  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
53  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
54  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR AND CONTRIBUTORS BE LIABLE
55  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
56  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
57  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
58  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
59  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
60  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
61  * SUCH DAMAGE.
62  */
63 
64 #include <sys/cdefs.h>
65 __KERNEL_RCSID(0, "$NetBSD: cpu.c,v 1.210 2024/04/22 23:07:47 andvar Exp $");
66 
67 #include "opt_ddb.h"
68 #include "opt_mpbios.h"		/* for MPDEBUG */
69 #include "opt_mtrr.h"
70 #include "opt_multiprocessor.h"
71 #include "opt_svs.h"
72 
73 #include "lapic.h"
74 #include "ioapic.h"
75 #include "acpica.h"
76 #include "hpet.h"
77 
78 #include <sys/param.h>
79 #include <sys/proc.h>
80 #include <sys/systm.h>
81 #include <sys/device.h>
82 #include <sys/cpu.h>
83 #include <sys/cpufreq.h>
84 #include <sys/idle.h>
85 #include <sys/atomic.h>
86 #include <sys/reboot.h>
87 #include <sys/csan.h>
88 
89 #include <uvm/uvm.h>
90 
91 #include "acpica.h"		/* for NACPICA, for mp_verbose */
92 
93 #include <x86/machdep.h>
94 #include <machine/cpufunc.h>
95 #include <machine/cpuvar.h>
96 #include <machine/pmap.h>
97 #include <machine/vmparam.h>
98 #if defined(MULTIPROCESSOR)
99 #include <machine/mpbiosvar.h>
100 #endif
101 #include <machine/mpconfig.h>		/* for mp_verbose */
102 #include <machine/pcb.h>
103 #include <machine/specialreg.h>
104 #include <machine/segments.h>
105 #include <machine/gdt.h>
106 #include <machine/mtrr.h>
107 #include <machine/pio.h>
108 #include <machine/cpu_counter.h>
109 #include <machine/pmap_private.h>
110 
111 #include <x86/fpu.h>
112 
113 #if NACPICA > 0
114 #include <dev/acpi/acpi_srat.h>
115 #endif
116 
117 #if NLAPIC > 0
118 #include <machine/apicvar.h>
119 #include <machine/i82489reg.h>
120 #include <machine/i82489var.h>
121 #endif
122 
123 #include <dev/ic/mc146818reg.h>
124 #include <dev/ic/hpetvar.h>
125 #include <i386/isa/nvram.h>
126 #include <dev/isa/isareg.h>
127 
128 #include "tsc.h"
129 
130 #ifndef XENPV
131 #include "hyperv.h"
132 #if NHYPERV > 0
133 #include <x86/x86/hypervvar.h>
134 #endif
135 #endif
136 
137 #ifdef XEN
138 #include <xen/hypervisor.h>
139 #endif
140 
141 static int	cpu_match(device_t, cfdata_t, void *);
142 static void	cpu_attach(device_t, device_t, void *);
143 static void	cpu_defer(device_t);
144 static int	cpu_rescan(device_t, const char *, const int *);
145 static void	cpu_childdetached(device_t, device_t);
146 static bool	cpu_stop(device_t);
147 static bool	cpu_suspend(device_t, const pmf_qual_t *);
148 static bool	cpu_resume(device_t, const pmf_qual_t *);
149 static bool	cpu_shutdown(device_t, int);
150 
151 struct cpu_softc {
152 	device_t sc_dev;		/* device tree glue */
153 	struct cpu_info *sc_info;	/* pointer to CPU info */
154 	bool sc_wasonline;
155 };
156 
157 #ifdef MULTIPROCESSOR
158 int mp_cpu_start(struct cpu_info *, paddr_t);
159 void mp_cpu_start_cleanup(struct cpu_info *);
160 const struct cpu_functions mp_cpu_funcs = { mp_cpu_start, NULL,
161 					    mp_cpu_start_cleanup };
162 #endif
163 
164 
165 CFATTACH_DECL2_NEW(cpu, sizeof(struct cpu_softc),
166     cpu_match, cpu_attach, NULL, NULL, cpu_rescan, cpu_childdetached);
167 
168 /*
169  * Statically-allocated CPU info for the primary CPU (or the only
170  * CPU, on uniprocessors).  The CPU info list is initialized to
171  * point at it.
172  */
173 struct cpu_info cpu_info_primary __aligned(CACHE_LINE_SIZE) = {
174 	.ci_dev = 0,
175 	.ci_self = &cpu_info_primary,
176 	.ci_idepth = -1,
177 	.ci_curlwp = &lwp0,
178 	.ci_curldt = -1,
179 	.ci_kfpu_spl = -1,
180 };
181 
182 struct cpu_info *cpu_info_list = &cpu_info_primary;
183 
184 #ifdef i386
185 void		cpu_set_tss_gates(struct cpu_info *);
186 #endif
187 
188 static void	cpu_init_idle_lwp(struct cpu_info *);
189 
190 uint32_t cpu_feature[7] __read_mostly; /* X86 CPUID feature bits */
191 			/* [0] basic features cpuid.1:%edx
192 			 * [1] basic features cpuid.1:%ecx (CPUID2_xxx bits)
193 			 * [2] extended features cpuid:80000001:%edx
194 			 * [3] extended features cpuid:80000001:%ecx
195 			 * [4] VIA padlock features
196 			 * [5] structured extended features cpuid.7:%ebx
197 			 * [6] structured extended features cpuid.7:%ecx
198 			 */
199 
200 #ifdef MULTIPROCESSOR
201 bool x86_mp_online;
202 paddr_t mp_trampoline_paddr = MP_TRAMPOLINE;
203 #endif
204 #if NLAPIC > 0
205 static vaddr_t cmos_data_mapping;
206 #endif
207 struct cpu_info *cpu_starting;
208 
209 #ifdef MULTIPROCESSOR
210 void		cpu_hatch(void *);
211 static void	cpu_boot_secondary(struct cpu_info *ci);
212 static void	cpu_start_secondary(struct cpu_info *ci);
213 #if NLAPIC > 0
214 static void	cpu_copy_trampoline(paddr_t);
215 #endif
216 #endif /* MULTIPROCESSOR */
217 
218 /*
219  * Runs once per boot once multiprocessor goo has been detected and
220  * the local APIC on the boot processor has been mapped.
221  *
222  * Called from lapic_boot_init() (from mpbios_scan()).
223  */
224 #if NLAPIC > 0
225 void
cpu_init_first(void)226 cpu_init_first(void)
227 {
228 
229 	cpu_info_primary.ci_cpuid = lapic_cpu_number();
230 
231 	cmos_data_mapping = uvm_km_alloc(kernel_map, PAGE_SIZE, 0, UVM_KMF_VAONLY);
232 	if (cmos_data_mapping == 0)
233 		panic("No KVA for page 0");
234 	pmap_kenter_pa(cmos_data_mapping, 0, VM_PROT_READ|VM_PROT_WRITE, 0);
235 	pmap_update(pmap_kernel());
236 }
237 #endif
238 
239 static int
cpu_match(device_t parent,cfdata_t match,void * aux)240 cpu_match(device_t parent, cfdata_t match, void *aux)
241 {
242 
243 	return 1;
244 }
245 
246 #ifdef __HAVE_PCPU_AREA
247 void
cpu_pcpuarea_init(struct cpu_info * ci)248 cpu_pcpuarea_init(struct cpu_info *ci)
249 {
250 	struct vm_page *pg;
251 	size_t i, npages;
252 	vaddr_t base, va;
253 	paddr_t pa;
254 
255 	CTASSERT(sizeof(struct pcpu_entry) % PAGE_SIZE == 0);
256 
257 	npages = sizeof(struct pcpu_entry) / PAGE_SIZE;
258 	base = (vaddr_t)&pcpuarea->ent[cpu_index(ci)];
259 
260 	for (i = 0; i < npages; i++) {
261 		pg = uvm_pagealloc(NULL, 0, NULL, UVM_PGA_ZERO);
262 		if (pg == NULL) {
263 			panic("failed to allocate pcpu PA");
264 		}
265 
266 		va = base + i * PAGE_SIZE;
267 		pa = VM_PAGE_TO_PHYS(pg);
268 
269 		pmap_kenter_pa(va, pa, VM_PROT_READ|VM_PROT_WRITE, 0);
270 	}
271 
272 	pmap_update(pmap_kernel());
273 }
274 #endif
275 
276 static void
cpu_vm_init(struct cpu_info * ci)277 cpu_vm_init(struct cpu_info *ci)
278 {
279 	unsigned int ncolors = 2;
280 
281 	/*
282 	 * XXX: for AP's the cache info has not been initialized yet
283 	 * but that does not matter because uvm only pays attention at
284 	 * the maximum only. We should fix it once cpus have different
285 	 * cache sizes.
286 	 */
287 	for (unsigned int i = CAI_ICACHE; i <= CAI_L2CACHE; i++) {
288 		struct x86_cache_info *cai;
289 		unsigned int tcolors;
290 
291 		cai = &ci->ci_cinfo[i];
292 
293 		tcolors = atop(cai->cai_totalsize);
294 		switch (cai->cai_associativity) {
295 		case 0xff:
296 			tcolors = 1; /* fully associative */
297 			break;
298 		case 0:
299 		case 1:
300 			break;
301 		default:
302 			tcolors /= cai->cai_associativity;
303 		}
304 		if (tcolors <= ncolors)
305 			continue;
306 		ncolors = tcolors;
307 	}
308 
309 	/*
310 	 * If the desired number of colors is not a power of
311 	 * two, it won't be good.  Find the greatest power of
312 	 * two which is an even divisor of the number of colors,
313 	 * to preserve even coloring of pages.
314 	 */
315 	if (ncolors & (ncolors - 1) ) {
316 		unsigned int try, picked = 1;
317 		for (try = 1; try < ncolors; try *= 2) {
318 			if (ncolors % try == 0) picked = try;
319 		}
320 		if (picked == 1) {
321 			panic("desired number of cache colors %u is "
322 			" > 1, but not even!", ncolors);
323 		}
324 		ncolors = picked;
325 	}
326 
327 	/*
328 	 * Knowing the size of the largest cache on this CPU, potentially
329 	 * re-color our pages.
330 	 */
331 	aprint_debug_dev(ci->ci_dev, "%d page colors\n", ncolors);
332 	uvm_page_recolor(ncolors);
333 
334 	pmap_tlb_cpu_init(ci);
335 #ifndef __HAVE_DIRECT_MAP
336 	pmap_vpage_cpu_init(ci);
337 #endif
338 }
339 
340 static void
cpu_attach(device_t parent,device_t self,void * aux)341 cpu_attach(device_t parent, device_t self, void *aux)
342 {
343 	struct cpu_softc *sc = device_private(self);
344 	struct cpu_attach_args *caa = aux;
345 	struct cpu_info *ci;
346 	uintptr_t ptr;
347 #if NLAPIC > 0
348 	int cpunum = caa->cpu_number;
349 #endif
350 	static bool again;
351 
352 	sc->sc_dev = self;
353 
354 	if (ncpu > maxcpus) {
355 #ifndef _LP64
356 		aprint_error(": too many CPUs, please use NetBSD/amd64\n");
357 #else
358 		aprint_error(": too many CPUs\n");
359 #endif
360 		return;
361 	}
362 
363 	/*
364 	 * If we're an Application Processor, allocate a cpu_info
365 	 * structure, otherwise use the primary's.
366 	 */
367 	if (caa->cpu_role == CPU_ROLE_AP) {
368 		if ((boothowto & RB_MD1) != 0) {
369 			aprint_error(": multiprocessor boot disabled\n");
370 			if (!pmf_device_register(self, NULL, NULL))
371 				aprint_error_dev(self,
372 				    "couldn't establish power handler\n");
373 			return;
374 		}
375 		aprint_naive(": Application Processor\n");
376 		ptr = (uintptr_t)uvm_km_alloc(kernel_map,
377 		    sizeof(*ci) + CACHE_LINE_SIZE - 1, 0,
378 		    UVM_KMF_WIRED|UVM_KMF_ZERO);
379 		ci = (struct cpu_info *)roundup2(ptr, CACHE_LINE_SIZE);
380 		ci->ci_curldt = -1;
381 	} else {
382 		aprint_naive(": %s Processor\n",
383 		    caa->cpu_role == CPU_ROLE_SP ? "Single" : "Boot");
384 		ci = &cpu_info_primary;
385 #if NLAPIC > 0
386 		if (cpunum != lapic_cpu_number()) {
387 			/* XXX should be done earlier. */
388 			uint32_t reg;
389 			aprint_verbose("\n");
390 			aprint_verbose_dev(self, "running CPU at apic %d"
391 			    " instead of at expected %d", lapic_cpu_number(),
392 			    cpunum);
393 			reg = lapic_readreg(LAPIC_ID);
394 			lapic_writereg(LAPIC_ID, (reg & ~LAPIC_ID_MASK) |
395 			    (cpunum << LAPIC_ID_SHIFT));
396 		}
397 		if (cpunum != lapic_cpu_number()) {
398 			aprint_error_dev(self, "unable to reset apic id\n");
399 		}
400 #endif
401 	}
402 
403 	ci->ci_self = ci;
404 	sc->sc_info = ci;
405 	ci->ci_dev = self;
406 	ci->ci_acpiid = caa->cpu_id;
407 	ci->ci_cpuid = caa->cpu_number;
408 	ci->ci_func = caa->cpu_func;
409 	ci->ci_kfpu_spl = -1;
410 	aprint_normal("\n");
411 
412 	/* Must be before mi_cpu_attach(). */
413 	cpu_vm_init(ci);
414 
415 	if (caa->cpu_role == CPU_ROLE_AP) {
416 		int error;
417 
418 		error = mi_cpu_attach(ci);
419 		if (error != 0) {
420 			aprint_error_dev(self,
421 			    "mi_cpu_attach failed with %d\n", error);
422 			return;
423 		}
424 #ifdef __HAVE_PCPU_AREA
425 		cpu_pcpuarea_init(ci);
426 #endif
427 		cpu_init_tss(ci);
428 	} else {
429 		KASSERT(ci->ci_data.cpu_idlelwp != NULL);
430 #if NACPICA > 0
431 		/* Parse out NUMA info for cpu_identify(). */
432 		acpisrat_init();
433 #endif
434 	}
435 
436 #ifdef SVS
437 	cpu_svs_init(ci);
438 #endif
439 
440 	pmap_reference(pmap_kernel());
441 	ci->ci_pmap = pmap_kernel();
442 	ci->ci_tlbstate = TLBSTATE_STALE;
443 
444 	/*
445 	 * Boot processor may not be attached first, but the below
446 	 * must be done to allow booting other processors.
447 	 */
448 	if (!again) {
449 		/* Make sure DELAY() (likely i8254_delay()) is initialized. */
450 		DELAY(1);
451 
452 		/*
453 		 * Basic init.  Compute an approximate frequency for the TSC
454 		 * using the i8254.  If there's a HPET we'll redo it later.
455 		 */
456 		atomic_or_32(&ci->ci_flags, CPUF_PRESENT | CPUF_PRIMARY);
457 		cpu_intr_init(ci);
458 		tsc_setfunc(ci);
459 		cpu_get_tsc_freq(ci);
460 		cpu_init(ci);
461 #ifdef i386
462 		cpu_set_tss_gates(ci);
463 #endif
464 		pmap_cpu_init_late(ci);
465 #if NLAPIC > 0
466 		if (caa->cpu_role != CPU_ROLE_SP) {
467 			/* Enable lapic. */
468 			lapic_enable();
469 			lapic_set_lvt();
470 			if (!vm_guest_is_xenpvh_or_pvhvm())
471 				lapic_calibrate_timer(false);
472 		}
473 #endif
474 		kcsan_cpu_init(ci);
475 		again = true;
476 	}
477 
478 	/* further PCB init done later. */
479 
480 	switch (caa->cpu_role) {
481 	case CPU_ROLE_SP:
482 		atomic_or_32(&ci->ci_flags, CPUF_SP);
483 		cpu_identify(ci);
484 		x86_errata();
485 		x86_cpu_idle_init();
486 #ifdef XENPVHVM
487 		xen_hvm_init_cpu(ci);
488 #endif
489 		break;
490 
491 	case CPU_ROLE_BP:
492 		atomic_or_32(&ci->ci_flags, CPUF_BSP);
493 		cpu_identify(ci);
494 		x86_errata();
495 		x86_cpu_idle_init();
496 #ifdef XENPVHVM
497 		xen_hvm_init_cpu(ci);
498 #endif
499 		break;
500 
501 #ifdef MULTIPROCESSOR
502 	case CPU_ROLE_AP:
503 		/*
504 		 * report on an AP
505 		 */
506 		cpu_intr_init(ci);
507 		idt_vec_init_cpu_md(&ci->ci_idtvec, cpu_index(ci));
508 		gdt_alloc_cpu(ci);
509 #ifdef i386
510 		cpu_set_tss_gates(ci);
511 #endif
512 		pmap_cpu_init_late(ci);
513 		cpu_start_secondary(ci);
514 		if (ci->ci_flags & CPUF_PRESENT) {
515 			struct cpu_info *tmp;
516 
517 			cpu_identify(ci);
518 			tmp = cpu_info_list;
519 			while (tmp->ci_next)
520 				tmp = tmp->ci_next;
521 
522 			tmp->ci_next = ci;
523 		}
524 		break;
525 #endif
526 
527 	default:
528 		panic("unknown processor type??\n");
529 	}
530 
531 	pat_init(ci);
532 
533 	if (!pmf_device_register1(self, cpu_suspend, cpu_resume, cpu_shutdown))
534 		aprint_error_dev(self, "couldn't establish power handler\n");
535 
536 #ifdef MULTIPROCESSOR
537 	if (mp_verbose) {
538 		struct lwp *l = ci->ci_data.cpu_idlelwp;
539 		struct pcb *pcb = lwp_getpcb(l);
540 
541 		aprint_verbose_dev(self,
542 		    "idle lwp at %p, idle sp at %p\n",
543 		    l,
544 #ifdef i386
545 		    (void *)pcb->pcb_esp
546 #else
547 		    (void *)pcb->pcb_rsp
548 #endif
549 		);
550 	}
551 #endif
552 
553 	/*
554 	 * Postpone the "cpufeaturebus" scan.
555 	 * It is safe to scan the pseudo-bus
556 	 * only after all CPUs have attached.
557 	 */
558 	(void)config_defer(self, cpu_defer);
559 }
560 
561 static void
cpu_defer(device_t self)562 cpu_defer(device_t self)
563 {
564 	cpu_rescan(self, NULL, NULL);
565 }
566 
567 static int
cpu_rescan(device_t self,const char * ifattr,const int * locators)568 cpu_rescan(device_t self, const char *ifattr, const int *locators)
569 {
570 	struct cpu_softc *sc = device_private(self);
571 	struct cpufeature_attach_args cfaa;
572 	struct cpu_info *ci = sc->sc_info;
573 
574 	/*
575 	 * If we booted with RB_MD1 to disable multiprocessor, the
576 	 * auto-configuration data still contains the additional
577 	 * CPUs.   But their initialization was mostly bypassed
578 	 * during attach, so we have to make sure we don't look at
579 	 * their featurebus info, since it wasn't retrieved.
580 	 */
581 	if (ci == NULL)
582 		return 0;
583 
584 	memset(&cfaa, 0, sizeof(cfaa));
585 	cfaa.ci = ci;
586 
587 	if (ifattr_match(ifattr, "cpufeaturebus")) {
588 		if (ci->ci_frequency == NULL) {
589 			cfaa.name = "frequency";
590 			ci->ci_frequency =
591 			    config_found(self, &cfaa, NULL,
592 					 CFARGS(.iattr = "cpufeaturebus"));
593 		}
594 
595 		if (ci->ci_padlock == NULL) {
596 			cfaa.name = "padlock";
597 			ci->ci_padlock =
598 			    config_found(self, &cfaa, NULL,
599 					 CFARGS(.iattr = "cpufeaturebus"));
600 		}
601 
602 		if (ci->ci_temperature == NULL) {
603 			cfaa.name = "temperature";
604 			ci->ci_temperature =
605 			    config_found(self, &cfaa, NULL,
606 					 CFARGS(.iattr = "cpufeaturebus"));
607 		}
608 
609 		if (ci->ci_vm == NULL) {
610 			cfaa.name = "vm";
611 			ci->ci_vm =
612 			    config_found(self, &cfaa, NULL,
613 					 CFARGS(.iattr = "cpufeaturebus"));
614 		}
615 	}
616 
617 	return 0;
618 }
619 
620 static void
cpu_childdetached(device_t self,device_t child)621 cpu_childdetached(device_t self, device_t child)
622 {
623 	struct cpu_softc *sc = device_private(self);
624 	struct cpu_info *ci = sc->sc_info;
625 
626 	if (ci->ci_frequency == child)
627 		ci->ci_frequency = NULL;
628 
629 	if (ci->ci_padlock == child)
630 		ci->ci_padlock = NULL;
631 
632 	if (ci->ci_temperature == child)
633 		ci->ci_temperature = NULL;
634 
635 	if (ci->ci_vm == child)
636 		ci->ci_vm = NULL;
637 }
638 
639 /*
640  * Initialize the processor appropriately.
641  */
642 
643 void
cpu_init(struct cpu_info * ci)644 cpu_init(struct cpu_info *ci)
645 {
646 	extern int x86_fpu_save;
647 	uint32_t cr4 = 0;
648 
649 	lcr0(rcr0() | CR0_WP);
650 
651 	/* If global TLB caching is supported, enable it */
652 	if (cpu_feature[0] & CPUID_PGE)
653 		cr4 |= CR4_PGE;
654 
655 	/*
656 	 * If we have FXSAVE/FXRESTOR, use them.
657 	 */
658 	if (cpu_feature[0] & CPUID_FXSR) {
659 		cr4 |= CR4_OSFXSR;
660 
661 		/*
662 		 * If we have SSE/SSE2, enable XMM exceptions.
663 		 */
664 		if (cpu_feature[0] & (CPUID_SSE|CPUID_SSE2))
665 			cr4 |= CR4_OSXMMEXCPT;
666 	}
667 
668 	/* If xsave is supported, enable it */
669 	if (cpu_feature[1] & CPUID2_XSAVE)
670 		cr4 |= CR4_OSXSAVE;
671 
672 	/* If SMEP is supported, enable it */
673 	if (cpu_feature[5] & CPUID_SEF_SMEP)
674 		cr4 |= CR4_SMEP;
675 
676 	/* If SMAP is supported, enable it */
677 	if (cpu_feature[5] & CPUID_SEF_SMAP)
678 		cr4 |= CR4_SMAP;
679 
680 #ifdef SVS
681 	/* If PCID is supported, enable it */
682 	if (svs_pcid)
683 		cr4 |= CR4_PCIDE;
684 #endif
685 
686 	if (cr4) {
687 		cr4 |= rcr4();
688 		lcr4(cr4);
689 	}
690 
691 	/*
692 	 * Changing CR4 register may change cpuid values. For example, setting
693 	 * CR4_OSXSAVE sets CPUID2_OSXSAVE. The CPUID2_OSXSAVE is in
694 	 * ci_feat_val[1], so update it.
695 	 * XXX Other than ci_feat_val[1] might be changed.
696 	 */
697 	if (cpuid_level >= 1) {
698 		u_int descs[4];
699 
700 		x86_cpuid(1, descs);
701 		ci->ci_feat_val[1] = descs[2];
702 	}
703 
704 	if (CPU_IS_PRIMARY(ci) &&
705 	    x86_fpu_save >= FPU_SAVE_FXSAVE) {
706 		fpuinit_mxcsr_mask();
707 	}
708 
709 	/* If xsave is enabled, enable all fpu features */
710 	if (cr4 & CR4_OSXSAVE)
711 		wrxcr(0, x86_xsave_features & XCR0_FPU);
712 
713 #ifdef MTRR
714 	/*
715 	 * On a P6 or above, initialize MTRR's if the hardware supports them.
716 	 */
717 	if (cpu_feature[0] & CPUID_MTRR) {
718 		if ((ci->ci_flags & CPUF_AP) == 0)
719 			i686_mtrr_init_first();
720 		mtrr_init_cpu(ci);
721 	}
722 
723 #ifdef i386
724 	if (strcmp((char *)(ci->ci_vendor), "AuthenticAMD") == 0) {
725 		/*
726 		 * Must be a K6-2 Step >= 7 or a K6-III.
727 		 */
728 		if (CPUID_TO_FAMILY(ci->ci_signature) == 5) {
729 			if (CPUID_TO_MODEL(ci->ci_signature) > 8 ||
730 			    (CPUID_TO_MODEL(ci->ci_signature) == 8 &&
731 			     CPUID_TO_STEPPING(ci->ci_signature) >= 7)) {
732 				mtrr_funcs = &k6_mtrr_funcs;
733 				k6_mtrr_init_first();
734 				mtrr_init_cpu(ci);
735 			}
736 		}
737 	}
738 #endif	/* i386 */
739 #endif /* MTRR */
740 
741 	if (ci != &cpu_info_primary) {
742 		/* Synchronize TSC */
743 		atomic_or_32(&ci->ci_flags, CPUF_RUNNING);
744 		tsc_sync_ap(ci);
745 	} else {
746 		atomic_or_32(&ci->ci_flags, CPUF_RUNNING);
747 	}
748 }
749 
750 #ifdef MULTIPROCESSOR
751 void
cpu_boot_secondary_processors(void)752 cpu_boot_secondary_processors(void)
753 {
754 	struct cpu_info *ci;
755 	kcpuset_t *cpus;
756 	u_long i;
757 
758 	/* Now that we know the number of CPUs, patch the text segment. */
759 	x86_patch(false);
760 
761 #if NACPICA > 0
762 	/* Finished with NUMA info for now. */
763 	acpisrat_exit();
764 #endif
765 
766 	kcpuset_create(&cpus, true);
767 	kcpuset_set(cpus, cpu_index(curcpu()));
768 	for (i = 0; i < maxcpus; i++) {
769 		ci = cpu_lookup(i);
770 		if (ci == NULL)
771 			continue;
772 		if (ci->ci_data.cpu_idlelwp == NULL)
773 			continue;
774 		if ((ci->ci_flags & CPUF_PRESENT) == 0)
775 			continue;
776 		if (ci->ci_flags & (CPUF_BSP|CPUF_SP|CPUF_PRIMARY))
777 			continue;
778 		cpu_boot_secondary(ci);
779 		kcpuset_set(cpus, cpu_index(ci));
780 	}
781 	while (!kcpuset_match(cpus, kcpuset_running))
782 		;
783 	kcpuset_destroy(cpus);
784 
785 	x86_mp_online = true;
786 
787 	/* Now that we know about the TSC, attach the timecounter. */
788 	tsc_tc_init();
789 }
790 #endif
791 
792 static void
cpu_init_idle_lwp(struct cpu_info * ci)793 cpu_init_idle_lwp(struct cpu_info *ci)
794 {
795 	struct lwp *l = ci->ci_data.cpu_idlelwp;
796 	struct pcb *pcb = lwp_getpcb(l);
797 
798 	pcb->pcb_cr0 = rcr0();
799 }
800 
801 void
cpu_init_idle_lwps(void)802 cpu_init_idle_lwps(void)
803 {
804 	struct cpu_info *ci;
805 	u_long i;
806 
807 	for (i = 0; i < maxcpus; i++) {
808 		ci = cpu_lookup(i);
809 		if (ci == NULL)
810 			continue;
811 		if (ci->ci_data.cpu_idlelwp == NULL)
812 			continue;
813 		if ((ci->ci_flags & CPUF_PRESENT) == 0)
814 			continue;
815 		cpu_init_idle_lwp(ci);
816 	}
817 }
818 
819 #ifdef MULTIPROCESSOR
820 void
cpu_start_secondary(struct cpu_info * ci)821 cpu_start_secondary(struct cpu_info *ci)
822 {
823 	u_long psl;
824 	int i;
825 
826 #if NLAPIC > 0
827 	paddr_t mp_pdirpa;
828 	mp_pdirpa = pmap_init_tmp_pgtbl(mp_trampoline_paddr);
829 	cpu_copy_trampoline(mp_pdirpa);
830 #endif
831 
832 	atomic_or_32(&ci->ci_flags, CPUF_AP);
833 	ci->ci_curlwp = ci->ci_data.cpu_idlelwp;
834 	if (CPU_STARTUP(ci, mp_trampoline_paddr) != 0) {
835 		return;
836 	}
837 
838 	/*
839 	 * Wait for it to become ready.   Setting cpu_starting opens the
840 	 * initial gate and allows the AP to start soft initialization.
841 	 */
842 	KASSERT(cpu_starting == NULL);
843 	cpu_starting = ci;
844 	for (i = 100000; (!(ci->ci_flags & CPUF_PRESENT)) && i > 0; i--) {
845 		delay_func(10);
846 	}
847 
848 	if ((ci->ci_flags & CPUF_PRESENT) == 0) {
849 		aprint_error_dev(ci->ci_dev, "failed to become ready\n");
850 #if defined(MPDEBUG) && defined(DDB)
851 		printf("dropping into debugger; continue from here to resume boot\n");
852 		Debugger();
853 #endif
854 	} else {
855 		/*
856 		 * Synchronize time stamp counters. Invalidate cache and do
857 		 * twice (in tsc_sync_bp) to minimize possible cache effects.
858 		 * Disable interrupts to try and rule out any external
859 		 * interference.
860 		 */
861 		psl = x86_read_psl();
862 		x86_disable_intr();
863 		tsc_sync_bp(ci);
864 		x86_write_psl(psl);
865 	}
866 
867 	CPU_START_CLEANUP(ci);
868 	cpu_starting = NULL;
869 }
870 
871 void
cpu_boot_secondary(struct cpu_info * ci)872 cpu_boot_secondary(struct cpu_info *ci)
873 {
874 	int64_t drift;
875 	u_long psl;
876 	int i;
877 
878 	atomic_or_32(&ci->ci_flags, CPUF_GO);
879 	for (i = 100000; (!(ci->ci_flags & CPUF_RUNNING)) && i > 0; i--) {
880 		delay_func(10);
881 	}
882 	if ((ci->ci_flags & CPUF_RUNNING) == 0) {
883 		aprint_error_dev(ci->ci_dev, "failed to start\n");
884 #if defined(MPDEBUG) && defined(DDB)
885 		printf("dropping into debugger; continue from here to resume boot\n");
886 		Debugger();
887 #endif
888 	} else {
889 		/* Synchronize TSC again, check for drift. */
890 		drift = ci->ci_data.cpu_cc_skew;
891 		psl = x86_read_psl();
892 		x86_disable_intr();
893 		tsc_sync_bp(ci);
894 		x86_write_psl(psl);
895 		drift -= ci->ci_data.cpu_cc_skew;
896 		aprint_debug_dev(ci->ci_dev, "TSC skew=%lld drift=%lld\n",
897 		    (long long)ci->ci_data.cpu_cc_skew, (long long)drift);
898 		tsc_sync_drift(drift);
899 	}
900 }
901 
902 /*
903  * The CPU ends up here when it's ready to run.
904  * This is called from code in mptramp.s; at this point, we are running
905  * in the idle pcb/idle stack of the new CPU.  When this function returns,
906  * this processor will enter the idle loop and start looking for work.
907  */
908 void
cpu_hatch(void * v)909 cpu_hatch(void *v)
910 {
911 	struct cpu_info *ci = (struct cpu_info *)v;
912 	struct pcb *pcb;
913 	int s, i;
914 
915 	/* ------------------------------------------------------------- */
916 
917 	/*
918 	 * This section of code must be compiled with SSP disabled, to
919 	 * prevent a race against cpu0. See sys/conf/ssp.mk.
920 	 */
921 
922 	/*
923 	 * Initialize MSRs on this CPU:
924 	 *
925 	 * - On amd64: Enables SYSCALL/SYSRET.
926 	 *
927 	 * - On amd64: Sets up %fs and %gs so that %gs points to the
928 	 *   current struct cpu_info as needed for CPUVAR(...),
929 	 *   curcpu(), and curlwp.
930 	 *
931 	 *   (On i386, CPUVAR(...), curcpu(), and curlwp are made to
932 	 *   work first by the conifguration of segment descriptors in
933 	 *   the Global Descriptor Table (GDT) in initgdt.)
934 	 *
935 	 * - Enables the no-execute bit if supported.
936 	 *
937 	 * Thus, after this point, CPUVAR(...), curcpu(), and curlwp
938 	 * will work on this CPU.
939 	 *
940 	 * Note: The call to cpu_init_msrs for cpu0 happens in
941 	 * init386/init_x86_64.
942 	 */
943 	cpu_init_msrs(ci, true);
944 
945 	cpu_probe(ci);
946 	cpu_speculation_init(ci);
947 #if NHYPERV > 0
948 	hyperv_init_cpu(ci);
949 #endif
950 
951 	ci->ci_data.cpu_cc_freq = cpu_info_primary.ci_data.cpu_cc_freq;
952 	/* cpu_get_tsc_freq(ci); */
953 
954 	KDASSERT((ci->ci_flags & CPUF_PRESENT) == 0);
955 
956 	/*
957 	 * Synchronize the TSC for the first time. Note that interrupts are
958 	 * off at this point.
959 	 */
960 	atomic_or_32(&ci->ci_flags, CPUF_PRESENT);
961 	tsc_sync_ap(ci);
962 
963 	/* ------------------------------------------------------------- */
964 
965 	/*
966 	 * Wait to be brought online.
967 	 *
968 	 * Use MONITOR/MWAIT if available. These instructions put the CPU in
969 	 * a low consumption mode (C-state), and if the TSC is not invariant,
970 	 * this causes the TSC to drift. We want this to happen, so that we
971 	 * can later detect (in tsc_tc_init) any abnormal drift with invariant
972 	 * TSCs. That's just for safety; by definition such drifts should
973 	 * never occur with invariant TSCs.
974 	 *
975 	 * If not available, try PAUSE. We'd like to use HLT, but we have
976 	 * interrupts off.
977 	 */
978 	while ((ci->ci_flags & CPUF_GO) == 0) {
979 		if ((cpu_feature[1] & CPUID2_MONITOR) != 0) {
980 			x86_monitor(&ci->ci_flags, 0, 0);
981 			if ((ci->ci_flags & CPUF_GO) != 0) {
982 				continue;
983 			}
984 			x86_mwait(0, 0);
985 		} else {
986 	/*
987 	 * XXX The loop repetition count could be a lot higher, but
988 	 * XXX currently qemu emulator takes a _very_long_time_ to
989 	 * XXX execute the pause instruction.  So for now, use a low
990 	 * XXX value to allow the cpu to hatch before timing out.
991 	 */
992 			for (i = 50; i != 0; i--) {
993 				x86_pause();
994 			}
995 		}
996 	}
997 
998 	/* Because the text may have been patched in x86_patch(). */
999 	wbinvd();
1000 	x86_flush();
1001 	tlbflushg();
1002 
1003 	KASSERT((ci->ci_flags & CPUF_RUNNING) == 0);
1004 
1005 #ifdef PAE
1006 	pd_entry_t * l3_pd = ci->ci_pae_l3_pdir;
1007 	for (i = 0 ; i < PDP_SIZE; i++) {
1008 		l3_pd[i] = pmap_kernel()->pm_pdirpa[i] | PTE_P;
1009 	}
1010 	lcr3(ci->ci_pae_l3_pdirpa);
1011 #else
1012 	lcr3(pmap_pdirpa(pmap_kernel(), 0));
1013 #endif
1014 
1015 	pcb = lwp_getpcb(curlwp);
1016 	pcb->pcb_cr3 = rcr3();
1017 	pcb = lwp_getpcb(ci->ci_data.cpu_idlelwp);
1018 	lcr0(pcb->pcb_cr0);
1019 
1020 	cpu_init_idt(ci);
1021 	gdt_init_cpu(ci);
1022 #if NLAPIC > 0
1023 	lapic_enable();
1024 	lapic_set_lvt();
1025 #endif
1026 
1027 	fpuinit(ci);
1028 	lldt(GSYSSEL(GLDT_SEL, SEL_KPL));
1029 	ltr(ci->ci_tss_sel);
1030 
1031 	/*
1032 	 * cpu_init will re-synchronize the TSC, and will detect any abnormal
1033 	 * drift that would have been caused by the use of MONITOR/MWAIT
1034 	 * above.
1035 	 */
1036 	cpu_init(ci);
1037 #ifdef XENPVHVM
1038 	xen_hvm_init_cpu(ci);
1039 #endif
1040 	(*x86_initclock_func)();
1041 	cpu_get_tsc_freq(ci);
1042 
1043 	s = splhigh();
1044 #if NLAPIC > 0
1045 	lapic_write_tpri(0);
1046 #endif
1047 	x86_enable_intr();
1048 	splx(s);
1049 	x86_errata();
1050 
1051 	aprint_debug_dev(ci->ci_dev, "running\n");
1052 
1053 	kcsan_cpu_init(ci);
1054 
1055 	idle_loop(NULL);
1056 	KASSERT(false);
1057 }
1058 #endif
1059 
1060 #if defined(DDB)
1061 
1062 #include <ddb/db_output.h>
1063 #include <machine/db_machdep.h>
1064 
1065 /*
1066  * Dump CPU information from ddb.
1067  */
1068 void
cpu_debug_dump(void)1069 cpu_debug_dump(void)
1070 {
1071 	struct cpu_info *ci;
1072 	CPU_INFO_ITERATOR cii;
1073 	const char sixtyfour64space[] =
1074 #ifdef _LP64
1075 			   "        "
1076 #endif
1077 			   "";
1078 
1079 	db_printf("addr		%sdev	id	flags	ipis	spl curlwp 		"
1080 		  "\n", sixtyfour64space);
1081 	for (CPU_INFO_FOREACH(cii, ci)) {
1082 		db_printf("%p	%s	%ld	%x	%x	%d  %10p\n",
1083 		    ci,
1084 		    ci->ci_dev == NULL ? "BOOT" : device_xname(ci->ci_dev),
1085 		    (long)ci->ci_cpuid,
1086 		    ci->ci_flags, ci->ci_ipis, ci->ci_ilevel,
1087 		    ci->ci_curlwp);
1088 	}
1089 }
1090 #endif
1091 
1092 #ifdef MULTIPROCESSOR
1093 #if NLAPIC > 0
1094 static void
cpu_copy_trampoline(paddr_t pdir_pa)1095 cpu_copy_trampoline(paddr_t pdir_pa)
1096 {
1097 	extern uint32_t nox_flag;
1098 	extern u_char cpu_spinup_trampoline[];
1099 	extern u_char cpu_spinup_trampoline_end[];
1100 	vaddr_t mp_trampoline_vaddr;
1101 	struct {
1102 		uint32_t large;
1103 		uint32_t nox;
1104 		uint32_t pdir;
1105 	} smp_data;
1106 	CTASSERT(sizeof(smp_data) == 3 * 4);
1107 
1108 	smp_data.large = (pmap_largepages != 0);
1109 	smp_data.nox = nox_flag;
1110 	smp_data.pdir = (uint32_t)(pdir_pa & 0xFFFFFFFF);
1111 
1112 	/* Enter the physical address */
1113 	mp_trampoline_vaddr = uvm_km_alloc(kernel_map, PAGE_SIZE, 0,
1114 	    UVM_KMF_VAONLY);
1115 	pmap_kenter_pa(mp_trampoline_vaddr, mp_trampoline_paddr,
1116 	    VM_PROT_READ | VM_PROT_WRITE, 0);
1117 	pmap_update(pmap_kernel());
1118 
1119 	/* Copy boot code */
1120 	memcpy((void *)mp_trampoline_vaddr,
1121 	    cpu_spinup_trampoline,
1122 	    cpu_spinup_trampoline_end - cpu_spinup_trampoline);
1123 
1124 	/* Copy smp_data at the end */
1125 	memcpy((void *)(mp_trampoline_vaddr + PAGE_SIZE - sizeof(smp_data)),
1126 	    &smp_data, sizeof(smp_data));
1127 
1128 	pmap_kremove(mp_trampoline_vaddr, PAGE_SIZE);
1129 	pmap_update(pmap_kernel());
1130 	uvm_km_free(kernel_map, mp_trampoline_vaddr, PAGE_SIZE, UVM_KMF_VAONLY);
1131 }
1132 #endif
1133 
1134 int
mp_cpu_start(struct cpu_info * ci,paddr_t target)1135 mp_cpu_start(struct cpu_info *ci, paddr_t target)
1136 {
1137 #if NLAPIC > 0
1138 	int error;
1139 
1140 	/*
1141 	 * Bootstrap code must be addressable in real mode
1142 	 * and it must be page aligned.
1143 	 */
1144 	KASSERT(target < 0x10000 && target % PAGE_SIZE == 0);
1145 
1146 	/*
1147 	 * "The BSP must initialize CMOS shutdown code to 0Ah ..."
1148 	 */
1149 
1150 	outb(IO_RTC, NVRAM_RESET);
1151 	outb(IO_RTC+1, NVRAM_RESET_JUMP);
1152 
1153 	/*
1154 	 * "and the warm reset vector (DWORD based at 40:67) to point
1155 	 * to the AP startup code ..."
1156 	 */
1157 	unsigned short dwordptr[2];
1158 	dwordptr[0] = 0;
1159 	dwordptr[1] = target >> 4;
1160 
1161 	memcpy((uint8_t *)cmos_data_mapping + 0x467, dwordptr, 4);
1162 
1163 	if ((cpu_feature[0] & CPUID_APIC) == 0) {
1164 		aprint_error("mp_cpu_start: CPU does not have APIC\n");
1165 		return ENODEV;
1166 	}
1167 
1168 	/*
1169 	 * ... prior to executing the following sequence:".  We'll also add in
1170 	 * local cache flush, in case the BIOS has left the AP with its cache
1171 	 * disabled.  It may not be able to cope with MP coherency.
1172 	 */
1173 	wbinvd();
1174 
1175 	if (ci->ci_flags & CPUF_AP) {
1176 		error = x86_ipi_init(ci->ci_cpuid);
1177 		if (error != 0) {
1178 			aprint_error_dev(ci->ci_dev, "%s: IPI not taken (1)\n",
1179 			    __func__);
1180 			return error;
1181 		}
1182 		delay_func(10000);
1183 
1184 		error = x86_ipi_startup(ci->ci_cpuid, target / PAGE_SIZE);
1185 		if (error != 0) {
1186 			aprint_error_dev(ci->ci_dev, "%s: IPI not taken (2)\n",
1187 			    __func__);
1188 			return error;
1189 		}
1190 		delay_func(200);
1191 
1192 		error = x86_ipi_startup(ci->ci_cpuid, target / PAGE_SIZE);
1193 		if (error != 0) {
1194 			aprint_error_dev(ci->ci_dev, "%s: IPI not taken (3)\n",
1195 			    __func__);
1196 			return error;
1197 		}
1198 		delay_func(200);
1199 	}
1200 
1201 	return 0;
1202 #else
1203 	return ENODEV;
1204 #endif /* NLAPIC > 0 */
1205 }
1206 
1207 void
mp_cpu_start_cleanup(struct cpu_info * ci)1208 mp_cpu_start_cleanup(struct cpu_info *ci)
1209 {
1210 	/*
1211 	 * Ensure the NVRAM reset byte contains something vaguely sane.
1212 	 */
1213 
1214 	outb(IO_RTC, NVRAM_RESET);
1215 	outb(IO_RTC+1, NVRAM_RESET_RST);
1216 }
1217 #endif
1218 
1219 #ifdef __x86_64__
1220 typedef void (vector)(void);
1221 extern vector Xsyscall, Xsyscall32, Xsyscall_svs;
1222 #endif
1223 
1224 /*
1225  * cpu_init_msrs(ci, full)
1226  *
1227  *	Initialize some Model-Specific Registers (MSRs) on the current
1228  *	CPU, whose struct cpu_info pointer is ci, for:
1229  *
1230  *	- SYSCALL/SYSRET.
1231  *	- %fs/%gs on amd64 if `full' is true; needed to make
1232  *	  CPUVAR(...), curcpu(), and curlwp work.  (We do this at boot,
1233  *	  but skip it on ACPI wakeup.)
1234  *	- No-execute bit, if supported.
1235  *
1236  *	References:
1237  *
1238  *	- Intel 64 and IA-32 Architectures Software Developer's Manual,
1239  *	  Volume 3: System Programming Guide, Order Number 325384,
1240  *	  April 2022, Sec. 5.8.8 `Fast System Calls in 64-Bit Mode',
1241  *	  pp. 5-22 through 5-23.
1242  *
1243  *	- Intel 64 and IA-32 Architectures Software Developer's Manual,
1244  *	  Volume 4: Model-Specific Registers, Order Number 335592,
1245  *	  April 2022, Sec. 2.1 `Architectural MSRs', Table 2-2,
1246  *	  pp. 2-60 through 2-61.
1247  */
1248 void
cpu_init_msrs(struct cpu_info * ci,bool full)1249 cpu_init_msrs(struct cpu_info *ci, bool full)
1250 {
1251 #ifdef __x86_64__
1252 	/*
1253 	 * On amd64, set up the syscall target address registers
1254 	 * for SYSCALL/SYSRET:
1255 	 *
1256 	 * - IA32_STAR, c000_0081h (MSR_STAR): System Call Target
1257 	 *   Address.  Code and stack segment selectors for SYSRET
1258 	 *   (bits 48:63) and SYSCALL (bits 32:47).
1259 	 *
1260 	 * - IA32_LSTAR, c000_0082h (MSR_LSTAR): IA-32e Mode System
1261 	 *   Call Target Address.  Target rip for SYSCALL when executed
1262 	 *   in 64-bit mode.
1263 	 *
1264 	 * - IA32_CSTAR, c000_0083h (MSR_CSTAR): IA-32e Mode System
1265 	 *   Call Target Address.  Target rip for SYSCALL when executed
1266 	 *   in compatibility mode.  (XXX Manual says this is `[n]ot
1267 	 *   used, as the SYSCALL instruction is not recognized in
1268 	 *   compatibility mode', so why do we set it?)
1269 	 *
1270 	 * - IA32_FMASK, c000_0084h (MSR_SFMASK): System Call Flag
1271 	 *   Mask.  Mask for the RFLAGS register on SYSCALL.
1272 	 */
1273 	wrmsr(MSR_STAR,
1274 	    ((uint64_t)GSEL(GCODE_SEL, SEL_KPL) << 32) |
1275 	    ((uint64_t)LSEL(LSYSRETBASE_SEL, SEL_UPL) << 48));
1276 	wrmsr(MSR_LSTAR, (uint64_t)Xsyscall);
1277 	wrmsr(MSR_CSTAR, (uint64_t)Xsyscall32);
1278 	wrmsr(MSR_SFMASK, PSL_NT|PSL_T|PSL_I|PSL_C|PSL_D|PSL_AC);
1279 
1280 #ifdef SVS
1281 	if (svs_enabled)
1282 		wrmsr(MSR_LSTAR, (uint64_t)Xsyscall_svs);
1283 #endif
1284 
1285 	/*
1286 	 * On amd64 if `full' is true -- used at boot, but not on ACPI
1287 	 * wakeup -- then additionally set up %fs and %gs:
1288 	 *
1289 	 * - IA32_FS_BASE, c000_0100h (MSR_FSBASE): Base address of
1290 	 *   %fs.  Not used in NetBSD kernel, so zero it.
1291 	 *
1292 	 * - IA32_GS_BASE, c000_0101h (MSR_GSBASE): Base address of
1293 	 *   %gs.  Used in NetBSD kernel by CPUVAR(...), curcpu(), and
1294 	 *   curlwp for access to the CPU-local area, so set it to ci.
1295 	 *
1296 	 * - IA32_KERNEL_GS_BASE, c000_0102h (MSR_KERNELGSBASE): Base
1297 	 *   address of what swapgs will leave in %gs when switching to
1298 	 *   userland.  Zero for now; will be set to pcb->pcb_gs in
1299 	 *   cpu_switchto for user threads.
1300 	 */
1301 	if (full) {
1302 		wrmsr(MSR_FSBASE, 0);
1303 		wrmsr(MSR_GSBASE, (uint64_t)ci);
1304 		wrmsr(MSR_KERNELGSBASE, 0);
1305 	}
1306 #endif	/* __x86_64__ */
1307 
1308 	/*
1309 	 * If the no-execute bit is supported, enable it in:
1310 	 *
1311 	 * - IA32_EFER, c000_0080h (MSR_EFER): Extended Feature
1312          *   Enables.
1313 	 */
1314 	if (cpu_feature[2] & CPUID_NOX)
1315 		wrmsr(MSR_EFER, rdmsr(MSR_EFER) | EFER_NXE);
1316 }
1317 
1318 void
cpu_offline_md(void)1319 cpu_offline_md(void)
1320 {
1321 	return;
1322 }
1323 
1324 /* XXX joerg restructure and restart CPUs individually */
1325 static bool
cpu_stop(device_t dv)1326 cpu_stop(device_t dv)
1327 {
1328 	struct cpu_softc *sc = device_private(dv);
1329 	struct cpu_info *ci = sc->sc_info;
1330 	int err;
1331 
1332 	KASSERT((ci->ci_flags & CPUF_PRESENT) != 0);
1333 
1334 	if (CPU_IS_PRIMARY(ci))
1335 		return true;
1336 
1337 	if (ci->ci_data.cpu_idlelwp == NULL)
1338 		return true;
1339 
1340 	sc->sc_wasonline = !(ci->ci_schedstate.spc_flags & SPCF_OFFLINE);
1341 
1342 	if (sc->sc_wasonline) {
1343 		mutex_enter(&cpu_lock);
1344 		err = cpu_setstate(ci, false);
1345 		mutex_exit(&cpu_lock);
1346 
1347 		if (err != 0)
1348 			return false;
1349 	}
1350 
1351 	return true;
1352 }
1353 
1354 static bool
cpu_suspend(device_t dv,const pmf_qual_t * qual)1355 cpu_suspend(device_t dv, const pmf_qual_t *qual)
1356 {
1357 	struct cpu_softc *sc = device_private(dv);
1358 	struct cpu_info *ci = sc->sc_info;
1359 
1360 	if ((ci->ci_flags & CPUF_PRESENT) == 0)
1361 		return true;
1362 	else {
1363 		cpufreq_suspend(ci);
1364 	}
1365 
1366 	return cpu_stop(dv);
1367 }
1368 
1369 static bool
cpu_resume(device_t dv,const pmf_qual_t * qual)1370 cpu_resume(device_t dv, const pmf_qual_t *qual)
1371 {
1372 	struct cpu_softc *sc = device_private(dv);
1373 	struct cpu_info *ci = sc->sc_info;
1374 	int err = 0;
1375 
1376 	if ((ci->ci_flags & CPUF_PRESENT) == 0)
1377 		return true;
1378 
1379 	if (CPU_IS_PRIMARY(ci))
1380 		goto out;
1381 
1382 	if (ci->ci_data.cpu_idlelwp == NULL)
1383 		goto out;
1384 
1385 	if (sc->sc_wasonline) {
1386 		mutex_enter(&cpu_lock);
1387 		err = cpu_setstate(ci, true);
1388 		mutex_exit(&cpu_lock);
1389 	}
1390 
1391 out:
1392 	if (err != 0)
1393 		return false;
1394 
1395 	cpufreq_resume(ci);
1396 
1397 	return true;
1398 }
1399 
1400 static bool
cpu_shutdown(device_t dv,int how)1401 cpu_shutdown(device_t dv, int how)
1402 {
1403 	struct cpu_softc *sc = device_private(dv);
1404 	struct cpu_info *ci = sc->sc_info;
1405 
1406 	if ((ci->ci_flags & CPUF_BSP) != 0)
1407 		return false;
1408 
1409 	if ((ci->ci_flags & CPUF_PRESENT) == 0)
1410 		return true;
1411 
1412 	return cpu_stop(dv);
1413 }
1414 
1415 /* Get the TSC frequency and set it to ci->ci_data.cpu_cc_freq. */
1416 void
cpu_get_tsc_freq(struct cpu_info * ci)1417 cpu_get_tsc_freq(struct cpu_info *ci)
1418 {
1419 	uint64_t freq = 0, freq_from_cpuid, t0, t1;
1420 	int64_t overhead;
1421 
1422 	if (CPU_IS_PRIMARY(ci) && cpu_hascounter()) {
1423 		/*
1424 		 * If it's the first call of this function, try to get TSC
1425 		 * freq from CPUID by calling cpu_tsc_freq_cpuid().
1426 		 * The function also set lapic_per_second variable if it's
1427 		 * known. This is required for Intel's Comet Lake and newer
1428 		 * processors to set LAPIC timer correctly.
1429 		 */
1430 		if (ci->ci_data.cpu_cc_freq == 0)
1431 			freq = freq_from_cpuid = cpu_tsc_freq_cpuid(ci);
1432 		if (freq != 0)
1433 			aprint_debug_dev(ci->ci_dev, "TSC freq "
1434 			    "from CPUID %" PRIu64 " Hz\n", freq);
1435 #if NHPET > 0
1436 		if (freq == 0) {
1437 			freq = hpet_tsc_freq();
1438 			if (freq != 0)
1439 				aprint_debug_dev(ci->ci_dev, "TSC freq "
1440 				    "from HPET %" PRIu64 " Hz\n", freq);
1441 		}
1442 #endif
1443 		if (freq == 0) {
1444 			/*
1445 			 * Work out the approximate overhead involved below.
1446 			 * Discard the result of the first go around the
1447 			 * loop.
1448 			 */
1449 			overhead = 0;
1450 			for (int i = 0; i <= 8; i++) {
1451 				const int s = splhigh();
1452 				t0 = cpu_counter();
1453 				delay_func(0);
1454 				t1 = cpu_counter();
1455 				splx(s);
1456 				if (i > 0) {
1457 					overhead += (t1 - t0);
1458 				}
1459 			}
1460 			overhead >>= 3;
1461 
1462 			/*
1463 			 * Now do the calibration.
1464 			 */
1465 			freq = 0;
1466 			for (int i = 0; i < 1000; i++) {
1467 				const int s = splhigh();
1468 				t0 = cpu_counter();
1469 				delay_func(100);
1470 				t1 = cpu_counter();
1471 				splx(s);
1472 				freq += t1 - t0 - overhead;
1473 			}
1474 			freq = freq * 10;
1475 
1476 			aprint_debug_dev(ci->ci_dev, "TSC freq "
1477 			    "from delay %" PRIu64 " Hz\n", freq);
1478 		}
1479 		if (ci->ci_data.cpu_cc_freq != 0) {
1480 			freq_from_cpuid = cpu_tsc_freq_cpuid(ci);
1481 			if ((freq_from_cpuid != 0)
1482 			    && (freq != freq_from_cpuid))
1483 				aprint_verbose_dev(ci->ci_dev, "TSC freq "
1484 				    "calibrated %" PRIu64 " Hz\n", freq);
1485 		}
1486 	} else {
1487 		freq = cpu_info_primary.ci_data.cpu_cc_freq;
1488 	}
1489 
1490 	ci->ci_data.cpu_cc_freq = freq;
1491 }
1492 
1493 void
x86_cpu_idle_mwait(void)1494 x86_cpu_idle_mwait(void)
1495 {
1496 	struct cpu_info *ci = curcpu();
1497 
1498 	KASSERT(ci->ci_ilevel == IPL_NONE);
1499 
1500 	x86_monitor(&ci->ci_want_resched, 0, 0);
1501 	if (__predict_false(ci->ci_want_resched)) {
1502 		return;
1503 	}
1504 	x86_mwait(0, 0);
1505 }
1506 
1507 void
x86_cpu_idle_halt(void)1508 x86_cpu_idle_halt(void)
1509 {
1510 	struct cpu_info *ci = curcpu();
1511 
1512 	KASSERT(ci->ci_ilevel == IPL_NONE);
1513 
1514 	x86_disable_intr();
1515 	if (!__predict_false(ci->ci_want_resched)) {
1516 		x86_stihlt();
1517 	} else {
1518 		x86_enable_intr();
1519 	}
1520 }
1521 
1522 /*
1523  * Loads pmap for the current CPU.
1524  */
1525 void
cpu_load_pmap(struct pmap * pmap,struct pmap * oldpmap)1526 cpu_load_pmap(struct pmap *pmap, struct pmap *oldpmap)
1527 {
1528 
1529 	KASSERT(kpreempt_disabled());
1530 
1531 #ifdef SVS
1532 	if (svs_enabled && pmap_is_user(pmap)) {
1533 		svs_pdir_switch(pmap);
1534 	}
1535 #endif
1536 
1537 #ifdef PAE
1538 	struct cpu_info *ci = curcpu();
1539 	bool interrupts_enabled;
1540 	pd_entry_t *l3_pd = ci->ci_pae_l3_pdir;
1541 	int i;
1542 
1543 	/*
1544 	 * disable interrupts to block TLB shootdowns, which can reload cr3.
1545 	 * while this doesn't block NMIs, it's probably ok as NMIs unlikely
1546 	 * reload cr3.
1547 	 */
1548 	interrupts_enabled = (x86_read_flags() & PSL_I) != 0;
1549 	if (interrupts_enabled)
1550 		x86_disable_intr();
1551 
1552 	for (i = 0 ; i < PDP_SIZE; i++) {
1553 		l3_pd[i] = pmap->pm_pdirpa[i] | PTE_P;
1554 	}
1555 
1556 	if (interrupts_enabled)
1557 		x86_enable_intr();
1558 	tlbflush();
1559 #else
1560 	lcr3(pmap_pdirpa(pmap, 0));
1561 #endif
1562 }
1563 
1564 /*
1565  * Notify all other cpus to halt.
1566  */
1567 
1568 void
cpu_broadcast_halt(void)1569 cpu_broadcast_halt(void)
1570 {
1571 	x86_broadcast_ipi(X86_IPI_HALT);
1572 }
1573 
1574 /*
1575  * Send a dummy ipi to a cpu to force it to run splraise()/spllower(),
1576  * and trigger an AST on the running LWP.
1577  */
1578 
1579 void
cpu_kick(struct cpu_info * ci)1580 cpu_kick(struct cpu_info *ci)
1581 {
1582 	x86_send_ipi(ci, X86_IPI_AST);
1583 }
1584