xref: /netbsd-src/sys/arch/xen/x86/cpu.c (revision bdc22b2e01993381dcefeff2bc9b56ca75a4235c)
1 /*	$NetBSD: cpu.c,v 1.122 2018/06/23 10:30:22 jdolecek Exp $	*/
2 
3 /*-
4  * Copyright (c) 2000 The NetBSD Foundation, Inc.
5  * Copyright (c) 2002, 2006, 2007 YAMAMOTO Takashi,
6  * All rights reserved.
7  *
8  * This code is derived from software contributed to The NetBSD Foundation
9  * by RedBack Networks Inc.
10  *
11  * Author: Bill Sommerfeld
12  *
13  * Redistribution and use in source and binary forms, with or without
14  * modification, are permitted provided that the following conditions
15  * are met:
16  * 1. Redistributions of source code must retain the above copyright
17  *    notice, this list of conditions and the following disclaimer.
18  * 2. Redistributions in binary form must reproduce the above copyright
19  *    notice, this list of conditions and the following disclaimer in the
20  *    documentation and/or other materials provided with the distribution.
21  *
22  * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS
23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED
24  * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR
25  * PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS
26  * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR
27  * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF
28  * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS
29  * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN
30  * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE)
31  * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE
32  * POSSIBILITY OF SUCH DAMAGE.
33  */
34 
35 /*
36  * Copyright (c) 1999 Stefan Grefen
37  *
38  * Redistribution and use in source and binary forms, with or without
39  * modification, are permitted provided that the following conditions
40  * are met:
41  * 1. Redistributions of source code must retain the above copyright
42  *    notice, this list of conditions and the following disclaimer.
43  * 2. Redistributions in binary form must reproduce the above copyright
44  *    notice, this list of conditions and the following disclaimer in the
45  *    documentation and/or other materials provided with the distribution.
46  * 3. All advertising materials mentioning features or use of this software
47  *    must display the following acknowledgement:
48  *      This product includes software developed by the NetBSD
49  *      Foundation, Inc. and its contributors.
50  * 4. Neither the name of The NetBSD Foundation nor the names of its
51  *    contributors may be used to endorse or promote products derived
52  *    from this software without specific prior written permission.
53  *
54  * THIS SOFTWARE IS PROVIDED BY AUTHOR AND CONTRIBUTORS ``AS IS'' AND ANY
55  * EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
56  * IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE
57  * ARE DISCLAIMED.  IN NO EVENT SHALL THE AUTHOR AND CONTRIBUTORS BE LIABLE
58  * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
59  * DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS
60  * OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION)
61  * HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT
62  * LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY
63  * OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
64  * SUCH DAMAGE.
65  */
66 
67 #include <sys/cdefs.h>
68 __KERNEL_RCSID(0, "$NetBSD: cpu.c,v 1.122 2018/06/23 10:30:22 jdolecek Exp $");
69 
70 #include "opt_ddb.h"
71 #include "opt_multiprocessor.h"
72 #include "opt_mpbios.h"		/* for MPDEBUG */
73 #include "opt_mtrr.h"
74 #include "opt_xen.h"
75 
76 #include "lapic.h"
77 #include "ioapic.h"
78 
79 #include <sys/param.h>
80 #include <sys/proc.h>
81 #include <sys/systm.h>
82 #include <sys/device.h>
83 #include <sys/kmem.h>
84 #include <sys/cpu.h>
85 #include <sys/cpufreq.h>
86 #include <sys/atomic.h>
87 #include <sys/reboot.h>
88 #include <sys/idle.h>
89 
90 #include <uvm/uvm.h>
91 
92 #include <machine/cpu.h>
93 #include <machine/cpufunc.h>
94 #include <machine/cpuvar.h>
95 #include <machine/pmap.h>
96 #include <machine/vmparam.h>
97 #include <machine/mpbiosvar.h>
98 #include <machine/pcb.h>
99 #include <machine/specialreg.h>
100 #include <machine/segments.h>
101 #include <machine/gdt.h>
102 #include <machine/mtrr.h>
103 #include <machine/pio.h>
104 
105 #include <x86/fpu.h>
106 
107 #include <xen/xen.h>
108 #include <xen/xen-public/vcpu.h>
109 #include <xen/vcpuvar.h>
110 
111 #if NLAPIC > 0
112 #include <machine/apicvar.h>
113 #include <machine/i82489reg.h>
114 #include <machine/i82489var.h>
115 #endif
116 
117 #include <dev/ic/mc146818reg.h>
118 #include <dev/isa/isareg.h>
119 
120 static int	cpu_match(device_t, cfdata_t, void *);
121 static void	cpu_attach(device_t, device_t, void *);
122 static void	cpu_defer(device_t);
123 static int	cpu_rescan(device_t, const char *, const int *);
124 static void	cpu_childdetached(device_t, device_t);
125 static int	vcpu_match(device_t, cfdata_t, void *);
126 static void	vcpu_attach(device_t, device_t, void *);
127 static void	cpu_attach_common(device_t, device_t, void *);
128 void		cpu_offline_md(void);
129 
130 struct cpu_softc {
131 	device_t sc_dev;		/* device tree glue */
132 	struct cpu_info *sc_info;	/* pointer to CPU info */
133 	bool sc_wasonline;
134 };
135 
136 int mp_cpu_start(struct cpu_info *, vaddr_t);
137 void mp_cpu_start_cleanup(struct cpu_info *);
138 const struct cpu_functions mp_cpu_funcs = { mp_cpu_start, NULL,
139 				      mp_cpu_start_cleanup };
140 
141 CFATTACH_DECL2_NEW(cpu, sizeof(struct cpu_softc),
142     cpu_match, cpu_attach, NULL, NULL, cpu_rescan, cpu_childdetached);
143 
144 CFATTACH_DECL_NEW(vcpu, sizeof(struct cpu_softc),
145     vcpu_match, vcpu_attach, NULL, NULL);
146 
147 /*
148  * Statically-allocated CPU info for the primary CPU (or the only
149  * CPU, on uniprocessors).  The CPU info list is initialized to
150  * point at it.
151  */
152 struct cpu_info cpu_info_primary __aligned(CACHE_LINE_SIZE) = {
153 	.ci_dev = 0,
154 	.ci_self = &cpu_info_primary,
155 	.ci_idepth = -1,
156 	.ci_curlwp = &lwp0,
157 	.ci_curldt = -1,
158 };
159 struct cpu_info phycpu_info_primary __aligned(CACHE_LINE_SIZE) = {
160 	.ci_dev = 0,
161 	.ci_self = &phycpu_info_primary,
162 };
163 
164 struct cpu_info *cpu_info_list = &cpu_info_primary;
165 struct cpu_info *phycpu_info_list = &phycpu_info_primary;
166 
167 uint32_t cpu_feature[7] __read_mostly; /* X86 CPUID feature bits
168 			  *	[0] basic features %edx
169 			  *	[1] basic features %ecx
170 			  *	[2] extended features %edx
171 			  *	[3] extended features %ecx
172 			  *	[4] VIA padlock features
173 			  *	[5] structured extended features cpuid.7:%ebx
174 			  *	[6] structured extended features cpuid.7:%ecx
175 			  */
176 
177 bool x86_mp_online;
178 paddr_t mp_trampoline_paddr = MP_TRAMPOLINE;
179 
180 #if defined(MULTIPROCESSOR)
181 void    	cpu_hatch(void *);
182 static void    	cpu_boot_secondary(struct cpu_info *ci);
183 static void    	cpu_start_secondary(struct cpu_info *ci);
184 #endif	/* MULTIPROCESSOR */
185 
186 static int
187 cpu_match(device_t parent, cfdata_t match, void *aux)
188 {
189 
190 	return 1;
191 }
192 
193 static void
194 cpu_attach(device_t parent, device_t self, void *aux)
195 {
196 	struct cpu_softc *sc = device_private(self);
197 	struct cpu_attach_args *caa = aux;
198 	struct cpu_info *ci;
199 	uintptr_t ptr;
200 	static int nphycpu = 0;
201 
202 	sc->sc_dev = self;
203 
204 	/*
205 	 * If we're an Application Processor, allocate a cpu_info
206 	 * If we're the first attached CPU use the primary cpu_info,
207 	 * otherwise allocate a new one
208 	 */
209 	aprint_naive("\n");
210 	aprint_normal("\n");
211 	if (nphycpu > 0) {
212 		struct cpu_info *tmp;
213 		ptr = (uintptr_t)kmem_zalloc(sizeof(*ci) + CACHE_LINE_SIZE - 1,
214 		    KM_SLEEP);
215 		ci = (struct cpu_info *)roundup2(ptr, CACHE_LINE_SIZE);
216 		ci->ci_curldt = -1;
217 
218 		tmp = phycpu_info_list;
219 		while (tmp->ci_next)
220 			tmp = tmp->ci_next;
221 
222 		tmp->ci_next = ci;
223 	} else {
224 		ci = &phycpu_info_primary;
225 	}
226 
227 	ci->ci_self = ci;
228 	sc->sc_info = ci;
229 
230 	ci->ci_dev = self;
231 	ci->ci_acpiid = caa->cpu_id;
232 	ci->ci_cpuid = caa->cpu_number;
233 	ci->ci_vcpu = NULL;
234 	ci->ci_index = nphycpu++;
235 
236 	if (!pmf_device_register(self, NULL, NULL))
237 		aprint_error_dev(self, "couldn't establish power handler\n");
238 
239 	(void)config_defer(self, cpu_defer);
240 }
241 
242 static void
243 cpu_defer(device_t self)
244 {
245 	cpu_rescan(self, NULL, NULL);
246 }
247 
248 static int
249 cpu_rescan(device_t self, const char *ifattr, const int *locators)
250 {
251 	struct cpu_softc *sc = device_private(self);
252 	struct cpufeature_attach_args cfaa;
253 	struct cpu_info *ci = sc->sc_info;
254 
255 	memset(&cfaa, 0, sizeof(cfaa));
256 	cfaa.ci = ci;
257 
258 	if (ifattr_match(ifattr, "cpufeaturebus")) {
259 
260 		if (ci->ci_frequency == NULL) {
261 			cfaa.name = "frequency";
262 			ci->ci_frequency = config_found_ia(self,
263 			    "cpufeaturebus", &cfaa, NULL);
264 		}
265 	}
266 
267 	return 0;
268 }
269 
270 static void
271 cpu_childdetached(device_t self, device_t child)
272 {
273 	struct cpu_softc *sc = device_private(self);
274 	struct cpu_info *ci = sc->sc_info;
275 
276 	if (ci->ci_frequency == child)
277 		ci->ci_frequency = NULL;
278 }
279 
280 static int
281 vcpu_match(device_t parent, cfdata_t match, void *aux)
282 {
283 	struct vcpu_attach_args *vcaa = aux;
284 	struct vcpu_runstate_info vcr;
285 	int error;
286 
287 	if (strcmp(vcaa->vcaa_name, match->cf_name) == 0) {
288 		error = HYPERVISOR_vcpu_op(VCPUOP_get_runstate_info,
289 		    vcaa->vcaa_caa.cpu_number, &vcr);
290 		switch (error) {
291 		case 0:
292 			return 1;
293 		case -ENOENT:
294 			return 0;
295 		default:
296 			panic("Unknown hypervisor error %d returned on vcpu runstate probe\n", error);
297 		}
298 	}
299 
300 	return 0;
301 }
302 
303 static void
304 vcpu_attach(device_t parent, device_t self, void *aux)
305 {
306 	struct vcpu_attach_args *vcaa = aux;
307 
308 	KASSERT(vcaa->vcaa_caa.cpu_func == NULL);
309 	vcaa->vcaa_caa.cpu_func = &mp_cpu_funcs;
310 	cpu_attach_common(parent, self, &vcaa->vcaa_caa);
311 
312 	if (!pmf_device_register(self, NULL, NULL))
313 		aprint_error_dev(self, "couldn't establish power handler\n");
314 }
315 
316 static int
317 vcpu_is_up(struct cpu_info *ci)
318 {
319 	KASSERT(ci != NULL);
320 	return HYPERVISOR_vcpu_op(VCPUOP_is_up, ci->ci_cpuid, NULL);
321 }
322 
323 static void
324 cpu_vm_init(struct cpu_info *ci)
325 {
326 	int ncolors = 2, i;
327 
328 	for (i = CAI_ICACHE; i <= CAI_L2CACHE; i++) {
329 		struct x86_cache_info *cai;
330 		int tcolors;
331 
332 		cai = &ci->ci_cinfo[i];
333 
334 		tcolors = atop(cai->cai_totalsize);
335 		switch (cai->cai_associativity) {
336 		case 0xff:
337 			tcolors = 1; /* fully associative */
338 			break;
339 		case 0:
340 		case 1:
341 			break;
342 		default:
343 			tcolors /= cai->cai_associativity;
344 		}
345 		ncolors = max(ncolors, tcolors);
346 	}
347 
348 	/*
349 	 * Knowing the size of the largest cache on this CPU, potentially
350 	 * re-color our pages.
351 	 */
352 	aprint_debug_dev(ci->ci_dev, "%d page colors\n", ncolors);
353 	uvm_page_recolor(ncolors);
354 	pmap_tlb_cpu_init(ci);
355 #ifndef __HAVE_DIRECT_MAP
356 	pmap_vpage_cpu_init(ci);
357 #endif
358 }
359 
360 static void
361 cpu_attach_common(device_t parent, device_t self, void *aux)
362 {
363 	struct cpu_softc *sc = device_private(self);
364 	struct cpu_attach_args *caa = aux;
365 	struct cpu_info *ci;
366 	uintptr_t ptr;
367 	int cpunum = caa->cpu_number;
368 	static bool again = false;
369 
370 	sc->sc_dev = self;
371 
372 	/*
373 	 * If we're an Application Processor, allocate a cpu_info
374 	 * structure, otherwise use the primary's.
375 	 */
376 	if (caa->cpu_role == CPU_ROLE_AP) {
377 		aprint_naive(": Application Processor\n");
378 		ptr = (uintptr_t)kmem_alloc(sizeof(*ci) + CACHE_LINE_SIZE - 1,
379 		    KM_SLEEP);
380 		ci = (struct cpu_info *)roundup2(ptr, CACHE_LINE_SIZE);
381 		memset(ci, 0, sizeof(*ci));
382 		cpu_init_tss(ci);
383 	} else {
384 		aprint_naive(": %s Processor\n",
385 		    caa->cpu_role == CPU_ROLE_SP ? "Single" : "Boot");
386 		ci = &cpu_info_primary;
387 	}
388 
389 	ci->ci_self = ci;
390 	sc->sc_info = ci;
391 	ci->ci_dev = self;
392 	ci->ci_cpuid = cpunum;
393 
394 	KASSERT(HYPERVISOR_shared_info != NULL);
395 	KASSERT(cpunum < XEN_LEGACY_MAX_VCPUS);
396 	ci->ci_vcpu = &HYPERVISOR_shared_info->vcpu_info[cpunum];
397 
398 	KASSERT(ci->ci_func == 0);
399 	ci->ci_func = caa->cpu_func;
400 	aprint_normal("\n");
401 
402 	/* Must be called before mi_cpu_attach(). */
403 	cpu_vm_init(ci);
404 
405 	if (caa->cpu_role == CPU_ROLE_AP) {
406 		int error;
407 
408 		error = mi_cpu_attach(ci);
409 
410 		KASSERT(ci->ci_data.cpu_idlelwp != NULL);
411 		if (error != 0) {
412 			aprint_error_dev(self,
413 			    "mi_cpu_attach failed with %d\n", error);
414 			return;
415 		}
416 
417 	} else {
418 		KASSERT(ci->ci_data.cpu_idlelwp != NULL);
419 	}
420 
421 	KASSERT(ci->ci_cpuid == ci->ci_index);
422 #ifdef __x86_64__
423 	/* No user PGD mapped for this CPU yet */
424 	ci->ci_xen_current_user_pgd = 0;
425 #endif
426 #if defined(__x86_64__) || defined(PAE)
427 	mutex_init(&ci->ci_kpm_mtx, MUTEX_DEFAULT, IPL_VM);
428 #endif
429 	pmap_reference(pmap_kernel());
430 	ci->ci_pmap = pmap_kernel();
431 	ci->ci_tlbstate = TLBSTATE_STALE;
432 
433 	/*
434 	 * Boot processor may not be attached first, but the below
435 	 * must be done to allow booting other processors.
436 	 */
437 	if (!again) {
438 		atomic_or_32(&ci->ci_flags, CPUF_PRESENT | CPUF_PRIMARY);
439 		/* Basic init. */
440 		cpu_intr_init(ci);
441 		cpu_get_tsc_freq(ci);
442 		cpu_init(ci);
443 		pmap_cpu_init_late(ci);
444 
445 		/* Every processor needs to init its own ipi h/w (similar to lapic) */
446 		xen_ipi_init();
447 
448 		/* Make sure DELAY() is initialized. */
449 		DELAY(1);
450 		again = true;
451 	}
452 
453 	/* further PCB init done later. */
454 
455 	switch (caa->cpu_role) {
456 	case CPU_ROLE_SP:
457 		atomic_or_32(&ci->ci_flags, CPUF_SP);
458 		cpu_identify(ci);
459 		x86_cpu_idle_init();
460 		break;
461 
462 	case CPU_ROLE_BP:
463 		atomic_or_32(&ci->ci_flags, CPUF_BSP);
464 		cpu_identify(ci);
465 		x86_cpu_idle_init();
466 		break;
467 
468 	case CPU_ROLE_AP:
469 		atomic_or_32(&ci->ci_flags, CPUF_AP);
470 
471 		/*
472 		 * report on an AP
473 		 */
474 
475 #if defined(MULTIPROCESSOR)
476 		/* interrupt handler stack */
477 		cpu_intr_init(ci);
478 
479 		/* Setup per-cpu memory for gdt */
480 		gdt_alloc_cpu(ci);
481 
482 		pmap_cpu_init_late(ci);
483 		cpu_start_secondary(ci);
484 
485 		if (ci->ci_flags & CPUF_PRESENT) {
486 			struct cpu_info *tmp;
487 
488 			cpu_identify(ci);
489 			tmp = cpu_info_list;
490 			while (tmp->ci_next)
491 				tmp = tmp->ci_next;
492 
493 			tmp->ci_next = ci;
494 		}
495 #else
496 		aprint_error_dev(ci->ci_dev, "not started\n");
497 #endif
498 		break;
499 
500 	default:
501 		panic("unknown processor type??\n");
502 	}
503 
504 #ifdef MPVERBOSE
505 	if (mp_verbose) {
506 		struct lwp *l = ci->ci_data.cpu_idlelwp;
507 		struct pcb *pcb = lwp_getpcb(l);
508 
509 		aprint_verbose_dev(self,
510 		    "idle lwp at %p, idle sp at 0x%p\n",
511 		    l,
512 #ifdef i386
513 		    (void *)pcb->pcb_esp
514 #else
515 		    (void *)pcb->pcb_rsp
516 #endif
517 		);
518 
519 	}
520 #endif /* MPVERBOSE */
521 }
522 
523 /*
524  * Initialize the processor appropriately.
525  */
526 
527 void
528 cpu_init(struct cpu_info *ci)
529 {
530 	uint32_t cr4 = 0;
531 
532 	/*
533 	 * If we have FXSAVE/FXRESTOR, use them.
534 	 */
535 	if (cpu_feature[0] & CPUID_FXSR) {
536 		cr4 |= CR4_OSFXSR;
537 
538 		/*
539 		 * If we have SSE/SSE2, enable XMM exceptions.
540 		 */
541 		if (cpu_feature[0] & (CPUID_SSE|CPUID_SSE2))
542 			cr4 |= CR4_OSXMMEXCPT;
543 	}
544 
545 	/* If xsave is supported, enable it */
546 	if (cpu_feature[1] & CPUID2_XSAVE && x86_fpu_save >= FPU_SAVE_XSAVE)
547 		cr4 |= CR4_OSXSAVE;
548 
549 	if (cr4) {
550 		cr4 |= rcr4();
551 		lcr4(cr4);
552 	}
553 
554 	if (x86_fpu_save >= FPU_SAVE_FXSAVE) {
555 		fpuinit_mxcsr_mask();
556 	}
557 
558 	/*
559 	 * Changing CR4 register may change cpuid values. For example, setting
560 	 * CR4_OSXSAVE sets CPUID2_OSXSAVE. The CPUID2_OSXSAVE is in
561 	 * ci_feat_val[1], so update it.
562 	 * XXX Other than ci_feat_val[1] might be changed.
563 	 */
564 	if (cpuid_level >= 1) {
565 		u_int descs[4];
566 
567 		x86_cpuid(1, descs);
568 		ci->ci_feat_val[1] = descs[2];
569 	}
570 
571 	/* If xsave is enabled, enable all fpu features */
572 	if (cr4 & CR4_OSXSAVE) {
573 		wrxcr(0, x86_xsave_features & XCR0_FPU);
574 	}
575 
576 	atomic_or_32(&ci->ci_flags, CPUF_RUNNING);
577 }
578 
579 
580 #ifdef MULTIPROCESSOR
581 
582 void
583 cpu_boot_secondary_processors(void)
584 {
585 	struct cpu_info *ci;
586 	u_long i;
587 	for (i = 0; i < maxcpus; i++) {
588 		ci = cpu_lookup(i);
589 		if (ci == NULL)
590 			continue;
591 		if (ci->ci_data.cpu_idlelwp == NULL)
592 			continue;
593 		if ((ci->ci_flags & CPUF_PRESENT) == 0)
594 			continue;
595 		if (ci->ci_flags & (CPUF_BSP|CPUF_SP|CPUF_PRIMARY))
596 			continue;
597 		cpu_boot_secondary(ci);
598 	}
599 
600 	x86_mp_online = true;
601 }
602 
603 static void
604 cpu_init_idle_lwp(struct cpu_info *ci)
605 {
606 	struct lwp *l = ci->ci_data.cpu_idlelwp;
607 	struct pcb *pcb = lwp_getpcb(l);
608 
609 	pcb->pcb_cr0 = rcr0();
610 }
611 
612 void
613 cpu_init_idle_lwps(void)
614 {
615 	struct cpu_info *ci;
616 	u_long i;
617 
618 	for (i = 0; i < maxcpus; i++) {
619 		ci = cpu_lookup(i);
620 		if (ci == NULL)
621 			continue;
622 		if (ci->ci_data.cpu_idlelwp == NULL)
623 			continue;
624 		if ((ci->ci_flags & CPUF_PRESENT) == 0)
625 			continue;
626 		cpu_init_idle_lwp(ci);
627 	}
628 }
629 
630 static void
631 cpu_start_secondary(struct cpu_info *ci)
632 {
633 	int i;
634 
635 	aprint_debug_dev(ci->ci_dev, "starting\n");
636 
637 	ci->ci_curlwp = ci->ci_data.cpu_idlelwp;
638 
639 	if (CPU_STARTUP(ci, (vaddr_t) cpu_hatch) != 0) {
640 		return;
641 	}
642 
643 	/*
644 	 * wait for it to become ready
645 	 */
646 	for (i = 100000; (!(ci->ci_flags & CPUF_PRESENT)) && i > 0; i--) {
647 		delay(10);
648 	}
649 	if ((ci->ci_flags & CPUF_PRESENT) == 0) {
650 		aprint_error_dev(ci->ci_dev, "failed to become ready\n");
651 #if defined(MPDEBUG) && defined(DDB)
652 		printf("dropping into debugger; continue from here to resume boot\n");
653 		Debugger();
654 #endif
655 	}
656 
657 	CPU_START_CLEANUP(ci);
658 }
659 
660 void
661 cpu_boot_secondary(struct cpu_info *ci)
662 {
663 	int i;
664 	atomic_or_32(&ci->ci_flags, CPUF_GO);
665 	for (i = 100000; (!(ci->ci_flags & CPUF_RUNNING)) && i > 0; i--) {
666 		delay(10);
667 	}
668 	if ((ci->ci_flags & CPUF_RUNNING) == 0) {
669 		aprint_error_dev(ci->ci_dev, "CPU failed to start\n");
670 #if defined(MPDEBUG) && defined(DDB)
671 		printf("dropping into debugger; continue from here to resume boot\n");
672 		Debugger();
673 #endif
674 	}
675 }
676 
677 /*
678  * APs end up here immediately after initialisation and VCPUOP_up in
679  * mp_cpu_start().
680  * At this point, we are running in the idle pcb/idle stack of the new
681  * CPU.  This function jumps to the idle loop and starts looking for
682  * work.
683  */
684 extern void x86_64_tls_switch(struct lwp *);
685 void
686 cpu_hatch(void *v)
687 {
688 	struct cpu_info *ci = (struct cpu_info *)v;
689 	struct pcb *pcb;
690 	int s, i;
691 
692 	/* Setup TLS and kernel GS/FS */
693 	cpu_init_msrs(ci, true);
694 	cpu_init_idt();
695 	gdt_init_cpu(ci);
696 
697 	cpu_probe(ci);
698 
699 	atomic_or_32(&ci->ci_flags, CPUF_PRESENT);
700 
701 	while ((ci->ci_flags & CPUF_GO) == 0) {
702 		/* Don't use delay, boot CPU may be patching the text. */
703 		for (i = 10000; i != 0; i--)
704 			x86_pause();
705 	}
706 
707 	/* Because the text may have been patched in x86_patch(). */
708 	x86_flush();
709 	tlbflushg();
710 
711 	KASSERT((ci->ci_flags & CPUF_RUNNING) == 0);
712 
713 	pcb = lwp_getpcb(curlwp);
714 	pcb->pcb_cr3 = pmap_pdirpa(pmap_kernel(), 0);
715 	pcb = lwp_getpcb(ci->ci_data.cpu_idlelwp);
716 
717 	xen_ipi_init();
718 
719 	xen_initclocks();
720 
721 #ifdef __x86_64__
722 	fpuinit(ci);
723 #endif
724 
725 	lldt(GSEL(GLDT_SEL, SEL_KPL));
726 
727 	cpu_init(ci);
728 	cpu_get_tsc_freq(ci);
729 
730 	s = splhigh();
731 	x86_enable_intr();
732 	splx(s);
733 
734 	aprint_debug_dev(ci->ci_dev, "running\n");
735 
736 	cpu_switchto(NULL, ci->ci_data.cpu_idlelwp, true);
737 
738 	idle_loop(NULL);
739 	KASSERT(false);
740 }
741 
742 #if defined(DDB)
743 
744 #include <ddb/db_output.h>
745 #include <machine/db_machdep.h>
746 
747 /*
748  * Dump CPU information from ddb.
749  */
750 void
751 cpu_debug_dump(void)
752 {
753 	struct cpu_info *ci;
754 	CPU_INFO_ITERATOR cii;
755 
756 	db_printf("addr		dev	id	flags	ipis	curlwp 		fpcurlwp\n");
757 	for (CPU_INFO_FOREACH(cii, ci)) {
758 		db_printf("%p	%s	%ld	%x	%x	%10p	%10p\n",
759 		    ci,
760 		    ci->ci_dev == NULL ? "BOOT" : device_xname(ci->ci_dev),
761 		    (long)ci->ci_cpuid,
762 		    ci->ci_flags, ci->ci_ipis,
763 		    ci->ci_curlwp,
764 		    ci->ci_fpcurlwp);
765 	}
766 }
767 #endif /* DDB */
768 
769 #endif /* MULTIPROCESSOR */
770 
771 extern void hypervisor_callback(void);
772 extern void failsafe_callback(void);
773 #ifdef __x86_64__
774 typedef void (vector)(void);
775 extern vector Xsyscall, Xsyscall32;
776 #endif
777 
778 /*
779  * Setup the "trampoline". On Xen, we setup nearly all cpu context
780  * outside a trampoline, so we prototype and call targetip like so:
781  * void targetip(struct cpu_info *);
782  */
783 
784 static void
785 gdt_prepframes(paddr_t *frames, vaddr_t base, uint32_t entries)
786 {
787 	int i;
788 	for (i = 0; i < entries; i++) {
789 		frames[i] = ((paddr_t)xpmap_ptetomach(
790 		    (pt_entry_t *)(base + (i << PAGE_SHIFT)))) >> PAGE_SHIFT;
791 
792 		/* Mark Read-only */
793 		pmap_pte_clearbits(kvtopte(base + (i << PAGE_SHIFT)),
794 		    PG_RW);
795 	}
796 }
797 
798 #ifdef __x86_64__
799 extern char *ldtstore;
800 
801 static void
802 xen_init_amd64_vcpuctxt(struct cpu_info *ci, struct vcpu_guest_context *initctx,
803     void targetrip(struct cpu_info *))
804 {
805 	/* page frames to point at GDT */
806 	extern int gdt_size;
807 	paddr_t frames[16];
808 	psize_t gdt_ents;
809 
810 	struct lwp *l;
811 	struct pcb *pcb;
812 
813 	volatile struct vcpu_info *vci;
814 
815 	KASSERT(ci != NULL);
816 	KASSERT(ci != &cpu_info_primary);
817 	KASSERT(initctx != NULL);
818 	KASSERT(targetrip != NULL);
819 
820 	memset(initctx, 0, sizeof(*initctx));
821 
822 	gdt_ents = roundup(gdt_size, PAGE_SIZE) >> PAGE_SHIFT;
823 	KASSERT(gdt_ents <= 16);
824 
825 	gdt_prepframes(frames, (vaddr_t)ci->ci_gdt, gdt_ents);
826 
827 	/* Initialise the vcpu context: We use idle_loop()'s pcb context. */
828 
829 	l = ci->ci_data.cpu_idlelwp;
830 
831 	KASSERT(l != NULL);
832 	pcb = lwp_getpcb(l);
833 	KASSERT(pcb != NULL);
834 
835 	/* resume with interrupts off */
836 	vci = ci->ci_vcpu;
837 	vci->evtchn_upcall_mask = 1;
838 	xen_mb();
839 
840 	/* resume in kernel-mode */
841 	initctx->flags = VGCF_in_kernel | VGCF_online;
842 
843 	/* Stack and entry points:
844 	 * We arrange for the stack frame for cpu_hatch() to
845 	 * appear as a callee frame of lwp_trampoline(). Being a
846 	 * leaf frame prevents trampling on any of the MD stack setup
847 	 * that x86/vm_machdep.c:cpu_lwp_fork() does for idle_loop()
848 	 */
849 
850 	initctx->user_regs.rdi = (uint64_t) ci; /* targetrip(ci); */
851 	initctx->user_regs.rip = (vaddr_t) targetrip;
852 
853 	initctx->user_regs.cs = GSEL(GCODE_SEL, SEL_KPL);
854 
855 	initctx->user_regs.rflags = pcb->pcb_flags;
856 	initctx->user_regs.rsp = pcb->pcb_rsp;
857 
858 	/* Data segments */
859 	initctx->user_regs.ss = GSEL(GDATA_SEL, SEL_KPL);
860 	initctx->user_regs.es = GSEL(GDATA_SEL, SEL_KPL);
861 	initctx->user_regs.ds = GSEL(GDATA_SEL, SEL_KPL);
862 
863 	/* GDT */
864 	memcpy(initctx->gdt_frames, frames, sizeof(frames));
865 	initctx->gdt_ents = gdt_ents;
866 
867 	/* LDT */
868 	initctx->ldt_base = (unsigned long)ldtstore;
869 	initctx->ldt_ents = LDT_SIZE >> 3;
870 
871 	/* Kernel context state */
872 	initctx->kernel_ss = GSEL(GDATA_SEL, SEL_KPL);
873 	initctx->kernel_sp = pcb->pcb_rsp0;
874 	initctx->ctrlreg[0] = pcb->pcb_cr0;
875 	initctx->ctrlreg[1] = 0; /* "resuming" from kernel - no User cr3. */
876 	initctx->ctrlreg[2] = (vaddr_t)targetrip;
877 	/*
878 	 * Use pmap_kernel() L4 PD directly, until we setup the
879 	 * per-cpu L4 PD in pmap_cpu_init_late()
880 	 */
881 	initctx->ctrlreg[3] = xen_pfn_to_cr3(x86_btop(xpmap_ptom(ci->ci_kpm_pdirpa)));
882 	initctx->ctrlreg[4] = CR4_PAE | CR4_OSFXSR | CR4_OSXMMEXCPT;
883 
884 	/* Xen callbacks */
885 	initctx->event_callback_eip = (unsigned long)hypervisor_callback;
886 	initctx->failsafe_callback_eip = (unsigned long)failsafe_callback;
887 	initctx->syscall_callback_eip = (unsigned long)Xsyscall;
888 
889 	return;
890 }
891 #else /* i386 */
892 extern union descriptor *ldtstore;
893 extern void Xsyscall(void);
894 
895 static void
896 xen_init_i386_vcpuctxt(struct cpu_info *ci, struct vcpu_guest_context *initctx,
897     void targeteip(struct cpu_info *))
898 {
899 	/* page frames to point at GDT */
900 	extern int gdt_size;
901 	paddr_t frames[16];
902 	psize_t gdt_ents;
903 
904 	struct lwp *l;
905 	struct pcb *pcb;
906 
907 	volatile struct vcpu_info *vci;
908 
909 	KASSERT(ci != NULL);
910 	KASSERT(ci != &cpu_info_primary);
911 	KASSERT(initctx != NULL);
912 	KASSERT(targeteip != NULL);
913 
914 	memset(initctx, 0, sizeof(*initctx));
915 
916 	gdt_ents = roundup(gdt_size, PAGE_SIZE) >> PAGE_SHIFT;
917 	KASSERT(gdt_ents <= 16);
918 
919 	gdt_prepframes(frames, (vaddr_t)ci->ci_gdt, gdt_ents);
920 
921 	/*
922 	 * Initialise the vcpu context:
923 	 * We use this cpu's idle_loop() pcb context.
924 	 */
925 
926 	l = ci->ci_data.cpu_idlelwp;
927 
928 	KASSERT(l != NULL);
929 	pcb = lwp_getpcb(l);
930 	KASSERT(pcb != NULL);
931 
932 	/* resume with interrupts off */
933 	vci = ci->ci_vcpu;
934 	vci->evtchn_upcall_mask = 1;
935 	xen_mb();
936 
937 	/* resume in kernel-mode */
938 	initctx->flags = VGCF_in_kernel | VGCF_online;
939 
940 	/* Stack frame setup for cpu_hatch():
941 	 * We arrange for the stack frame for cpu_hatch() to
942 	 * appear as a callee frame of lwp_trampoline(). Being a
943 	 * leaf frame prevents trampling on any of the MD stack setup
944 	 * that x86/vm_machdep.c:cpu_lwp_fork() does for idle_loop()
945 	 */
946 
947 	initctx->user_regs.esp = pcb->pcb_esp - 4; /* Leave word for
948 						      arg1 */
949 	{
950 		/* targeteip(ci); */
951 		uint32_t *arg = (uint32_t *)initctx->user_regs.esp;
952 		arg[1] = (uint32_t)ci; /* arg1 */
953 	}
954 
955 	initctx->user_regs.eip = (vaddr_t)targeteip;
956 	initctx->user_regs.cs = GSEL(GCODE_SEL, SEL_KPL);
957 	initctx->user_regs.eflags |= pcb->pcb_iopl;
958 
959 	/* Data segments */
960 	initctx->user_regs.ss = GSEL(GDATA_SEL, SEL_KPL);
961 	initctx->user_regs.es = GSEL(GDATA_SEL, SEL_KPL);
962 	initctx->user_regs.ds = GSEL(GDATA_SEL, SEL_KPL);
963 	initctx->user_regs.fs = GSEL(GDATA_SEL, SEL_KPL);
964 
965 	/* GDT */
966 	memcpy(initctx->gdt_frames, frames, sizeof(frames));
967 	initctx->gdt_ents = gdt_ents;
968 
969 	/* LDT */
970 	initctx->ldt_base = (unsigned long)ldtstore;
971 	initctx->ldt_ents = NLDT;
972 
973 	/* Kernel context state */
974 	initctx->kernel_ss = GSEL(GDATA_SEL, SEL_KPL);
975 	initctx->kernel_sp = pcb->pcb_esp0;
976 	initctx->ctrlreg[0] = pcb->pcb_cr0;
977 	initctx->ctrlreg[1] = 0; /* "resuming" from kernel - no User cr3. */
978 	initctx->ctrlreg[2] = (vaddr_t)targeteip;
979 #ifdef PAE
980 	initctx->ctrlreg[3] = xen_pfn_to_cr3(x86_btop(xpmap_ptom(ci->ci_pae_l3_pdirpa)));
981 #else
982 	initctx->ctrlreg[3] = xen_pfn_to_cr3(x86_btop(xpmap_ptom(pcb->pcb_cr3)));
983 #endif
984 	initctx->ctrlreg[4] = /* CR4_PAE | */CR4_OSFXSR | CR4_OSXMMEXCPT;
985 
986 	/* Xen callbacks */
987 	initctx->event_callback_eip = (unsigned long)hypervisor_callback;
988 	initctx->event_callback_cs = GSEL(GCODE_SEL, SEL_KPL);
989 	initctx->failsafe_callback_eip = (unsigned long)failsafe_callback;
990 	initctx->failsafe_callback_cs = GSEL(GCODE_SEL, SEL_KPL);
991 
992 	return;
993 }
994 #endif /* __x86_64__ */
995 
996 int
997 mp_cpu_start(struct cpu_info *ci, vaddr_t target)
998 {
999 	int hyperror;
1000 	struct vcpu_guest_context vcpuctx;
1001 
1002 	KASSERT(ci != NULL);
1003 	KASSERT(ci != &cpu_info_primary);
1004 	KASSERT(ci->ci_flags & CPUF_AP);
1005 
1006 #ifdef __x86_64__
1007 	xen_init_amd64_vcpuctxt(ci, &vcpuctx, (void (*)(struct cpu_info *))target);
1008 #else
1009 	xen_init_i386_vcpuctxt(ci, &vcpuctx, (void (*)(struct cpu_info *))target);
1010 #endif
1011 
1012 	/* Initialise the given vcpu to execute cpu_hatch(ci); */
1013 	if ((hyperror = HYPERVISOR_vcpu_op(VCPUOP_initialise, ci->ci_cpuid, &vcpuctx))) {
1014 		aprint_error(": context initialisation failed. errno = %d\n", hyperror);
1015 		return hyperror;
1016 	}
1017 
1018 	/* Start it up */
1019 
1020 	/* First bring it down */
1021 	if ((hyperror = HYPERVISOR_vcpu_op(VCPUOP_down, ci->ci_cpuid, NULL))) {
1022 		aprint_error(": VCPUOP_down hypervisor command failed. errno = %d\n", hyperror);
1023 		return hyperror;
1024 	}
1025 
1026 	if ((hyperror = HYPERVISOR_vcpu_op(VCPUOP_up, ci->ci_cpuid, NULL))) {
1027 		aprint_error(": VCPUOP_up hypervisor command failed. errno = %d\n", hyperror);
1028 		return hyperror;
1029 	}
1030 
1031 	if (!vcpu_is_up(ci)) {
1032 		aprint_error(": did not come up\n");
1033 		return -1;
1034 	}
1035 
1036 	return 0;
1037 }
1038 
1039 void
1040 mp_cpu_start_cleanup(struct cpu_info *ci)
1041 {
1042 	if (vcpu_is_up(ci)) {
1043 		aprint_debug_dev(ci->ci_dev, "is started.\n");
1044 	} else {
1045 		aprint_error_dev(ci->ci_dev, "did not start up.\n");
1046 	}
1047 }
1048 
1049 void
1050 cpu_init_msrs(struct cpu_info *ci, bool full)
1051 {
1052 #ifdef __x86_64__
1053 	if (full) {
1054 		HYPERVISOR_set_segment_base(SEGBASE_FS, 0);
1055 		HYPERVISOR_set_segment_base(SEGBASE_GS_KERNEL, (uint64_t)ci);
1056 		HYPERVISOR_set_segment_base(SEGBASE_GS_USER, 0);
1057 	}
1058 #endif
1059 
1060 	if (cpu_feature[2] & CPUID_NOX)
1061 		wrmsr(MSR_EFER, rdmsr(MSR_EFER) | EFER_NXE);
1062 }
1063 
1064 void
1065 cpu_offline_md(void)
1066 {
1067 	int s;
1068 
1069 	s = splhigh();
1070 	fpusave_cpu(true);
1071 	splx(s);
1072 }
1073 
1074 void
1075 cpu_get_tsc_freq(struct cpu_info *ci)
1076 {
1077 	uint32_t vcpu_tversion;
1078 	const volatile vcpu_time_info_t *tinfo = &ci->ci_vcpu->time;
1079 
1080 	vcpu_tversion = tinfo->version;
1081 	while (tinfo->version == vcpu_tversion); /* Wait for a time update. XXX: timeout ? */
1082 
1083 	uint64_t freq = 1000000000ULL << 32;
1084 	freq = freq / (uint64_t)tinfo->tsc_to_system_mul;
1085 	if (tinfo->tsc_shift < 0)
1086 		freq = freq << -tinfo->tsc_shift;
1087 	else
1088 		freq = freq >> tinfo->tsc_shift;
1089 	ci->ci_data.cpu_cc_freq = freq;
1090 }
1091 
1092 void
1093 x86_cpu_idle_xen(void)
1094 {
1095 	struct cpu_info *ci = curcpu();
1096 
1097 	KASSERT(ci->ci_ilevel == IPL_NONE);
1098 
1099 	x86_disable_intr();
1100 	if (!__predict_false(ci->ci_want_resched)) {
1101 		idle_block();
1102 	} else {
1103 		x86_enable_intr();
1104 	}
1105 }
1106 
1107 /*
1108  * Loads pmap for the current CPU.
1109  */
1110 void
1111 cpu_load_pmap(struct pmap *pmap, struct pmap *oldpmap)
1112 {
1113 	KASSERT(pmap != pmap_kernel());
1114 
1115 #if defined(__x86_64__) || defined(PAE)
1116 	struct cpu_info *ci = curcpu();
1117 	cpuid_t cid = cpu_index(ci);
1118 
1119 	mutex_enter(&ci->ci_kpm_mtx);
1120 	/* make new pmap visible to xen_kpm_sync() */
1121 	kcpuset_atomic_set(pmap->pm_xen_ptp_cpus, cid);
1122 #endif
1123 
1124 #ifdef i386
1125 #ifdef PAE
1126 	{
1127 		int i;
1128 		paddr_t l3_pd = xpmap_ptom_masked(ci->ci_pae_l3_pdirpa);
1129 		/* don't update the kernel L3 slot */
1130 		for (i = 0 ; i < PDP_SIZE - 1; i++) {
1131 			xpq_queue_pte_update(l3_pd + i * sizeof(pd_entry_t),
1132 			    xpmap_ptom(pmap->pm_pdirpa[i]) | PG_V);
1133 		}
1134 		tlbflush();
1135 	}
1136 #else /* PAE */
1137 	lcr3(pmap_pdirpa(pmap, 0));
1138 #endif /* PAE */
1139 #endif /* i386 */
1140 
1141 #ifdef __x86_64__
1142 	{
1143 		int i;
1144 		pd_entry_t *new_pgd;
1145 		paddr_t l4_pd_ma;
1146 
1147 		l4_pd_ma = xpmap_ptom_masked(ci->ci_kpm_pdirpa);
1148 
1149 		/*
1150 		 * Map user space address in kernel space and load
1151 		 * user cr3
1152 		 */
1153 		new_pgd = pmap->pm_pdir;
1154 		KASSERT(pmap == ci->ci_pmap);
1155 
1156 		/* Copy user pmap L4 PDEs (in user addr. range) to per-cpu L4 */
1157 		for (i = 0; i < PDIR_SLOT_PTE; i++) {
1158 			KASSERT(pmap != pmap_kernel() || new_pgd[i] == 0);
1159 			if (ci->ci_kpm_pdir[i] != new_pgd[i]) {
1160 				xpq_queue_pte_update(
1161 				    l4_pd_ma + i * sizeof(pd_entry_t),
1162 				    new_pgd[i]);
1163 			}
1164 		}
1165 
1166 		xen_set_user_pgd(pmap_pdirpa(pmap, 0));
1167 		ci->ci_xen_current_user_pgd = pmap_pdirpa(pmap, 0);
1168 
1169 		tlbflush();
1170 	}
1171 #endif /* __x86_64__ */
1172 
1173 #if defined(__x86_64__) || defined(PAE)
1174 	/* old pmap no longer visible to xen_kpm_sync() */
1175 	if (oldpmap != pmap_kernel()) {
1176 		kcpuset_atomic_clear(oldpmap->pm_xen_ptp_cpus, cid);
1177 	}
1178 	mutex_exit(&ci->ci_kpm_mtx);
1179 #endif
1180 }
1181 
1182 /*
1183  * pmap_cpu_init_late: perform late per-CPU initialization.
1184  *
1185  * Short note about percpu PDIR pages. Both the PAE and __x86_64__ architectures
1186  * have per-cpu PDIR tables, for two different reasons:
1187  *  - on PAE, this is to get around Xen's pagetable setup constraints (multiple
1188  *    L3[3]s cannot point to the same L2 - Xen will refuse to pin a table set up
1189  *    this way).
1190  *  - on __x86_64__, this is for multiple CPUs to map in different user pmaps
1191  *    (see cpu_load_pmap()).
1192  *
1193  * What this means for us is that the PDIR of the pmap_kernel() is considered
1194  * to be a canonical "SHADOW" PDIR with the following properties:
1195  *  - its recursive mapping points to itself
1196  *  - per-cpu recursive mappings point to themselves on __x86_64__
1197  *  - per-cpu L4 pages' kernel entries are expected to be in sync with
1198  *    the shadow
1199  */
1200 
1201 void
1202 pmap_cpu_init_late(struct cpu_info *ci)
1203 {
1204 #if defined(PAE) || defined(__x86_64__)
1205 	/*
1206 	 * The BP has already its own PD page allocated during early
1207 	 * MD startup.
1208 	 */
1209 
1210 #if defined(__x86_64__)
1211 	/* Setup per-cpu normal_pdes */
1212 	int i;
1213 	extern pd_entry_t * const normal_pdes[];
1214 	for (i = 0;i < PTP_LEVELS - 1;i++) {
1215 		ci->ci_normal_pdes[i] = normal_pdes[i];
1216 	}
1217 #endif /* __x86_64__ */
1218 
1219 	if (ci == &cpu_info_primary)
1220 		return;
1221 
1222 	KASSERT(ci != NULL);
1223 
1224 #if defined(PAE)
1225 	cpu_alloc_l3_page(ci);
1226 	KASSERT(ci->ci_pae_l3_pdirpa != 0);
1227 
1228 	/* Initialise L2 entries 0 - 2: Point them to pmap_kernel() */
1229 	int i;
1230 	for (i = 0 ; i < PDP_SIZE - 1; i++) {
1231 		ci->ci_pae_l3_pdir[i] =
1232 		    xpmap_ptom_masked(pmap_kernel()->pm_pdirpa[i]) | PG_V;
1233 	}
1234 #endif /* PAE */
1235 
1236 	ci->ci_kpm_pdir = (pd_entry_t *)uvm_km_alloc(kernel_map, PAGE_SIZE, 0,
1237 	    UVM_KMF_WIRED | UVM_KMF_ZERO | UVM_KMF_NOWAIT);
1238 
1239 	if (ci->ci_kpm_pdir == NULL) {
1240 		panic("%s: failed to allocate L4 per-cpu PD for CPU %d\n",
1241 		    __func__, cpu_index(ci));
1242 	}
1243 	ci->ci_kpm_pdirpa = vtophys((vaddr_t)ci->ci_kpm_pdir);
1244 	KASSERT(ci->ci_kpm_pdirpa != 0);
1245 
1246 #if defined(__x86_64__)
1247 	extern pt_entry_t xpmap_pg_nx;
1248 
1249 	/* Copy over the pmap_kernel() shadow L4 entries */
1250 	memcpy(ci->ci_kpm_pdir, pmap_kernel()->pm_pdir, PAGE_SIZE);
1251 
1252 	/* Recursive kernel mapping */
1253 	ci->ci_kpm_pdir[PDIR_SLOT_PTE] = xpmap_ptom_masked(ci->ci_kpm_pdirpa)
1254 	    | PG_V | xpmap_pg_nx;
1255 #elif defined(PAE)
1256 	/* Copy over the pmap_kernel() shadow L2 entries */
1257 	memcpy(ci->ci_kpm_pdir, pmap_kernel()->pm_pdir + PDIR_SLOT_KERN,
1258 	    nkptp[PTP_LEVELS - 1] * sizeof(pd_entry_t));
1259 #endif
1260 
1261 	/* Xen wants a RO pdir. */
1262 	pmap_protect(pmap_kernel(), (vaddr_t)ci->ci_kpm_pdir,
1263 	    (vaddr_t)ci->ci_kpm_pdir + PAGE_SIZE, VM_PROT_READ);
1264 	pmap_update(pmap_kernel());
1265 #if defined(PAE)
1266 	/*
1267 	 * Initialize L3 entry 3. This mapping is shared across all pmaps and is
1268 	 * static, ie: loading a new pmap will not update this entry.
1269 	 */
1270 	ci->ci_pae_l3_pdir[3] = xpmap_ptom_masked(ci->ci_kpm_pdirpa) | PG_V;
1271 
1272 	/* Xen wants a RO L3. */
1273 	pmap_protect(pmap_kernel(), (vaddr_t)ci->ci_pae_l3_pdir,
1274 	    (vaddr_t)ci->ci_pae_l3_pdir + PAGE_SIZE, VM_PROT_READ);
1275 	pmap_update(pmap_kernel());
1276 
1277 	xpq_queue_pin_l3_table(xpmap_ptom_masked(ci->ci_pae_l3_pdirpa));
1278 
1279 #elif defined(__x86_64__)
1280 	xpq_queue_pin_l4_table(xpmap_ptom_masked(ci->ci_kpm_pdirpa));
1281 #endif /* PAE , __x86_64__ */
1282 #endif /* defined(PAE) || defined(__x86_64__) */
1283 }
1284 
1285 /*
1286  * Notify all other cpus to halt.
1287  */
1288 
1289 void
1290 cpu_broadcast_halt(void)
1291 {
1292 	xen_broadcast_ipi(XEN_IPI_HALT);
1293 }
1294 
1295 /*
1296  * Send a dummy ipi to a cpu.
1297  */
1298 
1299 void
1300 cpu_kick(struct cpu_info *ci)
1301 {
1302 	(void)xen_send_ipi(ci, XEN_IPI_KICK);
1303 }
1304