xref: /illumos-gate/usr/src/uts/i86pc/os/mp_pc.c (revision 86ef0a63)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 /*
22  * Copyright (c) 2007, 2010, Oracle and/or its affiliates. All rights reserved.
23  */
24 /*
25  * Copyright (c) 2010, Intel Corporation.
26  * All rights reserved.
27  */
28 /*
29  * Copyright 2019 Joyent, Inc.
30  */
31 
32 /*
33  * Welcome to the world of the "real mode platter".
34  * See also startup.c, mpcore.s and apic.c for related routines.
35  */
36 
37 #include <sys/types.h>
38 #include <sys/systm.h>
39 #include <sys/cpuvar.h>
40 #include <sys/cpu_module.h>
41 #include <sys/kmem.h>
42 #include <sys/archsystm.h>
43 #include <sys/machsystm.h>
44 #include <sys/controlregs.h>
45 #include <sys/x86_archext.h>
46 #include <sys/smp_impldefs.h>
47 #include <sys/sysmacros.h>
48 #include <sys/mach_mmu.h>
49 #include <sys/promif.h>
50 #include <sys/cpu.h>
51 #include <sys/cpu_event.h>
52 #include <sys/sunndi.h>
53 #include <sys/fs/dv_node.h>
54 #include <vm/hat_i86.h>
55 #include <vm/as.h>
56 
57 extern cpuset_t cpu_ready_set;
58 
59 extern int  mp_start_cpu_common(cpu_t *cp, boolean_t boot);
60 extern void real_mode_start_cpu(void);
61 extern void real_mode_start_cpu_end(void);
62 extern void real_mode_stop_cpu_stage1(void);
63 extern void real_mode_stop_cpu_stage1_end(void);
64 extern void real_mode_stop_cpu_stage2(void);
65 extern void real_mode_stop_cpu_stage2_end(void);
66 
67 void rmp_gdt_init(rm_platter_t *);
68 
69 /*
70  * Fill up the real mode platter to make it easy for real mode code to
71  * kick it off. This area should really be one passed by boot to kernel
72  * and guaranteed to be below 1MB and aligned to 16 bytes. Should also
73  * have identical physical and virtual address in paged mode.
74  */
75 static ushort_t *warm_reset_vector = NULL;
76 
77 int
mach_cpucontext_init(void)78 mach_cpucontext_init(void)
79 {
80 	ushort_t *vec;
81 	ulong_t addr;
82 	struct rm_platter *rm = (struct rm_platter *)rm_platter_va;
83 
84 	if (!(vec = (ushort_t *)psm_map_phys(WARM_RESET_VECTOR,
85 	    sizeof (vec), PROT_READ | PROT_WRITE)))
86 		return (-1);
87 
88 	/*
89 	 * setup secondary cpu bios boot up vector
90 	 * Write page offset to 0x467 and page frame number to 0x469.
91 	 */
92 	addr = (ulong_t)((caddr_t)rm->rm_code - (caddr_t)rm) + rm_platter_pa;
93 	vec[0] = (ushort_t)(addr & PAGEOFFSET);
94 	vec[1] = (ushort_t)((addr & (0xfffff & PAGEMASK)) >> 4);
95 	warm_reset_vector = vec;
96 
97 	/* Map real mode platter into kas so kernel can access it. */
98 	hat_devload(kas.a_hat,
99 	    (caddr_t)(uintptr_t)rm_platter_pa, MMU_PAGESIZE,
100 	    btop(rm_platter_pa), PROT_READ | PROT_WRITE | PROT_EXEC,
101 	    HAT_LOAD_NOCONSIST);
102 
103 	/* Copy CPU startup code to rm_platter if it's still during boot. */
104 	if (!plat_dr_enabled()) {
105 		ASSERT((size_t)real_mode_start_cpu_end -
106 		    (size_t)real_mode_start_cpu <= RM_PLATTER_CODE_SIZE);
107 		bcopy((caddr_t)real_mode_start_cpu, (caddr_t)rm->rm_code,
108 		    (size_t)real_mode_start_cpu_end -
109 		    (size_t)real_mode_start_cpu);
110 	}
111 
112 	return (0);
113 }
114 
115 void
mach_cpucontext_fini(void)116 mach_cpucontext_fini(void)
117 {
118 	if (warm_reset_vector)
119 		psm_unmap_phys((caddr_t)warm_reset_vector,
120 		    sizeof (warm_reset_vector));
121 	hat_unload(kas.a_hat, (caddr_t)(uintptr_t)rm_platter_pa, MMU_PAGESIZE,
122 	    HAT_UNLOAD);
123 }
124 
125 extern void *long_mode_64(void);
126 
127 /*ARGSUSED*/
128 void
rmp_gdt_init(rm_platter_t * rm)129 rmp_gdt_init(rm_platter_t *rm)
130 {
131 
132 	/* Use the kas address space for the CPU startup thread. */
133 	if (mmu_ptob(kas.a_hat->hat_htable->ht_pfn) > 0xffffffffUL) {
134 		panic("Cannot initialize CPUs; kernel's 64-bit page tables\n"
135 		    "located above 4G in physical memory (@ 0x%lx)",
136 		    mmu_ptob(kas.a_hat->hat_htable->ht_pfn));
137 	}
138 
139 	/*
140 	 * Setup pseudo-descriptors for temporary GDT and IDT for use ONLY
141 	 * by code in real_mode_start_cpu():
142 	 *
143 	 * GDT[0]:  NULL selector
144 	 * GDT[1]:  64-bit CS: Long = 1, Present = 1, bits 12, 11 = 1
145 	 *
146 	 * Clear the IDT as interrupts will be off and a limit of 0 will cause
147 	 * the CPU to triple fault and reset on an NMI, seemingly as reasonable
148 	 * a course of action as any other, though it may cause the entire
149 	 * platform to reset in some cases...
150 	 */
151 	rm->rm_temp_gdt[0] = 0;
152 	rm->rm_temp_gdt[TEMPGDT_KCODE64] = 0x20980000000000ULL;
153 
154 	rm->rm_temp_gdt_lim = (ushort_t)(sizeof (rm->rm_temp_gdt) - 1);
155 	rm->rm_temp_gdt_base = rm_platter_pa +
156 	    (uint32_t)offsetof(rm_platter_t, rm_temp_gdt);
157 	rm->rm_temp_idt_lim = 0;
158 	rm->rm_temp_idt_base = 0;
159 
160 	/*
161 	 * Since the CPU needs to jump to protected mode using an identity
162 	 * mapped address, we need to calculate it here.
163 	 */
164 	rm->rm_longmode64_addr = rm_platter_pa +
165 	    (uint32_t)((uintptr_t)long_mode_64 -
166 	    (uintptr_t)real_mode_start_cpu);
167 }
168 
169 static void *
mach_cpucontext_alloc_tables(struct cpu * cp)170 mach_cpucontext_alloc_tables(struct cpu *cp)
171 {
172 	tss_t *ntss;
173 	struct cpu_tables *ct;
174 	size_t ctsize;
175 
176 	/*
177 	 * Allocate space for stack, tss, gdt and idt. We round the size
178 	 * allotted for cpu_tables up, so that the TSS is on a unique page.
179 	 * This is more efficient when running in virtual machines.
180 	 */
181 	ctsize = P2ROUNDUP(sizeof (*ct), PAGESIZE);
182 	ct = kmem_zalloc(ctsize, KM_SLEEP);
183 	if ((uintptr_t)ct & PAGEOFFSET)
184 		panic("mach_cpucontext_alloc_tables: cpu%d misaligned tables",
185 		    cp->cpu_id);
186 
187 	ntss = cp->cpu_tss = &ct->ct_tss;
188 
189 	uintptr_t va;
190 	size_t len;
191 
192 	/*
193 	 * #DF (double fault).
194 	 */
195 	ntss->tss_ist1 = (uintptr_t)&ct->ct_stack1[sizeof (ct->ct_stack1)];
196 
197 	/*
198 	 * #NM (non-maskable interrupt)
199 	 */
200 	ntss->tss_ist2 = (uintptr_t)&ct->ct_stack2[sizeof (ct->ct_stack2)];
201 
202 	/*
203 	 * #MC (machine check exception / hardware error)
204 	 */
205 	ntss->tss_ist3 = (uintptr_t)&ct->ct_stack3[sizeof (ct->ct_stack3)];
206 
207 	/*
208 	 * #DB, #BP debug interrupts and KDI/kmdb
209 	 */
210 	ntss->tss_ist4 = (uintptr_t)&cp->cpu_m.mcpu_kpti_dbg.kf_tr_rsp;
211 
212 	if (kpti_enable == 1) {
213 		/*
214 		 * #GP, #PF, #SS fault interrupts
215 		 */
216 		ntss->tss_ist5 = (uintptr_t)&cp->cpu_m.mcpu_kpti_flt.kf_tr_rsp;
217 
218 		/*
219 		 * Used by all other interrupts
220 		 */
221 		ntss->tss_ist6 = (uint64_t)&cp->cpu_m.mcpu_kpti.kf_tr_rsp;
222 
223 		/*
224 		 * On AMD64 we need to make sure that all of the pages of the
225 		 * struct cpu_tables are punched through onto the user CPU for
226 		 * kpti.
227 		 *
228 		 * The final page will always be the TSS, so treat that
229 		 * separately.
230 		 */
231 		for (va = (uintptr_t)ct, len = ctsize - MMU_PAGESIZE;
232 		    len >= MMU_PAGESIZE;
233 		    len -= MMU_PAGESIZE, va += MMU_PAGESIZE) {
234 			/* The doublefault stack must be RW */
235 			hati_cpu_punchin(cp, va, PROT_READ | PROT_WRITE);
236 		}
237 		ASSERT3U((uintptr_t)ntss, ==, va);
238 		hati_cpu_punchin(cp, (uintptr_t)ntss, PROT_READ);
239 	}
240 
241 
242 	/*
243 	 * Set I/O bit map offset equal to size of TSS segment limit
244 	 * for no I/O permission map. This will cause all user I/O
245 	 * instructions to generate #gp fault.
246 	 */
247 	ntss->tss_bitmapbase = sizeof (*ntss);
248 
249 	/*
250 	 * Setup kernel tss.
251 	 */
252 	set_syssegd((system_desc_t *)&cp->cpu_gdt[GDT_KTSS], cp->cpu_tss,
253 	    sizeof (*cp->cpu_tss) - 1, SDT_SYSTSS, SEL_KPL);
254 
255 	return (ct);
256 }
257 
258 void *
mach_cpucontext_xalloc(struct cpu * cp,int optype)259 mach_cpucontext_xalloc(struct cpu *cp, int optype)
260 {
261 	size_t len;
262 	struct cpu_tables *ct;
263 	rm_platter_t *rm = (rm_platter_t *)rm_platter_va;
264 	static int cpu_halt_code_ready;
265 
266 	if (optype == MACH_CPUCONTEXT_OP_STOP) {
267 		ASSERT(plat_dr_enabled());
268 
269 		/*
270 		 * The WARM_RESET_VECTOR has a limitation that the physical
271 		 * address written to it must be page-aligned. To work around
272 		 * this limitation, the CPU stop code has been splitted into
273 		 * two stages.
274 		 * The stage 2 code, which implements the real logic to halt
275 		 * CPUs, is copied to the rm_cpu_halt_code field in the real
276 		 * mode platter. The stage 1 code, which simply jumps to the
277 		 * stage 2 code in the rm_cpu_halt_code field, is copied to
278 		 * rm_code field in the real mode platter and it may be
279 		 * overwritten after the CPU has been stopped.
280 		 */
281 		if (!cpu_halt_code_ready) {
282 			/*
283 			 * The rm_cpu_halt_code field in the real mode platter
284 			 * is used by the CPU stop code only. So only copy the
285 			 * CPU stop stage 2 code into the rm_cpu_halt_code
286 			 * field on the first call.
287 			 */
288 			len = (size_t)real_mode_stop_cpu_stage2_end -
289 			    (size_t)real_mode_stop_cpu_stage2;
290 			ASSERT(len <= RM_PLATTER_CPU_HALT_CODE_SIZE);
291 			bcopy((caddr_t)real_mode_stop_cpu_stage2,
292 			    (caddr_t)rm->rm_cpu_halt_code, len);
293 			cpu_halt_code_ready = 1;
294 		}
295 
296 		/*
297 		 * The rm_code field in the real mode platter is shared by
298 		 * the CPU start, CPU stop, CPR and fast reboot code. So copy
299 		 * the CPU stop stage 1 code into the rm_code field every time.
300 		 */
301 		len = (size_t)real_mode_stop_cpu_stage1_end -
302 		    (size_t)real_mode_stop_cpu_stage1;
303 		ASSERT(len <= RM_PLATTER_CODE_SIZE);
304 		bcopy((caddr_t)real_mode_stop_cpu_stage1,
305 		    (caddr_t)rm->rm_code, len);
306 		rm->rm_cpu_halted = 0;
307 
308 		return (cp->cpu_m.mcpu_mach_ctx_ptr);
309 	} else if (optype != MACH_CPUCONTEXT_OP_START) {
310 		return (NULL);
311 	}
312 
313 	/*
314 	 * Only need to allocate tables when starting CPU.
315 	 * Tables allocated when starting CPU will be reused when stopping CPU.
316 	 */
317 	ct = mach_cpucontext_alloc_tables(cp);
318 	if (ct == NULL) {
319 		return (NULL);
320 	}
321 
322 	/* Copy CPU startup code to rm_platter for CPU hot-add operations. */
323 	if (plat_dr_enabled()) {
324 		bcopy((caddr_t)real_mode_start_cpu, (caddr_t)rm->rm_code,
325 		    (size_t)real_mode_start_cpu_end -
326 		    (size_t)real_mode_start_cpu);
327 	}
328 
329 	/*
330 	 * Now copy all that we've set up onto the real mode platter
331 	 * for the real mode code to digest as part of starting the cpu.
332 	 */
333 	rm->rm_idt_base = cp->cpu_idt;
334 	rm->rm_idt_lim = sizeof (*cp->cpu_idt) * NIDT - 1;
335 	rm->rm_gdt_base = cp->cpu_gdt;
336 	rm->rm_gdt_lim = sizeof (*cp->cpu_gdt) * NGDT - 1;
337 
338 	/*
339 	 * CPU needs to access kernel address space after powering on.
340 	 */
341 	rm->rm_pdbr = MAKECR3(kas.a_hat->hat_htable->ht_pfn, PCID_NONE);
342 	rm->rm_cpu = cp->cpu_id;
343 
344 	/*
345 	 * We need to mask off any bits set on our boot CPU that can't apply
346 	 * while the subject CPU is initializing.  If appropriate, they are
347 	 * enabled later on.
348 	 */
349 	rm->rm_cr4 = getcr4();
350 	rm->rm_cr4 &= ~(CR4_MCE | CR4_PCE | CR4_PCIDE);
351 
352 	rmp_gdt_init(rm);
353 
354 	return (ct);
355 }
356 
357 void
mach_cpucontext_xfree(struct cpu * cp,void * arg,int err,int optype)358 mach_cpucontext_xfree(struct cpu *cp, void *arg, int err, int optype)
359 {
360 	struct cpu_tables *ct = arg;
361 
362 	ASSERT(&ct->ct_tss == cp->cpu_tss);
363 	if (optype == MACH_CPUCONTEXT_OP_START) {
364 		switch (err) {
365 		case 0:
366 			/*
367 			 * Save pointer for reuse when stopping CPU.
368 			 */
369 			cp->cpu_m.mcpu_mach_ctx_ptr = arg;
370 			break;
371 		case ETIMEDOUT:
372 			/*
373 			 * The processor was poked, but failed to start before
374 			 * we gave up waiting for it.  In case it starts later,
375 			 * don't free anything.
376 			 */
377 			cp->cpu_m.mcpu_mach_ctx_ptr = arg;
378 			break;
379 		default:
380 			/*
381 			 * Some other, passive, error occurred.
382 			 */
383 			kmem_free(ct, P2ROUNDUP(sizeof (*ct), PAGESIZE));
384 			cp->cpu_tss = NULL;
385 			break;
386 		}
387 	} else if (optype == MACH_CPUCONTEXT_OP_STOP) {
388 		switch (err) {
389 		case 0:
390 			/*
391 			 * Free resources allocated when starting CPU.
392 			 */
393 			kmem_free(ct, P2ROUNDUP(sizeof (*ct), PAGESIZE));
394 			cp->cpu_tss = NULL;
395 			cp->cpu_m.mcpu_mach_ctx_ptr = NULL;
396 			break;
397 		default:
398 			/*
399 			 * Don't touch table pointer in case of failure.
400 			 */
401 			break;
402 		}
403 	} else {
404 		ASSERT(0);
405 	}
406 }
407 
408 void *
mach_cpucontext_alloc(struct cpu * cp)409 mach_cpucontext_alloc(struct cpu *cp)
410 {
411 	return (mach_cpucontext_xalloc(cp, MACH_CPUCONTEXT_OP_START));
412 }
413 
414 void
mach_cpucontext_free(struct cpu * cp,void * arg,int err)415 mach_cpucontext_free(struct cpu *cp, void *arg, int err)
416 {
417 	mach_cpucontext_xfree(cp, arg, err, MACH_CPUCONTEXT_OP_START);
418 }
419 
420 /*
421  * "Enter monitor."  Called via cross-call from stop_other_cpus().
422  */
423 int
mach_cpu_halt(xc_arg_t arg1,xc_arg_t arg2 __unused,xc_arg_t arg3 __unused)424 mach_cpu_halt(xc_arg_t arg1, xc_arg_t arg2 __unused, xc_arg_t arg3 __unused)
425 {
426 	char *msg = (char *)arg1;
427 
428 	if (msg)
429 		prom_printf("%s\n", msg);
430 
431 	/*CONSTANTCONDITION*/
432 	while (1)
433 		;
434 	return (0);
435 }
436 
437 void
mach_cpu_idle(void)438 mach_cpu_idle(void)
439 {
440 	x86_md_clear();
441 	i86_halt();
442 }
443 
444 void
mach_cpu_pause(volatile char * safe)445 mach_cpu_pause(volatile char *safe)
446 {
447 	/*
448 	 * This cpu is now safe.
449 	 */
450 	*safe = PAUSE_WAIT;
451 	membar_enter(); /* make sure stores are flushed */
452 
453 	/*
454 	 * Now we wait.  When we are allowed to continue, safe
455 	 * will be set to PAUSE_IDLE.
456 	 */
457 	while (*safe != PAUSE_IDLE)
458 		SMT_PAUSE();
459 }
460 
461 /*
462  * Power on the target CPU.
463  */
464 int
mp_cpu_poweron(struct cpu * cp)465 mp_cpu_poweron(struct cpu *cp)
466 {
467 	int error;
468 	cpuset_t tempset;
469 	processorid_t cpuid;
470 
471 	ASSERT(cp != NULL);
472 	cpuid = cp->cpu_id;
473 	if (use_mp == 0 || plat_dr_support_cpu() == 0) {
474 		return (ENOTSUP);
475 	} else if (cpuid < 0 || cpuid >= max_ncpus) {
476 		return (EINVAL);
477 	}
478 
479 	/*
480 	 * The currrent x86 implementaiton of mp_cpu_configure() and
481 	 * mp_cpu_poweron() have a limitation that mp_cpu_poweron() could only
482 	 * be called once after calling mp_cpu_configure() for a specific CPU.
483 	 * It's because mp_cpu_poweron() will destroy data structure created
484 	 * by mp_cpu_configure(). So reject the request if the CPU has already
485 	 * been powered on once after calling mp_cpu_configure().
486 	 * This limitaiton only affects the p_online syscall and the DR driver
487 	 * won't be affected because the DR driver always invoke public CPU
488 	 * management interfaces in the predefined order:
489 	 * cpu_configure()->cpu_poweron()...->cpu_poweroff()->cpu_unconfigure()
490 	 */
491 	if (cpuid_checkpass(cp, 4) || cp->cpu_thread == cp->cpu_idle_thread) {
492 		return (ENOTSUP);
493 	}
494 
495 	/*
496 	 * Check if there's at least a Mbyte of kmem available
497 	 * before attempting to start the cpu.
498 	 */
499 	if (kmem_avail() < 1024 * 1024) {
500 		/*
501 		 * Kick off a reap in case that helps us with
502 		 * later attempts ..
503 		 */
504 		kmem_reap();
505 		return (ENOMEM);
506 	}
507 
508 	affinity_set(CPU->cpu_id);
509 
510 	/*
511 	 * Start the target CPU. No need to call mach_cpucontext_fini()
512 	 * if mach_cpucontext_init() fails.
513 	 */
514 	if ((error = mach_cpucontext_init()) == 0) {
515 		error = mp_start_cpu_common(cp, B_FALSE);
516 		mach_cpucontext_fini();
517 	}
518 	if (error != 0) {
519 		affinity_clear();
520 		return (error);
521 	}
522 
523 	/* Wait for the target cpu to reach READY state. */
524 	tempset = cpu_ready_set;
525 	while (!CPU_IN_SET(tempset, cpuid)) {
526 		delay(1);
527 		tempset = *((volatile cpuset_t *)&cpu_ready_set);
528 	}
529 
530 	/* Mark the target CPU as available for mp operation. */
531 	CPUSET_ATOMIC_ADD(mp_cpus, cpuid);
532 
533 	/* Free the space allocated to hold the microcode file */
534 	ucode_cleanup();
535 
536 	affinity_clear();
537 
538 	return (0);
539 }
540 
541 #define	MP_CPU_DETACH_MAX_TRIES		5
542 #define	MP_CPU_DETACH_DELAY		100
543 
544 static int
mp_cpu_detach_driver(dev_info_t * dip)545 mp_cpu_detach_driver(dev_info_t *dip)
546 {
547 	int i;
548 	int rv = EBUSY;
549 	dev_info_t *pdip;
550 
551 	pdip = ddi_get_parent(dip);
552 	ASSERT(pdip != NULL);
553 	/*
554 	 * Check if caller holds pdip busy - can cause deadlocks in
555 	 * e_ddi_branch_unconfigure(), which calls devfs_clean().
556 	 */
557 	if (DEVI_BUSY_OWNED(pdip)) {
558 		return (EDEADLOCK);
559 	}
560 
561 	for (i = 0; i < MP_CPU_DETACH_MAX_TRIES; i++) {
562 		if (e_ddi_branch_unconfigure(dip, NULL, 0) == 0) {
563 			rv = 0;
564 			break;
565 		}
566 		DELAY(MP_CPU_DETACH_DELAY);
567 	}
568 
569 	return (rv);
570 }
571 
572 /*
573  * Power off the target CPU.
574  * Note: cpu_lock will be released and then reacquired.
575  */
576 int
mp_cpu_poweroff(struct cpu * cp)577 mp_cpu_poweroff(struct cpu *cp)
578 {
579 	int rv = 0;
580 	void *ctx;
581 	dev_info_t *dip = NULL;
582 	rm_platter_t *rm = (rm_platter_t *)rm_platter_va;
583 	extern void cpupm_start(cpu_t *);
584 	extern void cpupm_stop(cpu_t *);
585 
586 	ASSERT(cp != NULL);
587 	ASSERT((cp->cpu_flags & CPU_OFFLINE) != 0);
588 	ASSERT((cp->cpu_flags & CPU_QUIESCED) != 0);
589 
590 	if (use_mp == 0 || plat_dr_support_cpu() == 0) {
591 		return (ENOTSUP);
592 	}
593 	/*
594 	 * There is no support for powering off cpu0 yet.
595 	 * There are many pieces of code which have a hard dependency on cpu0.
596 	 */
597 	if (cp->cpu_id == 0) {
598 		return (ENOTSUP);
599 	};
600 
601 	if (mach_cpu_get_device_node(cp, &dip) != PSM_SUCCESS) {
602 		return (ENXIO);
603 	}
604 	ASSERT(dip != NULL);
605 	if (mp_cpu_detach_driver(dip) != 0) {
606 		rv = EBUSY;
607 		goto out_online;
608 	}
609 
610 	/* Allocate CPU context for stopping */
611 	if (mach_cpucontext_init() != 0) {
612 		rv = ENXIO;
613 		goto out_online;
614 	}
615 	ctx = mach_cpucontext_xalloc(cp, MACH_CPUCONTEXT_OP_STOP);
616 	if (ctx == NULL) {
617 		rv = ENXIO;
618 		goto out_context_fini;
619 	}
620 
621 	cpupm_stop(cp);
622 	cpu_event_fini_cpu(cp);
623 
624 	if (cp->cpu_m.mcpu_cmi_hdl != NULL) {
625 		cmi_fini(cp->cpu_m.mcpu_cmi_hdl);
626 		cp->cpu_m.mcpu_cmi_hdl = NULL;
627 	}
628 
629 	rv = mach_cpu_stop(cp, ctx);
630 	if (rv != 0) {
631 		goto out_enable_cmi;
632 	}
633 
634 	/* Wait until the target CPU has been halted. */
635 	while (*(volatile ushort_t *)&(rm->rm_cpu_halted) != 0xdead) {
636 		delay(1);
637 	}
638 	rm->rm_cpu_halted = 0xffff;
639 
640 	/* CPU_READY has been cleared by mach_cpu_stop. */
641 	ASSERT((cp->cpu_flags & CPU_READY) == 0);
642 	ASSERT((cp->cpu_flags & CPU_RUNNING) == 0);
643 	cp->cpu_flags = CPU_OFFLINE | CPU_QUIESCED | CPU_POWEROFF;
644 	CPUSET_ATOMIC_DEL(mp_cpus, cp->cpu_id);
645 
646 	mach_cpucontext_xfree(cp, ctx, 0, MACH_CPUCONTEXT_OP_STOP);
647 	mach_cpucontext_fini();
648 
649 	return (0);
650 
651 out_enable_cmi:
652 	{
653 		cmi_hdl_t hdl;
654 
655 		if ((hdl = cmi_init(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp),
656 		    cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp))) != NULL) {
657 			if (is_x86_feature(x86_featureset, X86FSET_MCA))
658 				cmi_mca_init(hdl);
659 			cp->cpu_m.mcpu_cmi_hdl = hdl;
660 		}
661 	}
662 	cpu_event_init_cpu(cp);
663 	cpupm_start(cp);
664 	mach_cpucontext_xfree(cp, ctx, rv, MACH_CPUCONTEXT_OP_STOP);
665 
666 out_context_fini:
667 	mach_cpucontext_fini();
668 
669 out_online:
670 	(void) e_ddi_branch_configure(dip, NULL, 0);
671 
672 	if (rv != EAGAIN && rv != ETIME) {
673 		rv = ENXIO;
674 	}
675 
676 	return (rv);
677 }
678 
679 /*
680  * Return vcpu state, since this could be a virtual environment that we
681  * are unaware of, return "unknown".
682  */
683 /* ARGSUSED */
684 int
vcpu_on_pcpu(processorid_t cpu)685 vcpu_on_pcpu(processorid_t cpu)
686 {
687 	return (VCPU_STATE_UNKNOWN);
688 }
689