xref: /illumos-gate/usr/src/uts/i86pc/os/mp_startup.c (revision 6a074c93c5dee390d8ca2377f42e55418f0a9eb3)
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 /*
23  * Copyright 2007 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #pragma ident	"%Z%%M%	%I%	%E% SMI"
28 
29 #include <sys/types.h>
30 #include <sys/thread.h>
31 #include <sys/cpuvar.h>
32 #include <sys/t_lock.h>
33 #include <sys/param.h>
34 #include <sys/proc.h>
35 #include <sys/disp.h>
36 #include <sys/class.h>
37 #include <sys/cmn_err.h>
38 #include <sys/debug.h>
39 #include <sys/asm_linkage.h>
40 #include <sys/x_call.h>
41 #include <sys/systm.h>
42 #include <sys/var.h>
43 #include <sys/vtrace.h>
44 #include <vm/hat.h>
45 #include <vm/as.h>
46 #include <vm/seg_kmem.h>
47 #include <vm/seg_kp.h>
48 #include <sys/segments.h>
49 #include <sys/kmem.h>
50 #include <sys/stack.h>
51 #include <sys/smp_impldefs.h>
52 #include <sys/x86_archext.h>
53 #include <sys/machsystm.h>
54 #include <sys/traptrace.h>
55 #include <sys/clock.h>
56 #include <sys/cpc_impl.h>
57 #include <sys/pg.h>
58 #include <sys/cmt.h>
59 #include <sys/dtrace.h>
60 #include <sys/archsystm.h>
61 #include <sys/fp.h>
62 #include <sys/reboot.h>
63 #include <sys/kdi_machimpl.h>
64 #include <vm/hat_i86.h>
65 #include <sys/memnode.h>
66 #include <sys/pci_cfgspace.h>
67 #include <sys/mach_mmu.h>
68 #include <sys/sysmacros.h>
69 #if defined(__xpv)
70 #include <sys/hypervisor.h>
71 #endif
72 #include <sys/cpu_module.h>
73 
74 struct cpu	cpus[1];			/* CPU data */
75 struct cpu	*cpu[NCPU] = {&cpus[0]};	/* pointers to all CPUs */
76 cpu_core_t	cpu_core[NCPU];			/* cpu_core structures */
77 
78 /*
79  * Useful for disabling MP bring-up on a MP capable system.
80  */
81 int use_mp = 1;
82 
83 /*
84  * to be set by a PSM to indicate what cpus
85  * are sitting around on the system.
86  */
87 cpuset_t mp_cpus;
88 
89 /*
90  * This variable is used by the hat layer to decide whether or not
91  * critical sections are needed to prevent race conditions.  For sun4m,
92  * this variable is set once enough MP initialization has been done in
93  * order to allow cross calls.
94  */
95 int flushes_require_xcalls;
96 cpuset_t cpu_ready_set = 1;
97 
98 static 	void	mp_startup(void);
99 
100 static void cpu_sep_enable(void);
101 static void cpu_sep_disable(void);
102 static void cpu_asysc_enable(void);
103 static void cpu_asysc_disable(void);
104 
105 /*
106  * Init CPU info - get CPU type info for processor_info system call.
107  */
108 void
109 init_cpu_info(struct cpu *cp)
110 {
111 	processor_info_t *pi = &cp->cpu_type_info;
112 	char buf[CPU_IDSTRLEN];
113 
114 	/*
115 	 * Get clock-frequency property for the CPU.
116 	 */
117 	pi->pi_clock = cpu_freq;
118 
119 	/*
120 	 * Current frequency in Hz.
121 	 */
122 	cp->cpu_curr_clock = cpu_freq_hz;
123 
124 	/*
125 	 * Supported frequencies.
126 	 */
127 	cpu_set_supp_freqs(cp, NULL);
128 
129 	(void) strcpy(pi->pi_processor_type, "i386");
130 	if (fpu_exists)
131 		(void) strcpy(pi->pi_fputypes, "i387 compatible");
132 
133 	(void) cpuid_getidstr(cp, buf, sizeof (buf));
134 
135 	cp->cpu_idstr = kmem_alloc(strlen(buf) + 1, KM_SLEEP);
136 	(void) strcpy(cp->cpu_idstr, buf);
137 
138 	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_idstr);
139 
140 	(void) cpuid_getbrandstr(cp, buf, sizeof (buf));
141 	cp->cpu_brandstr = kmem_alloc(strlen(buf) + 1, KM_SLEEP);
142 	(void) strcpy(cp->cpu_brandstr, buf);
143 
144 	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_brandstr);
145 }
146 
147 /*
148  * Configure syscall support on this CPU.
149  */
150 /*ARGSUSED*/
151 static void
152 init_cpu_syscall(struct cpu *cp)
153 {
154 	kpreempt_disable();
155 
156 #if defined(__amd64)
157 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC)) {
158 
159 #if !defined(__lint)
160 		/*
161 		 * The syscall instruction imposes a certain ordering on
162 		 * segment selectors, so we double-check that ordering
163 		 * here.
164 		 */
165 		ASSERT(KDS_SEL == KCS_SEL + 8);
166 		ASSERT(UDS_SEL == U32CS_SEL + 8);
167 		ASSERT(UCS_SEL == U32CS_SEL + 16);
168 #endif
169 		/*
170 		 * Turn syscall/sysret extensions on.
171 		 */
172 		cpu_asysc_enable();
173 
174 		/*
175 		 * Program the magic registers ..
176 		 */
177 		wrmsr(MSR_AMD_STAR,
178 		    ((uint64_t)(U32CS_SEL << 16 | KCS_SEL)) << 32);
179 		wrmsr(MSR_AMD_LSTAR, (uint64_t)(uintptr_t)sys_syscall);
180 		wrmsr(MSR_AMD_CSTAR, (uint64_t)(uintptr_t)sys_syscall32);
181 
182 		/*
183 		 * This list of flags is masked off the incoming
184 		 * %rfl when we enter the kernel.
185 		 */
186 		wrmsr(MSR_AMD_SFMASK, (uint64_t)(uintptr_t)(PS_IE | PS_T));
187 	}
188 #endif
189 
190 	/*
191 	 * On 32-bit kernels, we use sysenter/sysexit because it's too
192 	 * hard to use syscall/sysret, and it is more portable anyway.
193 	 *
194 	 * On 64-bit kernels on Nocona machines, the 32-bit syscall
195 	 * variant isn't available to 32-bit applications, but sysenter is.
196 	 */
197 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP)) {
198 
199 #if !defined(__lint)
200 		/*
201 		 * The sysenter instruction imposes a certain ordering on
202 		 * segment selectors, so we double-check that ordering
203 		 * here. See "sysenter" in Intel document 245471-012, "IA-32
204 		 * Intel Architecture Software Developer's Manual Volume 2:
205 		 * Instruction Set Reference"
206 		 */
207 		ASSERT(KDS_SEL == KCS_SEL + 8);
208 
209 		ASSERT32(UCS_SEL == ((KCS_SEL + 16) | 3));
210 		ASSERT32(UDS_SEL == UCS_SEL + 8);
211 
212 		ASSERT64(U32CS_SEL == ((KCS_SEL + 16) | 3));
213 		ASSERT64(UDS_SEL == U32CS_SEL + 8);
214 #endif
215 
216 		cpu_sep_enable();
217 
218 		/*
219 		 * resume() sets this value to the base of the threads stack
220 		 * via a context handler.
221 		 */
222 		wrmsr(MSR_INTC_SEP_ESP, 0);
223 		wrmsr(MSR_INTC_SEP_EIP, (uint64_t)(uintptr_t)sys_sysenter);
224 	}
225 
226 	kpreempt_enable();
227 }
228 
229 /*
230  * Multiprocessor initialization.
231  *
232  * Allocate and initialize the cpu structure, TRAPTRACE buffer, and the
233  * startup and idle threads for the specified CPU.
234  */
235 struct cpu *
236 mp_startup_init(int cpun)
237 {
238 	struct cpu *cp;
239 	kthread_id_t tp;
240 	caddr_t	sp;
241 	proc_t *procp;
242 #if !defined(__xpv)
243 	extern int idle_cpu_prefer_mwait;
244 #endif
245 	extern void idle();
246 
247 #ifdef TRAPTRACE
248 	trap_trace_ctl_t *ttc = &trap_trace_ctl[cpun];
249 #endif
250 
251 	ASSERT(cpun < NCPU && cpu[cpun] == NULL);
252 
253 	cp = kmem_zalloc(sizeof (*cp), KM_SLEEP);
254 #if !defined(__xpv)
255 	if ((x86_feature & X86_MWAIT) && idle_cpu_prefer_mwait)
256 		cp->cpu_m.mcpu_mwait = cpuid_mwait_alloc(CPU);
257 #endif
258 
259 	procp = curthread->t_procp;
260 
261 	mutex_enter(&cpu_lock);
262 	/*
263 	 * Initialize the dispatcher first.
264 	 */
265 	disp_cpu_init(cp);
266 	mutex_exit(&cpu_lock);
267 
268 	cpu_vm_data_init(cp);
269 
270 	/*
271 	 * Allocate and initialize the startup thread for this CPU.
272 	 * Interrupt and process switch stacks get allocated later
273 	 * when the CPU starts running.
274 	 */
275 	tp = thread_create(NULL, 0, NULL, NULL, 0, procp,
276 	    TS_STOPPED, maxclsyspri);
277 
278 	/*
279 	 * Set state to TS_ONPROC since this thread will start running
280 	 * as soon as the CPU comes online.
281 	 *
282 	 * All the other fields of the thread structure are setup by
283 	 * thread_create().
284 	 */
285 	THREAD_ONPROC(tp, cp);
286 	tp->t_preempt = 1;
287 	tp->t_bound_cpu = cp;
288 	tp->t_affinitycnt = 1;
289 	tp->t_cpu = cp;
290 	tp->t_disp_queue = cp->cpu_disp;
291 
292 	/*
293 	 * Setup thread to start in mp_startup.
294 	 */
295 	sp = tp->t_stk;
296 	tp->t_pc = (uintptr_t)mp_startup;
297 	tp->t_sp = (uintptr_t)(sp - MINFRAME);
298 #if defined(__amd64)
299 	tp->t_sp -= STACK_ENTRY_ALIGN;		/* fake a call */
300 #endif
301 
302 	cp->cpu_id = cpun;
303 	cp->cpu_self = cp;
304 	cp->cpu_thread = tp;
305 	cp->cpu_lwp = NULL;
306 	cp->cpu_dispthread = tp;
307 	cp->cpu_dispatch_pri = DISP_PRIO(tp);
308 
309 	/*
310 	 * cpu_base_spl must be set explicitly here to prevent any blocking
311 	 * operations in mp_startup from causing the spl of the cpu to drop
312 	 * to 0 (allowing device interrupts before we're ready) in resume().
313 	 * cpu_base_spl MUST remain at LOCK_LEVEL until the cpu is CPU_READY.
314 	 * As an extra bit of security on DEBUG kernels, this is enforced with
315 	 * an assertion in mp_startup() -- before cpu_base_spl is set to its
316 	 * proper value.
317 	 */
318 	cp->cpu_base_spl = ipltospl(LOCK_LEVEL);
319 
320 	/*
321 	 * Now, initialize per-CPU idle thread for this CPU.
322 	 */
323 	tp = thread_create(NULL, PAGESIZE, idle, NULL, 0, procp, TS_ONPROC, -1);
324 
325 	cp->cpu_idle_thread = tp;
326 
327 	tp->t_preempt = 1;
328 	tp->t_bound_cpu = cp;
329 	tp->t_affinitycnt = 1;
330 	tp->t_cpu = cp;
331 	tp->t_disp_queue = cp->cpu_disp;
332 
333 	/*
334 	 * Bootstrap the CPU's PG data
335 	 */
336 	pg_cpu_bootstrap(cp);
337 
338 	/*
339 	 * Perform CPC initialization on the new CPU.
340 	 */
341 	kcpc_hw_init(cp);
342 
343 	/*
344 	 * Allocate virtual addresses for cpu_caddr1 and cpu_caddr2
345 	 * for each CPU.
346 	 */
347 	setup_vaddr_for_ppcopy(cp);
348 
349 	/*
350 	 * Allocate page for new GDT and initialize from current GDT.
351 	 */
352 #if !defined(__lint)
353 	ASSERT((sizeof (*cp->cpu_gdt) * NGDT) <= PAGESIZE);
354 #endif
355 	cp->cpu_m.mcpu_gdt = kmem_zalloc(PAGESIZE, KM_SLEEP);
356 	bcopy(CPU->cpu_m.mcpu_gdt, cp->cpu_m.mcpu_gdt,
357 	    (sizeof (*cp->cpu_m.mcpu_gdt) * NGDT));
358 
359 #if defined(__i386)
360 	/*
361 	 * setup kernel %gs.
362 	 */
363 	set_usegd(&cp->cpu_gdt[GDT_GS], cp, sizeof (struct cpu) -1, SDT_MEMRWA,
364 	    SEL_KPL, 0, 1);
365 #endif
366 
367 	/*
368 	 * If we have more than one node, each cpu gets a copy of IDT
369 	 * local to its node. If this is a Pentium box, we use cpu 0's
370 	 * IDT. cpu 0's IDT has been made read-only to workaround the
371 	 * cmpxchgl register bug
372 	 */
373 	if (system_hardware.hd_nodes && x86_type != X86_TYPE_P5) {
374 		struct machcpu *mcpu = &cp->cpu_m;
375 
376 		mcpu->mcpu_idt = kmem_alloc(sizeof (idt0), KM_SLEEP);
377 		bcopy(idt0, mcpu->mcpu_idt, sizeof (idt0));
378 	} else {
379 		cp->cpu_m.mcpu_idt = CPU->cpu_m.mcpu_idt;
380 	}
381 
382 	/*
383 	 * Get interrupt priority data from cpu 0.
384 	 */
385 	cp->cpu_pri_data = CPU->cpu_pri_data;
386 
387 	/*
388 	 * alloc space for cpuid info
389 	 */
390 	cpuid_alloc_space(cp);
391 
392 #if !defined(__xpv)
393 	/*
394 	 * alloc space for ucode_info
395 	 */
396 	ucode_alloc_space(cp);
397 #endif
398 
399 	hat_cpu_online(cp);
400 
401 #ifdef TRAPTRACE
402 	/*
403 	 * If this is a TRAPTRACE kernel, allocate TRAPTRACE buffers
404 	 */
405 	ttc->ttc_first = (uintptr_t)kmem_zalloc(trap_trace_bufsize, KM_SLEEP);
406 	ttc->ttc_next = ttc->ttc_first;
407 	ttc->ttc_limit = ttc->ttc_first + trap_trace_bufsize;
408 #endif
409 	/*
410 	 * Record that we have another CPU.
411 	 */
412 	mutex_enter(&cpu_lock);
413 	/*
414 	 * Initialize the interrupt threads for this CPU
415 	 */
416 	cpu_intr_alloc(cp, NINTR_THREADS);
417 	/*
418 	 * Add CPU to list of available CPUs.  It'll be on the active list
419 	 * after mp_startup().
420 	 */
421 	cpu_add_unit(cp);
422 	mutex_exit(&cpu_lock);
423 
424 	return (cp);
425 }
426 
427 /*
428  * Undo what was done in mp_startup_init
429  */
430 static void
431 mp_startup_fini(struct cpu *cp, int error)
432 {
433 	mutex_enter(&cpu_lock);
434 
435 	/*
436 	 * Remove the CPU from the list of available CPUs.
437 	 */
438 	cpu_del_unit(cp->cpu_id);
439 
440 	if (error == ETIMEDOUT) {
441 		/*
442 		 * The cpu was started, but never *seemed* to run any
443 		 * code in the kernel; it's probably off spinning in its
444 		 * own private world, though with potential references to
445 		 * our kmem-allocated IDTs and GDTs (for example).
446 		 *
447 		 * Worse still, it may actually wake up some time later,
448 		 * so rather than guess what it might or might not do, we
449 		 * leave the fundamental data structures intact.
450 		 */
451 		cp->cpu_flags = 0;
452 		mutex_exit(&cpu_lock);
453 		return;
454 	}
455 
456 	/*
457 	 * At this point, the only threads bound to this CPU should
458 	 * special per-cpu threads: it's idle thread, it's pause threads,
459 	 * and it's interrupt threads.  Clean these up.
460 	 */
461 	cpu_destroy_bound_threads(cp);
462 	cp->cpu_idle_thread = NULL;
463 
464 	/*
465 	 * Free the interrupt stack.
466 	 */
467 	segkp_release(segkp,
468 	    cp->cpu_intr_stack - (INTR_STACK_SIZE - SA(MINFRAME)));
469 
470 	mutex_exit(&cpu_lock);
471 
472 #ifdef TRAPTRACE
473 	/*
474 	 * Discard the trap trace buffer
475 	 */
476 	{
477 		trap_trace_ctl_t *ttc = &trap_trace_ctl[cp->cpu_id];
478 
479 		kmem_free((void *)ttc->ttc_first, trap_trace_bufsize);
480 		ttc->ttc_first = NULL;
481 	}
482 #endif
483 
484 	hat_cpu_offline(cp);
485 
486 	cpuid_free_space(cp);
487 
488 #if !defined(__xpv)
489 	ucode_free_space(cp);
490 #endif
491 
492 	if (cp->cpu_m.mcpu_idt != CPU->cpu_m.mcpu_idt)
493 		kmem_free(cp->cpu_m.mcpu_idt, sizeof (idt0));
494 	cp->cpu_m.mcpu_idt = NULL;
495 
496 	kmem_free(cp->cpu_m.mcpu_gdt, PAGESIZE);
497 	cp->cpu_m.mcpu_gdt = NULL;
498 
499 	teardown_vaddr_for_ppcopy(cp);
500 
501 	kcpc_hw_fini(cp);
502 
503 	cp->cpu_dispthread = NULL;
504 	cp->cpu_thread = NULL;	/* discarded by cpu_destroy_bound_threads() */
505 
506 	cpu_vm_data_destroy(cp);
507 
508 	mutex_enter(&cpu_lock);
509 	disp_cpu_fini(cp);
510 	mutex_exit(&cpu_lock);
511 
512 #if !defined(__xpv)
513 	if (cp->cpu_m.mcpu_mwait != NULL)
514 		cpuid_mwait_free(cp);
515 #endif
516 	kmem_free(cp, sizeof (*cp));
517 }
518 
519 /*
520  * Apply workarounds for known errata, and warn about those that are absent.
521  *
522  * System vendors occasionally create configurations which contain different
523  * revisions of the CPUs that are almost but not exactly the same.  At the
524  * time of writing, this meant that their clock rates were the same, their
525  * feature sets were the same, but the required workaround were -not-
526  * necessarily the same.  So, this routine is invoked on -every- CPU soon
527  * after starting to make sure that the resulting system contains the most
528  * pessimal set of workarounds needed to cope with *any* of the CPUs in the
529  * system.
530  *
531  * workaround_errata is invoked early in mlsetup() for CPU 0, and in
532  * mp_startup() for all slave CPUs. Slaves process workaround_errata prior
533  * to acknowledging their readiness to the master, so this routine will
534  * never be executed by multiple CPUs in parallel, thus making updates to
535  * global data safe.
536  *
537  * These workarounds are based on Rev 3.57 of the Revision Guide for
538  * AMD Athlon(tm) 64 and AMD Opteron(tm) Processors, August 2005.
539  */
540 
541 #if defined(OPTERON_ERRATUM_88)
542 int opteron_erratum_88;		/* if non-zero -> at least one cpu has it */
543 #endif
544 
545 #if defined(OPTERON_ERRATUM_91)
546 int opteron_erratum_91;		/* if non-zero -> at least one cpu has it */
547 #endif
548 
549 #if defined(OPTERON_ERRATUM_93)
550 int opteron_erratum_93;		/* if non-zero -> at least one cpu has it */
551 #endif
552 
553 #if defined(OPTERON_ERRATUM_95)
554 int opteron_erratum_95;		/* if non-zero -> at least one cpu has it */
555 #endif
556 
557 #if defined(OPTERON_ERRATUM_100)
558 int opteron_erratum_100;	/* if non-zero -> at least one cpu has it */
559 #endif
560 
561 #if defined(OPTERON_ERRATUM_108)
562 int opteron_erratum_108;	/* if non-zero -> at least one cpu has it */
563 #endif
564 
565 #if defined(OPTERON_ERRATUM_109)
566 int opteron_erratum_109;	/* if non-zero -> at least one cpu has it */
567 #endif
568 
569 #if defined(OPTERON_ERRATUM_121)
570 int opteron_erratum_121;	/* if non-zero -> at least one cpu has it */
571 #endif
572 
573 #if defined(OPTERON_ERRATUM_122)
574 int opteron_erratum_122;	/* if non-zero -> at least one cpu has it */
575 #endif
576 
577 #if defined(OPTERON_ERRATUM_123)
578 int opteron_erratum_123;	/* if non-zero -> at least one cpu has it */
579 #endif
580 
581 #if defined(OPTERON_ERRATUM_131)
582 int opteron_erratum_131;	/* if non-zero -> at least one cpu has it */
583 #endif
584 
585 #if defined(OPTERON_WORKAROUND_6336786)
586 int opteron_workaround_6336786;	/* non-zero -> WA relevant and applied */
587 int opteron_workaround_6336786_UP = 0;	/* Not needed for UP */
588 #endif
589 
590 #if defined(OPTERON_WORKAROUND_6323525)
591 int opteron_workaround_6323525;	/* if non-zero -> at least one cpu has it */
592 #endif
593 
594 static void
595 workaround_warning(cpu_t *cp, uint_t erratum)
596 {
597 	cmn_err(CE_WARN, "cpu%d: no workaround for erratum %u",
598 	    cp->cpu_id, erratum);
599 }
600 
601 static void
602 workaround_applied(uint_t erratum)
603 {
604 	if (erratum > 1000000)
605 		cmn_err(CE_CONT, "?workaround applied for cpu issue #%d\n",
606 		    erratum);
607 	else
608 		cmn_err(CE_CONT, "?workaround applied for cpu erratum #%d\n",
609 		    erratum);
610 }
611 
612 static void
613 msr_warning(cpu_t *cp, const char *rw, uint_t msr, int error)
614 {
615 	cmn_err(CE_WARN, "cpu%d: couldn't %smsr 0x%x, error %d",
616 	    cp->cpu_id, rw, msr, error);
617 }
618 
619 #if defined(__xpv)
620 
621 /*
622  * On dom0, we can determine the number of physical cpus on the machine.
623  * This number is important when figuring out what workarounds are
624  * appropriate, so compute it now.
625  */
626 static uint_t
627 xen_get_nphyscpus(void)
628 {
629 	static uint_t nphyscpus = 0;
630 
631 	ASSERT(DOMAIN_IS_INITDOMAIN(xen_info));
632 
633 	if (nphyscpus == 0) {
634 		xen_sysctl_t op;
635 		xen_sysctl_physinfo_t *pi = &op.u.physinfo;
636 
637 		op.cmd = XEN_SYSCTL_physinfo;
638 		op.interface_version = XEN_SYSCTL_INTERFACE_VERSION;
639 		if (HYPERVISOR_sysctl(&op) == 0)
640 			nphyscpus = pi->threads_per_core *
641 			    pi->cores_per_socket * pi->sockets_per_node *
642 			    pi->nr_nodes;
643 	}
644 	return (nphyscpus);
645 }
646 #endif
647 
648 uint_t
649 workaround_errata(struct cpu *cpu)
650 {
651 	uint_t missing = 0;
652 
653 	ASSERT(cpu == CPU);
654 
655 	/*LINTED*/
656 	if (cpuid_opteron_erratum(cpu, 88) > 0) {
657 		/*
658 		 * SWAPGS May Fail To Read Correct GS Base
659 		 */
660 #if defined(OPTERON_ERRATUM_88)
661 		/*
662 		 * The workaround is an mfence in the relevant assembler code
663 		 */
664 		opteron_erratum_88++;
665 #else
666 		workaround_warning(cpu, 88);
667 		missing++;
668 #endif
669 	}
670 
671 	if (cpuid_opteron_erratum(cpu, 91) > 0) {
672 		/*
673 		 * Software Prefetches May Report A Page Fault
674 		 */
675 #if defined(OPTERON_ERRATUM_91)
676 		/*
677 		 * fix is in trap.c
678 		 */
679 		opteron_erratum_91++;
680 #else
681 		workaround_warning(cpu, 91);
682 		missing++;
683 #endif
684 	}
685 
686 	if (cpuid_opteron_erratum(cpu, 93) > 0) {
687 		/*
688 		 * RSM Auto-Halt Restart Returns to Incorrect RIP
689 		 */
690 #if defined(OPTERON_ERRATUM_93)
691 		/*
692 		 * fix is in trap.c
693 		 */
694 		opteron_erratum_93++;
695 #else
696 		workaround_warning(cpu, 93);
697 		missing++;
698 #endif
699 	}
700 
701 	/*LINTED*/
702 	if (cpuid_opteron_erratum(cpu, 95) > 0) {
703 		/*
704 		 * RET Instruction May Return to Incorrect EIP
705 		 */
706 #if defined(OPTERON_ERRATUM_95)
707 #if defined(_LP64)
708 		/*
709 		 * Workaround this by ensuring that 32-bit user code and
710 		 * 64-bit kernel code never occupy the same address
711 		 * range mod 4G.
712 		 */
713 		if (_userlimit32 > 0xc0000000ul)
714 			*(uintptr_t *)&_userlimit32 = 0xc0000000ul;
715 
716 		/*LINTED*/
717 		ASSERT((uint32_t)COREHEAP_BASE == 0xc0000000u);
718 		opteron_erratum_95++;
719 #endif	/* _LP64 */
720 #else
721 		workaround_warning(cpu, 95);
722 		missing++;
723 #endif
724 	}
725 
726 	if (cpuid_opteron_erratum(cpu, 100) > 0) {
727 		/*
728 		 * Compatibility Mode Branches Transfer to Illegal Address
729 		 */
730 #if defined(OPTERON_ERRATUM_100)
731 		/*
732 		 * fix is in trap.c
733 		 */
734 		opteron_erratum_100++;
735 #else
736 		workaround_warning(cpu, 100);
737 		missing++;
738 #endif
739 	}
740 
741 	/*LINTED*/
742 	if (cpuid_opteron_erratum(cpu, 108) > 0) {
743 		/*
744 		 * CPUID Instruction May Return Incorrect Model Number In
745 		 * Some Processors
746 		 */
747 #if defined(OPTERON_ERRATUM_108)
748 		/*
749 		 * (Our cpuid-handling code corrects the model number on
750 		 * those processors)
751 		 */
752 #else
753 		workaround_warning(cpu, 108);
754 		missing++;
755 #endif
756 	}
757 
758 	/*LINTED*/
759 	if (cpuid_opteron_erratum(cpu, 109) > 0) do {
760 		/*
761 		 * Certain Reverse REP MOVS May Produce Unpredictable Behaviour
762 		 */
763 #if defined(OPTERON_ERRATUM_109)
764 		/*
765 		 * The "workaround" is to print a warning to upgrade the BIOS
766 		 */
767 		uint64_t value;
768 		const uint_t msr = MSR_AMD_PATCHLEVEL;
769 		int err;
770 
771 		if ((err = checked_rdmsr(msr, &value)) != 0) {
772 			msr_warning(cpu, "rd", msr, err);
773 			workaround_warning(cpu, 109);
774 			missing++;
775 		}
776 		if (value == 0)
777 			opteron_erratum_109++;
778 #else
779 		workaround_warning(cpu, 109);
780 		missing++;
781 #endif
782 	/*CONSTANTCONDITION*/
783 	} while (0);
784 
785 	/*LINTED*/
786 	if (cpuid_opteron_erratum(cpu, 121) > 0) {
787 		/*
788 		 * Sequential Execution Across Non_Canonical Boundary Caused
789 		 * Processor Hang
790 		 */
791 #if defined(OPTERON_ERRATUM_121)
792 #if defined(_LP64)
793 		/*
794 		 * Erratum 121 is only present in long (64 bit) mode.
795 		 * Workaround is to include the page immediately before the
796 		 * va hole to eliminate the possibility of system hangs due to
797 		 * sequential execution across the va hole boundary.
798 		 */
799 		if (opteron_erratum_121)
800 			opteron_erratum_121++;
801 		else {
802 			if (hole_start) {
803 				hole_start -= PAGESIZE;
804 			} else {
805 				/*
806 				 * hole_start not yet initialized by
807 				 * mmu_init. Initialize hole_start
808 				 * with value to be subtracted.
809 				 */
810 				hole_start = PAGESIZE;
811 			}
812 			opteron_erratum_121++;
813 		}
814 #endif	/* _LP64 */
815 #else
816 		workaround_warning(cpu, 121);
817 		missing++;
818 #endif
819 	}
820 
821 	/*LINTED*/
822 	if (cpuid_opteron_erratum(cpu, 122) > 0) do {
823 		/*
824 		 * TLB Flush Filter May Cause Coherency Problem in
825 		 * Multiprocessor Systems
826 		 */
827 #if defined(OPTERON_ERRATUM_122)
828 		uint64_t value;
829 		const uint_t msr = MSR_AMD_HWCR;
830 		int error;
831 
832 		/*
833 		 * Erratum 122 is only present in MP configurations (multi-core
834 		 * or multi-processor).
835 		 */
836 #if defined(__xpv)
837 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
838 			break;
839 		if (!opteron_erratum_122 && xen_get_nphyscpus() == 1)
840 			break;
841 #else
842 		if (!opteron_erratum_122 && lgrp_plat_node_cnt == 1 &&
843 		    cpuid_get_ncpu_per_chip(cpu) == 1)
844 			break;
845 #endif
846 		/* disable TLB Flush Filter */
847 
848 		if ((error = checked_rdmsr(msr, &value)) != 0) {
849 			msr_warning(cpu, "rd", msr, error);
850 			workaround_warning(cpu, 122);
851 			missing++;
852 		} else {
853 			value |= (uint64_t)AMD_HWCR_FFDIS;
854 			if ((error = checked_wrmsr(msr, value)) != 0) {
855 				msr_warning(cpu, "wr", msr, error);
856 				workaround_warning(cpu, 122);
857 				missing++;
858 			}
859 		}
860 		opteron_erratum_122++;
861 #else
862 		workaround_warning(cpu, 122);
863 		missing++;
864 #endif
865 	/*CONSTANTCONDITION*/
866 	} while (0);
867 
868 	/*LINTED*/
869 	if (cpuid_opteron_erratum(cpu, 123) > 0) do {
870 		/*
871 		 * Bypassed Reads May Cause Data Corruption of System Hang in
872 		 * Dual Core Processors
873 		 */
874 #if defined(OPTERON_ERRATUM_123)
875 		uint64_t value;
876 		const uint_t msr = MSR_AMD_PATCHLEVEL;
877 		int err;
878 
879 		/*
880 		 * Erratum 123 applies only to multi-core cpus.
881 		 */
882 		if (cpuid_get_ncpu_per_chip(cpu) < 2)
883 			break;
884 #if defined(__xpv)
885 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
886 			break;
887 #endif
888 		/*
889 		 * The "workaround" is to print a warning to upgrade the BIOS
890 		 */
891 		if ((err = checked_rdmsr(msr, &value)) != 0) {
892 			msr_warning(cpu, "rd", msr, err);
893 			workaround_warning(cpu, 123);
894 			missing++;
895 		}
896 		if (value == 0)
897 			opteron_erratum_123++;
898 #else
899 		workaround_warning(cpu, 123);
900 		missing++;
901 
902 #endif
903 	/*CONSTANTCONDITION*/
904 	} while (0);
905 
906 	/*LINTED*/
907 	if (cpuid_opteron_erratum(cpu, 131) > 0) do {
908 		/*
909 		 * Multiprocessor Systems with Four or More Cores May Deadlock
910 		 * Waiting for a Probe Response
911 		 */
912 #if defined(OPTERON_ERRATUM_131)
913 		uint64_t nbcfg;
914 		const uint_t msr = MSR_AMD_NB_CFG;
915 		const uint64_t wabits =
916 		    AMD_NB_CFG_SRQ_HEARTBEAT | AMD_NB_CFG_SRQ_SPR;
917 		int error;
918 
919 		/*
920 		 * Erratum 131 applies to any system with four or more cores.
921 		 */
922 		if (opteron_erratum_131)
923 			break;
924 #if defined(__xpv)
925 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
926 			break;
927 		if (xen_get_nphyscpus() < 4)
928 			break;
929 #else
930 		if (lgrp_plat_node_cnt * cpuid_get_ncpu_per_chip(cpu) < 4)
931 			break;
932 #endif
933 		/*
934 		 * Print a warning if neither of the workarounds for
935 		 * erratum 131 is present.
936 		 */
937 		if ((error = checked_rdmsr(msr, &nbcfg)) != 0) {
938 			msr_warning(cpu, "rd", msr, error);
939 			workaround_warning(cpu, 131);
940 			missing++;
941 		} else if ((nbcfg & wabits) == 0) {
942 			opteron_erratum_131++;
943 		} else {
944 			/* cannot have both workarounds set */
945 			ASSERT((nbcfg & wabits) != wabits);
946 		}
947 #else
948 		workaround_warning(cpu, 131);
949 		missing++;
950 #endif
951 	/*CONSTANTCONDITION*/
952 	} while (0);
953 
954 	/*
955 	 * This isn't really an erratum, but for convenience the
956 	 * detection/workaround code lives here and in cpuid_opteron_erratum.
957 	 */
958 	if (cpuid_opteron_erratum(cpu, 6336786) > 0) {
959 #if defined(OPTERON_WORKAROUND_6336786)
960 		/*
961 		 * Disable C1-Clock ramping on multi-core/multi-processor
962 		 * K8 platforms to guard against TSC drift.
963 		 */
964 		if (opteron_workaround_6336786) {
965 			opteron_workaround_6336786++;
966 #if defined(__xpv)
967 		} else if ((DOMAIN_IS_INITDOMAIN(xen_info) &&
968 		    xen_get_nphyscpus() > 1) ||
969 		    opteron_workaround_6336786_UP) {
970 			/*
971 			 * XXPV	Hmm.  We can't walk the set of lgrps on
972 			 *	the hypervisor; so just complain and drive
973 			 *	on.  This probably needs to be fixed in
974 			 *	the hypervisor itself.
975 			 */
976 			opteron_workaround_6336786++;
977 			workaround_warning(cpu, 6336786);
978 #else	/* __xpv */
979 		} else if ((lgrp_plat_node_cnt *
980 		    cpuid_get_ncpu_per_chip(cpu) > 1) ||
981 		    opteron_workaround_6336786_UP) {
982 			int	node;
983 			uint8_t data;
984 
985 			for (node = 0; node < lgrp_plat_node_cnt; node++) {
986 				/*
987 				 * Clear PMM7[1:0] (function 3, offset 0x87)
988 				 * Northbridge device is the node id + 24.
989 				 */
990 				data = pci_getb_func(0, node + 24, 3, 0x87);
991 				data &= 0xFC;
992 				pci_putb_func(0, node + 24, 3, 0x87, data);
993 			}
994 			opteron_workaround_6336786++;
995 #endif	/* __xpv */
996 		}
997 #else
998 		workaround_warning(cpu, 6336786);
999 		missing++;
1000 #endif
1001 	}
1002 
1003 	/*LINTED*/
1004 	/*
1005 	 * Mutex primitives don't work as expected.
1006 	 */
1007 	if (cpuid_opteron_erratum(cpu, 6323525) > 0) {
1008 #if defined(OPTERON_WORKAROUND_6323525)
1009 		/*
1010 		 * This problem only occurs with 2 or more cores. If bit in
1011 		 * MSR_BU_CFG set, then not applicable. The workaround
1012 		 * is to patch the semaphone routines with the lfence
1013 		 * instruction to provide necessary load memory barrier with
1014 		 * possible subsequent read-modify-write ops.
1015 		 *
1016 		 * It is too early in boot to call the patch routine so
1017 		 * set erratum variable to be done in startup_end().
1018 		 */
1019 		if (opteron_workaround_6323525) {
1020 			opteron_workaround_6323525++;
1021 #if defined(__xpv)
1022 		} else if (x86_feature & X86_SSE2) {
1023 			if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1024 				/*
1025 				 * XXPV	Use dom0_msr here when extended
1026 				 *	operations are supported?
1027 				 */
1028 				if (xen_get_nphyscpus() > 1)
1029 					opteron_workaround_6323525++;
1030 			} else {
1031 				/*
1032 				 * We have no way to tell how many physical
1033 				 * cpus there are, or even if this processor
1034 				 * has the problem, so enable the workaround
1035 				 * unconditionally (at some performance cost).
1036 				 */
1037 				opteron_workaround_6323525++;
1038 			}
1039 #else	/* __xpv */
1040 		} else if ((x86_feature & X86_SSE2) && ((lgrp_plat_node_cnt *
1041 		    cpuid_get_ncpu_per_chip(cpu)) > 1)) {
1042 			if ((xrdmsr(MSR_BU_CFG) & 0x02) == 0)
1043 				opteron_workaround_6323525++;
1044 #endif	/* __xpv */
1045 		}
1046 #else
1047 		workaround_warning(cpu, 6323525);
1048 		missing++;
1049 #endif
1050 	}
1051 
1052 #ifdef __xpv
1053 	return (0);
1054 #else
1055 	return (missing);
1056 #endif
1057 }
1058 
1059 void
1060 workaround_errata_end()
1061 {
1062 #if defined(OPTERON_ERRATUM_88)
1063 	if (opteron_erratum_88)
1064 		workaround_applied(88);
1065 #endif
1066 #if defined(OPTERON_ERRATUM_91)
1067 	if (opteron_erratum_91)
1068 		workaround_applied(91);
1069 #endif
1070 #if defined(OPTERON_ERRATUM_93)
1071 	if (opteron_erratum_93)
1072 		workaround_applied(93);
1073 #endif
1074 #if defined(OPTERON_ERRATUM_95)
1075 	if (opteron_erratum_95)
1076 		workaround_applied(95);
1077 #endif
1078 #if defined(OPTERON_ERRATUM_100)
1079 	if (opteron_erratum_100)
1080 		workaround_applied(100);
1081 #endif
1082 #if defined(OPTERON_ERRATUM_108)
1083 	if (opteron_erratum_108)
1084 		workaround_applied(108);
1085 #endif
1086 #if defined(OPTERON_ERRATUM_109)
1087 	if (opteron_erratum_109) {
1088 		cmn_err(CE_WARN,
1089 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1090 		    " processor\nerratum 109 was not detected; updating your"
1091 		    " system's BIOS to a version\ncontaining this"
1092 		    " microcode patch is HIGHLY recommended or erroneous"
1093 		    " system\noperation may occur.\n");
1094 	}
1095 #endif
1096 #if defined(OPTERON_ERRATUM_121)
1097 	if (opteron_erratum_121)
1098 		workaround_applied(121);
1099 #endif
1100 #if defined(OPTERON_ERRATUM_122)
1101 	if (opteron_erratum_122)
1102 		workaround_applied(122);
1103 #endif
1104 #if defined(OPTERON_ERRATUM_123)
1105 	if (opteron_erratum_123) {
1106 		cmn_err(CE_WARN,
1107 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1108 		    " processor\nerratum 123 was not detected; updating your"
1109 		    " system's BIOS to a version\ncontaining this"
1110 		    " microcode patch is HIGHLY recommended or erroneous"
1111 		    " system\noperation may occur.\n");
1112 	}
1113 #endif
1114 #if defined(OPTERON_ERRATUM_131)
1115 	if (opteron_erratum_131) {
1116 		cmn_err(CE_WARN,
1117 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1118 		    " processor\nerratum 131 was not detected; updating your"
1119 		    " system's BIOS to a version\ncontaining this"
1120 		    " microcode patch is HIGHLY recommended or erroneous"
1121 		    " system\noperation may occur.\n");
1122 	}
1123 #endif
1124 #if defined(OPTERON_WORKAROUND_6336786)
1125 	if (opteron_workaround_6336786)
1126 		workaround_applied(6336786);
1127 #endif
1128 #if defined(OPTERON_WORKAROUND_6323525)
1129 	if (opteron_workaround_6323525)
1130 		workaround_applied(6323525);
1131 #endif
1132 }
1133 
1134 static cpuset_t procset;
1135 
1136 /*
1137  * Start a single cpu, assuming that the kernel context is available
1138  * to successfully start another cpu.
1139  *
1140  * (For example, real mode code is mapped into the right place
1141  * in memory and is ready to be run.)
1142  */
1143 int
1144 start_cpu(processorid_t who)
1145 {
1146 	void *ctx;
1147 	cpu_t *cp;
1148 	int delays;
1149 	int error = 0;
1150 
1151 	ASSERT(who != 0);
1152 
1153 	/*
1154 	 * Check if there's at least a Mbyte of kmem available
1155 	 * before attempting to start the cpu.
1156 	 */
1157 	if (kmem_avail() < 1024 * 1024) {
1158 		/*
1159 		 * Kick off a reap in case that helps us with
1160 		 * later attempts ..
1161 		 */
1162 		kmem_reap();
1163 		return (ENOMEM);
1164 	}
1165 
1166 	cp = mp_startup_init(who);
1167 	if ((ctx = mach_cpucontext_alloc(cp)) == NULL ||
1168 	    (error = mach_cpu_start(cp, ctx)) != 0) {
1169 
1170 		/*
1171 		 * Something went wrong before we even started it
1172 		 */
1173 		if (ctx)
1174 			cmn_err(CE_WARN,
1175 			    "cpu%d: failed to start error %d",
1176 			    cp->cpu_id, error);
1177 		else
1178 			cmn_err(CE_WARN,
1179 			    "cpu%d: failed to allocate context", cp->cpu_id);
1180 
1181 		if (ctx)
1182 			mach_cpucontext_free(cp, ctx, error);
1183 		else
1184 			error = EAGAIN;		/* hmm. */
1185 		mp_startup_fini(cp, error);
1186 		return (error);
1187 	}
1188 
1189 	for (delays = 0; !CPU_IN_SET(procset, who); delays++) {
1190 		if (delays == 500) {
1191 			/*
1192 			 * After five seconds, things are probably looking
1193 			 * a bit bleak - explain the hang.
1194 			 */
1195 			cmn_err(CE_NOTE, "cpu%d: started, "
1196 			    "but not running in the kernel yet", who);
1197 		} else if (delays > 2000) {
1198 			/*
1199 			 * We waited at least 20 seconds, bail ..
1200 			 */
1201 			error = ETIMEDOUT;
1202 			cmn_err(CE_WARN, "cpu%d: timed out", who);
1203 			mach_cpucontext_free(cp, ctx, error);
1204 			mp_startup_fini(cp, error);
1205 			return (error);
1206 		}
1207 
1208 		/*
1209 		 * wait at least 10ms, then check again..
1210 		 */
1211 		delay(USEC_TO_TICK_ROUNDUP(10000));
1212 	}
1213 
1214 	mach_cpucontext_free(cp, ctx, 0);
1215 
1216 #ifndef __xpv
1217 	if (tsc_gethrtime_enable)
1218 		tsc_sync_master(who);
1219 #endif
1220 
1221 	if (dtrace_cpu_init != NULL) {
1222 		/*
1223 		 * DTrace CPU initialization expects cpu_lock to be held.
1224 		 */
1225 		mutex_enter(&cpu_lock);
1226 		(*dtrace_cpu_init)(who);
1227 		mutex_exit(&cpu_lock);
1228 	}
1229 
1230 	while (!CPU_IN_SET(cpu_ready_set, who))
1231 		delay(1);
1232 
1233 	return (0);
1234 }
1235 
1236 
1237 /*ARGSUSED*/
1238 void
1239 start_other_cpus(int cprboot)
1240 {
1241 	uint_t who;
1242 	uint_t skipped = 0;
1243 	uint_t bootcpuid = 0;
1244 
1245 	/*
1246 	 * Initialize our own cpu_info.
1247 	 */
1248 	init_cpu_info(CPU);
1249 
1250 	/*
1251 	 * Initialize our syscall handlers
1252 	 */
1253 	init_cpu_syscall(CPU);
1254 
1255 	/*
1256 	 * Take the boot cpu out of the mp_cpus set because we know
1257 	 * it's already running.  Add it to the cpu_ready_set for
1258 	 * precisely the same reason.
1259 	 */
1260 	CPUSET_DEL(mp_cpus, bootcpuid);
1261 	CPUSET_ADD(cpu_ready_set, bootcpuid);
1262 
1263 	/*
1264 	 * if only 1 cpu or not using MP, skip the rest of this
1265 	 */
1266 	if (CPUSET_ISNULL(mp_cpus) || use_mp == 0) {
1267 		if (use_mp == 0)
1268 			cmn_err(CE_CONT, "?***** Not in MP mode\n");
1269 		goto done;
1270 	}
1271 
1272 	/*
1273 	 * perform such initialization as is needed
1274 	 * to be able to take CPUs on- and off-line.
1275 	 */
1276 	cpu_pause_init();
1277 
1278 	xc_init();		/* initialize processor crosscalls */
1279 
1280 	if (mach_cpucontext_init() != 0)
1281 		goto done;
1282 
1283 	flushes_require_xcalls = 1;
1284 
1285 	/*
1286 	 * We lock our affinity to the master CPU to ensure that all slave CPUs
1287 	 * do their TSC syncs with the same CPU.
1288 	 */
1289 	affinity_set(CPU_CURRENT);
1290 
1291 	for (who = 0; who < NCPU; who++) {
1292 
1293 		if (!CPU_IN_SET(mp_cpus, who))
1294 			continue;
1295 		ASSERT(who != bootcpuid);
1296 		if (ncpus >= max_ncpus) {
1297 			skipped = who;
1298 			continue;
1299 		}
1300 		if (start_cpu(who) != 0)
1301 			CPUSET_DEL(mp_cpus, who);
1302 	}
1303 
1304 #if !defined(__xpv)
1305 	/* Free the space allocated to hold the microcode file */
1306 	ucode_free();
1307 #endif
1308 
1309 	affinity_clear();
1310 
1311 	if (skipped) {
1312 		cmn_err(CE_NOTE,
1313 		    "System detected %d cpus, but "
1314 		    "only %d cpu(s) were enabled during boot.",
1315 		    skipped + 1, ncpus);
1316 		cmn_err(CE_NOTE,
1317 		    "Use \"boot-ncpus\" parameter to enable more CPU(s). "
1318 		    "See eeprom(1M).");
1319 	}
1320 
1321 done:
1322 	workaround_errata_end();
1323 	mach_cpucontext_fini();
1324 
1325 	cmi_post_mpstartup();
1326 }
1327 
1328 /*
1329  * Dummy functions - no i86pc platforms support dynamic cpu allocation.
1330  */
1331 /*ARGSUSED*/
1332 int
1333 mp_cpu_configure(int cpuid)
1334 {
1335 	return (ENOTSUP);		/* not supported */
1336 }
1337 
1338 /*ARGSUSED*/
1339 int
1340 mp_cpu_unconfigure(int cpuid)
1341 {
1342 	return (ENOTSUP);		/* not supported */
1343 }
1344 
1345 /*
1346  * Startup function for 'other' CPUs (besides boot cpu).
1347  * Called from real_mode_start.
1348  *
1349  * WARNING: until CPU_READY is set, mp_startup and routines called by
1350  * mp_startup should not call routines (e.g. kmem_free) that could call
1351  * hat_unload which requires CPU_READY to be set.
1352  */
1353 void
1354 mp_startup(void)
1355 {
1356 	struct cpu *cp = CPU;
1357 	uint_t new_x86_feature;
1358 
1359 	/*
1360 	 * We need to get TSC on this proc synced (i.e., any delta
1361 	 * from cpu0 accounted for) as soon as we can, because many
1362 	 * many things use gethrtime/pc_gethrestime, including
1363 	 * interrupts, cmn_err, etc.
1364 	 */
1365 
1366 	/* Let cpu0 continue into tsc_sync_master() */
1367 	CPUSET_ATOMIC_ADD(procset, cp->cpu_id);
1368 
1369 #ifndef __xpv
1370 	if (tsc_gethrtime_enable)
1371 		tsc_sync_slave();
1372 #endif
1373 
1374 	/*
1375 	 * Once this was done from assembly, but it's safer here; if
1376 	 * it blocks, we need to be able to swtch() to and from, and
1377 	 * since we get here by calling t_pc, we need to do that call
1378 	 * before swtch() overwrites it.
1379 	 */
1380 
1381 	(void) (*ap_mlsetup)();
1382 
1383 	new_x86_feature = cpuid_pass1(cp);
1384 
1385 #ifndef __xpv
1386 	/*
1387 	 * We need to Sync MTRR with cpu0's MTRR. We have to do
1388 	 * this with interrupts disabled.
1389 	 */
1390 	if (x86_feature & X86_MTRR)
1391 		mtrr_sync();
1392 #endif
1393 
1394 	/*
1395 	 * Set up TSC_AUX to contain the cpuid for this processor
1396 	 * for the rdtscp instruction.
1397 	 */
1398 	if (x86_feature & X86_TSCP)
1399 		(void) wrmsr(MSR_AMD_TSCAUX, cp->cpu_id);
1400 
1401 	/*
1402 	 * Initialize this CPU's syscall handlers
1403 	 */
1404 	init_cpu_syscall(cp);
1405 
1406 	/*
1407 	 * Enable interrupts with spl set to LOCK_LEVEL. LOCK_LEVEL is the
1408 	 * highest level at which a routine is permitted to block on
1409 	 * an adaptive mutex (allows for cpu poke interrupt in case
1410 	 * the cpu is blocked on a mutex and halts). Setting LOCK_LEVEL blocks
1411 	 * device interrupts that may end up in the hat layer issuing cross
1412 	 * calls before CPU_READY is set.
1413 	 */
1414 	splx(ipltospl(LOCK_LEVEL));
1415 	sti();
1416 
1417 	/*
1418 	 * Do a sanity check to make sure this new CPU is a sane thing
1419 	 * to add to the collection of processors running this system.
1420 	 *
1421 	 * XXX	Clearly this needs to get more sophisticated, if x86
1422 	 * systems start to get built out of heterogenous CPUs; as is
1423 	 * likely to happen once the number of processors in a configuration
1424 	 * gets large enough.
1425 	 */
1426 	if ((x86_feature & new_x86_feature) != x86_feature) {
1427 		cmn_err(CE_CONT, "?cpu%d: %b\n",
1428 		    cp->cpu_id, new_x86_feature, FMT_X86_FEATURE);
1429 		cmn_err(CE_WARN, "cpu%d feature mismatch", cp->cpu_id);
1430 	}
1431 
1432 	/*
1433 	 * We do not support cpus with mixed monitor/mwait support if the
1434 	 * boot cpu supports monitor/mwait.
1435 	 */
1436 	if ((x86_feature & ~new_x86_feature) & X86_MWAIT)
1437 		panic("unsupported mixed cpu monitor/mwait support detected");
1438 
1439 	/*
1440 	 * We could be more sophisticated here, and just mark the CPU
1441 	 * as "faulted" but at this point we'll opt for the easier
1442 	 * answer of dieing horribly.  Provided the boot cpu is ok,
1443 	 * the system can be recovered by booting with use_mp set to zero.
1444 	 */
1445 	if (workaround_errata(cp) != 0)
1446 		panic("critical workaround(s) missing for cpu%d", cp->cpu_id);
1447 
1448 	cpuid_pass2(cp);
1449 	cpuid_pass3(cp);
1450 	(void) cpuid_pass4(cp);
1451 
1452 	init_cpu_info(cp);
1453 
1454 	mutex_enter(&cpu_lock);
1455 	/*
1456 	 * Processor group initialization for this CPU is dependent on the
1457 	 * cpuid probing, which must be done in the context of the current
1458 	 * CPU.
1459 	 */
1460 	pghw_physid_create(cp);
1461 	pg_cpu_init(cp);
1462 	pg_cmt_cpu_startup(cp);
1463 
1464 	cp->cpu_flags |= CPU_RUNNING | CPU_READY | CPU_ENABLE | CPU_EXISTS;
1465 	cpu_add_active(cp);
1466 
1467 	if (dtrace_cpu_init != NULL) {
1468 		(*dtrace_cpu_init)(cp->cpu_id);
1469 	}
1470 
1471 #if !defined(__xpv)
1472 	/*
1473 	 * Fill out cpu_ucode_info.  Update microcode if necessary.
1474 	 */
1475 	ucode_check(cp);
1476 #endif
1477 
1478 	mutex_exit(&cpu_lock);
1479 
1480 	/*
1481 	 * Enable preemption here so that contention for any locks acquired
1482 	 * later in mp_startup may be preempted if the thread owning those
1483 	 * locks is continously executing on other CPUs (for example, this
1484 	 * CPU must be preemptible to allow other CPUs to pause it during their
1485 	 * startup phases).  It's safe to enable preemption here because the
1486 	 * CPU state is pretty-much fully constructed.
1487 	 */
1488 	curthread->t_preempt = 0;
1489 
1490 	add_cpunode2devtree(cp->cpu_id, cp->cpu_m.mcpu_cpi);
1491 
1492 	/* The base spl should still be at LOCK LEVEL here */
1493 	ASSERT(cp->cpu_base_spl == ipltospl(LOCK_LEVEL));
1494 	set_base_spl();		/* Restore the spl to its proper value */
1495 
1496 	(void) spl0();				/* enable interrupts */
1497 
1498 	/*
1499 	 * Set up the CPU module for this CPU.  This can't be done before
1500 	 * this CPU is made CPU_READY, because we may (in heterogeneous systems)
1501 	 * need to go load another CPU module.  The act of attempting to load
1502 	 * a module may trigger a cross-call, which will ASSERT unless this
1503 	 * cpu is CPU_READY.
1504 	 */
1505 	cmi_init();
1506 
1507 	if (x86_feature & X86_MCA)
1508 		cmi_mca_init();
1509 
1510 	if (boothowto & RB_DEBUG)
1511 		kdi_cpu_init();
1512 
1513 	/*
1514 	 * Setting the bit in cpu_ready_set must be the last operation in
1515 	 * processor initialization; the boot CPU will continue to boot once
1516 	 * it sees this bit set for all active CPUs.
1517 	 */
1518 	CPUSET_ATOMIC_ADD(cpu_ready_set, cp->cpu_id);
1519 
1520 	/*
1521 	 * Because mp_startup() gets fired off after init() starts, we
1522 	 * can't use the '?' trick to do 'boot -v' printing - so we
1523 	 * always direct the 'cpu .. online' messages to the log.
1524 	 */
1525 	cmn_err(CE_CONT, "!cpu%d initialization complete - online\n",
1526 	    cp->cpu_id);
1527 
1528 	/*
1529 	 * Now we are done with the startup thread, so free it up.
1530 	 */
1531 	thread_exit();
1532 	panic("mp_startup: cannot return");
1533 	/*NOTREACHED*/
1534 }
1535 
1536 
1537 /*
1538  * Start CPU on user request.
1539  */
1540 /* ARGSUSED */
1541 int
1542 mp_cpu_start(struct cpu *cp)
1543 {
1544 	ASSERT(MUTEX_HELD(&cpu_lock));
1545 	return (0);
1546 }
1547 
1548 /*
1549  * Stop CPU on user request.
1550  */
1551 /* ARGSUSED */
1552 int
1553 mp_cpu_stop(struct cpu *cp)
1554 {
1555 	extern int cbe_psm_timer_mode;
1556 	ASSERT(MUTEX_HELD(&cpu_lock));
1557 
1558 #ifdef __xpv
1559 	/*
1560 	 * We can't offline vcpu0.
1561 	 */
1562 	if (cp->cpu_id == 0)
1563 		return (EBUSY);
1564 #endif
1565 
1566 	/*
1567 	 * If TIMER_PERIODIC mode is used, CPU0 is the one running it;
1568 	 * can't stop it.  (This is true only for machines with no TSC.)
1569 	 */
1570 
1571 	if ((cbe_psm_timer_mode == TIMER_PERIODIC) && (cp->cpu_id == 0))
1572 		return (EBUSY);
1573 
1574 	return (0);
1575 }
1576 
1577 /*
1578  * Take the specified CPU out of participation in interrupts.
1579  */
1580 int
1581 cpu_disable_intr(struct cpu *cp)
1582 {
1583 	if (psm_disable_intr(cp->cpu_id) != DDI_SUCCESS)
1584 		return (EBUSY);
1585 
1586 	cp->cpu_flags &= ~CPU_ENABLE;
1587 	return (0);
1588 }
1589 
1590 /*
1591  * Allow the specified CPU to participate in interrupts.
1592  */
1593 void
1594 cpu_enable_intr(struct cpu *cp)
1595 {
1596 	ASSERT(MUTEX_HELD(&cpu_lock));
1597 	cp->cpu_flags |= CPU_ENABLE;
1598 	psm_enable_intr(cp->cpu_id);
1599 }
1600 
1601 
1602 
1603 void
1604 mp_cpu_faulted_enter(struct cpu *cp)
1605 {
1606 	cmi_faulted_enter(cp);
1607 }
1608 
1609 void
1610 mp_cpu_faulted_exit(struct cpu *cp)
1611 {
1612 	cmi_faulted_exit(cp);
1613 }
1614 
1615 /*
1616  * The following two routines are used as context operators on threads belonging
1617  * to processes with a private LDT (see sysi86).  Due to the rarity of such
1618  * processes, these routines are currently written for best code readability and
1619  * organization rather than speed.  We could avoid checking x86_feature at every
1620  * context switch by installing different context ops, depending on the
1621  * x86_feature flags, at LDT creation time -- one for each combination of fast
1622  * syscall feature flags.
1623  */
1624 
1625 /*ARGSUSED*/
1626 void
1627 cpu_fast_syscall_disable(void *arg)
1628 {
1629 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP))
1630 		cpu_sep_disable();
1631 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC))
1632 		cpu_asysc_disable();
1633 }
1634 
1635 /*ARGSUSED*/
1636 void
1637 cpu_fast_syscall_enable(void *arg)
1638 {
1639 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP))
1640 		cpu_sep_enable();
1641 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC))
1642 		cpu_asysc_enable();
1643 }
1644 
1645 static void
1646 cpu_sep_enable(void)
1647 {
1648 	ASSERT(x86_feature & X86_SEP);
1649 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1650 
1651 	wrmsr(MSR_INTC_SEP_CS, (uint64_t)(uintptr_t)KCS_SEL);
1652 }
1653 
1654 static void
1655 cpu_sep_disable(void)
1656 {
1657 	ASSERT(x86_feature & X86_SEP);
1658 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1659 
1660 	/*
1661 	 * Setting the SYSENTER_CS_MSR register to 0 causes software executing
1662 	 * the sysenter or sysexit instruction to trigger a #gp fault.
1663 	 */
1664 	wrmsr(MSR_INTC_SEP_CS, 0);
1665 }
1666 
1667 static void
1668 cpu_asysc_enable(void)
1669 {
1670 	ASSERT(x86_feature & X86_ASYSC);
1671 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1672 
1673 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) |
1674 	    (uint64_t)(uintptr_t)AMD_EFER_SCE);
1675 }
1676 
1677 static void
1678 cpu_asysc_disable(void)
1679 {
1680 	ASSERT(x86_feature & X86_ASYSC);
1681 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1682 
1683 	/*
1684 	 * Turn off the SCE (syscall enable) bit in the EFER register. Software
1685 	 * executing syscall or sysret with this bit off will incur a #ud trap.
1686 	 */
1687 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) &
1688 	    ~((uint64_t)(uintptr_t)AMD_EFER_SCE));
1689 }
1690