xref: /titanic_51/usr/src/uts/i86pc/os/mp_startup.c (revision 8a8d276f8934b606981d1c46b57c2eda30c67610)
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 #include <sys/cpu_module.h>
70 
71 struct cpu	cpus[1];			/* CPU data */
72 struct cpu	*cpu[NCPU] = {&cpus[0]};	/* pointers to all CPUs */
73 cpu_core_t	cpu_core[NCPU];			/* cpu_core structures */
74 
75 /*
76  * Useful for disabling MP bring-up on a MP capable system.
77  */
78 int use_mp = 1;
79 
80 /*
81  * to be set by a PSM to indicate what cpus
82  * are sitting around on the system.
83  */
84 cpuset_t mp_cpus;
85 
86 /*
87  * This variable is used by the hat layer to decide whether or not
88  * critical sections are needed to prevent race conditions.  For sun4m,
89  * this variable is set once enough MP initialization has been done in
90  * order to allow cross calls.
91  */
92 int flushes_require_xcalls;
93 cpuset_t cpu_ready_set = 1;
94 
95 static 	void	mp_startup(void);
96 
97 static void cpu_sep_enable(void);
98 static void cpu_sep_disable(void);
99 static void cpu_asysc_enable(void);
100 static void cpu_asysc_disable(void);
101 
102 extern int tsc_gethrtime_enable;
103 
104 /*
105  * Init CPU info - get CPU type info for processor_info system call.
106  */
107 void
108 init_cpu_info(struct cpu *cp)
109 {
110 	processor_info_t *pi = &cp->cpu_type_info;
111 	char buf[CPU_IDSTRLEN];
112 
113 	/*
114 	 * Get clock-frequency property for the CPU.
115 	 */
116 	pi->pi_clock = cpu_freq;
117 
118 	(void) strcpy(pi->pi_processor_type, "i386");
119 	if (fpu_exists)
120 		(void) strcpy(pi->pi_fputypes, "i387 compatible");
121 
122 	(void) cpuid_getidstr(cp, buf, sizeof (buf));
123 
124 	cp->cpu_idstr = kmem_alloc(strlen(buf) + 1, KM_SLEEP);
125 	(void) strcpy(cp->cpu_idstr, buf);
126 
127 	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_idstr);
128 
129 	(void) cpuid_getbrandstr(cp, buf, sizeof (buf));
130 	cp->cpu_brandstr = kmem_alloc(strlen(buf) + 1, KM_SLEEP);
131 	(void) strcpy(cp->cpu_brandstr, buf);
132 
133 	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_brandstr);
134 }
135 
136 /*
137  * Configure syscall support on this CPU.
138  */
139 /*ARGSUSED*/
140 static void
141 init_cpu_syscall(struct cpu *cp)
142 {
143 	kpreempt_disable();
144 
145 #if defined(__amd64)
146 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC)) {
147 
148 #if !defined(__lint)
149 		/*
150 		 * The syscall instruction imposes a certain ordering on
151 		 * segment selectors, so we double-check that ordering
152 		 * here.
153 		 */
154 		ASSERT(KDS_SEL == KCS_SEL + 8);
155 		ASSERT(UDS_SEL == U32CS_SEL + 8);
156 		ASSERT(UCS_SEL == U32CS_SEL + 16);
157 #endif
158 		/*
159 		 * Turn syscall/sysret extensions on.
160 		 */
161 		cpu_asysc_enable();
162 
163 		/*
164 		 * Program the magic registers ..
165 		 */
166 		wrmsr(MSR_AMD_STAR,
167 		    ((uint64_t)(U32CS_SEL << 16 | KCS_SEL)) << 32);
168 		wrmsr(MSR_AMD_LSTAR, (uint64_t)(uintptr_t)sys_syscall);
169 		wrmsr(MSR_AMD_CSTAR, (uint64_t)(uintptr_t)sys_syscall32);
170 
171 		/*
172 		 * This list of flags is masked off the incoming
173 		 * %rfl when we enter the kernel.
174 		 */
175 		wrmsr(MSR_AMD_SFMASK, (uint64_t)(uintptr_t)(PS_IE | PS_T));
176 	}
177 #endif
178 
179 	/*
180 	 * On 32-bit kernels, we use sysenter/sysexit because it's too
181 	 * hard to use syscall/sysret, and it is more portable anyway.
182 	 *
183 	 * On 64-bit kernels on Nocona machines, the 32-bit syscall
184 	 * variant isn't available to 32-bit applications, but sysenter is.
185 	 */
186 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP)) {
187 
188 #if !defined(__lint)
189 		/*
190 		 * The sysenter instruction imposes a certain ordering on
191 		 * segment selectors, so we double-check that ordering
192 		 * here. See "sysenter" in Intel document 245471-012, "IA-32
193 		 * Intel Architecture Software Developer's Manual Volume 2:
194 		 * Instruction Set Reference"
195 		 */
196 		ASSERT(KDS_SEL == KCS_SEL + 8);
197 
198 		ASSERT32(UCS_SEL == ((KCS_SEL + 16) | 3));
199 		ASSERT32(UDS_SEL == UCS_SEL + 8);
200 
201 		ASSERT64(U32CS_SEL == ((KCS_SEL + 16) | 3));
202 		ASSERT64(UDS_SEL == U32CS_SEL + 8);
203 #endif
204 
205 		cpu_sep_enable();
206 
207 		/*
208 		 * resume() sets this value to the base of the threads stack
209 		 * via a context handler.
210 		 */
211 		wrmsr(MSR_INTC_SEP_ESP, 0);
212 		wrmsr(MSR_INTC_SEP_EIP, (uint64_t)(uintptr_t)sys_sysenter);
213 	}
214 
215 	kpreempt_enable();
216 }
217 
218 /*
219  * Multiprocessor initialization.
220  *
221  * Allocate and initialize the cpu structure, TRAPTRACE buffer, and the
222  * startup and idle threads for the specified CPU.
223  */
224 struct cpu *
225 mp_startup_init(int cpun)
226 {
227 	struct cpu *cp;
228 	kthread_id_t tp;
229 	caddr_t	sp;
230 	proc_t *procp;
231 	extern void idle();
232 
233 #ifdef TRAPTRACE
234 	trap_trace_ctl_t *ttc = &trap_trace_ctl[cpun];
235 #endif
236 
237 	ASSERT(cpun < NCPU && cpu[cpun] == NULL);
238 
239 	cp = kmem_zalloc(sizeof (*cp), KM_SLEEP);
240 	if (x86_feature & X86_MWAIT)
241 		cp->cpu_m.mcpu_mwait = mach_alloc_mwait(CPU);
242 
243 	procp = curthread->t_procp;
244 
245 	mutex_enter(&cpu_lock);
246 	/*
247 	 * Initialize the dispatcher first.
248 	 */
249 	disp_cpu_init(cp);
250 	mutex_exit(&cpu_lock);
251 
252 	cpu_vm_data_init(cp);
253 
254 	/*
255 	 * Allocate and initialize the startup thread for this CPU.
256 	 * Interrupt and process switch stacks get allocated later
257 	 * when the CPU starts running.
258 	 */
259 	tp = thread_create(NULL, 0, NULL, NULL, 0, procp,
260 	    TS_STOPPED, maxclsyspri);
261 
262 	/*
263 	 * Set state to TS_ONPROC since this thread will start running
264 	 * as soon as the CPU comes online.
265 	 *
266 	 * All the other fields of the thread structure are setup by
267 	 * thread_create().
268 	 */
269 	THREAD_ONPROC(tp, cp);
270 	tp->t_preempt = 1;
271 	tp->t_bound_cpu = cp;
272 	tp->t_affinitycnt = 1;
273 	tp->t_cpu = cp;
274 	tp->t_disp_queue = cp->cpu_disp;
275 
276 	/*
277 	 * Setup thread to start in mp_startup.
278 	 */
279 	sp = tp->t_stk;
280 	tp->t_pc = (uintptr_t)mp_startup;
281 	tp->t_sp = (uintptr_t)(sp - MINFRAME);
282 #if defined(__amd64)
283 	tp->t_sp -= STACK_ENTRY_ALIGN;		/* fake a call */
284 #endif
285 
286 	cp->cpu_id = cpun;
287 	cp->cpu_self = cp;
288 	cp->cpu_thread = tp;
289 	cp->cpu_lwp = NULL;
290 	cp->cpu_dispthread = tp;
291 	cp->cpu_dispatch_pri = DISP_PRIO(tp);
292 
293 	/*
294 	 * cpu_base_spl must be set explicitly here to prevent any blocking
295 	 * operations in mp_startup from causing the spl of the cpu to drop
296 	 * to 0 (allowing device interrupts before we're ready) in resume().
297 	 * cpu_base_spl MUST remain at LOCK_LEVEL until the cpu is CPU_READY.
298 	 * As an extra bit of security on DEBUG kernels, this is enforced with
299 	 * an assertion in mp_startup() -- before cpu_base_spl is set to its
300 	 * proper value.
301 	 */
302 	cp->cpu_base_spl = ipltospl(LOCK_LEVEL);
303 
304 	/*
305 	 * Now, initialize per-CPU idle thread for this CPU.
306 	 */
307 	tp = thread_create(NULL, PAGESIZE, idle, NULL, 0, procp, TS_ONPROC, -1);
308 
309 	cp->cpu_idle_thread = tp;
310 
311 	tp->t_preempt = 1;
312 	tp->t_bound_cpu = cp;
313 	tp->t_affinitycnt = 1;
314 	tp->t_cpu = cp;
315 	tp->t_disp_queue = cp->cpu_disp;
316 
317 	/*
318 	 * Bootstrap the CPU's PG data
319 	 */
320 	pg_cpu_bootstrap(cp);
321 
322 	/*
323 	 * Perform CPC initialization on the new CPU.
324 	 */
325 	kcpc_hw_init(cp);
326 
327 	/*
328 	 * Allocate virtual addresses for cpu_caddr1 and cpu_caddr2
329 	 * for each CPU.
330 	 */
331 	setup_vaddr_for_ppcopy(cp);
332 
333 	/*
334 	 * Allocate page for new GDT and initialize from current GDT.
335 	 */
336 #if !defined(__lint)
337 	ASSERT((sizeof (*cp->cpu_gdt) * NGDT) <= PAGESIZE);
338 #endif
339 	cp->cpu_m.mcpu_gdt = kmem_zalloc(PAGESIZE, KM_SLEEP);
340 	bcopy(CPU->cpu_m.mcpu_gdt, cp->cpu_m.mcpu_gdt,
341 	    (sizeof (*cp->cpu_m.mcpu_gdt) * NGDT));
342 
343 #if defined(__i386)
344 	/*
345 	 * setup kernel %gs.
346 	 */
347 	set_usegd(&cp->cpu_gdt[GDT_GS], cp, sizeof (struct cpu) -1, SDT_MEMRWA,
348 	    SEL_KPL, 0, 1);
349 #endif
350 
351 	/*
352 	 * If we have more than one node, each cpu gets a copy of IDT
353 	 * local to its node. If this is a Pentium box, we use cpu 0's
354 	 * IDT. cpu 0's IDT has been made read-only to workaround the
355 	 * cmpxchgl register bug
356 	 */
357 	if (system_hardware.hd_nodes && x86_type != X86_TYPE_P5) {
358 		struct machcpu *mcpu = &cp->cpu_m;
359 
360 		mcpu->mcpu_idt = kmem_alloc(sizeof (idt0), KM_SLEEP);
361 		bcopy(idt0, mcpu->mcpu_idt, sizeof (idt0));
362 	} else {
363 		cp->cpu_m.mcpu_idt = CPU->cpu_m.mcpu_idt;
364 	}
365 
366 	/*
367 	 * Get interrupt priority data from cpu 0.
368 	 */
369 	cp->cpu_pri_data = CPU->cpu_pri_data;
370 
371 	/*
372 	 * alloc space for cpuid info
373 	 */
374 	cpuid_alloc_space(cp);
375 
376 	/*
377 	 * alloc space for ucode_info
378 	 */
379 	ucode_alloc_space(cp);
380 
381 	hat_cpu_online(cp);
382 
383 #ifdef TRAPTRACE
384 	/*
385 	 * If this is a TRAPTRACE kernel, allocate TRAPTRACE buffers
386 	 */
387 	ttc->ttc_first = (uintptr_t)kmem_zalloc(trap_trace_bufsize, KM_SLEEP);
388 	ttc->ttc_next = ttc->ttc_first;
389 	ttc->ttc_limit = ttc->ttc_first + trap_trace_bufsize;
390 #endif
391 	/*
392 	 * Record that we have another CPU.
393 	 */
394 	mutex_enter(&cpu_lock);
395 	/*
396 	 * Initialize the interrupt threads for this CPU
397 	 */
398 	cpu_intr_alloc(cp, NINTR_THREADS);
399 	/*
400 	 * Add CPU to list of available CPUs.  It'll be on the active list
401 	 * after mp_startup().
402 	 */
403 	cpu_add_unit(cp);
404 	mutex_exit(&cpu_lock);
405 
406 	return (cp);
407 }
408 
409 /*
410  * Undo what was done in mp_startup_init
411  */
412 static void
413 mp_startup_fini(struct cpu *cp, int error)
414 {
415 	mutex_enter(&cpu_lock);
416 
417 	/*
418 	 * Remove the CPU from the list of available CPUs.
419 	 */
420 	cpu_del_unit(cp->cpu_id);
421 
422 	if (error == ETIMEDOUT) {
423 		/*
424 		 * The cpu was started, but never *seemed* to run any
425 		 * code in the kernel; it's probably off spinning in its
426 		 * own private world, though with potential references to
427 		 * our kmem-allocated IDTs and GDTs (for example).
428 		 *
429 		 * Worse still, it may actually wake up some time later,
430 		 * so rather than guess what it might or might not do, we
431 		 * leave the fundamental data structures intact.
432 		 */
433 		cp->cpu_flags = 0;
434 		mutex_exit(&cpu_lock);
435 		return;
436 	}
437 
438 	/*
439 	 * At this point, the only threads bound to this CPU should
440 	 * special per-cpu threads: it's idle thread, it's pause threads,
441 	 * and it's interrupt threads.  Clean these up.
442 	 */
443 	cpu_destroy_bound_threads(cp);
444 	cp->cpu_idle_thread = NULL;
445 
446 	/*
447 	 * Free the interrupt stack.
448 	 */
449 	segkp_release(segkp,
450 	    cp->cpu_intr_stack - (INTR_STACK_SIZE - SA(MINFRAME)));
451 
452 	mutex_exit(&cpu_lock);
453 
454 #ifdef TRAPTRACE
455 	/*
456 	 * Discard the trap trace buffer
457 	 */
458 	{
459 		trap_trace_ctl_t *ttc = &trap_trace_ctl[cp->cpu_id];
460 
461 		kmem_free((void *)ttc->ttc_first, trap_trace_bufsize);
462 		ttc->ttc_first = NULL;
463 	}
464 #endif
465 
466 	hat_cpu_offline(cp);
467 
468 	cpuid_free_space(cp);
469 
470 	ucode_free_space(cp);
471 
472 	if (cp->cpu_m.mcpu_idt != CPU->cpu_m.mcpu_idt)
473 		kmem_free(cp->cpu_m.mcpu_idt, sizeof (idt0));
474 	cp->cpu_m.mcpu_idt = NULL;
475 
476 	kmem_free(cp->cpu_m.mcpu_gdt, PAGESIZE);
477 	cp->cpu_m.mcpu_gdt = NULL;
478 
479 	teardown_vaddr_for_ppcopy(cp);
480 
481 	kcpc_hw_fini(cp);
482 
483 	cp->cpu_dispthread = NULL;
484 	cp->cpu_thread = NULL;	/* discarded by cpu_destroy_bound_threads() */
485 
486 	cpu_vm_data_destroy(cp);
487 
488 	mutex_enter(&cpu_lock);
489 	disp_cpu_fini(cp);
490 	mutex_exit(&cpu_lock);
491 
492 	kmem_free(cp, sizeof (*cp));
493 }
494 
495 /*
496  * Apply workarounds for known errata, and warn about those that are absent.
497  *
498  * System vendors occasionally create configurations which contain different
499  * revisions of the CPUs that are almost but not exactly the same.  At the
500  * time of writing, this meant that their clock rates were the same, their
501  * feature sets were the same, but the required workaround were -not-
502  * necessarily the same.  So, this routine is invoked on -every- CPU soon
503  * after starting to make sure that the resulting system contains the most
504  * pessimal set of workarounds needed to cope with *any* of the CPUs in the
505  * system.
506  *
507  * workaround_errata is invoked early in mlsetup() for CPU 0, and in
508  * mp_startup() for all slave CPUs. Slaves process workaround_errata prior
509  * to acknowledging their readiness to the master, so this routine will
510  * never be executed by multiple CPUs in parallel, thus making updates to
511  * global data safe.
512  *
513  * These workarounds are based on Rev 3.57 of the Revision Guide for
514  * AMD Athlon(tm) 64 and AMD Opteron(tm) Processors, August 2005.
515  */
516 
517 #if defined(OPTERON_ERRATUM_88)
518 int opteron_erratum_88;		/* if non-zero -> at least one cpu has it */
519 #endif
520 
521 #if defined(OPTERON_ERRATUM_91)
522 int opteron_erratum_91;		/* if non-zero -> at least one cpu has it */
523 #endif
524 
525 #if defined(OPTERON_ERRATUM_93)
526 int opteron_erratum_93;		/* if non-zero -> at least one cpu has it */
527 #endif
528 
529 #if defined(OPTERON_ERRATUM_95)
530 int opteron_erratum_95;		/* if non-zero -> at least one cpu has it */
531 #endif
532 
533 #if defined(OPTERON_ERRATUM_100)
534 int opteron_erratum_100;	/* if non-zero -> at least one cpu has it */
535 #endif
536 
537 #if defined(OPTERON_ERRATUM_108)
538 int opteron_erratum_108;	/* if non-zero -> at least one cpu has it */
539 #endif
540 
541 #if defined(OPTERON_ERRATUM_109)
542 int opteron_erratum_109;	/* if non-zero -> at least one cpu has it */
543 #endif
544 
545 #if defined(OPTERON_ERRATUM_121)
546 int opteron_erratum_121;	/* if non-zero -> at least one cpu has it */
547 #endif
548 
549 #if defined(OPTERON_ERRATUM_122)
550 int opteron_erratum_122;	/* if non-zero -> at least one cpu has it */
551 #endif
552 
553 #if defined(OPTERON_ERRATUM_123)
554 int opteron_erratum_123;	/* if non-zero -> at least one cpu has it */
555 #endif
556 
557 #if defined(OPTERON_ERRATUM_131)
558 int opteron_erratum_131;	/* if non-zero -> at least one cpu has it */
559 #endif
560 
561 #if defined(OPTERON_WORKAROUND_6336786)
562 int opteron_workaround_6336786;	/* non-zero -> WA relevant and applied */
563 int opteron_workaround_6336786_UP = 0;	/* Not needed for UP */
564 #endif
565 
566 #if defined(OPTERON_WORKAROUND_6323525)
567 int opteron_workaround_6323525;	/* if non-zero -> at least one cpu has it */
568 #endif
569 
570 static void
571 workaround_warning(cpu_t *cp, uint_t erratum)
572 {
573 	cmn_err(CE_WARN, "cpu%d: no workaround for erratum %u",
574 	    cp->cpu_id, erratum);
575 }
576 
577 static void
578 workaround_applied(uint_t erratum)
579 {
580 	if (erratum > 1000000)
581 		cmn_err(CE_CONT, "?workaround applied for cpu issue #%d\n",
582 		    erratum);
583 	else
584 		cmn_err(CE_CONT, "?workaround applied for cpu erratum #%d\n",
585 		    erratum);
586 }
587 
588 static void
589 msr_warning(cpu_t *cp, const char *rw, uint_t msr, int error)
590 {
591 	cmn_err(CE_WARN, "cpu%d: couldn't %smsr 0x%x, error %d",
592 	    cp->cpu_id, rw, msr, error);
593 }
594 
595 uint_t
596 workaround_errata(struct cpu *cpu)
597 {
598 	uint_t missing = 0;
599 
600 	ASSERT(cpu == CPU);
601 
602 	/*LINTED*/
603 	if (cpuid_opteron_erratum(cpu, 88) > 0) {
604 		/*
605 		 * SWAPGS May Fail To Read Correct GS Base
606 		 */
607 #if defined(OPTERON_ERRATUM_88)
608 		/*
609 		 * The workaround is an mfence in the relevant assembler code
610 		 */
611 		opteron_erratum_88++;
612 #else
613 		workaround_warning(cpu, 88);
614 		missing++;
615 #endif
616 	}
617 
618 	if (cpuid_opteron_erratum(cpu, 91) > 0) {
619 		/*
620 		 * Software Prefetches May Report A Page Fault
621 		 */
622 #if defined(OPTERON_ERRATUM_91)
623 		/*
624 		 * fix is in trap.c
625 		 */
626 		opteron_erratum_91++;
627 #else
628 		workaround_warning(cpu, 91);
629 		missing++;
630 #endif
631 	}
632 
633 	if (cpuid_opteron_erratum(cpu, 93) > 0) {
634 		/*
635 		 * RSM Auto-Halt Restart Returns to Incorrect RIP
636 		 */
637 #if defined(OPTERON_ERRATUM_93)
638 		/*
639 		 * fix is in trap.c
640 		 */
641 		opteron_erratum_93++;
642 #else
643 		workaround_warning(cpu, 93);
644 		missing++;
645 #endif
646 	}
647 
648 	/*LINTED*/
649 	if (cpuid_opteron_erratum(cpu, 95) > 0) {
650 		/*
651 		 * RET Instruction May Return to Incorrect EIP
652 		 */
653 #if defined(OPTERON_ERRATUM_95)
654 #if defined(_LP64)
655 		/*
656 		 * Workaround this by ensuring that 32-bit user code and
657 		 * 64-bit kernel code never occupy the same address
658 		 * range mod 4G.
659 		 */
660 		if (_userlimit32 > 0xc0000000ul)
661 			*(uintptr_t *)&_userlimit32 = 0xc0000000ul;
662 
663 		/*LINTED*/
664 		ASSERT((uint32_t)COREHEAP_BASE == 0xc0000000u);
665 		opteron_erratum_95++;
666 #endif	/* _LP64 */
667 #else
668 		workaround_warning(cpu, 95);
669 		missing++;
670 #endif
671 	}
672 
673 	if (cpuid_opteron_erratum(cpu, 100) > 0) {
674 		/*
675 		 * Compatibility Mode Branches Transfer to Illegal Address
676 		 */
677 #if defined(OPTERON_ERRATUM_100)
678 		/*
679 		 * fix is in trap.c
680 		 */
681 		opteron_erratum_100++;
682 #else
683 		workaround_warning(cpu, 100);
684 		missing++;
685 #endif
686 	}
687 
688 	/*LINTED*/
689 	if (cpuid_opteron_erratum(cpu, 108) > 0) {
690 		/*
691 		 * CPUID Instruction May Return Incorrect Model Number In
692 		 * Some Processors
693 		 */
694 #if defined(OPTERON_ERRATUM_108)
695 		/*
696 		 * (Our cpuid-handling code corrects the model number on
697 		 * those processors)
698 		 */
699 #else
700 		workaround_warning(cpu, 108);
701 		missing++;
702 #endif
703 	}
704 
705 	/*LINTED*/
706 	if (cpuid_opteron_erratum(cpu, 109) > 0) do {
707 		/*
708 		 * Certain Reverse REP MOVS May Produce Unpredictable Behaviour
709 		 */
710 #if defined(OPTERON_ERRATUM_109)
711 		/*
712 		 * The "workaround" is to print a warning to upgrade the BIOS
713 		 */
714 		uint64_t value;
715 		const uint_t msr = MSR_AMD_PATCHLEVEL;
716 		int err;
717 
718 		if ((err = checked_rdmsr(msr, &value)) != 0) {
719 			msr_warning(cpu, "rd", msr, err);
720 			workaround_warning(cpu, 109);
721 			missing++;
722 		}
723 		if (value == 0)
724 			opteron_erratum_109++;
725 #else
726 		workaround_warning(cpu, 109);
727 		missing++;
728 #endif
729 	/*CONSTANTCONDITION*/
730 	} while (0);
731 
732 	/*LINTED*/
733 	if (cpuid_opteron_erratum(cpu, 121) > 0) {
734 		/*
735 		 * Sequential Execution Across Non_Canonical Boundary Caused
736 		 * Processor Hang
737 		 */
738 #if defined(OPTERON_ERRATUM_121)
739 #if defined(_LP64)
740 		/*
741 		 * Erratum 121 is only present in long (64 bit) mode.
742 		 * Workaround is to include the page immediately before the
743 		 * va hole to eliminate the possibility of system hangs due to
744 		 * sequential execution across the va hole boundary.
745 		 */
746 		if (opteron_erratum_121)
747 			opteron_erratum_121++;
748 		else {
749 			if (hole_start) {
750 				hole_start -= PAGESIZE;
751 			} else {
752 				/*
753 				 * hole_start not yet initialized by
754 				 * mmu_init. Initialize hole_start
755 				 * with value to be subtracted.
756 				 */
757 				hole_start = PAGESIZE;
758 			}
759 			opteron_erratum_121++;
760 		}
761 #endif	/* _LP64 */
762 #else
763 		workaround_warning(cpu, 121);
764 		missing++;
765 #endif
766 	}
767 
768 	/*LINTED*/
769 	if (cpuid_opteron_erratum(cpu, 122) > 0) do {
770 		/*
771 		 * TLB Flush Filter May Cause Coherency Problem in
772 		 * Multiprocessor Systems
773 		 */
774 #if defined(OPTERON_ERRATUM_122)
775 		uint64_t value;
776 		const uint_t msr = MSR_AMD_HWCR;
777 		int error;
778 
779 		/*
780 		 * Erratum 122 is only present in MP configurations (multi-core
781 		 * or multi-processor).
782 		 */
783 		if (!opteron_erratum_122 && lgrp_plat_node_cnt == 1 &&
784 		    cpuid_get_ncpu_per_chip(cpu) == 1)
785 			break;
786 
787 		/* disable TLB Flush Filter */
788 
789 		if ((error = checked_rdmsr(msr, &value)) != 0) {
790 			msr_warning(cpu, "rd", msr, error);
791 			workaround_warning(cpu, 122);
792 			missing++;
793 		} else {
794 			value |= (uint64_t)AMD_HWCR_FFDIS;
795 			if ((error = checked_wrmsr(msr, value)) != 0) {
796 				msr_warning(cpu, "wr", msr, error);
797 				workaround_warning(cpu, 122);
798 				missing++;
799 			}
800 		}
801 		opteron_erratum_122++;
802 #else
803 		workaround_warning(cpu, 122);
804 		missing++;
805 #endif
806 	/*CONSTANTCONDITION*/
807 	} while (0);
808 
809 	/*LINTED*/
810 	if (cpuid_opteron_erratum(cpu, 123) > 0) do {
811 		/*
812 		 * Bypassed Reads May Cause Data Corruption of System Hang in
813 		 * Dual Core Processors
814 		 */
815 #if defined(OPTERON_ERRATUM_123)
816 		uint64_t value;
817 		const uint_t msr = MSR_AMD_PATCHLEVEL;
818 		int err;
819 
820 		/*
821 		 * Erratum 123 applies only to multi-core cpus.
822 		 */
823 		if (cpuid_get_ncpu_per_chip(cpu) < 2)
824 			break;
825 
826 		/*
827 		 * The "workaround" is to print a warning to upgrade the BIOS
828 		 */
829 		if ((err = checked_rdmsr(msr, &value)) != 0) {
830 			msr_warning(cpu, "rd", msr, err);
831 			workaround_warning(cpu, 123);
832 			missing++;
833 		}
834 		if (value == 0)
835 			opteron_erratum_123++;
836 #else
837 		workaround_warning(cpu, 123);
838 		missing++;
839 
840 #endif
841 	/*CONSTANTCONDITION*/
842 	} while (0);
843 
844 	/*LINTED*/
845 	if (cpuid_opteron_erratum(cpu, 131) > 0) do {
846 		/*
847 		 * Multiprocessor Systems with Four or More Cores May Deadlock
848 		 * Waiting for a Probe Response
849 		 */
850 #if defined(OPTERON_ERRATUM_131)
851 		uint64_t nbcfg;
852 		const uint_t msr = MSR_AMD_NB_CFG;
853 		const uint64_t wabits =
854 		    AMD_NB_CFG_SRQ_HEARTBEAT | AMD_NB_CFG_SRQ_SPR;
855 		int error;
856 
857 		/*
858 		 * Erratum 131 applies to any system with four or more cores.
859 		 */
860 		if (opteron_erratum_131)
861 			break;
862 
863 		if (lgrp_plat_node_cnt * cpuid_get_ncpu_per_chip(cpu) < 4)
864 			break;
865 
866 		/*
867 		 * Print a warning if neither of the workarounds for
868 		 * erratum 131 is present.
869 		 */
870 		if ((error = checked_rdmsr(msr, &nbcfg)) != 0) {
871 			msr_warning(cpu, "rd", msr, error);
872 			workaround_warning(cpu, 131);
873 			missing++;
874 		} else if ((nbcfg & wabits) == 0) {
875 			opteron_erratum_131++;
876 		} else {
877 			/* cannot have both workarounds set */
878 			ASSERT((nbcfg & wabits) != wabits);
879 		}
880 #else
881 		workaround_warning(cpu, 131);
882 		missing++;
883 #endif
884 	/*CONSTANTCONDITION*/
885 	} while (0);
886 
887 	/*
888 	 * This isn't really an erratum, but for convenience the
889 	 * detection/workaround code lives here and in cpuid_opteron_erratum.
890 	 */
891 	if (cpuid_opteron_erratum(cpu, 6336786) > 0) {
892 #if defined(OPTERON_WORKAROUND_6336786)
893 		/*
894 		 * Disable C1-Clock ramping on multi-core/multi-processor
895 		 * K8 platforms to guard against TSC drift.
896 		 */
897 		if (opteron_workaround_6336786) {
898 			opteron_workaround_6336786++;
899 		} else if ((lgrp_plat_node_cnt *
900 		    cpuid_get_ncpu_per_chip(cpu) > 1) ||
901 		    opteron_workaround_6336786_UP) {
902 			int	node;
903 			uint8_t data;
904 
905 			for (node = 0; node < lgrp_plat_node_cnt; node++) {
906 				/*
907 				 * Clear PMM7[1:0] (function 3, offset 0x87)
908 				 * Northbridge device is the node id + 24.
909 				 */
910 				data = pci_getb_func(0, node + 24, 3, 0x87);
911 				data &= 0xFC;
912 				pci_putb_func(0, node + 24, 3, 0x87, data);
913 			}
914 			opteron_workaround_6336786++;
915 		}
916 #else
917 		workaround_warning(cpu, 6336786);
918 		missing++;
919 #endif
920 	}
921 
922 	/*LINTED*/
923 	/*
924 	 * Mutex primitives don't work as expected.
925 	 */
926 	if (cpuid_opteron_erratum(cpu, 6323525) > 0) {
927 #if defined(OPTERON_WORKAROUND_6323525)
928 		/*
929 		 * This problem only occurs with 2 or more cores. If bit in
930 		 * MSR_BU_CFG set, then not applicable. The workaround
931 		 * is to patch the semaphone routines with the lfence
932 		 * instruction to provide necessary load memory barrier with
933 		 * possible subsequent read-modify-write ops.
934 		 *
935 		 * It is too early in boot to call the patch routine so
936 		 * set erratum variable to be done in startup_end().
937 		 */
938 		if (opteron_workaround_6323525) {
939 			opteron_workaround_6323525++;
940 		} else if ((x86_feature & X86_SSE2) && ((lgrp_plat_node_cnt *
941 		    cpuid_get_ncpu_per_chip(cpu)) > 1)) {
942 			if ((xrdmsr(MSR_BU_CFG) & 0x02) == 0)
943 				opteron_workaround_6323525++;
944 		}
945 #else
946 		workaround_warning(cpu, 6323525);
947 		missing++;
948 #endif
949 	}
950 
951 	return (missing);
952 }
953 
954 void
955 workaround_errata_end()
956 {
957 #if defined(OPTERON_ERRATUM_88)
958 	if (opteron_erratum_88)
959 		workaround_applied(88);
960 #endif
961 #if defined(OPTERON_ERRATUM_91)
962 	if (opteron_erratum_91)
963 		workaround_applied(91);
964 #endif
965 #if defined(OPTERON_ERRATUM_93)
966 	if (opteron_erratum_93)
967 		workaround_applied(93);
968 #endif
969 #if defined(OPTERON_ERRATUM_95)
970 	if (opteron_erratum_95)
971 		workaround_applied(95);
972 #endif
973 #if defined(OPTERON_ERRATUM_100)
974 	if (opteron_erratum_100)
975 		workaround_applied(100);
976 #endif
977 #if defined(OPTERON_ERRATUM_108)
978 	if (opteron_erratum_108)
979 		workaround_applied(108);
980 #endif
981 #if defined(OPTERON_ERRATUM_109)
982 	if (opteron_erratum_109) {
983 		cmn_err(CE_WARN,
984 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
985 		    " processor\nerratum 109 was not detected; updating your"
986 		    " system's BIOS to a version\ncontaining this"
987 		    " microcode patch is HIGHLY recommended or erroneous"
988 		    " system\noperation may occur.\n");
989 	}
990 #endif
991 #if defined(OPTERON_ERRATUM_121)
992 	if (opteron_erratum_121)
993 		workaround_applied(121);
994 #endif
995 #if defined(OPTERON_ERRATUM_122)
996 	if (opteron_erratum_122)
997 		workaround_applied(122);
998 #endif
999 #if defined(OPTERON_ERRATUM_123)
1000 	if (opteron_erratum_123) {
1001 		cmn_err(CE_WARN,
1002 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1003 		    " processor\nerratum 123 was not detected; updating your"
1004 		    " system's BIOS to a version\ncontaining this"
1005 		    " microcode patch is HIGHLY recommended or erroneous"
1006 		    " system\noperation may occur.\n");
1007 	}
1008 #endif
1009 #if defined(OPTERON_ERRATUM_131)
1010 	if (opteron_erratum_131) {
1011 		cmn_err(CE_WARN,
1012 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1013 		    " processor\nerratum 131 was not detected; updating your"
1014 		    " system's BIOS to a version\ncontaining this"
1015 		    " microcode patch is HIGHLY recommended or erroneous"
1016 		    " system\noperation may occur.\n");
1017 	}
1018 #endif
1019 #if defined(OPTERON_WORKAROUND_6336786)
1020 	if (opteron_workaround_6336786)
1021 		workaround_applied(6336786);
1022 #endif
1023 #if defined(OPTERON_WORKAROUND_6323525)
1024 	if (opteron_workaround_6323525)
1025 		workaround_applied(6323525);
1026 #endif
1027 }
1028 
1029 static cpuset_t procset;
1030 
1031 /*
1032  * Start a single cpu, assuming that the kernel context is available
1033  * to successfully start another cpu.
1034  *
1035  * (For example, real mode code is mapped into the right place
1036  * in memory and is ready to be run.)
1037  */
1038 int
1039 start_cpu(processorid_t who)
1040 {
1041 	void *ctx;
1042 	cpu_t *cp;
1043 	int delays;
1044 	int error = 0;
1045 
1046 	ASSERT(who != 0);
1047 
1048 	/*
1049 	 * Check if there's at least a Mbyte of kmem available
1050 	 * before attempting to start the cpu.
1051 	 */
1052 	if (kmem_avail() < 1024 * 1024) {
1053 		/*
1054 		 * Kick off a reap in case that helps us with
1055 		 * later attempts ..
1056 		 */
1057 		kmem_reap();
1058 		return (ENOMEM);
1059 	}
1060 
1061 	cp = mp_startup_init(who);
1062 	if ((ctx = mach_cpucontext_alloc(cp)) == NULL ||
1063 	    (error = mach_cpu_start(cp, ctx)) != 0) {
1064 
1065 		/*
1066 		 * Something went wrong before we even started it
1067 		 */
1068 		if (ctx)
1069 			cmn_err(CE_WARN,
1070 			    "cpu%d: failed to start error %d",
1071 			    cp->cpu_id, error);
1072 		else
1073 			cmn_err(CE_WARN,
1074 			    "cpu%d: failed to allocate context", cp->cpu_id);
1075 
1076 		if (ctx)
1077 			mach_cpucontext_free(cp, ctx, error);
1078 		else
1079 			error = EAGAIN;		/* hmm. */
1080 		mp_startup_fini(cp, error);
1081 		return (error);
1082 	}
1083 
1084 	for (delays = 0; !CPU_IN_SET(procset, who); delays++) {
1085 		if (delays == 500) {
1086 			/*
1087 			 * After five seconds, things are probably looking
1088 			 * a bit bleak - explain the hang.
1089 			 */
1090 			cmn_err(CE_NOTE, "cpu%d: started, "
1091 			    "but not running in the kernel yet", who);
1092 		} else if (delays > 2000) {
1093 			/*
1094 			 * We waited at least 20 seconds, bail ..
1095 			 */
1096 			error = ETIMEDOUT;
1097 			cmn_err(CE_WARN, "cpu%d: timed out", who);
1098 			mach_cpucontext_free(cp, ctx, error);
1099 			mp_startup_fini(cp, error);
1100 			return (error);
1101 		}
1102 
1103 		/*
1104 		 * wait at least 10ms, then check again..
1105 		 */
1106 		delay(USEC_TO_TICK_ROUNDUP(10000));
1107 	}
1108 
1109 	mach_cpucontext_free(cp, ctx, 0);
1110 
1111 	if (tsc_gethrtime_enable)
1112 		tsc_sync_master(who);
1113 
1114 	if (dtrace_cpu_init != NULL) {
1115 		/*
1116 		 * DTrace CPU initialization expects cpu_lock to be held.
1117 		 */
1118 		mutex_enter(&cpu_lock);
1119 		(*dtrace_cpu_init)(who);
1120 		mutex_exit(&cpu_lock);
1121 	}
1122 
1123 	while (!CPU_IN_SET(cpu_ready_set, who))
1124 		delay(1);
1125 
1126 	return (0);
1127 }
1128 
1129 
1130 /*ARGSUSED*/
1131 void
1132 start_other_cpus(int cprboot)
1133 {
1134 	uint_t who;
1135 	uint_t skipped = 0;
1136 	uint_t bootcpuid = 0;
1137 
1138 	/*
1139 	 * Initialize our own cpu_info.
1140 	 */
1141 	init_cpu_info(CPU);
1142 
1143 	/*
1144 	 * Initialize our syscall handlers
1145 	 */
1146 	init_cpu_syscall(CPU);
1147 
1148 	/*
1149 	 * Take the boot cpu out of the mp_cpus set because we know
1150 	 * it's already running.  Add it to the cpu_ready_set for
1151 	 * precisely the same reason.
1152 	 */
1153 	CPUSET_DEL(mp_cpus, bootcpuid);
1154 	CPUSET_ADD(cpu_ready_set, bootcpuid);
1155 
1156 	/*
1157 	 * if only 1 cpu or not using MP, skip the rest of this
1158 	 */
1159 	if (CPUSET_ISNULL(mp_cpus) || use_mp == 0) {
1160 		if (use_mp == 0)
1161 			cmn_err(CE_CONT, "?***** Not in MP mode\n");
1162 		goto done;
1163 	}
1164 
1165 	/*
1166 	 * perform such initialization as is needed
1167 	 * to be able to take CPUs on- and off-line.
1168 	 */
1169 	cpu_pause_init();
1170 
1171 	xc_init();		/* initialize processor crosscalls */
1172 
1173 	if (mach_cpucontext_init() != 0)
1174 		goto done;
1175 
1176 	flushes_require_xcalls = 1;
1177 
1178 	/*
1179 	 * We lock our affinity to the master CPU to ensure that all slave CPUs
1180 	 * do their TSC syncs with the same CPU.
1181 	 */
1182 	affinity_set(CPU_CURRENT);
1183 
1184 	for (who = 0; who < NCPU; who++) {
1185 
1186 		if (!CPU_IN_SET(mp_cpus, who))
1187 			continue;
1188 		ASSERT(who != bootcpuid);
1189 		if (ncpus >= max_ncpus) {
1190 			skipped = who;
1191 			continue;
1192 		}
1193 		if (start_cpu(who) != 0)
1194 			CPUSET_DEL(mp_cpus, who);
1195 	}
1196 
1197 	/* Free the space allocated to hold the microcode file */
1198 	ucode_free();
1199 
1200 	affinity_clear();
1201 
1202 	if (skipped) {
1203 		cmn_err(CE_NOTE,
1204 		    "System detected %d cpus, but "
1205 		    "only %d cpu(s) were enabled during boot.",
1206 		    skipped + 1, ncpus);
1207 		cmn_err(CE_NOTE,
1208 		    "Use \"boot-ncpus\" parameter to enable more CPU(s). "
1209 		    "See eeprom(1M).");
1210 	}
1211 
1212 done:
1213 	workaround_errata_end();
1214 	mach_cpucontext_fini();
1215 
1216 	cmi_post_mpstartup();
1217 }
1218 
1219 /*
1220  * Dummy functions - no i86pc platforms support dynamic cpu allocation.
1221  */
1222 /*ARGSUSED*/
1223 int
1224 mp_cpu_configure(int cpuid)
1225 {
1226 	return (ENOTSUP);		/* not supported */
1227 }
1228 
1229 /*ARGSUSED*/
1230 int
1231 mp_cpu_unconfigure(int cpuid)
1232 {
1233 	return (ENOTSUP);		/* not supported */
1234 }
1235 
1236 /*
1237  * Startup function for 'other' CPUs (besides boot cpu).
1238  * Called from real_mode_start.
1239  *
1240  * WARNING: until CPU_READY is set, mp_startup and routines called by
1241  * mp_startup should not call routines (e.g. kmem_free) that could call
1242  * hat_unload which requires CPU_READY to be set.
1243  */
1244 void
1245 mp_startup(void)
1246 {
1247 	struct cpu *cp = CPU;
1248 	uint_t new_x86_feature;
1249 
1250 	/*
1251 	 * We need to get TSC on this proc synced (i.e., any delta
1252 	 * from cpu0 accounted for) as soon as we can, because many
1253 	 * many things use gethrtime/pc_gethrestime, including
1254 	 * interrupts, cmn_err, etc.
1255 	 */
1256 
1257 	/* Let cpu0 continue into tsc_sync_master() */
1258 	CPUSET_ATOMIC_ADD(procset, cp->cpu_id);
1259 
1260 	if (tsc_gethrtime_enable)
1261 		tsc_sync_slave();
1262 
1263 	/*
1264 	 * Once this was done from assembly, but it's safer here; if
1265 	 * it blocks, we need to be able to swtch() to and from, and
1266 	 * since we get here by calling t_pc, we need to do that call
1267 	 * before swtch() overwrites it.
1268 	 */
1269 
1270 	(void) (*ap_mlsetup)();
1271 
1272 	new_x86_feature = cpuid_pass1(cp);
1273 
1274 	/*
1275 	 * We need to Sync MTRR with cpu0's MTRR. We have to do
1276 	 * this with interrupts disabled.
1277 	 */
1278 	if (x86_feature & X86_MTRR)
1279 		mtrr_sync();
1280 
1281 	/*
1282 	 * Set up TSC_AUX to contain the cpuid for this processor
1283 	 * for the rdtscp instruction.
1284 	 */
1285 	if (x86_feature & X86_TSCP)
1286 		(void) wrmsr(MSR_AMD_TSCAUX, cp->cpu_id);
1287 
1288 	/*
1289 	 * Initialize this CPU's syscall handlers
1290 	 */
1291 	init_cpu_syscall(cp);
1292 
1293 	/*
1294 	 * Enable interrupts with spl set to LOCK_LEVEL. LOCK_LEVEL is the
1295 	 * highest level at which a routine is permitted to block on
1296 	 * an adaptive mutex (allows for cpu poke interrupt in case
1297 	 * the cpu is blocked on a mutex and halts). Setting LOCK_LEVEL blocks
1298 	 * device interrupts that may end up in the hat layer issuing cross
1299 	 * calls before CPU_READY is set.
1300 	 */
1301 	splx(ipltospl(LOCK_LEVEL));
1302 	sti();
1303 
1304 	/*
1305 	 * Do a sanity check to make sure this new CPU is a sane thing
1306 	 * to add to the collection of processors running this system.
1307 	 *
1308 	 * XXX	Clearly this needs to get more sophisticated, if x86
1309 	 * systems start to get built out of heterogenous CPUs; as is
1310 	 * likely to happen once the number of processors in a configuration
1311 	 * gets large enough.
1312 	 */
1313 	if ((x86_feature & new_x86_feature) != x86_feature) {
1314 		cmn_err(CE_CONT, "?cpu%d: %b\n",
1315 		    cp->cpu_id, new_x86_feature, FMT_X86_FEATURE);
1316 		cmn_err(CE_WARN, "cpu%d feature mismatch", cp->cpu_id);
1317 	}
1318 
1319 	/*
1320 	 * We do not support cpus with mixed monitor/mwait support if the
1321 	 * boot cpu supports monitor/mwait.
1322 	 */
1323 	if ((x86_feature & ~new_x86_feature) & X86_MWAIT)
1324 		panic("unsupported mixed cpu monitor/mwait support detected");
1325 
1326 	/*
1327 	 * We could be more sophisticated here, and just mark the CPU
1328 	 * as "faulted" but at this point we'll opt for the easier
1329 	 * answer of dieing horribly.  Provided the boot cpu is ok,
1330 	 * the system can be recovered by booting with use_mp set to zero.
1331 	 */
1332 	if (workaround_errata(cp) != 0)
1333 		panic("critical workaround(s) missing for cpu%d", cp->cpu_id);
1334 
1335 	cpuid_pass2(cp);
1336 	cpuid_pass3(cp);
1337 	(void) cpuid_pass4(cp);
1338 
1339 	init_cpu_info(cp);
1340 
1341 	mutex_enter(&cpu_lock);
1342 	/*
1343 	 * Processor group initialization for this CPU is dependent on the
1344 	 * cpuid probing, which must be done in the context of the current
1345 	 * CPU.
1346 	 */
1347 	pghw_physid_create(cp);
1348 	pg_cpu_init(cp);
1349 	pg_cmt_cpu_startup(cp);
1350 
1351 	cp->cpu_flags |= CPU_RUNNING | CPU_READY | CPU_ENABLE | CPU_EXISTS;
1352 	cpu_add_active(cp);
1353 
1354 	if (dtrace_cpu_init != NULL) {
1355 		(*dtrace_cpu_init)(cp->cpu_id);
1356 	}
1357 
1358 	/*
1359 	 * Fill out cpu_ucode_info.  Update microcode if necessary.
1360 	 */
1361 	ucode_check(cp);
1362 
1363 	mutex_exit(&cpu_lock);
1364 
1365 	/*
1366 	 * Enable preemption here so that contention for any locks acquired
1367 	 * later in mp_startup may be preempted if the thread owning those
1368 	 * locks is continously executing on other CPUs (for example, this
1369 	 * CPU must be preemptible to allow other CPUs to pause it during their
1370 	 * startup phases).  It's safe to enable preemption here because the
1371 	 * CPU state is pretty-much fully constructed.
1372 	 */
1373 	curthread->t_preempt = 0;
1374 
1375 	add_cpunode2devtree(cp->cpu_id, cp->cpu_m.mcpu_cpi);
1376 
1377 	/* The base spl should still be at LOCK LEVEL here */
1378 	ASSERT(cp->cpu_base_spl == ipltospl(LOCK_LEVEL));
1379 	set_base_spl();		/* Restore the spl to its proper value */
1380 
1381 	(void) spl0();				/* enable interrupts */
1382 
1383 	/*
1384 	 * Set up the CPU module for this CPU.  This can't be done before
1385 	 * this CPU is made CPU_READY, because we may (in heterogeneous systems)
1386 	 * need to go load another CPU module.  The act of attempting to load
1387 	 * a module may trigger a cross-call, which will ASSERT unless this
1388 	 * cpu is CPU_READY.
1389 	 */
1390 	cmi_init();
1391 
1392 	if (x86_feature & X86_MCA)
1393 		cmi_mca_init();
1394 
1395 	if (boothowto & RB_DEBUG)
1396 		kdi_cpu_init();
1397 
1398 	/*
1399 	 * Setting the bit in cpu_ready_set must be the last operation in
1400 	 * processor initialization; the boot CPU will continue to boot once
1401 	 * it sees this bit set for all active CPUs.
1402 	 */
1403 	CPUSET_ATOMIC_ADD(cpu_ready_set, cp->cpu_id);
1404 
1405 	/*
1406 	 * Because mp_startup() gets fired off after init() starts, we
1407 	 * can't use the '?' trick to do 'boot -v' printing - so we
1408 	 * always direct the 'cpu .. online' messages to the log.
1409 	 */
1410 	cmn_err(CE_CONT, "!cpu%d initialization complete - online\n",
1411 	    cp->cpu_id);
1412 
1413 	/*
1414 	 * Now we are done with the startup thread, so free it up.
1415 	 */
1416 	thread_exit();
1417 	panic("mp_startup: cannot return");
1418 	/*NOTREACHED*/
1419 }
1420 
1421 
1422 /*
1423  * Start CPU on user request.
1424  */
1425 /* ARGSUSED */
1426 int
1427 mp_cpu_start(struct cpu *cp)
1428 {
1429 	ASSERT(MUTEX_HELD(&cpu_lock));
1430 	return (0);
1431 }
1432 
1433 /*
1434  * Stop CPU on user request.
1435  */
1436 /* ARGSUSED */
1437 int
1438 mp_cpu_stop(struct cpu *cp)
1439 {
1440 	extern int cbe_psm_timer_mode;
1441 	ASSERT(MUTEX_HELD(&cpu_lock));
1442 
1443 	/*
1444 	 * If TIMER_PERIODIC mode is used, CPU0 is the one running it;
1445 	 * can't stop it.  (This is true only for machines with no TSC.)
1446 	 */
1447 
1448 	if ((cbe_psm_timer_mode == TIMER_PERIODIC) && (cp->cpu_id == 0))
1449 		return (1);
1450 
1451 	return (0);
1452 }
1453 
1454 /*
1455  * Take the specified CPU out of participation in interrupts.
1456  */
1457 int
1458 cpu_disable_intr(struct cpu *cp)
1459 {
1460 	if (psm_disable_intr(cp->cpu_id) != DDI_SUCCESS)
1461 		return (EBUSY);
1462 
1463 	cp->cpu_flags &= ~CPU_ENABLE;
1464 	return (0);
1465 }
1466 
1467 /*
1468  * Allow the specified CPU to participate in interrupts.
1469  */
1470 void
1471 cpu_enable_intr(struct cpu *cp)
1472 {
1473 	ASSERT(MUTEX_HELD(&cpu_lock));
1474 	cp->cpu_flags |= CPU_ENABLE;
1475 	psm_enable_intr(cp->cpu_id);
1476 }
1477 
1478 
1479 
1480 void
1481 mp_cpu_faulted_enter(struct cpu *cp)
1482 {
1483 	cmi_faulted_enter(cp);
1484 }
1485 
1486 void
1487 mp_cpu_faulted_exit(struct cpu *cp)
1488 {
1489 	cmi_faulted_exit(cp);
1490 }
1491 
1492 /*
1493  * The following two routines are used as context operators on threads belonging
1494  * to processes with a private LDT (see sysi86).  Due to the rarity of such
1495  * processes, these routines are currently written for best code readability and
1496  * organization rather than speed.  We could avoid checking x86_feature at every
1497  * context switch by installing different context ops, depending on the
1498  * x86_feature flags, at LDT creation time -- one for each combination of fast
1499  * syscall feature flags.
1500  */
1501 
1502 /*ARGSUSED*/
1503 void
1504 cpu_fast_syscall_disable(void *arg)
1505 {
1506 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP))
1507 		cpu_sep_disable();
1508 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC))
1509 		cpu_asysc_disable();
1510 }
1511 
1512 /*ARGSUSED*/
1513 void
1514 cpu_fast_syscall_enable(void *arg)
1515 {
1516 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP))
1517 		cpu_sep_enable();
1518 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC))
1519 		cpu_asysc_enable();
1520 }
1521 
1522 static void
1523 cpu_sep_enable(void)
1524 {
1525 	ASSERT(x86_feature & X86_SEP);
1526 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1527 
1528 	wrmsr(MSR_INTC_SEP_CS, (uint64_t)(uintptr_t)KCS_SEL);
1529 }
1530 
1531 static void
1532 cpu_sep_disable(void)
1533 {
1534 	ASSERT(x86_feature & X86_SEP);
1535 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1536 
1537 	/*
1538 	 * Setting the SYSENTER_CS_MSR register to 0 causes software executing
1539 	 * the sysenter or sysexit instruction to trigger a #gp fault.
1540 	 */
1541 	wrmsr(MSR_INTC_SEP_CS, 0);
1542 }
1543 
1544 static void
1545 cpu_asysc_enable(void)
1546 {
1547 	ASSERT(x86_feature & X86_ASYSC);
1548 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1549 
1550 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) |
1551 	    (uint64_t)(uintptr_t)AMD_EFER_SCE);
1552 }
1553 
1554 static void
1555 cpu_asysc_disable(void)
1556 {
1557 	ASSERT(x86_feature & X86_ASYSC);
1558 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1559 
1560 	/*
1561 	 * Turn off the SCE (syscall enable) bit in the EFER register. Software
1562 	 * executing syscall or sysret with this bit off will incur a #ud trap.
1563 	 */
1564 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) &
1565 	    ~((uint64_t)(uintptr_t)AMD_EFER_SCE));
1566 }
1567