xref: /titanic_50/usr/src/uts/i86pc/os/mp_startup.c (revision da2e3ebdc1edfbc5028edf1354e7dd2fa69a7968)
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 	/*
119 	 * Current frequency in Hz.
120 	 */
121 	cp->cpu_curr_clock = cpu_freq_hz;
122 
123 	/*
124 	 * Supported frequencies.
125 	 */
126 	cpu_set_supp_freqs(cp, NULL);
127 
128 	(void) strcpy(pi->pi_processor_type, "i386");
129 	if (fpu_exists)
130 		(void) strcpy(pi->pi_fputypes, "i387 compatible");
131 
132 	(void) cpuid_getidstr(cp, buf, sizeof (buf));
133 
134 	cp->cpu_idstr = kmem_alloc(strlen(buf) + 1, KM_SLEEP);
135 	(void) strcpy(cp->cpu_idstr, buf);
136 
137 	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_idstr);
138 
139 	(void) cpuid_getbrandstr(cp, buf, sizeof (buf));
140 	cp->cpu_brandstr = kmem_alloc(strlen(buf) + 1, KM_SLEEP);
141 	(void) strcpy(cp->cpu_brandstr, buf);
142 
143 	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_brandstr);
144 }
145 
146 /*
147  * Configure syscall support on this CPU.
148  */
149 /*ARGSUSED*/
150 static void
151 init_cpu_syscall(struct cpu *cp)
152 {
153 	kpreempt_disable();
154 
155 #if defined(__amd64)
156 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC)) {
157 
158 #if !defined(__lint)
159 		/*
160 		 * The syscall instruction imposes a certain ordering on
161 		 * segment selectors, so we double-check that ordering
162 		 * here.
163 		 */
164 		ASSERT(KDS_SEL == KCS_SEL + 8);
165 		ASSERT(UDS_SEL == U32CS_SEL + 8);
166 		ASSERT(UCS_SEL == U32CS_SEL + 16);
167 #endif
168 		/*
169 		 * Turn syscall/sysret extensions on.
170 		 */
171 		cpu_asysc_enable();
172 
173 		/*
174 		 * Program the magic registers ..
175 		 */
176 		wrmsr(MSR_AMD_STAR,
177 		    ((uint64_t)(U32CS_SEL << 16 | KCS_SEL)) << 32);
178 		wrmsr(MSR_AMD_LSTAR, (uint64_t)(uintptr_t)sys_syscall);
179 		wrmsr(MSR_AMD_CSTAR, (uint64_t)(uintptr_t)sys_syscall32);
180 
181 		/*
182 		 * This list of flags is masked off the incoming
183 		 * %rfl when we enter the kernel.
184 		 */
185 		wrmsr(MSR_AMD_SFMASK, (uint64_t)(uintptr_t)(PS_IE | PS_T));
186 	}
187 #endif
188 
189 	/*
190 	 * On 32-bit kernels, we use sysenter/sysexit because it's too
191 	 * hard to use syscall/sysret, and it is more portable anyway.
192 	 *
193 	 * On 64-bit kernels on Nocona machines, the 32-bit syscall
194 	 * variant isn't available to 32-bit applications, but sysenter is.
195 	 */
196 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP)) {
197 
198 #if !defined(__lint)
199 		/*
200 		 * The sysenter instruction imposes a certain ordering on
201 		 * segment selectors, so we double-check that ordering
202 		 * here. See "sysenter" in Intel document 245471-012, "IA-32
203 		 * Intel Architecture Software Developer's Manual Volume 2:
204 		 * Instruction Set Reference"
205 		 */
206 		ASSERT(KDS_SEL == KCS_SEL + 8);
207 
208 		ASSERT32(UCS_SEL == ((KCS_SEL + 16) | 3));
209 		ASSERT32(UDS_SEL == UCS_SEL + 8);
210 
211 		ASSERT64(U32CS_SEL == ((KCS_SEL + 16) | 3));
212 		ASSERT64(UDS_SEL == U32CS_SEL + 8);
213 #endif
214 
215 		cpu_sep_enable();
216 
217 		/*
218 		 * resume() sets this value to the base of the threads stack
219 		 * via a context handler.
220 		 */
221 		wrmsr(MSR_INTC_SEP_ESP, 0);
222 		wrmsr(MSR_INTC_SEP_EIP, (uint64_t)(uintptr_t)sys_sysenter);
223 	}
224 
225 	kpreempt_enable();
226 }
227 
228 /*
229  * Multiprocessor initialization.
230  *
231  * Allocate and initialize the cpu structure, TRAPTRACE buffer, and the
232  * startup and idle threads for the specified CPU.
233  */
234 struct cpu *
235 mp_startup_init(int cpun)
236 {
237 	struct cpu *cp;
238 	kthread_id_t tp;
239 	caddr_t	sp;
240 	proc_t *procp;
241 	extern void idle();
242 
243 #ifdef TRAPTRACE
244 	trap_trace_ctl_t *ttc = &trap_trace_ctl[cpun];
245 #endif
246 
247 	ASSERT(cpun < NCPU && cpu[cpun] == NULL);
248 
249 	cp = kmem_zalloc(sizeof (*cp), KM_SLEEP);
250 	if (x86_feature & X86_MWAIT)
251 		cp->cpu_m.mcpu_mwait = mach_alloc_mwait(CPU);
252 
253 	procp = curthread->t_procp;
254 
255 	mutex_enter(&cpu_lock);
256 	/*
257 	 * Initialize the dispatcher first.
258 	 */
259 	disp_cpu_init(cp);
260 	mutex_exit(&cpu_lock);
261 
262 	cpu_vm_data_init(cp);
263 
264 	/*
265 	 * Allocate and initialize the startup thread for this CPU.
266 	 * Interrupt and process switch stacks get allocated later
267 	 * when the CPU starts running.
268 	 */
269 	tp = thread_create(NULL, 0, NULL, NULL, 0, procp,
270 	    TS_STOPPED, maxclsyspri);
271 
272 	/*
273 	 * Set state to TS_ONPROC since this thread will start running
274 	 * as soon as the CPU comes online.
275 	 *
276 	 * All the other fields of the thread structure are setup by
277 	 * thread_create().
278 	 */
279 	THREAD_ONPROC(tp, cp);
280 	tp->t_preempt = 1;
281 	tp->t_bound_cpu = cp;
282 	tp->t_affinitycnt = 1;
283 	tp->t_cpu = cp;
284 	tp->t_disp_queue = cp->cpu_disp;
285 
286 	/*
287 	 * Setup thread to start in mp_startup.
288 	 */
289 	sp = tp->t_stk;
290 	tp->t_pc = (uintptr_t)mp_startup;
291 	tp->t_sp = (uintptr_t)(sp - MINFRAME);
292 #if defined(__amd64)
293 	tp->t_sp -= STACK_ENTRY_ALIGN;		/* fake a call */
294 #endif
295 
296 	cp->cpu_id = cpun;
297 	cp->cpu_self = cp;
298 	cp->cpu_thread = tp;
299 	cp->cpu_lwp = NULL;
300 	cp->cpu_dispthread = tp;
301 	cp->cpu_dispatch_pri = DISP_PRIO(tp);
302 
303 	/*
304 	 * cpu_base_spl must be set explicitly here to prevent any blocking
305 	 * operations in mp_startup from causing the spl of the cpu to drop
306 	 * to 0 (allowing device interrupts before we're ready) in resume().
307 	 * cpu_base_spl MUST remain at LOCK_LEVEL until the cpu is CPU_READY.
308 	 * As an extra bit of security on DEBUG kernels, this is enforced with
309 	 * an assertion in mp_startup() -- before cpu_base_spl is set to its
310 	 * proper value.
311 	 */
312 	cp->cpu_base_spl = ipltospl(LOCK_LEVEL);
313 
314 	/*
315 	 * Now, initialize per-CPU idle thread for this CPU.
316 	 */
317 	tp = thread_create(NULL, PAGESIZE, idle, NULL, 0, procp, TS_ONPROC, -1);
318 
319 	cp->cpu_idle_thread = tp;
320 
321 	tp->t_preempt = 1;
322 	tp->t_bound_cpu = cp;
323 	tp->t_affinitycnt = 1;
324 	tp->t_cpu = cp;
325 	tp->t_disp_queue = cp->cpu_disp;
326 
327 	/*
328 	 * Bootstrap the CPU's PG data
329 	 */
330 	pg_cpu_bootstrap(cp);
331 
332 	/*
333 	 * Perform CPC initialization on the new CPU.
334 	 */
335 	kcpc_hw_init(cp);
336 
337 	/*
338 	 * Allocate virtual addresses for cpu_caddr1 and cpu_caddr2
339 	 * for each CPU.
340 	 */
341 	setup_vaddr_for_ppcopy(cp);
342 
343 	/*
344 	 * Allocate page for new GDT and initialize from current GDT.
345 	 */
346 #if !defined(__lint)
347 	ASSERT((sizeof (*cp->cpu_gdt) * NGDT) <= PAGESIZE);
348 #endif
349 	cp->cpu_m.mcpu_gdt = kmem_zalloc(PAGESIZE, KM_SLEEP);
350 	bcopy(CPU->cpu_m.mcpu_gdt, cp->cpu_m.mcpu_gdt,
351 	    (sizeof (*cp->cpu_m.mcpu_gdt) * NGDT));
352 
353 #if defined(__i386)
354 	/*
355 	 * setup kernel %gs.
356 	 */
357 	set_usegd(&cp->cpu_gdt[GDT_GS], cp, sizeof (struct cpu) -1, SDT_MEMRWA,
358 	    SEL_KPL, 0, 1);
359 #endif
360 
361 	/*
362 	 * If we have more than one node, each cpu gets a copy of IDT
363 	 * local to its node. If this is a Pentium box, we use cpu 0's
364 	 * IDT. cpu 0's IDT has been made read-only to workaround the
365 	 * cmpxchgl register bug
366 	 */
367 	if (system_hardware.hd_nodes && x86_type != X86_TYPE_P5) {
368 		struct machcpu *mcpu = &cp->cpu_m;
369 
370 		mcpu->mcpu_idt = kmem_alloc(sizeof (idt0), KM_SLEEP);
371 		bcopy(idt0, mcpu->mcpu_idt, sizeof (idt0));
372 	} else {
373 		cp->cpu_m.mcpu_idt = CPU->cpu_m.mcpu_idt;
374 	}
375 
376 	/*
377 	 * Get interrupt priority data from cpu 0.
378 	 */
379 	cp->cpu_pri_data = CPU->cpu_pri_data;
380 
381 	/*
382 	 * alloc space for cpuid info
383 	 */
384 	cpuid_alloc_space(cp);
385 
386 	/*
387 	 * alloc space for ucode_info
388 	 */
389 	ucode_alloc_space(cp);
390 
391 	hat_cpu_online(cp);
392 
393 #ifdef TRAPTRACE
394 	/*
395 	 * If this is a TRAPTRACE kernel, allocate TRAPTRACE buffers
396 	 */
397 	ttc->ttc_first = (uintptr_t)kmem_zalloc(trap_trace_bufsize, KM_SLEEP);
398 	ttc->ttc_next = ttc->ttc_first;
399 	ttc->ttc_limit = ttc->ttc_first + trap_trace_bufsize;
400 #endif
401 	/*
402 	 * Record that we have another CPU.
403 	 */
404 	mutex_enter(&cpu_lock);
405 	/*
406 	 * Initialize the interrupt threads for this CPU
407 	 */
408 	cpu_intr_alloc(cp, NINTR_THREADS);
409 	/*
410 	 * Add CPU to list of available CPUs.  It'll be on the active list
411 	 * after mp_startup().
412 	 */
413 	cpu_add_unit(cp);
414 	mutex_exit(&cpu_lock);
415 
416 	return (cp);
417 }
418 
419 /*
420  * Undo what was done in mp_startup_init
421  */
422 static void
423 mp_startup_fini(struct cpu *cp, int error)
424 {
425 	mutex_enter(&cpu_lock);
426 
427 	/*
428 	 * Remove the CPU from the list of available CPUs.
429 	 */
430 	cpu_del_unit(cp->cpu_id);
431 
432 	if (error == ETIMEDOUT) {
433 		/*
434 		 * The cpu was started, but never *seemed* to run any
435 		 * code in the kernel; it's probably off spinning in its
436 		 * own private world, though with potential references to
437 		 * our kmem-allocated IDTs and GDTs (for example).
438 		 *
439 		 * Worse still, it may actually wake up some time later,
440 		 * so rather than guess what it might or might not do, we
441 		 * leave the fundamental data structures intact.
442 		 */
443 		cp->cpu_flags = 0;
444 		mutex_exit(&cpu_lock);
445 		return;
446 	}
447 
448 	/*
449 	 * At this point, the only threads bound to this CPU should
450 	 * special per-cpu threads: it's idle thread, it's pause threads,
451 	 * and it's interrupt threads.  Clean these up.
452 	 */
453 	cpu_destroy_bound_threads(cp);
454 	cp->cpu_idle_thread = NULL;
455 
456 	/*
457 	 * Free the interrupt stack.
458 	 */
459 	segkp_release(segkp,
460 	    cp->cpu_intr_stack - (INTR_STACK_SIZE - SA(MINFRAME)));
461 
462 	mutex_exit(&cpu_lock);
463 
464 #ifdef TRAPTRACE
465 	/*
466 	 * Discard the trap trace buffer
467 	 */
468 	{
469 		trap_trace_ctl_t *ttc = &trap_trace_ctl[cp->cpu_id];
470 
471 		kmem_free((void *)ttc->ttc_first, trap_trace_bufsize);
472 		ttc->ttc_first = NULL;
473 	}
474 #endif
475 
476 	hat_cpu_offline(cp);
477 
478 	cpuid_free_space(cp);
479 
480 	ucode_free_space(cp);
481 
482 	if (cp->cpu_m.mcpu_idt != CPU->cpu_m.mcpu_idt)
483 		kmem_free(cp->cpu_m.mcpu_idt, sizeof (idt0));
484 	cp->cpu_m.mcpu_idt = NULL;
485 
486 	kmem_free(cp->cpu_m.mcpu_gdt, PAGESIZE);
487 	cp->cpu_m.mcpu_gdt = NULL;
488 
489 	teardown_vaddr_for_ppcopy(cp);
490 
491 	kcpc_hw_fini(cp);
492 
493 	cp->cpu_dispthread = NULL;
494 	cp->cpu_thread = NULL;	/* discarded by cpu_destroy_bound_threads() */
495 
496 	cpu_vm_data_destroy(cp);
497 
498 	mutex_enter(&cpu_lock);
499 	disp_cpu_fini(cp);
500 	mutex_exit(&cpu_lock);
501 
502 	kmem_free(cp, sizeof (*cp));
503 }
504 
505 /*
506  * Apply workarounds for known errata, and warn about those that are absent.
507  *
508  * System vendors occasionally create configurations which contain different
509  * revisions of the CPUs that are almost but not exactly the same.  At the
510  * time of writing, this meant that their clock rates were the same, their
511  * feature sets were the same, but the required workaround were -not-
512  * necessarily the same.  So, this routine is invoked on -every- CPU soon
513  * after starting to make sure that the resulting system contains the most
514  * pessimal set of workarounds needed to cope with *any* of the CPUs in the
515  * system.
516  *
517  * workaround_errata is invoked early in mlsetup() for CPU 0, and in
518  * mp_startup() for all slave CPUs. Slaves process workaround_errata prior
519  * to acknowledging their readiness to the master, so this routine will
520  * never be executed by multiple CPUs in parallel, thus making updates to
521  * global data safe.
522  *
523  * These workarounds are based on Rev 3.57 of the Revision Guide for
524  * AMD Athlon(tm) 64 and AMD Opteron(tm) Processors, August 2005.
525  */
526 
527 #if defined(OPTERON_ERRATUM_88)
528 int opteron_erratum_88;		/* if non-zero -> at least one cpu has it */
529 #endif
530 
531 #if defined(OPTERON_ERRATUM_91)
532 int opteron_erratum_91;		/* if non-zero -> at least one cpu has it */
533 #endif
534 
535 #if defined(OPTERON_ERRATUM_93)
536 int opteron_erratum_93;		/* if non-zero -> at least one cpu has it */
537 #endif
538 
539 #if defined(OPTERON_ERRATUM_95)
540 int opteron_erratum_95;		/* if non-zero -> at least one cpu has it */
541 #endif
542 
543 #if defined(OPTERON_ERRATUM_100)
544 int opteron_erratum_100;	/* if non-zero -> at least one cpu has it */
545 #endif
546 
547 #if defined(OPTERON_ERRATUM_108)
548 int opteron_erratum_108;	/* if non-zero -> at least one cpu has it */
549 #endif
550 
551 #if defined(OPTERON_ERRATUM_109)
552 int opteron_erratum_109;	/* if non-zero -> at least one cpu has it */
553 #endif
554 
555 #if defined(OPTERON_ERRATUM_121)
556 int opteron_erratum_121;	/* if non-zero -> at least one cpu has it */
557 #endif
558 
559 #if defined(OPTERON_ERRATUM_122)
560 int opteron_erratum_122;	/* if non-zero -> at least one cpu has it */
561 #endif
562 
563 #if defined(OPTERON_ERRATUM_123)
564 int opteron_erratum_123;	/* if non-zero -> at least one cpu has it */
565 #endif
566 
567 #if defined(OPTERON_ERRATUM_131)
568 int opteron_erratum_131;	/* if non-zero -> at least one cpu has it */
569 #endif
570 
571 #if defined(OPTERON_WORKAROUND_6336786)
572 int opteron_workaround_6336786;	/* non-zero -> WA relevant and applied */
573 int opteron_workaround_6336786_UP = 0;	/* Not needed for UP */
574 #endif
575 
576 #if defined(OPTERON_WORKAROUND_6323525)
577 int opteron_workaround_6323525;	/* if non-zero -> at least one cpu has it */
578 #endif
579 
580 static void
581 workaround_warning(cpu_t *cp, uint_t erratum)
582 {
583 	cmn_err(CE_WARN, "cpu%d: no workaround for erratum %u",
584 	    cp->cpu_id, erratum);
585 }
586 
587 static void
588 workaround_applied(uint_t erratum)
589 {
590 	if (erratum > 1000000)
591 		cmn_err(CE_CONT, "?workaround applied for cpu issue #%d\n",
592 		    erratum);
593 	else
594 		cmn_err(CE_CONT, "?workaround applied for cpu erratum #%d\n",
595 		    erratum);
596 }
597 
598 static void
599 msr_warning(cpu_t *cp, const char *rw, uint_t msr, int error)
600 {
601 	cmn_err(CE_WARN, "cpu%d: couldn't %smsr 0x%x, error %d",
602 	    cp->cpu_id, rw, msr, error);
603 }
604 
605 uint_t
606 workaround_errata(struct cpu *cpu)
607 {
608 	uint_t missing = 0;
609 
610 	ASSERT(cpu == CPU);
611 
612 	/*LINTED*/
613 	if (cpuid_opteron_erratum(cpu, 88) > 0) {
614 		/*
615 		 * SWAPGS May Fail To Read Correct GS Base
616 		 */
617 #if defined(OPTERON_ERRATUM_88)
618 		/*
619 		 * The workaround is an mfence in the relevant assembler code
620 		 */
621 		opteron_erratum_88++;
622 #else
623 		workaround_warning(cpu, 88);
624 		missing++;
625 #endif
626 	}
627 
628 	if (cpuid_opteron_erratum(cpu, 91) > 0) {
629 		/*
630 		 * Software Prefetches May Report A Page Fault
631 		 */
632 #if defined(OPTERON_ERRATUM_91)
633 		/*
634 		 * fix is in trap.c
635 		 */
636 		opteron_erratum_91++;
637 #else
638 		workaround_warning(cpu, 91);
639 		missing++;
640 #endif
641 	}
642 
643 	if (cpuid_opteron_erratum(cpu, 93) > 0) {
644 		/*
645 		 * RSM Auto-Halt Restart Returns to Incorrect RIP
646 		 */
647 #if defined(OPTERON_ERRATUM_93)
648 		/*
649 		 * fix is in trap.c
650 		 */
651 		opteron_erratum_93++;
652 #else
653 		workaround_warning(cpu, 93);
654 		missing++;
655 #endif
656 	}
657 
658 	/*LINTED*/
659 	if (cpuid_opteron_erratum(cpu, 95) > 0) {
660 		/*
661 		 * RET Instruction May Return to Incorrect EIP
662 		 */
663 #if defined(OPTERON_ERRATUM_95)
664 #if defined(_LP64)
665 		/*
666 		 * Workaround this by ensuring that 32-bit user code and
667 		 * 64-bit kernel code never occupy the same address
668 		 * range mod 4G.
669 		 */
670 		if (_userlimit32 > 0xc0000000ul)
671 			*(uintptr_t *)&_userlimit32 = 0xc0000000ul;
672 
673 		/*LINTED*/
674 		ASSERT((uint32_t)COREHEAP_BASE == 0xc0000000u);
675 		opteron_erratum_95++;
676 #endif	/* _LP64 */
677 #else
678 		workaround_warning(cpu, 95);
679 		missing++;
680 #endif
681 	}
682 
683 	if (cpuid_opteron_erratum(cpu, 100) > 0) {
684 		/*
685 		 * Compatibility Mode Branches Transfer to Illegal Address
686 		 */
687 #if defined(OPTERON_ERRATUM_100)
688 		/*
689 		 * fix is in trap.c
690 		 */
691 		opteron_erratum_100++;
692 #else
693 		workaround_warning(cpu, 100);
694 		missing++;
695 #endif
696 	}
697 
698 	/*LINTED*/
699 	if (cpuid_opteron_erratum(cpu, 108) > 0) {
700 		/*
701 		 * CPUID Instruction May Return Incorrect Model Number In
702 		 * Some Processors
703 		 */
704 #if defined(OPTERON_ERRATUM_108)
705 		/*
706 		 * (Our cpuid-handling code corrects the model number on
707 		 * those processors)
708 		 */
709 #else
710 		workaround_warning(cpu, 108);
711 		missing++;
712 #endif
713 	}
714 
715 	/*LINTED*/
716 	if (cpuid_opteron_erratum(cpu, 109) > 0) do {
717 		/*
718 		 * Certain Reverse REP MOVS May Produce Unpredictable Behaviour
719 		 */
720 #if defined(OPTERON_ERRATUM_109)
721 		/*
722 		 * The "workaround" is to print a warning to upgrade the BIOS
723 		 */
724 		uint64_t value;
725 		const uint_t msr = MSR_AMD_PATCHLEVEL;
726 		int err;
727 
728 		if ((err = checked_rdmsr(msr, &value)) != 0) {
729 			msr_warning(cpu, "rd", msr, err);
730 			workaround_warning(cpu, 109);
731 			missing++;
732 		}
733 		if (value == 0)
734 			opteron_erratum_109++;
735 #else
736 		workaround_warning(cpu, 109);
737 		missing++;
738 #endif
739 	/*CONSTANTCONDITION*/
740 	} while (0);
741 
742 	/*LINTED*/
743 	if (cpuid_opteron_erratum(cpu, 121) > 0) {
744 		/*
745 		 * Sequential Execution Across Non_Canonical Boundary Caused
746 		 * Processor Hang
747 		 */
748 #if defined(OPTERON_ERRATUM_121)
749 #if defined(_LP64)
750 		/*
751 		 * Erratum 121 is only present in long (64 bit) mode.
752 		 * Workaround is to include the page immediately before the
753 		 * va hole to eliminate the possibility of system hangs due to
754 		 * sequential execution across the va hole boundary.
755 		 */
756 		if (opteron_erratum_121)
757 			opteron_erratum_121++;
758 		else {
759 			if (hole_start) {
760 				hole_start -= PAGESIZE;
761 			} else {
762 				/*
763 				 * hole_start not yet initialized by
764 				 * mmu_init. Initialize hole_start
765 				 * with value to be subtracted.
766 				 */
767 				hole_start = PAGESIZE;
768 			}
769 			opteron_erratum_121++;
770 		}
771 #endif	/* _LP64 */
772 #else
773 		workaround_warning(cpu, 121);
774 		missing++;
775 #endif
776 	}
777 
778 	/*LINTED*/
779 	if (cpuid_opteron_erratum(cpu, 122) > 0) do {
780 		/*
781 		 * TLB Flush Filter May Cause Coherency Problem in
782 		 * Multiprocessor Systems
783 		 */
784 #if defined(OPTERON_ERRATUM_122)
785 		uint64_t value;
786 		const uint_t msr = MSR_AMD_HWCR;
787 		int error;
788 
789 		/*
790 		 * Erratum 122 is only present in MP configurations (multi-core
791 		 * or multi-processor).
792 		 */
793 		if (!opteron_erratum_122 && lgrp_plat_node_cnt == 1 &&
794 		    cpuid_get_ncpu_per_chip(cpu) == 1)
795 			break;
796 
797 		/* disable TLB Flush Filter */
798 
799 		if ((error = checked_rdmsr(msr, &value)) != 0) {
800 			msr_warning(cpu, "rd", msr, error);
801 			workaround_warning(cpu, 122);
802 			missing++;
803 		} else {
804 			value |= (uint64_t)AMD_HWCR_FFDIS;
805 			if ((error = checked_wrmsr(msr, value)) != 0) {
806 				msr_warning(cpu, "wr", msr, error);
807 				workaround_warning(cpu, 122);
808 				missing++;
809 			}
810 		}
811 		opteron_erratum_122++;
812 #else
813 		workaround_warning(cpu, 122);
814 		missing++;
815 #endif
816 	/*CONSTANTCONDITION*/
817 	} while (0);
818 
819 	/*LINTED*/
820 	if (cpuid_opteron_erratum(cpu, 123) > 0) do {
821 		/*
822 		 * Bypassed Reads May Cause Data Corruption of System Hang in
823 		 * Dual Core Processors
824 		 */
825 #if defined(OPTERON_ERRATUM_123)
826 		uint64_t value;
827 		const uint_t msr = MSR_AMD_PATCHLEVEL;
828 		int err;
829 
830 		/*
831 		 * Erratum 123 applies only to multi-core cpus.
832 		 */
833 		if (cpuid_get_ncpu_per_chip(cpu) < 2)
834 			break;
835 
836 		/*
837 		 * The "workaround" is to print a warning to upgrade the BIOS
838 		 */
839 		if ((err = checked_rdmsr(msr, &value)) != 0) {
840 			msr_warning(cpu, "rd", msr, err);
841 			workaround_warning(cpu, 123);
842 			missing++;
843 		}
844 		if (value == 0)
845 			opteron_erratum_123++;
846 #else
847 		workaround_warning(cpu, 123);
848 		missing++;
849 
850 #endif
851 	/*CONSTANTCONDITION*/
852 	} while (0);
853 
854 	/*LINTED*/
855 	if (cpuid_opteron_erratum(cpu, 131) > 0) do {
856 		/*
857 		 * Multiprocessor Systems with Four or More Cores May Deadlock
858 		 * Waiting for a Probe Response
859 		 */
860 #if defined(OPTERON_ERRATUM_131)
861 		uint64_t nbcfg;
862 		const uint_t msr = MSR_AMD_NB_CFG;
863 		const uint64_t wabits =
864 		    AMD_NB_CFG_SRQ_HEARTBEAT | AMD_NB_CFG_SRQ_SPR;
865 		int error;
866 
867 		/*
868 		 * Erratum 131 applies to any system with four or more cores.
869 		 */
870 		if (opteron_erratum_131)
871 			break;
872 
873 		if (lgrp_plat_node_cnt * cpuid_get_ncpu_per_chip(cpu) < 4)
874 			break;
875 
876 		/*
877 		 * Print a warning if neither of the workarounds for
878 		 * erratum 131 is present.
879 		 */
880 		if ((error = checked_rdmsr(msr, &nbcfg)) != 0) {
881 			msr_warning(cpu, "rd", msr, error);
882 			workaround_warning(cpu, 131);
883 			missing++;
884 		} else if ((nbcfg & wabits) == 0) {
885 			opteron_erratum_131++;
886 		} else {
887 			/* cannot have both workarounds set */
888 			ASSERT((nbcfg & wabits) != wabits);
889 		}
890 #else
891 		workaround_warning(cpu, 131);
892 		missing++;
893 #endif
894 	/*CONSTANTCONDITION*/
895 	} while (0);
896 
897 	/*
898 	 * This isn't really an erratum, but for convenience the
899 	 * detection/workaround code lives here and in cpuid_opteron_erratum.
900 	 */
901 	if (cpuid_opteron_erratum(cpu, 6336786) > 0) {
902 #if defined(OPTERON_WORKAROUND_6336786)
903 		/*
904 		 * Disable C1-Clock ramping on multi-core/multi-processor
905 		 * K8 platforms to guard against TSC drift.
906 		 */
907 		if (opteron_workaround_6336786) {
908 			opteron_workaround_6336786++;
909 		} else if ((lgrp_plat_node_cnt *
910 		    cpuid_get_ncpu_per_chip(cpu) > 1) ||
911 		    opteron_workaround_6336786_UP) {
912 			int	node;
913 			uint8_t data;
914 
915 			for (node = 0; node < lgrp_plat_node_cnt; node++) {
916 				/*
917 				 * Clear PMM7[1:0] (function 3, offset 0x87)
918 				 * Northbridge device is the node id + 24.
919 				 */
920 				data = pci_getb_func(0, node + 24, 3, 0x87);
921 				data &= 0xFC;
922 				pci_putb_func(0, node + 24, 3, 0x87, data);
923 			}
924 			opteron_workaround_6336786++;
925 		}
926 #else
927 		workaround_warning(cpu, 6336786);
928 		missing++;
929 #endif
930 	}
931 
932 	/*LINTED*/
933 	/*
934 	 * Mutex primitives don't work as expected.
935 	 */
936 	if (cpuid_opteron_erratum(cpu, 6323525) > 0) {
937 #if defined(OPTERON_WORKAROUND_6323525)
938 		/*
939 		 * This problem only occurs with 2 or more cores. If bit in
940 		 * MSR_BU_CFG set, then not applicable. The workaround
941 		 * is to patch the semaphone routines with the lfence
942 		 * instruction to provide necessary load memory barrier with
943 		 * possible subsequent read-modify-write ops.
944 		 *
945 		 * It is too early in boot to call the patch routine so
946 		 * set erratum variable to be done in startup_end().
947 		 */
948 		if (opteron_workaround_6323525) {
949 			opteron_workaround_6323525++;
950 		} else if ((x86_feature & X86_SSE2) && ((lgrp_plat_node_cnt *
951 		    cpuid_get_ncpu_per_chip(cpu)) > 1)) {
952 			if ((xrdmsr(MSR_BU_CFG) & 0x02) == 0)
953 				opteron_workaround_6323525++;
954 		}
955 #else
956 		workaround_warning(cpu, 6323525);
957 		missing++;
958 #endif
959 	}
960 
961 	return (missing);
962 }
963 
964 void
965 workaround_errata_end()
966 {
967 #if defined(OPTERON_ERRATUM_88)
968 	if (opteron_erratum_88)
969 		workaround_applied(88);
970 #endif
971 #if defined(OPTERON_ERRATUM_91)
972 	if (opteron_erratum_91)
973 		workaround_applied(91);
974 #endif
975 #if defined(OPTERON_ERRATUM_93)
976 	if (opteron_erratum_93)
977 		workaround_applied(93);
978 #endif
979 #if defined(OPTERON_ERRATUM_95)
980 	if (opteron_erratum_95)
981 		workaround_applied(95);
982 #endif
983 #if defined(OPTERON_ERRATUM_100)
984 	if (opteron_erratum_100)
985 		workaround_applied(100);
986 #endif
987 #if defined(OPTERON_ERRATUM_108)
988 	if (opteron_erratum_108)
989 		workaround_applied(108);
990 #endif
991 #if defined(OPTERON_ERRATUM_109)
992 	if (opteron_erratum_109) {
993 		cmn_err(CE_WARN,
994 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
995 		    " processor\nerratum 109 was not detected; updating your"
996 		    " system's BIOS to a version\ncontaining this"
997 		    " microcode patch is HIGHLY recommended or erroneous"
998 		    " system\noperation may occur.\n");
999 	}
1000 #endif
1001 #if defined(OPTERON_ERRATUM_121)
1002 	if (opteron_erratum_121)
1003 		workaround_applied(121);
1004 #endif
1005 #if defined(OPTERON_ERRATUM_122)
1006 	if (opteron_erratum_122)
1007 		workaround_applied(122);
1008 #endif
1009 #if defined(OPTERON_ERRATUM_123)
1010 	if (opteron_erratum_123) {
1011 		cmn_err(CE_WARN,
1012 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1013 		    " processor\nerratum 123 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_ERRATUM_131)
1020 	if (opteron_erratum_131) {
1021 		cmn_err(CE_WARN,
1022 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1023 		    " processor\nerratum 131 was not detected; updating your"
1024 		    " system's BIOS to a version\ncontaining this"
1025 		    " microcode patch is HIGHLY recommended or erroneous"
1026 		    " system\noperation may occur.\n");
1027 	}
1028 #endif
1029 #if defined(OPTERON_WORKAROUND_6336786)
1030 	if (opteron_workaround_6336786)
1031 		workaround_applied(6336786);
1032 #endif
1033 #if defined(OPTERON_WORKAROUND_6323525)
1034 	if (opteron_workaround_6323525)
1035 		workaround_applied(6323525);
1036 #endif
1037 }
1038 
1039 static cpuset_t procset;
1040 
1041 /*
1042  * Start a single cpu, assuming that the kernel context is available
1043  * to successfully start another cpu.
1044  *
1045  * (For example, real mode code is mapped into the right place
1046  * in memory and is ready to be run.)
1047  */
1048 int
1049 start_cpu(processorid_t who)
1050 {
1051 	void *ctx;
1052 	cpu_t *cp;
1053 	int delays;
1054 	int error = 0;
1055 
1056 	ASSERT(who != 0);
1057 
1058 	/*
1059 	 * Check if there's at least a Mbyte of kmem available
1060 	 * before attempting to start the cpu.
1061 	 */
1062 	if (kmem_avail() < 1024 * 1024) {
1063 		/*
1064 		 * Kick off a reap in case that helps us with
1065 		 * later attempts ..
1066 		 */
1067 		kmem_reap();
1068 		return (ENOMEM);
1069 	}
1070 
1071 	cp = mp_startup_init(who);
1072 	if ((ctx = mach_cpucontext_alloc(cp)) == NULL ||
1073 	    (error = mach_cpu_start(cp, ctx)) != 0) {
1074 
1075 		/*
1076 		 * Something went wrong before we even started it
1077 		 */
1078 		if (ctx)
1079 			cmn_err(CE_WARN,
1080 			    "cpu%d: failed to start error %d",
1081 			    cp->cpu_id, error);
1082 		else
1083 			cmn_err(CE_WARN,
1084 			    "cpu%d: failed to allocate context", cp->cpu_id);
1085 
1086 		if (ctx)
1087 			mach_cpucontext_free(cp, ctx, error);
1088 		else
1089 			error = EAGAIN;		/* hmm. */
1090 		mp_startup_fini(cp, error);
1091 		return (error);
1092 	}
1093 
1094 	for (delays = 0; !CPU_IN_SET(procset, who); delays++) {
1095 		if (delays == 500) {
1096 			/*
1097 			 * After five seconds, things are probably looking
1098 			 * a bit bleak - explain the hang.
1099 			 */
1100 			cmn_err(CE_NOTE, "cpu%d: started, "
1101 			    "but not running in the kernel yet", who);
1102 		} else if (delays > 2000) {
1103 			/*
1104 			 * We waited at least 20 seconds, bail ..
1105 			 */
1106 			error = ETIMEDOUT;
1107 			cmn_err(CE_WARN, "cpu%d: timed out", who);
1108 			mach_cpucontext_free(cp, ctx, error);
1109 			mp_startup_fini(cp, error);
1110 			return (error);
1111 		}
1112 
1113 		/*
1114 		 * wait at least 10ms, then check again..
1115 		 */
1116 		delay(USEC_TO_TICK_ROUNDUP(10000));
1117 	}
1118 
1119 	mach_cpucontext_free(cp, ctx, 0);
1120 
1121 	if (tsc_gethrtime_enable)
1122 		tsc_sync_master(who);
1123 
1124 	if (dtrace_cpu_init != NULL) {
1125 		/*
1126 		 * DTrace CPU initialization expects cpu_lock to be held.
1127 		 */
1128 		mutex_enter(&cpu_lock);
1129 		(*dtrace_cpu_init)(who);
1130 		mutex_exit(&cpu_lock);
1131 	}
1132 
1133 	while (!CPU_IN_SET(cpu_ready_set, who))
1134 		delay(1);
1135 
1136 	return (0);
1137 }
1138 
1139 
1140 /*ARGSUSED*/
1141 void
1142 start_other_cpus(int cprboot)
1143 {
1144 	uint_t who;
1145 	uint_t skipped = 0;
1146 	uint_t bootcpuid = 0;
1147 
1148 	/*
1149 	 * Initialize our own cpu_info.
1150 	 */
1151 	init_cpu_info(CPU);
1152 
1153 	/*
1154 	 * Initialize our syscall handlers
1155 	 */
1156 	init_cpu_syscall(CPU);
1157 
1158 	/*
1159 	 * Take the boot cpu out of the mp_cpus set because we know
1160 	 * it's already running.  Add it to the cpu_ready_set for
1161 	 * precisely the same reason.
1162 	 */
1163 	CPUSET_DEL(mp_cpus, bootcpuid);
1164 	CPUSET_ADD(cpu_ready_set, bootcpuid);
1165 
1166 	/*
1167 	 * if only 1 cpu or not using MP, skip the rest of this
1168 	 */
1169 	if (CPUSET_ISNULL(mp_cpus) || use_mp == 0) {
1170 		if (use_mp == 0)
1171 			cmn_err(CE_CONT, "?***** Not in MP mode\n");
1172 		goto done;
1173 	}
1174 
1175 	/*
1176 	 * perform such initialization as is needed
1177 	 * to be able to take CPUs on- and off-line.
1178 	 */
1179 	cpu_pause_init();
1180 
1181 	xc_init();		/* initialize processor crosscalls */
1182 
1183 	if (mach_cpucontext_init() != 0)
1184 		goto done;
1185 
1186 	flushes_require_xcalls = 1;
1187 
1188 	/*
1189 	 * We lock our affinity to the master CPU to ensure that all slave CPUs
1190 	 * do their TSC syncs with the same CPU.
1191 	 */
1192 	affinity_set(CPU_CURRENT);
1193 
1194 	for (who = 0; who < NCPU; who++) {
1195 
1196 		if (!CPU_IN_SET(mp_cpus, who))
1197 			continue;
1198 		ASSERT(who != bootcpuid);
1199 		if (ncpus >= max_ncpus) {
1200 			skipped = who;
1201 			continue;
1202 		}
1203 		if (start_cpu(who) != 0)
1204 			CPUSET_DEL(mp_cpus, who);
1205 	}
1206 
1207 	/* Free the space allocated to hold the microcode file */
1208 	ucode_free();
1209 
1210 	affinity_clear();
1211 
1212 	if (skipped) {
1213 		cmn_err(CE_NOTE,
1214 		    "System detected %d cpus, but "
1215 		    "only %d cpu(s) were enabled during boot.",
1216 		    skipped + 1, ncpus);
1217 		cmn_err(CE_NOTE,
1218 		    "Use \"boot-ncpus\" parameter to enable more CPU(s). "
1219 		    "See eeprom(1M).");
1220 	}
1221 
1222 done:
1223 	workaround_errata_end();
1224 	mach_cpucontext_fini();
1225 
1226 	cmi_post_mpstartup();
1227 }
1228 
1229 /*
1230  * Dummy functions - no i86pc platforms support dynamic cpu allocation.
1231  */
1232 /*ARGSUSED*/
1233 int
1234 mp_cpu_configure(int cpuid)
1235 {
1236 	return (ENOTSUP);		/* not supported */
1237 }
1238 
1239 /*ARGSUSED*/
1240 int
1241 mp_cpu_unconfigure(int cpuid)
1242 {
1243 	return (ENOTSUP);		/* not supported */
1244 }
1245 
1246 /*
1247  * Startup function for 'other' CPUs (besides boot cpu).
1248  * Called from real_mode_start.
1249  *
1250  * WARNING: until CPU_READY is set, mp_startup and routines called by
1251  * mp_startup should not call routines (e.g. kmem_free) that could call
1252  * hat_unload which requires CPU_READY to be set.
1253  */
1254 void
1255 mp_startup(void)
1256 {
1257 	struct cpu *cp = CPU;
1258 	uint_t new_x86_feature;
1259 
1260 	/*
1261 	 * We need to get TSC on this proc synced (i.e., any delta
1262 	 * from cpu0 accounted for) as soon as we can, because many
1263 	 * many things use gethrtime/pc_gethrestime, including
1264 	 * interrupts, cmn_err, etc.
1265 	 */
1266 
1267 	/* Let cpu0 continue into tsc_sync_master() */
1268 	CPUSET_ATOMIC_ADD(procset, cp->cpu_id);
1269 
1270 	if (tsc_gethrtime_enable)
1271 		tsc_sync_slave();
1272 
1273 	/*
1274 	 * Once this was done from assembly, but it's safer here; if
1275 	 * it blocks, we need to be able to swtch() to and from, and
1276 	 * since we get here by calling t_pc, we need to do that call
1277 	 * before swtch() overwrites it.
1278 	 */
1279 
1280 	(void) (*ap_mlsetup)();
1281 
1282 	new_x86_feature = cpuid_pass1(cp);
1283 
1284 	/*
1285 	 * We need to Sync MTRR with cpu0's MTRR. We have to do
1286 	 * this with interrupts disabled.
1287 	 */
1288 	if (x86_feature & X86_MTRR)
1289 		mtrr_sync();
1290 
1291 	/*
1292 	 * Set up TSC_AUX to contain the cpuid for this processor
1293 	 * for the rdtscp instruction.
1294 	 */
1295 	if (x86_feature & X86_TSCP)
1296 		(void) wrmsr(MSR_AMD_TSCAUX, cp->cpu_id);
1297 
1298 	/*
1299 	 * Initialize this CPU's syscall handlers
1300 	 */
1301 	init_cpu_syscall(cp);
1302 
1303 	/*
1304 	 * Enable interrupts with spl set to LOCK_LEVEL. LOCK_LEVEL is the
1305 	 * highest level at which a routine is permitted to block on
1306 	 * an adaptive mutex (allows for cpu poke interrupt in case
1307 	 * the cpu is blocked on a mutex and halts). Setting LOCK_LEVEL blocks
1308 	 * device interrupts that may end up in the hat layer issuing cross
1309 	 * calls before CPU_READY is set.
1310 	 */
1311 	splx(ipltospl(LOCK_LEVEL));
1312 	sti();
1313 
1314 	/*
1315 	 * Do a sanity check to make sure this new CPU is a sane thing
1316 	 * to add to the collection of processors running this system.
1317 	 *
1318 	 * XXX	Clearly this needs to get more sophisticated, if x86
1319 	 * systems start to get built out of heterogenous CPUs; as is
1320 	 * likely to happen once the number of processors in a configuration
1321 	 * gets large enough.
1322 	 */
1323 	if ((x86_feature & new_x86_feature) != x86_feature) {
1324 		cmn_err(CE_CONT, "?cpu%d: %b\n",
1325 		    cp->cpu_id, new_x86_feature, FMT_X86_FEATURE);
1326 		cmn_err(CE_WARN, "cpu%d feature mismatch", cp->cpu_id);
1327 	}
1328 
1329 	/*
1330 	 * We do not support cpus with mixed monitor/mwait support if the
1331 	 * boot cpu supports monitor/mwait.
1332 	 */
1333 	if ((x86_feature & ~new_x86_feature) & X86_MWAIT)
1334 		panic("unsupported mixed cpu monitor/mwait support detected");
1335 
1336 	/*
1337 	 * We could be more sophisticated here, and just mark the CPU
1338 	 * as "faulted" but at this point we'll opt for the easier
1339 	 * answer of dieing horribly.  Provided the boot cpu is ok,
1340 	 * the system can be recovered by booting with use_mp set to zero.
1341 	 */
1342 	if (workaround_errata(cp) != 0)
1343 		panic("critical workaround(s) missing for cpu%d", cp->cpu_id);
1344 
1345 	cpuid_pass2(cp);
1346 	cpuid_pass3(cp);
1347 	(void) cpuid_pass4(cp);
1348 
1349 	init_cpu_info(cp);
1350 
1351 	mutex_enter(&cpu_lock);
1352 	/*
1353 	 * Processor group initialization for this CPU is dependent on the
1354 	 * cpuid probing, which must be done in the context of the current
1355 	 * CPU.
1356 	 */
1357 	pghw_physid_create(cp);
1358 	pg_cpu_init(cp);
1359 	pg_cmt_cpu_startup(cp);
1360 
1361 	cp->cpu_flags |= CPU_RUNNING | CPU_READY | CPU_ENABLE | CPU_EXISTS;
1362 	cpu_add_active(cp);
1363 
1364 	if (dtrace_cpu_init != NULL) {
1365 		(*dtrace_cpu_init)(cp->cpu_id);
1366 	}
1367 
1368 	/*
1369 	 * Fill out cpu_ucode_info.  Update microcode if necessary.
1370 	 */
1371 	ucode_check(cp);
1372 
1373 	mutex_exit(&cpu_lock);
1374 
1375 	/*
1376 	 * Enable preemption here so that contention for any locks acquired
1377 	 * later in mp_startup may be preempted if the thread owning those
1378 	 * locks is continously executing on other CPUs (for example, this
1379 	 * CPU must be preemptible to allow other CPUs to pause it during their
1380 	 * startup phases).  It's safe to enable preemption here because the
1381 	 * CPU state is pretty-much fully constructed.
1382 	 */
1383 	curthread->t_preempt = 0;
1384 
1385 	add_cpunode2devtree(cp->cpu_id, cp->cpu_m.mcpu_cpi);
1386 
1387 	/* The base spl should still be at LOCK LEVEL here */
1388 	ASSERT(cp->cpu_base_spl == ipltospl(LOCK_LEVEL));
1389 	set_base_spl();		/* Restore the spl to its proper value */
1390 
1391 	(void) spl0();				/* enable interrupts */
1392 
1393 	/*
1394 	 * Set up the CPU module for this CPU.  This can't be done before
1395 	 * this CPU is made CPU_READY, because we may (in heterogeneous systems)
1396 	 * need to go load another CPU module.  The act of attempting to load
1397 	 * a module may trigger a cross-call, which will ASSERT unless this
1398 	 * cpu is CPU_READY.
1399 	 */
1400 	cmi_init();
1401 
1402 	if (x86_feature & X86_MCA)
1403 		cmi_mca_init();
1404 
1405 	if (boothowto & RB_DEBUG)
1406 		kdi_cpu_init();
1407 
1408 	/*
1409 	 * Setting the bit in cpu_ready_set must be the last operation in
1410 	 * processor initialization; the boot CPU will continue to boot once
1411 	 * it sees this bit set for all active CPUs.
1412 	 */
1413 	CPUSET_ATOMIC_ADD(cpu_ready_set, cp->cpu_id);
1414 
1415 	/*
1416 	 * Because mp_startup() gets fired off after init() starts, we
1417 	 * can't use the '?' trick to do 'boot -v' printing - so we
1418 	 * always direct the 'cpu .. online' messages to the log.
1419 	 */
1420 	cmn_err(CE_CONT, "!cpu%d initialization complete - online\n",
1421 	    cp->cpu_id);
1422 
1423 	/*
1424 	 * Now we are done with the startup thread, so free it up.
1425 	 */
1426 	thread_exit();
1427 	panic("mp_startup: cannot return");
1428 	/*NOTREACHED*/
1429 }
1430 
1431 
1432 /*
1433  * Start CPU on user request.
1434  */
1435 /* ARGSUSED */
1436 int
1437 mp_cpu_start(struct cpu *cp)
1438 {
1439 	ASSERT(MUTEX_HELD(&cpu_lock));
1440 	return (0);
1441 }
1442 
1443 /*
1444  * Stop CPU on user request.
1445  */
1446 /* ARGSUSED */
1447 int
1448 mp_cpu_stop(struct cpu *cp)
1449 {
1450 	extern int cbe_psm_timer_mode;
1451 	ASSERT(MUTEX_HELD(&cpu_lock));
1452 
1453 	/*
1454 	 * If TIMER_PERIODIC mode is used, CPU0 is the one running it;
1455 	 * can't stop it.  (This is true only for machines with no TSC.)
1456 	 */
1457 
1458 	if ((cbe_psm_timer_mode == TIMER_PERIODIC) && (cp->cpu_id == 0))
1459 		return (1);
1460 
1461 	return (0);
1462 }
1463 
1464 /*
1465  * Take the specified CPU out of participation in interrupts.
1466  */
1467 int
1468 cpu_disable_intr(struct cpu *cp)
1469 {
1470 	if (psm_disable_intr(cp->cpu_id) != DDI_SUCCESS)
1471 		return (EBUSY);
1472 
1473 	cp->cpu_flags &= ~CPU_ENABLE;
1474 	return (0);
1475 }
1476 
1477 /*
1478  * Allow the specified CPU to participate in interrupts.
1479  */
1480 void
1481 cpu_enable_intr(struct cpu *cp)
1482 {
1483 	ASSERT(MUTEX_HELD(&cpu_lock));
1484 	cp->cpu_flags |= CPU_ENABLE;
1485 	psm_enable_intr(cp->cpu_id);
1486 }
1487 
1488 
1489 
1490 void
1491 mp_cpu_faulted_enter(struct cpu *cp)
1492 {
1493 	cmi_faulted_enter(cp);
1494 }
1495 
1496 void
1497 mp_cpu_faulted_exit(struct cpu *cp)
1498 {
1499 	cmi_faulted_exit(cp);
1500 }
1501 
1502 /*
1503  * The following two routines are used as context operators on threads belonging
1504  * to processes with a private LDT (see sysi86).  Due to the rarity of such
1505  * processes, these routines are currently written for best code readability and
1506  * organization rather than speed.  We could avoid checking x86_feature at every
1507  * context switch by installing different context ops, depending on the
1508  * x86_feature flags, at LDT creation time -- one for each combination of fast
1509  * syscall feature flags.
1510  */
1511 
1512 /*ARGSUSED*/
1513 void
1514 cpu_fast_syscall_disable(void *arg)
1515 {
1516 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP))
1517 		cpu_sep_disable();
1518 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC))
1519 		cpu_asysc_disable();
1520 }
1521 
1522 /*ARGSUSED*/
1523 void
1524 cpu_fast_syscall_enable(void *arg)
1525 {
1526 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP))
1527 		cpu_sep_enable();
1528 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC))
1529 		cpu_asysc_enable();
1530 }
1531 
1532 static void
1533 cpu_sep_enable(void)
1534 {
1535 	ASSERT(x86_feature & X86_SEP);
1536 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1537 
1538 	wrmsr(MSR_INTC_SEP_CS, (uint64_t)(uintptr_t)KCS_SEL);
1539 }
1540 
1541 static void
1542 cpu_sep_disable(void)
1543 {
1544 	ASSERT(x86_feature & X86_SEP);
1545 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1546 
1547 	/*
1548 	 * Setting the SYSENTER_CS_MSR register to 0 causes software executing
1549 	 * the sysenter or sysexit instruction to trigger a #gp fault.
1550 	 */
1551 	wrmsr(MSR_INTC_SEP_CS, 0);
1552 }
1553 
1554 static void
1555 cpu_asysc_enable(void)
1556 {
1557 	ASSERT(x86_feature & X86_ASYSC);
1558 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1559 
1560 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) |
1561 	    (uint64_t)(uintptr_t)AMD_EFER_SCE);
1562 }
1563 
1564 static void
1565 cpu_asysc_disable(void)
1566 {
1567 	ASSERT(x86_feature & X86_ASYSC);
1568 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1569 
1570 	/*
1571 	 * Turn off the SCE (syscall enable) bit in the EFER register. Software
1572 	 * executing syscall or sysret with this bit off will incur a #ud trap.
1573 	 */
1574 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) &
1575 	    ~((uint64_t)(uintptr_t)AMD_EFER_SCE));
1576 }
1577