xref: /titanic_41/usr/src/uts/i86pc/os/mp_startup.c (revision 24db46411fd54f70c35b94bb952eb7ba040e43b4)
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 	procp = curthread->t_procp;
241 
242 	mutex_enter(&cpu_lock);
243 	/*
244 	 * Initialize the dispatcher first.
245 	 */
246 	disp_cpu_init(cp);
247 	mutex_exit(&cpu_lock);
248 
249 	cpu_vm_data_init(cp);
250 
251 	/*
252 	 * Allocate and initialize the startup thread for this CPU.
253 	 * Interrupt and process switch stacks get allocated later
254 	 * when the CPU starts running.
255 	 */
256 	tp = thread_create(NULL, 0, NULL, NULL, 0, procp,
257 	    TS_STOPPED, maxclsyspri);
258 
259 	/*
260 	 * Set state to TS_ONPROC since this thread will start running
261 	 * as soon as the CPU comes online.
262 	 *
263 	 * All the other fields of the thread structure are setup by
264 	 * thread_create().
265 	 */
266 	THREAD_ONPROC(tp, cp);
267 	tp->t_preempt = 1;
268 	tp->t_bound_cpu = cp;
269 	tp->t_affinitycnt = 1;
270 	tp->t_cpu = cp;
271 	tp->t_disp_queue = cp->cpu_disp;
272 
273 	/*
274 	 * Setup thread to start in mp_startup.
275 	 */
276 	sp = tp->t_stk;
277 	tp->t_pc = (uintptr_t)mp_startup;
278 	tp->t_sp = (uintptr_t)(sp - MINFRAME);
279 #if defined(__amd64)
280 	tp->t_sp -= STACK_ENTRY_ALIGN;		/* fake a call */
281 #endif
282 
283 	cp->cpu_id = cpun;
284 	cp->cpu_self = cp;
285 	cp->cpu_thread = tp;
286 	cp->cpu_lwp = NULL;
287 	cp->cpu_dispthread = tp;
288 	cp->cpu_dispatch_pri = DISP_PRIO(tp);
289 
290 	/*
291 	 * cpu_base_spl must be set explicitly here to prevent any blocking
292 	 * operations in mp_startup from causing the spl of the cpu to drop
293 	 * to 0 (allowing device interrupts before we're ready) in resume().
294 	 * cpu_base_spl MUST remain at LOCK_LEVEL until the cpu is CPU_READY.
295 	 * As an extra bit of security on DEBUG kernels, this is enforced with
296 	 * an assertion in mp_startup() -- before cpu_base_spl is set to its
297 	 * proper value.
298 	 */
299 	cp->cpu_base_spl = ipltospl(LOCK_LEVEL);
300 
301 	/*
302 	 * Now, initialize per-CPU idle thread for this CPU.
303 	 */
304 	tp = thread_create(NULL, PAGESIZE, idle, NULL, 0, procp, TS_ONPROC, -1);
305 
306 	cp->cpu_idle_thread = tp;
307 
308 	tp->t_preempt = 1;
309 	tp->t_bound_cpu = cp;
310 	tp->t_affinitycnt = 1;
311 	tp->t_cpu = cp;
312 	tp->t_disp_queue = cp->cpu_disp;
313 
314 	/*
315 	 * Bootstrap the CPU's PG data
316 	 */
317 	pg_cpu_bootstrap(cp);
318 
319 	/*
320 	 * Perform CPC initialization on the new CPU.
321 	 */
322 	kcpc_hw_init(cp);
323 
324 	/*
325 	 * Allocate virtual addresses for cpu_caddr1 and cpu_caddr2
326 	 * for each CPU.
327 	 */
328 	setup_vaddr_for_ppcopy(cp);
329 
330 	/*
331 	 * Allocate page for new GDT and initialize from current GDT.
332 	 */
333 #if !defined(__lint)
334 	ASSERT((sizeof (*cp->cpu_gdt) * NGDT) <= PAGESIZE);
335 #endif
336 	cp->cpu_m.mcpu_gdt = kmem_zalloc(PAGESIZE, KM_SLEEP);
337 	bcopy(CPU->cpu_m.mcpu_gdt, cp->cpu_m.mcpu_gdt,
338 	    (sizeof (*cp->cpu_m.mcpu_gdt) * NGDT));
339 
340 #if defined(__i386)
341 	/*
342 	 * setup kernel %gs.
343 	 */
344 	set_usegd(&cp->cpu_gdt[GDT_GS], cp, sizeof (struct cpu) -1, SDT_MEMRWA,
345 	    SEL_KPL, 0, 1);
346 #endif
347 
348 	/*
349 	 * If we have more than one node, each cpu gets a copy of IDT
350 	 * local to its node. If this is a Pentium box, we use cpu 0's
351 	 * IDT. cpu 0's IDT has been made read-only to workaround the
352 	 * cmpxchgl register bug
353 	 */
354 	if (system_hardware.hd_nodes && x86_type != X86_TYPE_P5) {
355 		struct machcpu *mcpu = &cp->cpu_m;
356 
357 		mcpu->mcpu_idt = kmem_alloc(sizeof (idt0), KM_SLEEP);
358 		bcopy(idt0, mcpu->mcpu_idt, sizeof (idt0));
359 	} else {
360 		cp->cpu_m.mcpu_idt = CPU->cpu_m.mcpu_idt;
361 	}
362 
363 	/*
364 	 * Get interrupt priority data from cpu 0.
365 	 */
366 	cp->cpu_pri_data = CPU->cpu_pri_data;
367 
368 	/*
369 	 * alloc space for cpuid info
370 	 */
371 	cpuid_alloc_space(cp);
372 
373 	hat_cpu_online(cp);
374 
375 #ifdef TRAPTRACE
376 	/*
377 	 * If this is a TRAPTRACE kernel, allocate TRAPTRACE buffers
378 	 */
379 	ttc->ttc_first = (uintptr_t)kmem_zalloc(trap_trace_bufsize, KM_SLEEP);
380 	ttc->ttc_next = ttc->ttc_first;
381 	ttc->ttc_limit = ttc->ttc_first + trap_trace_bufsize;
382 #endif
383 	/*
384 	 * Record that we have another CPU.
385 	 */
386 	mutex_enter(&cpu_lock);
387 	/*
388 	 * Initialize the interrupt threads for this CPU
389 	 */
390 	cpu_intr_alloc(cp, NINTR_THREADS);
391 	/*
392 	 * Add CPU to list of available CPUs.  It'll be on the active list
393 	 * after mp_startup().
394 	 */
395 	cpu_add_unit(cp);
396 	mutex_exit(&cpu_lock);
397 
398 	return (cp);
399 }
400 
401 /*
402  * Undo what was done in mp_startup_init
403  */
404 static void
405 mp_startup_fini(struct cpu *cp, int error)
406 {
407 	mutex_enter(&cpu_lock);
408 
409 	/*
410 	 * Remove the CPU from the list of available CPUs.
411 	 */
412 	cpu_del_unit(cp->cpu_id);
413 
414 	if (error == ETIMEDOUT) {
415 		/*
416 		 * The cpu was started, but never *seemed* to run any
417 		 * code in the kernel; it's probably off spinning in its
418 		 * own private world, though with potential references to
419 		 * our kmem-allocated IDTs and GDTs (for example).
420 		 *
421 		 * Worse still, it may actually wake up some time later,
422 		 * so rather than guess what it might or might not do, we
423 		 * leave the fundamental data structures intact.
424 		 */
425 		cp->cpu_flags = 0;
426 		mutex_exit(&cpu_lock);
427 		return;
428 	}
429 
430 	/*
431 	 * At this point, the only threads bound to this CPU should
432 	 * special per-cpu threads: it's idle thread, it's pause threads,
433 	 * and it's interrupt threads.  Clean these up.
434 	 */
435 	cpu_destroy_bound_threads(cp);
436 	cp->cpu_idle_thread = NULL;
437 
438 	/*
439 	 * Free the interrupt stack.
440 	 */
441 	segkp_release(segkp,
442 	    cp->cpu_intr_stack - (INTR_STACK_SIZE - SA(MINFRAME)));
443 
444 	mutex_exit(&cpu_lock);
445 
446 #ifdef TRAPTRACE
447 	/*
448 	 * Discard the trap trace buffer
449 	 */
450 	{
451 		trap_trace_ctl_t *ttc = &trap_trace_ctl[cp->cpu_id];
452 
453 		kmem_free((void *)ttc->ttc_first, trap_trace_bufsize);
454 		ttc->ttc_first = NULL;
455 	}
456 #endif
457 
458 	hat_cpu_offline(cp);
459 
460 	cpuid_free_space(cp);
461 
462 	if (cp->cpu_m.mcpu_idt != CPU->cpu_m.mcpu_idt)
463 		kmem_free(cp->cpu_m.mcpu_idt, sizeof (idt0));
464 	cp->cpu_m.mcpu_idt = NULL;
465 
466 	kmem_free(cp->cpu_m.mcpu_gdt, PAGESIZE);
467 	cp->cpu_m.mcpu_gdt = NULL;
468 
469 	teardown_vaddr_for_ppcopy(cp);
470 
471 	kcpc_hw_fini(cp);
472 
473 	cp->cpu_dispthread = NULL;
474 	cp->cpu_thread = NULL;	/* discarded by cpu_destroy_bound_threads() */
475 
476 	cpu_vm_data_destroy(cp);
477 
478 	mutex_enter(&cpu_lock);
479 	disp_cpu_fini(cp);
480 	mutex_exit(&cpu_lock);
481 
482 	kmem_free(cp, sizeof (*cp));
483 }
484 
485 /*
486  * Apply workarounds for known errata, and warn about those that are absent.
487  *
488  * System vendors occasionally create configurations which contain different
489  * revisions of the CPUs that are almost but not exactly the same.  At the
490  * time of writing, this meant that their clock rates were the same, their
491  * feature sets were the same, but the required workaround were -not-
492  * necessarily the same.  So, this routine is invoked on -every- CPU soon
493  * after starting to make sure that the resulting system contains the most
494  * pessimal set of workarounds needed to cope with *any* of the CPUs in the
495  * system.
496  *
497  * workaround_errata is invoked early in mlsetup() for CPU 0, and in
498  * mp_startup() for all slave CPUs. Slaves process workaround_errata prior
499  * to acknowledging their readiness to the master, so this routine will
500  * never be executed by multiple CPUs in parallel, thus making updates to
501  * global data safe.
502  *
503  * These workarounds are based on Rev 3.57 of the Revision Guide for
504  * AMD Athlon(tm) 64 and AMD Opteron(tm) Processors, August 2005.
505  */
506 
507 #if defined(OPTERON_ERRATUM_88)
508 int opteron_erratum_88;		/* if non-zero -> at least one cpu has it */
509 #endif
510 
511 #if defined(OPTERON_ERRATUM_91)
512 int opteron_erratum_91;		/* if non-zero -> at least one cpu has it */
513 #endif
514 
515 #if defined(OPTERON_ERRATUM_93)
516 int opteron_erratum_93;		/* if non-zero -> at least one cpu has it */
517 #endif
518 
519 #if defined(OPTERON_ERRATUM_95)
520 int opteron_erratum_95;		/* if non-zero -> at least one cpu has it */
521 #endif
522 
523 #if defined(OPTERON_ERRATUM_100)
524 int opteron_erratum_100;	/* if non-zero -> at least one cpu has it */
525 #endif
526 
527 #if defined(OPTERON_ERRATUM_108)
528 int opteron_erratum_108;	/* if non-zero -> at least one cpu has it */
529 #endif
530 
531 #if defined(OPTERON_ERRATUM_109)
532 int opteron_erratum_109;	/* if non-zero -> at least one cpu has it */
533 #endif
534 
535 #if defined(OPTERON_ERRATUM_121)
536 int opteron_erratum_121;	/* if non-zero -> at least one cpu has it */
537 #endif
538 
539 #if defined(OPTERON_ERRATUM_122)
540 int opteron_erratum_122;	/* if non-zero -> at least one cpu has it */
541 #endif
542 
543 #if defined(OPTERON_ERRATUM_123)
544 int opteron_erratum_123;	/* if non-zero -> at least one cpu has it */
545 #endif
546 
547 #if defined(OPTERON_ERRATUM_131)
548 int opteron_erratum_131;	/* if non-zero -> at least one cpu has it */
549 #endif
550 
551 #if defined(OPTERON_WORKAROUND_6336786)
552 int opteron_workaround_6336786;	/* non-zero -> WA relevant and applied */
553 int opteron_workaround_6336786_UP = 0;	/* Not needed for UP */
554 #endif
555 
556 #if defined(OPTERON_WORKAROUND_6323525)
557 int opteron_workaround_6323525;	/* if non-zero -> at least one cpu has it */
558 #endif
559 
560 static void
561 workaround_warning(cpu_t *cp, uint_t erratum)
562 {
563 	cmn_err(CE_WARN, "cpu%d: no workaround for erratum %u",
564 	    cp->cpu_id, erratum);
565 }
566 
567 static void
568 workaround_applied(uint_t erratum)
569 {
570 	if (erratum > 1000000)
571 		cmn_err(CE_CONT, "?workaround applied for cpu issue #%d\n",
572 		    erratum);
573 	else
574 		cmn_err(CE_CONT, "?workaround applied for cpu erratum #%d\n",
575 		    erratum);
576 }
577 
578 static void
579 msr_warning(cpu_t *cp, const char *rw, uint_t msr, int error)
580 {
581 	cmn_err(CE_WARN, "cpu%d: couldn't %smsr 0x%x, error %d",
582 	    cp->cpu_id, rw, msr, error);
583 }
584 
585 uint_t
586 workaround_errata(struct cpu *cpu)
587 {
588 	uint_t missing = 0;
589 
590 	ASSERT(cpu == CPU);
591 
592 	/*LINTED*/
593 	if (cpuid_opteron_erratum(cpu, 88) > 0) {
594 		/*
595 		 * SWAPGS May Fail To Read Correct GS Base
596 		 */
597 #if defined(OPTERON_ERRATUM_88)
598 		/*
599 		 * The workaround is an mfence in the relevant assembler code
600 		 */
601 		opteron_erratum_88++;
602 #else
603 		workaround_warning(cpu, 88);
604 		missing++;
605 #endif
606 	}
607 
608 	if (cpuid_opteron_erratum(cpu, 91) > 0) {
609 		/*
610 		 * Software Prefetches May Report A Page Fault
611 		 */
612 #if defined(OPTERON_ERRATUM_91)
613 		/*
614 		 * fix is in trap.c
615 		 */
616 		opteron_erratum_91++;
617 #else
618 		workaround_warning(cpu, 91);
619 		missing++;
620 #endif
621 	}
622 
623 	if (cpuid_opteron_erratum(cpu, 93) > 0) {
624 		/*
625 		 * RSM Auto-Halt Restart Returns to Incorrect RIP
626 		 */
627 #if defined(OPTERON_ERRATUM_93)
628 		/*
629 		 * fix is in trap.c
630 		 */
631 		opteron_erratum_93++;
632 #else
633 		workaround_warning(cpu, 93);
634 		missing++;
635 #endif
636 	}
637 
638 	/*LINTED*/
639 	if (cpuid_opteron_erratum(cpu, 95) > 0) {
640 		/*
641 		 * RET Instruction May Return to Incorrect EIP
642 		 */
643 #if defined(OPTERON_ERRATUM_95)
644 #if defined(_LP64)
645 		/*
646 		 * Workaround this by ensuring that 32-bit user code and
647 		 * 64-bit kernel code never occupy the same address
648 		 * range mod 4G.
649 		 */
650 		if (_userlimit32 > 0xc0000000ul)
651 			*(uintptr_t *)&_userlimit32 = 0xc0000000ul;
652 
653 		/*LINTED*/
654 		ASSERT((uint32_t)COREHEAP_BASE == 0xc0000000u);
655 		opteron_erratum_95++;
656 #endif	/* _LP64 */
657 #else
658 		workaround_warning(cpu, 95);
659 		missing++;
660 #endif
661 	}
662 
663 	if (cpuid_opteron_erratum(cpu, 100) > 0) {
664 		/*
665 		 * Compatibility Mode Branches Transfer to Illegal Address
666 		 */
667 #if defined(OPTERON_ERRATUM_100)
668 		/*
669 		 * fix is in trap.c
670 		 */
671 		opteron_erratum_100++;
672 #else
673 		workaround_warning(cpu, 100);
674 		missing++;
675 #endif
676 	}
677 
678 	/*LINTED*/
679 	if (cpuid_opteron_erratum(cpu, 108) > 0) {
680 		/*
681 		 * CPUID Instruction May Return Incorrect Model Number In
682 		 * Some Processors
683 		 */
684 #if defined(OPTERON_ERRATUM_108)
685 		/*
686 		 * (Our cpuid-handling code corrects the model number on
687 		 * those processors)
688 		 */
689 #else
690 		workaround_warning(cpu, 108);
691 		missing++;
692 #endif
693 	}
694 
695 	/*LINTED*/
696 	if (cpuid_opteron_erratum(cpu, 109) > 0) do {
697 		/*
698 		 * Certain Reverse REP MOVS May Produce Unpredictable Behaviour
699 		 */
700 #if defined(OPTERON_ERRATUM_109)
701 		/*
702 		 * The "workaround" is to print a warning to upgrade the BIOS
703 		 */
704 		uint64_t value;
705 		const uint_t msr = MSR_AMD_PATCHLEVEL;
706 		int err;
707 
708 		if ((err = checked_rdmsr(msr, &value)) != 0) {
709 			msr_warning(cpu, "rd", msr, err);
710 			workaround_warning(cpu, 109);
711 			missing++;
712 		}
713 		if (value == 0)
714 			opteron_erratum_109++;
715 #else
716 		workaround_warning(cpu, 109);
717 		missing++;
718 #endif
719 	/*CONSTANTCONDITION*/
720 	} while (0);
721 
722 	/*LINTED*/
723 	if (cpuid_opteron_erratum(cpu, 121) > 0) {
724 		/*
725 		 * Sequential Execution Across Non_Canonical Boundary Caused
726 		 * Processor Hang
727 		 */
728 #if defined(OPTERON_ERRATUM_121)
729 #if defined(_LP64)
730 		/*
731 		 * Erratum 121 is only present in long (64 bit) mode.
732 		 * Workaround is to include the page immediately before the
733 		 * va hole to eliminate the possibility of system hangs due to
734 		 * sequential execution across the va hole boundary.
735 		 */
736 		if (opteron_erratum_121)
737 			opteron_erratum_121++;
738 		else {
739 			if (hole_start) {
740 				hole_start -= PAGESIZE;
741 			} else {
742 				/*
743 				 * hole_start not yet initialized by
744 				 * mmu_init. Initialize hole_start
745 				 * with value to be subtracted.
746 				 */
747 				hole_start = PAGESIZE;
748 			}
749 			opteron_erratum_121++;
750 		}
751 #endif	/* _LP64 */
752 #else
753 		workaround_warning(cpu, 121);
754 		missing++;
755 #endif
756 	}
757 
758 	/*LINTED*/
759 	if (cpuid_opteron_erratum(cpu, 122) > 0) do {
760 		/*
761 		 * TLB Flush Filter May Cause Coherency Problem in
762 		 * Multiprocessor Systems
763 		 */
764 #if defined(OPTERON_ERRATUM_122)
765 		uint64_t value;
766 		const uint_t msr = MSR_AMD_HWCR;
767 		int error;
768 
769 		/*
770 		 * Erratum 122 is only present in MP configurations (multi-core
771 		 * or multi-processor).
772 		 */
773 		if (!opteron_erratum_122 && lgrp_plat_node_cnt == 1 &&
774 		    cpuid_get_ncpu_per_chip(cpu) == 1)
775 			break;
776 
777 		/* disable TLB Flush Filter */
778 
779 		if ((error = checked_rdmsr(msr, &value)) != 0) {
780 			msr_warning(cpu, "rd", msr, error);
781 			workaround_warning(cpu, 122);
782 			missing++;
783 		} else {
784 			value |= (uint64_t)AMD_HWCR_FFDIS;
785 			if ((error = checked_wrmsr(msr, value)) != 0) {
786 				msr_warning(cpu, "wr", msr, error);
787 				workaround_warning(cpu, 122);
788 				missing++;
789 			}
790 		}
791 		opteron_erratum_122++;
792 #else
793 		workaround_warning(cpu, 122);
794 		missing++;
795 #endif
796 	/*CONSTANTCONDITION*/
797 	} while (0);
798 
799 	/*LINTED*/
800 	if (cpuid_opteron_erratum(cpu, 123) > 0) do {
801 		/*
802 		 * Bypassed Reads May Cause Data Corruption of System Hang in
803 		 * Dual Core Processors
804 		 */
805 #if defined(OPTERON_ERRATUM_123)
806 		uint64_t value;
807 		const uint_t msr = MSR_AMD_PATCHLEVEL;
808 		int err;
809 
810 		/*
811 		 * Erratum 123 applies only to multi-core cpus.
812 		 */
813 		if (cpuid_get_ncpu_per_chip(cpu) < 2)
814 			break;
815 
816 		/*
817 		 * The "workaround" is to print a warning to upgrade the BIOS
818 		 */
819 		if ((err = checked_rdmsr(msr, &value)) != 0) {
820 			msr_warning(cpu, "rd", msr, err);
821 			workaround_warning(cpu, 123);
822 			missing++;
823 		}
824 		if (value == 0)
825 			opteron_erratum_123++;
826 #else
827 		workaround_warning(cpu, 123);
828 		missing++;
829 
830 #endif
831 	/*CONSTANTCONDITION*/
832 	} while (0);
833 
834 	/*LINTED*/
835 	if (cpuid_opteron_erratum(cpu, 131) > 0) do {
836 		/*
837 		 * Multiprocessor Systems with Four or More Cores May Deadlock
838 		 * Waiting for a Probe Response
839 		 */
840 #if defined(OPTERON_ERRATUM_131)
841 		uint64_t nbcfg;
842 		const uint_t msr = MSR_AMD_NB_CFG;
843 		const uint64_t wabits =
844 		    AMD_NB_CFG_SRQ_HEARTBEAT | AMD_NB_CFG_SRQ_SPR;
845 		int error;
846 
847 		/*
848 		 * Erratum 131 applies to any system with four or more cores.
849 		 */
850 		if (opteron_erratum_131)
851 			break;
852 
853 		if (lgrp_plat_node_cnt * cpuid_get_ncpu_per_chip(cpu) < 4)
854 			break;
855 
856 		/*
857 		 * Print a warning if neither of the workarounds for
858 		 * erratum 131 is present.
859 		 */
860 		if ((error = checked_rdmsr(msr, &nbcfg)) != 0) {
861 			msr_warning(cpu, "rd", msr, error);
862 			workaround_warning(cpu, 131);
863 			missing++;
864 		} else if ((nbcfg & wabits) == 0) {
865 			opteron_erratum_131++;
866 		} else {
867 			/* cannot have both workarounds set */
868 			ASSERT((nbcfg & wabits) != wabits);
869 		}
870 #else
871 		workaround_warning(cpu, 131);
872 		missing++;
873 #endif
874 	/*CONSTANTCONDITION*/
875 	} while (0);
876 
877 	/*
878 	 * This isn't really an erratum, but for convenience the
879 	 * detection/workaround code lives here and in cpuid_opteron_erratum.
880 	 */
881 	if (cpuid_opteron_erratum(cpu, 6336786) > 0) {
882 #if defined(OPTERON_WORKAROUND_6336786)
883 		/*
884 		 * Disable C1-Clock ramping on multi-core/multi-processor
885 		 * K8 platforms to guard against TSC drift.
886 		 */
887 		if (opteron_workaround_6336786) {
888 			opteron_workaround_6336786++;
889 		} else if ((lgrp_plat_node_cnt *
890 		    cpuid_get_ncpu_per_chip(cpu) > 1) ||
891 		    opteron_workaround_6336786_UP) {
892 			int	node;
893 			uint8_t data;
894 
895 			for (node = 0; node < lgrp_plat_node_cnt; node++) {
896 				/*
897 				 * Clear PMM7[1:0] (function 3, offset 0x87)
898 				 * Northbridge device is the node id + 24.
899 				 */
900 				data = pci_getb_func(0, node + 24, 3, 0x87);
901 				data &= 0xFC;
902 				pci_putb_func(0, node + 24, 3, 0x87, data);
903 			}
904 			opteron_workaround_6336786++;
905 		}
906 #else
907 		workaround_warning(cpu, 6336786);
908 		missing++;
909 #endif
910 	}
911 
912 	/*LINTED*/
913 	/*
914 	 * Mutex primitives don't work as expected.
915 	 */
916 	if (cpuid_opteron_erratum(cpu, 6323525) > 0) {
917 #if defined(OPTERON_WORKAROUND_6323525)
918 		/*
919 		 * This problem only occurs with 2 or more cores. If bit in
920 		 * MSR_BU_CFG set, then not applicable. The workaround
921 		 * is to patch the semaphone routines with the lfence
922 		 * instruction to provide necessary load memory barrier with
923 		 * possible subsequent read-modify-write ops.
924 		 *
925 		 * It is too early in boot to call the patch routine so
926 		 * set erratum variable to be done in startup_end().
927 		 */
928 		if (opteron_workaround_6323525) {
929 			opteron_workaround_6323525++;
930 		} else if ((x86_feature & X86_SSE2) && ((lgrp_plat_node_cnt *
931 		    cpuid_get_ncpu_per_chip(cpu)) > 1)) {
932 			if ((xrdmsr(MSR_BU_CFG) & 0x02) == 0)
933 				opteron_workaround_6323525++;
934 		}
935 #else
936 		workaround_warning(cpu, 6323525);
937 		missing++;
938 #endif
939 	}
940 
941 	return (missing);
942 }
943 
944 void
945 workaround_errata_end()
946 {
947 #if defined(OPTERON_ERRATUM_88)
948 	if (opteron_erratum_88)
949 		workaround_applied(88);
950 #endif
951 #if defined(OPTERON_ERRATUM_91)
952 	if (opteron_erratum_91)
953 		workaround_applied(91);
954 #endif
955 #if defined(OPTERON_ERRATUM_93)
956 	if (opteron_erratum_93)
957 		workaround_applied(93);
958 #endif
959 #if defined(OPTERON_ERRATUM_95)
960 	if (opteron_erratum_95)
961 		workaround_applied(95);
962 #endif
963 #if defined(OPTERON_ERRATUM_100)
964 	if (opteron_erratum_100)
965 		workaround_applied(100);
966 #endif
967 #if defined(OPTERON_ERRATUM_108)
968 	if (opteron_erratum_108)
969 		workaround_applied(108);
970 #endif
971 #if defined(OPTERON_ERRATUM_109)
972 	if (opteron_erratum_109) {
973 		cmn_err(CE_WARN,
974 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
975 		    " processor\nerratum 109 was not detected; updating your"
976 		    " system's BIOS to a version\ncontaining this"
977 		    " microcode patch is HIGHLY recommended or erroneous"
978 		    " system\noperation may occur.\n");
979 	}
980 #endif
981 #if defined(OPTERON_ERRATUM_121)
982 	if (opteron_erratum_121)
983 		workaround_applied(121);
984 #endif
985 #if defined(OPTERON_ERRATUM_122)
986 	if (opteron_erratum_122)
987 		workaround_applied(122);
988 #endif
989 #if defined(OPTERON_ERRATUM_123)
990 	if (opteron_erratum_123) {
991 		cmn_err(CE_WARN,
992 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
993 		    " processor\nerratum 123 was not detected; updating your"
994 		    " system's BIOS to a version\ncontaining this"
995 		    " microcode patch is HIGHLY recommended or erroneous"
996 		    " system\noperation may occur.\n");
997 	}
998 #endif
999 #if defined(OPTERON_ERRATUM_131)
1000 	if (opteron_erratum_131) {
1001 		cmn_err(CE_WARN,
1002 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1003 		    " processor\nerratum 131 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_WORKAROUND_6336786)
1010 	if (opteron_workaround_6336786)
1011 		workaround_applied(6336786);
1012 #endif
1013 #if defined(OPTERON_WORKAROUND_6323525)
1014 	if (opteron_workaround_6323525)
1015 		workaround_applied(6323525);
1016 #endif
1017 }
1018 
1019 static cpuset_t procset;
1020 
1021 /*
1022  * Start a single cpu, assuming that the kernel context is available
1023  * to successfully start another cpu.
1024  *
1025  * (For example, real mode code is mapped into the right place
1026  * in memory and is ready to be run.)
1027  */
1028 int
1029 start_cpu(processorid_t who)
1030 {
1031 	void *ctx;
1032 	cpu_t *cp;
1033 	int delays;
1034 	int error = 0;
1035 
1036 	ASSERT(who != 0);
1037 
1038 	/*
1039 	 * Check if there's at least a Mbyte of kmem available
1040 	 * before attempting to start the cpu.
1041 	 */
1042 	if (kmem_avail() < 1024 * 1024) {
1043 		/*
1044 		 * Kick off a reap in case that helps us with
1045 		 * later attempts ..
1046 		 */
1047 		kmem_reap();
1048 		return (ENOMEM);
1049 	}
1050 
1051 	cp = mp_startup_init(who);
1052 	if ((ctx = mach_cpucontext_alloc(cp)) == NULL ||
1053 	    (error = mach_cpu_start(cp, ctx)) != 0) {
1054 
1055 		/*
1056 		 * Something went wrong before we even started it
1057 		 */
1058 		if (ctx)
1059 			cmn_err(CE_WARN,
1060 			    "cpu%d: failed to start error %d",
1061 			    cp->cpu_id, error);
1062 		else
1063 			cmn_err(CE_WARN,
1064 			    "cpu%d: failed to allocate context", cp->cpu_id);
1065 
1066 		if (ctx)
1067 			mach_cpucontext_free(cp, ctx, error);
1068 		else
1069 			error = EAGAIN;		/* hmm. */
1070 		mp_startup_fini(cp, error);
1071 		return (error);
1072 	}
1073 
1074 	for (delays = 0; !CPU_IN_SET(procset, who); delays++) {
1075 		if (delays == 500) {
1076 			/*
1077 			 * After five seconds, things are probably looking
1078 			 * a bit bleak - explain the hang.
1079 			 */
1080 			cmn_err(CE_NOTE, "cpu%d: started, "
1081 			    "but not running in the kernel yet", who);
1082 		} else if (delays > 2000) {
1083 			/*
1084 			 * We waited at least 20 seconds, bail ..
1085 			 */
1086 			error = ETIMEDOUT;
1087 			cmn_err(CE_WARN, "cpu%d: timed out", who);
1088 			mach_cpucontext_free(cp, ctx, error);
1089 			mp_startup_fini(cp, error);
1090 			return (error);
1091 		}
1092 
1093 		/*
1094 		 * wait at least 10ms, then check again..
1095 		 */
1096 		delay(USEC_TO_TICK_ROUNDUP(10000));
1097 	}
1098 
1099 	mach_cpucontext_free(cp, ctx, 0);
1100 
1101 	if (tsc_gethrtime_enable)
1102 		tsc_sync_master(who);
1103 
1104 	if (dtrace_cpu_init != NULL) {
1105 		/*
1106 		 * DTrace CPU initialization expects cpu_lock to be held.
1107 		 */
1108 		mutex_enter(&cpu_lock);
1109 		(*dtrace_cpu_init)(who);
1110 		mutex_exit(&cpu_lock);
1111 	}
1112 
1113 	while (!CPU_IN_SET(cpu_ready_set, who))
1114 		delay(1);
1115 
1116 	return (0);
1117 }
1118 
1119 
1120 /*ARGSUSED*/
1121 void
1122 start_other_cpus(int cprboot)
1123 {
1124 	uint_t who;
1125 	uint_t skipped = 0;
1126 	uint_t bootcpuid = 0;
1127 
1128 	/*
1129 	 * Initialize our own cpu_info.
1130 	 */
1131 	init_cpu_info(CPU);
1132 
1133 	/*
1134 	 * Initialize our syscall handlers
1135 	 */
1136 	init_cpu_syscall(CPU);
1137 
1138 	/*
1139 	 * Take the boot cpu out of the mp_cpus set because we know
1140 	 * it's already running.  Add it to the cpu_ready_set for
1141 	 * precisely the same reason.
1142 	 */
1143 	CPUSET_DEL(mp_cpus, bootcpuid);
1144 	CPUSET_ADD(cpu_ready_set, bootcpuid);
1145 
1146 	/*
1147 	 * if only 1 cpu or not using MP, skip the rest of this
1148 	 */
1149 	if (CPUSET_ISNULL(mp_cpus) || use_mp == 0) {
1150 		if (use_mp == 0)
1151 			cmn_err(CE_CONT, "?***** Not in MP mode\n");
1152 		goto done;
1153 	}
1154 
1155 	/*
1156 	 * perform such initialization as is needed
1157 	 * to be able to take CPUs on- and off-line.
1158 	 */
1159 	cpu_pause_init();
1160 
1161 	xc_init();		/* initialize processor crosscalls */
1162 
1163 	if (mach_cpucontext_init() != 0)
1164 		goto done;
1165 
1166 	flushes_require_xcalls = 1;
1167 
1168 	/*
1169 	 * We lock our affinity to the master CPU to ensure that all slave CPUs
1170 	 * do their TSC syncs with the same CPU.
1171 	 */
1172 	affinity_set(CPU_CURRENT);
1173 
1174 	for (who = 0; who < NCPU; who++) {
1175 
1176 		if (!CPU_IN_SET(mp_cpus, who))
1177 			continue;
1178 		ASSERT(who != bootcpuid);
1179 		if (ncpus >= max_ncpus) {
1180 			skipped = who;
1181 			continue;
1182 		}
1183 		if (start_cpu(who) != 0)
1184 			CPUSET_DEL(mp_cpus, who);
1185 	}
1186 
1187 	affinity_clear();
1188 
1189 	if (skipped) {
1190 		cmn_err(CE_NOTE,
1191 		    "System detected %d cpus, but "
1192 		    "only %d cpu(s) were enabled during boot.",
1193 		    skipped + 1, ncpus);
1194 		cmn_err(CE_NOTE,
1195 		    "Use \"boot-ncpus\" parameter to enable more CPU(s). "
1196 		    "See eeprom(1M).");
1197 	}
1198 
1199 done:
1200 	workaround_errata_end();
1201 	mach_cpucontext_fini();
1202 
1203 	cmi_post_mpstartup();
1204 }
1205 
1206 /*
1207  * Dummy functions - no i86pc platforms support dynamic cpu allocation.
1208  */
1209 /*ARGSUSED*/
1210 int
1211 mp_cpu_configure(int cpuid)
1212 {
1213 	return (ENOTSUP);		/* not supported */
1214 }
1215 
1216 /*ARGSUSED*/
1217 int
1218 mp_cpu_unconfigure(int cpuid)
1219 {
1220 	return (ENOTSUP);		/* not supported */
1221 }
1222 
1223 /*
1224  * Startup function for 'other' CPUs (besides boot cpu).
1225  * Called from real_mode_start.
1226  *
1227  * WARNING: until CPU_READY is set, mp_startup and routines called by
1228  * mp_startup should not call routines (e.g. kmem_free) that could call
1229  * hat_unload which requires CPU_READY to be set.
1230  */
1231 void
1232 mp_startup(void)
1233 {
1234 	struct cpu *cp = CPU;
1235 	uint_t new_x86_feature;
1236 
1237 	/*
1238 	 * We need to get TSC on this proc synced (i.e., any delta
1239 	 * from cpu0 accounted for) as soon as we can, because many
1240 	 * many things use gethrtime/pc_gethrestime, including
1241 	 * interrupts, cmn_err, etc.
1242 	 */
1243 
1244 	/* Let cpu0 continue into tsc_sync_master() */
1245 	CPUSET_ATOMIC_ADD(procset, cp->cpu_id);
1246 
1247 	if (tsc_gethrtime_enable)
1248 		tsc_sync_slave();
1249 
1250 	/*
1251 	 * Once this was done from assembly, but it's safer here; if
1252 	 * it blocks, we need to be able to swtch() to and from, and
1253 	 * since we get here by calling t_pc, we need to do that call
1254 	 * before swtch() overwrites it.
1255 	 */
1256 
1257 	(void) (*ap_mlsetup)();
1258 
1259 	new_x86_feature = cpuid_pass1(cp);
1260 
1261 	/*
1262 	 * We need to Sync MTRR with cpu0's MTRR. We have to do
1263 	 * this with interrupts disabled.
1264 	 */
1265 	if (x86_feature & X86_MTRR)
1266 		mtrr_sync();
1267 
1268 	/*
1269 	 * Set up TSC_AUX to contain the cpuid for this processor
1270 	 * for the rdtscp instruction.
1271 	 */
1272 	if (x86_feature & X86_TSCP)
1273 		(void) wrmsr(MSR_AMD_TSCAUX, cp->cpu_id);
1274 
1275 	/*
1276 	 * Initialize this CPU's syscall handlers
1277 	 */
1278 	init_cpu_syscall(cp);
1279 
1280 	/*
1281 	 * Enable interrupts with spl set to LOCK_LEVEL. LOCK_LEVEL is the
1282 	 * highest level at which a routine is permitted to block on
1283 	 * an adaptive mutex (allows for cpu poke interrupt in case
1284 	 * the cpu is blocked on a mutex and halts). Setting LOCK_LEVEL blocks
1285 	 * device interrupts that may end up in the hat layer issuing cross
1286 	 * calls before CPU_READY is set.
1287 	 */
1288 	splx(ipltospl(LOCK_LEVEL));
1289 	sti();
1290 
1291 	/*
1292 	 * Do a sanity check to make sure this new CPU is a sane thing
1293 	 * to add to the collection of processors running this system.
1294 	 *
1295 	 * XXX	Clearly this needs to get more sophisticated, if x86
1296 	 * systems start to get built out of heterogenous CPUs; as is
1297 	 * likely to happen once the number of processors in a configuration
1298 	 * gets large enough.
1299 	 */
1300 	if ((x86_feature & new_x86_feature) != x86_feature) {
1301 		cmn_err(CE_CONT, "?cpu%d: %b\n",
1302 		    cp->cpu_id, new_x86_feature, FMT_X86_FEATURE);
1303 		cmn_err(CE_WARN, "cpu%d feature mismatch", cp->cpu_id);
1304 	}
1305 
1306 	/*
1307 	 * We could be more sophisticated here, and just mark the CPU
1308 	 * as "faulted" but at this point we'll opt for the easier
1309 	 * answer of dieing horribly.  Provided the boot cpu is ok,
1310 	 * the system can be recovered by booting with use_mp set to zero.
1311 	 */
1312 	if (workaround_errata(cp) != 0)
1313 		panic("critical workaround(s) missing for cpu%d", cp->cpu_id);
1314 
1315 	cpuid_pass2(cp);
1316 	cpuid_pass3(cp);
1317 	(void) cpuid_pass4(cp);
1318 
1319 	init_cpu_info(cp);
1320 
1321 	mutex_enter(&cpu_lock);
1322 	/*
1323 	 * Processor group initialization for this CPU is dependent on the
1324 	 * cpuid probing, which must be done in the context of the current
1325 	 * CPU.
1326 	 */
1327 	pghw_physid_create(cp);
1328 	pg_cpu_init(cp);
1329 	pg_cmt_cpu_startup(cp);
1330 
1331 	cp->cpu_flags |= CPU_RUNNING | CPU_READY | CPU_ENABLE | CPU_EXISTS;
1332 	cpu_add_active(cp);
1333 
1334 	if (dtrace_cpu_init != NULL) {
1335 		(*dtrace_cpu_init)(cp->cpu_id);
1336 	}
1337 
1338 	mutex_exit(&cpu_lock);
1339 
1340 	/*
1341 	 * Enable preemption here so that contention for any locks acquired
1342 	 * later in mp_startup may be preempted if the thread owning those
1343 	 * locks is continously executing on other CPUs (for example, this
1344 	 * CPU must be preemptible to allow other CPUs to pause it during their
1345 	 * startup phases).  It's safe to enable preemption here because the
1346 	 * CPU state is pretty-much fully constructed.
1347 	 */
1348 	curthread->t_preempt = 0;
1349 
1350 	add_cpunode2devtree(cp->cpu_id, cp->cpu_m.mcpu_cpi);
1351 
1352 	/* The base spl should still be at LOCK LEVEL here */
1353 	ASSERT(cp->cpu_base_spl == ipltospl(LOCK_LEVEL));
1354 	set_base_spl();		/* Restore the spl to its proper value */
1355 
1356 	(void) spl0();				/* enable interrupts */
1357 
1358 	/*
1359 	 * Set up the CPU module for this CPU.  This can't be done before
1360 	 * this CPU is made CPU_READY, because we may (in heterogeneous systems)
1361 	 * need to go load another CPU module.  The act of attempting to load
1362 	 * a module may trigger a cross-call, which will ASSERT unless this
1363 	 * cpu is CPU_READY.
1364 	 */
1365 	cmi_init();
1366 
1367 	if (x86_feature & X86_MCA)
1368 		cmi_mca_init();
1369 
1370 	if (boothowto & RB_DEBUG)
1371 		kdi_cpu_init();
1372 
1373 	/*
1374 	 * Setting the bit in cpu_ready_set must be the last operation in
1375 	 * processor initialization; the boot CPU will continue to boot once
1376 	 * it sees this bit set for all active CPUs.
1377 	 */
1378 	CPUSET_ATOMIC_ADD(cpu_ready_set, cp->cpu_id);
1379 
1380 	/*
1381 	 * Because mp_startup() gets fired off after init() starts, we
1382 	 * can't use the '?' trick to do 'boot -v' printing - so we
1383 	 * always direct the 'cpu .. online' messages to the log.
1384 	 */
1385 	cmn_err(CE_CONT, "!cpu%d initialization complete - online\n",
1386 	    cp->cpu_id);
1387 
1388 	/*
1389 	 * Now we are done with the startup thread, so free it up.
1390 	 */
1391 	thread_exit();
1392 	panic("mp_startup: cannot return");
1393 	/*NOTREACHED*/
1394 }
1395 
1396 
1397 /*
1398  * Start CPU on user request.
1399  */
1400 /* ARGSUSED */
1401 int
1402 mp_cpu_start(struct cpu *cp)
1403 {
1404 	ASSERT(MUTEX_HELD(&cpu_lock));
1405 	return (0);
1406 }
1407 
1408 /*
1409  * Stop CPU on user request.
1410  */
1411 /* ARGSUSED */
1412 int
1413 mp_cpu_stop(struct cpu *cp)
1414 {
1415 	extern int cbe_psm_timer_mode;
1416 	ASSERT(MUTEX_HELD(&cpu_lock));
1417 
1418 	/*
1419 	 * If TIMER_PERIODIC mode is used, CPU0 is the one running it;
1420 	 * can't stop it.  (This is true only for machines with no TSC.)
1421 	 */
1422 
1423 	if ((cbe_psm_timer_mode == TIMER_PERIODIC) && (cp->cpu_id == 0))
1424 		return (1);
1425 
1426 	return (0);
1427 }
1428 
1429 /*
1430  * Take the specified CPU out of participation in interrupts.
1431  */
1432 int
1433 cpu_disable_intr(struct cpu *cp)
1434 {
1435 	if (psm_disable_intr(cp->cpu_id) != DDI_SUCCESS)
1436 		return (EBUSY);
1437 
1438 	cp->cpu_flags &= ~CPU_ENABLE;
1439 	return (0);
1440 }
1441 
1442 /*
1443  * Allow the specified CPU to participate in interrupts.
1444  */
1445 void
1446 cpu_enable_intr(struct cpu *cp)
1447 {
1448 	ASSERT(MUTEX_HELD(&cpu_lock));
1449 	cp->cpu_flags |= CPU_ENABLE;
1450 	psm_enable_intr(cp->cpu_id);
1451 }
1452 
1453 
1454 
1455 void
1456 mp_cpu_faulted_enter(struct cpu *cp)
1457 {
1458 	cmi_faulted_enter(cp);
1459 }
1460 
1461 void
1462 mp_cpu_faulted_exit(struct cpu *cp)
1463 {
1464 	cmi_faulted_exit(cp);
1465 }
1466 
1467 /*
1468  * The following two routines are used as context operators on threads belonging
1469  * to processes with a private LDT (see sysi86).  Due to the rarity of such
1470  * processes, these routines are currently written for best code readability and
1471  * organization rather than speed.  We could avoid checking x86_feature at every
1472  * context switch by installing different context ops, depending on the
1473  * x86_feature flags, at LDT creation time -- one for each combination of fast
1474  * syscall feature flags.
1475  */
1476 
1477 /*ARGSUSED*/
1478 void
1479 cpu_fast_syscall_disable(void *arg)
1480 {
1481 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP))
1482 		cpu_sep_disable();
1483 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC))
1484 		cpu_asysc_disable();
1485 }
1486 
1487 /*ARGSUSED*/
1488 void
1489 cpu_fast_syscall_enable(void *arg)
1490 {
1491 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP))
1492 		cpu_sep_enable();
1493 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC))
1494 		cpu_asysc_enable();
1495 }
1496 
1497 static void
1498 cpu_sep_enable(void)
1499 {
1500 	ASSERT(x86_feature & X86_SEP);
1501 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1502 
1503 	wrmsr(MSR_INTC_SEP_CS, (uint64_t)(uintptr_t)KCS_SEL);
1504 }
1505 
1506 static void
1507 cpu_sep_disable(void)
1508 {
1509 	ASSERT(x86_feature & X86_SEP);
1510 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1511 
1512 	/*
1513 	 * Setting the SYSENTER_CS_MSR register to 0 causes software executing
1514 	 * the sysenter or sysexit instruction to trigger a #gp fault.
1515 	 */
1516 	wrmsr(MSR_INTC_SEP_CS, 0);
1517 }
1518 
1519 static void
1520 cpu_asysc_enable(void)
1521 {
1522 	ASSERT(x86_feature & X86_ASYSC);
1523 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1524 
1525 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) |
1526 	    (uint64_t)(uintptr_t)AMD_EFER_SCE);
1527 }
1528 
1529 static void
1530 cpu_asysc_disable(void)
1531 {
1532 	ASSERT(x86_feature & X86_ASYSC);
1533 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1534 
1535 	/*
1536 	 * Turn off the SCE (syscall enable) bit in the EFER register. Software
1537 	 * executing syscall or sysret with this bit off will incur a #ud trap.
1538 	 */
1539 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) &
1540 	    ~((uint64_t)(uintptr_t)AMD_EFER_SCE));
1541 }
1542