xref: /titanic_50/usr/src/uts/i86pc/os/mp_startup.c (revision 4a6b6ac40d4a465eac9d0db6decbd3cf7fdf9349)
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 	hat_cpu_online(cp);
377 
378 #ifdef TRAPTRACE
379 	/*
380 	 * If this is a TRAPTRACE kernel, allocate TRAPTRACE buffers
381 	 */
382 	ttc->ttc_first = (uintptr_t)kmem_zalloc(trap_trace_bufsize, KM_SLEEP);
383 	ttc->ttc_next = ttc->ttc_first;
384 	ttc->ttc_limit = ttc->ttc_first + trap_trace_bufsize;
385 #endif
386 	/*
387 	 * Record that we have another CPU.
388 	 */
389 	mutex_enter(&cpu_lock);
390 	/*
391 	 * Initialize the interrupt threads for this CPU
392 	 */
393 	cpu_intr_alloc(cp, NINTR_THREADS);
394 	/*
395 	 * Add CPU to list of available CPUs.  It'll be on the active list
396 	 * after mp_startup().
397 	 */
398 	cpu_add_unit(cp);
399 	mutex_exit(&cpu_lock);
400 
401 	return (cp);
402 }
403 
404 /*
405  * Undo what was done in mp_startup_init
406  */
407 static void
408 mp_startup_fini(struct cpu *cp, int error)
409 {
410 	mutex_enter(&cpu_lock);
411 
412 	/*
413 	 * Remove the CPU from the list of available CPUs.
414 	 */
415 	cpu_del_unit(cp->cpu_id);
416 
417 	if (error == ETIMEDOUT) {
418 		/*
419 		 * The cpu was started, but never *seemed* to run any
420 		 * code in the kernel; it's probably off spinning in its
421 		 * own private world, though with potential references to
422 		 * our kmem-allocated IDTs and GDTs (for example).
423 		 *
424 		 * Worse still, it may actually wake up some time later,
425 		 * so rather than guess what it might or might not do, we
426 		 * leave the fundamental data structures intact.
427 		 */
428 		cp->cpu_flags = 0;
429 		mutex_exit(&cpu_lock);
430 		return;
431 	}
432 
433 	/*
434 	 * At this point, the only threads bound to this CPU should
435 	 * special per-cpu threads: it's idle thread, it's pause threads,
436 	 * and it's interrupt threads.  Clean these up.
437 	 */
438 	cpu_destroy_bound_threads(cp);
439 	cp->cpu_idle_thread = NULL;
440 
441 	/*
442 	 * Free the interrupt stack.
443 	 */
444 	segkp_release(segkp,
445 	    cp->cpu_intr_stack - (INTR_STACK_SIZE - SA(MINFRAME)));
446 
447 	mutex_exit(&cpu_lock);
448 
449 #ifdef TRAPTRACE
450 	/*
451 	 * Discard the trap trace buffer
452 	 */
453 	{
454 		trap_trace_ctl_t *ttc = &trap_trace_ctl[cp->cpu_id];
455 
456 		kmem_free((void *)ttc->ttc_first, trap_trace_bufsize);
457 		ttc->ttc_first = NULL;
458 	}
459 #endif
460 
461 	hat_cpu_offline(cp);
462 
463 	cpuid_free_space(cp);
464 
465 	if (cp->cpu_m.mcpu_idt != CPU->cpu_m.mcpu_idt)
466 		kmem_free(cp->cpu_m.mcpu_idt, sizeof (idt0));
467 	cp->cpu_m.mcpu_idt = NULL;
468 
469 	kmem_free(cp->cpu_m.mcpu_gdt, PAGESIZE);
470 	cp->cpu_m.mcpu_gdt = NULL;
471 
472 	teardown_vaddr_for_ppcopy(cp);
473 
474 	kcpc_hw_fini(cp);
475 
476 	cp->cpu_dispthread = NULL;
477 	cp->cpu_thread = NULL;	/* discarded by cpu_destroy_bound_threads() */
478 
479 	cpu_vm_data_destroy(cp);
480 
481 	mutex_enter(&cpu_lock);
482 	disp_cpu_fini(cp);
483 	mutex_exit(&cpu_lock);
484 
485 	kmem_free(cp, sizeof (*cp));
486 }
487 
488 /*
489  * Apply workarounds for known errata, and warn about those that are absent.
490  *
491  * System vendors occasionally create configurations which contain different
492  * revisions of the CPUs that are almost but not exactly the same.  At the
493  * time of writing, this meant that their clock rates were the same, their
494  * feature sets were the same, but the required workaround were -not-
495  * necessarily the same.  So, this routine is invoked on -every- CPU soon
496  * after starting to make sure that the resulting system contains the most
497  * pessimal set of workarounds needed to cope with *any* of the CPUs in the
498  * system.
499  *
500  * workaround_errata is invoked early in mlsetup() for CPU 0, and in
501  * mp_startup() for all slave CPUs. Slaves process workaround_errata prior
502  * to acknowledging their readiness to the master, so this routine will
503  * never be executed by multiple CPUs in parallel, thus making updates to
504  * global data safe.
505  *
506  * These workarounds are based on Rev 3.57 of the Revision Guide for
507  * AMD Athlon(tm) 64 and AMD Opteron(tm) Processors, August 2005.
508  */
509 
510 #if defined(OPTERON_ERRATUM_88)
511 int opteron_erratum_88;		/* if non-zero -> at least one cpu has it */
512 #endif
513 
514 #if defined(OPTERON_ERRATUM_91)
515 int opteron_erratum_91;		/* if non-zero -> at least one cpu has it */
516 #endif
517 
518 #if defined(OPTERON_ERRATUM_93)
519 int opteron_erratum_93;		/* if non-zero -> at least one cpu has it */
520 #endif
521 
522 #if defined(OPTERON_ERRATUM_95)
523 int opteron_erratum_95;		/* if non-zero -> at least one cpu has it */
524 #endif
525 
526 #if defined(OPTERON_ERRATUM_100)
527 int opteron_erratum_100;	/* if non-zero -> at least one cpu has it */
528 #endif
529 
530 #if defined(OPTERON_ERRATUM_108)
531 int opteron_erratum_108;	/* if non-zero -> at least one cpu has it */
532 #endif
533 
534 #if defined(OPTERON_ERRATUM_109)
535 int opteron_erratum_109;	/* if non-zero -> at least one cpu has it */
536 #endif
537 
538 #if defined(OPTERON_ERRATUM_121)
539 int opteron_erratum_121;	/* if non-zero -> at least one cpu has it */
540 #endif
541 
542 #if defined(OPTERON_ERRATUM_122)
543 int opteron_erratum_122;	/* if non-zero -> at least one cpu has it */
544 #endif
545 
546 #if defined(OPTERON_ERRATUM_123)
547 int opteron_erratum_123;	/* if non-zero -> at least one cpu has it */
548 #endif
549 
550 #if defined(OPTERON_ERRATUM_131)
551 int opteron_erratum_131;	/* if non-zero -> at least one cpu has it */
552 #endif
553 
554 #if defined(OPTERON_WORKAROUND_6336786)
555 int opteron_workaround_6336786;	/* non-zero -> WA relevant and applied */
556 int opteron_workaround_6336786_UP = 0;	/* Not needed for UP */
557 #endif
558 
559 #if defined(OPTERON_WORKAROUND_6323525)
560 int opteron_workaround_6323525;	/* if non-zero -> at least one cpu has it */
561 #endif
562 
563 static void
564 workaround_warning(cpu_t *cp, uint_t erratum)
565 {
566 	cmn_err(CE_WARN, "cpu%d: no workaround for erratum %u",
567 	    cp->cpu_id, erratum);
568 }
569 
570 static void
571 workaround_applied(uint_t erratum)
572 {
573 	if (erratum > 1000000)
574 		cmn_err(CE_CONT, "?workaround applied for cpu issue #%d\n",
575 		    erratum);
576 	else
577 		cmn_err(CE_CONT, "?workaround applied for cpu erratum #%d\n",
578 		    erratum);
579 }
580 
581 static void
582 msr_warning(cpu_t *cp, const char *rw, uint_t msr, int error)
583 {
584 	cmn_err(CE_WARN, "cpu%d: couldn't %smsr 0x%x, error %d",
585 	    cp->cpu_id, rw, msr, error);
586 }
587 
588 uint_t
589 workaround_errata(struct cpu *cpu)
590 {
591 	uint_t missing = 0;
592 
593 	ASSERT(cpu == CPU);
594 
595 	/*LINTED*/
596 	if (cpuid_opteron_erratum(cpu, 88) > 0) {
597 		/*
598 		 * SWAPGS May Fail To Read Correct GS Base
599 		 */
600 #if defined(OPTERON_ERRATUM_88)
601 		/*
602 		 * The workaround is an mfence in the relevant assembler code
603 		 */
604 		opteron_erratum_88++;
605 #else
606 		workaround_warning(cpu, 88);
607 		missing++;
608 #endif
609 	}
610 
611 	if (cpuid_opteron_erratum(cpu, 91) > 0) {
612 		/*
613 		 * Software Prefetches May Report A Page Fault
614 		 */
615 #if defined(OPTERON_ERRATUM_91)
616 		/*
617 		 * fix is in trap.c
618 		 */
619 		opteron_erratum_91++;
620 #else
621 		workaround_warning(cpu, 91);
622 		missing++;
623 #endif
624 	}
625 
626 	if (cpuid_opteron_erratum(cpu, 93) > 0) {
627 		/*
628 		 * RSM Auto-Halt Restart Returns to Incorrect RIP
629 		 */
630 #if defined(OPTERON_ERRATUM_93)
631 		/*
632 		 * fix is in trap.c
633 		 */
634 		opteron_erratum_93++;
635 #else
636 		workaround_warning(cpu, 93);
637 		missing++;
638 #endif
639 	}
640 
641 	/*LINTED*/
642 	if (cpuid_opteron_erratum(cpu, 95) > 0) {
643 		/*
644 		 * RET Instruction May Return to Incorrect EIP
645 		 */
646 #if defined(OPTERON_ERRATUM_95)
647 #if defined(_LP64)
648 		/*
649 		 * Workaround this by ensuring that 32-bit user code and
650 		 * 64-bit kernel code never occupy the same address
651 		 * range mod 4G.
652 		 */
653 		if (_userlimit32 > 0xc0000000ul)
654 			*(uintptr_t *)&_userlimit32 = 0xc0000000ul;
655 
656 		/*LINTED*/
657 		ASSERT((uint32_t)COREHEAP_BASE == 0xc0000000u);
658 		opteron_erratum_95++;
659 #endif	/* _LP64 */
660 #else
661 		workaround_warning(cpu, 95);
662 		missing++;
663 #endif
664 	}
665 
666 	if (cpuid_opteron_erratum(cpu, 100) > 0) {
667 		/*
668 		 * Compatibility Mode Branches Transfer to Illegal Address
669 		 */
670 #if defined(OPTERON_ERRATUM_100)
671 		/*
672 		 * fix is in trap.c
673 		 */
674 		opteron_erratum_100++;
675 #else
676 		workaround_warning(cpu, 100);
677 		missing++;
678 #endif
679 	}
680 
681 	/*LINTED*/
682 	if (cpuid_opteron_erratum(cpu, 108) > 0) {
683 		/*
684 		 * CPUID Instruction May Return Incorrect Model Number In
685 		 * Some Processors
686 		 */
687 #if defined(OPTERON_ERRATUM_108)
688 		/*
689 		 * (Our cpuid-handling code corrects the model number on
690 		 * those processors)
691 		 */
692 #else
693 		workaround_warning(cpu, 108);
694 		missing++;
695 #endif
696 	}
697 
698 	/*LINTED*/
699 	if (cpuid_opteron_erratum(cpu, 109) > 0) do {
700 		/*
701 		 * Certain Reverse REP MOVS May Produce Unpredictable Behaviour
702 		 */
703 #if defined(OPTERON_ERRATUM_109)
704 		/*
705 		 * The "workaround" is to print a warning to upgrade the BIOS
706 		 */
707 		uint64_t value;
708 		const uint_t msr = MSR_AMD_PATCHLEVEL;
709 		int err;
710 
711 		if ((err = checked_rdmsr(msr, &value)) != 0) {
712 			msr_warning(cpu, "rd", msr, err);
713 			workaround_warning(cpu, 109);
714 			missing++;
715 		}
716 		if (value == 0)
717 			opteron_erratum_109++;
718 #else
719 		workaround_warning(cpu, 109);
720 		missing++;
721 #endif
722 	/*CONSTANTCONDITION*/
723 	} while (0);
724 
725 	/*LINTED*/
726 	if (cpuid_opteron_erratum(cpu, 121) > 0) {
727 		/*
728 		 * Sequential Execution Across Non_Canonical Boundary Caused
729 		 * Processor Hang
730 		 */
731 #if defined(OPTERON_ERRATUM_121)
732 #if defined(_LP64)
733 		/*
734 		 * Erratum 121 is only present in long (64 bit) mode.
735 		 * Workaround is to include the page immediately before the
736 		 * va hole to eliminate the possibility of system hangs due to
737 		 * sequential execution across the va hole boundary.
738 		 */
739 		if (opteron_erratum_121)
740 			opteron_erratum_121++;
741 		else {
742 			if (hole_start) {
743 				hole_start -= PAGESIZE;
744 			} else {
745 				/*
746 				 * hole_start not yet initialized by
747 				 * mmu_init. Initialize hole_start
748 				 * with value to be subtracted.
749 				 */
750 				hole_start = PAGESIZE;
751 			}
752 			opteron_erratum_121++;
753 		}
754 #endif	/* _LP64 */
755 #else
756 		workaround_warning(cpu, 121);
757 		missing++;
758 #endif
759 	}
760 
761 	/*LINTED*/
762 	if (cpuid_opteron_erratum(cpu, 122) > 0) do {
763 		/*
764 		 * TLB Flush Filter May Cause Coherency Problem in
765 		 * Multiprocessor Systems
766 		 */
767 #if defined(OPTERON_ERRATUM_122)
768 		uint64_t value;
769 		const uint_t msr = MSR_AMD_HWCR;
770 		int error;
771 
772 		/*
773 		 * Erratum 122 is only present in MP configurations (multi-core
774 		 * or multi-processor).
775 		 */
776 		if (!opteron_erratum_122 && lgrp_plat_node_cnt == 1 &&
777 		    cpuid_get_ncpu_per_chip(cpu) == 1)
778 			break;
779 
780 		/* disable TLB Flush Filter */
781 
782 		if ((error = checked_rdmsr(msr, &value)) != 0) {
783 			msr_warning(cpu, "rd", msr, error);
784 			workaround_warning(cpu, 122);
785 			missing++;
786 		} else {
787 			value |= (uint64_t)AMD_HWCR_FFDIS;
788 			if ((error = checked_wrmsr(msr, value)) != 0) {
789 				msr_warning(cpu, "wr", msr, error);
790 				workaround_warning(cpu, 122);
791 				missing++;
792 			}
793 		}
794 		opteron_erratum_122++;
795 #else
796 		workaround_warning(cpu, 122);
797 		missing++;
798 #endif
799 	/*CONSTANTCONDITION*/
800 	} while (0);
801 
802 	/*LINTED*/
803 	if (cpuid_opteron_erratum(cpu, 123) > 0) do {
804 		/*
805 		 * Bypassed Reads May Cause Data Corruption of System Hang in
806 		 * Dual Core Processors
807 		 */
808 #if defined(OPTERON_ERRATUM_123)
809 		uint64_t value;
810 		const uint_t msr = MSR_AMD_PATCHLEVEL;
811 		int err;
812 
813 		/*
814 		 * Erratum 123 applies only to multi-core cpus.
815 		 */
816 		if (cpuid_get_ncpu_per_chip(cpu) < 2)
817 			break;
818 
819 		/*
820 		 * The "workaround" is to print a warning to upgrade the BIOS
821 		 */
822 		if ((err = checked_rdmsr(msr, &value)) != 0) {
823 			msr_warning(cpu, "rd", msr, err);
824 			workaround_warning(cpu, 123);
825 			missing++;
826 		}
827 		if (value == 0)
828 			opteron_erratum_123++;
829 #else
830 		workaround_warning(cpu, 123);
831 		missing++;
832 
833 #endif
834 	/*CONSTANTCONDITION*/
835 	} while (0);
836 
837 	/*LINTED*/
838 	if (cpuid_opteron_erratum(cpu, 131) > 0) do {
839 		/*
840 		 * Multiprocessor Systems with Four or More Cores May Deadlock
841 		 * Waiting for a Probe Response
842 		 */
843 #if defined(OPTERON_ERRATUM_131)
844 		uint64_t nbcfg;
845 		const uint_t msr = MSR_AMD_NB_CFG;
846 		const uint64_t wabits =
847 		    AMD_NB_CFG_SRQ_HEARTBEAT | AMD_NB_CFG_SRQ_SPR;
848 		int error;
849 
850 		/*
851 		 * Erratum 131 applies to any system with four or more cores.
852 		 */
853 		if (opteron_erratum_131)
854 			break;
855 
856 		if (lgrp_plat_node_cnt * cpuid_get_ncpu_per_chip(cpu) < 4)
857 			break;
858 
859 		/*
860 		 * Print a warning if neither of the workarounds for
861 		 * erratum 131 is present.
862 		 */
863 		if ((error = checked_rdmsr(msr, &nbcfg)) != 0) {
864 			msr_warning(cpu, "rd", msr, error);
865 			workaround_warning(cpu, 131);
866 			missing++;
867 		} else if ((nbcfg & wabits) == 0) {
868 			opteron_erratum_131++;
869 		} else {
870 			/* cannot have both workarounds set */
871 			ASSERT((nbcfg & wabits) != wabits);
872 		}
873 #else
874 		workaround_warning(cpu, 131);
875 		missing++;
876 #endif
877 	/*CONSTANTCONDITION*/
878 	} while (0);
879 
880 	/*
881 	 * This isn't really an erratum, but for convenience the
882 	 * detection/workaround code lives here and in cpuid_opteron_erratum.
883 	 */
884 	if (cpuid_opteron_erratum(cpu, 6336786) > 0) {
885 #if defined(OPTERON_WORKAROUND_6336786)
886 		/*
887 		 * Disable C1-Clock ramping on multi-core/multi-processor
888 		 * K8 platforms to guard against TSC drift.
889 		 */
890 		if (opteron_workaround_6336786) {
891 			opteron_workaround_6336786++;
892 		} else if ((lgrp_plat_node_cnt *
893 		    cpuid_get_ncpu_per_chip(cpu) > 1) ||
894 		    opteron_workaround_6336786_UP) {
895 			int	node;
896 			uint8_t data;
897 
898 			for (node = 0; node < lgrp_plat_node_cnt; node++) {
899 				/*
900 				 * Clear PMM7[1:0] (function 3, offset 0x87)
901 				 * Northbridge device is the node id + 24.
902 				 */
903 				data = pci_getb_func(0, node + 24, 3, 0x87);
904 				data &= 0xFC;
905 				pci_putb_func(0, node + 24, 3, 0x87, data);
906 			}
907 			opteron_workaround_6336786++;
908 		}
909 #else
910 		workaround_warning(cpu, 6336786);
911 		missing++;
912 #endif
913 	}
914 
915 	/*LINTED*/
916 	/*
917 	 * Mutex primitives don't work as expected.
918 	 */
919 	if (cpuid_opteron_erratum(cpu, 6323525) > 0) {
920 #if defined(OPTERON_WORKAROUND_6323525)
921 		/*
922 		 * This problem only occurs with 2 or more cores. If bit in
923 		 * MSR_BU_CFG set, then not applicable. The workaround
924 		 * is to patch the semaphone routines with the lfence
925 		 * instruction to provide necessary load memory barrier with
926 		 * possible subsequent read-modify-write ops.
927 		 *
928 		 * It is too early in boot to call the patch routine so
929 		 * set erratum variable to be done in startup_end().
930 		 */
931 		if (opteron_workaround_6323525) {
932 			opteron_workaround_6323525++;
933 		} else if ((x86_feature & X86_SSE2) && ((lgrp_plat_node_cnt *
934 		    cpuid_get_ncpu_per_chip(cpu)) > 1)) {
935 			if ((xrdmsr(MSR_BU_CFG) & 0x02) == 0)
936 				opteron_workaround_6323525++;
937 		}
938 #else
939 		workaround_warning(cpu, 6323525);
940 		missing++;
941 #endif
942 	}
943 
944 	return (missing);
945 }
946 
947 void
948 workaround_errata_end()
949 {
950 #if defined(OPTERON_ERRATUM_88)
951 	if (opteron_erratum_88)
952 		workaround_applied(88);
953 #endif
954 #if defined(OPTERON_ERRATUM_91)
955 	if (opteron_erratum_91)
956 		workaround_applied(91);
957 #endif
958 #if defined(OPTERON_ERRATUM_93)
959 	if (opteron_erratum_93)
960 		workaround_applied(93);
961 #endif
962 #if defined(OPTERON_ERRATUM_95)
963 	if (opteron_erratum_95)
964 		workaround_applied(95);
965 #endif
966 #if defined(OPTERON_ERRATUM_100)
967 	if (opteron_erratum_100)
968 		workaround_applied(100);
969 #endif
970 #if defined(OPTERON_ERRATUM_108)
971 	if (opteron_erratum_108)
972 		workaround_applied(108);
973 #endif
974 #if defined(OPTERON_ERRATUM_109)
975 	if (opteron_erratum_109) {
976 		cmn_err(CE_WARN,
977 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
978 		    " processor\nerratum 109 was not detected; updating your"
979 		    " system's BIOS to a version\ncontaining this"
980 		    " microcode patch is HIGHLY recommended or erroneous"
981 		    " system\noperation may occur.\n");
982 	}
983 #endif
984 #if defined(OPTERON_ERRATUM_121)
985 	if (opteron_erratum_121)
986 		workaround_applied(121);
987 #endif
988 #if defined(OPTERON_ERRATUM_122)
989 	if (opteron_erratum_122)
990 		workaround_applied(122);
991 #endif
992 #if defined(OPTERON_ERRATUM_123)
993 	if (opteron_erratum_123) {
994 		cmn_err(CE_WARN,
995 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
996 		    " processor\nerratum 123 was not detected; updating your"
997 		    " system's BIOS to a version\ncontaining this"
998 		    " microcode patch is HIGHLY recommended or erroneous"
999 		    " system\noperation may occur.\n");
1000 	}
1001 #endif
1002 #if defined(OPTERON_ERRATUM_131)
1003 	if (opteron_erratum_131) {
1004 		cmn_err(CE_WARN,
1005 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1006 		    " processor\nerratum 131 was not detected; updating your"
1007 		    " system's BIOS to a version\ncontaining this"
1008 		    " microcode patch is HIGHLY recommended or erroneous"
1009 		    " system\noperation may occur.\n");
1010 	}
1011 #endif
1012 #if defined(OPTERON_WORKAROUND_6336786)
1013 	if (opteron_workaround_6336786)
1014 		workaround_applied(6336786);
1015 #endif
1016 #if defined(OPTERON_WORKAROUND_6323525)
1017 	if (opteron_workaround_6323525)
1018 		workaround_applied(6323525);
1019 #endif
1020 }
1021 
1022 static cpuset_t procset;
1023 
1024 /*
1025  * Start a single cpu, assuming that the kernel context is available
1026  * to successfully start another cpu.
1027  *
1028  * (For example, real mode code is mapped into the right place
1029  * in memory and is ready to be run.)
1030  */
1031 int
1032 start_cpu(processorid_t who)
1033 {
1034 	void *ctx;
1035 	cpu_t *cp;
1036 	int delays;
1037 	int error = 0;
1038 
1039 	ASSERT(who != 0);
1040 
1041 	/*
1042 	 * Check if there's at least a Mbyte of kmem available
1043 	 * before attempting to start the cpu.
1044 	 */
1045 	if (kmem_avail() < 1024 * 1024) {
1046 		/*
1047 		 * Kick off a reap in case that helps us with
1048 		 * later attempts ..
1049 		 */
1050 		kmem_reap();
1051 		return (ENOMEM);
1052 	}
1053 
1054 	cp = mp_startup_init(who);
1055 	if ((ctx = mach_cpucontext_alloc(cp)) == NULL ||
1056 	    (error = mach_cpu_start(cp, ctx)) != 0) {
1057 
1058 		/*
1059 		 * Something went wrong before we even started it
1060 		 */
1061 		if (ctx)
1062 			cmn_err(CE_WARN,
1063 			    "cpu%d: failed to start error %d",
1064 			    cp->cpu_id, error);
1065 		else
1066 			cmn_err(CE_WARN,
1067 			    "cpu%d: failed to allocate context", cp->cpu_id);
1068 
1069 		if (ctx)
1070 			mach_cpucontext_free(cp, ctx, error);
1071 		else
1072 			error = EAGAIN;		/* hmm. */
1073 		mp_startup_fini(cp, error);
1074 		return (error);
1075 	}
1076 
1077 	for (delays = 0; !CPU_IN_SET(procset, who); delays++) {
1078 		if (delays == 500) {
1079 			/*
1080 			 * After five seconds, things are probably looking
1081 			 * a bit bleak - explain the hang.
1082 			 */
1083 			cmn_err(CE_NOTE, "cpu%d: started, "
1084 			    "but not running in the kernel yet", who);
1085 		} else if (delays > 2000) {
1086 			/*
1087 			 * We waited at least 20 seconds, bail ..
1088 			 */
1089 			error = ETIMEDOUT;
1090 			cmn_err(CE_WARN, "cpu%d: timed out", who);
1091 			mach_cpucontext_free(cp, ctx, error);
1092 			mp_startup_fini(cp, error);
1093 			return (error);
1094 		}
1095 
1096 		/*
1097 		 * wait at least 10ms, then check again..
1098 		 */
1099 		delay(USEC_TO_TICK_ROUNDUP(10000));
1100 	}
1101 
1102 	mach_cpucontext_free(cp, ctx, 0);
1103 
1104 	if (tsc_gethrtime_enable)
1105 		tsc_sync_master(who);
1106 
1107 	if (dtrace_cpu_init != NULL) {
1108 		/*
1109 		 * DTrace CPU initialization expects cpu_lock to be held.
1110 		 */
1111 		mutex_enter(&cpu_lock);
1112 		(*dtrace_cpu_init)(who);
1113 		mutex_exit(&cpu_lock);
1114 	}
1115 
1116 	while (!CPU_IN_SET(cpu_ready_set, who))
1117 		delay(1);
1118 
1119 	return (0);
1120 }
1121 
1122 
1123 /*ARGSUSED*/
1124 void
1125 start_other_cpus(int cprboot)
1126 {
1127 	uint_t who;
1128 	uint_t skipped = 0;
1129 	uint_t bootcpuid = 0;
1130 
1131 	/*
1132 	 * Initialize our own cpu_info.
1133 	 */
1134 	init_cpu_info(CPU);
1135 
1136 	/*
1137 	 * Initialize our syscall handlers
1138 	 */
1139 	init_cpu_syscall(CPU);
1140 
1141 	/*
1142 	 * Take the boot cpu out of the mp_cpus set because we know
1143 	 * it's already running.  Add it to the cpu_ready_set for
1144 	 * precisely the same reason.
1145 	 */
1146 	CPUSET_DEL(mp_cpus, bootcpuid);
1147 	CPUSET_ADD(cpu_ready_set, bootcpuid);
1148 
1149 	/*
1150 	 * if only 1 cpu or not using MP, skip the rest of this
1151 	 */
1152 	if (CPUSET_ISNULL(mp_cpus) || use_mp == 0) {
1153 		if (use_mp == 0)
1154 			cmn_err(CE_CONT, "?***** Not in MP mode\n");
1155 		goto done;
1156 	}
1157 
1158 	/*
1159 	 * perform such initialization as is needed
1160 	 * to be able to take CPUs on- and off-line.
1161 	 */
1162 	cpu_pause_init();
1163 
1164 	xc_init();		/* initialize processor crosscalls */
1165 
1166 	if (mach_cpucontext_init() != 0)
1167 		goto done;
1168 
1169 	flushes_require_xcalls = 1;
1170 
1171 	/*
1172 	 * We lock our affinity to the master CPU to ensure that all slave CPUs
1173 	 * do their TSC syncs with the same CPU.
1174 	 */
1175 	affinity_set(CPU_CURRENT);
1176 
1177 	for (who = 0; who < NCPU; who++) {
1178 
1179 		if (!CPU_IN_SET(mp_cpus, who))
1180 			continue;
1181 		ASSERT(who != bootcpuid);
1182 		if (ncpus >= max_ncpus) {
1183 			skipped = who;
1184 			continue;
1185 		}
1186 		if (start_cpu(who) != 0)
1187 			CPUSET_DEL(mp_cpus, who);
1188 	}
1189 
1190 	affinity_clear();
1191 
1192 	if (skipped) {
1193 		cmn_err(CE_NOTE,
1194 		    "System detected %d cpus, but "
1195 		    "only %d cpu(s) were enabled during boot.",
1196 		    skipped + 1, ncpus);
1197 		cmn_err(CE_NOTE,
1198 		    "Use \"boot-ncpus\" parameter to enable more CPU(s). "
1199 		    "See eeprom(1M).");
1200 	}
1201 
1202 done:
1203 	workaround_errata_end();
1204 	mach_cpucontext_fini();
1205 
1206 	cmi_post_mpstartup();
1207 }
1208 
1209 /*
1210  * Dummy functions - no i86pc platforms support dynamic cpu allocation.
1211  */
1212 /*ARGSUSED*/
1213 int
1214 mp_cpu_configure(int cpuid)
1215 {
1216 	return (ENOTSUP);		/* not supported */
1217 }
1218 
1219 /*ARGSUSED*/
1220 int
1221 mp_cpu_unconfigure(int cpuid)
1222 {
1223 	return (ENOTSUP);		/* not supported */
1224 }
1225 
1226 /*
1227  * Startup function for 'other' CPUs (besides boot cpu).
1228  * Called from real_mode_start.
1229  *
1230  * WARNING: until CPU_READY is set, mp_startup and routines called by
1231  * mp_startup should not call routines (e.g. kmem_free) that could call
1232  * hat_unload which requires CPU_READY to be set.
1233  */
1234 void
1235 mp_startup(void)
1236 {
1237 	struct cpu *cp = CPU;
1238 	uint_t new_x86_feature;
1239 
1240 	/*
1241 	 * We need to get TSC on this proc synced (i.e., any delta
1242 	 * from cpu0 accounted for) as soon as we can, because many
1243 	 * many things use gethrtime/pc_gethrestime, including
1244 	 * interrupts, cmn_err, etc.
1245 	 */
1246 
1247 	/* Let cpu0 continue into tsc_sync_master() */
1248 	CPUSET_ATOMIC_ADD(procset, cp->cpu_id);
1249 
1250 	if (tsc_gethrtime_enable)
1251 		tsc_sync_slave();
1252 
1253 	/*
1254 	 * Once this was done from assembly, but it's safer here; if
1255 	 * it blocks, we need to be able to swtch() to and from, and
1256 	 * since we get here by calling t_pc, we need to do that call
1257 	 * before swtch() overwrites it.
1258 	 */
1259 
1260 	(void) (*ap_mlsetup)();
1261 
1262 	new_x86_feature = cpuid_pass1(cp);
1263 
1264 	/*
1265 	 * We need to Sync MTRR with cpu0's MTRR. We have to do
1266 	 * this with interrupts disabled.
1267 	 */
1268 	if (x86_feature & X86_MTRR)
1269 		mtrr_sync();
1270 
1271 	/*
1272 	 * Set up TSC_AUX to contain the cpuid for this processor
1273 	 * for the rdtscp instruction.
1274 	 */
1275 	if (x86_feature & X86_TSCP)
1276 		(void) wrmsr(MSR_AMD_TSCAUX, cp->cpu_id);
1277 
1278 	/*
1279 	 * Initialize this CPU's syscall handlers
1280 	 */
1281 	init_cpu_syscall(cp);
1282 
1283 	/*
1284 	 * Enable interrupts with spl set to LOCK_LEVEL. LOCK_LEVEL is the
1285 	 * highest level at which a routine is permitted to block on
1286 	 * an adaptive mutex (allows for cpu poke interrupt in case
1287 	 * the cpu is blocked on a mutex and halts). Setting LOCK_LEVEL blocks
1288 	 * device interrupts that may end up in the hat layer issuing cross
1289 	 * calls before CPU_READY is set.
1290 	 */
1291 	splx(ipltospl(LOCK_LEVEL));
1292 	sti();
1293 
1294 	/*
1295 	 * Do a sanity check to make sure this new CPU is a sane thing
1296 	 * to add to the collection of processors running this system.
1297 	 *
1298 	 * XXX	Clearly this needs to get more sophisticated, if x86
1299 	 * systems start to get built out of heterogenous CPUs; as is
1300 	 * likely to happen once the number of processors in a configuration
1301 	 * gets large enough.
1302 	 */
1303 	if ((x86_feature & new_x86_feature) != x86_feature) {
1304 		cmn_err(CE_CONT, "?cpu%d: %b\n",
1305 		    cp->cpu_id, new_x86_feature, FMT_X86_FEATURE);
1306 		cmn_err(CE_WARN, "cpu%d feature mismatch", cp->cpu_id);
1307 	}
1308 
1309 	/*
1310 	 * We do not support cpus with mixed monitor/mwait support if the
1311 	 * boot cpu supports monitor/mwait.
1312 	 */
1313 	if ((x86_feature & ~new_x86_feature) & X86_MWAIT)
1314 		panic("unsupported mixed cpu monitor/mwait support detected");
1315 
1316 	/*
1317 	 * We could be more sophisticated here, and just mark the CPU
1318 	 * as "faulted" but at this point we'll opt for the easier
1319 	 * answer of dieing horribly.  Provided the boot cpu is ok,
1320 	 * the system can be recovered by booting with use_mp set to zero.
1321 	 */
1322 	if (workaround_errata(cp) != 0)
1323 		panic("critical workaround(s) missing for cpu%d", cp->cpu_id);
1324 
1325 	cpuid_pass2(cp);
1326 	cpuid_pass3(cp);
1327 	(void) cpuid_pass4(cp);
1328 
1329 	init_cpu_info(cp);
1330 
1331 	mutex_enter(&cpu_lock);
1332 	/*
1333 	 * Processor group initialization for this CPU is dependent on the
1334 	 * cpuid probing, which must be done in the context of the current
1335 	 * CPU.
1336 	 */
1337 	pghw_physid_create(cp);
1338 	pg_cpu_init(cp);
1339 	pg_cmt_cpu_startup(cp);
1340 
1341 	cp->cpu_flags |= CPU_RUNNING | CPU_READY | CPU_ENABLE | CPU_EXISTS;
1342 	cpu_add_active(cp);
1343 
1344 	if (dtrace_cpu_init != NULL) {
1345 		(*dtrace_cpu_init)(cp->cpu_id);
1346 	}
1347 
1348 	mutex_exit(&cpu_lock);
1349 
1350 	/*
1351 	 * Enable preemption here so that contention for any locks acquired
1352 	 * later in mp_startup may be preempted if the thread owning those
1353 	 * locks is continously executing on other CPUs (for example, this
1354 	 * CPU must be preemptible to allow other CPUs to pause it during their
1355 	 * startup phases).  It's safe to enable preemption here because the
1356 	 * CPU state is pretty-much fully constructed.
1357 	 */
1358 	curthread->t_preempt = 0;
1359 
1360 	add_cpunode2devtree(cp->cpu_id, cp->cpu_m.mcpu_cpi);
1361 
1362 	/* The base spl should still be at LOCK LEVEL here */
1363 	ASSERT(cp->cpu_base_spl == ipltospl(LOCK_LEVEL));
1364 	set_base_spl();		/* Restore the spl to its proper value */
1365 
1366 	(void) spl0();				/* enable interrupts */
1367 
1368 	/*
1369 	 * Set up the CPU module for this CPU.  This can't be done before
1370 	 * this CPU is made CPU_READY, because we may (in heterogeneous systems)
1371 	 * need to go load another CPU module.  The act of attempting to load
1372 	 * a module may trigger a cross-call, which will ASSERT unless this
1373 	 * cpu is CPU_READY.
1374 	 */
1375 	cmi_init();
1376 
1377 	if (x86_feature & X86_MCA)
1378 		cmi_mca_init();
1379 
1380 	if (boothowto & RB_DEBUG)
1381 		kdi_cpu_init();
1382 
1383 	/*
1384 	 * Setting the bit in cpu_ready_set must be the last operation in
1385 	 * processor initialization; the boot CPU will continue to boot once
1386 	 * it sees this bit set for all active CPUs.
1387 	 */
1388 	CPUSET_ATOMIC_ADD(cpu_ready_set, cp->cpu_id);
1389 
1390 	/*
1391 	 * Because mp_startup() gets fired off after init() starts, we
1392 	 * can't use the '?' trick to do 'boot -v' printing - so we
1393 	 * always direct the 'cpu .. online' messages to the log.
1394 	 */
1395 	cmn_err(CE_CONT, "!cpu%d initialization complete - online\n",
1396 	    cp->cpu_id);
1397 
1398 	/*
1399 	 * Now we are done with the startup thread, so free it up.
1400 	 */
1401 	thread_exit();
1402 	panic("mp_startup: cannot return");
1403 	/*NOTREACHED*/
1404 }
1405 
1406 
1407 /*
1408  * Start CPU on user request.
1409  */
1410 /* ARGSUSED */
1411 int
1412 mp_cpu_start(struct cpu *cp)
1413 {
1414 	ASSERT(MUTEX_HELD(&cpu_lock));
1415 	return (0);
1416 }
1417 
1418 /*
1419  * Stop CPU on user request.
1420  */
1421 /* ARGSUSED */
1422 int
1423 mp_cpu_stop(struct cpu *cp)
1424 {
1425 	extern int cbe_psm_timer_mode;
1426 	ASSERT(MUTEX_HELD(&cpu_lock));
1427 
1428 	/*
1429 	 * If TIMER_PERIODIC mode is used, CPU0 is the one running it;
1430 	 * can't stop it.  (This is true only for machines with no TSC.)
1431 	 */
1432 
1433 	if ((cbe_psm_timer_mode == TIMER_PERIODIC) && (cp->cpu_id == 0))
1434 		return (1);
1435 
1436 	return (0);
1437 }
1438 
1439 /*
1440  * Take the specified CPU out of participation in interrupts.
1441  */
1442 int
1443 cpu_disable_intr(struct cpu *cp)
1444 {
1445 	if (psm_disable_intr(cp->cpu_id) != DDI_SUCCESS)
1446 		return (EBUSY);
1447 
1448 	cp->cpu_flags &= ~CPU_ENABLE;
1449 	return (0);
1450 }
1451 
1452 /*
1453  * Allow the specified CPU to participate in interrupts.
1454  */
1455 void
1456 cpu_enable_intr(struct cpu *cp)
1457 {
1458 	ASSERT(MUTEX_HELD(&cpu_lock));
1459 	cp->cpu_flags |= CPU_ENABLE;
1460 	psm_enable_intr(cp->cpu_id);
1461 }
1462 
1463 
1464 
1465 void
1466 mp_cpu_faulted_enter(struct cpu *cp)
1467 {
1468 	cmi_faulted_enter(cp);
1469 }
1470 
1471 void
1472 mp_cpu_faulted_exit(struct cpu *cp)
1473 {
1474 	cmi_faulted_exit(cp);
1475 }
1476 
1477 /*
1478  * The following two routines are used as context operators on threads belonging
1479  * to processes with a private LDT (see sysi86).  Due to the rarity of such
1480  * processes, these routines are currently written for best code readability and
1481  * organization rather than speed.  We could avoid checking x86_feature at every
1482  * context switch by installing different context ops, depending on the
1483  * x86_feature flags, at LDT creation time -- one for each combination of fast
1484  * syscall feature flags.
1485  */
1486 
1487 /*ARGSUSED*/
1488 void
1489 cpu_fast_syscall_disable(void *arg)
1490 {
1491 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP))
1492 		cpu_sep_disable();
1493 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC))
1494 		cpu_asysc_disable();
1495 }
1496 
1497 /*ARGSUSED*/
1498 void
1499 cpu_fast_syscall_enable(void *arg)
1500 {
1501 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP))
1502 		cpu_sep_enable();
1503 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC))
1504 		cpu_asysc_enable();
1505 }
1506 
1507 static void
1508 cpu_sep_enable(void)
1509 {
1510 	ASSERT(x86_feature & X86_SEP);
1511 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1512 
1513 	wrmsr(MSR_INTC_SEP_CS, (uint64_t)(uintptr_t)KCS_SEL);
1514 }
1515 
1516 static void
1517 cpu_sep_disable(void)
1518 {
1519 	ASSERT(x86_feature & X86_SEP);
1520 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1521 
1522 	/*
1523 	 * Setting the SYSENTER_CS_MSR register to 0 causes software executing
1524 	 * the sysenter or sysexit instruction to trigger a #gp fault.
1525 	 */
1526 	wrmsr(MSR_INTC_SEP_CS, 0);
1527 }
1528 
1529 static void
1530 cpu_asysc_enable(void)
1531 {
1532 	ASSERT(x86_feature & X86_ASYSC);
1533 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1534 
1535 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) |
1536 	    (uint64_t)(uintptr_t)AMD_EFER_SCE);
1537 }
1538 
1539 static void
1540 cpu_asysc_disable(void)
1541 {
1542 	ASSERT(x86_feature & X86_ASYSC);
1543 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1544 
1545 	/*
1546 	 * Turn off the SCE (syscall enable) bit in the EFER register. Software
1547 	 * executing syscall or sysret with this bit off will incur a #ud trap.
1548 	 */
1549 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) &
1550 	    ~((uint64_t)(uintptr_t)AMD_EFER_SCE));
1551 }
1552