xref: /illumos-gate/usr/src/uts/i86pc/os/mp_startup.c (revision 8b80e8cb6855118d46f605e91b5ed4ce83417395)
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 2008 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 #include <sys/types.h>
28 #include <sys/thread.h>
29 #include <sys/cpuvar.h>
30 #include <sys/t_lock.h>
31 #include <sys/param.h>
32 #include <sys/proc.h>
33 #include <sys/disp.h>
34 #include <sys/class.h>
35 #include <sys/cmn_err.h>
36 #include <sys/debug.h>
37 #include <sys/asm_linkage.h>
38 #include <sys/x_call.h>
39 #include <sys/systm.h>
40 #include <sys/var.h>
41 #include <sys/vtrace.h>
42 #include <vm/hat.h>
43 #include <vm/as.h>
44 #include <vm/seg_kmem.h>
45 #include <vm/seg_kp.h>
46 #include <sys/segments.h>
47 #include <sys/kmem.h>
48 #include <sys/stack.h>
49 #include <sys/smp_impldefs.h>
50 #include <sys/x86_archext.h>
51 #include <sys/machsystm.h>
52 #include <sys/traptrace.h>
53 #include <sys/clock.h>
54 #include <sys/cpc_impl.h>
55 #include <sys/pg.h>
56 #include <sys/cmt.h>
57 #include <sys/dtrace.h>
58 #include <sys/archsystm.h>
59 #include <sys/fp.h>
60 #include <sys/reboot.h>
61 #include <sys/kdi_machimpl.h>
62 #include <vm/hat_i86.h>
63 #include <sys/memnode.h>
64 #include <sys/pci_cfgspace.h>
65 #include <sys/mach_mmu.h>
66 #include <sys/sysmacros.h>
67 #if defined(__xpv)
68 #include <sys/hypervisor.h>
69 #endif
70 #include <sys/cpu_module.h>
71 
72 struct cpu	cpus[1];			/* CPU data */
73 struct cpu	*cpu[NCPU] = {&cpus[0]};	/* pointers to all CPUs */
74 cpu_core_t	cpu_core[NCPU];			/* cpu_core structures */
75 
76 /*
77  * Useful for disabling MP bring-up on a MP capable system.
78  */
79 int use_mp = 1;
80 
81 /*
82  * to be set by a PSM to indicate what cpus
83  * are sitting around on the system.
84  */
85 cpuset_t mp_cpus;
86 
87 /*
88  * This variable is used by the hat layer to decide whether or not
89  * critical sections are needed to prevent race conditions.  For sun4m,
90  * this variable is set once enough MP initialization has been done in
91  * order to allow cross calls.
92  */
93 int flushes_require_xcalls;
94 
95 cpuset_t cpu_ready_set;		/* initialized in startup() */
96 
97 static 	void	mp_startup(void);
98 
99 static void cpu_sep_enable(void);
100 static void cpu_sep_disable(void);
101 static void cpu_asysc_enable(void);
102 static void cpu_asysc_disable(void);
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 	(void) cpuid_getbrandstr(cp, buf, sizeof (buf));
138 	cp->cpu_brandstr = kmem_alloc(strlen(buf) + 1, KM_SLEEP);
139 	(void) strcpy(cp->cpu_brandstr, buf);
140 }
141 
142 /*
143  * Configure syscall support on this CPU.
144  */
145 /*ARGSUSED*/
146 void
147 init_cpu_syscall(struct cpu *cp)
148 {
149 	kpreempt_disable();
150 
151 #if defined(__amd64)
152 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC)) {
153 
154 #if !defined(__lint)
155 		/*
156 		 * The syscall instruction imposes a certain ordering on
157 		 * segment selectors, so we double-check that ordering
158 		 * here.
159 		 */
160 		ASSERT(KDS_SEL == KCS_SEL + 8);
161 		ASSERT(UDS_SEL == U32CS_SEL + 8);
162 		ASSERT(UCS_SEL == U32CS_SEL + 16);
163 #endif
164 		/*
165 		 * Turn syscall/sysret extensions on.
166 		 */
167 		cpu_asysc_enable();
168 
169 		/*
170 		 * Program the magic registers ..
171 		 */
172 		wrmsr(MSR_AMD_STAR,
173 		    ((uint64_t)(U32CS_SEL << 16 | KCS_SEL)) << 32);
174 		wrmsr(MSR_AMD_LSTAR, (uint64_t)(uintptr_t)sys_syscall);
175 		wrmsr(MSR_AMD_CSTAR, (uint64_t)(uintptr_t)sys_syscall32);
176 
177 		/*
178 		 * This list of flags is masked off the incoming
179 		 * %rfl when we enter the kernel.
180 		 */
181 		wrmsr(MSR_AMD_SFMASK, (uint64_t)(uintptr_t)(PS_IE | PS_T));
182 	}
183 #endif
184 
185 	/*
186 	 * On 32-bit kernels, we use sysenter/sysexit because it's too
187 	 * hard to use syscall/sysret, and it is more portable anyway.
188 	 *
189 	 * On 64-bit kernels on Nocona machines, the 32-bit syscall
190 	 * variant isn't available to 32-bit applications, but sysenter is.
191 	 */
192 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP)) {
193 
194 #if !defined(__lint)
195 		/*
196 		 * The sysenter instruction imposes a certain ordering on
197 		 * segment selectors, so we double-check that ordering
198 		 * here. See "sysenter" in Intel document 245471-012, "IA-32
199 		 * Intel Architecture Software Developer's Manual Volume 2:
200 		 * Instruction Set Reference"
201 		 */
202 		ASSERT(KDS_SEL == KCS_SEL + 8);
203 
204 		ASSERT32(UCS_SEL == ((KCS_SEL + 16) | 3));
205 		ASSERT32(UDS_SEL == UCS_SEL + 8);
206 
207 		ASSERT64(U32CS_SEL == ((KCS_SEL + 16) | 3));
208 		ASSERT64(UDS_SEL == U32CS_SEL + 8);
209 #endif
210 
211 		cpu_sep_enable();
212 
213 		/*
214 		 * resume() sets this value to the base of the threads stack
215 		 * via a context handler.
216 		 */
217 		wrmsr(MSR_INTC_SEP_ESP, 0);
218 		wrmsr(MSR_INTC_SEP_EIP, (uint64_t)(uintptr_t)sys_sysenter);
219 	}
220 
221 	kpreempt_enable();
222 }
223 
224 /*
225  * Multiprocessor initialization.
226  *
227  * Allocate and initialize the cpu structure, TRAPTRACE buffer, and the
228  * startup and idle threads for the specified CPU.
229  */
230 struct cpu *
231 mp_startup_init(int cpun)
232 {
233 	struct cpu *cp;
234 	kthread_id_t tp;
235 	caddr_t	sp;
236 	proc_t *procp;
237 #if !defined(__xpv)
238 	extern int idle_cpu_prefer_mwait;
239 #endif
240 	extern void idle();
241 
242 #ifdef TRAPTRACE
243 	trap_trace_ctl_t *ttc = &trap_trace_ctl[cpun];
244 #endif
245 
246 	ASSERT(cpun < NCPU && cpu[cpun] == NULL);
247 
248 	cp = kmem_zalloc(sizeof (*cp), KM_SLEEP);
249 #if !defined(__xpv)
250 	if ((x86_feature & X86_MWAIT) && idle_cpu_prefer_mwait)
251 		cp->cpu_m.mcpu_mwait = cpuid_mwait_alloc(CPU);
252 #endif
253 
254 	procp = curthread->t_procp;
255 
256 	mutex_enter(&cpu_lock);
257 	/*
258 	 * Initialize the dispatcher first.
259 	 */
260 	disp_cpu_init(cp);
261 	mutex_exit(&cpu_lock);
262 
263 	cpu_vm_data_init(cp);
264 
265 	/*
266 	 * Allocate and initialize the startup thread for this CPU.
267 	 * Interrupt and process switch stacks get allocated later
268 	 * when the CPU starts running.
269 	 */
270 	tp = thread_create(NULL, 0, NULL, NULL, 0, procp,
271 	    TS_STOPPED, maxclsyspri);
272 
273 	/*
274 	 * Set state to TS_ONPROC since this thread will start running
275 	 * as soon as the CPU comes online.
276 	 *
277 	 * All the other fields of the thread structure are setup by
278 	 * thread_create().
279 	 */
280 	THREAD_ONPROC(tp, cp);
281 	tp->t_preempt = 1;
282 	tp->t_bound_cpu = cp;
283 	tp->t_affinitycnt = 1;
284 	tp->t_cpu = cp;
285 	tp->t_disp_queue = cp->cpu_disp;
286 
287 	/*
288 	 * Setup thread to start in mp_startup.
289 	 */
290 	sp = tp->t_stk;
291 	tp->t_pc = (uintptr_t)mp_startup;
292 	tp->t_sp = (uintptr_t)(sp - MINFRAME);
293 #if defined(__amd64)
294 	tp->t_sp -= STACK_ENTRY_ALIGN;		/* fake a call */
295 #endif
296 
297 	cp->cpu_id = cpun;
298 	cp->cpu_self = cp;
299 	cp->cpu_thread = tp;
300 	cp->cpu_lwp = NULL;
301 	cp->cpu_dispthread = tp;
302 	cp->cpu_dispatch_pri = DISP_PRIO(tp);
303 
304 	/*
305 	 * cpu_base_spl must be set explicitly here to prevent any blocking
306 	 * operations in mp_startup from causing the spl of the cpu to drop
307 	 * to 0 (allowing device interrupts before we're ready) in resume().
308 	 * cpu_base_spl MUST remain at LOCK_LEVEL until the cpu is CPU_READY.
309 	 * As an extra bit of security on DEBUG kernels, this is enforced with
310 	 * an assertion in mp_startup() -- before cpu_base_spl is set to its
311 	 * proper value.
312 	 */
313 	cp->cpu_base_spl = ipltospl(LOCK_LEVEL);
314 
315 	/*
316 	 * Now, initialize per-CPU idle thread for this CPU.
317 	 */
318 	tp = thread_create(NULL, PAGESIZE, idle, NULL, 0, procp, TS_ONPROC, -1);
319 
320 	cp->cpu_idle_thread = tp;
321 
322 	tp->t_preempt = 1;
323 	tp->t_bound_cpu = cp;
324 	tp->t_affinitycnt = 1;
325 	tp->t_cpu = cp;
326 	tp->t_disp_queue = cp->cpu_disp;
327 
328 	/*
329 	 * Bootstrap the CPU's PG data
330 	 */
331 	pg_cpu_bootstrap(cp);
332 
333 	/*
334 	 * Perform CPC initialization on the new CPU.
335 	 */
336 	kcpc_hw_init(cp);
337 
338 	/*
339 	 * Allocate virtual addresses for cpu_caddr1 and cpu_caddr2
340 	 * for each CPU.
341 	 */
342 	setup_vaddr_for_ppcopy(cp);
343 
344 	/*
345 	 * Allocate page for new GDT and initialize from current GDT.
346 	 */
347 #if !defined(__lint)
348 	ASSERT((sizeof (*cp->cpu_gdt) * NGDT) <= PAGESIZE);
349 #endif
350 	cp->cpu_gdt = kmem_zalloc(PAGESIZE, KM_SLEEP);
351 	bcopy(CPU->cpu_gdt, cp->cpu_gdt, (sizeof (*cp->cpu_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 #if !defined(__lint)
369 		ASSERT((sizeof (*CPU->cpu_idt) * NIDT) <= PAGESIZE);
370 #endif
371 		cp->cpu_idt = kmem_zalloc(PAGESIZE, KM_SLEEP);
372 		bcopy(CPU->cpu_idt, cp->cpu_idt, PAGESIZE);
373 	} else {
374 		cp->cpu_idt = CPU->cpu_idt;
375 	}
376 
377 	/*
378 	 * Get interrupt priority data from cpu 0.
379 	 */
380 	cp->cpu_pri_data = CPU->cpu_pri_data;
381 
382 	/*
383 	 * alloc space for cpuid info
384 	 */
385 	cpuid_alloc_space(cp);
386 
387 	/*
388 	 * alloc space for ucode_info
389 	 */
390 	ucode_alloc_space(cp);
391 
392 	hat_cpu_online(cp);
393 
394 #ifdef TRAPTRACE
395 	/*
396 	 * If this is a TRAPTRACE kernel, allocate TRAPTRACE buffers
397 	 */
398 	ttc->ttc_first = (uintptr_t)kmem_zalloc(trap_trace_bufsize, KM_SLEEP);
399 	ttc->ttc_next = ttc->ttc_first;
400 	ttc->ttc_limit = ttc->ttc_first + trap_trace_bufsize;
401 #endif
402 	/*
403 	 * Record that we have another CPU.
404 	 */
405 	mutex_enter(&cpu_lock);
406 	/*
407 	 * Initialize the interrupt threads for this CPU
408 	 */
409 	cpu_intr_alloc(cp, NINTR_THREADS);
410 	/*
411 	 * Add CPU to list of available CPUs.  It'll be on the active list
412 	 * after mp_startup().
413 	 */
414 	cpu_add_unit(cp);
415 	mutex_exit(&cpu_lock);
416 
417 	return (cp);
418 }
419 
420 /*
421  * Undo what was done in mp_startup_init
422  */
423 static void
424 mp_startup_fini(struct cpu *cp, int error)
425 {
426 	mutex_enter(&cpu_lock);
427 
428 	/*
429 	 * Remove the CPU from the list of available CPUs.
430 	 */
431 	cpu_del_unit(cp->cpu_id);
432 
433 	if (error == ETIMEDOUT) {
434 		/*
435 		 * The cpu was started, but never *seemed* to run any
436 		 * code in the kernel; it's probably off spinning in its
437 		 * own private world, though with potential references to
438 		 * our kmem-allocated IDTs and GDTs (for example).
439 		 *
440 		 * Worse still, it may actually wake up some time later,
441 		 * so rather than guess what it might or might not do, we
442 		 * leave the fundamental data structures intact.
443 		 */
444 		cp->cpu_flags = 0;
445 		mutex_exit(&cpu_lock);
446 		return;
447 	}
448 
449 	/*
450 	 * At this point, the only threads bound to this CPU should
451 	 * special per-cpu threads: it's idle thread, it's pause threads,
452 	 * and it's interrupt threads.  Clean these up.
453 	 */
454 	cpu_destroy_bound_threads(cp);
455 	cp->cpu_idle_thread = NULL;
456 
457 	/*
458 	 * Free the interrupt stack.
459 	 */
460 	segkp_release(segkp,
461 	    cp->cpu_intr_stack - (INTR_STACK_SIZE - SA(MINFRAME)));
462 
463 	mutex_exit(&cpu_lock);
464 
465 #ifdef TRAPTRACE
466 	/*
467 	 * Discard the trap trace buffer
468 	 */
469 	{
470 		trap_trace_ctl_t *ttc = &trap_trace_ctl[cp->cpu_id];
471 
472 		kmem_free((void *)ttc->ttc_first, trap_trace_bufsize);
473 		ttc->ttc_first = NULL;
474 	}
475 #endif
476 
477 	hat_cpu_offline(cp);
478 
479 	cpuid_free_space(cp);
480 
481 	ucode_free_space(cp);
482 
483 	if (cp->cpu_idt != CPU->cpu_idt)
484 		kmem_free(cp->cpu_idt, PAGESIZE);
485 	cp->cpu_idt = NULL;
486 
487 	kmem_free(cp->cpu_gdt, PAGESIZE);
488 	cp->cpu_gdt = NULL;
489 
490 	teardown_vaddr_for_ppcopy(cp);
491 
492 	kcpc_hw_fini(cp);
493 
494 	cp->cpu_dispthread = NULL;
495 	cp->cpu_thread = NULL;	/* discarded by cpu_destroy_bound_threads() */
496 
497 	cpu_vm_data_destroy(cp);
498 
499 	mutex_enter(&cpu_lock);
500 	disp_cpu_fini(cp);
501 	mutex_exit(&cpu_lock);
502 
503 #if !defined(__xpv)
504 	if (cp->cpu_m.mcpu_mwait != NULL)
505 		cpuid_mwait_free(cp);
506 #endif
507 	kmem_free(cp, sizeof (*cp));
508 }
509 
510 /*
511  * Apply workarounds for known errata, and warn about those that are absent.
512  *
513  * System vendors occasionally create configurations which contain different
514  * revisions of the CPUs that are almost but not exactly the same.  At the
515  * time of writing, this meant that their clock rates were the same, their
516  * feature sets were the same, but the required workaround were -not-
517  * necessarily the same.  So, this routine is invoked on -every- CPU soon
518  * after starting to make sure that the resulting system contains the most
519  * pessimal set of workarounds needed to cope with *any* of the CPUs in the
520  * system.
521  *
522  * workaround_errata is invoked early in mlsetup() for CPU 0, and in
523  * mp_startup() for all slave CPUs. Slaves process workaround_errata prior
524  * to acknowledging their readiness to the master, so this routine will
525  * never be executed by multiple CPUs in parallel, thus making updates to
526  * global data safe.
527  *
528  * These workarounds are based on Rev 3.57 of the Revision Guide for
529  * AMD Athlon(tm) 64 and AMD Opteron(tm) Processors, August 2005.
530  */
531 
532 #if defined(OPTERON_ERRATUM_88)
533 int opteron_erratum_88;		/* if non-zero -> at least one cpu has it */
534 #endif
535 
536 #if defined(OPTERON_ERRATUM_91)
537 int opteron_erratum_91;		/* if non-zero -> at least one cpu has it */
538 #endif
539 
540 #if defined(OPTERON_ERRATUM_93)
541 int opteron_erratum_93;		/* if non-zero -> at least one cpu has it */
542 #endif
543 
544 #if defined(OPTERON_ERRATUM_95)
545 int opteron_erratum_95;		/* if non-zero -> at least one cpu has it */
546 #endif
547 
548 #if defined(OPTERON_ERRATUM_100)
549 int opteron_erratum_100;	/* if non-zero -> at least one cpu has it */
550 #endif
551 
552 #if defined(OPTERON_ERRATUM_108)
553 int opteron_erratum_108;	/* if non-zero -> at least one cpu has it */
554 #endif
555 
556 #if defined(OPTERON_ERRATUM_109)
557 int opteron_erratum_109;	/* if non-zero -> at least one cpu has it */
558 #endif
559 
560 #if defined(OPTERON_ERRATUM_121)
561 int opteron_erratum_121;	/* if non-zero -> at least one cpu has it */
562 #endif
563 
564 #if defined(OPTERON_ERRATUM_122)
565 int opteron_erratum_122;	/* if non-zero -> at least one cpu has it */
566 #endif
567 
568 #if defined(OPTERON_ERRATUM_123)
569 int opteron_erratum_123;	/* if non-zero -> at least one cpu has it */
570 #endif
571 
572 #if defined(OPTERON_ERRATUM_131)
573 int opteron_erratum_131;	/* if non-zero -> at least one cpu has it */
574 #endif
575 
576 #if defined(OPTERON_WORKAROUND_6336786)
577 int opteron_workaround_6336786;	/* non-zero -> WA relevant and applied */
578 int opteron_workaround_6336786_UP = 0;	/* Not needed for UP */
579 #endif
580 
581 #if defined(OPTERON_WORKAROUND_6323525)
582 int opteron_workaround_6323525;	/* if non-zero -> at least one cpu has it */
583 #endif
584 
585 #if defined(OPTERON_ERRATUM_298)
586 int opteron_erratum_298;
587 #endif
588 
589 static void
590 workaround_warning(cpu_t *cp, uint_t erratum)
591 {
592 	cmn_err(CE_WARN, "cpu%d: no workaround for erratum %u",
593 	    cp->cpu_id, erratum);
594 }
595 
596 static void
597 workaround_applied(uint_t erratum)
598 {
599 	if (erratum > 1000000)
600 		cmn_err(CE_CONT, "?workaround applied for cpu issue #%d\n",
601 		    erratum);
602 	else
603 		cmn_err(CE_CONT, "?workaround applied for cpu erratum #%d\n",
604 		    erratum);
605 }
606 
607 static void
608 msr_warning(cpu_t *cp, const char *rw, uint_t msr, int error)
609 {
610 	cmn_err(CE_WARN, "cpu%d: couldn't %smsr 0x%x, error %d",
611 	    cp->cpu_id, rw, msr, error);
612 }
613 
614 /*
615  * Determine the number of nodes in an Opteron / Greyhound family system.
616  */
617 static uint_t
618 opteron_get_nnodes(void)
619 {
620 	static uint_t nnodes = 0;
621 
622 #ifdef	DEBUG
623 	uint_t family;
624 
625 	family = cpuid_getfamily(CPU);
626 	ASSERT(family == 0xf || family == 0x10);
627 #endif	/* DEBUG */
628 
629 	if (nnodes == 0) {
630 		/*
631 		 * Obtain the number of nodes in the system from
632 		 * bits [6:4] of the Node ID register on node 0.
633 		 *
634 		 * The actual node count is NodeID[6:4] + 1
635 		 *
636 		 * The Node ID register is accessed via function 0,
637 		 * offset 0x60. Node 0 is device 24.
638 		 */
639 		nnodes = ((pci_getl_func(0, 24, 0, 0x60) & 0x70) >> 4) + 1;
640 	}
641 	return (nnodes);
642 }
643 
644 #if defined(__xpv)
645 
646 /*
647  * On dom0, we can determine the number of physical cpus on the machine.
648  * This number is important when figuring out what workarounds are
649  * appropriate, so compute it now.
650  */
651 uint_t
652 xen_get_nphyscpus(void)
653 {
654 	static uint_t nphyscpus = 0;
655 
656 	ASSERT(DOMAIN_IS_INITDOMAIN(xen_info));
657 
658 	if (nphyscpus == 0) {
659 		xen_sysctl_t op;
660 		xen_sysctl_physinfo_t *pi = &op.u.physinfo;
661 
662 		op.cmd = XEN_SYSCTL_physinfo;
663 		op.interface_version = XEN_SYSCTL_INTERFACE_VERSION;
664 		if (HYPERVISOR_sysctl(&op) == 0)
665 			nphyscpus = pi->threads_per_core *
666 			    pi->cores_per_socket * pi->sockets_per_node *
667 			    pi->nr_nodes;
668 	}
669 	return (nphyscpus);
670 }
671 #endif
672 
673 uint_t
674 do_erratum_298(struct cpu *cpu)
675 {
676 	static int	osvwrc = -3;
677 	extern int	osvw_opteron_erratum(cpu_t *, uint_t);
678 
679 	/*
680 	 * L2 Eviction May Occur During Processor Operation To Set
681 	 * Accessed or Dirty Bit.
682 	 */
683 	if (osvwrc == -3) {
684 		osvwrc = osvw_opteron_erratum(cpu, 298);
685 	} else {
686 		/* osvw return codes should be consistent for all cpus */
687 		ASSERT(osvwrc == osvw_opteron_erratum(cpu, 298));
688 	}
689 
690 	switch (osvwrc) {
691 	case 0:		/* erratum is not present: do nothing */
692 		break;
693 	case 1:		/* erratum is present: BIOS workaround applied */
694 		/*
695 		 * check if workaround is actually in place and issue warning
696 		 * if not.
697 		 */
698 		if (((rdmsr(MSR_AMD_HWCR) & AMD_HWCR_TLBCACHEDIS) == 0) ||
699 		    ((rdmsr(MSR_AMD_BU_CFG) & AMD_BU_CFG_E298) == 0)) {
700 #if defined(OPTERON_ERRATUM_298)
701 			opteron_erratum_298++;
702 #else
703 			workaround_warning(cpu, 298);
704 			return (1);
705 #endif
706 		}
707 		break;
708 	case -1:	/* cannot determine via osvw: check cpuid */
709 		if ((cpuid_opteron_erratum(cpu, 298) > 0) &&
710 		    (((rdmsr(MSR_AMD_HWCR) & AMD_HWCR_TLBCACHEDIS) == 0) ||
711 		    ((rdmsr(MSR_AMD_BU_CFG) & AMD_BU_CFG_E298) == 0))) {
712 #if defined(OPTERON_ERRATUM_298)
713 			opteron_erratum_298++;
714 #else
715 			workaround_warning(cpu, 298);
716 			return (1);
717 #endif
718 		}
719 		break;
720 	}
721 	return (0);
722 }
723 
724 uint_t
725 workaround_errata(struct cpu *cpu)
726 {
727 	uint_t missing = 0;
728 
729 	ASSERT(cpu == CPU);
730 
731 	/*LINTED*/
732 	if (cpuid_opteron_erratum(cpu, 88) > 0) {
733 		/*
734 		 * SWAPGS May Fail To Read Correct GS Base
735 		 */
736 #if defined(OPTERON_ERRATUM_88)
737 		/*
738 		 * The workaround is an mfence in the relevant assembler code
739 		 */
740 		opteron_erratum_88++;
741 #else
742 		workaround_warning(cpu, 88);
743 		missing++;
744 #endif
745 	}
746 
747 	if (cpuid_opteron_erratum(cpu, 91) > 0) {
748 		/*
749 		 * Software Prefetches May Report A Page Fault
750 		 */
751 #if defined(OPTERON_ERRATUM_91)
752 		/*
753 		 * fix is in trap.c
754 		 */
755 		opteron_erratum_91++;
756 #else
757 		workaround_warning(cpu, 91);
758 		missing++;
759 #endif
760 	}
761 
762 	if (cpuid_opteron_erratum(cpu, 93) > 0) {
763 		/*
764 		 * RSM Auto-Halt Restart Returns to Incorrect RIP
765 		 */
766 #if defined(OPTERON_ERRATUM_93)
767 		/*
768 		 * fix is in trap.c
769 		 */
770 		opteron_erratum_93++;
771 #else
772 		workaround_warning(cpu, 93);
773 		missing++;
774 #endif
775 	}
776 
777 	/*LINTED*/
778 	if (cpuid_opteron_erratum(cpu, 95) > 0) {
779 		/*
780 		 * RET Instruction May Return to Incorrect EIP
781 		 */
782 #if defined(OPTERON_ERRATUM_95)
783 #if defined(_LP64)
784 		/*
785 		 * Workaround this by ensuring that 32-bit user code and
786 		 * 64-bit kernel code never occupy the same address
787 		 * range mod 4G.
788 		 */
789 		if (_userlimit32 > 0xc0000000ul)
790 			*(uintptr_t *)&_userlimit32 = 0xc0000000ul;
791 
792 		/*LINTED*/
793 		ASSERT((uint32_t)COREHEAP_BASE == 0xc0000000u);
794 		opteron_erratum_95++;
795 #endif	/* _LP64 */
796 #else
797 		workaround_warning(cpu, 95);
798 		missing++;
799 #endif
800 	}
801 
802 	if (cpuid_opteron_erratum(cpu, 100) > 0) {
803 		/*
804 		 * Compatibility Mode Branches Transfer to Illegal Address
805 		 */
806 #if defined(OPTERON_ERRATUM_100)
807 		/*
808 		 * fix is in trap.c
809 		 */
810 		opteron_erratum_100++;
811 #else
812 		workaround_warning(cpu, 100);
813 		missing++;
814 #endif
815 	}
816 
817 	/*LINTED*/
818 	if (cpuid_opteron_erratum(cpu, 108) > 0) {
819 		/*
820 		 * CPUID Instruction May Return Incorrect Model Number In
821 		 * Some Processors
822 		 */
823 #if defined(OPTERON_ERRATUM_108)
824 		/*
825 		 * (Our cpuid-handling code corrects the model number on
826 		 * those processors)
827 		 */
828 #else
829 		workaround_warning(cpu, 108);
830 		missing++;
831 #endif
832 	}
833 
834 	/*LINTED*/
835 	if (cpuid_opteron_erratum(cpu, 109) > 0) do {
836 		/*
837 		 * Certain Reverse REP MOVS May Produce Unpredictable Behaviour
838 		 */
839 #if defined(OPTERON_ERRATUM_109)
840 		/*
841 		 * The "workaround" is to print a warning to upgrade the BIOS
842 		 */
843 		uint64_t value;
844 		const uint_t msr = MSR_AMD_PATCHLEVEL;
845 		int err;
846 
847 		if ((err = checked_rdmsr(msr, &value)) != 0) {
848 			msr_warning(cpu, "rd", msr, err);
849 			workaround_warning(cpu, 109);
850 			missing++;
851 		}
852 		if (value == 0)
853 			opteron_erratum_109++;
854 #else
855 		workaround_warning(cpu, 109);
856 		missing++;
857 #endif
858 	/*CONSTANTCONDITION*/
859 	} while (0);
860 
861 	/*LINTED*/
862 	if (cpuid_opteron_erratum(cpu, 121) > 0) {
863 		/*
864 		 * Sequential Execution Across Non_Canonical Boundary Caused
865 		 * Processor Hang
866 		 */
867 #if defined(OPTERON_ERRATUM_121)
868 #if defined(_LP64)
869 		/*
870 		 * Erratum 121 is only present in long (64 bit) mode.
871 		 * Workaround is to include the page immediately before the
872 		 * va hole to eliminate the possibility of system hangs due to
873 		 * sequential execution across the va hole boundary.
874 		 */
875 		if (opteron_erratum_121)
876 			opteron_erratum_121++;
877 		else {
878 			if (hole_start) {
879 				hole_start -= PAGESIZE;
880 			} else {
881 				/*
882 				 * hole_start not yet initialized by
883 				 * mmu_init. Initialize hole_start
884 				 * with value to be subtracted.
885 				 */
886 				hole_start = PAGESIZE;
887 			}
888 			opteron_erratum_121++;
889 		}
890 #endif	/* _LP64 */
891 #else
892 		workaround_warning(cpu, 121);
893 		missing++;
894 #endif
895 	}
896 
897 	/*LINTED*/
898 	if (cpuid_opteron_erratum(cpu, 122) > 0) do {
899 		/*
900 		 * TLB Flush Filter May Cause Coherency Problem in
901 		 * Multiprocessor Systems
902 		 */
903 #if defined(OPTERON_ERRATUM_122)
904 		uint64_t value;
905 		const uint_t msr = MSR_AMD_HWCR;
906 		int error;
907 
908 		/*
909 		 * Erratum 122 is only present in MP configurations (multi-core
910 		 * or multi-processor).
911 		 */
912 #if defined(__xpv)
913 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
914 			break;
915 		if (!opteron_erratum_122 && xen_get_nphyscpus() == 1)
916 			break;
917 #else
918 		if (!opteron_erratum_122 && opteron_get_nnodes() == 1 &&
919 		    cpuid_get_ncpu_per_chip(cpu) == 1)
920 			break;
921 #endif
922 		/* disable TLB Flush Filter */
923 
924 		if ((error = checked_rdmsr(msr, &value)) != 0) {
925 			msr_warning(cpu, "rd", msr, error);
926 			workaround_warning(cpu, 122);
927 			missing++;
928 		} else {
929 			value |= (uint64_t)AMD_HWCR_FFDIS;
930 			if ((error = checked_wrmsr(msr, value)) != 0) {
931 				msr_warning(cpu, "wr", msr, error);
932 				workaround_warning(cpu, 122);
933 				missing++;
934 			}
935 		}
936 		opteron_erratum_122++;
937 #else
938 		workaround_warning(cpu, 122);
939 		missing++;
940 #endif
941 	/*CONSTANTCONDITION*/
942 	} while (0);
943 
944 	/*LINTED*/
945 	if (cpuid_opteron_erratum(cpu, 123) > 0) do {
946 		/*
947 		 * Bypassed Reads May Cause Data Corruption of System Hang in
948 		 * Dual Core Processors
949 		 */
950 #if defined(OPTERON_ERRATUM_123)
951 		uint64_t value;
952 		const uint_t msr = MSR_AMD_PATCHLEVEL;
953 		int err;
954 
955 		/*
956 		 * Erratum 123 applies only to multi-core cpus.
957 		 */
958 		if (cpuid_get_ncpu_per_chip(cpu) < 2)
959 			break;
960 #if defined(__xpv)
961 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
962 			break;
963 #endif
964 		/*
965 		 * The "workaround" is to print a warning to upgrade the BIOS
966 		 */
967 		if ((err = checked_rdmsr(msr, &value)) != 0) {
968 			msr_warning(cpu, "rd", msr, err);
969 			workaround_warning(cpu, 123);
970 			missing++;
971 		}
972 		if (value == 0)
973 			opteron_erratum_123++;
974 #else
975 		workaround_warning(cpu, 123);
976 		missing++;
977 
978 #endif
979 	/*CONSTANTCONDITION*/
980 	} while (0);
981 
982 	/*LINTED*/
983 	if (cpuid_opteron_erratum(cpu, 131) > 0) do {
984 		/*
985 		 * Multiprocessor Systems with Four or More Cores May Deadlock
986 		 * Waiting for a Probe Response
987 		 */
988 #if defined(OPTERON_ERRATUM_131)
989 		uint64_t nbcfg;
990 		const uint_t msr = MSR_AMD_NB_CFG;
991 		const uint64_t wabits =
992 		    AMD_NB_CFG_SRQ_HEARTBEAT | AMD_NB_CFG_SRQ_SPR;
993 		int error;
994 
995 		/*
996 		 * Erratum 131 applies to any system with four or more cores.
997 		 */
998 		if (opteron_erratum_131)
999 			break;
1000 #if defined(__xpv)
1001 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
1002 			break;
1003 		if (xen_get_nphyscpus() < 4)
1004 			break;
1005 #else
1006 		if (opteron_get_nnodes() * cpuid_get_ncpu_per_chip(cpu) < 4)
1007 			break;
1008 #endif
1009 		/*
1010 		 * Print a warning if neither of the workarounds for
1011 		 * erratum 131 is present.
1012 		 */
1013 		if ((error = checked_rdmsr(msr, &nbcfg)) != 0) {
1014 			msr_warning(cpu, "rd", msr, error);
1015 			workaround_warning(cpu, 131);
1016 			missing++;
1017 		} else if ((nbcfg & wabits) == 0) {
1018 			opteron_erratum_131++;
1019 		} else {
1020 			/* cannot have both workarounds set */
1021 			ASSERT((nbcfg & wabits) != wabits);
1022 		}
1023 #else
1024 		workaround_warning(cpu, 131);
1025 		missing++;
1026 #endif
1027 	/*CONSTANTCONDITION*/
1028 	} while (0);
1029 
1030 	/*
1031 	 * This isn't really an erratum, but for convenience the
1032 	 * detection/workaround code lives here and in cpuid_opteron_erratum.
1033 	 */
1034 	if (cpuid_opteron_erratum(cpu, 6336786) > 0) {
1035 #if defined(OPTERON_WORKAROUND_6336786)
1036 		/*
1037 		 * Disable C1-Clock ramping on multi-core/multi-processor
1038 		 * K8 platforms to guard against TSC drift.
1039 		 */
1040 		if (opteron_workaround_6336786) {
1041 			opteron_workaround_6336786++;
1042 #if defined(__xpv)
1043 		} else if ((DOMAIN_IS_INITDOMAIN(xen_info) &&
1044 		    xen_get_nphyscpus() > 1) ||
1045 		    opteron_workaround_6336786_UP) {
1046 			/*
1047 			 * XXPV	Hmm.  We can't walk the Northbridges on
1048 			 *	the hypervisor; so just complain and drive
1049 			 *	on.  This probably needs to be fixed in
1050 			 *	the hypervisor itself.
1051 			 */
1052 			opteron_workaround_6336786++;
1053 			workaround_warning(cpu, 6336786);
1054 #else	/* __xpv */
1055 		} else if ((opteron_get_nnodes() *
1056 		    cpuid_get_ncpu_per_chip(cpu) > 1) ||
1057 		    opteron_workaround_6336786_UP) {
1058 
1059 			uint_t	node, nnodes;
1060 			uint8_t data;
1061 
1062 			nnodes = opteron_get_nnodes();
1063 			for (node = 0; node < nnodes; node++) {
1064 				/*
1065 				 * Clear PMM7[1:0] (function 3, offset 0x87)
1066 				 * Northbridge device is the node id + 24.
1067 				 */
1068 				data = pci_getb_func(0, node + 24, 3, 0x87);
1069 				data &= 0xFC;
1070 				pci_putb_func(0, node + 24, 3, 0x87, data);
1071 			}
1072 			opteron_workaround_6336786++;
1073 #endif	/* __xpv */
1074 		}
1075 #else
1076 		workaround_warning(cpu, 6336786);
1077 		missing++;
1078 #endif
1079 	}
1080 
1081 	/*LINTED*/
1082 	/*
1083 	 * Mutex primitives don't work as expected.
1084 	 */
1085 	if (cpuid_opteron_erratum(cpu, 6323525) > 0) {
1086 #if defined(OPTERON_WORKAROUND_6323525)
1087 		/*
1088 		 * This problem only occurs with 2 or more cores. If bit in
1089 		 * MSR_AMD_BU_CFG set, then not applicable. The workaround
1090 		 * is to patch the semaphone routines with the lfence
1091 		 * instruction to provide necessary load memory barrier with
1092 		 * possible subsequent read-modify-write ops.
1093 		 *
1094 		 * It is too early in boot to call the patch routine so
1095 		 * set erratum variable to be done in startup_end().
1096 		 */
1097 		if (opteron_workaround_6323525) {
1098 			opteron_workaround_6323525++;
1099 #if defined(__xpv)
1100 		} else if (x86_feature & X86_SSE2) {
1101 			if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1102 				/*
1103 				 * XXPV	Use dom0_msr here when extended
1104 				 *	operations are supported?
1105 				 */
1106 				if (xen_get_nphyscpus() > 1)
1107 					opteron_workaround_6323525++;
1108 			} else {
1109 				/*
1110 				 * We have no way to tell how many physical
1111 				 * cpus there are, or even if this processor
1112 				 * has the problem, so enable the workaround
1113 				 * unconditionally (at some performance cost).
1114 				 */
1115 				opteron_workaround_6323525++;
1116 			}
1117 #else	/* __xpv */
1118 		} else if ((x86_feature & X86_SSE2) && ((opteron_get_nnodes() *
1119 		    cpuid_get_ncpu_per_chip(cpu)) > 1)) {
1120 			if ((xrdmsr(MSR_AMD_BU_CFG) & 0x02) == 0)
1121 				opteron_workaround_6323525++;
1122 #endif	/* __xpv */
1123 		}
1124 #else
1125 		workaround_warning(cpu, 6323525);
1126 		missing++;
1127 #endif
1128 	}
1129 
1130 	missing += do_erratum_298(cpu);
1131 
1132 #ifdef __xpv
1133 	return (0);
1134 #else
1135 	return (missing);
1136 #endif
1137 }
1138 
1139 void
1140 workaround_errata_end()
1141 {
1142 #if defined(OPTERON_ERRATUM_88)
1143 	if (opteron_erratum_88)
1144 		workaround_applied(88);
1145 #endif
1146 #if defined(OPTERON_ERRATUM_91)
1147 	if (opteron_erratum_91)
1148 		workaround_applied(91);
1149 #endif
1150 #if defined(OPTERON_ERRATUM_93)
1151 	if (opteron_erratum_93)
1152 		workaround_applied(93);
1153 #endif
1154 #if defined(OPTERON_ERRATUM_95)
1155 	if (opteron_erratum_95)
1156 		workaround_applied(95);
1157 #endif
1158 #if defined(OPTERON_ERRATUM_100)
1159 	if (opteron_erratum_100)
1160 		workaround_applied(100);
1161 #endif
1162 #if defined(OPTERON_ERRATUM_108)
1163 	if (opteron_erratum_108)
1164 		workaround_applied(108);
1165 #endif
1166 #if defined(OPTERON_ERRATUM_109)
1167 	if (opteron_erratum_109) {
1168 		cmn_err(CE_WARN,
1169 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1170 		    " processor\nerratum 109 was not detected; updating your"
1171 		    " system's BIOS to a version\ncontaining this"
1172 		    " microcode patch is HIGHLY recommended or erroneous"
1173 		    " system\noperation may occur.\n");
1174 	}
1175 #endif
1176 #if defined(OPTERON_ERRATUM_121)
1177 	if (opteron_erratum_121)
1178 		workaround_applied(121);
1179 #endif
1180 #if defined(OPTERON_ERRATUM_122)
1181 	if (opteron_erratum_122)
1182 		workaround_applied(122);
1183 #endif
1184 #if defined(OPTERON_ERRATUM_123)
1185 	if (opteron_erratum_123) {
1186 		cmn_err(CE_WARN,
1187 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1188 		    " processor\nerratum 123 was not detected; updating your"
1189 		    " system's BIOS to a version\ncontaining this"
1190 		    " microcode patch is HIGHLY recommended or erroneous"
1191 		    " system\noperation may occur.\n");
1192 	}
1193 #endif
1194 #if defined(OPTERON_ERRATUM_131)
1195 	if (opteron_erratum_131) {
1196 		cmn_err(CE_WARN,
1197 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1198 		    " processor\nerratum 131 was not detected; updating your"
1199 		    " system's BIOS to a version\ncontaining this"
1200 		    " microcode patch is HIGHLY recommended or erroneous"
1201 		    " system\noperation may occur.\n");
1202 	}
1203 #endif
1204 #if defined(OPTERON_WORKAROUND_6336786)
1205 	if (opteron_workaround_6336786)
1206 		workaround_applied(6336786);
1207 #endif
1208 #if defined(OPTERON_WORKAROUND_6323525)
1209 	if (opteron_workaround_6323525)
1210 		workaround_applied(6323525);
1211 #endif
1212 #if defined(OPTERON_ERRATUM_298)
1213 	if (opteron_erratum_298) {
1214 		cmn_err(CE_WARN,
1215 		    "BIOS microcode patch for AMD 64/Opteron(tm)"
1216 		    " processor\nerratum 298 was not detected; updating your"
1217 		    " system's BIOS to a version\ncontaining this"
1218 		    " microcode patch is HIGHLY recommended or erroneous"
1219 		    " system\noperation may occur.\n");
1220 	}
1221 #endif
1222 }
1223 
1224 static cpuset_t procset;
1225 
1226 /*
1227  * Start a single cpu, assuming that the kernel context is available
1228  * to successfully start another cpu.
1229  *
1230  * (For example, real mode code is mapped into the right place
1231  * in memory and is ready to be run.)
1232  */
1233 int
1234 start_cpu(processorid_t who)
1235 {
1236 	void *ctx;
1237 	cpu_t *cp;
1238 	int delays;
1239 	int error = 0;
1240 
1241 	ASSERT(who != 0);
1242 
1243 	/*
1244 	 * Check if there's at least a Mbyte of kmem available
1245 	 * before attempting to start the cpu.
1246 	 */
1247 	if (kmem_avail() < 1024 * 1024) {
1248 		/*
1249 		 * Kick off a reap in case that helps us with
1250 		 * later attempts ..
1251 		 */
1252 		kmem_reap();
1253 		return (ENOMEM);
1254 	}
1255 
1256 	cp = mp_startup_init(who);
1257 	if ((ctx = mach_cpucontext_alloc(cp)) == NULL ||
1258 	    (error = mach_cpu_start(cp, ctx)) != 0) {
1259 
1260 		/*
1261 		 * Something went wrong before we even started it
1262 		 */
1263 		if (ctx)
1264 			cmn_err(CE_WARN,
1265 			    "cpu%d: failed to start error %d",
1266 			    cp->cpu_id, error);
1267 		else
1268 			cmn_err(CE_WARN,
1269 			    "cpu%d: failed to allocate context", cp->cpu_id);
1270 
1271 		if (ctx)
1272 			mach_cpucontext_free(cp, ctx, error);
1273 		else
1274 			error = EAGAIN;		/* hmm. */
1275 		mp_startup_fini(cp, error);
1276 		return (error);
1277 	}
1278 
1279 	for (delays = 0; !CPU_IN_SET(procset, who); delays++) {
1280 		if (delays == 500) {
1281 			/*
1282 			 * After five seconds, things are probably looking
1283 			 * a bit bleak - explain the hang.
1284 			 */
1285 			cmn_err(CE_NOTE, "cpu%d: started, "
1286 			    "but not running in the kernel yet", who);
1287 		} else if (delays > 2000) {
1288 			/*
1289 			 * We waited at least 20 seconds, bail ..
1290 			 */
1291 			error = ETIMEDOUT;
1292 			cmn_err(CE_WARN, "cpu%d: timed out", who);
1293 			mach_cpucontext_free(cp, ctx, error);
1294 			mp_startup_fini(cp, error);
1295 			return (error);
1296 		}
1297 
1298 		/*
1299 		 * wait at least 10ms, then check again..
1300 		 */
1301 		delay(USEC_TO_TICK_ROUNDUP(10000));
1302 	}
1303 
1304 	mach_cpucontext_free(cp, ctx, 0);
1305 
1306 #ifndef __xpv
1307 	if (tsc_gethrtime_enable)
1308 		tsc_sync_master(who);
1309 #endif
1310 
1311 	if (dtrace_cpu_init != NULL) {
1312 		/*
1313 		 * DTrace CPU initialization expects cpu_lock to be held.
1314 		 */
1315 		mutex_enter(&cpu_lock);
1316 		(*dtrace_cpu_init)(who);
1317 		mutex_exit(&cpu_lock);
1318 	}
1319 
1320 	while (!CPU_IN_SET(cpu_ready_set, who))
1321 		delay(1);
1322 
1323 	return (0);
1324 }
1325 
1326 
1327 /*ARGSUSED*/
1328 void
1329 start_other_cpus(int cprboot)
1330 {
1331 	uint_t who;
1332 	uint_t skipped = 0;
1333 	uint_t bootcpuid = 0;
1334 
1335 	/*
1336 	 * Initialize our own cpu_info.
1337 	 */
1338 	init_cpu_info(CPU);
1339 
1340 	cmn_err(CE_CONT, "?cpu%d: %s\n", CPU->cpu_id, CPU->cpu_idstr);
1341 	cmn_err(CE_CONT, "?cpu%d: %s\n", CPU->cpu_id, CPU->cpu_brandstr);
1342 
1343 	/*
1344 	 * Initialize our syscall handlers
1345 	 */
1346 	init_cpu_syscall(CPU);
1347 
1348 	/*
1349 	 * Take the boot cpu out of the mp_cpus set because we know
1350 	 * it's already running.  Add it to the cpu_ready_set for
1351 	 * precisely the same reason.
1352 	 */
1353 	CPUSET_DEL(mp_cpus, bootcpuid);
1354 	CPUSET_ADD(cpu_ready_set, bootcpuid);
1355 
1356 	/*
1357 	 * if only 1 cpu or not using MP, skip the rest of this
1358 	 */
1359 	if (CPUSET_ISNULL(mp_cpus) || use_mp == 0) {
1360 		if (use_mp == 0)
1361 			cmn_err(CE_CONT, "?***** Not in MP mode\n");
1362 		goto done;
1363 	}
1364 
1365 	/*
1366 	 * perform such initialization as is needed
1367 	 * to be able to take CPUs on- and off-line.
1368 	 */
1369 	cpu_pause_init();
1370 
1371 	xc_init();		/* initialize processor crosscalls */
1372 
1373 	if (mach_cpucontext_init() != 0)
1374 		goto done;
1375 
1376 	flushes_require_xcalls = 1;
1377 
1378 	/*
1379 	 * We lock our affinity to the master CPU to ensure that all slave CPUs
1380 	 * do their TSC syncs with the same CPU.
1381 	 */
1382 	affinity_set(CPU_CURRENT);
1383 
1384 	for (who = 0; who < NCPU; who++) {
1385 
1386 		if (!CPU_IN_SET(mp_cpus, who))
1387 			continue;
1388 		ASSERT(who != bootcpuid);
1389 		if (ncpus >= max_ncpus) {
1390 			skipped = who;
1391 			continue;
1392 		}
1393 		if (start_cpu(who) != 0)
1394 			CPUSET_DEL(mp_cpus, who);
1395 	}
1396 
1397 	/* Free the space allocated to hold the microcode file */
1398 	ucode_cleanup();
1399 
1400 	affinity_clear();
1401 
1402 	if (skipped) {
1403 		cmn_err(CE_NOTE,
1404 		    "System detected %d cpus, but "
1405 		    "only %d cpu(s) were enabled during boot.",
1406 		    skipped + 1, ncpus);
1407 		cmn_err(CE_NOTE,
1408 		    "Use \"boot-ncpus\" parameter to enable more CPU(s). "
1409 		    "See eeprom(1M).");
1410 	}
1411 
1412 done:
1413 	workaround_errata_end();
1414 	mach_cpucontext_fini();
1415 
1416 	cmi_post_mpstartup();
1417 }
1418 
1419 /*
1420  * Dummy functions - no i86pc platforms support dynamic cpu allocation.
1421  */
1422 /*ARGSUSED*/
1423 int
1424 mp_cpu_configure(int cpuid)
1425 {
1426 	return (ENOTSUP);		/* not supported */
1427 }
1428 
1429 /*ARGSUSED*/
1430 int
1431 mp_cpu_unconfigure(int cpuid)
1432 {
1433 	return (ENOTSUP);		/* not supported */
1434 }
1435 
1436 /*
1437  * Startup function for 'other' CPUs (besides boot cpu).
1438  * Called from real_mode_start.
1439  *
1440  * WARNING: until CPU_READY is set, mp_startup and routines called by
1441  * mp_startup should not call routines (e.g. kmem_free) that could call
1442  * hat_unload which requires CPU_READY to be set.
1443  */
1444 void
1445 mp_startup(void)
1446 {
1447 	struct cpu *cp = CPU;
1448 	uint_t new_x86_feature;
1449 
1450 	/*
1451 	 * We need to get TSC on this proc synced (i.e., any delta
1452 	 * from cpu0 accounted for) as soon as we can, because many
1453 	 * many things use gethrtime/pc_gethrestime, including
1454 	 * interrupts, cmn_err, etc.
1455 	 */
1456 
1457 	/* Let cpu0 continue into tsc_sync_master() */
1458 	CPUSET_ATOMIC_ADD(procset, cp->cpu_id);
1459 
1460 #ifndef __xpv
1461 	if (tsc_gethrtime_enable)
1462 		tsc_sync_slave();
1463 #endif
1464 
1465 	/*
1466 	 * Once this was done from assembly, but it's safer here; if
1467 	 * it blocks, we need to be able to swtch() to and from, and
1468 	 * since we get here by calling t_pc, we need to do that call
1469 	 * before swtch() overwrites it.
1470 	 */
1471 	(void) (*ap_mlsetup)();
1472 
1473 	new_x86_feature = cpuid_pass1(cp);
1474 
1475 #ifndef __xpv
1476 	/*
1477 	 * Program this cpu's PAT
1478 	 */
1479 	if (x86_feature & X86_PAT)
1480 		pat_sync();
1481 #endif
1482 
1483 	/*
1484 	 * Set up TSC_AUX to contain the cpuid for this processor
1485 	 * for the rdtscp instruction.
1486 	 */
1487 	if (x86_feature & X86_TSCP)
1488 		(void) wrmsr(MSR_AMD_TSCAUX, cp->cpu_id);
1489 
1490 	/*
1491 	 * Initialize this CPU's syscall handlers
1492 	 */
1493 	init_cpu_syscall(cp);
1494 
1495 	/*
1496 	 * Enable interrupts with spl set to LOCK_LEVEL. LOCK_LEVEL is the
1497 	 * highest level at which a routine is permitted to block on
1498 	 * an adaptive mutex (allows for cpu poke interrupt in case
1499 	 * the cpu is blocked on a mutex and halts). Setting LOCK_LEVEL blocks
1500 	 * device interrupts that may end up in the hat layer issuing cross
1501 	 * calls before CPU_READY is set.
1502 	 */
1503 	splx(ipltospl(LOCK_LEVEL));
1504 	sti();
1505 
1506 	/*
1507 	 * Do a sanity check to make sure this new CPU is a sane thing
1508 	 * to add to the collection of processors running this system.
1509 	 *
1510 	 * XXX	Clearly this needs to get more sophisticated, if x86
1511 	 * systems start to get built out of heterogenous CPUs; as is
1512 	 * likely to happen once the number of processors in a configuration
1513 	 * gets large enough.
1514 	 */
1515 	if ((x86_feature & new_x86_feature) != x86_feature) {
1516 		cmn_err(CE_CONT, "?cpu%d: %b\n",
1517 		    cp->cpu_id, new_x86_feature, FMT_X86_FEATURE);
1518 		cmn_err(CE_WARN, "cpu%d feature mismatch", cp->cpu_id);
1519 	}
1520 
1521 	/*
1522 	 * We do not support cpus with mixed monitor/mwait support if the
1523 	 * boot cpu supports monitor/mwait.
1524 	 */
1525 	if ((x86_feature & ~new_x86_feature) & X86_MWAIT)
1526 		panic("unsupported mixed cpu monitor/mwait support detected");
1527 
1528 	/*
1529 	 * We could be more sophisticated here, and just mark the CPU
1530 	 * as "faulted" but at this point we'll opt for the easier
1531 	 * answer of dieing horribly.  Provided the boot cpu is ok,
1532 	 * the system can be recovered by booting with use_mp set to zero.
1533 	 */
1534 	if (workaround_errata(cp) != 0)
1535 		panic("critical workaround(s) missing for cpu%d", cp->cpu_id);
1536 
1537 	cpuid_pass2(cp);
1538 	cpuid_pass3(cp);
1539 	(void) cpuid_pass4(cp);
1540 
1541 	init_cpu_info(cp);
1542 
1543 	mutex_enter(&cpu_lock);
1544 	/*
1545 	 * Processor group initialization for this CPU is dependent on the
1546 	 * cpuid probing, which must be done in the context of the current
1547 	 * CPU.
1548 	 */
1549 	pghw_physid_create(cp);
1550 	pg_cpu_init(cp);
1551 	pg_cmt_cpu_startup(cp);
1552 
1553 	cp->cpu_flags |= CPU_RUNNING | CPU_READY | CPU_EXISTS;
1554 
1555 	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_idstr);
1556 	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_brandstr);
1557 
1558 	if (dtrace_cpu_init != NULL) {
1559 		(*dtrace_cpu_init)(cp->cpu_id);
1560 	}
1561 
1562 	/*
1563 	 * Fill out cpu_ucode_info.  Update microcode if necessary.
1564 	 */
1565 	ucode_check(cp);
1566 
1567 	mutex_exit(&cpu_lock);
1568 
1569 	/*
1570 	 * Enable preemption here so that contention for any locks acquired
1571 	 * later in mp_startup may be preempted if the thread owning those
1572 	 * locks is continously executing on other CPUs (for example, this
1573 	 * CPU must be preemptible to allow other CPUs to pause it during their
1574 	 * startup phases).  It's safe to enable preemption here because the
1575 	 * CPU state is pretty-much fully constructed.
1576 	 */
1577 	curthread->t_preempt = 0;
1578 
1579 	/* The base spl should still be at LOCK LEVEL here */
1580 	ASSERT(cp->cpu_base_spl == ipltospl(LOCK_LEVEL));
1581 	set_base_spl();		/* Restore the spl to its proper value */
1582 
1583 	/* Enable interrupts */
1584 	(void) spl0();
1585 	mutex_enter(&cpu_lock);
1586 	cpu_enable_intr(cp);
1587 	cpu_add_active(cp);
1588 	mutex_exit(&cpu_lock);
1589 
1590 	add_cpunode2devtree(cp->cpu_id, cp->cpu_m.mcpu_cpi);
1591 
1592 #ifndef __xpv
1593 	{
1594 		/*
1595 		 * Set up the CPU module for this CPU.  This can't be done
1596 		 * before this CPU is made CPU_READY, because we may (in
1597 		 * heterogeneous systems) need to go load another CPU module.
1598 		 * The act of attempting to load a module may trigger a
1599 		 * cross-call, which will ASSERT unless this cpu is CPU_READY.
1600 		 */
1601 		cmi_hdl_t hdl;
1602 
1603 		if ((hdl = cmi_init(CMI_HDL_NATIVE, cmi_ntv_hwchipid(CPU),
1604 		    cmi_ntv_hwcoreid(CPU), cmi_ntv_hwstrandid(CPU))) != NULL) {
1605 			if (x86_feature & X86_MCA)
1606 				cmi_mca_init(hdl);
1607 		}
1608 	}
1609 #endif /* __xpv */
1610 
1611 	if (boothowto & RB_DEBUG)
1612 		kdi_cpu_init();
1613 
1614 	/*
1615 	 * Setting the bit in cpu_ready_set must be the last operation in
1616 	 * processor initialization; the boot CPU will continue to boot once
1617 	 * it sees this bit set for all active CPUs.
1618 	 */
1619 	CPUSET_ATOMIC_ADD(cpu_ready_set, cp->cpu_id);
1620 
1621 	/*
1622 	 * Because mp_startup() gets fired off after init() starts, we
1623 	 * can't use the '?' trick to do 'boot -v' printing - so we
1624 	 * always direct the 'cpu .. online' messages to the log.
1625 	 */
1626 	cmn_err(CE_CONT, "!cpu%d initialization complete - online\n",
1627 	    cp->cpu_id);
1628 
1629 	/*
1630 	 * Now we are done with the startup thread, so free it up.
1631 	 */
1632 	thread_exit();
1633 	panic("mp_startup: cannot return");
1634 	/*NOTREACHED*/
1635 }
1636 
1637 
1638 /*
1639  * Start CPU on user request.
1640  */
1641 /* ARGSUSED */
1642 int
1643 mp_cpu_start(struct cpu *cp)
1644 {
1645 	ASSERT(MUTEX_HELD(&cpu_lock));
1646 	return (0);
1647 }
1648 
1649 /*
1650  * Stop CPU on user request.
1651  */
1652 /* ARGSUSED */
1653 int
1654 mp_cpu_stop(struct cpu *cp)
1655 {
1656 	extern int cbe_psm_timer_mode;
1657 	ASSERT(MUTEX_HELD(&cpu_lock));
1658 
1659 #ifdef __xpv
1660 	/*
1661 	 * We can't offline vcpu0.
1662 	 */
1663 	if (cp->cpu_id == 0)
1664 		return (EBUSY);
1665 #endif
1666 
1667 	/*
1668 	 * If TIMER_PERIODIC mode is used, CPU0 is the one running it;
1669 	 * can't stop it.  (This is true only for machines with no TSC.)
1670 	 */
1671 
1672 	if ((cbe_psm_timer_mode == TIMER_PERIODIC) && (cp->cpu_id == 0))
1673 		return (EBUSY);
1674 
1675 	return (0);
1676 }
1677 
1678 /*
1679  * Take the specified CPU out of participation in interrupts.
1680  */
1681 int
1682 cpu_disable_intr(struct cpu *cp)
1683 {
1684 	if (psm_disable_intr(cp->cpu_id) != DDI_SUCCESS)
1685 		return (EBUSY);
1686 
1687 	cp->cpu_flags &= ~CPU_ENABLE;
1688 	return (0);
1689 }
1690 
1691 /*
1692  * Allow the specified CPU to participate in interrupts.
1693  */
1694 void
1695 cpu_enable_intr(struct cpu *cp)
1696 {
1697 	ASSERT(MUTEX_HELD(&cpu_lock));
1698 	cp->cpu_flags |= CPU_ENABLE;
1699 	psm_enable_intr(cp->cpu_id);
1700 }
1701 
1702 
1703 /*ARGSUSED*/
1704 void
1705 mp_cpu_faulted_enter(struct cpu *cp)
1706 {
1707 #ifndef __xpv
1708 	cmi_hdl_t hdl = cmi_hdl_lookup(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp),
1709 	    cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp));
1710 
1711 	if (hdl != NULL) {
1712 		cmi_faulted_enter(hdl);
1713 		cmi_hdl_rele(hdl);
1714 	}
1715 #endif
1716 }
1717 
1718 /*ARGSUSED*/
1719 void
1720 mp_cpu_faulted_exit(struct cpu *cp)
1721 {
1722 #ifndef __xpv
1723 	cmi_hdl_t hdl = cmi_hdl_lookup(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp),
1724 	    cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp));
1725 
1726 	if (hdl != NULL) {
1727 		cmi_faulted_exit(hdl);
1728 		cmi_hdl_rele(hdl);
1729 	}
1730 #endif
1731 }
1732 
1733 /*
1734  * The following two routines are used as context operators on threads belonging
1735  * to processes with a private LDT (see sysi86).  Due to the rarity of such
1736  * processes, these routines are currently written for best code readability and
1737  * organization rather than speed.  We could avoid checking x86_feature at every
1738  * context switch by installing different context ops, depending on the
1739  * x86_feature flags, at LDT creation time -- one for each combination of fast
1740  * syscall feature flags.
1741  */
1742 
1743 /*ARGSUSED*/
1744 void
1745 cpu_fast_syscall_disable(void *arg)
1746 {
1747 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP))
1748 		cpu_sep_disable();
1749 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC))
1750 		cpu_asysc_disable();
1751 }
1752 
1753 /*ARGSUSED*/
1754 void
1755 cpu_fast_syscall_enable(void *arg)
1756 {
1757 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP))
1758 		cpu_sep_enable();
1759 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC))
1760 		cpu_asysc_enable();
1761 }
1762 
1763 static void
1764 cpu_sep_enable(void)
1765 {
1766 	ASSERT(x86_feature & X86_SEP);
1767 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1768 
1769 	wrmsr(MSR_INTC_SEP_CS, (uint64_t)(uintptr_t)KCS_SEL);
1770 }
1771 
1772 static void
1773 cpu_sep_disable(void)
1774 {
1775 	ASSERT(x86_feature & X86_SEP);
1776 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1777 
1778 	/*
1779 	 * Setting the SYSENTER_CS_MSR register to 0 causes software executing
1780 	 * the sysenter or sysexit instruction to trigger a #gp fault.
1781 	 */
1782 	wrmsr(MSR_INTC_SEP_CS, 0);
1783 }
1784 
1785 static void
1786 cpu_asysc_enable(void)
1787 {
1788 	ASSERT(x86_feature & X86_ASYSC);
1789 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1790 
1791 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) |
1792 	    (uint64_t)(uintptr_t)AMD_EFER_SCE);
1793 }
1794 
1795 static void
1796 cpu_asysc_disable(void)
1797 {
1798 	ASSERT(x86_feature & X86_ASYSC);
1799 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1800 
1801 	/*
1802 	 * Turn off the SCE (syscall enable) bit in the EFER register. Software
1803 	 * executing syscall or sysret with this bit off will incur a #ud trap.
1804 	 */
1805 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) &
1806 	    ~((uint64_t)(uintptr_t)AMD_EFER_SCE));
1807 }
1808