xref: /illumos-gate/usr/src/uts/i86pc/os/mp_startup.c (revision ddf7fe95b8ad67aa16deb427a0b78f4dd4ff22b1)
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 #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 #if defined(__xpv)
70 #include <sys/hypervisor.h>
71 #endif
72 #include <sys/cpu_module.h>
73 
74 struct cpu	cpus[1];			/* CPU data */
75 struct cpu	*cpu[NCPU] = {&cpus[0]};	/* pointers to all CPUs */
76 cpu_core_t	cpu_core[NCPU];			/* cpu_core structures */
77 
78 /*
79  * Useful for disabling MP bring-up on a MP capable system.
80  */
81 int use_mp = 1;
82 
83 /*
84  * to be set by a PSM to indicate what cpus
85  * are sitting around on the system.
86  */
87 cpuset_t mp_cpus;
88 
89 /*
90  * This variable is used by the hat layer to decide whether or not
91  * critical sections are needed to prevent race conditions.  For sun4m,
92  * this variable is set once enough MP initialization has been done in
93  * order to allow cross calls.
94  */
95 int flushes_require_xcalls;
96 cpuset_t cpu_ready_set = 1;
97 
98 static 	void	mp_startup(void);
99 
100 static void cpu_sep_enable(void);
101 static void cpu_sep_disable(void);
102 static void cpu_asysc_enable(void);
103 static void cpu_asysc_disable(void);
104 
105 /*
106  * Init CPU info - get CPU type info for processor_info system call.
107  */
108 void
109 init_cpu_info(struct cpu *cp)
110 {
111 	processor_info_t *pi = &cp->cpu_type_info;
112 	char buf[CPU_IDSTRLEN];
113 
114 	/*
115 	 * Get clock-frequency property for the CPU.
116 	 */
117 	pi->pi_clock = cpu_freq;
118 
119 	/*
120 	 * Current frequency in Hz.
121 	 */
122 	cp->cpu_curr_clock = cpu_freq_hz;
123 
124 	/*
125 	 * Supported frequencies.
126 	 */
127 	cpu_set_supp_freqs(cp, NULL);
128 
129 	(void) strcpy(pi->pi_processor_type, "i386");
130 	if (fpu_exists)
131 		(void) strcpy(pi->pi_fputypes, "i387 compatible");
132 
133 	(void) cpuid_getidstr(cp, buf, sizeof (buf));
134 
135 	cp->cpu_idstr = kmem_alloc(strlen(buf) + 1, KM_SLEEP);
136 	(void) strcpy(cp->cpu_idstr, buf);
137 
138 	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_idstr);
139 
140 	(void) cpuid_getbrandstr(cp, buf, sizeof (buf));
141 	cp->cpu_brandstr = kmem_alloc(strlen(buf) + 1, KM_SLEEP);
142 	(void) strcpy(cp->cpu_brandstr, buf);
143 
144 	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_brandstr);
145 }
146 
147 /*
148  * Configure syscall support on this CPU.
149  */
150 /*ARGSUSED*/
151 void
152 init_cpu_syscall(struct cpu *cp)
153 {
154 	kpreempt_disable();
155 
156 #if defined(__amd64)
157 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC)) {
158 
159 #if !defined(__lint)
160 		/*
161 		 * The syscall instruction imposes a certain ordering on
162 		 * segment selectors, so we double-check that ordering
163 		 * here.
164 		 */
165 		ASSERT(KDS_SEL == KCS_SEL + 8);
166 		ASSERT(UDS_SEL == U32CS_SEL + 8);
167 		ASSERT(UCS_SEL == U32CS_SEL + 16);
168 #endif
169 		/*
170 		 * Turn syscall/sysret extensions on.
171 		 */
172 		cpu_asysc_enable();
173 
174 		/*
175 		 * Program the magic registers ..
176 		 */
177 		wrmsr(MSR_AMD_STAR,
178 		    ((uint64_t)(U32CS_SEL << 16 | KCS_SEL)) << 32);
179 		wrmsr(MSR_AMD_LSTAR, (uint64_t)(uintptr_t)sys_syscall);
180 		wrmsr(MSR_AMD_CSTAR, (uint64_t)(uintptr_t)sys_syscall32);
181 
182 		/*
183 		 * This list of flags is masked off the incoming
184 		 * %rfl when we enter the kernel.
185 		 */
186 		wrmsr(MSR_AMD_SFMASK, (uint64_t)(uintptr_t)(PS_IE | PS_T));
187 	}
188 #endif
189 
190 	/*
191 	 * On 32-bit kernels, we use sysenter/sysexit because it's too
192 	 * hard to use syscall/sysret, and it is more portable anyway.
193 	 *
194 	 * On 64-bit kernels on Nocona machines, the 32-bit syscall
195 	 * variant isn't available to 32-bit applications, but sysenter is.
196 	 */
197 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP)) {
198 
199 #if !defined(__lint)
200 		/*
201 		 * The sysenter instruction imposes a certain ordering on
202 		 * segment selectors, so we double-check that ordering
203 		 * here. See "sysenter" in Intel document 245471-012, "IA-32
204 		 * Intel Architecture Software Developer's Manual Volume 2:
205 		 * Instruction Set Reference"
206 		 */
207 		ASSERT(KDS_SEL == KCS_SEL + 8);
208 
209 		ASSERT32(UCS_SEL == ((KCS_SEL + 16) | 3));
210 		ASSERT32(UDS_SEL == UCS_SEL + 8);
211 
212 		ASSERT64(U32CS_SEL == ((KCS_SEL + 16) | 3));
213 		ASSERT64(UDS_SEL == U32CS_SEL + 8);
214 #endif
215 
216 		cpu_sep_enable();
217 
218 		/*
219 		 * resume() sets this value to the base of the threads stack
220 		 * via a context handler.
221 		 */
222 		wrmsr(MSR_INTC_SEP_ESP, 0);
223 		wrmsr(MSR_INTC_SEP_EIP, (uint64_t)(uintptr_t)sys_sysenter);
224 	}
225 
226 	kpreempt_enable();
227 }
228 
229 /*
230  * Multiprocessor initialization.
231  *
232  * Allocate and initialize the cpu structure, TRAPTRACE buffer, and the
233  * startup and idle threads for the specified CPU.
234  */
235 struct cpu *
236 mp_startup_init(int cpun)
237 {
238 	struct cpu *cp;
239 	kthread_id_t tp;
240 	caddr_t	sp;
241 	proc_t *procp;
242 #if !defined(__xpv)
243 	extern int idle_cpu_prefer_mwait;
244 #endif
245 	extern void idle();
246 
247 #ifdef TRAPTRACE
248 	trap_trace_ctl_t *ttc = &trap_trace_ctl[cpun];
249 #endif
250 
251 	ASSERT(cpun < NCPU && cpu[cpun] == NULL);
252 
253 	cp = kmem_zalloc(sizeof (*cp), KM_SLEEP);
254 #if !defined(__xpv)
255 	if ((x86_feature & X86_MWAIT) && idle_cpu_prefer_mwait)
256 		cp->cpu_m.mcpu_mwait = cpuid_mwait_alloc(CPU);
257 #endif
258 
259 	procp = curthread->t_procp;
260 
261 	mutex_enter(&cpu_lock);
262 	/*
263 	 * Initialize the dispatcher first.
264 	 */
265 	disp_cpu_init(cp);
266 	mutex_exit(&cpu_lock);
267 
268 	cpu_vm_data_init(cp);
269 
270 	/*
271 	 * Allocate and initialize the startup thread for this CPU.
272 	 * Interrupt and process switch stacks get allocated later
273 	 * when the CPU starts running.
274 	 */
275 	tp = thread_create(NULL, 0, NULL, NULL, 0, procp,
276 	    TS_STOPPED, maxclsyspri);
277 
278 	/*
279 	 * Set state to TS_ONPROC since this thread will start running
280 	 * as soon as the CPU comes online.
281 	 *
282 	 * All the other fields of the thread structure are setup by
283 	 * thread_create().
284 	 */
285 	THREAD_ONPROC(tp, cp);
286 	tp->t_preempt = 1;
287 	tp->t_bound_cpu = cp;
288 	tp->t_affinitycnt = 1;
289 	tp->t_cpu = cp;
290 	tp->t_disp_queue = cp->cpu_disp;
291 
292 	/*
293 	 * Setup thread to start in mp_startup.
294 	 */
295 	sp = tp->t_stk;
296 	tp->t_pc = (uintptr_t)mp_startup;
297 	tp->t_sp = (uintptr_t)(sp - MINFRAME);
298 #if defined(__amd64)
299 	tp->t_sp -= STACK_ENTRY_ALIGN;		/* fake a call */
300 #endif
301 
302 	cp->cpu_id = cpun;
303 	cp->cpu_self = cp;
304 	cp->cpu_thread = tp;
305 	cp->cpu_lwp = NULL;
306 	cp->cpu_dispthread = tp;
307 	cp->cpu_dispatch_pri = DISP_PRIO(tp);
308 
309 	/*
310 	 * cpu_base_spl must be set explicitly here to prevent any blocking
311 	 * operations in mp_startup from causing the spl of the cpu to drop
312 	 * to 0 (allowing device interrupts before we're ready) in resume().
313 	 * cpu_base_spl MUST remain at LOCK_LEVEL until the cpu is CPU_READY.
314 	 * As an extra bit of security on DEBUG kernels, this is enforced with
315 	 * an assertion in mp_startup() -- before cpu_base_spl is set to its
316 	 * proper value.
317 	 */
318 	cp->cpu_base_spl = ipltospl(LOCK_LEVEL);
319 
320 	/*
321 	 * Now, initialize per-CPU idle thread for this CPU.
322 	 */
323 	tp = thread_create(NULL, PAGESIZE, idle, NULL, 0, procp, TS_ONPROC, -1);
324 
325 	cp->cpu_idle_thread = tp;
326 
327 	tp->t_preempt = 1;
328 	tp->t_bound_cpu = cp;
329 	tp->t_affinitycnt = 1;
330 	tp->t_cpu = cp;
331 	tp->t_disp_queue = cp->cpu_disp;
332 
333 	/*
334 	 * Bootstrap the CPU's PG data
335 	 */
336 	pg_cpu_bootstrap(cp);
337 
338 	/*
339 	 * Perform CPC initialization on the new CPU.
340 	 */
341 	kcpc_hw_init(cp);
342 
343 	/*
344 	 * Allocate virtual addresses for cpu_caddr1 and cpu_caddr2
345 	 * for each CPU.
346 	 */
347 	setup_vaddr_for_ppcopy(cp);
348 
349 	/*
350 	 * Allocate page for new GDT and initialize from current GDT.
351 	 */
352 #if !defined(__lint)
353 	ASSERT((sizeof (*cp->cpu_gdt) * NGDT) <= PAGESIZE);
354 #endif
355 	cp->cpu_gdt = kmem_zalloc(PAGESIZE, KM_SLEEP);
356 	bcopy(CPU->cpu_gdt, cp->cpu_gdt, (sizeof (*cp->cpu_gdt) * NGDT));
357 
358 #if defined(__i386)
359 	/*
360 	 * setup kernel %gs.
361 	 */
362 	set_usegd(&cp->cpu_gdt[GDT_GS], cp, sizeof (struct cpu) -1, SDT_MEMRWA,
363 	    SEL_KPL, 0, 1);
364 #endif
365 
366 	/*
367 	 * If we have more than one node, each cpu gets a copy of IDT
368 	 * local to its node. If this is a Pentium box, we use cpu 0's
369 	 * IDT. cpu 0's IDT has been made read-only to workaround the
370 	 * cmpxchgl register bug
371 	 */
372 	if (system_hardware.hd_nodes && x86_type != X86_TYPE_P5) {
373 #if !defined(__lint)
374 		ASSERT((sizeof (*CPU->cpu_idt) * NIDT) <= PAGESIZE);
375 #endif
376 		cp->cpu_idt = kmem_zalloc(PAGESIZE, KM_SLEEP);
377 		bcopy(CPU->cpu_idt, cp->cpu_idt, PAGESIZE);
378 	} else {
379 		cp->cpu_idt = CPU->cpu_idt;
380 	}
381 
382 	/*
383 	 * Get interrupt priority data from cpu 0.
384 	 */
385 	cp->cpu_pri_data = CPU->cpu_pri_data;
386 
387 	/*
388 	 * alloc space for cpuid info
389 	 */
390 	cpuid_alloc_space(cp);
391 
392 #if !defined(__xpv)
393 	/*
394 	 * alloc space for ucode_info
395 	 */
396 	ucode_alloc_space(cp);
397 #endif
398 
399 	hat_cpu_online(cp);
400 
401 #ifdef TRAPTRACE
402 	/*
403 	 * If this is a TRAPTRACE kernel, allocate TRAPTRACE buffers
404 	 */
405 	ttc->ttc_first = (uintptr_t)kmem_zalloc(trap_trace_bufsize, KM_SLEEP);
406 	ttc->ttc_next = ttc->ttc_first;
407 	ttc->ttc_limit = ttc->ttc_first + trap_trace_bufsize;
408 #endif
409 	/*
410 	 * Record that we have another CPU.
411 	 */
412 	mutex_enter(&cpu_lock);
413 	/*
414 	 * Initialize the interrupt threads for this CPU
415 	 */
416 	cpu_intr_alloc(cp, NINTR_THREADS);
417 	/*
418 	 * Add CPU to list of available CPUs.  It'll be on the active list
419 	 * after mp_startup().
420 	 */
421 	cpu_add_unit(cp);
422 	mutex_exit(&cpu_lock);
423 
424 	return (cp);
425 }
426 
427 /*
428  * Undo what was done in mp_startup_init
429  */
430 static void
431 mp_startup_fini(struct cpu *cp, int error)
432 {
433 	mutex_enter(&cpu_lock);
434 
435 	/*
436 	 * Remove the CPU from the list of available CPUs.
437 	 */
438 	cpu_del_unit(cp->cpu_id);
439 
440 	if (error == ETIMEDOUT) {
441 		/*
442 		 * The cpu was started, but never *seemed* to run any
443 		 * code in the kernel; it's probably off spinning in its
444 		 * own private world, though with potential references to
445 		 * our kmem-allocated IDTs and GDTs (for example).
446 		 *
447 		 * Worse still, it may actually wake up some time later,
448 		 * so rather than guess what it might or might not do, we
449 		 * leave the fundamental data structures intact.
450 		 */
451 		cp->cpu_flags = 0;
452 		mutex_exit(&cpu_lock);
453 		return;
454 	}
455 
456 	/*
457 	 * At this point, the only threads bound to this CPU should
458 	 * special per-cpu threads: it's idle thread, it's pause threads,
459 	 * and it's interrupt threads.  Clean these up.
460 	 */
461 	cpu_destroy_bound_threads(cp);
462 	cp->cpu_idle_thread = NULL;
463 
464 	/*
465 	 * Free the interrupt stack.
466 	 */
467 	segkp_release(segkp,
468 	    cp->cpu_intr_stack - (INTR_STACK_SIZE - SA(MINFRAME)));
469 
470 	mutex_exit(&cpu_lock);
471 
472 #ifdef TRAPTRACE
473 	/*
474 	 * Discard the trap trace buffer
475 	 */
476 	{
477 		trap_trace_ctl_t *ttc = &trap_trace_ctl[cp->cpu_id];
478 
479 		kmem_free((void *)ttc->ttc_first, trap_trace_bufsize);
480 		ttc->ttc_first = NULL;
481 	}
482 #endif
483 
484 	hat_cpu_offline(cp);
485 
486 	cpuid_free_space(cp);
487 
488 #if !defined(__xpv)
489 	ucode_free_space(cp);
490 #endif
491 
492 	if (cp->cpu_idt != CPU->cpu_idt)
493 		kmem_free(cp->cpu_idt, PAGESIZE);
494 	cp->cpu_idt = NULL;
495 
496 	kmem_free(cp->cpu_gdt, PAGESIZE);
497 	cp->cpu_gdt = NULL;
498 
499 	teardown_vaddr_for_ppcopy(cp);
500 
501 	kcpc_hw_fini(cp);
502 
503 	cp->cpu_dispthread = NULL;
504 	cp->cpu_thread = NULL;	/* discarded by cpu_destroy_bound_threads() */
505 
506 	cpu_vm_data_destroy(cp);
507 
508 	mutex_enter(&cpu_lock);
509 	disp_cpu_fini(cp);
510 	mutex_exit(&cpu_lock);
511 
512 #if !defined(__xpv)
513 	if (cp->cpu_m.mcpu_mwait != NULL)
514 		cpuid_mwait_free(cp);
515 #endif
516 	kmem_free(cp, sizeof (*cp));
517 }
518 
519 /*
520  * Apply workarounds for known errata, and warn about those that are absent.
521  *
522  * System vendors occasionally create configurations which contain different
523  * revisions of the CPUs that are almost but not exactly the same.  At the
524  * time of writing, this meant that their clock rates were the same, their
525  * feature sets were the same, but the required workaround were -not-
526  * necessarily the same.  So, this routine is invoked on -every- CPU soon
527  * after starting to make sure that the resulting system contains the most
528  * pessimal set of workarounds needed to cope with *any* of the CPUs in the
529  * system.
530  *
531  * workaround_errata is invoked early in mlsetup() for CPU 0, and in
532  * mp_startup() for all slave CPUs. Slaves process workaround_errata prior
533  * to acknowledging their readiness to the master, so this routine will
534  * never be executed by multiple CPUs in parallel, thus making updates to
535  * global data safe.
536  *
537  * These workarounds are based on Rev 3.57 of the Revision Guide for
538  * AMD Athlon(tm) 64 and AMD Opteron(tm) Processors, August 2005.
539  */
540 
541 #if defined(OPTERON_ERRATUM_88)
542 int opteron_erratum_88;		/* if non-zero -> at least one cpu has it */
543 #endif
544 
545 #if defined(OPTERON_ERRATUM_91)
546 int opteron_erratum_91;		/* if non-zero -> at least one cpu has it */
547 #endif
548 
549 #if defined(OPTERON_ERRATUM_93)
550 int opteron_erratum_93;		/* if non-zero -> at least one cpu has it */
551 #endif
552 
553 #if defined(OPTERON_ERRATUM_95)
554 int opteron_erratum_95;		/* if non-zero -> at least one cpu has it */
555 #endif
556 
557 #if defined(OPTERON_ERRATUM_100)
558 int opteron_erratum_100;	/* if non-zero -> at least one cpu has it */
559 #endif
560 
561 #if defined(OPTERON_ERRATUM_108)
562 int opteron_erratum_108;	/* if non-zero -> at least one cpu has it */
563 #endif
564 
565 #if defined(OPTERON_ERRATUM_109)
566 int opteron_erratum_109;	/* if non-zero -> at least one cpu has it */
567 #endif
568 
569 #if defined(OPTERON_ERRATUM_121)
570 int opteron_erratum_121;	/* if non-zero -> at least one cpu has it */
571 #endif
572 
573 #if defined(OPTERON_ERRATUM_122)
574 int opteron_erratum_122;	/* if non-zero -> at least one cpu has it */
575 #endif
576 
577 #if defined(OPTERON_ERRATUM_123)
578 int opteron_erratum_123;	/* if non-zero -> at least one cpu has it */
579 #endif
580 
581 #if defined(OPTERON_ERRATUM_131)
582 int opteron_erratum_131;	/* if non-zero -> at least one cpu has it */
583 #endif
584 
585 #if defined(OPTERON_WORKAROUND_6336786)
586 int opteron_workaround_6336786;	/* non-zero -> WA relevant and applied */
587 int opteron_workaround_6336786_UP = 0;	/* Not needed for UP */
588 #endif
589 
590 #if defined(OPTERON_WORKAROUND_6323525)
591 int opteron_workaround_6323525;	/* if non-zero -> at least one cpu has it */
592 #endif
593 
594 static void
595 workaround_warning(cpu_t *cp, uint_t erratum)
596 {
597 	cmn_err(CE_WARN, "cpu%d: no workaround for erratum %u",
598 	    cp->cpu_id, erratum);
599 }
600 
601 static void
602 workaround_applied(uint_t erratum)
603 {
604 	if (erratum > 1000000)
605 		cmn_err(CE_CONT, "?workaround applied for cpu issue #%d\n",
606 		    erratum);
607 	else
608 		cmn_err(CE_CONT, "?workaround applied for cpu erratum #%d\n",
609 		    erratum);
610 }
611 
612 static void
613 msr_warning(cpu_t *cp, const char *rw, uint_t msr, int error)
614 {
615 	cmn_err(CE_WARN, "cpu%d: couldn't %smsr 0x%x, error %d",
616 	    cp->cpu_id, rw, msr, error);
617 }
618 
619 /*
620  * Determine the number of nodes in an Opteron / Greyhound family system.
621  */
622 static uint_t
623 opteron_get_nnodes(void)
624 {
625 	static uint_t nnodes = 0;
626 
627 #ifdef	DEBUG
628 	uint_t family;
629 
630 	family = cpuid_getfamily(CPU);
631 	ASSERT(family == 0xf || family == 0x10);
632 #endif	/* DEBUG */
633 
634 	if (nnodes == 0) {
635 		/*
636 		 * Obtain the number of nodes in the system from
637 		 * bits [6:4] of the Node ID register on node 0.
638 		 *
639 		 * The actual node count is NodeID[6:4] + 1
640 		 *
641 		 * The Node ID register is accessed via function 0,
642 		 * offset 0x60. Node 0 is device 24.
643 		 */
644 		nnodes = ((pci_getl_func(0, 24, 0, 0x60) & 0x70) >> 4) + 1;
645 	}
646 	return (nnodes);
647 }
648 
649 #if defined(__xpv)
650 
651 /*
652  * On dom0, we can determine the number of physical cpus on the machine.
653  * This number is important when figuring out what workarounds are
654  * appropriate, so compute it now.
655  */
656 static uint_t
657 xen_get_nphyscpus(void)
658 {
659 	static uint_t nphyscpus = 0;
660 
661 	ASSERT(DOMAIN_IS_INITDOMAIN(xen_info));
662 
663 	if (nphyscpus == 0) {
664 		xen_sysctl_t op;
665 		xen_sysctl_physinfo_t *pi = &op.u.physinfo;
666 
667 		op.cmd = XEN_SYSCTL_physinfo;
668 		op.interface_version = XEN_SYSCTL_INTERFACE_VERSION;
669 		if (HYPERVISOR_sysctl(&op) == 0)
670 			nphyscpus = pi->threads_per_core *
671 			    pi->cores_per_socket * pi->sockets_per_node *
672 			    pi->nr_nodes;
673 	}
674 	return (nphyscpus);
675 }
676 #endif
677 
678 uint_t
679 workaround_errata(struct cpu *cpu)
680 {
681 	uint_t missing = 0;
682 
683 	ASSERT(cpu == CPU);
684 
685 	/*LINTED*/
686 	if (cpuid_opteron_erratum(cpu, 88) > 0) {
687 		/*
688 		 * SWAPGS May Fail To Read Correct GS Base
689 		 */
690 #if defined(OPTERON_ERRATUM_88)
691 		/*
692 		 * The workaround is an mfence in the relevant assembler code
693 		 */
694 		opteron_erratum_88++;
695 #else
696 		workaround_warning(cpu, 88);
697 		missing++;
698 #endif
699 	}
700 
701 	if (cpuid_opteron_erratum(cpu, 91) > 0) {
702 		/*
703 		 * Software Prefetches May Report A Page Fault
704 		 */
705 #if defined(OPTERON_ERRATUM_91)
706 		/*
707 		 * fix is in trap.c
708 		 */
709 		opteron_erratum_91++;
710 #else
711 		workaround_warning(cpu, 91);
712 		missing++;
713 #endif
714 	}
715 
716 	if (cpuid_opteron_erratum(cpu, 93) > 0) {
717 		/*
718 		 * RSM Auto-Halt Restart Returns to Incorrect RIP
719 		 */
720 #if defined(OPTERON_ERRATUM_93)
721 		/*
722 		 * fix is in trap.c
723 		 */
724 		opteron_erratum_93++;
725 #else
726 		workaround_warning(cpu, 93);
727 		missing++;
728 #endif
729 	}
730 
731 	/*LINTED*/
732 	if (cpuid_opteron_erratum(cpu, 95) > 0) {
733 		/*
734 		 * RET Instruction May Return to Incorrect EIP
735 		 */
736 #if defined(OPTERON_ERRATUM_95)
737 #if defined(_LP64)
738 		/*
739 		 * Workaround this by ensuring that 32-bit user code and
740 		 * 64-bit kernel code never occupy the same address
741 		 * range mod 4G.
742 		 */
743 		if (_userlimit32 > 0xc0000000ul)
744 			*(uintptr_t *)&_userlimit32 = 0xc0000000ul;
745 
746 		/*LINTED*/
747 		ASSERT((uint32_t)COREHEAP_BASE == 0xc0000000u);
748 		opteron_erratum_95++;
749 #endif	/* _LP64 */
750 #else
751 		workaround_warning(cpu, 95);
752 		missing++;
753 #endif
754 	}
755 
756 	if (cpuid_opteron_erratum(cpu, 100) > 0) {
757 		/*
758 		 * Compatibility Mode Branches Transfer to Illegal Address
759 		 */
760 #if defined(OPTERON_ERRATUM_100)
761 		/*
762 		 * fix is in trap.c
763 		 */
764 		opteron_erratum_100++;
765 #else
766 		workaround_warning(cpu, 100);
767 		missing++;
768 #endif
769 	}
770 
771 	/*LINTED*/
772 	if (cpuid_opteron_erratum(cpu, 108) > 0) {
773 		/*
774 		 * CPUID Instruction May Return Incorrect Model Number In
775 		 * Some Processors
776 		 */
777 #if defined(OPTERON_ERRATUM_108)
778 		/*
779 		 * (Our cpuid-handling code corrects the model number on
780 		 * those processors)
781 		 */
782 #else
783 		workaround_warning(cpu, 108);
784 		missing++;
785 #endif
786 	}
787 
788 	/*LINTED*/
789 	if (cpuid_opteron_erratum(cpu, 109) > 0) do {
790 		/*
791 		 * Certain Reverse REP MOVS May Produce Unpredictable Behaviour
792 		 */
793 #if defined(OPTERON_ERRATUM_109)
794 		/*
795 		 * The "workaround" is to print a warning to upgrade the BIOS
796 		 */
797 		uint64_t value;
798 		const uint_t msr = MSR_AMD_PATCHLEVEL;
799 		int err;
800 
801 		if ((err = checked_rdmsr(msr, &value)) != 0) {
802 			msr_warning(cpu, "rd", msr, err);
803 			workaround_warning(cpu, 109);
804 			missing++;
805 		}
806 		if (value == 0)
807 			opteron_erratum_109++;
808 #else
809 		workaround_warning(cpu, 109);
810 		missing++;
811 #endif
812 	/*CONSTANTCONDITION*/
813 	} while (0);
814 
815 	/*LINTED*/
816 	if (cpuid_opteron_erratum(cpu, 121) > 0) {
817 		/*
818 		 * Sequential Execution Across Non_Canonical Boundary Caused
819 		 * Processor Hang
820 		 */
821 #if defined(OPTERON_ERRATUM_121)
822 #if defined(_LP64)
823 		/*
824 		 * Erratum 121 is only present in long (64 bit) mode.
825 		 * Workaround is to include the page immediately before the
826 		 * va hole to eliminate the possibility of system hangs due to
827 		 * sequential execution across the va hole boundary.
828 		 */
829 		if (opteron_erratum_121)
830 			opteron_erratum_121++;
831 		else {
832 			if (hole_start) {
833 				hole_start -= PAGESIZE;
834 			} else {
835 				/*
836 				 * hole_start not yet initialized by
837 				 * mmu_init. Initialize hole_start
838 				 * with value to be subtracted.
839 				 */
840 				hole_start = PAGESIZE;
841 			}
842 			opteron_erratum_121++;
843 		}
844 #endif	/* _LP64 */
845 #else
846 		workaround_warning(cpu, 121);
847 		missing++;
848 #endif
849 	}
850 
851 	/*LINTED*/
852 	if (cpuid_opteron_erratum(cpu, 122) > 0) do {
853 		/*
854 		 * TLB Flush Filter May Cause Coherency Problem in
855 		 * Multiprocessor Systems
856 		 */
857 #if defined(OPTERON_ERRATUM_122)
858 		uint64_t value;
859 		const uint_t msr = MSR_AMD_HWCR;
860 		int error;
861 
862 		/*
863 		 * Erratum 122 is only present in MP configurations (multi-core
864 		 * or multi-processor).
865 		 */
866 #if defined(__xpv)
867 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
868 			break;
869 		if (!opteron_erratum_122 && xen_get_nphyscpus() == 1)
870 			break;
871 #else
872 		if (!opteron_erratum_122 && opteron_get_nnodes() == 1 &&
873 		    cpuid_get_ncpu_per_chip(cpu) == 1)
874 			break;
875 #endif
876 		/* disable TLB Flush Filter */
877 
878 		if ((error = checked_rdmsr(msr, &value)) != 0) {
879 			msr_warning(cpu, "rd", msr, error);
880 			workaround_warning(cpu, 122);
881 			missing++;
882 		} else {
883 			value |= (uint64_t)AMD_HWCR_FFDIS;
884 			if ((error = checked_wrmsr(msr, value)) != 0) {
885 				msr_warning(cpu, "wr", msr, error);
886 				workaround_warning(cpu, 122);
887 				missing++;
888 			}
889 		}
890 		opteron_erratum_122++;
891 #else
892 		workaround_warning(cpu, 122);
893 		missing++;
894 #endif
895 	/*CONSTANTCONDITION*/
896 	} while (0);
897 
898 	/*LINTED*/
899 	if (cpuid_opteron_erratum(cpu, 123) > 0) do {
900 		/*
901 		 * Bypassed Reads May Cause Data Corruption of System Hang in
902 		 * Dual Core Processors
903 		 */
904 #if defined(OPTERON_ERRATUM_123)
905 		uint64_t value;
906 		const uint_t msr = MSR_AMD_PATCHLEVEL;
907 		int err;
908 
909 		/*
910 		 * Erratum 123 applies only to multi-core cpus.
911 		 */
912 		if (cpuid_get_ncpu_per_chip(cpu) < 2)
913 			break;
914 #if defined(__xpv)
915 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
916 			break;
917 #endif
918 		/*
919 		 * The "workaround" is to print a warning to upgrade the BIOS
920 		 */
921 		if ((err = checked_rdmsr(msr, &value)) != 0) {
922 			msr_warning(cpu, "rd", msr, err);
923 			workaround_warning(cpu, 123);
924 			missing++;
925 		}
926 		if (value == 0)
927 			opteron_erratum_123++;
928 #else
929 		workaround_warning(cpu, 123);
930 		missing++;
931 
932 #endif
933 	/*CONSTANTCONDITION*/
934 	} while (0);
935 
936 	/*LINTED*/
937 	if (cpuid_opteron_erratum(cpu, 131) > 0) do {
938 		/*
939 		 * Multiprocessor Systems with Four or More Cores May Deadlock
940 		 * Waiting for a Probe Response
941 		 */
942 #if defined(OPTERON_ERRATUM_131)
943 		uint64_t nbcfg;
944 		const uint_t msr = MSR_AMD_NB_CFG;
945 		const uint64_t wabits =
946 		    AMD_NB_CFG_SRQ_HEARTBEAT | AMD_NB_CFG_SRQ_SPR;
947 		int error;
948 
949 		/*
950 		 * Erratum 131 applies to any system with four or more cores.
951 		 */
952 		if (opteron_erratum_131)
953 			break;
954 #if defined(__xpv)
955 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
956 			break;
957 		if (xen_get_nphyscpus() < 4)
958 			break;
959 #else
960 		if (opteron_get_nnodes() * cpuid_get_ncpu_per_chip(cpu) < 4)
961 			break;
962 #endif
963 		/*
964 		 * Print a warning if neither of the workarounds for
965 		 * erratum 131 is present.
966 		 */
967 		if ((error = checked_rdmsr(msr, &nbcfg)) != 0) {
968 			msr_warning(cpu, "rd", msr, error);
969 			workaround_warning(cpu, 131);
970 			missing++;
971 		} else if ((nbcfg & wabits) == 0) {
972 			opteron_erratum_131++;
973 		} else {
974 			/* cannot have both workarounds set */
975 			ASSERT((nbcfg & wabits) != wabits);
976 		}
977 #else
978 		workaround_warning(cpu, 131);
979 		missing++;
980 #endif
981 	/*CONSTANTCONDITION*/
982 	} while (0);
983 
984 	/*
985 	 * This isn't really an erratum, but for convenience the
986 	 * detection/workaround code lives here and in cpuid_opteron_erratum.
987 	 */
988 	if (cpuid_opteron_erratum(cpu, 6336786) > 0) {
989 #if defined(OPTERON_WORKAROUND_6336786)
990 		/*
991 		 * Disable C1-Clock ramping on multi-core/multi-processor
992 		 * K8 platforms to guard against TSC drift.
993 		 */
994 		if (opteron_workaround_6336786) {
995 			opteron_workaround_6336786++;
996 #if defined(__xpv)
997 		} else if ((DOMAIN_IS_INITDOMAIN(xen_info) &&
998 		    xen_get_nphyscpus() > 1) ||
999 		    opteron_workaround_6336786_UP) {
1000 			/*
1001 			 * XXPV	Hmm.  We can't walk the Northbridges on
1002 			 *	the hypervisor; so just complain and drive
1003 			 *	on.  This probably needs to be fixed in
1004 			 *	the hypervisor itself.
1005 			 */
1006 			opteron_workaround_6336786++;
1007 			workaround_warning(cpu, 6336786);
1008 #else	/* __xpv */
1009 		} else if ((opteron_get_nnodes() *
1010 		    cpuid_get_ncpu_per_chip(cpu) > 1) ||
1011 		    opteron_workaround_6336786_UP) {
1012 
1013 			uint_t	node, nnodes;
1014 			uint8_t data;
1015 
1016 			nnodes = opteron_get_nnodes();
1017 			for (node = 0; node < nnodes; node++) {
1018 				/*
1019 				 * Clear PMM7[1:0] (function 3, offset 0x87)
1020 				 * Northbridge device is the node id + 24.
1021 				 */
1022 				data = pci_getb_func(0, node + 24, 3, 0x87);
1023 				data &= 0xFC;
1024 				pci_putb_func(0, node + 24, 3, 0x87, data);
1025 			}
1026 			opteron_workaround_6336786++;
1027 #endif	/* __xpv */
1028 		}
1029 #else
1030 		workaround_warning(cpu, 6336786);
1031 		missing++;
1032 #endif
1033 	}
1034 
1035 	/*LINTED*/
1036 	/*
1037 	 * Mutex primitives don't work as expected.
1038 	 */
1039 	if (cpuid_opteron_erratum(cpu, 6323525) > 0) {
1040 #if defined(OPTERON_WORKAROUND_6323525)
1041 		/*
1042 		 * This problem only occurs with 2 or more cores. If bit in
1043 		 * MSR_BU_CFG set, then not applicable. The workaround
1044 		 * is to patch the semaphone routines with the lfence
1045 		 * instruction to provide necessary load memory barrier with
1046 		 * possible subsequent read-modify-write ops.
1047 		 *
1048 		 * It is too early in boot to call the patch routine so
1049 		 * set erratum variable to be done in startup_end().
1050 		 */
1051 		if (opteron_workaround_6323525) {
1052 			opteron_workaround_6323525++;
1053 #if defined(__xpv)
1054 		} else if (x86_feature & X86_SSE2) {
1055 			if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1056 				/*
1057 				 * XXPV	Use dom0_msr here when extended
1058 				 *	operations are supported?
1059 				 */
1060 				if (xen_get_nphyscpus() > 1)
1061 					opteron_workaround_6323525++;
1062 			} else {
1063 				/*
1064 				 * We have no way to tell how many physical
1065 				 * cpus there are, or even if this processor
1066 				 * has the problem, so enable the workaround
1067 				 * unconditionally (at some performance cost).
1068 				 */
1069 				opteron_workaround_6323525++;
1070 			}
1071 #else	/* __xpv */
1072 		} else if ((x86_feature & X86_SSE2) && ((opteron_get_nnodes() *
1073 		    cpuid_get_ncpu_per_chip(cpu)) > 1)) {
1074 			if ((xrdmsr(MSR_BU_CFG) & 0x02) == 0)
1075 				opteron_workaround_6323525++;
1076 #endif	/* __xpv */
1077 		}
1078 #else
1079 		workaround_warning(cpu, 6323525);
1080 		missing++;
1081 #endif
1082 	}
1083 
1084 #ifdef __xpv
1085 	return (0);
1086 #else
1087 	return (missing);
1088 #endif
1089 }
1090 
1091 void
1092 workaround_errata_end()
1093 {
1094 #if defined(OPTERON_ERRATUM_88)
1095 	if (opteron_erratum_88)
1096 		workaround_applied(88);
1097 #endif
1098 #if defined(OPTERON_ERRATUM_91)
1099 	if (opteron_erratum_91)
1100 		workaround_applied(91);
1101 #endif
1102 #if defined(OPTERON_ERRATUM_93)
1103 	if (opteron_erratum_93)
1104 		workaround_applied(93);
1105 #endif
1106 #if defined(OPTERON_ERRATUM_95)
1107 	if (opteron_erratum_95)
1108 		workaround_applied(95);
1109 #endif
1110 #if defined(OPTERON_ERRATUM_100)
1111 	if (opteron_erratum_100)
1112 		workaround_applied(100);
1113 #endif
1114 #if defined(OPTERON_ERRATUM_108)
1115 	if (opteron_erratum_108)
1116 		workaround_applied(108);
1117 #endif
1118 #if defined(OPTERON_ERRATUM_109)
1119 	if (opteron_erratum_109) {
1120 		cmn_err(CE_WARN,
1121 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1122 		    " processor\nerratum 109 was not detected; updating your"
1123 		    " system's BIOS to a version\ncontaining this"
1124 		    " microcode patch is HIGHLY recommended or erroneous"
1125 		    " system\noperation may occur.\n");
1126 	}
1127 #endif
1128 #if defined(OPTERON_ERRATUM_121)
1129 	if (opteron_erratum_121)
1130 		workaround_applied(121);
1131 #endif
1132 #if defined(OPTERON_ERRATUM_122)
1133 	if (opteron_erratum_122)
1134 		workaround_applied(122);
1135 #endif
1136 #if defined(OPTERON_ERRATUM_123)
1137 	if (opteron_erratum_123) {
1138 		cmn_err(CE_WARN,
1139 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1140 		    " processor\nerratum 123 was not detected; updating your"
1141 		    " system's BIOS to a version\ncontaining this"
1142 		    " microcode patch is HIGHLY recommended or erroneous"
1143 		    " system\noperation may occur.\n");
1144 	}
1145 #endif
1146 #if defined(OPTERON_ERRATUM_131)
1147 	if (opteron_erratum_131) {
1148 		cmn_err(CE_WARN,
1149 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1150 		    " processor\nerratum 131 was not detected; updating your"
1151 		    " system's BIOS to a version\ncontaining this"
1152 		    " microcode patch is HIGHLY recommended or erroneous"
1153 		    " system\noperation may occur.\n");
1154 	}
1155 #endif
1156 #if defined(OPTERON_WORKAROUND_6336786)
1157 	if (opteron_workaround_6336786)
1158 		workaround_applied(6336786);
1159 #endif
1160 #if defined(OPTERON_WORKAROUND_6323525)
1161 	if (opteron_workaround_6323525)
1162 		workaround_applied(6323525);
1163 #endif
1164 }
1165 
1166 static cpuset_t procset;
1167 
1168 /*
1169  * Start a single cpu, assuming that the kernel context is available
1170  * to successfully start another cpu.
1171  *
1172  * (For example, real mode code is mapped into the right place
1173  * in memory and is ready to be run.)
1174  */
1175 int
1176 start_cpu(processorid_t who)
1177 {
1178 	void *ctx;
1179 	cpu_t *cp;
1180 	int delays;
1181 	int error = 0;
1182 
1183 	ASSERT(who != 0);
1184 
1185 	/*
1186 	 * Check if there's at least a Mbyte of kmem available
1187 	 * before attempting to start the cpu.
1188 	 */
1189 	if (kmem_avail() < 1024 * 1024) {
1190 		/*
1191 		 * Kick off a reap in case that helps us with
1192 		 * later attempts ..
1193 		 */
1194 		kmem_reap();
1195 		return (ENOMEM);
1196 	}
1197 
1198 	cp = mp_startup_init(who);
1199 	if ((ctx = mach_cpucontext_alloc(cp)) == NULL ||
1200 	    (error = mach_cpu_start(cp, ctx)) != 0) {
1201 
1202 		/*
1203 		 * Something went wrong before we even started it
1204 		 */
1205 		if (ctx)
1206 			cmn_err(CE_WARN,
1207 			    "cpu%d: failed to start error %d",
1208 			    cp->cpu_id, error);
1209 		else
1210 			cmn_err(CE_WARN,
1211 			    "cpu%d: failed to allocate context", cp->cpu_id);
1212 
1213 		if (ctx)
1214 			mach_cpucontext_free(cp, ctx, error);
1215 		else
1216 			error = EAGAIN;		/* hmm. */
1217 		mp_startup_fini(cp, error);
1218 		return (error);
1219 	}
1220 
1221 	for (delays = 0; !CPU_IN_SET(procset, who); delays++) {
1222 		if (delays == 500) {
1223 			/*
1224 			 * After five seconds, things are probably looking
1225 			 * a bit bleak - explain the hang.
1226 			 */
1227 			cmn_err(CE_NOTE, "cpu%d: started, "
1228 			    "but not running in the kernel yet", who);
1229 		} else if (delays > 2000) {
1230 			/*
1231 			 * We waited at least 20 seconds, bail ..
1232 			 */
1233 			error = ETIMEDOUT;
1234 			cmn_err(CE_WARN, "cpu%d: timed out", who);
1235 			mach_cpucontext_free(cp, ctx, error);
1236 			mp_startup_fini(cp, error);
1237 			return (error);
1238 		}
1239 
1240 		/*
1241 		 * wait at least 10ms, then check again..
1242 		 */
1243 		delay(USEC_TO_TICK_ROUNDUP(10000));
1244 	}
1245 
1246 	mach_cpucontext_free(cp, ctx, 0);
1247 
1248 #ifndef __xpv
1249 	if (tsc_gethrtime_enable)
1250 		tsc_sync_master(who);
1251 #endif
1252 
1253 	if (dtrace_cpu_init != NULL) {
1254 		/*
1255 		 * DTrace CPU initialization expects cpu_lock to be held.
1256 		 */
1257 		mutex_enter(&cpu_lock);
1258 		(*dtrace_cpu_init)(who);
1259 		mutex_exit(&cpu_lock);
1260 	}
1261 
1262 	while (!CPU_IN_SET(cpu_ready_set, who))
1263 		delay(1);
1264 
1265 	return (0);
1266 }
1267 
1268 
1269 /*ARGSUSED*/
1270 void
1271 start_other_cpus(int cprboot)
1272 {
1273 	uint_t who;
1274 	uint_t skipped = 0;
1275 	uint_t bootcpuid = 0;
1276 
1277 	/*
1278 	 * Initialize our own cpu_info.
1279 	 */
1280 	init_cpu_info(CPU);
1281 
1282 	/*
1283 	 * Initialize our syscall handlers
1284 	 */
1285 	init_cpu_syscall(CPU);
1286 
1287 	/*
1288 	 * Take the boot cpu out of the mp_cpus set because we know
1289 	 * it's already running.  Add it to the cpu_ready_set for
1290 	 * precisely the same reason.
1291 	 */
1292 	CPUSET_DEL(mp_cpus, bootcpuid);
1293 	CPUSET_ADD(cpu_ready_set, bootcpuid);
1294 
1295 	/*
1296 	 * if only 1 cpu or not using MP, skip the rest of this
1297 	 */
1298 	if (CPUSET_ISNULL(mp_cpus) || use_mp == 0) {
1299 		if (use_mp == 0)
1300 			cmn_err(CE_CONT, "?***** Not in MP mode\n");
1301 		goto done;
1302 	}
1303 
1304 	/*
1305 	 * perform such initialization as is needed
1306 	 * to be able to take CPUs on- and off-line.
1307 	 */
1308 	cpu_pause_init();
1309 
1310 	xc_init();		/* initialize processor crosscalls */
1311 
1312 	if (mach_cpucontext_init() != 0)
1313 		goto done;
1314 
1315 	flushes_require_xcalls = 1;
1316 
1317 	/*
1318 	 * We lock our affinity to the master CPU to ensure that all slave CPUs
1319 	 * do their TSC syncs with the same CPU.
1320 	 */
1321 	affinity_set(CPU_CURRENT);
1322 
1323 	for (who = 0; who < NCPU; who++) {
1324 
1325 		if (!CPU_IN_SET(mp_cpus, who))
1326 			continue;
1327 		ASSERT(who != bootcpuid);
1328 		if (ncpus >= max_ncpus) {
1329 			skipped = who;
1330 			continue;
1331 		}
1332 		if (start_cpu(who) != 0)
1333 			CPUSET_DEL(mp_cpus, who);
1334 	}
1335 
1336 #if !defined(__xpv)
1337 	/* Free the space allocated to hold the microcode file */
1338 	ucode_free();
1339 #endif
1340 
1341 	affinity_clear();
1342 
1343 	if (skipped) {
1344 		cmn_err(CE_NOTE,
1345 		    "System detected %d cpus, but "
1346 		    "only %d cpu(s) were enabled during boot.",
1347 		    skipped + 1, ncpus);
1348 		cmn_err(CE_NOTE,
1349 		    "Use \"boot-ncpus\" parameter to enable more CPU(s). "
1350 		    "See eeprom(1M).");
1351 	}
1352 
1353 done:
1354 	workaround_errata_end();
1355 	mach_cpucontext_fini();
1356 
1357 	cmi_post_mpstartup();
1358 }
1359 
1360 /*
1361  * Dummy functions - no i86pc platforms support dynamic cpu allocation.
1362  */
1363 /*ARGSUSED*/
1364 int
1365 mp_cpu_configure(int cpuid)
1366 {
1367 	return (ENOTSUP);		/* not supported */
1368 }
1369 
1370 /*ARGSUSED*/
1371 int
1372 mp_cpu_unconfigure(int cpuid)
1373 {
1374 	return (ENOTSUP);		/* not supported */
1375 }
1376 
1377 /*
1378  * Startup function for 'other' CPUs (besides boot cpu).
1379  * Called from real_mode_start.
1380  *
1381  * WARNING: until CPU_READY is set, mp_startup and routines called by
1382  * mp_startup should not call routines (e.g. kmem_free) that could call
1383  * hat_unload which requires CPU_READY to be set.
1384  */
1385 void
1386 mp_startup(void)
1387 {
1388 	struct cpu *cp = CPU;
1389 	uint_t new_x86_feature;
1390 
1391 	/*
1392 	 * We need to get TSC on this proc synced (i.e., any delta
1393 	 * from cpu0 accounted for) as soon as we can, because many
1394 	 * many things use gethrtime/pc_gethrestime, including
1395 	 * interrupts, cmn_err, etc.
1396 	 */
1397 
1398 	/* Let cpu0 continue into tsc_sync_master() */
1399 	CPUSET_ATOMIC_ADD(procset, cp->cpu_id);
1400 
1401 #ifndef __xpv
1402 	if (tsc_gethrtime_enable)
1403 		tsc_sync_slave();
1404 #endif
1405 
1406 	/*
1407 	 * Once this was done from assembly, but it's safer here; if
1408 	 * it blocks, we need to be able to swtch() to and from, and
1409 	 * since we get here by calling t_pc, we need to do that call
1410 	 * before swtch() overwrites it.
1411 	 */
1412 
1413 	(void) (*ap_mlsetup)();
1414 
1415 	new_x86_feature = cpuid_pass1(cp);
1416 
1417 #ifndef __xpv
1418 	/*
1419 	 * Program this cpu's PAT
1420 	 */
1421 	if (x86_feature & X86_PAT)
1422 		pat_sync();
1423 #endif
1424 
1425 	/*
1426 	 * Set up TSC_AUX to contain the cpuid for this processor
1427 	 * for the rdtscp instruction.
1428 	 */
1429 	if (x86_feature & X86_TSCP)
1430 		(void) wrmsr(MSR_AMD_TSCAUX, cp->cpu_id);
1431 
1432 	/*
1433 	 * Initialize this CPU's syscall handlers
1434 	 */
1435 	init_cpu_syscall(cp);
1436 
1437 	/*
1438 	 * Enable interrupts with spl set to LOCK_LEVEL. LOCK_LEVEL is the
1439 	 * highest level at which a routine is permitted to block on
1440 	 * an adaptive mutex (allows for cpu poke interrupt in case
1441 	 * the cpu is blocked on a mutex and halts). Setting LOCK_LEVEL blocks
1442 	 * device interrupts that may end up in the hat layer issuing cross
1443 	 * calls before CPU_READY is set.
1444 	 */
1445 	splx(ipltospl(LOCK_LEVEL));
1446 	sti();
1447 
1448 	/*
1449 	 * Do a sanity check to make sure this new CPU is a sane thing
1450 	 * to add to the collection of processors running this system.
1451 	 *
1452 	 * XXX	Clearly this needs to get more sophisticated, if x86
1453 	 * systems start to get built out of heterogenous CPUs; as is
1454 	 * likely to happen once the number of processors in a configuration
1455 	 * gets large enough.
1456 	 */
1457 	if ((x86_feature & new_x86_feature) != x86_feature) {
1458 		cmn_err(CE_CONT, "?cpu%d: %b\n",
1459 		    cp->cpu_id, new_x86_feature, FMT_X86_FEATURE);
1460 		cmn_err(CE_WARN, "cpu%d feature mismatch", cp->cpu_id);
1461 	}
1462 
1463 	/*
1464 	 * We do not support cpus with mixed monitor/mwait support if the
1465 	 * boot cpu supports monitor/mwait.
1466 	 */
1467 	if ((x86_feature & ~new_x86_feature) & X86_MWAIT)
1468 		panic("unsupported mixed cpu monitor/mwait support detected");
1469 
1470 	/*
1471 	 * We could be more sophisticated here, and just mark the CPU
1472 	 * as "faulted" but at this point we'll opt for the easier
1473 	 * answer of dieing horribly.  Provided the boot cpu is ok,
1474 	 * the system can be recovered by booting with use_mp set to zero.
1475 	 */
1476 	if (workaround_errata(cp) != 0)
1477 		panic("critical workaround(s) missing for cpu%d", cp->cpu_id);
1478 
1479 	cpuid_pass2(cp);
1480 	cpuid_pass3(cp);
1481 	(void) cpuid_pass4(cp);
1482 
1483 	init_cpu_info(cp);
1484 
1485 	mutex_enter(&cpu_lock);
1486 	/*
1487 	 * Processor group initialization for this CPU is dependent on the
1488 	 * cpuid probing, which must be done in the context of the current
1489 	 * CPU.
1490 	 */
1491 	pghw_physid_create(cp);
1492 	pg_cpu_init(cp);
1493 	pg_cmt_cpu_startup(cp);
1494 
1495 	cp->cpu_flags |= CPU_RUNNING | CPU_READY | CPU_ENABLE | CPU_EXISTS;
1496 	cpu_add_active(cp);
1497 
1498 	if (dtrace_cpu_init != NULL) {
1499 		(*dtrace_cpu_init)(cp->cpu_id);
1500 	}
1501 
1502 #if !defined(__xpv)
1503 	/*
1504 	 * Fill out cpu_ucode_info.  Update microcode if necessary.
1505 	 */
1506 	ucode_check(cp);
1507 #endif
1508 
1509 	mutex_exit(&cpu_lock);
1510 
1511 	/*
1512 	 * Enable preemption here so that contention for any locks acquired
1513 	 * later in mp_startup may be preempted if the thread owning those
1514 	 * locks is continously executing on other CPUs (for example, this
1515 	 * CPU must be preemptible to allow other CPUs to pause it during their
1516 	 * startup phases).  It's safe to enable preemption here because the
1517 	 * CPU state is pretty-much fully constructed.
1518 	 */
1519 	curthread->t_preempt = 0;
1520 
1521 	add_cpunode2devtree(cp->cpu_id, cp->cpu_m.mcpu_cpi);
1522 
1523 	/* The base spl should still be at LOCK LEVEL here */
1524 	ASSERT(cp->cpu_base_spl == ipltospl(LOCK_LEVEL));
1525 	set_base_spl();		/* Restore the spl to its proper value */
1526 
1527 	(void) spl0();				/* enable interrupts */
1528 
1529 #ifndef __xpv
1530 	{
1531 		/*
1532 		 * Set up the CPU module for this CPU.  This can't be done
1533 		 * before this CPU is made CPU_READY, because we may (in
1534 		 * heterogeneous systems) need to go load another CPU module.
1535 		 * The act of attempting to load a module may trigger a
1536 		 * cross-call, which will ASSERT unless this cpu is CPU_READY.
1537 		 */
1538 		cmi_hdl_t hdl;
1539 
1540 		if ((hdl = cmi_init(CMI_HDL_NATIVE, cmi_ntv_hwchipid(CPU),
1541 		    cmi_ntv_hwcoreid(CPU), cmi_ntv_hwstrandid(CPU))) != NULL) {
1542 			if (x86_feature & X86_MCA)
1543 				cmi_mca_init(hdl);
1544 		}
1545 	}
1546 #endif /* __xpv */
1547 
1548 	if (boothowto & RB_DEBUG)
1549 		kdi_cpu_init();
1550 
1551 	/*
1552 	 * Setting the bit in cpu_ready_set must be the last operation in
1553 	 * processor initialization; the boot CPU will continue to boot once
1554 	 * it sees this bit set for all active CPUs.
1555 	 */
1556 	CPUSET_ATOMIC_ADD(cpu_ready_set, cp->cpu_id);
1557 
1558 	/*
1559 	 * Because mp_startup() gets fired off after init() starts, we
1560 	 * can't use the '?' trick to do 'boot -v' printing - so we
1561 	 * always direct the 'cpu .. online' messages to the log.
1562 	 */
1563 	cmn_err(CE_CONT, "!cpu%d initialization complete - online\n",
1564 	    cp->cpu_id);
1565 
1566 	/*
1567 	 * Now we are done with the startup thread, so free it up.
1568 	 */
1569 	thread_exit();
1570 	panic("mp_startup: cannot return");
1571 	/*NOTREACHED*/
1572 }
1573 
1574 
1575 /*
1576  * Start CPU on user request.
1577  */
1578 /* ARGSUSED */
1579 int
1580 mp_cpu_start(struct cpu *cp)
1581 {
1582 	ASSERT(MUTEX_HELD(&cpu_lock));
1583 	return (0);
1584 }
1585 
1586 /*
1587  * Stop CPU on user request.
1588  */
1589 /* ARGSUSED */
1590 int
1591 mp_cpu_stop(struct cpu *cp)
1592 {
1593 	extern int cbe_psm_timer_mode;
1594 	ASSERT(MUTEX_HELD(&cpu_lock));
1595 
1596 #ifdef __xpv
1597 	/*
1598 	 * We can't offline vcpu0.
1599 	 */
1600 	if (cp->cpu_id == 0)
1601 		return (EBUSY);
1602 #endif
1603 
1604 	/*
1605 	 * If TIMER_PERIODIC mode is used, CPU0 is the one running it;
1606 	 * can't stop it.  (This is true only for machines with no TSC.)
1607 	 */
1608 
1609 	if ((cbe_psm_timer_mode == TIMER_PERIODIC) && (cp->cpu_id == 0))
1610 		return (EBUSY);
1611 
1612 	return (0);
1613 }
1614 
1615 /*
1616  * Take the specified CPU out of participation in interrupts.
1617  */
1618 int
1619 cpu_disable_intr(struct cpu *cp)
1620 {
1621 	if (psm_disable_intr(cp->cpu_id) != DDI_SUCCESS)
1622 		return (EBUSY);
1623 
1624 	cp->cpu_flags &= ~CPU_ENABLE;
1625 	return (0);
1626 }
1627 
1628 /*
1629  * Allow the specified CPU to participate in interrupts.
1630  */
1631 void
1632 cpu_enable_intr(struct cpu *cp)
1633 {
1634 	ASSERT(MUTEX_HELD(&cpu_lock));
1635 	cp->cpu_flags |= CPU_ENABLE;
1636 	psm_enable_intr(cp->cpu_id);
1637 }
1638 
1639 
1640 /*ARGSUSED*/
1641 void
1642 mp_cpu_faulted_enter(struct cpu *cp)
1643 {
1644 #ifndef __xpv
1645 	cmi_hdl_t hdl = cmi_hdl_lookup(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp),
1646 	    cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp));
1647 
1648 	if (hdl != NULL) {
1649 		cmi_faulted_enter(hdl);
1650 		cmi_hdl_rele(hdl);
1651 	}
1652 #endif
1653 }
1654 
1655 /*ARGSUSED*/
1656 void
1657 mp_cpu_faulted_exit(struct cpu *cp)
1658 {
1659 #ifndef __xpv
1660 	cmi_hdl_t hdl = cmi_hdl_lookup(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp),
1661 	    cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp));
1662 
1663 	if (hdl != NULL) {
1664 		cmi_faulted_exit(hdl);
1665 		cmi_hdl_rele(hdl);
1666 	}
1667 #endif
1668 }
1669 
1670 /*
1671  * The following two routines are used as context operators on threads belonging
1672  * to processes with a private LDT (see sysi86).  Due to the rarity of such
1673  * processes, these routines are currently written for best code readability and
1674  * organization rather than speed.  We could avoid checking x86_feature at every
1675  * context switch by installing different context ops, depending on the
1676  * x86_feature flags, at LDT creation time -- one for each combination of fast
1677  * syscall feature flags.
1678  */
1679 
1680 /*ARGSUSED*/
1681 void
1682 cpu_fast_syscall_disable(void *arg)
1683 {
1684 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP))
1685 		cpu_sep_disable();
1686 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC))
1687 		cpu_asysc_disable();
1688 }
1689 
1690 /*ARGSUSED*/
1691 void
1692 cpu_fast_syscall_enable(void *arg)
1693 {
1694 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP))
1695 		cpu_sep_enable();
1696 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC))
1697 		cpu_asysc_enable();
1698 }
1699 
1700 static void
1701 cpu_sep_enable(void)
1702 {
1703 	ASSERT(x86_feature & X86_SEP);
1704 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1705 
1706 	wrmsr(MSR_INTC_SEP_CS, (uint64_t)(uintptr_t)KCS_SEL);
1707 }
1708 
1709 static void
1710 cpu_sep_disable(void)
1711 {
1712 	ASSERT(x86_feature & X86_SEP);
1713 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1714 
1715 	/*
1716 	 * Setting the SYSENTER_CS_MSR register to 0 causes software executing
1717 	 * the sysenter or sysexit instruction to trigger a #gp fault.
1718 	 */
1719 	wrmsr(MSR_INTC_SEP_CS, 0);
1720 }
1721 
1722 static void
1723 cpu_asysc_enable(void)
1724 {
1725 	ASSERT(x86_feature & X86_ASYSC);
1726 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1727 
1728 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) |
1729 	    (uint64_t)(uintptr_t)AMD_EFER_SCE);
1730 }
1731 
1732 static void
1733 cpu_asysc_disable(void)
1734 {
1735 	ASSERT(x86_feature & X86_ASYSC);
1736 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1737 
1738 	/*
1739 	 * Turn off the SCE (syscall enable) bit in the EFER register. Software
1740 	 * executing syscall or sysret with this bit off will incur a #ud trap.
1741 	 */
1742 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) &
1743 	    ~((uint64_t)(uintptr_t)AMD_EFER_SCE));
1744 }
1745