xref: /titanic_41/usr/src/uts/i86pc/os/mp_startup.c (revision d15447b6c777a1b2223924443bf36c9c8efb2ea4)
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 
97 cpuset_t cpu_ready_set;		/* initialized in startup() */
98 
99 static 	void	mp_startup(void);
100 
101 static void cpu_sep_enable(void);
102 static void cpu_sep_disable(void);
103 static void cpu_asysc_enable(void);
104 static void cpu_asysc_disable(void);
105 
106 /*
107  * Init CPU info - get CPU type info for processor_info system call.
108  */
109 void
110 init_cpu_info(struct cpu *cp)
111 {
112 	processor_info_t *pi = &cp->cpu_type_info;
113 	char buf[CPU_IDSTRLEN];
114 
115 	/*
116 	 * Get clock-frequency property for the CPU.
117 	 */
118 	pi->pi_clock = cpu_freq;
119 
120 	/*
121 	 * Current frequency in Hz.
122 	 */
123 	cp->cpu_curr_clock = cpu_freq_hz;
124 
125 	/*
126 	 * Supported frequencies.
127 	 */
128 	cpu_set_supp_freqs(cp, NULL);
129 
130 	(void) strcpy(pi->pi_processor_type, "i386");
131 	if (fpu_exists)
132 		(void) strcpy(pi->pi_fputypes, "i387 compatible");
133 
134 	(void) cpuid_getidstr(cp, buf, sizeof (buf));
135 
136 	cp->cpu_idstr = kmem_alloc(strlen(buf) + 1, KM_SLEEP);
137 	(void) strcpy(cp->cpu_idstr, buf);
138 
139 	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_idstr);
140 
141 	(void) cpuid_getbrandstr(cp, buf, sizeof (buf));
142 	cp->cpu_brandstr = kmem_alloc(strlen(buf) + 1, KM_SLEEP);
143 	(void) strcpy(cp->cpu_brandstr, buf);
144 
145 	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_brandstr);
146 }
147 
148 /*
149  * Configure syscall support on this CPU.
150  */
151 /*ARGSUSED*/
152 void
153 init_cpu_syscall(struct cpu *cp)
154 {
155 	kpreempt_disable();
156 
157 #if defined(__amd64)
158 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC)) {
159 
160 #if !defined(__lint)
161 		/*
162 		 * The syscall instruction imposes a certain ordering on
163 		 * segment selectors, so we double-check that ordering
164 		 * here.
165 		 */
166 		ASSERT(KDS_SEL == KCS_SEL + 8);
167 		ASSERT(UDS_SEL == U32CS_SEL + 8);
168 		ASSERT(UCS_SEL == U32CS_SEL + 16);
169 #endif
170 		/*
171 		 * Turn syscall/sysret extensions on.
172 		 */
173 		cpu_asysc_enable();
174 
175 		/*
176 		 * Program the magic registers ..
177 		 */
178 		wrmsr(MSR_AMD_STAR,
179 		    ((uint64_t)(U32CS_SEL << 16 | KCS_SEL)) << 32);
180 		wrmsr(MSR_AMD_LSTAR, (uint64_t)(uintptr_t)sys_syscall);
181 		wrmsr(MSR_AMD_CSTAR, (uint64_t)(uintptr_t)sys_syscall32);
182 
183 		/*
184 		 * This list of flags is masked off the incoming
185 		 * %rfl when we enter the kernel.
186 		 */
187 		wrmsr(MSR_AMD_SFMASK, (uint64_t)(uintptr_t)(PS_IE | PS_T));
188 	}
189 #endif
190 
191 	/*
192 	 * On 32-bit kernels, we use sysenter/sysexit because it's too
193 	 * hard to use syscall/sysret, and it is more portable anyway.
194 	 *
195 	 * On 64-bit kernels on Nocona machines, the 32-bit syscall
196 	 * variant isn't available to 32-bit applications, but sysenter is.
197 	 */
198 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP)) {
199 
200 #if !defined(__lint)
201 		/*
202 		 * The sysenter instruction imposes a certain ordering on
203 		 * segment selectors, so we double-check that ordering
204 		 * here. See "sysenter" in Intel document 245471-012, "IA-32
205 		 * Intel Architecture Software Developer's Manual Volume 2:
206 		 * Instruction Set Reference"
207 		 */
208 		ASSERT(KDS_SEL == KCS_SEL + 8);
209 
210 		ASSERT32(UCS_SEL == ((KCS_SEL + 16) | 3));
211 		ASSERT32(UDS_SEL == UCS_SEL + 8);
212 
213 		ASSERT64(U32CS_SEL == ((KCS_SEL + 16) | 3));
214 		ASSERT64(UDS_SEL == U32CS_SEL + 8);
215 #endif
216 
217 		cpu_sep_enable();
218 
219 		/*
220 		 * resume() sets this value to the base of the threads stack
221 		 * via a context handler.
222 		 */
223 		wrmsr(MSR_INTC_SEP_ESP, 0);
224 		wrmsr(MSR_INTC_SEP_EIP, (uint64_t)(uintptr_t)sys_sysenter);
225 	}
226 
227 	kpreempt_enable();
228 }
229 
230 /*
231  * Multiprocessor initialization.
232  *
233  * Allocate and initialize the cpu structure, TRAPTRACE buffer, and the
234  * startup and idle threads for the specified CPU.
235  */
236 struct cpu *
237 mp_startup_init(int cpun)
238 {
239 	struct cpu *cp;
240 	kthread_id_t tp;
241 	caddr_t	sp;
242 	proc_t *procp;
243 #if !defined(__xpv)
244 	extern int idle_cpu_prefer_mwait;
245 #endif
246 	extern void idle();
247 
248 #ifdef TRAPTRACE
249 	trap_trace_ctl_t *ttc = &trap_trace_ctl[cpun];
250 #endif
251 
252 	ASSERT(cpun < NCPU && cpu[cpun] == NULL);
253 
254 	cp = kmem_zalloc(sizeof (*cp), KM_SLEEP);
255 #if !defined(__xpv)
256 	if ((x86_feature & X86_MWAIT) && idle_cpu_prefer_mwait)
257 		cp->cpu_m.mcpu_mwait = cpuid_mwait_alloc(CPU);
258 #endif
259 
260 	procp = curthread->t_procp;
261 
262 	mutex_enter(&cpu_lock);
263 	/*
264 	 * Initialize the dispatcher first.
265 	 */
266 	disp_cpu_init(cp);
267 	mutex_exit(&cpu_lock);
268 
269 	cpu_vm_data_init(cp);
270 
271 	/*
272 	 * Allocate and initialize the startup thread for this CPU.
273 	 * Interrupt and process switch stacks get allocated later
274 	 * when the CPU starts running.
275 	 */
276 	tp = thread_create(NULL, 0, NULL, NULL, 0, procp,
277 	    TS_STOPPED, maxclsyspri);
278 
279 	/*
280 	 * Set state to TS_ONPROC since this thread will start running
281 	 * as soon as the CPU comes online.
282 	 *
283 	 * All the other fields of the thread structure are setup by
284 	 * thread_create().
285 	 */
286 	THREAD_ONPROC(tp, cp);
287 	tp->t_preempt = 1;
288 	tp->t_bound_cpu = cp;
289 	tp->t_affinitycnt = 1;
290 	tp->t_cpu = cp;
291 	tp->t_disp_queue = cp->cpu_disp;
292 
293 	/*
294 	 * Setup thread to start in mp_startup.
295 	 */
296 	sp = tp->t_stk;
297 	tp->t_pc = (uintptr_t)mp_startup;
298 	tp->t_sp = (uintptr_t)(sp - MINFRAME);
299 #if defined(__amd64)
300 	tp->t_sp -= STACK_ENTRY_ALIGN;		/* fake a call */
301 #endif
302 
303 	cp->cpu_id = cpun;
304 	cp->cpu_self = cp;
305 	cp->cpu_thread = tp;
306 	cp->cpu_lwp = NULL;
307 	cp->cpu_dispthread = tp;
308 	cp->cpu_dispatch_pri = DISP_PRIO(tp);
309 
310 	/*
311 	 * cpu_base_spl must be set explicitly here to prevent any blocking
312 	 * operations in mp_startup from causing the spl of the cpu to drop
313 	 * to 0 (allowing device interrupts before we're ready) in resume().
314 	 * cpu_base_spl MUST remain at LOCK_LEVEL until the cpu is CPU_READY.
315 	 * As an extra bit of security on DEBUG kernels, this is enforced with
316 	 * an assertion in mp_startup() -- before cpu_base_spl is set to its
317 	 * proper value.
318 	 */
319 	cp->cpu_base_spl = ipltospl(LOCK_LEVEL);
320 
321 	/*
322 	 * Now, initialize per-CPU idle thread for this CPU.
323 	 */
324 	tp = thread_create(NULL, PAGESIZE, idle, NULL, 0, procp, TS_ONPROC, -1);
325 
326 	cp->cpu_idle_thread = tp;
327 
328 	tp->t_preempt = 1;
329 	tp->t_bound_cpu = cp;
330 	tp->t_affinitycnt = 1;
331 	tp->t_cpu = cp;
332 	tp->t_disp_queue = cp->cpu_disp;
333 
334 	/*
335 	 * Bootstrap the CPU's PG data
336 	 */
337 	pg_cpu_bootstrap(cp);
338 
339 	/*
340 	 * Perform CPC initialization on the new CPU.
341 	 */
342 	kcpc_hw_init(cp);
343 
344 	/*
345 	 * Allocate virtual addresses for cpu_caddr1 and cpu_caddr2
346 	 * for each CPU.
347 	 */
348 	setup_vaddr_for_ppcopy(cp);
349 
350 	/*
351 	 * Allocate page for new GDT and initialize from current GDT.
352 	 */
353 #if !defined(__lint)
354 	ASSERT((sizeof (*cp->cpu_gdt) * NGDT) <= PAGESIZE);
355 #endif
356 	cp->cpu_gdt = kmem_zalloc(PAGESIZE, KM_SLEEP);
357 	bcopy(CPU->cpu_gdt, cp->cpu_gdt, (sizeof (*cp->cpu_gdt) * NGDT));
358 
359 #if defined(__i386)
360 	/*
361 	 * setup kernel %gs.
362 	 */
363 	set_usegd(&cp->cpu_gdt[GDT_GS], cp, sizeof (struct cpu) -1, SDT_MEMRWA,
364 	    SEL_KPL, 0, 1);
365 #endif
366 
367 	/*
368 	 * If we have more than one node, each cpu gets a copy of IDT
369 	 * local to its node. If this is a Pentium box, we use cpu 0's
370 	 * IDT. cpu 0's IDT has been made read-only to workaround the
371 	 * cmpxchgl register bug
372 	 */
373 	if (system_hardware.hd_nodes && x86_type != X86_TYPE_P5) {
374 #if !defined(__lint)
375 		ASSERT((sizeof (*CPU->cpu_idt) * NIDT) <= PAGESIZE);
376 #endif
377 		cp->cpu_idt = kmem_zalloc(PAGESIZE, KM_SLEEP);
378 		bcopy(CPU->cpu_idt, cp->cpu_idt, PAGESIZE);
379 	} else {
380 		cp->cpu_idt = CPU->cpu_idt;
381 	}
382 
383 	/*
384 	 * Get interrupt priority data from cpu 0.
385 	 */
386 	cp->cpu_pri_data = CPU->cpu_pri_data;
387 
388 	/*
389 	 * alloc space for cpuid info
390 	 */
391 	cpuid_alloc_space(cp);
392 
393 #if !defined(__xpv)
394 	/*
395 	 * alloc space for ucode_info
396 	 */
397 	ucode_alloc_space(cp);
398 #endif
399 
400 	hat_cpu_online(cp);
401 
402 #ifdef TRAPTRACE
403 	/*
404 	 * If this is a TRAPTRACE kernel, allocate TRAPTRACE buffers
405 	 */
406 	ttc->ttc_first = (uintptr_t)kmem_zalloc(trap_trace_bufsize, KM_SLEEP);
407 	ttc->ttc_next = ttc->ttc_first;
408 	ttc->ttc_limit = ttc->ttc_first + trap_trace_bufsize;
409 #endif
410 	/*
411 	 * Record that we have another CPU.
412 	 */
413 	mutex_enter(&cpu_lock);
414 	/*
415 	 * Initialize the interrupt threads for this CPU
416 	 */
417 	cpu_intr_alloc(cp, NINTR_THREADS);
418 	/*
419 	 * Add CPU to list of available CPUs.  It'll be on the active list
420 	 * after mp_startup().
421 	 */
422 	cpu_add_unit(cp);
423 	mutex_exit(&cpu_lock);
424 
425 	return (cp);
426 }
427 
428 /*
429  * Undo what was done in mp_startup_init
430  */
431 static void
432 mp_startup_fini(struct cpu *cp, int error)
433 {
434 	mutex_enter(&cpu_lock);
435 
436 	/*
437 	 * Remove the CPU from the list of available CPUs.
438 	 */
439 	cpu_del_unit(cp->cpu_id);
440 
441 	if (error == ETIMEDOUT) {
442 		/*
443 		 * The cpu was started, but never *seemed* to run any
444 		 * code in the kernel; it's probably off spinning in its
445 		 * own private world, though with potential references to
446 		 * our kmem-allocated IDTs and GDTs (for example).
447 		 *
448 		 * Worse still, it may actually wake up some time later,
449 		 * so rather than guess what it might or might not do, we
450 		 * leave the fundamental data structures intact.
451 		 */
452 		cp->cpu_flags = 0;
453 		mutex_exit(&cpu_lock);
454 		return;
455 	}
456 
457 	/*
458 	 * At this point, the only threads bound to this CPU should
459 	 * special per-cpu threads: it's idle thread, it's pause threads,
460 	 * and it's interrupt threads.  Clean these up.
461 	 */
462 	cpu_destroy_bound_threads(cp);
463 	cp->cpu_idle_thread = NULL;
464 
465 	/*
466 	 * Free the interrupt stack.
467 	 */
468 	segkp_release(segkp,
469 	    cp->cpu_intr_stack - (INTR_STACK_SIZE - SA(MINFRAME)));
470 
471 	mutex_exit(&cpu_lock);
472 
473 #ifdef TRAPTRACE
474 	/*
475 	 * Discard the trap trace buffer
476 	 */
477 	{
478 		trap_trace_ctl_t *ttc = &trap_trace_ctl[cp->cpu_id];
479 
480 		kmem_free((void *)ttc->ttc_first, trap_trace_bufsize);
481 		ttc->ttc_first = NULL;
482 	}
483 #endif
484 
485 	hat_cpu_offline(cp);
486 
487 	cpuid_free_space(cp);
488 
489 #if !defined(__xpv)
490 	ucode_free_space(cp);
491 #endif
492 
493 	if (cp->cpu_idt != CPU->cpu_idt)
494 		kmem_free(cp->cpu_idt, PAGESIZE);
495 	cp->cpu_idt = NULL;
496 
497 	kmem_free(cp->cpu_gdt, PAGESIZE);
498 	cp->cpu_gdt = NULL;
499 
500 	teardown_vaddr_for_ppcopy(cp);
501 
502 	kcpc_hw_fini(cp);
503 
504 	cp->cpu_dispthread = NULL;
505 	cp->cpu_thread = NULL;	/* discarded by cpu_destroy_bound_threads() */
506 
507 	cpu_vm_data_destroy(cp);
508 
509 	mutex_enter(&cpu_lock);
510 	disp_cpu_fini(cp);
511 	mutex_exit(&cpu_lock);
512 
513 #if !defined(__xpv)
514 	if (cp->cpu_m.mcpu_mwait != NULL)
515 		cpuid_mwait_free(cp);
516 #endif
517 	kmem_free(cp, sizeof (*cp));
518 }
519 
520 /*
521  * Apply workarounds for known errata, and warn about those that are absent.
522  *
523  * System vendors occasionally create configurations which contain different
524  * revisions of the CPUs that are almost but not exactly the same.  At the
525  * time of writing, this meant that their clock rates were the same, their
526  * feature sets were the same, but the required workaround were -not-
527  * necessarily the same.  So, this routine is invoked on -every- CPU soon
528  * after starting to make sure that the resulting system contains the most
529  * pessimal set of workarounds needed to cope with *any* of the CPUs in the
530  * system.
531  *
532  * workaround_errata is invoked early in mlsetup() for CPU 0, and in
533  * mp_startup() for all slave CPUs. Slaves process workaround_errata prior
534  * to acknowledging their readiness to the master, so this routine will
535  * never be executed by multiple CPUs in parallel, thus making updates to
536  * global data safe.
537  *
538  * These workarounds are based on Rev 3.57 of the Revision Guide for
539  * AMD Athlon(tm) 64 and AMD Opteron(tm) Processors, August 2005.
540  */
541 
542 #if defined(OPTERON_ERRATUM_88)
543 int opteron_erratum_88;		/* if non-zero -> at least one cpu has it */
544 #endif
545 
546 #if defined(OPTERON_ERRATUM_91)
547 int opteron_erratum_91;		/* if non-zero -> at least one cpu has it */
548 #endif
549 
550 #if defined(OPTERON_ERRATUM_93)
551 int opteron_erratum_93;		/* if non-zero -> at least one cpu has it */
552 #endif
553 
554 #if defined(OPTERON_ERRATUM_95)
555 int opteron_erratum_95;		/* if non-zero -> at least one cpu has it */
556 #endif
557 
558 #if defined(OPTERON_ERRATUM_100)
559 int opteron_erratum_100;	/* if non-zero -> at least one cpu has it */
560 #endif
561 
562 #if defined(OPTERON_ERRATUM_108)
563 int opteron_erratum_108;	/* if non-zero -> at least one cpu has it */
564 #endif
565 
566 #if defined(OPTERON_ERRATUM_109)
567 int opteron_erratum_109;	/* if non-zero -> at least one cpu has it */
568 #endif
569 
570 #if defined(OPTERON_ERRATUM_121)
571 int opteron_erratum_121;	/* if non-zero -> at least one cpu has it */
572 #endif
573 
574 #if defined(OPTERON_ERRATUM_122)
575 int opteron_erratum_122;	/* if non-zero -> at least one cpu has it */
576 #endif
577 
578 #if defined(OPTERON_ERRATUM_123)
579 int opteron_erratum_123;	/* if non-zero -> at least one cpu has it */
580 #endif
581 
582 #if defined(OPTERON_ERRATUM_131)
583 int opteron_erratum_131;	/* if non-zero -> at least one cpu has it */
584 #endif
585 
586 #if defined(OPTERON_WORKAROUND_6336786)
587 int opteron_workaround_6336786;	/* non-zero -> WA relevant and applied */
588 int opteron_workaround_6336786_UP = 0;	/* Not needed for UP */
589 #endif
590 
591 #if defined(OPTERON_WORKAROUND_6323525)
592 int opteron_workaround_6323525;	/* if non-zero -> at least one cpu has it */
593 #endif
594 
595 #if defined(OPTERON_ERRATUM_298)
596 int opteron_erratum_298;
597 #endif
598 
599 static void
600 workaround_warning(cpu_t *cp, uint_t erratum)
601 {
602 	cmn_err(CE_WARN, "cpu%d: no workaround for erratum %u",
603 	    cp->cpu_id, erratum);
604 }
605 
606 static void
607 workaround_applied(uint_t erratum)
608 {
609 	if (erratum > 1000000)
610 		cmn_err(CE_CONT, "?workaround applied for cpu issue #%d\n",
611 		    erratum);
612 	else
613 		cmn_err(CE_CONT, "?workaround applied for cpu erratum #%d\n",
614 		    erratum);
615 }
616 
617 static void
618 msr_warning(cpu_t *cp, const char *rw, uint_t msr, int error)
619 {
620 	cmn_err(CE_WARN, "cpu%d: couldn't %smsr 0x%x, error %d",
621 	    cp->cpu_id, rw, msr, error);
622 }
623 
624 /*
625  * Determine the number of nodes in an Opteron / Greyhound family system.
626  */
627 static uint_t
628 opteron_get_nnodes(void)
629 {
630 	static uint_t nnodes = 0;
631 
632 #ifdef	DEBUG
633 	uint_t family;
634 
635 	family = cpuid_getfamily(CPU);
636 	ASSERT(family == 0xf || family == 0x10);
637 #endif	/* DEBUG */
638 
639 	if (nnodes == 0) {
640 		/*
641 		 * Obtain the number of nodes in the system from
642 		 * bits [6:4] of the Node ID register on node 0.
643 		 *
644 		 * The actual node count is NodeID[6:4] + 1
645 		 *
646 		 * The Node ID register is accessed via function 0,
647 		 * offset 0x60. Node 0 is device 24.
648 		 */
649 		nnodes = ((pci_getl_func(0, 24, 0, 0x60) & 0x70) >> 4) + 1;
650 	}
651 	return (nnodes);
652 }
653 
654 #if defined(__xpv)
655 
656 /*
657  * On dom0, we can determine the number of physical cpus on the machine.
658  * This number is important when figuring out what workarounds are
659  * appropriate, so compute it now.
660  */
661 uint_t
662 xen_get_nphyscpus(void)
663 {
664 	static uint_t nphyscpus = 0;
665 
666 	ASSERT(DOMAIN_IS_INITDOMAIN(xen_info));
667 
668 	if (nphyscpus == 0) {
669 		xen_sysctl_t op;
670 		xen_sysctl_physinfo_t *pi = &op.u.physinfo;
671 
672 		op.cmd = XEN_SYSCTL_physinfo;
673 		op.interface_version = XEN_SYSCTL_INTERFACE_VERSION;
674 		if (HYPERVISOR_sysctl(&op) == 0)
675 			nphyscpus = pi->threads_per_core *
676 			    pi->cores_per_socket * pi->sockets_per_node *
677 			    pi->nr_nodes;
678 	}
679 	return (nphyscpus);
680 }
681 #endif
682 
683 uint_t
684 do_erratum_298(struct cpu *cpu)
685 {
686 	static int	osvwrc = -3;
687 	extern int	osvw_opteron_erratum(cpu_t *, uint_t);
688 
689 	/*
690 	 * L2 Eviction May Occur During Processor Operation To Set
691 	 * Accessed or Dirty Bit.
692 	 */
693 	if (osvwrc == -3) {
694 		osvwrc = osvw_opteron_erratum(cpu, 298);
695 	} else {
696 		/* osvw return codes should be consistent for all cpus */
697 		ASSERT(osvwrc == osvw_opteron_erratum(cpu, 298));
698 	}
699 
700 	switch (osvwrc) {
701 	case 0:		/* erratum is not present: do nothing */
702 		break;
703 	case 1:		/* erratum is present: BIOS workaround applied */
704 		/*
705 		 * check if workaround is actually in place and issue warning
706 		 * if not.
707 		 */
708 		if (((rdmsr(MSR_AMD_HWCR) & AMD_HWCR_TLBCACHEDIS) == 0) ||
709 		    ((rdmsr(MSR_AMD_BU_CFG) & AMD_BU_CFG_E298) == 0)) {
710 #if defined(OPTERON_ERRATUM_298)
711 			opteron_erratum_298++;
712 #else
713 			workaround_warning(cpu, 298);
714 			return (1);
715 #endif
716 		}
717 		break;
718 	case -1:	/* cannot determine via osvw: check cpuid */
719 		if ((cpuid_opteron_erratum(cpu, 298) > 0) &&
720 		    (((rdmsr(MSR_AMD_HWCR) & AMD_HWCR_TLBCACHEDIS) == 0) ||
721 		    ((rdmsr(MSR_AMD_BU_CFG) & AMD_BU_CFG_E298) == 0))) {
722 #if defined(OPTERON_ERRATUM_298)
723 			opteron_erratum_298++;
724 #else
725 			workaround_warning(cpu, 298);
726 			return (1);
727 #endif
728 		}
729 		break;
730 	}
731 	return (0);
732 }
733 
734 uint_t
735 workaround_errata(struct cpu *cpu)
736 {
737 	uint_t missing = 0;
738 
739 	ASSERT(cpu == CPU);
740 
741 	/*LINTED*/
742 	if (cpuid_opteron_erratum(cpu, 88) > 0) {
743 		/*
744 		 * SWAPGS May Fail To Read Correct GS Base
745 		 */
746 #if defined(OPTERON_ERRATUM_88)
747 		/*
748 		 * The workaround is an mfence in the relevant assembler code
749 		 */
750 		opteron_erratum_88++;
751 #else
752 		workaround_warning(cpu, 88);
753 		missing++;
754 #endif
755 	}
756 
757 	if (cpuid_opteron_erratum(cpu, 91) > 0) {
758 		/*
759 		 * Software Prefetches May Report A Page Fault
760 		 */
761 #if defined(OPTERON_ERRATUM_91)
762 		/*
763 		 * fix is in trap.c
764 		 */
765 		opteron_erratum_91++;
766 #else
767 		workaround_warning(cpu, 91);
768 		missing++;
769 #endif
770 	}
771 
772 	if (cpuid_opteron_erratum(cpu, 93) > 0) {
773 		/*
774 		 * RSM Auto-Halt Restart Returns to Incorrect RIP
775 		 */
776 #if defined(OPTERON_ERRATUM_93)
777 		/*
778 		 * fix is in trap.c
779 		 */
780 		opteron_erratum_93++;
781 #else
782 		workaround_warning(cpu, 93);
783 		missing++;
784 #endif
785 	}
786 
787 	/*LINTED*/
788 	if (cpuid_opteron_erratum(cpu, 95) > 0) {
789 		/*
790 		 * RET Instruction May Return to Incorrect EIP
791 		 */
792 #if defined(OPTERON_ERRATUM_95)
793 #if defined(_LP64)
794 		/*
795 		 * Workaround this by ensuring that 32-bit user code and
796 		 * 64-bit kernel code never occupy the same address
797 		 * range mod 4G.
798 		 */
799 		if (_userlimit32 > 0xc0000000ul)
800 			*(uintptr_t *)&_userlimit32 = 0xc0000000ul;
801 
802 		/*LINTED*/
803 		ASSERT((uint32_t)COREHEAP_BASE == 0xc0000000u);
804 		opteron_erratum_95++;
805 #endif	/* _LP64 */
806 #else
807 		workaround_warning(cpu, 95);
808 		missing++;
809 #endif
810 	}
811 
812 	if (cpuid_opteron_erratum(cpu, 100) > 0) {
813 		/*
814 		 * Compatibility Mode Branches Transfer to Illegal Address
815 		 */
816 #if defined(OPTERON_ERRATUM_100)
817 		/*
818 		 * fix is in trap.c
819 		 */
820 		opteron_erratum_100++;
821 #else
822 		workaround_warning(cpu, 100);
823 		missing++;
824 #endif
825 	}
826 
827 	/*LINTED*/
828 	if (cpuid_opteron_erratum(cpu, 108) > 0) {
829 		/*
830 		 * CPUID Instruction May Return Incorrect Model Number In
831 		 * Some Processors
832 		 */
833 #if defined(OPTERON_ERRATUM_108)
834 		/*
835 		 * (Our cpuid-handling code corrects the model number on
836 		 * those processors)
837 		 */
838 #else
839 		workaround_warning(cpu, 108);
840 		missing++;
841 #endif
842 	}
843 
844 	/*LINTED*/
845 	if (cpuid_opteron_erratum(cpu, 109) > 0) do {
846 		/*
847 		 * Certain Reverse REP MOVS May Produce Unpredictable Behaviour
848 		 */
849 #if defined(OPTERON_ERRATUM_109)
850 		/*
851 		 * The "workaround" is to print a warning to upgrade the BIOS
852 		 */
853 		uint64_t value;
854 		const uint_t msr = MSR_AMD_PATCHLEVEL;
855 		int err;
856 
857 		if ((err = checked_rdmsr(msr, &value)) != 0) {
858 			msr_warning(cpu, "rd", msr, err);
859 			workaround_warning(cpu, 109);
860 			missing++;
861 		}
862 		if (value == 0)
863 			opteron_erratum_109++;
864 #else
865 		workaround_warning(cpu, 109);
866 		missing++;
867 #endif
868 	/*CONSTANTCONDITION*/
869 	} while (0);
870 
871 	/*LINTED*/
872 	if (cpuid_opteron_erratum(cpu, 121) > 0) {
873 		/*
874 		 * Sequential Execution Across Non_Canonical Boundary Caused
875 		 * Processor Hang
876 		 */
877 #if defined(OPTERON_ERRATUM_121)
878 #if defined(_LP64)
879 		/*
880 		 * Erratum 121 is only present in long (64 bit) mode.
881 		 * Workaround is to include the page immediately before the
882 		 * va hole to eliminate the possibility of system hangs due to
883 		 * sequential execution across the va hole boundary.
884 		 */
885 		if (opteron_erratum_121)
886 			opteron_erratum_121++;
887 		else {
888 			if (hole_start) {
889 				hole_start -= PAGESIZE;
890 			} else {
891 				/*
892 				 * hole_start not yet initialized by
893 				 * mmu_init. Initialize hole_start
894 				 * with value to be subtracted.
895 				 */
896 				hole_start = PAGESIZE;
897 			}
898 			opteron_erratum_121++;
899 		}
900 #endif	/* _LP64 */
901 #else
902 		workaround_warning(cpu, 121);
903 		missing++;
904 #endif
905 	}
906 
907 	/*LINTED*/
908 	if (cpuid_opteron_erratum(cpu, 122) > 0) do {
909 		/*
910 		 * TLB Flush Filter May Cause Coherency Problem in
911 		 * Multiprocessor Systems
912 		 */
913 #if defined(OPTERON_ERRATUM_122)
914 		uint64_t value;
915 		const uint_t msr = MSR_AMD_HWCR;
916 		int error;
917 
918 		/*
919 		 * Erratum 122 is only present in MP configurations (multi-core
920 		 * or multi-processor).
921 		 */
922 #if defined(__xpv)
923 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
924 			break;
925 		if (!opteron_erratum_122 && xen_get_nphyscpus() == 1)
926 			break;
927 #else
928 		if (!opteron_erratum_122 && opteron_get_nnodes() == 1 &&
929 		    cpuid_get_ncpu_per_chip(cpu) == 1)
930 			break;
931 #endif
932 		/* disable TLB Flush Filter */
933 
934 		if ((error = checked_rdmsr(msr, &value)) != 0) {
935 			msr_warning(cpu, "rd", msr, error);
936 			workaround_warning(cpu, 122);
937 			missing++;
938 		} else {
939 			value |= (uint64_t)AMD_HWCR_FFDIS;
940 			if ((error = checked_wrmsr(msr, value)) != 0) {
941 				msr_warning(cpu, "wr", msr, error);
942 				workaround_warning(cpu, 122);
943 				missing++;
944 			}
945 		}
946 		opteron_erratum_122++;
947 #else
948 		workaround_warning(cpu, 122);
949 		missing++;
950 #endif
951 	/*CONSTANTCONDITION*/
952 	} while (0);
953 
954 	/*LINTED*/
955 	if (cpuid_opteron_erratum(cpu, 123) > 0) do {
956 		/*
957 		 * Bypassed Reads May Cause Data Corruption of System Hang in
958 		 * Dual Core Processors
959 		 */
960 #if defined(OPTERON_ERRATUM_123)
961 		uint64_t value;
962 		const uint_t msr = MSR_AMD_PATCHLEVEL;
963 		int err;
964 
965 		/*
966 		 * Erratum 123 applies only to multi-core cpus.
967 		 */
968 		if (cpuid_get_ncpu_per_chip(cpu) < 2)
969 			break;
970 #if defined(__xpv)
971 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
972 			break;
973 #endif
974 		/*
975 		 * The "workaround" is to print a warning to upgrade the BIOS
976 		 */
977 		if ((err = checked_rdmsr(msr, &value)) != 0) {
978 			msr_warning(cpu, "rd", msr, err);
979 			workaround_warning(cpu, 123);
980 			missing++;
981 		}
982 		if (value == 0)
983 			opteron_erratum_123++;
984 #else
985 		workaround_warning(cpu, 123);
986 		missing++;
987 
988 #endif
989 	/*CONSTANTCONDITION*/
990 	} while (0);
991 
992 	/*LINTED*/
993 	if (cpuid_opteron_erratum(cpu, 131) > 0) do {
994 		/*
995 		 * Multiprocessor Systems with Four or More Cores May Deadlock
996 		 * Waiting for a Probe Response
997 		 */
998 #if defined(OPTERON_ERRATUM_131)
999 		uint64_t nbcfg;
1000 		const uint_t msr = MSR_AMD_NB_CFG;
1001 		const uint64_t wabits =
1002 		    AMD_NB_CFG_SRQ_HEARTBEAT | AMD_NB_CFG_SRQ_SPR;
1003 		int error;
1004 
1005 		/*
1006 		 * Erratum 131 applies to any system with four or more cores.
1007 		 */
1008 		if (opteron_erratum_131)
1009 			break;
1010 #if defined(__xpv)
1011 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
1012 			break;
1013 		if (xen_get_nphyscpus() < 4)
1014 			break;
1015 #else
1016 		if (opteron_get_nnodes() * cpuid_get_ncpu_per_chip(cpu) < 4)
1017 			break;
1018 #endif
1019 		/*
1020 		 * Print a warning if neither of the workarounds for
1021 		 * erratum 131 is present.
1022 		 */
1023 		if ((error = checked_rdmsr(msr, &nbcfg)) != 0) {
1024 			msr_warning(cpu, "rd", msr, error);
1025 			workaround_warning(cpu, 131);
1026 			missing++;
1027 		} else if ((nbcfg & wabits) == 0) {
1028 			opteron_erratum_131++;
1029 		} else {
1030 			/* cannot have both workarounds set */
1031 			ASSERT((nbcfg & wabits) != wabits);
1032 		}
1033 #else
1034 		workaround_warning(cpu, 131);
1035 		missing++;
1036 #endif
1037 	/*CONSTANTCONDITION*/
1038 	} while (0);
1039 
1040 	/*
1041 	 * This isn't really an erratum, but for convenience the
1042 	 * detection/workaround code lives here and in cpuid_opteron_erratum.
1043 	 */
1044 	if (cpuid_opteron_erratum(cpu, 6336786) > 0) {
1045 #if defined(OPTERON_WORKAROUND_6336786)
1046 		/*
1047 		 * Disable C1-Clock ramping on multi-core/multi-processor
1048 		 * K8 platforms to guard against TSC drift.
1049 		 */
1050 		if (opteron_workaround_6336786) {
1051 			opteron_workaround_6336786++;
1052 #if defined(__xpv)
1053 		} else if ((DOMAIN_IS_INITDOMAIN(xen_info) &&
1054 		    xen_get_nphyscpus() > 1) ||
1055 		    opteron_workaround_6336786_UP) {
1056 			/*
1057 			 * XXPV	Hmm.  We can't walk the Northbridges on
1058 			 *	the hypervisor; so just complain and drive
1059 			 *	on.  This probably needs to be fixed in
1060 			 *	the hypervisor itself.
1061 			 */
1062 			opteron_workaround_6336786++;
1063 			workaround_warning(cpu, 6336786);
1064 #else	/* __xpv */
1065 		} else if ((opteron_get_nnodes() *
1066 		    cpuid_get_ncpu_per_chip(cpu) > 1) ||
1067 		    opteron_workaround_6336786_UP) {
1068 
1069 			uint_t	node, nnodes;
1070 			uint8_t data;
1071 
1072 			nnodes = opteron_get_nnodes();
1073 			for (node = 0; node < nnodes; node++) {
1074 				/*
1075 				 * Clear PMM7[1:0] (function 3, offset 0x87)
1076 				 * Northbridge device is the node id + 24.
1077 				 */
1078 				data = pci_getb_func(0, node + 24, 3, 0x87);
1079 				data &= 0xFC;
1080 				pci_putb_func(0, node + 24, 3, 0x87, data);
1081 			}
1082 			opteron_workaround_6336786++;
1083 #endif	/* __xpv */
1084 		}
1085 #else
1086 		workaround_warning(cpu, 6336786);
1087 		missing++;
1088 #endif
1089 	}
1090 
1091 	/*LINTED*/
1092 	/*
1093 	 * Mutex primitives don't work as expected.
1094 	 */
1095 	if (cpuid_opteron_erratum(cpu, 6323525) > 0) {
1096 #if defined(OPTERON_WORKAROUND_6323525)
1097 		/*
1098 		 * This problem only occurs with 2 or more cores. If bit in
1099 		 * MSR_AMD_BU_CFG set, then not applicable. The workaround
1100 		 * is to patch the semaphone routines with the lfence
1101 		 * instruction to provide necessary load memory barrier with
1102 		 * possible subsequent read-modify-write ops.
1103 		 *
1104 		 * It is too early in boot to call the patch routine so
1105 		 * set erratum variable to be done in startup_end().
1106 		 */
1107 		if (opteron_workaround_6323525) {
1108 			opteron_workaround_6323525++;
1109 #if defined(__xpv)
1110 		} else if (x86_feature & X86_SSE2) {
1111 			if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1112 				/*
1113 				 * XXPV	Use dom0_msr here when extended
1114 				 *	operations are supported?
1115 				 */
1116 				if (xen_get_nphyscpus() > 1)
1117 					opteron_workaround_6323525++;
1118 			} else {
1119 				/*
1120 				 * We have no way to tell how many physical
1121 				 * cpus there are, or even if this processor
1122 				 * has the problem, so enable the workaround
1123 				 * unconditionally (at some performance cost).
1124 				 */
1125 				opteron_workaround_6323525++;
1126 			}
1127 #else	/* __xpv */
1128 		} else if ((x86_feature & X86_SSE2) && ((opteron_get_nnodes() *
1129 		    cpuid_get_ncpu_per_chip(cpu)) > 1)) {
1130 			if ((xrdmsr(MSR_AMD_BU_CFG) & 0x02) == 0)
1131 				opteron_workaround_6323525++;
1132 #endif	/* __xpv */
1133 		}
1134 #else
1135 		workaround_warning(cpu, 6323525);
1136 		missing++;
1137 #endif
1138 	}
1139 
1140 	missing += do_erratum_298(cpu);
1141 
1142 #ifdef __xpv
1143 	return (0);
1144 #else
1145 	return (missing);
1146 #endif
1147 }
1148 
1149 void
1150 workaround_errata_end()
1151 {
1152 #if defined(OPTERON_ERRATUM_88)
1153 	if (opteron_erratum_88)
1154 		workaround_applied(88);
1155 #endif
1156 #if defined(OPTERON_ERRATUM_91)
1157 	if (opteron_erratum_91)
1158 		workaround_applied(91);
1159 #endif
1160 #if defined(OPTERON_ERRATUM_93)
1161 	if (opteron_erratum_93)
1162 		workaround_applied(93);
1163 #endif
1164 #if defined(OPTERON_ERRATUM_95)
1165 	if (opteron_erratum_95)
1166 		workaround_applied(95);
1167 #endif
1168 #if defined(OPTERON_ERRATUM_100)
1169 	if (opteron_erratum_100)
1170 		workaround_applied(100);
1171 #endif
1172 #if defined(OPTERON_ERRATUM_108)
1173 	if (opteron_erratum_108)
1174 		workaround_applied(108);
1175 #endif
1176 #if defined(OPTERON_ERRATUM_109)
1177 	if (opteron_erratum_109) {
1178 		cmn_err(CE_WARN,
1179 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1180 		    " processor\nerratum 109 was not detected; updating your"
1181 		    " system's BIOS to a version\ncontaining this"
1182 		    " microcode patch is HIGHLY recommended or erroneous"
1183 		    " system\noperation may occur.\n");
1184 	}
1185 #endif
1186 #if defined(OPTERON_ERRATUM_121)
1187 	if (opteron_erratum_121)
1188 		workaround_applied(121);
1189 #endif
1190 #if defined(OPTERON_ERRATUM_122)
1191 	if (opteron_erratum_122)
1192 		workaround_applied(122);
1193 #endif
1194 #if defined(OPTERON_ERRATUM_123)
1195 	if (opteron_erratum_123) {
1196 		cmn_err(CE_WARN,
1197 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1198 		    " processor\nerratum 123 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_ERRATUM_131)
1205 	if (opteron_erratum_131) {
1206 		cmn_err(CE_WARN,
1207 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1208 		    " processor\nerratum 131 was not detected; updating your"
1209 		    " system's BIOS to a version\ncontaining this"
1210 		    " microcode patch is HIGHLY recommended or erroneous"
1211 		    " system\noperation may occur.\n");
1212 	}
1213 #endif
1214 #if defined(OPTERON_WORKAROUND_6336786)
1215 	if (opteron_workaround_6336786)
1216 		workaround_applied(6336786);
1217 #endif
1218 #if defined(OPTERON_WORKAROUND_6323525)
1219 	if (opteron_workaround_6323525)
1220 		workaround_applied(6323525);
1221 #endif
1222 #if defined(OPTERON_ERRATUM_298)
1223 	if (opteron_erratum_298) {
1224 		cmn_err(CE_WARN,
1225 		    "BIOS microcode patch for AMD 64/Opteron(tm)"
1226 		    " processor\nerratum 298 was not detected; updating your"
1227 		    " system's BIOS to a version\ncontaining this"
1228 		    " microcode patch is HIGHLY recommended or erroneous"
1229 		    " system\noperation may occur.\n");
1230 	}
1231 #endif
1232 }
1233 
1234 static cpuset_t procset;
1235 
1236 /*
1237  * Start a single cpu, assuming that the kernel context is available
1238  * to successfully start another cpu.
1239  *
1240  * (For example, real mode code is mapped into the right place
1241  * in memory and is ready to be run.)
1242  */
1243 int
1244 start_cpu(processorid_t who)
1245 {
1246 	void *ctx;
1247 	cpu_t *cp;
1248 	int delays;
1249 	int error = 0;
1250 
1251 	ASSERT(who != 0);
1252 
1253 	/*
1254 	 * Check if there's at least a Mbyte of kmem available
1255 	 * before attempting to start the cpu.
1256 	 */
1257 	if (kmem_avail() < 1024 * 1024) {
1258 		/*
1259 		 * Kick off a reap in case that helps us with
1260 		 * later attempts ..
1261 		 */
1262 		kmem_reap();
1263 		return (ENOMEM);
1264 	}
1265 
1266 	cp = mp_startup_init(who);
1267 	if ((ctx = mach_cpucontext_alloc(cp)) == NULL ||
1268 	    (error = mach_cpu_start(cp, ctx)) != 0) {
1269 
1270 		/*
1271 		 * Something went wrong before we even started it
1272 		 */
1273 		if (ctx)
1274 			cmn_err(CE_WARN,
1275 			    "cpu%d: failed to start error %d",
1276 			    cp->cpu_id, error);
1277 		else
1278 			cmn_err(CE_WARN,
1279 			    "cpu%d: failed to allocate context", cp->cpu_id);
1280 
1281 		if (ctx)
1282 			mach_cpucontext_free(cp, ctx, error);
1283 		else
1284 			error = EAGAIN;		/* hmm. */
1285 		mp_startup_fini(cp, error);
1286 		return (error);
1287 	}
1288 
1289 	for (delays = 0; !CPU_IN_SET(procset, who); delays++) {
1290 		if (delays == 500) {
1291 			/*
1292 			 * After five seconds, things are probably looking
1293 			 * a bit bleak - explain the hang.
1294 			 */
1295 			cmn_err(CE_NOTE, "cpu%d: started, "
1296 			    "but not running in the kernel yet", who);
1297 		} else if (delays > 2000) {
1298 			/*
1299 			 * We waited at least 20 seconds, bail ..
1300 			 */
1301 			error = ETIMEDOUT;
1302 			cmn_err(CE_WARN, "cpu%d: timed out", who);
1303 			mach_cpucontext_free(cp, ctx, error);
1304 			mp_startup_fini(cp, error);
1305 			return (error);
1306 		}
1307 
1308 		/*
1309 		 * wait at least 10ms, then check again..
1310 		 */
1311 		delay(USEC_TO_TICK_ROUNDUP(10000));
1312 	}
1313 
1314 	mach_cpucontext_free(cp, ctx, 0);
1315 
1316 #ifndef __xpv
1317 	if (tsc_gethrtime_enable)
1318 		tsc_sync_master(who);
1319 #endif
1320 
1321 	if (dtrace_cpu_init != NULL) {
1322 		/*
1323 		 * DTrace CPU initialization expects cpu_lock to be held.
1324 		 */
1325 		mutex_enter(&cpu_lock);
1326 		(*dtrace_cpu_init)(who);
1327 		mutex_exit(&cpu_lock);
1328 	}
1329 
1330 	while (!CPU_IN_SET(cpu_ready_set, who))
1331 		delay(1);
1332 
1333 	return (0);
1334 }
1335 
1336 
1337 /*ARGSUSED*/
1338 void
1339 start_other_cpus(int cprboot)
1340 {
1341 	uint_t who;
1342 	uint_t skipped = 0;
1343 	uint_t bootcpuid = 0;
1344 
1345 	/*
1346 	 * Initialize our own cpu_info.
1347 	 */
1348 	init_cpu_info(CPU);
1349 
1350 	/*
1351 	 * Initialize our syscall handlers
1352 	 */
1353 	init_cpu_syscall(CPU);
1354 
1355 	/*
1356 	 * Take the boot cpu out of the mp_cpus set because we know
1357 	 * it's already running.  Add it to the cpu_ready_set for
1358 	 * precisely the same reason.
1359 	 */
1360 	CPUSET_DEL(mp_cpus, bootcpuid);
1361 	CPUSET_ADD(cpu_ready_set, bootcpuid);
1362 
1363 	/*
1364 	 * if only 1 cpu or not using MP, skip the rest of this
1365 	 */
1366 	if (CPUSET_ISNULL(mp_cpus) || use_mp == 0) {
1367 		if (use_mp == 0)
1368 			cmn_err(CE_CONT, "?***** Not in MP mode\n");
1369 		goto done;
1370 	}
1371 
1372 	/*
1373 	 * perform such initialization as is needed
1374 	 * to be able to take CPUs on- and off-line.
1375 	 */
1376 	cpu_pause_init();
1377 
1378 	xc_init();		/* initialize processor crosscalls */
1379 
1380 	if (mach_cpucontext_init() != 0)
1381 		goto done;
1382 
1383 	flushes_require_xcalls = 1;
1384 
1385 	/*
1386 	 * We lock our affinity to the master CPU to ensure that all slave CPUs
1387 	 * do their TSC syncs with the same CPU.
1388 	 */
1389 	affinity_set(CPU_CURRENT);
1390 
1391 	for (who = 0; who < NCPU; who++) {
1392 
1393 		if (!CPU_IN_SET(mp_cpus, who))
1394 			continue;
1395 		ASSERT(who != bootcpuid);
1396 		if (ncpus >= max_ncpus) {
1397 			skipped = who;
1398 			continue;
1399 		}
1400 		if (start_cpu(who) != 0)
1401 			CPUSET_DEL(mp_cpus, who);
1402 	}
1403 
1404 #if !defined(__xpv)
1405 	/* Free the space allocated to hold the microcode file */
1406 	ucode_free();
1407 #endif
1408 
1409 	affinity_clear();
1410 
1411 	if (skipped) {
1412 		cmn_err(CE_NOTE,
1413 		    "System detected %d cpus, but "
1414 		    "only %d cpu(s) were enabled during boot.",
1415 		    skipped + 1, ncpus);
1416 		cmn_err(CE_NOTE,
1417 		    "Use \"boot-ncpus\" parameter to enable more CPU(s). "
1418 		    "See eeprom(1M).");
1419 	}
1420 
1421 done:
1422 	workaround_errata_end();
1423 	mach_cpucontext_fini();
1424 
1425 	cmi_post_mpstartup();
1426 }
1427 
1428 /*
1429  * Dummy functions - no i86pc platforms support dynamic cpu allocation.
1430  */
1431 /*ARGSUSED*/
1432 int
1433 mp_cpu_configure(int cpuid)
1434 {
1435 	return (ENOTSUP);		/* not supported */
1436 }
1437 
1438 /*ARGSUSED*/
1439 int
1440 mp_cpu_unconfigure(int cpuid)
1441 {
1442 	return (ENOTSUP);		/* not supported */
1443 }
1444 
1445 /*
1446  * Startup function for 'other' CPUs (besides boot cpu).
1447  * Called from real_mode_start.
1448  *
1449  * WARNING: until CPU_READY is set, mp_startup and routines called by
1450  * mp_startup should not call routines (e.g. kmem_free) that could call
1451  * hat_unload which requires CPU_READY to be set.
1452  */
1453 void
1454 mp_startup(void)
1455 {
1456 	struct cpu *cp = CPU;
1457 	uint_t new_x86_feature;
1458 
1459 	/*
1460 	 * We need to get TSC on this proc synced (i.e., any delta
1461 	 * from cpu0 accounted for) as soon as we can, because many
1462 	 * many things use gethrtime/pc_gethrestime, including
1463 	 * interrupts, cmn_err, etc.
1464 	 */
1465 
1466 	/* Let cpu0 continue into tsc_sync_master() */
1467 	CPUSET_ATOMIC_ADD(procset, cp->cpu_id);
1468 
1469 #ifndef __xpv
1470 	if (tsc_gethrtime_enable)
1471 		tsc_sync_slave();
1472 #endif
1473 
1474 	/*
1475 	 * Once this was done from assembly, but it's safer here; if
1476 	 * it blocks, we need to be able to swtch() to and from, and
1477 	 * since we get here by calling t_pc, we need to do that call
1478 	 * before swtch() overwrites it.
1479 	 */
1480 	(void) (*ap_mlsetup)();
1481 
1482 	new_x86_feature = cpuid_pass1(cp);
1483 
1484 #ifndef __xpv
1485 	/*
1486 	 * Program this cpu's PAT
1487 	 */
1488 	if (x86_feature & X86_PAT)
1489 		pat_sync();
1490 #endif
1491 
1492 	/*
1493 	 * Set up TSC_AUX to contain the cpuid for this processor
1494 	 * for the rdtscp instruction.
1495 	 */
1496 	if (x86_feature & X86_TSCP)
1497 		(void) wrmsr(MSR_AMD_TSCAUX, cp->cpu_id);
1498 
1499 	/*
1500 	 * Initialize this CPU's syscall handlers
1501 	 */
1502 	init_cpu_syscall(cp);
1503 
1504 	/*
1505 	 * Enable interrupts with spl set to LOCK_LEVEL. LOCK_LEVEL is the
1506 	 * highest level at which a routine is permitted to block on
1507 	 * an adaptive mutex (allows for cpu poke interrupt in case
1508 	 * the cpu is blocked on a mutex and halts). Setting LOCK_LEVEL blocks
1509 	 * device interrupts that may end up in the hat layer issuing cross
1510 	 * calls before CPU_READY is set.
1511 	 */
1512 	splx(ipltospl(LOCK_LEVEL));
1513 	sti();
1514 
1515 	/*
1516 	 * Do a sanity check to make sure this new CPU is a sane thing
1517 	 * to add to the collection of processors running this system.
1518 	 *
1519 	 * XXX	Clearly this needs to get more sophisticated, if x86
1520 	 * systems start to get built out of heterogenous CPUs; as is
1521 	 * likely to happen once the number of processors in a configuration
1522 	 * gets large enough.
1523 	 */
1524 	if ((x86_feature & new_x86_feature) != x86_feature) {
1525 		cmn_err(CE_CONT, "?cpu%d: %b\n",
1526 		    cp->cpu_id, new_x86_feature, FMT_X86_FEATURE);
1527 		cmn_err(CE_WARN, "cpu%d feature mismatch", cp->cpu_id);
1528 	}
1529 
1530 	/*
1531 	 * We do not support cpus with mixed monitor/mwait support if the
1532 	 * boot cpu supports monitor/mwait.
1533 	 */
1534 	if ((x86_feature & ~new_x86_feature) & X86_MWAIT)
1535 		panic("unsupported mixed cpu monitor/mwait support detected");
1536 
1537 	/*
1538 	 * We could be more sophisticated here, and just mark the CPU
1539 	 * as "faulted" but at this point we'll opt for the easier
1540 	 * answer of dieing horribly.  Provided the boot cpu is ok,
1541 	 * the system can be recovered by booting with use_mp set to zero.
1542 	 */
1543 	if (workaround_errata(cp) != 0)
1544 		panic("critical workaround(s) missing for cpu%d", cp->cpu_id);
1545 
1546 	cpuid_pass2(cp);
1547 	cpuid_pass3(cp);
1548 	(void) cpuid_pass4(cp);
1549 
1550 	init_cpu_info(cp);
1551 
1552 	mutex_enter(&cpu_lock);
1553 	/*
1554 	 * Processor group initialization for this CPU is dependent on the
1555 	 * cpuid probing, which must be done in the context of the current
1556 	 * CPU.
1557 	 */
1558 	pghw_physid_create(cp);
1559 	pg_cpu_init(cp);
1560 	pg_cmt_cpu_startup(cp);
1561 
1562 	cp->cpu_flags |= CPU_RUNNING | CPU_READY | CPU_EXISTS;
1563 
1564 	if (dtrace_cpu_init != NULL) {
1565 		(*dtrace_cpu_init)(cp->cpu_id);
1566 	}
1567 
1568 #if !defined(__xpv)
1569 	/*
1570 	 * Fill out cpu_ucode_info.  Update microcode if necessary.
1571 	 */
1572 	ucode_check(cp);
1573 #endif
1574 
1575 	mutex_exit(&cpu_lock);
1576 
1577 	/*
1578 	 * Enable preemption here so that contention for any locks acquired
1579 	 * later in mp_startup may be preempted if the thread owning those
1580 	 * locks is continously executing on other CPUs (for example, this
1581 	 * CPU must be preemptible to allow other CPUs to pause it during their
1582 	 * startup phases).  It's safe to enable preemption here because the
1583 	 * CPU state is pretty-much fully constructed.
1584 	 */
1585 	curthread->t_preempt = 0;
1586 
1587 	/* The base spl should still be at LOCK LEVEL here */
1588 	ASSERT(cp->cpu_base_spl == ipltospl(LOCK_LEVEL));
1589 	set_base_spl();		/* Restore the spl to its proper value */
1590 
1591 	/* Enable interrupts */
1592 	(void) spl0();
1593 	mutex_enter(&cpu_lock);
1594 	cpu_enable_intr(cp);
1595 	cpu_add_active(cp);
1596 	mutex_exit(&cpu_lock);
1597 
1598 	add_cpunode2devtree(cp->cpu_id, cp->cpu_m.mcpu_cpi);
1599 
1600 #ifndef __xpv
1601 	{
1602 		/*
1603 		 * Set up the CPU module for this CPU.  This can't be done
1604 		 * before this CPU is made CPU_READY, because we may (in
1605 		 * heterogeneous systems) need to go load another CPU module.
1606 		 * The act of attempting to load a module may trigger a
1607 		 * cross-call, which will ASSERT unless this cpu is CPU_READY.
1608 		 */
1609 		cmi_hdl_t hdl;
1610 
1611 		if ((hdl = cmi_init(CMI_HDL_NATIVE, cmi_ntv_hwchipid(CPU),
1612 		    cmi_ntv_hwcoreid(CPU), cmi_ntv_hwstrandid(CPU))) != NULL) {
1613 			if (x86_feature & X86_MCA)
1614 				cmi_mca_init(hdl);
1615 		}
1616 	}
1617 #endif /* __xpv */
1618 
1619 	if (boothowto & RB_DEBUG)
1620 		kdi_cpu_init();
1621 
1622 	/*
1623 	 * Setting the bit in cpu_ready_set must be the last operation in
1624 	 * processor initialization; the boot CPU will continue to boot once
1625 	 * it sees this bit set for all active CPUs.
1626 	 */
1627 	CPUSET_ATOMIC_ADD(cpu_ready_set, cp->cpu_id);
1628 
1629 	/*
1630 	 * Because mp_startup() gets fired off after init() starts, we
1631 	 * can't use the '?' trick to do 'boot -v' printing - so we
1632 	 * always direct the 'cpu .. online' messages to the log.
1633 	 */
1634 	cmn_err(CE_CONT, "!cpu%d initialization complete - online\n",
1635 	    cp->cpu_id);
1636 
1637 	/*
1638 	 * Now we are done with the startup thread, so free it up.
1639 	 */
1640 	thread_exit();
1641 	panic("mp_startup: cannot return");
1642 	/*NOTREACHED*/
1643 }
1644 
1645 
1646 /*
1647  * Start CPU on user request.
1648  */
1649 /* ARGSUSED */
1650 int
1651 mp_cpu_start(struct cpu *cp)
1652 {
1653 	ASSERT(MUTEX_HELD(&cpu_lock));
1654 	return (0);
1655 }
1656 
1657 /*
1658  * Stop CPU on user request.
1659  */
1660 /* ARGSUSED */
1661 int
1662 mp_cpu_stop(struct cpu *cp)
1663 {
1664 	extern int cbe_psm_timer_mode;
1665 	ASSERT(MUTEX_HELD(&cpu_lock));
1666 
1667 #ifdef __xpv
1668 	/*
1669 	 * We can't offline vcpu0.
1670 	 */
1671 	if (cp->cpu_id == 0)
1672 		return (EBUSY);
1673 #endif
1674 
1675 	/*
1676 	 * If TIMER_PERIODIC mode is used, CPU0 is the one running it;
1677 	 * can't stop it.  (This is true only for machines with no TSC.)
1678 	 */
1679 
1680 	if ((cbe_psm_timer_mode == TIMER_PERIODIC) && (cp->cpu_id == 0))
1681 		return (EBUSY);
1682 
1683 	return (0);
1684 }
1685 
1686 /*
1687  * Take the specified CPU out of participation in interrupts.
1688  */
1689 int
1690 cpu_disable_intr(struct cpu *cp)
1691 {
1692 	if (psm_disable_intr(cp->cpu_id) != DDI_SUCCESS)
1693 		return (EBUSY);
1694 
1695 	cp->cpu_flags &= ~CPU_ENABLE;
1696 	return (0);
1697 }
1698 
1699 /*
1700  * Allow the specified CPU to participate in interrupts.
1701  */
1702 void
1703 cpu_enable_intr(struct cpu *cp)
1704 {
1705 	ASSERT(MUTEX_HELD(&cpu_lock));
1706 	cp->cpu_flags |= CPU_ENABLE;
1707 	psm_enable_intr(cp->cpu_id);
1708 }
1709 
1710 
1711 /*ARGSUSED*/
1712 void
1713 mp_cpu_faulted_enter(struct cpu *cp)
1714 {
1715 #ifndef __xpv
1716 	cmi_hdl_t hdl = cmi_hdl_lookup(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp),
1717 	    cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp));
1718 
1719 	if (hdl != NULL) {
1720 		cmi_faulted_enter(hdl);
1721 		cmi_hdl_rele(hdl);
1722 	}
1723 #endif
1724 }
1725 
1726 /*ARGSUSED*/
1727 void
1728 mp_cpu_faulted_exit(struct cpu *cp)
1729 {
1730 #ifndef __xpv
1731 	cmi_hdl_t hdl = cmi_hdl_lookup(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp),
1732 	    cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp));
1733 
1734 	if (hdl != NULL) {
1735 		cmi_faulted_exit(hdl);
1736 		cmi_hdl_rele(hdl);
1737 	}
1738 #endif
1739 }
1740 
1741 /*
1742  * The following two routines are used as context operators on threads belonging
1743  * to processes with a private LDT (see sysi86).  Due to the rarity of such
1744  * processes, these routines are currently written for best code readability and
1745  * organization rather than speed.  We could avoid checking x86_feature at every
1746  * context switch by installing different context ops, depending on the
1747  * x86_feature flags, at LDT creation time -- one for each combination of fast
1748  * syscall feature flags.
1749  */
1750 
1751 /*ARGSUSED*/
1752 void
1753 cpu_fast_syscall_disable(void *arg)
1754 {
1755 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP))
1756 		cpu_sep_disable();
1757 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC))
1758 		cpu_asysc_disable();
1759 }
1760 
1761 /*ARGSUSED*/
1762 void
1763 cpu_fast_syscall_enable(void *arg)
1764 {
1765 	if ((x86_feature & (X86_MSR | X86_SEP)) == (X86_MSR | X86_SEP))
1766 		cpu_sep_enable();
1767 	if ((x86_feature & (X86_MSR | X86_ASYSC)) == (X86_MSR | X86_ASYSC))
1768 		cpu_asysc_enable();
1769 }
1770 
1771 static void
1772 cpu_sep_enable(void)
1773 {
1774 	ASSERT(x86_feature & X86_SEP);
1775 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1776 
1777 	wrmsr(MSR_INTC_SEP_CS, (uint64_t)(uintptr_t)KCS_SEL);
1778 }
1779 
1780 static void
1781 cpu_sep_disable(void)
1782 {
1783 	ASSERT(x86_feature & X86_SEP);
1784 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1785 
1786 	/*
1787 	 * Setting the SYSENTER_CS_MSR register to 0 causes software executing
1788 	 * the sysenter or sysexit instruction to trigger a #gp fault.
1789 	 */
1790 	wrmsr(MSR_INTC_SEP_CS, 0);
1791 }
1792 
1793 static void
1794 cpu_asysc_enable(void)
1795 {
1796 	ASSERT(x86_feature & X86_ASYSC);
1797 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1798 
1799 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) |
1800 	    (uint64_t)(uintptr_t)AMD_EFER_SCE);
1801 }
1802 
1803 static void
1804 cpu_asysc_disable(void)
1805 {
1806 	ASSERT(x86_feature & X86_ASYSC);
1807 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
1808 
1809 	/*
1810 	 * Turn off the SCE (syscall enable) bit in the EFER register. Software
1811 	 * executing syscall or sysret with this bit off will incur a #ud trap.
1812 	 */
1813 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) &
1814 	    ~((uint64_t)(uintptr_t)AMD_EFER_SCE));
1815 }
1816