xref: /illumos-gate/usr/src/uts/i86pc/os/mp_startup.c (revision 7a6d80f1660abd4755c68cbd094d4a914681d26e)
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 (c) 1992, 2010, Oracle and/or its affiliates. All rights reserved.
24  */
25 /*
26  * Copyright (c) 2010, Intel Corporation.
27  * All rights reserved.
28  */
29 /*
30  * Copyright 2020 Joyent, Inc.
31  * Copyright 2013 Nexenta Systems, Inc.  All rights reserved.
32  * Copyright 2018 OmniOS Community Edition (OmniOSce) Association.
33  * Copyright 2022 Oxide Computer Company
34  */
35 
36 #include <sys/types.h>
37 #include <sys/thread.h>
38 #include <sys/cpuvar.h>
39 #include <sys/cpu.h>
40 #include <sys/t_lock.h>
41 #include <sys/param.h>
42 #include <sys/proc.h>
43 #include <sys/disp.h>
44 #include <sys/class.h>
45 #include <sys/cmn_err.h>
46 #include <sys/debug.h>
47 #include <sys/note.h>
48 #include <sys/asm_linkage.h>
49 #include <sys/x_call.h>
50 #include <sys/systm.h>
51 #include <sys/var.h>
52 #include <sys/vtrace.h>
53 #include <vm/hat.h>
54 #include <vm/as.h>
55 #include <vm/seg_kmem.h>
56 #include <vm/seg_kp.h>
57 #include <sys/segments.h>
58 #include <sys/kmem.h>
59 #include <sys/stack.h>
60 #include <sys/smp_impldefs.h>
61 #include <sys/x86_archext.h>
62 #include <sys/machsystm.h>
63 #include <sys/traptrace.h>
64 #include <sys/clock.h>
65 #include <sys/cpc_impl.h>
66 #include <sys/pg.h>
67 #include <sys/cmt.h>
68 #include <sys/dtrace.h>
69 #include <sys/archsystm.h>
70 #include <sys/fp.h>
71 #include <sys/reboot.h>
72 #include <sys/kdi_machimpl.h>
73 #include <vm/hat_i86.h>
74 #include <vm/vm_dep.h>
75 #include <sys/memnode.h>
76 #include <sys/pci_cfgspace.h>
77 #include <sys/mach_mmu.h>
78 #include <sys/sysmacros.h>
79 #if defined(__xpv)
80 #include <sys/hypervisor.h>
81 #else
82 #include <sys/hma.h>
83 #endif
84 #include <sys/cpu_module.h>
85 #include <sys/ontrap.h>
86 
87 struct cpu	cpus[1] __aligned(MMU_PAGESIZE);
88 struct cpu	*cpu[NCPU] = {&cpus[0]};
89 struct cpu	*cpu_free_list;
90 cpu_core_t	cpu_core[NCPU];
91 
92 #define	cpu_next_free	cpu_prev
93 
94 /*
95  * Useful for disabling MP bring-up on a MP capable system.
96  */
97 int use_mp = 1;
98 
99 /*
100  * to be set by a PSM to indicate what cpus
101  * are sitting around on the system.
102  */
103 cpuset_t mp_cpus;
104 
105 /*
106  * This variable is used by the hat layer to decide whether or not
107  * critical sections are needed to prevent race conditions.  For sun4m,
108  * this variable is set once enough MP initialization has been done in
109  * order to allow cross calls.
110  */
111 int flushes_require_xcalls;
112 
113 cpuset_t cpu_ready_set;		/* initialized in startup() */
114 
115 static void mp_startup_boot(void);
116 static void mp_startup_hotplug(void);
117 
118 static void cpu_sep_enable(void);
119 static void cpu_sep_disable(void);
120 static void cpu_asysc_enable(void);
121 static void cpu_asysc_disable(void);
122 
123 /*
124  * Init CPU info - get CPU type info for processor_info system call.
125  */
126 void
127 init_cpu_info(struct cpu *cp)
128 {
129 	processor_info_t *pi = &cp->cpu_type_info;
130 
131 	/*
132 	 * Get clock-frequency property for the CPU.
133 	 */
134 	pi->pi_clock = cpu_freq;
135 
136 	/*
137 	 * Current frequency in Hz.
138 	 */
139 	cp->cpu_curr_clock = cpu_freq_hz;
140 
141 	/*
142 	 * Supported frequencies.
143 	 */
144 	if (cp->cpu_supp_freqs == NULL) {
145 		cpu_set_supp_freqs(cp, NULL);
146 	}
147 
148 	(void) strcpy(pi->pi_processor_type, "i386");
149 	if (fpu_exists)
150 		(void) strcpy(pi->pi_fputypes, "i387 compatible");
151 
152 	cp->cpu_idstr = kmem_zalloc(CPU_IDSTRLEN, KM_SLEEP);
153 	cp->cpu_brandstr = kmem_zalloc(CPU_IDSTRLEN, KM_SLEEP);
154 
155 	/*
156 	 * If called for the BSP, cp is equal to current CPU.
157 	 * For non-BSPs, cpuid info of cp is not ready yet, so use cpuid info
158 	 * of current CPU as default values for cpu_idstr and cpu_brandstr.
159 	 * They will be corrected in mp_startup_common() after
160 	 * CPUID_PASS_DYNAMIC has been invoked on target CPU.
161 	 */
162 	(void) cpuid_getidstr(CPU, cp->cpu_idstr, CPU_IDSTRLEN);
163 	(void) cpuid_getbrandstr(CPU, cp->cpu_brandstr, CPU_IDSTRLEN);
164 }
165 
166 /*
167  * Configure syscall support on this CPU.
168  */
169 /*ARGSUSED*/
170 void
171 init_cpu_syscall(struct cpu *cp)
172 {
173 	kpreempt_disable();
174 
175 	if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
176 	    is_x86_feature(x86_featureset, X86FSET_ASYSC)) {
177 		uint64_t flags;
178 
179 #if !defined(__xpv)
180 		/*
181 		 * The syscall instruction imposes a certain ordering on
182 		 * segment selectors, so we double-check that ordering
183 		 * here.
184 		 */
185 		CTASSERT(KDS_SEL == KCS_SEL + 8);
186 		CTASSERT(UDS_SEL == U32CS_SEL + 8);
187 		CTASSERT(UCS_SEL == U32CS_SEL + 16);
188 #endif
189 
190 		/*
191 		 * Turn syscall/sysret extensions on.
192 		 */
193 		cpu_asysc_enable();
194 
195 		/*
196 		 * Program the magic registers ..
197 		 */
198 		wrmsr(MSR_AMD_STAR,
199 		    ((uint64_t)(U32CS_SEL << 16 | KCS_SEL)) << 32);
200 		if (kpti_enable == 1) {
201 			wrmsr(MSR_AMD_LSTAR,
202 			    (uint64_t)(uintptr_t)tr_sys_syscall);
203 			wrmsr(MSR_AMD_CSTAR,
204 			    (uint64_t)(uintptr_t)tr_sys_syscall32);
205 		} else {
206 			wrmsr(MSR_AMD_LSTAR,
207 			    (uint64_t)(uintptr_t)sys_syscall);
208 			wrmsr(MSR_AMD_CSTAR,
209 			    (uint64_t)(uintptr_t)sys_syscall32);
210 		}
211 
212 		/*
213 		 * This list of flags is masked off the incoming
214 		 * %rfl when we enter the kernel.
215 		 */
216 		flags = PS_IE | PS_T;
217 		if (is_x86_feature(x86_featureset, X86FSET_SMAP) == B_TRUE)
218 			flags |= PS_ACHK;
219 		wrmsr(MSR_AMD_SFMASK, flags);
220 	}
221 
222 	/*
223 	 * On 64-bit kernels on Nocona machines, the 32-bit syscall
224 	 * variant isn't available to 32-bit applications, but sysenter is.
225 	 */
226 	if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
227 	    is_x86_feature(x86_featureset, X86FSET_SEP)) {
228 
229 #if !defined(__xpv)
230 		/*
231 		 * The sysenter instruction imposes a certain ordering on
232 		 * segment selectors, so we double-check that ordering
233 		 * here. See "sysenter" in Intel document 245471-012, "IA-32
234 		 * Intel Architecture Software Developer's Manual Volume 2:
235 		 * Instruction Set Reference"
236 		 */
237 		CTASSERT(KDS_SEL == KCS_SEL + 8);
238 
239 		CTASSERT(U32CS_SEL == ((KCS_SEL + 16) | 3));
240 		CTASSERT(UDS_SEL == U32CS_SEL + 8);
241 #endif
242 
243 		cpu_sep_enable();
244 
245 		/*
246 		 * resume() sets this value to the base of the threads stack
247 		 * via a context handler.
248 		 */
249 		wrmsr(MSR_INTC_SEP_ESP, 0);
250 
251 		if (kpti_enable == 1) {
252 			wrmsr(MSR_INTC_SEP_EIP,
253 			    (uint64_t)(uintptr_t)tr_sys_sysenter);
254 		} else {
255 			wrmsr(MSR_INTC_SEP_EIP,
256 			    (uint64_t)(uintptr_t)sys_sysenter);
257 		}
258 	}
259 
260 	kpreempt_enable();
261 }
262 
263 #if !defined(__xpv)
264 /*
265  * Configure per-cpu ID GDT
266  */
267 static void
268 init_cpu_id_gdt(struct cpu *cp)
269 {
270 	/* Write cpu_id into limit field of GDT for usermode retrieval */
271 	set_usegd(&cp->cpu_gdt[GDT_CPUID], SDP_SHORT, NULL, cp->cpu_id,
272 	    SDT_MEMRODA, SEL_UPL, SDP_BYTES, SDP_OP32);
273 }
274 #endif /* !defined(__xpv) */
275 
276 /*
277  * Multiprocessor initialization.
278  *
279  * Allocate and initialize the cpu structure, TRAPTRACE buffer, and the
280  * startup and idle threads for the specified CPU.
281  * Parameter boot is true for boot time operations and is false for CPU
282  * DR operations.
283  */
284 static struct cpu *
285 mp_cpu_configure_common(int cpun, boolean_t boot)
286 {
287 	struct cpu *cp;
288 	kthread_id_t tp;
289 	caddr_t	sp;
290 	proc_t *procp;
291 #if !defined(__xpv)
292 	extern int idle_cpu_prefer_mwait;
293 	extern void cpu_idle_mwait();
294 #endif
295 	extern void idle();
296 	extern void cpu_idle();
297 
298 #ifdef TRAPTRACE
299 	trap_trace_ctl_t *ttc = &trap_trace_ctl[cpun];
300 #endif
301 
302 	ASSERT(MUTEX_HELD(&cpu_lock));
303 	ASSERT(cpun < NCPU && cpu[cpun] == NULL);
304 
305 	if (cpu_free_list == NULL) {
306 		cp = kmem_zalloc(sizeof (*cp), KM_SLEEP);
307 	} else {
308 		cp = cpu_free_list;
309 		cpu_free_list = cp->cpu_next_free;
310 	}
311 
312 	cp->cpu_m.mcpu_istamp = cpun << 16;
313 
314 	/* Create per CPU specific threads in the process p0. */
315 	procp = &p0;
316 
317 	/*
318 	 * Initialize the dispatcher first.
319 	 */
320 	disp_cpu_init(cp);
321 
322 	cpu_vm_data_init(cp);
323 
324 	/*
325 	 * Allocate and initialize the startup thread for this CPU.
326 	 * Interrupt and process switch stacks get allocated later
327 	 * when the CPU starts running.
328 	 */
329 	tp = thread_create(NULL, 0, NULL, NULL, 0, procp,
330 	    TS_STOPPED, maxclsyspri);
331 
332 	/*
333 	 * Set state to TS_ONPROC since this thread will start running
334 	 * as soon as the CPU comes online.
335 	 *
336 	 * All the other fields of the thread structure are setup by
337 	 * thread_create().
338 	 */
339 	THREAD_ONPROC(tp, cp);
340 	tp->t_preempt = 1;
341 	tp->t_bound_cpu = cp;
342 	tp->t_affinitycnt = 1;
343 	tp->t_cpu = cp;
344 	tp->t_disp_queue = cp->cpu_disp;
345 
346 	/*
347 	 * Setup thread to start in mp_startup_common.
348 	 */
349 	sp = tp->t_stk;
350 	tp->t_sp = (uintptr_t)(sp - MINFRAME);
351 	tp->t_sp -= STACK_ENTRY_ALIGN;		/* fake a call */
352 	/*
353 	 * Setup thread start entry point for boot or hotplug.
354 	 */
355 	if (boot) {
356 		tp->t_pc = (uintptr_t)mp_startup_boot;
357 	} else {
358 		tp->t_pc = (uintptr_t)mp_startup_hotplug;
359 	}
360 
361 	cp->cpu_id = cpun;
362 	cp->cpu_self = cp;
363 	cp->cpu_thread = tp;
364 	cp->cpu_lwp = NULL;
365 	cp->cpu_dispthread = tp;
366 	cp->cpu_dispatch_pri = DISP_PRIO(tp);
367 
368 	/*
369 	 * cpu_base_spl must be set explicitly here to prevent any blocking
370 	 * operations in mp_startup_common from causing the spl of the cpu
371 	 * to drop to 0 (allowing device interrupts before we're ready) in
372 	 * resume().
373 	 * cpu_base_spl MUST remain at LOCK_LEVEL until the cpu is CPU_READY.
374 	 * As an extra bit of security on DEBUG kernels, this is enforced with
375 	 * an assertion in mp_startup_common() -- before cpu_base_spl is set
376 	 * to its proper value.
377 	 */
378 	cp->cpu_base_spl = ipltospl(LOCK_LEVEL);
379 
380 	/*
381 	 * Now, initialize per-CPU idle thread for this CPU.
382 	 */
383 	tp = thread_create(NULL, PAGESIZE, idle, NULL, 0, procp, TS_ONPROC, -1);
384 
385 	cp->cpu_idle_thread = tp;
386 
387 	tp->t_preempt = 1;
388 	tp->t_bound_cpu = cp;
389 	tp->t_affinitycnt = 1;
390 	tp->t_cpu = cp;
391 	tp->t_disp_queue = cp->cpu_disp;
392 
393 	/*
394 	 * Bootstrap the CPU's PG data
395 	 */
396 	pg_cpu_bootstrap(cp);
397 
398 	/*
399 	 * Perform CPC initialization on the new CPU.
400 	 */
401 	kcpc_hw_init(cp);
402 
403 	/*
404 	 * Allocate virtual addresses for cpu_caddr1 and cpu_caddr2
405 	 * for each CPU.
406 	 */
407 	setup_vaddr_for_ppcopy(cp);
408 
409 	/*
410 	 * Allocate page for new GDT and initialize from current GDT.
411 	 */
412 #if !defined(__lint)
413 	ASSERT((sizeof (*cp->cpu_gdt) * NGDT) <= PAGESIZE);
414 #endif
415 	cp->cpu_gdt = kmem_zalloc(PAGESIZE, KM_SLEEP);
416 	bcopy(CPU->cpu_gdt, cp->cpu_gdt, (sizeof (*cp->cpu_gdt) * NGDT));
417 
418 
419 	/*
420 	 * Allocate pages for the CPU LDT.
421 	 */
422 	cp->cpu_m.mcpu_ldt = kmem_zalloc(LDT_CPU_SIZE, KM_SLEEP);
423 	cp->cpu_m.mcpu_ldt_len = 0;
424 
425 	/*
426 	 * Allocate a per-CPU IDT and initialize the new IDT to the currently
427 	 * runing CPU.
428 	 */
429 #if !defined(__lint)
430 	ASSERT((sizeof (*CPU->cpu_idt) * NIDT) <= PAGESIZE);
431 #endif
432 	cp->cpu_idt = kmem_alloc(PAGESIZE, KM_SLEEP);
433 	bcopy(CPU->cpu_idt, cp->cpu_idt, PAGESIZE);
434 
435 	/*
436 	 * alloc space for cpuid info
437 	 */
438 	cpuid_alloc_space(cp);
439 #if !defined(__xpv)
440 	if (is_x86_feature(x86_featureset, X86FSET_MWAIT) &&
441 	    idle_cpu_prefer_mwait) {
442 		cp->cpu_m.mcpu_mwait = cpuid_mwait_alloc(cp);
443 		cp->cpu_m.mcpu_idle_cpu = cpu_idle_mwait;
444 	} else
445 #endif
446 		cp->cpu_m.mcpu_idle_cpu = cpu_idle;
447 
448 	init_cpu_info(cp);
449 
450 #if !defined(__xpv)
451 	init_cpu_id_gdt(cp);
452 #endif
453 
454 	/*
455 	 * alloc space for ucode_info
456 	 */
457 	ucode_alloc_space(cp);
458 	xc_init_cpu(cp);
459 	hat_cpu_online(cp);
460 
461 #ifdef TRAPTRACE
462 	/*
463 	 * If this is a TRAPTRACE kernel, allocate TRAPTRACE buffers
464 	 */
465 	ttc->ttc_first = (uintptr_t)kmem_zalloc(trap_trace_bufsize, KM_SLEEP);
466 	ttc->ttc_next = ttc->ttc_first;
467 	ttc->ttc_limit = ttc->ttc_first + trap_trace_bufsize;
468 #endif
469 
470 	/*
471 	 * Record that we have another CPU.
472 	 */
473 	/*
474 	 * Initialize the interrupt threads for this CPU
475 	 */
476 	cpu_intr_alloc(cp, NINTR_THREADS);
477 
478 	cp->cpu_flags = CPU_OFFLINE | CPU_QUIESCED | CPU_POWEROFF;
479 	cpu_set_state(cp);
480 
481 	/*
482 	 * Add CPU to list of available CPUs.  It'll be on the active list
483 	 * after mp_startup_common().
484 	 */
485 	cpu_add_unit(cp);
486 
487 	return (cp);
488 }
489 
490 /*
491  * Undo what was done in mp_cpu_configure_common
492  */
493 static void
494 mp_cpu_unconfigure_common(struct cpu *cp, int error)
495 {
496 	ASSERT(MUTEX_HELD(&cpu_lock));
497 
498 	/*
499 	 * Remove the CPU from the list of available CPUs.
500 	 */
501 	cpu_del_unit(cp->cpu_id);
502 
503 	if (error == ETIMEDOUT) {
504 		/*
505 		 * The cpu was started, but never *seemed* to run any
506 		 * code in the kernel; it's probably off spinning in its
507 		 * own private world, though with potential references to
508 		 * our kmem-allocated IDTs and GDTs (for example).
509 		 *
510 		 * Worse still, it may actually wake up some time later,
511 		 * so rather than guess what it might or might not do, we
512 		 * leave the fundamental data structures intact.
513 		 */
514 		cp->cpu_flags = 0;
515 		return;
516 	}
517 
518 	/*
519 	 * At this point, the only threads bound to this CPU should
520 	 * special per-cpu threads: it's idle thread, it's pause threads,
521 	 * and it's interrupt threads.  Clean these up.
522 	 */
523 	cpu_destroy_bound_threads(cp);
524 	cp->cpu_idle_thread = NULL;
525 
526 	/*
527 	 * Free the interrupt stack.
528 	 */
529 	segkp_release(segkp,
530 	    cp->cpu_intr_stack - (INTR_STACK_SIZE - SA(MINFRAME)));
531 	cp->cpu_intr_stack = NULL;
532 
533 #ifdef TRAPTRACE
534 	/*
535 	 * Discard the trap trace buffer
536 	 */
537 	{
538 		trap_trace_ctl_t *ttc = &trap_trace_ctl[cp->cpu_id];
539 
540 		kmem_free((void *)ttc->ttc_first, trap_trace_bufsize);
541 		ttc->ttc_first = (uintptr_t)NULL;
542 	}
543 #endif
544 
545 	hat_cpu_offline(cp);
546 
547 	ucode_free_space(cp);
548 
549 	/* Free CPU ID string and brand string. */
550 	if (cp->cpu_idstr) {
551 		kmem_free(cp->cpu_idstr, CPU_IDSTRLEN);
552 		cp->cpu_idstr = NULL;
553 	}
554 	if (cp->cpu_brandstr) {
555 		kmem_free(cp->cpu_brandstr, CPU_IDSTRLEN);
556 		cp->cpu_brandstr = NULL;
557 	}
558 
559 #if !defined(__xpv)
560 	if (cp->cpu_m.mcpu_mwait != NULL) {
561 		cpuid_mwait_free(cp);
562 		cp->cpu_m.mcpu_mwait = NULL;
563 	}
564 #endif
565 	cpuid_free_space(cp);
566 
567 	if (cp->cpu_idt != CPU->cpu_idt)
568 		kmem_free(cp->cpu_idt, PAGESIZE);
569 	cp->cpu_idt = NULL;
570 
571 	kmem_free(cp->cpu_m.mcpu_ldt, LDT_CPU_SIZE);
572 	cp->cpu_m.mcpu_ldt = NULL;
573 	cp->cpu_m.mcpu_ldt_len = 0;
574 
575 	kmem_free(cp->cpu_gdt, PAGESIZE);
576 	cp->cpu_gdt = NULL;
577 
578 	if (cp->cpu_supp_freqs != NULL) {
579 		size_t len = strlen(cp->cpu_supp_freqs) + 1;
580 		kmem_free(cp->cpu_supp_freqs, len);
581 		cp->cpu_supp_freqs = NULL;
582 	}
583 
584 	teardown_vaddr_for_ppcopy(cp);
585 
586 	kcpc_hw_fini(cp);
587 
588 	cp->cpu_dispthread = NULL;
589 	cp->cpu_thread = NULL;	/* discarded by cpu_destroy_bound_threads() */
590 
591 	cpu_vm_data_destroy(cp);
592 
593 	xc_fini_cpu(cp);
594 	disp_cpu_fini(cp);
595 
596 	ASSERT(cp != CPU0);
597 	bzero(cp, sizeof (*cp));
598 	cp->cpu_next_free = cpu_free_list;
599 	cpu_free_list = cp;
600 }
601 
602 /*
603  * Apply workarounds for known errata, and warn about those that are absent.
604  *
605  * System vendors occasionally create configurations which contain different
606  * revisions of the CPUs that are almost but not exactly the same.  At the
607  * time of writing, this meant that their clock rates were the same, their
608  * feature sets were the same, but the required workaround were -not-
609  * necessarily the same.  So, this routine is invoked on -every- CPU soon
610  * after starting to make sure that the resulting system contains the most
611  * pessimal set of workarounds needed to cope with *any* of the CPUs in the
612  * system.
613  *
614  * workaround_errata is invoked early in mlsetup() for CPU 0, and in
615  * mp_startup_common() for all slave CPUs. Slaves process workaround_errata
616  * prior to acknowledging their readiness to the master, so this routine will
617  * never be executed by multiple CPUs in parallel, thus making updates to
618  * global data safe.
619  *
620  * These workarounds are based on Rev 3.57 of the Revision Guide for
621  * AMD Athlon(tm) 64 and AMD Opteron(tm) Processors, August 2005.
622  */
623 
624 #if defined(OPTERON_ERRATUM_88)
625 int opteron_erratum_88;		/* if non-zero -> at least one cpu has it */
626 #endif
627 
628 #if defined(OPTERON_ERRATUM_91)
629 int opteron_erratum_91;		/* if non-zero -> at least one cpu has it */
630 #endif
631 
632 #if defined(OPTERON_ERRATUM_93)
633 int opteron_erratum_93;		/* if non-zero -> at least one cpu has it */
634 #endif
635 
636 #if defined(OPTERON_ERRATUM_95)
637 int opteron_erratum_95;		/* if non-zero -> at least one cpu has it */
638 #endif
639 
640 #if defined(OPTERON_ERRATUM_100)
641 int opteron_erratum_100;	/* if non-zero -> at least one cpu has it */
642 #endif
643 
644 #if defined(OPTERON_ERRATUM_108)
645 int opteron_erratum_108;	/* if non-zero -> at least one cpu has it */
646 #endif
647 
648 #if defined(OPTERON_ERRATUM_109)
649 int opteron_erratum_109;	/* if non-zero -> at least one cpu has it */
650 #endif
651 
652 #if defined(OPTERON_ERRATUM_121)
653 int opteron_erratum_121;	/* if non-zero -> at least one cpu has it */
654 #endif
655 
656 #if defined(OPTERON_ERRATUM_122)
657 int opteron_erratum_122;	/* if non-zero -> at least one cpu has it */
658 #endif
659 
660 #if defined(OPTERON_ERRATUM_123)
661 int opteron_erratum_123;	/* if non-zero -> at least one cpu has it */
662 #endif
663 
664 #if defined(OPTERON_ERRATUM_131)
665 int opteron_erratum_131;	/* if non-zero -> at least one cpu has it */
666 #endif
667 
668 #if defined(OPTERON_WORKAROUND_6336786)
669 int opteron_workaround_6336786;	/* non-zero -> WA relevant and applied */
670 int opteron_workaround_6336786_UP = 0;	/* Not needed for UP */
671 #endif
672 
673 #if defined(OPTERON_ERRATUM_147)
674 int opteron_erratum_147;	/* if non-zero -> at least one cpu has it */
675 #endif
676 
677 #if defined(OPTERON_ERRATUM_298)
678 int opteron_erratum_298;
679 #endif
680 
681 #if defined(OPTERON_ERRATUM_721)
682 int opteron_erratum_721;
683 #endif
684 
685 static void
686 workaround_warning(cpu_t *cp, uint_t erratum)
687 {
688 	cmn_err(CE_WARN, "cpu%d: no workaround for erratum %u",
689 	    cp->cpu_id, erratum);
690 }
691 
692 static void
693 workaround_applied(uint_t erratum)
694 {
695 	if (erratum > 1000000)
696 		cmn_err(CE_CONT, "?workaround applied for cpu issue #%d\n",
697 		    erratum);
698 	else
699 		cmn_err(CE_CONT, "?workaround applied for cpu erratum #%d\n",
700 		    erratum);
701 }
702 
703 static void
704 msr_warning(cpu_t *cp, const char *rw, uint_t msr, int error)
705 {
706 	cmn_err(CE_WARN, "cpu%d: couldn't %smsr 0x%x, error %d",
707 	    cp->cpu_id, rw, msr, error);
708 }
709 
710 /*
711  * Determine the number of nodes in a Hammer / Greyhound / Griffin family
712  * system.
713  */
714 static uint_t
715 opteron_get_nnodes(void)
716 {
717 	static uint_t nnodes = 0;
718 
719 	if (nnodes == 0) {
720 #ifdef	DEBUG
721 		uint_t family;
722 
723 		/*
724 		 * This routine uses a PCI config space based mechanism
725 		 * for retrieving the number of nodes in the system.
726 		 * Device 24, function 0, offset 0x60 as used here is not
727 		 * AMD processor architectural, and may not work on processor
728 		 * families other than those listed below.
729 		 *
730 		 * Callers of this routine must ensure that we're running on
731 		 * a processor which supports this mechanism.
732 		 * The assertion below is meant to catch calls on unsupported
733 		 * processors.
734 		 */
735 		family = cpuid_getfamily(CPU);
736 		ASSERT(family == 0xf || family == 0x10 || family == 0x11);
737 #endif	/* DEBUG */
738 
739 		/*
740 		 * Obtain the number of nodes in the system from
741 		 * bits [6:4] of the Node ID register on node 0.
742 		 *
743 		 * The actual node count is NodeID[6:4] + 1
744 		 *
745 		 * The Node ID register is accessed via function 0,
746 		 * offset 0x60. Node 0 is device 24.
747 		 */
748 		nnodes = ((pci_getl_func(0, 24, 0, 0x60) & 0x70) >> 4) + 1;
749 	}
750 	return (nnodes);
751 }
752 
753 uint_t
754 do_erratum_298(struct cpu *cpu)
755 {
756 	static int	osvwrc = -3;
757 	extern int	osvw_opteron_erratum(cpu_t *, uint_t);
758 
759 	/*
760 	 * L2 Eviction May Occur During Processor Operation To Set
761 	 * Accessed or Dirty Bit.
762 	 */
763 	if (osvwrc == -3) {
764 		osvwrc = osvw_opteron_erratum(cpu, 298);
765 	} else {
766 		/* osvw return codes should be consistent for all cpus */
767 		ASSERT(osvwrc == osvw_opteron_erratum(cpu, 298));
768 	}
769 
770 	switch (osvwrc) {
771 	case 0:		/* erratum is not present: do nothing */
772 		break;
773 	case 1:		/* erratum is present: BIOS workaround applied */
774 		/*
775 		 * check if workaround is actually in place and issue warning
776 		 * if not.
777 		 */
778 		if (((rdmsr(MSR_AMD_HWCR) & AMD_HWCR_TLBCACHEDIS) == 0) ||
779 		    ((rdmsr(MSR_AMD_BU_CFG) & AMD_BU_CFG_E298) == 0)) {
780 #if defined(OPTERON_ERRATUM_298)
781 			opteron_erratum_298++;
782 #else
783 			workaround_warning(cpu, 298);
784 			return (1);
785 #endif
786 		}
787 		break;
788 	case -1:	/* cannot determine via osvw: check cpuid */
789 		if ((cpuid_opteron_erratum(cpu, 298) > 0) &&
790 		    (((rdmsr(MSR_AMD_HWCR) & AMD_HWCR_TLBCACHEDIS) == 0) ||
791 		    ((rdmsr(MSR_AMD_BU_CFG) & AMD_BU_CFG_E298) == 0))) {
792 #if defined(OPTERON_ERRATUM_298)
793 			opteron_erratum_298++;
794 #else
795 			workaround_warning(cpu, 298);
796 			return (1);
797 #endif
798 		}
799 		break;
800 	}
801 	return (0);
802 }
803 
804 uint_t
805 workaround_errata(struct cpu *cpu)
806 {
807 	volatile uint_t missing = 0;
808 
809 	ASSERT(cpu == CPU);
810 
811 	/*LINTED*/
812 	if (cpuid_opteron_erratum(cpu, 88) > 0) {
813 		/*
814 		 * SWAPGS May Fail To Read Correct GS Base
815 		 */
816 #if defined(OPTERON_ERRATUM_88)
817 		/*
818 		 * The workaround is an mfence in the relevant assembler code
819 		 */
820 		opteron_erratum_88++;
821 #else
822 		workaround_warning(cpu, 88);
823 		missing++;
824 #endif
825 	}
826 
827 	if (cpuid_opteron_erratum(cpu, 91) > 0) {
828 		/*
829 		 * Software Prefetches May Report A Page Fault
830 		 */
831 #if defined(OPTERON_ERRATUM_91)
832 		/*
833 		 * fix is in trap.c
834 		 */
835 		opteron_erratum_91++;
836 #else
837 		workaround_warning(cpu, 91);
838 		missing++;
839 #endif
840 	}
841 
842 	if (cpuid_opteron_erratum(cpu, 93) > 0) {
843 		/*
844 		 * RSM Auto-Halt Restart Returns to Incorrect RIP
845 		 */
846 #if defined(OPTERON_ERRATUM_93)
847 		/*
848 		 * fix is in trap.c
849 		 */
850 		opteron_erratum_93++;
851 #else
852 		workaround_warning(cpu, 93);
853 		missing++;
854 #endif
855 	}
856 
857 	/*LINTED*/
858 	if (cpuid_opteron_erratum(cpu, 95) > 0) {
859 		/*
860 		 * RET Instruction May Return to Incorrect EIP
861 		 */
862 #if defined(OPTERON_ERRATUM_95)
863 #if defined(_LP64)
864 		/*
865 		 * Workaround this by ensuring that 32-bit user code and
866 		 * 64-bit kernel code never occupy the same address
867 		 * range mod 4G.
868 		 */
869 		if (_userlimit32 > 0xc0000000ul)
870 			*(uintptr_t *)&_userlimit32 = 0xc0000000ul;
871 
872 		/*LINTED*/
873 		ASSERT((uint32_t)COREHEAP_BASE == 0xc0000000u);
874 		opteron_erratum_95++;
875 #endif	/* _LP64 */
876 #else
877 		workaround_warning(cpu, 95);
878 		missing++;
879 #endif
880 	}
881 
882 	if (cpuid_opteron_erratum(cpu, 100) > 0) {
883 		/*
884 		 * Compatibility Mode Branches Transfer to Illegal Address
885 		 */
886 #if defined(OPTERON_ERRATUM_100)
887 		/*
888 		 * fix is in trap.c
889 		 */
890 		opteron_erratum_100++;
891 #else
892 		workaround_warning(cpu, 100);
893 		missing++;
894 #endif
895 	}
896 
897 	/*LINTED*/
898 	if (cpuid_opteron_erratum(cpu, 108) > 0) {
899 		/*
900 		 * CPUID Instruction May Return Incorrect Model Number In
901 		 * Some Processors
902 		 */
903 #if defined(OPTERON_ERRATUM_108)
904 		/*
905 		 * (Our cpuid-handling code corrects the model number on
906 		 * those processors)
907 		 */
908 #else
909 		workaround_warning(cpu, 108);
910 		missing++;
911 #endif
912 	}
913 
914 	/*LINTED*/
915 	if (cpuid_opteron_erratum(cpu, 109) > 0) do {
916 		/*
917 		 * Certain Reverse REP MOVS May Produce Unpredictable Behavior
918 		 */
919 #if defined(OPTERON_ERRATUM_109)
920 		/*
921 		 * The "workaround" is to print a warning to upgrade the BIOS
922 		 */
923 		uint64_t value;
924 		const uint_t msr = MSR_AMD_PATCHLEVEL;
925 		int err;
926 
927 		if ((err = checked_rdmsr(msr, &value)) != 0) {
928 			msr_warning(cpu, "rd", msr, err);
929 			workaround_warning(cpu, 109);
930 			missing++;
931 		}
932 		if (value == 0)
933 			opteron_erratum_109++;
934 #else
935 		workaround_warning(cpu, 109);
936 		missing++;
937 #endif
938 	/*CONSTANTCONDITION*/
939 	} while (0);
940 
941 	/*LINTED*/
942 	if (cpuid_opteron_erratum(cpu, 121) > 0) {
943 		/*
944 		 * Sequential Execution Across Non_Canonical Boundary Caused
945 		 * Processor Hang
946 		 */
947 #if defined(OPTERON_ERRATUM_121)
948 #if defined(_LP64)
949 		/*
950 		 * Erratum 121 is only present in long (64 bit) mode.
951 		 * Workaround is to include the page immediately before the
952 		 * va hole to eliminate the possibility of system hangs due to
953 		 * sequential execution across the va hole boundary.
954 		 */
955 		if (opteron_erratum_121)
956 			opteron_erratum_121++;
957 		else {
958 			if (hole_start) {
959 				hole_start -= PAGESIZE;
960 			} else {
961 				/*
962 				 * hole_start not yet initialized by
963 				 * mmu_init. Initialize hole_start
964 				 * with value to be subtracted.
965 				 */
966 				hole_start = PAGESIZE;
967 			}
968 			opteron_erratum_121++;
969 		}
970 #endif	/* _LP64 */
971 #else
972 		workaround_warning(cpu, 121);
973 		missing++;
974 #endif
975 	}
976 
977 	/*LINTED*/
978 	if (cpuid_opteron_erratum(cpu, 122) > 0) do {
979 		/*
980 		 * TLB Flush Filter May Cause Coherency Problem in
981 		 * Multiprocessor Systems
982 		 */
983 #if defined(OPTERON_ERRATUM_122)
984 		uint64_t value;
985 		const uint_t msr = MSR_AMD_HWCR;
986 		int error;
987 
988 		/*
989 		 * Erratum 122 is only present in MP configurations (multi-core
990 		 * or multi-processor).
991 		 */
992 #if defined(__xpv)
993 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
994 			break;
995 		if (!opteron_erratum_122 && xpv_nr_phys_cpus() == 1)
996 			break;
997 #else
998 		if (!opteron_erratum_122 && opteron_get_nnodes() == 1 &&
999 		    cpuid_get_ncpu_per_chip(cpu) == 1)
1000 			break;
1001 #endif
1002 		/* disable TLB Flush Filter */
1003 
1004 		if ((error = checked_rdmsr(msr, &value)) != 0) {
1005 			msr_warning(cpu, "rd", msr, error);
1006 			workaround_warning(cpu, 122);
1007 			missing++;
1008 		} else {
1009 			value |= (uint64_t)AMD_HWCR_FFDIS;
1010 			if ((error = checked_wrmsr(msr, value)) != 0) {
1011 				msr_warning(cpu, "wr", msr, error);
1012 				workaround_warning(cpu, 122);
1013 				missing++;
1014 			}
1015 		}
1016 		opteron_erratum_122++;
1017 #else
1018 		workaround_warning(cpu, 122);
1019 		missing++;
1020 #endif
1021 	/*CONSTANTCONDITION*/
1022 	} while (0);
1023 
1024 	/*LINTED*/
1025 	if (cpuid_opteron_erratum(cpu, 123) > 0) do {
1026 		/*
1027 		 * Bypassed Reads May Cause Data Corruption of System Hang in
1028 		 * Dual Core Processors
1029 		 */
1030 #if defined(OPTERON_ERRATUM_123)
1031 		uint64_t value;
1032 		const uint_t msr = MSR_AMD_PATCHLEVEL;
1033 		int err;
1034 
1035 		/*
1036 		 * Erratum 123 applies only to multi-core cpus.
1037 		 */
1038 		if (cpuid_get_ncpu_per_chip(cpu) < 2)
1039 			break;
1040 #if defined(__xpv)
1041 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
1042 			break;
1043 #endif
1044 		/*
1045 		 * The "workaround" is to print a warning to upgrade the BIOS
1046 		 */
1047 		if ((err = checked_rdmsr(msr, &value)) != 0) {
1048 			msr_warning(cpu, "rd", msr, err);
1049 			workaround_warning(cpu, 123);
1050 			missing++;
1051 		}
1052 		if (value == 0)
1053 			opteron_erratum_123++;
1054 #else
1055 		workaround_warning(cpu, 123);
1056 		missing++;
1057 
1058 #endif
1059 	/*CONSTANTCONDITION*/
1060 	} while (0);
1061 
1062 	/*LINTED*/
1063 	if (cpuid_opteron_erratum(cpu, 131) > 0) do {
1064 		/*
1065 		 * Multiprocessor Systems with Four or More Cores May Deadlock
1066 		 * Waiting for a Probe Response
1067 		 */
1068 #if defined(OPTERON_ERRATUM_131)
1069 		uint64_t nbcfg;
1070 		const uint_t msr = MSR_AMD_NB_CFG;
1071 		const uint64_t wabits =
1072 		    AMD_NB_CFG_SRQ_HEARTBEAT | AMD_NB_CFG_SRQ_SPR;
1073 		int error;
1074 
1075 		/*
1076 		 * Erratum 131 applies to any system with four or more cores.
1077 		 */
1078 		if (opteron_erratum_131)
1079 			break;
1080 #if defined(__xpv)
1081 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
1082 			break;
1083 		if (xpv_nr_phys_cpus() < 4)
1084 			break;
1085 #else
1086 		if (opteron_get_nnodes() * cpuid_get_ncpu_per_chip(cpu) < 4)
1087 			break;
1088 #endif
1089 		/*
1090 		 * Print a warning if neither of the workarounds for
1091 		 * erratum 131 is present.
1092 		 */
1093 		if ((error = checked_rdmsr(msr, &nbcfg)) != 0) {
1094 			msr_warning(cpu, "rd", msr, error);
1095 			workaround_warning(cpu, 131);
1096 			missing++;
1097 		} else if ((nbcfg & wabits) == 0) {
1098 			opteron_erratum_131++;
1099 		} else {
1100 			/* cannot have both workarounds set */
1101 			ASSERT((nbcfg & wabits) != wabits);
1102 		}
1103 #else
1104 		workaround_warning(cpu, 131);
1105 		missing++;
1106 #endif
1107 	/*CONSTANTCONDITION*/
1108 	} while (0);
1109 
1110 	/*
1111 	 * This isn't really an erratum, but for convenience the
1112 	 * detection/workaround code lives here and in cpuid_opteron_erratum.
1113 	 * Note, the technique only is valid on families before 12h and
1114 	 * certainly doesn't work when we're virtualized. This is checked for in
1115 	 * the erratum workaround.
1116 	 */
1117 	if (cpuid_opteron_erratum(cpu, 6336786) > 0) {
1118 #if defined(OPTERON_WORKAROUND_6336786)
1119 		/*
1120 		 * Disable C1-Clock ramping on multi-core/multi-processor
1121 		 * K8 platforms to guard against TSC drift.
1122 		 */
1123 		if (opteron_workaround_6336786) {
1124 			opteron_workaround_6336786++;
1125 #if defined(__xpv)
1126 		} else if ((DOMAIN_IS_INITDOMAIN(xen_info) &&
1127 		    xpv_nr_phys_cpus() > 1) ||
1128 		    opteron_workaround_6336786_UP) {
1129 			/*
1130 			 * XXPV	Hmm.  We can't walk the Northbridges on
1131 			 *	the hypervisor; so just complain and drive
1132 			 *	on.  This probably needs to be fixed in
1133 			 *	the hypervisor itself.
1134 			 */
1135 			opteron_workaround_6336786++;
1136 			workaround_warning(cpu, 6336786);
1137 #else	/* __xpv */
1138 		} else if ((opteron_get_nnodes() *
1139 		    cpuid_get_ncpu_per_chip(cpu) > 1) ||
1140 		    opteron_workaround_6336786_UP) {
1141 
1142 			uint_t	node, nnodes;
1143 			uint8_t data;
1144 
1145 			nnodes = opteron_get_nnodes();
1146 			for (node = 0; node < nnodes; node++) {
1147 				/*
1148 				 * Clear PMM7[1:0] (function 3, offset 0x87)
1149 				 * Northbridge device is the node id + 24.
1150 				 */
1151 				data = pci_getb_func(0, node + 24, 3, 0x87);
1152 				data &= 0xFC;
1153 				pci_putb_func(0, node + 24, 3, 0x87, data);
1154 			}
1155 			opteron_workaround_6336786++;
1156 #endif	/* __xpv */
1157 		}
1158 #else
1159 		workaround_warning(cpu, 6336786);
1160 		missing++;
1161 #endif
1162 	}
1163 
1164 	/*LINTED*/
1165 	/*
1166 	 * Mutex primitives don't work as expected. This is erratum #147 from
1167 	 * 'Revision Guide for AMD Athlon 64 and AMD Opteron Processors'
1168 	 * document 25759.
1169 	 */
1170 	if (cpuid_opteron_erratum(cpu, 147) > 0) {
1171 #if defined(OPTERON_ERRATUM_147)
1172 		/*
1173 		 * This problem only occurs with 2 or more cores. If bit in
1174 		 * MSR_AMD_BU_CFG set, then not applicable. The workaround
1175 		 * is to patch the semaphone routines with the lfence
1176 		 * instruction to provide necessary load memory barrier with
1177 		 * possible subsequent read-modify-write ops.
1178 		 *
1179 		 * It is too early in boot to call the patch routine so
1180 		 * set erratum variable to be done in startup_end().
1181 		 */
1182 		if (opteron_erratum_147) {
1183 			opteron_erratum_147++;
1184 #if defined(__xpv)
1185 		} else if (is_x86_feature(x86_featureset, X86FSET_SSE2)) {
1186 			if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1187 				/*
1188 				 * XXPV	Use dom0_msr here when extended
1189 				 *	operations are supported?
1190 				 */
1191 				if (xpv_nr_phys_cpus() > 1)
1192 					opteron_erratum_147++;
1193 			} else {
1194 				/*
1195 				 * We have no way to tell how many physical
1196 				 * cpus there are, or even if this processor
1197 				 * has the problem, so enable the workaround
1198 				 * unconditionally (at some performance cost).
1199 				 */
1200 				opteron_erratum_147++;
1201 			}
1202 #else	/* __xpv */
1203 		} else if (is_x86_feature(x86_featureset, X86FSET_SSE2) &&
1204 		    ((opteron_get_nnodes() *
1205 		    cpuid_get_ncpu_per_chip(cpu)) > 1)) {
1206 			if ((xrdmsr(MSR_AMD_BU_CFG) & (UINT64_C(1) << 33)) == 0)
1207 				opteron_erratum_147++;
1208 #endif	/* __xpv */
1209 		}
1210 #else
1211 		workaround_warning(cpu, 147);
1212 		missing++;
1213 #endif
1214 	}
1215 
1216 	missing += do_erratum_298(cpu);
1217 
1218 	if (cpuid_opteron_erratum(cpu, 721) > 0) {
1219 #if defined(OPTERON_ERRATUM_721)
1220 		on_trap_data_t otd;
1221 
1222 		if (!on_trap(&otd, OT_DATA_ACCESS))
1223 			wrmsr(MSR_AMD_DE_CFG,
1224 			    rdmsr(MSR_AMD_DE_CFG) | AMD_DE_CFG_E721);
1225 		no_trap();
1226 
1227 		opteron_erratum_721++;
1228 #else
1229 		workaround_warning(cpu, 721);
1230 		missing++;
1231 #endif
1232 	}
1233 
1234 #ifdef __xpv
1235 	return (0);
1236 #else
1237 	return (missing);
1238 #endif
1239 }
1240 
1241 void
1242 workaround_errata_end()
1243 {
1244 #if defined(OPTERON_ERRATUM_88)
1245 	if (opteron_erratum_88)
1246 		workaround_applied(88);
1247 #endif
1248 #if defined(OPTERON_ERRATUM_91)
1249 	if (opteron_erratum_91)
1250 		workaround_applied(91);
1251 #endif
1252 #if defined(OPTERON_ERRATUM_93)
1253 	if (opteron_erratum_93)
1254 		workaround_applied(93);
1255 #endif
1256 #if defined(OPTERON_ERRATUM_95)
1257 	if (opteron_erratum_95)
1258 		workaround_applied(95);
1259 #endif
1260 #if defined(OPTERON_ERRATUM_100)
1261 	if (opteron_erratum_100)
1262 		workaround_applied(100);
1263 #endif
1264 #if defined(OPTERON_ERRATUM_108)
1265 	if (opteron_erratum_108)
1266 		workaround_applied(108);
1267 #endif
1268 #if defined(OPTERON_ERRATUM_109)
1269 	if (opteron_erratum_109) {
1270 		cmn_err(CE_WARN,
1271 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1272 		    " processor\nerratum 109 was not detected; updating your"
1273 		    " system's BIOS to a version\ncontaining this"
1274 		    " microcode patch is HIGHLY recommended or erroneous"
1275 		    " system\noperation may occur.\n");
1276 	}
1277 #endif
1278 #if defined(OPTERON_ERRATUM_121)
1279 	if (opteron_erratum_121)
1280 		workaround_applied(121);
1281 #endif
1282 #if defined(OPTERON_ERRATUM_122)
1283 	if (opteron_erratum_122)
1284 		workaround_applied(122);
1285 #endif
1286 #if defined(OPTERON_ERRATUM_123)
1287 	if (opteron_erratum_123) {
1288 		cmn_err(CE_WARN,
1289 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1290 		    " processor\nerratum 123 was not detected; updating your"
1291 		    " system's BIOS to a version\ncontaining this"
1292 		    " microcode patch is HIGHLY recommended or erroneous"
1293 		    " system\noperation may occur.\n");
1294 	}
1295 #endif
1296 #if defined(OPTERON_ERRATUM_131)
1297 	if (opteron_erratum_131) {
1298 		cmn_err(CE_WARN,
1299 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1300 		    " processor\nerratum 131 was not detected; updating your"
1301 		    " system's BIOS to a version\ncontaining this"
1302 		    " microcode patch is HIGHLY recommended or erroneous"
1303 		    " system\noperation may occur.\n");
1304 	}
1305 #endif
1306 #if defined(OPTERON_WORKAROUND_6336786)
1307 	if (opteron_workaround_6336786)
1308 		workaround_applied(6336786);
1309 #endif
1310 #if defined(OPTERON_ERRATUM_147)
1311 	if (opteron_erratum_147)
1312 		workaround_applied(147);
1313 #endif
1314 #if defined(OPTERON_ERRATUM_298)
1315 	if (opteron_erratum_298) {
1316 		cmn_err(CE_WARN,
1317 		    "BIOS microcode patch for AMD 64/Opteron(tm)"
1318 		    " processor\nerratum 298 was not detected; updating your"
1319 		    " system's BIOS to a version\ncontaining this"
1320 		    " microcode patch is HIGHLY recommended or erroneous"
1321 		    " system\noperation may occur.\n");
1322 	}
1323 #endif
1324 #if defined(OPTERON_ERRATUM_721)
1325 	if (opteron_erratum_721)
1326 		workaround_applied(721);
1327 #endif
1328 }
1329 
1330 /*
1331  * The procset_slave and procset_master are used to synchronize
1332  * between the control CPU and the target CPU when starting CPUs.
1333  */
1334 static cpuset_t procset_slave, procset_master;
1335 
1336 static void
1337 mp_startup_wait(cpuset_t *sp, processorid_t cpuid)
1338 {
1339 	cpuset_t tempset;
1340 
1341 	for (tempset = *sp; !CPU_IN_SET(tempset, cpuid);
1342 	    tempset = *(volatile cpuset_t *)sp) {
1343 		SMT_PAUSE();
1344 	}
1345 	CPUSET_ATOMIC_DEL(*(cpuset_t *)sp, cpuid);
1346 }
1347 
1348 static void
1349 mp_startup_signal(cpuset_t *sp, processorid_t cpuid)
1350 {
1351 	cpuset_t tempset;
1352 
1353 	CPUSET_ATOMIC_ADD(*(cpuset_t *)sp, cpuid);
1354 	for (tempset = *sp; CPU_IN_SET(tempset, cpuid);
1355 	    tempset = *(volatile cpuset_t *)sp) {
1356 		SMT_PAUSE();
1357 	}
1358 }
1359 
1360 int
1361 mp_start_cpu_common(cpu_t *cp, boolean_t boot)
1362 {
1363 	_NOTE(ARGUNUSED(boot));
1364 
1365 	void *ctx;
1366 	int delays;
1367 	int error = 0;
1368 	cpuset_t tempset;
1369 	processorid_t cpuid;
1370 #ifndef __xpv
1371 	extern void cpupm_init(cpu_t *);
1372 #endif
1373 
1374 	ASSERT(cp != NULL);
1375 	cpuid = cp->cpu_id;
1376 	ctx = mach_cpucontext_alloc(cp);
1377 	if (ctx == NULL) {
1378 		cmn_err(CE_WARN,
1379 		    "cpu%d: failed to allocate context", cp->cpu_id);
1380 		return (EAGAIN);
1381 	}
1382 	error = mach_cpu_start(cp, ctx);
1383 	if (error != 0) {
1384 		cmn_err(CE_WARN,
1385 		    "cpu%d: failed to start, error %d", cp->cpu_id, error);
1386 		mach_cpucontext_free(cp, ctx, error);
1387 		return (error);
1388 	}
1389 
1390 	for (delays = 0, tempset = procset_slave; !CPU_IN_SET(tempset, cpuid);
1391 	    delays++) {
1392 		if (delays == 500) {
1393 			/*
1394 			 * After five seconds, things are probably looking
1395 			 * a bit bleak - explain the hang.
1396 			 */
1397 			cmn_err(CE_NOTE, "cpu%d: started, "
1398 			    "but not running in the kernel yet", cpuid);
1399 		} else if (delays > 2000) {
1400 			/*
1401 			 * We waited at least 20 seconds, bail ..
1402 			 */
1403 			error = ETIMEDOUT;
1404 			cmn_err(CE_WARN, "cpu%d: timed out", cpuid);
1405 			mach_cpucontext_free(cp, ctx, error);
1406 			return (error);
1407 		}
1408 
1409 		/*
1410 		 * wait at least 10ms, then check again..
1411 		 */
1412 		delay(USEC_TO_TICK_ROUNDUP(10000));
1413 		tempset = *((volatile cpuset_t *)&procset_slave);
1414 	}
1415 	CPUSET_ATOMIC_DEL(procset_slave, cpuid);
1416 
1417 	mach_cpucontext_free(cp, ctx, 0);
1418 
1419 #ifndef __xpv
1420 	if (tsc_gethrtime_enable)
1421 		tsc_sync_master(cpuid);
1422 #endif
1423 
1424 	if (dtrace_cpu_init != NULL) {
1425 		(*dtrace_cpu_init)(cpuid);
1426 	}
1427 
1428 	/*
1429 	 * During CPU DR operations, the cpu_lock is held by current
1430 	 * (the control) thread. We can't release the cpu_lock here
1431 	 * because that will break the CPU DR logic.
1432 	 * On the other hand, CPUPM and processor group initialization
1433 	 * routines need to access the cpu_lock. So we invoke those
1434 	 * routines here on behalf of mp_startup_common().
1435 	 *
1436 	 * CPUPM and processor group initialization routines depend
1437 	 * on the cpuid probing results. Wait for mp_startup_common()
1438 	 * to signal that cpuid probing is done.
1439 	 */
1440 	mp_startup_wait(&procset_slave, cpuid);
1441 #ifndef __xpv
1442 	cpupm_init(cp);
1443 #endif
1444 	(void) pg_cpu_init(cp, B_FALSE);
1445 	cpu_set_state(cp);
1446 	mp_startup_signal(&procset_master, cpuid);
1447 
1448 	return (0);
1449 }
1450 
1451 /*
1452  * Start a single cpu, assuming that the kernel context is available
1453  * to successfully start another cpu.
1454  *
1455  * (For example, real mode code is mapped into the right place
1456  * in memory and is ready to be run.)
1457  */
1458 int
1459 start_cpu(processorid_t who)
1460 {
1461 	cpu_t *cp;
1462 	int error = 0;
1463 	cpuset_t tempset;
1464 
1465 	ASSERT(who != 0);
1466 
1467 	/*
1468 	 * Check if there's at least a Mbyte of kmem available
1469 	 * before attempting to start the cpu.
1470 	 */
1471 	if (kmem_avail() < 1024 * 1024) {
1472 		/*
1473 		 * Kick off a reap in case that helps us with
1474 		 * later attempts ..
1475 		 */
1476 		kmem_reap();
1477 		return (ENOMEM);
1478 	}
1479 
1480 	/*
1481 	 * First configure cpu.
1482 	 */
1483 	cp = mp_cpu_configure_common(who, B_TRUE);
1484 	ASSERT(cp != NULL);
1485 
1486 	/*
1487 	 * Then start cpu.
1488 	 */
1489 	error = mp_start_cpu_common(cp, B_TRUE);
1490 	if (error != 0) {
1491 		mp_cpu_unconfigure_common(cp, error);
1492 		return (error);
1493 	}
1494 
1495 	mutex_exit(&cpu_lock);
1496 	tempset = cpu_ready_set;
1497 	while (!CPU_IN_SET(tempset, who)) {
1498 		drv_usecwait(1);
1499 		tempset = *((volatile cpuset_t *)&cpu_ready_set);
1500 	}
1501 	mutex_enter(&cpu_lock);
1502 
1503 	return (0);
1504 }
1505 
1506 void
1507 start_other_cpus(int cprboot)
1508 {
1509 	_NOTE(ARGUNUSED(cprboot));
1510 
1511 	uint_t who;
1512 	uint_t bootcpuid = 0;
1513 
1514 	/*
1515 	 * Initialize our own cpu_info.
1516 	 */
1517 	init_cpu_info(CPU);
1518 
1519 #if !defined(__xpv)
1520 	init_cpu_id_gdt(CPU);
1521 #endif
1522 
1523 	cmn_err(CE_CONT, "?cpu%d: %s\n", CPU->cpu_id, CPU->cpu_idstr);
1524 	cmn_err(CE_CONT, "?cpu%d: %s\n", CPU->cpu_id, CPU->cpu_brandstr);
1525 
1526 	/*
1527 	 * KPTI initialisation happens very early in boot, before logging is
1528 	 * set up. Output a status message now as the boot CPU comes online.
1529 	 */
1530 	cmn_err(CE_CONT, "?KPTI %s (PCID %s, INVPCID %s)\n",
1531 	    kpti_enable ? "enabled" : "disabled",
1532 	    x86_use_pcid == 1 ? "in use" :
1533 	    (is_x86_feature(x86_featureset, X86FSET_PCID) ? "disabled" :
1534 	    "not supported"),
1535 	    x86_use_pcid == 1 && x86_use_invpcid == 1 ? "in use" :
1536 	    (is_x86_feature(x86_featureset, X86FSET_INVPCID) ? "disabled" :
1537 	    "not supported"));
1538 
1539 	/*
1540 	 * Initialize our syscall handlers
1541 	 */
1542 	init_cpu_syscall(CPU);
1543 
1544 	/*
1545 	 * Take the boot cpu out of the mp_cpus set because we know
1546 	 * it's already running.  Add it to the cpu_ready_set for
1547 	 * precisely the same reason.
1548 	 */
1549 	CPUSET_DEL(mp_cpus, bootcpuid);
1550 	CPUSET_ADD(cpu_ready_set, bootcpuid);
1551 
1552 	/*
1553 	 * skip the rest of this if
1554 	 * . only 1 cpu dectected and system isn't hotplug-capable
1555 	 * . not using MP
1556 	 */
1557 	if ((CPUSET_ISNULL(mp_cpus) && plat_dr_support_cpu() == 0) ||
1558 	    use_mp == 0) {
1559 		if (use_mp == 0)
1560 			cmn_err(CE_CONT, "?***** Not in MP mode\n");
1561 		goto done;
1562 	}
1563 
1564 	/*
1565 	 * perform such initialization as is needed
1566 	 * to be able to take CPUs on- and off-line.
1567 	 */
1568 	cpu_pause_init();
1569 
1570 	xc_init_cpu(CPU);		/* initialize processor crosscalls */
1571 
1572 	if (mach_cpucontext_init() != 0)
1573 		goto done;
1574 
1575 	flushes_require_xcalls = 1;
1576 
1577 	/*
1578 	 * We lock our affinity to the master CPU to ensure that all slave CPUs
1579 	 * do their TSC syncs with the same CPU.
1580 	 */
1581 	affinity_set(CPU_CURRENT);
1582 
1583 	for (who = 0; who < NCPU; who++) {
1584 		if (!CPU_IN_SET(mp_cpus, who))
1585 			continue;
1586 		ASSERT(who != bootcpuid);
1587 
1588 		mutex_enter(&cpu_lock);
1589 		if (start_cpu(who) != 0)
1590 			CPUSET_DEL(mp_cpus, who);
1591 		cpu_state_change_notify(who, CPU_SETUP);
1592 		mutex_exit(&cpu_lock);
1593 	}
1594 
1595 	/* Free the space allocated to hold the microcode file */
1596 	ucode_cleanup();
1597 
1598 	affinity_clear();
1599 
1600 	mach_cpucontext_fini();
1601 
1602 done:
1603 	if (get_hwenv() == HW_NATIVE)
1604 		workaround_errata_end();
1605 	cmi_post_mpstartup();
1606 
1607 #if !defined(__xpv)
1608 	/*
1609 	 * Once other CPUs have completed startup procedures, perform
1610 	 * initialization of hypervisor resources for HMA.
1611 	 */
1612 	hma_init();
1613 #endif
1614 
1615 	if (use_mp && ncpus != boot_max_ncpus) {
1616 		cmn_err(CE_NOTE,
1617 		    "System detected %d cpus, but "
1618 		    "only %d cpu(s) were enabled during boot.",
1619 		    boot_max_ncpus, ncpus);
1620 		cmn_err(CE_NOTE,
1621 		    "Use \"boot-ncpus\" parameter to enable more CPU(s). "
1622 		    "See eeprom(8).");
1623 	}
1624 }
1625 
1626 int
1627 mp_cpu_configure(int cpuid)
1628 {
1629 	cpu_t *cp;
1630 
1631 	if (use_mp == 0 || plat_dr_support_cpu() == 0) {
1632 		return (ENOTSUP);
1633 	}
1634 
1635 	cp = cpu_get(cpuid);
1636 	if (cp != NULL) {
1637 		return (EALREADY);
1638 	}
1639 
1640 	/*
1641 	 * Check if there's at least a Mbyte of kmem available
1642 	 * before attempting to start the cpu.
1643 	 */
1644 	if (kmem_avail() < 1024 * 1024) {
1645 		/*
1646 		 * Kick off a reap in case that helps us with
1647 		 * later attempts ..
1648 		 */
1649 		kmem_reap();
1650 		return (ENOMEM);
1651 	}
1652 
1653 	cp = mp_cpu_configure_common(cpuid, B_FALSE);
1654 	ASSERT(cp != NULL && cpu_get(cpuid) == cp);
1655 
1656 	return (cp != NULL ? 0 : EAGAIN);
1657 }
1658 
1659 int
1660 mp_cpu_unconfigure(int cpuid)
1661 {
1662 	cpu_t *cp;
1663 
1664 	if (use_mp == 0 || plat_dr_support_cpu() == 0) {
1665 		return (ENOTSUP);
1666 	} else if (cpuid < 0 || cpuid >= max_ncpus) {
1667 		return (EINVAL);
1668 	}
1669 
1670 	cp = cpu_get(cpuid);
1671 	if (cp == NULL) {
1672 		return (ENODEV);
1673 	}
1674 	mp_cpu_unconfigure_common(cp, 0);
1675 
1676 	return (0);
1677 }
1678 
1679 /*
1680  * Startup function for 'other' CPUs (besides boot cpu).
1681  * Called from real_mode_start.
1682  *
1683  * WARNING: until CPU_READY is set, mp_startup_common and routines called by
1684  * mp_startup_common should not call routines (e.g. kmem_free) that could call
1685  * hat_unload which requires CPU_READY to be set.
1686  */
1687 static void
1688 mp_startup_common(boolean_t boot)
1689 {
1690 	cpu_t *cp = CPU;
1691 	uchar_t new_x86_featureset[BT_SIZEOFMAP(NUM_X86_FEATURES)];
1692 	extern void cpu_event_init_cpu(cpu_t *);
1693 
1694 	/*
1695 	 * We need to get TSC on this proc synced (i.e., any delta
1696 	 * from cpu0 accounted for) as soon as we can, because many
1697 	 * many things use gethrtime/pc_gethrestime, including
1698 	 * interrupts, cmn_err, etc.  Before we can do that, we want to
1699 	 * clear TSC if we're on a buggy Sandy/Ivy Bridge CPU, so do that
1700 	 * right away.
1701 	 */
1702 	bzero(new_x86_featureset, BT_SIZEOFMAP(NUM_X86_FEATURES));
1703 	cpuid_execpass(cp, CPUID_PASS_PRELUDE, new_x86_featureset);
1704 	cpuid_execpass(cp, CPUID_PASS_IDENT, NULL);
1705 	cpuid_execpass(cp, CPUID_PASS_BASIC, new_x86_featureset);
1706 
1707 	if (boot && get_hwenv() == HW_NATIVE &&
1708 	    cpuid_getvendor(CPU) == X86_VENDOR_Intel &&
1709 	    cpuid_getfamily(CPU) == 6 &&
1710 	    (cpuid_getmodel(CPU) == 0x2d || cpuid_getmodel(CPU) == 0x3e) &&
1711 	    is_x86_feature(new_x86_featureset, X86FSET_TSC)) {
1712 		(void) wrmsr(REG_TSC, 0UL);
1713 	}
1714 
1715 	/* Let the control CPU continue into tsc_sync_master() */
1716 	mp_startup_signal(&procset_slave, cp->cpu_id);
1717 
1718 #ifndef __xpv
1719 	if (tsc_gethrtime_enable)
1720 		tsc_sync_slave();
1721 #endif
1722 
1723 	/*
1724 	 * Once this was done from assembly, but it's safer here; if
1725 	 * it blocks, we need to be able to swtch() to and from, and
1726 	 * since we get here by calling t_pc, we need to do that call
1727 	 * before swtch() overwrites it.
1728 	 */
1729 	(void) (*ap_mlsetup)();
1730 
1731 #ifndef __xpv
1732 	/*
1733 	 * Program this cpu's PAT
1734 	 */
1735 	pat_sync();
1736 #endif
1737 
1738 	/*
1739 	 * Set up TSC_AUX to contain the cpuid for this processor
1740 	 * for the rdtscp instruction.
1741 	 */
1742 	if (is_x86_feature(x86_featureset, X86FSET_TSCP))
1743 		(void) wrmsr(MSR_AMD_TSCAUX, cp->cpu_id);
1744 
1745 	/*
1746 	 * Initialize this CPU's syscall handlers
1747 	 */
1748 	init_cpu_syscall(cp);
1749 
1750 	/*
1751 	 * Enable interrupts with spl set to LOCK_LEVEL. LOCK_LEVEL is the
1752 	 * highest level at which a routine is permitted to block on
1753 	 * an adaptive mutex (allows for cpu poke interrupt in case
1754 	 * the cpu is blocked on a mutex and halts). Setting LOCK_LEVEL blocks
1755 	 * device interrupts that may end up in the hat layer issuing cross
1756 	 * calls before CPU_READY is set.
1757 	 */
1758 	splx(ipltospl(LOCK_LEVEL));
1759 	sti();
1760 
1761 	/*
1762 	 * There exists a small subset of systems which expose differing
1763 	 * MWAIT/MONITOR support between CPUs.  If MWAIT support is absent from
1764 	 * the boot CPU, but is found on a later CPU, the system continues to
1765 	 * operate as if no MWAIT support is available.
1766 	 *
1767 	 * The reverse case, where MWAIT is available on the boot CPU but not
1768 	 * on a subsequently initialized CPU, is not presently allowed and will
1769 	 * result in a panic.
1770 	 */
1771 	if (is_x86_feature(x86_featureset, X86FSET_MWAIT) !=
1772 	    is_x86_feature(new_x86_featureset, X86FSET_MWAIT)) {
1773 		if (!is_x86_feature(x86_featureset, X86FSET_MWAIT)) {
1774 			remove_x86_feature(new_x86_featureset, X86FSET_MWAIT);
1775 		} else {
1776 			panic("unsupported mixed cpu mwait support detected");
1777 		}
1778 	}
1779 
1780 	/*
1781 	 * We could be more sophisticated here, and just mark the CPU
1782 	 * as "faulted" but at this point we'll opt for the easier
1783 	 * answer of dying horribly.  Provided the boot cpu is ok,
1784 	 * the system can be recovered by booting with use_mp set to zero.
1785 	 */
1786 	if (workaround_errata(cp) != 0)
1787 		panic("critical workaround(s) missing for cpu%d", cp->cpu_id);
1788 
1789 	/*
1790 	 * We can touch cpu_flags here without acquiring the cpu_lock here
1791 	 * because the cpu_lock is held by the control CPU which is running
1792 	 * mp_start_cpu_common().
1793 	 * Need to clear CPU_QUIESCED flag before calling any function which
1794 	 * may cause thread context switching, such as kmem_alloc() etc.
1795 	 * The idle thread checks for CPU_QUIESCED flag and loops for ever if
1796 	 * it's set. So the startup thread may have no chance to switch back
1797 	 * again if it's switched away with CPU_QUIESCED set.
1798 	 */
1799 	cp->cpu_flags &= ~(CPU_POWEROFF | CPU_QUIESCED);
1800 
1801 	enable_pcid();
1802 
1803 	/*
1804 	 * Setup this processor for XSAVE.
1805 	 */
1806 	if (fp_save_mech == FP_XSAVE) {
1807 		xsave_setup_msr(cp);
1808 	}
1809 
1810 	cpuid_execpass(cp, CPUID_PASS_EXTENDED, NULL);
1811 	cpuid_execpass(cp, CPUID_PASS_DYNAMIC, NULL);
1812 	cpuid_execpass(cp, CPUID_PASS_RESOLVE, NULL);
1813 
1814 	/*
1815 	 * Correct cpu_idstr and cpu_brandstr on target CPU after
1816 	 * CPUID_PASS_DYNAMIC is done.
1817 	 */
1818 	(void) cpuid_getidstr(cp, cp->cpu_idstr, CPU_IDSTRLEN);
1819 	(void) cpuid_getbrandstr(cp, cp->cpu_brandstr, CPU_IDSTRLEN);
1820 
1821 	cp->cpu_flags |= CPU_RUNNING | CPU_READY | CPU_EXISTS;
1822 
1823 	post_startup_cpu_fixups();
1824 
1825 	cpu_event_init_cpu(cp);
1826 
1827 	/*
1828 	 * Enable preemption here so that contention for any locks acquired
1829 	 * later in mp_startup_common may be preempted if the thread owning
1830 	 * those locks is continuously executing on other CPUs (for example,
1831 	 * this CPU must be preemptible to allow other CPUs to pause it during
1832 	 * their startup phases).  It's safe to enable preemption here because
1833 	 * the CPU state is pretty-much fully constructed.
1834 	 */
1835 	curthread->t_preempt = 0;
1836 
1837 	/* The base spl should still be at LOCK LEVEL here */
1838 	ASSERT(cp->cpu_base_spl == ipltospl(LOCK_LEVEL));
1839 	set_base_spl();		/* Restore the spl to its proper value */
1840 
1841 	pghw_physid_create(cp);
1842 	/*
1843 	 * Delegate initialization tasks, which need to access the cpu_lock,
1844 	 * to mp_start_cpu_common() because we can't acquire the cpu_lock here
1845 	 * during CPU DR operations.
1846 	 */
1847 	mp_startup_signal(&procset_slave, cp->cpu_id);
1848 	mp_startup_wait(&procset_master, cp->cpu_id);
1849 	pg_cmt_cpu_startup(cp);
1850 
1851 	if (boot) {
1852 		mutex_enter(&cpu_lock);
1853 		cp->cpu_flags &= ~CPU_OFFLINE;
1854 		cpu_enable_intr(cp);
1855 		cpu_add_active(cp);
1856 		mutex_exit(&cpu_lock);
1857 	}
1858 
1859 	/* Enable interrupts */
1860 	(void) spl0();
1861 
1862 	/*
1863 	 * Fill out cpu_ucode_info.  Update microcode if necessary. Note that
1864 	 * this is done after pass1 on the boot CPU, but it needs to be later on
1865 	 * for the other CPUs.
1866 	 */
1867 	ucode_check(cp);
1868 	cpuid_pass_ucode(cp, new_x86_featureset);
1869 
1870 	/*
1871 	 * Do a sanity check to make sure this new CPU is a sane thing
1872 	 * to add to the collection of processors running this system.
1873 	 *
1874 	 * XXX	Clearly this needs to get more sophisticated, if x86
1875 	 * systems start to get built out of heterogenous CPUs; as is
1876 	 * likely to happen once the number of processors in a configuration
1877 	 * gets large enough.
1878 	 */
1879 	if (compare_x86_featureset(x86_featureset, new_x86_featureset) ==
1880 	    B_FALSE) {
1881 		cmn_err(CE_CONT, "cpu%d: featureset\n", cp->cpu_id);
1882 		print_x86_featureset(new_x86_featureset);
1883 		cmn_err(CE_WARN, "cpu%d feature mismatch", cp->cpu_id);
1884 	}
1885 
1886 #ifndef __xpv
1887 	{
1888 		/*
1889 		 * Set up the CPU module for this CPU.  This can't be done
1890 		 * before this CPU is made CPU_READY, because we may (in
1891 		 * heterogeneous systems) need to go load another CPU module.
1892 		 * The act of attempting to load a module may trigger a
1893 		 * cross-call, which will ASSERT unless this cpu is CPU_READY.
1894 		 */
1895 		cmi_hdl_t hdl;
1896 
1897 		if ((hdl = cmi_init(CMI_HDL_NATIVE, cmi_ntv_hwchipid(CPU),
1898 		    cmi_ntv_hwcoreid(CPU), cmi_ntv_hwstrandid(CPU))) != NULL) {
1899 			if (is_x86_feature(x86_featureset, X86FSET_MCA))
1900 				cmi_mca_init(hdl);
1901 			cp->cpu_m.mcpu_cmi_hdl = hdl;
1902 		}
1903 	}
1904 #endif /* __xpv */
1905 
1906 	if (boothowto & RB_DEBUG)
1907 		kdi_cpu_init();
1908 
1909 	(void) mach_cpu_create_device_node(cp, NULL);
1910 
1911 	/*
1912 	 * Setting the bit in cpu_ready_set must be the last operation in
1913 	 * processor initialization; the boot CPU will continue to boot once
1914 	 * it sees this bit set for all active CPUs.
1915 	 */
1916 	CPUSET_ATOMIC_ADD(cpu_ready_set, cp->cpu_id);
1917 
1918 	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_idstr);
1919 	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_brandstr);
1920 	cmn_err(CE_CONT, "?cpu%d initialization complete - online\n",
1921 	    cp->cpu_id);
1922 
1923 	/*
1924 	 * Now we are done with the startup thread, so free it up.
1925 	 */
1926 	thread_exit();
1927 	/*NOTREACHED*/
1928 }
1929 
1930 /*
1931  * Startup function for 'other' CPUs at boot time (besides boot cpu).
1932  */
1933 static void
1934 mp_startup_boot(void)
1935 {
1936 	mp_startup_common(B_TRUE);
1937 }
1938 
1939 /*
1940  * Startup function for hotplug CPUs at runtime.
1941  */
1942 void
1943 mp_startup_hotplug(void)
1944 {
1945 	mp_startup_common(B_FALSE);
1946 }
1947 
1948 /*
1949  * Start CPU on user request.
1950  */
1951 /* ARGSUSED */
1952 int
1953 mp_cpu_start(struct cpu *cp)
1954 {
1955 	ASSERT(MUTEX_HELD(&cpu_lock));
1956 	return (0);
1957 }
1958 
1959 /*
1960  * Stop CPU on user request.
1961  */
1962 int
1963 mp_cpu_stop(struct cpu *cp)
1964 {
1965 	extern int cbe_psm_timer_mode;
1966 	ASSERT(MUTEX_HELD(&cpu_lock));
1967 
1968 #ifdef __xpv
1969 	/*
1970 	 * We can't offline vcpu0.
1971 	 */
1972 	if (cp->cpu_id == 0)
1973 		return (EBUSY);
1974 #endif
1975 
1976 	/*
1977 	 * If TIMER_PERIODIC mode is used, CPU0 is the one running it;
1978 	 * can't stop it.  (This is true only for machines with no TSC.)
1979 	 */
1980 
1981 	if ((cbe_psm_timer_mode == TIMER_PERIODIC) && (cp->cpu_id == 0))
1982 		return (EBUSY);
1983 
1984 	return (0);
1985 }
1986 
1987 /*
1988  * Take the specified CPU out of participation in interrupts.
1989  *
1990  * Usually, we hold cpu_lock. But we cannot assert as such due to the
1991  * exception - i_cpr_save_context() - where we have mutual exclusion via a
1992  * separate mechanism.
1993  */
1994 int
1995 cpu_disable_intr(struct cpu *cp)
1996 {
1997 	if (psm_disable_intr(cp->cpu_id) != DDI_SUCCESS)
1998 		return (EBUSY);
1999 
2000 	cp->cpu_flags &= ~CPU_ENABLE;
2001 	ncpus_intr_enabled--;
2002 	return (0);
2003 }
2004 
2005 /*
2006  * Allow the specified CPU to participate in interrupts.
2007  */
2008 void
2009 cpu_enable_intr(struct cpu *cp)
2010 {
2011 	ASSERT(MUTEX_HELD(&cpu_lock));
2012 	cp->cpu_flags |= CPU_ENABLE;
2013 	ncpus_intr_enabled++;
2014 	psm_enable_intr(cp->cpu_id);
2015 }
2016 
2017 void
2018 mp_cpu_faulted_enter(struct cpu *cp)
2019 {
2020 #ifdef __xpv
2021 	_NOTE(ARGUNUSED(cp));
2022 #else
2023 	cmi_hdl_t hdl = cp->cpu_m.mcpu_cmi_hdl;
2024 
2025 	if (hdl != NULL) {
2026 		cmi_hdl_hold(hdl);
2027 	} else {
2028 		hdl = cmi_hdl_lookup(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp),
2029 		    cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp));
2030 	}
2031 	if (hdl != NULL) {
2032 		cmi_faulted_enter(hdl);
2033 		cmi_hdl_rele(hdl);
2034 	}
2035 #endif
2036 }
2037 
2038 void
2039 mp_cpu_faulted_exit(struct cpu *cp)
2040 {
2041 #ifdef __xpv
2042 	_NOTE(ARGUNUSED(cp));
2043 #else
2044 	cmi_hdl_t hdl = cp->cpu_m.mcpu_cmi_hdl;
2045 
2046 	if (hdl != NULL) {
2047 		cmi_hdl_hold(hdl);
2048 	} else {
2049 		hdl = cmi_hdl_lookup(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp),
2050 		    cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp));
2051 	}
2052 	if (hdl != NULL) {
2053 		cmi_faulted_exit(hdl);
2054 		cmi_hdl_rele(hdl);
2055 	}
2056 #endif
2057 }
2058 
2059 /*
2060  * The following two routines are used as context operators on threads belonging
2061  * to processes with a private LDT (see sysi86).  Due to the rarity of such
2062  * processes, these routines are currently written for best code readability and
2063  * organization rather than speed.  We could avoid checking x86_featureset at
2064  * every context switch by installing different context ops, depending on
2065  * x86_featureset, at LDT creation time -- one for each combination of fast
2066  * syscall features.
2067  */
2068 
2069 void
2070 cpu_fast_syscall_disable(void)
2071 {
2072 	if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
2073 	    is_x86_feature(x86_featureset, X86FSET_SEP))
2074 		cpu_sep_disable();
2075 	if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
2076 	    is_x86_feature(x86_featureset, X86FSET_ASYSC))
2077 		cpu_asysc_disable();
2078 }
2079 
2080 void
2081 cpu_fast_syscall_enable(void)
2082 {
2083 	if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
2084 	    is_x86_feature(x86_featureset, X86FSET_SEP))
2085 		cpu_sep_enable();
2086 	if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
2087 	    is_x86_feature(x86_featureset, X86FSET_ASYSC))
2088 		cpu_asysc_enable();
2089 }
2090 
2091 static void
2092 cpu_sep_enable(void)
2093 {
2094 	ASSERT(is_x86_feature(x86_featureset, X86FSET_SEP));
2095 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
2096 
2097 	wrmsr(MSR_INTC_SEP_CS, (uint64_t)(uintptr_t)KCS_SEL);
2098 
2099 	CPU->cpu_m.mcpu_fast_syscall_state |= FSS_SEP_ENABLED;
2100 }
2101 
2102 static void
2103 cpu_sep_disable(void)
2104 {
2105 	ASSERT(is_x86_feature(x86_featureset, X86FSET_SEP));
2106 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
2107 
2108 	/*
2109 	 * Setting the SYSENTER_CS_MSR register to 0 causes software executing
2110 	 * the sysenter or sysexit instruction to trigger a #gp fault.
2111 	 */
2112 	wrmsr(MSR_INTC_SEP_CS, 0);
2113 
2114 	CPU->cpu_m.mcpu_fast_syscall_state &= ~FSS_SEP_ENABLED;
2115 }
2116 
2117 static void
2118 cpu_asysc_enable(void)
2119 {
2120 	ASSERT(is_x86_feature(x86_featureset, X86FSET_ASYSC));
2121 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
2122 
2123 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) |
2124 	    (uint64_t)(uintptr_t)AMD_EFER_SCE);
2125 
2126 	CPU->cpu_m.mcpu_fast_syscall_state |= FSS_ASYSC_ENABLED;
2127 }
2128 
2129 static void
2130 cpu_asysc_disable(void)
2131 {
2132 	ASSERT(is_x86_feature(x86_featureset, X86FSET_ASYSC));
2133 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
2134 
2135 	/*
2136 	 * Turn off the SCE (syscall enable) bit in the EFER register. Software
2137 	 * executing syscall or sysret with this bit off will incur a #ud trap.
2138 	 */
2139 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) &
2140 	    ~((uint64_t)(uintptr_t)AMD_EFER_SCE));
2141 
2142 	CPU->cpu_m.mcpu_fast_syscall_state &= ~FSS_ASYSC_ENABLED;
2143 }
2144