xref: /titanic_41/usr/src/uts/i86pc/os/mp_startup.c (revision d9fc8ba670791dc3b8398347e67e1c5825d5e341)
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 2016 Joyent, Inc.
31  * Copyright 2013 Nexenta Systems, Inc.  All rights reserved.
32  */
33 
34 #include <sys/types.h>
35 #include <sys/thread.h>
36 #include <sys/cpuvar.h>
37 #include <sys/cpu.h>
38 #include <sys/t_lock.h>
39 #include <sys/param.h>
40 #include <sys/proc.h>
41 #include <sys/disp.h>
42 #include <sys/class.h>
43 #include <sys/cmn_err.h>
44 #include <sys/debug.h>
45 #include <sys/note.h>
46 #include <sys/asm_linkage.h>
47 #include <sys/x_call.h>
48 #include <sys/systm.h>
49 #include <sys/var.h>
50 #include <sys/vtrace.h>
51 #include <vm/hat.h>
52 #include <vm/as.h>
53 #include <vm/seg_kmem.h>
54 #include <vm/seg_kp.h>
55 #include <sys/segments.h>
56 #include <sys/kmem.h>
57 #include <sys/stack.h>
58 #include <sys/smp_impldefs.h>
59 #include <sys/x86_archext.h>
60 #include <sys/machsystm.h>
61 #include <sys/traptrace.h>
62 #include <sys/clock.h>
63 #include <sys/cpc_impl.h>
64 #include <sys/pg.h>
65 #include <sys/cmt.h>
66 #include <sys/dtrace.h>
67 #include <sys/archsystm.h>
68 #include <sys/fp.h>
69 #include <sys/reboot.h>
70 #include <sys/kdi_machimpl.h>
71 #include <vm/hat_i86.h>
72 #include <vm/vm_dep.h>
73 #include <sys/memnode.h>
74 #include <sys/pci_cfgspace.h>
75 #include <sys/mach_mmu.h>
76 #include <sys/sysmacros.h>
77 #if defined(__xpv)
78 #include <sys/hypervisor.h>
79 #endif
80 #include <sys/cpu_module.h>
81 #include <sys/ontrap.h>
82 
83 struct cpu	cpus[1];			/* CPU data */
84 struct cpu	*cpu[NCPU] = {&cpus[0]};	/* pointers to all CPUs */
85 struct cpu	*cpu_free_list;			/* list for released CPUs */
86 cpu_core_t	cpu_core[NCPU];			/* cpu_core structures */
87 
88 #define	cpu_next_free	cpu_prev
89 
90 /*
91  * Useful for disabling MP bring-up on a MP capable system.
92  */
93 int use_mp = 1;
94 
95 /*
96  * to be set by a PSM to indicate what cpus
97  * are sitting around on the system.
98  */
99 cpuset_t mp_cpus;
100 
101 /*
102  * This variable is used by the hat layer to decide whether or not
103  * critical sections are needed to prevent race conditions.  For sun4m,
104  * this variable is set once enough MP initialization has been done in
105  * order to allow cross calls.
106  */
107 int flushes_require_xcalls;
108 
109 cpuset_t cpu_ready_set;		/* initialized in startup() */
110 
111 static void mp_startup_boot(void);
112 static void mp_startup_hotplug(void);
113 
114 static void cpu_sep_enable(void);
115 static void cpu_sep_disable(void);
116 static void cpu_asysc_enable(void);
117 static void cpu_asysc_disable(void);
118 
119 /*
120  * Init CPU info - get CPU type info for processor_info system call.
121  */
122 void
123 init_cpu_info(struct cpu *cp)
124 {
125 	processor_info_t *pi = &cp->cpu_type_info;
126 
127 	/*
128 	 * Get clock-frequency property for the CPU.
129 	 */
130 	pi->pi_clock = cpu_freq;
131 
132 	/*
133 	 * Current frequency in Hz.
134 	 */
135 	cp->cpu_curr_clock = cpu_freq_hz;
136 
137 	/*
138 	 * Supported frequencies.
139 	 */
140 	if (cp->cpu_supp_freqs == NULL) {
141 		cpu_set_supp_freqs(cp, NULL);
142 	}
143 
144 	(void) strcpy(pi->pi_processor_type, "i386");
145 	if (fpu_exists)
146 		(void) strcpy(pi->pi_fputypes, "i387 compatible");
147 
148 	cp->cpu_idstr = kmem_zalloc(CPU_IDSTRLEN, KM_SLEEP);
149 	cp->cpu_brandstr = kmem_zalloc(CPU_IDSTRLEN, KM_SLEEP);
150 
151 	/*
152 	 * If called for the BSP, cp is equal to current CPU.
153 	 * For non-BSPs, cpuid info of cp is not ready yet, so use cpuid info
154 	 * of current CPU as default values for cpu_idstr and cpu_brandstr.
155 	 * They will be corrected in mp_startup_common() after cpuid_pass1()
156 	 * has been invoked on target CPU.
157 	 */
158 	(void) cpuid_getidstr(CPU, cp->cpu_idstr, CPU_IDSTRLEN);
159 	(void) cpuid_getbrandstr(CPU, cp->cpu_brandstr, CPU_IDSTRLEN);
160 }
161 
162 /*
163  * Configure syscall support on this CPU.
164  */
165 /*ARGSUSED*/
166 void
167 init_cpu_syscall(struct cpu *cp)
168 {
169 	kpreempt_disable();
170 
171 #if defined(__amd64)
172 	if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
173 	    is_x86_feature(x86_featureset, X86FSET_ASYSC)) {
174 
175 #if !defined(__lint)
176 		/*
177 		 * The syscall instruction imposes a certain ordering on
178 		 * segment selectors, so we double-check that ordering
179 		 * here.
180 		 */
181 		ASSERT(KDS_SEL == KCS_SEL + 8);
182 		ASSERT(UDS_SEL == U32CS_SEL + 8);
183 		ASSERT(UCS_SEL == U32CS_SEL + 16);
184 #endif
185 		/*
186 		 * Turn syscall/sysret extensions on.
187 		 */
188 		cpu_asysc_enable();
189 
190 		/*
191 		 * Program the magic registers ..
192 		 */
193 		wrmsr(MSR_AMD_STAR,
194 		    ((uint64_t)(U32CS_SEL << 16 | KCS_SEL)) << 32);
195 		wrmsr(MSR_AMD_LSTAR, (uint64_t)(uintptr_t)sys_syscall);
196 		wrmsr(MSR_AMD_CSTAR, (uint64_t)(uintptr_t)sys_syscall32);
197 
198 		/*
199 		 * This list of flags is masked off the incoming
200 		 * %rfl when we enter the kernel.
201 		 */
202 		wrmsr(MSR_AMD_SFMASK, (uint64_t)(uintptr_t)(PS_IE | PS_T));
203 	}
204 #endif
205 
206 	/*
207 	 * On 32-bit kernels, we use sysenter/sysexit because it's too
208 	 * hard to use syscall/sysret, and it is more portable anyway.
209 	 *
210 	 * On 64-bit kernels on Nocona machines, the 32-bit syscall
211 	 * variant isn't available to 32-bit applications, but sysenter is.
212 	 */
213 	if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
214 	    is_x86_feature(x86_featureset, X86FSET_SEP)) {
215 
216 #if !defined(__lint)
217 		/*
218 		 * The sysenter instruction imposes a certain ordering on
219 		 * segment selectors, so we double-check that ordering
220 		 * here. See "sysenter" in Intel document 245471-012, "IA-32
221 		 * Intel Architecture Software Developer's Manual Volume 2:
222 		 * Instruction Set Reference"
223 		 */
224 		ASSERT(KDS_SEL == KCS_SEL + 8);
225 
226 		ASSERT32(UCS_SEL == ((KCS_SEL + 16) | 3));
227 		ASSERT32(UDS_SEL == UCS_SEL + 8);
228 
229 		ASSERT64(U32CS_SEL == ((KCS_SEL + 16) | 3));
230 		ASSERT64(UDS_SEL == U32CS_SEL + 8);
231 #endif
232 
233 		cpu_sep_enable();
234 
235 		/*
236 		 * resume() sets this value to the base of the threads stack
237 		 * via a context handler.
238 		 */
239 		wrmsr(MSR_INTC_SEP_ESP, 0);
240 		wrmsr(MSR_INTC_SEP_EIP, (uint64_t)(uintptr_t)sys_sysenter);
241 	}
242 
243 	kpreempt_enable();
244 }
245 
246 #if !defined(__xpv)
247 /*
248  * Configure per-cpu ID GDT
249  */
250 static void
251 init_cpu_id_gdt(struct cpu *cp)
252 {
253 	/* Write cpu_id into limit field of GDT for usermode retrieval */
254 #if defined(__amd64)
255 	set_usegd(&cp->cpu_gdt[GDT_CPUID], SDP_SHORT, NULL, cp->cpu_id,
256 	    SDT_MEMRODA, SEL_UPL, SDP_BYTES, SDP_OP32);
257 #elif defined(__i386)
258 	set_usegd(&cp->cpu_gdt[GDT_CPUID], NULL, cp->cpu_id, SDT_MEMRODA,
259 	    SEL_UPL, SDP_BYTES, SDP_OP32);
260 #endif
261 }
262 #endif /* !defined(__xpv) */
263 
264 /*
265  * Multiprocessor initialization.
266  *
267  * Allocate and initialize the cpu structure, TRAPTRACE buffer, and the
268  * startup and idle threads for the specified CPU.
269  * Parameter boot is true for boot time operations and is false for CPU
270  * DR operations.
271  */
272 static struct cpu *
273 mp_cpu_configure_common(int cpun, boolean_t boot)
274 {
275 	struct cpu *cp;
276 	kthread_id_t tp;
277 	caddr_t	sp;
278 	proc_t *procp;
279 #if !defined(__xpv)
280 	extern int idle_cpu_prefer_mwait;
281 	extern void cpu_idle_mwait();
282 #endif
283 	extern void idle();
284 	extern void cpu_idle();
285 
286 #ifdef TRAPTRACE
287 	trap_trace_ctl_t *ttc = &trap_trace_ctl[cpun];
288 #endif
289 
290 	ASSERT(MUTEX_HELD(&cpu_lock));
291 	ASSERT(cpun < NCPU && cpu[cpun] == NULL);
292 
293 	if (cpu_free_list == NULL) {
294 		cp = kmem_zalloc(sizeof (*cp), KM_SLEEP);
295 	} else {
296 		cp = cpu_free_list;
297 		cpu_free_list = cp->cpu_next_free;
298 	}
299 
300 	cp->cpu_m.mcpu_istamp = cpun << 16;
301 
302 	/* Create per CPU specific threads in the process p0. */
303 	procp = &p0;
304 
305 	/*
306 	 * Initialize the dispatcher first.
307 	 */
308 	disp_cpu_init(cp);
309 
310 	cpu_vm_data_init(cp);
311 
312 	/*
313 	 * Allocate and initialize the startup thread for this CPU.
314 	 * Interrupt and process switch stacks get allocated later
315 	 * when the CPU starts running.
316 	 */
317 	tp = thread_create(NULL, 0, NULL, NULL, 0, procp,
318 	    TS_STOPPED, maxclsyspri);
319 
320 	/*
321 	 * Set state to TS_ONPROC since this thread will start running
322 	 * as soon as the CPU comes online.
323 	 *
324 	 * All the other fields of the thread structure are setup by
325 	 * thread_create().
326 	 */
327 	THREAD_ONPROC(tp, cp);
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 	 * Setup thread to start in mp_startup_common.
336 	 */
337 	sp = tp->t_stk;
338 	tp->t_sp = (uintptr_t)(sp - MINFRAME);
339 #if defined(__amd64)
340 	tp->t_sp -= STACK_ENTRY_ALIGN;		/* fake a call */
341 #endif
342 	/*
343 	 * Setup thread start entry point for boot or hotplug.
344 	 */
345 	if (boot) {
346 		tp->t_pc = (uintptr_t)mp_startup_boot;
347 	} else {
348 		tp->t_pc = (uintptr_t)mp_startup_hotplug;
349 	}
350 
351 	cp->cpu_id = cpun;
352 	cp->cpu_self = cp;
353 	cp->cpu_thread = tp;
354 	cp->cpu_lwp = NULL;
355 	cp->cpu_dispthread = tp;
356 	cp->cpu_dispatch_pri = DISP_PRIO(tp);
357 
358 	/*
359 	 * cpu_base_spl must be set explicitly here to prevent any blocking
360 	 * operations in mp_startup_common from causing the spl of the cpu
361 	 * to drop to 0 (allowing device interrupts before we're ready) in
362 	 * resume().
363 	 * cpu_base_spl MUST remain at LOCK_LEVEL until the cpu is CPU_READY.
364 	 * As an extra bit of security on DEBUG kernels, this is enforced with
365 	 * an assertion in mp_startup_common() -- before cpu_base_spl is set
366 	 * to its proper value.
367 	 */
368 	cp->cpu_base_spl = ipltospl(LOCK_LEVEL);
369 
370 	/*
371 	 * Now, initialize per-CPU idle thread for this CPU.
372 	 */
373 	tp = thread_create(NULL, PAGESIZE, idle, NULL, 0, procp, TS_ONPROC, -1);
374 
375 	cp->cpu_idle_thread = tp;
376 
377 	tp->t_preempt = 1;
378 	tp->t_bound_cpu = cp;
379 	tp->t_affinitycnt = 1;
380 	tp->t_cpu = cp;
381 	tp->t_disp_queue = cp->cpu_disp;
382 
383 	/*
384 	 * Bootstrap the CPU's PG data
385 	 */
386 	pg_cpu_bootstrap(cp);
387 
388 	/*
389 	 * Perform CPC initialization on the new CPU.
390 	 */
391 	kcpc_hw_init(cp);
392 
393 	/*
394 	 * Allocate virtual addresses for cpu_caddr1 and cpu_caddr2
395 	 * for each CPU.
396 	 */
397 	setup_vaddr_for_ppcopy(cp);
398 
399 	/*
400 	 * Allocate page for new GDT and initialize from current GDT.
401 	 */
402 #if !defined(__lint)
403 	ASSERT((sizeof (*cp->cpu_gdt) * NGDT) <= PAGESIZE);
404 #endif
405 	cp->cpu_gdt = kmem_zalloc(PAGESIZE, KM_SLEEP);
406 	bcopy(CPU->cpu_gdt, cp->cpu_gdt, (sizeof (*cp->cpu_gdt) * NGDT));
407 
408 #if defined(__i386)
409 	/*
410 	 * setup kernel %gs.
411 	 */
412 	set_usegd(&cp->cpu_gdt[GDT_GS], cp, sizeof (struct cpu) -1, SDT_MEMRWA,
413 	    SEL_KPL, 0, 1);
414 #endif
415 
416 	/*
417 	 * If we have more than one node, each cpu gets a copy of IDT
418 	 * local to its node. If this is a Pentium box, we use cpu 0's
419 	 * IDT. cpu 0's IDT has been made read-only to workaround the
420 	 * cmpxchgl register bug
421 	 */
422 	if (system_hardware.hd_nodes && x86_type != X86_TYPE_P5) {
423 #if !defined(__lint)
424 		ASSERT((sizeof (*CPU->cpu_idt) * NIDT) <= PAGESIZE);
425 #endif
426 		cp->cpu_idt = kmem_zalloc(PAGESIZE, KM_SLEEP);
427 		bcopy(CPU->cpu_idt, cp->cpu_idt, PAGESIZE);
428 	} else {
429 		cp->cpu_idt = CPU->cpu_idt;
430 	}
431 
432 	/*
433 	 * alloc space for cpuid info
434 	 */
435 	cpuid_alloc_space(cp);
436 #if !defined(__xpv)
437 	if (is_x86_feature(x86_featureset, X86FSET_MWAIT) &&
438 	    idle_cpu_prefer_mwait) {
439 		cp->cpu_m.mcpu_mwait = cpuid_mwait_alloc(cp);
440 		cp->cpu_m.mcpu_idle_cpu = cpu_idle_mwait;
441 	} else
442 #endif
443 		cp->cpu_m.mcpu_idle_cpu = cpu_idle;
444 
445 	init_cpu_info(cp);
446 
447 #if !defined(__xpv)
448 	init_cpu_id_gdt(cp);
449 #endif
450 
451 	/*
452 	 * alloc space for ucode_info
453 	 */
454 	ucode_alloc_space(cp);
455 	xc_init_cpu(cp);
456 	hat_cpu_online(cp);
457 
458 #ifdef TRAPTRACE
459 	/*
460 	 * If this is a TRAPTRACE kernel, allocate TRAPTRACE buffers
461 	 */
462 	ttc->ttc_first = (uintptr_t)kmem_zalloc(trap_trace_bufsize, KM_SLEEP);
463 	ttc->ttc_next = ttc->ttc_first;
464 	ttc->ttc_limit = ttc->ttc_first + trap_trace_bufsize;
465 #endif
466 
467 	/*
468 	 * Record that we have another CPU.
469 	 */
470 	/*
471 	 * Initialize the interrupt threads for this CPU
472 	 */
473 	cpu_intr_alloc(cp, NINTR_THREADS);
474 
475 	cp->cpu_flags = CPU_OFFLINE | CPU_QUIESCED | CPU_POWEROFF;
476 	cpu_set_state(cp);
477 
478 	/*
479 	 * Add CPU to list of available CPUs.  It'll be on the active list
480 	 * after mp_startup_common().
481 	 */
482 	cpu_add_unit(cp);
483 
484 	return (cp);
485 }
486 
487 /*
488  * Undo what was done in mp_cpu_configure_common
489  */
490 static void
491 mp_cpu_unconfigure_common(struct cpu *cp, int error)
492 {
493 	ASSERT(MUTEX_HELD(&cpu_lock));
494 
495 	/*
496 	 * Remove the CPU from the list of available CPUs.
497 	 */
498 	cpu_del_unit(cp->cpu_id);
499 
500 	if (error == ETIMEDOUT) {
501 		/*
502 		 * The cpu was started, but never *seemed* to run any
503 		 * code in the kernel; it's probably off spinning in its
504 		 * own private world, though with potential references to
505 		 * our kmem-allocated IDTs and GDTs (for example).
506 		 *
507 		 * Worse still, it may actually wake up some time later,
508 		 * so rather than guess what it might or might not do, we
509 		 * leave the fundamental data structures intact.
510 		 */
511 		cp->cpu_flags = 0;
512 		return;
513 	}
514 
515 	/*
516 	 * At this point, the only threads bound to this CPU should
517 	 * special per-cpu threads: it's idle thread, it's pause threads,
518 	 * and it's interrupt threads.  Clean these up.
519 	 */
520 	cpu_destroy_bound_threads(cp);
521 	cp->cpu_idle_thread = NULL;
522 
523 	/*
524 	 * Free the interrupt stack.
525 	 */
526 	segkp_release(segkp,
527 	    cp->cpu_intr_stack - (INTR_STACK_SIZE - SA(MINFRAME)));
528 	cp->cpu_intr_stack = NULL;
529 
530 #ifdef TRAPTRACE
531 	/*
532 	 * Discard the trap trace buffer
533 	 */
534 	{
535 		trap_trace_ctl_t *ttc = &trap_trace_ctl[cp->cpu_id];
536 
537 		kmem_free((void *)ttc->ttc_first, trap_trace_bufsize);
538 		ttc->ttc_first = NULL;
539 	}
540 #endif
541 
542 	hat_cpu_offline(cp);
543 
544 	ucode_free_space(cp);
545 
546 	/* Free CPU ID string and brand string. */
547 	if (cp->cpu_idstr) {
548 		kmem_free(cp->cpu_idstr, CPU_IDSTRLEN);
549 		cp->cpu_idstr = NULL;
550 	}
551 	if (cp->cpu_brandstr) {
552 		kmem_free(cp->cpu_brandstr, CPU_IDSTRLEN);
553 		cp->cpu_brandstr = NULL;
554 	}
555 
556 #if !defined(__xpv)
557 	if (cp->cpu_m.mcpu_mwait != NULL) {
558 		cpuid_mwait_free(cp);
559 		cp->cpu_m.mcpu_mwait = NULL;
560 	}
561 #endif
562 	cpuid_free_space(cp);
563 
564 	if (cp->cpu_idt != CPU->cpu_idt)
565 		kmem_free(cp->cpu_idt, PAGESIZE);
566 	cp->cpu_idt = NULL;
567 
568 	kmem_free(cp->cpu_gdt, PAGESIZE);
569 	cp->cpu_gdt = NULL;
570 
571 	if (cp->cpu_supp_freqs != NULL) {
572 		size_t len = strlen(cp->cpu_supp_freqs) + 1;
573 		kmem_free(cp->cpu_supp_freqs, len);
574 		cp->cpu_supp_freqs = NULL;
575 	}
576 
577 	teardown_vaddr_for_ppcopy(cp);
578 
579 	kcpc_hw_fini(cp);
580 
581 	cp->cpu_dispthread = NULL;
582 	cp->cpu_thread = NULL;	/* discarded by cpu_destroy_bound_threads() */
583 
584 	cpu_vm_data_destroy(cp);
585 
586 	xc_fini_cpu(cp);
587 	disp_cpu_fini(cp);
588 
589 	ASSERT(cp != CPU0);
590 	bzero(cp, sizeof (*cp));
591 	cp->cpu_next_free = cpu_free_list;
592 	cpu_free_list = cp;
593 }
594 
595 /*
596  * Apply workarounds for known errata, and warn about those that are absent.
597  *
598  * System vendors occasionally create configurations which contain different
599  * revisions of the CPUs that are almost but not exactly the same.  At the
600  * time of writing, this meant that their clock rates were the same, their
601  * feature sets were the same, but the required workaround were -not-
602  * necessarily the same.  So, this routine is invoked on -every- CPU soon
603  * after starting to make sure that the resulting system contains the most
604  * pessimal set of workarounds needed to cope with *any* of the CPUs in the
605  * system.
606  *
607  * workaround_errata is invoked early in mlsetup() for CPU 0, and in
608  * mp_startup_common() for all slave CPUs. Slaves process workaround_errata
609  * prior to acknowledging their readiness to the master, so this routine will
610  * never be executed by multiple CPUs in parallel, thus making updates to
611  * global data safe.
612  *
613  * These workarounds are based on Rev 3.57 of the Revision Guide for
614  * AMD Athlon(tm) 64 and AMD Opteron(tm) Processors, August 2005.
615  */
616 
617 #if defined(OPTERON_ERRATUM_88)
618 int opteron_erratum_88;		/* if non-zero -> at least one cpu has it */
619 #endif
620 
621 #if defined(OPTERON_ERRATUM_91)
622 int opteron_erratum_91;		/* if non-zero -> at least one cpu has it */
623 #endif
624 
625 #if defined(OPTERON_ERRATUM_93)
626 int opteron_erratum_93;		/* if non-zero -> at least one cpu has it */
627 #endif
628 
629 #if defined(OPTERON_ERRATUM_95)
630 int opteron_erratum_95;		/* if non-zero -> at least one cpu has it */
631 #endif
632 
633 #if defined(OPTERON_ERRATUM_100)
634 int opteron_erratum_100;	/* if non-zero -> at least one cpu has it */
635 #endif
636 
637 #if defined(OPTERON_ERRATUM_108)
638 int opteron_erratum_108;	/* if non-zero -> at least one cpu has it */
639 #endif
640 
641 #if defined(OPTERON_ERRATUM_109)
642 int opteron_erratum_109;	/* if non-zero -> at least one cpu has it */
643 #endif
644 
645 #if defined(OPTERON_ERRATUM_121)
646 int opteron_erratum_121;	/* if non-zero -> at least one cpu has it */
647 #endif
648 
649 #if defined(OPTERON_ERRATUM_122)
650 int opteron_erratum_122;	/* if non-zero -> at least one cpu has it */
651 #endif
652 
653 #if defined(OPTERON_ERRATUM_123)
654 int opteron_erratum_123;	/* if non-zero -> at least one cpu has it */
655 #endif
656 
657 #if defined(OPTERON_ERRATUM_131)
658 int opteron_erratum_131;	/* if non-zero -> at least one cpu has it */
659 #endif
660 
661 #if defined(OPTERON_WORKAROUND_6336786)
662 int opteron_workaround_6336786;	/* non-zero -> WA relevant and applied */
663 int opteron_workaround_6336786_UP = 0;	/* Not needed for UP */
664 #endif
665 
666 #if defined(OPTERON_WORKAROUND_6323525)
667 int opteron_workaround_6323525;	/* if non-zero -> at least one cpu has it */
668 #endif
669 
670 #if defined(OPTERON_ERRATUM_298)
671 int opteron_erratum_298;
672 #endif
673 
674 #if defined(OPTERON_ERRATUM_721)
675 int opteron_erratum_721;
676 #endif
677 
678 static void
679 workaround_warning(cpu_t *cp, uint_t erratum)
680 {
681 	cmn_err(CE_WARN, "cpu%d: no workaround for erratum %u",
682 	    cp->cpu_id, erratum);
683 }
684 
685 static void
686 workaround_applied(uint_t erratum)
687 {
688 	if (erratum > 1000000)
689 		cmn_err(CE_CONT, "?workaround applied for cpu issue #%d\n",
690 		    erratum);
691 	else
692 		cmn_err(CE_CONT, "?workaround applied for cpu erratum #%d\n",
693 		    erratum);
694 }
695 
696 static void
697 msr_warning(cpu_t *cp, const char *rw, uint_t msr, int error)
698 {
699 	cmn_err(CE_WARN, "cpu%d: couldn't %smsr 0x%x, error %d",
700 	    cp->cpu_id, rw, msr, error);
701 }
702 
703 /*
704  * Determine the number of nodes in a Hammer / Greyhound / Griffin family
705  * system.
706  */
707 static uint_t
708 opteron_get_nnodes(void)
709 {
710 	static uint_t nnodes = 0;
711 
712 	if (nnodes == 0) {
713 #ifdef	DEBUG
714 		uint_t family;
715 
716 		/*
717 		 * This routine uses a PCI config space based mechanism
718 		 * for retrieving the number of nodes in the system.
719 		 * Device 24, function 0, offset 0x60 as used here is not
720 		 * AMD processor architectural, and may not work on processor
721 		 * families other than those listed below.
722 		 *
723 		 * Callers of this routine must ensure that we're running on
724 		 * a processor which supports this mechanism.
725 		 * The assertion below is meant to catch calls on unsupported
726 		 * processors.
727 		 */
728 		family = cpuid_getfamily(CPU);
729 		ASSERT(family == 0xf || family == 0x10 || family == 0x11);
730 #endif	/* DEBUG */
731 
732 		/*
733 		 * Obtain the number of nodes in the system from
734 		 * bits [6:4] of the Node ID register on node 0.
735 		 *
736 		 * The actual node count is NodeID[6:4] + 1
737 		 *
738 		 * The Node ID register is accessed via function 0,
739 		 * offset 0x60. Node 0 is device 24.
740 		 */
741 		nnodes = ((pci_getl_func(0, 24, 0, 0x60) & 0x70) >> 4) + 1;
742 	}
743 	return (nnodes);
744 }
745 
746 uint_t
747 do_erratum_298(struct cpu *cpu)
748 {
749 	static int	osvwrc = -3;
750 	extern int	osvw_opteron_erratum(cpu_t *, uint_t);
751 
752 	/*
753 	 * L2 Eviction May Occur During Processor Operation To Set
754 	 * Accessed or Dirty Bit.
755 	 */
756 	if (osvwrc == -3) {
757 		osvwrc = osvw_opteron_erratum(cpu, 298);
758 	} else {
759 		/* osvw return codes should be consistent for all cpus */
760 		ASSERT(osvwrc == osvw_opteron_erratum(cpu, 298));
761 	}
762 
763 	switch (osvwrc) {
764 	case 0:		/* erratum is not present: do nothing */
765 		break;
766 	case 1:		/* erratum is present: BIOS workaround applied */
767 		/*
768 		 * check if workaround is actually in place and issue warning
769 		 * if not.
770 		 */
771 		if (((rdmsr(MSR_AMD_HWCR) & AMD_HWCR_TLBCACHEDIS) == 0) ||
772 		    ((rdmsr(MSR_AMD_BU_CFG) & AMD_BU_CFG_E298) == 0)) {
773 #if defined(OPTERON_ERRATUM_298)
774 			opteron_erratum_298++;
775 #else
776 			workaround_warning(cpu, 298);
777 			return (1);
778 #endif
779 		}
780 		break;
781 	case -1:	/* cannot determine via osvw: check cpuid */
782 		if ((cpuid_opteron_erratum(cpu, 298) > 0) &&
783 		    (((rdmsr(MSR_AMD_HWCR) & AMD_HWCR_TLBCACHEDIS) == 0) ||
784 		    ((rdmsr(MSR_AMD_BU_CFG) & AMD_BU_CFG_E298) == 0))) {
785 #if defined(OPTERON_ERRATUM_298)
786 			opteron_erratum_298++;
787 #else
788 			workaround_warning(cpu, 298);
789 			return (1);
790 #endif
791 		}
792 		break;
793 	}
794 	return (0);
795 }
796 
797 uint_t
798 workaround_errata(struct cpu *cpu)
799 {
800 	uint_t missing = 0;
801 
802 	ASSERT(cpu == CPU);
803 
804 	/*LINTED*/
805 	if (cpuid_opteron_erratum(cpu, 88) > 0) {
806 		/*
807 		 * SWAPGS May Fail To Read Correct GS Base
808 		 */
809 #if defined(OPTERON_ERRATUM_88)
810 		/*
811 		 * The workaround is an mfence in the relevant assembler code
812 		 */
813 		opteron_erratum_88++;
814 #else
815 		workaround_warning(cpu, 88);
816 		missing++;
817 #endif
818 	}
819 
820 	if (cpuid_opteron_erratum(cpu, 91) > 0) {
821 		/*
822 		 * Software Prefetches May Report A Page Fault
823 		 */
824 #if defined(OPTERON_ERRATUM_91)
825 		/*
826 		 * fix is in trap.c
827 		 */
828 		opteron_erratum_91++;
829 #else
830 		workaround_warning(cpu, 91);
831 		missing++;
832 #endif
833 	}
834 
835 	if (cpuid_opteron_erratum(cpu, 93) > 0) {
836 		/*
837 		 * RSM Auto-Halt Restart Returns to Incorrect RIP
838 		 */
839 #if defined(OPTERON_ERRATUM_93)
840 		/*
841 		 * fix is in trap.c
842 		 */
843 		opteron_erratum_93++;
844 #else
845 		workaround_warning(cpu, 93);
846 		missing++;
847 #endif
848 	}
849 
850 	/*LINTED*/
851 	if (cpuid_opteron_erratum(cpu, 95) > 0) {
852 		/*
853 		 * RET Instruction May Return to Incorrect EIP
854 		 */
855 #if defined(OPTERON_ERRATUM_95)
856 #if defined(_LP64)
857 		/*
858 		 * Workaround this by ensuring that 32-bit user code and
859 		 * 64-bit kernel code never occupy the same address
860 		 * range mod 4G.
861 		 */
862 		if (_userlimit32 > 0xc0000000ul)
863 			*(uintptr_t *)&_userlimit32 = 0xc0000000ul;
864 
865 		/*LINTED*/
866 		ASSERT((uint32_t)COREHEAP_BASE == 0xc0000000u);
867 		opteron_erratum_95++;
868 #endif	/* _LP64 */
869 #else
870 		workaround_warning(cpu, 95);
871 		missing++;
872 #endif
873 	}
874 
875 	if (cpuid_opteron_erratum(cpu, 100) > 0) {
876 		/*
877 		 * Compatibility Mode Branches Transfer to Illegal Address
878 		 */
879 #if defined(OPTERON_ERRATUM_100)
880 		/*
881 		 * fix is in trap.c
882 		 */
883 		opteron_erratum_100++;
884 #else
885 		workaround_warning(cpu, 100);
886 		missing++;
887 #endif
888 	}
889 
890 	/*LINTED*/
891 	if (cpuid_opteron_erratum(cpu, 108) > 0) {
892 		/*
893 		 * CPUID Instruction May Return Incorrect Model Number In
894 		 * Some Processors
895 		 */
896 #if defined(OPTERON_ERRATUM_108)
897 		/*
898 		 * (Our cpuid-handling code corrects the model number on
899 		 * those processors)
900 		 */
901 #else
902 		workaround_warning(cpu, 108);
903 		missing++;
904 #endif
905 	}
906 
907 	/*LINTED*/
908 	if (cpuid_opteron_erratum(cpu, 109) > 0) do {
909 		/*
910 		 * Certain Reverse REP MOVS May Produce Unpredictable Behavior
911 		 */
912 #if defined(OPTERON_ERRATUM_109)
913 		/*
914 		 * The "workaround" is to print a warning to upgrade the BIOS
915 		 */
916 		uint64_t value;
917 		const uint_t msr = MSR_AMD_PATCHLEVEL;
918 		int err;
919 
920 		if ((err = checked_rdmsr(msr, &value)) != 0) {
921 			msr_warning(cpu, "rd", msr, err);
922 			workaround_warning(cpu, 109);
923 			missing++;
924 		}
925 		if (value == 0)
926 			opteron_erratum_109++;
927 #else
928 		workaround_warning(cpu, 109);
929 		missing++;
930 #endif
931 	/*CONSTANTCONDITION*/
932 	} while (0);
933 
934 	/*LINTED*/
935 	if (cpuid_opteron_erratum(cpu, 121) > 0) {
936 		/*
937 		 * Sequential Execution Across Non_Canonical Boundary Caused
938 		 * Processor Hang
939 		 */
940 #if defined(OPTERON_ERRATUM_121)
941 #if defined(_LP64)
942 		/*
943 		 * Erratum 121 is only present in long (64 bit) mode.
944 		 * Workaround is to include the page immediately before the
945 		 * va hole to eliminate the possibility of system hangs due to
946 		 * sequential execution across the va hole boundary.
947 		 */
948 		if (opteron_erratum_121)
949 			opteron_erratum_121++;
950 		else {
951 			if (hole_start) {
952 				hole_start -= PAGESIZE;
953 			} else {
954 				/*
955 				 * hole_start not yet initialized by
956 				 * mmu_init. Initialize hole_start
957 				 * with value to be subtracted.
958 				 */
959 				hole_start = PAGESIZE;
960 			}
961 			opteron_erratum_121++;
962 		}
963 #endif	/* _LP64 */
964 #else
965 		workaround_warning(cpu, 121);
966 		missing++;
967 #endif
968 	}
969 
970 	/*LINTED*/
971 	if (cpuid_opteron_erratum(cpu, 122) > 0) do {
972 		/*
973 		 * TLB Flush Filter May Cause Coherency Problem in
974 		 * Multiprocessor Systems
975 		 */
976 #if defined(OPTERON_ERRATUM_122)
977 		uint64_t value;
978 		const uint_t msr = MSR_AMD_HWCR;
979 		int error;
980 
981 		/*
982 		 * Erratum 122 is only present in MP configurations (multi-core
983 		 * or multi-processor).
984 		 */
985 #if defined(__xpv)
986 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
987 			break;
988 		if (!opteron_erratum_122 && xpv_nr_phys_cpus() == 1)
989 			break;
990 #else
991 		if (!opteron_erratum_122 && opteron_get_nnodes() == 1 &&
992 		    cpuid_get_ncpu_per_chip(cpu) == 1)
993 			break;
994 #endif
995 		/* disable TLB Flush Filter */
996 
997 		if ((error = checked_rdmsr(msr, &value)) != 0) {
998 			msr_warning(cpu, "rd", msr, error);
999 			workaround_warning(cpu, 122);
1000 			missing++;
1001 		} else {
1002 			value |= (uint64_t)AMD_HWCR_FFDIS;
1003 			if ((error = checked_wrmsr(msr, value)) != 0) {
1004 				msr_warning(cpu, "wr", msr, error);
1005 				workaround_warning(cpu, 122);
1006 				missing++;
1007 			}
1008 		}
1009 		opteron_erratum_122++;
1010 #else
1011 		workaround_warning(cpu, 122);
1012 		missing++;
1013 #endif
1014 	/*CONSTANTCONDITION*/
1015 	} while (0);
1016 
1017 	/*LINTED*/
1018 	if (cpuid_opteron_erratum(cpu, 123) > 0) do {
1019 		/*
1020 		 * Bypassed Reads May Cause Data Corruption of System Hang in
1021 		 * Dual Core Processors
1022 		 */
1023 #if defined(OPTERON_ERRATUM_123)
1024 		uint64_t value;
1025 		const uint_t msr = MSR_AMD_PATCHLEVEL;
1026 		int err;
1027 
1028 		/*
1029 		 * Erratum 123 applies only to multi-core cpus.
1030 		 */
1031 		if (cpuid_get_ncpu_per_chip(cpu) < 2)
1032 			break;
1033 #if defined(__xpv)
1034 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
1035 			break;
1036 #endif
1037 		/*
1038 		 * The "workaround" is to print a warning to upgrade the BIOS
1039 		 */
1040 		if ((err = checked_rdmsr(msr, &value)) != 0) {
1041 			msr_warning(cpu, "rd", msr, err);
1042 			workaround_warning(cpu, 123);
1043 			missing++;
1044 		}
1045 		if (value == 0)
1046 			opteron_erratum_123++;
1047 #else
1048 		workaround_warning(cpu, 123);
1049 		missing++;
1050 
1051 #endif
1052 	/*CONSTANTCONDITION*/
1053 	} while (0);
1054 
1055 	/*LINTED*/
1056 	if (cpuid_opteron_erratum(cpu, 131) > 0) do {
1057 		/*
1058 		 * Multiprocessor Systems with Four or More Cores May Deadlock
1059 		 * Waiting for a Probe Response
1060 		 */
1061 #if defined(OPTERON_ERRATUM_131)
1062 		uint64_t nbcfg;
1063 		const uint_t msr = MSR_AMD_NB_CFG;
1064 		const uint64_t wabits =
1065 		    AMD_NB_CFG_SRQ_HEARTBEAT | AMD_NB_CFG_SRQ_SPR;
1066 		int error;
1067 
1068 		/*
1069 		 * Erratum 131 applies to any system with four or more cores.
1070 		 */
1071 		if (opteron_erratum_131)
1072 			break;
1073 #if defined(__xpv)
1074 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
1075 			break;
1076 		if (xpv_nr_phys_cpus() < 4)
1077 			break;
1078 #else
1079 		if (opteron_get_nnodes() * cpuid_get_ncpu_per_chip(cpu) < 4)
1080 			break;
1081 #endif
1082 		/*
1083 		 * Print a warning if neither of the workarounds for
1084 		 * erratum 131 is present.
1085 		 */
1086 		if ((error = checked_rdmsr(msr, &nbcfg)) != 0) {
1087 			msr_warning(cpu, "rd", msr, error);
1088 			workaround_warning(cpu, 131);
1089 			missing++;
1090 		} else if ((nbcfg & wabits) == 0) {
1091 			opteron_erratum_131++;
1092 		} else {
1093 			/* cannot have both workarounds set */
1094 			ASSERT((nbcfg & wabits) != wabits);
1095 		}
1096 #else
1097 		workaround_warning(cpu, 131);
1098 		missing++;
1099 #endif
1100 	/*CONSTANTCONDITION*/
1101 	} while (0);
1102 
1103 	/*
1104 	 * This isn't really an erratum, but for convenience the
1105 	 * detection/workaround code lives here and in cpuid_opteron_erratum.
1106 	 */
1107 	if (cpuid_opteron_erratum(cpu, 6336786) > 0) {
1108 #if defined(OPTERON_WORKAROUND_6336786)
1109 		/*
1110 		 * Disable C1-Clock ramping on multi-core/multi-processor
1111 		 * K8 platforms to guard against TSC drift.
1112 		 */
1113 		if (opteron_workaround_6336786) {
1114 			opteron_workaround_6336786++;
1115 #if defined(__xpv)
1116 		} else if ((DOMAIN_IS_INITDOMAIN(xen_info) &&
1117 		    xpv_nr_phys_cpus() > 1) ||
1118 		    opteron_workaround_6336786_UP) {
1119 			/*
1120 			 * XXPV	Hmm.  We can't walk the Northbridges on
1121 			 *	the hypervisor; so just complain and drive
1122 			 *	on.  This probably needs to be fixed in
1123 			 *	the hypervisor itself.
1124 			 */
1125 			opteron_workaround_6336786++;
1126 			workaround_warning(cpu, 6336786);
1127 #else	/* __xpv */
1128 		} else if ((opteron_get_nnodes() *
1129 		    cpuid_get_ncpu_per_chip(cpu) > 1) ||
1130 		    opteron_workaround_6336786_UP) {
1131 
1132 			uint_t	node, nnodes;
1133 			uint8_t data;
1134 
1135 			nnodes = opteron_get_nnodes();
1136 			for (node = 0; node < nnodes; node++) {
1137 				/*
1138 				 * Clear PMM7[1:0] (function 3, offset 0x87)
1139 				 * Northbridge device is the node id + 24.
1140 				 */
1141 				data = pci_getb_func(0, node + 24, 3, 0x87);
1142 				data &= 0xFC;
1143 				pci_putb_func(0, node + 24, 3, 0x87, data);
1144 			}
1145 			opteron_workaround_6336786++;
1146 #endif	/* __xpv */
1147 		}
1148 #else
1149 		workaround_warning(cpu, 6336786);
1150 		missing++;
1151 #endif
1152 	}
1153 
1154 	/*LINTED*/
1155 	/*
1156 	 * Mutex primitives don't work as expected.
1157 	 */
1158 	if (cpuid_opteron_erratum(cpu, 6323525) > 0) {
1159 #if defined(OPTERON_WORKAROUND_6323525)
1160 		/*
1161 		 * This problem only occurs with 2 or more cores. If bit in
1162 		 * MSR_AMD_BU_CFG set, then not applicable. The workaround
1163 		 * is to patch the semaphone routines with the lfence
1164 		 * instruction to provide necessary load memory barrier with
1165 		 * possible subsequent read-modify-write ops.
1166 		 *
1167 		 * It is too early in boot to call the patch routine so
1168 		 * set erratum variable to be done in startup_end().
1169 		 */
1170 		if (opteron_workaround_6323525) {
1171 			opteron_workaround_6323525++;
1172 #if defined(__xpv)
1173 		} else if (is_x86_feature(x86_featureset, X86FSET_SSE2)) {
1174 			if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1175 				/*
1176 				 * XXPV	Use dom0_msr here when extended
1177 				 *	operations are supported?
1178 				 */
1179 				if (xpv_nr_phys_cpus() > 1)
1180 					opteron_workaround_6323525++;
1181 			} else {
1182 				/*
1183 				 * We have no way to tell how many physical
1184 				 * cpus there are, or even if this processor
1185 				 * has the problem, so enable the workaround
1186 				 * unconditionally (at some performance cost).
1187 				 */
1188 				opteron_workaround_6323525++;
1189 			}
1190 #else	/* __xpv */
1191 		} else if (is_x86_feature(x86_featureset, X86FSET_SSE2) &&
1192 		    ((opteron_get_nnodes() *
1193 		    cpuid_get_ncpu_per_chip(cpu)) > 1)) {
1194 			if ((xrdmsr(MSR_AMD_BU_CFG) & (UINT64_C(1) << 33)) == 0)
1195 				opteron_workaround_6323525++;
1196 #endif	/* __xpv */
1197 		}
1198 #else
1199 		workaround_warning(cpu, 6323525);
1200 		missing++;
1201 #endif
1202 	}
1203 
1204 	missing += do_erratum_298(cpu);
1205 
1206 	if (cpuid_opteron_erratum(cpu, 721) > 0) {
1207 #if defined(OPTERON_ERRATUM_721)
1208 		on_trap_data_t otd;
1209 
1210 		if (!on_trap(&otd, OT_DATA_ACCESS))
1211 			wrmsr(MSR_AMD_DE_CFG,
1212 			    rdmsr(MSR_AMD_DE_CFG) | AMD_DE_CFG_E721);
1213 		no_trap();
1214 
1215 		opteron_erratum_721++;
1216 #else
1217 		workaround_warning(cpu, 721);
1218 		missing++;
1219 #endif
1220 	}
1221 
1222 #ifdef __xpv
1223 	return (0);
1224 #else
1225 	return (missing);
1226 #endif
1227 }
1228 
1229 void
1230 workaround_errata_end()
1231 {
1232 #if defined(OPTERON_ERRATUM_88)
1233 	if (opteron_erratum_88)
1234 		workaround_applied(88);
1235 #endif
1236 #if defined(OPTERON_ERRATUM_91)
1237 	if (opteron_erratum_91)
1238 		workaround_applied(91);
1239 #endif
1240 #if defined(OPTERON_ERRATUM_93)
1241 	if (opteron_erratum_93)
1242 		workaround_applied(93);
1243 #endif
1244 #if defined(OPTERON_ERRATUM_95)
1245 	if (opteron_erratum_95)
1246 		workaround_applied(95);
1247 #endif
1248 #if defined(OPTERON_ERRATUM_100)
1249 	if (opteron_erratum_100)
1250 		workaround_applied(100);
1251 #endif
1252 #if defined(OPTERON_ERRATUM_108)
1253 	if (opteron_erratum_108)
1254 		workaround_applied(108);
1255 #endif
1256 #if defined(OPTERON_ERRATUM_109)
1257 	if (opteron_erratum_109) {
1258 		cmn_err(CE_WARN,
1259 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1260 		    " processor\nerratum 109 was not detected; updating your"
1261 		    " system's BIOS to a version\ncontaining this"
1262 		    " microcode patch is HIGHLY recommended or erroneous"
1263 		    " system\noperation may occur.\n");
1264 	}
1265 #endif
1266 #if defined(OPTERON_ERRATUM_121)
1267 	if (opteron_erratum_121)
1268 		workaround_applied(121);
1269 #endif
1270 #if defined(OPTERON_ERRATUM_122)
1271 	if (opteron_erratum_122)
1272 		workaround_applied(122);
1273 #endif
1274 #if defined(OPTERON_ERRATUM_123)
1275 	if (opteron_erratum_123) {
1276 		cmn_err(CE_WARN,
1277 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1278 		    " processor\nerratum 123 was not detected; updating your"
1279 		    " system's BIOS to a version\ncontaining this"
1280 		    " microcode patch is HIGHLY recommended or erroneous"
1281 		    " system\noperation may occur.\n");
1282 	}
1283 #endif
1284 #if defined(OPTERON_ERRATUM_131)
1285 	if (opteron_erratum_131) {
1286 		cmn_err(CE_WARN,
1287 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1288 		    " processor\nerratum 131 was not detected; updating your"
1289 		    " system's BIOS to a version\ncontaining this"
1290 		    " microcode patch is HIGHLY recommended or erroneous"
1291 		    " system\noperation may occur.\n");
1292 	}
1293 #endif
1294 #if defined(OPTERON_WORKAROUND_6336786)
1295 	if (opteron_workaround_6336786)
1296 		workaround_applied(6336786);
1297 #endif
1298 #if defined(OPTERON_WORKAROUND_6323525)
1299 	if (opteron_workaround_6323525)
1300 		workaround_applied(6323525);
1301 #endif
1302 #if defined(OPTERON_ERRATUM_298)
1303 	if (opteron_erratum_298) {
1304 		cmn_err(CE_WARN,
1305 		    "BIOS microcode patch for AMD 64/Opteron(tm)"
1306 		    " processor\nerratum 298 was not detected; updating your"
1307 		    " system's BIOS to a version\ncontaining this"
1308 		    " microcode patch is HIGHLY recommended or erroneous"
1309 		    " system\noperation may occur.\n");
1310 	}
1311 #endif
1312 #if defined(OPTERON_ERRATUM_721)
1313 	if (opteron_erratum_721)
1314 		workaround_applied(721);
1315 #endif
1316 }
1317 
1318 /*
1319  * The procset_slave and procset_master are used to synchronize
1320  * between the control CPU and the target CPU when starting CPUs.
1321  */
1322 static cpuset_t procset_slave, procset_master;
1323 
1324 static void
1325 mp_startup_wait(cpuset_t *sp, processorid_t cpuid)
1326 {
1327 	cpuset_t tempset;
1328 
1329 	for (tempset = *sp; !CPU_IN_SET(tempset, cpuid);
1330 	    tempset = *(volatile cpuset_t *)sp) {
1331 		SMT_PAUSE();
1332 	}
1333 	CPUSET_ATOMIC_DEL(*(cpuset_t *)sp, cpuid);
1334 }
1335 
1336 static void
1337 mp_startup_signal(cpuset_t *sp, processorid_t cpuid)
1338 {
1339 	cpuset_t tempset;
1340 
1341 	CPUSET_ATOMIC_ADD(*(cpuset_t *)sp, cpuid);
1342 	for (tempset = *sp; CPU_IN_SET(tempset, cpuid);
1343 	    tempset = *(volatile cpuset_t *)sp) {
1344 		SMT_PAUSE();
1345 	}
1346 }
1347 
1348 int
1349 mp_start_cpu_common(cpu_t *cp, boolean_t boot)
1350 {
1351 	_NOTE(ARGUNUSED(boot));
1352 
1353 	void *ctx;
1354 	int delays;
1355 	int error = 0;
1356 	cpuset_t tempset;
1357 	processorid_t cpuid;
1358 #ifndef __xpv
1359 	extern void cpupm_init(cpu_t *);
1360 #endif
1361 
1362 	ASSERT(cp != NULL);
1363 	cpuid = cp->cpu_id;
1364 	ctx = mach_cpucontext_alloc(cp);
1365 	if (ctx == NULL) {
1366 		cmn_err(CE_WARN,
1367 		    "cpu%d: failed to allocate context", cp->cpu_id);
1368 		return (EAGAIN);
1369 	}
1370 	error = mach_cpu_start(cp, ctx);
1371 	if (error != 0) {
1372 		cmn_err(CE_WARN,
1373 		    "cpu%d: failed to start, error %d", cp->cpu_id, error);
1374 		mach_cpucontext_free(cp, ctx, error);
1375 		return (error);
1376 	}
1377 
1378 	for (delays = 0, tempset = procset_slave; !CPU_IN_SET(tempset, cpuid);
1379 	    delays++) {
1380 		if (delays == 500) {
1381 			/*
1382 			 * After five seconds, things are probably looking
1383 			 * a bit bleak - explain the hang.
1384 			 */
1385 			cmn_err(CE_NOTE, "cpu%d: started, "
1386 			    "but not running in the kernel yet", cpuid);
1387 		} else if (delays > 2000) {
1388 			/*
1389 			 * We waited at least 20 seconds, bail ..
1390 			 */
1391 			error = ETIMEDOUT;
1392 			cmn_err(CE_WARN, "cpu%d: timed out", cpuid);
1393 			mach_cpucontext_free(cp, ctx, error);
1394 			return (error);
1395 		}
1396 
1397 		/*
1398 		 * wait at least 10ms, then check again..
1399 		 */
1400 		delay(USEC_TO_TICK_ROUNDUP(10000));
1401 		tempset = *((volatile cpuset_t *)&procset_slave);
1402 	}
1403 	CPUSET_ATOMIC_DEL(procset_slave, cpuid);
1404 
1405 	mach_cpucontext_free(cp, ctx, 0);
1406 
1407 #ifndef __xpv
1408 	if (tsc_gethrtime_enable)
1409 		tsc_sync_master(cpuid);
1410 #endif
1411 
1412 	if (dtrace_cpu_init != NULL) {
1413 		(*dtrace_cpu_init)(cpuid);
1414 	}
1415 
1416 	/*
1417 	 * During CPU DR operations, the cpu_lock is held by current
1418 	 * (the control) thread. We can't release the cpu_lock here
1419 	 * because that will break the CPU DR logic.
1420 	 * On the other hand, CPUPM and processor group initialization
1421 	 * routines need to access the cpu_lock. So we invoke those
1422 	 * routines here on behalf of mp_startup_common().
1423 	 *
1424 	 * CPUPM and processor group initialization routines depend
1425 	 * on the cpuid probing results. Wait for mp_startup_common()
1426 	 * to signal that cpuid probing is done.
1427 	 */
1428 	mp_startup_wait(&procset_slave, cpuid);
1429 #ifndef __xpv
1430 	cpupm_init(cp);
1431 #endif
1432 	(void) pg_cpu_init(cp, B_FALSE);
1433 	cpu_set_state(cp);
1434 	mp_startup_signal(&procset_master, cpuid);
1435 
1436 	return (0);
1437 }
1438 
1439 /*
1440  * Start a single cpu, assuming that the kernel context is available
1441  * to successfully start another cpu.
1442  *
1443  * (For example, real mode code is mapped into the right place
1444  * in memory and is ready to be run.)
1445  */
1446 int
1447 start_cpu(processorid_t who)
1448 {
1449 	cpu_t *cp;
1450 	int error = 0;
1451 	cpuset_t tempset;
1452 
1453 	ASSERT(who != 0);
1454 
1455 	/*
1456 	 * Check if there's at least a Mbyte of kmem available
1457 	 * before attempting to start the cpu.
1458 	 */
1459 	if (kmem_avail() < 1024 * 1024) {
1460 		/*
1461 		 * Kick off a reap in case that helps us with
1462 		 * later attempts ..
1463 		 */
1464 		kmem_reap();
1465 		return (ENOMEM);
1466 	}
1467 
1468 	/*
1469 	 * First configure cpu.
1470 	 */
1471 	cp = mp_cpu_configure_common(who, B_TRUE);
1472 	ASSERT(cp != NULL);
1473 
1474 	/*
1475 	 * Then start cpu.
1476 	 */
1477 	error = mp_start_cpu_common(cp, B_TRUE);
1478 	if (error != 0) {
1479 		mp_cpu_unconfigure_common(cp, error);
1480 		return (error);
1481 	}
1482 
1483 	mutex_exit(&cpu_lock);
1484 	tempset = cpu_ready_set;
1485 	while (!CPU_IN_SET(tempset, who)) {
1486 		drv_usecwait(1);
1487 		tempset = *((volatile cpuset_t *)&cpu_ready_set);
1488 	}
1489 	mutex_enter(&cpu_lock);
1490 
1491 	return (0);
1492 }
1493 
1494 void
1495 start_other_cpus(int cprboot)
1496 {
1497 	_NOTE(ARGUNUSED(cprboot));
1498 
1499 	uint_t who;
1500 	uint_t bootcpuid = 0;
1501 
1502 	/*
1503 	 * Initialize our own cpu_info.
1504 	 */
1505 	init_cpu_info(CPU);
1506 
1507 #if !defined(__xpv)
1508 	init_cpu_id_gdt(CPU);
1509 #endif
1510 
1511 	cmn_err(CE_CONT, "?cpu%d: %s\n", CPU->cpu_id, CPU->cpu_idstr);
1512 	cmn_err(CE_CONT, "?cpu%d: %s\n", CPU->cpu_id, CPU->cpu_brandstr);
1513 
1514 	/*
1515 	 * Initialize our syscall handlers
1516 	 */
1517 	init_cpu_syscall(CPU);
1518 
1519 	/*
1520 	 * Take the boot cpu out of the mp_cpus set because we know
1521 	 * it's already running.  Add it to the cpu_ready_set for
1522 	 * precisely the same reason.
1523 	 */
1524 	CPUSET_DEL(mp_cpus, bootcpuid);
1525 	CPUSET_ADD(cpu_ready_set, bootcpuid);
1526 
1527 	/*
1528 	 * skip the rest of this if
1529 	 * . only 1 cpu dectected and system isn't hotplug-capable
1530 	 * . not using MP
1531 	 */
1532 	if ((CPUSET_ISNULL(mp_cpus) && plat_dr_support_cpu() == 0) ||
1533 	    use_mp == 0) {
1534 		if (use_mp == 0)
1535 			cmn_err(CE_CONT, "?***** Not in MP mode\n");
1536 		goto done;
1537 	}
1538 
1539 	/*
1540 	 * perform such initialization as is needed
1541 	 * to be able to take CPUs on- and off-line.
1542 	 */
1543 	cpu_pause_init();
1544 
1545 	xc_init_cpu(CPU);		/* initialize processor crosscalls */
1546 
1547 	if (mach_cpucontext_init() != 0)
1548 		goto done;
1549 
1550 	flushes_require_xcalls = 1;
1551 
1552 	/*
1553 	 * We lock our affinity to the master CPU to ensure that all slave CPUs
1554 	 * do their TSC syncs with the same CPU.
1555 	 */
1556 	affinity_set(CPU_CURRENT);
1557 
1558 	for (who = 0; who < NCPU; who++) {
1559 		if (!CPU_IN_SET(mp_cpus, who))
1560 			continue;
1561 		ASSERT(who != bootcpuid);
1562 
1563 		mutex_enter(&cpu_lock);
1564 		if (start_cpu(who) != 0)
1565 			CPUSET_DEL(mp_cpus, who);
1566 		cpu_state_change_notify(who, CPU_SETUP);
1567 		mutex_exit(&cpu_lock);
1568 	}
1569 
1570 	/* Free the space allocated to hold the microcode file */
1571 	ucode_cleanup();
1572 
1573 	affinity_clear();
1574 
1575 	mach_cpucontext_fini();
1576 
1577 done:
1578 	if (get_hwenv() == HW_NATIVE)
1579 		workaround_errata_end();
1580 	cmi_post_mpstartup();
1581 
1582 	if (use_mp && ncpus != boot_max_ncpus) {
1583 		cmn_err(CE_NOTE,
1584 		    "System detected %d cpus, but "
1585 		    "only %d cpu(s) were enabled during boot.",
1586 		    boot_max_ncpus, ncpus);
1587 		cmn_err(CE_NOTE,
1588 		    "Use \"boot-ncpus\" parameter to enable more CPU(s). "
1589 		    "See eeprom(1M).");
1590 	}
1591 }
1592 
1593 int
1594 mp_cpu_configure(int cpuid)
1595 {
1596 	cpu_t *cp;
1597 
1598 	if (use_mp == 0 || plat_dr_support_cpu() == 0) {
1599 		return (ENOTSUP);
1600 	}
1601 
1602 	cp = cpu_get(cpuid);
1603 	if (cp != NULL) {
1604 		return (EALREADY);
1605 	}
1606 
1607 	/*
1608 	 * Check if there's at least a Mbyte of kmem available
1609 	 * before attempting to start the cpu.
1610 	 */
1611 	if (kmem_avail() < 1024 * 1024) {
1612 		/*
1613 		 * Kick off a reap in case that helps us with
1614 		 * later attempts ..
1615 		 */
1616 		kmem_reap();
1617 		return (ENOMEM);
1618 	}
1619 
1620 	cp = mp_cpu_configure_common(cpuid, B_FALSE);
1621 	ASSERT(cp != NULL && cpu_get(cpuid) == cp);
1622 
1623 	return (cp != NULL ? 0 : EAGAIN);
1624 }
1625 
1626 int
1627 mp_cpu_unconfigure(int cpuid)
1628 {
1629 	cpu_t *cp;
1630 
1631 	if (use_mp == 0 || plat_dr_support_cpu() == 0) {
1632 		return (ENOTSUP);
1633 	} else if (cpuid < 0 || cpuid >= max_ncpus) {
1634 		return (EINVAL);
1635 	}
1636 
1637 	cp = cpu_get(cpuid);
1638 	if (cp == NULL) {
1639 		return (ENODEV);
1640 	}
1641 	mp_cpu_unconfigure_common(cp, 0);
1642 
1643 	return (0);
1644 }
1645 
1646 /*
1647  * Startup function for 'other' CPUs (besides boot cpu).
1648  * Called from real_mode_start.
1649  *
1650  * WARNING: until CPU_READY is set, mp_startup_common and routines called by
1651  * mp_startup_common should not call routines (e.g. kmem_free) that could call
1652  * hat_unload which requires CPU_READY to be set.
1653  */
1654 static void
1655 mp_startup_common(boolean_t boot)
1656 {
1657 	cpu_t *cp = CPU;
1658 	uchar_t new_x86_featureset[BT_SIZEOFMAP(NUM_X86_FEATURES)];
1659 	extern void cpu_event_init_cpu(cpu_t *);
1660 
1661 	/*
1662 	 * We need to get TSC on this proc synced (i.e., any delta
1663 	 * from cpu0 accounted for) as soon as we can, because many
1664 	 * many things use gethrtime/pc_gethrestime, including
1665 	 * interrupts, cmn_err, etc.  Before we can do that, we want to
1666 	 * clear TSC if we're on a buggy Sandy/Ivy Bridge CPU, so do that
1667 	 * right away.
1668 	 */
1669 	bzero(new_x86_featureset, BT_SIZEOFMAP(NUM_X86_FEATURES));
1670 	cpuid_pass1(cp, new_x86_featureset);
1671 
1672 	if (boot && get_hwenv() == HW_NATIVE &&
1673 	    cpuid_getvendor(CPU) == X86_VENDOR_Intel &&
1674 	    cpuid_getfamily(CPU) == 6 &&
1675 	    (cpuid_getmodel(CPU) == 0x2d || cpuid_getmodel(CPU) == 0x3e) &&
1676 	    is_x86_feature(new_x86_featureset, X86FSET_TSC)) {
1677 		(void) wrmsr(REG_TSC, 0UL);
1678 	}
1679 
1680 	/* Let the control CPU continue into tsc_sync_master() */
1681 	mp_startup_signal(&procset_slave, cp->cpu_id);
1682 
1683 #ifndef __xpv
1684 	if (tsc_gethrtime_enable)
1685 		tsc_sync_slave();
1686 #endif
1687 
1688 	/*
1689 	 * Once this was done from assembly, but it's safer here; if
1690 	 * it blocks, we need to be able to swtch() to and from, and
1691 	 * since we get here by calling t_pc, we need to do that call
1692 	 * before swtch() overwrites it.
1693 	 */
1694 	(void) (*ap_mlsetup)();
1695 
1696 #ifndef __xpv
1697 	/*
1698 	 * Program this cpu's PAT
1699 	 */
1700 	pat_sync();
1701 #endif
1702 
1703 	/*
1704 	 * Set up TSC_AUX to contain the cpuid for this processor
1705 	 * for the rdtscp instruction.
1706 	 */
1707 	if (is_x86_feature(x86_featureset, X86FSET_TSCP))
1708 		(void) wrmsr(MSR_AMD_TSCAUX, cp->cpu_id);
1709 
1710 	/*
1711 	 * Initialize this CPU's syscall handlers
1712 	 */
1713 	init_cpu_syscall(cp);
1714 
1715 	/*
1716 	 * Enable interrupts with spl set to LOCK_LEVEL. LOCK_LEVEL is the
1717 	 * highest level at which a routine is permitted to block on
1718 	 * an adaptive mutex (allows for cpu poke interrupt in case
1719 	 * the cpu is blocked on a mutex and halts). Setting LOCK_LEVEL blocks
1720 	 * device interrupts that may end up in the hat layer issuing cross
1721 	 * calls before CPU_READY is set.
1722 	 */
1723 	splx(ipltospl(LOCK_LEVEL));
1724 	sti();
1725 
1726 	/*
1727 	 * Do a sanity check to make sure this new CPU is a sane thing
1728 	 * to add to the collection of processors running this system.
1729 	 *
1730 	 * XXX	Clearly this needs to get more sophisticated, if x86
1731 	 * systems start to get built out of heterogenous CPUs; as is
1732 	 * likely to happen once the number of processors in a configuration
1733 	 * gets large enough.
1734 	 */
1735 	if (compare_x86_featureset(x86_featureset, new_x86_featureset) ==
1736 	    B_FALSE) {
1737 		cmn_err(CE_CONT, "cpu%d: featureset\n", cp->cpu_id);
1738 		print_x86_featureset(new_x86_featureset);
1739 		cmn_err(CE_WARN, "cpu%d feature mismatch", cp->cpu_id);
1740 	}
1741 
1742 	/*
1743 	 * We do not support cpus with mixed monitor/mwait support if the
1744 	 * boot cpu supports monitor/mwait.
1745 	 */
1746 	if (is_x86_feature(x86_featureset, X86FSET_MWAIT) !=
1747 	    is_x86_feature(new_x86_featureset, X86FSET_MWAIT))
1748 		panic("unsupported mixed cpu monitor/mwait support detected");
1749 
1750 	/*
1751 	 * We could be more sophisticated here, and just mark the CPU
1752 	 * as "faulted" but at this point we'll opt for the easier
1753 	 * answer of dying horribly.  Provided the boot cpu is ok,
1754 	 * the system can be recovered by booting with use_mp set to zero.
1755 	 */
1756 	if (workaround_errata(cp) != 0)
1757 		panic("critical workaround(s) missing for cpu%d", cp->cpu_id);
1758 
1759 	/*
1760 	 * We can touch cpu_flags here without acquiring the cpu_lock here
1761 	 * because the cpu_lock is held by the control CPU which is running
1762 	 * mp_start_cpu_common().
1763 	 * Need to clear CPU_QUIESCED flag before calling any function which
1764 	 * may cause thread context switching, such as kmem_alloc() etc.
1765 	 * The idle thread checks for CPU_QUIESCED flag and loops for ever if
1766 	 * it's set. So the startup thread may have no chance to switch back
1767 	 * again if it's switched away with CPU_QUIESCED set.
1768 	 */
1769 	cp->cpu_flags &= ~(CPU_POWEROFF | CPU_QUIESCED);
1770 
1771 	/*
1772 	 * Setup this processor for XSAVE.
1773 	 */
1774 	if (fp_save_mech == FP_XSAVE) {
1775 		xsave_setup_msr(cp);
1776 	}
1777 
1778 	cpuid_pass2(cp);
1779 	cpuid_pass3(cp);
1780 	cpuid_pass4(cp, NULL);
1781 
1782 	/*
1783 	 * Correct cpu_idstr and cpu_brandstr on target CPU after
1784 	 * cpuid_pass1() is done.
1785 	 */
1786 	(void) cpuid_getidstr(cp, cp->cpu_idstr, CPU_IDSTRLEN);
1787 	(void) cpuid_getbrandstr(cp, cp->cpu_brandstr, CPU_IDSTRLEN);
1788 
1789 	cp->cpu_flags |= CPU_RUNNING | CPU_READY | CPU_EXISTS;
1790 
1791 	post_startup_cpu_fixups();
1792 
1793 	cpu_event_init_cpu(cp);
1794 
1795 	/*
1796 	 * Enable preemption here so that contention for any locks acquired
1797 	 * later in mp_startup_common may be preempted if the thread owning
1798 	 * those locks is continuously executing on other CPUs (for example,
1799 	 * this CPU must be preemptible to allow other CPUs to pause it during
1800 	 * their startup phases).  It's safe to enable preemption here because
1801 	 * the CPU state is pretty-much fully constructed.
1802 	 */
1803 	curthread->t_preempt = 0;
1804 
1805 	/* The base spl should still be at LOCK LEVEL here */
1806 	ASSERT(cp->cpu_base_spl == ipltospl(LOCK_LEVEL));
1807 	set_base_spl();		/* Restore the spl to its proper value */
1808 
1809 	pghw_physid_create(cp);
1810 	/*
1811 	 * Delegate initialization tasks, which need to access the cpu_lock,
1812 	 * to mp_start_cpu_common() because we can't acquire the cpu_lock here
1813 	 * during CPU DR operations.
1814 	 */
1815 	mp_startup_signal(&procset_slave, cp->cpu_id);
1816 	mp_startup_wait(&procset_master, cp->cpu_id);
1817 	pg_cmt_cpu_startup(cp);
1818 
1819 	if (boot) {
1820 		mutex_enter(&cpu_lock);
1821 		cp->cpu_flags &= ~CPU_OFFLINE;
1822 		cpu_enable_intr(cp);
1823 		cpu_add_active(cp);
1824 		mutex_exit(&cpu_lock);
1825 	}
1826 
1827 	/* Enable interrupts */
1828 	(void) spl0();
1829 
1830 	/*
1831 	 * Fill out cpu_ucode_info.  Update microcode if necessary.
1832 	 */
1833 	ucode_check(cp);
1834 
1835 #ifndef __xpv
1836 	{
1837 		/*
1838 		 * Set up the CPU module for this CPU.  This can't be done
1839 		 * before this CPU is made CPU_READY, because we may (in
1840 		 * heterogeneous systems) need to go load another CPU module.
1841 		 * The act of attempting to load a module may trigger a
1842 		 * cross-call, which will ASSERT unless this cpu is CPU_READY.
1843 		 */
1844 		cmi_hdl_t hdl;
1845 
1846 		if ((hdl = cmi_init(CMI_HDL_NATIVE, cmi_ntv_hwchipid(CPU),
1847 		    cmi_ntv_hwcoreid(CPU), cmi_ntv_hwstrandid(CPU))) != NULL) {
1848 			if (is_x86_feature(x86_featureset, X86FSET_MCA))
1849 				cmi_mca_init(hdl);
1850 			cp->cpu_m.mcpu_cmi_hdl = hdl;
1851 		}
1852 	}
1853 #endif /* __xpv */
1854 
1855 	if (boothowto & RB_DEBUG)
1856 		kdi_cpu_init();
1857 
1858 	(void) mach_cpu_create_device_node(cp, NULL);
1859 
1860 	/*
1861 	 * Setting the bit in cpu_ready_set must be the last operation in
1862 	 * processor initialization; the boot CPU will continue to boot once
1863 	 * it sees this bit set for all active CPUs.
1864 	 */
1865 	CPUSET_ATOMIC_ADD(cpu_ready_set, cp->cpu_id);
1866 
1867 	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_idstr);
1868 	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_brandstr);
1869 	cmn_err(CE_CONT, "?cpu%d initialization complete - online\n",
1870 	    cp->cpu_id);
1871 
1872 	/*
1873 	 * Now we are done with the startup thread, so free it up.
1874 	 */
1875 	thread_exit();
1876 	panic("mp_startup: cannot return");
1877 	/*NOTREACHED*/
1878 }
1879 
1880 /*
1881  * Startup function for 'other' CPUs at boot time (besides boot cpu).
1882  */
1883 static void
1884 mp_startup_boot(void)
1885 {
1886 	mp_startup_common(B_TRUE);
1887 }
1888 
1889 /*
1890  * Startup function for hotplug CPUs at runtime.
1891  */
1892 void
1893 mp_startup_hotplug(void)
1894 {
1895 	mp_startup_common(B_FALSE);
1896 }
1897 
1898 /*
1899  * Start CPU on user request.
1900  */
1901 /* ARGSUSED */
1902 int
1903 mp_cpu_start(struct cpu *cp)
1904 {
1905 	ASSERT(MUTEX_HELD(&cpu_lock));
1906 	return (0);
1907 }
1908 
1909 /*
1910  * Stop CPU on user request.
1911  */
1912 int
1913 mp_cpu_stop(struct cpu *cp)
1914 {
1915 	extern int cbe_psm_timer_mode;
1916 	ASSERT(MUTEX_HELD(&cpu_lock));
1917 
1918 #ifdef __xpv
1919 	/*
1920 	 * We can't offline vcpu0.
1921 	 */
1922 	if (cp->cpu_id == 0)
1923 		return (EBUSY);
1924 #endif
1925 
1926 	/*
1927 	 * If TIMER_PERIODIC mode is used, CPU0 is the one running it;
1928 	 * can't stop it.  (This is true only for machines with no TSC.)
1929 	 */
1930 
1931 	if ((cbe_psm_timer_mode == TIMER_PERIODIC) && (cp->cpu_id == 0))
1932 		return (EBUSY);
1933 
1934 	return (0);
1935 }
1936 
1937 /*
1938  * Take the specified CPU out of participation in interrupts.
1939  */
1940 int
1941 cpu_disable_intr(struct cpu *cp)
1942 {
1943 	if (psm_disable_intr(cp->cpu_id) != DDI_SUCCESS)
1944 		return (EBUSY);
1945 
1946 	cp->cpu_flags &= ~CPU_ENABLE;
1947 	return (0);
1948 }
1949 
1950 /*
1951  * Allow the specified CPU to participate in interrupts.
1952  */
1953 void
1954 cpu_enable_intr(struct cpu *cp)
1955 {
1956 	ASSERT(MUTEX_HELD(&cpu_lock));
1957 	cp->cpu_flags |= CPU_ENABLE;
1958 	psm_enable_intr(cp->cpu_id);
1959 }
1960 
1961 void
1962 mp_cpu_faulted_enter(struct cpu *cp)
1963 {
1964 #ifdef __xpv
1965 	_NOTE(ARGUNUSED(cp));
1966 #else
1967 	cmi_hdl_t hdl = cp->cpu_m.mcpu_cmi_hdl;
1968 
1969 	if (hdl != NULL) {
1970 		cmi_hdl_hold(hdl);
1971 	} else {
1972 		hdl = cmi_hdl_lookup(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp),
1973 		    cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp));
1974 	}
1975 	if (hdl != NULL) {
1976 		cmi_faulted_enter(hdl);
1977 		cmi_hdl_rele(hdl);
1978 	}
1979 #endif
1980 }
1981 
1982 void
1983 mp_cpu_faulted_exit(struct cpu *cp)
1984 {
1985 #ifdef __xpv
1986 	_NOTE(ARGUNUSED(cp));
1987 #else
1988 	cmi_hdl_t hdl = cp->cpu_m.mcpu_cmi_hdl;
1989 
1990 	if (hdl != NULL) {
1991 		cmi_hdl_hold(hdl);
1992 	} else {
1993 		hdl = cmi_hdl_lookup(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp),
1994 		    cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp));
1995 	}
1996 	if (hdl != NULL) {
1997 		cmi_faulted_exit(hdl);
1998 		cmi_hdl_rele(hdl);
1999 	}
2000 #endif
2001 }
2002 
2003 /*
2004  * The following two routines are used as context operators on threads belonging
2005  * to processes with a private LDT (see sysi86).  Due to the rarity of such
2006  * processes, these routines are currently written for best code readability and
2007  * organization rather than speed.  We could avoid checking x86_featureset at
2008  * every context switch by installing different context ops, depending on
2009  * x86_featureset, at LDT creation time -- one for each combination of fast
2010  * syscall features.
2011  */
2012 
2013 /*ARGSUSED*/
2014 void
2015 cpu_fast_syscall_disable(void *arg)
2016 {
2017 	if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
2018 	    is_x86_feature(x86_featureset, X86FSET_SEP))
2019 		cpu_sep_disable();
2020 	if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
2021 	    is_x86_feature(x86_featureset, X86FSET_ASYSC))
2022 		cpu_asysc_disable();
2023 }
2024 
2025 /*ARGSUSED*/
2026 void
2027 cpu_fast_syscall_enable(void *arg)
2028 {
2029 	if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
2030 	    is_x86_feature(x86_featureset, X86FSET_SEP))
2031 		cpu_sep_enable();
2032 	if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
2033 	    is_x86_feature(x86_featureset, X86FSET_ASYSC))
2034 		cpu_asysc_enable();
2035 }
2036 
2037 static void
2038 cpu_sep_enable(void)
2039 {
2040 	ASSERT(is_x86_feature(x86_featureset, X86FSET_SEP));
2041 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
2042 
2043 	wrmsr(MSR_INTC_SEP_CS, (uint64_t)(uintptr_t)KCS_SEL);
2044 }
2045 
2046 static void
2047 cpu_sep_disable(void)
2048 {
2049 	ASSERT(is_x86_feature(x86_featureset, X86FSET_SEP));
2050 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
2051 
2052 	/*
2053 	 * Setting the SYSENTER_CS_MSR register to 0 causes software executing
2054 	 * the sysenter or sysexit instruction to trigger a #gp fault.
2055 	 */
2056 	wrmsr(MSR_INTC_SEP_CS, 0);
2057 }
2058 
2059 static void
2060 cpu_asysc_enable(void)
2061 {
2062 	ASSERT(is_x86_feature(x86_featureset, X86FSET_ASYSC));
2063 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
2064 
2065 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) |
2066 	    (uint64_t)(uintptr_t)AMD_EFER_SCE);
2067 }
2068 
2069 static void
2070 cpu_asysc_disable(void)
2071 {
2072 	ASSERT(is_x86_feature(x86_featureset, X86FSET_ASYSC));
2073 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
2074 
2075 	/*
2076 	 * Turn off the SCE (syscall enable) bit in the EFER register. Software
2077 	 * executing syscall or sysret with this bit off will incur a #ud trap.
2078 	 */
2079 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) &
2080 	    ~((uint64_t)(uintptr_t)AMD_EFER_SCE));
2081 }
2082