xref: /illumos-gate/usr/src/uts/i86pc/os/mp_startup.c (revision 8c69cc8fbe729fa7b091e901c4b50508ccc6bb33)
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 		uint64_t flags;
175 
176 #if !defined(__lint)
177 		/*
178 		 * The syscall instruction imposes a certain ordering on
179 		 * segment selectors, so we double-check that ordering
180 		 * here.
181 		 */
182 		ASSERT(KDS_SEL == KCS_SEL + 8);
183 		ASSERT(UDS_SEL == U32CS_SEL + 8);
184 		ASSERT(UCS_SEL == U32CS_SEL + 16);
185 #endif
186 		/*
187 		 * Turn syscall/sysret extensions on.
188 		 */
189 		cpu_asysc_enable();
190 
191 		/*
192 		 * Program the magic registers ..
193 		 */
194 		wrmsr(MSR_AMD_STAR,
195 		    ((uint64_t)(U32CS_SEL << 16 | KCS_SEL)) << 32);
196 		wrmsr(MSR_AMD_LSTAR, (uint64_t)(uintptr_t)sys_syscall);
197 		wrmsr(MSR_AMD_CSTAR, (uint64_t)(uintptr_t)sys_syscall32);
198 
199 		/*
200 		 * This list of flags is masked off the incoming
201 		 * %rfl when we enter the kernel.
202 		 */
203 		flags = PS_IE | PS_T;
204 		if (is_x86_feature(x86_featureset, X86FSET_SMAP) == B_TRUE)
205 			flags |= PS_ACHK;
206 		wrmsr(MSR_AMD_SFMASK, flags);
207 	}
208 #endif
209 
210 	/*
211 	 * On 32-bit kernels, we use sysenter/sysexit because it's too
212 	 * hard to use syscall/sysret, and it is more portable anyway.
213 	 *
214 	 * On 64-bit kernels on Nocona machines, the 32-bit syscall
215 	 * variant isn't available to 32-bit applications, but sysenter is.
216 	 */
217 	if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
218 	    is_x86_feature(x86_featureset, X86FSET_SEP)) {
219 
220 #if !defined(__lint)
221 		/*
222 		 * The sysenter instruction imposes a certain ordering on
223 		 * segment selectors, so we double-check that ordering
224 		 * here. See "sysenter" in Intel document 245471-012, "IA-32
225 		 * Intel Architecture Software Developer's Manual Volume 2:
226 		 * Instruction Set Reference"
227 		 */
228 		ASSERT(KDS_SEL == KCS_SEL + 8);
229 
230 		ASSERT32(UCS_SEL == ((KCS_SEL + 16) | 3));
231 		ASSERT32(UDS_SEL == UCS_SEL + 8);
232 
233 		ASSERT64(U32CS_SEL == ((KCS_SEL + 16) | 3));
234 		ASSERT64(UDS_SEL == U32CS_SEL + 8);
235 #endif
236 
237 		cpu_sep_enable();
238 
239 		/*
240 		 * resume() sets this value to the base of the threads stack
241 		 * via a context handler.
242 		 */
243 		wrmsr(MSR_INTC_SEP_ESP, 0);
244 		wrmsr(MSR_INTC_SEP_EIP, (uint64_t)(uintptr_t)sys_sysenter);
245 	}
246 
247 	kpreempt_enable();
248 }
249 
250 #if !defined(__xpv)
251 /*
252  * Configure per-cpu ID GDT
253  */
254 static void
255 init_cpu_id_gdt(struct cpu *cp)
256 {
257 	/* Write cpu_id into limit field of GDT for usermode retrieval */
258 #if defined(__amd64)
259 	set_usegd(&cp->cpu_gdt[GDT_CPUID], SDP_SHORT, NULL, cp->cpu_id,
260 	    SDT_MEMRODA, SEL_UPL, SDP_BYTES, SDP_OP32);
261 #elif defined(__i386)
262 	set_usegd(&cp->cpu_gdt[GDT_CPUID], NULL, cp->cpu_id, SDT_MEMRODA,
263 	    SEL_UPL, SDP_BYTES, SDP_OP32);
264 #endif
265 }
266 #endif /* !defined(__xpv) */
267 
268 /*
269  * Multiprocessor initialization.
270  *
271  * Allocate and initialize the cpu structure, TRAPTRACE buffer, and the
272  * startup and idle threads for the specified CPU.
273  * Parameter boot is true for boot time operations and is false for CPU
274  * DR operations.
275  */
276 static struct cpu *
277 mp_cpu_configure_common(int cpun, boolean_t boot)
278 {
279 	struct cpu *cp;
280 	kthread_id_t tp;
281 	caddr_t	sp;
282 	proc_t *procp;
283 #if !defined(__xpv)
284 	extern int idle_cpu_prefer_mwait;
285 	extern void cpu_idle_mwait();
286 #endif
287 	extern void idle();
288 	extern void cpu_idle();
289 
290 #ifdef TRAPTRACE
291 	trap_trace_ctl_t *ttc = &trap_trace_ctl[cpun];
292 #endif
293 
294 	ASSERT(MUTEX_HELD(&cpu_lock));
295 	ASSERT(cpun < NCPU && cpu[cpun] == NULL);
296 
297 	if (cpu_free_list == NULL) {
298 		cp = kmem_zalloc(sizeof (*cp), KM_SLEEP);
299 	} else {
300 		cp = cpu_free_list;
301 		cpu_free_list = cp->cpu_next_free;
302 	}
303 
304 	cp->cpu_m.mcpu_istamp = cpun << 16;
305 
306 	/* Create per CPU specific threads in the process p0. */
307 	procp = &p0;
308 
309 	/*
310 	 * Initialize the dispatcher first.
311 	 */
312 	disp_cpu_init(cp);
313 
314 	cpu_vm_data_init(cp);
315 
316 	/*
317 	 * Allocate and initialize the startup thread for this CPU.
318 	 * Interrupt and process switch stacks get allocated later
319 	 * when the CPU starts running.
320 	 */
321 	tp = thread_create(NULL, 0, NULL, NULL, 0, procp,
322 	    TS_STOPPED, maxclsyspri);
323 
324 	/*
325 	 * Set state to TS_ONPROC since this thread will start running
326 	 * as soon as the CPU comes online.
327 	 *
328 	 * All the other fields of the thread structure are setup by
329 	 * thread_create().
330 	 */
331 	THREAD_ONPROC(tp, cp);
332 	tp->t_preempt = 1;
333 	tp->t_bound_cpu = cp;
334 	tp->t_affinitycnt = 1;
335 	tp->t_cpu = cp;
336 	tp->t_disp_queue = cp->cpu_disp;
337 
338 	/*
339 	 * Setup thread to start in mp_startup_common.
340 	 */
341 	sp = tp->t_stk;
342 	tp->t_sp = (uintptr_t)(sp - MINFRAME);
343 #if defined(__amd64)
344 	tp->t_sp -= STACK_ENTRY_ALIGN;		/* fake a call */
345 #endif
346 	/*
347 	 * Setup thread start entry point for boot or hotplug.
348 	 */
349 	if (boot) {
350 		tp->t_pc = (uintptr_t)mp_startup_boot;
351 	} else {
352 		tp->t_pc = (uintptr_t)mp_startup_hotplug;
353 	}
354 
355 	cp->cpu_id = cpun;
356 	cp->cpu_self = cp;
357 	cp->cpu_thread = tp;
358 	cp->cpu_lwp = NULL;
359 	cp->cpu_dispthread = tp;
360 	cp->cpu_dispatch_pri = DISP_PRIO(tp);
361 
362 	/*
363 	 * cpu_base_spl must be set explicitly here to prevent any blocking
364 	 * operations in mp_startup_common from causing the spl of the cpu
365 	 * to drop to 0 (allowing device interrupts before we're ready) in
366 	 * resume().
367 	 * cpu_base_spl MUST remain at LOCK_LEVEL until the cpu is CPU_READY.
368 	 * As an extra bit of security on DEBUG kernels, this is enforced with
369 	 * an assertion in mp_startup_common() -- before cpu_base_spl is set
370 	 * to its proper value.
371 	 */
372 	cp->cpu_base_spl = ipltospl(LOCK_LEVEL);
373 
374 	/*
375 	 * Now, initialize per-CPU idle thread for this CPU.
376 	 */
377 	tp = thread_create(NULL, PAGESIZE, idle, NULL, 0, procp, TS_ONPROC, -1);
378 
379 	cp->cpu_idle_thread = tp;
380 
381 	tp->t_preempt = 1;
382 	tp->t_bound_cpu = cp;
383 	tp->t_affinitycnt = 1;
384 	tp->t_cpu = cp;
385 	tp->t_disp_queue = cp->cpu_disp;
386 
387 	/*
388 	 * Bootstrap the CPU's PG data
389 	 */
390 	pg_cpu_bootstrap(cp);
391 
392 	/*
393 	 * Perform CPC initialization on the new CPU.
394 	 */
395 	kcpc_hw_init(cp);
396 
397 	/*
398 	 * Allocate virtual addresses for cpu_caddr1 and cpu_caddr2
399 	 * for each CPU.
400 	 */
401 	setup_vaddr_for_ppcopy(cp);
402 
403 	/*
404 	 * Allocate page for new GDT and initialize from current GDT.
405 	 */
406 #if !defined(__lint)
407 	ASSERT((sizeof (*cp->cpu_gdt) * NGDT) <= PAGESIZE);
408 #endif
409 	cp->cpu_gdt = kmem_zalloc(PAGESIZE, KM_SLEEP);
410 	bcopy(CPU->cpu_gdt, cp->cpu_gdt, (sizeof (*cp->cpu_gdt) * NGDT));
411 
412 #if defined(__i386)
413 	/*
414 	 * setup kernel %gs.
415 	 */
416 	set_usegd(&cp->cpu_gdt[GDT_GS], cp, sizeof (struct cpu) -1, SDT_MEMRWA,
417 	    SEL_KPL, 0, 1);
418 #endif
419 
420 	/*
421 	 * If we have more than one node, each cpu gets a copy of IDT
422 	 * local to its node. If this is a Pentium box, we use cpu 0's
423 	 * IDT. cpu 0's IDT has been made read-only to workaround the
424 	 * cmpxchgl register bug
425 	 */
426 	if (system_hardware.hd_nodes && x86_type != X86_TYPE_P5) {
427 #if !defined(__lint)
428 		ASSERT((sizeof (*CPU->cpu_idt) * NIDT) <= PAGESIZE);
429 #endif
430 		cp->cpu_idt = kmem_zalloc(PAGESIZE, KM_SLEEP);
431 		bcopy(CPU->cpu_idt, cp->cpu_idt, PAGESIZE);
432 	} else {
433 		cp->cpu_idt = CPU->cpu_idt;
434 	}
435 
436 	/*
437 	 * alloc space for cpuid info
438 	 */
439 	cpuid_alloc_space(cp);
440 #if !defined(__xpv)
441 	if (is_x86_feature(x86_featureset, X86FSET_MWAIT) &&
442 	    idle_cpu_prefer_mwait) {
443 		cp->cpu_m.mcpu_mwait = cpuid_mwait_alloc(cp);
444 		cp->cpu_m.mcpu_idle_cpu = cpu_idle_mwait;
445 	} else
446 #endif
447 		cp->cpu_m.mcpu_idle_cpu = cpu_idle;
448 
449 	init_cpu_info(cp);
450 
451 #if !defined(__xpv)
452 	init_cpu_id_gdt(cp);
453 #endif
454 
455 	/*
456 	 * alloc space for ucode_info
457 	 */
458 	ucode_alloc_space(cp);
459 	xc_init_cpu(cp);
460 	hat_cpu_online(cp);
461 
462 #ifdef TRAPTRACE
463 	/*
464 	 * If this is a TRAPTRACE kernel, allocate TRAPTRACE buffers
465 	 */
466 	ttc->ttc_first = (uintptr_t)kmem_zalloc(trap_trace_bufsize, KM_SLEEP);
467 	ttc->ttc_next = ttc->ttc_first;
468 	ttc->ttc_limit = ttc->ttc_first + trap_trace_bufsize;
469 #endif
470 
471 	/*
472 	 * Record that we have another CPU.
473 	 */
474 	/*
475 	 * Initialize the interrupt threads for this CPU
476 	 */
477 	cpu_intr_alloc(cp, NINTR_THREADS);
478 
479 	cp->cpu_flags = CPU_OFFLINE | CPU_QUIESCED | CPU_POWEROFF;
480 	cpu_set_state(cp);
481 
482 	/*
483 	 * Add CPU to list of available CPUs.  It'll be on the active list
484 	 * after mp_startup_common().
485 	 */
486 	cpu_add_unit(cp);
487 
488 	return (cp);
489 }
490 
491 /*
492  * Undo what was done in mp_cpu_configure_common
493  */
494 static void
495 mp_cpu_unconfigure_common(struct cpu *cp, int error)
496 {
497 	ASSERT(MUTEX_HELD(&cpu_lock));
498 
499 	/*
500 	 * Remove the CPU from the list of available CPUs.
501 	 */
502 	cpu_del_unit(cp->cpu_id);
503 
504 	if (error == ETIMEDOUT) {
505 		/*
506 		 * The cpu was started, but never *seemed* to run any
507 		 * code in the kernel; it's probably off spinning in its
508 		 * own private world, though with potential references to
509 		 * our kmem-allocated IDTs and GDTs (for example).
510 		 *
511 		 * Worse still, it may actually wake up some time later,
512 		 * so rather than guess what it might or might not do, we
513 		 * leave the fundamental data structures intact.
514 		 */
515 		cp->cpu_flags = 0;
516 		return;
517 	}
518 
519 	/*
520 	 * At this point, the only threads bound to this CPU should
521 	 * special per-cpu threads: it's idle thread, it's pause threads,
522 	 * and it's interrupt threads.  Clean these up.
523 	 */
524 	cpu_destroy_bound_threads(cp);
525 	cp->cpu_idle_thread = NULL;
526 
527 	/*
528 	 * Free the interrupt stack.
529 	 */
530 	segkp_release(segkp,
531 	    cp->cpu_intr_stack - (INTR_STACK_SIZE - SA(MINFRAME)));
532 	cp->cpu_intr_stack = NULL;
533 
534 #ifdef TRAPTRACE
535 	/*
536 	 * Discard the trap trace buffer
537 	 */
538 	{
539 		trap_trace_ctl_t *ttc = &trap_trace_ctl[cp->cpu_id];
540 
541 		kmem_free((void *)ttc->ttc_first, trap_trace_bufsize);
542 		ttc->ttc_first = NULL;
543 	}
544 #endif
545 
546 	hat_cpu_offline(cp);
547 
548 	ucode_free_space(cp);
549 
550 	/* Free CPU ID string and brand string. */
551 	if (cp->cpu_idstr) {
552 		kmem_free(cp->cpu_idstr, CPU_IDSTRLEN);
553 		cp->cpu_idstr = NULL;
554 	}
555 	if (cp->cpu_brandstr) {
556 		kmem_free(cp->cpu_brandstr, CPU_IDSTRLEN);
557 		cp->cpu_brandstr = NULL;
558 	}
559 
560 #if !defined(__xpv)
561 	if (cp->cpu_m.mcpu_mwait != NULL) {
562 		cpuid_mwait_free(cp);
563 		cp->cpu_m.mcpu_mwait = NULL;
564 	}
565 #endif
566 	cpuid_free_space(cp);
567 
568 	if (cp->cpu_idt != CPU->cpu_idt)
569 		kmem_free(cp->cpu_idt, PAGESIZE);
570 	cp->cpu_idt = NULL;
571 
572 	kmem_free(cp->cpu_gdt, PAGESIZE);
573 	cp->cpu_gdt = NULL;
574 
575 	if (cp->cpu_supp_freqs != NULL) {
576 		size_t len = strlen(cp->cpu_supp_freqs) + 1;
577 		kmem_free(cp->cpu_supp_freqs, len);
578 		cp->cpu_supp_freqs = NULL;
579 	}
580 
581 	teardown_vaddr_for_ppcopy(cp);
582 
583 	kcpc_hw_fini(cp);
584 
585 	cp->cpu_dispthread = NULL;
586 	cp->cpu_thread = NULL;	/* discarded by cpu_destroy_bound_threads() */
587 
588 	cpu_vm_data_destroy(cp);
589 
590 	xc_fini_cpu(cp);
591 	disp_cpu_fini(cp);
592 
593 	ASSERT(cp != CPU0);
594 	bzero(cp, sizeof (*cp));
595 	cp->cpu_next_free = cpu_free_list;
596 	cpu_free_list = cp;
597 }
598 
599 /*
600  * Apply workarounds for known errata, and warn about those that are absent.
601  *
602  * System vendors occasionally create configurations which contain different
603  * revisions of the CPUs that are almost but not exactly the same.  At the
604  * time of writing, this meant that their clock rates were the same, their
605  * feature sets were the same, but the required workaround were -not-
606  * necessarily the same.  So, this routine is invoked on -every- CPU soon
607  * after starting to make sure that the resulting system contains the most
608  * pessimal set of workarounds needed to cope with *any* of the CPUs in the
609  * system.
610  *
611  * workaround_errata is invoked early in mlsetup() for CPU 0, and in
612  * mp_startup_common() for all slave CPUs. Slaves process workaround_errata
613  * prior to acknowledging their readiness to the master, so this routine will
614  * never be executed by multiple CPUs in parallel, thus making updates to
615  * global data safe.
616  *
617  * These workarounds are based on Rev 3.57 of the Revision Guide for
618  * AMD Athlon(tm) 64 and AMD Opteron(tm) Processors, August 2005.
619  */
620 
621 #if defined(OPTERON_ERRATUM_88)
622 int opteron_erratum_88;		/* if non-zero -> at least one cpu has it */
623 #endif
624 
625 #if defined(OPTERON_ERRATUM_91)
626 int opteron_erratum_91;		/* if non-zero -> at least one cpu has it */
627 #endif
628 
629 #if defined(OPTERON_ERRATUM_93)
630 int opteron_erratum_93;		/* if non-zero -> at least one cpu has it */
631 #endif
632 
633 #if defined(OPTERON_ERRATUM_95)
634 int opteron_erratum_95;		/* if non-zero -> at least one cpu has it */
635 #endif
636 
637 #if defined(OPTERON_ERRATUM_100)
638 int opteron_erratum_100;	/* if non-zero -> at least one cpu has it */
639 #endif
640 
641 #if defined(OPTERON_ERRATUM_108)
642 int opteron_erratum_108;	/* if non-zero -> at least one cpu has it */
643 #endif
644 
645 #if defined(OPTERON_ERRATUM_109)
646 int opteron_erratum_109;	/* if non-zero -> at least one cpu has it */
647 #endif
648 
649 #if defined(OPTERON_ERRATUM_121)
650 int opteron_erratum_121;	/* if non-zero -> at least one cpu has it */
651 #endif
652 
653 #if defined(OPTERON_ERRATUM_122)
654 int opteron_erratum_122;	/* if non-zero -> at least one cpu has it */
655 #endif
656 
657 #if defined(OPTERON_ERRATUM_123)
658 int opteron_erratum_123;	/* if non-zero -> at least one cpu has it */
659 #endif
660 
661 #if defined(OPTERON_ERRATUM_131)
662 int opteron_erratum_131;	/* if non-zero -> at least one cpu has it */
663 #endif
664 
665 #if defined(OPTERON_WORKAROUND_6336786)
666 int opteron_workaround_6336786;	/* non-zero -> WA relevant and applied */
667 int opteron_workaround_6336786_UP = 0;	/* Not needed for UP */
668 #endif
669 
670 #if defined(OPTERON_WORKAROUND_6323525)
671 int opteron_workaround_6323525;	/* if non-zero -> at least one cpu has it */
672 #endif
673 
674 #if defined(OPTERON_ERRATUM_298)
675 int opteron_erratum_298;
676 #endif
677 
678 #if defined(OPTERON_ERRATUM_721)
679 int opteron_erratum_721;
680 #endif
681 
682 static void
683 workaround_warning(cpu_t *cp, uint_t erratum)
684 {
685 	cmn_err(CE_WARN, "cpu%d: no workaround for erratum %u",
686 	    cp->cpu_id, erratum);
687 }
688 
689 static void
690 workaround_applied(uint_t erratum)
691 {
692 	if (erratum > 1000000)
693 		cmn_err(CE_CONT, "?workaround applied for cpu issue #%d\n",
694 		    erratum);
695 	else
696 		cmn_err(CE_CONT, "?workaround applied for cpu erratum #%d\n",
697 		    erratum);
698 }
699 
700 static void
701 msr_warning(cpu_t *cp, const char *rw, uint_t msr, int error)
702 {
703 	cmn_err(CE_WARN, "cpu%d: couldn't %smsr 0x%x, error %d",
704 	    cp->cpu_id, rw, msr, error);
705 }
706 
707 /*
708  * Determine the number of nodes in a Hammer / Greyhound / Griffin family
709  * system.
710  */
711 static uint_t
712 opteron_get_nnodes(void)
713 {
714 	static uint_t nnodes = 0;
715 
716 	if (nnodes == 0) {
717 #ifdef	DEBUG
718 		uint_t family;
719 
720 		/*
721 		 * This routine uses a PCI config space based mechanism
722 		 * for retrieving the number of nodes in the system.
723 		 * Device 24, function 0, offset 0x60 as used here is not
724 		 * AMD processor architectural, and may not work on processor
725 		 * families other than those listed below.
726 		 *
727 		 * Callers of this routine must ensure that we're running on
728 		 * a processor which supports this mechanism.
729 		 * The assertion below is meant to catch calls on unsupported
730 		 * processors.
731 		 */
732 		family = cpuid_getfamily(CPU);
733 		ASSERT(family == 0xf || family == 0x10 || family == 0x11);
734 #endif	/* DEBUG */
735 
736 		/*
737 		 * Obtain the number of nodes in the system from
738 		 * bits [6:4] of the Node ID register on node 0.
739 		 *
740 		 * The actual node count is NodeID[6:4] + 1
741 		 *
742 		 * The Node ID register is accessed via function 0,
743 		 * offset 0x60. Node 0 is device 24.
744 		 */
745 		nnodes = ((pci_getl_func(0, 24, 0, 0x60) & 0x70) >> 4) + 1;
746 	}
747 	return (nnodes);
748 }
749 
750 uint_t
751 do_erratum_298(struct cpu *cpu)
752 {
753 	static int	osvwrc = -3;
754 	extern int	osvw_opteron_erratum(cpu_t *, uint_t);
755 
756 	/*
757 	 * L2 Eviction May Occur During Processor Operation To Set
758 	 * Accessed or Dirty Bit.
759 	 */
760 	if (osvwrc == -3) {
761 		osvwrc = osvw_opteron_erratum(cpu, 298);
762 	} else {
763 		/* osvw return codes should be consistent for all cpus */
764 		ASSERT(osvwrc == osvw_opteron_erratum(cpu, 298));
765 	}
766 
767 	switch (osvwrc) {
768 	case 0:		/* erratum is not present: do nothing */
769 		break;
770 	case 1:		/* erratum is present: BIOS workaround applied */
771 		/*
772 		 * check if workaround is actually in place and issue warning
773 		 * if not.
774 		 */
775 		if (((rdmsr(MSR_AMD_HWCR) & AMD_HWCR_TLBCACHEDIS) == 0) ||
776 		    ((rdmsr(MSR_AMD_BU_CFG) & AMD_BU_CFG_E298) == 0)) {
777 #if defined(OPTERON_ERRATUM_298)
778 			opteron_erratum_298++;
779 #else
780 			workaround_warning(cpu, 298);
781 			return (1);
782 #endif
783 		}
784 		break;
785 	case -1:	/* cannot determine via osvw: check cpuid */
786 		if ((cpuid_opteron_erratum(cpu, 298) > 0) &&
787 		    (((rdmsr(MSR_AMD_HWCR) & AMD_HWCR_TLBCACHEDIS) == 0) ||
788 		    ((rdmsr(MSR_AMD_BU_CFG) & AMD_BU_CFG_E298) == 0))) {
789 #if defined(OPTERON_ERRATUM_298)
790 			opteron_erratum_298++;
791 #else
792 			workaround_warning(cpu, 298);
793 			return (1);
794 #endif
795 		}
796 		break;
797 	}
798 	return (0);
799 }
800 
801 uint_t
802 workaround_errata(struct cpu *cpu)
803 {
804 	uint_t missing = 0;
805 
806 	ASSERT(cpu == CPU);
807 
808 	/*LINTED*/
809 	if (cpuid_opteron_erratum(cpu, 88) > 0) {
810 		/*
811 		 * SWAPGS May Fail To Read Correct GS Base
812 		 */
813 #if defined(OPTERON_ERRATUM_88)
814 		/*
815 		 * The workaround is an mfence in the relevant assembler code
816 		 */
817 		opteron_erratum_88++;
818 #else
819 		workaround_warning(cpu, 88);
820 		missing++;
821 #endif
822 	}
823 
824 	if (cpuid_opteron_erratum(cpu, 91) > 0) {
825 		/*
826 		 * Software Prefetches May Report A Page Fault
827 		 */
828 #if defined(OPTERON_ERRATUM_91)
829 		/*
830 		 * fix is in trap.c
831 		 */
832 		opteron_erratum_91++;
833 #else
834 		workaround_warning(cpu, 91);
835 		missing++;
836 #endif
837 	}
838 
839 	if (cpuid_opteron_erratum(cpu, 93) > 0) {
840 		/*
841 		 * RSM Auto-Halt Restart Returns to Incorrect RIP
842 		 */
843 #if defined(OPTERON_ERRATUM_93)
844 		/*
845 		 * fix is in trap.c
846 		 */
847 		opteron_erratum_93++;
848 #else
849 		workaround_warning(cpu, 93);
850 		missing++;
851 #endif
852 	}
853 
854 	/*LINTED*/
855 	if (cpuid_opteron_erratum(cpu, 95) > 0) {
856 		/*
857 		 * RET Instruction May Return to Incorrect EIP
858 		 */
859 #if defined(OPTERON_ERRATUM_95)
860 #if defined(_LP64)
861 		/*
862 		 * Workaround this by ensuring that 32-bit user code and
863 		 * 64-bit kernel code never occupy the same address
864 		 * range mod 4G.
865 		 */
866 		if (_userlimit32 > 0xc0000000ul)
867 			*(uintptr_t *)&_userlimit32 = 0xc0000000ul;
868 
869 		/*LINTED*/
870 		ASSERT((uint32_t)COREHEAP_BASE == 0xc0000000u);
871 		opteron_erratum_95++;
872 #endif	/* _LP64 */
873 #else
874 		workaround_warning(cpu, 95);
875 		missing++;
876 #endif
877 	}
878 
879 	if (cpuid_opteron_erratum(cpu, 100) > 0) {
880 		/*
881 		 * Compatibility Mode Branches Transfer to Illegal Address
882 		 */
883 #if defined(OPTERON_ERRATUM_100)
884 		/*
885 		 * fix is in trap.c
886 		 */
887 		opteron_erratum_100++;
888 #else
889 		workaround_warning(cpu, 100);
890 		missing++;
891 #endif
892 	}
893 
894 	/*LINTED*/
895 	if (cpuid_opteron_erratum(cpu, 108) > 0) {
896 		/*
897 		 * CPUID Instruction May Return Incorrect Model Number In
898 		 * Some Processors
899 		 */
900 #if defined(OPTERON_ERRATUM_108)
901 		/*
902 		 * (Our cpuid-handling code corrects the model number on
903 		 * those processors)
904 		 */
905 #else
906 		workaround_warning(cpu, 108);
907 		missing++;
908 #endif
909 	}
910 
911 	/*LINTED*/
912 	if (cpuid_opteron_erratum(cpu, 109) > 0) do {
913 		/*
914 		 * Certain Reverse REP MOVS May Produce Unpredictable Behavior
915 		 */
916 #if defined(OPTERON_ERRATUM_109)
917 		/*
918 		 * The "workaround" is to print a warning to upgrade the BIOS
919 		 */
920 		uint64_t value;
921 		const uint_t msr = MSR_AMD_PATCHLEVEL;
922 		int err;
923 
924 		if ((err = checked_rdmsr(msr, &value)) != 0) {
925 			msr_warning(cpu, "rd", msr, err);
926 			workaround_warning(cpu, 109);
927 			missing++;
928 		}
929 		if (value == 0)
930 			opteron_erratum_109++;
931 #else
932 		workaround_warning(cpu, 109);
933 		missing++;
934 #endif
935 	/*CONSTANTCONDITION*/
936 	} while (0);
937 
938 	/*LINTED*/
939 	if (cpuid_opteron_erratum(cpu, 121) > 0) {
940 		/*
941 		 * Sequential Execution Across Non_Canonical Boundary Caused
942 		 * Processor Hang
943 		 */
944 #if defined(OPTERON_ERRATUM_121)
945 #if defined(_LP64)
946 		/*
947 		 * Erratum 121 is only present in long (64 bit) mode.
948 		 * Workaround is to include the page immediately before the
949 		 * va hole to eliminate the possibility of system hangs due to
950 		 * sequential execution across the va hole boundary.
951 		 */
952 		if (opteron_erratum_121)
953 			opteron_erratum_121++;
954 		else {
955 			if (hole_start) {
956 				hole_start -= PAGESIZE;
957 			} else {
958 				/*
959 				 * hole_start not yet initialized by
960 				 * mmu_init. Initialize hole_start
961 				 * with value to be subtracted.
962 				 */
963 				hole_start = PAGESIZE;
964 			}
965 			opteron_erratum_121++;
966 		}
967 #endif	/* _LP64 */
968 #else
969 		workaround_warning(cpu, 121);
970 		missing++;
971 #endif
972 	}
973 
974 	/*LINTED*/
975 	if (cpuid_opteron_erratum(cpu, 122) > 0) do {
976 		/*
977 		 * TLB Flush Filter May Cause Coherency Problem in
978 		 * Multiprocessor Systems
979 		 */
980 #if defined(OPTERON_ERRATUM_122)
981 		uint64_t value;
982 		const uint_t msr = MSR_AMD_HWCR;
983 		int error;
984 
985 		/*
986 		 * Erratum 122 is only present in MP configurations (multi-core
987 		 * or multi-processor).
988 		 */
989 #if defined(__xpv)
990 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
991 			break;
992 		if (!opteron_erratum_122 && xpv_nr_phys_cpus() == 1)
993 			break;
994 #else
995 		if (!opteron_erratum_122 && opteron_get_nnodes() == 1 &&
996 		    cpuid_get_ncpu_per_chip(cpu) == 1)
997 			break;
998 #endif
999 		/* disable TLB Flush Filter */
1000 
1001 		if ((error = checked_rdmsr(msr, &value)) != 0) {
1002 			msr_warning(cpu, "rd", msr, error);
1003 			workaround_warning(cpu, 122);
1004 			missing++;
1005 		} else {
1006 			value |= (uint64_t)AMD_HWCR_FFDIS;
1007 			if ((error = checked_wrmsr(msr, value)) != 0) {
1008 				msr_warning(cpu, "wr", msr, error);
1009 				workaround_warning(cpu, 122);
1010 				missing++;
1011 			}
1012 		}
1013 		opteron_erratum_122++;
1014 #else
1015 		workaround_warning(cpu, 122);
1016 		missing++;
1017 #endif
1018 	/*CONSTANTCONDITION*/
1019 	} while (0);
1020 
1021 	/*LINTED*/
1022 	if (cpuid_opteron_erratum(cpu, 123) > 0) do {
1023 		/*
1024 		 * Bypassed Reads May Cause Data Corruption of System Hang in
1025 		 * Dual Core Processors
1026 		 */
1027 #if defined(OPTERON_ERRATUM_123)
1028 		uint64_t value;
1029 		const uint_t msr = MSR_AMD_PATCHLEVEL;
1030 		int err;
1031 
1032 		/*
1033 		 * Erratum 123 applies only to multi-core cpus.
1034 		 */
1035 		if (cpuid_get_ncpu_per_chip(cpu) < 2)
1036 			break;
1037 #if defined(__xpv)
1038 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
1039 			break;
1040 #endif
1041 		/*
1042 		 * The "workaround" is to print a warning to upgrade the BIOS
1043 		 */
1044 		if ((err = checked_rdmsr(msr, &value)) != 0) {
1045 			msr_warning(cpu, "rd", msr, err);
1046 			workaround_warning(cpu, 123);
1047 			missing++;
1048 		}
1049 		if (value == 0)
1050 			opteron_erratum_123++;
1051 #else
1052 		workaround_warning(cpu, 123);
1053 		missing++;
1054 
1055 #endif
1056 	/*CONSTANTCONDITION*/
1057 	} while (0);
1058 
1059 	/*LINTED*/
1060 	if (cpuid_opteron_erratum(cpu, 131) > 0) do {
1061 		/*
1062 		 * Multiprocessor Systems with Four or More Cores May Deadlock
1063 		 * Waiting for a Probe Response
1064 		 */
1065 #if defined(OPTERON_ERRATUM_131)
1066 		uint64_t nbcfg;
1067 		const uint_t msr = MSR_AMD_NB_CFG;
1068 		const uint64_t wabits =
1069 		    AMD_NB_CFG_SRQ_HEARTBEAT | AMD_NB_CFG_SRQ_SPR;
1070 		int error;
1071 
1072 		/*
1073 		 * Erratum 131 applies to any system with four or more cores.
1074 		 */
1075 		if (opteron_erratum_131)
1076 			break;
1077 #if defined(__xpv)
1078 		if (!DOMAIN_IS_INITDOMAIN(xen_info))
1079 			break;
1080 		if (xpv_nr_phys_cpus() < 4)
1081 			break;
1082 #else
1083 		if (opteron_get_nnodes() * cpuid_get_ncpu_per_chip(cpu) < 4)
1084 			break;
1085 #endif
1086 		/*
1087 		 * Print a warning if neither of the workarounds for
1088 		 * erratum 131 is present.
1089 		 */
1090 		if ((error = checked_rdmsr(msr, &nbcfg)) != 0) {
1091 			msr_warning(cpu, "rd", msr, error);
1092 			workaround_warning(cpu, 131);
1093 			missing++;
1094 		} else if ((nbcfg & wabits) == 0) {
1095 			opteron_erratum_131++;
1096 		} else {
1097 			/* cannot have both workarounds set */
1098 			ASSERT((nbcfg & wabits) != wabits);
1099 		}
1100 #else
1101 		workaround_warning(cpu, 131);
1102 		missing++;
1103 #endif
1104 	/*CONSTANTCONDITION*/
1105 	} while (0);
1106 
1107 	/*
1108 	 * This isn't really an erratum, but for convenience the
1109 	 * detection/workaround code lives here and in cpuid_opteron_erratum.
1110 	 */
1111 	if (cpuid_opteron_erratum(cpu, 6336786) > 0) {
1112 #if defined(OPTERON_WORKAROUND_6336786)
1113 		/*
1114 		 * Disable C1-Clock ramping on multi-core/multi-processor
1115 		 * K8 platforms to guard against TSC drift.
1116 		 */
1117 		if (opteron_workaround_6336786) {
1118 			opteron_workaround_6336786++;
1119 #if defined(__xpv)
1120 		} else if ((DOMAIN_IS_INITDOMAIN(xen_info) &&
1121 		    xpv_nr_phys_cpus() > 1) ||
1122 		    opteron_workaround_6336786_UP) {
1123 			/*
1124 			 * XXPV	Hmm.  We can't walk the Northbridges on
1125 			 *	the hypervisor; so just complain and drive
1126 			 *	on.  This probably needs to be fixed in
1127 			 *	the hypervisor itself.
1128 			 */
1129 			opteron_workaround_6336786++;
1130 			workaround_warning(cpu, 6336786);
1131 #else	/* __xpv */
1132 		} else if ((opteron_get_nnodes() *
1133 		    cpuid_get_ncpu_per_chip(cpu) > 1) ||
1134 		    opteron_workaround_6336786_UP) {
1135 
1136 			uint_t	node, nnodes;
1137 			uint8_t data;
1138 
1139 			nnodes = opteron_get_nnodes();
1140 			for (node = 0; node < nnodes; node++) {
1141 				/*
1142 				 * Clear PMM7[1:0] (function 3, offset 0x87)
1143 				 * Northbridge device is the node id + 24.
1144 				 */
1145 				data = pci_getb_func(0, node + 24, 3, 0x87);
1146 				data &= 0xFC;
1147 				pci_putb_func(0, node + 24, 3, 0x87, data);
1148 			}
1149 			opteron_workaround_6336786++;
1150 #endif	/* __xpv */
1151 		}
1152 #else
1153 		workaround_warning(cpu, 6336786);
1154 		missing++;
1155 #endif
1156 	}
1157 
1158 	/*LINTED*/
1159 	/*
1160 	 * Mutex primitives don't work as expected.
1161 	 */
1162 	if (cpuid_opteron_erratum(cpu, 6323525) > 0) {
1163 #if defined(OPTERON_WORKAROUND_6323525)
1164 		/*
1165 		 * This problem only occurs with 2 or more cores. If bit in
1166 		 * MSR_AMD_BU_CFG set, then not applicable. The workaround
1167 		 * is to patch the semaphone routines with the lfence
1168 		 * instruction to provide necessary load memory barrier with
1169 		 * possible subsequent read-modify-write ops.
1170 		 *
1171 		 * It is too early in boot to call the patch routine so
1172 		 * set erratum variable to be done in startup_end().
1173 		 */
1174 		if (opteron_workaround_6323525) {
1175 			opteron_workaround_6323525++;
1176 #if defined(__xpv)
1177 		} else if (is_x86_feature(x86_featureset, X86FSET_SSE2)) {
1178 			if (DOMAIN_IS_INITDOMAIN(xen_info)) {
1179 				/*
1180 				 * XXPV	Use dom0_msr here when extended
1181 				 *	operations are supported?
1182 				 */
1183 				if (xpv_nr_phys_cpus() > 1)
1184 					opteron_workaround_6323525++;
1185 			} else {
1186 				/*
1187 				 * We have no way to tell how many physical
1188 				 * cpus there are, or even if this processor
1189 				 * has the problem, so enable the workaround
1190 				 * unconditionally (at some performance cost).
1191 				 */
1192 				opteron_workaround_6323525++;
1193 			}
1194 #else	/* __xpv */
1195 		} else if (is_x86_feature(x86_featureset, X86FSET_SSE2) &&
1196 		    ((opteron_get_nnodes() *
1197 		    cpuid_get_ncpu_per_chip(cpu)) > 1)) {
1198 			if ((xrdmsr(MSR_AMD_BU_CFG) & (UINT64_C(1) << 33)) == 0)
1199 				opteron_workaround_6323525++;
1200 #endif	/* __xpv */
1201 		}
1202 #else
1203 		workaround_warning(cpu, 6323525);
1204 		missing++;
1205 #endif
1206 	}
1207 
1208 	missing += do_erratum_298(cpu);
1209 
1210 	if (cpuid_opteron_erratum(cpu, 721) > 0) {
1211 #if defined(OPTERON_ERRATUM_721)
1212 		on_trap_data_t otd;
1213 
1214 		if (!on_trap(&otd, OT_DATA_ACCESS))
1215 			wrmsr(MSR_AMD_DE_CFG,
1216 			    rdmsr(MSR_AMD_DE_CFG) | AMD_DE_CFG_E721);
1217 		no_trap();
1218 
1219 		opteron_erratum_721++;
1220 #else
1221 		workaround_warning(cpu, 721);
1222 		missing++;
1223 #endif
1224 	}
1225 
1226 #ifdef __xpv
1227 	return (0);
1228 #else
1229 	return (missing);
1230 #endif
1231 }
1232 
1233 void
1234 workaround_errata_end()
1235 {
1236 #if defined(OPTERON_ERRATUM_88)
1237 	if (opteron_erratum_88)
1238 		workaround_applied(88);
1239 #endif
1240 #if defined(OPTERON_ERRATUM_91)
1241 	if (opteron_erratum_91)
1242 		workaround_applied(91);
1243 #endif
1244 #if defined(OPTERON_ERRATUM_93)
1245 	if (opteron_erratum_93)
1246 		workaround_applied(93);
1247 #endif
1248 #if defined(OPTERON_ERRATUM_95)
1249 	if (opteron_erratum_95)
1250 		workaround_applied(95);
1251 #endif
1252 #if defined(OPTERON_ERRATUM_100)
1253 	if (opteron_erratum_100)
1254 		workaround_applied(100);
1255 #endif
1256 #if defined(OPTERON_ERRATUM_108)
1257 	if (opteron_erratum_108)
1258 		workaround_applied(108);
1259 #endif
1260 #if defined(OPTERON_ERRATUM_109)
1261 	if (opteron_erratum_109) {
1262 		cmn_err(CE_WARN,
1263 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1264 		    " processor\nerratum 109 was not detected; updating your"
1265 		    " system's BIOS to a version\ncontaining this"
1266 		    " microcode patch is HIGHLY recommended or erroneous"
1267 		    " system\noperation may occur.\n");
1268 	}
1269 #endif
1270 #if defined(OPTERON_ERRATUM_121)
1271 	if (opteron_erratum_121)
1272 		workaround_applied(121);
1273 #endif
1274 #if defined(OPTERON_ERRATUM_122)
1275 	if (opteron_erratum_122)
1276 		workaround_applied(122);
1277 #endif
1278 #if defined(OPTERON_ERRATUM_123)
1279 	if (opteron_erratum_123) {
1280 		cmn_err(CE_WARN,
1281 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1282 		    " processor\nerratum 123 was not detected; updating your"
1283 		    " system's BIOS to a version\ncontaining this"
1284 		    " microcode patch is HIGHLY recommended or erroneous"
1285 		    " system\noperation may occur.\n");
1286 	}
1287 #endif
1288 #if defined(OPTERON_ERRATUM_131)
1289 	if (opteron_erratum_131) {
1290 		cmn_err(CE_WARN,
1291 		    "BIOS microcode patch for AMD Athlon(tm) 64/Opteron(tm)"
1292 		    " processor\nerratum 131 was not detected; updating your"
1293 		    " system's BIOS to a version\ncontaining this"
1294 		    " microcode patch is HIGHLY recommended or erroneous"
1295 		    " system\noperation may occur.\n");
1296 	}
1297 #endif
1298 #if defined(OPTERON_WORKAROUND_6336786)
1299 	if (opteron_workaround_6336786)
1300 		workaround_applied(6336786);
1301 #endif
1302 #if defined(OPTERON_WORKAROUND_6323525)
1303 	if (opteron_workaround_6323525)
1304 		workaround_applied(6323525);
1305 #endif
1306 #if defined(OPTERON_ERRATUM_298)
1307 	if (opteron_erratum_298) {
1308 		cmn_err(CE_WARN,
1309 		    "BIOS microcode patch for AMD 64/Opteron(tm)"
1310 		    " processor\nerratum 298 was not detected; updating your"
1311 		    " system's BIOS to a version\ncontaining this"
1312 		    " microcode patch is HIGHLY recommended or erroneous"
1313 		    " system\noperation may occur.\n");
1314 	}
1315 #endif
1316 #if defined(OPTERON_ERRATUM_721)
1317 	if (opteron_erratum_721)
1318 		workaround_applied(721);
1319 #endif
1320 }
1321 
1322 /*
1323  * The procset_slave and procset_master are used to synchronize
1324  * between the control CPU and the target CPU when starting CPUs.
1325  */
1326 static cpuset_t procset_slave, procset_master;
1327 
1328 static void
1329 mp_startup_wait(cpuset_t *sp, processorid_t cpuid)
1330 {
1331 	cpuset_t tempset;
1332 
1333 	for (tempset = *sp; !CPU_IN_SET(tempset, cpuid);
1334 	    tempset = *(volatile cpuset_t *)sp) {
1335 		SMT_PAUSE();
1336 	}
1337 	CPUSET_ATOMIC_DEL(*(cpuset_t *)sp, cpuid);
1338 }
1339 
1340 static void
1341 mp_startup_signal(cpuset_t *sp, processorid_t cpuid)
1342 {
1343 	cpuset_t tempset;
1344 
1345 	CPUSET_ATOMIC_ADD(*(cpuset_t *)sp, cpuid);
1346 	for (tempset = *sp; CPU_IN_SET(tempset, cpuid);
1347 	    tempset = *(volatile cpuset_t *)sp) {
1348 		SMT_PAUSE();
1349 	}
1350 }
1351 
1352 int
1353 mp_start_cpu_common(cpu_t *cp, boolean_t boot)
1354 {
1355 	_NOTE(ARGUNUSED(boot));
1356 
1357 	void *ctx;
1358 	int delays;
1359 	int error = 0;
1360 	cpuset_t tempset;
1361 	processorid_t cpuid;
1362 #ifndef __xpv
1363 	extern void cpupm_init(cpu_t *);
1364 #endif
1365 
1366 	ASSERT(cp != NULL);
1367 	cpuid = cp->cpu_id;
1368 	ctx = mach_cpucontext_alloc(cp);
1369 	if (ctx == NULL) {
1370 		cmn_err(CE_WARN,
1371 		    "cpu%d: failed to allocate context", cp->cpu_id);
1372 		return (EAGAIN);
1373 	}
1374 	error = mach_cpu_start(cp, ctx);
1375 	if (error != 0) {
1376 		cmn_err(CE_WARN,
1377 		    "cpu%d: failed to start, error %d", cp->cpu_id, error);
1378 		mach_cpucontext_free(cp, ctx, error);
1379 		return (error);
1380 	}
1381 
1382 	for (delays = 0, tempset = procset_slave; !CPU_IN_SET(tempset, cpuid);
1383 	    delays++) {
1384 		if (delays == 500) {
1385 			/*
1386 			 * After five seconds, things are probably looking
1387 			 * a bit bleak - explain the hang.
1388 			 */
1389 			cmn_err(CE_NOTE, "cpu%d: started, "
1390 			    "but not running in the kernel yet", cpuid);
1391 		} else if (delays > 2000) {
1392 			/*
1393 			 * We waited at least 20 seconds, bail ..
1394 			 */
1395 			error = ETIMEDOUT;
1396 			cmn_err(CE_WARN, "cpu%d: timed out", cpuid);
1397 			mach_cpucontext_free(cp, ctx, error);
1398 			return (error);
1399 		}
1400 
1401 		/*
1402 		 * wait at least 10ms, then check again..
1403 		 */
1404 		delay(USEC_TO_TICK_ROUNDUP(10000));
1405 		tempset = *((volatile cpuset_t *)&procset_slave);
1406 	}
1407 	CPUSET_ATOMIC_DEL(procset_slave, cpuid);
1408 
1409 	mach_cpucontext_free(cp, ctx, 0);
1410 
1411 #ifndef __xpv
1412 	if (tsc_gethrtime_enable)
1413 		tsc_sync_master(cpuid);
1414 #endif
1415 
1416 	if (dtrace_cpu_init != NULL) {
1417 		(*dtrace_cpu_init)(cpuid);
1418 	}
1419 
1420 	/*
1421 	 * During CPU DR operations, the cpu_lock is held by current
1422 	 * (the control) thread. We can't release the cpu_lock here
1423 	 * because that will break the CPU DR logic.
1424 	 * On the other hand, CPUPM and processor group initialization
1425 	 * routines need to access the cpu_lock. So we invoke those
1426 	 * routines here on behalf of mp_startup_common().
1427 	 *
1428 	 * CPUPM and processor group initialization routines depend
1429 	 * on the cpuid probing results. Wait for mp_startup_common()
1430 	 * to signal that cpuid probing is done.
1431 	 */
1432 	mp_startup_wait(&procset_slave, cpuid);
1433 #ifndef __xpv
1434 	cpupm_init(cp);
1435 #endif
1436 	(void) pg_cpu_init(cp, B_FALSE);
1437 	cpu_set_state(cp);
1438 	mp_startup_signal(&procset_master, cpuid);
1439 
1440 	return (0);
1441 }
1442 
1443 /*
1444  * Start a single cpu, assuming that the kernel context is available
1445  * to successfully start another cpu.
1446  *
1447  * (For example, real mode code is mapped into the right place
1448  * in memory and is ready to be run.)
1449  */
1450 int
1451 start_cpu(processorid_t who)
1452 {
1453 	cpu_t *cp;
1454 	int error = 0;
1455 	cpuset_t tempset;
1456 
1457 	ASSERT(who != 0);
1458 
1459 	/*
1460 	 * Check if there's at least a Mbyte of kmem available
1461 	 * before attempting to start the cpu.
1462 	 */
1463 	if (kmem_avail() < 1024 * 1024) {
1464 		/*
1465 		 * Kick off a reap in case that helps us with
1466 		 * later attempts ..
1467 		 */
1468 		kmem_reap();
1469 		return (ENOMEM);
1470 	}
1471 
1472 	/*
1473 	 * First configure cpu.
1474 	 */
1475 	cp = mp_cpu_configure_common(who, B_TRUE);
1476 	ASSERT(cp != NULL);
1477 
1478 	/*
1479 	 * Then start cpu.
1480 	 */
1481 	error = mp_start_cpu_common(cp, B_TRUE);
1482 	if (error != 0) {
1483 		mp_cpu_unconfigure_common(cp, error);
1484 		return (error);
1485 	}
1486 
1487 	mutex_exit(&cpu_lock);
1488 	tempset = cpu_ready_set;
1489 	while (!CPU_IN_SET(tempset, who)) {
1490 		drv_usecwait(1);
1491 		tempset = *((volatile cpuset_t *)&cpu_ready_set);
1492 	}
1493 	mutex_enter(&cpu_lock);
1494 
1495 	return (0);
1496 }
1497 
1498 void
1499 start_other_cpus(int cprboot)
1500 {
1501 	_NOTE(ARGUNUSED(cprboot));
1502 
1503 	uint_t who;
1504 	uint_t bootcpuid = 0;
1505 
1506 	/*
1507 	 * Initialize our own cpu_info.
1508 	 */
1509 	init_cpu_info(CPU);
1510 
1511 #if !defined(__xpv)
1512 	init_cpu_id_gdt(CPU);
1513 #endif
1514 
1515 	cmn_err(CE_CONT, "?cpu%d: %s\n", CPU->cpu_id, CPU->cpu_idstr);
1516 	cmn_err(CE_CONT, "?cpu%d: %s\n", CPU->cpu_id, CPU->cpu_brandstr);
1517 
1518 	/*
1519 	 * Initialize our syscall handlers
1520 	 */
1521 	init_cpu_syscall(CPU);
1522 
1523 	/*
1524 	 * Take the boot cpu out of the mp_cpus set because we know
1525 	 * it's already running.  Add it to the cpu_ready_set for
1526 	 * precisely the same reason.
1527 	 */
1528 	CPUSET_DEL(mp_cpus, bootcpuid);
1529 	CPUSET_ADD(cpu_ready_set, bootcpuid);
1530 
1531 	/*
1532 	 * skip the rest of this if
1533 	 * . only 1 cpu dectected and system isn't hotplug-capable
1534 	 * . not using MP
1535 	 */
1536 	if ((CPUSET_ISNULL(mp_cpus) && plat_dr_support_cpu() == 0) ||
1537 	    use_mp == 0) {
1538 		if (use_mp == 0)
1539 			cmn_err(CE_CONT, "?***** Not in MP mode\n");
1540 		goto done;
1541 	}
1542 
1543 	/*
1544 	 * perform such initialization as is needed
1545 	 * to be able to take CPUs on- and off-line.
1546 	 */
1547 	cpu_pause_init();
1548 
1549 	xc_init_cpu(CPU);		/* initialize processor crosscalls */
1550 
1551 	if (mach_cpucontext_init() != 0)
1552 		goto done;
1553 
1554 	flushes_require_xcalls = 1;
1555 
1556 	/*
1557 	 * We lock our affinity to the master CPU to ensure that all slave CPUs
1558 	 * do their TSC syncs with the same CPU.
1559 	 */
1560 	affinity_set(CPU_CURRENT);
1561 
1562 	for (who = 0; who < NCPU; who++) {
1563 		if (!CPU_IN_SET(mp_cpus, who))
1564 			continue;
1565 		ASSERT(who != bootcpuid);
1566 
1567 		mutex_enter(&cpu_lock);
1568 		if (start_cpu(who) != 0)
1569 			CPUSET_DEL(mp_cpus, who);
1570 		cpu_state_change_notify(who, CPU_SETUP);
1571 		mutex_exit(&cpu_lock);
1572 	}
1573 
1574 	/* Free the space allocated to hold the microcode file */
1575 	ucode_cleanup();
1576 
1577 	affinity_clear();
1578 
1579 	mach_cpucontext_fini();
1580 
1581 done:
1582 	if (get_hwenv() == HW_NATIVE)
1583 		workaround_errata_end();
1584 	cmi_post_mpstartup();
1585 
1586 	if (use_mp && ncpus != boot_max_ncpus) {
1587 		cmn_err(CE_NOTE,
1588 		    "System detected %d cpus, but "
1589 		    "only %d cpu(s) were enabled during boot.",
1590 		    boot_max_ncpus, ncpus);
1591 		cmn_err(CE_NOTE,
1592 		    "Use \"boot-ncpus\" parameter to enable more CPU(s). "
1593 		    "See eeprom(1M).");
1594 	}
1595 }
1596 
1597 int
1598 mp_cpu_configure(int cpuid)
1599 {
1600 	cpu_t *cp;
1601 
1602 	if (use_mp == 0 || plat_dr_support_cpu() == 0) {
1603 		return (ENOTSUP);
1604 	}
1605 
1606 	cp = cpu_get(cpuid);
1607 	if (cp != NULL) {
1608 		return (EALREADY);
1609 	}
1610 
1611 	/*
1612 	 * Check if there's at least a Mbyte of kmem available
1613 	 * before attempting to start the cpu.
1614 	 */
1615 	if (kmem_avail() < 1024 * 1024) {
1616 		/*
1617 		 * Kick off a reap in case that helps us with
1618 		 * later attempts ..
1619 		 */
1620 		kmem_reap();
1621 		return (ENOMEM);
1622 	}
1623 
1624 	cp = mp_cpu_configure_common(cpuid, B_FALSE);
1625 	ASSERT(cp != NULL && cpu_get(cpuid) == cp);
1626 
1627 	return (cp != NULL ? 0 : EAGAIN);
1628 }
1629 
1630 int
1631 mp_cpu_unconfigure(int cpuid)
1632 {
1633 	cpu_t *cp;
1634 
1635 	if (use_mp == 0 || plat_dr_support_cpu() == 0) {
1636 		return (ENOTSUP);
1637 	} else if (cpuid < 0 || cpuid >= max_ncpus) {
1638 		return (EINVAL);
1639 	}
1640 
1641 	cp = cpu_get(cpuid);
1642 	if (cp == NULL) {
1643 		return (ENODEV);
1644 	}
1645 	mp_cpu_unconfigure_common(cp, 0);
1646 
1647 	return (0);
1648 }
1649 
1650 /*
1651  * Startup function for 'other' CPUs (besides boot cpu).
1652  * Called from real_mode_start.
1653  *
1654  * WARNING: until CPU_READY is set, mp_startup_common and routines called by
1655  * mp_startup_common should not call routines (e.g. kmem_free) that could call
1656  * hat_unload which requires CPU_READY to be set.
1657  */
1658 static void
1659 mp_startup_common(boolean_t boot)
1660 {
1661 	cpu_t *cp = CPU;
1662 	uchar_t new_x86_featureset[BT_SIZEOFMAP(NUM_X86_FEATURES)];
1663 	extern void cpu_event_init_cpu(cpu_t *);
1664 
1665 	/*
1666 	 * We need to get TSC on this proc synced (i.e., any delta
1667 	 * from cpu0 accounted for) as soon as we can, because many
1668 	 * many things use gethrtime/pc_gethrestime, including
1669 	 * interrupts, cmn_err, etc.  Before we can do that, we want to
1670 	 * clear TSC if we're on a buggy Sandy/Ivy Bridge CPU, so do that
1671 	 * right away.
1672 	 */
1673 	bzero(new_x86_featureset, BT_SIZEOFMAP(NUM_X86_FEATURES));
1674 	cpuid_pass1(cp, new_x86_featureset);
1675 
1676 	if (boot && get_hwenv() == HW_NATIVE &&
1677 	    cpuid_getvendor(CPU) == X86_VENDOR_Intel &&
1678 	    cpuid_getfamily(CPU) == 6 &&
1679 	    (cpuid_getmodel(CPU) == 0x2d || cpuid_getmodel(CPU) == 0x3e) &&
1680 	    is_x86_feature(new_x86_featureset, X86FSET_TSC)) {
1681 		(void) wrmsr(REG_TSC, 0UL);
1682 	}
1683 
1684 	/* Let the control CPU continue into tsc_sync_master() */
1685 	mp_startup_signal(&procset_slave, cp->cpu_id);
1686 
1687 #ifndef __xpv
1688 	if (tsc_gethrtime_enable)
1689 		tsc_sync_slave();
1690 #endif
1691 
1692 	/*
1693 	 * Once this was done from assembly, but it's safer here; if
1694 	 * it blocks, we need to be able to swtch() to and from, and
1695 	 * since we get here by calling t_pc, we need to do that call
1696 	 * before swtch() overwrites it.
1697 	 */
1698 	(void) (*ap_mlsetup)();
1699 
1700 #ifndef __xpv
1701 	/*
1702 	 * Program this cpu's PAT
1703 	 */
1704 	pat_sync();
1705 #endif
1706 
1707 	/*
1708 	 * Set up TSC_AUX to contain the cpuid for this processor
1709 	 * for the rdtscp instruction.
1710 	 */
1711 	if (is_x86_feature(x86_featureset, X86FSET_TSCP))
1712 		(void) wrmsr(MSR_AMD_TSCAUX, cp->cpu_id);
1713 
1714 	/*
1715 	 * Initialize this CPU's syscall handlers
1716 	 */
1717 	init_cpu_syscall(cp);
1718 
1719 	/*
1720 	 * Enable interrupts with spl set to LOCK_LEVEL. LOCK_LEVEL is the
1721 	 * highest level at which a routine is permitted to block on
1722 	 * an adaptive mutex (allows for cpu poke interrupt in case
1723 	 * the cpu is blocked on a mutex and halts). Setting LOCK_LEVEL blocks
1724 	 * device interrupts that may end up in the hat layer issuing cross
1725 	 * calls before CPU_READY is set.
1726 	 */
1727 	splx(ipltospl(LOCK_LEVEL));
1728 	sti();
1729 
1730 	/*
1731 	 * Do a sanity check to make sure this new CPU is a sane thing
1732 	 * to add to the collection of processors running this system.
1733 	 *
1734 	 * XXX	Clearly this needs to get more sophisticated, if x86
1735 	 * systems start to get built out of heterogenous CPUs; as is
1736 	 * likely to happen once the number of processors in a configuration
1737 	 * gets large enough.
1738 	 */
1739 	if (compare_x86_featureset(x86_featureset, new_x86_featureset) ==
1740 	    B_FALSE) {
1741 		cmn_err(CE_CONT, "cpu%d: featureset\n", cp->cpu_id);
1742 		print_x86_featureset(new_x86_featureset);
1743 		cmn_err(CE_WARN, "cpu%d feature mismatch", cp->cpu_id);
1744 	}
1745 
1746 	/*
1747 	 * There exists a small subset of systems which expose differing
1748 	 * MWAIT/MONITOR support between CPUs.  If MWAIT support is absent from
1749 	 * the boot CPU, but is found on a later CPU, the system continues to
1750 	 * operate as if no MWAIT support is available.
1751 	 *
1752 	 * The reverse case, where MWAIT is available on the boot CPU but not
1753 	 * on a subsequently initialized CPU, is not presently allowed and will
1754 	 * result in a panic.
1755 	 */
1756 	if (is_x86_feature(x86_featureset, X86FSET_MWAIT) !=
1757 	    is_x86_feature(new_x86_featureset, X86FSET_MWAIT)) {
1758 		if (!is_x86_feature(x86_featureset, X86FSET_MWAIT)) {
1759 			remove_x86_feature(new_x86_featureset, X86FSET_MWAIT);
1760 		} else {
1761 			panic("unsupported mixed cpu mwait support detected");
1762 		}
1763 	}
1764 
1765 	/*
1766 	 * We could be more sophisticated here, and just mark the CPU
1767 	 * as "faulted" but at this point we'll opt for the easier
1768 	 * answer of dying horribly.  Provided the boot cpu is ok,
1769 	 * the system can be recovered by booting with use_mp set to zero.
1770 	 */
1771 	if (workaround_errata(cp) != 0)
1772 		panic("critical workaround(s) missing for cpu%d", cp->cpu_id);
1773 
1774 	/*
1775 	 * We can touch cpu_flags here without acquiring the cpu_lock here
1776 	 * because the cpu_lock is held by the control CPU which is running
1777 	 * mp_start_cpu_common().
1778 	 * Need to clear CPU_QUIESCED flag before calling any function which
1779 	 * may cause thread context switching, such as kmem_alloc() etc.
1780 	 * The idle thread checks for CPU_QUIESCED flag and loops for ever if
1781 	 * it's set. So the startup thread may have no chance to switch back
1782 	 * again if it's switched away with CPU_QUIESCED set.
1783 	 */
1784 	cp->cpu_flags &= ~(CPU_POWEROFF | CPU_QUIESCED);
1785 
1786 	/*
1787 	 * Setup this processor for XSAVE.
1788 	 */
1789 	if (fp_save_mech == FP_XSAVE) {
1790 		xsave_setup_msr(cp);
1791 	}
1792 
1793 	cpuid_pass2(cp);
1794 	cpuid_pass3(cp);
1795 	cpuid_pass4(cp, NULL);
1796 
1797 	/*
1798 	 * Correct cpu_idstr and cpu_brandstr on target CPU after
1799 	 * cpuid_pass1() is done.
1800 	 */
1801 	(void) cpuid_getidstr(cp, cp->cpu_idstr, CPU_IDSTRLEN);
1802 	(void) cpuid_getbrandstr(cp, cp->cpu_brandstr, CPU_IDSTRLEN);
1803 
1804 	cp->cpu_flags |= CPU_RUNNING | CPU_READY | CPU_EXISTS;
1805 
1806 	post_startup_cpu_fixups();
1807 
1808 	cpu_event_init_cpu(cp);
1809 
1810 	/*
1811 	 * Enable preemption here so that contention for any locks acquired
1812 	 * later in mp_startup_common may be preempted if the thread owning
1813 	 * those locks is continuously executing on other CPUs (for example,
1814 	 * this CPU must be preemptible to allow other CPUs to pause it during
1815 	 * their startup phases).  It's safe to enable preemption here because
1816 	 * the CPU state is pretty-much fully constructed.
1817 	 */
1818 	curthread->t_preempt = 0;
1819 
1820 	/* The base spl should still be at LOCK LEVEL here */
1821 	ASSERT(cp->cpu_base_spl == ipltospl(LOCK_LEVEL));
1822 	set_base_spl();		/* Restore the spl to its proper value */
1823 
1824 	pghw_physid_create(cp);
1825 	/*
1826 	 * Delegate initialization tasks, which need to access the cpu_lock,
1827 	 * to mp_start_cpu_common() because we can't acquire the cpu_lock here
1828 	 * during CPU DR operations.
1829 	 */
1830 	mp_startup_signal(&procset_slave, cp->cpu_id);
1831 	mp_startup_wait(&procset_master, cp->cpu_id);
1832 	pg_cmt_cpu_startup(cp);
1833 
1834 	if (boot) {
1835 		mutex_enter(&cpu_lock);
1836 		cp->cpu_flags &= ~CPU_OFFLINE;
1837 		cpu_enable_intr(cp);
1838 		cpu_add_active(cp);
1839 		mutex_exit(&cpu_lock);
1840 	}
1841 
1842 	/* Enable interrupts */
1843 	(void) spl0();
1844 
1845 	/*
1846 	 * Fill out cpu_ucode_info.  Update microcode if necessary.
1847 	 */
1848 	ucode_check(cp);
1849 
1850 #ifndef __xpv
1851 	{
1852 		/*
1853 		 * Set up the CPU module for this CPU.  This can't be done
1854 		 * before this CPU is made CPU_READY, because we may (in
1855 		 * heterogeneous systems) need to go load another CPU module.
1856 		 * The act of attempting to load a module may trigger a
1857 		 * cross-call, which will ASSERT unless this cpu is CPU_READY.
1858 		 */
1859 		cmi_hdl_t hdl;
1860 
1861 		if ((hdl = cmi_init(CMI_HDL_NATIVE, cmi_ntv_hwchipid(CPU),
1862 		    cmi_ntv_hwcoreid(CPU), cmi_ntv_hwstrandid(CPU))) != NULL) {
1863 			if (is_x86_feature(x86_featureset, X86FSET_MCA))
1864 				cmi_mca_init(hdl);
1865 			cp->cpu_m.mcpu_cmi_hdl = hdl;
1866 		}
1867 	}
1868 #endif /* __xpv */
1869 
1870 	if (boothowto & RB_DEBUG)
1871 		kdi_cpu_init();
1872 
1873 	/*
1874 	 * Setting the bit in cpu_ready_set must be the last operation in
1875 	 * processor initialization; the boot CPU will continue to boot once
1876 	 * it sees this bit set for all active CPUs.
1877 	 */
1878 	CPUSET_ATOMIC_ADD(cpu_ready_set, cp->cpu_id);
1879 
1880 	(void) mach_cpu_create_device_node(cp, NULL);
1881 
1882 	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_idstr);
1883 	cmn_err(CE_CONT, "?cpu%d: %s\n", cp->cpu_id, cp->cpu_brandstr);
1884 	cmn_err(CE_CONT, "?cpu%d initialization complete - online\n",
1885 	    cp->cpu_id);
1886 
1887 	/*
1888 	 * Now we are done with the startup thread, so free it up.
1889 	 */
1890 	thread_exit();
1891 	panic("mp_startup: cannot return");
1892 	/*NOTREACHED*/
1893 }
1894 
1895 /*
1896  * Startup function for 'other' CPUs at boot time (besides boot cpu).
1897  */
1898 static void
1899 mp_startup_boot(void)
1900 {
1901 	mp_startup_common(B_TRUE);
1902 }
1903 
1904 /*
1905  * Startup function for hotplug CPUs at runtime.
1906  */
1907 void
1908 mp_startup_hotplug(void)
1909 {
1910 	mp_startup_common(B_FALSE);
1911 }
1912 
1913 /*
1914  * Start CPU on user request.
1915  */
1916 /* ARGSUSED */
1917 int
1918 mp_cpu_start(struct cpu *cp)
1919 {
1920 	ASSERT(MUTEX_HELD(&cpu_lock));
1921 	return (0);
1922 }
1923 
1924 /*
1925  * Stop CPU on user request.
1926  */
1927 int
1928 mp_cpu_stop(struct cpu *cp)
1929 {
1930 	extern int cbe_psm_timer_mode;
1931 	ASSERT(MUTEX_HELD(&cpu_lock));
1932 
1933 #ifdef __xpv
1934 	/*
1935 	 * We can't offline vcpu0.
1936 	 */
1937 	if (cp->cpu_id == 0)
1938 		return (EBUSY);
1939 #endif
1940 
1941 	/*
1942 	 * If TIMER_PERIODIC mode is used, CPU0 is the one running it;
1943 	 * can't stop it.  (This is true only for machines with no TSC.)
1944 	 */
1945 
1946 	if ((cbe_psm_timer_mode == TIMER_PERIODIC) && (cp->cpu_id == 0))
1947 		return (EBUSY);
1948 
1949 	return (0);
1950 }
1951 
1952 /*
1953  * Take the specified CPU out of participation in interrupts.
1954  */
1955 int
1956 cpu_disable_intr(struct cpu *cp)
1957 {
1958 	if (psm_disable_intr(cp->cpu_id) != DDI_SUCCESS)
1959 		return (EBUSY);
1960 
1961 	cp->cpu_flags &= ~CPU_ENABLE;
1962 	return (0);
1963 }
1964 
1965 /*
1966  * Allow the specified CPU to participate in interrupts.
1967  */
1968 void
1969 cpu_enable_intr(struct cpu *cp)
1970 {
1971 	ASSERT(MUTEX_HELD(&cpu_lock));
1972 	cp->cpu_flags |= CPU_ENABLE;
1973 	psm_enable_intr(cp->cpu_id);
1974 }
1975 
1976 void
1977 mp_cpu_faulted_enter(struct cpu *cp)
1978 {
1979 #ifdef __xpv
1980 	_NOTE(ARGUNUSED(cp));
1981 #else
1982 	cmi_hdl_t hdl = cp->cpu_m.mcpu_cmi_hdl;
1983 
1984 	if (hdl != NULL) {
1985 		cmi_hdl_hold(hdl);
1986 	} else {
1987 		hdl = cmi_hdl_lookup(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp),
1988 		    cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp));
1989 	}
1990 	if (hdl != NULL) {
1991 		cmi_faulted_enter(hdl);
1992 		cmi_hdl_rele(hdl);
1993 	}
1994 #endif
1995 }
1996 
1997 void
1998 mp_cpu_faulted_exit(struct cpu *cp)
1999 {
2000 #ifdef __xpv
2001 	_NOTE(ARGUNUSED(cp));
2002 #else
2003 	cmi_hdl_t hdl = cp->cpu_m.mcpu_cmi_hdl;
2004 
2005 	if (hdl != NULL) {
2006 		cmi_hdl_hold(hdl);
2007 	} else {
2008 		hdl = cmi_hdl_lookup(CMI_HDL_NATIVE, cmi_ntv_hwchipid(cp),
2009 		    cmi_ntv_hwcoreid(cp), cmi_ntv_hwstrandid(cp));
2010 	}
2011 	if (hdl != NULL) {
2012 		cmi_faulted_exit(hdl);
2013 		cmi_hdl_rele(hdl);
2014 	}
2015 #endif
2016 }
2017 
2018 /*
2019  * The following two routines are used as context operators on threads belonging
2020  * to processes with a private LDT (see sysi86).  Due to the rarity of such
2021  * processes, these routines are currently written for best code readability and
2022  * organization rather than speed.  We could avoid checking x86_featureset at
2023  * every context switch by installing different context ops, depending on
2024  * x86_featureset, at LDT creation time -- one for each combination of fast
2025  * syscall features.
2026  */
2027 
2028 /*ARGSUSED*/
2029 void
2030 cpu_fast_syscall_disable(void *arg)
2031 {
2032 	if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
2033 	    is_x86_feature(x86_featureset, X86FSET_SEP))
2034 		cpu_sep_disable();
2035 	if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
2036 	    is_x86_feature(x86_featureset, X86FSET_ASYSC))
2037 		cpu_asysc_disable();
2038 }
2039 
2040 /*ARGSUSED*/
2041 void
2042 cpu_fast_syscall_enable(void *arg)
2043 {
2044 	if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
2045 	    is_x86_feature(x86_featureset, X86FSET_SEP))
2046 		cpu_sep_enable();
2047 	if (is_x86_feature(x86_featureset, X86FSET_MSR) &&
2048 	    is_x86_feature(x86_featureset, X86FSET_ASYSC))
2049 		cpu_asysc_enable();
2050 }
2051 
2052 static void
2053 cpu_sep_enable(void)
2054 {
2055 	ASSERT(is_x86_feature(x86_featureset, X86FSET_SEP));
2056 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
2057 
2058 	wrmsr(MSR_INTC_SEP_CS, (uint64_t)(uintptr_t)KCS_SEL);
2059 }
2060 
2061 static void
2062 cpu_sep_disable(void)
2063 {
2064 	ASSERT(is_x86_feature(x86_featureset, X86FSET_SEP));
2065 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
2066 
2067 	/*
2068 	 * Setting the SYSENTER_CS_MSR register to 0 causes software executing
2069 	 * the sysenter or sysexit instruction to trigger a #gp fault.
2070 	 */
2071 	wrmsr(MSR_INTC_SEP_CS, 0);
2072 }
2073 
2074 static void
2075 cpu_asysc_enable(void)
2076 {
2077 	ASSERT(is_x86_feature(x86_featureset, X86FSET_ASYSC));
2078 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
2079 
2080 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) |
2081 	    (uint64_t)(uintptr_t)AMD_EFER_SCE);
2082 }
2083 
2084 static void
2085 cpu_asysc_disable(void)
2086 {
2087 	ASSERT(is_x86_feature(x86_featureset, X86FSET_ASYSC));
2088 	ASSERT(curthread->t_preempt || getpil() >= LOCK_LEVEL);
2089 
2090 	/*
2091 	 * Turn off the SCE (syscall enable) bit in the EFER register. Software
2092 	 * executing syscall or sysret with this bit off will incur a #ud trap.
2093 	 */
2094 	wrmsr(MSR_AMD_EFER, rdmsr(MSR_AMD_EFER) &
2095 	    ~((uint64_t)(uintptr_t)AMD_EFER_SCE));
2096 }
2097