xref: /illumos-gate/usr/src/uts/i86pc/os/mp_machdep.c (revision b1e2e3fb17324e9ddf43db264a0c64da7756d9e6)
1 
2 /*
3  * CDDL HEADER START
4  *
5  * The contents of this file are subject to the terms of the
6  * Common Development and Distribution License (the "License").
7  * You may not use this file except in compliance with the License.
8  *
9  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
10  * or http://www.opensolaris.org/os/licensing.
11  * See the License for the specific language governing permissions
12  * and limitations under the License.
13  *
14  * When distributing Covered Code, include this CDDL HEADER in each
15  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
16  * If applicable, add the following below this CDDL HEADER, with the
17  * fields enclosed by brackets "[]" replaced with your own identifying
18  * information: Portions Copyright [yyyy] [name of copyright owner]
19  *
20  * CDDL HEADER END
21  */
22 /*
23  * Copyright (c) 1992, 2010, Oracle and/or its affiliates. All rights reserved.
24  */
25 /*
26  * Copyright (c) 2009-2010, Intel Corporation.
27  * All rights reserved.
28  * Copyright 2018 Joyent, Inc.
29  */
30 
31 #define	PSMI_1_7
32 #include <sys/smp_impldefs.h>
33 #include <sys/psm.h>
34 #include <sys/psm_modctl.h>
35 #include <sys/pit.h>
36 #include <sys/cmn_err.h>
37 #include <sys/strlog.h>
38 #include <sys/clock.h>
39 #include <sys/debug.h>
40 #include <sys/rtc.h>
41 #include <sys/x86_archext.h>
42 #include <sys/cpupart.h>
43 #include <sys/cpuvar.h>
44 #include <sys/cpu_event.h>
45 #include <sys/cmt.h>
46 #include <sys/cpu.h>
47 #include <sys/disp.h>
48 #include <sys/archsystm.h>
49 #include <sys/machsystm.h>
50 #include <sys/sysmacros.h>
51 #include <sys/memlist.h>
52 #include <sys/param.h>
53 #include <sys/promif.h>
54 #include <sys/cpu_pm.h>
55 #if defined(__xpv)
56 #include <sys/hypervisor.h>
57 #endif
58 #include <sys/mach_intr.h>
59 #include <vm/hat_i86.h>
60 #include <sys/kdi_machimpl.h>
61 #include <sys/sdt.h>
62 #include <sys/hpet.h>
63 #include <sys/sunddi.h>
64 #include <sys/sunndi.h>
65 #include <sys/cpc_pcbe.h>
66 
67 #define	OFFSETOF(s, m)		(size_t)(&(((s *)0)->m))
68 
69 /*
70  *	Local function prototypes
71  */
72 static int mp_disable_intr(processorid_t cpun);
73 static void mp_enable_intr(processorid_t cpun);
74 static void mach_init();
75 static void mach_picinit();
76 static int machhztomhz(uint64_t cpu_freq_hz);
77 static uint64_t mach_getcpufreq(void);
78 static void mach_fixcpufreq(void);
79 static int mach_clkinit(int, int *);
80 static void mach_smpinit(void);
81 static int mach_softlvl_to_vect(int ipl);
82 static void mach_get_platform(int owner);
83 static void mach_construct_info();
84 static int mach_translate_irq(dev_info_t *dip, int irqno);
85 static int mach_intr_ops(dev_info_t *, ddi_intr_handle_impl_t *,
86     psm_intr_op_t, int *);
87 static void mach_notify_error(int level, char *errmsg);
88 static hrtime_t dummy_hrtime(void);
89 static void dummy_scalehrtime(hrtime_t *);
90 static uint64_t dummy_unscalehrtime(hrtime_t);
91 void cpu_idle(void);
92 static void cpu_wakeup(cpu_t *, int);
93 #ifndef __xpv
94 void cpu_idle_mwait(void);
95 static void cpu_wakeup_mwait(cpu_t *, int);
96 #endif
97 static int mach_cpu_create_devinfo(cpu_t *cp, dev_info_t **dipp);
98 
99 /*
100  *	External reference functions
101  */
102 extern void return_instr();
103 extern uint64_t freq_tsc(uint32_t *);
104 #if defined(__i386)
105 extern uint64_t freq_notsc(uint32_t *);
106 #endif
107 extern void pc_gethrestime(timestruc_t *);
108 extern int cpuid_get_coreid(cpu_t *);
109 extern int cpuid_get_chipid(cpu_t *);
110 
111 /*
112  *	PSM functions initialization
113  */
114 void (*psm_shutdownf)(int, int)	= (void (*)(int, int))return_instr;
115 void (*psm_preshutdownf)(int, int) = (void (*)(int, int))return_instr;
116 void (*psm_notifyf)(int)	= (void (*)(int))return_instr;
117 void (*psm_set_idle_cpuf)(int)	= (void (*)(int))return_instr;
118 void (*psm_unset_idle_cpuf)(int) = (void (*)(int))return_instr;
119 void (*psminitf)()		= mach_init;
120 void (*picinitf)()		= return_instr;
121 int (*clkinitf)(int, int *)	= (int (*)(int, int *))return_instr;
122 int (*ap_mlsetup)()		= (int (*)(void))return_instr;
123 void (*send_dirintf)()		= return_instr;
124 void (*setspl)(int)		= (void (*)(int))return_instr;
125 int (*addspl)(int, int, int, int) = (int (*)(int, int, int, int))return_instr;
126 int (*delspl)(int, int, int, int) = (int (*)(int, int, int, int))return_instr;
127 int (*get_pending_spl)(void)	= (int (*)(void))return_instr;
128 int (*addintr)(void *, int, avfunc, char *, int, caddr_t, caddr_t,
129     uint64_t *, dev_info_t *) = NULL;
130 void (*remintr)(void *, int, avfunc, int) = NULL;
131 void (*kdisetsoftint)(int, struct av_softinfo *)=
132 	(void (*)(int, struct av_softinfo *))return_instr;
133 void (*setsoftint)(int, struct av_softinfo *)=
134 	(void (*)(int, struct av_softinfo *))return_instr;
135 int (*slvltovect)(int)		= (int (*)(int))return_instr;
136 int (*setlvl)(int, int *)	= (int (*)(int, int *))return_instr;
137 void (*setlvlx)(int, int)	= (void (*)(int, int))return_instr;
138 int (*psm_disable_intr)(int)	= mp_disable_intr;
139 void (*psm_enable_intr)(int)	= mp_enable_intr;
140 hrtime_t (*gethrtimef)(void)	= dummy_hrtime;
141 hrtime_t (*gethrtimeunscaledf)(void)	= dummy_hrtime;
142 void (*scalehrtimef)(hrtime_t *)	= dummy_scalehrtime;
143 uint64_t (*unscalehrtimef)(hrtime_t)	= dummy_unscalehrtime;
144 int (*psm_translate_irq)(dev_info_t *, int) = mach_translate_irq;
145 void (*gethrestimef)(timestruc_t *) = pc_gethrestime;
146 void (*psm_notify_error)(int, char *) = (void (*)(int, char *))NULL;
147 int (*psm_get_clockirq)(int) = NULL;
148 int (*psm_get_ipivect)(int, int) = NULL;
149 uchar_t (*psm_get_ioapicid)(uchar_t) = NULL;
150 uint32_t (*psm_get_localapicid)(uint32_t) = NULL;
151 uchar_t (*psm_xlate_vector_by_irq)(uchar_t) = NULL;
152 int (*psm_get_pir_ipivect)(void) = NULL;
153 void (*psm_send_pir_ipi)(processorid_t) = NULL;
154 void (*psm_cmci_setup)(processorid_t, boolean_t) = NULL;
155 
156 int (*psm_clkinit)(int) = NULL;
157 void (*psm_timer_reprogram)(hrtime_t) = NULL;
158 void (*psm_timer_enable)(void) = NULL;
159 void (*psm_timer_disable)(void) = NULL;
160 void (*psm_post_cyclic_setup)(void *arg) = NULL;
161 int (*psm_intr_ops)(dev_info_t *, ddi_intr_handle_impl_t *, psm_intr_op_t,
162     int *) = mach_intr_ops;
163 int (*psm_state)(psm_state_request_t *) = (int (*)(psm_state_request_t *))
164     return_instr;
165 
166 void (*notify_error)(int, char *) = (void (*)(int, char *))return_instr;
167 void (*hrtime_tick)(void)	= return_instr;
168 
169 int (*psm_cpu_create_devinfo)(cpu_t *, dev_info_t **) = mach_cpu_create_devinfo;
170 int (*psm_cpu_get_devinfo)(cpu_t *, dev_info_t **) = NULL;
171 
172 /* global IRM pool for APIX (PSM) module */
173 ddi_irm_pool_t *apix_irm_pool_p = NULL;
174 
175 /*
176  * True if the generic TSC code is our source of hrtime, rather than whatever
177  * the PSM can provide.
178  */
179 #ifdef __xpv
180 int tsc_gethrtime_enable = 0;
181 #else
182 int tsc_gethrtime_enable = 1;
183 #endif
184 int tsc_gethrtime_initted = 0;
185 
186 /*
187  * True if the hrtime implementation is "hires"; namely, better than microdata.
188  */
189 int gethrtime_hires = 0;
190 
191 /*
192  * Local Static Data
193  */
194 static struct psm_ops mach_ops;
195 static struct psm_ops *mach_set[4] = {&mach_ops, NULL, NULL, NULL};
196 static ushort_t mach_ver[4] = {0, 0, 0, 0};
197 
198 /*
199  * virtualization support for psm
200  */
201 void *psm_vt_ops = NULL;
202 /*
203  * If non-zero, idle cpus will become "halted" when there's
204  * no work to do.
205  */
206 int	idle_cpu_use_hlt = 1;
207 
208 #ifndef __xpv
209 /*
210  * If non-zero, idle cpus will use mwait if available to halt instead of hlt.
211  */
212 int	idle_cpu_prefer_mwait = 1;
213 /*
214  * Set to 0 to avoid MONITOR+CLFLUSH assertion.
215  */
216 int	idle_cpu_assert_cflush_monitor = 1;
217 
218 /*
219  * If non-zero, idle cpus will not use power saving Deep C-States idle loop.
220  */
221 int	idle_cpu_no_deep_c = 0;
222 /*
223  * Non-power saving idle loop and wakeup pointers.
224  * Allows user to toggle Deep Idle power saving feature on/off.
225  */
226 void	(*non_deep_idle_cpu)() = cpu_idle;
227 void	(*non_deep_idle_disp_enq_thread)(cpu_t *, int);
228 
229 /*
230  * Object for the kernel to access the HPET.
231  */
232 hpet_t hpet;
233 
234 #endif	/* ifndef __xpv */
235 
236 uint_t cp_haltset_fanout = 0;
237 
238 /*ARGSUSED*/
239 int
240 pg_plat_hw_shared(cpu_t *cp, pghw_type_t hw)
241 {
242 	switch (hw) {
243 	case PGHW_IPIPE:
244 		if (is_x86_feature(x86_featureset, X86FSET_HTT)) {
245 			/*
246 			 * Hyper-threading is SMT
247 			 */
248 			return (1);
249 		} else {
250 			return (0);
251 		}
252 	case PGHW_FPU:
253 		if (cpuid_get_cores_per_compunit(cp) > 1)
254 			return (1);
255 		else
256 			return (0);
257 	case PGHW_PROCNODE:
258 		if (cpuid_get_procnodes_per_pkg(cp) > 1)
259 			return (1);
260 		else
261 			return (0);
262 	case PGHW_CHIP:
263 		if (is_x86_feature(x86_featureset, X86FSET_CMP) ||
264 		    is_x86_feature(x86_featureset, X86FSET_HTT))
265 			return (1);
266 		else
267 			return (0);
268 	case PGHW_CACHE:
269 		if (cpuid_get_ncpu_sharing_last_cache(cp) > 1)
270 			return (1);
271 		else
272 			return (0);
273 	case PGHW_POW_ACTIVE:
274 		if (cpupm_domain_id(cp, CPUPM_DTYPE_ACTIVE) != (id_t)-1)
275 			return (1);
276 		else
277 			return (0);
278 	case PGHW_POW_IDLE:
279 		if (cpupm_domain_id(cp, CPUPM_DTYPE_IDLE) != (id_t)-1)
280 			return (1);
281 		else
282 			return (0);
283 	default:
284 		return (0);
285 	}
286 }
287 
288 /*
289  * Compare two CPUs and see if they have a pghw_type_t sharing relationship
290  * If pghw_type_t is an unsupported hardware type, then return -1
291  */
292 int
293 pg_plat_cpus_share(cpu_t *cpu_a, cpu_t *cpu_b, pghw_type_t hw)
294 {
295 	id_t pgp_a, pgp_b;
296 
297 	pgp_a = pg_plat_hw_instance_id(cpu_a, hw);
298 	pgp_b = pg_plat_hw_instance_id(cpu_b, hw);
299 
300 	if (pgp_a == -1 || pgp_b == -1)
301 		return (-1);
302 
303 	return (pgp_a == pgp_b);
304 }
305 
306 /*
307  * Return a physical instance identifier for known hardware sharing
308  * relationships
309  */
310 id_t
311 pg_plat_hw_instance_id(cpu_t *cpu, pghw_type_t hw)
312 {
313 	switch (hw) {
314 	case PGHW_IPIPE:
315 		return (cpuid_get_coreid(cpu));
316 	case PGHW_CACHE:
317 		return (cpuid_get_last_lvl_cacheid(cpu));
318 	case PGHW_FPU:
319 		return (cpuid_get_compunitid(cpu));
320 	case PGHW_PROCNODE:
321 		return (cpuid_get_procnodeid(cpu));
322 	case PGHW_CHIP:
323 		return (cpuid_get_chipid(cpu));
324 	case PGHW_POW_ACTIVE:
325 		return (cpupm_domain_id(cpu, CPUPM_DTYPE_ACTIVE));
326 	case PGHW_POW_IDLE:
327 		return (cpupm_domain_id(cpu, CPUPM_DTYPE_IDLE));
328 	default:
329 		return (-1);
330 	}
331 }
332 
333 /*
334  * Express preference for optimizing for sharing relationship
335  * hw1 vs hw2
336  */
337 pghw_type_t
338 pg_plat_hw_rank(pghw_type_t hw1, pghw_type_t hw2)
339 {
340 	int i, rank1, rank2;
341 
342 	static pghw_type_t hw_hier[] = {
343 		PGHW_IPIPE,
344 		PGHW_CACHE,
345 		PGHW_FPU,
346 		PGHW_PROCNODE,
347 		PGHW_CHIP,
348 		PGHW_POW_IDLE,
349 		PGHW_POW_ACTIVE,
350 		PGHW_NUM_COMPONENTS
351 	};
352 
353 	for (i = 0; hw_hier[i] != PGHW_NUM_COMPONENTS; i++) {
354 		if (hw_hier[i] == hw1)
355 			rank1 = i;
356 		if (hw_hier[i] == hw2)
357 			rank2 = i;
358 	}
359 
360 	if (rank1 > rank2)
361 		return (hw1);
362 	else
363 		return (hw2);
364 }
365 
366 /*
367  * Override the default CMT dispatcher policy for the specified
368  * hardware sharing relationship
369  */
370 pg_cmt_policy_t
371 pg_plat_cmt_policy(pghw_type_t hw)
372 {
373 	/*
374 	 * For shared caches, also load balance across them to
375 	 * maximize aggregate cache capacity
376 	 *
377 	 * On AMD family 0x15 CPUs, cores come in pairs called
378 	 * compute units, sharing the FPU and the I$ and L2
379 	 * caches. Use balancing and cache affinity.
380 	 */
381 	switch (hw) {
382 	case PGHW_FPU:
383 	case PGHW_CACHE:
384 		return (CMT_BALANCE|CMT_AFFINITY);
385 	default:
386 		return (CMT_NO_POLICY);
387 	}
388 }
389 
390 id_t
391 pg_plat_get_core_id(cpu_t *cpu)
392 {
393 	return ((id_t)cpuid_get_coreid(cpu));
394 }
395 
396 void
397 cmp_set_nosteal_interval(void)
398 {
399 	/* Set the nosteal interval (used by disp_getbest()) to 100us */
400 	nosteal_nsec = 100000UL;
401 }
402 
403 /*
404  * Routine to ensure initial callers to hrtime gets 0 as return
405  */
406 static hrtime_t
407 dummy_hrtime(void)
408 {
409 	return (0);
410 }
411 
412 /* ARGSUSED */
413 static void
414 dummy_scalehrtime(hrtime_t *ticks)
415 {}
416 
417 static uint64_t
418 dummy_unscalehrtime(hrtime_t nsecs)
419 {
420 	return ((uint64_t)nsecs);
421 }
422 
423 /*
424  * Supports Deep C-State power saving idle loop.
425  */
426 void
427 cpu_idle_adaptive(void)
428 {
429 	(*CPU->cpu_m.mcpu_idle_cpu)();
430 }
431 
432 /*
433  * Function called by CPU idle notification framework to check whether CPU
434  * has been awakened. It will be called with interrupt disabled.
435  * If CPU has been awakened, call cpu_idle_exit() to notify CPU idle
436  * notification framework.
437  */
438 /*ARGSUSED*/
439 static void
440 cpu_idle_check_wakeup(void *arg)
441 {
442 	/*
443 	 * Toggle interrupt flag to detect pending interrupts.
444 	 * If interrupt happened, do_interrupt() will notify CPU idle
445 	 * notification framework so no need to call cpu_idle_exit() here.
446 	 */
447 	sti();
448 	SMT_PAUSE();
449 	cli();
450 }
451 
452 /*
453  * Idle the present CPU until wakened via an interrupt
454  */
455 void
456 cpu_idle(void)
457 {
458 	cpu_t		*cpup = CPU;
459 	processorid_t	cpu_sid = cpup->cpu_seqid;
460 	cpupart_t	*cp = cpup->cpu_part;
461 	int		hset_update = 1;
462 
463 	/*
464 	 * If this CPU is online, and there's multiple CPUs
465 	 * in the system, then we should notate our halting
466 	 * by adding ourselves to the partition's halted CPU
467 	 * bitmap. This allows other CPUs to find/awaken us when
468 	 * work becomes available.
469 	 */
470 	if (cpup->cpu_flags & CPU_OFFLINE || ncpus == 1)
471 		hset_update = 0;
472 
473 	/*
474 	 * Add ourselves to the partition's halted CPUs bitmap
475 	 * and set our HALTED flag, if necessary.
476 	 *
477 	 * When a thread becomes runnable, it is placed on the queue
478 	 * and then the halted CPU bitmap is checked to determine who
479 	 * (if anyone) should be awakened. We therefore need to first
480 	 * add ourselves to the bitmap, and and then check if there
481 	 * is any work available. The order is important to prevent a race
482 	 * that can lead to work languishing on a run queue somewhere while
483 	 * this CPU remains halted.
484 	 *
485 	 * Either the producing CPU will see we're halted and will awaken us,
486 	 * or this CPU will see the work available in disp_anywork().
487 	 *
488 	 * Note that memory barriers after updating the HALTED flag
489 	 * are not necessary since an atomic operation (updating the bitset)
490 	 * immediately follows. On x86 the atomic operation acts as a
491 	 * memory barrier for the update of cpu_disp_flags.
492 	 */
493 	if (hset_update) {
494 		cpup->cpu_disp_flags |= CPU_DISP_HALTED;
495 		bitset_atomic_add(&cp->cp_haltset, cpu_sid);
496 	}
497 
498 	/*
499 	 * Check to make sure there's really nothing to do.
500 	 * Work destined for this CPU may become available after
501 	 * this check. We'll be notified through the clearing of our
502 	 * bit in the halted CPU bitmap, and a poke.
503 	 */
504 	if (disp_anywork()) {
505 		if (hset_update) {
506 			cpup->cpu_disp_flags &= ~CPU_DISP_HALTED;
507 			bitset_atomic_del(&cp->cp_haltset, cpu_sid);
508 		}
509 		return;
510 	}
511 
512 	/*
513 	 * We're on our way to being halted.
514 	 *
515 	 * Disable interrupts now, so that we'll awaken immediately
516 	 * after halting if someone tries to poke us between now and
517 	 * the time we actually halt.
518 	 *
519 	 * We check for the presence of our bit after disabling interrupts.
520 	 * If it's cleared, we'll return. If the bit is cleared after
521 	 * we check then the poke will pop us out of the halted state.
522 	 *
523 	 * This means that the ordering of the poke and the clearing
524 	 * of the bit by cpu_wakeup is important.
525 	 * cpu_wakeup() must clear, then poke.
526 	 * cpu_idle() must disable interrupts, then check for the bit.
527 	 */
528 	cli();
529 
530 	if (hset_update && bitset_in_set(&cp->cp_haltset, cpu_sid) == 0) {
531 		cpup->cpu_disp_flags &= ~CPU_DISP_HALTED;
532 		sti();
533 		return;
534 	}
535 
536 	/*
537 	 * The check for anything locally runnable is here for performance
538 	 * and isn't needed for correctness. disp_nrunnable ought to be
539 	 * in our cache still, so it's inexpensive to check, and if there
540 	 * is anything runnable we won't have to wait for the poke.
541 	 */
542 	if (cpup->cpu_disp->disp_nrunnable != 0) {
543 		if (hset_update) {
544 			cpup->cpu_disp_flags &= ~CPU_DISP_HALTED;
545 			bitset_atomic_del(&cp->cp_haltset, cpu_sid);
546 		}
547 		sti();
548 		return;
549 	}
550 
551 	if (cpu_idle_enter(IDLE_STATE_C1, 0,
552 	    cpu_idle_check_wakeup, NULL) == 0) {
553 		mach_cpu_idle();
554 		cpu_idle_exit(CPU_IDLE_CB_FLAG_IDLE);
555 	}
556 
557 	/*
558 	 * We're no longer halted
559 	 */
560 	if (hset_update) {
561 		cpup->cpu_disp_flags &= ~CPU_DISP_HALTED;
562 		bitset_atomic_del(&cp->cp_haltset, cpu_sid);
563 	}
564 }
565 
566 
567 /*
568  * If "cpu" is halted, then wake it up clearing its halted bit in advance.
569  * Otherwise, see if other CPUs in the cpu partition are halted and need to
570  * be woken up so that they can steal the thread we placed on this CPU.
571  * This function is only used on MP systems.
572  */
573 static void
574 cpu_wakeup(cpu_t *cpu, int bound)
575 {
576 	uint_t		cpu_found;
577 	processorid_t	cpu_sid;
578 	cpupart_t	*cp;
579 
580 	cp = cpu->cpu_part;
581 	cpu_sid = cpu->cpu_seqid;
582 	if (bitset_in_set(&cp->cp_haltset, cpu_sid)) {
583 		/*
584 		 * Clear the halted bit for that CPU since it will be
585 		 * poked in a moment.
586 		 */
587 		bitset_atomic_del(&cp->cp_haltset, cpu_sid);
588 		/*
589 		 * We may find the current CPU present in the halted cpuset
590 		 * if we're in the context of an interrupt that occurred
591 		 * before we had a chance to clear our bit in cpu_idle().
592 		 * Poking ourself is obviously unnecessary, since if
593 		 * we're here, we're not halted.
594 		 */
595 		if (cpu != CPU)
596 			poke_cpu(cpu->cpu_id);
597 		return;
598 	} else {
599 		/*
600 		 * This cpu isn't halted, but it's idle or undergoing a
601 		 * context switch. No need to awaken anyone else.
602 		 */
603 		if (cpu->cpu_thread == cpu->cpu_idle_thread ||
604 		    cpu->cpu_disp_flags & CPU_DISP_DONTSTEAL)
605 			return;
606 	}
607 
608 	/*
609 	 * No need to wake up other CPUs if this is for a bound thread.
610 	 */
611 	if (bound)
612 		return;
613 
614 	/*
615 	 * The CPU specified for wakeup isn't currently halted, so check
616 	 * to see if there are any other halted CPUs in the partition,
617 	 * and if there are then awaken one.
618 	 */
619 	do {
620 		cpu_found = bitset_find(&cp->cp_haltset);
621 		if (cpu_found == (uint_t)-1)
622 			return;
623 	} while (bitset_atomic_test_and_del(&cp->cp_haltset, cpu_found) < 0);
624 
625 	if (cpu_found != CPU->cpu_seqid) {
626 		poke_cpu(cpu_seq[cpu_found]->cpu_id);
627 	}
628 }
629 
630 #ifndef __xpv
631 /*
632  * Function called by CPU idle notification framework to check whether CPU
633  * has been awakened. It will be called with interrupt disabled.
634  * If CPU has been awakened, call cpu_idle_exit() to notify CPU idle
635  * notification framework.
636  */
637 static void
638 cpu_idle_mwait_check_wakeup(void *arg)
639 {
640 	volatile uint32_t *mcpu_mwait = (volatile uint32_t *)arg;
641 
642 	ASSERT(arg != NULL);
643 	if (*mcpu_mwait != MWAIT_HALTED) {
644 		/*
645 		 * CPU has been awakened, notify CPU idle notification system.
646 		 */
647 		cpu_idle_exit(CPU_IDLE_CB_FLAG_IDLE);
648 	} else {
649 		/*
650 		 * Toggle interrupt flag to detect pending interrupts.
651 		 * If interrupt happened, do_interrupt() will notify CPU idle
652 		 * notification framework so no need to call cpu_idle_exit()
653 		 * here.
654 		 */
655 		sti();
656 		SMT_PAUSE();
657 		cli();
658 	}
659 }
660 
661 /*
662  * Idle the present CPU until awakened via touching its monitored line
663  */
664 void
665 cpu_idle_mwait(void)
666 {
667 	volatile uint32_t	*mcpu_mwait = CPU->cpu_m.mcpu_mwait;
668 	cpu_t			*cpup = CPU;
669 	processorid_t		cpu_sid = cpup->cpu_seqid;
670 	cpupart_t		*cp = cpup->cpu_part;
671 	int			hset_update = 1;
672 
673 	/*
674 	 * Set our mcpu_mwait here, so we can tell if anyone tries to
675 	 * wake us between now and when we call mwait.  No other cpu will
676 	 * attempt to set our mcpu_mwait until we add ourself to the halted
677 	 * CPU bitmap.
678 	 */
679 	*mcpu_mwait = MWAIT_HALTED;
680 
681 	/*
682 	 * If this CPU is online, and there's multiple CPUs
683 	 * in the system, then we should note our halting
684 	 * by adding ourselves to the partition's halted CPU
685 	 * bitmap. This allows other CPUs to find/awaken us when
686 	 * work becomes available.
687 	 */
688 	if (cpup->cpu_flags & CPU_OFFLINE || ncpus == 1)
689 		hset_update = 0;
690 
691 	/*
692 	 * Add ourselves to the partition's halted CPUs bitmap
693 	 * and set our HALTED flag, if necessary.
694 	 *
695 	 * When a thread becomes runnable, it is placed on the queue
696 	 * and then the halted CPU bitmap is checked to determine who
697 	 * (if anyone) should be awakened. We therefore need to first
698 	 * add ourselves to the bitmap, and and then check if there
699 	 * is any work available.
700 	 *
701 	 * Note that memory barriers after updating the HALTED flag
702 	 * are not necessary since an atomic operation (updating the bitmap)
703 	 * immediately follows. On x86 the atomic operation acts as a
704 	 * memory barrier for the update of cpu_disp_flags.
705 	 */
706 	if (hset_update) {
707 		cpup->cpu_disp_flags |= CPU_DISP_HALTED;
708 		bitset_atomic_add(&cp->cp_haltset, cpu_sid);
709 	}
710 
711 	/*
712 	 * Check to make sure there's really nothing to do.
713 	 * Work destined for this CPU may become available after
714 	 * this check. We'll be notified through the clearing of our
715 	 * bit in the halted CPU bitmap, and a write to our mcpu_mwait.
716 	 *
717 	 * disp_anywork() checks disp_nrunnable, so we do not have to later.
718 	 */
719 	if (disp_anywork()) {
720 		if (hset_update) {
721 			cpup->cpu_disp_flags &= ~CPU_DISP_HALTED;
722 			bitset_atomic_del(&cp->cp_haltset, cpu_sid);
723 		}
724 		return;
725 	}
726 
727 	/*
728 	 * We're on our way to being halted.
729 	 * To avoid a lost wakeup, arm the monitor before checking if another
730 	 * cpu wrote to mcpu_mwait to wake us up.
731 	 */
732 	i86_monitor(mcpu_mwait, 0, 0);
733 	if (*mcpu_mwait == MWAIT_HALTED) {
734 		if (cpu_idle_enter(IDLE_STATE_C1, 0,
735 		    cpu_idle_mwait_check_wakeup, (void *)mcpu_mwait) == 0) {
736 			if (*mcpu_mwait == MWAIT_HALTED) {
737 				i86_mwait(0, 0);
738 			}
739 			cpu_idle_exit(CPU_IDLE_CB_FLAG_IDLE);
740 		}
741 	}
742 
743 	/*
744 	 * We're no longer halted
745 	 */
746 	if (hset_update) {
747 		cpup->cpu_disp_flags &= ~CPU_DISP_HALTED;
748 		bitset_atomic_del(&cp->cp_haltset, cpu_sid);
749 	}
750 }
751 
752 /*
753  * If "cpu" is halted in mwait, then wake it up clearing its halted bit in
754  * advance.  Otherwise, see if other CPUs in the cpu partition are halted and
755  * need to be woken up so that they can steal the thread we placed on this CPU.
756  * This function is only used on MP systems.
757  */
758 static void
759 cpu_wakeup_mwait(cpu_t *cp, int bound)
760 {
761 	cpupart_t	*cpu_part;
762 	uint_t		cpu_found;
763 	processorid_t	cpu_sid;
764 
765 	cpu_part = cp->cpu_part;
766 	cpu_sid = cp->cpu_seqid;
767 
768 	/*
769 	 * Clear the halted bit for that CPU since it will be woken up
770 	 * in a moment.
771 	 */
772 	if (bitset_in_set(&cpu_part->cp_haltset, cpu_sid)) {
773 		/*
774 		 * Clear the halted bit for that CPU since it will be
775 		 * poked in a moment.
776 		 */
777 		bitset_atomic_del(&cpu_part->cp_haltset, cpu_sid);
778 		/*
779 		 * We may find the current CPU present in the halted cpuset
780 		 * if we're in the context of an interrupt that occurred
781 		 * before we had a chance to clear our bit in cpu_idle().
782 		 * Waking ourself is obviously unnecessary, since if
783 		 * we're here, we're not halted.
784 		 *
785 		 * monitor/mwait wakeup via writing to our cache line is
786 		 * harmless and less expensive than always checking if we
787 		 * are waking ourself which is an uncommon case.
788 		 */
789 		MWAIT_WAKEUP(cp);	/* write to monitored line */
790 		return;
791 	} else {
792 		/*
793 		 * This cpu isn't halted, but it's idle or undergoing a
794 		 * context switch. No need to awaken anyone else.
795 		 */
796 		if (cp->cpu_thread == cp->cpu_idle_thread ||
797 		    cp->cpu_disp_flags & CPU_DISP_DONTSTEAL)
798 			return;
799 	}
800 
801 	/*
802 	 * No need to wake up other CPUs if the thread we just enqueued
803 	 * is bound.
804 	 */
805 	if (bound || ncpus == 1)
806 		return;
807 
808 	/*
809 	 * See if there's any other halted CPUs. If there are, then
810 	 * select one, and awaken it.
811 	 * It's possible that after we find a CPU, somebody else
812 	 * will awaken it before we get the chance.
813 	 * In that case, look again.
814 	 */
815 	do {
816 		cpu_found = bitset_find(&cpu_part->cp_haltset);
817 		if (cpu_found == (uint_t)-1)
818 			return;
819 	} while (bitset_atomic_test_and_del(&cpu_part->cp_haltset,
820 	    cpu_found) < 0);
821 
822 	/*
823 	 * Do not check if cpu_found is ourself as monitor/mwait
824 	 * wakeup is cheap.
825 	 */
826 	MWAIT_WAKEUP(cpu_seq[cpu_found]); /* write to monitored line */
827 }
828 
829 #endif
830 
831 void (*cpu_pause_handler)(volatile char *) = NULL;
832 
833 static int
834 mp_disable_intr(int cpun)
835 {
836 	/*
837 	 * switch to the offline cpu
838 	 */
839 	affinity_set(cpun);
840 	/*
841 	 * raise ipl to just below cross call
842 	 */
843 	splx(XC_SYS_PIL - 1);
844 	/*
845 	 *	set base spl to prevent the next swtch to idle from
846 	 *	lowering back to ipl 0
847 	 */
848 	CPU->cpu_intr_actv |= (1 << (XC_SYS_PIL - 1));
849 	set_base_spl();
850 	affinity_clear();
851 	return (DDI_SUCCESS);
852 }
853 
854 static void
855 mp_enable_intr(int cpun)
856 {
857 	/*
858 	 * switch to the online cpu
859 	 */
860 	affinity_set(cpun);
861 	/*
862 	 * clear the interrupt active mask
863 	 */
864 	CPU->cpu_intr_actv &= ~(1 << (XC_SYS_PIL - 1));
865 	set_base_spl();
866 	(void) spl0();
867 	affinity_clear();
868 }
869 
870 static void
871 mach_get_platform(int owner)
872 {
873 	void		**srv_opsp;
874 	void		**clt_opsp;
875 	int		i;
876 	int		total_ops;
877 
878 	/* fix up psm ops */
879 	srv_opsp = (void **)mach_set[0];
880 	clt_opsp = (void **)mach_set[owner];
881 	if (mach_ver[owner] == (ushort_t)PSM_INFO_VER01)
882 		total_ops = sizeof (struct psm_ops_ver01) /
883 		    sizeof (void (*)(void));
884 	else if (mach_ver[owner] == (ushort_t)PSM_INFO_VER01_1)
885 		/* no psm_notify_func */
886 		total_ops = OFFSETOF(struct psm_ops, psm_notify_func) /
887 		    sizeof (void (*)(void));
888 	else if (mach_ver[owner] == (ushort_t)PSM_INFO_VER01_2)
889 		/* no psm_timer funcs */
890 		total_ops = OFFSETOF(struct psm_ops, psm_timer_reprogram) /
891 		    sizeof (void (*)(void));
892 	else if (mach_ver[owner] == (ushort_t)PSM_INFO_VER01_3)
893 		/* no psm_preshutdown function */
894 		total_ops = OFFSETOF(struct psm_ops, psm_preshutdown) /
895 		    sizeof (void (*)(void));
896 	else if (mach_ver[owner] == (ushort_t)PSM_INFO_VER01_4)
897 		/* no psm_intr_ops function */
898 		total_ops = OFFSETOF(struct psm_ops, psm_intr_ops) /
899 		    sizeof (void (*)(void));
900 	else if (mach_ver[owner] == (ushort_t)PSM_INFO_VER01_5)
901 		/* no psm_state function */
902 		total_ops = OFFSETOF(struct psm_ops, psm_state) /
903 		    sizeof (void (*)(void));
904 	else if (mach_ver[owner] == (ushort_t)PSM_INFO_VER01_6)
905 		/* no psm_cpu_ops function */
906 		total_ops = OFFSETOF(struct psm_ops, psm_cpu_ops) /
907 		    sizeof (void (*)(void));
908 	else
909 		total_ops = sizeof (struct psm_ops) / sizeof (void (*)(void));
910 
911 	/*
912 	 * Save the version of the PSM module, in case we need to
913 	 * behave differently based on version.
914 	 */
915 	mach_ver[0] = mach_ver[owner];
916 
917 	for (i = 0; i < total_ops; i++)
918 		if (clt_opsp[i] != NULL)
919 			srv_opsp[i] = clt_opsp[i];
920 }
921 
922 static void
923 mach_construct_info()
924 {
925 	struct psm_sw *swp;
926 	int	mach_cnt[PSM_OWN_OVERRIDE+1] = {0};
927 	int	conflict_owner = 0;
928 
929 	if (psmsw->psw_forw == psmsw)
930 		panic("No valid PSM modules found");
931 	mutex_enter(&psmsw_lock);
932 	for (swp = psmsw->psw_forw; swp != psmsw; swp = swp->psw_forw) {
933 		if (!(swp->psw_flag & PSM_MOD_IDENTIFY))
934 			continue;
935 		mach_set[swp->psw_infop->p_owner] = swp->psw_infop->p_ops;
936 		mach_ver[swp->psw_infop->p_owner] = swp->psw_infop->p_version;
937 		mach_cnt[swp->psw_infop->p_owner]++;
938 	}
939 	mutex_exit(&psmsw_lock);
940 
941 	mach_get_platform(PSM_OWN_SYS_DEFAULT);
942 
943 	/* check to see are there any conflicts */
944 	if (mach_cnt[PSM_OWN_EXCLUSIVE] > 1)
945 		conflict_owner = PSM_OWN_EXCLUSIVE;
946 	if (mach_cnt[PSM_OWN_OVERRIDE] > 1)
947 		conflict_owner = PSM_OWN_OVERRIDE;
948 	if (conflict_owner) {
949 		/* remove all psm modules except uppc */
950 		cmn_err(CE_WARN,
951 		    "Conflicts detected on the following PSM modules:");
952 		mutex_enter(&psmsw_lock);
953 		for (swp = psmsw->psw_forw; swp != psmsw; swp = swp->psw_forw) {
954 			if (swp->psw_infop->p_owner == conflict_owner)
955 				cmn_err(CE_WARN, "%s ",
956 				    swp->psw_infop->p_mach_idstring);
957 		}
958 		mutex_exit(&psmsw_lock);
959 		cmn_err(CE_WARN,
960 		    "Setting the system back to SINGLE processor mode!");
961 		cmn_err(CE_WARN,
962 		    "Please edit /etc/mach to remove the invalid PSM module.");
963 		return;
964 	}
965 
966 	if (mach_set[PSM_OWN_EXCLUSIVE])
967 		mach_get_platform(PSM_OWN_EXCLUSIVE);
968 
969 	if (mach_set[PSM_OWN_OVERRIDE])
970 		mach_get_platform(PSM_OWN_OVERRIDE);
971 }
972 
973 static void
974 mach_init()
975 {
976 	struct psm_ops  *pops;
977 
978 	mach_construct_info();
979 
980 	pops = mach_set[0];
981 
982 	/* register the interrupt and clock initialization rotuines */
983 	picinitf = mach_picinit;
984 	clkinitf = mach_clkinit;
985 	psm_get_clockirq = pops->psm_get_clockirq;
986 
987 	/* register the interrupt setup code */
988 	slvltovect = mach_softlvl_to_vect;
989 	addspl	= pops->psm_addspl;
990 	delspl	= pops->psm_delspl;
991 
992 	if (pops->psm_translate_irq)
993 		psm_translate_irq = pops->psm_translate_irq;
994 	if (pops->psm_intr_ops)
995 		psm_intr_ops = pops->psm_intr_ops;
996 
997 #if defined(PSMI_1_2) || defined(PSMI_1_3) || defined(PSMI_1_4)
998 	/*
999 	 * Time-of-day functionality now handled in TOD modules.
1000 	 * (Warn about PSM modules that think that we're going to use
1001 	 * their ops vectors.)
1002 	 */
1003 	if (pops->psm_tod_get)
1004 		cmn_err(CE_WARN, "obsolete psm_tod_get op %p",
1005 		    (void *)pops->psm_tod_get);
1006 
1007 	if (pops->psm_tod_set)
1008 		cmn_err(CE_WARN, "obsolete psm_tod_set op %p",
1009 		    (void *)pops->psm_tod_set);
1010 #endif
1011 
1012 	if (pops->psm_notify_error) {
1013 		psm_notify_error = mach_notify_error;
1014 		notify_error = pops->psm_notify_error;
1015 	}
1016 
1017 	(*pops->psm_softinit)();
1018 
1019 	/*
1020 	 * Initialize the dispatcher's function hooks to enable CPU halting
1021 	 * when idle.  Set both the deep-idle and non-deep-idle hooks.
1022 	 *
1023 	 * Assume we can use power saving deep-idle loop cpu_idle_adaptive.
1024 	 * Platform deep-idle driver will reset our idle loop to
1025 	 * non_deep_idle_cpu if power saving deep-idle feature is not available.
1026 	 *
1027 	 * Do not use monitor/mwait if idle_cpu_use_hlt is not set(spin idle)
1028 	 * or idle_cpu_prefer_mwait is not set.
1029 	 * Allocate monitor/mwait buffer for cpu0.
1030 	 */
1031 #ifndef __xpv
1032 	non_deep_idle_disp_enq_thread = disp_enq_thread;
1033 #endif
1034 	if (idle_cpu_use_hlt) {
1035 		idle_cpu = cpu_idle_adaptive;
1036 		CPU->cpu_m.mcpu_idle_cpu = cpu_idle;
1037 #ifndef __xpv
1038 		if (is_x86_feature(x86_featureset, X86FSET_MWAIT) &&
1039 		    idle_cpu_prefer_mwait) {
1040 			CPU->cpu_m.mcpu_mwait = cpuid_mwait_alloc(CPU);
1041 			/*
1042 			 * Protect ourself from insane mwait size.
1043 			 */
1044 			if (CPU->cpu_m.mcpu_mwait == NULL) {
1045 #ifdef DEBUG
1046 				cmn_err(CE_NOTE, "Using hlt idle.  Cannot "
1047 				    "handle cpu 0 mwait size.");
1048 #endif
1049 				idle_cpu_prefer_mwait = 0;
1050 				CPU->cpu_m.mcpu_idle_cpu = cpu_idle;
1051 			} else {
1052 				CPU->cpu_m.mcpu_idle_cpu = cpu_idle_mwait;
1053 			}
1054 		} else {
1055 			CPU->cpu_m.mcpu_idle_cpu = cpu_idle;
1056 		}
1057 		non_deep_idle_cpu = CPU->cpu_m.mcpu_idle_cpu;
1058 
1059 		/*
1060 		 * Disable power saving deep idle loop?
1061 		 */
1062 		if (idle_cpu_no_deep_c) {
1063 			idle_cpu = non_deep_idle_cpu;
1064 		}
1065 #endif
1066 	}
1067 
1068 	mach_smpinit();
1069 }
1070 
1071 static void
1072 mach_smpinit(void)
1073 {
1074 	struct psm_ops  *pops;
1075 	processorid_t cpu_id;
1076 	int cnt;
1077 	cpuset_t cpumask;
1078 
1079 	pops = mach_set[0];
1080 	CPUSET_ZERO(cpumask);
1081 
1082 	cpu_id = -1;
1083 	cpu_id = (*pops->psm_get_next_processorid)(cpu_id);
1084 	/*
1085 	 * Only add boot_ncpus CPUs to mp_cpus. Other CPUs will be handled
1086 	 * by CPU DR driver at runtime.
1087 	 */
1088 	for (cnt = 0; cpu_id != -1 && cnt < boot_ncpus; cnt++) {
1089 		CPUSET_ADD(cpumask, cpu_id);
1090 		cpu_id = (*pops->psm_get_next_processorid)(cpu_id);
1091 	}
1092 
1093 	mp_cpus = cpumask;
1094 
1095 	/* MP related routines */
1096 	ap_mlsetup = pops->psm_post_cpu_start;
1097 	send_dirintf = pops->psm_send_ipi;
1098 
1099 	/* optional MP related routines */
1100 	if (pops->psm_shutdown)
1101 		psm_shutdownf = pops->psm_shutdown;
1102 	if (pops->psm_preshutdown)
1103 		psm_preshutdownf = pops->psm_preshutdown;
1104 	if (pops->psm_notify_func)
1105 		psm_notifyf = pops->psm_notify_func;
1106 	if (pops->psm_set_idlecpu)
1107 		psm_set_idle_cpuf = pops->psm_set_idlecpu;
1108 	if (pops->psm_unset_idlecpu)
1109 		psm_unset_idle_cpuf = pops->psm_unset_idlecpu;
1110 
1111 	psm_clkinit = pops->psm_clkinit;
1112 
1113 	if (pops->psm_timer_reprogram)
1114 		psm_timer_reprogram = pops->psm_timer_reprogram;
1115 
1116 	if (pops->psm_timer_enable)
1117 		psm_timer_enable = pops->psm_timer_enable;
1118 
1119 	if (pops->psm_timer_disable)
1120 		psm_timer_disable = pops->psm_timer_disable;
1121 
1122 	if (pops->psm_post_cyclic_setup)
1123 		psm_post_cyclic_setup = pops->psm_post_cyclic_setup;
1124 
1125 	if (pops->psm_state)
1126 		psm_state = pops->psm_state;
1127 
1128 	/*
1129 	 * Set these vectors here so they can be used by Suspend/Resume
1130 	 * on UP machines.
1131 	 */
1132 	if (pops->psm_disable_intr)
1133 		psm_disable_intr = pops->psm_disable_intr;
1134 	if (pops->psm_enable_intr)
1135 		psm_enable_intr  = pops->psm_enable_intr;
1136 
1137 	/* check for multiple CPUs */
1138 	if (cnt < 2 && plat_dr_support_cpu() == B_FALSE)
1139 		return;
1140 
1141 	/* check for MP platforms */
1142 	if (pops->psm_cpu_start == NULL)
1143 		return;
1144 
1145 	/*
1146 	 * Set the dispatcher hook to enable cpu "wake up"
1147 	 * when a thread becomes runnable.
1148 	 */
1149 	if (idle_cpu_use_hlt) {
1150 		disp_enq_thread = cpu_wakeup;
1151 #ifndef __xpv
1152 		if (is_x86_feature(x86_featureset, X86FSET_MWAIT) &&
1153 		    idle_cpu_prefer_mwait)
1154 			disp_enq_thread = cpu_wakeup_mwait;
1155 		non_deep_idle_disp_enq_thread = disp_enq_thread;
1156 #endif
1157 	}
1158 
1159 	psm_get_ipivect = pops->psm_get_ipivect;
1160 	psm_get_pir_ipivect = pops->psm_get_pir_ipivect;
1161 	psm_send_pir_ipi = pops->psm_send_pir_ipi;
1162 	psm_cmci_setup = pops->psm_cmci_setup;
1163 
1164 
1165 	(void) add_avintr((void *)NULL, XC_HI_PIL, xc_serv, "xc_intr",
1166 	    (*pops->psm_get_ipivect)(XC_HI_PIL, PSM_INTR_IPI_HI),
1167 	    NULL, NULL, NULL, NULL);
1168 
1169 	(void) (*pops->psm_get_ipivect)(XC_CPUPOKE_PIL, PSM_INTR_POKE);
1170 }
1171 
1172 static void
1173 mach_picinit()
1174 {
1175 	struct psm_ops  *pops;
1176 
1177 	pops = mach_set[0];
1178 
1179 	/* register the interrupt handlers */
1180 	setlvl = pops->psm_intr_enter;
1181 	setlvlx = pops->psm_intr_exit;
1182 
1183 	/* initialize the interrupt hardware */
1184 	(*pops->psm_picinit)();
1185 
1186 	/* set interrupt mask for current ipl */
1187 	setspl = pops->psm_setspl;
1188 	cli();
1189 	setspl(CPU->cpu_pri);
1190 }
1191 
1192 uint_t	cpu_freq;	/* MHz */
1193 uint64_t cpu_freq_hz;	/* measured (in hertz) */
1194 
1195 #define	MEGA_HZ		1000000
1196 
1197 #ifdef __xpv
1198 
1199 int xpv_cpufreq_workaround = 1;
1200 int xpv_cpufreq_verbose = 0;
1201 
1202 #else	/* __xpv */
1203 
1204 static uint64_t
1205 mach_calchz(uint32_t pit_counter, uint64_t *processor_clks)
1206 {
1207 	uint64_t cpu_hz;
1208 
1209 	if ((pit_counter == 0) || (*processor_clks == 0) ||
1210 	    (*processor_clks > (((uint64_t)-1) / PIT_HZ)))
1211 		return (0);
1212 
1213 	cpu_hz = ((uint64_t)PIT_HZ * *processor_clks) / pit_counter;
1214 
1215 	return (cpu_hz);
1216 }
1217 
1218 #endif	/* __xpv */
1219 
1220 static uint64_t
1221 mach_getcpufreq(void)
1222 {
1223 #if defined(__xpv)
1224 	vcpu_time_info_t *vti = &CPU->cpu_m.mcpu_vcpu_info->time;
1225 	uint64_t cpu_hz;
1226 
1227 	/*
1228 	 * During dom0 bringup, it was noted that on at least one older
1229 	 * Intel HT machine, the hypervisor initially gives a tsc_to_system_mul
1230 	 * value that is quite wrong (the 3.06GHz clock was reported
1231 	 * as 4.77GHz)
1232 	 *
1233 	 * The curious thing is, that if you stop the kernel at entry,
1234 	 * breakpoint here and inspect the value with kmdb, the value
1235 	 * is correct - but if you don't stop and simply enable the
1236 	 * printf statement (below), you can see the bad value printed
1237 	 * here.  Almost as if something kmdb did caused the hypervisor to
1238 	 * figure it out correctly.  And, note that the hypervisor
1239 	 * eventually -does- figure it out correctly ... if you look at
1240 	 * the field later in the life of dom0, it is correct.
1241 	 *
1242 	 * For now, on dom0, we employ a slightly cheesy workaround of
1243 	 * using the DOM0_PHYSINFO hypercall.
1244 	 */
1245 	if (DOMAIN_IS_INITDOMAIN(xen_info) && xpv_cpufreq_workaround) {
1246 		cpu_hz = 1000 * xpv_cpu_khz();
1247 	} else {
1248 		cpu_hz = (UINT64_C(1000000000) << 32) / vti->tsc_to_system_mul;
1249 
1250 		if (vti->tsc_shift < 0)
1251 			cpu_hz <<= -vti->tsc_shift;
1252 		else
1253 			cpu_hz >>= vti->tsc_shift;
1254 	}
1255 
1256 	if (xpv_cpufreq_verbose)
1257 		printf("mach_getcpufreq: system_mul 0x%x, shift %d, "
1258 		    "cpu_hz %" PRId64 "Hz\n",
1259 		    vti->tsc_to_system_mul, vti->tsc_shift, cpu_hz);
1260 
1261 	return (cpu_hz);
1262 #else	/* __xpv */
1263 	uint32_t pit_counter;
1264 	uint64_t processor_clks;
1265 
1266 	if (is_x86_feature(x86_featureset, X86FSET_TSC)) {
1267 		/*
1268 		 * We have a TSC. freq_tsc() knows how to measure the number
1269 		 * of clock cycles sampled against the PIT.
1270 		 */
1271 		ulong_t flags = clear_int_flag();
1272 		processor_clks = freq_tsc(&pit_counter);
1273 		restore_int_flag(flags);
1274 		return (mach_calchz(pit_counter, &processor_clks));
1275 	} else if (x86_vendor == X86_VENDOR_Cyrix || x86_type == X86_TYPE_P5) {
1276 #if defined(__amd64)
1277 		panic("mach_getcpufreq: no TSC!");
1278 #elif defined(__i386)
1279 		/*
1280 		 * We are a Cyrix based on a 6x86 core or an Intel Pentium
1281 		 * for which freq_notsc() knows how to measure the number of
1282 		 * elapsed clock cycles sampled against the PIT
1283 		 */
1284 		ulong_t flags = clear_int_flag();
1285 		processor_clks = freq_notsc(&pit_counter);
1286 		restore_int_flag(flags);
1287 		return (mach_calchz(pit_counter, &processor_clks));
1288 #endif	/* __i386 */
1289 	}
1290 
1291 	/* We do not know how to calculate cpu frequency for this cpu. */
1292 	return (0);
1293 #endif	/* __xpv */
1294 }
1295 
1296 /*
1297  * If the clock speed of a cpu is found to be reported incorrectly, do not add
1298  * to this array, instead improve the accuracy of the algorithm that determines
1299  * the clock speed of the processor or extend the implementation to support the
1300  * vendor as appropriate. This is here only to support adjusting the speed on
1301  * older slower processors that mach_fixcpufreq() would not be able to account
1302  * for otherwise.
1303  */
1304 static int x86_cpu_freq[] = { 60, 75, 80, 90, 120, 160, 166, 175, 180, 233 };
1305 
1306 /*
1307  * On fast processors the clock frequency that is measured may be off by
1308  * a few MHz from the value printed on the part. This is a combination of
1309  * the factors that for such fast parts being off by this much is within
1310  * the tolerances for manufacture and because of the difficulties in the
1311  * measurement that can lead to small error. This function uses some
1312  * heuristics in order to tweak the value that was measured to match what
1313  * is most likely printed on the part.
1314  *
1315  * Some examples:
1316  *	AMD Athlon 1000 mhz measured as 998 mhz
1317  *	Intel Pentium III Xeon 733 mhz measured as 731 mhz
1318  *	Intel Pentium IV 1500 mhz measured as 1495mhz
1319  *
1320  * If in the future this function is no longer sufficient to correct
1321  * for the error in the measurement, then the algorithm used to perform
1322  * the measurement will have to be improved in order to increase accuracy
1323  * rather than adding horrible and questionable kludges here.
1324  *
1325  * This is called after the cyclics subsystem because of the potential
1326  * that the heuristics within may give a worse estimate of the clock
1327  * frequency than the value that was measured.
1328  */
1329 static void
1330 mach_fixcpufreq(void)
1331 {
1332 	uint32_t freq, mul, near66, delta66, near50, delta50, fixed, delta, i;
1333 
1334 	freq = (uint32_t)cpu_freq;
1335 
1336 	/*
1337 	 * Find the nearest integer multiple of 200/3 (about 66) MHz to the
1338 	 * measured speed taking into account that the 667 MHz parts were
1339 	 * the first to round-up.
1340 	 */
1341 	mul = (uint32_t)((3 * (uint64_t)freq + 100) / 200);
1342 	near66 = (uint32_t)((200 * (uint64_t)mul + ((mul >= 10) ? 1 : 0)) / 3);
1343 	delta66 = (near66 > freq) ? (near66 - freq) : (freq - near66);
1344 
1345 	/* Find the nearest integer multiple of 50 MHz to the measured speed */
1346 	mul = (freq + 25) / 50;
1347 	near50 = mul * 50;
1348 	delta50 = (near50 > freq) ? (near50 - freq) : (freq - near50);
1349 
1350 	/* Find the closer of the two */
1351 	if (delta66 < delta50) {
1352 		fixed = near66;
1353 		delta = delta66;
1354 	} else {
1355 		fixed = near50;
1356 		delta = delta50;
1357 	}
1358 
1359 	if (fixed > INT_MAX)
1360 		return;
1361 
1362 	/*
1363 	 * Some older parts have a core clock frequency that is not an
1364 	 * integral multiple of 50 or 66 MHz. Check if one of the old
1365 	 * clock frequencies is closer to the measured value than any
1366 	 * of the integral multiples of 50 an 66, and if so set fixed
1367 	 * and delta appropriately to represent the closest value.
1368 	 */
1369 	i = sizeof (x86_cpu_freq) / sizeof (int);
1370 	while (i > 0) {
1371 		i--;
1372 
1373 		if (x86_cpu_freq[i] <= freq) {
1374 			mul = freq - x86_cpu_freq[i];
1375 
1376 			if (mul < delta) {
1377 				fixed = x86_cpu_freq[i];
1378 				delta = mul;
1379 			}
1380 
1381 			break;
1382 		}
1383 
1384 		mul = x86_cpu_freq[i] - freq;
1385 
1386 		if (mul < delta) {
1387 			fixed = x86_cpu_freq[i];
1388 			delta = mul;
1389 		}
1390 	}
1391 
1392 	/*
1393 	 * Set a reasonable maximum for how much to correct the measured
1394 	 * result by. This check is here to prevent the adjustment made
1395 	 * by this function from being more harm than good. It is entirely
1396 	 * possible that in the future parts will be made that are not
1397 	 * integral multiples of 66 or 50 in clock frequency or that
1398 	 * someone may overclock a part to some odd frequency. If the
1399 	 * measured value is farther from the corrected value than
1400 	 * allowed, then assume the corrected value is in error and use
1401 	 * the measured value.
1402 	 */
1403 	if (6 < delta)
1404 		return;
1405 
1406 	cpu_freq = (int)fixed;
1407 }
1408 
1409 
1410 static int
1411 machhztomhz(uint64_t cpu_freq_hz)
1412 {
1413 	uint64_t cpu_mhz;
1414 
1415 	/* Round to nearest MHZ */
1416 	cpu_mhz = (cpu_freq_hz + (MEGA_HZ / 2)) / MEGA_HZ;
1417 
1418 	if (cpu_mhz > INT_MAX)
1419 		return (0);
1420 
1421 	return ((int)cpu_mhz);
1422 
1423 }
1424 
1425 
1426 static int
1427 mach_clkinit(int preferred_mode, int *set_mode)
1428 {
1429 	struct psm_ops  *pops;
1430 	int resolution;
1431 
1432 	pops = mach_set[0];
1433 
1434 	cpu_freq_hz = mach_getcpufreq();
1435 
1436 	cpu_freq = machhztomhz(cpu_freq_hz);
1437 
1438 	if (!is_x86_feature(x86_featureset, X86FSET_TSC) || (cpu_freq == 0))
1439 		tsc_gethrtime_enable = 0;
1440 
1441 #ifndef __xpv
1442 	if (tsc_gethrtime_enable) {
1443 		tsc_hrtimeinit(cpu_freq_hz);
1444 	} else
1445 #endif
1446 	{
1447 		if (pops->psm_hrtimeinit)
1448 			(*pops->psm_hrtimeinit)();
1449 		gethrtimef = pops->psm_gethrtime;
1450 		gethrtimeunscaledf = gethrtimef;
1451 		/* scalehrtimef will remain dummy */
1452 	}
1453 
1454 	mach_fixcpufreq();
1455 
1456 	if (mach_ver[0] >= PSM_INFO_VER01_3) {
1457 		if (preferred_mode == TIMER_ONESHOT) {
1458 
1459 			resolution = (*pops->psm_clkinit)(0);
1460 			if (resolution != 0)  {
1461 				*set_mode = TIMER_ONESHOT;
1462 				return (resolution);
1463 			}
1464 		}
1465 
1466 		/*
1467 		 * either periodic mode was requested or could not set to
1468 		 * one-shot mode
1469 		 */
1470 		resolution = (*pops->psm_clkinit)(hz);
1471 		/*
1472 		 * psm should be able to do periodic, so we do not check
1473 		 * for return value of psm_clkinit here.
1474 		 */
1475 		*set_mode = TIMER_PERIODIC;
1476 		return (resolution);
1477 	} else {
1478 		/*
1479 		 * PSMI interface prior to PSMI_3 does not define a return
1480 		 * value for psm_clkinit, so the return value is ignored.
1481 		 */
1482 		(void) (*pops->psm_clkinit)(hz);
1483 		*set_mode = TIMER_PERIODIC;
1484 		return (nsec_per_tick);
1485 	}
1486 }
1487 
1488 
1489 /*ARGSUSED*/
1490 static int
1491 mach_softlvl_to_vect(int ipl)
1492 {
1493 	setsoftint = av_set_softint_pending;
1494 	kdisetsoftint = kdi_av_set_softint_pending;
1495 
1496 	return (PSM_SV_SOFTWARE);
1497 }
1498 
1499 #ifdef DEBUG
1500 /*
1501  * This is here to allow us to simulate cpus that refuse to start.
1502  */
1503 cpuset_t cpufailset;
1504 #endif
1505 
1506 int
1507 mach_cpu_start(struct cpu *cp, void *ctx)
1508 {
1509 	struct psm_ops *pops = mach_set[0];
1510 	processorid_t id = cp->cpu_id;
1511 
1512 #ifdef DEBUG
1513 	if (CPU_IN_SET(cpufailset, id))
1514 		return (0);
1515 #endif
1516 	return ((*pops->psm_cpu_start)(id, ctx));
1517 }
1518 
1519 int
1520 mach_cpuid_start(processorid_t id, void *ctx)
1521 {
1522 	struct psm_ops *pops = mach_set[0];
1523 
1524 #ifdef DEBUG
1525 	if (CPU_IN_SET(cpufailset, id))
1526 		return (0);
1527 #endif
1528 	return ((*pops->psm_cpu_start)(id, ctx));
1529 }
1530 
1531 int
1532 mach_cpu_stop(cpu_t *cp, void *ctx)
1533 {
1534 	struct psm_ops *pops = mach_set[0];
1535 	psm_cpu_request_t request;
1536 
1537 	if (pops->psm_cpu_ops == NULL) {
1538 		return (ENOTSUP);
1539 	}
1540 
1541 	ASSERT(cp->cpu_id != -1);
1542 	request.pcr_cmd = PSM_CPU_STOP;
1543 	request.req.cpu_stop.cpuid = cp->cpu_id;
1544 	request.req.cpu_stop.ctx = ctx;
1545 
1546 	return ((*pops->psm_cpu_ops)(&request));
1547 }
1548 
1549 int
1550 mach_cpu_add(mach_cpu_add_arg_t *argp, processorid_t *cpuidp)
1551 {
1552 	int rc;
1553 	struct psm_ops *pops = mach_set[0];
1554 	psm_cpu_request_t request;
1555 
1556 	if (pops->psm_cpu_ops == NULL) {
1557 		return (ENOTSUP);
1558 	}
1559 
1560 	request.pcr_cmd = PSM_CPU_ADD;
1561 	request.req.cpu_add.argp = argp;
1562 	request.req.cpu_add.cpuid = -1;
1563 	rc = (*pops->psm_cpu_ops)(&request);
1564 	if (rc == 0) {
1565 		ASSERT(request.req.cpu_add.cpuid != -1);
1566 		*cpuidp = request.req.cpu_add.cpuid;
1567 	}
1568 
1569 	return (rc);
1570 }
1571 
1572 int
1573 mach_cpu_remove(processorid_t cpuid)
1574 {
1575 	struct psm_ops *pops = mach_set[0];
1576 	psm_cpu_request_t request;
1577 
1578 	if (pops->psm_cpu_ops == NULL) {
1579 		return (ENOTSUP);
1580 	}
1581 
1582 	request.pcr_cmd = PSM_CPU_REMOVE;
1583 	request.req.cpu_remove.cpuid = cpuid;
1584 
1585 	return ((*pops->psm_cpu_ops)(&request));
1586 }
1587 
1588 /*
1589  * Default handler to create device node for CPU.
1590  * One reference count will be held on created device node.
1591  */
1592 static int
1593 mach_cpu_create_devinfo(cpu_t *cp, dev_info_t **dipp)
1594 {
1595 	int rv, circ;
1596 	dev_info_t *dip;
1597 	static kmutex_t cpu_node_lock;
1598 	static dev_info_t *cpu_nex_devi = NULL;
1599 
1600 	ASSERT(cp != NULL);
1601 	ASSERT(dipp != NULL);
1602 	*dipp = NULL;
1603 
1604 	if (cpu_nex_devi == NULL) {
1605 		mutex_enter(&cpu_node_lock);
1606 		/* First check whether cpus exists. */
1607 		cpu_nex_devi = ddi_find_devinfo("cpus", -1, 0);
1608 		/* Create cpus if it doesn't exist. */
1609 		if (cpu_nex_devi == NULL) {
1610 			ndi_devi_enter(ddi_root_node(), &circ);
1611 			rv = ndi_devi_alloc(ddi_root_node(), "cpus",
1612 			    (pnode_t)DEVI_SID_NODEID, &dip);
1613 			if (rv != NDI_SUCCESS) {
1614 				mutex_exit(&cpu_node_lock);
1615 				cmn_err(CE_CONT,
1616 				    "?failed to create cpu nexus device.\n");
1617 				return (PSM_FAILURE);
1618 			}
1619 			ASSERT(dip != NULL);
1620 			(void) ndi_devi_online(dip, 0);
1621 			ndi_devi_exit(ddi_root_node(), circ);
1622 			cpu_nex_devi = dip;
1623 		}
1624 		mutex_exit(&cpu_node_lock);
1625 	}
1626 
1627 	/*
1628 	 * create a child node for cpu identified as 'cpu_id'
1629 	 */
1630 	ndi_devi_enter(cpu_nex_devi, &circ);
1631 	dip = ddi_add_child(cpu_nex_devi, "cpu", DEVI_SID_NODEID, -1);
1632 	if (dip == NULL) {
1633 		cmn_err(CE_CONT,
1634 		    "?failed to create device node for cpu%d.\n", cp->cpu_id);
1635 		rv = PSM_FAILURE;
1636 	} else {
1637 		*dipp = dip;
1638 		(void) ndi_hold_devi(dip);
1639 		rv = PSM_SUCCESS;
1640 	}
1641 	ndi_devi_exit(cpu_nex_devi, circ);
1642 
1643 	return (rv);
1644 }
1645 
1646 /*
1647  * Create cpu device node in device tree and online it.
1648  * Return created dip with reference count held if requested.
1649  */
1650 int
1651 mach_cpu_create_device_node(struct cpu *cp, dev_info_t **dipp)
1652 {
1653 	int rv;
1654 	dev_info_t *dip = NULL;
1655 
1656 	ASSERT(psm_cpu_create_devinfo != NULL);
1657 	rv = psm_cpu_create_devinfo(cp, &dip);
1658 	if (rv == PSM_SUCCESS) {
1659 		cpuid_set_cpu_properties(dip, cp->cpu_id, cp->cpu_m.mcpu_cpi);
1660 		/* Recursively attach driver for parent nexus device. */
1661 		if (i_ddi_attach_node_hierarchy(ddi_get_parent(dip)) ==
1662 		    DDI_SUCCESS) {
1663 			/* Configure cpu itself and descendants. */
1664 			(void) ndi_devi_online(dip,
1665 			    NDI_ONLINE_ATTACH | NDI_CONFIG);
1666 		}
1667 		if (dipp != NULL) {
1668 			*dipp = dip;
1669 		} else {
1670 			(void) ndi_rele_devi(dip);
1671 		}
1672 	}
1673 
1674 	return (rv);
1675 }
1676 
1677 /*
1678  * The dipp contains one of following values on return:
1679  * - NULL if no device node found
1680  * - pointer to device node if found
1681  */
1682 int
1683 mach_cpu_get_device_node(struct cpu *cp, dev_info_t **dipp)
1684 {
1685 	*dipp = NULL;
1686 	if (psm_cpu_get_devinfo != NULL) {
1687 		if (psm_cpu_get_devinfo(cp, dipp) == PSM_SUCCESS) {
1688 			return (PSM_SUCCESS);
1689 		}
1690 	}
1691 
1692 	return (PSM_FAILURE);
1693 }
1694 
1695 /*ARGSUSED*/
1696 static int
1697 mach_translate_irq(dev_info_t *dip, int irqno)
1698 {
1699 	return (irqno);	/* default to NO translation */
1700 }
1701 
1702 static void
1703 mach_notify_error(int level, char *errmsg)
1704 {
1705 	/*
1706 	 * SL_FATAL is pass in once panicstr is set, deliver it
1707 	 * as CE_PANIC.  Also, translate SL_ codes back to CE_
1708 	 * codes for the psmi handler
1709 	 */
1710 	if (level & SL_FATAL)
1711 		(*notify_error)(CE_PANIC, errmsg);
1712 	else if (level & SL_WARN)
1713 		(*notify_error)(CE_WARN, errmsg);
1714 	else if (level & SL_NOTE)
1715 		(*notify_error)(CE_NOTE, errmsg);
1716 	else if (level & SL_CONSOLE)
1717 		(*notify_error)(CE_CONT, errmsg);
1718 }
1719 
1720 /*
1721  * It provides the default basic intr_ops interface for the new DDI
1722  * interrupt framework if the PSM doesn't have one.
1723  *
1724  * Input:
1725  * dip     - pointer to the dev_info structure of the requested device
1726  * hdlp    - pointer to the internal interrupt handle structure for the
1727  *	     requested interrupt
1728  * intr_op - opcode for this call
1729  * result  - pointer to the integer that will hold the result to be
1730  *	     passed back if return value is PSM_SUCCESS
1731  *
1732  * Output:
1733  * return value is either PSM_SUCCESS or PSM_FAILURE
1734  */
1735 static int
1736 mach_intr_ops(dev_info_t *dip, ddi_intr_handle_impl_t *hdlp,
1737     psm_intr_op_t intr_op, int *result)
1738 {
1739 	struct intrspec *ispec;
1740 
1741 	switch (intr_op) {
1742 	case PSM_INTR_OP_CHECK_MSI:
1743 		*result = hdlp->ih_type & ~(DDI_INTR_TYPE_MSI |
1744 		    DDI_INTR_TYPE_MSIX);
1745 		break;
1746 	case PSM_INTR_OP_ALLOC_VECTORS:
1747 		if (hdlp->ih_type == DDI_INTR_TYPE_FIXED)
1748 			*result = 1;
1749 		else
1750 			*result = 0;
1751 		break;
1752 	case PSM_INTR_OP_FREE_VECTORS:
1753 		break;
1754 	case PSM_INTR_OP_NAVAIL_VECTORS:
1755 		if (hdlp->ih_type == DDI_INTR_TYPE_FIXED)
1756 			*result = 1;
1757 		else
1758 			*result = 0;
1759 		break;
1760 	case PSM_INTR_OP_XLATE_VECTOR:
1761 		ispec = ((ihdl_plat_t *)hdlp->ih_private)->ip_ispecp;
1762 		*result = psm_translate_irq(dip, ispec->intrspec_vec);
1763 		break;
1764 	case PSM_INTR_OP_GET_CAP:
1765 		*result = 0;
1766 		break;
1767 	case PSM_INTR_OP_GET_PENDING:
1768 	case PSM_INTR_OP_CLEAR_MASK:
1769 	case PSM_INTR_OP_SET_MASK:
1770 	case PSM_INTR_OP_GET_SHARED:
1771 	case PSM_INTR_OP_SET_PRI:
1772 	case PSM_INTR_OP_SET_CAP:
1773 	case PSM_INTR_OP_SET_CPU:
1774 	case PSM_INTR_OP_GET_INTR:
1775 	default:
1776 		return (PSM_FAILURE);
1777 	}
1778 	return (PSM_SUCCESS);
1779 }
1780 /*
1781  * Return 1 if CMT load balancing policies should be
1782  * implemented across instances of the specified hardware
1783  * sharing relationship.
1784  */
1785 int
1786 pg_cmt_load_bal_hw(pghw_type_t hw)
1787 {
1788 	if (hw == PGHW_IPIPE ||
1789 	    hw == PGHW_FPU ||
1790 	    hw == PGHW_PROCNODE ||
1791 	    hw == PGHW_CHIP)
1792 		return (1);
1793 	else
1794 		return (0);
1795 }
1796 /*
1797  * Return 1 if thread affinity polices should be implemented
1798  * for instances of the specifed hardware sharing relationship.
1799  */
1800 int
1801 pg_cmt_affinity_hw(pghw_type_t hw)
1802 {
1803 	if (hw == PGHW_CACHE)
1804 		return (1);
1805 	else
1806 		return (0);
1807 }
1808 
1809 /*
1810  * Return number of counter events requested to measure hardware capacity and
1811  * utilization and setup CPC requests for specified CPU as needed
1812  *
1813  * May return 0 when platform or processor specific code knows that no CPC
1814  * events should be programmed on this CPU or -1 when platform or processor
1815  * specific code doesn't know which counter events are best to use and common
1816  * code should decide for itself
1817  */
1818 int
1819 /* LINTED E_FUNC_ARG_UNUSED */
1820 cu_plat_cpc_init(cpu_t *cp, kcpc_request_list_t *reqs, int nreqs)
1821 {
1822 	const char	*impl_name;
1823 
1824 	/*
1825 	 * Return error if pcbe_ops not set
1826 	 */
1827 	if (pcbe_ops == NULL)
1828 		return (-1);
1829 
1830 	/*
1831 	 * Return that no CPC events should be programmed on hyperthreaded
1832 	 * Pentium 4 and return error for all other x86 processors to tell
1833 	 * common code to decide what counter events to program on those CPUs
1834 	 * for measuring hardware capacity and utilization
1835 	 */
1836 	impl_name = pcbe_ops->pcbe_impl_name();
1837 	if (impl_name != NULL && strcmp(impl_name, PCBE_IMPL_NAME_P4HT) == 0)
1838 		return (0);
1839 	else
1840 		return (-1);
1841 }
1842