xref: /xnu-8020.140.41/osfmk/arm64/cpu.c (revision 27b03b360a988dfd3dfdf34262bb0042026747cc)
1 /*
2  * Copyright (c) 2007-2021 Apple Inc. All rights reserved.
3  *
4  * @APPLE_OSREFERENCE_LICENSE_HEADER_START@
5  *
6  * This file contains Original Code and/or Modifications of Original Code
7  * as defined in and that are subject to the Apple Public Source License
8  * Version 2.0 (the 'License'). You may not use this file except in
9  * compliance with the License. The rights granted to you under the License
10  * may not be used to create, or enable the creation or redistribution of,
11  * unlawful or unlicensed copies of an Apple operating system, or to
12  * circumvent, violate, or enable the circumvention or violation of, any
13  * terms of an Apple operating system software license agreement.
14  *
15  * Please obtain a copy of the License at
16  * http://www.opensource.apple.com/apsl/ and read it before using this file.
17  *
18  * The Original Code and all software distributed under the License are
19  * distributed on an 'AS IS' basis, WITHOUT WARRANTY OF ANY KIND, EITHER
20  * EXPRESS OR IMPLIED, AND APPLE HEREBY DISCLAIMS ALL SUCH WARRANTIES,
21  * INCLUDING WITHOUT LIMITATION, ANY WARRANTIES OF MERCHANTABILITY,
22  * FITNESS FOR A PARTICULAR PURPOSE, QUIET ENJOYMENT OR NON-INFRINGEMENT.
23  * Please see the License for the specific language governing rights and
24  * limitations under the License.
25  *
26  * @APPLE_OSREFERENCE_LICENSE_HEADER_END@
27  */
28 /*
29  *	File:	arm64/cpu.c
30  *
31  *	cpu specific routines
32  */
33 
34 #include <pexpert/arm64/board_config.h>
35 #include <kern/kalloc.h>
36 #include <kern/machine.h>
37 #include <kern/cpu_number.h>
38 #include <kern/percpu.h>
39 #include <kern/thread.h>
40 #include <kern/timer_queue.h>
41 #include <arm/cpu_data.h>
42 #include <arm/cpuid.h>
43 #include <arm/caches_internal.h>
44 #include <arm/cpu_data_internal.h>
45 #include <arm/cpu_internal.h>
46 #include <arm/misc_protos.h>
47 #include <arm/machine_cpu.h>
48 #include <arm/rtclock.h>
49 #include <arm64/proc_reg.h>
50 #include <mach/processor_info.h>
51 #include <vm/pmap.h>
52 #include <vm/vm_kern.h>
53 #include <vm/vm_map.h>
54 #include <pexpert/arm/protos.h>
55 #include <pexpert/device_tree.h>
56 #include <sys/kdebug.h>
57 #include <arm/machine_routines.h>
58 
59 #include <machine/atomic.h>
60 
61 #include <san/kasan.h>
62 
63 #if KPC
64 #include <kern/kpc.h>
65 #endif
66 
67 #if MONOTONIC
68 #include <kern/monotonic.h>
69 #endif /* MONOTONIC */
70 
71 #if KPERF
72 #include <kperf/kptimer.h>
73 #endif /* KPERF */
74 
75 #if HIBERNATION
76 #include <IOKit/IOPlatformExpert.h>
77 #include <IOKit/IOHibernatePrivate.h>
78 #endif /* HIBERNATION */
79 
80 
81 #include <libkern/section_keywords.h>
82 
83 extern boolean_t        idle_enable;
84 extern uint64_t         wake_abstime;
85 
86 #if WITH_CLASSIC_S2R
87 void sleep_token_buffer_init(void);
88 #endif
89 
90 
91 extern uintptr_t resume_idle_cpu;
92 extern uintptr_t start_cpu;
93 
94 #if __ARM_KERNEL_PROTECT__
95 extern void exc_vectors_table;
96 #endif /* __ARM_KERNEL_PROTECT__ */
97 
98 extern void __attribute__((noreturn)) arm64_prepare_for_sleep(boolean_t deep_sleep, unsigned int cpu, uint64_t entry_pa);
99 extern void arm64_force_wfi_clock_gate(void);
100 #if defined(APPLETYPHOON)
101 // <rdar://problem/15827409>
102 extern void typhoon_prepare_for_wfi(void);
103 extern void typhoon_return_from_wfi(void);
104 #endif
105 
106 #if HAS_RETENTION_STATE
107 extern void arm64_retention_wfi(void);
108 #endif
109 
110 vm_address_t   start_cpu_paddr;
111 
112 sysreg_restore_t sysreg_restore __attribute__((section("__DATA, __const"))) = {
113 	.tcr_el1 = TCR_EL1_BOOT,
114 };
115 
116 // wfi - wfi mode
117 //  0 : disabled
118 //  1 : normal
119 //  2 : overhead simulation (delay & flags)
120 TUNABLE(unsigned int, wfi, "wfi", 1);
121 #if DEVELOPMENT || DEBUG
122 
123 // wfi_flags
124 //  1 << 0 : flush L1s
125 //  1 << 1 : flush TLBs
126 static int wfi_flags = 0;
127 
128 // wfi_delay - delay ticks after wfi exit
129 static uint64_t wfi_delay = 0;
130 
131 #endif /* DEVELOPMENT || DEBUG */
132 
133 #define CPUPM_IDLE_WFE 0x5310300
134 #define CPUPM_IDLE_TIMER_WFE 0x5310304
135 
136 /* When recommended, issue WFE with [FI]IRQ unmasked in the idle
137  * loop. The default.
138  */
139 uint32_t idle_proximate_io_wfe_unmasked = 1;
140 #if DEVELOPMENT || DEBUG
141 uint32_t idle_proximate_timer_wfe = 1;
142 uint32_t idle_proximate_io_wfe_masked = 0;
143 #else
144 /* Issue WFE in lieu of WFI when awaiting a proximate timer. */
145 static uint32_t idle_proximate_timer_wfe = 1;
146 /* When recommended, issue WFE with [FI]IRQ masked in the idle loop.
147  * Non-default, retained for experimentation.
148  */
149 static uint32_t idle_proximate_io_wfe_masked = 0;
150 #endif
151 
152 #if __ARM_GLOBAL_SLEEP_BIT__
153 volatile boolean_t arm64_stall_sleep = TRUE;
154 #endif
155 
156 #if WITH_CLASSIC_S2R
157 /*
158  * These must be aligned to avoid issues with calling bcopy_phys on them before
159  * we are done with pmap initialization.
160  */
161 static const uint8_t __attribute__ ((aligned(8))) suspend_signature[] = {'X', 'S', 'O', 'M', 'P', 'S', 'U', 'S'};
162 static const uint8_t __attribute__ ((aligned(8))) running_signature[] = {'X', 'S', 'O', 'M', 'N', 'N', 'U', 'R'};
163 #endif
164 
165 #if WITH_CLASSIC_S2R
166 static vm_offset_t sleepTokenBuffer = (vm_offset_t)NULL;
167 #endif
168 static boolean_t coresight_debug_enabled = FALSE;
169 
170 #if defined(CONFIG_XNUPOST)
171 void arm64_ipi_test_callback(void *);
172 void arm64_immediate_ipi_test_callback(void *);
173 
174 void
arm64_ipi_test_callback(void * parm)175 arm64_ipi_test_callback(void *parm)
176 {
177 	volatile uint64_t *ipi_test_data = parm;
178 	cpu_data_t *cpu_data;
179 
180 	cpu_data = getCpuDatap();
181 
182 	*ipi_test_data = cpu_data->cpu_number;
183 }
184 
185 void
arm64_immediate_ipi_test_callback(void * parm)186 arm64_immediate_ipi_test_callback(void *parm)
187 {
188 	volatile uint64_t *ipi_test_data = parm;
189 	cpu_data_t *cpu_data;
190 
191 	cpu_data = getCpuDatap();
192 
193 	*ipi_test_data = cpu_data->cpu_number + MAX_CPUS;
194 }
195 
196 uint64_t arm64_ipi_test_data[MAX_CPUS * 2];
197 
198 MACHINE_TIMEOUT(arm64_ipi_test_timeout, "arm64-ipi-test", 100, MACHINE_TIMEOUT_UNIT_MSEC, NULL);
199 
200 void
arm64_ipi_test()201 arm64_ipi_test()
202 {
203 	volatile uint64_t *ipi_test_data, *immediate_ipi_test_data;
204 	uint64_t timeout_ms = os_atomic_load(&arm64_ipi_test_timeout, relaxed);
205 	uint64_t then, now, delta;
206 	int current_cpu_number = getCpuDatap()->cpu_number;
207 
208 	/*
209 	 * probably the only way to have this on most systems is with the
210 	 * cpus=1 boot-arg, but nonetheless, if we only have 1 CPU active,
211 	 * IPI is not available
212 	 */
213 	if (real_ncpus == 1) {
214 		return;
215 	}
216 
217 	const unsigned int max_cpu_id = ml_get_max_cpu_number();
218 	for (unsigned int i = 0; i <= max_cpu_id; ++i) {
219 		ipi_test_data = &arm64_ipi_test_data[i];
220 		immediate_ipi_test_data = &arm64_ipi_test_data[i + MAX_CPUS];
221 		*ipi_test_data = ~i;
222 		kern_return_t error = cpu_xcall((int)i, (void *)arm64_ipi_test_callback, (void *)(uintptr_t)ipi_test_data);
223 		if (error != KERN_SUCCESS) {
224 			panic("CPU %d was unable to IPI CPU %u: error %d", current_cpu_number, i, error);
225 		}
226 
227 		while ((error = cpu_immediate_xcall((int)i, (void *)arm64_immediate_ipi_test_callback,
228 		    (void *)(uintptr_t)immediate_ipi_test_data)) == KERN_ALREADY_WAITING) {
229 			now = mach_absolute_time();
230 			absolutetime_to_nanoseconds(now - then, &delta);
231 			if ((delta / NSEC_PER_MSEC) > timeout_ms) {
232 				panic("CPU %d was unable to immediate-IPI CPU %u within %lldms", current_cpu_number, i, timeout_ms);
233 			}
234 		}
235 
236 		if (error != KERN_SUCCESS) {
237 			panic("CPU %d was unable to immediate-IPI CPU %u: error %d", current_cpu_number, i, error);
238 		}
239 
240 		then = mach_absolute_time();
241 
242 		while ((*ipi_test_data != i) || (*immediate_ipi_test_data != (i + MAX_CPUS))) {
243 			now = mach_absolute_time();
244 			absolutetime_to_nanoseconds(now - then, &delta);
245 			if ((delta / NSEC_PER_MSEC) > timeout_ms) {
246 				panic("CPU %d tried to IPI CPU %d but didn't get correct responses within %lldms, responses: %llx, %llx",
247 				    current_cpu_number, i, timeout_ms, *ipi_test_data, *immediate_ipi_test_data);
248 			}
249 		}
250 	}
251 }
252 #endif /* defined(CONFIG_XNUPOST) */
253 
254 static void
configure_coresight_registers(cpu_data_t * cdp)255 configure_coresight_registers(cpu_data_t *cdp)
256 {
257 	int             i;
258 
259 	assert(cdp);
260 	vm_offset_t     coresight_regs = ml_get_topology_info()->cpus[cdp->cpu_number].coresight_regs;
261 
262 	/*
263 	 * ARMv8 coresight registers are optional. If the device tree did not
264 	 * provide either cpu_regmap_paddr (from the legacy "reg-private" EDT property)
265 	 * or coresight_regs (from the new "coresight-reg" property), assume that
266 	 * coresight registers are not supported.
267 	 */
268 	if (cdp->cpu_regmap_paddr || coresight_regs) {
269 		for (i = 0; i < CORESIGHT_REGIONS; ++i) {
270 			if (i == CORESIGHT_CTI) {
271 				continue;
272 			}
273 			/* Skip debug-only registers on production chips */
274 			if (((i == CORESIGHT_ED) || (i == CORESIGHT_UTT)) && !coresight_debug_enabled) {
275 				continue;
276 			}
277 
278 			if (!cdp->coresight_base[i]) {
279 				if (coresight_regs) {
280 					cdp->coresight_base[i] = coresight_regs + CORESIGHT_OFFSET(i);
281 				} else {
282 					uint64_t addr = cdp->cpu_regmap_paddr + CORESIGHT_OFFSET(i);
283 					cdp->coresight_base[i] = (vm_offset_t)ml_io_map(addr, CORESIGHT_SIZE);
284 				}
285 
286 				/*
287 				 * At this point, failing to io map the
288 				 * registers is considered as an error.
289 				 */
290 				if (!cdp->coresight_base[i]) {
291 					panic("unable to ml_io_map coresight regions");
292 				}
293 			}
294 			/* Unlock EDLAR, CTILAR, PMLAR */
295 			if (i != CORESIGHT_UTT) {
296 				*(volatile uint32_t *)(cdp->coresight_base[i] + ARM_DEBUG_OFFSET_DBGLAR) = ARM_DBG_LOCK_ACCESS_KEY;
297 			}
298 		}
299 	}
300 }
301 
302 
303 /*
304  *	Routine:	cpu_bootstrap
305  *	Function:
306  */
307 void
cpu_bootstrap(void)308 cpu_bootstrap(void)
309 {
310 }
311 
312 /*
313  *	Routine:	cpu_sleep
314  *	Function:
315  */
316 void
cpu_sleep(void)317 cpu_sleep(void)
318 {
319 	cpu_data_t     *cpu_data_ptr = getCpuDatap();
320 
321 	cpu_data_ptr->cpu_active_thread = current_thread();
322 	cpu_data_ptr->cpu_reset_handler = (uintptr_t) start_cpu_paddr;
323 	cpu_data_ptr->cpu_flags |= SleepState;
324 
325 	if (cpu_data_ptr->cpu_user_debug != NULL) {
326 		arm_debug_set(NULL);
327 	}
328 
329 #if KPC
330 	kpc_idle();
331 #endif /* KPC */
332 #if MONOTONIC
333 	mt_cpu_down(cpu_data_ptr);
334 #endif /* MONOTONIC */
335 #if KPERF
336 	kptimer_stop_curcpu();
337 #endif /* KPERF */
338 
339 	CleanPoC_Dcache();
340 
341 #if USE_APPLEARMSMP
342 	if (ml_is_quiescing()) {
343 		PE_cpu_machine_quiesce(cpu_data_ptr->cpu_id);
344 	} else {
345 		bool deep_sleep = PE_cpu_down(cpu_data_ptr->cpu_id);
346 		cpu_data_ptr->cpu_sleep_token = ARM_CPU_ON_SLEEP_PATH;
347 		// hang CPU on spurious wakeup
348 		cpu_data_ptr->cpu_reset_handler = (uintptr_t)0;
349 		__builtin_arm_dsb(DSB_ISH);
350 		CleanPoU_Dcache();
351 		arm64_prepare_for_sleep(deep_sleep, cpu_data_ptr->cpu_number, ml_vtophys((vm_offset_t)&LowResetVectorBase));
352 	}
353 #else
354 	PE_cpu_machine_quiesce(cpu_data_ptr->cpu_id);
355 #endif
356 	/*NOTREACHED*/
357 }
358 
359 /*
360  *	Routine:	cpu_interrupt_is_pending
361  *	Function:	Returns a bool signifying a non-zero ISR_EL1,
362  *			indicating a pending IRQ, FIQ or external abort.
363  */
364 
365 bool
cpu_interrupt_is_pending(void)366 cpu_interrupt_is_pending(void)
367 {
368 	uint64_t isr_value;
369 	isr_value = __builtin_arm_rsr64("ISR_EL1");
370 	return isr_value != 0;
371 }
372 
373 static bool
cpu_proximate_timer(void)374 cpu_proximate_timer(void)
375 {
376 	return !SetIdlePop();
377 }
378 
379 #ifdef ARM64_BOARD_CONFIG_T6000
380 int wfe_allowed = 0;
381 #else
382 int wfe_allowed = 1;
383 #endif /* ARM64_BOARD_CONFIG_T6000 */
384 
385 #if DEVELOPMENT || DEBUG
386 #define WFE_STAT(x)     \
387 	do {            \
388 	        (x);    \
389 	} while(0)
390 #else
391 #define WFE_STAT(x)     do {} while(0)
392 #endif /* DEVELOPMENT || DEBUG */
393 
394 bool
wfe_to_deadline_or_interrupt(uint32_t cid,uint64_t wfe_deadline,cpu_data_t * cdp,bool unmask)395 wfe_to_deadline_or_interrupt(uint32_t cid, uint64_t wfe_deadline, cpu_data_t *cdp, bool unmask)
396 {
397 	bool ipending = false;
398 	uint64_t irqc = 0, nirqc = 0;
399 
400 	/* The ARMv8 architecture permits a processor dwelling in WFE
401 	 * with F/IRQ masked to ignore a pending interrupt, i.e.
402 	 * not classify it as an 'event'. This is potentially
403 	 * problematic with AICv2's IRQ distribution model, as
404 	 * a transient interrupt masked interval can cause an SIQ
405 	 * query rejection, possibly routing the interrupt to
406 	 * another core/cluster in a powergated state.
407 	 * Hence, optionally unmask IRQs+FIQs across WFE.
408 	 */
409 	if (unmask) {
410 		/* Latch SW IRQ+FIQ counter prior to unmasking
411 		 * interrupts.
412 		 */
413 		irqc = nirqc = os_atomic_load(&cdp->cpu_stat.irq_ex_cnt_wake, relaxed);
414 		/* Unmask IRQ+FIQ. Mirrors mask used by machine_idle()
415 		 * with ASYNCF omission. Consider that this could
416 		 * delay recognition of an async abort, including
417 		 * those triggered by ISRs
418 		 */
419 		__builtin_arm_wsr("DAIFClr", (DAIFSC_IRQF | DAIFSC_FIQF));
420 	}
421 
422 	while ((ipending = (cpu_interrupt_is_pending())) == false) {
423 		if (unmask) {
424 			/* If WFE was issued with IRQs unmasked, an
425 			 * interrupt may have been processed.
426 			 * Consult the SW IRQ counter to determine
427 			 * whether the 'idle loop' must be
428 			 * re-evaluated.
429 			 */
430 			nirqc = os_atomic_load(&cdp->cpu_stat.irq_ex_cnt_wake, relaxed);
431 			if (nirqc != irqc) {
432 				break;
433 			}
434 		}
435 
436 		if (__probable(wfe_allowed)) {
437 			/*
438 			 * If IRQs are unmasked, there's a small window
439 			 * where an 'extra' WFE may be issued after
440 			 * the consultation of the SW interrupt counter
441 			 * and new interrupt arrival. Hence this WFE
442 			 * relies on the [FI]RQ interrupt handler
443 			 * epilogue issuing a 'SEVL', to post an
444 			 * event which causes the next WFE on the same
445 			 * PE to retire immediately.
446 			 */
447 
448 			__builtin_arm_wfe();
449 		}
450 
451 		WFE_STAT(cdp->wfe_count++);
452 		if (wfe_deadline != ~0ULL) {
453 			WFE_STAT(cdp->wfe_deadline_checks++);
454 			/* Check if the WFE recommendation has expired.
455 			 * We do not recompute the deadline here.
456 			 */
457 			if ((ml_cluster_wfe_timeout(cid) == 0) ||
458 			    mach_absolute_time() >= wfe_deadline) {
459 				WFE_STAT(cdp->wfe_terminations++);
460 				break;
461 			}
462 		}
463 	}
464 
465 	if (unmask) {
466 		/* Re-mask IRQ+FIQ
467 		 * Mirrors mask used by machine_idle(), with ASYNCF
468 		 * omission
469 		 */
470 		__builtin_arm_wsr64("DAIFSet", (DAIFSC_IRQF | DAIFSC_FIQF));
471 		/* Refetch SW interrupt counter with IRQs masked
472 		 * It is important that this routine accurately flags
473 		 * any observed interrupts via its return value,
474 		 * inaccuracy may lead to an erroneous WFI fallback.
475 		 */
476 		nirqc = os_atomic_load(&cdp->cpu_stat.irq_ex_cnt_wake, relaxed);
477 	}
478 
479 	return ipending || (nirqc != irqc);
480 }
481 
482 /*
483  *	Routine:	cpu_idle
484  *	Function:
485  */
486 void __attribute__((noreturn))
cpu_idle(void)487 cpu_idle(void)
488 {
489 	cpu_data_t     *cpu_data_ptr = getCpuDatap();
490 	uint64_t        new_idle_timeout_ticks = 0x0ULL, lastPop;
491 	bool idle_disallowed = false;
492 
493 	if (__improbable((!idle_enable))) {
494 		idle_disallowed = true;
495 	} else if (__improbable(cpu_data_ptr->cpu_signal & SIGPdisabled)) {
496 		idle_disallowed = true;
497 	}
498 
499 	if (__improbable(idle_disallowed)) {
500 		Idle_load_context();
501 	}
502 
503 	bool ipending = false;
504 	uint32_t cid = cpu_data_ptr->cpu_cluster_id;
505 
506 	if (idle_proximate_io_wfe_masked == 1) {
507 		uint64_t wfe_deadline = 0;
508 		/* Check for an active perf. controller generated
509 		 * WFE recommendation for this cluster.
510 		 */
511 		uint64_t wfe_ttd = 0;
512 		if ((wfe_ttd = ml_cluster_wfe_timeout(cid)) != 0) {
513 			wfe_deadline = mach_absolute_time() + wfe_ttd;
514 		}
515 
516 		if (wfe_deadline != 0) {
517 			/* Poll issuing event-bounded WFEs until an interrupt
518 			 * arrives or the WFE recommendation expires
519 			 */
520 #if DEVELOPMENT || DEBUG
521 			KDBG(CPUPM_IDLE_WFE | DBG_FUNC_START, ipending, cpu_data_ptr->wfe_count, wfe_ttd, cid);
522 #endif
523 			ipending = wfe_to_deadline_or_interrupt(cid, wfe_deadline, cpu_data_ptr, false);
524 #if DEVELOPMENT || DEBUG
525 			KDBG(CPUPM_IDLE_WFE | DBG_FUNC_END, ipending, cpu_data_ptr->wfe_count, wfe_deadline, 0);
526 #endif
527 			if (ipending == true) {
528 				/* Back to machine_idle() */
529 				Idle_load_context();
530 			}
531 		}
532 	}
533 
534 	if (__improbable(cpu_proximate_timer())) {
535 		if (idle_proximate_timer_wfe == 1) {
536 			/* Poll issuing WFEs until the expected
537 			 * timer FIQ arrives.
538 			 */
539 #if DEVELOPMENT || DEBUG
540 			KDBG(CPUPM_IDLE_TIMER_WFE | DBG_FUNC_START, ipending, cpu_data_ptr->wfe_count, ~0ULL, 0);
541 #endif
542 			ipending = wfe_to_deadline_or_interrupt(cid, ~0ULL, cpu_data_ptr, false);
543 #if DEVELOPMENT || DEBUG
544 			KDBG(CPUPM_IDLE_TIMER_WFE | DBG_FUNC_END, ipending, cpu_data_ptr->wfe_count, ~0ULL, 0);
545 #endif
546 			assert(ipending == true);
547 		}
548 		Idle_load_context();
549 	}
550 
551 	lastPop = cpu_data_ptr->rtcPop;
552 
553 	cpu_data_ptr->cpu_active_thread = current_thread();
554 
555 	if (wfi && (cpu_data_ptr->cpu_idle_notify != NULL)) {
556 		cpu_data_ptr->cpu_idle_notify(cpu_data_ptr->cpu_id, TRUE, &new_idle_timeout_ticks);
557 	}
558 
559 	if (cpu_data_ptr->idle_timer_notify != NULL) {
560 		if (new_idle_timeout_ticks == 0x0ULL) {
561 			/* turn off the idle timer */
562 			cpu_data_ptr->idle_timer_deadline = 0x0ULL;
563 		} else {
564 			/* set the new idle timeout */
565 			clock_absolutetime_interval_to_deadline(new_idle_timeout_ticks, &cpu_data_ptr->idle_timer_deadline);
566 		}
567 		timer_resync_deadlines();
568 		if (cpu_data_ptr->rtcPop != lastPop) {
569 			SetIdlePop();
570 		}
571 	}
572 
573 #if KPC
574 	kpc_idle();
575 #endif
576 #if MONOTONIC
577 	mt_cpu_idle(cpu_data_ptr);
578 #endif /* MONOTONIC */
579 
580 	if (wfi) {
581 #if !defined(APPLE_ARM64_ARCH_FAMILY)
582 		platform_cache_idle_enter();
583 #endif
584 
585 #if DEVELOPMENT || DEBUG
586 		// When simulating wfi overhead,
587 		// force wfi to clock gating only
588 		if (wfi == 2) {
589 			arm64_force_wfi_clock_gate();
590 		}
591 #endif /* DEVELOPMENT || DEBUG */
592 
593 #if defined(APPLETYPHOON)
594 		// <rdar://problem/15827409> CPU1 Stuck in WFIWT Because of MMU Prefetch
595 		typhoon_prepare_for_wfi();
596 #endif
597 		__builtin_arm_dsb(DSB_SY);
598 #if HAS_RETENTION_STATE
599 		arm64_retention_wfi();
600 #else
601 		__builtin_arm_wfi();
602 #endif
603 
604 #if defined(APPLETYPHOON)
605 		// <rdar://problem/15827409> CPU1 Stuck in WFIWT Because of MMU Prefetch
606 		typhoon_return_from_wfi();
607 #endif
608 
609 #if DEVELOPMENT || DEBUG
610 		// Handle wfi overhead simulation
611 		if (wfi == 2) {
612 			uint64_t deadline;
613 
614 			// Calculate wfi delay deadline
615 			clock_absolutetime_interval_to_deadline(wfi_delay, &deadline);
616 
617 			// Flush L1 caches
618 			if ((wfi_flags & 1) != 0) {
619 				InvalidatePoU_Icache();
620 				FlushPoC_Dcache();
621 			}
622 
623 			// Flush TLBs
624 			if ((wfi_flags & 2) != 0) {
625 				flush_core_tlb();
626 			}
627 
628 			// Wait for the ballance of the wfi delay
629 			clock_delay_until(deadline);
630 		}
631 #endif /* DEVELOPMENT || DEBUG */
632 #if !defined(APPLE_ARM64_ARCH_FAMILY)
633 		platform_cache_idle_exit();
634 #endif
635 	}
636 
637 	ClearIdlePop(TRUE);
638 
639 	cpu_idle_exit(FALSE);
640 }
641 
642 /*
643  *	Routine:	cpu_idle_exit
644  *	Function:
645  */
646 void
cpu_idle_exit(boolean_t from_reset)647 cpu_idle_exit(boolean_t from_reset)
648 {
649 	uint64_t        new_idle_timeout_ticks = 0x0ULL;
650 	cpu_data_t     *cpu_data_ptr = getCpuDatap();
651 
652 	assert(exception_stack_pointer() != 0);
653 
654 	/* Back from WFI, unlock OSLAR and EDLAR. */
655 	if (from_reset) {
656 		configure_coresight_registers(cpu_data_ptr);
657 	}
658 
659 #if KPC
660 	kpc_idle_exit();
661 #endif
662 
663 #if MONOTONIC
664 	mt_cpu_run(cpu_data_ptr);
665 #endif /* MONOTONIC */
666 
667 	if (wfi && (cpu_data_ptr->cpu_idle_notify != NULL)) {
668 		cpu_data_ptr->cpu_idle_notify(cpu_data_ptr->cpu_id, FALSE, &new_idle_timeout_ticks);
669 	}
670 
671 	if (cpu_data_ptr->idle_timer_notify != NULL) {
672 		if (new_idle_timeout_ticks == 0x0ULL) {
673 			/* turn off the idle timer */
674 			cpu_data_ptr->idle_timer_deadline = 0x0ULL;
675 		} else {
676 			/* set the new idle timeout */
677 			clock_absolutetime_interval_to_deadline(new_idle_timeout_ticks, &cpu_data_ptr->idle_timer_deadline);
678 		}
679 		timer_resync_deadlines();
680 	}
681 
682 	Idle_load_context();
683 }
684 
685 void
cpu_init(void)686 cpu_init(void)
687 {
688 	cpu_data_t     *cdp = getCpuDatap();
689 	arm_cpu_info_t *cpu_info_p;
690 
691 	assert(exception_stack_pointer() != 0);
692 
693 	if (cdp->cpu_type != CPU_TYPE_ARM64) {
694 		cdp->cpu_type = CPU_TYPE_ARM64;
695 
696 		timer_call_queue_init(&cdp->rtclock_timer.queue);
697 		cdp->rtclock_timer.deadline = EndOfAllTime;
698 
699 		if (cdp == &BootCpuData) {
700 			do_cpuid();
701 			do_mvfpid();
702 		} else {
703 			/*
704 			 * We initialize non-boot CPUs here; the boot CPU is
705 			 * dealt with as part of pmap_bootstrap.
706 			 */
707 			pmap_cpu_data_init();
708 		}
709 
710 		do_cacheid();
711 
712 		/* ARM_SMP: Assuming identical cpu */
713 		do_debugid();
714 
715 		cpu_info_p = cpuid_info();
716 
717 		/* switch based on CPU's reported architecture */
718 		switch (cpu_info_p->arm_info.arm_arch) {
719 		case CPU_ARCH_ARMv8:
720 			cdp->cpu_subtype = CPU_SUBTYPE_ARM64_V8;
721 			break;
722 		case CPU_ARCH_ARMv8E:
723 			cdp->cpu_subtype = CPU_SUBTYPE_ARM64E;
724 			break;
725 		default:
726 			//cdp->cpu_subtype = CPU_SUBTYPE_ARM64_ALL;
727 			/* this panic doesn't work this early in startup */
728 			panic("Unknown CPU subtype...");
729 			break;
730 		}
731 
732 		cdp->cpu_threadtype = CPU_THREADTYPE_NONE;
733 	}
734 	cdp->cpu_stat.irq_ex_cnt_wake = 0;
735 	cdp->cpu_stat.ipi_cnt_wake = 0;
736 #if MONOTONIC
737 	cdp->cpu_stat.pmi_cnt_wake = 0;
738 #endif /* MONOTONIC */
739 	cdp->cpu_running = TRUE;
740 	cdp->cpu_sleep_token_last = cdp->cpu_sleep_token;
741 	cdp->cpu_sleep_token = 0x0UL;
742 #if KPC
743 	kpc_idle_exit();
744 #endif /* KPC */
745 #if MONOTONIC
746 	mt_cpu_up(cdp);
747 #endif /* MONOTONIC */
748 }
749 
750 void
cpu_stack_alloc(cpu_data_t * cpu_data_ptr)751 cpu_stack_alloc(cpu_data_t *cpu_data_ptr)
752 {
753 	vm_offset_t irq_stack = 0;
754 	vm_offset_t exc_stack = 0;
755 
756 	kmem_alloc(kernel_map, &irq_stack,
757 	    INTSTACK_SIZE + ptoa(2), KMA_NOFAIL | KMA_PERMANENT | KMA_ZERO |
758 	    KMA_GUARD_FIRST | KMA_GUARD_LAST | KMA_KSTACK | KMA_KOBJECT,
759 	    VM_KERN_MEMORY_STACK);
760 
761 	cpu_data_ptr->intstack_top = irq_stack + PAGE_SIZE + INTSTACK_SIZE;
762 	cpu_data_ptr->istackptr = cpu_data_ptr->intstack_top;
763 
764 	kmem_alloc(kernel_map, &exc_stack,
765 	    EXCEPSTACK_SIZE + ptoa(2), KMA_NOFAIL | KMA_PERMANENT | KMA_ZERO |
766 	    KMA_GUARD_FIRST | KMA_GUARD_LAST | KMA_KSTACK | KMA_KOBJECT,
767 	    VM_KERN_MEMORY_STACK);
768 
769 	cpu_data_ptr->excepstack_top = exc_stack + PAGE_SIZE + EXCEPSTACK_SIZE;
770 	cpu_data_ptr->excepstackptr = cpu_data_ptr->excepstack_top;
771 }
772 
773 void
cpu_data_free(cpu_data_t * cpu_data_ptr)774 cpu_data_free(cpu_data_t *cpu_data_ptr)
775 {
776 	if ((cpu_data_ptr == NULL) || (cpu_data_ptr == &BootCpuData)) {
777 		return;
778 	}
779 
780 	int cpu_number = cpu_data_ptr->cpu_number;
781 
782 	if (CpuDataEntries[cpu_number].cpu_data_vaddr == cpu_data_ptr) {
783 		CpuDataEntries[cpu_number].cpu_data_vaddr = NULL;
784 		CpuDataEntries[cpu_number].cpu_data_paddr = 0;
785 		__builtin_arm_dmb(DMB_ISH); // Ensure prior stores to cpu array are visible
786 	}
787 	kmem_free(kernel_map,
788 	    cpu_data_ptr->intstack_top - INTSTACK_SIZE - PAGE_SIZE,
789 	    INTSTACK_SIZE + 2 * PAGE_SIZE);
790 	kmem_free(kernel_map,
791 	    cpu_data_ptr->excepstack_top - EXCEPSTACK_SIZE - PAGE_SIZE,
792 	    EXCEPSTACK_SIZE + 2 * PAGE_SIZE);
793 }
794 
795 void
cpu_data_init(cpu_data_t * cpu_data_ptr)796 cpu_data_init(cpu_data_t *cpu_data_ptr)
797 {
798 	uint32_t i;
799 
800 	cpu_data_ptr->cpu_flags = 0;
801 	cpu_data_ptr->cpu_int_state = 0;
802 	cpu_data_ptr->cpu_pending_ast = AST_NONE;
803 	cpu_data_ptr->cpu_cache_dispatch = NULL;
804 	cpu_data_ptr->rtcPop = EndOfAllTime;
805 	cpu_data_ptr->rtclock_datap = &RTClockData;
806 	cpu_data_ptr->cpu_user_debug = NULL;
807 
808 
809 	cpu_data_ptr->cpu_base_timebase = 0;
810 	cpu_data_ptr->cpu_idle_notify = NULL;
811 	cpu_data_ptr->cpu_idle_latency = 0x0ULL;
812 	cpu_data_ptr->cpu_idle_pop = 0x0ULL;
813 	cpu_data_ptr->cpu_reset_type = 0x0UL;
814 	cpu_data_ptr->cpu_reset_handler = 0x0UL;
815 	cpu_data_ptr->cpu_reset_assist = 0x0UL;
816 	cpu_data_ptr->cpu_regmap_paddr = 0x0ULL;
817 	cpu_data_ptr->cpu_phys_id = 0x0UL;
818 	cpu_data_ptr->cpu_l2_access_penalty = 0;
819 	cpu_data_ptr->cpu_cluster_type = CLUSTER_TYPE_SMP;
820 	cpu_data_ptr->cpu_cluster_id = 0;
821 	cpu_data_ptr->cpu_l2_id = 0;
822 	cpu_data_ptr->cpu_l2_size = 0;
823 	cpu_data_ptr->cpu_l3_id = 0;
824 	cpu_data_ptr->cpu_l3_size = 0;
825 
826 	cpu_data_ptr->cpu_signal = SIGPdisabled;
827 
828 	cpu_data_ptr->cpu_get_fiq_handler = NULL;
829 	cpu_data_ptr->cpu_tbd_hardware_addr = NULL;
830 	cpu_data_ptr->cpu_tbd_hardware_val = NULL;
831 	cpu_data_ptr->cpu_get_decrementer_func = NULL;
832 	cpu_data_ptr->cpu_set_decrementer_func = NULL;
833 	cpu_data_ptr->cpu_sleep_token = ARM_CPU_ON_SLEEP_PATH;
834 	cpu_data_ptr->cpu_sleep_token_last = 0x00000000UL;
835 	cpu_data_ptr->cpu_xcall_p0 = NULL;
836 	cpu_data_ptr->cpu_xcall_p1 = NULL;
837 	cpu_data_ptr->cpu_imm_xcall_p0 = NULL;
838 	cpu_data_ptr->cpu_imm_xcall_p1 = NULL;
839 
840 	for (i = 0; i < CORESIGHT_REGIONS; ++i) {
841 		cpu_data_ptr->coresight_base[i] = 0;
842 	}
843 
844 #if !XNU_MONITOR
845 	pmap_cpu_data_t * pmap_cpu_data_ptr = &cpu_data_ptr->cpu_pmap_cpu_data;
846 
847 	pmap_cpu_data_ptr->cpu_nested_pmap = (struct pmap *) NULL;
848 	pmap_cpu_data_ptr->cpu_number = PMAP_INVALID_CPU_NUM;
849 	pmap_cpu_data_ptr->pv_free.list = NULL;
850 	pmap_cpu_data_ptr->pv_free.count = 0;
851 	pmap_cpu_data_ptr->pv_free_spill_marker = NULL;
852 
853 	bzero(&(pmap_cpu_data_ptr->cpu_sw_asids[0]), sizeof(pmap_cpu_data_ptr->cpu_sw_asids));
854 #endif
855 	cpu_data_ptr->halt_status = CPU_NOT_HALTED;
856 #if __ARM_KERNEL_PROTECT__
857 	cpu_data_ptr->cpu_exc_vectors = (vm_offset_t)&exc_vectors_table;
858 #endif /* __ARM_KERNEL_PROTECT__ */
859 
860 #if defined(HAS_APPLE_PAC)
861 	cpu_data_ptr->rop_key = 0;
862 	cpu_data_ptr->jop_key = ml_default_jop_pid();
863 #endif
864 }
865 
866 kern_return_t
cpu_data_register(cpu_data_t * cpu_data_ptr)867 cpu_data_register(cpu_data_t *cpu_data_ptr)
868 {
869 	int     cpu = cpu_data_ptr->cpu_number;
870 
871 #if KASAN
872 	for (int i = 0; i < CPUWINDOWS_MAX; i++) {
873 		kasan_notify_address_nopoison(pmap_cpu_windows_copy_addr(cpu, i), PAGE_SIZE);
874 	}
875 #endif
876 
877 	__builtin_arm_dmb(DMB_ISH); // Ensure prior stores to cpu data are visible
878 	CpuDataEntries[cpu].cpu_data_vaddr = cpu_data_ptr;
879 	CpuDataEntries[cpu].cpu_data_paddr = (void *)ml_vtophys((vm_offset_t)cpu_data_ptr);
880 	return KERN_SUCCESS;
881 }
882 
883 #if defined(KERNEL_INTEGRITY_CTRR)
884 /* Hibernation needs to reset this state, so data and text are in the hib segment;
885  * this allows them be accessed and executed early.
886  */
887 LCK_GRP_DECLARE(ctrr_cpu_start_lock_grp, "ctrr_cpu_start_lock");
888 LCK_SPIN_DECLARE(ctrr_cpu_start_lck, &ctrr_cpu_start_lock_grp);
889 enum ctrr_cluster_states ctrr_cluster_locked[MAX_CPU_CLUSTERS] MARK_AS_HIBERNATE_DATA;
890 
891 MARK_AS_HIBERNATE_TEXT
892 void
init_ctrr_cluster_states(void)893 init_ctrr_cluster_states(void)
894 {
895 	for (int i = 0; i < MAX_CPU_CLUSTERS; i++) {
896 		ctrr_cluster_locked[i] = CTRR_UNLOCKED;
897 	}
898 }
899 #endif
900 
901 kern_return_t
cpu_start(int cpu)902 cpu_start(int cpu)
903 {
904 	cpu_data_t *cpu_data_ptr = CpuDataEntries[cpu].cpu_data_vaddr;
905 
906 	kprintf("cpu_start() cpu: %d\n", cpu);
907 
908 	if (cpu == cpu_number()) {
909 		cpu_machine_init();
910 		configure_coresight_registers(cpu_data_ptr);
911 	} else {
912 		thread_t first_thread;
913 		processor_t processor;
914 
915 		cpu_data_ptr->cpu_reset_handler = (vm_offset_t) start_cpu_paddr;
916 
917 #if !XNU_MONITOR
918 		cpu_data_ptr->cpu_pmap_cpu_data.cpu_nested_pmap = NULL;
919 #endif
920 
921 		processor = PERCPU_GET_RELATIVE(processor, cpu_data, cpu_data_ptr);
922 		if (processor->startup_thread != THREAD_NULL) {
923 			first_thread = processor->startup_thread;
924 		} else {
925 			first_thread = processor->idle_thread;
926 		}
927 		cpu_data_ptr->cpu_active_thread = first_thread;
928 		first_thread->machine.CpuDatap = cpu_data_ptr;
929 		first_thread->machine.pcpu_data_base =
930 		    (vm_address_t)cpu_data_ptr - __PERCPU_ADDR(cpu_data);
931 
932 		configure_coresight_registers(cpu_data_ptr);
933 
934 		flush_dcache((vm_offset_t)&CpuDataEntries[cpu], sizeof(cpu_data_entry_t), FALSE);
935 		flush_dcache((vm_offset_t)cpu_data_ptr, sizeof(cpu_data_t), FALSE);
936 #if defined(KERNEL_INTEGRITY_CTRR)
937 
938 		/* First CPU being started within a cluster goes ahead to lock CTRR for cluster;
939 		 * other CPUs block until cluster is locked. */
940 		lck_spin_lock(&ctrr_cpu_start_lck);
941 		switch (ctrr_cluster_locked[cpu_data_ptr->cpu_cluster_id]) {
942 		case CTRR_UNLOCKED:
943 			ctrr_cluster_locked[cpu_data_ptr->cpu_cluster_id] = CTRR_LOCKING;
944 			lck_spin_unlock(&ctrr_cpu_start_lck);
945 			break;
946 		case CTRR_LOCKING:
947 			assert_wait(&ctrr_cluster_locked[cpu_data_ptr->cpu_cluster_id], THREAD_UNINT);
948 			lck_spin_unlock(&ctrr_cpu_start_lck);
949 			thread_block(THREAD_CONTINUE_NULL);
950 			assert(ctrr_cluster_locked[cpu_data_ptr->cpu_cluster_id] != CTRR_LOCKING);
951 			break;
952 		default:         // CTRR_LOCKED
953 			lck_spin_unlock(&ctrr_cpu_start_lck);
954 			break;
955 		}
956 #endif
957 		(void) PE_cpu_start(cpu_data_ptr->cpu_id, (vm_offset_t)NULL, (vm_offset_t)NULL);
958 	}
959 
960 	return KERN_SUCCESS;
961 }
962 
963 
964 void
cpu_timebase_init(boolean_t from_boot)965 cpu_timebase_init(boolean_t from_boot)
966 {
967 	cpu_data_t *cdp = getCpuDatap();
968 
969 	if (cdp->cpu_get_fiq_handler == NULL) {
970 		cdp->cpu_get_fiq_handler = rtclock_timebase_func.tbd_fiq_handler;
971 		cdp->cpu_get_decrementer_func = rtclock_timebase_func.tbd_get_decrementer;
972 		cdp->cpu_set_decrementer_func = rtclock_timebase_func.tbd_set_decrementer;
973 		cdp->cpu_tbd_hardware_addr = (void *)rtclock_timebase_addr;
974 		cdp->cpu_tbd_hardware_val = (void *)rtclock_timebase_val;
975 	}
976 
977 	if (!from_boot && (cdp == &BootCpuData)) {
978 		/*
979 		 * When we wake from sleep, we have no guarantee about the state
980 		 * of the hardware timebase.  It may have kept ticking across sleep, or
981 		 * it may have reset.
982 		 *
983 		 * To deal with this, we calculate an offset to the clock that will
984 		 * produce a timebase value wake_abstime at the point the boot
985 		 * CPU calls cpu_timebase_init on wake.
986 		 *
987 		 * This ensures that mach_absolute_time() stops ticking across sleep.
988 		 */
989 		rtclock_base_abstime = wake_abstime - ml_get_hwclock();
990 	} else if (from_boot) {
991 		/* On initial boot, initialize time_since_reset to CNTPCT_EL0. */
992 		ml_set_reset_time(ml_get_hwclock());
993 	}
994 
995 	cdp->cpu_decrementer = 0x7FFFFFFFUL;
996 	cdp->cpu_timebase = 0x0UL;
997 	cdp->cpu_base_timebase = rtclock_base_abstime;
998 }
999 
1000 int
cpu_cluster_id(void)1001 cpu_cluster_id(void)
1002 {
1003 	return getCpuDatap()->cpu_cluster_id;
1004 }
1005 
1006 __attribute__((noreturn))
1007 void
ml_arm_sleep(void)1008 ml_arm_sleep(void)
1009 {
1010 	cpu_data_t              *cpu_data_ptr = getCpuDatap();
1011 
1012 	if (cpu_data_ptr == &BootCpuData) {
1013 		cpu_data_t      *target_cdp;
1014 		int             cpu;
1015 		int             max_cpu;
1016 
1017 		max_cpu = ml_get_max_cpu_number();
1018 		for (cpu = 0; cpu <= max_cpu; cpu++) {
1019 			target_cdp = (cpu_data_t *)CpuDataEntries[cpu].cpu_data_vaddr;
1020 
1021 			if ((target_cdp == NULL) || (target_cdp == cpu_data_ptr)) {
1022 				continue;
1023 			}
1024 
1025 			while (target_cdp->cpu_sleep_token != ARM_CPU_ON_SLEEP_PATH) {
1026 				;
1027 			}
1028 		}
1029 
1030 		/*
1031 		 * Now that the other cores have entered the sleep path, set
1032 		 * the abstime value we'll use when we resume.
1033 		 */
1034 		wake_abstime = ml_get_timebase();
1035 		ml_set_reset_time(UINT64_MAX);
1036 	} else {
1037 		CleanPoU_Dcache();
1038 	}
1039 
1040 	cpu_data_ptr->cpu_sleep_token = ARM_CPU_ON_SLEEP_PATH;
1041 
1042 	if (cpu_data_ptr == &BootCpuData) {
1043 #if WITH_CLASSIC_S2R
1044 		// Classic suspend to RAM writes the suspend signature into the
1045 		// sleep token buffer so that iBoot knows that it's on the warm
1046 		// boot (wake) path (as opposed to the cold boot path). Newer SoC
1047 		// do not go through SecureROM/iBoot on the warm boot path. The
1048 		// reconfig engine script brings the CPU out of reset at the kernel's
1049 		// reset vector which points to the warm boot initialization code.
1050 		if (sleepTokenBuffer != (vm_offset_t) NULL) {
1051 			platform_cache_shutdown();
1052 			bcopy((const void *)suspend_signature, (void *)sleepTokenBuffer, sizeof(SleepToken));
1053 		} else {
1054 			panic("No sleep token buffer");
1055 		}
1056 #endif
1057 
1058 #if __ARM_GLOBAL_SLEEP_BIT__
1059 		/* Allow other CPUs to go to sleep. */
1060 		arm64_stall_sleep = FALSE;
1061 		__builtin_arm_dmb(DMB_ISH);
1062 #endif
1063 
1064 		/* Architectural debug state: <rdar://problem/12390433>:
1065 		 *      Grab debug lock EDLAR and clear bit 0 in EDPRCR,
1066 		 *      tell debugger to not prevent power gating .
1067 		 */
1068 		if (cpu_data_ptr->coresight_base[CORESIGHT_ED]) {
1069 			*(volatile uint32_t *)(cpu_data_ptr->coresight_base[CORESIGHT_ED] + ARM_DEBUG_OFFSET_DBGLAR) = ARM_DBG_LOCK_ACCESS_KEY;
1070 			*(volatile uint32_t *)(cpu_data_ptr->coresight_base[CORESIGHT_ED] + ARM_DEBUG_OFFSET_DBGPRCR) = 0;
1071 		}
1072 
1073 #if HIBERNATION
1074 		uint32_t mode = hibernate_write_image();
1075 		if (mode == kIOHibernatePostWriteHalt) {
1076 			HIBLOG("powering off after writing hibernation image\n");
1077 			int halt_result = -1;
1078 			if (PE_halt_restart) {
1079 				/**
1080 				 * Drain serial FIFOs now as the normal call further down won't
1081 				 * be hit when the CPU halts here for hibernation. Here, it'll
1082 				 * make sure the preceding HIBLOG is flushed as well.
1083 				 */
1084 				serial_go_to_sleep();
1085 				halt_result = (*PE_halt_restart)(kPEHaltCPU);
1086 			}
1087 			panic("can't shutdown: PE_halt_restart returned %d", halt_result);
1088 		}
1089 #endif /* HIBERNATION */
1090 
1091 		serial_go_to_sleep();
1092 
1093 #if MONOTONIC
1094 		mt_sleep();
1095 #endif /* MONOTONIC */
1096 		/* ARM64-specific preparation */
1097 		arm64_prepare_for_sleep(true, cpu_data_ptr->cpu_number, ml_vtophys((vm_offset_t)&LowResetVectorBase));
1098 	} else {
1099 #if __ARM_GLOBAL_SLEEP_BIT__
1100 		/*
1101 		 * With the exception of the CPU revisions listed above, our ARM64 CPUs have a
1102 		 * global register to manage entering deep sleep, as opposed to a per-CPU
1103 		 * register.  We cannot update this register until all CPUs are ready to enter
1104 		 * deep sleep, because if a CPU executes WFI outside of the deep sleep context
1105 		 * (by idling), it will hang (due to the side effects of enabling deep sleep),
1106 		 * which can hang the sleep process or cause memory corruption on wake.
1107 		 *
1108 		 * To avoid these issues, we'll stall on this global value, which CPU0 will
1109 		 * manage.
1110 		 */
1111 		while (arm64_stall_sleep) {
1112 			__builtin_arm_wfe();
1113 		}
1114 #endif
1115 		CleanPoU_DcacheRegion((vm_offset_t) cpu_data_ptr, sizeof(cpu_data_t));
1116 
1117 		/* Architectural debug state: <rdar://problem/12390433>:
1118 		 *      Grab debug lock EDLAR and clear bit 0 in EDPRCR,
1119 		 *      tell debugger to not prevent power gating .
1120 		 */
1121 		if (cpu_data_ptr->coresight_base[CORESIGHT_ED]) {
1122 			*(volatile uint32_t *)(cpu_data_ptr->coresight_base[CORESIGHT_ED] + ARM_DEBUG_OFFSET_DBGLAR) = ARM_DBG_LOCK_ACCESS_KEY;
1123 			*(volatile uint32_t *)(cpu_data_ptr->coresight_base[CORESIGHT_ED] + ARM_DEBUG_OFFSET_DBGPRCR) = 0;
1124 		}
1125 
1126 		/* ARM64-specific preparation */
1127 		arm64_prepare_for_sleep(true, cpu_data_ptr->cpu_number, ml_vtophys((vm_offset_t)&LowResetVectorBase));
1128 	}
1129 }
1130 
1131 void
cpu_machine_idle_init(boolean_t from_boot)1132 cpu_machine_idle_init(boolean_t from_boot)
1133 {
1134 	static vm_address_t     resume_idle_cpu_paddr = (vm_address_t)NULL;
1135 	cpu_data_t              *cpu_data_ptr   = getCpuDatap();
1136 
1137 	if (from_boot) {
1138 		uint32_t        production = 1;
1139 		DTEntry         entry;
1140 
1141 		unsigned long   jtag = 0;
1142 
1143 		if (PE_parse_boot_argn("jtag", &jtag, sizeof(jtag))) {
1144 			if (jtag != 0) {
1145 				idle_enable = FALSE;
1146 			} else {
1147 				idle_enable = TRUE;
1148 			}
1149 		} else {
1150 			idle_enable = TRUE;
1151 		}
1152 
1153 #if DEVELOPMENT || DEBUG
1154 		uint32_t wfe_mode = 0;
1155 		if (PE_parse_boot_argn("wfe_mode", &wfe_mode, sizeof(wfe_mode))) {
1156 			idle_proximate_timer_wfe = ((wfe_mode & 1) == 1);
1157 			idle_proximate_io_wfe_masked = ((wfe_mode & 2) == 2);
1158 			extern uint32_t idle_proximate_io_wfe_unmasked;
1159 			idle_proximate_io_wfe_unmasked = ((wfe_mode & 4) == 4);
1160 		}
1161 #endif
1162 
1163 		// bits 7..0 give the wfi type
1164 		switch (wfi & 0xff) {
1165 		case 0:
1166 			// disable wfi
1167 			wfi = 0;
1168 			break;
1169 
1170 #if DEVELOPMENT || DEBUG
1171 		case 2:
1172 			// wfi overhead simulation
1173 			// 31..16 - wfi delay is us
1174 			// 15..8  - flags
1175 			// 7..0   - 2
1176 			wfi = 2;
1177 			wfi_flags = (wfi >> 8) & 0xFF;
1178 			nanoseconds_to_absolutetime(((wfi >> 16) & 0xFFFF) * NSEC_PER_MSEC, &wfi_delay);
1179 			break;
1180 #endif /* DEVELOPMENT || DEBUG */
1181 
1182 		case 1:
1183 		default:
1184 			// do nothing
1185 			break;
1186 		}
1187 
1188 		ResetHandlerData.assist_reset_handler = 0;
1189 		ResetHandlerData.cpu_data_entries = ml_static_vtop((vm_offset_t)CpuDataEntries);
1190 
1191 #ifdef MONITOR
1192 		monitor_call(MONITOR_SET_ENTRY, (uintptr_t)ml_static_vtop((vm_offset_t)&LowResetVectorBase), 0, 0);
1193 #elif !defined(NO_MONITOR)
1194 #error MONITOR undefined, WFI power gating may not operate correctly
1195 #endif /* MONITOR */
1196 
1197 		// Determine if we are on production or debug chip
1198 		if (kSuccess == SecureDTLookupEntry(NULL, "/chosen", &entry)) {
1199 			unsigned int    size;
1200 			void const      *prop;
1201 
1202 			if (kSuccess == SecureDTGetProperty(entry, "effective-production-status-ap", &prop, &size)) {
1203 				if (size == 4) {
1204 					bcopy(prop, &production, size);
1205 				}
1206 			}
1207 		}
1208 		if (!production) {
1209 #if defined(APPLE_ARM64_ARCH_FAMILY)
1210 			// Enable coresight debug registers on debug-fused chips
1211 			coresight_debug_enabled = TRUE;
1212 #endif
1213 		}
1214 
1215 		start_cpu_paddr = ml_static_vtop((vm_offset_t)&start_cpu);
1216 		resume_idle_cpu_paddr = ml_static_vtop((vm_offset_t)&resume_idle_cpu);
1217 	}
1218 
1219 #if WITH_CLASSIC_S2R
1220 	if (cpu_data_ptr == &BootCpuData) {
1221 		static addr64_t SleepToken_low_paddr = (addr64_t)NULL;
1222 		if (sleepTokenBuffer != (vm_offset_t) NULL) {
1223 			SleepToken_low_paddr = ml_vtophys(sleepTokenBuffer);
1224 		} else {
1225 			panic("No sleep token buffer");
1226 		}
1227 
1228 		bcopy_phys((addr64_t)ml_static_vtop((vm_offset_t)running_signature),
1229 		    SleepToken_low_paddr, sizeof(SleepToken));
1230 		flush_dcache((vm_offset_t)SleepToken, sizeof(SleepToken), TRUE);
1231 	}
1232 	;
1233 #endif
1234 
1235 	cpu_data_ptr->cpu_reset_handler = resume_idle_cpu_paddr;
1236 	clean_dcache((vm_offset_t)cpu_data_ptr, sizeof(cpu_data_t), FALSE);
1237 }
1238 
1239 _Atomic uint32_t cpu_idle_count = 0;
1240 
1241 void
machine_track_platform_idle(boolean_t entry)1242 machine_track_platform_idle(boolean_t entry)
1243 {
1244 	if (entry) {
1245 		os_atomic_inc(&cpu_idle_count, relaxed);
1246 	} else {
1247 		os_atomic_dec(&cpu_idle_count, relaxed);
1248 	}
1249 }
1250 
1251 #if WITH_CLASSIC_S2R
1252 void
sleep_token_buffer_init(void)1253 sleep_token_buffer_init(void)
1254 {
1255 	cpu_data_t      *cpu_data_ptr = getCpuDatap();
1256 	DTEntry         entry;
1257 	size_t          size;
1258 	void const * const *prop;
1259 
1260 	if ((cpu_data_ptr == &BootCpuData) && (sleepTokenBuffer == (vm_offset_t) NULL)) {
1261 		/* Find the stpage node in the device tree */
1262 		if (kSuccess != SecureDTLookupEntry(0, "stram", &entry)) {
1263 			return;
1264 		}
1265 
1266 		if (kSuccess != SecureDTGetProperty(entry, "reg", (const void **)&prop, (unsigned int *)&size)) {
1267 			return;
1268 		}
1269 
1270 		/* Map the page into the kernel space */
1271 		sleepTokenBuffer = ml_io_map(((vm_offset_t const *)prop)[0], ((vm_size_t const *)prop)[1]);
1272 	}
1273 }
1274 #endif
1275