xref: /xnu-11417.121.6/osfmk/arm/machine_routines_common.c (revision a1e26a70f38d1d7daa7b49b258e2f8538ad81650)
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 #include <arm/machine_cpu.h>
30 #include <arm/cpu_internal.h>
31 #include <arm/cpuid.h>
32 #include <arm/cpuid_internal.h>
33 #include <arm/cpu_data.h>
34 #include <arm/cpu_data_internal.h>
35 #include <arm/misc_protos.h>
36 #include <arm/machdep_call.h>
37 #include <arm/machine_routines.h>
38 #include <arm/rtclock.h>
39 #include <kern/machine.h>
40 #include <kern/thread.h>
41 #include <kern/thread_group.h>
42 #include <kern/policy_internal.h>
43 #include <kern/sched_hygiene.h>
44 #include <kern/startup.h>
45 #include <kern/monotonic.h>
46 #include <machine/config.h>
47 #include <machine/atomic.h>
48 #include <machine/monotonic.h>
49 #include <pexpert/pexpert.h>
50 #include <pexpert/device_tree.h>
51 #include <pexpert/arm64/apple_arm64_cpu.h>
52 
53 #include <mach/machine.h>
54 #include <mach/machine/sdt.h>
55 
56 #if !HAS_CONTINUOUS_HWCLOCK
57 extern uint64_t mach_absolutetime_asleep;
58 #else
59 extern uint64_t wake_abstime;
60 static uint64_t wake_conttime = UINT64_MAX;
61 #endif
62 
63 extern volatile uint32_t debug_enabled;
64 extern _Atomic unsigned int cluster_type_num_active_cpus[MAX_CPU_TYPES];
65 const char *cluster_type_names[MAX_CPU_TYPES] = {
66 	[CLUSTER_TYPE_SMP] = "Standard",
67 	[CLUSTER_TYPE_P] = "Performance",
68 	[CLUSTER_TYPE_E] = "Efficiency",
69 };
70 
71 static int max_cpus_initialized = 0;
72 #define MAX_CPUS_SET    0x1
73 #define MAX_CPUS_WAIT   0x2
74 
75 LCK_GRP_DECLARE(max_cpus_grp, "max_cpus");
76 LCK_MTX_DECLARE(max_cpus_lock, &max_cpus_grp);
77 uint32_t lockdown_done = 0;
78 boolean_t is_clock_configured = FALSE;
79 
80 static void
sched_perfcontrol_oncore_default(perfcontrol_state_t new_thread_state __unused,going_on_core_t on __unused)81 sched_perfcontrol_oncore_default(perfcontrol_state_t new_thread_state __unused, going_on_core_t on __unused)
82 {
83 }
84 
85 static void
sched_perfcontrol_switch_default(perfcontrol_state_t old_thread_state __unused,perfcontrol_state_t new_thread_state __unused)86 sched_perfcontrol_switch_default(perfcontrol_state_t old_thread_state __unused, perfcontrol_state_t new_thread_state __unused)
87 {
88 }
89 
90 static void
sched_perfcontrol_offcore_default(perfcontrol_state_t old_thread_state __unused,going_off_core_t off __unused,boolean_t thread_terminating __unused)91 sched_perfcontrol_offcore_default(perfcontrol_state_t old_thread_state __unused, going_off_core_t off __unused, boolean_t thread_terminating __unused)
92 {
93 }
94 
95 static void
sched_perfcontrol_thread_group_default(thread_group_data_t data __unused)96 sched_perfcontrol_thread_group_default(thread_group_data_t data __unused)
97 {
98 }
99 
100 static void
sched_perfcontrol_max_runnable_latency_default(perfcontrol_max_runnable_latency_t latencies __unused)101 sched_perfcontrol_max_runnable_latency_default(perfcontrol_max_runnable_latency_t latencies __unused)
102 {
103 }
104 
105 static void
sched_perfcontrol_work_interval_notify_default(perfcontrol_state_t thread_state __unused,perfcontrol_work_interval_t work_interval __unused)106 sched_perfcontrol_work_interval_notify_default(perfcontrol_state_t thread_state __unused,
107     perfcontrol_work_interval_t work_interval __unused)
108 {
109 }
110 
111 static void
sched_perfcontrol_work_interval_ctl_default(perfcontrol_state_t thread_state __unused,perfcontrol_work_interval_instance_t instance __unused)112 sched_perfcontrol_work_interval_ctl_default(perfcontrol_state_t thread_state __unused,
113     perfcontrol_work_interval_instance_t instance __unused)
114 {
115 }
116 
117 static void
sched_perfcontrol_deadline_passed_default(__unused uint64_t deadline)118 sched_perfcontrol_deadline_passed_default(__unused uint64_t deadline)
119 {
120 }
121 
122 static void
sched_perfcontrol_csw_default(__unused perfcontrol_event event,__unused uint32_t cpu_id,__unused uint64_t timestamp,__unused uint32_t flags,__unused struct perfcontrol_thread_data * offcore,__unused struct perfcontrol_thread_data * oncore,__unused struct perfcontrol_cpu_counters * cpu_counters,__unused void * unused)123 sched_perfcontrol_csw_default(
124 	__unused perfcontrol_event event, __unused uint32_t cpu_id, __unused uint64_t timestamp,
125 	__unused uint32_t flags, __unused struct perfcontrol_thread_data *offcore,
126 	__unused struct perfcontrol_thread_data *oncore,
127 	__unused struct perfcontrol_cpu_counters *cpu_counters, __unused void *unused)
128 {
129 }
130 
131 static void
sched_perfcontrol_state_update_default(__unused perfcontrol_event event,__unused uint32_t cpu_id,__unused uint64_t timestamp,__unused uint32_t flags,__unused struct perfcontrol_thread_data * thr_data,__unused void * unused)132 sched_perfcontrol_state_update_default(
133 	__unused perfcontrol_event event, __unused uint32_t cpu_id, __unused uint64_t timestamp,
134 	__unused uint32_t flags, __unused struct perfcontrol_thread_data *thr_data,
135 	__unused void *unused)
136 {
137 }
138 
139 static void
sched_perfcontrol_thread_group_blocked_default(__unused thread_group_data_t blocked_tg,__unused thread_group_data_t blocking_tg,__unused uint32_t flags,__unused perfcontrol_state_t blocked_thr_state)140 sched_perfcontrol_thread_group_blocked_default(
141 	__unused thread_group_data_t blocked_tg, __unused thread_group_data_t blocking_tg,
142 	__unused uint32_t flags, __unused perfcontrol_state_t blocked_thr_state)
143 {
144 }
145 
146 static void
sched_perfcontrol_thread_group_unblocked_default(__unused thread_group_data_t unblocked_tg,__unused thread_group_data_t unblocking_tg,__unused uint32_t flags,__unused perfcontrol_state_t unblocked_thr_state)147 sched_perfcontrol_thread_group_unblocked_default(
148 	__unused thread_group_data_t unblocked_tg, __unused thread_group_data_t unblocking_tg,
149 	__unused uint32_t flags, __unused perfcontrol_state_t unblocked_thr_state)
150 {
151 }
152 
153 sched_perfcontrol_offcore_t                     sched_perfcontrol_offcore = sched_perfcontrol_offcore_default;
154 sched_perfcontrol_context_switch_t              sched_perfcontrol_switch = sched_perfcontrol_switch_default;
155 sched_perfcontrol_oncore_t                      sched_perfcontrol_oncore = sched_perfcontrol_oncore_default;
156 sched_perfcontrol_thread_group_init_t           sched_perfcontrol_thread_group_init = sched_perfcontrol_thread_group_default;
157 sched_perfcontrol_thread_group_deinit_t         sched_perfcontrol_thread_group_deinit = sched_perfcontrol_thread_group_default;
158 sched_perfcontrol_thread_group_flags_update_t   sched_perfcontrol_thread_group_flags_update = sched_perfcontrol_thread_group_default;
159 sched_perfcontrol_max_runnable_latency_t        sched_perfcontrol_max_runnable_latency = sched_perfcontrol_max_runnable_latency_default;
160 sched_perfcontrol_work_interval_notify_t        sched_perfcontrol_work_interval_notify = sched_perfcontrol_work_interval_notify_default;
161 sched_perfcontrol_work_interval_ctl_t           sched_perfcontrol_work_interval_ctl = sched_perfcontrol_work_interval_ctl_default;
162 sched_perfcontrol_deadline_passed_t             sched_perfcontrol_deadline_passed = sched_perfcontrol_deadline_passed_default;
163 sched_perfcontrol_csw_t                         sched_perfcontrol_csw = sched_perfcontrol_csw_default;
164 sched_perfcontrol_state_update_t                sched_perfcontrol_state_update = sched_perfcontrol_state_update_default;
165 sched_perfcontrol_thread_group_blocked_t        sched_perfcontrol_thread_group_blocked = sched_perfcontrol_thread_group_blocked_default;
166 sched_perfcontrol_thread_group_unblocked_t      sched_perfcontrol_thread_group_unblocked = sched_perfcontrol_thread_group_unblocked_default;
167 boolean_t sched_perfcontrol_thread_shared_rsrc_flags_enabled = false;
168 
169 void
sched_perfcontrol_register_callbacks(sched_perfcontrol_callbacks_t callbacks,unsigned long size_of_state)170 sched_perfcontrol_register_callbacks(sched_perfcontrol_callbacks_t callbacks, unsigned long size_of_state)
171 {
172 	assert(callbacks == NULL || callbacks->version >= SCHED_PERFCONTROL_CALLBACKS_VERSION_2);
173 
174 	if (size_of_state > sizeof(struct perfcontrol_state)) {
175 		panic("%s: Invalid required state size %lu", __FUNCTION__, size_of_state);
176 	}
177 
178 	if (callbacks) {
179 #if CONFIG_THREAD_GROUPS
180 		if (callbacks->version >= SCHED_PERFCONTROL_CALLBACKS_VERSION_3) {
181 			if (callbacks->thread_group_init != NULL) {
182 				sched_perfcontrol_thread_group_init = callbacks->thread_group_init;
183 			} else {
184 				sched_perfcontrol_thread_group_init = sched_perfcontrol_thread_group_default;
185 			}
186 			if (callbacks->thread_group_deinit != NULL) {
187 				sched_perfcontrol_thread_group_deinit = callbacks->thread_group_deinit;
188 			} else {
189 				sched_perfcontrol_thread_group_deinit = sched_perfcontrol_thread_group_default;
190 			}
191 			// tell CLPC about existing thread groups
192 			thread_group_resync(TRUE);
193 		}
194 
195 		if (callbacks->version >= SCHED_PERFCONTROL_CALLBACKS_VERSION_6) {
196 			if (callbacks->thread_group_flags_update != NULL) {
197 				sched_perfcontrol_thread_group_flags_update = callbacks->thread_group_flags_update;
198 			} else {
199 				sched_perfcontrol_thread_group_flags_update = sched_perfcontrol_thread_group_default;
200 			}
201 		}
202 
203 		if (callbacks->version >= SCHED_PERFCONTROL_CALLBACKS_VERSION_8) {
204 			if (callbacks->thread_group_blocked != NULL) {
205 				sched_perfcontrol_thread_group_blocked = callbacks->thread_group_blocked;
206 			} else {
207 				sched_perfcontrol_thread_group_blocked = sched_perfcontrol_thread_group_blocked_default;
208 			}
209 
210 			if (callbacks->thread_group_unblocked != NULL) {
211 				sched_perfcontrol_thread_group_unblocked = callbacks->thread_group_unblocked;
212 			} else {
213 				sched_perfcontrol_thread_group_unblocked = sched_perfcontrol_thread_group_unblocked_default;
214 			}
215 		}
216 #endif
217 		if (callbacks->version >= SCHED_PERFCONTROL_CALLBACKS_VERSION_9) {
218 			sched_perfcontrol_thread_shared_rsrc_flags_enabled = true;
219 		}
220 
221 		if (callbacks->version >= SCHED_PERFCONTROL_CALLBACKS_VERSION_7) {
222 			if (callbacks->work_interval_ctl != NULL) {
223 				sched_perfcontrol_work_interval_ctl = callbacks->work_interval_ctl;
224 			} else {
225 				sched_perfcontrol_work_interval_ctl = sched_perfcontrol_work_interval_ctl_default;
226 			}
227 		}
228 
229 		if (callbacks->version >= SCHED_PERFCONTROL_CALLBACKS_VERSION_5) {
230 			if (callbacks->csw != NULL) {
231 				sched_perfcontrol_csw = callbacks->csw;
232 			} else {
233 				sched_perfcontrol_csw = sched_perfcontrol_csw_default;
234 			}
235 
236 			if (callbacks->state_update != NULL) {
237 				sched_perfcontrol_state_update = callbacks->state_update;
238 			} else {
239 				sched_perfcontrol_state_update = sched_perfcontrol_state_update_default;
240 			}
241 		}
242 
243 		if (callbacks->version >= SCHED_PERFCONTROL_CALLBACKS_VERSION_4) {
244 			if (callbacks->deadline_passed != NULL) {
245 				sched_perfcontrol_deadline_passed = callbacks->deadline_passed;
246 			} else {
247 				sched_perfcontrol_deadline_passed = sched_perfcontrol_deadline_passed_default;
248 			}
249 		}
250 
251 		if (callbacks->offcore != NULL) {
252 			sched_perfcontrol_offcore = callbacks->offcore;
253 		} else {
254 			sched_perfcontrol_offcore = sched_perfcontrol_offcore_default;
255 		}
256 
257 		if (callbacks->context_switch != NULL) {
258 			sched_perfcontrol_switch = callbacks->context_switch;
259 		} else {
260 			sched_perfcontrol_switch = sched_perfcontrol_switch_default;
261 		}
262 
263 		if (callbacks->oncore != NULL) {
264 			sched_perfcontrol_oncore = callbacks->oncore;
265 		} else {
266 			sched_perfcontrol_oncore = sched_perfcontrol_oncore_default;
267 		}
268 
269 		if (callbacks->max_runnable_latency != NULL) {
270 			sched_perfcontrol_max_runnable_latency = callbacks->max_runnable_latency;
271 		} else {
272 			sched_perfcontrol_max_runnable_latency = sched_perfcontrol_max_runnable_latency_default;
273 		}
274 
275 		if (callbacks->work_interval_notify != NULL) {
276 			sched_perfcontrol_work_interval_notify = callbacks->work_interval_notify;
277 		} else {
278 			sched_perfcontrol_work_interval_notify = sched_perfcontrol_work_interval_notify_default;
279 		}
280 	} else {
281 		/* reset to defaults */
282 #if CONFIG_THREAD_GROUPS
283 		thread_group_resync(FALSE);
284 #endif
285 		sched_perfcontrol_offcore = sched_perfcontrol_offcore_default;
286 		sched_perfcontrol_switch = sched_perfcontrol_switch_default;
287 		sched_perfcontrol_oncore = sched_perfcontrol_oncore_default;
288 		sched_perfcontrol_thread_group_init = sched_perfcontrol_thread_group_default;
289 		sched_perfcontrol_thread_group_deinit = sched_perfcontrol_thread_group_default;
290 		sched_perfcontrol_thread_group_flags_update = sched_perfcontrol_thread_group_default;
291 		sched_perfcontrol_max_runnable_latency = sched_perfcontrol_max_runnable_latency_default;
292 		sched_perfcontrol_work_interval_notify = sched_perfcontrol_work_interval_notify_default;
293 		sched_perfcontrol_work_interval_ctl = sched_perfcontrol_work_interval_ctl_default;
294 		sched_perfcontrol_csw = sched_perfcontrol_csw_default;
295 		sched_perfcontrol_state_update = sched_perfcontrol_state_update_default;
296 		sched_perfcontrol_thread_group_blocked = sched_perfcontrol_thread_group_blocked_default;
297 		sched_perfcontrol_thread_group_unblocked = sched_perfcontrol_thread_group_unblocked_default;
298 	}
299 }
300 
301 
302 static void
machine_switch_populate_perfcontrol_thread_data(struct perfcontrol_thread_data * data,thread_t thread,uint64_t same_pri_latency)303 machine_switch_populate_perfcontrol_thread_data(struct perfcontrol_thread_data *data,
304     thread_t thread,
305     uint64_t same_pri_latency)
306 {
307 	bzero(data, sizeof(struct perfcontrol_thread_data));
308 	data->perfctl_class = thread_get_perfcontrol_class(thread);
309 	data->energy_estimate_nj = 0;
310 	data->thread_id = thread->thread_id;
311 #if CONFIG_THREAD_GROUPS
312 	struct thread_group *tg = thread_group_get(thread);
313 	data->thread_group_id = thread_group_get_id(tg);
314 	data->thread_group_data = thread_group_get_machine_data(tg);
315 #endif
316 	data->scheduling_latency_at_same_basepri = same_pri_latency;
317 	data->perfctl_state = FIND_PERFCONTROL_STATE(thread);
318 }
319 
320 static void
machine_switch_populate_perfcontrol_cpu_counters(struct perfcontrol_cpu_counters * cpu_counters)321 machine_switch_populate_perfcontrol_cpu_counters(struct perfcontrol_cpu_counters *cpu_counters)
322 {
323 #if CONFIG_CPU_COUNTERS
324 	mt_perfcontrol(&cpu_counters->instructions, &cpu_counters->cycles);
325 #else /* CONFIG_CPU_COUNTERS */
326 	cpu_counters->instructions = 0;
327 	cpu_counters->cycles = 0;
328 #endif /* !CONFIG_CPU_COUNTERS */
329 }
330 
331 int perfcontrol_callout_stats_enabled = 0;
332 static _Atomic uint64_t perfcontrol_callout_stats[PERFCONTROL_CALLOUT_MAX][PERFCONTROL_STAT_MAX];
333 static _Atomic uint64_t perfcontrol_callout_count[PERFCONTROL_CALLOUT_MAX];
334 
335 #if CONFIG_CPU_COUNTERS
336 static inline
337 bool
perfcontrol_callout_counters_begin(uint64_t * counters)338 perfcontrol_callout_counters_begin(uint64_t *counters)
339 {
340 	if (!perfcontrol_callout_stats_enabled) {
341 		return false;
342 	}
343 	mt_fixed_counts(counters);
344 	return true;
345 }
346 
347 static inline
348 void
perfcontrol_callout_counters_end(uint64_t * start_counters,perfcontrol_callout_type_t type)349 perfcontrol_callout_counters_end(uint64_t *start_counters,
350     perfcontrol_callout_type_t type)
351 {
352 	uint64_t end_counters[MT_CORE_NFIXED];
353 	mt_fixed_counts(end_counters);
354 	os_atomic_add(&perfcontrol_callout_stats[type][PERFCONTROL_STAT_CYCLES],
355 	    end_counters[MT_CORE_CYCLES] - start_counters[MT_CORE_CYCLES], relaxed);
356 	os_atomic_add(&perfcontrol_callout_stats[type][PERFCONTROL_STAT_INSTRS],
357 	    end_counters[MT_CORE_INSTRS] - start_counters[MT_CORE_INSTRS], relaxed);
358 	os_atomic_inc(&perfcontrol_callout_count[type], relaxed);
359 }
360 #endif /* CONFIG_CPU_COUNTERS */
361 
362 uint64_t
perfcontrol_callout_stat_avg(perfcontrol_callout_type_t type,perfcontrol_callout_stat_t stat)363 perfcontrol_callout_stat_avg(perfcontrol_callout_type_t type,
364     perfcontrol_callout_stat_t stat)
365 {
366 	if (!perfcontrol_callout_stats_enabled) {
367 		return 0;
368 	}
369 	return os_atomic_load_wide(&perfcontrol_callout_stats[type][stat], relaxed) /
370 	       os_atomic_load_wide(&perfcontrol_callout_count[type], relaxed);
371 }
372 
373 #if CONFIG_SCHED_EDGE
374 
375 /*
376  * The Edge scheduler allows the performance controller to update properties about the
377  * threads as part of the callouts. These properties typically include shared cluster
378  * resource usage. This allows the scheduler to manage specific threads within the
379  * workload more optimally.
380  */
381 static void
sched_perfcontrol_thread_flags_update(thread_t thread,struct perfcontrol_thread_data * thread_data,shared_rsrc_policy_agent_t agent)382 sched_perfcontrol_thread_flags_update(thread_t thread,
383     struct perfcontrol_thread_data *thread_data,
384     shared_rsrc_policy_agent_t agent)
385 {
386 	kern_return_t kr = KERN_SUCCESS;
387 	if (thread_data->thread_flags_mask & PERFCTL_THREAD_FLAGS_MASK_CLUSTER_SHARED_RSRC_RR) {
388 		if (thread_data->thread_flags & PERFCTL_THREAD_FLAGS_MASK_CLUSTER_SHARED_RSRC_RR) {
389 			kr = thread_shared_rsrc_policy_set(thread, 0, CLUSTER_SHARED_RSRC_TYPE_RR, agent);
390 		} else {
391 			kr = thread_shared_rsrc_policy_clear(thread, CLUSTER_SHARED_RSRC_TYPE_RR, agent);
392 		}
393 	}
394 	if (thread_data->thread_flags_mask & PERFCTL_THREAD_FLAGS_MASK_CLUSTER_SHARED_RSRC_NATIVE_FIRST) {
395 		if (thread_data->thread_flags & PERFCTL_THREAD_FLAGS_MASK_CLUSTER_SHARED_RSRC_NATIVE_FIRST) {
396 			kr = thread_shared_rsrc_policy_set(thread, 0, CLUSTER_SHARED_RSRC_TYPE_NATIVE_FIRST, agent);
397 		} else {
398 			kr = thread_shared_rsrc_policy_clear(thread, CLUSTER_SHARED_RSRC_TYPE_NATIVE_FIRST, agent);
399 		}
400 	}
401 	/*
402 	 * The thread_shared_rsrc_policy_* routines only fail if the performance controller is
403 	 * attempting to double set/clear a policy on the thread.
404 	 */
405 	assert(kr == KERN_SUCCESS);
406 }
407 
408 #endif /* CONFIG_SCHED_EDGE */
409 
410 void
machine_switch_perfcontrol_context(perfcontrol_event event,uint64_t timestamp,uint32_t flags,uint64_t new_thread_same_pri_latency,thread_t old,thread_t new)411 machine_switch_perfcontrol_context(perfcontrol_event event,
412     uint64_t timestamp,
413     uint32_t flags,
414     uint64_t new_thread_same_pri_latency,
415     thread_t old,
416     thread_t new)
417 {
418 
419 	if (sched_perfcontrol_switch != sched_perfcontrol_switch_default) {
420 		perfcontrol_state_t old_perfcontrol_state = FIND_PERFCONTROL_STATE(old);
421 		perfcontrol_state_t new_perfcontrol_state = FIND_PERFCONTROL_STATE(new);
422 		sched_perfcontrol_switch(old_perfcontrol_state, new_perfcontrol_state);
423 	}
424 
425 	if (sched_perfcontrol_csw != sched_perfcontrol_csw_default) {
426 		uint32_t cpu_id = (uint32_t)cpu_number();
427 		struct perfcontrol_cpu_counters cpu_counters;
428 		struct perfcontrol_thread_data offcore, oncore;
429 		machine_switch_populate_perfcontrol_thread_data(&offcore, old, 0);
430 		machine_switch_populate_perfcontrol_thread_data(&oncore, new,
431 		    new_thread_same_pri_latency);
432 		machine_switch_populate_perfcontrol_cpu_counters(&cpu_counters);
433 
434 #if CONFIG_CPU_COUNTERS
435 		uint64_t counters[MT_CORE_NFIXED];
436 		bool ctrs_enabled = perfcontrol_callout_counters_begin(counters);
437 #endif /* CONFIG_CPU_COUNTERS */
438 		sched_perfcontrol_csw(event, cpu_id, timestamp, flags,
439 		    &offcore, &oncore, &cpu_counters, NULL);
440 #if CONFIG_CPU_COUNTERS
441 		if (ctrs_enabled) {
442 			perfcontrol_callout_counters_end(counters, PERFCONTROL_CALLOUT_CONTEXT);
443 		}
444 #endif /* CONFIG_CPU_COUNTERS */
445 
446 		recount_add_energy(old, get_threadtask(old),
447 		    offcore.energy_estimate_nj);
448 
449 #if CONFIG_SCHED_EDGE
450 		if (sched_perfcontrol_thread_shared_rsrc_flags_enabled) {
451 			sched_perfcontrol_thread_flags_update(old, &offcore, SHARED_RSRC_POLICY_AGENT_PERFCTL_CSW);
452 		}
453 #endif /* CONFIG_SCHED_EDGE */
454 	}
455 }
456 
457 void
machine_switch_perfcontrol_state_update(perfcontrol_event event,uint64_t timestamp,uint32_t flags,thread_t thread)458 machine_switch_perfcontrol_state_update(perfcontrol_event event,
459     uint64_t timestamp,
460     uint32_t flags,
461     thread_t thread)
462 {
463 
464 	if (sched_perfcontrol_state_update == sched_perfcontrol_state_update_default) {
465 		return;
466 	}
467 	uint32_t cpu_id = (uint32_t)cpu_number();
468 	struct perfcontrol_thread_data data;
469 	machine_switch_populate_perfcontrol_thread_data(&data, thread, 0);
470 
471 #if CONFIG_CPU_COUNTERS
472 	uint64_t counters[MT_CORE_NFIXED];
473 	bool ctrs_enabled = perfcontrol_callout_counters_begin(counters);
474 #endif /* CONFIG_CPU_COUNTERS */
475 	sched_perfcontrol_state_update(event, cpu_id, timestamp, flags,
476 	    &data, NULL);
477 #if CONFIG_CPU_COUNTERS
478 	if (ctrs_enabled) {
479 		perfcontrol_callout_counters_end(counters, PERFCONTROL_CALLOUT_STATE_UPDATE);
480 	}
481 #endif /* CONFIG_CPU_COUNTERS */
482 
483 #if CONFIG_PERVASIVE_ENERGY
484 	recount_add_energy(thread, get_threadtask(thread), data.energy_estimate_nj);
485 #endif /* CONFIG_PERVASIVE_ENERGY */
486 
487 #if CONFIG_SCHED_EDGE
488 	if (sched_perfcontrol_thread_shared_rsrc_flags_enabled && (event == QUANTUM_EXPIRY)) {
489 		sched_perfcontrol_thread_flags_update(thread, &data, SHARED_RSRC_POLICY_AGENT_PERFCTL_QUANTUM);
490 	} else {
491 		assert(data.thread_flags_mask == 0);
492 	}
493 #endif /* CONFIG_SCHED_EDGE */
494 }
495 
496 void
machine_thread_going_on_core(thread_t new_thread,thread_urgency_t urgency,uint64_t sched_latency,uint64_t same_pri_latency,uint64_t timestamp)497 machine_thread_going_on_core(thread_t   new_thread,
498     thread_urgency_t        urgency,
499     uint64_t   sched_latency,
500     uint64_t   same_pri_latency,
501     uint64_t   timestamp)
502 {
503 	if (sched_perfcontrol_oncore == sched_perfcontrol_oncore_default) {
504 		return;
505 	}
506 	struct going_on_core on_core;
507 	perfcontrol_state_t state = FIND_PERFCONTROL_STATE(new_thread);
508 
509 	on_core.thread_id = new_thread->thread_id;
510 	on_core.energy_estimate_nj = 0;
511 	on_core.qos_class = (uint16_t)proc_get_effective_thread_policy(new_thread, TASK_POLICY_QOS);
512 	on_core.urgency = (uint16_t)urgency;
513 	on_core.is_32_bit = thread_is_64bit_data(new_thread) ? FALSE : TRUE;
514 	on_core.is_kernel_thread = get_threadtask(new_thread) == kernel_task;
515 #if CONFIG_THREAD_GROUPS
516 	struct thread_group *tg = thread_group_get(new_thread);
517 	on_core.thread_group_id = thread_group_get_id(tg);
518 	on_core.thread_group_data = thread_group_get_machine_data(tg);
519 #endif
520 	on_core.scheduling_latency = sched_latency;
521 	on_core.start_time = timestamp;
522 	on_core.scheduling_latency_at_same_basepri = same_pri_latency;
523 
524 #if CONFIG_CPU_COUNTERS
525 	uint64_t counters[MT_CORE_NFIXED];
526 	bool ctrs_enabled = perfcontrol_callout_counters_begin(counters);
527 #endif /* CONFIG_CPU_COUNTERS */
528 	sched_perfcontrol_oncore(state, &on_core);
529 #if CONFIG_CPU_COUNTERS
530 	if (ctrs_enabled) {
531 		perfcontrol_callout_counters_end(counters, PERFCONTROL_CALLOUT_ON_CORE);
532 	}
533 #endif /* CONFIG_CPU_COUNTERS */
534 }
535 
536 void
machine_thread_going_off_core(thread_t old_thread,boolean_t thread_terminating,uint64_t last_dispatch,__unused boolean_t thread_runnable)537 machine_thread_going_off_core(thread_t old_thread, boolean_t thread_terminating,
538     uint64_t last_dispatch, __unused boolean_t thread_runnable)
539 {
540 	if (sched_perfcontrol_offcore == sched_perfcontrol_offcore_default) {
541 		return;
542 	}
543 	struct going_off_core off_core;
544 	perfcontrol_state_t state = FIND_PERFCONTROL_STATE(old_thread);
545 
546 	off_core.thread_id = old_thread->thread_id;
547 	off_core.energy_estimate_nj = 0;
548 	off_core.end_time = last_dispatch;
549 #if CONFIG_THREAD_GROUPS
550 	struct thread_group *tg = thread_group_get(old_thread);
551 	off_core.thread_group_id = thread_group_get_id(tg);
552 	off_core.thread_group_data = thread_group_get_machine_data(tg);
553 #endif
554 
555 #if CONFIG_CPU_COUNTERS
556 	uint64_t counters[MT_CORE_NFIXED];
557 	bool ctrs_enabled = perfcontrol_callout_counters_begin(counters);
558 #endif /* CONFIG_CPU_COUNTERS */
559 	sched_perfcontrol_offcore(state, &off_core, thread_terminating);
560 #if CONFIG_CPU_COUNTERS
561 	if (ctrs_enabled) {
562 		perfcontrol_callout_counters_end(counters, PERFCONTROL_CALLOUT_OFF_CORE);
563 	}
564 #endif /* CONFIG_CPU_COUNTERS */
565 }
566 
567 #if CONFIG_THREAD_GROUPS
568 void
machine_thread_group_init(struct thread_group * tg)569 machine_thread_group_init(struct thread_group *tg)
570 {
571 	if (sched_perfcontrol_thread_group_init == sched_perfcontrol_thread_group_default) {
572 		return;
573 	}
574 	struct thread_group_data data;
575 	data.thread_group_id = thread_group_get_id(tg);
576 	data.thread_group_data = thread_group_get_machine_data(tg);
577 	data.thread_group_size = thread_group_machine_data_size();
578 	data.thread_group_flags = thread_group_get_flags(tg);
579 	sched_perfcontrol_thread_group_init(&data);
580 }
581 
582 void
machine_thread_group_deinit(struct thread_group * tg)583 machine_thread_group_deinit(struct thread_group *tg)
584 {
585 	if (sched_perfcontrol_thread_group_deinit == sched_perfcontrol_thread_group_default) {
586 		return;
587 	}
588 	struct thread_group_data data;
589 	data.thread_group_id = thread_group_get_id(tg);
590 	data.thread_group_data = thread_group_get_machine_data(tg);
591 	data.thread_group_size = thread_group_machine_data_size();
592 	data.thread_group_flags = thread_group_get_flags(tg);
593 	sched_perfcontrol_thread_group_deinit(&data);
594 }
595 
596 void
machine_thread_group_flags_update(struct thread_group * tg,uint32_t flags)597 machine_thread_group_flags_update(struct thread_group *tg, uint32_t flags)
598 {
599 	if (sched_perfcontrol_thread_group_flags_update == sched_perfcontrol_thread_group_default) {
600 		return;
601 	}
602 	struct thread_group_data data;
603 	data.thread_group_id = thread_group_get_id(tg);
604 	data.thread_group_data = thread_group_get_machine_data(tg);
605 	data.thread_group_size = thread_group_machine_data_size();
606 	data.thread_group_flags = flags;
607 	sched_perfcontrol_thread_group_flags_update(&data);
608 }
609 
610 void
machine_thread_group_blocked(struct thread_group * blocked_tg,struct thread_group * blocking_tg,uint32_t flags,thread_t blocked_thread)611 machine_thread_group_blocked(struct thread_group *blocked_tg,
612     struct thread_group *blocking_tg,
613     uint32_t flags,
614     thread_t blocked_thread)
615 {
616 	if (sched_perfcontrol_thread_group_blocked == sched_perfcontrol_thread_group_blocked_default) {
617 		return;
618 	}
619 
620 	spl_t s = splsched();
621 
622 	perfcontrol_state_t state = FIND_PERFCONTROL_STATE(blocked_thread);
623 	struct thread_group_data blocked_data;
624 	assert(blocked_tg != NULL);
625 
626 	blocked_data.thread_group_id = thread_group_get_id(blocked_tg);
627 	blocked_data.thread_group_data = thread_group_get_machine_data(blocked_tg);
628 	blocked_data.thread_group_size = thread_group_machine_data_size();
629 
630 	if (blocking_tg == NULL) {
631 		/*
632 		 * For special cases such as the render server, the blocking TG is a
633 		 * well known TG. Only in that case, the blocking_tg should be NULL.
634 		 */
635 		assert(flags & PERFCONTROL_CALLOUT_BLOCKING_TG_RENDER_SERVER);
636 		sched_perfcontrol_thread_group_blocked(&blocked_data, NULL, flags, state);
637 	} else {
638 		struct thread_group_data blocking_data;
639 		blocking_data.thread_group_id = thread_group_get_id(blocking_tg);
640 		blocking_data.thread_group_data = thread_group_get_machine_data(blocking_tg);
641 		blocking_data.thread_group_size = thread_group_machine_data_size();
642 		sched_perfcontrol_thread_group_blocked(&blocked_data, &blocking_data, flags, state);
643 	}
644 	KDBG(MACHDBG_CODE(DBG_MACH_THREAD_GROUP, MACH_THREAD_GROUP_BLOCK) | DBG_FUNC_START,
645 	    thread_tid(blocked_thread), thread_group_get_id(blocked_tg),
646 	    blocking_tg ? thread_group_get_id(blocking_tg) : THREAD_GROUP_INVALID,
647 	    flags);
648 
649 	splx(s);
650 }
651 
652 void
machine_thread_group_unblocked(struct thread_group * unblocked_tg,struct thread_group * unblocking_tg,uint32_t flags,thread_t unblocked_thread)653 machine_thread_group_unblocked(struct thread_group *unblocked_tg,
654     struct thread_group *unblocking_tg,
655     uint32_t flags,
656     thread_t unblocked_thread)
657 {
658 	if (sched_perfcontrol_thread_group_unblocked == sched_perfcontrol_thread_group_unblocked_default) {
659 		return;
660 	}
661 
662 	spl_t s = splsched();
663 
664 	perfcontrol_state_t state = FIND_PERFCONTROL_STATE(unblocked_thread);
665 	struct thread_group_data unblocked_data;
666 	assert(unblocked_tg != NULL);
667 
668 	unblocked_data.thread_group_id = thread_group_get_id(unblocked_tg);
669 	unblocked_data.thread_group_data = thread_group_get_machine_data(unblocked_tg);
670 	unblocked_data.thread_group_size = thread_group_machine_data_size();
671 
672 	if (unblocking_tg == NULL) {
673 		/*
674 		 * For special cases such as the render server, the unblocking TG is a
675 		 * well known TG. Only in that case, the unblocking_tg should be NULL.
676 		 */
677 		assert(flags & PERFCONTROL_CALLOUT_BLOCKING_TG_RENDER_SERVER);
678 		sched_perfcontrol_thread_group_unblocked(&unblocked_data, NULL, flags, state);
679 	} else {
680 		struct thread_group_data unblocking_data;
681 		unblocking_data.thread_group_id = thread_group_get_id(unblocking_tg);
682 		unblocking_data.thread_group_data = thread_group_get_machine_data(unblocking_tg);
683 		unblocking_data.thread_group_size = thread_group_machine_data_size();
684 		sched_perfcontrol_thread_group_unblocked(&unblocked_data, &unblocking_data, flags, state);
685 	}
686 	KDBG(MACHDBG_CODE(DBG_MACH_THREAD_GROUP, MACH_THREAD_GROUP_BLOCK) | DBG_FUNC_END,
687 	    thread_tid(unblocked_thread), thread_group_get_id(unblocked_tg),
688 	    unblocking_tg ? thread_group_get_id(unblocking_tg) : THREAD_GROUP_INVALID,
689 	    flags);
690 
691 	splx(s);
692 }
693 
694 #endif /* CONFIG_THREAD_GROUPS */
695 
696 void
machine_max_runnable_latency(uint64_t bg_max_latency,uint64_t default_max_latency,uint64_t realtime_max_latency)697 machine_max_runnable_latency(uint64_t bg_max_latency,
698     uint64_t default_max_latency,
699     uint64_t realtime_max_latency)
700 {
701 	if (sched_perfcontrol_max_runnable_latency == sched_perfcontrol_max_runnable_latency_default) {
702 		return;
703 	}
704 	struct perfcontrol_max_runnable_latency latencies = {
705 		.max_scheduling_latencies = {
706 			[THREAD_URGENCY_NONE] = 0,
707 			[THREAD_URGENCY_BACKGROUND] = bg_max_latency,
708 			[THREAD_URGENCY_NORMAL] = default_max_latency,
709 			[THREAD_URGENCY_REAL_TIME] = realtime_max_latency
710 		}
711 	};
712 
713 	sched_perfcontrol_max_runnable_latency(&latencies);
714 }
715 
716 void
machine_work_interval_notify(thread_t thread,struct kern_work_interval_args * kwi_args)717 machine_work_interval_notify(thread_t thread,
718     struct kern_work_interval_args* kwi_args)
719 {
720 	if (sched_perfcontrol_work_interval_notify == sched_perfcontrol_work_interval_notify_default) {
721 		return;
722 	}
723 	perfcontrol_state_t state = FIND_PERFCONTROL_STATE(thread);
724 	struct perfcontrol_work_interval work_interval = {
725 		.thread_id      = thread->thread_id,
726 		.qos_class      = (uint16_t)proc_get_effective_thread_policy(thread, TASK_POLICY_QOS),
727 		.urgency        = kwi_args->urgency,
728 		.flags          = kwi_args->notify_flags,
729 		.work_interval_id = kwi_args->work_interval_id,
730 		.start          = kwi_args->start,
731 		.finish         = kwi_args->finish,
732 		.deadline       = kwi_args->deadline,
733 		.next_start     = kwi_args->next_start,
734 		.create_flags   = kwi_args->create_flags,
735 	};
736 #if CONFIG_THREAD_GROUPS
737 	struct thread_group *tg;
738 	tg = thread_group_get(thread);
739 	work_interval.thread_group_id = thread_group_get_id(tg);
740 	work_interval.thread_group_data = thread_group_get_machine_data(tg);
741 #endif
742 	sched_perfcontrol_work_interval_notify(state, &work_interval);
743 }
744 
745 
746 void
machine_perfcontrol_deadline_passed(uint64_t deadline)747 machine_perfcontrol_deadline_passed(uint64_t deadline)
748 {
749 	if (sched_perfcontrol_deadline_passed != sched_perfcontrol_deadline_passed_default) {
750 		sched_perfcontrol_deadline_passed(deadline);
751 	}
752 }
753 
754 #if SCHED_HYGIENE_DEBUG
755 
756 __options_decl(int_mask_hygiene_flags_t, uint8_t, {
757 	INT_MASK_BASE = 0x00,
758 	INT_MASK_FROM_HANDLER = 0x01,
759 	INT_MASK_IS_STACKSHOT = 0x02,
760 });
761 
762 /*
763  * ml_spin_debug_reset()
764  * Reset the timestamp on a thread that has been unscheduled
765  * to avoid false alarms. Alarm will go off if interrupts are held
766  * disabled for too long, starting from now.
767  *
768  * Call ml_get_timebase() directly to prevent extra overhead on newer
769  * platforms that's enabled in DEVELOPMENT kernel configurations.
770  */
771 void
ml_spin_debug_reset(thread_t thread)772 ml_spin_debug_reset(thread_t thread)
773 {
774 	if (thread->machine.intmask_timestamp) {
775 		thread->machine.intmask_timestamp = ml_get_sched_hygiene_timebase();
776 		INTERRUPT_MASKED_DEBUG_CAPTURE_PMC(thread);
777 	}
778 }
779 
780 /*
781  * ml_spin_debug_clear()
782  * Clear the timestamp and cycle/instruction counts on a thread that
783  * has been unscheduled to avoid false alarms
784  */
785 void
ml_spin_debug_clear(thread_t thread)786 ml_spin_debug_clear(thread_t thread)
787 {
788 	thread->machine.intmask_timestamp = 0;
789 	thread->machine.intmask_cycles = 0;
790 	thread->machine.intmask_instr = 0;
791 }
792 
793 /*
794  * ml_spin_debug_clear_self()
795  * Clear the timestamp on the current thread to prevent
796  * false alarms
797  */
798 void
ml_spin_debug_clear_self(void)799 ml_spin_debug_clear_self(void)
800 {
801 	ml_spin_debug_clear(current_thread());
802 }
803 
804 #ifndef KASAN
805 
806 /*
807  * Get a character representing the provided thread's kind of CPU.
808  */
809 #if !CONFIG_CPU_COUNTERS
810 __unused
811 #endif // !CONFIG_CPU_COUNTERS
812 static char
__ml_interrupts_disabled_cpu_kind(thread_t thread)813 __ml_interrupts_disabled_cpu_kind(thread_t thread)
814 {
815 #if __AMP__
816 	processor_t processor = thread->last_processor;
817 	if (!processor) {
818 		return '!';
819 	}
820 	switch (processor->processor_set->pset_cluster_type) {
821 	case PSET_AMP_P:
822 		return 'P';
823 	case PSET_AMP_E:
824 		return 'E';
825 	default:
826 		return '?';
827 	}
828 #else // __AMP__
829 #pragma unused(thread)
830 	return '-';
831 #endif // !__AMP__
832 }
833 
834 #define EXTRA_INFO_STRING_SIZE 256
835 #define LOW_FREQ_THRESHOLD_MHZ 500
836 #define HIGH_CPI_THRESHOLD     3
837 
838 static void
__ml_trigger_interrupts_disabled_handle(thread_t thread,uint64_t start,uint64_t now,uint64_t timeout,int_mask_hygiene_flags_t flags)839 __ml_trigger_interrupts_disabled_handle(thread_t thread, uint64_t start, uint64_t now, uint64_t timeout, int_mask_hygiene_flags_t flags)
840 {
841 	mach_timebase_info_data_t timebase;
842 	clock_timebase_info(&timebase);
843 	bool is_int_handler = flags & INT_MASK_FROM_HANDLER;
844 	bool is_stackshot = flags & INT_MASK_IS_STACKSHOT;
845 
846 	const uint64_t time_elapsed = now - start;
847 	const uint64_t time_elapsed_ns = (time_elapsed * timebase.numer) / timebase.denom;
848 
849 #if __AMP__
850 	if (is_stackshot && interrupt_masked_debug_mode == SCHED_HYGIENE_MODE_PANIC) {
851 		/*
852 		 * If there are no recommended performance cores, we double the timeout to compensate
853 		 * for the difference in time it takes Stackshot to run on efficiency cores, and then
854 		 * recheck if we still exceeded the adjusted timeout.
855 		 */
856 		int cpu;
857 		int max_cpu;
858 
859 		max_cpu = ml_get_max_cpu_number();
860 		for (cpu = 0; cpu <= max_cpu; cpu++) {
861 			processor_t processor = cpu_to_processor(cpu);
862 			if (processor->is_recommended &&
863 			    processor->processor_set->pset_cluster_type == PSET_AMP_P) {
864 				break;
865 			}
866 		}
867 		if (cpu > max_cpu) {
868 			if (time_elapsed < timeout * 2) {
869 				return;
870 			}
871 		}
872 	}
873 #endif /* __AMP__ */
874 
875 	uint64_t current_cycles = 0, current_instrs = 0;
876 
877 #if CONFIG_CPU_COUNTERS
878 	if (static_if(sched_debug_pmc)) {
879 		mt_cur_cpu_cycles_instrs_speculative(&current_cycles, &current_instrs);
880 	}
881 #endif // CONFIG_CPU_COUNTERS
882 
883 	const uint64_t cycles_elapsed = current_cycles - thread->machine.intmask_cycles;
884 	const uint64_t instrs_elapsed = current_instrs - thread->machine.intmask_instr;
885 
886 	if (interrupt_masked_debug_mode == SCHED_HYGIENE_MODE_PANIC) {
887 		const uint64_t timeout_ns = ((timeout * debug_cpu_performance_degradation_factor) * timebase.numer) / timebase.denom;
888 		char extra_info_string[EXTRA_INFO_STRING_SIZE] = { '\0' };
889 #if CONFIG_CPU_COUNTERS
890 		if (static_if(sched_debug_pmc)) {
891 			const uint64_t time_elapsed_us = time_elapsed_ns / 1000;
892 			const uint64_t average_freq_mhz = cycles_elapsed / time_elapsed_us;
893 			const uint64_t average_cpi_whole = cycles_elapsed / instrs_elapsed;
894 			const uint64_t average_cpi_fractional = ((cycles_elapsed * 100) / instrs_elapsed) % 100;
895 			bool high_cpi = average_cpi_whole >= HIGH_CPI_THRESHOLD;
896 			char core_kind = __ml_interrupts_disabled_cpu_kind(thread);
897 			bool low_mhz = average_freq_mhz < LOW_FREQ_THRESHOLD_MHZ;
898 
899 			snprintf(extra_info_string, EXTRA_INFO_STRING_SIZE,
900 			    ", %sfreq = %llu MHz, %sCPI = %llu.%llu, CPU kind = %c",
901 			    low_mhz ? "low " : "",
902 			    average_freq_mhz,
903 			    high_cpi ? "high " : "",
904 			    average_cpi_whole,
905 			    average_cpi_fractional,
906 			    core_kind);
907 		}
908 #endif // CONFIG_CPU_COUNTERS
909 
910 		if (is_int_handler) {
911 			panic("Processing of an interrupt (type = %u, handler address = %p, vector = %p) "
912 			    "took %llu nanoseconds (start = %llu, now = %llu, timeout = %llu ns%s)",
913 			    thread->machine.int_type, (void *)thread->machine.int_handler_addr, (void *)thread->machine.int_vector,
914 			    time_elapsed_ns, start, now, timeout_ns, extra_info_string);
915 		} else {
916 			panic("%s for %llu nanoseconds (start = %llu, now = %llu, timeout = %llu ns%s)",
917 			    is_stackshot ? "Stackshot disabled interrupts" : "Interrupts held disabled",
918 			    time_elapsed_ns, start, now, timeout_ns, extra_info_string);
919 		}
920 	} else if (interrupt_masked_debug_mode == SCHED_HYGIENE_MODE_TRACE) {
921 		if (is_int_handler) {
922 			static const uint32_t interrupt_handled_dbgid =
923 			    MACHDBG_CODE(DBG_MACH_SCHED, MACH_INT_HANDLED_EXPIRED);
924 			DTRACE_SCHED3(interrupt_handled_dbgid, uint64_t, time_elapsed,
925 			    uint64_t, cycles_elapsed, uint64_t, instrs_elapsed);
926 			KDBG(interrupt_handled_dbgid, time_elapsed,
927 			    cycles_elapsed, instrs_elapsed);
928 		} else {
929 			static const uint32_t interrupt_masked_dbgid =
930 			    MACHDBG_CODE(DBG_MACH_SCHED, MACH_INT_MASKED_EXPIRED);
931 			DTRACE_SCHED3(interrupt_masked_dbgid, uint64_t, time_elapsed,
932 			    uint64_t, cycles_elapsed, uint64_t, instrs_elapsed);
933 			KDBG(interrupt_masked_dbgid, time_elapsed,
934 			    cycles_elapsed, instrs_elapsed);
935 		}
936 	}
937 }
938 #endif // !defined(KASAN)
939 
940 static inline void
__ml_handle_interrupts_disabled_duration(thread_t thread,uint64_t timeout,bool is_int_handler)941 __ml_handle_interrupts_disabled_duration(thread_t thread, uint64_t timeout, bool is_int_handler)
942 {
943 	if (timeout == 0) {
944 		return; // 0 means timeout disabled.
945 	}
946 	uint64_t start = is_int_handler ? thread->machine.inthandler_timestamp : thread->machine.intmask_timestamp;
947 	if (start != 0) {
948 		uint64_t now = ml_get_sched_hygiene_timebase();
949 
950 		if (interrupt_masked_debug_mode &&
951 		    ((now - start) > timeout * debug_cpu_performance_degradation_factor) &&
952 		    !thread->machine.inthandler_abandon) {
953 			/*
954 			 * Disable the actual panic for KASAN due to the overhead of KASAN itself, leave the rest of the
955 			 * mechanism enabled so that KASAN can catch any bugs in the mechanism itself.
956 			 */
957 #ifndef KASAN
958 			__ml_trigger_interrupts_disabled_handle(thread, start, now, timeout, is_int_handler);
959 #endif
960 		}
961 
962 		if (is_int_handler) {
963 			uint64_t const duration = now - start;
964 			/*
965 			 * No need for an atomic add, the only thread modifying
966 			 * this is ourselves. Other threads querying will just see
967 			 * either the old or the new value. (This will also just
968 			 * resolve to regular loads and stores on relevant
969 			 * platforms.)
970 			 */
971 			uint64_t const old_duration = os_atomic_load_wide(&thread->machine.int_time_mt, relaxed);
972 			os_atomic_store_wide(&thread->machine.int_time_mt, old_duration + duration, relaxed);
973 		}
974 	}
975 }
976 
977 void
ml_handle_interrupts_disabled_duration(thread_t thread)978 ml_handle_interrupts_disabled_duration(thread_t thread)
979 {
980 	__ml_handle_interrupts_disabled_duration(thread, os_atomic_load(&interrupt_masked_timeout, relaxed), INT_MASK_BASE);
981 }
982 
983 void
ml_handle_stackshot_interrupt_disabled_duration(thread_t thread)984 ml_handle_stackshot_interrupt_disabled_duration(thread_t thread)
985 {
986 	/* Use MAX() to let the user bump the timeout further if needed */
987 	uint64_t stackshot_timeout = os_atomic_load(&stackshot_interrupt_masked_timeout, relaxed);
988 	uint64_t normal_timeout = os_atomic_load(&interrupt_masked_timeout, relaxed);
989 	uint64_t timeout = MAX(stackshot_timeout, normal_timeout);
990 	__ml_handle_interrupts_disabled_duration(thread, timeout, INT_MASK_IS_STACKSHOT);
991 }
992 
993 void
ml_handle_interrupt_handler_duration(thread_t thread)994 ml_handle_interrupt_handler_duration(thread_t thread)
995 {
996 	__ml_handle_interrupts_disabled_duration(thread, os_atomic_load(&interrupt_masked_timeout, relaxed), INT_MASK_FROM_HANDLER);
997 }
998 
999 void
ml_irq_debug_start(uintptr_t handler,uintptr_t vector)1000 ml_irq_debug_start(uintptr_t handler, uintptr_t vector)
1001 {
1002 	INTERRUPT_MASKED_DEBUG_START(handler, DBG_INTR_TYPE_OTHER);
1003 	current_thread()->machine.int_vector = (uintptr_t)VM_KERNEL_STRIP_PTR(vector);
1004 }
1005 
1006 void
ml_irq_debug_end()1007 ml_irq_debug_end()
1008 {
1009 	INTERRUPT_MASKED_DEBUG_END();
1010 }
1011 
1012 /*
1013  * Abandon a potential timeout when handling an interrupt. It is important to
1014  * continue to keep track of the interrupt time so the time-stamp can't be
1015  * reset. (Interrupt time is subtracted from preemption time to maintain
1016  * accurate preemption time measurement).
1017  * When `inthandler_abandon` is true, a timeout will be ignored when the
1018  * interrupt handler finishes.
1019  */
1020 void
ml_irq_debug_abandon(void)1021 ml_irq_debug_abandon(void)
1022 {
1023 	assert(!ml_get_interrupts_enabled());
1024 
1025 	thread_t t = current_thread();
1026 	if (t->machine.inthandler_timestamp != 0) {
1027 		t->machine.inthandler_abandon = true;
1028 	}
1029 }
1030 #endif // SCHED_HYGIENE_DEBUG
1031 
1032 #if SCHED_HYGIENE_DEBUG
1033 __attribute__((noinline))
1034 static void
ml_interrupt_masked_debug_timestamp(thread_t thread)1035 ml_interrupt_masked_debug_timestamp(thread_t thread)
1036 {
1037 	thread->machine.intmask_timestamp = ml_get_sched_hygiene_timebase();
1038 	INTERRUPT_MASKED_DEBUG_CAPTURE_PMC(thread);
1039 }
1040 #endif
1041 
1042 boolean_t
ml_set_interrupts_enabled_with_debug(boolean_t enable,boolean_t __unused debug)1043 ml_set_interrupts_enabled_with_debug(boolean_t enable, boolean_t __unused debug)
1044 {
1045 	thread_t        thread;
1046 	uint64_t        state;
1047 
1048 	thread = current_thread();
1049 
1050 	state = __builtin_arm_rsr("DAIF");
1051 
1052 	if (__improbable(!(state & DAIF_DEBUGF))) {
1053 		panic("%s: debug exceptions enabled in kernel mode", __func__);
1054 	}
1055 	if (enable && (state & DAIF_STANDARD_DISABLE)) {
1056 		assert3u(state & DAIF_STANDARD_DISABLE, ==, DAIF_STANDARD_DISABLE);
1057 		assert(getCpuDatap()->cpu_int_state == NULL); // Make sure we're not enabling interrupts from primary interrupt context
1058 #if SCHED_HYGIENE_DEBUG
1059 		if (__probable(debug && static_if(sched_debug_interrupt_disable))) {
1060 			// Interrupts are currently masked, we will enable them (after finishing this check)
1061 			if (stackshot_active()) {
1062 				ml_handle_stackshot_interrupt_disabled_duration(thread);
1063 			} else {
1064 				ml_handle_interrupts_disabled_duration(thread);
1065 			}
1066 			thread->machine.intmask_timestamp = 0;
1067 			thread->machine.intmask_cycles = 0;
1068 			thread->machine.intmask_instr = 0;
1069 		}
1070 #endif  // SCHED_HYGIENE_DEBUG
1071 		if (get_preemption_level() == 0) {
1072 			while (thread->machine.CpuDatap->cpu_pending_ast & AST_URGENT) {
1073 #if __ARM_USER_PROTECT__
1074 				uintptr_t up = arm_user_protect_begin(thread);
1075 #endif
1076 				ast_taken_kernel();
1077 #if __ARM_USER_PROTECT__
1078 				arm_user_protect_end(thread, up, FALSE);
1079 #endif
1080 			}
1081 		}
1082 		__builtin_arm_wsr("DAIFClr", DAIFSC_STANDARD_DISABLE);
1083 	} else if (!enable && ((state & DAIF_STANDARD_DISABLE) != DAIF_STANDARD_DISABLE)) {
1084 		assert3u(state & DAIF_STANDARD_DISABLE, ==, 0);
1085 		__builtin_arm_wsr("DAIFSet", DAIFSC_STANDARD_DISABLE);
1086 
1087 #if SCHED_HYGIENE_DEBUG
1088 		if (__probable(debug && static_if(sched_debug_interrupt_disable))) {
1089 			// Interrupts were enabled, we just masked them
1090 			ml_interrupt_masked_debug_timestamp(thread);
1091 		}
1092 #endif
1093 	}
1094 	return (state & DAIF_STANDARD_DISABLE) != DAIF_STANDARD_DISABLE;
1095 }
1096 
1097 boolean_t
ml_set_interrupts_enabled(boolean_t enable)1098 ml_set_interrupts_enabled(boolean_t enable)
1099 {
1100 	return ml_set_interrupts_enabled_with_debug(enable, true);
1101 }
1102 
1103 boolean_t
ml_early_set_interrupts_enabled(boolean_t enable)1104 ml_early_set_interrupts_enabled(boolean_t enable)
1105 {
1106 	return ml_set_interrupts_enabled(enable);
1107 }
1108 
1109 /*
1110  * Interrupt enable function exported for AppleCLPC without
1111  * measurements enabled.
1112  *
1113  * Only for AppleCLPC!
1114  */
1115 boolean_t
sched_perfcontrol_ml_set_interrupts_without_measurement(boolean_t enable)1116 sched_perfcontrol_ml_set_interrupts_without_measurement(boolean_t enable)
1117 {
1118 	return ml_set_interrupts_enabled_with_debug(enable, false);
1119 }
1120 
1121 /*
1122  *	Routine:        ml_at_interrupt_context
1123  *	Function:	Check if running at interrupt context
1124  */
1125 boolean_t
ml_at_interrupt_context(void)1126 ml_at_interrupt_context(void)
1127 {
1128 	/* Do not use a stack-based check here, as the top-level exception handler
1129 	 * is free to use some other stack besides the per-CPU interrupt stack.
1130 	 * Interrupts should always be disabled if we're at interrupt context.
1131 	 * Check that first, as we may be in a preemptible non-interrupt context, in
1132 	 * which case we could be migrated to a different CPU between obtaining
1133 	 * the per-cpu data pointer and loading cpu_int_state.  We then might end
1134 	 * up checking the interrupt state of a different CPU, resulting in a false
1135 	 * positive.  But if interrupts are disabled, we also know we cannot be
1136 	 * preempted. */
1137 	return !ml_get_interrupts_enabled() && (getCpuDatap()->cpu_int_state != NULL);
1138 }
1139 
1140 /*
1141  * This answers the question
1142  * "after returning from this interrupt handler with the AST_URGENT bit set,
1143  * will I end up in ast_taken_user or ast_taken_kernel?"
1144  *
1145  * If it's called in non-interrupt context (e.g. regular syscall), it should
1146  * return false.
1147  *
1148  * Must be called with interrupts disabled.
1149  */
1150 bool
ml_did_interrupt_userspace(void)1151 ml_did_interrupt_userspace(void)
1152 {
1153 	assert(ml_get_interrupts_enabled() == false);
1154 
1155 	struct arm_saved_state *state = getCpuDatap()->cpu_int_state;
1156 
1157 	return state && PSR64_IS_USER(get_saved_state_cpsr(state));
1158 }
1159 
1160 
1161 vm_offset_t
ml_stack_remaining(void)1162 ml_stack_remaining(void)
1163 {
1164 	uintptr_t local = (uintptr_t) &local;
1165 	vm_offset_t     intstack_top_ptr;
1166 
1167 	/* Since this is a stack-based check, we don't need to worry about
1168 	 * preemption as we do in ml_at_interrupt_context().  If we are preemptible,
1169 	 * then the sp should never be within any CPU's interrupt stack unless
1170 	 * something has gone horribly wrong. */
1171 	intstack_top_ptr = getCpuDatap()->intstack_top;
1172 	if ((local < intstack_top_ptr) && (local > intstack_top_ptr - INTSTACK_SIZE)) {
1173 		return local - (getCpuDatap()->intstack_top - INTSTACK_SIZE);
1174 	} else {
1175 		return local - current_thread()->kernel_stack;
1176 	}
1177 }
1178 
1179 static boolean_t ml_quiescing = FALSE;
1180 
1181 void
ml_set_is_quiescing(boolean_t quiescing)1182 ml_set_is_quiescing(boolean_t quiescing)
1183 {
1184 	assert(ml_quiescing != quiescing);
1185 	ml_quiescing = quiescing;
1186 	os_atomic_thread_fence(release);
1187 }
1188 
1189 boolean_t
ml_is_quiescing(void)1190 ml_is_quiescing(void)
1191 {
1192 	os_atomic_thread_fence(acquire);
1193 	return ml_quiescing;
1194 }
1195 
1196 uint64_t
ml_get_booter_memory_size(void)1197 ml_get_booter_memory_size(void)
1198 {
1199 #if CONFIG_SPTM
1200 	extern uint64_t memSize;
1201 #endif /* CONFIG_SPTM */
1202 	uint64_t size;
1203 	uint64_t roundsize = 512 * 1024 * 1024ULL;
1204 	size = BootArgs->memSizeActual;
1205 	if (!size) {
1206 #if CONFIG_SPTM
1207 		/*
1208 		 * SPTM systems cache [memSize] in a CTRR-protected variable rather
1209 		 * than relying on [BootArgs]. This is to enable the possibility
1210 		 * for XNU to modify it before machine lockdown, which happens in
1211 		 * KASAN kernels. If we did not do this, XNU would fault on the first
1212 		 * attempt to overwrite [BootArgs->memSize].
1213 		 */
1214 		size = memSize;
1215 #else
1216 		size = BootArgs->memSize;
1217 #endif /* CONFIG_SPTM */
1218 		if (size < (2 * roundsize)) {
1219 			roundsize >>= 1;
1220 		}
1221 		size  = (size + roundsize - 1) & ~(roundsize - 1);
1222 	}
1223 
1224 #if CONFIG_SPTM
1225 	size -= memSize;
1226 #else
1227 	size -= BootArgs->memSize;
1228 #endif /* CONFIG_SPTM */
1229 
1230 	return size;
1231 }
1232 
1233 uint64_t
ml_get_abstime_offset(void)1234 ml_get_abstime_offset(void)
1235 {
1236 	return rtclock_base_abstime;
1237 }
1238 
1239 uint64_t
ml_get_conttime_offset(void)1240 ml_get_conttime_offset(void)
1241 {
1242 #if HIBERNATION && HAS_CONTINUOUS_HWCLOCK
1243 	return hwclock_conttime_offset;
1244 #elif HAS_CONTINUOUS_HWCLOCK
1245 	return 0;
1246 #else
1247 	return rtclock_base_abstime + mach_absolutetime_asleep;
1248 #endif
1249 }
1250 
1251 uint64_t
ml_get_time_since_reset(void)1252 ml_get_time_since_reset(void)
1253 {
1254 #if HAS_CONTINUOUS_HWCLOCK
1255 	if (wake_conttime == UINT64_MAX) {
1256 		return UINT64_MAX;
1257 	} else {
1258 		return mach_continuous_time() - wake_conttime;
1259 	}
1260 #else
1261 	/* The timebase resets across S2R, so just return the raw value. */
1262 	return ml_get_hwclock();
1263 #endif
1264 }
1265 
1266 void
ml_set_reset_time(__unused uint64_t wake_time)1267 ml_set_reset_time(__unused uint64_t wake_time)
1268 {
1269 #if HAS_CONTINUOUS_HWCLOCK
1270 	wake_conttime = wake_time;
1271 #endif
1272 }
1273 
1274 uint64_t
ml_get_conttime_wake_time(void)1275 ml_get_conttime_wake_time(void)
1276 {
1277 #if HAS_CONTINUOUS_HWCLOCK
1278 	/*
1279 	 * For now, we will reconstitute the timebase value from
1280 	 * cpu_timebase_init and use it as the wake time.
1281 	 */
1282 	return wake_abstime - ml_get_abstime_offset();
1283 #else /* HAS_CONTINOUS_HWCLOCK */
1284 	/* The wake time is simply our continuous time offset. */
1285 	return ml_get_conttime_offset();
1286 #endif /* HAS_CONTINOUS_HWCLOCK */
1287 }
1288 
1289 /*
1290  * ml_snoop_thread_is_on_core(thread_t thread)
1291  * Check if the given thread is currently on core.  This function does not take
1292  * locks, disable preemption, or otherwise guarantee synchronization.  The
1293  * result should be considered advisory.
1294  */
1295 bool
ml_snoop_thread_is_on_core(thread_t thread)1296 ml_snoop_thread_is_on_core(thread_t thread)
1297 {
1298 	unsigned int cur_cpu_num = 0;
1299 	const unsigned int max_cpu_id = ml_get_max_cpu_number();
1300 
1301 	for (cur_cpu_num = 0; cur_cpu_num <= max_cpu_id; cur_cpu_num++) {
1302 		if (CpuDataEntries[cur_cpu_num].cpu_data_vaddr) {
1303 			if (CpuDataEntries[cur_cpu_num].cpu_data_vaddr->cpu_active_thread == thread) {
1304 				return true;
1305 			}
1306 		}
1307 	}
1308 
1309 	return false;
1310 }
1311 
1312 int
ml_early_cpu_max_number(void)1313 ml_early_cpu_max_number(void)
1314 {
1315 	assert(startup_phase >= STARTUP_SUB_TUNABLES);
1316 	return ml_get_max_cpu_number();
1317 }
1318 
1319 void
ml_set_max_cpus(unsigned int max_cpus __unused)1320 ml_set_max_cpus(unsigned int max_cpus __unused)
1321 {
1322 	lck_mtx_lock(&max_cpus_lock);
1323 	if (max_cpus_initialized != MAX_CPUS_SET) {
1324 		if (max_cpus_initialized == MAX_CPUS_WAIT) {
1325 			thread_wakeup((event_t) &max_cpus_initialized);
1326 		}
1327 		max_cpus_initialized = MAX_CPUS_SET;
1328 	}
1329 	lck_mtx_unlock(&max_cpus_lock);
1330 }
1331 
1332 unsigned int
ml_wait_max_cpus(void)1333 ml_wait_max_cpus(void)
1334 {
1335 	assert(lockdown_done);
1336 	lck_mtx_lock(&max_cpus_lock);
1337 	while (max_cpus_initialized != MAX_CPUS_SET) {
1338 		max_cpus_initialized = MAX_CPUS_WAIT;
1339 		lck_mtx_sleep(&max_cpus_lock, LCK_SLEEP_DEFAULT, &max_cpus_initialized, THREAD_UNINT);
1340 	}
1341 	lck_mtx_unlock(&max_cpus_lock);
1342 	return machine_info.max_cpus;
1343 }
1344 
1345 void
ml_cpu_get_info_type(ml_cpu_info_t * ml_cpu_info,cluster_type_t cluster_type)1346 ml_cpu_get_info_type(ml_cpu_info_t * ml_cpu_info, cluster_type_t cluster_type)
1347 {
1348 	cache_info_t   *cpuid_cache_info;
1349 
1350 	cpuid_cache_info = cache_info_type(cluster_type);
1351 	ml_cpu_info->vector_unit = 0;
1352 	ml_cpu_info->cache_line_size = cpuid_cache_info->c_linesz;
1353 	ml_cpu_info->l1_icache_size = cpuid_cache_info->c_isize;
1354 	ml_cpu_info->l1_dcache_size = cpuid_cache_info->c_dsize;
1355 
1356 #if (__ARM_ARCH__ >= 8)
1357 	ml_cpu_info->l2_settings = 1;
1358 	ml_cpu_info->l2_cache_size = cpuid_cache_info->c_l2size;
1359 #else
1360 #error Unsupported arch
1361 #endif
1362 	ml_cpu_info->l3_settings = 0;
1363 	ml_cpu_info->l3_cache_size = 0xFFFFFFFF;
1364 }
1365 
1366 /*
1367  *	Routine:        ml_cpu_get_info
1368  *	Function: Fill out the ml_cpu_info_t structure with parameters associated
1369  *	with the boot cluster.
1370  */
1371 void
ml_cpu_get_info(ml_cpu_info_t * ml_cpu_info)1372 ml_cpu_get_info(ml_cpu_info_t * ml_cpu_info)
1373 {
1374 	ml_cpu_get_info_type(ml_cpu_info, ml_get_topology_info()->boot_cpu->cluster_type);
1375 }
1376 
1377 unsigned int
ml_get_cpu_number_type(cluster_type_t cluster_type,bool logical,bool available)1378 ml_get_cpu_number_type(cluster_type_t cluster_type, bool logical, bool available)
1379 {
1380 	/*
1381 	 * At present no supported ARM system features SMT, so the "logical"
1382 	 * parameter doesn't have an impact on the result.
1383 	 */
1384 	if (logical && available) {
1385 		return os_atomic_load(&cluster_type_num_active_cpus[cluster_type], relaxed);
1386 	} else if (logical && !available) {
1387 		return ml_get_topology_info()->cluster_type_num_cpus[cluster_type];
1388 	} else if (!logical && available) {
1389 		return os_atomic_load(&cluster_type_num_active_cpus[cluster_type], relaxed);
1390 	} else {
1391 		return ml_get_topology_info()->cluster_type_num_cpus[cluster_type];
1392 	}
1393 }
1394 
1395 void
ml_get_cluster_type_name(cluster_type_t cluster_type,char * name,size_t name_size)1396 ml_get_cluster_type_name(cluster_type_t cluster_type, char *name, size_t name_size)
1397 {
1398 	strlcpy(name, cluster_type_names[cluster_type], name_size);
1399 }
1400 
1401 unsigned int
ml_get_cluster_number_type(cluster_type_t cluster_type)1402 ml_get_cluster_number_type(cluster_type_t cluster_type)
1403 {
1404 	return ml_get_topology_info()->cluster_type_num_clusters[cluster_type];
1405 }
1406 
1407 unsigned int
ml_cpu_cache_sharing(unsigned int level,cluster_type_t cluster_type,bool include_all_cpu_types __unused)1408 ml_cpu_cache_sharing(unsigned int level, cluster_type_t cluster_type, bool include_all_cpu_types __unused)
1409 {
1410 	unsigned int cpu_number = 0, cluster_types = 0;
1411 
1412 	/*
1413 	 * Level 0 corresponds to main memory, which is shared across all cores.
1414 	 */
1415 	if (level == 0) {
1416 		return ml_get_topology_info()->num_cpus;
1417 	}
1418 
1419 	/*
1420 	 * At present no supported ARM system features more than 2 levels of caches.
1421 	 */
1422 	if (level > 2) {
1423 		return 0;
1424 	}
1425 
1426 	/*
1427 	 * L1 caches are always per core.
1428 	 */
1429 	if (level == 1) {
1430 		return 1;
1431 	}
1432 
1433 	cluster_types = (1 << cluster_type);
1434 
1435 	/*
1436 	 * Traverse clusters until we find the one(s) of the desired type(s).
1437 	 */
1438 	for (int i = 0; i < ml_get_topology_info()->num_clusters; i++) {
1439 		ml_topology_cluster_t *cluster = &ml_get_topology_info()->clusters[i];
1440 		if ((1 << cluster->cluster_type) & cluster_types) {
1441 			cpu_number += cluster->num_cpus;
1442 			cluster_types &= ~(1 << cluster->cluster_type);
1443 			if (!cluster_types) {
1444 				break;
1445 			}
1446 		}
1447 	}
1448 
1449 	return cpu_number;
1450 }
1451 
1452 unsigned int
ml_get_cpu_types(void)1453 ml_get_cpu_types(void)
1454 {
1455 	return ml_get_topology_info()->cluster_types;
1456 }
1457 
1458 void
machine_conf(void)1459 machine_conf(void)
1460 {
1461 	/*
1462 	 * This is known to be inaccurate. mem_size should always be capped at 2 GB
1463 	 */
1464 	machine_info.memory_size = (uint32_t)mem_size;
1465 
1466 	// rdar://problem/58285685: Userland expects _COMM_PAGE_LOGICAL_CPUS to report
1467 	// (max_cpu_id+1) rather than a literal *count* of logical CPUs.
1468 	unsigned int num_cpus = ml_get_topology_info()->max_cpu_id + 1;
1469 	machine_info.max_cpus = num_cpus;
1470 	machine_info.physical_cpu_max = num_cpus;
1471 	machine_info.logical_cpu_max = num_cpus;
1472 }
1473 
1474 void
machine_init(void)1475 machine_init(void)
1476 {
1477 	debug_log_init();
1478 	clock_config();
1479 	is_clock_configured = TRUE;
1480 	if (debug_enabled) {
1481 		pmap_map_globals();
1482 	}
1483 	ml_lockdown_init();
1484 }
1485