xref: /linux/arch/x86/kernel/smpboot.c (revision 6a4aee277740d04ac0fd54cfa17cc28261932ddc)
1 // SPDX-License-Identifier: GPL-2.0-or-later
2  /*
3  *	x86 SMP booting functions
4  *
5  *	(c) 1995 Alan Cox, Building #3 <alan@lxorguk.ukuu.org.uk>
6  *	(c) 1998, 1999, 2000, 2009 Ingo Molnar <mingo@redhat.com>
7  *	Copyright 2001 Andi Kleen, SuSE Labs.
8  *
9  *	Much of the core SMP work is based on previous work by Thomas Radke, to
10  *	whom a great many thanks are extended.
11  *
12  *	Thanks to Intel for making available several different Pentium,
13  *	Pentium Pro and Pentium-II/Xeon MP machines.
14  *	Original development of Linux SMP code supported by Caldera.
15  *
16  *	Fixes
17  *		Felix Koop	:	NR_CPUS used properly
18  *		Jose Renau	:	Handle single CPU case.
19  *		Alan Cox	:	By repeated request 8) - Total BogoMIPS report.
20  *		Greg Wright	:	Fix for kernel stacks panic.
21  *		Erich Boleyn	:	MP v1.4 and additional changes.
22  *	Matthias Sattler	:	Changes for 2.1 kernel map.
23  *	Michel Lespinasse	:	Changes for 2.1 kernel map.
24  *	Michael Chastain	:	Change trampoline.S to gnu as.
25  *		Alan Cox	:	Dumb bug: 'B' step PPro's are fine
26  *		Ingo Molnar	:	Added APIC timers, based on code
27  *					from Jose Renau
28  *		Ingo Molnar	:	various cleanups and rewrites
29  *		Tigran Aivazian	:	fixed "0.00 in /proc/uptime on SMP" bug.
30  *	Maciej W. Rozycki	:	Bits for genuine 82489DX APICs
31  *	Andi Kleen		:	Changed for SMP boot into long mode.
32  *		Martin J. Bligh	: 	Added support for multi-quad systems
33  *		Dave Jones	:	Report invalid combinations of Athlon CPUs.
34  *		Rusty Russell	:	Hacked into shape for new "hotplug" boot process.
35  *      Andi Kleen              :       Converted to new state machine.
36  *	Ashok Raj		: 	CPU hotplug support
37  *	Glauber Costa		:	i386 and x86_64 integration
38  */
39 
40 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
41 
42 #include <linux/init.h>
43 #include <linux/smp.h>
44 #include <linux/export.h>
45 #include <linux/sched.h>
46 #include <linux/sched/topology.h>
47 #include <linux/sched/hotplug.h>
48 #include <linux/sched/task_stack.h>
49 #include <linux/percpu.h>
50 #include <linux/memblock.h>
51 #include <linux/err.h>
52 #include <linux/nmi.h>
53 #include <linux/tboot.h>
54 #include <linux/gfp.h>
55 #include <linux/cpuidle.h>
56 #include <linux/kexec.h>
57 #include <linux/numa.h>
58 #include <linux/pgtable.h>
59 #include <linux/overflow.h>
60 #include <linux/stackprotector.h>
61 #include <linux/cpuhotplug.h>
62 #include <linux/mc146818rtc.h>
63 
64 #include <asm/acpi.h>
65 #include <asm/cacheinfo.h>
66 #include <asm/desc.h>
67 #include <asm/nmi.h>
68 #include <asm/irq.h>
69 #include <asm/realmode.h>
70 #include <asm/cpu.h>
71 #include <asm/numa.h>
72 #include <asm/tlbflush.h>
73 #include <asm/mtrr.h>
74 #include <asm/mwait.h>
75 #include <asm/apic.h>
76 #include <asm/io_apic.h>
77 #include <asm/fpu/api.h>
78 #include <asm/setup.h>
79 #include <asm/uv/uv.h>
80 #include <asm/microcode.h>
81 #include <asm/i8259.h>
82 #include <asm/misc.h>
83 #include <asm/qspinlock.h>
84 #include <asm/intel-family.h>
85 #include <asm/cpu_device_id.h>
86 #include <asm/spec-ctrl.h>
87 #include <asm/hw_irq.h>
88 #include <asm/stackprotector.h>
89 #include <asm/sev.h>
90 #include <asm/spec-ctrl.h>
91 
92 /* representing HT siblings of each logical CPU */
93 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
94 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
95 
96 /* representing HT and core siblings of each logical CPU */
97 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
98 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
99 
100 /* representing HT, core, and die siblings of each logical CPU */
101 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_die_map);
102 EXPORT_PER_CPU_SYMBOL(cpu_die_map);
103 
104 /* CPUs which are the primary SMT threads */
105 struct cpumask __cpu_primary_thread_mask __read_mostly;
106 
107 /* Representing CPUs for which sibling maps can be computed */
108 static cpumask_var_t cpu_sibling_setup_mask;
109 
110 struct mwait_cpu_dead {
111 	unsigned int	control;
112 	unsigned int	status;
113 };
114 
115 #define CPUDEAD_MWAIT_WAIT	0xDEADBEEF
116 #define CPUDEAD_MWAIT_KEXEC_HLT	0x4A17DEAD
117 
118 /*
119  * Cache line aligned data for mwait_play_dead(). Separate on purpose so
120  * that it's unlikely to be touched by other CPUs.
121  */
122 static DEFINE_PER_CPU_ALIGNED(struct mwait_cpu_dead, mwait_cpu_dead);
123 
124 /* Maximum number of SMT threads on any online core */
125 int __read_mostly __max_smt_threads = 1;
126 
127 /* Flag to indicate if a complete sched domain rebuild is required */
128 bool x86_topology_update;
129 
130 int arch_update_cpu_topology(void)
131 {
132 	int retval = x86_topology_update;
133 
134 	x86_topology_update = false;
135 	return retval;
136 }
137 
138 static unsigned int smpboot_warm_reset_vector_count;
139 
140 static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
141 {
142 	unsigned long flags;
143 
144 	spin_lock_irqsave(&rtc_lock, flags);
145 	if (!smpboot_warm_reset_vector_count++) {
146 		CMOS_WRITE(0xa, 0xf);
147 		*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) = start_eip >> 4;
148 		*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = start_eip & 0xf;
149 	}
150 	spin_unlock_irqrestore(&rtc_lock, flags);
151 }
152 
153 static inline void smpboot_restore_warm_reset_vector(void)
154 {
155 	unsigned long flags;
156 
157 	/*
158 	 * Paranoid:  Set warm reset code and vector here back
159 	 * to default values.
160 	 */
161 	spin_lock_irqsave(&rtc_lock, flags);
162 	if (!--smpboot_warm_reset_vector_count) {
163 		CMOS_WRITE(0, 0xf);
164 		*((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
165 	}
166 	spin_unlock_irqrestore(&rtc_lock, flags);
167 
168 }
169 
170 /* Run the next set of setup steps for the upcoming CPU */
171 static void ap_starting(void)
172 {
173 	int cpuid = smp_processor_id();
174 
175 	/* Mop up eventual mwait_play_dead() wreckage */
176 	this_cpu_write(mwait_cpu_dead.status, 0);
177 	this_cpu_write(mwait_cpu_dead.control, 0);
178 
179 	/*
180 	 * If woken up by an INIT in an 82489DX configuration the alive
181 	 * synchronization guarantees that the CPU does not reach this
182 	 * point before an INIT_deassert IPI reaches the local APIC, so it
183 	 * is now safe to touch the local APIC.
184 	 *
185 	 * Set up this CPU, first the APIC, which is probably redundant on
186 	 * most boards.
187 	 */
188 	apic_ap_setup();
189 
190 	/* Save the processor parameters. */
191 	smp_store_cpu_info(cpuid);
192 
193 	/*
194 	 * The topology information must be up to date before
195 	 * notify_cpu_starting().
196 	 */
197 	set_cpu_sibling_map(cpuid);
198 
199 	ap_init_aperfmperf();
200 
201 	pr_debug("Stack at about %p\n", &cpuid);
202 
203 	wmb();
204 
205 	/*
206 	 * This runs the AP through all the cpuhp states to its target
207 	 * state CPUHP_ONLINE.
208 	 */
209 	notify_cpu_starting(cpuid);
210 }
211 
212 static void ap_calibrate_delay(void)
213 {
214 	/*
215 	 * Calibrate the delay loop and update loops_per_jiffy in cpu_data.
216 	 * smp_store_cpu_info() stored a value that is close but not as
217 	 * accurate as the value just calculated.
218 	 *
219 	 * As this is invoked after the TSC synchronization check,
220 	 * calibrate_delay_is_known() will skip the calibration routine
221 	 * when TSC is synchronized across sockets.
222 	 */
223 	calibrate_delay();
224 	cpu_data(smp_processor_id()).loops_per_jiffy = loops_per_jiffy;
225 }
226 
227 /*
228  * Activate a secondary processor.
229  */
230 static void notrace start_secondary(void *unused)
231 {
232 	/*
233 	 * Don't put *anything* except direct CPU state initialization
234 	 * before cpu_init(), SMP booting is too fragile that we want to
235 	 * limit the things done here to the most necessary things.
236 	 */
237 	cr4_init();
238 
239 	/*
240 	 * 32-bit specific. 64-bit reaches this code with the correct page
241 	 * table established. Yet another historical divergence.
242 	 */
243 	if (IS_ENABLED(CONFIG_X86_32)) {
244 		/* switch away from the initial page table */
245 		load_cr3(swapper_pg_dir);
246 		__flush_tlb_all();
247 	}
248 
249 	cpu_init_exception_handling();
250 
251 	/*
252 	 * Load the microcode before reaching the AP alive synchronization
253 	 * point below so it is not part of the full per CPU serialized
254 	 * bringup part when "parallel" bringup is enabled.
255 	 *
256 	 * That's even safe when hyperthreading is enabled in the CPU as
257 	 * the core code starts the primary threads first and leaves the
258 	 * secondary threads waiting for SIPI. Loading microcode on
259 	 * physical cores concurrently is a safe operation.
260 	 *
261 	 * This covers both the Intel specific issue that concurrent
262 	 * microcode loading on SMT siblings must be prohibited and the
263 	 * vendor independent issue`that microcode loading which changes
264 	 * CPUID, MSRs etc. must be strictly serialized to maintain
265 	 * software state correctness.
266 	 */
267 	load_ucode_ap();
268 
269 	/*
270 	 * Synchronization point with the hotplug core. Sets this CPUs
271 	 * synchronization state to ALIVE and spin-waits for the control CPU to
272 	 * release this CPU for further bringup.
273 	 */
274 	cpuhp_ap_sync_alive();
275 
276 	cpu_init();
277 	fpu__init_cpu();
278 	rcutree_report_cpu_starting(raw_smp_processor_id());
279 	x86_cpuinit.early_percpu_clock_init();
280 
281 	ap_starting();
282 
283 	/* Check TSC synchronization with the control CPU. */
284 	check_tsc_sync_target();
285 
286 	/*
287 	 * Calibrate the delay loop after the TSC synchronization check.
288 	 * This allows to skip the calibration when TSC is synchronized
289 	 * across sockets.
290 	 */
291 	ap_calibrate_delay();
292 
293 	speculative_store_bypass_ht_init();
294 
295 	/*
296 	 * Lock vector_lock, set CPU online and bring the vector
297 	 * allocator online. Online must be set with vector_lock held
298 	 * to prevent a concurrent irq setup/teardown from seeing a
299 	 * half valid vector space.
300 	 */
301 	lock_vector_lock();
302 	set_cpu_online(smp_processor_id(), true);
303 	lapic_online();
304 	unlock_vector_lock();
305 	x86_platform.nmi_init();
306 
307 	/* enable local interrupts */
308 	local_irq_enable();
309 
310 	x86_cpuinit.setup_percpu_clockev();
311 
312 	wmb();
313 	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
314 }
315 
316 static void __init smp_store_boot_cpu_info(void)
317 {
318 	struct cpuinfo_x86 *c = &cpu_data(0);
319 
320 	*c = boot_cpu_data;
321 	c->initialized = true;
322 }
323 
324 /*
325  * The bootstrap kernel entry code has set these up. Save them for
326  * a given CPU
327  */
328 void smp_store_cpu_info(int id)
329 {
330 	struct cpuinfo_x86 *c = &cpu_data(id);
331 
332 	/* Copy boot_cpu_data only on the first bringup */
333 	if (!c->initialized)
334 		*c = boot_cpu_data;
335 	c->cpu_index = id;
336 	/*
337 	 * During boot time, CPU0 has this setup already. Save the info when
338 	 * bringing up an AP.
339 	 */
340 	identify_secondary_cpu(c);
341 	c->initialized = true;
342 }
343 
344 static bool
345 topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
346 {
347 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
348 
349 	return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
350 }
351 
352 static bool
353 topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
354 {
355 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
356 
357 	return !WARN_ONCE(!topology_same_node(c, o),
358 		"sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
359 		"[node: %d != %d]. Ignoring dependency.\n",
360 		cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
361 }
362 
363 #define link_mask(mfunc, c1, c2)					\
364 do {									\
365 	cpumask_set_cpu((c1), mfunc(c2));				\
366 	cpumask_set_cpu((c2), mfunc(c1));				\
367 } while (0)
368 
369 static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
370 {
371 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
372 		int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
373 
374 		if (c->topo.pkg_id == o->topo.pkg_id &&
375 		    c->topo.die_id == o->topo.die_id &&
376 		    c->topo.amd_node_id == o->topo.amd_node_id &&
377 		    per_cpu_llc_id(cpu1) == per_cpu_llc_id(cpu2)) {
378 			if (c->topo.core_id == o->topo.core_id)
379 				return topology_sane(c, o, "smt");
380 
381 			if ((c->topo.cu_id != 0xff) &&
382 			    (o->topo.cu_id != 0xff) &&
383 			    (c->topo.cu_id == o->topo.cu_id))
384 				return topology_sane(c, o, "smt");
385 		}
386 
387 	} else if (c->topo.pkg_id == o->topo.pkg_id &&
388 		   c->topo.die_id == o->topo.die_id &&
389 		   c->topo.core_id == o->topo.core_id) {
390 		return topology_sane(c, o, "smt");
391 	}
392 
393 	return false;
394 }
395 
396 static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
397 {
398 	if (c->topo.pkg_id != o->topo.pkg_id || c->topo.die_id != o->topo.die_id)
399 		return false;
400 
401 	if (cpu_feature_enabled(X86_FEATURE_TOPOEXT) && topology_amd_nodes_per_pkg() > 1)
402 		return c->topo.amd_node_id == o->topo.amd_node_id;
403 
404 	return true;
405 }
406 
407 static bool match_l2c(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
408 {
409 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
410 
411 	/* If the arch didn't set up l2c_id, fall back to SMT */
412 	if (per_cpu_l2c_id(cpu1) == BAD_APICID)
413 		return match_smt(c, o);
414 
415 	/* Do not match if L2 cache id does not match: */
416 	if (per_cpu_l2c_id(cpu1) != per_cpu_l2c_id(cpu2))
417 		return false;
418 
419 	return topology_sane(c, o, "l2c");
420 }
421 
422 /*
423  * Unlike the other levels, we do not enforce keeping a
424  * multicore group inside a NUMA node.  If this happens, we will
425  * discard the MC level of the topology later.
426  */
427 static bool match_pkg(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
428 {
429 	if (c->topo.pkg_id == o->topo.pkg_id)
430 		return true;
431 	return false;
432 }
433 
434 /*
435  * Define intel_cod_cpu[] for Intel COD (Cluster-on-Die) CPUs.
436  *
437  * Any Intel CPU that has multiple nodes per package and does not
438  * match intel_cod_cpu[] has the SNC (Sub-NUMA Cluster) topology.
439  *
440  * When in SNC mode, these CPUs enumerate an LLC that is shared
441  * by multiple NUMA nodes. The LLC is shared for off-package data
442  * access but private to the NUMA node (half of the package) for
443  * on-package access. CPUID (the source of the information about
444  * the LLC) can only enumerate the cache as shared or unshared,
445  * but not this particular configuration.
446  */
447 
448 static const struct x86_cpu_id intel_cod_cpu[] = {
449 	X86_MATCH_INTEL_FAM6_MODEL(HASWELL_X, 0),	/* COD */
450 	X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_X, 0),	/* COD */
451 	X86_MATCH_INTEL_FAM6_MODEL(ANY, 1),		/* SNC */
452 	{}
453 };
454 
455 static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
456 {
457 	const struct x86_cpu_id *id = x86_match_cpu(intel_cod_cpu);
458 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
459 	bool intel_snc = id && id->driver_data;
460 
461 	/* Do not match if we do not have a valid APICID for cpu: */
462 	if (per_cpu_llc_id(cpu1) == BAD_APICID)
463 		return false;
464 
465 	/* Do not match if LLC id does not match: */
466 	if (per_cpu_llc_id(cpu1) != per_cpu_llc_id(cpu2))
467 		return false;
468 
469 	/*
470 	 * Allow the SNC topology without warning. Return of false
471 	 * means 'c' does not share the LLC of 'o'. This will be
472 	 * reflected to userspace.
473 	 */
474 	if (match_pkg(c, o) && !topology_same_node(c, o) && intel_snc)
475 		return false;
476 
477 	return topology_sane(c, o, "llc");
478 }
479 
480 
481 static inline int x86_sched_itmt_flags(void)
482 {
483 	return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
484 }
485 
486 #ifdef CONFIG_SCHED_MC
487 static int x86_core_flags(void)
488 {
489 	return cpu_core_flags() | x86_sched_itmt_flags();
490 }
491 #endif
492 #ifdef CONFIG_SCHED_SMT
493 static int x86_smt_flags(void)
494 {
495 	return cpu_smt_flags();
496 }
497 #endif
498 #ifdef CONFIG_SCHED_CLUSTER
499 static int x86_cluster_flags(void)
500 {
501 	return cpu_cluster_flags() | x86_sched_itmt_flags();
502 }
503 #endif
504 
505 static int x86_die_flags(void)
506 {
507 	if (cpu_feature_enabled(X86_FEATURE_HYBRID_CPU))
508 	       return x86_sched_itmt_flags();
509 
510 	return 0;
511 }
512 
513 /*
514  * Set if a package/die has multiple NUMA nodes inside.
515  * AMD Magny-Cours, Intel Cluster-on-Die, and Intel
516  * Sub-NUMA Clustering have this.
517  */
518 static bool x86_has_numa_in_package;
519 
520 static struct sched_domain_topology_level x86_topology[6];
521 
522 static void __init build_sched_topology(void)
523 {
524 	int i = 0;
525 
526 #ifdef CONFIG_SCHED_SMT
527 	x86_topology[i++] = (struct sched_domain_topology_level){
528 		cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT)
529 	};
530 #endif
531 #ifdef CONFIG_SCHED_CLUSTER
532 	x86_topology[i++] = (struct sched_domain_topology_level){
533 		cpu_clustergroup_mask, x86_cluster_flags, SD_INIT_NAME(CLS)
534 	};
535 #endif
536 #ifdef CONFIG_SCHED_MC
537 	x86_topology[i++] = (struct sched_domain_topology_level){
538 		cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC)
539 	};
540 #endif
541 	/*
542 	 * When there is NUMA topology inside the package skip the PKG domain
543 	 * since the NUMA domains will auto-magically create the right spanning
544 	 * domains based on the SLIT.
545 	 */
546 	if (!x86_has_numa_in_package) {
547 		x86_topology[i++] = (struct sched_domain_topology_level){
548 			cpu_cpu_mask, x86_die_flags, SD_INIT_NAME(PKG)
549 		};
550 	}
551 
552 	/*
553 	 * There must be one trailing NULL entry left.
554 	 */
555 	BUG_ON(i >= ARRAY_SIZE(x86_topology)-1);
556 
557 	set_sched_topology(x86_topology);
558 }
559 
560 void set_cpu_sibling_map(int cpu)
561 {
562 	bool has_smt = __max_threads_per_core > 1;
563 	bool has_mp = has_smt || topology_num_cores_per_package() > 1;
564 	struct cpuinfo_x86 *c = &cpu_data(cpu);
565 	struct cpuinfo_x86 *o;
566 	int i, threads;
567 
568 	cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
569 
570 	if (!has_mp) {
571 		cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
572 		cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
573 		cpumask_set_cpu(cpu, cpu_l2c_shared_mask(cpu));
574 		cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
575 		cpumask_set_cpu(cpu, topology_die_cpumask(cpu));
576 		c->booted_cores = 1;
577 		return;
578 	}
579 
580 	for_each_cpu(i, cpu_sibling_setup_mask) {
581 		o = &cpu_data(i);
582 
583 		if (match_pkg(c, o) && !topology_same_node(c, o))
584 			x86_has_numa_in_package = true;
585 
586 		if ((i == cpu) || (has_smt && match_smt(c, o)))
587 			link_mask(topology_sibling_cpumask, cpu, i);
588 
589 		if ((i == cpu) || (has_mp && match_llc(c, o)))
590 			link_mask(cpu_llc_shared_mask, cpu, i);
591 
592 		if ((i == cpu) || (has_mp && match_l2c(c, o)))
593 			link_mask(cpu_l2c_shared_mask, cpu, i);
594 
595 		if ((i == cpu) || (has_mp && match_die(c, o)))
596 			link_mask(topology_die_cpumask, cpu, i);
597 	}
598 
599 	threads = cpumask_weight(topology_sibling_cpumask(cpu));
600 	if (threads > __max_smt_threads)
601 		__max_smt_threads = threads;
602 
603 	for_each_cpu(i, topology_sibling_cpumask(cpu))
604 		cpu_data(i).smt_active = threads > 1;
605 
606 	/*
607 	 * This needs a separate iteration over the cpus because we rely on all
608 	 * topology_sibling_cpumask links to be set-up.
609 	 */
610 	for_each_cpu(i, cpu_sibling_setup_mask) {
611 		o = &cpu_data(i);
612 
613 		if ((i == cpu) || (has_mp && match_pkg(c, o))) {
614 			link_mask(topology_core_cpumask, cpu, i);
615 
616 			/*
617 			 *  Does this new cpu bringup a new core?
618 			 */
619 			if (threads == 1) {
620 				/*
621 				 * for each core in package, increment
622 				 * the booted_cores for this new cpu
623 				 */
624 				if (cpumask_first(
625 				    topology_sibling_cpumask(i)) == i)
626 					c->booted_cores++;
627 				/*
628 				 * increment the core count for all
629 				 * the other cpus in this package
630 				 */
631 				if (i != cpu)
632 					cpu_data(i).booted_cores++;
633 			} else if (i != cpu && !c->booted_cores)
634 				c->booted_cores = cpu_data(i).booted_cores;
635 		}
636 	}
637 }
638 
639 /* maps the cpu to the sched domain representing multi-core */
640 const struct cpumask *cpu_coregroup_mask(int cpu)
641 {
642 	return cpu_llc_shared_mask(cpu);
643 }
644 
645 const struct cpumask *cpu_clustergroup_mask(int cpu)
646 {
647 	return cpu_l2c_shared_mask(cpu);
648 }
649 EXPORT_SYMBOL_GPL(cpu_clustergroup_mask);
650 
651 static void impress_friends(void)
652 {
653 	int cpu;
654 	unsigned long bogosum = 0;
655 	/*
656 	 * Allow the user to impress friends.
657 	 */
658 	pr_debug("Before bogomips\n");
659 	for_each_online_cpu(cpu)
660 		bogosum += cpu_data(cpu).loops_per_jiffy;
661 
662 	pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
663 		num_online_cpus(),
664 		bogosum/(500000/HZ),
665 		(bogosum/(5000/HZ))%100);
666 
667 	pr_debug("Before bogocount - setting activated=1\n");
668 }
669 
670 /*
671  * The Multiprocessor Specification 1.4 (1997) example code suggests
672  * that there should be a 10ms delay between the BSP asserting INIT
673  * and de-asserting INIT, when starting a remote processor.
674  * But that slows boot and resume on modern processors, which include
675  * many cores and don't require that delay.
676  *
677  * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
678  * Modern processor families are quirked to remove the delay entirely.
679  */
680 #define UDELAY_10MS_DEFAULT 10000
681 
682 static unsigned int init_udelay = UINT_MAX;
683 
684 static int __init cpu_init_udelay(char *str)
685 {
686 	get_option(&str, &init_udelay);
687 
688 	return 0;
689 }
690 early_param("cpu_init_udelay", cpu_init_udelay);
691 
692 static void __init smp_quirk_init_udelay(void)
693 {
694 	/* if cmdline changed it from default, leave it alone */
695 	if (init_udelay != UINT_MAX)
696 		return;
697 
698 	/* if modern processor, use no delay */
699 	if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
700 	    ((boot_cpu_data.x86_vendor == X86_VENDOR_HYGON) && (boot_cpu_data.x86 >= 0x18)) ||
701 	    ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
702 		init_udelay = 0;
703 		return;
704 	}
705 	/* else, use legacy delay */
706 	init_udelay = UDELAY_10MS_DEFAULT;
707 }
708 
709 /*
710  * Wake up AP by INIT, INIT, STARTUP sequence.
711  */
712 static void send_init_sequence(u32 phys_apicid)
713 {
714 	int maxlvt = lapic_get_maxlvt();
715 
716 	/* Be paranoid about clearing APIC errors. */
717 	if (APIC_INTEGRATED(boot_cpu_apic_version)) {
718 		/* Due to the Pentium erratum 3AP.  */
719 		if (maxlvt > 3)
720 			apic_write(APIC_ESR, 0);
721 		apic_read(APIC_ESR);
722 	}
723 
724 	/* Assert INIT on the target CPU */
725 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT, phys_apicid);
726 	safe_apic_wait_icr_idle();
727 
728 	udelay(init_udelay);
729 
730 	/* Deassert INIT on the target CPU */
731 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
732 	safe_apic_wait_icr_idle();
733 }
734 
735 /*
736  * Wake up AP by INIT, INIT, STARTUP sequence.
737  */
738 static int wakeup_secondary_cpu_via_init(u32 phys_apicid, unsigned long start_eip)
739 {
740 	unsigned long send_status = 0, accept_status = 0;
741 	int num_starts, j, maxlvt;
742 
743 	preempt_disable();
744 	maxlvt = lapic_get_maxlvt();
745 	send_init_sequence(phys_apicid);
746 
747 	mb();
748 
749 	/*
750 	 * Should we send STARTUP IPIs ?
751 	 *
752 	 * Determine this based on the APIC version.
753 	 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
754 	 */
755 	if (APIC_INTEGRATED(boot_cpu_apic_version))
756 		num_starts = 2;
757 	else
758 		num_starts = 0;
759 
760 	/*
761 	 * Run STARTUP IPI loop.
762 	 */
763 	pr_debug("#startup loops: %d\n", num_starts);
764 
765 	for (j = 1; j <= num_starts; j++) {
766 		pr_debug("Sending STARTUP #%d\n", j);
767 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
768 			apic_write(APIC_ESR, 0);
769 		apic_read(APIC_ESR);
770 		pr_debug("After apic_write\n");
771 
772 		/*
773 		 * STARTUP IPI
774 		 */
775 
776 		/* Target chip */
777 		/* Boot on the stack */
778 		/* Kick the second */
779 		apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
780 			       phys_apicid);
781 
782 		/*
783 		 * Give the other CPU some time to accept the IPI.
784 		 */
785 		if (init_udelay == 0)
786 			udelay(10);
787 		else
788 			udelay(300);
789 
790 		pr_debug("Startup point 1\n");
791 
792 		pr_debug("Waiting for send to finish...\n");
793 		send_status = safe_apic_wait_icr_idle();
794 
795 		/*
796 		 * Give the other CPU some time to accept the IPI.
797 		 */
798 		if (init_udelay == 0)
799 			udelay(10);
800 		else
801 			udelay(200);
802 
803 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
804 			apic_write(APIC_ESR, 0);
805 		accept_status = (apic_read(APIC_ESR) & 0xEF);
806 		if (send_status || accept_status)
807 			break;
808 	}
809 	pr_debug("After Startup\n");
810 
811 	if (send_status)
812 		pr_err("APIC never delivered???\n");
813 	if (accept_status)
814 		pr_err("APIC delivery error (%lx)\n", accept_status);
815 
816 	preempt_enable();
817 	return (send_status | accept_status);
818 }
819 
820 /* reduce the number of lines printed when booting a large cpu count system */
821 static void announce_cpu(int cpu, int apicid)
822 {
823 	static int width, node_width, first = 1;
824 	static int current_node = NUMA_NO_NODE;
825 	int node = early_cpu_to_node(cpu);
826 
827 	if (!width)
828 		width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
829 
830 	if (!node_width)
831 		node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
832 
833 	if (system_state < SYSTEM_RUNNING) {
834 		if (first)
835 			pr_info("x86: Booting SMP configuration:\n");
836 
837 		if (node != current_node) {
838 			if (current_node > (-1))
839 				pr_cont("\n");
840 			current_node = node;
841 
842 			printk(KERN_INFO ".... node %*s#%d, CPUs:  ",
843 			       node_width - num_digits(node), " ", node);
844 		}
845 
846 		/* Add padding for the BSP */
847 		if (first)
848 			pr_cont("%*s", width + 1, " ");
849 		first = 0;
850 
851 		pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
852 	} else
853 		pr_info("Booting Node %d Processor %d APIC 0x%x\n",
854 			node, cpu, apicid);
855 }
856 
857 int common_cpu_up(unsigned int cpu, struct task_struct *idle)
858 {
859 	int ret;
860 
861 	/* Just in case we booted with a single CPU. */
862 	alternatives_enable_smp();
863 
864 	per_cpu(pcpu_hot.current_task, cpu) = idle;
865 	cpu_init_stack_canary(cpu, idle);
866 
867 	/* Initialize the interrupt stack(s) */
868 	ret = irq_init_percpu_irqstack(cpu);
869 	if (ret)
870 		return ret;
871 
872 #ifdef CONFIG_X86_32
873 	/* Stack for startup_32 can be just as for start_secondary onwards */
874 	per_cpu(pcpu_hot.top_of_stack, cpu) = task_top_of_stack(idle);
875 #endif
876 	return 0;
877 }
878 
879 /*
880  * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
881  * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
882  * Returns zero if startup was successfully sent, else error code from
883  * ->wakeup_secondary_cpu.
884  */
885 static int do_boot_cpu(u32 apicid, int cpu, struct task_struct *idle)
886 {
887 	unsigned long start_ip = real_mode_header->trampoline_start;
888 	int ret;
889 
890 #ifdef CONFIG_X86_64
891 	/* If 64-bit wakeup method exists, use the 64-bit mode trampoline IP */
892 	if (apic->wakeup_secondary_cpu_64)
893 		start_ip = real_mode_header->trampoline_start64;
894 #endif
895 	idle->thread.sp = (unsigned long)task_pt_regs(idle);
896 	initial_code = (unsigned long)start_secondary;
897 
898 	if (IS_ENABLED(CONFIG_X86_32)) {
899 		early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
900 		initial_stack  = idle->thread.sp;
901 	} else if (!(smpboot_control & STARTUP_PARALLEL_MASK)) {
902 		smpboot_control = cpu;
903 	}
904 
905 	/* Enable the espfix hack for this CPU */
906 	init_espfix_ap(cpu);
907 
908 	/* So we see what's up */
909 	announce_cpu(cpu, apicid);
910 
911 	/*
912 	 * This grunge runs the startup process for
913 	 * the targeted processor.
914 	 */
915 	if (x86_platform.legacy.warm_reset) {
916 
917 		pr_debug("Setting warm reset code and vector.\n");
918 
919 		smpboot_setup_warm_reset_vector(start_ip);
920 		/*
921 		 * Be paranoid about clearing APIC errors.
922 		*/
923 		if (APIC_INTEGRATED(boot_cpu_apic_version)) {
924 			apic_write(APIC_ESR, 0);
925 			apic_read(APIC_ESR);
926 		}
927 	}
928 
929 	smp_mb();
930 
931 	/*
932 	 * Wake up a CPU in difference cases:
933 	 * - Use a method from the APIC driver if one defined, with wakeup
934 	 *   straight to 64-bit mode preferred over wakeup to RM.
935 	 * Otherwise,
936 	 * - Use an INIT boot APIC message
937 	 */
938 	if (apic->wakeup_secondary_cpu_64)
939 		ret = apic->wakeup_secondary_cpu_64(apicid, start_ip);
940 	else if (apic->wakeup_secondary_cpu)
941 		ret = apic->wakeup_secondary_cpu(apicid, start_ip);
942 	else
943 		ret = wakeup_secondary_cpu_via_init(apicid, start_ip);
944 
945 	/* If the wakeup mechanism failed, cleanup the warm reset vector */
946 	if (ret)
947 		arch_cpuhp_cleanup_kick_cpu(cpu);
948 	return ret;
949 }
950 
951 int native_kick_ap(unsigned int cpu, struct task_struct *tidle)
952 {
953 	u32 apicid = apic->cpu_present_to_apicid(cpu);
954 	int err;
955 
956 	lockdep_assert_irqs_enabled();
957 
958 	pr_debug("++++++++++++++++++++=_---CPU UP  %u\n", cpu);
959 
960 	if (apicid == BAD_APICID || !apic_id_valid(apicid)) {
961 		pr_err("CPU %u has invalid APIC ID %x. Aborting bringup\n", cpu, apicid);
962 		return -EINVAL;
963 	}
964 
965 	if (!test_bit(apicid, phys_cpu_present_map)) {
966 		pr_err("CPU %u APIC ID %x is not present. Aborting bringup\n", cpu, apicid);
967 		return -EINVAL;
968 	}
969 
970 	/*
971 	 * Save current MTRR state in case it was changed since early boot
972 	 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
973 	 */
974 	mtrr_save_state();
975 
976 	/* the FPU context is blank, nobody can own it */
977 	per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
978 
979 	err = common_cpu_up(cpu, tidle);
980 	if (err)
981 		return err;
982 
983 	err = do_boot_cpu(apicid, cpu, tidle);
984 	if (err)
985 		pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
986 
987 	return err;
988 }
989 
990 int arch_cpuhp_kick_ap_alive(unsigned int cpu, struct task_struct *tidle)
991 {
992 	return smp_ops.kick_ap_alive(cpu, tidle);
993 }
994 
995 void arch_cpuhp_cleanup_kick_cpu(unsigned int cpu)
996 {
997 	/* Cleanup possible dangling ends... */
998 	if (smp_ops.kick_ap_alive == native_kick_ap && x86_platform.legacy.warm_reset)
999 		smpboot_restore_warm_reset_vector();
1000 }
1001 
1002 void arch_cpuhp_cleanup_dead_cpu(unsigned int cpu)
1003 {
1004 	if (smp_ops.cleanup_dead_cpu)
1005 		smp_ops.cleanup_dead_cpu(cpu);
1006 
1007 	if (system_state == SYSTEM_RUNNING)
1008 		pr_info("CPU %u is now offline\n", cpu);
1009 }
1010 
1011 void arch_cpuhp_sync_state_poll(void)
1012 {
1013 	if (smp_ops.poll_sync_state)
1014 		smp_ops.poll_sync_state();
1015 }
1016 
1017 /**
1018  * arch_disable_smp_support() - Disables SMP support for x86 at boottime
1019  */
1020 void __init arch_disable_smp_support(void)
1021 {
1022 	disable_ioapic_support();
1023 }
1024 
1025 /*
1026  * Fall back to non SMP mode after errors.
1027  *
1028  * RED-PEN audit/test this more. I bet there is more state messed up here.
1029  */
1030 static __init void disable_smp(void)
1031 {
1032 	pr_info("SMP disabled\n");
1033 
1034 	disable_ioapic_support();
1035 	topology_reset_possible_cpus_up();
1036 
1037 	cpumask_set_cpu(0, topology_sibling_cpumask(0));
1038 	cpumask_set_cpu(0, topology_core_cpumask(0));
1039 	cpumask_set_cpu(0, topology_die_cpumask(0));
1040 }
1041 
1042 static void __init smp_cpu_index_default(void)
1043 {
1044 	int i;
1045 	struct cpuinfo_x86 *c;
1046 
1047 	for_each_possible_cpu(i) {
1048 		c = &cpu_data(i);
1049 		/* mark all to hotplug */
1050 		c->cpu_index = nr_cpu_ids;
1051 	}
1052 }
1053 
1054 void __init smp_prepare_cpus_common(void)
1055 {
1056 	unsigned int i;
1057 
1058 	smp_cpu_index_default();
1059 
1060 	/*
1061 	 * Setup boot CPU information
1062 	 */
1063 	smp_store_boot_cpu_info(); /* Final full version of the data */
1064 	mb();
1065 
1066 	for_each_possible_cpu(i) {
1067 		zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1068 		zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1069 		zalloc_cpumask_var(&per_cpu(cpu_die_map, i), GFP_KERNEL);
1070 		zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1071 		zalloc_cpumask_var(&per_cpu(cpu_l2c_shared_map, i), GFP_KERNEL);
1072 	}
1073 
1074 	set_cpu_sibling_map(0);
1075 }
1076 
1077 void __init smp_prepare_boot_cpu(void)
1078 {
1079 	smp_ops.smp_prepare_boot_cpu();
1080 }
1081 
1082 #ifdef CONFIG_X86_64
1083 /* Establish whether parallel bringup can be supported. */
1084 bool __init arch_cpuhp_init_parallel_bringup(void)
1085 {
1086 	if (!x86_cpuinit.parallel_bringup) {
1087 		pr_info("Parallel CPU startup disabled by the platform\n");
1088 		return false;
1089 	}
1090 
1091 	smpboot_control = STARTUP_READ_APICID;
1092 	pr_debug("Parallel CPU startup enabled: 0x%08x\n", smpboot_control);
1093 	return true;
1094 }
1095 #endif
1096 
1097 /*
1098  * Prepare for SMP bootup.
1099  * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1100  *            for common interface support.
1101  */
1102 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1103 {
1104 	smp_prepare_cpus_common();
1105 
1106 	switch (apic_intr_mode) {
1107 	case APIC_PIC:
1108 	case APIC_VIRTUAL_WIRE_NO_CONFIG:
1109 		disable_smp();
1110 		return;
1111 	case APIC_SYMMETRIC_IO_NO_ROUTING:
1112 		disable_smp();
1113 		/* Setup local timer */
1114 		x86_init.timers.setup_percpu_clockev();
1115 		return;
1116 	case APIC_VIRTUAL_WIRE:
1117 	case APIC_SYMMETRIC_IO:
1118 		break;
1119 	}
1120 
1121 	/* Setup local timer */
1122 	x86_init.timers.setup_percpu_clockev();
1123 
1124 	pr_info("CPU0: ");
1125 	print_cpu_info(&cpu_data(0));
1126 
1127 	uv_system_init();
1128 
1129 	smp_quirk_init_udelay();
1130 
1131 	speculative_store_bypass_ht_init();
1132 
1133 	snp_set_wakeup_secondary_cpu();
1134 }
1135 
1136 void arch_thaw_secondary_cpus_begin(void)
1137 {
1138 	set_cache_aps_delayed_init(true);
1139 }
1140 
1141 void arch_thaw_secondary_cpus_end(void)
1142 {
1143 	cache_aps_init();
1144 }
1145 
1146 /*
1147  * Early setup to make printk work.
1148  */
1149 void __init native_smp_prepare_boot_cpu(void)
1150 {
1151 	int me = smp_processor_id();
1152 
1153 	/* SMP handles this from setup_per_cpu_areas() */
1154 	if (!IS_ENABLED(CONFIG_SMP))
1155 		switch_gdt_and_percpu_base(me);
1156 
1157 	native_pv_lock_init();
1158 }
1159 
1160 void __init native_smp_cpus_done(unsigned int max_cpus)
1161 {
1162 	pr_debug("Boot done\n");
1163 
1164 	build_sched_topology();
1165 	nmi_selftest();
1166 	impress_friends();
1167 	cache_aps_init();
1168 }
1169 
1170 /* correctly size the local cpu masks */
1171 void __init setup_cpu_local_masks(void)
1172 {
1173 	alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
1174 }
1175 
1176 #ifdef CONFIG_HOTPLUG_CPU
1177 
1178 /* Recompute SMT state for all CPUs on offline */
1179 static void recompute_smt_state(void)
1180 {
1181 	int max_threads, cpu;
1182 
1183 	max_threads = 0;
1184 	for_each_online_cpu (cpu) {
1185 		int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1186 
1187 		if (threads > max_threads)
1188 			max_threads = threads;
1189 	}
1190 	__max_smt_threads = max_threads;
1191 }
1192 
1193 static void remove_siblinginfo(int cpu)
1194 {
1195 	int sibling;
1196 	struct cpuinfo_x86 *c = &cpu_data(cpu);
1197 
1198 	for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1199 		cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1200 		/*/
1201 		 * last thread sibling in this cpu core going down
1202 		 */
1203 		if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1204 			cpu_data(sibling).booted_cores--;
1205 	}
1206 
1207 	for_each_cpu(sibling, topology_die_cpumask(cpu))
1208 		cpumask_clear_cpu(cpu, topology_die_cpumask(sibling));
1209 
1210 	for_each_cpu(sibling, topology_sibling_cpumask(cpu)) {
1211 		cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1212 		if (cpumask_weight(topology_sibling_cpumask(sibling)) == 1)
1213 			cpu_data(sibling).smt_active = false;
1214 	}
1215 
1216 	for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1217 		cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1218 	for_each_cpu(sibling, cpu_l2c_shared_mask(cpu))
1219 		cpumask_clear_cpu(cpu, cpu_l2c_shared_mask(sibling));
1220 	cpumask_clear(cpu_llc_shared_mask(cpu));
1221 	cpumask_clear(cpu_l2c_shared_mask(cpu));
1222 	cpumask_clear(topology_sibling_cpumask(cpu));
1223 	cpumask_clear(topology_core_cpumask(cpu));
1224 	cpumask_clear(topology_die_cpumask(cpu));
1225 	c->topo.core_id = 0;
1226 	c->booted_cores = 0;
1227 	cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1228 	recompute_smt_state();
1229 }
1230 
1231 static void remove_cpu_from_maps(int cpu)
1232 {
1233 	set_cpu_online(cpu, false);
1234 	numa_remove_cpu(cpu);
1235 }
1236 
1237 void cpu_disable_common(void)
1238 {
1239 	int cpu = smp_processor_id();
1240 
1241 	remove_siblinginfo(cpu);
1242 
1243 	/* It's now safe to remove this processor from the online map */
1244 	lock_vector_lock();
1245 	remove_cpu_from_maps(cpu);
1246 	unlock_vector_lock();
1247 	fixup_irqs();
1248 	lapic_offline();
1249 }
1250 
1251 int native_cpu_disable(void)
1252 {
1253 	int ret;
1254 
1255 	ret = lapic_can_unplug_cpu();
1256 	if (ret)
1257 		return ret;
1258 
1259 	cpu_disable_common();
1260 
1261         /*
1262          * Disable the local APIC. Otherwise IPI broadcasts will reach
1263          * it. It still responds normally to INIT, NMI, SMI, and SIPI
1264          * messages.
1265          *
1266          * Disabling the APIC must happen after cpu_disable_common()
1267          * which invokes fixup_irqs().
1268          *
1269          * Disabling the APIC preserves already set bits in IRR, but
1270          * an interrupt arriving after disabling the local APIC does not
1271          * set the corresponding IRR bit.
1272          *
1273          * fixup_irqs() scans IRR for set bits so it can raise a not
1274          * yet handled interrupt on the new destination CPU via an IPI
1275          * but obviously it can't do so for IRR bits which are not set.
1276          * IOW, interrupts arriving after disabling the local APIC will
1277          * be lost.
1278          */
1279 	apic_soft_disable();
1280 
1281 	return 0;
1282 }
1283 
1284 void play_dead_common(void)
1285 {
1286 	idle_task_exit();
1287 
1288 	cpuhp_ap_report_dead();
1289 
1290 	local_irq_disable();
1291 }
1292 
1293 /*
1294  * We need to flush the caches before going to sleep, lest we have
1295  * dirty data in our caches when we come back up.
1296  */
1297 static inline void mwait_play_dead(void)
1298 {
1299 	struct mwait_cpu_dead *md = this_cpu_ptr(&mwait_cpu_dead);
1300 	unsigned int eax, ebx, ecx, edx;
1301 	unsigned int highest_cstate = 0;
1302 	unsigned int highest_subcstate = 0;
1303 	int i;
1304 
1305 	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD ||
1306 	    boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
1307 		return;
1308 	if (!this_cpu_has(X86_FEATURE_MWAIT))
1309 		return;
1310 	if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1311 		return;
1312 	if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1313 		return;
1314 
1315 	eax = CPUID_MWAIT_LEAF;
1316 	ecx = 0;
1317 	native_cpuid(&eax, &ebx, &ecx, &edx);
1318 
1319 	/*
1320 	 * eax will be 0 if EDX enumeration is not valid.
1321 	 * Initialized below to cstate, sub_cstate value when EDX is valid.
1322 	 */
1323 	if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1324 		eax = 0;
1325 	} else {
1326 		edx >>= MWAIT_SUBSTATE_SIZE;
1327 		for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1328 			if (edx & MWAIT_SUBSTATE_MASK) {
1329 				highest_cstate = i;
1330 				highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1331 			}
1332 		}
1333 		eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1334 			(highest_subcstate - 1);
1335 	}
1336 
1337 	/* Set up state for the kexec() hack below */
1338 	md->status = CPUDEAD_MWAIT_WAIT;
1339 	md->control = CPUDEAD_MWAIT_WAIT;
1340 
1341 	wbinvd();
1342 
1343 	while (1) {
1344 		/*
1345 		 * The CLFLUSH is a workaround for erratum AAI65 for
1346 		 * the Xeon 7400 series.  It's not clear it is actually
1347 		 * needed, but it should be harmless in either case.
1348 		 * The WBINVD is insufficient due to the spurious-wakeup
1349 		 * case where we return around the loop.
1350 		 */
1351 		mb();
1352 		clflush(md);
1353 		mb();
1354 		__monitor(md, 0, 0);
1355 		mb();
1356 		__mwait(eax, 0);
1357 
1358 		if (READ_ONCE(md->control) == CPUDEAD_MWAIT_KEXEC_HLT) {
1359 			/*
1360 			 * Kexec is about to happen. Don't go back into mwait() as
1361 			 * the kexec kernel might overwrite text and data including
1362 			 * page tables and stack. So mwait() would resume when the
1363 			 * monitor cache line is written to and then the CPU goes
1364 			 * south due to overwritten text, page tables and stack.
1365 			 *
1366 			 * Note: This does _NOT_ protect against a stray MCE, NMI,
1367 			 * SMI. They will resume execution at the instruction
1368 			 * following the HLT instruction and run into the problem
1369 			 * which this is trying to prevent.
1370 			 */
1371 			WRITE_ONCE(md->status, CPUDEAD_MWAIT_KEXEC_HLT);
1372 			while(1)
1373 				native_halt();
1374 		}
1375 	}
1376 }
1377 
1378 /*
1379  * Kick all "offline" CPUs out of mwait on kexec(). See comment in
1380  * mwait_play_dead().
1381  */
1382 void smp_kick_mwait_play_dead(void)
1383 {
1384 	u32 newstate = CPUDEAD_MWAIT_KEXEC_HLT;
1385 	struct mwait_cpu_dead *md;
1386 	unsigned int cpu, i;
1387 
1388 	for_each_cpu_andnot(cpu, cpu_present_mask, cpu_online_mask) {
1389 		md = per_cpu_ptr(&mwait_cpu_dead, cpu);
1390 
1391 		/* Does it sit in mwait_play_dead() ? */
1392 		if (READ_ONCE(md->status) != CPUDEAD_MWAIT_WAIT)
1393 			continue;
1394 
1395 		/* Wait up to 5ms */
1396 		for (i = 0; READ_ONCE(md->status) != newstate && i < 1000; i++) {
1397 			/* Bring it out of mwait */
1398 			WRITE_ONCE(md->control, newstate);
1399 			udelay(5);
1400 		}
1401 
1402 		if (READ_ONCE(md->status) != newstate)
1403 			pr_err_once("CPU%u is stuck in mwait_play_dead()\n", cpu);
1404 	}
1405 }
1406 
1407 void __noreturn hlt_play_dead(void)
1408 {
1409 	if (__this_cpu_read(cpu_info.x86) >= 4)
1410 		wbinvd();
1411 
1412 	while (1)
1413 		native_halt();
1414 }
1415 
1416 /*
1417  * native_play_dead() is essentially a __noreturn function, but it can't
1418  * be marked as such as the compiler may complain about it.
1419  */
1420 void native_play_dead(void)
1421 {
1422 	if (cpu_feature_enabled(X86_FEATURE_KERNEL_IBRS))
1423 		__update_spec_ctrl(0);
1424 
1425 	play_dead_common();
1426 	tboot_shutdown(TB_SHUTDOWN_WFS);
1427 
1428 	mwait_play_dead();
1429 	if (cpuidle_play_dead())
1430 		hlt_play_dead();
1431 }
1432 
1433 #else /* ... !CONFIG_HOTPLUG_CPU */
1434 int native_cpu_disable(void)
1435 {
1436 	return -ENOSYS;
1437 }
1438 
1439 void native_play_dead(void)
1440 {
1441 	BUG();
1442 }
1443 
1444 #endif
1445