xref: /openbmc/linux/arch/x86/kernel/smpboot.c (revision 046b212a)
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 
91 /* representing HT siblings of each logical CPU */
92 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
93 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
94 
95 /* representing HT and core siblings of each logical CPU */
96 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
97 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
98 
99 /* representing HT, core, and die siblings of each logical CPU */
100 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_die_map);
101 EXPORT_PER_CPU_SYMBOL(cpu_die_map);
102 
103 /* Per CPU bogomips and other parameters */
104 DEFINE_PER_CPU_READ_MOSTLY(struct cpuinfo_x86, cpu_info);
105 EXPORT_PER_CPU_SYMBOL(cpu_info);
106 
107 /* CPUs which are the primary SMT threads */
108 struct cpumask __cpu_primary_thread_mask __read_mostly;
109 
110 /* Representing CPUs for which sibling maps can be computed */
111 static cpumask_var_t cpu_sibling_setup_mask;
112 
113 struct mwait_cpu_dead {
114 	unsigned int	control;
115 	unsigned int	status;
116 };
117 
118 #define CPUDEAD_MWAIT_WAIT	0xDEADBEEF
119 #define CPUDEAD_MWAIT_KEXEC_HLT	0x4A17DEAD
120 
121 /*
122  * Cache line aligned data for mwait_play_dead(). Separate on purpose so
123  * that it's unlikely to be touched by other CPUs.
124  */
125 static DEFINE_PER_CPU_ALIGNED(struct mwait_cpu_dead, mwait_cpu_dead);
126 
127 /* Logical package management. We might want to allocate that dynamically */
128 unsigned int __max_logical_packages __read_mostly;
129 EXPORT_SYMBOL(__max_logical_packages);
130 static unsigned int logical_packages __read_mostly;
131 static unsigned int logical_die __read_mostly;
132 
133 /* Maximum number of SMT threads on any online core */
134 int __read_mostly __max_smt_threads = 1;
135 
136 /* Flag to indicate if a complete sched domain rebuild is required */
137 bool x86_topology_update;
138 
139 int arch_update_cpu_topology(void)
140 {
141 	int retval = x86_topology_update;
142 
143 	x86_topology_update = false;
144 	return retval;
145 }
146 
147 static unsigned int smpboot_warm_reset_vector_count;
148 
149 static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
150 {
151 	unsigned long flags;
152 
153 	spin_lock_irqsave(&rtc_lock, flags);
154 	if (!smpboot_warm_reset_vector_count++) {
155 		CMOS_WRITE(0xa, 0xf);
156 		*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) = start_eip >> 4;
157 		*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = start_eip & 0xf;
158 	}
159 	spin_unlock_irqrestore(&rtc_lock, flags);
160 }
161 
162 static inline void smpboot_restore_warm_reset_vector(void)
163 {
164 	unsigned long flags;
165 
166 	/*
167 	 * Paranoid:  Set warm reset code and vector here back
168 	 * to default values.
169 	 */
170 	spin_lock_irqsave(&rtc_lock, flags);
171 	if (!--smpboot_warm_reset_vector_count) {
172 		CMOS_WRITE(0, 0xf);
173 		*((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
174 	}
175 	spin_unlock_irqrestore(&rtc_lock, flags);
176 
177 }
178 
179 /* Run the next set of setup steps for the upcoming CPU */
180 static void ap_starting(void)
181 {
182 	int cpuid = smp_processor_id();
183 
184 	/* Mop up eventual mwait_play_dead() wreckage */
185 	this_cpu_write(mwait_cpu_dead.status, 0);
186 	this_cpu_write(mwait_cpu_dead.control, 0);
187 
188 	/*
189 	 * If woken up by an INIT in an 82489DX configuration the alive
190 	 * synchronization guarantees that the CPU does not reach this
191 	 * point before an INIT_deassert IPI reaches the local APIC, so it
192 	 * is now safe to touch the local APIC.
193 	 *
194 	 * Set up this CPU, first the APIC, which is probably redundant on
195 	 * most boards.
196 	 */
197 	apic_ap_setup();
198 
199 	/* Save the processor parameters. */
200 	smp_store_cpu_info(cpuid);
201 
202 	/*
203 	 * The topology information must be up to date before
204 	 * notify_cpu_starting().
205 	 */
206 	set_cpu_sibling_map(cpuid);
207 
208 	ap_init_aperfmperf();
209 
210 	pr_debug("Stack at about %p\n", &cpuid);
211 
212 	wmb();
213 
214 	/*
215 	 * This runs the AP through all the cpuhp states to its target
216 	 * state CPUHP_ONLINE.
217 	 */
218 	notify_cpu_starting(cpuid);
219 }
220 
221 static void ap_calibrate_delay(void)
222 {
223 	/*
224 	 * Calibrate the delay loop and update loops_per_jiffy in cpu_data.
225 	 * smp_store_cpu_info() stored a value that is close but not as
226 	 * accurate as the value just calculated.
227 	 *
228 	 * As this is invoked after the TSC synchronization check,
229 	 * calibrate_delay_is_known() will skip the calibration routine
230 	 * when TSC is synchronized across sockets.
231 	 */
232 	calibrate_delay();
233 	cpu_data(smp_processor_id()).loops_per_jiffy = loops_per_jiffy;
234 }
235 
236 /*
237  * Activate a secondary processor.
238  */
239 static void notrace start_secondary(void *unused)
240 {
241 	/*
242 	 * Don't put *anything* except direct CPU state initialization
243 	 * before cpu_init(), SMP booting is too fragile that we want to
244 	 * limit the things done here to the most necessary things.
245 	 */
246 	cr4_init();
247 
248 	/*
249 	 * 32-bit specific. 64-bit reaches this code with the correct page
250 	 * table established. Yet another historical divergence.
251 	 */
252 	if (IS_ENABLED(CONFIG_X86_32)) {
253 		/* switch away from the initial page table */
254 		load_cr3(swapper_pg_dir);
255 		__flush_tlb_all();
256 	}
257 
258 	cpu_init_exception_handling();
259 
260 	/*
261 	 * 32-bit systems load the microcode from the ASM startup code for
262 	 * historical reasons.
263 	 *
264 	 * On 64-bit systems load it before reaching the AP alive
265 	 * synchronization point below so it is not part of the full per
266 	 * CPU serialized bringup part when "parallel" bringup is enabled.
267 	 *
268 	 * That's even safe when hyperthreading is enabled in the CPU as
269 	 * the core code starts the primary threads first and leaves the
270 	 * secondary threads waiting for SIPI. Loading microcode on
271 	 * physical cores concurrently is a safe operation.
272 	 *
273 	 * This covers both the Intel specific issue that concurrent
274 	 * microcode loading on SMT siblings must be prohibited and the
275 	 * vendor independent issue`that microcode loading which changes
276 	 * CPUID, MSRs etc. must be strictly serialized to maintain
277 	 * software state correctness.
278 	 */
279 	if (IS_ENABLED(CONFIG_X86_64))
280 		load_ucode_ap();
281 
282 	/*
283 	 * Synchronization point with the hotplug core. Sets this CPUs
284 	 * synchronization state to ALIVE and spin-waits for the control CPU to
285 	 * release this CPU for further bringup.
286 	 */
287 	cpuhp_ap_sync_alive();
288 
289 	cpu_init();
290 	fpu__init_cpu();
291 	rcu_cpu_starting(raw_smp_processor_id());
292 	x86_cpuinit.early_percpu_clock_init();
293 
294 	ap_starting();
295 
296 	/* Check TSC synchronization with the control CPU. */
297 	check_tsc_sync_target();
298 
299 	/*
300 	 * Calibrate the delay loop after the TSC synchronization check.
301 	 * This allows to skip the calibration when TSC is synchronized
302 	 * across sockets.
303 	 */
304 	ap_calibrate_delay();
305 
306 	speculative_store_bypass_ht_init();
307 
308 	/*
309 	 * Lock vector_lock, set CPU online and bring the vector
310 	 * allocator online. Online must be set with vector_lock held
311 	 * to prevent a concurrent irq setup/teardown from seeing a
312 	 * half valid vector space.
313 	 */
314 	lock_vector_lock();
315 	set_cpu_online(smp_processor_id(), true);
316 	lapic_online();
317 	unlock_vector_lock();
318 	x86_platform.nmi_init();
319 
320 	/* enable local interrupts */
321 	local_irq_enable();
322 
323 	x86_cpuinit.setup_percpu_clockev();
324 
325 	wmb();
326 	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
327 }
328 
329 /**
330  * topology_phys_to_logical_pkg - Map a physical package id to a logical
331  * @phys_pkg:	The physical package id to map
332  *
333  * Returns logical package id or -1 if not found
334  */
335 int topology_phys_to_logical_pkg(unsigned int phys_pkg)
336 {
337 	int cpu;
338 
339 	for_each_possible_cpu(cpu) {
340 		struct cpuinfo_x86 *c = &cpu_data(cpu);
341 
342 		if (c->initialized && c->phys_proc_id == phys_pkg)
343 			return c->logical_proc_id;
344 	}
345 	return -1;
346 }
347 EXPORT_SYMBOL(topology_phys_to_logical_pkg);
348 
349 /**
350  * topology_phys_to_logical_die - Map a physical die id to logical
351  * @die_id:	The physical die id to map
352  * @cur_cpu:	The CPU for which the mapping is done
353  *
354  * Returns logical die id or -1 if not found
355  */
356 static int topology_phys_to_logical_die(unsigned int die_id, unsigned int cur_cpu)
357 {
358 	int cpu, proc_id = cpu_data(cur_cpu).phys_proc_id;
359 
360 	for_each_possible_cpu(cpu) {
361 		struct cpuinfo_x86 *c = &cpu_data(cpu);
362 
363 		if (c->initialized && c->cpu_die_id == die_id &&
364 		    c->phys_proc_id == proc_id)
365 			return c->logical_die_id;
366 	}
367 	return -1;
368 }
369 
370 /**
371  * topology_update_package_map - Update the physical to logical package map
372  * @pkg:	The physical package id as retrieved via CPUID
373  * @cpu:	The cpu for which this is updated
374  */
375 int topology_update_package_map(unsigned int pkg, unsigned int cpu)
376 {
377 	int new;
378 
379 	/* Already available somewhere? */
380 	new = topology_phys_to_logical_pkg(pkg);
381 	if (new >= 0)
382 		goto found;
383 
384 	new = logical_packages++;
385 	if (new != pkg) {
386 		pr_info("CPU %u Converting physical %u to logical package %u\n",
387 			cpu, pkg, new);
388 	}
389 found:
390 	cpu_data(cpu).logical_proc_id = new;
391 	return 0;
392 }
393 /**
394  * topology_update_die_map - Update the physical to logical die map
395  * @die:	The die id as retrieved via CPUID
396  * @cpu:	The cpu for which this is updated
397  */
398 int topology_update_die_map(unsigned int die, unsigned int cpu)
399 {
400 	int new;
401 
402 	/* Already available somewhere? */
403 	new = topology_phys_to_logical_die(die, cpu);
404 	if (new >= 0)
405 		goto found;
406 
407 	new = logical_die++;
408 	if (new != die) {
409 		pr_info("CPU %u Converting physical %u to logical die %u\n",
410 			cpu, die, new);
411 	}
412 found:
413 	cpu_data(cpu).logical_die_id = new;
414 	return 0;
415 }
416 
417 static void __init smp_store_boot_cpu_info(void)
418 {
419 	int id = 0; /* CPU 0 */
420 	struct cpuinfo_x86 *c = &cpu_data(id);
421 
422 	*c = boot_cpu_data;
423 	c->cpu_index = id;
424 	topology_update_package_map(c->phys_proc_id, id);
425 	topology_update_die_map(c->cpu_die_id, id);
426 	c->initialized = true;
427 }
428 
429 /*
430  * The bootstrap kernel entry code has set these up. Save them for
431  * a given CPU
432  */
433 void smp_store_cpu_info(int id)
434 {
435 	struct cpuinfo_x86 *c = &cpu_data(id);
436 
437 	/* Copy boot_cpu_data only on the first bringup */
438 	if (!c->initialized)
439 		*c = boot_cpu_data;
440 	c->cpu_index = id;
441 	/*
442 	 * During boot time, CPU0 has this setup already. Save the info when
443 	 * bringing up an AP.
444 	 */
445 	identify_secondary_cpu(c);
446 	c->initialized = true;
447 }
448 
449 static bool
450 topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
451 {
452 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
453 
454 	return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
455 }
456 
457 static bool
458 topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
459 {
460 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
461 
462 	return !WARN_ONCE(!topology_same_node(c, o),
463 		"sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
464 		"[node: %d != %d]. Ignoring dependency.\n",
465 		cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
466 }
467 
468 #define link_mask(mfunc, c1, c2)					\
469 do {									\
470 	cpumask_set_cpu((c1), mfunc(c2));				\
471 	cpumask_set_cpu((c2), mfunc(c1));				\
472 } while (0)
473 
474 static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
475 {
476 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
477 		int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
478 
479 		if (c->phys_proc_id == o->phys_proc_id &&
480 		    c->cpu_die_id == o->cpu_die_id &&
481 		    per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2)) {
482 			if (c->cpu_core_id == o->cpu_core_id)
483 				return topology_sane(c, o, "smt");
484 
485 			if ((c->cu_id != 0xff) &&
486 			    (o->cu_id != 0xff) &&
487 			    (c->cu_id == o->cu_id))
488 				return topology_sane(c, o, "smt");
489 		}
490 
491 	} else if (c->phys_proc_id == o->phys_proc_id &&
492 		   c->cpu_die_id == o->cpu_die_id &&
493 		   c->cpu_core_id == o->cpu_core_id) {
494 		return topology_sane(c, o, "smt");
495 	}
496 
497 	return false;
498 }
499 
500 static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
501 {
502 	if (c->phys_proc_id == o->phys_proc_id &&
503 	    c->cpu_die_id == o->cpu_die_id)
504 		return true;
505 	return false;
506 }
507 
508 static bool match_l2c(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
509 {
510 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
511 
512 	/* If the arch didn't set up l2c_id, fall back to SMT */
513 	if (per_cpu(cpu_l2c_id, cpu1) == BAD_APICID)
514 		return match_smt(c, o);
515 
516 	/* Do not match if L2 cache id does not match: */
517 	if (per_cpu(cpu_l2c_id, cpu1) != per_cpu(cpu_l2c_id, cpu2))
518 		return false;
519 
520 	return topology_sane(c, o, "l2c");
521 }
522 
523 /*
524  * Unlike the other levels, we do not enforce keeping a
525  * multicore group inside a NUMA node.  If this happens, we will
526  * discard the MC level of the topology later.
527  */
528 static bool match_pkg(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
529 {
530 	if (c->phys_proc_id == o->phys_proc_id)
531 		return true;
532 	return false;
533 }
534 
535 /*
536  * Define intel_cod_cpu[] for Intel COD (Cluster-on-Die) CPUs.
537  *
538  * Any Intel CPU that has multiple nodes per package and does not
539  * match intel_cod_cpu[] has the SNC (Sub-NUMA Cluster) topology.
540  *
541  * When in SNC mode, these CPUs enumerate an LLC that is shared
542  * by multiple NUMA nodes. The LLC is shared for off-package data
543  * access but private to the NUMA node (half of the package) for
544  * on-package access. CPUID (the source of the information about
545  * the LLC) can only enumerate the cache as shared or unshared,
546  * but not this particular configuration.
547  */
548 
549 static const struct x86_cpu_id intel_cod_cpu[] = {
550 	X86_MATCH_INTEL_FAM6_MODEL(HASWELL_X, 0),	/* COD */
551 	X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_X, 0),	/* COD */
552 	X86_MATCH_INTEL_FAM6_MODEL(ANY, 1),		/* SNC */
553 	{}
554 };
555 
556 static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
557 {
558 	const struct x86_cpu_id *id = x86_match_cpu(intel_cod_cpu);
559 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
560 	bool intel_snc = id && id->driver_data;
561 
562 	/* Do not match if we do not have a valid APICID for cpu: */
563 	if (per_cpu(cpu_llc_id, cpu1) == BAD_APICID)
564 		return false;
565 
566 	/* Do not match if LLC id does not match: */
567 	if (per_cpu(cpu_llc_id, cpu1) != per_cpu(cpu_llc_id, cpu2))
568 		return false;
569 
570 	/*
571 	 * Allow the SNC topology without warning. Return of false
572 	 * means 'c' does not share the LLC of 'o'. This will be
573 	 * reflected to userspace.
574 	 */
575 	if (match_pkg(c, o) && !topology_same_node(c, o) && intel_snc)
576 		return false;
577 
578 	return topology_sane(c, o, "llc");
579 }
580 
581 
582 #if defined(CONFIG_SCHED_SMT) || defined(CONFIG_SCHED_CLUSTER) || defined(CONFIG_SCHED_MC)
583 static inline int x86_sched_itmt_flags(void)
584 {
585 	return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
586 }
587 
588 #ifdef CONFIG_SCHED_MC
589 static int x86_core_flags(void)
590 {
591 	return cpu_core_flags() | x86_sched_itmt_flags();
592 }
593 #endif
594 #ifdef CONFIG_SCHED_SMT
595 static int x86_smt_flags(void)
596 {
597 	return cpu_smt_flags();
598 }
599 #endif
600 #ifdef CONFIG_SCHED_CLUSTER
601 static int x86_cluster_flags(void)
602 {
603 	return cpu_cluster_flags() | x86_sched_itmt_flags();
604 }
605 #endif
606 #endif
607 
608 /*
609  * Set if a package/die has multiple NUMA nodes inside.
610  * AMD Magny-Cours, Intel Cluster-on-Die, and Intel
611  * Sub-NUMA Clustering have this.
612  */
613 static bool x86_has_numa_in_package;
614 
615 static struct sched_domain_topology_level x86_topology[6];
616 
617 static void __init build_sched_topology(void)
618 {
619 	int i = 0;
620 
621 #ifdef CONFIG_SCHED_SMT
622 	x86_topology[i++] = (struct sched_domain_topology_level){
623 		cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT)
624 	};
625 #endif
626 #ifdef CONFIG_SCHED_CLUSTER
627 	x86_topology[i++] = (struct sched_domain_topology_level){
628 		cpu_clustergroup_mask, x86_cluster_flags, SD_INIT_NAME(CLS)
629 	};
630 #endif
631 #ifdef CONFIG_SCHED_MC
632 	x86_topology[i++] = (struct sched_domain_topology_level){
633 		cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC)
634 	};
635 #endif
636 	/*
637 	 * When there is NUMA topology inside the package skip the DIE domain
638 	 * since the NUMA domains will auto-magically create the right spanning
639 	 * domains based on the SLIT.
640 	 */
641 	if (!x86_has_numa_in_package) {
642 		x86_topology[i++] = (struct sched_domain_topology_level){
643 			cpu_cpu_mask, SD_INIT_NAME(DIE)
644 		};
645 	}
646 
647 	/*
648 	 * There must be one trailing NULL entry left.
649 	 */
650 	BUG_ON(i >= ARRAY_SIZE(x86_topology)-1);
651 
652 	set_sched_topology(x86_topology);
653 }
654 
655 void set_cpu_sibling_map(int cpu)
656 {
657 	bool has_smt = smp_num_siblings > 1;
658 	bool has_mp = has_smt || boot_cpu_data.x86_max_cores > 1;
659 	struct cpuinfo_x86 *c = &cpu_data(cpu);
660 	struct cpuinfo_x86 *o;
661 	int i, threads;
662 
663 	cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
664 
665 	if (!has_mp) {
666 		cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
667 		cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
668 		cpumask_set_cpu(cpu, cpu_l2c_shared_mask(cpu));
669 		cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
670 		cpumask_set_cpu(cpu, topology_die_cpumask(cpu));
671 		c->booted_cores = 1;
672 		return;
673 	}
674 
675 	for_each_cpu(i, cpu_sibling_setup_mask) {
676 		o = &cpu_data(i);
677 
678 		if (match_pkg(c, o) && !topology_same_node(c, o))
679 			x86_has_numa_in_package = true;
680 
681 		if ((i == cpu) || (has_smt && match_smt(c, o)))
682 			link_mask(topology_sibling_cpumask, cpu, i);
683 
684 		if ((i == cpu) || (has_mp && match_llc(c, o)))
685 			link_mask(cpu_llc_shared_mask, cpu, i);
686 
687 		if ((i == cpu) || (has_mp && match_l2c(c, o)))
688 			link_mask(cpu_l2c_shared_mask, cpu, i);
689 
690 		if ((i == cpu) || (has_mp && match_die(c, o)))
691 			link_mask(topology_die_cpumask, cpu, i);
692 	}
693 
694 	threads = cpumask_weight(topology_sibling_cpumask(cpu));
695 	if (threads > __max_smt_threads)
696 		__max_smt_threads = threads;
697 
698 	for_each_cpu(i, topology_sibling_cpumask(cpu))
699 		cpu_data(i).smt_active = threads > 1;
700 
701 	/*
702 	 * This needs a separate iteration over the cpus because we rely on all
703 	 * topology_sibling_cpumask links to be set-up.
704 	 */
705 	for_each_cpu(i, cpu_sibling_setup_mask) {
706 		o = &cpu_data(i);
707 
708 		if ((i == cpu) || (has_mp && match_pkg(c, o))) {
709 			link_mask(topology_core_cpumask, cpu, i);
710 
711 			/*
712 			 *  Does this new cpu bringup a new core?
713 			 */
714 			if (threads == 1) {
715 				/*
716 				 * for each core in package, increment
717 				 * the booted_cores for this new cpu
718 				 */
719 				if (cpumask_first(
720 				    topology_sibling_cpumask(i)) == i)
721 					c->booted_cores++;
722 				/*
723 				 * increment the core count for all
724 				 * the other cpus in this package
725 				 */
726 				if (i != cpu)
727 					cpu_data(i).booted_cores++;
728 			} else if (i != cpu && !c->booted_cores)
729 				c->booted_cores = cpu_data(i).booted_cores;
730 		}
731 	}
732 }
733 
734 /* maps the cpu to the sched domain representing multi-core */
735 const struct cpumask *cpu_coregroup_mask(int cpu)
736 {
737 	return cpu_llc_shared_mask(cpu);
738 }
739 
740 const struct cpumask *cpu_clustergroup_mask(int cpu)
741 {
742 	return cpu_l2c_shared_mask(cpu);
743 }
744 
745 static void impress_friends(void)
746 {
747 	int cpu;
748 	unsigned long bogosum = 0;
749 	/*
750 	 * Allow the user to impress friends.
751 	 */
752 	pr_debug("Before bogomips\n");
753 	for_each_online_cpu(cpu)
754 		bogosum += cpu_data(cpu).loops_per_jiffy;
755 
756 	pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
757 		num_online_cpus(),
758 		bogosum/(500000/HZ),
759 		(bogosum/(5000/HZ))%100);
760 
761 	pr_debug("Before bogocount - setting activated=1\n");
762 }
763 
764 /*
765  * The Multiprocessor Specification 1.4 (1997) example code suggests
766  * that there should be a 10ms delay between the BSP asserting INIT
767  * and de-asserting INIT, when starting a remote processor.
768  * But that slows boot and resume on modern processors, which include
769  * many cores and don't require that delay.
770  *
771  * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
772  * Modern processor families are quirked to remove the delay entirely.
773  */
774 #define UDELAY_10MS_DEFAULT 10000
775 
776 static unsigned int init_udelay = UINT_MAX;
777 
778 static int __init cpu_init_udelay(char *str)
779 {
780 	get_option(&str, &init_udelay);
781 
782 	return 0;
783 }
784 early_param("cpu_init_udelay", cpu_init_udelay);
785 
786 static void __init smp_quirk_init_udelay(void)
787 {
788 	/* if cmdline changed it from default, leave it alone */
789 	if (init_udelay != UINT_MAX)
790 		return;
791 
792 	/* if modern processor, use no delay */
793 	if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
794 	    ((boot_cpu_data.x86_vendor == X86_VENDOR_HYGON) && (boot_cpu_data.x86 >= 0x18)) ||
795 	    ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
796 		init_udelay = 0;
797 		return;
798 	}
799 	/* else, use legacy delay */
800 	init_udelay = UDELAY_10MS_DEFAULT;
801 }
802 
803 /*
804  * Wake up AP by INIT, INIT, STARTUP sequence.
805  */
806 static void send_init_sequence(int phys_apicid)
807 {
808 	int maxlvt = lapic_get_maxlvt();
809 
810 	/* Be paranoid about clearing APIC errors. */
811 	if (APIC_INTEGRATED(boot_cpu_apic_version)) {
812 		/* Due to the Pentium erratum 3AP.  */
813 		if (maxlvt > 3)
814 			apic_write(APIC_ESR, 0);
815 		apic_read(APIC_ESR);
816 	}
817 
818 	/* Assert INIT on the target CPU */
819 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT, phys_apicid);
820 	safe_apic_wait_icr_idle();
821 
822 	udelay(init_udelay);
823 
824 	/* Deassert INIT on the target CPU */
825 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
826 	safe_apic_wait_icr_idle();
827 }
828 
829 /*
830  * Wake up AP by INIT, INIT, STARTUP sequence.
831  */
832 static int wakeup_secondary_cpu_via_init(int phys_apicid, unsigned long start_eip)
833 {
834 	unsigned long send_status = 0, accept_status = 0;
835 	int num_starts, j, maxlvt;
836 
837 	preempt_disable();
838 	maxlvt = lapic_get_maxlvt();
839 	send_init_sequence(phys_apicid);
840 
841 	mb();
842 
843 	/*
844 	 * Should we send STARTUP IPIs ?
845 	 *
846 	 * Determine this based on the APIC version.
847 	 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
848 	 */
849 	if (APIC_INTEGRATED(boot_cpu_apic_version))
850 		num_starts = 2;
851 	else
852 		num_starts = 0;
853 
854 	/*
855 	 * Run STARTUP IPI loop.
856 	 */
857 	pr_debug("#startup loops: %d\n", num_starts);
858 
859 	for (j = 1; j <= num_starts; j++) {
860 		pr_debug("Sending STARTUP #%d\n", j);
861 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
862 			apic_write(APIC_ESR, 0);
863 		apic_read(APIC_ESR);
864 		pr_debug("After apic_write\n");
865 
866 		/*
867 		 * STARTUP IPI
868 		 */
869 
870 		/* Target chip */
871 		/* Boot on the stack */
872 		/* Kick the second */
873 		apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
874 			       phys_apicid);
875 
876 		/*
877 		 * Give the other CPU some time to accept the IPI.
878 		 */
879 		if (init_udelay == 0)
880 			udelay(10);
881 		else
882 			udelay(300);
883 
884 		pr_debug("Startup point 1\n");
885 
886 		pr_debug("Waiting for send to finish...\n");
887 		send_status = safe_apic_wait_icr_idle();
888 
889 		/*
890 		 * Give the other CPU some time to accept the IPI.
891 		 */
892 		if (init_udelay == 0)
893 			udelay(10);
894 		else
895 			udelay(200);
896 
897 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
898 			apic_write(APIC_ESR, 0);
899 		accept_status = (apic_read(APIC_ESR) & 0xEF);
900 		if (send_status || accept_status)
901 			break;
902 	}
903 	pr_debug("After Startup\n");
904 
905 	if (send_status)
906 		pr_err("APIC never delivered???\n");
907 	if (accept_status)
908 		pr_err("APIC delivery error (%lx)\n", accept_status);
909 
910 	preempt_enable();
911 	return (send_status | accept_status);
912 }
913 
914 /* reduce the number of lines printed when booting a large cpu count system */
915 static void announce_cpu(int cpu, int apicid)
916 {
917 	static int width, node_width, first = 1;
918 	static int current_node = NUMA_NO_NODE;
919 	int node = early_cpu_to_node(cpu);
920 
921 	if (!width)
922 		width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
923 
924 	if (!node_width)
925 		node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
926 
927 	if (system_state < SYSTEM_RUNNING) {
928 		if (first)
929 			pr_info("x86: Booting SMP configuration:\n");
930 
931 		if (node != current_node) {
932 			if (current_node > (-1))
933 				pr_cont("\n");
934 			current_node = node;
935 
936 			printk(KERN_INFO ".... node %*s#%d, CPUs:  ",
937 			       node_width - num_digits(node), " ", node);
938 		}
939 
940 		/* Add padding for the BSP */
941 		if (first)
942 			pr_cont("%*s", width + 1, " ");
943 		first = 0;
944 
945 		pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
946 	} else
947 		pr_info("Booting Node %d Processor %d APIC 0x%x\n",
948 			node, cpu, apicid);
949 }
950 
951 int common_cpu_up(unsigned int cpu, struct task_struct *idle)
952 {
953 	int ret;
954 
955 	/* Just in case we booted with a single CPU. */
956 	alternatives_enable_smp();
957 
958 	per_cpu(pcpu_hot.current_task, cpu) = idle;
959 	cpu_init_stack_canary(cpu, idle);
960 
961 	/* Initialize the interrupt stack(s) */
962 	ret = irq_init_percpu_irqstack(cpu);
963 	if (ret)
964 		return ret;
965 
966 #ifdef CONFIG_X86_32
967 	/* Stack for startup_32 can be just as for start_secondary onwards */
968 	per_cpu(pcpu_hot.top_of_stack, cpu) = task_top_of_stack(idle);
969 #endif
970 	return 0;
971 }
972 
973 /*
974  * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
975  * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
976  * Returns zero if startup was successfully sent, else error code from
977  * ->wakeup_secondary_cpu.
978  */
979 static int do_boot_cpu(int apicid, int cpu, struct task_struct *idle)
980 {
981 	unsigned long start_ip = real_mode_header->trampoline_start;
982 	int ret;
983 
984 #ifdef CONFIG_X86_64
985 	/* If 64-bit wakeup method exists, use the 64-bit mode trampoline IP */
986 	if (apic->wakeup_secondary_cpu_64)
987 		start_ip = real_mode_header->trampoline_start64;
988 #endif
989 	idle->thread.sp = (unsigned long)task_pt_regs(idle);
990 	initial_code = (unsigned long)start_secondary;
991 
992 	if (IS_ENABLED(CONFIG_X86_32)) {
993 		early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
994 		initial_stack  = idle->thread.sp;
995 	} else if (!(smpboot_control & STARTUP_PARALLEL_MASK)) {
996 		smpboot_control = cpu;
997 	}
998 
999 	/* Enable the espfix hack for this CPU */
1000 	init_espfix_ap(cpu);
1001 
1002 	/* So we see what's up */
1003 	announce_cpu(cpu, apicid);
1004 
1005 	/*
1006 	 * This grunge runs the startup process for
1007 	 * the targeted processor.
1008 	 */
1009 	if (x86_platform.legacy.warm_reset) {
1010 
1011 		pr_debug("Setting warm reset code and vector.\n");
1012 
1013 		smpboot_setup_warm_reset_vector(start_ip);
1014 		/*
1015 		 * Be paranoid about clearing APIC errors.
1016 		*/
1017 		if (APIC_INTEGRATED(boot_cpu_apic_version)) {
1018 			apic_write(APIC_ESR, 0);
1019 			apic_read(APIC_ESR);
1020 		}
1021 	}
1022 
1023 	smp_mb();
1024 
1025 	/*
1026 	 * Wake up a CPU in difference cases:
1027 	 * - Use a method from the APIC driver if one defined, with wakeup
1028 	 *   straight to 64-bit mode preferred over wakeup to RM.
1029 	 * Otherwise,
1030 	 * - Use an INIT boot APIC message
1031 	 */
1032 	if (apic->wakeup_secondary_cpu_64)
1033 		ret = apic->wakeup_secondary_cpu_64(apicid, start_ip);
1034 	else if (apic->wakeup_secondary_cpu)
1035 		ret = apic->wakeup_secondary_cpu(apicid, start_ip);
1036 	else
1037 		ret = wakeup_secondary_cpu_via_init(apicid, start_ip);
1038 
1039 	/* If the wakeup mechanism failed, cleanup the warm reset vector */
1040 	if (ret)
1041 		arch_cpuhp_cleanup_kick_cpu(cpu);
1042 	return ret;
1043 }
1044 
1045 int native_kick_ap(unsigned int cpu, struct task_struct *tidle)
1046 {
1047 	int apicid = apic->cpu_present_to_apicid(cpu);
1048 	int err;
1049 
1050 	lockdep_assert_irqs_enabled();
1051 
1052 	pr_debug("++++++++++++++++++++=_---CPU UP  %u\n", cpu);
1053 
1054 	if (apicid == BAD_APICID || !physid_isset(apicid, phys_cpu_present_map) ||
1055 	    !apic_id_valid(apicid)) {
1056 		pr_err("%s: bad cpu %d\n", __func__, cpu);
1057 		return -EINVAL;
1058 	}
1059 
1060 	/*
1061 	 * Save current MTRR state in case it was changed since early boot
1062 	 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1063 	 */
1064 	mtrr_save_state();
1065 
1066 	/* the FPU context is blank, nobody can own it */
1067 	per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
1068 
1069 	err = common_cpu_up(cpu, tidle);
1070 	if (err)
1071 		return err;
1072 
1073 	err = do_boot_cpu(apicid, cpu, tidle);
1074 	if (err)
1075 		pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
1076 
1077 	return err;
1078 }
1079 
1080 int arch_cpuhp_kick_ap_alive(unsigned int cpu, struct task_struct *tidle)
1081 {
1082 	return smp_ops.kick_ap_alive(cpu, tidle);
1083 }
1084 
1085 void arch_cpuhp_cleanup_kick_cpu(unsigned int cpu)
1086 {
1087 	/* Cleanup possible dangling ends... */
1088 	if (smp_ops.kick_ap_alive == native_kick_ap && x86_platform.legacy.warm_reset)
1089 		smpboot_restore_warm_reset_vector();
1090 }
1091 
1092 void arch_cpuhp_cleanup_dead_cpu(unsigned int cpu)
1093 {
1094 	if (smp_ops.cleanup_dead_cpu)
1095 		smp_ops.cleanup_dead_cpu(cpu);
1096 
1097 	if (system_state == SYSTEM_RUNNING)
1098 		pr_info("CPU %u is now offline\n", cpu);
1099 }
1100 
1101 void arch_cpuhp_sync_state_poll(void)
1102 {
1103 	if (smp_ops.poll_sync_state)
1104 		smp_ops.poll_sync_state();
1105 }
1106 
1107 /**
1108  * arch_disable_smp_support() - Disables SMP support for x86 at boottime
1109  */
1110 void __init arch_disable_smp_support(void)
1111 {
1112 	disable_ioapic_support();
1113 }
1114 
1115 /*
1116  * Fall back to non SMP mode after errors.
1117  *
1118  * RED-PEN audit/test this more. I bet there is more state messed up here.
1119  */
1120 static __init void disable_smp(void)
1121 {
1122 	pr_info("SMP disabled\n");
1123 
1124 	disable_ioapic_support();
1125 
1126 	init_cpu_present(cpumask_of(0));
1127 	init_cpu_possible(cpumask_of(0));
1128 
1129 	if (smp_found_config)
1130 		physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
1131 	else
1132 		physid_set_mask_of_physid(0, &phys_cpu_present_map);
1133 	cpumask_set_cpu(0, topology_sibling_cpumask(0));
1134 	cpumask_set_cpu(0, topology_core_cpumask(0));
1135 	cpumask_set_cpu(0, topology_die_cpumask(0));
1136 }
1137 
1138 static void __init smp_cpu_index_default(void)
1139 {
1140 	int i;
1141 	struct cpuinfo_x86 *c;
1142 
1143 	for_each_possible_cpu(i) {
1144 		c = &cpu_data(i);
1145 		/* mark all to hotplug */
1146 		c->cpu_index = nr_cpu_ids;
1147 	}
1148 }
1149 
1150 void __init smp_prepare_cpus_common(void)
1151 {
1152 	unsigned int i;
1153 
1154 	smp_cpu_index_default();
1155 
1156 	/*
1157 	 * Setup boot CPU information
1158 	 */
1159 	smp_store_boot_cpu_info(); /* Final full version of the data */
1160 	mb();
1161 
1162 	for_each_possible_cpu(i) {
1163 		zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1164 		zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1165 		zalloc_cpumask_var(&per_cpu(cpu_die_map, i), GFP_KERNEL);
1166 		zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1167 		zalloc_cpumask_var(&per_cpu(cpu_l2c_shared_map, i), GFP_KERNEL);
1168 	}
1169 
1170 	set_cpu_sibling_map(0);
1171 }
1172 
1173 #ifdef CONFIG_X86_64
1174 /* Establish whether parallel bringup can be supported. */
1175 bool __init arch_cpuhp_init_parallel_bringup(void)
1176 {
1177 	if (!x86_cpuinit.parallel_bringup) {
1178 		pr_info("Parallel CPU startup disabled by the platform\n");
1179 		return false;
1180 	}
1181 
1182 	smpboot_control = STARTUP_READ_APICID;
1183 	pr_debug("Parallel CPU startup enabled: 0x%08x\n", smpboot_control);
1184 	return true;
1185 }
1186 #endif
1187 
1188 /*
1189  * Prepare for SMP bootup.
1190  * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1191  *            for common interface support.
1192  */
1193 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1194 {
1195 	smp_prepare_cpus_common();
1196 
1197 	switch (apic_intr_mode) {
1198 	case APIC_PIC:
1199 	case APIC_VIRTUAL_WIRE_NO_CONFIG:
1200 		disable_smp();
1201 		return;
1202 	case APIC_SYMMETRIC_IO_NO_ROUTING:
1203 		disable_smp();
1204 		/* Setup local timer */
1205 		x86_init.timers.setup_percpu_clockev();
1206 		return;
1207 	case APIC_VIRTUAL_WIRE:
1208 	case APIC_SYMMETRIC_IO:
1209 		break;
1210 	}
1211 
1212 	/* Setup local timer */
1213 	x86_init.timers.setup_percpu_clockev();
1214 
1215 	pr_info("CPU0: ");
1216 	print_cpu_info(&cpu_data(0));
1217 
1218 	uv_system_init();
1219 
1220 	smp_quirk_init_udelay();
1221 
1222 	speculative_store_bypass_ht_init();
1223 
1224 	snp_set_wakeup_secondary_cpu();
1225 }
1226 
1227 void arch_thaw_secondary_cpus_begin(void)
1228 {
1229 	set_cache_aps_delayed_init(true);
1230 }
1231 
1232 void arch_thaw_secondary_cpus_end(void)
1233 {
1234 	cache_aps_init();
1235 }
1236 
1237 bool smp_park_other_cpus_in_init(void)
1238 {
1239 	unsigned int cpu, this_cpu = smp_processor_id();
1240 	unsigned int apicid;
1241 
1242 	if (apic->wakeup_secondary_cpu_64 || apic->wakeup_secondary_cpu)
1243 		return false;
1244 
1245 	/*
1246 	 * If this is a crash stop which does not execute on the boot CPU,
1247 	 * then this cannot use the INIT mechanism because INIT to the boot
1248 	 * CPU will reset the machine.
1249 	 */
1250 	if (this_cpu)
1251 		return false;
1252 
1253 	for_each_cpu_and(cpu, &cpus_booted_once_mask, cpu_present_mask) {
1254 		if (cpu == this_cpu)
1255 			continue;
1256 		apicid = apic->cpu_present_to_apicid(cpu);
1257 		if (apicid == BAD_APICID)
1258 			continue;
1259 		send_init_sequence(apicid);
1260 	}
1261 	return true;
1262 }
1263 
1264 /*
1265  * Early setup to make printk work.
1266  */
1267 void __init native_smp_prepare_boot_cpu(void)
1268 {
1269 	int me = smp_processor_id();
1270 
1271 	/* SMP handles this from setup_per_cpu_areas() */
1272 	if (!IS_ENABLED(CONFIG_SMP))
1273 		switch_gdt_and_percpu_base(me);
1274 
1275 	native_pv_lock_init();
1276 }
1277 
1278 void __init calculate_max_logical_packages(void)
1279 {
1280 	int ncpus;
1281 
1282 	/*
1283 	 * Today neither Intel nor AMD support heterogeneous systems so
1284 	 * extrapolate the boot cpu's data to all packages.
1285 	 */
1286 	ncpus = cpu_data(0).booted_cores * topology_max_smt_threads();
1287 	__max_logical_packages = DIV_ROUND_UP(total_cpus, ncpus);
1288 	pr_info("Max logical packages: %u\n", __max_logical_packages);
1289 }
1290 
1291 void __init native_smp_cpus_done(unsigned int max_cpus)
1292 {
1293 	pr_debug("Boot done\n");
1294 
1295 	calculate_max_logical_packages();
1296 	build_sched_topology();
1297 	nmi_selftest();
1298 	impress_friends();
1299 	cache_aps_init();
1300 }
1301 
1302 static int __initdata setup_possible_cpus = -1;
1303 static int __init _setup_possible_cpus(char *str)
1304 {
1305 	get_option(&str, &setup_possible_cpus);
1306 	return 0;
1307 }
1308 early_param("possible_cpus", _setup_possible_cpus);
1309 
1310 
1311 /*
1312  * cpu_possible_mask should be static, it cannot change as cpu's
1313  * are onlined, or offlined. The reason is per-cpu data-structures
1314  * are allocated by some modules at init time, and don't expect to
1315  * do this dynamically on cpu arrival/departure.
1316  * cpu_present_mask on the other hand can change dynamically.
1317  * In case when cpu_hotplug is not compiled, then we resort to current
1318  * behaviour, which is cpu_possible == cpu_present.
1319  * - Ashok Raj
1320  *
1321  * Three ways to find out the number of additional hotplug CPUs:
1322  * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1323  * - The user can overwrite it with possible_cpus=NUM
1324  * - Otherwise don't reserve additional CPUs.
1325  * We do this because additional CPUs waste a lot of memory.
1326  * -AK
1327  */
1328 __init void prefill_possible_map(void)
1329 {
1330 	int i, possible;
1331 
1332 	i = setup_max_cpus ?: 1;
1333 	if (setup_possible_cpus == -1) {
1334 		possible = num_processors;
1335 #ifdef CONFIG_HOTPLUG_CPU
1336 		if (setup_max_cpus)
1337 			possible += disabled_cpus;
1338 #else
1339 		if (possible > i)
1340 			possible = i;
1341 #endif
1342 	} else
1343 		possible = setup_possible_cpus;
1344 
1345 	total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1346 
1347 	/* nr_cpu_ids could be reduced via nr_cpus= */
1348 	if (possible > nr_cpu_ids) {
1349 		pr_warn("%d Processors exceeds NR_CPUS limit of %u\n",
1350 			possible, nr_cpu_ids);
1351 		possible = nr_cpu_ids;
1352 	}
1353 
1354 #ifdef CONFIG_HOTPLUG_CPU
1355 	if (!setup_max_cpus)
1356 #endif
1357 	if (possible > i) {
1358 		pr_warn("%d Processors exceeds max_cpus limit of %u\n",
1359 			possible, setup_max_cpus);
1360 		possible = i;
1361 	}
1362 
1363 	set_nr_cpu_ids(possible);
1364 
1365 	pr_info("Allowing %d CPUs, %d hotplug CPUs\n",
1366 		possible, max_t(int, possible - num_processors, 0));
1367 
1368 	reset_cpu_possible_mask();
1369 
1370 	for (i = 0; i < possible; i++)
1371 		set_cpu_possible(i, true);
1372 }
1373 
1374 /* correctly size the local cpu masks */
1375 void __init setup_cpu_local_masks(void)
1376 {
1377 	alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
1378 }
1379 
1380 #ifdef CONFIG_HOTPLUG_CPU
1381 
1382 /* Recompute SMT state for all CPUs on offline */
1383 static void recompute_smt_state(void)
1384 {
1385 	int max_threads, cpu;
1386 
1387 	max_threads = 0;
1388 	for_each_online_cpu (cpu) {
1389 		int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1390 
1391 		if (threads > max_threads)
1392 			max_threads = threads;
1393 	}
1394 	__max_smt_threads = max_threads;
1395 }
1396 
1397 static void remove_siblinginfo(int cpu)
1398 {
1399 	int sibling;
1400 	struct cpuinfo_x86 *c = &cpu_data(cpu);
1401 
1402 	for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1403 		cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1404 		/*/
1405 		 * last thread sibling in this cpu core going down
1406 		 */
1407 		if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1408 			cpu_data(sibling).booted_cores--;
1409 	}
1410 
1411 	for_each_cpu(sibling, topology_die_cpumask(cpu))
1412 		cpumask_clear_cpu(cpu, topology_die_cpumask(sibling));
1413 
1414 	for_each_cpu(sibling, topology_sibling_cpumask(cpu)) {
1415 		cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1416 		if (cpumask_weight(topology_sibling_cpumask(sibling)) == 1)
1417 			cpu_data(sibling).smt_active = false;
1418 	}
1419 
1420 	for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1421 		cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1422 	for_each_cpu(sibling, cpu_l2c_shared_mask(cpu))
1423 		cpumask_clear_cpu(cpu, cpu_l2c_shared_mask(sibling));
1424 	cpumask_clear(cpu_llc_shared_mask(cpu));
1425 	cpumask_clear(cpu_l2c_shared_mask(cpu));
1426 	cpumask_clear(topology_sibling_cpumask(cpu));
1427 	cpumask_clear(topology_core_cpumask(cpu));
1428 	cpumask_clear(topology_die_cpumask(cpu));
1429 	c->cpu_core_id = 0;
1430 	c->booted_cores = 0;
1431 	cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1432 	recompute_smt_state();
1433 }
1434 
1435 static void remove_cpu_from_maps(int cpu)
1436 {
1437 	set_cpu_online(cpu, false);
1438 	numa_remove_cpu(cpu);
1439 }
1440 
1441 void cpu_disable_common(void)
1442 {
1443 	int cpu = smp_processor_id();
1444 
1445 	remove_siblinginfo(cpu);
1446 
1447 	/* It's now safe to remove this processor from the online map */
1448 	lock_vector_lock();
1449 	remove_cpu_from_maps(cpu);
1450 	unlock_vector_lock();
1451 	fixup_irqs();
1452 	lapic_offline();
1453 }
1454 
1455 int native_cpu_disable(void)
1456 {
1457 	int ret;
1458 
1459 	ret = lapic_can_unplug_cpu();
1460 	if (ret)
1461 		return ret;
1462 
1463 	cpu_disable_common();
1464 
1465         /*
1466          * Disable the local APIC. Otherwise IPI broadcasts will reach
1467          * it. It still responds normally to INIT, NMI, SMI, and SIPI
1468          * messages.
1469          *
1470          * Disabling the APIC must happen after cpu_disable_common()
1471          * which invokes fixup_irqs().
1472          *
1473          * Disabling the APIC preserves already set bits in IRR, but
1474          * an interrupt arriving after disabling the local APIC does not
1475          * set the corresponding IRR bit.
1476          *
1477          * fixup_irqs() scans IRR for set bits so it can raise a not
1478          * yet handled interrupt on the new destination CPU via an IPI
1479          * but obviously it can't do so for IRR bits which are not set.
1480          * IOW, interrupts arriving after disabling the local APIC will
1481          * be lost.
1482          */
1483 	apic_soft_disable();
1484 
1485 	return 0;
1486 }
1487 
1488 void play_dead_common(void)
1489 {
1490 	idle_task_exit();
1491 
1492 	cpuhp_ap_report_dead();
1493 
1494 	local_irq_disable();
1495 }
1496 
1497 /*
1498  * We need to flush the caches before going to sleep, lest we have
1499  * dirty data in our caches when we come back up.
1500  */
1501 static inline void mwait_play_dead(void)
1502 {
1503 	struct mwait_cpu_dead *md = this_cpu_ptr(&mwait_cpu_dead);
1504 	unsigned int eax, ebx, ecx, edx;
1505 	unsigned int highest_cstate = 0;
1506 	unsigned int highest_subcstate = 0;
1507 	int i;
1508 
1509 	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD ||
1510 	    boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
1511 		return;
1512 	if (!this_cpu_has(X86_FEATURE_MWAIT))
1513 		return;
1514 	if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1515 		return;
1516 	if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1517 		return;
1518 
1519 	eax = CPUID_MWAIT_LEAF;
1520 	ecx = 0;
1521 	native_cpuid(&eax, &ebx, &ecx, &edx);
1522 
1523 	/*
1524 	 * eax will be 0 if EDX enumeration is not valid.
1525 	 * Initialized below to cstate, sub_cstate value when EDX is valid.
1526 	 */
1527 	if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1528 		eax = 0;
1529 	} else {
1530 		edx >>= MWAIT_SUBSTATE_SIZE;
1531 		for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1532 			if (edx & MWAIT_SUBSTATE_MASK) {
1533 				highest_cstate = i;
1534 				highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1535 			}
1536 		}
1537 		eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1538 			(highest_subcstate - 1);
1539 	}
1540 
1541 	/* Set up state for the kexec() hack below */
1542 	md->status = CPUDEAD_MWAIT_WAIT;
1543 	md->control = CPUDEAD_MWAIT_WAIT;
1544 
1545 	wbinvd();
1546 
1547 	while (1) {
1548 		/*
1549 		 * The CLFLUSH is a workaround for erratum AAI65 for
1550 		 * the Xeon 7400 series.  It's not clear it is actually
1551 		 * needed, but it should be harmless in either case.
1552 		 * The WBINVD is insufficient due to the spurious-wakeup
1553 		 * case where we return around the loop.
1554 		 */
1555 		mb();
1556 		clflush(md);
1557 		mb();
1558 		__monitor(md, 0, 0);
1559 		mb();
1560 		__mwait(eax, 0);
1561 
1562 		if (READ_ONCE(md->control) == CPUDEAD_MWAIT_KEXEC_HLT) {
1563 			/*
1564 			 * Kexec is about to happen. Don't go back into mwait() as
1565 			 * the kexec kernel might overwrite text and data including
1566 			 * page tables and stack. So mwait() would resume when the
1567 			 * monitor cache line is written to and then the CPU goes
1568 			 * south due to overwritten text, page tables and stack.
1569 			 *
1570 			 * Note: This does _NOT_ protect against a stray MCE, NMI,
1571 			 * SMI. They will resume execution at the instruction
1572 			 * following the HLT instruction and run into the problem
1573 			 * which this is trying to prevent.
1574 			 */
1575 			WRITE_ONCE(md->status, CPUDEAD_MWAIT_KEXEC_HLT);
1576 			while(1)
1577 				native_halt();
1578 		}
1579 	}
1580 }
1581 
1582 /*
1583  * Kick all "offline" CPUs out of mwait on kexec(). See comment in
1584  * mwait_play_dead().
1585  */
1586 void smp_kick_mwait_play_dead(void)
1587 {
1588 	u32 newstate = CPUDEAD_MWAIT_KEXEC_HLT;
1589 	struct mwait_cpu_dead *md;
1590 	unsigned int cpu, i;
1591 
1592 	for_each_cpu_andnot(cpu, cpu_present_mask, cpu_online_mask) {
1593 		md = per_cpu_ptr(&mwait_cpu_dead, cpu);
1594 
1595 		/* Does it sit in mwait_play_dead() ? */
1596 		if (READ_ONCE(md->status) != CPUDEAD_MWAIT_WAIT)
1597 			continue;
1598 
1599 		/* Wait up to 5ms */
1600 		for (i = 0; READ_ONCE(md->status) != newstate && i < 1000; i++) {
1601 			/* Bring it out of mwait */
1602 			WRITE_ONCE(md->control, newstate);
1603 			udelay(5);
1604 		}
1605 
1606 		if (READ_ONCE(md->status) != newstate)
1607 			pr_err_once("CPU%u is stuck in mwait_play_dead()\n", cpu);
1608 	}
1609 }
1610 
1611 void __noreturn hlt_play_dead(void)
1612 {
1613 	if (__this_cpu_read(cpu_info.x86) >= 4)
1614 		wbinvd();
1615 
1616 	while (1)
1617 		native_halt();
1618 }
1619 
1620 void native_play_dead(void)
1621 {
1622 	play_dead_common();
1623 	tboot_shutdown(TB_SHUTDOWN_WFS);
1624 
1625 	mwait_play_dead();
1626 	if (cpuidle_play_dead())
1627 		hlt_play_dead();
1628 }
1629 
1630 #else /* ... !CONFIG_HOTPLUG_CPU */
1631 int native_cpu_disable(void)
1632 {
1633 	return -ENOSYS;
1634 }
1635 
1636 void native_play_dead(void)
1637 {
1638 	BUG();
1639 }
1640 
1641 #endif
1642