xref: /linux/arch/x86/kernel/smpboot.c (revision b0402403e54ae9eb94ce1cbb53c7def776e97426)
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 /* Per CPU bogomips and other parameters */
105 DEFINE_PER_CPU_READ_MOSTLY(struct cpuinfo_x86, cpu_info);
106 EXPORT_PER_CPU_SYMBOL(cpu_info);
107 
108 /* CPUs which are the primary SMT threads */
109 struct cpumask __cpu_primary_thread_mask __read_mostly;
110 
111 /* Representing CPUs for which sibling maps can be computed */
112 static cpumask_var_t cpu_sibling_setup_mask;
113 
114 struct mwait_cpu_dead {
115 	unsigned int	control;
116 	unsigned int	status;
117 };
118 
119 #define CPUDEAD_MWAIT_WAIT	0xDEADBEEF
120 #define CPUDEAD_MWAIT_KEXEC_HLT	0x4A17DEAD
121 
122 /*
123  * Cache line aligned data for mwait_play_dead(). Separate on purpose so
124  * that it's unlikely to be touched by other CPUs.
125  */
126 static DEFINE_PER_CPU_ALIGNED(struct mwait_cpu_dead, mwait_cpu_dead);
127 
128 /* Maximum number of SMT threads on any online core */
129 int __read_mostly __max_smt_threads = 1;
130 
131 /* Flag to indicate if a complete sched domain rebuild is required */
132 bool x86_topology_update;
133 
134 int arch_update_cpu_topology(void)
135 {
136 	int retval = x86_topology_update;
137 
138 	x86_topology_update = false;
139 	return retval;
140 }
141 
142 static unsigned int smpboot_warm_reset_vector_count;
143 
144 static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
145 {
146 	unsigned long flags;
147 
148 	spin_lock_irqsave(&rtc_lock, flags);
149 	if (!smpboot_warm_reset_vector_count++) {
150 		CMOS_WRITE(0xa, 0xf);
151 		*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) = start_eip >> 4;
152 		*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = start_eip & 0xf;
153 	}
154 	spin_unlock_irqrestore(&rtc_lock, flags);
155 }
156 
157 static inline void smpboot_restore_warm_reset_vector(void)
158 {
159 	unsigned long flags;
160 
161 	/*
162 	 * Paranoid:  Set warm reset code and vector here back
163 	 * to default values.
164 	 */
165 	spin_lock_irqsave(&rtc_lock, flags);
166 	if (!--smpboot_warm_reset_vector_count) {
167 		CMOS_WRITE(0, 0xf);
168 		*((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
169 	}
170 	spin_unlock_irqrestore(&rtc_lock, flags);
171 
172 }
173 
174 /* Run the next set of setup steps for the upcoming CPU */
175 static void ap_starting(void)
176 {
177 	int cpuid = smp_processor_id();
178 
179 	/* Mop up eventual mwait_play_dead() wreckage */
180 	this_cpu_write(mwait_cpu_dead.status, 0);
181 	this_cpu_write(mwait_cpu_dead.control, 0);
182 
183 	/*
184 	 * If woken up by an INIT in an 82489DX configuration the alive
185 	 * synchronization guarantees that the CPU does not reach this
186 	 * point before an INIT_deassert IPI reaches the local APIC, so it
187 	 * is now safe to touch the local APIC.
188 	 *
189 	 * Set up this CPU, first the APIC, which is probably redundant on
190 	 * most boards.
191 	 */
192 	apic_ap_setup();
193 
194 	/* Save the processor parameters. */
195 	smp_store_cpu_info(cpuid);
196 
197 	/*
198 	 * The topology information must be up to date before
199 	 * notify_cpu_starting().
200 	 */
201 	set_cpu_sibling_map(cpuid);
202 
203 	ap_init_aperfmperf();
204 
205 	pr_debug("Stack at about %p\n", &cpuid);
206 
207 	wmb();
208 
209 	/*
210 	 * This runs the AP through all the cpuhp states to its target
211 	 * state CPUHP_ONLINE.
212 	 */
213 	notify_cpu_starting(cpuid);
214 }
215 
216 static void ap_calibrate_delay(void)
217 {
218 	/*
219 	 * Calibrate the delay loop and update loops_per_jiffy in cpu_data.
220 	 * smp_store_cpu_info() stored a value that is close but not as
221 	 * accurate as the value just calculated.
222 	 *
223 	 * As this is invoked after the TSC synchronization check,
224 	 * calibrate_delay_is_known() will skip the calibration routine
225 	 * when TSC is synchronized across sockets.
226 	 */
227 	calibrate_delay();
228 	cpu_data(smp_processor_id()).loops_per_jiffy = loops_per_jiffy;
229 }
230 
231 /*
232  * Activate a secondary processor.
233  */
234 static void notrace start_secondary(void *unused)
235 {
236 	/*
237 	 * Don't put *anything* except direct CPU state initialization
238 	 * before cpu_init(), SMP booting is too fragile that we want to
239 	 * limit the things done here to the most necessary things.
240 	 */
241 	cr4_init();
242 
243 	/*
244 	 * 32-bit specific. 64-bit reaches this code with the correct page
245 	 * table established. Yet another historical divergence.
246 	 */
247 	if (IS_ENABLED(CONFIG_X86_32)) {
248 		/* switch away from the initial page table */
249 		load_cr3(swapper_pg_dir);
250 		__flush_tlb_all();
251 	}
252 
253 	cpu_init_exception_handling();
254 
255 	/*
256 	 * Load the microcode before reaching the AP alive synchronization
257 	 * point below so it is not part of the full per CPU serialized
258 	 * bringup part when "parallel" bringup is enabled.
259 	 *
260 	 * That's even safe when hyperthreading is enabled in the CPU as
261 	 * the core code starts the primary threads first and leaves the
262 	 * secondary threads waiting for SIPI. Loading microcode on
263 	 * physical cores concurrently is a safe operation.
264 	 *
265 	 * This covers both the Intel specific issue that concurrent
266 	 * microcode loading on SMT siblings must be prohibited and the
267 	 * vendor independent issue`that microcode loading which changes
268 	 * CPUID, MSRs etc. must be strictly serialized to maintain
269 	 * software state correctness.
270 	 */
271 	load_ucode_ap();
272 
273 	/*
274 	 * Synchronization point with the hotplug core. Sets this CPUs
275 	 * synchronization state to ALIVE and spin-waits for the control CPU to
276 	 * release this CPU for further bringup.
277 	 */
278 	cpuhp_ap_sync_alive();
279 
280 	cpu_init();
281 	fpu__init_cpu();
282 	rcutree_report_cpu_starting(raw_smp_processor_id());
283 	x86_cpuinit.early_percpu_clock_init();
284 
285 	ap_starting();
286 
287 	/* Check TSC synchronization with the control CPU. */
288 	check_tsc_sync_target();
289 
290 	/*
291 	 * Calibrate the delay loop after the TSC synchronization check.
292 	 * This allows to skip the calibration when TSC is synchronized
293 	 * across sockets.
294 	 */
295 	ap_calibrate_delay();
296 
297 	speculative_store_bypass_ht_init();
298 
299 	/*
300 	 * Lock vector_lock, set CPU online and bring the vector
301 	 * allocator online. Online must be set with vector_lock held
302 	 * to prevent a concurrent irq setup/teardown from seeing a
303 	 * half valid vector space.
304 	 */
305 	lock_vector_lock();
306 	set_cpu_online(smp_processor_id(), true);
307 	lapic_online();
308 	unlock_vector_lock();
309 	x86_platform.nmi_init();
310 
311 	/* enable local interrupts */
312 	local_irq_enable();
313 
314 	x86_cpuinit.setup_percpu_clockev();
315 
316 	wmb();
317 	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
318 }
319 
320 static void __init smp_store_boot_cpu_info(void)
321 {
322 	struct cpuinfo_x86 *c = &cpu_data(0);
323 
324 	*c = boot_cpu_data;
325 	c->initialized = true;
326 }
327 
328 /*
329  * The bootstrap kernel entry code has set these up. Save them for
330  * a given CPU
331  */
332 void smp_store_cpu_info(int id)
333 {
334 	struct cpuinfo_x86 *c = &cpu_data(id);
335 
336 	/* Copy boot_cpu_data only on the first bringup */
337 	if (!c->initialized)
338 		*c = boot_cpu_data;
339 	c->cpu_index = id;
340 	/*
341 	 * During boot time, CPU0 has this setup already. Save the info when
342 	 * bringing up an AP.
343 	 */
344 	identify_secondary_cpu(c);
345 	c->initialized = true;
346 }
347 
348 static bool
349 topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
350 {
351 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
352 
353 	return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
354 }
355 
356 static bool
357 topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
358 {
359 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
360 
361 	return !WARN_ONCE(!topology_same_node(c, o),
362 		"sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
363 		"[node: %d != %d]. Ignoring dependency.\n",
364 		cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
365 }
366 
367 #define link_mask(mfunc, c1, c2)					\
368 do {									\
369 	cpumask_set_cpu((c1), mfunc(c2));				\
370 	cpumask_set_cpu((c2), mfunc(c1));				\
371 } while (0)
372 
373 static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
374 {
375 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
376 		int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
377 
378 		if (c->topo.pkg_id == o->topo.pkg_id &&
379 		    c->topo.die_id == o->topo.die_id &&
380 		    c->topo.amd_node_id == o->topo.amd_node_id &&
381 		    per_cpu_llc_id(cpu1) == per_cpu_llc_id(cpu2)) {
382 			if (c->topo.core_id == o->topo.core_id)
383 				return topology_sane(c, o, "smt");
384 
385 			if ((c->topo.cu_id != 0xff) &&
386 			    (o->topo.cu_id != 0xff) &&
387 			    (c->topo.cu_id == o->topo.cu_id))
388 				return topology_sane(c, o, "smt");
389 		}
390 
391 	} else if (c->topo.pkg_id == o->topo.pkg_id &&
392 		   c->topo.die_id == o->topo.die_id &&
393 		   c->topo.core_id == o->topo.core_id) {
394 		return topology_sane(c, o, "smt");
395 	}
396 
397 	return false;
398 }
399 
400 static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
401 {
402 	if (c->topo.pkg_id != o->topo.pkg_id || c->topo.die_id != o->topo.die_id)
403 		return false;
404 
405 	if (cpu_feature_enabled(X86_FEATURE_TOPOEXT) && topology_amd_nodes_per_pkg() > 1)
406 		return c->topo.amd_node_id == o->topo.amd_node_id;
407 
408 	return true;
409 }
410 
411 static bool match_l2c(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
412 {
413 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
414 
415 	/* If the arch didn't set up l2c_id, fall back to SMT */
416 	if (per_cpu_l2c_id(cpu1) == BAD_APICID)
417 		return match_smt(c, o);
418 
419 	/* Do not match if L2 cache id does not match: */
420 	if (per_cpu_l2c_id(cpu1) != per_cpu_l2c_id(cpu2))
421 		return false;
422 
423 	return topology_sane(c, o, "l2c");
424 }
425 
426 /*
427  * Unlike the other levels, we do not enforce keeping a
428  * multicore group inside a NUMA node.  If this happens, we will
429  * discard the MC level of the topology later.
430  */
431 static bool match_pkg(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
432 {
433 	if (c->topo.pkg_id == o->topo.pkg_id)
434 		return true;
435 	return false;
436 }
437 
438 /*
439  * Define intel_cod_cpu[] for Intel COD (Cluster-on-Die) CPUs.
440  *
441  * Any Intel CPU that has multiple nodes per package and does not
442  * match intel_cod_cpu[] has the SNC (Sub-NUMA Cluster) topology.
443  *
444  * When in SNC mode, these CPUs enumerate an LLC that is shared
445  * by multiple NUMA nodes. The LLC is shared for off-package data
446  * access but private to the NUMA node (half of the package) for
447  * on-package access. CPUID (the source of the information about
448  * the LLC) can only enumerate the cache as shared or unshared,
449  * but not this particular configuration.
450  */
451 
452 static const struct x86_cpu_id intel_cod_cpu[] = {
453 	X86_MATCH_INTEL_FAM6_MODEL(HASWELL_X, 0),	/* COD */
454 	X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_X, 0),	/* COD */
455 	X86_MATCH_INTEL_FAM6_MODEL(ANY, 1),		/* SNC */
456 	{}
457 };
458 
459 static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
460 {
461 	const struct x86_cpu_id *id = x86_match_cpu(intel_cod_cpu);
462 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
463 	bool intel_snc = id && id->driver_data;
464 
465 	/* Do not match if we do not have a valid APICID for cpu: */
466 	if (per_cpu_llc_id(cpu1) == BAD_APICID)
467 		return false;
468 
469 	/* Do not match if LLC id does not match: */
470 	if (per_cpu_llc_id(cpu1) != per_cpu_llc_id(cpu2))
471 		return false;
472 
473 	/*
474 	 * Allow the SNC topology without warning. Return of false
475 	 * means 'c' does not share the LLC of 'o'. This will be
476 	 * reflected to userspace.
477 	 */
478 	if (match_pkg(c, o) && !topology_same_node(c, o) && intel_snc)
479 		return false;
480 
481 	return topology_sane(c, o, "llc");
482 }
483 
484 
485 static inline int x86_sched_itmt_flags(void)
486 {
487 	return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
488 }
489 
490 #ifdef CONFIG_SCHED_MC
491 static int x86_core_flags(void)
492 {
493 	return cpu_core_flags() | x86_sched_itmt_flags();
494 }
495 #endif
496 #ifdef CONFIG_SCHED_SMT
497 static int x86_smt_flags(void)
498 {
499 	return cpu_smt_flags();
500 }
501 #endif
502 #ifdef CONFIG_SCHED_CLUSTER
503 static int x86_cluster_flags(void)
504 {
505 	return cpu_cluster_flags() | x86_sched_itmt_flags();
506 }
507 #endif
508 
509 static int x86_die_flags(void)
510 {
511 	if (cpu_feature_enabled(X86_FEATURE_HYBRID_CPU))
512 	       return x86_sched_itmt_flags();
513 
514 	return 0;
515 }
516 
517 /*
518  * Set if a package/die has multiple NUMA nodes inside.
519  * AMD Magny-Cours, Intel Cluster-on-Die, and Intel
520  * Sub-NUMA Clustering have this.
521  */
522 static bool x86_has_numa_in_package;
523 
524 static struct sched_domain_topology_level x86_topology[6];
525 
526 static void __init build_sched_topology(void)
527 {
528 	int i = 0;
529 
530 #ifdef CONFIG_SCHED_SMT
531 	x86_topology[i++] = (struct sched_domain_topology_level){
532 		cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT)
533 	};
534 #endif
535 #ifdef CONFIG_SCHED_CLUSTER
536 	x86_topology[i++] = (struct sched_domain_topology_level){
537 		cpu_clustergroup_mask, x86_cluster_flags, SD_INIT_NAME(CLS)
538 	};
539 #endif
540 #ifdef CONFIG_SCHED_MC
541 	x86_topology[i++] = (struct sched_domain_topology_level){
542 		cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC)
543 	};
544 #endif
545 	/*
546 	 * When there is NUMA topology inside the package skip the PKG domain
547 	 * since the NUMA domains will auto-magically create the right spanning
548 	 * domains based on the SLIT.
549 	 */
550 	if (!x86_has_numa_in_package) {
551 		x86_topology[i++] = (struct sched_domain_topology_level){
552 			cpu_cpu_mask, x86_die_flags, SD_INIT_NAME(PKG)
553 		};
554 	}
555 
556 	/*
557 	 * There must be one trailing NULL entry left.
558 	 */
559 	BUG_ON(i >= ARRAY_SIZE(x86_topology)-1);
560 
561 	set_sched_topology(x86_topology);
562 }
563 
564 void set_cpu_sibling_map(int cpu)
565 {
566 	bool has_smt = __max_threads_per_core > 1;
567 	bool has_mp = has_smt || topology_num_cores_per_package() > 1;
568 	struct cpuinfo_x86 *c = &cpu_data(cpu);
569 	struct cpuinfo_x86 *o;
570 	int i, threads;
571 
572 	cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
573 
574 	if (!has_mp) {
575 		cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
576 		cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
577 		cpumask_set_cpu(cpu, cpu_l2c_shared_mask(cpu));
578 		cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
579 		cpumask_set_cpu(cpu, topology_die_cpumask(cpu));
580 		c->booted_cores = 1;
581 		return;
582 	}
583 
584 	for_each_cpu(i, cpu_sibling_setup_mask) {
585 		o = &cpu_data(i);
586 
587 		if (match_pkg(c, o) && !topology_same_node(c, o))
588 			x86_has_numa_in_package = true;
589 
590 		if ((i == cpu) || (has_smt && match_smt(c, o)))
591 			link_mask(topology_sibling_cpumask, cpu, i);
592 
593 		if ((i == cpu) || (has_mp && match_llc(c, o)))
594 			link_mask(cpu_llc_shared_mask, cpu, i);
595 
596 		if ((i == cpu) || (has_mp && match_l2c(c, o)))
597 			link_mask(cpu_l2c_shared_mask, cpu, i);
598 
599 		if ((i == cpu) || (has_mp && match_die(c, o)))
600 			link_mask(topology_die_cpumask, cpu, i);
601 	}
602 
603 	threads = cpumask_weight(topology_sibling_cpumask(cpu));
604 	if (threads > __max_smt_threads)
605 		__max_smt_threads = threads;
606 
607 	for_each_cpu(i, topology_sibling_cpumask(cpu))
608 		cpu_data(i).smt_active = threads > 1;
609 
610 	/*
611 	 * This needs a separate iteration over the cpus because we rely on all
612 	 * topology_sibling_cpumask links to be set-up.
613 	 */
614 	for_each_cpu(i, cpu_sibling_setup_mask) {
615 		o = &cpu_data(i);
616 
617 		if ((i == cpu) || (has_mp && match_pkg(c, o))) {
618 			link_mask(topology_core_cpumask, cpu, i);
619 
620 			/*
621 			 *  Does this new cpu bringup a new core?
622 			 */
623 			if (threads == 1) {
624 				/*
625 				 * for each core in package, increment
626 				 * the booted_cores for this new cpu
627 				 */
628 				if (cpumask_first(
629 				    topology_sibling_cpumask(i)) == i)
630 					c->booted_cores++;
631 				/*
632 				 * increment the core count for all
633 				 * the other cpus in this package
634 				 */
635 				if (i != cpu)
636 					cpu_data(i).booted_cores++;
637 			} else if (i != cpu && !c->booted_cores)
638 				c->booted_cores = cpu_data(i).booted_cores;
639 		}
640 	}
641 }
642 
643 /* maps the cpu to the sched domain representing multi-core */
644 const struct cpumask *cpu_coregroup_mask(int cpu)
645 {
646 	return cpu_llc_shared_mask(cpu);
647 }
648 
649 const struct cpumask *cpu_clustergroup_mask(int cpu)
650 {
651 	return cpu_l2c_shared_mask(cpu);
652 }
653 EXPORT_SYMBOL_GPL(cpu_clustergroup_mask);
654 
655 static void impress_friends(void)
656 {
657 	int cpu;
658 	unsigned long bogosum = 0;
659 	/*
660 	 * Allow the user to impress friends.
661 	 */
662 	pr_debug("Before bogomips\n");
663 	for_each_online_cpu(cpu)
664 		bogosum += cpu_data(cpu).loops_per_jiffy;
665 
666 	pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
667 		num_online_cpus(),
668 		bogosum/(500000/HZ),
669 		(bogosum/(5000/HZ))%100);
670 
671 	pr_debug("Before bogocount - setting activated=1\n");
672 }
673 
674 /*
675  * The Multiprocessor Specification 1.4 (1997) example code suggests
676  * that there should be a 10ms delay between the BSP asserting INIT
677  * and de-asserting INIT, when starting a remote processor.
678  * But that slows boot and resume on modern processors, which include
679  * many cores and don't require that delay.
680  *
681  * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
682  * Modern processor families are quirked to remove the delay entirely.
683  */
684 #define UDELAY_10MS_DEFAULT 10000
685 
686 static unsigned int init_udelay = UINT_MAX;
687 
688 static int __init cpu_init_udelay(char *str)
689 {
690 	get_option(&str, &init_udelay);
691 
692 	return 0;
693 }
694 early_param("cpu_init_udelay", cpu_init_udelay);
695 
696 static void __init smp_quirk_init_udelay(void)
697 {
698 	/* if cmdline changed it from default, leave it alone */
699 	if (init_udelay != UINT_MAX)
700 		return;
701 
702 	/* if modern processor, use no delay */
703 	if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
704 	    ((boot_cpu_data.x86_vendor == X86_VENDOR_HYGON) && (boot_cpu_data.x86 >= 0x18)) ||
705 	    ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
706 		init_udelay = 0;
707 		return;
708 	}
709 	/* else, use legacy delay */
710 	init_udelay = UDELAY_10MS_DEFAULT;
711 }
712 
713 /*
714  * Wake up AP by INIT, INIT, STARTUP sequence.
715  */
716 static void send_init_sequence(u32 phys_apicid)
717 {
718 	int maxlvt = lapic_get_maxlvt();
719 
720 	/* Be paranoid about clearing APIC errors. */
721 	if (APIC_INTEGRATED(boot_cpu_apic_version)) {
722 		/* Due to the Pentium erratum 3AP.  */
723 		if (maxlvt > 3)
724 			apic_write(APIC_ESR, 0);
725 		apic_read(APIC_ESR);
726 	}
727 
728 	/* Assert INIT on the target CPU */
729 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT, phys_apicid);
730 	safe_apic_wait_icr_idle();
731 
732 	udelay(init_udelay);
733 
734 	/* Deassert INIT on the target CPU */
735 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
736 	safe_apic_wait_icr_idle();
737 }
738 
739 /*
740  * Wake up AP by INIT, INIT, STARTUP sequence.
741  */
742 static int wakeup_secondary_cpu_via_init(u32 phys_apicid, unsigned long start_eip)
743 {
744 	unsigned long send_status = 0, accept_status = 0;
745 	int num_starts, j, maxlvt;
746 
747 	preempt_disable();
748 	maxlvt = lapic_get_maxlvt();
749 	send_init_sequence(phys_apicid);
750 
751 	mb();
752 
753 	/*
754 	 * Should we send STARTUP IPIs ?
755 	 *
756 	 * Determine this based on the APIC version.
757 	 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
758 	 */
759 	if (APIC_INTEGRATED(boot_cpu_apic_version))
760 		num_starts = 2;
761 	else
762 		num_starts = 0;
763 
764 	/*
765 	 * Run STARTUP IPI loop.
766 	 */
767 	pr_debug("#startup loops: %d\n", num_starts);
768 
769 	for (j = 1; j <= num_starts; j++) {
770 		pr_debug("Sending STARTUP #%d\n", j);
771 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
772 			apic_write(APIC_ESR, 0);
773 		apic_read(APIC_ESR);
774 		pr_debug("After apic_write\n");
775 
776 		/*
777 		 * STARTUP IPI
778 		 */
779 
780 		/* Target chip */
781 		/* Boot on the stack */
782 		/* Kick the second */
783 		apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
784 			       phys_apicid);
785 
786 		/*
787 		 * Give the other CPU some time to accept the IPI.
788 		 */
789 		if (init_udelay == 0)
790 			udelay(10);
791 		else
792 			udelay(300);
793 
794 		pr_debug("Startup point 1\n");
795 
796 		pr_debug("Waiting for send to finish...\n");
797 		send_status = safe_apic_wait_icr_idle();
798 
799 		/*
800 		 * Give the other CPU some time to accept the IPI.
801 		 */
802 		if (init_udelay == 0)
803 			udelay(10);
804 		else
805 			udelay(200);
806 
807 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
808 			apic_write(APIC_ESR, 0);
809 		accept_status = (apic_read(APIC_ESR) & 0xEF);
810 		if (send_status || accept_status)
811 			break;
812 	}
813 	pr_debug("After Startup\n");
814 
815 	if (send_status)
816 		pr_err("APIC never delivered???\n");
817 	if (accept_status)
818 		pr_err("APIC delivery error (%lx)\n", accept_status);
819 
820 	preempt_enable();
821 	return (send_status | accept_status);
822 }
823 
824 /* reduce the number of lines printed when booting a large cpu count system */
825 static void announce_cpu(int cpu, int apicid)
826 {
827 	static int width, node_width, first = 1;
828 	static int current_node = NUMA_NO_NODE;
829 	int node = early_cpu_to_node(cpu);
830 
831 	if (!width)
832 		width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
833 
834 	if (!node_width)
835 		node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
836 
837 	if (system_state < SYSTEM_RUNNING) {
838 		if (first)
839 			pr_info("x86: Booting SMP configuration:\n");
840 
841 		if (node != current_node) {
842 			if (current_node > (-1))
843 				pr_cont("\n");
844 			current_node = node;
845 
846 			printk(KERN_INFO ".... node %*s#%d, CPUs:  ",
847 			       node_width - num_digits(node), " ", node);
848 		}
849 
850 		/* Add padding for the BSP */
851 		if (first)
852 			pr_cont("%*s", width + 1, " ");
853 		first = 0;
854 
855 		pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
856 	} else
857 		pr_info("Booting Node %d Processor %d APIC 0x%x\n",
858 			node, cpu, apicid);
859 }
860 
861 int common_cpu_up(unsigned int cpu, struct task_struct *idle)
862 {
863 	int ret;
864 
865 	/* Just in case we booted with a single CPU. */
866 	alternatives_enable_smp();
867 
868 	per_cpu(pcpu_hot.current_task, cpu) = idle;
869 	cpu_init_stack_canary(cpu, idle);
870 
871 	/* Initialize the interrupt stack(s) */
872 	ret = irq_init_percpu_irqstack(cpu);
873 	if (ret)
874 		return ret;
875 
876 #ifdef CONFIG_X86_32
877 	/* Stack for startup_32 can be just as for start_secondary onwards */
878 	per_cpu(pcpu_hot.top_of_stack, cpu) = task_top_of_stack(idle);
879 #endif
880 	return 0;
881 }
882 
883 /*
884  * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
885  * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
886  * Returns zero if startup was successfully sent, else error code from
887  * ->wakeup_secondary_cpu.
888  */
889 static int do_boot_cpu(u32 apicid, int cpu, struct task_struct *idle)
890 {
891 	unsigned long start_ip = real_mode_header->trampoline_start;
892 	int ret;
893 
894 #ifdef CONFIG_X86_64
895 	/* If 64-bit wakeup method exists, use the 64-bit mode trampoline IP */
896 	if (apic->wakeup_secondary_cpu_64)
897 		start_ip = real_mode_header->trampoline_start64;
898 #endif
899 	idle->thread.sp = (unsigned long)task_pt_regs(idle);
900 	initial_code = (unsigned long)start_secondary;
901 
902 	if (IS_ENABLED(CONFIG_X86_32)) {
903 		early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
904 		initial_stack  = idle->thread.sp;
905 	} else if (!(smpboot_control & STARTUP_PARALLEL_MASK)) {
906 		smpboot_control = cpu;
907 	}
908 
909 	/* Enable the espfix hack for this CPU */
910 	init_espfix_ap(cpu);
911 
912 	/* So we see what's up */
913 	announce_cpu(cpu, apicid);
914 
915 	/*
916 	 * This grunge runs the startup process for
917 	 * the targeted processor.
918 	 */
919 	if (x86_platform.legacy.warm_reset) {
920 
921 		pr_debug("Setting warm reset code and vector.\n");
922 
923 		smpboot_setup_warm_reset_vector(start_ip);
924 		/*
925 		 * Be paranoid about clearing APIC errors.
926 		*/
927 		if (APIC_INTEGRATED(boot_cpu_apic_version)) {
928 			apic_write(APIC_ESR, 0);
929 			apic_read(APIC_ESR);
930 		}
931 	}
932 
933 	smp_mb();
934 
935 	/*
936 	 * Wake up a CPU in difference cases:
937 	 * - Use a method from the APIC driver if one defined, with wakeup
938 	 *   straight to 64-bit mode preferred over wakeup to RM.
939 	 * Otherwise,
940 	 * - Use an INIT boot APIC message
941 	 */
942 	if (apic->wakeup_secondary_cpu_64)
943 		ret = apic->wakeup_secondary_cpu_64(apicid, start_ip);
944 	else if (apic->wakeup_secondary_cpu)
945 		ret = apic->wakeup_secondary_cpu(apicid, start_ip);
946 	else
947 		ret = wakeup_secondary_cpu_via_init(apicid, start_ip);
948 
949 	/* If the wakeup mechanism failed, cleanup the warm reset vector */
950 	if (ret)
951 		arch_cpuhp_cleanup_kick_cpu(cpu);
952 	return ret;
953 }
954 
955 int native_kick_ap(unsigned int cpu, struct task_struct *tidle)
956 {
957 	u32 apicid = apic->cpu_present_to_apicid(cpu);
958 	int err;
959 
960 	lockdep_assert_irqs_enabled();
961 
962 	pr_debug("++++++++++++++++++++=_---CPU UP  %u\n", cpu);
963 
964 	if (apicid == BAD_APICID || !apic_id_valid(apicid)) {
965 		pr_err("CPU %u has invalid APIC ID %x. Aborting bringup\n", cpu, apicid);
966 		return -EINVAL;
967 	}
968 
969 	if (!test_bit(apicid, phys_cpu_present_map)) {
970 		pr_err("CPU %u APIC ID %x is not present. Aborting bringup\n", cpu, apicid);
971 		return -EINVAL;
972 	}
973 
974 	/*
975 	 * Save current MTRR state in case it was changed since early boot
976 	 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
977 	 */
978 	mtrr_save_state();
979 
980 	/* the FPU context is blank, nobody can own it */
981 	per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
982 
983 	err = common_cpu_up(cpu, tidle);
984 	if (err)
985 		return err;
986 
987 	err = do_boot_cpu(apicid, cpu, tidle);
988 	if (err)
989 		pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
990 
991 	return err;
992 }
993 
994 int arch_cpuhp_kick_ap_alive(unsigned int cpu, struct task_struct *tidle)
995 {
996 	return smp_ops.kick_ap_alive(cpu, tidle);
997 }
998 
999 void arch_cpuhp_cleanup_kick_cpu(unsigned int cpu)
1000 {
1001 	/* Cleanup possible dangling ends... */
1002 	if (smp_ops.kick_ap_alive == native_kick_ap && x86_platform.legacy.warm_reset)
1003 		smpboot_restore_warm_reset_vector();
1004 }
1005 
1006 void arch_cpuhp_cleanup_dead_cpu(unsigned int cpu)
1007 {
1008 	if (smp_ops.cleanup_dead_cpu)
1009 		smp_ops.cleanup_dead_cpu(cpu);
1010 
1011 	if (system_state == SYSTEM_RUNNING)
1012 		pr_info("CPU %u is now offline\n", cpu);
1013 }
1014 
1015 void arch_cpuhp_sync_state_poll(void)
1016 {
1017 	if (smp_ops.poll_sync_state)
1018 		smp_ops.poll_sync_state();
1019 }
1020 
1021 /**
1022  * arch_disable_smp_support() - Disables SMP support for x86 at boottime
1023  */
1024 void __init arch_disable_smp_support(void)
1025 {
1026 	disable_ioapic_support();
1027 }
1028 
1029 /*
1030  * Fall back to non SMP mode after errors.
1031  *
1032  * RED-PEN audit/test this more. I bet there is more state messed up here.
1033  */
1034 static __init void disable_smp(void)
1035 {
1036 	pr_info("SMP disabled\n");
1037 
1038 	disable_ioapic_support();
1039 	topology_reset_possible_cpus_up();
1040 
1041 	cpumask_set_cpu(0, topology_sibling_cpumask(0));
1042 	cpumask_set_cpu(0, topology_core_cpumask(0));
1043 	cpumask_set_cpu(0, topology_die_cpumask(0));
1044 }
1045 
1046 static void __init smp_cpu_index_default(void)
1047 {
1048 	int i;
1049 	struct cpuinfo_x86 *c;
1050 
1051 	for_each_possible_cpu(i) {
1052 		c = &cpu_data(i);
1053 		/* mark all to hotplug */
1054 		c->cpu_index = nr_cpu_ids;
1055 	}
1056 }
1057 
1058 void __init smp_prepare_cpus_common(void)
1059 {
1060 	unsigned int i;
1061 
1062 	smp_cpu_index_default();
1063 
1064 	/*
1065 	 * Setup boot CPU information
1066 	 */
1067 	smp_store_boot_cpu_info(); /* Final full version of the data */
1068 	mb();
1069 
1070 	for_each_possible_cpu(i) {
1071 		zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1072 		zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1073 		zalloc_cpumask_var(&per_cpu(cpu_die_map, i), GFP_KERNEL);
1074 		zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1075 		zalloc_cpumask_var(&per_cpu(cpu_l2c_shared_map, i), GFP_KERNEL);
1076 	}
1077 
1078 	set_cpu_sibling_map(0);
1079 }
1080 
1081 #ifdef CONFIG_X86_64
1082 /* Establish whether parallel bringup can be supported. */
1083 bool __init arch_cpuhp_init_parallel_bringup(void)
1084 {
1085 	if (!x86_cpuinit.parallel_bringup) {
1086 		pr_info("Parallel CPU startup disabled by the platform\n");
1087 		return false;
1088 	}
1089 
1090 	smpboot_control = STARTUP_READ_APICID;
1091 	pr_debug("Parallel CPU startup enabled: 0x%08x\n", smpboot_control);
1092 	return true;
1093 }
1094 #endif
1095 
1096 /*
1097  * Prepare for SMP bootup.
1098  * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1099  *            for common interface support.
1100  */
1101 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1102 {
1103 	smp_prepare_cpus_common();
1104 
1105 	switch (apic_intr_mode) {
1106 	case APIC_PIC:
1107 	case APIC_VIRTUAL_WIRE_NO_CONFIG:
1108 		disable_smp();
1109 		return;
1110 	case APIC_SYMMETRIC_IO_NO_ROUTING:
1111 		disable_smp();
1112 		/* Setup local timer */
1113 		x86_init.timers.setup_percpu_clockev();
1114 		return;
1115 	case APIC_VIRTUAL_WIRE:
1116 	case APIC_SYMMETRIC_IO:
1117 		break;
1118 	}
1119 
1120 	/* Setup local timer */
1121 	x86_init.timers.setup_percpu_clockev();
1122 
1123 	pr_info("CPU0: ");
1124 	print_cpu_info(&cpu_data(0));
1125 
1126 	uv_system_init();
1127 
1128 	smp_quirk_init_udelay();
1129 
1130 	speculative_store_bypass_ht_init();
1131 
1132 	snp_set_wakeup_secondary_cpu();
1133 }
1134 
1135 void arch_thaw_secondary_cpus_begin(void)
1136 {
1137 	set_cache_aps_delayed_init(true);
1138 }
1139 
1140 void arch_thaw_secondary_cpus_end(void)
1141 {
1142 	cache_aps_init();
1143 }
1144 
1145 /*
1146  * Early setup to make printk work.
1147  */
1148 void __init native_smp_prepare_boot_cpu(void)
1149 {
1150 	int me = smp_processor_id();
1151 
1152 	/* SMP handles this from setup_per_cpu_areas() */
1153 	if (!IS_ENABLED(CONFIG_SMP))
1154 		switch_gdt_and_percpu_base(me);
1155 
1156 	native_pv_lock_init();
1157 }
1158 
1159 void __init native_smp_cpus_done(unsigned int max_cpus)
1160 {
1161 	pr_debug("Boot done\n");
1162 
1163 	build_sched_topology();
1164 	nmi_selftest();
1165 	impress_friends();
1166 	cache_aps_init();
1167 }
1168 
1169 /* correctly size the local cpu masks */
1170 void __init setup_cpu_local_masks(void)
1171 {
1172 	alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
1173 }
1174 
1175 #ifdef CONFIG_HOTPLUG_CPU
1176 
1177 /* Recompute SMT state for all CPUs on offline */
1178 static void recompute_smt_state(void)
1179 {
1180 	int max_threads, cpu;
1181 
1182 	max_threads = 0;
1183 	for_each_online_cpu (cpu) {
1184 		int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1185 
1186 		if (threads > max_threads)
1187 			max_threads = threads;
1188 	}
1189 	__max_smt_threads = max_threads;
1190 }
1191 
1192 static void remove_siblinginfo(int cpu)
1193 {
1194 	int sibling;
1195 	struct cpuinfo_x86 *c = &cpu_data(cpu);
1196 
1197 	for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1198 		cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1199 		/*/
1200 		 * last thread sibling in this cpu core going down
1201 		 */
1202 		if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1203 			cpu_data(sibling).booted_cores--;
1204 	}
1205 
1206 	for_each_cpu(sibling, topology_die_cpumask(cpu))
1207 		cpumask_clear_cpu(cpu, topology_die_cpumask(sibling));
1208 
1209 	for_each_cpu(sibling, topology_sibling_cpumask(cpu)) {
1210 		cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1211 		if (cpumask_weight(topology_sibling_cpumask(sibling)) == 1)
1212 			cpu_data(sibling).smt_active = false;
1213 	}
1214 
1215 	for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1216 		cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1217 	for_each_cpu(sibling, cpu_l2c_shared_mask(cpu))
1218 		cpumask_clear_cpu(cpu, cpu_l2c_shared_mask(sibling));
1219 	cpumask_clear(cpu_llc_shared_mask(cpu));
1220 	cpumask_clear(cpu_l2c_shared_mask(cpu));
1221 	cpumask_clear(topology_sibling_cpumask(cpu));
1222 	cpumask_clear(topology_core_cpumask(cpu));
1223 	cpumask_clear(topology_die_cpumask(cpu));
1224 	c->topo.core_id = 0;
1225 	c->booted_cores = 0;
1226 	cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1227 	recompute_smt_state();
1228 }
1229 
1230 static void remove_cpu_from_maps(int cpu)
1231 {
1232 	set_cpu_online(cpu, false);
1233 	numa_remove_cpu(cpu);
1234 }
1235 
1236 void cpu_disable_common(void)
1237 {
1238 	int cpu = smp_processor_id();
1239 
1240 	remove_siblinginfo(cpu);
1241 
1242 	/* It's now safe to remove this processor from the online map */
1243 	lock_vector_lock();
1244 	remove_cpu_from_maps(cpu);
1245 	unlock_vector_lock();
1246 	fixup_irqs();
1247 	lapic_offline();
1248 }
1249 
1250 int native_cpu_disable(void)
1251 {
1252 	int ret;
1253 
1254 	ret = lapic_can_unplug_cpu();
1255 	if (ret)
1256 		return ret;
1257 
1258 	cpu_disable_common();
1259 
1260         /*
1261          * Disable the local APIC. Otherwise IPI broadcasts will reach
1262          * it. It still responds normally to INIT, NMI, SMI, and SIPI
1263          * messages.
1264          *
1265          * Disabling the APIC must happen after cpu_disable_common()
1266          * which invokes fixup_irqs().
1267          *
1268          * Disabling the APIC preserves already set bits in IRR, but
1269          * an interrupt arriving after disabling the local APIC does not
1270          * set the corresponding IRR bit.
1271          *
1272          * fixup_irqs() scans IRR for set bits so it can raise a not
1273          * yet handled interrupt on the new destination CPU via an IPI
1274          * but obviously it can't do so for IRR bits which are not set.
1275          * IOW, interrupts arriving after disabling the local APIC will
1276          * be lost.
1277          */
1278 	apic_soft_disable();
1279 
1280 	return 0;
1281 }
1282 
1283 void play_dead_common(void)
1284 {
1285 	idle_task_exit();
1286 
1287 	cpuhp_ap_report_dead();
1288 
1289 	local_irq_disable();
1290 }
1291 
1292 /*
1293  * We need to flush the caches before going to sleep, lest we have
1294  * dirty data in our caches when we come back up.
1295  */
1296 static inline void mwait_play_dead(void)
1297 {
1298 	struct mwait_cpu_dead *md = this_cpu_ptr(&mwait_cpu_dead);
1299 	unsigned int eax, ebx, ecx, edx;
1300 	unsigned int highest_cstate = 0;
1301 	unsigned int highest_subcstate = 0;
1302 	int i;
1303 
1304 	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD ||
1305 	    boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
1306 		return;
1307 	if (!this_cpu_has(X86_FEATURE_MWAIT))
1308 		return;
1309 	if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1310 		return;
1311 	if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1312 		return;
1313 
1314 	eax = CPUID_MWAIT_LEAF;
1315 	ecx = 0;
1316 	native_cpuid(&eax, &ebx, &ecx, &edx);
1317 
1318 	/*
1319 	 * eax will be 0 if EDX enumeration is not valid.
1320 	 * Initialized below to cstate, sub_cstate value when EDX is valid.
1321 	 */
1322 	if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1323 		eax = 0;
1324 	} else {
1325 		edx >>= MWAIT_SUBSTATE_SIZE;
1326 		for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1327 			if (edx & MWAIT_SUBSTATE_MASK) {
1328 				highest_cstate = i;
1329 				highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1330 			}
1331 		}
1332 		eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1333 			(highest_subcstate - 1);
1334 	}
1335 
1336 	/* Set up state for the kexec() hack below */
1337 	md->status = CPUDEAD_MWAIT_WAIT;
1338 	md->control = CPUDEAD_MWAIT_WAIT;
1339 
1340 	wbinvd();
1341 
1342 	while (1) {
1343 		/*
1344 		 * The CLFLUSH is a workaround for erratum AAI65 for
1345 		 * the Xeon 7400 series.  It's not clear it is actually
1346 		 * needed, but it should be harmless in either case.
1347 		 * The WBINVD is insufficient due to the spurious-wakeup
1348 		 * case where we return around the loop.
1349 		 */
1350 		mb();
1351 		clflush(md);
1352 		mb();
1353 		__monitor(md, 0, 0);
1354 		mb();
1355 		__mwait(eax, 0);
1356 
1357 		if (READ_ONCE(md->control) == CPUDEAD_MWAIT_KEXEC_HLT) {
1358 			/*
1359 			 * Kexec is about to happen. Don't go back into mwait() as
1360 			 * the kexec kernel might overwrite text and data including
1361 			 * page tables and stack. So mwait() would resume when the
1362 			 * monitor cache line is written to and then the CPU goes
1363 			 * south due to overwritten text, page tables and stack.
1364 			 *
1365 			 * Note: This does _NOT_ protect against a stray MCE, NMI,
1366 			 * SMI. They will resume execution at the instruction
1367 			 * following the HLT instruction and run into the problem
1368 			 * which this is trying to prevent.
1369 			 */
1370 			WRITE_ONCE(md->status, CPUDEAD_MWAIT_KEXEC_HLT);
1371 			while(1)
1372 				native_halt();
1373 		}
1374 	}
1375 }
1376 
1377 /*
1378  * Kick all "offline" CPUs out of mwait on kexec(). See comment in
1379  * mwait_play_dead().
1380  */
1381 void smp_kick_mwait_play_dead(void)
1382 {
1383 	u32 newstate = CPUDEAD_MWAIT_KEXEC_HLT;
1384 	struct mwait_cpu_dead *md;
1385 	unsigned int cpu, i;
1386 
1387 	for_each_cpu_andnot(cpu, cpu_present_mask, cpu_online_mask) {
1388 		md = per_cpu_ptr(&mwait_cpu_dead, cpu);
1389 
1390 		/* Does it sit in mwait_play_dead() ? */
1391 		if (READ_ONCE(md->status) != CPUDEAD_MWAIT_WAIT)
1392 			continue;
1393 
1394 		/* Wait up to 5ms */
1395 		for (i = 0; READ_ONCE(md->status) != newstate && i < 1000; i++) {
1396 			/* Bring it out of mwait */
1397 			WRITE_ONCE(md->control, newstate);
1398 			udelay(5);
1399 		}
1400 
1401 		if (READ_ONCE(md->status) != newstate)
1402 			pr_err_once("CPU%u is stuck in mwait_play_dead()\n", cpu);
1403 	}
1404 }
1405 
1406 void __noreturn hlt_play_dead(void)
1407 {
1408 	if (__this_cpu_read(cpu_info.x86) >= 4)
1409 		wbinvd();
1410 
1411 	while (1)
1412 		native_halt();
1413 }
1414 
1415 /*
1416  * native_play_dead() is essentially a __noreturn function, but it can't
1417  * be marked as such as the compiler may complain about it.
1418  */
1419 void native_play_dead(void)
1420 {
1421 	if (cpu_feature_enabled(X86_FEATURE_KERNEL_IBRS))
1422 		__update_spec_ctrl(0);
1423 
1424 	play_dead_common();
1425 	tboot_shutdown(TB_SHUTDOWN_WFS);
1426 
1427 	mwait_play_dead();
1428 	if (cpuidle_play_dead())
1429 		hlt_play_dead();
1430 }
1431 
1432 #else /* ... !CONFIG_HOTPLUG_CPU */
1433 int native_cpu_disable(void)
1434 {
1435 	return -ENOSYS;
1436 }
1437 
1438 void native_play_dead(void)
1439 {
1440 	BUG();
1441 }
1442 
1443 #endif
1444