xref: /linux/arch/x86/kernel/smpboot.c (revision 2dbbdeda77a61b39dc4a34dfce873907cfea2c4b)
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 #include <linux/acpi.h>
64 
65 #include <asm/acpi.h>
66 #include <asm/cacheinfo.h>
67 #include <asm/cpuid/api.h>
68 #include <asm/desc.h>
69 #include <asm/nmi.h>
70 #include <asm/irq.h>
71 #include <asm/realmode.h>
72 #include <asm/cpu.h>
73 #include <asm/numa.h>
74 #include <asm/tlbflush.h>
75 #include <asm/mtrr.h>
76 #include <asm/mwait.h>
77 #include <asm/apic.h>
78 #include <asm/io_apic.h>
79 #include <asm/fpu/api.h>
80 #include <asm/setup.h>
81 #include <asm/uv/uv.h>
82 #include <asm/microcode.h>
83 #include <asm/i8259.h>
84 #include <asm/misc.h>
85 #include <asm/qspinlock.h>
86 #include <asm/intel-family.h>
87 #include <asm/cpu_device_id.h>
88 #include <asm/spec-ctrl.h>
89 #include <asm/hw_irq.h>
90 #include <asm/stackprotector.h>
91 #include <asm/sev.h>
92 #include <asm/spec-ctrl.h>
93 
94 /* representing HT siblings of each logical CPU */
95 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
96 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
97 
98 /* representing HT and core siblings of each logical CPU */
99 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
100 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
101 
102 /* representing HT, core, and die siblings of each logical CPU */
103 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_die_map);
104 EXPORT_PER_CPU_SYMBOL(cpu_die_map);
105 
106 /* CPUs which are the primary SMT threads */
107 struct cpumask __cpu_primary_thread_mask __read_mostly;
108 
109 /* Representing CPUs for which sibling maps can be computed */
110 static cpumask_var_t cpu_sibling_setup_mask;
111 
112 struct mwait_cpu_dead {
113 	unsigned int	control;
114 	unsigned int	status;
115 };
116 
117 #define CPUDEAD_MWAIT_WAIT	0xDEADBEEF
118 #define CPUDEAD_MWAIT_KEXEC_HLT	0x4A17DEAD
119 
120 /*
121  * Cache line aligned data for mwait_play_dead(). Separate on purpose so
122  * that it's unlikely to be touched by other CPUs.
123  */
124 static DEFINE_PER_CPU_ALIGNED(struct mwait_cpu_dead, mwait_cpu_dead);
125 
126 /* Maximum number of SMT threads on any online core */
127 int __read_mostly __max_smt_threads = 1;
128 
129 /* Flag to indicate if a complete sched domain rebuild is required */
130 bool x86_topology_update;
131 
132 int arch_update_cpu_topology(void)
133 {
134 	int retval = x86_topology_update;
135 
136 	x86_topology_update = false;
137 	return retval;
138 }
139 
140 static unsigned int smpboot_warm_reset_vector_count;
141 
142 static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
143 {
144 	unsigned long flags;
145 
146 	spin_lock_irqsave(&rtc_lock, flags);
147 	if (!smpboot_warm_reset_vector_count++) {
148 		CMOS_WRITE(0xa, 0xf);
149 		*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) = start_eip >> 4;
150 		*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = start_eip & 0xf;
151 	}
152 	spin_unlock_irqrestore(&rtc_lock, flags);
153 }
154 
155 static inline void smpboot_restore_warm_reset_vector(void)
156 {
157 	unsigned long flags;
158 
159 	/*
160 	 * Paranoid:  Set warm reset code and vector here back
161 	 * to default values.
162 	 */
163 	spin_lock_irqsave(&rtc_lock, flags);
164 	if (!--smpboot_warm_reset_vector_count) {
165 		CMOS_WRITE(0, 0xf);
166 		*((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
167 	}
168 	spin_unlock_irqrestore(&rtc_lock, flags);
169 
170 }
171 
172 /* Run the next set of setup steps for the upcoming CPU */
173 static void ap_starting(void)
174 {
175 	int cpuid = smp_processor_id();
176 
177 	/* Mop up eventual mwait_play_dead() wreckage */
178 	this_cpu_write(mwait_cpu_dead.status, 0);
179 	this_cpu_write(mwait_cpu_dead.control, 0);
180 
181 	/*
182 	 * If woken up by an INIT in an 82489DX configuration the alive
183 	 * synchronization guarantees that the CPU does not reach this
184 	 * point before an INIT_deassert IPI reaches the local APIC, so it
185 	 * is now safe to touch the local APIC.
186 	 *
187 	 * Set up this CPU, first the APIC, which is probably redundant on
188 	 * most boards.
189 	 */
190 	apic_ap_setup();
191 
192 	/* Save the processor parameters. */
193 	identify_secondary_cpu(cpuid);
194 
195 	/*
196 	 * The topology information must be up to date before
197 	 * notify_cpu_starting().
198 	 */
199 	set_cpu_sibling_map(cpuid);
200 
201 	ap_init_aperfmperf();
202 
203 	pr_debug("Stack at about %p\n", &cpuid);
204 
205 	wmb();
206 
207 	/*
208 	 * This runs the AP through all the cpuhp states to its target
209 	 * state CPUHP_ONLINE.
210 	 */
211 	notify_cpu_starting(cpuid);
212 }
213 
214 static void ap_calibrate_delay(void)
215 {
216 	/*
217 	 * Calibrate the delay loop and update loops_per_jiffy in cpu_data.
218 	 * identify_secondary_cpu() stored a value that is close but not as
219 	 * accurate as the value just calculated.
220 	 *
221 	 * As this is invoked after the TSC synchronization check,
222 	 * calibrate_delay_is_known() will skip the calibration routine
223 	 * when TSC is synchronized across sockets.
224 	 */
225 	calibrate_delay();
226 	cpu_data(smp_processor_id()).loops_per_jiffy = loops_per_jiffy;
227 }
228 
229 /*
230  * Activate a secondary processor.
231  */
232 static void notrace __noendbr start_secondary(void *unused)
233 {
234 	/*
235 	 * Don't put *anything* except direct CPU state initialization
236 	 * before cpu_init(), SMP booting is too fragile that we want to
237 	 * limit the things done here to the most necessary things.
238 	 */
239 	cr4_init();
240 
241 	/*
242 	 * 32-bit specific. 64-bit reaches this code with the correct page
243 	 * table established. Yet another historical divergence.
244 	 */
245 	if (IS_ENABLED(CONFIG_X86_32)) {
246 		/* switch away from the initial page table */
247 		load_cr3(swapper_pg_dir);
248 		__flush_tlb_all();
249 	}
250 
251 	cpu_init_exception_handling(false);
252 
253 	/*
254 	 * Load the microcode before reaching the AP alive synchronization
255 	 * point below so it is not part of the full per CPU serialized
256 	 * bringup part when "parallel" bringup is enabled.
257 	 *
258 	 * That's even safe when hyperthreading is enabled in the CPU as
259 	 * the core code starts the primary threads first and leaves the
260 	 * secondary threads waiting for SIPI. Loading microcode on
261 	 * physical cores concurrently is a safe operation.
262 	 *
263 	 * This covers both the Intel specific issue that concurrent
264 	 * microcode loading on SMT siblings must be prohibited and the
265 	 * vendor independent issue`that microcode loading which changes
266 	 * CPUID, MSRs etc. must be strictly serialized to maintain
267 	 * software state correctness.
268 	 */
269 	load_ucode_ap();
270 
271 	/*
272 	 * Synchronization point with the hotplug core. Sets this CPUs
273 	 * synchronization state to ALIVE and spin-waits for the control CPU to
274 	 * release this CPU for further bringup.
275 	 */
276 	cpuhp_ap_sync_alive();
277 
278 	cpu_init();
279 	fpu__init_cpu();
280 	rcutree_report_cpu_starting(raw_smp_processor_id());
281 	x86_cpuinit.early_percpu_clock_init();
282 
283 	ap_starting();
284 
285 	/* Check TSC synchronization with the control CPU. */
286 	check_tsc_sync_target();
287 
288 	/*
289 	 * Calibrate the delay loop after the TSC synchronization check.
290 	 * This allows to skip the calibration when TSC is synchronized
291 	 * across sockets.
292 	 */
293 	ap_calibrate_delay();
294 
295 	speculative_store_bypass_ht_init();
296 
297 	/*
298 	 * Lock vector_lock, set CPU online and bring the vector
299 	 * allocator online. Online must be set with vector_lock held
300 	 * to prevent a concurrent irq setup/teardown from seeing a
301 	 * half valid vector space.
302 	 */
303 	lock_vector_lock();
304 	set_cpu_online(smp_processor_id(), true);
305 	lapic_online();
306 	unlock_vector_lock();
307 	x86_platform.nmi_init();
308 
309 	/* enable local interrupts */
310 	local_irq_enable();
311 
312 	x86_cpuinit.setup_percpu_clockev();
313 
314 	wmb();
315 	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
316 }
317 ANNOTATE_NOENDBR_SYM(start_secondary);
318 
319 static bool
320 topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
321 {
322 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
323 
324 	return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
325 }
326 
327 static bool
328 topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
329 {
330 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
331 
332 	return !WARN_ONCE(!topology_same_node(c, o),
333 		"sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
334 		"[node: %d != %d]. Ignoring dependency.\n",
335 		cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
336 }
337 
338 #define link_mask(mfunc, c1, c2)					\
339 do {									\
340 	cpumask_set_cpu((c1), mfunc(c2));				\
341 	cpumask_set_cpu((c2), mfunc(c1));				\
342 } while (0)
343 
344 static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
345 {
346 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
347 		int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
348 
349 		if (c->topo.pkg_id == o->topo.pkg_id &&
350 		    c->topo.die_id == o->topo.die_id &&
351 		    c->topo.amd_node_id == o->topo.amd_node_id &&
352 		    per_cpu_llc_id(cpu1) == per_cpu_llc_id(cpu2)) {
353 			if (c->topo.core_id == o->topo.core_id)
354 				return topology_sane(c, o, "smt");
355 
356 			if ((c->topo.cu_id != 0xff) &&
357 			    (o->topo.cu_id != 0xff) &&
358 			    (c->topo.cu_id == o->topo.cu_id))
359 				return topology_sane(c, o, "smt");
360 		}
361 
362 	} else if (c->topo.pkg_id == o->topo.pkg_id &&
363 		   c->topo.die_id == o->topo.die_id &&
364 		   c->topo.core_id == o->topo.core_id) {
365 		return topology_sane(c, o, "smt");
366 	}
367 
368 	return false;
369 }
370 
371 static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
372 {
373 	if (c->topo.pkg_id != o->topo.pkg_id || c->topo.die_id != o->topo.die_id)
374 		return false;
375 
376 	if (cpu_feature_enabled(X86_FEATURE_TOPOEXT) && topology_amd_nodes_per_pkg() > 1)
377 		return c->topo.amd_node_id == o->topo.amd_node_id;
378 
379 	return true;
380 }
381 
382 static bool match_l2c(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
383 {
384 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
385 
386 	/* If the arch didn't set up l2c_id, fall back to SMT */
387 	if (per_cpu_l2c_id(cpu1) == BAD_APICID)
388 		return match_smt(c, o);
389 
390 	/* Do not match if L2 cache id does not match: */
391 	if (per_cpu_l2c_id(cpu1) != per_cpu_l2c_id(cpu2))
392 		return false;
393 
394 	return topology_sane(c, o, "l2c");
395 }
396 
397 /*
398  * Unlike the other levels, we do not enforce keeping a
399  * multicore group inside a NUMA node.  If this happens, we will
400  * discard the MC level of the topology later.
401  */
402 static bool match_pkg(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
403 {
404 	if (c->topo.pkg_id == o->topo.pkg_id)
405 		return true;
406 	return false;
407 }
408 
409 /*
410  * Define intel_cod_cpu[] for Intel COD (Cluster-on-Die) CPUs.
411  *
412  * Any Intel CPU that has multiple nodes per package and does not
413  * match intel_cod_cpu[] has the SNC (Sub-NUMA Cluster) topology.
414  *
415  * When in SNC mode, these CPUs enumerate an LLC that is shared
416  * by multiple NUMA nodes. The LLC is shared for off-package data
417  * access but private to the NUMA node (half of the package) for
418  * on-package access. CPUID (the source of the information about
419  * the LLC) can only enumerate the cache as shared or unshared,
420  * but not this particular configuration.
421  */
422 
423 static const struct x86_cpu_id intel_cod_cpu[] = {
424 	X86_MATCH_VFM(INTEL_HASWELL_X,	 0),	/* COD */
425 	X86_MATCH_VFM(INTEL_BROADWELL_X, 0),	/* COD */
426 	X86_MATCH_VFM(INTEL_ANY,	 1),	/* SNC */
427 	{}
428 };
429 
430 static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
431 {
432 	const struct x86_cpu_id *id = x86_match_cpu(intel_cod_cpu);
433 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
434 	bool intel_snc = id && id->driver_data;
435 
436 	/* Do not match if we do not have a valid APICID for cpu: */
437 	if (per_cpu_llc_id(cpu1) == BAD_APICID)
438 		return false;
439 
440 	/* Do not match if LLC id does not match: */
441 	if (per_cpu_llc_id(cpu1) != per_cpu_llc_id(cpu2))
442 		return false;
443 
444 	/*
445 	 * Allow the SNC topology without warning. Return of false
446 	 * means 'c' does not share the LLC of 'o'. This will be
447 	 * reflected to userspace.
448 	 */
449 	if (match_pkg(c, o) && !topology_same_node(c, o) && intel_snc)
450 		return false;
451 
452 	return topology_sane(c, o, "llc");
453 }
454 
455 
456 static inline int x86_sched_itmt_flags(void)
457 {
458 	return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
459 }
460 
461 #ifdef CONFIG_SCHED_MC
462 static int x86_core_flags(void)
463 {
464 	return cpu_core_flags() | x86_sched_itmt_flags();
465 }
466 #endif
467 #ifdef CONFIG_SCHED_CLUSTER
468 static int x86_cluster_flags(void)
469 {
470 	return cpu_cluster_flags() | x86_sched_itmt_flags();
471 }
472 #endif
473 
474 /*
475  * Set if a package/die has multiple NUMA nodes inside.
476  * AMD Magny-Cours, Intel Cluster-on-Die, and Intel
477  * Sub-NUMA Clustering have this.
478  */
479 static bool x86_has_numa_in_package;
480 
481 static struct sched_domain_topology_level x86_topology[] = {
482 	SDTL_INIT(tl_smt_mask, cpu_smt_flags, SMT),
483 #ifdef CONFIG_SCHED_CLUSTER
484 	SDTL_INIT(tl_cls_mask, x86_cluster_flags, CLS),
485 #endif
486 #ifdef CONFIG_SCHED_MC
487 	SDTL_INIT(tl_mc_mask, x86_core_flags, MC),
488 #endif
489 	SDTL_INIT(tl_pkg_mask, x86_sched_itmt_flags, PKG),
490 	{ NULL },
491 };
492 
493 static void __init build_sched_topology(void)
494 {
495 	struct sched_domain_topology_level *topology = x86_topology;
496 
497 	/*
498 	 * When there is NUMA topology inside the package invalidate the
499 	 * PKG domain since the NUMA domains will auto-magically create the
500 	 * right spanning domains based on the SLIT.
501 	 */
502 	if (x86_has_numa_in_package) {
503 		unsigned int pkgdom = ARRAY_SIZE(x86_topology) - 2;
504 
505 		memset(&x86_topology[pkgdom], 0, sizeof(x86_topology[pkgdom]));
506 	}
507 
508 	/*
509 	 * Drop the SMT domains if there is only one thread per-core
510 	 * since it'll get degenerated by the scheduler anyways.
511 	 */
512 	if (cpu_smt_num_threads <= 1)
513 		++topology;
514 
515 	set_sched_topology(topology);
516 }
517 
518 #ifdef CONFIG_NUMA
519 static int sched_avg_remote_distance;
520 static int avg_remote_numa_distance(void)
521 {
522 	int i, j;
523 	int distance, nr_remote, total_distance;
524 
525 	if (sched_avg_remote_distance > 0)
526 		return sched_avg_remote_distance;
527 
528 	nr_remote = 0;
529 	total_distance = 0;
530 	for_each_node_state(i, N_CPU) {
531 		for_each_node_state(j, N_CPU) {
532 			distance = node_distance(i, j);
533 
534 			if (distance >= REMOTE_DISTANCE) {
535 				nr_remote++;
536 				total_distance += distance;
537 			}
538 		}
539 	}
540 	if (nr_remote)
541 		sched_avg_remote_distance = total_distance / nr_remote;
542 	else
543 		sched_avg_remote_distance = REMOTE_DISTANCE;
544 
545 	return sched_avg_remote_distance;
546 }
547 
548 int arch_sched_node_distance(int from, int to)
549 {
550 	int d = node_distance(from, to);
551 
552 	switch (boot_cpu_data.x86_vfm) {
553 	case INTEL_GRANITERAPIDS_X:
554 	case INTEL_ATOM_DARKMONT_X:
555 
556 		if (!x86_has_numa_in_package || topology_max_packages() == 1 ||
557 		    d < REMOTE_DISTANCE)
558 			return d;
559 
560 		/*
561 		 * With SNC enabled, there could be too many levels of remote
562 		 * NUMA node distances, creating NUMA domain levels
563 		 * including local nodes and partial remote nodes.
564 		 *
565 		 * Trim finer distance tuning for NUMA nodes in remote package
566 		 * for the purpose of building sched domains. Group NUMA nodes
567 		 * in the remote package in the same sched group.
568 		 * Simplify NUMA domains and avoid extra NUMA levels including
569 		 * different remote NUMA nodes and local nodes.
570 		 *
571 		 * GNR and CWF don't expect systems with more than 2 packages
572 		 * and more than 2 hops between packages. Single average remote
573 		 * distance won't be appropriate if there are more than 2
574 		 * packages as average distance to different remote packages
575 		 * could be different.
576 		 */
577 		WARN_ONCE(topology_max_packages() > 2,
578 			  "sched: Expect only up to 2 packages for GNR or CWF, "
579 			  "but saw %d packages when building sched domains.",
580 			  topology_max_packages());
581 
582 		d = avg_remote_numa_distance();
583 	}
584 	return d;
585 }
586 #endif /* CONFIG_NUMA */
587 
588 void set_cpu_sibling_map(int cpu)
589 {
590 	bool has_smt = __max_threads_per_core > 1;
591 	bool has_mp = has_smt || topology_num_cores_per_package() > 1;
592 	struct cpuinfo_x86 *c = &cpu_data(cpu);
593 	struct cpuinfo_x86 *o;
594 	int i, threads;
595 
596 	cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
597 
598 	if (!has_mp) {
599 		cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
600 		cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
601 		cpumask_set_cpu(cpu, cpu_l2c_shared_mask(cpu));
602 		cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
603 		cpumask_set_cpu(cpu, topology_die_cpumask(cpu));
604 		c->booted_cores = 1;
605 		return;
606 	}
607 
608 	for_each_cpu(i, cpu_sibling_setup_mask) {
609 		o = &cpu_data(i);
610 
611 		if (match_pkg(c, o) && !topology_same_node(c, o))
612 			x86_has_numa_in_package = true;
613 
614 		if ((i == cpu) || (has_smt && match_smt(c, o)))
615 			link_mask(topology_sibling_cpumask, cpu, i);
616 
617 		if ((i == cpu) || (has_mp && match_llc(c, o)))
618 			link_mask(cpu_llc_shared_mask, cpu, i);
619 
620 		if ((i == cpu) || (has_mp && match_l2c(c, o)))
621 			link_mask(cpu_l2c_shared_mask, cpu, i);
622 
623 		if ((i == cpu) || (has_mp && match_die(c, o)))
624 			link_mask(topology_die_cpumask, cpu, i);
625 	}
626 
627 	threads = cpumask_weight(topology_sibling_cpumask(cpu));
628 	if (threads > __max_smt_threads)
629 		__max_smt_threads = threads;
630 
631 	for_each_cpu(i, topology_sibling_cpumask(cpu))
632 		cpu_data(i).smt_active = threads > 1;
633 
634 	/*
635 	 * This needs a separate iteration over the cpus because we rely on all
636 	 * topology_sibling_cpumask links to be set-up.
637 	 */
638 	for_each_cpu(i, cpu_sibling_setup_mask) {
639 		o = &cpu_data(i);
640 
641 		if ((i == cpu) || (has_mp && match_pkg(c, o))) {
642 			link_mask(topology_core_cpumask, cpu, i);
643 
644 			/*
645 			 *  Does this new cpu bringup a new core?
646 			 */
647 			if (threads == 1) {
648 				/*
649 				 * for each core in package, increment
650 				 * the booted_cores for this new cpu
651 				 */
652 				if (cpumask_first(
653 				    topology_sibling_cpumask(i)) == i)
654 					c->booted_cores++;
655 				/*
656 				 * increment the core count for all
657 				 * the other cpus in this package
658 				 */
659 				if (i != cpu)
660 					cpu_data(i).booted_cores++;
661 			} else if (i != cpu && !c->booted_cores)
662 				c->booted_cores = cpu_data(i).booted_cores;
663 		}
664 	}
665 }
666 
667 /* maps the cpu to the sched domain representing multi-core */
668 const struct cpumask *cpu_coregroup_mask(int cpu)
669 {
670 	return cpu_llc_shared_mask(cpu);
671 }
672 
673 const struct cpumask *cpu_clustergroup_mask(int cpu)
674 {
675 	return cpu_l2c_shared_mask(cpu);
676 }
677 EXPORT_SYMBOL_GPL(cpu_clustergroup_mask);
678 
679 static void impress_friends(void)
680 {
681 	int cpu;
682 	unsigned long bogosum = 0;
683 	/*
684 	 * Allow the user to impress friends.
685 	 */
686 	pr_debug("Before bogomips\n");
687 	for_each_online_cpu(cpu)
688 		bogosum += cpu_data(cpu).loops_per_jiffy;
689 
690 	pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
691 		num_online_cpus(),
692 		bogosum/(500000/HZ),
693 		(bogosum/(5000/HZ))%100);
694 
695 	pr_debug("Before bogocount - setting activated=1\n");
696 }
697 
698 /*
699  * The Multiprocessor Specification 1.4 (1997) example code suggests
700  * that there should be a 10ms delay between the BSP asserting INIT
701  * and de-asserting INIT, when starting a remote processor.
702  * But that slows boot and resume on modern processors, which include
703  * many cores and don't require that delay.
704  *
705  * Cmdline "cpu_init_udelay=" is available to override this delay.
706  */
707 #define UDELAY_10MS_LEGACY 10000
708 
709 static unsigned int init_udelay = UINT_MAX;
710 
711 static int __init cpu_init_udelay(char *str)
712 {
713 	get_option(&str, &init_udelay);
714 
715 	return 0;
716 }
717 early_param("cpu_init_udelay", cpu_init_udelay);
718 
719 static void __init smp_set_init_udelay(void)
720 {
721 	/* if cmdline changed it from default, leave it alone */
722 	if (init_udelay != UINT_MAX)
723 		return;
724 
725 	/* if modern processor, use no delay */
726 	if ((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL && boot_cpu_data.x86_vfm >= INTEL_PENTIUM_PRO) ||
727 	    (boot_cpu_data.x86_vendor == X86_VENDOR_HYGON && boot_cpu_data.x86 >= 0x18) ||
728 	    (boot_cpu_data.x86_vendor == X86_VENDOR_AMD   && boot_cpu_data.x86 >= 0xF)) {
729 		init_udelay = 0;
730 		return;
731 	}
732 	/* else, use legacy delay */
733 	init_udelay = UDELAY_10MS_LEGACY;
734 }
735 
736 /*
737  * Wake up AP by INIT, INIT, STARTUP sequence.
738  */
739 static void send_init_sequence(u32 phys_apicid)
740 {
741 	int maxlvt = lapic_get_maxlvt();
742 
743 	/* Be paranoid about clearing APIC errors. */
744 	if (APIC_INTEGRATED(boot_cpu_apic_version)) {
745 		/* Due to the Pentium erratum 3AP.  */
746 		if (maxlvt > 3)
747 			apic_write(APIC_ESR, 0);
748 		apic_read(APIC_ESR);
749 	}
750 
751 	/* Assert INIT on the target CPU */
752 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT, phys_apicid);
753 	safe_apic_wait_icr_idle();
754 
755 	udelay(init_udelay);
756 
757 	/* Deassert INIT on the target CPU */
758 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
759 	safe_apic_wait_icr_idle();
760 }
761 
762 /*
763  * Wake up AP by INIT, INIT, STARTUP sequence.
764  */
765 static int wakeup_secondary_cpu_via_init(u32 phys_apicid, unsigned long start_eip, unsigned int cpu)
766 {
767 	unsigned long send_status = 0, accept_status = 0;
768 	int num_starts, j, maxlvt;
769 
770 	preempt_disable();
771 	maxlvt = lapic_get_maxlvt();
772 	send_init_sequence(phys_apicid);
773 
774 	mb();
775 
776 	/*
777 	 * Should we send STARTUP IPIs ?
778 	 *
779 	 * Determine this based on the APIC version.
780 	 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
781 	 */
782 	if (APIC_INTEGRATED(boot_cpu_apic_version))
783 		num_starts = 2;
784 	else
785 		num_starts = 0;
786 
787 	/*
788 	 * Run STARTUP IPI loop.
789 	 */
790 	pr_debug("#startup loops: %d\n", num_starts);
791 
792 	for (j = 1; j <= num_starts; j++) {
793 		pr_debug("Sending STARTUP #%d\n", j);
794 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
795 			apic_write(APIC_ESR, 0);
796 		apic_read(APIC_ESR);
797 		pr_debug("After apic_write\n");
798 
799 		/*
800 		 * STARTUP IPI
801 		 */
802 
803 		/* Target chip */
804 		/* Boot on the stack */
805 		/* Kick the second */
806 		apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
807 			       phys_apicid);
808 
809 		/*
810 		 * Give the other CPU some time to accept the IPI.
811 		 */
812 		if (init_udelay == 0)
813 			udelay(10);
814 		else
815 			udelay(300);
816 
817 		pr_debug("Startup point 1\n");
818 
819 		pr_debug("Waiting for send to finish...\n");
820 		send_status = safe_apic_wait_icr_idle();
821 
822 		/*
823 		 * Give the other CPU some time to accept the IPI.
824 		 */
825 		if (init_udelay == 0)
826 			udelay(10);
827 		else
828 			udelay(200);
829 
830 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
831 			apic_write(APIC_ESR, 0);
832 		accept_status = (apic_read(APIC_ESR) & 0xEF);
833 		if (send_status || accept_status)
834 			break;
835 	}
836 	pr_debug("After Startup\n");
837 
838 	if (send_status)
839 		pr_err("APIC never delivered???\n");
840 	if (accept_status)
841 		pr_err("APIC delivery error (%lx)\n", accept_status);
842 
843 	preempt_enable();
844 	return (send_status | accept_status);
845 }
846 
847 /* reduce the number of lines printed when booting a large cpu count system */
848 static void announce_cpu(int cpu, int apicid)
849 {
850 	static int width, node_width, first = 1;
851 	static int current_node = NUMA_NO_NODE;
852 	int node = early_cpu_to_node(cpu);
853 
854 	if (!width)
855 		width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
856 
857 	if (!node_width)
858 		node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
859 
860 	if (system_state < SYSTEM_RUNNING) {
861 		if (first)
862 			pr_info("x86: Booting SMP configuration:\n");
863 
864 		if (node != current_node) {
865 			if (current_node > (-1))
866 				pr_cont("\n");
867 			current_node = node;
868 
869 			printk(KERN_INFO ".... node %*s#%d, CPUs:  ",
870 			       node_width - num_digits(node), " ", node);
871 		}
872 
873 		/* Add padding for the BSP */
874 		if (first)
875 			pr_cont("%*s", width + 1, " ");
876 		first = 0;
877 
878 		pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
879 	} else
880 		pr_info("Booting Node %d Processor %d APIC 0x%x\n",
881 			node, cpu, apicid);
882 }
883 
884 int common_cpu_up(unsigned int cpu, struct task_struct *idle)
885 {
886 	int ret;
887 
888 	/* Just in case we booted with a single CPU. */
889 	alternatives_enable_smp();
890 
891 	per_cpu(current_task, cpu) = idle;
892 	cpu_init_stack_canary(cpu, idle);
893 
894 	/* Initialize the interrupt stack(s) */
895 	ret = irq_init_percpu_irqstack(cpu);
896 	if (ret)
897 		return ret;
898 
899 #ifdef CONFIG_X86_32
900 	/* Stack for startup_32 can be just as for start_secondary onwards */
901 	per_cpu(cpu_current_top_of_stack, cpu) = task_top_of_stack(idle);
902 #endif
903 	return 0;
904 }
905 
906 /*
907  * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
908  * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
909  * Returns zero if startup was successfully sent, else error code from
910  * ->wakeup_secondary_cpu.
911  */
912 static int do_boot_cpu(u32 apicid, unsigned int cpu, struct task_struct *idle)
913 {
914 	unsigned long start_ip = real_mode_header->trampoline_start;
915 	int ret;
916 
917 #ifdef CONFIG_X86_64
918 	/* If 64-bit wakeup method exists, use the 64-bit mode trampoline IP */
919 	if (apic->wakeup_secondary_cpu_64)
920 		start_ip = real_mode_header->trampoline_start64;
921 #endif
922 	idle->thread.sp = (unsigned long)task_pt_regs(idle);
923 	initial_code = (unsigned long)start_secondary;
924 
925 	if (IS_ENABLED(CONFIG_X86_32)) {
926 		early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
927 		initial_stack  = idle->thread.sp;
928 	} else if (!(smpboot_control & STARTUP_PARALLEL_MASK)) {
929 		smpboot_control = cpu;
930 	}
931 
932 	/* Enable the espfix hack for this CPU */
933 	init_espfix_ap(cpu);
934 
935 	/* So we see what's up */
936 	announce_cpu(cpu, apicid);
937 
938 	/*
939 	 * This grunge runs the startup process for
940 	 * the targeted processor.
941 	 */
942 	if (x86_platform.legacy.warm_reset) {
943 
944 		pr_debug("Setting warm reset code and vector.\n");
945 
946 		smpboot_setup_warm_reset_vector(start_ip);
947 		/*
948 		 * Be paranoid about clearing APIC errors.
949 		*/
950 		if (APIC_INTEGRATED(boot_cpu_apic_version)) {
951 			apic_write(APIC_ESR, 0);
952 			apic_read(APIC_ESR);
953 		}
954 	}
955 
956 	smp_mb();
957 
958 	/*
959 	 * Wake up a CPU in difference cases:
960 	 * - Use a method from the APIC driver if one defined, with wakeup
961 	 *   straight to 64-bit mode preferred over wakeup to RM.
962 	 * Otherwise,
963 	 * - Use an INIT boot APIC message
964 	 */
965 	if (apic->wakeup_secondary_cpu_64)
966 		ret = apic->wakeup_secondary_cpu_64(apicid, start_ip, cpu);
967 	else if (apic->wakeup_secondary_cpu)
968 		ret = apic->wakeup_secondary_cpu(apicid, start_ip, cpu);
969 	else
970 		ret = wakeup_secondary_cpu_via_init(apicid, start_ip, cpu);
971 
972 	/* If the wakeup mechanism failed, cleanup the warm reset vector */
973 	if (ret)
974 		arch_cpuhp_cleanup_kick_cpu(cpu);
975 	return ret;
976 }
977 
978 int native_kick_ap(unsigned int cpu, struct task_struct *tidle)
979 {
980 	u32 apicid = apic->cpu_present_to_apicid(cpu);
981 	int err;
982 
983 	lockdep_assert_irqs_enabled();
984 
985 	pr_debug("++++++++++++++++++++=_---CPU UP  %u\n", cpu);
986 
987 	if (apicid == BAD_APICID || !apic_id_valid(apicid)) {
988 		pr_err("CPU %u has invalid APIC ID %x. Aborting bringup\n", cpu, apicid);
989 		return -EINVAL;
990 	}
991 
992 	if (!test_bit(apicid, phys_cpu_present_map)) {
993 		pr_err("CPU %u APIC ID %x is not present. Aborting bringup\n", cpu, apicid);
994 		return -EINVAL;
995 	}
996 
997 	/*
998 	 * Save current MTRR state in case it was changed since early boot
999 	 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1000 	 */
1001 	mtrr_save_state();
1002 
1003 	/* the FPU context is blank, nobody can own it */
1004 	per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
1005 
1006 	err = common_cpu_up(cpu, tidle);
1007 	if (err)
1008 		return err;
1009 
1010 	err = do_boot_cpu(apicid, cpu, tidle);
1011 	if (err)
1012 		pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
1013 
1014 	return err;
1015 }
1016 
1017 int arch_cpuhp_kick_ap_alive(unsigned int cpu, struct task_struct *tidle)
1018 {
1019 	return smp_ops.kick_ap_alive(cpu, tidle);
1020 }
1021 
1022 void arch_cpuhp_cleanup_kick_cpu(unsigned int cpu)
1023 {
1024 	/* Cleanup possible dangling ends... */
1025 	if (smp_ops.kick_ap_alive == native_kick_ap && x86_platform.legacy.warm_reset)
1026 		smpboot_restore_warm_reset_vector();
1027 }
1028 
1029 void arch_cpuhp_cleanup_dead_cpu(unsigned int cpu)
1030 {
1031 	if (smp_ops.cleanup_dead_cpu)
1032 		smp_ops.cleanup_dead_cpu(cpu);
1033 
1034 	if (system_state == SYSTEM_RUNNING)
1035 		pr_info("CPU %u is now offline\n", cpu);
1036 }
1037 
1038 void arch_cpuhp_sync_state_poll(void)
1039 {
1040 	if (smp_ops.poll_sync_state)
1041 		smp_ops.poll_sync_state();
1042 }
1043 
1044 /**
1045  * arch_disable_smp_support() - Disables SMP support for x86 at boottime
1046  */
1047 void __init arch_disable_smp_support(void)
1048 {
1049 	disable_ioapic_support();
1050 }
1051 
1052 /*
1053  * Fall back to non SMP mode after errors.
1054  *
1055  * RED-PEN audit/test this more. I bet there is more state messed up here.
1056  */
1057 static __init void disable_smp(void)
1058 {
1059 	pr_info("SMP disabled\n");
1060 
1061 	disable_ioapic_support();
1062 	topology_reset_possible_cpus_up();
1063 
1064 	cpumask_set_cpu(0, topology_sibling_cpumask(0));
1065 	cpumask_set_cpu(0, topology_core_cpumask(0));
1066 	cpumask_set_cpu(0, topology_die_cpumask(0));
1067 }
1068 
1069 void __init smp_prepare_cpus_common(void)
1070 {
1071 	unsigned int cpu, node;
1072 
1073 	/* Mark all except the boot CPU as hotpluggable */
1074 	for_each_possible_cpu(cpu) {
1075 		if (cpu)
1076 			per_cpu(cpu_info.cpu_index, cpu) = nr_cpu_ids;
1077 	}
1078 
1079 	for_each_possible_cpu(cpu) {
1080 		node = cpu_to_node(cpu);
1081 
1082 		zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map,    cpu), GFP_KERNEL, node);
1083 		zalloc_cpumask_var_node(&per_cpu(cpu_core_map,       cpu), GFP_KERNEL, node);
1084 		zalloc_cpumask_var_node(&per_cpu(cpu_die_map,        cpu), GFP_KERNEL, node);
1085 		zalloc_cpumask_var_node(&per_cpu(cpu_llc_shared_map, cpu), GFP_KERNEL, node);
1086 		zalloc_cpumask_var_node(&per_cpu(cpu_l2c_shared_map, cpu), GFP_KERNEL, node);
1087 	}
1088 
1089 	set_cpu_sibling_map(0);
1090 }
1091 
1092 void __init smp_prepare_boot_cpu(void)
1093 {
1094 	smp_ops.smp_prepare_boot_cpu();
1095 }
1096 
1097 #ifdef CONFIG_X86_64
1098 /* Establish whether parallel bringup can be supported. */
1099 bool __init arch_cpuhp_init_parallel_bringup(void)
1100 {
1101 	if (!x86_cpuinit.parallel_bringup) {
1102 		pr_info("Parallel CPU startup disabled by the platform\n");
1103 		return false;
1104 	}
1105 
1106 	smpboot_control = STARTUP_READ_APICID;
1107 	pr_debug("Parallel CPU startup enabled: 0x%08x\n", smpboot_control);
1108 	return true;
1109 }
1110 #endif
1111 
1112 /*
1113  * Prepare for SMP bootup.
1114  * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1115  *            for common interface support.
1116  */
1117 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1118 {
1119 	smp_prepare_cpus_common();
1120 
1121 	switch (apic_intr_mode) {
1122 	case APIC_PIC:
1123 	case APIC_VIRTUAL_WIRE_NO_CONFIG:
1124 		disable_smp();
1125 		return;
1126 	case APIC_SYMMETRIC_IO_NO_ROUTING:
1127 		disable_smp();
1128 		/* Setup local timer */
1129 		x86_init.timers.setup_percpu_clockev();
1130 		return;
1131 	case APIC_VIRTUAL_WIRE:
1132 	case APIC_SYMMETRIC_IO:
1133 		break;
1134 	}
1135 
1136 	/* Setup local timer */
1137 	x86_init.timers.setup_percpu_clockev();
1138 
1139 	pr_info("CPU0: ");
1140 	print_cpu_info(&cpu_data(0));
1141 
1142 	uv_system_init();
1143 
1144 	smp_set_init_udelay();
1145 
1146 	speculative_store_bypass_ht_init();
1147 
1148 	snp_set_wakeup_secondary_cpu();
1149 }
1150 
1151 void arch_thaw_secondary_cpus_begin(void)
1152 {
1153 	set_cache_aps_delayed_init(true);
1154 }
1155 
1156 void arch_thaw_secondary_cpus_end(void)
1157 {
1158 	cache_aps_init();
1159 }
1160 
1161 /*
1162  * Early setup to make printk work.
1163  */
1164 void __init native_smp_prepare_boot_cpu(void)
1165 {
1166 	int me = smp_processor_id();
1167 
1168 	/* SMP handles this from setup_per_cpu_areas() */
1169 	if (!IS_ENABLED(CONFIG_SMP))
1170 		switch_gdt_and_percpu_base(me);
1171 
1172 	native_pv_lock_init();
1173 }
1174 
1175 void __init native_smp_cpus_done(unsigned int max_cpus)
1176 {
1177 	pr_debug("Boot done\n");
1178 
1179 	build_sched_topology();
1180 	nmi_selftest();
1181 	impress_friends();
1182 	cache_aps_init();
1183 }
1184 
1185 /* correctly size the local cpu masks */
1186 void __init setup_cpu_local_masks(void)
1187 {
1188 	alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
1189 }
1190 
1191 #ifdef CONFIG_HOTPLUG_CPU
1192 
1193 /* Recompute SMT state for all CPUs on offline */
1194 static void recompute_smt_state(void)
1195 {
1196 	int max_threads, cpu;
1197 
1198 	max_threads = 0;
1199 	for_each_online_cpu (cpu) {
1200 		int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1201 
1202 		if (threads > max_threads)
1203 			max_threads = threads;
1204 	}
1205 	__max_smt_threads = max_threads;
1206 }
1207 
1208 static void remove_siblinginfo(int cpu)
1209 {
1210 	int sibling;
1211 	struct cpuinfo_x86 *c = &cpu_data(cpu);
1212 
1213 	for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1214 		cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1215 		/*/
1216 		 * last thread sibling in this cpu core going down
1217 		 */
1218 		if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1219 			cpu_data(sibling).booted_cores--;
1220 	}
1221 
1222 	for_each_cpu(sibling, topology_die_cpumask(cpu))
1223 		cpumask_clear_cpu(cpu, topology_die_cpumask(sibling));
1224 
1225 	for_each_cpu(sibling, topology_sibling_cpumask(cpu)) {
1226 		cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1227 		if (cpumask_weight(topology_sibling_cpumask(sibling)) == 1)
1228 			cpu_data(sibling).smt_active = false;
1229 	}
1230 
1231 	for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1232 		cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1233 	for_each_cpu(sibling, cpu_l2c_shared_mask(cpu))
1234 		cpumask_clear_cpu(cpu, cpu_l2c_shared_mask(sibling));
1235 	cpumask_clear(cpu_llc_shared_mask(cpu));
1236 	cpumask_clear(cpu_l2c_shared_mask(cpu));
1237 	cpumask_clear(topology_sibling_cpumask(cpu));
1238 	cpumask_clear(topology_core_cpumask(cpu));
1239 	cpumask_clear(topology_die_cpumask(cpu));
1240 	c->topo.core_id = 0;
1241 	c->booted_cores = 0;
1242 	cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1243 	recompute_smt_state();
1244 }
1245 
1246 static void remove_cpu_from_maps(int cpu)
1247 {
1248 	set_cpu_online(cpu, false);
1249 	numa_remove_cpu(cpu);
1250 }
1251 
1252 void cpu_disable_common(void)
1253 {
1254 	int cpu = smp_processor_id();
1255 
1256 	remove_siblinginfo(cpu);
1257 
1258 	/*
1259 	 * Stop allowing kernel-mode FPU. This is needed so that if the CPU is
1260 	 * brought online again, the initial state is not allowed:
1261 	 */
1262 	this_cpu_write(kernel_fpu_allowed, false);
1263 
1264 	/* It's now safe to remove this processor from the online map */
1265 	lock_vector_lock();
1266 	remove_cpu_from_maps(cpu);
1267 	unlock_vector_lock();
1268 	fixup_irqs();
1269 	lapic_offline();
1270 }
1271 
1272 int native_cpu_disable(void)
1273 {
1274 	int ret;
1275 
1276 	ret = lapic_can_unplug_cpu();
1277 	if (ret)
1278 		return ret;
1279 
1280 	cpu_disable_common();
1281 
1282         /*
1283          * Disable the local APIC. Otherwise IPI broadcasts will reach
1284          * it. It still responds normally to INIT, NMI, SMI, and SIPI
1285          * messages.
1286          *
1287          * Disabling the APIC must happen after cpu_disable_common()
1288          * which invokes fixup_irqs().
1289          *
1290          * Disabling the APIC preserves already set bits in IRR, but
1291          * an interrupt arriving after disabling the local APIC does not
1292          * set the corresponding IRR bit.
1293          *
1294          * fixup_irqs() scans IRR for set bits so it can raise a not
1295          * yet handled interrupt on the new destination CPU via an IPI
1296          * but obviously it can't do so for IRR bits which are not set.
1297          * IOW, interrupts arriving after disabling the local APIC will
1298          * be lost.
1299          */
1300 	apic_soft_disable();
1301 
1302 	return 0;
1303 }
1304 
1305 void play_dead_common(void)
1306 {
1307 	idle_task_exit();
1308 
1309 	cpuhp_ap_report_dead();
1310 
1311 	local_irq_disable();
1312 }
1313 
1314 /*
1315  * We need to flush the caches before going to sleep, lest we have
1316  * dirty data in our caches when we come back up.
1317  */
1318 void __noreturn mwait_play_dead(unsigned int eax_hint)
1319 {
1320 	struct mwait_cpu_dead *md = this_cpu_ptr(&mwait_cpu_dead);
1321 
1322 	/* Set up state for the kexec() hack below */
1323 	md->status = CPUDEAD_MWAIT_WAIT;
1324 	md->control = CPUDEAD_MWAIT_WAIT;
1325 
1326 	wbinvd();
1327 
1328 	while (1) {
1329 		/*
1330 		 * The CLFLUSH is a workaround for erratum AAI65 for
1331 		 * the Xeon 7400 series.  It's not clear it is actually
1332 		 * needed, but it should be harmless in either case.
1333 		 * The WBINVD is insufficient due to the spurious-wakeup
1334 		 * case where we return around the loop.
1335 		 */
1336 		mb();
1337 		clflush(md);
1338 		mb();
1339 		__monitor(md, 0, 0);
1340 		mb();
1341 		__mwait(eax_hint, 0);
1342 
1343 		if (READ_ONCE(md->control) == CPUDEAD_MWAIT_KEXEC_HLT) {
1344 			/*
1345 			 * Kexec is about to happen. Don't go back into mwait() as
1346 			 * the kexec kernel might overwrite text and data including
1347 			 * page tables and stack. So mwait() would resume when the
1348 			 * monitor cache line is written to and then the CPU goes
1349 			 * south due to overwritten text, page tables and stack.
1350 			 *
1351 			 * Note: This does _NOT_ protect against a stray MCE, NMI,
1352 			 * SMI. They will resume execution at the instruction
1353 			 * following the HLT instruction and run into the problem
1354 			 * which this is trying to prevent.
1355 			 */
1356 			WRITE_ONCE(md->status, CPUDEAD_MWAIT_KEXEC_HLT);
1357 			while(1)
1358 				native_halt();
1359 		}
1360 	}
1361 }
1362 
1363 /*
1364  * Kick all "offline" CPUs out of mwait on kexec(). See comment in
1365  * mwait_play_dead().
1366  */
1367 void smp_kick_mwait_play_dead(void)
1368 {
1369 	u32 newstate = CPUDEAD_MWAIT_KEXEC_HLT;
1370 	struct mwait_cpu_dead *md;
1371 	unsigned int cpu, i;
1372 
1373 	for_each_cpu_andnot(cpu, cpu_present_mask, cpu_online_mask) {
1374 		md = per_cpu_ptr(&mwait_cpu_dead, cpu);
1375 
1376 		/* Does it sit in mwait_play_dead() ? */
1377 		if (READ_ONCE(md->status) != CPUDEAD_MWAIT_WAIT)
1378 			continue;
1379 
1380 		/* Wait up to 5ms */
1381 		for (i = 0; READ_ONCE(md->status) != newstate && i < 1000; i++) {
1382 			/* Bring it out of mwait */
1383 			WRITE_ONCE(md->control, newstate);
1384 			udelay(5);
1385 		}
1386 
1387 		if (READ_ONCE(md->status) != newstate)
1388 			pr_err_once("CPU%u is stuck in mwait_play_dead()\n", cpu);
1389 	}
1390 }
1391 
1392 void __noreturn hlt_play_dead(void)
1393 {
1394 	if (__this_cpu_read(cpu_info.x86) >= 4)
1395 		wbinvd();
1396 
1397 	while (1)
1398 		native_halt();
1399 }
1400 
1401 /*
1402  * native_play_dead() is essentially a __noreturn function, but it can't
1403  * be marked as such as the compiler may complain about it.
1404  */
1405 void native_play_dead(void)
1406 {
1407 	if (cpu_feature_enabled(X86_FEATURE_KERNEL_IBRS))
1408 		__update_spec_ctrl(0);
1409 
1410 	play_dead_common();
1411 	tboot_shutdown(TB_SHUTDOWN_WFS);
1412 
1413 	/* Below returns only on error. */
1414 	cpuidle_play_dead();
1415 	hlt_play_dead();
1416 }
1417 
1418 #else /* ... !CONFIG_HOTPLUG_CPU */
1419 int native_cpu_disable(void)
1420 {
1421 	return -ENOSYS;
1422 }
1423 
1424 void native_play_dead(void)
1425 {
1426 	BUG();
1427 }
1428 
1429 #endif
1430