xref: /openbmc/linux/arch/x86/kernel/smpboot.c (revision 240e6d25)
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/numa.h>
57 #include <linux/pgtable.h>
58 #include <linux/overflow.h>
59 #include <linux/syscore_ops.h>
60 
61 #include <asm/acpi.h>
62 #include <asm/desc.h>
63 #include <asm/nmi.h>
64 #include <asm/irq.h>
65 #include <asm/realmode.h>
66 #include <asm/cpu.h>
67 #include <asm/numa.h>
68 #include <asm/tlbflush.h>
69 #include <asm/mtrr.h>
70 #include <asm/mwait.h>
71 #include <asm/apic.h>
72 #include <asm/io_apic.h>
73 #include <asm/fpu/api.h>
74 #include <asm/setup.h>
75 #include <asm/uv/uv.h>
76 #include <linux/mc146818rtc.h>
77 #include <asm/i8259.h>
78 #include <asm/misc.h>
79 #include <asm/qspinlock.h>
80 #include <asm/intel-family.h>
81 #include <asm/cpu_device_id.h>
82 #include <asm/spec-ctrl.h>
83 #include <asm/hw_irq.h>
84 #include <asm/stackprotector.h>
85 
86 #ifdef CONFIG_ACPI_CPPC_LIB
87 #include <acpi/cppc_acpi.h>
88 #endif
89 
90 /* representing HT siblings of each logical CPU */
91 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
92 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
93 
94 /* representing HT and core siblings of each logical CPU */
95 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
96 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
97 
98 /* representing HT, core, and die siblings of each logical CPU */
99 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_die_map);
100 EXPORT_PER_CPU_SYMBOL(cpu_die_map);
101 
102 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_llc_shared_map);
103 
104 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_l2c_shared_map);
105 
106 /* Per CPU bogomips and other parameters */
107 DEFINE_PER_CPU_READ_MOSTLY(struct cpuinfo_x86, cpu_info);
108 EXPORT_PER_CPU_SYMBOL(cpu_info);
109 
110 /* Logical package management. We might want to allocate that dynamically */
111 unsigned int __max_logical_packages __read_mostly;
112 EXPORT_SYMBOL(__max_logical_packages);
113 static unsigned int logical_packages __read_mostly;
114 static unsigned int logical_die __read_mostly;
115 
116 /* Maximum number of SMT threads on any online core */
117 int __read_mostly __max_smt_threads = 1;
118 
119 /* Flag to indicate if a complete sched domain rebuild is required */
120 bool x86_topology_update;
121 
122 int arch_update_cpu_topology(void)
123 {
124 	int retval = x86_topology_update;
125 
126 	x86_topology_update = false;
127 	return retval;
128 }
129 
130 static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
131 {
132 	unsigned long flags;
133 
134 	spin_lock_irqsave(&rtc_lock, flags);
135 	CMOS_WRITE(0xa, 0xf);
136 	spin_unlock_irqrestore(&rtc_lock, flags);
137 	*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) =
138 							start_eip >> 4;
139 	*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) =
140 							start_eip & 0xf;
141 }
142 
143 static inline void smpboot_restore_warm_reset_vector(void)
144 {
145 	unsigned long flags;
146 
147 	/*
148 	 * Paranoid:  Set warm reset code and vector here back
149 	 * to default values.
150 	 */
151 	spin_lock_irqsave(&rtc_lock, flags);
152 	CMOS_WRITE(0, 0xf);
153 	spin_unlock_irqrestore(&rtc_lock, flags);
154 
155 	*((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
156 }
157 
158 static void init_freq_invariance(bool secondary, bool cppc_ready);
159 
160 /*
161  * Report back to the Boot Processor during boot time or to the caller processor
162  * during CPU online.
163  */
164 static void smp_callin(void)
165 {
166 	int cpuid;
167 
168 	/*
169 	 * If waken up by an INIT in an 82489DX configuration
170 	 * cpu_callout_mask guarantees we don't get here before
171 	 * an INIT_deassert IPI reaches our local APIC, so it is
172 	 * now safe to touch our local APIC.
173 	 */
174 	cpuid = smp_processor_id();
175 
176 	/*
177 	 * the boot CPU has finished the init stage and is spinning
178 	 * on callin_map until we finish. We are free to set up this
179 	 * CPU, first the APIC. (this is probably redundant on most
180 	 * boards)
181 	 */
182 	apic_ap_setup();
183 
184 	/*
185 	 * Save our processor parameters. Note: this information
186 	 * is needed for clock calibration.
187 	 */
188 	smp_store_cpu_info(cpuid);
189 
190 	/*
191 	 * The topology information must be up to date before
192 	 * calibrate_delay() and notify_cpu_starting().
193 	 */
194 	set_cpu_sibling_map(raw_smp_processor_id());
195 
196 	init_freq_invariance(true, false);
197 
198 	/*
199 	 * Get our bogomips.
200 	 * Update loops_per_jiffy in cpu_data. Previous call to
201 	 * smp_store_cpu_info() stored a value that is close but not as
202 	 * accurate as the value just calculated.
203 	 */
204 	calibrate_delay();
205 	cpu_data(cpuid).loops_per_jiffy = loops_per_jiffy;
206 	pr_debug("Stack at about %p\n", &cpuid);
207 
208 	wmb();
209 
210 	notify_cpu_starting(cpuid);
211 
212 	/*
213 	 * Allow the master to continue.
214 	 */
215 	cpumask_set_cpu(cpuid, cpu_callin_mask);
216 }
217 
218 static int cpu0_logical_apicid;
219 static int enable_start_cpu0;
220 /*
221  * Activate a secondary processor.
222  */
223 static void notrace start_secondary(void *unused)
224 {
225 	/*
226 	 * Don't put *anything* except direct CPU state initialization
227 	 * before cpu_init(), SMP booting is too fragile that we want to
228 	 * limit the things done here to the most necessary things.
229 	 */
230 	cr4_init();
231 
232 #ifdef CONFIG_X86_32
233 	/* switch away from the initial page table */
234 	load_cr3(swapper_pg_dir);
235 	__flush_tlb_all();
236 #endif
237 	cpu_init_secondary();
238 	rcu_cpu_starting(raw_smp_processor_id());
239 	x86_cpuinit.early_percpu_clock_init();
240 	smp_callin();
241 
242 	enable_start_cpu0 = 0;
243 
244 	/* otherwise gcc will move up smp_processor_id before the cpu_init */
245 	barrier();
246 	/*
247 	 * Check TSC synchronization with the boot CPU:
248 	 */
249 	check_tsc_sync_target();
250 
251 	speculative_store_bypass_ht_init();
252 
253 	/*
254 	 * Lock vector_lock, set CPU online and bring the vector
255 	 * allocator online. Online must be set with vector_lock held
256 	 * to prevent a concurrent irq setup/teardown from seeing a
257 	 * half valid vector space.
258 	 */
259 	lock_vector_lock();
260 	set_cpu_online(smp_processor_id(), true);
261 	lapic_online();
262 	unlock_vector_lock();
263 	cpu_set_state_online(smp_processor_id());
264 	x86_platform.nmi_init();
265 
266 	/* enable local interrupts */
267 	local_irq_enable();
268 
269 	x86_cpuinit.setup_percpu_clockev();
270 
271 	wmb();
272 	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
273 }
274 
275 /**
276  * topology_is_primary_thread - Check whether CPU is the primary SMT thread
277  * @cpu:	CPU to check
278  */
279 bool topology_is_primary_thread(unsigned int cpu)
280 {
281 	return apic_id_is_primary_thread(per_cpu(x86_cpu_to_apicid, cpu));
282 }
283 
284 /**
285  * topology_smt_supported - Check whether SMT is supported by the CPUs
286  */
287 bool topology_smt_supported(void)
288 {
289 	return smp_num_siblings > 1;
290 }
291 
292 /**
293  * topology_phys_to_logical_pkg - Map a physical package id to a logical
294  *
295  * Returns logical package id or -1 if not found
296  */
297 int topology_phys_to_logical_pkg(unsigned int phys_pkg)
298 {
299 	int cpu;
300 
301 	for_each_possible_cpu(cpu) {
302 		struct cpuinfo_x86 *c = &cpu_data(cpu);
303 
304 		if (c->initialized && c->phys_proc_id == phys_pkg)
305 			return c->logical_proc_id;
306 	}
307 	return -1;
308 }
309 EXPORT_SYMBOL(topology_phys_to_logical_pkg);
310 /**
311  * topology_phys_to_logical_die - Map a physical die id to logical
312  *
313  * Returns logical die id or -1 if not found
314  */
315 int topology_phys_to_logical_die(unsigned int die_id, unsigned int cur_cpu)
316 {
317 	int cpu;
318 	int proc_id = cpu_data(cur_cpu).phys_proc_id;
319 
320 	for_each_possible_cpu(cpu) {
321 		struct cpuinfo_x86 *c = &cpu_data(cpu);
322 
323 		if (c->initialized && c->cpu_die_id == die_id &&
324 		    c->phys_proc_id == proc_id)
325 			return c->logical_die_id;
326 	}
327 	return -1;
328 }
329 EXPORT_SYMBOL(topology_phys_to_logical_die);
330 
331 /**
332  * topology_update_package_map - Update the physical to logical package map
333  * @pkg:	The physical package id as retrieved via CPUID
334  * @cpu:	The cpu for which this is updated
335  */
336 int topology_update_package_map(unsigned int pkg, unsigned int cpu)
337 {
338 	int new;
339 
340 	/* Already available somewhere? */
341 	new = topology_phys_to_logical_pkg(pkg);
342 	if (new >= 0)
343 		goto found;
344 
345 	new = logical_packages++;
346 	if (new != pkg) {
347 		pr_info("CPU %u Converting physical %u to logical package %u\n",
348 			cpu, pkg, new);
349 	}
350 found:
351 	cpu_data(cpu).logical_proc_id = new;
352 	return 0;
353 }
354 /**
355  * topology_update_die_map - Update the physical to logical die map
356  * @die:	The die id as retrieved via CPUID
357  * @cpu:	The cpu for which this is updated
358  */
359 int topology_update_die_map(unsigned int die, unsigned int cpu)
360 {
361 	int new;
362 
363 	/* Already available somewhere? */
364 	new = topology_phys_to_logical_die(die, cpu);
365 	if (new >= 0)
366 		goto found;
367 
368 	new = logical_die++;
369 	if (new != die) {
370 		pr_info("CPU %u Converting physical %u to logical die %u\n",
371 			cpu, die, new);
372 	}
373 found:
374 	cpu_data(cpu).logical_die_id = new;
375 	return 0;
376 }
377 
378 void __init smp_store_boot_cpu_info(void)
379 {
380 	int id = 0; /* CPU 0 */
381 	struct cpuinfo_x86 *c = &cpu_data(id);
382 
383 	*c = boot_cpu_data;
384 	c->cpu_index = id;
385 	topology_update_package_map(c->phys_proc_id, id);
386 	topology_update_die_map(c->cpu_die_id, id);
387 	c->initialized = true;
388 }
389 
390 /*
391  * The bootstrap kernel entry code has set these up. Save them for
392  * a given CPU
393  */
394 void smp_store_cpu_info(int id)
395 {
396 	struct cpuinfo_x86 *c = &cpu_data(id);
397 
398 	/* Copy boot_cpu_data only on the first bringup */
399 	if (!c->initialized)
400 		*c = boot_cpu_data;
401 	c->cpu_index = id;
402 	/*
403 	 * During boot time, CPU0 has this setup already. Save the info when
404 	 * bringing up AP or offlined CPU0.
405 	 */
406 	identify_secondary_cpu(c);
407 	c->initialized = true;
408 }
409 
410 static bool
411 topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
412 {
413 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
414 
415 	return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
416 }
417 
418 static bool
419 topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
420 {
421 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
422 
423 	return !WARN_ONCE(!topology_same_node(c, o),
424 		"sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
425 		"[node: %d != %d]. Ignoring dependency.\n",
426 		cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
427 }
428 
429 #define link_mask(mfunc, c1, c2)					\
430 do {									\
431 	cpumask_set_cpu((c1), mfunc(c2));				\
432 	cpumask_set_cpu((c2), mfunc(c1));				\
433 } while (0)
434 
435 static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
436 {
437 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
438 		int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
439 
440 		if (c->phys_proc_id == o->phys_proc_id &&
441 		    c->cpu_die_id == o->cpu_die_id &&
442 		    per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2)) {
443 			if (c->cpu_core_id == o->cpu_core_id)
444 				return topology_sane(c, o, "smt");
445 
446 			if ((c->cu_id != 0xff) &&
447 			    (o->cu_id != 0xff) &&
448 			    (c->cu_id == o->cu_id))
449 				return topology_sane(c, o, "smt");
450 		}
451 
452 	} else if (c->phys_proc_id == o->phys_proc_id &&
453 		   c->cpu_die_id == o->cpu_die_id &&
454 		   c->cpu_core_id == o->cpu_core_id) {
455 		return topology_sane(c, o, "smt");
456 	}
457 
458 	return false;
459 }
460 
461 static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
462 {
463 	if (c->phys_proc_id == o->phys_proc_id &&
464 	    c->cpu_die_id == o->cpu_die_id)
465 		return true;
466 	return false;
467 }
468 
469 static bool match_l2c(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
470 {
471 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
472 
473 	/* If the arch didn't set up l2c_id, fall back to SMT */
474 	if (per_cpu(cpu_l2c_id, cpu1) == BAD_APICID)
475 		return match_smt(c, o);
476 
477 	/* Do not match if L2 cache id does not match: */
478 	if (per_cpu(cpu_l2c_id, cpu1) != per_cpu(cpu_l2c_id, cpu2))
479 		return false;
480 
481 	return topology_sane(c, o, "l2c");
482 }
483 
484 /*
485  * Unlike the other levels, we do not enforce keeping a
486  * multicore group inside a NUMA node.  If this happens, we will
487  * discard the MC level of the topology later.
488  */
489 static bool match_pkg(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
490 {
491 	if (c->phys_proc_id == o->phys_proc_id)
492 		return true;
493 	return false;
494 }
495 
496 /*
497  * Define intel_cod_cpu[] for Intel COD (Cluster-on-Die) CPUs.
498  *
499  * Any Intel CPU that has multiple nodes per package and does not
500  * match intel_cod_cpu[] has the SNC (Sub-NUMA Cluster) topology.
501  *
502  * When in SNC mode, these CPUs enumerate an LLC that is shared
503  * by multiple NUMA nodes. The LLC is shared for off-package data
504  * access but private to the NUMA node (half of the package) for
505  * on-package access. CPUID (the source of the information about
506  * the LLC) can only enumerate the cache as shared or unshared,
507  * but not this particular configuration.
508  */
509 
510 static const struct x86_cpu_id intel_cod_cpu[] = {
511 	X86_MATCH_INTEL_FAM6_MODEL(HASWELL_X, 0),	/* COD */
512 	X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_X, 0),	/* COD */
513 	X86_MATCH_INTEL_FAM6_MODEL(ANY, 1),		/* SNC */
514 	{}
515 };
516 
517 static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
518 {
519 	const struct x86_cpu_id *id = x86_match_cpu(intel_cod_cpu);
520 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
521 	bool intel_snc = id && id->driver_data;
522 
523 	/* Do not match if we do not have a valid APICID for cpu: */
524 	if (per_cpu(cpu_llc_id, cpu1) == BAD_APICID)
525 		return false;
526 
527 	/* Do not match if LLC id does not match: */
528 	if (per_cpu(cpu_llc_id, cpu1) != per_cpu(cpu_llc_id, cpu2))
529 		return false;
530 
531 	/*
532 	 * Allow the SNC topology without warning. Return of false
533 	 * means 'c' does not share the LLC of 'o'. This will be
534 	 * reflected to userspace.
535 	 */
536 	if (match_pkg(c, o) && !topology_same_node(c, o) && intel_snc)
537 		return false;
538 
539 	return topology_sane(c, o, "llc");
540 }
541 
542 
543 #if defined(CONFIG_SCHED_SMT) || defined(CONFIG_SCHED_CLUSTER) || defined(CONFIG_SCHED_MC)
544 static inline int x86_sched_itmt_flags(void)
545 {
546 	return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
547 }
548 
549 #ifdef CONFIG_SCHED_MC
550 static int x86_core_flags(void)
551 {
552 	return cpu_core_flags() | x86_sched_itmt_flags();
553 }
554 #endif
555 #ifdef CONFIG_SCHED_SMT
556 static int x86_smt_flags(void)
557 {
558 	return cpu_smt_flags() | x86_sched_itmt_flags();
559 }
560 #endif
561 #ifdef CONFIG_SCHED_CLUSTER
562 static int x86_cluster_flags(void)
563 {
564 	return cpu_cluster_flags() | x86_sched_itmt_flags();
565 }
566 #endif
567 #endif
568 
569 static struct sched_domain_topology_level x86_numa_in_package_topology[] = {
570 #ifdef CONFIG_SCHED_SMT
571 	{ cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
572 #endif
573 #ifdef CONFIG_SCHED_CLUSTER
574 	{ cpu_clustergroup_mask, x86_cluster_flags, SD_INIT_NAME(CLS) },
575 #endif
576 #ifdef CONFIG_SCHED_MC
577 	{ cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
578 #endif
579 	{ NULL, },
580 };
581 
582 static struct sched_domain_topology_level x86_topology[] = {
583 #ifdef CONFIG_SCHED_SMT
584 	{ cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
585 #endif
586 #ifdef CONFIG_SCHED_CLUSTER
587 	{ cpu_clustergroup_mask, x86_cluster_flags, SD_INIT_NAME(CLS) },
588 #endif
589 #ifdef CONFIG_SCHED_MC
590 	{ cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
591 #endif
592 	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
593 	{ NULL, },
594 };
595 
596 /*
597  * Set if a package/die has multiple NUMA nodes inside.
598  * AMD Magny-Cours, Intel Cluster-on-Die, and Intel
599  * Sub-NUMA Clustering have this.
600  */
601 static bool x86_has_numa_in_package;
602 
603 void set_cpu_sibling_map(int cpu)
604 {
605 	bool has_smt = smp_num_siblings > 1;
606 	bool has_mp = has_smt || boot_cpu_data.x86_max_cores > 1;
607 	struct cpuinfo_x86 *c = &cpu_data(cpu);
608 	struct cpuinfo_x86 *o;
609 	int i, threads;
610 
611 	cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
612 
613 	if (!has_mp) {
614 		cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
615 		cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
616 		cpumask_set_cpu(cpu, cpu_l2c_shared_mask(cpu));
617 		cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
618 		cpumask_set_cpu(cpu, topology_die_cpumask(cpu));
619 		c->booted_cores = 1;
620 		return;
621 	}
622 
623 	for_each_cpu(i, cpu_sibling_setup_mask) {
624 		o = &cpu_data(i);
625 
626 		if (match_pkg(c, o) && !topology_same_node(c, o))
627 			x86_has_numa_in_package = true;
628 
629 		if ((i == cpu) || (has_smt && match_smt(c, o)))
630 			link_mask(topology_sibling_cpumask, cpu, i);
631 
632 		if ((i == cpu) || (has_mp && match_llc(c, o)))
633 			link_mask(cpu_llc_shared_mask, cpu, i);
634 
635 		if ((i == cpu) || (has_mp && match_l2c(c, o)))
636 			link_mask(cpu_l2c_shared_mask, cpu, i);
637 
638 		if ((i == cpu) || (has_mp && match_die(c, o)))
639 			link_mask(topology_die_cpumask, cpu, i);
640 	}
641 
642 	threads = cpumask_weight(topology_sibling_cpumask(cpu));
643 	if (threads > __max_smt_threads)
644 		__max_smt_threads = threads;
645 
646 	for_each_cpu(i, topology_sibling_cpumask(cpu))
647 		cpu_data(i).smt_active = threads > 1;
648 
649 	/*
650 	 * This needs a separate iteration over the cpus because we rely on all
651 	 * topology_sibling_cpumask links to be set-up.
652 	 */
653 	for_each_cpu(i, cpu_sibling_setup_mask) {
654 		o = &cpu_data(i);
655 
656 		if ((i == cpu) || (has_mp && match_pkg(c, o))) {
657 			link_mask(topology_core_cpumask, cpu, i);
658 
659 			/*
660 			 *  Does this new cpu bringup a new core?
661 			 */
662 			if (threads == 1) {
663 				/*
664 				 * for each core in package, increment
665 				 * the booted_cores for this new cpu
666 				 */
667 				if (cpumask_first(
668 				    topology_sibling_cpumask(i)) == i)
669 					c->booted_cores++;
670 				/*
671 				 * increment the core count for all
672 				 * the other cpus in this package
673 				 */
674 				if (i != cpu)
675 					cpu_data(i).booted_cores++;
676 			} else if (i != cpu && !c->booted_cores)
677 				c->booted_cores = cpu_data(i).booted_cores;
678 		}
679 	}
680 }
681 
682 /* maps the cpu to the sched domain representing multi-core */
683 const struct cpumask *cpu_coregroup_mask(int cpu)
684 {
685 	return cpu_llc_shared_mask(cpu);
686 }
687 
688 const struct cpumask *cpu_clustergroup_mask(int cpu)
689 {
690 	return cpu_l2c_shared_mask(cpu);
691 }
692 
693 static void impress_friends(void)
694 {
695 	int cpu;
696 	unsigned long bogosum = 0;
697 	/*
698 	 * Allow the user to impress friends.
699 	 */
700 	pr_debug("Before bogomips\n");
701 	for_each_possible_cpu(cpu)
702 		if (cpumask_test_cpu(cpu, cpu_callout_mask))
703 			bogosum += cpu_data(cpu).loops_per_jiffy;
704 	pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
705 		num_online_cpus(),
706 		bogosum/(500000/HZ),
707 		(bogosum/(5000/HZ))%100);
708 
709 	pr_debug("Before bogocount - setting activated=1\n");
710 }
711 
712 void __inquire_remote_apic(int apicid)
713 {
714 	unsigned i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
715 	const char * const names[] = { "ID", "VERSION", "SPIV" };
716 	int timeout;
717 	u32 status;
718 
719 	pr_info("Inquiring remote APIC 0x%x...\n", apicid);
720 
721 	for (i = 0; i < ARRAY_SIZE(regs); i++) {
722 		pr_info("... APIC 0x%x %s: ", apicid, names[i]);
723 
724 		/*
725 		 * Wait for idle.
726 		 */
727 		status = safe_apic_wait_icr_idle();
728 		if (status)
729 			pr_cont("a previous APIC delivery may have failed\n");
730 
731 		apic_icr_write(APIC_DM_REMRD | regs[i], apicid);
732 
733 		timeout = 0;
734 		do {
735 			udelay(100);
736 			status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
737 		} while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
738 
739 		switch (status) {
740 		case APIC_ICR_RR_VALID:
741 			status = apic_read(APIC_RRR);
742 			pr_cont("%08x\n", status);
743 			break;
744 		default:
745 			pr_cont("failed\n");
746 		}
747 	}
748 }
749 
750 /*
751  * The Multiprocessor Specification 1.4 (1997) example code suggests
752  * that there should be a 10ms delay between the BSP asserting INIT
753  * and de-asserting INIT, when starting a remote processor.
754  * But that slows boot and resume on modern processors, which include
755  * many cores and don't require that delay.
756  *
757  * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
758  * Modern processor families are quirked to remove the delay entirely.
759  */
760 #define UDELAY_10MS_DEFAULT 10000
761 
762 static unsigned int init_udelay = UINT_MAX;
763 
764 static int __init cpu_init_udelay(char *str)
765 {
766 	get_option(&str, &init_udelay);
767 
768 	return 0;
769 }
770 early_param("cpu_init_udelay", cpu_init_udelay);
771 
772 static void __init smp_quirk_init_udelay(void)
773 {
774 	/* if cmdline changed it from default, leave it alone */
775 	if (init_udelay != UINT_MAX)
776 		return;
777 
778 	/* if modern processor, use no delay */
779 	if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
780 	    ((boot_cpu_data.x86_vendor == X86_VENDOR_HYGON) && (boot_cpu_data.x86 >= 0x18)) ||
781 	    ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
782 		init_udelay = 0;
783 		return;
784 	}
785 	/* else, use legacy delay */
786 	init_udelay = UDELAY_10MS_DEFAULT;
787 }
788 
789 /*
790  * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
791  * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
792  * won't ... remember to clear down the APIC, etc later.
793  */
794 int
795 wakeup_secondary_cpu_via_nmi(int apicid, unsigned long start_eip)
796 {
797 	u32 dm = apic->dest_mode_logical ? APIC_DEST_LOGICAL : APIC_DEST_PHYSICAL;
798 	unsigned long send_status, accept_status = 0;
799 	int maxlvt;
800 
801 	/* Target chip */
802 	/* Boot on the stack */
803 	/* Kick the second */
804 	apic_icr_write(APIC_DM_NMI | dm, apicid);
805 
806 	pr_debug("Waiting for send to finish...\n");
807 	send_status = safe_apic_wait_icr_idle();
808 
809 	/*
810 	 * Give the other CPU some time to accept the IPI.
811 	 */
812 	udelay(200);
813 	if (APIC_INTEGRATED(boot_cpu_apic_version)) {
814 		maxlvt = lapic_get_maxlvt();
815 		if (maxlvt > 3)			/* Due to the Pentium erratum 3AP.  */
816 			apic_write(APIC_ESR, 0);
817 		accept_status = (apic_read(APIC_ESR) & 0xEF);
818 	}
819 	pr_debug("NMI sent\n");
820 
821 	if (send_status)
822 		pr_err("APIC never delivered???\n");
823 	if (accept_status)
824 		pr_err("APIC delivery error (%lx)\n", accept_status);
825 
826 	return (send_status | accept_status);
827 }
828 
829 static int
830 wakeup_secondary_cpu_via_init(int phys_apicid, unsigned long start_eip)
831 {
832 	unsigned long send_status = 0, accept_status = 0;
833 	int maxlvt, num_starts, j;
834 
835 	maxlvt = lapic_get_maxlvt();
836 
837 	/*
838 	 * Be paranoid about clearing APIC errors.
839 	 */
840 	if (APIC_INTEGRATED(boot_cpu_apic_version)) {
841 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
842 			apic_write(APIC_ESR, 0);
843 		apic_read(APIC_ESR);
844 	}
845 
846 	pr_debug("Asserting INIT\n");
847 
848 	/*
849 	 * Turn INIT on target chip
850 	 */
851 	/*
852 	 * Send IPI
853 	 */
854 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT,
855 		       phys_apicid);
856 
857 	pr_debug("Waiting for send to finish...\n");
858 	send_status = safe_apic_wait_icr_idle();
859 
860 	udelay(init_udelay);
861 
862 	pr_debug("Deasserting INIT\n");
863 
864 	/* Target chip */
865 	/* Send IPI */
866 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
867 
868 	pr_debug("Waiting for send to finish...\n");
869 	send_status = safe_apic_wait_icr_idle();
870 
871 	mb();
872 
873 	/*
874 	 * Should we send STARTUP IPIs ?
875 	 *
876 	 * Determine this based on the APIC version.
877 	 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
878 	 */
879 	if (APIC_INTEGRATED(boot_cpu_apic_version))
880 		num_starts = 2;
881 	else
882 		num_starts = 0;
883 
884 	/*
885 	 * Run STARTUP IPI loop.
886 	 */
887 	pr_debug("#startup loops: %d\n", num_starts);
888 
889 	for (j = 1; j <= num_starts; j++) {
890 		pr_debug("Sending STARTUP #%d\n", j);
891 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
892 			apic_write(APIC_ESR, 0);
893 		apic_read(APIC_ESR);
894 		pr_debug("After apic_write\n");
895 
896 		/*
897 		 * STARTUP IPI
898 		 */
899 
900 		/* Target chip */
901 		/* Boot on the stack */
902 		/* Kick the second */
903 		apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
904 			       phys_apicid);
905 
906 		/*
907 		 * Give the other CPU some time to accept the IPI.
908 		 */
909 		if (init_udelay == 0)
910 			udelay(10);
911 		else
912 			udelay(300);
913 
914 		pr_debug("Startup point 1\n");
915 
916 		pr_debug("Waiting for send to finish...\n");
917 		send_status = safe_apic_wait_icr_idle();
918 
919 		/*
920 		 * Give the other CPU some time to accept the IPI.
921 		 */
922 		if (init_udelay == 0)
923 			udelay(10);
924 		else
925 			udelay(200);
926 
927 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
928 			apic_write(APIC_ESR, 0);
929 		accept_status = (apic_read(APIC_ESR) & 0xEF);
930 		if (send_status || accept_status)
931 			break;
932 	}
933 	pr_debug("After Startup\n");
934 
935 	if (send_status)
936 		pr_err("APIC never delivered???\n");
937 	if (accept_status)
938 		pr_err("APIC delivery error (%lx)\n", accept_status);
939 
940 	return (send_status | accept_status);
941 }
942 
943 /* reduce the number of lines printed when booting a large cpu count system */
944 static void announce_cpu(int cpu, int apicid)
945 {
946 	static int current_node = NUMA_NO_NODE;
947 	int node = early_cpu_to_node(cpu);
948 	static int width, node_width;
949 
950 	if (!width)
951 		width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
952 
953 	if (!node_width)
954 		node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
955 
956 	if (cpu == 1)
957 		printk(KERN_INFO "x86: Booting SMP configuration:\n");
958 
959 	if (system_state < SYSTEM_RUNNING) {
960 		if (node != current_node) {
961 			if (current_node > (-1))
962 				pr_cont("\n");
963 			current_node = node;
964 
965 			printk(KERN_INFO ".... node %*s#%d, CPUs:  ",
966 			       node_width - num_digits(node), " ", node);
967 		}
968 
969 		/* Add padding for the BSP */
970 		if (cpu == 1)
971 			pr_cont("%*s", width + 1, " ");
972 
973 		pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
974 
975 	} else
976 		pr_info("Booting Node %d Processor %d APIC 0x%x\n",
977 			node, cpu, apicid);
978 }
979 
980 static int wakeup_cpu0_nmi(unsigned int cmd, struct pt_regs *regs)
981 {
982 	int cpu;
983 
984 	cpu = smp_processor_id();
985 	if (cpu == 0 && !cpu_online(cpu) && enable_start_cpu0)
986 		return NMI_HANDLED;
987 
988 	return NMI_DONE;
989 }
990 
991 /*
992  * Wake up AP by INIT, INIT, STARTUP sequence.
993  *
994  * Instead of waiting for STARTUP after INITs, BSP will execute the BIOS
995  * boot-strap code which is not a desired behavior for waking up BSP. To
996  * void the boot-strap code, wake up CPU0 by NMI instead.
997  *
998  * This works to wake up soft offlined CPU0 only. If CPU0 is hard offlined
999  * (i.e. physically hot removed and then hot added), NMI won't wake it up.
1000  * We'll change this code in the future to wake up hard offlined CPU0 if
1001  * real platform and request are available.
1002  */
1003 static int
1004 wakeup_cpu_via_init_nmi(int cpu, unsigned long start_ip, int apicid,
1005 	       int *cpu0_nmi_registered)
1006 {
1007 	int id;
1008 	int boot_error;
1009 
1010 	preempt_disable();
1011 
1012 	/*
1013 	 * Wake up AP by INIT, INIT, STARTUP sequence.
1014 	 */
1015 	if (cpu) {
1016 		boot_error = wakeup_secondary_cpu_via_init(apicid, start_ip);
1017 		goto out;
1018 	}
1019 
1020 	/*
1021 	 * Wake up BSP by nmi.
1022 	 *
1023 	 * Register a NMI handler to help wake up CPU0.
1024 	 */
1025 	boot_error = register_nmi_handler(NMI_LOCAL,
1026 					  wakeup_cpu0_nmi, 0, "wake_cpu0");
1027 
1028 	if (!boot_error) {
1029 		enable_start_cpu0 = 1;
1030 		*cpu0_nmi_registered = 1;
1031 		id = apic->dest_mode_logical ? cpu0_logical_apicid : apicid;
1032 		boot_error = wakeup_secondary_cpu_via_nmi(id, start_ip);
1033 	}
1034 
1035 out:
1036 	preempt_enable();
1037 
1038 	return boot_error;
1039 }
1040 
1041 int common_cpu_up(unsigned int cpu, struct task_struct *idle)
1042 {
1043 	int ret;
1044 
1045 	/* Just in case we booted with a single CPU. */
1046 	alternatives_enable_smp();
1047 
1048 	per_cpu(current_task, cpu) = idle;
1049 	cpu_init_stack_canary(cpu, idle);
1050 
1051 	/* Initialize the interrupt stack(s) */
1052 	ret = irq_init_percpu_irqstack(cpu);
1053 	if (ret)
1054 		return ret;
1055 
1056 #ifdef CONFIG_X86_32
1057 	/* Stack for startup_32 can be just as for start_secondary onwards */
1058 	per_cpu(cpu_current_top_of_stack, cpu) = task_top_of_stack(idle);
1059 #else
1060 	initial_gs = per_cpu_offset(cpu);
1061 #endif
1062 	return 0;
1063 }
1064 
1065 /*
1066  * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
1067  * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
1068  * Returns zero if CPU booted OK, else error code from
1069  * ->wakeup_secondary_cpu.
1070  */
1071 static int do_boot_cpu(int apicid, int cpu, struct task_struct *idle,
1072 		       int *cpu0_nmi_registered)
1073 {
1074 	/* start_ip had better be page-aligned! */
1075 	unsigned long start_ip = real_mode_header->trampoline_start;
1076 
1077 	unsigned long boot_error = 0;
1078 	unsigned long timeout;
1079 
1080 	idle->thread.sp = (unsigned long)task_pt_regs(idle);
1081 	early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
1082 	initial_code = (unsigned long)start_secondary;
1083 	initial_stack  = idle->thread.sp;
1084 
1085 	/* Enable the espfix hack for this CPU */
1086 	init_espfix_ap(cpu);
1087 
1088 	/* So we see what's up */
1089 	announce_cpu(cpu, apicid);
1090 
1091 	/*
1092 	 * This grunge runs the startup process for
1093 	 * the targeted processor.
1094 	 */
1095 
1096 	if (x86_platform.legacy.warm_reset) {
1097 
1098 		pr_debug("Setting warm reset code and vector.\n");
1099 
1100 		smpboot_setup_warm_reset_vector(start_ip);
1101 		/*
1102 		 * Be paranoid about clearing APIC errors.
1103 		*/
1104 		if (APIC_INTEGRATED(boot_cpu_apic_version)) {
1105 			apic_write(APIC_ESR, 0);
1106 			apic_read(APIC_ESR);
1107 		}
1108 	}
1109 
1110 	/*
1111 	 * AP might wait on cpu_callout_mask in cpu_init() with
1112 	 * cpu_initialized_mask set if previous attempt to online
1113 	 * it timed-out. Clear cpu_initialized_mask so that after
1114 	 * INIT/SIPI it could start with a clean state.
1115 	 */
1116 	cpumask_clear_cpu(cpu, cpu_initialized_mask);
1117 	smp_mb();
1118 
1119 	/*
1120 	 * Wake up a CPU in difference cases:
1121 	 * - Use the method in the APIC driver if it's defined
1122 	 * Otherwise,
1123 	 * - Use an INIT boot APIC message for APs or NMI for BSP.
1124 	 */
1125 	if (apic->wakeup_secondary_cpu)
1126 		boot_error = apic->wakeup_secondary_cpu(apicid, start_ip);
1127 	else
1128 		boot_error = wakeup_cpu_via_init_nmi(cpu, start_ip, apicid,
1129 						     cpu0_nmi_registered);
1130 
1131 	if (!boot_error) {
1132 		/*
1133 		 * Wait 10s total for first sign of life from AP
1134 		 */
1135 		boot_error = -1;
1136 		timeout = jiffies + 10*HZ;
1137 		while (time_before(jiffies, timeout)) {
1138 			if (cpumask_test_cpu(cpu, cpu_initialized_mask)) {
1139 				/*
1140 				 * Tell AP to proceed with initialization
1141 				 */
1142 				cpumask_set_cpu(cpu, cpu_callout_mask);
1143 				boot_error = 0;
1144 				break;
1145 			}
1146 			schedule();
1147 		}
1148 	}
1149 
1150 	if (!boot_error) {
1151 		/*
1152 		 * Wait till AP completes initial initialization
1153 		 */
1154 		while (!cpumask_test_cpu(cpu, cpu_callin_mask)) {
1155 			/*
1156 			 * Allow other tasks to run while we wait for the
1157 			 * AP to come online. This also gives a chance
1158 			 * for the MTRR work(triggered by the AP coming online)
1159 			 * to be completed in the stop machine context.
1160 			 */
1161 			schedule();
1162 		}
1163 	}
1164 
1165 	if (x86_platform.legacy.warm_reset) {
1166 		/*
1167 		 * Cleanup possible dangling ends...
1168 		 */
1169 		smpboot_restore_warm_reset_vector();
1170 	}
1171 
1172 	return boot_error;
1173 }
1174 
1175 int native_cpu_up(unsigned int cpu, struct task_struct *tidle)
1176 {
1177 	int apicid = apic->cpu_present_to_apicid(cpu);
1178 	int cpu0_nmi_registered = 0;
1179 	unsigned long flags;
1180 	int err, ret = 0;
1181 
1182 	lockdep_assert_irqs_enabled();
1183 
1184 	pr_debug("++++++++++++++++++++=_---CPU UP  %u\n", cpu);
1185 
1186 	if (apicid == BAD_APICID ||
1187 	    !physid_isset(apicid, phys_cpu_present_map) ||
1188 	    !apic->apic_id_valid(apicid)) {
1189 		pr_err("%s: bad cpu %d\n", __func__, cpu);
1190 		return -EINVAL;
1191 	}
1192 
1193 	/*
1194 	 * Already booted CPU?
1195 	 */
1196 	if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
1197 		pr_debug("do_boot_cpu %d Already started\n", cpu);
1198 		return -ENOSYS;
1199 	}
1200 
1201 	/*
1202 	 * Save current MTRR state in case it was changed since early boot
1203 	 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1204 	 */
1205 	mtrr_save_state();
1206 
1207 	/* x86 CPUs take themselves offline, so delayed offline is OK. */
1208 	err = cpu_check_up_prepare(cpu);
1209 	if (err && err != -EBUSY)
1210 		return err;
1211 
1212 	/* the FPU context is blank, nobody can own it */
1213 	per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
1214 
1215 	err = common_cpu_up(cpu, tidle);
1216 	if (err)
1217 		return err;
1218 
1219 	err = do_boot_cpu(apicid, cpu, tidle, &cpu0_nmi_registered);
1220 	if (err) {
1221 		pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
1222 		ret = -EIO;
1223 		goto unreg_nmi;
1224 	}
1225 
1226 	/*
1227 	 * Check TSC synchronization with the AP (keep irqs disabled
1228 	 * while doing so):
1229 	 */
1230 	local_irq_save(flags);
1231 	check_tsc_sync_source(cpu);
1232 	local_irq_restore(flags);
1233 
1234 	while (!cpu_online(cpu)) {
1235 		cpu_relax();
1236 		touch_nmi_watchdog();
1237 	}
1238 
1239 unreg_nmi:
1240 	/*
1241 	 * Clean up the nmi handler. Do this after the callin and callout sync
1242 	 * to avoid impact of possible long unregister time.
1243 	 */
1244 	if (cpu0_nmi_registered)
1245 		unregister_nmi_handler(NMI_LOCAL, "wake_cpu0");
1246 
1247 	return ret;
1248 }
1249 
1250 /**
1251  * arch_disable_smp_support() - disables SMP support for x86 at runtime
1252  */
1253 void arch_disable_smp_support(void)
1254 {
1255 	disable_ioapic_support();
1256 }
1257 
1258 /*
1259  * Fall back to non SMP mode after errors.
1260  *
1261  * RED-PEN audit/test this more. I bet there is more state messed up here.
1262  */
1263 static __init void disable_smp(void)
1264 {
1265 	pr_info("SMP disabled\n");
1266 
1267 	disable_ioapic_support();
1268 
1269 	init_cpu_present(cpumask_of(0));
1270 	init_cpu_possible(cpumask_of(0));
1271 
1272 	if (smp_found_config)
1273 		physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
1274 	else
1275 		physid_set_mask_of_physid(0, &phys_cpu_present_map);
1276 	cpumask_set_cpu(0, topology_sibling_cpumask(0));
1277 	cpumask_set_cpu(0, topology_core_cpumask(0));
1278 	cpumask_set_cpu(0, topology_die_cpumask(0));
1279 }
1280 
1281 /*
1282  * Various sanity checks.
1283  */
1284 static void __init smp_sanity_check(void)
1285 {
1286 	preempt_disable();
1287 
1288 #if !defined(CONFIG_X86_BIGSMP) && defined(CONFIG_X86_32)
1289 	if (def_to_bigsmp && nr_cpu_ids > 8) {
1290 		unsigned int cpu;
1291 		unsigned nr;
1292 
1293 		pr_warn("More than 8 CPUs detected - skipping them\n"
1294 			"Use CONFIG_X86_BIGSMP\n");
1295 
1296 		nr = 0;
1297 		for_each_present_cpu(cpu) {
1298 			if (nr >= 8)
1299 				set_cpu_present(cpu, false);
1300 			nr++;
1301 		}
1302 
1303 		nr = 0;
1304 		for_each_possible_cpu(cpu) {
1305 			if (nr >= 8)
1306 				set_cpu_possible(cpu, false);
1307 			nr++;
1308 		}
1309 
1310 		nr_cpu_ids = 8;
1311 	}
1312 #endif
1313 
1314 	if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map)) {
1315 		pr_warn("weird, boot CPU (#%d) not listed by the BIOS\n",
1316 			hard_smp_processor_id());
1317 
1318 		physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1319 	}
1320 
1321 	/*
1322 	 * Should not be necessary because the MP table should list the boot
1323 	 * CPU too, but we do it for the sake of robustness anyway.
1324 	 */
1325 	if (!apic->check_phys_apicid_present(boot_cpu_physical_apicid)) {
1326 		pr_notice("weird, boot CPU (#%d) not listed by the BIOS\n",
1327 			  boot_cpu_physical_apicid);
1328 		physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1329 	}
1330 	preempt_enable();
1331 }
1332 
1333 static void __init smp_cpu_index_default(void)
1334 {
1335 	int i;
1336 	struct cpuinfo_x86 *c;
1337 
1338 	for_each_possible_cpu(i) {
1339 		c = &cpu_data(i);
1340 		/* mark all to hotplug */
1341 		c->cpu_index = nr_cpu_ids;
1342 	}
1343 }
1344 
1345 static void __init smp_get_logical_apicid(void)
1346 {
1347 	if (x2apic_mode)
1348 		cpu0_logical_apicid = apic_read(APIC_LDR);
1349 	else
1350 		cpu0_logical_apicid = GET_APIC_LOGICAL_ID(apic_read(APIC_LDR));
1351 }
1352 
1353 void __init smp_prepare_cpus_common(void)
1354 {
1355 	unsigned int i;
1356 
1357 	smp_cpu_index_default();
1358 
1359 	/*
1360 	 * Setup boot CPU information
1361 	 */
1362 	smp_store_boot_cpu_info(); /* Final full version of the data */
1363 	cpumask_copy(cpu_callin_mask, cpumask_of(0));
1364 	mb();
1365 
1366 	for_each_possible_cpu(i) {
1367 		zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1368 		zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1369 		zalloc_cpumask_var(&per_cpu(cpu_die_map, i), GFP_KERNEL);
1370 		zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1371 		zalloc_cpumask_var(&per_cpu(cpu_l2c_shared_map, i), GFP_KERNEL);
1372 	}
1373 
1374 	/*
1375 	 * Set 'default' x86 topology, this matches default_topology() in that
1376 	 * it has NUMA nodes as a topology level. See also
1377 	 * native_smp_cpus_done().
1378 	 *
1379 	 * Must be done before set_cpus_sibling_map() is ran.
1380 	 */
1381 	set_sched_topology(x86_topology);
1382 
1383 	set_cpu_sibling_map(0);
1384 }
1385 
1386 /*
1387  * Prepare for SMP bootup.
1388  * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1389  *            for common interface support.
1390  */
1391 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1392 {
1393 	smp_prepare_cpus_common();
1394 
1395 	init_freq_invariance(false, false);
1396 	smp_sanity_check();
1397 
1398 	switch (apic_intr_mode) {
1399 	case APIC_PIC:
1400 	case APIC_VIRTUAL_WIRE_NO_CONFIG:
1401 		disable_smp();
1402 		return;
1403 	case APIC_SYMMETRIC_IO_NO_ROUTING:
1404 		disable_smp();
1405 		/* Setup local timer */
1406 		x86_init.timers.setup_percpu_clockev();
1407 		return;
1408 	case APIC_VIRTUAL_WIRE:
1409 	case APIC_SYMMETRIC_IO:
1410 		break;
1411 	}
1412 
1413 	/* Setup local timer */
1414 	x86_init.timers.setup_percpu_clockev();
1415 
1416 	smp_get_logical_apicid();
1417 
1418 	pr_info("CPU0: ");
1419 	print_cpu_info(&cpu_data(0));
1420 
1421 	uv_system_init();
1422 
1423 	set_mtrr_aps_delayed_init();
1424 
1425 	smp_quirk_init_udelay();
1426 
1427 	speculative_store_bypass_ht_init();
1428 }
1429 
1430 void arch_thaw_secondary_cpus_begin(void)
1431 {
1432 	set_mtrr_aps_delayed_init();
1433 }
1434 
1435 void arch_thaw_secondary_cpus_end(void)
1436 {
1437 	mtrr_aps_init();
1438 }
1439 
1440 /*
1441  * Early setup to make printk work.
1442  */
1443 void __init native_smp_prepare_boot_cpu(void)
1444 {
1445 	int me = smp_processor_id();
1446 	switch_to_new_gdt(me);
1447 	/* already set me in cpu_online_mask in boot_cpu_init() */
1448 	cpumask_set_cpu(me, cpu_callout_mask);
1449 	cpu_set_state_online(me);
1450 	native_pv_lock_init();
1451 }
1452 
1453 void __init calculate_max_logical_packages(void)
1454 {
1455 	int ncpus;
1456 
1457 	/*
1458 	 * Today neither Intel nor AMD support heterogeneous systems so
1459 	 * extrapolate the boot cpu's data to all packages.
1460 	 */
1461 	ncpus = cpu_data(0).booted_cores * topology_max_smt_threads();
1462 	__max_logical_packages = DIV_ROUND_UP(total_cpus, ncpus);
1463 	pr_info("Max logical packages: %u\n", __max_logical_packages);
1464 }
1465 
1466 void __init native_smp_cpus_done(unsigned int max_cpus)
1467 {
1468 	pr_debug("Boot done\n");
1469 
1470 	calculate_max_logical_packages();
1471 
1472 	if (x86_has_numa_in_package)
1473 		set_sched_topology(x86_numa_in_package_topology);
1474 
1475 	nmi_selftest();
1476 	impress_friends();
1477 	mtrr_aps_init();
1478 }
1479 
1480 static int __initdata setup_possible_cpus = -1;
1481 static int __init _setup_possible_cpus(char *str)
1482 {
1483 	get_option(&str, &setup_possible_cpus);
1484 	return 0;
1485 }
1486 early_param("possible_cpus", _setup_possible_cpus);
1487 
1488 
1489 /*
1490  * cpu_possible_mask should be static, it cannot change as cpu's
1491  * are onlined, or offlined. The reason is per-cpu data-structures
1492  * are allocated by some modules at init time, and don't expect to
1493  * do this dynamically on cpu arrival/departure.
1494  * cpu_present_mask on the other hand can change dynamically.
1495  * In case when cpu_hotplug is not compiled, then we resort to current
1496  * behaviour, which is cpu_possible == cpu_present.
1497  * - Ashok Raj
1498  *
1499  * Three ways to find out the number of additional hotplug CPUs:
1500  * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1501  * - The user can overwrite it with possible_cpus=NUM
1502  * - Otherwise don't reserve additional CPUs.
1503  * We do this because additional CPUs waste a lot of memory.
1504  * -AK
1505  */
1506 __init void prefill_possible_map(void)
1507 {
1508 	int i, possible;
1509 
1510 	/* No boot processor was found in mptable or ACPI MADT */
1511 	if (!num_processors) {
1512 		if (boot_cpu_has(X86_FEATURE_APIC)) {
1513 			int apicid = boot_cpu_physical_apicid;
1514 			int cpu = hard_smp_processor_id();
1515 
1516 			pr_warn("Boot CPU (id %d) not listed by BIOS\n", cpu);
1517 
1518 			/* Make sure boot cpu is enumerated */
1519 			if (apic->cpu_present_to_apicid(0) == BAD_APICID &&
1520 			    apic->apic_id_valid(apicid))
1521 				generic_processor_info(apicid, boot_cpu_apic_version);
1522 		}
1523 
1524 		if (!num_processors)
1525 			num_processors = 1;
1526 	}
1527 
1528 	i = setup_max_cpus ?: 1;
1529 	if (setup_possible_cpus == -1) {
1530 		possible = num_processors;
1531 #ifdef CONFIG_HOTPLUG_CPU
1532 		if (setup_max_cpus)
1533 			possible += disabled_cpus;
1534 #else
1535 		if (possible > i)
1536 			possible = i;
1537 #endif
1538 	} else
1539 		possible = setup_possible_cpus;
1540 
1541 	total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1542 
1543 	/* nr_cpu_ids could be reduced via nr_cpus= */
1544 	if (possible > nr_cpu_ids) {
1545 		pr_warn("%d Processors exceeds NR_CPUS limit of %u\n",
1546 			possible, nr_cpu_ids);
1547 		possible = nr_cpu_ids;
1548 	}
1549 
1550 #ifdef CONFIG_HOTPLUG_CPU
1551 	if (!setup_max_cpus)
1552 #endif
1553 	if (possible > i) {
1554 		pr_warn("%d Processors exceeds max_cpus limit of %u\n",
1555 			possible, setup_max_cpus);
1556 		possible = i;
1557 	}
1558 
1559 	nr_cpu_ids = possible;
1560 
1561 	pr_info("Allowing %d CPUs, %d hotplug CPUs\n",
1562 		possible, max_t(int, possible - num_processors, 0));
1563 
1564 	reset_cpu_possible_mask();
1565 
1566 	for (i = 0; i < possible; i++)
1567 		set_cpu_possible(i, true);
1568 }
1569 
1570 #ifdef CONFIG_HOTPLUG_CPU
1571 
1572 /* Recompute SMT state for all CPUs on offline */
1573 static void recompute_smt_state(void)
1574 {
1575 	int max_threads, cpu;
1576 
1577 	max_threads = 0;
1578 	for_each_online_cpu (cpu) {
1579 		int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1580 
1581 		if (threads > max_threads)
1582 			max_threads = threads;
1583 	}
1584 	__max_smt_threads = max_threads;
1585 }
1586 
1587 static void remove_siblinginfo(int cpu)
1588 {
1589 	int sibling;
1590 	struct cpuinfo_x86 *c = &cpu_data(cpu);
1591 
1592 	for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1593 		cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1594 		/*/
1595 		 * last thread sibling in this cpu core going down
1596 		 */
1597 		if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1598 			cpu_data(sibling).booted_cores--;
1599 	}
1600 
1601 	for_each_cpu(sibling, topology_die_cpumask(cpu))
1602 		cpumask_clear_cpu(cpu, topology_die_cpumask(sibling));
1603 
1604 	for_each_cpu(sibling, topology_sibling_cpumask(cpu)) {
1605 		cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1606 		if (cpumask_weight(topology_sibling_cpumask(sibling)) == 1)
1607 			cpu_data(sibling).smt_active = false;
1608 	}
1609 
1610 	for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1611 		cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1612 	for_each_cpu(sibling, cpu_l2c_shared_mask(cpu))
1613 		cpumask_clear_cpu(cpu, cpu_l2c_shared_mask(sibling));
1614 	cpumask_clear(cpu_llc_shared_mask(cpu));
1615 	cpumask_clear(cpu_l2c_shared_mask(cpu));
1616 	cpumask_clear(topology_sibling_cpumask(cpu));
1617 	cpumask_clear(topology_core_cpumask(cpu));
1618 	cpumask_clear(topology_die_cpumask(cpu));
1619 	c->cpu_core_id = 0;
1620 	c->booted_cores = 0;
1621 	cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1622 	recompute_smt_state();
1623 }
1624 
1625 static void remove_cpu_from_maps(int cpu)
1626 {
1627 	set_cpu_online(cpu, false);
1628 	cpumask_clear_cpu(cpu, cpu_callout_mask);
1629 	cpumask_clear_cpu(cpu, cpu_callin_mask);
1630 	/* was set by cpu_init() */
1631 	cpumask_clear_cpu(cpu, cpu_initialized_mask);
1632 	numa_remove_cpu(cpu);
1633 }
1634 
1635 void cpu_disable_common(void)
1636 {
1637 	int cpu = smp_processor_id();
1638 
1639 	remove_siblinginfo(cpu);
1640 
1641 	/* It's now safe to remove this processor from the online map */
1642 	lock_vector_lock();
1643 	remove_cpu_from_maps(cpu);
1644 	unlock_vector_lock();
1645 	fixup_irqs();
1646 	lapic_offline();
1647 }
1648 
1649 int native_cpu_disable(void)
1650 {
1651 	int ret;
1652 
1653 	ret = lapic_can_unplug_cpu();
1654 	if (ret)
1655 		return ret;
1656 
1657 	cpu_disable_common();
1658 
1659         /*
1660          * Disable the local APIC. Otherwise IPI broadcasts will reach
1661          * it. It still responds normally to INIT, NMI, SMI, and SIPI
1662          * messages.
1663          *
1664          * Disabling the APIC must happen after cpu_disable_common()
1665          * which invokes fixup_irqs().
1666          *
1667          * Disabling the APIC preserves already set bits in IRR, but
1668          * an interrupt arriving after disabling the local APIC does not
1669          * set the corresponding IRR bit.
1670          *
1671          * fixup_irqs() scans IRR for set bits so it can raise a not
1672          * yet handled interrupt on the new destination CPU via an IPI
1673          * but obviously it can't do so for IRR bits which are not set.
1674          * IOW, interrupts arriving after disabling the local APIC will
1675          * be lost.
1676          */
1677 	apic_soft_disable();
1678 
1679 	return 0;
1680 }
1681 
1682 int common_cpu_die(unsigned int cpu)
1683 {
1684 	int ret = 0;
1685 
1686 	/* We don't do anything here: idle task is faking death itself. */
1687 
1688 	/* They ack this in play_dead() by setting CPU_DEAD */
1689 	if (cpu_wait_death(cpu, 5)) {
1690 		if (system_state == SYSTEM_RUNNING)
1691 			pr_info("CPU %u is now offline\n", cpu);
1692 	} else {
1693 		pr_err("CPU %u didn't die...\n", cpu);
1694 		ret = -1;
1695 	}
1696 
1697 	return ret;
1698 }
1699 
1700 void native_cpu_die(unsigned int cpu)
1701 {
1702 	common_cpu_die(cpu);
1703 }
1704 
1705 void play_dead_common(void)
1706 {
1707 	idle_task_exit();
1708 
1709 	/* Ack it */
1710 	(void)cpu_report_death();
1711 
1712 	/*
1713 	 * With physical CPU hotplug, we should halt the cpu
1714 	 */
1715 	local_irq_disable();
1716 }
1717 
1718 /**
1719  * cond_wakeup_cpu0 - Wake up CPU0 if needed.
1720  *
1721  * If NMI wants to wake up CPU0, start CPU0.
1722  */
1723 void cond_wakeup_cpu0(void)
1724 {
1725 	if (smp_processor_id() == 0 && enable_start_cpu0)
1726 		start_cpu0();
1727 }
1728 EXPORT_SYMBOL_GPL(cond_wakeup_cpu0);
1729 
1730 /*
1731  * We need to flush the caches before going to sleep, lest we have
1732  * dirty data in our caches when we come back up.
1733  */
1734 static inline void mwait_play_dead(void)
1735 {
1736 	unsigned int eax, ebx, ecx, edx;
1737 	unsigned int highest_cstate = 0;
1738 	unsigned int highest_subcstate = 0;
1739 	void *mwait_ptr;
1740 	int i;
1741 
1742 	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD ||
1743 	    boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
1744 		return;
1745 	if (!this_cpu_has(X86_FEATURE_MWAIT))
1746 		return;
1747 	if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1748 		return;
1749 	if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1750 		return;
1751 
1752 	eax = CPUID_MWAIT_LEAF;
1753 	ecx = 0;
1754 	native_cpuid(&eax, &ebx, &ecx, &edx);
1755 
1756 	/*
1757 	 * eax will be 0 if EDX enumeration is not valid.
1758 	 * Initialized below to cstate, sub_cstate value when EDX is valid.
1759 	 */
1760 	if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1761 		eax = 0;
1762 	} else {
1763 		edx >>= MWAIT_SUBSTATE_SIZE;
1764 		for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1765 			if (edx & MWAIT_SUBSTATE_MASK) {
1766 				highest_cstate = i;
1767 				highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1768 			}
1769 		}
1770 		eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1771 			(highest_subcstate - 1);
1772 	}
1773 
1774 	/*
1775 	 * This should be a memory location in a cache line which is
1776 	 * unlikely to be touched by other processors.  The actual
1777 	 * content is immaterial as it is not actually modified in any way.
1778 	 */
1779 	mwait_ptr = &current_thread_info()->flags;
1780 
1781 	wbinvd();
1782 
1783 	while (1) {
1784 		/*
1785 		 * The CLFLUSH is a workaround for erratum AAI65 for
1786 		 * the Xeon 7400 series.  It's not clear it is actually
1787 		 * needed, but it should be harmless in either case.
1788 		 * The WBINVD is insufficient due to the spurious-wakeup
1789 		 * case where we return around the loop.
1790 		 */
1791 		mb();
1792 		clflush(mwait_ptr);
1793 		mb();
1794 		__monitor(mwait_ptr, 0, 0);
1795 		mb();
1796 		__mwait(eax, 0);
1797 
1798 		cond_wakeup_cpu0();
1799 	}
1800 }
1801 
1802 void hlt_play_dead(void)
1803 {
1804 	if (__this_cpu_read(cpu_info.x86) >= 4)
1805 		wbinvd();
1806 
1807 	while (1) {
1808 		native_halt();
1809 
1810 		cond_wakeup_cpu0();
1811 	}
1812 }
1813 
1814 void native_play_dead(void)
1815 {
1816 	play_dead_common();
1817 	tboot_shutdown(TB_SHUTDOWN_WFS);
1818 
1819 	mwait_play_dead();	/* Only returns on failure */
1820 	if (cpuidle_play_dead())
1821 		hlt_play_dead();
1822 }
1823 
1824 #else /* ... !CONFIG_HOTPLUG_CPU */
1825 int native_cpu_disable(void)
1826 {
1827 	return -ENOSYS;
1828 }
1829 
1830 void native_cpu_die(unsigned int cpu)
1831 {
1832 	/* We said "no" in __cpu_disable */
1833 	BUG();
1834 }
1835 
1836 void native_play_dead(void)
1837 {
1838 	BUG();
1839 }
1840 
1841 #endif
1842 
1843 #ifdef CONFIG_X86_64
1844 /*
1845  * APERF/MPERF frequency ratio computation.
1846  *
1847  * The scheduler wants to do frequency invariant accounting and needs a <1
1848  * ratio to account for the 'current' frequency, corresponding to
1849  * freq_curr / freq_max.
1850  *
1851  * Since the frequency freq_curr on x86 is controlled by micro-controller and
1852  * our P-state setting is little more than a request/hint, we need to observe
1853  * the effective frequency 'BusyMHz', i.e. the average frequency over a time
1854  * interval after discarding idle time. This is given by:
1855  *
1856  *   BusyMHz = delta_APERF / delta_MPERF * freq_base
1857  *
1858  * where freq_base is the max non-turbo P-state.
1859  *
1860  * The freq_max term has to be set to a somewhat arbitrary value, because we
1861  * can't know which turbo states will be available at a given point in time:
1862  * it all depends on the thermal headroom of the entire package. We set it to
1863  * the turbo level with 4 cores active.
1864  *
1865  * Benchmarks show that's a good compromise between the 1C turbo ratio
1866  * (freq_curr/freq_max would rarely reach 1) and something close to freq_base,
1867  * which would ignore the entire turbo range (a conspicuous part, making
1868  * freq_curr/freq_max always maxed out).
1869  *
1870  * An exception to the heuristic above is the Atom uarch, where we choose the
1871  * highest turbo level for freq_max since Atom's are generally oriented towards
1872  * power efficiency.
1873  *
1874  * Setting freq_max to anything less than the 1C turbo ratio makes the ratio
1875  * freq_curr / freq_max to eventually grow >1, in which case we clip it to 1.
1876  */
1877 
1878 DEFINE_STATIC_KEY_FALSE(arch_scale_freq_key);
1879 
1880 static DEFINE_PER_CPU(u64, arch_prev_aperf);
1881 static DEFINE_PER_CPU(u64, arch_prev_mperf);
1882 static u64 arch_turbo_freq_ratio = SCHED_CAPACITY_SCALE;
1883 static u64 arch_max_freq_ratio = SCHED_CAPACITY_SCALE;
1884 
1885 void arch_set_max_freq_ratio(bool turbo_disabled)
1886 {
1887 	arch_max_freq_ratio = turbo_disabled ? SCHED_CAPACITY_SCALE :
1888 					arch_turbo_freq_ratio;
1889 }
1890 EXPORT_SYMBOL_GPL(arch_set_max_freq_ratio);
1891 
1892 static bool turbo_disabled(void)
1893 {
1894 	u64 misc_en;
1895 	int err;
1896 
1897 	err = rdmsrl_safe(MSR_IA32_MISC_ENABLE, &misc_en);
1898 	if (err)
1899 		return false;
1900 
1901 	return (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
1902 }
1903 
1904 static bool slv_set_max_freq_ratio(u64 *base_freq, u64 *turbo_freq)
1905 {
1906 	int err;
1907 
1908 	err = rdmsrl_safe(MSR_ATOM_CORE_RATIOS, base_freq);
1909 	if (err)
1910 		return false;
1911 
1912 	err = rdmsrl_safe(MSR_ATOM_CORE_TURBO_RATIOS, turbo_freq);
1913 	if (err)
1914 		return false;
1915 
1916 	*base_freq = (*base_freq >> 16) & 0x3F;     /* max P state */
1917 	*turbo_freq = *turbo_freq & 0x3F;           /* 1C turbo    */
1918 
1919 	return true;
1920 }
1921 
1922 #define X86_MATCH(model)					\
1923 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6,		\
1924 		INTEL_FAM6_##model, X86_FEATURE_APERFMPERF, NULL)
1925 
1926 static const struct x86_cpu_id has_knl_turbo_ratio_limits[] = {
1927 	X86_MATCH(XEON_PHI_KNL),
1928 	X86_MATCH(XEON_PHI_KNM),
1929 	{}
1930 };
1931 
1932 static const struct x86_cpu_id has_skx_turbo_ratio_limits[] = {
1933 	X86_MATCH(SKYLAKE_X),
1934 	{}
1935 };
1936 
1937 static const struct x86_cpu_id has_glm_turbo_ratio_limits[] = {
1938 	X86_MATCH(ATOM_GOLDMONT),
1939 	X86_MATCH(ATOM_GOLDMONT_D),
1940 	X86_MATCH(ATOM_GOLDMONT_PLUS),
1941 	{}
1942 };
1943 
1944 static bool knl_set_max_freq_ratio(u64 *base_freq, u64 *turbo_freq,
1945 				int num_delta_fratio)
1946 {
1947 	int fratio, delta_fratio, found;
1948 	int err, i;
1949 	u64 msr;
1950 
1951 	err = rdmsrl_safe(MSR_PLATFORM_INFO, base_freq);
1952 	if (err)
1953 		return false;
1954 
1955 	*base_freq = (*base_freq >> 8) & 0xFF;	    /* max P state */
1956 
1957 	err = rdmsrl_safe(MSR_TURBO_RATIO_LIMIT, &msr);
1958 	if (err)
1959 		return false;
1960 
1961 	fratio = (msr >> 8) & 0xFF;
1962 	i = 16;
1963 	found = 0;
1964 	do {
1965 		if (found >= num_delta_fratio) {
1966 			*turbo_freq = fratio;
1967 			return true;
1968 		}
1969 
1970 		delta_fratio = (msr >> (i + 5)) & 0x7;
1971 
1972 		if (delta_fratio) {
1973 			found += 1;
1974 			fratio -= delta_fratio;
1975 		}
1976 
1977 		i += 8;
1978 	} while (i < 64);
1979 
1980 	return true;
1981 }
1982 
1983 static bool skx_set_max_freq_ratio(u64 *base_freq, u64 *turbo_freq, int size)
1984 {
1985 	u64 ratios, counts;
1986 	u32 group_size;
1987 	int err, i;
1988 
1989 	err = rdmsrl_safe(MSR_PLATFORM_INFO, base_freq);
1990 	if (err)
1991 		return false;
1992 
1993 	*base_freq = (*base_freq >> 8) & 0xFF;      /* max P state */
1994 
1995 	err = rdmsrl_safe(MSR_TURBO_RATIO_LIMIT, &ratios);
1996 	if (err)
1997 		return false;
1998 
1999 	err = rdmsrl_safe(MSR_TURBO_RATIO_LIMIT1, &counts);
2000 	if (err)
2001 		return false;
2002 
2003 	for (i = 0; i < 64; i += 8) {
2004 		group_size = (counts >> i) & 0xFF;
2005 		if (group_size >= size) {
2006 			*turbo_freq = (ratios >> i) & 0xFF;
2007 			return true;
2008 		}
2009 	}
2010 
2011 	return false;
2012 }
2013 
2014 static bool core_set_max_freq_ratio(u64 *base_freq, u64 *turbo_freq)
2015 {
2016 	u64 msr;
2017 	int err;
2018 
2019 	err = rdmsrl_safe(MSR_PLATFORM_INFO, base_freq);
2020 	if (err)
2021 		return false;
2022 
2023 	err = rdmsrl_safe(MSR_TURBO_RATIO_LIMIT, &msr);
2024 	if (err)
2025 		return false;
2026 
2027 	*base_freq = (*base_freq >> 8) & 0xFF;    /* max P state */
2028 	*turbo_freq = (msr >> 24) & 0xFF;         /* 4C turbo    */
2029 
2030 	/* The CPU may have less than 4 cores */
2031 	if (!*turbo_freq)
2032 		*turbo_freq = msr & 0xFF;         /* 1C turbo    */
2033 
2034 	return true;
2035 }
2036 
2037 static bool intel_set_max_freq_ratio(void)
2038 {
2039 	u64 base_freq, turbo_freq;
2040 	u64 turbo_ratio;
2041 
2042 	if (slv_set_max_freq_ratio(&base_freq, &turbo_freq))
2043 		goto out;
2044 
2045 	if (x86_match_cpu(has_glm_turbo_ratio_limits) &&
2046 	    skx_set_max_freq_ratio(&base_freq, &turbo_freq, 1))
2047 		goto out;
2048 
2049 	if (x86_match_cpu(has_knl_turbo_ratio_limits) &&
2050 	    knl_set_max_freq_ratio(&base_freq, &turbo_freq, 1))
2051 		goto out;
2052 
2053 	if (x86_match_cpu(has_skx_turbo_ratio_limits) &&
2054 	    skx_set_max_freq_ratio(&base_freq, &turbo_freq, 4))
2055 		goto out;
2056 
2057 	if (core_set_max_freq_ratio(&base_freq, &turbo_freq))
2058 		goto out;
2059 
2060 	return false;
2061 
2062 out:
2063 	/*
2064 	 * Some hypervisors advertise X86_FEATURE_APERFMPERF
2065 	 * but then fill all MSR's with zeroes.
2066 	 * Some CPUs have turbo boost but don't declare any turbo ratio
2067 	 * in MSR_TURBO_RATIO_LIMIT.
2068 	 */
2069 	if (!base_freq || !turbo_freq) {
2070 		pr_debug("Couldn't determine cpu base or turbo frequency, necessary for scale-invariant accounting.\n");
2071 		return false;
2072 	}
2073 
2074 	turbo_ratio = div_u64(turbo_freq * SCHED_CAPACITY_SCALE, base_freq);
2075 	if (!turbo_ratio) {
2076 		pr_debug("Non-zero turbo and base frequencies led to a 0 ratio.\n");
2077 		return false;
2078 	}
2079 
2080 	arch_turbo_freq_ratio = turbo_ratio;
2081 	arch_set_max_freq_ratio(turbo_disabled());
2082 
2083 	return true;
2084 }
2085 
2086 #ifdef CONFIG_ACPI_CPPC_LIB
2087 static bool amd_set_max_freq_ratio(void)
2088 {
2089 	struct cppc_perf_caps perf_caps;
2090 	u64 highest_perf, nominal_perf;
2091 	u64 perf_ratio;
2092 	int rc;
2093 
2094 	rc = cppc_get_perf_caps(0, &perf_caps);
2095 	if (rc) {
2096 		pr_debug("Could not retrieve perf counters (%d)\n", rc);
2097 		return false;
2098 	}
2099 
2100 	highest_perf = amd_get_highest_perf();
2101 	nominal_perf = perf_caps.nominal_perf;
2102 
2103 	if (!highest_perf || !nominal_perf) {
2104 		pr_debug("Could not retrieve highest or nominal performance\n");
2105 		return false;
2106 	}
2107 
2108 	perf_ratio = div_u64(highest_perf * SCHED_CAPACITY_SCALE, nominal_perf);
2109 	/* midpoint between max_boost and max_P */
2110 	perf_ratio = (perf_ratio + SCHED_CAPACITY_SCALE) >> 1;
2111 	if (!perf_ratio) {
2112 		pr_debug("Non-zero highest/nominal perf values led to a 0 ratio\n");
2113 		return false;
2114 	}
2115 
2116 	arch_turbo_freq_ratio = perf_ratio;
2117 	arch_set_max_freq_ratio(false);
2118 
2119 	return true;
2120 }
2121 #else
2122 static bool amd_set_max_freq_ratio(void)
2123 {
2124 	return false;
2125 }
2126 #endif
2127 
2128 static void init_counter_refs(void)
2129 {
2130 	u64 aperf, mperf;
2131 
2132 	rdmsrl(MSR_IA32_APERF, aperf);
2133 	rdmsrl(MSR_IA32_MPERF, mperf);
2134 
2135 	this_cpu_write(arch_prev_aperf, aperf);
2136 	this_cpu_write(arch_prev_mperf, mperf);
2137 }
2138 
2139 #ifdef CONFIG_PM_SLEEP
2140 static struct syscore_ops freq_invariance_syscore_ops = {
2141 	.resume = init_counter_refs,
2142 };
2143 
2144 static void register_freq_invariance_syscore_ops(void)
2145 {
2146 	/* Bail out if registered already. */
2147 	if (freq_invariance_syscore_ops.node.prev)
2148 		return;
2149 
2150 	register_syscore_ops(&freq_invariance_syscore_ops);
2151 }
2152 #else
2153 static inline void register_freq_invariance_syscore_ops(void) {}
2154 #endif
2155 
2156 static void init_freq_invariance(bool secondary, bool cppc_ready)
2157 {
2158 	bool ret = false;
2159 
2160 	if (!boot_cpu_has(X86_FEATURE_APERFMPERF))
2161 		return;
2162 
2163 	if (secondary) {
2164 		if (static_branch_likely(&arch_scale_freq_key)) {
2165 			init_counter_refs();
2166 		}
2167 		return;
2168 	}
2169 
2170 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
2171 		ret = intel_set_max_freq_ratio();
2172 	else if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD) {
2173 		if (!cppc_ready) {
2174 			return;
2175 		}
2176 		ret = amd_set_max_freq_ratio();
2177 	}
2178 
2179 	if (ret) {
2180 		init_counter_refs();
2181 		static_branch_enable(&arch_scale_freq_key);
2182 		register_freq_invariance_syscore_ops();
2183 		pr_info("Estimated ratio of average max frequency by base frequency (times 1024): %llu\n", arch_max_freq_ratio);
2184 	} else {
2185 		pr_debug("Couldn't determine max cpu frequency, necessary for scale-invariant accounting.\n");
2186 	}
2187 }
2188 
2189 #ifdef CONFIG_ACPI_CPPC_LIB
2190 static DEFINE_MUTEX(freq_invariance_lock);
2191 
2192 void init_freq_invariance_cppc(void)
2193 {
2194 	static bool secondary;
2195 
2196 	mutex_lock(&freq_invariance_lock);
2197 
2198 	init_freq_invariance(secondary, true);
2199 	secondary = true;
2200 
2201 	mutex_unlock(&freq_invariance_lock);
2202 }
2203 #endif
2204 
2205 static void disable_freq_invariance_workfn(struct work_struct *work)
2206 {
2207 	static_branch_disable(&arch_scale_freq_key);
2208 }
2209 
2210 static DECLARE_WORK(disable_freq_invariance_work,
2211 		    disable_freq_invariance_workfn);
2212 
2213 DEFINE_PER_CPU(unsigned long, arch_freq_scale) = SCHED_CAPACITY_SCALE;
2214 
2215 void arch_scale_freq_tick(void)
2216 {
2217 	u64 freq_scale;
2218 	u64 aperf, mperf;
2219 	u64 acnt, mcnt;
2220 
2221 	if (!arch_scale_freq_invariant())
2222 		return;
2223 
2224 	rdmsrl(MSR_IA32_APERF, aperf);
2225 	rdmsrl(MSR_IA32_MPERF, mperf);
2226 
2227 	acnt = aperf - this_cpu_read(arch_prev_aperf);
2228 	mcnt = mperf - this_cpu_read(arch_prev_mperf);
2229 
2230 	this_cpu_write(arch_prev_aperf, aperf);
2231 	this_cpu_write(arch_prev_mperf, mperf);
2232 
2233 	if (check_shl_overflow(acnt, 2*SCHED_CAPACITY_SHIFT, &acnt))
2234 		goto error;
2235 
2236 	if (check_mul_overflow(mcnt, arch_max_freq_ratio, &mcnt) || !mcnt)
2237 		goto error;
2238 
2239 	freq_scale = div64_u64(acnt, mcnt);
2240 	if (!freq_scale)
2241 		goto error;
2242 
2243 	if (freq_scale > SCHED_CAPACITY_SCALE)
2244 		freq_scale = SCHED_CAPACITY_SCALE;
2245 
2246 	this_cpu_write(arch_freq_scale, freq_scale);
2247 	return;
2248 
2249 error:
2250 	pr_warn("Scheduler frequency invariance went wobbly, disabling!\n");
2251 	schedule_work(&disable_freq_invariance_work);
2252 }
2253 #else
2254 static inline void init_freq_invariance(bool secondary, bool cppc_ready)
2255 {
2256 }
2257 #endif /* CONFIG_X86_64 */
2258