xref: /openbmc/linux/arch/x86/kernel/smpboot.c (revision 2cf1c348)
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_hybrid_topology[] = {
583 #ifdef CONFIG_SCHED_SMT
584 	{ cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
585 #endif
586 #ifdef CONFIG_SCHED_MC
587 	{ cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
588 #endif
589 	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
590 	{ NULL, },
591 };
592 
593 static struct sched_domain_topology_level x86_topology[] = {
594 #ifdef CONFIG_SCHED_SMT
595 	{ cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
596 #endif
597 #ifdef CONFIG_SCHED_CLUSTER
598 	{ cpu_clustergroup_mask, x86_cluster_flags, SD_INIT_NAME(CLS) },
599 #endif
600 #ifdef CONFIG_SCHED_MC
601 	{ cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
602 #endif
603 	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
604 	{ NULL, },
605 };
606 
607 /*
608  * Set if a package/die has multiple NUMA nodes inside.
609  * AMD Magny-Cours, Intel Cluster-on-Die, and Intel
610  * Sub-NUMA Clustering have this.
611  */
612 static bool x86_has_numa_in_package;
613 
614 void set_cpu_sibling_map(int cpu)
615 {
616 	bool has_smt = smp_num_siblings > 1;
617 	bool has_mp = has_smt || boot_cpu_data.x86_max_cores > 1;
618 	struct cpuinfo_x86 *c = &cpu_data(cpu);
619 	struct cpuinfo_x86 *o;
620 	int i, threads;
621 
622 	cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
623 
624 	if (!has_mp) {
625 		cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
626 		cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
627 		cpumask_set_cpu(cpu, cpu_l2c_shared_mask(cpu));
628 		cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
629 		cpumask_set_cpu(cpu, topology_die_cpumask(cpu));
630 		c->booted_cores = 1;
631 		return;
632 	}
633 
634 	for_each_cpu(i, cpu_sibling_setup_mask) {
635 		o = &cpu_data(i);
636 
637 		if (match_pkg(c, o) && !topology_same_node(c, o))
638 			x86_has_numa_in_package = true;
639 
640 		if ((i == cpu) || (has_smt && match_smt(c, o)))
641 			link_mask(topology_sibling_cpumask, cpu, i);
642 
643 		if ((i == cpu) || (has_mp && match_llc(c, o)))
644 			link_mask(cpu_llc_shared_mask, cpu, i);
645 
646 		if ((i == cpu) || (has_mp && match_l2c(c, o)))
647 			link_mask(cpu_l2c_shared_mask, cpu, i);
648 
649 		if ((i == cpu) || (has_mp && match_die(c, o)))
650 			link_mask(topology_die_cpumask, cpu, i);
651 	}
652 
653 	threads = cpumask_weight(topology_sibling_cpumask(cpu));
654 	if (threads > __max_smt_threads)
655 		__max_smt_threads = threads;
656 
657 	for_each_cpu(i, topology_sibling_cpumask(cpu))
658 		cpu_data(i).smt_active = threads > 1;
659 
660 	/*
661 	 * This needs a separate iteration over the cpus because we rely on all
662 	 * topology_sibling_cpumask links to be set-up.
663 	 */
664 	for_each_cpu(i, cpu_sibling_setup_mask) {
665 		o = &cpu_data(i);
666 
667 		if ((i == cpu) || (has_mp && match_pkg(c, o))) {
668 			link_mask(topology_core_cpumask, cpu, i);
669 
670 			/*
671 			 *  Does this new cpu bringup a new core?
672 			 */
673 			if (threads == 1) {
674 				/*
675 				 * for each core in package, increment
676 				 * the booted_cores for this new cpu
677 				 */
678 				if (cpumask_first(
679 				    topology_sibling_cpumask(i)) == i)
680 					c->booted_cores++;
681 				/*
682 				 * increment the core count for all
683 				 * the other cpus in this package
684 				 */
685 				if (i != cpu)
686 					cpu_data(i).booted_cores++;
687 			} else if (i != cpu && !c->booted_cores)
688 				c->booted_cores = cpu_data(i).booted_cores;
689 		}
690 	}
691 }
692 
693 /* maps the cpu to the sched domain representing multi-core */
694 const struct cpumask *cpu_coregroup_mask(int cpu)
695 {
696 	return cpu_llc_shared_mask(cpu);
697 }
698 
699 const struct cpumask *cpu_clustergroup_mask(int cpu)
700 {
701 	return cpu_l2c_shared_mask(cpu);
702 }
703 
704 static void impress_friends(void)
705 {
706 	int cpu;
707 	unsigned long bogosum = 0;
708 	/*
709 	 * Allow the user to impress friends.
710 	 */
711 	pr_debug("Before bogomips\n");
712 	for_each_possible_cpu(cpu)
713 		if (cpumask_test_cpu(cpu, cpu_callout_mask))
714 			bogosum += cpu_data(cpu).loops_per_jiffy;
715 	pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
716 		num_online_cpus(),
717 		bogosum/(500000/HZ),
718 		(bogosum/(5000/HZ))%100);
719 
720 	pr_debug("Before bogocount - setting activated=1\n");
721 }
722 
723 void __inquire_remote_apic(int apicid)
724 {
725 	unsigned i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
726 	const char * const names[] = { "ID", "VERSION", "SPIV" };
727 	int timeout;
728 	u32 status;
729 
730 	pr_info("Inquiring remote APIC 0x%x...\n", apicid);
731 
732 	for (i = 0; i < ARRAY_SIZE(regs); i++) {
733 		pr_info("... APIC 0x%x %s: ", apicid, names[i]);
734 
735 		/*
736 		 * Wait for idle.
737 		 */
738 		status = safe_apic_wait_icr_idle();
739 		if (status)
740 			pr_cont("a previous APIC delivery may have failed\n");
741 
742 		apic_icr_write(APIC_DM_REMRD | regs[i], apicid);
743 
744 		timeout = 0;
745 		do {
746 			udelay(100);
747 			status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
748 		} while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
749 
750 		switch (status) {
751 		case APIC_ICR_RR_VALID:
752 			status = apic_read(APIC_RRR);
753 			pr_cont("%08x\n", status);
754 			break;
755 		default:
756 			pr_cont("failed\n");
757 		}
758 	}
759 }
760 
761 /*
762  * The Multiprocessor Specification 1.4 (1997) example code suggests
763  * that there should be a 10ms delay between the BSP asserting INIT
764  * and de-asserting INIT, when starting a remote processor.
765  * But that slows boot and resume on modern processors, which include
766  * many cores and don't require that delay.
767  *
768  * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
769  * Modern processor families are quirked to remove the delay entirely.
770  */
771 #define UDELAY_10MS_DEFAULT 10000
772 
773 static unsigned int init_udelay = UINT_MAX;
774 
775 static int __init cpu_init_udelay(char *str)
776 {
777 	get_option(&str, &init_udelay);
778 
779 	return 0;
780 }
781 early_param("cpu_init_udelay", cpu_init_udelay);
782 
783 static void __init smp_quirk_init_udelay(void)
784 {
785 	/* if cmdline changed it from default, leave it alone */
786 	if (init_udelay != UINT_MAX)
787 		return;
788 
789 	/* if modern processor, use no delay */
790 	if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
791 	    ((boot_cpu_data.x86_vendor == X86_VENDOR_HYGON) && (boot_cpu_data.x86 >= 0x18)) ||
792 	    ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
793 		init_udelay = 0;
794 		return;
795 	}
796 	/* else, use legacy delay */
797 	init_udelay = UDELAY_10MS_DEFAULT;
798 }
799 
800 /*
801  * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
802  * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
803  * won't ... remember to clear down the APIC, etc later.
804  */
805 int
806 wakeup_secondary_cpu_via_nmi(int apicid, unsigned long start_eip)
807 {
808 	u32 dm = apic->dest_mode_logical ? APIC_DEST_LOGICAL : APIC_DEST_PHYSICAL;
809 	unsigned long send_status, accept_status = 0;
810 	int maxlvt;
811 
812 	/* Target chip */
813 	/* Boot on the stack */
814 	/* Kick the second */
815 	apic_icr_write(APIC_DM_NMI | dm, apicid);
816 
817 	pr_debug("Waiting for send to finish...\n");
818 	send_status = safe_apic_wait_icr_idle();
819 
820 	/*
821 	 * Give the other CPU some time to accept the IPI.
822 	 */
823 	udelay(200);
824 	if (APIC_INTEGRATED(boot_cpu_apic_version)) {
825 		maxlvt = lapic_get_maxlvt();
826 		if (maxlvt > 3)			/* Due to the Pentium erratum 3AP.  */
827 			apic_write(APIC_ESR, 0);
828 		accept_status = (apic_read(APIC_ESR) & 0xEF);
829 	}
830 	pr_debug("NMI sent\n");
831 
832 	if (send_status)
833 		pr_err("APIC never delivered???\n");
834 	if (accept_status)
835 		pr_err("APIC delivery error (%lx)\n", accept_status);
836 
837 	return (send_status | accept_status);
838 }
839 
840 static int
841 wakeup_secondary_cpu_via_init(int phys_apicid, unsigned long start_eip)
842 {
843 	unsigned long send_status = 0, accept_status = 0;
844 	int maxlvt, num_starts, j;
845 
846 	maxlvt = lapic_get_maxlvt();
847 
848 	/*
849 	 * Be paranoid about clearing APIC errors.
850 	 */
851 	if (APIC_INTEGRATED(boot_cpu_apic_version)) {
852 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
853 			apic_write(APIC_ESR, 0);
854 		apic_read(APIC_ESR);
855 	}
856 
857 	pr_debug("Asserting INIT\n");
858 
859 	/*
860 	 * Turn INIT on target chip
861 	 */
862 	/*
863 	 * Send IPI
864 	 */
865 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT,
866 		       phys_apicid);
867 
868 	pr_debug("Waiting for send to finish...\n");
869 	send_status = safe_apic_wait_icr_idle();
870 
871 	udelay(init_udelay);
872 
873 	pr_debug("Deasserting INIT\n");
874 
875 	/* Target chip */
876 	/* Send IPI */
877 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
878 
879 	pr_debug("Waiting for send to finish...\n");
880 	send_status = safe_apic_wait_icr_idle();
881 
882 	mb();
883 
884 	/*
885 	 * Should we send STARTUP IPIs ?
886 	 *
887 	 * Determine this based on the APIC version.
888 	 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
889 	 */
890 	if (APIC_INTEGRATED(boot_cpu_apic_version))
891 		num_starts = 2;
892 	else
893 		num_starts = 0;
894 
895 	/*
896 	 * Run STARTUP IPI loop.
897 	 */
898 	pr_debug("#startup loops: %d\n", num_starts);
899 
900 	for (j = 1; j <= num_starts; j++) {
901 		pr_debug("Sending STARTUP #%d\n", j);
902 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
903 			apic_write(APIC_ESR, 0);
904 		apic_read(APIC_ESR);
905 		pr_debug("After apic_write\n");
906 
907 		/*
908 		 * STARTUP IPI
909 		 */
910 
911 		/* Target chip */
912 		/* Boot on the stack */
913 		/* Kick the second */
914 		apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
915 			       phys_apicid);
916 
917 		/*
918 		 * Give the other CPU some time to accept the IPI.
919 		 */
920 		if (init_udelay == 0)
921 			udelay(10);
922 		else
923 			udelay(300);
924 
925 		pr_debug("Startup point 1\n");
926 
927 		pr_debug("Waiting for send to finish...\n");
928 		send_status = safe_apic_wait_icr_idle();
929 
930 		/*
931 		 * Give the other CPU some time to accept the IPI.
932 		 */
933 		if (init_udelay == 0)
934 			udelay(10);
935 		else
936 			udelay(200);
937 
938 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
939 			apic_write(APIC_ESR, 0);
940 		accept_status = (apic_read(APIC_ESR) & 0xEF);
941 		if (send_status || accept_status)
942 			break;
943 	}
944 	pr_debug("After Startup\n");
945 
946 	if (send_status)
947 		pr_err("APIC never delivered???\n");
948 	if (accept_status)
949 		pr_err("APIC delivery error (%lx)\n", accept_status);
950 
951 	return (send_status | accept_status);
952 }
953 
954 /* reduce the number of lines printed when booting a large cpu count system */
955 static void announce_cpu(int cpu, int apicid)
956 {
957 	static int current_node = NUMA_NO_NODE;
958 	int node = early_cpu_to_node(cpu);
959 	static int width, node_width;
960 
961 	if (!width)
962 		width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
963 
964 	if (!node_width)
965 		node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
966 
967 	if (cpu == 1)
968 		printk(KERN_INFO "x86: Booting SMP configuration:\n");
969 
970 	if (system_state < SYSTEM_RUNNING) {
971 		if (node != current_node) {
972 			if (current_node > (-1))
973 				pr_cont("\n");
974 			current_node = node;
975 
976 			printk(KERN_INFO ".... node %*s#%d, CPUs:  ",
977 			       node_width - num_digits(node), " ", node);
978 		}
979 
980 		/* Add padding for the BSP */
981 		if (cpu == 1)
982 			pr_cont("%*s", width + 1, " ");
983 
984 		pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
985 
986 	} else
987 		pr_info("Booting Node %d Processor %d APIC 0x%x\n",
988 			node, cpu, apicid);
989 }
990 
991 static int wakeup_cpu0_nmi(unsigned int cmd, struct pt_regs *regs)
992 {
993 	int cpu;
994 
995 	cpu = smp_processor_id();
996 	if (cpu == 0 && !cpu_online(cpu) && enable_start_cpu0)
997 		return NMI_HANDLED;
998 
999 	return NMI_DONE;
1000 }
1001 
1002 /*
1003  * Wake up AP by INIT, INIT, STARTUP sequence.
1004  *
1005  * Instead of waiting for STARTUP after INITs, BSP will execute the BIOS
1006  * boot-strap code which is not a desired behavior for waking up BSP. To
1007  * void the boot-strap code, wake up CPU0 by NMI instead.
1008  *
1009  * This works to wake up soft offlined CPU0 only. If CPU0 is hard offlined
1010  * (i.e. physically hot removed and then hot added), NMI won't wake it up.
1011  * We'll change this code in the future to wake up hard offlined CPU0 if
1012  * real platform and request are available.
1013  */
1014 static int
1015 wakeup_cpu_via_init_nmi(int cpu, unsigned long start_ip, int apicid,
1016 	       int *cpu0_nmi_registered)
1017 {
1018 	int id;
1019 	int boot_error;
1020 
1021 	preempt_disable();
1022 
1023 	/*
1024 	 * Wake up AP by INIT, INIT, STARTUP sequence.
1025 	 */
1026 	if (cpu) {
1027 		boot_error = wakeup_secondary_cpu_via_init(apicid, start_ip);
1028 		goto out;
1029 	}
1030 
1031 	/*
1032 	 * Wake up BSP by nmi.
1033 	 *
1034 	 * Register a NMI handler to help wake up CPU0.
1035 	 */
1036 	boot_error = register_nmi_handler(NMI_LOCAL,
1037 					  wakeup_cpu0_nmi, 0, "wake_cpu0");
1038 
1039 	if (!boot_error) {
1040 		enable_start_cpu0 = 1;
1041 		*cpu0_nmi_registered = 1;
1042 		id = apic->dest_mode_logical ? cpu0_logical_apicid : apicid;
1043 		boot_error = wakeup_secondary_cpu_via_nmi(id, start_ip);
1044 	}
1045 
1046 out:
1047 	preempt_enable();
1048 
1049 	return boot_error;
1050 }
1051 
1052 int common_cpu_up(unsigned int cpu, struct task_struct *idle)
1053 {
1054 	int ret;
1055 
1056 	/* Just in case we booted with a single CPU. */
1057 	alternatives_enable_smp();
1058 
1059 	per_cpu(current_task, cpu) = idle;
1060 	cpu_init_stack_canary(cpu, idle);
1061 
1062 	/* Initialize the interrupt stack(s) */
1063 	ret = irq_init_percpu_irqstack(cpu);
1064 	if (ret)
1065 		return ret;
1066 
1067 #ifdef CONFIG_X86_32
1068 	/* Stack for startup_32 can be just as for start_secondary onwards */
1069 	per_cpu(cpu_current_top_of_stack, cpu) = task_top_of_stack(idle);
1070 #else
1071 	initial_gs = per_cpu_offset(cpu);
1072 #endif
1073 	return 0;
1074 }
1075 
1076 /*
1077  * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
1078  * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
1079  * Returns zero if CPU booted OK, else error code from
1080  * ->wakeup_secondary_cpu.
1081  */
1082 static int do_boot_cpu(int apicid, int cpu, struct task_struct *idle,
1083 		       int *cpu0_nmi_registered)
1084 {
1085 	/* start_ip had better be page-aligned! */
1086 	unsigned long start_ip = real_mode_header->trampoline_start;
1087 
1088 	unsigned long boot_error = 0;
1089 	unsigned long timeout;
1090 
1091 	idle->thread.sp = (unsigned long)task_pt_regs(idle);
1092 	early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
1093 	initial_code = (unsigned long)start_secondary;
1094 	initial_stack  = idle->thread.sp;
1095 
1096 	/* Enable the espfix hack for this CPU */
1097 	init_espfix_ap(cpu);
1098 
1099 	/* So we see what's up */
1100 	announce_cpu(cpu, apicid);
1101 
1102 	/*
1103 	 * This grunge runs the startup process for
1104 	 * the targeted processor.
1105 	 */
1106 
1107 	if (x86_platform.legacy.warm_reset) {
1108 
1109 		pr_debug("Setting warm reset code and vector.\n");
1110 
1111 		smpboot_setup_warm_reset_vector(start_ip);
1112 		/*
1113 		 * Be paranoid about clearing APIC errors.
1114 		*/
1115 		if (APIC_INTEGRATED(boot_cpu_apic_version)) {
1116 			apic_write(APIC_ESR, 0);
1117 			apic_read(APIC_ESR);
1118 		}
1119 	}
1120 
1121 	/*
1122 	 * AP might wait on cpu_callout_mask in cpu_init() with
1123 	 * cpu_initialized_mask set if previous attempt to online
1124 	 * it timed-out. Clear cpu_initialized_mask so that after
1125 	 * INIT/SIPI it could start with a clean state.
1126 	 */
1127 	cpumask_clear_cpu(cpu, cpu_initialized_mask);
1128 	smp_mb();
1129 
1130 	/*
1131 	 * Wake up a CPU in difference cases:
1132 	 * - Use the method in the APIC driver if it's defined
1133 	 * Otherwise,
1134 	 * - Use an INIT boot APIC message for APs or NMI for BSP.
1135 	 */
1136 	if (apic->wakeup_secondary_cpu)
1137 		boot_error = apic->wakeup_secondary_cpu(apicid, start_ip);
1138 	else
1139 		boot_error = wakeup_cpu_via_init_nmi(cpu, start_ip, apicid,
1140 						     cpu0_nmi_registered);
1141 
1142 	if (!boot_error) {
1143 		/*
1144 		 * Wait 10s total for first sign of life from AP
1145 		 */
1146 		boot_error = -1;
1147 		timeout = jiffies + 10*HZ;
1148 		while (time_before(jiffies, timeout)) {
1149 			if (cpumask_test_cpu(cpu, cpu_initialized_mask)) {
1150 				/*
1151 				 * Tell AP to proceed with initialization
1152 				 */
1153 				cpumask_set_cpu(cpu, cpu_callout_mask);
1154 				boot_error = 0;
1155 				break;
1156 			}
1157 			schedule();
1158 		}
1159 	}
1160 
1161 	if (!boot_error) {
1162 		/*
1163 		 * Wait till AP completes initial initialization
1164 		 */
1165 		while (!cpumask_test_cpu(cpu, cpu_callin_mask)) {
1166 			/*
1167 			 * Allow other tasks to run while we wait for the
1168 			 * AP to come online. This also gives a chance
1169 			 * for the MTRR work(triggered by the AP coming online)
1170 			 * to be completed in the stop machine context.
1171 			 */
1172 			schedule();
1173 		}
1174 	}
1175 
1176 	if (x86_platform.legacy.warm_reset) {
1177 		/*
1178 		 * Cleanup possible dangling ends...
1179 		 */
1180 		smpboot_restore_warm_reset_vector();
1181 	}
1182 
1183 	return boot_error;
1184 }
1185 
1186 int native_cpu_up(unsigned int cpu, struct task_struct *tidle)
1187 {
1188 	int apicid = apic->cpu_present_to_apicid(cpu);
1189 	int cpu0_nmi_registered = 0;
1190 	unsigned long flags;
1191 	int err, ret = 0;
1192 
1193 	lockdep_assert_irqs_enabled();
1194 
1195 	pr_debug("++++++++++++++++++++=_---CPU UP  %u\n", cpu);
1196 
1197 	if (apicid == BAD_APICID ||
1198 	    !physid_isset(apicid, phys_cpu_present_map) ||
1199 	    !apic->apic_id_valid(apicid)) {
1200 		pr_err("%s: bad cpu %d\n", __func__, cpu);
1201 		return -EINVAL;
1202 	}
1203 
1204 	/*
1205 	 * Already booted CPU?
1206 	 */
1207 	if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
1208 		pr_debug("do_boot_cpu %d Already started\n", cpu);
1209 		return -ENOSYS;
1210 	}
1211 
1212 	/*
1213 	 * Save current MTRR state in case it was changed since early boot
1214 	 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1215 	 */
1216 	mtrr_save_state();
1217 
1218 	/* x86 CPUs take themselves offline, so delayed offline is OK. */
1219 	err = cpu_check_up_prepare(cpu);
1220 	if (err && err != -EBUSY)
1221 		return err;
1222 
1223 	/* the FPU context is blank, nobody can own it */
1224 	per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
1225 
1226 	err = common_cpu_up(cpu, tidle);
1227 	if (err)
1228 		return err;
1229 
1230 	err = do_boot_cpu(apicid, cpu, tidle, &cpu0_nmi_registered);
1231 	if (err) {
1232 		pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
1233 		ret = -EIO;
1234 		goto unreg_nmi;
1235 	}
1236 
1237 	/*
1238 	 * Check TSC synchronization with the AP (keep irqs disabled
1239 	 * while doing so):
1240 	 */
1241 	local_irq_save(flags);
1242 	check_tsc_sync_source(cpu);
1243 	local_irq_restore(flags);
1244 
1245 	while (!cpu_online(cpu)) {
1246 		cpu_relax();
1247 		touch_nmi_watchdog();
1248 	}
1249 
1250 unreg_nmi:
1251 	/*
1252 	 * Clean up the nmi handler. Do this after the callin and callout sync
1253 	 * to avoid impact of possible long unregister time.
1254 	 */
1255 	if (cpu0_nmi_registered)
1256 		unregister_nmi_handler(NMI_LOCAL, "wake_cpu0");
1257 
1258 	return ret;
1259 }
1260 
1261 /**
1262  * arch_disable_smp_support() - disables SMP support for x86 at runtime
1263  */
1264 void arch_disable_smp_support(void)
1265 {
1266 	disable_ioapic_support();
1267 }
1268 
1269 /*
1270  * Fall back to non SMP mode after errors.
1271  *
1272  * RED-PEN audit/test this more. I bet there is more state messed up here.
1273  */
1274 static __init void disable_smp(void)
1275 {
1276 	pr_info("SMP disabled\n");
1277 
1278 	disable_ioapic_support();
1279 
1280 	init_cpu_present(cpumask_of(0));
1281 	init_cpu_possible(cpumask_of(0));
1282 
1283 	if (smp_found_config)
1284 		physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
1285 	else
1286 		physid_set_mask_of_physid(0, &phys_cpu_present_map);
1287 	cpumask_set_cpu(0, topology_sibling_cpumask(0));
1288 	cpumask_set_cpu(0, topology_core_cpumask(0));
1289 	cpumask_set_cpu(0, topology_die_cpumask(0));
1290 }
1291 
1292 /*
1293  * Various sanity checks.
1294  */
1295 static void __init smp_sanity_check(void)
1296 {
1297 	preempt_disable();
1298 
1299 #if !defined(CONFIG_X86_BIGSMP) && defined(CONFIG_X86_32)
1300 	if (def_to_bigsmp && nr_cpu_ids > 8) {
1301 		unsigned int cpu;
1302 		unsigned nr;
1303 
1304 		pr_warn("More than 8 CPUs detected - skipping them\n"
1305 			"Use CONFIG_X86_BIGSMP\n");
1306 
1307 		nr = 0;
1308 		for_each_present_cpu(cpu) {
1309 			if (nr >= 8)
1310 				set_cpu_present(cpu, false);
1311 			nr++;
1312 		}
1313 
1314 		nr = 0;
1315 		for_each_possible_cpu(cpu) {
1316 			if (nr >= 8)
1317 				set_cpu_possible(cpu, false);
1318 			nr++;
1319 		}
1320 
1321 		nr_cpu_ids = 8;
1322 	}
1323 #endif
1324 
1325 	if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map)) {
1326 		pr_warn("weird, boot CPU (#%d) not listed by the BIOS\n",
1327 			hard_smp_processor_id());
1328 
1329 		physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1330 	}
1331 
1332 	/*
1333 	 * Should not be necessary because the MP table should list the boot
1334 	 * CPU too, but we do it for the sake of robustness anyway.
1335 	 */
1336 	if (!apic->check_phys_apicid_present(boot_cpu_physical_apicid)) {
1337 		pr_notice("weird, boot CPU (#%d) not listed by the BIOS\n",
1338 			  boot_cpu_physical_apicid);
1339 		physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1340 	}
1341 	preempt_enable();
1342 }
1343 
1344 static void __init smp_cpu_index_default(void)
1345 {
1346 	int i;
1347 	struct cpuinfo_x86 *c;
1348 
1349 	for_each_possible_cpu(i) {
1350 		c = &cpu_data(i);
1351 		/* mark all to hotplug */
1352 		c->cpu_index = nr_cpu_ids;
1353 	}
1354 }
1355 
1356 static void __init smp_get_logical_apicid(void)
1357 {
1358 	if (x2apic_mode)
1359 		cpu0_logical_apicid = apic_read(APIC_LDR);
1360 	else
1361 		cpu0_logical_apicid = GET_APIC_LOGICAL_ID(apic_read(APIC_LDR));
1362 }
1363 
1364 void __init smp_prepare_cpus_common(void)
1365 {
1366 	unsigned int i;
1367 
1368 	smp_cpu_index_default();
1369 
1370 	/*
1371 	 * Setup boot CPU information
1372 	 */
1373 	smp_store_boot_cpu_info(); /* Final full version of the data */
1374 	cpumask_copy(cpu_callin_mask, cpumask_of(0));
1375 	mb();
1376 
1377 	for_each_possible_cpu(i) {
1378 		zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1379 		zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1380 		zalloc_cpumask_var(&per_cpu(cpu_die_map, i), GFP_KERNEL);
1381 		zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1382 		zalloc_cpumask_var(&per_cpu(cpu_l2c_shared_map, i), GFP_KERNEL);
1383 	}
1384 
1385 	/*
1386 	 * Set 'default' x86 topology, this matches default_topology() in that
1387 	 * it has NUMA nodes as a topology level. See also
1388 	 * native_smp_cpus_done().
1389 	 *
1390 	 * Must be done before set_cpus_sibling_map() is ran.
1391 	 */
1392 	set_sched_topology(x86_topology);
1393 
1394 	set_cpu_sibling_map(0);
1395 }
1396 
1397 /*
1398  * Prepare for SMP bootup.
1399  * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1400  *            for common interface support.
1401  */
1402 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1403 {
1404 	smp_prepare_cpus_common();
1405 
1406 	init_freq_invariance(false, false);
1407 	smp_sanity_check();
1408 
1409 	switch (apic_intr_mode) {
1410 	case APIC_PIC:
1411 	case APIC_VIRTUAL_WIRE_NO_CONFIG:
1412 		disable_smp();
1413 		return;
1414 	case APIC_SYMMETRIC_IO_NO_ROUTING:
1415 		disable_smp();
1416 		/* Setup local timer */
1417 		x86_init.timers.setup_percpu_clockev();
1418 		return;
1419 	case APIC_VIRTUAL_WIRE:
1420 	case APIC_SYMMETRIC_IO:
1421 		break;
1422 	}
1423 
1424 	/* Setup local timer */
1425 	x86_init.timers.setup_percpu_clockev();
1426 
1427 	smp_get_logical_apicid();
1428 
1429 	pr_info("CPU0: ");
1430 	print_cpu_info(&cpu_data(0));
1431 
1432 	uv_system_init();
1433 
1434 	set_mtrr_aps_delayed_init();
1435 
1436 	smp_quirk_init_udelay();
1437 
1438 	speculative_store_bypass_ht_init();
1439 }
1440 
1441 void arch_thaw_secondary_cpus_begin(void)
1442 {
1443 	set_mtrr_aps_delayed_init();
1444 }
1445 
1446 void arch_thaw_secondary_cpus_end(void)
1447 {
1448 	mtrr_aps_init();
1449 }
1450 
1451 /*
1452  * Early setup to make printk work.
1453  */
1454 void __init native_smp_prepare_boot_cpu(void)
1455 {
1456 	int me = smp_processor_id();
1457 	switch_to_new_gdt(me);
1458 	/* already set me in cpu_online_mask in boot_cpu_init() */
1459 	cpumask_set_cpu(me, cpu_callout_mask);
1460 	cpu_set_state_online(me);
1461 	native_pv_lock_init();
1462 }
1463 
1464 void __init calculate_max_logical_packages(void)
1465 {
1466 	int ncpus;
1467 
1468 	/*
1469 	 * Today neither Intel nor AMD support heterogeneous systems so
1470 	 * extrapolate the boot cpu's data to all packages.
1471 	 */
1472 	ncpus = cpu_data(0).booted_cores * topology_max_smt_threads();
1473 	__max_logical_packages = DIV_ROUND_UP(total_cpus, ncpus);
1474 	pr_info("Max logical packages: %u\n", __max_logical_packages);
1475 }
1476 
1477 void __init native_smp_cpus_done(unsigned int max_cpus)
1478 {
1479 	pr_debug("Boot done\n");
1480 
1481 	calculate_max_logical_packages();
1482 
1483 	/* XXX for now assume numa-in-package and hybrid don't overlap */
1484 	if (x86_has_numa_in_package)
1485 		set_sched_topology(x86_numa_in_package_topology);
1486 	if (cpu_feature_enabled(X86_FEATURE_HYBRID_CPU))
1487 		set_sched_topology(x86_hybrid_topology);
1488 
1489 	nmi_selftest();
1490 	impress_friends();
1491 	mtrr_aps_init();
1492 }
1493 
1494 static int __initdata setup_possible_cpus = -1;
1495 static int __init _setup_possible_cpus(char *str)
1496 {
1497 	get_option(&str, &setup_possible_cpus);
1498 	return 0;
1499 }
1500 early_param("possible_cpus", _setup_possible_cpus);
1501 
1502 
1503 /*
1504  * cpu_possible_mask should be static, it cannot change as cpu's
1505  * are onlined, or offlined. The reason is per-cpu data-structures
1506  * are allocated by some modules at init time, and don't expect to
1507  * do this dynamically on cpu arrival/departure.
1508  * cpu_present_mask on the other hand can change dynamically.
1509  * In case when cpu_hotplug is not compiled, then we resort to current
1510  * behaviour, which is cpu_possible == cpu_present.
1511  * - Ashok Raj
1512  *
1513  * Three ways to find out the number of additional hotplug CPUs:
1514  * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1515  * - The user can overwrite it with possible_cpus=NUM
1516  * - Otherwise don't reserve additional CPUs.
1517  * We do this because additional CPUs waste a lot of memory.
1518  * -AK
1519  */
1520 __init void prefill_possible_map(void)
1521 {
1522 	int i, possible;
1523 
1524 	/* No boot processor was found in mptable or ACPI MADT */
1525 	if (!num_processors) {
1526 		if (boot_cpu_has(X86_FEATURE_APIC)) {
1527 			int apicid = boot_cpu_physical_apicid;
1528 			int cpu = hard_smp_processor_id();
1529 
1530 			pr_warn("Boot CPU (id %d) not listed by BIOS\n", cpu);
1531 
1532 			/* Make sure boot cpu is enumerated */
1533 			if (apic->cpu_present_to_apicid(0) == BAD_APICID &&
1534 			    apic->apic_id_valid(apicid))
1535 				generic_processor_info(apicid, boot_cpu_apic_version);
1536 		}
1537 
1538 		if (!num_processors)
1539 			num_processors = 1;
1540 	}
1541 
1542 	i = setup_max_cpus ?: 1;
1543 	if (setup_possible_cpus == -1) {
1544 		possible = num_processors;
1545 #ifdef CONFIG_HOTPLUG_CPU
1546 		if (setup_max_cpus)
1547 			possible += disabled_cpus;
1548 #else
1549 		if (possible > i)
1550 			possible = i;
1551 #endif
1552 	} else
1553 		possible = setup_possible_cpus;
1554 
1555 	total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1556 
1557 	/* nr_cpu_ids could be reduced via nr_cpus= */
1558 	if (possible > nr_cpu_ids) {
1559 		pr_warn("%d Processors exceeds NR_CPUS limit of %u\n",
1560 			possible, nr_cpu_ids);
1561 		possible = nr_cpu_ids;
1562 	}
1563 
1564 #ifdef CONFIG_HOTPLUG_CPU
1565 	if (!setup_max_cpus)
1566 #endif
1567 	if (possible > i) {
1568 		pr_warn("%d Processors exceeds max_cpus limit of %u\n",
1569 			possible, setup_max_cpus);
1570 		possible = i;
1571 	}
1572 
1573 	nr_cpu_ids = possible;
1574 
1575 	pr_info("Allowing %d CPUs, %d hotplug CPUs\n",
1576 		possible, max_t(int, possible - num_processors, 0));
1577 
1578 	reset_cpu_possible_mask();
1579 
1580 	for (i = 0; i < possible; i++)
1581 		set_cpu_possible(i, true);
1582 }
1583 
1584 #ifdef CONFIG_HOTPLUG_CPU
1585 
1586 /* Recompute SMT state for all CPUs on offline */
1587 static void recompute_smt_state(void)
1588 {
1589 	int max_threads, cpu;
1590 
1591 	max_threads = 0;
1592 	for_each_online_cpu (cpu) {
1593 		int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1594 
1595 		if (threads > max_threads)
1596 			max_threads = threads;
1597 	}
1598 	__max_smt_threads = max_threads;
1599 }
1600 
1601 static void remove_siblinginfo(int cpu)
1602 {
1603 	int sibling;
1604 	struct cpuinfo_x86 *c = &cpu_data(cpu);
1605 
1606 	for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1607 		cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1608 		/*/
1609 		 * last thread sibling in this cpu core going down
1610 		 */
1611 		if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1612 			cpu_data(sibling).booted_cores--;
1613 	}
1614 
1615 	for_each_cpu(sibling, topology_die_cpumask(cpu))
1616 		cpumask_clear_cpu(cpu, topology_die_cpumask(sibling));
1617 
1618 	for_each_cpu(sibling, topology_sibling_cpumask(cpu)) {
1619 		cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1620 		if (cpumask_weight(topology_sibling_cpumask(sibling)) == 1)
1621 			cpu_data(sibling).smt_active = false;
1622 	}
1623 
1624 	for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1625 		cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1626 	for_each_cpu(sibling, cpu_l2c_shared_mask(cpu))
1627 		cpumask_clear_cpu(cpu, cpu_l2c_shared_mask(sibling));
1628 	cpumask_clear(cpu_llc_shared_mask(cpu));
1629 	cpumask_clear(cpu_l2c_shared_mask(cpu));
1630 	cpumask_clear(topology_sibling_cpumask(cpu));
1631 	cpumask_clear(topology_core_cpumask(cpu));
1632 	cpumask_clear(topology_die_cpumask(cpu));
1633 	c->cpu_core_id = 0;
1634 	c->booted_cores = 0;
1635 	cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1636 	recompute_smt_state();
1637 }
1638 
1639 static void remove_cpu_from_maps(int cpu)
1640 {
1641 	set_cpu_online(cpu, false);
1642 	cpumask_clear_cpu(cpu, cpu_callout_mask);
1643 	cpumask_clear_cpu(cpu, cpu_callin_mask);
1644 	/* was set by cpu_init() */
1645 	cpumask_clear_cpu(cpu, cpu_initialized_mask);
1646 	numa_remove_cpu(cpu);
1647 }
1648 
1649 void cpu_disable_common(void)
1650 {
1651 	int cpu = smp_processor_id();
1652 
1653 	remove_siblinginfo(cpu);
1654 
1655 	/* It's now safe to remove this processor from the online map */
1656 	lock_vector_lock();
1657 	remove_cpu_from_maps(cpu);
1658 	unlock_vector_lock();
1659 	fixup_irqs();
1660 	lapic_offline();
1661 }
1662 
1663 int native_cpu_disable(void)
1664 {
1665 	int ret;
1666 
1667 	ret = lapic_can_unplug_cpu();
1668 	if (ret)
1669 		return ret;
1670 
1671 	cpu_disable_common();
1672 
1673         /*
1674          * Disable the local APIC. Otherwise IPI broadcasts will reach
1675          * it. It still responds normally to INIT, NMI, SMI, and SIPI
1676          * messages.
1677          *
1678          * Disabling the APIC must happen after cpu_disable_common()
1679          * which invokes fixup_irqs().
1680          *
1681          * Disabling the APIC preserves already set bits in IRR, but
1682          * an interrupt arriving after disabling the local APIC does not
1683          * set the corresponding IRR bit.
1684          *
1685          * fixup_irqs() scans IRR for set bits so it can raise a not
1686          * yet handled interrupt on the new destination CPU via an IPI
1687          * but obviously it can't do so for IRR bits which are not set.
1688          * IOW, interrupts arriving after disabling the local APIC will
1689          * be lost.
1690          */
1691 	apic_soft_disable();
1692 
1693 	return 0;
1694 }
1695 
1696 int common_cpu_die(unsigned int cpu)
1697 {
1698 	int ret = 0;
1699 
1700 	/* We don't do anything here: idle task is faking death itself. */
1701 
1702 	/* They ack this in play_dead() by setting CPU_DEAD */
1703 	if (cpu_wait_death(cpu, 5)) {
1704 		if (system_state == SYSTEM_RUNNING)
1705 			pr_info("CPU %u is now offline\n", cpu);
1706 	} else {
1707 		pr_err("CPU %u didn't die...\n", cpu);
1708 		ret = -1;
1709 	}
1710 
1711 	return ret;
1712 }
1713 
1714 void native_cpu_die(unsigned int cpu)
1715 {
1716 	common_cpu_die(cpu);
1717 }
1718 
1719 void play_dead_common(void)
1720 {
1721 	idle_task_exit();
1722 
1723 	/* Ack it */
1724 	(void)cpu_report_death();
1725 
1726 	/*
1727 	 * With physical CPU hotplug, we should halt the cpu
1728 	 */
1729 	local_irq_disable();
1730 }
1731 
1732 /**
1733  * cond_wakeup_cpu0 - Wake up CPU0 if needed.
1734  *
1735  * If NMI wants to wake up CPU0, start CPU0.
1736  */
1737 void cond_wakeup_cpu0(void)
1738 {
1739 	if (smp_processor_id() == 0 && enable_start_cpu0)
1740 		start_cpu0();
1741 }
1742 EXPORT_SYMBOL_GPL(cond_wakeup_cpu0);
1743 
1744 /*
1745  * We need to flush the caches before going to sleep, lest we have
1746  * dirty data in our caches when we come back up.
1747  */
1748 static inline void mwait_play_dead(void)
1749 {
1750 	unsigned int eax, ebx, ecx, edx;
1751 	unsigned int highest_cstate = 0;
1752 	unsigned int highest_subcstate = 0;
1753 	void *mwait_ptr;
1754 	int i;
1755 
1756 	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD ||
1757 	    boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
1758 		return;
1759 	if (!this_cpu_has(X86_FEATURE_MWAIT))
1760 		return;
1761 	if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1762 		return;
1763 	if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1764 		return;
1765 
1766 	eax = CPUID_MWAIT_LEAF;
1767 	ecx = 0;
1768 	native_cpuid(&eax, &ebx, &ecx, &edx);
1769 
1770 	/*
1771 	 * eax will be 0 if EDX enumeration is not valid.
1772 	 * Initialized below to cstate, sub_cstate value when EDX is valid.
1773 	 */
1774 	if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1775 		eax = 0;
1776 	} else {
1777 		edx >>= MWAIT_SUBSTATE_SIZE;
1778 		for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1779 			if (edx & MWAIT_SUBSTATE_MASK) {
1780 				highest_cstate = i;
1781 				highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1782 			}
1783 		}
1784 		eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1785 			(highest_subcstate - 1);
1786 	}
1787 
1788 	/*
1789 	 * This should be a memory location in a cache line which is
1790 	 * unlikely to be touched by other processors.  The actual
1791 	 * content is immaterial as it is not actually modified in any way.
1792 	 */
1793 	mwait_ptr = &current_thread_info()->flags;
1794 
1795 	wbinvd();
1796 
1797 	while (1) {
1798 		/*
1799 		 * The CLFLUSH is a workaround for erratum AAI65 for
1800 		 * the Xeon 7400 series.  It's not clear it is actually
1801 		 * needed, but it should be harmless in either case.
1802 		 * The WBINVD is insufficient due to the spurious-wakeup
1803 		 * case where we return around the loop.
1804 		 */
1805 		mb();
1806 		clflush(mwait_ptr);
1807 		mb();
1808 		__monitor(mwait_ptr, 0, 0);
1809 		mb();
1810 		__mwait(eax, 0);
1811 
1812 		cond_wakeup_cpu0();
1813 	}
1814 }
1815 
1816 void hlt_play_dead(void)
1817 {
1818 	if (__this_cpu_read(cpu_info.x86) >= 4)
1819 		wbinvd();
1820 
1821 	while (1) {
1822 		native_halt();
1823 
1824 		cond_wakeup_cpu0();
1825 	}
1826 }
1827 
1828 void native_play_dead(void)
1829 {
1830 	play_dead_common();
1831 	tboot_shutdown(TB_SHUTDOWN_WFS);
1832 
1833 	mwait_play_dead();	/* Only returns on failure */
1834 	if (cpuidle_play_dead())
1835 		hlt_play_dead();
1836 }
1837 
1838 #else /* ... !CONFIG_HOTPLUG_CPU */
1839 int native_cpu_disable(void)
1840 {
1841 	return -ENOSYS;
1842 }
1843 
1844 void native_cpu_die(unsigned int cpu)
1845 {
1846 	/* We said "no" in __cpu_disable */
1847 	BUG();
1848 }
1849 
1850 void native_play_dead(void)
1851 {
1852 	BUG();
1853 }
1854 
1855 #endif
1856 
1857 #ifdef CONFIG_X86_64
1858 /*
1859  * APERF/MPERF frequency ratio computation.
1860  *
1861  * The scheduler wants to do frequency invariant accounting and needs a <1
1862  * ratio to account for the 'current' frequency, corresponding to
1863  * freq_curr / freq_max.
1864  *
1865  * Since the frequency freq_curr on x86 is controlled by micro-controller and
1866  * our P-state setting is little more than a request/hint, we need to observe
1867  * the effective frequency 'BusyMHz', i.e. the average frequency over a time
1868  * interval after discarding idle time. This is given by:
1869  *
1870  *   BusyMHz = delta_APERF / delta_MPERF * freq_base
1871  *
1872  * where freq_base is the max non-turbo P-state.
1873  *
1874  * The freq_max term has to be set to a somewhat arbitrary value, because we
1875  * can't know which turbo states will be available at a given point in time:
1876  * it all depends on the thermal headroom of the entire package. We set it to
1877  * the turbo level with 4 cores active.
1878  *
1879  * Benchmarks show that's a good compromise between the 1C turbo ratio
1880  * (freq_curr/freq_max would rarely reach 1) and something close to freq_base,
1881  * which would ignore the entire turbo range (a conspicuous part, making
1882  * freq_curr/freq_max always maxed out).
1883  *
1884  * An exception to the heuristic above is the Atom uarch, where we choose the
1885  * highest turbo level for freq_max since Atom's are generally oriented towards
1886  * power efficiency.
1887  *
1888  * Setting freq_max to anything less than the 1C turbo ratio makes the ratio
1889  * freq_curr / freq_max to eventually grow >1, in which case we clip it to 1.
1890  */
1891 
1892 DEFINE_STATIC_KEY_FALSE(arch_scale_freq_key);
1893 
1894 static DEFINE_PER_CPU(u64, arch_prev_aperf);
1895 static DEFINE_PER_CPU(u64, arch_prev_mperf);
1896 static u64 arch_turbo_freq_ratio = SCHED_CAPACITY_SCALE;
1897 static u64 arch_max_freq_ratio = SCHED_CAPACITY_SCALE;
1898 
1899 void arch_set_max_freq_ratio(bool turbo_disabled)
1900 {
1901 	arch_max_freq_ratio = turbo_disabled ? SCHED_CAPACITY_SCALE :
1902 					arch_turbo_freq_ratio;
1903 }
1904 EXPORT_SYMBOL_GPL(arch_set_max_freq_ratio);
1905 
1906 static bool turbo_disabled(void)
1907 {
1908 	u64 misc_en;
1909 	int err;
1910 
1911 	err = rdmsrl_safe(MSR_IA32_MISC_ENABLE, &misc_en);
1912 	if (err)
1913 		return false;
1914 
1915 	return (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
1916 }
1917 
1918 static bool slv_set_max_freq_ratio(u64 *base_freq, u64 *turbo_freq)
1919 {
1920 	int err;
1921 
1922 	err = rdmsrl_safe(MSR_ATOM_CORE_RATIOS, base_freq);
1923 	if (err)
1924 		return false;
1925 
1926 	err = rdmsrl_safe(MSR_ATOM_CORE_TURBO_RATIOS, turbo_freq);
1927 	if (err)
1928 		return false;
1929 
1930 	*base_freq = (*base_freq >> 16) & 0x3F;     /* max P state */
1931 	*turbo_freq = *turbo_freq & 0x3F;           /* 1C turbo    */
1932 
1933 	return true;
1934 }
1935 
1936 #define X86_MATCH(model)					\
1937 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6,		\
1938 		INTEL_FAM6_##model, X86_FEATURE_APERFMPERF, NULL)
1939 
1940 static const struct x86_cpu_id has_knl_turbo_ratio_limits[] = {
1941 	X86_MATCH(XEON_PHI_KNL),
1942 	X86_MATCH(XEON_PHI_KNM),
1943 	{}
1944 };
1945 
1946 static const struct x86_cpu_id has_skx_turbo_ratio_limits[] = {
1947 	X86_MATCH(SKYLAKE_X),
1948 	{}
1949 };
1950 
1951 static const struct x86_cpu_id has_glm_turbo_ratio_limits[] = {
1952 	X86_MATCH(ATOM_GOLDMONT),
1953 	X86_MATCH(ATOM_GOLDMONT_D),
1954 	X86_MATCH(ATOM_GOLDMONT_PLUS),
1955 	{}
1956 };
1957 
1958 static bool knl_set_max_freq_ratio(u64 *base_freq, u64 *turbo_freq,
1959 				int num_delta_fratio)
1960 {
1961 	int fratio, delta_fratio, found;
1962 	int err, i;
1963 	u64 msr;
1964 
1965 	err = rdmsrl_safe(MSR_PLATFORM_INFO, base_freq);
1966 	if (err)
1967 		return false;
1968 
1969 	*base_freq = (*base_freq >> 8) & 0xFF;	    /* max P state */
1970 
1971 	err = rdmsrl_safe(MSR_TURBO_RATIO_LIMIT, &msr);
1972 	if (err)
1973 		return false;
1974 
1975 	fratio = (msr >> 8) & 0xFF;
1976 	i = 16;
1977 	found = 0;
1978 	do {
1979 		if (found >= num_delta_fratio) {
1980 			*turbo_freq = fratio;
1981 			return true;
1982 		}
1983 
1984 		delta_fratio = (msr >> (i + 5)) & 0x7;
1985 
1986 		if (delta_fratio) {
1987 			found += 1;
1988 			fratio -= delta_fratio;
1989 		}
1990 
1991 		i += 8;
1992 	} while (i < 64);
1993 
1994 	return true;
1995 }
1996 
1997 static bool skx_set_max_freq_ratio(u64 *base_freq, u64 *turbo_freq, int size)
1998 {
1999 	u64 ratios, counts;
2000 	u32 group_size;
2001 	int err, i;
2002 
2003 	err = rdmsrl_safe(MSR_PLATFORM_INFO, base_freq);
2004 	if (err)
2005 		return false;
2006 
2007 	*base_freq = (*base_freq >> 8) & 0xFF;      /* max P state */
2008 
2009 	err = rdmsrl_safe(MSR_TURBO_RATIO_LIMIT, &ratios);
2010 	if (err)
2011 		return false;
2012 
2013 	err = rdmsrl_safe(MSR_TURBO_RATIO_LIMIT1, &counts);
2014 	if (err)
2015 		return false;
2016 
2017 	for (i = 0; i < 64; i += 8) {
2018 		group_size = (counts >> i) & 0xFF;
2019 		if (group_size >= size) {
2020 			*turbo_freq = (ratios >> i) & 0xFF;
2021 			return true;
2022 		}
2023 	}
2024 
2025 	return false;
2026 }
2027 
2028 static bool core_set_max_freq_ratio(u64 *base_freq, u64 *turbo_freq)
2029 {
2030 	u64 msr;
2031 	int err;
2032 
2033 	err = rdmsrl_safe(MSR_PLATFORM_INFO, base_freq);
2034 	if (err)
2035 		return false;
2036 
2037 	err = rdmsrl_safe(MSR_TURBO_RATIO_LIMIT, &msr);
2038 	if (err)
2039 		return false;
2040 
2041 	*base_freq = (*base_freq >> 8) & 0xFF;    /* max P state */
2042 	*turbo_freq = (msr >> 24) & 0xFF;         /* 4C turbo    */
2043 
2044 	/* The CPU may have less than 4 cores */
2045 	if (!*turbo_freq)
2046 		*turbo_freq = msr & 0xFF;         /* 1C turbo    */
2047 
2048 	return true;
2049 }
2050 
2051 static bool intel_set_max_freq_ratio(void)
2052 {
2053 	u64 base_freq, turbo_freq;
2054 	u64 turbo_ratio;
2055 
2056 	if (slv_set_max_freq_ratio(&base_freq, &turbo_freq))
2057 		goto out;
2058 
2059 	if (x86_match_cpu(has_glm_turbo_ratio_limits) &&
2060 	    skx_set_max_freq_ratio(&base_freq, &turbo_freq, 1))
2061 		goto out;
2062 
2063 	if (x86_match_cpu(has_knl_turbo_ratio_limits) &&
2064 	    knl_set_max_freq_ratio(&base_freq, &turbo_freq, 1))
2065 		goto out;
2066 
2067 	if (x86_match_cpu(has_skx_turbo_ratio_limits) &&
2068 	    skx_set_max_freq_ratio(&base_freq, &turbo_freq, 4))
2069 		goto out;
2070 
2071 	if (core_set_max_freq_ratio(&base_freq, &turbo_freq))
2072 		goto out;
2073 
2074 	return false;
2075 
2076 out:
2077 	/*
2078 	 * Some hypervisors advertise X86_FEATURE_APERFMPERF
2079 	 * but then fill all MSR's with zeroes.
2080 	 * Some CPUs have turbo boost but don't declare any turbo ratio
2081 	 * in MSR_TURBO_RATIO_LIMIT.
2082 	 */
2083 	if (!base_freq || !turbo_freq) {
2084 		pr_debug("Couldn't determine cpu base or turbo frequency, necessary for scale-invariant accounting.\n");
2085 		return false;
2086 	}
2087 
2088 	turbo_ratio = div_u64(turbo_freq * SCHED_CAPACITY_SCALE, base_freq);
2089 	if (!turbo_ratio) {
2090 		pr_debug("Non-zero turbo and base frequencies led to a 0 ratio.\n");
2091 		return false;
2092 	}
2093 
2094 	arch_turbo_freq_ratio = turbo_ratio;
2095 	arch_set_max_freq_ratio(turbo_disabled());
2096 
2097 	return true;
2098 }
2099 
2100 #ifdef CONFIG_ACPI_CPPC_LIB
2101 static bool amd_set_max_freq_ratio(void)
2102 {
2103 	struct cppc_perf_caps perf_caps;
2104 	u64 highest_perf, nominal_perf;
2105 	u64 perf_ratio;
2106 	int rc;
2107 
2108 	rc = cppc_get_perf_caps(0, &perf_caps);
2109 	if (rc) {
2110 		pr_debug("Could not retrieve perf counters (%d)\n", rc);
2111 		return false;
2112 	}
2113 
2114 	highest_perf = amd_get_highest_perf();
2115 	nominal_perf = perf_caps.nominal_perf;
2116 
2117 	if (!highest_perf || !nominal_perf) {
2118 		pr_debug("Could not retrieve highest or nominal performance\n");
2119 		return false;
2120 	}
2121 
2122 	perf_ratio = div_u64(highest_perf * SCHED_CAPACITY_SCALE, nominal_perf);
2123 	/* midpoint between max_boost and max_P */
2124 	perf_ratio = (perf_ratio + SCHED_CAPACITY_SCALE) >> 1;
2125 	if (!perf_ratio) {
2126 		pr_debug("Non-zero highest/nominal perf values led to a 0 ratio\n");
2127 		return false;
2128 	}
2129 
2130 	arch_turbo_freq_ratio = perf_ratio;
2131 	arch_set_max_freq_ratio(false);
2132 
2133 	return true;
2134 }
2135 #else
2136 static bool amd_set_max_freq_ratio(void)
2137 {
2138 	return false;
2139 }
2140 #endif
2141 
2142 static void init_counter_refs(void)
2143 {
2144 	u64 aperf, mperf;
2145 
2146 	rdmsrl(MSR_IA32_APERF, aperf);
2147 	rdmsrl(MSR_IA32_MPERF, mperf);
2148 
2149 	this_cpu_write(arch_prev_aperf, aperf);
2150 	this_cpu_write(arch_prev_mperf, mperf);
2151 }
2152 
2153 #ifdef CONFIG_PM_SLEEP
2154 static struct syscore_ops freq_invariance_syscore_ops = {
2155 	.resume = init_counter_refs,
2156 };
2157 
2158 static void register_freq_invariance_syscore_ops(void)
2159 {
2160 	/* Bail out if registered already. */
2161 	if (freq_invariance_syscore_ops.node.prev)
2162 		return;
2163 
2164 	register_syscore_ops(&freq_invariance_syscore_ops);
2165 }
2166 #else
2167 static inline void register_freq_invariance_syscore_ops(void) {}
2168 #endif
2169 
2170 static void init_freq_invariance(bool secondary, bool cppc_ready)
2171 {
2172 	bool ret = false;
2173 
2174 	if (!boot_cpu_has(X86_FEATURE_APERFMPERF))
2175 		return;
2176 
2177 	if (secondary) {
2178 		if (static_branch_likely(&arch_scale_freq_key)) {
2179 			init_counter_refs();
2180 		}
2181 		return;
2182 	}
2183 
2184 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
2185 		ret = intel_set_max_freq_ratio();
2186 	else if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD) {
2187 		if (!cppc_ready) {
2188 			return;
2189 		}
2190 		ret = amd_set_max_freq_ratio();
2191 	}
2192 
2193 	if (ret) {
2194 		init_counter_refs();
2195 		static_branch_enable(&arch_scale_freq_key);
2196 		register_freq_invariance_syscore_ops();
2197 		pr_info("Estimated ratio of average max frequency by base frequency (times 1024): %llu\n", arch_max_freq_ratio);
2198 	} else {
2199 		pr_debug("Couldn't determine max cpu frequency, necessary for scale-invariant accounting.\n");
2200 	}
2201 }
2202 
2203 #ifdef CONFIG_ACPI_CPPC_LIB
2204 static DEFINE_MUTEX(freq_invariance_lock);
2205 
2206 void init_freq_invariance_cppc(void)
2207 {
2208 	static bool secondary;
2209 
2210 	mutex_lock(&freq_invariance_lock);
2211 
2212 	init_freq_invariance(secondary, true);
2213 	secondary = true;
2214 
2215 	mutex_unlock(&freq_invariance_lock);
2216 }
2217 #endif
2218 
2219 static void disable_freq_invariance_workfn(struct work_struct *work)
2220 {
2221 	static_branch_disable(&arch_scale_freq_key);
2222 }
2223 
2224 static DECLARE_WORK(disable_freq_invariance_work,
2225 		    disable_freq_invariance_workfn);
2226 
2227 DEFINE_PER_CPU(unsigned long, arch_freq_scale) = SCHED_CAPACITY_SCALE;
2228 
2229 void arch_scale_freq_tick(void)
2230 {
2231 	u64 freq_scale;
2232 	u64 aperf, mperf;
2233 	u64 acnt, mcnt;
2234 
2235 	if (!arch_scale_freq_invariant())
2236 		return;
2237 
2238 	rdmsrl(MSR_IA32_APERF, aperf);
2239 	rdmsrl(MSR_IA32_MPERF, mperf);
2240 
2241 	acnt = aperf - this_cpu_read(arch_prev_aperf);
2242 	mcnt = mperf - this_cpu_read(arch_prev_mperf);
2243 
2244 	this_cpu_write(arch_prev_aperf, aperf);
2245 	this_cpu_write(arch_prev_mperf, mperf);
2246 
2247 	if (check_shl_overflow(acnt, 2*SCHED_CAPACITY_SHIFT, &acnt))
2248 		goto error;
2249 
2250 	if (check_mul_overflow(mcnt, arch_max_freq_ratio, &mcnt) || !mcnt)
2251 		goto error;
2252 
2253 	freq_scale = div64_u64(acnt, mcnt);
2254 	if (!freq_scale)
2255 		goto error;
2256 
2257 	if (freq_scale > SCHED_CAPACITY_SCALE)
2258 		freq_scale = SCHED_CAPACITY_SCALE;
2259 
2260 	this_cpu_write(arch_freq_scale, freq_scale);
2261 	return;
2262 
2263 error:
2264 	pr_warn("Scheduler frequency invariance went wobbly, disabling!\n");
2265 	schedule_work(&disable_freq_invariance_work);
2266 }
2267 #else
2268 static inline void init_freq_invariance(bool secondary, bool cppc_ready)
2269 {
2270 }
2271 #endif /* CONFIG_X86_64 */
2272