xref: /openbmc/linux/arch/x86/kernel/smpboot.c (revision c0c45238)
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 /* representing HT siblings of each logical CPU */
87 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
88 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
89 
90 /* representing HT and core siblings of each logical CPU */
91 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
92 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
93 
94 /* representing HT, core, and die siblings of each logical CPU */
95 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_die_map);
96 EXPORT_PER_CPU_SYMBOL(cpu_die_map);
97 
98 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_llc_shared_map);
99 
100 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_l2c_shared_map);
101 
102 /* Per CPU bogomips and other parameters */
103 DEFINE_PER_CPU_READ_MOSTLY(struct cpuinfo_x86, cpu_info);
104 EXPORT_PER_CPU_SYMBOL(cpu_info);
105 
106 /* Logical package management. We might want to allocate that dynamically */
107 unsigned int __max_logical_packages __read_mostly;
108 EXPORT_SYMBOL(__max_logical_packages);
109 static unsigned int logical_packages __read_mostly;
110 static unsigned int logical_die __read_mostly;
111 
112 /* Maximum number of SMT threads on any online core */
113 int __read_mostly __max_smt_threads = 1;
114 
115 /* Flag to indicate if a complete sched domain rebuild is required */
116 bool x86_topology_update;
117 
118 int arch_update_cpu_topology(void)
119 {
120 	int retval = x86_topology_update;
121 
122 	x86_topology_update = false;
123 	return retval;
124 }
125 
126 static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
127 {
128 	unsigned long flags;
129 
130 	spin_lock_irqsave(&rtc_lock, flags);
131 	CMOS_WRITE(0xa, 0xf);
132 	spin_unlock_irqrestore(&rtc_lock, flags);
133 	*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) =
134 							start_eip >> 4;
135 	*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) =
136 							start_eip & 0xf;
137 }
138 
139 static inline void smpboot_restore_warm_reset_vector(void)
140 {
141 	unsigned long flags;
142 
143 	/*
144 	 * Paranoid:  Set warm reset code and vector here back
145 	 * to default values.
146 	 */
147 	spin_lock_irqsave(&rtc_lock, flags);
148 	CMOS_WRITE(0, 0xf);
149 	spin_unlock_irqrestore(&rtc_lock, flags);
150 
151 	*((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
152 }
153 
154 /*
155  * Report back to the Boot Processor during boot time or to the caller processor
156  * during CPU online.
157  */
158 static void smp_callin(void)
159 {
160 	int cpuid;
161 
162 	/*
163 	 * If waken up by an INIT in an 82489DX configuration
164 	 * cpu_callout_mask guarantees we don't get here before
165 	 * an INIT_deassert IPI reaches our local APIC, so it is
166 	 * now safe to touch our local APIC.
167 	 */
168 	cpuid = smp_processor_id();
169 
170 	/*
171 	 * the boot CPU has finished the init stage and is spinning
172 	 * on callin_map until we finish. We are free to set up this
173 	 * CPU, first the APIC. (this is probably redundant on most
174 	 * boards)
175 	 */
176 	apic_ap_setup();
177 
178 	/*
179 	 * Save our processor parameters. Note: this information
180 	 * is needed for clock calibration.
181 	 */
182 	smp_store_cpu_info(cpuid);
183 
184 	/*
185 	 * The topology information must be up to date before
186 	 * calibrate_delay() and notify_cpu_starting().
187 	 */
188 	set_cpu_sibling_map(raw_smp_processor_id());
189 
190 	init_freq_invariance(true, false);
191 
192 	/*
193 	 * Get our bogomips.
194 	 * Update loops_per_jiffy in cpu_data. Previous call to
195 	 * smp_store_cpu_info() stored a value that is close but not as
196 	 * accurate as the value just calculated.
197 	 */
198 	calibrate_delay();
199 	cpu_data(cpuid).loops_per_jiffy = loops_per_jiffy;
200 	pr_debug("Stack at about %p\n", &cpuid);
201 
202 	wmb();
203 
204 	notify_cpu_starting(cpuid);
205 
206 	/*
207 	 * Allow the master to continue.
208 	 */
209 	cpumask_set_cpu(cpuid, cpu_callin_mask);
210 }
211 
212 static int cpu0_logical_apicid;
213 static int enable_start_cpu0;
214 /*
215  * Activate a secondary processor.
216  */
217 static void notrace start_secondary(void *unused)
218 {
219 	/*
220 	 * Don't put *anything* except direct CPU state initialization
221 	 * before cpu_init(), SMP booting is too fragile that we want to
222 	 * limit the things done here to the most necessary things.
223 	 */
224 	cr4_init();
225 
226 #ifdef CONFIG_X86_32
227 	/* switch away from the initial page table */
228 	load_cr3(swapper_pg_dir);
229 	__flush_tlb_all();
230 #endif
231 	cpu_init_secondary();
232 	rcu_cpu_starting(raw_smp_processor_id());
233 	x86_cpuinit.early_percpu_clock_init();
234 	smp_callin();
235 
236 	enable_start_cpu0 = 0;
237 
238 	/* otherwise gcc will move up smp_processor_id before the cpu_init */
239 	barrier();
240 	/*
241 	 * Check TSC synchronization with the boot CPU:
242 	 */
243 	check_tsc_sync_target();
244 
245 	speculative_store_bypass_ht_init();
246 
247 	/*
248 	 * Lock vector_lock, set CPU online and bring the vector
249 	 * allocator online. Online must be set with vector_lock held
250 	 * to prevent a concurrent irq setup/teardown from seeing a
251 	 * half valid vector space.
252 	 */
253 	lock_vector_lock();
254 	set_cpu_online(smp_processor_id(), true);
255 	lapic_online();
256 	unlock_vector_lock();
257 	cpu_set_state_online(smp_processor_id());
258 	x86_platform.nmi_init();
259 
260 	/* enable local interrupts */
261 	local_irq_enable();
262 
263 	x86_cpuinit.setup_percpu_clockev();
264 
265 	wmb();
266 	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
267 }
268 
269 /**
270  * topology_is_primary_thread - Check whether CPU is the primary SMT thread
271  * @cpu:	CPU to check
272  */
273 bool topology_is_primary_thread(unsigned int cpu)
274 {
275 	return apic_id_is_primary_thread(per_cpu(x86_cpu_to_apicid, cpu));
276 }
277 
278 /**
279  * topology_smt_supported - Check whether SMT is supported by the CPUs
280  */
281 bool topology_smt_supported(void)
282 {
283 	return smp_num_siblings > 1;
284 }
285 
286 /**
287  * topology_phys_to_logical_pkg - Map a physical package id to a logical
288  *
289  * Returns logical package id or -1 if not found
290  */
291 int topology_phys_to_logical_pkg(unsigned int phys_pkg)
292 {
293 	int cpu;
294 
295 	for_each_possible_cpu(cpu) {
296 		struct cpuinfo_x86 *c = &cpu_data(cpu);
297 
298 		if (c->initialized && c->phys_proc_id == phys_pkg)
299 			return c->logical_proc_id;
300 	}
301 	return -1;
302 }
303 EXPORT_SYMBOL(topology_phys_to_logical_pkg);
304 /**
305  * topology_phys_to_logical_die - Map a physical die id to logical
306  *
307  * Returns logical die id or -1 if not found
308  */
309 int topology_phys_to_logical_die(unsigned int die_id, unsigned int cur_cpu)
310 {
311 	int cpu;
312 	int proc_id = cpu_data(cur_cpu).phys_proc_id;
313 
314 	for_each_possible_cpu(cpu) {
315 		struct cpuinfo_x86 *c = &cpu_data(cpu);
316 
317 		if (c->initialized && c->cpu_die_id == die_id &&
318 		    c->phys_proc_id == proc_id)
319 			return c->logical_die_id;
320 	}
321 	return -1;
322 }
323 EXPORT_SYMBOL(topology_phys_to_logical_die);
324 
325 /**
326  * topology_update_package_map - Update the physical to logical package map
327  * @pkg:	The physical package id as retrieved via CPUID
328  * @cpu:	The cpu for which this is updated
329  */
330 int topology_update_package_map(unsigned int pkg, unsigned int cpu)
331 {
332 	int new;
333 
334 	/* Already available somewhere? */
335 	new = topology_phys_to_logical_pkg(pkg);
336 	if (new >= 0)
337 		goto found;
338 
339 	new = logical_packages++;
340 	if (new != pkg) {
341 		pr_info("CPU %u Converting physical %u to logical package %u\n",
342 			cpu, pkg, new);
343 	}
344 found:
345 	cpu_data(cpu).logical_proc_id = new;
346 	return 0;
347 }
348 /**
349  * topology_update_die_map - Update the physical to logical die map
350  * @die:	The die id as retrieved via CPUID
351  * @cpu:	The cpu for which this is updated
352  */
353 int topology_update_die_map(unsigned int die, unsigned int cpu)
354 {
355 	int new;
356 
357 	/* Already available somewhere? */
358 	new = topology_phys_to_logical_die(die, cpu);
359 	if (new >= 0)
360 		goto found;
361 
362 	new = logical_die++;
363 	if (new != die) {
364 		pr_info("CPU %u Converting physical %u to logical die %u\n",
365 			cpu, die, new);
366 	}
367 found:
368 	cpu_data(cpu).logical_die_id = new;
369 	return 0;
370 }
371 
372 void __init smp_store_boot_cpu_info(void)
373 {
374 	int id = 0; /* CPU 0 */
375 	struct cpuinfo_x86 *c = &cpu_data(id);
376 
377 	*c = boot_cpu_data;
378 	c->cpu_index = id;
379 	topology_update_package_map(c->phys_proc_id, id);
380 	topology_update_die_map(c->cpu_die_id, id);
381 	c->initialized = true;
382 }
383 
384 /*
385  * The bootstrap kernel entry code has set these up. Save them for
386  * a given CPU
387  */
388 void smp_store_cpu_info(int id)
389 {
390 	struct cpuinfo_x86 *c = &cpu_data(id);
391 
392 	/* Copy boot_cpu_data only on the first bringup */
393 	if (!c->initialized)
394 		*c = boot_cpu_data;
395 	c->cpu_index = id;
396 	/*
397 	 * During boot time, CPU0 has this setup already. Save the info when
398 	 * bringing up AP or offlined CPU0.
399 	 */
400 	identify_secondary_cpu(c);
401 	c->initialized = true;
402 }
403 
404 static bool
405 topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
406 {
407 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
408 
409 	return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
410 }
411 
412 static bool
413 topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
414 {
415 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
416 
417 	return !WARN_ONCE(!topology_same_node(c, o),
418 		"sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
419 		"[node: %d != %d]. Ignoring dependency.\n",
420 		cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
421 }
422 
423 #define link_mask(mfunc, c1, c2)					\
424 do {									\
425 	cpumask_set_cpu((c1), mfunc(c2));				\
426 	cpumask_set_cpu((c2), mfunc(c1));				\
427 } while (0)
428 
429 static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
430 {
431 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
432 		int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
433 
434 		if (c->phys_proc_id == o->phys_proc_id &&
435 		    c->cpu_die_id == o->cpu_die_id &&
436 		    per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2)) {
437 			if (c->cpu_core_id == o->cpu_core_id)
438 				return topology_sane(c, o, "smt");
439 
440 			if ((c->cu_id != 0xff) &&
441 			    (o->cu_id != 0xff) &&
442 			    (c->cu_id == o->cu_id))
443 				return topology_sane(c, o, "smt");
444 		}
445 
446 	} else if (c->phys_proc_id == o->phys_proc_id &&
447 		   c->cpu_die_id == o->cpu_die_id &&
448 		   c->cpu_core_id == o->cpu_core_id) {
449 		return topology_sane(c, o, "smt");
450 	}
451 
452 	return false;
453 }
454 
455 static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
456 {
457 	if (c->phys_proc_id == o->phys_proc_id &&
458 	    c->cpu_die_id == o->cpu_die_id)
459 		return true;
460 	return false;
461 }
462 
463 static bool match_l2c(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
464 {
465 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
466 
467 	/* If the arch didn't set up l2c_id, fall back to SMT */
468 	if (per_cpu(cpu_l2c_id, cpu1) == BAD_APICID)
469 		return match_smt(c, o);
470 
471 	/* Do not match if L2 cache id does not match: */
472 	if (per_cpu(cpu_l2c_id, cpu1) != per_cpu(cpu_l2c_id, cpu2))
473 		return false;
474 
475 	return topology_sane(c, o, "l2c");
476 }
477 
478 /*
479  * Unlike the other levels, we do not enforce keeping a
480  * multicore group inside a NUMA node.  If this happens, we will
481  * discard the MC level of the topology later.
482  */
483 static bool match_pkg(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
484 {
485 	if (c->phys_proc_id == o->phys_proc_id)
486 		return true;
487 	return false;
488 }
489 
490 /*
491  * Define intel_cod_cpu[] for Intel COD (Cluster-on-Die) CPUs.
492  *
493  * Any Intel CPU that has multiple nodes per package and does not
494  * match intel_cod_cpu[] has the SNC (Sub-NUMA Cluster) topology.
495  *
496  * When in SNC mode, these CPUs enumerate an LLC that is shared
497  * by multiple NUMA nodes. The LLC is shared for off-package data
498  * access but private to the NUMA node (half of the package) for
499  * on-package access. CPUID (the source of the information about
500  * the LLC) can only enumerate the cache as shared or unshared,
501  * but not this particular configuration.
502  */
503 
504 static const struct x86_cpu_id intel_cod_cpu[] = {
505 	X86_MATCH_INTEL_FAM6_MODEL(HASWELL_X, 0),	/* COD */
506 	X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_X, 0),	/* COD */
507 	X86_MATCH_INTEL_FAM6_MODEL(ANY, 1),		/* SNC */
508 	{}
509 };
510 
511 static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
512 {
513 	const struct x86_cpu_id *id = x86_match_cpu(intel_cod_cpu);
514 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
515 	bool intel_snc = id && id->driver_data;
516 
517 	/* Do not match if we do not have a valid APICID for cpu: */
518 	if (per_cpu(cpu_llc_id, cpu1) == BAD_APICID)
519 		return false;
520 
521 	/* Do not match if LLC id does not match: */
522 	if (per_cpu(cpu_llc_id, cpu1) != per_cpu(cpu_llc_id, cpu2))
523 		return false;
524 
525 	/*
526 	 * Allow the SNC topology without warning. Return of false
527 	 * means 'c' does not share the LLC of 'o'. This will be
528 	 * reflected to userspace.
529 	 */
530 	if (match_pkg(c, o) && !topology_same_node(c, o) && intel_snc)
531 		return false;
532 
533 	return topology_sane(c, o, "llc");
534 }
535 
536 
537 #if defined(CONFIG_SCHED_SMT) || defined(CONFIG_SCHED_CLUSTER) || defined(CONFIG_SCHED_MC)
538 static inline int x86_sched_itmt_flags(void)
539 {
540 	return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
541 }
542 
543 #ifdef CONFIG_SCHED_MC
544 static int x86_core_flags(void)
545 {
546 	return cpu_core_flags() | x86_sched_itmt_flags();
547 }
548 #endif
549 #ifdef CONFIG_SCHED_SMT
550 static int x86_smt_flags(void)
551 {
552 	return cpu_smt_flags() | x86_sched_itmt_flags();
553 }
554 #endif
555 #ifdef CONFIG_SCHED_CLUSTER
556 static int x86_cluster_flags(void)
557 {
558 	return cpu_cluster_flags() | x86_sched_itmt_flags();
559 }
560 #endif
561 #endif
562 
563 static struct sched_domain_topology_level x86_numa_in_package_topology[] = {
564 #ifdef CONFIG_SCHED_SMT
565 	{ cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
566 #endif
567 #ifdef CONFIG_SCHED_CLUSTER
568 	{ cpu_clustergroup_mask, x86_cluster_flags, SD_INIT_NAME(CLS) },
569 #endif
570 #ifdef CONFIG_SCHED_MC
571 	{ cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
572 #endif
573 	{ NULL, },
574 };
575 
576 static struct sched_domain_topology_level x86_hybrid_topology[] = {
577 #ifdef CONFIG_SCHED_SMT
578 	{ cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
579 #endif
580 #ifdef CONFIG_SCHED_MC
581 	{ cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
582 #endif
583 	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
584 	{ NULL, },
585 };
586 
587 static struct sched_domain_topology_level x86_topology[] = {
588 #ifdef CONFIG_SCHED_SMT
589 	{ cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
590 #endif
591 #ifdef CONFIG_SCHED_CLUSTER
592 	{ cpu_clustergroup_mask, x86_cluster_flags, SD_INIT_NAME(CLS) },
593 #endif
594 #ifdef CONFIG_SCHED_MC
595 	{ cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
596 #endif
597 	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
598 	{ NULL, },
599 };
600 
601 /*
602  * Set if a package/die has multiple NUMA nodes inside.
603  * AMD Magny-Cours, Intel Cluster-on-Die, and Intel
604  * Sub-NUMA Clustering have this.
605  */
606 static bool x86_has_numa_in_package;
607 
608 void set_cpu_sibling_map(int cpu)
609 {
610 	bool has_smt = smp_num_siblings > 1;
611 	bool has_mp = has_smt || boot_cpu_data.x86_max_cores > 1;
612 	struct cpuinfo_x86 *c = &cpu_data(cpu);
613 	struct cpuinfo_x86 *o;
614 	int i, threads;
615 
616 	cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
617 
618 	if (!has_mp) {
619 		cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
620 		cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
621 		cpumask_set_cpu(cpu, cpu_l2c_shared_mask(cpu));
622 		cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
623 		cpumask_set_cpu(cpu, topology_die_cpumask(cpu));
624 		c->booted_cores = 1;
625 		return;
626 	}
627 
628 	for_each_cpu(i, cpu_sibling_setup_mask) {
629 		o = &cpu_data(i);
630 
631 		if (match_pkg(c, o) && !topology_same_node(c, o))
632 			x86_has_numa_in_package = true;
633 
634 		if ((i == cpu) || (has_smt && match_smt(c, o)))
635 			link_mask(topology_sibling_cpumask, cpu, i);
636 
637 		if ((i == cpu) || (has_mp && match_llc(c, o)))
638 			link_mask(cpu_llc_shared_mask, cpu, i);
639 
640 		if ((i == cpu) || (has_mp && match_l2c(c, o)))
641 			link_mask(cpu_l2c_shared_mask, cpu, i);
642 
643 		if ((i == cpu) || (has_mp && match_die(c, o)))
644 			link_mask(topology_die_cpumask, cpu, i);
645 	}
646 
647 	threads = cpumask_weight(topology_sibling_cpumask(cpu));
648 	if (threads > __max_smt_threads)
649 		__max_smt_threads = threads;
650 
651 	for_each_cpu(i, topology_sibling_cpumask(cpu))
652 		cpu_data(i).smt_active = threads > 1;
653 
654 	/*
655 	 * This needs a separate iteration over the cpus because we rely on all
656 	 * topology_sibling_cpumask links to be set-up.
657 	 */
658 	for_each_cpu(i, cpu_sibling_setup_mask) {
659 		o = &cpu_data(i);
660 
661 		if ((i == cpu) || (has_mp && match_pkg(c, o))) {
662 			link_mask(topology_core_cpumask, cpu, i);
663 
664 			/*
665 			 *  Does this new cpu bringup a new core?
666 			 */
667 			if (threads == 1) {
668 				/*
669 				 * for each core in package, increment
670 				 * the booted_cores for this new cpu
671 				 */
672 				if (cpumask_first(
673 				    topology_sibling_cpumask(i)) == i)
674 					c->booted_cores++;
675 				/*
676 				 * increment the core count for all
677 				 * the other cpus in this package
678 				 */
679 				if (i != cpu)
680 					cpu_data(i).booted_cores++;
681 			} else if (i != cpu && !c->booted_cores)
682 				c->booted_cores = cpu_data(i).booted_cores;
683 		}
684 	}
685 }
686 
687 /* maps the cpu to the sched domain representing multi-core */
688 const struct cpumask *cpu_coregroup_mask(int cpu)
689 {
690 	return cpu_llc_shared_mask(cpu);
691 }
692 
693 const struct cpumask *cpu_clustergroup_mask(int cpu)
694 {
695 	return cpu_l2c_shared_mask(cpu);
696 }
697 
698 static void impress_friends(void)
699 {
700 	int cpu;
701 	unsigned long bogosum = 0;
702 	/*
703 	 * Allow the user to impress friends.
704 	 */
705 	pr_debug("Before bogomips\n");
706 	for_each_possible_cpu(cpu)
707 		if (cpumask_test_cpu(cpu, cpu_callout_mask))
708 			bogosum += cpu_data(cpu).loops_per_jiffy;
709 	pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
710 		num_online_cpus(),
711 		bogosum/(500000/HZ),
712 		(bogosum/(5000/HZ))%100);
713 
714 	pr_debug("Before bogocount - setting activated=1\n");
715 }
716 
717 void __inquire_remote_apic(int apicid)
718 {
719 	unsigned i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
720 	const char * const names[] = { "ID", "VERSION", "SPIV" };
721 	int timeout;
722 	u32 status;
723 
724 	pr_info("Inquiring remote APIC 0x%x...\n", apicid);
725 
726 	for (i = 0; i < ARRAY_SIZE(regs); i++) {
727 		pr_info("... APIC 0x%x %s: ", apicid, names[i]);
728 
729 		/*
730 		 * Wait for idle.
731 		 */
732 		status = safe_apic_wait_icr_idle();
733 		if (status)
734 			pr_cont("a previous APIC delivery may have failed\n");
735 
736 		apic_icr_write(APIC_DM_REMRD | regs[i], apicid);
737 
738 		timeout = 0;
739 		do {
740 			udelay(100);
741 			status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
742 		} while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
743 
744 		switch (status) {
745 		case APIC_ICR_RR_VALID:
746 			status = apic_read(APIC_RRR);
747 			pr_cont("%08x\n", status);
748 			break;
749 		default:
750 			pr_cont("failed\n");
751 		}
752 	}
753 }
754 
755 /*
756  * The Multiprocessor Specification 1.4 (1997) example code suggests
757  * that there should be a 10ms delay between the BSP asserting INIT
758  * and de-asserting INIT, when starting a remote processor.
759  * But that slows boot and resume on modern processors, which include
760  * many cores and don't require that delay.
761  *
762  * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
763  * Modern processor families are quirked to remove the delay entirely.
764  */
765 #define UDELAY_10MS_DEFAULT 10000
766 
767 static unsigned int init_udelay = UINT_MAX;
768 
769 static int __init cpu_init_udelay(char *str)
770 {
771 	get_option(&str, &init_udelay);
772 
773 	return 0;
774 }
775 early_param("cpu_init_udelay", cpu_init_udelay);
776 
777 static void __init smp_quirk_init_udelay(void)
778 {
779 	/* if cmdline changed it from default, leave it alone */
780 	if (init_udelay != UINT_MAX)
781 		return;
782 
783 	/* if modern processor, use no delay */
784 	if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
785 	    ((boot_cpu_data.x86_vendor == X86_VENDOR_HYGON) && (boot_cpu_data.x86 >= 0x18)) ||
786 	    ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
787 		init_udelay = 0;
788 		return;
789 	}
790 	/* else, use legacy delay */
791 	init_udelay = UDELAY_10MS_DEFAULT;
792 }
793 
794 /*
795  * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
796  * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
797  * won't ... remember to clear down the APIC, etc later.
798  */
799 int
800 wakeup_secondary_cpu_via_nmi(int apicid, unsigned long start_eip)
801 {
802 	u32 dm = apic->dest_mode_logical ? APIC_DEST_LOGICAL : APIC_DEST_PHYSICAL;
803 	unsigned long send_status, accept_status = 0;
804 	int maxlvt;
805 
806 	/* Target chip */
807 	/* Boot on the stack */
808 	/* Kick the second */
809 	apic_icr_write(APIC_DM_NMI | dm, apicid);
810 
811 	pr_debug("Waiting for send to finish...\n");
812 	send_status = safe_apic_wait_icr_idle();
813 
814 	/*
815 	 * Give the other CPU some time to accept the IPI.
816 	 */
817 	udelay(200);
818 	if (APIC_INTEGRATED(boot_cpu_apic_version)) {
819 		maxlvt = lapic_get_maxlvt();
820 		if (maxlvt > 3)			/* Due to the Pentium erratum 3AP.  */
821 			apic_write(APIC_ESR, 0);
822 		accept_status = (apic_read(APIC_ESR) & 0xEF);
823 	}
824 	pr_debug("NMI sent\n");
825 
826 	if (send_status)
827 		pr_err("APIC never delivered???\n");
828 	if (accept_status)
829 		pr_err("APIC delivery error (%lx)\n", accept_status);
830 
831 	return (send_status | accept_status);
832 }
833 
834 static int
835 wakeup_secondary_cpu_via_init(int phys_apicid, unsigned long start_eip)
836 {
837 	unsigned long send_status = 0, accept_status = 0;
838 	int maxlvt, num_starts, j;
839 
840 	maxlvt = lapic_get_maxlvt();
841 
842 	/*
843 	 * Be paranoid about clearing APIC errors.
844 	 */
845 	if (APIC_INTEGRATED(boot_cpu_apic_version)) {
846 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
847 			apic_write(APIC_ESR, 0);
848 		apic_read(APIC_ESR);
849 	}
850 
851 	pr_debug("Asserting INIT\n");
852 
853 	/*
854 	 * Turn INIT on target chip
855 	 */
856 	/*
857 	 * Send IPI
858 	 */
859 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT,
860 		       phys_apicid);
861 
862 	pr_debug("Waiting for send to finish...\n");
863 	send_status = safe_apic_wait_icr_idle();
864 
865 	udelay(init_udelay);
866 
867 	pr_debug("Deasserting INIT\n");
868 
869 	/* Target chip */
870 	/* Send IPI */
871 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
872 
873 	pr_debug("Waiting for send to finish...\n");
874 	send_status = safe_apic_wait_icr_idle();
875 
876 	mb();
877 
878 	/*
879 	 * Should we send STARTUP IPIs ?
880 	 *
881 	 * Determine this based on the APIC version.
882 	 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
883 	 */
884 	if (APIC_INTEGRATED(boot_cpu_apic_version))
885 		num_starts = 2;
886 	else
887 		num_starts = 0;
888 
889 	/*
890 	 * Run STARTUP IPI loop.
891 	 */
892 	pr_debug("#startup loops: %d\n", num_starts);
893 
894 	for (j = 1; j <= num_starts; j++) {
895 		pr_debug("Sending STARTUP #%d\n", j);
896 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
897 			apic_write(APIC_ESR, 0);
898 		apic_read(APIC_ESR);
899 		pr_debug("After apic_write\n");
900 
901 		/*
902 		 * STARTUP IPI
903 		 */
904 
905 		/* Target chip */
906 		/* Boot on the stack */
907 		/* Kick the second */
908 		apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
909 			       phys_apicid);
910 
911 		/*
912 		 * Give the other CPU some time to accept the IPI.
913 		 */
914 		if (init_udelay == 0)
915 			udelay(10);
916 		else
917 			udelay(300);
918 
919 		pr_debug("Startup point 1\n");
920 
921 		pr_debug("Waiting for send to finish...\n");
922 		send_status = safe_apic_wait_icr_idle();
923 
924 		/*
925 		 * Give the other CPU some time to accept the IPI.
926 		 */
927 		if (init_udelay == 0)
928 			udelay(10);
929 		else
930 			udelay(200);
931 
932 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
933 			apic_write(APIC_ESR, 0);
934 		accept_status = (apic_read(APIC_ESR) & 0xEF);
935 		if (send_status || accept_status)
936 			break;
937 	}
938 	pr_debug("After Startup\n");
939 
940 	if (send_status)
941 		pr_err("APIC never delivered???\n");
942 	if (accept_status)
943 		pr_err("APIC delivery error (%lx)\n", accept_status);
944 
945 	return (send_status | accept_status);
946 }
947 
948 /* reduce the number of lines printed when booting a large cpu count system */
949 static void announce_cpu(int cpu, int apicid)
950 {
951 	static int current_node = NUMA_NO_NODE;
952 	int node = early_cpu_to_node(cpu);
953 	static int width, node_width;
954 
955 	if (!width)
956 		width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
957 
958 	if (!node_width)
959 		node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
960 
961 	if (cpu == 1)
962 		printk(KERN_INFO "x86: Booting SMP configuration:\n");
963 
964 	if (system_state < SYSTEM_RUNNING) {
965 		if (node != current_node) {
966 			if (current_node > (-1))
967 				pr_cont("\n");
968 			current_node = node;
969 
970 			printk(KERN_INFO ".... node %*s#%d, CPUs:  ",
971 			       node_width - num_digits(node), " ", node);
972 		}
973 
974 		/* Add padding for the BSP */
975 		if (cpu == 1)
976 			pr_cont("%*s", width + 1, " ");
977 
978 		pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
979 
980 	} else
981 		pr_info("Booting Node %d Processor %d APIC 0x%x\n",
982 			node, cpu, apicid);
983 }
984 
985 static int wakeup_cpu0_nmi(unsigned int cmd, struct pt_regs *regs)
986 {
987 	int cpu;
988 
989 	cpu = smp_processor_id();
990 	if (cpu == 0 && !cpu_online(cpu) && enable_start_cpu0)
991 		return NMI_HANDLED;
992 
993 	return NMI_DONE;
994 }
995 
996 /*
997  * Wake up AP by INIT, INIT, STARTUP sequence.
998  *
999  * Instead of waiting for STARTUP after INITs, BSP will execute the BIOS
1000  * boot-strap code which is not a desired behavior for waking up BSP. To
1001  * void the boot-strap code, wake up CPU0 by NMI instead.
1002  *
1003  * This works to wake up soft offlined CPU0 only. If CPU0 is hard offlined
1004  * (i.e. physically hot removed and then hot added), NMI won't wake it up.
1005  * We'll change this code in the future to wake up hard offlined CPU0 if
1006  * real platform and request are available.
1007  */
1008 static int
1009 wakeup_cpu_via_init_nmi(int cpu, unsigned long start_ip, int apicid,
1010 	       int *cpu0_nmi_registered)
1011 {
1012 	int id;
1013 	int boot_error;
1014 
1015 	preempt_disable();
1016 
1017 	/*
1018 	 * Wake up AP by INIT, INIT, STARTUP sequence.
1019 	 */
1020 	if (cpu) {
1021 		boot_error = wakeup_secondary_cpu_via_init(apicid, start_ip);
1022 		goto out;
1023 	}
1024 
1025 	/*
1026 	 * Wake up BSP by nmi.
1027 	 *
1028 	 * Register a NMI handler to help wake up CPU0.
1029 	 */
1030 	boot_error = register_nmi_handler(NMI_LOCAL,
1031 					  wakeup_cpu0_nmi, 0, "wake_cpu0");
1032 
1033 	if (!boot_error) {
1034 		enable_start_cpu0 = 1;
1035 		*cpu0_nmi_registered = 1;
1036 		id = apic->dest_mode_logical ? cpu0_logical_apicid : apicid;
1037 		boot_error = wakeup_secondary_cpu_via_nmi(id, start_ip);
1038 	}
1039 
1040 out:
1041 	preempt_enable();
1042 
1043 	return boot_error;
1044 }
1045 
1046 int common_cpu_up(unsigned int cpu, struct task_struct *idle)
1047 {
1048 	int ret;
1049 
1050 	/* Just in case we booted with a single CPU. */
1051 	alternatives_enable_smp();
1052 
1053 	per_cpu(current_task, cpu) = idle;
1054 	cpu_init_stack_canary(cpu, idle);
1055 
1056 	/* Initialize the interrupt stack(s) */
1057 	ret = irq_init_percpu_irqstack(cpu);
1058 	if (ret)
1059 		return ret;
1060 
1061 #ifdef CONFIG_X86_32
1062 	/* Stack for startup_32 can be just as for start_secondary onwards */
1063 	per_cpu(cpu_current_top_of_stack, cpu) = task_top_of_stack(idle);
1064 #else
1065 	initial_gs = per_cpu_offset(cpu);
1066 #endif
1067 	return 0;
1068 }
1069 
1070 /*
1071  * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
1072  * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
1073  * Returns zero if CPU booted OK, else error code from
1074  * ->wakeup_secondary_cpu.
1075  */
1076 static int do_boot_cpu(int apicid, int cpu, struct task_struct *idle,
1077 		       int *cpu0_nmi_registered)
1078 {
1079 	/* start_ip had better be page-aligned! */
1080 	unsigned long start_ip = real_mode_header->trampoline_start;
1081 
1082 	unsigned long boot_error = 0;
1083 	unsigned long timeout;
1084 
1085 	idle->thread.sp = (unsigned long)task_pt_regs(idle);
1086 	early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
1087 	initial_code = (unsigned long)start_secondary;
1088 	initial_stack  = idle->thread.sp;
1089 
1090 	/* Enable the espfix hack for this CPU */
1091 	init_espfix_ap(cpu);
1092 
1093 	/* So we see what's up */
1094 	announce_cpu(cpu, apicid);
1095 
1096 	/*
1097 	 * This grunge runs the startup process for
1098 	 * the targeted processor.
1099 	 */
1100 
1101 	if (x86_platform.legacy.warm_reset) {
1102 
1103 		pr_debug("Setting warm reset code and vector.\n");
1104 
1105 		smpboot_setup_warm_reset_vector(start_ip);
1106 		/*
1107 		 * Be paranoid about clearing APIC errors.
1108 		*/
1109 		if (APIC_INTEGRATED(boot_cpu_apic_version)) {
1110 			apic_write(APIC_ESR, 0);
1111 			apic_read(APIC_ESR);
1112 		}
1113 	}
1114 
1115 	/*
1116 	 * AP might wait on cpu_callout_mask in cpu_init() with
1117 	 * cpu_initialized_mask set if previous attempt to online
1118 	 * it timed-out. Clear cpu_initialized_mask so that after
1119 	 * INIT/SIPI it could start with a clean state.
1120 	 */
1121 	cpumask_clear_cpu(cpu, cpu_initialized_mask);
1122 	smp_mb();
1123 
1124 	/*
1125 	 * Wake up a CPU in difference cases:
1126 	 * - Use the method in the APIC driver if it's defined
1127 	 * Otherwise,
1128 	 * - Use an INIT boot APIC message for APs or NMI for BSP.
1129 	 */
1130 	if (apic->wakeup_secondary_cpu)
1131 		boot_error = apic->wakeup_secondary_cpu(apicid, start_ip);
1132 	else
1133 		boot_error = wakeup_cpu_via_init_nmi(cpu, start_ip, apicid,
1134 						     cpu0_nmi_registered);
1135 
1136 	if (!boot_error) {
1137 		/*
1138 		 * Wait 10s total for first sign of life from AP
1139 		 */
1140 		boot_error = -1;
1141 		timeout = jiffies + 10*HZ;
1142 		while (time_before(jiffies, timeout)) {
1143 			if (cpumask_test_cpu(cpu, cpu_initialized_mask)) {
1144 				/*
1145 				 * Tell AP to proceed with initialization
1146 				 */
1147 				cpumask_set_cpu(cpu, cpu_callout_mask);
1148 				boot_error = 0;
1149 				break;
1150 			}
1151 			schedule();
1152 		}
1153 	}
1154 
1155 	if (!boot_error) {
1156 		/*
1157 		 * Wait till AP completes initial initialization
1158 		 */
1159 		while (!cpumask_test_cpu(cpu, cpu_callin_mask)) {
1160 			/*
1161 			 * Allow other tasks to run while we wait for the
1162 			 * AP to come online. This also gives a chance
1163 			 * for the MTRR work(triggered by the AP coming online)
1164 			 * to be completed in the stop machine context.
1165 			 */
1166 			schedule();
1167 		}
1168 	}
1169 
1170 	if (x86_platform.legacy.warm_reset) {
1171 		/*
1172 		 * Cleanup possible dangling ends...
1173 		 */
1174 		smpboot_restore_warm_reset_vector();
1175 	}
1176 
1177 	return boot_error;
1178 }
1179 
1180 int native_cpu_up(unsigned int cpu, struct task_struct *tidle)
1181 {
1182 	int apicid = apic->cpu_present_to_apicid(cpu);
1183 	int cpu0_nmi_registered = 0;
1184 	unsigned long flags;
1185 	int err, ret = 0;
1186 
1187 	lockdep_assert_irqs_enabled();
1188 
1189 	pr_debug("++++++++++++++++++++=_---CPU UP  %u\n", cpu);
1190 
1191 	if (apicid == BAD_APICID ||
1192 	    !physid_isset(apicid, phys_cpu_present_map) ||
1193 	    !apic->apic_id_valid(apicid)) {
1194 		pr_err("%s: bad cpu %d\n", __func__, cpu);
1195 		return -EINVAL;
1196 	}
1197 
1198 	/*
1199 	 * Already booted CPU?
1200 	 */
1201 	if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
1202 		pr_debug("do_boot_cpu %d Already started\n", cpu);
1203 		return -ENOSYS;
1204 	}
1205 
1206 	/*
1207 	 * Save current MTRR state in case it was changed since early boot
1208 	 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1209 	 */
1210 	mtrr_save_state();
1211 
1212 	/* x86 CPUs take themselves offline, so delayed offline is OK. */
1213 	err = cpu_check_up_prepare(cpu);
1214 	if (err && err != -EBUSY)
1215 		return err;
1216 
1217 	/* the FPU context is blank, nobody can own it */
1218 	per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
1219 
1220 	err = common_cpu_up(cpu, tidle);
1221 	if (err)
1222 		return err;
1223 
1224 	err = do_boot_cpu(apicid, cpu, tidle, &cpu0_nmi_registered);
1225 	if (err) {
1226 		pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
1227 		ret = -EIO;
1228 		goto unreg_nmi;
1229 	}
1230 
1231 	/*
1232 	 * Check TSC synchronization with the AP (keep irqs disabled
1233 	 * while doing so):
1234 	 */
1235 	local_irq_save(flags);
1236 	check_tsc_sync_source(cpu);
1237 	local_irq_restore(flags);
1238 
1239 	while (!cpu_online(cpu)) {
1240 		cpu_relax();
1241 		touch_nmi_watchdog();
1242 	}
1243 
1244 unreg_nmi:
1245 	/*
1246 	 * Clean up the nmi handler. Do this after the callin and callout sync
1247 	 * to avoid impact of possible long unregister time.
1248 	 */
1249 	if (cpu0_nmi_registered)
1250 		unregister_nmi_handler(NMI_LOCAL, "wake_cpu0");
1251 
1252 	return ret;
1253 }
1254 
1255 /**
1256  * arch_disable_smp_support() - disables SMP support for x86 at runtime
1257  */
1258 void arch_disable_smp_support(void)
1259 {
1260 	disable_ioapic_support();
1261 }
1262 
1263 /*
1264  * Fall back to non SMP mode after errors.
1265  *
1266  * RED-PEN audit/test this more. I bet there is more state messed up here.
1267  */
1268 static __init void disable_smp(void)
1269 {
1270 	pr_info("SMP disabled\n");
1271 
1272 	disable_ioapic_support();
1273 
1274 	init_cpu_present(cpumask_of(0));
1275 	init_cpu_possible(cpumask_of(0));
1276 
1277 	if (smp_found_config)
1278 		physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
1279 	else
1280 		physid_set_mask_of_physid(0, &phys_cpu_present_map);
1281 	cpumask_set_cpu(0, topology_sibling_cpumask(0));
1282 	cpumask_set_cpu(0, topology_core_cpumask(0));
1283 	cpumask_set_cpu(0, topology_die_cpumask(0));
1284 }
1285 
1286 /*
1287  * Various sanity checks.
1288  */
1289 static void __init smp_sanity_check(void)
1290 {
1291 	preempt_disable();
1292 
1293 #if !defined(CONFIG_X86_BIGSMP) && defined(CONFIG_X86_32)
1294 	if (def_to_bigsmp && nr_cpu_ids > 8) {
1295 		unsigned int cpu;
1296 		unsigned nr;
1297 
1298 		pr_warn("More than 8 CPUs detected - skipping them\n"
1299 			"Use CONFIG_X86_BIGSMP\n");
1300 
1301 		nr = 0;
1302 		for_each_present_cpu(cpu) {
1303 			if (nr >= 8)
1304 				set_cpu_present(cpu, false);
1305 			nr++;
1306 		}
1307 
1308 		nr = 0;
1309 		for_each_possible_cpu(cpu) {
1310 			if (nr >= 8)
1311 				set_cpu_possible(cpu, false);
1312 			nr++;
1313 		}
1314 
1315 		nr_cpu_ids = 8;
1316 	}
1317 #endif
1318 
1319 	if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map)) {
1320 		pr_warn("weird, boot CPU (#%d) not listed by the BIOS\n",
1321 			hard_smp_processor_id());
1322 
1323 		physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1324 	}
1325 
1326 	/*
1327 	 * Should not be necessary because the MP table should list the boot
1328 	 * CPU too, but we do it for the sake of robustness anyway.
1329 	 */
1330 	if (!apic->check_phys_apicid_present(boot_cpu_physical_apicid)) {
1331 		pr_notice("weird, boot CPU (#%d) not listed by the BIOS\n",
1332 			  boot_cpu_physical_apicid);
1333 		physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1334 	}
1335 	preempt_enable();
1336 }
1337 
1338 static void __init smp_cpu_index_default(void)
1339 {
1340 	int i;
1341 	struct cpuinfo_x86 *c;
1342 
1343 	for_each_possible_cpu(i) {
1344 		c = &cpu_data(i);
1345 		/* mark all to hotplug */
1346 		c->cpu_index = nr_cpu_ids;
1347 	}
1348 }
1349 
1350 static void __init smp_get_logical_apicid(void)
1351 {
1352 	if (x2apic_mode)
1353 		cpu0_logical_apicid = apic_read(APIC_LDR);
1354 	else
1355 		cpu0_logical_apicid = GET_APIC_LOGICAL_ID(apic_read(APIC_LDR));
1356 }
1357 
1358 void __init smp_prepare_cpus_common(void)
1359 {
1360 	unsigned int i;
1361 
1362 	smp_cpu_index_default();
1363 
1364 	/*
1365 	 * Setup boot CPU information
1366 	 */
1367 	smp_store_boot_cpu_info(); /* Final full version of the data */
1368 	cpumask_copy(cpu_callin_mask, cpumask_of(0));
1369 	mb();
1370 
1371 	for_each_possible_cpu(i) {
1372 		zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1373 		zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1374 		zalloc_cpumask_var(&per_cpu(cpu_die_map, i), GFP_KERNEL);
1375 		zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1376 		zalloc_cpumask_var(&per_cpu(cpu_l2c_shared_map, i), GFP_KERNEL);
1377 	}
1378 
1379 	/*
1380 	 * Set 'default' x86 topology, this matches default_topology() in that
1381 	 * it has NUMA nodes as a topology level. See also
1382 	 * native_smp_cpus_done().
1383 	 *
1384 	 * Must be done before set_cpus_sibling_map() is ran.
1385 	 */
1386 	set_sched_topology(x86_topology);
1387 
1388 	set_cpu_sibling_map(0);
1389 }
1390 
1391 /*
1392  * Prepare for SMP bootup.
1393  * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1394  *            for common interface support.
1395  */
1396 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1397 {
1398 	smp_prepare_cpus_common();
1399 
1400 	init_freq_invariance(false, false);
1401 	smp_sanity_check();
1402 
1403 	switch (apic_intr_mode) {
1404 	case APIC_PIC:
1405 	case APIC_VIRTUAL_WIRE_NO_CONFIG:
1406 		disable_smp();
1407 		return;
1408 	case APIC_SYMMETRIC_IO_NO_ROUTING:
1409 		disable_smp();
1410 		/* Setup local timer */
1411 		x86_init.timers.setup_percpu_clockev();
1412 		return;
1413 	case APIC_VIRTUAL_WIRE:
1414 	case APIC_SYMMETRIC_IO:
1415 		break;
1416 	}
1417 
1418 	/* Setup local timer */
1419 	x86_init.timers.setup_percpu_clockev();
1420 
1421 	smp_get_logical_apicid();
1422 
1423 	pr_info("CPU0: ");
1424 	print_cpu_info(&cpu_data(0));
1425 
1426 	uv_system_init();
1427 
1428 	set_mtrr_aps_delayed_init();
1429 
1430 	smp_quirk_init_udelay();
1431 
1432 	speculative_store_bypass_ht_init();
1433 }
1434 
1435 void arch_thaw_secondary_cpus_begin(void)
1436 {
1437 	set_mtrr_aps_delayed_init();
1438 }
1439 
1440 void arch_thaw_secondary_cpus_end(void)
1441 {
1442 	mtrr_aps_init();
1443 }
1444 
1445 /*
1446  * Early setup to make printk work.
1447  */
1448 void __init native_smp_prepare_boot_cpu(void)
1449 {
1450 	int me = smp_processor_id();
1451 	switch_to_new_gdt(me);
1452 	/* already set me in cpu_online_mask in boot_cpu_init() */
1453 	cpumask_set_cpu(me, cpu_callout_mask);
1454 	cpu_set_state_online(me);
1455 	native_pv_lock_init();
1456 }
1457 
1458 void __init calculate_max_logical_packages(void)
1459 {
1460 	int ncpus;
1461 
1462 	/*
1463 	 * Today neither Intel nor AMD support heterogeneous systems so
1464 	 * extrapolate the boot cpu's data to all packages.
1465 	 */
1466 	ncpus = cpu_data(0).booted_cores * topology_max_smt_threads();
1467 	__max_logical_packages = DIV_ROUND_UP(total_cpus, ncpus);
1468 	pr_info("Max logical packages: %u\n", __max_logical_packages);
1469 }
1470 
1471 void __init native_smp_cpus_done(unsigned int max_cpus)
1472 {
1473 	pr_debug("Boot done\n");
1474 
1475 	calculate_max_logical_packages();
1476 
1477 	/* XXX for now assume numa-in-package and hybrid don't overlap */
1478 	if (x86_has_numa_in_package)
1479 		set_sched_topology(x86_numa_in_package_topology);
1480 	if (cpu_feature_enabled(X86_FEATURE_HYBRID_CPU))
1481 		set_sched_topology(x86_hybrid_topology);
1482 
1483 	nmi_selftest();
1484 	impress_friends();
1485 	mtrr_aps_init();
1486 }
1487 
1488 static int __initdata setup_possible_cpus = -1;
1489 static int __init _setup_possible_cpus(char *str)
1490 {
1491 	get_option(&str, &setup_possible_cpus);
1492 	return 0;
1493 }
1494 early_param("possible_cpus", _setup_possible_cpus);
1495 
1496 
1497 /*
1498  * cpu_possible_mask should be static, it cannot change as cpu's
1499  * are onlined, or offlined. The reason is per-cpu data-structures
1500  * are allocated by some modules at init time, and don't expect to
1501  * do this dynamically on cpu arrival/departure.
1502  * cpu_present_mask on the other hand can change dynamically.
1503  * In case when cpu_hotplug is not compiled, then we resort to current
1504  * behaviour, which is cpu_possible == cpu_present.
1505  * - Ashok Raj
1506  *
1507  * Three ways to find out the number of additional hotplug CPUs:
1508  * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1509  * - The user can overwrite it with possible_cpus=NUM
1510  * - Otherwise don't reserve additional CPUs.
1511  * We do this because additional CPUs waste a lot of memory.
1512  * -AK
1513  */
1514 __init void prefill_possible_map(void)
1515 {
1516 	int i, possible;
1517 
1518 	/* No boot processor was found in mptable or ACPI MADT */
1519 	if (!num_processors) {
1520 		if (boot_cpu_has(X86_FEATURE_APIC)) {
1521 			int apicid = boot_cpu_physical_apicid;
1522 			int cpu = hard_smp_processor_id();
1523 
1524 			pr_warn("Boot CPU (id %d) not listed by BIOS\n", cpu);
1525 
1526 			/* Make sure boot cpu is enumerated */
1527 			if (apic->cpu_present_to_apicid(0) == BAD_APICID &&
1528 			    apic->apic_id_valid(apicid))
1529 				generic_processor_info(apicid, boot_cpu_apic_version);
1530 		}
1531 
1532 		if (!num_processors)
1533 			num_processors = 1;
1534 	}
1535 
1536 	i = setup_max_cpus ?: 1;
1537 	if (setup_possible_cpus == -1) {
1538 		possible = num_processors;
1539 #ifdef CONFIG_HOTPLUG_CPU
1540 		if (setup_max_cpus)
1541 			possible += disabled_cpus;
1542 #else
1543 		if (possible > i)
1544 			possible = i;
1545 #endif
1546 	} else
1547 		possible = setup_possible_cpus;
1548 
1549 	total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1550 
1551 	/* nr_cpu_ids could be reduced via nr_cpus= */
1552 	if (possible > nr_cpu_ids) {
1553 		pr_warn("%d Processors exceeds NR_CPUS limit of %u\n",
1554 			possible, nr_cpu_ids);
1555 		possible = nr_cpu_ids;
1556 	}
1557 
1558 #ifdef CONFIG_HOTPLUG_CPU
1559 	if (!setup_max_cpus)
1560 #endif
1561 	if (possible > i) {
1562 		pr_warn("%d Processors exceeds max_cpus limit of %u\n",
1563 			possible, setup_max_cpus);
1564 		possible = i;
1565 	}
1566 
1567 	nr_cpu_ids = possible;
1568 
1569 	pr_info("Allowing %d CPUs, %d hotplug CPUs\n",
1570 		possible, max_t(int, possible - num_processors, 0));
1571 
1572 	reset_cpu_possible_mask();
1573 
1574 	for (i = 0; i < possible; i++)
1575 		set_cpu_possible(i, true);
1576 }
1577 
1578 #ifdef CONFIG_HOTPLUG_CPU
1579 
1580 /* Recompute SMT state for all CPUs on offline */
1581 static void recompute_smt_state(void)
1582 {
1583 	int max_threads, cpu;
1584 
1585 	max_threads = 0;
1586 	for_each_online_cpu (cpu) {
1587 		int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1588 
1589 		if (threads > max_threads)
1590 			max_threads = threads;
1591 	}
1592 	__max_smt_threads = max_threads;
1593 }
1594 
1595 static void remove_siblinginfo(int cpu)
1596 {
1597 	int sibling;
1598 	struct cpuinfo_x86 *c = &cpu_data(cpu);
1599 
1600 	for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1601 		cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1602 		/*/
1603 		 * last thread sibling in this cpu core going down
1604 		 */
1605 		if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1606 			cpu_data(sibling).booted_cores--;
1607 	}
1608 
1609 	for_each_cpu(sibling, topology_die_cpumask(cpu))
1610 		cpumask_clear_cpu(cpu, topology_die_cpumask(sibling));
1611 
1612 	for_each_cpu(sibling, topology_sibling_cpumask(cpu)) {
1613 		cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1614 		if (cpumask_weight(topology_sibling_cpumask(sibling)) == 1)
1615 			cpu_data(sibling).smt_active = false;
1616 	}
1617 
1618 	for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1619 		cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1620 	for_each_cpu(sibling, cpu_l2c_shared_mask(cpu))
1621 		cpumask_clear_cpu(cpu, cpu_l2c_shared_mask(sibling));
1622 	cpumask_clear(cpu_llc_shared_mask(cpu));
1623 	cpumask_clear(cpu_l2c_shared_mask(cpu));
1624 	cpumask_clear(topology_sibling_cpumask(cpu));
1625 	cpumask_clear(topology_core_cpumask(cpu));
1626 	cpumask_clear(topology_die_cpumask(cpu));
1627 	c->cpu_core_id = 0;
1628 	c->booted_cores = 0;
1629 	cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1630 	recompute_smt_state();
1631 }
1632 
1633 static void remove_cpu_from_maps(int cpu)
1634 {
1635 	set_cpu_online(cpu, false);
1636 	cpumask_clear_cpu(cpu, cpu_callout_mask);
1637 	cpumask_clear_cpu(cpu, cpu_callin_mask);
1638 	/* was set by cpu_init() */
1639 	cpumask_clear_cpu(cpu, cpu_initialized_mask);
1640 	numa_remove_cpu(cpu);
1641 }
1642 
1643 void cpu_disable_common(void)
1644 {
1645 	int cpu = smp_processor_id();
1646 
1647 	remove_siblinginfo(cpu);
1648 
1649 	/* It's now safe to remove this processor from the online map */
1650 	lock_vector_lock();
1651 	remove_cpu_from_maps(cpu);
1652 	unlock_vector_lock();
1653 	fixup_irqs();
1654 	lapic_offline();
1655 }
1656 
1657 int native_cpu_disable(void)
1658 {
1659 	int ret;
1660 
1661 	ret = lapic_can_unplug_cpu();
1662 	if (ret)
1663 		return ret;
1664 
1665 	cpu_disable_common();
1666 
1667         /*
1668          * Disable the local APIC. Otherwise IPI broadcasts will reach
1669          * it. It still responds normally to INIT, NMI, SMI, and SIPI
1670          * messages.
1671          *
1672          * Disabling the APIC must happen after cpu_disable_common()
1673          * which invokes fixup_irqs().
1674          *
1675          * Disabling the APIC preserves already set bits in IRR, but
1676          * an interrupt arriving after disabling the local APIC does not
1677          * set the corresponding IRR bit.
1678          *
1679          * fixup_irqs() scans IRR for set bits so it can raise a not
1680          * yet handled interrupt on the new destination CPU via an IPI
1681          * but obviously it can't do so for IRR bits which are not set.
1682          * IOW, interrupts arriving after disabling the local APIC will
1683          * be lost.
1684          */
1685 	apic_soft_disable();
1686 
1687 	return 0;
1688 }
1689 
1690 int common_cpu_die(unsigned int cpu)
1691 {
1692 	int ret = 0;
1693 
1694 	/* We don't do anything here: idle task is faking death itself. */
1695 
1696 	/* They ack this in play_dead() by setting CPU_DEAD */
1697 	if (cpu_wait_death(cpu, 5)) {
1698 		if (system_state == SYSTEM_RUNNING)
1699 			pr_info("CPU %u is now offline\n", cpu);
1700 	} else {
1701 		pr_err("CPU %u didn't die...\n", cpu);
1702 		ret = -1;
1703 	}
1704 
1705 	return ret;
1706 }
1707 
1708 void native_cpu_die(unsigned int cpu)
1709 {
1710 	common_cpu_die(cpu);
1711 }
1712 
1713 void play_dead_common(void)
1714 {
1715 	idle_task_exit();
1716 
1717 	/* Ack it */
1718 	(void)cpu_report_death();
1719 
1720 	/*
1721 	 * With physical CPU hotplug, we should halt the cpu
1722 	 */
1723 	local_irq_disable();
1724 }
1725 
1726 /**
1727  * cond_wakeup_cpu0 - Wake up CPU0 if needed.
1728  *
1729  * If NMI wants to wake up CPU0, start CPU0.
1730  */
1731 void cond_wakeup_cpu0(void)
1732 {
1733 	if (smp_processor_id() == 0 && enable_start_cpu0)
1734 		start_cpu0();
1735 }
1736 EXPORT_SYMBOL_GPL(cond_wakeup_cpu0);
1737 
1738 /*
1739  * We need to flush the caches before going to sleep, lest we have
1740  * dirty data in our caches when we come back up.
1741  */
1742 static inline void mwait_play_dead(void)
1743 {
1744 	unsigned int eax, ebx, ecx, edx;
1745 	unsigned int highest_cstate = 0;
1746 	unsigned int highest_subcstate = 0;
1747 	void *mwait_ptr;
1748 	int i;
1749 
1750 	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD ||
1751 	    boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
1752 		return;
1753 	if (!this_cpu_has(X86_FEATURE_MWAIT))
1754 		return;
1755 	if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1756 		return;
1757 	if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1758 		return;
1759 
1760 	eax = CPUID_MWAIT_LEAF;
1761 	ecx = 0;
1762 	native_cpuid(&eax, &ebx, &ecx, &edx);
1763 
1764 	/*
1765 	 * eax will be 0 if EDX enumeration is not valid.
1766 	 * Initialized below to cstate, sub_cstate value when EDX is valid.
1767 	 */
1768 	if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1769 		eax = 0;
1770 	} else {
1771 		edx >>= MWAIT_SUBSTATE_SIZE;
1772 		for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1773 			if (edx & MWAIT_SUBSTATE_MASK) {
1774 				highest_cstate = i;
1775 				highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1776 			}
1777 		}
1778 		eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1779 			(highest_subcstate - 1);
1780 	}
1781 
1782 	/*
1783 	 * This should be a memory location in a cache line which is
1784 	 * unlikely to be touched by other processors.  The actual
1785 	 * content is immaterial as it is not actually modified in any way.
1786 	 */
1787 	mwait_ptr = &current_thread_info()->flags;
1788 
1789 	wbinvd();
1790 
1791 	while (1) {
1792 		/*
1793 		 * The CLFLUSH is a workaround for erratum AAI65 for
1794 		 * the Xeon 7400 series.  It's not clear it is actually
1795 		 * needed, but it should be harmless in either case.
1796 		 * The WBINVD is insufficient due to the spurious-wakeup
1797 		 * case where we return around the loop.
1798 		 */
1799 		mb();
1800 		clflush(mwait_ptr);
1801 		mb();
1802 		__monitor(mwait_ptr, 0, 0);
1803 		mb();
1804 		__mwait(eax, 0);
1805 
1806 		cond_wakeup_cpu0();
1807 	}
1808 }
1809 
1810 void hlt_play_dead(void)
1811 {
1812 	if (__this_cpu_read(cpu_info.x86) >= 4)
1813 		wbinvd();
1814 
1815 	while (1) {
1816 		native_halt();
1817 
1818 		cond_wakeup_cpu0();
1819 	}
1820 }
1821 
1822 void native_play_dead(void)
1823 {
1824 	play_dead_common();
1825 	tboot_shutdown(TB_SHUTDOWN_WFS);
1826 
1827 	mwait_play_dead();	/* Only returns on failure */
1828 	if (cpuidle_play_dead())
1829 		hlt_play_dead();
1830 }
1831 
1832 #else /* ... !CONFIG_HOTPLUG_CPU */
1833 int native_cpu_disable(void)
1834 {
1835 	return -ENOSYS;
1836 }
1837 
1838 void native_cpu_die(unsigned int cpu)
1839 {
1840 	/* We said "no" in __cpu_disable */
1841 	BUG();
1842 }
1843 
1844 void native_play_dead(void)
1845 {
1846 	BUG();
1847 }
1848 
1849 #endif
1850 
1851 #ifdef CONFIG_X86_64
1852 /*
1853  * APERF/MPERF frequency ratio computation.
1854  *
1855  * The scheduler wants to do frequency invariant accounting and needs a <1
1856  * ratio to account for the 'current' frequency, corresponding to
1857  * freq_curr / freq_max.
1858  *
1859  * Since the frequency freq_curr on x86 is controlled by micro-controller and
1860  * our P-state setting is little more than a request/hint, we need to observe
1861  * the effective frequency 'BusyMHz', i.e. the average frequency over a time
1862  * interval after discarding idle time. This is given by:
1863  *
1864  *   BusyMHz = delta_APERF / delta_MPERF * freq_base
1865  *
1866  * where freq_base is the max non-turbo P-state.
1867  *
1868  * The freq_max term has to be set to a somewhat arbitrary value, because we
1869  * can't know which turbo states will be available at a given point in time:
1870  * it all depends on the thermal headroom of the entire package. We set it to
1871  * the turbo level with 4 cores active.
1872  *
1873  * Benchmarks show that's a good compromise between the 1C turbo ratio
1874  * (freq_curr/freq_max would rarely reach 1) and something close to freq_base,
1875  * which would ignore the entire turbo range (a conspicuous part, making
1876  * freq_curr/freq_max always maxed out).
1877  *
1878  * An exception to the heuristic above is the Atom uarch, where we choose the
1879  * highest turbo level for freq_max since Atom's are generally oriented towards
1880  * power efficiency.
1881  *
1882  * Setting freq_max to anything less than the 1C turbo ratio makes the ratio
1883  * freq_curr / freq_max to eventually grow >1, in which case we clip it to 1.
1884  */
1885 
1886 DEFINE_STATIC_KEY_FALSE(arch_scale_freq_key);
1887 
1888 static DEFINE_PER_CPU(u64, arch_prev_aperf);
1889 static DEFINE_PER_CPU(u64, arch_prev_mperf);
1890 static u64 arch_turbo_freq_ratio = SCHED_CAPACITY_SCALE;
1891 static u64 arch_max_freq_ratio = SCHED_CAPACITY_SCALE;
1892 
1893 void arch_set_max_freq_ratio(bool turbo_disabled)
1894 {
1895 	arch_max_freq_ratio = turbo_disabled ? SCHED_CAPACITY_SCALE :
1896 					arch_turbo_freq_ratio;
1897 }
1898 EXPORT_SYMBOL_GPL(arch_set_max_freq_ratio);
1899 
1900 static bool turbo_disabled(void)
1901 {
1902 	u64 misc_en;
1903 	int err;
1904 
1905 	err = rdmsrl_safe(MSR_IA32_MISC_ENABLE, &misc_en);
1906 	if (err)
1907 		return false;
1908 
1909 	return (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
1910 }
1911 
1912 static bool slv_set_max_freq_ratio(u64 *base_freq, u64 *turbo_freq)
1913 {
1914 	int err;
1915 
1916 	err = rdmsrl_safe(MSR_ATOM_CORE_RATIOS, base_freq);
1917 	if (err)
1918 		return false;
1919 
1920 	err = rdmsrl_safe(MSR_ATOM_CORE_TURBO_RATIOS, turbo_freq);
1921 	if (err)
1922 		return false;
1923 
1924 	*base_freq = (*base_freq >> 16) & 0x3F;     /* max P state */
1925 	*turbo_freq = *turbo_freq & 0x3F;           /* 1C turbo    */
1926 
1927 	return true;
1928 }
1929 
1930 #define X86_MATCH(model)					\
1931 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6,		\
1932 		INTEL_FAM6_##model, X86_FEATURE_APERFMPERF, NULL)
1933 
1934 static const struct x86_cpu_id has_knl_turbo_ratio_limits[] = {
1935 	X86_MATCH(XEON_PHI_KNL),
1936 	X86_MATCH(XEON_PHI_KNM),
1937 	{}
1938 };
1939 
1940 static const struct x86_cpu_id has_skx_turbo_ratio_limits[] = {
1941 	X86_MATCH(SKYLAKE_X),
1942 	{}
1943 };
1944 
1945 static const struct x86_cpu_id has_glm_turbo_ratio_limits[] = {
1946 	X86_MATCH(ATOM_GOLDMONT),
1947 	X86_MATCH(ATOM_GOLDMONT_D),
1948 	X86_MATCH(ATOM_GOLDMONT_PLUS),
1949 	{}
1950 };
1951 
1952 static bool knl_set_max_freq_ratio(u64 *base_freq, u64 *turbo_freq,
1953 				int num_delta_fratio)
1954 {
1955 	int fratio, delta_fratio, found;
1956 	int err, i;
1957 	u64 msr;
1958 
1959 	err = rdmsrl_safe(MSR_PLATFORM_INFO, base_freq);
1960 	if (err)
1961 		return false;
1962 
1963 	*base_freq = (*base_freq >> 8) & 0xFF;	    /* max P state */
1964 
1965 	err = rdmsrl_safe(MSR_TURBO_RATIO_LIMIT, &msr);
1966 	if (err)
1967 		return false;
1968 
1969 	fratio = (msr >> 8) & 0xFF;
1970 	i = 16;
1971 	found = 0;
1972 	do {
1973 		if (found >= num_delta_fratio) {
1974 			*turbo_freq = fratio;
1975 			return true;
1976 		}
1977 
1978 		delta_fratio = (msr >> (i + 5)) & 0x7;
1979 
1980 		if (delta_fratio) {
1981 			found += 1;
1982 			fratio -= delta_fratio;
1983 		}
1984 
1985 		i += 8;
1986 	} while (i < 64);
1987 
1988 	return true;
1989 }
1990 
1991 static bool skx_set_max_freq_ratio(u64 *base_freq, u64 *turbo_freq, int size)
1992 {
1993 	u64 ratios, counts;
1994 	u32 group_size;
1995 	int err, i;
1996 
1997 	err = rdmsrl_safe(MSR_PLATFORM_INFO, base_freq);
1998 	if (err)
1999 		return false;
2000 
2001 	*base_freq = (*base_freq >> 8) & 0xFF;      /* max P state */
2002 
2003 	err = rdmsrl_safe(MSR_TURBO_RATIO_LIMIT, &ratios);
2004 	if (err)
2005 		return false;
2006 
2007 	err = rdmsrl_safe(MSR_TURBO_RATIO_LIMIT1, &counts);
2008 	if (err)
2009 		return false;
2010 
2011 	for (i = 0; i < 64; i += 8) {
2012 		group_size = (counts >> i) & 0xFF;
2013 		if (group_size >= size) {
2014 			*turbo_freq = (ratios >> i) & 0xFF;
2015 			return true;
2016 		}
2017 	}
2018 
2019 	return false;
2020 }
2021 
2022 static bool core_set_max_freq_ratio(u64 *base_freq, u64 *turbo_freq)
2023 {
2024 	u64 msr;
2025 	int err;
2026 
2027 	err = rdmsrl_safe(MSR_PLATFORM_INFO, base_freq);
2028 	if (err)
2029 		return false;
2030 
2031 	err = rdmsrl_safe(MSR_TURBO_RATIO_LIMIT, &msr);
2032 	if (err)
2033 		return false;
2034 
2035 	*base_freq = (*base_freq >> 8) & 0xFF;    /* max P state */
2036 	*turbo_freq = (msr >> 24) & 0xFF;         /* 4C turbo    */
2037 
2038 	/* The CPU may have less than 4 cores */
2039 	if (!*turbo_freq)
2040 		*turbo_freq = msr & 0xFF;         /* 1C turbo    */
2041 
2042 	return true;
2043 }
2044 
2045 static bool intel_set_max_freq_ratio(void)
2046 {
2047 	u64 base_freq, turbo_freq;
2048 	u64 turbo_ratio;
2049 
2050 	if (slv_set_max_freq_ratio(&base_freq, &turbo_freq))
2051 		goto out;
2052 
2053 	if (x86_match_cpu(has_glm_turbo_ratio_limits) &&
2054 	    skx_set_max_freq_ratio(&base_freq, &turbo_freq, 1))
2055 		goto out;
2056 
2057 	if (x86_match_cpu(has_knl_turbo_ratio_limits) &&
2058 	    knl_set_max_freq_ratio(&base_freq, &turbo_freq, 1))
2059 		goto out;
2060 
2061 	if (x86_match_cpu(has_skx_turbo_ratio_limits) &&
2062 	    skx_set_max_freq_ratio(&base_freq, &turbo_freq, 4))
2063 		goto out;
2064 
2065 	if (core_set_max_freq_ratio(&base_freq, &turbo_freq))
2066 		goto out;
2067 
2068 	return false;
2069 
2070 out:
2071 	/*
2072 	 * Some hypervisors advertise X86_FEATURE_APERFMPERF
2073 	 * but then fill all MSR's with zeroes.
2074 	 * Some CPUs have turbo boost but don't declare any turbo ratio
2075 	 * in MSR_TURBO_RATIO_LIMIT.
2076 	 */
2077 	if (!base_freq || !turbo_freq) {
2078 		pr_debug("Couldn't determine cpu base or turbo frequency, necessary for scale-invariant accounting.\n");
2079 		return false;
2080 	}
2081 
2082 	turbo_ratio = div_u64(turbo_freq * SCHED_CAPACITY_SCALE, base_freq);
2083 	if (!turbo_ratio) {
2084 		pr_debug("Non-zero turbo and base frequencies led to a 0 ratio.\n");
2085 		return false;
2086 	}
2087 
2088 	arch_turbo_freq_ratio = turbo_ratio;
2089 	arch_set_max_freq_ratio(turbo_disabled());
2090 
2091 	return true;
2092 }
2093 
2094 static void init_counter_refs(void)
2095 {
2096 	u64 aperf, mperf;
2097 
2098 	rdmsrl(MSR_IA32_APERF, aperf);
2099 	rdmsrl(MSR_IA32_MPERF, mperf);
2100 
2101 	this_cpu_write(arch_prev_aperf, aperf);
2102 	this_cpu_write(arch_prev_mperf, mperf);
2103 }
2104 
2105 #ifdef CONFIG_PM_SLEEP
2106 static struct syscore_ops freq_invariance_syscore_ops = {
2107 	.resume = init_counter_refs,
2108 };
2109 
2110 static void register_freq_invariance_syscore_ops(void)
2111 {
2112 	/* Bail out if registered already. */
2113 	if (freq_invariance_syscore_ops.node.prev)
2114 		return;
2115 
2116 	register_syscore_ops(&freq_invariance_syscore_ops);
2117 }
2118 #else
2119 static inline void register_freq_invariance_syscore_ops(void) {}
2120 #endif
2121 
2122 void init_freq_invariance(bool secondary, bool cppc_ready)
2123 {
2124 	bool ret = false;
2125 
2126 	if (!boot_cpu_has(X86_FEATURE_APERFMPERF))
2127 		return;
2128 
2129 	if (secondary) {
2130 		if (static_branch_likely(&arch_scale_freq_key)) {
2131 			init_counter_refs();
2132 		}
2133 		return;
2134 	}
2135 
2136 	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL)
2137 		ret = intel_set_max_freq_ratio();
2138 	else if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD) {
2139 		if (!cppc_ready) {
2140 			return;
2141 		}
2142 		ret = amd_set_max_freq_ratio(&arch_turbo_freq_ratio);
2143 	}
2144 
2145 	if (ret) {
2146 		init_counter_refs();
2147 		static_branch_enable(&arch_scale_freq_key);
2148 		register_freq_invariance_syscore_ops();
2149 		pr_info("Estimated ratio of average max frequency by base frequency (times 1024): %llu\n", arch_max_freq_ratio);
2150 	} else {
2151 		pr_debug("Couldn't determine max cpu frequency, necessary for scale-invariant accounting.\n");
2152 	}
2153 }
2154 
2155 static void disable_freq_invariance_workfn(struct work_struct *work)
2156 {
2157 	static_branch_disable(&arch_scale_freq_key);
2158 }
2159 
2160 static DECLARE_WORK(disable_freq_invariance_work,
2161 		    disable_freq_invariance_workfn);
2162 
2163 DEFINE_PER_CPU(unsigned long, arch_freq_scale) = SCHED_CAPACITY_SCALE;
2164 
2165 void arch_scale_freq_tick(void)
2166 {
2167 	u64 freq_scale;
2168 	u64 aperf, mperf;
2169 	u64 acnt, mcnt;
2170 
2171 	if (!arch_scale_freq_invariant())
2172 		return;
2173 
2174 	rdmsrl(MSR_IA32_APERF, aperf);
2175 	rdmsrl(MSR_IA32_MPERF, mperf);
2176 
2177 	acnt = aperf - this_cpu_read(arch_prev_aperf);
2178 	mcnt = mperf - this_cpu_read(arch_prev_mperf);
2179 
2180 	this_cpu_write(arch_prev_aperf, aperf);
2181 	this_cpu_write(arch_prev_mperf, mperf);
2182 
2183 	if (check_shl_overflow(acnt, 2*SCHED_CAPACITY_SHIFT, &acnt))
2184 		goto error;
2185 
2186 	if (check_mul_overflow(mcnt, arch_max_freq_ratio, &mcnt) || !mcnt)
2187 		goto error;
2188 
2189 	freq_scale = div64_u64(acnt, mcnt);
2190 	if (!freq_scale)
2191 		goto error;
2192 
2193 	if (freq_scale > SCHED_CAPACITY_SCALE)
2194 		freq_scale = SCHED_CAPACITY_SCALE;
2195 
2196 	this_cpu_write(arch_freq_scale, freq_scale);
2197 	return;
2198 
2199 error:
2200 	pr_warn("Scheduler frequency invariance went wobbly, disabling!\n");
2201 	schedule_work(&disable_freq_invariance_work);
2202 }
2203 #endif /* CONFIG_X86_64 */
2204