xref: /openbmc/linux/arch/x86/kernel/smpboot.c (revision 9be08a27)
1  /*
2  *	x86 SMP booting functions
3  *
4  *	(c) 1995 Alan Cox, Building #3 <alan@lxorguk.ukuu.org.uk>
5  *	(c) 1998, 1999, 2000, 2009 Ingo Molnar <mingo@redhat.com>
6  *	Copyright 2001 Andi Kleen, SuSE Labs.
7  *
8  *	Much of the core SMP work is based on previous work by Thomas Radke, to
9  *	whom a great many thanks are extended.
10  *
11  *	Thanks to Intel for making available several different Pentium,
12  *	Pentium Pro and Pentium-II/Xeon MP machines.
13  *	Original development of Linux SMP code supported by Caldera.
14  *
15  *	This code is released under the GNU General Public License version 2 or
16  *	later.
17  *
18  *	Fixes
19  *		Felix Koop	:	NR_CPUS used properly
20  *		Jose Renau	:	Handle single CPU case.
21  *		Alan Cox	:	By repeated request 8) - Total BogoMIPS report.
22  *		Greg Wright	:	Fix for kernel stacks panic.
23  *		Erich Boleyn	:	MP v1.4 and additional changes.
24  *	Matthias Sattler	:	Changes for 2.1 kernel map.
25  *	Michel Lespinasse	:	Changes for 2.1 kernel map.
26  *	Michael Chastain	:	Change trampoline.S to gnu as.
27  *		Alan Cox	:	Dumb bug: 'B' step PPro's are fine
28  *		Ingo Molnar	:	Added APIC timers, based on code
29  *					from Jose Renau
30  *		Ingo Molnar	:	various cleanups and rewrites
31  *		Tigran Aivazian	:	fixed "0.00 in /proc/uptime on SMP" bug.
32  *	Maciej W. Rozycki	:	Bits for genuine 82489DX APICs
33  *	Andi Kleen		:	Changed for SMP boot into long mode.
34  *		Martin J. Bligh	: 	Added support for multi-quad systems
35  *		Dave Jones	:	Report invalid combinations of Athlon CPUs.
36  *		Rusty Russell	:	Hacked into shape for new "hotplug" boot process.
37  *      Andi Kleen              :       Converted to new state machine.
38  *	Ashok Raj		: 	CPU hotplug support
39  *	Glauber Costa		:	i386 and x86_64 integration
40  */
41 
42 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
43 
44 #include <linux/init.h>
45 #include <linux/smp.h>
46 #include <linux/export.h>
47 #include <linux/sched.h>
48 #include <linux/sched/topology.h>
49 #include <linux/sched/hotplug.h>
50 #include <linux/sched/task_stack.h>
51 #include <linux/percpu.h>
52 #include <linux/bootmem.h>
53 #include <linux/err.h>
54 #include <linux/nmi.h>
55 #include <linux/tboot.h>
56 #include <linux/stackprotector.h>
57 #include <linux/gfp.h>
58 #include <linux/cpuidle.h>
59 
60 #include <asm/acpi.h>
61 #include <asm/desc.h>
62 #include <asm/nmi.h>
63 #include <asm/irq.h>
64 #include <asm/realmode.h>
65 #include <asm/cpu.h>
66 #include <asm/numa.h>
67 #include <asm/pgtable.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/internal.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 
85 /* representing HT siblings of each logical CPU */
86 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
87 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
88 
89 /* representing HT and core siblings of each logical CPU */
90 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
91 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
92 
93 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_llc_shared_map);
94 
95 /* Per CPU bogomips and other parameters */
96 DEFINE_PER_CPU_READ_MOSTLY(struct cpuinfo_x86, cpu_info);
97 EXPORT_PER_CPU_SYMBOL(cpu_info);
98 
99 /* Logical package management. We might want to allocate that dynamically */
100 unsigned int __max_logical_packages __read_mostly;
101 EXPORT_SYMBOL(__max_logical_packages);
102 static unsigned int logical_packages __read_mostly;
103 
104 /* Maximum number of SMT threads on any online core */
105 int __read_mostly __max_smt_threads = 1;
106 
107 /* Flag to indicate if a complete sched domain rebuild is required */
108 bool x86_topology_update;
109 
110 int arch_update_cpu_topology(void)
111 {
112 	int retval = x86_topology_update;
113 
114 	x86_topology_update = false;
115 	return retval;
116 }
117 
118 static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
119 {
120 	unsigned long flags;
121 
122 	spin_lock_irqsave(&rtc_lock, flags);
123 	CMOS_WRITE(0xa, 0xf);
124 	spin_unlock_irqrestore(&rtc_lock, flags);
125 	*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) =
126 							start_eip >> 4;
127 	*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) =
128 							start_eip & 0xf;
129 }
130 
131 static inline void smpboot_restore_warm_reset_vector(void)
132 {
133 	unsigned long flags;
134 
135 	/*
136 	 * Paranoid:  Set warm reset code and vector here back
137 	 * to default values.
138 	 */
139 	spin_lock_irqsave(&rtc_lock, flags);
140 	CMOS_WRITE(0, 0xf);
141 	spin_unlock_irqrestore(&rtc_lock, flags);
142 
143 	*((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
144 }
145 
146 /*
147  * Report back to the Boot Processor during boot time or to the caller processor
148  * during CPU online.
149  */
150 static void smp_callin(void)
151 {
152 	int cpuid, phys_id;
153 
154 	/*
155 	 * If waken up by an INIT in an 82489DX configuration
156 	 * cpu_callout_mask guarantees we don't get here before
157 	 * an INIT_deassert IPI reaches our local APIC, so it is
158 	 * now safe to touch our local APIC.
159 	 */
160 	cpuid = smp_processor_id();
161 
162 	/*
163 	 * (This works even if the APIC is not enabled.)
164 	 */
165 	phys_id = read_apic_id();
166 
167 	/*
168 	 * the boot CPU has finished the init stage and is spinning
169 	 * on callin_map until we finish. We are free to set up this
170 	 * CPU, first the APIC. (this is probably redundant on most
171 	 * boards)
172 	 */
173 	apic_ap_setup();
174 
175 	/*
176 	 * Save our processor parameters. Note: this information
177 	 * is needed for clock calibration.
178 	 */
179 	smp_store_cpu_info(cpuid);
180 
181 	/*
182 	 * The topology information must be up to date before
183 	 * calibrate_delay() and notify_cpu_starting().
184 	 */
185 	set_cpu_sibling_map(raw_smp_processor_id());
186 
187 	/*
188 	 * Get our bogomips.
189 	 * Update loops_per_jiffy in cpu_data. Previous call to
190 	 * smp_store_cpu_info() stored a value that is close but not as
191 	 * accurate as the value just calculated.
192 	 */
193 	calibrate_delay();
194 	cpu_data(cpuid).loops_per_jiffy = loops_per_jiffy;
195 	pr_debug("Stack at about %p\n", &cpuid);
196 
197 	wmb();
198 
199 	notify_cpu_starting(cpuid);
200 
201 	/*
202 	 * Allow the master to continue.
203 	 */
204 	cpumask_set_cpu(cpuid, cpu_callin_mask);
205 }
206 
207 static int cpu0_logical_apicid;
208 static int enable_start_cpu0;
209 /*
210  * Activate a secondary processor.
211  */
212 static void notrace start_secondary(void *unused)
213 {
214 	/*
215 	 * Don't put *anything* except direct CPU state initialization
216 	 * before cpu_init(), SMP booting is too fragile that we want to
217 	 * limit the things done here to the most necessary things.
218 	 */
219 	if (boot_cpu_has(X86_FEATURE_PCID))
220 		__write_cr4(__read_cr4() | X86_CR4_PCIDE);
221 
222 #ifdef CONFIG_X86_32
223 	/* switch away from the initial page table */
224 	load_cr3(swapper_pg_dir);
225 	/*
226 	 * Initialize the CR4 shadow before doing anything that could
227 	 * try to read it.
228 	 */
229 	cr4_init_shadow();
230 	__flush_tlb_all();
231 #endif
232 	load_current_idt();
233 	cpu_init();
234 	x86_cpuinit.early_percpu_clock_init();
235 	preempt_disable();
236 	smp_callin();
237 
238 	enable_start_cpu0 = 0;
239 
240 	/* otherwise gcc will move up smp_processor_id before the cpu_init */
241 	barrier();
242 	/*
243 	 * Check TSC synchronization with the boot CPU:
244 	 */
245 	check_tsc_sync_target();
246 
247 	speculative_store_bypass_ht_init();
248 
249 	/*
250 	 * Lock vector_lock, set CPU online and bring the vector
251 	 * allocator online. Online must be set with vector_lock held
252 	 * to prevent a concurrent irq setup/teardown from seeing a
253 	 * half valid vector space.
254 	 */
255 	lock_vector_lock();
256 	set_cpu_online(smp_processor_id(), true);
257 	lapic_online();
258 	unlock_vector_lock();
259 	cpu_set_state_online(smp_processor_id());
260 	x86_platform.nmi_init();
261 
262 	/* enable local interrupts */
263 	local_irq_enable();
264 
265 	/* to prevent fake stack check failure in clock setup */
266 	boot_init_stack_canary();
267 
268 	x86_cpuinit.setup_percpu_clockev();
269 
270 	wmb();
271 	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
272 }
273 
274 /**
275  * topology_is_primary_thread - Check whether CPU is the primary SMT thread
276  * @cpu:	CPU to check
277  */
278 bool topology_is_primary_thread(unsigned int cpu)
279 {
280 	return apic_id_is_primary_thread(per_cpu(x86_cpu_to_apicid, cpu));
281 }
282 
283 /**
284  * topology_smt_supported - Check whether SMT is supported by the CPUs
285  */
286 bool topology_smt_supported(void)
287 {
288 	return smp_num_siblings > 1;
289 }
290 
291 /**
292  * topology_phys_to_logical_pkg - Map a physical package id to a logical
293  *
294  * Returns logical package id or -1 if not found
295  */
296 int topology_phys_to_logical_pkg(unsigned int phys_pkg)
297 {
298 	int cpu;
299 
300 	for_each_possible_cpu(cpu) {
301 		struct cpuinfo_x86 *c = &cpu_data(cpu);
302 
303 		if (c->initialized && c->phys_proc_id == phys_pkg)
304 			return c->logical_proc_id;
305 	}
306 	return -1;
307 }
308 EXPORT_SYMBOL(topology_phys_to_logical_pkg);
309 
310 /**
311  * topology_update_package_map - Update the physical to logical package map
312  * @pkg:	The physical package id as retrieved via CPUID
313  * @cpu:	The cpu for which this is updated
314  */
315 int topology_update_package_map(unsigned int pkg, unsigned int cpu)
316 {
317 	int new;
318 
319 	/* Already available somewhere? */
320 	new = topology_phys_to_logical_pkg(pkg);
321 	if (new >= 0)
322 		goto found;
323 
324 	new = logical_packages++;
325 	if (new != pkg) {
326 		pr_info("CPU %u Converting physical %u to logical package %u\n",
327 			cpu, pkg, new);
328 	}
329 found:
330 	cpu_data(cpu).logical_proc_id = new;
331 	return 0;
332 }
333 
334 void __init smp_store_boot_cpu_info(void)
335 {
336 	int id = 0; /* CPU 0 */
337 	struct cpuinfo_x86 *c = &cpu_data(id);
338 
339 	*c = boot_cpu_data;
340 	c->cpu_index = id;
341 	topology_update_package_map(c->phys_proc_id, id);
342 	c->initialized = true;
343 }
344 
345 /*
346  * The bootstrap kernel entry code has set these up. Save them for
347  * a given CPU
348  */
349 void smp_store_cpu_info(int id)
350 {
351 	struct cpuinfo_x86 *c = &cpu_data(id);
352 
353 	/* Copy boot_cpu_data only on the first bringup */
354 	if (!c->initialized)
355 		*c = boot_cpu_data;
356 	c->cpu_index = id;
357 	/*
358 	 * During boot time, CPU0 has this setup already. Save the info when
359 	 * bringing up AP or offlined CPU0.
360 	 */
361 	identify_secondary_cpu(c);
362 	c->initialized = true;
363 }
364 
365 static bool
366 topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
367 {
368 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
369 
370 	return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
371 }
372 
373 static bool
374 topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
375 {
376 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
377 
378 	return !WARN_ONCE(!topology_same_node(c, o),
379 		"sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
380 		"[node: %d != %d]. Ignoring dependency.\n",
381 		cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
382 }
383 
384 #define link_mask(mfunc, c1, c2)					\
385 do {									\
386 	cpumask_set_cpu((c1), mfunc(c2));				\
387 	cpumask_set_cpu((c2), mfunc(c1));				\
388 } while (0)
389 
390 static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
391 {
392 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
393 		int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
394 
395 		if (c->phys_proc_id == o->phys_proc_id &&
396 		    per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2)) {
397 			if (c->cpu_core_id == o->cpu_core_id)
398 				return topology_sane(c, o, "smt");
399 
400 			if ((c->cu_id != 0xff) &&
401 			    (o->cu_id != 0xff) &&
402 			    (c->cu_id == o->cu_id))
403 				return topology_sane(c, o, "smt");
404 		}
405 
406 	} else if (c->phys_proc_id == o->phys_proc_id &&
407 		   c->cpu_core_id == o->cpu_core_id) {
408 		return topology_sane(c, o, "smt");
409 	}
410 
411 	return false;
412 }
413 
414 /*
415  * Define snc_cpu[] for SNC (Sub-NUMA Cluster) CPUs.
416  *
417  * These are Intel CPUs that enumerate an LLC that is shared by
418  * multiple NUMA nodes. The LLC on these systems is shared for
419  * off-package data access but private to the NUMA node (half
420  * of the package) for on-package access.
421  *
422  * CPUID (the source of the information about the LLC) can only
423  * enumerate the cache as being shared *or* unshared, but not
424  * this particular configuration. The CPU in this case enumerates
425  * the cache to be shared across the entire package (spanning both
426  * NUMA nodes).
427  */
428 
429 static const struct x86_cpu_id snc_cpu[] = {
430 	{ X86_VENDOR_INTEL, 6, INTEL_FAM6_SKYLAKE_X },
431 	{}
432 };
433 
434 static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
435 {
436 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
437 
438 	/* Do not match if we do not have a valid APICID for cpu: */
439 	if (per_cpu(cpu_llc_id, cpu1) == BAD_APICID)
440 		return false;
441 
442 	/* Do not match if LLC id does not match: */
443 	if (per_cpu(cpu_llc_id, cpu1) != per_cpu(cpu_llc_id, cpu2))
444 		return false;
445 
446 	/*
447 	 * Allow the SNC topology without warning. Return of false
448 	 * means 'c' does not share the LLC of 'o'. This will be
449 	 * reflected to userspace.
450 	 */
451 	if (!topology_same_node(c, o) && x86_match_cpu(snc_cpu))
452 		return false;
453 
454 	return topology_sane(c, o, "llc");
455 }
456 
457 /*
458  * Unlike the other levels, we do not enforce keeping a
459  * multicore group inside a NUMA node.  If this happens, we will
460  * discard the MC level of the topology later.
461  */
462 static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
463 {
464 	if (c->phys_proc_id == o->phys_proc_id)
465 		return true;
466 	return false;
467 }
468 
469 #if defined(CONFIG_SCHED_SMT) || defined(CONFIG_SCHED_MC)
470 static inline int x86_sched_itmt_flags(void)
471 {
472 	return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
473 }
474 
475 #ifdef CONFIG_SCHED_MC
476 static int x86_core_flags(void)
477 {
478 	return cpu_core_flags() | x86_sched_itmt_flags();
479 }
480 #endif
481 #ifdef CONFIG_SCHED_SMT
482 static int x86_smt_flags(void)
483 {
484 	return cpu_smt_flags() | x86_sched_itmt_flags();
485 }
486 #endif
487 #endif
488 
489 static struct sched_domain_topology_level x86_numa_in_package_topology[] = {
490 #ifdef CONFIG_SCHED_SMT
491 	{ cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
492 #endif
493 #ifdef CONFIG_SCHED_MC
494 	{ cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
495 #endif
496 	{ NULL, },
497 };
498 
499 static struct sched_domain_topology_level x86_topology[] = {
500 #ifdef CONFIG_SCHED_SMT
501 	{ cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
502 #endif
503 #ifdef CONFIG_SCHED_MC
504 	{ cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
505 #endif
506 	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
507 	{ NULL, },
508 };
509 
510 /*
511  * Set if a package/die has multiple NUMA nodes inside.
512  * AMD Magny-Cours, Intel Cluster-on-Die, and Intel
513  * Sub-NUMA Clustering have this.
514  */
515 static bool x86_has_numa_in_package;
516 
517 void set_cpu_sibling_map(int cpu)
518 {
519 	bool has_smt = smp_num_siblings > 1;
520 	bool has_mp = has_smt || boot_cpu_data.x86_max_cores > 1;
521 	struct cpuinfo_x86 *c = &cpu_data(cpu);
522 	struct cpuinfo_x86 *o;
523 	int i, threads;
524 
525 	cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
526 
527 	if (!has_mp) {
528 		cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
529 		cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
530 		cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
531 		c->booted_cores = 1;
532 		return;
533 	}
534 
535 	for_each_cpu(i, cpu_sibling_setup_mask) {
536 		o = &cpu_data(i);
537 
538 		if ((i == cpu) || (has_smt && match_smt(c, o)))
539 			link_mask(topology_sibling_cpumask, cpu, i);
540 
541 		if ((i == cpu) || (has_mp && match_llc(c, o)))
542 			link_mask(cpu_llc_shared_mask, cpu, i);
543 
544 	}
545 
546 	/*
547 	 * This needs a separate iteration over the cpus because we rely on all
548 	 * topology_sibling_cpumask links to be set-up.
549 	 */
550 	for_each_cpu(i, cpu_sibling_setup_mask) {
551 		o = &cpu_data(i);
552 
553 		if ((i == cpu) || (has_mp && match_die(c, o))) {
554 			link_mask(topology_core_cpumask, cpu, i);
555 
556 			/*
557 			 *  Does this new cpu bringup a new core?
558 			 */
559 			if (cpumask_weight(
560 			    topology_sibling_cpumask(cpu)) == 1) {
561 				/*
562 				 * for each core in package, increment
563 				 * the booted_cores for this new cpu
564 				 */
565 				if (cpumask_first(
566 				    topology_sibling_cpumask(i)) == i)
567 					c->booted_cores++;
568 				/*
569 				 * increment the core count for all
570 				 * the other cpus in this package
571 				 */
572 				if (i != cpu)
573 					cpu_data(i).booted_cores++;
574 			} else if (i != cpu && !c->booted_cores)
575 				c->booted_cores = cpu_data(i).booted_cores;
576 		}
577 		if (match_die(c, o) && !topology_same_node(c, o))
578 			x86_has_numa_in_package = true;
579 	}
580 
581 	threads = cpumask_weight(topology_sibling_cpumask(cpu));
582 	if (threads > __max_smt_threads)
583 		__max_smt_threads = threads;
584 }
585 
586 /* maps the cpu to the sched domain representing multi-core */
587 const struct cpumask *cpu_coregroup_mask(int cpu)
588 {
589 	return cpu_llc_shared_mask(cpu);
590 }
591 
592 static void impress_friends(void)
593 {
594 	int cpu;
595 	unsigned long bogosum = 0;
596 	/*
597 	 * Allow the user to impress friends.
598 	 */
599 	pr_debug("Before bogomips\n");
600 	for_each_possible_cpu(cpu)
601 		if (cpumask_test_cpu(cpu, cpu_callout_mask))
602 			bogosum += cpu_data(cpu).loops_per_jiffy;
603 	pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
604 		num_online_cpus(),
605 		bogosum/(500000/HZ),
606 		(bogosum/(5000/HZ))%100);
607 
608 	pr_debug("Before bogocount - setting activated=1\n");
609 }
610 
611 void __inquire_remote_apic(int apicid)
612 {
613 	unsigned i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
614 	const char * const names[] = { "ID", "VERSION", "SPIV" };
615 	int timeout;
616 	u32 status;
617 
618 	pr_info("Inquiring remote APIC 0x%x...\n", apicid);
619 
620 	for (i = 0; i < ARRAY_SIZE(regs); i++) {
621 		pr_info("... APIC 0x%x %s: ", apicid, names[i]);
622 
623 		/*
624 		 * Wait for idle.
625 		 */
626 		status = safe_apic_wait_icr_idle();
627 		if (status)
628 			pr_cont("a previous APIC delivery may have failed\n");
629 
630 		apic_icr_write(APIC_DM_REMRD | regs[i], apicid);
631 
632 		timeout = 0;
633 		do {
634 			udelay(100);
635 			status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
636 		} while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
637 
638 		switch (status) {
639 		case APIC_ICR_RR_VALID:
640 			status = apic_read(APIC_RRR);
641 			pr_cont("%08x\n", status);
642 			break;
643 		default:
644 			pr_cont("failed\n");
645 		}
646 	}
647 }
648 
649 /*
650  * The Multiprocessor Specification 1.4 (1997) example code suggests
651  * that there should be a 10ms delay between the BSP asserting INIT
652  * and de-asserting INIT, when starting a remote processor.
653  * But that slows boot and resume on modern processors, which include
654  * many cores and don't require that delay.
655  *
656  * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
657  * Modern processor families are quirked to remove the delay entirely.
658  */
659 #define UDELAY_10MS_DEFAULT 10000
660 
661 static unsigned int init_udelay = UINT_MAX;
662 
663 static int __init cpu_init_udelay(char *str)
664 {
665 	get_option(&str, &init_udelay);
666 
667 	return 0;
668 }
669 early_param("cpu_init_udelay", cpu_init_udelay);
670 
671 static void __init smp_quirk_init_udelay(void)
672 {
673 	/* if cmdline changed it from default, leave it alone */
674 	if (init_udelay != UINT_MAX)
675 		return;
676 
677 	/* if modern processor, use no delay */
678 	if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
679 	    ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
680 		init_udelay = 0;
681 		return;
682 	}
683 	/* else, use legacy delay */
684 	init_udelay = UDELAY_10MS_DEFAULT;
685 }
686 
687 /*
688  * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
689  * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
690  * won't ... remember to clear down the APIC, etc later.
691  */
692 int
693 wakeup_secondary_cpu_via_nmi(int apicid, unsigned long start_eip)
694 {
695 	unsigned long send_status, accept_status = 0;
696 	int maxlvt;
697 
698 	/* Target chip */
699 	/* Boot on the stack */
700 	/* Kick the second */
701 	apic_icr_write(APIC_DM_NMI | apic->dest_logical, apicid);
702 
703 	pr_debug("Waiting for send to finish...\n");
704 	send_status = safe_apic_wait_icr_idle();
705 
706 	/*
707 	 * Give the other CPU some time to accept the IPI.
708 	 */
709 	udelay(200);
710 	if (APIC_INTEGRATED(boot_cpu_apic_version)) {
711 		maxlvt = lapic_get_maxlvt();
712 		if (maxlvt > 3)			/* Due to the Pentium erratum 3AP.  */
713 			apic_write(APIC_ESR, 0);
714 		accept_status = (apic_read(APIC_ESR) & 0xEF);
715 	}
716 	pr_debug("NMI sent\n");
717 
718 	if (send_status)
719 		pr_err("APIC never delivered???\n");
720 	if (accept_status)
721 		pr_err("APIC delivery error (%lx)\n", accept_status);
722 
723 	return (send_status | accept_status);
724 }
725 
726 static int
727 wakeup_secondary_cpu_via_init(int phys_apicid, unsigned long start_eip)
728 {
729 	unsigned long send_status = 0, accept_status = 0;
730 	int maxlvt, num_starts, j;
731 
732 	maxlvt = lapic_get_maxlvt();
733 
734 	/*
735 	 * Be paranoid about clearing APIC errors.
736 	 */
737 	if (APIC_INTEGRATED(boot_cpu_apic_version)) {
738 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
739 			apic_write(APIC_ESR, 0);
740 		apic_read(APIC_ESR);
741 	}
742 
743 	pr_debug("Asserting INIT\n");
744 
745 	/*
746 	 * Turn INIT on target chip
747 	 */
748 	/*
749 	 * Send IPI
750 	 */
751 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT,
752 		       phys_apicid);
753 
754 	pr_debug("Waiting for send to finish...\n");
755 	send_status = safe_apic_wait_icr_idle();
756 
757 	udelay(init_udelay);
758 
759 	pr_debug("Deasserting INIT\n");
760 
761 	/* Target chip */
762 	/* Send IPI */
763 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
764 
765 	pr_debug("Waiting for send to finish...\n");
766 	send_status = safe_apic_wait_icr_idle();
767 
768 	mb();
769 
770 	/*
771 	 * Should we send STARTUP IPIs ?
772 	 *
773 	 * Determine this based on the APIC version.
774 	 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
775 	 */
776 	if (APIC_INTEGRATED(boot_cpu_apic_version))
777 		num_starts = 2;
778 	else
779 		num_starts = 0;
780 
781 	/*
782 	 * Run STARTUP IPI loop.
783 	 */
784 	pr_debug("#startup loops: %d\n", num_starts);
785 
786 	for (j = 1; j <= num_starts; j++) {
787 		pr_debug("Sending STARTUP #%d\n", j);
788 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
789 			apic_write(APIC_ESR, 0);
790 		apic_read(APIC_ESR);
791 		pr_debug("After apic_write\n");
792 
793 		/*
794 		 * STARTUP IPI
795 		 */
796 
797 		/* Target chip */
798 		/* Boot on the stack */
799 		/* Kick the second */
800 		apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
801 			       phys_apicid);
802 
803 		/*
804 		 * Give the other CPU some time to accept the IPI.
805 		 */
806 		if (init_udelay == 0)
807 			udelay(10);
808 		else
809 			udelay(300);
810 
811 		pr_debug("Startup point 1\n");
812 
813 		pr_debug("Waiting for send to finish...\n");
814 		send_status = safe_apic_wait_icr_idle();
815 
816 		/*
817 		 * Give the other CPU some time to accept the IPI.
818 		 */
819 		if (init_udelay == 0)
820 			udelay(10);
821 		else
822 			udelay(200);
823 
824 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
825 			apic_write(APIC_ESR, 0);
826 		accept_status = (apic_read(APIC_ESR) & 0xEF);
827 		if (send_status || accept_status)
828 			break;
829 	}
830 	pr_debug("After Startup\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 /* reduce the number of lines printed when booting a large cpu count system */
841 static void announce_cpu(int cpu, int apicid)
842 {
843 	static int current_node = -1;
844 	int node = early_cpu_to_node(cpu);
845 	static int width, node_width;
846 
847 	if (!width)
848 		width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
849 
850 	if (!node_width)
851 		node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
852 
853 	if (cpu == 1)
854 		printk(KERN_INFO "x86: Booting SMP configuration:\n");
855 
856 	if (system_state < SYSTEM_RUNNING) {
857 		if (node != current_node) {
858 			if (current_node > (-1))
859 				pr_cont("\n");
860 			current_node = node;
861 
862 			printk(KERN_INFO ".... node %*s#%d, CPUs:  ",
863 			       node_width - num_digits(node), " ", node);
864 		}
865 
866 		/* Add padding for the BSP */
867 		if (cpu == 1)
868 			pr_cont("%*s", width + 1, " ");
869 
870 		pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
871 
872 	} else
873 		pr_info("Booting Node %d Processor %d APIC 0x%x\n",
874 			node, cpu, apicid);
875 }
876 
877 static int wakeup_cpu0_nmi(unsigned int cmd, struct pt_regs *regs)
878 {
879 	int cpu;
880 
881 	cpu = smp_processor_id();
882 	if (cpu == 0 && !cpu_online(cpu) && enable_start_cpu0)
883 		return NMI_HANDLED;
884 
885 	return NMI_DONE;
886 }
887 
888 /*
889  * Wake up AP by INIT, INIT, STARTUP sequence.
890  *
891  * Instead of waiting for STARTUP after INITs, BSP will execute the BIOS
892  * boot-strap code which is not a desired behavior for waking up BSP. To
893  * void the boot-strap code, wake up CPU0 by NMI instead.
894  *
895  * This works to wake up soft offlined CPU0 only. If CPU0 is hard offlined
896  * (i.e. physically hot removed and then hot added), NMI won't wake it up.
897  * We'll change this code in the future to wake up hard offlined CPU0 if
898  * real platform and request are available.
899  */
900 static int
901 wakeup_cpu_via_init_nmi(int cpu, unsigned long start_ip, int apicid,
902 	       int *cpu0_nmi_registered)
903 {
904 	int id;
905 	int boot_error;
906 
907 	preempt_disable();
908 
909 	/*
910 	 * Wake up AP by INIT, INIT, STARTUP sequence.
911 	 */
912 	if (cpu) {
913 		boot_error = wakeup_secondary_cpu_via_init(apicid, start_ip);
914 		goto out;
915 	}
916 
917 	/*
918 	 * Wake up BSP by nmi.
919 	 *
920 	 * Register a NMI handler to help wake up CPU0.
921 	 */
922 	boot_error = register_nmi_handler(NMI_LOCAL,
923 					  wakeup_cpu0_nmi, 0, "wake_cpu0");
924 
925 	if (!boot_error) {
926 		enable_start_cpu0 = 1;
927 		*cpu0_nmi_registered = 1;
928 		if (apic->dest_logical == APIC_DEST_LOGICAL)
929 			id = cpu0_logical_apicid;
930 		else
931 			id = apicid;
932 		boot_error = wakeup_secondary_cpu_via_nmi(id, start_ip);
933 	}
934 
935 out:
936 	preempt_enable();
937 
938 	return boot_error;
939 }
940 
941 void common_cpu_up(unsigned int cpu, struct task_struct *idle)
942 {
943 	/* Just in case we booted with a single CPU. */
944 	alternatives_enable_smp();
945 
946 	per_cpu(current_task, cpu) = idle;
947 
948 #ifdef CONFIG_X86_32
949 	/* Stack for startup_32 can be just as for start_secondary onwards */
950 	irq_ctx_init(cpu);
951 	per_cpu(cpu_current_top_of_stack, cpu) = task_top_of_stack(idle);
952 #else
953 	initial_gs = per_cpu_offset(cpu);
954 #endif
955 }
956 
957 /*
958  * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
959  * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
960  * Returns zero if CPU booted OK, else error code from
961  * ->wakeup_secondary_cpu.
962  */
963 static int do_boot_cpu(int apicid, int cpu, struct task_struct *idle,
964 		       int *cpu0_nmi_registered)
965 {
966 	volatile u32 *trampoline_status =
967 		(volatile u32 *) __va(real_mode_header->trampoline_status);
968 	/* start_ip had better be page-aligned! */
969 	unsigned long start_ip = real_mode_header->trampoline_start;
970 
971 	unsigned long boot_error = 0;
972 	unsigned long timeout;
973 
974 	idle->thread.sp = (unsigned long)task_pt_regs(idle);
975 	early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
976 	initial_code = (unsigned long)start_secondary;
977 	initial_stack  = idle->thread.sp;
978 
979 	/* Enable the espfix hack for this CPU */
980 	init_espfix_ap(cpu);
981 
982 	/* So we see what's up */
983 	announce_cpu(cpu, apicid);
984 
985 	/*
986 	 * This grunge runs the startup process for
987 	 * the targeted processor.
988 	 */
989 
990 	if (x86_platform.legacy.warm_reset) {
991 
992 		pr_debug("Setting warm reset code and vector.\n");
993 
994 		smpboot_setup_warm_reset_vector(start_ip);
995 		/*
996 		 * Be paranoid about clearing APIC errors.
997 		*/
998 		if (APIC_INTEGRATED(boot_cpu_apic_version)) {
999 			apic_write(APIC_ESR, 0);
1000 			apic_read(APIC_ESR);
1001 		}
1002 	}
1003 
1004 	/*
1005 	 * AP might wait on cpu_callout_mask in cpu_init() with
1006 	 * cpu_initialized_mask set if previous attempt to online
1007 	 * it timed-out. Clear cpu_initialized_mask so that after
1008 	 * INIT/SIPI it could start with a clean state.
1009 	 */
1010 	cpumask_clear_cpu(cpu, cpu_initialized_mask);
1011 	smp_mb();
1012 
1013 	/*
1014 	 * Wake up a CPU in difference cases:
1015 	 * - Use the method in the APIC driver if it's defined
1016 	 * Otherwise,
1017 	 * - Use an INIT boot APIC message for APs or NMI for BSP.
1018 	 */
1019 	if (apic->wakeup_secondary_cpu)
1020 		boot_error = apic->wakeup_secondary_cpu(apicid, start_ip);
1021 	else
1022 		boot_error = wakeup_cpu_via_init_nmi(cpu, start_ip, apicid,
1023 						     cpu0_nmi_registered);
1024 
1025 	if (!boot_error) {
1026 		/*
1027 		 * Wait 10s total for first sign of life from AP
1028 		 */
1029 		boot_error = -1;
1030 		timeout = jiffies + 10*HZ;
1031 		while (time_before(jiffies, timeout)) {
1032 			if (cpumask_test_cpu(cpu, cpu_initialized_mask)) {
1033 				/*
1034 				 * Tell AP to proceed with initialization
1035 				 */
1036 				cpumask_set_cpu(cpu, cpu_callout_mask);
1037 				boot_error = 0;
1038 				break;
1039 			}
1040 			schedule();
1041 		}
1042 	}
1043 
1044 	if (!boot_error) {
1045 		/*
1046 		 * Wait till AP completes initial initialization
1047 		 */
1048 		while (!cpumask_test_cpu(cpu, cpu_callin_mask)) {
1049 			/*
1050 			 * Allow other tasks to run while we wait for the
1051 			 * AP to come online. This also gives a chance
1052 			 * for the MTRR work(triggered by the AP coming online)
1053 			 * to be completed in the stop machine context.
1054 			 */
1055 			schedule();
1056 		}
1057 	}
1058 
1059 	/* mark "stuck" area as not stuck */
1060 	*trampoline_status = 0;
1061 
1062 	if (x86_platform.legacy.warm_reset) {
1063 		/*
1064 		 * Cleanup possible dangling ends...
1065 		 */
1066 		smpboot_restore_warm_reset_vector();
1067 	}
1068 
1069 	return boot_error;
1070 }
1071 
1072 int native_cpu_up(unsigned int cpu, struct task_struct *tidle)
1073 {
1074 	int apicid = apic->cpu_present_to_apicid(cpu);
1075 	int cpu0_nmi_registered = 0;
1076 	unsigned long flags;
1077 	int err, ret = 0;
1078 
1079 	lockdep_assert_irqs_enabled();
1080 
1081 	pr_debug("++++++++++++++++++++=_---CPU UP  %u\n", cpu);
1082 
1083 	if (apicid == BAD_APICID ||
1084 	    !physid_isset(apicid, phys_cpu_present_map) ||
1085 	    !apic->apic_id_valid(apicid)) {
1086 		pr_err("%s: bad cpu %d\n", __func__, cpu);
1087 		return -EINVAL;
1088 	}
1089 
1090 	/*
1091 	 * Already booted CPU?
1092 	 */
1093 	if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
1094 		pr_debug("do_boot_cpu %d Already started\n", cpu);
1095 		return -ENOSYS;
1096 	}
1097 
1098 	/*
1099 	 * Save current MTRR state in case it was changed since early boot
1100 	 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1101 	 */
1102 	mtrr_save_state();
1103 
1104 	/* x86 CPUs take themselves offline, so delayed offline is OK. */
1105 	err = cpu_check_up_prepare(cpu);
1106 	if (err && err != -EBUSY)
1107 		return err;
1108 
1109 	/* the FPU context is blank, nobody can own it */
1110 	per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
1111 
1112 	common_cpu_up(cpu, tidle);
1113 
1114 	err = do_boot_cpu(apicid, cpu, tidle, &cpu0_nmi_registered);
1115 	if (err) {
1116 		pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
1117 		ret = -EIO;
1118 		goto unreg_nmi;
1119 	}
1120 
1121 	/*
1122 	 * Check TSC synchronization with the AP (keep irqs disabled
1123 	 * while doing so):
1124 	 */
1125 	local_irq_save(flags);
1126 	check_tsc_sync_source(cpu);
1127 	local_irq_restore(flags);
1128 
1129 	while (!cpu_online(cpu)) {
1130 		cpu_relax();
1131 		touch_nmi_watchdog();
1132 	}
1133 
1134 unreg_nmi:
1135 	/*
1136 	 * Clean up the nmi handler. Do this after the callin and callout sync
1137 	 * to avoid impact of possible long unregister time.
1138 	 */
1139 	if (cpu0_nmi_registered)
1140 		unregister_nmi_handler(NMI_LOCAL, "wake_cpu0");
1141 
1142 	return ret;
1143 }
1144 
1145 /**
1146  * arch_disable_smp_support() - disables SMP support for x86 at runtime
1147  */
1148 void arch_disable_smp_support(void)
1149 {
1150 	disable_ioapic_support();
1151 }
1152 
1153 /*
1154  * Fall back to non SMP mode after errors.
1155  *
1156  * RED-PEN audit/test this more. I bet there is more state messed up here.
1157  */
1158 static __init void disable_smp(void)
1159 {
1160 	pr_info("SMP disabled\n");
1161 
1162 	disable_ioapic_support();
1163 
1164 	init_cpu_present(cpumask_of(0));
1165 	init_cpu_possible(cpumask_of(0));
1166 
1167 	if (smp_found_config)
1168 		physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
1169 	else
1170 		physid_set_mask_of_physid(0, &phys_cpu_present_map);
1171 	cpumask_set_cpu(0, topology_sibling_cpumask(0));
1172 	cpumask_set_cpu(0, topology_core_cpumask(0));
1173 }
1174 
1175 /*
1176  * Various sanity checks.
1177  */
1178 static void __init smp_sanity_check(void)
1179 {
1180 	preempt_disable();
1181 
1182 #if !defined(CONFIG_X86_BIGSMP) && defined(CONFIG_X86_32)
1183 	if (def_to_bigsmp && nr_cpu_ids > 8) {
1184 		unsigned int cpu;
1185 		unsigned nr;
1186 
1187 		pr_warn("More than 8 CPUs detected - skipping them\n"
1188 			"Use CONFIG_X86_BIGSMP\n");
1189 
1190 		nr = 0;
1191 		for_each_present_cpu(cpu) {
1192 			if (nr >= 8)
1193 				set_cpu_present(cpu, false);
1194 			nr++;
1195 		}
1196 
1197 		nr = 0;
1198 		for_each_possible_cpu(cpu) {
1199 			if (nr >= 8)
1200 				set_cpu_possible(cpu, false);
1201 			nr++;
1202 		}
1203 
1204 		nr_cpu_ids = 8;
1205 	}
1206 #endif
1207 
1208 	if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map)) {
1209 		pr_warn("weird, boot CPU (#%d) not listed by the BIOS\n",
1210 			hard_smp_processor_id());
1211 
1212 		physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1213 	}
1214 
1215 	/*
1216 	 * Should not be necessary because the MP table should list the boot
1217 	 * CPU too, but we do it for the sake of robustness anyway.
1218 	 */
1219 	if (!apic->check_phys_apicid_present(boot_cpu_physical_apicid)) {
1220 		pr_notice("weird, boot CPU (#%d) not listed by the BIOS\n",
1221 			  boot_cpu_physical_apicid);
1222 		physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1223 	}
1224 	preempt_enable();
1225 }
1226 
1227 static void __init smp_cpu_index_default(void)
1228 {
1229 	int i;
1230 	struct cpuinfo_x86 *c;
1231 
1232 	for_each_possible_cpu(i) {
1233 		c = &cpu_data(i);
1234 		/* mark all to hotplug */
1235 		c->cpu_index = nr_cpu_ids;
1236 	}
1237 }
1238 
1239 static void __init smp_get_logical_apicid(void)
1240 {
1241 	if (x2apic_mode)
1242 		cpu0_logical_apicid = apic_read(APIC_LDR);
1243 	else
1244 		cpu0_logical_apicid = GET_APIC_LOGICAL_ID(apic_read(APIC_LDR));
1245 }
1246 
1247 /*
1248  * Prepare for SMP bootup.
1249  * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1250  *            for common interface support.
1251  */
1252 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1253 {
1254 	unsigned int i;
1255 
1256 	smp_cpu_index_default();
1257 
1258 	/*
1259 	 * Setup boot CPU information
1260 	 */
1261 	smp_store_boot_cpu_info(); /* Final full version of the data */
1262 	cpumask_copy(cpu_callin_mask, cpumask_of(0));
1263 	mb();
1264 
1265 	for_each_possible_cpu(i) {
1266 		zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1267 		zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1268 		zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1269 	}
1270 
1271 	/*
1272 	 * Set 'default' x86 topology, this matches default_topology() in that
1273 	 * it has NUMA nodes as a topology level. See also
1274 	 * native_smp_cpus_done().
1275 	 *
1276 	 * Must be done before set_cpus_sibling_map() is ran.
1277 	 */
1278 	set_sched_topology(x86_topology);
1279 
1280 	set_cpu_sibling_map(0);
1281 
1282 	smp_sanity_check();
1283 
1284 	switch (apic_intr_mode) {
1285 	case APIC_PIC:
1286 	case APIC_VIRTUAL_WIRE_NO_CONFIG:
1287 		disable_smp();
1288 		return;
1289 	case APIC_SYMMETRIC_IO_NO_ROUTING:
1290 		disable_smp();
1291 		/* Setup local timer */
1292 		x86_init.timers.setup_percpu_clockev();
1293 		return;
1294 	case APIC_VIRTUAL_WIRE:
1295 	case APIC_SYMMETRIC_IO:
1296 		break;
1297 	}
1298 
1299 	/* Setup local timer */
1300 	x86_init.timers.setup_percpu_clockev();
1301 
1302 	smp_get_logical_apicid();
1303 
1304 	pr_info("CPU0: ");
1305 	print_cpu_info(&cpu_data(0));
1306 
1307 	native_pv_lock_init();
1308 
1309 	uv_system_init();
1310 
1311 	set_mtrr_aps_delayed_init();
1312 
1313 	smp_quirk_init_udelay();
1314 
1315 	speculative_store_bypass_ht_init();
1316 }
1317 
1318 void arch_enable_nonboot_cpus_begin(void)
1319 {
1320 	set_mtrr_aps_delayed_init();
1321 }
1322 
1323 void arch_enable_nonboot_cpus_end(void)
1324 {
1325 	mtrr_aps_init();
1326 }
1327 
1328 /*
1329  * Early setup to make printk work.
1330  */
1331 void __init native_smp_prepare_boot_cpu(void)
1332 {
1333 	int me = smp_processor_id();
1334 	switch_to_new_gdt(me);
1335 	/* already set me in cpu_online_mask in boot_cpu_init() */
1336 	cpumask_set_cpu(me, cpu_callout_mask);
1337 	cpu_set_state_online(me);
1338 }
1339 
1340 void __init calculate_max_logical_packages(void)
1341 {
1342 	int ncpus;
1343 
1344 	/*
1345 	 * Today neither Intel nor AMD support heterogenous systems so
1346 	 * extrapolate the boot cpu's data to all packages.
1347 	 */
1348 	ncpus = cpu_data(0).booted_cores * topology_max_smt_threads();
1349 	__max_logical_packages = DIV_ROUND_UP(nr_cpu_ids, ncpus);
1350 	pr_info("Max logical packages: %u\n", __max_logical_packages);
1351 }
1352 
1353 void __init native_smp_cpus_done(unsigned int max_cpus)
1354 {
1355 	pr_debug("Boot done\n");
1356 
1357 	calculate_max_logical_packages();
1358 
1359 	if (x86_has_numa_in_package)
1360 		set_sched_topology(x86_numa_in_package_topology);
1361 
1362 	nmi_selftest();
1363 	impress_friends();
1364 	mtrr_aps_init();
1365 }
1366 
1367 static int __initdata setup_possible_cpus = -1;
1368 static int __init _setup_possible_cpus(char *str)
1369 {
1370 	get_option(&str, &setup_possible_cpus);
1371 	return 0;
1372 }
1373 early_param("possible_cpus", _setup_possible_cpus);
1374 
1375 
1376 /*
1377  * cpu_possible_mask should be static, it cannot change as cpu's
1378  * are onlined, or offlined. The reason is per-cpu data-structures
1379  * are allocated by some modules at init time, and dont expect to
1380  * do this dynamically on cpu arrival/departure.
1381  * cpu_present_mask on the other hand can change dynamically.
1382  * In case when cpu_hotplug is not compiled, then we resort to current
1383  * behaviour, which is cpu_possible == cpu_present.
1384  * - Ashok Raj
1385  *
1386  * Three ways to find out the number of additional hotplug CPUs:
1387  * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1388  * - The user can overwrite it with possible_cpus=NUM
1389  * - Otherwise don't reserve additional CPUs.
1390  * We do this because additional CPUs waste a lot of memory.
1391  * -AK
1392  */
1393 __init void prefill_possible_map(void)
1394 {
1395 	int i, possible;
1396 
1397 	/* No boot processor was found in mptable or ACPI MADT */
1398 	if (!num_processors) {
1399 		if (boot_cpu_has(X86_FEATURE_APIC)) {
1400 			int apicid = boot_cpu_physical_apicid;
1401 			int cpu = hard_smp_processor_id();
1402 
1403 			pr_warn("Boot CPU (id %d) not listed by BIOS\n", cpu);
1404 
1405 			/* Make sure boot cpu is enumerated */
1406 			if (apic->cpu_present_to_apicid(0) == BAD_APICID &&
1407 			    apic->apic_id_valid(apicid))
1408 				generic_processor_info(apicid, boot_cpu_apic_version);
1409 		}
1410 
1411 		if (!num_processors)
1412 			num_processors = 1;
1413 	}
1414 
1415 	i = setup_max_cpus ?: 1;
1416 	if (setup_possible_cpus == -1) {
1417 		possible = num_processors;
1418 #ifdef CONFIG_HOTPLUG_CPU
1419 		if (setup_max_cpus)
1420 			possible += disabled_cpus;
1421 #else
1422 		if (possible > i)
1423 			possible = i;
1424 #endif
1425 	} else
1426 		possible = setup_possible_cpus;
1427 
1428 	total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1429 
1430 	/* nr_cpu_ids could be reduced via nr_cpus= */
1431 	if (possible > nr_cpu_ids) {
1432 		pr_warn("%d Processors exceeds NR_CPUS limit of %u\n",
1433 			possible, nr_cpu_ids);
1434 		possible = nr_cpu_ids;
1435 	}
1436 
1437 #ifdef CONFIG_HOTPLUG_CPU
1438 	if (!setup_max_cpus)
1439 #endif
1440 	if (possible > i) {
1441 		pr_warn("%d Processors exceeds max_cpus limit of %u\n",
1442 			possible, setup_max_cpus);
1443 		possible = i;
1444 	}
1445 
1446 	nr_cpu_ids = possible;
1447 
1448 	pr_info("Allowing %d CPUs, %d hotplug CPUs\n",
1449 		possible, max_t(int, possible - num_processors, 0));
1450 
1451 	reset_cpu_possible_mask();
1452 
1453 	for (i = 0; i < possible; i++)
1454 		set_cpu_possible(i, true);
1455 }
1456 
1457 #ifdef CONFIG_HOTPLUG_CPU
1458 
1459 /* Recompute SMT state for all CPUs on offline */
1460 static void recompute_smt_state(void)
1461 {
1462 	int max_threads, cpu;
1463 
1464 	max_threads = 0;
1465 	for_each_online_cpu (cpu) {
1466 		int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1467 
1468 		if (threads > max_threads)
1469 			max_threads = threads;
1470 	}
1471 	__max_smt_threads = max_threads;
1472 }
1473 
1474 static void remove_siblinginfo(int cpu)
1475 {
1476 	int sibling;
1477 	struct cpuinfo_x86 *c = &cpu_data(cpu);
1478 
1479 	for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1480 		cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1481 		/*/
1482 		 * last thread sibling in this cpu core going down
1483 		 */
1484 		if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1485 			cpu_data(sibling).booted_cores--;
1486 	}
1487 
1488 	for_each_cpu(sibling, topology_sibling_cpumask(cpu))
1489 		cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1490 	for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1491 		cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1492 	cpumask_clear(cpu_llc_shared_mask(cpu));
1493 	cpumask_clear(topology_sibling_cpumask(cpu));
1494 	cpumask_clear(topology_core_cpumask(cpu));
1495 	c->cpu_core_id = 0;
1496 	c->booted_cores = 0;
1497 	cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1498 	recompute_smt_state();
1499 }
1500 
1501 static void remove_cpu_from_maps(int cpu)
1502 {
1503 	set_cpu_online(cpu, false);
1504 	cpumask_clear_cpu(cpu, cpu_callout_mask);
1505 	cpumask_clear_cpu(cpu, cpu_callin_mask);
1506 	/* was set by cpu_init() */
1507 	cpumask_clear_cpu(cpu, cpu_initialized_mask);
1508 	numa_remove_cpu(cpu);
1509 }
1510 
1511 void cpu_disable_common(void)
1512 {
1513 	int cpu = smp_processor_id();
1514 
1515 	remove_siblinginfo(cpu);
1516 
1517 	/* It's now safe to remove this processor from the online map */
1518 	lock_vector_lock();
1519 	remove_cpu_from_maps(cpu);
1520 	unlock_vector_lock();
1521 	fixup_irqs();
1522 	lapic_offline();
1523 }
1524 
1525 int native_cpu_disable(void)
1526 {
1527 	int ret;
1528 
1529 	ret = lapic_can_unplug_cpu();
1530 	if (ret)
1531 		return ret;
1532 
1533 	clear_local_APIC();
1534 	cpu_disable_common();
1535 
1536 	return 0;
1537 }
1538 
1539 int common_cpu_die(unsigned int cpu)
1540 {
1541 	int ret = 0;
1542 
1543 	/* We don't do anything here: idle task is faking death itself. */
1544 
1545 	/* They ack this in play_dead() by setting CPU_DEAD */
1546 	if (cpu_wait_death(cpu, 5)) {
1547 		if (system_state == SYSTEM_RUNNING)
1548 			pr_info("CPU %u is now offline\n", cpu);
1549 	} else {
1550 		pr_err("CPU %u didn't die...\n", cpu);
1551 		ret = -1;
1552 	}
1553 
1554 	return ret;
1555 }
1556 
1557 void native_cpu_die(unsigned int cpu)
1558 {
1559 	common_cpu_die(cpu);
1560 }
1561 
1562 void play_dead_common(void)
1563 {
1564 	idle_task_exit();
1565 
1566 	/* Ack it */
1567 	(void)cpu_report_death();
1568 
1569 	/*
1570 	 * With physical CPU hotplug, we should halt the cpu
1571 	 */
1572 	local_irq_disable();
1573 }
1574 
1575 static bool wakeup_cpu0(void)
1576 {
1577 	if (smp_processor_id() == 0 && enable_start_cpu0)
1578 		return true;
1579 
1580 	return false;
1581 }
1582 
1583 /*
1584  * We need to flush the caches before going to sleep, lest we have
1585  * dirty data in our caches when we come back up.
1586  */
1587 static inline void mwait_play_dead(void)
1588 {
1589 	unsigned int eax, ebx, ecx, edx;
1590 	unsigned int highest_cstate = 0;
1591 	unsigned int highest_subcstate = 0;
1592 	void *mwait_ptr;
1593 	int i;
1594 
1595 	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
1596 		return;
1597 	if (!this_cpu_has(X86_FEATURE_MWAIT))
1598 		return;
1599 	if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1600 		return;
1601 	if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1602 		return;
1603 
1604 	eax = CPUID_MWAIT_LEAF;
1605 	ecx = 0;
1606 	native_cpuid(&eax, &ebx, &ecx, &edx);
1607 
1608 	/*
1609 	 * eax will be 0 if EDX enumeration is not valid.
1610 	 * Initialized below to cstate, sub_cstate value when EDX is valid.
1611 	 */
1612 	if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1613 		eax = 0;
1614 	} else {
1615 		edx >>= MWAIT_SUBSTATE_SIZE;
1616 		for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1617 			if (edx & MWAIT_SUBSTATE_MASK) {
1618 				highest_cstate = i;
1619 				highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1620 			}
1621 		}
1622 		eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1623 			(highest_subcstate - 1);
1624 	}
1625 
1626 	/*
1627 	 * This should be a memory location in a cache line which is
1628 	 * unlikely to be touched by other processors.  The actual
1629 	 * content is immaterial as it is not actually modified in any way.
1630 	 */
1631 	mwait_ptr = &current_thread_info()->flags;
1632 
1633 	wbinvd();
1634 
1635 	while (1) {
1636 		/*
1637 		 * The CLFLUSH is a workaround for erratum AAI65 for
1638 		 * the Xeon 7400 series.  It's not clear it is actually
1639 		 * needed, but it should be harmless in either case.
1640 		 * The WBINVD is insufficient due to the spurious-wakeup
1641 		 * case where we return around the loop.
1642 		 */
1643 		mb();
1644 		clflush(mwait_ptr);
1645 		mb();
1646 		__monitor(mwait_ptr, 0, 0);
1647 		mb();
1648 		__mwait(eax, 0);
1649 		/*
1650 		 * If NMI wants to wake up CPU0, start CPU0.
1651 		 */
1652 		if (wakeup_cpu0())
1653 			start_cpu0();
1654 	}
1655 }
1656 
1657 void hlt_play_dead(void)
1658 {
1659 	if (__this_cpu_read(cpu_info.x86) >= 4)
1660 		wbinvd();
1661 
1662 	while (1) {
1663 		native_halt();
1664 		/*
1665 		 * If NMI wants to wake up CPU0, start CPU0.
1666 		 */
1667 		if (wakeup_cpu0())
1668 			start_cpu0();
1669 	}
1670 }
1671 
1672 void native_play_dead(void)
1673 {
1674 	play_dead_common();
1675 	tboot_shutdown(TB_SHUTDOWN_WFS);
1676 
1677 	mwait_play_dead();	/* Only returns on failure */
1678 	if (cpuidle_play_dead())
1679 		hlt_play_dead();
1680 }
1681 
1682 #else /* ... !CONFIG_HOTPLUG_CPU */
1683 int native_cpu_disable(void)
1684 {
1685 	return -ENOSYS;
1686 }
1687 
1688 void native_cpu_die(unsigned int cpu)
1689 {
1690 	/* We said "no" in __cpu_disable */
1691 	BUG();
1692 }
1693 
1694 void native_play_dead(void)
1695 {
1696 	BUG();
1697 }
1698 
1699 #endif
1700