xref: /linux/arch/x86/kernel/smpboot.c (revision 372e2db7210df7c45ead46429aeb1443ba148060)
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/percpu.h>
49 #include <linux/bootmem.h>
50 #include <linux/err.h>
51 #include <linux/nmi.h>
52 #include <linux/tboot.h>
53 #include <linux/stackprotector.h>
54 #include <linux/gfp.h>
55 #include <linux/cpuidle.h>
56 
57 #include <asm/acpi.h>
58 #include <asm/desc.h>
59 #include <asm/nmi.h>
60 #include <asm/irq.h>
61 #include <asm/realmode.h>
62 #include <asm/cpu.h>
63 #include <asm/numa.h>
64 #include <asm/pgtable.h>
65 #include <asm/tlbflush.h>
66 #include <asm/mtrr.h>
67 #include <asm/mwait.h>
68 #include <asm/apic.h>
69 #include <asm/io_apic.h>
70 #include <asm/fpu/internal.h>
71 #include <asm/setup.h>
72 #include <asm/uv/uv.h>
73 #include <linux/mc146818rtc.h>
74 #include <asm/i8259.h>
75 #include <asm/realmode.h>
76 #include <asm/misc.h>
77 
78 /* Number of siblings per CPU package */
79 int smp_num_siblings = 1;
80 EXPORT_SYMBOL(smp_num_siblings);
81 
82 /* Last level cache ID of each logical CPU */
83 DEFINE_PER_CPU_READ_MOSTLY(u16, cpu_llc_id) = BAD_APICID;
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 static int *physical_to_logical_pkg __read_mostly;
101 static unsigned long *physical_package_map __read_mostly;;
102 static unsigned int max_physical_pkg_id __read_mostly;
103 unsigned int __max_logical_packages __read_mostly;
104 EXPORT_SYMBOL(__max_logical_packages);
105 static unsigned int logical_packages __read_mostly;
106 static bool logical_packages_frozen __read_mostly;
107 
108 /* Maximum number of SMT threads on any online core */
109 int __max_smt_threads __read_mostly;
110 
111 /* Flag to indicate if a complete sched domain rebuild is required */
112 bool x86_topology_update;
113 
114 int arch_update_cpu_topology(void)
115 {
116 	int retval = x86_topology_update;
117 
118 	x86_topology_update = false;
119 	return retval;
120 }
121 
122 static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
123 {
124 	unsigned long flags;
125 
126 	spin_lock_irqsave(&rtc_lock, flags);
127 	CMOS_WRITE(0xa, 0xf);
128 	spin_unlock_irqrestore(&rtc_lock, flags);
129 	local_flush_tlb();
130 	pr_debug("1.\n");
131 	*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) =
132 							start_eip >> 4;
133 	pr_debug("2.\n");
134 	*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) =
135 							start_eip & 0xf;
136 	pr_debug("3.\n");
137 }
138 
139 static inline void smpboot_restore_warm_reset_vector(void)
140 {
141 	unsigned long flags;
142 
143 	/*
144 	 * Install writable page 0 entry to set BIOS data area.
145 	 */
146 	local_flush_tlb();
147 
148 	/*
149 	 * Paranoid:  Set warm reset code and vector here back
150 	 * to default values.
151 	 */
152 	spin_lock_irqsave(&rtc_lock, flags);
153 	CMOS_WRITE(0, 0xf);
154 	spin_unlock_irqrestore(&rtc_lock, flags);
155 
156 	*((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
157 }
158 
159 /*
160  * Report back to the Boot Processor during boot time or to the caller processor
161  * during CPU online.
162  */
163 static void smp_callin(void)
164 {
165 	int cpuid, phys_id;
166 
167 	/*
168 	 * If waken up by an INIT in an 82489DX configuration
169 	 * cpu_callout_mask guarantees we don't get here before
170 	 * an INIT_deassert IPI reaches our local APIC, so it is
171 	 * now safe to touch our local APIC.
172 	 */
173 	cpuid = smp_processor_id();
174 
175 	/*
176 	 * (This works even if the APIC is not enabled.)
177 	 */
178 	phys_id = read_apic_id();
179 
180 	/*
181 	 * the boot CPU has finished the init stage and is spinning
182 	 * on callin_map until we finish. We are free to set up this
183 	 * CPU, first the APIC. (this is probably redundant on most
184 	 * boards)
185 	 */
186 	apic_ap_setup();
187 
188 	/*
189 	 * Save our processor parameters. Note: this information
190 	 * is needed for clock calibration.
191 	 */
192 	smp_store_cpu_info(cpuid);
193 
194 	/*
195 	 * Get our bogomips.
196 	 * Update loops_per_jiffy in cpu_data. Previous call to
197 	 * smp_store_cpu_info() stored a value that is close but not as
198 	 * accurate as the value just calculated.
199 	 */
200 	calibrate_delay();
201 	cpu_data(cpuid).loops_per_jiffy = loops_per_jiffy;
202 	pr_debug("Stack at about %p\n", &cpuid);
203 
204 	/*
205 	 * This must be done before setting cpu_online_mask
206 	 * or calling notify_cpu_starting.
207 	 */
208 	set_cpu_sibling_map(raw_smp_processor_id());
209 	wmb();
210 
211 	notify_cpu_starting(cpuid);
212 
213 	/*
214 	 * Allow the master to continue.
215 	 */
216 	cpumask_set_cpu(cpuid, cpu_callin_mask);
217 }
218 
219 static int cpu0_logical_apicid;
220 static int enable_start_cpu0;
221 /*
222  * Activate a secondary processor.
223  */
224 static void notrace start_secondary(void *unused)
225 {
226 	/*
227 	 * Don't put *anything* before cpu_init(), SMP booting is too
228 	 * fragile that we want to limit the things done here to the
229 	 * most necessary things.
230 	 */
231 	cpu_init();
232 	x86_cpuinit.early_percpu_clock_init();
233 	preempt_disable();
234 	smp_callin();
235 
236 	enable_start_cpu0 = 0;
237 
238 #ifdef CONFIG_X86_32
239 	/* switch away from the initial page table */
240 	load_cr3(swapper_pg_dir);
241 	__flush_tlb_all();
242 #endif
243 
244 	/* otherwise gcc will move up smp_processor_id before the cpu_init */
245 	barrier();
246 	/*
247 	 * Check TSC synchronization with the BP:
248 	 */
249 	check_tsc_sync_target();
250 
251 	/*
252 	 * Lock vector_lock and initialize the vectors on this cpu
253 	 * before setting the cpu online. We must set it online with
254 	 * vector_lock held to prevent a concurrent setup/teardown
255 	 * from seeing a half valid vector space.
256 	 */
257 	lock_vector_lock();
258 	setup_vector_irq(smp_processor_id());
259 	set_cpu_online(smp_processor_id(), true);
260 	unlock_vector_lock();
261 	cpu_set_state_online(smp_processor_id());
262 	x86_platform.nmi_init();
263 
264 	/* enable local interrupts */
265 	local_irq_enable();
266 
267 	/* to prevent fake stack check failure in clock setup */
268 	boot_init_stack_canary();
269 
270 	x86_cpuinit.setup_percpu_clockev();
271 
272 	wmb();
273 	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
274 }
275 
276 int topology_update_package_map(unsigned int apicid, unsigned int cpu)
277 {
278 	unsigned int new, pkg = apicid >> boot_cpu_data.x86_coreid_bits;
279 
280 	/* Called from early boot ? */
281 	if (!physical_package_map)
282 		return 0;
283 
284 	if (pkg >= max_physical_pkg_id)
285 		return -EINVAL;
286 
287 	/* Set the logical package id */
288 	if (test_and_set_bit(pkg, physical_package_map))
289 		goto found;
290 
291 	if (logical_packages_frozen) {
292 		physical_to_logical_pkg[pkg] = -1;
293 		pr_warn("APIC(%x) Package %u exceeds logical package max\n",
294 			apicid, pkg);
295 		return -ENOSPC;
296 	}
297 
298 	new = logical_packages++;
299 	pr_info("APIC(%x) Converting physical %u to logical package %u\n",
300 		apicid, pkg, new);
301 	physical_to_logical_pkg[pkg] = new;
302 
303 found:
304 	cpu_data(cpu).logical_proc_id = physical_to_logical_pkg[pkg];
305 	return 0;
306 }
307 
308 /**
309  * topology_phys_to_logical_pkg - Map a physical package id to a logical
310  *
311  * Returns logical package id or -1 if not found
312  */
313 int topology_phys_to_logical_pkg(unsigned int phys_pkg)
314 {
315 	if (phys_pkg >= max_physical_pkg_id)
316 		return -1;
317 	return physical_to_logical_pkg[phys_pkg];
318 }
319 EXPORT_SYMBOL(topology_phys_to_logical_pkg);
320 
321 static void __init smp_init_package_map(void)
322 {
323 	unsigned int ncpus, cpu;
324 	size_t size;
325 
326 	/*
327 	 * Today neither Intel nor AMD support heterogenous systems. That
328 	 * might change in the future....
329 	 *
330 	 * While ideally we'd want '* smp_num_siblings' in the below @ncpus
331 	 * computation, this won't actually work since some Intel BIOSes
332 	 * report inconsistent HT data when they disable HT.
333 	 *
334 	 * In particular, they reduce the APIC-IDs to only include the cores,
335 	 * but leave the CPUID topology to say there are (2) siblings.
336 	 * This means we don't know how many threads there will be until
337 	 * after the APIC enumeration.
338 	 *
339 	 * By not including this we'll sometimes over-estimate the number of
340 	 * logical packages by the amount of !present siblings, but this is
341 	 * still better than MAX_LOCAL_APIC.
342 	 *
343 	 * We use total_cpus not nr_cpu_ids because nr_cpu_ids can be limited
344 	 * on the command line leading to a similar issue as the HT disable
345 	 * problem because the hyperthreads are usually enumerated after the
346 	 * primary cores.
347 	 */
348 	ncpus = boot_cpu_data.x86_max_cores;
349 	if (!ncpus) {
350 		pr_warn("x86_max_cores == zero !?!?");
351 		ncpus = 1;
352 	}
353 
354 	__max_logical_packages = DIV_ROUND_UP(total_cpus, ncpus);
355 	logical_packages = 0;
356 
357 	/*
358 	 * Possibly larger than what we need as the number of apic ids per
359 	 * package can be smaller than the actual used apic ids.
360 	 */
361 	max_physical_pkg_id = DIV_ROUND_UP(MAX_LOCAL_APIC, ncpus);
362 	size = max_physical_pkg_id * sizeof(unsigned int);
363 	physical_to_logical_pkg = kmalloc(size, GFP_KERNEL);
364 	memset(physical_to_logical_pkg, 0xff, size);
365 	size = BITS_TO_LONGS(max_physical_pkg_id) * sizeof(unsigned long);
366 	physical_package_map = kzalloc(size, GFP_KERNEL);
367 
368 	for_each_present_cpu(cpu) {
369 		unsigned int apicid = apic->cpu_present_to_apicid(cpu);
370 
371 		if (apicid == BAD_APICID || !apic->apic_id_valid(apicid))
372 			continue;
373 		if (!topology_update_package_map(apicid, cpu))
374 			continue;
375 		pr_warn("CPU %u APICId %x disabled\n", cpu, apicid);
376 		per_cpu(x86_bios_cpu_apicid, cpu) = BAD_APICID;
377 		set_cpu_possible(cpu, false);
378 		set_cpu_present(cpu, false);
379 	}
380 
381 	if (logical_packages > __max_logical_packages) {
382 		pr_warn("Detected more packages (%u), then computed by BIOS data (%u).\n",
383 			logical_packages, __max_logical_packages);
384 		logical_packages_frozen = true;
385 		__max_logical_packages  = logical_packages;
386 	}
387 
388 	pr_info("Max logical packages: %u\n", __max_logical_packages);
389 }
390 
391 void __init smp_store_boot_cpu_info(void)
392 {
393 	int id = 0; /* CPU 0 */
394 	struct cpuinfo_x86 *c = &cpu_data(id);
395 
396 	*c = boot_cpu_data;
397 	c->cpu_index = id;
398 	smp_init_package_map();
399 }
400 
401 /*
402  * The bootstrap kernel entry code has set these up. Save them for
403  * a given CPU
404  */
405 void smp_store_cpu_info(int id)
406 {
407 	struct cpuinfo_x86 *c = &cpu_data(id);
408 
409 	*c = boot_cpu_data;
410 	c->cpu_index = id;
411 	/*
412 	 * During boot time, CPU0 has this setup already. Save the info when
413 	 * bringing up AP or offlined CPU0.
414 	 */
415 	identify_secondary_cpu(c);
416 }
417 
418 static bool
419 topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
420 {
421 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
422 
423 	return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
424 }
425 
426 static bool
427 topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
428 {
429 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
430 
431 	return !WARN_ONCE(!topology_same_node(c, o),
432 		"sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
433 		"[node: %d != %d]. Ignoring dependency.\n",
434 		cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
435 }
436 
437 #define link_mask(mfunc, c1, c2)					\
438 do {									\
439 	cpumask_set_cpu((c1), mfunc(c2));				\
440 	cpumask_set_cpu((c2), mfunc(c1));				\
441 } while (0)
442 
443 static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
444 {
445 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
446 		int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
447 
448 		if (c->phys_proc_id == o->phys_proc_id &&
449 		    per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2) &&
450 		    c->cpu_core_id == o->cpu_core_id)
451 			return topology_sane(c, o, "smt");
452 
453 	} else if (c->phys_proc_id == o->phys_proc_id &&
454 		   c->cpu_core_id == o->cpu_core_id) {
455 		return topology_sane(c, o, "smt");
456 	}
457 
458 	return false;
459 }
460 
461 static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
462 {
463 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
464 
465 	if (per_cpu(cpu_llc_id, cpu1) != BAD_APICID &&
466 	    per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2))
467 		return topology_sane(c, o, "llc");
468 
469 	return false;
470 }
471 
472 /*
473  * Unlike the other levels, we do not enforce keeping a
474  * multicore group inside a NUMA node.  If this happens, we will
475  * discard the MC level of the topology later.
476  */
477 static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
478 {
479 	if (c->phys_proc_id == o->phys_proc_id)
480 		return true;
481 	return false;
482 }
483 
484 #if defined(CONFIG_SCHED_SMT) || defined(CONFIG_SCHED_MC)
485 static inline int x86_sched_itmt_flags(void)
486 {
487 	return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
488 }
489 
490 #ifdef CONFIG_SCHED_MC
491 static int x86_core_flags(void)
492 {
493 	return cpu_core_flags() | x86_sched_itmt_flags();
494 }
495 #endif
496 #ifdef CONFIG_SCHED_SMT
497 static int x86_smt_flags(void)
498 {
499 	return cpu_smt_flags() | x86_sched_itmt_flags();
500 }
501 #endif
502 #endif
503 
504 static struct sched_domain_topology_level x86_numa_in_package_topology[] = {
505 #ifdef CONFIG_SCHED_SMT
506 	{ cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
507 #endif
508 #ifdef CONFIG_SCHED_MC
509 	{ cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
510 #endif
511 	{ NULL, },
512 };
513 
514 static struct sched_domain_topology_level x86_topology[] = {
515 #ifdef CONFIG_SCHED_SMT
516 	{ cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
517 #endif
518 #ifdef CONFIG_SCHED_MC
519 	{ cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
520 #endif
521 	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
522 	{ NULL, },
523 };
524 
525 /*
526  * Set if a package/die has multiple NUMA nodes inside.
527  * AMD Magny-Cours and Intel Cluster-on-Die have this.
528  */
529 static bool x86_has_numa_in_package;
530 
531 void set_cpu_sibling_map(int cpu)
532 {
533 	bool has_smt = smp_num_siblings > 1;
534 	bool has_mp = has_smt || boot_cpu_data.x86_max_cores > 1;
535 	struct cpuinfo_x86 *c = &cpu_data(cpu);
536 	struct cpuinfo_x86 *o;
537 	int i, threads;
538 
539 	cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
540 
541 	if (!has_mp) {
542 		cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
543 		cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
544 		cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
545 		c->booted_cores = 1;
546 		return;
547 	}
548 
549 	for_each_cpu(i, cpu_sibling_setup_mask) {
550 		o = &cpu_data(i);
551 
552 		if ((i == cpu) || (has_smt && match_smt(c, o)))
553 			link_mask(topology_sibling_cpumask, cpu, i);
554 
555 		if ((i == cpu) || (has_mp && match_llc(c, o)))
556 			link_mask(cpu_llc_shared_mask, cpu, i);
557 
558 	}
559 
560 	/*
561 	 * This needs a separate iteration over the cpus because we rely on all
562 	 * topology_sibling_cpumask links to be set-up.
563 	 */
564 	for_each_cpu(i, cpu_sibling_setup_mask) {
565 		o = &cpu_data(i);
566 
567 		if ((i == cpu) || (has_mp && match_die(c, o))) {
568 			link_mask(topology_core_cpumask, cpu, i);
569 
570 			/*
571 			 *  Does this new cpu bringup a new core?
572 			 */
573 			if (cpumask_weight(
574 			    topology_sibling_cpumask(cpu)) == 1) {
575 				/*
576 				 * for each core in package, increment
577 				 * the booted_cores for this new cpu
578 				 */
579 				if (cpumask_first(
580 				    topology_sibling_cpumask(i)) == i)
581 					c->booted_cores++;
582 				/*
583 				 * increment the core count for all
584 				 * the other cpus in this package
585 				 */
586 				if (i != cpu)
587 					cpu_data(i).booted_cores++;
588 			} else if (i != cpu && !c->booted_cores)
589 				c->booted_cores = cpu_data(i).booted_cores;
590 		}
591 		if (match_die(c, o) && !topology_same_node(c, o))
592 			x86_has_numa_in_package = true;
593 	}
594 
595 	threads = cpumask_weight(topology_sibling_cpumask(cpu));
596 	if (threads > __max_smt_threads)
597 		__max_smt_threads = threads;
598 }
599 
600 /* maps the cpu to the sched domain representing multi-core */
601 const struct cpumask *cpu_coregroup_mask(int cpu)
602 {
603 	return cpu_llc_shared_mask(cpu);
604 }
605 
606 static void impress_friends(void)
607 {
608 	int cpu;
609 	unsigned long bogosum = 0;
610 	/*
611 	 * Allow the user to impress friends.
612 	 */
613 	pr_debug("Before bogomips\n");
614 	for_each_possible_cpu(cpu)
615 		if (cpumask_test_cpu(cpu, cpu_callout_mask))
616 			bogosum += cpu_data(cpu).loops_per_jiffy;
617 	pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
618 		num_online_cpus(),
619 		bogosum/(500000/HZ),
620 		(bogosum/(5000/HZ))%100);
621 
622 	pr_debug("Before bogocount - setting activated=1\n");
623 }
624 
625 void __inquire_remote_apic(int apicid)
626 {
627 	unsigned i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
628 	const char * const names[] = { "ID", "VERSION", "SPIV" };
629 	int timeout;
630 	u32 status;
631 
632 	pr_info("Inquiring remote APIC 0x%x...\n", apicid);
633 
634 	for (i = 0; i < ARRAY_SIZE(regs); i++) {
635 		pr_info("... APIC 0x%x %s: ", apicid, names[i]);
636 
637 		/*
638 		 * Wait for idle.
639 		 */
640 		status = safe_apic_wait_icr_idle();
641 		if (status)
642 			pr_cont("a previous APIC delivery may have failed\n");
643 
644 		apic_icr_write(APIC_DM_REMRD | regs[i], apicid);
645 
646 		timeout = 0;
647 		do {
648 			udelay(100);
649 			status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
650 		} while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
651 
652 		switch (status) {
653 		case APIC_ICR_RR_VALID:
654 			status = apic_read(APIC_RRR);
655 			pr_cont("%08x\n", status);
656 			break;
657 		default:
658 			pr_cont("failed\n");
659 		}
660 	}
661 }
662 
663 /*
664  * The Multiprocessor Specification 1.4 (1997) example code suggests
665  * that there should be a 10ms delay between the BSP asserting INIT
666  * and de-asserting INIT, when starting a remote processor.
667  * But that slows boot and resume on modern processors, which include
668  * many cores and don't require that delay.
669  *
670  * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
671  * Modern processor families are quirked to remove the delay entirely.
672  */
673 #define UDELAY_10MS_DEFAULT 10000
674 
675 static unsigned int init_udelay = UINT_MAX;
676 
677 static int __init cpu_init_udelay(char *str)
678 {
679 	get_option(&str, &init_udelay);
680 
681 	return 0;
682 }
683 early_param("cpu_init_udelay", cpu_init_udelay);
684 
685 static void __init smp_quirk_init_udelay(void)
686 {
687 	/* if cmdline changed it from default, leave it alone */
688 	if (init_udelay != UINT_MAX)
689 		return;
690 
691 	/* if modern processor, use no delay */
692 	if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
693 	    ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
694 		init_udelay = 0;
695 		return;
696 	}
697 	/* else, use legacy delay */
698 	init_udelay = UDELAY_10MS_DEFAULT;
699 }
700 
701 /*
702  * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
703  * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
704  * won't ... remember to clear down the APIC, etc later.
705  */
706 int
707 wakeup_secondary_cpu_via_nmi(int apicid, unsigned long start_eip)
708 {
709 	unsigned long send_status, accept_status = 0;
710 	int maxlvt;
711 
712 	/* Target chip */
713 	/* Boot on the stack */
714 	/* Kick the second */
715 	apic_icr_write(APIC_DM_NMI | apic->dest_logical, apicid);
716 
717 	pr_debug("Waiting for send to finish...\n");
718 	send_status = safe_apic_wait_icr_idle();
719 
720 	/*
721 	 * Give the other CPU some time to accept the IPI.
722 	 */
723 	udelay(200);
724 	if (APIC_INTEGRATED(boot_cpu_apic_version)) {
725 		maxlvt = lapic_get_maxlvt();
726 		if (maxlvt > 3)			/* Due to the Pentium erratum 3AP.  */
727 			apic_write(APIC_ESR, 0);
728 		accept_status = (apic_read(APIC_ESR) & 0xEF);
729 	}
730 	pr_debug("NMI sent\n");
731 
732 	if (send_status)
733 		pr_err("APIC never delivered???\n");
734 	if (accept_status)
735 		pr_err("APIC delivery error (%lx)\n", accept_status);
736 
737 	return (send_status | accept_status);
738 }
739 
740 static int
741 wakeup_secondary_cpu_via_init(int phys_apicid, unsigned long start_eip)
742 {
743 	unsigned long send_status = 0, accept_status = 0;
744 	int maxlvt, num_starts, j;
745 
746 	maxlvt = lapic_get_maxlvt();
747 
748 	/*
749 	 * Be paranoid about clearing APIC errors.
750 	 */
751 	if (APIC_INTEGRATED(boot_cpu_apic_version)) {
752 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
753 			apic_write(APIC_ESR, 0);
754 		apic_read(APIC_ESR);
755 	}
756 
757 	pr_debug("Asserting INIT\n");
758 
759 	/*
760 	 * Turn INIT on target chip
761 	 */
762 	/*
763 	 * Send IPI
764 	 */
765 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT,
766 		       phys_apicid);
767 
768 	pr_debug("Waiting for send to finish...\n");
769 	send_status = safe_apic_wait_icr_idle();
770 
771 	udelay(init_udelay);
772 
773 	pr_debug("Deasserting INIT\n");
774 
775 	/* Target chip */
776 	/* Send IPI */
777 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
778 
779 	pr_debug("Waiting for send to finish...\n");
780 	send_status = safe_apic_wait_icr_idle();
781 
782 	mb();
783 
784 	/*
785 	 * Should we send STARTUP IPIs ?
786 	 *
787 	 * Determine this based on the APIC version.
788 	 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
789 	 */
790 	if (APIC_INTEGRATED(boot_cpu_apic_version))
791 		num_starts = 2;
792 	else
793 		num_starts = 0;
794 
795 	/*
796 	 * Run STARTUP IPI loop.
797 	 */
798 	pr_debug("#startup loops: %d\n", num_starts);
799 
800 	for (j = 1; j <= num_starts; j++) {
801 		pr_debug("Sending STARTUP #%d\n", j);
802 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
803 			apic_write(APIC_ESR, 0);
804 		apic_read(APIC_ESR);
805 		pr_debug("After apic_write\n");
806 
807 		/*
808 		 * STARTUP IPI
809 		 */
810 
811 		/* Target chip */
812 		/* Boot on the stack */
813 		/* Kick the second */
814 		apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
815 			       phys_apicid);
816 
817 		/*
818 		 * Give the other CPU some time to accept the IPI.
819 		 */
820 		if (init_udelay == 0)
821 			udelay(10);
822 		else
823 			udelay(300);
824 
825 		pr_debug("Startup point 1\n");
826 
827 		pr_debug("Waiting for send to finish...\n");
828 		send_status = safe_apic_wait_icr_idle();
829 
830 		/*
831 		 * Give the other CPU some time to accept the IPI.
832 		 */
833 		if (init_udelay == 0)
834 			udelay(10);
835 		else
836 			udelay(200);
837 
838 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
839 			apic_write(APIC_ESR, 0);
840 		accept_status = (apic_read(APIC_ESR) & 0xEF);
841 		if (send_status || accept_status)
842 			break;
843 	}
844 	pr_debug("After Startup\n");
845 
846 	if (send_status)
847 		pr_err("APIC never delivered???\n");
848 	if (accept_status)
849 		pr_err("APIC delivery error (%lx)\n", accept_status);
850 
851 	return (send_status | accept_status);
852 }
853 
854 /* reduce the number of lines printed when booting a large cpu count system */
855 static void announce_cpu(int cpu, int apicid)
856 {
857 	static int current_node = -1;
858 	int node = early_cpu_to_node(cpu);
859 	static int width, node_width;
860 
861 	if (!width)
862 		width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
863 
864 	if (!node_width)
865 		node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
866 
867 	if (cpu == 1)
868 		printk(KERN_INFO "x86: Booting SMP configuration:\n");
869 
870 	if (system_state == SYSTEM_BOOTING) {
871 		if (node != current_node) {
872 			if (current_node > (-1))
873 				pr_cont("\n");
874 			current_node = node;
875 
876 			printk(KERN_INFO ".... node %*s#%d, CPUs:  ",
877 			       node_width - num_digits(node), " ", node);
878 		}
879 
880 		/* Add padding for the BSP */
881 		if (cpu == 1)
882 			pr_cont("%*s", width + 1, " ");
883 
884 		pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
885 
886 	} else
887 		pr_info("Booting Node %d Processor %d APIC 0x%x\n",
888 			node, cpu, apicid);
889 }
890 
891 static int wakeup_cpu0_nmi(unsigned int cmd, struct pt_regs *regs)
892 {
893 	int cpu;
894 
895 	cpu = smp_processor_id();
896 	if (cpu == 0 && !cpu_online(cpu) && enable_start_cpu0)
897 		return NMI_HANDLED;
898 
899 	return NMI_DONE;
900 }
901 
902 /*
903  * Wake up AP by INIT, INIT, STARTUP sequence.
904  *
905  * Instead of waiting for STARTUP after INITs, BSP will execute the BIOS
906  * boot-strap code which is not a desired behavior for waking up BSP. To
907  * void the boot-strap code, wake up CPU0 by NMI instead.
908  *
909  * This works to wake up soft offlined CPU0 only. If CPU0 is hard offlined
910  * (i.e. physically hot removed and then hot added), NMI won't wake it up.
911  * We'll change this code in the future to wake up hard offlined CPU0 if
912  * real platform and request are available.
913  */
914 static int
915 wakeup_cpu_via_init_nmi(int cpu, unsigned long start_ip, int apicid,
916 	       int *cpu0_nmi_registered)
917 {
918 	int id;
919 	int boot_error;
920 
921 	preempt_disable();
922 
923 	/*
924 	 * Wake up AP by INIT, INIT, STARTUP sequence.
925 	 */
926 	if (cpu) {
927 		boot_error = wakeup_secondary_cpu_via_init(apicid, start_ip);
928 		goto out;
929 	}
930 
931 	/*
932 	 * Wake up BSP by nmi.
933 	 *
934 	 * Register a NMI handler to help wake up CPU0.
935 	 */
936 	boot_error = register_nmi_handler(NMI_LOCAL,
937 					  wakeup_cpu0_nmi, 0, "wake_cpu0");
938 
939 	if (!boot_error) {
940 		enable_start_cpu0 = 1;
941 		*cpu0_nmi_registered = 1;
942 		if (apic->dest_logical == APIC_DEST_LOGICAL)
943 			id = cpu0_logical_apicid;
944 		else
945 			id = apicid;
946 		boot_error = wakeup_secondary_cpu_via_nmi(id, start_ip);
947 	}
948 
949 out:
950 	preempt_enable();
951 
952 	return boot_error;
953 }
954 
955 void common_cpu_up(unsigned int cpu, struct task_struct *idle)
956 {
957 	/* Just in case we booted with a single CPU. */
958 	alternatives_enable_smp();
959 
960 	per_cpu(current_task, cpu) = idle;
961 
962 #ifdef CONFIG_X86_32
963 	/* Stack for startup_32 can be just as for start_secondary onwards */
964 	irq_ctx_init(cpu);
965 	per_cpu(cpu_current_top_of_stack, cpu) =
966 		(unsigned long)task_stack_page(idle) + THREAD_SIZE;
967 #else
968 	initial_gs = per_cpu_offset(cpu);
969 #endif
970 }
971 
972 /*
973  * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
974  * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
975  * Returns zero if CPU booted OK, else error code from
976  * ->wakeup_secondary_cpu.
977  */
978 static int do_boot_cpu(int apicid, int cpu, struct task_struct *idle)
979 {
980 	volatile u32 *trampoline_status =
981 		(volatile u32 *) __va(real_mode_header->trampoline_status);
982 	/* start_ip had better be page-aligned! */
983 	unsigned long start_ip = real_mode_header->trampoline_start;
984 
985 	unsigned long boot_error = 0;
986 	int cpu0_nmi_registered = 0;
987 	unsigned long timeout;
988 
989 	idle->thread.sp = (unsigned long)task_pt_regs(idle);
990 	early_gdt_descr.address = (unsigned long)get_cpu_gdt_table(cpu);
991 	initial_code = (unsigned long)start_secondary;
992 	initial_stack  = idle->thread.sp;
993 
994 	/*
995 	 * Enable the espfix hack for this CPU
996 	*/
997 #ifdef CONFIG_X86_ESPFIX64
998 	init_espfix_ap(cpu);
999 #endif
1000 
1001 	/* So we see what's up */
1002 	announce_cpu(cpu, apicid);
1003 
1004 	/*
1005 	 * This grunge runs the startup process for
1006 	 * the targeted processor.
1007 	 */
1008 
1009 	if (get_uv_system_type() != UV_NON_UNIQUE_APIC) {
1010 
1011 		pr_debug("Setting warm reset code and vector.\n");
1012 
1013 		smpboot_setup_warm_reset_vector(start_ip);
1014 		/*
1015 		 * Be paranoid about clearing APIC errors.
1016 		*/
1017 		if (APIC_INTEGRATED(boot_cpu_apic_version)) {
1018 			apic_write(APIC_ESR, 0);
1019 			apic_read(APIC_ESR);
1020 		}
1021 	}
1022 
1023 	/*
1024 	 * AP might wait on cpu_callout_mask in cpu_init() with
1025 	 * cpu_initialized_mask set if previous attempt to online
1026 	 * it timed-out. Clear cpu_initialized_mask so that after
1027 	 * INIT/SIPI it could start with a clean state.
1028 	 */
1029 	cpumask_clear_cpu(cpu, cpu_initialized_mask);
1030 	smp_mb();
1031 
1032 	/*
1033 	 * Wake up a CPU in difference cases:
1034 	 * - Use the method in the APIC driver if it's defined
1035 	 * Otherwise,
1036 	 * - Use an INIT boot APIC message for APs or NMI for BSP.
1037 	 */
1038 	if (apic->wakeup_secondary_cpu)
1039 		boot_error = apic->wakeup_secondary_cpu(apicid, start_ip);
1040 	else
1041 		boot_error = wakeup_cpu_via_init_nmi(cpu, start_ip, apicid,
1042 						     &cpu0_nmi_registered);
1043 
1044 	if (!boot_error) {
1045 		/*
1046 		 * Wait 10s total for first sign of life from AP
1047 		 */
1048 		boot_error = -1;
1049 		timeout = jiffies + 10*HZ;
1050 		while (time_before(jiffies, timeout)) {
1051 			if (cpumask_test_cpu(cpu, cpu_initialized_mask)) {
1052 				/*
1053 				 * Tell AP to proceed with initialization
1054 				 */
1055 				cpumask_set_cpu(cpu, cpu_callout_mask);
1056 				boot_error = 0;
1057 				break;
1058 			}
1059 			schedule();
1060 		}
1061 	}
1062 
1063 	if (!boot_error) {
1064 		/*
1065 		 * Wait till AP completes initial initialization
1066 		 */
1067 		while (!cpumask_test_cpu(cpu, cpu_callin_mask)) {
1068 			/*
1069 			 * Allow other tasks to run while we wait for the
1070 			 * AP to come online. This also gives a chance
1071 			 * for the MTRR work(triggered by the AP coming online)
1072 			 * to be completed in the stop machine context.
1073 			 */
1074 			schedule();
1075 		}
1076 	}
1077 
1078 	/* mark "stuck" area as not stuck */
1079 	*trampoline_status = 0;
1080 
1081 	if (get_uv_system_type() != UV_NON_UNIQUE_APIC) {
1082 		/*
1083 		 * Cleanup possible dangling ends...
1084 		 */
1085 		smpboot_restore_warm_reset_vector();
1086 	}
1087 	/*
1088 	 * Clean up the nmi handler. Do this after the callin and callout sync
1089 	 * to avoid impact of possible long unregister time.
1090 	 */
1091 	if (cpu0_nmi_registered)
1092 		unregister_nmi_handler(NMI_LOCAL, "wake_cpu0");
1093 
1094 	return boot_error;
1095 }
1096 
1097 int native_cpu_up(unsigned int cpu, struct task_struct *tidle)
1098 {
1099 	int apicid = apic->cpu_present_to_apicid(cpu);
1100 	unsigned long flags;
1101 	int err;
1102 
1103 	WARN_ON(irqs_disabled());
1104 
1105 	pr_debug("++++++++++++++++++++=_---CPU UP  %u\n", cpu);
1106 
1107 	if (apicid == BAD_APICID ||
1108 	    !physid_isset(apicid, phys_cpu_present_map) ||
1109 	    !apic->apic_id_valid(apicid)) {
1110 		pr_err("%s: bad cpu %d\n", __func__, cpu);
1111 		return -EINVAL;
1112 	}
1113 
1114 	/*
1115 	 * Already booted CPU?
1116 	 */
1117 	if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
1118 		pr_debug("do_boot_cpu %d Already started\n", cpu);
1119 		return -ENOSYS;
1120 	}
1121 
1122 	/*
1123 	 * Save current MTRR state in case it was changed since early boot
1124 	 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1125 	 */
1126 	mtrr_save_state();
1127 
1128 	/* x86 CPUs take themselves offline, so delayed offline is OK. */
1129 	err = cpu_check_up_prepare(cpu);
1130 	if (err && err != -EBUSY)
1131 		return err;
1132 
1133 	/* the FPU context is blank, nobody can own it */
1134 	per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
1135 
1136 	common_cpu_up(cpu, tidle);
1137 
1138 	err = do_boot_cpu(apicid, cpu, tidle);
1139 	if (err) {
1140 		pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
1141 		return -EIO;
1142 	}
1143 
1144 	/*
1145 	 * Check TSC synchronization with the AP (keep irqs disabled
1146 	 * while doing so):
1147 	 */
1148 	local_irq_save(flags);
1149 	check_tsc_sync_source(cpu);
1150 	local_irq_restore(flags);
1151 
1152 	while (!cpu_online(cpu)) {
1153 		cpu_relax();
1154 		touch_nmi_watchdog();
1155 	}
1156 
1157 	return 0;
1158 }
1159 
1160 /**
1161  * arch_disable_smp_support() - disables SMP support for x86 at runtime
1162  */
1163 void arch_disable_smp_support(void)
1164 {
1165 	disable_ioapic_support();
1166 }
1167 
1168 /*
1169  * Fall back to non SMP mode after errors.
1170  *
1171  * RED-PEN audit/test this more. I bet there is more state messed up here.
1172  */
1173 static __init void disable_smp(void)
1174 {
1175 	pr_info("SMP disabled\n");
1176 
1177 	disable_ioapic_support();
1178 
1179 	init_cpu_present(cpumask_of(0));
1180 	init_cpu_possible(cpumask_of(0));
1181 
1182 	if (smp_found_config)
1183 		physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
1184 	else
1185 		physid_set_mask_of_physid(0, &phys_cpu_present_map);
1186 	cpumask_set_cpu(0, topology_sibling_cpumask(0));
1187 	cpumask_set_cpu(0, topology_core_cpumask(0));
1188 }
1189 
1190 enum {
1191 	SMP_OK,
1192 	SMP_NO_CONFIG,
1193 	SMP_NO_APIC,
1194 	SMP_FORCE_UP,
1195 };
1196 
1197 /*
1198  * Various sanity checks.
1199  */
1200 static int __init smp_sanity_check(unsigned max_cpus)
1201 {
1202 	preempt_disable();
1203 
1204 #if !defined(CONFIG_X86_BIGSMP) && defined(CONFIG_X86_32)
1205 	if (def_to_bigsmp && nr_cpu_ids > 8) {
1206 		unsigned int cpu;
1207 		unsigned nr;
1208 
1209 		pr_warn("More than 8 CPUs detected - skipping them\n"
1210 			"Use CONFIG_X86_BIGSMP\n");
1211 
1212 		nr = 0;
1213 		for_each_present_cpu(cpu) {
1214 			if (nr >= 8)
1215 				set_cpu_present(cpu, false);
1216 			nr++;
1217 		}
1218 
1219 		nr = 0;
1220 		for_each_possible_cpu(cpu) {
1221 			if (nr >= 8)
1222 				set_cpu_possible(cpu, false);
1223 			nr++;
1224 		}
1225 
1226 		nr_cpu_ids = 8;
1227 	}
1228 #endif
1229 
1230 	if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map)) {
1231 		pr_warn("weird, boot CPU (#%d) not listed by the BIOS\n",
1232 			hard_smp_processor_id());
1233 
1234 		physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1235 	}
1236 
1237 	/*
1238 	 * If we couldn't find an SMP configuration at boot time,
1239 	 * get out of here now!
1240 	 */
1241 	if (!smp_found_config && !acpi_lapic) {
1242 		preempt_enable();
1243 		pr_notice("SMP motherboard not detected\n");
1244 		return SMP_NO_CONFIG;
1245 	}
1246 
1247 	/*
1248 	 * Should not be necessary because the MP table should list the boot
1249 	 * CPU too, but we do it for the sake of robustness anyway.
1250 	 */
1251 	if (!apic->check_phys_apicid_present(boot_cpu_physical_apicid)) {
1252 		pr_notice("weird, boot CPU (#%d) not listed by the BIOS\n",
1253 			  boot_cpu_physical_apicid);
1254 		physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1255 	}
1256 	preempt_enable();
1257 
1258 	/*
1259 	 * If we couldn't find a local APIC, then get out of here now!
1260 	 */
1261 	if (APIC_INTEGRATED(boot_cpu_apic_version) &&
1262 	    !boot_cpu_has(X86_FEATURE_APIC)) {
1263 		if (!disable_apic) {
1264 			pr_err("BIOS bug, local APIC #%d not detected!...\n",
1265 				boot_cpu_physical_apicid);
1266 			pr_err("... forcing use of dummy APIC emulation (tell your hw vendor)\n");
1267 		}
1268 		return SMP_NO_APIC;
1269 	}
1270 
1271 	/*
1272 	 * If SMP should be disabled, then really disable it!
1273 	 */
1274 	if (!max_cpus) {
1275 		pr_info("SMP mode deactivated\n");
1276 		return SMP_FORCE_UP;
1277 	}
1278 
1279 	return SMP_OK;
1280 }
1281 
1282 static void __init smp_cpu_index_default(void)
1283 {
1284 	int i;
1285 	struct cpuinfo_x86 *c;
1286 
1287 	for_each_possible_cpu(i) {
1288 		c = &cpu_data(i);
1289 		/* mark all to hotplug */
1290 		c->cpu_index = nr_cpu_ids;
1291 	}
1292 }
1293 
1294 /*
1295  * Prepare for SMP bootup.  The MP table or ACPI has been read
1296  * earlier.  Just do some sanity checking here and enable APIC mode.
1297  */
1298 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1299 {
1300 	unsigned int i;
1301 
1302 	smp_cpu_index_default();
1303 
1304 	/*
1305 	 * Setup boot CPU information
1306 	 */
1307 	smp_store_boot_cpu_info(); /* Final full version of the data */
1308 	cpumask_copy(cpu_callin_mask, cpumask_of(0));
1309 	mb();
1310 
1311 	for_each_possible_cpu(i) {
1312 		zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1313 		zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1314 		zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1315 	}
1316 
1317 	/*
1318 	 * Set 'default' x86 topology, this matches default_topology() in that
1319 	 * it has NUMA nodes as a topology level. See also
1320 	 * native_smp_cpus_done().
1321 	 *
1322 	 * Must be done before set_cpus_sibling_map() is ran.
1323 	 */
1324 	set_sched_topology(x86_topology);
1325 
1326 	set_cpu_sibling_map(0);
1327 
1328 	switch (smp_sanity_check(max_cpus)) {
1329 	case SMP_NO_CONFIG:
1330 		disable_smp();
1331 		if (APIC_init_uniprocessor())
1332 			pr_notice("Local APIC not detected. Using dummy APIC emulation.\n");
1333 		return;
1334 	case SMP_NO_APIC:
1335 		disable_smp();
1336 		return;
1337 	case SMP_FORCE_UP:
1338 		disable_smp();
1339 		apic_bsp_setup(false);
1340 		return;
1341 	case SMP_OK:
1342 		break;
1343 	}
1344 
1345 	if (read_apic_id() != boot_cpu_physical_apicid) {
1346 		panic("Boot APIC ID in local APIC unexpected (%d vs %d)",
1347 		     read_apic_id(), boot_cpu_physical_apicid);
1348 		/* Or can we switch back to PIC here? */
1349 	}
1350 
1351 	default_setup_apic_routing();
1352 	cpu0_logical_apicid = apic_bsp_setup(false);
1353 
1354 	pr_info("CPU0: ");
1355 	print_cpu_info(&cpu_data(0));
1356 
1357 	if (is_uv_system())
1358 		uv_system_init();
1359 
1360 	set_mtrr_aps_delayed_init();
1361 
1362 	smp_quirk_init_udelay();
1363 }
1364 
1365 void arch_enable_nonboot_cpus_begin(void)
1366 {
1367 	set_mtrr_aps_delayed_init();
1368 }
1369 
1370 void arch_enable_nonboot_cpus_end(void)
1371 {
1372 	mtrr_aps_init();
1373 }
1374 
1375 /*
1376  * Early setup to make printk work.
1377  */
1378 void __init native_smp_prepare_boot_cpu(void)
1379 {
1380 	int me = smp_processor_id();
1381 	switch_to_new_gdt(me);
1382 	/* already set me in cpu_online_mask in boot_cpu_init() */
1383 	cpumask_set_cpu(me, cpu_callout_mask);
1384 	cpu_set_state_online(me);
1385 }
1386 
1387 void __init native_smp_cpus_done(unsigned int max_cpus)
1388 {
1389 	pr_debug("Boot done\n");
1390 
1391 	if (x86_has_numa_in_package)
1392 		set_sched_topology(x86_numa_in_package_topology);
1393 
1394 	nmi_selftest();
1395 	impress_friends();
1396 	setup_ioapic_dest();
1397 	mtrr_aps_init();
1398 }
1399 
1400 static int __initdata setup_possible_cpus = -1;
1401 static int __init _setup_possible_cpus(char *str)
1402 {
1403 	get_option(&str, &setup_possible_cpus);
1404 	return 0;
1405 }
1406 early_param("possible_cpus", _setup_possible_cpus);
1407 
1408 
1409 /*
1410  * cpu_possible_mask should be static, it cannot change as cpu's
1411  * are onlined, or offlined. The reason is per-cpu data-structures
1412  * are allocated by some modules at init time, and dont expect to
1413  * do this dynamically on cpu arrival/departure.
1414  * cpu_present_mask on the other hand can change dynamically.
1415  * In case when cpu_hotplug is not compiled, then we resort to current
1416  * behaviour, which is cpu_possible == cpu_present.
1417  * - Ashok Raj
1418  *
1419  * Three ways to find out the number of additional hotplug CPUs:
1420  * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1421  * - The user can overwrite it with possible_cpus=NUM
1422  * - Otherwise don't reserve additional CPUs.
1423  * We do this because additional CPUs waste a lot of memory.
1424  * -AK
1425  */
1426 __init void prefill_possible_map(void)
1427 {
1428 	int i, possible;
1429 
1430 	/* No boot processor was found in mptable or ACPI MADT */
1431 	if (!num_processors) {
1432 		if (boot_cpu_has(X86_FEATURE_APIC)) {
1433 			int apicid = boot_cpu_physical_apicid;
1434 			int cpu = hard_smp_processor_id();
1435 
1436 			pr_warn("Boot CPU (id %d) not listed by BIOS\n", cpu);
1437 
1438 			/* Make sure boot cpu is enumerated */
1439 			if (apic->cpu_present_to_apicid(0) == BAD_APICID &&
1440 			    apic->apic_id_valid(apicid))
1441 				generic_processor_info(apicid, boot_cpu_apic_version);
1442 		}
1443 
1444 		if (!num_processors)
1445 			num_processors = 1;
1446 	}
1447 
1448 	i = setup_max_cpus ?: 1;
1449 	if (setup_possible_cpus == -1) {
1450 		possible = num_processors;
1451 #ifdef CONFIG_HOTPLUG_CPU
1452 		if (setup_max_cpus)
1453 			possible += disabled_cpus;
1454 #else
1455 		if (possible > i)
1456 			possible = i;
1457 #endif
1458 	} else
1459 		possible = setup_possible_cpus;
1460 
1461 	total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1462 
1463 	/* nr_cpu_ids could be reduced via nr_cpus= */
1464 	if (possible > nr_cpu_ids) {
1465 		pr_warn("%d Processors exceeds NR_CPUS limit of %d\n",
1466 			possible, nr_cpu_ids);
1467 		possible = nr_cpu_ids;
1468 	}
1469 
1470 #ifdef CONFIG_HOTPLUG_CPU
1471 	if (!setup_max_cpus)
1472 #endif
1473 	if (possible > i) {
1474 		pr_warn("%d Processors exceeds max_cpus limit of %u\n",
1475 			possible, setup_max_cpus);
1476 		possible = i;
1477 	}
1478 
1479 	pr_info("Allowing %d CPUs, %d hotplug CPUs\n",
1480 		possible, max_t(int, possible - num_processors, 0));
1481 
1482 	for (i = 0; i < possible; i++)
1483 		set_cpu_possible(i, true);
1484 	for (; i < NR_CPUS; i++)
1485 		set_cpu_possible(i, false);
1486 
1487 	nr_cpu_ids = possible;
1488 }
1489 
1490 #ifdef CONFIG_HOTPLUG_CPU
1491 
1492 /* Recompute SMT state for all CPUs on offline */
1493 static void recompute_smt_state(void)
1494 {
1495 	int max_threads, cpu;
1496 
1497 	max_threads = 0;
1498 	for_each_online_cpu (cpu) {
1499 		int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1500 
1501 		if (threads > max_threads)
1502 			max_threads = threads;
1503 	}
1504 	__max_smt_threads = max_threads;
1505 }
1506 
1507 static void remove_siblinginfo(int cpu)
1508 {
1509 	int sibling;
1510 	struct cpuinfo_x86 *c = &cpu_data(cpu);
1511 
1512 	for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1513 		cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1514 		/*/
1515 		 * last thread sibling in this cpu core going down
1516 		 */
1517 		if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1518 			cpu_data(sibling).booted_cores--;
1519 	}
1520 
1521 	for_each_cpu(sibling, topology_sibling_cpumask(cpu))
1522 		cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1523 	for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1524 		cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1525 	cpumask_clear(cpu_llc_shared_mask(cpu));
1526 	cpumask_clear(topology_sibling_cpumask(cpu));
1527 	cpumask_clear(topology_core_cpumask(cpu));
1528 	c->phys_proc_id = 0;
1529 	c->cpu_core_id = 0;
1530 	cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1531 	recompute_smt_state();
1532 }
1533 
1534 static void remove_cpu_from_maps(int cpu)
1535 {
1536 	set_cpu_online(cpu, false);
1537 	cpumask_clear_cpu(cpu, cpu_callout_mask);
1538 	cpumask_clear_cpu(cpu, cpu_callin_mask);
1539 	/* was set by cpu_init() */
1540 	cpumask_clear_cpu(cpu, cpu_initialized_mask);
1541 	numa_remove_cpu(cpu);
1542 }
1543 
1544 void cpu_disable_common(void)
1545 {
1546 	int cpu = smp_processor_id();
1547 
1548 	remove_siblinginfo(cpu);
1549 
1550 	/* It's now safe to remove this processor from the online map */
1551 	lock_vector_lock();
1552 	remove_cpu_from_maps(cpu);
1553 	unlock_vector_lock();
1554 	fixup_irqs();
1555 }
1556 
1557 int native_cpu_disable(void)
1558 {
1559 	int ret;
1560 
1561 	ret = check_irq_vectors_for_cpu_disable();
1562 	if (ret)
1563 		return ret;
1564 
1565 	clear_local_APIC();
1566 	cpu_disable_common();
1567 
1568 	return 0;
1569 }
1570 
1571 int common_cpu_die(unsigned int cpu)
1572 {
1573 	int ret = 0;
1574 
1575 	/* We don't do anything here: idle task is faking death itself. */
1576 
1577 	/* They ack this in play_dead() by setting CPU_DEAD */
1578 	if (cpu_wait_death(cpu, 5)) {
1579 		if (system_state == SYSTEM_RUNNING)
1580 			pr_info("CPU %u is now offline\n", cpu);
1581 	} else {
1582 		pr_err("CPU %u didn't die...\n", cpu);
1583 		ret = -1;
1584 	}
1585 
1586 	return ret;
1587 }
1588 
1589 void native_cpu_die(unsigned int cpu)
1590 {
1591 	common_cpu_die(cpu);
1592 }
1593 
1594 void play_dead_common(void)
1595 {
1596 	idle_task_exit();
1597 	reset_lazy_tlbstate();
1598 
1599 	/* Ack it */
1600 	(void)cpu_report_death();
1601 
1602 	/*
1603 	 * With physical CPU hotplug, we should halt the cpu
1604 	 */
1605 	local_irq_disable();
1606 }
1607 
1608 static bool wakeup_cpu0(void)
1609 {
1610 	if (smp_processor_id() == 0 && enable_start_cpu0)
1611 		return true;
1612 
1613 	return false;
1614 }
1615 
1616 /*
1617  * We need to flush the caches before going to sleep, lest we have
1618  * dirty data in our caches when we come back up.
1619  */
1620 static inline void mwait_play_dead(void)
1621 {
1622 	unsigned int eax, ebx, ecx, edx;
1623 	unsigned int highest_cstate = 0;
1624 	unsigned int highest_subcstate = 0;
1625 	void *mwait_ptr;
1626 	int i;
1627 
1628 	if (!this_cpu_has(X86_FEATURE_MWAIT))
1629 		return;
1630 	if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1631 		return;
1632 	if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1633 		return;
1634 
1635 	eax = CPUID_MWAIT_LEAF;
1636 	ecx = 0;
1637 	native_cpuid(&eax, &ebx, &ecx, &edx);
1638 
1639 	/*
1640 	 * eax will be 0 if EDX enumeration is not valid.
1641 	 * Initialized below to cstate, sub_cstate value when EDX is valid.
1642 	 */
1643 	if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1644 		eax = 0;
1645 	} else {
1646 		edx >>= MWAIT_SUBSTATE_SIZE;
1647 		for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1648 			if (edx & MWAIT_SUBSTATE_MASK) {
1649 				highest_cstate = i;
1650 				highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1651 			}
1652 		}
1653 		eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1654 			(highest_subcstate - 1);
1655 	}
1656 
1657 	/*
1658 	 * This should be a memory location in a cache line which is
1659 	 * unlikely to be touched by other processors.  The actual
1660 	 * content is immaterial as it is not actually modified in any way.
1661 	 */
1662 	mwait_ptr = &current_thread_info()->flags;
1663 
1664 	wbinvd();
1665 
1666 	while (1) {
1667 		/*
1668 		 * The CLFLUSH is a workaround for erratum AAI65 for
1669 		 * the Xeon 7400 series.  It's not clear it is actually
1670 		 * needed, but it should be harmless in either case.
1671 		 * The WBINVD is insufficient due to the spurious-wakeup
1672 		 * case where we return around the loop.
1673 		 */
1674 		mb();
1675 		clflush(mwait_ptr);
1676 		mb();
1677 		__monitor(mwait_ptr, 0, 0);
1678 		mb();
1679 		__mwait(eax, 0);
1680 		/*
1681 		 * If NMI wants to wake up CPU0, start CPU0.
1682 		 */
1683 		if (wakeup_cpu0())
1684 			start_cpu0();
1685 	}
1686 }
1687 
1688 void hlt_play_dead(void)
1689 {
1690 	if (__this_cpu_read(cpu_info.x86) >= 4)
1691 		wbinvd();
1692 
1693 	while (1) {
1694 		native_halt();
1695 		/*
1696 		 * If NMI wants to wake up CPU0, start CPU0.
1697 		 */
1698 		if (wakeup_cpu0())
1699 			start_cpu0();
1700 	}
1701 }
1702 
1703 void native_play_dead(void)
1704 {
1705 	play_dead_common();
1706 	tboot_shutdown(TB_SHUTDOWN_WFS);
1707 
1708 	mwait_play_dead();	/* Only returns on failure */
1709 	if (cpuidle_play_dead())
1710 		hlt_play_dead();
1711 }
1712 
1713 #else /* ... !CONFIG_HOTPLUG_CPU */
1714 int native_cpu_disable(void)
1715 {
1716 	return -ENOSYS;
1717 }
1718 
1719 void native_cpu_die(unsigned int cpu)
1720 {
1721 	/* We said "no" in __cpu_disable */
1722 	BUG();
1723 }
1724 
1725 void native_play_dead(void)
1726 {
1727 	BUG();
1728 }
1729 
1730 #endif
1731