xref: /linux/arch/x86/kernel/smpboot.c (revision 6faadbbb7f9da70ce484f98f72223c20125a1009)
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/realmode.h>
79 #include <asm/misc.h>
80 
81 /* Number of siblings per CPU package */
82 int smp_num_siblings = 1;
83 EXPORT_SYMBOL(smp_num_siblings);
84 
85 /* Last level cache ID of each logical CPU */
86 DEFINE_PER_CPU_READ_MOSTLY(u16, cpu_llc_id) = BAD_APICID;
87 
88 /* representing HT siblings of each logical CPU */
89 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
90 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
91 
92 /* representing HT and core siblings of each logical CPU */
93 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
94 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
95 
96 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_llc_shared_map);
97 
98 /* Per CPU bogomips and other parameters */
99 DEFINE_PER_CPU_READ_MOSTLY(struct cpuinfo_x86, cpu_info);
100 EXPORT_PER_CPU_SYMBOL(cpu_info);
101 
102 /* Logical package management. We might want to allocate that dynamically */
103 static int *physical_to_logical_pkg __read_mostly;
104 static unsigned long *physical_package_map __read_mostly;;
105 static unsigned int max_physical_pkg_id __read_mostly;
106 unsigned int __max_logical_packages __read_mostly;
107 EXPORT_SYMBOL(__max_logical_packages);
108 static unsigned int logical_packages __read_mostly;
109 
110 /* Maximum number of SMT threads on any online core */
111 int __max_smt_threads __read_mostly;
112 
113 /* Flag to indicate if a complete sched domain rebuild is required */
114 bool x86_topology_update;
115 
116 int arch_update_cpu_topology(void)
117 {
118 	int retval = x86_topology_update;
119 
120 	x86_topology_update = false;
121 	return retval;
122 }
123 
124 static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
125 {
126 	unsigned long flags;
127 
128 	spin_lock_irqsave(&rtc_lock, flags);
129 	CMOS_WRITE(0xa, 0xf);
130 	spin_unlock_irqrestore(&rtc_lock, flags);
131 	local_flush_tlb();
132 	pr_debug("1.\n");
133 	*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) =
134 							start_eip >> 4;
135 	pr_debug("2.\n");
136 	*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) =
137 							start_eip & 0xf;
138 	pr_debug("3.\n");
139 }
140 
141 static inline void smpboot_restore_warm_reset_vector(void)
142 {
143 	unsigned long flags;
144 
145 	/*
146 	 * Install writable page 0 entry to set BIOS data area.
147 	 */
148 	local_flush_tlb();
149 
150 	/*
151 	 * Paranoid:  Set warm reset code and vector here back
152 	 * to default values.
153 	 */
154 	spin_lock_irqsave(&rtc_lock, flags);
155 	CMOS_WRITE(0, 0xf);
156 	spin_unlock_irqrestore(&rtc_lock, flags);
157 
158 	*((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
159 }
160 
161 /*
162  * Report back to the Boot Processor during boot time or to the caller processor
163  * during CPU online.
164  */
165 static void smp_callin(void)
166 {
167 	int cpuid, phys_id;
168 
169 	/*
170 	 * If waken up by an INIT in an 82489DX configuration
171 	 * cpu_callout_mask guarantees we don't get here before
172 	 * an INIT_deassert IPI reaches our local APIC, so it is
173 	 * now safe to touch our local APIC.
174 	 */
175 	cpuid = smp_processor_id();
176 
177 	/*
178 	 * (This works even if the APIC is not enabled.)
179 	 */
180 	phys_id = read_apic_id();
181 
182 	/*
183 	 * the boot CPU has finished the init stage and is spinning
184 	 * on callin_map until we finish. We are free to set up this
185 	 * CPU, first the APIC. (this is probably redundant on most
186 	 * boards)
187 	 */
188 	apic_ap_setup();
189 
190 	/*
191 	 * Save our processor parameters. Note: this information
192 	 * is needed for clock calibration.
193 	 */
194 	smp_store_cpu_info(cpuid);
195 
196 	/*
197 	 * Get our bogomips.
198 	 * Update loops_per_jiffy in cpu_data. Previous call to
199 	 * smp_store_cpu_info() stored a value that is close but not as
200 	 * accurate as the value just calculated.
201 	 */
202 	calibrate_delay();
203 	cpu_data(cpuid).loops_per_jiffy = loops_per_jiffy;
204 	pr_debug("Stack at about %p\n", &cpuid);
205 
206 	/*
207 	 * This must be done before setting cpu_online_mask
208 	 * or calling notify_cpu_starting.
209 	 */
210 	set_cpu_sibling_map(raw_smp_processor_id());
211 	wmb();
212 
213 	notify_cpu_starting(cpuid);
214 
215 	/*
216 	 * Allow the master to continue.
217 	 */
218 	cpumask_set_cpu(cpuid, cpu_callin_mask);
219 }
220 
221 static int cpu0_logical_apicid;
222 static int enable_start_cpu0;
223 /*
224  * Activate a secondary processor.
225  */
226 static void notrace start_secondary(void *unused)
227 {
228 	/*
229 	 * Don't put *anything* except direct CPU state initialization
230 	 * before cpu_init(), SMP booting is too fragile that we want to
231 	 * limit the things done here to the most necessary things.
232 	 */
233 	if (boot_cpu_has(X86_FEATURE_PCID))
234 		__write_cr4(__read_cr4() | X86_CR4_PCIDE);
235 	cpu_init();
236 	x86_cpuinit.early_percpu_clock_init();
237 	preempt_disable();
238 	smp_callin();
239 
240 	enable_start_cpu0 = 0;
241 
242 #ifdef CONFIG_X86_32
243 	/* switch away from the initial page table */
244 	load_cr3(swapper_pg_dir);
245 	__flush_tlb_all();
246 #endif
247 
248 	/* otherwise gcc will move up smp_processor_id before the cpu_init */
249 	barrier();
250 	/*
251 	 * Check TSC synchronization with the BP:
252 	 */
253 	check_tsc_sync_target();
254 
255 	/*
256 	 * Lock vector_lock and initialize the vectors on this cpu
257 	 * before setting the cpu online. We must set it online with
258 	 * vector_lock held to prevent a concurrent setup/teardown
259 	 * from seeing a half valid vector space.
260 	 */
261 	lock_vector_lock();
262 	setup_vector_irq(smp_processor_id());
263 	set_cpu_online(smp_processor_id(), true);
264 	unlock_vector_lock();
265 	cpu_set_state_online(smp_processor_id());
266 	x86_platform.nmi_init();
267 
268 	/* enable local interrupts */
269 	local_irq_enable();
270 
271 	/* to prevent fake stack check failure in clock setup */
272 	boot_init_stack_canary();
273 
274 	x86_cpuinit.setup_percpu_clockev();
275 
276 	wmb();
277 	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
278 }
279 
280 /**
281  * topology_update_package_map - Update the physical to logical package map
282  * @pkg:	The physical package id as retrieved via CPUID
283  * @cpu:	The cpu for which this is updated
284  */
285 int topology_update_package_map(unsigned int pkg, unsigned int cpu)
286 {
287 	unsigned int new;
288 
289 	/* Called from early boot ? */
290 	if (!physical_package_map)
291 		return 0;
292 
293 	if (pkg >= max_physical_pkg_id)
294 		return -EINVAL;
295 
296 	/* Set the logical package id */
297 	if (test_and_set_bit(pkg, physical_package_map))
298 		goto found;
299 
300 	if (logical_packages >= __max_logical_packages) {
301 		pr_warn("Package %u of CPU %u exceeds BIOS package data %u.\n",
302 			logical_packages, cpu, __max_logical_packages);
303 		return -ENOSPC;
304 	}
305 
306 	new = logical_packages++;
307 	if (new != pkg) {
308 		pr_info("CPU %u Converting physical %u to logical package %u\n",
309 			cpu, pkg, new);
310 	}
311 	physical_to_logical_pkg[pkg] = new;
312 
313 found:
314 	cpu_data(cpu).logical_proc_id = physical_to_logical_pkg[pkg];
315 	return 0;
316 }
317 
318 /**
319  * topology_phys_to_logical_pkg - Map a physical package id to a logical
320  *
321  * Returns logical package id or -1 if not found
322  */
323 int topology_phys_to_logical_pkg(unsigned int phys_pkg)
324 {
325 	if (phys_pkg >= max_physical_pkg_id)
326 		return -1;
327 	return physical_to_logical_pkg[phys_pkg];
328 }
329 EXPORT_SYMBOL(topology_phys_to_logical_pkg);
330 
331 static void __init smp_init_package_map(struct cpuinfo_x86 *c, unsigned int cpu)
332 {
333 	unsigned int ncpus;
334 	size_t size;
335 
336 	/*
337 	 * Today neither Intel nor AMD support heterogenous systems. That
338 	 * might change in the future....
339 	 *
340 	 * While ideally we'd want '* smp_num_siblings' in the below @ncpus
341 	 * computation, this won't actually work since some Intel BIOSes
342 	 * report inconsistent HT data when they disable HT.
343 	 *
344 	 * In particular, they reduce the APIC-IDs to only include the cores,
345 	 * but leave the CPUID topology to say there are (2) siblings.
346 	 * This means we don't know how many threads there will be until
347 	 * after the APIC enumeration.
348 	 *
349 	 * By not including this we'll sometimes over-estimate the number of
350 	 * logical packages by the amount of !present siblings, but this is
351 	 * still better than MAX_LOCAL_APIC.
352 	 *
353 	 * We use total_cpus not nr_cpu_ids because nr_cpu_ids can be limited
354 	 * on the command line leading to a similar issue as the HT disable
355 	 * problem because the hyperthreads are usually enumerated after the
356 	 * primary cores.
357 	 */
358 	ncpus = boot_cpu_data.x86_max_cores;
359 	if (!ncpus) {
360 		pr_warn("x86_max_cores == zero !?!?");
361 		ncpus = 1;
362 	}
363 
364 	__max_logical_packages = DIV_ROUND_UP(total_cpus, ncpus);
365 	logical_packages = 0;
366 
367 	/*
368 	 * Possibly larger than what we need as the number of apic ids per
369 	 * package can be smaller than the actual used apic ids.
370 	 */
371 	max_physical_pkg_id = DIV_ROUND_UP(MAX_LOCAL_APIC, ncpus);
372 	size = max_physical_pkg_id * sizeof(unsigned int);
373 	physical_to_logical_pkg = kmalloc(size, GFP_KERNEL);
374 	memset(physical_to_logical_pkg, 0xff, size);
375 	size = BITS_TO_LONGS(max_physical_pkg_id) * sizeof(unsigned long);
376 	physical_package_map = kzalloc(size, GFP_KERNEL);
377 
378 	pr_info("Max logical packages: %u\n", __max_logical_packages);
379 
380 	topology_update_package_map(c->phys_proc_id, cpu);
381 }
382 
383 void __init smp_store_boot_cpu_info(void)
384 {
385 	int id = 0; /* CPU 0 */
386 	struct cpuinfo_x86 *c = &cpu_data(id);
387 
388 	*c = boot_cpu_data;
389 	c->cpu_index = id;
390 	smp_init_package_map(c, id);
391 }
392 
393 /*
394  * The bootstrap kernel entry code has set these up. Save them for
395  * a given CPU
396  */
397 void smp_store_cpu_info(int id)
398 {
399 	struct cpuinfo_x86 *c = &cpu_data(id);
400 
401 	*c = boot_cpu_data;
402 	c->cpu_index = id;
403 	/*
404 	 * During boot time, CPU0 has this setup already. Save the info when
405 	 * bringing up AP or offlined CPU0.
406 	 */
407 	identify_secondary_cpu(c);
408 }
409 
410 static bool
411 topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
412 {
413 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
414 
415 	return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
416 }
417 
418 static bool
419 topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
420 {
421 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
422 
423 	return !WARN_ONCE(!topology_same_node(c, o),
424 		"sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
425 		"[node: %d != %d]. Ignoring dependency.\n",
426 		cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
427 }
428 
429 #define link_mask(mfunc, c1, c2)					\
430 do {									\
431 	cpumask_set_cpu((c1), mfunc(c2));				\
432 	cpumask_set_cpu((c2), mfunc(c1));				\
433 } while (0)
434 
435 static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
436 {
437 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
438 		int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
439 
440 		if (c->phys_proc_id == o->phys_proc_id &&
441 		    per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2)) {
442 			if (c->cpu_core_id == o->cpu_core_id)
443 				return topology_sane(c, o, "smt");
444 
445 			if ((c->cu_id != 0xff) &&
446 			    (o->cu_id != 0xff) &&
447 			    (c->cu_id == o->cu_id))
448 				return topology_sane(c, o, "smt");
449 		}
450 
451 	} else if (c->phys_proc_id == o->phys_proc_id &&
452 		   c->cpu_core_id == o->cpu_core_id) {
453 		return topology_sane(c, o, "smt");
454 	}
455 
456 	return false;
457 }
458 
459 static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
460 {
461 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
462 
463 	if (per_cpu(cpu_llc_id, cpu1) != BAD_APICID &&
464 	    per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2))
465 		return topology_sane(c, o, "llc");
466 
467 	return false;
468 }
469 
470 /*
471  * Unlike the other levels, we do not enforce keeping a
472  * multicore group inside a NUMA node.  If this happens, we will
473  * discard the MC level of the topology later.
474  */
475 static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
476 {
477 	if (c->phys_proc_id == o->phys_proc_id)
478 		return true;
479 	return false;
480 }
481 
482 #if defined(CONFIG_SCHED_SMT) || defined(CONFIG_SCHED_MC)
483 static inline int x86_sched_itmt_flags(void)
484 {
485 	return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
486 }
487 
488 #ifdef CONFIG_SCHED_MC
489 static int x86_core_flags(void)
490 {
491 	return cpu_core_flags() | x86_sched_itmt_flags();
492 }
493 #endif
494 #ifdef CONFIG_SCHED_SMT
495 static int x86_smt_flags(void)
496 {
497 	return cpu_smt_flags() | x86_sched_itmt_flags();
498 }
499 #endif
500 #endif
501 
502 static struct sched_domain_topology_level x86_numa_in_package_topology[] = {
503 #ifdef CONFIG_SCHED_SMT
504 	{ cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
505 #endif
506 #ifdef CONFIG_SCHED_MC
507 	{ cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
508 #endif
509 	{ NULL, },
510 };
511 
512 static struct sched_domain_topology_level x86_topology[] = {
513 #ifdef CONFIG_SCHED_SMT
514 	{ cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
515 #endif
516 #ifdef CONFIG_SCHED_MC
517 	{ cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
518 #endif
519 	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
520 	{ NULL, },
521 };
522 
523 /*
524  * Set if a package/die has multiple NUMA nodes inside.
525  * AMD Magny-Cours and Intel Cluster-on-Die have this.
526  */
527 static bool x86_has_numa_in_package;
528 
529 void set_cpu_sibling_map(int cpu)
530 {
531 	bool has_smt = smp_num_siblings > 1;
532 	bool has_mp = has_smt || boot_cpu_data.x86_max_cores > 1;
533 	struct cpuinfo_x86 *c = &cpu_data(cpu);
534 	struct cpuinfo_x86 *o;
535 	int i, threads;
536 
537 	cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
538 
539 	if (!has_mp) {
540 		cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
541 		cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
542 		cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
543 		c->booted_cores = 1;
544 		return;
545 	}
546 
547 	for_each_cpu(i, cpu_sibling_setup_mask) {
548 		o = &cpu_data(i);
549 
550 		if ((i == cpu) || (has_smt && match_smt(c, o)))
551 			link_mask(topology_sibling_cpumask, cpu, i);
552 
553 		if ((i == cpu) || (has_mp && match_llc(c, o)))
554 			link_mask(cpu_llc_shared_mask, cpu, i);
555 
556 	}
557 
558 	/*
559 	 * This needs a separate iteration over the cpus because we rely on all
560 	 * topology_sibling_cpumask links to be set-up.
561 	 */
562 	for_each_cpu(i, cpu_sibling_setup_mask) {
563 		o = &cpu_data(i);
564 
565 		if ((i == cpu) || (has_mp && match_die(c, o))) {
566 			link_mask(topology_core_cpumask, cpu, i);
567 
568 			/*
569 			 *  Does this new cpu bringup a new core?
570 			 */
571 			if (cpumask_weight(
572 			    topology_sibling_cpumask(cpu)) == 1) {
573 				/*
574 				 * for each core in package, increment
575 				 * the booted_cores for this new cpu
576 				 */
577 				if (cpumask_first(
578 				    topology_sibling_cpumask(i)) == i)
579 					c->booted_cores++;
580 				/*
581 				 * increment the core count for all
582 				 * the other cpus in this package
583 				 */
584 				if (i != cpu)
585 					cpu_data(i).booted_cores++;
586 			} else if (i != cpu && !c->booted_cores)
587 				c->booted_cores = cpu_data(i).booted_cores;
588 		}
589 		if (match_die(c, o) && !topology_same_node(c, o))
590 			x86_has_numa_in_package = true;
591 	}
592 
593 	threads = cpumask_weight(topology_sibling_cpumask(cpu));
594 	if (threads > __max_smt_threads)
595 		__max_smt_threads = threads;
596 }
597 
598 /* maps the cpu to the sched domain representing multi-core */
599 const struct cpumask *cpu_coregroup_mask(int cpu)
600 {
601 	return cpu_llc_shared_mask(cpu);
602 }
603 
604 static void impress_friends(void)
605 {
606 	int cpu;
607 	unsigned long bogosum = 0;
608 	/*
609 	 * Allow the user to impress friends.
610 	 */
611 	pr_debug("Before bogomips\n");
612 	for_each_possible_cpu(cpu)
613 		if (cpumask_test_cpu(cpu, cpu_callout_mask))
614 			bogosum += cpu_data(cpu).loops_per_jiffy;
615 	pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
616 		num_online_cpus(),
617 		bogosum/(500000/HZ),
618 		(bogosum/(5000/HZ))%100);
619 
620 	pr_debug("Before bogocount - setting activated=1\n");
621 }
622 
623 void __inquire_remote_apic(int apicid)
624 {
625 	unsigned i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
626 	const char * const names[] = { "ID", "VERSION", "SPIV" };
627 	int timeout;
628 	u32 status;
629 
630 	pr_info("Inquiring remote APIC 0x%x...\n", apicid);
631 
632 	for (i = 0; i < ARRAY_SIZE(regs); i++) {
633 		pr_info("... APIC 0x%x %s: ", apicid, names[i]);
634 
635 		/*
636 		 * Wait for idle.
637 		 */
638 		status = safe_apic_wait_icr_idle();
639 		if (status)
640 			pr_cont("a previous APIC delivery may have failed\n");
641 
642 		apic_icr_write(APIC_DM_REMRD | regs[i], apicid);
643 
644 		timeout = 0;
645 		do {
646 			udelay(100);
647 			status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
648 		} while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
649 
650 		switch (status) {
651 		case APIC_ICR_RR_VALID:
652 			status = apic_read(APIC_RRR);
653 			pr_cont("%08x\n", status);
654 			break;
655 		default:
656 			pr_cont("failed\n");
657 		}
658 	}
659 }
660 
661 /*
662  * The Multiprocessor Specification 1.4 (1997) example code suggests
663  * that there should be a 10ms delay between the BSP asserting INIT
664  * and de-asserting INIT, when starting a remote processor.
665  * But that slows boot and resume on modern processors, which include
666  * many cores and don't require that delay.
667  *
668  * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
669  * Modern processor families are quirked to remove the delay entirely.
670  */
671 #define UDELAY_10MS_DEFAULT 10000
672 
673 static unsigned int init_udelay = UINT_MAX;
674 
675 static int __init cpu_init_udelay(char *str)
676 {
677 	get_option(&str, &init_udelay);
678 
679 	return 0;
680 }
681 early_param("cpu_init_udelay", cpu_init_udelay);
682 
683 static void __init smp_quirk_init_udelay(void)
684 {
685 	/* if cmdline changed it from default, leave it alone */
686 	if (init_udelay != UINT_MAX)
687 		return;
688 
689 	/* if modern processor, use no delay */
690 	if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
691 	    ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
692 		init_udelay = 0;
693 		return;
694 	}
695 	/* else, use legacy delay */
696 	init_udelay = UDELAY_10MS_DEFAULT;
697 }
698 
699 /*
700  * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
701  * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
702  * won't ... remember to clear down the APIC, etc later.
703  */
704 int
705 wakeup_secondary_cpu_via_nmi(int apicid, unsigned long start_eip)
706 {
707 	unsigned long send_status, accept_status = 0;
708 	int maxlvt;
709 
710 	/* Target chip */
711 	/* Boot on the stack */
712 	/* Kick the second */
713 	apic_icr_write(APIC_DM_NMI | apic->dest_logical, apicid);
714 
715 	pr_debug("Waiting for send to finish...\n");
716 	send_status = safe_apic_wait_icr_idle();
717 
718 	/*
719 	 * Give the other CPU some time to accept the IPI.
720 	 */
721 	udelay(200);
722 	if (APIC_INTEGRATED(boot_cpu_apic_version)) {
723 		maxlvt = lapic_get_maxlvt();
724 		if (maxlvt > 3)			/* Due to the Pentium erratum 3AP.  */
725 			apic_write(APIC_ESR, 0);
726 		accept_status = (apic_read(APIC_ESR) & 0xEF);
727 	}
728 	pr_debug("NMI sent\n");
729 
730 	if (send_status)
731 		pr_err("APIC never delivered???\n");
732 	if (accept_status)
733 		pr_err("APIC delivery error (%lx)\n", accept_status);
734 
735 	return (send_status | accept_status);
736 }
737 
738 static int
739 wakeup_secondary_cpu_via_init(int phys_apicid, unsigned long start_eip)
740 {
741 	unsigned long send_status = 0, accept_status = 0;
742 	int maxlvt, num_starts, j;
743 
744 	maxlvt = lapic_get_maxlvt();
745 
746 	/*
747 	 * Be paranoid about clearing APIC errors.
748 	 */
749 	if (APIC_INTEGRATED(boot_cpu_apic_version)) {
750 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
751 			apic_write(APIC_ESR, 0);
752 		apic_read(APIC_ESR);
753 	}
754 
755 	pr_debug("Asserting INIT\n");
756 
757 	/*
758 	 * Turn INIT on target chip
759 	 */
760 	/*
761 	 * Send IPI
762 	 */
763 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT,
764 		       phys_apicid);
765 
766 	pr_debug("Waiting for send to finish...\n");
767 	send_status = safe_apic_wait_icr_idle();
768 
769 	udelay(init_udelay);
770 
771 	pr_debug("Deasserting INIT\n");
772 
773 	/* Target chip */
774 	/* Send IPI */
775 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
776 
777 	pr_debug("Waiting for send to finish...\n");
778 	send_status = safe_apic_wait_icr_idle();
779 
780 	mb();
781 
782 	/*
783 	 * Should we send STARTUP IPIs ?
784 	 *
785 	 * Determine this based on the APIC version.
786 	 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
787 	 */
788 	if (APIC_INTEGRATED(boot_cpu_apic_version))
789 		num_starts = 2;
790 	else
791 		num_starts = 0;
792 
793 	/*
794 	 * Run STARTUP IPI loop.
795 	 */
796 	pr_debug("#startup loops: %d\n", num_starts);
797 
798 	for (j = 1; j <= num_starts; j++) {
799 		pr_debug("Sending STARTUP #%d\n", j);
800 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
801 			apic_write(APIC_ESR, 0);
802 		apic_read(APIC_ESR);
803 		pr_debug("After apic_write\n");
804 
805 		/*
806 		 * STARTUP IPI
807 		 */
808 
809 		/* Target chip */
810 		/* Boot on the stack */
811 		/* Kick the second */
812 		apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
813 			       phys_apicid);
814 
815 		/*
816 		 * Give the other CPU some time to accept the IPI.
817 		 */
818 		if (init_udelay == 0)
819 			udelay(10);
820 		else
821 			udelay(300);
822 
823 		pr_debug("Startup point 1\n");
824 
825 		pr_debug("Waiting for send to finish...\n");
826 		send_status = safe_apic_wait_icr_idle();
827 
828 		/*
829 		 * Give the other CPU some time to accept the IPI.
830 		 */
831 		if (init_udelay == 0)
832 			udelay(10);
833 		else
834 			udelay(200);
835 
836 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
837 			apic_write(APIC_ESR, 0);
838 		accept_status = (apic_read(APIC_ESR) & 0xEF);
839 		if (send_status || accept_status)
840 			break;
841 	}
842 	pr_debug("After Startup\n");
843 
844 	if (send_status)
845 		pr_err("APIC never delivered???\n");
846 	if (accept_status)
847 		pr_err("APIC delivery error (%lx)\n", accept_status);
848 
849 	return (send_status | accept_status);
850 }
851 
852 /* reduce the number of lines printed when booting a large cpu count system */
853 static void announce_cpu(int cpu, int apicid)
854 {
855 	static int current_node = -1;
856 	int node = early_cpu_to_node(cpu);
857 	static int width, node_width;
858 
859 	if (!width)
860 		width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
861 
862 	if (!node_width)
863 		node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
864 
865 	if (cpu == 1)
866 		printk(KERN_INFO "x86: Booting SMP configuration:\n");
867 
868 	if (system_state < SYSTEM_RUNNING) {
869 		if (node != current_node) {
870 			if (current_node > (-1))
871 				pr_cont("\n");
872 			current_node = node;
873 
874 			printk(KERN_INFO ".... node %*s#%d, CPUs:  ",
875 			       node_width - num_digits(node), " ", node);
876 		}
877 
878 		/* Add padding for the BSP */
879 		if (cpu == 1)
880 			pr_cont("%*s", width + 1, " ");
881 
882 		pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
883 
884 	} else
885 		pr_info("Booting Node %d Processor %d APIC 0x%x\n",
886 			node, cpu, apicid);
887 }
888 
889 static int wakeup_cpu0_nmi(unsigned int cmd, struct pt_regs *regs)
890 {
891 	int cpu;
892 
893 	cpu = smp_processor_id();
894 	if (cpu == 0 && !cpu_online(cpu) && enable_start_cpu0)
895 		return NMI_HANDLED;
896 
897 	return NMI_DONE;
898 }
899 
900 /*
901  * Wake up AP by INIT, INIT, STARTUP sequence.
902  *
903  * Instead of waiting for STARTUP after INITs, BSP will execute the BIOS
904  * boot-strap code which is not a desired behavior for waking up BSP. To
905  * void the boot-strap code, wake up CPU0 by NMI instead.
906  *
907  * This works to wake up soft offlined CPU0 only. If CPU0 is hard offlined
908  * (i.e. physically hot removed and then hot added), NMI won't wake it up.
909  * We'll change this code in the future to wake up hard offlined CPU0 if
910  * real platform and request are available.
911  */
912 static int
913 wakeup_cpu_via_init_nmi(int cpu, unsigned long start_ip, int apicid,
914 	       int *cpu0_nmi_registered)
915 {
916 	int id;
917 	int boot_error;
918 
919 	preempt_disable();
920 
921 	/*
922 	 * Wake up AP by INIT, INIT, STARTUP sequence.
923 	 */
924 	if (cpu) {
925 		boot_error = wakeup_secondary_cpu_via_init(apicid, start_ip);
926 		goto out;
927 	}
928 
929 	/*
930 	 * Wake up BSP by nmi.
931 	 *
932 	 * Register a NMI handler to help wake up CPU0.
933 	 */
934 	boot_error = register_nmi_handler(NMI_LOCAL,
935 					  wakeup_cpu0_nmi, 0, "wake_cpu0");
936 
937 	if (!boot_error) {
938 		enable_start_cpu0 = 1;
939 		*cpu0_nmi_registered = 1;
940 		if (apic->dest_logical == APIC_DEST_LOGICAL)
941 			id = cpu0_logical_apicid;
942 		else
943 			id = apicid;
944 		boot_error = wakeup_secondary_cpu_via_nmi(id, start_ip);
945 	}
946 
947 out:
948 	preempt_enable();
949 
950 	return boot_error;
951 }
952 
953 void common_cpu_up(unsigned int cpu, struct task_struct *idle)
954 {
955 	/* Just in case we booted with a single CPU. */
956 	alternatives_enable_smp();
957 
958 	per_cpu(current_task, cpu) = idle;
959 
960 #ifdef CONFIG_X86_32
961 	/* Stack for startup_32 can be just as for start_secondary onwards */
962 	irq_ctx_init(cpu);
963 	per_cpu(cpu_current_top_of_stack, cpu) =
964 		(unsigned long)task_stack_page(idle) + THREAD_SIZE;
965 #else
966 	initial_gs = per_cpu_offset(cpu);
967 #endif
968 }
969 
970 /*
971  * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
972  * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
973  * Returns zero if CPU booted OK, else error code from
974  * ->wakeup_secondary_cpu.
975  */
976 static int do_boot_cpu(int apicid, int cpu, struct task_struct *idle,
977 		       int *cpu0_nmi_registered)
978 {
979 	volatile u32 *trampoline_status =
980 		(volatile u32 *) __va(real_mode_header->trampoline_status);
981 	/* start_ip had better be page-aligned! */
982 	unsigned long start_ip = real_mode_header->trampoline_start;
983 
984 	unsigned long boot_error = 0;
985 	unsigned long timeout;
986 
987 	idle->thread.sp = (unsigned long)task_pt_regs(idle);
988 	early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
989 	initial_code = (unsigned long)start_secondary;
990 	initial_stack  = idle->thread.sp;
991 
992 	/*
993 	 * Enable the espfix hack for this CPU
994 	*/
995 #ifdef CONFIG_X86_ESPFIX64
996 	init_espfix_ap(cpu);
997 #endif
998 
999 	/* So we see what's up */
1000 	announce_cpu(cpu, apicid);
1001 
1002 	/*
1003 	 * This grunge runs the startup process for
1004 	 * the targeted processor.
1005 	 */
1006 
1007 	if (get_uv_system_type() != UV_NON_UNIQUE_APIC) {
1008 
1009 		pr_debug("Setting warm reset code and vector.\n");
1010 
1011 		smpboot_setup_warm_reset_vector(start_ip);
1012 		/*
1013 		 * Be paranoid about clearing APIC errors.
1014 		*/
1015 		if (APIC_INTEGRATED(boot_cpu_apic_version)) {
1016 			apic_write(APIC_ESR, 0);
1017 			apic_read(APIC_ESR);
1018 		}
1019 	}
1020 
1021 	/*
1022 	 * AP might wait on cpu_callout_mask in cpu_init() with
1023 	 * cpu_initialized_mask set if previous attempt to online
1024 	 * it timed-out. Clear cpu_initialized_mask so that after
1025 	 * INIT/SIPI it could start with a clean state.
1026 	 */
1027 	cpumask_clear_cpu(cpu, cpu_initialized_mask);
1028 	smp_mb();
1029 
1030 	/*
1031 	 * Wake up a CPU in difference cases:
1032 	 * - Use the method in the APIC driver if it's defined
1033 	 * Otherwise,
1034 	 * - Use an INIT boot APIC message for APs or NMI for BSP.
1035 	 */
1036 	if (apic->wakeup_secondary_cpu)
1037 		boot_error = apic->wakeup_secondary_cpu(apicid, start_ip);
1038 	else
1039 		boot_error = wakeup_cpu_via_init_nmi(cpu, start_ip, apicid,
1040 						     cpu0_nmi_registered);
1041 
1042 	if (!boot_error) {
1043 		/*
1044 		 * Wait 10s total for first sign of life from AP
1045 		 */
1046 		boot_error = -1;
1047 		timeout = jiffies + 10*HZ;
1048 		while (time_before(jiffies, timeout)) {
1049 			if (cpumask_test_cpu(cpu, cpu_initialized_mask)) {
1050 				/*
1051 				 * Tell AP to proceed with initialization
1052 				 */
1053 				cpumask_set_cpu(cpu, cpu_callout_mask);
1054 				boot_error = 0;
1055 				break;
1056 			}
1057 			schedule();
1058 		}
1059 	}
1060 
1061 	if (!boot_error) {
1062 		/*
1063 		 * Wait till AP completes initial initialization
1064 		 */
1065 		while (!cpumask_test_cpu(cpu, cpu_callin_mask)) {
1066 			/*
1067 			 * Allow other tasks to run while we wait for the
1068 			 * AP to come online. This also gives a chance
1069 			 * for the MTRR work(triggered by the AP coming online)
1070 			 * to be completed in the stop machine context.
1071 			 */
1072 			schedule();
1073 		}
1074 	}
1075 
1076 	/* mark "stuck" area as not stuck */
1077 	*trampoline_status = 0;
1078 
1079 	if (get_uv_system_type() != UV_NON_UNIQUE_APIC) {
1080 		/*
1081 		 * Cleanup possible dangling ends...
1082 		 */
1083 		smpboot_restore_warm_reset_vector();
1084 	}
1085 
1086 	return boot_error;
1087 }
1088 
1089 int native_cpu_up(unsigned int cpu, struct task_struct *tidle)
1090 {
1091 	int apicid = apic->cpu_present_to_apicid(cpu);
1092 	int cpu0_nmi_registered = 0;
1093 	unsigned long flags;
1094 	int err, ret = 0;
1095 
1096 	WARN_ON(irqs_disabled());
1097 
1098 	pr_debug("++++++++++++++++++++=_---CPU UP  %u\n", cpu);
1099 
1100 	if (apicid == BAD_APICID ||
1101 	    !physid_isset(apicid, phys_cpu_present_map) ||
1102 	    !apic->apic_id_valid(apicid)) {
1103 		pr_err("%s: bad cpu %d\n", __func__, cpu);
1104 		return -EINVAL;
1105 	}
1106 
1107 	/*
1108 	 * Already booted CPU?
1109 	 */
1110 	if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
1111 		pr_debug("do_boot_cpu %d Already started\n", cpu);
1112 		return -ENOSYS;
1113 	}
1114 
1115 	/*
1116 	 * Save current MTRR state in case it was changed since early boot
1117 	 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1118 	 */
1119 	mtrr_save_state();
1120 
1121 	/* x86 CPUs take themselves offline, so delayed offline is OK. */
1122 	err = cpu_check_up_prepare(cpu);
1123 	if (err && err != -EBUSY)
1124 		return err;
1125 
1126 	/* the FPU context is blank, nobody can own it */
1127 	per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
1128 
1129 	common_cpu_up(cpu, tidle);
1130 
1131 	err = do_boot_cpu(apicid, cpu, tidle, &cpu0_nmi_registered);
1132 	if (err) {
1133 		pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
1134 		ret = -EIO;
1135 		goto unreg_nmi;
1136 	}
1137 
1138 	/*
1139 	 * Check TSC synchronization with the AP (keep irqs disabled
1140 	 * while doing so):
1141 	 */
1142 	local_irq_save(flags);
1143 	check_tsc_sync_source(cpu);
1144 	local_irq_restore(flags);
1145 
1146 	while (!cpu_online(cpu)) {
1147 		cpu_relax();
1148 		touch_nmi_watchdog();
1149 	}
1150 
1151 unreg_nmi:
1152 	/*
1153 	 * Clean up the nmi handler. Do this after the callin and callout sync
1154 	 * to avoid impact of possible long unregister time.
1155 	 */
1156 	if (cpu0_nmi_registered)
1157 		unregister_nmi_handler(NMI_LOCAL, "wake_cpu0");
1158 
1159 	return ret;
1160 }
1161 
1162 /**
1163  * arch_disable_smp_support() - disables SMP support for x86 at runtime
1164  */
1165 void arch_disable_smp_support(void)
1166 {
1167 	disable_ioapic_support();
1168 }
1169 
1170 /*
1171  * Fall back to non SMP mode after errors.
1172  *
1173  * RED-PEN audit/test this more. I bet there is more state messed up here.
1174  */
1175 static __init void disable_smp(void)
1176 {
1177 	pr_info("SMP disabled\n");
1178 
1179 	disable_ioapic_support();
1180 
1181 	init_cpu_present(cpumask_of(0));
1182 	init_cpu_possible(cpumask_of(0));
1183 
1184 	if (smp_found_config)
1185 		physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
1186 	else
1187 		physid_set_mask_of_physid(0, &phys_cpu_present_map);
1188 	cpumask_set_cpu(0, topology_sibling_cpumask(0));
1189 	cpumask_set_cpu(0, topology_core_cpumask(0));
1190 }
1191 
1192 enum {
1193 	SMP_OK,
1194 	SMP_NO_CONFIG,
1195 	SMP_NO_APIC,
1196 	SMP_FORCE_UP,
1197 };
1198 
1199 /*
1200  * Various sanity checks.
1201  */
1202 static int __init smp_sanity_check(unsigned max_cpus)
1203 {
1204 	preempt_disable();
1205 
1206 #if !defined(CONFIG_X86_BIGSMP) && defined(CONFIG_X86_32)
1207 	if (def_to_bigsmp && nr_cpu_ids > 8) {
1208 		unsigned int cpu;
1209 		unsigned nr;
1210 
1211 		pr_warn("More than 8 CPUs detected - skipping them\n"
1212 			"Use CONFIG_X86_BIGSMP\n");
1213 
1214 		nr = 0;
1215 		for_each_present_cpu(cpu) {
1216 			if (nr >= 8)
1217 				set_cpu_present(cpu, false);
1218 			nr++;
1219 		}
1220 
1221 		nr = 0;
1222 		for_each_possible_cpu(cpu) {
1223 			if (nr >= 8)
1224 				set_cpu_possible(cpu, false);
1225 			nr++;
1226 		}
1227 
1228 		nr_cpu_ids = 8;
1229 	}
1230 #endif
1231 
1232 	if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map)) {
1233 		pr_warn("weird, boot CPU (#%d) not listed by the BIOS\n",
1234 			hard_smp_processor_id());
1235 
1236 		physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1237 	}
1238 
1239 	/*
1240 	 * If we couldn't find an SMP configuration at boot time,
1241 	 * get out of here now!
1242 	 */
1243 	if (!smp_found_config && !acpi_lapic) {
1244 		preempt_enable();
1245 		pr_notice("SMP motherboard not detected\n");
1246 		return SMP_NO_CONFIG;
1247 	}
1248 
1249 	/*
1250 	 * Should not be necessary because the MP table should list the boot
1251 	 * CPU too, but we do it for the sake of robustness anyway.
1252 	 */
1253 	if (!apic->check_phys_apicid_present(boot_cpu_physical_apicid)) {
1254 		pr_notice("weird, boot CPU (#%d) not listed by the BIOS\n",
1255 			  boot_cpu_physical_apicid);
1256 		physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1257 	}
1258 	preempt_enable();
1259 
1260 	/*
1261 	 * If we couldn't find a local APIC, then get out of here now!
1262 	 */
1263 	if (APIC_INTEGRATED(boot_cpu_apic_version) &&
1264 	    !boot_cpu_has(X86_FEATURE_APIC)) {
1265 		if (!disable_apic) {
1266 			pr_err("BIOS bug, local APIC #%d not detected!...\n",
1267 				boot_cpu_physical_apicid);
1268 			pr_err("... forcing use of dummy APIC emulation (tell your hw vendor)\n");
1269 		}
1270 		return SMP_NO_APIC;
1271 	}
1272 
1273 	/*
1274 	 * If SMP should be disabled, then really disable it!
1275 	 */
1276 	if (!max_cpus) {
1277 		pr_info("SMP mode deactivated\n");
1278 		return SMP_FORCE_UP;
1279 	}
1280 
1281 	return SMP_OK;
1282 }
1283 
1284 static void __init smp_cpu_index_default(void)
1285 {
1286 	int i;
1287 	struct cpuinfo_x86 *c;
1288 
1289 	for_each_possible_cpu(i) {
1290 		c = &cpu_data(i);
1291 		/* mark all to hotplug */
1292 		c->cpu_index = nr_cpu_ids;
1293 	}
1294 }
1295 
1296 /*
1297  * Prepare for SMP bootup.  The MP table or ACPI has been read
1298  * earlier.  Just do some sanity checking here and enable APIC mode.
1299  */
1300 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1301 {
1302 	unsigned int i;
1303 
1304 	smp_cpu_index_default();
1305 
1306 	/*
1307 	 * Setup boot CPU information
1308 	 */
1309 	smp_store_boot_cpu_info(); /* Final full version of the data */
1310 	cpumask_copy(cpu_callin_mask, cpumask_of(0));
1311 	mb();
1312 
1313 	for_each_possible_cpu(i) {
1314 		zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1315 		zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1316 		zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1317 	}
1318 
1319 	/*
1320 	 * Set 'default' x86 topology, this matches default_topology() in that
1321 	 * it has NUMA nodes as a topology level. See also
1322 	 * native_smp_cpus_done().
1323 	 *
1324 	 * Must be done before set_cpus_sibling_map() is ran.
1325 	 */
1326 	set_sched_topology(x86_topology);
1327 
1328 	set_cpu_sibling_map(0);
1329 
1330 	switch (smp_sanity_check(max_cpus)) {
1331 	case SMP_NO_CONFIG:
1332 		disable_smp();
1333 		if (APIC_init_uniprocessor())
1334 			pr_notice("Local APIC not detected. Using dummy APIC emulation.\n");
1335 		return;
1336 	case SMP_NO_APIC:
1337 		disable_smp();
1338 		return;
1339 	case SMP_FORCE_UP:
1340 		disable_smp();
1341 		apic_bsp_setup(false);
1342 		return;
1343 	case SMP_OK:
1344 		break;
1345 	}
1346 
1347 	if (read_apic_id() != boot_cpu_physical_apicid) {
1348 		panic("Boot APIC ID in local APIC unexpected (%d vs %d)",
1349 		     read_apic_id(), boot_cpu_physical_apicid);
1350 		/* Or can we switch back to PIC here? */
1351 	}
1352 
1353 	default_setup_apic_routing();
1354 	cpu0_logical_apicid = apic_bsp_setup(false);
1355 
1356 	pr_info("CPU0: ");
1357 	print_cpu_info(&cpu_data(0));
1358 
1359 	uv_system_init();
1360 
1361 	set_mtrr_aps_delayed_init();
1362 
1363 	smp_quirk_init_udelay();
1364 }
1365 
1366 void arch_enable_nonboot_cpus_begin(void)
1367 {
1368 	set_mtrr_aps_delayed_init();
1369 }
1370 
1371 void arch_enable_nonboot_cpus_end(void)
1372 {
1373 	mtrr_aps_init();
1374 }
1375 
1376 /*
1377  * Early setup to make printk work.
1378  */
1379 void __init native_smp_prepare_boot_cpu(void)
1380 {
1381 	int me = smp_processor_id();
1382 	switch_to_new_gdt(me);
1383 	/* already set me in cpu_online_mask in boot_cpu_init() */
1384 	cpumask_set_cpu(me, cpu_callout_mask);
1385 	cpu_set_state_online(me);
1386 }
1387 
1388 void __init native_smp_cpus_done(unsigned int max_cpus)
1389 {
1390 	pr_debug("Boot done\n");
1391 
1392 	if (x86_has_numa_in_package)
1393 		set_sched_topology(x86_numa_in_package_topology);
1394 
1395 	nmi_selftest();
1396 	impress_friends();
1397 	setup_ioapic_dest();
1398 	mtrr_aps_init();
1399 }
1400 
1401 static int __initdata setup_possible_cpus = -1;
1402 static int __init _setup_possible_cpus(char *str)
1403 {
1404 	get_option(&str, &setup_possible_cpus);
1405 	return 0;
1406 }
1407 early_param("possible_cpus", _setup_possible_cpus);
1408 
1409 
1410 /*
1411  * cpu_possible_mask should be static, it cannot change as cpu's
1412  * are onlined, or offlined. The reason is per-cpu data-structures
1413  * are allocated by some modules at init time, and dont expect to
1414  * do this dynamically on cpu arrival/departure.
1415  * cpu_present_mask on the other hand can change dynamically.
1416  * In case when cpu_hotplug is not compiled, then we resort to current
1417  * behaviour, which is cpu_possible == cpu_present.
1418  * - Ashok Raj
1419  *
1420  * Three ways to find out the number of additional hotplug CPUs:
1421  * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1422  * - The user can overwrite it with possible_cpus=NUM
1423  * - Otherwise don't reserve additional CPUs.
1424  * We do this because additional CPUs waste a lot of memory.
1425  * -AK
1426  */
1427 __init void prefill_possible_map(void)
1428 {
1429 	int i, possible;
1430 
1431 	/* No boot processor was found in mptable or ACPI MADT */
1432 	if (!num_processors) {
1433 		if (boot_cpu_has(X86_FEATURE_APIC)) {
1434 			int apicid = boot_cpu_physical_apicid;
1435 			int cpu = hard_smp_processor_id();
1436 
1437 			pr_warn("Boot CPU (id %d) not listed by BIOS\n", cpu);
1438 
1439 			/* Make sure boot cpu is enumerated */
1440 			if (apic->cpu_present_to_apicid(0) == BAD_APICID &&
1441 			    apic->apic_id_valid(apicid))
1442 				generic_processor_info(apicid, boot_cpu_apic_version);
1443 		}
1444 
1445 		if (!num_processors)
1446 			num_processors = 1;
1447 	}
1448 
1449 	i = setup_max_cpus ?: 1;
1450 	if (setup_possible_cpus == -1) {
1451 		possible = num_processors;
1452 #ifdef CONFIG_HOTPLUG_CPU
1453 		if (setup_max_cpus)
1454 			possible += disabled_cpus;
1455 #else
1456 		if (possible > i)
1457 			possible = i;
1458 #endif
1459 	} else
1460 		possible = setup_possible_cpus;
1461 
1462 	total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1463 
1464 	/* nr_cpu_ids could be reduced via nr_cpus= */
1465 	if (possible > nr_cpu_ids) {
1466 		pr_warn("%d Processors exceeds NR_CPUS limit of %u\n",
1467 			possible, nr_cpu_ids);
1468 		possible = nr_cpu_ids;
1469 	}
1470 
1471 #ifdef CONFIG_HOTPLUG_CPU
1472 	if (!setup_max_cpus)
1473 #endif
1474 	if (possible > i) {
1475 		pr_warn("%d Processors exceeds max_cpus limit of %u\n",
1476 			possible, setup_max_cpus);
1477 		possible = i;
1478 	}
1479 
1480 	nr_cpu_ids = possible;
1481 
1482 	pr_info("Allowing %d CPUs, %d hotplug CPUs\n",
1483 		possible, max_t(int, possible - num_processors, 0));
1484 
1485 	reset_cpu_possible_mask();
1486 
1487 	for (i = 0; i < possible; i++)
1488 		set_cpu_possible(i, true);
1489 }
1490 
1491 #ifdef CONFIG_HOTPLUG_CPU
1492 
1493 /* Recompute SMT state for all CPUs on offline */
1494 static void recompute_smt_state(void)
1495 {
1496 	int max_threads, cpu;
1497 
1498 	max_threads = 0;
1499 	for_each_online_cpu (cpu) {
1500 		int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1501 
1502 		if (threads > max_threads)
1503 			max_threads = threads;
1504 	}
1505 	__max_smt_threads = max_threads;
1506 }
1507 
1508 static void remove_siblinginfo(int cpu)
1509 {
1510 	int sibling;
1511 	struct cpuinfo_x86 *c = &cpu_data(cpu);
1512 
1513 	for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1514 		cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1515 		/*/
1516 		 * last thread sibling in this cpu core going down
1517 		 */
1518 		if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1519 			cpu_data(sibling).booted_cores--;
1520 	}
1521 
1522 	for_each_cpu(sibling, topology_sibling_cpumask(cpu))
1523 		cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1524 	for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1525 		cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1526 	cpumask_clear(cpu_llc_shared_mask(cpu));
1527 	cpumask_clear(topology_sibling_cpumask(cpu));
1528 	cpumask_clear(topology_core_cpumask(cpu));
1529 	c->phys_proc_id = 0;
1530 	c->cpu_core_id = 0;
1531 	cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1532 	recompute_smt_state();
1533 }
1534 
1535 static void remove_cpu_from_maps(int cpu)
1536 {
1537 	set_cpu_online(cpu, false);
1538 	cpumask_clear_cpu(cpu, cpu_callout_mask);
1539 	cpumask_clear_cpu(cpu, cpu_callin_mask);
1540 	/* was set by cpu_init() */
1541 	cpumask_clear_cpu(cpu, cpu_initialized_mask);
1542 	numa_remove_cpu(cpu);
1543 }
1544 
1545 void cpu_disable_common(void)
1546 {
1547 	int cpu = smp_processor_id();
1548 
1549 	remove_siblinginfo(cpu);
1550 
1551 	/* It's now safe to remove this processor from the online map */
1552 	lock_vector_lock();
1553 	remove_cpu_from_maps(cpu);
1554 	unlock_vector_lock();
1555 	fixup_irqs();
1556 }
1557 
1558 int native_cpu_disable(void)
1559 {
1560 	int ret;
1561 
1562 	ret = check_irq_vectors_for_cpu_disable();
1563 	if (ret)
1564 		return ret;
1565 
1566 	clear_local_APIC();
1567 	cpu_disable_common();
1568 
1569 	return 0;
1570 }
1571 
1572 int common_cpu_die(unsigned int cpu)
1573 {
1574 	int ret = 0;
1575 
1576 	/* We don't do anything here: idle task is faking death itself. */
1577 
1578 	/* They ack this in play_dead() by setting CPU_DEAD */
1579 	if (cpu_wait_death(cpu, 5)) {
1580 		if (system_state == SYSTEM_RUNNING)
1581 			pr_info("CPU %u is now offline\n", cpu);
1582 	} else {
1583 		pr_err("CPU %u didn't die...\n", cpu);
1584 		ret = -1;
1585 	}
1586 
1587 	return ret;
1588 }
1589 
1590 void native_cpu_die(unsigned int cpu)
1591 {
1592 	common_cpu_die(cpu);
1593 }
1594 
1595 void play_dead_common(void)
1596 {
1597 	idle_task_exit();
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