xref: /openbmc/linux/arch/x86/kernel/smpboot.c (revision 77ab8d5d)
1  /*
2  *	x86 SMP booting functions
3  *
4  *	(c) 1995 Alan Cox, Building #3 <alan@lxorguk.ukuu.org.uk>
5  *	(c) 1998, 1999, 2000, 2009 Ingo Molnar <mingo@redhat.com>
6  *	Copyright 2001 Andi Kleen, SuSE Labs.
7  *
8  *	Much of the core SMP work is based on previous work by Thomas Radke, to
9  *	whom a great many thanks are extended.
10  *
11  *	Thanks to Intel for making available several different Pentium,
12  *	Pentium Pro and Pentium-II/Xeon MP machines.
13  *	Original development of Linux SMP code supported by Caldera.
14  *
15  *	This code is released under the GNU General Public License version 2 or
16  *	later.
17  *
18  *	Fixes
19  *		Felix Koop	:	NR_CPUS used properly
20  *		Jose Renau	:	Handle single CPU case.
21  *		Alan Cox	:	By repeated request 8) - Total BogoMIPS report.
22  *		Greg Wright	:	Fix for kernel stacks panic.
23  *		Erich Boleyn	:	MP v1.4 and additional changes.
24  *	Matthias Sattler	:	Changes for 2.1 kernel map.
25  *	Michel Lespinasse	:	Changes for 2.1 kernel map.
26  *	Michael Chastain	:	Change trampoline.S to gnu as.
27  *		Alan Cox	:	Dumb bug: 'B' step PPro's are fine
28  *		Ingo Molnar	:	Added APIC timers, based on code
29  *					from Jose Renau
30  *		Ingo Molnar	:	various cleanups and rewrites
31  *		Tigran Aivazian	:	fixed "0.00 in /proc/uptime on SMP" bug.
32  *	Maciej W. Rozycki	:	Bits for genuine 82489DX APICs
33  *	Andi Kleen		:	Changed for SMP boot into long mode.
34  *		Martin J. Bligh	: 	Added support for multi-quad systems
35  *		Dave Jones	:	Report invalid combinations of Athlon CPUs.
36  *		Rusty Russell	:	Hacked into shape for new "hotplug" boot process.
37  *      Andi Kleen              :       Converted to new state machine.
38  *	Ashok Raj		: 	CPU hotplug support
39  *	Glauber Costa		:	i386 and x86_64 integration
40  */
41 
42 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
43 
44 #include <linux/init.h>
45 #include <linux/smp.h>
46 #include <linux/export.h>
47 #include <linux/sched.h>
48 #include <linux/sched/topology.h>
49 #include <linux/sched/hotplug.h>
50 #include <linux/sched/task_stack.h>
51 #include <linux/percpu.h>
52 #include <linux/bootmem.h>
53 #include <linux/err.h>
54 #include <linux/nmi.h>
55 #include <linux/tboot.h>
56 #include <linux/stackprotector.h>
57 #include <linux/gfp.h>
58 #include <linux/cpuidle.h>
59 
60 #include <asm/acpi.h>
61 #include <asm/desc.h>
62 #include <asm/nmi.h>
63 #include <asm/irq.h>
64 #include <asm/realmode.h>
65 #include <asm/cpu.h>
66 #include <asm/numa.h>
67 #include <asm/pgtable.h>
68 #include <asm/tlbflush.h>
69 #include <asm/mtrr.h>
70 #include <asm/mwait.h>
71 #include <asm/apic.h>
72 #include <asm/io_apic.h>
73 #include <asm/fpu/internal.h>
74 #include <asm/setup.h>
75 #include <asm/uv/uv.h>
76 #include <linux/mc146818rtc.h>
77 #include <asm/i8259.h>
78 #include <asm/misc.h>
79 #include <asm/qspinlock.h>
80 #include <asm/intel-family.h>
81 #include <asm/cpu_device_id.h>
82 #include <asm/spec-ctrl.h>
83 
84 /* Number of siblings per CPU package */
85 int smp_num_siblings = 1;
86 EXPORT_SYMBOL(smp_num_siblings);
87 
88 /* Last level cache ID of each logical CPU */
89 DEFINE_PER_CPU_READ_MOSTLY(u16, cpu_llc_id) = BAD_APICID;
90 
91 /* representing HT siblings of each logical CPU */
92 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
93 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
94 
95 /* representing HT and core siblings of each logical CPU */
96 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
97 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
98 
99 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_llc_shared_map);
100 
101 /* Per CPU bogomips and other parameters */
102 DEFINE_PER_CPU_READ_MOSTLY(struct cpuinfo_x86, cpu_info);
103 EXPORT_PER_CPU_SYMBOL(cpu_info);
104 
105 /* Logical package management. We might want to allocate that dynamically */
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 __read_mostly __max_smt_threads = 1;
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 	*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) =
132 							start_eip >> 4;
133 	*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) =
134 							start_eip & 0xf;
135 }
136 
137 static inline void smpboot_restore_warm_reset_vector(void)
138 {
139 	unsigned long flags;
140 
141 	/*
142 	 * Paranoid:  Set warm reset code and vector here back
143 	 * to default values.
144 	 */
145 	spin_lock_irqsave(&rtc_lock, flags);
146 	CMOS_WRITE(0, 0xf);
147 	spin_unlock_irqrestore(&rtc_lock, flags);
148 
149 	*((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
150 }
151 
152 /*
153  * Report back to the Boot Processor during boot time or to the caller processor
154  * during CPU online.
155  */
156 static void smp_callin(void)
157 {
158 	int cpuid, phys_id;
159 
160 	/*
161 	 * If waken up by an INIT in an 82489DX configuration
162 	 * cpu_callout_mask guarantees we don't get here before
163 	 * an INIT_deassert IPI reaches our local APIC, so it is
164 	 * now safe to touch our local APIC.
165 	 */
166 	cpuid = smp_processor_id();
167 
168 	/*
169 	 * (This works even if the APIC is not enabled.)
170 	 */
171 	phys_id = read_apic_id();
172 
173 	/*
174 	 * the boot CPU has finished the init stage and is spinning
175 	 * on callin_map until we finish. We are free to set up this
176 	 * CPU, first the APIC. (this is probably redundant on most
177 	 * boards)
178 	 */
179 	apic_ap_setup();
180 
181 	/*
182 	 * Save our processor parameters. Note: this information
183 	 * is needed for clock calibration.
184 	 */
185 	smp_store_cpu_info(cpuid);
186 
187 	/*
188 	 * The topology information must be up to date before
189 	 * calibrate_delay() and notify_cpu_starting().
190 	 */
191 	set_cpu_sibling_map(raw_smp_processor_id());
192 
193 	/*
194 	 * Get our bogomips.
195 	 * Update loops_per_jiffy in cpu_data. Previous call to
196 	 * smp_store_cpu_info() stored a value that is close but not as
197 	 * accurate as the value just calculated.
198 	 */
199 	calibrate_delay();
200 	cpu_data(cpuid).loops_per_jiffy = loops_per_jiffy;
201 	pr_debug("Stack at about %p\n", &cpuid);
202 
203 	wmb();
204 
205 	notify_cpu_starting(cpuid);
206 
207 	/*
208 	 * Allow the master to continue.
209 	 */
210 	cpumask_set_cpu(cpuid, cpu_callin_mask);
211 }
212 
213 static int cpu0_logical_apicid;
214 static int enable_start_cpu0;
215 /*
216  * Activate a secondary processor.
217  */
218 static void notrace start_secondary(void *unused)
219 {
220 	/*
221 	 * Don't put *anything* except direct CPU state initialization
222 	 * before cpu_init(), SMP booting is too fragile that we want to
223 	 * limit the things done here to the most necessary things.
224 	 */
225 	if (boot_cpu_has(X86_FEATURE_PCID))
226 		__write_cr4(__read_cr4() | X86_CR4_PCIDE);
227 
228 #ifdef CONFIG_X86_32
229 	/* switch away from the initial page table */
230 	load_cr3(swapper_pg_dir);
231 	__flush_tlb_all();
232 #endif
233 	load_current_idt();
234 	cpu_init();
235 	x86_cpuinit.early_percpu_clock_init();
236 	preempt_disable();
237 	smp_callin();
238 
239 	enable_start_cpu0 = 0;
240 
241 	/* otherwise gcc will move up smp_processor_id before the cpu_init */
242 	barrier();
243 	/*
244 	 * Check TSC synchronization with the boot CPU:
245 	 */
246 	check_tsc_sync_target();
247 
248 	speculative_store_bypass_ht_init();
249 
250 	/*
251 	 * Lock vector_lock, set CPU online and bring the vector
252 	 * allocator online. Online must be set with vector_lock held
253 	 * to prevent a concurrent irq setup/teardown from seeing a
254 	 * half valid vector space.
255 	 */
256 	lock_vector_lock();
257 	set_cpu_online(smp_processor_id(), true);
258 	lapic_online();
259 	unlock_vector_lock();
260 	cpu_set_state_online(smp_processor_id());
261 	x86_platform.nmi_init();
262 
263 	/* enable local interrupts */
264 	local_irq_enable();
265 
266 	/* to prevent fake stack check failure in clock setup */
267 	boot_init_stack_canary();
268 
269 	x86_cpuinit.setup_percpu_clockev();
270 
271 	wmb();
272 	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
273 }
274 
275 /**
276  * topology_phys_to_logical_pkg - Map a physical package id to a logical
277  *
278  * Returns logical package id or -1 if not found
279  */
280 int topology_phys_to_logical_pkg(unsigned int phys_pkg)
281 {
282 	int cpu;
283 
284 	for_each_possible_cpu(cpu) {
285 		struct cpuinfo_x86 *c = &cpu_data(cpu);
286 
287 		if (c->initialized && c->phys_proc_id == phys_pkg)
288 			return c->logical_proc_id;
289 	}
290 	return -1;
291 }
292 EXPORT_SYMBOL(topology_phys_to_logical_pkg);
293 
294 /**
295  * topology_update_package_map - Update the physical to logical package map
296  * @pkg:	The physical package id as retrieved via CPUID
297  * @cpu:	The cpu for which this is updated
298  */
299 int topology_update_package_map(unsigned int pkg, unsigned int cpu)
300 {
301 	int new;
302 
303 	/* Already available somewhere? */
304 	new = topology_phys_to_logical_pkg(pkg);
305 	if (new >= 0)
306 		goto found;
307 
308 	new = logical_packages++;
309 	if (new != pkg) {
310 		pr_info("CPU %u Converting physical %u to logical package %u\n",
311 			cpu, pkg, new);
312 	}
313 found:
314 	cpu_data(cpu).logical_proc_id = new;
315 	return 0;
316 }
317 
318 void __init smp_store_boot_cpu_info(void)
319 {
320 	int id = 0; /* CPU 0 */
321 	struct cpuinfo_x86 *c = &cpu_data(id);
322 
323 	*c = boot_cpu_data;
324 	c->cpu_index = id;
325 	topology_update_package_map(c->phys_proc_id, id);
326 	c->initialized = true;
327 }
328 
329 /*
330  * The bootstrap kernel entry code has set these up. Save them for
331  * a given CPU
332  */
333 void smp_store_cpu_info(int id)
334 {
335 	struct cpuinfo_x86 *c = &cpu_data(id);
336 
337 	/* Copy boot_cpu_data only on the first bringup */
338 	if (!c->initialized)
339 		*c = boot_cpu_data;
340 	c->cpu_index = id;
341 	/*
342 	 * During boot time, CPU0 has this setup already. Save the info when
343 	 * bringing up AP or offlined CPU0.
344 	 */
345 	identify_secondary_cpu(c);
346 	c->initialized = true;
347 }
348 
349 static bool
350 topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
351 {
352 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
353 
354 	return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
355 }
356 
357 static bool
358 topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
359 {
360 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
361 
362 	return !WARN_ONCE(!topology_same_node(c, o),
363 		"sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
364 		"[node: %d != %d]. Ignoring dependency.\n",
365 		cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
366 }
367 
368 #define link_mask(mfunc, c1, c2)					\
369 do {									\
370 	cpumask_set_cpu((c1), mfunc(c2));				\
371 	cpumask_set_cpu((c2), mfunc(c1));				\
372 } while (0)
373 
374 static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
375 {
376 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
377 		int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
378 
379 		if (c->phys_proc_id == o->phys_proc_id &&
380 		    per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2)) {
381 			if (c->cpu_core_id == o->cpu_core_id)
382 				return topology_sane(c, o, "smt");
383 
384 			if ((c->cu_id != 0xff) &&
385 			    (o->cu_id != 0xff) &&
386 			    (c->cu_id == o->cu_id))
387 				return topology_sane(c, o, "smt");
388 		}
389 
390 	} else if (c->phys_proc_id == o->phys_proc_id &&
391 		   c->cpu_core_id == o->cpu_core_id) {
392 		return topology_sane(c, o, "smt");
393 	}
394 
395 	return false;
396 }
397 
398 /*
399  * Define snc_cpu[] for SNC (Sub-NUMA Cluster) CPUs.
400  *
401  * These are Intel CPUs that enumerate an LLC that is shared by
402  * multiple NUMA nodes. The LLC on these systems is shared for
403  * off-package data access but private to the NUMA node (half
404  * of the package) for on-package access.
405  *
406  * CPUID (the source of the information about the LLC) can only
407  * enumerate the cache as being shared *or* unshared, but not
408  * this particular configuration. The CPU in this case enumerates
409  * the cache to be shared across the entire package (spanning both
410  * NUMA nodes).
411  */
412 
413 static const struct x86_cpu_id snc_cpu[] = {
414 	{ X86_VENDOR_INTEL, 6, INTEL_FAM6_SKYLAKE_X },
415 	{}
416 };
417 
418 static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
419 {
420 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
421 
422 	/* Do not match if we do not have a valid APICID for cpu: */
423 	if (per_cpu(cpu_llc_id, cpu1) == BAD_APICID)
424 		return false;
425 
426 	/* Do not match if LLC id does not match: */
427 	if (per_cpu(cpu_llc_id, cpu1) != per_cpu(cpu_llc_id, cpu2))
428 		return false;
429 
430 	/*
431 	 * Allow the SNC topology without warning. Return of false
432 	 * means 'c' does not share the LLC of 'o'. This will be
433 	 * reflected to userspace.
434 	 */
435 	if (!topology_same_node(c, o) && x86_match_cpu(snc_cpu))
436 		return false;
437 
438 	return topology_sane(c, o, "llc");
439 }
440 
441 /*
442  * Unlike the other levels, we do not enforce keeping a
443  * multicore group inside a NUMA node.  If this happens, we will
444  * discard the MC level of the topology later.
445  */
446 static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
447 {
448 	if (c->phys_proc_id == o->phys_proc_id)
449 		return true;
450 	return false;
451 }
452 
453 #if defined(CONFIG_SCHED_SMT) || defined(CONFIG_SCHED_MC)
454 static inline int x86_sched_itmt_flags(void)
455 {
456 	return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
457 }
458 
459 #ifdef CONFIG_SCHED_MC
460 static int x86_core_flags(void)
461 {
462 	return cpu_core_flags() | x86_sched_itmt_flags();
463 }
464 #endif
465 #ifdef CONFIG_SCHED_SMT
466 static int x86_smt_flags(void)
467 {
468 	return cpu_smt_flags() | x86_sched_itmt_flags();
469 }
470 #endif
471 #endif
472 
473 static struct sched_domain_topology_level x86_numa_in_package_topology[] = {
474 #ifdef CONFIG_SCHED_SMT
475 	{ cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
476 #endif
477 #ifdef CONFIG_SCHED_MC
478 	{ cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
479 #endif
480 	{ NULL, },
481 };
482 
483 static struct sched_domain_topology_level x86_topology[] = {
484 #ifdef CONFIG_SCHED_SMT
485 	{ cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
486 #endif
487 #ifdef CONFIG_SCHED_MC
488 	{ cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
489 #endif
490 	{ cpu_cpu_mask, SD_INIT_NAME(DIE) },
491 	{ NULL, },
492 };
493 
494 /*
495  * Set if a package/die has multiple NUMA nodes inside.
496  * AMD Magny-Cours, Intel Cluster-on-Die, and Intel
497  * Sub-NUMA Clustering have this.
498  */
499 static bool x86_has_numa_in_package;
500 
501 void set_cpu_sibling_map(int cpu)
502 {
503 	bool has_smt = smp_num_siblings > 1;
504 	bool has_mp = has_smt || boot_cpu_data.x86_max_cores > 1;
505 	struct cpuinfo_x86 *c = &cpu_data(cpu);
506 	struct cpuinfo_x86 *o;
507 	int i, threads;
508 
509 	cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
510 
511 	if (!has_mp) {
512 		cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
513 		cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
514 		cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
515 		c->booted_cores = 1;
516 		return;
517 	}
518 
519 	for_each_cpu(i, cpu_sibling_setup_mask) {
520 		o = &cpu_data(i);
521 
522 		if ((i == cpu) || (has_smt && match_smt(c, o)))
523 			link_mask(topology_sibling_cpumask, cpu, i);
524 
525 		if ((i == cpu) || (has_mp && match_llc(c, o)))
526 			link_mask(cpu_llc_shared_mask, cpu, i);
527 
528 	}
529 
530 	/*
531 	 * This needs a separate iteration over the cpus because we rely on all
532 	 * topology_sibling_cpumask links to be set-up.
533 	 */
534 	for_each_cpu(i, cpu_sibling_setup_mask) {
535 		o = &cpu_data(i);
536 
537 		if ((i == cpu) || (has_mp && match_die(c, o))) {
538 			link_mask(topology_core_cpumask, cpu, i);
539 
540 			/*
541 			 *  Does this new cpu bringup a new core?
542 			 */
543 			if (cpumask_weight(
544 			    topology_sibling_cpumask(cpu)) == 1) {
545 				/*
546 				 * for each core in package, increment
547 				 * the booted_cores for this new cpu
548 				 */
549 				if (cpumask_first(
550 				    topology_sibling_cpumask(i)) == i)
551 					c->booted_cores++;
552 				/*
553 				 * increment the core count for all
554 				 * the other cpus in this package
555 				 */
556 				if (i != cpu)
557 					cpu_data(i).booted_cores++;
558 			} else if (i != cpu && !c->booted_cores)
559 				c->booted_cores = cpu_data(i).booted_cores;
560 		}
561 		if (match_die(c, o) && !topology_same_node(c, o))
562 			x86_has_numa_in_package = true;
563 	}
564 
565 	threads = cpumask_weight(topology_sibling_cpumask(cpu));
566 	if (threads > __max_smt_threads)
567 		__max_smt_threads = threads;
568 }
569 
570 /* maps the cpu to the sched domain representing multi-core */
571 const struct cpumask *cpu_coregroup_mask(int cpu)
572 {
573 	return cpu_llc_shared_mask(cpu);
574 }
575 
576 static void impress_friends(void)
577 {
578 	int cpu;
579 	unsigned long bogosum = 0;
580 	/*
581 	 * Allow the user to impress friends.
582 	 */
583 	pr_debug("Before bogomips\n");
584 	for_each_possible_cpu(cpu)
585 		if (cpumask_test_cpu(cpu, cpu_callout_mask))
586 			bogosum += cpu_data(cpu).loops_per_jiffy;
587 	pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
588 		num_online_cpus(),
589 		bogosum/(500000/HZ),
590 		(bogosum/(5000/HZ))%100);
591 
592 	pr_debug("Before bogocount - setting activated=1\n");
593 }
594 
595 void __inquire_remote_apic(int apicid)
596 {
597 	unsigned i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
598 	const char * const names[] = { "ID", "VERSION", "SPIV" };
599 	int timeout;
600 	u32 status;
601 
602 	pr_info("Inquiring remote APIC 0x%x...\n", apicid);
603 
604 	for (i = 0; i < ARRAY_SIZE(regs); i++) {
605 		pr_info("... APIC 0x%x %s: ", apicid, names[i]);
606 
607 		/*
608 		 * Wait for idle.
609 		 */
610 		status = safe_apic_wait_icr_idle();
611 		if (status)
612 			pr_cont("a previous APIC delivery may have failed\n");
613 
614 		apic_icr_write(APIC_DM_REMRD | regs[i], apicid);
615 
616 		timeout = 0;
617 		do {
618 			udelay(100);
619 			status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
620 		} while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
621 
622 		switch (status) {
623 		case APIC_ICR_RR_VALID:
624 			status = apic_read(APIC_RRR);
625 			pr_cont("%08x\n", status);
626 			break;
627 		default:
628 			pr_cont("failed\n");
629 		}
630 	}
631 }
632 
633 /*
634  * The Multiprocessor Specification 1.4 (1997) example code suggests
635  * that there should be a 10ms delay between the BSP asserting INIT
636  * and de-asserting INIT, when starting a remote processor.
637  * But that slows boot and resume on modern processors, which include
638  * many cores and don't require that delay.
639  *
640  * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
641  * Modern processor families are quirked to remove the delay entirely.
642  */
643 #define UDELAY_10MS_DEFAULT 10000
644 
645 static unsigned int init_udelay = UINT_MAX;
646 
647 static int __init cpu_init_udelay(char *str)
648 {
649 	get_option(&str, &init_udelay);
650 
651 	return 0;
652 }
653 early_param("cpu_init_udelay", cpu_init_udelay);
654 
655 static void __init smp_quirk_init_udelay(void)
656 {
657 	/* if cmdline changed it from default, leave it alone */
658 	if (init_udelay != UINT_MAX)
659 		return;
660 
661 	/* if modern processor, use no delay */
662 	if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
663 	    ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
664 		init_udelay = 0;
665 		return;
666 	}
667 	/* else, use legacy delay */
668 	init_udelay = UDELAY_10MS_DEFAULT;
669 }
670 
671 /*
672  * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
673  * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
674  * won't ... remember to clear down the APIC, etc later.
675  */
676 int
677 wakeup_secondary_cpu_via_nmi(int apicid, unsigned long start_eip)
678 {
679 	unsigned long send_status, accept_status = 0;
680 	int maxlvt;
681 
682 	/* Target chip */
683 	/* Boot on the stack */
684 	/* Kick the second */
685 	apic_icr_write(APIC_DM_NMI | apic->dest_logical, apicid);
686 
687 	pr_debug("Waiting for send to finish...\n");
688 	send_status = safe_apic_wait_icr_idle();
689 
690 	/*
691 	 * Give the other CPU some time to accept the IPI.
692 	 */
693 	udelay(200);
694 	if (APIC_INTEGRATED(boot_cpu_apic_version)) {
695 		maxlvt = lapic_get_maxlvt();
696 		if (maxlvt > 3)			/* Due to the Pentium erratum 3AP.  */
697 			apic_write(APIC_ESR, 0);
698 		accept_status = (apic_read(APIC_ESR) & 0xEF);
699 	}
700 	pr_debug("NMI sent\n");
701 
702 	if (send_status)
703 		pr_err("APIC never delivered???\n");
704 	if (accept_status)
705 		pr_err("APIC delivery error (%lx)\n", accept_status);
706 
707 	return (send_status | accept_status);
708 }
709 
710 static int
711 wakeup_secondary_cpu_via_init(int phys_apicid, unsigned long start_eip)
712 {
713 	unsigned long send_status = 0, accept_status = 0;
714 	int maxlvt, num_starts, j;
715 
716 	maxlvt = lapic_get_maxlvt();
717 
718 	/*
719 	 * Be paranoid about clearing APIC errors.
720 	 */
721 	if (APIC_INTEGRATED(boot_cpu_apic_version)) {
722 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
723 			apic_write(APIC_ESR, 0);
724 		apic_read(APIC_ESR);
725 	}
726 
727 	pr_debug("Asserting INIT\n");
728 
729 	/*
730 	 * Turn INIT on target chip
731 	 */
732 	/*
733 	 * Send IPI
734 	 */
735 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT,
736 		       phys_apicid);
737 
738 	pr_debug("Waiting for send to finish...\n");
739 	send_status = safe_apic_wait_icr_idle();
740 
741 	udelay(init_udelay);
742 
743 	pr_debug("Deasserting INIT\n");
744 
745 	/* Target chip */
746 	/* Send IPI */
747 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
748 
749 	pr_debug("Waiting for send to finish...\n");
750 	send_status = safe_apic_wait_icr_idle();
751 
752 	mb();
753 
754 	/*
755 	 * Should we send STARTUP IPIs ?
756 	 *
757 	 * Determine this based on the APIC version.
758 	 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
759 	 */
760 	if (APIC_INTEGRATED(boot_cpu_apic_version))
761 		num_starts = 2;
762 	else
763 		num_starts = 0;
764 
765 	/*
766 	 * Run STARTUP IPI loop.
767 	 */
768 	pr_debug("#startup loops: %d\n", num_starts);
769 
770 	for (j = 1; j <= num_starts; j++) {
771 		pr_debug("Sending STARTUP #%d\n", j);
772 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
773 			apic_write(APIC_ESR, 0);
774 		apic_read(APIC_ESR);
775 		pr_debug("After apic_write\n");
776 
777 		/*
778 		 * STARTUP IPI
779 		 */
780 
781 		/* Target chip */
782 		/* Boot on the stack */
783 		/* Kick the second */
784 		apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
785 			       phys_apicid);
786 
787 		/*
788 		 * Give the other CPU some time to accept the IPI.
789 		 */
790 		if (init_udelay == 0)
791 			udelay(10);
792 		else
793 			udelay(300);
794 
795 		pr_debug("Startup point 1\n");
796 
797 		pr_debug("Waiting for send to finish...\n");
798 		send_status = safe_apic_wait_icr_idle();
799 
800 		/*
801 		 * Give the other CPU some time to accept the IPI.
802 		 */
803 		if (init_udelay == 0)
804 			udelay(10);
805 		else
806 			udelay(200);
807 
808 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
809 			apic_write(APIC_ESR, 0);
810 		accept_status = (apic_read(APIC_ESR) & 0xEF);
811 		if (send_status || accept_status)
812 			break;
813 	}
814 	pr_debug("After Startup\n");
815 
816 	if (send_status)
817 		pr_err("APIC never delivered???\n");
818 	if (accept_status)
819 		pr_err("APIC delivery error (%lx)\n", accept_status);
820 
821 	return (send_status | accept_status);
822 }
823 
824 /* reduce the number of lines printed when booting a large cpu count system */
825 static void announce_cpu(int cpu, int apicid)
826 {
827 	static int current_node = -1;
828 	int node = early_cpu_to_node(cpu);
829 	static int width, node_width;
830 
831 	if (!width)
832 		width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
833 
834 	if (!node_width)
835 		node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
836 
837 	if (cpu == 1)
838 		printk(KERN_INFO "x86: Booting SMP configuration:\n");
839 
840 	if (system_state < SYSTEM_RUNNING) {
841 		if (node != current_node) {
842 			if (current_node > (-1))
843 				pr_cont("\n");
844 			current_node = node;
845 
846 			printk(KERN_INFO ".... node %*s#%d, CPUs:  ",
847 			       node_width - num_digits(node), " ", node);
848 		}
849 
850 		/* Add padding for the BSP */
851 		if (cpu == 1)
852 			pr_cont("%*s", width + 1, " ");
853 
854 		pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
855 
856 	} else
857 		pr_info("Booting Node %d Processor %d APIC 0x%x\n",
858 			node, cpu, apicid);
859 }
860 
861 static int wakeup_cpu0_nmi(unsigned int cmd, struct pt_regs *regs)
862 {
863 	int cpu;
864 
865 	cpu = smp_processor_id();
866 	if (cpu == 0 && !cpu_online(cpu) && enable_start_cpu0)
867 		return NMI_HANDLED;
868 
869 	return NMI_DONE;
870 }
871 
872 /*
873  * Wake up AP by INIT, INIT, STARTUP sequence.
874  *
875  * Instead of waiting for STARTUP after INITs, BSP will execute the BIOS
876  * boot-strap code which is not a desired behavior for waking up BSP. To
877  * void the boot-strap code, wake up CPU0 by NMI instead.
878  *
879  * This works to wake up soft offlined CPU0 only. If CPU0 is hard offlined
880  * (i.e. physically hot removed and then hot added), NMI won't wake it up.
881  * We'll change this code in the future to wake up hard offlined CPU0 if
882  * real platform and request are available.
883  */
884 static int
885 wakeup_cpu_via_init_nmi(int cpu, unsigned long start_ip, int apicid,
886 	       int *cpu0_nmi_registered)
887 {
888 	int id;
889 	int boot_error;
890 
891 	preempt_disable();
892 
893 	/*
894 	 * Wake up AP by INIT, INIT, STARTUP sequence.
895 	 */
896 	if (cpu) {
897 		boot_error = wakeup_secondary_cpu_via_init(apicid, start_ip);
898 		goto out;
899 	}
900 
901 	/*
902 	 * Wake up BSP by nmi.
903 	 *
904 	 * Register a NMI handler to help wake up CPU0.
905 	 */
906 	boot_error = register_nmi_handler(NMI_LOCAL,
907 					  wakeup_cpu0_nmi, 0, "wake_cpu0");
908 
909 	if (!boot_error) {
910 		enable_start_cpu0 = 1;
911 		*cpu0_nmi_registered = 1;
912 		if (apic->dest_logical == APIC_DEST_LOGICAL)
913 			id = cpu0_logical_apicid;
914 		else
915 			id = apicid;
916 		boot_error = wakeup_secondary_cpu_via_nmi(id, start_ip);
917 	}
918 
919 out:
920 	preempt_enable();
921 
922 	return boot_error;
923 }
924 
925 void common_cpu_up(unsigned int cpu, struct task_struct *idle)
926 {
927 	/* Just in case we booted with a single CPU. */
928 	alternatives_enable_smp();
929 
930 	per_cpu(current_task, cpu) = idle;
931 
932 #ifdef CONFIG_X86_32
933 	/* Stack for startup_32 can be just as for start_secondary onwards */
934 	irq_ctx_init(cpu);
935 	per_cpu(cpu_current_top_of_stack, cpu) = task_top_of_stack(idle);
936 #else
937 	initial_gs = per_cpu_offset(cpu);
938 #endif
939 }
940 
941 /*
942  * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
943  * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
944  * Returns zero if CPU booted OK, else error code from
945  * ->wakeup_secondary_cpu.
946  */
947 static int do_boot_cpu(int apicid, int cpu, struct task_struct *idle,
948 		       int *cpu0_nmi_registered)
949 {
950 	volatile u32 *trampoline_status =
951 		(volatile u32 *) __va(real_mode_header->trampoline_status);
952 	/* start_ip had better be page-aligned! */
953 	unsigned long start_ip = real_mode_header->trampoline_start;
954 
955 	unsigned long boot_error = 0;
956 	unsigned long timeout;
957 
958 	idle->thread.sp = (unsigned long)task_pt_regs(idle);
959 	early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
960 	initial_code = (unsigned long)start_secondary;
961 	initial_stack  = idle->thread.sp;
962 
963 	/* Enable the espfix hack for this CPU */
964 	init_espfix_ap(cpu);
965 
966 	/* So we see what's up */
967 	announce_cpu(cpu, apicid);
968 
969 	/*
970 	 * This grunge runs the startup process for
971 	 * the targeted processor.
972 	 */
973 
974 	if (x86_platform.legacy.warm_reset) {
975 
976 		pr_debug("Setting warm reset code and vector.\n");
977 
978 		smpboot_setup_warm_reset_vector(start_ip);
979 		/*
980 		 * Be paranoid about clearing APIC errors.
981 		*/
982 		if (APIC_INTEGRATED(boot_cpu_apic_version)) {
983 			apic_write(APIC_ESR, 0);
984 			apic_read(APIC_ESR);
985 		}
986 	}
987 
988 	/*
989 	 * AP might wait on cpu_callout_mask in cpu_init() with
990 	 * cpu_initialized_mask set if previous attempt to online
991 	 * it timed-out. Clear cpu_initialized_mask so that after
992 	 * INIT/SIPI it could start with a clean state.
993 	 */
994 	cpumask_clear_cpu(cpu, cpu_initialized_mask);
995 	smp_mb();
996 
997 	/*
998 	 * Wake up a CPU in difference cases:
999 	 * - Use the method in the APIC driver if it's defined
1000 	 * Otherwise,
1001 	 * - Use an INIT boot APIC message for APs or NMI for BSP.
1002 	 */
1003 	if (apic->wakeup_secondary_cpu)
1004 		boot_error = apic->wakeup_secondary_cpu(apicid, start_ip);
1005 	else
1006 		boot_error = wakeup_cpu_via_init_nmi(cpu, start_ip, apicid,
1007 						     cpu0_nmi_registered);
1008 
1009 	if (!boot_error) {
1010 		/*
1011 		 * Wait 10s total for first sign of life from AP
1012 		 */
1013 		boot_error = -1;
1014 		timeout = jiffies + 10*HZ;
1015 		while (time_before(jiffies, timeout)) {
1016 			if (cpumask_test_cpu(cpu, cpu_initialized_mask)) {
1017 				/*
1018 				 * Tell AP to proceed with initialization
1019 				 */
1020 				cpumask_set_cpu(cpu, cpu_callout_mask);
1021 				boot_error = 0;
1022 				break;
1023 			}
1024 			schedule();
1025 		}
1026 	}
1027 
1028 	if (!boot_error) {
1029 		/*
1030 		 * Wait till AP completes initial initialization
1031 		 */
1032 		while (!cpumask_test_cpu(cpu, cpu_callin_mask)) {
1033 			/*
1034 			 * Allow other tasks to run while we wait for the
1035 			 * AP to come online. This also gives a chance
1036 			 * for the MTRR work(triggered by the AP coming online)
1037 			 * to be completed in the stop machine context.
1038 			 */
1039 			schedule();
1040 		}
1041 	}
1042 
1043 	/* mark "stuck" area as not stuck */
1044 	*trampoline_status = 0;
1045 
1046 	if (x86_platform.legacy.warm_reset) {
1047 		/*
1048 		 * Cleanup possible dangling ends...
1049 		 */
1050 		smpboot_restore_warm_reset_vector();
1051 	}
1052 
1053 	return boot_error;
1054 }
1055 
1056 int native_cpu_up(unsigned int cpu, struct task_struct *tidle)
1057 {
1058 	int apicid = apic->cpu_present_to_apicid(cpu);
1059 	int cpu0_nmi_registered = 0;
1060 	unsigned long flags;
1061 	int err, ret = 0;
1062 
1063 	lockdep_assert_irqs_enabled();
1064 
1065 	pr_debug("++++++++++++++++++++=_---CPU UP  %u\n", cpu);
1066 
1067 	if (apicid == BAD_APICID ||
1068 	    !physid_isset(apicid, phys_cpu_present_map) ||
1069 	    !apic->apic_id_valid(apicid)) {
1070 		pr_err("%s: bad cpu %d\n", __func__, cpu);
1071 		return -EINVAL;
1072 	}
1073 
1074 	/*
1075 	 * Already booted CPU?
1076 	 */
1077 	if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
1078 		pr_debug("do_boot_cpu %d Already started\n", cpu);
1079 		return -ENOSYS;
1080 	}
1081 
1082 	/*
1083 	 * Save current MTRR state in case it was changed since early boot
1084 	 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1085 	 */
1086 	mtrr_save_state();
1087 
1088 	/* x86 CPUs take themselves offline, so delayed offline is OK. */
1089 	err = cpu_check_up_prepare(cpu);
1090 	if (err && err != -EBUSY)
1091 		return err;
1092 
1093 	/* the FPU context is blank, nobody can own it */
1094 	per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
1095 
1096 	common_cpu_up(cpu, tidle);
1097 
1098 	err = do_boot_cpu(apicid, cpu, tidle, &cpu0_nmi_registered);
1099 	if (err) {
1100 		pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
1101 		ret = -EIO;
1102 		goto unreg_nmi;
1103 	}
1104 
1105 	/*
1106 	 * Check TSC synchronization with the AP (keep irqs disabled
1107 	 * while doing so):
1108 	 */
1109 	local_irq_save(flags);
1110 	check_tsc_sync_source(cpu);
1111 	local_irq_restore(flags);
1112 
1113 	while (!cpu_online(cpu)) {
1114 		cpu_relax();
1115 		touch_nmi_watchdog();
1116 	}
1117 
1118 unreg_nmi:
1119 	/*
1120 	 * Clean up the nmi handler. Do this after the callin and callout sync
1121 	 * to avoid impact of possible long unregister time.
1122 	 */
1123 	if (cpu0_nmi_registered)
1124 		unregister_nmi_handler(NMI_LOCAL, "wake_cpu0");
1125 
1126 	return ret;
1127 }
1128 
1129 /**
1130  * arch_disable_smp_support() - disables SMP support for x86 at runtime
1131  */
1132 void arch_disable_smp_support(void)
1133 {
1134 	disable_ioapic_support();
1135 }
1136 
1137 /*
1138  * Fall back to non SMP mode after errors.
1139  *
1140  * RED-PEN audit/test this more. I bet there is more state messed up here.
1141  */
1142 static __init void disable_smp(void)
1143 {
1144 	pr_info("SMP disabled\n");
1145 
1146 	disable_ioapic_support();
1147 
1148 	init_cpu_present(cpumask_of(0));
1149 	init_cpu_possible(cpumask_of(0));
1150 
1151 	if (smp_found_config)
1152 		physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
1153 	else
1154 		physid_set_mask_of_physid(0, &phys_cpu_present_map);
1155 	cpumask_set_cpu(0, topology_sibling_cpumask(0));
1156 	cpumask_set_cpu(0, topology_core_cpumask(0));
1157 }
1158 
1159 /*
1160  * Various sanity checks.
1161  */
1162 static void __init smp_sanity_check(void)
1163 {
1164 	preempt_disable();
1165 
1166 #if !defined(CONFIG_X86_BIGSMP) && defined(CONFIG_X86_32)
1167 	if (def_to_bigsmp && nr_cpu_ids > 8) {
1168 		unsigned int cpu;
1169 		unsigned nr;
1170 
1171 		pr_warn("More than 8 CPUs detected - skipping them\n"
1172 			"Use CONFIG_X86_BIGSMP\n");
1173 
1174 		nr = 0;
1175 		for_each_present_cpu(cpu) {
1176 			if (nr >= 8)
1177 				set_cpu_present(cpu, false);
1178 			nr++;
1179 		}
1180 
1181 		nr = 0;
1182 		for_each_possible_cpu(cpu) {
1183 			if (nr >= 8)
1184 				set_cpu_possible(cpu, false);
1185 			nr++;
1186 		}
1187 
1188 		nr_cpu_ids = 8;
1189 	}
1190 #endif
1191 
1192 	if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map)) {
1193 		pr_warn("weird, boot CPU (#%d) not listed by the BIOS\n",
1194 			hard_smp_processor_id());
1195 
1196 		physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1197 	}
1198 
1199 	/*
1200 	 * Should not be necessary because the MP table should list the boot
1201 	 * CPU too, but we do it for the sake of robustness anyway.
1202 	 */
1203 	if (!apic->check_phys_apicid_present(boot_cpu_physical_apicid)) {
1204 		pr_notice("weird, boot CPU (#%d) not listed by the BIOS\n",
1205 			  boot_cpu_physical_apicid);
1206 		physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1207 	}
1208 	preempt_enable();
1209 }
1210 
1211 static void __init smp_cpu_index_default(void)
1212 {
1213 	int i;
1214 	struct cpuinfo_x86 *c;
1215 
1216 	for_each_possible_cpu(i) {
1217 		c = &cpu_data(i);
1218 		/* mark all to hotplug */
1219 		c->cpu_index = nr_cpu_ids;
1220 	}
1221 }
1222 
1223 static void __init smp_get_logical_apicid(void)
1224 {
1225 	if (x2apic_mode)
1226 		cpu0_logical_apicid = apic_read(APIC_LDR);
1227 	else
1228 		cpu0_logical_apicid = GET_APIC_LOGICAL_ID(apic_read(APIC_LDR));
1229 }
1230 
1231 /*
1232  * Prepare for SMP bootup.
1233  * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1234  *            for common interface support.
1235  */
1236 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1237 {
1238 	unsigned int i;
1239 
1240 	smp_cpu_index_default();
1241 
1242 	/*
1243 	 * Setup boot CPU information
1244 	 */
1245 	smp_store_boot_cpu_info(); /* Final full version of the data */
1246 	cpumask_copy(cpu_callin_mask, cpumask_of(0));
1247 	mb();
1248 
1249 	for_each_possible_cpu(i) {
1250 		zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1251 		zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1252 		zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1253 	}
1254 
1255 	/*
1256 	 * Set 'default' x86 topology, this matches default_topology() in that
1257 	 * it has NUMA nodes as a topology level. See also
1258 	 * native_smp_cpus_done().
1259 	 *
1260 	 * Must be done before set_cpus_sibling_map() is ran.
1261 	 */
1262 	set_sched_topology(x86_topology);
1263 
1264 	set_cpu_sibling_map(0);
1265 
1266 	smp_sanity_check();
1267 
1268 	switch (apic_intr_mode) {
1269 	case APIC_PIC:
1270 	case APIC_VIRTUAL_WIRE_NO_CONFIG:
1271 		disable_smp();
1272 		return;
1273 	case APIC_SYMMETRIC_IO_NO_ROUTING:
1274 		disable_smp();
1275 		/* Setup local timer */
1276 		x86_init.timers.setup_percpu_clockev();
1277 		return;
1278 	case APIC_VIRTUAL_WIRE:
1279 	case APIC_SYMMETRIC_IO:
1280 		break;
1281 	}
1282 
1283 	/* Setup local timer */
1284 	x86_init.timers.setup_percpu_clockev();
1285 
1286 	smp_get_logical_apicid();
1287 
1288 	pr_info("CPU0: ");
1289 	print_cpu_info(&cpu_data(0));
1290 
1291 	native_pv_lock_init();
1292 
1293 	uv_system_init();
1294 
1295 	set_mtrr_aps_delayed_init();
1296 
1297 	smp_quirk_init_udelay();
1298 
1299 	speculative_store_bypass_ht_init();
1300 }
1301 
1302 void arch_enable_nonboot_cpus_begin(void)
1303 {
1304 	set_mtrr_aps_delayed_init();
1305 }
1306 
1307 void arch_enable_nonboot_cpus_end(void)
1308 {
1309 	mtrr_aps_init();
1310 }
1311 
1312 /*
1313  * Early setup to make printk work.
1314  */
1315 void __init native_smp_prepare_boot_cpu(void)
1316 {
1317 	int me = smp_processor_id();
1318 	switch_to_new_gdt(me);
1319 	/* already set me in cpu_online_mask in boot_cpu_init() */
1320 	cpumask_set_cpu(me, cpu_callout_mask);
1321 	cpu_set_state_online(me);
1322 }
1323 
1324 void __init calculate_max_logical_packages(void)
1325 {
1326 	int ncpus;
1327 
1328 	/*
1329 	 * Today neither Intel nor AMD support heterogenous systems so
1330 	 * extrapolate the boot cpu's data to all packages.
1331 	 */
1332 	ncpus = cpu_data(0).booted_cores * topology_max_smt_threads();
1333 	__max_logical_packages = DIV_ROUND_UP(nr_cpu_ids, ncpus);
1334 	pr_info("Max logical packages: %u\n", __max_logical_packages);
1335 }
1336 
1337 void __init native_smp_cpus_done(unsigned int max_cpus)
1338 {
1339 	pr_debug("Boot done\n");
1340 
1341 	calculate_max_logical_packages();
1342 
1343 	if (x86_has_numa_in_package)
1344 		set_sched_topology(x86_numa_in_package_topology);
1345 
1346 	nmi_selftest();
1347 	impress_friends();
1348 	mtrr_aps_init();
1349 }
1350 
1351 static int __initdata setup_possible_cpus = -1;
1352 static int __init _setup_possible_cpus(char *str)
1353 {
1354 	get_option(&str, &setup_possible_cpus);
1355 	return 0;
1356 }
1357 early_param("possible_cpus", _setup_possible_cpus);
1358 
1359 
1360 /*
1361  * cpu_possible_mask should be static, it cannot change as cpu's
1362  * are onlined, or offlined. The reason is per-cpu data-structures
1363  * are allocated by some modules at init time, and dont expect to
1364  * do this dynamically on cpu arrival/departure.
1365  * cpu_present_mask on the other hand can change dynamically.
1366  * In case when cpu_hotplug is not compiled, then we resort to current
1367  * behaviour, which is cpu_possible == cpu_present.
1368  * - Ashok Raj
1369  *
1370  * Three ways to find out the number of additional hotplug CPUs:
1371  * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1372  * - The user can overwrite it with possible_cpus=NUM
1373  * - Otherwise don't reserve additional CPUs.
1374  * We do this because additional CPUs waste a lot of memory.
1375  * -AK
1376  */
1377 __init void prefill_possible_map(void)
1378 {
1379 	int i, possible;
1380 
1381 	/* No boot processor was found in mptable or ACPI MADT */
1382 	if (!num_processors) {
1383 		if (boot_cpu_has(X86_FEATURE_APIC)) {
1384 			int apicid = boot_cpu_physical_apicid;
1385 			int cpu = hard_smp_processor_id();
1386 
1387 			pr_warn("Boot CPU (id %d) not listed by BIOS\n", cpu);
1388 
1389 			/* Make sure boot cpu is enumerated */
1390 			if (apic->cpu_present_to_apicid(0) == BAD_APICID &&
1391 			    apic->apic_id_valid(apicid))
1392 				generic_processor_info(apicid, boot_cpu_apic_version);
1393 		}
1394 
1395 		if (!num_processors)
1396 			num_processors = 1;
1397 	}
1398 
1399 	i = setup_max_cpus ?: 1;
1400 	if (setup_possible_cpus == -1) {
1401 		possible = num_processors;
1402 #ifdef CONFIG_HOTPLUG_CPU
1403 		if (setup_max_cpus)
1404 			possible += disabled_cpus;
1405 #else
1406 		if (possible > i)
1407 			possible = i;
1408 #endif
1409 	} else
1410 		possible = setup_possible_cpus;
1411 
1412 	total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1413 
1414 	/* nr_cpu_ids could be reduced via nr_cpus= */
1415 	if (possible > nr_cpu_ids) {
1416 		pr_warn("%d Processors exceeds NR_CPUS limit of %u\n",
1417 			possible, nr_cpu_ids);
1418 		possible = nr_cpu_ids;
1419 	}
1420 
1421 #ifdef CONFIG_HOTPLUG_CPU
1422 	if (!setup_max_cpus)
1423 #endif
1424 	if (possible > i) {
1425 		pr_warn("%d Processors exceeds max_cpus limit of %u\n",
1426 			possible, setup_max_cpus);
1427 		possible = i;
1428 	}
1429 
1430 	nr_cpu_ids = possible;
1431 
1432 	pr_info("Allowing %d CPUs, %d hotplug CPUs\n",
1433 		possible, max_t(int, possible - num_processors, 0));
1434 
1435 	reset_cpu_possible_mask();
1436 
1437 	for (i = 0; i < possible; i++)
1438 		set_cpu_possible(i, true);
1439 }
1440 
1441 #ifdef CONFIG_HOTPLUG_CPU
1442 
1443 /* Recompute SMT state for all CPUs on offline */
1444 static void recompute_smt_state(void)
1445 {
1446 	int max_threads, cpu;
1447 
1448 	max_threads = 0;
1449 	for_each_online_cpu (cpu) {
1450 		int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1451 
1452 		if (threads > max_threads)
1453 			max_threads = threads;
1454 	}
1455 	__max_smt_threads = max_threads;
1456 }
1457 
1458 static void remove_siblinginfo(int cpu)
1459 {
1460 	int sibling;
1461 	struct cpuinfo_x86 *c = &cpu_data(cpu);
1462 
1463 	for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1464 		cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1465 		/*/
1466 		 * last thread sibling in this cpu core going down
1467 		 */
1468 		if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1469 			cpu_data(sibling).booted_cores--;
1470 	}
1471 
1472 	for_each_cpu(sibling, topology_sibling_cpumask(cpu))
1473 		cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1474 	for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1475 		cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1476 	cpumask_clear(cpu_llc_shared_mask(cpu));
1477 	cpumask_clear(topology_sibling_cpumask(cpu));
1478 	cpumask_clear(topology_core_cpumask(cpu));
1479 	c->cpu_core_id = 0;
1480 	c->booted_cores = 0;
1481 	cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1482 	recompute_smt_state();
1483 }
1484 
1485 static void remove_cpu_from_maps(int cpu)
1486 {
1487 	set_cpu_online(cpu, false);
1488 	cpumask_clear_cpu(cpu, cpu_callout_mask);
1489 	cpumask_clear_cpu(cpu, cpu_callin_mask);
1490 	/* was set by cpu_init() */
1491 	cpumask_clear_cpu(cpu, cpu_initialized_mask);
1492 	numa_remove_cpu(cpu);
1493 }
1494 
1495 void cpu_disable_common(void)
1496 {
1497 	int cpu = smp_processor_id();
1498 
1499 	remove_siblinginfo(cpu);
1500 
1501 	/* It's now safe to remove this processor from the online map */
1502 	lock_vector_lock();
1503 	remove_cpu_from_maps(cpu);
1504 	unlock_vector_lock();
1505 	fixup_irqs();
1506 	lapic_offline();
1507 }
1508 
1509 int native_cpu_disable(void)
1510 {
1511 	int ret;
1512 
1513 	ret = lapic_can_unplug_cpu();
1514 	if (ret)
1515 		return ret;
1516 
1517 	clear_local_APIC();
1518 	cpu_disable_common();
1519 
1520 	return 0;
1521 }
1522 
1523 int common_cpu_die(unsigned int cpu)
1524 {
1525 	int ret = 0;
1526 
1527 	/* We don't do anything here: idle task is faking death itself. */
1528 
1529 	/* They ack this in play_dead() by setting CPU_DEAD */
1530 	if (cpu_wait_death(cpu, 5)) {
1531 		if (system_state == SYSTEM_RUNNING)
1532 			pr_info("CPU %u is now offline\n", cpu);
1533 	} else {
1534 		pr_err("CPU %u didn't die...\n", cpu);
1535 		ret = -1;
1536 	}
1537 
1538 	return ret;
1539 }
1540 
1541 void native_cpu_die(unsigned int cpu)
1542 {
1543 	common_cpu_die(cpu);
1544 }
1545 
1546 void play_dead_common(void)
1547 {
1548 	idle_task_exit();
1549 
1550 	/* Ack it */
1551 	(void)cpu_report_death();
1552 
1553 	/*
1554 	 * With physical CPU hotplug, we should halt the cpu
1555 	 */
1556 	local_irq_disable();
1557 }
1558 
1559 static bool wakeup_cpu0(void)
1560 {
1561 	if (smp_processor_id() == 0 && enable_start_cpu0)
1562 		return true;
1563 
1564 	return false;
1565 }
1566 
1567 /*
1568  * We need to flush the caches before going to sleep, lest we have
1569  * dirty data in our caches when we come back up.
1570  */
1571 static inline void mwait_play_dead(void)
1572 {
1573 	unsigned int eax, ebx, ecx, edx;
1574 	unsigned int highest_cstate = 0;
1575 	unsigned int highest_subcstate = 0;
1576 	void *mwait_ptr;
1577 	int i;
1578 
1579 	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
1580 		return;
1581 	if (!this_cpu_has(X86_FEATURE_MWAIT))
1582 		return;
1583 	if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1584 		return;
1585 	if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1586 		return;
1587 
1588 	eax = CPUID_MWAIT_LEAF;
1589 	ecx = 0;
1590 	native_cpuid(&eax, &ebx, &ecx, &edx);
1591 
1592 	/*
1593 	 * eax will be 0 if EDX enumeration is not valid.
1594 	 * Initialized below to cstate, sub_cstate value when EDX is valid.
1595 	 */
1596 	if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1597 		eax = 0;
1598 	} else {
1599 		edx >>= MWAIT_SUBSTATE_SIZE;
1600 		for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1601 			if (edx & MWAIT_SUBSTATE_MASK) {
1602 				highest_cstate = i;
1603 				highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1604 			}
1605 		}
1606 		eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1607 			(highest_subcstate - 1);
1608 	}
1609 
1610 	/*
1611 	 * This should be a memory location in a cache line which is
1612 	 * unlikely to be touched by other processors.  The actual
1613 	 * content is immaterial as it is not actually modified in any way.
1614 	 */
1615 	mwait_ptr = &current_thread_info()->flags;
1616 
1617 	wbinvd();
1618 
1619 	while (1) {
1620 		/*
1621 		 * The CLFLUSH is a workaround for erratum AAI65 for
1622 		 * the Xeon 7400 series.  It's not clear it is actually
1623 		 * needed, but it should be harmless in either case.
1624 		 * The WBINVD is insufficient due to the spurious-wakeup
1625 		 * case where we return around the loop.
1626 		 */
1627 		mb();
1628 		clflush(mwait_ptr);
1629 		mb();
1630 		__monitor(mwait_ptr, 0, 0);
1631 		mb();
1632 		__mwait(eax, 0);
1633 		/*
1634 		 * If NMI wants to wake up CPU0, start CPU0.
1635 		 */
1636 		if (wakeup_cpu0())
1637 			start_cpu0();
1638 	}
1639 }
1640 
1641 void hlt_play_dead(void)
1642 {
1643 	if (__this_cpu_read(cpu_info.x86) >= 4)
1644 		wbinvd();
1645 
1646 	while (1) {
1647 		native_halt();
1648 		/*
1649 		 * If NMI wants to wake up CPU0, start CPU0.
1650 		 */
1651 		if (wakeup_cpu0())
1652 			start_cpu0();
1653 	}
1654 }
1655 
1656 void native_play_dead(void)
1657 {
1658 	play_dead_common();
1659 	tboot_shutdown(TB_SHUTDOWN_WFS);
1660 
1661 	mwait_play_dead();	/* Only returns on failure */
1662 	if (cpuidle_play_dead())
1663 		hlt_play_dead();
1664 }
1665 
1666 #else /* ... !CONFIG_HOTPLUG_CPU */
1667 int native_cpu_disable(void)
1668 {
1669 	return -ENOSYS;
1670 }
1671 
1672 void native_cpu_die(unsigned int cpu)
1673 {
1674 	/* We said "no" in __cpu_disable */
1675 	BUG();
1676 }
1677 
1678 void native_play_dead(void)
1679 {
1680 	BUG();
1681 }
1682 
1683 #endif
1684