xref: /linux/arch/x86/kernel/smpboot.c (revision 10accd2e6890b57db8e717e9aee91b791f90fe14)
1  /*
2  *	x86 SMP booting functions
3  *
4  *	(c) 1995 Alan Cox, Building #3 <alan@lxorguk.ukuu.org.uk>
5  *	(c) 1998, 1999, 2000, 2009 Ingo Molnar <mingo@redhat.com>
6  *	Copyright 2001 Andi Kleen, SuSE Labs.
7  *
8  *	Much of the core SMP work is based on previous work by Thomas Radke, to
9  *	whom a great many thanks are extended.
10  *
11  *	Thanks to Intel for making available several different Pentium,
12  *	Pentium Pro and Pentium-II/Xeon MP machines.
13  *	Original development of Linux SMP code supported by Caldera.
14  *
15  *	This code is released under the GNU General Public License version 2 or
16  *	later.
17  *
18  *	Fixes
19  *		Felix Koop	:	NR_CPUS used properly
20  *		Jose Renau	:	Handle single CPU case.
21  *		Alan Cox	:	By repeated request 8) - Total BogoMIPS report.
22  *		Greg Wright	:	Fix for kernel stacks panic.
23  *		Erich Boleyn	:	MP v1.4 and additional changes.
24  *	Matthias Sattler	:	Changes for 2.1 kernel map.
25  *	Michel Lespinasse	:	Changes for 2.1 kernel map.
26  *	Michael Chastain	:	Change trampoline.S to gnu as.
27  *		Alan Cox	:	Dumb bug: 'B' step PPro's are fine
28  *		Ingo Molnar	:	Added APIC timers, based on code
29  *					from Jose Renau
30  *		Ingo Molnar	:	various cleanups and rewrites
31  *		Tigran Aivazian	:	fixed "0.00 in /proc/uptime on SMP" bug.
32  *	Maciej W. Rozycki	:	Bits for genuine 82489DX APICs
33  *	Andi Kleen		:	Changed for SMP boot into long mode.
34  *		Martin J. Bligh	: 	Added support for multi-quad systems
35  *		Dave Jones	:	Report invalid combinations of Athlon CPUs.
36  *		Rusty Russell	:	Hacked into shape for new "hotplug" boot process.
37  *      Andi Kleen              :       Converted to new state machine.
38  *	Ashok Raj		: 	CPU hotplug support
39  *	Glauber Costa		:	i386 and x86_64 integration
40  */
41 
42 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
43 
44 #include <linux/init.h>
45 #include <linux/smp.h>
46 #include <linux/export.h>
47 #include <linux/sched.h>
48 #include <linux/percpu.h>
49 #include <linux/bootmem.h>
50 #include <linux/err.h>
51 #include <linux/nmi.h>
52 #include <linux/tboot.h>
53 #include <linux/stackprotector.h>
54 #include <linux/gfp.h>
55 #include <linux/cpuidle.h>
56 
57 #include <asm/acpi.h>
58 #include <asm/desc.h>
59 #include <asm/nmi.h>
60 #include <asm/irq.h>
61 #include <asm/idle.h>
62 #include <asm/realmode.h>
63 #include <asm/cpu.h>
64 #include <asm/numa.h>
65 #include <asm/pgtable.h>
66 #include <asm/tlbflush.h>
67 #include <asm/mtrr.h>
68 #include <asm/mwait.h>
69 #include <asm/apic.h>
70 #include <asm/io_apic.h>
71 #include <asm/fpu/internal.h>
72 #include <asm/setup.h>
73 #include <asm/uv/uv.h>
74 #include <linux/mc146818rtc.h>
75 #include <asm/i8259.h>
76 #include <asm/realmode.h>
77 #include <asm/misc.h>
78 
79 /* Number of siblings per CPU package */
80 int smp_num_siblings = 1;
81 EXPORT_SYMBOL(smp_num_siblings);
82 
83 /* Last level cache ID of each logical CPU */
84 DEFINE_PER_CPU_READ_MOSTLY(u16, cpu_llc_id) = BAD_APICID;
85 
86 /* representing HT siblings of each logical CPU */
87 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
88 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
89 
90 /* representing HT and core siblings of each logical CPU */
91 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
92 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
93 
94 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_llc_shared_map);
95 
96 /* Per CPU bogomips and other parameters */
97 DEFINE_PER_CPU_READ_MOSTLY(struct cpuinfo_x86, cpu_info);
98 EXPORT_PER_CPU_SYMBOL(cpu_info);
99 
100 /* Logical package management. We might want to allocate that dynamically */
101 static int *physical_to_logical_pkg __read_mostly;
102 static unsigned long *physical_package_map __read_mostly;;
103 static unsigned long *logical_package_map  __read_mostly;
104 static unsigned int max_physical_pkg_id __read_mostly;
105 unsigned int __max_logical_packages __read_mostly;
106 EXPORT_SYMBOL(__max_logical_packages);
107 
108 /* Maximum number of SMT threads on any online core */
109 int __max_smt_threads __read_mostly;
110 
111 static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
112 {
113 	unsigned long flags;
114 
115 	spin_lock_irqsave(&rtc_lock, flags);
116 	CMOS_WRITE(0xa, 0xf);
117 	spin_unlock_irqrestore(&rtc_lock, flags);
118 	local_flush_tlb();
119 	pr_debug("1.\n");
120 	*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) =
121 							start_eip >> 4;
122 	pr_debug("2.\n");
123 	*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) =
124 							start_eip & 0xf;
125 	pr_debug("3.\n");
126 }
127 
128 static inline void smpboot_restore_warm_reset_vector(void)
129 {
130 	unsigned long flags;
131 
132 	/*
133 	 * Install writable page 0 entry to set BIOS data area.
134 	 */
135 	local_flush_tlb();
136 
137 	/*
138 	 * Paranoid:  Set warm reset code and vector here back
139 	 * to default values.
140 	 */
141 	spin_lock_irqsave(&rtc_lock, flags);
142 	CMOS_WRITE(0, 0xf);
143 	spin_unlock_irqrestore(&rtc_lock, flags);
144 
145 	*((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
146 }
147 
148 /*
149  * Report back to the Boot Processor during boot time or to the caller processor
150  * during CPU online.
151  */
152 static void smp_callin(void)
153 {
154 	int cpuid, phys_id;
155 
156 	/*
157 	 * If waken up by an INIT in an 82489DX configuration
158 	 * cpu_callout_mask guarantees we don't get here before
159 	 * an INIT_deassert IPI reaches our local APIC, so it is
160 	 * now safe to touch our local APIC.
161 	 */
162 	cpuid = smp_processor_id();
163 
164 	/*
165 	 * (This works even if the APIC is not enabled.)
166 	 */
167 	phys_id = read_apic_id();
168 
169 	/*
170 	 * the boot CPU has finished the init stage and is spinning
171 	 * on callin_map until we finish. We are free to set up this
172 	 * CPU, first the APIC. (this is probably redundant on most
173 	 * boards)
174 	 */
175 	apic_ap_setup();
176 
177 	/*
178 	 * Save our processor parameters. Note: this information
179 	 * is needed for clock calibration.
180 	 */
181 	smp_store_cpu_info(cpuid);
182 
183 	/*
184 	 * Get our bogomips.
185 	 * Update loops_per_jiffy in cpu_data. Previous call to
186 	 * smp_store_cpu_info() stored a value that is close but not as
187 	 * accurate as the value just calculated.
188 	 */
189 	calibrate_delay();
190 	cpu_data(cpuid).loops_per_jiffy = loops_per_jiffy;
191 	pr_debug("Stack at about %p\n", &cpuid);
192 
193 	/*
194 	 * This must be done before setting cpu_online_mask
195 	 * or calling notify_cpu_starting.
196 	 */
197 	set_cpu_sibling_map(raw_smp_processor_id());
198 	wmb();
199 
200 	notify_cpu_starting(cpuid);
201 
202 	/*
203 	 * Allow the master to continue.
204 	 */
205 	cpumask_set_cpu(cpuid, cpu_callin_mask);
206 }
207 
208 static int cpu0_logical_apicid;
209 static int enable_start_cpu0;
210 /*
211  * Activate a secondary processor.
212  */
213 static void notrace start_secondary(void *unused)
214 {
215 	/*
216 	 * Don't put *anything* before cpu_init(), SMP booting is too
217 	 * fragile that we want to limit the things done here to the
218 	 * most necessary things.
219 	 */
220 	cpu_init();
221 	x86_cpuinit.early_percpu_clock_init();
222 	preempt_disable();
223 	smp_callin();
224 
225 	enable_start_cpu0 = 0;
226 
227 #ifdef CONFIG_X86_32
228 	/* switch away from the initial page table */
229 	load_cr3(swapper_pg_dir);
230 	__flush_tlb_all();
231 #endif
232 
233 	/* otherwise gcc will move up smp_processor_id before the cpu_init */
234 	barrier();
235 	/*
236 	 * Check TSC synchronization with the BP:
237 	 */
238 	check_tsc_sync_target();
239 
240 	/*
241 	 * Lock vector_lock and initialize the vectors on this cpu
242 	 * before setting the cpu online. We must set it online with
243 	 * vector_lock held to prevent a concurrent setup/teardown
244 	 * from seeing a half valid vector space.
245 	 */
246 	lock_vector_lock();
247 	setup_vector_irq(smp_processor_id());
248 	set_cpu_online(smp_processor_id(), true);
249 	unlock_vector_lock();
250 	cpu_set_state_online(smp_processor_id());
251 	x86_platform.nmi_init();
252 
253 	/* enable local interrupts */
254 	local_irq_enable();
255 
256 	/* to prevent fake stack check failure in clock setup */
257 	boot_init_stack_canary();
258 
259 	x86_cpuinit.setup_percpu_clockev();
260 
261 	wmb();
262 	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
263 }
264 
265 int topology_update_package_map(unsigned int apicid, unsigned int cpu)
266 {
267 	unsigned int new, pkg = apicid >> boot_cpu_data.x86_coreid_bits;
268 
269 	/* Called from early boot ? */
270 	if (!physical_package_map)
271 		return 0;
272 
273 	if (pkg >= max_physical_pkg_id)
274 		return -EINVAL;
275 
276 	/* Set the logical package id */
277 	if (test_and_set_bit(pkg, physical_package_map))
278 		goto found;
279 
280 	new = find_first_zero_bit(logical_package_map, __max_logical_packages);
281 	if (new >= __max_logical_packages) {
282 		physical_to_logical_pkg[pkg] = -1;
283 		pr_warn("APIC(%x) Package %u exceeds logical package map\n",
284 			apicid, pkg);
285 		return -ENOSPC;
286 	}
287 	set_bit(new, logical_package_map);
288 	pr_info("APIC(%x) Converting physical %u to logical package %u\n",
289 		apicid, pkg, new);
290 	physical_to_logical_pkg[pkg] = new;
291 
292 found:
293 	cpu_data(cpu).logical_proc_id = physical_to_logical_pkg[pkg];
294 	return 0;
295 }
296 
297 /**
298  * topology_phys_to_logical_pkg - Map a physical package id to a logical
299  *
300  * Returns logical package id or -1 if not found
301  */
302 int topology_phys_to_logical_pkg(unsigned int phys_pkg)
303 {
304 	if (phys_pkg >= max_physical_pkg_id)
305 		return -1;
306 	return physical_to_logical_pkg[phys_pkg];
307 }
308 EXPORT_SYMBOL(topology_phys_to_logical_pkg);
309 
310 static void __init smp_init_package_map(void)
311 {
312 	unsigned int ncpus, cpu;
313 	size_t size;
314 
315 	/*
316 	 * Today neither Intel nor AMD support heterogenous systems. That
317 	 * might change in the future....
318 	 *
319 	 * While ideally we'd want '* smp_num_siblings' in the below @ncpus
320 	 * computation, this won't actually work since some Intel BIOSes
321 	 * report inconsistent HT data when they disable HT.
322 	 *
323 	 * In particular, they reduce the APIC-IDs to only include the cores,
324 	 * but leave the CPUID topology to say there are (2) siblings.
325 	 * This means we don't know how many threads there will be until
326 	 * after the APIC enumeration.
327 	 *
328 	 * By not including this we'll sometimes over-estimate the number of
329 	 * logical packages by the amount of !present siblings, but this is
330 	 * still better than MAX_LOCAL_APIC.
331 	 *
332 	 * We use total_cpus not nr_cpu_ids because nr_cpu_ids can be limited
333 	 * on the command line leading to a similar issue as the HT disable
334 	 * problem because the hyperthreads are usually enumerated after the
335 	 * primary cores.
336 	 */
337 	ncpus = boot_cpu_data.x86_max_cores;
338 	if (!ncpus) {
339 		pr_warn("x86_max_cores == zero !?!?");
340 		ncpus = 1;
341 	}
342 
343 	__max_logical_packages = DIV_ROUND_UP(total_cpus, ncpus);
344 
345 	/*
346 	 * Possibly larger than what we need as the number of apic ids per
347 	 * package can be smaller than the actual used apic ids.
348 	 */
349 	max_physical_pkg_id = DIV_ROUND_UP(MAX_LOCAL_APIC, ncpus);
350 	size = max_physical_pkg_id * sizeof(unsigned int);
351 	physical_to_logical_pkg = kmalloc(size, GFP_KERNEL);
352 	memset(physical_to_logical_pkg, 0xff, size);
353 	size = BITS_TO_LONGS(max_physical_pkg_id) * sizeof(unsigned long);
354 	physical_package_map = kzalloc(size, GFP_KERNEL);
355 	size = BITS_TO_LONGS(__max_logical_packages) * sizeof(unsigned long);
356 	logical_package_map = kzalloc(size, GFP_KERNEL);
357 
358 	pr_info("Max logical packages: %u\n", __max_logical_packages);
359 
360 	for_each_present_cpu(cpu) {
361 		unsigned int apicid = apic->cpu_present_to_apicid(cpu);
362 
363 		if (apicid == BAD_APICID || !apic->apic_id_valid(apicid))
364 			continue;
365 		if (!topology_update_package_map(apicid, cpu))
366 			continue;
367 		pr_warn("CPU %u APICId %x disabled\n", cpu, apicid);
368 		per_cpu(x86_bios_cpu_apicid, cpu) = BAD_APICID;
369 		set_cpu_possible(cpu, false);
370 		set_cpu_present(cpu, false);
371 	}
372 }
373 
374 void __init smp_store_boot_cpu_info(void)
375 {
376 	int id = 0; /* CPU 0 */
377 	struct cpuinfo_x86 *c = &cpu_data(id);
378 
379 	*c = boot_cpu_data;
380 	c->cpu_index = id;
381 	smp_init_package_map();
382 }
383 
384 /*
385  * The bootstrap kernel entry code has set these up. Save them for
386  * a given CPU
387  */
388 void smp_store_cpu_info(int id)
389 {
390 	struct cpuinfo_x86 *c = &cpu_data(id);
391 
392 	*c = boot_cpu_data;
393 	c->cpu_index = id;
394 	/*
395 	 * During boot time, CPU0 has this setup already. Save the info when
396 	 * bringing up AP or offlined CPU0.
397 	 */
398 	identify_secondary_cpu(c);
399 }
400 
401 static bool
402 topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
403 {
404 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
405 
406 	return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
407 }
408 
409 static bool
410 topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
411 {
412 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
413 
414 	return !WARN_ONCE(!topology_same_node(c, o),
415 		"sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
416 		"[node: %d != %d]. Ignoring dependency.\n",
417 		cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
418 }
419 
420 #define link_mask(mfunc, c1, c2)					\
421 do {									\
422 	cpumask_set_cpu((c1), mfunc(c2));				\
423 	cpumask_set_cpu((c2), mfunc(c1));				\
424 } while (0)
425 
426 static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
427 {
428 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
429 		int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
430 
431 		if (c->phys_proc_id == o->phys_proc_id &&
432 		    per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2) &&
433 		    c->cpu_core_id == o->cpu_core_id)
434 			return topology_sane(c, o, "smt");
435 
436 	} else if (c->phys_proc_id == o->phys_proc_id &&
437 		   c->cpu_core_id == o->cpu_core_id) {
438 		return topology_sane(c, o, "smt");
439 	}
440 
441 	return false;
442 }
443 
444 static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
445 {
446 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
447 
448 	if (per_cpu(cpu_llc_id, cpu1) != BAD_APICID &&
449 	    per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2))
450 		return topology_sane(c, o, "llc");
451 
452 	return false;
453 }
454 
455 /*
456  * Unlike the other levels, we do not enforce keeping a
457  * multicore group inside a NUMA node.  If this happens, we will
458  * discard the MC level of the topology later.
459  */
460 static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
461 {
462 	if (c->phys_proc_id == o->phys_proc_id)
463 		return true;
464 	return false;
465 }
466 
467 static struct sched_domain_topology_level numa_inside_package_topology[] = {
468 #ifdef CONFIG_SCHED_SMT
469 	{ cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
470 #endif
471 #ifdef CONFIG_SCHED_MC
472 	{ cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
473 #endif
474 	{ NULL, },
475 };
476 /*
477  * set_sched_topology() sets the topology internal to a CPU.  The
478  * NUMA topologies are layered on top of it to build the full
479  * system topology.
480  *
481  * If NUMA nodes are observed to occur within a CPU package, this
482  * function should be called.  It forces the sched domain code to
483  * only use the SMT level for the CPU portion of the topology.
484  * This essentially falls back to relying on NUMA information
485  * from the SRAT table to describe the entire system topology
486  * (except for hyperthreads).
487  */
488 static void primarily_use_numa_for_topology(void)
489 {
490 	set_sched_topology(numa_inside_package_topology);
491 }
492 
493 void set_cpu_sibling_map(int cpu)
494 {
495 	bool has_smt = smp_num_siblings > 1;
496 	bool has_mp = has_smt || boot_cpu_data.x86_max_cores > 1;
497 	struct cpuinfo_x86 *c = &cpu_data(cpu);
498 	struct cpuinfo_x86 *o;
499 	int i, threads;
500 
501 	cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
502 
503 	if (!has_mp) {
504 		cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
505 		cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
506 		cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
507 		c->booted_cores = 1;
508 		return;
509 	}
510 
511 	for_each_cpu(i, cpu_sibling_setup_mask) {
512 		o = &cpu_data(i);
513 
514 		if ((i == cpu) || (has_smt && match_smt(c, o)))
515 			link_mask(topology_sibling_cpumask, cpu, i);
516 
517 		if ((i == cpu) || (has_mp && match_llc(c, o)))
518 			link_mask(cpu_llc_shared_mask, cpu, i);
519 
520 	}
521 
522 	/*
523 	 * This needs a separate iteration over the cpus because we rely on all
524 	 * topology_sibling_cpumask links to be set-up.
525 	 */
526 	for_each_cpu(i, cpu_sibling_setup_mask) {
527 		o = &cpu_data(i);
528 
529 		if ((i == cpu) || (has_mp && match_die(c, o))) {
530 			link_mask(topology_core_cpumask, cpu, i);
531 
532 			/*
533 			 *  Does this new cpu bringup a new core?
534 			 */
535 			if (cpumask_weight(
536 			    topology_sibling_cpumask(cpu)) == 1) {
537 				/*
538 				 * for each core in package, increment
539 				 * the booted_cores for this new cpu
540 				 */
541 				if (cpumask_first(
542 				    topology_sibling_cpumask(i)) == i)
543 					c->booted_cores++;
544 				/*
545 				 * increment the core count for all
546 				 * the other cpus in this package
547 				 */
548 				if (i != cpu)
549 					cpu_data(i).booted_cores++;
550 			} else if (i != cpu && !c->booted_cores)
551 				c->booted_cores = cpu_data(i).booted_cores;
552 		}
553 		if (match_die(c, o) && !topology_same_node(c, o))
554 			primarily_use_numa_for_topology();
555 	}
556 
557 	threads = cpumask_weight(topology_sibling_cpumask(cpu));
558 	if (threads > __max_smt_threads)
559 		__max_smt_threads = threads;
560 }
561 
562 /* maps the cpu to the sched domain representing multi-core */
563 const struct cpumask *cpu_coregroup_mask(int cpu)
564 {
565 	return cpu_llc_shared_mask(cpu);
566 }
567 
568 static void impress_friends(void)
569 {
570 	int cpu;
571 	unsigned long bogosum = 0;
572 	/*
573 	 * Allow the user to impress friends.
574 	 */
575 	pr_debug("Before bogomips\n");
576 	for_each_possible_cpu(cpu)
577 		if (cpumask_test_cpu(cpu, cpu_callout_mask))
578 			bogosum += cpu_data(cpu).loops_per_jiffy;
579 	pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
580 		num_online_cpus(),
581 		bogosum/(500000/HZ),
582 		(bogosum/(5000/HZ))%100);
583 
584 	pr_debug("Before bogocount - setting activated=1\n");
585 }
586 
587 void __inquire_remote_apic(int apicid)
588 {
589 	unsigned i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
590 	const char * const names[] = { "ID", "VERSION", "SPIV" };
591 	int timeout;
592 	u32 status;
593 
594 	pr_info("Inquiring remote APIC 0x%x...\n", apicid);
595 
596 	for (i = 0; i < ARRAY_SIZE(regs); i++) {
597 		pr_info("... APIC 0x%x %s: ", apicid, names[i]);
598 
599 		/*
600 		 * Wait for idle.
601 		 */
602 		status = safe_apic_wait_icr_idle();
603 		if (status)
604 			pr_cont("a previous APIC delivery may have failed\n");
605 
606 		apic_icr_write(APIC_DM_REMRD | regs[i], apicid);
607 
608 		timeout = 0;
609 		do {
610 			udelay(100);
611 			status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
612 		} while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
613 
614 		switch (status) {
615 		case APIC_ICR_RR_VALID:
616 			status = apic_read(APIC_RRR);
617 			pr_cont("%08x\n", status);
618 			break;
619 		default:
620 			pr_cont("failed\n");
621 		}
622 	}
623 }
624 
625 /*
626  * The Multiprocessor Specification 1.4 (1997) example code suggests
627  * that there should be a 10ms delay between the BSP asserting INIT
628  * and de-asserting INIT, when starting a remote processor.
629  * But that slows boot and resume on modern processors, which include
630  * many cores and don't require that delay.
631  *
632  * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
633  * Modern processor families are quirked to remove the delay entirely.
634  */
635 #define UDELAY_10MS_DEFAULT 10000
636 
637 static unsigned int init_udelay = UINT_MAX;
638 
639 static int __init cpu_init_udelay(char *str)
640 {
641 	get_option(&str, &init_udelay);
642 
643 	return 0;
644 }
645 early_param("cpu_init_udelay", cpu_init_udelay);
646 
647 static void __init smp_quirk_init_udelay(void)
648 {
649 	/* if cmdline changed it from default, leave it alone */
650 	if (init_udelay != UINT_MAX)
651 		return;
652 
653 	/* if modern processor, use no delay */
654 	if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
655 	    ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
656 		init_udelay = 0;
657 		return;
658 	}
659 	/* else, use legacy delay */
660 	init_udelay = UDELAY_10MS_DEFAULT;
661 }
662 
663 /*
664  * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
665  * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
666  * won't ... remember to clear down the APIC, etc later.
667  */
668 int
669 wakeup_secondary_cpu_via_nmi(int apicid, unsigned long start_eip)
670 {
671 	unsigned long send_status, accept_status = 0;
672 	int maxlvt;
673 
674 	/* Target chip */
675 	/* Boot on the stack */
676 	/* Kick the second */
677 	apic_icr_write(APIC_DM_NMI | apic->dest_logical, apicid);
678 
679 	pr_debug("Waiting for send to finish...\n");
680 	send_status = safe_apic_wait_icr_idle();
681 
682 	/*
683 	 * Give the other CPU some time to accept the IPI.
684 	 */
685 	udelay(200);
686 	if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid])) {
687 		maxlvt = lapic_get_maxlvt();
688 		if (maxlvt > 3)			/* Due to the Pentium erratum 3AP.  */
689 			apic_write(APIC_ESR, 0);
690 		accept_status = (apic_read(APIC_ESR) & 0xEF);
691 	}
692 	pr_debug("NMI sent\n");
693 
694 	if (send_status)
695 		pr_err("APIC never delivered???\n");
696 	if (accept_status)
697 		pr_err("APIC delivery error (%lx)\n", accept_status);
698 
699 	return (send_status | accept_status);
700 }
701 
702 static int
703 wakeup_secondary_cpu_via_init(int phys_apicid, unsigned long start_eip)
704 {
705 	unsigned long send_status = 0, accept_status = 0;
706 	int maxlvt, num_starts, j;
707 
708 	maxlvt = lapic_get_maxlvt();
709 
710 	/*
711 	 * Be paranoid about clearing APIC errors.
712 	 */
713 	if (APIC_INTEGRATED(apic_version[phys_apicid])) {
714 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
715 			apic_write(APIC_ESR, 0);
716 		apic_read(APIC_ESR);
717 	}
718 
719 	pr_debug("Asserting INIT\n");
720 
721 	/*
722 	 * Turn INIT on target chip
723 	 */
724 	/*
725 	 * Send IPI
726 	 */
727 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT,
728 		       phys_apicid);
729 
730 	pr_debug("Waiting for send to finish...\n");
731 	send_status = safe_apic_wait_icr_idle();
732 
733 	udelay(init_udelay);
734 
735 	pr_debug("Deasserting INIT\n");
736 
737 	/* Target chip */
738 	/* Send IPI */
739 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
740 
741 	pr_debug("Waiting for send to finish...\n");
742 	send_status = safe_apic_wait_icr_idle();
743 
744 	mb();
745 
746 	/*
747 	 * Should we send STARTUP IPIs ?
748 	 *
749 	 * Determine this based on the APIC version.
750 	 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
751 	 */
752 	if (APIC_INTEGRATED(apic_version[phys_apicid]))
753 		num_starts = 2;
754 	else
755 		num_starts = 0;
756 
757 	/*
758 	 * Run STARTUP IPI loop.
759 	 */
760 	pr_debug("#startup loops: %d\n", num_starts);
761 
762 	for (j = 1; j <= num_starts; j++) {
763 		pr_debug("Sending STARTUP #%d\n", j);
764 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
765 			apic_write(APIC_ESR, 0);
766 		apic_read(APIC_ESR);
767 		pr_debug("After apic_write\n");
768 
769 		/*
770 		 * STARTUP IPI
771 		 */
772 
773 		/* Target chip */
774 		/* Boot on the stack */
775 		/* Kick the second */
776 		apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
777 			       phys_apicid);
778 
779 		/*
780 		 * Give the other CPU some time to accept the IPI.
781 		 */
782 		if (init_udelay == 0)
783 			udelay(10);
784 		else
785 			udelay(300);
786 
787 		pr_debug("Startup point 1\n");
788 
789 		pr_debug("Waiting for send to finish...\n");
790 		send_status = safe_apic_wait_icr_idle();
791 
792 		/*
793 		 * Give the other CPU some time to accept the IPI.
794 		 */
795 		if (init_udelay == 0)
796 			udelay(10);
797 		else
798 			udelay(200);
799 
800 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
801 			apic_write(APIC_ESR, 0);
802 		accept_status = (apic_read(APIC_ESR) & 0xEF);
803 		if (send_status || accept_status)
804 			break;
805 	}
806 	pr_debug("After Startup\n");
807 
808 	if (send_status)
809 		pr_err("APIC never delivered???\n");
810 	if (accept_status)
811 		pr_err("APIC delivery error (%lx)\n", accept_status);
812 
813 	return (send_status | accept_status);
814 }
815 
816 void smp_announce(void)
817 {
818 	int num_nodes = num_online_nodes();
819 
820 	printk(KERN_INFO "x86: Booted up %d node%s, %d CPUs\n",
821 	       num_nodes, (num_nodes > 1 ? "s" : ""), num_online_cpus());
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_BOOTING) {
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) =
936 		(unsigned long)task_stack_page(idle) + THREAD_SIZE;
937 #else
938 	clear_tsk_thread_flag(idle, TIF_FORK);
939 	initial_gs = per_cpu_offset(cpu);
940 #endif
941 }
942 
943 /*
944  * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
945  * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
946  * Returns zero if CPU booted OK, else error code from
947  * ->wakeup_secondary_cpu.
948  */
949 static int do_boot_cpu(int apicid, int cpu, struct task_struct *idle)
950 {
951 	volatile u32 *trampoline_status =
952 		(volatile u32 *) __va(real_mode_header->trampoline_status);
953 	/* start_ip had better be page-aligned! */
954 	unsigned long start_ip = real_mode_header->trampoline_start;
955 
956 	unsigned long boot_error = 0;
957 	int cpu0_nmi_registered = 0;
958 	unsigned long timeout;
959 
960 	idle->thread.sp = (unsigned long) (((struct pt_regs *)
961 			  (THREAD_SIZE +  task_stack_page(idle))) - 1);
962 
963 	early_gdt_descr.address = (unsigned long)get_cpu_gdt_table(cpu);
964 	initial_code = (unsigned long)start_secondary;
965 	stack_start  = idle->thread.sp;
966 
967 	/*
968 	 * Enable the espfix hack for this CPU
969 	*/
970 #ifdef CONFIG_X86_ESPFIX64
971 	init_espfix_ap(cpu);
972 #endif
973 
974 	/* So we see what's up */
975 	announce_cpu(cpu, apicid);
976 
977 	/*
978 	 * This grunge runs the startup process for
979 	 * the targeted processor.
980 	 */
981 
982 	if (get_uv_system_type() != UV_NON_UNIQUE_APIC) {
983 
984 		pr_debug("Setting warm reset code and vector.\n");
985 
986 		smpboot_setup_warm_reset_vector(start_ip);
987 		/*
988 		 * Be paranoid about clearing APIC errors.
989 		*/
990 		if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid])) {
991 			apic_write(APIC_ESR, 0);
992 			apic_read(APIC_ESR);
993 		}
994 	}
995 
996 	/*
997 	 * AP might wait on cpu_callout_mask in cpu_init() with
998 	 * cpu_initialized_mask set if previous attempt to online
999 	 * it timed-out. Clear cpu_initialized_mask so that after
1000 	 * INIT/SIPI it could start with a clean state.
1001 	 */
1002 	cpumask_clear_cpu(cpu, cpu_initialized_mask);
1003 	smp_mb();
1004 
1005 	/*
1006 	 * Wake up a CPU in difference cases:
1007 	 * - Use the method in the APIC driver if it's defined
1008 	 * Otherwise,
1009 	 * - Use an INIT boot APIC message for APs or NMI for BSP.
1010 	 */
1011 	if (apic->wakeup_secondary_cpu)
1012 		boot_error = apic->wakeup_secondary_cpu(apicid, start_ip);
1013 	else
1014 		boot_error = wakeup_cpu_via_init_nmi(cpu, start_ip, apicid,
1015 						     &cpu0_nmi_registered);
1016 
1017 	if (!boot_error) {
1018 		/*
1019 		 * Wait 10s total for first sign of life from AP
1020 		 */
1021 		boot_error = -1;
1022 		timeout = jiffies + 10*HZ;
1023 		while (time_before(jiffies, timeout)) {
1024 			if (cpumask_test_cpu(cpu, cpu_initialized_mask)) {
1025 				/*
1026 				 * Tell AP to proceed with initialization
1027 				 */
1028 				cpumask_set_cpu(cpu, cpu_callout_mask);
1029 				boot_error = 0;
1030 				break;
1031 			}
1032 			schedule();
1033 		}
1034 	}
1035 
1036 	if (!boot_error) {
1037 		/*
1038 		 * Wait till AP completes initial initialization
1039 		 */
1040 		while (!cpumask_test_cpu(cpu, cpu_callin_mask)) {
1041 			/*
1042 			 * Allow other tasks to run while we wait for the
1043 			 * AP to come online. This also gives a chance
1044 			 * for the MTRR work(triggered by the AP coming online)
1045 			 * to be completed in the stop machine context.
1046 			 */
1047 			schedule();
1048 		}
1049 	}
1050 
1051 	/* mark "stuck" area as not stuck */
1052 	*trampoline_status = 0;
1053 
1054 	if (get_uv_system_type() != UV_NON_UNIQUE_APIC) {
1055 		/*
1056 		 * Cleanup possible dangling ends...
1057 		 */
1058 		smpboot_restore_warm_reset_vector();
1059 	}
1060 	/*
1061 	 * Clean up the nmi handler. Do this after the callin and callout sync
1062 	 * to avoid impact of possible long unregister time.
1063 	 */
1064 	if (cpu0_nmi_registered)
1065 		unregister_nmi_handler(NMI_LOCAL, "wake_cpu0");
1066 
1067 	return boot_error;
1068 }
1069 
1070 int native_cpu_up(unsigned int cpu, struct task_struct *tidle)
1071 {
1072 	int apicid = apic->cpu_present_to_apicid(cpu);
1073 	unsigned long flags;
1074 	int err;
1075 
1076 	WARN_ON(irqs_disabled());
1077 
1078 	pr_debug("++++++++++++++++++++=_---CPU UP  %u\n", cpu);
1079 
1080 	if (apicid == BAD_APICID ||
1081 	    !physid_isset(apicid, phys_cpu_present_map) ||
1082 	    !apic->apic_id_valid(apicid)) {
1083 		pr_err("%s: bad cpu %d\n", __func__, cpu);
1084 		return -EINVAL;
1085 	}
1086 
1087 	/*
1088 	 * Already booted CPU?
1089 	 */
1090 	if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
1091 		pr_debug("do_boot_cpu %d Already started\n", cpu);
1092 		return -ENOSYS;
1093 	}
1094 
1095 	/*
1096 	 * Save current MTRR state in case it was changed since early boot
1097 	 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1098 	 */
1099 	mtrr_save_state();
1100 
1101 	/* x86 CPUs take themselves offline, so delayed offline is OK. */
1102 	err = cpu_check_up_prepare(cpu);
1103 	if (err && err != -EBUSY)
1104 		return err;
1105 
1106 	/* the FPU context is blank, nobody can own it */
1107 	__cpu_disable_lazy_restore(cpu);
1108 
1109 	common_cpu_up(cpu, tidle);
1110 
1111 	/*
1112 	 * We have to walk the irq descriptors to setup the vector
1113 	 * space for the cpu which comes online.  Prevent irq
1114 	 * alloc/free across the bringup.
1115 	 */
1116 	irq_lock_sparse();
1117 
1118 	err = do_boot_cpu(apicid, cpu, tidle);
1119 
1120 	if (err) {
1121 		irq_unlock_sparse();
1122 		pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
1123 		return -EIO;
1124 	}
1125 
1126 	/*
1127 	 * Check TSC synchronization with the AP (keep irqs disabled
1128 	 * while doing so):
1129 	 */
1130 	local_irq_save(flags);
1131 	check_tsc_sync_source(cpu);
1132 	local_irq_restore(flags);
1133 
1134 	while (!cpu_online(cpu)) {
1135 		cpu_relax();
1136 		touch_nmi_watchdog();
1137 	}
1138 
1139 	irq_unlock_sparse();
1140 
1141 	return 0;
1142 }
1143 
1144 /**
1145  * arch_disable_smp_support() - disables SMP support for x86 at runtime
1146  */
1147 void arch_disable_smp_support(void)
1148 {
1149 	disable_ioapic_support();
1150 }
1151 
1152 /*
1153  * Fall back to non SMP mode after errors.
1154  *
1155  * RED-PEN audit/test this more. I bet there is more state messed up here.
1156  */
1157 static __init void disable_smp(void)
1158 {
1159 	pr_info("SMP disabled\n");
1160 
1161 	disable_ioapic_support();
1162 
1163 	init_cpu_present(cpumask_of(0));
1164 	init_cpu_possible(cpumask_of(0));
1165 
1166 	if (smp_found_config)
1167 		physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
1168 	else
1169 		physid_set_mask_of_physid(0, &phys_cpu_present_map);
1170 	cpumask_set_cpu(0, topology_sibling_cpumask(0));
1171 	cpumask_set_cpu(0, topology_core_cpumask(0));
1172 }
1173 
1174 enum {
1175 	SMP_OK,
1176 	SMP_NO_CONFIG,
1177 	SMP_NO_APIC,
1178 	SMP_FORCE_UP,
1179 };
1180 
1181 /*
1182  * Various sanity checks.
1183  */
1184 static int __init smp_sanity_check(unsigned max_cpus)
1185 {
1186 	preempt_disable();
1187 
1188 #if !defined(CONFIG_X86_BIGSMP) && defined(CONFIG_X86_32)
1189 	if (def_to_bigsmp && nr_cpu_ids > 8) {
1190 		unsigned int cpu;
1191 		unsigned nr;
1192 
1193 		pr_warn("More than 8 CPUs detected - skipping them\n"
1194 			"Use CONFIG_X86_BIGSMP\n");
1195 
1196 		nr = 0;
1197 		for_each_present_cpu(cpu) {
1198 			if (nr >= 8)
1199 				set_cpu_present(cpu, false);
1200 			nr++;
1201 		}
1202 
1203 		nr = 0;
1204 		for_each_possible_cpu(cpu) {
1205 			if (nr >= 8)
1206 				set_cpu_possible(cpu, false);
1207 			nr++;
1208 		}
1209 
1210 		nr_cpu_ids = 8;
1211 	}
1212 #endif
1213 
1214 	if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map)) {
1215 		pr_warn("weird, boot CPU (#%d) not listed by the BIOS\n",
1216 			hard_smp_processor_id());
1217 
1218 		physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1219 	}
1220 
1221 	/*
1222 	 * If we couldn't find an SMP configuration at boot time,
1223 	 * get out of here now!
1224 	 */
1225 	if (!smp_found_config && !acpi_lapic) {
1226 		preempt_enable();
1227 		pr_notice("SMP motherboard not detected\n");
1228 		return SMP_NO_CONFIG;
1229 	}
1230 
1231 	/*
1232 	 * Should not be necessary because the MP table should list the boot
1233 	 * CPU too, but we do it for the sake of robustness anyway.
1234 	 */
1235 	if (!apic->check_phys_apicid_present(boot_cpu_physical_apicid)) {
1236 		pr_notice("weird, boot CPU (#%d) not listed by the BIOS\n",
1237 			  boot_cpu_physical_apicid);
1238 		physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1239 	}
1240 	preempt_enable();
1241 
1242 	/*
1243 	 * If we couldn't find a local APIC, then get out of here now!
1244 	 */
1245 	if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid]) &&
1246 	    !boot_cpu_has(X86_FEATURE_APIC)) {
1247 		if (!disable_apic) {
1248 			pr_err("BIOS bug, local APIC #%d not detected!...\n",
1249 				boot_cpu_physical_apicid);
1250 			pr_err("... forcing use of dummy APIC emulation (tell your hw vendor)\n");
1251 		}
1252 		return SMP_NO_APIC;
1253 	}
1254 
1255 	/*
1256 	 * If SMP should be disabled, then really disable it!
1257 	 */
1258 	if (!max_cpus) {
1259 		pr_info("SMP mode deactivated\n");
1260 		return SMP_FORCE_UP;
1261 	}
1262 
1263 	return SMP_OK;
1264 }
1265 
1266 static void __init smp_cpu_index_default(void)
1267 {
1268 	int i;
1269 	struct cpuinfo_x86 *c;
1270 
1271 	for_each_possible_cpu(i) {
1272 		c = &cpu_data(i);
1273 		/* mark all to hotplug */
1274 		c->cpu_index = nr_cpu_ids;
1275 	}
1276 }
1277 
1278 /*
1279  * Prepare for SMP bootup.  The MP table or ACPI has been read
1280  * earlier.  Just do some sanity checking here and enable APIC mode.
1281  */
1282 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1283 {
1284 	unsigned int i;
1285 
1286 	smp_cpu_index_default();
1287 
1288 	/*
1289 	 * Setup boot CPU information
1290 	 */
1291 	smp_store_boot_cpu_info(); /* Final full version of the data */
1292 	cpumask_copy(cpu_callin_mask, cpumask_of(0));
1293 	mb();
1294 
1295 	for_each_possible_cpu(i) {
1296 		zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1297 		zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1298 		zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1299 	}
1300 	set_cpu_sibling_map(0);
1301 
1302 	switch (smp_sanity_check(max_cpus)) {
1303 	case SMP_NO_CONFIG:
1304 		disable_smp();
1305 		if (APIC_init_uniprocessor())
1306 			pr_notice("Local APIC not detected. Using dummy APIC emulation.\n");
1307 		return;
1308 	case SMP_NO_APIC:
1309 		disable_smp();
1310 		return;
1311 	case SMP_FORCE_UP:
1312 		disable_smp();
1313 		apic_bsp_setup(false);
1314 		return;
1315 	case SMP_OK:
1316 		break;
1317 	}
1318 
1319 	default_setup_apic_routing();
1320 
1321 	if (read_apic_id() != boot_cpu_physical_apicid) {
1322 		panic("Boot APIC ID in local APIC unexpected (%d vs %d)",
1323 		     read_apic_id(), boot_cpu_physical_apicid);
1324 		/* Or can we switch back to PIC here? */
1325 	}
1326 
1327 	cpu0_logical_apicid = apic_bsp_setup(false);
1328 
1329 	pr_info("CPU%d: ", 0);
1330 	print_cpu_info(&cpu_data(0));
1331 
1332 	if (is_uv_system())
1333 		uv_system_init();
1334 
1335 	set_mtrr_aps_delayed_init();
1336 
1337 	smp_quirk_init_udelay();
1338 }
1339 
1340 void arch_enable_nonboot_cpus_begin(void)
1341 {
1342 	set_mtrr_aps_delayed_init();
1343 }
1344 
1345 void arch_enable_nonboot_cpus_end(void)
1346 {
1347 	mtrr_aps_init();
1348 }
1349 
1350 /*
1351  * Early setup to make printk work.
1352  */
1353 void __init native_smp_prepare_boot_cpu(void)
1354 {
1355 	int me = smp_processor_id();
1356 	switch_to_new_gdt(me);
1357 	/* already set me in cpu_online_mask in boot_cpu_init() */
1358 	cpumask_set_cpu(me, cpu_callout_mask);
1359 	cpu_set_state_online(me);
1360 }
1361 
1362 void __init native_smp_cpus_done(unsigned int max_cpus)
1363 {
1364 	pr_debug("Boot done\n");
1365 
1366 	nmi_selftest();
1367 	impress_friends();
1368 	setup_ioapic_dest();
1369 	mtrr_aps_init();
1370 }
1371 
1372 static int __initdata setup_possible_cpus = -1;
1373 static int __init _setup_possible_cpus(char *str)
1374 {
1375 	get_option(&str, &setup_possible_cpus);
1376 	return 0;
1377 }
1378 early_param("possible_cpus", _setup_possible_cpus);
1379 
1380 
1381 /*
1382  * cpu_possible_mask should be static, it cannot change as cpu's
1383  * are onlined, or offlined. The reason is per-cpu data-structures
1384  * are allocated by some modules at init time, and dont expect to
1385  * do this dynamically on cpu arrival/departure.
1386  * cpu_present_mask on the other hand can change dynamically.
1387  * In case when cpu_hotplug is not compiled, then we resort to current
1388  * behaviour, which is cpu_possible == cpu_present.
1389  * - Ashok Raj
1390  *
1391  * Three ways to find out the number of additional hotplug CPUs:
1392  * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1393  * - The user can overwrite it with possible_cpus=NUM
1394  * - Otherwise don't reserve additional CPUs.
1395  * We do this because additional CPUs waste a lot of memory.
1396  * -AK
1397  */
1398 __init void prefill_possible_map(void)
1399 {
1400 	int i, possible;
1401 
1402 	/* no processor from mptable or madt */
1403 	if (!num_processors)
1404 		num_processors = 1;
1405 
1406 	i = setup_max_cpus ?: 1;
1407 	if (setup_possible_cpus == -1) {
1408 		possible = num_processors;
1409 #ifdef CONFIG_HOTPLUG_CPU
1410 		if (setup_max_cpus)
1411 			possible += disabled_cpus;
1412 #else
1413 		if (possible > i)
1414 			possible = i;
1415 #endif
1416 	} else
1417 		possible = setup_possible_cpus;
1418 
1419 	total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1420 
1421 	/* nr_cpu_ids could be reduced via nr_cpus= */
1422 	if (possible > nr_cpu_ids) {
1423 		pr_warn("%d Processors exceeds NR_CPUS limit of %d\n",
1424 			possible, nr_cpu_ids);
1425 		possible = nr_cpu_ids;
1426 	}
1427 
1428 #ifdef CONFIG_HOTPLUG_CPU
1429 	if (!setup_max_cpus)
1430 #endif
1431 	if (possible > i) {
1432 		pr_warn("%d Processors exceeds max_cpus limit of %u\n",
1433 			possible, setup_max_cpus);
1434 		possible = i;
1435 	}
1436 
1437 	pr_info("Allowing %d CPUs, %d hotplug CPUs\n",
1438 		possible, max_t(int, possible - num_processors, 0));
1439 
1440 	for (i = 0; i < possible; i++)
1441 		set_cpu_possible(i, true);
1442 	for (; i < NR_CPUS; i++)
1443 		set_cpu_possible(i, false);
1444 
1445 	nr_cpu_ids = possible;
1446 }
1447 
1448 #ifdef CONFIG_HOTPLUG_CPU
1449 
1450 /* Recompute SMT state for all CPUs on offline */
1451 static void recompute_smt_state(void)
1452 {
1453 	int max_threads, cpu;
1454 
1455 	max_threads = 0;
1456 	for_each_online_cpu (cpu) {
1457 		int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1458 
1459 		if (threads > max_threads)
1460 			max_threads = threads;
1461 	}
1462 	__max_smt_threads = max_threads;
1463 }
1464 
1465 static void remove_siblinginfo(int cpu)
1466 {
1467 	int sibling;
1468 	struct cpuinfo_x86 *c = &cpu_data(cpu);
1469 
1470 	for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1471 		cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1472 		/*/
1473 		 * last thread sibling in this cpu core going down
1474 		 */
1475 		if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1476 			cpu_data(sibling).booted_cores--;
1477 	}
1478 
1479 	for_each_cpu(sibling, topology_sibling_cpumask(cpu))
1480 		cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1481 	for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1482 		cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1483 	cpumask_clear(cpu_llc_shared_mask(cpu));
1484 	cpumask_clear(topology_sibling_cpumask(cpu));
1485 	cpumask_clear(topology_core_cpumask(cpu));
1486 	c->phys_proc_id = 0;
1487 	c->cpu_core_id = 0;
1488 	cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1489 	recompute_smt_state();
1490 }
1491 
1492 static void remove_cpu_from_maps(int cpu)
1493 {
1494 	set_cpu_online(cpu, false);
1495 	cpumask_clear_cpu(cpu, cpu_callout_mask);
1496 	cpumask_clear_cpu(cpu, cpu_callin_mask);
1497 	/* was set by cpu_init() */
1498 	cpumask_clear_cpu(cpu, cpu_initialized_mask);
1499 	numa_remove_cpu(cpu);
1500 }
1501 
1502 void cpu_disable_common(void)
1503 {
1504 	int cpu = smp_processor_id();
1505 
1506 	remove_siblinginfo(cpu);
1507 
1508 	/* It's now safe to remove this processor from the online map */
1509 	lock_vector_lock();
1510 	remove_cpu_from_maps(cpu);
1511 	unlock_vector_lock();
1512 	fixup_irqs();
1513 }
1514 
1515 int native_cpu_disable(void)
1516 {
1517 	int ret;
1518 
1519 	ret = check_irq_vectors_for_cpu_disable();
1520 	if (ret)
1521 		return ret;
1522 
1523 	clear_local_APIC();
1524 	cpu_disable_common();
1525 
1526 	return 0;
1527 }
1528 
1529 int common_cpu_die(unsigned int cpu)
1530 {
1531 	int ret = 0;
1532 
1533 	/* We don't do anything here: idle task is faking death itself. */
1534 
1535 	/* They ack this in play_dead() by setting CPU_DEAD */
1536 	if (cpu_wait_death(cpu, 5)) {
1537 		if (system_state == SYSTEM_RUNNING)
1538 			pr_info("CPU %u is now offline\n", cpu);
1539 	} else {
1540 		pr_err("CPU %u didn't die...\n", cpu);
1541 		ret = -1;
1542 	}
1543 
1544 	return ret;
1545 }
1546 
1547 void native_cpu_die(unsigned int cpu)
1548 {
1549 	common_cpu_die(cpu);
1550 }
1551 
1552 void play_dead_common(void)
1553 {
1554 	idle_task_exit();
1555 	reset_lazy_tlbstate();
1556 	amd_e400_remove_cpu(raw_smp_processor_id());
1557 
1558 	/* Ack it */
1559 	(void)cpu_report_death();
1560 
1561 	/*
1562 	 * With physical CPU hotplug, we should halt the cpu
1563 	 */
1564 	local_irq_disable();
1565 }
1566 
1567 static bool wakeup_cpu0(void)
1568 {
1569 	if (smp_processor_id() == 0 && enable_start_cpu0)
1570 		return true;
1571 
1572 	return false;
1573 }
1574 
1575 /*
1576  * We need to flush the caches before going to sleep, lest we have
1577  * dirty data in our caches when we come back up.
1578  */
1579 static inline void mwait_play_dead(void)
1580 {
1581 	unsigned int eax, ebx, ecx, edx;
1582 	unsigned int highest_cstate = 0;
1583 	unsigned int highest_subcstate = 0;
1584 	void *mwait_ptr;
1585 	int i;
1586 
1587 	if (!this_cpu_has(X86_FEATURE_MWAIT))
1588 		return;
1589 	if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1590 		return;
1591 	if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1592 		return;
1593 
1594 	eax = CPUID_MWAIT_LEAF;
1595 	ecx = 0;
1596 	native_cpuid(&eax, &ebx, &ecx, &edx);
1597 
1598 	/*
1599 	 * eax will be 0 if EDX enumeration is not valid.
1600 	 * Initialized below to cstate, sub_cstate value when EDX is valid.
1601 	 */
1602 	if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1603 		eax = 0;
1604 	} else {
1605 		edx >>= MWAIT_SUBSTATE_SIZE;
1606 		for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1607 			if (edx & MWAIT_SUBSTATE_MASK) {
1608 				highest_cstate = i;
1609 				highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1610 			}
1611 		}
1612 		eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1613 			(highest_subcstate - 1);
1614 	}
1615 
1616 	/*
1617 	 * This should be a memory location in a cache line which is
1618 	 * unlikely to be touched by other processors.  The actual
1619 	 * content is immaterial as it is not actually modified in any way.
1620 	 */
1621 	mwait_ptr = &current_thread_info()->flags;
1622 
1623 	wbinvd();
1624 
1625 	while (1) {
1626 		/*
1627 		 * The CLFLUSH is a workaround for erratum AAI65 for
1628 		 * the Xeon 7400 series.  It's not clear it is actually
1629 		 * needed, but it should be harmless in either case.
1630 		 * The WBINVD is insufficient due to the spurious-wakeup
1631 		 * case where we return around the loop.
1632 		 */
1633 		mb();
1634 		clflush(mwait_ptr);
1635 		mb();
1636 		__monitor(mwait_ptr, 0, 0);
1637 		mb();
1638 		__mwait(eax, 0);
1639 		/*
1640 		 * If NMI wants to wake up CPU0, start CPU0.
1641 		 */
1642 		if (wakeup_cpu0())
1643 			start_cpu0();
1644 	}
1645 }
1646 
1647 void hlt_play_dead(void)
1648 {
1649 	if (__this_cpu_read(cpu_info.x86) >= 4)
1650 		wbinvd();
1651 
1652 	while (1) {
1653 		native_halt();
1654 		/*
1655 		 * If NMI wants to wake up CPU0, start CPU0.
1656 		 */
1657 		if (wakeup_cpu0())
1658 			start_cpu0();
1659 	}
1660 }
1661 
1662 void native_play_dead(void)
1663 {
1664 	play_dead_common();
1665 	tboot_shutdown(TB_SHUTDOWN_WFS);
1666 
1667 	mwait_play_dead();	/* Only returns on failure */
1668 	if (cpuidle_play_dead())
1669 		hlt_play_dead();
1670 }
1671 
1672 #else /* ... !CONFIG_HOTPLUG_CPU */
1673 int native_cpu_disable(void)
1674 {
1675 	return -ENOSYS;
1676 }
1677 
1678 void native_cpu_die(unsigned int cpu)
1679 {
1680 	/* We said "no" in __cpu_disable */
1681 	BUG();
1682 }
1683 
1684 void native_play_dead(void)
1685 {
1686 	BUG();
1687 }
1688 
1689 #endif
1690