Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * x86 SMP booting functions
4 *
5 * (c) 1995 Alan Cox, Building #3 <alan@lxorguk.ukuu.org.uk>
6 * (c) 1998, 1999, 2000, 2009 Ingo Molnar <mingo@redhat.com>
7 * Copyright 2001 Andi Kleen, SuSE Labs.
8 *
9 * Much of the core SMP work is based on previous work by Thomas Radke, to
10 * whom a great many thanks are extended.
11 *
12 * Thanks to Intel for making available several different Pentium,
13 * Pentium Pro and Pentium-II/Xeon MP machines.
14 * Original development of Linux SMP code supported by Caldera.
15 *
16 * Fixes
17 * Felix Koop : NR_CPUS used properly
18 * Jose Renau : Handle single CPU case.
19 * Alan Cox : By repeated request 8) - Total BogoMIPS report.
20 * Greg Wright : Fix for kernel stacks panic.
21 * Erich Boleyn : MP v1.4 and additional changes.
22 * Matthias Sattler : Changes for 2.1 kernel map.
23 * Michel Lespinasse : Changes for 2.1 kernel map.
24 * Michael Chastain : Change trampoline.S to gnu as.
25 * Alan Cox : Dumb bug: 'B' step PPro's are fine
26 * Ingo Molnar : Added APIC timers, based on code
27 * from Jose Renau
28 * Ingo Molnar : various cleanups and rewrites
29 * Tigran Aivazian : fixed "0.00 in /proc/uptime on SMP" bug.
30 * Maciej W. Rozycki : Bits for genuine 82489DX APICs
31 * Andi Kleen : Changed for SMP boot into long mode.
32 * Martin J. Bligh : Added support for multi-quad systems
33 * Dave Jones : Report invalid combinations of Athlon CPUs.
34 * Rusty Russell : Hacked into shape for new "hotplug" boot process.
35 * Andi Kleen : Converted to new state machine.
36 * Ashok Raj : CPU hotplug support
37 * Glauber Costa : i386 and x86_64 integration
38 */
39
40#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
41
42#include <linux/init.h>
43#include <linux/smp.h>
44#include <linux/export.h>
45#include <linux/sched.h>
46#include <linux/sched/topology.h>
47#include <linux/sched/hotplug.h>
48#include <linux/sched/task_stack.h>
49#include <linux/percpu.h>
50#include <linux/memblock.h>
51#include <linux/err.h>
52#include <linux/nmi.h>
53#include <linux/tboot.h>
54#include <linux/stackprotector.h>
55#include <linux/gfp.h>
56#include <linux/cpuidle.h>
57#include <linux/numa.h>
58
59#include <asm/acpi.h>
60#include <asm/desc.h>
61#include <asm/nmi.h>
62#include <asm/irq.h>
63#include <asm/realmode.h>
64#include <asm/cpu.h>
65#include <asm/numa.h>
66#include <asm/pgtable.h>
67#include <asm/tlbflush.h>
68#include <asm/mtrr.h>
69#include <asm/mwait.h>
70#include <asm/apic.h>
71#include <asm/io_apic.h>
72#include <asm/fpu/internal.h>
73#include <asm/setup.h>
74#include <asm/uv/uv.h>
75#include <linux/mc146818rtc.h>
76#include <asm/i8259.h>
77#include <asm/misc.h>
78#include <asm/qspinlock.h>
79#include <asm/intel-family.h>
80#include <asm/cpu_device_id.h>
81#include <asm/spec-ctrl.h>
82#include <asm/hw_irq.h>
83
84/* representing HT siblings of each logical CPU */
85DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
86EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
87
88/* representing HT and core siblings of each logical CPU */
89DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
90EXPORT_PER_CPU_SYMBOL(cpu_core_map);
91
92/* representing HT, core, and die siblings of each logical CPU */
93DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_die_map);
94EXPORT_PER_CPU_SYMBOL(cpu_die_map);
95
96DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_llc_shared_map);
97
98/* Per CPU bogomips and other parameters */
99DEFINE_PER_CPU_READ_MOSTLY(struct cpuinfo_x86, cpu_info);
100EXPORT_PER_CPU_SYMBOL(cpu_info);
101
102/* Logical package management. We might want to allocate that dynamically */
103unsigned int __max_logical_packages __read_mostly;
104EXPORT_SYMBOL(__max_logical_packages);
105static unsigned int logical_packages __read_mostly;
106static unsigned int logical_die __read_mostly;
107
108/* Maximum number of SMT threads on any online core */
109int __read_mostly __max_smt_threads = 1;
110
111/* Flag to indicate if a complete sched domain rebuild is required */
112bool x86_topology_update;
113
114int arch_update_cpu_topology(void)
115{
116 int retval = x86_topology_update;
117
118 x86_topology_update = false;
119 return retval;
120}
121
122static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
123{
124 unsigned long flags;
125
126 spin_lock_irqsave(&rtc_lock, flags);
127 CMOS_WRITE(0xa, 0xf);
128 spin_unlock_irqrestore(&rtc_lock, flags);
129 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) =
130 start_eip >> 4;
131 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) =
132 start_eip & 0xf;
133}
134
135static inline void smpboot_restore_warm_reset_vector(void)
136{
137 unsigned long flags;
138
139 /*
140 * Paranoid: Set warm reset code and vector here back
141 * to default values.
142 */
143 spin_lock_irqsave(&rtc_lock, flags);
144 CMOS_WRITE(0, 0xf);
145 spin_unlock_irqrestore(&rtc_lock, flags);
146
147 *((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
148}
149
150/*
151 * Report back to the Boot Processor during boot time or to the caller processor
152 * during CPU online.
153 */
154static void smp_callin(void)
155{
156 int cpuid;
157
158 /*
159 * If waken up by an INIT in an 82489DX configuration
160 * cpu_callout_mask guarantees we don't get here before
161 * an INIT_deassert IPI reaches our local APIC, so it is
162 * now safe to touch our local APIC.
163 */
164 cpuid = smp_processor_id();
165
166 /*
167 * the boot CPU has finished the init stage and is spinning
168 * on callin_map until we finish. We are free to set up this
169 * CPU, first the APIC. (this is probably redundant on most
170 * boards)
171 */
172 apic_ap_setup();
173
174 /*
175 * Save our processor parameters. Note: this information
176 * is needed for clock calibration.
177 */
178 smp_store_cpu_info(cpuid);
179
180 /*
181 * The topology information must be up to date before
182 * calibrate_delay() and notify_cpu_starting().
183 */
184 set_cpu_sibling_map(raw_smp_processor_id());
185
186 /*
187 * Get our bogomips.
188 * Update loops_per_jiffy in cpu_data. Previous call to
189 * smp_store_cpu_info() stored a value that is close but not as
190 * accurate as the value just calculated.
191 */
192 calibrate_delay();
193 cpu_data(cpuid).loops_per_jiffy = loops_per_jiffy;
194 pr_debug("Stack at about %p\n", &cpuid);
195
196 wmb();
197
198 notify_cpu_starting(cpuid);
199
200 /*
201 * Allow the master to continue.
202 */
203 cpumask_set_cpu(cpuid, cpu_callin_mask);
204}
205
206static int cpu0_logical_apicid;
207static int enable_start_cpu0;
208/*
209 * Activate a secondary processor.
210 */
211static void notrace start_secondary(void *unused)
212{
213 /*
214 * Don't put *anything* except direct CPU state initialization
215 * before cpu_init(), SMP booting is too fragile that we want to
216 * limit the things done here to the most necessary things.
217 */
218 cr4_init();
219
220#ifdef CONFIG_X86_32
221 /* switch away from the initial page table */
222 load_cr3(swapper_pg_dir);
223 __flush_tlb_all();
224#endif
225 load_current_idt();
226 cpu_init();
227 x86_cpuinit.early_percpu_clock_init();
228 preempt_disable();
229 smp_callin();
230
231 enable_start_cpu0 = 0;
232
233 /* otherwise gcc will move up smp_processor_id before the cpu_init */
234 barrier();
235 /*
236 * Check TSC synchronization with the boot CPU:
237 */
238 check_tsc_sync_target();
239
240 speculative_store_bypass_ht_init();
241
242 /*
243 * Lock vector_lock, set CPU online and bring the vector
244 * allocator online. Online must be set with vector_lock held
245 * to prevent a concurrent irq setup/teardown from seeing a
246 * half valid vector space.
247 */
248 lock_vector_lock();
249 set_cpu_online(smp_processor_id(), true);
250 lapic_online();
251 unlock_vector_lock();
252 cpu_set_state_online(smp_processor_id());
253 x86_platform.nmi_init();
254
255 /* enable local interrupts */
256 local_irq_enable();
257
258 /* to prevent fake stack check failure in clock setup */
259 boot_init_stack_canary();
260
261 x86_cpuinit.setup_percpu_clockev();
262
263 wmb();
264 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
265}
266
267/**
268 * topology_is_primary_thread - Check whether CPU is the primary SMT thread
269 * @cpu: CPU to check
270 */
271bool topology_is_primary_thread(unsigned int cpu)
272{
273 return apic_id_is_primary_thread(per_cpu(x86_cpu_to_apicid, cpu));
274}
275
276/**
277 * topology_smt_supported - Check whether SMT is supported by the CPUs
278 */
279bool topology_smt_supported(void)
280{
281 return smp_num_siblings > 1;
282}
283
284/**
285 * topology_phys_to_logical_pkg - Map a physical package id to a logical
286 *
287 * Returns logical package id or -1 if not found
288 */
289int topology_phys_to_logical_pkg(unsigned int phys_pkg)
290{
291 int cpu;
292
293 for_each_possible_cpu(cpu) {
294 struct cpuinfo_x86 *c = &cpu_data(cpu);
295
296 if (c->initialized && c->phys_proc_id == phys_pkg)
297 return c->logical_proc_id;
298 }
299 return -1;
300}
301EXPORT_SYMBOL(topology_phys_to_logical_pkg);
302/**
303 * topology_phys_to_logical_die - Map a physical die id to logical
304 *
305 * Returns logical die id or -1 if not found
306 */
307int topology_phys_to_logical_die(unsigned int die_id, unsigned int cur_cpu)
308{
309 int cpu;
310 int proc_id = cpu_data(cur_cpu).phys_proc_id;
311
312 for_each_possible_cpu(cpu) {
313 struct cpuinfo_x86 *c = &cpu_data(cpu);
314
315 if (c->initialized && c->cpu_die_id == die_id &&
316 c->phys_proc_id == proc_id)
317 return c->logical_die_id;
318 }
319 return -1;
320}
321EXPORT_SYMBOL(topology_phys_to_logical_die);
322
323/**
324 * topology_update_package_map - Update the physical to logical package map
325 * @pkg: The physical package id as retrieved via CPUID
326 * @cpu: The cpu for which this is updated
327 */
328int topology_update_package_map(unsigned int pkg, unsigned int cpu)
329{
330 int new;
331
332 /* Already available somewhere? */
333 new = topology_phys_to_logical_pkg(pkg);
334 if (new >= 0)
335 goto found;
336
337 new = logical_packages++;
338 if (new != pkg) {
339 pr_info("CPU %u Converting physical %u to logical package %u\n",
340 cpu, pkg, new);
341 }
342found:
343 cpu_data(cpu).logical_proc_id = new;
344 return 0;
345}
346/**
347 * topology_update_die_map - Update the physical to logical die map
348 * @die: The die id as retrieved via CPUID
349 * @cpu: The cpu for which this is updated
350 */
351int topology_update_die_map(unsigned int die, unsigned int cpu)
352{
353 int new;
354
355 /* Already available somewhere? */
356 new = topology_phys_to_logical_die(die, cpu);
357 if (new >= 0)
358 goto found;
359
360 new = logical_die++;
361 if (new != die) {
362 pr_info("CPU %u Converting physical %u to logical die %u\n",
363 cpu, die, new);
364 }
365found:
366 cpu_data(cpu).logical_die_id = new;
367 return 0;
368}
369
370void __init smp_store_boot_cpu_info(void)
371{
372 int id = 0; /* CPU 0 */
373 struct cpuinfo_x86 *c = &cpu_data(id);
374
375 *c = boot_cpu_data;
376 c->cpu_index = id;
377 topology_update_package_map(c->phys_proc_id, id);
378 topology_update_die_map(c->cpu_die_id, id);
379 c->initialized = true;
380}
381
382/*
383 * The bootstrap kernel entry code has set these up. Save them for
384 * a given CPU
385 */
386void smp_store_cpu_info(int id)
387{
388 struct cpuinfo_x86 *c = &cpu_data(id);
389
390 /* Copy boot_cpu_data only on the first bringup */
391 if (!c->initialized)
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 c->initialized = true;
400}
401
402static bool
403topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
404{
405 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
406
407 return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
408}
409
410static bool
411topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
412{
413 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
414
415 return !WARN_ONCE(!topology_same_node(c, o),
416 "sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
417 "[node: %d != %d]. Ignoring dependency.\n",
418 cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
419}
420
421#define link_mask(mfunc, c1, c2) \
422do { \
423 cpumask_set_cpu((c1), mfunc(c2)); \
424 cpumask_set_cpu((c2), mfunc(c1)); \
425} while (0)
426
427static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
428{
429 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
430 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
431
432 if (c->phys_proc_id == o->phys_proc_id &&
433 c->cpu_die_id == o->cpu_die_id &&
434 per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2)) {
435 if (c->cpu_core_id == o->cpu_core_id)
436 return topology_sane(c, o, "smt");
437
438 if ((c->cu_id != 0xff) &&
439 (o->cu_id != 0xff) &&
440 (c->cu_id == o->cu_id))
441 return topology_sane(c, o, "smt");
442 }
443
444 } else if (c->phys_proc_id == o->phys_proc_id &&
445 c->cpu_die_id == o->cpu_die_id &&
446 c->cpu_core_id == o->cpu_core_id) {
447 return topology_sane(c, o, "smt");
448 }
449
450 return false;
451}
452
453/*
454 * Define snc_cpu[] for SNC (Sub-NUMA Cluster) CPUs.
455 *
456 * These are Intel CPUs that enumerate an LLC that is shared by
457 * multiple NUMA nodes. The LLC on these systems is shared for
458 * off-package data access but private to the NUMA node (half
459 * of the package) for on-package access.
460 *
461 * CPUID (the source of the information about the LLC) can only
462 * enumerate the cache as being shared *or* unshared, but not
463 * this particular configuration. The CPU in this case enumerates
464 * the cache to be shared across the entire package (spanning both
465 * NUMA nodes).
466 */
467
468static const struct x86_cpu_id snc_cpu[] = {
469 { X86_VENDOR_INTEL, 6, INTEL_FAM6_SKYLAKE_X },
470 {}
471};
472
473static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
474{
475 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
476
477 /* Do not match if we do not have a valid APICID for cpu: */
478 if (per_cpu(cpu_llc_id, cpu1) == BAD_APICID)
479 return false;
480
481 /* Do not match if LLC id does not match: */
482 if (per_cpu(cpu_llc_id, cpu1) != per_cpu(cpu_llc_id, cpu2))
483 return false;
484
485 /*
486 * Allow the SNC topology without warning. Return of false
487 * means 'c' does not share the LLC of 'o'. This will be
488 * reflected to userspace.
489 */
490 if (!topology_same_node(c, o) && x86_match_cpu(snc_cpu))
491 return false;
492
493 return topology_sane(c, o, "llc");
494}
495
496/*
497 * Unlike the other levels, we do not enforce keeping a
498 * multicore group inside a NUMA node. If this happens, we will
499 * discard the MC level of the topology later.
500 */
501static bool match_pkg(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
502{
503 if (c->phys_proc_id == o->phys_proc_id)
504 return true;
505 return false;
506}
507
508static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
509{
510 if ((c->phys_proc_id == o->phys_proc_id) &&
511 (c->cpu_die_id == o->cpu_die_id))
512 return true;
513 return false;
514}
515
516
517#if defined(CONFIG_SCHED_SMT) || defined(CONFIG_SCHED_MC)
518static inline int x86_sched_itmt_flags(void)
519{
520 return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
521}
522
523#ifdef CONFIG_SCHED_MC
524static int x86_core_flags(void)
525{
526 return cpu_core_flags() | x86_sched_itmt_flags();
527}
528#endif
529#ifdef CONFIG_SCHED_SMT
530static int x86_smt_flags(void)
531{
532 return cpu_smt_flags() | x86_sched_itmt_flags();
533}
534#endif
535#endif
536
537static struct sched_domain_topology_level x86_numa_in_package_topology[] = {
538#ifdef CONFIG_SCHED_SMT
539 { cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
540#endif
541#ifdef CONFIG_SCHED_MC
542 { cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
543#endif
544 { NULL, },
545};
546
547static struct sched_domain_topology_level x86_topology[] = {
548#ifdef CONFIG_SCHED_SMT
549 { cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
550#endif
551#ifdef CONFIG_SCHED_MC
552 { cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
553#endif
554 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
555 { NULL, },
556};
557
558/*
559 * Set if a package/die has multiple NUMA nodes inside.
560 * AMD Magny-Cours, Intel Cluster-on-Die, and Intel
561 * Sub-NUMA Clustering have this.
562 */
563static bool x86_has_numa_in_package;
564
565void set_cpu_sibling_map(int cpu)
566{
567 bool has_smt = smp_num_siblings > 1;
568 bool has_mp = has_smt || boot_cpu_data.x86_max_cores > 1;
569 struct cpuinfo_x86 *c = &cpu_data(cpu);
570 struct cpuinfo_x86 *o;
571 int i, threads;
572
573 cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
574
575 if (!has_mp) {
576 cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
577 cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
578 cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
579 cpumask_set_cpu(cpu, topology_die_cpumask(cpu));
580 c->booted_cores = 1;
581 return;
582 }
583
584 for_each_cpu(i, cpu_sibling_setup_mask) {
585 o = &cpu_data(i);
586
587 if ((i == cpu) || (has_smt && match_smt(c, o)))
588 link_mask(topology_sibling_cpumask, cpu, i);
589
590 if ((i == cpu) || (has_mp && match_llc(c, o)))
591 link_mask(cpu_llc_shared_mask, cpu, i);
592
593 }
594
595 /*
596 * This needs a separate iteration over the cpus because we rely on all
597 * topology_sibling_cpumask links to be set-up.
598 */
599 for_each_cpu(i, cpu_sibling_setup_mask) {
600 o = &cpu_data(i);
601
602 if ((i == cpu) || (has_mp && match_pkg(c, o))) {
603 link_mask(topology_core_cpumask, cpu, i);
604
605 /*
606 * Does this new cpu bringup a new core?
607 */
608 if (cpumask_weight(
609 topology_sibling_cpumask(cpu)) == 1) {
610 /*
611 * for each core in package, increment
612 * the booted_cores for this new cpu
613 */
614 if (cpumask_first(
615 topology_sibling_cpumask(i)) == i)
616 c->booted_cores++;
617 /*
618 * increment the core count for all
619 * the other cpus in this package
620 */
621 if (i != cpu)
622 cpu_data(i).booted_cores++;
623 } else if (i != cpu && !c->booted_cores)
624 c->booted_cores = cpu_data(i).booted_cores;
625 }
626 if (match_pkg(c, o) && !topology_same_node(c, o))
627 x86_has_numa_in_package = true;
628
629 if ((i == cpu) || (has_mp && match_die(c, o)))
630 link_mask(topology_die_cpumask, cpu, i);
631 }
632
633 threads = cpumask_weight(topology_sibling_cpumask(cpu));
634 if (threads > __max_smt_threads)
635 __max_smt_threads = threads;
636}
637
638/* maps the cpu to the sched domain representing multi-core */
639const struct cpumask *cpu_coregroup_mask(int cpu)
640{
641 return cpu_llc_shared_mask(cpu);
642}
643
644static void impress_friends(void)
645{
646 int cpu;
647 unsigned long bogosum = 0;
648 /*
649 * Allow the user to impress friends.
650 */
651 pr_debug("Before bogomips\n");
652 for_each_possible_cpu(cpu)
653 if (cpumask_test_cpu(cpu, cpu_callout_mask))
654 bogosum += cpu_data(cpu).loops_per_jiffy;
655 pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
656 num_online_cpus(),
657 bogosum/(500000/HZ),
658 (bogosum/(5000/HZ))%100);
659
660 pr_debug("Before bogocount - setting activated=1\n");
661}
662
663void __inquire_remote_apic(int apicid)
664{
665 unsigned i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
666 const char * const names[] = { "ID", "VERSION", "SPIV" };
667 int timeout;
668 u32 status;
669
670 pr_info("Inquiring remote APIC 0x%x...\n", apicid);
671
672 for (i = 0; i < ARRAY_SIZE(regs); i++) {
673 pr_info("... APIC 0x%x %s: ", apicid, names[i]);
674
675 /*
676 * Wait for idle.
677 */
678 status = safe_apic_wait_icr_idle();
679 if (status)
680 pr_cont("a previous APIC delivery may have failed\n");
681
682 apic_icr_write(APIC_DM_REMRD | regs[i], apicid);
683
684 timeout = 0;
685 do {
686 udelay(100);
687 status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
688 } while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
689
690 switch (status) {
691 case APIC_ICR_RR_VALID:
692 status = apic_read(APIC_RRR);
693 pr_cont("%08x\n", status);
694 break;
695 default:
696 pr_cont("failed\n");
697 }
698 }
699}
700
701/*
702 * The Multiprocessor Specification 1.4 (1997) example code suggests
703 * that there should be a 10ms delay between the BSP asserting INIT
704 * and de-asserting INIT, when starting a remote processor.
705 * But that slows boot and resume on modern processors, which include
706 * many cores and don't require that delay.
707 *
708 * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
709 * Modern processor families are quirked to remove the delay entirely.
710 */
711#define UDELAY_10MS_DEFAULT 10000
712
713static unsigned int init_udelay = UINT_MAX;
714
715static int __init cpu_init_udelay(char *str)
716{
717 get_option(&str, &init_udelay);
718
719 return 0;
720}
721early_param("cpu_init_udelay", cpu_init_udelay);
722
723static void __init smp_quirk_init_udelay(void)
724{
725 /* if cmdline changed it from default, leave it alone */
726 if (init_udelay != UINT_MAX)
727 return;
728
729 /* if modern processor, use no delay */
730 if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
731 ((boot_cpu_data.x86_vendor == X86_VENDOR_HYGON) && (boot_cpu_data.x86 >= 0x18)) ||
732 ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
733 init_udelay = 0;
734 return;
735 }
736 /* else, use legacy delay */
737 init_udelay = UDELAY_10MS_DEFAULT;
738}
739
740/*
741 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
742 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
743 * won't ... remember to clear down the APIC, etc later.
744 */
745int
746wakeup_secondary_cpu_via_nmi(int apicid, unsigned long start_eip)
747{
748 unsigned long send_status, accept_status = 0;
749 int maxlvt;
750
751 /* Target chip */
752 /* Boot on the stack */
753 /* Kick the second */
754 apic_icr_write(APIC_DM_NMI | apic->dest_logical, apicid);
755
756 pr_debug("Waiting for send to finish...\n");
757 send_status = safe_apic_wait_icr_idle();
758
759 /*
760 * Give the other CPU some time to accept the IPI.
761 */
762 udelay(200);
763 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
764 maxlvt = lapic_get_maxlvt();
765 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
766 apic_write(APIC_ESR, 0);
767 accept_status = (apic_read(APIC_ESR) & 0xEF);
768 }
769 pr_debug("NMI sent\n");
770
771 if (send_status)
772 pr_err("APIC never delivered???\n");
773 if (accept_status)
774 pr_err("APIC delivery error (%lx)\n", accept_status);
775
776 return (send_status | accept_status);
777}
778
779static int
780wakeup_secondary_cpu_via_init(int phys_apicid, unsigned long start_eip)
781{
782 unsigned long send_status = 0, accept_status = 0;
783 int maxlvt, num_starts, j;
784
785 maxlvt = lapic_get_maxlvt();
786
787 /*
788 * Be paranoid about clearing APIC errors.
789 */
790 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
791 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
792 apic_write(APIC_ESR, 0);
793 apic_read(APIC_ESR);
794 }
795
796 pr_debug("Asserting INIT\n");
797
798 /*
799 * Turn INIT on target chip
800 */
801 /*
802 * Send IPI
803 */
804 apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT,
805 phys_apicid);
806
807 pr_debug("Waiting for send to finish...\n");
808 send_status = safe_apic_wait_icr_idle();
809
810 udelay(init_udelay);
811
812 pr_debug("Deasserting INIT\n");
813
814 /* Target chip */
815 /* Send IPI */
816 apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
817
818 pr_debug("Waiting for send to finish...\n");
819 send_status = safe_apic_wait_icr_idle();
820
821 mb();
822
823 /*
824 * Should we send STARTUP IPIs ?
825 *
826 * Determine this based on the APIC version.
827 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
828 */
829 if (APIC_INTEGRATED(boot_cpu_apic_version))
830 num_starts = 2;
831 else
832 num_starts = 0;
833
834 /*
835 * Run STARTUP IPI loop.
836 */
837 pr_debug("#startup loops: %d\n", num_starts);
838
839 for (j = 1; j <= num_starts; j++) {
840 pr_debug("Sending STARTUP #%d\n", j);
841 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
842 apic_write(APIC_ESR, 0);
843 apic_read(APIC_ESR);
844 pr_debug("After apic_write\n");
845
846 /*
847 * STARTUP IPI
848 */
849
850 /* Target chip */
851 /* Boot on the stack */
852 /* Kick the second */
853 apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
854 phys_apicid);
855
856 /*
857 * Give the other CPU some time to accept the IPI.
858 */
859 if (init_udelay == 0)
860 udelay(10);
861 else
862 udelay(300);
863
864 pr_debug("Startup point 1\n");
865
866 pr_debug("Waiting for send to finish...\n");
867 send_status = safe_apic_wait_icr_idle();
868
869 /*
870 * Give the other CPU some time to accept the IPI.
871 */
872 if (init_udelay == 0)
873 udelay(10);
874 else
875 udelay(200);
876
877 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
878 apic_write(APIC_ESR, 0);
879 accept_status = (apic_read(APIC_ESR) & 0xEF);
880 if (send_status || accept_status)
881 break;
882 }
883 pr_debug("After Startup\n");
884
885 if (send_status)
886 pr_err("APIC never delivered???\n");
887 if (accept_status)
888 pr_err("APIC delivery error (%lx)\n", accept_status);
889
890 return (send_status | accept_status);
891}
892
893/* reduce the number of lines printed when booting a large cpu count system */
894static void announce_cpu(int cpu, int apicid)
895{
896 static int current_node = NUMA_NO_NODE;
897 int node = early_cpu_to_node(cpu);
898 static int width, node_width;
899
900 if (!width)
901 width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
902
903 if (!node_width)
904 node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
905
906 if (cpu == 1)
907 printk(KERN_INFO "x86: Booting SMP configuration:\n");
908
909 if (system_state < SYSTEM_RUNNING) {
910 if (node != current_node) {
911 if (current_node > (-1))
912 pr_cont("\n");
913 current_node = node;
914
915 printk(KERN_INFO ".... node %*s#%d, CPUs: ",
916 node_width - num_digits(node), " ", node);
917 }
918
919 /* Add padding for the BSP */
920 if (cpu == 1)
921 pr_cont("%*s", width + 1, " ");
922
923 pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
924
925 } else
926 pr_info("Booting Node %d Processor %d APIC 0x%x\n",
927 node, cpu, apicid);
928}
929
930static int wakeup_cpu0_nmi(unsigned int cmd, struct pt_regs *regs)
931{
932 int cpu;
933
934 cpu = smp_processor_id();
935 if (cpu == 0 && !cpu_online(cpu) && enable_start_cpu0)
936 return NMI_HANDLED;
937
938 return NMI_DONE;
939}
940
941/*
942 * Wake up AP by INIT, INIT, STARTUP sequence.
943 *
944 * Instead of waiting for STARTUP after INITs, BSP will execute the BIOS
945 * boot-strap code which is not a desired behavior for waking up BSP. To
946 * void the boot-strap code, wake up CPU0 by NMI instead.
947 *
948 * This works to wake up soft offlined CPU0 only. If CPU0 is hard offlined
949 * (i.e. physically hot removed and then hot added), NMI won't wake it up.
950 * We'll change this code in the future to wake up hard offlined CPU0 if
951 * real platform and request are available.
952 */
953static int
954wakeup_cpu_via_init_nmi(int cpu, unsigned long start_ip, int apicid,
955 int *cpu0_nmi_registered)
956{
957 int id;
958 int boot_error;
959
960 preempt_disable();
961
962 /*
963 * Wake up AP by INIT, INIT, STARTUP sequence.
964 */
965 if (cpu) {
966 boot_error = wakeup_secondary_cpu_via_init(apicid, start_ip);
967 goto out;
968 }
969
970 /*
971 * Wake up BSP by nmi.
972 *
973 * Register a NMI handler to help wake up CPU0.
974 */
975 boot_error = register_nmi_handler(NMI_LOCAL,
976 wakeup_cpu0_nmi, 0, "wake_cpu0");
977
978 if (!boot_error) {
979 enable_start_cpu0 = 1;
980 *cpu0_nmi_registered = 1;
981 if (apic->dest_logical == APIC_DEST_LOGICAL)
982 id = cpu0_logical_apicid;
983 else
984 id = apicid;
985 boot_error = wakeup_secondary_cpu_via_nmi(id, start_ip);
986 }
987
988out:
989 preempt_enable();
990
991 return boot_error;
992}
993
994int common_cpu_up(unsigned int cpu, struct task_struct *idle)
995{
996 int ret;
997
998 /* Just in case we booted with a single CPU. */
999 alternatives_enable_smp();
1000
1001 per_cpu(current_task, cpu) = idle;
1002
1003 /* Initialize the interrupt stack(s) */
1004 ret = irq_init_percpu_irqstack(cpu);
1005 if (ret)
1006 return ret;
1007
1008#ifdef CONFIG_X86_32
1009 /* Stack for startup_32 can be just as for start_secondary onwards */
1010 per_cpu(cpu_current_top_of_stack, cpu) = task_top_of_stack(idle);
1011#else
1012 initial_gs = per_cpu_offset(cpu);
1013#endif
1014 return 0;
1015}
1016
1017/*
1018 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
1019 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
1020 * Returns zero if CPU booted OK, else error code from
1021 * ->wakeup_secondary_cpu.
1022 */
1023static int do_boot_cpu(int apicid, int cpu, struct task_struct *idle,
1024 int *cpu0_nmi_registered)
1025{
1026 /* start_ip had better be page-aligned! */
1027 unsigned long start_ip = real_mode_header->trampoline_start;
1028
1029 unsigned long boot_error = 0;
1030 unsigned long timeout;
1031
1032 idle->thread.sp = (unsigned long)task_pt_regs(idle);
1033 early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
1034 initial_code = (unsigned long)start_secondary;
1035 initial_stack = idle->thread.sp;
1036
1037 /* Enable the espfix hack for this CPU */
1038 init_espfix_ap(cpu);
1039
1040 /* So we see what's up */
1041 announce_cpu(cpu, apicid);
1042
1043 /*
1044 * This grunge runs the startup process for
1045 * the targeted processor.
1046 */
1047
1048 if (x86_platform.legacy.warm_reset) {
1049
1050 pr_debug("Setting warm reset code and vector.\n");
1051
1052 smpboot_setup_warm_reset_vector(start_ip);
1053 /*
1054 * Be paranoid about clearing APIC errors.
1055 */
1056 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
1057 apic_write(APIC_ESR, 0);
1058 apic_read(APIC_ESR);
1059 }
1060 }
1061
1062 /*
1063 * AP might wait on cpu_callout_mask in cpu_init() with
1064 * cpu_initialized_mask set if previous attempt to online
1065 * it timed-out. Clear cpu_initialized_mask so that after
1066 * INIT/SIPI it could start with a clean state.
1067 */
1068 cpumask_clear_cpu(cpu, cpu_initialized_mask);
1069 smp_mb();
1070
1071 /*
1072 * Wake up a CPU in difference cases:
1073 * - Use the method in the APIC driver if it's defined
1074 * Otherwise,
1075 * - Use an INIT boot APIC message for APs or NMI for BSP.
1076 */
1077 if (apic->wakeup_secondary_cpu)
1078 boot_error = apic->wakeup_secondary_cpu(apicid, start_ip);
1079 else
1080 boot_error = wakeup_cpu_via_init_nmi(cpu, start_ip, apicid,
1081 cpu0_nmi_registered);
1082
1083 if (!boot_error) {
1084 /*
1085 * Wait 10s total for first sign of life from AP
1086 */
1087 boot_error = -1;
1088 timeout = jiffies + 10*HZ;
1089 while (time_before(jiffies, timeout)) {
1090 if (cpumask_test_cpu(cpu, cpu_initialized_mask)) {
1091 /*
1092 * Tell AP to proceed with initialization
1093 */
1094 cpumask_set_cpu(cpu, cpu_callout_mask);
1095 boot_error = 0;
1096 break;
1097 }
1098 schedule();
1099 }
1100 }
1101
1102 if (!boot_error) {
1103 /*
1104 * Wait till AP completes initial initialization
1105 */
1106 while (!cpumask_test_cpu(cpu, cpu_callin_mask)) {
1107 /*
1108 * Allow other tasks to run while we wait for the
1109 * AP to come online. This also gives a chance
1110 * for the MTRR work(triggered by the AP coming online)
1111 * to be completed in the stop machine context.
1112 */
1113 schedule();
1114 }
1115 }
1116
1117 if (x86_platform.legacy.warm_reset) {
1118 /*
1119 * Cleanup possible dangling ends...
1120 */
1121 smpboot_restore_warm_reset_vector();
1122 }
1123
1124 return boot_error;
1125}
1126
1127int native_cpu_up(unsigned int cpu, struct task_struct *tidle)
1128{
1129 int apicid = apic->cpu_present_to_apicid(cpu);
1130 int cpu0_nmi_registered = 0;
1131 unsigned long flags;
1132 int err, ret = 0;
1133
1134 lockdep_assert_irqs_enabled();
1135
1136 pr_debug("++++++++++++++++++++=_---CPU UP %u\n", cpu);
1137
1138 if (apicid == BAD_APICID ||
1139 !physid_isset(apicid, phys_cpu_present_map) ||
1140 !apic->apic_id_valid(apicid)) {
1141 pr_err("%s: bad cpu %d\n", __func__, cpu);
1142 return -EINVAL;
1143 }
1144
1145 /*
1146 * Already booted CPU?
1147 */
1148 if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
1149 pr_debug("do_boot_cpu %d Already started\n", cpu);
1150 return -ENOSYS;
1151 }
1152
1153 /*
1154 * Save current MTRR state in case it was changed since early boot
1155 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1156 */
1157 mtrr_save_state();
1158
1159 /* x86 CPUs take themselves offline, so delayed offline is OK. */
1160 err = cpu_check_up_prepare(cpu);
1161 if (err && err != -EBUSY)
1162 return err;
1163
1164 /* the FPU context is blank, nobody can own it */
1165 per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
1166
1167 err = common_cpu_up(cpu, tidle);
1168 if (err)
1169 return err;
1170
1171 err = do_boot_cpu(apicid, cpu, tidle, &cpu0_nmi_registered);
1172 if (err) {
1173 pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
1174 ret = -EIO;
1175 goto unreg_nmi;
1176 }
1177
1178 /*
1179 * Check TSC synchronization with the AP (keep irqs disabled
1180 * while doing so):
1181 */
1182 local_irq_save(flags);
1183 check_tsc_sync_source(cpu);
1184 local_irq_restore(flags);
1185
1186 while (!cpu_online(cpu)) {
1187 cpu_relax();
1188 touch_nmi_watchdog();
1189 }
1190
1191unreg_nmi:
1192 /*
1193 * Clean up the nmi handler. Do this after the callin and callout sync
1194 * to avoid impact of possible long unregister time.
1195 */
1196 if (cpu0_nmi_registered)
1197 unregister_nmi_handler(NMI_LOCAL, "wake_cpu0");
1198
1199 return ret;
1200}
1201
1202/**
1203 * arch_disable_smp_support() - disables SMP support for x86 at runtime
1204 */
1205void arch_disable_smp_support(void)
1206{
1207 disable_ioapic_support();
1208}
1209
1210/*
1211 * Fall back to non SMP mode after errors.
1212 *
1213 * RED-PEN audit/test this more. I bet there is more state messed up here.
1214 */
1215static __init void disable_smp(void)
1216{
1217 pr_info("SMP disabled\n");
1218
1219 disable_ioapic_support();
1220
1221 init_cpu_present(cpumask_of(0));
1222 init_cpu_possible(cpumask_of(0));
1223
1224 if (smp_found_config)
1225 physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
1226 else
1227 physid_set_mask_of_physid(0, &phys_cpu_present_map);
1228 cpumask_set_cpu(0, topology_sibling_cpumask(0));
1229 cpumask_set_cpu(0, topology_core_cpumask(0));
1230 cpumask_set_cpu(0, topology_die_cpumask(0));
1231}
1232
1233/*
1234 * Various sanity checks.
1235 */
1236static void __init smp_sanity_check(void)
1237{
1238 preempt_disable();
1239
1240#if !defined(CONFIG_X86_BIGSMP) && defined(CONFIG_X86_32)
1241 if (def_to_bigsmp && nr_cpu_ids > 8) {
1242 unsigned int cpu;
1243 unsigned nr;
1244
1245 pr_warn("More than 8 CPUs detected - skipping them\n"
1246 "Use CONFIG_X86_BIGSMP\n");
1247
1248 nr = 0;
1249 for_each_present_cpu(cpu) {
1250 if (nr >= 8)
1251 set_cpu_present(cpu, false);
1252 nr++;
1253 }
1254
1255 nr = 0;
1256 for_each_possible_cpu(cpu) {
1257 if (nr >= 8)
1258 set_cpu_possible(cpu, false);
1259 nr++;
1260 }
1261
1262 nr_cpu_ids = 8;
1263 }
1264#endif
1265
1266 if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map)) {
1267 pr_warn("weird, boot CPU (#%d) not listed by the BIOS\n",
1268 hard_smp_processor_id());
1269
1270 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1271 }
1272
1273 /*
1274 * Should not be necessary because the MP table should list the boot
1275 * CPU too, but we do it for the sake of robustness anyway.
1276 */
1277 if (!apic->check_phys_apicid_present(boot_cpu_physical_apicid)) {
1278 pr_notice("weird, boot CPU (#%d) not listed by the BIOS\n",
1279 boot_cpu_physical_apicid);
1280 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1281 }
1282 preempt_enable();
1283}
1284
1285static void __init smp_cpu_index_default(void)
1286{
1287 int i;
1288 struct cpuinfo_x86 *c;
1289
1290 for_each_possible_cpu(i) {
1291 c = &cpu_data(i);
1292 /* mark all to hotplug */
1293 c->cpu_index = nr_cpu_ids;
1294 }
1295}
1296
1297static void __init smp_get_logical_apicid(void)
1298{
1299 if (x2apic_mode)
1300 cpu0_logical_apicid = apic_read(APIC_LDR);
1301 else
1302 cpu0_logical_apicid = GET_APIC_LOGICAL_ID(apic_read(APIC_LDR));
1303}
1304
1305/*
1306 * Prepare for SMP bootup.
1307 * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1308 * for common interface support.
1309 */
1310void __init native_smp_prepare_cpus(unsigned int max_cpus)
1311{
1312 unsigned int i;
1313
1314 smp_cpu_index_default();
1315
1316 /*
1317 * Setup boot CPU information
1318 */
1319 smp_store_boot_cpu_info(); /* Final full version of the data */
1320 cpumask_copy(cpu_callin_mask, cpumask_of(0));
1321 mb();
1322
1323 for_each_possible_cpu(i) {
1324 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1325 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1326 zalloc_cpumask_var(&per_cpu(cpu_die_map, i), GFP_KERNEL);
1327 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1328 }
1329
1330 /*
1331 * Set 'default' x86 topology, this matches default_topology() in that
1332 * it has NUMA nodes as a topology level. See also
1333 * native_smp_cpus_done().
1334 *
1335 * Must be done before set_cpus_sibling_map() is ran.
1336 */
1337 set_sched_topology(x86_topology);
1338
1339 set_cpu_sibling_map(0);
1340
1341 smp_sanity_check();
1342
1343 switch (apic_intr_mode) {
1344 case APIC_PIC:
1345 case APIC_VIRTUAL_WIRE_NO_CONFIG:
1346 disable_smp();
1347 return;
1348 case APIC_SYMMETRIC_IO_NO_ROUTING:
1349 disable_smp();
1350 /* Setup local timer */
1351 x86_init.timers.setup_percpu_clockev();
1352 return;
1353 case APIC_VIRTUAL_WIRE:
1354 case APIC_SYMMETRIC_IO:
1355 break;
1356 }
1357
1358 /* Setup local timer */
1359 x86_init.timers.setup_percpu_clockev();
1360
1361 smp_get_logical_apicid();
1362
1363 pr_info("CPU0: ");
1364 print_cpu_info(&cpu_data(0));
1365
1366 uv_system_init();
1367
1368 set_mtrr_aps_delayed_init();
1369
1370 smp_quirk_init_udelay();
1371
1372 speculative_store_bypass_ht_init();
1373}
1374
1375void arch_enable_nonboot_cpus_begin(void)
1376{
1377 set_mtrr_aps_delayed_init();
1378}
1379
1380void arch_enable_nonboot_cpus_end(void)
1381{
1382 mtrr_aps_init();
1383}
1384
1385/*
1386 * Early setup to make printk work.
1387 */
1388void __init native_smp_prepare_boot_cpu(void)
1389{
1390 int me = smp_processor_id();
1391 switch_to_new_gdt(me);
1392 /* already set me in cpu_online_mask in boot_cpu_init() */
1393 cpumask_set_cpu(me, cpu_callout_mask);
1394 cpu_set_state_online(me);
1395 native_pv_lock_init();
1396}
1397
1398void __init calculate_max_logical_packages(void)
1399{
1400 int ncpus;
1401
1402 /*
1403 * Today neither Intel nor AMD support heterogenous systems so
1404 * extrapolate the boot cpu's data to all packages.
1405 */
1406 ncpus = cpu_data(0).booted_cores * topology_max_smt_threads();
1407 __max_logical_packages = DIV_ROUND_UP(total_cpus, ncpus);
1408 pr_info("Max logical packages: %u\n", __max_logical_packages);
1409}
1410
1411void __init native_smp_cpus_done(unsigned int max_cpus)
1412{
1413 pr_debug("Boot done\n");
1414
1415 calculate_max_logical_packages();
1416
1417 if (x86_has_numa_in_package)
1418 set_sched_topology(x86_numa_in_package_topology);
1419
1420 nmi_selftest();
1421 impress_friends();
1422 mtrr_aps_init();
1423}
1424
1425static int __initdata setup_possible_cpus = -1;
1426static int __init _setup_possible_cpus(char *str)
1427{
1428 get_option(&str, &setup_possible_cpus);
1429 return 0;
1430}
1431early_param("possible_cpus", _setup_possible_cpus);
1432
1433
1434/*
1435 * cpu_possible_mask should be static, it cannot change as cpu's
1436 * are onlined, or offlined. The reason is per-cpu data-structures
1437 * are allocated by some modules at init time, and dont expect to
1438 * do this dynamically on cpu arrival/departure.
1439 * cpu_present_mask on the other hand can change dynamically.
1440 * In case when cpu_hotplug is not compiled, then we resort to current
1441 * behaviour, which is cpu_possible == cpu_present.
1442 * - Ashok Raj
1443 *
1444 * Three ways to find out the number of additional hotplug CPUs:
1445 * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1446 * - The user can overwrite it with possible_cpus=NUM
1447 * - Otherwise don't reserve additional CPUs.
1448 * We do this because additional CPUs waste a lot of memory.
1449 * -AK
1450 */
1451__init void prefill_possible_map(void)
1452{
1453 int i, possible;
1454
1455 /* No boot processor was found in mptable or ACPI MADT */
1456 if (!num_processors) {
1457 if (boot_cpu_has(X86_FEATURE_APIC)) {
1458 int apicid = boot_cpu_physical_apicid;
1459 int cpu = hard_smp_processor_id();
1460
1461 pr_warn("Boot CPU (id %d) not listed by BIOS\n", cpu);
1462
1463 /* Make sure boot cpu is enumerated */
1464 if (apic->cpu_present_to_apicid(0) == BAD_APICID &&
1465 apic->apic_id_valid(apicid))
1466 generic_processor_info(apicid, boot_cpu_apic_version);
1467 }
1468
1469 if (!num_processors)
1470 num_processors = 1;
1471 }
1472
1473 i = setup_max_cpus ?: 1;
1474 if (setup_possible_cpus == -1) {
1475 possible = num_processors;
1476#ifdef CONFIG_HOTPLUG_CPU
1477 if (setup_max_cpus)
1478 possible += disabled_cpus;
1479#else
1480 if (possible > i)
1481 possible = i;
1482#endif
1483 } else
1484 possible = setup_possible_cpus;
1485
1486 total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1487
1488 /* nr_cpu_ids could be reduced via nr_cpus= */
1489 if (possible > nr_cpu_ids) {
1490 pr_warn("%d Processors exceeds NR_CPUS limit of %u\n",
1491 possible, nr_cpu_ids);
1492 possible = nr_cpu_ids;
1493 }
1494
1495#ifdef CONFIG_HOTPLUG_CPU
1496 if (!setup_max_cpus)
1497#endif
1498 if (possible > i) {
1499 pr_warn("%d Processors exceeds max_cpus limit of %u\n",
1500 possible, setup_max_cpus);
1501 possible = i;
1502 }
1503
1504 nr_cpu_ids = possible;
1505
1506 pr_info("Allowing %d CPUs, %d hotplug CPUs\n",
1507 possible, max_t(int, possible - num_processors, 0));
1508
1509 reset_cpu_possible_mask();
1510
1511 for (i = 0; i < possible; i++)
1512 set_cpu_possible(i, true);
1513}
1514
1515#ifdef CONFIG_HOTPLUG_CPU
1516
1517/* Recompute SMT state for all CPUs on offline */
1518static void recompute_smt_state(void)
1519{
1520 int max_threads, cpu;
1521
1522 max_threads = 0;
1523 for_each_online_cpu (cpu) {
1524 int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1525
1526 if (threads > max_threads)
1527 max_threads = threads;
1528 }
1529 __max_smt_threads = max_threads;
1530}
1531
1532static void remove_siblinginfo(int cpu)
1533{
1534 int sibling;
1535 struct cpuinfo_x86 *c = &cpu_data(cpu);
1536
1537 for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1538 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1539 /*/
1540 * last thread sibling in this cpu core going down
1541 */
1542 if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1543 cpu_data(sibling).booted_cores--;
1544 }
1545
1546 for_each_cpu(sibling, topology_die_cpumask(cpu))
1547 cpumask_clear_cpu(cpu, topology_die_cpumask(sibling));
1548 for_each_cpu(sibling, topology_sibling_cpumask(cpu))
1549 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1550 for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1551 cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1552 cpumask_clear(cpu_llc_shared_mask(cpu));
1553 cpumask_clear(topology_sibling_cpumask(cpu));
1554 cpumask_clear(topology_core_cpumask(cpu));
1555 cpumask_clear(topology_die_cpumask(cpu));
1556 c->cpu_core_id = 0;
1557 c->booted_cores = 0;
1558 cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1559 recompute_smt_state();
1560}
1561
1562static void remove_cpu_from_maps(int cpu)
1563{
1564 set_cpu_online(cpu, false);
1565 cpumask_clear_cpu(cpu, cpu_callout_mask);
1566 cpumask_clear_cpu(cpu, cpu_callin_mask);
1567 /* was set by cpu_init() */
1568 cpumask_clear_cpu(cpu, cpu_initialized_mask);
1569 numa_remove_cpu(cpu);
1570}
1571
1572void cpu_disable_common(void)
1573{
1574 int cpu = smp_processor_id();
1575
1576 remove_siblinginfo(cpu);
1577
1578 /* It's now safe to remove this processor from the online map */
1579 lock_vector_lock();
1580 remove_cpu_from_maps(cpu);
1581 unlock_vector_lock();
1582 fixup_irqs();
1583 lapic_offline();
1584}
1585
1586int native_cpu_disable(void)
1587{
1588 int ret;
1589
1590 ret = lapic_can_unplug_cpu();
1591 if (ret)
1592 return ret;
1593
1594 /*
1595 * Disable the local APIC. Otherwise IPI broadcasts will reach
1596 * it. It still responds normally to INIT, NMI, SMI, and SIPI
1597 * messages.
1598 */
1599 apic_soft_disable();
1600 cpu_disable_common();
1601
1602 return 0;
1603}
1604
1605int common_cpu_die(unsigned int cpu)
1606{
1607 int ret = 0;
1608
1609 /* We don't do anything here: idle task is faking death itself. */
1610
1611 /* They ack this in play_dead() by setting CPU_DEAD */
1612 if (cpu_wait_death(cpu, 5)) {
1613 if (system_state == SYSTEM_RUNNING)
1614 pr_info("CPU %u is now offline\n", cpu);
1615 } else {
1616 pr_err("CPU %u didn't die...\n", cpu);
1617 ret = -1;
1618 }
1619
1620 return ret;
1621}
1622
1623void native_cpu_die(unsigned int cpu)
1624{
1625 common_cpu_die(cpu);
1626}
1627
1628void play_dead_common(void)
1629{
1630 idle_task_exit();
1631
1632 /* Ack it */
1633 (void)cpu_report_death();
1634
1635 /*
1636 * With physical CPU hotplug, we should halt the cpu
1637 */
1638 local_irq_disable();
1639}
1640
1641static bool wakeup_cpu0(void)
1642{
1643 if (smp_processor_id() == 0 && enable_start_cpu0)
1644 return true;
1645
1646 return false;
1647}
1648
1649/*
1650 * We need to flush the caches before going to sleep, lest we have
1651 * dirty data in our caches when we come back up.
1652 */
1653static inline void mwait_play_dead(void)
1654{
1655 unsigned int eax, ebx, ecx, edx;
1656 unsigned int highest_cstate = 0;
1657 unsigned int highest_subcstate = 0;
1658 void *mwait_ptr;
1659 int i;
1660
1661 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD ||
1662 boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
1663 return;
1664 if (!this_cpu_has(X86_FEATURE_MWAIT))
1665 return;
1666 if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1667 return;
1668 if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1669 return;
1670
1671 eax = CPUID_MWAIT_LEAF;
1672 ecx = 0;
1673 native_cpuid(&eax, &ebx, &ecx, &edx);
1674
1675 /*
1676 * eax will be 0 if EDX enumeration is not valid.
1677 * Initialized below to cstate, sub_cstate value when EDX is valid.
1678 */
1679 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1680 eax = 0;
1681 } else {
1682 edx >>= MWAIT_SUBSTATE_SIZE;
1683 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1684 if (edx & MWAIT_SUBSTATE_MASK) {
1685 highest_cstate = i;
1686 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1687 }
1688 }
1689 eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1690 (highest_subcstate - 1);
1691 }
1692
1693 /*
1694 * This should be a memory location in a cache line which is
1695 * unlikely to be touched by other processors. The actual
1696 * content is immaterial as it is not actually modified in any way.
1697 */
1698 mwait_ptr = ¤t_thread_info()->flags;
1699
1700 wbinvd();
1701
1702 while (1) {
1703 /*
1704 * The CLFLUSH is a workaround for erratum AAI65 for
1705 * the Xeon 7400 series. It's not clear it is actually
1706 * needed, but it should be harmless in either case.
1707 * The WBINVD is insufficient due to the spurious-wakeup
1708 * case where we return around the loop.
1709 */
1710 mb();
1711 clflush(mwait_ptr);
1712 mb();
1713 __monitor(mwait_ptr, 0, 0);
1714 mb();
1715 __mwait(eax, 0);
1716 /*
1717 * If NMI wants to wake up CPU0, start CPU0.
1718 */
1719 if (wakeup_cpu0())
1720 start_cpu0();
1721 }
1722}
1723
1724void hlt_play_dead(void)
1725{
1726 if (__this_cpu_read(cpu_info.x86) >= 4)
1727 wbinvd();
1728
1729 while (1) {
1730 native_halt();
1731 /*
1732 * If NMI wants to wake up CPU0, start CPU0.
1733 */
1734 if (wakeup_cpu0())
1735 start_cpu0();
1736 }
1737}
1738
1739void native_play_dead(void)
1740{
1741 play_dead_common();
1742 tboot_shutdown(TB_SHUTDOWN_WFS);
1743
1744 mwait_play_dead(); /* Only returns on failure */
1745 if (cpuidle_play_dead())
1746 hlt_play_dead();
1747}
1748
1749#else /* ... !CONFIG_HOTPLUG_CPU */
1750int native_cpu_disable(void)
1751{
1752 return -ENOSYS;
1753}
1754
1755void native_cpu_die(unsigned int cpu)
1756{
1757 /* We said "no" in __cpu_disable */
1758 BUG();
1759}
1760
1761void native_play_dead(void)
1762{
1763 BUG();
1764}
1765
1766#endif
1 /*
2 * x86 SMP booting functions
3 *
4 * (c) 1995 Alan Cox, Building #3 <alan@lxorguk.ukuu.org.uk>
5 * (c) 1998, 1999, 2000, 2009 Ingo Molnar <mingo@redhat.com>
6 * Copyright 2001 Andi Kleen, SuSE Labs.
7 *
8 * Much of the core SMP work is based on previous work by Thomas Radke, to
9 * whom a great many thanks are extended.
10 *
11 * Thanks to Intel for making available several different Pentium,
12 * Pentium Pro and Pentium-II/Xeon MP machines.
13 * Original development of Linux SMP code supported by Caldera.
14 *
15 * This code is released under the GNU General Public License version 2 or
16 * later.
17 *
18 * Fixes
19 * Felix Koop : NR_CPUS used properly
20 * Jose Renau : Handle single CPU case.
21 * Alan Cox : By repeated request 8) - Total BogoMIPS report.
22 * Greg Wright : Fix for kernel stacks panic.
23 * Erich Boleyn : MP v1.4 and additional changes.
24 * Matthias Sattler : Changes for 2.1 kernel map.
25 * Michel Lespinasse : Changes for 2.1 kernel map.
26 * Michael Chastain : Change trampoline.S to gnu as.
27 * Alan Cox : Dumb bug: 'B' step PPro's are fine
28 * Ingo Molnar : Added APIC timers, based on code
29 * from Jose Renau
30 * Ingo Molnar : various cleanups and rewrites
31 * Tigran Aivazian : fixed "0.00 in /proc/uptime on SMP" bug.
32 * Maciej W. Rozycki : Bits for genuine 82489DX APICs
33 * Andi Kleen : Changed for SMP boot into long mode.
34 * Martin J. Bligh : Added support for multi-quad systems
35 * Dave Jones : Report invalid combinations of Athlon CPUs.
36 * Rusty Russell : Hacked into shape for new "hotplug" boot process.
37 * Andi Kleen : Converted to new state machine.
38 * Ashok Raj : CPU hotplug support
39 * Glauber Costa : i386 and x86_64 integration
40 */
41
42#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
43
44#include <linux/init.h>
45#include <linux/smp.h>
46#include <linux/export.h>
47#include <linux/sched.h>
48#include <linux/percpu.h>
49#include <linux/bootmem.h>
50#include <linux/err.h>
51#include <linux/nmi.h>
52#include <linux/tboot.h>
53#include <linux/stackprotector.h>
54#include <linux/gfp.h>
55#include <linux/cpuidle.h>
56
57#include <asm/acpi.h>
58#include <asm/desc.h>
59#include <asm/nmi.h>
60#include <asm/irq.h>
61#include <asm/realmode.h>
62#include <asm/cpu.h>
63#include <asm/numa.h>
64#include <asm/pgtable.h>
65#include <asm/tlbflush.h>
66#include <asm/mtrr.h>
67#include <asm/mwait.h>
68#include <asm/apic.h>
69#include <asm/io_apic.h>
70#include <asm/fpu/internal.h>
71#include <asm/setup.h>
72#include <asm/uv/uv.h>
73#include <linux/mc146818rtc.h>
74#include <asm/i8259.h>
75#include <asm/realmode.h>
76#include <asm/misc.h>
77
78/* Number of siblings per CPU package */
79int smp_num_siblings = 1;
80EXPORT_SYMBOL(smp_num_siblings);
81
82/* Last level cache ID of each logical CPU */
83DEFINE_PER_CPU_READ_MOSTLY(u16, cpu_llc_id) = BAD_APICID;
84
85/* representing HT siblings of each logical CPU */
86DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
87EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
88
89/* representing HT and core siblings of each logical CPU */
90DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
91EXPORT_PER_CPU_SYMBOL(cpu_core_map);
92
93DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_llc_shared_map);
94
95/* Per CPU bogomips and other parameters */
96DEFINE_PER_CPU_READ_MOSTLY(struct cpuinfo_x86, cpu_info);
97EXPORT_PER_CPU_SYMBOL(cpu_info);
98
99/* Logical package management. We might want to allocate that dynamically */
100static int *physical_to_logical_pkg __read_mostly;
101static unsigned long *physical_package_map __read_mostly;;
102static unsigned int max_physical_pkg_id __read_mostly;
103unsigned int __max_logical_packages __read_mostly;
104EXPORT_SYMBOL(__max_logical_packages);
105static unsigned int logical_packages __read_mostly;
106
107/* Maximum number of SMT threads on any online core */
108int __max_smt_threads __read_mostly;
109
110/* Flag to indicate if a complete sched domain rebuild is required */
111bool x86_topology_update;
112
113int arch_update_cpu_topology(void)
114{
115 int retval = x86_topology_update;
116
117 x86_topology_update = false;
118 return retval;
119}
120
121static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
122{
123 unsigned long flags;
124
125 spin_lock_irqsave(&rtc_lock, flags);
126 CMOS_WRITE(0xa, 0xf);
127 spin_unlock_irqrestore(&rtc_lock, flags);
128 local_flush_tlb();
129 pr_debug("1.\n");
130 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) =
131 start_eip >> 4;
132 pr_debug("2.\n");
133 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) =
134 start_eip & 0xf;
135 pr_debug("3.\n");
136}
137
138static inline void smpboot_restore_warm_reset_vector(void)
139{
140 unsigned long flags;
141
142 /*
143 * Install writable page 0 entry to set BIOS data area.
144 */
145 local_flush_tlb();
146
147 /*
148 * Paranoid: Set warm reset code and vector here back
149 * to default values.
150 */
151 spin_lock_irqsave(&rtc_lock, flags);
152 CMOS_WRITE(0, 0xf);
153 spin_unlock_irqrestore(&rtc_lock, flags);
154
155 *((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
156}
157
158/*
159 * Report back to the Boot Processor during boot time or to the caller processor
160 * during CPU online.
161 */
162static void smp_callin(void)
163{
164 int cpuid, phys_id;
165
166 /*
167 * If waken up by an INIT in an 82489DX configuration
168 * cpu_callout_mask guarantees we don't get here before
169 * an INIT_deassert IPI reaches our local APIC, so it is
170 * now safe to touch our local APIC.
171 */
172 cpuid = smp_processor_id();
173
174 /*
175 * (This works even if the APIC is not enabled.)
176 */
177 phys_id = read_apic_id();
178
179 /*
180 * the boot CPU has finished the init stage and is spinning
181 * on callin_map until we finish. We are free to set up this
182 * CPU, first the APIC. (this is probably redundant on most
183 * boards)
184 */
185 apic_ap_setup();
186
187 /*
188 * Save our processor parameters. Note: this information
189 * is needed for clock calibration.
190 */
191 smp_store_cpu_info(cpuid);
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 /*
204 * This must be done before setting cpu_online_mask
205 * or calling notify_cpu_starting.
206 */
207 set_cpu_sibling_map(raw_smp_processor_id());
208 wmb();
209
210 notify_cpu_starting(cpuid);
211
212 /*
213 * Allow the master to continue.
214 */
215 cpumask_set_cpu(cpuid, cpu_callin_mask);
216}
217
218static int cpu0_logical_apicid;
219static int enable_start_cpu0;
220/*
221 * Activate a secondary processor.
222 */
223static void notrace start_secondary(void *unused)
224{
225 /*
226 * Don't put *anything* before cpu_init(), SMP booting is too
227 * fragile that we want to limit the things done here to the
228 * most necessary things.
229 */
230 cpu_init();
231 x86_cpuinit.early_percpu_clock_init();
232 preempt_disable();
233 smp_callin();
234
235 enable_start_cpu0 = 0;
236
237#ifdef CONFIG_X86_32
238 /* switch away from the initial page table */
239 load_cr3(swapper_pg_dir);
240 __flush_tlb_all();
241#endif
242
243 /* otherwise gcc will move up smp_processor_id before the cpu_init */
244 barrier();
245 /*
246 * Check TSC synchronization with the BP:
247 */
248 check_tsc_sync_target();
249
250 /*
251 * Lock vector_lock and initialize the vectors on this cpu
252 * before setting the cpu online. We must set it online with
253 * vector_lock held to prevent a concurrent setup/teardown
254 * from seeing a half valid vector space.
255 */
256 lock_vector_lock();
257 setup_vector_irq(smp_processor_id());
258 set_cpu_online(smp_processor_id(), true);
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_update_package_map - Update the physical to logical package map
277 * @pkg: The physical package id as retrieved via CPUID
278 * @cpu: The cpu for which this is updated
279 */
280int topology_update_package_map(unsigned int pkg, unsigned int cpu)
281{
282 unsigned int new;
283
284 /* Called from early boot ? */
285 if (!physical_package_map)
286 return 0;
287
288 if (pkg >= max_physical_pkg_id)
289 return -EINVAL;
290
291 /* Set the logical package id */
292 if (test_and_set_bit(pkg, physical_package_map))
293 goto found;
294
295 if (logical_packages >= __max_logical_packages) {
296 pr_warn("Package %u of CPU %u exceeds BIOS package data %u.\n",
297 logical_packages, cpu, __max_logical_packages);
298 return -ENOSPC;
299 }
300
301 new = logical_packages++;
302 if (new != pkg) {
303 pr_info("CPU %u Converting physical %u to logical package %u\n",
304 cpu, pkg, new);
305 }
306 physical_to_logical_pkg[pkg] = new;
307
308found:
309 cpu_data(cpu).logical_proc_id = physical_to_logical_pkg[pkg];
310 return 0;
311}
312
313/**
314 * topology_phys_to_logical_pkg - Map a physical package id to a logical
315 *
316 * Returns logical package id or -1 if not found
317 */
318int topology_phys_to_logical_pkg(unsigned int phys_pkg)
319{
320 if (phys_pkg >= max_physical_pkg_id)
321 return -1;
322 return physical_to_logical_pkg[phys_pkg];
323}
324EXPORT_SYMBOL(topology_phys_to_logical_pkg);
325
326static void __init smp_init_package_map(struct cpuinfo_x86 *c, unsigned int cpu)
327{
328 unsigned int ncpus;
329 size_t size;
330
331 /*
332 * Today neither Intel nor AMD support heterogenous systems. That
333 * might change in the future....
334 *
335 * While ideally we'd want '* smp_num_siblings' in the below @ncpus
336 * computation, this won't actually work since some Intel BIOSes
337 * report inconsistent HT data when they disable HT.
338 *
339 * In particular, they reduce the APIC-IDs to only include the cores,
340 * but leave the CPUID topology to say there are (2) siblings.
341 * This means we don't know how many threads there will be until
342 * after the APIC enumeration.
343 *
344 * By not including this we'll sometimes over-estimate the number of
345 * logical packages by the amount of !present siblings, but this is
346 * still better than MAX_LOCAL_APIC.
347 *
348 * We use total_cpus not nr_cpu_ids because nr_cpu_ids can be limited
349 * on the command line leading to a similar issue as the HT disable
350 * problem because the hyperthreads are usually enumerated after the
351 * primary cores.
352 */
353 ncpus = boot_cpu_data.x86_max_cores;
354 if (!ncpus) {
355 pr_warn("x86_max_cores == zero !?!?");
356 ncpus = 1;
357 }
358
359 __max_logical_packages = DIV_ROUND_UP(total_cpus, ncpus);
360 logical_packages = 0;
361
362 /*
363 * Possibly larger than what we need as the number of apic ids per
364 * package can be smaller than the actual used apic ids.
365 */
366 max_physical_pkg_id = DIV_ROUND_UP(MAX_LOCAL_APIC, ncpus);
367 size = max_physical_pkg_id * sizeof(unsigned int);
368 physical_to_logical_pkg = kmalloc(size, GFP_KERNEL);
369 memset(physical_to_logical_pkg, 0xff, size);
370 size = BITS_TO_LONGS(max_physical_pkg_id) * sizeof(unsigned long);
371 physical_package_map = kzalloc(size, GFP_KERNEL);
372
373 pr_info("Max logical packages: %u\n", __max_logical_packages);
374
375 topology_update_package_map(c->phys_proc_id, cpu);
376}
377
378void __init smp_store_boot_cpu_info(void)
379{
380 int id = 0; /* CPU 0 */
381 struct cpuinfo_x86 *c = &cpu_data(id);
382
383 *c = boot_cpu_data;
384 c->cpu_index = id;
385 smp_init_package_map(c, id);
386}
387
388/*
389 * The bootstrap kernel entry code has set these up. Save them for
390 * a given CPU
391 */
392void smp_store_cpu_info(int id)
393{
394 struct cpuinfo_x86 *c = &cpu_data(id);
395
396 *c = boot_cpu_data;
397 c->cpu_index = id;
398 /*
399 * During boot time, CPU0 has this setup already. Save the info when
400 * bringing up AP or offlined CPU0.
401 */
402 identify_secondary_cpu(c);
403}
404
405static bool
406topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
407{
408 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
409
410 return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
411}
412
413static bool
414topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
415{
416 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
417
418 return !WARN_ONCE(!topology_same_node(c, o),
419 "sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
420 "[node: %d != %d]. Ignoring dependency.\n",
421 cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
422}
423
424#define link_mask(mfunc, c1, c2) \
425do { \
426 cpumask_set_cpu((c1), mfunc(c2)); \
427 cpumask_set_cpu((c2), mfunc(c1)); \
428} while (0)
429
430static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
431{
432 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
433 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
434
435 if (c->phys_proc_id == o->phys_proc_id &&
436 per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2)) {
437 if (c->cpu_core_id == o->cpu_core_id)
438 return topology_sane(c, o, "smt");
439
440 if ((c->cu_id != 0xff) &&
441 (o->cu_id != 0xff) &&
442 (c->cu_id == o->cu_id))
443 return topology_sane(c, o, "smt");
444 }
445
446 } else if (c->phys_proc_id == o->phys_proc_id &&
447 c->cpu_core_id == o->cpu_core_id) {
448 return topology_sane(c, o, "smt");
449 }
450
451 return false;
452}
453
454static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
455{
456 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
457
458 if (per_cpu(cpu_llc_id, cpu1) != BAD_APICID &&
459 per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2))
460 return topology_sane(c, o, "llc");
461
462 return false;
463}
464
465/*
466 * Unlike the other levels, we do not enforce keeping a
467 * multicore group inside a NUMA node. If this happens, we will
468 * discard the MC level of the topology later.
469 */
470static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
471{
472 if (c->phys_proc_id == o->phys_proc_id)
473 return true;
474 return false;
475}
476
477#if defined(CONFIG_SCHED_SMT) || defined(CONFIG_SCHED_MC)
478static inline int x86_sched_itmt_flags(void)
479{
480 return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
481}
482
483#ifdef CONFIG_SCHED_MC
484static int x86_core_flags(void)
485{
486 return cpu_core_flags() | x86_sched_itmt_flags();
487}
488#endif
489#ifdef CONFIG_SCHED_SMT
490static int x86_smt_flags(void)
491{
492 return cpu_smt_flags() | x86_sched_itmt_flags();
493}
494#endif
495#endif
496
497static struct sched_domain_topology_level x86_numa_in_package_topology[] = {
498#ifdef CONFIG_SCHED_SMT
499 { cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
500#endif
501#ifdef CONFIG_SCHED_MC
502 { cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
503#endif
504 { NULL, },
505};
506
507static struct sched_domain_topology_level x86_topology[] = {
508#ifdef CONFIG_SCHED_SMT
509 { cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
510#endif
511#ifdef CONFIG_SCHED_MC
512 { cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
513#endif
514 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
515 { NULL, },
516};
517
518/*
519 * Set if a package/die has multiple NUMA nodes inside.
520 * AMD Magny-Cours and Intel Cluster-on-Die have this.
521 */
522static bool x86_has_numa_in_package;
523
524void set_cpu_sibling_map(int cpu)
525{
526 bool has_smt = smp_num_siblings > 1;
527 bool has_mp = has_smt || boot_cpu_data.x86_max_cores > 1;
528 struct cpuinfo_x86 *c = &cpu_data(cpu);
529 struct cpuinfo_x86 *o;
530 int i, threads;
531
532 cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
533
534 if (!has_mp) {
535 cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
536 cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
537 cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
538 c->booted_cores = 1;
539 return;
540 }
541
542 for_each_cpu(i, cpu_sibling_setup_mask) {
543 o = &cpu_data(i);
544
545 if ((i == cpu) || (has_smt && match_smt(c, o)))
546 link_mask(topology_sibling_cpumask, cpu, i);
547
548 if ((i == cpu) || (has_mp && match_llc(c, o)))
549 link_mask(cpu_llc_shared_mask, cpu, i);
550
551 }
552
553 /*
554 * This needs a separate iteration over the cpus because we rely on all
555 * topology_sibling_cpumask links to be set-up.
556 */
557 for_each_cpu(i, cpu_sibling_setup_mask) {
558 o = &cpu_data(i);
559
560 if ((i == cpu) || (has_mp && match_die(c, o))) {
561 link_mask(topology_core_cpumask, cpu, i);
562
563 /*
564 * Does this new cpu bringup a new core?
565 */
566 if (cpumask_weight(
567 topology_sibling_cpumask(cpu)) == 1) {
568 /*
569 * for each core in package, increment
570 * the booted_cores for this new cpu
571 */
572 if (cpumask_first(
573 topology_sibling_cpumask(i)) == i)
574 c->booted_cores++;
575 /*
576 * increment the core count for all
577 * the other cpus in this package
578 */
579 if (i != cpu)
580 cpu_data(i).booted_cores++;
581 } else if (i != cpu && !c->booted_cores)
582 c->booted_cores = cpu_data(i).booted_cores;
583 }
584 if (match_die(c, o) && !topology_same_node(c, o))
585 x86_has_numa_in_package = true;
586 }
587
588 threads = cpumask_weight(topology_sibling_cpumask(cpu));
589 if (threads > __max_smt_threads)
590 __max_smt_threads = threads;
591}
592
593/* maps the cpu to the sched domain representing multi-core */
594const struct cpumask *cpu_coregroup_mask(int cpu)
595{
596 return cpu_llc_shared_mask(cpu);
597}
598
599static void impress_friends(void)
600{
601 int cpu;
602 unsigned long bogosum = 0;
603 /*
604 * Allow the user to impress friends.
605 */
606 pr_debug("Before bogomips\n");
607 for_each_possible_cpu(cpu)
608 if (cpumask_test_cpu(cpu, cpu_callout_mask))
609 bogosum += cpu_data(cpu).loops_per_jiffy;
610 pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
611 num_online_cpus(),
612 bogosum/(500000/HZ),
613 (bogosum/(5000/HZ))%100);
614
615 pr_debug("Before bogocount - setting activated=1\n");
616}
617
618void __inquire_remote_apic(int apicid)
619{
620 unsigned i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
621 const char * const names[] = { "ID", "VERSION", "SPIV" };
622 int timeout;
623 u32 status;
624
625 pr_info("Inquiring remote APIC 0x%x...\n", apicid);
626
627 for (i = 0; i < ARRAY_SIZE(regs); i++) {
628 pr_info("... APIC 0x%x %s: ", apicid, names[i]);
629
630 /*
631 * Wait for idle.
632 */
633 status = safe_apic_wait_icr_idle();
634 if (status)
635 pr_cont("a previous APIC delivery may have failed\n");
636
637 apic_icr_write(APIC_DM_REMRD | regs[i], apicid);
638
639 timeout = 0;
640 do {
641 udelay(100);
642 status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
643 } while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
644
645 switch (status) {
646 case APIC_ICR_RR_VALID:
647 status = apic_read(APIC_RRR);
648 pr_cont("%08x\n", status);
649 break;
650 default:
651 pr_cont("failed\n");
652 }
653 }
654}
655
656/*
657 * The Multiprocessor Specification 1.4 (1997) example code suggests
658 * that there should be a 10ms delay between the BSP asserting INIT
659 * and de-asserting INIT, when starting a remote processor.
660 * But that slows boot and resume on modern processors, which include
661 * many cores and don't require that delay.
662 *
663 * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
664 * Modern processor families are quirked to remove the delay entirely.
665 */
666#define UDELAY_10MS_DEFAULT 10000
667
668static unsigned int init_udelay = UINT_MAX;
669
670static int __init cpu_init_udelay(char *str)
671{
672 get_option(&str, &init_udelay);
673
674 return 0;
675}
676early_param("cpu_init_udelay", cpu_init_udelay);
677
678static void __init smp_quirk_init_udelay(void)
679{
680 /* if cmdline changed it from default, leave it alone */
681 if (init_udelay != UINT_MAX)
682 return;
683
684 /* if modern processor, use no delay */
685 if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
686 ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
687 init_udelay = 0;
688 return;
689 }
690 /* else, use legacy delay */
691 init_udelay = UDELAY_10MS_DEFAULT;
692}
693
694/*
695 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
696 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
697 * won't ... remember to clear down the APIC, etc later.
698 */
699int
700wakeup_secondary_cpu_via_nmi(int apicid, unsigned long start_eip)
701{
702 unsigned long send_status, accept_status = 0;
703 int maxlvt;
704
705 /* Target chip */
706 /* Boot on the stack */
707 /* Kick the second */
708 apic_icr_write(APIC_DM_NMI | apic->dest_logical, apicid);
709
710 pr_debug("Waiting for send to finish...\n");
711 send_status = safe_apic_wait_icr_idle();
712
713 /*
714 * Give the other CPU some time to accept the IPI.
715 */
716 udelay(200);
717 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
718 maxlvt = lapic_get_maxlvt();
719 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
720 apic_write(APIC_ESR, 0);
721 accept_status = (apic_read(APIC_ESR) & 0xEF);
722 }
723 pr_debug("NMI sent\n");
724
725 if (send_status)
726 pr_err("APIC never delivered???\n");
727 if (accept_status)
728 pr_err("APIC delivery error (%lx)\n", accept_status);
729
730 return (send_status | accept_status);
731}
732
733static int
734wakeup_secondary_cpu_via_init(int phys_apicid, unsigned long start_eip)
735{
736 unsigned long send_status = 0, accept_status = 0;
737 int maxlvt, num_starts, j;
738
739 maxlvt = lapic_get_maxlvt();
740
741 /*
742 * Be paranoid about clearing APIC errors.
743 */
744 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
745 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
746 apic_write(APIC_ESR, 0);
747 apic_read(APIC_ESR);
748 }
749
750 pr_debug("Asserting INIT\n");
751
752 /*
753 * Turn INIT on target chip
754 */
755 /*
756 * Send IPI
757 */
758 apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT,
759 phys_apicid);
760
761 pr_debug("Waiting for send to finish...\n");
762 send_status = safe_apic_wait_icr_idle();
763
764 udelay(init_udelay);
765
766 pr_debug("Deasserting INIT\n");
767
768 /* Target chip */
769 /* Send IPI */
770 apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
771
772 pr_debug("Waiting for send to finish...\n");
773 send_status = safe_apic_wait_icr_idle();
774
775 mb();
776
777 /*
778 * Should we send STARTUP IPIs ?
779 *
780 * Determine this based on the APIC version.
781 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
782 */
783 if (APIC_INTEGRATED(boot_cpu_apic_version))
784 num_starts = 2;
785 else
786 num_starts = 0;
787
788 /*
789 * Run STARTUP IPI loop.
790 */
791 pr_debug("#startup loops: %d\n", num_starts);
792
793 for (j = 1; j <= num_starts; j++) {
794 pr_debug("Sending STARTUP #%d\n", j);
795 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
796 apic_write(APIC_ESR, 0);
797 apic_read(APIC_ESR);
798 pr_debug("After apic_write\n");
799
800 /*
801 * STARTUP IPI
802 */
803
804 /* Target chip */
805 /* Boot on the stack */
806 /* Kick the second */
807 apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
808 phys_apicid);
809
810 /*
811 * Give the other CPU some time to accept the IPI.
812 */
813 if (init_udelay == 0)
814 udelay(10);
815 else
816 udelay(300);
817
818 pr_debug("Startup point 1\n");
819
820 pr_debug("Waiting for send to finish...\n");
821 send_status = safe_apic_wait_icr_idle();
822
823 /*
824 * Give the other CPU some time to accept the IPI.
825 */
826 if (init_udelay == 0)
827 udelay(10);
828 else
829 udelay(200);
830
831 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
832 apic_write(APIC_ESR, 0);
833 accept_status = (apic_read(APIC_ESR) & 0xEF);
834 if (send_status || accept_status)
835 break;
836 }
837 pr_debug("After Startup\n");
838
839 if (send_status)
840 pr_err("APIC never delivered???\n");
841 if (accept_status)
842 pr_err("APIC delivery error (%lx)\n", accept_status);
843
844 return (send_status | accept_status);
845}
846
847/* reduce the number of lines printed when booting a large cpu count system */
848static void announce_cpu(int cpu, int apicid)
849{
850 static int current_node = -1;
851 int node = early_cpu_to_node(cpu);
852 static int width, node_width;
853
854 if (!width)
855 width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
856
857 if (!node_width)
858 node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
859
860 if (cpu == 1)
861 printk(KERN_INFO "x86: Booting SMP configuration:\n");
862
863 if (system_state == SYSTEM_BOOTING) {
864 if (node != current_node) {
865 if (current_node > (-1))
866 pr_cont("\n");
867 current_node = node;
868
869 printk(KERN_INFO ".... node %*s#%d, CPUs: ",
870 node_width - num_digits(node), " ", node);
871 }
872
873 /* Add padding for the BSP */
874 if (cpu == 1)
875 pr_cont("%*s", width + 1, " ");
876
877 pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
878
879 } else
880 pr_info("Booting Node %d Processor %d APIC 0x%x\n",
881 node, cpu, apicid);
882}
883
884static int wakeup_cpu0_nmi(unsigned int cmd, struct pt_regs *regs)
885{
886 int cpu;
887
888 cpu = smp_processor_id();
889 if (cpu == 0 && !cpu_online(cpu) && enable_start_cpu0)
890 return NMI_HANDLED;
891
892 return NMI_DONE;
893}
894
895/*
896 * Wake up AP by INIT, INIT, STARTUP sequence.
897 *
898 * Instead of waiting for STARTUP after INITs, BSP will execute the BIOS
899 * boot-strap code which is not a desired behavior for waking up BSP. To
900 * void the boot-strap code, wake up CPU0 by NMI instead.
901 *
902 * This works to wake up soft offlined CPU0 only. If CPU0 is hard offlined
903 * (i.e. physically hot removed and then hot added), NMI won't wake it up.
904 * We'll change this code in the future to wake up hard offlined CPU0 if
905 * real platform and request are available.
906 */
907static int
908wakeup_cpu_via_init_nmi(int cpu, unsigned long start_ip, int apicid,
909 int *cpu0_nmi_registered)
910{
911 int id;
912 int boot_error;
913
914 preempt_disable();
915
916 /*
917 * Wake up AP by INIT, INIT, STARTUP sequence.
918 */
919 if (cpu) {
920 boot_error = wakeup_secondary_cpu_via_init(apicid, start_ip);
921 goto out;
922 }
923
924 /*
925 * Wake up BSP by nmi.
926 *
927 * Register a NMI handler to help wake up CPU0.
928 */
929 boot_error = register_nmi_handler(NMI_LOCAL,
930 wakeup_cpu0_nmi, 0, "wake_cpu0");
931
932 if (!boot_error) {
933 enable_start_cpu0 = 1;
934 *cpu0_nmi_registered = 1;
935 if (apic->dest_logical == APIC_DEST_LOGICAL)
936 id = cpu0_logical_apicid;
937 else
938 id = apicid;
939 boot_error = wakeup_secondary_cpu_via_nmi(id, start_ip);
940 }
941
942out:
943 preempt_enable();
944
945 return boot_error;
946}
947
948void common_cpu_up(unsigned int cpu, struct task_struct *idle)
949{
950 /* Just in case we booted with a single CPU. */
951 alternatives_enable_smp();
952
953 per_cpu(current_task, cpu) = idle;
954
955#ifdef CONFIG_X86_32
956 /* Stack for startup_32 can be just as for start_secondary onwards */
957 irq_ctx_init(cpu);
958 per_cpu(cpu_current_top_of_stack, cpu) =
959 (unsigned long)task_stack_page(idle) + THREAD_SIZE;
960#else
961 initial_gs = per_cpu_offset(cpu);
962#endif
963}
964
965/*
966 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
967 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
968 * Returns zero if CPU booted OK, else error code from
969 * ->wakeup_secondary_cpu.
970 */
971static int do_boot_cpu(int apicid, int cpu, struct task_struct *idle)
972{
973 volatile u32 *trampoline_status =
974 (volatile u32 *) __va(real_mode_header->trampoline_status);
975 /* start_ip had better be page-aligned! */
976 unsigned long start_ip = real_mode_header->trampoline_start;
977
978 unsigned long boot_error = 0;
979 int cpu0_nmi_registered = 0;
980 unsigned long timeout;
981
982 idle->thread.sp = (unsigned long)task_pt_regs(idle);
983 early_gdt_descr.address = (unsigned long)get_cpu_gdt_table(cpu);
984 initial_code = (unsigned long)start_secondary;
985 initial_stack = idle->thread.sp;
986
987 /*
988 * Enable the espfix hack for this CPU
989 */
990#ifdef CONFIG_X86_ESPFIX64
991 init_espfix_ap(cpu);
992#endif
993
994 /* So we see what's up */
995 announce_cpu(cpu, apicid);
996
997 /*
998 * This grunge runs the startup process for
999 * the targeted processor.
1000 */
1001
1002 if (get_uv_system_type() != UV_NON_UNIQUE_APIC) {
1003
1004 pr_debug("Setting warm reset code and vector.\n");
1005
1006 smpboot_setup_warm_reset_vector(start_ip);
1007 /*
1008 * Be paranoid about clearing APIC errors.
1009 */
1010 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
1011 apic_write(APIC_ESR, 0);
1012 apic_read(APIC_ESR);
1013 }
1014 }
1015
1016 /*
1017 * AP might wait on cpu_callout_mask in cpu_init() with
1018 * cpu_initialized_mask set if previous attempt to online
1019 * it timed-out. Clear cpu_initialized_mask so that after
1020 * INIT/SIPI it could start with a clean state.
1021 */
1022 cpumask_clear_cpu(cpu, cpu_initialized_mask);
1023 smp_mb();
1024
1025 /*
1026 * Wake up a CPU in difference cases:
1027 * - Use the method in the APIC driver if it's defined
1028 * Otherwise,
1029 * - Use an INIT boot APIC message for APs or NMI for BSP.
1030 */
1031 if (apic->wakeup_secondary_cpu)
1032 boot_error = apic->wakeup_secondary_cpu(apicid, start_ip);
1033 else
1034 boot_error = wakeup_cpu_via_init_nmi(cpu, start_ip, apicid,
1035 &cpu0_nmi_registered);
1036
1037 if (!boot_error) {
1038 /*
1039 * Wait 10s total for first sign of life from AP
1040 */
1041 boot_error = -1;
1042 timeout = jiffies + 10*HZ;
1043 while (time_before(jiffies, timeout)) {
1044 if (cpumask_test_cpu(cpu, cpu_initialized_mask)) {
1045 /*
1046 * Tell AP to proceed with initialization
1047 */
1048 cpumask_set_cpu(cpu, cpu_callout_mask);
1049 boot_error = 0;
1050 break;
1051 }
1052 schedule();
1053 }
1054 }
1055
1056 if (!boot_error) {
1057 /*
1058 * Wait till AP completes initial initialization
1059 */
1060 while (!cpumask_test_cpu(cpu, cpu_callin_mask)) {
1061 /*
1062 * Allow other tasks to run while we wait for the
1063 * AP to come online. This also gives a chance
1064 * for the MTRR work(triggered by the AP coming online)
1065 * to be completed in the stop machine context.
1066 */
1067 schedule();
1068 }
1069 }
1070
1071 /* mark "stuck" area as not stuck */
1072 *trampoline_status = 0;
1073
1074 if (get_uv_system_type() != UV_NON_UNIQUE_APIC) {
1075 /*
1076 * Cleanup possible dangling ends...
1077 */
1078 smpboot_restore_warm_reset_vector();
1079 }
1080 /*
1081 * Clean up the nmi handler. Do this after the callin and callout sync
1082 * to avoid impact of possible long unregister time.
1083 */
1084 if (cpu0_nmi_registered)
1085 unregister_nmi_handler(NMI_LOCAL, "wake_cpu0");
1086
1087 return boot_error;
1088}
1089
1090int native_cpu_up(unsigned int cpu, struct task_struct *tidle)
1091{
1092 int apicid = apic->cpu_present_to_apicid(cpu);
1093 unsigned long flags;
1094 int err;
1095
1096 WARN_ON(irqs_disabled());
1097
1098 pr_debug("++++++++++++++++++++=_---CPU UP %u\n", cpu);
1099
1100 if (apicid == BAD_APICID ||
1101 !physid_isset(apicid, phys_cpu_present_map) ||
1102 !apic->apic_id_valid(apicid)) {
1103 pr_err("%s: bad cpu %d\n", __func__, cpu);
1104 return -EINVAL;
1105 }
1106
1107 /*
1108 * Already booted CPU?
1109 */
1110 if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
1111 pr_debug("do_boot_cpu %d Already started\n", cpu);
1112 return -ENOSYS;
1113 }
1114
1115 /*
1116 * Save current MTRR state in case it was changed since early boot
1117 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1118 */
1119 mtrr_save_state();
1120
1121 /* x86 CPUs take themselves offline, so delayed offline is OK. */
1122 err = cpu_check_up_prepare(cpu);
1123 if (err && err != -EBUSY)
1124 return err;
1125
1126 /* the FPU context is blank, nobody can own it */
1127 per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
1128
1129 common_cpu_up(cpu, tidle);
1130
1131 err = do_boot_cpu(apicid, cpu, tidle);
1132 if (err) {
1133 pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
1134 return -EIO;
1135 }
1136
1137 /*
1138 * Check TSC synchronization with the AP (keep irqs disabled
1139 * while doing so):
1140 */
1141 local_irq_save(flags);
1142 check_tsc_sync_source(cpu);
1143 local_irq_restore(flags);
1144
1145 while (!cpu_online(cpu)) {
1146 cpu_relax();
1147 touch_nmi_watchdog();
1148 }
1149
1150 return 0;
1151}
1152
1153/**
1154 * arch_disable_smp_support() - disables SMP support for x86 at runtime
1155 */
1156void arch_disable_smp_support(void)
1157{
1158 disable_ioapic_support();
1159}
1160
1161/*
1162 * Fall back to non SMP mode after errors.
1163 *
1164 * RED-PEN audit/test this more. I bet there is more state messed up here.
1165 */
1166static __init void disable_smp(void)
1167{
1168 pr_info("SMP disabled\n");
1169
1170 disable_ioapic_support();
1171
1172 init_cpu_present(cpumask_of(0));
1173 init_cpu_possible(cpumask_of(0));
1174
1175 if (smp_found_config)
1176 physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
1177 else
1178 physid_set_mask_of_physid(0, &phys_cpu_present_map);
1179 cpumask_set_cpu(0, topology_sibling_cpumask(0));
1180 cpumask_set_cpu(0, topology_core_cpumask(0));
1181}
1182
1183enum {
1184 SMP_OK,
1185 SMP_NO_CONFIG,
1186 SMP_NO_APIC,
1187 SMP_FORCE_UP,
1188};
1189
1190/*
1191 * Various sanity checks.
1192 */
1193static int __init smp_sanity_check(unsigned max_cpus)
1194{
1195 preempt_disable();
1196
1197#if !defined(CONFIG_X86_BIGSMP) && defined(CONFIG_X86_32)
1198 if (def_to_bigsmp && nr_cpu_ids > 8) {
1199 unsigned int cpu;
1200 unsigned nr;
1201
1202 pr_warn("More than 8 CPUs detected - skipping them\n"
1203 "Use CONFIG_X86_BIGSMP\n");
1204
1205 nr = 0;
1206 for_each_present_cpu(cpu) {
1207 if (nr >= 8)
1208 set_cpu_present(cpu, false);
1209 nr++;
1210 }
1211
1212 nr = 0;
1213 for_each_possible_cpu(cpu) {
1214 if (nr >= 8)
1215 set_cpu_possible(cpu, false);
1216 nr++;
1217 }
1218
1219 nr_cpu_ids = 8;
1220 }
1221#endif
1222
1223 if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map)) {
1224 pr_warn("weird, boot CPU (#%d) not listed by the BIOS\n",
1225 hard_smp_processor_id());
1226
1227 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1228 }
1229
1230 /*
1231 * If we couldn't find an SMP configuration at boot time,
1232 * get out of here now!
1233 */
1234 if (!smp_found_config && !acpi_lapic) {
1235 preempt_enable();
1236 pr_notice("SMP motherboard not detected\n");
1237 return SMP_NO_CONFIG;
1238 }
1239
1240 /*
1241 * Should not be necessary because the MP table should list the boot
1242 * CPU too, but we do it for the sake of robustness anyway.
1243 */
1244 if (!apic->check_phys_apicid_present(boot_cpu_physical_apicid)) {
1245 pr_notice("weird, boot CPU (#%d) not listed by the BIOS\n",
1246 boot_cpu_physical_apicid);
1247 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1248 }
1249 preempt_enable();
1250
1251 /*
1252 * If we couldn't find a local APIC, then get out of here now!
1253 */
1254 if (APIC_INTEGRATED(boot_cpu_apic_version) &&
1255 !boot_cpu_has(X86_FEATURE_APIC)) {
1256 if (!disable_apic) {
1257 pr_err("BIOS bug, local APIC #%d not detected!...\n",
1258 boot_cpu_physical_apicid);
1259 pr_err("... forcing use of dummy APIC emulation (tell your hw vendor)\n");
1260 }
1261 return SMP_NO_APIC;
1262 }
1263
1264 /*
1265 * If SMP should be disabled, then really disable it!
1266 */
1267 if (!max_cpus) {
1268 pr_info("SMP mode deactivated\n");
1269 return SMP_FORCE_UP;
1270 }
1271
1272 return SMP_OK;
1273}
1274
1275static void __init smp_cpu_index_default(void)
1276{
1277 int i;
1278 struct cpuinfo_x86 *c;
1279
1280 for_each_possible_cpu(i) {
1281 c = &cpu_data(i);
1282 /* mark all to hotplug */
1283 c->cpu_index = nr_cpu_ids;
1284 }
1285}
1286
1287/*
1288 * Prepare for SMP bootup. The MP table or ACPI has been read
1289 * earlier. Just do some sanity checking here and enable APIC mode.
1290 */
1291void __init native_smp_prepare_cpus(unsigned int max_cpus)
1292{
1293 unsigned int i;
1294
1295 smp_cpu_index_default();
1296
1297 /*
1298 * Setup boot CPU information
1299 */
1300 smp_store_boot_cpu_info(); /* Final full version of the data */
1301 cpumask_copy(cpu_callin_mask, cpumask_of(0));
1302 mb();
1303
1304 for_each_possible_cpu(i) {
1305 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1306 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1307 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1308 }
1309
1310 /*
1311 * Set 'default' x86 topology, this matches default_topology() in that
1312 * it has NUMA nodes as a topology level. See also
1313 * native_smp_cpus_done().
1314 *
1315 * Must be done before set_cpus_sibling_map() is ran.
1316 */
1317 set_sched_topology(x86_topology);
1318
1319 set_cpu_sibling_map(0);
1320
1321 switch (smp_sanity_check(max_cpus)) {
1322 case SMP_NO_CONFIG:
1323 disable_smp();
1324 if (APIC_init_uniprocessor())
1325 pr_notice("Local APIC not detected. Using dummy APIC emulation.\n");
1326 return;
1327 case SMP_NO_APIC:
1328 disable_smp();
1329 return;
1330 case SMP_FORCE_UP:
1331 disable_smp();
1332 apic_bsp_setup(false);
1333 return;
1334 case SMP_OK:
1335 break;
1336 }
1337
1338 if (read_apic_id() != boot_cpu_physical_apicid) {
1339 panic("Boot APIC ID in local APIC unexpected (%d vs %d)",
1340 read_apic_id(), boot_cpu_physical_apicid);
1341 /* Or can we switch back to PIC here? */
1342 }
1343
1344 default_setup_apic_routing();
1345 cpu0_logical_apicid = apic_bsp_setup(false);
1346
1347 pr_info("CPU0: ");
1348 print_cpu_info(&cpu_data(0));
1349
1350 if (is_uv_system())
1351 uv_system_init();
1352
1353 set_mtrr_aps_delayed_init();
1354
1355 smp_quirk_init_udelay();
1356}
1357
1358void arch_enable_nonboot_cpus_begin(void)
1359{
1360 set_mtrr_aps_delayed_init();
1361}
1362
1363void arch_enable_nonboot_cpus_end(void)
1364{
1365 mtrr_aps_init();
1366}
1367
1368/*
1369 * Early setup to make printk work.
1370 */
1371void __init native_smp_prepare_boot_cpu(void)
1372{
1373 int me = smp_processor_id();
1374 switch_to_new_gdt(me);
1375 /* already set me in cpu_online_mask in boot_cpu_init() */
1376 cpumask_set_cpu(me, cpu_callout_mask);
1377 cpu_set_state_online(me);
1378}
1379
1380void __init native_smp_cpus_done(unsigned int max_cpus)
1381{
1382 pr_debug("Boot done\n");
1383
1384 if (x86_has_numa_in_package)
1385 set_sched_topology(x86_numa_in_package_topology);
1386
1387 nmi_selftest();
1388 impress_friends();
1389 setup_ioapic_dest();
1390 mtrr_aps_init();
1391}
1392
1393static int __initdata setup_possible_cpus = -1;
1394static int __init _setup_possible_cpus(char *str)
1395{
1396 get_option(&str, &setup_possible_cpus);
1397 return 0;
1398}
1399early_param("possible_cpus", _setup_possible_cpus);
1400
1401
1402/*
1403 * cpu_possible_mask should be static, it cannot change as cpu's
1404 * are onlined, or offlined. The reason is per-cpu data-structures
1405 * are allocated by some modules at init time, and dont expect to
1406 * do this dynamically on cpu arrival/departure.
1407 * cpu_present_mask on the other hand can change dynamically.
1408 * In case when cpu_hotplug is not compiled, then we resort to current
1409 * behaviour, which is cpu_possible == cpu_present.
1410 * - Ashok Raj
1411 *
1412 * Three ways to find out the number of additional hotplug CPUs:
1413 * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1414 * - The user can overwrite it with possible_cpus=NUM
1415 * - Otherwise don't reserve additional CPUs.
1416 * We do this because additional CPUs waste a lot of memory.
1417 * -AK
1418 */
1419__init void prefill_possible_map(void)
1420{
1421 int i, possible;
1422
1423 /* No boot processor was found in mptable or ACPI MADT */
1424 if (!num_processors) {
1425 if (boot_cpu_has(X86_FEATURE_APIC)) {
1426 int apicid = boot_cpu_physical_apicid;
1427 int cpu = hard_smp_processor_id();
1428
1429 pr_warn("Boot CPU (id %d) not listed by BIOS\n", cpu);
1430
1431 /* Make sure boot cpu is enumerated */
1432 if (apic->cpu_present_to_apicid(0) == BAD_APICID &&
1433 apic->apic_id_valid(apicid))
1434 generic_processor_info(apicid, boot_cpu_apic_version);
1435 }
1436
1437 if (!num_processors)
1438 num_processors = 1;
1439 }
1440
1441 i = setup_max_cpus ?: 1;
1442 if (setup_possible_cpus == -1) {
1443 possible = num_processors;
1444#ifdef CONFIG_HOTPLUG_CPU
1445 if (setup_max_cpus)
1446 possible += disabled_cpus;
1447#else
1448 if (possible > i)
1449 possible = i;
1450#endif
1451 } else
1452 possible = setup_possible_cpus;
1453
1454 total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1455
1456 /* nr_cpu_ids could be reduced via nr_cpus= */
1457 if (possible > nr_cpu_ids) {
1458 pr_warn("%d Processors exceeds NR_CPUS limit of %d\n",
1459 possible, nr_cpu_ids);
1460 possible = nr_cpu_ids;
1461 }
1462
1463#ifdef CONFIG_HOTPLUG_CPU
1464 if (!setup_max_cpus)
1465#endif
1466 if (possible > i) {
1467 pr_warn("%d Processors exceeds max_cpus limit of %u\n",
1468 possible, setup_max_cpus);
1469 possible = i;
1470 }
1471
1472 nr_cpu_ids = possible;
1473
1474 pr_info("Allowing %d CPUs, %d hotplug CPUs\n",
1475 possible, max_t(int, possible - num_processors, 0));
1476
1477 reset_cpu_possible_mask();
1478
1479 for (i = 0; i < possible; i++)
1480 set_cpu_possible(i, true);
1481}
1482
1483#ifdef CONFIG_HOTPLUG_CPU
1484
1485/* Recompute SMT state for all CPUs on offline */
1486static void recompute_smt_state(void)
1487{
1488 int max_threads, cpu;
1489
1490 max_threads = 0;
1491 for_each_online_cpu (cpu) {
1492 int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1493
1494 if (threads > max_threads)
1495 max_threads = threads;
1496 }
1497 __max_smt_threads = max_threads;
1498}
1499
1500static void remove_siblinginfo(int cpu)
1501{
1502 int sibling;
1503 struct cpuinfo_x86 *c = &cpu_data(cpu);
1504
1505 for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1506 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1507 /*/
1508 * last thread sibling in this cpu core going down
1509 */
1510 if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1511 cpu_data(sibling).booted_cores--;
1512 }
1513
1514 for_each_cpu(sibling, topology_sibling_cpumask(cpu))
1515 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1516 for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1517 cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1518 cpumask_clear(cpu_llc_shared_mask(cpu));
1519 cpumask_clear(topology_sibling_cpumask(cpu));
1520 cpumask_clear(topology_core_cpumask(cpu));
1521 c->phys_proc_id = 0;
1522 c->cpu_core_id = 0;
1523 cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1524 recompute_smt_state();
1525}
1526
1527static void remove_cpu_from_maps(int cpu)
1528{
1529 set_cpu_online(cpu, false);
1530 cpumask_clear_cpu(cpu, cpu_callout_mask);
1531 cpumask_clear_cpu(cpu, cpu_callin_mask);
1532 /* was set by cpu_init() */
1533 cpumask_clear_cpu(cpu, cpu_initialized_mask);
1534 numa_remove_cpu(cpu);
1535}
1536
1537void cpu_disable_common(void)
1538{
1539 int cpu = smp_processor_id();
1540
1541 remove_siblinginfo(cpu);
1542
1543 /* It's now safe to remove this processor from the online map */
1544 lock_vector_lock();
1545 remove_cpu_from_maps(cpu);
1546 unlock_vector_lock();
1547 fixup_irqs();
1548}
1549
1550int native_cpu_disable(void)
1551{
1552 int ret;
1553
1554 ret = check_irq_vectors_for_cpu_disable();
1555 if (ret)
1556 return ret;
1557
1558 clear_local_APIC();
1559 cpu_disable_common();
1560
1561 return 0;
1562}
1563
1564int common_cpu_die(unsigned int cpu)
1565{
1566 int ret = 0;
1567
1568 /* We don't do anything here: idle task is faking death itself. */
1569
1570 /* They ack this in play_dead() by setting CPU_DEAD */
1571 if (cpu_wait_death(cpu, 5)) {
1572 if (system_state == SYSTEM_RUNNING)
1573 pr_info("CPU %u is now offline\n", cpu);
1574 } else {
1575 pr_err("CPU %u didn't die...\n", cpu);
1576 ret = -1;
1577 }
1578
1579 return ret;
1580}
1581
1582void native_cpu_die(unsigned int cpu)
1583{
1584 common_cpu_die(cpu);
1585}
1586
1587void play_dead_common(void)
1588{
1589 idle_task_exit();
1590 reset_lazy_tlbstate();
1591
1592 /* Ack it */
1593 (void)cpu_report_death();
1594
1595 /*
1596 * With physical CPU hotplug, we should halt the cpu
1597 */
1598 local_irq_disable();
1599}
1600
1601static bool wakeup_cpu0(void)
1602{
1603 if (smp_processor_id() == 0 && enable_start_cpu0)
1604 return true;
1605
1606 return false;
1607}
1608
1609/*
1610 * We need to flush the caches before going to sleep, lest we have
1611 * dirty data in our caches when we come back up.
1612 */
1613static inline void mwait_play_dead(void)
1614{
1615 unsigned int eax, ebx, ecx, edx;
1616 unsigned int highest_cstate = 0;
1617 unsigned int highest_subcstate = 0;
1618 void *mwait_ptr;
1619 int i;
1620
1621 if (!this_cpu_has(X86_FEATURE_MWAIT))
1622 return;
1623 if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1624 return;
1625 if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1626 return;
1627
1628 eax = CPUID_MWAIT_LEAF;
1629 ecx = 0;
1630 native_cpuid(&eax, &ebx, &ecx, &edx);
1631
1632 /*
1633 * eax will be 0 if EDX enumeration is not valid.
1634 * Initialized below to cstate, sub_cstate value when EDX is valid.
1635 */
1636 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1637 eax = 0;
1638 } else {
1639 edx >>= MWAIT_SUBSTATE_SIZE;
1640 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1641 if (edx & MWAIT_SUBSTATE_MASK) {
1642 highest_cstate = i;
1643 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1644 }
1645 }
1646 eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1647 (highest_subcstate - 1);
1648 }
1649
1650 /*
1651 * This should be a memory location in a cache line which is
1652 * unlikely to be touched by other processors. The actual
1653 * content is immaterial as it is not actually modified in any way.
1654 */
1655 mwait_ptr = ¤t_thread_info()->flags;
1656
1657 wbinvd();
1658
1659 while (1) {
1660 /*
1661 * The CLFLUSH is a workaround for erratum AAI65 for
1662 * the Xeon 7400 series. It's not clear it is actually
1663 * needed, but it should be harmless in either case.
1664 * The WBINVD is insufficient due to the spurious-wakeup
1665 * case where we return around the loop.
1666 */
1667 mb();
1668 clflush(mwait_ptr);
1669 mb();
1670 __monitor(mwait_ptr, 0, 0);
1671 mb();
1672 __mwait(eax, 0);
1673 /*
1674 * If NMI wants to wake up CPU0, start CPU0.
1675 */
1676 if (wakeup_cpu0())
1677 start_cpu0();
1678 }
1679}
1680
1681void hlt_play_dead(void)
1682{
1683 if (__this_cpu_read(cpu_info.x86) >= 4)
1684 wbinvd();
1685
1686 while (1) {
1687 native_halt();
1688 /*
1689 * If NMI wants to wake up CPU0, start CPU0.
1690 */
1691 if (wakeup_cpu0())
1692 start_cpu0();
1693 }
1694}
1695
1696void native_play_dead(void)
1697{
1698 play_dead_common();
1699 tboot_shutdown(TB_SHUTDOWN_WFS);
1700
1701 mwait_play_dead(); /* Only returns on failure */
1702 if (cpuidle_play_dead())
1703 hlt_play_dead();
1704}
1705
1706#else /* ... !CONFIG_HOTPLUG_CPU */
1707int native_cpu_disable(void)
1708{
1709 return -ENOSYS;
1710}
1711
1712void native_cpu_die(unsigned int cpu)
1713{
1714 /* We said "no" in __cpu_disable */
1715 BUG();
1716}
1717
1718void native_play_dead(void)
1719{
1720 BUG();
1721}
1722
1723#endif