Loading...
1 /*
2 * x86 SMP booting functions
3 *
4 * (c) 1995 Alan Cox, Building #3 <alan@lxorguk.ukuu.org.uk>
5 * (c) 1998, 1999, 2000, 2009 Ingo Molnar <mingo@redhat.com>
6 * Copyright 2001 Andi Kleen, SuSE Labs.
7 *
8 * Much of the core SMP work is based on previous work by Thomas Radke, to
9 * whom a great many thanks are extended.
10 *
11 * Thanks to Intel for making available several different Pentium,
12 * Pentium Pro and Pentium-II/Xeon MP machines.
13 * Original development of Linux SMP code supported by Caldera.
14 *
15 * This code is released under the GNU General Public License version 2 or
16 * later.
17 *
18 * Fixes
19 * Felix Koop : NR_CPUS used properly
20 * Jose Renau : Handle single CPU case.
21 * Alan Cox : By repeated request 8) - Total BogoMIPS report.
22 * Greg Wright : Fix for kernel stacks panic.
23 * Erich Boleyn : MP v1.4 and additional changes.
24 * Matthias Sattler : Changes for 2.1 kernel map.
25 * Michel Lespinasse : Changes for 2.1 kernel map.
26 * Michael Chastain : Change trampoline.S to gnu as.
27 * Alan Cox : Dumb bug: 'B' step PPro's are fine
28 * Ingo Molnar : Added APIC timers, based on code
29 * from Jose Renau
30 * Ingo Molnar : various cleanups and rewrites
31 * Tigran Aivazian : fixed "0.00 in /proc/uptime on SMP" bug.
32 * Maciej W. Rozycki : Bits for genuine 82489DX APICs
33 * Andi Kleen : Changed for SMP boot into long mode.
34 * Martin J. Bligh : Added support for multi-quad systems
35 * Dave Jones : Report invalid combinations of Athlon CPUs.
36 * Rusty Russell : Hacked into shape for new "hotplug" boot process.
37 * Andi Kleen : Converted to new state machine.
38 * Ashok Raj : CPU hotplug support
39 * Glauber Costa : i386 and x86_64 integration
40 */
41
42#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
43
44#include <linux/init.h>
45#include <linux/smp.h>
46#include <linux/export.h>
47#include <linux/sched.h>
48#include <linux/sched/topology.h>
49#include <linux/sched/hotplug.h>
50#include <linux/sched/task_stack.h>
51#include <linux/percpu.h>
52#include <linux/bootmem.h>
53#include <linux/err.h>
54#include <linux/nmi.h>
55#include <linux/tboot.h>
56#include <linux/stackprotector.h>
57#include <linux/gfp.h>
58#include <linux/cpuidle.h>
59
60#include <asm/acpi.h>
61#include <asm/desc.h>
62#include <asm/nmi.h>
63#include <asm/irq.h>
64#include <asm/realmode.h>
65#include <asm/cpu.h>
66#include <asm/numa.h>
67#include <asm/pgtable.h>
68#include <asm/tlbflush.h>
69#include <asm/mtrr.h>
70#include <asm/mwait.h>
71#include <asm/apic.h>
72#include <asm/io_apic.h>
73#include <asm/fpu/internal.h>
74#include <asm/setup.h>
75#include <asm/uv/uv.h>
76#include <linux/mc146818rtc.h>
77#include <asm/i8259.h>
78#include <asm/misc.h>
79#include <asm/qspinlock.h>
80#include <asm/intel-family.h>
81#include <asm/cpu_device_id.h>
82#include <asm/spec-ctrl.h>
83
84/* Number of siblings per CPU package */
85int smp_num_siblings = 1;
86EXPORT_SYMBOL(smp_num_siblings);
87
88/* Last level cache ID of each logical CPU */
89DEFINE_PER_CPU_READ_MOSTLY(u16, cpu_llc_id) = BAD_APICID;
90
91/* representing HT siblings of each logical CPU */
92DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
93EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
94
95/* representing HT and core siblings of each logical CPU */
96DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
97EXPORT_PER_CPU_SYMBOL(cpu_core_map);
98
99DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_llc_shared_map);
100
101/* Per CPU bogomips and other parameters */
102DEFINE_PER_CPU_READ_MOSTLY(struct cpuinfo_x86, cpu_info);
103EXPORT_PER_CPU_SYMBOL(cpu_info);
104
105/* Logical package management. We might want to allocate that dynamically */
106unsigned int __max_logical_packages __read_mostly;
107EXPORT_SYMBOL(__max_logical_packages);
108static unsigned int logical_packages __read_mostly;
109
110/* Maximum number of SMT threads on any online core */
111int __read_mostly __max_smt_threads = 1;
112
113/* Flag to indicate if a complete sched domain rebuild is required */
114bool x86_topology_update;
115
116int arch_update_cpu_topology(void)
117{
118 int retval = x86_topology_update;
119
120 x86_topology_update = false;
121 return retval;
122}
123
124static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
125{
126 unsigned long flags;
127
128 spin_lock_irqsave(&rtc_lock, flags);
129 CMOS_WRITE(0xa, 0xf);
130 spin_unlock_irqrestore(&rtc_lock, flags);
131 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) =
132 start_eip >> 4;
133 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) =
134 start_eip & 0xf;
135}
136
137static inline void smpboot_restore_warm_reset_vector(void)
138{
139 unsigned long flags;
140
141 /*
142 * Paranoid: Set warm reset code and vector here back
143 * to default values.
144 */
145 spin_lock_irqsave(&rtc_lock, flags);
146 CMOS_WRITE(0, 0xf);
147 spin_unlock_irqrestore(&rtc_lock, flags);
148
149 *((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
150}
151
152/*
153 * Report back to the Boot Processor during boot time or to the caller processor
154 * during CPU online.
155 */
156static void smp_callin(void)
157{
158 int cpuid, phys_id;
159
160 /*
161 * If waken up by an INIT in an 82489DX configuration
162 * cpu_callout_mask guarantees we don't get here before
163 * an INIT_deassert IPI reaches our local APIC, so it is
164 * now safe to touch our local APIC.
165 */
166 cpuid = smp_processor_id();
167
168 /*
169 * (This works even if the APIC is not enabled.)
170 */
171 phys_id = read_apic_id();
172
173 /*
174 * the boot CPU has finished the init stage and is spinning
175 * on callin_map until we finish. We are free to set up this
176 * CPU, first the APIC. (this is probably redundant on most
177 * boards)
178 */
179 apic_ap_setup();
180
181 /*
182 * Save our processor parameters. Note: this information
183 * is needed for clock calibration.
184 */
185 smp_store_cpu_info(cpuid);
186
187 /*
188 * The topology information must be up to date before
189 * calibrate_delay() and notify_cpu_starting().
190 */
191 set_cpu_sibling_map(raw_smp_processor_id());
192
193 /*
194 * Get our bogomips.
195 * Update loops_per_jiffy in cpu_data. Previous call to
196 * smp_store_cpu_info() stored a value that is close but not as
197 * accurate as the value just calculated.
198 */
199 calibrate_delay();
200 cpu_data(cpuid).loops_per_jiffy = loops_per_jiffy;
201 pr_debug("Stack at about %p\n", &cpuid);
202
203 wmb();
204
205 notify_cpu_starting(cpuid);
206
207 /*
208 * Allow the master to continue.
209 */
210 cpumask_set_cpu(cpuid, cpu_callin_mask);
211}
212
213static int cpu0_logical_apicid;
214static int enable_start_cpu0;
215/*
216 * Activate a secondary processor.
217 */
218static void notrace start_secondary(void *unused)
219{
220 /*
221 * Don't put *anything* except direct CPU state initialization
222 * before cpu_init(), SMP booting is too fragile that we want to
223 * limit the things done here to the most necessary things.
224 */
225 if (boot_cpu_has(X86_FEATURE_PCID))
226 __write_cr4(__read_cr4() | X86_CR4_PCIDE);
227
228#ifdef CONFIG_X86_32
229 /* switch away from the initial page table */
230 load_cr3(swapper_pg_dir);
231 __flush_tlb_all();
232#endif
233 load_current_idt();
234 cpu_init();
235 x86_cpuinit.early_percpu_clock_init();
236 preempt_disable();
237 smp_callin();
238
239 enable_start_cpu0 = 0;
240
241 /* otherwise gcc will move up smp_processor_id before the cpu_init */
242 barrier();
243 /*
244 * Check TSC synchronization with the boot CPU:
245 */
246 check_tsc_sync_target();
247
248 speculative_store_bypass_ht_init();
249
250 /*
251 * Lock vector_lock, set CPU online and bring the vector
252 * allocator online. Online must be set with vector_lock held
253 * to prevent a concurrent irq setup/teardown from seeing a
254 * half valid vector space.
255 */
256 lock_vector_lock();
257 set_cpu_online(smp_processor_id(), true);
258 lapic_online();
259 unlock_vector_lock();
260 cpu_set_state_online(smp_processor_id());
261 x86_platform.nmi_init();
262
263 /* enable local interrupts */
264 local_irq_enable();
265
266 /* to prevent fake stack check failure in clock setup */
267 boot_init_stack_canary();
268
269 x86_cpuinit.setup_percpu_clockev();
270
271 wmb();
272 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
273}
274
275/**
276 * topology_phys_to_logical_pkg - Map a physical package id to a logical
277 *
278 * Returns logical package id or -1 if not found
279 */
280int topology_phys_to_logical_pkg(unsigned int phys_pkg)
281{
282 int cpu;
283
284 for_each_possible_cpu(cpu) {
285 struct cpuinfo_x86 *c = &cpu_data(cpu);
286
287 if (c->initialized && c->phys_proc_id == phys_pkg)
288 return c->logical_proc_id;
289 }
290 return -1;
291}
292EXPORT_SYMBOL(topology_phys_to_logical_pkg);
293
294/**
295 * topology_update_package_map - Update the physical to logical package map
296 * @pkg: The physical package id as retrieved via CPUID
297 * @cpu: The cpu for which this is updated
298 */
299int topology_update_package_map(unsigned int pkg, unsigned int cpu)
300{
301 int new;
302
303 /* Already available somewhere? */
304 new = topology_phys_to_logical_pkg(pkg);
305 if (new >= 0)
306 goto found;
307
308 new = logical_packages++;
309 if (new != pkg) {
310 pr_info("CPU %u Converting physical %u to logical package %u\n",
311 cpu, pkg, new);
312 }
313found:
314 cpu_data(cpu).logical_proc_id = new;
315 return 0;
316}
317
318void __init smp_store_boot_cpu_info(void)
319{
320 int id = 0; /* CPU 0 */
321 struct cpuinfo_x86 *c = &cpu_data(id);
322
323 *c = boot_cpu_data;
324 c->cpu_index = id;
325 topology_update_package_map(c->phys_proc_id, id);
326 c->initialized = true;
327}
328
329/*
330 * The bootstrap kernel entry code has set these up. Save them for
331 * a given CPU
332 */
333void smp_store_cpu_info(int id)
334{
335 struct cpuinfo_x86 *c = &cpu_data(id);
336
337 /* Copy boot_cpu_data only on the first bringup */
338 if (!c->initialized)
339 *c = boot_cpu_data;
340 c->cpu_index = id;
341 /*
342 * During boot time, CPU0 has this setup already. Save the info when
343 * bringing up AP or offlined CPU0.
344 */
345 identify_secondary_cpu(c);
346 c->initialized = true;
347}
348
349static bool
350topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
351{
352 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
353
354 return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
355}
356
357static bool
358topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
359{
360 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
361
362 return !WARN_ONCE(!topology_same_node(c, o),
363 "sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
364 "[node: %d != %d]. Ignoring dependency.\n",
365 cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
366}
367
368#define link_mask(mfunc, c1, c2) \
369do { \
370 cpumask_set_cpu((c1), mfunc(c2)); \
371 cpumask_set_cpu((c2), mfunc(c1)); \
372} while (0)
373
374static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
375{
376 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
377 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
378
379 if (c->phys_proc_id == o->phys_proc_id &&
380 per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2)) {
381 if (c->cpu_core_id == o->cpu_core_id)
382 return topology_sane(c, o, "smt");
383
384 if ((c->cu_id != 0xff) &&
385 (o->cu_id != 0xff) &&
386 (c->cu_id == o->cu_id))
387 return topology_sane(c, o, "smt");
388 }
389
390 } else if (c->phys_proc_id == o->phys_proc_id &&
391 c->cpu_core_id == o->cpu_core_id) {
392 return topology_sane(c, o, "smt");
393 }
394
395 return false;
396}
397
398/*
399 * Define snc_cpu[] for SNC (Sub-NUMA Cluster) CPUs.
400 *
401 * These are Intel CPUs that enumerate an LLC that is shared by
402 * multiple NUMA nodes. The LLC on these systems is shared for
403 * off-package data access but private to the NUMA node (half
404 * of the package) for on-package access.
405 *
406 * CPUID (the source of the information about the LLC) can only
407 * enumerate the cache as being shared *or* unshared, but not
408 * this particular configuration. The CPU in this case enumerates
409 * the cache to be shared across the entire package (spanning both
410 * NUMA nodes).
411 */
412
413static const struct x86_cpu_id snc_cpu[] = {
414 { X86_VENDOR_INTEL, 6, INTEL_FAM6_SKYLAKE_X },
415 {}
416};
417
418static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
419{
420 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
421
422 /* Do not match if we do not have a valid APICID for cpu: */
423 if (per_cpu(cpu_llc_id, cpu1) == BAD_APICID)
424 return false;
425
426 /* Do not match if LLC id does not match: */
427 if (per_cpu(cpu_llc_id, cpu1) != per_cpu(cpu_llc_id, cpu2))
428 return false;
429
430 /*
431 * Allow the SNC topology without warning. Return of false
432 * means 'c' does not share the LLC of 'o'. This will be
433 * reflected to userspace.
434 */
435 if (!topology_same_node(c, o) && x86_match_cpu(snc_cpu))
436 return false;
437
438 return topology_sane(c, o, "llc");
439}
440
441/*
442 * Unlike the other levels, we do not enforce keeping a
443 * multicore group inside a NUMA node. If this happens, we will
444 * discard the MC level of the topology later.
445 */
446static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
447{
448 if (c->phys_proc_id == o->phys_proc_id)
449 return true;
450 return false;
451}
452
453#if defined(CONFIG_SCHED_SMT) || defined(CONFIG_SCHED_MC)
454static inline int x86_sched_itmt_flags(void)
455{
456 return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
457}
458
459#ifdef CONFIG_SCHED_MC
460static int x86_core_flags(void)
461{
462 return cpu_core_flags() | x86_sched_itmt_flags();
463}
464#endif
465#ifdef CONFIG_SCHED_SMT
466static int x86_smt_flags(void)
467{
468 return cpu_smt_flags() | x86_sched_itmt_flags();
469}
470#endif
471#endif
472
473static struct sched_domain_topology_level x86_numa_in_package_topology[] = {
474#ifdef CONFIG_SCHED_SMT
475 { cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
476#endif
477#ifdef CONFIG_SCHED_MC
478 { cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
479#endif
480 { NULL, },
481};
482
483static struct sched_domain_topology_level x86_topology[] = {
484#ifdef CONFIG_SCHED_SMT
485 { cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
486#endif
487#ifdef CONFIG_SCHED_MC
488 { cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
489#endif
490 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
491 { NULL, },
492};
493
494/*
495 * Set if a package/die has multiple NUMA nodes inside.
496 * AMD Magny-Cours, Intel Cluster-on-Die, and Intel
497 * Sub-NUMA Clustering have this.
498 */
499static bool x86_has_numa_in_package;
500
501void set_cpu_sibling_map(int cpu)
502{
503 bool has_smt = smp_num_siblings > 1;
504 bool has_mp = has_smt || boot_cpu_data.x86_max_cores > 1;
505 struct cpuinfo_x86 *c = &cpu_data(cpu);
506 struct cpuinfo_x86 *o;
507 int i, threads;
508
509 cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
510
511 if (!has_mp) {
512 cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
513 cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
514 cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
515 c->booted_cores = 1;
516 return;
517 }
518
519 for_each_cpu(i, cpu_sibling_setup_mask) {
520 o = &cpu_data(i);
521
522 if ((i == cpu) || (has_smt && match_smt(c, o)))
523 link_mask(topology_sibling_cpumask, cpu, i);
524
525 if ((i == cpu) || (has_mp && match_llc(c, o)))
526 link_mask(cpu_llc_shared_mask, cpu, i);
527
528 }
529
530 /*
531 * This needs a separate iteration over the cpus because we rely on all
532 * topology_sibling_cpumask links to be set-up.
533 */
534 for_each_cpu(i, cpu_sibling_setup_mask) {
535 o = &cpu_data(i);
536
537 if ((i == cpu) || (has_mp && match_die(c, o))) {
538 link_mask(topology_core_cpumask, cpu, i);
539
540 /*
541 * Does this new cpu bringup a new core?
542 */
543 if (cpumask_weight(
544 topology_sibling_cpumask(cpu)) == 1) {
545 /*
546 * for each core in package, increment
547 * the booted_cores for this new cpu
548 */
549 if (cpumask_first(
550 topology_sibling_cpumask(i)) == i)
551 c->booted_cores++;
552 /*
553 * increment the core count for all
554 * the other cpus in this package
555 */
556 if (i != cpu)
557 cpu_data(i).booted_cores++;
558 } else if (i != cpu && !c->booted_cores)
559 c->booted_cores = cpu_data(i).booted_cores;
560 }
561 if (match_die(c, o) && !topology_same_node(c, o))
562 x86_has_numa_in_package = true;
563 }
564
565 threads = cpumask_weight(topology_sibling_cpumask(cpu));
566 if (threads > __max_smt_threads)
567 __max_smt_threads = threads;
568}
569
570/* maps the cpu to the sched domain representing multi-core */
571const struct cpumask *cpu_coregroup_mask(int cpu)
572{
573 return cpu_llc_shared_mask(cpu);
574}
575
576static void impress_friends(void)
577{
578 int cpu;
579 unsigned long bogosum = 0;
580 /*
581 * Allow the user to impress friends.
582 */
583 pr_debug("Before bogomips\n");
584 for_each_possible_cpu(cpu)
585 if (cpumask_test_cpu(cpu, cpu_callout_mask))
586 bogosum += cpu_data(cpu).loops_per_jiffy;
587 pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
588 num_online_cpus(),
589 bogosum/(500000/HZ),
590 (bogosum/(5000/HZ))%100);
591
592 pr_debug("Before bogocount - setting activated=1\n");
593}
594
595void __inquire_remote_apic(int apicid)
596{
597 unsigned i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
598 const char * const names[] = { "ID", "VERSION", "SPIV" };
599 int timeout;
600 u32 status;
601
602 pr_info("Inquiring remote APIC 0x%x...\n", apicid);
603
604 for (i = 0; i < ARRAY_SIZE(regs); i++) {
605 pr_info("... APIC 0x%x %s: ", apicid, names[i]);
606
607 /*
608 * Wait for idle.
609 */
610 status = safe_apic_wait_icr_idle();
611 if (status)
612 pr_cont("a previous APIC delivery may have failed\n");
613
614 apic_icr_write(APIC_DM_REMRD | regs[i], apicid);
615
616 timeout = 0;
617 do {
618 udelay(100);
619 status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
620 } while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
621
622 switch (status) {
623 case APIC_ICR_RR_VALID:
624 status = apic_read(APIC_RRR);
625 pr_cont("%08x\n", status);
626 break;
627 default:
628 pr_cont("failed\n");
629 }
630 }
631}
632
633/*
634 * The Multiprocessor Specification 1.4 (1997) example code suggests
635 * that there should be a 10ms delay between the BSP asserting INIT
636 * and de-asserting INIT, when starting a remote processor.
637 * But that slows boot and resume on modern processors, which include
638 * many cores and don't require that delay.
639 *
640 * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
641 * Modern processor families are quirked to remove the delay entirely.
642 */
643#define UDELAY_10MS_DEFAULT 10000
644
645static unsigned int init_udelay = UINT_MAX;
646
647static int __init cpu_init_udelay(char *str)
648{
649 get_option(&str, &init_udelay);
650
651 return 0;
652}
653early_param("cpu_init_udelay", cpu_init_udelay);
654
655static void __init smp_quirk_init_udelay(void)
656{
657 /* if cmdline changed it from default, leave it alone */
658 if (init_udelay != UINT_MAX)
659 return;
660
661 /* if modern processor, use no delay */
662 if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
663 ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
664 init_udelay = 0;
665 return;
666 }
667 /* else, use legacy delay */
668 init_udelay = UDELAY_10MS_DEFAULT;
669}
670
671/*
672 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
673 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
674 * won't ... remember to clear down the APIC, etc later.
675 */
676int
677wakeup_secondary_cpu_via_nmi(int apicid, unsigned long start_eip)
678{
679 unsigned long send_status, accept_status = 0;
680 int maxlvt;
681
682 /* Target chip */
683 /* Boot on the stack */
684 /* Kick the second */
685 apic_icr_write(APIC_DM_NMI | apic->dest_logical, apicid);
686
687 pr_debug("Waiting for send to finish...\n");
688 send_status = safe_apic_wait_icr_idle();
689
690 /*
691 * Give the other CPU some time to accept the IPI.
692 */
693 udelay(200);
694 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
695 maxlvt = lapic_get_maxlvt();
696 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
697 apic_write(APIC_ESR, 0);
698 accept_status = (apic_read(APIC_ESR) & 0xEF);
699 }
700 pr_debug("NMI sent\n");
701
702 if (send_status)
703 pr_err("APIC never delivered???\n");
704 if (accept_status)
705 pr_err("APIC delivery error (%lx)\n", accept_status);
706
707 return (send_status | accept_status);
708}
709
710static int
711wakeup_secondary_cpu_via_init(int phys_apicid, unsigned long start_eip)
712{
713 unsigned long send_status = 0, accept_status = 0;
714 int maxlvt, num_starts, j;
715
716 maxlvt = lapic_get_maxlvt();
717
718 /*
719 * Be paranoid about clearing APIC errors.
720 */
721 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
722 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
723 apic_write(APIC_ESR, 0);
724 apic_read(APIC_ESR);
725 }
726
727 pr_debug("Asserting INIT\n");
728
729 /*
730 * Turn INIT on target chip
731 */
732 /*
733 * Send IPI
734 */
735 apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT,
736 phys_apicid);
737
738 pr_debug("Waiting for send to finish...\n");
739 send_status = safe_apic_wait_icr_idle();
740
741 udelay(init_udelay);
742
743 pr_debug("Deasserting INIT\n");
744
745 /* Target chip */
746 /* Send IPI */
747 apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
748
749 pr_debug("Waiting for send to finish...\n");
750 send_status = safe_apic_wait_icr_idle();
751
752 mb();
753
754 /*
755 * Should we send STARTUP IPIs ?
756 *
757 * Determine this based on the APIC version.
758 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
759 */
760 if (APIC_INTEGRATED(boot_cpu_apic_version))
761 num_starts = 2;
762 else
763 num_starts = 0;
764
765 /*
766 * Run STARTUP IPI loop.
767 */
768 pr_debug("#startup loops: %d\n", num_starts);
769
770 for (j = 1; j <= num_starts; j++) {
771 pr_debug("Sending STARTUP #%d\n", j);
772 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
773 apic_write(APIC_ESR, 0);
774 apic_read(APIC_ESR);
775 pr_debug("After apic_write\n");
776
777 /*
778 * STARTUP IPI
779 */
780
781 /* Target chip */
782 /* Boot on the stack */
783 /* Kick the second */
784 apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
785 phys_apicid);
786
787 /*
788 * Give the other CPU some time to accept the IPI.
789 */
790 if (init_udelay == 0)
791 udelay(10);
792 else
793 udelay(300);
794
795 pr_debug("Startup point 1\n");
796
797 pr_debug("Waiting for send to finish...\n");
798 send_status = safe_apic_wait_icr_idle();
799
800 /*
801 * Give the other CPU some time to accept the IPI.
802 */
803 if (init_udelay == 0)
804 udelay(10);
805 else
806 udelay(200);
807
808 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
809 apic_write(APIC_ESR, 0);
810 accept_status = (apic_read(APIC_ESR) & 0xEF);
811 if (send_status || accept_status)
812 break;
813 }
814 pr_debug("After Startup\n");
815
816 if (send_status)
817 pr_err("APIC never delivered???\n");
818 if (accept_status)
819 pr_err("APIC delivery error (%lx)\n", accept_status);
820
821 return (send_status | accept_status);
822}
823
824/* reduce the number of lines printed when booting a large cpu count system */
825static void announce_cpu(int cpu, int apicid)
826{
827 static int current_node = -1;
828 int node = early_cpu_to_node(cpu);
829 static int width, node_width;
830
831 if (!width)
832 width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
833
834 if (!node_width)
835 node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
836
837 if (cpu == 1)
838 printk(KERN_INFO "x86: Booting SMP configuration:\n");
839
840 if (system_state < SYSTEM_RUNNING) {
841 if (node != current_node) {
842 if (current_node > (-1))
843 pr_cont("\n");
844 current_node = node;
845
846 printk(KERN_INFO ".... node %*s#%d, CPUs: ",
847 node_width - num_digits(node), " ", node);
848 }
849
850 /* Add padding for the BSP */
851 if (cpu == 1)
852 pr_cont("%*s", width + 1, " ");
853
854 pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
855
856 } else
857 pr_info("Booting Node %d Processor %d APIC 0x%x\n",
858 node, cpu, apicid);
859}
860
861static int wakeup_cpu0_nmi(unsigned int cmd, struct pt_regs *regs)
862{
863 int cpu;
864
865 cpu = smp_processor_id();
866 if (cpu == 0 && !cpu_online(cpu) && enable_start_cpu0)
867 return NMI_HANDLED;
868
869 return NMI_DONE;
870}
871
872/*
873 * Wake up AP by INIT, INIT, STARTUP sequence.
874 *
875 * Instead of waiting for STARTUP after INITs, BSP will execute the BIOS
876 * boot-strap code which is not a desired behavior for waking up BSP. To
877 * void the boot-strap code, wake up CPU0 by NMI instead.
878 *
879 * This works to wake up soft offlined CPU0 only. If CPU0 is hard offlined
880 * (i.e. physically hot removed and then hot added), NMI won't wake it up.
881 * We'll change this code in the future to wake up hard offlined CPU0 if
882 * real platform and request are available.
883 */
884static int
885wakeup_cpu_via_init_nmi(int cpu, unsigned long start_ip, int apicid,
886 int *cpu0_nmi_registered)
887{
888 int id;
889 int boot_error;
890
891 preempt_disable();
892
893 /*
894 * Wake up AP by INIT, INIT, STARTUP sequence.
895 */
896 if (cpu) {
897 boot_error = wakeup_secondary_cpu_via_init(apicid, start_ip);
898 goto out;
899 }
900
901 /*
902 * Wake up BSP by nmi.
903 *
904 * Register a NMI handler to help wake up CPU0.
905 */
906 boot_error = register_nmi_handler(NMI_LOCAL,
907 wakeup_cpu0_nmi, 0, "wake_cpu0");
908
909 if (!boot_error) {
910 enable_start_cpu0 = 1;
911 *cpu0_nmi_registered = 1;
912 if (apic->dest_logical == APIC_DEST_LOGICAL)
913 id = cpu0_logical_apicid;
914 else
915 id = apicid;
916 boot_error = wakeup_secondary_cpu_via_nmi(id, start_ip);
917 }
918
919out:
920 preempt_enable();
921
922 return boot_error;
923}
924
925void common_cpu_up(unsigned int cpu, struct task_struct *idle)
926{
927 /* Just in case we booted with a single CPU. */
928 alternatives_enable_smp();
929
930 per_cpu(current_task, cpu) = idle;
931
932#ifdef CONFIG_X86_32
933 /* Stack for startup_32 can be just as for start_secondary onwards */
934 irq_ctx_init(cpu);
935 per_cpu(cpu_current_top_of_stack, cpu) = task_top_of_stack(idle);
936#else
937 initial_gs = per_cpu_offset(cpu);
938#endif
939}
940
941/*
942 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
943 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
944 * Returns zero if CPU booted OK, else error code from
945 * ->wakeup_secondary_cpu.
946 */
947static int do_boot_cpu(int apicid, int cpu, struct task_struct *idle,
948 int *cpu0_nmi_registered)
949{
950 volatile u32 *trampoline_status =
951 (volatile u32 *) __va(real_mode_header->trampoline_status);
952 /* start_ip had better be page-aligned! */
953 unsigned long start_ip = real_mode_header->trampoline_start;
954
955 unsigned long boot_error = 0;
956 unsigned long timeout;
957
958 idle->thread.sp = (unsigned long)task_pt_regs(idle);
959 early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
960 initial_code = (unsigned long)start_secondary;
961 initial_stack = idle->thread.sp;
962
963 /* Enable the espfix hack for this CPU */
964 init_espfix_ap(cpu);
965
966 /* So we see what's up */
967 announce_cpu(cpu, apicid);
968
969 /*
970 * This grunge runs the startup process for
971 * the targeted processor.
972 */
973
974 if (x86_platform.legacy.warm_reset) {
975
976 pr_debug("Setting warm reset code and vector.\n");
977
978 smpboot_setup_warm_reset_vector(start_ip);
979 /*
980 * Be paranoid about clearing APIC errors.
981 */
982 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
983 apic_write(APIC_ESR, 0);
984 apic_read(APIC_ESR);
985 }
986 }
987
988 /*
989 * AP might wait on cpu_callout_mask in cpu_init() with
990 * cpu_initialized_mask set if previous attempt to online
991 * it timed-out. Clear cpu_initialized_mask so that after
992 * INIT/SIPI it could start with a clean state.
993 */
994 cpumask_clear_cpu(cpu, cpu_initialized_mask);
995 smp_mb();
996
997 /*
998 * Wake up a CPU in difference cases:
999 * - Use the method in the APIC driver if it's defined
1000 * Otherwise,
1001 * - Use an INIT boot APIC message for APs or NMI for BSP.
1002 */
1003 if (apic->wakeup_secondary_cpu)
1004 boot_error = apic->wakeup_secondary_cpu(apicid, start_ip);
1005 else
1006 boot_error = wakeup_cpu_via_init_nmi(cpu, start_ip, apicid,
1007 cpu0_nmi_registered);
1008
1009 if (!boot_error) {
1010 /*
1011 * Wait 10s total for first sign of life from AP
1012 */
1013 boot_error = -1;
1014 timeout = jiffies + 10*HZ;
1015 while (time_before(jiffies, timeout)) {
1016 if (cpumask_test_cpu(cpu, cpu_initialized_mask)) {
1017 /*
1018 * Tell AP to proceed with initialization
1019 */
1020 cpumask_set_cpu(cpu, cpu_callout_mask);
1021 boot_error = 0;
1022 break;
1023 }
1024 schedule();
1025 }
1026 }
1027
1028 if (!boot_error) {
1029 /*
1030 * Wait till AP completes initial initialization
1031 */
1032 while (!cpumask_test_cpu(cpu, cpu_callin_mask)) {
1033 /*
1034 * Allow other tasks to run while we wait for the
1035 * AP to come online. This also gives a chance
1036 * for the MTRR work(triggered by the AP coming online)
1037 * to be completed in the stop machine context.
1038 */
1039 schedule();
1040 }
1041 }
1042
1043 /* mark "stuck" area as not stuck */
1044 *trampoline_status = 0;
1045
1046 if (x86_platform.legacy.warm_reset) {
1047 /*
1048 * Cleanup possible dangling ends...
1049 */
1050 smpboot_restore_warm_reset_vector();
1051 }
1052
1053 return boot_error;
1054}
1055
1056int native_cpu_up(unsigned int cpu, struct task_struct *tidle)
1057{
1058 int apicid = apic->cpu_present_to_apicid(cpu);
1059 int cpu0_nmi_registered = 0;
1060 unsigned long flags;
1061 int err, ret = 0;
1062
1063 lockdep_assert_irqs_enabled();
1064
1065 pr_debug("++++++++++++++++++++=_---CPU UP %u\n", cpu);
1066
1067 if (apicid == BAD_APICID ||
1068 !physid_isset(apicid, phys_cpu_present_map) ||
1069 !apic->apic_id_valid(apicid)) {
1070 pr_err("%s: bad cpu %d\n", __func__, cpu);
1071 return -EINVAL;
1072 }
1073
1074 /*
1075 * Already booted CPU?
1076 */
1077 if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
1078 pr_debug("do_boot_cpu %d Already started\n", cpu);
1079 return -ENOSYS;
1080 }
1081
1082 /*
1083 * Save current MTRR state in case it was changed since early boot
1084 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1085 */
1086 mtrr_save_state();
1087
1088 /* x86 CPUs take themselves offline, so delayed offline is OK. */
1089 err = cpu_check_up_prepare(cpu);
1090 if (err && err != -EBUSY)
1091 return err;
1092
1093 /* the FPU context is blank, nobody can own it */
1094 per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
1095
1096 common_cpu_up(cpu, tidle);
1097
1098 err = do_boot_cpu(apicid, cpu, tidle, &cpu0_nmi_registered);
1099 if (err) {
1100 pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
1101 ret = -EIO;
1102 goto unreg_nmi;
1103 }
1104
1105 /*
1106 * Check TSC synchronization with the AP (keep irqs disabled
1107 * while doing so):
1108 */
1109 local_irq_save(flags);
1110 check_tsc_sync_source(cpu);
1111 local_irq_restore(flags);
1112
1113 while (!cpu_online(cpu)) {
1114 cpu_relax();
1115 touch_nmi_watchdog();
1116 }
1117
1118unreg_nmi:
1119 /*
1120 * Clean up the nmi handler. Do this after the callin and callout sync
1121 * to avoid impact of possible long unregister time.
1122 */
1123 if (cpu0_nmi_registered)
1124 unregister_nmi_handler(NMI_LOCAL, "wake_cpu0");
1125
1126 return ret;
1127}
1128
1129/**
1130 * arch_disable_smp_support() - disables SMP support for x86 at runtime
1131 */
1132void arch_disable_smp_support(void)
1133{
1134 disable_ioapic_support();
1135}
1136
1137/*
1138 * Fall back to non SMP mode after errors.
1139 *
1140 * RED-PEN audit/test this more. I bet there is more state messed up here.
1141 */
1142static __init void disable_smp(void)
1143{
1144 pr_info("SMP disabled\n");
1145
1146 disable_ioapic_support();
1147
1148 init_cpu_present(cpumask_of(0));
1149 init_cpu_possible(cpumask_of(0));
1150
1151 if (smp_found_config)
1152 physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
1153 else
1154 physid_set_mask_of_physid(0, &phys_cpu_present_map);
1155 cpumask_set_cpu(0, topology_sibling_cpumask(0));
1156 cpumask_set_cpu(0, topology_core_cpumask(0));
1157}
1158
1159/*
1160 * Various sanity checks.
1161 */
1162static void __init smp_sanity_check(void)
1163{
1164 preempt_disable();
1165
1166#if !defined(CONFIG_X86_BIGSMP) && defined(CONFIG_X86_32)
1167 if (def_to_bigsmp && nr_cpu_ids > 8) {
1168 unsigned int cpu;
1169 unsigned nr;
1170
1171 pr_warn("More than 8 CPUs detected - skipping them\n"
1172 "Use CONFIG_X86_BIGSMP\n");
1173
1174 nr = 0;
1175 for_each_present_cpu(cpu) {
1176 if (nr >= 8)
1177 set_cpu_present(cpu, false);
1178 nr++;
1179 }
1180
1181 nr = 0;
1182 for_each_possible_cpu(cpu) {
1183 if (nr >= 8)
1184 set_cpu_possible(cpu, false);
1185 nr++;
1186 }
1187
1188 nr_cpu_ids = 8;
1189 }
1190#endif
1191
1192 if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map)) {
1193 pr_warn("weird, boot CPU (#%d) not listed by the BIOS\n",
1194 hard_smp_processor_id());
1195
1196 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1197 }
1198
1199 /*
1200 * Should not be necessary because the MP table should list the boot
1201 * CPU too, but we do it for the sake of robustness anyway.
1202 */
1203 if (!apic->check_phys_apicid_present(boot_cpu_physical_apicid)) {
1204 pr_notice("weird, boot CPU (#%d) not listed by the BIOS\n",
1205 boot_cpu_physical_apicid);
1206 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1207 }
1208 preempt_enable();
1209}
1210
1211static void __init smp_cpu_index_default(void)
1212{
1213 int i;
1214 struct cpuinfo_x86 *c;
1215
1216 for_each_possible_cpu(i) {
1217 c = &cpu_data(i);
1218 /* mark all to hotplug */
1219 c->cpu_index = nr_cpu_ids;
1220 }
1221}
1222
1223static void __init smp_get_logical_apicid(void)
1224{
1225 if (x2apic_mode)
1226 cpu0_logical_apicid = apic_read(APIC_LDR);
1227 else
1228 cpu0_logical_apicid = GET_APIC_LOGICAL_ID(apic_read(APIC_LDR));
1229}
1230
1231/*
1232 * Prepare for SMP bootup.
1233 * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1234 * for common interface support.
1235 */
1236void __init native_smp_prepare_cpus(unsigned int max_cpus)
1237{
1238 unsigned int i;
1239
1240 smp_cpu_index_default();
1241
1242 /*
1243 * Setup boot CPU information
1244 */
1245 smp_store_boot_cpu_info(); /* Final full version of the data */
1246 cpumask_copy(cpu_callin_mask, cpumask_of(0));
1247 mb();
1248
1249 for_each_possible_cpu(i) {
1250 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1251 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1252 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1253 }
1254
1255 /*
1256 * Set 'default' x86 topology, this matches default_topology() in that
1257 * it has NUMA nodes as a topology level. See also
1258 * native_smp_cpus_done().
1259 *
1260 * Must be done before set_cpus_sibling_map() is ran.
1261 */
1262 set_sched_topology(x86_topology);
1263
1264 set_cpu_sibling_map(0);
1265
1266 smp_sanity_check();
1267
1268 switch (apic_intr_mode) {
1269 case APIC_PIC:
1270 case APIC_VIRTUAL_WIRE_NO_CONFIG:
1271 disable_smp();
1272 return;
1273 case APIC_SYMMETRIC_IO_NO_ROUTING:
1274 disable_smp();
1275 /* Setup local timer */
1276 x86_init.timers.setup_percpu_clockev();
1277 return;
1278 case APIC_VIRTUAL_WIRE:
1279 case APIC_SYMMETRIC_IO:
1280 break;
1281 }
1282
1283 /* Setup local timer */
1284 x86_init.timers.setup_percpu_clockev();
1285
1286 smp_get_logical_apicid();
1287
1288 pr_info("CPU0: ");
1289 print_cpu_info(&cpu_data(0));
1290
1291 native_pv_lock_init();
1292
1293 uv_system_init();
1294
1295 set_mtrr_aps_delayed_init();
1296
1297 smp_quirk_init_udelay();
1298
1299 speculative_store_bypass_ht_init();
1300}
1301
1302void arch_enable_nonboot_cpus_begin(void)
1303{
1304 set_mtrr_aps_delayed_init();
1305}
1306
1307void arch_enable_nonboot_cpus_end(void)
1308{
1309 mtrr_aps_init();
1310}
1311
1312/*
1313 * Early setup to make printk work.
1314 */
1315void __init native_smp_prepare_boot_cpu(void)
1316{
1317 int me = smp_processor_id();
1318 switch_to_new_gdt(me);
1319 /* already set me in cpu_online_mask in boot_cpu_init() */
1320 cpumask_set_cpu(me, cpu_callout_mask);
1321 cpu_set_state_online(me);
1322}
1323
1324void __init calculate_max_logical_packages(void)
1325{
1326 int ncpus;
1327
1328 /*
1329 * Today neither Intel nor AMD support heterogenous systems so
1330 * extrapolate the boot cpu's data to all packages.
1331 */
1332 ncpus = cpu_data(0).booted_cores * topology_max_smt_threads();
1333 __max_logical_packages = DIV_ROUND_UP(nr_cpu_ids, ncpus);
1334 pr_info("Max logical packages: %u\n", __max_logical_packages);
1335}
1336
1337void __init native_smp_cpus_done(unsigned int max_cpus)
1338{
1339 pr_debug("Boot done\n");
1340
1341 calculate_max_logical_packages();
1342
1343 if (x86_has_numa_in_package)
1344 set_sched_topology(x86_numa_in_package_topology);
1345
1346 nmi_selftest();
1347 impress_friends();
1348 mtrr_aps_init();
1349}
1350
1351static int __initdata setup_possible_cpus = -1;
1352static int __init _setup_possible_cpus(char *str)
1353{
1354 get_option(&str, &setup_possible_cpus);
1355 return 0;
1356}
1357early_param("possible_cpus", _setup_possible_cpus);
1358
1359
1360/*
1361 * cpu_possible_mask should be static, it cannot change as cpu's
1362 * are onlined, or offlined. The reason is per-cpu data-structures
1363 * are allocated by some modules at init time, and dont expect to
1364 * do this dynamically on cpu arrival/departure.
1365 * cpu_present_mask on the other hand can change dynamically.
1366 * In case when cpu_hotplug is not compiled, then we resort to current
1367 * behaviour, which is cpu_possible == cpu_present.
1368 * - Ashok Raj
1369 *
1370 * Three ways to find out the number of additional hotplug CPUs:
1371 * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1372 * - The user can overwrite it with possible_cpus=NUM
1373 * - Otherwise don't reserve additional CPUs.
1374 * We do this because additional CPUs waste a lot of memory.
1375 * -AK
1376 */
1377__init void prefill_possible_map(void)
1378{
1379 int i, possible;
1380
1381 /* No boot processor was found in mptable or ACPI MADT */
1382 if (!num_processors) {
1383 if (boot_cpu_has(X86_FEATURE_APIC)) {
1384 int apicid = boot_cpu_physical_apicid;
1385 int cpu = hard_smp_processor_id();
1386
1387 pr_warn("Boot CPU (id %d) not listed by BIOS\n", cpu);
1388
1389 /* Make sure boot cpu is enumerated */
1390 if (apic->cpu_present_to_apicid(0) == BAD_APICID &&
1391 apic->apic_id_valid(apicid))
1392 generic_processor_info(apicid, boot_cpu_apic_version);
1393 }
1394
1395 if (!num_processors)
1396 num_processors = 1;
1397 }
1398
1399 i = setup_max_cpus ?: 1;
1400 if (setup_possible_cpus == -1) {
1401 possible = num_processors;
1402#ifdef CONFIG_HOTPLUG_CPU
1403 if (setup_max_cpus)
1404 possible += disabled_cpus;
1405#else
1406 if (possible > i)
1407 possible = i;
1408#endif
1409 } else
1410 possible = setup_possible_cpus;
1411
1412 total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1413
1414 /* nr_cpu_ids could be reduced via nr_cpus= */
1415 if (possible > nr_cpu_ids) {
1416 pr_warn("%d Processors exceeds NR_CPUS limit of %u\n",
1417 possible, nr_cpu_ids);
1418 possible = nr_cpu_ids;
1419 }
1420
1421#ifdef CONFIG_HOTPLUG_CPU
1422 if (!setup_max_cpus)
1423#endif
1424 if (possible > i) {
1425 pr_warn("%d Processors exceeds max_cpus limit of %u\n",
1426 possible, setup_max_cpus);
1427 possible = i;
1428 }
1429
1430 nr_cpu_ids = possible;
1431
1432 pr_info("Allowing %d CPUs, %d hotplug CPUs\n",
1433 possible, max_t(int, possible - num_processors, 0));
1434
1435 reset_cpu_possible_mask();
1436
1437 for (i = 0; i < possible; i++)
1438 set_cpu_possible(i, true);
1439}
1440
1441#ifdef CONFIG_HOTPLUG_CPU
1442
1443/* Recompute SMT state for all CPUs on offline */
1444static void recompute_smt_state(void)
1445{
1446 int max_threads, cpu;
1447
1448 max_threads = 0;
1449 for_each_online_cpu (cpu) {
1450 int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1451
1452 if (threads > max_threads)
1453 max_threads = threads;
1454 }
1455 __max_smt_threads = max_threads;
1456}
1457
1458static void remove_siblinginfo(int cpu)
1459{
1460 int sibling;
1461 struct cpuinfo_x86 *c = &cpu_data(cpu);
1462
1463 for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1464 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1465 /*/
1466 * last thread sibling in this cpu core going down
1467 */
1468 if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1469 cpu_data(sibling).booted_cores--;
1470 }
1471
1472 for_each_cpu(sibling, topology_sibling_cpumask(cpu))
1473 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1474 for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1475 cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1476 cpumask_clear(cpu_llc_shared_mask(cpu));
1477 cpumask_clear(topology_sibling_cpumask(cpu));
1478 cpumask_clear(topology_core_cpumask(cpu));
1479 c->cpu_core_id = 0;
1480 c->booted_cores = 0;
1481 cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1482 recompute_smt_state();
1483}
1484
1485static void remove_cpu_from_maps(int cpu)
1486{
1487 set_cpu_online(cpu, false);
1488 cpumask_clear_cpu(cpu, cpu_callout_mask);
1489 cpumask_clear_cpu(cpu, cpu_callin_mask);
1490 /* was set by cpu_init() */
1491 cpumask_clear_cpu(cpu, cpu_initialized_mask);
1492 numa_remove_cpu(cpu);
1493}
1494
1495void cpu_disable_common(void)
1496{
1497 int cpu = smp_processor_id();
1498
1499 remove_siblinginfo(cpu);
1500
1501 /* It's now safe to remove this processor from the online map */
1502 lock_vector_lock();
1503 remove_cpu_from_maps(cpu);
1504 unlock_vector_lock();
1505 fixup_irqs();
1506 lapic_offline();
1507}
1508
1509int native_cpu_disable(void)
1510{
1511 int ret;
1512
1513 ret = lapic_can_unplug_cpu();
1514 if (ret)
1515 return ret;
1516
1517 clear_local_APIC();
1518 cpu_disable_common();
1519
1520 return 0;
1521}
1522
1523int common_cpu_die(unsigned int cpu)
1524{
1525 int ret = 0;
1526
1527 /* We don't do anything here: idle task is faking death itself. */
1528
1529 /* They ack this in play_dead() by setting CPU_DEAD */
1530 if (cpu_wait_death(cpu, 5)) {
1531 if (system_state == SYSTEM_RUNNING)
1532 pr_info("CPU %u is now offline\n", cpu);
1533 } else {
1534 pr_err("CPU %u didn't die...\n", cpu);
1535 ret = -1;
1536 }
1537
1538 return ret;
1539}
1540
1541void native_cpu_die(unsigned int cpu)
1542{
1543 common_cpu_die(cpu);
1544}
1545
1546void play_dead_common(void)
1547{
1548 idle_task_exit();
1549
1550 /* Ack it */
1551 (void)cpu_report_death();
1552
1553 /*
1554 * With physical CPU hotplug, we should halt the cpu
1555 */
1556 local_irq_disable();
1557}
1558
1559static bool wakeup_cpu0(void)
1560{
1561 if (smp_processor_id() == 0 && enable_start_cpu0)
1562 return true;
1563
1564 return false;
1565}
1566
1567/*
1568 * We need to flush the caches before going to sleep, lest we have
1569 * dirty data in our caches when we come back up.
1570 */
1571static inline void mwait_play_dead(void)
1572{
1573 unsigned int eax, ebx, ecx, edx;
1574 unsigned int highest_cstate = 0;
1575 unsigned int highest_subcstate = 0;
1576 void *mwait_ptr;
1577 int i;
1578
1579 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
1580 return;
1581 if (!this_cpu_has(X86_FEATURE_MWAIT))
1582 return;
1583 if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1584 return;
1585 if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1586 return;
1587
1588 eax = CPUID_MWAIT_LEAF;
1589 ecx = 0;
1590 native_cpuid(&eax, &ebx, &ecx, &edx);
1591
1592 /*
1593 * eax will be 0 if EDX enumeration is not valid.
1594 * Initialized below to cstate, sub_cstate value when EDX is valid.
1595 */
1596 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1597 eax = 0;
1598 } else {
1599 edx >>= MWAIT_SUBSTATE_SIZE;
1600 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1601 if (edx & MWAIT_SUBSTATE_MASK) {
1602 highest_cstate = i;
1603 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1604 }
1605 }
1606 eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1607 (highest_subcstate - 1);
1608 }
1609
1610 /*
1611 * This should be a memory location in a cache line which is
1612 * unlikely to be touched by other processors. The actual
1613 * content is immaterial as it is not actually modified in any way.
1614 */
1615 mwait_ptr = ¤t_thread_info()->flags;
1616
1617 wbinvd();
1618
1619 while (1) {
1620 /*
1621 * The CLFLUSH is a workaround for erratum AAI65 for
1622 * the Xeon 7400 series. It's not clear it is actually
1623 * needed, but it should be harmless in either case.
1624 * The WBINVD is insufficient due to the spurious-wakeup
1625 * case where we return around the loop.
1626 */
1627 mb();
1628 clflush(mwait_ptr);
1629 mb();
1630 __monitor(mwait_ptr, 0, 0);
1631 mb();
1632 __mwait(eax, 0);
1633 /*
1634 * If NMI wants to wake up CPU0, start CPU0.
1635 */
1636 if (wakeup_cpu0())
1637 start_cpu0();
1638 }
1639}
1640
1641void hlt_play_dead(void)
1642{
1643 if (__this_cpu_read(cpu_info.x86) >= 4)
1644 wbinvd();
1645
1646 while (1) {
1647 native_halt();
1648 /*
1649 * If NMI wants to wake up CPU0, start CPU0.
1650 */
1651 if (wakeup_cpu0())
1652 start_cpu0();
1653 }
1654}
1655
1656void native_play_dead(void)
1657{
1658 play_dead_common();
1659 tboot_shutdown(TB_SHUTDOWN_WFS);
1660
1661 mwait_play_dead(); /* Only returns on failure */
1662 if (cpuidle_play_dead())
1663 hlt_play_dead();
1664}
1665
1666#else /* ... !CONFIG_HOTPLUG_CPU */
1667int native_cpu_disable(void)
1668{
1669 return -ENOSYS;
1670}
1671
1672void native_cpu_die(unsigned int cpu)
1673{
1674 /* We said "no" in __cpu_disable */
1675 BUG();
1676}
1677
1678void native_play_dead(void)
1679{
1680 BUG();
1681}
1682
1683#endif
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/gfp.h>
55#include <linux/cpuidle.h>
56#include <linux/kexec.h>
57#include <linux/numa.h>
58#include <linux/pgtable.h>
59#include <linux/overflow.h>
60#include <linux/stackprotector.h>
61#include <linux/cpuhotplug.h>
62#include <linux/mc146818rtc.h>
63#include <linux/acpi.h>
64
65#include <asm/acpi.h>
66#include <asm/cacheinfo.h>
67#include <asm/desc.h>
68#include <asm/nmi.h>
69#include <asm/irq.h>
70#include <asm/realmode.h>
71#include <asm/cpu.h>
72#include <asm/numa.h>
73#include <asm/tlbflush.h>
74#include <asm/mtrr.h>
75#include <asm/mwait.h>
76#include <asm/apic.h>
77#include <asm/io_apic.h>
78#include <asm/fpu/api.h>
79#include <asm/setup.h>
80#include <asm/uv/uv.h>
81#include <asm/microcode.h>
82#include <asm/i8259.h>
83#include <asm/misc.h>
84#include <asm/qspinlock.h>
85#include <asm/intel-family.h>
86#include <asm/cpu_device_id.h>
87#include <asm/spec-ctrl.h>
88#include <asm/hw_irq.h>
89#include <asm/stackprotector.h>
90#include <asm/sev.h>
91#include <asm/spec-ctrl.h>
92
93/* representing HT siblings of each logical CPU */
94DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
95EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
96
97/* representing HT and core siblings of each logical CPU */
98DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
99EXPORT_PER_CPU_SYMBOL(cpu_core_map);
100
101/* representing HT, core, and die siblings of each logical CPU */
102DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_die_map);
103EXPORT_PER_CPU_SYMBOL(cpu_die_map);
104
105/* CPUs which are the primary SMT threads */
106struct cpumask __cpu_primary_thread_mask __read_mostly;
107
108/* Representing CPUs for which sibling maps can be computed */
109static cpumask_var_t cpu_sibling_setup_mask;
110
111struct mwait_cpu_dead {
112 unsigned int control;
113 unsigned int status;
114};
115
116#define CPUDEAD_MWAIT_WAIT 0xDEADBEEF
117#define CPUDEAD_MWAIT_KEXEC_HLT 0x4A17DEAD
118
119/*
120 * Cache line aligned data for mwait_play_dead(). Separate on purpose so
121 * that it's unlikely to be touched by other CPUs.
122 */
123static DEFINE_PER_CPU_ALIGNED(struct mwait_cpu_dead, mwait_cpu_dead);
124
125/* Maximum number of SMT threads on any online core */
126int __read_mostly __max_smt_threads = 1;
127
128/* Flag to indicate if a complete sched domain rebuild is required */
129bool x86_topology_update;
130
131int arch_update_cpu_topology(void)
132{
133 int retval = x86_topology_update;
134
135 x86_topology_update = false;
136 return retval;
137}
138
139static unsigned int smpboot_warm_reset_vector_count;
140
141static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
142{
143 unsigned long flags;
144
145 spin_lock_irqsave(&rtc_lock, flags);
146 if (!smpboot_warm_reset_vector_count++) {
147 CMOS_WRITE(0xa, 0xf);
148 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) = start_eip >> 4;
149 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = start_eip & 0xf;
150 }
151 spin_unlock_irqrestore(&rtc_lock, flags);
152}
153
154static inline void smpboot_restore_warm_reset_vector(void)
155{
156 unsigned long flags;
157
158 /*
159 * Paranoid: Set warm reset code and vector here back
160 * to default values.
161 */
162 spin_lock_irqsave(&rtc_lock, flags);
163 if (!--smpboot_warm_reset_vector_count) {
164 CMOS_WRITE(0, 0xf);
165 *((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
166 }
167 spin_unlock_irqrestore(&rtc_lock, flags);
168
169}
170
171/* Run the next set of setup steps for the upcoming CPU */
172static void ap_starting(void)
173{
174 int cpuid = smp_processor_id();
175
176 /* Mop up eventual mwait_play_dead() wreckage */
177 this_cpu_write(mwait_cpu_dead.status, 0);
178 this_cpu_write(mwait_cpu_dead.control, 0);
179
180 /*
181 * If woken up by an INIT in an 82489DX configuration the alive
182 * synchronization guarantees that the CPU does not reach this
183 * point before an INIT_deassert IPI reaches the local APIC, so it
184 * is now safe to touch the local APIC.
185 *
186 * Set up this CPU, first the APIC, which is probably redundant on
187 * most boards.
188 */
189 apic_ap_setup();
190
191 /* Save the processor parameters. */
192 smp_store_cpu_info(cpuid);
193
194 /*
195 * The topology information must be up to date before
196 * notify_cpu_starting().
197 */
198 set_cpu_sibling_map(cpuid);
199
200 ap_init_aperfmperf();
201
202 pr_debug("Stack at about %p\n", &cpuid);
203
204 wmb();
205
206 /*
207 * This runs the AP through all the cpuhp states to its target
208 * state CPUHP_ONLINE.
209 */
210 notify_cpu_starting(cpuid);
211}
212
213static void ap_calibrate_delay(void)
214{
215 /*
216 * Calibrate the delay loop and update loops_per_jiffy in cpu_data.
217 * smp_store_cpu_info() stored a value that is close but not as
218 * accurate as the value just calculated.
219 *
220 * As this is invoked after the TSC synchronization check,
221 * calibrate_delay_is_known() will skip the calibration routine
222 * when TSC is synchronized across sockets.
223 */
224 calibrate_delay();
225 cpu_data(smp_processor_id()).loops_per_jiffy = loops_per_jiffy;
226}
227
228/*
229 * Activate a secondary processor.
230 */
231static void notrace start_secondary(void *unused)
232{
233 /*
234 * Don't put *anything* except direct CPU state initialization
235 * before cpu_init(), SMP booting is too fragile that we want to
236 * limit the things done here to the most necessary things.
237 */
238 cr4_init();
239
240 /*
241 * 32-bit specific. 64-bit reaches this code with the correct page
242 * table established. Yet another historical divergence.
243 */
244 if (IS_ENABLED(CONFIG_X86_32)) {
245 /* switch away from the initial page table */
246 load_cr3(swapper_pg_dir);
247 __flush_tlb_all();
248 }
249
250 cpu_init_exception_handling(false);
251
252 /*
253 * Load the microcode before reaching the AP alive synchronization
254 * point below so it is not part of the full per CPU serialized
255 * bringup part when "parallel" bringup is enabled.
256 *
257 * That's even safe when hyperthreading is enabled in the CPU as
258 * the core code starts the primary threads first and leaves the
259 * secondary threads waiting for SIPI. Loading microcode on
260 * physical cores concurrently is a safe operation.
261 *
262 * This covers both the Intel specific issue that concurrent
263 * microcode loading on SMT siblings must be prohibited and the
264 * vendor independent issue`that microcode loading which changes
265 * CPUID, MSRs etc. must be strictly serialized to maintain
266 * software state correctness.
267 */
268 load_ucode_ap();
269
270 /*
271 * Synchronization point with the hotplug core. Sets this CPUs
272 * synchronization state to ALIVE and spin-waits for the control CPU to
273 * release this CPU for further bringup.
274 */
275 cpuhp_ap_sync_alive();
276
277 cpu_init();
278 fpu__init_cpu();
279 rcutree_report_cpu_starting(raw_smp_processor_id());
280 x86_cpuinit.early_percpu_clock_init();
281
282 ap_starting();
283
284 /* Check TSC synchronization with the control CPU. */
285 check_tsc_sync_target();
286
287 /*
288 * Calibrate the delay loop after the TSC synchronization check.
289 * This allows to skip the calibration when TSC is synchronized
290 * across sockets.
291 */
292 ap_calibrate_delay();
293
294 speculative_store_bypass_ht_init();
295
296 /*
297 * Lock vector_lock, set CPU online and bring the vector
298 * allocator online. Online must be set with vector_lock held
299 * to prevent a concurrent irq setup/teardown from seeing a
300 * half valid vector space.
301 */
302 lock_vector_lock();
303 set_cpu_online(smp_processor_id(), true);
304 lapic_online();
305 unlock_vector_lock();
306 x86_platform.nmi_init();
307
308 /* enable local interrupts */
309 local_irq_enable();
310
311 x86_cpuinit.setup_percpu_clockev();
312
313 wmb();
314 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
315}
316
317/*
318 * The bootstrap kernel entry code has set these up. Save them for
319 * a given CPU
320 */
321void smp_store_cpu_info(int id)
322{
323 struct cpuinfo_x86 *c = &cpu_data(id);
324
325 /* Copy boot_cpu_data only on the first bringup */
326 if (!c->initialized)
327 *c = boot_cpu_data;
328 c->cpu_index = id;
329 /*
330 * During boot time, CPU0 has this setup already. Save the info when
331 * bringing up an AP.
332 */
333 identify_secondary_cpu(c);
334 c->initialized = true;
335}
336
337static bool
338topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
339{
340 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
341
342 return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
343}
344
345static bool
346topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
347{
348 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
349
350 return !WARN_ONCE(!topology_same_node(c, o),
351 "sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
352 "[node: %d != %d]. Ignoring dependency.\n",
353 cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
354}
355
356#define link_mask(mfunc, c1, c2) \
357do { \
358 cpumask_set_cpu((c1), mfunc(c2)); \
359 cpumask_set_cpu((c2), mfunc(c1)); \
360} while (0)
361
362static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
363{
364 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
365 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
366
367 if (c->topo.pkg_id == o->topo.pkg_id &&
368 c->topo.die_id == o->topo.die_id &&
369 c->topo.amd_node_id == o->topo.amd_node_id &&
370 per_cpu_llc_id(cpu1) == per_cpu_llc_id(cpu2)) {
371 if (c->topo.core_id == o->topo.core_id)
372 return topology_sane(c, o, "smt");
373
374 if ((c->topo.cu_id != 0xff) &&
375 (o->topo.cu_id != 0xff) &&
376 (c->topo.cu_id == o->topo.cu_id))
377 return topology_sane(c, o, "smt");
378 }
379
380 } else if (c->topo.pkg_id == o->topo.pkg_id &&
381 c->topo.die_id == o->topo.die_id &&
382 c->topo.core_id == o->topo.core_id) {
383 return topology_sane(c, o, "smt");
384 }
385
386 return false;
387}
388
389static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
390{
391 if (c->topo.pkg_id != o->topo.pkg_id || c->topo.die_id != o->topo.die_id)
392 return false;
393
394 if (cpu_feature_enabled(X86_FEATURE_TOPOEXT) && topology_amd_nodes_per_pkg() > 1)
395 return c->topo.amd_node_id == o->topo.amd_node_id;
396
397 return true;
398}
399
400static bool match_l2c(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
401{
402 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
403
404 /* If the arch didn't set up l2c_id, fall back to SMT */
405 if (per_cpu_l2c_id(cpu1) == BAD_APICID)
406 return match_smt(c, o);
407
408 /* Do not match if L2 cache id does not match: */
409 if (per_cpu_l2c_id(cpu1) != per_cpu_l2c_id(cpu2))
410 return false;
411
412 return topology_sane(c, o, "l2c");
413}
414
415/*
416 * Unlike the other levels, we do not enforce keeping a
417 * multicore group inside a NUMA node. If this happens, we will
418 * discard the MC level of the topology later.
419 */
420static bool match_pkg(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
421{
422 if (c->topo.pkg_id == o->topo.pkg_id)
423 return true;
424 return false;
425}
426
427/*
428 * Define intel_cod_cpu[] for Intel COD (Cluster-on-Die) CPUs.
429 *
430 * Any Intel CPU that has multiple nodes per package and does not
431 * match intel_cod_cpu[] has the SNC (Sub-NUMA Cluster) topology.
432 *
433 * When in SNC mode, these CPUs enumerate an LLC that is shared
434 * by multiple NUMA nodes. The LLC is shared for off-package data
435 * access but private to the NUMA node (half of the package) for
436 * on-package access. CPUID (the source of the information about
437 * the LLC) can only enumerate the cache as shared or unshared,
438 * but not this particular configuration.
439 */
440
441static const struct x86_cpu_id intel_cod_cpu[] = {
442 X86_MATCH_VFM(INTEL_HASWELL_X, 0), /* COD */
443 X86_MATCH_VFM(INTEL_BROADWELL_X, 0), /* COD */
444 X86_MATCH_VFM(INTEL_ANY, 1), /* SNC */
445 {}
446};
447
448static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
449{
450 const struct x86_cpu_id *id = x86_match_cpu(intel_cod_cpu);
451 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
452 bool intel_snc = id && id->driver_data;
453
454 /* Do not match if we do not have a valid APICID for cpu: */
455 if (per_cpu_llc_id(cpu1) == BAD_APICID)
456 return false;
457
458 /* Do not match if LLC id does not match: */
459 if (per_cpu_llc_id(cpu1) != per_cpu_llc_id(cpu2))
460 return false;
461
462 /*
463 * Allow the SNC topology without warning. Return of false
464 * means 'c' does not share the LLC of 'o'. This will be
465 * reflected to userspace.
466 */
467 if (match_pkg(c, o) && !topology_same_node(c, o) && intel_snc)
468 return false;
469
470 return topology_sane(c, o, "llc");
471}
472
473
474static inline int x86_sched_itmt_flags(void)
475{
476 return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
477}
478
479#ifdef CONFIG_SCHED_MC
480static int x86_core_flags(void)
481{
482 return cpu_core_flags() | x86_sched_itmt_flags();
483}
484#endif
485#ifdef CONFIG_SCHED_SMT
486static int x86_smt_flags(void)
487{
488 return cpu_smt_flags();
489}
490#endif
491#ifdef CONFIG_SCHED_CLUSTER
492static int x86_cluster_flags(void)
493{
494 return cpu_cluster_flags() | x86_sched_itmt_flags();
495}
496#endif
497
498/*
499 * Set if a package/die has multiple NUMA nodes inside.
500 * AMD Magny-Cours, Intel Cluster-on-Die, and Intel
501 * Sub-NUMA Clustering have this.
502 */
503static bool x86_has_numa_in_package;
504
505static struct sched_domain_topology_level x86_topology[6];
506
507static void __init build_sched_topology(void)
508{
509 int i = 0;
510
511#ifdef CONFIG_SCHED_SMT
512 x86_topology[i++] = (struct sched_domain_topology_level){
513 cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT)
514 };
515#endif
516#ifdef CONFIG_SCHED_CLUSTER
517 x86_topology[i++] = (struct sched_domain_topology_level){
518 cpu_clustergroup_mask, x86_cluster_flags, SD_INIT_NAME(CLS)
519 };
520#endif
521#ifdef CONFIG_SCHED_MC
522 x86_topology[i++] = (struct sched_domain_topology_level){
523 cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC)
524 };
525#endif
526 /*
527 * When there is NUMA topology inside the package skip the PKG domain
528 * since the NUMA domains will auto-magically create the right spanning
529 * domains based on the SLIT.
530 */
531 if (!x86_has_numa_in_package) {
532 x86_topology[i++] = (struct sched_domain_topology_level){
533 cpu_cpu_mask, x86_sched_itmt_flags, SD_INIT_NAME(PKG)
534 };
535 }
536
537 /*
538 * There must be one trailing NULL entry left.
539 */
540 BUG_ON(i >= ARRAY_SIZE(x86_topology)-1);
541
542 set_sched_topology(x86_topology);
543}
544
545void set_cpu_sibling_map(int cpu)
546{
547 bool has_smt = __max_threads_per_core > 1;
548 bool has_mp = has_smt || topology_num_cores_per_package() > 1;
549 struct cpuinfo_x86 *c = &cpu_data(cpu);
550 struct cpuinfo_x86 *o;
551 int i, threads;
552
553 cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
554
555 if (!has_mp) {
556 cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
557 cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
558 cpumask_set_cpu(cpu, cpu_l2c_shared_mask(cpu));
559 cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
560 cpumask_set_cpu(cpu, topology_die_cpumask(cpu));
561 c->booted_cores = 1;
562 return;
563 }
564
565 for_each_cpu(i, cpu_sibling_setup_mask) {
566 o = &cpu_data(i);
567
568 if (match_pkg(c, o) && !topology_same_node(c, o))
569 x86_has_numa_in_package = true;
570
571 if ((i == cpu) || (has_smt && match_smt(c, o)))
572 link_mask(topology_sibling_cpumask, cpu, i);
573
574 if ((i == cpu) || (has_mp && match_llc(c, o)))
575 link_mask(cpu_llc_shared_mask, cpu, i);
576
577 if ((i == cpu) || (has_mp && match_l2c(c, o)))
578 link_mask(cpu_l2c_shared_mask, cpu, i);
579
580 if ((i == cpu) || (has_mp && match_die(c, o)))
581 link_mask(topology_die_cpumask, cpu, i);
582 }
583
584 threads = cpumask_weight(topology_sibling_cpumask(cpu));
585 if (threads > __max_smt_threads)
586 __max_smt_threads = threads;
587
588 for_each_cpu(i, topology_sibling_cpumask(cpu))
589 cpu_data(i).smt_active = threads > 1;
590
591 /*
592 * This needs a separate iteration over the cpus because we rely on all
593 * topology_sibling_cpumask links to be set-up.
594 */
595 for_each_cpu(i, cpu_sibling_setup_mask) {
596 o = &cpu_data(i);
597
598 if ((i == cpu) || (has_mp && match_pkg(c, o))) {
599 link_mask(topology_core_cpumask, cpu, i);
600
601 /*
602 * Does this new cpu bringup a new core?
603 */
604 if (threads == 1) {
605 /*
606 * for each core in package, increment
607 * the booted_cores for this new cpu
608 */
609 if (cpumask_first(
610 topology_sibling_cpumask(i)) == i)
611 c->booted_cores++;
612 /*
613 * increment the core count for all
614 * the other cpus in this package
615 */
616 if (i != cpu)
617 cpu_data(i).booted_cores++;
618 } else if (i != cpu && !c->booted_cores)
619 c->booted_cores = cpu_data(i).booted_cores;
620 }
621 }
622}
623
624/* maps the cpu to the sched domain representing multi-core */
625const struct cpumask *cpu_coregroup_mask(int cpu)
626{
627 return cpu_llc_shared_mask(cpu);
628}
629
630const struct cpumask *cpu_clustergroup_mask(int cpu)
631{
632 return cpu_l2c_shared_mask(cpu);
633}
634EXPORT_SYMBOL_GPL(cpu_clustergroup_mask);
635
636static void impress_friends(void)
637{
638 int cpu;
639 unsigned long bogosum = 0;
640 /*
641 * Allow the user to impress friends.
642 */
643 pr_debug("Before bogomips\n");
644 for_each_online_cpu(cpu)
645 bogosum += cpu_data(cpu).loops_per_jiffy;
646
647 pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
648 num_online_cpus(),
649 bogosum/(500000/HZ),
650 (bogosum/(5000/HZ))%100);
651
652 pr_debug("Before bogocount - setting activated=1\n");
653}
654
655/*
656 * The Multiprocessor Specification 1.4 (1997) example code suggests
657 * that there should be a 10ms delay between the BSP asserting INIT
658 * and de-asserting INIT, when starting a remote processor.
659 * But that slows boot and resume on modern processors, which include
660 * many cores and don't require that delay.
661 *
662 * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
663 * Modern processor families are quirked to remove the delay entirely.
664 */
665#define UDELAY_10MS_DEFAULT 10000
666
667static unsigned int init_udelay = UINT_MAX;
668
669static int __init cpu_init_udelay(char *str)
670{
671 get_option(&str, &init_udelay);
672
673 return 0;
674}
675early_param("cpu_init_udelay", cpu_init_udelay);
676
677static void __init smp_quirk_init_udelay(void)
678{
679 /* if cmdline changed it from default, leave it alone */
680 if (init_udelay != UINT_MAX)
681 return;
682
683 /* if modern processor, use no delay */
684 if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
685 ((boot_cpu_data.x86_vendor == X86_VENDOR_HYGON) && (boot_cpu_data.x86 >= 0x18)) ||
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 * Wake up AP by INIT, INIT, STARTUP sequence.
696 */
697static void send_init_sequence(u32 phys_apicid)
698{
699 int maxlvt = lapic_get_maxlvt();
700
701 /* Be paranoid about clearing APIC errors. */
702 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
703 /* Due to the Pentium erratum 3AP. */
704 if (maxlvt > 3)
705 apic_write(APIC_ESR, 0);
706 apic_read(APIC_ESR);
707 }
708
709 /* Assert INIT on the target CPU */
710 apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT, phys_apicid);
711 safe_apic_wait_icr_idle();
712
713 udelay(init_udelay);
714
715 /* Deassert INIT on the target CPU */
716 apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
717 safe_apic_wait_icr_idle();
718}
719
720/*
721 * Wake up AP by INIT, INIT, STARTUP sequence.
722 */
723static int wakeup_secondary_cpu_via_init(u32 phys_apicid, unsigned long start_eip)
724{
725 unsigned long send_status = 0, accept_status = 0;
726 int num_starts, j, maxlvt;
727
728 preempt_disable();
729 maxlvt = lapic_get_maxlvt();
730 send_init_sequence(phys_apicid);
731
732 mb();
733
734 /*
735 * Should we send STARTUP IPIs ?
736 *
737 * Determine this based on the APIC version.
738 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
739 */
740 if (APIC_INTEGRATED(boot_cpu_apic_version))
741 num_starts = 2;
742 else
743 num_starts = 0;
744
745 /*
746 * Run STARTUP IPI loop.
747 */
748 pr_debug("#startup loops: %d\n", num_starts);
749
750 for (j = 1; j <= num_starts; j++) {
751 pr_debug("Sending STARTUP #%d\n", j);
752 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
753 apic_write(APIC_ESR, 0);
754 apic_read(APIC_ESR);
755 pr_debug("After apic_write\n");
756
757 /*
758 * STARTUP IPI
759 */
760
761 /* Target chip */
762 /* Boot on the stack */
763 /* Kick the second */
764 apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
765 phys_apicid);
766
767 /*
768 * Give the other CPU some time to accept the IPI.
769 */
770 if (init_udelay == 0)
771 udelay(10);
772 else
773 udelay(300);
774
775 pr_debug("Startup point 1\n");
776
777 pr_debug("Waiting for send to finish...\n");
778 send_status = safe_apic_wait_icr_idle();
779
780 /*
781 * Give the other CPU some time to accept the IPI.
782 */
783 if (init_udelay == 0)
784 udelay(10);
785 else
786 udelay(200);
787
788 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
789 apic_write(APIC_ESR, 0);
790 accept_status = (apic_read(APIC_ESR) & 0xEF);
791 if (send_status || accept_status)
792 break;
793 }
794 pr_debug("After Startup\n");
795
796 if (send_status)
797 pr_err("APIC never delivered???\n");
798 if (accept_status)
799 pr_err("APIC delivery error (%lx)\n", accept_status);
800
801 preempt_enable();
802 return (send_status | accept_status);
803}
804
805/* reduce the number of lines printed when booting a large cpu count system */
806static void announce_cpu(int cpu, int apicid)
807{
808 static int width, node_width, first = 1;
809 static int current_node = NUMA_NO_NODE;
810 int node = early_cpu_to_node(cpu);
811
812 if (!width)
813 width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
814
815 if (!node_width)
816 node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
817
818 if (system_state < SYSTEM_RUNNING) {
819 if (first)
820 pr_info("x86: Booting SMP configuration:\n");
821
822 if (node != current_node) {
823 if (current_node > (-1))
824 pr_cont("\n");
825 current_node = node;
826
827 printk(KERN_INFO ".... node %*s#%d, CPUs: ",
828 node_width - num_digits(node), " ", node);
829 }
830
831 /* Add padding for the BSP */
832 if (first)
833 pr_cont("%*s", width + 1, " ");
834 first = 0;
835
836 pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
837 } else
838 pr_info("Booting Node %d Processor %d APIC 0x%x\n",
839 node, cpu, apicid);
840}
841
842int common_cpu_up(unsigned int cpu, struct task_struct *idle)
843{
844 int ret;
845
846 /* Just in case we booted with a single CPU. */
847 alternatives_enable_smp();
848
849 per_cpu(pcpu_hot.current_task, cpu) = idle;
850 cpu_init_stack_canary(cpu, idle);
851
852 /* Initialize the interrupt stack(s) */
853 ret = irq_init_percpu_irqstack(cpu);
854 if (ret)
855 return ret;
856
857#ifdef CONFIG_X86_32
858 /* Stack for startup_32 can be just as for start_secondary onwards */
859 per_cpu(pcpu_hot.top_of_stack, cpu) = task_top_of_stack(idle);
860#endif
861 return 0;
862}
863
864/*
865 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
866 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
867 * Returns zero if startup was successfully sent, else error code from
868 * ->wakeup_secondary_cpu.
869 */
870static int do_boot_cpu(u32 apicid, int cpu, struct task_struct *idle)
871{
872 unsigned long start_ip = real_mode_header->trampoline_start;
873 int ret;
874
875#ifdef CONFIG_X86_64
876 /* If 64-bit wakeup method exists, use the 64-bit mode trampoline IP */
877 if (apic->wakeup_secondary_cpu_64)
878 start_ip = real_mode_header->trampoline_start64;
879#endif
880 idle->thread.sp = (unsigned long)task_pt_regs(idle);
881 initial_code = (unsigned long)start_secondary;
882
883 if (IS_ENABLED(CONFIG_X86_32)) {
884 early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
885 initial_stack = idle->thread.sp;
886 } else if (!(smpboot_control & STARTUP_PARALLEL_MASK)) {
887 smpboot_control = cpu;
888 }
889
890 /* Enable the espfix hack for this CPU */
891 init_espfix_ap(cpu);
892
893 /* So we see what's up */
894 announce_cpu(cpu, apicid);
895
896 /*
897 * This grunge runs the startup process for
898 * the targeted processor.
899 */
900 if (x86_platform.legacy.warm_reset) {
901
902 pr_debug("Setting warm reset code and vector.\n");
903
904 smpboot_setup_warm_reset_vector(start_ip);
905 /*
906 * Be paranoid about clearing APIC errors.
907 */
908 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
909 apic_write(APIC_ESR, 0);
910 apic_read(APIC_ESR);
911 }
912 }
913
914 smp_mb();
915
916 /*
917 * Wake up a CPU in difference cases:
918 * - Use a method from the APIC driver if one defined, with wakeup
919 * straight to 64-bit mode preferred over wakeup to RM.
920 * Otherwise,
921 * - Use an INIT boot APIC message
922 */
923 if (apic->wakeup_secondary_cpu_64)
924 ret = apic->wakeup_secondary_cpu_64(apicid, start_ip);
925 else if (apic->wakeup_secondary_cpu)
926 ret = apic->wakeup_secondary_cpu(apicid, start_ip);
927 else
928 ret = wakeup_secondary_cpu_via_init(apicid, start_ip);
929
930 /* If the wakeup mechanism failed, cleanup the warm reset vector */
931 if (ret)
932 arch_cpuhp_cleanup_kick_cpu(cpu);
933 return ret;
934}
935
936int native_kick_ap(unsigned int cpu, struct task_struct *tidle)
937{
938 u32 apicid = apic->cpu_present_to_apicid(cpu);
939 int err;
940
941 lockdep_assert_irqs_enabled();
942
943 pr_debug("++++++++++++++++++++=_---CPU UP %u\n", cpu);
944
945 if (apicid == BAD_APICID || !apic_id_valid(apicid)) {
946 pr_err("CPU %u has invalid APIC ID %x. Aborting bringup\n", cpu, apicid);
947 return -EINVAL;
948 }
949
950 if (!test_bit(apicid, phys_cpu_present_map)) {
951 pr_err("CPU %u APIC ID %x is not present. Aborting bringup\n", cpu, apicid);
952 return -EINVAL;
953 }
954
955 /*
956 * Save current MTRR state in case it was changed since early boot
957 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
958 */
959 mtrr_save_state();
960
961 /* the FPU context is blank, nobody can own it */
962 per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
963
964 err = common_cpu_up(cpu, tidle);
965 if (err)
966 return err;
967
968 err = do_boot_cpu(apicid, cpu, tidle);
969 if (err)
970 pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
971
972 return err;
973}
974
975int arch_cpuhp_kick_ap_alive(unsigned int cpu, struct task_struct *tidle)
976{
977 return smp_ops.kick_ap_alive(cpu, tidle);
978}
979
980void arch_cpuhp_cleanup_kick_cpu(unsigned int cpu)
981{
982 /* Cleanup possible dangling ends... */
983 if (smp_ops.kick_ap_alive == native_kick_ap && x86_platform.legacy.warm_reset)
984 smpboot_restore_warm_reset_vector();
985}
986
987void arch_cpuhp_cleanup_dead_cpu(unsigned int cpu)
988{
989 if (smp_ops.cleanup_dead_cpu)
990 smp_ops.cleanup_dead_cpu(cpu);
991
992 if (system_state == SYSTEM_RUNNING)
993 pr_info("CPU %u is now offline\n", cpu);
994}
995
996void arch_cpuhp_sync_state_poll(void)
997{
998 if (smp_ops.poll_sync_state)
999 smp_ops.poll_sync_state();
1000}
1001
1002/**
1003 * arch_disable_smp_support() - Disables SMP support for x86 at boottime
1004 */
1005void __init arch_disable_smp_support(void)
1006{
1007 disable_ioapic_support();
1008}
1009
1010/*
1011 * Fall back to non SMP mode after errors.
1012 *
1013 * RED-PEN audit/test this more. I bet there is more state messed up here.
1014 */
1015static __init void disable_smp(void)
1016{
1017 pr_info("SMP disabled\n");
1018
1019 disable_ioapic_support();
1020 topology_reset_possible_cpus_up();
1021
1022 cpumask_set_cpu(0, topology_sibling_cpumask(0));
1023 cpumask_set_cpu(0, topology_core_cpumask(0));
1024 cpumask_set_cpu(0, topology_die_cpumask(0));
1025}
1026
1027void __init smp_prepare_cpus_common(void)
1028{
1029 unsigned int cpu, node;
1030
1031 /* Mark all except the boot CPU as hotpluggable */
1032 for_each_possible_cpu(cpu) {
1033 if (cpu)
1034 per_cpu(cpu_info.cpu_index, cpu) = nr_cpu_ids;
1035 }
1036
1037 for_each_possible_cpu(cpu) {
1038 node = cpu_to_node(cpu);
1039
1040 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu), GFP_KERNEL, node);
1041 zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu), GFP_KERNEL, node);
1042 zalloc_cpumask_var_node(&per_cpu(cpu_die_map, cpu), GFP_KERNEL, node);
1043 zalloc_cpumask_var_node(&per_cpu(cpu_llc_shared_map, cpu), GFP_KERNEL, node);
1044 zalloc_cpumask_var_node(&per_cpu(cpu_l2c_shared_map, cpu), GFP_KERNEL, node);
1045 }
1046
1047 set_cpu_sibling_map(0);
1048}
1049
1050void __init smp_prepare_boot_cpu(void)
1051{
1052 smp_ops.smp_prepare_boot_cpu();
1053}
1054
1055#ifdef CONFIG_X86_64
1056/* Establish whether parallel bringup can be supported. */
1057bool __init arch_cpuhp_init_parallel_bringup(void)
1058{
1059 if (!x86_cpuinit.parallel_bringup) {
1060 pr_info("Parallel CPU startup disabled by the platform\n");
1061 return false;
1062 }
1063
1064 smpboot_control = STARTUP_READ_APICID;
1065 pr_debug("Parallel CPU startup enabled: 0x%08x\n", smpboot_control);
1066 return true;
1067}
1068#endif
1069
1070/*
1071 * Prepare for SMP bootup.
1072 * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1073 * for common interface support.
1074 */
1075void __init native_smp_prepare_cpus(unsigned int max_cpus)
1076{
1077 smp_prepare_cpus_common();
1078
1079 switch (apic_intr_mode) {
1080 case APIC_PIC:
1081 case APIC_VIRTUAL_WIRE_NO_CONFIG:
1082 disable_smp();
1083 return;
1084 case APIC_SYMMETRIC_IO_NO_ROUTING:
1085 disable_smp();
1086 /* Setup local timer */
1087 x86_init.timers.setup_percpu_clockev();
1088 return;
1089 case APIC_VIRTUAL_WIRE:
1090 case APIC_SYMMETRIC_IO:
1091 break;
1092 }
1093
1094 /* Setup local timer */
1095 x86_init.timers.setup_percpu_clockev();
1096
1097 pr_info("CPU0: ");
1098 print_cpu_info(&cpu_data(0));
1099
1100 uv_system_init();
1101
1102 smp_quirk_init_udelay();
1103
1104 speculative_store_bypass_ht_init();
1105
1106 snp_set_wakeup_secondary_cpu();
1107}
1108
1109void arch_thaw_secondary_cpus_begin(void)
1110{
1111 set_cache_aps_delayed_init(true);
1112}
1113
1114void arch_thaw_secondary_cpus_end(void)
1115{
1116 cache_aps_init();
1117}
1118
1119/*
1120 * Early setup to make printk work.
1121 */
1122void __init native_smp_prepare_boot_cpu(void)
1123{
1124 int me = smp_processor_id();
1125
1126 /* SMP handles this from setup_per_cpu_areas() */
1127 if (!IS_ENABLED(CONFIG_SMP))
1128 switch_gdt_and_percpu_base(me);
1129
1130 native_pv_lock_init();
1131}
1132
1133void __init native_smp_cpus_done(unsigned int max_cpus)
1134{
1135 pr_debug("Boot done\n");
1136
1137 build_sched_topology();
1138 nmi_selftest();
1139 impress_friends();
1140 cache_aps_init();
1141}
1142
1143/* correctly size the local cpu masks */
1144void __init setup_cpu_local_masks(void)
1145{
1146 alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
1147}
1148
1149#ifdef CONFIG_HOTPLUG_CPU
1150
1151/* Recompute SMT state for all CPUs on offline */
1152static void recompute_smt_state(void)
1153{
1154 int max_threads, cpu;
1155
1156 max_threads = 0;
1157 for_each_online_cpu (cpu) {
1158 int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1159
1160 if (threads > max_threads)
1161 max_threads = threads;
1162 }
1163 __max_smt_threads = max_threads;
1164}
1165
1166static void remove_siblinginfo(int cpu)
1167{
1168 int sibling;
1169 struct cpuinfo_x86 *c = &cpu_data(cpu);
1170
1171 for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1172 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1173 /*/
1174 * last thread sibling in this cpu core going down
1175 */
1176 if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1177 cpu_data(sibling).booted_cores--;
1178 }
1179
1180 for_each_cpu(sibling, topology_die_cpumask(cpu))
1181 cpumask_clear_cpu(cpu, topology_die_cpumask(sibling));
1182
1183 for_each_cpu(sibling, topology_sibling_cpumask(cpu)) {
1184 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1185 if (cpumask_weight(topology_sibling_cpumask(sibling)) == 1)
1186 cpu_data(sibling).smt_active = false;
1187 }
1188
1189 for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1190 cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1191 for_each_cpu(sibling, cpu_l2c_shared_mask(cpu))
1192 cpumask_clear_cpu(cpu, cpu_l2c_shared_mask(sibling));
1193 cpumask_clear(cpu_llc_shared_mask(cpu));
1194 cpumask_clear(cpu_l2c_shared_mask(cpu));
1195 cpumask_clear(topology_sibling_cpumask(cpu));
1196 cpumask_clear(topology_core_cpumask(cpu));
1197 cpumask_clear(topology_die_cpumask(cpu));
1198 c->topo.core_id = 0;
1199 c->booted_cores = 0;
1200 cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1201 recompute_smt_state();
1202}
1203
1204static void remove_cpu_from_maps(int cpu)
1205{
1206 set_cpu_online(cpu, false);
1207 numa_remove_cpu(cpu);
1208}
1209
1210void cpu_disable_common(void)
1211{
1212 int cpu = smp_processor_id();
1213
1214 remove_siblinginfo(cpu);
1215
1216 /* It's now safe to remove this processor from the online map */
1217 lock_vector_lock();
1218 remove_cpu_from_maps(cpu);
1219 unlock_vector_lock();
1220 fixup_irqs();
1221 lapic_offline();
1222}
1223
1224int native_cpu_disable(void)
1225{
1226 int ret;
1227
1228 ret = lapic_can_unplug_cpu();
1229 if (ret)
1230 return ret;
1231
1232 cpu_disable_common();
1233
1234 /*
1235 * Disable the local APIC. Otherwise IPI broadcasts will reach
1236 * it. It still responds normally to INIT, NMI, SMI, and SIPI
1237 * messages.
1238 *
1239 * Disabling the APIC must happen after cpu_disable_common()
1240 * which invokes fixup_irqs().
1241 *
1242 * Disabling the APIC preserves already set bits in IRR, but
1243 * an interrupt arriving after disabling the local APIC does not
1244 * set the corresponding IRR bit.
1245 *
1246 * fixup_irqs() scans IRR for set bits so it can raise a not
1247 * yet handled interrupt on the new destination CPU via an IPI
1248 * but obviously it can't do so for IRR bits which are not set.
1249 * IOW, interrupts arriving after disabling the local APIC will
1250 * be lost.
1251 */
1252 apic_soft_disable();
1253
1254 return 0;
1255}
1256
1257void play_dead_common(void)
1258{
1259 idle_task_exit();
1260
1261 cpuhp_ap_report_dead();
1262
1263 local_irq_disable();
1264}
1265
1266/*
1267 * We need to flush the caches before going to sleep, lest we have
1268 * dirty data in our caches when we come back up.
1269 */
1270static inline void mwait_play_dead(void)
1271{
1272 struct mwait_cpu_dead *md = this_cpu_ptr(&mwait_cpu_dead);
1273 unsigned int eax, ebx, ecx, edx;
1274 unsigned int highest_cstate = 0;
1275 unsigned int highest_subcstate = 0;
1276 int i;
1277
1278 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD ||
1279 boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
1280 return;
1281 if (!this_cpu_has(X86_FEATURE_MWAIT))
1282 return;
1283 if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1284 return;
1285 if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1286 return;
1287
1288 eax = CPUID_MWAIT_LEAF;
1289 ecx = 0;
1290 native_cpuid(&eax, &ebx, &ecx, &edx);
1291
1292 /*
1293 * eax will be 0 if EDX enumeration is not valid.
1294 * Initialized below to cstate, sub_cstate value when EDX is valid.
1295 */
1296 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1297 eax = 0;
1298 } else {
1299 edx >>= MWAIT_SUBSTATE_SIZE;
1300 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1301 if (edx & MWAIT_SUBSTATE_MASK) {
1302 highest_cstate = i;
1303 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1304 }
1305 }
1306 eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1307 (highest_subcstate - 1);
1308 }
1309
1310 /* Set up state for the kexec() hack below */
1311 md->status = CPUDEAD_MWAIT_WAIT;
1312 md->control = CPUDEAD_MWAIT_WAIT;
1313
1314 wbinvd();
1315
1316 while (1) {
1317 /*
1318 * The CLFLUSH is a workaround for erratum AAI65 for
1319 * the Xeon 7400 series. It's not clear it is actually
1320 * needed, but it should be harmless in either case.
1321 * The WBINVD is insufficient due to the spurious-wakeup
1322 * case where we return around the loop.
1323 */
1324 mb();
1325 clflush(md);
1326 mb();
1327 __monitor(md, 0, 0);
1328 mb();
1329 __mwait(eax, 0);
1330
1331 if (READ_ONCE(md->control) == CPUDEAD_MWAIT_KEXEC_HLT) {
1332 /*
1333 * Kexec is about to happen. Don't go back into mwait() as
1334 * the kexec kernel might overwrite text and data including
1335 * page tables and stack. So mwait() would resume when the
1336 * monitor cache line is written to and then the CPU goes
1337 * south due to overwritten text, page tables and stack.
1338 *
1339 * Note: This does _NOT_ protect against a stray MCE, NMI,
1340 * SMI. They will resume execution at the instruction
1341 * following the HLT instruction and run into the problem
1342 * which this is trying to prevent.
1343 */
1344 WRITE_ONCE(md->status, CPUDEAD_MWAIT_KEXEC_HLT);
1345 while(1)
1346 native_halt();
1347 }
1348 }
1349}
1350
1351/*
1352 * Kick all "offline" CPUs out of mwait on kexec(). See comment in
1353 * mwait_play_dead().
1354 */
1355void smp_kick_mwait_play_dead(void)
1356{
1357 u32 newstate = CPUDEAD_MWAIT_KEXEC_HLT;
1358 struct mwait_cpu_dead *md;
1359 unsigned int cpu, i;
1360
1361 for_each_cpu_andnot(cpu, cpu_present_mask, cpu_online_mask) {
1362 md = per_cpu_ptr(&mwait_cpu_dead, cpu);
1363
1364 /* Does it sit in mwait_play_dead() ? */
1365 if (READ_ONCE(md->status) != CPUDEAD_MWAIT_WAIT)
1366 continue;
1367
1368 /* Wait up to 5ms */
1369 for (i = 0; READ_ONCE(md->status) != newstate && i < 1000; i++) {
1370 /* Bring it out of mwait */
1371 WRITE_ONCE(md->control, newstate);
1372 udelay(5);
1373 }
1374
1375 if (READ_ONCE(md->status) != newstate)
1376 pr_err_once("CPU%u is stuck in mwait_play_dead()\n", cpu);
1377 }
1378}
1379
1380void __noreturn hlt_play_dead(void)
1381{
1382 if (__this_cpu_read(cpu_info.x86) >= 4)
1383 wbinvd();
1384
1385 while (1)
1386 native_halt();
1387}
1388
1389/*
1390 * native_play_dead() is essentially a __noreturn function, but it can't
1391 * be marked as such as the compiler may complain about it.
1392 */
1393void native_play_dead(void)
1394{
1395 if (cpu_feature_enabled(X86_FEATURE_KERNEL_IBRS))
1396 __update_spec_ctrl(0);
1397
1398 play_dead_common();
1399 tboot_shutdown(TB_SHUTDOWN_WFS);
1400
1401 mwait_play_dead();
1402 if (cpuidle_play_dead())
1403 hlt_play_dead();
1404}
1405
1406#else /* ... !CONFIG_HOTPLUG_CPU */
1407int native_cpu_disable(void)
1408{
1409 return -ENOSYS;
1410}
1411
1412void native_play_dead(void)
1413{
1414 BUG();
1415}
1416
1417#endif