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/module.h>
47#include <linux/sched.h>
48#include <linux/percpu.h>
49#include <linux/bootmem.h>
50#include <linux/err.h>
51#include <linux/nmi.h>
52#include <linux/tboot.h>
53#include <linux/stackprotector.h>
54#include <linux/gfp.h>
55#include <linux/cpuidle.h>
56
57#include <asm/acpi.h>
58#include <asm/desc.h>
59#include <asm/nmi.h>
60#include <asm/irq.h>
61#include <asm/idle.h>
62#include <asm/realmode.h>
63#include <asm/cpu.h>
64#include <asm/numa.h>
65#include <asm/pgtable.h>
66#include <asm/tlbflush.h>
67#include <asm/mtrr.h>
68#include <asm/mwait.h>
69#include <asm/apic.h>
70#include <asm/io_apic.h>
71#include <asm/fpu/internal.h>
72#include <asm/setup.h>
73#include <asm/uv/uv.h>
74#include <linux/mc146818rtc.h>
75#include <asm/i8259.h>
76#include <asm/realmode.h>
77#include <asm/misc.h>
78
79/* Number of siblings per CPU package */
80int smp_num_siblings = 1;
81EXPORT_SYMBOL(smp_num_siblings);
82
83/* Last level cache ID of each logical CPU */
84DEFINE_PER_CPU_READ_MOSTLY(u16, cpu_llc_id) = BAD_APICID;
85
86/* representing HT siblings of each logical CPU */
87DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
88EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
89
90/* representing HT and core siblings of each logical CPU */
91DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
92EXPORT_PER_CPU_SYMBOL(cpu_core_map);
93
94DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_llc_shared_map);
95
96/* Per CPU bogomips and other parameters */
97DEFINE_PER_CPU_READ_MOSTLY(struct cpuinfo_x86, cpu_info);
98EXPORT_PER_CPU_SYMBOL(cpu_info);
99
100/* Logical package management. We might want to allocate that dynamically */
101static int *physical_to_logical_pkg __read_mostly;
102static unsigned long *physical_package_map __read_mostly;;
103static unsigned long *logical_package_map __read_mostly;
104static unsigned int max_physical_pkg_id __read_mostly;
105unsigned int __max_logical_packages __read_mostly;
106EXPORT_SYMBOL(__max_logical_packages);
107
108static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
109{
110 unsigned long flags;
111
112 spin_lock_irqsave(&rtc_lock, flags);
113 CMOS_WRITE(0xa, 0xf);
114 spin_unlock_irqrestore(&rtc_lock, flags);
115 local_flush_tlb();
116 pr_debug("1.\n");
117 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) =
118 start_eip >> 4;
119 pr_debug("2.\n");
120 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) =
121 start_eip & 0xf;
122 pr_debug("3.\n");
123}
124
125static inline void smpboot_restore_warm_reset_vector(void)
126{
127 unsigned long flags;
128
129 /*
130 * Install writable page 0 entry to set BIOS data area.
131 */
132 local_flush_tlb();
133
134 /*
135 * Paranoid: Set warm reset code and vector here back
136 * to default values.
137 */
138 spin_lock_irqsave(&rtc_lock, flags);
139 CMOS_WRITE(0, 0xf);
140 spin_unlock_irqrestore(&rtc_lock, flags);
141
142 *((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
143}
144
145/*
146 * Report back to the Boot Processor during boot time or to the caller processor
147 * during CPU online.
148 */
149static void smp_callin(void)
150{
151 int cpuid, phys_id;
152
153 /*
154 * If waken up by an INIT in an 82489DX configuration
155 * cpu_callout_mask guarantees we don't get here before
156 * an INIT_deassert IPI reaches our local APIC, so it is
157 * now safe to touch our local APIC.
158 */
159 cpuid = smp_processor_id();
160
161 /*
162 * (This works even if the APIC is not enabled.)
163 */
164 phys_id = read_apic_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 * Get our bogomips.
182 * Update loops_per_jiffy in cpu_data. Previous call to
183 * smp_store_cpu_info() stored a value that is close but not as
184 * accurate as the value just calculated.
185 */
186 calibrate_delay();
187 cpu_data(cpuid).loops_per_jiffy = loops_per_jiffy;
188 pr_debug("Stack at about %p\n", &cpuid);
189
190 /*
191 * This must be done before setting cpu_online_mask
192 * or calling notify_cpu_starting.
193 */
194 set_cpu_sibling_map(raw_smp_processor_id());
195 wmb();
196
197 notify_cpu_starting(cpuid);
198
199 /*
200 * Allow the master to continue.
201 */
202 cpumask_set_cpu(cpuid, cpu_callin_mask);
203}
204
205static int cpu0_logical_apicid;
206static int enable_start_cpu0;
207/*
208 * Activate a secondary processor.
209 */
210static void notrace start_secondary(void *unused)
211{
212 /*
213 * Don't put *anything* before cpu_init(), SMP booting is too
214 * fragile that we want to limit the things done here to the
215 * most necessary things.
216 */
217 cpu_init();
218 x86_cpuinit.early_percpu_clock_init();
219 preempt_disable();
220 smp_callin();
221
222 enable_start_cpu0 = 0;
223
224#ifdef CONFIG_X86_32
225 /* switch away from the initial page table */
226 load_cr3(swapper_pg_dir);
227 __flush_tlb_all();
228#endif
229
230 /* otherwise gcc will move up smp_processor_id before the cpu_init */
231 barrier();
232 /*
233 * Check TSC synchronization with the BP:
234 */
235 check_tsc_sync_target();
236
237 /*
238 * Lock vector_lock and initialize the vectors on this cpu
239 * before setting the cpu online. We must set it online with
240 * vector_lock held to prevent a concurrent setup/teardown
241 * from seeing a half valid vector space.
242 */
243 lock_vector_lock();
244 setup_vector_irq(smp_processor_id());
245 set_cpu_online(smp_processor_id(), true);
246 unlock_vector_lock();
247 cpu_set_state_online(smp_processor_id());
248 x86_platform.nmi_init();
249
250 /* enable local interrupts */
251 local_irq_enable();
252
253 /* to prevent fake stack check failure in clock setup */
254 boot_init_stack_canary();
255
256 x86_cpuinit.setup_percpu_clockev();
257
258 wmb();
259 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
260}
261
262int topology_update_package_map(unsigned int apicid, unsigned int cpu)
263{
264 unsigned int new, pkg = apicid >> boot_cpu_data.x86_coreid_bits;
265
266 /* Called from early boot ? */
267 if (!physical_package_map)
268 return 0;
269
270 if (pkg >= max_physical_pkg_id)
271 return -EINVAL;
272
273 /* Set the logical package id */
274 if (test_and_set_bit(pkg, physical_package_map))
275 goto found;
276
277 new = find_first_zero_bit(logical_package_map, __max_logical_packages);
278 if (new >= __max_logical_packages) {
279 physical_to_logical_pkg[pkg] = -1;
280 pr_warn("APIC(%x) Package %u exceeds logical package map\n",
281 apicid, pkg);
282 return -ENOSPC;
283 }
284 set_bit(new, logical_package_map);
285 pr_info("APIC(%x) Converting physical %u to logical package %u\n",
286 apicid, pkg, new);
287 physical_to_logical_pkg[pkg] = new;
288
289found:
290 cpu_data(cpu).logical_proc_id = physical_to_logical_pkg[pkg];
291 return 0;
292}
293
294/**
295 * topology_phys_to_logical_pkg - Map a physical package id to a logical
296 *
297 * Returns logical package id or -1 if not found
298 */
299int topology_phys_to_logical_pkg(unsigned int phys_pkg)
300{
301 if (phys_pkg >= max_physical_pkg_id)
302 return -1;
303 return physical_to_logical_pkg[phys_pkg];
304}
305EXPORT_SYMBOL(topology_phys_to_logical_pkg);
306
307static void __init smp_init_package_map(void)
308{
309 unsigned int ncpus, cpu;
310 size_t size;
311
312 /*
313 * Today neither Intel nor AMD support heterogenous systems. That
314 * might change in the future....
315 *
316 * While ideally we'd want '* smp_num_siblings' in the below @ncpus
317 * computation, this won't actually work since some Intel BIOSes
318 * report inconsistent HT data when they disable HT.
319 *
320 * In particular, they reduce the APIC-IDs to only include the cores,
321 * but leave the CPUID topology to say there are (2) siblings.
322 * This means we don't know how many threads there will be until
323 * after the APIC enumeration.
324 *
325 * By not including this we'll sometimes over-estimate the number of
326 * logical packages by the amount of !present siblings, but this is
327 * still better than MAX_LOCAL_APIC.
328 *
329 * We use total_cpus not nr_cpu_ids because nr_cpu_ids can be limited
330 * on the command line leading to a similar issue as the HT disable
331 * problem because the hyperthreads are usually enumerated after the
332 * primary cores.
333 */
334 ncpus = boot_cpu_data.x86_max_cores;
335 if (!ncpus) {
336 pr_warn("x86_max_cores == zero !?!?");
337 ncpus = 1;
338 }
339
340 __max_logical_packages = DIV_ROUND_UP(total_cpus, ncpus);
341
342 /*
343 * Possibly larger than what we need as the number of apic ids per
344 * package can be smaller than the actual used apic ids.
345 */
346 max_physical_pkg_id = DIV_ROUND_UP(MAX_LOCAL_APIC, ncpus);
347 size = max_physical_pkg_id * sizeof(unsigned int);
348 physical_to_logical_pkg = kmalloc(size, GFP_KERNEL);
349 memset(physical_to_logical_pkg, 0xff, size);
350 size = BITS_TO_LONGS(max_physical_pkg_id) * sizeof(unsigned long);
351 physical_package_map = kzalloc(size, GFP_KERNEL);
352 size = BITS_TO_LONGS(__max_logical_packages) * sizeof(unsigned long);
353 logical_package_map = kzalloc(size, GFP_KERNEL);
354
355 pr_info("Max logical packages: %u\n", __max_logical_packages);
356
357 for_each_present_cpu(cpu) {
358 unsigned int apicid = apic->cpu_present_to_apicid(cpu);
359
360 if (apicid == BAD_APICID || !apic->apic_id_valid(apicid))
361 continue;
362 if (!topology_update_package_map(apicid, cpu))
363 continue;
364 pr_warn("CPU %u APICId %x disabled\n", cpu, apicid);
365 per_cpu(x86_bios_cpu_apicid, cpu) = BAD_APICID;
366 set_cpu_possible(cpu, false);
367 set_cpu_present(cpu, false);
368 }
369}
370
371void __init smp_store_boot_cpu_info(void)
372{
373 int id = 0; /* CPU 0 */
374 struct cpuinfo_x86 *c = &cpu_data(id);
375
376 *c = boot_cpu_data;
377 c->cpu_index = id;
378 smp_init_package_map();
379}
380
381/*
382 * The bootstrap kernel entry code has set these up. Save them for
383 * a given CPU
384 */
385void smp_store_cpu_info(int id)
386{
387 struct cpuinfo_x86 *c = &cpu_data(id);
388
389 *c = boot_cpu_data;
390 c->cpu_index = id;
391 /*
392 * During boot time, CPU0 has this setup already. Save the info when
393 * bringing up AP or offlined CPU0.
394 */
395 identify_secondary_cpu(c);
396}
397
398static bool
399topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
400{
401 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
402
403 return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
404}
405
406static bool
407topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
408{
409 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
410
411 return !WARN_ONCE(!topology_same_node(c, o),
412 "sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
413 "[node: %d != %d]. Ignoring dependency.\n",
414 cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
415}
416
417#define link_mask(mfunc, c1, c2) \
418do { \
419 cpumask_set_cpu((c1), mfunc(c2)); \
420 cpumask_set_cpu((c2), mfunc(c1)); \
421} while (0)
422
423static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
424{
425 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
426 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
427
428 if (c->phys_proc_id == o->phys_proc_id &&
429 per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2) &&
430 c->cpu_core_id == o->cpu_core_id)
431 return topology_sane(c, o, "smt");
432
433 } else if (c->phys_proc_id == o->phys_proc_id &&
434 c->cpu_core_id == o->cpu_core_id) {
435 return topology_sane(c, o, "smt");
436 }
437
438 return false;
439}
440
441static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
442{
443 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
444
445 if (per_cpu(cpu_llc_id, cpu1) != BAD_APICID &&
446 per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2))
447 return topology_sane(c, o, "llc");
448
449 return false;
450}
451
452/*
453 * Unlike the other levels, we do not enforce keeping a
454 * multicore group inside a NUMA node. If this happens, we will
455 * discard the MC level of the topology later.
456 */
457static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
458{
459 if (c->phys_proc_id == o->phys_proc_id)
460 return true;
461 return false;
462}
463
464static struct sched_domain_topology_level numa_inside_package_topology[] = {
465#ifdef CONFIG_SCHED_SMT
466 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
467#endif
468#ifdef CONFIG_SCHED_MC
469 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
470#endif
471 { NULL, },
472};
473/*
474 * set_sched_topology() sets the topology internal to a CPU. The
475 * NUMA topologies are layered on top of it to build the full
476 * system topology.
477 *
478 * If NUMA nodes are observed to occur within a CPU package, this
479 * function should be called. It forces the sched domain code to
480 * only use the SMT level for the CPU portion of the topology.
481 * This essentially falls back to relying on NUMA information
482 * from the SRAT table to describe the entire system topology
483 * (except for hyperthreads).
484 */
485static void primarily_use_numa_for_topology(void)
486{
487 set_sched_topology(numa_inside_package_topology);
488}
489
490void set_cpu_sibling_map(int cpu)
491{
492 bool has_smt = smp_num_siblings > 1;
493 bool has_mp = has_smt || boot_cpu_data.x86_max_cores > 1;
494 struct cpuinfo_x86 *c = &cpu_data(cpu);
495 struct cpuinfo_x86 *o;
496 int i;
497
498 cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
499
500 if (!has_mp) {
501 cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
502 cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
503 cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
504 c->booted_cores = 1;
505 return;
506 }
507
508 for_each_cpu(i, cpu_sibling_setup_mask) {
509 o = &cpu_data(i);
510
511 if ((i == cpu) || (has_smt && match_smt(c, o)))
512 link_mask(topology_sibling_cpumask, cpu, i);
513
514 if ((i == cpu) || (has_mp && match_llc(c, o)))
515 link_mask(cpu_llc_shared_mask, cpu, i);
516
517 }
518
519 /*
520 * This needs a separate iteration over the cpus because we rely on all
521 * topology_sibling_cpumask links to be set-up.
522 */
523 for_each_cpu(i, cpu_sibling_setup_mask) {
524 o = &cpu_data(i);
525
526 if ((i == cpu) || (has_mp && match_die(c, o))) {
527 link_mask(topology_core_cpumask, cpu, i);
528
529 /*
530 * Does this new cpu bringup a new core?
531 */
532 if (cpumask_weight(
533 topology_sibling_cpumask(cpu)) == 1) {
534 /*
535 * for each core in package, increment
536 * the booted_cores for this new cpu
537 */
538 if (cpumask_first(
539 topology_sibling_cpumask(i)) == i)
540 c->booted_cores++;
541 /*
542 * increment the core count for all
543 * the other cpus in this package
544 */
545 if (i != cpu)
546 cpu_data(i).booted_cores++;
547 } else if (i != cpu && !c->booted_cores)
548 c->booted_cores = cpu_data(i).booted_cores;
549 }
550 if (match_die(c, o) && !topology_same_node(c, o))
551 primarily_use_numa_for_topology();
552 }
553}
554
555/* maps the cpu to the sched domain representing multi-core */
556const struct cpumask *cpu_coregroup_mask(int cpu)
557{
558 return cpu_llc_shared_mask(cpu);
559}
560
561static void impress_friends(void)
562{
563 int cpu;
564 unsigned long bogosum = 0;
565 /*
566 * Allow the user to impress friends.
567 */
568 pr_debug("Before bogomips\n");
569 for_each_possible_cpu(cpu)
570 if (cpumask_test_cpu(cpu, cpu_callout_mask))
571 bogosum += cpu_data(cpu).loops_per_jiffy;
572 pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
573 num_online_cpus(),
574 bogosum/(500000/HZ),
575 (bogosum/(5000/HZ))%100);
576
577 pr_debug("Before bogocount - setting activated=1\n");
578}
579
580void __inquire_remote_apic(int apicid)
581{
582 unsigned i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
583 const char * const names[] = { "ID", "VERSION", "SPIV" };
584 int timeout;
585 u32 status;
586
587 pr_info("Inquiring remote APIC 0x%x...\n", apicid);
588
589 for (i = 0; i < ARRAY_SIZE(regs); i++) {
590 pr_info("... APIC 0x%x %s: ", apicid, names[i]);
591
592 /*
593 * Wait for idle.
594 */
595 status = safe_apic_wait_icr_idle();
596 if (status)
597 pr_cont("a previous APIC delivery may have failed\n");
598
599 apic_icr_write(APIC_DM_REMRD | regs[i], apicid);
600
601 timeout = 0;
602 do {
603 udelay(100);
604 status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
605 } while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
606
607 switch (status) {
608 case APIC_ICR_RR_VALID:
609 status = apic_read(APIC_RRR);
610 pr_cont("%08x\n", status);
611 break;
612 default:
613 pr_cont("failed\n");
614 }
615 }
616}
617
618/*
619 * The Multiprocessor Specification 1.4 (1997) example code suggests
620 * that there should be a 10ms delay between the BSP asserting INIT
621 * and de-asserting INIT, when starting a remote processor.
622 * But that slows boot and resume on modern processors, which include
623 * many cores and don't require that delay.
624 *
625 * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
626 * Modern processor families are quirked to remove the delay entirely.
627 */
628#define UDELAY_10MS_DEFAULT 10000
629
630static unsigned int init_udelay = UINT_MAX;
631
632static int __init cpu_init_udelay(char *str)
633{
634 get_option(&str, &init_udelay);
635
636 return 0;
637}
638early_param("cpu_init_udelay", cpu_init_udelay);
639
640static void __init smp_quirk_init_udelay(void)
641{
642 /* if cmdline changed it from default, leave it alone */
643 if (init_udelay != UINT_MAX)
644 return;
645
646 /* if modern processor, use no delay */
647 if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
648 ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
649 init_udelay = 0;
650 return;
651 }
652 /* else, use legacy delay */
653 init_udelay = UDELAY_10MS_DEFAULT;
654}
655
656/*
657 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
658 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
659 * won't ... remember to clear down the APIC, etc later.
660 */
661int
662wakeup_secondary_cpu_via_nmi(int apicid, unsigned long start_eip)
663{
664 unsigned long send_status, accept_status = 0;
665 int maxlvt;
666
667 /* Target chip */
668 /* Boot on the stack */
669 /* Kick the second */
670 apic_icr_write(APIC_DM_NMI | apic->dest_logical, apicid);
671
672 pr_debug("Waiting for send to finish...\n");
673 send_status = safe_apic_wait_icr_idle();
674
675 /*
676 * Give the other CPU some time to accept the IPI.
677 */
678 udelay(200);
679 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid])) {
680 maxlvt = lapic_get_maxlvt();
681 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
682 apic_write(APIC_ESR, 0);
683 accept_status = (apic_read(APIC_ESR) & 0xEF);
684 }
685 pr_debug("NMI sent\n");
686
687 if (send_status)
688 pr_err("APIC never delivered???\n");
689 if (accept_status)
690 pr_err("APIC delivery error (%lx)\n", accept_status);
691
692 return (send_status | accept_status);
693}
694
695static int
696wakeup_secondary_cpu_via_init(int phys_apicid, unsigned long start_eip)
697{
698 unsigned long send_status = 0, accept_status = 0;
699 int maxlvt, num_starts, j;
700
701 maxlvt = lapic_get_maxlvt();
702
703 /*
704 * Be paranoid about clearing APIC errors.
705 */
706 if (APIC_INTEGRATED(apic_version[phys_apicid])) {
707 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
708 apic_write(APIC_ESR, 0);
709 apic_read(APIC_ESR);
710 }
711
712 pr_debug("Asserting INIT\n");
713
714 /*
715 * Turn INIT on target chip
716 */
717 /*
718 * Send IPI
719 */
720 apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT,
721 phys_apicid);
722
723 pr_debug("Waiting for send to finish...\n");
724 send_status = safe_apic_wait_icr_idle();
725
726 udelay(init_udelay);
727
728 pr_debug("Deasserting INIT\n");
729
730 /* Target chip */
731 /* Send IPI */
732 apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
733
734 pr_debug("Waiting for send to finish...\n");
735 send_status = safe_apic_wait_icr_idle();
736
737 mb();
738
739 /*
740 * Should we send STARTUP IPIs ?
741 *
742 * Determine this based on the APIC version.
743 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
744 */
745 if (APIC_INTEGRATED(apic_version[phys_apicid]))
746 num_starts = 2;
747 else
748 num_starts = 0;
749
750 /*
751 * Run STARTUP IPI loop.
752 */
753 pr_debug("#startup loops: %d\n", num_starts);
754
755 for (j = 1; j <= num_starts; j++) {
756 pr_debug("Sending STARTUP #%d\n", j);
757 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
758 apic_write(APIC_ESR, 0);
759 apic_read(APIC_ESR);
760 pr_debug("After apic_write\n");
761
762 /*
763 * STARTUP IPI
764 */
765
766 /* Target chip */
767 /* Boot on the stack */
768 /* Kick the second */
769 apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
770 phys_apicid);
771
772 /*
773 * Give the other CPU some time to accept the IPI.
774 */
775 if (init_udelay == 0)
776 udelay(10);
777 else
778 udelay(300);
779
780 pr_debug("Startup point 1\n");
781
782 pr_debug("Waiting for send to finish...\n");
783 send_status = safe_apic_wait_icr_idle();
784
785 /*
786 * Give the other CPU some time to accept the IPI.
787 */
788 if (init_udelay == 0)
789 udelay(10);
790 else
791 udelay(200);
792
793 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
794 apic_write(APIC_ESR, 0);
795 accept_status = (apic_read(APIC_ESR) & 0xEF);
796 if (send_status || accept_status)
797 break;
798 }
799 pr_debug("After Startup\n");
800
801 if (send_status)
802 pr_err("APIC never delivered???\n");
803 if (accept_status)
804 pr_err("APIC delivery error (%lx)\n", accept_status);
805
806 return (send_status | accept_status);
807}
808
809void smp_announce(void)
810{
811 int num_nodes = num_online_nodes();
812
813 printk(KERN_INFO "x86: Booted up %d node%s, %d CPUs\n",
814 num_nodes, (num_nodes > 1 ? "s" : ""), num_online_cpus());
815}
816
817/* reduce the number of lines printed when booting a large cpu count system */
818static void announce_cpu(int cpu, int apicid)
819{
820 static int current_node = -1;
821 int node = early_cpu_to_node(cpu);
822 static int width, node_width;
823
824 if (!width)
825 width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
826
827 if (!node_width)
828 node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
829
830 if (cpu == 1)
831 printk(KERN_INFO "x86: Booting SMP configuration:\n");
832
833 if (system_state == SYSTEM_BOOTING) {
834 if (node != current_node) {
835 if (current_node > (-1))
836 pr_cont("\n");
837 current_node = node;
838
839 printk(KERN_INFO ".... node %*s#%d, CPUs: ",
840 node_width - num_digits(node), " ", node);
841 }
842
843 /* Add padding for the BSP */
844 if (cpu == 1)
845 pr_cont("%*s", width + 1, " ");
846
847 pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
848
849 } else
850 pr_info("Booting Node %d Processor %d APIC 0x%x\n",
851 node, cpu, apicid);
852}
853
854static int wakeup_cpu0_nmi(unsigned int cmd, struct pt_regs *regs)
855{
856 int cpu;
857
858 cpu = smp_processor_id();
859 if (cpu == 0 && !cpu_online(cpu) && enable_start_cpu0)
860 return NMI_HANDLED;
861
862 return NMI_DONE;
863}
864
865/*
866 * Wake up AP by INIT, INIT, STARTUP sequence.
867 *
868 * Instead of waiting for STARTUP after INITs, BSP will execute the BIOS
869 * boot-strap code which is not a desired behavior for waking up BSP. To
870 * void the boot-strap code, wake up CPU0 by NMI instead.
871 *
872 * This works to wake up soft offlined CPU0 only. If CPU0 is hard offlined
873 * (i.e. physically hot removed and then hot added), NMI won't wake it up.
874 * We'll change this code in the future to wake up hard offlined CPU0 if
875 * real platform and request are available.
876 */
877static int
878wakeup_cpu_via_init_nmi(int cpu, unsigned long start_ip, int apicid,
879 int *cpu0_nmi_registered)
880{
881 int id;
882 int boot_error;
883
884 preempt_disable();
885
886 /*
887 * Wake up AP by INIT, INIT, STARTUP sequence.
888 */
889 if (cpu) {
890 boot_error = wakeup_secondary_cpu_via_init(apicid, start_ip);
891 goto out;
892 }
893
894 /*
895 * Wake up BSP by nmi.
896 *
897 * Register a NMI handler to help wake up CPU0.
898 */
899 boot_error = register_nmi_handler(NMI_LOCAL,
900 wakeup_cpu0_nmi, 0, "wake_cpu0");
901
902 if (!boot_error) {
903 enable_start_cpu0 = 1;
904 *cpu0_nmi_registered = 1;
905 if (apic->dest_logical == APIC_DEST_LOGICAL)
906 id = cpu0_logical_apicid;
907 else
908 id = apicid;
909 boot_error = wakeup_secondary_cpu_via_nmi(id, start_ip);
910 }
911
912out:
913 preempt_enable();
914
915 return boot_error;
916}
917
918void common_cpu_up(unsigned int cpu, struct task_struct *idle)
919{
920 /* Just in case we booted with a single CPU. */
921 alternatives_enable_smp();
922
923 per_cpu(current_task, cpu) = idle;
924
925#ifdef CONFIG_X86_32
926 /* Stack for startup_32 can be just as for start_secondary onwards */
927 irq_ctx_init(cpu);
928 per_cpu(cpu_current_top_of_stack, cpu) =
929 (unsigned long)task_stack_page(idle) + THREAD_SIZE;
930#else
931 clear_tsk_thread_flag(idle, TIF_FORK);
932 initial_gs = per_cpu_offset(cpu);
933#endif
934}
935
936/*
937 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
938 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
939 * Returns zero if CPU booted OK, else error code from
940 * ->wakeup_secondary_cpu.
941 */
942static int do_boot_cpu(int apicid, int cpu, struct task_struct *idle)
943{
944 volatile u32 *trampoline_status =
945 (volatile u32 *) __va(real_mode_header->trampoline_status);
946 /* start_ip had better be page-aligned! */
947 unsigned long start_ip = real_mode_header->trampoline_start;
948
949 unsigned long boot_error = 0;
950 int cpu0_nmi_registered = 0;
951 unsigned long timeout;
952
953 idle->thread.sp = (unsigned long) (((struct pt_regs *)
954 (THREAD_SIZE + task_stack_page(idle))) - 1);
955
956 early_gdt_descr.address = (unsigned long)get_cpu_gdt_table(cpu);
957 initial_code = (unsigned long)start_secondary;
958 stack_start = idle->thread.sp;
959
960 /*
961 * Enable the espfix hack for this CPU
962 */
963#ifdef CONFIG_X86_ESPFIX64
964 init_espfix_ap(cpu);
965#endif
966
967 /* So we see what's up */
968 announce_cpu(cpu, apicid);
969
970 /*
971 * This grunge runs the startup process for
972 * the targeted processor.
973 */
974
975 if (get_uv_system_type() != UV_NON_UNIQUE_APIC) {
976
977 pr_debug("Setting warm reset code and vector.\n");
978
979 smpboot_setup_warm_reset_vector(start_ip);
980 /*
981 * Be paranoid about clearing APIC errors.
982 */
983 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid])) {
984 apic_write(APIC_ESR, 0);
985 apic_read(APIC_ESR);
986 }
987 }
988
989 /*
990 * AP might wait on cpu_callout_mask in cpu_init() with
991 * cpu_initialized_mask set if previous attempt to online
992 * it timed-out. Clear cpu_initialized_mask so that after
993 * INIT/SIPI it could start with a clean state.
994 */
995 cpumask_clear_cpu(cpu, cpu_initialized_mask);
996 smp_mb();
997
998 /*
999 * Wake up a CPU in difference cases:
1000 * - Use the method in the APIC driver if it's defined
1001 * Otherwise,
1002 * - Use an INIT boot APIC message for APs or NMI for BSP.
1003 */
1004 if (apic->wakeup_secondary_cpu)
1005 boot_error = apic->wakeup_secondary_cpu(apicid, start_ip);
1006 else
1007 boot_error = wakeup_cpu_via_init_nmi(cpu, start_ip, apicid,
1008 &cpu0_nmi_registered);
1009
1010 if (!boot_error) {
1011 /*
1012 * Wait 10s total for first sign of life from AP
1013 */
1014 boot_error = -1;
1015 timeout = jiffies + 10*HZ;
1016 while (time_before(jiffies, timeout)) {
1017 if (cpumask_test_cpu(cpu, cpu_initialized_mask)) {
1018 /*
1019 * Tell AP to proceed with initialization
1020 */
1021 cpumask_set_cpu(cpu, cpu_callout_mask);
1022 boot_error = 0;
1023 break;
1024 }
1025 schedule();
1026 }
1027 }
1028
1029 if (!boot_error) {
1030 /*
1031 * Wait till AP completes initial initialization
1032 */
1033 while (!cpumask_test_cpu(cpu, cpu_callin_mask)) {
1034 /*
1035 * Allow other tasks to run while we wait for the
1036 * AP to come online. This also gives a chance
1037 * for the MTRR work(triggered by the AP coming online)
1038 * to be completed in the stop machine context.
1039 */
1040 schedule();
1041 }
1042 }
1043
1044 /* mark "stuck" area as not stuck */
1045 *trampoline_status = 0;
1046
1047 if (get_uv_system_type() != UV_NON_UNIQUE_APIC) {
1048 /*
1049 * Cleanup possible dangling ends...
1050 */
1051 smpboot_restore_warm_reset_vector();
1052 }
1053 /*
1054 * Clean up the nmi handler. Do this after the callin and callout sync
1055 * to avoid impact of possible long unregister time.
1056 */
1057 if (cpu0_nmi_registered)
1058 unregister_nmi_handler(NMI_LOCAL, "wake_cpu0");
1059
1060 return boot_error;
1061}
1062
1063int native_cpu_up(unsigned int cpu, struct task_struct *tidle)
1064{
1065 int apicid = apic->cpu_present_to_apicid(cpu);
1066 unsigned long flags;
1067 int err;
1068
1069 WARN_ON(irqs_disabled());
1070
1071 pr_debug("++++++++++++++++++++=_---CPU UP %u\n", cpu);
1072
1073 if (apicid == BAD_APICID ||
1074 !physid_isset(apicid, phys_cpu_present_map) ||
1075 !apic->apic_id_valid(apicid)) {
1076 pr_err("%s: bad cpu %d\n", __func__, cpu);
1077 return -EINVAL;
1078 }
1079
1080 /*
1081 * Already booted CPU?
1082 */
1083 if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
1084 pr_debug("do_boot_cpu %d Already started\n", cpu);
1085 return -ENOSYS;
1086 }
1087
1088 /*
1089 * Save current MTRR state in case it was changed since early boot
1090 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1091 */
1092 mtrr_save_state();
1093
1094 /* x86 CPUs take themselves offline, so delayed offline is OK. */
1095 err = cpu_check_up_prepare(cpu);
1096 if (err && err != -EBUSY)
1097 return err;
1098
1099 /* the FPU context is blank, nobody can own it */
1100 __cpu_disable_lazy_restore(cpu);
1101
1102 common_cpu_up(cpu, tidle);
1103
1104 /*
1105 * We have to walk the irq descriptors to setup the vector
1106 * space for the cpu which comes online. Prevent irq
1107 * alloc/free across the bringup.
1108 */
1109 irq_lock_sparse();
1110
1111 err = do_boot_cpu(apicid, cpu, tidle);
1112
1113 if (err) {
1114 irq_unlock_sparse();
1115 pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
1116 return -EIO;
1117 }
1118
1119 /*
1120 * Check TSC synchronization with the AP (keep irqs disabled
1121 * while doing so):
1122 */
1123 local_irq_save(flags);
1124 check_tsc_sync_source(cpu);
1125 local_irq_restore(flags);
1126
1127 while (!cpu_online(cpu)) {
1128 cpu_relax();
1129 touch_nmi_watchdog();
1130 }
1131
1132 irq_unlock_sparse();
1133
1134 return 0;
1135}
1136
1137/**
1138 * arch_disable_smp_support() - disables SMP support for x86 at runtime
1139 */
1140void arch_disable_smp_support(void)
1141{
1142 disable_ioapic_support();
1143}
1144
1145/*
1146 * Fall back to non SMP mode after errors.
1147 *
1148 * RED-PEN audit/test this more. I bet there is more state messed up here.
1149 */
1150static __init void disable_smp(void)
1151{
1152 pr_info("SMP disabled\n");
1153
1154 disable_ioapic_support();
1155
1156 init_cpu_present(cpumask_of(0));
1157 init_cpu_possible(cpumask_of(0));
1158
1159 if (smp_found_config)
1160 physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
1161 else
1162 physid_set_mask_of_physid(0, &phys_cpu_present_map);
1163 cpumask_set_cpu(0, topology_sibling_cpumask(0));
1164 cpumask_set_cpu(0, topology_core_cpumask(0));
1165}
1166
1167enum {
1168 SMP_OK,
1169 SMP_NO_CONFIG,
1170 SMP_NO_APIC,
1171 SMP_FORCE_UP,
1172};
1173
1174/*
1175 * Various sanity checks.
1176 */
1177static int __init smp_sanity_check(unsigned max_cpus)
1178{
1179 preempt_disable();
1180
1181#if !defined(CONFIG_X86_BIGSMP) && defined(CONFIG_X86_32)
1182 if (def_to_bigsmp && nr_cpu_ids > 8) {
1183 unsigned int cpu;
1184 unsigned nr;
1185
1186 pr_warn("More than 8 CPUs detected - skipping them\n"
1187 "Use CONFIG_X86_BIGSMP\n");
1188
1189 nr = 0;
1190 for_each_present_cpu(cpu) {
1191 if (nr >= 8)
1192 set_cpu_present(cpu, false);
1193 nr++;
1194 }
1195
1196 nr = 0;
1197 for_each_possible_cpu(cpu) {
1198 if (nr >= 8)
1199 set_cpu_possible(cpu, false);
1200 nr++;
1201 }
1202
1203 nr_cpu_ids = 8;
1204 }
1205#endif
1206
1207 if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map)) {
1208 pr_warn("weird, boot CPU (#%d) not listed by the BIOS\n",
1209 hard_smp_processor_id());
1210
1211 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1212 }
1213
1214 /*
1215 * If we couldn't find an SMP configuration at boot time,
1216 * get out of here now!
1217 */
1218 if (!smp_found_config && !acpi_lapic) {
1219 preempt_enable();
1220 pr_notice("SMP motherboard not detected\n");
1221 return SMP_NO_CONFIG;
1222 }
1223
1224 /*
1225 * Should not be necessary because the MP table should list the boot
1226 * CPU too, but we do it for the sake of robustness anyway.
1227 */
1228 if (!apic->check_phys_apicid_present(boot_cpu_physical_apicid)) {
1229 pr_notice("weird, boot CPU (#%d) not listed by the BIOS\n",
1230 boot_cpu_physical_apicid);
1231 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1232 }
1233 preempt_enable();
1234
1235 /*
1236 * If we couldn't find a local APIC, then get out of here now!
1237 */
1238 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid]) &&
1239 !cpu_has_apic) {
1240 if (!disable_apic) {
1241 pr_err("BIOS bug, local APIC #%d not detected!...\n",
1242 boot_cpu_physical_apicid);
1243 pr_err("... forcing use of dummy APIC emulation (tell your hw vendor)\n");
1244 }
1245 return SMP_NO_APIC;
1246 }
1247
1248 /*
1249 * If SMP should be disabled, then really disable it!
1250 */
1251 if (!max_cpus) {
1252 pr_info("SMP mode deactivated\n");
1253 return SMP_FORCE_UP;
1254 }
1255
1256 return SMP_OK;
1257}
1258
1259static void __init smp_cpu_index_default(void)
1260{
1261 int i;
1262 struct cpuinfo_x86 *c;
1263
1264 for_each_possible_cpu(i) {
1265 c = &cpu_data(i);
1266 /* mark all to hotplug */
1267 c->cpu_index = nr_cpu_ids;
1268 }
1269}
1270
1271/*
1272 * Prepare for SMP bootup. The MP table or ACPI has been read
1273 * earlier. Just do some sanity checking here and enable APIC mode.
1274 */
1275void __init native_smp_prepare_cpus(unsigned int max_cpus)
1276{
1277 unsigned int i;
1278
1279 smp_cpu_index_default();
1280
1281 /*
1282 * Setup boot CPU information
1283 */
1284 smp_store_boot_cpu_info(); /* Final full version of the data */
1285 cpumask_copy(cpu_callin_mask, cpumask_of(0));
1286 mb();
1287
1288 current_thread_info()->cpu = 0; /* needed? */
1289 for_each_possible_cpu(i) {
1290 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1291 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1292 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1293 }
1294 set_cpu_sibling_map(0);
1295
1296 switch (smp_sanity_check(max_cpus)) {
1297 case SMP_NO_CONFIG:
1298 disable_smp();
1299 if (APIC_init_uniprocessor())
1300 pr_notice("Local APIC not detected. Using dummy APIC emulation.\n");
1301 return;
1302 case SMP_NO_APIC:
1303 disable_smp();
1304 return;
1305 case SMP_FORCE_UP:
1306 disable_smp();
1307 apic_bsp_setup(false);
1308 return;
1309 case SMP_OK:
1310 break;
1311 }
1312
1313 default_setup_apic_routing();
1314
1315 if (read_apic_id() != boot_cpu_physical_apicid) {
1316 panic("Boot APIC ID in local APIC unexpected (%d vs %d)",
1317 read_apic_id(), boot_cpu_physical_apicid);
1318 /* Or can we switch back to PIC here? */
1319 }
1320
1321 cpu0_logical_apicid = apic_bsp_setup(false);
1322
1323 pr_info("CPU%d: ", 0);
1324 print_cpu_info(&cpu_data(0));
1325
1326 if (is_uv_system())
1327 uv_system_init();
1328
1329 set_mtrr_aps_delayed_init();
1330
1331 smp_quirk_init_udelay();
1332}
1333
1334void arch_enable_nonboot_cpus_begin(void)
1335{
1336 set_mtrr_aps_delayed_init();
1337}
1338
1339void arch_enable_nonboot_cpus_end(void)
1340{
1341 mtrr_aps_init();
1342}
1343
1344/*
1345 * Early setup to make printk work.
1346 */
1347void __init native_smp_prepare_boot_cpu(void)
1348{
1349 int me = smp_processor_id();
1350 switch_to_new_gdt(me);
1351 /* already set me in cpu_online_mask in boot_cpu_init() */
1352 cpumask_set_cpu(me, cpu_callout_mask);
1353 cpu_set_state_online(me);
1354}
1355
1356void __init native_smp_cpus_done(unsigned int max_cpus)
1357{
1358 pr_debug("Boot done\n");
1359
1360 nmi_selftest();
1361 impress_friends();
1362 setup_ioapic_dest();
1363 mtrr_aps_init();
1364}
1365
1366static int __initdata setup_possible_cpus = -1;
1367static int __init _setup_possible_cpus(char *str)
1368{
1369 get_option(&str, &setup_possible_cpus);
1370 return 0;
1371}
1372early_param("possible_cpus", _setup_possible_cpus);
1373
1374
1375/*
1376 * cpu_possible_mask should be static, it cannot change as cpu's
1377 * are onlined, or offlined. The reason is per-cpu data-structures
1378 * are allocated by some modules at init time, and dont expect to
1379 * do this dynamically on cpu arrival/departure.
1380 * cpu_present_mask on the other hand can change dynamically.
1381 * In case when cpu_hotplug is not compiled, then we resort to current
1382 * behaviour, which is cpu_possible == cpu_present.
1383 * - Ashok Raj
1384 *
1385 * Three ways to find out the number of additional hotplug CPUs:
1386 * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1387 * - The user can overwrite it with possible_cpus=NUM
1388 * - Otherwise don't reserve additional CPUs.
1389 * We do this because additional CPUs waste a lot of memory.
1390 * -AK
1391 */
1392__init void prefill_possible_map(void)
1393{
1394 int i, possible;
1395
1396 /* no processor from mptable or madt */
1397 if (!num_processors)
1398 num_processors = 1;
1399
1400 i = setup_max_cpus ?: 1;
1401 if (setup_possible_cpus == -1) {
1402 possible = num_processors;
1403#ifdef CONFIG_HOTPLUG_CPU
1404 if (setup_max_cpus)
1405 possible += disabled_cpus;
1406#else
1407 if (possible > i)
1408 possible = i;
1409#endif
1410 } else
1411 possible = setup_possible_cpus;
1412
1413 total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1414
1415 /* nr_cpu_ids could be reduced via nr_cpus= */
1416 if (possible > nr_cpu_ids) {
1417 pr_warn("%d Processors exceeds NR_CPUS limit of %d\n",
1418 possible, nr_cpu_ids);
1419 possible = nr_cpu_ids;
1420 }
1421
1422#ifdef CONFIG_HOTPLUG_CPU
1423 if (!setup_max_cpus)
1424#endif
1425 if (possible > i) {
1426 pr_warn("%d Processors exceeds max_cpus limit of %u\n",
1427 possible, setup_max_cpus);
1428 possible = i;
1429 }
1430
1431 pr_info("Allowing %d CPUs, %d hotplug CPUs\n",
1432 possible, max_t(int, possible - num_processors, 0));
1433
1434 for (i = 0; i < possible; i++)
1435 set_cpu_possible(i, true);
1436 for (; i < NR_CPUS; i++)
1437 set_cpu_possible(i, false);
1438
1439 nr_cpu_ids = possible;
1440}
1441
1442#ifdef CONFIG_HOTPLUG_CPU
1443
1444static void remove_siblinginfo(int cpu)
1445{
1446 int sibling;
1447 struct cpuinfo_x86 *c = &cpu_data(cpu);
1448
1449 for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1450 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1451 /*/
1452 * last thread sibling in this cpu core going down
1453 */
1454 if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1455 cpu_data(sibling).booted_cores--;
1456 }
1457
1458 for_each_cpu(sibling, topology_sibling_cpumask(cpu))
1459 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1460 for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1461 cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1462 cpumask_clear(cpu_llc_shared_mask(cpu));
1463 cpumask_clear(topology_sibling_cpumask(cpu));
1464 cpumask_clear(topology_core_cpumask(cpu));
1465 c->phys_proc_id = 0;
1466 c->cpu_core_id = 0;
1467 cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1468}
1469
1470static void remove_cpu_from_maps(int cpu)
1471{
1472 set_cpu_online(cpu, false);
1473 cpumask_clear_cpu(cpu, cpu_callout_mask);
1474 cpumask_clear_cpu(cpu, cpu_callin_mask);
1475 /* was set by cpu_init() */
1476 cpumask_clear_cpu(cpu, cpu_initialized_mask);
1477 numa_remove_cpu(cpu);
1478}
1479
1480void cpu_disable_common(void)
1481{
1482 int cpu = smp_processor_id();
1483
1484 remove_siblinginfo(cpu);
1485
1486 /* It's now safe to remove this processor from the online map */
1487 lock_vector_lock();
1488 remove_cpu_from_maps(cpu);
1489 unlock_vector_lock();
1490 fixup_irqs();
1491}
1492
1493int native_cpu_disable(void)
1494{
1495 int ret;
1496
1497 ret = check_irq_vectors_for_cpu_disable();
1498 if (ret)
1499 return ret;
1500
1501 clear_local_APIC();
1502 cpu_disable_common();
1503
1504 return 0;
1505}
1506
1507int common_cpu_die(unsigned int cpu)
1508{
1509 int ret = 0;
1510
1511 /* We don't do anything here: idle task is faking death itself. */
1512
1513 /* They ack this in play_dead() by setting CPU_DEAD */
1514 if (cpu_wait_death(cpu, 5)) {
1515 if (system_state == SYSTEM_RUNNING)
1516 pr_info("CPU %u is now offline\n", cpu);
1517 } else {
1518 pr_err("CPU %u didn't die...\n", cpu);
1519 ret = -1;
1520 }
1521
1522 return ret;
1523}
1524
1525void native_cpu_die(unsigned int cpu)
1526{
1527 common_cpu_die(cpu);
1528}
1529
1530void play_dead_common(void)
1531{
1532 idle_task_exit();
1533 reset_lazy_tlbstate();
1534 amd_e400_remove_cpu(raw_smp_processor_id());
1535
1536 /* Ack it */
1537 (void)cpu_report_death();
1538
1539 /*
1540 * With physical CPU hotplug, we should halt the cpu
1541 */
1542 local_irq_disable();
1543}
1544
1545static bool wakeup_cpu0(void)
1546{
1547 if (smp_processor_id() == 0 && enable_start_cpu0)
1548 return true;
1549
1550 return false;
1551}
1552
1553/*
1554 * We need to flush the caches before going to sleep, lest we have
1555 * dirty data in our caches when we come back up.
1556 */
1557static inline void mwait_play_dead(void)
1558{
1559 unsigned int eax, ebx, ecx, edx;
1560 unsigned int highest_cstate = 0;
1561 unsigned int highest_subcstate = 0;
1562 void *mwait_ptr;
1563 int i;
1564
1565 if (!this_cpu_has(X86_FEATURE_MWAIT))
1566 return;
1567 if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1568 return;
1569 if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1570 return;
1571
1572 eax = CPUID_MWAIT_LEAF;
1573 ecx = 0;
1574 native_cpuid(&eax, &ebx, &ecx, &edx);
1575
1576 /*
1577 * eax will be 0 if EDX enumeration is not valid.
1578 * Initialized below to cstate, sub_cstate value when EDX is valid.
1579 */
1580 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1581 eax = 0;
1582 } else {
1583 edx >>= MWAIT_SUBSTATE_SIZE;
1584 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1585 if (edx & MWAIT_SUBSTATE_MASK) {
1586 highest_cstate = i;
1587 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1588 }
1589 }
1590 eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1591 (highest_subcstate - 1);
1592 }
1593
1594 /*
1595 * This should be a memory location in a cache line which is
1596 * unlikely to be touched by other processors. The actual
1597 * content is immaterial as it is not actually modified in any way.
1598 */
1599 mwait_ptr = ¤t_thread_info()->flags;
1600
1601 wbinvd();
1602
1603 while (1) {
1604 /*
1605 * The CLFLUSH is a workaround for erratum AAI65 for
1606 * the Xeon 7400 series. It's not clear it is actually
1607 * needed, but it should be harmless in either case.
1608 * The WBINVD is insufficient due to the spurious-wakeup
1609 * case where we return around the loop.
1610 */
1611 mb();
1612 clflush(mwait_ptr);
1613 mb();
1614 __monitor(mwait_ptr, 0, 0);
1615 mb();
1616 __mwait(eax, 0);
1617 /*
1618 * If NMI wants to wake up CPU0, start CPU0.
1619 */
1620 if (wakeup_cpu0())
1621 start_cpu0();
1622 }
1623}
1624
1625static inline void hlt_play_dead(void)
1626{
1627 if (__this_cpu_read(cpu_info.x86) >= 4)
1628 wbinvd();
1629
1630 while (1) {
1631 native_halt();
1632 /*
1633 * If NMI wants to wake up CPU0, start CPU0.
1634 */
1635 if (wakeup_cpu0())
1636 start_cpu0();
1637 }
1638}
1639
1640void native_play_dead(void)
1641{
1642 play_dead_common();
1643 tboot_shutdown(TB_SHUTDOWN_WFS);
1644
1645 mwait_play_dead(); /* Only returns on failure */
1646 if (cpuidle_play_dead())
1647 hlt_play_dead();
1648}
1649
1650#else /* ... !CONFIG_HOTPLUG_CPU */
1651int native_cpu_disable(void)
1652{
1653 return -ENOSYS;
1654}
1655
1656void native_cpu_die(unsigned int cpu)
1657{
1658 /* We said "no" in __cpu_disable */
1659 BUG();
1660}
1661
1662void native_play_dead(void)
1663{
1664 BUG();
1665}
1666
1667#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
64#include <asm/acpi.h>
65#include <asm/cacheinfo.h>
66#include <asm/desc.h>
67#include <asm/nmi.h>
68#include <asm/irq.h>
69#include <asm/realmode.h>
70#include <asm/cpu.h>
71#include <asm/numa.h>
72#include <asm/tlbflush.h>
73#include <asm/mtrr.h>
74#include <asm/mwait.h>
75#include <asm/apic.h>
76#include <asm/io_apic.h>
77#include <asm/fpu/api.h>
78#include <asm/setup.h>
79#include <asm/uv/uv.h>
80#include <asm/microcode.h>
81#include <asm/i8259.h>
82#include <asm/misc.h>
83#include <asm/qspinlock.h>
84#include <asm/intel-family.h>
85#include <asm/cpu_device_id.h>
86#include <asm/spec-ctrl.h>
87#include <asm/hw_irq.h>
88#include <asm/stackprotector.h>
89#include <asm/sev.h>
90#include <asm/spec-ctrl.h>
91
92/* representing HT siblings of each logical CPU */
93DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
94EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
95
96/* representing HT and core siblings of each logical CPU */
97DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
98EXPORT_PER_CPU_SYMBOL(cpu_core_map);
99
100/* representing HT, core, and die siblings of each logical CPU */
101DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_die_map);
102EXPORT_PER_CPU_SYMBOL(cpu_die_map);
103
104/* CPUs which are the primary SMT threads */
105struct cpumask __cpu_primary_thread_mask __read_mostly;
106
107/* Representing CPUs for which sibling maps can be computed */
108static cpumask_var_t cpu_sibling_setup_mask;
109
110struct mwait_cpu_dead {
111 unsigned int control;
112 unsigned int status;
113};
114
115#define CPUDEAD_MWAIT_WAIT 0xDEADBEEF
116#define CPUDEAD_MWAIT_KEXEC_HLT 0x4A17DEAD
117
118/*
119 * Cache line aligned data for mwait_play_dead(). Separate on purpose so
120 * that it's unlikely to be touched by other CPUs.
121 */
122static DEFINE_PER_CPU_ALIGNED(struct mwait_cpu_dead, mwait_cpu_dead);
123
124/* Maximum number of SMT threads on any online core */
125int __read_mostly __max_smt_threads = 1;
126
127/* Flag to indicate if a complete sched domain rebuild is required */
128bool x86_topology_update;
129
130int arch_update_cpu_topology(void)
131{
132 int retval = x86_topology_update;
133
134 x86_topology_update = false;
135 return retval;
136}
137
138static unsigned int smpboot_warm_reset_vector_count;
139
140static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
141{
142 unsigned long flags;
143
144 spin_lock_irqsave(&rtc_lock, flags);
145 if (!smpboot_warm_reset_vector_count++) {
146 CMOS_WRITE(0xa, 0xf);
147 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) = start_eip >> 4;
148 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = start_eip & 0xf;
149 }
150 spin_unlock_irqrestore(&rtc_lock, flags);
151}
152
153static inline void smpboot_restore_warm_reset_vector(void)
154{
155 unsigned long flags;
156
157 /*
158 * Paranoid: Set warm reset code and vector here back
159 * to default values.
160 */
161 spin_lock_irqsave(&rtc_lock, flags);
162 if (!--smpboot_warm_reset_vector_count) {
163 CMOS_WRITE(0, 0xf);
164 *((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
165 }
166 spin_unlock_irqrestore(&rtc_lock, flags);
167
168}
169
170/* Run the next set of setup steps for the upcoming CPU */
171static void ap_starting(void)
172{
173 int cpuid = smp_processor_id();
174
175 /* Mop up eventual mwait_play_dead() wreckage */
176 this_cpu_write(mwait_cpu_dead.status, 0);
177 this_cpu_write(mwait_cpu_dead.control, 0);
178
179 /*
180 * If woken up by an INIT in an 82489DX configuration the alive
181 * synchronization guarantees that the CPU does not reach this
182 * point before an INIT_deassert IPI reaches the local APIC, so it
183 * is now safe to touch the local APIC.
184 *
185 * Set up this CPU, first the APIC, which is probably redundant on
186 * most boards.
187 */
188 apic_ap_setup();
189
190 /* Save the processor parameters. */
191 smp_store_cpu_info(cpuid);
192
193 /*
194 * The topology information must be up to date before
195 * notify_cpu_starting().
196 */
197 set_cpu_sibling_map(cpuid);
198
199 ap_init_aperfmperf();
200
201 pr_debug("Stack at about %p\n", &cpuid);
202
203 wmb();
204
205 /*
206 * This runs the AP through all the cpuhp states to its target
207 * state CPUHP_ONLINE.
208 */
209 notify_cpu_starting(cpuid);
210}
211
212static void ap_calibrate_delay(void)
213{
214 /*
215 * Calibrate the delay loop and update loops_per_jiffy in cpu_data.
216 * smp_store_cpu_info() stored a value that is close but not as
217 * accurate as the value just calculated.
218 *
219 * As this is invoked after the TSC synchronization check,
220 * calibrate_delay_is_known() will skip the calibration routine
221 * when TSC is synchronized across sockets.
222 */
223 calibrate_delay();
224 cpu_data(smp_processor_id()).loops_per_jiffy = loops_per_jiffy;
225}
226
227/*
228 * Activate a secondary processor.
229 */
230static void notrace start_secondary(void *unused)
231{
232 /*
233 * Don't put *anything* except direct CPU state initialization
234 * before cpu_init(), SMP booting is too fragile that we want to
235 * limit the things done here to the most necessary things.
236 */
237 cr4_init();
238
239 /*
240 * 32-bit specific. 64-bit reaches this code with the correct page
241 * table established. Yet another historical divergence.
242 */
243 if (IS_ENABLED(CONFIG_X86_32)) {
244 /* switch away from the initial page table */
245 load_cr3(swapper_pg_dir);
246 __flush_tlb_all();
247 }
248
249 cpu_init_exception_handling();
250
251 /*
252 * Load the microcode before reaching the AP alive synchronization
253 * point below so it is not part of the full per CPU serialized
254 * bringup part when "parallel" bringup is enabled.
255 *
256 * That's even safe when hyperthreading is enabled in the CPU as
257 * the core code starts the primary threads first and leaves the
258 * secondary threads waiting for SIPI. Loading microcode on
259 * physical cores concurrently is a safe operation.
260 *
261 * This covers both the Intel specific issue that concurrent
262 * microcode loading on SMT siblings must be prohibited and the
263 * vendor independent issue`that microcode loading which changes
264 * CPUID, MSRs etc. must be strictly serialized to maintain
265 * software state correctness.
266 */
267 load_ucode_ap();
268
269 /*
270 * Synchronization point with the hotplug core. Sets this CPUs
271 * synchronization state to ALIVE and spin-waits for the control CPU to
272 * release this CPU for further bringup.
273 */
274 cpuhp_ap_sync_alive();
275
276 cpu_init();
277 fpu__init_cpu();
278 rcutree_report_cpu_starting(raw_smp_processor_id());
279 x86_cpuinit.early_percpu_clock_init();
280
281 ap_starting();
282
283 /* Check TSC synchronization with the control CPU. */
284 check_tsc_sync_target();
285
286 /*
287 * Calibrate the delay loop after the TSC synchronization check.
288 * This allows to skip the calibration when TSC is synchronized
289 * across sockets.
290 */
291 ap_calibrate_delay();
292
293 speculative_store_bypass_ht_init();
294
295 /*
296 * Lock vector_lock, set CPU online and bring the vector
297 * allocator online. Online must be set with vector_lock held
298 * to prevent a concurrent irq setup/teardown from seeing a
299 * half valid vector space.
300 */
301 lock_vector_lock();
302 set_cpu_online(smp_processor_id(), true);
303 lapic_online();
304 unlock_vector_lock();
305 x86_platform.nmi_init();
306
307 /* enable local interrupts */
308 local_irq_enable();
309
310 x86_cpuinit.setup_percpu_clockev();
311
312 wmb();
313 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
314}
315
316/*
317 * The bootstrap kernel entry code has set these up. Save them for
318 * a given CPU
319 */
320void smp_store_cpu_info(int id)
321{
322 struct cpuinfo_x86 *c = &cpu_data(id);
323
324 /* Copy boot_cpu_data only on the first bringup */
325 if (!c->initialized)
326 *c = boot_cpu_data;
327 c->cpu_index = id;
328 /*
329 * During boot time, CPU0 has this setup already. Save the info when
330 * bringing up an AP.
331 */
332 identify_secondary_cpu(c);
333 c->initialized = true;
334}
335
336static bool
337topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
338{
339 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
340
341 return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
342}
343
344static bool
345topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
346{
347 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
348
349 return !WARN_ONCE(!topology_same_node(c, o),
350 "sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
351 "[node: %d != %d]. Ignoring dependency.\n",
352 cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
353}
354
355#define link_mask(mfunc, c1, c2) \
356do { \
357 cpumask_set_cpu((c1), mfunc(c2)); \
358 cpumask_set_cpu((c2), mfunc(c1)); \
359} while (0)
360
361static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
362{
363 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
364 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
365
366 if (c->topo.pkg_id == o->topo.pkg_id &&
367 c->topo.die_id == o->topo.die_id &&
368 c->topo.amd_node_id == o->topo.amd_node_id &&
369 per_cpu_llc_id(cpu1) == per_cpu_llc_id(cpu2)) {
370 if (c->topo.core_id == o->topo.core_id)
371 return topology_sane(c, o, "smt");
372
373 if ((c->topo.cu_id != 0xff) &&
374 (o->topo.cu_id != 0xff) &&
375 (c->topo.cu_id == o->topo.cu_id))
376 return topology_sane(c, o, "smt");
377 }
378
379 } else if (c->topo.pkg_id == o->topo.pkg_id &&
380 c->topo.die_id == o->topo.die_id &&
381 c->topo.core_id == o->topo.core_id) {
382 return topology_sane(c, o, "smt");
383 }
384
385 return false;
386}
387
388static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
389{
390 if (c->topo.pkg_id != o->topo.pkg_id || c->topo.die_id != o->topo.die_id)
391 return false;
392
393 if (cpu_feature_enabled(X86_FEATURE_TOPOEXT) && topology_amd_nodes_per_pkg() > 1)
394 return c->topo.amd_node_id == o->topo.amd_node_id;
395
396 return true;
397}
398
399static bool match_l2c(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
400{
401 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
402
403 /* If the arch didn't set up l2c_id, fall back to SMT */
404 if (per_cpu_l2c_id(cpu1) == BAD_APICID)
405 return match_smt(c, o);
406
407 /* Do not match if L2 cache id does not match: */
408 if (per_cpu_l2c_id(cpu1) != per_cpu_l2c_id(cpu2))
409 return false;
410
411 return topology_sane(c, o, "l2c");
412}
413
414/*
415 * Unlike the other levels, we do not enforce keeping a
416 * multicore group inside a NUMA node. If this happens, we will
417 * discard the MC level of the topology later.
418 */
419static bool match_pkg(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
420{
421 if (c->topo.pkg_id == o->topo.pkg_id)
422 return true;
423 return false;
424}
425
426/*
427 * Define intel_cod_cpu[] for Intel COD (Cluster-on-Die) CPUs.
428 *
429 * Any Intel CPU that has multiple nodes per package and does not
430 * match intel_cod_cpu[] has the SNC (Sub-NUMA Cluster) topology.
431 *
432 * When in SNC mode, these CPUs enumerate an LLC that is shared
433 * by multiple NUMA nodes. The LLC is shared for off-package data
434 * access but private to the NUMA node (half of the package) for
435 * on-package access. CPUID (the source of the information about
436 * the LLC) can only enumerate the cache as shared or unshared,
437 * but not this particular configuration.
438 */
439
440static const struct x86_cpu_id intel_cod_cpu[] = {
441 X86_MATCH_INTEL_FAM6_MODEL(HASWELL_X, 0), /* COD */
442 X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_X, 0), /* COD */
443 X86_MATCH_INTEL_FAM6_MODEL(ANY, 1), /* SNC */
444 {}
445};
446
447static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
448{
449 const struct x86_cpu_id *id = x86_match_cpu(intel_cod_cpu);
450 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
451 bool intel_snc = id && id->driver_data;
452
453 /* Do not match if we do not have a valid APICID for cpu: */
454 if (per_cpu_llc_id(cpu1) == BAD_APICID)
455 return false;
456
457 /* Do not match if LLC id does not match: */
458 if (per_cpu_llc_id(cpu1) != per_cpu_llc_id(cpu2))
459 return false;
460
461 /*
462 * Allow the SNC topology without warning. Return of false
463 * means 'c' does not share the LLC of 'o'. This will be
464 * reflected to userspace.
465 */
466 if (match_pkg(c, o) && !topology_same_node(c, o) && intel_snc)
467 return false;
468
469 return topology_sane(c, o, "llc");
470}
471
472
473static inline int x86_sched_itmt_flags(void)
474{
475 return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
476}
477
478#ifdef CONFIG_SCHED_MC
479static int x86_core_flags(void)
480{
481 return cpu_core_flags() | x86_sched_itmt_flags();
482}
483#endif
484#ifdef CONFIG_SCHED_SMT
485static int x86_smt_flags(void)
486{
487 return cpu_smt_flags();
488}
489#endif
490#ifdef CONFIG_SCHED_CLUSTER
491static int x86_cluster_flags(void)
492{
493 return cpu_cluster_flags() | x86_sched_itmt_flags();
494}
495#endif
496
497static int x86_die_flags(void)
498{
499 if (cpu_feature_enabled(X86_FEATURE_HYBRID_CPU))
500 return x86_sched_itmt_flags();
501
502 return 0;
503}
504
505/*
506 * Set if a package/die has multiple NUMA nodes inside.
507 * AMD Magny-Cours, Intel Cluster-on-Die, and Intel
508 * Sub-NUMA Clustering have this.
509 */
510static bool x86_has_numa_in_package;
511
512static struct sched_domain_topology_level x86_topology[6];
513
514static void __init build_sched_topology(void)
515{
516 int i = 0;
517
518#ifdef CONFIG_SCHED_SMT
519 x86_topology[i++] = (struct sched_domain_topology_level){
520 cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT)
521 };
522#endif
523#ifdef CONFIG_SCHED_CLUSTER
524 x86_topology[i++] = (struct sched_domain_topology_level){
525 cpu_clustergroup_mask, x86_cluster_flags, SD_INIT_NAME(CLS)
526 };
527#endif
528#ifdef CONFIG_SCHED_MC
529 x86_topology[i++] = (struct sched_domain_topology_level){
530 cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC)
531 };
532#endif
533 /*
534 * When there is NUMA topology inside the package skip the PKG domain
535 * since the NUMA domains will auto-magically create the right spanning
536 * domains based on the SLIT.
537 */
538 if (!x86_has_numa_in_package) {
539 x86_topology[i++] = (struct sched_domain_topology_level){
540 cpu_cpu_mask, x86_die_flags, SD_INIT_NAME(PKG)
541 };
542 }
543
544 /*
545 * There must be one trailing NULL entry left.
546 */
547 BUG_ON(i >= ARRAY_SIZE(x86_topology)-1);
548
549 set_sched_topology(x86_topology);
550}
551
552void set_cpu_sibling_map(int cpu)
553{
554 bool has_smt = __max_threads_per_core > 1;
555 bool has_mp = has_smt || topology_num_cores_per_package() > 1;
556 struct cpuinfo_x86 *c = &cpu_data(cpu);
557 struct cpuinfo_x86 *o;
558 int i, threads;
559
560 cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
561
562 if (!has_mp) {
563 cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
564 cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
565 cpumask_set_cpu(cpu, cpu_l2c_shared_mask(cpu));
566 cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
567 cpumask_set_cpu(cpu, topology_die_cpumask(cpu));
568 c->booted_cores = 1;
569 return;
570 }
571
572 for_each_cpu(i, cpu_sibling_setup_mask) {
573 o = &cpu_data(i);
574
575 if (match_pkg(c, o) && !topology_same_node(c, o))
576 x86_has_numa_in_package = true;
577
578 if ((i == cpu) || (has_smt && match_smt(c, o)))
579 link_mask(topology_sibling_cpumask, cpu, i);
580
581 if ((i == cpu) || (has_mp && match_llc(c, o)))
582 link_mask(cpu_llc_shared_mask, cpu, i);
583
584 if ((i == cpu) || (has_mp && match_l2c(c, o)))
585 link_mask(cpu_l2c_shared_mask, cpu, i);
586
587 if ((i == cpu) || (has_mp && match_die(c, o)))
588 link_mask(topology_die_cpumask, cpu, i);
589 }
590
591 threads = cpumask_weight(topology_sibling_cpumask(cpu));
592 if (threads > __max_smt_threads)
593 __max_smt_threads = threads;
594
595 for_each_cpu(i, topology_sibling_cpumask(cpu))
596 cpu_data(i).smt_active = threads > 1;
597
598 /*
599 * This needs a separate iteration over the cpus because we rely on all
600 * topology_sibling_cpumask links to be set-up.
601 */
602 for_each_cpu(i, cpu_sibling_setup_mask) {
603 o = &cpu_data(i);
604
605 if ((i == cpu) || (has_mp && match_pkg(c, o))) {
606 link_mask(topology_core_cpumask, cpu, i);
607
608 /*
609 * Does this new cpu bringup a new core?
610 */
611 if (threads == 1) {
612 /*
613 * for each core in package, increment
614 * the booted_cores for this new cpu
615 */
616 if (cpumask_first(
617 topology_sibling_cpumask(i)) == i)
618 c->booted_cores++;
619 /*
620 * increment the core count for all
621 * the other cpus in this package
622 */
623 if (i != cpu)
624 cpu_data(i).booted_cores++;
625 } else if (i != cpu && !c->booted_cores)
626 c->booted_cores = cpu_data(i).booted_cores;
627 }
628 }
629}
630
631/* maps the cpu to the sched domain representing multi-core */
632const struct cpumask *cpu_coregroup_mask(int cpu)
633{
634 return cpu_llc_shared_mask(cpu);
635}
636
637const struct cpumask *cpu_clustergroup_mask(int cpu)
638{
639 return cpu_l2c_shared_mask(cpu);
640}
641EXPORT_SYMBOL_GPL(cpu_clustergroup_mask);
642
643static void impress_friends(void)
644{
645 int cpu;
646 unsigned long bogosum = 0;
647 /*
648 * Allow the user to impress friends.
649 */
650 pr_debug("Before bogomips\n");
651 for_each_online_cpu(cpu)
652 bogosum += cpu_data(cpu).loops_per_jiffy;
653
654 pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
655 num_online_cpus(),
656 bogosum/(500000/HZ),
657 (bogosum/(5000/HZ))%100);
658
659 pr_debug("Before bogocount - setting activated=1\n");
660}
661
662/*
663 * The Multiprocessor Specification 1.4 (1997) example code suggests
664 * that there should be a 10ms delay between the BSP asserting INIT
665 * and de-asserting INIT, when starting a remote processor.
666 * But that slows boot and resume on modern processors, which include
667 * many cores and don't require that delay.
668 *
669 * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
670 * Modern processor families are quirked to remove the delay entirely.
671 */
672#define UDELAY_10MS_DEFAULT 10000
673
674static unsigned int init_udelay = UINT_MAX;
675
676static int __init cpu_init_udelay(char *str)
677{
678 get_option(&str, &init_udelay);
679
680 return 0;
681}
682early_param("cpu_init_udelay", cpu_init_udelay);
683
684static void __init smp_quirk_init_udelay(void)
685{
686 /* if cmdline changed it from default, leave it alone */
687 if (init_udelay != UINT_MAX)
688 return;
689
690 /* if modern processor, use no delay */
691 if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
692 ((boot_cpu_data.x86_vendor == X86_VENDOR_HYGON) && (boot_cpu_data.x86 >= 0x18)) ||
693 ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
694 init_udelay = 0;
695 return;
696 }
697 /* else, use legacy delay */
698 init_udelay = UDELAY_10MS_DEFAULT;
699}
700
701/*
702 * Wake up AP by INIT, INIT, STARTUP sequence.
703 */
704static void send_init_sequence(u32 phys_apicid)
705{
706 int maxlvt = lapic_get_maxlvt();
707
708 /* Be paranoid about clearing APIC errors. */
709 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
710 /* Due to the Pentium erratum 3AP. */
711 if (maxlvt > 3)
712 apic_write(APIC_ESR, 0);
713 apic_read(APIC_ESR);
714 }
715
716 /* Assert INIT on the target CPU */
717 apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT, phys_apicid);
718 safe_apic_wait_icr_idle();
719
720 udelay(init_udelay);
721
722 /* Deassert INIT on the target CPU */
723 apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
724 safe_apic_wait_icr_idle();
725}
726
727/*
728 * Wake up AP by INIT, INIT, STARTUP sequence.
729 */
730static int wakeup_secondary_cpu_via_init(u32 phys_apicid, unsigned long start_eip)
731{
732 unsigned long send_status = 0, accept_status = 0;
733 int num_starts, j, maxlvt;
734
735 preempt_disable();
736 maxlvt = lapic_get_maxlvt();
737 send_init_sequence(phys_apicid);
738
739 mb();
740
741 /*
742 * Should we send STARTUP IPIs ?
743 *
744 * Determine this based on the APIC version.
745 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
746 */
747 if (APIC_INTEGRATED(boot_cpu_apic_version))
748 num_starts = 2;
749 else
750 num_starts = 0;
751
752 /*
753 * Run STARTUP IPI loop.
754 */
755 pr_debug("#startup loops: %d\n", num_starts);
756
757 for (j = 1; j <= num_starts; j++) {
758 pr_debug("Sending STARTUP #%d\n", j);
759 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
760 apic_write(APIC_ESR, 0);
761 apic_read(APIC_ESR);
762 pr_debug("After apic_write\n");
763
764 /*
765 * STARTUP IPI
766 */
767
768 /* Target chip */
769 /* Boot on the stack */
770 /* Kick the second */
771 apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
772 phys_apicid);
773
774 /*
775 * Give the other CPU some time to accept the IPI.
776 */
777 if (init_udelay == 0)
778 udelay(10);
779 else
780 udelay(300);
781
782 pr_debug("Startup point 1\n");
783
784 pr_debug("Waiting for send to finish...\n");
785 send_status = safe_apic_wait_icr_idle();
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(200);
794
795 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
796 apic_write(APIC_ESR, 0);
797 accept_status = (apic_read(APIC_ESR) & 0xEF);
798 if (send_status || accept_status)
799 break;
800 }
801 pr_debug("After Startup\n");
802
803 if (send_status)
804 pr_err("APIC never delivered???\n");
805 if (accept_status)
806 pr_err("APIC delivery error (%lx)\n", accept_status);
807
808 preempt_enable();
809 return (send_status | accept_status);
810}
811
812/* reduce the number of lines printed when booting a large cpu count system */
813static void announce_cpu(int cpu, int apicid)
814{
815 static int width, node_width, first = 1;
816 static int current_node = NUMA_NO_NODE;
817 int node = early_cpu_to_node(cpu);
818
819 if (!width)
820 width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
821
822 if (!node_width)
823 node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
824
825 if (system_state < SYSTEM_RUNNING) {
826 if (first)
827 pr_info("x86: Booting SMP configuration:\n");
828
829 if (node != current_node) {
830 if (current_node > (-1))
831 pr_cont("\n");
832 current_node = node;
833
834 printk(KERN_INFO ".... node %*s#%d, CPUs: ",
835 node_width - num_digits(node), " ", node);
836 }
837
838 /* Add padding for the BSP */
839 if (first)
840 pr_cont("%*s", width + 1, " ");
841 first = 0;
842
843 pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
844 } else
845 pr_info("Booting Node %d Processor %d APIC 0x%x\n",
846 node, cpu, apicid);
847}
848
849int common_cpu_up(unsigned int cpu, struct task_struct *idle)
850{
851 int ret;
852
853 /* Just in case we booted with a single CPU. */
854 alternatives_enable_smp();
855
856 per_cpu(pcpu_hot.current_task, cpu) = idle;
857 cpu_init_stack_canary(cpu, idle);
858
859 /* Initialize the interrupt stack(s) */
860 ret = irq_init_percpu_irqstack(cpu);
861 if (ret)
862 return ret;
863
864#ifdef CONFIG_X86_32
865 /* Stack for startup_32 can be just as for start_secondary onwards */
866 per_cpu(pcpu_hot.top_of_stack, cpu) = task_top_of_stack(idle);
867#endif
868 return 0;
869}
870
871/*
872 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
873 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
874 * Returns zero if startup was successfully sent, else error code from
875 * ->wakeup_secondary_cpu.
876 */
877static int do_boot_cpu(u32 apicid, int cpu, struct task_struct *idle)
878{
879 unsigned long start_ip = real_mode_header->trampoline_start;
880 int ret;
881
882#ifdef CONFIG_X86_64
883 /* If 64-bit wakeup method exists, use the 64-bit mode trampoline IP */
884 if (apic->wakeup_secondary_cpu_64)
885 start_ip = real_mode_header->trampoline_start64;
886#endif
887 idle->thread.sp = (unsigned long)task_pt_regs(idle);
888 initial_code = (unsigned long)start_secondary;
889
890 if (IS_ENABLED(CONFIG_X86_32)) {
891 early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
892 initial_stack = idle->thread.sp;
893 } else if (!(smpboot_control & STARTUP_PARALLEL_MASK)) {
894 smpboot_control = cpu;
895 }
896
897 /* Enable the espfix hack for this CPU */
898 init_espfix_ap(cpu);
899
900 /* So we see what's up */
901 announce_cpu(cpu, apicid);
902
903 /*
904 * This grunge runs the startup process for
905 * the targeted processor.
906 */
907 if (x86_platform.legacy.warm_reset) {
908
909 pr_debug("Setting warm reset code and vector.\n");
910
911 smpboot_setup_warm_reset_vector(start_ip);
912 /*
913 * Be paranoid about clearing APIC errors.
914 */
915 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
916 apic_write(APIC_ESR, 0);
917 apic_read(APIC_ESR);
918 }
919 }
920
921 smp_mb();
922
923 /*
924 * Wake up a CPU in difference cases:
925 * - Use a method from the APIC driver if one defined, with wakeup
926 * straight to 64-bit mode preferred over wakeup to RM.
927 * Otherwise,
928 * - Use an INIT boot APIC message
929 */
930 if (apic->wakeup_secondary_cpu_64)
931 ret = apic->wakeup_secondary_cpu_64(apicid, start_ip);
932 else if (apic->wakeup_secondary_cpu)
933 ret = apic->wakeup_secondary_cpu(apicid, start_ip);
934 else
935 ret = wakeup_secondary_cpu_via_init(apicid, start_ip);
936
937 /* If the wakeup mechanism failed, cleanup the warm reset vector */
938 if (ret)
939 arch_cpuhp_cleanup_kick_cpu(cpu);
940 return ret;
941}
942
943int native_kick_ap(unsigned int cpu, struct task_struct *tidle)
944{
945 u32 apicid = apic->cpu_present_to_apicid(cpu);
946 int err;
947
948 lockdep_assert_irqs_enabled();
949
950 pr_debug("++++++++++++++++++++=_---CPU UP %u\n", cpu);
951
952 if (apicid == BAD_APICID || !apic_id_valid(apicid)) {
953 pr_err("CPU %u has invalid APIC ID %x. Aborting bringup\n", cpu, apicid);
954 return -EINVAL;
955 }
956
957 if (!test_bit(apicid, phys_cpu_present_map)) {
958 pr_err("CPU %u APIC ID %x is not present. Aborting bringup\n", cpu, apicid);
959 return -EINVAL;
960 }
961
962 /*
963 * Save current MTRR state in case it was changed since early boot
964 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
965 */
966 mtrr_save_state();
967
968 /* the FPU context is blank, nobody can own it */
969 per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
970
971 err = common_cpu_up(cpu, tidle);
972 if (err)
973 return err;
974
975 err = do_boot_cpu(apicid, cpu, tidle);
976 if (err)
977 pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
978
979 return err;
980}
981
982int arch_cpuhp_kick_ap_alive(unsigned int cpu, struct task_struct *tidle)
983{
984 return smp_ops.kick_ap_alive(cpu, tidle);
985}
986
987void arch_cpuhp_cleanup_kick_cpu(unsigned int cpu)
988{
989 /* Cleanup possible dangling ends... */
990 if (smp_ops.kick_ap_alive == native_kick_ap && x86_platform.legacy.warm_reset)
991 smpboot_restore_warm_reset_vector();
992}
993
994void arch_cpuhp_cleanup_dead_cpu(unsigned int cpu)
995{
996 if (smp_ops.cleanup_dead_cpu)
997 smp_ops.cleanup_dead_cpu(cpu);
998
999 if (system_state == SYSTEM_RUNNING)
1000 pr_info("CPU %u is now offline\n", cpu);
1001}
1002
1003void arch_cpuhp_sync_state_poll(void)
1004{
1005 if (smp_ops.poll_sync_state)
1006 smp_ops.poll_sync_state();
1007}
1008
1009/**
1010 * arch_disable_smp_support() - Disables SMP support for x86 at boottime
1011 */
1012void __init arch_disable_smp_support(void)
1013{
1014 disable_ioapic_support();
1015}
1016
1017/*
1018 * Fall back to non SMP mode after errors.
1019 *
1020 * RED-PEN audit/test this more. I bet there is more state messed up here.
1021 */
1022static __init void disable_smp(void)
1023{
1024 pr_info("SMP disabled\n");
1025
1026 disable_ioapic_support();
1027 topology_reset_possible_cpus_up();
1028
1029 cpumask_set_cpu(0, topology_sibling_cpumask(0));
1030 cpumask_set_cpu(0, topology_core_cpumask(0));
1031 cpumask_set_cpu(0, topology_die_cpumask(0));
1032}
1033
1034void __init smp_prepare_cpus_common(void)
1035{
1036 unsigned int i;
1037
1038 /* Mark all except the boot CPU as hotpluggable */
1039 for_each_possible_cpu(i) {
1040 if (i)
1041 per_cpu(cpu_info.cpu_index, i) = nr_cpu_ids;
1042 }
1043
1044 for_each_possible_cpu(i) {
1045 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1046 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1047 zalloc_cpumask_var(&per_cpu(cpu_die_map, i), GFP_KERNEL);
1048 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1049 zalloc_cpumask_var(&per_cpu(cpu_l2c_shared_map, i), GFP_KERNEL);
1050 }
1051
1052 set_cpu_sibling_map(0);
1053}
1054
1055void __init smp_prepare_boot_cpu(void)
1056{
1057 smp_ops.smp_prepare_boot_cpu();
1058}
1059
1060#ifdef CONFIG_X86_64
1061/* Establish whether parallel bringup can be supported. */
1062bool __init arch_cpuhp_init_parallel_bringup(void)
1063{
1064 if (!x86_cpuinit.parallel_bringup) {
1065 pr_info("Parallel CPU startup disabled by the platform\n");
1066 return false;
1067 }
1068
1069 smpboot_control = STARTUP_READ_APICID;
1070 pr_debug("Parallel CPU startup enabled: 0x%08x\n", smpboot_control);
1071 return true;
1072}
1073#endif
1074
1075/*
1076 * Prepare for SMP bootup.
1077 * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1078 * for common interface support.
1079 */
1080void __init native_smp_prepare_cpus(unsigned int max_cpus)
1081{
1082 smp_prepare_cpus_common();
1083
1084 switch (apic_intr_mode) {
1085 case APIC_PIC:
1086 case APIC_VIRTUAL_WIRE_NO_CONFIG:
1087 disable_smp();
1088 return;
1089 case APIC_SYMMETRIC_IO_NO_ROUTING:
1090 disable_smp();
1091 /* Setup local timer */
1092 x86_init.timers.setup_percpu_clockev();
1093 return;
1094 case APIC_VIRTUAL_WIRE:
1095 case APIC_SYMMETRIC_IO:
1096 break;
1097 }
1098
1099 /* Setup local timer */
1100 x86_init.timers.setup_percpu_clockev();
1101
1102 pr_info("CPU0: ");
1103 print_cpu_info(&cpu_data(0));
1104
1105 uv_system_init();
1106
1107 smp_quirk_init_udelay();
1108
1109 speculative_store_bypass_ht_init();
1110
1111 snp_set_wakeup_secondary_cpu();
1112}
1113
1114void arch_thaw_secondary_cpus_begin(void)
1115{
1116 set_cache_aps_delayed_init(true);
1117}
1118
1119void arch_thaw_secondary_cpus_end(void)
1120{
1121 cache_aps_init();
1122}
1123
1124/*
1125 * Early setup to make printk work.
1126 */
1127void __init native_smp_prepare_boot_cpu(void)
1128{
1129 int me = smp_processor_id();
1130
1131 /* SMP handles this from setup_per_cpu_areas() */
1132 if (!IS_ENABLED(CONFIG_SMP))
1133 switch_gdt_and_percpu_base(me);
1134
1135 native_pv_lock_init();
1136}
1137
1138void __init native_smp_cpus_done(unsigned int max_cpus)
1139{
1140 pr_debug("Boot done\n");
1141
1142 build_sched_topology();
1143 nmi_selftest();
1144 impress_friends();
1145 cache_aps_init();
1146}
1147
1148/* correctly size the local cpu masks */
1149void __init setup_cpu_local_masks(void)
1150{
1151 alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
1152}
1153
1154#ifdef CONFIG_HOTPLUG_CPU
1155
1156/* Recompute SMT state for all CPUs on offline */
1157static void recompute_smt_state(void)
1158{
1159 int max_threads, cpu;
1160
1161 max_threads = 0;
1162 for_each_online_cpu (cpu) {
1163 int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1164
1165 if (threads > max_threads)
1166 max_threads = threads;
1167 }
1168 __max_smt_threads = max_threads;
1169}
1170
1171static void remove_siblinginfo(int cpu)
1172{
1173 int sibling;
1174 struct cpuinfo_x86 *c = &cpu_data(cpu);
1175
1176 for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1177 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1178 /*/
1179 * last thread sibling in this cpu core going down
1180 */
1181 if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1182 cpu_data(sibling).booted_cores--;
1183 }
1184
1185 for_each_cpu(sibling, topology_die_cpumask(cpu))
1186 cpumask_clear_cpu(cpu, topology_die_cpumask(sibling));
1187
1188 for_each_cpu(sibling, topology_sibling_cpumask(cpu)) {
1189 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1190 if (cpumask_weight(topology_sibling_cpumask(sibling)) == 1)
1191 cpu_data(sibling).smt_active = false;
1192 }
1193
1194 for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1195 cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1196 for_each_cpu(sibling, cpu_l2c_shared_mask(cpu))
1197 cpumask_clear_cpu(cpu, cpu_l2c_shared_mask(sibling));
1198 cpumask_clear(cpu_llc_shared_mask(cpu));
1199 cpumask_clear(cpu_l2c_shared_mask(cpu));
1200 cpumask_clear(topology_sibling_cpumask(cpu));
1201 cpumask_clear(topology_core_cpumask(cpu));
1202 cpumask_clear(topology_die_cpumask(cpu));
1203 c->topo.core_id = 0;
1204 c->booted_cores = 0;
1205 cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1206 recompute_smt_state();
1207}
1208
1209static void remove_cpu_from_maps(int cpu)
1210{
1211 set_cpu_online(cpu, false);
1212 numa_remove_cpu(cpu);
1213}
1214
1215void cpu_disable_common(void)
1216{
1217 int cpu = smp_processor_id();
1218
1219 remove_siblinginfo(cpu);
1220
1221 /* It's now safe to remove this processor from the online map */
1222 lock_vector_lock();
1223 remove_cpu_from_maps(cpu);
1224 unlock_vector_lock();
1225 fixup_irqs();
1226 lapic_offline();
1227}
1228
1229int native_cpu_disable(void)
1230{
1231 int ret;
1232
1233 ret = lapic_can_unplug_cpu();
1234 if (ret)
1235 return ret;
1236
1237 cpu_disable_common();
1238
1239 /*
1240 * Disable the local APIC. Otherwise IPI broadcasts will reach
1241 * it. It still responds normally to INIT, NMI, SMI, and SIPI
1242 * messages.
1243 *
1244 * Disabling the APIC must happen after cpu_disable_common()
1245 * which invokes fixup_irqs().
1246 *
1247 * Disabling the APIC preserves already set bits in IRR, but
1248 * an interrupt arriving after disabling the local APIC does not
1249 * set the corresponding IRR bit.
1250 *
1251 * fixup_irqs() scans IRR for set bits so it can raise a not
1252 * yet handled interrupt on the new destination CPU via an IPI
1253 * but obviously it can't do so for IRR bits which are not set.
1254 * IOW, interrupts arriving after disabling the local APIC will
1255 * be lost.
1256 */
1257 apic_soft_disable();
1258
1259 return 0;
1260}
1261
1262void play_dead_common(void)
1263{
1264 idle_task_exit();
1265
1266 cpuhp_ap_report_dead();
1267
1268 local_irq_disable();
1269}
1270
1271/*
1272 * We need to flush the caches before going to sleep, lest we have
1273 * dirty data in our caches when we come back up.
1274 */
1275static inline void mwait_play_dead(void)
1276{
1277 struct mwait_cpu_dead *md = this_cpu_ptr(&mwait_cpu_dead);
1278 unsigned int eax, ebx, ecx, edx;
1279 unsigned int highest_cstate = 0;
1280 unsigned int highest_subcstate = 0;
1281 int i;
1282
1283 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD ||
1284 boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
1285 return;
1286 if (!this_cpu_has(X86_FEATURE_MWAIT))
1287 return;
1288 if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1289 return;
1290 if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1291 return;
1292
1293 eax = CPUID_MWAIT_LEAF;
1294 ecx = 0;
1295 native_cpuid(&eax, &ebx, &ecx, &edx);
1296
1297 /*
1298 * eax will be 0 if EDX enumeration is not valid.
1299 * Initialized below to cstate, sub_cstate value when EDX is valid.
1300 */
1301 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1302 eax = 0;
1303 } else {
1304 edx >>= MWAIT_SUBSTATE_SIZE;
1305 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1306 if (edx & MWAIT_SUBSTATE_MASK) {
1307 highest_cstate = i;
1308 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1309 }
1310 }
1311 eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1312 (highest_subcstate - 1);
1313 }
1314
1315 /* Set up state for the kexec() hack below */
1316 md->status = CPUDEAD_MWAIT_WAIT;
1317 md->control = CPUDEAD_MWAIT_WAIT;
1318
1319 wbinvd();
1320
1321 while (1) {
1322 /*
1323 * The CLFLUSH is a workaround for erratum AAI65 for
1324 * the Xeon 7400 series. It's not clear it is actually
1325 * needed, but it should be harmless in either case.
1326 * The WBINVD is insufficient due to the spurious-wakeup
1327 * case where we return around the loop.
1328 */
1329 mb();
1330 clflush(md);
1331 mb();
1332 __monitor(md, 0, 0);
1333 mb();
1334 __mwait(eax, 0);
1335
1336 if (READ_ONCE(md->control) == CPUDEAD_MWAIT_KEXEC_HLT) {
1337 /*
1338 * Kexec is about to happen. Don't go back into mwait() as
1339 * the kexec kernel might overwrite text and data including
1340 * page tables and stack. So mwait() would resume when the
1341 * monitor cache line is written to and then the CPU goes
1342 * south due to overwritten text, page tables and stack.
1343 *
1344 * Note: This does _NOT_ protect against a stray MCE, NMI,
1345 * SMI. They will resume execution at the instruction
1346 * following the HLT instruction and run into the problem
1347 * which this is trying to prevent.
1348 */
1349 WRITE_ONCE(md->status, CPUDEAD_MWAIT_KEXEC_HLT);
1350 while(1)
1351 native_halt();
1352 }
1353 }
1354}
1355
1356/*
1357 * Kick all "offline" CPUs out of mwait on kexec(). See comment in
1358 * mwait_play_dead().
1359 */
1360void smp_kick_mwait_play_dead(void)
1361{
1362 u32 newstate = CPUDEAD_MWAIT_KEXEC_HLT;
1363 struct mwait_cpu_dead *md;
1364 unsigned int cpu, i;
1365
1366 for_each_cpu_andnot(cpu, cpu_present_mask, cpu_online_mask) {
1367 md = per_cpu_ptr(&mwait_cpu_dead, cpu);
1368
1369 /* Does it sit in mwait_play_dead() ? */
1370 if (READ_ONCE(md->status) != CPUDEAD_MWAIT_WAIT)
1371 continue;
1372
1373 /* Wait up to 5ms */
1374 for (i = 0; READ_ONCE(md->status) != newstate && i < 1000; i++) {
1375 /* Bring it out of mwait */
1376 WRITE_ONCE(md->control, newstate);
1377 udelay(5);
1378 }
1379
1380 if (READ_ONCE(md->status) != newstate)
1381 pr_err_once("CPU%u is stuck in mwait_play_dead()\n", cpu);
1382 }
1383}
1384
1385void __noreturn hlt_play_dead(void)
1386{
1387 if (__this_cpu_read(cpu_info.x86) >= 4)
1388 wbinvd();
1389
1390 while (1)
1391 native_halt();
1392}
1393
1394/*
1395 * native_play_dead() is essentially a __noreturn function, but it can't
1396 * be marked as such as the compiler may complain about it.
1397 */
1398void native_play_dead(void)
1399{
1400 if (cpu_feature_enabled(X86_FEATURE_KERNEL_IBRS))
1401 __update_spec_ctrl(0);
1402
1403 play_dead_common();
1404 tboot_shutdown(TB_SHUTDOWN_WFS);
1405
1406 mwait_play_dead();
1407 if (cpuidle_play_dead())
1408 hlt_play_dead();
1409}
1410
1411#else /* ... !CONFIG_HOTPLUG_CPU */
1412int native_cpu_disable(void)
1413{
1414 return -ENOSYS;
1415}
1416
1417void native_play_dead(void)
1418{
1419 BUG();
1420}
1421
1422#endif