Loading...
1// SPDX-License-Identifier: GPL-2.0-or-later
2 /*
3 * x86 SMP booting functions
4 *
5 * (c) 1995 Alan Cox, Building #3 <alan@lxorguk.ukuu.org.uk>
6 * (c) 1998, 1999, 2000, 2009 Ingo Molnar <mingo@redhat.com>
7 * Copyright 2001 Andi Kleen, SuSE Labs.
8 *
9 * Much of the core SMP work is based on previous work by Thomas Radke, to
10 * whom a great many thanks are extended.
11 *
12 * Thanks to Intel for making available several different Pentium,
13 * Pentium Pro and Pentium-II/Xeon MP machines.
14 * Original development of Linux SMP code supported by Caldera.
15 *
16 * Fixes
17 * Felix Koop : NR_CPUS used properly
18 * Jose Renau : Handle single CPU case.
19 * Alan Cox : By repeated request 8) - Total BogoMIPS report.
20 * Greg Wright : Fix for kernel stacks panic.
21 * Erich Boleyn : MP v1.4 and additional changes.
22 * Matthias Sattler : Changes for 2.1 kernel map.
23 * Michel Lespinasse : Changes for 2.1 kernel map.
24 * Michael Chastain : Change trampoline.S to gnu as.
25 * Alan Cox : Dumb bug: 'B' step PPro's are fine
26 * Ingo Molnar : Added APIC timers, based on code
27 * from Jose Renau
28 * Ingo Molnar : various cleanups and rewrites
29 * Tigran Aivazian : fixed "0.00 in /proc/uptime on SMP" bug.
30 * Maciej W. Rozycki : Bits for genuine 82489DX APICs
31 * Andi Kleen : Changed for SMP boot into long mode.
32 * Martin J. Bligh : Added support for multi-quad systems
33 * Dave Jones : Report invalid combinations of Athlon CPUs.
34 * Rusty Russell : Hacked into shape for new "hotplug" boot process.
35 * Andi Kleen : Converted to new state machine.
36 * Ashok Raj : CPU hotplug support
37 * Glauber Costa : i386 and x86_64 integration
38 */
39
40#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
41
42#include <linux/init.h>
43#include <linux/smp.h>
44#include <linux/export.h>
45#include <linux/sched.h>
46#include <linux/sched/topology.h>
47#include <linux/sched/hotplug.h>
48#include <linux/sched/task_stack.h>
49#include <linux/percpu.h>
50#include <linux/memblock.h>
51#include <linux/err.h>
52#include <linux/nmi.h>
53#include <linux/tboot.h>
54#include <linux/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
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/* Per CPU bogomips and other parameters */
105DEFINE_PER_CPU_READ_MOSTLY(struct cpuinfo_x86, cpu_info);
106EXPORT_PER_CPU_SYMBOL(cpu_info);
107
108/* CPUs which are the primary SMT threads */
109struct cpumask __cpu_primary_thread_mask __read_mostly;
110
111/* Representing CPUs for which sibling maps can be computed */
112static cpumask_var_t cpu_sibling_setup_mask;
113
114struct mwait_cpu_dead {
115 unsigned int control;
116 unsigned int status;
117};
118
119#define CPUDEAD_MWAIT_WAIT 0xDEADBEEF
120#define CPUDEAD_MWAIT_KEXEC_HLT 0x4A17DEAD
121
122/*
123 * Cache line aligned data for mwait_play_dead(). Separate on purpose so
124 * that it's unlikely to be touched by other CPUs.
125 */
126static DEFINE_PER_CPU_ALIGNED(struct mwait_cpu_dead, mwait_cpu_dead);
127
128/* Logical package management. */
129struct logical_maps {
130 u32 phys_pkg_id;
131 u32 phys_die_id;
132 u32 logical_pkg_id;
133 u32 logical_die_id;
134};
135
136/* Temporary workaround until the full topology mechanics is in place */
137static DEFINE_PER_CPU_READ_MOSTLY(struct logical_maps, logical_maps) = {
138 .phys_pkg_id = U32_MAX,
139 .phys_die_id = U32_MAX,
140};
141
142unsigned int __max_logical_packages __read_mostly;
143EXPORT_SYMBOL(__max_logical_packages);
144static unsigned int logical_packages __read_mostly;
145static unsigned int logical_die __read_mostly;
146
147/* Maximum number of SMT threads on any online core */
148int __read_mostly __max_smt_threads = 1;
149
150/* Flag to indicate if a complete sched domain rebuild is required */
151bool x86_topology_update;
152
153int arch_update_cpu_topology(void)
154{
155 int retval = x86_topology_update;
156
157 x86_topology_update = false;
158 return retval;
159}
160
161static unsigned int smpboot_warm_reset_vector_count;
162
163static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
164{
165 unsigned long flags;
166
167 spin_lock_irqsave(&rtc_lock, flags);
168 if (!smpboot_warm_reset_vector_count++) {
169 CMOS_WRITE(0xa, 0xf);
170 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) = start_eip >> 4;
171 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = start_eip & 0xf;
172 }
173 spin_unlock_irqrestore(&rtc_lock, flags);
174}
175
176static inline void smpboot_restore_warm_reset_vector(void)
177{
178 unsigned long flags;
179
180 /*
181 * Paranoid: Set warm reset code and vector here back
182 * to default values.
183 */
184 spin_lock_irqsave(&rtc_lock, flags);
185 if (!--smpboot_warm_reset_vector_count) {
186 CMOS_WRITE(0, 0xf);
187 *((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
188 }
189 spin_unlock_irqrestore(&rtc_lock, flags);
190
191}
192
193/* Run the next set of setup steps for the upcoming CPU */
194static void ap_starting(void)
195{
196 int cpuid = smp_processor_id();
197
198 /* Mop up eventual mwait_play_dead() wreckage */
199 this_cpu_write(mwait_cpu_dead.status, 0);
200 this_cpu_write(mwait_cpu_dead.control, 0);
201
202 /*
203 * If woken up by an INIT in an 82489DX configuration the alive
204 * synchronization guarantees that the CPU does not reach this
205 * point before an INIT_deassert IPI reaches the local APIC, so it
206 * is now safe to touch the local APIC.
207 *
208 * Set up this CPU, first the APIC, which is probably redundant on
209 * most boards.
210 */
211 apic_ap_setup();
212
213 /* Save the processor parameters. */
214 smp_store_cpu_info(cpuid);
215
216 /*
217 * The topology information must be up to date before
218 * notify_cpu_starting().
219 */
220 set_cpu_sibling_map(cpuid);
221
222 ap_init_aperfmperf();
223
224 pr_debug("Stack at about %p\n", &cpuid);
225
226 wmb();
227
228 /*
229 * This runs the AP through all the cpuhp states to its target
230 * state CPUHP_ONLINE.
231 */
232 notify_cpu_starting(cpuid);
233}
234
235static void ap_calibrate_delay(void)
236{
237 /*
238 * Calibrate the delay loop and update loops_per_jiffy in cpu_data.
239 * smp_store_cpu_info() stored a value that is close but not as
240 * accurate as the value just calculated.
241 *
242 * As this is invoked after the TSC synchronization check,
243 * calibrate_delay_is_known() will skip the calibration routine
244 * when TSC is synchronized across sockets.
245 */
246 calibrate_delay();
247 cpu_data(smp_processor_id()).loops_per_jiffy = loops_per_jiffy;
248}
249
250/*
251 * Activate a secondary processor.
252 */
253static void notrace start_secondary(void *unused)
254{
255 /*
256 * Don't put *anything* except direct CPU state initialization
257 * before cpu_init(), SMP booting is too fragile that we want to
258 * limit the things done here to the most necessary things.
259 */
260 cr4_init();
261
262 /*
263 * 32-bit specific. 64-bit reaches this code with the correct page
264 * table established. Yet another historical divergence.
265 */
266 if (IS_ENABLED(CONFIG_X86_32)) {
267 /* switch away from the initial page table */
268 load_cr3(swapper_pg_dir);
269 __flush_tlb_all();
270 }
271
272 cpu_init_exception_handling();
273
274 /*
275 * Load the microcode before reaching the AP alive synchronization
276 * point below so it is not part of the full per CPU serialized
277 * bringup part when "parallel" bringup is enabled.
278 *
279 * That's even safe when hyperthreading is enabled in the CPU as
280 * the core code starts the primary threads first and leaves the
281 * secondary threads waiting for SIPI. Loading microcode on
282 * physical cores concurrently is a safe operation.
283 *
284 * This covers both the Intel specific issue that concurrent
285 * microcode loading on SMT siblings must be prohibited and the
286 * vendor independent issue`that microcode loading which changes
287 * CPUID, MSRs etc. must be strictly serialized to maintain
288 * software state correctness.
289 */
290 load_ucode_ap();
291
292 /*
293 * Synchronization point with the hotplug core. Sets this CPUs
294 * synchronization state to ALIVE and spin-waits for the control CPU to
295 * release this CPU for further bringup.
296 */
297 cpuhp_ap_sync_alive();
298
299 cpu_init();
300 fpu__init_cpu();
301 rcutree_report_cpu_starting(raw_smp_processor_id());
302 x86_cpuinit.early_percpu_clock_init();
303
304 ap_starting();
305
306 /* Check TSC synchronization with the control CPU. */
307 check_tsc_sync_target();
308
309 /*
310 * Calibrate the delay loop after the TSC synchronization check.
311 * This allows to skip the calibration when TSC is synchronized
312 * across sockets.
313 */
314 ap_calibrate_delay();
315
316 speculative_store_bypass_ht_init();
317
318 /*
319 * Lock vector_lock, set CPU online and bring the vector
320 * allocator online. Online must be set with vector_lock held
321 * to prevent a concurrent irq setup/teardown from seeing a
322 * half valid vector space.
323 */
324 lock_vector_lock();
325 set_cpu_online(smp_processor_id(), true);
326 lapic_online();
327 unlock_vector_lock();
328 x86_platform.nmi_init();
329
330 /* enable local interrupts */
331 local_irq_enable();
332
333 x86_cpuinit.setup_percpu_clockev();
334
335 wmb();
336 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
337}
338
339/**
340 * topology_phys_to_logical_pkg - Map a physical package id to a logical
341 * @phys_pkg: The physical package id to map
342 *
343 * Returns logical package id or -1 if not found
344 */
345int topology_phys_to_logical_pkg(unsigned int phys_pkg)
346{
347 int cpu;
348
349 for_each_possible_cpu(cpu) {
350 if (per_cpu(logical_maps.phys_pkg_id, cpu) == phys_pkg)
351 return per_cpu(logical_maps.logical_pkg_id, cpu);
352 }
353 return -1;
354}
355EXPORT_SYMBOL(topology_phys_to_logical_pkg);
356
357/**
358 * topology_phys_to_logical_die - Map a physical die id to logical
359 * @die_id: The physical die id to map
360 * @cur_cpu: The CPU for which the mapping is done
361 *
362 * Returns logical die id or -1 if not found
363 */
364static int topology_phys_to_logical_die(unsigned int die_id, unsigned int cur_cpu)
365{
366 int cpu, proc_id = cpu_data(cur_cpu).topo.pkg_id;
367
368 for_each_possible_cpu(cpu) {
369 if (per_cpu(logical_maps.phys_pkg_id, cpu) == proc_id &&
370 per_cpu(logical_maps.phys_die_id, cpu) == die_id)
371 return per_cpu(logical_maps.logical_die_id, cpu);
372 }
373 return -1;
374}
375
376/**
377 * topology_update_package_map - Update the physical to logical package map
378 * @pkg: The physical package id as retrieved via CPUID
379 * @cpu: The cpu for which this is updated
380 */
381int topology_update_package_map(unsigned int pkg, unsigned int cpu)
382{
383 int new;
384
385 /* Already available somewhere? */
386 new = topology_phys_to_logical_pkg(pkg);
387 if (new >= 0)
388 goto found;
389
390 new = logical_packages++;
391 if (new != pkg) {
392 pr_info("CPU %u Converting physical %u to logical package %u\n",
393 cpu, pkg, new);
394 }
395found:
396 per_cpu(logical_maps.phys_pkg_id, cpu) = pkg;
397 per_cpu(logical_maps.logical_pkg_id, cpu) = new;
398 cpu_data(cpu).topo.logical_pkg_id = new;
399 return 0;
400}
401/**
402 * topology_update_die_map - Update the physical to logical die map
403 * @die: The die id as retrieved via CPUID
404 * @cpu: The cpu for which this is updated
405 */
406int topology_update_die_map(unsigned int die, unsigned int cpu)
407{
408 int new;
409
410 /* Already available somewhere? */
411 new = topology_phys_to_logical_die(die, cpu);
412 if (new >= 0)
413 goto found;
414
415 new = logical_die++;
416 if (new != die) {
417 pr_info("CPU %u Converting physical %u to logical die %u\n",
418 cpu, die, new);
419 }
420found:
421 per_cpu(logical_maps.phys_die_id, cpu) = die;
422 per_cpu(logical_maps.logical_die_id, cpu) = new;
423 cpu_data(cpu).topo.logical_die_id = new;
424 return 0;
425}
426
427static void __init smp_store_boot_cpu_info(void)
428{
429 int id = 0; /* CPU 0 */
430 struct cpuinfo_x86 *c = &cpu_data(id);
431
432 *c = boot_cpu_data;
433 c->cpu_index = id;
434 topology_update_package_map(c->topo.pkg_id, id);
435 topology_update_die_map(c->topo.die_id, id);
436 c->initialized = true;
437}
438
439/*
440 * The bootstrap kernel entry code has set these up. Save them for
441 * a given CPU
442 */
443void smp_store_cpu_info(int id)
444{
445 struct cpuinfo_x86 *c = &cpu_data(id);
446
447 /* Copy boot_cpu_data only on the first bringup */
448 if (!c->initialized)
449 *c = boot_cpu_data;
450 c->cpu_index = id;
451 /*
452 * During boot time, CPU0 has this setup already. Save the info when
453 * bringing up an AP.
454 */
455 identify_secondary_cpu(c);
456 c->initialized = true;
457}
458
459static bool
460topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
461{
462 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
463
464 return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
465}
466
467static bool
468topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
469{
470 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
471
472 return !WARN_ONCE(!topology_same_node(c, o),
473 "sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
474 "[node: %d != %d]. Ignoring dependency.\n",
475 cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
476}
477
478#define link_mask(mfunc, c1, c2) \
479do { \
480 cpumask_set_cpu((c1), mfunc(c2)); \
481 cpumask_set_cpu((c2), mfunc(c1)); \
482} while (0)
483
484static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
485{
486 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
487 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
488
489 if (c->topo.pkg_id == o->topo.pkg_id &&
490 c->topo.die_id == o->topo.die_id &&
491 per_cpu_llc_id(cpu1) == per_cpu_llc_id(cpu2)) {
492 if (c->topo.core_id == o->topo.core_id)
493 return topology_sane(c, o, "smt");
494
495 if ((c->topo.cu_id != 0xff) &&
496 (o->topo.cu_id != 0xff) &&
497 (c->topo.cu_id == o->topo.cu_id))
498 return topology_sane(c, o, "smt");
499 }
500
501 } else if (c->topo.pkg_id == o->topo.pkg_id &&
502 c->topo.die_id == o->topo.die_id &&
503 c->topo.core_id == o->topo.core_id) {
504 return topology_sane(c, o, "smt");
505 }
506
507 return false;
508}
509
510static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
511{
512 if (c->topo.pkg_id == o->topo.pkg_id &&
513 c->topo.die_id == o->topo.die_id)
514 return true;
515 return false;
516}
517
518static bool match_l2c(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
519{
520 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
521
522 /* If the arch didn't set up l2c_id, fall back to SMT */
523 if (per_cpu_l2c_id(cpu1) == BAD_APICID)
524 return match_smt(c, o);
525
526 /* Do not match if L2 cache id does not match: */
527 if (per_cpu_l2c_id(cpu1) != per_cpu_l2c_id(cpu2))
528 return false;
529
530 return topology_sane(c, o, "l2c");
531}
532
533/*
534 * Unlike the other levels, we do not enforce keeping a
535 * multicore group inside a NUMA node. If this happens, we will
536 * discard the MC level of the topology later.
537 */
538static bool match_pkg(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
539{
540 if (c->topo.pkg_id == o->topo.pkg_id)
541 return true;
542 return false;
543}
544
545/*
546 * Define intel_cod_cpu[] for Intel COD (Cluster-on-Die) CPUs.
547 *
548 * Any Intel CPU that has multiple nodes per package and does not
549 * match intel_cod_cpu[] has the SNC (Sub-NUMA Cluster) topology.
550 *
551 * When in SNC mode, these CPUs enumerate an LLC that is shared
552 * by multiple NUMA nodes. The LLC is shared for off-package data
553 * access but private to the NUMA node (half of the package) for
554 * on-package access. CPUID (the source of the information about
555 * the LLC) can only enumerate the cache as shared or unshared,
556 * but not this particular configuration.
557 */
558
559static const struct x86_cpu_id intel_cod_cpu[] = {
560 X86_MATCH_INTEL_FAM6_MODEL(HASWELL_X, 0), /* COD */
561 X86_MATCH_INTEL_FAM6_MODEL(BROADWELL_X, 0), /* COD */
562 X86_MATCH_INTEL_FAM6_MODEL(ANY, 1), /* SNC */
563 {}
564};
565
566static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
567{
568 const struct x86_cpu_id *id = x86_match_cpu(intel_cod_cpu);
569 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
570 bool intel_snc = id && id->driver_data;
571
572 /* Do not match if we do not have a valid APICID for cpu: */
573 if (per_cpu_llc_id(cpu1) == BAD_APICID)
574 return false;
575
576 /* Do not match if LLC id does not match: */
577 if (per_cpu_llc_id(cpu1) != per_cpu_llc_id(cpu2))
578 return false;
579
580 /*
581 * Allow the SNC topology without warning. Return of false
582 * means 'c' does not share the LLC of 'o'. This will be
583 * reflected to userspace.
584 */
585 if (match_pkg(c, o) && !topology_same_node(c, o) && intel_snc)
586 return false;
587
588 return topology_sane(c, o, "llc");
589}
590
591
592static inline int x86_sched_itmt_flags(void)
593{
594 return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
595}
596
597#ifdef CONFIG_SCHED_MC
598static int x86_core_flags(void)
599{
600 return cpu_core_flags() | x86_sched_itmt_flags();
601}
602#endif
603#ifdef CONFIG_SCHED_SMT
604static int x86_smt_flags(void)
605{
606 return cpu_smt_flags();
607}
608#endif
609#ifdef CONFIG_SCHED_CLUSTER
610static int x86_cluster_flags(void)
611{
612 return cpu_cluster_flags() | x86_sched_itmt_flags();
613}
614#endif
615
616static int x86_die_flags(void)
617{
618 if (cpu_feature_enabled(X86_FEATURE_HYBRID_CPU))
619 return x86_sched_itmt_flags();
620
621 return 0;
622}
623
624/*
625 * Set if a package/die has multiple NUMA nodes inside.
626 * AMD Magny-Cours, Intel Cluster-on-Die, and Intel
627 * Sub-NUMA Clustering have this.
628 */
629static bool x86_has_numa_in_package;
630
631static struct sched_domain_topology_level x86_topology[6];
632
633static void __init build_sched_topology(void)
634{
635 int i = 0;
636
637#ifdef CONFIG_SCHED_SMT
638 x86_topology[i++] = (struct sched_domain_topology_level){
639 cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT)
640 };
641#endif
642#ifdef CONFIG_SCHED_CLUSTER
643 x86_topology[i++] = (struct sched_domain_topology_level){
644 cpu_clustergroup_mask, x86_cluster_flags, SD_INIT_NAME(CLS)
645 };
646#endif
647#ifdef CONFIG_SCHED_MC
648 x86_topology[i++] = (struct sched_domain_topology_level){
649 cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC)
650 };
651#endif
652 /*
653 * When there is NUMA topology inside the package skip the PKG domain
654 * since the NUMA domains will auto-magically create the right spanning
655 * domains based on the SLIT.
656 */
657 if (!x86_has_numa_in_package) {
658 x86_topology[i++] = (struct sched_domain_topology_level){
659 cpu_cpu_mask, x86_die_flags, SD_INIT_NAME(PKG)
660 };
661 }
662
663 /*
664 * There must be one trailing NULL entry left.
665 */
666 BUG_ON(i >= ARRAY_SIZE(x86_topology)-1);
667
668 set_sched_topology(x86_topology);
669}
670
671void set_cpu_sibling_map(int cpu)
672{
673 bool has_smt = smp_num_siblings > 1;
674 bool has_mp = has_smt || boot_cpu_data.x86_max_cores > 1;
675 struct cpuinfo_x86 *c = &cpu_data(cpu);
676 struct cpuinfo_x86 *o;
677 int i, threads;
678
679 cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
680
681 if (!has_mp) {
682 cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
683 cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
684 cpumask_set_cpu(cpu, cpu_l2c_shared_mask(cpu));
685 cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
686 cpumask_set_cpu(cpu, topology_die_cpumask(cpu));
687 c->booted_cores = 1;
688 return;
689 }
690
691 for_each_cpu(i, cpu_sibling_setup_mask) {
692 o = &cpu_data(i);
693
694 if (match_pkg(c, o) && !topology_same_node(c, o))
695 x86_has_numa_in_package = true;
696
697 if ((i == cpu) || (has_smt && match_smt(c, o)))
698 link_mask(topology_sibling_cpumask, cpu, i);
699
700 if ((i == cpu) || (has_mp && match_llc(c, o)))
701 link_mask(cpu_llc_shared_mask, cpu, i);
702
703 if ((i == cpu) || (has_mp && match_l2c(c, o)))
704 link_mask(cpu_l2c_shared_mask, cpu, i);
705
706 if ((i == cpu) || (has_mp && match_die(c, o)))
707 link_mask(topology_die_cpumask, cpu, i);
708 }
709
710 threads = cpumask_weight(topology_sibling_cpumask(cpu));
711 if (threads > __max_smt_threads)
712 __max_smt_threads = threads;
713
714 for_each_cpu(i, topology_sibling_cpumask(cpu))
715 cpu_data(i).smt_active = threads > 1;
716
717 /*
718 * This needs a separate iteration over the cpus because we rely on all
719 * topology_sibling_cpumask links to be set-up.
720 */
721 for_each_cpu(i, cpu_sibling_setup_mask) {
722 o = &cpu_data(i);
723
724 if ((i == cpu) || (has_mp && match_pkg(c, o))) {
725 link_mask(topology_core_cpumask, cpu, i);
726
727 /*
728 * Does this new cpu bringup a new core?
729 */
730 if (threads == 1) {
731 /*
732 * for each core in package, increment
733 * the booted_cores for this new cpu
734 */
735 if (cpumask_first(
736 topology_sibling_cpumask(i)) == i)
737 c->booted_cores++;
738 /*
739 * increment the core count for all
740 * the other cpus in this package
741 */
742 if (i != cpu)
743 cpu_data(i).booted_cores++;
744 } else if (i != cpu && !c->booted_cores)
745 c->booted_cores = cpu_data(i).booted_cores;
746 }
747 }
748}
749
750/* maps the cpu to the sched domain representing multi-core */
751const struct cpumask *cpu_coregroup_mask(int cpu)
752{
753 return cpu_llc_shared_mask(cpu);
754}
755
756const struct cpumask *cpu_clustergroup_mask(int cpu)
757{
758 return cpu_l2c_shared_mask(cpu);
759}
760EXPORT_SYMBOL_GPL(cpu_clustergroup_mask);
761
762static void impress_friends(void)
763{
764 int cpu;
765 unsigned long bogosum = 0;
766 /*
767 * Allow the user to impress friends.
768 */
769 pr_debug("Before bogomips\n");
770 for_each_online_cpu(cpu)
771 bogosum += cpu_data(cpu).loops_per_jiffy;
772
773 pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
774 num_online_cpus(),
775 bogosum/(500000/HZ),
776 (bogosum/(5000/HZ))%100);
777
778 pr_debug("Before bogocount - setting activated=1\n");
779}
780
781/*
782 * The Multiprocessor Specification 1.4 (1997) example code suggests
783 * that there should be a 10ms delay between the BSP asserting INIT
784 * and de-asserting INIT, when starting a remote processor.
785 * But that slows boot and resume on modern processors, which include
786 * many cores and don't require that delay.
787 *
788 * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
789 * Modern processor families are quirked to remove the delay entirely.
790 */
791#define UDELAY_10MS_DEFAULT 10000
792
793static unsigned int init_udelay = UINT_MAX;
794
795static int __init cpu_init_udelay(char *str)
796{
797 get_option(&str, &init_udelay);
798
799 return 0;
800}
801early_param("cpu_init_udelay", cpu_init_udelay);
802
803static void __init smp_quirk_init_udelay(void)
804{
805 /* if cmdline changed it from default, leave it alone */
806 if (init_udelay != UINT_MAX)
807 return;
808
809 /* if modern processor, use no delay */
810 if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
811 ((boot_cpu_data.x86_vendor == X86_VENDOR_HYGON) && (boot_cpu_data.x86 >= 0x18)) ||
812 ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
813 init_udelay = 0;
814 return;
815 }
816 /* else, use legacy delay */
817 init_udelay = UDELAY_10MS_DEFAULT;
818}
819
820/*
821 * Wake up AP by INIT, INIT, STARTUP sequence.
822 */
823static void send_init_sequence(u32 phys_apicid)
824{
825 int maxlvt = lapic_get_maxlvt();
826
827 /* Be paranoid about clearing APIC errors. */
828 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
829 /* Due to the Pentium erratum 3AP. */
830 if (maxlvt > 3)
831 apic_write(APIC_ESR, 0);
832 apic_read(APIC_ESR);
833 }
834
835 /* Assert INIT on the target CPU */
836 apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT, phys_apicid);
837 safe_apic_wait_icr_idle();
838
839 udelay(init_udelay);
840
841 /* Deassert INIT on the target CPU */
842 apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
843 safe_apic_wait_icr_idle();
844}
845
846/*
847 * Wake up AP by INIT, INIT, STARTUP sequence.
848 */
849static int wakeup_secondary_cpu_via_init(u32 phys_apicid, unsigned long start_eip)
850{
851 unsigned long send_status = 0, accept_status = 0;
852 int num_starts, j, maxlvt;
853
854 preempt_disable();
855 maxlvt = lapic_get_maxlvt();
856 send_init_sequence(phys_apicid);
857
858 mb();
859
860 /*
861 * Should we send STARTUP IPIs ?
862 *
863 * Determine this based on the APIC version.
864 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
865 */
866 if (APIC_INTEGRATED(boot_cpu_apic_version))
867 num_starts = 2;
868 else
869 num_starts = 0;
870
871 /*
872 * Run STARTUP IPI loop.
873 */
874 pr_debug("#startup loops: %d\n", num_starts);
875
876 for (j = 1; j <= num_starts; j++) {
877 pr_debug("Sending STARTUP #%d\n", j);
878 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
879 apic_write(APIC_ESR, 0);
880 apic_read(APIC_ESR);
881 pr_debug("After apic_write\n");
882
883 /*
884 * STARTUP IPI
885 */
886
887 /* Target chip */
888 /* Boot on the stack */
889 /* Kick the second */
890 apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
891 phys_apicid);
892
893 /*
894 * Give the other CPU some time to accept the IPI.
895 */
896 if (init_udelay == 0)
897 udelay(10);
898 else
899 udelay(300);
900
901 pr_debug("Startup point 1\n");
902
903 pr_debug("Waiting for send to finish...\n");
904 send_status = safe_apic_wait_icr_idle();
905
906 /*
907 * Give the other CPU some time to accept the IPI.
908 */
909 if (init_udelay == 0)
910 udelay(10);
911 else
912 udelay(200);
913
914 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
915 apic_write(APIC_ESR, 0);
916 accept_status = (apic_read(APIC_ESR) & 0xEF);
917 if (send_status || accept_status)
918 break;
919 }
920 pr_debug("After Startup\n");
921
922 if (send_status)
923 pr_err("APIC never delivered???\n");
924 if (accept_status)
925 pr_err("APIC delivery error (%lx)\n", accept_status);
926
927 preempt_enable();
928 return (send_status | accept_status);
929}
930
931/* reduce the number of lines printed when booting a large cpu count system */
932static void announce_cpu(int cpu, int apicid)
933{
934 static int width, node_width, first = 1;
935 static int current_node = NUMA_NO_NODE;
936 int node = early_cpu_to_node(cpu);
937
938 if (!width)
939 width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
940
941 if (!node_width)
942 node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
943
944 if (system_state < SYSTEM_RUNNING) {
945 if (first)
946 pr_info("x86: Booting SMP configuration:\n");
947
948 if (node != current_node) {
949 if (current_node > (-1))
950 pr_cont("\n");
951 current_node = node;
952
953 printk(KERN_INFO ".... node %*s#%d, CPUs: ",
954 node_width - num_digits(node), " ", node);
955 }
956
957 /* Add padding for the BSP */
958 if (first)
959 pr_cont("%*s", width + 1, " ");
960 first = 0;
961
962 pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
963 } else
964 pr_info("Booting Node %d Processor %d APIC 0x%x\n",
965 node, cpu, apicid);
966}
967
968int common_cpu_up(unsigned int cpu, struct task_struct *idle)
969{
970 int ret;
971
972 /* Just in case we booted with a single CPU. */
973 alternatives_enable_smp();
974
975 per_cpu(pcpu_hot.current_task, cpu) = idle;
976 cpu_init_stack_canary(cpu, idle);
977
978 /* Initialize the interrupt stack(s) */
979 ret = irq_init_percpu_irqstack(cpu);
980 if (ret)
981 return ret;
982
983#ifdef CONFIG_X86_32
984 /* Stack for startup_32 can be just as for start_secondary onwards */
985 per_cpu(pcpu_hot.top_of_stack, cpu) = task_top_of_stack(idle);
986#endif
987 return 0;
988}
989
990/*
991 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
992 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
993 * Returns zero if startup was successfully sent, else error code from
994 * ->wakeup_secondary_cpu.
995 */
996static int do_boot_cpu(u32 apicid, int cpu, struct task_struct *idle)
997{
998 unsigned long start_ip = real_mode_header->trampoline_start;
999 int ret;
1000
1001#ifdef CONFIG_X86_64
1002 /* If 64-bit wakeup method exists, use the 64-bit mode trampoline IP */
1003 if (apic->wakeup_secondary_cpu_64)
1004 start_ip = real_mode_header->trampoline_start64;
1005#endif
1006 idle->thread.sp = (unsigned long)task_pt_regs(idle);
1007 initial_code = (unsigned long)start_secondary;
1008
1009 if (IS_ENABLED(CONFIG_X86_32)) {
1010 early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
1011 initial_stack = idle->thread.sp;
1012 } else if (!(smpboot_control & STARTUP_PARALLEL_MASK)) {
1013 smpboot_control = cpu;
1014 }
1015
1016 /* Enable the espfix hack for this CPU */
1017 init_espfix_ap(cpu);
1018
1019 /* So we see what's up */
1020 announce_cpu(cpu, apicid);
1021
1022 /*
1023 * This grunge runs the startup process for
1024 * the targeted processor.
1025 */
1026 if (x86_platform.legacy.warm_reset) {
1027
1028 pr_debug("Setting warm reset code and vector.\n");
1029
1030 smpboot_setup_warm_reset_vector(start_ip);
1031 /*
1032 * Be paranoid about clearing APIC errors.
1033 */
1034 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
1035 apic_write(APIC_ESR, 0);
1036 apic_read(APIC_ESR);
1037 }
1038 }
1039
1040 smp_mb();
1041
1042 /*
1043 * Wake up a CPU in difference cases:
1044 * - Use a method from the APIC driver if one defined, with wakeup
1045 * straight to 64-bit mode preferred over wakeup to RM.
1046 * Otherwise,
1047 * - Use an INIT boot APIC message
1048 */
1049 if (apic->wakeup_secondary_cpu_64)
1050 ret = apic->wakeup_secondary_cpu_64(apicid, start_ip);
1051 else if (apic->wakeup_secondary_cpu)
1052 ret = apic->wakeup_secondary_cpu(apicid, start_ip);
1053 else
1054 ret = wakeup_secondary_cpu_via_init(apicid, start_ip);
1055
1056 /* If the wakeup mechanism failed, cleanup the warm reset vector */
1057 if (ret)
1058 arch_cpuhp_cleanup_kick_cpu(cpu);
1059 return ret;
1060}
1061
1062int native_kick_ap(unsigned int cpu, struct task_struct *tidle)
1063{
1064 u32 apicid = apic->cpu_present_to_apicid(cpu);
1065 int err;
1066
1067 lockdep_assert_irqs_enabled();
1068
1069 pr_debug("++++++++++++++++++++=_---CPU UP %u\n", cpu);
1070
1071 if (apicid == BAD_APICID || !physid_isset(apicid, phys_cpu_present_map) ||
1072 !apic_id_valid(apicid)) {
1073 pr_err("%s: bad cpu %d\n", __func__, cpu);
1074 return -EINVAL;
1075 }
1076
1077 /*
1078 * Save current MTRR state in case it was changed since early boot
1079 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1080 */
1081 mtrr_save_state();
1082
1083 /* the FPU context is blank, nobody can own it */
1084 per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
1085
1086 err = common_cpu_up(cpu, tidle);
1087 if (err)
1088 return err;
1089
1090 err = do_boot_cpu(apicid, cpu, tidle);
1091 if (err)
1092 pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
1093
1094 return err;
1095}
1096
1097int arch_cpuhp_kick_ap_alive(unsigned int cpu, struct task_struct *tidle)
1098{
1099 return smp_ops.kick_ap_alive(cpu, tidle);
1100}
1101
1102void arch_cpuhp_cleanup_kick_cpu(unsigned int cpu)
1103{
1104 /* Cleanup possible dangling ends... */
1105 if (smp_ops.kick_ap_alive == native_kick_ap && x86_platform.legacy.warm_reset)
1106 smpboot_restore_warm_reset_vector();
1107}
1108
1109void arch_cpuhp_cleanup_dead_cpu(unsigned int cpu)
1110{
1111 if (smp_ops.cleanup_dead_cpu)
1112 smp_ops.cleanup_dead_cpu(cpu);
1113
1114 if (system_state == SYSTEM_RUNNING)
1115 pr_info("CPU %u is now offline\n", cpu);
1116}
1117
1118void arch_cpuhp_sync_state_poll(void)
1119{
1120 if (smp_ops.poll_sync_state)
1121 smp_ops.poll_sync_state();
1122}
1123
1124/**
1125 * arch_disable_smp_support() - Disables SMP support for x86 at boottime
1126 */
1127void __init arch_disable_smp_support(void)
1128{
1129 disable_ioapic_support();
1130}
1131
1132/*
1133 * Fall back to non SMP mode after errors.
1134 *
1135 * RED-PEN audit/test this more. I bet there is more state messed up here.
1136 */
1137static __init void disable_smp(void)
1138{
1139 pr_info("SMP disabled\n");
1140
1141 disable_ioapic_support();
1142
1143 init_cpu_present(cpumask_of(0));
1144 init_cpu_possible(cpumask_of(0));
1145
1146 if (smp_found_config)
1147 physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
1148 else
1149 physid_set_mask_of_physid(0, &phys_cpu_present_map);
1150 cpumask_set_cpu(0, topology_sibling_cpumask(0));
1151 cpumask_set_cpu(0, topology_core_cpumask(0));
1152 cpumask_set_cpu(0, topology_die_cpumask(0));
1153}
1154
1155static void __init smp_cpu_index_default(void)
1156{
1157 int i;
1158 struct cpuinfo_x86 *c;
1159
1160 for_each_possible_cpu(i) {
1161 c = &cpu_data(i);
1162 /* mark all to hotplug */
1163 c->cpu_index = nr_cpu_ids;
1164 }
1165}
1166
1167void __init smp_prepare_cpus_common(void)
1168{
1169 unsigned int i;
1170
1171 smp_cpu_index_default();
1172
1173 /*
1174 * Setup boot CPU information
1175 */
1176 smp_store_boot_cpu_info(); /* Final full version of the data */
1177 mb();
1178
1179 for_each_possible_cpu(i) {
1180 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1181 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1182 zalloc_cpumask_var(&per_cpu(cpu_die_map, i), GFP_KERNEL);
1183 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1184 zalloc_cpumask_var(&per_cpu(cpu_l2c_shared_map, i), GFP_KERNEL);
1185 }
1186
1187 set_cpu_sibling_map(0);
1188}
1189
1190#ifdef CONFIG_X86_64
1191/* Establish whether parallel bringup can be supported. */
1192bool __init arch_cpuhp_init_parallel_bringup(void)
1193{
1194 if (!x86_cpuinit.parallel_bringup) {
1195 pr_info("Parallel CPU startup disabled by the platform\n");
1196 return false;
1197 }
1198
1199 smpboot_control = STARTUP_READ_APICID;
1200 pr_debug("Parallel CPU startup enabled: 0x%08x\n", smpboot_control);
1201 return true;
1202}
1203#endif
1204
1205/*
1206 * Prepare for SMP bootup.
1207 * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1208 * for common interface support.
1209 */
1210void __init native_smp_prepare_cpus(unsigned int max_cpus)
1211{
1212 smp_prepare_cpus_common();
1213
1214 switch (apic_intr_mode) {
1215 case APIC_PIC:
1216 case APIC_VIRTUAL_WIRE_NO_CONFIG:
1217 disable_smp();
1218 return;
1219 case APIC_SYMMETRIC_IO_NO_ROUTING:
1220 disable_smp();
1221 /* Setup local timer */
1222 x86_init.timers.setup_percpu_clockev();
1223 return;
1224 case APIC_VIRTUAL_WIRE:
1225 case APIC_SYMMETRIC_IO:
1226 break;
1227 }
1228
1229 /* Setup local timer */
1230 x86_init.timers.setup_percpu_clockev();
1231
1232 pr_info("CPU0: ");
1233 print_cpu_info(&cpu_data(0));
1234
1235 uv_system_init();
1236
1237 smp_quirk_init_udelay();
1238
1239 speculative_store_bypass_ht_init();
1240
1241 snp_set_wakeup_secondary_cpu();
1242}
1243
1244void arch_thaw_secondary_cpus_begin(void)
1245{
1246 set_cache_aps_delayed_init(true);
1247}
1248
1249void arch_thaw_secondary_cpus_end(void)
1250{
1251 cache_aps_init();
1252}
1253
1254/*
1255 * Early setup to make printk work.
1256 */
1257void __init native_smp_prepare_boot_cpu(void)
1258{
1259 int me = smp_processor_id();
1260
1261 /* SMP handles this from setup_per_cpu_areas() */
1262 if (!IS_ENABLED(CONFIG_SMP))
1263 switch_gdt_and_percpu_base(me);
1264
1265 native_pv_lock_init();
1266}
1267
1268void __init calculate_max_logical_packages(void)
1269{
1270 int ncpus;
1271
1272 /*
1273 * Today neither Intel nor AMD support heterogeneous systems so
1274 * extrapolate the boot cpu's data to all packages.
1275 */
1276 ncpus = cpu_data(0).booted_cores * topology_max_smt_threads();
1277 __max_logical_packages = DIV_ROUND_UP(total_cpus, ncpus);
1278 pr_info("Max logical packages: %u\n", __max_logical_packages);
1279}
1280
1281void __init native_smp_cpus_done(unsigned int max_cpus)
1282{
1283 pr_debug("Boot done\n");
1284
1285 calculate_max_logical_packages();
1286 build_sched_topology();
1287 nmi_selftest();
1288 impress_friends();
1289 cache_aps_init();
1290}
1291
1292static int __initdata setup_possible_cpus = -1;
1293static int __init _setup_possible_cpus(char *str)
1294{
1295 get_option(&str, &setup_possible_cpus);
1296 return 0;
1297}
1298early_param("possible_cpus", _setup_possible_cpus);
1299
1300
1301/*
1302 * cpu_possible_mask should be static, it cannot change as cpu's
1303 * are onlined, or offlined. The reason is per-cpu data-structures
1304 * are allocated by some modules at init time, and don't expect to
1305 * do this dynamically on cpu arrival/departure.
1306 * cpu_present_mask on the other hand can change dynamically.
1307 * In case when cpu_hotplug is not compiled, then we resort to current
1308 * behaviour, which is cpu_possible == cpu_present.
1309 * - Ashok Raj
1310 *
1311 * Three ways to find out the number of additional hotplug CPUs:
1312 * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1313 * - The user can overwrite it with possible_cpus=NUM
1314 * - Otherwise don't reserve additional CPUs.
1315 * We do this because additional CPUs waste a lot of memory.
1316 * -AK
1317 */
1318__init void prefill_possible_map(void)
1319{
1320 int i, possible;
1321
1322 i = setup_max_cpus ?: 1;
1323 if (setup_possible_cpus == -1) {
1324 possible = num_processors;
1325#ifdef CONFIG_HOTPLUG_CPU
1326 if (setup_max_cpus)
1327 possible += disabled_cpus;
1328#else
1329 if (possible > i)
1330 possible = i;
1331#endif
1332 } else
1333 possible = setup_possible_cpus;
1334
1335 total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1336
1337 /* nr_cpu_ids could be reduced via nr_cpus= */
1338 if (possible > nr_cpu_ids) {
1339 pr_warn("%d Processors exceeds NR_CPUS limit of %u\n",
1340 possible, nr_cpu_ids);
1341 possible = nr_cpu_ids;
1342 }
1343
1344#ifdef CONFIG_HOTPLUG_CPU
1345 if (!setup_max_cpus)
1346#endif
1347 if (possible > i) {
1348 pr_warn("%d Processors exceeds max_cpus limit of %u\n",
1349 possible, setup_max_cpus);
1350 possible = i;
1351 }
1352
1353 set_nr_cpu_ids(possible);
1354
1355 pr_info("Allowing %d CPUs, %d hotplug CPUs\n",
1356 possible, max_t(int, possible - num_processors, 0));
1357
1358 reset_cpu_possible_mask();
1359
1360 for (i = 0; i < possible; i++)
1361 set_cpu_possible(i, true);
1362}
1363
1364/* correctly size the local cpu masks */
1365void __init setup_cpu_local_masks(void)
1366{
1367 alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
1368}
1369
1370#ifdef CONFIG_HOTPLUG_CPU
1371
1372/* Recompute SMT state for all CPUs on offline */
1373static void recompute_smt_state(void)
1374{
1375 int max_threads, cpu;
1376
1377 max_threads = 0;
1378 for_each_online_cpu (cpu) {
1379 int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1380
1381 if (threads > max_threads)
1382 max_threads = threads;
1383 }
1384 __max_smt_threads = max_threads;
1385}
1386
1387static void remove_siblinginfo(int cpu)
1388{
1389 int sibling;
1390 struct cpuinfo_x86 *c = &cpu_data(cpu);
1391
1392 for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1393 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1394 /*/
1395 * last thread sibling in this cpu core going down
1396 */
1397 if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1398 cpu_data(sibling).booted_cores--;
1399 }
1400
1401 for_each_cpu(sibling, topology_die_cpumask(cpu))
1402 cpumask_clear_cpu(cpu, topology_die_cpumask(sibling));
1403
1404 for_each_cpu(sibling, topology_sibling_cpumask(cpu)) {
1405 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1406 if (cpumask_weight(topology_sibling_cpumask(sibling)) == 1)
1407 cpu_data(sibling).smt_active = false;
1408 }
1409
1410 for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1411 cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1412 for_each_cpu(sibling, cpu_l2c_shared_mask(cpu))
1413 cpumask_clear_cpu(cpu, cpu_l2c_shared_mask(sibling));
1414 cpumask_clear(cpu_llc_shared_mask(cpu));
1415 cpumask_clear(cpu_l2c_shared_mask(cpu));
1416 cpumask_clear(topology_sibling_cpumask(cpu));
1417 cpumask_clear(topology_core_cpumask(cpu));
1418 cpumask_clear(topology_die_cpumask(cpu));
1419 c->topo.core_id = 0;
1420 c->booted_cores = 0;
1421 cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1422 recompute_smt_state();
1423}
1424
1425static void remove_cpu_from_maps(int cpu)
1426{
1427 set_cpu_online(cpu, false);
1428 numa_remove_cpu(cpu);
1429}
1430
1431void cpu_disable_common(void)
1432{
1433 int cpu = smp_processor_id();
1434
1435 remove_siblinginfo(cpu);
1436
1437 /* It's now safe to remove this processor from the online map */
1438 lock_vector_lock();
1439 remove_cpu_from_maps(cpu);
1440 unlock_vector_lock();
1441 fixup_irqs();
1442 lapic_offline();
1443}
1444
1445int native_cpu_disable(void)
1446{
1447 int ret;
1448
1449 ret = lapic_can_unplug_cpu();
1450 if (ret)
1451 return ret;
1452
1453 cpu_disable_common();
1454
1455 /*
1456 * Disable the local APIC. Otherwise IPI broadcasts will reach
1457 * it. It still responds normally to INIT, NMI, SMI, and SIPI
1458 * messages.
1459 *
1460 * Disabling the APIC must happen after cpu_disable_common()
1461 * which invokes fixup_irqs().
1462 *
1463 * Disabling the APIC preserves already set bits in IRR, but
1464 * an interrupt arriving after disabling the local APIC does not
1465 * set the corresponding IRR bit.
1466 *
1467 * fixup_irqs() scans IRR for set bits so it can raise a not
1468 * yet handled interrupt on the new destination CPU via an IPI
1469 * but obviously it can't do so for IRR bits which are not set.
1470 * IOW, interrupts arriving after disabling the local APIC will
1471 * be lost.
1472 */
1473 apic_soft_disable();
1474
1475 return 0;
1476}
1477
1478void play_dead_common(void)
1479{
1480 idle_task_exit();
1481
1482 cpuhp_ap_report_dead();
1483
1484 local_irq_disable();
1485}
1486
1487/*
1488 * We need to flush the caches before going to sleep, lest we have
1489 * dirty data in our caches when we come back up.
1490 */
1491static inline void mwait_play_dead(void)
1492{
1493 struct mwait_cpu_dead *md = this_cpu_ptr(&mwait_cpu_dead);
1494 unsigned int eax, ebx, ecx, edx;
1495 unsigned int highest_cstate = 0;
1496 unsigned int highest_subcstate = 0;
1497 int i;
1498
1499 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD ||
1500 boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
1501 return;
1502 if (!this_cpu_has(X86_FEATURE_MWAIT))
1503 return;
1504 if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1505 return;
1506 if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1507 return;
1508
1509 eax = CPUID_MWAIT_LEAF;
1510 ecx = 0;
1511 native_cpuid(&eax, &ebx, &ecx, &edx);
1512
1513 /*
1514 * eax will be 0 if EDX enumeration is not valid.
1515 * Initialized below to cstate, sub_cstate value when EDX is valid.
1516 */
1517 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1518 eax = 0;
1519 } else {
1520 edx >>= MWAIT_SUBSTATE_SIZE;
1521 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1522 if (edx & MWAIT_SUBSTATE_MASK) {
1523 highest_cstate = i;
1524 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1525 }
1526 }
1527 eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1528 (highest_subcstate - 1);
1529 }
1530
1531 /* Set up state for the kexec() hack below */
1532 md->status = CPUDEAD_MWAIT_WAIT;
1533 md->control = CPUDEAD_MWAIT_WAIT;
1534
1535 wbinvd();
1536
1537 while (1) {
1538 /*
1539 * The CLFLUSH is a workaround for erratum AAI65 for
1540 * the Xeon 7400 series. It's not clear it is actually
1541 * needed, but it should be harmless in either case.
1542 * The WBINVD is insufficient due to the spurious-wakeup
1543 * case where we return around the loop.
1544 */
1545 mb();
1546 clflush(md);
1547 mb();
1548 __monitor(md, 0, 0);
1549 mb();
1550 __mwait(eax, 0);
1551
1552 if (READ_ONCE(md->control) == CPUDEAD_MWAIT_KEXEC_HLT) {
1553 /*
1554 * Kexec is about to happen. Don't go back into mwait() as
1555 * the kexec kernel might overwrite text and data including
1556 * page tables and stack. So mwait() would resume when the
1557 * monitor cache line is written to and then the CPU goes
1558 * south due to overwritten text, page tables and stack.
1559 *
1560 * Note: This does _NOT_ protect against a stray MCE, NMI,
1561 * SMI. They will resume execution at the instruction
1562 * following the HLT instruction and run into the problem
1563 * which this is trying to prevent.
1564 */
1565 WRITE_ONCE(md->status, CPUDEAD_MWAIT_KEXEC_HLT);
1566 while(1)
1567 native_halt();
1568 }
1569 }
1570}
1571
1572/*
1573 * Kick all "offline" CPUs out of mwait on kexec(). See comment in
1574 * mwait_play_dead().
1575 */
1576void smp_kick_mwait_play_dead(void)
1577{
1578 u32 newstate = CPUDEAD_MWAIT_KEXEC_HLT;
1579 struct mwait_cpu_dead *md;
1580 unsigned int cpu, i;
1581
1582 for_each_cpu_andnot(cpu, cpu_present_mask, cpu_online_mask) {
1583 md = per_cpu_ptr(&mwait_cpu_dead, cpu);
1584
1585 /* Does it sit in mwait_play_dead() ? */
1586 if (READ_ONCE(md->status) != CPUDEAD_MWAIT_WAIT)
1587 continue;
1588
1589 /* Wait up to 5ms */
1590 for (i = 0; READ_ONCE(md->status) != newstate && i < 1000; i++) {
1591 /* Bring it out of mwait */
1592 WRITE_ONCE(md->control, newstate);
1593 udelay(5);
1594 }
1595
1596 if (READ_ONCE(md->status) != newstate)
1597 pr_err_once("CPU%u is stuck in mwait_play_dead()\n", cpu);
1598 }
1599}
1600
1601void __noreturn hlt_play_dead(void)
1602{
1603 if (__this_cpu_read(cpu_info.x86) >= 4)
1604 wbinvd();
1605
1606 while (1)
1607 native_halt();
1608}
1609
1610/*
1611 * native_play_dead() is essentially a __noreturn function, but it can't
1612 * be marked as such as the compiler may complain about it.
1613 */
1614void native_play_dead(void)
1615{
1616 if (cpu_feature_enabled(X86_FEATURE_KERNEL_IBRS))
1617 __update_spec_ctrl(0);
1618
1619 play_dead_common();
1620 tboot_shutdown(TB_SHUTDOWN_WFS);
1621
1622 mwait_play_dead();
1623 if (cpuidle_play_dead())
1624 hlt_play_dead();
1625}
1626
1627#else /* ... !CONFIG_HOTPLUG_CPU */
1628int native_cpu_disable(void)
1629{
1630 return -ENOSYS;
1631}
1632
1633void native_play_dead(void)
1634{
1635 BUG();
1636}
1637
1638#endif