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 */
93 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
94 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
95
96 /* representing HT and core siblings of each logical CPU */
97 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
98 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
99
100 /* representing HT, core, and die siblings of each logical CPU */
101 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_die_map);
102 EXPORT_PER_CPU_SYMBOL(cpu_die_map);
103
104 /* CPUs which are the primary SMT threads */
105 struct cpumask __cpu_primary_thread_mask __read_mostly;
106
107 /* Representing CPUs for which sibling maps can be computed */
108 static cpumask_var_t cpu_sibling_setup_mask;
109
110 struct 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 */
122 static DEFINE_PER_CPU_ALIGNED(struct mwait_cpu_dead, mwait_cpu_dead);
123
124 /* Maximum number of SMT threads on any online core */
125 int __read_mostly __max_smt_threads = 1;
126
127 /* Flag to indicate if a complete sched domain rebuild is required */
128 bool x86_topology_update;
129
arch_update_cpu_topology(void)130 int arch_update_cpu_topology(void)
131 {
132 int retval = x86_topology_update;
133
134 x86_topology_update = false;
135 return retval;
136 }
137
138 static unsigned int smpboot_warm_reset_vector_count;
139
smpboot_setup_warm_reset_vector(unsigned long start_eip)140 static 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
smpboot_restore_warm_reset_vector(void)153 static 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 */
ap_starting(void)171 static 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
ap_calibrate_delay(void)212 static 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 */
start_secondary(void * unused)230 static 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 */
smp_store_cpu_info(int id)320 void 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
336 static bool
topology_same_node(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o)337 topology_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
344 static bool
topology_sane(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o,const char * name)345 topology_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) \
356 do { \
357 cpumask_set_cpu((c1), mfunc(c2)); \
358 cpumask_set_cpu((c2), mfunc(c1)); \
359 } while (0)
360
match_smt(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o)361 static 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
match_die(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o)388 static 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
match_l2c(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o)399 static 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 */
match_pkg(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o)419 static 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
440 static const struct x86_cpu_id intel_cod_cpu[] = {
441 X86_MATCH_VFM(INTEL_HASWELL_X, 0), /* COD */
442 X86_MATCH_VFM(INTEL_BROADWELL_X, 0), /* COD */
443 X86_MATCH_VFM(INTEL_ANY, 1), /* SNC */
444 {}
445 };
446
match_llc(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o)447 static 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
x86_sched_itmt_flags(void)473 static 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
x86_core_flags(void)479 static int x86_core_flags(void)
480 {
481 return cpu_core_flags() | x86_sched_itmt_flags();
482 }
483 #endif
484 #ifdef CONFIG_SCHED_SMT
x86_smt_flags(void)485 static int x86_smt_flags(void)
486 {
487 return cpu_smt_flags();
488 }
489 #endif
490 #ifdef CONFIG_SCHED_CLUSTER
x86_cluster_flags(void)491 static int x86_cluster_flags(void)
492 {
493 return cpu_cluster_flags() | x86_sched_itmt_flags();
494 }
495 #endif
496
x86_die_flags(void)497 static 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 */
510 static bool x86_has_numa_in_package;
511
512 static struct sched_domain_topology_level x86_topology[6];
513
build_sched_topology(void)514 static 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
set_cpu_sibling_map(int cpu)552 void 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 */
cpu_coregroup_mask(int cpu)632 const struct cpumask *cpu_coregroup_mask(int cpu)
633 {
634 return cpu_llc_shared_mask(cpu);
635 }
636
cpu_clustergroup_mask(int cpu)637 const struct cpumask *cpu_clustergroup_mask(int cpu)
638 {
639 return cpu_l2c_shared_mask(cpu);
640 }
641 EXPORT_SYMBOL_GPL(cpu_clustergroup_mask);
642
impress_friends(void)643 static 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
674 static unsigned int init_udelay = UINT_MAX;
675
cpu_init_udelay(char * str)676 static int __init cpu_init_udelay(char *str)
677 {
678 get_option(&str, &init_udelay);
679
680 return 0;
681 }
682 early_param("cpu_init_udelay", cpu_init_udelay);
683
smp_quirk_init_udelay(void)684 static 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 */
send_init_sequence(u32 phys_apicid)704 static 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 */
wakeup_secondary_cpu_via_init(u32 phys_apicid,unsigned long start_eip)730 static 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 */
announce_cpu(int cpu,int apicid)813 static 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
common_cpu_up(unsigned int cpu,struct task_struct * idle)849 int 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 */
do_boot_cpu(u32 apicid,int cpu,struct task_struct * idle)877 static 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
native_kick_ap(unsigned int cpu,struct task_struct * tidle)943 int 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
arch_cpuhp_kick_ap_alive(unsigned int cpu,struct task_struct * tidle)982 int arch_cpuhp_kick_ap_alive(unsigned int cpu, struct task_struct *tidle)
983 {
984 return smp_ops.kick_ap_alive(cpu, tidle);
985 }
986
arch_cpuhp_cleanup_kick_cpu(unsigned int cpu)987 void 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
arch_cpuhp_cleanup_dead_cpu(unsigned int cpu)994 void 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
arch_cpuhp_sync_state_poll(void)1003 void 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 */
arch_disable_smp_support(void)1012 void __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 */
disable_smp(void)1022 static __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
smp_prepare_cpus_common(void)1034 void __init smp_prepare_cpus_common(void)
1035 {
1036 unsigned int cpu, node;
1037
1038 /* Mark all except the boot CPU as hotpluggable */
1039 for_each_possible_cpu(cpu) {
1040 if (cpu)
1041 per_cpu(cpu_info.cpu_index, cpu) = nr_cpu_ids;
1042 }
1043
1044 for_each_possible_cpu(cpu) {
1045 node = cpu_to_node(cpu);
1046
1047 zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map, cpu), GFP_KERNEL, node);
1048 zalloc_cpumask_var_node(&per_cpu(cpu_core_map, cpu), GFP_KERNEL, node);
1049 zalloc_cpumask_var_node(&per_cpu(cpu_die_map, cpu), GFP_KERNEL, node);
1050 zalloc_cpumask_var_node(&per_cpu(cpu_llc_shared_map, cpu), GFP_KERNEL, node);
1051 zalloc_cpumask_var_node(&per_cpu(cpu_l2c_shared_map, cpu), GFP_KERNEL, node);
1052 }
1053
1054 set_cpu_sibling_map(0);
1055 }
1056
smp_prepare_boot_cpu(void)1057 void __init smp_prepare_boot_cpu(void)
1058 {
1059 smp_ops.smp_prepare_boot_cpu();
1060 }
1061
1062 #ifdef CONFIG_X86_64
1063 /* Establish whether parallel bringup can be supported. */
arch_cpuhp_init_parallel_bringup(void)1064 bool __init arch_cpuhp_init_parallel_bringup(void)
1065 {
1066 if (!x86_cpuinit.parallel_bringup) {
1067 pr_info("Parallel CPU startup disabled by the platform\n");
1068 return false;
1069 }
1070
1071 smpboot_control = STARTUP_READ_APICID;
1072 pr_debug("Parallel CPU startup enabled: 0x%08x\n", smpboot_control);
1073 return true;
1074 }
1075 #endif
1076
1077 /*
1078 * Prepare for SMP bootup.
1079 * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1080 * for common interface support.
1081 */
native_smp_prepare_cpus(unsigned int max_cpus)1082 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1083 {
1084 smp_prepare_cpus_common();
1085
1086 switch (apic_intr_mode) {
1087 case APIC_PIC:
1088 case APIC_VIRTUAL_WIRE_NO_CONFIG:
1089 disable_smp();
1090 return;
1091 case APIC_SYMMETRIC_IO_NO_ROUTING:
1092 disable_smp();
1093 /* Setup local timer */
1094 x86_init.timers.setup_percpu_clockev();
1095 return;
1096 case APIC_VIRTUAL_WIRE:
1097 case APIC_SYMMETRIC_IO:
1098 break;
1099 }
1100
1101 /* Setup local timer */
1102 x86_init.timers.setup_percpu_clockev();
1103
1104 pr_info("CPU0: ");
1105 print_cpu_info(&cpu_data(0));
1106
1107 uv_system_init();
1108
1109 smp_quirk_init_udelay();
1110
1111 speculative_store_bypass_ht_init();
1112
1113 snp_set_wakeup_secondary_cpu();
1114 }
1115
arch_thaw_secondary_cpus_begin(void)1116 void arch_thaw_secondary_cpus_begin(void)
1117 {
1118 set_cache_aps_delayed_init(true);
1119 }
1120
arch_thaw_secondary_cpus_end(void)1121 void arch_thaw_secondary_cpus_end(void)
1122 {
1123 cache_aps_init();
1124 }
1125
1126 /*
1127 * Early setup to make printk work.
1128 */
native_smp_prepare_boot_cpu(void)1129 void __init native_smp_prepare_boot_cpu(void)
1130 {
1131 int me = smp_processor_id();
1132
1133 /* SMP handles this from setup_per_cpu_areas() */
1134 if (!IS_ENABLED(CONFIG_SMP))
1135 switch_gdt_and_percpu_base(me);
1136
1137 native_pv_lock_init();
1138 }
1139
native_smp_cpus_done(unsigned int max_cpus)1140 void __init native_smp_cpus_done(unsigned int max_cpus)
1141 {
1142 pr_debug("Boot done\n");
1143
1144 build_sched_topology();
1145 nmi_selftest();
1146 impress_friends();
1147 cache_aps_init();
1148 }
1149
1150 /* correctly size the local cpu masks */
setup_cpu_local_masks(void)1151 void __init setup_cpu_local_masks(void)
1152 {
1153 alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
1154 }
1155
1156 #ifdef CONFIG_HOTPLUG_CPU
1157
1158 /* Recompute SMT state for all CPUs on offline */
recompute_smt_state(void)1159 static void recompute_smt_state(void)
1160 {
1161 int max_threads, cpu;
1162
1163 max_threads = 0;
1164 for_each_online_cpu (cpu) {
1165 int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1166
1167 if (threads > max_threads)
1168 max_threads = threads;
1169 }
1170 __max_smt_threads = max_threads;
1171 }
1172
remove_siblinginfo(int cpu)1173 static void remove_siblinginfo(int cpu)
1174 {
1175 int sibling;
1176 struct cpuinfo_x86 *c = &cpu_data(cpu);
1177
1178 for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1179 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1180 /*/
1181 * last thread sibling in this cpu core going down
1182 */
1183 if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1184 cpu_data(sibling).booted_cores--;
1185 }
1186
1187 for_each_cpu(sibling, topology_die_cpumask(cpu))
1188 cpumask_clear_cpu(cpu, topology_die_cpumask(sibling));
1189
1190 for_each_cpu(sibling, topology_sibling_cpumask(cpu)) {
1191 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1192 if (cpumask_weight(topology_sibling_cpumask(sibling)) == 1)
1193 cpu_data(sibling).smt_active = false;
1194 }
1195
1196 for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1197 cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1198 for_each_cpu(sibling, cpu_l2c_shared_mask(cpu))
1199 cpumask_clear_cpu(cpu, cpu_l2c_shared_mask(sibling));
1200 cpumask_clear(cpu_llc_shared_mask(cpu));
1201 cpumask_clear(cpu_l2c_shared_mask(cpu));
1202 cpumask_clear(topology_sibling_cpumask(cpu));
1203 cpumask_clear(topology_core_cpumask(cpu));
1204 cpumask_clear(topology_die_cpumask(cpu));
1205 c->topo.core_id = 0;
1206 c->booted_cores = 0;
1207 cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1208 recompute_smt_state();
1209 }
1210
remove_cpu_from_maps(int cpu)1211 static void remove_cpu_from_maps(int cpu)
1212 {
1213 set_cpu_online(cpu, false);
1214 numa_remove_cpu(cpu);
1215 }
1216
cpu_disable_common(void)1217 void cpu_disable_common(void)
1218 {
1219 int cpu = smp_processor_id();
1220
1221 remove_siblinginfo(cpu);
1222
1223 /* It's now safe to remove this processor from the online map */
1224 lock_vector_lock();
1225 remove_cpu_from_maps(cpu);
1226 unlock_vector_lock();
1227 fixup_irqs();
1228 lapic_offline();
1229 }
1230
native_cpu_disable(void)1231 int native_cpu_disable(void)
1232 {
1233 int ret;
1234
1235 ret = lapic_can_unplug_cpu();
1236 if (ret)
1237 return ret;
1238
1239 cpu_disable_common();
1240
1241 /*
1242 * Disable the local APIC. Otherwise IPI broadcasts will reach
1243 * it. It still responds normally to INIT, NMI, SMI, and SIPI
1244 * messages.
1245 *
1246 * Disabling the APIC must happen after cpu_disable_common()
1247 * which invokes fixup_irqs().
1248 *
1249 * Disabling the APIC preserves already set bits in IRR, but
1250 * an interrupt arriving after disabling the local APIC does not
1251 * set the corresponding IRR bit.
1252 *
1253 * fixup_irqs() scans IRR for set bits so it can raise a not
1254 * yet handled interrupt on the new destination CPU via an IPI
1255 * but obviously it can't do so for IRR bits which are not set.
1256 * IOW, interrupts arriving after disabling the local APIC will
1257 * be lost.
1258 */
1259 apic_soft_disable();
1260
1261 return 0;
1262 }
1263
play_dead_common(void)1264 void play_dead_common(void)
1265 {
1266 idle_task_exit();
1267
1268 cpuhp_ap_report_dead();
1269
1270 local_irq_disable();
1271 }
1272
1273 /*
1274 * We need to flush the caches before going to sleep, lest we have
1275 * dirty data in our caches when we come back up.
1276 */
mwait_play_dead(void)1277 static inline void mwait_play_dead(void)
1278 {
1279 struct mwait_cpu_dead *md = this_cpu_ptr(&mwait_cpu_dead);
1280 unsigned int eax, ebx, ecx, edx;
1281 unsigned int highest_cstate = 0;
1282 unsigned int highest_subcstate = 0;
1283 int i;
1284
1285 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD ||
1286 boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
1287 return;
1288 if (!this_cpu_has(X86_FEATURE_MWAIT))
1289 return;
1290 if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1291 return;
1292 if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1293 return;
1294
1295 eax = CPUID_MWAIT_LEAF;
1296 ecx = 0;
1297 native_cpuid(&eax, &ebx, &ecx, &edx);
1298
1299 /*
1300 * eax will be 0 if EDX enumeration is not valid.
1301 * Initialized below to cstate, sub_cstate value when EDX is valid.
1302 */
1303 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1304 eax = 0;
1305 } else {
1306 edx >>= MWAIT_SUBSTATE_SIZE;
1307 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1308 if (edx & MWAIT_SUBSTATE_MASK) {
1309 highest_cstate = i;
1310 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1311 }
1312 }
1313 eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1314 (highest_subcstate - 1);
1315 }
1316
1317 /* Set up state for the kexec() hack below */
1318 md->status = CPUDEAD_MWAIT_WAIT;
1319 md->control = CPUDEAD_MWAIT_WAIT;
1320
1321 wbinvd();
1322
1323 while (1) {
1324 /*
1325 * The CLFLUSH is a workaround for erratum AAI65 for
1326 * the Xeon 7400 series. It's not clear it is actually
1327 * needed, but it should be harmless in either case.
1328 * The WBINVD is insufficient due to the spurious-wakeup
1329 * case where we return around the loop.
1330 */
1331 mb();
1332 clflush(md);
1333 mb();
1334 __monitor(md, 0, 0);
1335 mb();
1336 __mwait(eax, 0);
1337
1338 if (READ_ONCE(md->control) == CPUDEAD_MWAIT_KEXEC_HLT) {
1339 /*
1340 * Kexec is about to happen. Don't go back into mwait() as
1341 * the kexec kernel might overwrite text and data including
1342 * page tables and stack. So mwait() would resume when the
1343 * monitor cache line is written to and then the CPU goes
1344 * south due to overwritten text, page tables and stack.
1345 *
1346 * Note: This does _NOT_ protect against a stray MCE, NMI,
1347 * SMI. They will resume execution at the instruction
1348 * following the HLT instruction and run into the problem
1349 * which this is trying to prevent.
1350 */
1351 WRITE_ONCE(md->status, CPUDEAD_MWAIT_KEXEC_HLT);
1352 while(1)
1353 native_halt();
1354 }
1355 }
1356 }
1357
1358 /*
1359 * Kick all "offline" CPUs out of mwait on kexec(). See comment in
1360 * mwait_play_dead().
1361 */
smp_kick_mwait_play_dead(void)1362 void smp_kick_mwait_play_dead(void)
1363 {
1364 u32 newstate = CPUDEAD_MWAIT_KEXEC_HLT;
1365 struct mwait_cpu_dead *md;
1366 unsigned int cpu, i;
1367
1368 for_each_cpu_andnot(cpu, cpu_present_mask, cpu_online_mask) {
1369 md = per_cpu_ptr(&mwait_cpu_dead, cpu);
1370
1371 /* Does it sit in mwait_play_dead() ? */
1372 if (READ_ONCE(md->status) != CPUDEAD_MWAIT_WAIT)
1373 continue;
1374
1375 /* Wait up to 5ms */
1376 for (i = 0; READ_ONCE(md->status) != newstate && i < 1000; i++) {
1377 /* Bring it out of mwait */
1378 WRITE_ONCE(md->control, newstate);
1379 udelay(5);
1380 }
1381
1382 if (READ_ONCE(md->status) != newstate)
1383 pr_err_once("CPU%u is stuck in mwait_play_dead()\n", cpu);
1384 }
1385 }
1386
hlt_play_dead(void)1387 void __noreturn hlt_play_dead(void)
1388 {
1389 if (__this_cpu_read(cpu_info.x86) >= 4)
1390 wbinvd();
1391
1392 while (1)
1393 native_halt();
1394 }
1395
1396 /*
1397 * native_play_dead() is essentially a __noreturn function, but it can't
1398 * be marked as such as the compiler may complain about it.
1399 */
native_play_dead(void)1400 void native_play_dead(void)
1401 {
1402 if (cpu_feature_enabled(X86_FEATURE_KERNEL_IBRS))
1403 __update_spec_ctrl(0);
1404
1405 play_dead_common();
1406 tboot_shutdown(TB_SHUTDOWN_WFS);
1407
1408 mwait_play_dead();
1409 if (cpuidle_play_dead())
1410 hlt_play_dead();
1411 }
1412
1413 #else /* ... !CONFIG_HOTPLUG_CPU */
native_cpu_disable(void)1414 int native_cpu_disable(void)
1415 {
1416 return -ENOSYS;
1417 }
1418
native_play_dead(void)1419 void native_play_dead(void)
1420 {
1421 BUG();
1422 }
1423
1424 #endif
1425