xref: /linux/arch/x86/kernel/smpboot.c (revision bf76f23aa1c178e9115eba17f699fa726aed669b)
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/cpuid/api.h>
68 #include <asm/desc.h>
69 #include <asm/nmi.h>
70 #include <asm/irq.h>
71 #include <asm/realmode.h>
72 #include <asm/cpu.h>
73 #include <asm/numa.h>
74 #include <asm/tlbflush.h>
75 #include <asm/mtrr.h>
76 #include <asm/mwait.h>
77 #include <asm/apic.h>
78 #include <asm/io_apic.h>
79 #include <asm/fpu/api.h>
80 #include <asm/setup.h>
81 #include <asm/uv/uv.h>
82 #include <asm/microcode.h>
83 #include <asm/i8259.h>
84 #include <asm/misc.h>
85 #include <asm/qspinlock.h>
86 #include <asm/intel-family.h>
87 #include <asm/cpu_device_id.h>
88 #include <asm/spec-ctrl.h>
89 #include <asm/hw_irq.h>
90 #include <asm/stackprotector.h>
91 #include <asm/sev.h>
92 #include <asm/spec-ctrl.h>
93 
94 /* representing HT siblings of each logical CPU */
95 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
96 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
97 
98 /* representing HT and core siblings of each logical CPU */
99 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
100 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
101 
102 /* representing HT, core, and die siblings of each logical CPU */
103 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_die_map);
104 EXPORT_PER_CPU_SYMBOL(cpu_die_map);
105 
106 /* CPUs which are the primary SMT threads */
107 struct cpumask __cpu_primary_thread_mask __read_mostly;
108 
109 /* Representing CPUs for which sibling maps can be computed */
110 static cpumask_var_t cpu_sibling_setup_mask;
111 
112 struct mwait_cpu_dead {
113 	unsigned int	control;
114 	unsigned int	status;
115 };
116 
117 #define CPUDEAD_MWAIT_WAIT	0xDEADBEEF
118 #define CPUDEAD_MWAIT_KEXEC_HLT	0x4A17DEAD
119 
120 /*
121  * Cache line aligned data for mwait_play_dead(). Separate on purpose so
122  * that it's unlikely to be touched by other CPUs.
123  */
124 static DEFINE_PER_CPU_ALIGNED(struct mwait_cpu_dead, mwait_cpu_dead);
125 
126 /* Maximum number of SMT threads on any online core */
127 int __read_mostly __max_smt_threads = 1;
128 
129 /* Flag to indicate if a complete sched domain rebuild is required */
130 bool x86_topology_update;
131 
arch_update_cpu_topology(void)132 int arch_update_cpu_topology(void)
133 {
134 	int retval = x86_topology_update;
135 
136 	x86_topology_update = false;
137 	return retval;
138 }
139 
140 static unsigned int smpboot_warm_reset_vector_count;
141 
smpboot_setup_warm_reset_vector(unsigned long start_eip)142 static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
143 {
144 	unsigned long flags;
145 
146 	spin_lock_irqsave(&rtc_lock, flags);
147 	if (!smpboot_warm_reset_vector_count++) {
148 		CMOS_WRITE(0xa, 0xf);
149 		*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) = start_eip >> 4;
150 		*((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = start_eip & 0xf;
151 	}
152 	spin_unlock_irqrestore(&rtc_lock, flags);
153 }
154 
smpboot_restore_warm_reset_vector(void)155 static inline void smpboot_restore_warm_reset_vector(void)
156 {
157 	unsigned long flags;
158 
159 	/*
160 	 * Paranoid:  Set warm reset code and vector here back
161 	 * to default values.
162 	 */
163 	spin_lock_irqsave(&rtc_lock, flags);
164 	if (!--smpboot_warm_reset_vector_count) {
165 		CMOS_WRITE(0, 0xf);
166 		*((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
167 	}
168 	spin_unlock_irqrestore(&rtc_lock, flags);
169 
170 }
171 
172 /* Run the next set of setup steps for the upcoming CPU */
ap_starting(void)173 static void ap_starting(void)
174 {
175 	int cpuid = smp_processor_id();
176 
177 	/* Mop up eventual mwait_play_dead() wreckage */
178 	this_cpu_write(mwait_cpu_dead.status, 0);
179 	this_cpu_write(mwait_cpu_dead.control, 0);
180 
181 	/*
182 	 * If woken up by an INIT in an 82489DX configuration the alive
183 	 * synchronization guarantees that the CPU does not reach this
184 	 * point before an INIT_deassert IPI reaches the local APIC, so it
185 	 * is now safe to touch the local APIC.
186 	 *
187 	 * Set up this CPU, first the APIC, which is probably redundant on
188 	 * most boards.
189 	 */
190 	apic_ap_setup();
191 
192 	/* Save the processor parameters. */
193 	identify_secondary_cpu(cpuid);
194 
195 	/*
196 	 * The topology information must be up to date before
197 	 * notify_cpu_starting().
198 	 */
199 	set_cpu_sibling_map(cpuid);
200 
201 	ap_init_aperfmperf();
202 
203 	pr_debug("Stack at about %p\n", &cpuid);
204 
205 	wmb();
206 
207 	/*
208 	 * This runs the AP through all the cpuhp states to its target
209 	 * state CPUHP_ONLINE.
210 	 */
211 	notify_cpu_starting(cpuid);
212 }
213 
ap_calibrate_delay(void)214 static void ap_calibrate_delay(void)
215 {
216 	/*
217 	 * Calibrate the delay loop and update loops_per_jiffy in cpu_data.
218 	 * identify_secondary_cpu() stored a value that is close but not as
219 	 * accurate as the value just calculated.
220 	 *
221 	 * As this is invoked after the TSC synchronization check,
222 	 * calibrate_delay_is_known() will skip the calibration routine
223 	 * when TSC is synchronized across sockets.
224 	 */
225 	calibrate_delay();
226 	cpu_data(smp_processor_id()).loops_per_jiffy = loops_per_jiffy;
227 }
228 
229 /*
230  * Activate a secondary processor.
231  */
start_secondary(void * unused)232 static void notrace __noendbr start_secondary(void *unused)
233 {
234 	/*
235 	 * Don't put *anything* except direct CPU state initialization
236 	 * before cpu_init(), SMP booting is too fragile that we want to
237 	 * limit the things done here to the most necessary things.
238 	 */
239 	cr4_init();
240 
241 	/*
242 	 * 32-bit specific. 64-bit reaches this code with the correct page
243 	 * table established. Yet another historical divergence.
244 	 */
245 	if (IS_ENABLED(CONFIG_X86_32)) {
246 		/* switch away from the initial page table */
247 		load_cr3(swapper_pg_dir);
248 		__flush_tlb_all();
249 	}
250 
251 	cpu_init_exception_handling(false);
252 
253 	/*
254 	 * Load the microcode before reaching the AP alive synchronization
255 	 * point below so it is not part of the full per CPU serialized
256 	 * bringup part when "parallel" bringup is enabled.
257 	 *
258 	 * That's even safe when hyperthreading is enabled in the CPU as
259 	 * the core code starts the primary threads first and leaves the
260 	 * secondary threads waiting for SIPI. Loading microcode on
261 	 * physical cores concurrently is a safe operation.
262 	 *
263 	 * This covers both the Intel specific issue that concurrent
264 	 * microcode loading on SMT siblings must be prohibited and the
265 	 * vendor independent issue`that microcode loading which changes
266 	 * CPUID, MSRs etc. must be strictly serialized to maintain
267 	 * software state correctness.
268 	 */
269 	load_ucode_ap();
270 
271 	/*
272 	 * Synchronization point with the hotplug core. Sets this CPUs
273 	 * synchronization state to ALIVE and spin-waits for the control CPU to
274 	 * release this CPU for further bringup.
275 	 */
276 	cpuhp_ap_sync_alive();
277 
278 	cpu_init();
279 	fpu__init_cpu();
280 	rcutree_report_cpu_starting(raw_smp_processor_id());
281 	x86_cpuinit.early_percpu_clock_init();
282 
283 	ap_starting();
284 
285 	/* Check TSC synchronization with the control CPU. */
286 	check_tsc_sync_target();
287 
288 	/*
289 	 * Calibrate the delay loop after the TSC synchronization check.
290 	 * This allows to skip the calibration when TSC is synchronized
291 	 * across sockets.
292 	 */
293 	ap_calibrate_delay();
294 
295 	speculative_store_bypass_ht_init();
296 
297 	/*
298 	 * Lock vector_lock, set CPU online and bring the vector
299 	 * allocator online. Online must be set with vector_lock held
300 	 * to prevent a concurrent irq setup/teardown from seeing a
301 	 * half valid vector space.
302 	 */
303 	lock_vector_lock();
304 	set_cpu_online(smp_processor_id(), true);
305 	lapic_online();
306 	unlock_vector_lock();
307 	x86_platform.nmi_init();
308 
309 	/* enable local interrupts */
310 	local_irq_enable();
311 
312 	x86_cpuinit.setup_percpu_clockev();
313 
314 	wmb();
315 	cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
316 }
317 ANNOTATE_NOENDBR_SYM(start_secondary);
318 
319 static bool
topology_same_node(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o)320 topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
321 {
322 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
323 
324 	return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
325 }
326 
327 static bool
topology_sane(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o,const char * name)328 topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
329 {
330 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
331 
332 	return !WARN_ONCE(!topology_same_node(c, o),
333 		"sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
334 		"[node: %d != %d]. Ignoring dependency.\n",
335 		cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
336 }
337 
338 #define link_mask(mfunc, c1, c2)					\
339 do {									\
340 	cpumask_set_cpu((c1), mfunc(c2));				\
341 	cpumask_set_cpu((c2), mfunc(c1));				\
342 } while (0)
343 
match_smt(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o)344 static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
345 {
346 	if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
347 		int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
348 
349 		if (c->topo.pkg_id == o->topo.pkg_id &&
350 		    c->topo.die_id == o->topo.die_id &&
351 		    c->topo.amd_node_id == o->topo.amd_node_id &&
352 		    per_cpu_llc_id(cpu1) == per_cpu_llc_id(cpu2)) {
353 			if (c->topo.core_id == o->topo.core_id)
354 				return topology_sane(c, o, "smt");
355 
356 			if ((c->topo.cu_id != 0xff) &&
357 			    (o->topo.cu_id != 0xff) &&
358 			    (c->topo.cu_id == o->topo.cu_id))
359 				return topology_sane(c, o, "smt");
360 		}
361 
362 	} else if (c->topo.pkg_id == o->topo.pkg_id &&
363 		   c->topo.die_id == o->topo.die_id &&
364 		   c->topo.core_id == o->topo.core_id) {
365 		return topology_sane(c, o, "smt");
366 	}
367 
368 	return false;
369 }
370 
match_die(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o)371 static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
372 {
373 	if (c->topo.pkg_id != o->topo.pkg_id || c->topo.die_id != o->topo.die_id)
374 		return false;
375 
376 	if (cpu_feature_enabled(X86_FEATURE_TOPOEXT) && topology_amd_nodes_per_pkg() > 1)
377 		return c->topo.amd_node_id == o->topo.amd_node_id;
378 
379 	return true;
380 }
381 
match_l2c(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o)382 static bool match_l2c(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
383 {
384 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
385 
386 	/* If the arch didn't set up l2c_id, fall back to SMT */
387 	if (per_cpu_l2c_id(cpu1) == BAD_APICID)
388 		return match_smt(c, o);
389 
390 	/* Do not match if L2 cache id does not match: */
391 	if (per_cpu_l2c_id(cpu1) != per_cpu_l2c_id(cpu2))
392 		return false;
393 
394 	return topology_sane(c, o, "l2c");
395 }
396 
397 /*
398  * Unlike the other levels, we do not enforce keeping a
399  * multicore group inside a NUMA node.  If this happens, we will
400  * discard the MC level of the topology later.
401  */
match_pkg(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o)402 static bool match_pkg(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
403 {
404 	if (c->topo.pkg_id == o->topo.pkg_id)
405 		return true;
406 	return false;
407 }
408 
409 /*
410  * Define intel_cod_cpu[] for Intel COD (Cluster-on-Die) CPUs.
411  *
412  * Any Intel CPU that has multiple nodes per package and does not
413  * match intel_cod_cpu[] has the SNC (Sub-NUMA Cluster) topology.
414  *
415  * When in SNC mode, these CPUs enumerate an LLC that is shared
416  * by multiple NUMA nodes. The LLC is shared for off-package data
417  * access but private to the NUMA node (half of the package) for
418  * on-package access. CPUID (the source of the information about
419  * the LLC) can only enumerate the cache as shared or unshared,
420  * but not this particular configuration.
421  */
422 
423 static const struct x86_cpu_id intel_cod_cpu[] = {
424 	X86_MATCH_VFM(INTEL_HASWELL_X,	 0),	/* COD */
425 	X86_MATCH_VFM(INTEL_BROADWELL_X, 0),	/* COD */
426 	X86_MATCH_VFM(INTEL_ANY,	 1),	/* SNC */
427 	{}
428 };
429 
match_llc(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o)430 static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
431 {
432 	const struct x86_cpu_id *id = x86_match_cpu(intel_cod_cpu);
433 	int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
434 	bool intel_snc = id && id->driver_data;
435 
436 	/* Do not match if we do not have a valid APICID for cpu: */
437 	if (per_cpu_llc_id(cpu1) == BAD_APICID)
438 		return false;
439 
440 	/* Do not match if LLC id does not match: */
441 	if (per_cpu_llc_id(cpu1) != per_cpu_llc_id(cpu2))
442 		return false;
443 
444 	/*
445 	 * Allow the SNC topology without warning. Return of false
446 	 * means 'c' does not share the LLC of 'o'. This will be
447 	 * reflected to userspace.
448 	 */
449 	if (match_pkg(c, o) && !topology_same_node(c, o) && intel_snc)
450 		return false;
451 
452 	return topology_sane(c, o, "llc");
453 }
454 
455 
x86_sched_itmt_flags(void)456 static inline int x86_sched_itmt_flags(void)
457 {
458 	return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
459 }
460 
461 #ifdef CONFIG_SCHED_MC
x86_core_flags(void)462 static int x86_core_flags(void)
463 {
464 	return cpu_core_flags() | x86_sched_itmt_flags();
465 }
466 #endif
467 #ifdef CONFIG_SCHED_CLUSTER
x86_cluster_flags(void)468 static int x86_cluster_flags(void)
469 {
470 	return cpu_cluster_flags() | x86_sched_itmt_flags();
471 }
472 #endif
473 
474 /*
475  * Set if a package/die has multiple NUMA nodes inside.
476  * AMD Magny-Cours, Intel Cluster-on-Die, and Intel
477  * Sub-NUMA Clustering have this.
478  */
479 static bool x86_has_numa_in_package;
480 
481 static struct sched_domain_topology_level x86_topology[] = {
482 	SDTL_INIT(cpu_smt_mask, cpu_smt_flags, SMT),
483 #ifdef CONFIG_SCHED_CLUSTER
484 	SDTL_INIT(cpu_clustergroup_mask, x86_cluster_flags, CLS),
485 #endif
486 #ifdef CONFIG_SCHED_MC
487 	SDTL_INIT(cpu_coregroup_mask, x86_core_flags, MC),
488 #endif
489 	SDTL_INIT(cpu_cpu_mask, x86_sched_itmt_flags, PKG),
490 	{ NULL },
491 };
492 
build_sched_topology(void)493 static void __init build_sched_topology(void)
494 {
495 	struct sched_domain_topology_level *topology = x86_topology;
496 
497 	/*
498 	 * When there is NUMA topology inside the package invalidate the
499 	 * PKG domain since the NUMA domains will auto-magically create the
500 	 * right spanning domains based on the SLIT.
501 	 */
502 	if (x86_has_numa_in_package) {
503 		unsigned int pkgdom = ARRAY_SIZE(x86_topology) - 2;
504 
505 		memset(&x86_topology[pkgdom], 0, sizeof(x86_topology[pkgdom]));
506 	}
507 
508 	/*
509 	 * Drop the SMT domains if there is only one thread per-core
510 	 * since it'll get degenerated by the scheduler anyways.
511 	 */
512 	if (cpu_smt_num_threads <= 1)
513 		++topology;
514 
515 	set_sched_topology(topology);
516 }
517 
set_cpu_sibling_map(int cpu)518 void set_cpu_sibling_map(int cpu)
519 {
520 	bool has_smt = __max_threads_per_core > 1;
521 	bool has_mp = has_smt || topology_num_cores_per_package() > 1;
522 	struct cpuinfo_x86 *c = &cpu_data(cpu);
523 	struct cpuinfo_x86 *o;
524 	int i, threads;
525 
526 	cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
527 
528 	if (!has_mp) {
529 		cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
530 		cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
531 		cpumask_set_cpu(cpu, cpu_l2c_shared_mask(cpu));
532 		cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
533 		cpumask_set_cpu(cpu, topology_die_cpumask(cpu));
534 		c->booted_cores = 1;
535 		return;
536 	}
537 
538 	for_each_cpu(i, cpu_sibling_setup_mask) {
539 		o = &cpu_data(i);
540 
541 		if (match_pkg(c, o) && !topology_same_node(c, o))
542 			x86_has_numa_in_package = true;
543 
544 		if ((i == cpu) || (has_smt && match_smt(c, o)))
545 			link_mask(topology_sibling_cpumask, cpu, i);
546 
547 		if ((i == cpu) || (has_mp && match_llc(c, o)))
548 			link_mask(cpu_llc_shared_mask, cpu, i);
549 
550 		if ((i == cpu) || (has_mp && match_l2c(c, o)))
551 			link_mask(cpu_l2c_shared_mask, cpu, i);
552 
553 		if ((i == cpu) || (has_mp && match_die(c, o)))
554 			link_mask(topology_die_cpumask, cpu, i);
555 	}
556 
557 	threads = cpumask_weight(topology_sibling_cpumask(cpu));
558 	if (threads > __max_smt_threads)
559 		__max_smt_threads = threads;
560 
561 	for_each_cpu(i, topology_sibling_cpumask(cpu))
562 		cpu_data(i).smt_active = threads > 1;
563 
564 	/*
565 	 * This needs a separate iteration over the cpus because we rely on all
566 	 * topology_sibling_cpumask links to be set-up.
567 	 */
568 	for_each_cpu(i, cpu_sibling_setup_mask) {
569 		o = &cpu_data(i);
570 
571 		if ((i == cpu) || (has_mp && match_pkg(c, o))) {
572 			link_mask(topology_core_cpumask, cpu, i);
573 
574 			/*
575 			 *  Does this new cpu bringup a new core?
576 			 */
577 			if (threads == 1) {
578 				/*
579 				 * for each core in package, increment
580 				 * the booted_cores for this new cpu
581 				 */
582 				if (cpumask_first(
583 				    topology_sibling_cpumask(i)) == i)
584 					c->booted_cores++;
585 				/*
586 				 * increment the core count for all
587 				 * the other cpus in this package
588 				 */
589 				if (i != cpu)
590 					cpu_data(i).booted_cores++;
591 			} else if (i != cpu && !c->booted_cores)
592 				c->booted_cores = cpu_data(i).booted_cores;
593 		}
594 	}
595 }
596 
597 /* maps the cpu to the sched domain representing multi-core */
cpu_coregroup_mask(int cpu)598 const struct cpumask *cpu_coregroup_mask(int cpu)
599 {
600 	return cpu_llc_shared_mask(cpu);
601 }
602 
cpu_clustergroup_mask(int cpu)603 const struct cpumask *cpu_clustergroup_mask(int cpu)
604 {
605 	return cpu_l2c_shared_mask(cpu);
606 }
607 EXPORT_SYMBOL_GPL(cpu_clustergroup_mask);
608 
impress_friends(void)609 static void impress_friends(void)
610 {
611 	int cpu;
612 	unsigned long bogosum = 0;
613 	/*
614 	 * Allow the user to impress friends.
615 	 */
616 	pr_debug("Before bogomips\n");
617 	for_each_online_cpu(cpu)
618 		bogosum += cpu_data(cpu).loops_per_jiffy;
619 
620 	pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
621 		num_online_cpus(),
622 		bogosum/(500000/HZ),
623 		(bogosum/(5000/HZ))%100);
624 
625 	pr_debug("Before bogocount - setting activated=1\n");
626 }
627 
628 /*
629  * The Multiprocessor Specification 1.4 (1997) example code suggests
630  * that there should be a 10ms delay between the BSP asserting INIT
631  * and de-asserting INIT, when starting a remote processor.
632  * But that slows boot and resume on modern processors, which include
633  * many cores and don't require that delay.
634  *
635  * Cmdline "cpu_init_udelay=" is available to override this delay.
636  */
637 #define UDELAY_10MS_LEGACY 10000
638 
639 static unsigned int init_udelay = UINT_MAX;
640 
cpu_init_udelay(char * str)641 static int __init cpu_init_udelay(char *str)
642 {
643 	get_option(&str, &init_udelay);
644 
645 	return 0;
646 }
647 early_param("cpu_init_udelay", cpu_init_udelay);
648 
smp_set_init_udelay(void)649 static void __init smp_set_init_udelay(void)
650 {
651 	/* if cmdline changed it from default, leave it alone */
652 	if (init_udelay != UINT_MAX)
653 		return;
654 
655 	/* if modern processor, use no delay */
656 	if ((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL && boot_cpu_data.x86_vfm >= INTEL_PENTIUM_PRO) ||
657 	    (boot_cpu_data.x86_vendor == X86_VENDOR_HYGON && boot_cpu_data.x86 >= 0x18) ||
658 	    (boot_cpu_data.x86_vendor == X86_VENDOR_AMD   && boot_cpu_data.x86 >= 0xF)) {
659 		init_udelay = 0;
660 		return;
661 	}
662 	/* else, use legacy delay */
663 	init_udelay = UDELAY_10MS_LEGACY;
664 }
665 
666 /*
667  * Wake up AP by INIT, INIT, STARTUP sequence.
668  */
send_init_sequence(u32 phys_apicid)669 static void send_init_sequence(u32 phys_apicid)
670 {
671 	int maxlvt = lapic_get_maxlvt();
672 
673 	/* Be paranoid about clearing APIC errors. */
674 	if (APIC_INTEGRATED(boot_cpu_apic_version)) {
675 		/* Due to the Pentium erratum 3AP.  */
676 		if (maxlvt > 3)
677 			apic_write(APIC_ESR, 0);
678 		apic_read(APIC_ESR);
679 	}
680 
681 	/* Assert INIT on the target CPU */
682 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT, phys_apicid);
683 	safe_apic_wait_icr_idle();
684 
685 	udelay(init_udelay);
686 
687 	/* Deassert INIT on the target CPU */
688 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
689 	safe_apic_wait_icr_idle();
690 }
691 
692 /*
693  * Wake up AP by INIT, INIT, STARTUP sequence.
694  */
wakeup_secondary_cpu_via_init(u32 phys_apicid,unsigned long start_eip,unsigned int cpu)695 static int wakeup_secondary_cpu_via_init(u32 phys_apicid, unsigned long start_eip, unsigned int cpu)
696 {
697 	unsigned long send_status = 0, accept_status = 0;
698 	int num_starts, j, maxlvt;
699 
700 	preempt_disable();
701 	maxlvt = lapic_get_maxlvt();
702 	send_init_sequence(phys_apicid);
703 
704 	mb();
705 
706 	/*
707 	 * Should we send STARTUP IPIs ?
708 	 *
709 	 * Determine this based on the APIC version.
710 	 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
711 	 */
712 	if (APIC_INTEGRATED(boot_cpu_apic_version))
713 		num_starts = 2;
714 	else
715 		num_starts = 0;
716 
717 	/*
718 	 * Run STARTUP IPI loop.
719 	 */
720 	pr_debug("#startup loops: %d\n", num_starts);
721 
722 	for (j = 1; j <= num_starts; j++) {
723 		pr_debug("Sending STARTUP #%d\n", j);
724 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
725 			apic_write(APIC_ESR, 0);
726 		apic_read(APIC_ESR);
727 		pr_debug("After apic_write\n");
728 
729 		/*
730 		 * STARTUP IPI
731 		 */
732 
733 		/* Target chip */
734 		/* Boot on the stack */
735 		/* Kick the second */
736 		apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
737 			       phys_apicid);
738 
739 		/*
740 		 * Give the other CPU some time to accept the IPI.
741 		 */
742 		if (init_udelay == 0)
743 			udelay(10);
744 		else
745 			udelay(300);
746 
747 		pr_debug("Startup point 1\n");
748 
749 		pr_debug("Waiting for send to finish...\n");
750 		send_status = safe_apic_wait_icr_idle();
751 
752 		/*
753 		 * Give the other CPU some time to accept the IPI.
754 		 */
755 		if (init_udelay == 0)
756 			udelay(10);
757 		else
758 			udelay(200);
759 
760 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
761 			apic_write(APIC_ESR, 0);
762 		accept_status = (apic_read(APIC_ESR) & 0xEF);
763 		if (send_status || accept_status)
764 			break;
765 	}
766 	pr_debug("After Startup\n");
767 
768 	if (send_status)
769 		pr_err("APIC never delivered???\n");
770 	if (accept_status)
771 		pr_err("APIC delivery error (%lx)\n", accept_status);
772 
773 	preempt_enable();
774 	return (send_status | accept_status);
775 }
776 
777 /* reduce the number of lines printed when booting a large cpu count system */
announce_cpu(int cpu,int apicid)778 static void announce_cpu(int cpu, int apicid)
779 {
780 	static int width, node_width, first = 1;
781 	static int current_node = NUMA_NO_NODE;
782 	int node = early_cpu_to_node(cpu);
783 
784 	if (!width)
785 		width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
786 
787 	if (!node_width)
788 		node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
789 
790 	if (system_state < SYSTEM_RUNNING) {
791 		if (first)
792 			pr_info("x86: Booting SMP configuration:\n");
793 
794 		if (node != current_node) {
795 			if (current_node > (-1))
796 				pr_cont("\n");
797 			current_node = node;
798 
799 			printk(KERN_INFO ".... node %*s#%d, CPUs:  ",
800 			       node_width - num_digits(node), " ", node);
801 		}
802 
803 		/* Add padding for the BSP */
804 		if (first)
805 			pr_cont("%*s", width + 1, " ");
806 		first = 0;
807 
808 		pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
809 	} else
810 		pr_info("Booting Node %d Processor %d APIC 0x%x\n",
811 			node, cpu, apicid);
812 }
813 
common_cpu_up(unsigned int cpu,struct task_struct * idle)814 int common_cpu_up(unsigned int cpu, struct task_struct *idle)
815 {
816 	int ret;
817 
818 	/* Just in case we booted with a single CPU. */
819 	alternatives_enable_smp();
820 
821 	per_cpu(current_task, cpu) = idle;
822 	cpu_init_stack_canary(cpu, idle);
823 
824 	/* Initialize the interrupt stack(s) */
825 	ret = irq_init_percpu_irqstack(cpu);
826 	if (ret)
827 		return ret;
828 
829 #ifdef CONFIG_X86_32
830 	/* Stack for startup_32 can be just as for start_secondary onwards */
831 	per_cpu(cpu_current_top_of_stack, cpu) = task_top_of_stack(idle);
832 #endif
833 	return 0;
834 }
835 
836 /*
837  * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
838  * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
839  * Returns zero if startup was successfully sent, else error code from
840  * ->wakeup_secondary_cpu.
841  */
do_boot_cpu(u32 apicid,unsigned int cpu,struct task_struct * idle)842 static int do_boot_cpu(u32 apicid, unsigned int cpu, struct task_struct *idle)
843 {
844 	unsigned long start_ip = real_mode_header->trampoline_start;
845 	int ret;
846 
847 #ifdef CONFIG_X86_64
848 	/* If 64-bit wakeup method exists, use the 64-bit mode trampoline IP */
849 	if (apic->wakeup_secondary_cpu_64)
850 		start_ip = real_mode_header->trampoline_start64;
851 #endif
852 	idle->thread.sp = (unsigned long)task_pt_regs(idle);
853 	initial_code = (unsigned long)start_secondary;
854 
855 	if (IS_ENABLED(CONFIG_X86_32)) {
856 		early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
857 		initial_stack  = idle->thread.sp;
858 	} else if (!(smpboot_control & STARTUP_PARALLEL_MASK)) {
859 		smpboot_control = cpu;
860 	}
861 
862 	/* Enable the espfix hack for this CPU */
863 	init_espfix_ap(cpu);
864 
865 	/* So we see what's up */
866 	announce_cpu(cpu, apicid);
867 
868 	/*
869 	 * This grunge runs the startup process for
870 	 * the targeted processor.
871 	 */
872 	if (x86_platform.legacy.warm_reset) {
873 
874 		pr_debug("Setting warm reset code and vector.\n");
875 
876 		smpboot_setup_warm_reset_vector(start_ip);
877 		/*
878 		 * Be paranoid about clearing APIC errors.
879 		*/
880 		if (APIC_INTEGRATED(boot_cpu_apic_version)) {
881 			apic_write(APIC_ESR, 0);
882 			apic_read(APIC_ESR);
883 		}
884 	}
885 
886 	smp_mb();
887 
888 	/*
889 	 * Wake up a CPU in difference cases:
890 	 * - Use a method from the APIC driver if one defined, with wakeup
891 	 *   straight to 64-bit mode preferred over wakeup to RM.
892 	 * Otherwise,
893 	 * - Use an INIT boot APIC message
894 	 */
895 	if (apic->wakeup_secondary_cpu_64)
896 		ret = apic->wakeup_secondary_cpu_64(apicid, start_ip, cpu);
897 	else if (apic->wakeup_secondary_cpu)
898 		ret = apic->wakeup_secondary_cpu(apicid, start_ip, cpu);
899 	else
900 		ret = wakeup_secondary_cpu_via_init(apicid, start_ip, cpu);
901 
902 	/* If the wakeup mechanism failed, cleanup the warm reset vector */
903 	if (ret)
904 		arch_cpuhp_cleanup_kick_cpu(cpu);
905 	return ret;
906 }
907 
native_kick_ap(unsigned int cpu,struct task_struct * tidle)908 int native_kick_ap(unsigned int cpu, struct task_struct *tidle)
909 {
910 	u32 apicid = apic->cpu_present_to_apicid(cpu);
911 	int err;
912 
913 	lockdep_assert_irqs_enabled();
914 
915 	pr_debug("++++++++++++++++++++=_---CPU UP  %u\n", cpu);
916 
917 	if (apicid == BAD_APICID || !apic_id_valid(apicid)) {
918 		pr_err("CPU %u has invalid APIC ID %x. Aborting bringup\n", cpu, apicid);
919 		return -EINVAL;
920 	}
921 
922 	if (!test_bit(apicid, phys_cpu_present_map)) {
923 		pr_err("CPU %u APIC ID %x is not present. Aborting bringup\n", cpu, apicid);
924 		return -EINVAL;
925 	}
926 
927 	/*
928 	 * Save current MTRR state in case it was changed since early boot
929 	 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
930 	 */
931 	mtrr_save_state();
932 
933 	/* the FPU context is blank, nobody can own it */
934 	per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
935 
936 	err = common_cpu_up(cpu, tidle);
937 	if (err)
938 		return err;
939 
940 	err = do_boot_cpu(apicid, cpu, tidle);
941 	if (err)
942 		pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
943 
944 	return err;
945 }
946 
arch_cpuhp_kick_ap_alive(unsigned int cpu,struct task_struct * tidle)947 int arch_cpuhp_kick_ap_alive(unsigned int cpu, struct task_struct *tidle)
948 {
949 	return smp_ops.kick_ap_alive(cpu, tidle);
950 }
951 
arch_cpuhp_cleanup_kick_cpu(unsigned int cpu)952 void arch_cpuhp_cleanup_kick_cpu(unsigned int cpu)
953 {
954 	/* Cleanup possible dangling ends... */
955 	if (smp_ops.kick_ap_alive == native_kick_ap && x86_platform.legacy.warm_reset)
956 		smpboot_restore_warm_reset_vector();
957 }
958 
arch_cpuhp_cleanup_dead_cpu(unsigned int cpu)959 void arch_cpuhp_cleanup_dead_cpu(unsigned int cpu)
960 {
961 	if (smp_ops.cleanup_dead_cpu)
962 		smp_ops.cleanup_dead_cpu(cpu);
963 
964 	if (system_state == SYSTEM_RUNNING)
965 		pr_info("CPU %u is now offline\n", cpu);
966 }
967 
arch_cpuhp_sync_state_poll(void)968 void arch_cpuhp_sync_state_poll(void)
969 {
970 	if (smp_ops.poll_sync_state)
971 		smp_ops.poll_sync_state();
972 }
973 
974 /**
975  * arch_disable_smp_support() - Disables SMP support for x86 at boottime
976  */
arch_disable_smp_support(void)977 void __init arch_disable_smp_support(void)
978 {
979 	disable_ioapic_support();
980 }
981 
982 /*
983  * Fall back to non SMP mode after errors.
984  *
985  * RED-PEN audit/test this more. I bet there is more state messed up here.
986  */
disable_smp(void)987 static __init void disable_smp(void)
988 {
989 	pr_info("SMP disabled\n");
990 
991 	disable_ioapic_support();
992 	topology_reset_possible_cpus_up();
993 
994 	cpumask_set_cpu(0, topology_sibling_cpumask(0));
995 	cpumask_set_cpu(0, topology_core_cpumask(0));
996 	cpumask_set_cpu(0, topology_die_cpumask(0));
997 }
998 
smp_prepare_cpus_common(void)999 void __init smp_prepare_cpus_common(void)
1000 {
1001 	unsigned int cpu, node;
1002 
1003 	/* Mark all except the boot CPU as hotpluggable */
1004 	for_each_possible_cpu(cpu) {
1005 		if (cpu)
1006 			per_cpu(cpu_info.cpu_index, cpu) = nr_cpu_ids;
1007 	}
1008 
1009 	for_each_possible_cpu(cpu) {
1010 		node = cpu_to_node(cpu);
1011 
1012 		zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map,    cpu), GFP_KERNEL, node);
1013 		zalloc_cpumask_var_node(&per_cpu(cpu_core_map,       cpu), GFP_KERNEL, node);
1014 		zalloc_cpumask_var_node(&per_cpu(cpu_die_map,        cpu), GFP_KERNEL, node);
1015 		zalloc_cpumask_var_node(&per_cpu(cpu_llc_shared_map, cpu), GFP_KERNEL, node);
1016 		zalloc_cpumask_var_node(&per_cpu(cpu_l2c_shared_map, cpu), GFP_KERNEL, node);
1017 	}
1018 
1019 	set_cpu_sibling_map(0);
1020 }
1021 
smp_prepare_boot_cpu(void)1022 void __init smp_prepare_boot_cpu(void)
1023 {
1024 	smp_ops.smp_prepare_boot_cpu();
1025 }
1026 
1027 #ifdef CONFIG_X86_64
1028 /* Establish whether parallel bringup can be supported. */
arch_cpuhp_init_parallel_bringup(void)1029 bool __init arch_cpuhp_init_parallel_bringup(void)
1030 {
1031 	if (!x86_cpuinit.parallel_bringup) {
1032 		pr_info("Parallel CPU startup disabled by the platform\n");
1033 		return false;
1034 	}
1035 
1036 	smpboot_control = STARTUP_READ_APICID;
1037 	pr_debug("Parallel CPU startup enabled: 0x%08x\n", smpboot_control);
1038 	return true;
1039 }
1040 #endif
1041 
1042 /*
1043  * Prepare for SMP bootup.
1044  * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1045  *            for common interface support.
1046  */
native_smp_prepare_cpus(unsigned int max_cpus)1047 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1048 {
1049 	smp_prepare_cpus_common();
1050 
1051 	switch (apic_intr_mode) {
1052 	case APIC_PIC:
1053 	case APIC_VIRTUAL_WIRE_NO_CONFIG:
1054 		disable_smp();
1055 		return;
1056 	case APIC_SYMMETRIC_IO_NO_ROUTING:
1057 		disable_smp();
1058 		/* Setup local timer */
1059 		x86_init.timers.setup_percpu_clockev();
1060 		return;
1061 	case APIC_VIRTUAL_WIRE:
1062 	case APIC_SYMMETRIC_IO:
1063 		break;
1064 	}
1065 
1066 	/* Setup local timer */
1067 	x86_init.timers.setup_percpu_clockev();
1068 
1069 	pr_info("CPU0: ");
1070 	print_cpu_info(&cpu_data(0));
1071 
1072 	uv_system_init();
1073 
1074 	smp_set_init_udelay();
1075 
1076 	speculative_store_bypass_ht_init();
1077 
1078 	snp_set_wakeup_secondary_cpu();
1079 }
1080 
arch_thaw_secondary_cpus_begin(void)1081 void arch_thaw_secondary_cpus_begin(void)
1082 {
1083 	set_cache_aps_delayed_init(true);
1084 }
1085 
arch_thaw_secondary_cpus_end(void)1086 void arch_thaw_secondary_cpus_end(void)
1087 {
1088 	cache_aps_init();
1089 }
1090 
1091 /*
1092  * Early setup to make printk work.
1093  */
native_smp_prepare_boot_cpu(void)1094 void __init native_smp_prepare_boot_cpu(void)
1095 {
1096 	int me = smp_processor_id();
1097 
1098 	/* SMP handles this from setup_per_cpu_areas() */
1099 	if (!IS_ENABLED(CONFIG_SMP))
1100 		switch_gdt_and_percpu_base(me);
1101 
1102 	native_pv_lock_init();
1103 }
1104 
native_smp_cpus_done(unsigned int max_cpus)1105 void __init native_smp_cpus_done(unsigned int max_cpus)
1106 {
1107 	pr_debug("Boot done\n");
1108 
1109 	build_sched_topology();
1110 	nmi_selftest();
1111 	impress_friends();
1112 	cache_aps_init();
1113 }
1114 
1115 /* correctly size the local cpu masks */
setup_cpu_local_masks(void)1116 void __init setup_cpu_local_masks(void)
1117 {
1118 	alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
1119 }
1120 
1121 #ifdef CONFIG_HOTPLUG_CPU
1122 
1123 /* Recompute SMT state for all CPUs on offline */
recompute_smt_state(void)1124 static void recompute_smt_state(void)
1125 {
1126 	int max_threads, cpu;
1127 
1128 	max_threads = 0;
1129 	for_each_online_cpu (cpu) {
1130 		int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1131 
1132 		if (threads > max_threads)
1133 			max_threads = threads;
1134 	}
1135 	__max_smt_threads = max_threads;
1136 }
1137 
remove_siblinginfo(int cpu)1138 static void remove_siblinginfo(int cpu)
1139 {
1140 	int sibling;
1141 	struct cpuinfo_x86 *c = &cpu_data(cpu);
1142 
1143 	for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1144 		cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1145 		/*/
1146 		 * last thread sibling in this cpu core going down
1147 		 */
1148 		if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1149 			cpu_data(sibling).booted_cores--;
1150 	}
1151 
1152 	for_each_cpu(sibling, topology_die_cpumask(cpu))
1153 		cpumask_clear_cpu(cpu, topology_die_cpumask(sibling));
1154 
1155 	for_each_cpu(sibling, topology_sibling_cpumask(cpu)) {
1156 		cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1157 		if (cpumask_weight(topology_sibling_cpumask(sibling)) == 1)
1158 			cpu_data(sibling).smt_active = false;
1159 	}
1160 
1161 	for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1162 		cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1163 	for_each_cpu(sibling, cpu_l2c_shared_mask(cpu))
1164 		cpumask_clear_cpu(cpu, cpu_l2c_shared_mask(sibling));
1165 	cpumask_clear(cpu_llc_shared_mask(cpu));
1166 	cpumask_clear(cpu_l2c_shared_mask(cpu));
1167 	cpumask_clear(topology_sibling_cpumask(cpu));
1168 	cpumask_clear(topology_core_cpumask(cpu));
1169 	cpumask_clear(topology_die_cpumask(cpu));
1170 	c->topo.core_id = 0;
1171 	c->booted_cores = 0;
1172 	cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1173 	recompute_smt_state();
1174 }
1175 
remove_cpu_from_maps(int cpu)1176 static void remove_cpu_from_maps(int cpu)
1177 {
1178 	set_cpu_online(cpu, false);
1179 	numa_remove_cpu(cpu);
1180 }
1181 
cpu_disable_common(void)1182 void cpu_disable_common(void)
1183 {
1184 	int cpu = smp_processor_id();
1185 
1186 	remove_siblinginfo(cpu);
1187 
1188 	/*
1189 	 * Stop allowing kernel-mode FPU. This is needed so that if the CPU is
1190 	 * brought online again, the initial state is not allowed:
1191 	 */
1192 	this_cpu_write(kernel_fpu_allowed, false);
1193 
1194 	/* It's now safe to remove this processor from the online map */
1195 	lock_vector_lock();
1196 	remove_cpu_from_maps(cpu);
1197 	unlock_vector_lock();
1198 	fixup_irqs();
1199 	lapic_offline();
1200 }
1201 
native_cpu_disable(void)1202 int native_cpu_disable(void)
1203 {
1204 	int ret;
1205 
1206 	ret = lapic_can_unplug_cpu();
1207 	if (ret)
1208 		return ret;
1209 
1210 	cpu_disable_common();
1211 
1212         /*
1213          * Disable the local APIC. Otherwise IPI broadcasts will reach
1214          * it. It still responds normally to INIT, NMI, SMI, and SIPI
1215          * messages.
1216          *
1217          * Disabling the APIC must happen after cpu_disable_common()
1218          * which invokes fixup_irqs().
1219          *
1220          * Disabling the APIC preserves already set bits in IRR, but
1221          * an interrupt arriving after disabling the local APIC does not
1222          * set the corresponding IRR bit.
1223          *
1224          * fixup_irqs() scans IRR for set bits so it can raise a not
1225          * yet handled interrupt on the new destination CPU via an IPI
1226          * but obviously it can't do so for IRR bits which are not set.
1227          * IOW, interrupts arriving after disabling the local APIC will
1228          * be lost.
1229          */
1230 	apic_soft_disable();
1231 
1232 	return 0;
1233 }
1234 
play_dead_common(void)1235 void play_dead_common(void)
1236 {
1237 	idle_task_exit();
1238 
1239 	cpuhp_ap_report_dead();
1240 
1241 	local_irq_disable();
1242 }
1243 
1244 /*
1245  * We need to flush the caches before going to sleep, lest we have
1246  * dirty data in our caches when we come back up.
1247  */
mwait_play_dead(unsigned int eax_hint)1248 void __noreturn mwait_play_dead(unsigned int eax_hint)
1249 {
1250 	struct mwait_cpu_dead *md = this_cpu_ptr(&mwait_cpu_dead);
1251 
1252 	/* Set up state for the kexec() hack below */
1253 	md->status = CPUDEAD_MWAIT_WAIT;
1254 	md->control = CPUDEAD_MWAIT_WAIT;
1255 
1256 	wbinvd();
1257 
1258 	while (1) {
1259 		/*
1260 		 * The CLFLUSH is a workaround for erratum AAI65 for
1261 		 * the Xeon 7400 series.  It's not clear it is actually
1262 		 * needed, but it should be harmless in either case.
1263 		 * The WBINVD is insufficient due to the spurious-wakeup
1264 		 * case where we return around the loop.
1265 		 */
1266 		mb();
1267 		clflush(md);
1268 		mb();
1269 		__monitor(md, 0, 0);
1270 		mb();
1271 		__mwait(eax_hint, 0);
1272 
1273 		if (READ_ONCE(md->control) == CPUDEAD_MWAIT_KEXEC_HLT) {
1274 			/*
1275 			 * Kexec is about to happen. Don't go back into mwait() as
1276 			 * the kexec kernel might overwrite text and data including
1277 			 * page tables and stack. So mwait() would resume when the
1278 			 * monitor cache line is written to and then the CPU goes
1279 			 * south due to overwritten text, page tables and stack.
1280 			 *
1281 			 * Note: This does _NOT_ protect against a stray MCE, NMI,
1282 			 * SMI. They will resume execution at the instruction
1283 			 * following the HLT instruction and run into the problem
1284 			 * which this is trying to prevent.
1285 			 */
1286 			WRITE_ONCE(md->status, CPUDEAD_MWAIT_KEXEC_HLT);
1287 			while(1)
1288 				native_halt();
1289 		}
1290 	}
1291 }
1292 
1293 /*
1294  * Kick all "offline" CPUs out of mwait on kexec(). See comment in
1295  * mwait_play_dead().
1296  */
smp_kick_mwait_play_dead(void)1297 void smp_kick_mwait_play_dead(void)
1298 {
1299 	u32 newstate = CPUDEAD_MWAIT_KEXEC_HLT;
1300 	struct mwait_cpu_dead *md;
1301 	unsigned int cpu, i;
1302 
1303 	for_each_cpu_andnot(cpu, cpu_present_mask, cpu_online_mask) {
1304 		md = per_cpu_ptr(&mwait_cpu_dead, cpu);
1305 
1306 		/* Does it sit in mwait_play_dead() ? */
1307 		if (READ_ONCE(md->status) != CPUDEAD_MWAIT_WAIT)
1308 			continue;
1309 
1310 		/* Wait up to 5ms */
1311 		for (i = 0; READ_ONCE(md->status) != newstate && i < 1000; i++) {
1312 			/* Bring it out of mwait */
1313 			WRITE_ONCE(md->control, newstate);
1314 			udelay(5);
1315 		}
1316 
1317 		if (READ_ONCE(md->status) != newstate)
1318 			pr_err_once("CPU%u is stuck in mwait_play_dead()\n", cpu);
1319 	}
1320 }
1321 
hlt_play_dead(void)1322 void __noreturn hlt_play_dead(void)
1323 {
1324 	if (__this_cpu_read(cpu_info.x86) >= 4)
1325 		wbinvd();
1326 
1327 	while (1)
1328 		native_halt();
1329 }
1330 
1331 /*
1332  * native_play_dead() is essentially a __noreturn function, but it can't
1333  * be marked as such as the compiler may complain about it.
1334  */
native_play_dead(void)1335 void native_play_dead(void)
1336 {
1337 	if (cpu_feature_enabled(X86_FEATURE_KERNEL_IBRS))
1338 		__update_spec_ctrl(0);
1339 
1340 	play_dead_common();
1341 	tboot_shutdown(TB_SHUTDOWN_WFS);
1342 
1343 	/* Below returns only on error. */
1344 	cpuidle_play_dead();
1345 	hlt_play_dead();
1346 }
1347 
1348 #else /* ... !CONFIG_HOTPLUG_CPU */
native_cpu_disable(void)1349 int native_cpu_disable(void)
1350 {
1351 	return -ENOSYS;
1352 }
1353 
native_play_dead(void)1354 void native_play_dead(void)
1355 {
1356 	BUG();
1357 }
1358 
1359 #endif
1360