xref: /linux/arch/x86/kernel/smpboot.c (revision 3a39d672e7f48b8d6b91a09afa4b55352773b4b5)
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 */
94 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
95 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
96 
97 /* representing HT and core siblings of each logical CPU */
98 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
99 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
100 
101 /* representing HT, core, and die siblings of each logical CPU */
102 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_die_map);
103 EXPORT_PER_CPU_SYMBOL(cpu_die_map);
104 
105 /* CPUs which are the primary SMT threads */
106 struct cpumask __cpu_primary_thread_mask __read_mostly;
107 
108 /* Representing CPUs for which sibling maps can be computed */
109 static cpumask_var_t cpu_sibling_setup_mask;
110 
111 struct 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  */
123 static DEFINE_PER_CPU_ALIGNED(struct mwait_cpu_dead, mwait_cpu_dead);
124 
125 /* Maximum number of SMT threads on any online core */
126 int __read_mostly __max_smt_threads = 1;
127 
128 /* Flag to indicate if a complete sched domain rebuild is required */
129 bool x86_topology_update;
130 
arch_update_cpu_topology(void)131 int arch_update_cpu_topology(void)
132 {
133 	int retval = x86_topology_update;
134 
135 	x86_topology_update = false;
136 	return retval;
137 }
138 
139 static unsigned int smpboot_warm_reset_vector_count;
140 
smpboot_setup_warm_reset_vector(unsigned long start_eip)141 static 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 
smpboot_restore_warm_reset_vector(void)154 static 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 */
ap_starting(void)172 static 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 
ap_calibrate_delay(void)213 static 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  */
start_secondary(void * unused)231 static 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  */
smp_store_cpu_info(int id)321 void 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 
337 static bool
topology_same_node(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o)338 topology_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 
345 static bool
topology_sane(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o,const char * name)346 topology_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)					\
357 do {									\
358 	cpumask_set_cpu((c1), mfunc(c2));				\
359 	cpumask_set_cpu((c2), mfunc(c1));				\
360 } while (0)
361 
match_smt(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o)362 static 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 
match_die(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o)389 static 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 
match_l2c(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o)400 static 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  */
match_pkg(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o)420 static 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 
441 static 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 
match_llc(struct cpuinfo_x86 * c,struct cpuinfo_x86 * o)448 static 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 
x86_sched_itmt_flags(void)474 static 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
x86_core_flags(void)480 static int x86_core_flags(void)
481 {
482 	return cpu_core_flags() | x86_sched_itmt_flags();
483 }
484 #endif
485 #ifdef CONFIG_SCHED_SMT
x86_smt_flags(void)486 static int x86_smt_flags(void)
487 {
488 	return cpu_smt_flags();
489 }
490 #endif
491 #ifdef CONFIG_SCHED_CLUSTER
x86_cluster_flags(void)492 static int x86_cluster_flags(void)
493 {
494 	return cpu_cluster_flags() | x86_sched_itmt_flags();
495 }
496 #endif
497 
x86_die_flags(void)498 static int x86_die_flags(void)
499 {
500 	if (cpu_feature_enabled(X86_FEATURE_HYBRID_CPU))
501 	       return x86_sched_itmt_flags();
502 
503 	return 0;
504 }
505 
506 /*
507  * Set if a package/die has multiple NUMA nodes inside.
508  * AMD Magny-Cours, Intel Cluster-on-Die, and Intel
509  * Sub-NUMA Clustering have this.
510  */
511 static bool x86_has_numa_in_package;
512 
513 static struct sched_domain_topology_level x86_topology[6];
514 
build_sched_topology(void)515 static void __init build_sched_topology(void)
516 {
517 	int i = 0;
518 
519 #ifdef CONFIG_SCHED_SMT
520 	x86_topology[i++] = (struct sched_domain_topology_level){
521 		cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT)
522 	};
523 #endif
524 #ifdef CONFIG_SCHED_CLUSTER
525 	x86_topology[i++] = (struct sched_domain_topology_level){
526 		cpu_clustergroup_mask, x86_cluster_flags, SD_INIT_NAME(CLS)
527 	};
528 #endif
529 #ifdef CONFIG_SCHED_MC
530 	x86_topology[i++] = (struct sched_domain_topology_level){
531 		cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC)
532 	};
533 #endif
534 	/*
535 	 * When there is NUMA topology inside the package skip the PKG domain
536 	 * since the NUMA domains will auto-magically create the right spanning
537 	 * domains based on the SLIT.
538 	 */
539 	if (!x86_has_numa_in_package) {
540 		x86_topology[i++] = (struct sched_domain_topology_level){
541 			cpu_cpu_mask, x86_die_flags, SD_INIT_NAME(PKG)
542 		};
543 	}
544 
545 	/*
546 	 * There must be one trailing NULL entry left.
547 	 */
548 	BUG_ON(i >= ARRAY_SIZE(x86_topology)-1);
549 
550 	set_sched_topology(x86_topology);
551 }
552 
set_cpu_sibling_map(int cpu)553 void set_cpu_sibling_map(int cpu)
554 {
555 	bool has_smt = __max_threads_per_core > 1;
556 	bool has_mp = has_smt || topology_num_cores_per_package() > 1;
557 	struct cpuinfo_x86 *c = &cpu_data(cpu);
558 	struct cpuinfo_x86 *o;
559 	int i, threads;
560 
561 	cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
562 
563 	if (!has_mp) {
564 		cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
565 		cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
566 		cpumask_set_cpu(cpu, cpu_l2c_shared_mask(cpu));
567 		cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
568 		cpumask_set_cpu(cpu, topology_die_cpumask(cpu));
569 		c->booted_cores = 1;
570 		return;
571 	}
572 
573 	for_each_cpu(i, cpu_sibling_setup_mask) {
574 		o = &cpu_data(i);
575 
576 		if (match_pkg(c, o) && !topology_same_node(c, o))
577 			x86_has_numa_in_package = true;
578 
579 		if ((i == cpu) || (has_smt && match_smt(c, o)))
580 			link_mask(topology_sibling_cpumask, cpu, i);
581 
582 		if ((i == cpu) || (has_mp && match_llc(c, o)))
583 			link_mask(cpu_llc_shared_mask, cpu, i);
584 
585 		if ((i == cpu) || (has_mp && match_l2c(c, o)))
586 			link_mask(cpu_l2c_shared_mask, cpu, i);
587 
588 		if ((i == cpu) || (has_mp && match_die(c, o)))
589 			link_mask(topology_die_cpumask, cpu, i);
590 	}
591 
592 	threads = cpumask_weight(topology_sibling_cpumask(cpu));
593 	if (threads > __max_smt_threads)
594 		__max_smt_threads = threads;
595 
596 	for_each_cpu(i, topology_sibling_cpumask(cpu))
597 		cpu_data(i).smt_active = threads > 1;
598 
599 	/*
600 	 * This needs a separate iteration over the cpus because we rely on all
601 	 * topology_sibling_cpumask links to be set-up.
602 	 */
603 	for_each_cpu(i, cpu_sibling_setup_mask) {
604 		o = &cpu_data(i);
605 
606 		if ((i == cpu) || (has_mp && match_pkg(c, o))) {
607 			link_mask(topology_core_cpumask, cpu, i);
608 
609 			/*
610 			 *  Does this new cpu bringup a new core?
611 			 */
612 			if (threads == 1) {
613 				/*
614 				 * for each core in package, increment
615 				 * the booted_cores for this new cpu
616 				 */
617 				if (cpumask_first(
618 				    topology_sibling_cpumask(i)) == i)
619 					c->booted_cores++;
620 				/*
621 				 * increment the core count for all
622 				 * the other cpus in this package
623 				 */
624 				if (i != cpu)
625 					cpu_data(i).booted_cores++;
626 			} else if (i != cpu && !c->booted_cores)
627 				c->booted_cores = cpu_data(i).booted_cores;
628 		}
629 	}
630 }
631 
632 /* maps the cpu to the sched domain representing multi-core */
cpu_coregroup_mask(int cpu)633 const struct cpumask *cpu_coregroup_mask(int cpu)
634 {
635 	return cpu_llc_shared_mask(cpu);
636 }
637 
cpu_clustergroup_mask(int cpu)638 const struct cpumask *cpu_clustergroup_mask(int cpu)
639 {
640 	return cpu_l2c_shared_mask(cpu);
641 }
642 EXPORT_SYMBOL_GPL(cpu_clustergroup_mask);
643 
impress_friends(void)644 static void impress_friends(void)
645 {
646 	int cpu;
647 	unsigned long bogosum = 0;
648 	/*
649 	 * Allow the user to impress friends.
650 	 */
651 	pr_debug("Before bogomips\n");
652 	for_each_online_cpu(cpu)
653 		bogosum += cpu_data(cpu).loops_per_jiffy;
654 
655 	pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
656 		num_online_cpus(),
657 		bogosum/(500000/HZ),
658 		(bogosum/(5000/HZ))%100);
659 
660 	pr_debug("Before bogocount - setting activated=1\n");
661 }
662 
663 /*
664  * The Multiprocessor Specification 1.4 (1997) example code suggests
665  * that there should be a 10ms delay between the BSP asserting INIT
666  * and de-asserting INIT, when starting a remote processor.
667  * But that slows boot and resume on modern processors, which include
668  * many cores and don't require that delay.
669  *
670  * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
671  * Modern processor families are quirked to remove the delay entirely.
672  */
673 #define UDELAY_10MS_DEFAULT 10000
674 
675 static unsigned int init_udelay = UINT_MAX;
676 
cpu_init_udelay(char * str)677 static int __init cpu_init_udelay(char *str)
678 {
679 	get_option(&str, &init_udelay);
680 
681 	return 0;
682 }
683 early_param("cpu_init_udelay", cpu_init_udelay);
684 
smp_quirk_init_udelay(void)685 static void __init smp_quirk_init_udelay(void)
686 {
687 	/* if cmdline changed it from default, leave it alone */
688 	if (init_udelay != UINT_MAX)
689 		return;
690 
691 	/* if modern processor, use no delay */
692 	if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
693 	    ((boot_cpu_data.x86_vendor == X86_VENDOR_HYGON) && (boot_cpu_data.x86 >= 0x18)) ||
694 	    ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
695 		init_udelay = 0;
696 		return;
697 	}
698 	/* else, use legacy delay */
699 	init_udelay = UDELAY_10MS_DEFAULT;
700 }
701 
702 /*
703  * Wake up AP by INIT, INIT, STARTUP sequence.
704  */
send_init_sequence(u32 phys_apicid)705 static void send_init_sequence(u32 phys_apicid)
706 {
707 	int maxlvt = lapic_get_maxlvt();
708 
709 	/* Be paranoid about clearing APIC errors. */
710 	if (APIC_INTEGRATED(boot_cpu_apic_version)) {
711 		/* Due to the Pentium erratum 3AP.  */
712 		if (maxlvt > 3)
713 			apic_write(APIC_ESR, 0);
714 		apic_read(APIC_ESR);
715 	}
716 
717 	/* Assert INIT on the target CPU */
718 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT, phys_apicid);
719 	safe_apic_wait_icr_idle();
720 
721 	udelay(init_udelay);
722 
723 	/* Deassert INIT on the target CPU */
724 	apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
725 	safe_apic_wait_icr_idle();
726 }
727 
728 /*
729  * Wake up AP by INIT, INIT, STARTUP sequence.
730  */
wakeup_secondary_cpu_via_init(u32 phys_apicid,unsigned long start_eip)731 static int wakeup_secondary_cpu_via_init(u32 phys_apicid, unsigned long start_eip)
732 {
733 	unsigned long send_status = 0, accept_status = 0;
734 	int num_starts, j, maxlvt;
735 
736 	preempt_disable();
737 	maxlvt = lapic_get_maxlvt();
738 	send_init_sequence(phys_apicid);
739 
740 	mb();
741 
742 	/*
743 	 * Should we send STARTUP IPIs ?
744 	 *
745 	 * Determine this based on the APIC version.
746 	 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
747 	 */
748 	if (APIC_INTEGRATED(boot_cpu_apic_version))
749 		num_starts = 2;
750 	else
751 		num_starts = 0;
752 
753 	/*
754 	 * Run STARTUP IPI loop.
755 	 */
756 	pr_debug("#startup loops: %d\n", num_starts);
757 
758 	for (j = 1; j <= num_starts; j++) {
759 		pr_debug("Sending STARTUP #%d\n", j);
760 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
761 			apic_write(APIC_ESR, 0);
762 		apic_read(APIC_ESR);
763 		pr_debug("After apic_write\n");
764 
765 		/*
766 		 * STARTUP IPI
767 		 */
768 
769 		/* Target chip */
770 		/* Boot on the stack */
771 		/* Kick the second */
772 		apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
773 			       phys_apicid);
774 
775 		/*
776 		 * Give the other CPU some time to accept the IPI.
777 		 */
778 		if (init_udelay == 0)
779 			udelay(10);
780 		else
781 			udelay(300);
782 
783 		pr_debug("Startup point 1\n");
784 
785 		pr_debug("Waiting for send to finish...\n");
786 		send_status = safe_apic_wait_icr_idle();
787 
788 		/*
789 		 * Give the other CPU some time to accept the IPI.
790 		 */
791 		if (init_udelay == 0)
792 			udelay(10);
793 		else
794 			udelay(200);
795 
796 		if (maxlvt > 3)		/* Due to the Pentium erratum 3AP.  */
797 			apic_write(APIC_ESR, 0);
798 		accept_status = (apic_read(APIC_ESR) & 0xEF);
799 		if (send_status || accept_status)
800 			break;
801 	}
802 	pr_debug("After Startup\n");
803 
804 	if (send_status)
805 		pr_err("APIC never delivered???\n");
806 	if (accept_status)
807 		pr_err("APIC delivery error (%lx)\n", accept_status);
808 
809 	preempt_enable();
810 	return (send_status | accept_status);
811 }
812 
813 /* reduce the number of lines printed when booting a large cpu count system */
announce_cpu(int cpu,int apicid)814 static void announce_cpu(int cpu, int apicid)
815 {
816 	static int width, node_width, first = 1;
817 	static int current_node = NUMA_NO_NODE;
818 	int node = early_cpu_to_node(cpu);
819 
820 	if (!width)
821 		width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
822 
823 	if (!node_width)
824 		node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
825 
826 	if (system_state < SYSTEM_RUNNING) {
827 		if (first)
828 			pr_info("x86: Booting SMP configuration:\n");
829 
830 		if (node != current_node) {
831 			if (current_node > (-1))
832 				pr_cont("\n");
833 			current_node = node;
834 
835 			printk(KERN_INFO ".... node %*s#%d, CPUs:  ",
836 			       node_width - num_digits(node), " ", node);
837 		}
838 
839 		/* Add padding for the BSP */
840 		if (first)
841 			pr_cont("%*s", width + 1, " ");
842 		first = 0;
843 
844 		pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
845 	} else
846 		pr_info("Booting Node %d Processor %d APIC 0x%x\n",
847 			node, cpu, apicid);
848 }
849 
common_cpu_up(unsigned int cpu,struct task_struct * idle)850 int common_cpu_up(unsigned int cpu, struct task_struct *idle)
851 {
852 	int ret;
853 
854 	/* Just in case we booted with a single CPU. */
855 	alternatives_enable_smp();
856 
857 	per_cpu(pcpu_hot.current_task, cpu) = idle;
858 	cpu_init_stack_canary(cpu, idle);
859 
860 	/* Initialize the interrupt stack(s) */
861 	ret = irq_init_percpu_irqstack(cpu);
862 	if (ret)
863 		return ret;
864 
865 #ifdef CONFIG_X86_32
866 	/* Stack for startup_32 can be just as for start_secondary onwards */
867 	per_cpu(pcpu_hot.top_of_stack, cpu) = task_top_of_stack(idle);
868 #endif
869 	return 0;
870 }
871 
872 /*
873  * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
874  * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
875  * Returns zero if startup was successfully sent, else error code from
876  * ->wakeup_secondary_cpu.
877  */
do_boot_cpu(u32 apicid,int cpu,struct task_struct * idle)878 static int do_boot_cpu(u32 apicid, int cpu, struct task_struct *idle)
879 {
880 	unsigned long start_ip = real_mode_header->trampoline_start;
881 	int ret;
882 
883 #ifdef CONFIG_X86_64
884 	/* If 64-bit wakeup method exists, use the 64-bit mode trampoline IP */
885 	if (apic->wakeup_secondary_cpu_64)
886 		start_ip = real_mode_header->trampoline_start64;
887 #endif
888 	idle->thread.sp = (unsigned long)task_pt_regs(idle);
889 	initial_code = (unsigned long)start_secondary;
890 
891 	if (IS_ENABLED(CONFIG_X86_32)) {
892 		early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
893 		initial_stack  = idle->thread.sp;
894 	} else if (!(smpboot_control & STARTUP_PARALLEL_MASK)) {
895 		smpboot_control = cpu;
896 	}
897 
898 	/* Enable the espfix hack for this CPU */
899 	init_espfix_ap(cpu);
900 
901 	/* So we see what's up */
902 	announce_cpu(cpu, apicid);
903 
904 	/*
905 	 * This grunge runs the startup process for
906 	 * the targeted processor.
907 	 */
908 	if (x86_platform.legacy.warm_reset) {
909 
910 		pr_debug("Setting warm reset code and vector.\n");
911 
912 		smpboot_setup_warm_reset_vector(start_ip);
913 		/*
914 		 * Be paranoid about clearing APIC errors.
915 		*/
916 		if (APIC_INTEGRATED(boot_cpu_apic_version)) {
917 			apic_write(APIC_ESR, 0);
918 			apic_read(APIC_ESR);
919 		}
920 	}
921 
922 	smp_mb();
923 
924 	/*
925 	 * Wake up a CPU in difference cases:
926 	 * - Use a method from the APIC driver if one defined, with wakeup
927 	 *   straight to 64-bit mode preferred over wakeup to RM.
928 	 * Otherwise,
929 	 * - Use an INIT boot APIC message
930 	 */
931 	if (apic->wakeup_secondary_cpu_64)
932 		ret = apic->wakeup_secondary_cpu_64(apicid, start_ip);
933 	else if (apic->wakeup_secondary_cpu)
934 		ret = apic->wakeup_secondary_cpu(apicid, start_ip);
935 	else
936 		ret = wakeup_secondary_cpu_via_init(apicid, start_ip);
937 
938 	/* If the wakeup mechanism failed, cleanup the warm reset vector */
939 	if (ret)
940 		arch_cpuhp_cleanup_kick_cpu(cpu);
941 	return ret;
942 }
943 
native_kick_ap(unsigned int cpu,struct task_struct * tidle)944 int native_kick_ap(unsigned int cpu, struct task_struct *tidle)
945 {
946 	u32 apicid = apic->cpu_present_to_apicid(cpu);
947 	int err;
948 
949 	lockdep_assert_irqs_enabled();
950 
951 	pr_debug("++++++++++++++++++++=_---CPU UP  %u\n", cpu);
952 
953 	if (apicid == BAD_APICID || !apic_id_valid(apicid)) {
954 		pr_err("CPU %u has invalid APIC ID %x. Aborting bringup\n", cpu, apicid);
955 		return -EINVAL;
956 	}
957 
958 	if (!test_bit(apicid, phys_cpu_present_map)) {
959 		pr_err("CPU %u APIC ID %x is not present. Aborting bringup\n", cpu, apicid);
960 		return -EINVAL;
961 	}
962 
963 	/*
964 	 * Save current MTRR state in case it was changed since early boot
965 	 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
966 	 */
967 	mtrr_save_state();
968 
969 	/* the FPU context is blank, nobody can own it */
970 	per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
971 
972 	err = common_cpu_up(cpu, tidle);
973 	if (err)
974 		return err;
975 
976 	err = do_boot_cpu(apicid, cpu, tidle);
977 	if (err)
978 		pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
979 
980 	return err;
981 }
982 
arch_cpuhp_kick_ap_alive(unsigned int cpu,struct task_struct * tidle)983 int arch_cpuhp_kick_ap_alive(unsigned int cpu, struct task_struct *tidle)
984 {
985 	return smp_ops.kick_ap_alive(cpu, tidle);
986 }
987 
arch_cpuhp_cleanup_kick_cpu(unsigned int cpu)988 void arch_cpuhp_cleanup_kick_cpu(unsigned int cpu)
989 {
990 	/* Cleanup possible dangling ends... */
991 	if (smp_ops.kick_ap_alive == native_kick_ap && x86_platform.legacy.warm_reset)
992 		smpboot_restore_warm_reset_vector();
993 }
994 
arch_cpuhp_cleanup_dead_cpu(unsigned int cpu)995 void arch_cpuhp_cleanup_dead_cpu(unsigned int cpu)
996 {
997 	if (smp_ops.cleanup_dead_cpu)
998 		smp_ops.cleanup_dead_cpu(cpu);
999 
1000 	if (system_state == SYSTEM_RUNNING)
1001 		pr_info("CPU %u is now offline\n", cpu);
1002 }
1003 
arch_cpuhp_sync_state_poll(void)1004 void arch_cpuhp_sync_state_poll(void)
1005 {
1006 	if (smp_ops.poll_sync_state)
1007 		smp_ops.poll_sync_state();
1008 }
1009 
1010 /**
1011  * arch_disable_smp_support() - Disables SMP support for x86 at boottime
1012  */
arch_disable_smp_support(void)1013 void __init arch_disable_smp_support(void)
1014 {
1015 	disable_ioapic_support();
1016 }
1017 
1018 /*
1019  * Fall back to non SMP mode after errors.
1020  *
1021  * RED-PEN audit/test this more. I bet there is more state messed up here.
1022  */
disable_smp(void)1023 static __init void disable_smp(void)
1024 {
1025 	pr_info("SMP disabled\n");
1026 
1027 	disable_ioapic_support();
1028 	topology_reset_possible_cpus_up();
1029 
1030 	cpumask_set_cpu(0, topology_sibling_cpumask(0));
1031 	cpumask_set_cpu(0, topology_core_cpumask(0));
1032 	cpumask_set_cpu(0, topology_die_cpumask(0));
1033 }
1034 
smp_prepare_cpus_common(void)1035 void __init smp_prepare_cpus_common(void)
1036 {
1037 	unsigned int cpu, node;
1038 
1039 	/* Mark all except the boot CPU as hotpluggable */
1040 	for_each_possible_cpu(cpu) {
1041 		if (cpu)
1042 			per_cpu(cpu_info.cpu_index, cpu) = nr_cpu_ids;
1043 	}
1044 
1045 	for_each_possible_cpu(cpu) {
1046 		node = cpu_to_node(cpu);
1047 
1048 		zalloc_cpumask_var_node(&per_cpu(cpu_sibling_map,    cpu), GFP_KERNEL, node);
1049 		zalloc_cpumask_var_node(&per_cpu(cpu_core_map,       cpu), GFP_KERNEL, node);
1050 		zalloc_cpumask_var_node(&per_cpu(cpu_die_map,        cpu), GFP_KERNEL, node);
1051 		zalloc_cpumask_var_node(&per_cpu(cpu_llc_shared_map, cpu), GFP_KERNEL, node);
1052 		zalloc_cpumask_var_node(&per_cpu(cpu_l2c_shared_map, cpu), GFP_KERNEL, node);
1053 	}
1054 
1055 	set_cpu_sibling_map(0);
1056 }
1057 
smp_prepare_boot_cpu(void)1058 void __init smp_prepare_boot_cpu(void)
1059 {
1060 	smp_ops.smp_prepare_boot_cpu();
1061 }
1062 
1063 #ifdef CONFIG_X86_64
1064 /* Establish whether parallel bringup can be supported. */
arch_cpuhp_init_parallel_bringup(void)1065 bool __init arch_cpuhp_init_parallel_bringup(void)
1066 {
1067 	if (!x86_cpuinit.parallel_bringup) {
1068 		pr_info("Parallel CPU startup disabled by the platform\n");
1069 		return false;
1070 	}
1071 
1072 	smpboot_control = STARTUP_READ_APICID;
1073 	pr_debug("Parallel CPU startup enabled: 0x%08x\n", smpboot_control);
1074 	return true;
1075 }
1076 #endif
1077 
1078 /*
1079  * Prepare for SMP bootup.
1080  * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1081  *            for common interface support.
1082  */
native_smp_prepare_cpus(unsigned int max_cpus)1083 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1084 {
1085 	smp_prepare_cpus_common();
1086 
1087 	switch (apic_intr_mode) {
1088 	case APIC_PIC:
1089 	case APIC_VIRTUAL_WIRE_NO_CONFIG:
1090 		disable_smp();
1091 		return;
1092 	case APIC_SYMMETRIC_IO_NO_ROUTING:
1093 		disable_smp();
1094 		/* Setup local timer */
1095 		x86_init.timers.setup_percpu_clockev();
1096 		return;
1097 	case APIC_VIRTUAL_WIRE:
1098 	case APIC_SYMMETRIC_IO:
1099 		break;
1100 	}
1101 
1102 	/* Setup local timer */
1103 	x86_init.timers.setup_percpu_clockev();
1104 
1105 	pr_info("CPU0: ");
1106 	print_cpu_info(&cpu_data(0));
1107 
1108 	uv_system_init();
1109 
1110 	smp_quirk_init_udelay();
1111 
1112 	speculative_store_bypass_ht_init();
1113 
1114 	snp_set_wakeup_secondary_cpu();
1115 }
1116 
arch_thaw_secondary_cpus_begin(void)1117 void arch_thaw_secondary_cpus_begin(void)
1118 {
1119 	set_cache_aps_delayed_init(true);
1120 }
1121 
arch_thaw_secondary_cpus_end(void)1122 void arch_thaw_secondary_cpus_end(void)
1123 {
1124 	cache_aps_init();
1125 }
1126 
1127 /*
1128  * Early setup to make printk work.
1129  */
native_smp_prepare_boot_cpu(void)1130 void __init native_smp_prepare_boot_cpu(void)
1131 {
1132 	int me = smp_processor_id();
1133 
1134 	/* SMP handles this from setup_per_cpu_areas() */
1135 	if (!IS_ENABLED(CONFIG_SMP))
1136 		switch_gdt_and_percpu_base(me);
1137 
1138 	native_pv_lock_init();
1139 }
1140 
native_smp_cpus_done(unsigned int max_cpus)1141 void __init native_smp_cpus_done(unsigned int max_cpus)
1142 {
1143 	pr_debug("Boot done\n");
1144 
1145 	build_sched_topology();
1146 	nmi_selftest();
1147 	impress_friends();
1148 	cache_aps_init();
1149 }
1150 
1151 /* correctly size the local cpu masks */
setup_cpu_local_masks(void)1152 void __init setup_cpu_local_masks(void)
1153 {
1154 	alloc_bootmem_cpumask_var(&cpu_sibling_setup_mask);
1155 }
1156 
1157 #ifdef CONFIG_HOTPLUG_CPU
1158 
1159 /* Recompute SMT state for all CPUs on offline */
recompute_smt_state(void)1160 static void recompute_smt_state(void)
1161 {
1162 	int max_threads, cpu;
1163 
1164 	max_threads = 0;
1165 	for_each_online_cpu (cpu) {
1166 		int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1167 
1168 		if (threads > max_threads)
1169 			max_threads = threads;
1170 	}
1171 	__max_smt_threads = max_threads;
1172 }
1173 
remove_siblinginfo(int cpu)1174 static void remove_siblinginfo(int cpu)
1175 {
1176 	int sibling;
1177 	struct cpuinfo_x86 *c = &cpu_data(cpu);
1178 
1179 	for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1180 		cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1181 		/*/
1182 		 * last thread sibling in this cpu core going down
1183 		 */
1184 		if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1185 			cpu_data(sibling).booted_cores--;
1186 	}
1187 
1188 	for_each_cpu(sibling, topology_die_cpumask(cpu))
1189 		cpumask_clear_cpu(cpu, topology_die_cpumask(sibling));
1190 
1191 	for_each_cpu(sibling, topology_sibling_cpumask(cpu)) {
1192 		cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1193 		if (cpumask_weight(topology_sibling_cpumask(sibling)) == 1)
1194 			cpu_data(sibling).smt_active = false;
1195 	}
1196 
1197 	for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1198 		cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1199 	for_each_cpu(sibling, cpu_l2c_shared_mask(cpu))
1200 		cpumask_clear_cpu(cpu, cpu_l2c_shared_mask(sibling));
1201 	cpumask_clear(cpu_llc_shared_mask(cpu));
1202 	cpumask_clear(cpu_l2c_shared_mask(cpu));
1203 	cpumask_clear(topology_sibling_cpumask(cpu));
1204 	cpumask_clear(topology_core_cpumask(cpu));
1205 	cpumask_clear(topology_die_cpumask(cpu));
1206 	c->topo.core_id = 0;
1207 	c->booted_cores = 0;
1208 	cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1209 	recompute_smt_state();
1210 }
1211 
remove_cpu_from_maps(int cpu)1212 static void remove_cpu_from_maps(int cpu)
1213 {
1214 	set_cpu_online(cpu, false);
1215 	numa_remove_cpu(cpu);
1216 }
1217 
cpu_disable_common(void)1218 void cpu_disable_common(void)
1219 {
1220 	int cpu = smp_processor_id();
1221 
1222 	remove_siblinginfo(cpu);
1223 
1224 	/* It's now safe to remove this processor from the online map */
1225 	lock_vector_lock();
1226 	remove_cpu_from_maps(cpu);
1227 	unlock_vector_lock();
1228 	fixup_irqs();
1229 	lapic_offline();
1230 }
1231 
native_cpu_disable(void)1232 int native_cpu_disable(void)
1233 {
1234 	int ret;
1235 
1236 	ret = lapic_can_unplug_cpu();
1237 	if (ret)
1238 		return ret;
1239 
1240 	cpu_disable_common();
1241 
1242         /*
1243          * Disable the local APIC. Otherwise IPI broadcasts will reach
1244          * it. It still responds normally to INIT, NMI, SMI, and SIPI
1245          * messages.
1246          *
1247          * Disabling the APIC must happen after cpu_disable_common()
1248          * which invokes fixup_irqs().
1249          *
1250          * Disabling the APIC preserves already set bits in IRR, but
1251          * an interrupt arriving after disabling the local APIC does not
1252          * set the corresponding IRR bit.
1253          *
1254          * fixup_irqs() scans IRR for set bits so it can raise a not
1255          * yet handled interrupt on the new destination CPU via an IPI
1256          * but obviously it can't do so for IRR bits which are not set.
1257          * IOW, interrupts arriving after disabling the local APIC will
1258          * be lost.
1259          */
1260 	apic_soft_disable();
1261 
1262 	return 0;
1263 }
1264 
play_dead_common(void)1265 void play_dead_common(void)
1266 {
1267 	idle_task_exit();
1268 
1269 	cpuhp_ap_report_dead();
1270 
1271 	local_irq_disable();
1272 }
1273 
1274 /*
1275  * We need to flush the caches before going to sleep, lest we have
1276  * dirty data in our caches when we come back up.
1277  */
mwait_play_dead(void)1278 static inline void mwait_play_dead(void)
1279 {
1280 	struct mwait_cpu_dead *md = this_cpu_ptr(&mwait_cpu_dead);
1281 	unsigned int eax, ebx, ecx, edx;
1282 	unsigned int highest_cstate = 0;
1283 	unsigned int highest_subcstate = 0;
1284 	int i;
1285 
1286 	if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD ||
1287 	    boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
1288 		return;
1289 	if (!this_cpu_has(X86_FEATURE_MWAIT))
1290 		return;
1291 	if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1292 		return;
1293 	if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1294 		return;
1295 
1296 	eax = CPUID_MWAIT_LEAF;
1297 	ecx = 0;
1298 	native_cpuid(&eax, &ebx, &ecx, &edx);
1299 
1300 	/*
1301 	 * eax will be 0 if EDX enumeration is not valid.
1302 	 * Initialized below to cstate, sub_cstate value when EDX is valid.
1303 	 */
1304 	if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1305 		eax = 0;
1306 	} else {
1307 		edx >>= MWAIT_SUBSTATE_SIZE;
1308 		for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1309 			if (edx & MWAIT_SUBSTATE_MASK) {
1310 				highest_cstate = i;
1311 				highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1312 			}
1313 		}
1314 		eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1315 			(highest_subcstate - 1);
1316 	}
1317 
1318 	/* Set up state for the kexec() hack below */
1319 	md->status = CPUDEAD_MWAIT_WAIT;
1320 	md->control = CPUDEAD_MWAIT_WAIT;
1321 
1322 	wbinvd();
1323 
1324 	while (1) {
1325 		/*
1326 		 * The CLFLUSH is a workaround for erratum AAI65 for
1327 		 * the Xeon 7400 series.  It's not clear it is actually
1328 		 * needed, but it should be harmless in either case.
1329 		 * The WBINVD is insufficient due to the spurious-wakeup
1330 		 * case where we return around the loop.
1331 		 */
1332 		mb();
1333 		clflush(md);
1334 		mb();
1335 		__monitor(md, 0, 0);
1336 		mb();
1337 		__mwait(eax, 0);
1338 
1339 		if (READ_ONCE(md->control) == CPUDEAD_MWAIT_KEXEC_HLT) {
1340 			/*
1341 			 * Kexec is about to happen. Don't go back into mwait() as
1342 			 * the kexec kernel might overwrite text and data including
1343 			 * page tables and stack. So mwait() would resume when the
1344 			 * monitor cache line is written to and then the CPU goes
1345 			 * south due to overwritten text, page tables and stack.
1346 			 *
1347 			 * Note: This does _NOT_ protect against a stray MCE, NMI,
1348 			 * SMI. They will resume execution at the instruction
1349 			 * following the HLT instruction and run into the problem
1350 			 * which this is trying to prevent.
1351 			 */
1352 			WRITE_ONCE(md->status, CPUDEAD_MWAIT_KEXEC_HLT);
1353 			while(1)
1354 				native_halt();
1355 		}
1356 	}
1357 }
1358 
1359 /*
1360  * Kick all "offline" CPUs out of mwait on kexec(). See comment in
1361  * mwait_play_dead().
1362  */
smp_kick_mwait_play_dead(void)1363 void smp_kick_mwait_play_dead(void)
1364 {
1365 	u32 newstate = CPUDEAD_MWAIT_KEXEC_HLT;
1366 	struct mwait_cpu_dead *md;
1367 	unsigned int cpu, i;
1368 
1369 	for_each_cpu_andnot(cpu, cpu_present_mask, cpu_online_mask) {
1370 		md = per_cpu_ptr(&mwait_cpu_dead, cpu);
1371 
1372 		/* Does it sit in mwait_play_dead() ? */
1373 		if (READ_ONCE(md->status) != CPUDEAD_MWAIT_WAIT)
1374 			continue;
1375 
1376 		/* Wait up to 5ms */
1377 		for (i = 0; READ_ONCE(md->status) != newstate && i < 1000; i++) {
1378 			/* Bring it out of mwait */
1379 			WRITE_ONCE(md->control, newstate);
1380 			udelay(5);
1381 		}
1382 
1383 		if (READ_ONCE(md->status) != newstate)
1384 			pr_err_once("CPU%u is stuck in mwait_play_dead()\n", cpu);
1385 	}
1386 }
1387 
hlt_play_dead(void)1388 void __noreturn hlt_play_dead(void)
1389 {
1390 	if (__this_cpu_read(cpu_info.x86) >= 4)
1391 		wbinvd();
1392 
1393 	while (1)
1394 		native_halt();
1395 }
1396 
1397 /*
1398  * native_play_dead() is essentially a __noreturn function, but it can't
1399  * be marked as such as the compiler may complain about it.
1400  */
native_play_dead(void)1401 void native_play_dead(void)
1402 {
1403 	if (cpu_feature_enabled(X86_FEATURE_KERNEL_IBRS))
1404 		__update_spec_ctrl(0);
1405 
1406 	play_dead_common();
1407 	tboot_shutdown(TB_SHUTDOWN_WFS);
1408 
1409 	mwait_play_dead();
1410 	if (cpuidle_play_dead())
1411 		hlt_play_dead();
1412 }
1413 
1414 #else /* ... !CONFIG_HOTPLUG_CPU */
native_cpu_disable(void)1415 int native_cpu_disable(void)
1416 {
1417 	return -ENOSYS;
1418 }
1419 
native_play_dead(void)1420 void native_play_dead(void)
1421 {
1422 	BUG();
1423 }
1424 
1425 #endif
1426