1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3 * CPU/APIC topology
4 *
5 * The APIC IDs describe the system topology in multiple domain levels.
6 * The CPUID topology parser provides the information which part of the
7 * APIC ID is associated to the individual levels:
8 *
9 * [PACKAGE][DIEGRP][DIE][TILE][MODULE][CORE][THREAD]
10 *
11 * The root space contains the package (socket) IDs.
12 *
13 * Not enumerated levels consume 0 bits space, but conceptually they are
14 * always represented. If e.g. only CORE and THREAD levels are enumerated
15 * then the DIE, MODULE and TILE have the same physical ID as the PACKAGE.
16 *
17 * If SMT is not supported, then the THREAD domain is still used. It then
18 * has the same physical ID as the CORE domain and is the only child of
19 * the core domain.
20 *
21 * This allows a unified view on the system independent of the enumerated
22 * domain levels without requiring any conditionals in the code.
23 */
24 #define pr_fmt(fmt) "CPU topo: " fmt
25 #include <linux/cpu.h>
26
27 #include <xen/xen.h>
28
29 #include <asm/apic.h>
30 #include <asm/hypervisor.h>
31 #include <asm/io_apic.h>
32 #include <asm/mpspec.h>
33 #include <asm/smp.h>
34
35 #include "cpu.h"
36
37 /*
38 * Map cpu index to physical APIC ID
39 */
40 DEFINE_EARLY_PER_CPU_READ_MOSTLY(u32, x86_cpu_to_apicid, BAD_APICID);
41 DEFINE_EARLY_PER_CPU_READ_MOSTLY(u32, x86_cpu_to_acpiid, CPU_ACPIID_INVALID);
42 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_apicid);
43 EXPORT_EARLY_PER_CPU_SYMBOL(x86_cpu_to_acpiid);
44
45 /* Bitmap of physically present CPUs. */
46 DECLARE_BITMAP(phys_cpu_present_map, MAX_LOCAL_APIC) __read_mostly;
47
48 /* Used for CPU number allocation and parallel CPU bringup */
49 u32 cpuid_to_apicid[] __ro_after_init = { [0 ... NR_CPUS - 1] = BAD_APICID, };
50
51 /* Bitmaps to mark registered APICs at each topology domain */
52 static struct { DECLARE_BITMAP(map, MAX_LOCAL_APIC); } apic_maps[TOPO_MAX_DOMAIN] __ro_after_init;
53
54 /*
55 * Keep track of assigned, disabled and rejected CPUs. Present assigned
56 * with 1 as CPU #0 is reserved for the boot CPU.
57 */
58 static struct {
59 unsigned int nr_assigned_cpus;
60 unsigned int nr_disabled_cpus;
61 unsigned int nr_rejected_cpus;
62 u32 boot_cpu_apic_id;
63 u32 real_bsp_apic_id;
64 } topo_info __ro_after_init = {
65 .nr_assigned_cpus = 1,
66 .boot_cpu_apic_id = BAD_APICID,
67 .real_bsp_apic_id = BAD_APICID,
68 };
69
70 #define domain_weight(_dom) bitmap_weight(apic_maps[_dom].map, MAX_LOCAL_APIC)
71
arch_match_cpu_phys_id(int cpu,u64 phys_id)72 bool arch_match_cpu_phys_id(int cpu, u64 phys_id)
73 {
74 return phys_id == (u64)cpuid_to_apicid[cpu];
75 }
76
77 #ifdef CONFIG_SMP
cpu_mark_primary_thread(unsigned int cpu,unsigned int apicid)78 static void cpu_mark_primary_thread(unsigned int cpu, unsigned int apicid)
79 {
80 if (!(apicid & (__max_threads_per_core - 1)))
81 cpumask_set_cpu(cpu, &__cpu_primary_thread_mask);
82 }
83 #else
cpu_mark_primary_thread(unsigned int cpu,unsigned int apicid)84 static inline void cpu_mark_primary_thread(unsigned int cpu, unsigned int apicid) { }
85 #endif
86
87 /*
88 * Convert the APIC ID to a domain level ID by masking out the low bits
89 * below the domain level @dom.
90 */
topo_apicid(u32 apicid,enum x86_topology_domains dom)91 static inline u32 topo_apicid(u32 apicid, enum x86_topology_domains dom)
92 {
93 if (dom == TOPO_SMT_DOMAIN)
94 return apicid;
95 return apicid & (UINT_MAX << x86_topo_system.dom_shifts[dom - 1]);
96 }
97
topo_lookup_cpuid(u32 apic_id)98 static int topo_lookup_cpuid(u32 apic_id)
99 {
100 int i;
101
102 /* CPU# to APICID mapping is persistent once it is established */
103 for (i = 0; i < topo_info.nr_assigned_cpus; i++) {
104 if (cpuid_to_apicid[i] == apic_id)
105 return i;
106 }
107 return -ENODEV;
108 }
109
topo_get_cpunr(u32 apic_id)110 static __init int topo_get_cpunr(u32 apic_id)
111 {
112 int cpu = topo_lookup_cpuid(apic_id);
113
114 if (cpu >= 0)
115 return cpu;
116
117 return topo_info.nr_assigned_cpus++;
118 }
119
topo_set_cpuids(unsigned int cpu,u32 apic_id,u32 acpi_id)120 static void topo_set_cpuids(unsigned int cpu, u32 apic_id, u32 acpi_id)
121 {
122 #if defined(CONFIG_SMP) || defined(CONFIG_X86_64)
123 early_per_cpu(x86_cpu_to_apicid, cpu) = apic_id;
124 early_per_cpu(x86_cpu_to_acpiid, cpu) = acpi_id;
125 #endif
126 set_cpu_present(cpu, true);
127 }
128
check_for_real_bsp(u32 apic_id)129 static __init bool check_for_real_bsp(u32 apic_id)
130 {
131 bool is_bsp = false, has_apic_base = boot_cpu_data.x86 >= 6;
132 u64 msr;
133
134 /*
135 * There is no real good way to detect whether this a kdump()
136 * kernel, but except on the Voyager SMP monstrosity which is not
137 * longer supported, the real BSP APIC ID is the first one which is
138 * enumerated by firmware. That allows to detect whether the boot
139 * CPU is the real BSP. If it is not, then do not register the APIC
140 * because sending INIT to the real BSP would reset the whole
141 * system.
142 *
143 * The first APIC ID which is enumerated by firmware is detectable
144 * because the boot CPU APIC ID is registered before that without
145 * invoking this code.
146 */
147 if (topo_info.real_bsp_apic_id != BAD_APICID)
148 return false;
149
150 /*
151 * Check whether the enumeration order is broken by evaluating the
152 * BSP bit in the APICBASE MSR. If the CPU does not have the
153 * APICBASE MSR then the BSP detection is not possible and the
154 * kernel must rely on the firmware enumeration order.
155 */
156 if (has_apic_base) {
157 rdmsrl(MSR_IA32_APICBASE, msr);
158 is_bsp = !!(msr & MSR_IA32_APICBASE_BSP);
159 }
160
161 if (apic_id == topo_info.boot_cpu_apic_id) {
162 /*
163 * If the boot CPU has the APIC BSP bit set then the
164 * firmware enumeration is agreeing. If the CPU does not
165 * have the APICBASE MSR then the only choice is to trust
166 * the enumeration order.
167 */
168 if (is_bsp || !has_apic_base) {
169 topo_info.real_bsp_apic_id = apic_id;
170 return false;
171 }
172 /*
173 * If the boot APIC is enumerated first, but the APICBASE
174 * MSR does not have the BSP bit set, then there is no way
175 * to discover the real BSP here. Assume a crash kernel and
176 * limit the number of CPUs to 1 as an INIT to the real BSP
177 * would reset the machine.
178 */
179 pr_warn("Enumerated BSP APIC %x is not marked in APICBASE MSR\n", apic_id);
180 pr_warn("Assuming crash kernel. Limiting to one CPU to prevent machine INIT\n");
181 set_nr_cpu_ids(1);
182 goto fwbug;
183 }
184
185 pr_warn("Boot CPU APIC ID not the first enumerated APIC ID: %x != %x\n",
186 topo_info.boot_cpu_apic_id, apic_id);
187
188 if (is_bsp) {
189 /*
190 * The boot CPU has the APIC BSP bit set. Use it and complain
191 * about the broken firmware enumeration.
192 */
193 topo_info.real_bsp_apic_id = topo_info.boot_cpu_apic_id;
194 goto fwbug;
195 }
196
197 pr_warn("Crash kernel detected. Disabling real BSP to prevent machine INIT\n");
198
199 topo_info.real_bsp_apic_id = apic_id;
200 return true;
201
202 fwbug:
203 pr_warn(FW_BUG "APIC enumeration order not specification compliant\n");
204 return false;
205 }
206
topo_unit_count(u32 lvlid,enum x86_topology_domains at_level,unsigned long * map)207 static unsigned int topo_unit_count(u32 lvlid, enum x86_topology_domains at_level,
208 unsigned long *map)
209 {
210 unsigned int id, end, cnt = 0;
211
212 /* Calculate the exclusive end */
213 end = lvlid + (1U << x86_topo_system.dom_shifts[at_level]);
214
215 /* Unfortunately there is no bitmap_weight_range() */
216 for (id = find_next_bit(map, end, lvlid); id < end; id = find_next_bit(map, end, ++id))
217 cnt++;
218 return cnt;
219 }
220
topo_register_apic(u32 apic_id,u32 acpi_id,bool present)221 static __init void topo_register_apic(u32 apic_id, u32 acpi_id, bool present)
222 {
223 int cpu, dom;
224
225 if (present) {
226 set_bit(apic_id, phys_cpu_present_map);
227
228 /*
229 * Double registration is valid in case of the boot CPU
230 * APIC because that is registered before the enumeration
231 * of the APICs via firmware parsers or VM guest
232 * mechanisms.
233 */
234 if (apic_id == topo_info.boot_cpu_apic_id)
235 cpu = 0;
236 else
237 cpu = topo_get_cpunr(apic_id);
238
239 cpuid_to_apicid[cpu] = apic_id;
240 topo_set_cpuids(cpu, apic_id, acpi_id);
241 } else {
242 u32 pkgid = topo_apicid(apic_id, TOPO_PKG_DOMAIN);
243
244 /*
245 * Check for present APICs in the same package when running
246 * on bare metal. Allow the bogosity in a guest.
247 */
248 if (hypervisor_is_type(X86_HYPER_NATIVE) &&
249 topo_unit_count(pkgid, TOPO_PKG_DOMAIN, phys_cpu_present_map)) {
250 pr_info_once("Ignoring hot-pluggable APIC ID %x in present package.\n",
251 apic_id);
252 topo_info.nr_rejected_cpus++;
253 return;
254 }
255
256 topo_info.nr_disabled_cpus++;
257 }
258
259 /*
260 * Register present and possible CPUs in the domain
261 * maps. cpu_possible_map will be updated in
262 * topology_init_possible_cpus() after enumeration is done.
263 */
264 for (dom = TOPO_SMT_DOMAIN; dom < TOPO_MAX_DOMAIN; dom++)
265 set_bit(topo_apicid(apic_id, dom), apic_maps[dom].map);
266 }
267
268 /**
269 * topology_register_apic - Register an APIC in early topology maps
270 * @apic_id: The APIC ID to set up
271 * @acpi_id: The ACPI ID associated to the APIC
272 * @present: True if the corresponding CPU is present
273 */
topology_register_apic(u32 apic_id,u32 acpi_id,bool present)274 void __init topology_register_apic(u32 apic_id, u32 acpi_id, bool present)
275 {
276 if (apic_id >= MAX_LOCAL_APIC) {
277 pr_err_once("APIC ID %x exceeds kernel limit of: %x\n", apic_id, MAX_LOCAL_APIC - 1);
278 topo_info.nr_rejected_cpus++;
279 return;
280 }
281
282 if (check_for_real_bsp(apic_id)) {
283 topo_info.nr_rejected_cpus++;
284 return;
285 }
286
287 /* CPU numbers exhausted? */
288 if (apic_id != topo_info.boot_cpu_apic_id && topo_info.nr_assigned_cpus >= nr_cpu_ids) {
289 pr_warn_once("CPU limit of %d reached. Ignoring further CPUs\n", nr_cpu_ids);
290 topo_info.nr_rejected_cpus++;
291 return;
292 }
293
294 topo_register_apic(apic_id, acpi_id, present);
295 }
296
297 /**
298 * topology_register_boot_apic - Register the boot CPU APIC
299 * @apic_id: The APIC ID to set up
300 *
301 * Separate so CPU #0 can be assigned
302 */
topology_register_boot_apic(u32 apic_id)303 void __init topology_register_boot_apic(u32 apic_id)
304 {
305 WARN_ON_ONCE(topo_info.boot_cpu_apic_id != BAD_APICID);
306
307 topo_info.boot_cpu_apic_id = apic_id;
308 topo_register_apic(apic_id, CPU_ACPIID_INVALID, true);
309 }
310
311 /**
312 * topology_get_logical_id - Retrieve the logical ID at a given topology domain level
313 * @apicid: The APIC ID for which to lookup the logical ID
314 * @at_level: The topology domain level to use
315 *
316 * @apicid must be a full APIC ID, not the normalized variant. It's valid to have
317 * all bits below the domain level specified by @at_level to be clear. So both
318 * real APIC IDs and backshifted normalized APIC IDs work correctly.
319 *
320 * Returns:
321 * - >= 0: The requested logical ID
322 * - -ERANGE: @apicid is out of range
323 * - -ENODEV: @apicid is not registered
324 */
topology_get_logical_id(u32 apicid,enum x86_topology_domains at_level)325 int topology_get_logical_id(u32 apicid, enum x86_topology_domains at_level)
326 {
327 /* Remove the bits below @at_level to get the proper level ID of @apicid */
328 unsigned int lvlid = topo_apicid(apicid, at_level);
329
330 if (lvlid >= MAX_LOCAL_APIC)
331 return -ERANGE;
332 if (!test_bit(lvlid, apic_maps[at_level].map))
333 return -ENODEV;
334 /* Get the number of set bits before @lvlid. */
335 return bitmap_weight(apic_maps[at_level].map, lvlid);
336 }
337 EXPORT_SYMBOL_GPL(topology_get_logical_id);
338
339 /**
340 * topology_unit_count - Retrieve the count of specified units at a given topology domain level
341 * @apicid: The APIC ID which specifies the search range
342 * @which_units: The domain level specifying the units to count
343 * @at_level: The domain level at which @which_units have to be counted
344 *
345 * This returns the number of possible units according to the enumerated
346 * information.
347 *
348 * E.g. topology_count_units(apicid, TOPO_CORE_DOMAIN, TOPO_PKG_DOMAIN)
349 * counts the number of possible cores in the package to which @apicid
350 * belongs.
351 *
352 * @at_level must obviously be greater than @which_level to produce useful
353 * results. If @at_level is equal to @which_units the result is
354 * unsurprisingly 1. If @at_level is less than @which_units the results
355 * is by definition undefined and the function returns 0.
356 */
topology_unit_count(u32 apicid,enum x86_topology_domains which_units,enum x86_topology_domains at_level)357 unsigned int topology_unit_count(u32 apicid, enum x86_topology_domains which_units,
358 enum x86_topology_domains at_level)
359 {
360 /* Remove the bits below @at_level to get the proper level ID of @apicid */
361 unsigned int lvlid = topo_apicid(apicid, at_level);
362
363 if (lvlid >= MAX_LOCAL_APIC)
364 return 0;
365 if (!test_bit(lvlid, apic_maps[at_level].map))
366 return 0;
367 if (which_units > at_level)
368 return 0;
369 if (which_units == at_level)
370 return 1;
371 return topo_unit_count(lvlid, at_level, apic_maps[which_units].map);
372 }
373
374 #ifdef CONFIG_ACPI_HOTPLUG_CPU
375 /**
376 * topology_hotplug_apic - Handle a physical hotplugged APIC after boot
377 * @apic_id: The APIC ID to set up
378 * @acpi_id: The ACPI ID associated to the APIC
379 */
topology_hotplug_apic(u32 apic_id,u32 acpi_id)380 int topology_hotplug_apic(u32 apic_id, u32 acpi_id)
381 {
382 int cpu;
383
384 if (apic_id >= MAX_LOCAL_APIC)
385 return -EINVAL;
386
387 /* Reject if the APIC ID was not registered during enumeration. */
388 if (!test_bit(apic_id, apic_maps[TOPO_SMT_DOMAIN].map))
389 return -ENODEV;
390
391 cpu = topo_lookup_cpuid(apic_id);
392 if (cpu < 0)
393 return -ENOSPC;
394
395 set_bit(apic_id, phys_cpu_present_map);
396 topo_set_cpuids(cpu, apic_id, acpi_id);
397 cpu_mark_primary_thread(cpu, apic_id);
398 return cpu;
399 }
400
401 /**
402 * topology_hotunplug_apic - Remove a physical hotplugged APIC after boot
403 * @cpu: The CPU number for which the APIC ID is removed
404 */
topology_hotunplug_apic(unsigned int cpu)405 void topology_hotunplug_apic(unsigned int cpu)
406 {
407 u32 apic_id = cpuid_to_apicid[cpu];
408
409 if (apic_id == BAD_APICID)
410 return;
411
412 per_cpu(x86_cpu_to_apicid, cpu) = BAD_APICID;
413 clear_bit(apic_id, phys_cpu_present_map);
414 set_cpu_present(cpu, false);
415 }
416 #endif
417
418 #ifdef CONFIG_X86_LOCAL_APIC
419 static unsigned int max_possible_cpus __initdata = NR_CPUS;
420
421 /**
422 * topology_apply_cmdline_limits_early - Apply topology command line limits early
423 *
424 * Ensure that command line limits are in effect before firmware parsing
425 * takes place.
426 */
topology_apply_cmdline_limits_early(void)427 void __init topology_apply_cmdline_limits_early(void)
428 {
429 unsigned int possible = nr_cpu_ids;
430
431 /* 'maxcpus=0' 'nosmp' 'nolapic' 'disableapic' 'noapic' */
432 if (!setup_max_cpus || ioapic_is_disabled || apic_is_disabled)
433 possible = 1;
434
435 /* 'possible_cpus=N' */
436 possible = min_t(unsigned int, max_possible_cpus, possible);
437
438 if (possible < nr_cpu_ids) {
439 pr_info("Limiting to %u possible CPUs\n", possible);
440 set_nr_cpu_ids(possible);
441 }
442 }
443
restrict_to_up(void)444 static __init bool restrict_to_up(void)
445 {
446 if (!smp_found_config || ioapic_is_disabled)
447 return true;
448 /*
449 * XEN PV is special as it does not advertise the local APIC
450 * properly, but provides a fake topology for it so that the
451 * infrastructure works. So don't apply the restrictions vs. APIC
452 * here.
453 */
454 if (xen_pv_domain())
455 return false;
456
457 return apic_is_disabled;
458 }
459
topology_init_possible_cpus(void)460 void __init topology_init_possible_cpus(void)
461 {
462 unsigned int assigned = topo_info.nr_assigned_cpus;
463 unsigned int disabled = topo_info.nr_disabled_cpus;
464 unsigned int cnta, cntb, cpu, allowed = 1;
465 unsigned int total = assigned + disabled;
466 u32 apicid, firstid;
467
468 /*
469 * If there was no APIC registered, then fake one so that the
470 * topology bitmap is populated. That ensures that the code below
471 * is valid and the various query interfaces can be used
472 * unconditionally. This does not affect the actual APIC code in
473 * any way because either the local APIC address has not been
474 * registered or the local APIC was disabled on the command line.
475 */
476 if (topo_info.boot_cpu_apic_id == BAD_APICID)
477 topology_register_boot_apic(0);
478
479 if (!restrict_to_up()) {
480 if (WARN_ON_ONCE(assigned > nr_cpu_ids)) {
481 disabled += assigned - nr_cpu_ids;
482 assigned = nr_cpu_ids;
483 }
484 allowed = min_t(unsigned int, total, nr_cpu_ids);
485 }
486
487 if (total > allowed)
488 pr_warn("%u possible CPUs exceed the limit of %u\n", total, allowed);
489
490 assigned = min_t(unsigned int, allowed, assigned);
491 disabled = allowed - assigned;
492
493 topo_info.nr_assigned_cpus = assigned;
494 topo_info.nr_disabled_cpus = disabled;
495
496 total_cpus = allowed;
497 set_nr_cpu_ids(allowed);
498
499 cnta = domain_weight(TOPO_PKG_DOMAIN);
500 cntb = domain_weight(TOPO_DIE_DOMAIN);
501 __max_logical_packages = cnta;
502 __max_dies_per_package = 1U << (get_count_order(cntb) - get_count_order(cnta));
503
504 pr_info("Max. logical packages: %3u\n", cnta);
505 pr_info("Max. logical dies: %3u\n", cntb);
506 pr_info("Max. dies per package: %3u\n", __max_dies_per_package);
507
508 cnta = domain_weight(TOPO_CORE_DOMAIN);
509 cntb = domain_weight(TOPO_SMT_DOMAIN);
510 /*
511 * Can't use order delta here as order(cnta) can be equal
512 * order(cntb) even if cnta != cntb.
513 */
514 __max_threads_per_core = DIV_ROUND_UP(cntb, cnta);
515 pr_info("Max. threads per core: %3u\n", __max_threads_per_core);
516
517 firstid = find_first_bit(apic_maps[TOPO_SMT_DOMAIN].map, MAX_LOCAL_APIC);
518 __num_cores_per_package = topology_unit_count(firstid, TOPO_CORE_DOMAIN, TOPO_PKG_DOMAIN);
519 pr_info("Num. cores per package: %3u\n", __num_cores_per_package);
520 __num_threads_per_package = topology_unit_count(firstid, TOPO_SMT_DOMAIN, TOPO_PKG_DOMAIN);
521 pr_info("Num. threads per package: %3u\n", __num_threads_per_package);
522
523 pr_info("Allowing %u present CPUs plus %u hotplug CPUs\n", assigned, disabled);
524 if (topo_info.nr_rejected_cpus)
525 pr_info("Rejected CPUs %u\n", topo_info.nr_rejected_cpus);
526
527 init_cpu_present(cpumask_of(0));
528 init_cpu_possible(cpumask_of(0));
529
530 /* Assign CPU numbers to non-present CPUs */
531 for (apicid = 0; disabled; disabled--, apicid++) {
532 apicid = find_next_andnot_bit(apic_maps[TOPO_SMT_DOMAIN].map, phys_cpu_present_map,
533 MAX_LOCAL_APIC, apicid);
534 if (apicid >= MAX_LOCAL_APIC)
535 break;
536 cpuid_to_apicid[topo_info.nr_assigned_cpus++] = apicid;
537 }
538
539 for (cpu = 0; cpu < allowed; cpu++) {
540 apicid = cpuid_to_apicid[cpu];
541
542 set_cpu_possible(cpu, true);
543
544 if (apicid == BAD_APICID)
545 continue;
546
547 cpu_mark_primary_thread(cpu, apicid);
548 set_cpu_present(cpu, test_bit(apicid, phys_cpu_present_map));
549 }
550 }
551
552 /*
553 * Late SMP disable after sizing CPU masks when APIC/IOAPIC setup failed.
554 */
topology_reset_possible_cpus_up(void)555 void __init topology_reset_possible_cpus_up(void)
556 {
557 init_cpu_present(cpumask_of(0));
558 init_cpu_possible(cpumask_of(0));
559
560 bitmap_zero(phys_cpu_present_map, MAX_LOCAL_APIC);
561 if (topo_info.boot_cpu_apic_id != BAD_APICID)
562 set_bit(topo_info.boot_cpu_apic_id, phys_cpu_present_map);
563 }
564
setup_possible_cpus(char * str)565 static int __init setup_possible_cpus(char *str)
566 {
567 get_option(&str, &max_possible_cpus);
568 return 0;
569 }
570 early_param("possible_cpus", setup_possible_cpus);
571 #endif
572