1 // SPDX-License-Identifier: GPL-2.0-or-later
3 * x86 SMP booting functions
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.
9 * Much of the core SMP work is based on previous work by Thomas Radke, to
10 * whom a great many thanks are extended.
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.
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
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
40 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
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/stackprotector.h>
55 #include <linux/gfp.h>
56 #include <linux/cpuidle.h>
57 #include <linux/numa.h>
63 #include <asm/realmode.h>
66 #include <asm/pgtable.h>
67 #include <asm/tlbflush.h>
69 #include <asm/mwait.h>
71 #include <asm/io_apic.h>
72 #include <asm/fpu/internal.h>
73 #include <asm/setup.h>
74 #include <asm/uv/uv.h>
75 #include <linux/mc146818rtc.h>
76 #include <asm/i8259.h>
78 #include <asm/qspinlock.h>
79 #include <asm/intel-family.h>
80 #include <asm/cpu_device_id.h>
81 #include <asm/spec-ctrl.h>
82 #include <asm/hw_irq.h>
84 /* representing HT siblings of each logical CPU */
85 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
86 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
88 /* representing HT and core siblings of each logical CPU */
89 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
90 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
92 /* representing HT, core, and die siblings of each logical CPU */
93 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_die_map);
94 EXPORT_PER_CPU_SYMBOL(cpu_die_map);
96 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_llc_shared_map);
98 /* Per CPU bogomips and other parameters */
99 DEFINE_PER_CPU_READ_MOSTLY(struct cpuinfo_x86, cpu_info);
100 EXPORT_PER_CPU_SYMBOL(cpu_info);
102 /* Logical package management. We might want to allocate that dynamically */
103 unsigned int __max_logical_packages __read_mostly;
104 EXPORT_SYMBOL(__max_logical_packages);
105 static unsigned int logical_packages __read_mostly;
106 static unsigned int logical_die __read_mostly;
108 /* Maximum number of SMT threads on any online core */
109 int __read_mostly __max_smt_threads = 1;
111 /* Flag to indicate if a complete sched domain rebuild is required */
112 bool x86_topology_update;
114 int arch_update_cpu_topology(void)
116 int retval = x86_topology_update;
118 x86_topology_update = false;
122 static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
126 spin_lock_irqsave(&rtc_lock, flags);
127 CMOS_WRITE(0xa, 0xf);
128 spin_unlock_irqrestore(&rtc_lock, flags);
129 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) =
131 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) =
135 static inline void smpboot_restore_warm_reset_vector(void)
140 * Paranoid: Set warm reset code and vector here back
143 spin_lock_irqsave(&rtc_lock, flags);
145 spin_unlock_irqrestore(&rtc_lock, flags);
147 *((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
151 * Report back to the Boot Processor during boot time or to the caller processor
154 static void smp_callin(void)
159 * If waken up by an INIT in an 82489DX configuration
160 * cpu_callout_mask guarantees we don't get here before
161 * an INIT_deassert IPI reaches our local APIC, so it is
162 * now safe to touch our local APIC.
164 cpuid = smp_processor_id();
167 * the boot CPU has finished the init stage and is spinning
168 * on callin_map until we finish. We are free to set up this
169 * CPU, first the APIC. (this is probably redundant on most
175 * Save our processor parameters. Note: this information
176 * is needed for clock calibration.
178 smp_store_cpu_info(cpuid);
181 * The topology information must be up to date before
182 * calibrate_delay() and notify_cpu_starting().
184 set_cpu_sibling_map(raw_smp_processor_id());
188 * Update loops_per_jiffy in cpu_data. Previous call to
189 * smp_store_cpu_info() stored a value that is close but not as
190 * accurate as the value just calculated.
193 cpu_data(cpuid).loops_per_jiffy = loops_per_jiffy;
194 pr_debug("Stack at about %p\n", &cpuid);
198 notify_cpu_starting(cpuid);
201 * Allow the master to continue.
203 cpumask_set_cpu(cpuid, cpu_callin_mask);
206 static int cpu0_logical_apicid;
207 static int enable_start_cpu0;
209 * Activate a secondary processor.
211 static void notrace start_secondary(void *unused)
213 unsigned long cr4 = __read_cr4();
216 * Don't put *anything* except direct CPU state initialization
217 * before cpu_init(), SMP booting is too fragile that we want to
218 * limit the things done here to the most necessary things.
220 if (boot_cpu_has(X86_FEATURE_PCID))
221 cr4 |= X86_CR4_PCIDE;
222 if (static_branch_likely(&cr_pinning))
223 cr4 |= cr4_pinned_bits;
228 /* switch away from the initial page table */
229 load_cr3(swapper_pg_dir);
231 * Initialize the CR4 shadow before doing anything that could
239 x86_cpuinit.early_percpu_clock_init();
243 enable_start_cpu0 = 0;
245 /* otherwise gcc will move up smp_processor_id before the cpu_init */
248 * Check TSC synchronization with the boot CPU:
250 check_tsc_sync_target();
252 speculative_store_bypass_ht_init();
255 * Lock vector_lock, set CPU online and bring the vector
256 * allocator online. Online must be set with vector_lock held
257 * to prevent a concurrent irq setup/teardown from seeing a
258 * half valid vector space.
261 set_cpu_online(smp_processor_id(), true);
263 unlock_vector_lock();
264 cpu_set_state_online(smp_processor_id());
265 x86_platform.nmi_init();
267 /* enable local interrupts */
270 /* to prevent fake stack check failure in clock setup */
271 boot_init_stack_canary();
273 x86_cpuinit.setup_percpu_clockev();
276 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
280 * topology_is_primary_thread - Check whether CPU is the primary SMT thread
283 bool topology_is_primary_thread(unsigned int cpu)
285 return apic_id_is_primary_thread(per_cpu(x86_cpu_to_apicid, cpu));
289 * topology_smt_supported - Check whether SMT is supported by the CPUs
291 bool topology_smt_supported(void)
293 return smp_num_siblings > 1;
297 * topology_phys_to_logical_pkg - Map a physical package id to a logical
299 * Returns logical package id or -1 if not found
301 int topology_phys_to_logical_pkg(unsigned int phys_pkg)
305 for_each_possible_cpu(cpu) {
306 struct cpuinfo_x86 *c = &cpu_data(cpu);
308 if (c->initialized && c->phys_proc_id == phys_pkg)
309 return c->logical_proc_id;
313 EXPORT_SYMBOL(topology_phys_to_logical_pkg);
315 * topology_phys_to_logical_die - Map a physical die id to logical
317 * Returns logical die id or -1 if not found
319 int topology_phys_to_logical_die(unsigned int die_id, unsigned int cur_cpu)
322 int proc_id = cpu_data(cur_cpu).phys_proc_id;
324 for_each_possible_cpu(cpu) {
325 struct cpuinfo_x86 *c = &cpu_data(cpu);
327 if (c->initialized && c->cpu_die_id == die_id &&
328 c->phys_proc_id == proc_id)
329 return c->logical_die_id;
333 EXPORT_SYMBOL(topology_phys_to_logical_die);
336 * topology_update_package_map - Update the physical to logical package map
337 * @pkg: The physical package id as retrieved via CPUID
338 * @cpu: The cpu for which this is updated
340 int topology_update_package_map(unsigned int pkg, unsigned int cpu)
344 /* Already available somewhere? */
345 new = topology_phys_to_logical_pkg(pkg);
349 new = logical_packages++;
351 pr_info("CPU %u Converting physical %u to logical package %u\n",
355 cpu_data(cpu).logical_proc_id = new;
359 * topology_update_die_map - Update the physical to logical die map
360 * @die: The die id as retrieved via CPUID
361 * @cpu: The cpu for which this is updated
363 int topology_update_die_map(unsigned int die, unsigned int cpu)
367 /* Already available somewhere? */
368 new = topology_phys_to_logical_die(die, cpu);
374 pr_info("CPU %u Converting physical %u to logical die %u\n",
378 cpu_data(cpu).logical_die_id = new;
382 void __init smp_store_boot_cpu_info(void)
384 int id = 0; /* CPU 0 */
385 struct cpuinfo_x86 *c = &cpu_data(id);
389 topology_update_package_map(c->phys_proc_id, id);
390 topology_update_die_map(c->cpu_die_id, id);
391 c->initialized = true;
395 * The bootstrap kernel entry code has set these up. Save them for
398 void smp_store_cpu_info(int id)
400 struct cpuinfo_x86 *c = &cpu_data(id);
402 /* Copy boot_cpu_data only on the first bringup */
407 * During boot time, CPU0 has this setup already. Save the info when
408 * bringing up AP or offlined CPU0.
410 identify_secondary_cpu(c);
411 c->initialized = true;
415 topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
417 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
419 return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
423 topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
425 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
427 return !WARN_ONCE(!topology_same_node(c, o),
428 "sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
429 "[node: %d != %d]. Ignoring dependency.\n",
430 cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
433 #define link_mask(mfunc, c1, c2) \
435 cpumask_set_cpu((c1), mfunc(c2)); \
436 cpumask_set_cpu((c2), mfunc(c1)); \
439 static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
441 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
442 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
444 if (c->phys_proc_id == o->phys_proc_id &&
445 c->cpu_die_id == o->cpu_die_id &&
446 per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2)) {
447 if (c->cpu_core_id == o->cpu_core_id)
448 return topology_sane(c, o, "smt");
450 if ((c->cu_id != 0xff) &&
451 (o->cu_id != 0xff) &&
452 (c->cu_id == o->cu_id))
453 return topology_sane(c, o, "smt");
456 } else if (c->phys_proc_id == o->phys_proc_id &&
457 c->cpu_die_id == o->cpu_die_id &&
458 c->cpu_core_id == o->cpu_core_id) {
459 return topology_sane(c, o, "smt");
466 * Define snc_cpu[] for SNC (Sub-NUMA Cluster) CPUs.
468 * These are Intel CPUs that enumerate an LLC that is shared by
469 * multiple NUMA nodes. The LLC on these systems is shared for
470 * off-package data access but private to the NUMA node (half
471 * of the package) for on-package access.
473 * CPUID (the source of the information about the LLC) can only
474 * enumerate the cache as being shared *or* unshared, but not
475 * this particular configuration. The CPU in this case enumerates
476 * the cache to be shared across the entire package (spanning both
480 static const struct x86_cpu_id snc_cpu[] = {
481 { X86_VENDOR_INTEL, 6, INTEL_FAM6_SKYLAKE_X },
485 static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
487 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
489 /* Do not match if we do not have a valid APICID for cpu: */
490 if (per_cpu(cpu_llc_id, cpu1) == BAD_APICID)
493 /* Do not match if LLC id does not match: */
494 if (per_cpu(cpu_llc_id, cpu1) != per_cpu(cpu_llc_id, cpu2))
498 * Allow the SNC topology without warning. Return of false
499 * means 'c' does not share the LLC of 'o'. This will be
500 * reflected to userspace.
502 if (!topology_same_node(c, o) && x86_match_cpu(snc_cpu))
505 return topology_sane(c, o, "llc");
509 * Unlike the other levels, we do not enforce keeping a
510 * multicore group inside a NUMA node. If this happens, we will
511 * discard the MC level of the topology later.
513 static bool match_pkg(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
515 if (c->phys_proc_id == o->phys_proc_id)
520 static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
522 if ((c->phys_proc_id == o->phys_proc_id) &&
523 (c->cpu_die_id == o->cpu_die_id))
529 #if defined(CONFIG_SCHED_SMT) || defined(CONFIG_SCHED_MC)
530 static inline int x86_sched_itmt_flags(void)
532 return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
535 #ifdef CONFIG_SCHED_MC
536 static int x86_core_flags(void)
538 return cpu_core_flags() | x86_sched_itmt_flags();
541 #ifdef CONFIG_SCHED_SMT
542 static int x86_smt_flags(void)
544 return cpu_smt_flags() | x86_sched_itmt_flags();
549 static struct sched_domain_topology_level x86_numa_in_package_topology[] = {
550 #ifdef CONFIG_SCHED_SMT
551 { cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
553 #ifdef CONFIG_SCHED_MC
554 { cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
559 static struct sched_domain_topology_level x86_topology[] = {
560 #ifdef CONFIG_SCHED_SMT
561 { cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
563 #ifdef CONFIG_SCHED_MC
564 { cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
566 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
571 * Set if a package/die has multiple NUMA nodes inside.
572 * AMD Magny-Cours, Intel Cluster-on-Die, and Intel
573 * Sub-NUMA Clustering have this.
575 static bool x86_has_numa_in_package;
577 void set_cpu_sibling_map(int cpu)
579 bool has_smt = smp_num_siblings > 1;
580 bool has_mp = has_smt || boot_cpu_data.x86_max_cores > 1;
581 struct cpuinfo_x86 *c = &cpu_data(cpu);
582 struct cpuinfo_x86 *o;
585 cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
588 cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
589 cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
590 cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
591 cpumask_set_cpu(cpu, topology_die_cpumask(cpu));
596 for_each_cpu(i, cpu_sibling_setup_mask) {
599 if ((i == cpu) || (has_smt && match_smt(c, o)))
600 link_mask(topology_sibling_cpumask, cpu, i);
602 if ((i == cpu) || (has_mp && match_llc(c, o)))
603 link_mask(cpu_llc_shared_mask, cpu, i);
608 * This needs a separate iteration over the cpus because we rely on all
609 * topology_sibling_cpumask links to be set-up.
611 for_each_cpu(i, cpu_sibling_setup_mask) {
614 if ((i == cpu) || (has_mp && match_pkg(c, o))) {
615 link_mask(topology_core_cpumask, cpu, i);
618 * Does this new cpu bringup a new core?
621 topology_sibling_cpumask(cpu)) == 1) {
623 * for each core in package, increment
624 * the booted_cores for this new cpu
627 topology_sibling_cpumask(i)) == i)
630 * increment the core count for all
631 * the other cpus in this package
634 cpu_data(i).booted_cores++;
635 } else if (i != cpu && !c->booted_cores)
636 c->booted_cores = cpu_data(i).booted_cores;
638 if (match_pkg(c, o) && !topology_same_node(c, o))
639 x86_has_numa_in_package = true;
641 if ((i == cpu) || (has_mp && match_die(c, o)))
642 link_mask(topology_die_cpumask, cpu, i);
645 threads = cpumask_weight(topology_sibling_cpumask(cpu));
646 if (threads > __max_smt_threads)
647 __max_smt_threads = threads;
650 /* maps the cpu to the sched domain representing multi-core */
651 const struct cpumask *cpu_coregroup_mask(int cpu)
653 return cpu_llc_shared_mask(cpu);
656 static void impress_friends(void)
659 unsigned long bogosum = 0;
661 * Allow the user to impress friends.
663 pr_debug("Before bogomips\n");
664 for_each_possible_cpu(cpu)
665 if (cpumask_test_cpu(cpu, cpu_callout_mask))
666 bogosum += cpu_data(cpu).loops_per_jiffy;
667 pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
670 (bogosum/(5000/HZ))%100);
672 pr_debug("Before bogocount - setting activated=1\n");
675 void __inquire_remote_apic(int apicid)
677 unsigned i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
678 const char * const names[] = { "ID", "VERSION", "SPIV" };
682 pr_info("Inquiring remote APIC 0x%x...\n", apicid);
684 for (i = 0; i < ARRAY_SIZE(regs); i++) {
685 pr_info("... APIC 0x%x %s: ", apicid, names[i]);
690 status = safe_apic_wait_icr_idle();
692 pr_cont("a previous APIC delivery may have failed\n");
694 apic_icr_write(APIC_DM_REMRD | regs[i], apicid);
699 status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
700 } while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
703 case APIC_ICR_RR_VALID:
704 status = apic_read(APIC_RRR);
705 pr_cont("%08x\n", status);
714 * The Multiprocessor Specification 1.4 (1997) example code suggests
715 * that there should be a 10ms delay between the BSP asserting INIT
716 * and de-asserting INIT, when starting a remote processor.
717 * But that slows boot and resume on modern processors, which include
718 * many cores and don't require that delay.
720 * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
721 * Modern processor families are quirked to remove the delay entirely.
723 #define UDELAY_10MS_DEFAULT 10000
725 static unsigned int init_udelay = UINT_MAX;
727 static int __init cpu_init_udelay(char *str)
729 get_option(&str, &init_udelay);
733 early_param("cpu_init_udelay", cpu_init_udelay);
735 static void __init smp_quirk_init_udelay(void)
737 /* if cmdline changed it from default, leave it alone */
738 if (init_udelay != UINT_MAX)
741 /* if modern processor, use no delay */
742 if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
743 ((boot_cpu_data.x86_vendor == X86_VENDOR_HYGON) && (boot_cpu_data.x86 >= 0x18)) ||
744 ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
748 /* else, use legacy delay */
749 init_udelay = UDELAY_10MS_DEFAULT;
753 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
754 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
755 * won't ... remember to clear down the APIC, etc later.
758 wakeup_secondary_cpu_via_nmi(int apicid, unsigned long start_eip)
760 unsigned long send_status, accept_status = 0;
764 /* Boot on the stack */
765 /* Kick the second */
766 apic_icr_write(APIC_DM_NMI | apic->dest_logical, apicid);
768 pr_debug("Waiting for send to finish...\n");
769 send_status = safe_apic_wait_icr_idle();
772 * Give the other CPU some time to accept the IPI.
775 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
776 maxlvt = lapic_get_maxlvt();
777 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
778 apic_write(APIC_ESR, 0);
779 accept_status = (apic_read(APIC_ESR) & 0xEF);
781 pr_debug("NMI sent\n");
784 pr_err("APIC never delivered???\n");
786 pr_err("APIC delivery error (%lx)\n", accept_status);
788 return (send_status | accept_status);
792 wakeup_secondary_cpu_via_init(int phys_apicid, unsigned long start_eip)
794 unsigned long send_status = 0, accept_status = 0;
795 int maxlvt, num_starts, j;
797 maxlvt = lapic_get_maxlvt();
800 * Be paranoid about clearing APIC errors.
802 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
803 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
804 apic_write(APIC_ESR, 0);
808 pr_debug("Asserting INIT\n");
811 * Turn INIT on target chip
816 apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT,
819 pr_debug("Waiting for send to finish...\n");
820 send_status = safe_apic_wait_icr_idle();
824 pr_debug("Deasserting INIT\n");
828 apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
830 pr_debug("Waiting for send to finish...\n");
831 send_status = safe_apic_wait_icr_idle();
836 * Should we send STARTUP IPIs ?
838 * Determine this based on the APIC version.
839 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
841 if (APIC_INTEGRATED(boot_cpu_apic_version))
847 * Run STARTUP IPI loop.
849 pr_debug("#startup loops: %d\n", num_starts);
851 for (j = 1; j <= num_starts; j++) {
852 pr_debug("Sending STARTUP #%d\n", j);
853 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
854 apic_write(APIC_ESR, 0);
856 pr_debug("After apic_write\n");
863 /* Boot on the stack */
864 /* Kick the second */
865 apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
869 * Give the other CPU some time to accept the IPI.
871 if (init_udelay == 0)
876 pr_debug("Startup point 1\n");
878 pr_debug("Waiting for send to finish...\n");
879 send_status = safe_apic_wait_icr_idle();
882 * Give the other CPU some time to accept the IPI.
884 if (init_udelay == 0)
889 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
890 apic_write(APIC_ESR, 0);
891 accept_status = (apic_read(APIC_ESR) & 0xEF);
892 if (send_status || accept_status)
895 pr_debug("After Startup\n");
898 pr_err("APIC never delivered???\n");
900 pr_err("APIC delivery error (%lx)\n", accept_status);
902 return (send_status | accept_status);
905 /* reduce the number of lines printed when booting a large cpu count system */
906 static void announce_cpu(int cpu, int apicid)
908 static int current_node = NUMA_NO_NODE;
909 int node = early_cpu_to_node(cpu);
910 static int width, node_width;
913 width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
916 node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
919 printk(KERN_INFO "x86: Booting SMP configuration:\n");
921 if (system_state < SYSTEM_RUNNING) {
922 if (node != current_node) {
923 if (current_node > (-1))
927 printk(KERN_INFO ".... node %*s#%d, CPUs: ",
928 node_width - num_digits(node), " ", node);
931 /* Add padding for the BSP */
933 pr_cont("%*s", width + 1, " ");
935 pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
938 pr_info("Booting Node %d Processor %d APIC 0x%x\n",
942 static int wakeup_cpu0_nmi(unsigned int cmd, struct pt_regs *regs)
946 cpu = smp_processor_id();
947 if (cpu == 0 && !cpu_online(cpu) && enable_start_cpu0)
954 * Wake up AP by INIT, INIT, STARTUP sequence.
956 * Instead of waiting for STARTUP after INITs, BSP will execute the BIOS
957 * boot-strap code which is not a desired behavior for waking up BSP. To
958 * void the boot-strap code, wake up CPU0 by NMI instead.
960 * This works to wake up soft offlined CPU0 only. If CPU0 is hard offlined
961 * (i.e. physically hot removed and then hot added), NMI won't wake it up.
962 * We'll change this code in the future to wake up hard offlined CPU0 if
963 * real platform and request are available.
966 wakeup_cpu_via_init_nmi(int cpu, unsigned long start_ip, int apicid,
967 int *cpu0_nmi_registered)
975 * Wake up AP by INIT, INIT, STARTUP sequence.
978 boot_error = wakeup_secondary_cpu_via_init(apicid, start_ip);
983 * Wake up BSP by nmi.
985 * Register a NMI handler to help wake up CPU0.
987 boot_error = register_nmi_handler(NMI_LOCAL,
988 wakeup_cpu0_nmi, 0, "wake_cpu0");
991 enable_start_cpu0 = 1;
992 *cpu0_nmi_registered = 1;
993 if (apic->dest_logical == APIC_DEST_LOGICAL)
994 id = cpu0_logical_apicid;
997 boot_error = wakeup_secondary_cpu_via_nmi(id, start_ip);
1006 int common_cpu_up(unsigned int cpu, struct task_struct *idle)
1010 /* Just in case we booted with a single CPU. */
1011 alternatives_enable_smp();
1013 per_cpu(current_task, cpu) = idle;
1015 /* Initialize the interrupt stack(s) */
1016 ret = irq_init_percpu_irqstack(cpu);
1020 #ifdef CONFIG_X86_32
1021 /* Stack for startup_32 can be just as for start_secondary onwards */
1022 per_cpu(cpu_current_top_of_stack, cpu) = task_top_of_stack(idle);
1024 initial_gs = per_cpu_offset(cpu);
1030 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
1031 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
1032 * Returns zero if CPU booted OK, else error code from
1033 * ->wakeup_secondary_cpu.
1035 static int do_boot_cpu(int apicid, int cpu, struct task_struct *idle,
1036 int *cpu0_nmi_registered)
1038 volatile u32 *trampoline_status =
1039 (volatile u32 *) __va(real_mode_header->trampoline_status);
1040 /* start_ip had better be page-aligned! */
1041 unsigned long start_ip = real_mode_header->trampoline_start;
1043 unsigned long boot_error = 0;
1044 unsigned long timeout;
1046 idle->thread.sp = (unsigned long)task_pt_regs(idle);
1047 early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
1048 initial_code = (unsigned long)start_secondary;
1049 initial_stack = idle->thread.sp;
1051 /* Enable the espfix hack for this CPU */
1052 init_espfix_ap(cpu);
1054 /* So we see what's up */
1055 announce_cpu(cpu, apicid);
1058 * This grunge runs the startup process for
1059 * the targeted processor.
1062 if (x86_platform.legacy.warm_reset) {
1064 pr_debug("Setting warm reset code and vector.\n");
1066 smpboot_setup_warm_reset_vector(start_ip);
1068 * Be paranoid about clearing APIC errors.
1070 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
1071 apic_write(APIC_ESR, 0);
1072 apic_read(APIC_ESR);
1077 * AP might wait on cpu_callout_mask in cpu_init() with
1078 * cpu_initialized_mask set if previous attempt to online
1079 * it timed-out. Clear cpu_initialized_mask so that after
1080 * INIT/SIPI it could start with a clean state.
1082 cpumask_clear_cpu(cpu, cpu_initialized_mask);
1086 * Wake up a CPU in difference cases:
1087 * - Use the method in the APIC driver if it's defined
1089 * - Use an INIT boot APIC message for APs or NMI for BSP.
1091 if (apic->wakeup_secondary_cpu)
1092 boot_error = apic->wakeup_secondary_cpu(apicid, start_ip);
1094 boot_error = wakeup_cpu_via_init_nmi(cpu, start_ip, apicid,
1095 cpu0_nmi_registered);
1099 * Wait 10s total for first sign of life from AP
1102 timeout = jiffies + 10*HZ;
1103 while (time_before(jiffies, timeout)) {
1104 if (cpumask_test_cpu(cpu, cpu_initialized_mask)) {
1106 * Tell AP to proceed with initialization
1108 cpumask_set_cpu(cpu, cpu_callout_mask);
1118 * Wait till AP completes initial initialization
1120 while (!cpumask_test_cpu(cpu, cpu_callin_mask)) {
1122 * Allow other tasks to run while we wait for the
1123 * AP to come online. This also gives a chance
1124 * for the MTRR work(triggered by the AP coming online)
1125 * to be completed in the stop machine context.
1131 /* mark "stuck" area as not stuck */
1132 *trampoline_status = 0;
1134 if (x86_platform.legacy.warm_reset) {
1136 * Cleanup possible dangling ends...
1138 smpboot_restore_warm_reset_vector();
1144 int native_cpu_up(unsigned int cpu, struct task_struct *tidle)
1146 int apicid = apic->cpu_present_to_apicid(cpu);
1147 int cpu0_nmi_registered = 0;
1148 unsigned long flags;
1151 lockdep_assert_irqs_enabled();
1153 pr_debug("++++++++++++++++++++=_---CPU UP %u\n", cpu);
1155 if (apicid == BAD_APICID ||
1156 !physid_isset(apicid, phys_cpu_present_map) ||
1157 !apic->apic_id_valid(apicid)) {
1158 pr_err("%s: bad cpu %d\n", __func__, cpu);
1163 * Already booted CPU?
1165 if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
1166 pr_debug("do_boot_cpu %d Already started\n", cpu);
1171 * Save current MTRR state in case it was changed since early boot
1172 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1176 /* x86 CPUs take themselves offline, so delayed offline is OK. */
1177 err = cpu_check_up_prepare(cpu);
1178 if (err && err != -EBUSY)
1181 /* the FPU context is blank, nobody can own it */
1182 per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
1184 err = common_cpu_up(cpu, tidle);
1188 err = do_boot_cpu(apicid, cpu, tidle, &cpu0_nmi_registered);
1190 pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
1196 * Check TSC synchronization with the AP (keep irqs disabled
1199 local_irq_save(flags);
1200 check_tsc_sync_source(cpu);
1201 local_irq_restore(flags);
1203 while (!cpu_online(cpu)) {
1205 touch_nmi_watchdog();
1210 * Clean up the nmi handler. Do this after the callin and callout sync
1211 * to avoid impact of possible long unregister time.
1213 if (cpu0_nmi_registered)
1214 unregister_nmi_handler(NMI_LOCAL, "wake_cpu0");
1220 * arch_disable_smp_support() - disables SMP support for x86 at runtime
1222 void arch_disable_smp_support(void)
1224 disable_ioapic_support();
1228 * Fall back to non SMP mode after errors.
1230 * RED-PEN audit/test this more. I bet there is more state messed up here.
1232 static __init void disable_smp(void)
1234 pr_info("SMP disabled\n");
1236 disable_ioapic_support();
1238 init_cpu_present(cpumask_of(0));
1239 init_cpu_possible(cpumask_of(0));
1241 if (smp_found_config)
1242 physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
1244 physid_set_mask_of_physid(0, &phys_cpu_present_map);
1245 cpumask_set_cpu(0, topology_sibling_cpumask(0));
1246 cpumask_set_cpu(0, topology_core_cpumask(0));
1247 cpumask_set_cpu(0, topology_die_cpumask(0));
1251 * Various sanity checks.
1253 static void __init smp_sanity_check(void)
1257 #if !defined(CONFIG_X86_BIGSMP) && defined(CONFIG_X86_32)
1258 if (def_to_bigsmp && nr_cpu_ids > 8) {
1262 pr_warn("More than 8 CPUs detected - skipping them\n"
1263 "Use CONFIG_X86_BIGSMP\n");
1266 for_each_present_cpu(cpu) {
1268 set_cpu_present(cpu, false);
1273 for_each_possible_cpu(cpu) {
1275 set_cpu_possible(cpu, false);
1283 if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map)) {
1284 pr_warn("weird, boot CPU (#%d) not listed by the BIOS\n",
1285 hard_smp_processor_id());
1287 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1291 * Should not be necessary because the MP table should list the boot
1292 * CPU too, but we do it for the sake of robustness anyway.
1294 if (!apic->check_phys_apicid_present(boot_cpu_physical_apicid)) {
1295 pr_notice("weird, boot CPU (#%d) not listed by the BIOS\n",
1296 boot_cpu_physical_apicid);
1297 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1302 static void __init smp_cpu_index_default(void)
1305 struct cpuinfo_x86 *c;
1307 for_each_possible_cpu(i) {
1309 /* mark all to hotplug */
1310 c->cpu_index = nr_cpu_ids;
1314 static void __init smp_get_logical_apicid(void)
1317 cpu0_logical_apicid = apic_read(APIC_LDR);
1319 cpu0_logical_apicid = GET_APIC_LOGICAL_ID(apic_read(APIC_LDR));
1323 * Prepare for SMP bootup.
1324 * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1325 * for common interface support.
1327 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1331 smp_cpu_index_default();
1334 * Setup boot CPU information
1336 smp_store_boot_cpu_info(); /* Final full version of the data */
1337 cpumask_copy(cpu_callin_mask, cpumask_of(0));
1340 for_each_possible_cpu(i) {
1341 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1342 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1343 zalloc_cpumask_var(&per_cpu(cpu_die_map, i), GFP_KERNEL);
1344 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1348 * Set 'default' x86 topology, this matches default_topology() in that
1349 * it has NUMA nodes as a topology level. See also
1350 * native_smp_cpus_done().
1352 * Must be done before set_cpus_sibling_map() is ran.
1354 set_sched_topology(x86_topology);
1356 set_cpu_sibling_map(0);
1360 switch (apic_intr_mode) {
1362 case APIC_VIRTUAL_WIRE_NO_CONFIG:
1365 case APIC_SYMMETRIC_IO_NO_ROUTING:
1367 /* Setup local timer */
1368 x86_init.timers.setup_percpu_clockev();
1370 case APIC_VIRTUAL_WIRE:
1371 case APIC_SYMMETRIC_IO:
1375 /* Setup local timer */
1376 x86_init.timers.setup_percpu_clockev();
1378 smp_get_logical_apicid();
1381 print_cpu_info(&cpu_data(0));
1383 native_pv_lock_init();
1387 set_mtrr_aps_delayed_init();
1389 smp_quirk_init_udelay();
1391 speculative_store_bypass_ht_init();
1394 void arch_enable_nonboot_cpus_begin(void)
1396 set_mtrr_aps_delayed_init();
1399 void arch_enable_nonboot_cpus_end(void)
1405 * Early setup to make printk work.
1407 void __init native_smp_prepare_boot_cpu(void)
1409 int me = smp_processor_id();
1410 switch_to_new_gdt(me);
1411 /* already set me in cpu_online_mask in boot_cpu_init() */
1412 cpumask_set_cpu(me, cpu_callout_mask);
1413 cpu_set_state_online(me);
1416 void __init calculate_max_logical_packages(void)
1421 * Today neither Intel nor AMD support heterogenous systems so
1422 * extrapolate the boot cpu's data to all packages.
1424 ncpus = cpu_data(0).booted_cores * topology_max_smt_threads();
1425 __max_logical_packages = DIV_ROUND_UP(total_cpus, ncpus);
1426 pr_info("Max logical packages: %u\n", __max_logical_packages);
1429 void __init native_smp_cpus_done(unsigned int max_cpus)
1431 pr_debug("Boot done\n");
1433 calculate_max_logical_packages();
1435 if (x86_has_numa_in_package)
1436 set_sched_topology(x86_numa_in_package_topology);
1443 static int __initdata setup_possible_cpus = -1;
1444 static int __init _setup_possible_cpus(char *str)
1446 get_option(&str, &setup_possible_cpus);
1449 early_param("possible_cpus", _setup_possible_cpus);
1453 * cpu_possible_mask should be static, it cannot change as cpu's
1454 * are onlined, or offlined. The reason is per-cpu data-structures
1455 * are allocated by some modules at init time, and dont expect to
1456 * do this dynamically on cpu arrival/departure.
1457 * cpu_present_mask on the other hand can change dynamically.
1458 * In case when cpu_hotplug is not compiled, then we resort to current
1459 * behaviour, which is cpu_possible == cpu_present.
1462 * Three ways to find out the number of additional hotplug CPUs:
1463 * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1464 * - The user can overwrite it with possible_cpus=NUM
1465 * - Otherwise don't reserve additional CPUs.
1466 * We do this because additional CPUs waste a lot of memory.
1469 __init void prefill_possible_map(void)
1473 /* No boot processor was found in mptable or ACPI MADT */
1474 if (!num_processors) {
1475 if (boot_cpu_has(X86_FEATURE_APIC)) {
1476 int apicid = boot_cpu_physical_apicid;
1477 int cpu = hard_smp_processor_id();
1479 pr_warn("Boot CPU (id %d) not listed by BIOS\n", cpu);
1481 /* Make sure boot cpu is enumerated */
1482 if (apic->cpu_present_to_apicid(0) == BAD_APICID &&
1483 apic->apic_id_valid(apicid))
1484 generic_processor_info(apicid, boot_cpu_apic_version);
1487 if (!num_processors)
1491 i = setup_max_cpus ?: 1;
1492 if (setup_possible_cpus == -1) {
1493 possible = num_processors;
1494 #ifdef CONFIG_HOTPLUG_CPU
1496 possible += disabled_cpus;
1502 possible = setup_possible_cpus;
1504 total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1506 /* nr_cpu_ids could be reduced via nr_cpus= */
1507 if (possible > nr_cpu_ids) {
1508 pr_warn("%d Processors exceeds NR_CPUS limit of %u\n",
1509 possible, nr_cpu_ids);
1510 possible = nr_cpu_ids;
1513 #ifdef CONFIG_HOTPLUG_CPU
1514 if (!setup_max_cpus)
1517 pr_warn("%d Processors exceeds max_cpus limit of %u\n",
1518 possible, setup_max_cpus);
1522 nr_cpu_ids = possible;
1524 pr_info("Allowing %d CPUs, %d hotplug CPUs\n",
1525 possible, max_t(int, possible - num_processors, 0));
1527 reset_cpu_possible_mask();
1529 for (i = 0; i < possible; i++)
1530 set_cpu_possible(i, true);
1533 #ifdef CONFIG_HOTPLUG_CPU
1535 /* Recompute SMT state for all CPUs on offline */
1536 static void recompute_smt_state(void)
1538 int max_threads, cpu;
1541 for_each_online_cpu (cpu) {
1542 int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1544 if (threads > max_threads)
1545 max_threads = threads;
1547 __max_smt_threads = max_threads;
1550 static void remove_siblinginfo(int cpu)
1553 struct cpuinfo_x86 *c = &cpu_data(cpu);
1555 for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1556 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1558 * last thread sibling in this cpu core going down
1560 if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1561 cpu_data(sibling).booted_cores--;
1564 for_each_cpu(sibling, topology_die_cpumask(cpu))
1565 cpumask_clear_cpu(cpu, topology_die_cpumask(sibling));
1566 for_each_cpu(sibling, topology_sibling_cpumask(cpu))
1567 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1568 for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1569 cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1570 cpumask_clear(cpu_llc_shared_mask(cpu));
1571 cpumask_clear(topology_sibling_cpumask(cpu));
1572 cpumask_clear(topology_core_cpumask(cpu));
1573 cpumask_clear(topology_die_cpumask(cpu));
1575 c->booted_cores = 0;
1576 cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1577 recompute_smt_state();
1580 static void remove_cpu_from_maps(int cpu)
1582 set_cpu_online(cpu, false);
1583 cpumask_clear_cpu(cpu, cpu_callout_mask);
1584 cpumask_clear_cpu(cpu, cpu_callin_mask);
1585 /* was set by cpu_init() */
1586 cpumask_clear_cpu(cpu, cpu_initialized_mask);
1587 numa_remove_cpu(cpu);
1590 void cpu_disable_common(void)
1592 int cpu = smp_processor_id();
1594 remove_siblinginfo(cpu);
1596 /* It's now safe to remove this processor from the online map */
1598 remove_cpu_from_maps(cpu);
1599 unlock_vector_lock();
1604 int native_cpu_disable(void)
1608 ret = lapic_can_unplug_cpu();
1613 cpu_disable_common();
1618 int common_cpu_die(unsigned int cpu)
1622 /* We don't do anything here: idle task is faking death itself. */
1624 /* They ack this in play_dead() by setting CPU_DEAD */
1625 if (cpu_wait_death(cpu, 5)) {
1626 if (system_state == SYSTEM_RUNNING)
1627 pr_info("CPU %u is now offline\n", cpu);
1629 pr_err("CPU %u didn't die...\n", cpu);
1636 void native_cpu_die(unsigned int cpu)
1638 common_cpu_die(cpu);
1641 void play_dead_common(void)
1646 (void)cpu_report_death();
1649 * With physical CPU hotplug, we should halt the cpu
1651 local_irq_disable();
1654 static bool wakeup_cpu0(void)
1656 if (smp_processor_id() == 0 && enable_start_cpu0)
1663 * We need to flush the caches before going to sleep, lest we have
1664 * dirty data in our caches when we come back up.
1666 static inline void mwait_play_dead(void)
1668 unsigned int eax, ebx, ecx, edx;
1669 unsigned int highest_cstate = 0;
1670 unsigned int highest_subcstate = 0;
1674 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD ||
1675 boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
1677 if (!this_cpu_has(X86_FEATURE_MWAIT))
1679 if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1681 if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1684 eax = CPUID_MWAIT_LEAF;
1686 native_cpuid(&eax, &ebx, &ecx, &edx);
1689 * eax will be 0 if EDX enumeration is not valid.
1690 * Initialized below to cstate, sub_cstate value when EDX is valid.
1692 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1695 edx >>= MWAIT_SUBSTATE_SIZE;
1696 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1697 if (edx & MWAIT_SUBSTATE_MASK) {
1699 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1702 eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1703 (highest_subcstate - 1);
1707 * This should be a memory location in a cache line which is
1708 * unlikely to be touched by other processors. The actual
1709 * content is immaterial as it is not actually modified in any way.
1711 mwait_ptr = ¤t_thread_info()->flags;
1717 * The CLFLUSH is a workaround for erratum AAI65 for
1718 * the Xeon 7400 series. It's not clear it is actually
1719 * needed, but it should be harmless in either case.
1720 * The WBINVD is insufficient due to the spurious-wakeup
1721 * case where we return around the loop.
1726 __monitor(mwait_ptr, 0, 0);
1730 * If NMI wants to wake up CPU0, start CPU0.
1737 void hlt_play_dead(void)
1739 if (__this_cpu_read(cpu_info.x86) >= 4)
1745 * If NMI wants to wake up CPU0, start CPU0.
1752 void native_play_dead(void)
1755 tboot_shutdown(TB_SHUTDOWN_WFS);
1757 mwait_play_dead(); /* Only returns on failure */
1758 if (cpuidle_play_dead())
1762 #else /* ... !CONFIG_HOTPLUG_CPU */
1763 int native_cpu_disable(void)
1768 void native_cpu_die(unsigned int cpu)
1770 /* We said "no" in __cpu_disable */
1774 void native_play_dead(void)