2 * x86 SMP booting functions
4 * (c) 1995 Alan Cox, Building #3 <alan@lxorguk.ukuu.org.uk>
5 * (c) 1998, 1999, 2000, 2009 Ingo Molnar <mingo@redhat.com>
6 * Copyright 2001 Andi Kleen, SuSE Labs.
8 * Much of the core SMP work is based on previous work by Thomas Radke, to
9 * whom a great many thanks are extended.
11 * Thanks to Intel for making available several different Pentium,
12 * Pentium Pro and Pentium-II/Xeon MP machines.
13 * Original development of Linux SMP code supported by Caldera.
15 * This code is released under the GNU General Public License version 2 or
19 * Felix Koop : NR_CPUS used properly
20 * Jose Renau : Handle single CPU case.
21 * Alan Cox : By repeated request 8) - Total BogoMIPS report.
22 * Greg Wright : Fix for kernel stacks panic.
23 * Erich Boleyn : MP v1.4 and additional changes.
24 * Matthias Sattler : Changes for 2.1 kernel map.
25 * Michel Lespinasse : Changes for 2.1 kernel map.
26 * Michael Chastain : Change trampoline.S to gnu as.
27 * Alan Cox : Dumb bug: 'B' step PPro's are fine
28 * Ingo Molnar : Added APIC timers, based on code
30 * Ingo Molnar : various cleanups and rewrites
31 * Tigran Aivazian : fixed "0.00 in /proc/uptime on SMP" bug.
32 * Maciej W. Rozycki : Bits for genuine 82489DX APICs
33 * Andi Kleen : Changed for SMP boot into long mode.
34 * Martin J. Bligh : Added support for multi-quad systems
35 * Dave Jones : Report invalid combinations of Athlon CPUs.
36 * Rusty Russell : Hacked into shape for new "hotplug" boot process.
37 * Andi Kleen : Converted to new state machine.
38 * Ashok Raj : CPU hotplug support
39 * Glauber Costa : i386 and x86_64 integration
42 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
44 #include <linux/init.h>
45 #include <linux/smp.h>
46 #include <linux/export.h>
47 #include <linux/sched.h>
48 #include <linux/sched/topology.h>
49 #include <linux/sched/hotplug.h>
50 #include <linux/sched/task_stack.h>
51 #include <linux/percpu.h>
52 #include <linux/memblock.h>
53 #include <linux/err.h>
54 #include <linux/nmi.h>
55 #include <linux/tboot.h>
56 #include <linux/stackprotector.h>
57 #include <linux/gfp.h>
58 #include <linux/cpuidle.h>
64 #include <asm/realmode.h>
67 #include <asm/pgtable.h>
68 #include <asm/tlbflush.h>
70 #include <asm/mwait.h>
72 #include <asm/io_apic.h>
73 #include <asm/fpu/internal.h>
74 #include <asm/setup.h>
75 #include <asm/uv/uv.h>
76 #include <linux/mc146818rtc.h>
77 #include <asm/i8259.h>
79 #include <asm/qspinlock.h>
80 #include <asm/intel-family.h>
81 #include <asm/cpu_device_id.h>
82 #include <asm/spec-ctrl.h>
83 #include <asm/hw_irq.h>
85 /* representing HT siblings of each logical CPU */
86 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
87 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
89 /* representing HT and core siblings of each logical CPU */
90 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
91 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
93 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_llc_shared_map);
95 /* Per CPU bogomips and other parameters */
96 DEFINE_PER_CPU_READ_MOSTLY(struct cpuinfo_x86, cpu_info);
97 EXPORT_PER_CPU_SYMBOL(cpu_info);
99 /* Logical package management. We might want to allocate that dynamically */
100 unsigned int __max_logical_packages __read_mostly;
101 EXPORT_SYMBOL(__max_logical_packages);
102 static unsigned int logical_packages __read_mostly;
104 /* Maximum number of SMT threads on any online core */
105 int __read_mostly __max_smt_threads = 1;
107 /* Flag to indicate if a complete sched domain rebuild is required */
108 bool x86_topology_update;
110 int arch_update_cpu_topology(void)
112 int retval = x86_topology_update;
114 x86_topology_update = false;
118 static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
122 spin_lock_irqsave(&rtc_lock, flags);
123 CMOS_WRITE(0xa, 0xf);
124 spin_unlock_irqrestore(&rtc_lock, flags);
125 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) =
127 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) =
131 static inline void smpboot_restore_warm_reset_vector(void)
136 * Paranoid: Set warm reset code and vector here back
139 spin_lock_irqsave(&rtc_lock, flags);
141 spin_unlock_irqrestore(&rtc_lock, flags);
143 *((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
147 * Report back to the Boot Processor during boot time or to the caller processor
150 static void smp_callin(void)
155 * If waken up by an INIT in an 82489DX configuration
156 * cpu_callout_mask guarantees we don't get here before
157 * an INIT_deassert IPI reaches our local APIC, so it is
158 * now safe to touch our local APIC.
160 cpuid = smp_processor_id();
163 * the boot CPU has finished the init stage and is spinning
164 * on callin_map until we finish. We are free to set up this
165 * CPU, first the APIC. (this is probably redundant on most
171 * Save our processor parameters. Note: this information
172 * is needed for clock calibration.
174 smp_store_cpu_info(cpuid);
177 * The topology information must be up to date before
178 * calibrate_delay() and notify_cpu_starting().
180 set_cpu_sibling_map(raw_smp_processor_id());
184 * Update loops_per_jiffy in cpu_data. Previous call to
185 * smp_store_cpu_info() stored a value that is close but not as
186 * accurate as the value just calculated.
189 cpu_data(cpuid).loops_per_jiffy = loops_per_jiffy;
190 pr_debug("Stack at about %p\n", &cpuid);
194 notify_cpu_starting(cpuid);
197 * Allow the master to continue.
199 cpumask_set_cpu(cpuid, cpu_callin_mask);
202 static int cpu0_logical_apicid;
203 static int enable_start_cpu0;
205 * Activate a secondary processor.
207 static void notrace start_secondary(void *unused)
210 * Don't put *anything* except direct CPU state initialization
211 * before cpu_init(), SMP booting is too fragile that we want to
212 * limit the things done here to the most necessary things.
214 if (boot_cpu_has(X86_FEATURE_PCID))
215 __write_cr4(__read_cr4() | X86_CR4_PCIDE);
218 /* switch away from the initial page table */
219 load_cr3(swapper_pg_dir);
221 * Initialize the CR4 shadow before doing anything that could
229 x86_cpuinit.early_percpu_clock_init();
233 enable_start_cpu0 = 0;
235 /* otherwise gcc will move up smp_processor_id before the cpu_init */
238 * Check TSC synchronization with the boot CPU:
240 check_tsc_sync_target();
242 speculative_store_bypass_ht_init();
245 * Lock vector_lock, set CPU online and bring the vector
246 * allocator online. Online must be set with vector_lock held
247 * to prevent a concurrent irq setup/teardown from seeing a
248 * half valid vector space.
251 set_cpu_online(smp_processor_id(), true);
253 unlock_vector_lock();
254 cpu_set_state_online(smp_processor_id());
255 x86_platform.nmi_init();
257 /* enable local interrupts */
260 /* to prevent fake stack check failure in clock setup */
261 boot_init_stack_canary();
263 x86_cpuinit.setup_percpu_clockev();
266 cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
270 * topology_is_primary_thread - Check whether CPU is the primary SMT thread
273 bool topology_is_primary_thread(unsigned int cpu)
275 return apic_id_is_primary_thread(per_cpu(x86_cpu_to_apicid, cpu));
279 * topology_smt_supported - Check whether SMT is supported by the CPUs
281 bool topology_smt_supported(void)
283 return smp_num_siblings > 1;
287 * topology_phys_to_logical_pkg - Map a physical package id to a logical
289 * Returns logical package id or -1 if not found
291 int topology_phys_to_logical_pkg(unsigned int phys_pkg)
295 for_each_possible_cpu(cpu) {
296 struct cpuinfo_x86 *c = &cpu_data(cpu);
298 if (c->initialized && c->phys_proc_id == phys_pkg)
299 return c->logical_proc_id;
303 EXPORT_SYMBOL(topology_phys_to_logical_pkg);
306 * topology_update_package_map - Update the physical to logical package map
307 * @pkg: The physical package id as retrieved via CPUID
308 * @cpu: The cpu for which this is updated
310 int topology_update_package_map(unsigned int pkg, unsigned int cpu)
314 /* Already available somewhere? */
315 new = topology_phys_to_logical_pkg(pkg);
319 new = logical_packages++;
321 pr_info("CPU %u Converting physical %u to logical package %u\n",
325 cpu_data(cpu).logical_proc_id = new;
329 void __init smp_store_boot_cpu_info(void)
331 int id = 0; /* CPU 0 */
332 struct cpuinfo_x86 *c = &cpu_data(id);
336 topology_update_package_map(c->phys_proc_id, id);
337 c->initialized = true;
341 * The bootstrap kernel entry code has set these up. Save them for
344 void smp_store_cpu_info(int id)
346 struct cpuinfo_x86 *c = &cpu_data(id);
348 /* Copy boot_cpu_data only on the first bringup */
353 * During boot time, CPU0 has this setup already. Save the info when
354 * bringing up AP or offlined CPU0.
356 identify_secondary_cpu(c);
357 c->initialized = true;
361 topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
363 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
365 return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
369 topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
371 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
373 return !WARN_ONCE(!topology_same_node(c, o),
374 "sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
375 "[node: %d != %d]. Ignoring dependency.\n",
376 cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
379 #define link_mask(mfunc, c1, c2) \
381 cpumask_set_cpu((c1), mfunc(c2)); \
382 cpumask_set_cpu((c2), mfunc(c1)); \
385 static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
387 if (boot_cpu_has(X86_FEATURE_TOPOEXT)) {
388 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
390 if (c->phys_proc_id == o->phys_proc_id &&
391 per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2)) {
392 if (c->cpu_core_id == o->cpu_core_id)
393 return topology_sane(c, o, "smt");
395 if ((c->cu_id != 0xff) &&
396 (o->cu_id != 0xff) &&
397 (c->cu_id == o->cu_id))
398 return topology_sane(c, o, "smt");
401 } else if (c->phys_proc_id == o->phys_proc_id &&
402 c->cpu_core_id == o->cpu_core_id) {
403 return topology_sane(c, o, "smt");
410 * Define snc_cpu[] for SNC (Sub-NUMA Cluster) CPUs.
412 * These are Intel CPUs that enumerate an LLC that is shared by
413 * multiple NUMA nodes. The LLC on these systems is shared for
414 * off-package data access but private to the NUMA node (half
415 * of the package) for on-package access.
417 * CPUID (the source of the information about the LLC) can only
418 * enumerate the cache as being shared *or* unshared, but not
419 * this particular configuration. The CPU in this case enumerates
420 * the cache to be shared across the entire package (spanning both
424 static const struct x86_cpu_id snc_cpu[] = {
425 { X86_VENDOR_INTEL, 6, INTEL_FAM6_SKYLAKE_X },
429 static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
431 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
433 /* Do not match if we do not have a valid APICID for cpu: */
434 if (per_cpu(cpu_llc_id, cpu1) == BAD_APICID)
437 /* Do not match if LLC id does not match: */
438 if (per_cpu(cpu_llc_id, cpu1) != per_cpu(cpu_llc_id, cpu2))
442 * Allow the SNC topology without warning. Return of false
443 * means 'c' does not share the LLC of 'o'. This will be
444 * reflected to userspace.
446 if (!topology_same_node(c, o) && x86_match_cpu(snc_cpu))
449 return topology_sane(c, o, "llc");
453 * Unlike the other levels, we do not enforce keeping a
454 * multicore group inside a NUMA node. If this happens, we will
455 * discard the MC level of the topology later.
457 static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
459 if (c->phys_proc_id == o->phys_proc_id)
464 #if defined(CONFIG_SCHED_SMT) || defined(CONFIG_SCHED_MC)
465 static inline int x86_sched_itmt_flags(void)
467 return sysctl_sched_itmt_enabled ? SD_ASYM_PACKING : 0;
470 #ifdef CONFIG_SCHED_MC
471 static int x86_core_flags(void)
473 return cpu_core_flags() | x86_sched_itmt_flags();
476 #ifdef CONFIG_SCHED_SMT
477 static int x86_smt_flags(void)
479 return cpu_smt_flags() | x86_sched_itmt_flags();
484 static struct sched_domain_topology_level x86_numa_in_package_topology[] = {
485 #ifdef CONFIG_SCHED_SMT
486 { cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
488 #ifdef CONFIG_SCHED_MC
489 { cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
494 static struct sched_domain_topology_level x86_topology[] = {
495 #ifdef CONFIG_SCHED_SMT
496 { cpu_smt_mask, x86_smt_flags, SD_INIT_NAME(SMT) },
498 #ifdef CONFIG_SCHED_MC
499 { cpu_coregroup_mask, x86_core_flags, SD_INIT_NAME(MC) },
501 { cpu_cpu_mask, SD_INIT_NAME(DIE) },
506 * Set if a package/die has multiple NUMA nodes inside.
507 * AMD Magny-Cours, Intel Cluster-on-Die, and Intel
508 * Sub-NUMA Clustering have this.
510 static bool x86_has_numa_in_package;
512 void set_cpu_sibling_map(int cpu)
514 bool has_smt = smp_num_siblings > 1;
515 bool has_mp = has_smt || boot_cpu_data.x86_max_cores > 1;
516 struct cpuinfo_x86 *c = &cpu_data(cpu);
517 struct cpuinfo_x86 *o;
520 cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
523 cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
524 cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
525 cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
530 for_each_cpu(i, cpu_sibling_setup_mask) {
533 if ((i == cpu) || (has_smt && match_smt(c, o)))
534 link_mask(topology_sibling_cpumask, cpu, i);
536 if ((i == cpu) || (has_mp && match_llc(c, o)))
537 link_mask(cpu_llc_shared_mask, cpu, i);
542 * This needs a separate iteration over the cpus because we rely on all
543 * topology_sibling_cpumask links to be set-up.
545 for_each_cpu(i, cpu_sibling_setup_mask) {
548 if ((i == cpu) || (has_mp && match_die(c, o))) {
549 link_mask(topology_core_cpumask, cpu, i);
552 * Does this new cpu bringup a new core?
555 topology_sibling_cpumask(cpu)) == 1) {
557 * for each core in package, increment
558 * the booted_cores for this new cpu
561 topology_sibling_cpumask(i)) == i)
564 * increment the core count for all
565 * the other cpus in this package
568 cpu_data(i).booted_cores++;
569 } else if (i != cpu && !c->booted_cores)
570 c->booted_cores = cpu_data(i).booted_cores;
572 if (match_die(c, o) && !topology_same_node(c, o))
573 x86_has_numa_in_package = true;
576 threads = cpumask_weight(topology_sibling_cpumask(cpu));
577 if (threads > __max_smt_threads)
578 __max_smt_threads = threads;
581 /* maps the cpu to the sched domain representing multi-core */
582 const struct cpumask *cpu_coregroup_mask(int cpu)
584 return cpu_llc_shared_mask(cpu);
587 static void impress_friends(void)
590 unsigned long bogosum = 0;
592 * Allow the user to impress friends.
594 pr_debug("Before bogomips\n");
595 for_each_possible_cpu(cpu)
596 if (cpumask_test_cpu(cpu, cpu_callout_mask))
597 bogosum += cpu_data(cpu).loops_per_jiffy;
598 pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
601 (bogosum/(5000/HZ))%100);
603 pr_debug("Before bogocount - setting activated=1\n");
606 void __inquire_remote_apic(int apicid)
608 unsigned i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
609 const char * const names[] = { "ID", "VERSION", "SPIV" };
613 pr_info("Inquiring remote APIC 0x%x...\n", apicid);
615 for (i = 0; i < ARRAY_SIZE(regs); i++) {
616 pr_info("... APIC 0x%x %s: ", apicid, names[i]);
621 status = safe_apic_wait_icr_idle();
623 pr_cont("a previous APIC delivery may have failed\n");
625 apic_icr_write(APIC_DM_REMRD | regs[i], apicid);
630 status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
631 } while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
634 case APIC_ICR_RR_VALID:
635 status = apic_read(APIC_RRR);
636 pr_cont("%08x\n", status);
645 * The Multiprocessor Specification 1.4 (1997) example code suggests
646 * that there should be a 10ms delay between the BSP asserting INIT
647 * and de-asserting INIT, when starting a remote processor.
648 * But that slows boot and resume on modern processors, which include
649 * many cores and don't require that delay.
651 * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
652 * Modern processor families are quirked to remove the delay entirely.
654 #define UDELAY_10MS_DEFAULT 10000
656 static unsigned int init_udelay = UINT_MAX;
658 static int __init cpu_init_udelay(char *str)
660 get_option(&str, &init_udelay);
664 early_param("cpu_init_udelay", cpu_init_udelay);
666 static void __init smp_quirk_init_udelay(void)
668 /* if cmdline changed it from default, leave it alone */
669 if (init_udelay != UINT_MAX)
672 /* if modern processor, use no delay */
673 if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
674 ((boot_cpu_data.x86_vendor == X86_VENDOR_HYGON) && (boot_cpu_data.x86 >= 0x18)) ||
675 ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF))) {
679 /* else, use legacy delay */
680 init_udelay = UDELAY_10MS_DEFAULT;
684 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
685 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
686 * won't ... remember to clear down the APIC, etc later.
689 wakeup_secondary_cpu_via_nmi(int apicid, unsigned long start_eip)
691 unsigned long send_status, accept_status = 0;
695 /* Boot on the stack */
696 /* Kick the second */
697 apic_icr_write(APIC_DM_NMI | apic->dest_logical, apicid);
699 pr_debug("Waiting for send to finish...\n");
700 send_status = safe_apic_wait_icr_idle();
703 * Give the other CPU some time to accept the IPI.
706 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
707 maxlvt = lapic_get_maxlvt();
708 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
709 apic_write(APIC_ESR, 0);
710 accept_status = (apic_read(APIC_ESR) & 0xEF);
712 pr_debug("NMI sent\n");
715 pr_err("APIC never delivered???\n");
717 pr_err("APIC delivery error (%lx)\n", accept_status);
719 return (send_status | accept_status);
723 wakeup_secondary_cpu_via_init(int phys_apicid, unsigned long start_eip)
725 unsigned long send_status = 0, accept_status = 0;
726 int maxlvt, num_starts, j;
728 maxlvt = lapic_get_maxlvt();
731 * Be paranoid about clearing APIC errors.
733 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
734 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
735 apic_write(APIC_ESR, 0);
739 pr_debug("Asserting INIT\n");
742 * Turn INIT on target chip
747 apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT,
750 pr_debug("Waiting for send to finish...\n");
751 send_status = safe_apic_wait_icr_idle();
755 pr_debug("Deasserting INIT\n");
759 apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
761 pr_debug("Waiting for send to finish...\n");
762 send_status = safe_apic_wait_icr_idle();
767 * Should we send STARTUP IPIs ?
769 * Determine this based on the APIC version.
770 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
772 if (APIC_INTEGRATED(boot_cpu_apic_version))
778 * Run STARTUP IPI loop.
780 pr_debug("#startup loops: %d\n", num_starts);
782 for (j = 1; j <= num_starts; j++) {
783 pr_debug("Sending STARTUP #%d\n", j);
784 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
785 apic_write(APIC_ESR, 0);
787 pr_debug("After apic_write\n");
794 /* Boot on the stack */
795 /* Kick the second */
796 apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
800 * Give the other CPU some time to accept the IPI.
802 if (init_udelay == 0)
807 pr_debug("Startup point 1\n");
809 pr_debug("Waiting for send to finish...\n");
810 send_status = safe_apic_wait_icr_idle();
813 * Give the other CPU some time to accept the IPI.
815 if (init_udelay == 0)
820 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
821 apic_write(APIC_ESR, 0);
822 accept_status = (apic_read(APIC_ESR) & 0xEF);
823 if (send_status || accept_status)
826 pr_debug("After Startup\n");
829 pr_err("APIC never delivered???\n");
831 pr_err("APIC delivery error (%lx)\n", accept_status);
833 return (send_status | accept_status);
836 /* reduce the number of lines printed when booting a large cpu count system */
837 static void announce_cpu(int cpu, int apicid)
839 static int current_node = -1;
840 int node = early_cpu_to_node(cpu);
841 static int width, node_width;
844 width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
847 node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
850 printk(KERN_INFO "x86: Booting SMP configuration:\n");
852 if (system_state < SYSTEM_RUNNING) {
853 if (node != current_node) {
854 if (current_node > (-1))
858 printk(KERN_INFO ".... node %*s#%d, CPUs: ",
859 node_width - num_digits(node), " ", node);
862 /* Add padding for the BSP */
864 pr_cont("%*s", width + 1, " ");
866 pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
869 pr_info("Booting Node %d Processor %d APIC 0x%x\n",
873 static int wakeup_cpu0_nmi(unsigned int cmd, struct pt_regs *regs)
877 cpu = smp_processor_id();
878 if (cpu == 0 && !cpu_online(cpu) && enable_start_cpu0)
885 * Wake up AP by INIT, INIT, STARTUP sequence.
887 * Instead of waiting for STARTUP after INITs, BSP will execute the BIOS
888 * boot-strap code which is not a desired behavior for waking up BSP. To
889 * void the boot-strap code, wake up CPU0 by NMI instead.
891 * This works to wake up soft offlined CPU0 only. If CPU0 is hard offlined
892 * (i.e. physically hot removed and then hot added), NMI won't wake it up.
893 * We'll change this code in the future to wake up hard offlined CPU0 if
894 * real platform and request are available.
897 wakeup_cpu_via_init_nmi(int cpu, unsigned long start_ip, int apicid,
898 int *cpu0_nmi_registered)
906 * Wake up AP by INIT, INIT, STARTUP sequence.
909 boot_error = wakeup_secondary_cpu_via_init(apicid, start_ip);
914 * Wake up BSP by nmi.
916 * Register a NMI handler to help wake up CPU0.
918 boot_error = register_nmi_handler(NMI_LOCAL,
919 wakeup_cpu0_nmi, 0, "wake_cpu0");
922 enable_start_cpu0 = 1;
923 *cpu0_nmi_registered = 1;
924 if (apic->dest_logical == APIC_DEST_LOGICAL)
925 id = cpu0_logical_apicid;
928 boot_error = wakeup_secondary_cpu_via_nmi(id, start_ip);
937 void common_cpu_up(unsigned int cpu, struct task_struct *idle)
939 /* Just in case we booted with a single CPU. */
940 alternatives_enable_smp();
942 per_cpu(current_task, cpu) = idle;
945 /* Stack for startup_32 can be just as for start_secondary onwards */
947 per_cpu(cpu_current_top_of_stack, cpu) = task_top_of_stack(idle);
949 initial_gs = per_cpu_offset(cpu);
954 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
955 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
956 * Returns zero if CPU booted OK, else error code from
957 * ->wakeup_secondary_cpu.
959 static int do_boot_cpu(int apicid, int cpu, struct task_struct *idle,
960 int *cpu0_nmi_registered)
962 volatile u32 *trampoline_status =
963 (volatile u32 *) __va(real_mode_header->trampoline_status);
964 /* start_ip had better be page-aligned! */
965 unsigned long start_ip = real_mode_header->trampoline_start;
967 unsigned long boot_error = 0;
968 unsigned long timeout;
970 idle->thread.sp = (unsigned long)task_pt_regs(idle);
971 early_gdt_descr.address = (unsigned long)get_cpu_gdt_rw(cpu);
972 initial_code = (unsigned long)start_secondary;
973 initial_stack = idle->thread.sp;
975 /* Enable the espfix hack for this CPU */
978 /* So we see what's up */
979 announce_cpu(cpu, apicid);
982 * This grunge runs the startup process for
983 * the targeted processor.
986 if (x86_platform.legacy.warm_reset) {
988 pr_debug("Setting warm reset code and vector.\n");
990 smpboot_setup_warm_reset_vector(start_ip);
992 * Be paranoid about clearing APIC errors.
994 if (APIC_INTEGRATED(boot_cpu_apic_version)) {
995 apic_write(APIC_ESR, 0);
1001 * AP might wait on cpu_callout_mask in cpu_init() with
1002 * cpu_initialized_mask set if previous attempt to online
1003 * it timed-out. Clear cpu_initialized_mask so that after
1004 * INIT/SIPI it could start with a clean state.
1006 cpumask_clear_cpu(cpu, cpu_initialized_mask);
1010 * Wake up a CPU in difference cases:
1011 * - Use the method in the APIC driver if it's defined
1013 * - Use an INIT boot APIC message for APs or NMI for BSP.
1015 if (apic->wakeup_secondary_cpu)
1016 boot_error = apic->wakeup_secondary_cpu(apicid, start_ip);
1018 boot_error = wakeup_cpu_via_init_nmi(cpu, start_ip, apicid,
1019 cpu0_nmi_registered);
1023 * Wait 10s total for first sign of life from AP
1026 timeout = jiffies + 10*HZ;
1027 while (time_before(jiffies, timeout)) {
1028 if (cpumask_test_cpu(cpu, cpu_initialized_mask)) {
1030 * Tell AP to proceed with initialization
1032 cpumask_set_cpu(cpu, cpu_callout_mask);
1042 * Wait till AP completes initial initialization
1044 while (!cpumask_test_cpu(cpu, cpu_callin_mask)) {
1046 * Allow other tasks to run while we wait for the
1047 * AP to come online. This also gives a chance
1048 * for the MTRR work(triggered by the AP coming online)
1049 * to be completed in the stop machine context.
1055 /* mark "stuck" area as not stuck */
1056 *trampoline_status = 0;
1058 if (x86_platform.legacy.warm_reset) {
1060 * Cleanup possible dangling ends...
1062 smpboot_restore_warm_reset_vector();
1068 int native_cpu_up(unsigned int cpu, struct task_struct *tidle)
1070 int apicid = apic->cpu_present_to_apicid(cpu);
1071 int cpu0_nmi_registered = 0;
1072 unsigned long flags;
1075 lockdep_assert_irqs_enabled();
1077 pr_debug("++++++++++++++++++++=_---CPU UP %u\n", cpu);
1079 if (apicid == BAD_APICID ||
1080 !physid_isset(apicid, phys_cpu_present_map) ||
1081 !apic->apic_id_valid(apicid)) {
1082 pr_err("%s: bad cpu %d\n", __func__, cpu);
1087 * Already booted CPU?
1089 if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
1090 pr_debug("do_boot_cpu %d Already started\n", cpu);
1095 * Save current MTRR state in case it was changed since early boot
1096 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1100 /* x86 CPUs take themselves offline, so delayed offline is OK. */
1101 err = cpu_check_up_prepare(cpu);
1102 if (err && err != -EBUSY)
1105 /* the FPU context is blank, nobody can own it */
1106 per_cpu(fpu_fpregs_owner_ctx, cpu) = NULL;
1108 common_cpu_up(cpu, tidle);
1110 err = do_boot_cpu(apicid, cpu, tidle, &cpu0_nmi_registered);
1112 pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
1118 * Check TSC synchronization with the AP (keep irqs disabled
1121 local_irq_save(flags);
1122 check_tsc_sync_source(cpu);
1123 local_irq_restore(flags);
1125 while (!cpu_online(cpu)) {
1127 touch_nmi_watchdog();
1132 * Clean up the nmi handler. Do this after the callin and callout sync
1133 * to avoid impact of possible long unregister time.
1135 if (cpu0_nmi_registered)
1136 unregister_nmi_handler(NMI_LOCAL, "wake_cpu0");
1142 * arch_disable_smp_support() - disables SMP support for x86 at runtime
1144 void arch_disable_smp_support(void)
1146 disable_ioapic_support();
1150 * Fall back to non SMP mode after errors.
1152 * RED-PEN audit/test this more. I bet there is more state messed up here.
1154 static __init void disable_smp(void)
1156 pr_info("SMP disabled\n");
1158 disable_ioapic_support();
1160 init_cpu_present(cpumask_of(0));
1161 init_cpu_possible(cpumask_of(0));
1163 if (smp_found_config)
1164 physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
1166 physid_set_mask_of_physid(0, &phys_cpu_present_map);
1167 cpumask_set_cpu(0, topology_sibling_cpumask(0));
1168 cpumask_set_cpu(0, topology_core_cpumask(0));
1172 * Various sanity checks.
1174 static void __init smp_sanity_check(void)
1178 #if !defined(CONFIG_X86_BIGSMP) && defined(CONFIG_X86_32)
1179 if (def_to_bigsmp && nr_cpu_ids > 8) {
1183 pr_warn("More than 8 CPUs detected - skipping them\n"
1184 "Use CONFIG_X86_BIGSMP\n");
1187 for_each_present_cpu(cpu) {
1189 set_cpu_present(cpu, false);
1194 for_each_possible_cpu(cpu) {
1196 set_cpu_possible(cpu, false);
1204 if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map)) {
1205 pr_warn("weird, boot CPU (#%d) not listed by the BIOS\n",
1206 hard_smp_processor_id());
1208 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1212 * Should not be necessary because the MP table should list the boot
1213 * CPU too, but we do it for the sake of robustness anyway.
1215 if (!apic->check_phys_apicid_present(boot_cpu_physical_apicid)) {
1216 pr_notice("weird, boot CPU (#%d) not listed by the BIOS\n",
1217 boot_cpu_physical_apicid);
1218 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1223 static void __init smp_cpu_index_default(void)
1226 struct cpuinfo_x86 *c;
1228 for_each_possible_cpu(i) {
1230 /* mark all to hotplug */
1231 c->cpu_index = nr_cpu_ids;
1235 static void __init smp_get_logical_apicid(void)
1238 cpu0_logical_apicid = apic_read(APIC_LDR);
1240 cpu0_logical_apicid = GET_APIC_LOGICAL_ID(apic_read(APIC_LDR));
1244 * Prepare for SMP bootup.
1245 * @max_cpus: configured maximum number of CPUs, It is a legacy parameter
1246 * for common interface support.
1248 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1252 smp_cpu_index_default();
1255 * Setup boot CPU information
1257 smp_store_boot_cpu_info(); /* Final full version of the data */
1258 cpumask_copy(cpu_callin_mask, cpumask_of(0));
1261 for_each_possible_cpu(i) {
1262 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1263 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1264 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1268 * Set 'default' x86 topology, this matches default_topology() in that
1269 * it has NUMA nodes as a topology level. See also
1270 * native_smp_cpus_done().
1272 * Must be done before set_cpus_sibling_map() is ran.
1274 set_sched_topology(x86_topology);
1276 set_cpu_sibling_map(0);
1280 switch (apic_intr_mode) {
1282 case APIC_VIRTUAL_WIRE_NO_CONFIG:
1285 case APIC_SYMMETRIC_IO_NO_ROUTING:
1287 /* Setup local timer */
1288 x86_init.timers.setup_percpu_clockev();
1290 case APIC_VIRTUAL_WIRE:
1291 case APIC_SYMMETRIC_IO:
1295 /* Setup local timer */
1296 x86_init.timers.setup_percpu_clockev();
1298 smp_get_logical_apicid();
1301 print_cpu_info(&cpu_data(0));
1303 native_pv_lock_init();
1307 set_mtrr_aps_delayed_init();
1309 smp_quirk_init_udelay();
1311 speculative_store_bypass_ht_init();
1314 void arch_enable_nonboot_cpus_begin(void)
1316 set_mtrr_aps_delayed_init();
1319 void arch_enable_nonboot_cpus_end(void)
1325 * Early setup to make printk work.
1327 void __init native_smp_prepare_boot_cpu(void)
1329 int me = smp_processor_id();
1330 switch_to_new_gdt(me);
1331 /* already set me in cpu_online_mask in boot_cpu_init() */
1332 cpumask_set_cpu(me, cpu_callout_mask);
1333 cpu_set_state_online(me);
1336 void __init calculate_max_logical_packages(void)
1341 * Today neither Intel nor AMD support heterogenous systems so
1342 * extrapolate the boot cpu's data to all packages.
1344 ncpus = cpu_data(0).booted_cores * topology_max_smt_threads();
1345 __max_logical_packages = DIV_ROUND_UP(total_cpus, ncpus);
1346 pr_info("Max logical packages: %u\n", __max_logical_packages);
1349 void __init native_smp_cpus_done(unsigned int max_cpus)
1351 pr_debug("Boot done\n");
1353 calculate_max_logical_packages();
1355 if (x86_has_numa_in_package)
1356 set_sched_topology(x86_numa_in_package_topology);
1363 static int __initdata setup_possible_cpus = -1;
1364 static int __init _setup_possible_cpus(char *str)
1366 get_option(&str, &setup_possible_cpus);
1369 early_param("possible_cpus", _setup_possible_cpus);
1373 * cpu_possible_mask should be static, it cannot change as cpu's
1374 * are onlined, or offlined. The reason is per-cpu data-structures
1375 * are allocated by some modules at init time, and dont expect to
1376 * do this dynamically on cpu arrival/departure.
1377 * cpu_present_mask on the other hand can change dynamically.
1378 * In case when cpu_hotplug is not compiled, then we resort to current
1379 * behaviour, which is cpu_possible == cpu_present.
1382 * Three ways to find out the number of additional hotplug CPUs:
1383 * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1384 * - The user can overwrite it with possible_cpus=NUM
1385 * - Otherwise don't reserve additional CPUs.
1386 * We do this because additional CPUs waste a lot of memory.
1389 __init void prefill_possible_map(void)
1393 /* No boot processor was found in mptable or ACPI MADT */
1394 if (!num_processors) {
1395 if (boot_cpu_has(X86_FEATURE_APIC)) {
1396 int apicid = boot_cpu_physical_apicid;
1397 int cpu = hard_smp_processor_id();
1399 pr_warn("Boot CPU (id %d) not listed by BIOS\n", cpu);
1401 /* Make sure boot cpu is enumerated */
1402 if (apic->cpu_present_to_apicid(0) == BAD_APICID &&
1403 apic->apic_id_valid(apicid))
1404 generic_processor_info(apicid, boot_cpu_apic_version);
1407 if (!num_processors)
1411 i = setup_max_cpus ?: 1;
1412 if (setup_possible_cpus == -1) {
1413 possible = num_processors;
1414 #ifdef CONFIG_HOTPLUG_CPU
1416 possible += disabled_cpus;
1422 possible = setup_possible_cpus;
1424 total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1426 /* nr_cpu_ids could be reduced via nr_cpus= */
1427 if (possible > nr_cpu_ids) {
1428 pr_warn("%d Processors exceeds NR_CPUS limit of %u\n",
1429 possible, nr_cpu_ids);
1430 possible = nr_cpu_ids;
1433 #ifdef CONFIG_HOTPLUG_CPU
1434 if (!setup_max_cpus)
1437 pr_warn("%d Processors exceeds max_cpus limit of %u\n",
1438 possible, setup_max_cpus);
1442 nr_cpu_ids = possible;
1444 pr_info("Allowing %d CPUs, %d hotplug CPUs\n",
1445 possible, max_t(int, possible - num_processors, 0));
1447 reset_cpu_possible_mask();
1449 for (i = 0; i < possible; i++)
1450 set_cpu_possible(i, true);
1453 #ifdef CONFIG_HOTPLUG_CPU
1455 /* Recompute SMT state for all CPUs on offline */
1456 static void recompute_smt_state(void)
1458 int max_threads, cpu;
1461 for_each_online_cpu (cpu) {
1462 int threads = cpumask_weight(topology_sibling_cpumask(cpu));
1464 if (threads > max_threads)
1465 max_threads = threads;
1467 __max_smt_threads = max_threads;
1470 static void remove_siblinginfo(int cpu)
1473 struct cpuinfo_x86 *c = &cpu_data(cpu);
1475 for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1476 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1478 * last thread sibling in this cpu core going down
1480 if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1481 cpu_data(sibling).booted_cores--;
1484 for_each_cpu(sibling, topology_sibling_cpumask(cpu))
1485 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1486 for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1487 cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1488 cpumask_clear(cpu_llc_shared_mask(cpu));
1489 cpumask_clear(topology_sibling_cpumask(cpu));
1490 cpumask_clear(topology_core_cpumask(cpu));
1492 c->booted_cores = 0;
1493 cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1494 recompute_smt_state();
1497 static void remove_cpu_from_maps(int cpu)
1499 set_cpu_online(cpu, false);
1500 cpumask_clear_cpu(cpu, cpu_callout_mask);
1501 cpumask_clear_cpu(cpu, cpu_callin_mask);
1502 /* was set by cpu_init() */
1503 cpumask_clear_cpu(cpu, cpu_initialized_mask);
1504 numa_remove_cpu(cpu);
1507 void cpu_disable_common(void)
1509 int cpu = smp_processor_id();
1511 remove_siblinginfo(cpu);
1513 /* It's now safe to remove this processor from the online map */
1515 remove_cpu_from_maps(cpu);
1516 unlock_vector_lock();
1521 int native_cpu_disable(void)
1525 ret = lapic_can_unplug_cpu();
1530 cpu_disable_common();
1535 int common_cpu_die(unsigned int cpu)
1539 /* We don't do anything here: idle task is faking death itself. */
1541 /* They ack this in play_dead() by setting CPU_DEAD */
1542 if (cpu_wait_death(cpu, 5)) {
1543 if (system_state == SYSTEM_RUNNING)
1544 pr_info("CPU %u is now offline\n", cpu);
1546 pr_err("CPU %u didn't die...\n", cpu);
1553 void native_cpu_die(unsigned int cpu)
1555 common_cpu_die(cpu);
1558 void play_dead_common(void)
1563 (void)cpu_report_death();
1566 * With physical CPU hotplug, we should halt the cpu
1568 local_irq_disable();
1571 static bool wakeup_cpu0(void)
1573 if (smp_processor_id() == 0 && enable_start_cpu0)
1580 * We need to flush the caches before going to sleep, lest we have
1581 * dirty data in our caches when we come back up.
1583 static inline void mwait_play_dead(void)
1585 unsigned int eax, ebx, ecx, edx;
1586 unsigned int highest_cstate = 0;
1587 unsigned int highest_subcstate = 0;
1591 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD ||
1592 boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
1594 if (!this_cpu_has(X86_FEATURE_MWAIT))
1596 if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1598 if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1601 eax = CPUID_MWAIT_LEAF;
1603 native_cpuid(&eax, &ebx, &ecx, &edx);
1606 * eax will be 0 if EDX enumeration is not valid.
1607 * Initialized below to cstate, sub_cstate value when EDX is valid.
1609 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1612 edx >>= MWAIT_SUBSTATE_SIZE;
1613 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1614 if (edx & MWAIT_SUBSTATE_MASK) {
1616 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1619 eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1620 (highest_subcstate - 1);
1624 * This should be a memory location in a cache line which is
1625 * unlikely to be touched by other processors. The actual
1626 * content is immaterial as it is not actually modified in any way.
1628 mwait_ptr = ¤t_thread_info()->flags;
1634 * The CLFLUSH is a workaround for erratum AAI65 for
1635 * the Xeon 7400 series. It's not clear it is actually
1636 * needed, but it should be harmless in either case.
1637 * The WBINVD is insufficient due to the spurious-wakeup
1638 * case where we return around the loop.
1643 __monitor(mwait_ptr, 0, 0);
1647 * If NMI wants to wake up CPU0, start CPU0.
1654 void hlt_play_dead(void)
1656 if (__this_cpu_read(cpu_info.x86) >= 4)
1662 * If NMI wants to wake up CPU0, start CPU0.
1669 void native_play_dead(void)
1672 tboot_shutdown(TB_SHUTDOWN_WFS);
1674 mwait_play_dead(); /* Only returns on failure */
1675 if (cpuidle_play_dead())
1679 #else /* ... !CONFIG_HOTPLUG_CPU */
1680 int native_cpu_disable(void)
1685 void native_cpu_die(unsigned int cpu)
1687 /* We said "no" in __cpu_disable */
1691 void native_play_dead(void)