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/module.h>
47 #include <linux/sched.h>
48 #include <linux/percpu.h>
49 #include <linux/bootmem.h>
50 #include <linux/err.h>
51 #include <linux/nmi.h>
52 #include <linux/tboot.h>
53 #include <linux/stackprotector.h>
54 #include <linux/gfp.h>
55 #include <linux/cpuidle.h>
62 #include <asm/realmode.h>
65 #include <asm/pgtable.h>
66 #include <asm/tlbflush.h>
68 #include <asm/mwait.h>
70 #include <asm/io_apic.h>
71 #include <asm/fpu/internal.h>
72 #include <asm/setup.h>
73 #include <asm/uv/uv.h>
74 #include <linux/mc146818rtc.h>
75 #include <asm/i8259.h>
76 #include <asm/realmode.h>
79 /* Number of siblings per CPU package */
80 int smp_num_siblings = 1;
81 EXPORT_SYMBOL(smp_num_siblings);
83 /* Last level cache ID of each logical CPU */
84 DEFINE_PER_CPU_READ_MOSTLY(u16, cpu_llc_id) = BAD_APICID;
86 /* representing HT siblings of each logical CPU */
87 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_sibling_map);
88 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
90 /* representing HT and core siblings of each logical CPU */
91 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_core_map);
92 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
94 DEFINE_PER_CPU_READ_MOSTLY(cpumask_var_t, cpu_llc_shared_map);
96 /* Per CPU bogomips and other parameters */
97 DEFINE_PER_CPU_READ_MOSTLY(struct cpuinfo_x86, cpu_info);
98 EXPORT_PER_CPU_SYMBOL(cpu_info);
100 atomic_t init_deasserted;
102 static inline void smpboot_setup_warm_reset_vector(unsigned long start_eip)
106 spin_lock_irqsave(&rtc_lock, flags);
107 CMOS_WRITE(0xa, 0xf);
108 spin_unlock_irqrestore(&rtc_lock, flags);
111 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_HIGH)) =
114 *((volatile unsigned short *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) =
119 static inline void smpboot_restore_warm_reset_vector(void)
124 * Install writable page 0 entry to set BIOS data area.
129 * Paranoid: Set warm reset code and vector here back
132 spin_lock_irqsave(&rtc_lock, flags);
134 spin_unlock_irqrestore(&rtc_lock, flags);
136 *((volatile u32 *)phys_to_virt(TRAMPOLINE_PHYS_LOW)) = 0;
140 * Report back to the Boot Processor during boot time or to the caller processor
143 static void smp_callin(void)
148 * If waken up by an INIT in an 82489DX configuration
149 * we may get here before an INIT-deassert IPI reaches
150 * our local APIC. We have to wait for the IPI or we'll
151 * lock up on an APIC access.
153 * Since CPU0 is not wakened up by INIT, it doesn't wait for the IPI.
155 cpuid = smp_processor_id();
156 if (apic->wait_for_init_deassert && cpuid)
157 while (!atomic_read(&init_deasserted))
161 * (This works even if the APIC is not enabled.)
163 phys_id = read_apic_id();
166 * the boot CPU has finished the init stage and is spinning
167 * on callin_map until we finish. We are free to set up this
168 * CPU, first the APIC. (this is probably redundant on most
174 * Need to setup vector mappings before we enable interrupts.
176 setup_vector_irq(smp_processor_id());
179 * Save our processor parameters. Note: this information
180 * is needed for clock calibration.
182 smp_store_cpu_info(cpuid);
186 * Update loops_per_jiffy in cpu_data. Previous call to
187 * smp_store_cpu_info() stored a value that is close but not as
188 * accurate as the value just calculated.
191 cpu_data(cpuid).loops_per_jiffy = loops_per_jiffy;
192 pr_debug("Stack at about %p\n", &cpuid);
195 * This must be done before setting cpu_online_mask
196 * or calling notify_cpu_starting.
198 set_cpu_sibling_map(raw_smp_processor_id());
201 notify_cpu_starting(cpuid);
204 * Allow the master to continue.
206 cpumask_set_cpu(cpuid, cpu_callin_mask);
209 static int cpu0_logical_apicid;
210 static int enable_start_cpu0;
212 * Activate a secondary processor.
214 static void notrace start_secondary(void *unused)
217 * Don't put *anything* before cpu_init(), SMP booting is too
218 * fragile that we want to limit the things done here to the
219 * most necessary things.
222 x86_cpuinit.early_percpu_clock_init();
226 enable_start_cpu0 = 0;
229 /* switch away from the initial page table */
230 load_cr3(swapper_pg_dir);
234 /* otherwise gcc will move up smp_processor_id before the cpu_init */
237 * Check TSC synchronization with the BP:
239 check_tsc_sync_target();
242 * Enable the espfix hack for this CPU
244 #ifdef CONFIG_X86_ESPFIX64
249 * We need to hold vector_lock so there the set of online cpus
250 * does not change while we are assigning vectors to cpus. Holding
251 * this lock ensures we don't half assign or remove an irq from a cpu.
254 set_cpu_online(smp_processor_id(), true);
255 unlock_vector_lock();
256 cpu_set_state_online(smp_processor_id());
257 x86_platform.nmi_init();
259 /* enable local interrupts */
262 /* to prevent fake stack check failure in clock setup */
263 boot_init_stack_canary();
265 x86_cpuinit.setup_percpu_clockev();
268 cpu_startup_entry(CPUHP_ONLINE);
271 void __init smp_store_boot_cpu_info(void)
273 int id = 0; /* CPU 0 */
274 struct cpuinfo_x86 *c = &cpu_data(id);
281 * The bootstrap kernel entry code has set these up. Save them for
284 void smp_store_cpu_info(int id)
286 struct cpuinfo_x86 *c = &cpu_data(id);
291 * During boot time, CPU0 has this setup already. Save the info when
292 * bringing up AP or offlined CPU0.
294 identify_secondary_cpu(c);
298 topology_same_node(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
300 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
302 return (cpu_to_node(cpu1) == cpu_to_node(cpu2));
306 topology_sane(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o, const char *name)
308 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
310 return !WARN_ONCE(!topology_same_node(c, o),
311 "sched: CPU #%d's %s-sibling CPU #%d is not on the same node! "
312 "[node: %d != %d]. Ignoring dependency.\n",
313 cpu1, name, cpu2, cpu_to_node(cpu1), cpu_to_node(cpu2));
316 #define link_mask(mfunc, c1, c2) \
318 cpumask_set_cpu((c1), mfunc(c2)); \
319 cpumask_set_cpu((c2), mfunc(c1)); \
322 static bool match_smt(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
324 if (cpu_has_topoext) {
325 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
327 if (c->phys_proc_id == o->phys_proc_id &&
328 per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2) &&
329 c->compute_unit_id == o->compute_unit_id)
330 return topology_sane(c, o, "smt");
332 } else if (c->phys_proc_id == o->phys_proc_id &&
333 c->cpu_core_id == o->cpu_core_id) {
334 return topology_sane(c, o, "smt");
340 static bool match_llc(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
342 int cpu1 = c->cpu_index, cpu2 = o->cpu_index;
344 if (per_cpu(cpu_llc_id, cpu1) != BAD_APICID &&
345 per_cpu(cpu_llc_id, cpu1) == per_cpu(cpu_llc_id, cpu2))
346 return topology_sane(c, o, "llc");
352 * Unlike the other levels, we do not enforce keeping a
353 * multicore group inside a NUMA node. If this happens, we will
354 * discard the MC level of the topology later.
356 static bool match_die(struct cpuinfo_x86 *c, struct cpuinfo_x86 *o)
358 if (c->phys_proc_id == o->phys_proc_id)
363 static struct sched_domain_topology_level numa_inside_package_topology[] = {
364 #ifdef CONFIG_SCHED_SMT
365 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) },
367 #ifdef CONFIG_SCHED_MC
368 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) },
373 * set_sched_topology() sets the topology internal to a CPU. The
374 * NUMA topologies are layered on top of it to build the full
377 * If NUMA nodes are observed to occur within a CPU package, this
378 * function should be called. It forces the sched domain code to
379 * only use the SMT level for the CPU portion of the topology.
380 * This essentially falls back to relying on NUMA information
381 * from the SRAT table to describe the entire system topology
382 * (except for hyperthreads).
384 static void primarily_use_numa_for_topology(void)
386 set_sched_topology(numa_inside_package_topology);
389 void set_cpu_sibling_map(int cpu)
391 bool has_smt = smp_num_siblings > 1;
392 bool has_mp = has_smt || boot_cpu_data.x86_max_cores > 1;
393 struct cpuinfo_x86 *c = &cpu_data(cpu);
394 struct cpuinfo_x86 *o;
397 cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
400 cpumask_set_cpu(cpu, topology_sibling_cpumask(cpu));
401 cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
402 cpumask_set_cpu(cpu, topology_core_cpumask(cpu));
407 for_each_cpu(i, cpu_sibling_setup_mask) {
410 if ((i == cpu) || (has_smt && match_smt(c, o)))
411 link_mask(topology_sibling_cpumask, cpu, i);
413 if ((i == cpu) || (has_mp && match_llc(c, o)))
414 link_mask(cpu_llc_shared_mask, cpu, i);
419 * This needs a separate iteration over the cpus because we rely on all
420 * topology_sibling_cpumask links to be set-up.
422 for_each_cpu(i, cpu_sibling_setup_mask) {
425 if ((i == cpu) || (has_mp && match_die(c, o))) {
426 link_mask(topology_core_cpumask, cpu, i);
429 * Does this new cpu bringup a new core?
432 topology_sibling_cpumask(cpu)) == 1) {
434 * for each core in package, increment
435 * the booted_cores for this new cpu
438 topology_sibling_cpumask(i)) == i)
441 * increment the core count for all
442 * the other cpus in this package
445 cpu_data(i).booted_cores++;
446 } else if (i != cpu && !c->booted_cores)
447 c->booted_cores = cpu_data(i).booted_cores;
449 if (match_die(c, o) && !topology_same_node(c, o))
450 primarily_use_numa_for_topology();
454 /* maps the cpu to the sched domain representing multi-core */
455 const struct cpumask *cpu_coregroup_mask(int cpu)
457 return cpu_llc_shared_mask(cpu);
460 static void impress_friends(void)
463 unsigned long bogosum = 0;
465 * Allow the user to impress friends.
467 pr_debug("Before bogomips\n");
468 for_each_possible_cpu(cpu)
469 if (cpumask_test_cpu(cpu, cpu_callout_mask))
470 bogosum += cpu_data(cpu).loops_per_jiffy;
471 pr_info("Total of %d processors activated (%lu.%02lu BogoMIPS)\n",
474 (bogosum/(5000/HZ))%100);
476 pr_debug("Before bogocount - setting activated=1\n");
479 void __inquire_remote_apic(int apicid)
481 unsigned i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
482 const char * const names[] = { "ID", "VERSION", "SPIV" };
486 pr_info("Inquiring remote APIC 0x%x...\n", apicid);
488 for (i = 0; i < ARRAY_SIZE(regs); i++) {
489 pr_info("... APIC 0x%x %s: ", apicid, names[i]);
494 status = safe_apic_wait_icr_idle();
496 pr_cont("a previous APIC delivery may have failed\n");
498 apic_icr_write(APIC_DM_REMRD | regs[i], apicid);
503 status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
504 } while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
507 case APIC_ICR_RR_VALID:
508 status = apic_read(APIC_RRR);
509 pr_cont("%08x\n", status);
518 * The Multiprocessor Specification 1.4 (1997) example code suggests
519 * that there should be a 10ms delay between the BSP asserting INIT
520 * and de-asserting INIT, when starting a remote processor.
521 * But that slows boot and resume on modern processors, which include
522 * many cores and don't require that delay.
524 * Cmdline "init_cpu_udelay=" is available to over-ride this delay.
525 * Modern processor families are quirked to remove the delay entirely.
527 #define UDELAY_10MS_DEFAULT 10000
529 static unsigned int init_udelay = UDELAY_10MS_DEFAULT;
531 static int __init cpu_init_udelay(char *str)
533 get_option(&str, &init_udelay);
537 early_param("cpu_init_udelay", cpu_init_udelay);
539 static void __init smp_quirk_init_udelay(void)
541 /* if cmdline changed it from default, leave it alone */
542 if (init_udelay != UDELAY_10MS_DEFAULT)
545 /* if modern processor, use no delay */
546 if (((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL) && (boot_cpu_data.x86 == 6)) ||
547 ((boot_cpu_data.x86_vendor == X86_VENDOR_AMD) && (boot_cpu_data.x86 >= 0xF)))
552 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
553 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
554 * won't ... remember to clear down the APIC, etc later.
557 wakeup_secondary_cpu_via_nmi(int apicid, unsigned long start_eip)
559 unsigned long send_status, accept_status = 0;
563 /* Boot on the stack */
564 /* Kick the second */
565 apic_icr_write(APIC_DM_NMI | apic->dest_logical, apicid);
567 pr_debug("Waiting for send to finish...\n");
568 send_status = safe_apic_wait_icr_idle();
571 * Give the other CPU some time to accept the IPI.
574 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid])) {
575 maxlvt = lapic_get_maxlvt();
576 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
577 apic_write(APIC_ESR, 0);
578 accept_status = (apic_read(APIC_ESR) & 0xEF);
580 pr_debug("NMI sent\n");
583 pr_err("APIC never delivered???\n");
585 pr_err("APIC delivery error (%lx)\n", accept_status);
587 return (send_status | accept_status);
591 wakeup_secondary_cpu_via_init(int phys_apicid, unsigned long start_eip)
593 unsigned long send_status = 0, accept_status = 0;
594 int maxlvt, num_starts, j;
596 maxlvt = lapic_get_maxlvt();
599 * Be paranoid about clearing APIC errors.
601 if (APIC_INTEGRATED(apic_version[phys_apicid])) {
602 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
603 apic_write(APIC_ESR, 0);
607 pr_debug("Asserting INIT\n");
610 * Turn INIT on target chip
615 apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT,
618 pr_debug("Waiting for send to finish...\n");
619 send_status = safe_apic_wait_icr_idle();
623 pr_debug("Deasserting INIT\n");
627 apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
629 pr_debug("Waiting for send to finish...\n");
630 send_status = safe_apic_wait_icr_idle();
633 atomic_set(&init_deasserted, 1);
636 * Should we send STARTUP IPIs ?
638 * Determine this based on the APIC version.
639 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
641 if (APIC_INTEGRATED(apic_version[phys_apicid]))
647 * Paravirt / VMI wants a startup IPI hook here to set up the
648 * target processor state.
650 startup_ipi_hook(phys_apicid, (unsigned long) start_secondary,
654 * Run STARTUP IPI loop.
656 pr_debug("#startup loops: %d\n", num_starts);
658 for (j = 1; j <= num_starts; j++) {
659 pr_debug("Sending STARTUP #%d\n", j);
660 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
661 apic_write(APIC_ESR, 0);
663 pr_debug("After apic_write\n");
670 /* Boot on the stack */
671 /* Kick the second */
672 apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
676 * Give the other CPU some time to accept the IPI.
680 pr_debug("Startup point 1\n");
682 pr_debug("Waiting for send to finish...\n");
683 send_status = safe_apic_wait_icr_idle();
686 * Give the other CPU some time to accept the IPI.
690 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
691 apic_write(APIC_ESR, 0);
692 accept_status = (apic_read(APIC_ESR) & 0xEF);
693 if (send_status || accept_status)
696 pr_debug("After Startup\n");
699 pr_err("APIC never delivered???\n");
701 pr_err("APIC delivery error (%lx)\n", accept_status);
703 return (send_status | accept_status);
706 void smp_announce(void)
708 int num_nodes = num_online_nodes();
710 printk(KERN_INFO "x86: Booted up %d node%s, %d CPUs\n",
711 num_nodes, (num_nodes > 1 ? "s" : ""), num_online_cpus());
714 /* reduce the number of lines printed when booting a large cpu count system */
715 static void announce_cpu(int cpu, int apicid)
717 static int current_node = -1;
718 int node = early_cpu_to_node(cpu);
719 static int width, node_width;
722 width = num_digits(num_possible_cpus()) + 1; /* + '#' sign */
725 node_width = num_digits(num_possible_nodes()) + 1; /* + '#' */
728 printk(KERN_INFO "x86: Booting SMP configuration:\n");
730 if (system_state == SYSTEM_BOOTING) {
731 if (node != current_node) {
732 if (current_node > (-1))
736 printk(KERN_INFO ".... node %*s#%d, CPUs: ",
737 node_width - num_digits(node), " ", node);
740 /* Add padding for the BSP */
742 pr_cont("%*s", width + 1, " ");
744 pr_cont("%*s#%d", width - num_digits(cpu), " ", cpu);
747 pr_info("Booting Node %d Processor %d APIC 0x%x\n",
751 static int wakeup_cpu0_nmi(unsigned int cmd, struct pt_regs *regs)
755 cpu = smp_processor_id();
756 if (cpu == 0 && !cpu_online(cpu) && enable_start_cpu0)
763 * Wake up AP by INIT, INIT, STARTUP sequence.
765 * Instead of waiting for STARTUP after INITs, BSP will execute the BIOS
766 * boot-strap code which is not a desired behavior for waking up BSP. To
767 * void the boot-strap code, wake up CPU0 by NMI instead.
769 * This works to wake up soft offlined CPU0 only. If CPU0 is hard offlined
770 * (i.e. physically hot removed and then hot added), NMI won't wake it up.
771 * We'll change this code in the future to wake up hard offlined CPU0 if
772 * real platform and request are available.
775 wakeup_cpu_via_init_nmi(int cpu, unsigned long start_ip, int apicid,
776 int *cpu0_nmi_registered)
784 * Wake up AP by INIT, INIT, STARTUP sequence.
787 boot_error = wakeup_secondary_cpu_via_init(apicid, start_ip);
792 * Wake up BSP by nmi.
794 * Register a NMI handler to help wake up CPU0.
796 boot_error = register_nmi_handler(NMI_LOCAL,
797 wakeup_cpu0_nmi, 0, "wake_cpu0");
800 enable_start_cpu0 = 1;
801 *cpu0_nmi_registered = 1;
802 if (apic->dest_logical == APIC_DEST_LOGICAL)
803 id = cpu0_logical_apicid;
806 boot_error = wakeup_secondary_cpu_via_nmi(id, start_ip);
815 void common_cpu_up(unsigned int cpu, struct task_struct *idle)
817 /* Just in case we booted with a single CPU. */
818 alternatives_enable_smp();
820 per_cpu(current_task, cpu) = idle;
823 /* Stack for startup_32 can be just as for start_secondary onwards */
825 per_cpu(cpu_current_top_of_stack, cpu) =
826 (unsigned long)task_stack_page(idle) + THREAD_SIZE;
828 clear_tsk_thread_flag(idle, TIF_FORK);
829 initial_gs = per_cpu_offset(cpu);
834 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
835 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
836 * Returns zero if CPU booted OK, else error code from
837 * ->wakeup_secondary_cpu.
839 static int do_boot_cpu(int apicid, int cpu, struct task_struct *idle)
841 volatile u32 *trampoline_status =
842 (volatile u32 *) __va(real_mode_header->trampoline_status);
843 /* start_ip had better be page-aligned! */
844 unsigned long start_ip = real_mode_header->trampoline_start;
846 unsigned long boot_error = 0;
847 int cpu0_nmi_registered = 0;
848 unsigned long timeout;
850 idle->thread.sp = (unsigned long) (((struct pt_regs *)
851 (THREAD_SIZE + task_stack_page(idle))) - 1);
853 early_gdt_descr.address = (unsigned long)get_cpu_gdt_table(cpu);
854 initial_code = (unsigned long)start_secondary;
855 stack_start = idle->thread.sp;
857 /* So we see what's up */
858 announce_cpu(cpu, apicid);
861 * This grunge runs the startup process for
862 * the targeted processor.
865 atomic_set(&init_deasserted, 0);
867 if (get_uv_system_type() != UV_NON_UNIQUE_APIC) {
869 pr_debug("Setting warm reset code and vector.\n");
871 smpboot_setup_warm_reset_vector(start_ip);
873 * Be paranoid about clearing APIC errors.
875 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid])) {
876 apic_write(APIC_ESR, 0);
882 * AP might wait on cpu_callout_mask in cpu_init() with
883 * cpu_initialized_mask set if previous attempt to online
884 * it timed-out. Clear cpu_initialized_mask so that after
885 * INIT/SIPI it could start with a clean state.
887 cpumask_clear_cpu(cpu, cpu_initialized_mask);
891 * Wake up a CPU in difference cases:
892 * - Use the method in the APIC driver if it's defined
894 * - Use an INIT boot APIC message for APs or NMI for BSP.
896 if (apic->wakeup_secondary_cpu)
897 boot_error = apic->wakeup_secondary_cpu(apicid, start_ip);
899 boot_error = wakeup_cpu_via_init_nmi(cpu, start_ip, apicid,
900 &cpu0_nmi_registered);
904 * Wait 10s total for a response from AP
907 timeout = jiffies + 10*HZ;
908 while (time_before(jiffies, timeout)) {
909 if (cpumask_test_cpu(cpu, cpu_initialized_mask)) {
911 * Tell AP to proceed with initialization
913 cpumask_set_cpu(cpu, cpu_callout_mask);
924 * Wait till AP completes initial initialization
926 while (!cpumask_test_cpu(cpu, cpu_callin_mask)) {
928 * Allow other tasks to run while we wait for the
929 * AP to come online. This also gives a chance
930 * for the MTRR work(triggered by the AP coming online)
931 * to be completed in the stop machine context.
938 /* mark "stuck" area as not stuck */
939 *trampoline_status = 0;
941 if (get_uv_system_type() != UV_NON_UNIQUE_APIC) {
943 * Cleanup possible dangling ends...
945 smpboot_restore_warm_reset_vector();
948 * Clean up the nmi handler. Do this after the callin and callout sync
949 * to avoid impact of possible long unregister time.
951 if (cpu0_nmi_registered)
952 unregister_nmi_handler(NMI_LOCAL, "wake_cpu0");
957 int native_cpu_up(unsigned int cpu, struct task_struct *tidle)
959 int apicid = apic->cpu_present_to_apicid(cpu);
963 WARN_ON(irqs_disabled());
965 pr_debug("++++++++++++++++++++=_---CPU UP %u\n", cpu);
967 if (apicid == BAD_APICID ||
968 !physid_isset(apicid, phys_cpu_present_map) ||
969 !apic->apic_id_valid(apicid)) {
970 pr_err("%s: bad cpu %d\n", __func__, cpu);
975 * Already booted CPU?
977 if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
978 pr_debug("do_boot_cpu %d Already started\n", cpu);
983 * Save current MTRR state in case it was changed since early boot
984 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
988 /* x86 CPUs take themselves offline, so delayed offline is OK. */
989 err = cpu_check_up_prepare(cpu);
990 if (err && err != -EBUSY)
993 /* the FPU context is blank, nobody can own it */
994 __cpu_disable_lazy_restore(cpu);
996 common_cpu_up(cpu, tidle);
998 err = do_boot_cpu(apicid, cpu, tidle);
1000 pr_err("do_boot_cpu failed(%d) to wakeup CPU#%u\n", err, cpu);
1005 * Check TSC synchronization with the AP (keep irqs disabled
1008 local_irq_save(flags);
1009 check_tsc_sync_source(cpu);
1010 local_irq_restore(flags);
1012 while (!cpu_online(cpu)) {
1014 touch_nmi_watchdog();
1021 * arch_disable_smp_support() - disables SMP support for x86 at runtime
1023 void arch_disable_smp_support(void)
1025 disable_ioapic_support();
1029 * Fall back to non SMP mode after errors.
1031 * RED-PEN audit/test this more. I bet there is more state messed up here.
1033 static __init void disable_smp(void)
1035 pr_info("SMP disabled\n");
1037 disable_ioapic_support();
1039 init_cpu_present(cpumask_of(0));
1040 init_cpu_possible(cpumask_of(0));
1042 if (smp_found_config)
1043 physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
1045 physid_set_mask_of_physid(0, &phys_cpu_present_map);
1046 cpumask_set_cpu(0, topology_sibling_cpumask(0));
1047 cpumask_set_cpu(0, topology_core_cpumask(0));
1058 * Various sanity checks.
1060 static int __init smp_sanity_check(unsigned max_cpus)
1064 #if !defined(CONFIG_X86_BIGSMP) && defined(CONFIG_X86_32)
1065 if (def_to_bigsmp && nr_cpu_ids > 8) {
1069 pr_warn("More than 8 CPUs detected - skipping them\n"
1070 "Use CONFIG_X86_BIGSMP\n");
1073 for_each_present_cpu(cpu) {
1075 set_cpu_present(cpu, false);
1080 for_each_possible_cpu(cpu) {
1082 set_cpu_possible(cpu, false);
1090 if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map)) {
1091 pr_warn("weird, boot CPU (#%d) not listed by the BIOS\n",
1092 hard_smp_processor_id());
1094 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1098 * If we couldn't find an SMP configuration at boot time,
1099 * get out of here now!
1101 if (!smp_found_config && !acpi_lapic) {
1103 pr_notice("SMP motherboard not detected\n");
1104 return SMP_NO_CONFIG;
1108 * Should not be necessary because the MP table should list the boot
1109 * CPU too, but we do it for the sake of robustness anyway.
1111 if (!apic->check_phys_apicid_present(boot_cpu_physical_apicid)) {
1112 pr_notice("weird, boot CPU (#%d) not listed by the BIOS\n",
1113 boot_cpu_physical_apicid);
1114 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1119 * If we couldn't find a local APIC, then get out of here now!
1121 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid]) &&
1123 if (!disable_apic) {
1124 pr_err("BIOS bug, local APIC #%d not detected!...\n",
1125 boot_cpu_physical_apicid);
1126 pr_err("... forcing use of dummy APIC emulation (tell your hw vendor)\n");
1132 * If SMP should be disabled, then really disable it!
1135 pr_info("SMP mode deactivated\n");
1136 return SMP_FORCE_UP;
1142 static void __init smp_cpu_index_default(void)
1145 struct cpuinfo_x86 *c;
1147 for_each_possible_cpu(i) {
1149 /* mark all to hotplug */
1150 c->cpu_index = nr_cpu_ids;
1155 * Prepare for SMP bootup. The MP table or ACPI has been read
1156 * earlier. Just do some sanity checking here and enable APIC mode.
1158 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1162 smp_cpu_index_default();
1165 * Setup boot CPU information
1167 smp_store_boot_cpu_info(); /* Final full version of the data */
1168 cpumask_copy(cpu_callin_mask, cpumask_of(0));
1171 current_thread_info()->cpu = 0; /* needed? */
1172 for_each_possible_cpu(i) {
1173 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1174 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1175 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1177 set_cpu_sibling_map(0);
1179 switch (smp_sanity_check(max_cpus)) {
1182 if (APIC_init_uniprocessor())
1183 pr_notice("Local APIC not detected. Using dummy APIC emulation.\n");
1190 apic_bsp_setup(false);
1196 default_setup_apic_routing();
1198 if (read_apic_id() != boot_cpu_physical_apicid) {
1199 panic("Boot APIC ID in local APIC unexpected (%d vs %d)",
1200 read_apic_id(), boot_cpu_physical_apicid);
1201 /* Or can we switch back to PIC here? */
1204 cpu0_logical_apicid = apic_bsp_setup(false);
1206 pr_info("CPU%d: ", 0);
1207 print_cpu_info(&cpu_data(0));
1212 set_mtrr_aps_delayed_init();
1214 smp_quirk_init_udelay();
1217 void arch_enable_nonboot_cpus_begin(void)
1219 set_mtrr_aps_delayed_init();
1222 void arch_enable_nonboot_cpus_end(void)
1228 * Early setup to make printk work.
1230 void __init native_smp_prepare_boot_cpu(void)
1232 int me = smp_processor_id();
1233 switch_to_new_gdt(me);
1234 /* already set me in cpu_online_mask in boot_cpu_init() */
1235 cpumask_set_cpu(me, cpu_callout_mask);
1236 cpu_set_state_online(me);
1239 void __init native_smp_cpus_done(unsigned int max_cpus)
1241 pr_debug("Boot done\n");
1245 setup_ioapic_dest();
1249 static int __initdata setup_possible_cpus = -1;
1250 static int __init _setup_possible_cpus(char *str)
1252 get_option(&str, &setup_possible_cpus);
1255 early_param("possible_cpus", _setup_possible_cpus);
1259 * cpu_possible_mask should be static, it cannot change as cpu's
1260 * are onlined, or offlined. The reason is per-cpu data-structures
1261 * are allocated by some modules at init time, and dont expect to
1262 * do this dynamically on cpu arrival/departure.
1263 * cpu_present_mask on the other hand can change dynamically.
1264 * In case when cpu_hotplug is not compiled, then we resort to current
1265 * behaviour, which is cpu_possible == cpu_present.
1268 * Three ways to find out the number of additional hotplug CPUs:
1269 * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1270 * - The user can overwrite it with possible_cpus=NUM
1271 * - Otherwise don't reserve additional CPUs.
1272 * We do this because additional CPUs waste a lot of memory.
1275 __init void prefill_possible_map(void)
1279 /* no processor from mptable or madt */
1280 if (!num_processors)
1283 i = setup_max_cpus ?: 1;
1284 if (setup_possible_cpus == -1) {
1285 possible = num_processors;
1286 #ifdef CONFIG_HOTPLUG_CPU
1288 possible += disabled_cpus;
1294 possible = setup_possible_cpus;
1296 total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1298 /* nr_cpu_ids could be reduced via nr_cpus= */
1299 if (possible > nr_cpu_ids) {
1300 pr_warn("%d Processors exceeds NR_CPUS limit of %d\n",
1301 possible, nr_cpu_ids);
1302 possible = nr_cpu_ids;
1305 #ifdef CONFIG_HOTPLUG_CPU
1306 if (!setup_max_cpus)
1309 pr_warn("%d Processors exceeds max_cpus limit of %u\n",
1310 possible, setup_max_cpus);
1314 pr_info("Allowing %d CPUs, %d hotplug CPUs\n",
1315 possible, max_t(int, possible - num_processors, 0));
1317 for (i = 0; i < possible; i++)
1318 set_cpu_possible(i, true);
1319 for (; i < NR_CPUS; i++)
1320 set_cpu_possible(i, false);
1322 nr_cpu_ids = possible;
1325 #ifdef CONFIG_HOTPLUG_CPU
1327 static void remove_siblinginfo(int cpu)
1330 struct cpuinfo_x86 *c = &cpu_data(cpu);
1332 for_each_cpu(sibling, topology_core_cpumask(cpu)) {
1333 cpumask_clear_cpu(cpu, topology_core_cpumask(sibling));
1335 * last thread sibling in this cpu core going down
1337 if (cpumask_weight(topology_sibling_cpumask(cpu)) == 1)
1338 cpu_data(sibling).booted_cores--;
1341 for_each_cpu(sibling, topology_sibling_cpumask(cpu))
1342 cpumask_clear_cpu(cpu, topology_sibling_cpumask(sibling));
1343 for_each_cpu(sibling, cpu_llc_shared_mask(cpu))
1344 cpumask_clear_cpu(cpu, cpu_llc_shared_mask(sibling));
1345 cpumask_clear(cpu_llc_shared_mask(cpu));
1346 cpumask_clear(topology_sibling_cpumask(cpu));
1347 cpumask_clear(topology_core_cpumask(cpu));
1348 c->phys_proc_id = 0;
1350 cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1353 static void __ref remove_cpu_from_maps(int cpu)
1355 set_cpu_online(cpu, false);
1356 cpumask_clear_cpu(cpu, cpu_callout_mask);
1357 cpumask_clear_cpu(cpu, cpu_callin_mask);
1358 /* was set by cpu_init() */
1359 cpumask_clear_cpu(cpu, cpu_initialized_mask);
1360 numa_remove_cpu(cpu);
1363 void cpu_disable_common(void)
1365 int cpu = smp_processor_id();
1367 remove_siblinginfo(cpu);
1369 /* It's now safe to remove this processor from the online map */
1371 remove_cpu_from_maps(cpu);
1372 unlock_vector_lock();
1376 int native_cpu_disable(void)
1380 ret = check_irq_vectors_for_cpu_disable();
1385 cpu_disable_common();
1390 int common_cpu_die(unsigned int cpu)
1394 /* We don't do anything here: idle task is faking death itself. */
1396 /* They ack this in play_dead() by setting CPU_DEAD */
1397 if (cpu_wait_death(cpu, 5)) {
1398 if (system_state == SYSTEM_RUNNING)
1399 pr_info("CPU %u is now offline\n", cpu);
1401 pr_err("CPU %u didn't die...\n", cpu);
1408 void native_cpu_die(unsigned int cpu)
1410 common_cpu_die(cpu);
1413 void play_dead_common(void)
1416 reset_lazy_tlbstate();
1417 amd_e400_remove_cpu(raw_smp_processor_id());
1420 (void)cpu_report_death();
1423 * With physical CPU hotplug, we should halt the cpu
1425 local_irq_disable();
1428 static bool wakeup_cpu0(void)
1430 if (smp_processor_id() == 0 && enable_start_cpu0)
1437 * We need to flush the caches before going to sleep, lest we have
1438 * dirty data in our caches when we come back up.
1440 static inline void mwait_play_dead(void)
1442 unsigned int eax, ebx, ecx, edx;
1443 unsigned int highest_cstate = 0;
1444 unsigned int highest_subcstate = 0;
1448 if (!this_cpu_has(X86_FEATURE_MWAIT))
1450 if (!this_cpu_has(X86_FEATURE_CLFLUSH))
1452 if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1455 eax = CPUID_MWAIT_LEAF;
1457 native_cpuid(&eax, &ebx, &ecx, &edx);
1460 * eax will be 0 if EDX enumeration is not valid.
1461 * Initialized below to cstate, sub_cstate value when EDX is valid.
1463 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1466 edx >>= MWAIT_SUBSTATE_SIZE;
1467 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1468 if (edx & MWAIT_SUBSTATE_MASK) {
1470 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1473 eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1474 (highest_subcstate - 1);
1478 * This should be a memory location in a cache line which is
1479 * unlikely to be touched by other processors. The actual
1480 * content is immaterial as it is not actually modified in any way.
1482 mwait_ptr = ¤t_thread_info()->flags;
1488 * The CLFLUSH is a workaround for erratum AAI65 for
1489 * the Xeon 7400 series. It's not clear it is actually
1490 * needed, but it should be harmless in either case.
1491 * The WBINVD is insufficient due to the spurious-wakeup
1492 * case where we return around the loop.
1497 __monitor(mwait_ptr, 0, 0);
1501 * If NMI wants to wake up CPU0, start CPU0.
1508 static inline void hlt_play_dead(void)
1510 if (__this_cpu_read(cpu_info.x86) >= 4)
1516 * If NMI wants to wake up CPU0, start CPU0.
1523 void native_play_dead(void)
1526 tboot_shutdown(TB_SHUTDOWN_WFS);
1528 mwait_play_dead(); /* Only returns on failure */
1529 if (cpuidle_play_dead())
1533 #else /* ... !CONFIG_HOTPLUG_CPU */
1534 int native_cpu_disable(void)
1539 void native_cpu_die(unsigned int cpu)
1541 /* We said "no" in __cpu_disable */
1545 void native_play_dead(void)