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 #include <linux/init.h>
43 #include <linux/smp.h>
44 #include <linux/module.h>
45 #include <linux/sched.h>
46 #include <linux/percpu.h>
47 #include <linux/bootmem.h>
48 #include <linux/err.h>
49 #include <linux/nmi.h>
50 #include <linux/tboot.h>
51 #include <linux/stackprotector.h>
52 #include <linux/gfp.h>
53 #include <linux/cpuidle.h>
60 #include <asm/trampoline.h>
63 #include <asm/pgtable.h>
64 #include <asm/tlbflush.h>
66 #include <asm/mwait.h>
68 #include <asm/io_apic.h>
70 #include <asm/fpu-internal.h>
71 #include <asm/setup.h>
72 #include <asm/uv/uv.h>
73 #include <linux/mc146818rtc.h>
75 #include <asm/smpboot_hooks.h>
76 #include <asm/i8259.h>
78 /* State of each CPU */
79 DEFINE_PER_CPU(int, cpu_state) = { 0 };
81 /* Store all idle threads, this can be reused instead of creating
82 * a new thread. Also avoids complicated thread destroy functionality
85 #ifdef CONFIG_HOTPLUG_CPU
87 * Needed only for CONFIG_HOTPLUG_CPU because __cpuinitdata is
88 * removed after init for !CONFIG_HOTPLUG_CPU.
90 static DEFINE_PER_CPU(struct task_struct *, idle_thread_array);
91 #define get_idle_for_cpu(x) (per_cpu(idle_thread_array, x))
92 #define set_idle_for_cpu(x, p) (per_cpu(idle_thread_array, x) = (p))
95 * We need this for trampoline_base protection from concurrent accesses when
96 * off- and onlining cores wildly.
98 static DEFINE_MUTEX(x86_cpu_hotplug_driver_mutex);
100 void cpu_hotplug_driver_lock(void)
102 mutex_lock(&x86_cpu_hotplug_driver_mutex);
105 void cpu_hotplug_driver_unlock(void)
107 mutex_unlock(&x86_cpu_hotplug_driver_mutex);
110 ssize_t arch_cpu_probe(const char *buf, size_t count) { return -1; }
111 ssize_t arch_cpu_release(const char *buf, size_t count) { return -1; }
113 static struct task_struct *idle_thread_array[NR_CPUS] __cpuinitdata ;
114 #define get_idle_for_cpu(x) (idle_thread_array[(x)])
115 #define set_idle_for_cpu(x, p) (idle_thread_array[(x)] = (p))
118 /* Number of siblings per CPU package */
119 int smp_num_siblings = 1;
120 EXPORT_SYMBOL(smp_num_siblings);
122 /* Last level cache ID of each logical CPU */
123 DEFINE_PER_CPU(u16, cpu_llc_id) = BAD_APICID;
125 /* representing HT siblings of each logical CPU */
126 DEFINE_PER_CPU(cpumask_var_t, cpu_sibling_map);
127 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
129 /* representing HT and core siblings of each logical CPU */
130 DEFINE_PER_CPU(cpumask_var_t, cpu_core_map);
131 EXPORT_PER_CPU_SYMBOL(cpu_core_map);
133 DEFINE_PER_CPU(cpumask_var_t, cpu_llc_shared_map);
135 /* Per CPU bogomips and other parameters */
136 DEFINE_PER_CPU_SHARED_ALIGNED(struct cpuinfo_x86, cpu_info);
137 EXPORT_PER_CPU_SYMBOL(cpu_info);
139 atomic_t init_deasserted;
142 * Report back to the Boot Processor.
145 static void __cpuinit smp_callin(void)
148 unsigned long timeout;
151 * If waken up by an INIT in an 82489DX configuration
152 * we may get here before an INIT-deassert IPI reaches
153 * our local APIC. We have to wait for the IPI or we'll
154 * lock up on an APIC access.
156 if (apic->wait_for_init_deassert)
157 apic->wait_for_init_deassert(&init_deasserted);
160 * (This works even if the APIC is not enabled.)
162 phys_id = read_apic_id();
163 cpuid = smp_processor_id();
164 if (cpumask_test_cpu(cpuid, cpu_callin_mask)) {
165 panic("%s: phys CPU#%d, CPU#%d already present??\n", __func__,
168 pr_debug("CPU#%d (phys ID: %d) waiting for CALLOUT\n", cpuid, phys_id);
171 * STARTUP IPIs are fragile beasts as they might sometimes
172 * trigger some glue motherboard logic. Complete APIC bus
173 * silence for 1 second, this overestimates the time the
174 * boot CPU is spending to send the up to 2 STARTUP IPIs
175 * by a factor of two. This should be enough.
179 * Waiting 2s total for startup (udelay is not yet working)
181 timeout = jiffies + 2*HZ;
182 while (time_before(jiffies, timeout)) {
184 * Has the boot CPU finished it's STARTUP sequence?
186 if (cpumask_test_cpu(cpuid, cpu_callout_mask))
191 if (!time_before(jiffies, timeout)) {
192 panic("%s: CPU%d started up but did not get a callout!\n",
197 * the boot CPU has finished the init stage and is spinning
198 * on callin_map until we finish. We are free to set up this
199 * CPU, first the APIC. (this is probably redundant on most
203 pr_debug("CALLIN, before setup_local_APIC().\n");
204 if (apic->smp_callin_clear_local_apic)
205 apic->smp_callin_clear_local_apic();
207 end_local_APIC_setup();
210 * Need to setup vector mappings before we enable interrupts.
212 setup_vector_irq(smp_processor_id());
215 * Save our processor parameters. Note: this information
216 * is needed for clock calibration.
218 smp_store_cpu_info(cpuid);
222 * Update loops_per_jiffy in cpu_data. Previous call to
223 * smp_store_cpu_info() stored a value that is close but not as
224 * accurate as the value just calculated.
227 cpu_data(cpuid).loops_per_jiffy = loops_per_jiffy;
228 pr_debug("Stack at about %p\n", &cpuid);
231 * This must be done before setting cpu_online_mask
232 * or calling notify_cpu_starting.
234 set_cpu_sibling_map(raw_smp_processor_id());
237 notify_cpu_starting(cpuid);
240 * Allow the master to continue.
242 cpumask_set_cpu(cpuid, cpu_callin_mask);
246 * Activate a secondary processor.
248 notrace static void __cpuinit start_secondary(void *unused)
251 * Don't put *anything* before cpu_init(), SMP booting is too
252 * fragile that we want to limit the things done here to the
253 * most necessary things.
256 x86_cpuinit.early_percpu_clock_init();
261 /* switch away from the initial page table */
262 load_cr3(swapper_pg_dir);
266 /* otherwise gcc will move up smp_processor_id before the cpu_init */
269 * Check TSC synchronization with the BP:
271 check_tsc_sync_target();
274 * Enable the espfix hack for this CPU
281 * We need to hold call_lock, so there is no inconsistency
282 * between the time smp_call_function() determines number of
283 * IPI recipients, and the time when the determination is made
284 * for which cpus receive the IPI. Holding this
285 * lock helps us to not include this cpu in a currently in progress
286 * smp_call_function().
288 * We need to hold vector_lock so there the set of online cpus
289 * does not change while we are assigning vectors to cpus. Holding
290 * this lock ensures we don't half assign or remove an irq from a cpu.
294 set_cpu_online(smp_processor_id(), true);
295 unlock_vector_lock();
297 per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
298 x86_platform.nmi_init();
300 /* enable local interrupts */
303 /* to prevent fake stack check failure in clock setup */
304 boot_init_stack_canary();
306 x86_cpuinit.setup_percpu_clockev();
313 * The bootstrap kernel entry code has set these up. Save them for
317 void __cpuinit smp_store_cpu_info(int id)
319 struct cpuinfo_x86 *c = &cpu_data(id);
324 identify_secondary_cpu(c);
327 static void __cpuinit link_thread_siblings(int cpu1, int cpu2)
329 cpumask_set_cpu(cpu1, cpu_sibling_mask(cpu2));
330 cpumask_set_cpu(cpu2, cpu_sibling_mask(cpu1));
331 cpumask_set_cpu(cpu1, cpu_core_mask(cpu2));
332 cpumask_set_cpu(cpu2, cpu_core_mask(cpu1));
333 cpumask_set_cpu(cpu1, cpu_llc_shared_mask(cpu2));
334 cpumask_set_cpu(cpu2, cpu_llc_shared_mask(cpu1));
338 void __cpuinit set_cpu_sibling_map(int cpu)
341 struct cpuinfo_x86 *c = &cpu_data(cpu);
343 cpumask_set_cpu(cpu, cpu_sibling_setup_mask);
345 if (smp_num_siblings > 1) {
346 for_each_cpu(i, cpu_sibling_setup_mask) {
347 struct cpuinfo_x86 *o = &cpu_data(i);
349 if (cpu_has(c, X86_FEATURE_TOPOEXT)) {
350 if (c->phys_proc_id == o->phys_proc_id &&
351 per_cpu(cpu_llc_id, cpu) == per_cpu(cpu_llc_id, i) &&
352 c->compute_unit_id == o->compute_unit_id)
353 link_thread_siblings(cpu, i);
354 } else if (c->phys_proc_id == o->phys_proc_id &&
355 c->cpu_core_id == o->cpu_core_id) {
356 link_thread_siblings(cpu, i);
360 cpumask_set_cpu(cpu, cpu_sibling_mask(cpu));
363 cpumask_set_cpu(cpu, cpu_llc_shared_mask(cpu));
365 if (__this_cpu_read(cpu_info.x86_max_cores) == 1) {
366 cpumask_copy(cpu_core_mask(cpu), cpu_sibling_mask(cpu));
371 for_each_cpu(i, cpu_sibling_setup_mask) {
372 if (per_cpu(cpu_llc_id, cpu) != BAD_APICID &&
373 per_cpu(cpu_llc_id, cpu) == per_cpu(cpu_llc_id, i)) {
374 cpumask_set_cpu(i, cpu_llc_shared_mask(cpu));
375 cpumask_set_cpu(cpu, cpu_llc_shared_mask(i));
377 if (c->phys_proc_id == cpu_data(i).phys_proc_id) {
378 cpumask_set_cpu(i, cpu_core_mask(cpu));
379 cpumask_set_cpu(cpu, cpu_core_mask(i));
381 * Does this new cpu bringup a new core?
383 if (cpumask_weight(cpu_sibling_mask(cpu)) == 1) {
385 * for each core in package, increment
386 * the booted_cores for this new cpu
388 if (cpumask_first(cpu_sibling_mask(i)) == i)
391 * increment the core count for all
392 * the other cpus in this package
395 cpu_data(i).booted_cores++;
396 } else if (i != cpu && !c->booted_cores)
397 c->booted_cores = cpu_data(i).booted_cores;
402 /* maps the cpu to the sched domain representing multi-core */
403 const struct cpumask *cpu_coregroup_mask(int cpu)
405 struct cpuinfo_x86 *c = &cpu_data(cpu);
407 * For perf, we return last level cache shared map.
408 * And for power savings, we return cpu_core_map
410 if ((sched_mc_power_savings || sched_smt_power_savings) &&
411 !(cpu_has(c, X86_FEATURE_AMD_DCM)))
412 return cpu_core_mask(cpu);
414 return cpu_llc_shared_mask(cpu);
417 static void impress_friends(void)
420 unsigned long bogosum = 0;
422 * Allow the user to impress friends.
424 pr_debug("Before bogomips.\n");
425 for_each_possible_cpu(cpu)
426 if (cpumask_test_cpu(cpu, cpu_callout_mask))
427 bogosum += cpu_data(cpu).loops_per_jiffy;
429 "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
432 (bogosum/(5000/HZ))%100);
434 pr_debug("Before bogocount - setting activated=1.\n");
437 void __inquire_remote_apic(int apicid)
439 unsigned i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
440 const char * const names[] = { "ID", "VERSION", "SPIV" };
444 printk(KERN_INFO "Inquiring remote APIC 0x%x...\n", apicid);
446 for (i = 0; i < ARRAY_SIZE(regs); i++) {
447 printk(KERN_INFO "... APIC 0x%x %s: ", apicid, names[i]);
452 status = safe_apic_wait_icr_idle();
455 "a previous APIC delivery may have failed\n");
457 apic_icr_write(APIC_DM_REMRD | regs[i], apicid);
462 status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
463 } while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
466 case APIC_ICR_RR_VALID:
467 status = apic_read(APIC_RRR);
468 printk(KERN_CONT "%08x\n", status);
471 printk(KERN_CONT "failed\n");
477 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
478 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
479 * won't ... remember to clear down the APIC, etc later.
482 wakeup_secondary_cpu_via_nmi(int logical_apicid, unsigned long start_eip)
484 unsigned long send_status, accept_status = 0;
488 /* Boot on the stack */
489 /* Kick the second */
490 apic_icr_write(APIC_DM_NMI | apic->dest_logical, logical_apicid);
492 pr_debug("Waiting for send to finish...\n");
493 send_status = safe_apic_wait_icr_idle();
496 * Give the other CPU some time to accept the IPI.
499 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid])) {
500 maxlvt = lapic_get_maxlvt();
501 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
502 apic_write(APIC_ESR, 0);
503 accept_status = (apic_read(APIC_ESR) & 0xEF);
505 pr_debug("NMI sent.\n");
508 printk(KERN_ERR "APIC never delivered???\n");
510 printk(KERN_ERR "APIC delivery error (%lx).\n", accept_status);
512 return (send_status | accept_status);
516 wakeup_secondary_cpu_via_init(int phys_apicid, unsigned long start_eip)
518 unsigned long send_status, accept_status = 0;
519 int maxlvt, num_starts, j;
521 maxlvt = lapic_get_maxlvt();
524 * Be paranoid about clearing APIC errors.
526 if (APIC_INTEGRATED(apic_version[phys_apicid])) {
527 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
528 apic_write(APIC_ESR, 0);
532 pr_debug("Asserting INIT.\n");
535 * Turn INIT on target chip
540 apic_icr_write(APIC_INT_LEVELTRIG | APIC_INT_ASSERT | APIC_DM_INIT,
543 pr_debug("Waiting for send to finish...\n");
544 send_status = safe_apic_wait_icr_idle();
548 pr_debug("Deasserting INIT.\n");
552 apic_icr_write(APIC_INT_LEVELTRIG | APIC_DM_INIT, phys_apicid);
554 pr_debug("Waiting for send to finish...\n");
555 send_status = safe_apic_wait_icr_idle();
558 atomic_set(&init_deasserted, 1);
561 * Should we send STARTUP IPIs ?
563 * Determine this based on the APIC version.
564 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
566 if (APIC_INTEGRATED(apic_version[phys_apicid]))
572 * Paravirt / VMI wants a startup IPI hook here to set up the
573 * target processor state.
575 startup_ipi_hook(phys_apicid, (unsigned long) start_secondary,
579 * Run STARTUP IPI loop.
581 pr_debug("#startup loops: %d.\n", num_starts);
583 for (j = 1; j <= num_starts; j++) {
584 pr_debug("Sending STARTUP #%d.\n", j);
585 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
586 apic_write(APIC_ESR, 0);
588 pr_debug("After apic_write.\n");
595 /* Boot on the stack */
596 /* Kick the second */
597 apic_icr_write(APIC_DM_STARTUP | (start_eip >> 12),
601 * Give the other CPU some time to accept the IPI.
605 pr_debug("Startup point 1.\n");
607 pr_debug("Waiting for send to finish...\n");
608 send_status = safe_apic_wait_icr_idle();
611 * Give the other CPU some time to accept the IPI.
614 if (maxlvt > 3) /* Due to the Pentium erratum 3AP. */
615 apic_write(APIC_ESR, 0);
616 accept_status = (apic_read(APIC_ESR) & 0xEF);
617 if (send_status || accept_status)
620 pr_debug("After Startup.\n");
623 printk(KERN_ERR "APIC never delivered???\n");
625 printk(KERN_ERR "APIC delivery error (%lx).\n", accept_status);
627 return (send_status | accept_status);
631 struct work_struct work;
632 struct task_struct *idle;
633 struct completion done;
637 static void __cpuinit do_fork_idle(struct work_struct *work)
639 struct create_idle *c_idle =
640 container_of(work, struct create_idle, work);
642 c_idle->idle = fork_idle(c_idle->cpu);
643 complete(&c_idle->done);
646 /* reduce the number of lines printed when booting a large cpu count system */
647 static void __cpuinit announce_cpu(int cpu, int apicid)
649 static int current_node = -1;
650 int node = early_cpu_to_node(cpu);
652 if (system_state == SYSTEM_BOOTING) {
653 if (node != current_node) {
654 if (current_node > (-1))
657 pr_info("Booting Node %3d, Processors ", node);
659 pr_cont(" #%d%s", cpu, cpu == (nr_cpu_ids - 1) ? " Ok.\n" : "");
662 pr_info("Booting Node %d Processor %d APIC 0x%x\n",
667 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
668 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
669 * Returns zero if CPU booted OK, else error code from
670 * ->wakeup_secondary_cpu.
672 static int __cpuinit do_boot_cpu(int apicid, int cpu)
674 unsigned long boot_error = 0;
675 unsigned long start_ip;
677 struct create_idle c_idle = {
679 .done = COMPLETION_INITIALIZER_ONSTACK(c_idle.done),
682 INIT_WORK_ONSTACK(&c_idle.work, do_fork_idle);
684 alternatives_smp_switch(1);
686 c_idle.idle = get_idle_for_cpu(cpu);
689 * We can't use kernel_thread since we must avoid to
690 * reschedule the child.
693 c_idle.idle->thread.sp = (unsigned long) (((struct pt_regs *)
694 (THREAD_SIZE + task_stack_page(c_idle.idle))) - 1);
695 init_idle(c_idle.idle, cpu);
699 schedule_work(&c_idle.work);
700 wait_for_completion(&c_idle.done);
702 if (IS_ERR(c_idle.idle)) {
703 printk("failed fork for CPU %d\n", cpu);
704 destroy_work_on_stack(&c_idle.work);
705 return PTR_ERR(c_idle.idle);
708 set_idle_for_cpu(cpu, c_idle.idle);
710 per_cpu(current_task, cpu) = c_idle.idle;
712 /* Stack for startup_32 can be just as for start_secondary onwards */
715 clear_tsk_thread_flag(c_idle.idle, TIF_FORK);
716 initial_gs = per_cpu_offset(cpu);
717 per_cpu(kernel_stack, cpu) =
718 (unsigned long)task_stack_page(c_idle.idle) -
719 KERNEL_STACK_OFFSET + THREAD_SIZE;
721 early_gdt_descr.address = (unsigned long)get_cpu_gdt_table(cpu);
722 initial_code = (unsigned long)start_secondary;
723 stack_start = c_idle.idle->thread.sp;
725 /* start_ip had better be page-aligned! */
726 start_ip = trampoline_address();
728 /* So we see what's up */
729 announce_cpu(cpu, apicid);
732 * This grunge runs the startup process for
733 * the targeted processor.
736 atomic_set(&init_deasserted, 0);
738 if (get_uv_system_type() != UV_NON_UNIQUE_APIC) {
740 pr_debug("Setting warm reset code and vector.\n");
742 smpboot_setup_warm_reset_vector(start_ip);
744 * Be paranoid about clearing APIC errors.
746 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid])) {
747 apic_write(APIC_ESR, 0);
753 * Kick the secondary CPU. Use the method in the APIC driver
754 * if it's defined - or use an INIT boot APIC message otherwise:
756 if (apic->wakeup_secondary_cpu)
757 boot_error = apic->wakeup_secondary_cpu(apicid, start_ip);
759 boot_error = wakeup_secondary_cpu_via_init(apicid, start_ip);
763 * allow APs to start initializing.
765 pr_debug("Before Callout %d.\n", cpu);
766 cpumask_set_cpu(cpu, cpu_callout_mask);
767 pr_debug("After Callout %d.\n", cpu);
770 * Wait 5s total for a response
772 for (timeout = 0; timeout < 50000; timeout++) {
773 if (cpumask_test_cpu(cpu, cpu_callin_mask))
774 break; /* It has booted */
777 * Allow other tasks to run while we wait for the
778 * AP to come online. This also gives a chance
779 * for the MTRR work(triggered by the AP coming online)
780 * to be completed in the stop machine context.
785 if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
786 print_cpu_msr(&cpu_data(cpu));
787 pr_debug("CPU%d: has booted.\n", cpu);
790 if (*(volatile u32 *)TRAMPOLINE_SYM(trampoline_status)
792 /* trampoline started but...? */
793 pr_err("CPU%d: Stuck ??\n", cpu);
795 /* trampoline code not run */
796 pr_err("CPU%d: Not responding.\n", cpu);
797 if (apic->inquire_remote_apic)
798 apic->inquire_remote_apic(apicid);
803 /* Try to put things back the way they were before ... */
804 numa_remove_cpu(cpu); /* was set by numa_add_cpu */
806 /* was set by do_boot_cpu() */
807 cpumask_clear_cpu(cpu, cpu_callout_mask);
809 /* was set by cpu_init() */
810 cpumask_clear_cpu(cpu, cpu_initialized_mask);
812 set_cpu_present(cpu, false);
813 per_cpu(x86_cpu_to_apicid, cpu) = BAD_APICID;
816 /* mark "stuck" area as not stuck */
817 *(volatile u32 *)TRAMPOLINE_SYM(trampoline_status) = 0;
819 if (get_uv_system_type() != UV_NON_UNIQUE_APIC) {
821 * Cleanup possible dangling ends...
823 smpboot_restore_warm_reset_vector();
826 destroy_work_on_stack(&c_idle.work);
830 int __cpuinit native_cpu_up(unsigned int cpu)
832 int apicid = apic->cpu_present_to_apicid(cpu);
836 WARN_ON(irqs_disabled());
838 pr_debug("++++++++++++++++++++=_---CPU UP %u\n", cpu);
840 if (apicid == BAD_APICID || apicid == boot_cpu_physical_apicid ||
841 !physid_isset(apicid, phys_cpu_present_map) ||
842 !apic->apic_id_valid(apicid)) {
843 printk(KERN_ERR "%s: bad cpu %d\n", __func__, cpu);
848 * Already booted CPU?
850 if (cpumask_test_cpu(cpu, cpu_callin_mask)) {
851 pr_debug("do_boot_cpu %d Already started\n", cpu);
856 * Save current MTRR state in case it was changed since early boot
857 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
861 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
863 /* the FPU context is blank, nobody can own it */
864 __cpu_disable_lazy_restore(cpu);
866 err = do_boot_cpu(apicid, cpu);
868 pr_debug("do_boot_cpu failed %d\n", err);
873 * Check TSC synchronization with the AP (keep irqs disabled
876 local_irq_save(flags);
877 check_tsc_sync_source(cpu);
878 local_irq_restore(flags);
880 while (!cpu_online(cpu)) {
882 touch_nmi_watchdog();
889 * arch_disable_smp_support() - disables SMP support for x86 at runtime
891 void arch_disable_smp_support(void)
893 disable_ioapic_support();
897 * Fall back to non SMP mode after errors.
899 * RED-PEN audit/test this more. I bet there is more state messed up here.
901 static __init void disable_smp(void)
903 init_cpu_present(cpumask_of(0));
904 init_cpu_possible(cpumask_of(0));
905 smpboot_clear_io_apic_irqs();
907 if (smp_found_config)
908 physid_set_mask_of_physid(boot_cpu_physical_apicid, &phys_cpu_present_map);
910 physid_set_mask_of_physid(0, &phys_cpu_present_map);
911 cpumask_set_cpu(0, cpu_sibling_mask(0));
912 cpumask_set_cpu(0, cpu_core_mask(0));
916 * Various sanity checks.
918 static int __init smp_sanity_check(unsigned max_cpus)
922 #if !defined(CONFIG_X86_BIGSMP) && defined(CONFIG_X86_32)
923 if (def_to_bigsmp && nr_cpu_ids > 8) {
928 "More than 8 CPUs detected - skipping them.\n"
929 "Use CONFIG_X86_BIGSMP.\n");
932 for_each_present_cpu(cpu) {
934 set_cpu_present(cpu, false);
939 for_each_possible_cpu(cpu) {
941 set_cpu_possible(cpu, false);
949 if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map)) {
951 "weird, boot CPU (#%d) not listed by the BIOS.\n",
952 hard_smp_processor_id());
954 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
958 * If we couldn't find an SMP configuration at boot time,
959 * get out of here now!
961 if (!smp_found_config && !acpi_lapic) {
963 printk(KERN_NOTICE "SMP motherboard not detected.\n");
965 if (APIC_init_uniprocessor())
966 printk(KERN_NOTICE "Local APIC not detected."
967 " Using dummy APIC emulation.\n");
972 * Should not be necessary because the MP table should list the boot
973 * CPU too, but we do it for the sake of robustness anyway.
975 if (!apic->check_phys_apicid_present(boot_cpu_physical_apicid)) {
977 "weird, boot CPU (#%d) not listed by the BIOS.\n",
978 boot_cpu_physical_apicid);
979 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
984 * If we couldn't find a local APIC, then get out of here now!
986 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid]) &&
989 pr_err("BIOS bug, local APIC #%d not detected!...\n",
990 boot_cpu_physical_apicid);
991 pr_err("... forcing use of dummy APIC emulation."
992 "(tell your hw vendor)\n");
994 smpboot_clear_io_apic();
995 disable_ioapic_support();
1002 * If SMP should be disabled, then really disable it!
1005 printk(KERN_INFO "SMP mode deactivated.\n");
1006 smpboot_clear_io_apic();
1010 bsp_end_local_APIC_setup();
1017 static void __init smp_cpu_index_default(void)
1020 struct cpuinfo_x86 *c;
1022 for_each_possible_cpu(i) {
1024 /* mark all to hotplug */
1025 c->cpu_index = nr_cpu_ids;
1030 * Prepare for SMP bootup. The MP table or ACPI has been read
1031 * earlier. Just do some sanity checking here and enable APIC mode.
1033 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1038 smp_cpu_index_default();
1041 * Setup boot CPU information
1043 smp_store_cpu_info(0); /* Final full version of the data */
1044 cpumask_copy(cpu_callin_mask, cpumask_of(0));
1047 current_thread_info()->cpu = 0; /* needed? */
1048 for_each_possible_cpu(i) {
1049 zalloc_cpumask_var(&per_cpu(cpu_sibling_map, i), GFP_KERNEL);
1050 zalloc_cpumask_var(&per_cpu(cpu_core_map, i), GFP_KERNEL);
1051 zalloc_cpumask_var(&per_cpu(cpu_llc_shared_map, i), GFP_KERNEL);
1053 set_cpu_sibling_map(0);
1056 if (smp_sanity_check(max_cpus) < 0) {
1057 printk(KERN_INFO "SMP disabled\n");
1062 default_setup_apic_routing();
1065 if (read_apic_id() != boot_cpu_physical_apicid) {
1066 panic("Boot APIC ID in local APIC unexpected (%d vs %d)",
1067 read_apic_id(), boot_cpu_physical_apicid);
1068 /* Or can we switch back to PIC here? */
1075 * Switch from PIC to APIC mode.
1080 * Enable IO APIC before setting up error vector
1082 if (!skip_ioapic_setup && nr_ioapics)
1085 bsp_end_local_APIC_setup();
1087 if (apic->setup_portio_remap)
1088 apic->setup_portio_remap();
1090 smpboot_setup_io_apic();
1092 * Set up local APIC timer on boot CPU.
1095 printk(KERN_INFO "CPU%d: ", 0);
1096 print_cpu_info(&cpu_data(0));
1097 x86_init.timers.setup_percpu_clockev();
1102 set_mtrr_aps_delayed_init();
1107 void arch_disable_nonboot_cpus_begin(void)
1110 * Avoid the smp alternatives switch during the disable_nonboot_cpus().
1111 * In the suspend path, we will be back in the SMP mode shortly anyways.
1113 skip_smp_alternatives = true;
1116 void arch_disable_nonboot_cpus_end(void)
1118 skip_smp_alternatives = false;
1121 void arch_enable_nonboot_cpus_begin(void)
1123 set_mtrr_aps_delayed_init();
1126 void arch_enable_nonboot_cpus_end(void)
1132 * Early setup to make printk work.
1134 void __init native_smp_prepare_boot_cpu(void)
1136 int me = smp_processor_id();
1137 switch_to_new_gdt(me);
1138 /* already set me in cpu_online_mask in boot_cpu_init() */
1139 cpumask_set_cpu(me, cpu_callout_mask);
1140 per_cpu(cpu_state, me) = CPU_ONLINE;
1143 void __init native_smp_cpus_done(unsigned int max_cpus)
1145 pr_debug("Boot done.\n");
1149 #ifdef CONFIG_X86_IO_APIC
1150 setup_ioapic_dest();
1155 static int __initdata setup_possible_cpus = -1;
1156 static int __init _setup_possible_cpus(char *str)
1158 get_option(&str, &setup_possible_cpus);
1161 early_param("possible_cpus", _setup_possible_cpus);
1165 * cpu_possible_mask should be static, it cannot change as cpu's
1166 * are onlined, or offlined. The reason is per-cpu data-structures
1167 * are allocated by some modules at init time, and dont expect to
1168 * do this dynamically on cpu arrival/departure.
1169 * cpu_present_mask on the other hand can change dynamically.
1170 * In case when cpu_hotplug is not compiled, then we resort to current
1171 * behaviour, which is cpu_possible == cpu_present.
1174 * Three ways to find out the number of additional hotplug CPUs:
1175 * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1176 * - The user can overwrite it with possible_cpus=NUM
1177 * - Otherwise don't reserve additional CPUs.
1178 * We do this because additional CPUs waste a lot of memory.
1181 __init void prefill_possible_map(void)
1185 /* no processor from mptable or madt */
1186 if (!num_processors)
1189 i = setup_max_cpus ?: 1;
1190 if (setup_possible_cpus == -1) {
1191 possible = num_processors;
1192 #ifdef CONFIG_HOTPLUG_CPU
1194 possible += disabled_cpus;
1200 possible = setup_possible_cpus;
1202 total_cpus = max_t(int, possible, num_processors + disabled_cpus);
1204 /* nr_cpu_ids could be reduced via nr_cpus= */
1205 if (possible > nr_cpu_ids) {
1207 "%d Processors exceeds NR_CPUS limit of %d\n",
1208 possible, nr_cpu_ids);
1209 possible = nr_cpu_ids;
1212 #ifdef CONFIG_HOTPLUG_CPU
1213 if (!setup_max_cpus)
1217 "%d Processors exceeds max_cpus limit of %u\n",
1218 possible, setup_max_cpus);
1222 printk(KERN_INFO "SMP: Allowing %d CPUs, %d hotplug CPUs\n",
1223 possible, max_t(int, possible - num_processors, 0));
1225 for (i = 0; i < possible; i++)
1226 set_cpu_possible(i, true);
1227 for (; i < NR_CPUS; i++)
1228 set_cpu_possible(i, false);
1230 nr_cpu_ids = possible;
1233 #ifdef CONFIG_HOTPLUG_CPU
1235 static void remove_siblinginfo(int cpu)
1238 struct cpuinfo_x86 *c = &cpu_data(cpu);
1240 for_each_cpu(sibling, cpu_core_mask(cpu)) {
1241 cpumask_clear_cpu(cpu, cpu_core_mask(sibling));
1243 * last thread sibling in this cpu core going down
1245 if (cpumask_weight(cpu_sibling_mask(cpu)) == 1)
1246 cpu_data(sibling).booted_cores--;
1249 for_each_cpu(sibling, cpu_sibling_mask(cpu))
1250 cpumask_clear_cpu(cpu, cpu_sibling_mask(sibling));
1251 cpumask_clear(cpu_sibling_mask(cpu));
1252 cpumask_clear(cpu_core_mask(cpu));
1253 c->phys_proc_id = 0;
1255 cpumask_clear_cpu(cpu, cpu_sibling_setup_mask);
1258 static void __ref remove_cpu_from_maps(int cpu)
1260 set_cpu_online(cpu, false);
1261 cpumask_clear_cpu(cpu, cpu_callout_mask);
1262 cpumask_clear_cpu(cpu, cpu_callin_mask);
1263 /* was set by cpu_init() */
1264 cpumask_clear_cpu(cpu, cpu_initialized_mask);
1265 numa_remove_cpu(cpu);
1268 void cpu_disable_common(void)
1270 int cpu = smp_processor_id();
1272 remove_siblinginfo(cpu);
1274 /* It's now safe to remove this processor from the online map */
1276 remove_cpu_from_maps(cpu);
1277 unlock_vector_lock();
1281 int native_cpu_disable(void)
1283 int cpu = smp_processor_id();
1286 * Perhaps use cpufreq to drop frequency, but that could go
1287 * into generic code.
1289 * We won't take down the boot processor on i386 due to some
1290 * interrupts only being able to be serviced by the BSP.
1291 * Especially so if we're not using an IOAPIC -zwane
1298 cpu_disable_common();
1302 void native_cpu_die(unsigned int cpu)
1304 /* We don't do anything here: idle task is faking death itself. */
1307 for (i = 0; i < 10; i++) {
1308 /* They ack this in play_dead by setting CPU_DEAD */
1309 if (per_cpu(cpu_state, cpu) == CPU_DEAD) {
1310 if (system_state == SYSTEM_RUNNING)
1311 pr_info("CPU %u is now offline\n", cpu);
1313 if (1 == num_online_cpus())
1314 alternatives_smp_switch(0);
1319 pr_err("CPU %u didn't die...\n", cpu);
1322 void play_dead_common(void)
1325 reset_lazy_tlbstate();
1326 amd_e400_remove_cpu(raw_smp_processor_id());
1330 __this_cpu_write(cpu_state, CPU_DEAD);
1333 * With physical CPU hotplug, we should halt the cpu
1335 local_irq_disable();
1339 * We need to flush the caches before going to sleep, lest we have
1340 * dirty data in our caches when we come back up.
1342 static inline void mwait_play_dead(void)
1344 unsigned int eax, ebx, ecx, edx;
1345 unsigned int highest_cstate = 0;
1346 unsigned int highest_subcstate = 0;
1349 struct cpuinfo_x86 *c = __this_cpu_ptr(&cpu_info);
1351 if (!(this_cpu_has(X86_FEATURE_MWAIT) && mwait_usable(c)))
1353 if (!this_cpu_has(X86_FEATURE_CLFLSH))
1355 if (__this_cpu_read(cpu_info.cpuid_level) < CPUID_MWAIT_LEAF)
1358 eax = CPUID_MWAIT_LEAF;
1360 native_cpuid(&eax, &ebx, &ecx, &edx);
1363 * eax will be 0 if EDX enumeration is not valid.
1364 * Initialized below to cstate, sub_cstate value when EDX is valid.
1366 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED)) {
1369 edx >>= MWAIT_SUBSTATE_SIZE;
1370 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
1371 if (edx & MWAIT_SUBSTATE_MASK) {
1373 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
1376 eax = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
1377 (highest_subcstate - 1);
1381 * This should be a memory location in a cache line which is
1382 * unlikely to be touched by other processors. The actual
1383 * content is immaterial as it is not actually modified in any way.
1385 mwait_ptr = ¤t_thread_info()->flags;
1391 * The CLFLUSH is a workaround for erratum AAI65 for
1392 * the Xeon 7400 series. It's not clear it is actually
1393 * needed, but it should be harmless in either case.
1394 * The WBINVD is insufficient due to the spurious-wakeup
1395 * case where we return around the loop.
1398 __monitor(mwait_ptr, 0, 0);
1404 static inline void hlt_play_dead(void)
1406 if (__this_cpu_read(cpu_info.x86) >= 4)
1414 void native_play_dead(void)
1417 tboot_shutdown(TB_SHUTDOWN_WFS);
1419 mwait_play_dead(); /* Only returns on failure */
1420 if (cpuidle_play_dead())
1424 #else /* ... !CONFIG_HOTPLUG_CPU */
1425 int native_cpu_disable(void)
1430 void native_cpu_die(unsigned int cpu)
1432 /* We said "no" in __cpu_disable */
1436 void native_play_dead(void)