2 * x86 SMP booting functions
4 * (c) 1995 Alan Cox, Building #3 <alan@redhat.com>
5 * (c) 1998, 1999, 2000 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>
56 #include <asm/trampoline.h>
59 #include <asm/pgtable.h>
60 #include <asm/tlbflush.h>
64 #include <linux/mc146818rtc.h>
66 #include <mach_apic.h>
67 #include <mach_wakecpu.h>
68 #include <smpboot_hooks.h>
71 * FIXME: For x86_64, those are defined in other files. But moving them here,
72 * would make the setup areas dependent on smp, which is a loss. When we
73 * integrate apic between arches, we can probably do a better job, but
74 * right now, they'll stay here -- glommer
77 /* which logical CPU number maps to which CPU (physical APIC ID) */
78 u16 x86_cpu_to_apicid_init[NR_CPUS] __initdata =
79 { [0 ... NR_CPUS-1] = BAD_APICID };
80 void *x86_cpu_to_apicid_early_ptr;
82 u16 x86_bios_cpu_apicid_init[NR_CPUS] __initdata
83 = { [0 ... NR_CPUS-1] = BAD_APICID };
84 void *x86_bios_cpu_apicid_early_ptr;
87 u8 apicid_2_node[MAX_APICID];
90 /* State of each CPU */
91 DEFINE_PER_CPU(int, cpu_state) = { 0 };
93 /* Store all idle threads, this can be reused instead of creating
94 * a new thread. Also avoids complicated thread destroy functionality
97 #ifdef CONFIG_HOTPLUG_CPU
99 * Needed only for CONFIG_HOTPLUG_CPU because __cpuinitdata is
100 * removed after init for !CONFIG_HOTPLUG_CPU.
102 static DEFINE_PER_CPU(struct task_struct *, idle_thread_array);
103 #define get_idle_for_cpu(x) (per_cpu(idle_thread_array, x))
104 #define set_idle_for_cpu(x, p) (per_cpu(idle_thread_array, x) = (p))
106 struct task_struct *idle_thread_array[NR_CPUS] __cpuinitdata ;
107 #define get_idle_for_cpu(x) (idle_thread_array[(x)])
108 #define set_idle_for_cpu(x, p) (idle_thread_array[(x)] = (p))
111 /* Number of siblings per CPU package */
112 int smp_num_siblings = 1;
113 EXPORT_SYMBOL(smp_num_siblings);
115 /* Last level cache ID of each logical CPU */
116 DEFINE_PER_CPU(u16, cpu_llc_id) = BAD_APICID;
118 /* bitmap of online cpus */
119 cpumask_t cpu_online_map __read_mostly;
120 EXPORT_SYMBOL(cpu_online_map);
122 cpumask_t cpu_callin_map;
123 cpumask_t cpu_callout_map;
124 cpumask_t cpu_possible_map;
125 EXPORT_SYMBOL(cpu_possible_map);
127 /* representing HT siblings of each logical CPU */
128 DEFINE_PER_CPU(cpumask_t, cpu_sibling_map);
129 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
131 /* representing HT and core siblings of each logical CPU */
132 DEFINE_PER_CPU(cpumask_t, cpu_core_map);
133 EXPORT_PER_CPU_SYMBOL(cpu_core_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 static atomic_t init_deasserted;
141 static int boot_cpu_logical_apicid;
143 /* representing cpus for which sibling maps can be computed */
144 static cpumask_t cpu_sibling_setup_map;
146 /* Set if we find a B stepping CPU */
147 int __cpuinitdata smp_b_stepping;
149 #if defined(CONFIG_NUMA) && defined(CONFIG_X86_32)
151 /* which logical CPUs are on which nodes */
152 cpumask_t node_to_cpumask_map[MAX_NUMNODES] __read_mostly =
153 { [0 ... MAX_NUMNODES-1] = CPU_MASK_NONE };
154 EXPORT_SYMBOL(node_to_cpumask_map);
155 /* which node each logical CPU is on */
156 int cpu_to_node_map[NR_CPUS] __read_mostly = { [0 ... NR_CPUS-1] = 0 };
157 EXPORT_SYMBOL(cpu_to_node_map);
159 /* set up a mapping between cpu and node. */
160 static void map_cpu_to_node(int cpu, int node)
162 printk(KERN_INFO "Mapping cpu %d to node %d\n", cpu, node);
163 cpu_set(cpu, node_to_cpumask_map[node]);
164 cpu_to_node_map[cpu] = node;
167 /* undo a mapping between cpu and node. */
168 static void unmap_cpu_to_node(int cpu)
172 printk(KERN_INFO "Unmapping cpu %d from all nodes\n", cpu);
173 for (node = 0; node < MAX_NUMNODES; node++)
174 cpu_clear(cpu, node_to_cpumask_map[node]);
175 cpu_to_node_map[cpu] = 0;
177 #else /* !(CONFIG_NUMA && CONFIG_X86_32) */
178 #define map_cpu_to_node(cpu, node) ({})
179 #define unmap_cpu_to_node(cpu) ({})
183 u8 cpu_2_logical_apicid[NR_CPUS] __read_mostly =
184 { [0 ... NR_CPUS-1] = BAD_APICID };
186 void map_cpu_to_logical_apicid(void)
188 int cpu = smp_processor_id();
189 int apicid = logical_smp_processor_id();
190 int node = apicid_to_node(apicid);
192 if (!node_online(node))
193 node = first_online_node;
195 cpu_2_logical_apicid[cpu] = apicid;
196 map_cpu_to_node(cpu, node);
199 void unmap_cpu_to_logical_apicid(int cpu)
201 cpu_2_logical_apicid[cpu] = BAD_APICID;
202 unmap_cpu_to_node(cpu);
205 #define unmap_cpu_to_logical_apicid(cpu) do {} while (0)
206 #define map_cpu_to_logical_apicid() do {} while (0)
210 * Report back to the Boot Processor.
213 void __cpuinit smp_callin(void)
216 unsigned long timeout;
219 * If waken up by an INIT in an 82489DX configuration
220 * we may get here before an INIT-deassert IPI reaches
221 * our local APIC. We have to wait for the IPI or we'll
222 * lock up on an APIC access.
224 wait_for_init_deassert(&init_deasserted);
227 * (This works even if the APIC is not enabled.)
229 phys_id = GET_APIC_ID(read_apic_id());
230 cpuid = smp_processor_id();
231 if (cpu_isset(cpuid, cpu_callin_map)) {
232 panic("%s: phys CPU#%d, CPU#%d already present??\n", __func__,
235 Dprintk("CPU#%d (phys ID: %d) waiting for CALLOUT\n", cpuid, phys_id);
238 * STARTUP IPIs are fragile beasts as they might sometimes
239 * trigger some glue motherboard logic. Complete APIC bus
240 * silence for 1 second, this overestimates the time the
241 * boot CPU is spending to send the up to 2 STARTUP IPIs
242 * by a factor of two. This should be enough.
246 * Waiting 2s total for startup (udelay is not yet working)
248 timeout = jiffies + 2*HZ;
249 while (time_before(jiffies, timeout)) {
251 * Has the boot CPU finished it's STARTUP sequence?
253 if (cpu_isset(cpuid, cpu_callout_map))
258 if (!time_before(jiffies, timeout)) {
259 panic("%s: CPU%d started up but did not get a callout!\n",
264 * the boot CPU has finished the init stage and is spinning
265 * on callin_map until we finish. We are free to set up this
266 * CPU, first the APIC. (this is probably redundant on most
270 Dprintk("CALLIN, before setup_local_APIC().\n");
271 smp_callin_clear_local_apic();
273 end_local_APIC_setup();
274 map_cpu_to_logical_apicid();
279 * Need to enable IRQs because it can take longer and then
280 * the NMI watchdog might kill us.
285 Dprintk("Stack at about %p\n", &cpuid);
288 * Save our processor parameters
290 smp_store_cpu_info(cpuid);
293 * Allow the master to continue.
295 cpu_set(cpuid, cpu_callin_map);
299 * Activate a secondary processor.
301 void __cpuinit start_secondary(void *unused)
304 * Don't put *anything* before cpu_init(), SMP booting is too
305 * fragile that we want to limit the things done here to the
306 * most necessary things.
315 /* otherwise gcc will move up smp_processor_id before the cpu_init */
318 * Check TSC synchronization with the BP:
320 check_tsc_sync_target();
322 if (nmi_watchdog == NMI_IO_APIC) {
323 disable_8259A_irq(0);
324 enable_NMI_through_LVT0();
328 /* This must be done before setting cpu_online_map */
329 set_cpu_sibling_map(raw_smp_processor_id());
333 * We need to hold call_lock, so there is no inconsistency
334 * between the time smp_call_function() determines number of
335 * IPI recipients, and the time when the determination is made
336 * for which cpus receive the IPI. Holding this
337 * lock helps us to not include this cpu in a currently in progress
338 * smp_call_function().
340 lock_ipi_call_lock();
342 spin_lock(&vector_lock);
344 /* Setup the per cpu irq handling data structures */
345 __setup_vector_irq(smp_processor_id());
347 * Allow the master to continue.
349 spin_unlock(&vector_lock);
351 cpu_set(smp_processor_id(), cpu_online_map);
352 unlock_ipi_call_lock();
353 per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
355 setup_secondary_clock();
363 * Everything has been set up for the secondary
364 * CPUs - they just need to reload everything
365 * from the task structure
366 * This function must not return.
368 void __devinit initialize_secondary(void)
371 * We don't actually need to load the full TSS,
372 * basically just the stack pointer and the ip.
379 :"m" (current->thread.sp), "m" (current->thread.ip));
383 static void __cpuinit smp_apply_quirks(struct cpuinfo_x86 *c)
387 * Mask B, Pentium, but not Pentium MMX
389 if (c->x86_vendor == X86_VENDOR_INTEL &&
391 c->x86_mask >= 1 && c->x86_mask <= 4 &&
394 * Remember we have B step Pentia with bugs
399 * Certain Athlons might work (for various values of 'work') in SMP
400 * but they are not certified as MP capable.
402 if ((c->x86_vendor == X86_VENDOR_AMD) && (c->x86 == 6)) {
404 if (num_possible_cpus() == 1)
407 /* Athlon 660/661 is valid. */
408 if ((c->x86_model == 6) && ((c->x86_mask == 0) ||
412 /* Duron 670 is valid */
413 if ((c->x86_model == 7) && (c->x86_mask == 0))
417 * Athlon 662, Duron 671, and Athlon >model 7 have capability
418 * bit. It's worth noting that the A5 stepping (662) of some
419 * Athlon XP's have the MP bit set.
420 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
423 if (((c->x86_model == 6) && (c->x86_mask >= 2)) ||
424 ((c->x86_model == 7) && (c->x86_mask >= 1)) ||
429 /* If we get here, not a certified SMP capable AMD system. */
430 add_taint(TAINT_UNSAFE_SMP);
438 void __cpuinit smp_checks(void)
441 printk(KERN_WARNING "WARNING: SMP operation may be unreliable"
442 "with B stepping processors.\n");
445 * Don't taint if we are running SMP kernel on a single non-MP
448 if (tainted & TAINT_UNSAFE_SMP) {
449 if (num_online_cpus())
450 printk(KERN_INFO "WARNING: This combination of AMD"
451 "processors is not suitable for SMP.\n");
453 tainted &= ~TAINT_UNSAFE_SMP;
458 * The bootstrap kernel entry code has set these up. Save them for
462 void __cpuinit smp_store_cpu_info(int id)
464 struct cpuinfo_x86 *c = &cpu_data(id);
469 identify_secondary_cpu(c);
474 void __cpuinit set_cpu_sibling_map(int cpu)
477 struct cpuinfo_x86 *c = &cpu_data(cpu);
479 cpu_set(cpu, cpu_sibling_setup_map);
481 if (smp_num_siblings > 1) {
482 for_each_cpu_mask(i, cpu_sibling_setup_map) {
483 if (c->phys_proc_id == cpu_data(i).phys_proc_id &&
484 c->cpu_core_id == cpu_data(i).cpu_core_id) {
485 cpu_set(i, per_cpu(cpu_sibling_map, cpu));
486 cpu_set(cpu, per_cpu(cpu_sibling_map, i));
487 cpu_set(i, per_cpu(cpu_core_map, cpu));
488 cpu_set(cpu, per_cpu(cpu_core_map, i));
489 cpu_set(i, c->llc_shared_map);
490 cpu_set(cpu, cpu_data(i).llc_shared_map);
494 cpu_set(cpu, per_cpu(cpu_sibling_map, cpu));
497 cpu_set(cpu, c->llc_shared_map);
499 if (current_cpu_data.x86_max_cores == 1) {
500 per_cpu(cpu_core_map, cpu) = per_cpu(cpu_sibling_map, cpu);
505 for_each_cpu_mask(i, cpu_sibling_setup_map) {
506 if (per_cpu(cpu_llc_id, cpu) != BAD_APICID &&
507 per_cpu(cpu_llc_id, cpu) == per_cpu(cpu_llc_id, i)) {
508 cpu_set(i, c->llc_shared_map);
509 cpu_set(cpu, cpu_data(i).llc_shared_map);
511 if (c->phys_proc_id == cpu_data(i).phys_proc_id) {
512 cpu_set(i, per_cpu(cpu_core_map, cpu));
513 cpu_set(cpu, per_cpu(cpu_core_map, i));
515 * Does this new cpu bringup a new core?
517 if (cpus_weight(per_cpu(cpu_sibling_map, cpu)) == 1) {
519 * for each core in package, increment
520 * the booted_cores for this new cpu
522 if (first_cpu(per_cpu(cpu_sibling_map, i)) == i)
525 * increment the core count for all
526 * the other cpus in this package
529 cpu_data(i).booted_cores++;
530 } else if (i != cpu && !c->booted_cores)
531 c->booted_cores = cpu_data(i).booted_cores;
536 /* maps the cpu to the sched domain representing multi-core */
537 cpumask_t cpu_coregroup_map(int cpu)
539 struct cpuinfo_x86 *c = &cpu_data(cpu);
541 * For perf, we return last level cache shared map.
542 * And for power savings, we return cpu_core_map
544 if (sched_mc_power_savings || sched_smt_power_savings)
545 return per_cpu(cpu_core_map, cpu);
547 return c->llc_shared_map;
552 * We are called very early to get the low memory for the
553 * SMP bootup trampoline page.
555 void __init smp_alloc_memory(void)
557 trampoline_base = alloc_bootmem_low_pages(PAGE_SIZE);
559 * Has to be in very low memory so we can execute
562 if (__pa(trampoline_base) >= 0x9F000)
567 void impress_friends(void)
570 unsigned long bogosum = 0;
572 * Allow the user to impress friends.
574 Dprintk("Before bogomips.\n");
575 for_each_possible_cpu(cpu)
576 if (cpu_isset(cpu, cpu_callout_map))
577 bogosum += cpu_data(cpu).loops_per_jiffy;
579 "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
582 (bogosum/(5000/HZ))%100);
584 Dprintk("Before bogocount - setting activated=1.\n");
587 static inline void __inquire_remote_apic(int apicid)
589 unsigned i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
590 char *names[] = { "ID", "VERSION", "SPIV" };
594 printk(KERN_INFO "Inquiring remote APIC #%d...\n", apicid);
596 for (i = 0; i < ARRAY_SIZE(regs); i++) {
597 printk(KERN_INFO "... APIC #%d %s: ", apicid, names[i]);
602 status = safe_apic_wait_icr_idle();
605 "a previous APIC delivery may have failed\n");
607 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(apicid));
608 apic_write_around(APIC_ICR, APIC_DM_REMRD | regs[i]);
613 status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
614 } while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
617 case APIC_ICR_RR_VALID:
618 status = apic_read(APIC_RRR);
619 printk(KERN_CONT "%08x\n", status);
622 printk(KERN_CONT "failed\n");
627 #ifdef WAKE_SECONDARY_VIA_NMI
629 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
630 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
631 * won't ... remember to clear down the APIC, etc later.
634 wakeup_secondary_cpu(int logical_apicid, unsigned long start_eip)
636 unsigned long send_status, accept_status = 0;
640 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(logical_apicid));
642 /* Boot on the stack */
643 /* Kick the second */
644 apic_write_around(APIC_ICR, APIC_DM_NMI | APIC_DEST_LOGICAL);
646 Dprintk("Waiting for send to finish...\n");
647 send_status = safe_apic_wait_icr_idle();
650 * Give the other CPU some time to accept the IPI.
654 * Due to the Pentium erratum 3AP.
656 maxlvt = lapic_get_maxlvt();
658 apic_read_around(APIC_SPIV);
659 apic_write(APIC_ESR, 0);
661 accept_status = (apic_read(APIC_ESR) & 0xEF);
662 Dprintk("NMI sent.\n");
665 printk(KERN_ERR "APIC never delivered???\n");
667 printk(KERN_ERR "APIC delivery error (%lx).\n", accept_status);
669 return (send_status | accept_status);
671 #endif /* WAKE_SECONDARY_VIA_NMI */
673 #ifdef WAKE_SECONDARY_VIA_INIT
675 wakeup_secondary_cpu(int phys_apicid, unsigned long start_eip)
677 unsigned long send_status, accept_status = 0;
678 int maxlvt, num_starts, j;
681 * Be paranoid about clearing APIC errors.
683 if (APIC_INTEGRATED(apic_version[phys_apicid])) {
684 apic_read_around(APIC_SPIV);
685 apic_write(APIC_ESR, 0);
689 Dprintk("Asserting INIT.\n");
692 * Turn INIT on target chip
694 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
699 apic_write_around(APIC_ICR, APIC_INT_LEVELTRIG | APIC_INT_ASSERT
702 Dprintk("Waiting for send to finish...\n");
703 send_status = safe_apic_wait_icr_idle();
707 Dprintk("Deasserting INIT.\n");
710 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
713 apic_write_around(APIC_ICR, APIC_INT_LEVELTRIG | APIC_DM_INIT);
715 Dprintk("Waiting for send to finish...\n");
716 send_status = safe_apic_wait_icr_idle();
719 atomic_set(&init_deasserted, 1);
722 * Should we send STARTUP IPIs ?
724 * Determine this based on the APIC version.
725 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
727 if (APIC_INTEGRATED(apic_version[phys_apicid]))
733 * Paravirt / VMI wants a startup IPI hook here to set up the
734 * target processor state.
736 startup_ipi_hook(phys_apicid, (unsigned long) start_secondary,
738 (unsigned long)init_rsp);
740 (unsigned long)stack_start.sp);
744 * Run STARTUP IPI loop.
746 Dprintk("#startup loops: %d.\n", num_starts);
748 maxlvt = lapic_get_maxlvt();
750 for (j = 1; j <= num_starts; j++) {
751 Dprintk("Sending STARTUP #%d.\n", j);
752 apic_read_around(APIC_SPIV);
753 apic_write(APIC_ESR, 0);
755 Dprintk("After apic_write.\n");
762 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
764 /* Boot on the stack */
765 /* Kick the second */
766 apic_write_around(APIC_ICR, APIC_DM_STARTUP
767 | (start_eip >> 12));
770 * Give the other CPU some time to accept the IPI.
774 Dprintk("Startup point 1.\n");
776 Dprintk("Waiting for send to finish...\n");
777 send_status = safe_apic_wait_icr_idle();
780 * Give the other CPU some time to accept the IPI.
784 * Due to the Pentium erratum 3AP.
787 apic_read_around(APIC_SPIV);
788 apic_write(APIC_ESR, 0);
790 accept_status = (apic_read(APIC_ESR) & 0xEF);
791 if (send_status || accept_status)
794 Dprintk("After Startup.\n");
797 printk(KERN_ERR "APIC never delivered???\n");
799 printk(KERN_ERR "APIC delivery error (%lx).\n", accept_status);
801 return (send_status | accept_status);
803 #endif /* WAKE_SECONDARY_VIA_INIT */
806 struct work_struct work;
807 struct task_struct *idle;
808 struct completion done;
812 static void __cpuinit do_fork_idle(struct work_struct *work)
814 struct create_idle *c_idle =
815 container_of(work, struct create_idle, work);
817 c_idle->idle = fork_idle(c_idle->cpu);
818 complete(&c_idle->done);
821 static int __cpuinit do_boot_cpu(int apicid, int cpu)
823 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
824 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
825 * Returns zero if CPU booted OK, else error code from wakeup_secondary_cpu.
828 unsigned long boot_error = 0;
830 unsigned long start_ip;
831 unsigned short nmi_high = 0, nmi_low = 0;
832 struct create_idle c_idle = {
834 .done = COMPLETION_INITIALIZER_ONSTACK(c_idle.done),
836 INIT_WORK(&c_idle.work, do_fork_idle);
838 /* allocate memory for gdts of secondary cpus. Hotplug is considered */
839 if (!cpu_gdt_descr[cpu].address &&
840 !(cpu_gdt_descr[cpu].address = get_zeroed_page(GFP_KERNEL))) {
841 printk(KERN_ERR "Failed to allocate GDT for CPU %d\n", cpu);
845 /* Allocate node local memory for AP pdas */
846 if (cpu_pda(cpu) == &boot_cpu_pda[cpu]) {
847 struct x8664_pda *newpda, *pda;
848 int node = cpu_to_node(cpu);
850 newpda = kmalloc_node(sizeof(struct x8664_pda), GFP_ATOMIC,
853 memcpy(newpda, pda, sizeof(struct x8664_pda));
854 cpu_pda(cpu) = newpda;
857 "Could not allocate node local PDA for CPU %d on node %d\n",
862 alternatives_smp_switch(1);
864 c_idle.idle = get_idle_for_cpu(cpu);
867 * We can't use kernel_thread since we must avoid to
868 * reschedule the child.
871 c_idle.idle->thread.sp = (unsigned long) (((struct pt_regs *)
872 (THREAD_SIZE + task_stack_page(c_idle.idle))) - 1);
873 init_idle(c_idle.idle, cpu);
877 if (!keventd_up() || current_is_keventd())
878 c_idle.work.func(&c_idle.work);
880 schedule_work(&c_idle.work);
881 wait_for_completion(&c_idle.done);
884 if (IS_ERR(c_idle.idle)) {
885 printk("failed fork for CPU %d\n", cpu);
886 return PTR_ERR(c_idle.idle);
889 set_idle_for_cpu(cpu, c_idle.idle);
892 per_cpu(current_task, cpu) = c_idle.idle;
894 early_gdt_descr.address = (unsigned long)get_cpu_gdt_table(cpu);
895 c_idle.idle->thread.ip = (unsigned long) start_secondary;
896 /* Stack for startup_32 can be just as for start_secondary onwards */
897 stack_start.sp = (void *) c_idle.idle->thread.sp;
900 cpu_pda(cpu)->pcurrent = c_idle.idle;
901 init_rsp = c_idle.idle->thread.sp;
902 load_sp0(&per_cpu(init_tss, cpu), &c_idle.idle->thread);
903 initial_code = (unsigned long)start_secondary;
904 clear_tsk_thread_flag(c_idle.idle, TIF_FORK);
907 /* start_ip had better be page-aligned! */
908 start_ip = setup_trampoline();
910 /* So we see what's up */
911 printk(KERN_INFO "Booting processor %d/%d ip %lx\n",
912 cpu, apicid, start_ip);
915 * This grunge runs the startup process for
916 * the targeted processor.
919 atomic_set(&init_deasserted, 0);
921 Dprintk("Setting warm reset code and vector.\n");
923 store_NMI_vector(&nmi_high, &nmi_low);
925 smpboot_setup_warm_reset_vector(start_ip);
927 * Be paranoid about clearing APIC errors.
929 apic_write(APIC_ESR, 0);
933 * Starting actual IPI sequence...
935 boot_error = wakeup_secondary_cpu(apicid, start_ip);
939 * allow APs to start initializing.
941 Dprintk("Before Callout %d.\n", cpu);
942 cpu_set(cpu, cpu_callout_map);
943 Dprintk("After Callout %d.\n", cpu);
946 * Wait 5s total for a response
948 for (timeout = 0; timeout < 50000; timeout++) {
949 if (cpu_isset(cpu, cpu_callin_map))
950 break; /* It has booted */
954 if (cpu_isset(cpu, cpu_callin_map)) {
955 /* number CPUs logically, starting from 1 (BSP is 0) */
957 printk(KERN_INFO "CPU%d: ", cpu);
958 print_cpu_info(&cpu_data(cpu));
959 Dprintk("CPU has booted.\n");
962 if (*((volatile unsigned char *)trampoline_base)
964 /* trampoline started but...? */
965 printk(KERN_ERR "Stuck ??\n");
967 /* trampoline code not run */
968 printk(KERN_ERR "Not responding.\n");
969 inquire_remote_apic(apicid);
974 /* Try to put things back the way they were before ... */
975 unmap_cpu_to_logical_apicid(cpu);
977 clear_node_cpumask(cpu); /* was set by numa_add_cpu */
979 cpu_clear(cpu, cpu_callout_map); /* was set by do_boot_cpu() */
980 cpu_clear(cpu, cpu_initialized); /* was set by cpu_init() */
981 cpu_clear(cpu, cpu_possible_map);
982 cpu_clear(cpu, cpu_present_map);
983 per_cpu(x86_cpu_to_apicid, cpu) = BAD_APICID;
986 /* mark "stuck" area as not stuck */
987 *((volatile unsigned long *)trampoline_base) = 0;
990 * Cleanup possible dangling ends...
992 smpboot_restore_warm_reset_vector();
997 int __cpuinit native_cpu_up(unsigned int cpu)
999 int apicid = cpu_present_to_apicid(cpu);
1000 unsigned long flags;
1003 WARN_ON(irqs_disabled());
1005 Dprintk("++++++++++++++++++++=_---CPU UP %u\n", cpu);
1007 if (apicid == BAD_APICID || apicid == boot_cpu_physical_apicid ||
1008 !physid_isset(apicid, phys_cpu_present_map)) {
1009 printk(KERN_ERR "%s: bad cpu %d\n", __func__, cpu);
1014 * Already booted CPU?
1016 if (cpu_isset(cpu, cpu_callin_map)) {
1017 Dprintk("do_boot_cpu %d Already started\n", cpu);
1022 * Save current MTRR state in case it was changed since early boot
1023 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1027 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
1029 #ifdef CONFIG_X86_32
1030 /* init low mem mapping */
1031 clone_pgd_range(swapper_pg_dir, swapper_pg_dir + USER_PGD_PTRS,
1032 min_t(unsigned long, KERNEL_PGD_PTRS, USER_PGD_PTRS));
1036 err = do_boot_cpu(apicid, cpu);
1038 Dprintk("do_boot_cpu failed %d\n", err);
1043 * Check TSC synchronization with the AP (keep irqs disabled
1046 local_irq_save(flags);
1047 check_tsc_sync_source(cpu);
1048 local_irq_restore(flags);
1050 while (!cpu_isset(cpu, cpu_online_map)) {
1052 touch_nmi_watchdog();
1059 * Fall back to non SMP mode after errors.
1061 * RED-PEN audit/test this more. I bet there is more state messed up here.
1063 static __init void disable_smp(void)
1065 cpu_present_map = cpumask_of_cpu(0);
1066 cpu_possible_map = cpumask_of_cpu(0);
1067 #ifdef CONFIG_X86_32
1068 smpboot_clear_io_apic_irqs();
1070 if (smp_found_config)
1071 phys_cpu_present_map =
1072 physid_mask_of_physid(boot_cpu_physical_apicid);
1074 phys_cpu_present_map = physid_mask_of_physid(0);
1075 map_cpu_to_logical_apicid();
1076 cpu_set(0, per_cpu(cpu_sibling_map, 0));
1077 cpu_set(0, per_cpu(cpu_core_map, 0));
1081 * Various sanity checks.
1083 static int __init smp_sanity_check(unsigned max_cpus)
1086 if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map)) {
1087 printk(KERN_WARNING "weird, boot CPU (#%d) not listed"
1088 "by the BIOS.\n", hard_smp_processor_id());
1089 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1093 * If we couldn't find an SMP configuration at boot time,
1094 * get out of here now!
1096 if (!smp_found_config && !acpi_lapic) {
1098 printk(KERN_NOTICE "SMP motherboard not detected.\n");
1100 if (APIC_init_uniprocessor())
1101 printk(KERN_NOTICE "Local APIC not detected."
1102 " Using dummy APIC emulation.\n");
1107 * Should not be necessary because the MP table should list the boot
1108 * CPU too, but we do it for the sake of robustness anyway.
1110 if (!check_phys_apicid_present(boot_cpu_physical_apicid)) {
1112 "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 printk(KERN_ERR "BIOS bug, local APIC #%d not detected!...\n",
1124 boot_cpu_physical_apicid);
1125 printk(KERN_ERR "... forcing use of dummy APIC emulation."
1126 "(tell your hw vendor)\n");
1127 smpboot_clear_io_apic();
1131 verify_local_APIC();
1134 * If SMP should be disabled, then really disable it!
1137 printk(KERN_INFO "SMP mode deactivated,"
1138 "forcing use of dummy APIC emulation.\n");
1139 smpboot_clear_io_apic();
1140 #ifdef CONFIG_X86_32
1141 if (nmi_watchdog == NMI_LOCAL_APIC) {
1142 printk(KERN_INFO "activating minimal APIC for"
1143 "NMI watchdog use.\n");
1146 end_local_APIC_setup();
1155 static void __init smp_cpu_index_default(void)
1158 struct cpuinfo_x86 *c;
1160 for_each_cpu_mask(i, cpu_possible_map) {
1162 /* mark all to hotplug */
1163 c->cpu_index = NR_CPUS;
1168 * Prepare for SMP bootup. The MP table or ACPI has been read
1169 * earlier. Just do some sanity checking here and enable APIC mode.
1171 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1173 nmi_watchdog_default();
1174 smp_cpu_index_default();
1175 current_cpu_data = boot_cpu_data;
1176 cpu_callin_map = cpumask_of_cpu(0);
1179 * Setup boot CPU information
1181 smp_store_cpu_info(0); /* Final full version of the data */
1182 boot_cpu_logical_apicid = logical_smp_processor_id();
1183 current_thread_info()->cpu = 0; /* needed? */
1184 set_cpu_sibling_map(0);
1186 if (smp_sanity_check(max_cpus) < 0) {
1187 printk(KERN_INFO "SMP disabled\n");
1193 if (GET_APIC_ID(read_apic_id()) != boot_cpu_physical_apicid) {
1194 panic("Boot APIC ID in local APIC unexpected (%d vs %d)",
1195 GET_APIC_ID(read_apic_id()), boot_cpu_physical_apicid);
1196 /* Or can we switch back to PIC here? */
1200 #ifdef CONFIG_X86_32
1204 * Switch from PIC to APIC mode.
1208 #ifdef CONFIG_X86_64
1210 * Enable IO APIC before setting up error vector
1212 if (!skip_ioapic_setup && nr_ioapics)
1215 end_local_APIC_setup();
1217 map_cpu_to_logical_apicid();
1219 setup_portio_remap();
1221 smpboot_setup_io_apic();
1223 * Set up local APIC timer on boot CPU.
1226 printk(KERN_INFO "CPU%d: ", 0);
1227 print_cpu_info(&cpu_data(0));
1231 * Early setup to make printk work.
1233 void __init native_smp_prepare_boot_cpu(void)
1235 int me = smp_processor_id();
1236 #ifdef CONFIG_X86_32
1238 switch_to_new_gdt();
1240 /* already set me in cpu_online_map in boot_cpu_init() */
1241 cpu_set(me, cpu_callout_map);
1242 per_cpu(cpu_state, me) = CPU_ONLINE;
1245 void __init native_smp_cpus_done(unsigned int max_cpus)
1247 Dprintk("Boot done.\n");
1251 #ifdef CONFIG_X86_IO_APIC
1252 setup_ioapic_dest();
1254 check_nmi_watchdog();
1255 #ifdef CONFIG_X86_32
1260 #ifdef CONFIG_HOTPLUG_CPU
1262 # ifdef CONFIG_X86_32
1263 void cpu_exit_clear(void)
1265 int cpu = raw_smp_processor_id();
1272 cpu_clear(cpu, cpu_callout_map);
1273 cpu_clear(cpu, cpu_callin_map);
1275 unmap_cpu_to_logical_apicid(cpu);
1277 # endif /* CONFIG_X86_32 */
1279 void remove_siblinginfo(int cpu)
1282 struct cpuinfo_x86 *c = &cpu_data(cpu);
1284 for_each_cpu_mask(sibling, per_cpu(cpu_core_map, cpu)) {
1285 cpu_clear(cpu, per_cpu(cpu_core_map, sibling));
1287 * last thread sibling in this cpu core going down
1289 if (cpus_weight(per_cpu(cpu_sibling_map, cpu)) == 1)
1290 cpu_data(sibling).booted_cores--;
1293 for_each_cpu_mask(sibling, per_cpu(cpu_sibling_map, cpu))
1294 cpu_clear(cpu, per_cpu(cpu_sibling_map, sibling));
1295 cpus_clear(per_cpu(cpu_sibling_map, cpu));
1296 cpus_clear(per_cpu(cpu_core_map, cpu));
1297 c->phys_proc_id = 0;
1299 cpu_clear(cpu, cpu_sibling_setup_map);
1302 int additional_cpus __initdata = -1;
1304 static __init int setup_additional_cpus(char *s)
1306 return s && get_option(&s, &additional_cpus) ? 0 : -EINVAL;
1308 early_param("additional_cpus", setup_additional_cpus);
1311 * cpu_possible_map should be static, it cannot change as cpu's
1312 * are onlined, or offlined. The reason is per-cpu data-structures
1313 * are allocated by some modules at init time, and dont expect to
1314 * do this dynamically on cpu arrival/departure.
1315 * cpu_present_map on the other hand can change dynamically.
1316 * In case when cpu_hotplug is not compiled, then we resort to current
1317 * behaviour, which is cpu_possible == cpu_present.
1320 * Three ways to find out the number of additional hotplug CPUs:
1321 * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1322 * - The user can overwrite it with additional_cpus=NUM
1323 * - Otherwise don't reserve additional CPUs.
1324 * We do this because additional CPUs waste a lot of memory.
1327 __init void prefill_possible_map(void)
1332 if (additional_cpus == -1) {
1333 if (disabled_cpus > 0)
1334 additional_cpus = disabled_cpus;
1336 additional_cpus = 0;
1338 possible = num_processors + additional_cpus;
1339 if (possible > NR_CPUS)
1342 printk(KERN_INFO "SMP: Allowing %d CPUs, %d hotplug CPUs\n",
1343 possible, max_t(int, possible - num_processors, 0));
1345 for (i = 0; i < possible; i++)
1346 cpu_set(i, cpu_possible_map);
1349 static void __ref remove_cpu_from_maps(int cpu)
1351 cpu_clear(cpu, cpu_online_map);
1352 #ifdef CONFIG_X86_64
1353 cpu_clear(cpu, cpu_callout_map);
1354 cpu_clear(cpu, cpu_callin_map);
1355 /* was set by cpu_init() */
1356 clear_bit(cpu, (unsigned long *)&cpu_initialized);
1357 clear_node_cpumask(cpu);
1361 int __cpu_disable(void)
1363 int cpu = smp_processor_id();
1366 * Perhaps use cpufreq to drop frequency, but that could go
1367 * into generic code.
1369 * We won't take down the boot processor on i386 due to some
1370 * interrupts only being able to be serviced by the BSP.
1371 * Especially so if we're not using an IOAPIC -zwane
1376 if (nmi_watchdog == NMI_LOCAL_APIC)
1377 stop_apic_nmi_watchdog(NULL);
1382 * Allow any queued timer interrupts to get serviced
1383 * This is only a temporary solution until we cleanup
1384 * fixup_irqs as we do for IA64.
1389 local_irq_disable();
1390 remove_siblinginfo(cpu);
1392 /* It's now safe to remove this processor from the online map */
1393 remove_cpu_from_maps(cpu);
1394 fixup_irqs(cpu_online_map);
1398 void __cpu_die(unsigned int cpu)
1400 /* We don't do anything here: idle task is faking death itself. */
1403 for (i = 0; i < 10; i++) {
1404 /* They ack this in play_dead by setting CPU_DEAD */
1405 if (per_cpu(cpu_state, cpu) == CPU_DEAD) {
1406 printk(KERN_INFO "CPU %d is now offline\n", cpu);
1407 if (1 == num_online_cpus())
1408 alternatives_smp_switch(0);
1413 printk(KERN_ERR "CPU %u didn't die...\n", cpu);
1415 #else /* ... !CONFIG_HOTPLUG_CPU */
1416 int __cpu_disable(void)
1421 void __cpu_die(unsigned int cpu)
1423 /* We said "no" in __cpu_disable */
1429 * If the BIOS enumerates physical processors before logical,
1430 * maxcpus=N at enumeration-time can be used to disable HT.
1432 static int __init parse_maxcpus(char *arg)
1434 extern unsigned int maxcpus;
1436 maxcpus = simple_strtoul(arg, NULL, 0);
1439 early_param("maxcpus", parse_maxcpus);