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>
7 * Much of the core SMP work is based on previous work by Thomas Radke, to
8 * whom a great many thanks are extended.
10 * Thanks to Intel for making available several different Pentium,
11 * Pentium Pro and Pentium-II/Xeon MP machines.
12 * Original development of Linux SMP code supported by Caldera.
14 * This code is released under the GNU General Public License version 2 or
18 * Felix Koop : NR_CPUS used properly
19 * Jose Renau : Handle single CPU case.
20 * Alan Cox : By repeated request 8) - Total BogoMIPS report.
21 * Greg Wright : Fix for kernel stacks panic.
22 * Erich Boleyn : MP v1.4 and additional changes.
23 * Matthias Sattler : Changes for 2.1 kernel map.
24 * Michel Lespinasse : Changes for 2.1 kernel map.
25 * Michael Chastain : Change trampoline.S to gnu as.
26 * Alan Cox : Dumb bug: 'B' step PPro's are fine
27 * Ingo Molnar : Added APIC timers, based on code
29 * Ingo Molnar : various cleanups and rewrites
30 * Tigran Aivazian : fixed "0.00 in /proc/uptime on SMP" bug.
31 * Maciej W. Rozycki : Bits for genuine 82489DX APICs
32 * Martin J. Bligh : Added support for multi-quad systems
33 * Dave Jones : Report invalid combinations of Athlon CPUs.
34 * Rusty Russell : Hacked into shape for new "hotplug" boot process. */
36 #include <linux/module.h>
37 #include <linux/init.h>
38 #include <linux/kernel.h>
41 #include <linux/sched.h>
42 #include <linux/kernel_stat.h>
43 #include <linux/smp_lock.h>
44 #include <linux/bootmem.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/percpu.h>
49 #include <linux/delay.h>
50 #include <linux/mc146818rtc.h>
51 #include <asm/tlbflush.h>
53 #include <asm/arch_hooks.h>
56 #include <mach_apic.h>
57 #include <mach_wakecpu.h>
58 #include <smpboot_hooks.h>
60 /* Set if we find a B stepping CPU */
61 static int __devinitdata smp_b_stepping;
63 /* Number of siblings per CPU package */
64 int smp_num_siblings = 1;
66 EXPORT_SYMBOL(smp_num_siblings);
69 /* Last level cache ID of each logical CPU */
70 int cpu_llc_id[NR_CPUS] __cpuinitdata = {[0 ... NR_CPUS-1] = BAD_APICID};
72 /* representing HT siblings of each logical CPU */
73 cpumask_t cpu_sibling_map[NR_CPUS] __read_mostly;
74 EXPORT_SYMBOL(cpu_sibling_map);
76 /* representing HT and core siblings of each logical CPU */
77 cpumask_t cpu_core_map[NR_CPUS] __read_mostly;
78 EXPORT_SYMBOL(cpu_core_map);
80 /* bitmap of online cpus */
81 cpumask_t cpu_online_map __read_mostly;
82 EXPORT_SYMBOL(cpu_online_map);
84 cpumask_t cpu_callin_map;
85 cpumask_t cpu_callout_map;
86 EXPORT_SYMBOL(cpu_callout_map);
87 cpumask_t cpu_possible_map;
88 EXPORT_SYMBOL(cpu_possible_map);
89 static cpumask_t smp_commenced_mask;
91 /* TSC's upper 32 bits can't be written in eariler CPU (before prescott), there
92 * is no way to resync one AP against BP. TBD: for prescott and above, we
93 * should use IA64's algorithm
95 static int __devinitdata tsc_sync_disabled;
97 /* Per CPU bogomips and other parameters */
98 struct cpuinfo_x86 cpu_data[NR_CPUS] __cacheline_aligned;
99 EXPORT_SYMBOL(cpu_data);
101 u8 x86_cpu_to_apicid[NR_CPUS] __read_mostly =
102 { [0 ... NR_CPUS-1] = 0xff };
103 EXPORT_SYMBOL(x86_cpu_to_apicid);
106 * Trampoline 80x86 program as an array.
109 extern unsigned char trampoline_data [];
110 extern unsigned char trampoline_end [];
111 static unsigned char *trampoline_base;
112 static int trampoline_exec;
114 static void map_cpu_to_logical_apicid(void);
116 /* State of each CPU. */
117 DEFINE_PER_CPU(int, cpu_state) = { 0 };
120 * Currently trivial. Write the real->protected mode
121 * bootstrap into the page concerned. The caller
122 * has made sure it's suitably aligned.
125 static unsigned long __devinit setup_trampoline(void)
127 memcpy(trampoline_base, trampoline_data, trampoline_end - trampoline_data);
128 return virt_to_phys(trampoline_base);
132 * We are called very early to get the low memory for the
133 * SMP bootup trampoline page.
135 void __init smp_alloc_memory(void)
137 trampoline_base = (void *) alloc_bootmem_low_pages(PAGE_SIZE);
139 * Has to be in very low memory so we can execute
142 if (__pa(trampoline_base) >= 0x9F000)
145 * Make the SMP trampoline executable:
147 trampoline_exec = set_kernel_exec((unsigned long)trampoline_base, 1);
151 * The bootstrap kernel entry code has set these up. Save them for
155 static void __devinit smp_store_cpu_info(int id)
157 struct cpuinfo_x86 *c = cpu_data + id;
163 * Mask B, Pentium, but not Pentium MMX
165 if (c->x86_vendor == X86_VENDOR_INTEL &&
167 c->x86_mask >= 1 && c->x86_mask <= 4 &&
170 * Remember we have B step Pentia with bugs
175 * Certain Athlons might work (for various values of 'work') in SMP
176 * but they are not certified as MP capable.
178 if ((c->x86_vendor == X86_VENDOR_AMD) && (c->x86 == 6)) {
180 /* Athlon 660/661 is valid. */
181 if ((c->x86_model==6) && ((c->x86_mask==0) || (c->x86_mask==1)))
184 /* Duron 670 is valid */
185 if ((c->x86_model==7) && (c->x86_mask==0))
189 * Athlon 662, Duron 671, and Athlon >model 7 have capability bit.
190 * It's worth noting that the A5 stepping (662) of some Athlon XP's
191 * have the MP bit set.
192 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for more.
194 if (((c->x86_model==6) && (c->x86_mask>=2)) ||
195 ((c->x86_model==7) && (c->x86_mask>=1)) ||
200 /* If we get here, it's not a certified SMP capable AMD system. */
201 add_taint(TAINT_UNSAFE_SMP);
209 * TSC synchronization.
211 * We first check whether all CPUs have their TSC's synchronized,
212 * then we print a warning if not, and always resync.
217 atomic_t count_start;
219 unsigned long long values[NR_CPUS];
221 .start_flag = ATOMIC_INIT(0),
222 .count_start = ATOMIC_INIT(0),
223 .count_stop = ATOMIC_INIT(0),
228 static void __init synchronize_tsc_bp(void)
231 unsigned long long t0;
232 unsigned long long sum, avg;
234 unsigned int one_usec;
237 printk(KERN_INFO "checking TSC synchronization across %u CPUs: ", num_booting_cpus());
239 /* convert from kcyc/sec to cyc/usec */
240 one_usec = cpu_khz / 1000;
242 atomic_set(&tsc.start_flag, 1);
246 * We loop a few times to get a primed instruction cache,
247 * then the last pass is more or less synchronized and
248 * the BP and APs set their cycle counters to zero all at
249 * once. This reduces the chance of having random offsets
250 * between the processors, and guarantees that the maximum
251 * delay between the cycle counters is never bigger than
252 * the latency of information-passing (cachelines) between
255 for (i = 0; i < NR_LOOPS; i++) {
257 * all APs synchronize but they loop on '== num_cpus'
259 while (atomic_read(&tsc.count_start) != num_booting_cpus()-1)
261 atomic_set(&tsc.count_stop, 0);
264 * this lets the APs save their current TSC:
266 atomic_inc(&tsc.count_start);
268 rdtscll(tsc.values[smp_processor_id()]);
270 * We clear the TSC in the last loop:
276 * Wait for all APs to leave the synchronization point:
278 while (atomic_read(&tsc.count_stop) != num_booting_cpus()-1)
280 atomic_set(&tsc.count_start, 0);
282 atomic_inc(&tsc.count_stop);
286 for (i = 0; i < NR_CPUS; i++) {
287 if (cpu_isset(i, cpu_callout_map)) {
293 do_div(avg, num_booting_cpus());
295 for (i = 0; i < NR_CPUS; i++) {
296 if (!cpu_isset(i, cpu_callout_map))
298 delta = tsc.values[i] - avg;
302 * We report bigger than 2 microseconds clock differences.
304 if (delta > 2*one_usec) {
312 do_div(realdelta, one_usec);
313 if (tsc.values[i] < avg)
314 realdelta = -realdelta;
317 printk(KERN_INFO "CPU#%d had %Ld usecs TSC "
318 "skew, fixed it up.\n", i, realdelta);
325 static void __init synchronize_tsc_ap(void)
330 * Not every cpu is online at the time
331 * this gets called, so we first wait for the BP to
332 * finish SMP initialization:
334 while (!atomic_read(&tsc.start_flag))
337 for (i = 0; i < NR_LOOPS; i++) {
338 atomic_inc(&tsc.count_start);
339 while (atomic_read(&tsc.count_start) != num_booting_cpus())
342 rdtscll(tsc.values[smp_processor_id()]);
346 atomic_inc(&tsc.count_stop);
347 while (atomic_read(&tsc.count_stop) != num_booting_cpus())
353 extern void calibrate_delay(void);
355 static atomic_t init_deasserted;
357 static void __devinit smp_callin(void)
360 unsigned long timeout;
363 * If waken up by an INIT in an 82489DX configuration
364 * we may get here before an INIT-deassert IPI reaches
365 * our local APIC. We have to wait for the IPI or we'll
366 * lock up on an APIC access.
368 wait_for_init_deassert(&init_deasserted);
371 * (This works even if the APIC is not enabled.)
373 phys_id = GET_APIC_ID(apic_read(APIC_ID));
374 cpuid = smp_processor_id();
375 if (cpu_isset(cpuid, cpu_callin_map)) {
376 printk("huh, phys CPU#%d, CPU#%d already present??\n",
380 Dprintk("CPU#%d (phys ID: %d) waiting for CALLOUT\n", cpuid, phys_id);
383 * STARTUP IPIs are fragile beasts as they might sometimes
384 * trigger some glue motherboard logic. Complete APIC bus
385 * silence for 1 second, this overestimates the time the
386 * boot CPU is spending to send the up to 2 STARTUP IPIs
387 * by a factor of two. This should be enough.
391 * Waiting 2s total for startup (udelay is not yet working)
393 timeout = jiffies + 2*HZ;
394 while (time_before(jiffies, timeout)) {
396 * Has the boot CPU finished it's STARTUP sequence?
398 if (cpu_isset(cpuid, cpu_callout_map))
403 if (!time_before(jiffies, timeout)) {
404 printk("BUG: CPU%d started up but did not get a callout!\n",
410 * the boot CPU has finished the init stage and is spinning
411 * on callin_map until we finish. We are free to set up this
412 * CPU, first the APIC. (this is probably redundant on most
416 Dprintk("CALLIN, before setup_local_APIC().\n");
417 smp_callin_clear_local_apic();
419 map_cpu_to_logical_apicid();
425 Dprintk("Stack at about %p\n",&cpuid);
428 * Save our processor parameters
430 smp_store_cpu_info(cpuid);
432 disable_APIC_timer();
435 * Allow the master to continue.
437 cpu_set(cpuid, cpu_callin_map);
440 * Synchronize the TSC with the BP
442 if (cpu_has_tsc && cpu_khz && !tsc_sync_disabled)
443 synchronize_tsc_ap();
448 /* maps the cpu to the sched domain representing multi-core */
449 cpumask_t cpu_coregroup_map(int cpu)
451 struct cpuinfo_x86 *c = cpu_data + cpu;
453 * For perf, we return last level cache shared map.
454 * And for power savings, we return cpu_core_map
456 if (sched_mc_power_savings || sched_smt_power_savings)
457 return cpu_core_map[cpu];
459 return c->llc_shared_map;
462 /* representing cpus for which sibling maps can be computed */
463 static cpumask_t cpu_sibling_setup_map;
466 set_cpu_sibling_map(int cpu)
469 struct cpuinfo_x86 *c = cpu_data;
471 cpu_set(cpu, cpu_sibling_setup_map);
473 if (smp_num_siblings > 1) {
474 for_each_cpu_mask(i, cpu_sibling_setup_map) {
475 if (c[cpu].phys_proc_id == c[i].phys_proc_id &&
476 c[cpu].cpu_core_id == c[i].cpu_core_id) {
477 cpu_set(i, cpu_sibling_map[cpu]);
478 cpu_set(cpu, cpu_sibling_map[i]);
479 cpu_set(i, cpu_core_map[cpu]);
480 cpu_set(cpu, cpu_core_map[i]);
481 cpu_set(i, c[cpu].llc_shared_map);
482 cpu_set(cpu, c[i].llc_shared_map);
486 cpu_set(cpu, cpu_sibling_map[cpu]);
489 cpu_set(cpu, c[cpu].llc_shared_map);
491 if (current_cpu_data.x86_max_cores == 1) {
492 cpu_core_map[cpu] = cpu_sibling_map[cpu];
493 c[cpu].booted_cores = 1;
497 for_each_cpu_mask(i, cpu_sibling_setup_map) {
498 if (cpu_llc_id[cpu] != BAD_APICID &&
499 cpu_llc_id[cpu] == cpu_llc_id[i]) {
500 cpu_set(i, c[cpu].llc_shared_map);
501 cpu_set(cpu, c[i].llc_shared_map);
503 if (c[cpu].phys_proc_id == c[i].phys_proc_id) {
504 cpu_set(i, cpu_core_map[cpu]);
505 cpu_set(cpu, cpu_core_map[i]);
507 * Does this new cpu bringup a new core?
509 if (cpus_weight(cpu_sibling_map[cpu]) == 1) {
511 * for each core in package, increment
512 * the booted_cores for this new cpu
514 if (first_cpu(cpu_sibling_map[i]) == i)
515 c[cpu].booted_cores++;
517 * increment the core count for all
518 * the other cpus in this package
522 } else if (i != cpu && !c[cpu].booted_cores)
523 c[cpu].booted_cores = c[i].booted_cores;
529 * Activate a secondary processor.
531 static void __devinit start_secondary(void *unused)
534 * Dont put anything before smp_callin(), SMP
535 * booting is too fragile that we want to limit the
536 * things done here to the most necessary things.
541 while (!cpu_isset(smp_processor_id(), smp_commenced_mask))
543 setup_secondary_APIC_clock();
544 if (nmi_watchdog == NMI_IO_APIC) {
545 disable_8259A_irq(0);
546 enable_NMI_through_LVT0(NULL);
551 * low-memory mappings have been cleared, flush them from
552 * the local TLBs too.
556 /* This must be done before setting cpu_online_map */
557 set_cpu_sibling_map(raw_smp_processor_id());
561 * We need to hold call_lock, so there is no inconsistency
562 * between the time smp_call_function() determines number of
563 * IPI receipients, and the time when the determination is made
564 * for which cpus receive the IPI. Holding this
565 * lock helps us to not include this cpu in a currently in progress
566 * smp_call_function().
568 lock_ipi_call_lock();
569 cpu_set(smp_processor_id(), cpu_online_map);
570 unlock_ipi_call_lock();
571 per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
573 /* We can take interrupts now: we're officially "up". */
581 * Everything has been set up for the secondary
582 * CPUs - they just need to reload everything
583 * from the task structure
584 * This function must not return.
586 void __devinit initialize_secondary(void)
589 * We don't actually need to load the full TSS,
590 * basically just the stack pointer and the eip.
597 :"r" (current->thread.esp),"r" (current->thread.eip));
607 /* which logical CPUs are on which nodes */
608 cpumask_t node_2_cpu_mask[MAX_NUMNODES] __read_mostly =
609 { [0 ... MAX_NUMNODES-1] = CPU_MASK_NONE };
610 /* which node each logical CPU is on */
611 int cpu_2_node[NR_CPUS] __read_mostly = { [0 ... NR_CPUS-1] = 0 };
612 EXPORT_SYMBOL(cpu_2_node);
614 /* set up a mapping between cpu and node. */
615 static inline void map_cpu_to_node(int cpu, int node)
617 printk("Mapping cpu %d to node %d\n", cpu, node);
618 cpu_set(cpu, node_2_cpu_mask[node]);
619 cpu_2_node[cpu] = node;
622 /* undo a mapping between cpu and node. */
623 static inline void unmap_cpu_to_node(int cpu)
627 printk("Unmapping cpu %d from all nodes\n", cpu);
628 for (node = 0; node < MAX_NUMNODES; node ++)
629 cpu_clear(cpu, node_2_cpu_mask[node]);
632 #else /* !CONFIG_NUMA */
634 #define map_cpu_to_node(cpu, node) ({})
635 #define unmap_cpu_to_node(cpu) ({})
637 #endif /* CONFIG_NUMA */
639 u8 cpu_2_logical_apicid[NR_CPUS] __read_mostly = { [0 ... NR_CPUS-1] = BAD_APICID };
641 static void map_cpu_to_logical_apicid(void)
643 int cpu = smp_processor_id();
644 int apicid = logical_smp_processor_id();
645 int node = apicid_to_node(apicid);
647 if (!node_online(node))
648 node = first_online_node;
650 cpu_2_logical_apicid[cpu] = apicid;
651 map_cpu_to_node(cpu, node);
654 static void unmap_cpu_to_logical_apicid(int cpu)
656 cpu_2_logical_apicid[cpu] = BAD_APICID;
657 unmap_cpu_to_node(cpu);
661 static inline void __inquire_remote_apic(int apicid)
663 int i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
664 char *names[] = { "ID", "VERSION", "SPIV" };
667 printk("Inquiring remote APIC #%d...\n", apicid);
669 for (i = 0; i < ARRAY_SIZE(regs); i++) {
670 printk("... APIC #%d %s: ", apicid, names[i]);
675 apic_wait_icr_idle();
677 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(apicid));
678 apic_write_around(APIC_ICR, APIC_DM_REMRD | regs[i]);
683 status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
684 } while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
687 case APIC_ICR_RR_VALID:
688 status = apic_read(APIC_RRR);
689 printk("%08x\n", status);
698 #ifdef WAKE_SECONDARY_VIA_NMI
700 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
701 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
702 * won't ... remember to clear down the APIC, etc later.
705 wakeup_secondary_cpu(int logical_apicid, unsigned long start_eip)
707 unsigned long send_status = 0, accept_status = 0;
711 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(logical_apicid));
713 /* Boot on the stack */
714 /* Kick the second */
715 apic_write_around(APIC_ICR, APIC_DM_NMI | APIC_DEST_LOGICAL);
717 Dprintk("Waiting for send to finish...\n");
722 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
723 } while (send_status && (timeout++ < 1000));
726 * Give the other CPU some time to accept the IPI.
730 * Due to the Pentium erratum 3AP.
732 maxlvt = get_maxlvt();
734 apic_read_around(APIC_SPIV);
735 apic_write(APIC_ESR, 0);
737 accept_status = (apic_read(APIC_ESR) & 0xEF);
738 Dprintk("NMI sent.\n");
741 printk("APIC never delivered???\n");
743 printk("APIC delivery error (%lx).\n", accept_status);
745 return (send_status | accept_status);
747 #endif /* WAKE_SECONDARY_VIA_NMI */
749 #ifdef WAKE_SECONDARY_VIA_INIT
751 wakeup_secondary_cpu(int phys_apicid, unsigned long start_eip)
753 unsigned long send_status = 0, accept_status = 0;
754 int maxlvt, timeout, num_starts, j;
757 * Be paranoid about clearing APIC errors.
759 if (APIC_INTEGRATED(apic_version[phys_apicid])) {
760 apic_read_around(APIC_SPIV);
761 apic_write(APIC_ESR, 0);
765 Dprintk("Asserting INIT.\n");
768 * Turn INIT on target chip
770 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
775 apic_write_around(APIC_ICR, APIC_INT_LEVELTRIG | APIC_INT_ASSERT
778 Dprintk("Waiting for send to finish...\n");
783 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
784 } while (send_status && (timeout++ < 1000));
788 Dprintk("Deasserting INIT.\n");
791 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
794 apic_write_around(APIC_ICR, APIC_INT_LEVELTRIG | APIC_DM_INIT);
796 Dprintk("Waiting for send to finish...\n");
801 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
802 } while (send_status && (timeout++ < 1000));
804 atomic_set(&init_deasserted, 1);
807 * Should we send STARTUP IPIs ?
809 * Determine this based on the APIC version.
810 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
812 if (APIC_INTEGRATED(apic_version[phys_apicid]))
818 * Run STARTUP IPI loop.
820 Dprintk("#startup loops: %d.\n", num_starts);
822 maxlvt = get_maxlvt();
824 for (j = 1; j <= num_starts; j++) {
825 Dprintk("Sending STARTUP #%d.\n",j);
826 apic_read_around(APIC_SPIV);
827 apic_write(APIC_ESR, 0);
829 Dprintk("After apic_write.\n");
836 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
838 /* Boot on the stack */
839 /* Kick the second */
840 apic_write_around(APIC_ICR, APIC_DM_STARTUP
841 | (start_eip >> 12));
844 * Give the other CPU some time to accept the IPI.
848 Dprintk("Startup point 1.\n");
850 Dprintk("Waiting for send to finish...\n");
855 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
856 } while (send_status && (timeout++ < 1000));
859 * Give the other CPU some time to accept the IPI.
863 * Due to the Pentium erratum 3AP.
866 apic_read_around(APIC_SPIV);
867 apic_write(APIC_ESR, 0);
869 accept_status = (apic_read(APIC_ESR) & 0xEF);
870 if (send_status || accept_status)
873 Dprintk("After Startup.\n");
876 printk("APIC never delivered???\n");
878 printk("APIC delivery error (%lx).\n", accept_status);
880 return (send_status | accept_status);
882 #endif /* WAKE_SECONDARY_VIA_INIT */
884 extern cpumask_t cpu_initialized;
885 static inline int alloc_cpu_id(void)
889 cpus_complement(tmp_map, cpu_present_map);
890 cpu = first_cpu(tmp_map);
896 #ifdef CONFIG_HOTPLUG_CPU
897 static struct task_struct * __devinitdata cpu_idle_tasks[NR_CPUS];
898 static inline struct task_struct * alloc_idle_task(int cpu)
900 struct task_struct *idle;
902 if ((idle = cpu_idle_tasks[cpu]) != NULL) {
903 /* initialize thread_struct. we really want to avoid destroy
906 idle->thread.esp = (unsigned long)task_pt_regs(idle);
907 init_idle(idle, cpu);
910 idle = fork_idle(cpu);
913 cpu_idle_tasks[cpu] = idle;
917 #define alloc_idle_task(cpu) fork_idle(cpu)
920 static int __devinit do_boot_cpu(int apicid, int cpu)
922 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
923 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
924 * Returns zero if CPU booted OK, else error code from wakeup_secondary_cpu.
927 struct task_struct *idle;
928 unsigned long boot_error;
930 unsigned long start_eip;
931 unsigned short nmi_high = 0, nmi_low = 0;
934 alternatives_smp_switch(1);
937 * We can't use kernel_thread since we must avoid to
938 * reschedule the child.
940 idle = alloc_idle_task(cpu);
942 panic("failed fork for CPU %d", cpu);
943 idle->thread.eip = (unsigned long) start_secondary;
944 /* start_eip had better be page-aligned! */
945 start_eip = setup_trampoline();
947 /* So we see what's up */
948 printk("Booting processor %d/%d eip %lx\n", cpu, apicid, start_eip);
949 /* Stack for startup_32 can be just as for start_secondary onwards */
950 stack_start.esp = (void *) idle->thread.esp;
955 * This grunge runs the startup process for
956 * the targeted processor.
959 atomic_set(&init_deasserted, 0);
961 Dprintk("Setting warm reset code and vector.\n");
963 store_NMI_vector(&nmi_high, &nmi_low);
965 smpboot_setup_warm_reset_vector(start_eip);
968 * Starting actual IPI sequence...
970 boot_error = wakeup_secondary_cpu(apicid, start_eip);
974 * allow APs to start initializing.
976 Dprintk("Before Callout %d.\n", cpu);
977 cpu_set(cpu, cpu_callout_map);
978 Dprintk("After Callout %d.\n", cpu);
981 * Wait 5s total for a response
983 for (timeout = 0; timeout < 50000; timeout++) {
984 if (cpu_isset(cpu, cpu_callin_map))
985 break; /* It has booted */
989 if (cpu_isset(cpu, cpu_callin_map)) {
990 /* number CPUs logically, starting from 1 (BSP is 0) */
992 printk("CPU%d: ", cpu);
993 print_cpu_info(&cpu_data[cpu]);
994 Dprintk("CPU has booted.\n");
997 if (*((volatile unsigned char *)trampoline_base)
999 /* trampoline started but...? */
1000 printk("Stuck ??\n");
1002 /* trampoline code not run */
1003 printk("Not responding.\n");
1004 inquire_remote_apic(apicid);
1009 /* Try to put things back the way they were before ... */
1010 unmap_cpu_to_logical_apicid(cpu);
1011 cpu_clear(cpu, cpu_callout_map); /* was set here (do_boot_cpu()) */
1012 cpu_clear(cpu, cpu_initialized); /* was set by cpu_init() */
1015 x86_cpu_to_apicid[cpu] = apicid;
1016 cpu_set(cpu, cpu_present_map);
1019 /* mark "stuck" area as not stuck */
1020 *((volatile unsigned long *)trampoline_base) = 0;
1025 #ifdef CONFIG_HOTPLUG_CPU
1026 void cpu_exit_clear(void)
1028 int cpu = raw_smp_processor_id();
1036 cpu_clear(cpu, cpu_callout_map);
1037 cpu_clear(cpu, cpu_callin_map);
1039 cpu_clear(cpu, smp_commenced_mask);
1040 unmap_cpu_to_logical_apicid(cpu);
1043 struct warm_boot_cpu_info {
1044 struct completion *complete;
1049 static void __cpuinit do_warm_boot_cpu(void *p)
1051 struct warm_boot_cpu_info *info = p;
1052 do_boot_cpu(info->apicid, info->cpu);
1053 complete(info->complete);
1056 static int __cpuinit __smp_prepare_cpu(int cpu)
1058 DECLARE_COMPLETION(done);
1059 struct warm_boot_cpu_info info;
1060 struct work_struct task;
1062 struct Xgt_desc_struct *cpu_gdt_descr = &per_cpu(cpu_gdt_descr, cpu);
1064 apicid = x86_cpu_to_apicid[cpu];
1065 if (apicid == BAD_APICID) {
1071 * the CPU isn't initialized at boot time, allocate gdt table here.
1072 * cpu_init will initialize it
1074 if (!cpu_gdt_descr->address) {
1075 cpu_gdt_descr->address = get_zeroed_page(GFP_KERNEL);
1076 if (!cpu_gdt_descr->address)
1077 printk(KERN_CRIT "CPU%d failed to allocate GDT\n", cpu);
1082 info.complete = &done;
1083 info.apicid = apicid;
1085 INIT_WORK(&task, do_warm_boot_cpu, &info);
1087 tsc_sync_disabled = 1;
1089 /* init low mem mapping */
1090 clone_pgd_range(swapper_pg_dir, swapper_pg_dir + USER_PGD_PTRS,
1093 schedule_work(&task);
1094 wait_for_completion(&done);
1096 tsc_sync_disabled = 0;
1104 static void smp_tune_scheduling (void)
1106 unsigned long cachesize; /* kB */
1107 unsigned long bandwidth = 350; /* MB/s */
1109 * Rough estimation for SMP scheduling, this is the number of
1110 * cycles it takes for a fully memory-limited process to flush
1111 * the SMP-local cache.
1113 * (For a P5 this pretty much means we will choose another idle
1114 * CPU almost always at wakeup time (this is due to the small
1115 * L1 cache), on PIIs it's around 50-100 usecs, depending on
1121 * this basically disables processor-affinity
1122 * scheduling on SMP without a TSC.
1126 cachesize = boot_cpu_data.x86_cache_size;
1127 if (cachesize == -1) {
1128 cachesize = 16; /* Pentiums, 2x8kB cache */
1131 max_cache_size = cachesize * 1024;
1136 * Cycle through the processors sending APIC IPIs to boot each.
1139 static int boot_cpu_logical_apicid;
1140 /* Where the IO area was mapped on multiquad, always 0 otherwise */
1142 #ifdef CONFIG_X86_NUMAQ
1143 EXPORT_SYMBOL(xquad_portio);
1146 static void __init smp_boot_cpus(unsigned int max_cpus)
1148 int apicid, cpu, bit, kicked;
1149 unsigned long bogosum = 0;
1152 * Setup boot CPU information
1154 smp_store_cpu_info(0); /* Final full version of the data */
1155 printk("CPU%d: ", 0);
1156 print_cpu_info(&cpu_data[0]);
1158 boot_cpu_physical_apicid = GET_APIC_ID(apic_read(APIC_ID));
1159 boot_cpu_logical_apicid = logical_smp_processor_id();
1160 x86_cpu_to_apicid[0] = boot_cpu_physical_apicid;
1162 current_thread_info()->cpu = 0;
1163 smp_tune_scheduling();
1165 set_cpu_sibling_map(0);
1168 * If we couldn't find an SMP configuration at boot time,
1169 * get out of here now!
1171 if (!smp_found_config && !acpi_lapic) {
1172 printk(KERN_NOTICE "SMP motherboard not detected.\n");
1173 smpboot_clear_io_apic_irqs();
1174 phys_cpu_present_map = physid_mask_of_physid(0);
1175 if (APIC_init_uniprocessor())
1176 printk(KERN_NOTICE "Local APIC not detected."
1177 " Using dummy APIC emulation.\n");
1178 map_cpu_to_logical_apicid();
1179 cpu_set(0, cpu_sibling_map[0]);
1180 cpu_set(0, cpu_core_map[0]);
1185 * Should not be necessary because the MP table should list the boot
1186 * CPU too, but we do it for the sake of robustness anyway.
1187 * Makes no sense to do this check in clustered apic mode, so skip it
1189 if (!check_phys_apicid_present(boot_cpu_physical_apicid)) {
1190 printk("weird, boot CPU (#%d) not listed by the BIOS.\n",
1191 boot_cpu_physical_apicid);
1192 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1196 * If we couldn't find a local APIC, then get out of here now!
1198 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid]) && !cpu_has_apic) {
1199 printk(KERN_ERR "BIOS bug, local APIC #%d not detected!...\n",
1200 boot_cpu_physical_apicid);
1201 printk(KERN_ERR "... forcing use of dummy APIC emulation. (tell your hw vendor)\n");
1202 smpboot_clear_io_apic_irqs();
1203 phys_cpu_present_map = physid_mask_of_physid(0);
1204 cpu_set(0, cpu_sibling_map[0]);
1205 cpu_set(0, cpu_core_map[0]);
1209 verify_local_APIC();
1212 * If SMP should be disabled, then really disable it!
1215 smp_found_config = 0;
1216 printk(KERN_INFO "SMP mode deactivated, forcing use of dummy APIC emulation.\n");
1217 smpboot_clear_io_apic_irqs();
1218 phys_cpu_present_map = physid_mask_of_physid(0);
1219 cpu_set(0, cpu_sibling_map[0]);
1220 cpu_set(0, cpu_core_map[0]);
1226 map_cpu_to_logical_apicid();
1229 setup_portio_remap();
1232 * Scan the CPU present map and fire up the other CPUs via do_boot_cpu
1234 * In clustered apic mode, phys_cpu_present_map is a constructed thus:
1235 * bits 0-3 are quad0, 4-7 are quad1, etc. A perverse twist on the
1236 * clustered apic ID.
1238 Dprintk("CPU present map: %lx\n", physids_coerce(phys_cpu_present_map));
1241 for (bit = 0; kicked < NR_CPUS && bit < MAX_APICS; bit++) {
1242 apicid = cpu_present_to_apicid(bit);
1244 * Don't even attempt to start the boot CPU!
1246 if ((apicid == boot_cpu_apicid) || (apicid == BAD_APICID))
1249 if (!check_apicid_present(bit))
1251 if (max_cpus <= cpucount+1)
1254 if (((cpu = alloc_cpu_id()) <= 0) || do_boot_cpu(apicid, cpu))
1255 printk("CPU #%d not responding - cannot use it.\n",
1262 * Cleanup possible dangling ends...
1264 smpboot_restore_warm_reset_vector();
1267 * Allow the user to impress friends.
1269 Dprintk("Before bogomips.\n");
1270 for (cpu = 0; cpu < NR_CPUS; cpu++)
1271 if (cpu_isset(cpu, cpu_callout_map))
1272 bogosum += cpu_data[cpu].loops_per_jiffy;
1274 "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
1276 bogosum/(500000/HZ),
1277 (bogosum/(5000/HZ))%100);
1279 Dprintk("Before bogocount - setting activated=1.\n");
1282 printk(KERN_WARNING "WARNING: SMP operation may be unreliable with B stepping processors.\n");
1285 * Don't taint if we are running SMP kernel on a single non-MP
1288 if (tainted & TAINT_UNSAFE_SMP) {
1290 printk (KERN_INFO "WARNING: This combination of AMD processors is not suitable for SMP.\n");
1292 tainted &= ~TAINT_UNSAFE_SMP;
1295 Dprintk("Boot done.\n");
1298 * construct cpu_sibling_map[], so that we can tell sibling CPUs
1301 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1302 cpus_clear(cpu_sibling_map[cpu]);
1303 cpus_clear(cpu_core_map[cpu]);
1306 cpu_set(0, cpu_sibling_map[0]);
1307 cpu_set(0, cpu_core_map[0]);
1309 smpboot_setup_io_apic();
1311 setup_boot_APIC_clock();
1314 * Synchronize the TSC with the AP
1316 if (cpu_has_tsc && cpucount && cpu_khz)
1317 synchronize_tsc_bp();
1320 /* These are wrappers to interface to the new boot process. Someone
1321 who understands all this stuff should rewrite it properly. --RR 15/Jul/02 */
1322 void __init smp_prepare_cpus(unsigned int max_cpus)
1324 smp_commenced_mask = cpumask_of_cpu(0);
1325 cpu_callin_map = cpumask_of_cpu(0);
1327 smp_boot_cpus(max_cpus);
1330 void __devinit smp_prepare_boot_cpu(void)
1332 cpu_set(smp_processor_id(), cpu_online_map);
1333 cpu_set(smp_processor_id(), cpu_callout_map);
1334 cpu_set(smp_processor_id(), cpu_present_map);
1335 cpu_set(smp_processor_id(), cpu_possible_map);
1336 per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
1339 #ifdef CONFIG_HOTPLUG_CPU
1341 remove_siblinginfo(int cpu)
1344 struct cpuinfo_x86 *c = cpu_data;
1346 for_each_cpu_mask(sibling, cpu_core_map[cpu]) {
1347 cpu_clear(cpu, cpu_core_map[sibling]);
1349 * last thread sibling in this cpu core going down
1351 if (cpus_weight(cpu_sibling_map[cpu]) == 1)
1352 c[sibling].booted_cores--;
1355 for_each_cpu_mask(sibling, cpu_sibling_map[cpu])
1356 cpu_clear(cpu, cpu_sibling_map[sibling]);
1357 cpus_clear(cpu_sibling_map[cpu]);
1358 cpus_clear(cpu_core_map[cpu]);
1359 c[cpu].phys_proc_id = 0;
1360 c[cpu].cpu_core_id = 0;
1361 cpu_clear(cpu, cpu_sibling_setup_map);
1364 int __cpu_disable(void)
1366 cpumask_t map = cpu_online_map;
1367 int cpu = smp_processor_id();
1370 * Perhaps use cpufreq to drop frequency, but that could go
1371 * into generic code.
1373 * We won't take down the boot processor on i386 due to some
1374 * interrupts only being able to be serviced by the BSP.
1375 * Especially so if we're not using an IOAPIC -zwane
1381 /* Allow any queued timer interrupts to get serviced */
1384 local_irq_disable();
1386 remove_siblinginfo(cpu);
1388 cpu_clear(cpu, map);
1390 /* It's now safe to remove this processor from the online map */
1391 cpu_clear(cpu, cpu_online_map);
1395 void __cpu_die(unsigned int cpu)
1397 /* We don't do anything here: idle task is faking death itself. */
1400 for (i = 0; i < 10; i++) {
1401 /* They ack this in play_dead by setting CPU_DEAD */
1402 if (per_cpu(cpu_state, cpu) == CPU_DEAD) {
1403 printk ("CPU %d is now offline\n", cpu);
1404 if (1 == num_online_cpus())
1405 alternatives_smp_switch(0);
1410 printk(KERN_ERR "CPU %u didn't die...\n", cpu);
1412 #else /* ... !CONFIG_HOTPLUG_CPU */
1413 int __cpu_disable(void)
1418 void __cpu_die(unsigned int cpu)
1420 /* We said "no" in __cpu_disable */
1423 #endif /* CONFIG_HOTPLUG_CPU */
1425 int __devinit __cpu_up(unsigned int cpu)
1427 #ifdef CONFIG_HOTPLUG_CPU
1431 * We do warm boot only on cpus that had booted earlier
1432 * Otherwise cold boot is all handled from smp_boot_cpus().
1433 * cpu_callin_map is set during AP kickstart process. Its reset
1434 * when a cpu is taken offline from cpu_exit_clear().
1436 if (!cpu_isset(cpu, cpu_callin_map))
1437 ret = __smp_prepare_cpu(cpu);
1443 /* In case one didn't come up */
1444 if (!cpu_isset(cpu, cpu_callin_map)) {
1445 printk(KERN_DEBUG "skipping cpu%d, didn't come online\n", cpu);
1451 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
1452 /* Unleash the CPU! */
1453 cpu_set(cpu, smp_commenced_mask);
1454 while (!cpu_isset(cpu, cpu_online_map))
1459 void __init smp_cpus_done(unsigned int max_cpus)
1461 #ifdef CONFIG_X86_IO_APIC
1462 setup_ioapic_dest();
1465 #ifndef CONFIG_HOTPLUG_CPU
1467 * Disable executability of the SMP trampoline:
1469 set_kernel_exec((unsigned long)trampoline_base, trampoline_exec);
1473 void __init smp_intr_init(void)
1476 * IRQ0 must be given a fixed assignment and initialized,
1477 * because it's used before the IO-APIC is set up.
1479 set_intr_gate(FIRST_DEVICE_VECTOR, interrupt[0]);
1482 * The reschedule interrupt is a CPU-to-CPU reschedule-helper
1483 * IPI, driven by wakeup.
1485 set_intr_gate(RESCHEDULE_VECTOR, reschedule_interrupt);
1487 /* IPI for invalidation */
1488 set_intr_gate(INVALIDATE_TLB_VECTOR, invalidate_interrupt);
1490 /* IPI for generic function call */
1491 set_intr_gate(CALL_FUNCTION_VECTOR, call_function_interrupt);