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/config.h>
38 #include <linux/init.h>
39 #include <linux/kernel.h>
42 #include <linux/sched.h>
43 #include <linux/kernel_stat.h>
44 #include <linux/smp_lock.h>
45 #include <linux/bootmem.h>
46 #include <linux/notifier.h>
47 #include <linux/cpu.h>
48 #include <linux/percpu.h>
50 #include <linux/delay.h>
51 #include <linux/mc146818rtc.h>
52 #include <asm/tlbflush.h>
54 #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 /* Package ID of each logical CPU */
70 int phys_proc_id[NR_CPUS] __read_mostly = {[0 ... NR_CPUS-1] = BAD_APICID};
72 /* Core ID of each logical CPU */
73 int cpu_core_id[NR_CPUS] __read_mostly = {[0 ... NR_CPUS-1] = BAD_APICID};
75 /* representing HT siblings of each logical CPU */
76 cpumask_t cpu_sibling_map[NR_CPUS] __read_mostly;
77 EXPORT_SYMBOL(cpu_sibling_map);
79 /* representing HT and core siblings of each logical CPU */
80 cpumask_t cpu_core_map[NR_CPUS] __read_mostly;
81 EXPORT_SYMBOL(cpu_core_map);
83 /* bitmap of online cpus */
84 cpumask_t cpu_online_map __read_mostly;
85 EXPORT_SYMBOL(cpu_online_map);
87 cpumask_t cpu_callin_map;
88 cpumask_t cpu_callout_map;
89 EXPORT_SYMBOL(cpu_callout_map);
90 #ifdef CONFIG_HOTPLUG_CPU
91 cpumask_t cpu_possible_map = CPU_MASK_ALL;
93 cpumask_t cpu_possible_map;
95 EXPORT_SYMBOL(cpu_possible_map);
96 static cpumask_t smp_commenced_mask;
98 /* TSC's upper 32 bits can't be written in eariler CPU (before prescott), there
99 * is no way to resync one AP against BP. TBD: for prescott and above, we
100 * should use IA64's algorithm
102 static int __devinitdata tsc_sync_disabled;
104 /* Per CPU bogomips and other parameters */
105 struct cpuinfo_x86 cpu_data[NR_CPUS] __cacheline_aligned;
106 EXPORT_SYMBOL(cpu_data);
108 u8 x86_cpu_to_apicid[NR_CPUS] __read_mostly =
109 { [0 ... NR_CPUS-1] = 0xff };
110 EXPORT_SYMBOL(x86_cpu_to_apicid);
113 * Trampoline 80x86 program as an array.
116 extern unsigned char trampoline_data [];
117 extern unsigned char trampoline_end [];
118 static unsigned char *trampoline_base;
119 static int trampoline_exec;
121 static void map_cpu_to_logical_apicid(void);
123 /* State of each CPU. */
124 DEFINE_PER_CPU(int, cpu_state) = { 0 };
127 * Currently trivial. Write the real->protected mode
128 * bootstrap into the page concerned. The caller
129 * has made sure it's suitably aligned.
132 static unsigned long __devinit setup_trampoline(void)
134 memcpy(trampoline_base, trampoline_data, trampoline_end - trampoline_data);
135 return virt_to_phys(trampoline_base);
139 * We are called very early to get the low memory for the
140 * SMP bootup trampoline page.
142 void __init smp_alloc_memory(void)
144 trampoline_base = (void *) alloc_bootmem_low_pages(PAGE_SIZE);
146 * Has to be in very low memory so we can execute
149 if (__pa(trampoline_base) >= 0x9F000)
152 * Make the SMP trampoline executable:
154 trampoline_exec = set_kernel_exec((unsigned long)trampoline_base, 1);
158 * The bootstrap kernel entry code has set these up. Save them for
162 static void __devinit smp_store_cpu_info(int id)
164 struct cpuinfo_x86 *c = cpu_data + id;
170 * Mask B, Pentium, but not Pentium MMX
172 if (c->x86_vendor == X86_VENDOR_INTEL &&
174 c->x86_mask >= 1 && c->x86_mask <= 4 &&
177 * Remember we have B step Pentia with bugs
182 * Certain Athlons might work (for various values of 'work') in SMP
183 * but they are not certified as MP capable.
185 if ((c->x86_vendor == X86_VENDOR_AMD) && (c->x86 == 6)) {
187 /* Athlon 660/661 is valid. */
188 if ((c->x86_model==6) && ((c->x86_mask==0) || (c->x86_mask==1)))
191 /* Duron 670 is valid */
192 if ((c->x86_model==7) && (c->x86_mask==0))
196 * Athlon 662, Duron 671, and Athlon >model 7 have capability bit.
197 * It's worth noting that the A5 stepping (662) of some Athlon XP's
198 * have the MP bit set.
199 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for more.
201 if (((c->x86_model==6) && (c->x86_mask>=2)) ||
202 ((c->x86_model==7) && (c->x86_mask>=1)) ||
207 /* If we get here, it's not a certified SMP capable AMD system. */
208 add_taint(TAINT_UNSAFE_SMP);
216 * TSC synchronization.
218 * We first check whether all CPUs have their TSC's synchronized,
219 * then we print a warning if not, and always resync.
222 static atomic_t tsc_start_flag = ATOMIC_INIT(0);
223 static atomic_t tsc_count_start = ATOMIC_INIT(0);
224 static atomic_t tsc_count_stop = ATOMIC_INIT(0);
225 static unsigned long long tsc_values[NR_CPUS];
229 static void __init synchronize_tsc_bp (void)
232 unsigned long long t0;
233 unsigned long long sum, avg;
235 unsigned int one_usec;
238 printk(KERN_INFO "checking TSC synchronization across %u CPUs: ", num_booting_cpus());
240 /* convert from kcyc/sec to cyc/usec */
241 one_usec = cpu_khz / 1000;
243 atomic_set(&tsc_start_flag, 1);
247 * We loop a few times to get a primed instruction cache,
248 * then the last pass is more or less synchronized and
249 * the BP and APs set their cycle counters to zero all at
250 * once. This reduces the chance of having random offsets
251 * between the processors, and guarantees that the maximum
252 * delay between the cycle counters is never bigger than
253 * the latency of information-passing (cachelines) between
256 for (i = 0; i < NR_LOOPS; i++) {
258 * all APs synchronize but they loop on '== num_cpus'
260 while (atomic_read(&tsc_count_start) != num_booting_cpus()-1)
262 atomic_set(&tsc_count_stop, 0);
265 * this lets the APs save their current TSC:
267 atomic_inc(&tsc_count_start);
269 rdtscll(tsc_values[smp_processor_id()]);
271 * We clear the TSC in the last loop:
277 * Wait for all APs to leave the synchronization point:
279 while (atomic_read(&tsc_count_stop) != num_booting_cpus()-1)
281 atomic_set(&tsc_count_start, 0);
283 atomic_inc(&tsc_count_stop);
287 for (i = 0; i < NR_CPUS; i++) {
288 if (cpu_isset(i, cpu_callout_map)) {
294 do_div(avg, num_booting_cpus());
297 for (i = 0; i < NR_CPUS; i++) {
298 if (!cpu_isset(i, cpu_callout_map))
300 delta = tsc_values[i] - avg;
304 * We report bigger than 2 microseconds clock differences.
306 if (delta > 2*one_usec) {
313 do_div(realdelta, one_usec);
314 if (tsc_values[i] < avg)
315 realdelta = -realdelta;
317 printk(KERN_INFO "CPU#%d had %ld usecs TSC skew, fixed it up.\n", i, realdelta);
326 static void __init synchronize_tsc_ap (void)
331 * Not every cpu is online at the time
332 * this gets called, so we first wait for the BP to
333 * finish SMP initialization:
335 while (!atomic_read(&tsc_start_flag)) mb();
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()) mb();
352 extern void calibrate_delay(void);
354 static atomic_t init_deasserted;
356 static void __devinit smp_callin(void)
359 unsigned long timeout;
362 * If waken up by an INIT in an 82489DX configuration
363 * we may get here before an INIT-deassert IPI reaches
364 * our local APIC. We have to wait for the IPI or we'll
365 * lock up on an APIC access.
367 wait_for_init_deassert(&init_deasserted);
370 * (This works even if the APIC is not enabled.)
372 phys_id = GET_APIC_ID(apic_read(APIC_ID));
373 cpuid = smp_processor_id();
374 if (cpu_isset(cpuid, cpu_callin_map)) {
375 printk("huh, phys CPU#%d, CPU#%d already present??\n",
379 Dprintk("CPU#%d (phys ID: %d) waiting for CALLOUT\n", cpuid, phys_id);
382 * STARTUP IPIs are fragile beasts as they might sometimes
383 * trigger some glue motherboard logic. Complete APIC bus
384 * silence for 1 second, this overestimates the time the
385 * boot CPU is spending to send the up to 2 STARTUP IPIs
386 * by a factor of two. This should be enough.
390 * Waiting 2s total for startup (udelay is not yet working)
392 timeout = jiffies + 2*HZ;
393 while (time_before(jiffies, timeout)) {
395 * Has the boot CPU finished it's STARTUP sequence?
397 if (cpu_isset(cpuid, cpu_callout_map))
402 if (!time_before(jiffies, timeout)) {
403 printk("BUG: CPU%d started up but did not get a callout!\n",
409 * the boot CPU has finished the init stage and is spinning
410 * on callin_map until we finish. We are free to set up this
411 * CPU, first the APIC. (this is probably redundant on most
415 Dprintk("CALLIN, before setup_local_APIC().\n");
416 smp_callin_clear_local_apic();
418 map_cpu_to_logical_apicid();
424 Dprintk("Stack at about %p\n",&cpuid);
427 * Save our processor parameters
429 smp_store_cpu_info(cpuid);
431 disable_APIC_timer();
434 * Allow the master to continue.
436 cpu_set(cpuid, cpu_callin_map);
439 * Synchronize the TSC with the BP
441 if (cpu_has_tsc && cpu_khz && !tsc_sync_disabled)
442 synchronize_tsc_ap();
447 /* representing cpus for which sibling maps can be computed */
448 static cpumask_t cpu_sibling_setup_map;
451 set_cpu_sibling_map(int cpu)
454 struct cpuinfo_x86 *c = cpu_data;
456 cpu_set(cpu, cpu_sibling_setup_map);
458 if (smp_num_siblings > 1) {
459 for_each_cpu_mask(i, cpu_sibling_setup_map) {
460 if (phys_proc_id[cpu] == phys_proc_id[i] &&
461 cpu_core_id[cpu] == cpu_core_id[i]) {
462 cpu_set(i, cpu_sibling_map[cpu]);
463 cpu_set(cpu, cpu_sibling_map[i]);
464 cpu_set(i, cpu_core_map[cpu]);
465 cpu_set(cpu, cpu_core_map[i]);
469 cpu_set(cpu, cpu_sibling_map[cpu]);
472 if (current_cpu_data.x86_max_cores == 1) {
473 cpu_core_map[cpu] = cpu_sibling_map[cpu];
474 c[cpu].booted_cores = 1;
478 for_each_cpu_mask(i, cpu_sibling_setup_map) {
479 if (phys_proc_id[cpu] == phys_proc_id[i]) {
480 cpu_set(i, cpu_core_map[cpu]);
481 cpu_set(cpu, cpu_core_map[i]);
483 * Does this new cpu bringup a new core?
485 if (cpus_weight(cpu_sibling_map[cpu]) == 1) {
487 * for each core in package, increment
488 * the booted_cores for this new cpu
490 if (first_cpu(cpu_sibling_map[i]) == i)
491 c[cpu].booted_cores++;
493 * increment the core count for all
494 * the other cpus in this package
498 } else if (i != cpu && !c[cpu].booted_cores)
499 c[cpu].booted_cores = c[i].booted_cores;
505 * Activate a secondary processor.
507 static void __devinit start_secondary(void *unused)
510 * Dont put anything before smp_callin(), SMP
511 * booting is too fragile that we want to limit the
512 * things done here to the most necessary things.
517 while (!cpu_isset(smp_processor_id(), smp_commenced_mask))
519 setup_secondary_APIC_clock();
520 if (nmi_watchdog == NMI_IO_APIC) {
521 disable_8259A_irq(0);
522 enable_NMI_through_LVT0(NULL);
527 * low-memory mappings have been cleared, flush them from
528 * the local TLBs too.
532 /* This must be done before setting cpu_online_map */
533 set_cpu_sibling_map(raw_smp_processor_id());
537 * We need to hold call_lock, so there is no inconsistency
538 * between the time smp_call_function() determines number of
539 * IPI receipients, and the time when the determination is made
540 * for which cpus receive the IPI. Holding this
541 * lock helps us to not include this cpu in a currently in progress
542 * smp_call_function().
544 lock_ipi_call_lock();
545 cpu_set(smp_processor_id(), cpu_online_map);
546 unlock_ipi_call_lock();
547 per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
549 /* We can take interrupts now: we're officially "up". */
557 * Everything has been set up for the secondary
558 * CPUs - they just need to reload everything
559 * from the task structure
560 * This function must not return.
562 void __devinit initialize_secondary(void)
565 * We don't actually need to load the full TSS,
566 * basically just the stack pointer and the eip.
573 :"r" (current->thread.esp),"r" (current->thread.eip));
583 /* which logical CPUs are on which nodes */
584 cpumask_t node_2_cpu_mask[MAX_NUMNODES] __read_mostly =
585 { [0 ... MAX_NUMNODES-1] = CPU_MASK_NONE };
586 /* which node each logical CPU is on */
587 int cpu_2_node[NR_CPUS] __read_mostly = { [0 ... NR_CPUS-1] = 0 };
588 EXPORT_SYMBOL(cpu_2_node);
590 /* set up a mapping between cpu and node. */
591 static inline void map_cpu_to_node(int cpu, int node)
593 printk("Mapping cpu %d to node %d\n", cpu, node);
594 cpu_set(cpu, node_2_cpu_mask[node]);
595 cpu_2_node[cpu] = node;
598 /* undo a mapping between cpu and node. */
599 static inline void unmap_cpu_to_node(int cpu)
603 printk("Unmapping cpu %d from all nodes\n", cpu);
604 for (node = 0; node < MAX_NUMNODES; node ++)
605 cpu_clear(cpu, node_2_cpu_mask[node]);
608 #else /* !CONFIG_NUMA */
610 #define map_cpu_to_node(cpu, node) ({})
611 #define unmap_cpu_to_node(cpu) ({})
613 #endif /* CONFIG_NUMA */
615 u8 cpu_2_logical_apicid[NR_CPUS] __read_mostly = { [0 ... NR_CPUS-1] = BAD_APICID };
617 static void map_cpu_to_logical_apicid(void)
619 int cpu = smp_processor_id();
620 int apicid = logical_smp_processor_id();
622 cpu_2_logical_apicid[cpu] = apicid;
623 map_cpu_to_node(cpu, apicid_to_node(apicid));
626 static void unmap_cpu_to_logical_apicid(int cpu)
628 cpu_2_logical_apicid[cpu] = BAD_APICID;
629 unmap_cpu_to_node(cpu);
633 static inline void __inquire_remote_apic(int apicid)
635 int i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
636 char *names[] = { "ID", "VERSION", "SPIV" };
639 printk("Inquiring remote APIC #%d...\n", apicid);
641 for (i = 0; i < ARRAY_SIZE(regs); i++) {
642 printk("... APIC #%d %s: ", apicid, names[i]);
647 apic_wait_icr_idle();
649 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(apicid));
650 apic_write_around(APIC_ICR, APIC_DM_REMRD | regs[i]);
655 status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
656 } while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
659 case APIC_ICR_RR_VALID:
660 status = apic_read(APIC_RRR);
661 printk("%08x\n", status);
670 #ifdef WAKE_SECONDARY_VIA_NMI
672 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
673 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
674 * won't ... remember to clear down the APIC, etc later.
677 wakeup_secondary_cpu(int logical_apicid, unsigned long start_eip)
679 unsigned long send_status = 0, accept_status = 0;
683 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(logical_apicid));
685 /* Boot on the stack */
686 /* Kick the second */
687 apic_write_around(APIC_ICR, APIC_DM_NMI | APIC_DEST_LOGICAL);
689 Dprintk("Waiting for send to finish...\n");
694 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
695 } while (send_status && (timeout++ < 1000));
698 * Give the other CPU some time to accept the IPI.
702 * Due to the Pentium erratum 3AP.
704 maxlvt = get_maxlvt();
706 apic_read_around(APIC_SPIV);
707 apic_write(APIC_ESR, 0);
709 accept_status = (apic_read(APIC_ESR) & 0xEF);
710 Dprintk("NMI sent.\n");
713 printk("APIC never delivered???\n");
715 printk("APIC delivery error (%lx).\n", accept_status);
717 return (send_status | accept_status);
719 #endif /* WAKE_SECONDARY_VIA_NMI */
721 #ifdef WAKE_SECONDARY_VIA_INIT
723 wakeup_secondary_cpu(int phys_apicid, unsigned long start_eip)
725 unsigned long send_status = 0, accept_status = 0;
726 int maxlvt, timeout, num_starts, j;
729 * Be paranoid about clearing APIC errors.
731 if (APIC_INTEGRATED(apic_version[phys_apicid])) {
732 apic_read_around(APIC_SPIV);
733 apic_write(APIC_ESR, 0);
737 Dprintk("Asserting INIT.\n");
740 * Turn INIT on target chip
742 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
747 apic_write_around(APIC_ICR, APIC_INT_LEVELTRIG | APIC_INT_ASSERT
750 Dprintk("Waiting for send to finish...\n");
755 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
756 } while (send_status && (timeout++ < 1000));
760 Dprintk("Deasserting INIT.\n");
763 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
766 apic_write_around(APIC_ICR, APIC_INT_LEVELTRIG | APIC_DM_INIT);
768 Dprintk("Waiting for send to finish...\n");
773 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
774 } while (send_status && (timeout++ < 1000));
776 atomic_set(&init_deasserted, 1);
779 * Should we send STARTUP IPIs ?
781 * Determine this based on the APIC version.
782 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
784 if (APIC_INTEGRATED(apic_version[phys_apicid]))
790 * Run STARTUP IPI loop.
792 Dprintk("#startup loops: %d.\n", num_starts);
794 maxlvt = get_maxlvt();
796 for (j = 1; j <= num_starts; j++) {
797 Dprintk("Sending STARTUP #%d.\n",j);
798 apic_read_around(APIC_SPIV);
799 apic_write(APIC_ESR, 0);
801 Dprintk("After apic_write.\n");
808 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
810 /* Boot on the stack */
811 /* Kick the second */
812 apic_write_around(APIC_ICR, APIC_DM_STARTUP
813 | (start_eip >> 12));
816 * Give the other CPU some time to accept the IPI.
820 Dprintk("Startup point 1.\n");
822 Dprintk("Waiting for send to finish...\n");
827 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
828 } while (send_status && (timeout++ < 1000));
831 * Give the other CPU some time to accept the IPI.
835 * Due to the Pentium erratum 3AP.
838 apic_read_around(APIC_SPIV);
839 apic_write(APIC_ESR, 0);
841 accept_status = (apic_read(APIC_ESR) & 0xEF);
842 if (send_status || accept_status)
845 Dprintk("After Startup.\n");
848 printk("APIC never delivered???\n");
850 printk("APIC delivery error (%lx).\n", accept_status);
852 return (send_status | accept_status);
854 #endif /* WAKE_SECONDARY_VIA_INIT */
856 extern cpumask_t cpu_initialized;
857 static inline int alloc_cpu_id(void)
861 cpus_complement(tmp_map, cpu_present_map);
862 cpu = first_cpu(tmp_map);
868 #ifdef CONFIG_HOTPLUG_CPU
869 static struct task_struct * __devinitdata cpu_idle_tasks[NR_CPUS];
870 static inline struct task_struct * alloc_idle_task(int cpu)
872 struct task_struct *idle;
874 if ((idle = cpu_idle_tasks[cpu]) != NULL) {
875 /* initialize thread_struct. we really want to avoid destroy
878 idle->thread.esp = (unsigned long)(((struct pt_regs *)
879 (THREAD_SIZE + (unsigned long) idle->thread_info)) - 1);
880 init_idle(idle, cpu);
883 idle = fork_idle(cpu);
886 cpu_idle_tasks[cpu] = idle;
890 #define alloc_idle_task(cpu) fork_idle(cpu)
893 static int __devinit do_boot_cpu(int apicid, int cpu)
895 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
896 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
897 * Returns zero if CPU booted OK, else error code from wakeup_secondary_cpu.
900 struct task_struct *idle;
901 unsigned long boot_error;
903 unsigned long start_eip;
904 unsigned short nmi_high = 0, nmi_low = 0;
906 if (!cpu_gdt_descr[cpu].address &&
907 !(cpu_gdt_descr[cpu].address = get_zeroed_page(GFP_KERNEL))) {
908 printk("Failed to allocate GDT for CPU %d\n", cpu);
915 * We can't use kernel_thread since we must avoid to
916 * reschedule the child.
918 idle = alloc_idle_task(cpu);
920 panic("failed fork for CPU %d", cpu);
921 idle->thread.eip = (unsigned long) start_secondary;
922 /* start_eip had better be page-aligned! */
923 start_eip = setup_trampoline();
925 /* So we see what's up */
926 printk("Booting processor %d/%d eip %lx\n", cpu, apicid, start_eip);
927 /* Stack for startup_32 can be just as for start_secondary onwards */
928 stack_start.esp = (void *) idle->thread.esp;
933 * This grunge runs the startup process for
934 * the targeted processor.
937 atomic_set(&init_deasserted, 0);
939 Dprintk("Setting warm reset code and vector.\n");
941 store_NMI_vector(&nmi_high, &nmi_low);
943 smpboot_setup_warm_reset_vector(start_eip);
946 * Starting actual IPI sequence...
948 boot_error = wakeup_secondary_cpu(apicid, start_eip);
952 * allow APs to start initializing.
954 Dprintk("Before Callout %d.\n", cpu);
955 cpu_set(cpu, cpu_callout_map);
956 Dprintk("After Callout %d.\n", cpu);
959 * Wait 5s total for a response
961 for (timeout = 0; timeout < 50000; timeout++) {
962 if (cpu_isset(cpu, cpu_callin_map))
963 break; /* It has booted */
967 if (cpu_isset(cpu, cpu_callin_map)) {
968 /* number CPUs logically, starting from 1 (BSP is 0) */
970 printk("CPU%d: ", cpu);
971 print_cpu_info(&cpu_data[cpu]);
972 Dprintk("CPU has booted.\n");
975 if (*((volatile unsigned char *)trampoline_base)
977 /* trampoline started but...? */
978 printk("Stuck ??\n");
980 /* trampoline code not run */
981 printk("Not responding.\n");
982 inquire_remote_apic(apicid);
987 /* Try to put things back the way they were before ... */
988 unmap_cpu_to_logical_apicid(cpu);
989 cpu_clear(cpu, cpu_callout_map); /* was set here (do_boot_cpu()) */
990 cpu_clear(cpu, cpu_initialized); /* was set by cpu_init() */
993 x86_cpu_to_apicid[cpu] = apicid;
994 cpu_set(cpu, cpu_present_map);
997 /* mark "stuck" area as not stuck */
998 *((volatile unsigned long *)trampoline_base) = 0;
1003 #ifdef CONFIG_HOTPLUG_CPU
1004 void cpu_exit_clear(void)
1006 int cpu = raw_smp_processor_id();
1014 cpu_clear(cpu, cpu_callout_map);
1015 cpu_clear(cpu, cpu_callin_map);
1016 cpu_clear(cpu, cpu_present_map);
1018 cpu_clear(cpu, smp_commenced_mask);
1019 unmap_cpu_to_logical_apicid(cpu);
1022 struct warm_boot_cpu_info {
1023 struct completion *complete;
1028 static void __devinit do_warm_boot_cpu(void *p)
1030 struct warm_boot_cpu_info *info = p;
1031 do_boot_cpu(info->apicid, info->cpu);
1032 complete(info->complete);
1035 int __devinit smp_prepare_cpu(int cpu)
1037 DECLARE_COMPLETION(done);
1038 struct warm_boot_cpu_info info;
1039 struct work_struct task;
1043 apicid = x86_cpu_to_apicid[cpu];
1044 if (apicid == BAD_APICID) {
1049 info.complete = &done;
1050 info.apicid = apicid;
1052 INIT_WORK(&task, do_warm_boot_cpu, &info);
1054 tsc_sync_disabled = 1;
1056 /* init low mem mapping */
1057 clone_pgd_range(swapper_pg_dir, swapper_pg_dir + USER_PGD_PTRS,
1060 schedule_work(&task);
1061 wait_for_completion(&done);
1063 tsc_sync_disabled = 0;
1067 unlock_cpu_hotplug();
1072 static void smp_tune_scheduling (void)
1074 unsigned long cachesize; /* kB */
1075 unsigned long bandwidth = 350; /* MB/s */
1077 * Rough estimation for SMP scheduling, this is the number of
1078 * cycles it takes for a fully memory-limited process to flush
1079 * the SMP-local cache.
1081 * (For a P5 this pretty much means we will choose another idle
1082 * CPU almost always at wakeup time (this is due to the small
1083 * L1 cache), on PIIs it's around 50-100 usecs, depending on
1089 * this basically disables processor-affinity
1090 * scheduling on SMP without a TSC.
1094 cachesize = boot_cpu_data.x86_cache_size;
1095 if (cachesize == -1) {
1096 cachesize = 16; /* Pentiums, 2x8kB cache */
1103 * Cycle through the processors sending APIC IPIs to boot each.
1106 static int boot_cpu_logical_apicid;
1107 /* Where the IO area was mapped on multiquad, always 0 otherwise */
1109 #ifdef CONFIG_X86_NUMAQ
1110 EXPORT_SYMBOL(xquad_portio);
1113 static void __init smp_boot_cpus(unsigned int max_cpus)
1115 int apicid, cpu, bit, kicked;
1116 unsigned long bogosum = 0;
1119 * Setup boot CPU information
1121 smp_store_cpu_info(0); /* Final full version of the data */
1122 printk("CPU%d: ", 0);
1123 print_cpu_info(&cpu_data[0]);
1125 boot_cpu_physical_apicid = GET_APIC_ID(apic_read(APIC_ID));
1126 boot_cpu_logical_apicid = logical_smp_processor_id();
1127 x86_cpu_to_apicid[0] = boot_cpu_physical_apicid;
1129 current_thread_info()->cpu = 0;
1130 smp_tune_scheduling();
1132 set_cpu_sibling_map(0);
1135 * If we couldn't find an SMP configuration at boot time,
1136 * get out of here now!
1138 if (!smp_found_config && !acpi_lapic) {
1139 printk(KERN_NOTICE "SMP motherboard not detected.\n");
1140 smpboot_clear_io_apic_irqs();
1141 phys_cpu_present_map = physid_mask_of_physid(0);
1142 if (APIC_init_uniprocessor())
1143 printk(KERN_NOTICE "Local APIC not detected."
1144 " Using dummy APIC emulation.\n");
1145 map_cpu_to_logical_apicid();
1146 cpu_set(0, cpu_sibling_map[0]);
1147 cpu_set(0, cpu_core_map[0]);
1152 * Should not be necessary because the MP table should list the boot
1153 * CPU too, but we do it for the sake of robustness anyway.
1154 * Makes no sense to do this check in clustered apic mode, so skip it
1156 if (!check_phys_apicid_present(boot_cpu_physical_apicid)) {
1157 printk("weird, boot CPU (#%d) not listed by the BIOS.\n",
1158 boot_cpu_physical_apicid);
1159 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1163 * If we couldn't find a local APIC, then get out of here now!
1165 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid]) && !cpu_has_apic) {
1166 printk(KERN_ERR "BIOS bug, local APIC #%d not detected!...\n",
1167 boot_cpu_physical_apicid);
1168 printk(KERN_ERR "... forcing use of dummy APIC emulation. (tell your hw vendor)\n");
1169 smpboot_clear_io_apic_irqs();
1170 phys_cpu_present_map = physid_mask_of_physid(0);
1171 cpu_set(0, cpu_sibling_map[0]);
1172 cpu_set(0, cpu_core_map[0]);
1176 verify_local_APIC();
1179 * If SMP should be disabled, then really disable it!
1182 smp_found_config = 0;
1183 printk(KERN_INFO "SMP mode deactivated, forcing use of dummy APIC emulation.\n");
1184 smpboot_clear_io_apic_irqs();
1185 phys_cpu_present_map = physid_mask_of_physid(0);
1186 cpu_set(0, cpu_sibling_map[0]);
1187 cpu_set(0, cpu_core_map[0]);
1193 map_cpu_to_logical_apicid();
1196 setup_portio_remap();
1199 * Scan the CPU present map and fire up the other CPUs via do_boot_cpu
1201 * In clustered apic mode, phys_cpu_present_map is a constructed thus:
1202 * bits 0-3 are quad0, 4-7 are quad1, etc. A perverse twist on the
1203 * clustered apic ID.
1205 Dprintk("CPU present map: %lx\n", physids_coerce(phys_cpu_present_map));
1208 for (bit = 0; kicked < NR_CPUS && bit < MAX_APICS; bit++) {
1209 apicid = cpu_present_to_apicid(bit);
1211 * Don't even attempt to start the boot CPU!
1213 if ((apicid == boot_cpu_apicid) || (apicid == BAD_APICID))
1216 if (!check_apicid_present(bit))
1218 if (max_cpus <= cpucount+1)
1221 if (((cpu = alloc_cpu_id()) <= 0) || do_boot_cpu(apicid, cpu))
1222 printk("CPU #%d not responding - cannot use it.\n",
1229 * Cleanup possible dangling ends...
1231 smpboot_restore_warm_reset_vector();
1234 * Allow the user to impress friends.
1236 Dprintk("Before bogomips.\n");
1237 for (cpu = 0; cpu < NR_CPUS; cpu++)
1238 if (cpu_isset(cpu, cpu_callout_map))
1239 bogosum += cpu_data[cpu].loops_per_jiffy;
1241 "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
1243 bogosum/(500000/HZ),
1244 (bogosum/(5000/HZ))%100);
1246 Dprintk("Before bogocount - setting activated=1.\n");
1249 printk(KERN_WARNING "WARNING: SMP operation may be unreliable with B stepping processors.\n");
1252 * Don't taint if we are running SMP kernel on a single non-MP
1255 if (tainted & TAINT_UNSAFE_SMP) {
1257 printk (KERN_INFO "WARNING: This combination of AMD processors is not suitable for SMP.\n");
1259 tainted &= ~TAINT_UNSAFE_SMP;
1262 Dprintk("Boot done.\n");
1265 * construct cpu_sibling_map[], so that we can tell sibling CPUs
1268 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1269 cpus_clear(cpu_sibling_map[cpu]);
1270 cpus_clear(cpu_core_map[cpu]);
1273 cpu_set(0, cpu_sibling_map[0]);
1274 cpu_set(0, cpu_core_map[0]);
1276 smpboot_setup_io_apic();
1278 setup_boot_APIC_clock();
1281 * Synchronize the TSC with the AP
1283 if (cpu_has_tsc && cpucount && cpu_khz)
1284 synchronize_tsc_bp();
1287 /* These are wrappers to interface to the new boot process. Someone
1288 who understands all this stuff should rewrite it properly. --RR 15/Jul/02 */
1289 void __init smp_prepare_cpus(unsigned int max_cpus)
1291 smp_commenced_mask = cpumask_of_cpu(0);
1292 cpu_callin_map = cpumask_of_cpu(0);
1294 smp_boot_cpus(max_cpus);
1297 void __devinit smp_prepare_boot_cpu(void)
1299 cpu_set(smp_processor_id(), cpu_online_map);
1300 cpu_set(smp_processor_id(), cpu_callout_map);
1301 cpu_set(smp_processor_id(), cpu_present_map);
1302 cpu_set(smp_processor_id(), cpu_possible_map);
1303 per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
1306 #ifdef CONFIG_HOTPLUG_CPU
1308 remove_siblinginfo(int cpu)
1311 struct cpuinfo_x86 *c = cpu_data;
1313 for_each_cpu_mask(sibling, cpu_core_map[cpu]) {
1314 cpu_clear(cpu, cpu_core_map[sibling]);
1316 * last thread sibling in this cpu core going down
1318 if (cpus_weight(cpu_sibling_map[cpu]) == 1)
1319 c[sibling].booted_cores--;
1322 for_each_cpu_mask(sibling, cpu_sibling_map[cpu])
1323 cpu_clear(cpu, cpu_sibling_map[sibling]);
1324 cpus_clear(cpu_sibling_map[cpu]);
1325 cpus_clear(cpu_core_map[cpu]);
1326 phys_proc_id[cpu] = BAD_APICID;
1327 cpu_core_id[cpu] = BAD_APICID;
1328 cpu_clear(cpu, cpu_sibling_setup_map);
1331 int __cpu_disable(void)
1333 cpumask_t map = cpu_online_map;
1334 int cpu = smp_processor_id();
1337 * Perhaps use cpufreq to drop frequency, but that could go
1338 * into generic code.
1340 * We won't take down the boot processor on i386 due to some
1341 * interrupts only being able to be serviced by the BSP.
1342 * Especially so if we're not using an IOAPIC -zwane
1348 /* Allow any queued timer interrupts to get serviced */
1351 local_irq_disable();
1353 remove_siblinginfo(cpu);
1355 cpu_clear(cpu, map);
1357 /* It's now safe to remove this processor from the online map */
1358 cpu_clear(cpu, cpu_online_map);
1362 void __cpu_die(unsigned int cpu)
1364 /* We don't do anything here: idle task is faking death itself. */
1367 for (i = 0; i < 10; i++) {
1368 /* They ack this in play_dead by setting CPU_DEAD */
1369 if (per_cpu(cpu_state, cpu) == CPU_DEAD) {
1370 printk ("CPU %d is now offline\n", cpu);
1375 printk(KERN_ERR "CPU %u didn't die...\n", cpu);
1377 #else /* ... !CONFIG_HOTPLUG_CPU */
1378 int __cpu_disable(void)
1383 void __cpu_die(unsigned int cpu)
1385 /* We said "no" in __cpu_disable */
1388 #endif /* CONFIG_HOTPLUG_CPU */
1390 int __devinit __cpu_up(unsigned int cpu)
1392 /* In case one didn't come up */
1393 if (!cpu_isset(cpu, cpu_callin_map)) {
1394 printk(KERN_DEBUG "skipping cpu%d, didn't come online\n", cpu);
1400 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
1401 /* Unleash the CPU! */
1402 cpu_set(cpu, smp_commenced_mask);
1403 while (!cpu_isset(cpu, cpu_online_map))
1408 void __init smp_cpus_done(unsigned int max_cpus)
1410 #ifdef CONFIG_X86_IO_APIC
1411 setup_ioapic_dest();
1414 #ifndef CONFIG_HOTPLUG_CPU
1416 * Disable executability of the SMP trampoline:
1418 set_kernel_exec((unsigned long)trampoline_base, trampoline_exec);
1422 void __init smp_intr_init(void)
1425 * IRQ0 must be given a fixed assignment and initialized,
1426 * because it's used before the IO-APIC is set up.
1428 set_intr_gate(FIRST_DEVICE_VECTOR, interrupt[0]);
1431 * The reschedule interrupt is a CPU-to-CPU reschedule-helper
1432 * IPI, driven by wakeup.
1434 set_intr_gate(RESCHEDULE_VECTOR, reschedule_interrupt);
1436 /* IPI for invalidation */
1437 set_intr_gate(INVALIDATE_TLB_VECTOR, invalidate_interrupt);
1439 /* IPI for generic function call */
1440 set_intr_gate(CALL_FUNCTION_VECTOR, call_function_interrupt);