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/irq.h>
46 #include <linux/bootmem.h>
47 #include <linux/notifier.h>
48 #include <linux/cpu.h>
49 #include <linux/percpu.h>
51 #include <linux/delay.h>
52 #include <linux/mc146818rtc.h>
53 #include <asm/tlbflush.h>
55 #include <asm/arch_hooks.h>
57 #include <mach_apic.h>
58 #include <mach_wakecpu.h>
59 #include <smpboot_hooks.h>
61 /* Set if we find a B stepping CPU */
62 static int __devinitdata smp_b_stepping;
64 /* Number of siblings per CPU package */
65 int smp_num_siblings = 1;
67 EXPORT_SYMBOL(smp_num_siblings);
70 /* Package ID of each logical CPU */
71 int phys_proc_id[NR_CPUS] = {[0 ... NR_CPUS-1] = BAD_APICID};
72 EXPORT_SYMBOL(phys_proc_id);
74 /* Core ID of each logical CPU */
75 int cpu_core_id[NR_CPUS] = {[0 ... NR_CPUS-1] = BAD_APICID};
76 EXPORT_SYMBOL(cpu_core_id);
78 cpumask_t cpu_sibling_map[NR_CPUS];
79 EXPORT_SYMBOL(cpu_sibling_map);
81 cpumask_t cpu_core_map[NR_CPUS];
82 EXPORT_SYMBOL(cpu_core_map);
84 /* bitmap of online cpus */
85 cpumask_t cpu_online_map;
86 EXPORT_SYMBOL(cpu_online_map);
88 cpumask_t cpu_callin_map;
89 cpumask_t cpu_callout_map;
90 EXPORT_SYMBOL(cpu_callout_map);
91 static cpumask_t smp_commenced_mask;
93 /* TSC's upper 32 bits can't be written in eariler CPU (before prescott), there
94 * is no way to resync one AP against BP. TBD: for prescott and above, we
95 * should use IA64's algorithm
97 static int __devinitdata tsc_sync_disabled;
99 /* Per CPU bogomips and other parameters */
100 struct cpuinfo_x86 cpu_data[NR_CPUS] __cacheline_aligned;
101 EXPORT_SYMBOL(cpu_data);
103 u8 x86_cpu_to_apicid[NR_CPUS] =
104 { [0 ... NR_CPUS-1] = 0xff };
105 EXPORT_SYMBOL(x86_cpu_to_apicid);
108 * Trampoline 80x86 program as an array.
111 extern unsigned char trampoline_data [];
112 extern unsigned char trampoline_end [];
113 static unsigned char *trampoline_base;
114 static int trampoline_exec;
116 static void map_cpu_to_logical_apicid(void);
118 /* State of each CPU. */
119 DEFINE_PER_CPU(int, cpu_state) = { 0 };
122 * Currently trivial. Write the real->protected mode
123 * bootstrap into the page concerned. The caller
124 * has made sure it's suitably aligned.
127 static unsigned long __devinit setup_trampoline(void)
129 memcpy(trampoline_base, trampoline_data, trampoline_end - trampoline_data);
130 return virt_to_phys(trampoline_base);
134 * We are called very early to get the low memory for the
135 * SMP bootup trampoline page.
137 void __init smp_alloc_memory(void)
139 trampoline_base = (void *) alloc_bootmem_low_pages(PAGE_SIZE);
141 * Has to be in very low memory so we can execute
144 if (__pa(trampoline_base) >= 0x9F000)
147 * Make the SMP trampoline executable:
149 trampoline_exec = set_kernel_exec((unsigned long)trampoline_base, 1);
153 * The bootstrap kernel entry code has set these up. Save them for
157 static void __devinit smp_store_cpu_info(int id)
159 struct cpuinfo_x86 *c = cpu_data + id;
165 * Mask B, Pentium, but not Pentium MMX
167 if (c->x86_vendor == X86_VENDOR_INTEL &&
169 c->x86_mask >= 1 && c->x86_mask <= 4 &&
172 * Remember we have B step Pentia with bugs
177 * Certain Athlons might work (for various values of 'work') in SMP
178 * but they are not certified as MP capable.
180 if ((c->x86_vendor == X86_VENDOR_AMD) && (c->x86 == 6)) {
182 /* Athlon 660/661 is valid. */
183 if ((c->x86_model==6) && ((c->x86_mask==0) || (c->x86_mask==1)))
186 /* Duron 670 is valid */
187 if ((c->x86_model==7) && (c->x86_mask==0))
191 * Athlon 662, Duron 671, and Athlon >model 7 have capability bit.
192 * It's worth noting that the A5 stepping (662) of some Athlon XP's
193 * have the MP bit set.
194 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for more.
196 if (((c->x86_model==6) && (c->x86_mask>=2)) ||
197 ((c->x86_model==7) && (c->x86_mask>=1)) ||
202 /* If we get here, it's not a certified SMP capable AMD system. */
203 tainted |= TAINT_UNSAFE_SMP;
211 * TSC synchronization.
213 * We first check whether all CPUs have their TSC's synchronized,
214 * then we print a warning if not, and always resync.
217 static atomic_t tsc_start_flag = ATOMIC_INIT(0);
218 static atomic_t tsc_count_start = ATOMIC_INIT(0);
219 static atomic_t tsc_count_stop = ATOMIC_INIT(0);
220 static unsigned long long tsc_values[NR_CPUS];
224 static void __init synchronize_tsc_bp (void)
227 unsigned long long t0;
228 unsigned long long sum, avg;
230 unsigned int one_usec;
233 printk(KERN_INFO "checking TSC synchronization across %u CPUs: ", num_booting_cpus());
235 /* convert from kcyc/sec to cyc/usec */
236 one_usec = cpu_khz / 1000;
238 atomic_set(&tsc_start_flag, 1);
242 * We loop a few times to get a primed instruction cache,
243 * then the last pass is more or less synchronized and
244 * the BP and APs set their cycle counters to zero all at
245 * once. This reduces the chance of having random offsets
246 * between the processors, and guarantees that the maximum
247 * delay between the cycle counters is never bigger than
248 * the latency of information-passing (cachelines) between
251 for (i = 0; i < NR_LOOPS; i++) {
253 * all APs synchronize but they loop on '== num_cpus'
255 while (atomic_read(&tsc_count_start) != num_booting_cpus()-1)
257 atomic_set(&tsc_count_stop, 0);
260 * this lets the APs save their current TSC:
262 atomic_inc(&tsc_count_start);
264 rdtscll(tsc_values[smp_processor_id()]);
266 * We clear the TSC in the last loop:
272 * Wait for all APs to leave the synchronization point:
274 while (atomic_read(&tsc_count_stop) != num_booting_cpus()-1)
276 atomic_set(&tsc_count_start, 0);
278 atomic_inc(&tsc_count_stop);
282 for (i = 0; i < NR_CPUS; i++) {
283 if (cpu_isset(i, cpu_callout_map)) {
289 do_div(avg, num_booting_cpus());
292 for (i = 0; i < NR_CPUS; i++) {
293 if (!cpu_isset(i, cpu_callout_map))
295 delta = tsc_values[i] - avg;
299 * We report bigger than 2 microseconds clock differences.
301 if (delta > 2*one_usec) {
308 do_div(realdelta, one_usec);
309 if (tsc_values[i] < avg)
310 realdelta = -realdelta;
312 printk(KERN_INFO "CPU#%d had %ld usecs TSC skew, fixed it up.\n", i, realdelta);
321 static void __init synchronize_tsc_ap (void)
326 * Not every cpu is online at the time
327 * this gets called, so we first wait for the BP to
328 * finish SMP initialization:
330 while (!atomic_read(&tsc_start_flag)) mb();
332 for (i = 0; i < NR_LOOPS; i++) {
333 atomic_inc(&tsc_count_start);
334 while (atomic_read(&tsc_count_start) != num_booting_cpus())
337 rdtscll(tsc_values[smp_processor_id()]);
341 atomic_inc(&tsc_count_stop);
342 while (atomic_read(&tsc_count_stop) != num_booting_cpus()) mb();
347 extern void calibrate_delay(void);
349 static atomic_t init_deasserted;
351 static void __devinit smp_callin(void)
354 unsigned long timeout;
357 * If waken up by an INIT in an 82489DX configuration
358 * we may get here before an INIT-deassert IPI reaches
359 * our local APIC. We have to wait for the IPI or we'll
360 * lock up on an APIC access.
362 wait_for_init_deassert(&init_deasserted);
365 * (This works even if the APIC is not enabled.)
367 phys_id = GET_APIC_ID(apic_read(APIC_ID));
368 cpuid = smp_processor_id();
369 if (cpu_isset(cpuid, cpu_callin_map)) {
370 printk("huh, phys CPU#%d, CPU#%d already present??\n",
374 Dprintk("CPU#%d (phys ID: %d) waiting for CALLOUT\n", cpuid, phys_id);
377 * STARTUP IPIs are fragile beasts as they might sometimes
378 * trigger some glue motherboard logic. Complete APIC bus
379 * silence for 1 second, this overestimates the time the
380 * boot CPU is spending to send the up to 2 STARTUP IPIs
381 * by a factor of two. This should be enough.
385 * Waiting 2s total for startup (udelay is not yet working)
387 timeout = jiffies + 2*HZ;
388 while (time_before(jiffies, timeout)) {
390 * Has the boot CPU finished it's STARTUP sequence?
392 if (cpu_isset(cpuid, cpu_callout_map))
397 if (!time_before(jiffies, timeout)) {
398 printk("BUG: CPU%d started up but did not get a callout!\n",
404 * the boot CPU has finished the init stage and is spinning
405 * on callin_map until we finish. We are free to set up this
406 * CPU, first the APIC. (this is probably redundant on most
410 Dprintk("CALLIN, before setup_local_APIC().\n");
411 smp_callin_clear_local_apic();
413 map_cpu_to_logical_apicid();
419 Dprintk("Stack at about %p\n",&cpuid);
422 * Save our processor parameters
424 smp_store_cpu_info(cpuid);
426 disable_APIC_timer();
429 * Allow the master to continue.
431 cpu_set(cpuid, cpu_callin_map);
434 * Synchronize the TSC with the BP
436 if (cpu_has_tsc && cpu_khz && !tsc_sync_disabled)
437 synchronize_tsc_ap();
443 set_cpu_sibling_map(int cpu)
447 if (smp_num_siblings > 1) {
448 for (i = 0; i < NR_CPUS; i++) {
449 if (!cpu_isset(i, cpu_callout_map))
451 if (cpu_core_id[cpu] == cpu_core_id[i]) {
452 cpu_set(i, cpu_sibling_map[cpu]);
453 cpu_set(cpu, cpu_sibling_map[i]);
457 cpu_set(cpu, cpu_sibling_map[cpu]);
460 if (current_cpu_data.x86_num_cores > 1) {
461 for (i = 0; i < NR_CPUS; i++) {
462 if (!cpu_isset(i, cpu_callout_map))
464 if (phys_proc_id[cpu] == phys_proc_id[i]) {
465 cpu_set(i, cpu_core_map[cpu]);
466 cpu_set(cpu, cpu_core_map[i]);
470 cpu_core_map[cpu] = cpu_sibling_map[cpu];
475 * Activate a secondary processor.
477 static void __devinit start_secondary(void *unused)
480 * Dont put anything before smp_callin(), SMP
481 * booting is too fragile that we want to limit the
482 * things done here to the most necessary things.
486 while (!cpu_isset(smp_processor_id(), smp_commenced_mask))
488 setup_secondary_APIC_clock();
489 if (nmi_watchdog == NMI_IO_APIC) {
490 disable_8259A_irq(0);
491 enable_NMI_through_LVT0(NULL);
496 * low-memory mappings have been cleared, flush them from
497 * the local TLBs too.
501 /* This must be done before setting cpu_online_map */
502 set_cpu_sibling_map(raw_smp_processor_id());
506 * We need to hold call_lock, so there is no inconsistency
507 * between the time smp_call_function() determines number of
508 * IPI receipients, and the time when the determination is made
509 * for which cpus receive the IPI. Holding this
510 * lock helps us to not include this cpu in a currently in progress
511 * smp_call_function().
513 lock_ipi_call_lock();
514 cpu_set(smp_processor_id(), cpu_online_map);
515 unlock_ipi_call_lock();
516 per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
518 /* We can take interrupts now: we're officially "up". */
526 * Everything has been set up for the secondary
527 * CPUs - they just need to reload everything
528 * from the task structure
529 * This function must not return.
531 void __devinit initialize_secondary(void)
534 * We don't actually need to load the full TSS,
535 * basically just the stack pointer and the eip.
542 :"r" (current->thread.esp),"r" (current->thread.eip));
552 /* which logical CPUs are on which nodes */
553 cpumask_t node_2_cpu_mask[MAX_NUMNODES] =
554 { [0 ... MAX_NUMNODES-1] = CPU_MASK_NONE };
555 /* which node each logical CPU is on */
556 int cpu_2_node[NR_CPUS] = { [0 ... NR_CPUS-1] = 0 };
557 EXPORT_SYMBOL(cpu_2_node);
559 /* set up a mapping between cpu and node. */
560 static inline void map_cpu_to_node(int cpu, int node)
562 printk("Mapping cpu %d to node %d\n", cpu, node);
563 cpu_set(cpu, node_2_cpu_mask[node]);
564 cpu_2_node[cpu] = node;
567 /* undo a mapping between cpu and node. */
568 static inline void unmap_cpu_to_node(int cpu)
572 printk("Unmapping cpu %d from all nodes\n", cpu);
573 for (node = 0; node < MAX_NUMNODES; node ++)
574 cpu_clear(cpu, node_2_cpu_mask[node]);
577 #else /* !CONFIG_NUMA */
579 #define map_cpu_to_node(cpu, node) ({})
580 #define unmap_cpu_to_node(cpu) ({})
582 #endif /* CONFIG_NUMA */
584 u8 cpu_2_logical_apicid[NR_CPUS] = { [0 ... NR_CPUS-1] = BAD_APICID };
586 static void map_cpu_to_logical_apicid(void)
588 int cpu = smp_processor_id();
589 int apicid = logical_smp_processor_id();
591 cpu_2_logical_apicid[cpu] = apicid;
592 map_cpu_to_node(cpu, apicid_to_node(apicid));
595 static void unmap_cpu_to_logical_apicid(int cpu)
597 cpu_2_logical_apicid[cpu] = BAD_APICID;
598 unmap_cpu_to_node(cpu);
602 static inline void __inquire_remote_apic(int apicid)
604 int i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
605 char *names[] = { "ID", "VERSION", "SPIV" };
608 printk("Inquiring remote APIC #%d...\n", apicid);
610 for (i = 0; i < sizeof(regs) / sizeof(*regs); i++) {
611 printk("... APIC #%d %s: ", apicid, names[i]);
616 apic_wait_icr_idle();
618 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(apicid));
619 apic_write_around(APIC_ICR, APIC_DM_REMRD | regs[i]);
624 status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
625 } while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
628 case APIC_ICR_RR_VALID:
629 status = apic_read(APIC_RRR);
630 printk("%08x\n", status);
639 #ifdef WAKE_SECONDARY_VIA_NMI
641 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
642 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
643 * won't ... remember to clear down the APIC, etc later.
646 wakeup_secondary_cpu(int logical_apicid, unsigned long start_eip)
648 unsigned long send_status = 0, accept_status = 0;
652 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(logical_apicid));
654 /* Boot on the stack */
655 /* Kick the second */
656 apic_write_around(APIC_ICR, APIC_DM_NMI | APIC_DEST_LOGICAL);
658 Dprintk("Waiting for send to finish...\n");
663 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
664 } while (send_status && (timeout++ < 1000));
667 * Give the other CPU some time to accept the IPI.
671 * Due to the Pentium erratum 3AP.
673 maxlvt = get_maxlvt();
675 apic_read_around(APIC_SPIV);
676 apic_write(APIC_ESR, 0);
678 accept_status = (apic_read(APIC_ESR) & 0xEF);
679 Dprintk("NMI sent.\n");
682 printk("APIC never delivered???\n");
684 printk("APIC delivery error (%lx).\n", accept_status);
686 return (send_status | accept_status);
688 #endif /* WAKE_SECONDARY_VIA_NMI */
690 #ifdef WAKE_SECONDARY_VIA_INIT
692 wakeup_secondary_cpu(int phys_apicid, unsigned long start_eip)
694 unsigned long send_status = 0, accept_status = 0;
695 int maxlvt, timeout, num_starts, j;
698 * Be paranoid about clearing APIC errors.
700 if (APIC_INTEGRATED(apic_version[phys_apicid])) {
701 apic_read_around(APIC_SPIV);
702 apic_write(APIC_ESR, 0);
706 Dprintk("Asserting INIT.\n");
709 * Turn INIT on target chip
711 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
716 apic_write_around(APIC_ICR, APIC_INT_LEVELTRIG | APIC_INT_ASSERT
719 Dprintk("Waiting for send to finish...\n");
724 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
725 } while (send_status && (timeout++ < 1000));
729 Dprintk("Deasserting INIT.\n");
732 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
735 apic_write_around(APIC_ICR, APIC_INT_LEVELTRIG | APIC_DM_INIT);
737 Dprintk("Waiting for send to finish...\n");
742 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
743 } while (send_status && (timeout++ < 1000));
745 atomic_set(&init_deasserted, 1);
748 * Should we send STARTUP IPIs ?
750 * Determine this based on the APIC version.
751 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
753 if (APIC_INTEGRATED(apic_version[phys_apicid]))
759 * Run STARTUP IPI loop.
761 Dprintk("#startup loops: %d.\n", num_starts);
763 maxlvt = get_maxlvt();
765 for (j = 1; j <= num_starts; j++) {
766 Dprintk("Sending STARTUP #%d.\n",j);
767 apic_read_around(APIC_SPIV);
768 apic_write(APIC_ESR, 0);
770 Dprintk("After apic_write.\n");
777 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
779 /* Boot on the stack */
780 /* Kick the second */
781 apic_write_around(APIC_ICR, APIC_DM_STARTUP
782 | (start_eip >> 12));
785 * Give the other CPU some time to accept the IPI.
789 Dprintk("Startup point 1.\n");
791 Dprintk("Waiting for send to finish...\n");
796 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
797 } while (send_status && (timeout++ < 1000));
800 * Give the other CPU some time to accept the IPI.
804 * Due to the Pentium erratum 3AP.
807 apic_read_around(APIC_SPIV);
808 apic_write(APIC_ESR, 0);
810 accept_status = (apic_read(APIC_ESR) & 0xEF);
811 if (send_status || accept_status)
814 Dprintk("After Startup.\n");
817 printk("APIC never delivered???\n");
819 printk("APIC delivery error (%lx).\n", accept_status);
821 return (send_status | accept_status);
823 #endif /* WAKE_SECONDARY_VIA_INIT */
825 extern cpumask_t cpu_initialized;
826 static inline int alloc_cpu_id(void)
830 cpus_complement(tmp_map, cpu_present_map);
831 cpu = first_cpu(tmp_map);
837 #ifdef CONFIG_HOTPLUG_CPU
838 static struct task_struct * __devinitdata cpu_idle_tasks[NR_CPUS];
839 static inline struct task_struct * alloc_idle_task(int cpu)
841 struct task_struct *idle;
843 if ((idle = cpu_idle_tasks[cpu]) != NULL) {
844 /* initialize thread_struct. we really want to avoid destroy
847 idle->thread.esp = (unsigned long)(((struct pt_regs *)
848 (THREAD_SIZE + (unsigned long) idle->thread_info)) - 1);
849 init_idle(idle, cpu);
852 idle = fork_idle(cpu);
855 cpu_idle_tasks[cpu] = idle;
859 #define alloc_idle_task(cpu) fork_idle(cpu)
862 static int __devinit do_boot_cpu(int apicid, int cpu)
864 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
865 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
866 * Returns zero if CPU booted OK, else error code from wakeup_secondary_cpu.
869 struct task_struct *idle;
870 unsigned long boot_error;
872 unsigned long start_eip;
873 unsigned short nmi_high = 0, nmi_low = 0;
878 * We can't use kernel_thread since we must avoid to
879 * reschedule the child.
881 idle = alloc_idle_task(cpu);
883 panic("failed fork for CPU %d", cpu);
884 idle->thread.eip = (unsigned long) start_secondary;
885 /* start_eip had better be page-aligned! */
886 start_eip = setup_trampoline();
888 /* So we see what's up */
889 printk("Booting processor %d/%d eip %lx\n", cpu, apicid, start_eip);
890 /* Stack for startup_32 can be just as for start_secondary onwards */
891 stack_start.esp = (void *) idle->thread.esp;
896 * This grunge runs the startup process for
897 * the targeted processor.
900 atomic_set(&init_deasserted, 0);
902 Dprintk("Setting warm reset code and vector.\n");
904 store_NMI_vector(&nmi_high, &nmi_low);
906 smpboot_setup_warm_reset_vector(start_eip);
909 * Starting actual IPI sequence...
911 boot_error = wakeup_secondary_cpu(apicid, start_eip);
915 * allow APs to start initializing.
917 Dprintk("Before Callout %d.\n", cpu);
918 cpu_set(cpu, cpu_callout_map);
919 Dprintk("After Callout %d.\n", cpu);
922 * Wait 5s total for a response
924 for (timeout = 0; timeout < 50000; timeout++) {
925 if (cpu_isset(cpu, cpu_callin_map))
926 break; /* It has booted */
930 if (cpu_isset(cpu, cpu_callin_map)) {
931 /* number CPUs logically, starting from 1 (BSP is 0) */
933 printk("CPU%d: ", cpu);
934 print_cpu_info(&cpu_data[cpu]);
935 Dprintk("CPU has booted.\n");
938 if (*((volatile unsigned char *)trampoline_base)
940 /* trampoline started but...? */
941 printk("Stuck ??\n");
943 /* trampoline code not run */
944 printk("Not responding.\n");
945 inquire_remote_apic(apicid);
950 /* Try to put things back the way they were before ... */
951 unmap_cpu_to_logical_apicid(cpu);
952 cpu_clear(cpu, cpu_callout_map); /* was set here (do_boot_cpu()) */
953 cpu_clear(cpu, cpu_initialized); /* was set by cpu_init() */
956 x86_cpu_to_apicid[cpu] = apicid;
957 cpu_set(cpu, cpu_present_map);
960 /* mark "stuck" area as not stuck */
961 *((volatile unsigned long *)trampoline_base) = 0;
966 #ifdef CONFIG_HOTPLUG_CPU
967 void cpu_exit_clear(void)
969 int cpu = raw_smp_processor_id();
977 cpu_clear(cpu, cpu_callout_map);
978 cpu_clear(cpu, cpu_callin_map);
979 cpu_clear(cpu, cpu_present_map);
981 cpu_clear(cpu, smp_commenced_mask);
982 unmap_cpu_to_logical_apicid(cpu);
985 struct warm_boot_cpu_info {
986 struct completion *complete;
991 static void __devinit do_warm_boot_cpu(void *p)
993 struct warm_boot_cpu_info *info = p;
994 do_boot_cpu(info->apicid, info->cpu);
995 complete(info->complete);
998 int __devinit smp_prepare_cpu(int cpu)
1000 DECLARE_COMPLETION(done);
1001 struct warm_boot_cpu_info info;
1002 struct work_struct task;
1006 apicid = x86_cpu_to_apicid[cpu];
1007 if (apicid == BAD_APICID) {
1012 info.complete = &done;
1013 info.apicid = apicid;
1015 INIT_WORK(&task, do_warm_boot_cpu, &info);
1017 tsc_sync_disabled = 1;
1019 /* init low mem mapping */
1020 memcpy(swapper_pg_dir, swapper_pg_dir + USER_PGD_PTRS,
1021 sizeof(swapper_pg_dir[0]) * KERNEL_PGD_PTRS);
1023 schedule_work(&task);
1024 wait_for_completion(&done);
1026 tsc_sync_disabled = 0;
1030 unlock_cpu_hotplug();
1035 static void smp_tune_scheduling (void)
1037 unsigned long cachesize; /* kB */
1038 unsigned long bandwidth = 350; /* MB/s */
1040 * Rough estimation for SMP scheduling, this is the number of
1041 * cycles it takes for a fully memory-limited process to flush
1042 * the SMP-local cache.
1044 * (For a P5 this pretty much means we will choose another idle
1045 * CPU almost always at wakeup time (this is due to the small
1046 * L1 cache), on PIIs it's around 50-100 usecs, depending on
1052 * this basically disables processor-affinity
1053 * scheduling on SMP without a TSC.
1057 cachesize = boot_cpu_data.x86_cache_size;
1058 if (cachesize == -1) {
1059 cachesize = 16; /* Pentiums, 2x8kB cache */
1066 * Cycle through the processors sending APIC IPIs to boot each.
1069 static int boot_cpu_logical_apicid;
1070 /* Where the IO area was mapped on multiquad, always 0 otherwise */
1072 #ifdef CONFIG_X86_NUMAQ
1073 EXPORT_SYMBOL(xquad_portio);
1076 static void __init smp_boot_cpus(unsigned int max_cpus)
1078 int apicid, cpu, bit, kicked;
1079 unsigned long bogosum = 0;
1082 * Setup boot CPU information
1084 smp_store_cpu_info(0); /* Final full version of the data */
1085 printk("CPU%d: ", 0);
1086 print_cpu_info(&cpu_data[0]);
1088 boot_cpu_physical_apicid = GET_APIC_ID(apic_read(APIC_ID));
1089 boot_cpu_logical_apicid = logical_smp_processor_id();
1090 x86_cpu_to_apicid[0] = boot_cpu_physical_apicid;
1092 current_thread_info()->cpu = 0;
1093 smp_tune_scheduling();
1094 cpus_clear(cpu_sibling_map[0]);
1095 cpu_set(0, cpu_sibling_map[0]);
1097 cpus_clear(cpu_core_map[0]);
1098 cpu_set(0, cpu_core_map[0]);
1101 * If we couldn't find an SMP configuration at boot time,
1102 * get out of here now!
1104 if (!smp_found_config && !acpi_lapic) {
1105 printk(KERN_NOTICE "SMP motherboard not detected.\n");
1106 smpboot_clear_io_apic_irqs();
1107 phys_cpu_present_map = physid_mask_of_physid(0);
1108 if (APIC_init_uniprocessor())
1109 printk(KERN_NOTICE "Local APIC not detected."
1110 " Using dummy APIC emulation.\n");
1111 map_cpu_to_logical_apicid();
1112 cpu_set(0, cpu_sibling_map[0]);
1113 cpu_set(0, cpu_core_map[0]);
1118 * Should not be necessary because the MP table should list the boot
1119 * CPU too, but we do it for the sake of robustness anyway.
1120 * Makes no sense to do this check in clustered apic mode, so skip it
1122 if (!check_phys_apicid_present(boot_cpu_physical_apicid)) {
1123 printk("weird, boot CPU (#%d) not listed by the BIOS.\n",
1124 boot_cpu_physical_apicid);
1125 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1129 * If we couldn't find a local APIC, then get out of here now!
1131 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid]) && !cpu_has_apic) {
1132 printk(KERN_ERR "BIOS bug, local APIC #%d not detected!...\n",
1133 boot_cpu_physical_apicid);
1134 printk(KERN_ERR "... forcing use of dummy APIC emulation. (tell your hw vendor)\n");
1135 smpboot_clear_io_apic_irqs();
1136 phys_cpu_present_map = physid_mask_of_physid(0);
1137 cpu_set(0, cpu_sibling_map[0]);
1138 cpu_set(0, cpu_core_map[0]);
1142 verify_local_APIC();
1145 * If SMP should be disabled, then really disable it!
1148 smp_found_config = 0;
1149 printk(KERN_INFO "SMP mode deactivated, forcing use of dummy APIC emulation.\n");
1150 smpboot_clear_io_apic_irqs();
1151 phys_cpu_present_map = physid_mask_of_physid(0);
1152 cpu_set(0, cpu_sibling_map[0]);
1153 cpu_set(0, cpu_core_map[0]);
1159 map_cpu_to_logical_apicid();
1162 setup_portio_remap();
1165 * Scan the CPU present map and fire up the other CPUs via do_boot_cpu
1167 * In clustered apic mode, phys_cpu_present_map is a constructed thus:
1168 * bits 0-3 are quad0, 4-7 are quad1, etc. A perverse twist on the
1169 * clustered apic ID.
1171 Dprintk("CPU present map: %lx\n", physids_coerce(phys_cpu_present_map));
1174 for (bit = 0; kicked < NR_CPUS && bit < MAX_APICS; bit++) {
1175 apicid = cpu_present_to_apicid(bit);
1177 * Don't even attempt to start the boot CPU!
1179 if ((apicid == boot_cpu_apicid) || (apicid == BAD_APICID))
1182 if (!check_apicid_present(bit))
1184 if (max_cpus <= cpucount+1)
1187 if (((cpu = alloc_cpu_id()) <= 0) || do_boot_cpu(apicid, cpu))
1188 printk("CPU #%d not responding - cannot use it.\n",
1195 * Cleanup possible dangling ends...
1197 smpboot_restore_warm_reset_vector();
1200 * Allow the user to impress friends.
1202 Dprintk("Before bogomips.\n");
1203 for (cpu = 0; cpu < NR_CPUS; cpu++)
1204 if (cpu_isset(cpu, cpu_callout_map))
1205 bogosum += cpu_data[cpu].loops_per_jiffy;
1207 "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
1209 bogosum/(500000/HZ),
1210 (bogosum/(5000/HZ))%100);
1212 Dprintk("Before bogocount - setting activated=1.\n");
1215 printk(KERN_WARNING "WARNING: SMP operation may be unreliable with B stepping processors.\n");
1218 * Don't taint if we are running SMP kernel on a single non-MP
1221 if (tainted & TAINT_UNSAFE_SMP) {
1223 printk (KERN_INFO "WARNING: This combination of AMD processors is not suitable for SMP.\n");
1225 tainted &= ~TAINT_UNSAFE_SMP;
1228 Dprintk("Boot done.\n");
1231 * construct cpu_sibling_map[], so that we can tell sibling CPUs
1234 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1235 cpus_clear(cpu_sibling_map[cpu]);
1236 cpus_clear(cpu_core_map[cpu]);
1239 cpu_set(0, cpu_sibling_map[0]);
1240 cpu_set(0, cpu_core_map[0]);
1242 smpboot_setup_io_apic();
1244 setup_boot_APIC_clock();
1247 * Synchronize the TSC with the AP
1249 if (cpu_has_tsc && cpucount && cpu_khz)
1250 synchronize_tsc_bp();
1253 /* These are wrappers to interface to the new boot process. Someone
1254 who understands all this stuff should rewrite it properly. --RR 15/Jul/02 */
1255 void __init smp_prepare_cpus(unsigned int max_cpus)
1257 smp_commenced_mask = cpumask_of_cpu(0);
1258 cpu_callin_map = cpumask_of_cpu(0);
1260 smp_boot_cpus(max_cpus);
1263 void __devinit smp_prepare_boot_cpu(void)
1265 cpu_set(smp_processor_id(), cpu_online_map);
1266 cpu_set(smp_processor_id(), cpu_callout_map);
1267 cpu_set(smp_processor_id(), cpu_present_map);
1268 per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
1271 #ifdef CONFIG_HOTPLUG_CPU
1273 remove_siblinginfo(int cpu)
1277 for_each_cpu_mask(sibling, cpu_sibling_map[cpu])
1278 cpu_clear(cpu, cpu_sibling_map[sibling]);
1279 for_each_cpu_mask(sibling, cpu_core_map[cpu])
1280 cpu_clear(cpu, cpu_core_map[sibling]);
1281 cpus_clear(cpu_sibling_map[cpu]);
1282 cpus_clear(cpu_core_map[cpu]);
1283 phys_proc_id[cpu] = BAD_APICID;
1284 cpu_core_id[cpu] = BAD_APICID;
1287 int __cpu_disable(void)
1289 cpumask_t map = cpu_online_map;
1290 int cpu = smp_processor_id();
1293 * Perhaps use cpufreq to drop frequency, but that could go
1294 * into generic code.
1296 * We won't take down the boot processor on i386 due to some
1297 * interrupts only being able to be serviced by the BSP.
1298 * Especially so if we're not using an IOAPIC -zwane
1303 /* We enable the timer again on the exit path of the death loop */
1304 disable_APIC_timer();
1305 /* Allow any queued timer interrupts to get serviced */
1308 local_irq_disable();
1310 remove_siblinginfo(cpu);
1312 cpu_clear(cpu, map);
1314 /* It's now safe to remove this processor from the online map */
1315 cpu_clear(cpu, cpu_online_map);
1319 void __cpu_die(unsigned int cpu)
1321 /* We don't do anything here: idle task is faking death itself. */
1324 for (i = 0; i < 10; i++) {
1325 /* They ack this in play_dead by setting CPU_DEAD */
1326 if (per_cpu(cpu_state, cpu) == CPU_DEAD) {
1327 printk ("CPU %d is now offline\n", cpu);
1330 current->state = TASK_UNINTERRUPTIBLE;
1331 schedule_timeout(HZ/10);
1333 printk(KERN_ERR "CPU %u didn't die...\n", cpu);
1335 #else /* ... !CONFIG_HOTPLUG_CPU */
1336 int __cpu_disable(void)
1341 void __cpu_die(unsigned int cpu)
1343 /* We said "no" in __cpu_disable */
1346 #endif /* CONFIG_HOTPLUG_CPU */
1348 int __devinit __cpu_up(unsigned int cpu)
1350 /* In case one didn't come up */
1351 if (!cpu_isset(cpu, cpu_callin_map)) {
1352 printk(KERN_DEBUG "skipping cpu%d, didn't come online\n", cpu);
1358 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
1359 /* Unleash the CPU! */
1360 cpu_set(cpu, smp_commenced_mask);
1361 while (!cpu_isset(cpu, cpu_online_map))
1366 void __init smp_cpus_done(unsigned int max_cpus)
1368 #ifdef CONFIG_X86_IO_APIC
1369 setup_ioapic_dest();
1372 #ifndef CONFIG_HOTPLUG_CPU
1374 * Disable executability of the SMP trampoline:
1376 set_kernel_exec((unsigned long)trampoline_base, trampoline_exec);
1380 void __init smp_intr_init(void)
1383 * IRQ0 must be given a fixed assignment and initialized,
1384 * because it's used before the IO-APIC is set up.
1386 set_intr_gate(FIRST_DEVICE_VECTOR, interrupt[0]);
1389 * The reschedule interrupt is a CPU-to-CPU reschedule-helper
1390 * IPI, driven by wakeup.
1392 set_intr_gate(RESCHEDULE_VECTOR, reschedule_interrupt);
1394 /* IPI for invalidation */
1395 set_intr_gate(INVALIDATE_TLB_VECTOR, invalidate_interrupt);
1397 /* IPI for generic function call */
1398 set_intr_gate(CALL_FUNCTION_VECTOR, call_function_interrupt);