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);
105 u8 apicid_2_node[MAX_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 if (num_possible_cpus() == 1)
185 /* Athlon 660/661 is valid. */
186 if ((c->x86_model==6) && ((c->x86_mask==0) || (c->x86_mask==1)))
189 /* Duron 670 is valid */
190 if ((c->x86_model==7) && (c->x86_mask==0))
194 * Athlon 662, Duron 671, and Athlon >model 7 have capability bit.
195 * It's worth noting that the A5 stepping (662) of some Athlon XP's
196 * have the MP bit set.
197 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for more.
199 if (((c->x86_model==6) && (c->x86_mask>=2)) ||
200 ((c->x86_model==7) && (c->x86_mask>=1)) ||
205 /* If we get here, it's not a certified SMP capable AMD system. */
206 add_taint(TAINT_UNSAFE_SMP);
214 * TSC synchronization.
216 * We first check whether all CPUs have their TSC's synchronized,
217 * then we print a warning if not, and always resync.
222 atomic_t count_start;
224 unsigned long long values[NR_CPUS];
226 .start_flag = ATOMIC_INIT(0),
227 .count_start = ATOMIC_INIT(0),
228 .count_stop = ATOMIC_INIT(0),
233 static void __init synchronize_tsc_bp(void)
236 unsigned long long t0;
237 unsigned long long sum, avg;
239 unsigned int one_usec;
242 printk(KERN_INFO "checking TSC synchronization across %u CPUs: ", num_booting_cpus());
244 /* convert from kcyc/sec to cyc/usec */
245 one_usec = cpu_khz / 1000;
247 atomic_set(&tsc.start_flag, 1);
251 * We loop a few times to get a primed instruction cache,
252 * then the last pass is more or less synchronized and
253 * the BP and APs set their cycle counters to zero all at
254 * once. This reduces the chance of having random offsets
255 * between the processors, and guarantees that the maximum
256 * delay between the cycle counters is never bigger than
257 * the latency of information-passing (cachelines) between
260 for (i = 0; i < NR_LOOPS; i++) {
262 * all APs synchronize but they loop on '== num_cpus'
264 while (atomic_read(&tsc.count_start) != num_booting_cpus()-1)
266 atomic_set(&tsc.count_stop, 0);
269 * this lets the APs save their current TSC:
271 atomic_inc(&tsc.count_start);
273 rdtscll(tsc.values[smp_processor_id()]);
275 * We clear the TSC in the last loop:
281 * Wait for all APs to leave the synchronization point:
283 while (atomic_read(&tsc.count_stop) != num_booting_cpus()-1)
285 atomic_set(&tsc.count_start, 0);
287 atomic_inc(&tsc.count_stop);
291 for (i = 0; i < NR_CPUS; i++) {
292 if (cpu_isset(i, cpu_callout_map)) {
298 do_div(avg, num_booting_cpus());
300 for (i = 0; i < NR_CPUS; i++) {
301 if (!cpu_isset(i, cpu_callout_map))
303 delta = tsc.values[i] - avg;
307 * We report bigger than 2 microseconds clock differences.
309 if (delta > 2*one_usec) {
317 do_div(realdelta, one_usec);
318 if (tsc.values[i] < avg)
319 realdelta = -realdelta;
322 printk(KERN_INFO "CPU#%d had %Ld usecs TSC "
323 "skew, fixed it up.\n", i, realdelta);
330 static void __init synchronize_tsc_ap(void)
335 * Not every cpu is online at the time
336 * this gets called, so we first wait for the BP to
337 * finish SMP initialization:
339 while (!atomic_read(&tsc.start_flag))
342 for (i = 0; i < NR_LOOPS; i++) {
343 atomic_inc(&tsc.count_start);
344 while (atomic_read(&tsc.count_start) != num_booting_cpus())
347 rdtscll(tsc.values[smp_processor_id()]);
351 atomic_inc(&tsc.count_stop);
352 while (atomic_read(&tsc.count_stop) != num_booting_cpus())
358 extern void calibrate_delay(void);
360 static atomic_t init_deasserted;
362 static void __devinit smp_callin(void)
365 unsigned long timeout;
368 * If waken up by an INIT in an 82489DX configuration
369 * we may get here before an INIT-deassert IPI reaches
370 * our local APIC. We have to wait for the IPI or we'll
371 * lock up on an APIC access.
373 wait_for_init_deassert(&init_deasserted);
376 * (This works even if the APIC is not enabled.)
378 phys_id = GET_APIC_ID(apic_read(APIC_ID));
379 cpuid = smp_processor_id();
380 if (cpu_isset(cpuid, cpu_callin_map)) {
381 printk("huh, phys CPU#%d, CPU#%d already present??\n",
385 Dprintk("CPU#%d (phys ID: %d) waiting for CALLOUT\n", cpuid, phys_id);
388 * STARTUP IPIs are fragile beasts as they might sometimes
389 * trigger some glue motherboard logic. Complete APIC bus
390 * silence for 1 second, this overestimates the time the
391 * boot CPU is spending to send the up to 2 STARTUP IPIs
392 * by a factor of two. This should be enough.
396 * Waiting 2s total for startup (udelay is not yet working)
398 timeout = jiffies + 2*HZ;
399 while (time_before(jiffies, timeout)) {
401 * Has the boot CPU finished it's STARTUP sequence?
403 if (cpu_isset(cpuid, cpu_callout_map))
408 if (!time_before(jiffies, timeout)) {
409 printk("BUG: CPU%d started up but did not get a callout!\n",
415 * the boot CPU has finished the init stage and is spinning
416 * on callin_map until we finish. We are free to set up this
417 * CPU, first the APIC. (this is probably redundant on most
421 Dprintk("CALLIN, before setup_local_APIC().\n");
422 smp_callin_clear_local_apic();
424 map_cpu_to_logical_apicid();
430 Dprintk("Stack at about %p\n",&cpuid);
433 * Save our processor parameters
435 smp_store_cpu_info(cpuid);
437 disable_APIC_timer();
440 * Allow the master to continue.
442 cpu_set(cpuid, cpu_callin_map);
445 * Synchronize the TSC with the BP
447 if (cpu_has_tsc && cpu_khz && !tsc_sync_disabled)
448 synchronize_tsc_ap();
453 /* maps the cpu to the sched domain representing multi-core */
454 cpumask_t cpu_coregroup_map(int cpu)
456 struct cpuinfo_x86 *c = cpu_data + cpu;
458 * For perf, we return last level cache shared map.
459 * And for power savings, we return cpu_core_map
461 if (sched_mc_power_savings || sched_smt_power_savings)
462 return cpu_core_map[cpu];
464 return c->llc_shared_map;
467 /* representing cpus for which sibling maps can be computed */
468 static cpumask_t cpu_sibling_setup_map;
471 set_cpu_sibling_map(int cpu)
474 struct cpuinfo_x86 *c = cpu_data;
476 cpu_set(cpu, cpu_sibling_setup_map);
478 if (smp_num_siblings > 1) {
479 for_each_cpu_mask(i, cpu_sibling_setup_map) {
480 if (c[cpu].phys_proc_id == c[i].phys_proc_id &&
481 c[cpu].cpu_core_id == c[i].cpu_core_id) {
482 cpu_set(i, cpu_sibling_map[cpu]);
483 cpu_set(cpu, cpu_sibling_map[i]);
484 cpu_set(i, cpu_core_map[cpu]);
485 cpu_set(cpu, cpu_core_map[i]);
486 cpu_set(i, c[cpu].llc_shared_map);
487 cpu_set(cpu, c[i].llc_shared_map);
491 cpu_set(cpu, cpu_sibling_map[cpu]);
494 cpu_set(cpu, c[cpu].llc_shared_map);
496 if (current_cpu_data.x86_max_cores == 1) {
497 cpu_core_map[cpu] = cpu_sibling_map[cpu];
498 c[cpu].booted_cores = 1;
502 for_each_cpu_mask(i, cpu_sibling_setup_map) {
503 if (cpu_llc_id[cpu] != BAD_APICID &&
504 cpu_llc_id[cpu] == cpu_llc_id[i]) {
505 cpu_set(i, c[cpu].llc_shared_map);
506 cpu_set(cpu, c[i].llc_shared_map);
508 if (c[cpu].phys_proc_id == c[i].phys_proc_id) {
509 cpu_set(i, cpu_core_map[cpu]);
510 cpu_set(cpu, cpu_core_map[i]);
512 * Does this new cpu bringup a new core?
514 if (cpus_weight(cpu_sibling_map[cpu]) == 1) {
516 * for each core in package, increment
517 * the booted_cores for this new cpu
519 if (first_cpu(cpu_sibling_map[i]) == i)
520 c[cpu].booted_cores++;
522 * increment the core count for all
523 * the other cpus in this package
527 } else if (i != cpu && !c[cpu].booted_cores)
528 c[cpu].booted_cores = c[i].booted_cores;
534 * Activate a secondary processor.
536 static void __devinit start_secondary(void *unused)
539 * Dont put anything before smp_callin(), SMP
540 * booting is too fragile that we want to limit the
541 * things done here to the most necessary things.
546 while (!cpu_isset(smp_processor_id(), smp_commenced_mask))
548 setup_secondary_APIC_clock();
549 if (nmi_watchdog == NMI_IO_APIC) {
550 disable_8259A_irq(0);
551 enable_NMI_through_LVT0(NULL);
556 * low-memory mappings have been cleared, flush them from
557 * the local TLBs too.
561 /* This must be done before setting cpu_online_map */
562 set_cpu_sibling_map(raw_smp_processor_id());
566 * We need to hold call_lock, so there is no inconsistency
567 * between the time smp_call_function() determines number of
568 * IPI receipients, and the time when the determination is made
569 * for which cpus receive the IPI. Holding this
570 * lock helps us to not include this cpu in a currently in progress
571 * smp_call_function().
573 lock_ipi_call_lock();
574 cpu_set(smp_processor_id(), cpu_online_map);
575 unlock_ipi_call_lock();
576 per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
578 /* We can take interrupts now: we're officially "up". */
586 * Everything has been set up for the secondary
587 * CPUs - they just need to reload everything
588 * from the task structure
589 * This function must not return.
591 void __devinit initialize_secondary(void)
594 * We don't actually need to load the full TSS,
595 * basically just the stack pointer and the eip.
602 :"r" (current->thread.esp),"r" (current->thread.eip));
612 /* which logical CPUs are on which nodes */
613 cpumask_t node_2_cpu_mask[MAX_NUMNODES] __read_mostly =
614 { [0 ... MAX_NUMNODES-1] = CPU_MASK_NONE };
615 EXPORT_SYMBOL(node_2_cpu_mask);
616 /* which node each logical CPU is on */
617 int cpu_2_node[NR_CPUS] __read_mostly = { [0 ... NR_CPUS-1] = 0 };
618 EXPORT_SYMBOL(cpu_2_node);
620 /* set up a mapping between cpu and node. */
621 static inline void map_cpu_to_node(int cpu, int node)
623 printk("Mapping cpu %d to node %d\n", cpu, node);
624 cpu_set(cpu, node_2_cpu_mask[node]);
625 cpu_2_node[cpu] = node;
628 /* undo a mapping between cpu and node. */
629 static inline void unmap_cpu_to_node(int cpu)
633 printk("Unmapping cpu %d from all nodes\n", cpu);
634 for (node = 0; node < MAX_NUMNODES; node ++)
635 cpu_clear(cpu, node_2_cpu_mask[node]);
638 #else /* !CONFIG_NUMA */
640 #define map_cpu_to_node(cpu, node) ({})
641 #define unmap_cpu_to_node(cpu) ({})
643 #endif /* CONFIG_NUMA */
645 u8 cpu_2_logical_apicid[NR_CPUS] __read_mostly = { [0 ... NR_CPUS-1] = BAD_APICID };
647 static void map_cpu_to_logical_apicid(void)
649 int cpu = smp_processor_id();
650 int apicid = logical_smp_processor_id();
651 int node = apicid_to_node(apicid);
653 if (!node_online(node))
654 node = first_online_node;
656 cpu_2_logical_apicid[cpu] = apicid;
657 map_cpu_to_node(cpu, node);
660 static void unmap_cpu_to_logical_apicid(int cpu)
662 cpu_2_logical_apicid[cpu] = BAD_APICID;
663 unmap_cpu_to_node(cpu);
667 static inline void __inquire_remote_apic(int apicid)
669 int i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
670 char *names[] = { "ID", "VERSION", "SPIV" };
673 printk("Inquiring remote APIC #%d...\n", apicid);
675 for (i = 0; i < ARRAY_SIZE(regs); i++) {
676 printk("... APIC #%d %s: ", apicid, names[i]);
681 apic_wait_icr_idle();
683 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(apicid));
684 apic_write_around(APIC_ICR, APIC_DM_REMRD | regs[i]);
689 status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
690 } while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
693 case APIC_ICR_RR_VALID:
694 status = apic_read(APIC_RRR);
695 printk("%08x\n", status);
704 #ifdef WAKE_SECONDARY_VIA_NMI
706 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
707 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
708 * won't ... remember to clear down the APIC, etc later.
711 wakeup_secondary_cpu(int logical_apicid, unsigned long start_eip)
713 unsigned long send_status = 0, accept_status = 0;
717 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(logical_apicid));
719 /* Boot on the stack */
720 /* Kick the second */
721 apic_write_around(APIC_ICR, APIC_DM_NMI | APIC_DEST_LOGICAL);
723 Dprintk("Waiting for send to finish...\n");
728 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
729 } while (send_status && (timeout++ < 1000));
732 * Give the other CPU some time to accept the IPI.
736 * Due to the Pentium erratum 3AP.
738 maxlvt = get_maxlvt();
740 apic_read_around(APIC_SPIV);
741 apic_write(APIC_ESR, 0);
743 accept_status = (apic_read(APIC_ESR) & 0xEF);
744 Dprintk("NMI sent.\n");
747 printk("APIC never delivered???\n");
749 printk("APIC delivery error (%lx).\n", accept_status);
751 return (send_status | accept_status);
753 #endif /* WAKE_SECONDARY_VIA_NMI */
755 #ifdef WAKE_SECONDARY_VIA_INIT
757 wakeup_secondary_cpu(int phys_apicid, unsigned long start_eip)
759 unsigned long send_status = 0, accept_status = 0;
760 int maxlvt, timeout, num_starts, j;
763 * Be paranoid about clearing APIC errors.
765 if (APIC_INTEGRATED(apic_version[phys_apicid])) {
766 apic_read_around(APIC_SPIV);
767 apic_write(APIC_ESR, 0);
771 Dprintk("Asserting INIT.\n");
774 * Turn INIT on target chip
776 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
781 apic_write_around(APIC_ICR, APIC_INT_LEVELTRIG | APIC_INT_ASSERT
784 Dprintk("Waiting for send to finish...\n");
789 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
790 } while (send_status && (timeout++ < 1000));
794 Dprintk("Deasserting INIT.\n");
797 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
800 apic_write_around(APIC_ICR, APIC_INT_LEVELTRIG | APIC_DM_INIT);
802 Dprintk("Waiting for send to finish...\n");
807 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
808 } while (send_status && (timeout++ < 1000));
810 atomic_set(&init_deasserted, 1);
813 * Should we send STARTUP IPIs ?
815 * Determine this based on the APIC version.
816 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
818 if (APIC_INTEGRATED(apic_version[phys_apicid]))
824 * Run STARTUP IPI loop.
826 Dprintk("#startup loops: %d.\n", num_starts);
828 maxlvt = get_maxlvt();
830 for (j = 1; j <= num_starts; j++) {
831 Dprintk("Sending STARTUP #%d.\n",j);
832 apic_read_around(APIC_SPIV);
833 apic_write(APIC_ESR, 0);
835 Dprintk("After apic_write.\n");
842 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
844 /* Boot on the stack */
845 /* Kick the second */
846 apic_write_around(APIC_ICR, APIC_DM_STARTUP
847 | (start_eip >> 12));
850 * Give the other CPU some time to accept the IPI.
854 Dprintk("Startup point 1.\n");
856 Dprintk("Waiting for send to finish...\n");
861 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
862 } while (send_status && (timeout++ < 1000));
865 * Give the other CPU some time to accept the IPI.
869 * Due to the Pentium erratum 3AP.
872 apic_read_around(APIC_SPIV);
873 apic_write(APIC_ESR, 0);
875 accept_status = (apic_read(APIC_ESR) & 0xEF);
876 if (send_status || accept_status)
879 Dprintk("After Startup.\n");
882 printk("APIC never delivered???\n");
884 printk("APIC delivery error (%lx).\n", accept_status);
886 return (send_status | accept_status);
888 #endif /* WAKE_SECONDARY_VIA_INIT */
890 extern cpumask_t cpu_initialized;
891 static inline int alloc_cpu_id(void)
895 cpus_complement(tmp_map, cpu_present_map);
896 cpu = first_cpu(tmp_map);
902 #ifdef CONFIG_HOTPLUG_CPU
903 static struct task_struct * __devinitdata cpu_idle_tasks[NR_CPUS];
904 static inline struct task_struct * alloc_idle_task(int cpu)
906 struct task_struct *idle;
908 if ((idle = cpu_idle_tasks[cpu]) != NULL) {
909 /* initialize thread_struct. we really want to avoid destroy
912 idle->thread.esp = (unsigned long)task_pt_regs(idle);
913 init_idle(idle, cpu);
916 idle = fork_idle(cpu);
919 cpu_idle_tasks[cpu] = idle;
923 #define alloc_idle_task(cpu) fork_idle(cpu)
926 static int __devinit do_boot_cpu(int apicid, int cpu)
928 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
929 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
930 * Returns zero if CPU booted OK, else error code from wakeup_secondary_cpu.
933 struct task_struct *idle;
934 unsigned long boot_error;
936 unsigned long start_eip;
937 unsigned short nmi_high = 0, nmi_low = 0;
940 alternatives_smp_switch(1);
943 * We can't use kernel_thread since we must avoid to
944 * reschedule the child.
946 idle = alloc_idle_task(cpu);
948 panic("failed fork for CPU %d", cpu);
949 idle->thread.eip = (unsigned long) start_secondary;
950 /* start_eip had better be page-aligned! */
951 start_eip = setup_trampoline();
953 /* So we see what's up */
954 printk("Booting processor %d/%d eip %lx\n", cpu, apicid, start_eip);
955 /* Stack for startup_32 can be just as for start_secondary onwards */
956 stack_start.esp = (void *) idle->thread.esp;
960 x86_cpu_to_apicid[cpu] = apicid;
962 * This grunge runs the startup process for
963 * the targeted processor.
966 atomic_set(&init_deasserted, 0);
968 Dprintk("Setting warm reset code and vector.\n");
970 store_NMI_vector(&nmi_high, &nmi_low);
972 smpboot_setup_warm_reset_vector(start_eip);
975 * Starting actual IPI sequence...
977 boot_error = wakeup_secondary_cpu(apicid, start_eip);
981 * allow APs to start initializing.
983 Dprintk("Before Callout %d.\n", cpu);
984 cpu_set(cpu, cpu_callout_map);
985 Dprintk("After Callout %d.\n", cpu);
988 * Wait 5s total for a response
990 for (timeout = 0; timeout < 50000; timeout++) {
991 if (cpu_isset(cpu, cpu_callin_map))
992 break; /* It has booted */
996 if (cpu_isset(cpu, cpu_callin_map)) {
997 /* number CPUs logically, starting from 1 (BSP is 0) */
999 printk("CPU%d: ", cpu);
1000 print_cpu_info(&cpu_data[cpu]);
1001 Dprintk("CPU has booted.\n");
1004 if (*((volatile unsigned char *)trampoline_base)
1006 /* trampoline started but...? */
1007 printk("Stuck ??\n");
1009 /* trampoline code not run */
1010 printk("Not responding.\n");
1011 inquire_remote_apic(apicid);
1016 /* Try to put things back the way they were before ... */
1017 unmap_cpu_to_logical_apicid(cpu);
1018 cpu_clear(cpu, cpu_callout_map); /* was set here (do_boot_cpu()) */
1019 cpu_clear(cpu, cpu_initialized); /* was set by cpu_init() */
1022 x86_cpu_to_apicid[cpu] = apicid;
1023 cpu_set(cpu, cpu_present_map);
1026 /* mark "stuck" area as not stuck */
1027 *((volatile unsigned long *)trampoline_base) = 0;
1032 #ifdef CONFIG_HOTPLUG_CPU
1033 void cpu_exit_clear(void)
1035 int cpu = raw_smp_processor_id();
1043 cpu_clear(cpu, cpu_callout_map);
1044 cpu_clear(cpu, cpu_callin_map);
1046 cpu_clear(cpu, smp_commenced_mask);
1047 unmap_cpu_to_logical_apicid(cpu);
1050 struct warm_boot_cpu_info {
1051 struct completion *complete;
1052 struct work_struct task;
1057 static void __cpuinit do_warm_boot_cpu(struct work_struct *work)
1059 struct warm_boot_cpu_info *info =
1060 container_of(work, struct warm_boot_cpu_info, task);
1061 do_boot_cpu(info->apicid, info->cpu);
1062 complete(info->complete);
1065 static int __cpuinit __smp_prepare_cpu(int cpu)
1067 DECLARE_COMPLETION_ONSTACK(done);
1068 struct warm_boot_cpu_info info;
1070 struct Xgt_desc_struct *cpu_gdt_descr = &per_cpu(cpu_gdt_descr, cpu);
1072 apicid = x86_cpu_to_apicid[cpu];
1073 if (apicid == BAD_APICID) {
1079 * the CPU isn't initialized at boot time, allocate gdt table here.
1080 * cpu_init will initialize it
1082 if (!cpu_gdt_descr->address) {
1083 cpu_gdt_descr->address = get_zeroed_page(GFP_KERNEL);
1084 if (!cpu_gdt_descr->address)
1085 printk(KERN_CRIT "CPU%d failed to allocate GDT\n", cpu);
1090 info.complete = &done;
1091 info.apicid = apicid;
1093 INIT_WORK(&info.task, do_warm_boot_cpu);
1095 tsc_sync_disabled = 1;
1097 /* init low mem mapping */
1098 clone_pgd_range(swapper_pg_dir, swapper_pg_dir + USER_PGD_PTRS,
1101 schedule_work(&info.task);
1102 wait_for_completion(&done);
1104 tsc_sync_disabled = 0;
1112 static void smp_tune_scheduling (void)
1114 unsigned long cachesize; /* kB */
1115 unsigned long bandwidth = 350; /* MB/s */
1117 * Rough estimation for SMP scheduling, this is the number of
1118 * cycles it takes for a fully memory-limited process to flush
1119 * the SMP-local cache.
1121 * (For a P5 this pretty much means we will choose another idle
1122 * CPU almost always at wakeup time (this is due to the small
1123 * L1 cache), on PIIs it's around 50-100 usecs, depending on
1129 * this basically disables processor-affinity
1130 * scheduling on SMP without a TSC.
1134 cachesize = boot_cpu_data.x86_cache_size;
1135 if (cachesize == -1) {
1136 cachesize = 16; /* Pentiums, 2x8kB cache */
1139 max_cache_size = cachesize * 1024;
1144 * Cycle through the processors sending APIC IPIs to boot each.
1147 static int boot_cpu_logical_apicid;
1148 /* Where the IO area was mapped on multiquad, always 0 otherwise */
1150 #ifdef CONFIG_X86_NUMAQ
1151 EXPORT_SYMBOL(xquad_portio);
1154 static void __init smp_boot_cpus(unsigned int max_cpus)
1156 int apicid, cpu, bit, kicked;
1157 unsigned long bogosum = 0;
1160 * Setup boot CPU information
1162 smp_store_cpu_info(0); /* Final full version of the data */
1163 printk("CPU%d: ", 0);
1164 print_cpu_info(&cpu_data[0]);
1166 boot_cpu_physical_apicid = GET_APIC_ID(apic_read(APIC_ID));
1167 boot_cpu_logical_apicid = logical_smp_processor_id();
1168 x86_cpu_to_apicid[0] = boot_cpu_physical_apicid;
1170 current_thread_info()->cpu = 0;
1171 smp_tune_scheduling();
1173 set_cpu_sibling_map(0);
1176 * If we couldn't find an SMP configuration at boot time,
1177 * get out of here now!
1179 if (!smp_found_config && !acpi_lapic) {
1180 printk(KERN_NOTICE "SMP motherboard not detected.\n");
1181 smpboot_clear_io_apic_irqs();
1182 phys_cpu_present_map = physid_mask_of_physid(0);
1183 if (APIC_init_uniprocessor())
1184 printk(KERN_NOTICE "Local APIC not detected."
1185 " Using dummy APIC emulation.\n");
1186 map_cpu_to_logical_apicid();
1187 cpu_set(0, cpu_sibling_map[0]);
1188 cpu_set(0, cpu_core_map[0]);
1193 * Should not be necessary because the MP table should list the boot
1194 * CPU too, but we do it for the sake of robustness anyway.
1195 * Makes no sense to do this check in clustered apic mode, so skip it
1197 if (!check_phys_apicid_present(boot_cpu_physical_apicid)) {
1198 printk("weird, boot CPU (#%d) not listed by the BIOS.\n",
1199 boot_cpu_physical_apicid);
1200 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1204 * If we couldn't find a local APIC, then get out of here now!
1206 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid]) && !cpu_has_apic) {
1207 printk(KERN_ERR "BIOS bug, local APIC #%d not detected!...\n",
1208 boot_cpu_physical_apicid);
1209 printk(KERN_ERR "... forcing use of dummy APIC emulation. (tell your hw vendor)\n");
1210 smpboot_clear_io_apic_irqs();
1211 phys_cpu_present_map = physid_mask_of_physid(0);
1212 cpu_set(0, cpu_sibling_map[0]);
1213 cpu_set(0, cpu_core_map[0]);
1217 verify_local_APIC();
1220 * If SMP should be disabled, then really disable it!
1223 smp_found_config = 0;
1224 printk(KERN_INFO "SMP mode deactivated, forcing use of dummy APIC emulation.\n");
1225 smpboot_clear_io_apic_irqs();
1226 phys_cpu_present_map = physid_mask_of_physid(0);
1227 cpu_set(0, cpu_sibling_map[0]);
1228 cpu_set(0, cpu_core_map[0]);
1234 map_cpu_to_logical_apicid();
1237 setup_portio_remap();
1240 * Scan the CPU present map and fire up the other CPUs via do_boot_cpu
1242 * In clustered apic mode, phys_cpu_present_map is a constructed thus:
1243 * bits 0-3 are quad0, 4-7 are quad1, etc. A perverse twist on the
1244 * clustered apic ID.
1246 Dprintk("CPU present map: %lx\n", physids_coerce(phys_cpu_present_map));
1249 for (bit = 0; kicked < NR_CPUS && bit < MAX_APICS; bit++) {
1250 apicid = cpu_present_to_apicid(bit);
1252 * Don't even attempt to start the boot CPU!
1254 if ((apicid == boot_cpu_apicid) || (apicid == BAD_APICID))
1257 if (!check_apicid_present(bit))
1259 if (max_cpus <= cpucount+1)
1262 if (((cpu = alloc_cpu_id()) <= 0) || do_boot_cpu(apicid, cpu))
1263 printk("CPU #%d not responding - cannot use it.\n",
1270 * Cleanup possible dangling ends...
1272 smpboot_restore_warm_reset_vector();
1275 * Allow the user to impress friends.
1277 Dprintk("Before bogomips.\n");
1278 for (cpu = 0; cpu < NR_CPUS; cpu++)
1279 if (cpu_isset(cpu, cpu_callout_map))
1280 bogosum += cpu_data[cpu].loops_per_jiffy;
1282 "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
1284 bogosum/(500000/HZ),
1285 (bogosum/(5000/HZ))%100);
1287 Dprintk("Before bogocount - setting activated=1.\n");
1290 printk(KERN_WARNING "WARNING: SMP operation may be unreliable with B stepping processors.\n");
1293 * Don't taint if we are running SMP kernel on a single non-MP
1296 if (tainted & TAINT_UNSAFE_SMP) {
1298 printk (KERN_INFO "WARNING: This combination of AMD processors is not suitable for SMP.\n");
1300 tainted &= ~TAINT_UNSAFE_SMP;
1303 Dprintk("Boot done.\n");
1306 * construct cpu_sibling_map[], so that we can tell sibling CPUs
1309 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1310 cpus_clear(cpu_sibling_map[cpu]);
1311 cpus_clear(cpu_core_map[cpu]);
1314 cpu_set(0, cpu_sibling_map[0]);
1315 cpu_set(0, cpu_core_map[0]);
1317 smpboot_setup_io_apic();
1319 setup_boot_APIC_clock();
1322 * Synchronize the TSC with the AP
1324 if (cpu_has_tsc && cpucount && cpu_khz)
1325 synchronize_tsc_bp();
1328 /* These are wrappers to interface to the new boot process. Someone
1329 who understands all this stuff should rewrite it properly. --RR 15/Jul/02 */
1330 void __init smp_prepare_cpus(unsigned int max_cpus)
1332 smp_commenced_mask = cpumask_of_cpu(0);
1333 cpu_callin_map = cpumask_of_cpu(0);
1335 smp_boot_cpus(max_cpus);
1338 void __devinit smp_prepare_boot_cpu(void)
1340 cpu_set(smp_processor_id(), cpu_online_map);
1341 cpu_set(smp_processor_id(), cpu_callout_map);
1342 cpu_set(smp_processor_id(), cpu_present_map);
1343 cpu_set(smp_processor_id(), cpu_possible_map);
1344 per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
1347 #ifdef CONFIG_HOTPLUG_CPU
1349 remove_siblinginfo(int cpu)
1352 struct cpuinfo_x86 *c = cpu_data;
1354 for_each_cpu_mask(sibling, cpu_core_map[cpu]) {
1355 cpu_clear(cpu, cpu_core_map[sibling]);
1357 * last thread sibling in this cpu core going down
1359 if (cpus_weight(cpu_sibling_map[cpu]) == 1)
1360 c[sibling].booted_cores--;
1363 for_each_cpu_mask(sibling, cpu_sibling_map[cpu])
1364 cpu_clear(cpu, cpu_sibling_map[sibling]);
1365 cpus_clear(cpu_sibling_map[cpu]);
1366 cpus_clear(cpu_core_map[cpu]);
1367 c[cpu].phys_proc_id = 0;
1368 c[cpu].cpu_core_id = 0;
1369 cpu_clear(cpu, cpu_sibling_setup_map);
1372 int __cpu_disable(void)
1374 cpumask_t map = cpu_online_map;
1375 int cpu = smp_processor_id();
1378 * Perhaps use cpufreq to drop frequency, but that could go
1379 * into generic code.
1381 * We won't take down the boot processor on i386 due to some
1382 * interrupts only being able to be serviced by the BSP.
1383 * Especially so if we're not using an IOAPIC -zwane
1387 if (nmi_watchdog == NMI_LOCAL_APIC)
1388 stop_apic_nmi_watchdog(NULL);
1390 /* Allow any queued timer interrupts to get serviced */
1393 local_irq_disable();
1395 remove_siblinginfo(cpu);
1397 cpu_clear(cpu, map);
1399 /* It's now safe to remove this processor from the online map */
1400 cpu_clear(cpu, cpu_online_map);
1404 void __cpu_die(unsigned int cpu)
1406 /* We don't do anything here: idle task is faking death itself. */
1409 for (i = 0; i < 10; i++) {
1410 /* They ack this in play_dead by setting CPU_DEAD */
1411 if (per_cpu(cpu_state, cpu) == CPU_DEAD) {
1412 printk ("CPU %d is now offline\n", cpu);
1413 if (1 == num_online_cpus())
1414 alternatives_smp_switch(0);
1419 printk(KERN_ERR "CPU %u didn't die...\n", cpu);
1421 #else /* ... !CONFIG_HOTPLUG_CPU */
1422 int __cpu_disable(void)
1427 void __cpu_die(unsigned int cpu)
1429 /* We said "no" in __cpu_disable */
1432 #endif /* CONFIG_HOTPLUG_CPU */
1434 int __devinit __cpu_up(unsigned int cpu)
1436 #ifdef CONFIG_HOTPLUG_CPU
1440 * We do warm boot only on cpus that had booted earlier
1441 * Otherwise cold boot is all handled from smp_boot_cpus().
1442 * cpu_callin_map is set during AP kickstart process. Its reset
1443 * when a cpu is taken offline from cpu_exit_clear().
1445 if (!cpu_isset(cpu, cpu_callin_map))
1446 ret = __smp_prepare_cpu(cpu);
1452 /* In case one didn't come up */
1453 if (!cpu_isset(cpu, cpu_callin_map)) {
1454 printk(KERN_DEBUG "skipping cpu%d, didn't come online\n", cpu);
1460 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
1461 /* Unleash the CPU! */
1462 cpu_set(cpu, smp_commenced_mask);
1463 while (!cpu_isset(cpu, cpu_online_map))
1468 void __init smp_cpus_done(unsigned int max_cpus)
1470 #ifdef CONFIG_X86_IO_APIC
1471 setup_ioapic_dest();
1474 #ifndef CONFIG_HOTPLUG_CPU
1476 * Disable executability of the SMP trampoline:
1478 set_kernel_exec((unsigned long)trampoline_base, trampoline_exec);
1482 void __init smp_intr_init(void)
1485 * IRQ0 must be given a fixed assignment and initialized,
1486 * because it's used before the IO-APIC is set up.
1488 set_intr_gate(FIRST_DEVICE_VECTOR, interrupt[0]);
1491 * The reschedule interrupt is a CPU-to-CPU reschedule-helper
1492 * IPI, driven by wakeup.
1494 set_intr_gate(RESCHEDULE_VECTOR, reschedule_interrupt);
1496 /* IPI for invalidation */
1497 set_intr_gate(INVALIDATE_TLB_VECTOR, invalidate_interrupt);
1499 /* IPI for generic function call */
1500 set_intr_gate(CALL_FUNCTION_VECTOR, call_function_interrupt);
1504 * If the BIOS enumerates physical processors before logical,
1505 * maxcpus=N at enumeration-time can be used to disable HT.
1507 static int __init parse_maxcpus(char *arg)
1509 extern unsigned int maxcpus;
1511 maxcpus = simple_strtoul(arg, NULL, 0);
1514 early_param("maxcpus", parse_maxcpus);