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] __read_mostly = {[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] __read_mostly = {[0 ... NR_CPUS-1] = BAD_APICID};
76 EXPORT_SYMBOL(cpu_core_id);
78 cpumask_t cpu_sibling_map[NR_CPUS] __read_mostly;
79 EXPORT_SYMBOL(cpu_sibling_map);
81 cpumask_t cpu_core_map[NR_CPUS] __read_mostly;
82 EXPORT_SYMBOL(cpu_core_map);
84 /* bitmap of online cpus */
85 cpumask_t cpu_online_map __read_mostly;
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 cpumask_t cpu_possible_map;
92 EXPORT_SYMBOL(cpu_possible_map);
93 static cpumask_t smp_commenced_mask;
95 /* TSC's upper 32 bits can't be written in eariler CPU (before prescott), there
96 * is no way to resync one AP against BP. TBD: for prescott and above, we
97 * should use IA64's algorithm
99 static int __devinitdata tsc_sync_disabled;
101 /* Per CPU bogomips and other parameters */
102 struct cpuinfo_x86 cpu_data[NR_CPUS] __cacheline_aligned;
103 EXPORT_SYMBOL(cpu_data);
105 u8 x86_cpu_to_apicid[NR_CPUS] __read_mostly =
106 { [0 ... NR_CPUS-1] = 0xff };
107 EXPORT_SYMBOL(x86_cpu_to_apicid);
110 * Trampoline 80x86 program as an array.
113 extern unsigned char trampoline_data [];
114 extern unsigned char trampoline_end [];
115 static unsigned char *trampoline_base;
116 static int trampoline_exec;
118 static void map_cpu_to_logical_apicid(void);
120 /* State of each CPU. */
121 DEFINE_PER_CPU(int, cpu_state) = { 0 };
124 * Currently trivial. Write the real->protected mode
125 * bootstrap into the page concerned. The caller
126 * has made sure it's suitably aligned.
129 static unsigned long __devinit setup_trampoline(void)
131 memcpy(trampoline_base, trampoline_data, trampoline_end - trampoline_data);
132 return virt_to_phys(trampoline_base);
136 * We are called very early to get the low memory for the
137 * SMP bootup trampoline page.
139 void __init smp_alloc_memory(void)
141 trampoline_base = (void *) alloc_bootmem_low_pages(PAGE_SIZE);
143 * Has to be in very low memory so we can execute
146 if (__pa(trampoline_base) >= 0x9F000)
149 * Make the SMP trampoline executable:
151 trampoline_exec = set_kernel_exec((unsigned long)trampoline_base, 1);
155 * The bootstrap kernel entry code has set these up. Save them for
159 static void __devinit smp_store_cpu_info(int id)
161 struct cpuinfo_x86 *c = cpu_data + id;
167 * Mask B, Pentium, but not Pentium MMX
169 if (c->x86_vendor == X86_VENDOR_INTEL &&
171 c->x86_mask >= 1 && c->x86_mask <= 4 &&
174 * Remember we have B step Pentia with bugs
179 * Certain Athlons might work (for various values of 'work') in SMP
180 * but they are not certified as MP capable.
182 if ((c->x86_vendor == X86_VENDOR_AMD) && (c->x86 == 6)) {
184 /* Athlon 660/661 is valid. */
185 if ((c->x86_model==6) && ((c->x86_mask==0) || (c->x86_mask==1)))
188 /* Duron 670 is valid */
189 if ((c->x86_model==7) && (c->x86_mask==0))
193 * Athlon 662, Duron 671, and Athlon >model 7 have capability bit.
194 * It's worth noting that the A5 stepping (662) of some Athlon XP's
195 * have the MP bit set.
196 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for more.
198 if (((c->x86_model==6) && (c->x86_mask>=2)) ||
199 ((c->x86_model==7) && (c->x86_mask>=1)) ||
204 /* If we get here, it's not a certified SMP capable AMD system. */
205 add_taint(TAINT_UNSAFE_SMP);
213 * TSC synchronization.
215 * We first check whether all CPUs have their TSC's synchronized,
216 * then we print a warning if not, and always resync.
219 static atomic_t tsc_start_flag = ATOMIC_INIT(0);
220 static atomic_t tsc_count_start = ATOMIC_INIT(0);
221 static atomic_t tsc_count_stop = ATOMIC_INIT(0);
222 static unsigned long long tsc_values[NR_CPUS];
226 static void __init synchronize_tsc_bp (void)
229 unsigned long long t0;
230 unsigned long long sum, avg;
232 unsigned int one_usec;
235 printk(KERN_INFO "checking TSC synchronization across %u CPUs: ", num_booting_cpus());
237 /* convert from kcyc/sec to cyc/usec */
238 one_usec = cpu_khz / 1000;
240 atomic_set(&tsc_start_flag, 1);
244 * We loop a few times to get a primed instruction cache,
245 * then the last pass is more or less synchronized and
246 * the BP and APs set their cycle counters to zero all at
247 * once. This reduces the chance of having random offsets
248 * between the processors, and guarantees that the maximum
249 * delay between the cycle counters is never bigger than
250 * the latency of information-passing (cachelines) between
253 for (i = 0; i < NR_LOOPS; i++) {
255 * all APs synchronize but they loop on '== num_cpus'
257 while (atomic_read(&tsc_count_start) != num_booting_cpus()-1)
259 atomic_set(&tsc_count_stop, 0);
262 * this lets the APs save their current TSC:
264 atomic_inc(&tsc_count_start);
266 rdtscll(tsc_values[smp_processor_id()]);
268 * We clear the TSC in the last loop:
274 * Wait for all APs to leave the synchronization point:
276 while (atomic_read(&tsc_count_stop) != num_booting_cpus()-1)
278 atomic_set(&tsc_count_start, 0);
280 atomic_inc(&tsc_count_stop);
284 for (i = 0; i < NR_CPUS; i++) {
285 if (cpu_isset(i, cpu_callout_map)) {
291 do_div(avg, num_booting_cpus());
294 for (i = 0; i < NR_CPUS; i++) {
295 if (!cpu_isset(i, cpu_callout_map))
297 delta = tsc_values[i] - avg;
301 * We report bigger than 2 microseconds clock differences.
303 if (delta > 2*one_usec) {
310 do_div(realdelta, one_usec);
311 if (tsc_values[i] < avg)
312 realdelta = -realdelta;
314 printk(KERN_INFO "CPU#%d had %ld usecs TSC skew, fixed it up.\n", i, realdelta);
323 static void __init synchronize_tsc_ap (void)
328 * Not every cpu is online at the time
329 * this gets called, so we first wait for the BP to
330 * finish SMP initialization:
332 while (!atomic_read(&tsc_start_flag)) mb();
334 for (i = 0; i < NR_LOOPS; i++) {
335 atomic_inc(&tsc_count_start);
336 while (atomic_read(&tsc_count_start) != num_booting_cpus())
339 rdtscll(tsc_values[smp_processor_id()]);
343 atomic_inc(&tsc_count_stop);
344 while (atomic_read(&tsc_count_stop) != num_booting_cpus()) mb();
349 extern void calibrate_delay(void);
351 static atomic_t init_deasserted;
353 static void __devinit smp_callin(void)
356 unsigned long timeout;
359 * If waken up by an INIT in an 82489DX configuration
360 * we may get here before an INIT-deassert IPI reaches
361 * our local APIC. We have to wait for the IPI or we'll
362 * lock up on an APIC access.
364 wait_for_init_deassert(&init_deasserted);
367 * (This works even if the APIC is not enabled.)
369 phys_id = GET_APIC_ID(apic_read(APIC_ID));
370 cpuid = smp_processor_id();
371 if (cpu_isset(cpuid, cpu_callin_map)) {
372 printk("huh, phys CPU#%d, CPU#%d already present??\n",
376 Dprintk("CPU#%d (phys ID: %d) waiting for CALLOUT\n", cpuid, phys_id);
379 * STARTUP IPIs are fragile beasts as they might sometimes
380 * trigger some glue motherboard logic. Complete APIC bus
381 * silence for 1 second, this overestimates the time the
382 * boot CPU is spending to send the up to 2 STARTUP IPIs
383 * by a factor of two. This should be enough.
387 * Waiting 2s total for startup (udelay is not yet working)
389 timeout = jiffies + 2*HZ;
390 while (time_before(jiffies, timeout)) {
392 * Has the boot CPU finished it's STARTUP sequence?
394 if (cpu_isset(cpuid, cpu_callout_map))
399 if (!time_before(jiffies, timeout)) {
400 printk("BUG: CPU%d started up but did not get a callout!\n",
406 * the boot CPU has finished the init stage and is spinning
407 * on callin_map until we finish. We are free to set up this
408 * CPU, first the APIC. (this is probably redundant on most
412 Dprintk("CALLIN, before setup_local_APIC().\n");
413 smp_callin_clear_local_apic();
415 map_cpu_to_logical_apicid();
421 Dprintk("Stack at about %p\n",&cpuid);
424 * Save our processor parameters
426 smp_store_cpu_info(cpuid);
428 disable_APIC_timer();
431 * Allow the master to continue.
433 cpu_set(cpuid, cpu_callin_map);
436 * Synchronize the TSC with the BP
438 if (cpu_has_tsc && cpu_khz && !tsc_sync_disabled)
439 synchronize_tsc_ap();
445 set_cpu_sibling_map(int cpu)
449 if (smp_num_siblings > 1) {
450 for (i = 0; i < NR_CPUS; i++) {
451 if (!cpu_isset(i, cpu_callout_map))
453 if (cpu_core_id[cpu] == cpu_core_id[i]) {
454 cpu_set(i, cpu_sibling_map[cpu]);
455 cpu_set(cpu, cpu_sibling_map[i]);
459 cpu_set(cpu, cpu_sibling_map[cpu]);
462 if (current_cpu_data.x86_num_cores > 1) {
463 for (i = 0; i < NR_CPUS; i++) {
464 if (!cpu_isset(i, cpu_callout_map))
466 if (phys_proc_id[cpu] == phys_proc_id[i]) {
467 cpu_set(i, cpu_core_map[cpu]);
468 cpu_set(cpu, cpu_core_map[i]);
472 cpu_core_map[cpu] = cpu_sibling_map[cpu];
477 * Activate a secondary processor.
479 static void __devinit start_secondary(void *unused)
482 * Dont put anything before smp_callin(), SMP
483 * booting is too fragile that we want to limit the
484 * things done here to the most necessary things.
488 while (!cpu_isset(smp_processor_id(), smp_commenced_mask))
490 setup_secondary_APIC_clock();
491 if (nmi_watchdog == NMI_IO_APIC) {
492 disable_8259A_irq(0);
493 enable_NMI_through_LVT0(NULL);
498 * low-memory mappings have been cleared, flush them from
499 * the local TLBs too.
503 /* This must be done before setting cpu_online_map */
504 set_cpu_sibling_map(raw_smp_processor_id());
508 * We need to hold call_lock, so there is no inconsistency
509 * between the time smp_call_function() determines number of
510 * IPI receipients, and the time when the determination is made
511 * for which cpus receive the IPI. Holding this
512 * lock helps us to not include this cpu in a currently in progress
513 * smp_call_function().
515 lock_ipi_call_lock();
516 cpu_set(smp_processor_id(), cpu_online_map);
517 unlock_ipi_call_lock();
518 per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
520 /* We can take interrupts now: we're officially "up". */
528 * Everything has been set up for the secondary
529 * CPUs - they just need to reload everything
530 * from the task structure
531 * This function must not return.
533 void __devinit initialize_secondary(void)
536 * We don't actually need to load the full TSS,
537 * basically just the stack pointer and the eip.
544 :"r" (current->thread.esp),"r" (current->thread.eip));
554 /* which logical CPUs are on which nodes */
555 cpumask_t node_2_cpu_mask[MAX_NUMNODES] __read_mostly =
556 { [0 ... MAX_NUMNODES-1] = CPU_MASK_NONE };
557 /* which node each logical CPU is on */
558 int cpu_2_node[NR_CPUS] __read_mostly = { [0 ... NR_CPUS-1] = 0 };
559 EXPORT_SYMBOL(cpu_2_node);
561 /* set up a mapping between cpu and node. */
562 static inline void map_cpu_to_node(int cpu, int node)
564 printk("Mapping cpu %d to node %d\n", cpu, node);
565 cpu_set(cpu, node_2_cpu_mask[node]);
566 cpu_2_node[cpu] = node;
569 /* undo a mapping between cpu and node. */
570 static inline void unmap_cpu_to_node(int cpu)
574 printk("Unmapping cpu %d from all nodes\n", cpu);
575 for (node = 0; node < MAX_NUMNODES; node ++)
576 cpu_clear(cpu, node_2_cpu_mask[node]);
579 #else /* !CONFIG_NUMA */
581 #define map_cpu_to_node(cpu, node) ({})
582 #define unmap_cpu_to_node(cpu) ({})
584 #endif /* CONFIG_NUMA */
586 u8 cpu_2_logical_apicid[NR_CPUS] __read_mostly = { [0 ... NR_CPUS-1] = BAD_APICID };
588 static void map_cpu_to_logical_apicid(void)
590 int cpu = smp_processor_id();
591 int apicid = logical_smp_processor_id();
593 cpu_2_logical_apicid[cpu] = apicid;
594 map_cpu_to_node(cpu, apicid_to_node(apicid));
597 static void unmap_cpu_to_logical_apicid(int cpu)
599 cpu_2_logical_apicid[cpu] = BAD_APICID;
600 unmap_cpu_to_node(cpu);
604 static inline void __inquire_remote_apic(int apicid)
606 int i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
607 char *names[] = { "ID", "VERSION", "SPIV" };
610 printk("Inquiring remote APIC #%d...\n", apicid);
612 for (i = 0; i < sizeof(regs) / sizeof(*regs); i++) {
613 printk("... APIC #%d %s: ", apicid, names[i]);
618 apic_wait_icr_idle();
620 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(apicid));
621 apic_write_around(APIC_ICR, APIC_DM_REMRD | regs[i]);
626 status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
627 } while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
630 case APIC_ICR_RR_VALID:
631 status = apic_read(APIC_RRR);
632 printk("%08x\n", status);
641 #ifdef WAKE_SECONDARY_VIA_NMI
643 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
644 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
645 * won't ... remember to clear down the APIC, etc later.
648 wakeup_secondary_cpu(int logical_apicid, unsigned long start_eip)
650 unsigned long send_status = 0, accept_status = 0;
654 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(logical_apicid));
656 /* Boot on the stack */
657 /* Kick the second */
658 apic_write_around(APIC_ICR, APIC_DM_NMI | APIC_DEST_LOGICAL);
660 Dprintk("Waiting for send to finish...\n");
665 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
666 } while (send_status && (timeout++ < 1000));
669 * Give the other CPU some time to accept the IPI.
673 * Due to the Pentium erratum 3AP.
675 maxlvt = get_maxlvt();
677 apic_read_around(APIC_SPIV);
678 apic_write(APIC_ESR, 0);
680 accept_status = (apic_read(APIC_ESR) & 0xEF);
681 Dprintk("NMI sent.\n");
684 printk("APIC never delivered???\n");
686 printk("APIC delivery error (%lx).\n", accept_status);
688 return (send_status | accept_status);
690 #endif /* WAKE_SECONDARY_VIA_NMI */
692 #ifdef WAKE_SECONDARY_VIA_INIT
694 wakeup_secondary_cpu(int phys_apicid, unsigned long start_eip)
696 unsigned long send_status = 0, accept_status = 0;
697 int maxlvt, timeout, num_starts, j;
700 * Be paranoid about clearing APIC errors.
702 if (APIC_INTEGRATED(apic_version[phys_apicid])) {
703 apic_read_around(APIC_SPIV);
704 apic_write(APIC_ESR, 0);
708 Dprintk("Asserting INIT.\n");
711 * Turn INIT on target chip
713 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
718 apic_write_around(APIC_ICR, APIC_INT_LEVELTRIG | APIC_INT_ASSERT
721 Dprintk("Waiting for send to finish...\n");
726 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
727 } while (send_status && (timeout++ < 1000));
731 Dprintk("Deasserting INIT.\n");
734 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
737 apic_write_around(APIC_ICR, APIC_INT_LEVELTRIG | APIC_DM_INIT);
739 Dprintk("Waiting for send to finish...\n");
744 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
745 } while (send_status && (timeout++ < 1000));
747 atomic_set(&init_deasserted, 1);
750 * Should we send STARTUP IPIs ?
752 * Determine this based on the APIC version.
753 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
755 if (APIC_INTEGRATED(apic_version[phys_apicid]))
761 * Run STARTUP IPI loop.
763 Dprintk("#startup loops: %d.\n", num_starts);
765 maxlvt = get_maxlvt();
767 for (j = 1; j <= num_starts; j++) {
768 Dprintk("Sending STARTUP #%d.\n",j);
769 apic_read_around(APIC_SPIV);
770 apic_write(APIC_ESR, 0);
772 Dprintk("After apic_write.\n");
779 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
781 /* Boot on the stack */
782 /* Kick the second */
783 apic_write_around(APIC_ICR, APIC_DM_STARTUP
784 | (start_eip >> 12));
787 * Give the other CPU some time to accept the IPI.
791 Dprintk("Startup point 1.\n");
793 Dprintk("Waiting for send to finish...\n");
798 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
799 } while (send_status && (timeout++ < 1000));
802 * Give the other CPU some time to accept the IPI.
806 * Due to the Pentium erratum 3AP.
809 apic_read_around(APIC_SPIV);
810 apic_write(APIC_ESR, 0);
812 accept_status = (apic_read(APIC_ESR) & 0xEF);
813 if (send_status || accept_status)
816 Dprintk("After Startup.\n");
819 printk("APIC never delivered???\n");
821 printk("APIC delivery error (%lx).\n", accept_status);
823 return (send_status | accept_status);
825 #endif /* WAKE_SECONDARY_VIA_INIT */
827 extern cpumask_t cpu_initialized;
828 static inline int alloc_cpu_id(void)
832 cpus_complement(tmp_map, cpu_present_map);
833 cpu = first_cpu(tmp_map);
839 #ifdef CONFIG_HOTPLUG_CPU
840 static struct task_struct * __devinitdata cpu_idle_tasks[NR_CPUS];
841 static inline struct task_struct * alloc_idle_task(int cpu)
843 struct task_struct *idle;
845 if ((idle = cpu_idle_tasks[cpu]) != NULL) {
846 /* initialize thread_struct. we really want to avoid destroy
849 idle->thread.esp = (unsigned long)(((struct pt_regs *)
850 (THREAD_SIZE + (unsigned long) idle->thread_info)) - 1);
851 init_idle(idle, cpu);
854 idle = fork_idle(cpu);
857 cpu_idle_tasks[cpu] = idle;
861 #define alloc_idle_task(cpu) fork_idle(cpu)
864 static int __devinit do_boot_cpu(int apicid, int cpu)
866 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
867 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
868 * Returns zero if CPU booted OK, else error code from wakeup_secondary_cpu.
871 struct task_struct *idle;
872 unsigned long boot_error;
874 unsigned long start_eip;
875 unsigned short nmi_high = 0, nmi_low = 0;
880 * We can't use kernel_thread since we must avoid to
881 * reschedule the child.
883 idle = alloc_idle_task(cpu);
885 panic("failed fork for CPU %d", cpu);
886 idle->thread.eip = (unsigned long) start_secondary;
887 /* start_eip had better be page-aligned! */
888 start_eip = setup_trampoline();
890 /* So we see what's up */
891 printk("Booting processor %d/%d eip %lx\n", cpu, apicid, start_eip);
892 /* Stack for startup_32 can be just as for start_secondary onwards */
893 stack_start.esp = (void *) idle->thread.esp;
898 * This grunge runs the startup process for
899 * the targeted processor.
902 atomic_set(&init_deasserted, 0);
904 Dprintk("Setting warm reset code and vector.\n");
906 store_NMI_vector(&nmi_high, &nmi_low);
908 smpboot_setup_warm_reset_vector(start_eip);
911 * Starting actual IPI sequence...
913 boot_error = wakeup_secondary_cpu(apicid, start_eip);
917 * allow APs to start initializing.
919 Dprintk("Before Callout %d.\n", cpu);
920 cpu_set(cpu, cpu_callout_map);
921 Dprintk("After Callout %d.\n", cpu);
924 * Wait 5s total for a response
926 for (timeout = 0; timeout < 50000; timeout++) {
927 if (cpu_isset(cpu, cpu_callin_map))
928 break; /* It has booted */
932 if (cpu_isset(cpu, cpu_callin_map)) {
933 /* number CPUs logically, starting from 1 (BSP is 0) */
935 printk("CPU%d: ", cpu);
936 print_cpu_info(&cpu_data[cpu]);
937 Dprintk("CPU has booted.\n");
940 if (*((volatile unsigned char *)trampoline_base)
942 /* trampoline started but...? */
943 printk("Stuck ??\n");
945 /* trampoline code not run */
946 printk("Not responding.\n");
947 inquire_remote_apic(apicid);
952 /* Try to put things back the way they were before ... */
953 unmap_cpu_to_logical_apicid(cpu);
954 cpu_clear(cpu, cpu_callout_map); /* was set here (do_boot_cpu()) */
955 cpu_clear(cpu, cpu_initialized); /* was set by cpu_init() */
958 x86_cpu_to_apicid[cpu] = apicid;
959 cpu_set(cpu, cpu_present_map);
962 /* mark "stuck" area as not stuck */
963 *((volatile unsigned long *)trampoline_base) = 0;
968 #ifdef CONFIG_HOTPLUG_CPU
969 void cpu_exit_clear(void)
971 int cpu = raw_smp_processor_id();
979 cpu_clear(cpu, cpu_callout_map);
980 cpu_clear(cpu, cpu_callin_map);
981 cpu_clear(cpu, cpu_present_map);
983 cpu_clear(cpu, smp_commenced_mask);
984 unmap_cpu_to_logical_apicid(cpu);
987 struct warm_boot_cpu_info {
988 struct completion *complete;
993 static void __devinit do_warm_boot_cpu(void *p)
995 struct warm_boot_cpu_info *info = p;
996 do_boot_cpu(info->apicid, info->cpu);
997 complete(info->complete);
1000 int __devinit smp_prepare_cpu(int cpu)
1002 DECLARE_COMPLETION(done);
1003 struct warm_boot_cpu_info info;
1004 struct work_struct task;
1008 apicid = x86_cpu_to_apicid[cpu];
1009 if (apicid == BAD_APICID) {
1014 info.complete = &done;
1015 info.apicid = apicid;
1017 INIT_WORK(&task, do_warm_boot_cpu, &info);
1019 tsc_sync_disabled = 1;
1021 /* init low mem mapping */
1022 clone_pgd_range(swapper_pg_dir, swapper_pg_dir + USER_PGD_PTRS,
1025 schedule_work(&task);
1026 wait_for_completion(&done);
1028 tsc_sync_disabled = 0;
1032 unlock_cpu_hotplug();
1037 static void smp_tune_scheduling (void)
1039 unsigned long cachesize; /* kB */
1040 unsigned long bandwidth = 350; /* MB/s */
1042 * Rough estimation for SMP scheduling, this is the number of
1043 * cycles it takes for a fully memory-limited process to flush
1044 * the SMP-local cache.
1046 * (For a P5 this pretty much means we will choose another idle
1047 * CPU almost always at wakeup time (this is due to the small
1048 * L1 cache), on PIIs it's around 50-100 usecs, depending on
1054 * this basically disables processor-affinity
1055 * scheduling on SMP without a TSC.
1059 cachesize = boot_cpu_data.x86_cache_size;
1060 if (cachesize == -1) {
1061 cachesize = 16; /* Pentiums, 2x8kB cache */
1068 * Cycle through the processors sending APIC IPIs to boot each.
1071 static int boot_cpu_logical_apicid;
1072 /* Where the IO area was mapped on multiquad, always 0 otherwise */
1074 #ifdef CONFIG_X86_NUMAQ
1075 EXPORT_SYMBOL(xquad_portio);
1078 static void __init smp_boot_cpus(unsigned int max_cpus)
1080 int apicid, cpu, bit, kicked;
1081 unsigned long bogosum = 0;
1084 * Setup boot CPU information
1086 smp_store_cpu_info(0); /* Final full version of the data */
1087 printk("CPU%d: ", 0);
1088 print_cpu_info(&cpu_data[0]);
1090 boot_cpu_physical_apicid = GET_APIC_ID(apic_read(APIC_ID));
1091 boot_cpu_logical_apicid = logical_smp_processor_id();
1092 x86_cpu_to_apicid[0] = boot_cpu_physical_apicid;
1094 current_thread_info()->cpu = 0;
1095 smp_tune_scheduling();
1096 cpus_clear(cpu_sibling_map[0]);
1097 cpu_set(0, cpu_sibling_map[0]);
1099 cpus_clear(cpu_core_map[0]);
1100 cpu_set(0, cpu_core_map[0]);
1103 * If we couldn't find an SMP configuration at boot time,
1104 * get out of here now!
1106 if (!smp_found_config && !acpi_lapic) {
1107 printk(KERN_NOTICE "SMP motherboard not detected.\n");
1108 smpboot_clear_io_apic_irqs();
1109 phys_cpu_present_map = physid_mask_of_physid(0);
1110 if (APIC_init_uniprocessor())
1111 printk(KERN_NOTICE "Local APIC not detected."
1112 " Using dummy APIC emulation.\n");
1113 map_cpu_to_logical_apicid();
1114 cpu_set(0, cpu_sibling_map[0]);
1115 cpu_set(0, cpu_core_map[0]);
1120 * Should not be necessary because the MP table should list the boot
1121 * CPU too, but we do it for the sake of robustness anyway.
1122 * Makes no sense to do this check in clustered apic mode, so skip it
1124 if (!check_phys_apicid_present(boot_cpu_physical_apicid)) {
1125 printk("weird, boot CPU (#%d) not listed by the BIOS.\n",
1126 boot_cpu_physical_apicid);
1127 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1131 * If we couldn't find a local APIC, then get out of here now!
1133 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid]) && !cpu_has_apic) {
1134 printk(KERN_ERR "BIOS bug, local APIC #%d not detected!...\n",
1135 boot_cpu_physical_apicid);
1136 printk(KERN_ERR "... forcing use of dummy APIC emulation. (tell your hw vendor)\n");
1137 smpboot_clear_io_apic_irqs();
1138 phys_cpu_present_map = physid_mask_of_physid(0);
1139 cpu_set(0, cpu_sibling_map[0]);
1140 cpu_set(0, cpu_core_map[0]);
1144 verify_local_APIC();
1147 * If SMP should be disabled, then really disable it!
1150 smp_found_config = 0;
1151 printk(KERN_INFO "SMP mode deactivated, forcing use of dummy APIC emulation.\n");
1152 smpboot_clear_io_apic_irqs();
1153 phys_cpu_present_map = physid_mask_of_physid(0);
1154 cpu_set(0, cpu_sibling_map[0]);
1155 cpu_set(0, cpu_core_map[0]);
1161 map_cpu_to_logical_apicid();
1164 setup_portio_remap();
1167 * Scan the CPU present map and fire up the other CPUs via do_boot_cpu
1169 * In clustered apic mode, phys_cpu_present_map is a constructed thus:
1170 * bits 0-3 are quad0, 4-7 are quad1, etc. A perverse twist on the
1171 * clustered apic ID.
1173 Dprintk("CPU present map: %lx\n", physids_coerce(phys_cpu_present_map));
1176 for (bit = 0; kicked < NR_CPUS && bit < MAX_APICS; bit++) {
1177 apicid = cpu_present_to_apicid(bit);
1179 * Don't even attempt to start the boot CPU!
1181 if ((apicid == boot_cpu_apicid) || (apicid == BAD_APICID))
1184 if (!check_apicid_present(bit))
1186 if (max_cpus <= cpucount+1)
1189 if (((cpu = alloc_cpu_id()) <= 0) || do_boot_cpu(apicid, cpu))
1190 printk("CPU #%d not responding - cannot use it.\n",
1197 * Cleanup possible dangling ends...
1199 smpboot_restore_warm_reset_vector();
1202 * Allow the user to impress friends.
1204 Dprintk("Before bogomips.\n");
1205 for (cpu = 0; cpu < NR_CPUS; cpu++)
1206 if (cpu_isset(cpu, cpu_callout_map))
1207 bogosum += cpu_data[cpu].loops_per_jiffy;
1209 "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
1211 bogosum/(500000/HZ),
1212 (bogosum/(5000/HZ))%100);
1214 Dprintk("Before bogocount - setting activated=1.\n");
1217 printk(KERN_WARNING "WARNING: SMP operation may be unreliable with B stepping processors.\n");
1220 * Don't taint if we are running SMP kernel on a single non-MP
1223 if (tainted & TAINT_UNSAFE_SMP) {
1225 printk (KERN_INFO "WARNING: This combination of AMD processors is not suitable for SMP.\n");
1227 tainted &= ~TAINT_UNSAFE_SMP;
1230 Dprintk("Boot done.\n");
1233 * construct cpu_sibling_map[], so that we can tell sibling CPUs
1236 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1237 cpus_clear(cpu_sibling_map[cpu]);
1238 cpus_clear(cpu_core_map[cpu]);
1241 cpu_set(0, cpu_sibling_map[0]);
1242 cpu_set(0, cpu_core_map[0]);
1244 smpboot_setup_io_apic();
1246 setup_boot_APIC_clock();
1249 * Synchronize the TSC with the AP
1251 if (cpu_has_tsc && cpucount && cpu_khz)
1252 synchronize_tsc_bp();
1255 /* These are wrappers to interface to the new boot process. Someone
1256 who understands all this stuff should rewrite it properly. --RR 15/Jul/02 */
1257 void __init smp_prepare_cpus(unsigned int max_cpus)
1259 smp_commenced_mask = cpumask_of_cpu(0);
1260 cpu_callin_map = cpumask_of_cpu(0);
1262 smp_boot_cpus(max_cpus);
1265 void __devinit smp_prepare_boot_cpu(void)
1267 cpu_set(smp_processor_id(), cpu_online_map);
1268 cpu_set(smp_processor_id(), cpu_callout_map);
1269 cpu_set(smp_processor_id(), cpu_present_map);
1270 cpu_set(smp_processor_id(), cpu_possible_map);
1271 per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
1274 #ifdef CONFIG_HOTPLUG_CPU
1276 remove_siblinginfo(int cpu)
1280 for_each_cpu_mask(sibling, cpu_sibling_map[cpu])
1281 cpu_clear(cpu, cpu_sibling_map[sibling]);
1282 for_each_cpu_mask(sibling, cpu_core_map[cpu])
1283 cpu_clear(cpu, cpu_core_map[sibling]);
1284 cpus_clear(cpu_sibling_map[cpu]);
1285 cpus_clear(cpu_core_map[cpu]);
1286 phys_proc_id[cpu] = BAD_APICID;
1287 cpu_core_id[cpu] = BAD_APICID;
1290 int __cpu_disable(void)
1292 cpumask_t map = cpu_online_map;
1293 int cpu = smp_processor_id();
1296 * Perhaps use cpufreq to drop frequency, but that could go
1297 * into generic code.
1299 * We won't take down the boot processor on i386 due to some
1300 * interrupts only being able to be serviced by the BSP.
1301 * Especially so if we're not using an IOAPIC -zwane
1306 /* We enable the timer again on the exit path of the death loop */
1307 disable_APIC_timer();
1308 /* Allow any queued timer interrupts to get serviced */
1311 local_irq_disable();
1313 remove_siblinginfo(cpu);
1315 cpu_clear(cpu, map);
1317 /* It's now safe to remove this processor from the online map */
1318 cpu_clear(cpu, cpu_online_map);
1322 void __cpu_die(unsigned int cpu)
1324 /* We don't do anything here: idle task is faking death itself. */
1327 for (i = 0; i < 10; i++) {
1328 /* They ack this in play_dead by setting CPU_DEAD */
1329 if (per_cpu(cpu_state, cpu) == CPU_DEAD) {
1330 printk ("CPU %d is now offline\n", cpu);
1335 printk(KERN_ERR "CPU %u didn't die...\n", cpu);
1337 #else /* ... !CONFIG_HOTPLUG_CPU */
1338 int __cpu_disable(void)
1343 void __cpu_die(unsigned int cpu)
1345 /* We said "no" in __cpu_disable */
1348 #endif /* CONFIG_HOTPLUG_CPU */
1350 int __devinit __cpu_up(unsigned int cpu)
1352 /* In case one didn't come up */
1353 if (!cpu_isset(cpu, cpu_callin_map)) {
1354 printk(KERN_DEBUG "skipping cpu%d, didn't come online\n", cpu);
1360 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
1361 /* Unleash the CPU! */
1362 cpu_set(cpu, smp_commenced_mask);
1363 while (!cpu_isset(cpu, cpu_online_map))
1368 void __init smp_cpus_done(unsigned int max_cpus)
1370 #ifdef CONFIG_X86_IO_APIC
1371 setup_ioapic_dest();
1374 #ifndef CONFIG_HOTPLUG_CPU
1376 * Disable executability of the SMP trampoline:
1378 set_kernel_exec((unsigned long)trampoline_base, trampoline_exec);
1382 void __init smp_intr_init(void)
1385 * IRQ0 must be given a fixed assignment and initialized,
1386 * because it's used before the IO-APIC is set up.
1388 set_intr_gate(FIRST_DEVICE_VECTOR, interrupt[0]);
1391 * The reschedule interrupt is a CPU-to-CPU reschedule-helper
1392 * IPI, driven by wakeup.
1394 set_intr_gate(RESCHEDULE_VECTOR, reschedule_interrupt);
1396 /* IPI for invalidation */
1397 set_intr_gate(INVALIDATE_TLB_VECTOR, invalidate_interrupt);
1399 /* IPI for generic function call */
1400 set_intr_gate(CALL_FUNCTION_VECTOR, call_function_interrupt);