2 * x86 SMP booting functions
4 * (c) 1995 Alan Cox, Building #3 <alan@redhat.com>
5 * (c) 1998, 1999, 2000 Ingo Molnar <mingo@redhat.com>
7 * Much of the core SMP work is based on previous work by Thomas Radke, to
8 * whom a great many thanks are extended.
10 * Thanks to Intel for making available several different Pentium,
11 * Pentium Pro and Pentium-II/Xeon MP machines.
12 * Original development of Linux SMP code supported by Caldera.
14 * This code is released under the GNU General Public License version 2 or
18 * Felix Koop : NR_CPUS used properly
19 * Jose Renau : Handle single CPU case.
20 * Alan Cox : By repeated request 8) - Total BogoMIPS report.
21 * Greg Wright : Fix for kernel stacks panic.
22 * Erich Boleyn : MP v1.4 and additional changes.
23 * Matthias Sattler : Changes for 2.1 kernel map.
24 * Michel Lespinasse : Changes for 2.1 kernel map.
25 * Michael Chastain : Change trampoline.S to gnu as.
26 * Alan Cox : Dumb bug: 'B' step PPro's are fine
27 * Ingo Molnar : Added APIC timers, based on code
29 * Ingo Molnar : various cleanups and rewrites
30 * Tigran Aivazian : fixed "0.00 in /proc/uptime on SMP" bug.
31 * Maciej W. Rozycki : Bits for genuine 82489DX APICs
32 * Martin J. Bligh : Added support for multi-quad systems
33 * Dave Jones : Report invalid combinations of Athlon CPUs.
34 * Rusty Russell : Hacked into shape for new "hotplug" boot process. */
36 #include <linux/module.h>
37 #include <linux/config.h>
38 #include <linux/init.h>
39 #include <linux/kernel.h>
42 #include <linux/sched.h>
43 #include <linux/kernel_stat.h>
44 #include <linux/smp_lock.h>
45 #include <linux/bootmem.h>
46 #include <linux/notifier.h>
47 #include <linux/cpu.h>
48 #include <linux/percpu.h>
50 #include <linux/delay.h>
51 #include <linux/mc146818rtc.h>
52 #include <asm/tlbflush.h>
54 #include <asm/arch_hooks.h>
56 #include <mach_apic.h>
57 #include <mach_wakecpu.h>
58 #include <smpboot_hooks.h>
60 /* Set if we find a B stepping CPU */
61 static int __devinitdata smp_b_stepping;
63 /* Number of siblings per CPU package */
64 int smp_num_siblings = 1;
66 EXPORT_SYMBOL(smp_num_siblings);
69 /* Package ID of each logical CPU */
70 int phys_proc_id[NR_CPUS] __read_mostly = {[0 ... NR_CPUS-1] = BAD_APICID};
72 /* Core ID of each logical CPU */
73 int cpu_core_id[NR_CPUS] __read_mostly = {[0 ... NR_CPUS-1] = BAD_APICID};
75 cpumask_t cpu_sibling_map[NR_CPUS] __read_mostly;
76 EXPORT_SYMBOL(cpu_sibling_map);
78 cpumask_t cpu_core_map[NR_CPUS] __read_mostly;
79 EXPORT_SYMBOL(cpu_core_map);
81 /* bitmap of online cpus */
82 cpumask_t cpu_online_map __read_mostly;
83 EXPORT_SYMBOL(cpu_online_map);
85 cpumask_t cpu_callin_map;
86 cpumask_t cpu_callout_map;
87 EXPORT_SYMBOL(cpu_callout_map);
88 #ifdef CONFIG_HOTPLUG_CPU
89 cpumask_t cpu_possible_map = CPU_MASK_ALL;
91 cpumask_t cpu_possible_map;
93 EXPORT_SYMBOL(cpu_possible_map);
94 static cpumask_t smp_commenced_mask;
96 /* TSC's upper 32 bits can't be written in eariler CPU (before prescott), there
97 * is no way to resync one AP against BP. TBD: for prescott and above, we
98 * should use IA64's algorithm
100 static int __devinitdata tsc_sync_disabled;
102 /* Per CPU bogomips and other parameters */
103 struct cpuinfo_x86 cpu_data[NR_CPUS] __cacheline_aligned;
104 EXPORT_SYMBOL(cpu_data);
106 u8 x86_cpu_to_apicid[NR_CPUS] __read_mostly =
107 { [0 ... NR_CPUS-1] = 0xff };
108 EXPORT_SYMBOL(x86_cpu_to_apicid);
111 * Trampoline 80x86 program as an array.
114 extern unsigned char trampoline_data [];
115 extern unsigned char trampoline_end [];
116 static unsigned char *trampoline_base;
117 static int trampoline_exec;
119 static void map_cpu_to_logical_apicid(void);
121 /* State of each CPU. */
122 DEFINE_PER_CPU(int, cpu_state) = { 0 };
125 * Currently trivial. Write the real->protected mode
126 * bootstrap into the page concerned. The caller
127 * has made sure it's suitably aligned.
130 static unsigned long __devinit setup_trampoline(void)
132 memcpy(trampoline_base, trampoline_data, trampoline_end - trampoline_data);
133 return virt_to_phys(trampoline_base);
137 * We are called very early to get the low memory for the
138 * SMP bootup trampoline page.
140 void __init smp_alloc_memory(void)
142 trampoline_base = (void *) alloc_bootmem_low_pages(PAGE_SIZE);
144 * Has to be in very low memory so we can execute
147 if (__pa(trampoline_base) >= 0x9F000)
150 * Make the SMP trampoline executable:
152 trampoline_exec = set_kernel_exec((unsigned long)trampoline_base, 1);
156 * The bootstrap kernel entry code has set these up. Save them for
160 static void __devinit smp_store_cpu_info(int id)
162 struct cpuinfo_x86 *c = cpu_data + id;
168 * Mask B, Pentium, but not Pentium MMX
170 if (c->x86_vendor == X86_VENDOR_INTEL &&
172 c->x86_mask >= 1 && c->x86_mask <= 4 &&
175 * Remember we have B step Pentia with bugs
180 * Certain Athlons might work (for various values of 'work') in SMP
181 * but they are not certified as MP capable.
183 if ((c->x86_vendor == X86_VENDOR_AMD) && (c->x86 == 6)) {
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.
220 static atomic_t tsc_start_flag = ATOMIC_INIT(0);
221 static atomic_t tsc_count_start = ATOMIC_INIT(0);
222 static atomic_t tsc_count_stop = ATOMIC_INIT(0);
223 static unsigned long long tsc_values[NR_CPUS];
227 static void __init synchronize_tsc_bp (void)
230 unsigned long long t0;
231 unsigned long long sum, avg;
233 unsigned int one_usec;
236 printk(KERN_INFO "checking TSC synchronization across %u CPUs: ", num_booting_cpus());
238 /* convert from kcyc/sec to cyc/usec */
239 one_usec = cpu_khz / 1000;
241 atomic_set(&tsc_start_flag, 1);
245 * We loop a few times to get a primed instruction cache,
246 * then the last pass is more or less synchronized and
247 * the BP and APs set their cycle counters to zero all at
248 * once. This reduces the chance of having random offsets
249 * between the processors, and guarantees that the maximum
250 * delay between the cycle counters is never bigger than
251 * the latency of information-passing (cachelines) between
254 for (i = 0; i < NR_LOOPS; i++) {
256 * all APs synchronize but they loop on '== num_cpus'
258 while (atomic_read(&tsc_count_start) != num_booting_cpus()-1)
260 atomic_set(&tsc_count_stop, 0);
263 * this lets the APs save their current TSC:
265 atomic_inc(&tsc_count_start);
267 rdtscll(tsc_values[smp_processor_id()]);
269 * We clear the TSC in the last loop:
275 * Wait for all APs to leave the synchronization point:
277 while (atomic_read(&tsc_count_stop) != num_booting_cpus()-1)
279 atomic_set(&tsc_count_start, 0);
281 atomic_inc(&tsc_count_stop);
285 for (i = 0; i < NR_CPUS; i++) {
286 if (cpu_isset(i, cpu_callout_map)) {
292 do_div(avg, num_booting_cpus());
295 for (i = 0; i < NR_CPUS; i++) {
296 if (!cpu_isset(i, cpu_callout_map))
298 delta = tsc_values[i] - avg;
302 * We report bigger than 2 microseconds clock differences.
304 if (delta > 2*one_usec) {
311 do_div(realdelta, one_usec);
312 if (tsc_values[i] < avg)
313 realdelta = -realdelta;
315 printk(KERN_INFO "CPU#%d had %ld usecs TSC skew, fixed it up.\n", i, realdelta);
324 static void __init synchronize_tsc_ap (void)
329 * Not every cpu is online at the time
330 * this gets called, so we first wait for the BP to
331 * finish SMP initialization:
333 while (!atomic_read(&tsc_start_flag)) mb();
335 for (i = 0; i < NR_LOOPS; i++) {
336 atomic_inc(&tsc_count_start);
337 while (atomic_read(&tsc_count_start) != num_booting_cpus())
340 rdtscll(tsc_values[smp_processor_id()]);
344 atomic_inc(&tsc_count_stop);
345 while (atomic_read(&tsc_count_stop) != num_booting_cpus()) mb();
350 extern void calibrate_delay(void);
352 static atomic_t init_deasserted;
354 static void __devinit smp_callin(void)
357 unsigned long timeout;
360 * If waken up by an INIT in an 82489DX configuration
361 * we may get here before an INIT-deassert IPI reaches
362 * our local APIC. We have to wait for the IPI or we'll
363 * lock up on an APIC access.
365 wait_for_init_deassert(&init_deasserted);
368 * (This works even if the APIC is not enabled.)
370 phys_id = GET_APIC_ID(apic_read(APIC_ID));
371 cpuid = smp_processor_id();
372 if (cpu_isset(cpuid, cpu_callin_map)) {
373 printk("huh, phys CPU#%d, CPU#%d already present??\n",
377 Dprintk("CPU#%d (phys ID: %d) waiting for CALLOUT\n", cpuid, phys_id);
380 * STARTUP IPIs are fragile beasts as they might sometimes
381 * trigger some glue motherboard logic. Complete APIC bus
382 * silence for 1 second, this overestimates the time the
383 * boot CPU is spending to send the up to 2 STARTUP IPIs
384 * by a factor of two. This should be enough.
388 * Waiting 2s total for startup (udelay is not yet working)
390 timeout = jiffies + 2*HZ;
391 while (time_before(jiffies, timeout)) {
393 * Has the boot CPU finished it's STARTUP sequence?
395 if (cpu_isset(cpuid, cpu_callout_map))
400 if (!time_before(jiffies, timeout)) {
401 printk("BUG: CPU%d started up but did not get a callout!\n",
407 * the boot CPU has finished the init stage and is spinning
408 * on callin_map until we finish. We are free to set up this
409 * CPU, first the APIC. (this is probably redundant on most
413 Dprintk("CALLIN, before setup_local_APIC().\n");
414 smp_callin_clear_local_apic();
416 map_cpu_to_logical_apicid();
422 Dprintk("Stack at about %p\n",&cpuid);
425 * Save our processor parameters
427 smp_store_cpu_info(cpuid);
429 disable_APIC_timer();
432 * Allow the master to continue.
434 cpu_set(cpuid, cpu_callin_map);
437 * Synchronize the TSC with the BP
439 if (cpu_has_tsc && cpu_khz && !tsc_sync_disabled)
440 synchronize_tsc_ap();
446 set_cpu_sibling_map(int cpu)
450 if (smp_num_siblings > 1) {
451 for (i = 0; i < NR_CPUS; i++) {
452 if (!cpu_isset(i, cpu_callout_map))
454 if (cpu_core_id[cpu] == cpu_core_id[i]) {
455 cpu_set(i, cpu_sibling_map[cpu]);
456 cpu_set(cpu, cpu_sibling_map[i]);
460 cpu_set(cpu, cpu_sibling_map[cpu]);
463 if (current_cpu_data.x86_num_cores > 1) {
464 for (i = 0; i < NR_CPUS; i++) {
465 if (!cpu_isset(i, cpu_callout_map))
467 if (phys_proc_id[cpu] == phys_proc_id[i]) {
468 cpu_set(i, cpu_core_map[cpu]);
469 cpu_set(cpu, cpu_core_map[i]);
473 cpu_core_map[cpu] = cpu_sibling_map[cpu];
478 * Activate a secondary processor.
480 static void __devinit start_secondary(void *unused)
483 * Dont put anything before smp_callin(), SMP
484 * booting is too fragile that we want to limit the
485 * things done here to the most necessary things.
490 while (!cpu_isset(smp_processor_id(), smp_commenced_mask))
492 setup_secondary_APIC_clock();
493 if (nmi_watchdog == NMI_IO_APIC) {
494 disable_8259A_irq(0);
495 enable_NMI_through_LVT0(NULL);
500 * low-memory mappings have been cleared, flush them from
501 * the local TLBs too.
505 /* This must be done before setting cpu_online_map */
506 set_cpu_sibling_map(raw_smp_processor_id());
510 * We need to hold call_lock, so there is no inconsistency
511 * between the time smp_call_function() determines number of
512 * IPI receipients, and the time when the determination is made
513 * for which cpus receive the IPI. Holding this
514 * lock helps us to not include this cpu in a currently in progress
515 * smp_call_function().
517 lock_ipi_call_lock();
518 cpu_set(smp_processor_id(), cpu_online_map);
519 unlock_ipi_call_lock();
520 per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
522 /* We can take interrupts now: we're officially "up". */
530 * Everything has been set up for the secondary
531 * CPUs - they just need to reload everything
532 * from the task structure
533 * This function must not return.
535 void __devinit initialize_secondary(void)
538 * We don't actually need to load the full TSS,
539 * basically just the stack pointer and the eip.
546 :"r" (current->thread.esp),"r" (current->thread.eip));
556 /* which logical CPUs are on which nodes */
557 cpumask_t node_2_cpu_mask[MAX_NUMNODES] __read_mostly =
558 { [0 ... MAX_NUMNODES-1] = CPU_MASK_NONE };
559 /* which node each logical CPU is on */
560 int cpu_2_node[NR_CPUS] __read_mostly = { [0 ... NR_CPUS-1] = 0 };
561 EXPORT_SYMBOL(cpu_2_node);
563 /* set up a mapping between cpu and node. */
564 static inline void map_cpu_to_node(int cpu, int node)
566 printk("Mapping cpu %d to node %d\n", cpu, node);
567 cpu_set(cpu, node_2_cpu_mask[node]);
568 cpu_2_node[cpu] = node;
571 /* undo a mapping between cpu and node. */
572 static inline void unmap_cpu_to_node(int cpu)
576 printk("Unmapping cpu %d from all nodes\n", cpu);
577 for (node = 0; node < MAX_NUMNODES; node ++)
578 cpu_clear(cpu, node_2_cpu_mask[node]);
581 #else /* !CONFIG_NUMA */
583 #define map_cpu_to_node(cpu, node) ({})
584 #define unmap_cpu_to_node(cpu) ({})
586 #endif /* CONFIG_NUMA */
588 u8 cpu_2_logical_apicid[NR_CPUS] __read_mostly = { [0 ... NR_CPUS-1] = BAD_APICID };
590 static void map_cpu_to_logical_apicid(void)
592 int cpu = smp_processor_id();
593 int apicid = logical_smp_processor_id();
595 cpu_2_logical_apicid[cpu] = apicid;
596 map_cpu_to_node(cpu, apicid_to_node(apicid));
599 static void unmap_cpu_to_logical_apicid(int cpu)
601 cpu_2_logical_apicid[cpu] = BAD_APICID;
602 unmap_cpu_to_node(cpu);
606 static inline void __inquire_remote_apic(int apicid)
608 int i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
609 char *names[] = { "ID", "VERSION", "SPIV" };
612 printk("Inquiring remote APIC #%d...\n", apicid);
614 for (i = 0; i < ARRAY_SIZE(regs); i++) {
615 printk("... APIC #%d %s: ", apicid, names[i]);
620 apic_wait_icr_idle();
622 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(apicid));
623 apic_write_around(APIC_ICR, APIC_DM_REMRD | regs[i]);
628 status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
629 } while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
632 case APIC_ICR_RR_VALID:
633 status = apic_read(APIC_RRR);
634 printk("%08x\n", status);
643 #ifdef WAKE_SECONDARY_VIA_NMI
645 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
646 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
647 * won't ... remember to clear down the APIC, etc later.
650 wakeup_secondary_cpu(int logical_apicid, unsigned long start_eip)
652 unsigned long send_status = 0, accept_status = 0;
656 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(logical_apicid));
658 /* Boot on the stack */
659 /* Kick the second */
660 apic_write_around(APIC_ICR, APIC_DM_NMI | APIC_DEST_LOGICAL);
662 Dprintk("Waiting for send to finish...\n");
667 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
668 } while (send_status && (timeout++ < 1000));
671 * Give the other CPU some time to accept the IPI.
675 * Due to the Pentium erratum 3AP.
677 maxlvt = get_maxlvt();
679 apic_read_around(APIC_SPIV);
680 apic_write(APIC_ESR, 0);
682 accept_status = (apic_read(APIC_ESR) & 0xEF);
683 Dprintk("NMI sent.\n");
686 printk("APIC never delivered???\n");
688 printk("APIC delivery error (%lx).\n", accept_status);
690 return (send_status | accept_status);
692 #endif /* WAKE_SECONDARY_VIA_NMI */
694 #ifdef WAKE_SECONDARY_VIA_INIT
696 wakeup_secondary_cpu(int phys_apicid, unsigned long start_eip)
698 unsigned long send_status = 0, accept_status = 0;
699 int maxlvt, timeout, num_starts, j;
702 * Be paranoid about clearing APIC errors.
704 if (APIC_INTEGRATED(apic_version[phys_apicid])) {
705 apic_read_around(APIC_SPIV);
706 apic_write(APIC_ESR, 0);
710 Dprintk("Asserting INIT.\n");
713 * Turn INIT on target chip
715 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
720 apic_write_around(APIC_ICR, APIC_INT_LEVELTRIG | APIC_INT_ASSERT
723 Dprintk("Waiting for send to finish...\n");
728 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
729 } while (send_status && (timeout++ < 1000));
733 Dprintk("Deasserting INIT.\n");
736 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
739 apic_write_around(APIC_ICR, APIC_INT_LEVELTRIG | APIC_DM_INIT);
741 Dprintk("Waiting for send to finish...\n");
746 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
747 } while (send_status && (timeout++ < 1000));
749 atomic_set(&init_deasserted, 1);
752 * Should we send STARTUP IPIs ?
754 * Determine this based on the APIC version.
755 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
757 if (APIC_INTEGRATED(apic_version[phys_apicid]))
763 * Run STARTUP IPI loop.
765 Dprintk("#startup loops: %d.\n", num_starts);
767 maxlvt = get_maxlvt();
769 for (j = 1; j <= num_starts; j++) {
770 Dprintk("Sending STARTUP #%d.\n",j);
771 apic_read_around(APIC_SPIV);
772 apic_write(APIC_ESR, 0);
774 Dprintk("After apic_write.\n");
781 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
783 /* Boot on the stack */
784 /* Kick the second */
785 apic_write_around(APIC_ICR, APIC_DM_STARTUP
786 | (start_eip >> 12));
789 * Give the other CPU some time to accept the IPI.
793 Dprintk("Startup point 1.\n");
795 Dprintk("Waiting for send to finish...\n");
800 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
801 } while (send_status && (timeout++ < 1000));
804 * Give the other CPU some time to accept the IPI.
808 * Due to the Pentium erratum 3AP.
811 apic_read_around(APIC_SPIV);
812 apic_write(APIC_ESR, 0);
814 accept_status = (apic_read(APIC_ESR) & 0xEF);
815 if (send_status || accept_status)
818 Dprintk("After Startup.\n");
821 printk("APIC never delivered???\n");
823 printk("APIC delivery error (%lx).\n", accept_status);
825 return (send_status | accept_status);
827 #endif /* WAKE_SECONDARY_VIA_INIT */
829 extern cpumask_t cpu_initialized;
830 static inline int alloc_cpu_id(void)
834 cpus_complement(tmp_map, cpu_present_map);
835 cpu = first_cpu(tmp_map);
841 #ifdef CONFIG_HOTPLUG_CPU
842 static struct task_struct * __devinitdata cpu_idle_tasks[NR_CPUS];
843 static inline struct task_struct * alloc_idle_task(int cpu)
845 struct task_struct *idle;
847 if ((idle = cpu_idle_tasks[cpu]) != NULL) {
848 /* initialize thread_struct. we really want to avoid destroy
851 idle->thread.esp = (unsigned long)(((struct pt_regs *)
852 (THREAD_SIZE + (unsigned long) idle->thread_info)) - 1);
853 init_idle(idle, cpu);
856 idle = fork_idle(cpu);
859 cpu_idle_tasks[cpu] = idle;
863 #define alloc_idle_task(cpu) fork_idle(cpu)
866 static int __devinit do_boot_cpu(int apicid, int cpu)
868 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
869 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
870 * Returns zero if CPU booted OK, else error code from wakeup_secondary_cpu.
873 struct task_struct *idle;
874 unsigned long boot_error;
876 unsigned long start_eip;
877 unsigned short nmi_high = 0, nmi_low = 0;
882 * We can't use kernel_thread since we must avoid to
883 * reschedule the child.
885 idle = alloc_idle_task(cpu);
887 panic("failed fork for CPU %d", cpu);
888 idle->thread.eip = (unsigned long) start_secondary;
889 /* start_eip had better be page-aligned! */
890 start_eip = setup_trampoline();
892 /* So we see what's up */
893 printk("Booting processor %d/%d eip %lx\n", cpu, apicid, start_eip);
894 /* Stack for startup_32 can be just as for start_secondary onwards */
895 stack_start.esp = (void *) idle->thread.esp;
900 * This grunge runs the startup process for
901 * the targeted processor.
904 atomic_set(&init_deasserted, 0);
906 Dprintk("Setting warm reset code and vector.\n");
908 store_NMI_vector(&nmi_high, &nmi_low);
910 smpboot_setup_warm_reset_vector(start_eip);
913 * Starting actual IPI sequence...
915 boot_error = wakeup_secondary_cpu(apicid, start_eip);
919 * allow APs to start initializing.
921 Dprintk("Before Callout %d.\n", cpu);
922 cpu_set(cpu, cpu_callout_map);
923 Dprintk("After Callout %d.\n", cpu);
926 * Wait 5s total for a response
928 for (timeout = 0; timeout < 50000; timeout++) {
929 if (cpu_isset(cpu, cpu_callin_map))
930 break; /* It has booted */
934 if (cpu_isset(cpu, cpu_callin_map)) {
935 /* number CPUs logically, starting from 1 (BSP is 0) */
937 printk("CPU%d: ", cpu);
938 print_cpu_info(&cpu_data[cpu]);
939 Dprintk("CPU has booted.\n");
942 if (*((volatile unsigned char *)trampoline_base)
944 /* trampoline started but...? */
945 printk("Stuck ??\n");
947 /* trampoline code not run */
948 printk("Not responding.\n");
949 inquire_remote_apic(apicid);
954 /* Try to put things back the way they were before ... */
955 unmap_cpu_to_logical_apicid(cpu);
956 cpu_clear(cpu, cpu_callout_map); /* was set here (do_boot_cpu()) */
957 cpu_clear(cpu, cpu_initialized); /* was set by cpu_init() */
960 x86_cpu_to_apicid[cpu] = apicid;
961 cpu_set(cpu, cpu_present_map);
964 /* mark "stuck" area as not stuck */
965 *((volatile unsigned long *)trampoline_base) = 0;
970 #ifdef CONFIG_HOTPLUG_CPU
971 void cpu_exit_clear(void)
973 int cpu = raw_smp_processor_id();
981 cpu_clear(cpu, cpu_callout_map);
982 cpu_clear(cpu, cpu_callin_map);
983 cpu_clear(cpu, cpu_present_map);
985 cpu_clear(cpu, smp_commenced_mask);
986 unmap_cpu_to_logical_apicid(cpu);
989 struct warm_boot_cpu_info {
990 struct completion *complete;
995 static void __devinit do_warm_boot_cpu(void *p)
997 struct warm_boot_cpu_info *info = p;
998 do_boot_cpu(info->apicid, info->cpu);
999 complete(info->complete);
1002 int __devinit smp_prepare_cpu(int cpu)
1004 DECLARE_COMPLETION(done);
1005 struct warm_boot_cpu_info info;
1006 struct work_struct task;
1010 apicid = x86_cpu_to_apicid[cpu];
1011 if (apicid == BAD_APICID) {
1016 info.complete = &done;
1017 info.apicid = apicid;
1019 INIT_WORK(&task, do_warm_boot_cpu, &info);
1021 tsc_sync_disabled = 1;
1023 /* init low mem mapping */
1024 clone_pgd_range(swapper_pg_dir, swapper_pg_dir + USER_PGD_PTRS,
1027 schedule_work(&task);
1028 wait_for_completion(&done);
1030 tsc_sync_disabled = 0;
1034 unlock_cpu_hotplug();
1039 static void smp_tune_scheduling (void)
1041 unsigned long cachesize; /* kB */
1042 unsigned long bandwidth = 350; /* MB/s */
1044 * Rough estimation for SMP scheduling, this is the number of
1045 * cycles it takes for a fully memory-limited process to flush
1046 * the SMP-local cache.
1048 * (For a P5 this pretty much means we will choose another idle
1049 * CPU almost always at wakeup time (this is due to the small
1050 * L1 cache), on PIIs it's around 50-100 usecs, depending on
1056 * this basically disables processor-affinity
1057 * scheduling on SMP without a TSC.
1061 cachesize = boot_cpu_data.x86_cache_size;
1062 if (cachesize == -1) {
1063 cachesize = 16; /* Pentiums, 2x8kB cache */
1070 * Cycle through the processors sending APIC IPIs to boot each.
1073 static int boot_cpu_logical_apicid;
1074 /* Where the IO area was mapped on multiquad, always 0 otherwise */
1076 #ifdef CONFIG_X86_NUMAQ
1077 EXPORT_SYMBOL(xquad_portio);
1080 static void __init smp_boot_cpus(unsigned int max_cpus)
1082 int apicid, cpu, bit, kicked;
1083 unsigned long bogosum = 0;
1086 * Setup boot CPU information
1088 smp_store_cpu_info(0); /* Final full version of the data */
1089 printk("CPU%d: ", 0);
1090 print_cpu_info(&cpu_data[0]);
1092 boot_cpu_physical_apicid = GET_APIC_ID(apic_read(APIC_ID));
1093 boot_cpu_logical_apicid = logical_smp_processor_id();
1094 x86_cpu_to_apicid[0] = boot_cpu_physical_apicid;
1096 current_thread_info()->cpu = 0;
1097 smp_tune_scheduling();
1098 cpus_clear(cpu_sibling_map[0]);
1099 cpu_set(0, cpu_sibling_map[0]);
1101 cpus_clear(cpu_core_map[0]);
1102 cpu_set(0, cpu_core_map[0]);
1105 * If we couldn't find an SMP configuration at boot time,
1106 * get out of here now!
1108 if (!smp_found_config && !acpi_lapic) {
1109 printk(KERN_NOTICE "SMP motherboard not detected.\n");
1110 smpboot_clear_io_apic_irqs();
1111 phys_cpu_present_map = physid_mask_of_physid(0);
1112 if (APIC_init_uniprocessor())
1113 printk(KERN_NOTICE "Local APIC not detected."
1114 " Using dummy APIC emulation.\n");
1115 map_cpu_to_logical_apicid();
1116 cpu_set(0, cpu_sibling_map[0]);
1117 cpu_set(0, cpu_core_map[0]);
1122 * Should not be necessary because the MP table should list the boot
1123 * CPU too, but we do it for the sake of robustness anyway.
1124 * Makes no sense to do this check in clustered apic mode, so skip it
1126 if (!check_phys_apicid_present(boot_cpu_physical_apicid)) {
1127 printk("weird, boot CPU (#%d) not listed by the BIOS.\n",
1128 boot_cpu_physical_apicid);
1129 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1133 * If we couldn't find a local APIC, then get out of here now!
1135 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid]) && !cpu_has_apic) {
1136 printk(KERN_ERR "BIOS bug, local APIC #%d not detected!...\n",
1137 boot_cpu_physical_apicid);
1138 printk(KERN_ERR "... forcing use of dummy APIC emulation. (tell your hw vendor)\n");
1139 smpboot_clear_io_apic_irqs();
1140 phys_cpu_present_map = physid_mask_of_physid(0);
1141 cpu_set(0, cpu_sibling_map[0]);
1142 cpu_set(0, cpu_core_map[0]);
1146 verify_local_APIC();
1149 * If SMP should be disabled, then really disable it!
1152 smp_found_config = 0;
1153 printk(KERN_INFO "SMP mode deactivated, forcing use of dummy APIC emulation.\n");
1154 smpboot_clear_io_apic_irqs();
1155 phys_cpu_present_map = physid_mask_of_physid(0);
1156 cpu_set(0, cpu_sibling_map[0]);
1157 cpu_set(0, cpu_core_map[0]);
1163 map_cpu_to_logical_apicid();
1166 setup_portio_remap();
1169 * Scan the CPU present map and fire up the other CPUs via do_boot_cpu
1171 * In clustered apic mode, phys_cpu_present_map is a constructed thus:
1172 * bits 0-3 are quad0, 4-7 are quad1, etc. A perverse twist on the
1173 * clustered apic ID.
1175 Dprintk("CPU present map: %lx\n", physids_coerce(phys_cpu_present_map));
1178 for (bit = 0; kicked < NR_CPUS && bit < MAX_APICS; bit++) {
1179 apicid = cpu_present_to_apicid(bit);
1181 * Don't even attempt to start the boot CPU!
1183 if ((apicid == boot_cpu_apicid) || (apicid == BAD_APICID))
1186 if (!check_apicid_present(bit))
1188 if (max_cpus <= cpucount+1)
1191 if (((cpu = alloc_cpu_id()) <= 0) || do_boot_cpu(apicid, cpu))
1192 printk("CPU #%d not responding - cannot use it.\n",
1199 * Cleanup possible dangling ends...
1201 smpboot_restore_warm_reset_vector();
1204 * Allow the user to impress friends.
1206 Dprintk("Before bogomips.\n");
1207 for (cpu = 0; cpu < NR_CPUS; cpu++)
1208 if (cpu_isset(cpu, cpu_callout_map))
1209 bogosum += cpu_data[cpu].loops_per_jiffy;
1211 "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
1213 bogosum/(500000/HZ),
1214 (bogosum/(5000/HZ))%100);
1216 Dprintk("Before bogocount - setting activated=1.\n");
1219 printk(KERN_WARNING "WARNING: SMP operation may be unreliable with B stepping processors.\n");
1222 * Don't taint if we are running SMP kernel on a single non-MP
1225 if (tainted & TAINT_UNSAFE_SMP) {
1227 printk (KERN_INFO "WARNING: This combination of AMD processors is not suitable for SMP.\n");
1229 tainted &= ~TAINT_UNSAFE_SMP;
1232 Dprintk("Boot done.\n");
1235 * construct cpu_sibling_map[], so that we can tell sibling CPUs
1238 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1239 cpus_clear(cpu_sibling_map[cpu]);
1240 cpus_clear(cpu_core_map[cpu]);
1243 cpu_set(0, cpu_sibling_map[0]);
1244 cpu_set(0, cpu_core_map[0]);
1246 smpboot_setup_io_apic();
1248 setup_boot_APIC_clock();
1251 * Synchronize the TSC with the AP
1253 if (cpu_has_tsc && cpucount && cpu_khz)
1254 synchronize_tsc_bp();
1257 /* These are wrappers to interface to the new boot process. Someone
1258 who understands all this stuff should rewrite it properly. --RR 15/Jul/02 */
1259 void __init smp_prepare_cpus(unsigned int max_cpus)
1261 smp_commenced_mask = cpumask_of_cpu(0);
1262 cpu_callin_map = cpumask_of_cpu(0);
1264 smp_boot_cpus(max_cpus);
1267 void __devinit smp_prepare_boot_cpu(void)
1269 cpu_set(smp_processor_id(), cpu_online_map);
1270 cpu_set(smp_processor_id(), cpu_callout_map);
1271 cpu_set(smp_processor_id(), cpu_present_map);
1272 cpu_set(smp_processor_id(), cpu_possible_map);
1273 per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
1276 #ifdef CONFIG_HOTPLUG_CPU
1278 remove_siblinginfo(int cpu)
1282 for_each_cpu_mask(sibling, cpu_sibling_map[cpu])
1283 cpu_clear(cpu, cpu_sibling_map[sibling]);
1284 for_each_cpu_mask(sibling, cpu_core_map[cpu])
1285 cpu_clear(cpu, cpu_core_map[sibling]);
1286 cpus_clear(cpu_sibling_map[cpu]);
1287 cpus_clear(cpu_core_map[cpu]);
1288 phys_proc_id[cpu] = BAD_APICID;
1289 cpu_core_id[cpu] = BAD_APICID;
1292 int __cpu_disable(void)
1294 cpumask_t map = cpu_online_map;
1295 int cpu = smp_processor_id();
1298 * Perhaps use cpufreq to drop frequency, but that could go
1299 * into generic code.
1301 * We won't take down the boot processor on i386 due to some
1302 * interrupts only being able to be serviced by the BSP.
1303 * Especially so if we're not using an IOAPIC -zwane
1308 /* We enable the timer again on the exit path of the death loop */
1309 disable_APIC_timer();
1310 /* Allow any queued timer interrupts to get serviced */
1313 local_irq_disable();
1315 remove_siblinginfo(cpu);
1317 cpu_clear(cpu, map);
1319 /* It's now safe to remove this processor from the online map */
1320 cpu_clear(cpu, cpu_online_map);
1324 void __cpu_die(unsigned int cpu)
1326 /* We don't do anything here: idle task is faking death itself. */
1329 for (i = 0; i < 10; i++) {
1330 /* They ack this in play_dead by setting CPU_DEAD */
1331 if (per_cpu(cpu_state, cpu) == CPU_DEAD) {
1332 printk ("CPU %d is now offline\n", cpu);
1337 printk(KERN_ERR "CPU %u didn't die...\n", cpu);
1339 #else /* ... !CONFIG_HOTPLUG_CPU */
1340 int __cpu_disable(void)
1345 void __cpu_die(unsigned int cpu)
1347 /* We said "no" in __cpu_disable */
1350 #endif /* CONFIG_HOTPLUG_CPU */
1352 int __devinit __cpu_up(unsigned int cpu)
1354 /* In case one didn't come up */
1355 if (!cpu_isset(cpu, cpu_callin_map)) {
1356 printk(KERN_DEBUG "skipping cpu%d, didn't come online\n", cpu);
1362 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
1363 /* Unleash the CPU! */
1364 cpu_set(cpu, smp_commenced_mask);
1365 while (!cpu_isset(cpu, cpu_online_map))
1370 void __init smp_cpus_done(unsigned int max_cpus)
1372 #ifdef CONFIG_X86_IO_APIC
1373 setup_ioapic_dest();
1376 #ifndef CONFIG_HOTPLUG_CPU
1378 * Disable executability of the SMP trampoline:
1380 set_kernel_exec((unsigned long)trampoline_base, trampoline_exec);
1384 void __init smp_intr_init(void)
1387 * IRQ0 must be given a fixed assignment and initialized,
1388 * because it's used before the IO-APIC is set up.
1390 set_intr_gate(FIRST_DEVICE_VECTOR, interrupt[0]);
1393 * The reschedule interrupt is a CPU-to-CPU reschedule-helper
1394 * IPI, driven by wakeup.
1396 set_intr_gate(RESCHEDULE_VECTOR, reschedule_interrupt);
1398 /* IPI for invalidation */
1399 set_intr_gate(INVALIDATE_TLB_VECTOR, invalidate_interrupt);
1401 /* IPI for generic function call */
1402 set_intr_gate(CALL_FUNCTION_VECTOR, call_function_interrupt);