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>
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};
73 /* Core ID of each logical CPU */
74 int cpu_core_id[NR_CPUS] __read_mostly = {[0 ... NR_CPUS-1] = BAD_APICID};
76 /* Last level cache ID of each logical CPU */
77 int cpu_llc_id[NR_CPUS] __cpuinitdata = {[0 ... NR_CPUS-1] = BAD_APICID};
79 /* representing HT siblings of each logical CPU */
80 cpumask_t cpu_sibling_map[NR_CPUS] __read_mostly;
81 EXPORT_SYMBOL(cpu_sibling_map);
83 /* representing HT and core siblings of each logical CPU */
84 cpumask_t cpu_core_map[NR_CPUS] __read_mostly;
85 EXPORT_SYMBOL(cpu_core_map);
87 /* bitmap of online cpus */
88 cpumask_t cpu_online_map __read_mostly;
89 EXPORT_SYMBOL(cpu_online_map);
91 cpumask_t cpu_callin_map;
92 cpumask_t cpu_callout_map;
93 EXPORT_SYMBOL(cpu_callout_map);
94 cpumask_t cpu_possible_map;
95 EXPORT_SYMBOL(cpu_possible_map);
96 static cpumask_t smp_commenced_mask;
98 /* TSC's upper 32 bits can't be written in eariler CPU (before prescott), there
99 * is no way to resync one AP against BP. TBD: for prescott and above, we
100 * should use IA64's algorithm
102 static int __devinitdata tsc_sync_disabled;
104 /* Per CPU bogomips and other parameters */
105 struct cpuinfo_x86 cpu_data[NR_CPUS] __cacheline_aligned;
106 EXPORT_SYMBOL(cpu_data);
108 u8 x86_cpu_to_apicid[NR_CPUS] __read_mostly =
109 { [0 ... NR_CPUS-1] = 0xff };
110 EXPORT_SYMBOL(x86_cpu_to_apicid);
113 * Trampoline 80x86 program as an array.
116 extern unsigned char trampoline_data [];
117 extern unsigned char trampoline_end [];
118 static unsigned char *trampoline_base;
119 static int trampoline_exec;
121 static void map_cpu_to_logical_apicid(void);
123 /* State of each CPU. */
124 DEFINE_PER_CPU(int, cpu_state) = { 0 };
127 * Currently trivial. Write the real->protected mode
128 * bootstrap into the page concerned. The caller
129 * has made sure it's suitably aligned.
132 static unsigned long __devinit setup_trampoline(void)
134 memcpy(trampoline_base, trampoline_data, trampoline_end - trampoline_data);
135 return virt_to_phys(trampoline_base);
139 * We are called very early to get the low memory for the
140 * SMP bootup trampoline page.
142 void __init smp_alloc_memory(void)
144 trampoline_base = (void *) alloc_bootmem_low_pages(PAGE_SIZE);
146 * Has to be in very low memory so we can execute
149 if (__pa(trampoline_base) >= 0x9F000)
152 * Make the SMP trampoline executable:
154 trampoline_exec = set_kernel_exec((unsigned long)trampoline_base, 1);
158 * The bootstrap kernel entry code has set these up. Save them for
162 static void __devinit smp_store_cpu_info(int id)
164 struct cpuinfo_x86 *c = cpu_data + id;
170 * Mask B, Pentium, but not Pentium MMX
172 if (c->x86_vendor == X86_VENDOR_INTEL &&
174 c->x86_mask >= 1 && c->x86_mask <= 4 &&
177 * Remember we have B step Pentia with bugs
182 * Certain Athlons might work (for various values of 'work') in SMP
183 * but they are not certified as MP capable.
185 if ((c->x86_vendor == X86_VENDOR_AMD) && (c->x86 == 6)) {
187 /* Athlon 660/661 is valid. */
188 if ((c->x86_model==6) && ((c->x86_mask==0) || (c->x86_mask==1)))
191 /* Duron 670 is valid */
192 if ((c->x86_model==7) && (c->x86_mask==0))
196 * Athlon 662, Duron 671, and Athlon >model 7 have capability bit.
197 * It's worth noting that the A5 stepping (662) of some Athlon XP's
198 * have the MP bit set.
199 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for more.
201 if (((c->x86_model==6) && (c->x86_mask>=2)) ||
202 ((c->x86_model==7) && (c->x86_mask>=1)) ||
207 /* If we get here, it's not a certified SMP capable AMD system. */
208 add_taint(TAINT_UNSAFE_SMP);
216 * TSC synchronization.
218 * We first check whether all CPUs have their TSC's synchronized,
219 * then we print a warning if not, and always resync.
222 static atomic_t tsc_start_flag = ATOMIC_INIT(0);
223 static atomic_t tsc_count_start = ATOMIC_INIT(0);
224 static atomic_t tsc_count_stop = ATOMIC_INIT(0);
225 static unsigned long long tsc_values[NR_CPUS];
229 static void __init synchronize_tsc_bp (void)
232 unsigned long long t0;
233 unsigned long long sum, avg;
235 unsigned int one_usec;
238 printk(KERN_INFO "checking TSC synchronization across %u CPUs: ", num_booting_cpus());
240 /* convert from kcyc/sec to cyc/usec */
241 one_usec = cpu_khz / 1000;
243 atomic_set(&tsc_start_flag, 1);
247 * We loop a few times to get a primed instruction cache,
248 * then the last pass is more or less synchronized and
249 * the BP and APs set their cycle counters to zero all at
250 * once. This reduces the chance of having random offsets
251 * between the processors, and guarantees that the maximum
252 * delay between the cycle counters is never bigger than
253 * the latency of information-passing (cachelines) between
256 for (i = 0; i < NR_LOOPS; i++) {
258 * all APs synchronize but they loop on '== num_cpus'
260 while (atomic_read(&tsc_count_start) != num_booting_cpus()-1)
262 atomic_set(&tsc_count_stop, 0);
265 * this lets the APs save their current TSC:
267 atomic_inc(&tsc_count_start);
269 rdtscll(tsc_values[smp_processor_id()]);
271 * We clear the TSC in the last loop:
277 * Wait for all APs to leave the synchronization point:
279 while (atomic_read(&tsc_count_stop) != num_booting_cpus()-1)
281 atomic_set(&tsc_count_start, 0);
283 atomic_inc(&tsc_count_stop);
287 for (i = 0; i < NR_CPUS; i++) {
288 if (cpu_isset(i, cpu_callout_map)) {
294 do_div(avg, num_booting_cpus());
297 for (i = 0; i < NR_CPUS; i++) {
298 if (!cpu_isset(i, cpu_callout_map))
300 delta = tsc_values[i] - avg;
304 * We report bigger than 2 microseconds clock differences.
306 if (delta > 2*one_usec) {
313 do_div(realdelta, one_usec);
314 if (tsc_values[i] < avg)
315 realdelta = -realdelta;
318 printk(KERN_INFO "CPU#%d had %ld usecs TSC "
319 "skew, fixed it up.\n", i, realdelta);
328 static void __init synchronize_tsc_ap (void)
333 * Not every cpu is online at the time
334 * this gets called, so we first wait for the BP to
335 * finish SMP initialization:
337 while (!atomic_read(&tsc_start_flag))
340 for (i = 0; i < NR_LOOPS; i++) {
341 atomic_inc(&tsc_count_start);
342 while (atomic_read(&tsc_count_start) != num_booting_cpus())
345 rdtscll(tsc_values[smp_processor_id()]);
349 atomic_inc(&tsc_count_stop);
350 while (atomic_read(&tsc_count_stop) != num_booting_cpus())
356 extern void calibrate_delay(void);
358 static atomic_t init_deasserted;
360 static void __devinit smp_callin(void)
363 unsigned long timeout;
366 * If waken up by an INIT in an 82489DX configuration
367 * we may get here before an INIT-deassert IPI reaches
368 * our local APIC. We have to wait for the IPI or we'll
369 * lock up on an APIC access.
371 wait_for_init_deassert(&init_deasserted);
374 * (This works even if the APIC is not enabled.)
376 phys_id = GET_APIC_ID(apic_read(APIC_ID));
377 cpuid = smp_processor_id();
378 if (cpu_isset(cpuid, cpu_callin_map)) {
379 printk("huh, phys CPU#%d, CPU#%d already present??\n",
383 Dprintk("CPU#%d (phys ID: %d) waiting for CALLOUT\n", cpuid, phys_id);
386 * STARTUP IPIs are fragile beasts as they might sometimes
387 * trigger some glue motherboard logic. Complete APIC bus
388 * silence for 1 second, this overestimates the time the
389 * boot CPU is spending to send the up to 2 STARTUP IPIs
390 * by a factor of two. This should be enough.
394 * Waiting 2s total for startup (udelay is not yet working)
396 timeout = jiffies + 2*HZ;
397 while (time_before(jiffies, timeout)) {
399 * Has the boot CPU finished it's STARTUP sequence?
401 if (cpu_isset(cpuid, cpu_callout_map))
406 if (!time_before(jiffies, timeout)) {
407 printk("BUG: CPU%d started up but did not get a callout!\n",
413 * the boot CPU has finished the init stage and is spinning
414 * on callin_map until we finish. We are free to set up this
415 * CPU, first the APIC. (this is probably redundant on most
419 Dprintk("CALLIN, before setup_local_APIC().\n");
420 smp_callin_clear_local_apic();
422 map_cpu_to_logical_apicid();
428 Dprintk("Stack at about %p\n",&cpuid);
431 * Save our processor parameters
433 smp_store_cpu_info(cpuid);
435 disable_APIC_timer();
438 * Allow the master to continue.
440 cpu_set(cpuid, cpu_callin_map);
443 * Synchronize the TSC with the BP
445 if (cpu_has_tsc && cpu_khz && !tsc_sync_disabled)
446 synchronize_tsc_ap();
451 /* maps the cpu to the sched domain representing multi-core */
452 cpumask_t cpu_coregroup_map(int cpu)
454 struct cpuinfo_x86 *c = cpu_data + cpu;
456 * For perf, we return last level cache shared map.
457 * TBD: when power saving sched policy is added, we will return
458 * cpu_core_map when power saving policy is enabled
460 return c->llc_shared_map;
463 /* representing cpus for which sibling maps can be computed */
464 static cpumask_t cpu_sibling_setup_map;
467 set_cpu_sibling_map(int cpu)
470 struct cpuinfo_x86 *c = cpu_data;
472 cpu_set(cpu, cpu_sibling_setup_map);
474 if (smp_num_siblings > 1) {
475 for_each_cpu_mask(i, cpu_sibling_setup_map) {
476 if (phys_proc_id[cpu] == phys_proc_id[i] &&
477 cpu_core_id[cpu] == cpu_core_id[i]) {
478 cpu_set(i, cpu_sibling_map[cpu]);
479 cpu_set(cpu, cpu_sibling_map[i]);
480 cpu_set(i, cpu_core_map[cpu]);
481 cpu_set(cpu, cpu_core_map[i]);
482 cpu_set(i, c[cpu].llc_shared_map);
483 cpu_set(cpu, c[i].llc_shared_map);
487 cpu_set(cpu, cpu_sibling_map[cpu]);
490 cpu_set(cpu, c[cpu].llc_shared_map);
492 if (current_cpu_data.x86_max_cores == 1) {
493 cpu_core_map[cpu] = cpu_sibling_map[cpu];
494 c[cpu].booted_cores = 1;
498 for_each_cpu_mask(i, cpu_sibling_setup_map) {
499 if (cpu_llc_id[cpu] != BAD_APICID &&
500 cpu_llc_id[cpu] == cpu_llc_id[i]) {
501 cpu_set(i, c[cpu].llc_shared_map);
502 cpu_set(cpu, c[i].llc_shared_map);
504 if (phys_proc_id[cpu] == phys_proc_id[i]) {
505 cpu_set(i, cpu_core_map[cpu]);
506 cpu_set(cpu, cpu_core_map[i]);
508 * Does this new cpu bringup a new core?
510 if (cpus_weight(cpu_sibling_map[cpu]) == 1) {
512 * for each core in package, increment
513 * the booted_cores for this new cpu
515 if (first_cpu(cpu_sibling_map[i]) == i)
516 c[cpu].booted_cores++;
518 * increment the core count for all
519 * the other cpus in this package
523 } else if (i != cpu && !c[cpu].booted_cores)
524 c[cpu].booted_cores = c[i].booted_cores;
530 * Activate a secondary processor.
532 static void __devinit start_secondary(void *unused)
535 * Dont put anything before smp_callin(), SMP
536 * booting is too fragile that we want to limit the
537 * things done here to the most necessary things.
542 while (!cpu_isset(smp_processor_id(), smp_commenced_mask))
544 setup_secondary_APIC_clock();
545 if (nmi_watchdog == NMI_IO_APIC) {
546 disable_8259A_irq(0);
547 enable_NMI_through_LVT0(NULL);
552 * low-memory mappings have been cleared, flush them from
553 * the local TLBs too.
557 /* This must be done before setting cpu_online_map */
558 set_cpu_sibling_map(raw_smp_processor_id());
562 * We need to hold call_lock, so there is no inconsistency
563 * between the time smp_call_function() determines number of
564 * IPI receipients, and the time when the determination is made
565 * for which cpus receive the IPI. Holding this
566 * lock helps us to not include this cpu in a currently in progress
567 * smp_call_function().
569 lock_ipi_call_lock();
570 cpu_set(smp_processor_id(), cpu_online_map);
571 unlock_ipi_call_lock();
572 per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
574 /* We can take interrupts now: we're officially "up". */
582 * Everything has been set up for the secondary
583 * CPUs - they just need to reload everything
584 * from the task structure
585 * This function must not return.
587 void __devinit initialize_secondary(void)
590 * We don't actually need to load the full TSS,
591 * basically just the stack pointer and the eip.
598 :"r" (current->thread.esp),"r" (current->thread.eip));
608 /* which logical CPUs are on which nodes */
609 cpumask_t node_2_cpu_mask[MAX_NUMNODES] __read_mostly =
610 { [0 ... MAX_NUMNODES-1] = CPU_MASK_NONE };
611 /* which node each logical CPU is on */
612 int cpu_2_node[NR_CPUS] __read_mostly = { [0 ... NR_CPUS-1] = 0 };
613 EXPORT_SYMBOL(cpu_2_node);
615 /* set up a mapping between cpu and node. */
616 static inline void map_cpu_to_node(int cpu, int node)
618 printk("Mapping cpu %d to node %d\n", cpu, node);
619 cpu_set(cpu, node_2_cpu_mask[node]);
620 cpu_2_node[cpu] = node;
623 /* undo a mapping between cpu and node. */
624 static inline void unmap_cpu_to_node(int cpu)
628 printk("Unmapping cpu %d from all nodes\n", cpu);
629 for (node = 0; node < MAX_NUMNODES; node ++)
630 cpu_clear(cpu, node_2_cpu_mask[node]);
633 #else /* !CONFIG_NUMA */
635 #define map_cpu_to_node(cpu, node) ({})
636 #define unmap_cpu_to_node(cpu) ({})
638 #endif /* CONFIG_NUMA */
640 u8 cpu_2_logical_apicid[NR_CPUS] __read_mostly = { [0 ... NR_CPUS-1] = BAD_APICID };
642 static void map_cpu_to_logical_apicid(void)
644 int cpu = smp_processor_id();
645 int apicid = logical_smp_processor_id();
647 cpu_2_logical_apicid[cpu] = apicid;
648 map_cpu_to_node(cpu, apicid_to_node(apicid));
651 static void unmap_cpu_to_logical_apicid(int cpu)
653 cpu_2_logical_apicid[cpu] = BAD_APICID;
654 unmap_cpu_to_node(cpu);
658 static inline void __inquire_remote_apic(int apicid)
660 int i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
661 char *names[] = { "ID", "VERSION", "SPIV" };
664 printk("Inquiring remote APIC #%d...\n", apicid);
666 for (i = 0; i < ARRAY_SIZE(regs); i++) {
667 printk("... APIC #%d %s: ", apicid, names[i]);
672 apic_wait_icr_idle();
674 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(apicid));
675 apic_write_around(APIC_ICR, APIC_DM_REMRD | regs[i]);
680 status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
681 } while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
684 case APIC_ICR_RR_VALID:
685 status = apic_read(APIC_RRR);
686 printk("%08x\n", status);
695 #ifdef WAKE_SECONDARY_VIA_NMI
697 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
698 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
699 * won't ... remember to clear down the APIC, etc later.
702 wakeup_secondary_cpu(int logical_apicid, unsigned long start_eip)
704 unsigned long send_status = 0, accept_status = 0;
708 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(logical_apicid));
710 /* Boot on the stack */
711 /* Kick the second */
712 apic_write_around(APIC_ICR, APIC_DM_NMI | APIC_DEST_LOGICAL);
714 Dprintk("Waiting for send to finish...\n");
719 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
720 } while (send_status && (timeout++ < 1000));
723 * Give the other CPU some time to accept the IPI.
727 * Due to the Pentium erratum 3AP.
729 maxlvt = get_maxlvt();
731 apic_read_around(APIC_SPIV);
732 apic_write(APIC_ESR, 0);
734 accept_status = (apic_read(APIC_ESR) & 0xEF);
735 Dprintk("NMI sent.\n");
738 printk("APIC never delivered???\n");
740 printk("APIC delivery error (%lx).\n", accept_status);
742 return (send_status | accept_status);
744 #endif /* WAKE_SECONDARY_VIA_NMI */
746 #ifdef WAKE_SECONDARY_VIA_INIT
748 wakeup_secondary_cpu(int phys_apicid, unsigned long start_eip)
750 unsigned long send_status = 0, accept_status = 0;
751 int maxlvt, timeout, num_starts, j;
754 * Be paranoid about clearing APIC errors.
756 if (APIC_INTEGRATED(apic_version[phys_apicid])) {
757 apic_read_around(APIC_SPIV);
758 apic_write(APIC_ESR, 0);
762 Dprintk("Asserting INIT.\n");
765 * Turn INIT on target chip
767 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
772 apic_write_around(APIC_ICR, APIC_INT_LEVELTRIG | APIC_INT_ASSERT
775 Dprintk("Waiting for send to finish...\n");
780 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
781 } while (send_status && (timeout++ < 1000));
785 Dprintk("Deasserting INIT.\n");
788 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
791 apic_write_around(APIC_ICR, APIC_INT_LEVELTRIG | APIC_DM_INIT);
793 Dprintk("Waiting for send to finish...\n");
798 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
799 } while (send_status && (timeout++ < 1000));
801 atomic_set(&init_deasserted, 1);
804 * Should we send STARTUP IPIs ?
806 * Determine this based on the APIC version.
807 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
809 if (APIC_INTEGRATED(apic_version[phys_apicid]))
815 * Run STARTUP IPI loop.
817 Dprintk("#startup loops: %d.\n", num_starts);
819 maxlvt = get_maxlvt();
821 for (j = 1; j <= num_starts; j++) {
822 Dprintk("Sending STARTUP #%d.\n",j);
823 apic_read_around(APIC_SPIV);
824 apic_write(APIC_ESR, 0);
826 Dprintk("After apic_write.\n");
833 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
835 /* Boot on the stack */
836 /* Kick the second */
837 apic_write_around(APIC_ICR, APIC_DM_STARTUP
838 | (start_eip >> 12));
841 * Give the other CPU some time to accept the IPI.
845 Dprintk("Startup point 1.\n");
847 Dprintk("Waiting for send to finish...\n");
852 send_status = apic_read(APIC_ICR) & APIC_ICR_BUSY;
853 } while (send_status && (timeout++ < 1000));
856 * Give the other CPU some time to accept the IPI.
860 * Due to the Pentium erratum 3AP.
863 apic_read_around(APIC_SPIV);
864 apic_write(APIC_ESR, 0);
866 accept_status = (apic_read(APIC_ESR) & 0xEF);
867 if (send_status || accept_status)
870 Dprintk("After Startup.\n");
873 printk("APIC never delivered???\n");
875 printk("APIC delivery error (%lx).\n", accept_status);
877 return (send_status | accept_status);
879 #endif /* WAKE_SECONDARY_VIA_INIT */
881 extern cpumask_t cpu_initialized;
882 static inline int alloc_cpu_id(void)
886 cpus_complement(tmp_map, cpu_present_map);
887 cpu = first_cpu(tmp_map);
893 #ifdef CONFIG_HOTPLUG_CPU
894 static struct task_struct * __devinitdata cpu_idle_tasks[NR_CPUS];
895 static inline struct task_struct * alloc_idle_task(int cpu)
897 struct task_struct *idle;
899 if ((idle = cpu_idle_tasks[cpu]) != NULL) {
900 /* initialize thread_struct. we really want to avoid destroy
903 idle->thread.esp = (unsigned long)task_pt_regs(idle);
904 init_idle(idle, cpu);
907 idle = fork_idle(cpu);
910 cpu_idle_tasks[cpu] = idle;
914 #define alloc_idle_task(cpu) fork_idle(cpu)
917 static int __devinit do_boot_cpu(int apicid, int cpu)
919 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
920 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
921 * Returns zero if CPU booted OK, else error code from wakeup_secondary_cpu.
924 struct task_struct *idle;
925 unsigned long boot_error;
927 unsigned long start_eip;
928 unsigned short nmi_high = 0, nmi_low = 0;
931 alternatives_smp_switch(1);
934 * We can't use kernel_thread since we must avoid to
935 * reschedule the child.
937 idle = alloc_idle_task(cpu);
939 panic("failed fork for CPU %d", cpu);
940 idle->thread.eip = (unsigned long) start_secondary;
941 /* start_eip had better be page-aligned! */
942 start_eip = setup_trampoline();
944 /* So we see what's up */
945 printk("Booting processor %d/%d eip %lx\n", cpu, apicid, start_eip);
946 /* Stack for startup_32 can be just as for start_secondary onwards */
947 stack_start.esp = (void *) idle->thread.esp;
952 * This grunge runs the startup process for
953 * the targeted processor.
956 atomic_set(&init_deasserted, 0);
958 Dprintk("Setting warm reset code and vector.\n");
960 store_NMI_vector(&nmi_high, &nmi_low);
962 smpboot_setup_warm_reset_vector(start_eip);
965 * Starting actual IPI sequence...
967 boot_error = wakeup_secondary_cpu(apicid, start_eip);
971 * allow APs to start initializing.
973 Dprintk("Before Callout %d.\n", cpu);
974 cpu_set(cpu, cpu_callout_map);
975 Dprintk("After Callout %d.\n", cpu);
978 * Wait 5s total for a response
980 for (timeout = 0; timeout < 50000; timeout++) {
981 if (cpu_isset(cpu, cpu_callin_map))
982 break; /* It has booted */
986 if (cpu_isset(cpu, cpu_callin_map)) {
987 /* number CPUs logically, starting from 1 (BSP is 0) */
989 printk("CPU%d: ", cpu);
990 print_cpu_info(&cpu_data[cpu]);
991 Dprintk("CPU has booted.\n");
994 if (*((volatile unsigned char *)trampoline_base)
996 /* trampoline started but...? */
997 printk("Stuck ??\n");
999 /* trampoline code not run */
1000 printk("Not responding.\n");
1001 inquire_remote_apic(apicid);
1006 /* Try to put things back the way they were before ... */
1007 unmap_cpu_to_logical_apicid(cpu);
1008 cpu_clear(cpu, cpu_callout_map); /* was set here (do_boot_cpu()) */
1009 cpu_clear(cpu, cpu_initialized); /* was set by cpu_init() */
1012 x86_cpu_to_apicid[cpu] = apicid;
1013 cpu_set(cpu, cpu_present_map);
1016 /* mark "stuck" area as not stuck */
1017 *((volatile unsigned long *)trampoline_base) = 0;
1022 #ifdef CONFIG_HOTPLUG_CPU
1023 void cpu_exit_clear(void)
1025 int cpu = raw_smp_processor_id();
1033 cpu_clear(cpu, cpu_callout_map);
1034 cpu_clear(cpu, cpu_callin_map);
1036 cpu_clear(cpu, smp_commenced_mask);
1037 unmap_cpu_to_logical_apicid(cpu);
1040 struct warm_boot_cpu_info {
1041 struct completion *complete;
1046 static void __cpuinit do_warm_boot_cpu(void *p)
1048 struct warm_boot_cpu_info *info = p;
1049 do_boot_cpu(info->apicid, info->cpu);
1050 complete(info->complete);
1053 static int __cpuinit __smp_prepare_cpu(int cpu)
1055 DECLARE_COMPLETION(done);
1056 struct warm_boot_cpu_info info;
1057 struct work_struct task;
1060 apicid = x86_cpu_to_apicid[cpu];
1061 if (apicid == BAD_APICID) {
1066 info.complete = &done;
1067 info.apicid = apicid;
1069 INIT_WORK(&task, do_warm_boot_cpu, &info);
1071 tsc_sync_disabled = 1;
1073 /* init low mem mapping */
1074 clone_pgd_range(swapper_pg_dir, swapper_pg_dir + USER_PGD_PTRS,
1077 schedule_work(&task);
1078 wait_for_completion(&done);
1080 tsc_sync_disabled = 0;
1088 static void smp_tune_scheduling (void)
1090 unsigned long cachesize; /* kB */
1091 unsigned long bandwidth = 350; /* MB/s */
1093 * Rough estimation for SMP scheduling, this is the number of
1094 * cycles it takes for a fully memory-limited process to flush
1095 * the SMP-local cache.
1097 * (For a P5 this pretty much means we will choose another idle
1098 * CPU almost always at wakeup time (this is due to the small
1099 * L1 cache), on PIIs it's around 50-100 usecs, depending on
1105 * this basically disables processor-affinity
1106 * scheduling on SMP without a TSC.
1110 cachesize = boot_cpu_data.x86_cache_size;
1111 if (cachesize == -1) {
1112 cachesize = 16; /* Pentiums, 2x8kB cache */
1115 max_cache_size = cachesize * 1024;
1120 * Cycle through the processors sending APIC IPIs to boot each.
1123 static int boot_cpu_logical_apicid;
1124 /* Where the IO area was mapped on multiquad, always 0 otherwise */
1126 #ifdef CONFIG_X86_NUMAQ
1127 EXPORT_SYMBOL(xquad_portio);
1130 static void __init smp_boot_cpus(unsigned int max_cpus)
1132 int apicid, cpu, bit, kicked;
1133 unsigned long bogosum = 0;
1136 * Setup boot CPU information
1138 smp_store_cpu_info(0); /* Final full version of the data */
1139 printk("CPU%d: ", 0);
1140 print_cpu_info(&cpu_data[0]);
1142 boot_cpu_physical_apicid = GET_APIC_ID(apic_read(APIC_ID));
1143 boot_cpu_logical_apicid = logical_smp_processor_id();
1144 x86_cpu_to_apicid[0] = boot_cpu_physical_apicid;
1146 current_thread_info()->cpu = 0;
1147 smp_tune_scheduling();
1149 set_cpu_sibling_map(0);
1152 * If we couldn't find an SMP configuration at boot time,
1153 * get out of here now!
1155 if (!smp_found_config && !acpi_lapic) {
1156 printk(KERN_NOTICE "SMP motherboard not detected.\n");
1157 smpboot_clear_io_apic_irqs();
1158 phys_cpu_present_map = physid_mask_of_physid(0);
1159 if (APIC_init_uniprocessor())
1160 printk(KERN_NOTICE "Local APIC not detected."
1161 " Using dummy APIC emulation.\n");
1162 map_cpu_to_logical_apicid();
1163 cpu_set(0, cpu_sibling_map[0]);
1164 cpu_set(0, cpu_core_map[0]);
1169 * Should not be necessary because the MP table should list the boot
1170 * CPU too, but we do it for the sake of robustness anyway.
1171 * Makes no sense to do this check in clustered apic mode, so skip it
1173 if (!check_phys_apicid_present(boot_cpu_physical_apicid)) {
1174 printk("weird, boot CPU (#%d) not listed by the BIOS.\n",
1175 boot_cpu_physical_apicid);
1176 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1180 * If we couldn't find a local APIC, then get out of here now!
1182 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid]) && !cpu_has_apic) {
1183 printk(KERN_ERR "BIOS bug, local APIC #%d not detected!...\n",
1184 boot_cpu_physical_apicid);
1185 printk(KERN_ERR "... forcing use of dummy APIC emulation. (tell your hw vendor)\n");
1186 smpboot_clear_io_apic_irqs();
1187 phys_cpu_present_map = physid_mask_of_physid(0);
1188 cpu_set(0, cpu_sibling_map[0]);
1189 cpu_set(0, cpu_core_map[0]);
1193 verify_local_APIC();
1196 * If SMP should be disabled, then really disable it!
1199 smp_found_config = 0;
1200 printk(KERN_INFO "SMP mode deactivated, forcing use of dummy APIC emulation.\n");
1201 smpboot_clear_io_apic_irqs();
1202 phys_cpu_present_map = physid_mask_of_physid(0);
1203 cpu_set(0, cpu_sibling_map[0]);
1204 cpu_set(0, cpu_core_map[0]);
1210 map_cpu_to_logical_apicid();
1213 setup_portio_remap();
1216 * Scan the CPU present map and fire up the other CPUs via do_boot_cpu
1218 * In clustered apic mode, phys_cpu_present_map is a constructed thus:
1219 * bits 0-3 are quad0, 4-7 are quad1, etc. A perverse twist on the
1220 * clustered apic ID.
1222 Dprintk("CPU present map: %lx\n", physids_coerce(phys_cpu_present_map));
1225 for (bit = 0; kicked < NR_CPUS && bit < MAX_APICS; bit++) {
1226 apicid = cpu_present_to_apicid(bit);
1228 * Don't even attempt to start the boot CPU!
1230 if ((apicid == boot_cpu_apicid) || (apicid == BAD_APICID))
1233 if (!check_apicid_present(bit))
1235 if (max_cpus <= cpucount+1)
1238 if (((cpu = alloc_cpu_id()) <= 0) || do_boot_cpu(apicid, cpu))
1239 printk("CPU #%d not responding - cannot use it.\n",
1246 * Cleanup possible dangling ends...
1248 smpboot_restore_warm_reset_vector();
1251 * Allow the user to impress friends.
1253 Dprintk("Before bogomips.\n");
1254 for (cpu = 0; cpu < NR_CPUS; cpu++)
1255 if (cpu_isset(cpu, cpu_callout_map))
1256 bogosum += cpu_data[cpu].loops_per_jiffy;
1258 "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
1260 bogosum/(500000/HZ),
1261 (bogosum/(5000/HZ))%100);
1263 Dprintk("Before bogocount - setting activated=1.\n");
1266 printk(KERN_WARNING "WARNING: SMP operation may be unreliable with B stepping processors.\n");
1269 * Don't taint if we are running SMP kernel on a single non-MP
1272 if (tainted & TAINT_UNSAFE_SMP) {
1274 printk (KERN_INFO "WARNING: This combination of AMD processors is not suitable for SMP.\n");
1276 tainted &= ~TAINT_UNSAFE_SMP;
1279 Dprintk("Boot done.\n");
1282 * construct cpu_sibling_map[], so that we can tell sibling CPUs
1285 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1286 cpus_clear(cpu_sibling_map[cpu]);
1287 cpus_clear(cpu_core_map[cpu]);
1290 cpu_set(0, cpu_sibling_map[0]);
1291 cpu_set(0, cpu_core_map[0]);
1293 smpboot_setup_io_apic();
1295 setup_boot_APIC_clock();
1298 * Synchronize the TSC with the AP
1300 if (cpu_has_tsc && cpucount && cpu_khz)
1301 synchronize_tsc_bp();
1304 /* These are wrappers to interface to the new boot process. Someone
1305 who understands all this stuff should rewrite it properly. --RR 15/Jul/02 */
1306 void __init smp_prepare_cpus(unsigned int max_cpus)
1308 smp_commenced_mask = cpumask_of_cpu(0);
1309 cpu_callin_map = cpumask_of_cpu(0);
1311 smp_boot_cpus(max_cpus);
1314 void __devinit smp_prepare_boot_cpu(void)
1316 cpu_set(smp_processor_id(), cpu_online_map);
1317 cpu_set(smp_processor_id(), cpu_callout_map);
1318 cpu_set(smp_processor_id(), cpu_present_map);
1319 cpu_set(smp_processor_id(), cpu_possible_map);
1320 per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
1323 #ifdef CONFIG_HOTPLUG_CPU
1325 remove_siblinginfo(int cpu)
1328 struct cpuinfo_x86 *c = cpu_data;
1330 for_each_cpu_mask(sibling, cpu_core_map[cpu]) {
1331 cpu_clear(cpu, cpu_core_map[sibling]);
1333 * last thread sibling in this cpu core going down
1335 if (cpus_weight(cpu_sibling_map[cpu]) == 1)
1336 c[sibling].booted_cores--;
1339 for_each_cpu_mask(sibling, cpu_sibling_map[cpu])
1340 cpu_clear(cpu, cpu_sibling_map[sibling]);
1341 cpus_clear(cpu_sibling_map[cpu]);
1342 cpus_clear(cpu_core_map[cpu]);
1343 phys_proc_id[cpu] = BAD_APICID;
1344 cpu_core_id[cpu] = BAD_APICID;
1345 cpu_clear(cpu, cpu_sibling_setup_map);
1348 int __cpu_disable(void)
1350 cpumask_t map = cpu_online_map;
1351 int cpu = smp_processor_id();
1354 * Perhaps use cpufreq to drop frequency, but that could go
1355 * into generic code.
1357 * We won't take down the boot processor on i386 due to some
1358 * interrupts only being able to be serviced by the BSP.
1359 * Especially so if we're not using an IOAPIC -zwane
1365 /* Allow any queued timer interrupts to get serviced */
1368 local_irq_disable();
1370 remove_siblinginfo(cpu);
1372 cpu_clear(cpu, map);
1374 /* It's now safe to remove this processor from the online map */
1375 cpu_clear(cpu, cpu_online_map);
1379 void __cpu_die(unsigned int cpu)
1381 /* We don't do anything here: idle task is faking death itself. */
1384 for (i = 0; i < 10; i++) {
1385 /* They ack this in play_dead by setting CPU_DEAD */
1386 if (per_cpu(cpu_state, cpu) == CPU_DEAD) {
1387 printk ("CPU %d is now offline\n", cpu);
1388 if (1 == num_online_cpus())
1389 alternatives_smp_switch(0);
1394 printk(KERN_ERR "CPU %u didn't die...\n", cpu);
1396 #else /* ... !CONFIG_HOTPLUG_CPU */
1397 int __cpu_disable(void)
1402 void __cpu_die(unsigned int cpu)
1404 /* We said "no" in __cpu_disable */
1407 #endif /* CONFIG_HOTPLUG_CPU */
1409 int __devinit __cpu_up(unsigned int cpu)
1411 #ifdef CONFIG_HOTPLUG_CPU
1415 * We do warm boot only on cpus that had booted earlier
1416 * Otherwise cold boot is all handled from smp_boot_cpus().
1417 * cpu_callin_map is set during AP kickstart process. Its reset
1418 * when a cpu is taken offline from cpu_exit_clear().
1420 if (!cpu_isset(cpu, cpu_callin_map))
1421 ret = __smp_prepare_cpu(cpu);
1427 /* In case one didn't come up */
1428 if (!cpu_isset(cpu, cpu_callin_map)) {
1429 printk(KERN_DEBUG "skipping cpu%d, didn't come online\n", cpu);
1435 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
1436 /* Unleash the CPU! */
1437 cpu_set(cpu, smp_commenced_mask);
1438 while (!cpu_isset(cpu, cpu_online_map))
1443 void __init smp_cpus_done(unsigned int max_cpus)
1445 #ifdef CONFIG_X86_IO_APIC
1446 setup_ioapic_dest();
1449 #ifndef CONFIG_HOTPLUG_CPU
1451 * Disable executability of the SMP trampoline:
1453 set_kernel_exec((unsigned long)trampoline_base, trampoline_exec);
1457 void __init smp_intr_init(void)
1460 * IRQ0 must be given a fixed assignment and initialized,
1461 * because it's used before the IO-APIC is set up.
1463 set_intr_gate(FIRST_DEVICE_VECTOR, interrupt[0]);
1466 * The reschedule interrupt is a CPU-to-CPU reschedule-helper
1467 * IPI, driven by wakeup.
1469 set_intr_gate(RESCHEDULE_VECTOR, reschedule_interrupt);
1471 /* IPI for invalidation */
1472 set_intr_gate(INVALIDATE_TLB_VECTOR, invalidate_interrupt);
1474 /* IPI for generic function call */
1475 set_intr_gate(CALL_FUNCTION_VECTOR, call_function_interrupt);