2 * x86 SMP booting functions
4 * (c) 1995 Alan Cox, Building #3 <alan@redhat.com>
5 * (c) 1998, 1999, 2000 Ingo Molnar <mingo@redhat.com>
7 * Much of the core SMP work is based on previous work by Thomas Radke, to
8 * whom a great many thanks are extended.
10 * Thanks to Intel for making available several different Pentium,
11 * Pentium Pro and Pentium-II/Xeon MP machines.
12 * Original development of Linux SMP code supported by Caldera.
14 * This code is released under the GNU General Public License version 2 or
18 * Felix Koop : NR_CPUS used properly
19 * Jose Renau : Handle single CPU case.
20 * Alan Cox : By repeated request 8) - Total BogoMIPS report.
21 * Greg Wright : Fix for kernel stacks panic.
22 * Erich Boleyn : MP v1.4 and additional changes.
23 * Matthias Sattler : Changes for 2.1 kernel map.
24 * Michel Lespinasse : Changes for 2.1 kernel map.
25 * Michael Chastain : Change trampoline.S to gnu as.
26 * Alan Cox : Dumb bug: 'B' step PPro's are fine
27 * Ingo Molnar : Added APIC timers, based on code
29 * Ingo Molnar : various cleanups and rewrites
30 * Tigran Aivazian : fixed "0.00 in /proc/uptime on SMP" bug.
31 * Maciej W. Rozycki : Bits for genuine 82489DX APICs
32 * Martin J. Bligh : Added support for multi-quad systems
33 * Dave Jones : Report invalid combinations of Athlon CPUs.
34 * Rusty Russell : Hacked into shape for new "hotplug" boot process. */
36 #include <linux/module.h>
37 #include <linux/init.h>
38 #include <linux/kernel.h>
41 #include <linux/sched.h>
42 #include <linux/kernel_stat.h>
43 #include <linux/smp_lock.h>
44 #include <linux/bootmem.h>
45 #include <linux/notifier.h>
46 #include <linux/cpu.h>
47 #include <linux/percpu.h>
48 #include <linux/nmi.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>
63 /* Set if we find a B stepping CPU */
64 static int __devinitdata smp_b_stepping;
66 /* Number of siblings per CPU package */
67 int smp_num_siblings = 1;
68 EXPORT_SYMBOL(smp_num_siblings);
70 /* Last level cache ID of each logical CPU */
71 int cpu_llc_id[NR_CPUS] __cpuinitdata = {[0 ... NR_CPUS-1] = BAD_APICID};
73 /* representing HT siblings of each logical CPU */
74 cpumask_t cpu_sibling_map[NR_CPUS] __read_mostly;
75 EXPORT_SYMBOL(cpu_sibling_map);
77 /* representing HT and core siblings of each logical CPU */
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 cpumask_t cpu_possible_map;
89 EXPORT_SYMBOL(cpu_possible_map);
90 static cpumask_t smp_commenced_mask;
92 /* Per CPU bogomips and other parameters */
93 struct cpuinfo_x86 cpu_data[NR_CPUS] __cacheline_aligned;
94 EXPORT_SYMBOL(cpu_data);
96 u8 x86_cpu_to_apicid[NR_CPUS] __read_mostly =
97 { [0 ... NR_CPUS-1] = 0xff };
98 EXPORT_SYMBOL(x86_cpu_to_apicid);
100 u8 apicid_2_node[MAX_APICID];
102 DEFINE_PER_CPU(unsigned long, this_cpu_off);
103 EXPORT_PER_CPU_SYMBOL(this_cpu_off);
106 * Trampoline 80x86 program as an array.
109 extern unsigned char trampoline_data [];
110 extern unsigned char trampoline_end [];
111 static unsigned char *trampoline_base;
112 static int trampoline_exec;
114 static void map_cpu_to_logical_apicid(void);
116 /* State of each CPU. */
117 DEFINE_PER_CPU(int, cpu_state) = { 0 };
120 * Currently trivial. Write the real->protected mode
121 * bootstrap into the page concerned. The caller
122 * has made sure it's suitably aligned.
125 static unsigned long __devinit setup_trampoline(void)
127 memcpy(trampoline_base, trampoline_data, trampoline_end - trampoline_data);
128 return virt_to_phys(trampoline_base);
132 * We are called very early to get the low memory for the
133 * SMP bootup trampoline page.
135 void __init smp_alloc_memory(void)
137 trampoline_base = (void *) alloc_bootmem_low_pages(PAGE_SIZE);
139 * Has to be in very low memory so we can execute
142 if (__pa(trampoline_base) >= 0x9F000)
145 * Make the SMP trampoline executable:
147 trampoline_exec = set_kernel_exec((unsigned long)trampoline_base, 1);
151 * The bootstrap kernel entry code has set these up. Save them for
155 static void __cpuinit smp_store_cpu_info(int id)
157 struct cpuinfo_x86 *c = cpu_data + id;
161 identify_secondary_cpu(c);
163 * Mask B, Pentium, but not Pentium MMX
165 if (c->x86_vendor == X86_VENDOR_INTEL &&
167 c->x86_mask >= 1 && c->x86_mask <= 4 &&
170 * Remember we have B step Pentia with bugs
175 * Certain Athlons might work (for various values of 'work') in SMP
176 * but they are not certified as MP capable.
178 if ((c->x86_vendor == X86_VENDOR_AMD) && (c->x86 == 6)) {
180 if (num_possible_cpus() == 1)
183 /* Athlon 660/661 is valid. */
184 if ((c->x86_model==6) && ((c->x86_mask==0) || (c->x86_mask==1)))
187 /* Duron 670 is valid */
188 if ((c->x86_model==7) && (c->x86_mask==0))
192 * Athlon 662, Duron 671, and Athlon >model 7 have capability bit.
193 * It's worth noting that the A5 stepping (662) of some Athlon XP's
194 * have the MP bit set.
195 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for more.
197 if (((c->x86_model==6) && (c->x86_mask>=2)) ||
198 ((c->x86_model==7) && (c->x86_mask>=1)) ||
203 /* If we get here, it's not a certified SMP capable AMD system. */
204 add_taint(TAINT_UNSAFE_SMP);
211 extern void calibrate_delay(void);
213 static atomic_t init_deasserted;
215 static void __cpuinit smp_callin(void)
218 unsigned long timeout;
221 * If waken up by an INIT in an 82489DX configuration
222 * we may get here before an INIT-deassert IPI reaches
223 * our local APIC. We have to wait for the IPI or we'll
224 * lock up on an APIC access.
226 wait_for_init_deassert(&init_deasserted);
229 * (This works even if the APIC is not enabled.)
231 phys_id = GET_APIC_ID(apic_read(APIC_ID));
232 cpuid = smp_processor_id();
233 if (cpu_isset(cpuid, cpu_callin_map)) {
234 printk("huh, phys CPU#%d, CPU#%d already present??\n",
238 Dprintk("CPU#%d (phys ID: %d) waiting for CALLOUT\n", cpuid, phys_id);
241 * STARTUP IPIs are fragile beasts as they might sometimes
242 * trigger some glue motherboard logic. Complete APIC bus
243 * silence for 1 second, this overestimates the time the
244 * boot CPU is spending to send the up to 2 STARTUP IPIs
245 * by a factor of two. This should be enough.
249 * Waiting 2s total for startup (udelay is not yet working)
251 timeout = jiffies + 2*HZ;
252 while (time_before(jiffies, timeout)) {
254 * Has the boot CPU finished it's STARTUP sequence?
256 if (cpu_isset(cpuid, cpu_callout_map))
261 if (!time_before(jiffies, timeout)) {
262 printk("BUG: CPU%d started up but did not get a callout!\n",
268 * the boot CPU has finished the init stage and is spinning
269 * on callin_map until we finish. We are free to set up this
270 * CPU, first the APIC. (this is probably redundant on most
274 Dprintk("CALLIN, before setup_local_APIC().\n");
275 smp_callin_clear_local_apic();
277 map_cpu_to_logical_apicid();
283 Dprintk("Stack at about %p\n",&cpuid);
286 * Save our processor parameters
288 smp_store_cpu_info(cpuid);
291 * Allow the master to continue.
293 cpu_set(cpuid, cpu_callin_map);
298 /* maps the cpu to the sched domain representing multi-core */
299 cpumask_t cpu_coregroup_map(int cpu)
301 struct cpuinfo_x86 *c = cpu_data + cpu;
303 * For perf, we return last level cache shared map.
304 * And for power savings, we return cpu_core_map
306 if (sched_mc_power_savings || sched_smt_power_savings)
307 return cpu_core_map[cpu];
309 return c->llc_shared_map;
312 /* representing cpus for which sibling maps can be computed */
313 static cpumask_t cpu_sibling_setup_map;
316 set_cpu_sibling_map(int cpu)
319 struct cpuinfo_x86 *c = cpu_data;
321 cpu_set(cpu, cpu_sibling_setup_map);
323 if (smp_num_siblings > 1) {
324 for_each_cpu_mask(i, cpu_sibling_setup_map) {
325 if (c[cpu].phys_proc_id == c[i].phys_proc_id &&
326 c[cpu].cpu_core_id == c[i].cpu_core_id) {
327 cpu_set(i, cpu_sibling_map[cpu]);
328 cpu_set(cpu, cpu_sibling_map[i]);
329 cpu_set(i, cpu_core_map[cpu]);
330 cpu_set(cpu, cpu_core_map[i]);
331 cpu_set(i, c[cpu].llc_shared_map);
332 cpu_set(cpu, c[i].llc_shared_map);
336 cpu_set(cpu, cpu_sibling_map[cpu]);
339 cpu_set(cpu, c[cpu].llc_shared_map);
341 if (current_cpu_data.x86_max_cores == 1) {
342 cpu_core_map[cpu] = cpu_sibling_map[cpu];
343 c[cpu].booted_cores = 1;
347 for_each_cpu_mask(i, cpu_sibling_setup_map) {
348 if (cpu_llc_id[cpu] != BAD_APICID &&
349 cpu_llc_id[cpu] == cpu_llc_id[i]) {
350 cpu_set(i, c[cpu].llc_shared_map);
351 cpu_set(cpu, c[i].llc_shared_map);
353 if (c[cpu].phys_proc_id == c[i].phys_proc_id) {
354 cpu_set(i, cpu_core_map[cpu]);
355 cpu_set(cpu, cpu_core_map[i]);
357 * Does this new cpu bringup a new core?
359 if (cpus_weight(cpu_sibling_map[cpu]) == 1) {
361 * for each core in package, increment
362 * the booted_cores for this new cpu
364 if (first_cpu(cpu_sibling_map[i]) == i)
365 c[cpu].booted_cores++;
367 * increment the core count for all
368 * the other cpus in this package
372 } else if (i != cpu && !c[cpu].booted_cores)
373 c[cpu].booted_cores = c[i].booted_cores;
379 * Activate a secondary processor.
381 static void __cpuinit start_secondary(void *unused)
384 * Don't put *anything* before cpu_init(), SMP booting is too
385 * fragile that we want to limit the things done here to the
386 * most necessary things.
394 while (!cpu_isset(smp_processor_id(), smp_commenced_mask))
397 * Check TSC synchronization with the BP:
399 check_tsc_sync_target();
401 setup_secondary_clock();
402 if (nmi_watchdog == NMI_IO_APIC) {
403 disable_8259A_irq(0);
404 enable_NMI_through_LVT0(NULL);
408 * low-memory mappings have been cleared, flush them from
409 * the local TLBs too.
413 /* This must be done before setting cpu_online_map */
414 set_cpu_sibling_map(raw_smp_processor_id());
418 * We need to hold call_lock, so there is no inconsistency
419 * between the time smp_call_function() determines number of
420 * IPI receipients, and the time when the determination is made
421 * for which cpus receive the IPI. Holding this
422 * lock helps us to not include this cpu in a currently in progress
423 * smp_call_function().
425 lock_ipi_call_lock();
426 cpu_set(smp_processor_id(), cpu_online_map);
427 unlock_ipi_call_lock();
428 per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
430 /* We can take interrupts now: we're officially "up". */
438 * Everything has been set up for the secondary
439 * CPUs - they just need to reload everything
440 * from the task structure
441 * This function must not return.
443 void __devinit initialize_secondary(void)
446 * We don't actually need to load the full TSS,
447 * basically just the stack pointer and the eip.
454 :"m" (current->thread.esp),"m" (current->thread.eip));
457 /* Static state in head.S used to set up a CPU */
465 /* which logical CPUs are on which nodes */
466 cpumask_t node_2_cpu_mask[MAX_NUMNODES] __read_mostly =
467 { [0 ... MAX_NUMNODES-1] = CPU_MASK_NONE };
468 EXPORT_SYMBOL(node_2_cpu_mask);
469 /* which node each logical CPU is on */
470 int cpu_2_node[NR_CPUS] __read_mostly = { [0 ... NR_CPUS-1] = 0 };
471 EXPORT_SYMBOL(cpu_2_node);
473 /* set up a mapping between cpu and node. */
474 static inline void map_cpu_to_node(int cpu, int node)
476 printk("Mapping cpu %d to node %d\n", cpu, node);
477 cpu_set(cpu, node_2_cpu_mask[node]);
478 cpu_2_node[cpu] = node;
481 /* undo a mapping between cpu and node. */
482 static inline void unmap_cpu_to_node(int cpu)
486 printk("Unmapping cpu %d from all nodes\n", cpu);
487 for (node = 0; node < MAX_NUMNODES; node ++)
488 cpu_clear(cpu, node_2_cpu_mask[node]);
491 #else /* !CONFIG_NUMA */
493 #define map_cpu_to_node(cpu, node) ({})
494 #define unmap_cpu_to_node(cpu) ({})
496 #endif /* CONFIG_NUMA */
498 u8 cpu_2_logical_apicid[NR_CPUS] __read_mostly = { [0 ... NR_CPUS-1] = BAD_APICID };
500 static void map_cpu_to_logical_apicid(void)
502 int cpu = smp_processor_id();
503 int apicid = logical_smp_processor_id();
504 int node = apicid_to_node(apicid);
506 if (!node_online(node))
507 node = first_online_node;
509 cpu_2_logical_apicid[cpu] = apicid;
510 map_cpu_to_node(cpu, node);
513 static void unmap_cpu_to_logical_apicid(int cpu)
515 cpu_2_logical_apicid[cpu] = BAD_APICID;
516 unmap_cpu_to_node(cpu);
519 static inline void __inquire_remote_apic(int apicid)
521 int i, regs[] = { APIC_ID >> 4, APIC_LVR >> 4, APIC_SPIV >> 4 };
522 char *names[] = { "ID", "VERSION", "SPIV" };
524 unsigned long status;
526 printk("Inquiring remote APIC #%d...\n", apicid);
528 for (i = 0; i < ARRAY_SIZE(regs); i++) {
529 printk("... APIC #%d %s: ", apicid, names[i]);
534 status = safe_apic_wait_icr_idle();
536 printk("a previous APIC delivery may have failed\n");
538 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(apicid));
539 apic_write_around(APIC_ICR, APIC_DM_REMRD | regs[i]);
544 status = apic_read(APIC_ICR) & APIC_ICR_RR_MASK;
545 } while (status == APIC_ICR_RR_INPROG && timeout++ < 1000);
548 case APIC_ICR_RR_VALID:
549 status = apic_read(APIC_RRR);
550 printk("%lx\n", status);
558 #ifdef WAKE_SECONDARY_VIA_NMI
560 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
561 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
562 * won't ... remember to clear down the APIC, etc later.
565 wakeup_secondary_cpu(int logical_apicid, unsigned long start_eip)
567 unsigned long send_status, accept_status = 0;
571 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(logical_apicid));
573 /* Boot on the stack */
574 /* Kick the second */
575 apic_write_around(APIC_ICR, APIC_DM_NMI | APIC_DEST_LOGICAL);
577 Dprintk("Waiting for send to finish...\n");
578 send_status = safe_apic_wait_icr_idle();
581 * Give the other CPU some time to accept the IPI.
585 * Due to the Pentium erratum 3AP.
587 maxlvt = lapic_get_maxlvt();
589 apic_read_around(APIC_SPIV);
590 apic_write(APIC_ESR, 0);
592 accept_status = (apic_read(APIC_ESR) & 0xEF);
593 Dprintk("NMI sent.\n");
596 printk("APIC never delivered???\n");
598 printk("APIC delivery error (%lx).\n", accept_status);
600 return (send_status | accept_status);
602 #endif /* WAKE_SECONDARY_VIA_NMI */
604 #ifdef WAKE_SECONDARY_VIA_INIT
606 wakeup_secondary_cpu(int phys_apicid, unsigned long start_eip)
608 unsigned long send_status, accept_status = 0;
609 int maxlvt, num_starts, j;
612 * Be paranoid about clearing APIC errors.
614 if (APIC_INTEGRATED(apic_version[phys_apicid])) {
615 apic_read_around(APIC_SPIV);
616 apic_write(APIC_ESR, 0);
620 Dprintk("Asserting INIT.\n");
623 * Turn INIT on target chip
625 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
630 apic_write_around(APIC_ICR, APIC_INT_LEVELTRIG | APIC_INT_ASSERT
633 Dprintk("Waiting for send to finish...\n");
634 send_status = safe_apic_wait_icr_idle();
638 Dprintk("Deasserting INIT.\n");
641 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
644 apic_write_around(APIC_ICR, APIC_INT_LEVELTRIG | APIC_DM_INIT);
646 Dprintk("Waiting for send to finish...\n");
647 send_status = safe_apic_wait_icr_idle();
649 atomic_set(&init_deasserted, 1);
652 * Should we send STARTUP IPIs ?
654 * Determine this based on the APIC version.
655 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
657 if (APIC_INTEGRATED(apic_version[phys_apicid]))
663 * Paravirt / VMI wants a startup IPI hook here to set up the
664 * target processor state.
666 startup_ipi_hook(phys_apicid, (unsigned long) start_secondary,
667 (unsigned long) stack_start.esp);
670 * Run STARTUP IPI loop.
672 Dprintk("#startup loops: %d.\n", num_starts);
674 maxlvt = lapic_get_maxlvt();
676 for (j = 1; j <= num_starts; j++) {
677 Dprintk("Sending STARTUP #%d.\n",j);
678 apic_read_around(APIC_SPIV);
679 apic_write(APIC_ESR, 0);
681 Dprintk("After apic_write.\n");
688 apic_write_around(APIC_ICR2, SET_APIC_DEST_FIELD(phys_apicid));
690 /* Boot on the stack */
691 /* Kick the second */
692 apic_write_around(APIC_ICR, APIC_DM_STARTUP
693 | (start_eip >> 12));
696 * Give the other CPU some time to accept the IPI.
700 Dprintk("Startup point 1.\n");
702 Dprintk("Waiting for send to finish...\n");
703 send_status = safe_apic_wait_icr_idle();
706 * Give the other CPU some time to accept the IPI.
710 * Due to the Pentium erratum 3AP.
713 apic_read_around(APIC_SPIV);
714 apic_write(APIC_ESR, 0);
716 accept_status = (apic_read(APIC_ESR) & 0xEF);
717 if (send_status || accept_status)
720 Dprintk("After Startup.\n");
723 printk("APIC never delivered???\n");
725 printk("APIC delivery error (%lx).\n", accept_status);
727 return (send_status | accept_status);
729 #endif /* WAKE_SECONDARY_VIA_INIT */
731 extern cpumask_t cpu_initialized;
732 static inline int alloc_cpu_id(void)
736 cpus_complement(tmp_map, cpu_present_map);
737 cpu = first_cpu(tmp_map);
743 #ifdef CONFIG_HOTPLUG_CPU
744 static struct task_struct * __devinitdata cpu_idle_tasks[NR_CPUS];
745 static inline struct task_struct * alloc_idle_task(int cpu)
747 struct task_struct *idle;
749 if ((idle = cpu_idle_tasks[cpu]) != NULL) {
750 /* initialize thread_struct. we really want to avoid destroy
753 idle->thread.esp = (unsigned long)task_pt_regs(idle);
754 init_idle(idle, cpu);
757 idle = fork_idle(cpu);
760 cpu_idle_tasks[cpu] = idle;
764 #define alloc_idle_task(cpu) fork_idle(cpu)
767 /* Initialize the CPU's GDT. This is either the boot CPU doing itself
768 (still using the master per-cpu area), or a CPU doing it for a
769 secondary which will soon come up. */
770 static __cpuinit void init_gdt(int cpu)
772 struct desc_struct *gdt = get_cpu_gdt_table(cpu);
774 pack_descriptor((u32 *)&gdt[GDT_ENTRY_PERCPU].a,
775 (u32 *)&gdt[GDT_ENTRY_PERCPU].b,
776 __per_cpu_offset[cpu], 0xFFFFF,
777 0x80 | DESCTYPE_S | 0x2, 0x8);
779 per_cpu(this_cpu_off, cpu) = __per_cpu_offset[cpu];
780 per_cpu(cpu_number, cpu) = cpu;
783 /* Defined in head.S */
784 extern struct Xgt_desc_struct early_gdt_descr;
786 static int __cpuinit do_boot_cpu(int apicid, int cpu)
788 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
789 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
790 * Returns zero if CPU booted OK, else error code from wakeup_secondary_cpu.
793 struct task_struct *idle;
794 unsigned long boot_error;
796 unsigned long start_eip;
797 unsigned short nmi_high = 0, nmi_low = 0;
800 * Save current MTRR state in case it was changed since early boot
801 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
806 * We can't use kernel_thread since we must avoid to
807 * reschedule the child.
809 idle = alloc_idle_task(cpu);
811 panic("failed fork for CPU %d", cpu);
814 per_cpu(current_task, cpu) = idle;
815 early_gdt_descr.address = (unsigned long)get_cpu_gdt_table(cpu);
817 idle->thread.eip = (unsigned long) start_secondary;
818 /* start_eip had better be page-aligned! */
819 start_eip = setup_trampoline();
822 alternatives_smp_switch(1);
824 /* So we see what's up */
825 printk("Booting processor %d/%d eip %lx\n", cpu, apicid, start_eip);
826 /* Stack for startup_32 can be just as for start_secondary onwards */
827 stack_start.esp = (void *) idle->thread.esp;
831 x86_cpu_to_apicid[cpu] = apicid;
833 * This grunge runs the startup process for
834 * the targeted processor.
837 atomic_set(&init_deasserted, 0);
839 Dprintk("Setting warm reset code and vector.\n");
841 store_NMI_vector(&nmi_high, &nmi_low);
843 smpboot_setup_warm_reset_vector(start_eip);
846 * Starting actual IPI sequence...
848 boot_error = wakeup_secondary_cpu(apicid, start_eip);
852 * allow APs to start initializing.
854 Dprintk("Before Callout %d.\n", cpu);
855 cpu_set(cpu, cpu_callout_map);
856 Dprintk("After Callout %d.\n", cpu);
859 * Wait 5s total for a response
861 for (timeout = 0; timeout < 50000; timeout++) {
862 if (cpu_isset(cpu, cpu_callin_map))
863 break; /* It has booted */
867 if (cpu_isset(cpu, cpu_callin_map)) {
868 /* number CPUs logically, starting from 1 (BSP is 0) */
870 printk("CPU%d: ", cpu);
871 print_cpu_info(&cpu_data[cpu]);
872 Dprintk("CPU has booted.\n");
875 if (*((volatile unsigned char *)trampoline_base)
877 /* trampoline started but...? */
878 printk("Stuck ??\n");
880 /* trampoline code not run */
881 printk("Not responding.\n");
882 inquire_remote_apic(apicid);
887 /* Try to put things back the way they were before ... */
888 unmap_cpu_to_logical_apicid(cpu);
889 cpu_clear(cpu, cpu_callout_map); /* was set here (do_boot_cpu()) */
890 cpu_clear(cpu, cpu_initialized); /* was set by cpu_init() */
893 x86_cpu_to_apicid[cpu] = apicid;
894 cpu_set(cpu, cpu_present_map);
897 /* mark "stuck" area as not stuck */
898 *((volatile unsigned long *)trampoline_base) = 0;
903 #ifdef CONFIG_HOTPLUG_CPU
904 void cpu_exit_clear(void)
906 int cpu = raw_smp_processor_id();
914 cpu_clear(cpu, cpu_callout_map);
915 cpu_clear(cpu, cpu_callin_map);
917 cpu_clear(cpu, smp_commenced_mask);
918 unmap_cpu_to_logical_apicid(cpu);
921 struct warm_boot_cpu_info {
922 struct completion *complete;
923 struct work_struct task;
928 static void __cpuinit do_warm_boot_cpu(struct work_struct *work)
930 struct warm_boot_cpu_info *info =
931 container_of(work, struct warm_boot_cpu_info, task);
932 do_boot_cpu(info->apicid, info->cpu);
933 complete(info->complete);
936 static int __cpuinit __smp_prepare_cpu(int cpu)
938 DECLARE_COMPLETION_ONSTACK(done);
939 struct warm_boot_cpu_info info;
942 apicid = x86_cpu_to_apicid[cpu];
943 if (apicid == BAD_APICID) {
948 info.complete = &done;
949 info.apicid = apicid;
951 INIT_WORK(&info.task, do_warm_boot_cpu);
953 /* init low mem mapping */
954 clone_pgd_range(swapper_pg_dir, swapper_pg_dir + USER_PGD_PTRS,
955 min_t(unsigned long, KERNEL_PGD_PTRS, USER_PGD_PTRS));
957 schedule_work(&info.task);
958 wait_for_completion(&done);
967 static void smp_tune_scheduling(void)
969 unsigned long cachesize; /* kB */
972 cachesize = boot_cpu_data.x86_cache_size;
975 max_cache_size = cachesize * 1024;
980 * Cycle through the processors sending APIC IPIs to boot each.
983 static int boot_cpu_logical_apicid;
984 /* Where the IO area was mapped on multiquad, always 0 otherwise */
986 #ifdef CONFIG_X86_NUMAQ
987 EXPORT_SYMBOL(xquad_portio);
990 static void __init smp_boot_cpus(unsigned int max_cpus)
992 int apicid, cpu, bit, kicked;
993 unsigned long bogosum = 0;
996 * Setup boot CPU information
998 smp_store_cpu_info(0); /* Final full version of the data */
999 printk("CPU%d: ", 0);
1000 print_cpu_info(&cpu_data[0]);
1002 boot_cpu_physical_apicid = GET_APIC_ID(apic_read(APIC_ID));
1003 boot_cpu_logical_apicid = logical_smp_processor_id();
1004 x86_cpu_to_apicid[0] = boot_cpu_physical_apicid;
1006 current_thread_info()->cpu = 0;
1007 smp_tune_scheduling();
1009 set_cpu_sibling_map(0);
1012 * If we couldn't find an SMP configuration at boot time,
1013 * get out of here now!
1015 if (!smp_found_config && !acpi_lapic) {
1016 printk(KERN_NOTICE "SMP motherboard not detected.\n");
1017 smpboot_clear_io_apic_irqs();
1018 phys_cpu_present_map = physid_mask_of_physid(0);
1019 if (APIC_init_uniprocessor())
1020 printk(KERN_NOTICE "Local APIC not detected."
1021 " Using dummy APIC emulation.\n");
1022 map_cpu_to_logical_apicid();
1023 cpu_set(0, cpu_sibling_map[0]);
1024 cpu_set(0, cpu_core_map[0]);
1029 * Should not be necessary because the MP table should list the boot
1030 * CPU too, but we do it for the sake of robustness anyway.
1031 * Makes no sense to do this check in clustered apic mode, so skip it
1033 if (!check_phys_apicid_present(boot_cpu_physical_apicid)) {
1034 printk("weird, boot CPU (#%d) not listed by the BIOS.\n",
1035 boot_cpu_physical_apicid);
1036 physid_set(hard_smp_processor_id(), phys_cpu_present_map);
1040 * If we couldn't find a local APIC, then get out of here now!
1042 if (APIC_INTEGRATED(apic_version[boot_cpu_physical_apicid]) && !cpu_has_apic) {
1043 printk(KERN_ERR "BIOS bug, local APIC #%d not detected!...\n",
1044 boot_cpu_physical_apicid);
1045 printk(KERN_ERR "... forcing use of dummy APIC emulation. (tell your hw vendor)\n");
1046 smpboot_clear_io_apic_irqs();
1047 phys_cpu_present_map = physid_mask_of_physid(0);
1048 cpu_set(0, cpu_sibling_map[0]);
1049 cpu_set(0, cpu_core_map[0]);
1053 verify_local_APIC();
1056 * If SMP should be disabled, then really disable it!
1059 smp_found_config = 0;
1060 printk(KERN_INFO "SMP mode deactivated, forcing use of dummy APIC emulation.\n");
1061 smpboot_clear_io_apic_irqs();
1062 phys_cpu_present_map = physid_mask_of_physid(0);
1063 cpu_set(0, cpu_sibling_map[0]);
1064 cpu_set(0, cpu_core_map[0]);
1070 map_cpu_to_logical_apicid();
1073 setup_portio_remap();
1076 * Scan the CPU present map and fire up the other CPUs via do_boot_cpu
1078 * In clustered apic mode, phys_cpu_present_map is a constructed thus:
1079 * bits 0-3 are quad0, 4-7 are quad1, etc. A perverse twist on the
1080 * clustered apic ID.
1082 Dprintk("CPU present map: %lx\n", physids_coerce(phys_cpu_present_map));
1085 for (bit = 0; kicked < NR_CPUS && bit < MAX_APICS; bit++) {
1086 apicid = cpu_present_to_apicid(bit);
1088 * Don't even attempt to start the boot CPU!
1090 if ((apicid == boot_cpu_apicid) || (apicid == BAD_APICID))
1093 if (!check_apicid_present(bit))
1095 if (max_cpus <= cpucount+1)
1098 if (((cpu = alloc_cpu_id()) <= 0) || do_boot_cpu(apicid, cpu))
1099 printk("CPU #%d not responding - cannot use it.\n",
1106 * Cleanup possible dangling ends...
1108 smpboot_restore_warm_reset_vector();
1111 * Allow the user to impress friends.
1113 Dprintk("Before bogomips.\n");
1114 for (cpu = 0; cpu < NR_CPUS; cpu++)
1115 if (cpu_isset(cpu, cpu_callout_map))
1116 bogosum += cpu_data[cpu].loops_per_jiffy;
1118 "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
1120 bogosum/(500000/HZ),
1121 (bogosum/(5000/HZ))%100);
1123 Dprintk("Before bogocount - setting activated=1.\n");
1126 printk(KERN_WARNING "WARNING: SMP operation may be unreliable with B stepping processors.\n");
1129 * Don't taint if we are running SMP kernel on a single non-MP
1132 if (tainted & TAINT_UNSAFE_SMP) {
1134 printk (KERN_INFO "WARNING: This combination of AMD processors is not suitable for SMP.\n");
1136 tainted &= ~TAINT_UNSAFE_SMP;
1139 Dprintk("Boot done.\n");
1142 * construct cpu_sibling_map[], so that we can tell sibling CPUs
1145 for (cpu = 0; cpu < NR_CPUS; cpu++) {
1146 cpus_clear(cpu_sibling_map[cpu]);
1147 cpus_clear(cpu_core_map[cpu]);
1150 cpu_set(0, cpu_sibling_map[0]);
1151 cpu_set(0, cpu_core_map[0]);
1153 smpboot_setup_io_apic();
1158 /* These are wrappers to interface to the new boot process. Someone
1159 who understands all this stuff should rewrite it properly. --RR 15/Jul/02 */
1160 void __init native_smp_prepare_cpus(unsigned int max_cpus)
1162 smp_commenced_mask = cpumask_of_cpu(0);
1163 cpu_callin_map = cpumask_of_cpu(0);
1165 smp_boot_cpus(max_cpus);
1168 void __init native_smp_prepare_boot_cpu(void)
1170 unsigned int cpu = smp_processor_id();
1173 switch_to_new_gdt();
1175 cpu_set(cpu, cpu_online_map);
1176 cpu_set(cpu, cpu_callout_map);
1177 cpu_set(cpu, cpu_present_map);
1178 cpu_set(cpu, cpu_possible_map);
1179 __get_cpu_var(cpu_state) = CPU_ONLINE;
1182 #ifdef CONFIG_HOTPLUG_CPU
1184 remove_siblinginfo(int cpu)
1187 struct cpuinfo_x86 *c = cpu_data;
1189 for_each_cpu_mask(sibling, cpu_core_map[cpu]) {
1190 cpu_clear(cpu, cpu_core_map[sibling]);
1192 * last thread sibling in this cpu core going down
1194 if (cpus_weight(cpu_sibling_map[cpu]) == 1)
1195 c[sibling].booted_cores--;
1198 for_each_cpu_mask(sibling, cpu_sibling_map[cpu])
1199 cpu_clear(cpu, cpu_sibling_map[sibling]);
1200 cpus_clear(cpu_sibling_map[cpu]);
1201 cpus_clear(cpu_core_map[cpu]);
1202 c[cpu].phys_proc_id = 0;
1203 c[cpu].cpu_core_id = 0;
1204 cpu_clear(cpu, cpu_sibling_setup_map);
1207 int __cpu_disable(void)
1209 cpumask_t map = cpu_online_map;
1210 int cpu = smp_processor_id();
1213 * Perhaps use cpufreq to drop frequency, but that could go
1214 * into generic code.
1216 * We won't take down the boot processor on i386 due to some
1217 * interrupts only being able to be serviced by the BSP.
1218 * Especially so if we're not using an IOAPIC -zwane
1222 if (nmi_watchdog == NMI_LOCAL_APIC)
1223 stop_apic_nmi_watchdog(NULL);
1225 /* Allow any queued timer interrupts to get serviced */
1228 local_irq_disable();
1230 remove_siblinginfo(cpu);
1232 cpu_clear(cpu, map);
1234 /* It's now safe to remove this processor from the online map */
1235 cpu_clear(cpu, cpu_online_map);
1239 void __cpu_die(unsigned int cpu)
1241 /* We don't do anything here: idle task is faking death itself. */
1244 for (i = 0; i < 10; i++) {
1245 /* They ack this in play_dead by setting CPU_DEAD */
1246 if (per_cpu(cpu_state, cpu) == CPU_DEAD) {
1247 printk ("CPU %d is now offline\n", cpu);
1248 if (1 == num_online_cpus())
1249 alternatives_smp_switch(0);
1254 printk(KERN_ERR "CPU %u didn't die...\n", cpu);
1256 #else /* ... !CONFIG_HOTPLUG_CPU */
1257 int __cpu_disable(void)
1262 void __cpu_die(unsigned int cpu)
1264 /* We said "no" in __cpu_disable */
1267 #endif /* CONFIG_HOTPLUG_CPU */
1269 int __cpuinit native_cpu_up(unsigned int cpu)
1271 unsigned long flags;
1272 #ifdef CONFIG_HOTPLUG_CPU
1276 * We do warm boot only on cpus that had booted earlier
1277 * Otherwise cold boot is all handled from smp_boot_cpus().
1278 * cpu_callin_map is set during AP kickstart process. Its reset
1279 * when a cpu is taken offline from cpu_exit_clear().
1281 if (!cpu_isset(cpu, cpu_callin_map))
1282 ret = __smp_prepare_cpu(cpu);
1288 /* In case one didn't come up */
1289 if (!cpu_isset(cpu, cpu_callin_map)) {
1290 printk(KERN_DEBUG "skipping cpu%d, didn't come online\n", cpu);
1294 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
1295 /* Unleash the CPU! */
1296 cpu_set(cpu, smp_commenced_mask);
1299 * Check TSC synchronization with the AP (keep irqs disabled
1302 local_irq_save(flags);
1303 check_tsc_sync_source(cpu);
1304 local_irq_restore(flags);
1306 while (!cpu_isset(cpu, cpu_online_map)) {
1308 touch_nmi_watchdog();
1314 void __init native_smp_cpus_done(unsigned int max_cpus)
1316 #ifdef CONFIG_X86_IO_APIC
1317 setup_ioapic_dest();
1320 #ifndef CONFIG_HOTPLUG_CPU
1322 * Disable executability of the SMP trampoline:
1324 set_kernel_exec((unsigned long)trampoline_base, trampoline_exec);
1328 void __init smp_intr_init(void)
1331 * IRQ0 must be given a fixed assignment and initialized,
1332 * because it's used before the IO-APIC is set up.
1334 set_intr_gate(FIRST_DEVICE_VECTOR, interrupt[0]);
1337 * The reschedule interrupt is a CPU-to-CPU reschedule-helper
1338 * IPI, driven by wakeup.
1340 set_intr_gate(RESCHEDULE_VECTOR, reschedule_interrupt);
1342 /* IPI for invalidation */
1343 set_intr_gate(INVALIDATE_TLB_VECTOR, invalidate_interrupt);
1345 /* IPI for generic function call */
1346 set_intr_gate(CALL_FUNCTION_VECTOR, call_function_interrupt);
1350 * If the BIOS enumerates physical processors before logical,
1351 * maxcpus=N at enumeration-time can be used to disable HT.
1353 static int __init parse_maxcpus(char *arg)
1355 extern unsigned int maxcpus;
1357 maxcpus = simple_strtoul(arg, NULL, 0);
1360 early_param("maxcpus", parse_maxcpus);