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>
6 * Copyright 2001 Andi Kleen, SuSE Labs.
8 * Much of the core SMP work is based on previous work by Thomas Radke, to
9 * whom a great many thanks are extended.
11 * Thanks to Intel for making available several different Pentium,
12 * Pentium Pro and Pentium-II/Xeon MP machines.
13 * Original development of Linux SMP code supported by Caldera.
15 * This code is released under the GNU General Public License version 2 or
19 * Felix Koop : NR_CPUS used properly
20 * Jose Renau : Handle single CPU case.
21 * Alan Cox : By repeated request 8) - Total BogoMIPS report.
22 * Greg Wright : Fix for kernel stacks panic.
23 * Erich Boleyn : MP v1.4 and additional changes.
24 * Matthias Sattler : Changes for 2.1 kernel map.
25 * Michel Lespinasse : Changes for 2.1 kernel map.
26 * Michael Chastain : Change trampoline.S to gnu as.
27 * Alan Cox : Dumb bug: 'B' step PPro's are fine
28 * Ingo Molnar : Added APIC timers, based on code
30 * Ingo Molnar : various cleanups and rewrites
31 * Tigran Aivazian : fixed "0.00 in /proc/uptime on SMP" bug.
32 * Maciej W. Rozycki : Bits for genuine 82489DX APICs
33 * Andi Kleen : Changed for SMP boot into long mode.
34 * Martin J. Bligh : Added support for multi-quad systems
35 * Dave Jones : Report invalid combinations of Athlon CPUs.
36 * Rusty Russell : Hacked into shape for new "hotplug" boot process.
37 * Andi Kleen : Converted to new state machine.
38 * Ashok Raj : CPU hotplug support
39 * Glauber Costa : i386 and x86_64 integration
42 #include <linux/init.h>
43 #include <linux/smp.h>
44 #include <linux/module.h>
45 #include <linux/sched.h>
46 #include <linux/percpu.h>
47 #include <linux/bootmem.h>
48 #include <linux/err.h>
49 #include <linux/nmi.h>
56 #include <asm/trampoline.h>
59 #include <asm/pgtable.h>
60 #include <asm/tlbflush.h>
63 #include <asm/genapic.h>
64 #include <linux/mc146818rtc.h>
66 #include <mach_apic.h>
67 #include <mach_wakecpu.h>
68 #include <smpboot_hooks.h>
71 u8 apicid_2_node
[MAX_APICID
];
72 static int low_mappings
;
75 /* State of each CPU */
76 DEFINE_PER_CPU(int, cpu_state
) = { 0 };
78 /* Store all idle threads, this can be reused instead of creating
79 * a new thread. Also avoids complicated thread destroy functionality
82 #ifdef CONFIG_HOTPLUG_CPU
84 * Needed only for CONFIG_HOTPLUG_CPU because __cpuinitdata is
85 * removed after init for !CONFIG_HOTPLUG_CPU.
87 static DEFINE_PER_CPU(struct task_struct
*, idle_thread_array
);
88 #define get_idle_for_cpu(x) (per_cpu(idle_thread_array, x))
89 #define set_idle_for_cpu(x, p) (per_cpu(idle_thread_array, x) = (p))
91 struct task_struct
*idle_thread_array
[NR_CPUS
] __cpuinitdata
;
92 #define get_idle_for_cpu(x) (idle_thread_array[(x)])
93 #define set_idle_for_cpu(x, p) (idle_thread_array[(x)] = (p))
96 /* Number of siblings per CPU package */
97 int smp_num_siblings
= 1;
98 EXPORT_SYMBOL(smp_num_siblings
);
100 /* Last level cache ID of each logical CPU */
101 DEFINE_PER_CPU(u16
, cpu_llc_id
) = BAD_APICID
;
103 /* bitmap of online cpus */
104 cpumask_t cpu_online_map __read_mostly
;
105 EXPORT_SYMBOL(cpu_online_map
);
107 cpumask_t cpu_callin_map
;
108 cpumask_t cpu_callout_map
;
109 cpumask_t cpu_possible_map
;
110 EXPORT_SYMBOL(cpu_possible_map
);
112 /* representing HT siblings of each logical CPU */
113 DEFINE_PER_CPU(cpumask_t
, cpu_sibling_map
);
114 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map
);
116 /* representing HT and core siblings of each logical CPU */
117 DEFINE_PER_CPU(cpumask_t
, cpu_core_map
);
118 EXPORT_PER_CPU_SYMBOL(cpu_core_map
);
120 /* Per CPU bogomips and other parameters */
121 DEFINE_PER_CPU_SHARED_ALIGNED(struct cpuinfo_x86
, cpu_info
);
122 EXPORT_PER_CPU_SYMBOL(cpu_info
);
124 static atomic_t init_deasserted
;
126 static int boot_cpu_logical_apicid
;
128 /* representing cpus for which sibling maps can be computed */
129 static cpumask_t cpu_sibling_setup_map
;
131 /* Set if we find a B stepping CPU */
132 int __cpuinitdata smp_b_stepping
;
134 #if defined(CONFIG_NUMA) && defined(CONFIG_X86_32)
136 /* which logical CPUs are on which nodes */
137 cpumask_t node_to_cpumask_map
[MAX_NUMNODES
] __read_mostly
=
138 { [0 ... MAX_NUMNODES
-1] = CPU_MASK_NONE
};
139 EXPORT_SYMBOL(node_to_cpumask_map
);
140 /* which node each logical CPU is on */
141 int cpu_to_node_map
[NR_CPUS
] __read_mostly
= { [0 ... NR_CPUS
-1] = 0 };
142 EXPORT_SYMBOL(cpu_to_node_map
);
144 /* set up a mapping between cpu and node. */
145 static void map_cpu_to_node(int cpu
, int node
)
147 printk(KERN_INFO
"Mapping cpu %d to node %d\n", cpu
, node
);
148 cpu_set(cpu
, node_to_cpumask_map
[node
]);
149 cpu_to_node_map
[cpu
] = node
;
152 /* undo a mapping between cpu and node. */
153 static void unmap_cpu_to_node(int cpu
)
157 printk(KERN_INFO
"Unmapping cpu %d from all nodes\n", cpu
);
158 for (node
= 0; node
< MAX_NUMNODES
; node
++)
159 cpu_clear(cpu
, node_to_cpumask_map
[node
]);
160 cpu_to_node_map
[cpu
] = 0;
162 #else /* !(CONFIG_NUMA && CONFIG_X86_32) */
163 #define map_cpu_to_node(cpu, node) ({})
164 #define unmap_cpu_to_node(cpu) ({})
168 u8 cpu_2_logical_apicid
[NR_CPUS
] __read_mostly
=
169 { [0 ... NR_CPUS
-1] = BAD_APICID
};
171 static void map_cpu_to_logical_apicid(void)
173 int cpu
= smp_processor_id();
174 int apicid
= logical_smp_processor_id();
175 int node
= apicid_to_node(apicid
);
177 if (!node_online(node
))
178 node
= first_online_node
;
180 cpu_2_logical_apicid
[cpu
] = apicid
;
181 map_cpu_to_node(cpu
, node
);
184 static void unmap_cpu_to_logical_apicid(int cpu
)
186 cpu_2_logical_apicid
[cpu
] = BAD_APICID
;
187 unmap_cpu_to_node(cpu
);
190 #define unmap_cpu_to_logical_apicid(cpu) do {} while (0)
191 #define map_cpu_to_logical_apicid() do {} while (0)
195 * Report back to the Boot Processor.
198 static void __cpuinit
smp_callin(void)
201 unsigned long timeout
;
204 * If waken up by an INIT in an 82489DX configuration
205 * we may get here before an INIT-deassert IPI reaches
206 * our local APIC. We have to wait for the IPI or we'll
207 * lock up on an APIC access.
209 wait_for_init_deassert(&init_deasserted
);
212 * (This works even if the APIC is not enabled.)
214 phys_id
= GET_APIC_ID(read_apic_id());
215 cpuid
= smp_processor_id();
216 if (cpu_isset(cpuid
, cpu_callin_map
)) {
217 panic("%s: phys CPU#%d, CPU#%d already present??\n", __func__
,
220 Dprintk("CPU#%d (phys ID: %d) waiting for CALLOUT\n", cpuid
, phys_id
);
223 * STARTUP IPIs are fragile beasts as they might sometimes
224 * trigger some glue motherboard logic. Complete APIC bus
225 * silence for 1 second, this overestimates the time the
226 * boot CPU is spending to send the up to 2 STARTUP IPIs
227 * by a factor of two. This should be enough.
231 * Waiting 2s total for startup (udelay is not yet working)
233 timeout
= jiffies
+ 2*HZ
;
234 while (time_before(jiffies
, timeout
)) {
236 * Has the boot CPU finished it's STARTUP sequence?
238 if (cpu_isset(cpuid
, cpu_callout_map
))
243 if (!time_before(jiffies
, timeout
)) {
244 panic("%s: CPU%d started up but did not get a callout!\n",
249 * the boot CPU has finished the init stage and is spinning
250 * on callin_map until we finish. We are free to set up this
251 * CPU, first the APIC. (this is probably redundant on most
255 Dprintk("CALLIN, before setup_local_APIC().\n");
256 smp_callin_clear_local_apic();
258 end_local_APIC_setup();
259 map_cpu_to_logical_apicid();
264 * Need to enable IRQs because it can take longer and then
265 * the NMI watchdog might kill us.
270 Dprintk("Stack at about %p\n", &cpuid
);
273 * Save our processor parameters
275 smp_store_cpu_info(cpuid
);
278 * Allow the master to continue.
280 cpu_set(cpuid
, cpu_callin_map
);
284 * Activate a secondary processor.
286 static void __cpuinit
start_secondary(void *unused
)
289 * Don't put *anything* before cpu_init(), SMP booting is too
290 * fragile that we want to limit the things done here to the
291 * most necessary things.
300 /* otherwise gcc will move up smp_processor_id before the cpu_init */
303 * Check TSC synchronization with the BP:
305 check_tsc_sync_target();
307 if (nmi_watchdog
== NMI_IO_APIC
) {
308 disable_8259A_irq(0);
309 enable_NMI_through_LVT0();
319 /* This must be done before setting cpu_online_map */
320 set_cpu_sibling_map(raw_smp_processor_id());
324 * We need to hold call_lock, so there is no inconsistency
325 * between the time smp_call_function() determines number of
326 * IPI recipients, and the time when the determination is made
327 * for which cpus receive the IPI. Holding this
328 * lock helps us to not include this cpu in a currently in progress
329 * smp_call_function().
331 lock_ipi_call_lock();
333 spin_lock(&vector_lock
);
335 /* Setup the per cpu irq handling data structures */
336 __setup_vector_irq(smp_processor_id());
338 * Allow the master to continue.
340 spin_unlock(&vector_lock
);
342 cpu_set(smp_processor_id(), cpu_online_map
);
343 unlock_ipi_call_lock();
344 per_cpu(cpu_state
, smp_processor_id()) = CPU_ONLINE
;
346 setup_secondary_clock();
352 static void __cpuinit
smp_apply_quirks(struct cpuinfo_x86
*c
)
356 * Mask B, Pentium, but not Pentium MMX
358 if (c
->x86_vendor
== X86_VENDOR_INTEL
&&
360 c
->x86_mask
>= 1 && c
->x86_mask
<= 4 &&
363 * Remember we have B step Pentia with bugs
368 * Certain Athlons might work (for various values of 'work') in SMP
369 * but they are not certified as MP capable.
371 if ((c
->x86_vendor
== X86_VENDOR_AMD
) && (c
->x86
== 6)) {
373 if (num_possible_cpus() == 1)
376 /* Athlon 660/661 is valid. */
377 if ((c
->x86_model
== 6) && ((c
->x86_mask
== 0) ||
381 /* Duron 670 is valid */
382 if ((c
->x86_model
== 7) && (c
->x86_mask
== 0))
386 * Athlon 662, Duron 671, and Athlon >model 7 have capability
387 * bit. It's worth noting that the A5 stepping (662) of some
388 * Athlon XP's have the MP bit set.
389 * See http://www.heise.de/newsticker/data/jow-18.10.01-000 for
392 if (((c
->x86_model
== 6) && (c
->x86_mask
>= 2)) ||
393 ((c
->x86_model
== 7) && (c
->x86_mask
>= 1)) ||
398 /* If we get here, not a certified SMP capable AMD system. */
399 add_taint(TAINT_UNSAFE_SMP
);
407 static void __cpuinit
smp_checks(void)
410 printk(KERN_WARNING
"WARNING: SMP operation may be unreliable"
411 "with B stepping processors.\n");
414 * Don't taint if we are running SMP kernel on a single non-MP
417 if (tainted
& TAINT_UNSAFE_SMP
) {
418 if (num_online_cpus())
419 printk(KERN_INFO
"WARNING: This combination of AMD"
420 "processors is not suitable for SMP.\n");
422 tainted
&= ~TAINT_UNSAFE_SMP
;
427 * The bootstrap kernel entry code has set these up. Save them for
431 void __cpuinit
smp_store_cpu_info(int id
)
433 struct cpuinfo_x86
*c
= &cpu_data(id
);
438 identify_secondary_cpu(c
);
443 void __cpuinit
set_cpu_sibling_map(int cpu
)
446 struct cpuinfo_x86
*c
= &cpu_data(cpu
);
448 cpu_set(cpu
, cpu_sibling_setup_map
);
450 if (smp_num_siblings
> 1) {
451 for_each_cpu_mask(i
, cpu_sibling_setup_map
) {
452 if (c
->phys_proc_id
== cpu_data(i
).phys_proc_id
&&
453 c
->cpu_core_id
== cpu_data(i
).cpu_core_id
) {
454 cpu_set(i
, per_cpu(cpu_sibling_map
, cpu
));
455 cpu_set(cpu
, per_cpu(cpu_sibling_map
, i
));
456 cpu_set(i
, per_cpu(cpu_core_map
, cpu
));
457 cpu_set(cpu
, per_cpu(cpu_core_map
, i
));
458 cpu_set(i
, c
->llc_shared_map
);
459 cpu_set(cpu
, cpu_data(i
).llc_shared_map
);
463 cpu_set(cpu
, per_cpu(cpu_sibling_map
, cpu
));
466 cpu_set(cpu
, c
->llc_shared_map
);
468 if (current_cpu_data
.x86_max_cores
== 1) {
469 per_cpu(cpu_core_map
, cpu
) = per_cpu(cpu_sibling_map
, cpu
);
474 for_each_cpu_mask(i
, cpu_sibling_setup_map
) {
475 if (per_cpu(cpu_llc_id
, cpu
) != BAD_APICID
&&
476 per_cpu(cpu_llc_id
, cpu
) == per_cpu(cpu_llc_id
, i
)) {
477 cpu_set(i
, c
->llc_shared_map
);
478 cpu_set(cpu
, cpu_data(i
).llc_shared_map
);
480 if (c
->phys_proc_id
== cpu_data(i
).phys_proc_id
) {
481 cpu_set(i
, per_cpu(cpu_core_map
, cpu
));
482 cpu_set(cpu
, per_cpu(cpu_core_map
, i
));
484 * Does this new cpu bringup a new core?
486 if (cpus_weight(per_cpu(cpu_sibling_map
, cpu
)) == 1) {
488 * for each core in package, increment
489 * the booted_cores for this new cpu
491 if (first_cpu(per_cpu(cpu_sibling_map
, i
)) == i
)
494 * increment the core count for all
495 * the other cpus in this package
498 cpu_data(i
).booted_cores
++;
499 } else if (i
!= cpu
&& !c
->booted_cores
)
500 c
->booted_cores
= cpu_data(i
).booted_cores
;
505 /* maps the cpu to the sched domain representing multi-core */
506 cpumask_t
cpu_coregroup_map(int cpu
)
508 struct cpuinfo_x86
*c
= &cpu_data(cpu
);
510 * For perf, we return last level cache shared map.
511 * And for power savings, we return cpu_core_map
513 if (sched_mc_power_savings
|| sched_smt_power_savings
)
514 return per_cpu(cpu_core_map
, cpu
);
516 return c
->llc_shared_map
;
519 static void impress_friends(void)
522 unsigned long bogosum
= 0;
524 * Allow the user to impress friends.
526 Dprintk("Before bogomips.\n");
527 for_each_possible_cpu(cpu
)
528 if (cpu_isset(cpu
, cpu_callout_map
))
529 bogosum
+= cpu_data(cpu
).loops_per_jiffy
;
531 "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
534 (bogosum
/(5000/HZ
))%100);
536 Dprintk("Before bogocount - setting activated=1.\n");
539 static inline void __inquire_remote_apic(int apicid
)
541 unsigned i
, regs
[] = { APIC_ID
>> 4, APIC_LVR
>> 4, APIC_SPIV
>> 4 };
542 char *names
[] = { "ID", "VERSION", "SPIV" };
546 printk(KERN_INFO
"Inquiring remote APIC #%d...\n", apicid
);
548 for (i
= 0; i
< ARRAY_SIZE(regs
); i
++) {
549 printk(KERN_INFO
"... APIC #%d %s: ", apicid
, names
[i
]);
554 status
= safe_apic_wait_icr_idle();
557 "a previous APIC delivery may have failed\n");
559 apic_write_around(APIC_ICR2
, SET_APIC_DEST_FIELD(apicid
));
560 apic_write_around(APIC_ICR
, APIC_DM_REMRD
| regs
[i
]);
565 status
= apic_read(APIC_ICR
) & APIC_ICR_RR_MASK
;
566 } while (status
== APIC_ICR_RR_INPROG
&& timeout
++ < 1000);
569 case APIC_ICR_RR_VALID
:
570 status
= apic_read(APIC_RRR
);
571 printk(KERN_CONT
"%08x\n", status
);
574 printk(KERN_CONT
"failed\n");
579 #ifdef WAKE_SECONDARY_VIA_NMI
581 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
582 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
583 * won't ... remember to clear down the APIC, etc later.
586 wakeup_secondary_cpu(int logical_apicid
, unsigned long start_eip
)
588 unsigned long send_status
, accept_status
= 0;
592 apic_write_around(APIC_ICR2
, SET_APIC_DEST_FIELD(logical_apicid
));
594 /* Boot on the stack */
595 /* Kick the second */
596 apic_write_around(APIC_ICR
, APIC_DM_NMI
| APIC_DEST_LOGICAL
);
598 Dprintk("Waiting for send to finish...\n");
599 send_status
= safe_apic_wait_icr_idle();
602 * Give the other CPU some time to accept the IPI.
606 * Due to the Pentium erratum 3AP.
608 maxlvt
= lapic_get_maxlvt();
610 apic_read_around(APIC_SPIV
);
611 apic_write(APIC_ESR
, 0);
613 accept_status
= (apic_read(APIC_ESR
) & 0xEF);
614 Dprintk("NMI sent.\n");
617 printk(KERN_ERR
"APIC never delivered???\n");
619 printk(KERN_ERR
"APIC delivery error (%lx).\n", accept_status
);
621 return (send_status
| accept_status
);
623 #endif /* WAKE_SECONDARY_VIA_NMI */
625 #ifdef WAKE_SECONDARY_VIA_INIT
627 wakeup_secondary_cpu(int phys_apicid
, unsigned long start_eip
)
629 unsigned long send_status
, accept_status
= 0;
630 int maxlvt
, num_starts
, j
;
632 if (get_uv_system_type() == UV_NON_UNIQUE_APIC
) {
633 send_status
= uv_wakeup_secondary(phys_apicid
, start_eip
);
634 atomic_set(&init_deasserted
, 1);
639 * Be paranoid about clearing APIC errors.
641 if (APIC_INTEGRATED(apic_version
[phys_apicid
])) {
642 apic_read_around(APIC_SPIV
);
643 apic_write(APIC_ESR
, 0);
647 Dprintk("Asserting INIT.\n");
650 * Turn INIT on target chip
652 apic_write_around(APIC_ICR2
, SET_APIC_DEST_FIELD(phys_apicid
));
657 apic_write_around(APIC_ICR
, APIC_INT_LEVELTRIG
| APIC_INT_ASSERT
660 Dprintk("Waiting for send to finish...\n");
661 send_status
= safe_apic_wait_icr_idle();
665 Dprintk("Deasserting INIT.\n");
668 apic_write_around(APIC_ICR2
, SET_APIC_DEST_FIELD(phys_apicid
));
671 apic_write_around(APIC_ICR
, APIC_INT_LEVELTRIG
| APIC_DM_INIT
);
673 Dprintk("Waiting for send to finish...\n");
674 send_status
= safe_apic_wait_icr_idle();
677 atomic_set(&init_deasserted
, 1);
680 * Should we send STARTUP IPIs ?
682 * Determine this based on the APIC version.
683 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
685 if (APIC_INTEGRATED(apic_version
[phys_apicid
]))
691 * Paravirt / VMI wants a startup IPI hook here to set up the
692 * target processor state.
694 startup_ipi_hook(phys_apicid
, (unsigned long) start_secondary
,
695 (unsigned long)stack_start
.sp
);
698 * Run STARTUP IPI loop.
700 Dprintk("#startup loops: %d.\n", num_starts
);
702 maxlvt
= lapic_get_maxlvt();
704 for (j
= 1; j
<= num_starts
; j
++) {
705 Dprintk("Sending STARTUP #%d.\n", j
);
706 apic_read_around(APIC_SPIV
);
707 apic_write(APIC_ESR
, 0);
709 Dprintk("After apic_write.\n");
716 apic_write_around(APIC_ICR2
, SET_APIC_DEST_FIELD(phys_apicid
));
718 /* Boot on the stack */
719 /* Kick the second */
720 apic_write_around(APIC_ICR
, APIC_DM_STARTUP
721 | (start_eip
>> 12));
724 * Give the other CPU some time to accept the IPI.
728 Dprintk("Startup point 1.\n");
730 Dprintk("Waiting for send to finish...\n");
731 send_status
= safe_apic_wait_icr_idle();
734 * Give the other CPU some time to accept the IPI.
738 * Due to the Pentium erratum 3AP.
741 apic_read_around(APIC_SPIV
);
742 apic_write(APIC_ESR
, 0);
744 accept_status
= (apic_read(APIC_ESR
) & 0xEF);
745 if (send_status
|| accept_status
)
748 Dprintk("After Startup.\n");
751 printk(KERN_ERR
"APIC never delivered???\n");
753 printk(KERN_ERR
"APIC delivery error (%lx).\n", accept_status
);
755 return (send_status
| accept_status
);
757 #endif /* WAKE_SECONDARY_VIA_INIT */
760 struct work_struct work
;
761 struct task_struct
*idle
;
762 struct completion done
;
766 static void __cpuinit
do_fork_idle(struct work_struct
*work
)
768 struct create_idle
*c_idle
=
769 container_of(work
, struct create_idle
, work
);
771 c_idle
->idle
= fork_idle(c_idle
->cpu
);
772 complete(&c_idle
->done
);
777 * Allocate node local memory for the AP pda.
779 * Must be called after the _cpu_pda pointer table is initialized.
781 static int __cpuinit
get_local_pda(int cpu
)
783 struct x8664_pda
*oldpda
, *newpda
;
784 unsigned long size
= sizeof(struct x8664_pda
);
785 int node
= cpu_to_node(cpu
);
787 if (cpu_pda(cpu
) && !cpu_pda(cpu
)->in_bootmem
)
790 oldpda
= cpu_pda(cpu
);
791 newpda
= kmalloc_node(size
, GFP_ATOMIC
, node
);
793 printk(KERN_ERR
"Could not allocate node local PDA "
794 "for CPU %d on node %d\n", cpu
, node
);
797 return 0; /* have a usable pda */
803 memcpy(newpda
, oldpda
, size
);
805 free_bootmem((unsigned long)oldpda
, size
);
808 newpda
->in_bootmem
= 0;
809 cpu_pda(cpu
) = newpda
;
812 #endif /* CONFIG_X86_64 */
814 static int __cpuinit
do_boot_cpu(int apicid
, int cpu
)
816 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
817 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
818 * Returns zero if CPU booted OK, else error code from wakeup_secondary_cpu.
821 unsigned long boot_error
= 0;
823 unsigned long start_ip
;
824 unsigned short nmi_high
= 0, nmi_low
= 0;
825 struct create_idle c_idle
= {
827 .done
= COMPLETION_INITIALIZER_ONSTACK(c_idle
.done
),
829 INIT_WORK(&c_idle
.work
, do_fork_idle
);
832 /* Allocate node local memory for AP pdas */
834 boot_error
= get_local_pda(cpu
);
837 /* if can't get pda memory, can't start cpu */
841 alternatives_smp_switch(1);
843 c_idle
.idle
= get_idle_for_cpu(cpu
);
846 * We can't use kernel_thread since we must avoid to
847 * reschedule the child.
850 c_idle
.idle
->thread
.sp
= (unsigned long) (((struct pt_regs
*)
851 (THREAD_SIZE
+ task_stack_page(c_idle
.idle
))) - 1);
852 init_idle(c_idle
.idle
, cpu
);
856 if (!keventd_up() || current_is_keventd())
857 c_idle
.work
.func(&c_idle
.work
);
859 schedule_work(&c_idle
.work
);
860 wait_for_completion(&c_idle
.done
);
863 if (IS_ERR(c_idle
.idle
)) {
864 printk("failed fork for CPU %d\n", cpu
);
865 return PTR_ERR(c_idle
.idle
);
868 set_idle_for_cpu(cpu
, c_idle
.idle
);
871 per_cpu(current_task
, cpu
) = c_idle
.idle
;
873 /* Stack for startup_32 can be just as for start_secondary onwards */
876 cpu_pda(cpu
)->pcurrent
= c_idle
.idle
;
877 load_sp0(&per_cpu(init_tss
, cpu
), &c_idle
.idle
->thread
);
878 clear_tsk_thread_flag(c_idle
.idle
, TIF_FORK
);
880 early_gdt_descr
.address
= (unsigned long)get_cpu_gdt_table(cpu
);
881 initial_code
= (unsigned long)start_secondary
;
882 stack_start
.sp
= (void *) c_idle
.idle
->thread
.sp
;
884 /* start_ip had better be page-aligned! */
885 start_ip
= setup_trampoline();
887 /* So we see what's up */
888 printk(KERN_INFO
"Booting processor %d/%d ip %lx\n",
889 cpu
, apicid
, start_ip
);
892 * This grunge runs the startup process for
893 * the targeted processor.
896 atomic_set(&init_deasserted
, 0);
898 if (get_uv_system_type() != UV_NON_UNIQUE_APIC
) {
900 Dprintk("Setting warm reset code and vector.\n");
902 store_NMI_vector(&nmi_high
, &nmi_low
);
904 smpboot_setup_warm_reset_vector(start_ip
);
906 * Be paranoid about clearing APIC errors.
908 apic_write(APIC_ESR
, 0);
913 * Starting actual IPI sequence...
915 boot_error
= wakeup_secondary_cpu(apicid
, start_ip
);
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(KERN_INFO
"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(KERN_ERR
"Stuck ??\n");
947 /* trampoline code not run */
948 printk(KERN_ERR
"Not responding.\n");
949 if (get_uv_system_type() != UV_NON_UNIQUE_APIC
)
950 inquire_remote_apic(apicid
);
957 /* Try to put things back the way they were before ... */
958 unmap_cpu_to_logical_apicid(cpu
);
960 numa_remove_cpu(cpu
); /* was set by numa_add_cpu */
962 cpu_clear(cpu
, cpu_callout_map
); /* was set by do_boot_cpu() */
963 cpu_clear(cpu
, cpu_initialized
); /* was set by cpu_init() */
964 cpu_clear(cpu
, cpu_present_map
);
965 per_cpu(x86_cpu_to_apicid
, cpu
) = BAD_APICID
;
968 /* mark "stuck" area as not stuck */
969 *((volatile unsigned long *)trampoline_base
) = 0;
972 * Cleanup possible dangling ends...
974 smpboot_restore_warm_reset_vector();
979 int __cpuinit
native_cpu_up(unsigned int cpu
)
981 int apicid
= cpu_present_to_apicid(cpu
);
985 WARN_ON(irqs_disabled());
987 Dprintk("++++++++++++++++++++=_---CPU UP %u\n", cpu
);
989 if (apicid
== BAD_APICID
|| apicid
== boot_cpu_physical_apicid
||
990 !physid_isset(apicid
, phys_cpu_present_map
)) {
991 printk(KERN_ERR
"%s: bad cpu %d\n", __func__
, cpu
);
996 * Already booted CPU?
998 if (cpu_isset(cpu
, cpu_callin_map
)) {
999 Dprintk("do_boot_cpu %d Already started\n", cpu
);
1004 * Save current MTRR state in case it was changed since early boot
1005 * (e.g. by the ACPI SMI) to initialize new CPUs with MTRRs in sync:
1009 per_cpu(cpu_state
, cpu
) = CPU_UP_PREPARE
;
1011 #ifdef CONFIG_X86_32
1012 /* init low mem mapping */
1013 clone_pgd_range(swapper_pg_dir
, swapper_pg_dir
+ KERNEL_PGD_BOUNDARY
,
1014 min_t(unsigned long, KERNEL_PGD_PTRS
, KERNEL_PGD_BOUNDARY
));
1018 err
= do_boot_cpu(apicid
, cpu
);
1023 err
= do_boot_cpu(apicid
, cpu
);
1026 Dprintk("do_boot_cpu failed %d\n", err
);
1031 * Check TSC synchronization with the AP (keep irqs disabled
1034 local_irq_save(flags
);
1035 check_tsc_sync_source(cpu
);
1036 local_irq_restore(flags
);
1038 while (!cpu_online(cpu
)) {
1040 touch_nmi_watchdog();
1047 * Fall back to non SMP mode after errors.
1049 * RED-PEN audit/test this more. I bet there is more state messed up here.
1051 static __init
void disable_smp(void)
1053 cpu_present_map
= cpumask_of_cpu(0);
1054 cpu_possible_map
= cpumask_of_cpu(0);
1055 smpboot_clear_io_apic_irqs();
1057 if (smp_found_config
)
1058 physid_set_mask_of_physid(boot_cpu_physical_apicid
, &phys_cpu_present_map
);
1060 physid_set_mask_of_physid(0, &phys_cpu_present_map
);
1061 map_cpu_to_logical_apicid();
1062 cpu_set(0, per_cpu(cpu_sibling_map
, 0));
1063 cpu_set(0, per_cpu(cpu_core_map
, 0));
1067 * Various sanity checks.
1069 static int __init
smp_sanity_check(unsigned max_cpus
)
1072 if (!physid_isset(hard_smp_processor_id(), phys_cpu_present_map
)) {
1073 printk(KERN_WARNING
"weird, boot CPU (#%d) not listed"
1074 "by the BIOS.\n", hard_smp_processor_id());
1075 physid_set(hard_smp_processor_id(), phys_cpu_present_map
);
1079 * If we couldn't find an SMP configuration at boot time,
1080 * get out of here now!
1082 if (!smp_found_config
&& !acpi_lapic
) {
1084 printk(KERN_NOTICE
"SMP motherboard not detected.\n");
1086 if (APIC_init_uniprocessor())
1087 printk(KERN_NOTICE
"Local APIC not detected."
1088 " Using dummy APIC emulation.\n");
1093 * Should not be necessary because the MP table should list the boot
1094 * CPU too, but we do it for the sake of robustness anyway.
1096 if (!check_phys_apicid_present(boot_cpu_physical_apicid
)) {
1098 "weird, boot CPU (#%d) not listed by the BIOS.\n",
1099 boot_cpu_physical_apicid
);
1100 physid_set(hard_smp_processor_id(), phys_cpu_present_map
);
1105 * If we couldn't find a local APIC, then get out of here now!
1107 if (APIC_INTEGRATED(apic_version
[boot_cpu_physical_apicid
]) &&
1109 printk(KERN_ERR
"BIOS bug, local APIC #%d not detected!...\n",
1110 boot_cpu_physical_apicid
);
1111 printk(KERN_ERR
"... forcing use of dummy APIC emulation."
1112 "(tell your hw vendor)\n");
1113 smpboot_clear_io_apic();
1117 verify_local_APIC();
1120 * If SMP should be disabled, then really disable it!
1123 printk(KERN_INFO
"SMP mode deactivated.\n");
1124 smpboot_clear_io_apic();
1126 localise_nmi_watchdog();
1128 #ifdef CONFIG_X86_32
1132 end_local_APIC_setup();
1139 static void __init
smp_cpu_index_default(void)
1142 struct cpuinfo_x86
*c
;
1144 for_each_possible_cpu(i
) {
1146 /* mark all to hotplug */
1147 c
->cpu_index
= NR_CPUS
;
1152 * Prepare for SMP bootup. The MP table or ACPI has been read
1153 * earlier. Just do some sanity checking here and enable APIC mode.
1155 void __init
native_smp_prepare_cpus(unsigned int max_cpus
)
1158 nmi_watchdog_default();
1159 smp_cpu_index_default();
1160 current_cpu_data
= boot_cpu_data
;
1161 cpu_callin_map
= cpumask_of_cpu(0);
1164 * Setup boot CPU information
1166 smp_store_cpu_info(0); /* Final full version of the data */
1167 boot_cpu_logical_apicid
= logical_smp_processor_id();
1168 current_thread_info()->cpu
= 0; /* needed? */
1169 set_cpu_sibling_map(0);
1171 if (smp_sanity_check(max_cpus
) < 0) {
1172 printk(KERN_INFO
"SMP disabled\n");
1178 if (GET_APIC_ID(read_apic_id()) != boot_cpu_physical_apicid
) {
1179 panic("Boot APIC ID in local APIC unexpected (%d vs %d)",
1180 GET_APIC_ID(read_apic_id()), boot_cpu_physical_apicid
);
1181 /* Or can we switch back to PIC here? */
1185 #ifdef CONFIG_X86_32
1189 * Switch from PIC to APIC mode.
1193 #ifdef CONFIG_X86_64
1195 * Enable IO APIC before setting up error vector
1197 if (!skip_ioapic_setup
&& nr_ioapics
)
1200 end_local_APIC_setup();
1202 map_cpu_to_logical_apicid();
1204 setup_portio_remap();
1206 smpboot_setup_io_apic();
1208 * Set up local APIC timer on boot CPU.
1211 printk(KERN_INFO
"CPU%d: ", 0);
1212 print_cpu_info(&cpu_data(0));
1218 * Early setup to make printk work.
1220 void __init
native_smp_prepare_boot_cpu(void)
1222 int me
= smp_processor_id();
1223 #ifdef CONFIG_X86_32
1226 switch_to_new_gdt();
1227 /* already set me in cpu_online_map in boot_cpu_init() */
1228 cpu_set(me
, cpu_callout_map
);
1229 per_cpu(cpu_state
, me
) = CPU_ONLINE
;
1232 void __init
native_smp_cpus_done(unsigned int max_cpus
)
1234 Dprintk("Boot done.\n");
1238 #ifdef CONFIG_X86_IO_APIC
1239 setup_ioapic_dest();
1241 check_nmi_watchdog();
1244 #ifdef CONFIG_HOTPLUG_CPU
1246 # ifdef CONFIG_X86_32
1247 void cpu_exit_clear(void)
1249 int cpu
= raw_smp_processor_id();
1256 cpu_clear(cpu
, cpu_callout_map
);
1257 cpu_clear(cpu
, cpu_callin_map
);
1259 unmap_cpu_to_logical_apicid(cpu
);
1261 # endif /* CONFIG_X86_32 */
1263 static void remove_siblinginfo(int cpu
)
1266 struct cpuinfo_x86
*c
= &cpu_data(cpu
);
1268 for_each_cpu_mask(sibling
, per_cpu(cpu_core_map
, cpu
)) {
1269 cpu_clear(cpu
, per_cpu(cpu_core_map
, sibling
));
1271 * last thread sibling in this cpu core going down
1273 if (cpus_weight(per_cpu(cpu_sibling_map
, cpu
)) == 1)
1274 cpu_data(sibling
).booted_cores
--;
1277 for_each_cpu_mask(sibling
, per_cpu(cpu_sibling_map
, cpu
))
1278 cpu_clear(cpu
, per_cpu(cpu_sibling_map
, sibling
));
1279 cpus_clear(per_cpu(cpu_sibling_map
, cpu
));
1280 cpus_clear(per_cpu(cpu_core_map
, cpu
));
1281 c
->phys_proc_id
= 0;
1283 cpu_clear(cpu
, cpu_sibling_setup_map
);
1286 static int additional_cpus __initdata
= -1;
1288 static __init
int setup_additional_cpus(char *s
)
1290 return s
&& get_option(&s
, &additional_cpus
) ? 0 : -EINVAL
;
1292 early_param("additional_cpus", setup_additional_cpus
);
1295 * cpu_possible_map should be static, it cannot change as cpu's
1296 * are onlined, or offlined. The reason is per-cpu data-structures
1297 * are allocated by some modules at init time, and dont expect to
1298 * do this dynamically on cpu arrival/departure.
1299 * cpu_present_map on the other hand can change dynamically.
1300 * In case when cpu_hotplug is not compiled, then we resort to current
1301 * behaviour, which is cpu_possible == cpu_present.
1304 * Three ways to find out the number of additional hotplug CPUs:
1305 * - If the BIOS specified disabled CPUs in ACPI/mptables use that.
1306 * - The user can overwrite it with additional_cpus=NUM
1307 * - Otherwise don't reserve additional CPUs.
1308 * We do this because additional CPUs waste a lot of memory.
1311 __init
void prefill_possible_map(void)
1316 if (additional_cpus
== -1) {
1317 if (disabled_cpus
> 0)
1318 additional_cpus
= disabled_cpus
;
1320 additional_cpus
= 0;
1322 possible
= num_processors
+ additional_cpus
;
1323 if (possible
> NR_CPUS
)
1326 printk(KERN_INFO
"SMP: Allowing %d CPUs, %d hotplug CPUs\n",
1327 possible
, max_t(int, possible
- num_processors
, 0));
1329 for (i
= 0; i
< possible
; i
++)
1330 cpu_set(i
, cpu_possible_map
);
1332 nr_cpu_ids
= possible
;
1335 static void __ref
remove_cpu_from_maps(int cpu
)
1337 cpu_clear(cpu
, cpu_online_map
);
1338 #ifdef CONFIG_X86_64
1339 cpu_clear(cpu
, cpu_callout_map
);
1340 cpu_clear(cpu
, cpu_callin_map
);
1341 /* was set by cpu_init() */
1342 clear_bit(cpu
, (unsigned long *)&cpu_initialized
);
1343 numa_remove_cpu(cpu
);
1347 int __cpu_disable(void)
1349 int cpu
= smp_processor_id();
1352 * Perhaps use cpufreq to drop frequency, but that could go
1353 * into generic code.
1355 * We won't take down the boot processor on i386 due to some
1356 * interrupts only being able to be serviced by the BSP.
1357 * Especially so if we're not using an IOAPIC -zwane
1362 if (nmi_watchdog
== NMI_LOCAL_APIC
)
1363 stop_apic_nmi_watchdog(NULL
);
1368 * Allow any queued timer interrupts to get serviced
1369 * This is only a temporary solution until we cleanup
1370 * fixup_irqs as we do for IA64.
1375 local_irq_disable();
1376 remove_siblinginfo(cpu
);
1378 /* It's now safe to remove this processor from the online map */
1379 remove_cpu_from_maps(cpu
);
1380 fixup_irqs(cpu_online_map
);
1384 void __cpu_die(unsigned int cpu
)
1386 /* We don't do anything here: idle task is faking death itself. */
1389 for (i
= 0; i
< 10; i
++) {
1390 /* They ack this in play_dead by setting CPU_DEAD */
1391 if (per_cpu(cpu_state
, cpu
) == CPU_DEAD
) {
1392 printk(KERN_INFO
"CPU %d is now offline\n", cpu
);
1393 if (1 == num_online_cpus())
1394 alternatives_smp_switch(0);
1399 printk(KERN_ERR
"CPU %u didn't die...\n", cpu
);
1401 #else /* ... !CONFIG_HOTPLUG_CPU */
1402 int __cpu_disable(void)
1407 void __cpu_die(unsigned int cpu
)
1409 /* We said "no" in __cpu_disable */
1415 * If the BIOS enumerates physical processors before logical,
1416 * maxcpus=N at enumeration-time can be used to disable HT.
1418 static int __init
parse_maxcpus(char *arg
)
1420 extern unsigned int maxcpus
;
1422 maxcpus
= simple_strtoul(arg
, NULL
, 0);
1425 early_param("maxcpus", parse_maxcpus
);