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>
49 #include <linux/delay.h>
50 #include <linux/mc146818rtc.h>
51 #include <asm/tlbflush.h>
53 #include <asm/arch_hooks.h>
56 #include <mach_apic.h>
57 #include <mach_wakecpu.h>
58 #include <smpboot_hooks.h>
60 /* Set if we find a B stepping CPU */
61 static int __devinitdata smp_b_stepping
;
63 /* Number of siblings per CPU package */
64 int smp_num_siblings
= 1;
66 EXPORT_SYMBOL(smp_num_siblings
);
69 /* Last level cache ID of each logical CPU */
70 int cpu_llc_id
[NR_CPUS
] __cpuinitdata
= {[0 ... NR_CPUS
-1] = BAD_APICID
};
72 /* representing HT siblings of each logical CPU */
73 cpumask_t cpu_sibling_map
[NR_CPUS
] __read_mostly
;
74 EXPORT_SYMBOL(cpu_sibling_map
);
76 /* representing HT and core siblings of each logical CPU */
77 cpumask_t cpu_core_map
[NR_CPUS
] __read_mostly
;
78 EXPORT_SYMBOL(cpu_core_map
);
80 /* bitmap of online cpus */
81 cpumask_t cpu_online_map __read_mostly
;
82 EXPORT_SYMBOL(cpu_online_map
);
84 cpumask_t cpu_callin_map
;
85 cpumask_t cpu_callout_map
;
86 EXPORT_SYMBOL(cpu_callout_map
);
87 cpumask_t cpu_possible_map
;
88 EXPORT_SYMBOL(cpu_possible_map
);
89 static cpumask_t smp_commenced_mask
;
91 /* TSC's upper 32 bits can't be written in eariler CPU (before prescott), there
92 * is no way to resync one AP against BP. TBD: for prescott and above, we
93 * should use IA64's algorithm
95 static int __devinitdata tsc_sync_disabled
;
97 /* Per CPU bogomips and other parameters */
98 struct cpuinfo_x86 cpu_data
[NR_CPUS
] __cacheline_aligned
;
99 EXPORT_SYMBOL(cpu_data
);
101 u8 x86_cpu_to_apicid
[NR_CPUS
] __read_mostly
=
102 { [0 ... NR_CPUS
-1] = 0xff };
103 EXPORT_SYMBOL(x86_cpu_to_apicid
);
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 __devinit
smp_store_cpu_info(int id
)
157 struct cpuinfo_x86
*c
= cpu_data
+ id
;
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
);
212 * TSC synchronization.
214 * We first check whether all CPUs have their TSC's synchronized,
215 * then we print a warning if not, and always resync.
220 atomic_t count_start
;
222 unsigned long long values
[NR_CPUS
];
224 .start_flag
= ATOMIC_INIT(0),
225 .count_start
= ATOMIC_INIT(0),
226 .count_stop
= ATOMIC_INIT(0),
231 static void __init
synchronize_tsc_bp(void)
234 unsigned long long t0
;
235 unsigned long long sum
, avg
;
237 unsigned int one_usec
;
240 printk(KERN_INFO
"checking TSC synchronization across %u CPUs: ", num_booting_cpus());
242 /* convert from kcyc/sec to cyc/usec */
243 one_usec
= cpu_khz
/ 1000;
245 atomic_set(&tsc
.start_flag
, 1);
249 * We loop a few times to get a primed instruction cache,
250 * then the last pass is more or less synchronized and
251 * the BP and APs set their cycle counters to zero all at
252 * once. This reduces the chance of having random offsets
253 * between the processors, and guarantees that the maximum
254 * delay between the cycle counters is never bigger than
255 * the latency of information-passing (cachelines) between
258 for (i
= 0; i
< NR_LOOPS
; i
++) {
260 * all APs synchronize but they loop on '== num_cpus'
262 while (atomic_read(&tsc
.count_start
) != num_booting_cpus()-1)
264 atomic_set(&tsc
.count_stop
, 0);
267 * this lets the APs save their current TSC:
269 atomic_inc(&tsc
.count_start
);
271 rdtscll(tsc
.values
[smp_processor_id()]);
273 * We clear the TSC in the last loop:
279 * Wait for all APs to leave the synchronization point:
281 while (atomic_read(&tsc
.count_stop
) != num_booting_cpus()-1)
283 atomic_set(&tsc
.count_start
, 0);
285 atomic_inc(&tsc
.count_stop
);
289 for (i
= 0; i
< NR_CPUS
; i
++) {
290 if (cpu_isset(i
, cpu_callout_map
)) {
296 do_div(avg
, num_booting_cpus());
298 for (i
= 0; i
< NR_CPUS
; i
++) {
299 if (!cpu_isset(i
, cpu_callout_map
))
301 delta
= tsc
.values
[i
] - avg
;
305 * We report bigger than 2 microseconds clock differences.
307 if (delta
> 2*one_usec
) {
315 do_div(realdelta
, one_usec
);
316 if (tsc
.values
[i
] < avg
)
317 realdelta
= -realdelta
;
320 printk(KERN_INFO
"CPU#%d had %Ld usecs TSC "
321 "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 * And for power savings, we return cpu_core_map
459 if (sched_mc_power_savings
|| sched_smt_power_savings
)
460 return cpu_core_map
[cpu
];
462 return c
->llc_shared_map
;
465 /* representing cpus for which sibling maps can be computed */
466 static cpumask_t cpu_sibling_setup_map
;
469 set_cpu_sibling_map(int cpu
)
472 struct cpuinfo_x86
*c
= cpu_data
;
474 cpu_set(cpu
, cpu_sibling_setup_map
);
476 if (smp_num_siblings
> 1) {
477 for_each_cpu_mask(i
, cpu_sibling_setup_map
) {
478 if (c
[cpu
].phys_proc_id
== c
[i
].phys_proc_id
&&
479 c
[cpu
].cpu_core_id
== c
[i
].cpu_core_id
) {
480 cpu_set(i
, cpu_sibling_map
[cpu
]);
481 cpu_set(cpu
, cpu_sibling_map
[i
]);
482 cpu_set(i
, cpu_core_map
[cpu
]);
483 cpu_set(cpu
, cpu_core_map
[i
]);
484 cpu_set(i
, c
[cpu
].llc_shared_map
);
485 cpu_set(cpu
, c
[i
].llc_shared_map
);
489 cpu_set(cpu
, cpu_sibling_map
[cpu
]);
492 cpu_set(cpu
, c
[cpu
].llc_shared_map
);
494 if (current_cpu_data
.x86_max_cores
== 1) {
495 cpu_core_map
[cpu
] = cpu_sibling_map
[cpu
];
496 c
[cpu
].booted_cores
= 1;
500 for_each_cpu_mask(i
, cpu_sibling_setup_map
) {
501 if (cpu_llc_id
[cpu
] != BAD_APICID
&&
502 cpu_llc_id
[cpu
] == cpu_llc_id
[i
]) {
503 cpu_set(i
, c
[cpu
].llc_shared_map
);
504 cpu_set(cpu
, c
[i
].llc_shared_map
);
506 if (c
[cpu
].phys_proc_id
== c
[i
].phys_proc_id
) {
507 cpu_set(i
, cpu_core_map
[cpu
]);
508 cpu_set(cpu
, cpu_core_map
[i
]);
510 * Does this new cpu bringup a new core?
512 if (cpus_weight(cpu_sibling_map
[cpu
]) == 1) {
514 * for each core in package, increment
515 * the booted_cores for this new cpu
517 if (first_cpu(cpu_sibling_map
[i
]) == i
)
518 c
[cpu
].booted_cores
++;
520 * increment the core count for all
521 * the other cpus in this package
525 } else if (i
!= cpu
&& !c
[cpu
].booted_cores
)
526 c
[cpu
].booted_cores
= c
[i
].booted_cores
;
532 * Activate a secondary processor.
534 static void __devinit
start_secondary(void *unused
)
537 * Dont put anything before smp_callin(), SMP
538 * booting is too fragile that we want to limit the
539 * things done here to the most necessary things.
544 while (!cpu_isset(smp_processor_id(), smp_commenced_mask
))
546 setup_secondary_APIC_clock();
547 if (nmi_watchdog
== NMI_IO_APIC
) {
548 disable_8259A_irq(0);
549 enable_NMI_through_LVT0(NULL
);
554 * low-memory mappings have been cleared, flush them from
555 * the local TLBs too.
559 /* This must be done before setting cpu_online_map */
560 set_cpu_sibling_map(raw_smp_processor_id());
564 * We need to hold call_lock, so there is no inconsistency
565 * between the time smp_call_function() determines number of
566 * IPI receipients, and the time when the determination is made
567 * for which cpus receive the IPI. Holding this
568 * lock helps us to not include this cpu in a currently in progress
569 * smp_call_function().
571 lock_ipi_call_lock();
572 cpu_set(smp_processor_id(), cpu_online_map
);
573 unlock_ipi_call_lock();
574 per_cpu(cpu_state
, smp_processor_id()) = CPU_ONLINE
;
576 /* We can take interrupts now: we're officially "up". */
584 * Everything has been set up for the secondary
585 * CPUs - they just need to reload everything
586 * from the task structure
587 * This function must not return.
589 void __devinit
initialize_secondary(void)
592 * We don't actually need to load the full TSS,
593 * basically just the stack pointer and the eip.
600 :"r" (current
->thread
.esp
),"r" (current
->thread
.eip
));
610 /* which logical CPUs are on which nodes */
611 cpumask_t node_2_cpu_mask
[MAX_NUMNODES
] __read_mostly
=
612 { [0 ... MAX_NUMNODES
-1] = CPU_MASK_NONE
};
613 /* which node each logical CPU is on */
614 int cpu_2_node
[NR_CPUS
] __read_mostly
= { [0 ... NR_CPUS
-1] = 0 };
615 EXPORT_SYMBOL(cpu_2_node
);
617 /* set up a mapping between cpu and node. */
618 static inline void map_cpu_to_node(int cpu
, int node
)
620 printk("Mapping cpu %d to node %d\n", cpu
, node
);
621 cpu_set(cpu
, node_2_cpu_mask
[node
]);
622 cpu_2_node
[cpu
] = node
;
625 /* undo a mapping between cpu and node. */
626 static inline void unmap_cpu_to_node(int cpu
)
630 printk("Unmapping cpu %d from all nodes\n", cpu
);
631 for (node
= 0; node
< MAX_NUMNODES
; node
++)
632 cpu_clear(cpu
, node_2_cpu_mask
[node
]);
635 #else /* !CONFIG_NUMA */
637 #define map_cpu_to_node(cpu, node) ({})
638 #define unmap_cpu_to_node(cpu) ({})
640 #endif /* CONFIG_NUMA */
642 u8 cpu_2_logical_apicid
[NR_CPUS
] __read_mostly
= { [0 ... NR_CPUS
-1] = BAD_APICID
};
644 static void map_cpu_to_logical_apicid(void)
646 int cpu
= smp_processor_id();
647 int apicid
= logical_smp_processor_id();
648 int node
= apicid_to_node(apicid
);
650 if (!node_online(node
))
651 node
= first_online_node
;
653 cpu_2_logical_apicid
[cpu
] = apicid
;
654 map_cpu_to_node(cpu
, node
);
657 static void unmap_cpu_to_logical_apicid(int cpu
)
659 cpu_2_logical_apicid
[cpu
] = BAD_APICID
;
660 unmap_cpu_to_node(cpu
);
664 static inline void __inquire_remote_apic(int apicid
)
666 int i
, regs
[] = { APIC_ID
>> 4, APIC_LVR
>> 4, APIC_SPIV
>> 4 };
667 char *names
[] = { "ID", "VERSION", "SPIV" };
670 printk("Inquiring remote APIC #%d...\n", apicid
);
672 for (i
= 0; i
< ARRAY_SIZE(regs
); i
++) {
673 printk("... APIC #%d %s: ", apicid
, names
[i
]);
678 apic_wait_icr_idle();
680 apic_write_around(APIC_ICR2
, SET_APIC_DEST_FIELD(apicid
));
681 apic_write_around(APIC_ICR
, APIC_DM_REMRD
| regs
[i
]);
686 status
= apic_read(APIC_ICR
) & APIC_ICR_RR_MASK
;
687 } while (status
== APIC_ICR_RR_INPROG
&& timeout
++ < 1000);
690 case APIC_ICR_RR_VALID
:
691 status
= apic_read(APIC_RRR
);
692 printk("%08x\n", status
);
701 #ifdef WAKE_SECONDARY_VIA_NMI
703 * Poke the other CPU in the eye via NMI to wake it up. Remember that the normal
704 * INIT, INIT, STARTUP sequence will reset the chip hard for us, and this
705 * won't ... remember to clear down the APIC, etc later.
708 wakeup_secondary_cpu(int logical_apicid
, unsigned long start_eip
)
710 unsigned long send_status
= 0, accept_status
= 0;
714 apic_write_around(APIC_ICR2
, SET_APIC_DEST_FIELD(logical_apicid
));
716 /* Boot on the stack */
717 /* Kick the second */
718 apic_write_around(APIC_ICR
, APIC_DM_NMI
| APIC_DEST_LOGICAL
);
720 Dprintk("Waiting for send to finish...\n");
725 send_status
= apic_read(APIC_ICR
) & APIC_ICR_BUSY
;
726 } while (send_status
&& (timeout
++ < 1000));
729 * Give the other CPU some time to accept the IPI.
733 * Due to the Pentium erratum 3AP.
735 maxlvt
= get_maxlvt();
737 apic_read_around(APIC_SPIV
);
738 apic_write(APIC_ESR
, 0);
740 accept_status
= (apic_read(APIC_ESR
) & 0xEF);
741 Dprintk("NMI sent.\n");
744 printk("APIC never delivered???\n");
746 printk("APIC delivery error (%lx).\n", accept_status
);
748 return (send_status
| accept_status
);
750 #endif /* WAKE_SECONDARY_VIA_NMI */
752 #ifdef WAKE_SECONDARY_VIA_INIT
754 wakeup_secondary_cpu(int phys_apicid
, unsigned long start_eip
)
756 unsigned long send_status
= 0, accept_status
= 0;
757 int maxlvt
, timeout
, num_starts
, j
;
760 * Be paranoid about clearing APIC errors.
762 if (APIC_INTEGRATED(apic_version
[phys_apicid
])) {
763 apic_read_around(APIC_SPIV
);
764 apic_write(APIC_ESR
, 0);
768 Dprintk("Asserting INIT.\n");
771 * Turn INIT on target chip
773 apic_write_around(APIC_ICR2
, SET_APIC_DEST_FIELD(phys_apicid
));
778 apic_write_around(APIC_ICR
, APIC_INT_LEVELTRIG
| APIC_INT_ASSERT
781 Dprintk("Waiting for send to finish...\n");
786 send_status
= apic_read(APIC_ICR
) & APIC_ICR_BUSY
;
787 } while (send_status
&& (timeout
++ < 1000));
791 Dprintk("Deasserting INIT.\n");
794 apic_write_around(APIC_ICR2
, SET_APIC_DEST_FIELD(phys_apicid
));
797 apic_write_around(APIC_ICR
, APIC_INT_LEVELTRIG
| APIC_DM_INIT
);
799 Dprintk("Waiting for send to finish...\n");
804 send_status
= apic_read(APIC_ICR
) & APIC_ICR_BUSY
;
805 } while (send_status
&& (timeout
++ < 1000));
807 atomic_set(&init_deasserted
, 1);
810 * Should we send STARTUP IPIs ?
812 * Determine this based on the APIC version.
813 * If we don't have an integrated APIC, don't send the STARTUP IPIs.
815 if (APIC_INTEGRATED(apic_version
[phys_apicid
]))
821 * Run STARTUP IPI loop.
823 Dprintk("#startup loops: %d.\n", num_starts
);
825 maxlvt
= get_maxlvt();
827 for (j
= 1; j
<= num_starts
; j
++) {
828 Dprintk("Sending STARTUP #%d.\n",j
);
829 apic_read_around(APIC_SPIV
);
830 apic_write(APIC_ESR
, 0);
832 Dprintk("After apic_write.\n");
839 apic_write_around(APIC_ICR2
, SET_APIC_DEST_FIELD(phys_apicid
));
841 /* Boot on the stack */
842 /* Kick the second */
843 apic_write_around(APIC_ICR
, APIC_DM_STARTUP
844 | (start_eip
>> 12));
847 * Give the other CPU some time to accept the IPI.
851 Dprintk("Startup point 1.\n");
853 Dprintk("Waiting for send to finish...\n");
858 send_status
= apic_read(APIC_ICR
) & APIC_ICR_BUSY
;
859 } while (send_status
&& (timeout
++ < 1000));
862 * Give the other CPU some time to accept the IPI.
866 * Due to the Pentium erratum 3AP.
869 apic_read_around(APIC_SPIV
);
870 apic_write(APIC_ESR
, 0);
872 accept_status
= (apic_read(APIC_ESR
) & 0xEF);
873 if (send_status
|| accept_status
)
876 Dprintk("After Startup.\n");
879 printk("APIC never delivered???\n");
881 printk("APIC delivery error (%lx).\n", accept_status
);
883 return (send_status
| accept_status
);
885 #endif /* WAKE_SECONDARY_VIA_INIT */
887 extern cpumask_t cpu_initialized
;
888 static inline int alloc_cpu_id(void)
892 cpus_complement(tmp_map
, cpu_present_map
);
893 cpu
= first_cpu(tmp_map
);
899 #ifdef CONFIG_HOTPLUG_CPU
900 static struct task_struct
* __devinitdata cpu_idle_tasks
[NR_CPUS
];
901 static inline struct task_struct
* alloc_idle_task(int cpu
)
903 struct task_struct
*idle
;
905 if ((idle
= cpu_idle_tasks
[cpu
]) != NULL
) {
906 /* initialize thread_struct. we really want to avoid destroy
909 idle
->thread
.esp
= (unsigned long)task_pt_regs(idle
);
910 init_idle(idle
, cpu
);
913 idle
= fork_idle(cpu
);
916 cpu_idle_tasks
[cpu
] = idle
;
920 #define alloc_idle_task(cpu) fork_idle(cpu)
923 static int __devinit
do_boot_cpu(int apicid
, int cpu
)
925 * NOTE - on most systems this is a PHYSICAL apic ID, but on multiquad
926 * (ie clustered apic addressing mode), this is a LOGICAL apic ID.
927 * Returns zero if CPU booted OK, else error code from wakeup_secondary_cpu.
930 struct task_struct
*idle
;
931 unsigned long boot_error
;
933 unsigned long start_eip
;
934 unsigned short nmi_high
= 0, nmi_low
= 0;
937 alternatives_smp_switch(1);
940 * We can't use kernel_thread since we must avoid to
941 * reschedule the child.
943 idle
= alloc_idle_task(cpu
);
945 panic("failed fork for CPU %d", cpu
);
946 idle
->thread
.eip
= (unsigned long) start_secondary
;
947 /* start_eip had better be page-aligned! */
948 start_eip
= setup_trampoline();
950 /* So we see what's up */
951 printk("Booting processor %d/%d eip %lx\n", cpu
, apicid
, start_eip
);
952 /* Stack for startup_32 can be just as for start_secondary onwards */
953 stack_start
.esp
= (void *) idle
->thread
.esp
;
958 * This grunge runs the startup process for
959 * the targeted processor.
962 atomic_set(&init_deasserted
, 0);
964 Dprintk("Setting warm reset code and vector.\n");
966 store_NMI_vector(&nmi_high
, &nmi_low
);
968 smpboot_setup_warm_reset_vector(start_eip
);
971 * Starting actual IPI sequence...
973 boot_error
= wakeup_secondary_cpu(apicid
, start_eip
);
977 * allow APs to start initializing.
979 Dprintk("Before Callout %d.\n", cpu
);
980 cpu_set(cpu
, cpu_callout_map
);
981 Dprintk("After Callout %d.\n", cpu
);
984 * Wait 5s total for a response
986 for (timeout
= 0; timeout
< 50000; timeout
++) {
987 if (cpu_isset(cpu
, cpu_callin_map
))
988 break; /* It has booted */
992 if (cpu_isset(cpu
, cpu_callin_map
)) {
993 /* number CPUs logically, starting from 1 (BSP is 0) */
995 printk("CPU%d: ", cpu
);
996 print_cpu_info(&cpu_data
[cpu
]);
997 Dprintk("CPU has booted.\n");
1000 if (*((volatile unsigned char *)trampoline_base
)
1002 /* trampoline started but...? */
1003 printk("Stuck ??\n");
1005 /* trampoline code not run */
1006 printk("Not responding.\n");
1007 inquire_remote_apic(apicid
);
1012 /* Try to put things back the way they were before ... */
1013 unmap_cpu_to_logical_apicid(cpu
);
1014 cpu_clear(cpu
, cpu_callout_map
); /* was set here (do_boot_cpu()) */
1015 cpu_clear(cpu
, cpu_initialized
); /* was set by cpu_init() */
1018 x86_cpu_to_apicid
[cpu
] = apicid
;
1019 cpu_set(cpu
, cpu_present_map
);
1022 /* mark "stuck" area as not stuck */
1023 *((volatile unsigned long *)trampoline_base
) = 0;
1028 #ifdef CONFIG_HOTPLUG_CPU
1029 void cpu_exit_clear(void)
1031 int cpu
= raw_smp_processor_id();
1039 cpu_clear(cpu
, cpu_callout_map
);
1040 cpu_clear(cpu
, cpu_callin_map
);
1042 cpu_clear(cpu
, smp_commenced_mask
);
1043 unmap_cpu_to_logical_apicid(cpu
);
1046 struct warm_boot_cpu_info
{
1047 struct completion
*complete
;
1052 static void __cpuinit
do_warm_boot_cpu(void *p
)
1054 struct warm_boot_cpu_info
*info
= p
;
1055 do_boot_cpu(info
->apicid
, info
->cpu
);
1056 complete(info
->complete
);
1059 static int __cpuinit
__smp_prepare_cpu(int cpu
)
1061 DECLARE_COMPLETION(done
);
1062 struct warm_boot_cpu_info info
;
1063 struct work_struct task
;
1065 struct Xgt_desc_struct
*cpu_gdt_descr
= &per_cpu(cpu_gdt_descr
, cpu
);
1067 apicid
= x86_cpu_to_apicid
[cpu
];
1068 if (apicid
== BAD_APICID
) {
1074 * the CPU isn't initialized at boot time, allocate gdt table here.
1075 * cpu_init will initialize it
1077 if (!cpu_gdt_descr
->address
) {
1078 cpu_gdt_descr
->address
= get_zeroed_page(GFP_KERNEL
);
1079 if (!cpu_gdt_descr
->address
)
1080 printk(KERN_CRIT
"CPU%d failed to allocate GDT\n", cpu
);
1085 info
.complete
= &done
;
1086 info
.apicid
= apicid
;
1088 INIT_WORK(&task
, do_warm_boot_cpu
, &info
);
1090 tsc_sync_disabled
= 1;
1092 /* init low mem mapping */
1093 clone_pgd_range(swapper_pg_dir
, swapper_pg_dir
+ USER_PGD_PTRS
,
1096 schedule_work(&task
);
1097 wait_for_completion(&done
);
1099 tsc_sync_disabled
= 0;
1107 static void smp_tune_scheduling (void)
1109 unsigned long cachesize
; /* kB */
1110 unsigned long bandwidth
= 350; /* MB/s */
1112 * Rough estimation for SMP scheduling, this is the number of
1113 * cycles it takes for a fully memory-limited process to flush
1114 * the SMP-local cache.
1116 * (For a P5 this pretty much means we will choose another idle
1117 * CPU almost always at wakeup time (this is due to the small
1118 * L1 cache), on PIIs it's around 50-100 usecs, depending on
1124 * this basically disables processor-affinity
1125 * scheduling on SMP without a TSC.
1129 cachesize
= boot_cpu_data
.x86_cache_size
;
1130 if (cachesize
== -1) {
1131 cachesize
= 16; /* Pentiums, 2x8kB cache */
1134 max_cache_size
= cachesize
* 1024;
1139 * Cycle through the processors sending APIC IPIs to boot each.
1142 static int boot_cpu_logical_apicid
;
1143 /* Where the IO area was mapped on multiquad, always 0 otherwise */
1145 #ifdef CONFIG_X86_NUMAQ
1146 EXPORT_SYMBOL(xquad_portio
);
1149 static void __init
smp_boot_cpus(unsigned int max_cpus
)
1151 int apicid
, cpu
, bit
, kicked
;
1152 unsigned long bogosum
= 0;
1155 * Setup boot CPU information
1157 smp_store_cpu_info(0); /* Final full version of the data */
1158 printk("CPU%d: ", 0);
1159 print_cpu_info(&cpu_data
[0]);
1161 boot_cpu_physical_apicid
= GET_APIC_ID(apic_read(APIC_ID
));
1162 boot_cpu_logical_apicid
= logical_smp_processor_id();
1163 x86_cpu_to_apicid
[0] = boot_cpu_physical_apicid
;
1165 current_thread_info()->cpu
= 0;
1166 smp_tune_scheduling();
1168 set_cpu_sibling_map(0);
1171 * If we couldn't find an SMP configuration at boot time,
1172 * get out of here now!
1174 if (!smp_found_config
&& !acpi_lapic
) {
1175 printk(KERN_NOTICE
"SMP motherboard not detected.\n");
1176 smpboot_clear_io_apic_irqs();
1177 phys_cpu_present_map
= physid_mask_of_physid(0);
1178 if (APIC_init_uniprocessor())
1179 printk(KERN_NOTICE
"Local APIC not detected."
1180 " Using dummy APIC emulation.\n");
1181 map_cpu_to_logical_apicid();
1182 cpu_set(0, cpu_sibling_map
[0]);
1183 cpu_set(0, cpu_core_map
[0]);
1188 * Should not be necessary because the MP table should list the boot
1189 * CPU too, but we do it for the sake of robustness anyway.
1190 * Makes no sense to do this check in clustered apic mode, so skip it
1192 if (!check_phys_apicid_present(boot_cpu_physical_apicid
)) {
1193 printk("weird, boot CPU (#%d) not listed by the BIOS.\n",
1194 boot_cpu_physical_apicid
);
1195 physid_set(hard_smp_processor_id(), phys_cpu_present_map
);
1199 * If we couldn't find a local APIC, then get out of here now!
1201 if (APIC_INTEGRATED(apic_version
[boot_cpu_physical_apicid
]) && !cpu_has_apic
) {
1202 printk(KERN_ERR
"BIOS bug, local APIC #%d not detected!...\n",
1203 boot_cpu_physical_apicid
);
1204 printk(KERN_ERR
"... forcing use of dummy APIC emulation. (tell your hw vendor)\n");
1205 smpboot_clear_io_apic_irqs();
1206 phys_cpu_present_map
= physid_mask_of_physid(0);
1207 cpu_set(0, cpu_sibling_map
[0]);
1208 cpu_set(0, cpu_core_map
[0]);
1212 verify_local_APIC();
1215 * If SMP should be disabled, then really disable it!
1218 smp_found_config
= 0;
1219 printk(KERN_INFO
"SMP mode deactivated, forcing use of dummy APIC emulation.\n");
1220 smpboot_clear_io_apic_irqs();
1221 phys_cpu_present_map
= physid_mask_of_physid(0);
1222 cpu_set(0, cpu_sibling_map
[0]);
1223 cpu_set(0, cpu_core_map
[0]);
1229 map_cpu_to_logical_apicid();
1232 setup_portio_remap();
1235 * Scan the CPU present map and fire up the other CPUs via do_boot_cpu
1237 * In clustered apic mode, phys_cpu_present_map is a constructed thus:
1238 * bits 0-3 are quad0, 4-7 are quad1, etc. A perverse twist on the
1239 * clustered apic ID.
1241 Dprintk("CPU present map: %lx\n", physids_coerce(phys_cpu_present_map
));
1244 for (bit
= 0; kicked
< NR_CPUS
&& bit
< MAX_APICS
; bit
++) {
1245 apicid
= cpu_present_to_apicid(bit
);
1247 * Don't even attempt to start the boot CPU!
1249 if ((apicid
== boot_cpu_apicid
) || (apicid
== BAD_APICID
))
1252 if (!check_apicid_present(bit
))
1254 if (max_cpus
<= cpucount
+1)
1257 if (((cpu
= alloc_cpu_id()) <= 0) || do_boot_cpu(apicid
, cpu
))
1258 printk("CPU #%d not responding - cannot use it.\n",
1265 * Cleanup possible dangling ends...
1267 smpboot_restore_warm_reset_vector();
1270 * Allow the user to impress friends.
1272 Dprintk("Before bogomips.\n");
1273 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++)
1274 if (cpu_isset(cpu
, cpu_callout_map
))
1275 bogosum
+= cpu_data
[cpu
].loops_per_jiffy
;
1277 "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
1279 bogosum
/(500000/HZ
),
1280 (bogosum
/(5000/HZ
))%100);
1282 Dprintk("Before bogocount - setting activated=1.\n");
1285 printk(KERN_WARNING
"WARNING: SMP operation may be unreliable with B stepping processors.\n");
1288 * Don't taint if we are running SMP kernel on a single non-MP
1291 if (tainted
& TAINT_UNSAFE_SMP
) {
1293 printk (KERN_INFO
"WARNING: This combination of AMD processors is not suitable for SMP.\n");
1295 tainted
&= ~TAINT_UNSAFE_SMP
;
1298 Dprintk("Boot done.\n");
1301 * construct cpu_sibling_map[], so that we can tell sibling CPUs
1304 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
1305 cpus_clear(cpu_sibling_map
[cpu
]);
1306 cpus_clear(cpu_core_map
[cpu
]);
1309 cpu_set(0, cpu_sibling_map
[0]);
1310 cpu_set(0, cpu_core_map
[0]);
1312 smpboot_setup_io_apic();
1314 setup_boot_APIC_clock();
1317 * Synchronize the TSC with the AP
1319 if (cpu_has_tsc
&& cpucount
&& cpu_khz
)
1320 synchronize_tsc_bp();
1323 /* These are wrappers to interface to the new boot process. Someone
1324 who understands all this stuff should rewrite it properly. --RR 15/Jul/02 */
1325 void __init
smp_prepare_cpus(unsigned int max_cpus
)
1327 smp_commenced_mask
= cpumask_of_cpu(0);
1328 cpu_callin_map
= cpumask_of_cpu(0);
1330 smp_boot_cpus(max_cpus
);
1333 void __devinit
smp_prepare_boot_cpu(void)
1335 cpu_set(smp_processor_id(), cpu_online_map
);
1336 cpu_set(smp_processor_id(), cpu_callout_map
);
1337 cpu_set(smp_processor_id(), cpu_present_map
);
1338 cpu_set(smp_processor_id(), cpu_possible_map
);
1339 per_cpu(cpu_state
, smp_processor_id()) = CPU_ONLINE
;
1342 #ifdef CONFIG_HOTPLUG_CPU
1344 remove_siblinginfo(int cpu
)
1347 struct cpuinfo_x86
*c
= cpu_data
;
1349 for_each_cpu_mask(sibling
, cpu_core_map
[cpu
]) {
1350 cpu_clear(cpu
, cpu_core_map
[sibling
]);
1352 * last thread sibling in this cpu core going down
1354 if (cpus_weight(cpu_sibling_map
[cpu
]) == 1)
1355 c
[sibling
].booted_cores
--;
1358 for_each_cpu_mask(sibling
, cpu_sibling_map
[cpu
])
1359 cpu_clear(cpu
, cpu_sibling_map
[sibling
]);
1360 cpus_clear(cpu_sibling_map
[cpu
]);
1361 cpus_clear(cpu_core_map
[cpu
]);
1362 c
[cpu
].phys_proc_id
= 0;
1363 c
[cpu
].cpu_core_id
= 0;
1364 cpu_clear(cpu
, cpu_sibling_setup_map
);
1367 int __cpu_disable(void)
1369 cpumask_t map
= cpu_online_map
;
1370 int cpu
= smp_processor_id();
1373 * Perhaps use cpufreq to drop frequency, but that could go
1374 * into generic code.
1376 * We won't take down the boot processor on i386 due to some
1377 * interrupts only being able to be serviced by the BSP.
1378 * Especially so if we're not using an IOAPIC -zwane
1382 if (nmi_watchdog
== NMI_LOCAL_APIC
)
1383 stop_apic_nmi_watchdog(NULL
);
1385 /* Allow any queued timer interrupts to get serviced */
1388 local_irq_disable();
1390 remove_siblinginfo(cpu
);
1392 cpu_clear(cpu
, map
);
1394 /* It's now safe to remove this processor from the online map */
1395 cpu_clear(cpu
, cpu_online_map
);
1399 void __cpu_die(unsigned int cpu
)
1401 /* We don't do anything here: idle task is faking death itself. */
1404 for (i
= 0; i
< 10; i
++) {
1405 /* They ack this in play_dead by setting CPU_DEAD */
1406 if (per_cpu(cpu_state
, cpu
) == CPU_DEAD
) {
1407 printk ("CPU %d is now offline\n", cpu
);
1408 if (1 == num_online_cpus())
1409 alternatives_smp_switch(0);
1414 printk(KERN_ERR
"CPU %u didn't die...\n", cpu
);
1416 #else /* ... !CONFIG_HOTPLUG_CPU */
1417 int __cpu_disable(void)
1422 void __cpu_die(unsigned int cpu
)
1424 /* We said "no" in __cpu_disable */
1427 #endif /* CONFIG_HOTPLUG_CPU */
1429 int __devinit
__cpu_up(unsigned int cpu
)
1431 #ifdef CONFIG_HOTPLUG_CPU
1435 * We do warm boot only on cpus that had booted earlier
1436 * Otherwise cold boot is all handled from smp_boot_cpus().
1437 * cpu_callin_map is set during AP kickstart process. Its reset
1438 * when a cpu is taken offline from cpu_exit_clear().
1440 if (!cpu_isset(cpu
, cpu_callin_map
))
1441 ret
= __smp_prepare_cpu(cpu
);
1447 /* In case one didn't come up */
1448 if (!cpu_isset(cpu
, cpu_callin_map
)) {
1449 printk(KERN_DEBUG
"skipping cpu%d, didn't come online\n", cpu
);
1455 per_cpu(cpu_state
, cpu
) = CPU_UP_PREPARE
;
1456 /* Unleash the CPU! */
1457 cpu_set(cpu
, smp_commenced_mask
);
1458 while (!cpu_isset(cpu
, cpu_online_map
))
1463 void __init
smp_cpus_done(unsigned int max_cpus
)
1465 #ifdef CONFIG_X86_IO_APIC
1466 setup_ioapic_dest();
1469 #ifndef CONFIG_HOTPLUG_CPU
1471 * Disable executability of the SMP trampoline:
1473 set_kernel_exec((unsigned long)trampoline_base
, trampoline_exec
);
1477 void __init
smp_intr_init(void)
1480 * IRQ0 must be given a fixed assignment and initialized,
1481 * because it's used before the IO-APIC is set up.
1483 set_intr_gate(FIRST_DEVICE_VECTOR
, interrupt
[0]);
1486 * The reschedule interrupt is a CPU-to-CPU reschedule-helper
1487 * IPI, driven by wakeup.
1489 set_intr_gate(RESCHEDULE_VECTOR
, reschedule_interrupt
);
1491 /* IPI for invalidation */
1492 set_intr_gate(INVALIDATE_TLB_VECTOR
, invalidate_interrupt
);
1494 /* IPI for generic function call */
1495 set_intr_gate(CALL_FUNCTION_VECTOR
, call_function_interrupt
);
1499 * If the BIOS enumerates physical processors before logical,
1500 * maxcpus=N at enumeration-time can be used to disable HT.
1502 static int __init
parse_maxcpus(char *arg
)
1504 extern unsigned int maxcpus
;
1506 maxcpus
= simple_strtoul(arg
, NULL
, 0);
1509 early_param("maxcpus", parse_maxcpus
);