2 * SMP boot-related support
4 * Copyright (C) 1998-2003 Hewlett-Packard Co
5 * David Mosberger-Tang <davidm@hpl.hp.com>
7 * 01/05/16 Rohit Seth <rohit.seth@intel.com> Moved SMP booting functions from smp.c to here.
8 * 01/04/27 David Mosberger <davidm@hpl.hp.com> Added ITC synching code.
9 * 02/07/31 David Mosberger <davidm@hpl.hp.com> Switch over to hotplug-CPU boot-sequence.
10 * smp_boot_cpus()/smp_commence() is replaced by
11 * smp_prepare_cpus()/__cpu_up()/smp_cpus_done().
12 * 04/06/21 Ashok Raj <ashok.raj@intel.com> Added CPU Hotplug Support
14 #include <linux/config.h>
16 #include <linux/module.h>
17 #include <linux/acpi.h>
18 #include <linux/bootmem.h>
19 #include <linux/cpu.h>
20 #include <linux/delay.h>
21 #include <linux/init.h>
22 #include <linux/interrupt.h>
23 #include <linux/irq.h>
24 #include <linux/kernel.h>
25 #include <linux/kernel_stat.h>
27 #include <linux/notifier.h>
28 #include <linux/smp.h>
29 #include <linux/smp_lock.h>
30 #include <linux/spinlock.h>
31 #include <linux/efi.h>
32 #include <linux/percpu.h>
33 #include <linux/bitops.h>
35 #include <asm/atomic.h>
36 #include <asm/cache.h>
37 #include <asm/current.h>
38 #include <asm/delay.h>
42 #include <asm/machvec.h>
45 #include <asm/pgalloc.h>
46 #include <asm/pgtable.h>
47 #include <asm/processor.h>
48 #include <asm/ptrace.h>
50 #include <asm/system.h>
51 #include <asm/tlbflush.h>
52 #include <asm/unistd.h>
57 #define Dprintk(x...) printk(x)
62 #ifdef CONFIG_HOTPLUG_CPU
64 * Store all idle threads, this can be reused instead of creating
65 * a new thread. Also avoids complicated thread destroy functionality
68 struct task_struct
*idle_thread_array
[NR_CPUS
];
71 * Global array allocated for NR_CPUS at boot time
73 struct sal_to_os_boot sal_boot_rendez_state
[NR_CPUS
];
76 * start_ap in head.S uses this to store current booting cpu
79 struct sal_to_os_boot
*sal_state_for_booting_cpu
= &sal_boot_rendez_state
[0];
81 #define set_brendez_area(x) (sal_state_for_booting_cpu = &sal_boot_rendez_state[(x)]);
83 #define get_idle_for_cpu(x) (idle_thread_array[(x)])
84 #define set_idle_for_cpu(x,p) (idle_thread_array[(x)] = (p))
88 #define get_idle_for_cpu(x) (NULL)
89 #define set_idle_for_cpu(x,p)
90 #define set_brendez_area(x)
95 * ITC synchronization related stuff:
98 #define SLAVE (SMP_CACHE_BYTES/8)
100 #define NUM_ROUNDS 64 /* magic value */
101 #define NUM_ITERS 5 /* likewise */
103 static DEFINE_SPINLOCK(itc_sync_lock
);
104 static volatile unsigned long go
[SLAVE
+ 1];
106 #define DEBUG_ITC_SYNC 0
108 extern void __devinit
calibrate_delay (void);
109 extern void start_ap (void);
110 extern unsigned long ia64_iobase
;
112 task_t
*task_for_booting_cpu
;
117 DEFINE_PER_CPU(int, cpu_state
);
119 /* Bitmasks of currently online, and possible CPUs */
120 cpumask_t cpu_online_map
;
121 EXPORT_SYMBOL(cpu_online_map
);
122 cpumask_t cpu_possible_map
;
123 EXPORT_SYMBOL(cpu_possible_map
);
125 /* which logical CPU number maps to which CPU (physical APIC ID) */
126 volatile int ia64_cpu_to_sapicid
[NR_CPUS
];
127 EXPORT_SYMBOL(ia64_cpu_to_sapicid
);
129 static volatile cpumask_t cpu_callin_map
;
131 struct smp_boot_data smp_boot_data __initdata
;
133 unsigned long ap_wakeup_vector
= -1; /* External Int use to wakeup APs */
135 char __initdata no_int_routing
;
137 unsigned char smp_int_redirect
; /* are INT and IPI redirectable by the chipset? */
140 nointroute (char *str
)
143 printk ("no_int_routing on\n");
147 __setup("nointroute", nointroute
);
150 sync_master (void *arg
)
152 unsigned long flags
, i
;
156 local_irq_save(flags
);
158 for (i
= 0; i
< NUM_ROUNDS
*NUM_ITERS
; ++i
) {
161 go
[SLAVE
] = ia64_get_itc();
164 local_irq_restore(flags
);
168 * Return the number of cycles by which our itc differs from the itc on the master
169 * (time-keeper) CPU. A positive number indicates our itc is ahead of the master,
170 * negative that it is behind.
173 get_delta (long *rt
, long *master
)
175 unsigned long best_t0
= 0, best_t1
= ~0UL, best_tm
= 0;
176 unsigned long tcenter
, t0
, t1
, tm
;
179 for (i
= 0; i
< NUM_ITERS
; ++i
) {
182 while (!(tm
= go
[SLAVE
]));
186 if (t1
- t0
< best_t1
- best_t0
)
187 best_t0
= t0
, best_t1
= t1
, best_tm
= tm
;
190 *rt
= best_t1
- best_t0
;
191 *master
= best_tm
- best_t0
;
193 /* average best_t0 and best_t1 without overflow: */
194 tcenter
= (best_t0
/2 + best_t1
/2);
195 if (best_t0
% 2 + best_t1
% 2 == 2)
197 return tcenter
- best_tm
;
201 * Synchronize ar.itc of the current (slave) CPU with the ar.itc of the MASTER CPU
202 * (normally the time-keeper CPU). We use a closed loop to eliminate the possibility of
203 * unaccounted-for errors (such as getting a machine check in the middle of a calibration
204 * step). The basic idea is for the slave to ask the master what itc value it has and to
205 * read its own itc before and after the master responds. Each iteration gives us three
219 * The goal is to adjust the slave's ar.itc such that tm falls exactly half-way between t0
220 * and t1. If we achieve this, the clocks are synchronized provided the interconnect
221 * between the slave and the master is symmetric. Even if the interconnect were
222 * asymmetric, we would still know that the synchronization error is smaller than the
223 * roundtrip latency (t0 - t1).
225 * When the interconnect is quiet and symmetric, this lets us synchronize the itc to
226 * within one or two cycles. However, we can only *guarantee* that the synchronization is
227 * accurate to within a round-trip time, which is typically in the range of several
228 * hundred cycles (e.g., ~500 cycles). In practice, this means that the itc's are usually
229 * almost perfectly synchronized, but we shouldn't assume that the accuracy is much better
230 * than half a micro second or so.
233 ia64_sync_itc (unsigned int master
)
235 long i
, delta
, adj
, adjust_latency
= 0, done
= 0;
236 unsigned long flags
, rt
, master_time_stamp
, bound
;
239 long rt
; /* roundtrip time */
240 long master
; /* master's timestamp */
241 long diff
; /* difference between midpoint and master's timestamp */
242 long lat
; /* estimate of itc adjustment latency */
247 * Make sure local timer ticks are disabled while we sync. If
248 * they were enabled, we'd have to worry about nasty issues
249 * like setting the ITC ahead of (or a long time before) the
250 * next scheduled tick.
252 BUG_ON((ia64_get_itv() & (1 << 16)) == 0);
256 if (smp_call_function_single(master
, sync_master
, NULL
, 1, 0) < 0) {
257 printk(KERN_ERR
"sync_itc: failed to get attention of CPU %u!\n", master
);
261 while (go
[MASTER
]); /* wait for master to be ready */
263 spin_lock_irqsave(&itc_sync_lock
, flags
);
265 for (i
= 0; i
< NUM_ROUNDS
; ++i
) {
266 delta
= get_delta(&rt
, &master_time_stamp
);
268 done
= 1; /* let's lock on to this... */
274 adjust_latency
+= -delta
;
275 adj
= -delta
+ adjust_latency
/4;
279 ia64_set_itc(ia64_get_itc() + adj
);
283 t
[i
].master
= master_time_stamp
;
285 t
[i
].lat
= adjust_latency
/4;
289 spin_unlock_irqrestore(&itc_sync_lock
, flags
);
292 for (i
= 0; i
< NUM_ROUNDS
; ++i
)
293 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
294 t
[i
].rt
, t
[i
].master
, t
[i
].diff
, t
[i
].lat
);
297 printk(KERN_INFO
"CPU %d: synchronized ITC with CPU %u (last diff %ld cycles, "
298 "maxerr %lu cycles)\n", smp_processor_id(), master
, delta
, rt
);
302 * Ideally sets up per-cpu profiling hooks. Doesn't do much now...
304 static inline void __devinit
305 smp_setup_percpu_timer (void)
309 static void __devinit
313 extern void ia64_init_itm(void);
315 #ifdef CONFIG_PERFMON
316 extern void pfm_init_percpu(void);
319 cpuid
= smp_processor_id();
320 phys_id
= hard_smp_processor_id();
322 if (cpu_online(cpuid
)) {
323 printk(KERN_ERR
"huh, phys CPU#0x%x, CPU#0x%x already present??\n",
329 cpu_set(cpuid
, cpu_online_map
);
330 unlock_ipi_calllock();
332 smp_setup_percpu_timer();
334 ia64_mca_cmc_vector_setup(); /* Setup vector on AP */
336 #ifdef CONFIG_PERFMON
342 if (!(sal_platform_features
& IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT
)) {
344 * Synchronize the ITC with the BP. Need to do this after irqs are
345 * enabled because ia64_sync_itc() calls smp_call_function_single(), which
346 * calls spin_unlock_bh(), which calls spin_unlock_bh(), which calls
347 * local_bh_enable(), which bugs out if irqs are not enabled...
349 Dprintk("Going to syncup ITC with BP.\n");
358 local_cpu_data
->loops_per_jiffy
= loops_per_jiffy
;
360 #ifdef CONFIG_IA32_SUPPORT
365 * Allow the master to continue.
367 cpu_set(cpuid
, cpu_callin_map
);
368 Dprintk("Stack on CPU %d at about %p\n",cpuid
, &cpuid
);
373 * Activate a secondary processor. head.S calls this.
376 start_secondary (void *unused
)
378 /* Early console may use I/O ports */
379 ia64_set_kr(IA64_KR_IO_BASE
, __pa(ia64_iobase
));
380 Dprintk("start_secondary: starting CPU 0x%x\n", hard_smp_processor_id());
389 struct pt_regs
* __devinit
idle_regs(struct pt_regs
*regs
)
395 struct task_struct
*idle
;
396 struct completion done
;
401 do_fork_idle(void *_c_idle
)
403 struct create_idle
*c_idle
= _c_idle
;
405 c_idle
->idle
= fork_idle(c_idle
->cpu
);
406 complete(&c_idle
->done
);
410 do_boot_cpu (int sapicid
, int cpu
)
413 struct create_idle c_idle
= {
415 .done
= COMPLETION_INITIALIZER(c_idle
.done
),
417 DECLARE_WORK(work
, do_fork_idle
, &c_idle
);
419 c_idle
.idle
= get_idle_for_cpu(cpu
);
421 init_idle(c_idle
.idle
, cpu
);
426 * We can't use kernel_thread since we must avoid to reschedule the child.
428 if (!keventd_up() || current_is_keventd())
429 work
.func(work
.data
);
431 schedule_work(&work
);
432 wait_for_completion(&c_idle
.done
);
435 if (IS_ERR(c_idle
.idle
))
436 panic("failed fork for CPU %d", cpu
);
438 set_idle_for_cpu(cpu
, c_idle
.idle
);
441 task_for_booting_cpu
= c_idle
.idle
;
443 Dprintk("Sending wakeup vector %lu to AP 0x%x/0x%x.\n", ap_wakeup_vector
, cpu
, sapicid
);
445 set_brendez_area(cpu
);
446 platform_send_ipi(cpu
, ap_wakeup_vector
, IA64_IPI_DM_INT
, 0);
449 * Wait 10s total for the AP to start
451 Dprintk("Waiting on callin_map ...");
452 for (timeout
= 0; timeout
< 100000; timeout
++) {
453 if (cpu_isset(cpu
, cpu_callin_map
))
454 break; /* It has booted */
459 if (!cpu_isset(cpu
, cpu_callin_map
)) {
460 printk(KERN_ERR
"Processor 0x%x/0x%x is stuck.\n", cpu
, sapicid
);
461 ia64_cpu_to_sapicid
[cpu
] = -1;
462 cpu_clear(cpu
, cpu_online_map
); /* was set in smp_callin() */
472 get_option (&str
, &ticks
);
476 __setup("decay=", decay
);
479 * Initialize the logical CPU number to SAPICID mapping
482 smp_build_cpu_map (void)
485 int boot_cpu_id
= hard_smp_processor_id();
487 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++) {
488 ia64_cpu_to_sapicid
[cpu
] = -1;
489 #ifdef CONFIG_HOTPLUG_CPU
490 cpu_set(cpu
, cpu_possible_map
);
494 ia64_cpu_to_sapicid
[0] = boot_cpu_id
;
495 cpus_clear(cpu_present_map
);
496 cpu_set(0, cpu_present_map
);
497 cpu_set(0, cpu_possible_map
);
498 for (cpu
= 1, i
= 0; i
< smp_boot_data
.cpu_count
; i
++) {
499 sapicid
= smp_boot_data
.cpu_phys_id
[i
];
500 if (sapicid
== boot_cpu_id
)
502 cpu_set(cpu
, cpu_present_map
);
503 cpu_set(cpu
, cpu_possible_map
);
504 ia64_cpu_to_sapicid
[cpu
] = sapicid
;
511 /* on which node is each logical CPU (one cacheline even for 64 CPUs) */
512 u8 cpu_to_node_map
[NR_CPUS
] __cacheline_aligned
;
513 EXPORT_SYMBOL(cpu_to_node_map
);
514 /* which logical CPUs are on which nodes */
515 cpumask_t node_to_cpu_mask
[MAX_NUMNODES
] __cacheline_aligned
;
518 * Build cpu to node mapping and initialize the per node cpu masks.
521 build_cpu_to_node_map (void)
525 for(node
=0; node
<MAX_NUMNODES
; node
++)
526 cpus_clear(node_to_cpu_mask
[node
]);
527 for(cpu
= 0; cpu
< NR_CPUS
; ++cpu
) {
529 * All Itanium NUMA platforms I know use ACPI, so maybe we
530 * can drop this ifdef completely. [EF]
532 #ifdef CONFIG_ACPI_NUMA
534 for (i
= 0; i
< NR_CPUS
; ++i
)
535 if (cpu_physical_id(cpu
) == node_cpuid
[i
].phys_id
) {
536 node
= node_cpuid
[i
].nid
;
540 # error Fixme: Dunno how to build CPU-to-node map.
542 cpu_to_node_map
[cpu
] = (node
>= 0) ? node
: 0;
544 cpu_set(cpu
, node_to_cpu_mask
[node
]);
548 #endif /* CONFIG_NUMA */
551 * Cycle through the APs sending Wakeup IPIs to boot each.
554 smp_prepare_cpus (unsigned int max_cpus
)
556 int boot_cpu_id
= hard_smp_processor_id();
559 * Initialize the per-CPU profiling counter/multiplier
562 smp_setup_percpu_timer();
565 * We have the boot CPU online for sure.
567 cpu_set(0, cpu_online_map
);
568 cpu_set(0, cpu_callin_map
);
570 local_cpu_data
->loops_per_jiffy
= loops_per_jiffy
;
571 ia64_cpu_to_sapicid
[0] = boot_cpu_id
;
573 printk(KERN_INFO
"Boot processor id 0x%x/0x%x\n", 0, boot_cpu_id
);
575 current_thread_info()->cpu
= 0;
578 * If SMP should be disabled, then really disable it!
581 printk(KERN_INFO
"SMP mode deactivated.\n");
582 cpus_clear(cpu_online_map
);
583 cpus_clear(cpu_present_map
);
584 cpus_clear(cpu_possible_map
);
585 cpu_set(0, cpu_online_map
);
586 cpu_set(0, cpu_present_map
);
587 cpu_set(0, cpu_possible_map
);
592 void __devinit
smp_prepare_boot_cpu(void)
594 cpu_set(smp_processor_id(), cpu_online_map
);
595 cpu_set(smp_processor_id(), cpu_callin_map
);
598 #ifdef CONFIG_HOTPLUG_CPU
599 extern void fixup_irqs(void);
600 /* must be called with cpucontrol mutex held */
601 int __cpu_disable(void)
603 int cpu
= smp_processor_id();
606 * dont permit boot processor for now
612 local_flush_tlb_all();
613 cpu_clear(cpu
, cpu_callin_map
);
617 void __cpu_die(unsigned int cpu
)
621 for (i
= 0; i
< 100; i
++) {
622 /* They ack this in play_dead by setting CPU_DEAD */
623 if (per_cpu(cpu_state
, cpu
) == CPU_DEAD
)
625 printk ("CPU %d is now offline\n", cpu
);
630 printk(KERN_ERR
"CPU %u didn't die...\n", cpu
);
632 #else /* !CONFIG_HOTPLUG_CPU */
633 int __cpu_disable(void)
638 void __cpu_die(unsigned int cpu
)
640 /* We said "no" in __cpu_disable */
643 #endif /* CONFIG_HOTPLUG_CPU */
646 smp_cpus_done (unsigned int dummy
)
649 unsigned long bogosum
= 0;
652 * Allow the user to impress friends.
655 for (cpu
= 0; cpu
< NR_CPUS
; cpu
++)
657 bogosum
+= cpu_data(cpu
)->loops_per_jiffy
;
659 printk(KERN_INFO
"Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
660 (int)num_online_cpus(), bogosum
/(500000/HZ
), (bogosum
/(5000/HZ
))%100);
664 __cpu_up (unsigned int cpu
)
669 sapicid
= ia64_cpu_to_sapicid
[cpu
];
674 * Already booted cpu? not valid anymore since we dont
675 * do idle loop tightspin anymore.
677 if (cpu_isset(cpu
, cpu_callin_map
))
680 /* Processor goes to start_secondary(), sets online flag */
681 ret
= do_boot_cpu(sapicid
, cpu
);
689 * Assume that CPU's have been discovered by some platform-dependent interface. For
690 * SoftSDV/Lion, that would be ACPI.
692 * Setup of the IPI irq handler is done in irq.c:init_IRQ_SMP().
695 init_smp_config(void)
703 /* Tell SAL where to drop the AP's. */
704 ap_startup
= (struct fptr
*) start_ap
;
705 sal_ret
= ia64_sal_set_vectors(SAL_VECTOR_OS_BOOT_RENDEZ
,
706 ia64_tpa(ap_startup
->fp
), ia64_tpa(ap_startup
->gp
), 0, 0, 0, 0);
708 printk(KERN_ERR
"SMP: Can't set SAL AP Boot Rendezvous: %s\n",
709 ia64_sal_strerror(sal_ret
));