ia64/xen-unstable

view xen/arch/ia64/linux-xen/smpboot.c @ 10888:5379548bfc79

[NET] Enable TCPv4 segmentation offload in front/back drivers.
Signed-off-by: Keir Fraser <keir@xensource.com>
author kfraser@localhost.localdomain
date Tue Aug 01 11:54:45 2006 +0100 (2006-08-01)
parents 75f791c710df
children 21f8c507da29
line source
1 /*
2 * SMP boot-related support
3 *
4 * Copyright (C) 1998-2003, 2005 Hewlett-Packard Co
5 * David Mosberger-Tang <davidm@hpl.hp.com>
6 * Copyright (C) 2001, 2004-2005 Intel Corp
7 * Rohit Seth <rohit.seth@intel.com>
8 * Suresh Siddha <suresh.b.siddha@intel.com>
9 * Gordon Jin <gordon.jin@intel.com>
10 * Ashok Raj <ashok.raj@intel.com>
11 *
12 * 01/05/16 Rohit Seth <rohit.seth@intel.com> Moved SMP booting functions from smp.c to here.
13 * 01/04/27 David Mosberger <davidm@hpl.hp.com> Added ITC synching code.
14 * 02/07/31 David Mosberger <davidm@hpl.hp.com> Switch over to hotplug-CPU boot-sequence.
15 * smp_boot_cpus()/smp_commence() is replaced by
16 * smp_prepare_cpus()/__cpu_up()/smp_cpus_done().
17 * 04/06/21 Ashok Raj <ashok.raj@intel.com> Added CPU Hotplug Support
18 * 04/12/26 Jin Gordon <gordon.jin@intel.com>
19 * 04/12/26 Rohit Seth <rohit.seth@intel.com>
20 * Add multi-threading and multi-core detection
21 * 05/01/30 Suresh Siddha <suresh.b.siddha@intel.com>
22 * Setup cpu_sibling_map and cpu_core_map
23 */
24 #include <linux/config.h>
26 #include <linux/module.h>
27 #include <linux/acpi.h>
28 #include <linux/bootmem.h>
29 #include <linux/cpu.h>
30 #include <linux/delay.h>
31 #include <linux/init.h>
32 #include <linux/interrupt.h>
33 #include <linux/irq.h>
34 #include <linux/kernel.h>
35 #include <linux/kernel_stat.h>
36 #include <linux/mm.h>
37 #include <linux/notifier.h> /* hg add me */
38 #include <linux/smp.h>
39 #include <linux/smp_lock.h>
40 #include <linux/spinlock.h>
41 #include <linux/efi.h>
42 #include <linux/percpu.h>
43 #include <linux/bitops.h>
45 #include <asm/atomic.h>
46 #include <asm/cache.h>
47 #include <asm/current.h>
48 #include <asm/delay.h>
49 #include <asm/ia32.h>
50 #include <asm/io.h>
51 #include <asm/irq.h>
52 #include <asm/machvec.h>
53 #include <asm/mca.h>
54 #include <asm/page.h>
55 #include <asm/pgalloc.h>
56 #include <asm/pgtable.h>
57 #include <asm/processor.h>
58 #include <asm/ptrace.h>
59 #include <asm/sal.h>
60 #include <asm/system.h>
61 #include <asm/tlbflush.h>
62 #include <asm/unistd.h>
64 #ifdef XEN
65 #include <xen/domain.h>
66 #include <asm/hw_irq.h>
67 #ifndef CONFIG_SMP
68 cpumask_t cpu_online_map = CPU_MASK_CPU0;
69 EXPORT_SYMBOL(cpu_online_map);
70 #endif
71 #endif
73 #ifdef CONFIG_SMP /* ifdef XEN */
75 #define SMP_DEBUG 0
77 #if SMP_DEBUG
78 #define Dprintk(x...) printk(x)
79 #else
80 #define Dprintk(x...)
81 #endif
83 #ifdef CONFIG_HOTPLUG_CPU
84 /*
85 * Store all idle threads, this can be reused instead of creating
86 * a new thread. Also avoids complicated thread destroy functionality
87 * for idle threads.
88 */
89 struct task_struct *idle_thread_array[NR_CPUS];
91 /*
92 * Global array allocated for NR_CPUS at boot time
93 */
94 struct sal_to_os_boot sal_boot_rendez_state[NR_CPUS];
96 /*
97 * start_ap in head.S uses this to store current booting cpu
98 * info.
99 */
100 struct sal_to_os_boot *sal_state_for_booting_cpu = &sal_boot_rendez_state[0];
102 #define set_brendez_area(x) (sal_state_for_booting_cpu = &sal_boot_rendez_state[(x)]);
104 #define get_idle_for_cpu(x) (idle_thread_array[(x)])
105 #define set_idle_for_cpu(x,p) (idle_thread_array[(x)] = (p))
107 #else
109 #define get_idle_for_cpu(x) (NULL)
110 #define set_idle_for_cpu(x,p)
111 #define set_brendez_area(x)
112 #endif
115 /*
116 * ITC synchronization related stuff:
117 */
118 #define MASTER 0
119 #define SLAVE (SMP_CACHE_BYTES/8)
121 #define NUM_ROUNDS 64 /* magic value */
122 #define NUM_ITERS 5 /* likewise */
124 static DEFINE_SPINLOCK(itc_sync_lock);
125 static volatile unsigned long go[SLAVE + 1];
127 #define DEBUG_ITC_SYNC 0
129 extern void __devinit calibrate_delay (void);
130 extern void start_ap (void);
131 extern unsigned long ia64_iobase;
133 task_t *task_for_booting_cpu;
135 /*
136 * State for each CPU
137 */
138 DEFINE_PER_CPU(int, cpu_state);
140 /* Bitmasks of currently online, and possible CPUs */
141 cpumask_t cpu_online_map;
142 EXPORT_SYMBOL(cpu_online_map);
143 cpumask_t cpu_possible_map;
144 EXPORT_SYMBOL(cpu_possible_map);
146 cpumask_t cpu_core_map[NR_CPUS] __cacheline_aligned;
147 cpumask_t cpu_sibling_map[NR_CPUS] __cacheline_aligned;
148 int smp_num_siblings = 1;
149 int smp_num_cpucores = 1;
151 /* which logical CPU number maps to which CPU (physical APIC ID) */
152 volatile int ia64_cpu_to_sapicid[NR_CPUS];
153 EXPORT_SYMBOL(ia64_cpu_to_sapicid);
155 static volatile cpumask_t cpu_callin_map;
157 struct smp_boot_data smp_boot_data __initdata;
159 unsigned long ap_wakeup_vector = -1; /* External Int use to wakeup APs */
161 char __initdata no_int_routing;
163 unsigned char smp_int_redirect; /* are INT and IPI redirectable by the chipset? */
165 static int __init
166 nointroute (char *str)
167 {
168 no_int_routing = 1;
169 printk ("no_int_routing on\n");
170 return 1;
171 }
173 __setup("nointroute", nointroute);
175 void
176 sync_master (void *arg)
177 {
178 unsigned long flags, i;
180 go[MASTER] = 0;
182 local_irq_save(flags);
183 {
184 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; ++i) {
185 while (!go[MASTER])
186 cpu_relax();
187 go[MASTER] = 0;
188 go[SLAVE] = ia64_get_itc();
189 }
190 }
191 local_irq_restore(flags);
192 }
194 /*
195 * Return the number of cycles by which our itc differs from the itc on the master
196 * (time-keeper) CPU. A positive number indicates our itc is ahead of the master,
197 * negative that it is behind.
198 */
199 static inline long
200 #ifdef XEN /* warning cleanup */
201 get_delta (unsigned long *rt, unsigned long *master)
202 #else
203 get_delta (long *rt, long *master)
204 #endif
205 {
206 unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
207 unsigned long tcenter, t0, t1, tm;
208 long i;
210 for (i = 0; i < NUM_ITERS; ++i) {
211 t0 = ia64_get_itc();
212 go[MASTER] = 1;
213 while (!(tm = go[SLAVE]))
214 cpu_relax();
215 go[SLAVE] = 0;
216 t1 = ia64_get_itc();
218 if (t1 - t0 < best_t1 - best_t0)
219 best_t0 = t0, best_t1 = t1, best_tm = tm;
220 }
222 *rt = best_t1 - best_t0;
223 *master = best_tm - best_t0;
225 /* average best_t0 and best_t1 without overflow: */
226 tcenter = (best_t0/2 + best_t1/2);
227 if (best_t0 % 2 + best_t1 % 2 == 2)
228 ++tcenter;
229 return tcenter - best_tm;
230 }
232 /*
233 * Synchronize ar.itc of the current (slave) CPU with the ar.itc of the MASTER CPU
234 * (normally the time-keeper CPU). We use a closed loop to eliminate the possibility of
235 * unaccounted-for errors (such as getting a machine check in the middle of a calibration
236 * step). The basic idea is for the slave to ask the master what itc value it has and to
237 * read its own itc before and after the master responds. Each iteration gives us three
238 * timestamps:
239 *
240 * slave master
241 *
242 * t0 ---\
243 * ---\
244 * --->
245 * tm
246 * /---
247 * /---
248 * t1 <---
249 *
250 *
251 * The goal is to adjust the slave's ar.itc such that tm falls exactly half-way between t0
252 * and t1. If we achieve this, the clocks are synchronized provided the interconnect
253 * between the slave and the master is symmetric. Even if the interconnect were
254 * asymmetric, we would still know that the synchronization error is smaller than the
255 * roundtrip latency (t0 - t1).
256 *
257 * When the interconnect is quiet and symmetric, this lets us synchronize the itc to
258 * within one or two cycles. However, we can only *guarantee* that the synchronization is
259 * accurate to within a round-trip time, which is typically in the range of several
260 * hundred cycles (e.g., ~500 cycles). In practice, this means that the itc's are usually
261 * almost perfectly synchronized, but we shouldn't assume that the accuracy is much better
262 * than half a micro second or so.
263 */
264 void
265 ia64_sync_itc (unsigned int master)
266 {
267 long i, delta, adj, adjust_latency = 0, done = 0;
268 unsigned long flags, rt, master_time_stamp, bound;
269 #if DEBUG_ITC_SYNC
270 struct {
271 long rt; /* roundtrip time */
272 long master; /* master's timestamp */
273 long diff; /* difference between midpoint and master's timestamp */
274 long lat; /* estimate of itc adjustment latency */
275 } t[NUM_ROUNDS];
276 #endif
278 /*
279 * Make sure local timer ticks are disabled while we sync. If
280 * they were enabled, we'd have to worry about nasty issues
281 * like setting the ITC ahead of (or a long time before) the
282 * next scheduled tick.
283 */
284 BUG_ON((ia64_get_itv() & (1 << 16)) == 0);
286 go[MASTER] = 1;
288 if (smp_call_function_single(master, sync_master, NULL, 1, 0) < 0) {
289 printk(KERN_ERR "sync_itc: failed to get attention of CPU %u!\n", master);
290 return;
291 }
293 while (go[MASTER])
294 cpu_relax(); /* wait for master to be ready */
296 spin_lock_irqsave(&itc_sync_lock, flags);
297 {
298 for (i = 0; i < NUM_ROUNDS; ++i) {
299 delta = get_delta(&rt, &master_time_stamp);
300 if (delta == 0) {
301 done = 1; /* let's lock on to this... */
302 bound = rt;
303 }
305 if (!done) {
306 if (i > 0) {
307 adjust_latency += -delta;
308 adj = -delta + adjust_latency/4;
309 } else
310 adj = -delta;
312 ia64_set_itc(ia64_get_itc() + adj);
313 }
314 #if DEBUG_ITC_SYNC
315 t[i].rt = rt;
316 t[i].master = master_time_stamp;
317 t[i].diff = delta;
318 t[i].lat = adjust_latency/4;
319 #endif
320 }
321 }
322 spin_unlock_irqrestore(&itc_sync_lock, flags);
324 #if DEBUG_ITC_SYNC
325 for (i = 0; i < NUM_ROUNDS; ++i)
326 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
327 t[i].rt, t[i].master, t[i].diff, t[i].lat);
328 #endif
330 printk(KERN_INFO "CPU %d: synchronized ITC with CPU %u (last diff %ld cycles, "
331 "maxerr %lu cycles)\n", smp_processor_id(), master, delta, rt);
332 }
334 /*
335 * Ideally sets up per-cpu profiling hooks. Doesn't do much now...
336 */
337 static inline void __devinit
338 smp_setup_percpu_timer (void)
339 {
340 }
342 static void __devinit
343 smp_callin (void)
344 {
345 int cpuid, phys_id;
346 extern void ia64_init_itm(void);
348 #ifdef CONFIG_PERFMON
349 extern void pfm_init_percpu(void);
350 #endif
352 cpuid = smp_processor_id();
353 phys_id = hard_smp_processor_id();
355 if (cpu_online(cpuid)) {
356 printk(KERN_ERR "huh, phys CPU#0x%x, CPU#0x%x already present??\n",
357 phys_id, cpuid);
358 BUG();
359 }
361 lock_ipi_calllock();
362 cpu_set(cpuid, cpu_online_map);
363 unlock_ipi_calllock();
364 per_cpu(cpu_state, cpuid) = CPU_ONLINE;
366 smp_setup_percpu_timer();
368 #ifndef XEN
369 ia64_mca_cmc_vector_setup(); /* Setup vector on AP */
370 #endif
372 #ifdef CONFIG_PERFMON
373 pfm_init_percpu();
374 #endif
376 local_irq_enable();
378 if (!(sal_platform_features & IA64_SAL_PLATFORM_FEATURE_ITC_DRIFT)) {
379 /*
380 * Synchronize the ITC with the BP. Need to do this after irqs are
381 * enabled because ia64_sync_itc() calls smp_call_function_single(), which
382 * calls spin_unlock_bh(), which calls spin_unlock_bh(), which calls
383 * local_bh_enable(), which bugs out if irqs are not enabled...
384 */
385 Dprintk("Going to syncup ITC with BP.\n");
386 ia64_sync_itc(0);
387 }
389 /*
390 * Get our bogomips.
391 */
392 ia64_init_itm();
393 #ifndef XEN
394 calibrate_delay();
395 #endif
396 local_cpu_data->loops_per_jiffy = loops_per_jiffy;
398 #ifdef CONFIG_IA32_SUPPORT
399 ia32_gdt_init();
400 #endif
402 /*
403 * Allow the master to continue.
404 */
405 cpu_set(cpuid, cpu_callin_map);
406 Dprintk("Stack on CPU %d at about %p\n",cpuid, &cpuid);
407 }
410 /*
411 * Activate a secondary processor. head.S calls this.
412 */
413 int __devinit
414 start_secondary (void *unused)
415 {
416 /* Early console may use I/O ports */
417 ia64_set_kr(IA64_KR_IO_BASE, __pa(ia64_iobase));
418 Dprintk("start_secondary: starting CPU 0x%x\n", hard_smp_processor_id());
419 efi_map_pal_code();
420 cpu_init();
421 smp_callin();
423 #ifdef XEN
424 startup_cpu_idle_loop();
425 #else
426 cpu_idle();
427 #endif
428 return 0;
429 }
431 struct pt_regs * __devinit idle_regs(struct pt_regs *regs)
432 {
433 return NULL;
434 }
436 #ifndef XEN
437 struct create_idle {
438 struct task_struct *idle;
439 struct completion done;
440 int cpu;
441 };
443 void
444 do_fork_idle(void *_c_idle)
445 {
446 struct create_idle *c_idle = _c_idle;
448 c_idle->idle = fork_idle(c_idle->cpu);
449 complete(&c_idle->done);
450 }
451 #endif
453 static int __devinit
454 do_boot_cpu (int sapicid, int cpu)
455 {
456 int timeout;
457 #ifndef XEN
458 struct create_idle c_idle = {
459 .cpu = cpu,
460 .done = COMPLETION_INITIALIZER(c_idle.done),
461 };
462 DECLARE_WORK(work, do_fork_idle, &c_idle);
464 c_idle.idle = get_idle_for_cpu(cpu);
465 if (c_idle.idle) {
466 init_idle(c_idle.idle, cpu);
467 goto do_rest;
468 }
470 /*
471 * We can't use kernel_thread since we must avoid to reschedule the child.
472 */
473 if (!keventd_up() || current_is_keventd())
474 work.func(work.data);
475 else {
476 schedule_work(&work);
477 wait_for_completion(&c_idle.done);
478 }
480 if (IS_ERR(c_idle.idle))
481 panic("failed fork for CPU %d", cpu);
483 set_idle_for_cpu(cpu, c_idle.idle);
485 do_rest:
486 task_for_booting_cpu = c_idle.idle;
487 #else
488 struct vcpu *v;
490 v = alloc_idle_vcpu(cpu);
491 BUG_ON(v == NULL);
493 //printf ("do_boot_cpu: cpu=%d, domain=%p, vcpu=%p\n", cpu, idle, v);
495 task_for_booting_cpu = (task_t *)v;
497 /* Set cpu number. */
498 get_thread_info(v)->cpu = cpu;
499 #endif
501 Dprintk("Sending wakeup vector %lu to AP 0x%x/0x%x.\n", ap_wakeup_vector, cpu, sapicid);
503 set_brendez_area(cpu);
504 platform_send_ipi(cpu, ap_wakeup_vector, IA64_IPI_DM_INT, 0);
506 /*
507 * Wait 10s total for the AP to start
508 */
509 Dprintk("Waiting on callin_map ...");
510 for (timeout = 0; timeout < 100000; timeout++) {
511 if (cpu_isset(cpu, cpu_callin_map))
512 break; /* It has booted */
513 udelay(100);
514 }
515 Dprintk("\n");
517 if (!cpu_isset(cpu, cpu_callin_map)) {
518 printk(KERN_ERR "Processor 0x%x/0x%x is stuck.\n", cpu, sapicid);
519 ia64_cpu_to_sapicid[cpu] = -1;
520 cpu_clear(cpu, cpu_online_map); /* was set in smp_callin() */
521 return -EINVAL;
522 }
523 return 0;
524 }
526 #ifndef XEN
527 static int __init
528 decay (char *str)
529 {
530 int ticks;
531 get_option (&str, &ticks);
532 return 1;
533 }
535 __setup("decay=", decay);
536 #endif
538 /*
539 * Initialize the logical CPU number to SAPICID mapping
540 */
541 void __init
542 smp_build_cpu_map (void)
543 {
544 int sapicid, cpu, i;
545 int boot_cpu_id = hard_smp_processor_id();
547 for (cpu = 0; cpu < NR_CPUS; cpu++) {
548 ia64_cpu_to_sapicid[cpu] = -1;
549 #ifdef CONFIG_HOTPLUG_CPU
550 cpu_set(cpu, cpu_possible_map);
551 #endif
552 }
554 ia64_cpu_to_sapicid[0] = boot_cpu_id;
555 cpus_clear(cpu_present_map);
556 cpu_set(0, cpu_present_map);
557 cpu_set(0, cpu_possible_map);
558 for (cpu = 1, i = 0; i < smp_boot_data.cpu_count; i++) {
559 sapicid = smp_boot_data.cpu_phys_id[i];
560 if (sapicid == boot_cpu_id)
561 continue;
562 cpu_set(cpu, cpu_present_map);
563 cpu_set(cpu, cpu_possible_map);
564 ia64_cpu_to_sapicid[cpu] = sapicid;
565 cpu++;
566 }
567 }
569 /*
570 * Cycle through the APs sending Wakeup IPIs to boot each.
571 */
572 void __init
573 smp_prepare_cpus (unsigned int max_cpus)
574 {
575 int boot_cpu_id = hard_smp_processor_id();
577 /*
578 * Initialize the per-CPU profiling counter/multiplier
579 */
581 smp_setup_percpu_timer();
583 /*
584 * We have the boot CPU online for sure.
585 */
586 cpu_set(0, cpu_online_map);
587 cpu_set(0, cpu_callin_map);
589 local_cpu_data->loops_per_jiffy = loops_per_jiffy;
590 ia64_cpu_to_sapicid[0] = boot_cpu_id;
592 printk(KERN_INFO "Boot processor id 0x%x/0x%x\n", 0, boot_cpu_id);
594 current_thread_info()->cpu = 0;
596 /*
597 * If SMP should be disabled, then really disable it!
598 */
599 if (!max_cpus) {
600 printk(KERN_INFO "SMP mode deactivated.\n");
601 cpus_clear(cpu_online_map);
602 cpus_clear(cpu_present_map);
603 cpus_clear(cpu_possible_map);
604 cpu_set(0, cpu_online_map);
605 cpu_set(0, cpu_present_map);
606 cpu_set(0, cpu_possible_map);
607 return;
608 }
609 }
611 void __devinit smp_prepare_boot_cpu(void)
612 {
613 cpu_set(smp_processor_id(), cpu_online_map);
614 cpu_set(smp_processor_id(), cpu_callin_map);
615 per_cpu(cpu_state, smp_processor_id()) = CPU_ONLINE;
616 }
618 /*
619 * mt_info[] is a temporary store for all info returned by
620 * PAL_LOGICAL_TO_PHYSICAL, to be copied into cpuinfo_ia64 when the
621 * specific cpu comes.
622 */
623 static struct {
624 __u32 socket_id;
625 __u16 core_id;
626 __u16 thread_id;
627 __u16 proc_fixed_addr;
628 __u8 valid;
629 } mt_info[NR_CPUS] __devinitdata;
631 #ifdef CONFIG_HOTPLUG_CPU
632 static inline void
633 remove_from_mtinfo(int cpu)
634 {
635 int i;
637 for_each_cpu(i)
638 if (mt_info[i].valid && mt_info[i].socket_id ==
639 cpu_data(cpu)->socket_id)
640 mt_info[i].valid = 0;
641 }
643 static inline void
644 clear_cpu_sibling_map(int cpu)
645 {
646 int i;
648 for_each_cpu_mask(i, cpu_sibling_map[cpu])
649 cpu_clear(cpu, cpu_sibling_map[i]);
650 for_each_cpu_mask(i, cpu_core_map[cpu])
651 cpu_clear(cpu, cpu_core_map[i]);
653 cpu_sibling_map[cpu] = cpu_core_map[cpu] = CPU_MASK_NONE;
654 }
656 static void
657 remove_siblinginfo(int cpu)
658 {
659 int last = 0;
661 if (cpu_data(cpu)->threads_per_core == 1 &&
662 cpu_data(cpu)->cores_per_socket == 1) {
663 cpu_clear(cpu, cpu_core_map[cpu]);
664 cpu_clear(cpu, cpu_sibling_map[cpu]);
665 return;
666 }
668 last = (cpus_weight(cpu_core_map[cpu]) == 1 ? 1 : 0);
670 /* remove it from all sibling map's */
671 clear_cpu_sibling_map(cpu);
673 /* if this cpu is the last in the core group, remove all its info
674 * from mt_info structure
675 */
676 if (last)
677 remove_from_mtinfo(cpu);
678 }
680 extern void fixup_irqs(void);
681 /* must be called with cpucontrol mutex held */
682 int __cpu_disable(void)
683 {
684 int cpu = smp_processor_id();
686 /*
687 * dont permit boot processor for now
688 */
689 if (cpu == 0)
690 return -EBUSY;
692 remove_siblinginfo(cpu);
693 cpu_clear(cpu, cpu_online_map);
694 fixup_irqs();
695 local_flush_tlb_all();
696 cpu_clear(cpu, cpu_callin_map);
697 return 0;
698 }
700 void __cpu_die(unsigned int cpu)
701 {
702 unsigned int i;
704 for (i = 0; i < 100; i++) {
705 /* They ack this in play_dead by setting CPU_DEAD */
706 if (per_cpu(cpu_state, cpu) == CPU_DEAD)
707 {
708 printk ("CPU %d is now offline\n", cpu);
709 return;
710 }
711 msleep(100);
712 }
713 printk(KERN_ERR "CPU %u didn't die...\n", cpu);
714 }
715 #else /* !CONFIG_HOTPLUG_CPU */
716 int __cpu_disable(void)
717 {
718 return -ENOSYS;
719 }
721 void __cpu_die(unsigned int cpu)
722 {
723 /* We said "no" in __cpu_disable */
724 BUG();
725 }
726 #endif /* CONFIG_HOTPLUG_CPU */
728 void
729 smp_cpus_done (unsigned int dummy)
730 {
731 int cpu;
732 unsigned long bogosum = 0;
734 /*
735 * Allow the user to impress friends.
736 */
738 for (cpu = 0; cpu < NR_CPUS; cpu++)
739 if (cpu_online(cpu))
740 bogosum += cpu_data(cpu)->loops_per_jiffy;
742 printk(KERN_INFO "Total of %d processors activated (%lu.%02lu BogoMIPS).\n",
743 (int)num_online_cpus(), bogosum/(500000/HZ), (bogosum/(5000/HZ))%100);
744 }
746 static inline void __devinit
747 set_cpu_sibling_map(int cpu)
748 {
749 int i;
751 for_each_online_cpu(i) {
752 if ((cpu_data(cpu)->socket_id == cpu_data(i)->socket_id)) {
753 cpu_set(i, cpu_core_map[cpu]);
754 cpu_set(cpu, cpu_core_map[i]);
755 if (cpu_data(cpu)->core_id == cpu_data(i)->core_id) {
756 cpu_set(i, cpu_sibling_map[cpu]);
757 cpu_set(cpu, cpu_sibling_map[i]);
758 }
759 }
760 }
761 }
763 int __devinit
764 __cpu_up (unsigned int cpu)
765 {
766 int ret;
767 int sapicid;
769 sapicid = ia64_cpu_to_sapicid[cpu];
770 if (sapicid == -1)
771 return -EINVAL;
773 /*
774 * Already booted cpu? not valid anymore since we dont
775 * do idle loop tightspin anymore.
776 */
777 if (cpu_isset(cpu, cpu_callin_map))
778 return -EINVAL;
780 per_cpu(cpu_state, cpu) = CPU_UP_PREPARE;
781 /* Processor goes to start_secondary(), sets online flag */
782 ret = do_boot_cpu(sapicid, cpu);
783 if (ret < 0)
784 return ret;
786 if (cpu_data(cpu)->threads_per_core == 1 &&
787 cpu_data(cpu)->cores_per_socket == 1) {
788 cpu_set(cpu, cpu_sibling_map[cpu]);
789 cpu_set(cpu, cpu_core_map[cpu]);
790 return 0;
791 }
793 set_cpu_sibling_map(cpu);
795 return 0;
796 }
798 /*
799 * Assume that CPU's have been discovered by some platform-dependent interface. For
800 * SoftSDV/Lion, that would be ACPI.
801 *
802 * Setup of the IPI irq handler is done in irq.c:init_IRQ_SMP().
803 */
804 void __init
805 init_smp_config(void)
806 {
807 struct fptr {
808 unsigned long fp;
809 unsigned long gp;
810 } *ap_startup;
811 long sal_ret;
813 /* Tell SAL where to drop the AP's. */
814 ap_startup = (struct fptr *) start_ap;
815 sal_ret = ia64_sal_set_vectors(SAL_VECTOR_OS_BOOT_RENDEZ,
816 ia64_tpa(ap_startup->fp), ia64_tpa(ap_startup->gp), 0, 0, 0, 0);
817 if (sal_ret < 0)
818 printk(KERN_ERR "SMP: Can't set SAL AP Boot Rendezvous: %s\n",
819 ia64_sal_strerror(sal_ret));
820 }
822 static inline int __devinit
823 check_for_mtinfo_index(void)
824 {
825 int i;
827 for_each_cpu(i)
828 if (!mt_info[i].valid)
829 return i;
831 return -1;
832 }
834 /*
835 * Search the mt_info to find out if this socket's cid/tid information is
836 * cached or not. If the socket exists, fill in the core_id and thread_id
837 * in cpuinfo
838 */
839 static int __devinit
840 check_for_new_socket(__u16 logical_address, struct cpuinfo_ia64 *c)
841 {
842 int i;
843 __u32 sid = c->socket_id;
845 for_each_cpu(i) {
846 if (mt_info[i].valid && mt_info[i].proc_fixed_addr == logical_address
847 && mt_info[i].socket_id == sid) {
848 c->core_id = mt_info[i].core_id;
849 c->thread_id = mt_info[i].thread_id;
850 return 1; /* not a new socket */
851 }
852 }
853 return 0;
854 }
856 /*
857 * identify_siblings(cpu) gets called from identify_cpu. This populates the
858 * information related to logical execution units in per_cpu_data structure.
859 */
860 void __devinit
861 identify_siblings(struct cpuinfo_ia64 *c)
862 {
863 s64 status;
864 u16 pltid;
865 u64 proc_fixed_addr;
866 int count, i;
867 pal_logical_to_physical_t info;
869 if (smp_num_cpucores == 1 && smp_num_siblings == 1)
870 return;
872 if ((status = ia64_pal_logical_to_phys(0, &info)) != PAL_STATUS_SUCCESS) {
873 printk(KERN_ERR "ia64_pal_logical_to_phys failed with %ld\n",
874 status);
875 return;
876 }
877 if ((status = ia64_sal_physical_id_info(&pltid)) != PAL_STATUS_SUCCESS) {
878 printk(KERN_ERR "ia64_sal_pltid failed with %ld\n", status);
879 return;
880 }
881 if ((status = ia64_pal_fixed_addr(&proc_fixed_addr)) != PAL_STATUS_SUCCESS) {
882 printk(KERN_ERR "ia64_pal_fixed_addr failed with %ld\n", status);
883 return;
884 }
886 c->socket_id = (pltid << 8) | info.overview_ppid;
887 c->cores_per_socket = info.overview_cpp;
888 c->threads_per_core = info.overview_tpc;
889 count = c->num_log = info.overview_num_log;
891 /* If the thread and core id information is already cached, then
892 * we will simply update cpu_info and return. Otherwise, we will
893 * do the PAL calls and cache core and thread id's of all the siblings.
894 */
895 if (check_for_new_socket(proc_fixed_addr, c))
896 return;
898 for (i = 0; i < count; i++) {
899 int index;
901 if (i && (status = ia64_pal_logical_to_phys(i, &info))
902 != PAL_STATUS_SUCCESS) {
903 printk(KERN_ERR "ia64_pal_logical_to_phys failed"
904 " with %ld\n", status);
905 return;
906 }
907 if (info.log2_la == proc_fixed_addr) {
908 c->core_id = info.log1_cid;
909 c->thread_id = info.log1_tid;
910 }
912 index = check_for_mtinfo_index();
913 /* We will not do the mt_info caching optimization in this case.
914 */
915 if (index < 0)
916 continue;
918 mt_info[index].valid = 1;
919 mt_info[index].socket_id = c->socket_id;
920 mt_info[index].core_id = info.log1_cid;
921 mt_info[index].thread_id = info.log1_tid;
922 mt_info[index].proc_fixed_addr = info.log2_la;
923 }
924 }
925 #endif /* CONFIG_SMP ifdef XEN */