Merge master.kernel.org:/home/rmk/linux-2.6-arm
[linux-2.6] / arch / sparc64 / kernel / smp.c
1 /* smp.c: Sparc64 SMP support.
2  *
3  * Copyright (C) 1997 David S. Miller (davem@caip.rutgers.edu)
4  */
5
6 #include <linux/module.h>
7 #include <linux/kernel.h>
8 #include <linux/sched.h>
9 #include <linux/mm.h>
10 #include <linux/pagemap.h>
11 #include <linux/threads.h>
12 #include <linux/smp.h>
13 #include <linux/smp_lock.h>
14 #include <linux/interrupt.h>
15 #include <linux/kernel_stat.h>
16 #include <linux/delay.h>
17 #include <linux/init.h>
18 #include <linux/spinlock.h>
19 #include <linux/fs.h>
20 #include <linux/seq_file.h>
21 #include <linux/cache.h>
22 #include <linux/jiffies.h>
23 #include <linux/profile.h>
24 #include <linux/bootmem.h>
25
26 #include <asm/head.h>
27 #include <asm/ptrace.h>
28 #include <asm/atomic.h>
29 #include <asm/tlbflush.h>
30 #include <asm/mmu_context.h>
31 #include <asm/cpudata.h>
32
33 #include <asm/irq.h>
34 #include <asm/page.h>
35 #include <asm/pgtable.h>
36 #include <asm/oplib.h>
37 #include <asm/uaccess.h>
38 #include <asm/timer.h>
39 #include <asm/starfire.h>
40 #include <asm/tlb.h>
41
42 extern int linux_num_cpus;
43 extern void calibrate_delay(void);
44
45 /* Please don't make this stuff initdata!!!  --DaveM */
46 static unsigned char boot_cpu_id;
47
48 cpumask_t cpu_online_map __read_mostly = CPU_MASK_NONE;
49 cpumask_t phys_cpu_present_map __read_mostly = CPU_MASK_NONE;
50 static cpumask_t smp_commenced_mask;
51 static cpumask_t cpu_callout_map;
52
53 void smp_info(struct seq_file *m)
54 {
55         int i;
56         
57         seq_printf(m, "State:\n");
58         for (i = 0; i < NR_CPUS; i++) {
59                 if (cpu_online(i))
60                         seq_printf(m,
61                                    "CPU%d:\t\tonline\n", i);
62         }
63 }
64
65 void smp_bogo(struct seq_file *m)
66 {
67         int i;
68         
69         for (i = 0; i < NR_CPUS; i++)
70                 if (cpu_online(i))
71                         seq_printf(m,
72                                    "Cpu%dBogo\t: %lu.%02lu\n"
73                                    "Cpu%dClkTck\t: %016lx\n",
74                                    i, cpu_data(i).udelay_val / (500000/HZ),
75                                    (cpu_data(i).udelay_val / (5000/HZ)) % 100,
76                                    i, cpu_data(i).clock_tick);
77 }
78
79 void __init smp_store_cpu_info(int id)
80 {
81         int cpu_node;
82
83         /* multiplier and counter set by
84            smp_setup_percpu_timer()  */
85         cpu_data(id).udelay_val                 = loops_per_jiffy;
86
87         cpu_find_by_mid(id, &cpu_node);
88         cpu_data(id).clock_tick = prom_getintdefault(cpu_node,
89                                                      "clock-frequency", 0);
90
91         cpu_data(id).pgcache_size               = 0;
92         cpu_data(id).pte_cache[0]               = NULL;
93         cpu_data(id).pte_cache[1]               = NULL;
94         cpu_data(id).pgd_cache                  = NULL;
95         cpu_data(id).idle_volume                = 1;
96
97         cpu_data(id).dcache_size = prom_getintdefault(cpu_node, "dcache-size",
98                                                       16 * 1024);
99         cpu_data(id).dcache_line_size =
100                 prom_getintdefault(cpu_node, "dcache-line-size", 32);
101         cpu_data(id).icache_size = prom_getintdefault(cpu_node, "icache-size",
102                                                       16 * 1024);
103         cpu_data(id).icache_line_size =
104                 prom_getintdefault(cpu_node, "icache-line-size", 32);
105         cpu_data(id).ecache_size = prom_getintdefault(cpu_node, "ecache-size",
106                                                       4 * 1024 * 1024);
107         cpu_data(id).ecache_line_size =
108                 prom_getintdefault(cpu_node, "ecache-line-size", 64);
109         printk("CPU[%d]: Caches "
110                "D[sz(%d):line_sz(%d)] "
111                "I[sz(%d):line_sz(%d)] "
112                "E[sz(%d):line_sz(%d)]\n",
113                id,
114                cpu_data(id).dcache_size, cpu_data(id).dcache_line_size,
115                cpu_data(id).icache_size, cpu_data(id).icache_line_size,
116                cpu_data(id).ecache_size, cpu_data(id).ecache_line_size);
117 }
118
119 static void smp_setup_percpu_timer(void);
120
121 static volatile unsigned long callin_flag = 0;
122
123 extern void inherit_locked_prom_mappings(int save_p);
124
125 static inline void cpu_setup_percpu_base(unsigned long cpu_id)
126 {
127         __asm__ __volatile__("mov       %0, %%g5\n\t"
128                              "stxa      %0, [%1] %2\n\t"
129                              "membar    #Sync"
130                              : /* no outputs */
131                              : "r" (__per_cpu_offset(cpu_id)),
132                                "r" (TSB_REG), "i" (ASI_IMMU));
133 }
134
135 void __init smp_callin(void)
136 {
137         int cpuid = hard_smp_processor_id();
138
139         inherit_locked_prom_mappings(0);
140
141         __flush_tlb_all();
142
143         cpu_setup_percpu_base(cpuid);
144
145         smp_setup_percpu_timer();
146
147         if (cheetah_pcache_forced_on)
148                 cheetah_enable_pcache();
149
150         local_irq_enable();
151
152         calibrate_delay();
153         smp_store_cpu_info(cpuid);
154         callin_flag = 1;
155         __asm__ __volatile__("membar #Sync\n\t"
156                              "flush  %%g6" : : : "memory");
157
158         /* Clear this or we will die instantly when we
159          * schedule back to this idler...
160          */
161         current_thread_info()->new_child = 0;
162
163         /* Attach to the address space of init_task. */
164         atomic_inc(&init_mm.mm_count);
165         current->active_mm = &init_mm;
166
167         while (!cpu_isset(cpuid, smp_commenced_mask))
168                 rmb();
169
170         cpu_set(cpuid, cpu_online_map);
171
172         /* idle thread is expected to have preempt disabled */
173         preempt_disable();
174 }
175
176 void cpu_panic(void)
177 {
178         printk("CPU[%d]: Returns from cpu_idle!\n", smp_processor_id());
179         panic("SMP bolixed\n");
180 }
181
182 static unsigned long current_tick_offset __read_mostly;
183
184 /* This tick register synchronization scheme is taken entirely from
185  * the ia64 port, see arch/ia64/kernel/smpboot.c for details and credit.
186  *
187  * The only change I've made is to rework it so that the master
188  * initiates the synchonization instead of the slave. -DaveM
189  */
190
191 #define MASTER  0
192 #define SLAVE   (SMP_CACHE_BYTES/sizeof(unsigned long))
193
194 #define NUM_ROUNDS      64      /* magic value */
195 #define NUM_ITERS       5       /* likewise */
196
197 static DEFINE_SPINLOCK(itc_sync_lock);
198 static unsigned long go[SLAVE + 1];
199
200 #define DEBUG_TICK_SYNC 0
201
202 static inline long get_delta (long *rt, long *master)
203 {
204         unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
205         unsigned long tcenter, t0, t1, tm;
206         unsigned long i;
207
208         for (i = 0; i < NUM_ITERS; i++) {
209                 t0 = tick_ops->get_tick();
210                 go[MASTER] = 1;
211                 membar_storeload();
212                 while (!(tm = go[SLAVE]))
213                         rmb();
214                 go[SLAVE] = 0;
215                 wmb();
216                 t1 = tick_ops->get_tick();
217
218                 if (t1 - t0 < best_t1 - best_t0)
219                         best_t0 = t0, best_t1 = t1, best_tm = tm;
220         }
221
222         *rt = best_t1 - best_t0;
223         *master = best_tm - best_t0;
224
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 }
231
232 void smp_synchronize_tick_client(void)
233 {
234         long i, delta, adj, adjust_latency = 0, done = 0;
235         unsigned long flags, rt, master_time_stamp, bound;
236 #if DEBUG_TICK_SYNC
237         struct {
238                 long rt;        /* roundtrip time */
239                 long master;    /* master's timestamp */
240                 long diff;      /* difference between midpoint and master's timestamp */
241                 long lat;       /* estimate of itc adjustment latency */
242         } t[NUM_ROUNDS];
243 #endif
244
245         go[MASTER] = 1;
246
247         while (go[MASTER])
248                 rmb();
249
250         local_irq_save(flags);
251         {
252                 for (i = 0; i < NUM_ROUNDS; i++) {
253                         delta = get_delta(&rt, &master_time_stamp);
254                         if (delta == 0) {
255                                 done = 1;       /* let's lock on to this... */
256                                 bound = rt;
257                         }
258
259                         if (!done) {
260                                 if (i > 0) {
261                                         adjust_latency += -delta;
262                                         adj = -delta + adjust_latency/4;
263                                 } else
264                                         adj = -delta;
265
266                                 tick_ops->add_tick(adj, current_tick_offset);
267                         }
268 #if DEBUG_TICK_SYNC
269                         t[i].rt = rt;
270                         t[i].master = master_time_stamp;
271                         t[i].diff = delta;
272                         t[i].lat = adjust_latency/4;
273 #endif
274                 }
275         }
276         local_irq_restore(flags);
277
278 #if DEBUG_TICK_SYNC
279         for (i = 0; i < NUM_ROUNDS; i++)
280                 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
281                        t[i].rt, t[i].master, t[i].diff, t[i].lat);
282 #endif
283
284         printk(KERN_INFO "CPU %d: synchronized TICK with master CPU (last diff %ld cycles,"
285                "maxerr %lu cycles)\n", smp_processor_id(), delta, rt);
286 }
287
288 static void smp_start_sync_tick_client(int cpu);
289
290 static void smp_synchronize_one_tick(int cpu)
291 {
292         unsigned long flags, i;
293
294         go[MASTER] = 0;
295
296         smp_start_sync_tick_client(cpu);
297
298         /* wait for client to be ready */
299         while (!go[MASTER])
300                 rmb();
301
302         /* now let the client proceed into his loop */
303         go[MASTER] = 0;
304         membar_storeload();
305
306         spin_lock_irqsave(&itc_sync_lock, flags);
307         {
308                 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; i++) {
309                         while (!go[MASTER])
310                                 rmb();
311                         go[MASTER] = 0;
312                         wmb();
313                         go[SLAVE] = tick_ops->get_tick();
314                         membar_storeload();
315                 }
316         }
317         spin_unlock_irqrestore(&itc_sync_lock, flags);
318 }
319
320 extern unsigned long sparc64_cpu_startup;
321
322 /* The OBP cpu startup callback truncates the 3rd arg cookie to
323  * 32-bits (I think) so to be safe we have it read the pointer
324  * contained here so we work on >4GB machines. -DaveM
325  */
326 static struct thread_info *cpu_new_thread = NULL;
327
328 static int __devinit smp_boot_one_cpu(unsigned int cpu)
329 {
330         unsigned long entry =
331                 (unsigned long)(&sparc64_cpu_startup);
332         unsigned long cookie =
333                 (unsigned long)(&cpu_new_thread);
334         struct task_struct *p;
335         int timeout, ret, cpu_node;
336
337         p = fork_idle(cpu);
338         callin_flag = 0;
339         cpu_new_thread = p->thread_info;
340         cpu_set(cpu, cpu_callout_map);
341
342         cpu_find_by_mid(cpu, &cpu_node);
343         prom_startcpu(cpu_node, entry, cookie);
344
345         for (timeout = 0; timeout < 5000000; timeout++) {
346                 if (callin_flag)
347                         break;
348                 udelay(100);
349         }
350         if (callin_flag) {
351                 ret = 0;
352         } else {
353                 printk("Processor %d is stuck.\n", cpu);
354                 cpu_clear(cpu, cpu_callout_map);
355                 ret = -ENODEV;
356         }
357         cpu_new_thread = NULL;
358
359         return ret;
360 }
361
362 static void spitfire_xcall_helper(u64 data0, u64 data1, u64 data2, u64 pstate, unsigned long cpu)
363 {
364         u64 result, target;
365         int stuck, tmp;
366
367         if (this_is_starfire) {
368                 /* map to real upaid */
369                 cpu = (((cpu & 0x3c) << 1) |
370                         ((cpu & 0x40) >> 4) |
371                         (cpu & 0x3));
372         }
373
374         target = (cpu << 14) | 0x70;
375 again:
376         /* Ok, this is the real Spitfire Errata #54.
377          * One must read back from a UDB internal register
378          * after writes to the UDB interrupt dispatch, but
379          * before the membar Sync for that write.
380          * So we use the high UDB control register (ASI 0x7f,
381          * ADDR 0x20) for the dummy read. -DaveM
382          */
383         tmp = 0x40;
384         __asm__ __volatile__(
385         "wrpr   %1, %2, %%pstate\n\t"
386         "stxa   %4, [%0] %3\n\t"
387         "stxa   %5, [%0+%8] %3\n\t"
388         "add    %0, %8, %0\n\t"
389         "stxa   %6, [%0+%8] %3\n\t"
390         "membar #Sync\n\t"
391         "stxa   %%g0, [%7] %3\n\t"
392         "membar #Sync\n\t"
393         "mov    0x20, %%g1\n\t"
394         "ldxa   [%%g1] 0x7f, %%g0\n\t"
395         "membar #Sync"
396         : "=r" (tmp)
397         : "r" (pstate), "i" (PSTATE_IE), "i" (ASI_INTR_W),
398           "r" (data0), "r" (data1), "r" (data2), "r" (target),
399           "r" (0x10), "0" (tmp)
400         : "g1");
401
402         /* NOTE: PSTATE_IE is still clear. */
403         stuck = 100000;
404         do {
405                 __asm__ __volatile__("ldxa [%%g0] %1, %0"
406                         : "=r" (result)
407                         : "i" (ASI_INTR_DISPATCH_STAT));
408                 if (result == 0) {
409                         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
410                                              : : "r" (pstate));
411                         return;
412                 }
413                 stuck -= 1;
414                 if (stuck == 0)
415                         break;
416         } while (result & 0x1);
417         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
418                              : : "r" (pstate));
419         if (stuck == 0) {
420                 printk("CPU[%d]: mondo stuckage result[%016lx]\n",
421                        smp_processor_id(), result);
422         } else {
423                 udelay(2);
424                 goto again;
425         }
426 }
427
428 static __inline__ void spitfire_xcall_deliver(u64 data0, u64 data1, u64 data2, cpumask_t mask)
429 {
430         u64 pstate;
431         int i;
432
433         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
434         for_each_cpu_mask(i, mask)
435                 spitfire_xcall_helper(data0, data1, data2, pstate, i);
436 }
437
438 /* Cheetah now allows to send the whole 64-bytes of data in the interrupt
439  * packet, but we have no use for that.  However we do take advantage of
440  * the new pipelining feature (ie. dispatch to multiple cpus simultaneously).
441  */
442 static void cheetah_xcall_deliver(u64 data0, u64 data1, u64 data2, cpumask_t mask)
443 {
444         u64 pstate, ver;
445         int nack_busy_id, is_jalapeno;
446
447         if (cpus_empty(mask))
448                 return;
449
450         /* Unfortunately, someone at Sun had the brilliant idea to make the
451          * busy/nack fields hard-coded by ITID number for this Ultra-III
452          * derivative processor.
453          */
454         __asm__ ("rdpr %%ver, %0" : "=r" (ver));
455         is_jalapeno = ((ver >> 32) == 0x003e0016);
456
457         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
458
459 retry:
460         __asm__ __volatile__("wrpr %0, %1, %%pstate\n\t"
461                              : : "r" (pstate), "i" (PSTATE_IE));
462
463         /* Setup the dispatch data registers. */
464         __asm__ __volatile__("stxa      %0, [%3] %6\n\t"
465                              "stxa      %1, [%4] %6\n\t"
466                              "stxa      %2, [%5] %6\n\t"
467                              "membar    #Sync\n\t"
468                              : /* no outputs */
469                              : "r" (data0), "r" (data1), "r" (data2),
470                                "r" (0x40), "r" (0x50), "r" (0x60),
471                                "i" (ASI_INTR_W));
472
473         nack_busy_id = 0;
474         {
475                 int i;
476
477                 for_each_cpu_mask(i, mask) {
478                         u64 target = (i << 14) | 0x70;
479
480                         if (!is_jalapeno)
481                                 target |= (nack_busy_id << 24);
482                         __asm__ __volatile__(
483                                 "stxa   %%g0, [%0] %1\n\t"
484                                 "membar #Sync\n\t"
485                                 : /* no outputs */
486                                 : "r" (target), "i" (ASI_INTR_W));
487                         nack_busy_id++;
488                 }
489         }
490
491         /* Now, poll for completion. */
492         {
493                 u64 dispatch_stat;
494                 long stuck;
495
496                 stuck = 100000 * nack_busy_id;
497                 do {
498                         __asm__ __volatile__("ldxa      [%%g0] %1, %0"
499                                              : "=r" (dispatch_stat)
500                                              : "i" (ASI_INTR_DISPATCH_STAT));
501                         if (dispatch_stat == 0UL) {
502                                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
503                                                      : : "r" (pstate));
504                                 return;
505                         }
506                         if (!--stuck)
507                                 break;
508                 } while (dispatch_stat & 0x5555555555555555UL);
509
510                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
511                                      : : "r" (pstate));
512
513                 if ((dispatch_stat & ~(0x5555555555555555UL)) == 0) {
514                         /* Busy bits will not clear, continue instead
515                          * of freezing up on this cpu.
516                          */
517                         printk("CPU[%d]: mondo stuckage result[%016lx]\n",
518                                smp_processor_id(), dispatch_stat);
519                 } else {
520                         int i, this_busy_nack = 0;
521
522                         /* Delay some random time with interrupts enabled
523                          * to prevent deadlock.
524                          */
525                         udelay(2 * nack_busy_id);
526
527                         /* Clear out the mask bits for cpus which did not
528                          * NACK us.
529                          */
530                         for_each_cpu_mask(i, mask) {
531                                 u64 check_mask;
532
533                                 if (is_jalapeno)
534                                         check_mask = (0x2UL << (2*i));
535                                 else
536                                         check_mask = (0x2UL <<
537                                                       this_busy_nack);
538                                 if ((dispatch_stat & check_mask) == 0)
539                                         cpu_clear(i, mask);
540                                 this_busy_nack += 2;
541                         }
542
543                         goto retry;
544                 }
545         }
546 }
547
548 /* Send cross call to all processors mentioned in MASK
549  * except self.
550  */
551 static void smp_cross_call_masked(unsigned long *func, u32 ctx, u64 data1, u64 data2, cpumask_t mask)
552 {
553         u64 data0 = (((u64)ctx)<<32 | (((u64)func) & 0xffffffff));
554         int this_cpu = get_cpu();
555
556         cpus_and(mask, mask, cpu_online_map);
557         cpu_clear(this_cpu, mask);
558
559         if (tlb_type == spitfire)
560                 spitfire_xcall_deliver(data0, data1, data2, mask);
561         else
562                 cheetah_xcall_deliver(data0, data1, data2, mask);
563         /* NOTE: Caller runs local copy on master. */
564
565         put_cpu();
566 }
567
568 extern unsigned long xcall_sync_tick;
569
570 static void smp_start_sync_tick_client(int cpu)
571 {
572         cpumask_t mask = cpumask_of_cpu(cpu);
573
574         smp_cross_call_masked(&xcall_sync_tick,
575                               0, 0, 0, mask);
576 }
577
578 /* Send cross call to all processors except self. */
579 #define smp_cross_call(func, ctx, data1, data2) \
580         smp_cross_call_masked(func, ctx, data1, data2, cpu_online_map)
581
582 struct call_data_struct {
583         void (*func) (void *info);
584         void *info;
585         atomic_t finished;
586         int wait;
587 };
588
589 static DEFINE_SPINLOCK(call_lock);
590 static struct call_data_struct *call_data;
591
592 extern unsigned long xcall_call_function;
593
594 /*
595  * You must not call this function with disabled interrupts or from a
596  * hardware interrupt handler or from a bottom half handler.
597  */
598 int smp_call_function(void (*func)(void *info), void *info,
599                       int nonatomic, int wait)
600 {
601         struct call_data_struct data;
602         int cpus = num_online_cpus() - 1;
603         long timeout;
604
605         if (!cpus)
606                 return 0;
607
608         /* Can deadlock when called with interrupts disabled */
609         WARN_ON(irqs_disabled());
610
611         data.func = func;
612         data.info = info;
613         atomic_set(&data.finished, 0);
614         data.wait = wait;
615
616         spin_lock(&call_lock);
617
618         call_data = &data;
619
620         smp_cross_call(&xcall_call_function, 0, 0, 0);
621
622         /* 
623          * Wait for other cpus to complete function or at
624          * least snap the call data.
625          */
626         timeout = 1000000;
627         while (atomic_read(&data.finished) != cpus) {
628                 if (--timeout <= 0)
629                         goto out_timeout;
630                 barrier();
631                 udelay(1);
632         }
633
634         spin_unlock(&call_lock);
635
636         return 0;
637
638 out_timeout:
639         spin_unlock(&call_lock);
640         printk("XCALL: Remote cpus not responding, ncpus=%ld finished=%ld\n",
641                (long) num_online_cpus() - 1L,
642                (long) atomic_read(&data.finished));
643         return 0;
644 }
645
646 void smp_call_function_client(int irq, struct pt_regs *regs)
647 {
648         void (*func) (void *info) = call_data->func;
649         void *info = call_data->info;
650
651         clear_softint(1 << irq);
652         if (call_data->wait) {
653                 /* let initiator proceed only after completion */
654                 func(info);
655                 atomic_inc(&call_data->finished);
656         } else {
657                 /* let initiator proceed after getting data */
658                 atomic_inc(&call_data->finished);
659                 func(info);
660         }
661 }
662
663 extern unsigned long xcall_flush_tlb_mm;
664 extern unsigned long xcall_flush_tlb_pending;
665 extern unsigned long xcall_flush_tlb_kernel_range;
666 extern unsigned long xcall_flush_tlb_all_spitfire;
667 extern unsigned long xcall_flush_tlb_all_cheetah;
668 extern unsigned long xcall_report_regs;
669 extern unsigned long xcall_receive_signal;
670
671 #ifdef DCACHE_ALIASING_POSSIBLE
672 extern unsigned long xcall_flush_dcache_page_cheetah;
673 #endif
674 extern unsigned long xcall_flush_dcache_page_spitfire;
675
676 #ifdef CONFIG_DEBUG_DCFLUSH
677 extern atomic_t dcpage_flushes;
678 extern atomic_t dcpage_flushes_xcall;
679 #endif
680
681 static __inline__ void __local_flush_dcache_page(struct page *page)
682 {
683 #ifdef DCACHE_ALIASING_POSSIBLE
684         __flush_dcache_page(page_address(page),
685                             ((tlb_type == spitfire) &&
686                              page_mapping(page) != NULL));
687 #else
688         if (page_mapping(page) != NULL &&
689             tlb_type == spitfire)
690                 __flush_icache_page(__pa(page_address(page)));
691 #endif
692 }
693
694 void smp_flush_dcache_page_impl(struct page *page, int cpu)
695 {
696         cpumask_t mask = cpumask_of_cpu(cpu);
697         int this_cpu = get_cpu();
698
699 #ifdef CONFIG_DEBUG_DCFLUSH
700         atomic_inc(&dcpage_flushes);
701 #endif
702         if (cpu == this_cpu) {
703                 __local_flush_dcache_page(page);
704         } else if (cpu_online(cpu)) {
705                 void *pg_addr = page_address(page);
706                 u64 data0;
707
708                 if (tlb_type == spitfire) {
709                         data0 =
710                                 ((u64)&xcall_flush_dcache_page_spitfire);
711                         if (page_mapping(page) != NULL)
712                                 data0 |= ((u64)1 << 32);
713                         spitfire_xcall_deliver(data0,
714                                                __pa(pg_addr),
715                                                (u64) pg_addr,
716                                                mask);
717                 } else {
718 #ifdef DCACHE_ALIASING_POSSIBLE
719                         data0 =
720                                 ((u64)&xcall_flush_dcache_page_cheetah);
721                         cheetah_xcall_deliver(data0,
722                                               __pa(pg_addr),
723                                               0, mask);
724 #endif
725                 }
726 #ifdef CONFIG_DEBUG_DCFLUSH
727                 atomic_inc(&dcpage_flushes_xcall);
728 #endif
729         }
730
731         put_cpu();
732 }
733
734 void flush_dcache_page_all(struct mm_struct *mm, struct page *page)
735 {
736         void *pg_addr = page_address(page);
737         cpumask_t mask = cpu_online_map;
738         u64 data0;
739         int this_cpu = get_cpu();
740
741         cpu_clear(this_cpu, mask);
742
743 #ifdef CONFIG_DEBUG_DCFLUSH
744         atomic_inc(&dcpage_flushes);
745 #endif
746         if (cpus_empty(mask))
747                 goto flush_self;
748         if (tlb_type == spitfire) {
749                 data0 = ((u64)&xcall_flush_dcache_page_spitfire);
750                 if (page_mapping(page) != NULL)
751                         data0 |= ((u64)1 << 32);
752                 spitfire_xcall_deliver(data0,
753                                        __pa(pg_addr),
754                                        (u64) pg_addr,
755                                        mask);
756         } else {
757 #ifdef DCACHE_ALIASING_POSSIBLE
758                 data0 = ((u64)&xcall_flush_dcache_page_cheetah);
759                 cheetah_xcall_deliver(data0,
760                                       __pa(pg_addr),
761                                       0, mask);
762 #endif
763         }
764 #ifdef CONFIG_DEBUG_DCFLUSH
765         atomic_inc(&dcpage_flushes_xcall);
766 #endif
767  flush_self:
768         __local_flush_dcache_page(page);
769
770         put_cpu();
771 }
772
773 void smp_receive_signal(int cpu)
774 {
775         cpumask_t mask = cpumask_of_cpu(cpu);
776
777         if (cpu_online(cpu)) {
778                 u64 data0 = (((u64)&xcall_receive_signal) & 0xffffffff);
779
780                 if (tlb_type == spitfire)
781                         spitfire_xcall_deliver(data0, 0, 0, mask);
782                 else
783                         cheetah_xcall_deliver(data0, 0, 0, mask);
784         }
785 }
786
787 void smp_receive_signal_client(int irq, struct pt_regs *regs)
788 {
789         /* Just return, rtrap takes care of the rest. */
790         clear_softint(1 << irq);
791 }
792
793 void smp_report_regs(void)
794 {
795         smp_cross_call(&xcall_report_regs, 0, 0, 0);
796 }
797
798 void smp_flush_tlb_all(void)
799 {
800         if (tlb_type == spitfire)
801                 smp_cross_call(&xcall_flush_tlb_all_spitfire, 0, 0, 0);
802         else
803                 smp_cross_call(&xcall_flush_tlb_all_cheetah, 0, 0, 0);
804         __flush_tlb_all();
805 }
806
807 /* We know that the window frames of the user have been flushed
808  * to the stack before we get here because all callers of us
809  * are flush_tlb_*() routines, and these run after flush_cache_*()
810  * which performs the flushw.
811  *
812  * The SMP TLB coherency scheme we use works as follows:
813  *
814  * 1) mm->cpu_vm_mask is a bit mask of which cpus an address
815  *    space has (potentially) executed on, this is the heuristic
816  *    we use to avoid doing cross calls.
817  *
818  *    Also, for flushing from kswapd and also for clones, we
819  *    use cpu_vm_mask as the list of cpus to make run the TLB.
820  *
821  * 2) TLB context numbers are shared globally across all processors
822  *    in the system, this allows us to play several games to avoid
823  *    cross calls.
824  *
825  *    One invariant is that when a cpu switches to a process, and
826  *    that processes tsk->active_mm->cpu_vm_mask does not have the
827  *    current cpu's bit set, that tlb context is flushed locally.
828  *
829  *    If the address space is non-shared (ie. mm->count == 1) we avoid
830  *    cross calls when we want to flush the currently running process's
831  *    tlb state.  This is done by clearing all cpu bits except the current
832  *    processor's in current->active_mm->cpu_vm_mask and performing the
833  *    flush locally only.  This will force any subsequent cpus which run
834  *    this task to flush the context from the local tlb if the process
835  *    migrates to another cpu (again).
836  *
837  * 3) For shared address spaces (threads) and swapping we bite the
838  *    bullet for most cases and perform the cross call (but only to
839  *    the cpus listed in cpu_vm_mask).
840  *
841  *    The performance gain from "optimizing" away the cross call for threads is
842  *    questionable (in theory the big win for threads is the massive sharing of
843  *    address space state across processors).
844  */
845
846 /* This currently is only used by the hugetlb arch pre-fault
847  * hook on UltraSPARC-III+ and later when changing the pagesize
848  * bits of the context register for an address space.
849  */
850 void smp_flush_tlb_mm(struct mm_struct *mm)
851 {
852         u32 ctx = CTX_HWBITS(mm->context);
853         int cpu = get_cpu();
854
855         if (atomic_read(&mm->mm_users) == 1) {
856                 mm->cpu_vm_mask = cpumask_of_cpu(cpu);
857                 goto local_flush_and_out;
858         }
859
860         smp_cross_call_masked(&xcall_flush_tlb_mm,
861                               ctx, 0, 0,
862                               mm->cpu_vm_mask);
863
864 local_flush_and_out:
865         __flush_tlb_mm(ctx, SECONDARY_CONTEXT);
866
867         put_cpu();
868 }
869
870 void smp_flush_tlb_pending(struct mm_struct *mm, unsigned long nr, unsigned long *vaddrs)
871 {
872         u32 ctx = CTX_HWBITS(mm->context);
873         int cpu = get_cpu();
874
875         if (mm == current->active_mm && atomic_read(&mm->mm_users) == 1)
876                 mm->cpu_vm_mask = cpumask_of_cpu(cpu);
877         else
878                 smp_cross_call_masked(&xcall_flush_tlb_pending,
879                                       ctx, nr, (unsigned long) vaddrs,
880                                       mm->cpu_vm_mask);
881
882         __flush_tlb_pending(ctx, nr, vaddrs);
883
884         put_cpu();
885 }
886
887 void smp_flush_tlb_kernel_range(unsigned long start, unsigned long end)
888 {
889         start &= PAGE_MASK;
890         end    = PAGE_ALIGN(end);
891         if (start != end) {
892                 smp_cross_call(&xcall_flush_tlb_kernel_range,
893                                0, start, end);
894
895                 __flush_tlb_kernel_range(start, end);
896         }
897 }
898
899 /* CPU capture. */
900 /* #define CAPTURE_DEBUG */
901 extern unsigned long xcall_capture;
902
903 static atomic_t smp_capture_depth = ATOMIC_INIT(0);
904 static atomic_t smp_capture_registry = ATOMIC_INIT(0);
905 static unsigned long penguins_are_doing_time;
906
907 void smp_capture(void)
908 {
909         int result = atomic_add_ret(1, &smp_capture_depth);
910
911         if (result == 1) {
912                 int ncpus = num_online_cpus();
913
914 #ifdef CAPTURE_DEBUG
915                 printk("CPU[%d]: Sending penguins to jail...",
916                        smp_processor_id());
917 #endif
918                 penguins_are_doing_time = 1;
919                 membar_storestore_loadstore();
920                 atomic_inc(&smp_capture_registry);
921                 smp_cross_call(&xcall_capture, 0, 0, 0);
922                 while (atomic_read(&smp_capture_registry) != ncpus)
923                         rmb();
924 #ifdef CAPTURE_DEBUG
925                 printk("done\n");
926 #endif
927         }
928 }
929
930 void smp_release(void)
931 {
932         if (atomic_dec_and_test(&smp_capture_depth)) {
933 #ifdef CAPTURE_DEBUG
934                 printk("CPU[%d]: Giving pardon to "
935                        "imprisoned penguins\n",
936                        smp_processor_id());
937 #endif
938                 penguins_are_doing_time = 0;
939                 membar_storeload_storestore();
940                 atomic_dec(&smp_capture_registry);
941         }
942 }
943
944 /* Imprisoned penguins run with %pil == 15, but PSTATE_IE set, so they
945  * can service tlb flush xcalls...
946  */
947 extern void prom_world(int);
948 extern void save_alternate_globals(unsigned long *);
949 extern void restore_alternate_globals(unsigned long *);
950 void smp_penguin_jailcell(int irq, struct pt_regs *regs)
951 {
952         unsigned long global_save[24];
953
954         clear_softint(1 << irq);
955
956         preempt_disable();
957
958         __asm__ __volatile__("flushw");
959         save_alternate_globals(global_save);
960         prom_world(1);
961         atomic_inc(&smp_capture_registry);
962         membar_storeload_storestore();
963         while (penguins_are_doing_time)
964                 rmb();
965         restore_alternate_globals(global_save);
966         atomic_dec(&smp_capture_registry);
967         prom_world(0);
968
969         preempt_enable();
970 }
971
972 #define prof_multiplier(__cpu)          cpu_data(__cpu).multiplier
973 #define prof_counter(__cpu)             cpu_data(__cpu).counter
974
975 void smp_percpu_timer_interrupt(struct pt_regs *regs)
976 {
977         unsigned long compare, tick, pstate;
978         int cpu = smp_processor_id();
979         int user = user_mode(regs);
980
981         /*
982          * Check for level 14 softint.
983          */
984         {
985                 unsigned long tick_mask = tick_ops->softint_mask;
986
987                 if (!(get_softint() & tick_mask)) {
988                         extern void handler_irq(int, struct pt_regs *);
989
990                         handler_irq(14, regs);
991                         return;
992                 }
993                 clear_softint(tick_mask);
994         }
995
996         do {
997                 profile_tick(CPU_PROFILING, regs);
998                 if (!--prof_counter(cpu)) {
999                         irq_enter();
1000
1001                         if (cpu == boot_cpu_id) {
1002                                 kstat_this_cpu.irqs[0]++;
1003                                 timer_tick_interrupt(regs);
1004                         }
1005
1006                         update_process_times(user);
1007
1008                         irq_exit();
1009
1010                         prof_counter(cpu) = prof_multiplier(cpu);
1011                 }
1012
1013                 /* Guarantee that the following sequences execute
1014                  * uninterrupted.
1015                  */
1016                 __asm__ __volatile__("rdpr      %%pstate, %0\n\t"
1017                                      "wrpr      %0, %1, %%pstate"
1018                                      : "=r" (pstate)
1019                                      : "i" (PSTATE_IE));
1020
1021                 compare = tick_ops->add_compare(current_tick_offset);
1022                 tick = tick_ops->get_tick();
1023
1024                 /* Restore PSTATE_IE. */
1025                 __asm__ __volatile__("wrpr      %0, 0x0, %%pstate"
1026                                      : /* no outputs */
1027                                      : "r" (pstate));
1028         } while (time_after_eq(tick, compare));
1029 }
1030
1031 static void __init smp_setup_percpu_timer(void)
1032 {
1033         int cpu = smp_processor_id();
1034         unsigned long pstate;
1035
1036         prof_counter(cpu) = prof_multiplier(cpu) = 1;
1037
1038         /* Guarantee that the following sequences execute
1039          * uninterrupted.
1040          */
1041         __asm__ __volatile__("rdpr      %%pstate, %0\n\t"
1042                              "wrpr      %0, %1, %%pstate"
1043                              : "=r" (pstate)
1044                              : "i" (PSTATE_IE));
1045
1046         tick_ops->init_tick(current_tick_offset);
1047
1048         /* Restore PSTATE_IE. */
1049         __asm__ __volatile__("wrpr      %0, 0x0, %%pstate"
1050                              : /* no outputs */
1051                              : "r" (pstate));
1052 }
1053
1054 void __init smp_tick_init(void)
1055 {
1056         boot_cpu_id = hard_smp_processor_id();
1057         current_tick_offset = timer_tick_offset;
1058
1059         cpu_set(boot_cpu_id, cpu_online_map);
1060         prof_counter(boot_cpu_id) = prof_multiplier(boot_cpu_id) = 1;
1061 }
1062
1063 /* /proc/profile writes can call this, don't __init it please. */
1064 static DEFINE_SPINLOCK(prof_setup_lock);
1065
1066 int setup_profiling_timer(unsigned int multiplier)
1067 {
1068         unsigned long flags;
1069         int i;
1070
1071         if ((!multiplier) || (timer_tick_offset / multiplier) < 1000)
1072                 return -EINVAL;
1073
1074         spin_lock_irqsave(&prof_setup_lock, flags);
1075         for (i = 0; i < NR_CPUS; i++)
1076                 prof_multiplier(i) = multiplier;
1077         current_tick_offset = (timer_tick_offset / multiplier);
1078         spin_unlock_irqrestore(&prof_setup_lock, flags);
1079
1080         return 0;
1081 }
1082
1083 void __init smp_prepare_cpus(unsigned int max_cpus)
1084 {
1085         int instance, mid;
1086
1087         instance = 0;
1088         while (!cpu_find_by_instance(instance, NULL, &mid)) {
1089                 if (mid < max_cpus)
1090                         cpu_set(mid, phys_cpu_present_map);
1091                 instance++;
1092         }
1093
1094         if (num_possible_cpus() > max_cpus) {
1095                 instance = 0;
1096                 while (!cpu_find_by_instance(instance, NULL, &mid)) {
1097                         if (mid != boot_cpu_id) {
1098                                 cpu_clear(mid, phys_cpu_present_map);
1099                                 if (num_possible_cpus() <= max_cpus)
1100                                         break;
1101                         }
1102                         instance++;
1103                 }
1104         }
1105
1106         smp_store_cpu_info(boot_cpu_id);
1107 }
1108
1109 void __devinit smp_prepare_boot_cpu(void)
1110 {
1111         if (hard_smp_processor_id() >= NR_CPUS) {
1112                 prom_printf("Serious problem, boot cpu id >= NR_CPUS\n");
1113                 prom_halt();
1114         }
1115
1116         current_thread_info()->cpu = hard_smp_processor_id();
1117
1118         cpu_set(smp_processor_id(), cpu_online_map);
1119         cpu_set(smp_processor_id(), phys_cpu_present_map);
1120 }
1121
1122 int __devinit __cpu_up(unsigned int cpu)
1123 {
1124         int ret = smp_boot_one_cpu(cpu);
1125
1126         if (!ret) {
1127                 cpu_set(cpu, smp_commenced_mask);
1128                 while (!cpu_isset(cpu, cpu_online_map))
1129                         mb();
1130                 if (!cpu_isset(cpu, cpu_online_map)) {
1131                         ret = -ENODEV;
1132                 } else {
1133                         smp_synchronize_one_tick(cpu);
1134                 }
1135         }
1136         return ret;
1137 }
1138
1139 void __init smp_cpus_done(unsigned int max_cpus)
1140 {
1141         unsigned long bogosum = 0;
1142         int i;
1143
1144         for (i = 0; i < NR_CPUS; i++) {
1145                 if (cpu_online(i))
1146                         bogosum += cpu_data(i).udelay_val;
1147         }
1148         printk("Total of %ld processors activated "
1149                "(%lu.%02lu BogoMIPS).\n",
1150                (long) num_online_cpus(),
1151                bogosum/(500000/HZ),
1152                (bogosum/(5000/HZ))%100);
1153 }
1154
1155 void smp_send_reschedule(int cpu)
1156 {
1157         smp_receive_signal(cpu);
1158 }
1159
1160 /* This is a nop because we capture all other cpus
1161  * anyways when making the PROM active.
1162  */
1163 void smp_send_stop(void)
1164 {
1165 }
1166
1167 unsigned long __per_cpu_base __read_mostly;
1168 unsigned long __per_cpu_shift __read_mostly;
1169
1170 EXPORT_SYMBOL(__per_cpu_base);
1171 EXPORT_SYMBOL(__per_cpu_shift);
1172
1173 void __init setup_per_cpu_areas(void)
1174 {
1175         unsigned long goal, size, i;
1176         char *ptr;
1177         /* Created by linker magic */
1178         extern char __per_cpu_start[], __per_cpu_end[];
1179
1180         /* Copy section for each CPU (we discard the original) */
1181         goal = ALIGN(__per_cpu_end - __per_cpu_start, PAGE_SIZE);
1182
1183 #ifdef CONFIG_MODULES
1184         if (goal < PERCPU_ENOUGH_ROOM)
1185                 goal = PERCPU_ENOUGH_ROOM;
1186 #endif
1187         __per_cpu_shift = 0;
1188         for (size = 1UL; size < goal; size <<= 1UL)
1189                 __per_cpu_shift++;
1190
1191         /* Make sure the resulting __per_cpu_base value
1192          * will fit in the 43-bit sign extended IMMU
1193          * TSB register.
1194          */
1195         ptr = __alloc_bootmem(size * NR_CPUS, PAGE_SIZE,
1196                               (unsigned long) __per_cpu_start);
1197
1198         __per_cpu_base = ptr - __per_cpu_start;
1199
1200         if ((__per_cpu_shift < PAGE_SHIFT) ||
1201             (__per_cpu_base & ~PAGE_MASK) ||
1202             (__per_cpu_base != (((long) __per_cpu_base << 20) >> 20))) {
1203                 prom_printf("PER_CPU: Invalid layout, "
1204                             "ptr[%p] shift[%lx] base[%lx]\n",
1205                             ptr, __per_cpu_shift, __per_cpu_base);
1206                 prom_halt();
1207         }
1208
1209         for (i = 0; i < NR_CPUS; i++, ptr += size)
1210                 memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start);
1211
1212         /* Finally, load in the boot cpu's base value.
1213          * We abuse the IMMU TSB register for trap handler
1214          * entry and exit loading of %g5.  That is why it
1215          * has to be page aligned.
1216          */
1217         cpu_setup_percpu_base(hard_smp_processor_id());
1218 }