Merge rsync://bughost.org/repos/ipw-delta/
[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
173 void cpu_panic(void)
174 {
175         printk("CPU[%d]: Returns from cpu_idle!\n", smp_processor_id());
176         panic("SMP bolixed\n");
177 }
178
179 static unsigned long current_tick_offset __read_mostly;
180
181 /* This tick register synchronization scheme is taken entirely from
182  * the ia64 port, see arch/ia64/kernel/smpboot.c for details and credit.
183  *
184  * The only change I've made is to rework it so that the master
185  * initiates the synchonization instead of the slave. -DaveM
186  */
187
188 #define MASTER  0
189 #define SLAVE   (SMP_CACHE_BYTES/sizeof(unsigned long))
190
191 #define NUM_ROUNDS      64      /* magic value */
192 #define NUM_ITERS       5       /* likewise */
193
194 static DEFINE_SPINLOCK(itc_sync_lock);
195 static unsigned long go[SLAVE + 1];
196
197 #define DEBUG_TICK_SYNC 0
198
199 static inline long get_delta (long *rt, long *master)
200 {
201         unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
202         unsigned long tcenter, t0, t1, tm;
203         unsigned long i;
204
205         for (i = 0; i < NUM_ITERS; i++) {
206                 t0 = tick_ops->get_tick();
207                 go[MASTER] = 1;
208                 membar_storeload();
209                 while (!(tm = go[SLAVE]))
210                         rmb();
211                 go[SLAVE] = 0;
212                 wmb();
213                 t1 = tick_ops->get_tick();
214
215                 if (t1 - t0 < best_t1 - best_t0)
216                         best_t0 = t0, best_t1 = t1, best_tm = tm;
217         }
218
219         *rt = best_t1 - best_t0;
220         *master = best_tm - best_t0;
221
222         /* average best_t0 and best_t1 without overflow: */
223         tcenter = (best_t0/2 + best_t1/2);
224         if (best_t0 % 2 + best_t1 % 2 == 2)
225                 tcenter++;
226         return tcenter - best_tm;
227 }
228
229 void smp_synchronize_tick_client(void)
230 {
231         long i, delta, adj, adjust_latency = 0, done = 0;
232         unsigned long flags, rt, master_time_stamp, bound;
233 #if DEBUG_TICK_SYNC
234         struct {
235                 long rt;        /* roundtrip time */
236                 long master;    /* master's timestamp */
237                 long diff;      /* difference between midpoint and master's timestamp */
238                 long lat;       /* estimate of itc adjustment latency */
239         } t[NUM_ROUNDS];
240 #endif
241
242         go[MASTER] = 1;
243
244         while (go[MASTER])
245                 rmb();
246
247         local_irq_save(flags);
248         {
249                 for (i = 0; i < NUM_ROUNDS; i++) {
250                         delta = get_delta(&rt, &master_time_stamp);
251                         if (delta == 0) {
252                                 done = 1;       /* let's lock on to this... */
253                                 bound = rt;
254                         }
255
256                         if (!done) {
257                                 if (i > 0) {
258                                         adjust_latency += -delta;
259                                         adj = -delta + adjust_latency/4;
260                                 } else
261                                         adj = -delta;
262
263                                 tick_ops->add_tick(adj, current_tick_offset);
264                         }
265 #if DEBUG_TICK_SYNC
266                         t[i].rt = rt;
267                         t[i].master = master_time_stamp;
268                         t[i].diff = delta;
269                         t[i].lat = adjust_latency/4;
270 #endif
271                 }
272         }
273         local_irq_restore(flags);
274
275 #if DEBUG_TICK_SYNC
276         for (i = 0; i < NUM_ROUNDS; i++)
277                 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
278                        t[i].rt, t[i].master, t[i].diff, t[i].lat);
279 #endif
280
281         printk(KERN_INFO "CPU %d: synchronized TICK with master CPU (last diff %ld cycles,"
282                "maxerr %lu cycles)\n", smp_processor_id(), delta, rt);
283 }
284
285 static void smp_start_sync_tick_client(int cpu);
286
287 static void smp_synchronize_one_tick(int cpu)
288 {
289         unsigned long flags, i;
290
291         go[MASTER] = 0;
292
293         smp_start_sync_tick_client(cpu);
294
295         /* wait for client to be ready */
296         while (!go[MASTER])
297                 rmb();
298
299         /* now let the client proceed into his loop */
300         go[MASTER] = 0;
301         membar_storeload();
302
303         spin_lock_irqsave(&itc_sync_lock, flags);
304         {
305                 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; i++) {
306                         while (!go[MASTER])
307                                 rmb();
308                         go[MASTER] = 0;
309                         wmb();
310                         go[SLAVE] = tick_ops->get_tick();
311                         membar_storeload();
312                 }
313         }
314         spin_unlock_irqrestore(&itc_sync_lock, flags);
315 }
316
317 extern unsigned long sparc64_cpu_startup;
318
319 /* The OBP cpu startup callback truncates the 3rd arg cookie to
320  * 32-bits (I think) so to be safe we have it read the pointer
321  * contained here so we work on >4GB machines. -DaveM
322  */
323 static struct thread_info *cpu_new_thread = NULL;
324
325 static int __devinit smp_boot_one_cpu(unsigned int cpu)
326 {
327         unsigned long entry =
328                 (unsigned long)(&sparc64_cpu_startup);
329         unsigned long cookie =
330                 (unsigned long)(&cpu_new_thread);
331         struct task_struct *p;
332         int timeout, ret, cpu_node;
333
334         p = fork_idle(cpu);
335         callin_flag = 0;
336         cpu_new_thread = p->thread_info;
337         cpu_set(cpu, cpu_callout_map);
338
339         cpu_find_by_mid(cpu, &cpu_node);
340         prom_startcpu(cpu_node, entry, cookie);
341
342         for (timeout = 0; timeout < 5000000; timeout++) {
343                 if (callin_flag)
344                         break;
345                 udelay(100);
346         }
347         if (callin_flag) {
348                 ret = 0;
349         } else {
350                 printk("Processor %d is stuck.\n", cpu);
351                 cpu_clear(cpu, cpu_callout_map);
352                 ret = -ENODEV;
353         }
354         cpu_new_thread = NULL;
355
356         return ret;
357 }
358
359 static void spitfire_xcall_helper(u64 data0, u64 data1, u64 data2, u64 pstate, unsigned long cpu)
360 {
361         u64 result, target;
362         int stuck, tmp;
363
364         if (this_is_starfire) {
365                 /* map to real upaid */
366                 cpu = (((cpu & 0x3c) << 1) |
367                         ((cpu & 0x40) >> 4) |
368                         (cpu & 0x3));
369         }
370
371         target = (cpu << 14) | 0x70;
372 again:
373         /* Ok, this is the real Spitfire Errata #54.
374          * One must read back from a UDB internal register
375          * after writes to the UDB interrupt dispatch, but
376          * before the membar Sync for that write.
377          * So we use the high UDB control register (ASI 0x7f,
378          * ADDR 0x20) for the dummy read. -DaveM
379          */
380         tmp = 0x40;
381         __asm__ __volatile__(
382         "wrpr   %1, %2, %%pstate\n\t"
383         "stxa   %4, [%0] %3\n\t"
384         "stxa   %5, [%0+%8] %3\n\t"
385         "add    %0, %8, %0\n\t"
386         "stxa   %6, [%0+%8] %3\n\t"
387         "membar #Sync\n\t"
388         "stxa   %%g0, [%7] %3\n\t"
389         "membar #Sync\n\t"
390         "mov    0x20, %%g1\n\t"
391         "ldxa   [%%g1] 0x7f, %%g0\n\t"
392         "membar #Sync"
393         : "=r" (tmp)
394         : "r" (pstate), "i" (PSTATE_IE), "i" (ASI_INTR_W),
395           "r" (data0), "r" (data1), "r" (data2), "r" (target),
396           "r" (0x10), "0" (tmp)
397         : "g1");
398
399         /* NOTE: PSTATE_IE is still clear. */
400         stuck = 100000;
401         do {
402                 __asm__ __volatile__("ldxa [%%g0] %1, %0"
403                         : "=r" (result)
404                         : "i" (ASI_INTR_DISPATCH_STAT));
405                 if (result == 0) {
406                         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
407                                              : : "r" (pstate));
408                         return;
409                 }
410                 stuck -= 1;
411                 if (stuck == 0)
412                         break;
413         } while (result & 0x1);
414         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
415                              : : "r" (pstate));
416         if (stuck == 0) {
417                 printk("CPU[%d]: mondo stuckage result[%016lx]\n",
418                        smp_processor_id(), result);
419         } else {
420                 udelay(2);
421                 goto again;
422         }
423 }
424
425 static __inline__ void spitfire_xcall_deliver(u64 data0, u64 data1, u64 data2, cpumask_t mask)
426 {
427         u64 pstate;
428         int i;
429
430         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
431         for_each_cpu_mask(i, mask)
432                 spitfire_xcall_helper(data0, data1, data2, pstate, i);
433 }
434
435 /* Cheetah now allows to send the whole 64-bytes of data in the interrupt
436  * packet, but we have no use for that.  However we do take advantage of
437  * the new pipelining feature (ie. dispatch to multiple cpus simultaneously).
438  */
439 static void cheetah_xcall_deliver(u64 data0, u64 data1, u64 data2, cpumask_t mask)
440 {
441         u64 pstate, ver;
442         int nack_busy_id, is_jalapeno;
443
444         if (cpus_empty(mask))
445                 return;
446
447         /* Unfortunately, someone at Sun had the brilliant idea to make the
448          * busy/nack fields hard-coded by ITID number for this Ultra-III
449          * derivative processor.
450          */
451         __asm__ ("rdpr %%ver, %0" : "=r" (ver));
452         is_jalapeno = ((ver >> 32) == 0x003e0016);
453
454         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
455
456 retry:
457         __asm__ __volatile__("wrpr %0, %1, %%pstate\n\t"
458                              : : "r" (pstate), "i" (PSTATE_IE));
459
460         /* Setup the dispatch data registers. */
461         __asm__ __volatile__("stxa      %0, [%3] %6\n\t"
462                              "stxa      %1, [%4] %6\n\t"
463                              "stxa      %2, [%5] %6\n\t"
464                              "membar    #Sync\n\t"
465                              : /* no outputs */
466                              : "r" (data0), "r" (data1), "r" (data2),
467                                "r" (0x40), "r" (0x50), "r" (0x60),
468                                "i" (ASI_INTR_W));
469
470         nack_busy_id = 0;
471         {
472                 int i;
473
474                 for_each_cpu_mask(i, mask) {
475                         u64 target = (i << 14) | 0x70;
476
477                         if (!is_jalapeno)
478                                 target |= (nack_busy_id << 24);
479                         __asm__ __volatile__(
480                                 "stxa   %%g0, [%0] %1\n\t"
481                                 "membar #Sync\n\t"
482                                 : /* no outputs */
483                                 : "r" (target), "i" (ASI_INTR_W));
484                         nack_busy_id++;
485                 }
486         }
487
488         /* Now, poll for completion. */
489         {
490                 u64 dispatch_stat;
491                 long stuck;
492
493                 stuck = 100000 * nack_busy_id;
494                 do {
495                         __asm__ __volatile__("ldxa      [%%g0] %1, %0"
496                                              : "=r" (dispatch_stat)
497                                              : "i" (ASI_INTR_DISPATCH_STAT));
498                         if (dispatch_stat == 0UL) {
499                                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
500                                                      : : "r" (pstate));
501                                 return;
502                         }
503                         if (!--stuck)
504                                 break;
505                 } while (dispatch_stat & 0x5555555555555555UL);
506
507                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
508                                      : : "r" (pstate));
509
510                 if ((dispatch_stat & ~(0x5555555555555555UL)) == 0) {
511                         /* Busy bits will not clear, continue instead
512                          * of freezing up on this cpu.
513                          */
514                         printk("CPU[%d]: mondo stuckage result[%016lx]\n",
515                                smp_processor_id(), dispatch_stat);
516                 } else {
517                         int i, this_busy_nack = 0;
518
519                         /* Delay some random time with interrupts enabled
520                          * to prevent deadlock.
521                          */
522                         udelay(2 * nack_busy_id);
523
524                         /* Clear out the mask bits for cpus which did not
525                          * NACK us.
526                          */
527                         for_each_cpu_mask(i, mask) {
528                                 u64 check_mask;
529
530                                 if (is_jalapeno)
531                                         check_mask = (0x2UL << (2*i));
532                                 else
533                                         check_mask = (0x2UL <<
534                                                       this_busy_nack);
535                                 if ((dispatch_stat & check_mask) == 0)
536                                         cpu_clear(i, mask);
537                                 this_busy_nack += 2;
538                         }
539
540                         goto retry;
541                 }
542         }
543 }
544
545 /* Send cross call to all processors mentioned in MASK
546  * except self.
547  */
548 static void smp_cross_call_masked(unsigned long *func, u32 ctx, u64 data1, u64 data2, cpumask_t mask)
549 {
550         u64 data0 = (((u64)ctx)<<32 | (((u64)func) & 0xffffffff));
551         int this_cpu = get_cpu();
552
553         cpus_and(mask, mask, cpu_online_map);
554         cpu_clear(this_cpu, mask);
555
556         if (tlb_type == spitfire)
557                 spitfire_xcall_deliver(data0, data1, data2, mask);
558         else
559                 cheetah_xcall_deliver(data0, data1, data2, mask);
560         /* NOTE: Caller runs local copy on master. */
561
562         put_cpu();
563 }
564
565 extern unsigned long xcall_sync_tick;
566
567 static void smp_start_sync_tick_client(int cpu)
568 {
569         cpumask_t mask = cpumask_of_cpu(cpu);
570
571         smp_cross_call_masked(&xcall_sync_tick,
572                               0, 0, 0, mask);
573 }
574
575 /* Send cross call to all processors except self. */
576 #define smp_cross_call(func, ctx, data1, data2) \
577         smp_cross_call_masked(func, ctx, data1, data2, cpu_online_map)
578
579 struct call_data_struct {
580         void (*func) (void *info);
581         void *info;
582         atomic_t finished;
583         int wait;
584 };
585
586 static DEFINE_SPINLOCK(call_lock);
587 static struct call_data_struct *call_data;
588
589 extern unsigned long xcall_call_function;
590
591 /*
592  * You must not call this function with disabled interrupts or from a
593  * hardware interrupt handler or from a bottom half handler.
594  */
595 int smp_call_function(void (*func)(void *info), void *info,
596                       int nonatomic, int wait)
597 {
598         struct call_data_struct data;
599         int cpus = num_online_cpus() - 1;
600         long timeout;
601
602         if (!cpus)
603                 return 0;
604
605         /* Can deadlock when called with interrupts disabled */
606         WARN_ON(irqs_disabled());
607
608         data.func = func;
609         data.info = info;
610         atomic_set(&data.finished, 0);
611         data.wait = wait;
612
613         spin_lock(&call_lock);
614
615         call_data = &data;
616
617         smp_cross_call(&xcall_call_function, 0, 0, 0);
618
619         /* 
620          * Wait for other cpus to complete function or at
621          * least snap the call data.
622          */
623         timeout = 1000000;
624         while (atomic_read(&data.finished) != cpus) {
625                 if (--timeout <= 0)
626                         goto out_timeout;
627                 barrier();
628                 udelay(1);
629         }
630
631         spin_unlock(&call_lock);
632
633         return 0;
634
635 out_timeout:
636         spin_unlock(&call_lock);
637         printk("XCALL: Remote cpus not responding, ncpus=%ld finished=%ld\n",
638                (long) num_online_cpus() - 1L,
639                (long) atomic_read(&data.finished));
640         return 0;
641 }
642
643 void smp_call_function_client(int irq, struct pt_regs *regs)
644 {
645         void (*func) (void *info) = call_data->func;
646         void *info = call_data->info;
647
648         clear_softint(1 << irq);
649         if (call_data->wait) {
650                 /* let initiator proceed only after completion */
651                 func(info);
652                 atomic_inc(&call_data->finished);
653         } else {
654                 /* let initiator proceed after getting data */
655                 atomic_inc(&call_data->finished);
656                 func(info);
657         }
658 }
659
660 extern unsigned long xcall_flush_tlb_mm;
661 extern unsigned long xcall_flush_tlb_pending;
662 extern unsigned long xcall_flush_tlb_kernel_range;
663 extern unsigned long xcall_flush_tlb_all_spitfire;
664 extern unsigned long xcall_flush_tlb_all_cheetah;
665 extern unsigned long xcall_report_regs;
666 extern unsigned long xcall_receive_signal;
667
668 #ifdef DCACHE_ALIASING_POSSIBLE
669 extern unsigned long xcall_flush_dcache_page_cheetah;
670 #endif
671 extern unsigned long xcall_flush_dcache_page_spitfire;
672
673 #ifdef CONFIG_DEBUG_DCFLUSH
674 extern atomic_t dcpage_flushes;
675 extern atomic_t dcpage_flushes_xcall;
676 #endif
677
678 static __inline__ void __local_flush_dcache_page(struct page *page)
679 {
680 #ifdef DCACHE_ALIASING_POSSIBLE
681         __flush_dcache_page(page_address(page),
682                             ((tlb_type == spitfire) &&
683                              page_mapping(page) != NULL));
684 #else
685         if (page_mapping(page) != NULL &&
686             tlb_type == spitfire)
687                 __flush_icache_page(__pa(page_address(page)));
688 #endif
689 }
690
691 void smp_flush_dcache_page_impl(struct page *page, int cpu)
692 {
693         cpumask_t mask = cpumask_of_cpu(cpu);
694         int this_cpu = get_cpu();
695
696 #ifdef CONFIG_DEBUG_DCFLUSH
697         atomic_inc(&dcpage_flushes);
698 #endif
699         if (cpu == this_cpu) {
700                 __local_flush_dcache_page(page);
701         } else if (cpu_online(cpu)) {
702                 void *pg_addr = page_address(page);
703                 u64 data0;
704
705                 if (tlb_type == spitfire) {
706                         data0 =
707                                 ((u64)&xcall_flush_dcache_page_spitfire);
708                         if (page_mapping(page) != NULL)
709                                 data0 |= ((u64)1 << 32);
710                         spitfire_xcall_deliver(data0,
711                                                __pa(pg_addr),
712                                                (u64) pg_addr,
713                                                mask);
714                 } else {
715 #ifdef DCACHE_ALIASING_POSSIBLE
716                         data0 =
717                                 ((u64)&xcall_flush_dcache_page_cheetah);
718                         cheetah_xcall_deliver(data0,
719                                               __pa(pg_addr),
720                                               0, mask);
721 #endif
722                 }
723 #ifdef CONFIG_DEBUG_DCFLUSH
724                 atomic_inc(&dcpage_flushes_xcall);
725 #endif
726         }
727
728         put_cpu();
729 }
730
731 void flush_dcache_page_all(struct mm_struct *mm, struct page *page)
732 {
733         void *pg_addr = page_address(page);
734         cpumask_t mask = cpu_online_map;
735         u64 data0;
736         int this_cpu = get_cpu();
737
738         cpu_clear(this_cpu, mask);
739
740 #ifdef CONFIG_DEBUG_DCFLUSH
741         atomic_inc(&dcpage_flushes);
742 #endif
743         if (cpus_empty(mask))
744                 goto flush_self;
745         if (tlb_type == spitfire) {
746                 data0 = ((u64)&xcall_flush_dcache_page_spitfire);
747                 if (page_mapping(page) != NULL)
748                         data0 |= ((u64)1 << 32);
749                 spitfire_xcall_deliver(data0,
750                                        __pa(pg_addr),
751                                        (u64) pg_addr,
752                                        mask);
753         } else {
754 #ifdef DCACHE_ALIASING_POSSIBLE
755                 data0 = ((u64)&xcall_flush_dcache_page_cheetah);
756                 cheetah_xcall_deliver(data0,
757                                       __pa(pg_addr),
758                                       0, mask);
759 #endif
760         }
761 #ifdef CONFIG_DEBUG_DCFLUSH
762         atomic_inc(&dcpage_flushes_xcall);
763 #endif
764  flush_self:
765         __local_flush_dcache_page(page);
766
767         put_cpu();
768 }
769
770 void smp_receive_signal(int cpu)
771 {
772         cpumask_t mask = cpumask_of_cpu(cpu);
773
774         if (cpu_online(cpu)) {
775                 u64 data0 = (((u64)&xcall_receive_signal) & 0xffffffff);
776
777                 if (tlb_type == spitfire)
778                         spitfire_xcall_deliver(data0, 0, 0, mask);
779                 else
780                         cheetah_xcall_deliver(data0, 0, 0, mask);
781         }
782 }
783
784 void smp_receive_signal_client(int irq, struct pt_regs *regs)
785 {
786         /* Just return, rtrap takes care of the rest. */
787         clear_softint(1 << irq);
788 }
789
790 void smp_report_regs(void)
791 {
792         smp_cross_call(&xcall_report_regs, 0, 0, 0);
793 }
794
795 void smp_flush_tlb_all(void)
796 {
797         if (tlb_type == spitfire)
798                 smp_cross_call(&xcall_flush_tlb_all_spitfire, 0, 0, 0);
799         else
800                 smp_cross_call(&xcall_flush_tlb_all_cheetah, 0, 0, 0);
801         __flush_tlb_all();
802 }
803
804 /* We know that the window frames of the user have been flushed
805  * to the stack before we get here because all callers of us
806  * are flush_tlb_*() routines, and these run after flush_cache_*()
807  * which performs the flushw.
808  *
809  * The SMP TLB coherency scheme we use works as follows:
810  *
811  * 1) mm->cpu_vm_mask is a bit mask of which cpus an address
812  *    space has (potentially) executed on, this is the heuristic
813  *    we use to avoid doing cross calls.
814  *
815  *    Also, for flushing from kswapd and also for clones, we
816  *    use cpu_vm_mask as the list of cpus to make run the TLB.
817  *
818  * 2) TLB context numbers are shared globally across all processors
819  *    in the system, this allows us to play several games to avoid
820  *    cross calls.
821  *
822  *    One invariant is that when a cpu switches to a process, and
823  *    that processes tsk->active_mm->cpu_vm_mask does not have the
824  *    current cpu's bit set, that tlb context is flushed locally.
825  *
826  *    If the address space is non-shared (ie. mm->count == 1) we avoid
827  *    cross calls when we want to flush the currently running process's
828  *    tlb state.  This is done by clearing all cpu bits except the current
829  *    processor's in current->active_mm->cpu_vm_mask and performing the
830  *    flush locally only.  This will force any subsequent cpus which run
831  *    this task to flush the context from the local tlb if the process
832  *    migrates to another cpu (again).
833  *
834  * 3) For shared address spaces (threads) and swapping we bite the
835  *    bullet for most cases and perform the cross call (but only to
836  *    the cpus listed in cpu_vm_mask).
837  *
838  *    The performance gain from "optimizing" away the cross call for threads is
839  *    questionable (in theory the big win for threads is the massive sharing of
840  *    address space state across processors).
841  */
842
843 /* This currently is only used by the hugetlb arch pre-fault
844  * hook on UltraSPARC-III+ and later when changing the pagesize
845  * bits of the context register for an address space.
846  */
847 void smp_flush_tlb_mm(struct mm_struct *mm)
848 {
849         u32 ctx = CTX_HWBITS(mm->context);
850         int cpu = get_cpu();
851
852         if (atomic_read(&mm->mm_users) == 1) {
853                 mm->cpu_vm_mask = cpumask_of_cpu(cpu);
854                 goto local_flush_and_out;
855         }
856
857         smp_cross_call_masked(&xcall_flush_tlb_mm,
858                               ctx, 0, 0,
859                               mm->cpu_vm_mask);
860
861 local_flush_and_out:
862         __flush_tlb_mm(ctx, SECONDARY_CONTEXT);
863
864         put_cpu();
865 }
866
867 void smp_flush_tlb_pending(struct mm_struct *mm, unsigned long nr, unsigned long *vaddrs)
868 {
869         u32 ctx = CTX_HWBITS(mm->context);
870         int cpu = get_cpu();
871
872         if (mm == current->active_mm && atomic_read(&mm->mm_users) == 1)
873                 mm->cpu_vm_mask = cpumask_of_cpu(cpu);
874         else
875                 smp_cross_call_masked(&xcall_flush_tlb_pending,
876                                       ctx, nr, (unsigned long) vaddrs,
877                                       mm->cpu_vm_mask);
878
879         __flush_tlb_pending(ctx, nr, vaddrs);
880
881         put_cpu();
882 }
883
884 void smp_flush_tlb_kernel_range(unsigned long start, unsigned long end)
885 {
886         start &= PAGE_MASK;
887         end    = PAGE_ALIGN(end);
888         if (start != end) {
889                 smp_cross_call(&xcall_flush_tlb_kernel_range,
890                                0, start, end);
891
892                 __flush_tlb_kernel_range(start, end);
893         }
894 }
895
896 /* CPU capture. */
897 /* #define CAPTURE_DEBUG */
898 extern unsigned long xcall_capture;
899
900 static atomic_t smp_capture_depth = ATOMIC_INIT(0);
901 static atomic_t smp_capture_registry = ATOMIC_INIT(0);
902 static unsigned long penguins_are_doing_time;
903
904 void smp_capture(void)
905 {
906         int result = atomic_add_ret(1, &smp_capture_depth);
907
908         if (result == 1) {
909                 int ncpus = num_online_cpus();
910
911 #ifdef CAPTURE_DEBUG
912                 printk("CPU[%d]: Sending penguins to jail...",
913                        smp_processor_id());
914 #endif
915                 penguins_are_doing_time = 1;
916                 membar_storestore_loadstore();
917                 atomic_inc(&smp_capture_registry);
918                 smp_cross_call(&xcall_capture, 0, 0, 0);
919                 while (atomic_read(&smp_capture_registry) != ncpus)
920                         rmb();
921 #ifdef CAPTURE_DEBUG
922                 printk("done\n");
923 #endif
924         }
925 }
926
927 void smp_release(void)
928 {
929         if (atomic_dec_and_test(&smp_capture_depth)) {
930 #ifdef CAPTURE_DEBUG
931                 printk("CPU[%d]: Giving pardon to "
932                        "imprisoned penguins\n",
933                        smp_processor_id());
934 #endif
935                 penguins_are_doing_time = 0;
936                 membar_storeload_storestore();
937                 atomic_dec(&smp_capture_registry);
938         }
939 }
940
941 /* Imprisoned penguins run with %pil == 15, but PSTATE_IE set, so they
942  * can service tlb flush xcalls...
943  */
944 extern void prom_world(int);
945 extern void save_alternate_globals(unsigned long *);
946 extern void restore_alternate_globals(unsigned long *);
947 void smp_penguin_jailcell(int irq, struct pt_regs *regs)
948 {
949         unsigned long global_save[24];
950
951         clear_softint(1 << irq);
952
953         preempt_disable();
954
955         __asm__ __volatile__("flushw");
956         save_alternate_globals(global_save);
957         prom_world(1);
958         atomic_inc(&smp_capture_registry);
959         membar_storeload_storestore();
960         while (penguins_are_doing_time)
961                 rmb();
962         restore_alternate_globals(global_save);
963         atomic_dec(&smp_capture_registry);
964         prom_world(0);
965
966         preempt_enable();
967 }
968
969 #define prof_multiplier(__cpu)          cpu_data(__cpu).multiplier
970 #define prof_counter(__cpu)             cpu_data(__cpu).counter
971
972 void smp_percpu_timer_interrupt(struct pt_regs *regs)
973 {
974         unsigned long compare, tick, pstate;
975         int cpu = smp_processor_id();
976         int user = user_mode(regs);
977
978         /*
979          * Check for level 14 softint.
980          */
981         {
982                 unsigned long tick_mask = tick_ops->softint_mask;
983
984                 if (!(get_softint() & tick_mask)) {
985                         extern void handler_irq(int, struct pt_regs *);
986
987                         handler_irq(14, regs);
988                         return;
989                 }
990                 clear_softint(tick_mask);
991         }
992
993         do {
994                 profile_tick(CPU_PROFILING, regs);
995                 if (!--prof_counter(cpu)) {
996                         irq_enter();
997
998                         if (cpu == boot_cpu_id) {
999                                 kstat_this_cpu.irqs[0]++;
1000                                 timer_tick_interrupt(regs);
1001                         }
1002
1003                         update_process_times(user);
1004
1005                         irq_exit();
1006
1007                         prof_counter(cpu) = prof_multiplier(cpu);
1008                 }
1009
1010                 /* Guarantee that the following sequences execute
1011                  * uninterrupted.
1012                  */
1013                 __asm__ __volatile__("rdpr      %%pstate, %0\n\t"
1014                                      "wrpr      %0, %1, %%pstate"
1015                                      : "=r" (pstate)
1016                                      : "i" (PSTATE_IE));
1017
1018                 compare = tick_ops->add_compare(current_tick_offset);
1019                 tick = tick_ops->get_tick();
1020
1021                 /* Restore PSTATE_IE. */
1022                 __asm__ __volatile__("wrpr      %0, 0x0, %%pstate"
1023                                      : /* no outputs */
1024                                      : "r" (pstate));
1025         } while (time_after_eq(tick, compare));
1026 }
1027
1028 static void __init smp_setup_percpu_timer(void)
1029 {
1030         int cpu = smp_processor_id();
1031         unsigned long pstate;
1032
1033         prof_counter(cpu) = prof_multiplier(cpu) = 1;
1034
1035         /* Guarantee that the following sequences execute
1036          * uninterrupted.
1037          */
1038         __asm__ __volatile__("rdpr      %%pstate, %0\n\t"
1039                              "wrpr      %0, %1, %%pstate"
1040                              : "=r" (pstate)
1041                              : "i" (PSTATE_IE));
1042
1043         tick_ops->init_tick(current_tick_offset);
1044
1045         /* Restore PSTATE_IE. */
1046         __asm__ __volatile__("wrpr      %0, 0x0, %%pstate"
1047                              : /* no outputs */
1048                              : "r" (pstate));
1049 }
1050
1051 void __init smp_tick_init(void)
1052 {
1053         boot_cpu_id = hard_smp_processor_id();
1054         current_tick_offset = timer_tick_offset;
1055
1056         cpu_set(boot_cpu_id, cpu_online_map);
1057         prof_counter(boot_cpu_id) = prof_multiplier(boot_cpu_id) = 1;
1058 }
1059
1060 /* /proc/profile writes can call this, don't __init it please. */
1061 static DEFINE_SPINLOCK(prof_setup_lock);
1062
1063 int setup_profiling_timer(unsigned int multiplier)
1064 {
1065         unsigned long flags;
1066         int i;
1067
1068         if ((!multiplier) || (timer_tick_offset / multiplier) < 1000)
1069                 return -EINVAL;
1070
1071         spin_lock_irqsave(&prof_setup_lock, flags);
1072         for (i = 0; i < NR_CPUS; i++)
1073                 prof_multiplier(i) = multiplier;
1074         current_tick_offset = (timer_tick_offset / multiplier);
1075         spin_unlock_irqrestore(&prof_setup_lock, flags);
1076
1077         return 0;
1078 }
1079
1080 void __init smp_prepare_cpus(unsigned int max_cpus)
1081 {
1082         int instance, mid;
1083
1084         instance = 0;
1085         while (!cpu_find_by_instance(instance, NULL, &mid)) {
1086                 if (mid < max_cpus)
1087                         cpu_set(mid, phys_cpu_present_map);
1088                 instance++;
1089         }
1090
1091         if (num_possible_cpus() > max_cpus) {
1092                 instance = 0;
1093                 while (!cpu_find_by_instance(instance, NULL, &mid)) {
1094                         if (mid != boot_cpu_id) {
1095                                 cpu_clear(mid, phys_cpu_present_map);
1096                                 if (num_possible_cpus() <= max_cpus)
1097                                         break;
1098                         }
1099                         instance++;
1100                 }
1101         }
1102
1103         smp_store_cpu_info(boot_cpu_id);
1104 }
1105
1106 void __devinit smp_prepare_boot_cpu(void)
1107 {
1108         if (hard_smp_processor_id() >= NR_CPUS) {
1109                 prom_printf("Serious problem, boot cpu id >= NR_CPUS\n");
1110                 prom_halt();
1111         }
1112
1113         current_thread_info()->cpu = hard_smp_processor_id();
1114
1115         cpu_set(smp_processor_id(), cpu_online_map);
1116         cpu_set(smp_processor_id(), phys_cpu_present_map);
1117 }
1118
1119 int __devinit __cpu_up(unsigned int cpu)
1120 {
1121         int ret = smp_boot_one_cpu(cpu);
1122
1123         if (!ret) {
1124                 cpu_set(cpu, smp_commenced_mask);
1125                 while (!cpu_isset(cpu, cpu_online_map))
1126                         mb();
1127                 if (!cpu_isset(cpu, cpu_online_map)) {
1128                         ret = -ENODEV;
1129                 } else {
1130                         smp_synchronize_one_tick(cpu);
1131                 }
1132         }
1133         return ret;
1134 }
1135
1136 void __init smp_cpus_done(unsigned int max_cpus)
1137 {
1138         unsigned long bogosum = 0;
1139         int i;
1140
1141         for (i = 0; i < NR_CPUS; i++) {
1142                 if (cpu_online(i))
1143                         bogosum += cpu_data(i).udelay_val;
1144         }
1145         printk("Total of %ld processors activated "
1146                "(%lu.%02lu BogoMIPS).\n",
1147                (long) num_online_cpus(),
1148                bogosum/(500000/HZ),
1149                (bogosum/(5000/HZ))%100);
1150 }
1151
1152 /* This needn't do anything as we do not sleep the cpu
1153  * inside of the idler task, so an interrupt is not needed
1154  * to get a clean fast response.
1155  *
1156  * XXX Reverify this assumption... -DaveM
1157  *
1158  * Addendum: We do want it to do something for the signal
1159  *           delivery case, we detect that by just seeing
1160  *           if we are trying to send this to an idler or not.
1161  */
1162 void smp_send_reschedule(int cpu)
1163 {
1164         if (cpu_data(cpu).idle_volume == 0)
1165                 smp_receive_signal(cpu);
1166 }
1167
1168 /* This is a nop because we capture all other cpus
1169  * anyways when making the PROM active.
1170  */
1171 void smp_send_stop(void)
1172 {
1173 }
1174
1175 unsigned long __per_cpu_base __read_mostly;
1176 unsigned long __per_cpu_shift __read_mostly;
1177
1178 EXPORT_SYMBOL(__per_cpu_base);
1179 EXPORT_SYMBOL(__per_cpu_shift);
1180
1181 void __init setup_per_cpu_areas(void)
1182 {
1183         unsigned long goal, size, i;
1184         char *ptr;
1185         /* Created by linker magic */
1186         extern char __per_cpu_start[], __per_cpu_end[];
1187
1188         /* Copy section for each CPU (we discard the original) */
1189         goal = ALIGN(__per_cpu_end - __per_cpu_start, PAGE_SIZE);
1190
1191 #ifdef CONFIG_MODULES
1192         if (goal < PERCPU_ENOUGH_ROOM)
1193                 goal = PERCPU_ENOUGH_ROOM;
1194 #endif
1195         __per_cpu_shift = 0;
1196         for (size = 1UL; size < goal; size <<= 1UL)
1197                 __per_cpu_shift++;
1198
1199         /* Make sure the resulting __per_cpu_base value
1200          * will fit in the 43-bit sign extended IMMU
1201          * TSB register.
1202          */
1203         ptr = __alloc_bootmem(size * NR_CPUS, PAGE_SIZE,
1204                               (unsigned long) __per_cpu_start);
1205
1206         __per_cpu_base = ptr - __per_cpu_start;
1207
1208         if ((__per_cpu_shift < PAGE_SHIFT) ||
1209             (__per_cpu_base & ~PAGE_MASK) ||
1210             (__per_cpu_base != (((long) __per_cpu_base << 20) >> 20))) {
1211                 prom_printf("PER_CPU: Invalid layout, "
1212                             "ptr[%p] shift[%lx] base[%lx]\n",
1213                             ptr, __per_cpu_shift, __per_cpu_base);
1214                 prom_halt();
1215         }
1216
1217         for (i = 0; i < NR_CPUS; i++, ptr += size)
1218                 memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start);
1219
1220         /* Finally, load in the boot cpu's base value.
1221          * We abuse the IMMU TSB register for trap handler
1222          * entry and exit loading of %g5.  That is why it
1223          * has to be page aligned.
1224          */
1225         cpu_setup_percpu_base(hard_smp_processor_id());
1226 }