Merge git://git.kernel.org/pub/scm/linux/kernel/git/jejb/scsi-misc-2.6
[linux-2.6] / arch / sparc / kernel / smp_64.c
1 /* smp.c: Sparc64 SMP support.
2  *
3  * Copyright (C) 1997, 2007, 2008 David S. Miller (davem@davemloft.net)
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/interrupt.h>
14 #include <linux/kernel_stat.h>
15 #include <linux/delay.h>
16 #include <linux/init.h>
17 #include <linux/spinlock.h>
18 #include <linux/fs.h>
19 #include <linux/seq_file.h>
20 #include <linux/cache.h>
21 #include <linux/jiffies.h>
22 #include <linux/profile.h>
23 #include <linux/lmb.h>
24 #include <linux/cpu.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 #include <asm/hvtramp.h>
33 #include <asm/io.h>
34 #include <asm/timer.h>
35
36 #include <asm/irq.h>
37 #include <asm/irq_regs.h>
38 #include <asm/page.h>
39 #include <asm/pgtable.h>
40 #include <asm/oplib.h>
41 #include <asm/uaccess.h>
42 #include <asm/starfire.h>
43 #include <asm/tlb.h>
44 #include <asm/sections.h>
45 #include <asm/prom.h>
46 #include <asm/mdesc.h>
47 #include <asm/ldc.h>
48 #include <asm/hypervisor.h>
49
50 int sparc64_multi_core __read_mostly;
51
52 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_NONE;
53 cpumask_t cpu_online_map __read_mostly = CPU_MASK_NONE;
54 DEFINE_PER_CPU(cpumask_t, cpu_sibling_map) = CPU_MASK_NONE;
55 cpumask_t cpu_core_map[NR_CPUS] __read_mostly =
56         { [0 ... NR_CPUS-1] = CPU_MASK_NONE };
57
58 EXPORT_SYMBOL(cpu_possible_map);
59 EXPORT_SYMBOL(cpu_online_map);
60 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
61 EXPORT_SYMBOL(cpu_core_map);
62
63 static cpumask_t smp_commenced_mask;
64
65 void smp_info(struct seq_file *m)
66 {
67         int i;
68         
69         seq_printf(m, "State:\n");
70         for_each_online_cpu(i)
71                 seq_printf(m, "CPU%d:\t\tonline\n", i);
72 }
73
74 void smp_bogo(struct seq_file *m)
75 {
76         int i;
77         
78         for_each_online_cpu(i)
79                 seq_printf(m,
80                            "Cpu%dClkTck\t: %016lx\n",
81                            i, cpu_data(i).clock_tick);
82 }
83
84 extern void setup_sparc64_timer(void);
85
86 static volatile unsigned long callin_flag = 0;
87
88 void __cpuinit smp_callin(void)
89 {
90         int cpuid = hard_smp_processor_id();
91
92         __local_per_cpu_offset = __per_cpu_offset(cpuid);
93
94         if (tlb_type == hypervisor)
95                 sun4v_ktsb_register();
96
97         __flush_tlb_all();
98
99         setup_sparc64_timer();
100
101         if (cheetah_pcache_forced_on)
102                 cheetah_enable_pcache();
103
104         local_irq_enable();
105
106         callin_flag = 1;
107         __asm__ __volatile__("membar #Sync\n\t"
108                              "flush  %%g6" : : : "memory");
109
110         /* Clear this or we will die instantly when we
111          * schedule back to this idler...
112          */
113         current_thread_info()->new_child = 0;
114
115         /* Attach to the address space of init_task. */
116         atomic_inc(&init_mm.mm_count);
117         current->active_mm = &init_mm;
118
119         /* inform the notifiers about the new cpu */
120         notify_cpu_starting(cpuid);
121
122         while (!cpu_isset(cpuid, smp_commenced_mask))
123                 rmb();
124
125         ipi_call_lock();
126         cpu_set(cpuid, cpu_online_map);
127         ipi_call_unlock();
128
129         /* idle thread is expected to have preempt disabled */
130         preempt_disable();
131 }
132
133 void cpu_panic(void)
134 {
135         printk("CPU[%d]: Returns from cpu_idle!\n", smp_processor_id());
136         panic("SMP bolixed\n");
137 }
138
139 /* This tick register synchronization scheme is taken entirely from
140  * the ia64 port, see arch/ia64/kernel/smpboot.c for details and credit.
141  *
142  * The only change I've made is to rework it so that the master
143  * initiates the synchonization instead of the slave. -DaveM
144  */
145
146 #define MASTER  0
147 #define SLAVE   (SMP_CACHE_BYTES/sizeof(unsigned long))
148
149 #define NUM_ROUNDS      64      /* magic value */
150 #define NUM_ITERS       5       /* likewise */
151
152 static DEFINE_SPINLOCK(itc_sync_lock);
153 static unsigned long go[SLAVE + 1];
154
155 #define DEBUG_TICK_SYNC 0
156
157 static inline long get_delta (long *rt, long *master)
158 {
159         unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
160         unsigned long tcenter, t0, t1, tm;
161         unsigned long i;
162
163         for (i = 0; i < NUM_ITERS; i++) {
164                 t0 = tick_ops->get_tick();
165                 go[MASTER] = 1;
166                 membar_safe("#StoreLoad");
167                 while (!(tm = go[SLAVE]))
168                         rmb();
169                 go[SLAVE] = 0;
170                 wmb();
171                 t1 = tick_ops->get_tick();
172
173                 if (t1 - t0 < best_t1 - best_t0)
174                         best_t0 = t0, best_t1 = t1, best_tm = tm;
175         }
176
177         *rt = best_t1 - best_t0;
178         *master = best_tm - best_t0;
179
180         /* average best_t0 and best_t1 without overflow: */
181         tcenter = (best_t0/2 + best_t1/2);
182         if (best_t0 % 2 + best_t1 % 2 == 2)
183                 tcenter++;
184         return tcenter - best_tm;
185 }
186
187 void smp_synchronize_tick_client(void)
188 {
189         long i, delta, adj, adjust_latency = 0, done = 0;
190         unsigned long flags, rt, master_time_stamp, bound;
191 #if DEBUG_TICK_SYNC
192         struct {
193                 long rt;        /* roundtrip time */
194                 long master;    /* master's timestamp */
195                 long diff;      /* difference between midpoint and master's timestamp */
196                 long lat;       /* estimate of itc adjustment latency */
197         } t[NUM_ROUNDS];
198 #endif
199
200         go[MASTER] = 1;
201
202         while (go[MASTER])
203                 rmb();
204
205         local_irq_save(flags);
206         {
207                 for (i = 0; i < NUM_ROUNDS; i++) {
208                         delta = get_delta(&rt, &master_time_stamp);
209                         if (delta == 0) {
210                                 done = 1;       /* let's lock on to this... */
211                                 bound = rt;
212                         }
213
214                         if (!done) {
215                                 if (i > 0) {
216                                         adjust_latency += -delta;
217                                         adj = -delta + adjust_latency/4;
218                                 } else
219                                         adj = -delta;
220
221                                 tick_ops->add_tick(adj);
222                         }
223 #if DEBUG_TICK_SYNC
224                         t[i].rt = rt;
225                         t[i].master = master_time_stamp;
226                         t[i].diff = delta;
227                         t[i].lat = adjust_latency/4;
228 #endif
229                 }
230         }
231         local_irq_restore(flags);
232
233 #if DEBUG_TICK_SYNC
234         for (i = 0; i < NUM_ROUNDS; i++)
235                 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
236                        t[i].rt, t[i].master, t[i].diff, t[i].lat);
237 #endif
238
239         printk(KERN_INFO "CPU %d: synchronized TICK with master CPU "
240                "(last diff %ld cycles, maxerr %lu cycles)\n",
241                smp_processor_id(), delta, rt);
242 }
243
244 static void smp_start_sync_tick_client(int cpu);
245
246 static void smp_synchronize_one_tick(int cpu)
247 {
248         unsigned long flags, i;
249
250         go[MASTER] = 0;
251
252         smp_start_sync_tick_client(cpu);
253
254         /* wait for client to be ready */
255         while (!go[MASTER])
256                 rmb();
257
258         /* now let the client proceed into his loop */
259         go[MASTER] = 0;
260         membar_safe("#StoreLoad");
261
262         spin_lock_irqsave(&itc_sync_lock, flags);
263         {
264                 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; i++) {
265                         while (!go[MASTER])
266                                 rmb();
267                         go[MASTER] = 0;
268                         wmb();
269                         go[SLAVE] = tick_ops->get_tick();
270                         membar_safe("#StoreLoad");
271                 }
272         }
273         spin_unlock_irqrestore(&itc_sync_lock, flags);
274 }
275
276 #if defined(CONFIG_SUN_LDOMS) && defined(CONFIG_HOTPLUG_CPU)
277 /* XXX Put this in some common place. XXX */
278 static unsigned long kimage_addr_to_ra(void *p)
279 {
280         unsigned long val = (unsigned long) p;
281
282         return kern_base + (val - KERNBASE);
283 }
284
285 static void __cpuinit ldom_startcpu_cpuid(unsigned int cpu, unsigned long thread_reg)
286 {
287         extern unsigned long sparc64_ttable_tl0;
288         extern unsigned long kern_locked_tte_data;
289         struct hvtramp_descr *hdesc;
290         unsigned long trampoline_ra;
291         struct trap_per_cpu *tb;
292         u64 tte_vaddr, tte_data;
293         unsigned long hv_err;
294         int i;
295
296         hdesc = kzalloc(sizeof(*hdesc) +
297                         (sizeof(struct hvtramp_mapping) *
298                          num_kernel_image_mappings - 1),
299                         GFP_KERNEL);
300         if (!hdesc) {
301                 printk(KERN_ERR "ldom_startcpu_cpuid: Cannot allocate "
302                        "hvtramp_descr.\n");
303                 return;
304         }
305
306         hdesc->cpu = cpu;
307         hdesc->num_mappings = num_kernel_image_mappings;
308
309         tb = &trap_block[cpu];
310         tb->hdesc = hdesc;
311
312         hdesc->fault_info_va = (unsigned long) &tb->fault_info;
313         hdesc->fault_info_pa = kimage_addr_to_ra(&tb->fault_info);
314
315         hdesc->thread_reg = thread_reg;
316
317         tte_vaddr = (unsigned long) KERNBASE;
318         tte_data = kern_locked_tte_data;
319
320         for (i = 0; i < hdesc->num_mappings; i++) {
321                 hdesc->maps[i].vaddr = tte_vaddr;
322                 hdesc->maps[i].tte   = tte_data;
323                 tte_vaddr += 0x400000;
324                 tte_data  += 0x400000;
325         }
326
327         trampoline_ra = kimage_addr_to_ra(hv_cpu_startup);
328
329         hv_err = sun4v_cpu_start(cpu, trampoline_ra,
330                                  kimage_addr_to_ra(&sparc64_ttable_tl0),
331                                  __pa(hdesc));
332         if (hv_err)
333                 printk(KERN_ERR "ldom_startcpu_cpuid: sun4v_cpu_start() "
334                        "gives error %lu\n", hv_err);
335 }
336 #endif
337
338 extern unsigned long sparc64_cpu_startup;
339
340 /* The OBP cpu startup callback truncates the 3rd arg cookie to
341  * 32-bits (I think) so to be safe we have it read the pointer
342  * contained here so we work on >4GB machines. -DaveM
343  */
344 static struct thread_info *cpu_new_thread = NULL;
345
346 static int __cpuinit smp_boot_one_cpu(unsigned int cpu)
347 {
348         struct trap_per_cpu *tb = &trap_block[cpu];
349         unsigned long entry =
350                 (unsigned long)(&sparc64_cpu_startup);
351         unsigned long cookie =
352                 (unsigned long)(&cpu_new_thread);
353         struct task_struct *p;
354         int timeout, ret;
355
356         p = fork_idle(cpu);
357         if (IS_ERR(p))
358                 return PTR_ERR(p);
359         callin_flag = 0;
360         cpu_new_thread = task_thread_info(p);
361
362         if (tlb_type == hypervisor) {
363 #if defined(CONFIG_SUN_LDOMS) && defined(CONFIG_HOTPLUG_CPU)
364                 if (ldom_domaining_enabled)
365                         ldom_startcpu_cpuid(cpu,
366                                             (unsigned long) cpu_new_thread);
367                 else
368 #endif
369                         prom_startcpu_cpuid(cpu, entry, cookie);
370         } else {
371                 struct device_node *dp = of_find_node_by_cpuid(cpu);
372
373                 prom_startcpu(dp->node, entry, cookie);
374         }
375
376         for (timeout = 0; timeout < 50000; timeout++) {
377                 if (callin_flag)
378                         break;
379                 udelay(100);
380         }
381
382         if (callin_flag) {
383                 ret = 0;
384         } else {
385                 printk("Processor %d is stuck.\n", cpu);
386                 ret = -ENODEV;
387         }
388         cpu_new_thread = NULL;
389
390         if (tb->hdesc) {
391                 kfree(tb->hdesc);
392                 tb->hdesc = NULL;
393         }
394
395         return ret;
396 }
397
398 static void spitfire_xcall_helper(u64 data0, u64 data1, u64 data2, u64 pstate, unsigned long cpu)
399 {
400         u64 result, target;
401         int stuck, tmp;
402
403         if (this_is_starfire) {
404                 /* map to real upaid */
405                 cpu = (((cpu & 0x3c) << 1) |
406                         ((cpu & 0x40) >> 4) |
407                         (cpu & 0x3));
408         }
409
410         target = (cpu << 14) | 0x70;
411 again:
412         /* Ok, this is the real Spitfire Errata #54.
413          * One must read back from a UDB internal register
414          * after writes to the UDB interrupt dispatch, but
415          * before the membar Sync for that write.
416          * So we use the high UDB control register (ASI 0x7f,
417          * ADDR 0x20) for the dummy read. -DaveM
418          */
419         tmp = 0x40;
420         __asm__ __volatile__(
421         "wrpr   %1, %2, %%pstate\n\t"
422         "stxa   %4, [%0] %3\n\t"
423         "stxa   %5, [%0+%8] %3\n\t"
424         "add    %0, %8, %0\n\t"
425         "stxa   %6, [%0+%8] %3\n\t"
426         "membar #Sync\n\t"
427         "stxa   %%g0, [%7] %3\n\t"
428         "membar #Sync\n\t"
429         "mov    0x20, %%g1\n\t"
430         "ldxa   [%%g1] 0x7f, %%g0\n\t"
431         "membar #Sync"
432         : "=r" (tmp)
433         : "r" (pstate), "i" (PSTATE_IE), "i" (ASI_INTR_W),
434           "r" (data0), "r" (data1), "r" (data2), "r" (target),
435           "r" (0x10), "0" (tmp)
436         : "g1");
437
438         /* NOTE: PSTATE_IE is still clear. */
439         stuck = 100000;
440         do {
441                 __asm__ __volatile__("ldxa [%%g0] %1, %0"
442                         : "=r" (result)
443                         : "i" (ASI_INTR_DISPATCH_STAT));
444                 if (result == 0) {
445                         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
446                                              : : "r" (pstate));
447                         return;
448                 }
449                 stuck -= 1;
450                 if (stuck == 0)
451                         break;
452         } while (result & 0x1);
453         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
454                              : : "r" (pstate));
455         if (stuck == 0) {
456                 printk("CPU[%d]: mondo stuckage result[%016lx]\n",
457                        smp_processor_id(), result);
458         } else {
459                 udelay(2);
460                 goto again;
461         }
462 }
463
464 static void spitfire_xcall_deliver(struct trap_per_cpu *tb, int cnt)
465 {
466         u64 *mondo, data0, data1, data2;
467         u16 *cpu_list;
468         u64 pstate;
469         int i;
470
471         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
472         cpu_list = __va(tb->cpu_list_pa);
473         mondo = __va(tb->cpu_mondo_block_pa);
474         data0 = mondo[0];
475         data1 = mondo[1];
476         data2 = mondo[2];
477         for (i = 0; i < cnt; i++)
478                 spitfire_xcall_helper(data0, data1, data2, pstate, cpu_list[i]);
479 }
480
481 /* Cheetah now allows to send the whole 64-bytes of data in the interrupt
482  * packet, but we have no use for that.  However we do take advantage of
483  * the new pipelining feature (ie. dispatch to multiple cpus simultaneously).
484  */
485 static void cheetah_xcall_deliver(struct trap_per_cpu *tb, int cnt)
486 {
487         int nack_busy_id, is_jbus, need_more;
488         u64 *mondo, pstate, ver, busy_mask;
489         u16 *cpu_list;
490
491         cpu_list = __va(tb->cpu_list_pa);
492         mondo = __va(tb->cpu_mondo_block_pa);
493
494         /* Unfortunately, someone at Sun had the brilliant idea to make the
495          * busy/nack fields hard-coded by ITID number for this Ultra-III
496          * derivative processor.
497          */
498         __asm__ ("rdpr %%ver, %0" : "=r" (ver));
499         is_jbus = ((ver >> 32) == __JALAPENO_ID ||
500                    (ver >> 32) == __SERRANO_ID);
501
502         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
503
504 retry:
505         need_more = 0;
506         __asm__ __volatile__("wrpr %0, %1, %%pstate\n\t"
507                              : : "r" (pstate), "i" (PSTATE_IE));
508
509         /* Setup the dispatch data registers. */
510         __asm__ __volatile__("stxa      %0, [%3] %6\n\t"
511                              "stxa      %1, [%4] %6\n\t"
512                              "stxa      %2, [%5] %6\n\t"
513                              "membar    #Sync\n\t"
514                              : /* no outputs */
515                              : "r" (mondo[0]), "r" (mondo[1]), "r" (mondo[2]),
516                                "r" (0x40), "r" (0x50), "r" (0x60),
517                                "i" (ASI_INTR_W));
518
519         nack_busy_id = 0;
520         busy_mask = 0;
521         {
522                 int i;
523
524                 for (i = 0; i < cnt; i++) {
525                         u64 target, nr;
526
527                         nr = cpu_list[i];
528                         if (nr == 0xffff)
529                                 continue;
530
531                         target = (nr << 14) | 0x70;
532                         if (is_jbus) {
533                                 busy_mask |= (0x1UL << (nr * 2));
534                         } else {
535                                 target |= (nack_busy_id << 24);
536                                 busy_mask |= (0x1UL <<
537                                               (nack_busy_id * 2));
538                         }
539                         __asm__ __volatile__(
540                                 "stxa   %%g0, [%0] %1\n\t"
541                                 "membar #Sync\n\t"
542                                 : /* no outputs */
543                                 : "r" (target), "i" (ASI_INTR_W));
544                         nack_busy_id++;
545                         if (nack_busy_id == 32) {
546                                 need_more = 1;
547                                 break;
548                         }
549                 }
550         }
551
552         /* Now, poll for completion. */
553         {
554                 u64 dispatch_stat, nack_mask;
555                 long stuck;
556
557                 stuck = 100000 * nack_busy_id;
558                 nack_mask = busy_mask << 1;
559                 do {
560                         __asm__ __volatile__("ldxa      [%%g0] %1, %0"
561                                              : "=r" (dispatch_stat)
562                                              : "i" (ASI_INTR_DISPATCH_STAT));
563                         if (!(dispatch_stat & (busy_mask | nack_mask))) {
564                                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
565                                                      : : "r" (pstate));
566                                 if (unlikely(need_more)) {
567                                         int i, this_cnt = 0;
568                                         for (i = 0; i < cnt; i++) {
569                                                 if (cpu_list[i] == 0xffff)
570                                                         continue;
571                                                 cpu_list[i] = 0xffff;
572                                                 this_cnt++;
573                                                 if (this_cnt == 32)
574                                                         break;
575                                         }
576                                         goto retry;
577                                 }
578                                 return;
579                         }
580                         if (!--stuck)
581                                 break;
582                 } while (dispatch_stat & busy_mask);
583
584                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
585                                      : : "r" (pstate));
586
587                 if (dispatch_stat & busy_mask) {
588                         /* Busy bits will not clear, continue instead
589                          * of freezing up on this cpu.
590                          */
591                         printk("CPU[%d]: mondo stuckage result[%016lx]\n",
592                                smp_processor_id(), dispatch_stat);
593                 } else {
594                         int i, this_busy_nack = 0;
595
596                         /* Delay some random time with interrupts enabled
597                          * to prevent deadlock.
598                          */
599                         udelay(2 * nack_busy_id);
600
601                         /* Clear out the mask bits for cpus which did not
602                          * NACK us.
603                          */
604                         for (i = 0; i < cnt; i++) {
605                                 u64 check_mask, nr;
606
607                                 nr = cpu_list[i];
608                                 if (nr == 0xffff)
609                                         continue;
610
611                                 if (is_jbus)
612                                         check_mask = (0x2UL << (2*nr));
613                                 else
614                                         check_mask = (0x2UL <<
615                                                       this_busy_nack);
616                                 if ((dispatch_stat & check_mask) == 0)
617                                         cpu_list[i] = 0xffff;
618                                 this_busy_nack += 2;
619                                 if (this_busy_nack == 64)
620                                         break;
621                         }
622
623                         goto retry;
624                 }
625         }
626 }
627
628 /* Multi-cpu list version.  */
629 static void hypervisor_xcall_deliver(struct trap_per_cpu *tb, int cnt)
630 {
631         int retries, this_cpu, prev_sent, i, saw_cpu_error;
632         unsigned long status;
633         u16 *cpu_list;
634
635         this_cpu = smp_processor_id();
636
637         cpu_list = __va(tb->cpu_list_pa);
638
639         saw_cpu_error = 0;
640         retries = 0;
641         prev_sent = 0;
642         do {
643                 int forward_progress, n_sent;
644
645                 status = sun4v_cpu_mondo_send(cnt,
646                                               tb->cpu_list_pa,
647                                               tb->cpu_mondo_block_pa);
648
649                 /* HV_EOK means all cpus received the xcall, we're done.  */
650                 if (likely(status == HV_EOK))
651                         break;
652
653                 /* First, see if we made any forward progress.
654                  *
655                  * The hypervisor indicates successful sends by setting
656                  * cpu list entries to the value 0xffff.
657                  */
658                 n_sent = 0;
659                 for (i = 0; i < cnt; i++) {
660                         if (likely(cpu_list[i] == 0xffff))
661                                 n_sent++;
662                 }
663
664                 forward_progress = 0;
665                 if (n_sent > prev_sent)
666                         forward_progress = 1;
667
668                 prev_sent = n_sent;
669
670                 /* If we get a HV_ECPUERROR, then one or more of the cpus
671                  * in the list are in error state.  Use the cpu_state()
672                  * hypervisor call to find out which cpus are in error state.
673                  */
674                 if (unlikely(status == HV_ECPUERROR)) {
675                         for (i = 0; i < cnt; i++) {
676                                 long err;
677                                 u16 cpu;
678
679                                 cpu = cpu_list[i];
680                                 if (cpu == 0xffff)
681                                         continue;
682
683                                 err = sun4v_cpu_state(cpu);
684                                 if (err == HV_CPU_STATE_ERROR) {
685                                         saw_cpu_error = (cpu + 1);
686                                         cpu_list[i] = 0xffff;
687                                 }
688                         }
689                 } else if (unlikely(status != HV_EWOULDBLOCK))
690                         goto fatal_mondo_error;
691
692                 /* Don't bother rewriting the CPU list, just leave the
693                  * 0xffff and non-0xffff entries in there and the
694                  * hypervisor will do the right thing.
695                  *
696                  * Only advance timeout state if we didn't make any
697                  * forward progress.
698                  */
699                 if (unlikely(!forward_progress)) {
700                         if (unlikely(++retries > 10000))
701                                 goto fatal_mondo_timeout;
702
703                         /* Delay a little bit to let other cpus catch up
704                          * on their cpu mondo queue work.
705                          */
706                         udelay(2 * cnt);
707                 }
708         } while (1);
709
710         if (unlikely(saw_cpu_error))
711                 goto fatal_mondo_cpu_error;
712
713         return;
714
715 fatal_mondo_cpu_error:
716         printk(KERN_CRIT "CPU[%d]: SUN4V mondo cpu error, some target cpus "
717                "(including %d) were in error state\n",
718                this_cpu, saw_cpu_error - 1);
719         return;
720
721 fatal_mondo_timeout:
722         printk(KERN_CRIT "CPU[%d]: SUN4V mondo timeout, no forward "
723                " progress after %d retries.\n",
724                this_cpu, retries);
725         goto dump_cpu_list_and_out;
726
727 fatal_mondo_error:
728         printk(KERN_CRIT "CPU[%d]: Unexpected SUN4V mondo error %lu\n",
729                this_cpu, status);
730         printk(KERN_CRIT "CPU[%d]: Args were cnt(%d) cpulist_pa(%lx) "
731                "mondo_block_pa(%lx)\n",
732                this_cpu, cnt, tb->cpu_list_pa, tb->cpu_mondo_block_pa);
733
734 dump_cpu_list_and_out:
735         printk(KERN_CRIT "CPU[%d]: CPU list [ ", this_cpu);
736         for (i = 0; i < cnt; i++)
737                 printk("%u ", cpu_list[i]);
738         printk("]\n");
739 }
740
741 static void (*xcall_deliver_impl)(struct trap_per_cpu *, int);
742
743 static void xcall_deliver(u64 data0, u64 data1, u64 data2, const cpumask_t *mask)
744 {
745         struct trap_per_cpu *tb;
746         int this_cpu, i, cnt;
747         unsigned long flags;
748         u16 *cpu_list;
749         u64 *mondo;
750
751         /* We have to do this whole thing with interrupts fully disabled.
752          * Otherwise if we send an xcall from interrupt context it will
753          * corrupt both our mondo block and cpu list state.
754          *
755          * One consequence of this is that we cannot use timeout mechanisms
756          * that depend upon interrupts being delivered locally.  So, for
757          * example, we cannot sample jiffies and expect it to advance.
758          *
759          * Fortunately, udelay() uses %stick/%tick so we can use that.
760          */
761         local_irq_save(flags);
762
763         this_cpu = smp_processor_id();
764         tb = &trap_block[this_cpu];
765
766         mondo = __va(tb->cpu_mondo_block_pa);
767         mondo[0] = data0;
768         mondo[1] = data1;
769         mondo[2] = data2;
770         wmb();
771
772         cpu_list = __va(tb->cpu_list_pa);
773
774         /* Setup the initial cpu list.  */
775         cnt = 0;
776         for_each_cpu(i, mask) {
777                 if (i == this_cpu || !cpu_online(i))
778                         continue;
779                 cpu_list[cnt++] = i;
780         }
781
782         if (cnt)
783                 xcall_deliver_impl(tb, cnt);
784
785         local_irq_restore(flags);
786 }
787
788 /* Send cross call to all processors mentioned in MASK_P
789  * except self.  Really, there are only two cases currently,
790  * "&cpu_online_map" and "&mm->cpu_vm_mask".
791  */
792 static void smp_cross_call_masked(unsigned long *func, u32 ctx, u64 data1, u64 data2, const cpumask_t *mask)
793 {
794         u64 data0 = (((u64)ctx)<<32 | (((u64)func) & 0xffffffff));
795
796         xcall_deliver(data0, data1, data2, mask);
797 }
798
799 /* Send cross call to all processors except self. */
800 static void smp_cross_call(unsigned long *func, u32 ctx, u64 data1, u64 data2)
801 {
802         smp_cross_call_masked(func, ctx, data1, data2, &cpu_online_map);
803 }
804
805 extern unsigned long xcall_sync_tick;
806
807 static void smp_start_sync_tick_client(int cpu)
808 {
809         xcall_deliver((u64) &xcall_sync_tick, 0, 0,
810                       &cpumask_of_cpu(cpu));
811 }
812
813 extern unsigned long xcall_call_function;
814
815 void arch_send_call_function_ipi(cpumask_t mask)
816 {
817         xcall_deliver((u64) &xcall_call_function, 0, 0, &mask);
818 }
819
820 extern unsigned long xcall_call_function_single;
821
822 void arch_send_call_function_single_ipi(int cpu)
823 {
824         xcall_deliver((u64) &xcall_call_function_single, 0, 0,
825                       &cpumask_of_cpu(cpu));
826 }
827
828 void smp_call_function_client(int irq, struct pt_regs *regs)
829 {
830         clear_softint(1 << irq);
831         generic_smp_call_function_interrupt();
832 }
833
834 void smp_call_function_single_client(int irq, struct pt_regs *regs)
835 {
836         clear_softint(1 << irq);
837         generic_smp_call_function_single_interrupt();
838 }
839
840 static void tsb_sync(void *info)
841 {
842         struct trap_per_cpu *tp = &trap_block[raw_smp_processor_id()];
843         struct mm_struct *mm = info;
844
845         /* It is not valid to test "currrent->active_mm == mm" here.
846          *
847          * The value of "current" is not changed atomically with
848          * switch_mm().  But that's OK, we just need to check the
849          * current cpu's trap block PGD physical address.
850          */
851         if (tp->pgd_paddr == __pa(mm->pgd))
852                 tsb_context_switch(mm);
853 }
854
855 void smp_tsb_sync(struct mm_struct *mm)
856 {
857         smp_call_function_mask(mm->cpu_vm_mask, tsb_sync, mm, 1);
858 }
859
860 extern unsigned long xcall_flush_tlb_mm;
861 extern unsigned long xcall_flush_tlb_pending;
862 extern unsigned long xcall_flush_tlb_kernel_range;
863 extern unsigned long xcall_fetch_glob_regs;
864 extern unsigned long xcall_receive_signal;
865 extern unsigned long xcall_new_mmu_context_version;
866 #ifdef CONFIG_KGDB
867 extern unsigned long xcall_kgdb_capture;
868 #endif
869
870 #ifdef DCACHE_ALIASING_POSSIBLE
871 extern unsigned long xcall_flush_dcache_page_cheetah;
872 #endif
873 extern unsigned long xcall_flush_dcache_page_spitfire;
874
875 #ifdef CONFIG_DEBUG_DCFLUSH
876 extern atomic_t dcpage_flushes;
877 extern atomic_t dcpage_flushes_xcall;
878 #endif
879
880 static inline void __local_flush_dcache_page(struct page *page)
881 {
882 #ifdef DCACHE_ALIASING_POSSIBLE
883         __flush_dcache_page(page_address(page),
884                             ((tlb_type == spitfire) &&
885                              page_mapping(page) != NULL));
886 #else
887         if (page_mapping(page) != NULL &&
888             tlb_type == spitfire)
889                 __flush_icache_page(__pa(page_address(page)));
890 #endif
891 }
892
893 void smp_flush_dcache_page_impl(struct page *page, int cpu)
894 {
895         int this_cpu;
896
897         if (tlb_type == hypervisor)
898                 return;
899
900 #ifdef CONFIG_DEBUG_DCFLUSH
901         atomic_inc(&dcpage_flushes);
902 #endif
903
904         this_cpu = get_cpu();
905
906         if (cpu == this_cpu) {
907                 __local_flush_dcache_page(page);
908         } else if (cpu_online(cpu)) {
909                 void *pg_addr = page_address(page);
910                 u64 data0 = 0;
911
912                 if (tlb_type == spitfire) {
913                         data0 = ((u64)&xcall_flush_dcache_page_spitfire);
914                         if (page_mapping(page) != NULL)
915                                 data0 |= ((u64)1 << 32);
916                 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
917 #ifdef DCACHE_ALIASING_POSSIBLE
918                         data0 = ((u64)&xcall_flush_dcache_page_cheetah);
919 #endif
920                 }
921                 if (data0) {
922                         xcall_deliver(data0, __pa(pg_addr),
923                                       (u64) pg_addr, &cpumask_of_cpu(cpu));
924 #ifdef CONFIG_DEBUG_DCFLUSH
925                         atomic_inc(&dcpage_flushes_xcall);
926 #endif
927                 }
928         }
929
930         put_cpu();
931 }
932
933 void flush_dcache_page_all(struct mm_struct *mm, struct page *page)
934 {
935         void *pg_addr;
936         int this_cpu;
937         u64 data0;
938
939         if (tlb_type == hypervisor)
940                 return;
941
942         this_cpu = get_cpu();
943
944 #ifdef CONFIG_DEBUG_DCFLUSH
945         atomic_inc(&dcpage_flushes);
946 #endif
947         data0 = 0;
948         pg_addr = page_address(page);
949         if (tlb_type == spitfire) {
950                 data0 = ((u64)&xcall_flush_dcache_page_spitfire);
951                 if (page_mapping(page) != NULL)
952                         data0 |= ((u64)1 << 32);
953         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
954 #ifdef DCACHE_ALIASING_POSSIBLE
955                 data0 = ((u64)&xcall_flush_dcache_page_cheetah);
956 #endif
957         }
958         if (data0) {
959                 xcall_deliver(data0, __pa(pg_addr),
960                               (u64) pg_addr, &cpu_online_map);
961 #ifdef CONFIG_DEBUG_DCFLUSH
962                 atomic_inc(&dcpage_flushes_xcall);
963 #endif
964         }
965         __local_flush_dcache_page(page);
966
967         put_cpu();
968 }
969
970 void smp_new_mmu_context_version_client(int irq, struct pt_regs *regs)
971 {
972         struct mm_struct *mm;
973         unsigned long flags;
974
975         clear_softint(1 << irq);
976
977         /* See if we need to allocate a new TLB context because
978          * the version of the one we are using is now out of date.
979          */
980         mm = current->active_mm;
981         if (unlikely(!mm || (mm == &init_mm)))
982                 return;
983
984         spin_lock_irqsave(&mm->context.lock, flags);
985
986         if (unlikely(!CTX_VALID(mm->context)))
987                 get_new_mmu_context(mm);
988
989         spin_unlock_irqrestore(&mm->context.lock, flags);
990
991         load_secondary_context(mm);
992         __flush_tlb_mm(CTX_HWBITS(mm->context),
993                        SECONDARY_CONTEXT);
994 }
995
996 void smp_new_mmu_context_version(void)
997 {
998         smp_cross_call(&xcall_new_mmu_context_version, 0, 0, 0);
999 }
1000
1001 #ifdef CONFIG_KGDB
1002 void kgdb_roundup_cpus(unsigned long flags)
1003 {
1004         smp_cross_call(&xcall_kgdb_capture, 0, 0, 0);
1005 }
1006 #endif
1007
1008 void smp_fetch_global_regs(void)
1009 {
1010         smp_cross_call(&xcall_fetch_glob_regs, 0, 0, 0);
1011 }
1012
1013 /* We know that the window frames of the user have been flushed
1014  * to the stack before we get here because all callers of us
1015  * are flush_tlb_*() routines, and these run after flush_cache_*()
1016  * which performs the flushw.
1017  *
1018  * The SMP TLB coherency scheme we use works as follows:
1019  *
1020  * 1) mm->cpu_vm_mask is a bit mask of which cpus an address
1021  *    space has (potentially) executed on, this is the heuristic
1022  *    we use to avoid doing cross calls.
1023  *
1024  *    Also, for flushing from kswapd and also for clones, we
1025  *    use cpu_vm_mask as the list of cpus to make run the TLB.
1026  *
1027  * 2) TLB context numbers are shared globally across all processors
1028  *    in the system, this allows us to play several games to avoid
1029  *    cross calls.
1030  *
1031  *    One invariant is that when a cpu switches to a process, and
1032  *    that processes tsk->active_mm->cpu_vm_mask does not have the
1033  *    current cpu's bit set, that tlb context is flushed locally.
1034  *
1035  *    If the address space is non-shared (ie. mm->count == 1) we avoid
1036  *    cross calls when we want to flush the currently running process's
1037  *    tlb state.  This is done by clearing all cpu bits except the current
1038  *    processor's in current->active_mm->cpu_vm_mask and performing the
1039  *    flush locally only.  This will force any subsequent cpus which run
1040  *    this task to flush the context from the local tlb if the process
1041  *    migrates to another cpu (again).
1042  *
1043  * 3) For shared address spaces (threads) and swapping we bite the
1044  *    bullet for most cases and perform the cross call (but only to
1045  *    the cpus listed in cpu_vm_mask).
1046  *
1047  *    The performance gain from "optimizing" away the cross call for threads is
1048  *    questionable (in theory the big win for threads is the massive sharing of
1049  *    address space state across processors).
1050  */
1051
1052 /* This currently is only used by the hugetlb arch pre-fault
1053  * hook on UltraSPARC-III+ and later when changing the pagesize
1054  * bits of the context register for an address space.
1055  */
1056 void smp_flush_tlb_mm(struct mm_struct *mm)
1057 {
1058         u32 ctx = CTX_HWBITS(mm->context);
1059         int cpu = get_cpu();
1060
1061         if (atomic_read(&mm->mm_users) == 1) {
1062                 mm->cpu_vm_mask = cpumask_of_cpu(cpu);
1063                 goto local_flush_and_out;
1064         }
1065
1066         smp_cross_call_masked(&xcall_flush_tlb_mm,
1067                               ctx, 0, 0,
1068                               &mm->cpu_vm_mask);
1069
1070 local_flush_and_out:
1071         __flush_tlb_mm(ctx, SECONDARY_CONTEXT);
1072
1073         put_cpu();
1074 }
1075
1076 void smp_flush_tlb_pending(struct mm_struct *mm, unsigned long nr, unsigned long *vaddrs)
1077 {
1078         u32 ctx = CTX_HWBITS(mm->context);
1079         int cpu = get_cpu();
1080
1081         if (mm == current->active_mm && atomic_read(&mm->mm_users) == 1)
1082                 mm->cpu_vm_mask = cpumask_of_cpu(cpu);
1083         else
1084                 smp_cross_call_masked(&xcall_flush_tlb_pending,
1085                                       ctx, nr, (unsigned long) vaddrs,
1086                                       &mm->cpu_vm_mask);
1087
1088         __flush_tlb_pending(ctx, nr, vaddrs);
1089
1090         put_cpu();
1091 }
1092
1093 void smp_flush_tlb_kernel_range(unsigned long start, unsigned long end)
1094 {
1095         start &= PAGE_MASK;
1096         end    = PAGE_ALIGN(end);
1097         if (start != end) {
1098                 smp_cross_call(&xcall_flush_tlb_kernel_range,
1099                                0, start, end);
1100
1101                 __flush_tlb_kernel_range(start, end);
1102         }
1103 }
1104
1105 /* CPU capture. */
1106 /* #define CAPTURE_DEBUG */
1107 extern unsigned long xcall_capture;
1108
1109 static atomic_t smp_capture_depth = ATOMIC_INIT(0);
1110 static atomic_t smp_capture_registry = ATOMIC_INIT(0);
1111 static unsigned long penguins_are_doing_time;
1112
1113 void smp_capture(void)
1114 {
1115         int result = atomic_add_ret(1, &smp_capture_depth);
1116
1117         if (result == 1) {
1118                 int ncpus = num_online_cpus();
1119
1120 #ifdef CAPTURE_DEBUG
1121                 printk("CPU[%d]: Sending penguins to jail...",
1122                        smp_processor_id());
1123 #endif
1124                 penguins_are_doing_time = 1;
1125                 atomic_inc(&smp_capture_registry);
1126                 smp_cross_call(&xcall_capture, 0, 0, 0);
1127                 while (atomic_read(&smp_capture_registry) != ncpus)
1128                         rmb();
1129 #ifdef CAPTURE_DEBUG
1130                 printk("done\n");
1131 #endif
1132         }
1133 }
1134
1135 void smp_release(void)
1136 {
1137         if (atomic_dec_and_test(&smp_capture_depth)) {
1138 #ifdef CAPTURE_DEBUG
1139                 printk("CPU[%d]: Giving pardon to "
1140                        "imprisoned penguins\n",
1141                        smp_processor_id());
1142 #endif
1143                 penguins_are_doing_time = 0;
1144                 membar_safe("#StoreLoad");
1145                 atomic_dec(&smp_capture_registry);
1146         }
1147 }
1148
1149 /* Imprisoned penguins run with %pil == PIL_NORMAL_MAX, but PSTATE_IE
1150  * set, so they can service tlb flush xcalls...
1151  */
1152 extern void prom_world(int);
1153
1154 void smp_penguin_jailcell(int irq, struct pt_regs *regs)
1155 {
1156         clear_softint(1 << irq);
1157
1158         preempt_disable();
1159
1160         __asm__ __volatile__("flushw");
1161         prom_world(1);
1162         atomic_inc(&smp_capture_registry);
1163         membar_safe("#StoreLoad");
1164         while (penguins_are_doing_time)
1165                 rmb();
1166         atomic_dec(&smp_capture_registry);
1167         prom_world(0);
1168
1169         preempt_enable();
1170 }
1171
1172 /* /proc/profile writes can call this, don't __init it please. */
1173 int setup_profiling_timer(unsigned int multiplier)
1174 {
1175         return -EINVAL;
1176 }
1177
1178 void __init smp_prepare_cpus(unsigned int max_cpus)
1179 {
1180 }
1181
1182 void __devinit smp_prepare_boot_cpu(void)
1183 {
1184 }
1185
1186 void __init smp_setup_processor_id(void)
1187 {
1188         if (tlb_type == spitfire)
1189                 xcall_deliver_impl = spitfire_xcall_deliver;
1190         else if (tlb_type == cheetah || tlb_type == cheetah_plus)
1191                 xcall_deliver_impl = cheetah_xcall_deliver;
1192         else
1193                 xcall_deliver_impl = hypervisor_xcall_deliver;
1194 }
1195
1196 void __devinit smp_fill_in_sib_core_maps(void)
1197 {
1198         unsigned int i;
1199
1200         for_each_present_cpu(i) {
1201                 unsigned int j;
1202
1203                 cpus_clear(cpu_core_map[i]);
1204                 if (cpu_data(i).core_id == 0) {
1205                         cpu_set(i, cpu_core_map[i]);
1206                         continue;
1207                 }
1208
1209                 for_each_present_cpu(j) {
1210                         if (cpu_data(i).core_id ==
1211                             cpu_data(j).core_id)
1212                                 cpu_set(j, cpu_core_map[i]);
1213                 }
1214         }
1215
1216         for_each_present_cpu(i) {
1217                 unsigned int j;
1218
1219                 cpus_clear(per_cpu(cpu_sibling_map, i));
1220                 if (cpu_data(i).proc_id == -1) {
1221                         cpu_set(i, per_cpu(cpu_sibling_map, i));
1222                         continue;
1223                 }
1224
1225                 for_each_present_cpu(j) {
1226                         if (cpu_data(i).proc_id ==
1227                             cpu_data(j).proc_id)
1228                                 cpu_set(j, per_cpu(cpu_sibling_map, i));
1229                 }
1230         }
1231 }
1232
1233 int __cpuinit __cpu_up(unsigned int cpu)
1234 {
1235         int ret = smp_boot_one_cpu(cpu);
1236
1237         if (!ret) {
1238                 cpu_set(cpu, smp_commenced_mask);
1239                 while (!cpu_isset(cpu, cpu_online_map))
1240                         mb();
1241                 if (!cpu_isset(cpu, cpu_online_map)) {
1242                         ret = -ENODEV;
1243                 } else {
1244                         /* On SUN4V, writes to %tick and %stick are
1245                          * not allowed.
1246                          */
1247                         if (tlb_type != hypervisor)
1248                                 smp_synchronize_one_tick(cpu);
1249                 }
1250         }
1251         return ret;
1252 }
1253
1254 #ifdef CONFIG_HOTPLUG_CPU
1255 void cpu_play_dead(void)
1256 {
1257         int cpu = smp_processor_id();
1258         unsigned long pstate;
1259
1260         idle_task_exit();
1261
1262         if (tlb_type == hypervisor) {
1263                 struct trap_per_cpu *tb = &trap_block[cpu];
1264
1265                 sun4v_cpu_qconf(HV_CPU_QUEUE_CPU_MONDO,
1266                                 tb->cpu_mondo_pa, 0);
1267                 sun4v_cpu_qconf(HV_CPU_QUEUE_DEVICE_MONDO,
1268                                 tb->dev_mondo_pa, 0);
1269                 sun4v_cpu_qconf(HV_CPU_QUEUE_RES_ERROR,
1270                                 tb->resum_mondo_pa, 0);
1271                 sun4v_cpu_qconf(HV_CPU_QUEUE_NONRES_ERROR,
1272                                 tb->nonresum_mondo_pa, 0);
1273         }
1274
1275         cpu_clear(cpu, smp_commenced_mask);
1276         membar_safe("#Sync");
1277
1278         local_irq_disable();
1279
1280         __asm__ __volatile__(
1281                 "rdpr   %%pstate, %0\n\t"
1282                 "wrpr   %0, %1, %%pstate"
1283                 : "=r" (pstate)
1284                 : "i" (PSTATE_IE));
1285
1286         while (1)
1287                 barrier();
1288 }
1289
1290 int __cpu_disable(void)
1291 {
1292         int cpu = smp_processor_id();
1293         cpuinfo_sparc *c;
1294         int i;
1295
1296         for_each_cpu_mask(i, cpu_core_map[cpu])
1297                 cpu_clear(cpu, cpu_core_map[i]);
1298         cpus_clear(cpu_core_map[cpu]);
1299
1300         for_each_cpu_mask(i, per_cpu(cpu_sibling_map, cpu))
1301                 cpu_clear(cpu, per_cpu(cpu_sibling_map, i));
1302         cpus_clear(per_cpu(cpu_sibling_map, cpu));
1303
1304         c = &cpu_data(cpu);
1305
1306         c->core_id = 0;
1307         c->proc_id = -1;
1308
1309         smp_wmb();
1310
1311         /* Make sure no interrupts point to this cpu.  */
1312         fixup_irqs();
1313
1314         local_irq_enable();
1315         mdelay(1);
1316         local_irq_disable();
1317
1318         ipi_call_lock();
1319         cpu_clear(cpu, cpu_online_map);
1320         ipi_call_unlock();
1321
1322         return 0;
1323 }
1324
1325 void __cpu_die(unsigned int cpu)
1326 {
1327         int i;
1328
1329         for (i = 0; i < 100; i++) {
1330                 smp_rmb();
1331                 if (!cpu_isset(cpu, smp_commenced_mask))
1332                         break;
1333                 msleep(100);
1334         }
1335         if (cpu_isset(cpu, smp_commenced_mask)) {
1336                 printk(KERN_ERR "CPU %u didn't die...\n", cpu);
1337         } else {
1338 #if defined(CONFIG_SUN_LDOMS)
1339                 unsigned long hv_err;
1340                 int limit = 100;
1341
1342                 do {
1343                         hv_err = sun4v_cpu_stop(cpu);
1344                         if (hv_err == HV_EOK) {
1345                                 cpu_clear(cpu, cpu_present_map);
1346                                 break;
1347                         }
1348                 } while (--limit > 0);
1349                 if (limit <= 0) {
1350                         printk(KERN_ERR "sun4v_cpu_stop() fails err=%lu\n",
1351                                hv_err);
1352                 }
1353 #endif
1354         }
1355 }
1356 #endif
1357
1358 void __init smp_cpus_done(unsigned int max_cpus)
1359 {
1360 }
1361
1362 void smp_send_reschedule(int cpu)
1363 {
1364         xcall_deliver((u64) &xcall_receive_signal, 0, 0,
1365                       &cpumask_of_cpu(cpu));
1366 }
1367
1368 void smp_receive_signal_client(int irq, struct pt_regs *regs)
1369 {
1370         clear_softint(1 << irq);
1371 }
1372
1373 /* This is a nop because we capture all other cpus
1374  * anyways when making the PROM active.
1375  */
1376 void smp_send_stop(void)
1377 {
1378 }
1379
1380 unsigned long __per_cpu_base __read_mostly;
1381 unsigned long __per_cpu_shift __read_mostly;
1382
1383 EXPORT_SYMBOL(__per_cpu_base);
1384 EXPORT_SYMBOL(__per_cpu_shift);
1385
1386 void __init real_setup_per_cpu_areas(void)
1387 {
1388         unsigned long paddr, goal, size, i;
1389         char *ptr;
1390
1391         /* Copy section for each CPU (we discard the original) */
1392         goal = PERCPU_ENOUGH_ROOM;
1393
1394         __per_cpu_shift = PAGE_SHIFT;
1395         for (size = PAGE_SIZE; size < goal; size <<= 1UL)
1396                 __per_cpu_shift++;
1397
1398         paddr = lmb_alloc(size * NR_CPUS, PAGE_SIZE);
1399         if (!paddr) {
1400                 prom_printf("Cannot allocate per-cpu memory.\n");
1401                 prom_halt();
1402         }
1403
1404         ptr = __va(paddr);
1405         __per_cpu_base = ptr - __per_cpu_start;
1406
1407         for (i = 0; i < NR_CPUS; i++, ptr += size)
1408                 memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start);
1409
1410         /* Setup %g5 for the boot cpu.  */
1411         __local_per_cpu_offset = __per_cpu_offset(smp_processor_id());
1412 }