include cleanup: Update gfp.h and slab.h includes to prepare for breaking implicit...
[linux-2.6.git] / 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/bootmem.h>
24 #include <linux/vmalloc.h>
25 #include <linux/cpu.h>
26 #include <linux/slab.h>
27
28 #include <asm/head.h>
29 #include <asm/ptrace.h>
30 #include <asm/atomic.h>
31 #include <asm/tlbflush.h>
32 #include <asm/mmu_context.h>
33 #include <asm/cpudata.h>
34 #include <asm/hvtramp.h>
35 #include <asm/io.h>
36 #include <asm/timer.h>
37
38 #include <asm/irq.h>
39 #include <asm/irq_regs.h>
40 #include <asm/page.h>
41 #include <asm/pgtable.h>
42 #include <asm/oplib.h>
43 #include <asm/uaccess.h>
44 #include <asm/starfire.h>
45 #include <asm/tlb.h>
46 #include <asm/sections.h>
47 #include <asm/prom.h>
48 #include <asm/mdesc.h>
49 #include <asm/ldc.h>
50 #include <asm/hypervisor.h>
51
52 #include "cpumap.h"
53
54 int sparc64_multi_core __read_mostly;
55
56 DEFINE_PER_CPU(cpumask_t, cpu_sibling_map) = CPU_MASK_NONE;
57 cpumask_t cpu_core_map[NR_CPUS] __read_mostly =
58         { [0 ... NR_CPUS-1] = CPU_MASK_NONE };
59
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_irq();
126         cpu_set(cpuid, cpu_online_map);
127         ipi_call_unlock_irq();
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, void **descrp)
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         *descrp = hdesc;
306
307         hdesc->cpu = cpu;
308         hdesc->num_mappings = num_kernel_image_mappings;
309
310         tb = &trap_block[cpu];
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         unsigned long entry =
349                 (unsigned long)(&sparc64_cpu_startup);
350         unsigned long cookie =
351                 (unsigned long)(&cpu_new_thread);
352         struct task_struct *p;
353         void *descr = NULL;
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                                             &descr);
368                 else
369 #endif
370                         prom_startcpu_cpuid(cpu, entry, cookie);
371         } else {
372                 struct device_node *dp = of_find_node_by_cpuid(cpu);
373
374                 prom_startcpu(dp->phandle, entry, cookie);
375         }
376
377         for (timeout = 0; timeout < 50000; timeout++) {
378                 if (callin_flag)
379                         break;
380                 udelay(100);
381         }
382
383         if (callin_flag) {
384                 ret = 0;
385         } else {
386                 printk("Processor %d is stuck.\n", cpu);
387                 ret = -ENODEV;
388         }
389         cpu_new_thread = NULL;
390
391         kfree(descr);
392
393         return ret;
394 }
395
396 static void spitfire_xcall_helper(u64 data0, u64 data1, u64 data2, u64 pstate, unsigned long cpu)
397 {
398         u64 result, target;
399         int stuck, tmp;
400
401         if (this_is_starfire) {
402                 /* map to real upaid */
403                 cpu = (((cpu & 0x3c) << 1) |
404                         ((cpu & 0x40) >> 4) |
405                         (cpu & 0x3));
406         }
407
408         target = (cpu << 14) | 0x70;
409 again:
410         /* Ok, this is the real Spitfire Errata #54.
411          * One must read back from a UDB internal register
412          * after writes to the UDB interrupt dispatch, but
413          * before the membar Sync for that write.
414          * So we use the high UDB control register (ASI 0x7f,
415          * ADDR 0x20) for the dummy read. -DaveM
416          */
417         tmp = 0x40;
418         __asm__ __volatile__(
419         "wrpr   %1, %2, %%pstate\n\t"
420         "stxa   %4, [%0] %3\n\t"
421         "stxa   %5, [%0+%8] %3\n\t"
422         "add    %0, %8, %0\n\t"
423         "stxa   %6, [%0+%8] %3\n\t"
424         "membar #Sync\n\t"
425         "stxa   %%g0, [%7] %3\n\t"
426         "membar #Sync\n\t"
427         "mov    0x20, %%g1\n\t"
428         "ldxa   [%%g1] 0x7f, %%g0\n\t"
429         "membar #Sync"
430         : "=r" (tmp)
431         : "r" (pstate), "i" (PSTATE_IE), "i" (ASI_INTR_W),
432           "r" (data0), "r" (data1), "r" (data2), "r" (target),
433           "r" (0x10), "0" (tmp)
434         : "g1");
435
436         /* NOTE: PSTATE_IE is still clear. */
437         stuck = 100000;
438         do {
439                 __asm__ __volatile__("ldxa [%%g0] %1, %0"
440                         : "=r" (result)
441                         : "i" (ASI_INTR_DISPATCH_STAT));
442                 if (result == 0) {
443                         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
444                                              : : "r" (pstate));
445                         return;
446                 }
447                 stuck -= 1;
448                 if (stuck == 0)
449                         break;
450         } while (result & 0x1);
451         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
452                              : : "r" (pstate));
453         if (stuck == 0) {
454                 printk("CPU[%d]: mondo stuckage result[%016llx]\n",
455                        smp_processor_id(), result);
456         } else {
457                 udelay(2);
458                 goto again;
459         }
460 }
461
462 static void spitfire_xcall_deliver(struct trap_per_cpu *tb, int cnt)
463 {
464         u64 *mondo, data0, data1, data2;
465         u16 *cpu_list;
466         u64 pstate;
467         int i;
468
469         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
470         cpu_list = __va(tb->cpu_list_pa);
471         mondo = __va(tb->cpu_mondo_block_pa);
472         data0 = mondo[0];
473         data1 = mondo[1];
474         data2 = mondo[2];
475         for (i = 0; i < cnt; i++)
476                 spitfire_xcall_helper(data0, data1, data2, pstate, cpu_list[i]);
477 }
478
479 /* Cheetah now allows to send the whole 64-bytes of data in the interrupt
480  * packet, but we have no use for that.  However we do take advantage of
481  * the new pipelining feature (ie. dispatch to multiple cpus simultaneously).
482  */
483 static void cheetah_xcall_deliver(struct trap_per_cpu *tb, int cnt)
484 {
485         int nack_busy_id, is_jbus, need_more;
486         u64 *mondo, pstate, ver, busy_mask;
487         u16 *cpu_list;
488
489         cpu_list = __va(tb->cpu_list_pa);
490         mondo = __va(tb->cpu_mondo_block_pa);
491
492         /* Unfortunately, someone at Sun had the brilliant idea to make the
493          * busy/nack fields hard-coded by ITID number for this Ultra-III
494          * derivative processor.
495          */
496         __asm__ ("rdpr %%ver, %0" : "=r" (ver));
497         is_jbus = ((ver >> 32) == __JALAPENO_ID ||
498                    (ver >> 32) == __SERRANO_ID);
499
500         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
501
502 retry:
503         need_more = 0;
504         __asm__ __volatile__("wrpr %0, %1, %%pstate\n\t"
505                              : : "r" (pstate), "i" (PSTATE_IE));
506
507         /* Setup the dispatch data registers. */
508         __asm__ __volatile__("stxa      %0, [%3] %6\n\t"
509                              "stxa      %1, [%4] %6\n\t"
510                              "stxa      %2, [%5] %6\n\t"
511                              "membar    #Sync\n\t"
512                              : /* no outputs */
513                              : "r" (mondo[0]), "r" (mondo[1]), "r" (mondo[2]),
514                                "r" (0x40), "r" (0x50), "r" (0x60),
515                                "i" (ASI_INTR_W));
516
517         nack_busy_id = 0;
518         busy_mask = 0;
519         {
520                 int i;
521
522                 for (i = 0; i < cnt; i++) {
523                         u64 target, nr;
524
525                         nr = cpu_list[i];
526                         if (nr == 0xffff)
527                                 continue;
528
529                         target = (nr << 14) | 0x70;
530                         if (is_jbus) {
531                                 busy_mask |= (0x1UL << (nr * 2));
532                         } else {
533                                 target |= (nack_busy_id << 24);
534                                 busy_mask |= (0x1UL <<
535                                               (nack_busy_id * 2));
536                         }
537                         __asm__ __volatile__(
538                                 "stxa   %%g0, [%0] %1\n\t"
539                                 "membar #Sync\n\t"
540                                 : /* no outputs */
541                                 : "r" (target), "i" (ASI_INTR_W));
542                         nack_busy_id++;
543                         if (nack_busy_id == 32) {
544                                 need_more = 1;
545                                 break;
546                         }
547                 }
548         }
549
550         /* Now, poll for completion. */
551         {
552                 u64 dispatch_stat, nack_mask;
553                 long stuck;
554
555                 stuck = 100000 * nack_busy_id;
556                 nack_mask = busy_mask << 1;
557                 do {
558                         __asm__ __volatile__("ldxa      [%%g0] %1, %0"
559                                              : "=r" (dispatch_stat)
560                                              : "i" (ASI_INTR_DISPATCH_STAT));
561                         if (!(dispatch_stat & (busy_mask | nack_mask))) {
562                                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
563                                                      : : "r" (pstate));
564                                 if (unlikely(need_more)) {
565                                         int i, this_cnt = 0;
566                                         for (i = 0; i < cnt; i++) {
567                                                 if (cpu_list[i] == 0xffff)
568                                                         continue;
569                                                 cpu_list[i] = 0xffff;
570                                                 this_cnt++;
571                                                 if (this_cnt == 32)
572                                                         break;
573                                         }
574                                         goto retry;
575                                 }
576                                 return;
577                         }
578                         if (!--stuck)
579                                 break;
580                 } while (dispatch_stat & busy_mask);
581
582                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
583                                      : : "r" (pstate));
584
585                 if (dispatch_stat & busy_mask) {
586                         /* Busy bits will not clear, continue instead
587                          * of freezing up on this cpu.
588                          */
589                         printk("CPU[%d]: mondo stuckage result[%016llx]\n",
590                                smp_processor_id(), dispatch_stat);
591                 } else {
592                         int i, this_busy_nack = 0;
593
594                         /* Delay some random time with interrupts enabled
595                          * to prevent deadlock.
596                          */
597                         udelay(2 * nack_busy_id);
598
599                         /* Clear out the mask bits for cpus which did not
600                          * NACK us.
601                          */
602                         for (i = 0; i < cnt; i++) {
603                                 u64 check_mask, nr;
604
605                                 nr = cpu_list[i];
606                                 if (nr == 0xffff)
607                                         continue;
608
609                                 if (is_jbus)
610                                         check_mask = (0x2UL << (2*nr));
611                                 else
612                                         check_mask = (0x2UL <<
613                                                       this_busy_nack);
614                                 if ((dispatch_stat & check_mask) == 0)
615                                         cpu_list[i] = 0xffff;
616                                 this_busy_nack += 2;
617                                 if (this_busy_nack == 64)
618                                         break;
619                         }
620
621                         goto retry;
622                 }
623         }
624 }
625
626 /* Multi-cpu list version.  */
627 static void hypervisor_xcall_deliver(struct trap_per_cpu *tb, int cnt)
628 {
629         int retries, this_cpu, prev_sent, i, saw_cpu_error;
630         unsigned long status;
631         u16 *cpu_list;
632
633         this_cpu = smp_processor_id();
634
635         cpu_list = __va(tb->cpu_list_pa);
636
637         saw_cpu_error = 0;
638         retries = 0;
639         prev_sent = 0;
640         do {
641                 int forward_progress, n_sent;
642
643                 status = sun4v_cpu_mondo_send(cnt,
644                                               tb->cpu_list_pa,
645                                               tb->cpu_mondo_block_pa);
646
647                 /* HV_EOK means all cpus received the xcall, we're done.  */
648                 if (likely(status == HV_EOK))
649                         break;
650
651                 /* First, see if we made any forward progress.
652                  *
653                  * The hypervisor indicates successful sends by setting
654                  * cpu list entries to the value 0xffff.
655                  */
656                 n_sent = 0;
657                 for (i = 0; i < cnt; i++) {
658                         if (likely(cpu_list[i] == 0xffff))
659                                 n_sent++;
660                 }
661
662                 forward_progress = 0;
663                 if (n_sent > prev_sent)
664                         forward_progress = 1;
665
666                 prev_sent = n_sent;
667
668                 /* If we get a HV_ECPUERROR, then one or more of the cpus
669                  * in the list are in error state.  Use the cpu_state()
670                  * hypervisor call to find out which cpus are in error state.
671                  */
672                 if (unlikely(status == HV_ECPUERROR)) {
673                         for (i = 0; i < cnt; i++) {
674                                 long err;
675                                 u16 cpu;
676
677                                 cpu = cpu_list[i];
678                                 if (cpu == 0xffff)
679                                         continue;
680
681                                 err = sun4v_cpu_state(cpu);
682                                 if (err == HV_CPU_STATE_ERROR) {
683                                         saw_cpu_error = (cpu + 1);
684                                         cpu_list[i] = 0xffff;
685                                 }
686                         }
687                 } else if (unlikely(status != HV_EWOULDBLOCK))
688                         goto fatal_mondo_error;
689
690                 /* Don't bother rewriting the CPU list, just leave the
691                  * 0xffff and non-0xffff entries in there and the
692                  * hypervisor will do the right thing.
693                  *
694                  * Only advance timeout state if we didn't make any
695                  * forward progress.
696                  */
697                 if (unlikely(!forward_progress)) {
698                         if (unlikely(++retries > 10000))
699                                 goto fatal_mondo_timeout;
700
701                         /* Delay a little bit to let other cpus catch up
702                          * on their cpu mondo queue work.
703                          */
704                         udelay(2 * cnt);
705                 }
706         } while (1);
707
708         if (unlikely(saw_cpu_error))
709                 goto fatal_mondo_cpu_error;
710
711         return;
712
713 fatal_mondo_cpu_error:
714         printk(KERN_CRIT "CPU[%d]: SUN4V mondo cpu error, some target cpus "
715                "(including %d) were in error state\n",
716                this_cpu, saw_cpu_error - 1);
717         return;
718
719 fatal_mondo_timeout:
720         printk(KERN_CRIT "CPU[%d]: SUN4V mondo timeout, no forward "
721                " progress after %d retries.\n",
722                this_cpu, retries);
723         goto dump_cpu_list_and_out;
724
725 fatal_mondo_error:
726         printk(KERN_CRIT "CPU[%d]: Unexpected SUN4V mondo error %lu\n",
727                this_cpu, status);
728         printk(KERN_CRIT "CPU[%d]: Args were cnt(%d) cpulist_pa(%lx) "
729                "mondo_block_pa(%lx)\n",
730                this_cpu, cnt, tb->cpu_list_pa, tb->cpu_mondo_block_pa);
731
732 dump_cpu_list_and_out:
733         printk(KERN_CRIT "CPU[%d]: CPU list [ ", this_cpu);
734         for (i = 0; i < cnt; i++)
735                 printk("%u ", cpu_list[i]);
736         printk("]\n");
737 }
738
739 static void (*xcall_deliver_impl)(struct trap_per_cpu *, int);
740
741 static void xcall_deliver(u64 data0, u64 data1, u64 data2, const cpumask_t *mask)
742 {
743         struct trap_per_cpu *tb;
744         int this_cpu, i, cnt;
745         unsigned long flags;
746         u16 *cpu_list;
747         u64 *mondo;
748
749         /* We have to do this whole thing with interrupts fully disabled.
750          * Otherwise if we send an xcall from interrupt context it will
751          * corrupt both our mondo block and cpu list state.
752          *
753          * One consequence of this is that we cannot use timeout mechanisms
754          * that depend upon interrupts being delivered locally.  So, for
755          * example, we cannot sample jiffies and expect it to advance.
756          *
757          * Fortunately, udelay() uses %stick/%tick so we can use that.
758          */
759         local_irq_save(flags);
760
761         this_cpu = smp_processor_id();
762         tb = &trap_block[this_cpu];
763
764         mondo = __va(tb->cpu_mondo_block_pa);
765         mondo[0] = data0;
766         mondo[1] = data1;
767         mondo[2] = data2;
768         wmb();
769
770         cpu_list = __va(tb->cpu_list_pa);
771
772         /* Setup the initial cpu list.  */
773         cnt = 0;
774         for_each_cpu(i, mask) {
775                 if (i == this_cpu || !cpu_online(i))
776                         continue;
777                 cpu_list[cnt++] = i;
778         }
779
780         if (cnt)
781                 xcall_deliver_impl(tb, cnt);
782
783         local_irq_restore(flags);
784 }
785
786 /* Send cross call to all processors mentioned in MASK_P
787  * except self.  Really, there are only two cases currently,
788  * "&cpu_online_map" and "&mm->cpu_vm_mask".
789  */
790 static void smp_cross_call_masked(unsigned long *func, u32 ctx, u64 data1, u64 data2, const cpumask_t *mask)
791 {
792         u64 data0 = (((u64)ctx)<<32 | (((u64)func) & 0xffffffff));
793
794         xcall_deliver(data0, data1, data2, mask);
795 }
796
797 /* Send cross call to all processors except self. */
798 static void smp_cross_call(unsigned long *func, u32 ctx, u64 data1, u64 data2)
799 {
800         smp_cross_call_masked(func, ctx, data1, data2, &cpu_online_map);
801 }
802
803 extern unsigned long xcall_sync_tick;
804
805 static void smp_start_sync_tick_client(int cpu)
806 {
807         xcall_deliver((u64) &xcall_sync_tick, 0, 0,
808                       &cpumask_of_cpu(cpu));
809 }
810
811 extern unsigned long xcall_call_function;
812
813 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
814 {
815         xcall_deliver((u64) &xcall_call_function, 0, 0, mask);
816 }
817
818 extern unsigned long xcall_call_function_single;
819
820 void arch_send_call_function_single_ipi(int cpu)
821 {
822         xcall_deliver((u64) &xcall_call_function_single, 0, 0,
823                       &cpumask_of_cpu(cpu));
824 }
825
826 void smp_call_function_client(int irq, struct pt_regs *regs)
827 {
828         clear_softint(1 << irq);
829         generic_smp_call_function_interrupt();
830 }
831
832 void smp_call_function_single_client(int irq, struct pt_regs *regs)
833 {
834         clear_softint(1 << irq);
835         generic_smp_call_function_single_interrupt();
836 }
837
838 static void tsb_sync(void *info)
839 {
840         struct trap_per_cpu *tp = &trap_block[raw_smp_processor_id()];
841         struct mm_struct *mm = info;
842
843         /* It is not valid to test "currrent->active_mm == mm" here.
844          *
845          * The value of "current" is not changed atomically with
846          * switch_mm().  But that's OK, we just need to check the
847          * current cpu's trap block PGD physical address.
848          */
849         if (tp->pgd_paddr == __pa(mm->pgd))
850                 tsb_context_switch(mm);
851 }
852
853 void smp_tsb_sync(struct mm_struct *mm)
854 {
855         smp_call_function_many(mm_cpumask(mm), tsb_sync, mm, 1);
856 }
857
858 extern unsigned long xcall_flush_tlb_mm;
859 extern unsigned long xcall_flush_tlb_pending;
860 extern unsigned long xcall_flush_tlb_kernel_range;
861 extern unsigned long xcall_fetch_glob_regs;
862 extern unsigned long xcall_receive_signal;
863 extern unsigned long xcall_new_mmu_context_version;
864 #ifdef CONFIG_KGDB
865 extern unsigned long xcall_kgdb_capture;
866 #endif
867
868 #ifdef DCACHE_ALIASING_POSSIBLE
869 extern unsigned long xcall_flush_dcache_page_cheetah;
870 #endif
871 extern unsigned long xcall_flush_dcache_page_spitfire;
872
873 #ifdef CONFIG_DEBUG_DCFLUSH
874 extern atomic_t dcpage_flushes;
875 extern atomic_t dcpage_flushes_xcall;
876 #endif
877
878 static inline void __local_flush_dcache_page(struct page *page)
879 {
880 #ifdef DCACHE_ALIASING_POSSIBLE
881         __flush_dcache_page(page_address(page),
882                             ((tlb_type == spitfire) &&
883                              page_mapping(page) != NULL));
884 #else
885         if (page_mapping(page) != NULL &&
886             tlb_type == spitfire)
887                 __flush_icache_page(__pa(page_address(page)));
888 #endif
889 }
890
891 void smp_flush_dcache_page_impl(struct page *page, int cpu)
892 {
893         int this_cpu;
894
895         if (tlb_type == hypervisor)
896                 return;
897
898 #ifdef CONFIG_DEBUG_DCFLUSH
899         atomic_inc(&dcpage_flushes);
900 #endif
901
902         this_cpu = get_cpu();
903
904         if (cpu == this_cpu) {
905                 __local_flush_dcache_page(page);
906         } else if (cpu_online(cpu)) {
907                 void *pg_addr = page_address(page);
908                 u64 data0 = 0;
909
910                 if (tlb_type == spitfire) {
911                         data0 = ((u64)&xcall_flush_dcache_page_spitfire);
912                         if (page_mapping(page) != NULL)
913                                 data0 |= ((u64)1 << 32);
914                 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
915 #ifdef DCACHE_ALIASING_POSSIBLE
916                         data0 = ((u64)&xcall_flush_dcache_page_cheetah);
917 #endif
918                 }
919                 if (data0) {
920                         xcall_deliver(data0, __pa(pg_addr),
921                                       (u64) pg_addr, &cpumask_of_cpu(cpu));
922 #ifdef CONFIG_DEBUG_DCFLUSH
923                         atomic_inc(&dcpage_flushes_xcall);
924 #endif
925                 }
926         }
927
928         put_cpu();
929 }
930
931 void flush_dcache_page_all(struct mm_struct *mm, struct page *page)
932 {
933         void *pg_addr;
934         int this_cpu;
935         u64 data0;
936
937         if (tlb_type == hypervisor)
938                 return;
939
940         this_cpu = get_cpu();
941
942 #ifdef CONFIG_DEBUG_DCFLUSH
943         atomic_inc(&dcpage_flushes);
944 #endif
945         data0 = 0;
946         pg_addr = page_address(page);
947         if (tlb_type == spitfire) {
948                 data0 = ((u64)&xcall_flush_dcache_page_spitfire);
949                 if (page_mapping(page) != NULL)
950                         data0 |= ((u64)1 << 32);
951         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
952 #ifdef DCACHE_ALIASING_POSSIBLE
953                 data0 = ((u64)&xcall_flush_dcache_page_cheetah);
954 #endif
955         }
956         if (data0) {
957                 xcall_deliver(data0, __pa(pg_addr),
958                               (u64) pg_addr, &cpu_online_map);
959 #ifdef CONFIG_DEBUG_DCFLUSH
960                 atomic_inc(&dcpage_flushes_xcall);
961 #endif
962         }
963         __local_flush_dcache_page(page);
964
965         put_cpu();
966 }
967
968 void smp_new_mmu_context_version_client(int irq, struct pt_regs *regs)
969 {
970         struct mm_struct *mm;
971         unsigned long flags;
972
973         clear_softint(1 << irq);
974
975         /* See if we need to allocate a new TLB context because
976          * the version of the one we are using is now out of date.
977          */
978         mm = current->active_mm;
979         if (unlikely(!mm || (mm == &init_mm)))
980                 return;
981
982         spin_lock_irqsave(&mm->context.lock, flags);
983
984         if (unlikely(!CTX_VALID(mm->context)))
985                 get_new_mmu_context(mm);
986
987         spin_unlock_irqrestore(&mm->context.lock, flags);
988
989         load_secondary_context(mm);
990         __flush_tlb_mm(CTX_HWBITS(mm->context),
991                        SECONDARY_CONTEXT);
992 }
993
994 void smp_new_mmu_context_version(void)
995 {
996         smp_cross_call(&xcall_new_mmu_context_version, 0, 0, 0);
997 }
998
999 #ifdef CONFIG_KGDB
1000 void kgdb_roundup_cpus(unsigned long flags)
1001 {
1002         smp_cross_call(&xcall_kgdb_capture, 0, 0, 0);
1003 }
1004 #endif
1005
1006 void smp_fetch_global_regs(void)
1007 {
1008         smp_cross_call(&xcall_fetch_glob_regs, 0, 0, 0);
1009 }
1010
1011 /* We know that the window frames of the user have been flushed
1012  * to the stack before we get here because all callers of us
1013  * are flush_tlb_*() routines, and these run after flush_cache_*()
1014  * which performs the flushw.
1015  *
1016  * The SMP TLB coherency scheme we use works as follows:
1017  *
1018  * 1) mm->cpu_vm_mask is a bit mask of which cpus an address
1019  *    space has (potentially) executed on, this is the heuristic
1020  *    we use to avoid doing cross calls.
1021  *
1022  *    Also, for flushing from kswapd and also for clones, we
1023  *    use cpu_vm_mask as the list of cpus to make run the TLB.
1024  *
1025  * 2) TLB context numbers are shared globally across all processors
1026  *    in the system, this allows us to play several games to avoid
1027  *    cross calls.
1028  *
1029  *    One invariant is that when a cpu switches to a process, and
1030  *    that processes tsk->active_mm->cpu_vm_mask does not have the
1031  *    current cpu's bit set, that tlb context is flushed locally.
1032  *
1033  *    If the address space is non-shared (ie. mm->count == 1) we avoid
1034  *    cross calls when we want to flush the currently running process's
1035  *    tlb state.  This is done by clearing all cpu bits except the current
1036  *    processor's in current->mm->cpu_vm_mask and performing the
1037  *    flush locally only.  This will force any subsequent cpus which run
1038  *    this task to flush the context from the local tlb if the process
1039  *    migrates to another cpu (again).
1040  *
1041  * 3) For shared address spaces (threads) and swapping we bite the
1042  *    bullet for most cases and perform the cross call (but only to
1043  *    the cpus listed in cpu_vm_mask).
1044  *
1045  *    The performance gain from "optimizing" away the cross call for threads is
1046  *    questionable (in theory the big win for threads is the massive sharing of
1047  *    address space state across processors).
1048  */
1049
1050 /* This currently is only used by the hugetlb arch pre-fault
1051  * hook on UltraSPARC-III+ and later when changing the pagesize
1052  * bits of the context register for an address space.
1053  */
1054 void smp_flush_tlb_mm(struct mm_struct *mm)
1055 {
1056         u32 ctx = CTX_HWBITS(mm->context);
1057         int cpu = get_cpu();
1058
1059         if (atomic_read(&mm->mm_users) == 1) {
1060                 cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
1061                 goto local_flush_and_out;
1062         }
1063
1064         smp_cross_call_masked(&xcall_flush_tlb_mm,
1065                               ctx, 0, 0,
1066                               mm_cpumask(mm));
1067
1068 local_flush_and_out:
1069         __flush_tlb_mm(ctx, SECONDARY_CONTEXT);
1070
1071         put_cpu();
1072 }
1073
1074 void smp_flush_tlb_pending(struct mm_struct *mm, unsigned long nr, unsigned long *vaddrs)
1075 {
1076         u32 ctx = CTX_HWBITS(mm->context);
1077         int cpu = get_cpu();
1078
1079         if (mm == current->mm && atomic_read(&mm->mm_users) == 1)
1080                 cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
1081         else
1082                 smp_cross_call_masked(&xcall_flush_tlb_pending,
1083                                       ctx, nr, (unsigned long) vaddrs,
1084                                       mm_cpumask(mm));
1085
1086         __flush_tlb_pending(ctx, nr, vaddrs);
1087
1088         put_cpu();
1089 }
1090
1091 void smp_flush_tlb_kernel_range(unsigned long start, unsigned long end)
1092 {
1093         start &= PAGE_MASK;
1094         end    = PAGE_ALIGN(end);
1095         if (start != end) {
1096                 smp_cross_call(&xcall_flush_tlb_kernel_range,
1097                                0, start, end);
1098
1099                 __flush_tlb_kernel_range(start, end);
1100         }
1101 }
1102
1103 /* CPU capture. */
1104 /* #define CAPTURE_DEBUG */
1105 extern unsigned long xcall_capture;
1106
1107 static atomic_t smp_capture_depth = ATOMIC_INIT(0);
1108 static atomic_t smp_capture_registry = ATOMIC_INIT(0);
1109 static unsigned long penguins_are_doing_time;
1110
1111 void smp_capture(void)
1112 {
1113         int result = atomic_add_ret(1, &smp_capture_depth);
1114
1115         if (result == 1) {
1116                 int ncpus = num_online_cpus();
1117
1118 #ifdef CAPTURE_DEBUG
1119                 printk("CPU[%d]: Sending penguins to jail...",
1120                        smp_processor_id());
1121 #endif
1122                 penguins_are_doing_time = 1;
1123                 atomic_inc(&smp_capture_registry);
1124                 smp_cross_call(&xcall_capture, 0, 0, 0);
1125                 while (atomic_read(&smp_capture_registry) != ncpus)
1126                         rmb();
1127 #ifdef CAPTURE_DEBUG
1128                 printk("done\n");
1129 #endif
1130         }
1131 }
1132
1133 void smp_release(void)
1134 {
1135         if (atomic_dec_and_test(&smp_capture_depth)) {
1136 #ifdef CAPTURE_DEBUG
1137                 printk("CPU[%d]: Giving pardon to "
1138                        "imprisoned penguins\n",
1139                        smp_processor_id());
1140 #endif
1141                 penguins_are_doing_time = 0;
1142                 membar_safe("#StoreLoad");
1143                 atomic_dec(&smp_capture_registry);
1144         }
1145 }
1146
1147 /* Imprisoned penguins run with %pil == PIL_NORMAL_MAX, but PSTATE_IE
1148  * set, so they can service tlb flush xcalls...
1149  */
1150 extern void prom_world(int);
1151
1152 void smp_penguin_jailcell(int irq, struct pt_regs *regs)
1153 {
1154         clear_softint(1 << irq);
1155
1156         preempt_disable();
1157
1158         __asm__ __volatile__("flushw");
1159         prom_world(1);
1160         atomic_inc(&smp_capture_registry);
1161         membar_safe("#StoreLoad");
1162         while (penguins_are_doing_time)
1163                 rmb();
1164         atomic_dec(&smp_capture_registry);
1165         prom_world(0);
1166
1167         preempt_enable();
1168 }
1169
1170 /* /proc/profile writes can call this, don't __init it please. */
1171 int setup_profiling_timer(unsigned int multiplier)
1172 {
1173         return -EINVAL;
1174 }
1175
1176 void __init smp_prepare_cpus(unsigned int max_cpus)
1177 {
1178 }
1179
1180 void __devinit smp_prepare_boot_cpu(void)
1181 {
1182 }
1183
1184 void __init smp_setup_processor_id(void)
1185 {
1186         if (tlb_type == spitfire)
1187                 xcall_deliver_impl = spitfire_xcall_deliver;
1188         else if (tlb_type == cheetah || tlb_type == cheetah_plus)
1189                 xcall_deliver_impl = cheetah_xcall_deliver;
1190         else
1191                 xcall_deliver_impl = hypervisor_xcall_deliver;
1192 }
1193
1194 void __devinit smp_fill_in_sib_core_maps(void)
1195 {
1196         unsigned int i;
1197
1198         for_each_present_cpu(i) {
1199                 unsigned int j;
1200
1201                 cpus_clear(cpu_core_map[i]);
1202                 if (cpu_data(i).core_id == 0) {
1203                         cpu_set(i, cpu_core_map[i]);
1204                         continue;
1205                 }
1206
1207                 for_each_present_cpu(j) {
1208                         if (cpu_data(i).core_id ==
1209                             cpu_data(j).core_id)
1210                                 cpu_set(j, cpu_core_map[i]);
1211                 }
1212         }
1213
1214         for_each_present_cpu(i) {
1215                 unsigned int j;
1216
1217                 cpus_clear(per_cpu(cpu_sibling_map, i));
1218                 if (cpu_data(i).proc_id == -1) {
1219                         cpu_set(i, per_cpu(cpu_sibling_map, i));
1220                         continue;
1221                 }
1222
1223                 for_each_present_cpu(j) {
1224                         if (cpu_data(i).proc_id ==
1225                             cpu_data(j).proc_id)
1226                                 cpu_set(j, per_cpu(cpu_sibling_map, i));
1227                 }
1228         }
1229 }
1230
1231 int __cpuinit __cpu_up(unsigned int cpu)
1232 {
1233         int ret = smp_boot_one_cpu(cpu);
1234
1235         if (!ret) {
1236                 cpu_set(cpu, smp_commenced_mask);
1237                 while (!cpu_isset(cpu, cpu_online_map))
1238                         mb();
1239                 if (!cpu_isset(cpu, cpu_online_map)) {
1240                         ret = -ENODEV;
1241                 } else {
1242                         /* On SUN4V, writes to %tick and %stick are
1243                          * not allowed.
1244                          */
1245                         if (tlb_type != hypervisor)
1246                                 smp_synchronize_one_tick(cpu);
1247                 }
1248         }
1249         return ret;
1250 }
1251
1252 #ifdef CONFIG_HOTPLUG_CPU
1253 void cpu_play_dead(void)
1254 {
1255         int cpu = smp_processor_id();
1256         unsigned long pstate;
1257
1258         idle_task_exit();
1259
1260         if (tlb_type == hypervisor) {
1261                 struct trap_per_cpu *tb = &trap_block[cpu];
1262
1263                 sun4v_cpu_qconf(HV_CPU_QUEUE_CPU_MONDO,
1264                                 tb->cpu_mondo_pa, 0);
1265                 sun4v_cpu_qconf(HV_CPU_QUEUE_DEVICE_MONDO,
1266                                 tb->dev_mondo_pa, 0);
1267                 sun4v_cpu_qconf(HV_CPU_QUEUE_RES_ERROR,
1268                                 tb->resum_mondo_pa, 0);
1269                 sun4v_cpu_qconf(HV_CPU_QUEUE_NONRES_ERROR,
1270                                 tb->nonresum_mondo_pa, 0);
1271         }
1272
1273         cpu_clear(cpu, smp_commenced_mask);
1274         membar_safe("#Sync");
1275
1276         local_irq_disable();
1277
1278         __asm__ __volatile__(
1279                 "rdpr   %%pstate, %0\n\t"
1280                 "wrpr   %0, %1, %%pstate"
1281                 : "=r" (pstate)
1282                 : "i" (PSTATE_IE));
1283
1284         while (1)
1285                 barrier();
1286 }
1287
1288 int __cpu_disable(void)
1289 {
1290         int cpu = smp_processor_id();
1291         cpuinfo_sparc *c;
1292         int i;
1293
1294         for_each_cpu_mask(i, cpu_core_map[cpu])
1295                 cpu_clear(cpu, cpu_core_map[i]);
1296         cpus_clear(cpu_core_map[cpu]);
1297
1298         for_each_cpu_mask(i, per_cpu(cpu_sibling_map, cpu))
1299                 cpu_clear(cpu, per_cpu(cpu_sibling_map, i));
1300         cpus_clear(per_cpu(cpu_sibling_map, cpu));
1301
1302         c = &cpu_data(cpu);
1303
1304         c->core_id = 0;
1305         c->proc_id = -1;
1306
1307         smp_wmb();
1308
1309         /* Make sure no interrupts point to this cpu.  */
1310         fixup_irqs();
1311
1312         local_irq_enable();
1313         mdelay(1);
1314         local_irq_disable();
1315
1316         ipi_call_lock();
1317         cpu_clear(cpu, cpu_online_map);
1318         ipi_call_unlock();
1319
1320         cpu_map_rebuild();
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 /**
1381  * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu
1382  * @cpu: cpu to allocate for
1383  * @size: size allocation in bytes
1384  * @align: alignment
1385  *
1386  * Allocate @size bytes aligned at @align for cpu @cpu.  This wrapper
1387  * does the right thing for NUMA regardless of the current
1388  * configuration.
1389  *
1390  * RETURNS:
1391  * Pointer to the allocated area on success, NULL on failure.
1392  */
1393 static void * __init pcpu_alloc_bootmem(unsigned int cpu, size_t size,
1394                                         size_t align)
1395 {
1396         const unsigned long goal = __pa(MAX_DMA_ADDRESS);
1397 #ifdef CONFIG_NEED_MULTIPLE_NODES
1398         int node = cpu_to_node(cpu);
1399         void *ptr;
1400
1401         if (!node_online(node) || !NODE_DATA(node)) {
1402                 ptr = __alloc_bootmem(size, align, goal);
1403                 pr_info("cpu %d has no node %d or node-local memory\n",
1404                         cpu, node);
1405                 pr_debug("per cpu data for cpu%d %lu bytes at %016lx\n",
1406                          cpu, size, __pa(ptr));
1407         } else {
1408                 ptr = __alloc_bootmem_node(NODE_DATA(node),
1409                                            size, align, goal);
1410                 pr_debug("per cpu data for cpu%d %lu bytes on node%d at "
1411                          "%016lx\n", cpu, size, node, __pa(ptr));
1412         }
1413         return ptr;
1414 #else
1415         return __alloc_bootmem(size, align, goal);
1416 #endif
1417 }
1418
1419 static void __init pcpu_free_bootmem(void *ptr, size_t size)
1420 {
1421         free_bootmem(__pa(ptr), size);
1422 }
1423
1424 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
1425 {
1426         if (cpu_to_node(from) == cpu_to_node(to))
1427                 return LOCAL_DISTANCE;
1428         else
1429                 return REMOTE_DISTANCE;
1430 }
1431
1432 static void __init pcpu_populate_pte(unsigned long addr)
1433 {
1434         pgd_t *pgd = pgd_offset_k(addr);
1435         pud_t *pud;
1436         pmd_t *pmd;
1437
1438         pud = pud_offset(pgd, addr);
1439         if (pud_none(*pud)) {
1440                 pmd_t *new;
1441
1442                 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1443                 pud_populate(&init_mm, pud, new);
1444         }
1445
1446         pmd = pmd_offset(pud, addr);
1447         if (!pmd_present(*pmd)) {
1448                 pte_t *new;
1449
1450                 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1451                 pmd_populate_kernel(&init_mm, pmd, new);
1452         }
1453 }
1454
1455 void __init setup_per_cpu_areas(void)
1456 {
1457         unsigned long delta;
1458         unsigned int cpu;
1459         int rc = -EINVAL;
1460
1461         if (pcpu_chosen_fc != PCPU_FC_PAGE) {
1462                 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
1463                                             PERCPU_DYNAMIC_RESERVE, 4 << 20,
1464                                             pcpu_cpu_distance,
1465                                             pcpu_alloc_bootmem,
1466                                             pcpu_free_bootmem);
1467                 if (rc)
1468                         pr_warning("PERCPU: %s allocator failed (%d), "
1469                                    "falling back to page size\n",
1470                                    pcpu_fc_names[pcpu_chosen_fc], rc);
1471         }
1472         if (rc < 0)
1473                 rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE,
1474                                            pcpu_alloc_bootmem,
1475                                            pcpu_free_bootmem,
1476                                            pcpu_populate_pte);
1477         if (rc < 0)
1478                 panic("cannot initialize percpu area (err=%d)", rc);
1479
1480         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1481         for_each_possible_cpu(cpu)
1482                 __per_cpu_offset(cpu) = delta + pcpu_unit_offsets[cpu];
1483
1484         /* Setup %g5 for the boot cpu.  */
1485         __local_per_cpu_offset = __per_cpu_offset(smp_processor_id());
1486
1487         of_fill_in_cpu_data();
1488         if (tlb_type == hypervisor)
1489                 mdesc_fill_in_cpu_data(cpu_all_mask);
1490 }