Merge branch 'master' of /home/davem/src/GIT/linux-2.6/
[linux-3.10.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/ftrace.h>
26 #include <linux/cpu.h>
27 #include <linux/slab.h>
28
29 #include <asm/head.h>
30 #include <asm/ptrace.h>
31 #include <asm/atomic.h>
32 #include <asm/tlbflush.h>
33 #include <asm/mmu_context.h>
34 #include <asm/cpudata.h>
35 #include <asm/hvtramp.h>
36 #include <asm/io.h>
37 #include <asm/timer.h>
38
39 #include <asm/irq.h>
40 #include <asm/irq_regs.h>
41 #include <asm/page.h>
42 #include <asm/pgtable.h>
43 #include <asm/oplib.h>
44 #include <asm/uaccess.h>
45 #include <asm/starfire.h>
46 #include <asm/tlb.h>
47 #include <asm/sections.h>
48 #include <asm/prom.h>
49 #include <asm/mdesc.h>
50 #include <asm/ldc.h>
51 #include <asm/hypervisor.h>
52
53 #include "cpumap.h"
54
55 int sparc64_multi_core __read_mostly;
56
57 DEFINE_PER_CPU(cpumask_t, cpu_sibling_map) = CPU_MASK_NONE;
58 cpumask_t cpu_core_map[NR_CPUS] __read_mostly =
59         { [0 ... NR_CPUS-1] = CPU_MASK_NONE };
60
61 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
62 EXPORT_SYMBOL(cpu_core_map);
63
64 static cpumask_t smp_commenced_mask;
65
66 void smp_info(struct seq_file *m)
67 {
68         int i;
69         
70         seq_printf(m, "State:\n");
71         for_each_online_cpu(i)
72                 seq_printf(m, "CPU%d:\t\tonline\n", i);
73 }
74
75 void smp_bogo(struct seq_file *m)
76 {
77         int i;
78         
79         for_each_online_cpu(i)
80                 seq_printf(m,
81                            "Cpu%dClkTck\t: %016lx\n",
82                            i, cpu_data(i).clock_tick);
83 }
84
85 extern void setup_sparc64_timer(void);
86
87 static volatile unsigned long callin_flag = 0;
88
89 void __cpuinit smp_callin(void)
90 {
91         int cpuid = hard_smp_processor_id();
92
93         __local_per_cpu_offset = __per_cpu_offset(cpuid);
94
95         if (tlb_type == hypervisor)
96                 sun4v_ktsb_register();
97
98         __flush_tlb_all();
99
100         setup_sparc64_timer();
101
102         if (cheetah_pcache_forced_on)
103                 cheetah_enable_pcache();
104
105         local_irq_enable();
106
107         callin_flag = 1;
108         __asm__ __volatile__("membar #Sync\n\t"
109                              "flush  %%g6" : : : "memory");
110
111         /* Clear this or we will die instantly when we
112          * schedule back to this idler...
113          */
114         current_thread_info()->new_child = 0;
115
116         /* Attach to the address space of init_task. */
117         atomic_inc(&init_mm.mm_count);
118         current->active_mm = &init_mm;
119
120         /* inform the notifiers about the new cpu */
121         notify_cpu_starting(cpuid);
122
123         while (!cpu_isset(cpuid, smp_commenced_mask))
124                 rmb();
125
126         ipi_call_lock_irq();
127         cpu_set(cpuid, cpu_online_map);
128         ipi_call_unlock_irq();
129
130         /* idle thread is expected to have preempt disabled */
131         preempt_disable();
132 }
133
134 void cpu_panic(void)
135 {
136         printk("CPU[%d]: Returns from cpu_idle!\n", smp_processor_id());
137         panic("SMP bolixed\n");
138 }
139
140 /* This tick register synchronization scheme is taken entirely from
141  * the ia64 port, see arch/ia64/kernel/smpboot.c for details and credit.
142  *
143  * The only change I've made is to rework it so that the master
144  * initiates the synchonization instead of the slave. -DaveM
145  */
146
147 #define MASTER  0
148 #define SLAVE   (SMP_CACHE_BYTES/sizeof(unsigned long))
149
150 #define NUM_ROUNDS      64      /* magic value */
151 #define NUM_ITERS       5       /* likewise */
152
153 static DEFINE_SPINLOCK(itc_sync_lock);
154 static unsigned long go[SLAVE + 1];
155
156 #define DEBUG_TICK_SYNC 0
157
158 static inline long get_delta (long *rt, long *master)
159 {
160         unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
161         unsigned long tcenter, t0, t1, tm;
162         unsigned long i;
163
164         for (i = 0; i < NUM_ITERS; i++) {
165                 t0 = tick_ops->get_tick();
166                 go[MASTER] = 1;
167                 membar_safe("#StoreLoad");
168                 while (!(tm = go[SLAVE]))
169                         rmb();
170                 go[SLAVE] = 0;
171                 wmb();
172                 t1 = tick_ops->get_tick();
173
174                 if (t1 - t0 < best_t1 - best_t0)
175                         best_t0 = t0, best_t1 = t1, best_tm = tm;
176         }
177
178         *rt = best_t1 - best_t0;
179         *master = best_tm - best_t0;
180
181         /* average best_t0 and best_t1 without overflow: */
182         tcenter = (best_t0/2 + best_t1/2);
183         if (best_t0 % 2 + best_t1 % 2 == 2)
184                 tcenter++;
185         return tcenter - best_tm;
186 }
187
188 void smp_synchronize_tick_client(void)
189 {
190         long i, delta, adj, adjust_latency = 0, done = 0;
191         unsigned long flags, rt, master_time_stamp, bound;
192 #if DEBUG_TICK_SYNC
193         struct {
194                 long rt;        /* roundtrip time */
195                 long master;    /* master's timestamp */
196                 long diff;      /* difference between midpoint and master's timestamp */
197                 long lat;       /* estimate of itc adjustment latency */
198         } t[NUM_ROUNDS];
199 #endif
200
201         go[MASTER] = 1;
202
203         while (go[MASTER])
204                 rmb();
205
206         local_irq_save(flags);
207         {
208                 for (i = 0; i < NUM_ROUNDS; i++) {
209                         delta = get_delta(&rt, &master_time_stamp);
210                         if (delta == 0) {
211                                 done = 1;       /* let's lock on to this... */
212                                 bound = rt;
213                         }
214
215                         if (!done) {
216                                 if (i > 0) {
217                                         adjust_latency += -delta;
218                                         adj = -delta + adjust_latency/4;
219                                 } else
220                                         adj = -delta;
221
222                                 tick_ops->add_tick(adj);
223                         }
224 #if DEBUG_TICK_SYNC
225                         t[i].rt = rt;
226                         t[i].master = master_time_stamp;
227                         t[i].diff = delta;
228                         t[i].lat = adjust_latency/4;
229 #endif
230                 }
231         }
232         local_irq_restore(flags);
233
234 #if DEBUG_TICK_SYNC
235         for (i = 0; i < NUM_ROUNDS; i++)
236                 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
237                        t[i].rt, t[i].master, t[i].diff, t[i].lat);
238 #endif
239
240         printk(KERN_INFO "CPU %d: synchronized TICK with master CPU "
241                "(last diff %ld cycles, maxerr %lu cycles)\n",
242                smp_processor_id(), delta, rt);
243 }
244
245 static void smp_start_sync_tick_client(int cpu);
246
247 static void smp_synchronize_one_tick(int cpu)
248 {
249         unsigned long flags, i;
250
251         go[MASTER] = 0;
252
253         smp_start_sync_tick_client(cpu);
254
255         /* wait for client to be ready */
256         while (!go[MASTER])
257                 rmb();
258
259         /* now let the client proceed into his loop */
260         go[MASTER] = 0;
261         membar_safe("#StoreLoad");
262
263         spin_lock_irqsave(&itc_sync_lock, flags);
264         {
265                 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; i++) {
266                         while (!go[MASTER])
267                                 rmb();
268                         go[MASTER] = 0;
269                         wmb();
270                         go[SLAVE] = tick_ops->get_tick();
271                         membar_safe("#StoreLoad");
272                 }
273         }
274         spin_unlock_irqrestore(&itc_sync_lock, flags);
275 }
276
277 #if defined(CONFIG_SUN_LDOMS) && defined(CONFIG_HOTPLUG_CPU)
278 /* XXX Put this in some common place. XXX */
279 static unsigned long kimage_addr_to_ra(void *p)
280 {
281         unsigned long val = (unsigned long) p;
282
283         return kern_base + (val - KERNBASE);
284 }
285
286 static void __cpuinit ldom_startcpu_cpuid(unsigned int cpu, unsigned long thread_reg, void **descrp)
287 {
288         extern unsigned long sparc64_ttable_tl0;
289         extern unsigned long kern_locked_tte_data;
290         struct hvtramp_descr *hdesc;
291         unsigned long trampoline_ra;
292         struct trap_per_cpu *tb;
293         u64 tte_vaddr, tte_data;
294         unsigned long hv_err;
295         int i;
296
297         hdesc = kzalloc(sizeof(*hdesc) +
298                         (sizeof(struct hvtramp_mapping) *
299                          num_kernel_image_mappings - 1),
300                         GFP_KERNEL);
301         if (!hdesc) {
302                 printk(KERN_ERR "ldom_startcpu_cpuid: Cannot allocate "
303                        "hvtramp_descr.\n");
304                 return;
305         }
306         *descrp = hdesc;
307
308         hdesc->cpu = cpu;
309         hdesc->num_mappings = num_kernel_image_mappings;
310
311         tb = &trap_block[cpu];
312
313         hdesc->fault_info_va = (unsigned long) &tb->fault_info;
314         hdesc->fault_info_pa = kimage_addr_to_ra(&tb->fault_info);
315
316         hdesc->thread_reg = thread_reg;
317
318         tte_vaddr = (unsigned long) KERNBASE;
319         tte_data = kern_locked_tte_data;
320
321         for (i = 0; i < hdesc->num_mappings; i++) {
322                 hdesc->maps[i].vaddr = tte_vaddr;
323                 hdesc->maps[i].tte   = tte_data;
324                 tte_vaddr += 0x400000;
325                 tte_data  += 0x400000;
326         }
327
328         trampoline_ra = kimage_addr_to_ra(hv_cpu_startup);
329
330         hv_err = sun4v_cpu_start(cpu, trampoline_ra,
331                                  kimage_addr_to_ra(&sparc64_ttable_tl0),
332                                  __pa(hdesc));
333         if (hv_err)
334                 printk(KERN_ERR "ldom_startcpu_cpuid: sun4v_cpu_start() "
335                        "gives error %lu\n", hv_err);
336 }
337 #endif
338
339 extern unsigned long sparc64_cpu_startup;
340
341 /* The OBP cpu startup callback truncates the 3rd arg cookie to
342  * 32-bits (I think) so to be safe we have it read the pointer
343  * contained here so we work on >4GB machines. -DaveM
344  */
345 static struct thread_info *cpu_new_thread = NULL;
346
347 static int __cpuinit smp_boot_one_cpu(unsigned int cpu)
348 {
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         void *descr = NULL;
355         int timeout, ret;
356
357         p = fork_idle(cpu);
358         if (IS_ERR(p))
359                 return PTR_ERR(p);
360         callin_flag = 0;
361         cpu_new_thread = task_thread_info(p);
362
363         if (tlb_type == hypervisor) {
364 #if defined(CONFIG_SUN_LDOMS) && defined(CONFIG_HOTPLUG_CPU)
365                 if (ldom_domaining_enabled)
366                         ldom_startcpu_cpuid(cpu,
367                                             (unsigned long) cpu_new_thread,
368                                             &descr);
369                 else
370 #endif
371                         prom_startcpu_cpuid(cpu, entry, cookie);
372         } else {
373                 struct device_node *dp = of_find_node_by_cpuid(cpu);
374
375                 prom_startcpu(dp->phandle, entry, cookie);
376         }
377
378         for (timeout = 0; timeout < 50000; timeout++) {
379                 if (callin_flag)
380                         break;
381                 udelay(100);
382         }
383
384         if (callin_flag) {
385                 ret = 0;
386         } else {
387                 printk("Processor %d is stuck.\n", cpu);
388                 ret = -ENODEV;
389         }
390         cpu_new_thread = NULL;
391
392         kfree(descr);
393
394         return ret;
395 }
396
397 static void spitfire_xcall_helper(u64 data0, u64 data1, u64 data2, u64 pstate, unsigned long cpu)
398 {
399         u64 result, target;
400         int stuck, tmp;
401
402         if (this_is_starfire) {
403                 /* map to real upaid */
404                 cpu = (((cpu & 0x3c) << 1) |
405                         ((cpu & 0x40) >> 4) |
406                         (cpu & 0x3));
407         }
408
409         target = (cpu << 14) | 0x70;
410 again:
411         /* Ok, this is the real Spitfire Errata #54.
412          * One must read back from a UDB internal register
413          * after writes to the UDB interrupt dispatch, but
414          * before the membar Sync for that write.
415          * So we use the high UDB control register (ASI 0x7f,
416          * ADDR 0x20) for the dummy read. -DaveM
417          */
418         tmp = 0x40;
419         __asm__ __volatile__(
420         "wrpr   %1, %2, %%pstate\n\t"
421         "stxa   %4, [%0] %3\n\t"
422         "stxa   %5, [%0+%8] %3\n\t"
423         "add    %0, %8, %0\n\t"
424         "stxa   %6, [%0+%8] %3\n\t"
425         "membar #Sync\n\t"
426         "stxa   %%g0, [%7] %3\n\t"
427         "membar #Sync\n\t"
428         "mov    0x20, %%g1\n\t"
429         "ldxa   [%%g1] 0x7f, %%g0\n\t"
430         "membar #Sync"
431         : "=r" (tmp)
432         : "r" (pstate), "i" (PSTATE_IE), "i" (ASI_INTR_W),
433           "r" (data0), "r" (data1), "r" (data2), "r" (target),
434           "r" (0x10), "0" (tmp)
435         : "g1");
436
437         /* NOTE: PSTATE_IE is still clear. */
438         stuck = 100000;
439         do {
440                 __asm__ __volatile__("ldxa [%%g0] %1, %0"
441                         : "=r" (result)
442                         : "i" (ASI_INTR_DISPATCH_STAT));
443                 if (result == 0) {
444                         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
445                                              : : "r" (pstate));
446                         return;
447                 }
448                 stuck -= 1;
449                 if (stuck == 0)
450                         break;
451         } while (result & 0x1);
452         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
453                              : : "r" (pstate));
454         if (stuck == 0) {
455                 printk("CPU[%d]: mondo stuckage result[%016llx]\n",
456                        smp_processor_id(), result);
457         } else {
458                 udelay(2);
459                 goto again;
460         }
461 }
462
463 static void spitfire_xcall_deliver(struct trap_per_cpu *tb, int cnt)
464 {
465         u64 *mondo, data0, data1, data2;
466         u16 *cpu_list;
467         u64 pstate;
468         int i;
469
470         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
471         cpu_list = __va(tb->cpu_list_pa);
472         mondo = __va(tb->cpu_mondo_block_pa);
473         data0 = mondo[0];
474         data1 = mondo[1];
475         data2 = mondo[2];
476         for (i = 0; i < cnt; i++)
477                 spitfire_xcall_helper(data0, data1, data2, pstate, cpu_list[i]);
478 }
479
480 /* Cheetah now allows to send the whole 64-bytes of data in the interrupt
481  * packet, but we have no use for that.  However we do take advantage of
482  * the new pipelining feature (ie. dispatch to multiple cpus simultaneously).
483  */
484 static void cheetah_xcall_deliver(struct trap_per_cpu *tb, int cnt)
485 {
486         int nack_busy_id, is_jbus, need_more;
487         u64 *mondo, pstate, ver, busy_mask;
488         u16 *cpu_list;
489
490         cpu_list = __va(tb->cpu_list_pa);
491         mondo = __va(tb->cpu_mondo_block_pa);
492
493         /* Unfortunately, someone at Sun had the brilliant idea to make the
494          * busy/nack fields hard-coded by ITID number for this Ultra-III
495          * derivative processor.
496          */
497         __asm__ ("rdpr %%ver, %0" : "=r" (ver));
498         is_jbus = ((ver >> 32) == __JALAPENO_ID ||
499                    (ver >> 32) == __SERRANO_ID);
500
501         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
502
503 retry:
504         need_more = 0;
505         __asm__ __volatile__("wrpr %0, %1, %%pstate\n\t"
506                              : : "r" (pstate), "i" (PSTATE_IE));
507
508         /* Setup the dispatch data registers. */
509         __asm__ __volatile__("stxa      %0, [%3] %6\n\t"
510                              "stxa      %1, [%4] %6\n\t"
511                              "stxa      %2, [%5] %6\n\t"
512                              "membar    #Sync\n\t"
513                              : /* no outputs */
514                              : "r" (mondo[0]), "r" (mondo[1]), "r" (mondo[2]),
515                                "r" (0x40), "r" (0x50), "r" (0x60),
516                                "i" (ASI_INTR_W));
517
518         nack_busy_id = 0;
519         busy_mask = 0;
520         {
521                 int i;
522
523                 for (i = 0; i < cnt; i++) {
524                         u64 target, nr;
525
526                         nr = cpu_list[i];
527                         if (nr == 0xffff)
528                                 continue;
529
530                         target = (nr << 14) | 0x70;
531                         if (is_jbus) {
532                                 busy_mask |= (0x1UL << (nr * 2));
533                         } else {
534                                 target |= (nack_busy_id << 24);
535                                 busy_mask |= (0x1UL <<
536                                               (nack_busy_id * 2));
537                         }
538                         __asm__ __volatile__(
539                                 "stxa   %%g0, [%0] %1\n\t"
540                                 "membar #Sync\n\t"
541                                 : /* no outputs */
542                                 : "r" (target), "i" (ASI_INTR_W));
543                         nack_busy_id++;
544                         if (nack_busy_id == 32) {
545                                 need_more = 1;
546                                 break;
547                         }
548                 }
549         }
550
551         /* Now, poll for completion. */
552         {
553                 u64 dispatch_stat, nack_mask;
554                 long stuck;
555
556                 stuck = 100000 * nack_busy_id;
557                 nack_mask = busy_mask << 1;
558                 do {
559                         __asm__ __volatile__("ldxa      [%%g0] %1, %0"
560                                              : "=r" (dispatch_stat)
561                                              : "i" (ASI_INTR_DISPATCH_STAT));
562                         if (!(dispatch_stat & (busy_mask | nack_mask))) {
563                                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
564                                                      : : "r" (pstate));
565                                 if (unlikely(need_more)) {
566                                         int i, this_cnt = 0;
567                                         for (i = 0; i < cnt; i++) {
568                                                 if (cpu_list[i] == 0xffff)
569                                                         continue;
570                                                 cpu_list[i] = 0xffff;
571                                                 this_cnt++;
572                                                 if (this_cnt == 32)
573                                                         break;
574                                         }
575                                         goto retry;
576                                 }
577                                 return;
578                         }
579                         if (!--stuck)
580                                 break;
581                 } while (dispatch_stat & busy_mask);
582
583                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
584                                      : : "r" (pstate));
585
586                 if (dispatch_stat & busy_mask) {
587                         /* Busy bits will not clear, continue instead
588                          * of freezing up on this cpu.
589                          */
590                         printk("CPU[%d]: mondo stuckage result[%016llx]\n",
591                                smp_processor_id(), dispatch_stat);
592                 } else {
593                         int i, this_busy_nack = 0;
594
595                         /* Delay some random time with interrupts enabled
596                          * to prevent deadlock.
597                          */
598                         udelay(2 * nack_busy_id);
599
600                         /* Clear out the mask bits for cpus which did not
601                          * NACK us.
602                          */
603                         for (i = 0; i < cnt; i++) {
604                                 u64 check_mask, nr;
605
606                                 nr = cpu_list[i];
607                                 if (nr == 0xffff)
608                                         continue;
609
610                                 if (is_jbus)
611                                         check_mask = (0x2UL << (2*nr));
612                                 else
613                                         check_mask = (0x2UL <<
614                                                       this_busy_nack);
615                                 if ((dispatch_stat & check_mask) == 0)
616                                         cpu_list[i] = 0xffff;
617                                 this_busy_nack += 2;
618                                 if (this_busy_nack == 64)
619                                         break;
620                         }
621
622                         goto retry;
623                 }
624         }
625 }
626
627 /* Multi-cpu list version.  */
628 static void hypervisor_xcall_deliver(struct trap_per_cpu *tb, int cnt)
629 {
630         int retries, this_cpu, prev_sent, i, saw_cpu_error;
631         unsigned long status;
632         u16 *cpu_list;
633
634         this_cpu = smp_processor_id();
635
636         cpu_list = __va(tb->cpu_list_pa);
637
638         saw_cpu_error = 0;
639         retries = 0;
640         prev_sent = 0;
641         do {
642                 int forward_progress, n_sent;
643
644                 status = sun4v_cpu_mondo_send(cnt,
645                                               tb->cpu_list_pa,
646                                               tb->cpu_mondo_block_pa);
647
648                 /* HV_EOK means all cpus received the xcall, we're done.  */
649                 if (likely(status == HV_EOK))
650                         break;
651
652                 /* First, see if we made any forward progress.
653                  *
654                  * The hypervisor indicates successful sends by setting
655                  * cpu list entries to the value 0xffff.
656                  */
657                 n_sent = 0;
658                 for (i = 0; i < cnt; i++) {
659                         if (likely(cpu_list[i] == 0xffff))
660                                 n_sent++;
661                 }
662
663                 forward_progress = 0;
664                 if (n_sent > prev_sent)
665                         forward_progress = 1;
666
667                 prev_sent = n_sent;
668
669                 /* If we get a HV_ECPUERROR, then one or more of the cpus
670                  * in the list are in error state.  Use the cpu_state()
671                  * hypervisor call to find out which cpus are in error state.
672                  */
673                 if (unlikely(status == HV_ECPUERROR)) {
674                         for (i = 0; i < cnt; i++) {
675                                 long err;
676                                 u16 cpu;
677
678                                 cpu = cpu_list[i];
679                                 if (cpu == 0xffff)
680                                         continue;
681
682                                 err = sun4v_cpu_state(cpu);
683                                 if (err == HV_CPU_STATE_ERROR) {
684                                         saw_cpu_error = (cpu + 1);
685                                         cpu_list[i] = 0xffff;
686                                 }
687                         }
688                 } else if (unlikely(status != HV_EWOULDBLOCK))
689                         goto fatal_mondo_error;
690
691                 /* Don't bother rewriting the CPU list, just leave the
692                  * 0xffff and non-0xffff entries in there and the
693                  * hypervisor will do the right thing.
694                  *
695                  * Only advance timeout state if we didn't make any
696                  * forward progress.
697                  */
698                 if (unlikely(!forward_progress)) {
699                         if (unlikely(++retries > 10000))
700                                 goto fatal_mondo_timeout;
701
702                         /* Delay a little bit to let other cpus catch up
703                          * on their cpu mondo queue work.
704                          */
705                         udelay(2 * cnt);
706                 }
707         } while (1);
708
709         if (unlikely(saw_cpu_error))
710                 goto fatal_mondo_cpu_error;
711
712         return;
713
714 fatal_mondo_cpu_error:
715         printk(KERN_CRIT "CPU[%d]: SUN4V mondo cpu error, some target cpus "
716                "(including %d) were in error state\n",
717                this_cpu, saw_cpu_error - 1);
718         return;
719
720 fatal_mondo_timeout:
721         printk(KERN_CRIT "CPU[%d]: SUN4V mondo timeout, no forward "
722                " progress after %d retries.\n",
723                this_cpu, retries);
724         goto dump_cpu_list_and_out;
725
726 fatal_mondo_error:
727         printk(KERN_CRIT "CPU[%d]: Unexpected SUN4V mondo error %lu\n",
728                this_cpu, status);
729         printk(KERN_CRIT "CPU[%d]: Args were cnt(%d) cpulist_pa(%lx) "
730                "mondo_block_pa(%lx)\n",
731                this_cpu, cnt, tb->cpu_list_pa, tb->cpu_mondo_block_pa);
732
733 dump_cpu_list_and_out:
734         printk(KERN_CRIT "CPU[%d]: CPU list [ ", this_cpu);
735         for (i = 0; i < cnt; i++)
736                 printk("%u ", cpu_list[i]);
737         printk("]\n");
738 }
739
740 static void (*xcall_deliver_impl)(struct trap_per_cpu *, int);
741
742 static void xcall_deliver(u64 data0, u64 data1, u64 data2, const cpumask_t *mask)
743 {
744         struct trap_per_cpu *tb;
745         int this_cpu, i, cnt;
746         unsigned long flags;
747         u16 *cpu_list;
748         u64 *mondo;
749
750         /* We have to do this whole thing with interrupts fully disabled.
751          * Otherwise if we send an xcall from interrupt context it will
752          * corrupt both our mondo block and cpu list state.
753          *
754          * One consequence of this is that we cannot use timeout mechanisms
755          * that depend upon interrupts being delivered locally.  So, for
756          * example, we cannot sample jiffies and expect it to advance.
757          *
758          * Fortunately, udelay() uses %stick/%tick so we can use that.
759          */
760         local_irq_save(flags);
761
762         this_cpu = smp_processor_id();
763         tb = &trap_block[this_cpu];
764
765         mondo = __va(tb->cpu_mondo_block_pa);
766         mondo[0] = data0;
767         mondo[1] = data1;
768         mondo[2] = data2;
769         wmb();
770
771         cpu_list = __va(tb->cpu_list_pa);
772
773         /* Setup the initial cpu list.  */
774         cnt = 0;
775         for_each_cpu(i, mask) {
776                 if (i == this_cpu || !cpu_online(i))
777                         continue;
778                 cpu_list[cnt++] = i;
779         }
780
781         if (cnt)
782                 xcall_deliver_impl(tb, cnt);
783
784         local_irq_restore(flags);
785 }
786
787 /* Send cross call to all processors mentioned in MASK_P
788  * except self.  Really, there are only two cases currently,
789  * "&cpu_online_map" and "&mm->cpu_vm_mask".
790  */
791 static void smp_cross_call_masked(unsigned long *func, u32 ctx, u64 data1, u64 data2, const cpumask_t *mask)
792 {
793         u64 data0 = (((u64)ctx)<<32 | (((u64)func) & 0xffffffff));
794
795         xcall_deliver(data0, data1, data2, mask);
796 }
797
798 /* Send cross call to all processors except self. */
799 static void smp_cross_call(unsigned long *func, u32 ctx, u64 data1, u64 data2)
800 {
801         smp_cross_call_masked(func, ctx, data1, data2, &cpu_online_map);
802 }
803
804 extern unsigned long xcall_sync_tick;
805
806 static void smp_start_sync_tick_client(int cpu)
807 {
808         xcall_deliver((u64) &xcall_sync_tick, 0, 0,
809                       &cpumask_of_cpu(cpu));
810 }
811
812 extern unsigned long xcall_call_function;
813
814 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
815 {
816         xcall_deliver((u64) &xcall_call_function, 0, 0, mask);
817 }
818
819 extern unsigned long xcall_call_function_single;
820
821 void arch_send_call_function_single_ipi(int cpu)
822 {
823         xcall_deliver((u64) &xcall_call_function_single, 0, 0,
824                       &cpumask_of_cpu(cpu));
825 }
826
827 void __irq_entry smp_call_function_client(int irq, struct pt_regs *regs)
828 {
829         clear_softint(1 << irq);
830         generic_smp_call_function_interrupt();
831 }
832
833 void __irq_entry smp_call_function_single_client(int irq, struct pt_regs *regs)
834 {
835         clear_softint(1 << irq);
836         generic_smp_call_function_single_interrupt();
837 }
838
839 static void tsb_sync(void *info)
840 {
841         struct trap_per_cpu *tp = &trap_block[raw_smp_processor_id()];
842         struct mm_struct *mm = info;
843
844         /* It is not valid to test "currrent->active_mm == mm" here.
845          *
846          * The value of "current" is not changed atomically with
847          * switch_mm().  But that's OK, we just need to check the
848          * current cpu's trap block PGD physical address.
849          */
850         if (tp->pgd_paddr == __pa(mm->pgd))
851                 tsb_context_switch(mm);
852 }
853
854 void smp_tsb_sync(struct mm_struct *mm)
855 {
856         smp_call_function_many(mm_cpumask(mm), tsb_sync, mm, 1);
857 }
858
859 extern unsigned long xcall_flush_tlb_mm;
860 extern unsigned long xcall_flush_tlb_pending;
861 extern unsigned long xcall_flush_tlb_kernel_range;
862 extern unsigned long xcall_fetch_glob_regs;
863 extern unsigned long xcall_receive_signal;
864 extern unsigned long xcall_new_mmu_context_version;
865 #ifdef CONFIG_KGDB
866 extern unsigned long xcall_kgdb_capture;
867 #endif
868
869 #ifdef DCACHE_ALIASING_POSSIBLE
870 extern unsigned long xcall_flush_dcache_page_cheetah;
871 #endif
872 extern unsigned long xcall_flush_dcache_page_spitfire;
873
874 #ifdef CONFIG_DEBUG_DCFLUSH
875 extern atomic_t dcpage_flushes;
876 extern atomic_t dcpage_flushes_xcall;
877 #endif
878
879 static inline void __local_flush_dcache_page(struct page *page)
880 {
881 #ifdef DCACHE_ALIASING_POSSIBLE
882         __flush_dcache_page(page_address(page),
883                             ((tlb_type == spitfire) &&
884                              page_mapping(page) != NULL));
885 #else
886         if (page_mapping(page) != NULL &&
887             tlb_type == spitfire)
888                 __flush_icache_page(__pa(page_address(page)));
889 #endif
890 }
891
892 void smp_flush_dcache_page_impl(struct page *page, int cpu)
893 {
894         int this_cpu;
895
896         if (tlb_type == hypervisor)
897                 return;
898
899 #ifdef CONFIG_DEBUG_DCFLUSH
900         atomic_inc(&dcpage_flushes);
901 #endif
902
903         this_cpu = get_cpu();
904
905         if (cpu == this_cpu) {
906                 __local_flush_dcache_page(page);
907         } else if (cpu_online(cpu)) {
908                 void *pg_addr = page_address(page);
909                 u64 data0 = 0;
910
911                 if (tlb_type == spitfire) {
912                         data0 = ((u64)&xcall_flush_dcache_page_spitfire);
913                         if (page_mapping(page) != NULL)
914                                 data0 |= ((u64)1 << 32);
915                 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
916 #ifdef DCACHE_ALIASING_POSSIBLE
917                         data0 = ((u64)&xcall_flush_dcache_page_cheetah);
918 #endif
919                 }
920                 if (data0) {
921                         xcall_deliver(data0, __pa(pg_addr),
922                                       (u64) pg_addr, &cpumask_of_cpu(cpu));
923 #ifdef CONFIG_DEBUG_DCFLUSH
924                         atomic_inc(&dcpage_flushes_xcall);
925 #endif
926                 }
927         }
928
929         put_cpu();
930 }
931
932 void flush_dcache_page_all(struct mm_struct *mm, struct page *page)
933 {
934         void *pg_addr;
935         int this_cpu;
936         u64 data0;
937
938         if (tlb_type == hypervisor)
939                 return;
940
941         this_cpu = get_cpu();
942
943 #ifdef CONFIG_DEBUG_DCFLUSH
944         atomic_inc(&dcpage_flushes);
945 #endif
946         data0 = 0;
947         pg_addr = page_address(page);
948         if (tlb_type == spitfire) {
949                 data0 = ((u64)&xcall_flush_dcache_page_spitfire);
950                 if (page_mapping(page) != NULL)
951                         data0 |= ((u64)1 << 32);
952         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
953 #ifdef DCACHE_ALIASING_POSSIBLE
954                 data0 = ((u64)&xcall_flush_dcache_page_cheetah);
955 #endif
956         }
957         if (data0) {
958                 xcall_deliver(data0, __pa(pg_addr),
959                               (u64) pg_addr, &cpu_online_map);
960 #ifdef CONFIG_DEBUG_DCFLUSH
961                 atomic_inc(&dcpage_flushes_xcall);
962 #endif
963         }
964         __local_flush_dcache_page(page);
965
966         put_cpu();
967 }
968
969 void __irq_entry smp_new_mmu_context_version_client(int irq, struct pt_regs *regs)
970 {
971         struct mm_struct *mm;
972         unsigned long flags;
973
974         clear_softint(1 << irq);
975
976         /* See if we need to allocate a new TLB context because
977          * the version of the one we are using is now out of date.
978          */
979         mm = current->active_mm;
980         if (unlikely(!mm || (mm == &init_mm)))
981                 return;
982
983         spin_lock_irqsave(&mm->context.lock, flags);
984
985         if (unlikely(!CTX_VALID(mm->context)))
986                 get_new_mmu_context(mm);
987
988         spin_unlock_irqrestore(&mm->context.lock, flags);
989
990         load_secondary_context(mm);
991         __flush_tlb_mm(CTX_HWBITS(mm->context),
992                        SECONDARY_CONTEXT);
993 }
994
995 void smp_new_mmu_context_version(void)
996 {
997         smp_cross_call(&xcall_new_mmu_context_version, 0, 0, 0);
998 }
999
1000 #ifdef CONFIG_KGDB
1001 void kgdb_roundup_cpus(unsigned long flags)
1002 {
1003         smp_cross_call(&xcall_kgdb_capture, 0, 0, 0);
1004 }
1005 #endif
1006
1007 void smp_fetch_global_regs(void)
1008 {
1009         smp_cross_call(&xcall_fetch_glob_regs, 0, 0, 0);
1010 }
1011
1012 /* We know that the window frames of the user have been flushed
1013  * to the stack before we get here because all callers of us
1014  * are flush_tlb_*() routines, and these run after flush_cache_*()
1015  * which performs the flushw.
1016  *
1017  * The SMP TLB coherency scheme we use works as follows:
1018  *
1019  * 1) mm->cpu_vm_mask is a bit mask of which cpus an address
1020  *    space has (potentially) executed on, this is the heuristic
1021  *    we use to avoid doing cross calls.
1022  *
1023  *    Also, for flushing from kswapd and also for clones, we
1024  *    use cpu_vm_mask as the list of cpus to make run the TLB.
1025  *
1026  * 2) TLB context numbers are shared globally across all processors
1027  *    in the system, this allows us to play several games to avoid
1028  *    cross calls.
1029  *
1030  *    One invariant is that when a cpu switches to a process, and
1031  *    that processes tsk->active_mm->cpu_vm_mask does not have the
1032  *    current cpu's bit set, that tlb context is flushed locally.
1033  *
1034  *    If the address space is non-shared (ie. mm->count == 1) we avoid
1035  *    cross calls when we want to flush the currently running process's
1036  *    tlb state.  This is done by clearing all cpu bits except the current
1037  *    processor's in current->mm->cpu_vm_mask and performing the
1038  *    flush locally only.  This will force any subsequent cpus which run
1039  *    this task to flush the context from the local tlb if the process
1040  *    migrates to another cpu (again).
1041  *
1042  * 3) For shared address spaces (threads) and swapping we bite the
1043  *    bullet for most cases and perform the cross call (but only to
1044  *    the cpus listed in cpu_vm_mask).
1045  *
1046  *    The performance gain from "optimizing" away the cross call for threads is
1047  *    questionable (in theory the big win for threads is the massive sharing of
1048  *    address space state across processors).
1049  */
1050
1051 /* This currently is only used by the hugetlb arch pre-fault
1052  * hook on UltraSPARC-III+ and later when changing the pagesize
1053  * bits of the context register for an address space.
1054  */
1055 void smp_flush_tlb_mm(struct mm_struct *mm)
1056 {
1057         u32 ctx = CTX_HWBITS(mm->context);
1058         int cpu = get_cpu();
1059
1060         if (atomic_read(&mm->mm_users) == 1) {
1061                 cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
1062                 goto local_flush_and_out;
1063         }
1064
1065         smp_cross_call_masked(&xcall_flush_tlb_mm,
1066                               ctx, 0, 0,
1067                               mm_cpumask(mm));
1068
1069 local_flush_and_out:
1070         __flush_tlb_mm(ctx, SECONDARY_CONTEXT);
1071
1072         put_cpu();
1073 }
1074
1075 void smp_flush_tlb_pending(struct mm_struct *mm, unsigned long nr, unsigned long *vaddrs)
1076 {
1077         u32 ctx = CTX_HWBITS(mm->context);
1078         int cpu = get_cpu();
1079
1080         if (mm == current->mm && atomic_read(&mm->mm_users) == 1)
1081                 cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
1082         else
1083                 smp_cross_call_masked(&xcall_flush_tlb_pending,
1084                                       ctx, nr, (unsigned long) vaddrs,
1085                                       mm_cpumask(mm));
1086
1087         __flush_tlb_pending(ctx, nr, vaddrs);
1088
1089         put_cpu();
1090 }
1091
1092 void smp_flush_tlb_kernel_range(unsigned long start, unsigned long end)
1093 {
1094         start &= PAGE_MASK;
1095         end    = PAGE_ALIGN(end);
1096         if (start != end) {
1097                 smp_cross_call(&xcall_flush_tlb_kernel_range,
1098                                0, start, end);
1099
1100                 __flush_tlb_kernel_range(start, end);
1101         }
1102 }
1103
1104 /* CPU capture. */
1105 /* #define CAPTURE_DEBUG */
1106 extern unsigned long xcall_capture;
1107
1108 static atomic_t smp_capture_depth = ATOMIC_INIT(0);
1109 static atomic_t smp_capture_registry = ATOMIC_INIT(0);
1110 static unsigned long penguins_are_doing_time;
1111
1112 void smp_capture(void)
1113 {
1114         int result = atomic_add_ret(1, &smp_capture_depth);
1115
1116         if (result == 1) {
1117                 int ncpus = num_online_cpus();
1118
1119 #ifdef CAPTURE_DEBUG
1120                 printk("CPU[%d]: Sending penguins to jail...",
1121                        smp_processor_id());
1122 #endif
1123                 penguins_are_doing_time = 1;
1124                 atomic_inc(&smp_capture_registry);
1125                 smp_cross_call(&xcall_capture, 0, 0, 0);
1126                 while (atomic_read(&smp_capture_registry) != ncpus)
1127                         rmb();
1128 #ifdef CAPTURE_DEBUG
1129                 printk("done\n");
1130 #endif
1131         }
1132 }
1133
1134 void smp_release(void)
1135 {
1136         if (atomic_dec_and_test(&smp_capture_depth)) {
1137 #ifdef CAPTURE_DEBUG
1138                 printk("CPU[%d]: Giving pardon to "
1139                        "imprisoned penguins\n",
1140                        smp_processor_id());
1141 #endif
1142                 penguins_are_doing_time = 0;
1143                 membar_safe("#StoreLoad");
1144                 atomic_dec(&smp_capture_registry);
1145         }
1146 }
1147
1148 /* Imprisoned penguins run with %pil == PIL_NORMAL_MAX, but PSTATE_IE
1149  * set, so they can service tlb flush xcalls...
1150  */
1151 extern void prom_world(int);
1152
1153 void __irq_entry smp_penguin_jailcell(int irq, struct pt_regs *regs)
1154 {
1155         clear_softint(1 << irq);
1156
1157         preempt_disable();
1158
1159         __asm__ __volatile__("flushw");
1160         prom_world(1);
1161         atomic_inc(&smp_capture_registry);
1162         membar_safe("#StoreLoad");
1163         while (penguins_are_doing_time)
1164                 rmb();
1165         atomic_dec(&smp_capture_registry);
1166         prom_world(0);
1167
1168         preempt_enable();
1169 }
1170
1171 /* /proc/profile writes can call this, don't __init it please. */
1172 int setup_profiling_timer(unsigned int multiplier)
1173 {
1174         return -EINVAL;
1175 }
1176
1177 void __init smp_prepare_cpus(unsigned int max_cpus)
1178 {
1179 }
1180
1181 void __devinit smp_prepare_boot_cpu(void)
1182 {
1183 }
1184
1185 void __init smp_setup_processor_id(void)
1186 {
1187         if (tlb_type == spitfire)
1188                 xcall_deliver_impl = spitfire_xcall_deliver;
1189         else if (tlb_type == cheetah || tlb_type == cheetah_plus)
1190                 xcall_deliver_impl = cheetah_xcall_deliver;
1191         else
1192                 xcall_deliver_impl = hypervisor_xcall_deliver;
1193 }
1194
1195 void __devinit smp_fill_in_sib_core_maps(void)
1196 {
1197         unsigned int i;
1198
1199         for_each_present_cpu(i) {
1200                 unsigned int j;
1201
1202                 cpus_clear(cpu_core_map[i]);
1203                 if (cpu_data(i).core_id == 0) {
1204                         cpu_set(i, cpu_core_map[i]);
1205                         continue;
1206                 }
1207
1208                 for_each_present_cpu(j) {
1209                         if (cpu_data(i).core_id ==
1210                             cpu_data(j).core_id)
1211                                 cpu_set(j, cpu_core_map[i]);
1212                 }
1213         }
1214
1215         for_each_present_cpu(i) {
1216                 unsigned int j;
1217
1218                 cpus_clear(per_cpu(cpu_sibling_map, i));
1219                 if (cpu_data(i).proc_id == -1) {
1220                         cpu_set(i, per_cpu(cpu_sibling_map, i));
1221                         continue;
1222                 }
1223
1224                 for_each_present_cpu(j) {
1225                         if (cpu_data(i).proc_id ==
1226                             cpu_data(j).proc_id)
1227                                 cpu_set(j, per_cpu(cpu_sibling_map, i));
1228                 }
1229         }
1230 }
1231
1232 int __cpuinit __cpu_up(unsigned int cpu)
1233 {
1234         int ret = smp_boot_one_cpu(cpu);
1235
1236         if (!ret) {
1237                 cpu_set(cpu, smp_commenced_mask);
1238                 while (!cpu_isset(cpu, cpu_online_map))
1239                         mb();
1240                 if (!cpu_isset(cpu, cpu_online_map)) {
1241                         ret = -ENODEV;
1242                 } else {
1243                         /* On SUN4V, writes to %tick and %stick are
1244                          * not allowed.
1245                          */
1246                         if (tlb_type != hypervisor)
1247                                 smp_synchronize_one_tick(cpu);
1248                 }
1249         }
1250         return ret;
1251 }
1252
1253 #ifdef CONFIG_HOTPLUG_CPU
1254 void cpu_play_dead(void)
1255 {
1256         int cpu = smp_processor_id();
1257         unsigned long pstate;
1258
1259         idle_task_exit();
1260
1261         if (tlb_type == hypervisor) {
1262                 struct trap_per_cpu *tb = &trap_block[cpu];
1263
1264                 sun4v_cpu_qconf(HV_CPU_QUEUE_CPU_MONDO,
1265                                 tb->cpu_mondo_pa, 0);
1266                 sun4v_cpu_qconf(HV_CPU_QUEUE_DEVICE_MONDO,
1267                                 tb->dev_mondo_pa, 0);
1268                 sun4v_cpu_qconf(HV_CPU_QUEUE_RES_ERROR,
1269                                 tb->resum_mondo_pa, 0);
1270                 sun4v_cpu_qconf(HV_CPU_QUEUE_NONRES_ERROR,
1271                                 tb->nonresum_mondo_pa, 0);
1272         }
1273
1274         cpu_clear(cpu, smp_commenced_mask);
1275         membar_safe("#Sync");
1276
1277         local_irq_disable();
1278
1279         __asm__ __volatile__(
1280                 "rdpr   %%pstate, %0\n\t"
1281                 "wrpr   %0, %1, %%pstate"
1282                 : "=r" (pstate)
1283                 : "i" (PSTATE_IE));
1284
1285         while (1)
1286                 barrier();
1287 }
1288
1289 int __cpu_disable(void)
1290 {
1291         int cpu = smp_processor_id();
1292         cpuinfo_sparc *c;
1293         int i;
1294
1295         for_each_cpu_mask(i, cpu_core_map[cpu])
1296                 cpu_clear(cpu, cpu_core_map[i]);
1297         cpus_clear(cpu_core_map[cpu]);
1298
1299         for_each_cpu_mask(i, per_cpu(cpu_sibling_map, cpu))
1300                 cpu_clear(cpu, per_cpu(cpu_sibling_map, i));
1301         cpus_clear(per_cpu(cpu_sibling_map, cpu));
1302
1303         c = &cpu_data(cpu);
1304
1305         c->core_id = 0;
1306         c->proc_id = -1;
1307
1308         smp_wmb();
1309
1310         /* Make sure no interrupts point to this cpu.  */
1311         fixup_irqs();
1312
1313         local_irq_enable();
1314         mdelay(1);
1315         local_irq_disable();
1316
1317         ipi_call_lock();
1318         cpu_clear(cpu, cpu_online_map);
1319         ipi_call_unlock();
1320
1321         cpu_map_rebuild();
1322
1323         return 0;
1324 }
1325
1326 void __cpu_die(unsigned int cpu)
1327 {
1328         int i;
1329
1330         for (i = 0; i < 100; i++) {
1331                 smp_rmb();
1332                 if (!cpu_isset(cpu, smp_commenced_mask))
1333                         break;
1334                 msleep(100);
1335         }
1336         if (cpu_isset(cpu, smp_commenced_mask)) {
1337                 printk(KERN_ERR "CPU %u didn't die...\n", cpu);
1338         } else {
1339 #if defined(CONFIG_SUN_LDOMS)
1340                 unsigned long hv_err;
1341                 int limit = 100;
1342
1343                 do {
1344                         hv_err = sun4v_cpu_stop(cpu);
1345                         if (hv_err == HV_EOK) {
1346                                 cpu_clear(cpu, cpu_present_map);
1347                                 break;
1348                         }
1349                 } while (--limit > 0);
1350                 if (limit <= 0) {
1351                         printk(KERN_ERR "sun4v_cpu_stop() fails err=%lu\n",
1352                                hv_err);
1353                 }
1354 #endif
1355         }
1356 }
1357 #endif
1358
1359 void __init smp_cpus_done(unsigned int max_cpus)
1360 {
1361 }
1362
1363 void smp_send_reschedule(int cpu)
1364 {
1365         xcall_deliver((u64) &xcall_receive_signal, 0, 0,
1366                       &cpumask_of_cpu(cpu));
1367 }
1368
1369 void __irq_entry smp_receive_signal_client(int irq, struct pt_regs *regs)
1370 {
1371         clear_softint(1 << irq);
1372 }
1373
1374 /* This is a nop because we capture all other cpus
1375  * anyways when making the PROM active.
1376  */
1377 void smp_send_stop(void)
1378 {
1379 }
1380
1381 /**
1382  * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu
1383  * @cpu: cpu to allocate for
1384  * @size: size allocation in bytes
1385  * @align: alignment
1386  *
1387  * Allocate @size bytes aligned at @align for cpu @cpu.  This wrapper
1388  * does the right thing for NUMA regardless of the current
1389  * configuration.
1390  *
1391  * RETURNS:
1392  * Pointer to the allocated area on success, NULL on failure.
1393  */
1394 static void * __init pcpu_alloc_bootmem(unsigned int cpu, size_t size,
1395                                         size_t align)
1396 {
1397         const unsigned long goal = __pa(MAX_DMA_ADDRESS);
1398 #ifdef CONFIG_NEED_MULTIPLE_NODES
1399         int node = cpu_to_node(cpu);
1400         void *ptr;
1401
1402         if (!node_online(node) || !NODE_DATA(node)) {
1403                 ptr = __alloc_bootmem(size, align, goal);
1404                 pr_info("cpu %d has no node %d or node-local memory\n",
1405                         cpu, node);
1406                 pr_debug("per cpu data for cpu%d %lu bytes at %016lx\n",
1407                          cpu, size, __pa(ptr));
1408         } else {
1409                 ptr = __alloc_bootmem_node(NODE_DATA(node),
1410                                            size, align, goal);
1411                 pr_debug("per cpu data for cpu%d %lu bytes on node%d at "
1412                          "%016lx\n", cpu, size, node, __pa(ptr));
1413         }
1414         return ptr;
1415 #else
1416         return __alloc_bootmem(size, align, goal);
1417 #endif
1418 }
1419
1420 static void __init pcpu_free_bootmem(void *ptr, size_t size)
1421 {
1422         free_bootmem(__pa(ptr), size);
1423 }
1424
1425 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
1426 {
1427         if (cpu_to_node(from) == cpu_to_node(to))
1428                 return LOCAL_DISTANCE;
1429         else
1430                 return REMOTE_DISTANCE;
1431 }
1432
1433 static void __init pcpu_populate_pte(unsigned long addr)
1434 {
1435         pgd_t *pgd = pgd_offset_k(addr);
1436         pud_t *pud;
1437         pmd_t *pmd;
1438
1439         pud = pud_offset(pgd, addr);
1440         if (pud_none(*pud)) {
1441                 pmd_t *new;
1442
1443                 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1444                 pud_populate(&init_mm, pud, new);
1445         }
1446
1447         pmd = pmd_offset(pud, addr);
1448         if (!pmd_present(*pmd)) {
1449                 pte_t *new;
1450
1451                 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1452                 pmd_populate_kernel(&init_mm, pmd, new);
1453         }
1454 }
1455
1456 void __init setup_per_cpu_areas(void)
1457 {
1458         unsigned long delta;
1459         unsigned int cpu;
1460         int rc = -EINVAL;
1461
1462         if (pcpu_chosen_fc != PCPU_FC_PAGE) {
1463                 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
1464                                             PERCPU_DYNAMIC_RESERVE, 4 << 20,
1465                                             pcpu_cpu_distance,
1466                                             pcpu_alloc_bootmem,
1467                                             pcpu_free_bootmem);
1468                 if (rc)
1469                         pr_warning("PERCPU: %s allocator failed (%d), "
1470                                    "falling back to page size\n",
1471                                    pcpu_fc_names[pcpu_chosen_fc], rc);
1472         }
1473         if (rc < 0)
1474                 rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE,
1475                                            pcpu_alloc_bootmem,
1476                                            pcpu_free_bootmem,
1477                                            pcpu_populate_pte);
1478         if (rc < 0)
1479                 panic("cannot initialize percpu area (err=%d)", rc);
1480
1481         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1482         for_each_possible_cpu(cpu)
1483                 __per_cpu_offset(cpu) = delta + pcpu_unit_offsets[cpu];
1484
1485         /* Setup %g5 for the boot cpu.  */
1486         __local_per_cpu_offset = __per_cpu_offset(smp_processor_id());
1487
1488         of_fill_in_cpu_data();
1489         if (tlb_type == hypervisor)
1490                 mdesc_fill_in_cpu_data(cpu_all_mask);
1491 }