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
  27 #include <asm/head.h>
  28 #include <asm/ptrace.h>
  29 #include <asm/atomic.h>
  30 #include <asm/tlbflush.h>
  31 #include <asm/mmu_context.h>
  32 #include <asm/cpudata.h>
  33 #include <asm/hvtramp.h>
  34 #include <asm/io.h>
  35 #include <asm/timer.h>
  36
  37 #include <asm/irq.h>
  38 #include <asm/irq_regs.h>
  39 #include <asm/page.h>
  40 #include <asm/pgtable.h>
  41 #include <asm/oplib.h>
  42 #include <asm/uaccess.h>
  43 #include <asm/starfire.h>
  44 #include <asm/tlb.h>
  45 #include <asm/sections.h>
  46 #include <asm/prom.h>
  47 #include <asm/mdesc.h>
  48 #include <asm/ldc.h>
  49 #include <asm/hypervisor.h>
  50
  51 #include "cpumap.h"
  52
  53 int sparc64_multi_core __read_mostly;
  54
  55 DEFINE_PER_CPU(cpumask_t, cpu_sibling_map) = CPU_MASK_NONE;
  56 cpumask_t cpu_core_map[NR_CPUS] __read_mostly =
  57         { [0 ... NR_CPUS-1] = CPU_MASK_NONE };
  58
  59 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
  60 EXPORT_SYMBOL(cpu_core_map);
  61
  62 static cpumask_t smp_commenced_mask;
  63
  64 void smp_info(struct seq_file *m)
  65 {
  66         int i;
  67
  68         seq_printf(m, "State:\n");
  69         for_each_online_cpu(i)
  70                 seq_printf(m, "CPU%d:\t\tonline\n", i);
  71 }
  72
  73 void smp_bogo(struct seq_file *m)
  74 {
  75         int i;
  76
  77         for_each_online_cpu(i)
  78                 seq_printf(m,
  79                            "Cpu%dClkTck\t: %016lx\n",
  80                            i, cpu_data(i).clock_tick);
  81 }
  82
  83 extern void setup_sparc64_timer(void);
  84
  85 static volatile unsigned long callin_flag = 0;
  86
  87 void __cpuinit smp_callin(void)
  88 {
  89         int cpuid = hard_smp_processor_id();
  90
  91         __local_per_cpu_offset = __per_cpu_offset(cpuid);
  92
  93         if (tlb_type == hypervisor)
  94                 sun4v_ktsb_register();
  95
  96         __flush_tlb_all();
  97
  98         setup_sparc64_timer();
  99
 100         if (cheetah_pcache_forced_on)
 101                 cheetah_enable_pcache();
 102
 103         local_irq_enable();
 104
 105         callin_flag = 1;
 106         __asm__ __volatile__("membar #Sync\n\t"
 107                              "flush  %%g6" : : : "memory");
 108
 109         /* Clear this or we will die instantly when we
 110          * schedule back to this idler...
 111          */
 112         current_thread_info()->new_child = 0;
 113
 114         /* Attach to the address space of init_task. */
 115         atomic_inc(&init_mm.mm_count);
 116         current->active_mm = &init_mm;
 117
 118         /* inform the notifiers about the new cpu */
 119         notify_cpu_starting(cpuid);
 120
 121         while (!cpu_isset(cpuid, smp_commenced_mask))
 122                 rmb();
 123
 124         ipi_call_lock_irq();
 125         cpu_set(cpuid, cpu_online_map);
 126         ipi_call_unlock_irq();
 127
 128         /* idle thread is expected to have preempt disabled */
 129         preempt_disable();
 130 }
 131
 132 void cpu_panic(void)
 133 {
 134         printk("CPU[%d]: Returns from cpu_idle!\n", smp_processor_id());
 135         panic("SMP bolixed\n");
 136 }
 137
 138 /* This tick register synchronization scheme is taken entirely from
 139  * the ia64 port, see arch/ia64/kernel/smpboot.c for details and credit.
 140  *
 141  * The only change I've made is to rework it so that the master
 142  * initiates the synchonization instead of the slave. -DaveM
 143  */
 144
 145 #define MASTER  0
 146 #define SLAVE   (SMP_CACHE_BYTES/sizeof(unsigned long))
 147
 148 #define NUM_ROUNDS      64      /* magic value */
 149 #define NUM_ITERS       5       /* likewise */
 150
 151 static DEFINE_SPINLOCK(itc_sync_lock);
 152 static unsigned long go[SLAVE + 1];
 153
 154 #define DEBUG_TICK_SYNC 0
 155
 156 static inline long get_delta (long *rt, long *master)
 157 {
 158         unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
 159         unsigned long tcenter, t0, t1, tm;
 160         unsigned long i;
 161
 162         for (i = 0; i < NUM_ITERS; i++) {
 163                 t0 = tick_ops->get_tick();
 164                 go[MASTER] = 1;
 165                 membar_safe("#StoreLoad");
 166                 while (!(tm = go[SLAVE]))
 167                         rmb();
 168                 go[SLAVE] = 0;
 169                 wmb();
 170                 t1 = tick_ops->get_tick();
 171
 172                 if (t1 - t0 < best_t1 - best_t0)
 173                         best_t0 = t0, best_t1 = t1, best_tm = tm;
 174         }
 175
 176         *rt = best_t1 - best_t0;
 177         *master = best_tm - best_t0;
 178
 179         /* average best_t0 and best_t1 without overflow: */
 180         tcenter = (best_t0/2 + best_t1/2);
 181         if (best_t0 % 2 + best_t1 % 2 == 2)
 182                 tcenter++;
 183         return tcenter - best_tm;
 184 }
 185
 186 void smp_synchronize_tick_client(void)
 187 {
 188         long i, delta, adj, adjust_latency = 0, done = 0;
 189         unsigned long flags, rt, master_time_stamp, bound;
 190 #if DEBUG_TICK_SYNC
 191         struct {
 192                 long rt;        /* roundtrip time */
 193                 long master;    /* master's timestamp */
 194                 long diff;      /* difference between midpoint and master's timestamp */
 195                 long lat;       /* estimate of itc adjustment latency */
 196         } t[NUM_ROUNDS];
 197 #endif
 198
 199         go[MASTER] = 1;
 200
 201         while (go[MASTER])
 202                 rmb();
 203
 204         local_irq_save(flags);
 205         {
 206                 for (i = 0; i < NUM_ROUNDS; i++) {
 207                         delta = get_delta(&rt, &master_time_stamp);
 208                         if (delta == 0) {
 209                                 done = 1;       /* let's lock on to this... */
 210                                 bound = rt;
 211                         }
 212
 213                         if (!done) {
 214                                 if (i > 0) {
 215                                         adjust_latency += -delta;
 216                                         adj = -delta + adjust_latency/4;
 217                                 } else
 218                                         adj = -delta;
 219
 220                                 tick_ops->add_tick(adj);
 221                         }
 222 #if DEBUG_TICK_SYNC
 223                         t[i].rt = rt;
 224                         t[i].master = master_time_stamp;
 225                         t[i].diff = delta;
 226                         t[i].lat = adjust_latency/4;
 227 #endif
 228                 }
 229         }
 230         local_irq_restore(flags);
 231
 232 #if DEBUG_TICK_SYNC
 233         for (i = 0; i < NUM_ROUNDS; i++)
 234                 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
 235                        t[i].rt, t[i].master, t[i].diff, t[i].lat);
 236 #endif
 237
 238         printk(KERN_INFO "CPU %d: synchronized TICK with master CPU "
 239                "(last diff %ld cycles, maxerr %lu cycles)\n",
 240                smp_processor_id(), delta, rt);
 241 }
 242
 243 static void smp_start_sync_tick_client(int cpu);
 244
 245 static void smp_synchronize_one_tick(int cpu)
 246 {
 247         unsigned long flags, i;
 248
 249         go[MASTER] = 0;
 250
 251         smp_start_sync_tick_client(cpu);
 252
 253         /* wait for client to be ready */
 254         while (!go[MASTER])
 255                 rmb();
 256
 257         /* now let the client proceed into his loop */
 258         go[MASTER] = 0;
 259         membar_safe("#StoreLoad");
 260
 261         spin_lock_irqsave(&itc_sync_lock, flags);
 262         {
 263                 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; i++) {
 264                         while (!go[MASTER])
 265                                 rmb();
 266                         go[MASTER] = 0;
 267                         wmb();
 268                         go[SLAVE] = tick_ops->get_tick();
 269                         membar_safe("#StoreLoad");
 270                 }
 271         }
 272         spin_unlock_irqrestore(&itc_sync_lock, flags);
 273 }
 274
 275 #if defined(CONFIG_SUN_LDOMS) && defined(CONFIG_HOTPLUG_CPU)
 276 /* XXX Put this in some common place. XXX */
 277 static unsigned long kimage_addr_to_ra(void *p)
 278 {
 279         unsigned long val = (unsigned long) p;
 280
 281         return kern_base + (val - KERNBASE);
 282 }
 283
 284 static void __cpuinit ldom_startcpu_cpuid(unsigned int cpu, unsigned long thread_reg, void **descrp)
 285 {
 286         extern unsigned long sparc64_ttable_tl0;
 287         extern unsigned long kern_locked_tte_data;
 288         struct hvtramp_descr *hdesc;
 289         unsigned long trampoline_ra;
 290         struct trap_per_cpu *tb;
 291         u64 tte_vaddr, tte_data;
 292         unsigned long hv_err;
 293         int i;
 294
 295         hdesc = kzalloc(sizeof(*hdesc) +
 296                         (sizeof(struct hvtramp_mapping) *
 297                          num_kernel_image_mappings - 1),
 298                         GFP_KERNEL);
 299         if (!hdesc) {
 300                 printk(KERN_ERR "ldom_startcpu_cpuid: Cannot allocate "
 301                        "hvtramp_descr.\n");
 302                 return;
 303         }
 304         *descrp = hdesc;
 305
 306         hdesc->cpu = cpu;
 307         hdesc->num_mappings = num_kernel_image_mappings;
 308
 309         tb = &trap_block[cpu];
 310
 311         hdesc->fault_info_va = (unsigned long) &tb->fault_info;
 312         hdesc->fault_info_pa = kimage_addr_to_ra(&tb->fault_info);
 313
 314         hdesc->thread_reg = thread_reg;
 315
 316         tte_vaddr = (unsigned long) KERNBASE;
 317         tte_data = kern_locked_tte_data;
 318
 319         for (i = 0; i < hdesc->num_mappings; i++) {
 320                 hdesc->maps[i].vaddr = tte_vaddr;
 321                 hdesc->maps[i].tte   = tte_data;
 322                 tte_vaddr += 0x400000;
 323                 tte_data  += 0x400000;
 324         }
 325
 326         trampoline_ra = kimage_addr_to_ra(hv_cpu_startup);
 327
 328         hv_err = sun4v_cpu_start(cpu, trampoline_ra,
 329                                  kimage_addr_to_ra(&sparc64_ttable_tl0),
 330                                  __pa(hdesc));
 331         if (hv_err)
 332                 printk(KERN_ERR "ldom_startcpu_cpuid: sun4v_cpu_start() "
 333                        "gives error %lu\n", hv_err);
 334 }
 335 #endif
 336
 337 extern unsigned long sparc64_cpu_startup;
 338
 339 /* The OBP cpu startup callback truncates the 3rd arg cookie to
 340  * 32-bits (I think) so to be safe we have it read the pointer
 341  * contained here so we work on >4GB machines. -DaveM
 342  */
 343 static struct thread_info *cpu_new_thread = NULL;
 344
 345 static int __cpuinit smp_boot_one_cpu(unsigned int cpu)
 346 {
 347         unsigned long entry =
 348                 (unsigned long)(&sparc64_cpu_startup);
 349         unsigned long cookie =
 350                 (unsigned long)(&cpu_new_thread);
 351         struct task_struct *p;
 352         void *descr = NULL;
 353         int timeout, ret;
 354
 355         p = fork_idle(cpu);
 356         if (IS_ERR(p))
 357                 return PTR_ERR(p);
 358         callin_flag = 0;
 359         cpu_new_thread = task_thread_info(p);
 360
 361         if (tlb_type == hypervisor) {
 362 #if defined(CONFIG_SUN_LDOMS) && defined(CONFIG_HOTPLUG_CPU)
 363                 if (ldom_domaining_enabled)
 364                         ldom_startcpu_cpuid(cpu,
 365                                             (unsigned long) cpu_new_thread,
 366                                             &descr);
 367                 else
 368 #endif
 369                         prom_startcpu_cpuid(cpu, entry, cookie);
 370         } else {
 371                 struct device_node *dp = of_find_node_by_cpuid(cpu);
 372
 373                 prom_startcpu(dp->node, entry, cookie);
 374         }
 375
 376         for (timeout = 0; timeout < 50000; timeout++) {
 377                 if (callin_flag)
 378                         break;
 379                 udelay(100);
 380         }
 381
 382         if (callin_flag) {
 383                 ret = 0;
 384         } else {
 385                 printk("Processor %d is stuck.\n", cpu);
 386                 ret = -ENODEV;
 387         }
 388         cpu_new_thread = NULL;
 389
 390         kfree(descr);
 391
 392         return ret;
 393 }
 394
 395 static void spitfire_xcall_helper(u64 data0, u64 data1, u64 data2, u64 pstate, unsigned long cpu)
 396 {
 397         u64 result, target;
 398         int stuck, tmp;
 399
 400         if (this_is_starfire) {
 401                 /* map to real upaid */
 402                 cpu = (((cpu & 0x3c) << 1) |
 403                         ((cpu & 0x40) >> 4) |
 404                         (cpu & 0x3));
 405         }
 406
 407         target = (cpu << 14) | 0x70;
 408 again:
 409         /* Ok, this is the real Spitfire Errata #54.
 410          * One must read back from a UDB internal register
 411          * after writes to the UDB interrupt dispatch, but
 412          * before the membar Sync for that write.
 413          * So we use the high UDB control register (ASI 0x7f,
 414          * ADDR 0x20) for the dummy read. -DaveM
 415          */
 416         tmp = 0x40;
 417         __asm__ __volatile__(
 418         "wrpr   %1, %2, %%pstate\n\t"
 419         "stxa   %4, [%0] %3\n\t"
 420         "stxa   %5, [%0+%8] %3\n\t"
 421         "add    %0, %8, %0\n\t"
 422         "stxa   %6, [%0+%8] %3\n\t"
 423         "membar #Sync\n\t"
 424         "stxa   %%g0, [%7] %3\n\t"
 425         "membar #Sync\n\t"
 426         "mov    0x20, %%g1\n\t"
 427         "ldxa   [%%g1] 0x7f, %%g0\n\t"
 428         "membar #Sync"
 429         : "=r" (tmp)
 430         : "r" (pstate), "i" (PSTATE_IE), "i" (ASI_INTR_W),
 431           "r" (data0), "r" (data1), "r" (data2), "r" (target),
 432           "r" (0x10), "0" (tmp)
 433         : "g1");
 434
 435         /* NOTE: PSTATE_IE is still clear. */
 436         stuck = 100000;
 437         do {
 438                 __asm__ __volatile__("ldxa [%%g0] %1, %0"
 439                         : "=r" (result)
 440                         : "i" (ASI_INTR_DISPATCH_STAT));
 441                 if (result == 0) {
 442                         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 443                                              : : "r" (pstate));
 444                         return;
 445                 }
 446                 stuck -= 1;
 447                 if (stuck == 0)
 448                         break;
 449         } while (result & 0x1);
 450         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 451                              : : "r" (pstate));
 452         if (stuck == 0) {
 453                 printk("CPU[%d]: mondo stuckage result[%016llx]\n",
 454                        smp_processor_id(), result);
 455         } else {
 456                 udelay(2);
 457                 goto again;
 458         }
 459 }
 460
 461 static void spitfire_xcall_deliver(struct trap_per_cpu *tb, int cnt)
 462 {
 463         u64 *mondo, data0, data1, data2;
 464         u16 *cpu_list;
 465         u64 pstate;
 466         int i;
 467
 468         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 469         cpu_list = __va(tb->cpu_list_pa);
 470         mondo = __va(tb->cpu_mondo_block_pa);
 471         data0 = mondo[0];
 472         data1 = mondo[1];
 473         data2 = mondo[2];
 474         for (i = 0; i < cnt; i++)
 475                 spitfire_xcall_helper(data0, data1, data2, pstate, cpu_list[i]);
 476 }
 477
 478 /* Cheetah now allows to send the whole 64-bytes of data in the interrupt
 479  * packet, but we have no use for that.  However we do take advantage of
 480  * the new pipelining feature (ie. dispatch to multiple cpus simultaneously).
 481  */
 482 static void cheetah_xcall_deliver(struct trap_per_cpu *tb, int cnt)
 483 {
 484         int nack_busy_id, is_jbus, need_more;
 485         u64 *mondo, pstate, ver, busy_mask;
 486         u16 *cpu_list;
 487
 488         cpu_list = __va(tb->cpu_list_pa);
 489         mondo = __va(tb->cpu_mondo_block_pa);
 490
 491         /* Unfortunately, someone at Sun had the brilliant idea to make the
 492          * busy/nack fields hard-coded by ITID number for this Ultra-III
 493          * derivative processor.
 494          */
 495         __asm__ ("rdpr %%ver, %0" : "=r" (ver));
 496         is_jbus = ((ver >> 32) == __JALAPENO_ID ||
 497                    (ver >> 32) == __SERRANO_ID);
 498
 499         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 500
 501 retry:
 502         need_more = 0;
 503         __asm__ __volatile__("wrpr %0, %1, %%pstate\n\t"
 504                              : : "r" (pstate), "i" (PSTATE_IE));
 505
 506         /* Setup the dispatch data registers. */
 507         __asm__ __volatile__("stxa      %0, [%3] %6\n\t"
 508                              "stxa      %1, [%4] %6\n\t"
 509                              "stxa      %2, [%5] %6\n\t"
 510                              "membar    #Sync\n\t"
 511                              : /* no outputs */
 512                              : "r" (mondo[0]), "r" (mondo[1]), "r" (mondo[2]),
 513                                "r" (0x40), "r" (0x50), "r" (0x60),
 514                                "i" (ASI_INTR_W));
 515
 516         nack_busy_id = 0;
 517         busy_mask = 0;
 518         {
 519                 int i;
 520
 521                 for (i = 0; i < cnt; i++) {
 522                         u64 target, nr;
 523
 524                         nr = cpu_list[i];
 525                         if (nr == 0xffff)
 526                                 continue;
 527
 528                         target = (nr << 14) | 0x70;
 529                         if (is_jbus) {
 530                                 busy_mask |= (0x1UL << (nr * 2));
 531                         } else {
 532                                 target |= (nack_busy_id << 24);
 533                                 busy_mask |= (0x1UL <<
 534                                               (nack_busy_id * 2));
 535                         }
 536                         __asm__ __volatile__(
 537                                 "stxa   %%g0, [%0] %1\n\t"
 538                                 "membar #Sync\n\t"
 539                                 : /* no outputs */
 540                                 : "r" (target), "i" (ASI_INTR_W));
 541                         nack_busy_id++;
 542                         if (nack_busy_id == 32) {
 543                                 need_more = 1;
 544                                 break;
 545                         }
 546                 }
 547         }
 548
 549         /* Now, poll for completion. */
 550         {
 551                 u64 dispatch_stat, nack_mask;
 552                 long stuck;
 553
 554                 stuck = 100000 * nack_busy_id;
 555                 nack_mask = busy_mask << 1;
 556                 do {
 557                         __asm__ __volatile__("ldxa      [%%g0] %1, %0"
 558                                              : "=r" (dispatch_stat)
 559                                              : "i" (ASI_INTR_DISPATCH_STAT));
 560                         if (!(dispatch_stat & (busy_mask | nack_mask))) {
 561                                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 562                                                      : : "r" (pstate));
 563                                 if (unlikely(need_more)) {
 564                                         int i, this_cnt = 0;
 565                                         for (i = 0; i < cnt; i++) {
 566                                                 if (cpu_list[i] == 0xffff)
 567                                                         continue;
 568                                                 cpu_list[i] = 0xffff;
 569                                                 this_cnt++;
 570                                                 if (this_cnt == 32)
 571                                                         break;
 572                                         }
 573                                         goto retry;
 574                                 }
 575                                 return;
 576                         }
 577                         if (!--stuck)
 578                                 break;
 579                 } while (dispatch_stat & busy_mask);
 580
 581                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 582                                      : : "r" (pstate));
 583
 584                 if (dispatch_stat & busy_mask) {
 585                         /* Busy bits will not clear, continue instead
 586                          * of freezing up on this cpu.
 587                          */
 588                         printk("CPU[%d]: mondo stuckage result[%016llx]\n",
 589                                smp_processor_id(), dispatch_stat);
 590                 } else {
 591                         int i, this_busy_nack = 0;
 592
 593                         /* Delay some random time with interrupts enabled
 594                          * to prevent deadlock.
 595                          */
 596                         udelay(2 * nack_busy_id);
 597
 598                         /* Clear out the mask bits for cpus which did not
 599                          * NACK us.
 600                          */
 601                         for (i = 0; i < cnt; i++) {
 602                                 u64 check_mask, nr;
 603
 604                                 nr = cpu_list[i];
 605                                 if (nr == 0xffff)
 606                                         continue;
 607
 608                                 if (is_jbus)
 609                                         check_mask = (0x2UL << (2*nr));
 610                                 else
 611                                         check_mask = (0x2UL <<
 612                                                       this_busy_nack);
 613                                 if ((dispatch_stat & check_mask) == 0)
 614                                         cpu_list[i] = 0xffff;
 615                                 this_busy_nack += 2;
 616                                 if (this_busy_nack == 64)
 617                                         break;
 618                         }
 619
 620                         goto retry;
 621                 }
 622         }
 623 }
 624
 625 /* Multi-cpu list version.  */
 626 static void hypervisor_xcall_deliver(struct trap_per_cpu *tb, int cnt)
 627 {
 628         int retries, this_cpu, prev_sent, i, saw_cpu_error;
 629         unsigned long status;
 630         u16 *cpu_list;
 631
 632         this_cpu = smp_processor_id();
 633
 634         cpu_list = __va(tb->cpu_list_pa);
 635
 636         saw_cpu_error = 0;
 637         retries = 0;
 638         prev_sent = 0;
 639         do {
 640                 int forward_progress, n_sent;
 641
 642                 status = sun4v_cpu_mondo_send(cnt,
 643                                               tb->cpu_list_pa,
 644                                               tb->cpu_mondo_block_pa);
 645
 646                 /* HV_EOK means all cpus received the xcall, we're done.  */
 647                 if (likely(status == HV_EOK))
 648                         break;
 649
 650                 /* First, see if we made any forward progress.
 651                  *
 652                  * The hypervisor indicates successful sends by setting
 653                  * cpu list entries to the value 0xffff.
 654                  */
 655                 n_sent = 0;
 656                 for (i = 0; i < cnt; i++) {
 657                         if (likely(cpu_list[i] == 0xffff))
 658                                 n_sent++;
 659                 }
 660
 661                 forward_progress = 0;
 662                 if (n_sent > prev_sent)
 663                         forward_progress = 1;
 664
 665                 prev_sent = n_sent;
 666
 667                 /* If we get a HV_ECPUERROR, then one or more of the cpus
 668                  * in the list are in error state.  Use the cpu_state()
 669                  * hypervisor call to find out which cpus are in error state.
 670                  */
 671                 if (unlikely(status == HV_ECPUERROR)) {
 672                         for (i = 0; i < cnt; i++) {
 673                                 long err;
 674                                 u16 cpu;
 675
 676                                 cpu = cpu_list[i];
 677                                 if (cpu == 0xffff)
 678                                         continue;
 679
 680                                 err = sun4v_cpu_state(cpu);
 681                                 if (err == HV_CPU_STATE_ERROR) {
 682                                         saw_cpu_error = (cpu + 1);
 683                                         cpu_list[i] = 0xffff;
 684                                 }
 685                         }
 686                 } else if (unlikely(status != HV_EWOULDBLOCK))
 687                         goto fatal_mondo_error;
 688
 689                 /* Don't bother rewriting the CPU list, just leave the
 690                  * 0xffff and non-0xffff entries in there and the
 691                  * hypervisor will do the right thing.
 692                  *
 693                  * Only advance timeout state if we didn't make any
 694                  * forward progress.
 695                  */
 696                 if (unlikely(!forward_progress)) {
 697                         if (unlikely(++retries > 10000))
 698                                 goto fatal_mondo_timeout;
 699
 700                         /* Delay a little bit to let other cpus catch up
 701                          * on their cpu mondo queue work.
 702                          */
 703                         udelay(2 * cnt);
 704                 }
 705         } while (1);
 706
 707         if (unlikely(saw_cpu_error))
 708                 goto fatal_mondo_cpu_error;
 709
 710         return;
 711
 712 fatal_mondo_cpu_error:
 713         printk(KERN_CRIT "CPU[%d]: SUN4V mondo cpu error, some target cpus "
 714                "(including %d) were in error state\n",
 715                this_cpu, saw_cpu_error - 1);
 716         return;
 717
 718 fatal_mondo_timeout:
 719         printk(KERN_CRIT "CPU[%d]: SUN4V mondo timeout, no forward "
 720                " progress after %d retries.\n",
 721                this_cpu, retries);
 722         goto dump_cpu_list_and_out;
 723
 724 fatal_mondo_error:
 725         printk(KERN_CRIT "CPU[%d]: Unexpected SUN4V mondo error %lu\n",
 726                this_cpu, status);
 727         printk(KERN_CRIT "CPU[%d]: Args were cnt(%d) cpulist_pa(%lx) "
 728                "mondo_block_pa(%lx)\n",
 729                this_cpu, cnt, tb->cpu_list_pa, tb->cpu_mondo_block_pa);
 730
 731 dump_cpu_list_and_out:
 732         printk(KERN_CRIT "CPU[%d]: CPU list [ ", this_cpu);
 733         for (i = 0; i < cnt; i++)
 734                 printk("%u ", cpu_list[i]);
 735         printk("]\n");
 736 }
 737
 738 static void (*xcall_deliver_impl)(struct trap_per_cpu *, int);
 739
 740 static void xcall_deliver(u64 data0, u64 data1, u64 data2, const cpumask_t *mask)
 741 {
 742         struct trap_per_cpu *tb;
 743         int this_cpu, i, cnt;
 744         unsigned long flags;
 745         u16 *cpu_list;
 746         u64 *mondo;
 747
 748         /* We have to do this whole thing with interrupts fully disabled.
 749          * Otherwise if we send an xcall from interrupt context it will
 750          * corrupt both our mondo block and cpu list state.
 751          *
 752          * One consequence of this is that we cannot use timeout mechanisms
 753          * that depend upon interrupts being delivered locally.  So, for
 754          * example, we cannot sample jiffies and expect it to advance.
 755          *
 756          * Fortunately, udelay() uses %stick/%tick so we can use that.
 757          */
 758         local_irq_save(flags);
 759
 760         this_cpu = smp_processor_id();
 761         tb = &trap_block[this_cpu];
 762
 763         mondo = __va(tb->cpu_mondo_block_pa);
 764         mondo[0] = data0;
 765         mondo[1] = data1;
 766         mondo[2] = data2;
 767         wmb();
 768
 769         cpu_list = __va(tb->cpu_list_pa);
 770
 771         /* Setup the initial cpu list.  */
 772         cnt = 0;
 773         for_each_cpu(i, mask) {
 774                 if (i == this_cpu || !cpu_online(i))
 775                         continue;
 776                 cpu_list[cnt++] = i;
 777         }
 778
 779         if (cnt)
 780                 xcall_deliver_impl(tb, cnt);
 781
 782         local_irq_restore(flags);
 783 }
 784
 785 /* Send cross call to all processors mentioned in MASK_P
 786  * except self.  Really, there are only two cases currently,
 787  * "&cpu_online_map" and "&mm->cpu_vm_mask".
 788  */
 789 static void smp_cross_call_masked(unsigned long *func, u32 ctx, u64 data1, u64 data2, const cpumask_t *mask)
 790 {
 791         u64 data0 = (((u64)ctx)<<32 | (((u64)func) & 0xffffffff));
 792
 793         xcall_deliver(data0, data1, data2, mask);
 794 }
 795
 796 /* Send cross call to all processors except self. */
 797 static void smp_cross_call(unsigned long *func, u32 ctx, u64 data1, u64 data2)
 798 {
 799         smp_cross_call_masked(func, ctx, data1, data2, &cpu_online_map);
 800 }
 801
 802 extern unsigned long xcall_sync_tick;
 803
 804 static void smp_start_sync_tick_client(int cpu)
 805 {
 806         xcall_deliver((u64) &xcall_sync_tick, 0, 0,
 807                       &cpumask_of_cpu(cpu));
 808 }
 809
 810 extern unsigned long xcall_call_function;
 811
 812 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
 813 {
 814         xcall_deliver((u64) &xcall_call_function, 0, 0, mask);
 815 }
 816
 817 extern unsigned long xcall_call_function_single;
 818
 819 void arch_send_call_function_single_ipi(int cpu)
 820 {
 821         xcall_deliver((u64) &xcall_call_function_single, 0, 0,
 822                       &cpumask_of_cpu(cpu));
 823 }
 824
 825 void smp_call_function_client(int irq, struct pt_regs *regs)
 826 {
 827         clear_softint(1 << irq);
 828         generic_smp_call_function_interrupt();
 829 }
 830
 831 void smp_call_function_single_client(int irq, struct pt_regs *regs)
 832 {
 833         clear_softint(1 << irq);
 834         generic_smp_call_function_single_interrupt();
 835 }
 836
 837 static void tsb_sync(void *info)
 838 {
 839         struct trap_per_cpu *tp = &trap_block[raw_smp_processor_id()];
 840         struct mm_struct *mm = info;
 841
 842         /* It is not valid to test "currrent->active_mm == mm" here.
 843          *
 844          * The value of "current" is not changed atomically with
 845          * switch_mm().  But that's OK, we just need to check the
 846          * current cpu's trap block PGD physical address.
 847          */
 848         if (tp->pgd_paddr == __pa(mm->pgd))
 849                 tsb_context_switch(mm);
 850 }
 851
 852 void smp_tsb_sync(struct mm_struct *mm)
 853 {
 854         smp_call_function_many(mm_cpumask(mm), tsb_sync, mm, 1);
 855 }
 856
 857 extern unsigned long xcall_flush_tlb_mm;
 858 extern unsigned long xcall_flush_tlb_pending;
 859 extern unsigned long xcall_flush_tlb_kernel_range;
 860 extern unsigned long xcall_fetch_glob_regs;
 861 extern unsigned long xcall_receive_signal;
 862 extern unsigned long xcall_new_mmu_context_version;
 863 #ifdef CONFIG_KGDB
 864 extern unsigned long xcall_kgdb_capture;
 865 #endif
 866
 867 #ifdef DCACHE_ALIASING_POSSIBLE
 868 extern unsigned long xcall_flush_dcache_page_cheetah;
 869 #endif
 870 extern unsigned long xcall_flush_dcache_page_spitfire;
 871
 872 #ifdef CONFIG_DEBUG_DCFLUSH
 873 extern atomic_t dcpage_flushes;
 874 extern atomic_t dcpage_flushes_xcall;
 875 #endif
 876
 877 static inline void __local_flush_dcache_page(struct page *page)
 878 {
 879 #ifdef DCACHE_ALIASING_POSSIBLE
 880         __flush_dcache_page(page_address(page),
 881                             ((tlb_type == spitfire) &&
 882                              page_mapping(page) != NULL));
 883 #else
 884         if (page_mapping(page) != NULL &&
 885             tlb_type == spitfire)
 886                 __flush_icache_page(__pa(page_address(page)));
 887 #endif
 888 }
 889
 890 void smp_flush_dcache_page_impl(struct page *page, int cpu)
 891 {
 892         int this_cpu;
 893
 894         if (tlb_type == hypervisor)
 895                 return;
 896
 897 #ifdef CONFIG_DEBUG_DCFLUSH
 898         atomic_inc(&dcpage_flushes);
 899 #endif
 900
 901         this_cpu = get_cpu();
 902
 903         if (cpu == this_cpu) {
 904                 __local_flush_dcache_page(page);
 905         } else if (cpu_online(cpu)) {
 906                 void *pg_addr = page_address(page);
 907                 u64 data0 = 0;
 908
 909                 if (tlb_type == spitfire) {
 910                         data0 = ((u64)&xcall_flush_dcache_page_spitfire);
 911                         if (page_mapping(page) != NULL)
 912                                 data0 |= ((u64)1 << 32);
 913                 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 914 #ifdef DCACHE_ALIASING_POSSIBLE
 915                         data0 = ((u64)&xcall_flush_dcache_page_cheetah);
 916 #endif
 917                 }
 918                 if (data0) {
 919                         xcall_deliver(data0, __pa(pg_addr),
 920                                       (u64) pg_addr, &cpumask_of_cpu(cpu));
 921 #ifdef CONFIG_DEBUG_DCFLUSH
 922                         atomic_inc(&dcpage_flushes_xcall);
 923 #endif
 924                 }
 925         }
 926
 927         put_cpu();
 928 }
 929
 930 void flush_dcache_page_all(struct mm_struct *mm, struct page *page)
 931 {
 932         void *pg_addr;
 933         int this_cpu;
 934         u64 data0;
 935
 936         if (tlb_type == hypervisor)
 937                 return;
 938
 939         this_cpu = get_cpu();
 940
 941 #ifdef CONFIG_DEBUG_DCFLUSH
 942         atomic_inc(&dcpage_flushes);
 943 #endif
 944         data0 = 0;
 945         pg_addr = page_address(page);
 946         if (tlb_type == spitfire) {
 947                 data0 = ((u64)&xcall_flush_dcache_page_spitfire);
 948                 if (page_mapping(page) != NULL)
 949                         data0 |= ((u64)1 << 32);
 950         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 951 #ifdef DCACHE_ALIASING_POSSIBLE
 952                 data0 = ((u64)&xcall_flush_dcache_page_cheetah);
 953 #endif
 954         }
 955         if (data0) {
 956                 xcall_deliver(data0, __pa(pg_addr),
 957                               (u64) pg_addr, &cpu_online_map);
 958 #ifdef CONFIG_DEBUG_DCFLUSH
 959                 atomic_inc(&dcpage_flushes_xcall);
 960 #endif
 961         }
 962         __local_flush_dcache_page(page);
 963
 964         put_cpu();
 965 }
 966
 967 void smp_new_mmu_context_version_client(int irq, struct pt_regs *regs)
 968 {
 969         struct mm_struct *mm;
 970         unsigned long flags;
 971
 972         clear_softint(1 << irq);
 973
 974         /* See if we need to allocate a new TLB context because
 975          * the version of the one we are using is now out of date.
 976          */
 977         mm = current->active_mm;
 978         if (unlikely(!mm || (mm == &init_mm)))
 979                 return;
 980
 981         spin_lock_irqsave(&mm->context.lock, flags);
 982
 983         if (unlikely(!CTX_VALID(mm->context)))
 984                 get_new_mmu_context(mm);
 985
 986         spin_unlock_irqrestore(&mm->context.lock, flags);
 987
 988         load_secondary_context(mm);
 989         __flush_tlb_mm(CTX_HWBITS(mm->context),
 990                        SECONDARY_CONTEXT);
 991 }
 992
 993 void smp_new_mmu_context_version(void)
 994 {
 995         smp_cross_call(&xcall_new_mmu_context_version, 0, 0, 0);
 996 }
 997
 998 #ifdef CONFIG_KGDB
 999 void kgdb_roundup_cpus(unsigned long flags)
1000 {
1001         smp_cross_call(&xcall_kgdb_capture, 0, 0, 0);
1002 }
1003 #endif
1004
1005 void smp_fetch_global_regs(void)
1006 {
1007         smp_cross_call(&xcall_fetch_glob_regs, 0, 0, 0);
1008 }
1009
1010 /* We know that the window frames of the user have been flushed
1011  * to the stack before we get here because all callers of us
1012  * are flush_tlb_*() routines, and these run after flush_cache_*()
1013  * which performs the flushw.
1014  *
1015  * The SMP TLB coherency scheme we use works as follows:
1016  *
1017  * 1) mm->cpu_vm_mask is a bit mask of which cpus an address
1018  *    space has (potentially) executed on, this is the heuristic
1019  *    we use to avoid doing cross calls.
1020  *
1021  *    Also, for flushing from kswapd and also for clones, we
1022  *    use cpu_vm_mask as the list of cpus to make run the TLB.
1023  *
1024  * 2) TLB context numbers are shared globally across all processors
1025  *    in the system, this allows us to play several games to avoid
1026  *    cross calls.
1027  *
1028  *    One invariant is that when a cpu switches to a process, and
1029  *    that processes tsk->active_mm->cpu_vm_mask does not have the
1030  *    current cpu's bit set, that tlb context is flushed locally.
1031  *
1032  *    If the address space is non-shared (ie. mm->count == 1) we avoid
1033  *    cross calls when we want to flush the currently running process's
1034  *    tlb state.  This is done by clearing all cpu bits except the current
1035  *    processor's in current->mm->cpu_vm_mask and performing the
1036  *    flush locally only.  This will force any subsequent cpus which run
1037  *    this task to flush the context from the local tlb if the process
1038  *    migrates to another cpu (again).
1039  *
1040  * 3) For shared address spaces (threads) and swapping we bite the
1041  *    bullet for most cases and perform the cross call (but only to
1042  *    the cpus listed in cpu_vm_mask).
1043  *
1044  *    The performance gain from "optimizing" away the cross call for threads is
1045  *    questionable (in theory the big win for threads is the massive sharing of
1046  *    address space state across processors).
1047  */
1048
1049 /* This currently is only used by the hugetlb arch pre-fault
1050  * hook on UltraSPARC-III+ and later when changing the pagesize
1051  * bits of the context register for an address space.
1052  */
1053 void smp_flush_tlb_mm(struct mm_struct *mm)
1054 {
1055         u32 ctx = CTX_HWBITS(mm->context);
1056         int cpu = get_cpu();
1057
1058         if (atomic_read(&mm->mm_users) == 1) {
1059                 cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
1060                 goto local_flush_and_out;
1061         }
1062
1063         smp_cross_call_masked(&xcall_flush_tlb_mm,
1064                               ctx, 0, 0,
1065                               mm_cpumask(mm));
1066
1067 local_flush_and_out:
1068         __flush_tlb_mm(ctx, SECONDARY_CONTEXT);
1069
1070         put_cpu();
1071 }
1072
1073 void smp_flush_tlb_pending(struct mm_struct *mm, unsigned long nr, unsigned long *vaddrs)
1074 {
1075         u32 ctx = CTX_HWBITS(mm->context);
1076         int cpu = get_cpu();
1077
1078         if (mm == current->mm && atomic_read(&mm->mm_users) == 1)
1079                 cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
1080         else
1081                 smp_cross_call_masked(&xcall_flush_tlb_pending,
1082                                       ctx, nr, (unsigned long) vaddrs,
1083                                       mm_cpumask(mm));
1084
1085         __flush_tlb_pending(ctx, nr, vaddrs);
1086
1087         put_cpu();
1088 }
1089
1090 void smp_flush_tlb_kernel_range(unsigned long start, unsigned long end)
1091 {
1092         start &= PAGE_MASK;
1093         end    = PAGE_ALIGN(end);
1094         if (start != end) {
1095                 smp_cross_call(&xcall_flush_tlb_kernel_range,
1096                                0, start, end);
1097
1098                 __flush_tlb_kernel_range(start, end);
1099         }
1100 }
1101
1102 /* CPU capture. */
1103 /* #define CAPTURE_DEBUG */
1104 extern unsigned long xcall_capture;
1105
1106 static atomic_t smp_capture_depth = ATOMIC_INIT(0);
1107 static atomic_t smp_capture_registry = ATOMIC_INIT(0);
1108 static unsigned long penguins_are_doing_time;
1109
1110 void smp_capture(void)
1111 {
1112         int result = atomic_add_ret(1, &smp_capture_depth);
1113
1114         if (result == 1) {
1115                 int ncpus = num_online_cpus();
1116
1117 #ifdef CAPTURE_DEBUG
1118                 printk("CPU[%d]: Sending penguins to jail...",
1119                        smp_processor_id());
1120 #endif
1121                 penguins_are_doing_time = 1;
1122                 atomic_inc(&smp_capture_registry);
1123                 smp_cross_call(&xcall_capture, 0, 0, 0);
1124                 while (atomic_read(&smp_capture_registry) != ncpus)
1125                         rmb();
1126 #ifdef CAPTURE_DEBUG
1127                 printk("done\n");
1128 #endif
1129         }
1130 }
1131
1132 void smp_release(void)
1133 {
1134         if (atomic_dec_and_test(&smp_capture_depth)) {
1135 #ifdef CAPTURE_DEBUG
1136                 printk("CPU[%d]: Giving pardon to "
1137                        "imprisoned penguins\n",
1138                        smp_processor_id());
1139 #endif
1140                 penguins_are_doing_time = 0;
1141                 membar_safe("#StoreLoad");
1142                 atomic_dec(&smp_capture_registry);
1143         }
1144 }
1145
1146 /* Imprisoned penguins run with %pil == PIL_NORMAL_MAX, but PSTATE_IE
1147  * set, so they can service tlb flush xcalls...
1148  */
1149 extern void prom_world(int);
1150
1151 void smp_penguin_jailcell(int irq, struct pt_regs *regs)
1152 {
1153         clear_softint(1 << irq);
1154
1155         preempt_disable();
1156
1157         __asm__ __volatile__("flushw");
1158         prom_world(1);
1159         atomic_inc(&smp_capture_registry);
1160         membar_safe("#StoreLoad");
1161         while (penguins_are_doing_time)
1162                 rmb();
1163         atomic_dec(&smp_capture_registry);
1164         prom_world(0);
1165
1166         preempt_enable();
1167 }
1168
1169 /* /proc/profile writes can call this, don't __init it please. */
1170 int setup_profiling_timer(unsigned int multiplier)
1171 {
1172         return -EINVAL;
1173 }
1174
1175 void __init smp_prepare_cpus(unsigned int max_cpus)
1176 {
1177 }
1178
1179 void __devinit smp_prepare_boot_cpu(void)
1180 {
1181 }
1182
1183 void __init smp_setup_processor_id(void)
1184 {
1185         if (tlb_type == spitfire)
1186                 xcall_deliver_impl = spitfire_xcall_deliver;
1187         else if (tlb_type == cheetah || tlb_type == cheetah_plus)
1188                 xcall_deliver_impl = cheetah_xcall_deliver;
1189         else
1190                 xcall_deliver_impl = hypervisor_xcall_deliver;
1191 }
1192
1193 void __devinit smp_fill_in_sib_core_maps(void)
1194 {
1195         unsigned int i;
1196
1197         for_each_present_cpu(i) {
1198                 unsigned int j;
1199
1200                 cpus_clear(cpu_core_map[i]);
1201                 if (cpu_data(i).core_id == 0) {
1202                         cpu_set(i, cpu_core_map[i]);
1203                         continue;
1204                 }
1205
1206                 for_each_present_cpu(j) {
1207                         if (cpu_data(i).core_id ==
1208                             cpu_data(j).core_id)
1209                                 cpu_set(j, cpu_core_map[i]);
1210                 }
1211         }
1212
1213         for_each_present_cpu(i) {
1214                 unsigned int j;
1215
1216                 cpus_clear(per_cpu(cpu_sibling_map, i));
1217                 if (cpu_data(i).proc_id == -1) {
1218                         cpu_set(i, per_cpu(cpu_sibling_map, i));
1219                         continue;
1220                 }
1221
1222                 for_each_present_cpu(j) {
1223                         if (cpu_data(i).proc_id ==
1224                             cpu_data(j).proc_id)
1225                                 cpu_set(j, per_cpu(cpu_sibling_map, i));
1226                 }
1227         }
1228 }
1229
1230 int __cpuinit __cpu_up(unsigned int cpu)
1231 {
1232         int ret = smp_boot_one_cpu(cpu);
1233
1234         if (!ret) {
1235                 cpu_set(cpu, smp_commenced_mask);
1236                 while (!cpu_isset(cpu, cpu_online_map))
1237                         mb();
1238                 if (!cpu_isset(cpu, cpu_online_map)) {
1239                         ret = -ENODEV;
1240                 } else {
1241                         /* On SUN4V, writes to %tick and %stick are
1242                          * not allowed.
1243                          */
1244                         if (tlb_type != hypervisor)
1245                                 smp_synchronize_one_tick(cpu);
1246                 }
1247         }
1248         return ret;
1249 }
1250
1251 #ifdef CONFIG_HOTPLUG_CPU
1252 void cpu_play_dead(void)
1253 {
1254         int cpu = smp_processor_id();
1255         unsigned long pstate;
1256
1257         idle_task_exit();
1258
1259         if (tlb_type == hypervisor) {
1260                 struct trap_per_cpu *tb = &trap_block[cpu];
1261
1262                 sun4v_cpu_qconf(HV_CPU_QUEUE_CPU_MONDO,
1263                                 tb->cpu_mondo_pa, 0);
1264                 sun4v_cpu_qconf(HV_CPU_QUEUE_DEVICE_MONDO,
1265                                 tb->dev_mondo_pa, 0);
1266                 sun4v_cpu_qconf(HV_CPU_QUEUE_RES_ERROR,
1267                                 tb->resum_mondo_pa, 0);
1268                 sun4v_cpu_qconf(HV_CPU_QUEUE_NONRES_ERROR,
1269                                 tb->nonresum_mondo_pa, 0);
1270         }
1271
1272         cpu_clear(cpu, smp_commenced_mask);
1273         membar_safe("#Sync");
1274
1275         local_irq_disable();
1276
1277         __asm__ __volatile__(
1278                 "rdpr   %%pstate, %0\n\t"
1279                 "wrpr   %0, %1, %%pstate"
1280                 : "=r" (pstate)
1281                 : "i" (PSTATE_IE));
1282
1283         while (1)
1284                 barrier();
1285 }
1286
1287 int __cpu_disable(void)
1288 {
1289         int cpu = smp_processor_id();
1290         cpuinfo_sparc *c;
1291         int i;
1292
1293         for_each_cpu_mask(i, cpu_core_map[cpu])
1294                 cpu_clear(cpu, cpu_core_map[i]);
1295         cpus_clear(cpu_core_map[cpu]);
1296
1297         for_each_cpu_mask(i, per_cpu(cpu_sibling_map, cpu))
1298                 cpu_clear(cpu, per_cpu(cpu_sibling_map, i));
1299         cpus_clear(per_cpu(cpu_sibling_map, cpu));
1300
1301         c = &cpu_data(cpu);
1302
1303         c->core_id = 0;
1304         c->proc_id = -1;
1305
1306         smp_wmb();
1307
1308         /* Make sure no interrupts point to this cpu.  */
1309         fixup_irqs();
1310
1311         local_irq_enable();
1312         mdelay(1);
1313         local_irq_disable();
1314
1315         ipi_call_lock();
1316         cpu_clear(cpu, cpu_online_map);
1317         ipi_call_unlock();
1318
1319         cpu_map_rebuild();
1320
1321         return 0;
1322 }
1323
1324 void __cpu_die(unsigned int cpu)
1325 {
1326         int i;
1327
1328         for (i = 0; i < 100; i++) {
1329                 smp_rmb();
1330                 if (!cpu_isset(cpu, smp_commenced_mask))
1331                         break;
1332                 msleep(100);
1333         }
1334         if (cpu_isset(cpu, smp_commenced_mask)) {
1335                 printk(KERN_ERR "CPU %u didn't die...\n", cpu);
1336         } else {
1337 #if defined(CONFIG_SUN_LDOMS)
1338                 unsigned long hv_err;
1339                 int limit = 100;
1340
1341                 do {
1342                         hv_err = sun4v_cpu_stop(cpu);
1343                         if (hv_err == HV_EOK) {
1344                                 cpu_clear(cpu, cpu_present_map);
1345                                 break;
1346                         }
1347                 } while (--limit > 0);
1348                 if (limit <= 0) {
1349                         printk(KERN_ERR "sun4v_cpu_stop() fails err=%lu\n",
1350                                hv_err);
1351                 }
1352 #endif
1353         }
1354 }
1355 #endif
1356
1357 void __init smp_cpus_done(unsigned int max_cpus)
1358 {
1359 }
1360
1361 void smp_send_reschedule(int cpu)
1362 {
1363         xcall_deliver((u64) &xcall_receive_signal, 0, 0,
1364                       &cpumask_of_cpu(cpu));
1365 }
1366
1367 void smp_receive_signal_client(int irq, struct pt_regs *regs)
1368 {
1369         clear_softint(1 << irq);
1370 }
1371
1372 /* This is a nop because we capture all other cpus
1373  * anyways when making the PROM active.
1374  */
1375 void smp_send_stop(void)
1376 {
1377 }
1378
1379 /**
1380  * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu
1381  * @cpu: cpu to allocate for
1382  * @size: size allocation in bytes
1383  * @align: alignment
1384  *
1385  * Allocate @size bytes aligned at @align for cpu @cpu.  This wrapper
1386  * does the right thing for NUMA regardless of the current
1387  * configuration.
1388  *
1389  * RETURNS:
1390  * Pointer to the allocated area on success, NULL on failure.
1391  */
1392 static void * __init pcpu_alloc_bootmem(unsigned int cpu, unsigned long size,
1393                                         unsigned long align)
1394 {
1395         const unsigned long goal = __pa(MAX_DMA_ADDRESS);
1396 #ifdef CONFIG_NEED_MULTIPLE_NODES
1397         int node = cpu_to_node(cpu);
1398         void *ptr;
1399
1400         if (!node_online(node) || !NODE_DATA(node)) {
1401                 ptr = __alloc_bootmem(size, align, goal);
1402                 pr_info("cpu %d has no node %d or node-local memory\n",
1403                         cpu, node);
1404                 pr_debug("per cpu data for cpu%d %lu bytes at %016lx\n",
1405                          cpu, size, __pa(ptr));
1406         } else {
1407                 ptr = __alloc_bootmem_node(NODE_DATA(node),
1408                                            size, align, goal);
1409                 pr_debug("per cpu data for cpu%d %lu bytes on node%d at "
1410                          "%016lx\n", cpu, size, node, __pa(ptr));
1411         }
1412         return ptr;
1413 #else
1414         return __alloc_bootmem(size, align, goal);
1415 #endif
1416 }
1417
1418 static size_t pcpur_size __initdata;
1419 static void **pcpur_ptrs __initdata;
1420
1421 static struct page * __init pcpur_get_page(unsigned int cpu, int pageno)
1422 {
1423         size_t off = (size_t)pageno << PAGE_SHIFT;
1424
1425         if (off >= pcpur_size)
1426                 return NULL;
1427
1428         return virt_to_page(pcpur_ptrs[cpu] + off);
1429 }
1430
1431 #define PCPU_CHUNK_SIZE (4UL * 1024UL * 1024UL)
1432
1433 static void __init pcpu_map_range(unsigned long start, unsigned long end,
1434                                   struct page *page)
1435 {
1436         unsigned long pfn = page_to_pfn(page);
1437         unsigned long pte_base;
1438
1439         BUG_ON((pfn<<PAGE_SHIFT)&(PCPU_CHUNK_SIZE - 1UL));
1440
1441         pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
1442                     _PAGE_CP_4U | _PAGE_CV_4U |
1443                     _PAGE_P_4U | _PAGE_W_4U);
1444         if (tlb_type == hypervisor)
1445                 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
1446                             _PAGE_CP_4V | _PAGE_CV_4V |
1447                             _PAGE_P_4V | _PAGE_W_4V);
1448
1449         while (start < end) {
1450                 pgd_t *pgd = pgd_offset_k(start);
1451                 unsigned long this_end;
1452                 pud_t *pud;
1453                 pmd_t *pmd;
1454                 pte_t *pte;
1455
1456                 pud = pud_offset(pgd, start);
1457                 if (pud_none(*pud)) {
1458                         pmd_t *new;
1459
1460                         new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1461                         pud_populate(&init_mm, pud, new);
1462                 }
1463
1464                 pmd = pmd_offset(pud, start);
1465                 if (!pmd_present(*pmd)) {
1466                         pte_t *new;
1467
1468                         new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1469                         pmd_populate_kernel(&init_mm, pmd, new);
1470                 }
1471
1472                 pte = pte_offset_kernel(pmd, start);
1473                 this_end = (start + PMD_SIZE) & PMD_MASK;
1474                 if (this_end > end)
1475                         this_end = end;
1476
1477                 while (start < this_end) {
1478                         unsigned long paddr = pfn << PAGE_SHIFT;
1479
1480                         pte_val(*pte) = (paddr | pte_base);
1481
1482                         start += PAGE_SIZE;
1483                         pte++;
1484                         pfn++;
1485                 }
1486         }
1487 }
1488
1489 void __init setup_per_cpu_areas(void)
1490 {
1491         size_t dyn_size, static_size = __per_cpu_end - __per_cpu_start;
1492         static struct vm_struct vm;
1493         unsigned long delta, cpu;
1494         size_t pcpu_unit_size;
1495         size_t ptrs_size;
1496
1497         pcpur_size = PFN_ALIGN(static_size + PERCPU_MODULE_RESERVE +
1498                                PERCPU_DYNAMIC_RESERVE);
1499         dyn_size = pcpur_size - static_size - PERCPU_MODULE_RESERVE;
1500
1501
1502         ptrs_size = PFN_ALIGN(nr_cpu_ids * sizeof(pcpur_ptrs[0]));
1503         pcpur_ptrs = alloc_bootmem(ptrs_size);
1504
1505         for_each_possible_cpu(cpu) {
1506                 pcpur_ptrs[cpu] = pcpu_alloc_bootmem(cpu, PCPU_CHUNK_SIZE,
1507                                                      PCPU_CHUNK_SIZE);
1508
1509                 free_bootmem(__pa(pcpur_ptrs[cpu] + pcpur_size),
1510                              PCPU_CHUNK_SIZE - pcpur_size);
1511
1512                 memcpy(pcpur_ptrs[cpu], __per_cpu_load, static_size);
1513         }
1514
1515         /* allocate address and map */
1516         vm.flags = VM_ALLOC;
1517         vm.size = nr_cpu_ids * PCPU_CHUNK_SIZE;
1518         vm_area_register_early(&vm, PCPU_CHUNK_SIZE);
1519
1520         for_each_possible_cpu(cpu) {
1521                 unsigned long start = (unsigned long) vm.addr;
1522                 unsigned long end;
1523
1524                 start += cpu * PCPU_CHUNK_SIZE;
1525                 end = start + PCPU_CHUNK_SIZE;
1526                 pcpu_map_range(start, end, virt_to_page(pcpur_ptrs[cpu]));
1527         }
1528
1529         pcpu_unit_size = pcpu_setup_first_chunk(pcpur_get_page, static_size,
1530                                                 PERCPU_MODULE_RESERVE, dyn_size,
1531                                                 PCPU_CHUNK_SIZE, vm.addr, NULL);
1532
1533         free_bootmem(__pa(pcpur_ptrs), ptrs_size);
1534
1535         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1536         for_each_possible_cpu(cpu) {
1537                 __per_cpu_offset(cpu) = delta + cpu * pcpu_unit_size;
1538         }
1539
1540         /* Setup %g5 for the boot cpu.  */
1541         __local_per_cpu_offset = __per_cpu_offset(smp_processor_id());
1542
1543         of_fill_in_cpu_data();
1544         if (tlb_type == hypervisor)
1545                 mdesc_fill_in_cpu_data(cpu_all_mask);
1546 }