git.oblomov.eu Git - linux-2.6/blob - arch/sparc64/kernel/smp.c

   1 /* smp.c: Sparc64 SMP support.
   2  *
   3  * Copyright (C) 1997, 2007 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
  25 #include <asm/head.h>
  26 #include <asm/ptrace.h>
  27 #include <asm/atomic.h>
  28 #include <asm/tlbflush.h>
  29 #include <asm/mmu_context.h>
  30 #include <asm/cpudata.h>
  31 #include <asm/hvtramp.h>
  32 #include <asm/io.h>
  33
  34 #include <asm/irq.h>
  35 #include <asm/irq_regs.h>
  36 #include <asm/page.h>
  37 #include <asm/pgtable.h>
  38 #include <asm/oplib.h>
  39 #include <asm/uaccess.h>
  40 #include <asm/timer.h>
  41 #include <asm/starfire.h>
  42 #include <asm/tlb.h>
  43 #include <asm/sections.h>
  44 #include <asm/prom.h>
  45 #include <asm/mdesc.h>
  46 #include <asm/ldc.h>
  47 #include <asm/hypervisor.h>
  48
  49 extern void calibrate_delay(void);
  50
  51 int sparc64_multi_core __read_mostly;
  52
  53 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_NONE;
  54 cpumask_t cpu_online_map __read_mostly = CPU_MASK_NONE;
  55 cpumask_t cpu_sibling_map[NR_CPUS] __read_mostly =
  56         { [0 ... NR_CPUS-1] = CPU_MASK_NONE };
  57 cpumask_t cpu_core_map[NR_CPUS] __read_mostly =
  58         { [0 ... NR_CPUS-1] = CPU_MASK_NONE };
  59
  60 EXPORT_SYMBOL(cpu_possible_map);
  61 EXPORT_SYMBOL(cpu_online_map);
  62 EXPORT_SYMBOL(cpu_sibling_map);
  63 EXPORT_SYMBOL(cpu_core_map);
  64
  65 static cpumask_t smp_commenced_mask;
  66
  67 void smp_info(struct seq_file *m)
  68 {
  69         int i;
  70
  71         seq_printf(m, "State:\n");
  72         for_each_online_cpu(i)
  73                 seq_printf(m, "CPU%d:\t\tonline\n", i);
  74 }
  75
  76 void smp_bogo(struct seq_file *m)
  77 {
  78         int i;
  79
  80         for_each_online_cpu(i)
  81                 seq_printf(m,
  82                            "Cpu%dClkTck\t: %016lx\n",
  83                            i, cpu_data(i).clock_tick);
  84 }
  85
  86 static __cacheline_aligned_in_smp DEFINE_SPINLOCK(call_lock);
  87
  88 extern void setup_sparc64_timer(void);
  89
  90 static volatile unsigned long callin_flag = 0;
  91
  92 void __devinit smp_callin(void)
  93 {
  94         int cpuid = hard_smp_processor_id();
  95
  96         __local_per_cpu_offset = __per_cpu_offset(cpuid);
  97
  98         if (tlb_type == hypervisor)
  99                 sun4v_ktsb_register();
 100
 101         __flush_tlb_all();
 102
 103         setup_sparc64_timer();
 104
 105         if (cheetah_pcache_forced_on)
 106                 cheetah_enable_pcache();
 107
 108         local_irq_enable();
 109
 110         callin_flag = 1;
 111         __asm__ __volatile__("membar #Sync\n\t"
 112                              "flush  %%g6" : : : "memory");
 113
 114         /* Clear this or we will die instantly when we
 115          * schedule back to this idler...
 116          */
 117         current_thread_info()->new_child = 0;
 118
 119         /* Attach to the address space of init_task. */
 120         atomic_inc(&init_mm.mm_count);
 121         current->active_mm = &init_mm;
 122
 123         while (!cpu_isset(cpuid, smp_commenced_mask))
 124                 rmb();
 125
 126         spin_lock(&call_lock);
 127         cpu_set(cpuid, cpu_online_map);
 128         spin_unlock(&call_lock);
 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_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 (last diff %ld cycles,"
 241                "maxerr %lu cycles)\n", smp_processor_id(), delta, rt);
 242 }
 243
 244 static void smp_start_sync_tick_client(int cpu);
 245
 246 static void smp_synchronize_one_tick(int cpu)
 247 {
 248         unsigned long flags, i;
 249
 250         go[MASTER] = 0;
 251
 252         smp_start_sync_tick_client(cpu);
 253
 254         /* wait for client to be ready */
 255         while (!go[MASTER])
 256                 rmb();
 257
 258         /* now let the client proceed into his loop */
 259         go[MASTER] = 0;
 260         membar_storeload();
 261
 262         spin_lock_irqsave(&itc_sync_lock, flags);
 263         {
 264                 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; i++) {
 265                         while (!go[MASTER])
 266                                 rmb();
 267                         go[MASTER] = 0;
 268                         wmb();
 269                         go[SLAVE] = tick_ops->get_tick();
 270                         membar_storeload();
 271                 }
 272         }
 273         spin_unlock_irqrestore(&itc_sync_lock, flags);
 274 }
 275
 276 #if defined(CONFIG_SUN_LDOMS) && defined(CONFIG_HOTPLUG_CPU)
 277 /* XXX Put this in some common place. XXX */
 278 static unsigned long kimage_addr_to_ra(void *p)
 279 {
 280         unsigned long val = (unsigned long) p;
 281
 282         return kern_base + (val - KERNBASE);
 283 }
 284
 285 static void ldom_startcpu_cpuid(unsigned int cpu, unsigned long thread_reg)
 286 {
 287         extern unsigned long sparc64_ttable_tl0;
 288         extern unsigned long kern_locked_tte_data;
 289         extern int bigkernel;
 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
 296         hdesc = kzalloc(sizeof(*hdesc), GFP_KERNEL);
 297         if (!hdesc) {
 298                 printk(KERN_ERR "ldom_startcpu_cpuid: Cannot allocate "
 299                        "hvtramp_descr.\n");
 300                 return;
 301         }
 302
 303         hdesc->cpu = cpu;
 304         hdesc->num_mappings = (bigkernel ? 2 : 1);
 305
 306         tb = &trap_block[cpu];
 307         tb->hdesc = hdesc;
 308
 309         hdesc->fault_info_va = (unsigned long) &tb->fault_info;
 310         hdesc->fault_info_pa = kimage_addr_to_ra(&tb->fault_info);
 311
 312         hdesc->thread_reg = thread_reg;
 313
 314         tte_vaddr = (unsigned long) KERNBASE;
 315         tte_data = kern_locked_tte_data;
 316
 317         hdesc->maps[0].vaddr = tte_vaddr;
 318         hdesc->maps[0].tte   = tte_data;
 319         if (bigkernel) {
 320                 tte_vaddr += 0x400000;
 321                 tte_data  += 0x400000;
 322                 hdesc->maps[1].vaddr = tte_vaddr;
 323                 hdesc->maps[1].tte   = tte_data;
 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 __devinit smp_boot_one_cpu(unsigned int cpu)
 346 {
 347         struct trap_per_cpu *tb = &trap_block[cpu];
 348         unsigned long entry =
 349                 (unsigned long)(&sparc64_cpu_startup);
 350         unsigned long cookie =
 351                 (unsigned long)(&cpu_new_thread);
 352         struct task_struct *p;
 353         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                 else
 367 #endif
 368                         prom_startcpu_cpuid(cpu, entry, cookie);
 369         } else {
 370                 struct device_node *dp = of_find_node_by_cpuid(cpu);
 371
 372                 prom_startcpu(dp->node, entry, cookie);
 373         }
 374
 375         for (timeout = 0; timeout < 50000; timeout++) {
 376                 if (callin_flag)
 377                         break;
 378                 udelay(100);
 379         }
 380
 381         if (callin_flag) {
 382                 ret = 0;
 383         } else {
 384                 printk("Processor %d is stuck.\n", cpu);
 385                 ret = -ENODEV;
 386         }
 387         cpu_new_thread = NULL;
 388
 389         if (tb->hdesc) {
 390                 kfree(tb->hdesc);
 391                 tb->hdesc = NULL;
 392         }
 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[%016lx]\n",
 456                        smp_processor_id(), result);
 457         } else {
 458                 udelay(2);
 459                 goto again;
 460         }
 461 }
 462
 463 static __inline__ void spitfire_xcall_deliver(u64 data0, u64 data1, u64 data2, cpumask_t mask)
 464 {
 465         u64 pstate;
 466         int i;
 467
 468         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 469         for_each_cpu_mask(i, mask)
 470                 spitfire_xcall_helper(data0, data1, data2, pstate, i);
 471 }
 472
 473 /* Cheetah now allows to send the whole 64-bytes of data in the interrupt
 474  * packet, but we have no use for that.  However we do take advantage of
 475  * the new pipelining feature (ie. dispatch to multiple cpus simultaneously).
 476  */
 477 static void cheetah_xcall_deliver(u64 data0, u64 data1, u64 data2, cpumask_t mask)
 478 {
 479         u64 pstate, ver;
 480         int nack_busy_id, is_jbus, need_more;
 481
 482         if (cpus_empty(mask))
 483                 return;
 484
 485         /* Unfortunately, someone at Sun had the brilliant idea to make the
 486          * busy/nack fields hard-coded by ITID number for this Ultra-III
 487          * derivative processor.
 488          */
 489         __asm__ ("rdpr %%ver, %0" : "=r" (ver));
 490         is_jbus = ((ver >> 32) == __JALAPENO_ID ||
 491                    (ver >> 32) == __SERRANO_ID);
 492
 493         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 494
 495 retry:
 496         need_more = 0;
 497         __asm__ __volatile__("wrpr %0, %1, %%pstate\n\t"
 498                              : : "r" (pstate), "i" (PSTATE_IE));
 499
 500         /* Setup the dispatch data registers. */
 501         __asm__ __volatile__("stxa      %0, [%3] %6\n\t"
 502                              "stxa      %1, [%4] %6\n\t"
 503                              "stxa      %2, [%5] %6\n\t"
 504                              "membar    #Sync\n\t"
 505                              : /* no outputs */
 506                              : "r" (data0), "r" (data1), "r" (data2),
 507                                "r" (0x40), "r" (0x50), "r" (0x60),
 508                                "i" (ASI_INTR_W));
 509
 510         nack_busy_id = 0;
 511         {
 512                 int i;
 513
 514                 for_each_cpu_mask(i, mask) {
 515                         u64 target = (i << 14) | 0x70;
 516
 517                         if (!is_jbus)
 518                                 target |= (nack_busy_id << 24);
 519                         __asm__ __volatile__(
 520                                 "stxa   %%g0, [%0] %1\n\t"
 521                                 "membar #Sync\n\t"
 522                                 : /* no outputs */
 523                                 : "r" (target), "i" (ASI_INTR_W));
 524                         nack_busy_id++;
 525                         if (nack_busy_id == 32) {
 526                                 need_more = 1;
 527                                 break;
 528                         }
 529                 }
 530         }
 531
 532         /* Now, poll for completion. */
 533         {
 534                 u64 dispatch_stat;
 535                 long stuck;
 536
 537                 stuck = 100000 * nack_busy_id;
 538                 do {
 539                         __asm__ __volatile__("ldxa      [%%g0] %1, %0"
 540                                              : "=r" (dispatch_stat)
 541                                              : "i" (ASI_INTR_DISPATCH_STAT));
 542                         if (dispatch_stat == 0UL) {
 543                                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 544                                                      : : "r" (pstate));
 545                                 if (unlikely(need_more)) {
 546                                         int i, cnt = 0;
 547                                         for_each_cpu_mask(i, mask) {
 548                                                 cpu_clear(i, mask);
 549                                                 cnt++;
 550                                                 if (cnt == 32)
 551                                                         break;
 552                                         }
 553                                         goto retry;
 554                                 }
 555                                 return;
 556                         }
 557                         if (!--stuck)
 558                                 break;
 559                 } while (dispatch_stat & 0x5555555555555555UL);
 560
 561                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 562                                      : : "r" (pstate));
 563
 564                 if ((dispatch_stat & ~(0x5555555555555555UL)) == 0) {
 565                         /* Busy bits will not clear, continue instead
 566                          * of freezing up on this cpu.
 567                          */
 568                         printk("CPU[%d]: mondo stuckage result[%016lx]\n",
 569                                smp_processor_id(), dispatch_stat);
 570                 } else {
 571                         int i, this_busy_nack = 0;
 572
 573                         /* Delay some random time with interrupts enabled
 574                          * to prevent deadlock.
 575                          */
 576                         udelay(2 * nack_busy_id);
 577
 578                         /* Clear out the mask bits for cpus which did not
 579                          * NACK us.
 580                          */
 581                         for_each_cpu_mask(i, mask) {
 582                                 u64 check_mask;
 583
 584                                 if (is_jbus)
 585                                         check_mask = (0x2UL << (2*i));
 586                                 else
 587                                         check_mask = (0x2UL <<
 588                                                       this_busy_nack);
 589                                 if ((dispatch_stat & check_mask) == 0)
 590                                         cpu_clear(i, mask);
 591                                 this_busy_nack += 2;
 592                                 if (this_busy_nack == 64)
 593                                         break;
 594                         }
 595
 596                         goto retry;
 597                 }
 598         }
 599 }
 600
 601 /* Multi-cpu list version.  */
 602 static void hypervisor_xcall_deliver(u64 data0, u64 data1, u64 data2, cpumask_t mask)
 603 {
 604         struct trap_per_cpu *tb;
 605         u16 *cpu_list;
 606         u64 *mondo;
 607         cpumask_t error_mask;
 608         unsigned long flags, status;
 609         int cnt, retries, this_cpu, prev_sent, i;
 610
 611         if (cpus_empty(mask))
 612                 return;
 613
 614         /* We have to do this whole thing with interrupts fully disabled.
 615          * Otherwise if we send an xcall from interrupt context it will
 616          * corrupt both our mondo block and cpu list state.
 617          *
 618          * One consequence of this is that we cannot use timeout mechanisms
 619          * that depend upon interrupts being delivered locally.  So, for
 620          * example, we cannot sample jiffies and expect it to advance.
 621          *
 622          * Fortunately, udelay() uses %stick/%tick so we can use that.
 623          */
 624         local_irq_save(flags);
 625
 626         this_cpu = smp_processor_id();
 627         tb = &trap_block[this_cpu];
 628
 629         mondo = __va(tb->cpu_mondo_block_pa);
 630         mondo[0] = data0;
 631         mondo[1] = data1;
 632         mondo[2] = data2;
 633         wmb();
 634
 635         cpu_list = __va(tb->cpu_list_pa);
 636
 637         /* Setup the initial cpu list.  */
 638         cnt = 0;
 639         for_each_cpu_mask(i, mask)
 640                 cpu_list[cnt++] = i;
 641
 642         cpus_clear(error_mask);
 643         retries = 0;
 644         prev_sent = 0;
 645         do {
 646                 int forward_progress, n_sent;
 647
 648                 status = sun4v_cpu_mondo_send(cnt,
 649                                               tb->cpu_list_pa,
 650                                               tb->cpu_mondo_block_pa);
 651
 652                 /* HV_EOK means all cpus received the xcall, we're done.  */
 653                 if (likely(status == HV_EOK))
 654                         break;
 655
 656                 /* First, see if we made any forward progress.
 657                  *
 658                  * The hypervisor indicates successful sends by setting
 659                  * cpu list entries to the value 0xffff.
 660                  */
 661                 n_sent = 0;
 662                 for (i = 0; i < cnt; i++) {
 663                         if (likely(cpu_list[i] == 0xffff))
 664                                 n_sent++;
 665                 }
 666
 667                 forward_progress = 0;
 668                 if (n_sent > prev_sent)
 669                         forward_progress = 1;
 670
 671                 prev_sent = n_sent;
 672
 673                 /* If we get a HV_ECPUERROR, then one or more of the cpus
 674                  * in the list are in error state.  Use the cpu_state()
 675                  * hypervisor call to find out which cpus are in error state.
 676                  */
 677                 if (unlikely(status == HV_ECPUERROR)) {
 678                         for (i = 0; i < cnt; i++) {
 679                                 long err;
 680                                 u16 cpu;
 681
 682                                 cpu = cpu_list[i];
 683                                 if (cpu == 0xffff)
 684                                         continue;
 685
 686                                 err = sun4v_cpu_state(cpu);
 687                                 if (err >= 0 &&
 688                                     err == HV_CPU_STATE_ERROR) {
 689                                         cpu_list[i] = 0xffff;
 690                                         cpu_set(cpu, error_mask);
 691                                 }
 692                         }
 693                 } else if (unlikely(status != HV_EWOULDBLOCK))
 694                         goto fatal_mondo_error;
 695
 696                 /* Don't bother rewriting the CPU list, just leave the
 697                  * 0xffff and non-0xffff entries in there and the
 698                  * hypervisor will do the right thing.
 699                  *
 700                  * Only advance timeout state if we didn't make any
 701                  * forward progress.
 702                  */
 703                 if (unlikely(!forward_progress)) {
 704                         if (unlikely(++retries > 10000))
 705                                 goto fatal_mondo_timeout;
 706
 707                         /* Delay a little bit to let other cpus catch up
 708                          * on their cpu mondo queue work.
 709                          */
 710                         udelay(2 * cnt);
 711                 }
 712         } while (1);
 713
 714         local_irq_restore(flags);
 715
 716         if (unlikely(!cpus_empty(error_mask)))
 717                 goto fatal_mondo_cpu_error;
 718
 719         return;
 720
 721 fatal_mondo_cpu_error:
 722         printk(KERN_CRIT "CPU[%d]: SUN4V mondo cpu error, some target cpus "
 723                "were in error state\n",
 724                this_cpu);
 725         printk(KERN_CRIT "CPU[%d]: Error mask [ ", this_cpu);
 726         for_each_cpu_mask(i, error_mask)
 727                 printk("%d ", i);
 728         printk("]\n");
 729         return;
 730
 731 fatal_mondo_timeout:
 732         local_irq_restore(flags);
 733         printk(KERN_CRIT "CPU[%d]: SUN4V mondo timeout, no forward "
 734                " progress after %d retries.\n",
 735                this_cpu, retries);
 736         goto dump_cpu_list_and_out;
 737
 738 fatal_mondo_error:
 739         local_irq_restore(flags);
 740         printk(KERN_CRIT "CPU[%d]: Unexpected SUN4V mondo error %lu\n",
 741                this_cpu, status);
 742         printk(KERN_CRIT "CPU[%d]: Args were cnt(%d) cpulist_pa(%lx) "
 743                "mondo_block_pa(%lx)\n",
 744                this_cpu, cnt, tb->cpu_list_pa, tb->cpu_mondo_block_pa);
 745
 746 dump_cpu_list_and_out:
 747         printk(KERN_CRIT "CPU[%d]: CPU list [ ", this_cpu);
 748         for (i = 0; i < cnt; i++)
 749                 printk("%u ", cpu_list[i]);
 750         printk("]\n");
 751 }
 752
 753 /* Send cross call to all processors mentioned in MASK
 754  * except self.
 755  */
 756 static void smp_cross_call_masked(unsigned long *func, u32 ctx, u64 data1, u64 data2, cpumask_t mask)
 757 {
 758         u64 data0 = (((u64)ctx)<<32 | (((u64)func) & 0xffffffff));
 759         int this_cpu = get_cpu();
 760
 761         cpus_and(mask, mask, cpu_online_map);
 762         cpu_clear(this_cpu, mask);
 763
 764         if (tlb_type == spitfire)
 765                 spitfire_xcall_deliver(data0, data1, data2, mask);
 766         else if (tlb_type == cheetah || tlb_type == cheetah_plus)
 767                 cheetah_xcall_deliver(data0, data1, data2, mask);
 768         else
 769                 hypervisor_xcall_deliver(data0, data1, data2, mask);
 770         /* NOTE: Caller runs local copy on master. */
 771
 772         put_cpu();
 773 }
 774
 775 extern unsigned long xcall_sync_tick;
 776
 777 static void smp_start_sync_tick_client(int cpu)
 778 {
 779         cpumask_t mask = cpumask_of_cpu(cpu);
 780
 781         smp_cross_call_masked(&xcall_sync_tick,
 782                               0, 0, 0, mask);
 783 }
 784
 785 /* Send cross call to all processors except self. */
 786 #define smp_cross_call(func, ctx, data1, data2) \
 787         smp_cross_call_masked(func, ctx, data1, data2, cpu_online_map)
 788
 789 struct call_data_struct {
 790         void (*func) (void *info);
 791         void *info;
 792         atomic_t finished;
 793         int wait;
 794 };
 795
 796 static struct call_data_struct *call_data;
 797
 798 extern unsigned long xcall_call_function;
 799
 800 /**
 801  * smp_call_function(): Run a function on all other CPUs.
 802  * @func: The function to run. This must be fast and non-blocking.
 803  * @info: An arbitrary pointer to pass to the function.
 804  * @nonatomic: currently unused.
 805  * @wait: If true, wait (atomically) until function has completed on other CPUs.
 806  *
 807  * Returns 0 on success, else a negative status code. Does not return until
 808  * remote CPUs are nearly ready to execute <<func>> or are or have executed.
 809  *
 810  * You must not call this function with disabled interrupts or from a
 811  * hardware interrupt handler or from a bottom half handler.
 812  */
 813 static int smp_call_function_mask(void (*func)(void *info), void *info,
 814                                   int nonatomic, int wait, cpumask_t mask)
 815 {
 816         struct call_data_struct data;
 817         int cpus;
 818
 819         /* Can deadlock when called with interrupts disabled */
 820         WARN_ON(irqs_disabled());
 821
 822         data.func = func;
 823         data.info = info;
 824         atomic_set(&data.finished, 0);
 825         data.wait = wait;
 826
 827         spin_lock(&call_lock);
 828
 829         cpu_clear(smp_processor_id(), mask);
 830         cpus = cpus_weight(mask);
 831         if (!cpus)
 832                 goto out_unlock;
 833
 834         call_data = &data;
 835         mb();
 836
 837         smp_cross_call_masked(&xcall_call_function, 0, 0, 0, mask);
 838
 839         /* Wait for response */
 840         while (atomic_read(&data.finished) != cpus)
 841                 cpu_relax();
 842
 843 out_unlock:
 844         spin_unlock(&call_lock);
 845
 846         return 0;
 847 }
 848
 849 int smp_call_function(void (*func)(void *info), void *info,
 850                       int nonatomic, int wait)
 851 {
 852         return smp_call_function_mask(func, info, nonatomic, wait,
 853                                       cpu_online_map);
 854 }
 855
 856 void smp_call_function_client(int irq, struct pt_regs *regs)
 857 {
 858         void (*func) (void *info) = call_data->func;
 859         void *info = call_data->info;
 860
 861         clear_softint(1 << irq);
 862         if (call_data->wait) {
 863                 /* let initiator proceed only after completion */
 864                 func(info);
 865                 atomic_inc(&call_data->finished);
 866         } else {
 867                 /* let initiator proceed after getting data */
 868                 atomic_inc(&call_data->finished);
 869                 func(info);
 870         }
 871 }
 872
 873 static void tsb_sync(void *info)
 874 {
 875         struct trap_per_cpu *tp = &trap_block[raw_smp_processor_id()];
 876         struct mm_struct *mm = info;
 877
 878         /* It is not valid to test "currrent->active_mm == mm" here.
 879          *
 880          * The value of "current" is not changed atomically with
 881          * switch_mm().  But that's OK, we just need to check the
 882          * current cpu's trap block PGD physical address.
 883          */
 884         if (tp->pgd_paddr == __pa(mm->pgd))
 885                 tsb_context_switch(mm);
 886 }
 887
 888 void smp_tsb_sync(struct mm_struct *mm)
 889 {
 890         smp_call_function_mask(tsb_sync, mm, 0, 1, mm->cpu_vm_mask);
 891 }
 892
 893 extern unsigned long xcall_flush_tlb_mm;
 894 extern unsigned long xcall_flush_tlb_pending;
 895 extern unsigned long xcall_flush_tlb_kernel_range;
 896 extern unsigned long xcall_report_regs;
 897 extern unsigned long xcall_receive_signal;
 898 extern unsigned long xcall_new_mmu_context_version;
 899
 900 #ifdef DCACHE_ALIASING_POSSIBLE
 901 extern unsigned long xcall_flush_dcache_page_cheetah;
 902 #endif
 903 extern unsigned long xcall_flush_dcache_page_spitfire;
 904
 905 #ifdef CONFIG_DEBUG_DCFLUSH
 906 extern atomic_t dcpage_flushes;
 907 extern atomic_t dcpage_flushes_xcall;
 908 #endif
 909
 910 static __inline__ void __local_flush_dcache_page(struct page *page)
 911 {
 912 #ifdef DCACHE_ALIASING_POSSIBLE
 913         __flush_dcache_page(page_address(page),
 914                             ((tlb_type == spitfire) &&
 915                              page_mapping(page) != NULL));
 916 #else
 917         if (page_mapping(page) != NULL &&
 918             tlb_type == spitfire)
 919                 __flush_icache_page(__pa(page_address(page)));
 920 #endif
 921 }
 922
 923 void smp_flush_dcache_page_impl(struct page *page, int cpu)
 924 {
 925         cpumask_t mask = cpumask_of_cpu(cpu);
 926         int this_cpu;
 927
 928         if (tlb_type == hypervisor)
 929                 return;
 930
 931 #ifdef CONFIG_DEBUG_DCFLUSH
 932         atomic_inc(&dcpage_flushes);
 933 #endif
 934
 935         this_cpu = get_cpu();
 936
 937         if (cpu == this_cpu) {
 938                 __local_flush_dcache_page(page);
 939         } else if (cpu_online(cpu)) {
 940                 void *pg_addr = page_address(page);
 941                 u64 data0;
 942
 943                 if (tlb_type == spitfire) {
 944                         data0 =
 945                                 ((u64)&xcall_flush_dcache_page_spitfire);
 946                         if (page_mapping(page) != NULL)
 947                                 data0 |= ((u64)1 << 32);
 948                         spitfire_xcall_deliver(data0,
 949                                                __pa(pg_addr),
 950                                                (u64) pg_addr,
 951                                                mask);
 952                 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 953 #ifdef DCACHE_ALIASING_POSSIBLE
 954                         data0 =
 955                                 ((u64)&xcall_flush_dcache_page_cheetah);
 956                         cheetah_xcall_deliver(data0,
 957                                               __pa(pg_addr),
 958                                               0, mask);
 959 #endif
 960                 }
 961 #ifdef CONFIG_DEBUG_DCFLUSH
 962                 atomic_inc(&dcpage_flushes_xcall);
 963 #endif
 964         }
 965
 966         put_cpu();
 967 }
 968
 969 void flush_dcache_page_all(struct mm_struct *mm, struct page *page)
 970 {
 971         void *pg_addr = page_address(page);
 972         cpumask_t mask = cpu_online_map;
 973         u64 data0;
 974         int this_cpu;
 975
 976         if (tlb_type == hypervisor)
 977                 return;
 978
 979         this_cpu = get_cpu();
 980
 981         cpu_clear(this_cpu, mask);
 982
 983 #ifdef CONFIG_DEBUG_DCFLUSH
 984         atomic_inc(&dcpage_flushes);
 985 #endif
 986         if (cpus_empty(mask))
 987                 goto flush_self;
 988         if (tlb_type == spitfire) {
 989                 data0 = ((u64)&xcall_flush_dcache_page_spitfire);
 990                 if (page_mapping(page) != NULL)
 991                         data0 |= ((u64)1 << 32);
 992                 spitfire_xcall_deliver(data0,
 993                                        __pa(pg_addr),
 994                                        (u64) pg_addr,
 995                                        mask);
 996         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 997 #ifdef DCACHE_ALIASING_POSSIBLE
 998                 data0 = ((u64)&xcall_flush_dcache_page_cheetah);
 999                 cheetah_xcall_deliver(data0,
1000                                       __pa(pg_addr),
1001                                       0, mask);
1002 #endif
1003         }
1004 #ifdef CONFIG_DEBUG_DCFLUSH
1005         atomic_inc(&dcpage_flushes_xcall);
1006 #endif
1007  flush_self:
1008         __local_flush_dcache_page(page);
1009
1010         put_cpu();
1011 }
1012
1013 static void __smp_receive_signal_mask(cpumask_t mask)
1014 {
1015         smp_cross_call_masked(&xcall_receive_signal, 0, 0, 0, mask);
1016 }
1017
1018 void smp_receive_signal(int cpu)
1019 {
1020         cpumask_t mask = cpumask_of_cpu(cpu);
1021
1022         if (cpu_online(cpu))
1023                 __smp_receive_signal_mask(mask);
1024 }
1025
1026 void smp_receive_signal_client(int irq, struct pt_regs *regs)
1027 {
1028         clear_softint(1 << irq);
1029 }
1030
1031 void smp_new_mmu_context_version_client(int irq, struct pt_regs *regs)
1032 {
1033         struct mm_struct *mm;
1034         unsigned long flags;
1035
1036         clear_softint(1 << irq);
1037
1038         /* See if we need to allocate a new TLB context because
1039          * the version of the one we are using is now out of date.
1040          */
1041         mm = current->active_mm;
1042         if (unlikely(!mm || (mm == &init_mm)))
1043                 return;
1044
1045         spin_lock_irqsave(&mm->context.lock, flags);
1046
1047         if (unlikely(!CTX_VALID(mm->context)))
1048                 get_new_mmu_context(mm);
1049
1050         spin_unlock_irqrestore(&mm->context.lock, flags);
1051
1052         load_secondary_context(mm);
1053         __flush_tlb_mm(CTX_HWBITS(mm->context),
1054                        SECONDARY_CONTEXT);
1055 }
1056
1057 void smp_new_mmu_context_version(void)
1058 {
1059         smp_cross_call(&xcall_new_mmu_context_version, 0, 0, 0);
1060 }
1061
1062 void smp_report_regs(void)
1063 {
1064         smp_cross_call(&xcall_report_regs, 0, 0, 0);
1065 }
1066
1067 /* We know that the window frames of the user have been flushed
1068  * to the stack before we get here because all callers of us
1069  * are flush_tlb_*() routines, and these run after flush_cache_*()
1070  * which performs the flushw.
1071  *
1072  * The SMP TLB coherency scheme we use works as follows:
1073  *
1074  * 1) mm->cpu_vm_mask is a bit mask of which cpus an address
1075  *    space has (potentially) executed on, this is the heuristic
1076  *    we use to avoid doing cross calls.
1077  *
1078  *    Also, for flushing from kswapd and also for clones, we
1079  *    use cpu_vm_mask as the list of cpus to make run the TLB.
1080  *
1081  * 2) TLB context numbers are shared globally across all processors
1082  *    in the system, this allows us to play several games to avoid
1083  *    cross calls.
1084  *
1085  *    One invariant is that when a cpu switches to a process, and
1086  *    that processes tsk->active_mm->cpu_vm_mask does not have the
1087  *    current cpu's bit set, that tlb context is flushed locally.
1088  *
1089  *    If the address space is non-shared (ie. mm->count == 1) we avoid
1090  *    cross calls when we want to flush the currently running process's
1091  *    tlb state.  This is done by clearing all cpu bits except the current
1092  *    processor's in current->active_mm->cpu_vm_mask and performing the
1093  *    flush locally only.  This will force any subsequent cpus which run
1094  *    this task to flush the context from the local tlb if the process
1095  *    migrates to another cpu (again).
1096  *
1097  * 3) For shared address spaces (threads) and swapping we bite the
1098  *    bullet for most cases and perform the cross call (but only to
1099  *    the cpus listed in cpu_vm_mask).
1100  *
1101  *    The performance gain from "optimizing" away the cross call for threads is
1102  *    questionable (in theory the big win for threads is the massive sharing of
1103  *    address space state across processors).
1104  */
1105
1106 /* This currently is only used by the hugetlb arch pre-fault
1107  * hook on UltraSPARC-III+ and later when changing the pagesize
1108  * bits of the context register for an address space.
1109  */
1110 void smp_flush_tlb_mm(struct mm_struct *mm)
1111 {
1112         u32 ctx = CTX_HWBITS(mm->context);
1113         int cpu = get_cpu();
1114
1115         if (atomic_read(&mm->mm_users) == 1) {
1116                 mm->cpu_vm_mask = cpumask_of_cpu(cpu);
1117                 goto local_flush_and_out;
1118         }
1119
1120         smp_cross_call_masked(&xcall_flush_tlb_mm,
1121                               ctx, 0, 0,
1122                               mm->cpu_vm_mask);
1123
1124 local_flush_and_out:
1125         __flush_tlb_mm(ctx, SECONDARY_CONTEXT);
1126
1127         put_cpu();
1128 }
1129
1130 void smp_flush_tlb_pending(struct mm_struct *mm, unsigned long nr, unsigned long *vaddrs)
1131 {
1132         u32 ctx = CTX_HWBITS(mm->context);
1133         int cpu = get_cpu();
1134
1135         if (mm == current->active_mm && atomic_read(&mm->mm_users) == 1)
1136                 mm->cpu_vm_mask = cpumask_of_cpu(cpu);
1137         else
1138                 smp_cross_call_masked(&xcall_flush_tlb_pending,
1139                                       ctx, nr, (unsigned long) vaddrs,
1140                                       mm->cpu_vm_mask);
1141
1142         __flush_tlb_pending(ctx, nr, vaddrs);
1143
1144         put_cpu();
1145 }
1146
1147 void smp_flush_tlb_kernel_range(unsigned long start, unsigned long end)
1148 {
1149         start &= PAGE_MASK;
1150         end    = PAGE_ALIGN(end);
1151         if (start != end) {
1152                 smp_cross_call(&xcall_flush_tlb_kernel_range,
1153                                0, start, end);
1154
1155                 __flush_tlb_kernel_range(start, end);
1156         }
1157 }
1158
1159 /* CPU capture. */
1160 /* #define CAPTURE_DEBUG */
1161 extern unsigned long xcall_capture;
1162
1163 static atomic_t smp_capture_depth = ATOMIC_INIT(0);
1164 static atomic_t smp_capture_registry = ATOMIC_INIT(0);
1165 static unsigned long penguins_are_doing_time;
1166
1167 void smp_capture(void)
1168 {
1169         int result = atomic_add_ret(1, &smp_capture_depth);
1170
1171         if (result == 1) {
1172                 int ncpus = num_online_cpus();
1173
1174 #ifdef CAPTURE_DEBUG
1175                 printk("CPU[%d]: Sending penguins to jail...",
1176                        smp_processor_id());
1177 #endif
1178                 penguins_are_doing_time = 1;
1179                 membar_storestore_loadstore();
1180                 atomic_inc(&smp_capture_registry);
1181                 smp_cross_call(&xcall_capture, 0, 0, 0);
1182                 while (atomic_read(&smp_capture_registry) != ncpus)
1183                         rmb();
1184 #ifdef CAPTURE_DEBUG
1185                 printk("done\n");
1186 #endif
1187         }
1188 }
1189
1190 void smp_release(void)
1191 {
1192         if (atomic_dec_and_test(&smp_capture_depth)) {
1193 #ifdef CAPTURE_DEBUG
1194                 printk("CPU[%d]: Giving pardon to "
1195                        "imprisoned penguins\n",
1196                        smp_processor_id());
1197 #endif
1198                 penguins_are_doing_time = 0;
1199                 membar_storeload_storestore();
1200                 atomic_dec(&smp_capture_registry);
1201         }
1202 }
1203
1204 /* Imprisoned penguins run with %pil == 15, but PSTATE_IE set, so they
1205  * can service tlb flush xcalls...
1206  */
1207 extern void prom_world(int);
1208
1209 void smp_penguin_jailcell(int irq, struct pt_regs *regs)
1210 {
1211         clear_softint(1 << irq);
1212
1213         preempt_disable();
1214
1215         __asm__ __volatile__("flushw");
1216         prom_world(1);
1217         atomic_inc(&smp_capture_registry);
1218         membar_storeload_storestore();
1219         while (penguins_are_doing_time)
1220                 rmb();
1221         atomic_dec(&smp_capture_registry);
1222         prom_world(0);
1223
1224         preempt_enable();
1225 }
1226
1227 /* /proc/profile writes can call this, don't __init it please. */
1228 int setup_profiling_timer(unsigned int multiplier)
1229 {
1230         return -EINVAL;
1231 }
1232
1233 void __init smp_prepare_cpus(unsigned int max_cpus)
1234 {
1235 }
1236
1237 void __devinit smp_prepare_boot_cpu(void)
1238 {
1239 }
1240
1241 void __devinit smp_fill_in_sib_core_maps(void)
1242 {
1243         unsigned int i;
1244
1245         for_each_present_cpu(i) {
1246                 unsigned int j;
1247
1248                 cpus_clear(cpu_core_map[i]);
1249                 if (cpu_data(i).core_id == 0) {
1250                         cpu_set(i, cpu_core_map[i]);
1251                         continue;
1252                 }
1253
1254                 for_each_present_cpu(j) {
1255                         if (cpu_data(i).core_id ==
1256                             cpu_data(j).core_id)
1257                                 cpu_set(j, cpu_core_map[i]);
1258                 }
1259         }
1260
1261         for_each_present_cpu(i) {
1262                 unsigned int j;
1263
1264                 cpus_clear(cpu_sibling_map[i]);
1265                 if (cpu_data(i).proc_id == -1) {
1266                         cpu_set(i, cpu_sibling_map[i]);
1267                         continue;
1268                 }
1269
1270                 for_each_present_cpu(j) {
1271                         if (cpu_data(i).proc_id ==
1272                             cpu_data(j).proc_id)
1273                                 cpu_set(j, cpu_sibling_map[i]);
1274                 }
1275         }
1276 }
1277
1278 int __cpuinit __cpu_up(unsigned int cpu)
1279 {
1280         int ret = smp_boot_one_cpu(cpu);
1281
1282         if (!ret) {
1283                 cpu_set(cpu, smp_commenced_mask);
1284                 while (!cpu_isset(cpu, cpu_online_map))
1285                         mb();
1286                 if (!cpu_isset(cpu, cpu_online_map)) {
1287                         ret = -ENODEV;
1288                 } else {
1289                         /* On SUN4V, writes to %tick and %stick are
1290                          * not allowed.
1291                          */
1292                         if (tlb_type != hypervisor)
1293                                 smp_synchronize_one_tick(cpu);
1294                 }
1295         }
1296         return ret;
1297 }
1298
1299 #ifdef CONFIG_HOTPLUG_CPU
1300 void cpu_play_dead(void)
1301 {
1302         int cpu = smp_processor_id();
1303         unsigned long pstate;
1304
1305         idle_task_exit();
1306
1307         if (tlb_type == hypervisor) {
1308                 struct trap_per_cpu *tb = &trap_block[cpu];
1309
1310                 sun4v_cpu_qconf(HV_CPU_QUEUE_CPU_MONDO,
1311                                 tb->cpu_mondo_pa, 0);
1312                 sun4v_cpu_qconf(HV_CPU_QUEUE_DEVICE_MONDO,
1313                                 tb->dev_mondo_pa, 0);
1314                 sun4v_cpu_qconf(HV_CPU_QUEUE_RES_ERROR,
1315                                 tb->resum_mondo_pa, 0);
1316                 sun4v_cpu_qconf(HV_CPU_QUEUE_NONRES_ERROR,
1317                                 tb->nonresum_mondo_pa, 0);
1318         }
1319
1320         cpu_clear(cpu, smp_commenced_mask);
1321         membar_safe("#Sync");
1322
1323         local_irq_disable();
1324
1325         __asm__ __volatile__(
1326                 "rdpr   %%pstate, %0\n\t"
1327                 "wrpr   %0, %1, %%pstate"
1328                 : "=r" (pstate)
1329                 : "i" (PSTATE_IE));
1330
1331         while (1)
1332                 barrier();
1333 }
1334
1335 int __cpu_disable(void)
1336 {
1337         int cpu = smp_processor_id();
1338         cpuinfo_sparc *c;
1339         int i;
1340
1341         for_each_cpu_mask(i, cpu_core_map[cpu])
1342                 cpu_clear(cpu, cpu_core_map[i]);
1343         cpus_clear(cpu_core_map[cpu]);
1344
1345         for_each_cpu_mask(i, cpu_sibling_map[cpu])
1346                 cpu_clear(cpu, cpu_sibling_map[i]);
1347         cpus_clear(cpu_sibling_map[cpu]);
1348
1349         c = &cpu_data(cpu);
1350
1351         c->core_id = 0;
1352         c->proc_id = -1;
1353
1354         spin_lock(&call_lock);
1355         cpu_clear(cpu, cpu_online_map);
1356         spin_unlock(&call_lock);
1357
1358         smp_wmb();
1359
1360         /* Make sure no interrupts point to this cpu.  */
1361         fixup_irqs();
1362
1363         local_irq_enable();
1364         mdelay(1);
1365         local_irq_disable();
1366
1367         return 0;
1368 }
1369
1370 void __cpu_die(unsigned int cpu)
1371 {
1372         int i;
1373
1374         for (i = 0; i < 100; i++) {
1375                 smp_rmb();
1376                 if (!cpu_isset(cpu, smp_commenced_mask))
1377                         break;
1378                 msleep(100);
1379         }
1380         if (cpu_isset(cpu, smp_commenced_mask)) {
1381                 printk(KERN_ERR "CPU %u didn't die...\n", cpu);
1382         } else {
1383 #if defined(CONFIG_SUN_LDOMS)
1384                 unsigned long hv_err;
1385                 int limit = 100;
1386
1387                 do {
1388                         hv_err = sun4v_cpu_stop(cpu);
1389                         if (hv_err == HV_EOK) {
1390                                 cpu_clear(cpu, cpu_present_map);
1391                                 break;
1392                         }
1393                 } while (--limit > 0);
1394                 if (limit <= 0) {
1395                         printk(KERN_ERR "sun4v_cpu_stop() fails err=%lu\n",
1396                                hv_err);
1397                 }
1398 #endif
1399         }
1400 }
1401 #endif
1402
1403 void __init smp_cpus_done(unsigned int max_cpus)
1404 {
1405 }
1406
1407 void smp_send_reschedule(int cpu)
1408 {
1409         smp_receive_signal(cpu);
1410 }
1411
1412 /* This is a nop because we capture all other cpus
1413  * anyways when making the PROM active.
1414  */
1415 void smp_send_stop(void)
1416 {
1417 }
1418
1419 unsigned long __per_cpu_base __read_mostly;
1420 unsigned long __per_cpu_shift __read_mostly;
1421
1422 EXPORT_SYMBOL(__per_cpu_base);
1423 EXPORT_SYMBOL(__per_cpu_shift);
1424
1425 void __init real_setup_per_cpu_areas(void)
1426 {
1427         unsigned long goal, size, i;
1428         char *ptr;
1429
1430         /* Copy section for each CPU (we discard the original) */
1431         goal = PERCPU_ENOUGH_ROOM;
1432
1433         __per_cpu_shift = PAGE_SHIFT;
1434         for (size = PAGE_SIZE; size < goal; size <<= 1UL)
1435                 __per_cpu_shift++;
1436
1437         ptr = alloc_bootmem_pages(size * NR_CPUS);
1438
1439         __per_cpu_base = ptr - __per_cpu_start;
1440
1441         for (i = 0; i < NR_CPUS; i++, ptr += size)
1442                 memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start);
1443
1444         /* Setup %g5 for the boot cpu.  */
1445         __local_per_cpu_offset = __per_cpu_offset(smp_processor_id());
1446 }