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 void sun4v_init_mondo_queues(int use_bootmem, int cpu, int alloc, int load);
 338
 339 extern unsigned long sparc64_cpu_startup;
 340
 341 /* The OBP cpu startup callback truncates the 3rd arg cookie to
 342  * 32-bits (I think) so to be safe we have it read the pointer
 343  * contained here so we work on >4GB machines. -DaveM
 344  */
 345 static struct thread_info *cpu_new_thread = NULL;
 346
 347 static int __devinit smp_boot_one_cpu(unsigned int cpu)
 348 {
 349         struct trap_per_cpu *tb = &trap_block[cpu];
 350         unsigned long entry =
 351                 (unsigned long)(&sparc64_cpu_startup);
 352         unsigned long cookie =
 353                 (unsigned long)(&cpu_new_thread);
 354         struct task_struct *p;
 355         int timeout, ret;
 356
 357         p = fork_idle(cpu);
 358         callin_flag = 0;
 359         cpu_new_thread = task_thread_info(p);
 360
 361         if (tlb_type == hypervisor) {
 362                 /* Alloc the mondo queues, cpu will load them.  */
 363                 sun4v_init_mondo_queues(0, cpu, 1, 0);
 364
 365 #if defined(CONFIG_SUN_LDOMS) && defined(CONFIG_HOTPLUG_CPU)
 366                 if (ldom_domaining_enabled)
 367                         ldom_startcpu_cpuid(cpu,
 368                                             (unsigned long) cpu_new_thread);
 369                 else
 370 #endif
 371                         prom_startcpu_cpuid(cpu, entry, cookie);
 372         } else {
 373                 struct device_node *dp = of_find_node_by_cpuid(cpu);
 374
 375                 prom_startcpu(dp->node, entry, cookie);
 376         }
 377
 378         for (timeout = 0; timeout < 50000; timeout++) {
 379                 if (callin_flag)
 380                         break;
 381                 udelay(100);
 382         }
 383
 384         if (callin_flag) {
 385                 ret = 0;
 386         } else {
 387                 printk("Processor %d is stuck.\n", cpu);
 388                 ret = -ENODEV;
 389         }
 390         cpu_new_thread = NULL;
 391
 392         if (tb->hdesc) {
 393                 kfree(tb->hdesc);
 394                 tb->hdesc = NULL;
 395         }
 396
 397         return ret;
 398 }
 399
 400 static void spitfire_xcall_helper(u64 data0, u64 data1, u64 data2, u64 pstate, unsigned long cpu)
 401 {
 402         u64 result, target;
 403         int stuck, tmp;
 404
 405         if (this_is_starfire) {
 406                 /* map to real upaid */
 407                 cpu = (((cpu & 0x3c) << 1) |
 408                         ((cpu & 0x40) >> 4) |
 409                         (cpu & 0x3));
 410         }
 411
 412         target = (cpu << 14) | 0x70;
 413 again:
 414         /* Ok, this is the real Spitfire Errata #54.
 415          * One must read back from a UDB internal register
 416          * after writes to the UDB interrupt dispatch, but
 417          * before the membar Sync for that write.
 418          * So we use the high UDB control register (ASI 0x7f,
 419          * ADDR 0x20) for the dummy read. -DaveM
 420          */
 421         tmp = 0x40;
 422         __asm__ __volatile__(
 423         "wrpr   %1, %2, %%pstate\n\t"
 424         "stxa   %4, [%0] %3\n\t"
 425         "stxa   %5, [%0+%8] %3\n\t"
 426         "add    %0, %8, %0\n\t"
 427         "stxa   %6, [%0+%8] %3\n\t"
 428         "membar #Sync\n\t"
 429         "stxa   %%g0, [%7] %3\n\t"
 430         "membar #Sync\n\t"
 431         "mov    0x20, %%g1\n\t"
 432         "ldxa   [%%g1] 0x7f, %%g0\n\t"
 433         "membar #Sync"
 434         : "=r" (tmp)
 435         : "r" (pstate), "i" (PSTATE_IE), "i" (ASI_INTR_W),
 436           "r" (data0), "r" (data1), "r" (data2), "r" (target),
 437           "r" (0x10), "0" (tmp)
 438         : "g1");
 439
 440         /* NOTE: PSTATE_IE is still clear. */
 441         stuck = 100000;
 442         do {
 443                 __asm__ __volatile__("ldxa [%%g0] %1, %0"
 444                         : "=r" (result)
 445                         : "i" (ASI_INTR_DISPATCH_STAT));
 446                 if (result == 0) {
 447                         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 448                                              : : "r" (pstate));
 449                         return;
 450                 }
 451                 stuck -= 1;
 452                 if (stuck == 0)
 453                         break;
 454         } while (result & 0x1);
 455         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 456                              : : "r" (pstate));
 457         if (stuck == 0) {
 458                 printk("CPU[%d]: mondo stuckage result[%016lx]\n",
 459                        smp_processor_id(), result);
 460         } else {
 461                 udelay(2);
 462                 goto again;
 463         }
 464 }
 465
 466 static __inline__ void spitfire_xcall_deliver(u64 data0, u64 data1, u64 data2, cpumask_t mask)
 467 {
 468         u64 pstate;
 469         int i;
 470
 471         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 472         for_each_cpu_mask(i, mask)
 473                 spitfire_xcall_helper(data0, data1, data2, pstate, i);
 474 }
 475
 476 /* Cheetah now allows to send the whole 64-bytes of data in the interrupt
 477  * packet, but we have no use for that.  However we do take advantage of
 478  * the new pipelining feature (ie. dispatch to multiple cpus simultaneously).
 479  */
 480 static void cheetah_xcall_deliver(u64 data0, u64 data1, u64 data2, cpumask_t mask)
 481 {
 482         u64 pstate, ver;
 483         int nack_busy_id, is_jbus, need_more;
 484
 485         if (cpus_empty(mask))
 486                 return;
 487
 488         /* Unfortunately, someone at Sun had the brilliant idea to make the
 489          * busy/nack fields hard-coded by ITID number for this Ultra-III
 490          * derivative processor.
 491          */
 492         __asm__ ("rdpr %%ver, %0" : "=r" (ver));
 493         is_jbus = ((ver >> 32) == __JALAPENO_ID ||
 494                    (ver >> 32) == __SERRANO_ID);
 495
 496         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 497
 498 retry:
 499         need_more = 0;
 500         __asm__ __volatile__("wrpr %0, %1, %%pstate\n\t"
 501                              : : "r" (pstate), "i" (PSTATE_IE));
 502
 503         /* Setup the dispatch data registers. */
 504         __asm__ __volatile__("stxa      %0, [%3] %6\n\t"
 505                              "stxa      %1, [%4] %6\n\t"
 506                              "stxa      %2, [%5] %6\n\t"
 507                              "membar    #Sync\n\t"
 508                              : /* no outputs */
 509                              : "r" (data0), "r" (data1), "r" (data2),
 510                                "r" (0x40), "r" (0x50), "r" (0x60),
 511                                "i" (ASI_INTR_W));
 512
 513         nack_busy_id = 0;
 514         {
 515                 int i;
 516
 517                 for_each_cpu_mask(i, mask) {
 518                         u64 target = (i << 14) | 0x70;
 519
 520                         if (!is_jbus)
 521                                 target |= (nack_busy_id << 24);
 522                         __asm__ __volatile__(
 523                                 "stxa   %%g0, [%0] %1\n\t"
 524                                 "membar #Sync\n\t"
 525                                 : /* no outputs */
 526                                 : "r" (target), "i" (ASI_INTR_W));
 527                         nack_busy_id++;
 528                         if (nack_busy_id == 32) {
 529                                 need_more = 1;
 530                                 break;
 531                         }
 532                 }
 533         }
 534
 535         /* Now, poll for completion. */
 536         {
 537                 u64 dispatch_stat;
 538                 long stuck;
 539
 540                 stuck = 100000 * nack_busy_id;
 541                 do {
 542                         __asm__ __volatile__("ldxa      [%%g0] %1, %0"
 543                                              : "=r" (dispatch_stat)
 544                                              : "i" (ASI_INTR_DISPATCH_STAT));
 545                         if (dispatch_stat == 0UL) {
 546                                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 547                                                      : : "r" (pstate));
 548                                 if (unlikely(need_more)) {
 549                                         int i, cnt = 0;
 550                                         for_each_cpu_mask(i, mask) {
 551                                                 cpu_clear(i, mask);
 552                                                 cnt++;
 553                                                 if (cnt == 32)
 554                                                         break;
 555                                         }
 556                                         goto retry;
 557                                 }
 558                                 return;
 559                         }
 560                         if (!--stuck)
 561                                 break;
 562                 } while (dispatch_stat & 0x5555555555555555UL);
 563
 564                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 565                                      : : "r" (pstate));
 566
 567                 if ((dispatch_stat & ~(0x5555555555555555UL)) == 0) {
 568                         /* Busy bits will not clear, continue instead
 569                          * of freezing up on this cpu.
 570                          */
 571                         printk("CPU[%d]: mondo stuckage result[%016lx]\n",
 572                                smp_processor_id(), dispatch_stat);
 573                 } else {
 574                         int i, this_busy_nack = 0;
 575
 576                         /* Delay some random time with interrupts enabled
 577                          * to prevent deadlock.
 578                          */
 579                         udelay(2 * nack_busy_id);
 580
 581                         /* Clear out the mask bits for cpus which did not
 582                          * NACK us.
 583                          */
 584                         for_each_cpu_mask(i, mask) {
 585                                 u64 check_mask;
 586
 587                                 if (is_jbus)
 588                                         check_mask = (0x2UL << (2*i));
 589                                 else
 590                                         check_mask = (0x2UL <<
 591                                                       this_busy_nack);
 592                                 if ((dispatch_stat & check_mask) == 0)
 593                                         cpu_clear(i, mask);
 594                                 this_busy_nack += 2;
 595                                 if (this_busy_nack == 64)
 596                                         break;
 597                         }
 598
 599                         goto retry;
 600                 }
 601         }
 602 }
 603
 604 /* Multi-cpu list version.  */
 605 static void hypervisor_xcall_deliver(u64 data0, u64 data1, u64 data2, cpumask_t mask)
 606 {
 607         struct trap_per_cpu *tb;
 608         u16 *cpu_list;
 609         u64 *mondo;
 610         cpumask_t error_mask;
 611         unsigned long flags, status;
 612         int cnt, retries, this_cpu, prev_sent, i;
 613
 614         if (cpus_empty(mask))
 615                 return;
 616
 617         /* We have to do this whole thing with interrupts fully disabled.
 618          * Otherwise if we send an xcall from interrupt context it will
 619          * corrupt both our mondo block and cpu list state.
 620          *
 621          * One consequence of this is that we cannot use timeout mechanisms
 622          * that depend upon interrupts being delivered locally.  So, for
 623          * example, we cannot sample jiffies and expect it to advance.
 624          *
 625          * Fortunately, udelay() uses %stick/%tick so we can use that.
 626          */
 627         local_irq_save(flags);
 628
 629         this_cpu = smp_processor_id();
 630         tb = &trap_block[this_cpu];
 631
 632         mondo = __va(tb->cpu_mondo_block_pa);
 633         mondo[0] = data0;
 634         mondo[1] = data1;
 635         mondo[2] = data2;
 636         wmb();
 637
 638         cpu_list = __va(tb->cpu_list_pa);
 639
 640         /* Setup the initial cpu list.  */
 641         cnt = 0;
 642         for_each_cpu_mask(i, mask)
 643                 cpu_list[cnt++] = i;
 644
 645         cpus_clear(error_mask);
 646         retries = 0;
 647         prev_sent = 0;
 648         do {
 649                 int forward_progress, n_sent;
 650
 651                 status = sun4v_cpu_mondo_send(cnt,
 652                                               tb->cpu_list_pa,
 653                                               tb->cpu_mondo_block_pa);
 654
 655                 /* HV_EOK means all cpus received the xcall, we're done.  */
 656                 if (likely(status == HV_EOK))
 657                         break;
 658
 659                 /* First, see if we made any forward progress.
 660                  *
 661                  * The hypervisor indicates successful sends by setting
 662                  * cpu list entries to the value 0xffff.
 663                  */
 664                 n_sent = 0;
 665                 for (i = 0; i < cnt; i++) {
 666                         if (likely(cpu_list[i] == 0xffff))
 667                                 n_sent++;
 668                 }
 669
 670                 forward_progress = 0;
 671                 if (n_sent > prev_sent)
 672                         forward_progress = 1;
 673
 674                 prev_sent = n_sent;
 675
 676                 /* If we get a HV_ECPUERROR, then one or more of the cpus
 677                  * in the list are in error state.  Use the cpu_state()
 678                  * hypervisor call to find out which cpus are in error state.
 679                  */
 680                 if (unlikely(status == HV_ECPUERROR)) {
 681                         for (i = 0; i < cnt; i++) {
 682                                 long err;
 683                                 u16 cpu;
 684
 685                                 cpu = cpu_list[i];
 686                                 if (cpu == 0xffff)
 687                                         continue;
 688
 689                                 err = sun4v_cpu_state(cpu);
 690                                 if (err >= 0 &&
 691                                     err == HV_CPU_STATE_ERROR) {
 692                                         cpu_list[i] = 0xffff;
 693                                         cpu_set(cpu, error_mask);
 694                                 }
 695                         }
 696                 } else if (unlikely(status != HV_EWOULDBLOCK))
 697                         goto fatal_mondo_error;
 698
 699                 /* Don't bother rewriting the CPU list, just leave the
 700                  * 0xffff and non-0xffff entries in there and the
 701                  * hypervisor will do the right thing.
 702                  *
 703                  * Only advance timeout state if we didn't make any
 704                  * forward progress.
 705                  */
 706                 if (unlikely(!forward_progress)) {
 707                         if (unlikely(++retries > 10000))
 708                                 goto fatal_mondo_timeout;
 709
 710                         /* Delay a little bit to let other cpus catch up
 711                          * on their cpu mondo queue work.
 712                          */
 713                         udelay(2 * cnt);
 714                 }
 715         } while (1);
 716
 717         local_irq_restore(flags);
 718
 719         if (unlikely(!cpus_empty(error_mask)))
 720                 goto fatal_mondo_cpu_error;
 721
 722         return;
 723
 724 fatal_mondo_cpu_error:
 725         printk(KERN_CRIT "CPU[%d]: SUN4V mondo cpu error, some target cpus "
 726                "were in error state\n",
 727                this_cpu);
 728         printk(KERN_CRIT "CPU[%d]: Error mask [ ", this_cpu);
 729         for_each_cpu_mask(i, error_mask)
 730                 printk("%d ", i);
 731         printk("]\n");
 732         return;
 733
 734 fatal_mondo_timeout:
 735         local_irq_restore(flags);
 736         printk(KERN_CRIT "CPU[%d]: SUN4V mondo timeout, no forward "
 737                " progress after %d retries.\n",
 738                this_cpu, retries);
 739         goto dump_cpu_list_and_out;
 740
 741 fatal_mondo_error:
 742         local_irq_restore(flags);
 743         printk(KERN_CRIT "CPU[%d]: Unexpected SUN4V mondo error %lu\n",
 744                this_cpu, status);
 745         printk(KERN_CRIT "CPU[%d]: Args were cnt(%d) cpulist_pa(%lx) "
 746                "mondo_block_pa(%lx)\n",
 747                this_cpu, cnt, tb->cpu_list_pa, tb->cpu_mondo_block_pa);
 748
 749 dump_cpu_list_and_out:
 750         printk(KERN_CRIT "CPU[%d]: CPU list [ ", this_cpu);
 751         for (i = 0; i < cnt; i++)
 752                 printk("%u ", cpu_list[i]);
 753         printk("]\n");
 754 }
 755
 756 /* Send cross call to all processors mentioned in MASK
 757  * except self.
 758  */
 759 static void smp_cross_call_masked(unsigned long *func, u32 ctx, u64 data1, u64 data2, cpumask_t mask)
 760 {
 761         u64 data0 = (((u64)ctx)<<32 | (((u64)func) & 0xffffffff));
 762         int this_cpu = get_cpu();
 763
 764         cpus_and(mask, mask, cpu_online_map);
 765         cpu_clear(this_cpu, mask);
 766
 767         if (tlb_type == spitfire)
 768                 spitfire_xcall_deliver(data0, data1, data2, mask);
 769         else if (tlb_type == cheetah || tlb_type == cheetah_plus)
 770                 cheetah_xcall_deliver(data0, data1, data2, mask);
 771         else
 772                 hypervisor_xcall_deliver(data0, data1, data2, mask);
 773         /* NOTE: Caller runs local copy on master. */
 774
 775         put_cpu();
 776 }
 777
 778 extern unsigned long xcall_sync_tick;
 779
 780 static void smp_start_sync_tick_client(int cpu)
 781 {
 782         cpumask_t mask = cpumask_of_cpu(cpu);
 783
 784         smp_cross_call_masked(&xcall_sync_tick,
 785                               0, 0, 0, mask);
 786 }
 787
 788 /* Send cross call to all processors except self. */
 789 #define smp_cross_call(func, ctx, data1, data2) \
 790         smp_cross_call_masked(func, ctx, data1, data2, cpu_online_map)
 791
 792 struct call_data_struct {
 793         void (*func) (void *info);
 794         void *info;
 795         atomic_t finished;
 796         int wait;
 797 };
 798
 799 static struct call_data_struct *call_data;
 800
 801 extern unsigned long xcall_call_function;
 802
 803 /**
 804  * smp_call_function(): Run a function on all other CPUs.
 805  * @func: The function to run. This must be fast and non-blocking.
 806  * @info: An arbitrary pointer to pass to the function.
 807  * @nonatomic: currently unused.
 808  * @wait: If true, wait (atomically) until function has completed on other CPUs.
 809  *
 810  * Returns 0 on success, else a negative status code. Does not return until
 811  * remote CPUs are nearly ready to execute <<func>> or are or have executed.
 812  *
 813  * You must not call this function with disabled interrupts or from a
 814  * hardware interrupt handler or from a bottom half handler.
 815  */
 816 static int smp_call_function_mask(void (*func)(void *info), void *info,
 817                                   int nonatomic, int wait, cpumask_t mask)
 818 {
 819         struct call_data_struct data;
 820         int cpus;
 821
 822         /* Can deadlock when called with interrupts disabled */
 823         WARN_ON(irqs_disabled());
 824
 825         data.func = func;
 826         data.info = info;
 827         atomic_set(&data.finished, 0);
 828         data.wait = wait;
 829
 830         spin_lock(&call_lock);
 831
 832         cpu_clear(smp_processor_id(), mask);
 833         cpus = cpus_weight(mask);
 834         if (!cpus)
 835                 goto out_unlock;
 836
 837         call_data = &data;
 838         mb();
 839
 840         smp_cross_call_masked(&xcall_call_function, 0, 0, 0, mask);
 841
 842         /* Wait for response */
 843         while (atomic_read(&data.finished) != cpus)
 844                 cpu_relax();
 845
 846 out_unlock:
 847         spin_unlock(&call_lock);
 848
 849         return 0;
 850 }
 851
 852 int smp_call_function(void (*func)(void *info), void *info,
 853                       int nonatomic, int wait)
 854 {
 855         return smp_call_function_mask(func, info, nonatomic, wait,
 856                                       cpu_online_map);
 857 }
 858
 859 void smp_call_function_client(int irq, struct pt_regs *regs)
 860 {
 861         void (*func) (void *info) = call_data->func;
 862         void *info = call_data->info;
 863
 864         clear_softint(1 << irq);
 865         if (call_data->wait) {
 866                 /* let initiator proceed only after completion */
 867                 func(info);
 868                 atomic_inc(&call_data->finished);
 869         } else {
 870                 /* let initiator proceed after getting data */
 871                 atomic_inc(&call_data->finished);
 872                 func(info);
 873         }
 874 }
 875
 876 static void tsb_sync(void *info)
 877 {
 878         struct trap_per_cpu *tp = &trap_block[raw_smp_processor_id()];
 879         struct mm_struct *mm = info;
 880
 881         /* It is not valid to test "currrent->active_mm == mm" here.
 882          *
 883          * The value of "current" is not changed atomically with
 884          * switch_mm().  But that's OK, we just need to check the
 885          * current cpu's trap block PGD physical address.
 886          */
 887         if (tp->pgd_paddr == __pa(mm->pgd))
 888                 tsb_context_switch(mm);
 889 }
 890
 891 void smp_tsb_sync(struct mm_struct *mm)
 892 {
 893         smp_call_function_mask(tsb_sync, mm, 0, 1, mm->cpu_vm_mask);
 894 }
 895
 896 extern unsigned long xcall_flush_tlb_mm;
 897 extern unsigned long xcall_flush_tlb_pending;
 898 extern unsigned long xcall_flush_tlb_kernel_range;
 899 extern unsigned long xcall_report_regs;
 900 extern unsigned long xcall_receive_signal;
 901 extern unsigned long xcall_new_mmu_context_version;
 902
 903 #ifdef DCACHE_ALIASING_POSSIBLE
 904 extern unsigned long xcall_flush_dcache_page_cheetah;
 905 #endif
 906 extern unsigned long xcall_flush_dcache_page_spitfire;
 907
 908 #ifdef CONFIG_DEBUG_DCFLUSH
 909 extern atomic_t dcpage_flushes;
 910 extern atomic_t dcpage_flushes_xcall;
 911 #endif
 912
 913 static __inline__ void __local_flush_dcache_page(struct page *page)
 914 {
 915 #ifdef DCACHE_ALIASING_POSSIBLE
 916         __flush_dcache_page(page_address(page),
 917                             ((tlb_type == spitfire) &&
 918                              page_mapping(page) != NULL));
 919 #else
 920         if (page_mapping(page) != NULL &&
 921             tlb_type == spitfire)
 922                 __flush_icache_page(__pa(page_address(page)));
 923 #endif
 924 }
 925
 926 void smp_flush_dcache_page_impl(struct page *page, int cpu)
 927 {
 928         cpumask_t mask = cpumask_of_cpu(cpu);
 929         int this_cpu;
 930
 931         if (tlb_type == hypervisor)
 932                 return;
 933
 934 #ifdef CONFIG_DEBUG_DCFLUSH
 935         atomic_inc(&dcpage_flushes);
 936 #endif
 937
 938         this_cpu = get_cpu();
 939
 940         if (cpu == this_cpu) {
 941                 __local_flush_dcache_page(page);
 942         } else if (cpu_online(cpu)) {
 943                 void *pg_addr = page_address(page);
 944                 u64 data0;
 945
 946                 if (tlb_type == spitfire) {
 947                         data0 =
 948                                 ((u64)&xcall_flush_dcache_page_spitfire);
 949                         if (page_mapping(page) != NULL)
 950                                 data0 |= ((u64)1 << 32);
 951                         spitfire_xcall_deliver(data0,
 952                                                __pa(pg_addr),
 953                                                (u64) pg_addr,
 954                                                mask);
 955                 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 956 #ifdef DCACHE_ALIASING_POSSIBLE
 957                         data0 =
 958                                 ((u64)&xcall_flush_dcache_page_cheetah);
 959                         cheetah_xcall_deliver(data0,
 960                                               __pa(pg_addr),
 961                                               0, mask);
 962 #endif
 963                 }
 964 #ifdef CONFIG_DEBUG_DCFLUSH
 965                 atomic_inc(&dcpage_flushes_xcall);
 966 #endif
 967         }
 968
 969         put_cpu();
 970 }
 971
 972 void flush_dcache_page_all(struct mm_struct *mm, struct page *page)
 973 {
 974         void *pg_addr = page_address(page);
 975         cpumask_t mask = cpu_online_map;
 976         u64 data0;
 977         int this_cpu;
 978
 979         if (tlb_type == hypervisor)
 980                 return;
 981
 982         this_cpu = get_cpu();
 983
 984         cpu_clear(this_cpu, mask);
 985
 986 #ifdef CONFIG_DEBUG_DCFLUSH
 987         atomic_inc(&dcpage_flushes);
 988 #endif
 989         if (cpus_empty(mask))
 990                 goto flush_self;
 991         if (tlb_type == spitfire) {
 992                 data0 = ((u64)&xcall_flush_dcache_page_spitfire);
 993                 if (page_mapping(page) != NULL)
 994                         data0 |= ((u64)1 << 32);
 995                 spitfire_xcall_deliver(data0,
 996                                        __pa(pg_addr),
 997                                        (u64) pg_addr,
 998                                        mask);
 999         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1000 #ifdef DCACHE_ALIASING_POSSIBLE
1001                 data0 = ((u64)&xcall_flush_dcache_page_cheetah);
1002                 cheetah_xcall_deliver(data0,
1003                                       __pa(pg_addr),
1004                                       0, mask);
1005 #endif
1006         }
1007 #ifdef CONFIG_DEBUG_DCFLUSH
1008         atomic_inc(&dcpage_flushes_xcall);
1009 #endif
1010  flush_self:
1011         __local_flush_dcache_page(page);
1012
1013         put_cpu();
1014 }
1015
1016 static void __smp_receive_signal_mask(cpumask_t mask)
1017 {
1018         smp_cross_call_masked(&xcall_receive_signal, 0, 0, 0, mask);
1019 }
1020
1021 void smp_receive_signal(int cpu)
1022 {
1023         cpumask_t mask = cpumask_of_cpu(cpu);
1024
1025         if (cpu_online(cpu))
1026                 __smp_receive_signal_mask(mask);
1027 }
1028
1029 void smp_receive_signal_client(int irq, struct pt_regs *regs)
1030 {
1031         clear_softint(1 << irq);
1032 }
1033
1034 void smp_new_mmu_context_version_client(int irq, struct pt_regs *regs)
1035 {
1036         struct mm_struct *mm;
1037         unsigned long flags;
1038
1039         clear_softint(1 << irq);
1040
1041         /* See if we need to allocate a new TLB context because
1042          * the version of the one we are using is now out of date.
1043          */
1044         mm = current->active_mm;
1045         if (unlikely(!mm || (mm == &init_mm)))
1046                 return;
1047
1048         spin_lock_irqsave(&mm->context.lock, flags);
1049
1050         if (unlikely(!CTX_VALID(mm->context)))
1051                 get_new_mmu_context(mm);
1052
1053         spin_unlock_irqrestore(&mm->context.lock, flags);
1054
1055         load_secondary_context(mm);
1056         __flush_tlb_mm(CTX_HWBITS(mm->context),
1057                        SECONDARY_CONTEXT);
1058 }
1059
1060 void smp_new_mmu_context_version(void)
1061 {
1062         smp_cross_call(&xcall_new_mmu_context_version, 0, 0, 0);
1063 }
1064
1065 void smp_report_regs(void)
1066 {
1067         smp_cross_call(&xcall_report_regs, 0, 0, 0);
1068 }
1069
1070 /* We know that the window frames of the user have been flushed
1071  * to the stack before we get here because all callers of us
1072  * are flush_tlb_*() routines, and these run after flush_cache_*()
1073  * which performs the flushw.
1074  *
1075  * The SMP TLB coherency scheme we use works as follows:
1076  *
1077  * 1) mm->cpu_vm_mask is a bit mask of which cpus an address
1078  *    space has (potentially) executed on, this is the heuristic
1079  *    we use to avoid doing cross calls.
1080  *
1081  *    Also, for flushing from kswapd and also for clones, we
1082  *    use cpu_vm_mask as the list of cpus to make run the TLB.
1083  *
1084  * 2) TLB context numbers are shared globally across all processors
1085  *    in the system, this allows us to play several games to avoid
1086  *    cross calls.
1087  *
1088  *    One invariant is that when a cpu switches to a process, and
1089  *    that processes tsk->active_mm->cpu_vm_mask does not have the
1090  *    current cpu's bit set, that tlb context is flushed locally.
1091  *
1092  *    If the address space is non-shared (ie. mm->count == 1) we avoid
1093  *    cross calls when we want to flush the currently running process's
1094  *    tlb state.  This is done by clearing all cpu bits except the current
1095  *    processor's in current->active_mm->cpu_vm_mask and performing the
1096  *    flush locally only.  This will force any subsequent cpus which run
1097  *    this task to flush the context from the local tlb if the process
1098  *    migrates to another cpu (again).
1099  *
1100  * 3) For shared address spaces (threads) and swapping we bite the
1101  *    bullet for most cases and perform the cross call (but only to
1102  *    the cpus listed in cpu_vm_mask).
1103  *
1104  *    The performance gain from "optimizing" away the cross call for threads is
1105  *    questionable (in theory the big win for threads is the massive sharing of
1106  *    address space state across processors).
1107  */
1108
1109 /* This currently is only used by the hugetlb arch pre-fault
1110  * hook on UltraSPARC-III+ and later when changing the pagesize
1111  * bits of the context register for an address space.
1112  */
1113 void smp_flush_tlb_mm(struct mm_struct *mm)
1114 {
1115         u32 ctx = CTX_HWBITS(mm->context);
1116         int cpu = get_cpu();
1117
1118         if (atomic_read(&mm->mm_users) == 1) {
1119                 mm->cpu_vm_mask = cpumask_of_cpu(cpu);
1120                 goto local_flush_and_out;
1121         }
1122
1123         smp_cross_call_masked(&xcall_flush_tlb_mm,
1124                               ctx, 0, 0,
1125                               mm->cpu_vm_mask);
1126
1127 local_flush_and_out:
1128         __flush_tlb_mm(ctx, SECONDARY_CONTEXT);
1129
1130         put_cpu();
1131 }
1132
1133 void smp_flush_tlb_pending(struct mm_struct *mm, unsigned long nr, unsigned long *vaddrs)
1134 {
1135         u32 ctx = CTX_HWBITS(mm->context);
1136         int cpu = get_cpu();
1137
1138         if (mm == current->active_mm && atomic_read(&mm->mm_users) == 1)
1139                 mm->cpu_vm_mask = cpumask_of_cpu(cpu);
1140         else
1141                 smp_cross_call_masked(&xcall_flush_tlb_pending,
1142                                       ctx, nr, (unsigned long) vaddrs,
1143                                       mm->cpu_vm_mask);
1144
1145         __flush_tlb_pending(ctx, nr, vaddrs);
1146
1147         put_cpu();
1148 }
1149
1150 void smp_flush_tlb_kernel_range(unsigned long start, unsigned long end)
1151 {
1152         start &= PAGE_MASK;
1153         end    = PAGE_ALIGN(end);
1154         if (start != end) {
1155                 smp_cross_call(&xcall_flush_tlb_kernel_range,
1156                                0, start, end);
1157
1158                 __flush_tlb_kernel_range(start, end);
1159         }
1160 }
1161
1162 /* CPU capture. */
1163 /* #define CAPTURE_DEBUG */
1164 extern unsigned long xcall_capture;
1165
1166 static atomic_t smp_capture_depth = ATOMIC_INIT(0);
1167 static atomic_t smp_capture_registry = ATOMIC_INIT(0);
1168 static unsigned long penguins_are_doing_time;
1169
1170 void smp_capture(void)
1171 {
1172         int result = atomic_add_ret(1, &smp_capture_depth);
1173
1174         if (result == 1) {
1175                 int ncpus = num_online_cpus();
1176
1177 #ifdef CAPTURE_DEBUG
1178                 printk("CPU[%d]: Sending penguins to jail...",
1179                        smp_processor_id());
1180 #endif
1181                 penguins_are_doing_time = 1;
1182                 membar_storestore_loadstore();
1183                 atomic_inc(&smp_capture_registry);
1184                 smp_cross_call(&xcall_capture, 0, 0, 0);
1185                 while (atomic_read(&smp_capture_registry) != ncpus)
1186                         rmb();
1187 #ifdef CAPTURE_DEBUG
1188                 printk("done\n");
1189 #endif
1190         }
1191 }
1192
1193 void smp_release(void)
1194 {
1195         if (atomic_dec_and_test(&smp_capture_depth)) {
1196 #ifdef CAPTURE_DEBUG
1197                 printk("CPU[%d]: Giving pardon to "
1198                        "imprisoned penguins\n",
1199                        smp_processor_id());
1200 #endif
1201                 penguins_are_doing_time = 0;
1202                 membar_storeload_storestore();
1203                 atomic_dec(&smp_capture_registry);
1204         }
1205 }
1206
1207 /* Imprisoned penguins run with %pil == 15, but PSTATE_IE set, so they
1208  * can service tlb flush xcalls...
1209  */
1210 extern void prom_world(int);
1211
1212 void smp_penguin_jailcell(int irq, struct pt_regs *regs)
1213 {
1214         clear_softint(1 << irq);
1215
1216         preempt_disable();
1217
1218         __asm__ __volatile__("flushw");
1219         prom_world(1);
1220         atomic_inc(&smp_capture_registry);
1221         membar_storeload_storestore();
1222         while (penguins_are_doing_time)
1223                 rmb();
1224         atomic_dec(&smp_capture_registry);
1225         prom_world(0);
1226
1227         preempt_enable();
1228 }
1229
1230 /* /proc/profile writes can call this, don't __init it please. */
1231 int setup_profiling_timer(unsigned int multiplier)
1232 {
1233         return -EINVAL;
1234 }
1235
1236 void __init smp_prepare_cpus(unsigned int max_cpus)
1237 {
1238 }
1239
1240 void __devinit smp_prepare_boot_cpu(void)
1241 {
1242 }
1243
1244 void __devinit smp_fill_in_sib_core_maps(void)
1245 {
1246         unsigned int i;
1247
1248         for_each_present_cpu(i) {
1249                 unsigned int j;
1250
1251                 cpus_clear(cpu_core_map[i]);
1252                 if (cpu_data(i).core_id == 0) {
1253                         cpu_set(i, cpu_core_map[i]);
1254                         continue;
1255                 }
1256
1257                 for_each_present_cpu(j) {
1258                         if (cpu_data(i).core_id ==
1259                             cpu_data(j).core_id)
1260                                 cpu_set(j, cpu_core_map[i]);
1261                 }
1262         }
1263
1264         for_each_present_cpu(i) {
1265                 unsigned int j;
1266
1267                 cpus_clear(cpu_sibling_map[i]);
1268                 if (cpu_data(i).proc_id == -1) {
1269                         cpu_set(i, cpu_sibling_map[i]);
1270                         continue;
1271                 }
1272
1273                 for_each_present_cpu(j) {
1274                         if (cpu_data(i).proc_id ==
1275                             cpu_data(j).proc_id)
1276                                 cpu_set(j, cpu_sibling_map[i]);
1277                 }
1278         }
1279 }
1280
1281 int __cpuinit __cpu_up(unsigned int cpu)
1282 {
1283         int ret = smp_boot_one_cpu(cpu);
1284
1285         if (!ret) {
1286                 cpu_set(cpu, smp_commenced_mask);
1287                 while (!cpu_isset(cpu, cpu_online_map))
1288                         mb();
1289                 if (!cpu_isset(cpu, cpu_online_map)) {
1290                         ret = -ENODEV;
1291                 } else {
1292                         /* On SUN4V, writes to %tick and %stick are
1293                          * not allowed.
1294                          */
1295                         if (tlb_type != hypervisor)
1296                                 smp_synchronize_one_tick(cpu);
1297                 }
1298         }
1299         return ret;
1300 }
1301
1302 #ifdef CONFIG_HOTPLUG_CPU
1303 void cpu_play_dead(void)
1304 {
1305         int cpu = smp_processor_id();
1306         unsigned long pstate;
1307
1308         idle_task_exit();
1309
1310         if (tlb_type == hypervisor) {
1311                 struct trap_per_cpu *tb = &trap_block[cpu];
1312
1313                 sun4v_cpu_qconf(HV_CPU_QUEUE_CPU_MONDO,
1314                                 tb->cpu_mondo_pa, 0);
1315                 sun4v_cpu_qconf(HV_CPU_QUEUE_DEVICE_MONDO,
1316                                 tb->dev_mondo_pa, 0);
1317                 sun4v_cpu_qconf(HV_CPU_QUEUE_RES_ERROR,
1318                                 tb->resum_mondo_pa, 0);
1319                 sun4v_cpu_qconf(HV_CPU_QUEUE_NONRES_ERROR,
1320                                 tb->nonresum_mondo_pa, 0);
1321         }
1322
1323         cpu_clear(cpu, smp_commenced_mask);
1324         membar_safe("#Sync");
1325
1326         local_irq_disable();
1327
1328         __asm__ __volatile__(
1329                 "rdpr   %%pstate, %0\n\t"
1330                 "wrpr   %0, %1, %%pstate"
1331                 : "=r" (pstate)
1332                 : "i" (PSTATE_IE));
1333
1334         while (1)
1335                 barrier();
1336 }
1337
1338 int __cpu_disable(void)
1339 {
1340         int cpu = smp_processor_id();
1341         cpuinfo_sparc *c;
1342         int i;
1343
1344         for_each_cpu_mask(i, cpu_core_map[cpu])
1345                 cpu_clear(cpu, cpu_core_map[i]);
1346         cpus_clear(cpu_core_map[cpu]);
1347
1348         for_each_cpu_mask(i, cpu_sibling_map[cpu])
1349                 cpu_clear(cpu, cpu_sibling_map[i]);
1350         cpus_clear(cpu_sibling_map[cpu]);
1351
1352         c = &cpu_data(cpu);
1353
1354         c->core_id = 0;
1355         c->proc_id = -1;
1356
1357         spin_lock(&call_lock);
1358         cpu_clear(cpu, cpu_online_map);
1359         spin_unlock(&call_lock);
1360
1361         smp_wmb();
1362
1363         /* Make sure no interrupts point to this cpu.  */
1364         fixup_irqs();
1365
1366         local_irq_enable();
1367         mdelay(1);
1368         local_irq_disable();
1369
1370         return 0;
1371 }
1372
1373 void __cpu_die(unsigned int cpu)
1374 {
1375         int i;
1376
1377         for (i = 0; i < 100; i++) {
1378                 smp_rmb();
1379                 if (!cpu_isset(cpu, smp_commenced_mask))
1380                         break;
1381                 msleep(100);
1382         }
1383         if (cpu_isset(cpu, smp_commenced_mask)) {
1384                 printk(KERN_ERR "CPU %u didn't die...\n", cpu);
1385         } else {
1386 #if defined(CONFIG_SUN_LDOMS)
1387                 unsigned long hv_err;
1388                 int limit = 100;
1389
1390                 do {
1391                         hv_err = sun4v_cpu_stop(cpu);
1392                         if (hv_err == HV_EOK) {
1393                                 cpu_clear(cpu, cpu_present_map);
1394                                 break;
1395                         }
1396                 } while (--limit > 0);
1397                 if (limit <= 0) {
1398                         printk(KERN_ERR "sun4v_cpu_stop() fails err=%lu\n",
1399                                hv_err);
1400                 }
1401 #endif
1402         }
1403 }
1404 #endif
1405
1406 void __init smp_cpus_done(unsigned int max_cpus)
1407 {
1408 }
1409
1410 void smp_send_reschedule(int cpu)
1411 {
1412         smp_receive_signal(cpu);
1413 }
1414
1415 /* This is a nop because we capture all other cpus
1416  * anyways when making the PROM active.
1417  */
1418 void smp_send_stop(void)
1419 {
1420 }
1421
1422 unsigned long __per_cpu_base __read_mostly;
1423 unsigned long __per_cpu_shift __read_mostly;
1424
1425 EXPORT_SYMBOL(__per_cpu_base);
1426 EXPORT_SYMBOL(__per_cpu_shift);
1427
1428 void __init real_setup_per_cpu_areas(void)
1429 {
1430         unsigned long goal, size, i;
1431         char *ptr;
1432
1433         /* Copy section for each CPU (we discard the original) */
1434         goal = PERCPU_ENOUGH_ROOM;
1435
1436         __per_cpu_shift = PAGE_SHIFT;
1437         for (size = PAGE_SIZE; size < goal; size <<= 1UL)
1438                 __per_cpu_shift++;
1439
1440         ptr = alloc_bootmem_pages(size * NR_CPUS);
1441
1442         __per_cpu_base = ptr - __per_cpu_start;
1443
1444         for (i = 0; i < NR_CPUS; i++, ptr += size)
1445                 memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start);
1446
1447         /* Setup %g5 for the boot cpu.  */
1448         __local_per_cpu_offset = __per_cpu_offset(smp_processor_id());
1449 }