git.oblomov.eu Git - linux-2.6/blob - arch/sparc/kernel/smp_64.c

   1 /* smp.c: Sparc64 SMP support.
   2  *
   3  * Copyright (C) 1997, 2007, 2008 David S. Miller (davem@davemloft.net)
   4  */
   5
   6 #include <linux/module.h>
   7 #include <linux/kernel.h>
   8 #include <linux/sched.h>
   9 #include <linux/mm.h>
  10 #include <linux/pagemap.h>
  11 #include <linux/threads.h>
  12 #include <linux/smp.h>
  13 #include <linux/interrupt.h>
  14 #include <linux/kernel_stat.h>
  15 #include <linux/delay.h>
  16 #include <linux/init.h>
  17 #include <linux/spinlock.h>
  18 #include <linux/fs.h>
  19 #include <linux/seq_file.h>
  20 #include <linux/cache.h>
  21 #include <linux/jiffies.h>
  22 #include <linux/profile.h>
  23 #include <linux/lmb.h>
  24 #include <linux/cpu.h>
  25
  26 #include <asm/head.h>
  27 #include <asm/ptrace.h>
  28 #include <asm/atomic.h>
  29 #include <asm/tlbflush.h>
  30 #include <asm/mmu_context.h>
  31 #include <asm/cpudata.h>
  32 #include <asm/hvtramp.h>
  33 #include <asm/io.h>
  34 #include <asm/timer.h>
  35
  36 #include <asm/irq.h>
  37 #include <asm/irq_regs.h>
  38 #include <asm/page.h>
  39 #include <asm/pgtable.h>
  40 #include <asm/oplib.h>
  41 #include <asm/uaccess.h>
  42 #include <asm/starfire.h>
  43 #include <asm/tlb.h>
  44 #include <asm/sections.h>
  45 #include <asm/prom.h>
  46 #include <asm/mdesc.h>
  47 #include <asm/ldc.h>
  48 #include <asm/hypervisor.h>
  49
  50 int sparc64_multi_core __read_mostly;
  51
  52 DEFINE_PER_CPU(cpumask_t, cpu_sibling_map) = CPU_MASK_NONE;
  53 cpumask_t cpu_core_map[NR_CPUS] __read_mostly =
  54         { [0 ... NR_CPUS-1] = CPU_MASK_NONE };
  55
  56 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
  57 EXPORT_SYMBOL(cpu_core_map);
  58
  59 static cpumask_t smp_commenced_mask;
  60
  61 void smp_info(struct seq_file *m)
  62 {
  63         int i;
  64
  65         seq_printf(m, "State:\n");
  66         for_each_online_cpu(i)
  67                 seq_printf(m, "CPU%d:\t\tonline\n", i);
  68 }
  69
  70 void smp_bogo(struct seq_file *m)
  71 {
  72         int i;
  73
  74         for_each_online_cpu(i)
  75                 seq_printf(m,
  76                            "Cpu%dClkTck\t: %016lx\n",
  77                            i, cpu_data(i).clock_tick);
  78 }
  79
  80 extern void setup_sparc64_timer(void);
  81
  82 static volatile unsigned long callin_flag = 0;
  83
  84 void __cpuinit smp_callin(void)
  85 {
  86         int cpuid = hard_smp_processor_id();
  87
  88         __local_per_cpu_offset = __per_cpu_offset(cpuid);
  89
  90         if (tlb_type == hypervisor)
  91                 sun4v_ktsb_register();
  92
  93         __flush_tlb_all();
  94
  95         setup_sparc64_timer();
  96
  97         if (cheetah_pcache_forced_on)
  98                 cheetah_enable_pcache();
  99
 100         local_irq_enable();
 101
 102         callin_flag = 1;
 103         __asm__ __volatile__("membar #Sync\n\t"
 104                              "flush  %%g6" : : : "memory");
 105
 106         /* Clear this or we will die instantly when we
 107          * schedule back to this idler...
 108          */
 109         current_thread_info()->new_child = 0;
 110
 111         /* Attach to the address space of init_task. */
 112         atomic_inc(&init_mm.mm_count);
 113         current->active_mm = &init_mm;
 114
 115         /* inform the notifiers about the new cpu */
 116         notify_cpu_starting(cpuid);
 117
 118         while (!cpu_isset(cpuid, smp_commenced_mask))
 119                 rmb();
 120
 121         ipi_call_lock_irq();
 122         cpu_set(cpuid, cpu_online_map);
 123         ipi_call_unlock_irq();
 124
 125         /* idle thread is expected to have preempt disabled */
 126         preempt_disable();
 127 }
 128
 129 void cpu_panic(void)
 130 {
 131         printk("CPU[%d]: Returns from cpu_idle!\n", smp_processor_id());
 132         panic("SMP bolixed\n");
 133 }
 134
 135 /* This tick register synchronization scheme is taken entirely from
 136  * the ia64 port, see arch/ia64/kernel/smpboot.c for details and credit.
 137  *
 138  * The only change I've made is to rework it so that the master
 139  * initiates the synchonization instead of the slave. -DaveM
 140  */
 141
 142 #define MASTER  0
 143 #define SLAVE   (SMP_CACHE_BYTES/sizeof(unsigned long))
 144
 145 #define NUM_ROUNDS      64      /* magic value */
 146 #define NUM_ITERS       5       /* likewise */
 147
 148 static DEFINE_SPINLOCK(itc_sync_lock);
 149 static unsigned long go[SLAVE + 1];
 150
 151 #define DEBUG_TICK_SYNC 0
 152
 153 static inline long get_delta (long *rt, long *master)
 154 {
 155         unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
 156         unsigned long tcenter, t0, t1, tm;
 157         unsigned long i;
 158
 159         for (i = 0; i < NUM_ITERS; i++) {
 160                 t0 = tick_ops->get_tick();
 161                 go[MASTER] = 1;
 162                 membar_safe("#StoreLoad");
 163                 while (!(tm = go[SLAVE]))
 164                         rmb();
 165                 go[SLAVE] = 0;
 166                 wmb();
 167                 t1 = tick_ops->get_tick();
 168
 169                 if (t1 - t0 < best_t1 - best_t0)
 170                         best_t0 = t0, best_t1 = t1, best_tm = tm;
 171         }
 172
 173         *rt = best_t1 - best_t0;
 174         *master = best_tm - best_t0;
 175
 176         /* average best_t0 and best_t1 without overflow: */
 177         tcenter = (best_t0/2 + best_t1/2);
 178         if (best_t0 % 2 + best_t1 % 2 == 2)
 179                 tcenter++;
 180         return tcenter - best_tm;
 181 }
 182
 183 void smp_synchronize_tick_client(void)
 184 {
 185         long i, delta, adj, adjust_latency = 0, done = 0;
 186         unsigned long flags, rt, master_time_stamp, bound;
 187 #if DEBUG_TICK_SYNC
 188         struct {
 189                 long rt;        /* roundtrip time */
 190                 long master;    /* master's timestamp */
 191                 long diff;      /* difference between midpoint and master's timestamp */
 192                 long lat;       /* estimate of itc adjustment latency */
 193         } t[NUM_ROUNDS];
 194 #endif
 195
 196         go[MASTER] = 1;
 197
 198         while (go[MASTER])
 199                 rmb();
 200
 201         local_irq_save(flags);
 202         {
 203                 for (i = 0; i < NUM_ROUNDS; i++) {
 204                         delta = get_delta(&rt, &master_time_stamp);
 205                         if (delta == 0) {
 206                                 done = 1;       /* let's lock on to this... */
 207                                 bound = rt;
 208                         }
 209
 210                         if (!done) {
 211                                 if (i > 0) {
 212                                         adjust_latency += -delta;
 213                                         adj = -delta + adjust_latency/4;
 214                                 } else
 215                                         adj = -delta;
 216
 217                                 tick_ops->add_tick(adj);
 218                         }
 219 #if DEBUG_TICK_SYNC
 220                         t[i].rt = rt;
 221                         t[i].master = master_time_stamp;
 222                         t[i].diff = delta;
 223                         t[i].lat = adjust_latency/4;
 224 #endif
 225                 }
 226         }
 227         local_irq_restore(flags);
 228
 229 #if DEBUG_TICK_SYNC
 230         for (i = 0; i < NUM_ROUNDS; i++)
 231                 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
 232                        t[i].rt, t[i].master, t[i].diff, t[i].lat);
 233 #endif
 234
 235         printk(KERN_INFO "CPU %d: synchronized TICK with master CPU "
 236                "(last diff %ld cycles, maxerr %lu cycles)\n",
 237                smp_processor_id(), delta, rt);
 238 }
 239
 240 static void smp_start_sync_tick_client(int cpu);
 241
 242 static void smp_synchronize_one_tick(int cpu)
 243 {
 244         unsigned long flags, i;
 245
 246         go[MASTER] = 0;
 247
 248         smp_start_sync_tick_client(cpu);
 249
 250         /* wait for client to be ready */
 251         while (!go[MASTER])
 252                 rmb();
 253
 254         /* now let the client proceed into his loop */
 255         go[MASTER] = 0;
 256         membar_safe("#StoreLoad");
 257
 258         spin_lock_irqsave(&itc_sync_lock, flags);
 259         {
 260                 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; i++) {
 261                         while (!go[MASTER])
 262                                 rmb();
 263                         go[MASTER] = 0;
 264                         wmb();
 265                         go[SLAVE] = tick_ops->get_tick();
 266                         membar_safe("#StoreLoad");
 267                 }
 268         }
 269         spin_unlock_irqrestore(&itc_sync_lock, flags);
 270 }
 271
 272 #if defined(CONFIG_SUN_LDOMS) && defined(CONFIG_HOTPLUG_CPU)
 273 /* XXX Put this in some common place. XXX */
 274 static unsigned long kimage_addr_to_ra(void *p)
 275 {
 276         unsigned long val = (unsigned long) p;
 277
 278         return kern_base + (val - KERNBASE);
 279 }
 280
 281 static void __cpuinit ldom_startcpu_cpuid(unsigned int cpu, unsigned long thread_reg)
 282 {
 283         extern unsigned long sparc64_ttable_tl0;
 284         extern unsigned long kern_locked_tte_data;
 285         struct hvtramp_descr *hdesc;
 286         unsigned long trampoline_ra;
 287         struct trap_per_cpu *tb;
 288         u64 tte_vaddr, tte_data;
 289         unsigned long hv_err;
 290         int i;
 291
 292         hdesc = kzalloc(sizeof(*hdesc) +
 293                         (sizeof(struct hvtramp_mapping) *
 294                          num_kernel_image_mappings - 1),
 295                         GFP_KERNEL);
 296         if (!hdesc) {
 297                 printk(KERN_ERR "ldom_startcpu_cpuid: Cannot allocate "
 298                        "hvtramp_descr.\n");
 299                 return;
 300         }
 301
 302         hdesc->cpu = cpu;
 303         hdesc->num_mappings = num_kernel_image_mappings;
 304
 305         tb = &trap_block[cpu];
 306         tb->hdesc = hdesc;
 307
 308         hdesc->fault_info_va = (unsigned long) &tb->fault_info;
 309         hdesc->fault_info_pa = kimage_addr_to_ra(&tb->fault_info);
 310
 311         hdesc->thread_reg = thread_reg;
 312
 313         tte_vaddr = (unsigned long) KERNBASE;
 314         tte_data = kern_locked_tte_data;
 315
 316         for (i = 0; i < hdesc->num_mappings; i++) {
 317                 hdesc->maps[i].vaddr = tte_vaddr;
 318                 hdesc->maps[i].tte   = tte_data;
 319                 tte_vaddr += 0x400000;
 320                 tte_data  += 0x400000;
 321         }
 322
 323         trampoline_ra = kimage_addr_to_ra(hv_cpu_startup);
 324
 325         hv_err = sun4v_cpu_start(cpu, trampoline_ra,
 326                                  kimage_addr_to_ra(&sparc64_ttable_tl0),
 327                                  __pa(hdesc));
 328         if (hv_err)
 329                 printk(KERN_ERR "ldom_startcpu_cpuid: sun4v_cpu_start() "
 330                        "gives error %lu\n", hv_err);
 331 }
 332 #endif
 333
 334 extern unsigned long sparc64_cpu_startup;
 335
 336 /* The OBP cpu startup callback truncates the 3rd arg cookie to
 337  * 32-bits (I think) so to be safe we have it read the pointer
 338  * contained here so we work on >4GB machines. -DaveM
 339  */
 340 static struct thread_info *cpu_new_thread = NULL;
 341
 342 static int __cpuinit smp_boot_one_cpu(unsigned int cpu)
 343 {
 344         struct trap_per_cpu *tb = &trap_block[cpu];
 345         unsigned long entry =
 346                 (unsigned long)(&sparc64_cpu_startup);
 347         unsigned long cookie =
 348                 (unsigned long)(&cpu_new_thread);
 349         struct task_struct *p;
 350         int timeout, ret;
 351
 352         p = fork_idle(cpu);
 353         if (IS_ERR(p))
 354                 return PTR_ERR(p);
 355         callin_flag = 0;
 356         cpu_new_thread = task_thread_info(p);
 357
 358         if (tlb_type == hypervisor) {
 359 #if defined(CONFIG_SUN_LDOMS) && defined(CONFIG_HOTPLUG_CPU)
 360                 if (ldom_domaining_enabled)
 361                         ldom_startcpu_cpuid(cpu,
 362                                             (unsigned long) cpu_new_thread);
 363                 else
 364 #endif
 365                         prom_startcpu_cpuid(cpu, entry, cookie);
 366         } else {
 367                 struct device_node *dp = of_find_node_by_cpuid(cpu);
 368
 369                 prom_startcpu(dp->node, entry, cookie);
 370         }
 371
 372         for (timeout = 0; timeout < 50000; timeout++) {
 373                 if (callin_flag)
 374                         break;
 375                 udelay(100);
 376         }
 377
 378         if (callin_flag) {
 379                 ret = 0;
 380         } else {
 381                 printk("Processor %d is stuck.\n", cpu);
 382                 ret = -ENODEV;
 383         }
 384         cpu_new_thread = NULL;
 385
 386         if (tb->hdesc) {
 387                 kfree(tb->hdesc);
 388                 tb->hdesc = NULL;
 389         }
 390
 391         return ret;
 392 }
 393
 394 static void spitfire_xcall_helper(u64 data0, u64 data1, u64 data2, u64 pstate, unsigned long cpu)
 395 {
 396         u64 result, target;
 397         int stuck, tmp;
 398
 399         if (this_is_starfire) {
 400                 /* map to real upaid */
 401                 cpu = (((cpu & 0x3c) << 1) |
 402                         ((cpu & 0x40) >> 4) |
 403                         (cpu & 0x3));
 404         }
 405
 406         target = (cpu << 14) | 0x70;
 407 again:
 408         /* Ok, this is the real Spitfire Errata #54.
 409          * One must read back from a UDB internal register
 410          * after writes to the UDB interrupt dispatch, but
 411          * before the membar Sync for that write.
 412          * So we use the high UDB control register (ASI 0x7f,
 413          * ADDR 0x20) for the dummy read. -DaveM
 414          */
 415         tmp = 0x40;
 416         __asm__ __volatile__(
 417         "wrpr   %1, %2, %%pstate\n\t"
 418         "stxa   %4, [%0] %3\n\t"
 419         "stxa   %5, [%0+%8] %3\n\t"
 420         "add    %0, %8, %0\n\t"
 421         "stxa   %6, [%0+%8] %3\n\t"
 422         "membar #Sync\n\t"
 423         "stxa   %%g0, [%7] %3\n\t"
 424         "membar #Sync\n\t"
 425         "mov    0x20, %%g1\n\t"
 426         "ldxa   [%%g1] 0x7f, %%g0\n\t"
 427         "membar #Sync"
 428         : "=r" (tmp)
 429         : "r" (pstate), "i" (PSTATE_IE), "i" (ASI_INTR_W),
 430           "r" (data0), "r" (data1), "r" (data2), "r" (target),
 431           "r" (0x10), "0" (tmp)
 432         : "g1");
 433
 434         /* NOTE: PSTATE_IE is still clear. */
 435         stuck = 100000;
 436         do {
 437                 __asm__ __volatile__("ldxa [%%g0] %1, %0"
 438                         : "=r" (result)
 439                         : "i" (ASI_INTR_DISPATCH_STAT));
 440                 if (result == 0) {
 441                         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 442                                              : : "r" (pstate));
 443                         return;
 444                 }
 445                 stuck -= 1;
 446                 if (stuck == 0)
 447                         break;
 448         } while (result & 0x1);
 449         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 450                              : : "r" (pstate));
 451         if (stuck == 0) {
 452                 printk("CPU[%d]: mondo stuckage result[%016llx]\n",
 453                        smp_processor_id(), result);
 454         } else {
 455                 udelay(2);
 456                 goto again;
 457         }
 458 }
 459
 460 static void spitfire_xcall_deliver(struct trap_per_cpu *tb, int cnt)
 461 {
 462         u64 *mondo, data0, data1, data2;
 463         u16 *cpu_list;
 464         u64 pstate;
 465         int i;
 466
 467         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 468         cpu_list = __va(tb->cpu_list_pa);
 469         mondo = __va(tb->cpu_mondo_block_pa);
 470         data0 = mondo[0];
 471         data1 = mondo[1];
 472         data2 = mondo[2];
 473         for (i = 0; i < cnt; i++)
 474                 spitfire_xcall_helper(data0, data1, data2, pstate, cpu_list[i]);
 475 }
 476
 477 /* Cheetah now allows to send the whole 64-bytes of data in the interrupt
 478  * packet, but we have no use for that.  However we do take advantage of
 479  * the new pipelining feature (ie. dispatch to multiple cpus simultaneously).
 480  */
 481 static void cheetah_xcall_deliver(struct trap_per_cpu *tb, int cnt)
 482 {
 483         int nack_busy_id, is_jbus, need_more;
 484         u64 *mondo, pstate, ver, busy_mask;
 485         u16 *cpu_list;
 486
 487         cpu_list = __va(tb->cpu_list_pa);
 488         mondo = __va(tb->cpu_mondo_block_pa);
 489
 490         /* Unfortunately, someone at Sun had the brilliant idea to make the
 491          * busy/nack fields hard-coded by ITID number for this Ultra-III
 492          * derivative processor.
 493          */
 494         __asm__ ("rdpr %%ver, %0" : "=r" (ver));
 495         is_jbus = ((ver >> 32) == __JALAPENO_ID ||
 496                    (ver >> 32) == __SERRANO_ID);
 497
 498         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 499
 500 retry:
 501         need_more = 0;
 502         __asm__ __volatile__("wrpr %0, %1, %%pstate\n\t"
 503                              : : "r" (pstate), "i" (PSTATE_IE));
 504
 505         /* Setup the dispatch data registers. */
 506         __asm__ __volatile__("stxa      %0, [%3] %6\n\t"
 507                              "stxa      %1, [%4] %6\n\t"
 508                              "stxa      %2, [%5] %6\n\t"
 509                              "membar    #Sync\n\t"
 510                              : /* no outputs */
 511                              : "r" (mondo[0]), "r" (mondo[1]), "r" (mondo[2]),
 512                                "r" (0x40), "r" (0x50), "r" (0x60),
 513                                "i" (ASI_INTR_W));
 514
 515         nack_busy_id = 0;
 516         busy_mask = 0;
 517         {
 518                 int i;
 519
 520                 for (i = 0; i < cnt; i++) {
 521                         u64 target, nr;
 522
 523                         nr = cpu_list[i];
 524                         if (nr == 0xffff)
 525                                 continue;
 526
 527                         target = (nr << 14) | 0x70;
 528                         if (is_jbus) {
 529                                 busy_mask |= (0x1UL << (nr * 2));
 530                         } else {
 531                                 target |= (nack_busy_id << 24);
 532                                 busy_mask |= (0x1UL <<
 533                                               (nack_busy_id * 2));
 534                         }
 535                         __asm__ __volatile__(
 536                                 "stxa   %%g0, [%0] %1\n\t"
 537                                 "membar #Sync\n\t"
 538                                 : /* no outputs */
 539                                 : "r" (target), "i" (ASI_INTR_W));
 540                         nack_busy_id++;
 541                         if (nack_busy_id == 32) {
 542                                 need_more = 1;
 543                                 break;
 544                         }
 545                 }
 546         }
 547
 548         /* Now, poll for completion. */
 549         {
 550                 u64 dispatch_stat, nack_mask;
 551                 long stuck;
 552
 553                 stuck = 100000 * nack_busy_id;
 554                 nack_mask = busy_mask << 1;
 555                 do {
 556                         __asm__ __volatile__("ldxa      [%%g0] %1, %0"
 557                                              : "=r" (dispatch_stat)
 558                                              : "i" (ASI_INTR_DISPATCH_STAT));
 559                         if (!(dispatch_stat & (busy_mask | nack_mask))) {
 560                                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 561                                                      : : "r" (pstate));
 562                                 if (unlikely(need_more)) {
 563                                         int i, this_cnt = 0;
 564                                         for (i = 0; i < cnt; i++) {
 565                                                 if (cpu_list[i] == 0xffff)
 566                                                         continue;
 567                                                 cpu_list[i] = 0xffff;
 568                                                 this_cnt++;
 569                                                 if (this_cnt == 32)
 570                                                         break;
 571                                         }
 572                                         goto retry;
 573                                 }
 574                                 return;
 575                         }
 576                         if (!--stuck)
 577                                 break;
 578                 } while (dispatch_stat & busy_mask);
 579
 580                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 581                                      : : "r" (pstate));
 582
 583                 if (dispatch_stat & busy_mask) {
 584                         /* Busy bits will not clear, continue instead
 585                          * of freezing up on this cpu.
 586                          */
 587                         printk("CPU[%d]: mondo stuckage result[%016llx]\n",
 588                                smp_processor_id(), dispatch_stat);
 589                 } else {
 590                         int i, this_busy_nack = 0;
 591
 592                         /* Delay some random time with interrupts enabled
 593                          * to prevent deadlock.
 594                          */
 595                         udelay(2 * nack_busy_id);
 596
 597                         /* Clear out the mask bits for cpus which did not
 598                          * NACK us.
 599                          */
 600                         for (i = 0; i < cnt; i++) {
 601                                 u64 check_mask, nr;
 602
 603                                 nr = cpu_list[i];
 604                                 if (nr == 0xffff)
 605                                         continue;
 606
 607                                 if (is_jbus)
 608                                         check_mask = (0x2UL << (2*nr));
 609                                 else
 610                                         check_mask = (0x2UL <<
 611                                                       this_busy_nack);
 612                                 if ((dispatch_stat & check_mask) == 0)
 613                                         cpu_list[i] = 0xffff;
 614                                 this_busy_nack += 2;
 615                                 if (this_busy_nack == 64)
 616                                         break;
 617                         }
 618
 619                         goto retry;
 620                 }
 621         }
 622 }
 623
 624 /* Multi-cpu list version.  */
 625 static void hypervisor_xcall_deliver(struct trap_per_cpu *tb, int cnt)
 626 {
 627         int retries, this_cpu, prev_sent, i, saw_cpu_error;
 628         unsigned long status;
 629         u16 *cpu_list;
 630
 631         this_cpu = smp_processor_id();
 632
 633         cpu_list = __va(tb->cpu_list_pa);
 634
 635         saw_cpu_error = 0;
 636         retries = 0;
 637         prev_sent = 0;
 638         do {
 639                 int forward_progress, n_sent;
 640
 641                 status = sun4v_cpu_mondo_send(cnt,
 642                                               tb->cpu_list_pa,
 643                                               tb->cpu_mondo_block_pa);
 644
 645                 /* HV_EOK means all cpus received the xcall, we're done.  */
 646                 if (likely(status == HV_EOK))
 647                         break;
 648
 649                 /* First, see if we made any forward progress.
 650                  *
 651                  * The hypervisor indicates successful sends by setting
 652                  * cpu list entries to the value 0xffff.
 653                  */
 654                 n_sent = 0;
 655                 for (i = 0; i < cnt; i++) {
 656                         if (likely(cpu_list[i] == 0xffff))
 657                                 n_sent++;
 658                 }
 659
 660                 forward_progress = 0;
 661                 if (n_sent > prev_sent)
 662                         forward_progress = 1;
 663
 664                 prev_sent = n_sent;
 665
 666                 /* If we get a HV_ECPUERROR, then one or more of the cpus
 667                  * in the list are in error state.  Use the cpu_state()
 668                  * hypervisor call to find out which cpus are in error state.
 669                  */
 670                 if (unlikely(status == HV_ECPUERROR)) {
 671                         for (i = 0; i < cnt; i++) {
 672                                 long err;
 673                                 u16 cpu;
 674
 675                                 cpu = cpu_list[i];
 676                                 if (cpu == 0xffff)
 677                                         continue;
 678
 679                                 err = sun4v_cpu_state(cpu);
 680                                 if (err == HV_CPU_STATE_ERROR) {
 681                                         saw_cpu_error = (cpu + 1);
 682                                         cpu_list[i] = 0xffff;
 683                                 }
 684                         }
 685                 } else if (unlikely(status != HV_EWOULDBLOCK))
 686                         goto fatal_mondo_error;
 687
 688                 /* Don't bother rewriting the CPU list, just leave the
 689                  * 0xffff and non-0xffff entries in there and the
 690                  * hypervisor will do the right thing.
 691                  *
 692                  * Only advance timeout state if we didn't make any
 693                  * forward progress.
 694                  */
 695                 if (unlikely(!forward_progress)) {
 696                         if (unlikely(++retries > 10000))
 697                                 goto fatal_mondo_timeout;
 698
 699                         /* Delay a little bit to let other cpus catch up
 700                          * on their cpu mondo queue work.
 701                          */
 702                         udelay(2 * cnt);
 703                 }
 704         } while (1);
 705
 706         if (unlikely(saw_cpu_error))
 707                 goto fatal_mondo_cpu_error;
 708
 709         return;
 710
 711 fatal_mondo_cpu_error:
 712         printk(KERN_CRIT "CPU[%d]: SUN4V mondo cpu error, some target cpus "
 713                "(including %d) were in error state\n",
 714                this_cpu, saw_cpu_error - 1);
 715         return;
 716
 717 fatal_mondo_timeout:
 718         printk(KERN_CRIT "CPU[%d]: SUN4V mondo timeout, no forward "
 719                " progress after %d retries.\n",
 720                this_cpu, retries);
 721         goto dump_cpu_list_and_out;
 722
 723 fatal_mondo_error:
 724         printk(KERN_CRIT "CPU[%d]: Unexpected SUN4V mondo error %lu\n",
 725                this_cpu, status);
 726         printk(KERN_CRIT "CPU[%d]: Args were cnt(%d) cpulist_pa(%lx) "
 727                "mondo_block_pa(%lx)\n",
 728                this_cpu, cnt, tb->cpu_list_pa, tb->cpu_mondo_block_pa);
 729
 730 dump_cpu_list_and_out:
 731         printk(KERN_CRIT "CPU[%d]: CPU list [ ", this_cpu);
 732         for (i = 0; i < cnt; i++)
 733                 printk("%u ", cpu_list[i]);
 734         printk("]\n");
 735 }
 736
 737 static void (*xcall_deliver_impl)(struct trap_per_cpu *, int);
 738
 739 static void xcall_deliver(u64 data0, u64 data1, u64 data2, const cpumask_t *mask)
 740 {
 741         struct trap_per_cpu *tb;
 742         int this_cpu, i, cnt;
 743         unsigned long flags;
 744         u16 *cpu_list;
 745         u64 *mondo;
 746
 747         /* We have to do this whole thing with interrupts fully disabled.
 748          * Otherwise if we send an xcall from interrupt context it will
 749          * corrupt both our mondo block and cpu list state.
 750          *
 751          * One consequence of this is that we cannot use timeout mechanisms
 752          * that depend upon interrupts being delivered locally.  So, for
 753          * example, we cannot sample jiffies and expect it to advance.
 754          *
 755          * Fortunately, udelay() uses %stick/%tick so we can use that.
 756          */
 757         local_irq_save(flags);
 758
 759         this_cpu = smp_processor_id();
 760         tb = &trap_block[this_cpu];
 761
 762         mondo = __va(tb->cpu_mondo_block_pa);
 763         mondo[0] = data0;
 764         mondo[1] = data1;
 765         mondo[2] = data2;
 766         wmb();
 767
 768         cpu_list = __va(tb->cpu_list_pa);
 769
 770         /* Setup the initial cpu list.  */
 771         cnt = 0;
 772         for_each_cpu(i, mask) {
 773                 if (i == this_cpu || !cpu_online(i))
 774                         continue;
 775                 cpu_list[cnt++] = i;
 776         }
 777
 778         if (cnt)
 779                 xcall_deliver_impl(tb, cnt);
 780
 781         local_irq_restore(flags);
 782 }
 783
 784 /* Send cross call to all processors mentioned in MASK_P
 785  * except self.  Really, there are only two cases currently,
 786  * "&cpu_online_map" and "&mm->cpu_vm_mask".
 787  */
 788 static void smp_cross_call_masked(unsigned long *func, u32 ctx, u64 data1, u64 data2, const cpumask_t *mask)
 789 {
 790         u64 data0 = (((u64)ctx)<<32 | (((u64)func) & 0xffffffff));
 791
 792         xcall_deliver(data0, data1, data2, mask);
 793 }
 794
 795 /* Send cross call to all processors except self. */
 796 static void smp_cross_call(unsigned long *func, u32 ctx, u64 data1, u64 data2)
 797 {
 798         smp_cross_call_masked(func, ctx, data1, data2, &cpu_online_map);
 799 }
 800
 801 extern unsigned long xcall_sync_tick;
 802
 803 static void smp_start_sync_tick_client(int cpu)
 804 {
 805         xcall_deliver((u64) &xcall_sync_tick, 0, 0,
 806                       &cpumask_of_cpu(cpu));
 807 }
 808
 809 extern unsigned long xcall_call_function;
 810
 811 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
 812 {
 813         xcall_deliver((u64) &xcall_call_function, 0, 0, mask);
 814 }
 815
 816 extern unsigned long xcall_call_function_single;
 817
 818 void arch_send_call_function_single_ipi(int cpu)
 819 {
 820         xcall_deliver((u64) &xcall_call_function_single, 0, 0,
 821                       &cpumask_of_cpu(cpu));
 822 }
 823
 824 void smp_call_function_client(int irq, struct pt_regs *regs)
 825 {
 826         clear_softint(1 << irq);
 827         generic_smp_call_function_interrupt();
 828 }
 829
 830 void smp_call_function_single_client(int irq, struct pt_regs *regs)
 831 {
 832         clear_softint(1 << irq);
 833         generic_smp_call_function_single_interrupt();
 834 }
 835
 836 static void tsb_sync(void *info)
 837 {
 838         struct trap_per_cpu *tp = &trap_block[raw_smp_processor_id()];
 839         struct mm_struct *mm = info;
 840
 841         /* It is not valid to test "currrent->active_mm == mm" here.
 842          *
 843          * The value of "current" is not changed atomically with
 844          * switch_mm().  But that's OK, we just need to check the
 845          * current cpu's trap block PGD physical address.
 846          */
 847         if (tp->pgd_paddr == __pa(mm->pgd))
 848                 tsb_context_switch(mm);
 849 }
 850
 851 void smp_tsb_sync(struct mm_struct *mm)
 852 {
 853         smp_call_function_many(mm_cpumask(mm), tsb_sync, mm, 1);
 854 }
 855
 856 extern unsigned long xcall_flush_tlb_mm;
 857 extern unsigned long xcall_flush_tlb_pending;
 858 extern unsigned long xcall_flush_tlb_kernel_range;
 859 extern unsigned long xcall_fetch_glob_regs;
 860 extern unsigned long xcall_receive_signal;
 861 extern unsigned long xcall_new_mmu_context_version;
 862 #ifdef CONFIG_KGDB
 863 extern unsigned long xcall_kgdb_capture;
 864 #endif
 865
 866 #ifdef DCACHE_ALIASING_POSSIBLE
 867 extern unsigned long xcall_flush_dcache_page_cheetah;
 868 #endif
 869 extern unsigned long xcall_flush_dcache_page_spitfire;
 870
 871 #ifdef CONFIG_DEBUG_DCFLUSH
 872 extern atomic_t dcpage_flushes;
 873 extern atomic_t dcpage_flushes_xcall;
 874 #endif
 875
 876 static inline void __local_flush_dcache_page(struct page *page)
 877 {
 878 #ifdef DCACHE_ALIASING_POSSIBLE
 879         __flush_dcache_page(page_address(page),
 880                             ((tlb_type == spitfire) &&
 881                              page_mapping(page) != NULL));
 882 #else
 883         if (page_mapping(page) != NULL &&
 884             tlb_type == spitfire)
 885                 __flush_icache_page(__pa(page_address(page)));
 886 #endif
 887 }
 888
 889 void smp_flush_dcache_page_impl(struct page *page, int cpu)
 890 {
 891         int this_cpu;
 892
 893         if (tlb_type == hypervisor)
 894                 return;
 895
 896 #ifdef CONFIG_DEBUG_DCFLUSH
 897         atomic_inc(&dcpage_flushes);
 898 #endif
 899
 900         this_cpu = get_cpu();
 901
 902         if (cpu == this_cpu) {
 903                 __local_flush_dcache_page(page);
 904         } else if (cpu_online(cpu)) {
 905                 void *pg_addr = page_address(page);
 906                 u64 data0 = 0;
 907
 908                 if (tlb_type == spitfire) {
 909                         data0 = ((u64)&xcall_flush_dcache_page_spitfire);
 910                         if (page_mapping(page) != NULL)
 911                                 data0 |= ((u64)1 << 32);
 912                 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 913 #ifdef DCACHE_ALIASING_POSSIBLE
 914                         data0 = ((u64)&xcall_flush_dcache_page_cheetah);
 915 #endif
 916                 }
 917                 if (data0) {
 918                         xcall_deliver(data0, __pa(pg_addr),
 919                                       (u64) pg_addr, &cpumask_of_cpu(cpu));
 920 #ifdef CONFIG_DEBUG_DCFLUSH
 921                         atomic_inc(&dcpage_flushes_xcall);
 922 #endif
 923                 }
 924         }
 925
 926         put_cpu();
 927 }
 928
 929 void flush_dcache_page_all(struct mm_struct *mm, struct page *page)
 930 {
 931         void *pg_addr;
 932         int this_cpu;
 933         u64 data0;
 934
 935         if (tlb_type == hypervisor)
 936                 return;
 937
 938         this_cpu = get_cpu();
 939
 940 #ifdef CONFIG_DEBUG_DCFLUSH
 941         atomic_inc(&dcpage_flushes);
 942 #endif
 943         data0 = 0;
 944         pg_addr = page_address(page);
 945         if (tlb_type == spitfire) {
 946                 data0 = ((u64)&xcall_flush_dcache_page_spitfire);
 947                 if (page_mapping(page) != NULL)
 948                         data0 |= ((u64)1 << 32);
 949         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 950 #ifdef DCACHE_ALIASING_POSSIBLE
 951                 data0 = ((u64)&xcall_flush_dcache_page_cheetah);
 952 #endif
 953         }
 954         if (data0) {
 955                 xcall_deliver(data0, __pa(pg_addr),
 956                               (u64) pg_addr, &cpu_online_map);
 957 #ifdef CONFIG_DEBUG_DCFLUSH
 958                 atomic_inc(&dcpage_flushes_xcall);
 959 #endif
 960         }
 961         __local_flush_dcache_page(page);
 962
 963         put_cpu();
 964 }
 965
 966 void smp_new_mmu_context_version_client(int irq, struct pt_regs *regs)
 967 {
 968         struct mm_struct *mm;
 969         unsigned long flags;
 970
 971         clear_softint(1 << irq);
 972
 973         /* See if we need to allocate a new TLB context because
 974          * the version of the one we are using is now out of date.
 975          */
 976         mm = current->active_mm;
 977         if (unlikely(!mm || (mm == &init_mm)))
 978                 return;
 979
 980         spin_lock_irqsave(&mm->context.lock, flags);
 981
 982         if (unlikely(!CTX_VALID(mm->context)))
 983                 get_new_mmu_context(mm);
 984
 985         spin_unlock_irqrestore(&mm->context.lock, flags);
 986
 987         load_secondary_context(mm);
 988         __flush_tlb_mm(CTX_HWBITS(mm->context),
 989                        SECONDARY_CONTEXT);
 990 }
 991
 992 void smp_new_mmu_context_version(void)
 993 {
 994         smp_cross_call(&xcall_new_mmu_context_version, 0, 0, 0);
 995 }
 996
 997 #ifdef CONFIG_KGDB
 998 void kgdb_roundup_cpus(unsigned long flags)
 999 {
1000         smp_cross_call(&xcall_kgdb_capture, 0, 0, 0);
1001 }
1002 #endif
1003
1004 void smp_fetch_global_regs(void)
1005 {
1006         smp_cross_call(&xcall_fetch_glob_regs, 0, 0, 0);
1007 }
1008
1009 /* We know that the window frames of the user have been flushed
1010  * to the stack before we get here because all callers of us
1011  * are flush_tlb_*() routines, and these run after flush_cache_*()
1012  * which performs the flushw.
1013  *
1014  * The SMP TLB coherency scheme we use works as follows:
1015  *
1016  * 1) mm->cpu_vm_mask is a bit mask of which cpus an address
1017  *    space has (potentially) executed on, this is the heuristic
1018  *    we use to avoid doing cross calls.
1019  *
1020  *    Also, for flushing from kswapd and also for clones, we
1021  *    use cpu_vm_mask as the list of cpus to make run the TLB.
1022  *
1023  * 2) TLB context numbers are shared globally across all processors
1024  *    in the system, this allows us to play several games to avoid
1025  *    cross calls.
1026  *
1027  *    One invariant is that when a cpu switches to a process, and
1028  *    that processes tsk->active_mm->cpu_vm_mask does not have the
1029  *    current cpu's bit set, that tlb context is flushed locally.
1030  *
1031  *    If the address space is non-shared (ie. mm->count == 1) we avoid
1032  *    cross calls when we want to flush the currently running process's
1033  *    tlb state.  This is done by clearing all cpu bits except the current
1034  *    processor's in current->mm->cpu_vm_mask and performing the
1035  *    flush locally only.  This will force any subsequent cpus which run
1036  *    this task to flush the context from the local tlb if the process
1037  *    migrates to another cpu (again).
1038  *
1039  * 3) For shared address spaces (threads) and swapping we bite the
1040  *    bullet for most cases and perform the cross call (but only to
1041  *    the cpus listed in cpu_vm_mask).
1042  *
1043  *    The performance gain from "optimizing" away the cross call for threads is
1044  *    questionable (in theory the big win for threads is the massive sharing of
1045  *    address space state across processors).
1046  */
1047
1048 /* This currently is only used by the hugetlb arch pre-fault
1049  * hook on UltraSPARC-III+ and later when changing the pagesize
1050  * bits of the context register for an address space.
1051  */
1052 void smp_flush_tlb_mm(struct mm_struct *mm)
1053 {
1054         u32 ctx = CTX_HWBITS(mm->context);
1055         int cpu = get_cpu();
1056
1057         if (atomic_read(&mm->mm_users) == 1) {
1058                 cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
1059                 goto local_flush_and_out;
1060         }
1061
1062         smp_cross_call_masked(&xcall_flush_tlb_mm,
1063                               ctx, 0, 0,
1064                               mm_cpumask(mm));
1065
1066 local_flush_and_out:
1067         __flush_tlb_mm(ctx, SECONDARY_CONTEXT);
1068
1069         put_cpu();
1070 }
1071
1072 void smp_flush_tlb_pending(struct mm_struct *mm, unsigned long nr, unsigned long *vaddrs)
1073 {
1074         u32 ctx = CTX_HWBITS(mm->context);
1075         int cpu = get_cpu();
1076
1077         if (mm == current->mm && atomic_read(&mm->mm_users) == 1)
1078                 cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
1079         else
1080                 smp_cross_call_masked(&xcall_flush_tlb_pending,
1081                                       ctx, nr, (unsigned long) vaddrs,
1082                                       mm_cpumask(mm));
1083
1084         __flush_tlb_pending(ctx, nr, vaddrs);
1085
1086         put_cpu();
1087 }
1088
1089 void smp_flush_tlb_kernel_range(unsigned long start, unsigned long end)
1090 {
1091         start &= PAGE_MASK;
1092         end    = PAGE_ALIGN(end);
1093         if (start != end) {
1094                 smp_cross_call(&xcall_flush_tlb_kernel_range,
1095                                0, start, end);
1096
1097                 __flush_tlb_kernel_range(start, end);
1098         }
1099 }
1100
1101 /* CPU capture. */
1102 /* #define CAPTURE_DEBUG */
1103 extern unsigned long xcall_capture;
1104
1105 static atomic_t smp_capture_depth = ATOMIC_INIT(0);
1106 static atomic_t smp_capture_registry = ATOMIC_INIT(0);
1107 static unsigned long penguins_are_doing_time;
1108
1109 void smp_capture(void)
1110 {
1111         int result = atomic_add_ret(1, &smp_capture_depth);
1112
1113         if (result == 1) {
1114                 int ncpus = num_online_cpus();
1115
1116 #ifdef CAPTURE_DEBUG
1117                 printk("CPU[%d]: Sending penguins to jail...",
1118                        smp_processor_id());
1119 #endif
1120                 penguins_are_doing_time = 1;
1121                 atomic_inc(&smp_capture_registry);
1122                 smp_cross_call(&xcall_capture, 0, 0, 0);
1123                 while (atomic_read(&smp_capture_registry) != ncpus)
1124                         rmb();
1125 #ifdef CAPTURE_DEBUG
1126                 printk("done\n");
1127 #endif
1128         }
1129 }
1130
1131 void smp_release(void)
1132 {
1133         if (atomic_dec_and_test(&smp_capture_depth)) {
1134 #ifdef CAPTURE_DEBUG
1135                 printk("CPU[%d]: Giving pardon to "
1136                        "imprisoned penguins\n",
1137                        smp_processor_id());
1138 #endif
1139                 penguins_are_doing_time = 0;
1140                 membar_safe("#StoreLoad");
1141                 atomic_dec(&smp_capture_registry);
1142         }
1143 }
1144
1145 /* Imprisoned penguins run with %pil == PIL_NORMAL_MAX, but PSTATE_IE
1146  * set, so they can service tlb flush xcalls...
1147  */
1148 extern void prom_world(int);
1149
1150 void smp_penguin_jailcell(int irq, struct pt_regs *regs)
1151 {
1152         clear_softint(1 << irq);
1153
1154         preempt_disable();
1155
1156         __asm__ __volatile__("flushw");
1157         prom_world(1);
1158         atomic_inc(&smp_capture_registry);
1159         membar_safe("#StoreLoad");
1160         while (penguins_are_doing_time)
1161                 rmb();
1162         atomic_dec(&smp_capture_registry);
1163         prom_world(0);
1164
1165         preempt_enable();
1166 }
1167
1168 /* /proc/profile writes can call this, don't __init it please. */
1169 int setup_profiling_timer(unsigned int multiplier)
1170 {
1171         return -EINVAL;
1172 }
1173
1174 void __init smp_prepare_cpus(unsigned int max_cpus)
1175 {
1176 }
1177
1178 void __devinit smp_prepare_boot_cpu(void)
1179 {
1180 }
1181
1182 void __init smp_setup_processor_id(void)
1183 {
1184         if (tlb_type == spitfire)
1185                 xcall_deliver_impl = spitfire_xcall_deliver;
1186         else if (tlb_type == cheetah || tlb_type == cheetah_plus)
1187                 xcall_deliver_impl = cheetah_xcall_deliver;
1188         else
1189                 xcall_deliver_impl = hypervisor_xcall_deliver;
1190 }
1191
1192 void __devinit smp_fill_in_sib_core_maps(void)
1193 {
1194         unsigned int i;
1195
1196         for_each_present_cpu(i) {
1197                 unsigned int j;
1198
1199                 cpus_clear(cpu_core_map[i]);
1200                 if (cpu_data(i).core_id == 0) {
1201                         cpu_set(i, cpu_core_map[i]);
1202                         continue;
1203                 }
1204
1205                 for_each_present_cpu(j) {
1206                         if (cpu_data(i).core_id ==
1207                             cpu_data(j).core_id)
1208                                 cpu_set(j, cpu_core_map[i]);
1209                 }
1210         }
1211
1212         for_each_present_cpu(i) {
1213                 unsigned int j;
1214
1215                 cpus_clear(per_cpu(cpu_sibling_map, i));
1216                 if (cpu_data(i).proc_id == -1) {
1217                         cpu_set(i, per_cpu(cpu_sibling_map, i));
1218                         continue;
1219                 }
1220
1221                 for_each_present_cpu(j) {
1222                         if (cpu_data(i).proc_id ==
1223                             cpu_data(j).proc_id)
1224                                 cpu_set(j, per_cpu(cpu_sibling_map, i));
1225                 }
1226         }
1227 }
1228
1229 int __cpuinit __cpu_up(unsigned int cpu)
1230 {
1231         int ret = smp_boot_one_cpu(cpu);
1232
1233         if (!ret) {
1234                 cpu_set(cpu, smp_commenced_mask);
1235                 while (!cpu_isset(cpu, cpu_online_map))
1236                         mb();
1237                 if (!cpu_isset(cpu, cpu_online_map)) {
1238                         ret = -ENODEV;
1239                 } else {
1240                         /* On SUN4V, writes to %tick and %stick are
1241                          * not allowed.
1242                          */
1243                         if (tlb_type != hypervisor)
1244                                 smp_synchronize_one_tick(cpu);
1245                 }
1246         }
1247         return ret;
1248 }
1249
1250 #ifdef CONFIG_HOTPLUG_CPU
1251 void cpu_play_dead(void)
1252 {
1253         int cpu = smp_processor_id();
1254         unsigned long pstate;
1255
1256         idle_task_exit();
1257
1258         if (tlb_type == hypervisor) {
1259                 struct trap_per_cpu *tb = &trap_block[cpu];
1260
1261                 sun4v_cpu_qconf(HV_CPU_QUEUE_CPU_MONDO,
1262                                 tb->cpu_mondo_pa, 0);
1263                 sun4v_cpu_qconf(HV_CPU_QUEUE_DEVICE_MONDO,
1264                                 tb->dev_mondo_pa, 0);
1265                 sun4v_cpu_qconf(HV_CPU_QUEUE_RES_ERROR,
1266                                 tb->resum_mondo_pa, 0);
1267                 sun4v_cpu_qconf(HV_CPU_QUEUE_NONRES_ERROR,
1268                                 tb->nonresum_mondo_pa, 0);
1269         }
1270
1271         cpu_clear(cpu, smp_commenced_mask);
1272         membar_safe("#Sync");
1273
1274         local_irq_disable();
1275
1276         __asm__ __volatile__(
1277                 "rdpr   %%pstate, %0\n\t"
1278                 "wrpr   %0, %1, %%pstate"
1279                 : "=r" (pstate)
1280                 : "i" (PSTATE_IE));
1281
1282         while (1)
1283                 barrier();
1284 }
1285
1286 int __cpu_disable(void)
1287 {
1288         int cpu = smp_processor_id();
1289         cpuinfo_sparc *c;
1290         int i;
1291
1292         for_each_cpu_mask(i, cpu_core_map[cpu])
1293                 cpu_clear(cpu, cpu_core_map[i]);
1294         cpus_clear(cpu_core_map[cpu]);
1295
1296         for_each_cpu_mask(i, per_cpu(cpu_sibling_map, cpu))
1297                 cpu_clear(cpu, per_cpu(cpu_sibling_map, i));
1298         cpus_clear(per_cpu(cpu_sibling_map, cpu));
1299
1300         c = &cpu_data(cpu);
1301
1302         c->core_id = 0;
1303         c->proc_id = -1;
1304
1305         smp_wmb();
1306
1307         /* Make sure no interrupts point to this cpu.  */
1308         fixup_irqs();
1309
1310         local_irq_enable();
1311         mdelay(1);
1312         local_irq_disable();
1313
1314         ipi_call_lock();
1315         cpu_clear(cpu, cpu_online_map);
1316         ipi_call_unlock();
1317
1318         return 0;
1319 }
1320
1321 void __cpu_die(unsigned int cpu)
1322 {
1323         int i;
1324
1325         for (i = 0; i < 100; i++) {
1326                 smp_rmb();
1327                 if (!cpu_isset(cpu, smp_commenced_mask))
1328                         break;
1329                 msleep(100);
1330         }
1331         if (cpu_isset(cpu, smp_commenced_mask)) {
1332                 printk(KERN_ERR "CPU %u didn't die...\n", cpu);
1333         } else {
1334 #if defined(CONFIG_SUN_LDOMS)
1335                 unsigned long hv_err;
1336                 int limit = 100;
1337
1338                 do {
1339                         hv_err = sun4v_cpu_stop(cpu);
1340                         if (hv_err == HV_EOK) {
1341                                 cpu_clear(cpu, cpu_present_map);
1342                                 break;
1343                         }
1344                 } while (--limit > 0);
1345                 if (limit <= 0) {
1346                         printk(KERN_ERR "sun4v_cpu_stop() fails err=%lu\n",
1347                                hv_err);
1348                 }
1349 #endif
1350         }
1351 }
1352 #endif
1353
1354 void __init smp_cpus_done(unsigned int max_cpus)
1355 {
1356 }
1357
1358 void smp_send_reschedule(int cpu)
1359 {
1360         xcall_deliver((u64) &xcall_receive_signal, 0, 0,
1361                       &cpumask_of_cpu(cpu));
1362 }
1363
1364 void smp_receive_signal_client(int irq, struct pt_regs *regs)
1365 {
1366         clear_softint(1 << irq);
1367 }
1368
1369 /* This is a nop because we capture all other cpus
1370  * anyways when making the PROM active.
1371  */
1372 void smp_send_stop(void)
1373 {
1374 }
1375
1376 unsigned long __per_cpu_base __read_mostly;
1377 unsigned long __per_cpu_shift __read_mostly;
1378
1379 EXPORT_SYMBOL(__per_cpu_base);
1380 EXPORT_SYMBOL(__per_cpu_shift);
1381
1382 void __init real_setup_per_cpu_areas(void)
1383 {
1384         unsigned long paddr, goal, size, i;
1385         char *ptr;
1386
1387         /* Copy section for each CPU (we discard the original) */
1388         goal = PERCPU_ENOUGH_ROOM;
1389
1390         __per_cpu_shift = PAGE_SHIFT;
1391         for (size = PAGE_SIZE; size < goal; size <<= 1UL)
1392                 __per_cpu_shift++;
1393
1394         paddr = lmb_alloc(size * NR_CPUS, PAGE_SIZE);
1395         if (!paddr) {
1396                 prom_printf("Cannot allocate per-cpu memory.\n");
1397                 prom_halt();
1398         }
1399
1400         ptr = __va(paddr);
1401         __per_cpu_base = ptr - __per_cpu_start;
1402
1403         for (i = 0; i < NR_CPUS; i++, ptr += size)
1404                 memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start);
1405
1406         /* Setup %g5 for the boot cpu.  */
1407         __local_per_cpu_offset = __per_cpu_offset(smp_processor_id());
1408 }