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/export.h>
   7 #include <linux/kernel.h>
   8 #include <linux/sched.h>
   9 #include <linux/mm.h>
  10 #include <linux/pagemap.h>
  11 #include <linux/threads.h>
  12 #include <linux/smp.h>
  13 #include <linux/interrupt.h>
  14 #include <linux/kernel_stat.h>
  15 #include <linux/delay.h>
  16 #include <linux/init.h>
  17 #include <linux/spinlock.h>
  18 #include <linux/fs.h>
  19 #include <linux/seq_file.h>
  20 #include <linux/cache.h>
  21 #include <linux/jiffies.h>
  22 #include <linux/profile.h>
  23 #include <linux/bootmem.h>
  24 #include <linux/vmalloc.h>
  25 #include <linux/ftrace.h>
  26 #include <linux/cpu.h>
  27 #include <linux/slab.h>
  28
  29 #include <asm/head.h>
  30 #include <asm/ptrace.h>
  31 #include <linux/atomic.h>
  32 #include <asm/tlbflush.h>
  33 #include <asm/mmu_context.h>
  34 #include <asm/cpudata.h>
  35 #include <asm/hvtramp.h>
  36 #include <asm/io.h>
  37 #include <asm/timer.h>
  38
  39 #include <asm/irq.h>
  40 #include <asm/irq_regs.h>
  41 #include <asm/page.h>
  42 #include <asm/pgtable.h>
  43 #include <asm/oplib.h>
  44 #include <asm/uaccess.h>
  45 #include <asm/starfire.h>
  46 #include <asm/tlb.h>
  47 #include <asm/sections.h>
  48 #include <asm/prom.h>
  49 #include <asm/mdesc.h>
  50 #include <asm/ldc.h>
  51 #include <asm/hypervisor.h>
  52 #include <asm/pcr.h>
  53
  54 #include "cpumap.h"
  55
  56 int sparc64_multi_core __read_mostly;
  57
  58 DEFINE_PER_CPU(cpumask_t, cpu_sibling_map) = CPU_MASK_NONE;
  59 cpumask_t cpu_core_map[NR_CPUS] __read_mostly =
  60         { [0 ... NR_CPUS-1] = CPU_MASK_NONE };
  61
  62 EXPORT_PER_CPU_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 extern void setup_sparc64_timer(void);
  87
  88 static volatile unsigned long callin_flag = 0;
  89
  90 void __cpuinit smp_callin(void)
  91 {
  92         int cpuid = hard_smp_processor_id();
  93
  94         __local_per_cpu_offset = __per_cpu_offset(cpuid);
  95
  96         if (tlb_type == hypervisor)
  97                 sun4v_ktsb_register();
  98
  99         __flush_tlb_all();
 100
 101         setup_sparc64_timer();
 102
 103         if (cheetah_pcache_forced_on)
 104                 cheetah_enable_pcache();
 105
 106         callin_flag = 1;
 107         __asm__ __volatile__("membar #Sync\n\t"
 108                              "flush  %%g6" : : : "memory");
 109
 110         /* Clear this or we will die instantly when we
 111          * schedule back to this idler...
 112          */
 113         current_thread_info()->new_child = 0;
 114
 115         /* Attach to the address space of init_task. */
 116         atomic_inc(&init_mm.mm_count);
 117         current->active_mm = &init_mm;
 118
 119         /* inform the notifiers about the new cpu */
 120         notify_cpu_starting(cpuid);
 121
 122         while (!cpumask_test_cpu(cpuid, &smp_commenced_mask))
 123                 rmb();
 124
 125         set_cpu_online(cpuid, true);
 126         local_irq_enable();
 127
 128         /* idle thread is expected to have preempt disabled */
 129         preempt_disable();
 130 }
 131
 132 void cpu_panic(void)
 133 {
 134         printk("CPU[%d]: Returns from cpu_idle!\n", smp_processor_id());
 135         panic("SMP bolixed\n");
 136 }
 137
 138 /* This tick register synchronization scheme is taken entirely from
 139  * the ia64 port, see arch/ia64/kernel/smpboot.c for details and credit.
 140  *
 141  * The only change I've made is to rework it so that the master
 142  * initiates the synchonization instead of the slave. -DaveM
 143  */
 144
 145 #define MASTER  0
 146 #define SLAVE   (SMP_CACHE_BYTES/sizeof(unsigned long))
 147
 148 #define NUM_ROUNDS      64      /* magic value */
 149 #define NUM_ITERS       5       /* likewise */
 150
 151 static DEFINE_SPINLOCK(itc_sync_lock);
 152 static unsigned long go[SLAVE + 1];
 153
 154 #define DEBUG_TICK_SYNC 0
 155
 156 static inline long get_delta (long *rt, long *master)
 157 {
 158         unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
 159         unsigned long tcenter, t0, t1, tm;
 160         unsigned long i;
 161
 162         for (i = 0; i < NUM_ITERS; i++) {
 163                 t0 = tick_ops->get_tick();
 164                 go[MASTER] = 1;
 165                 membar_safe("#StoreLoad");
 166                 while (!(tm = go[SLAVE]))
 167                         rmb();
 168                 go[SLAVE] = 0;
 169                 wmb();
 170                 t1 = tick_ops->get_tick();
 171
 172                 if (t1 - t0 < best_t1 - best_t0)
 173                         best_t0 = t0, best_t1 = t1, best_tm = tm;
 174         }
 175
 176         *rt = best_t1 - best_t0;
 177         *master = best_tm - best_t0;
 178
 179         /* average best_t0 and best_t1 without overflow: */
 180         tcenter = (best_t0/2 + best_t1/2);
 181         if (best_t0 % 2 + best_t1 % 2 == 2)
 182                 tcenter++;
 183         return tcenter - best_tm;
 184 }
 185
 186 void smp_synchronize_tick_client(void)
 187 {
 188         long i, delta, adj, adjust_latency = 0, done = 0;
 189         unsigned long flags, rt, master_time_stamp;
 190 #if DEBUG_TICK_SYNC
 191         struct {
 192                 long rt;        /* roundtrip time */
 193                 long master;    /* master's timestamp */
 194                 long diff;      /* difference between midpoint and master's timestamp */
 195                 long lat;       /* estimate of itc adjustment latency */
 196         } t[NUM_ROUNDS];
 197 #endif
 198
 199         go[MASTER] = 1;
 200
 201         while (go[MASTER])
 202                 rmb();
 203
 204         local_irq_save(flags);
 205         {
 206                 for (i = 0; i < NUM_ROUNDS; i++) {
 207                         delta = get_delta(&rt, &master_time_stamp);
 208                         if (delta == 0)
 209                                 done = 1;       /* let's lock on to this... */
 210
 211                         if (!done) {
 212                                 if (i > 0) {
 213                                         adjust_latency += -delta;
 214                                         adj = -delta + adjust_latency/4;
 215                                 } else
 216                                         adj = -delta;
 217
 218                                 tick_ops->add_tick(adj);
 219                         }
 220 #if DEBUG_TICK_SYNC
 221                         t[i].rt = rt;
 222                         t[i].master = master_time_stamp;
 223                         t[i].diff = delta;
 224                         t[i].lat = adjust_latency/4;
 225 #endif
 226                 }
 227         }
 228         local_irq_restore(flags);
 229
 230 #if DEBUG_TICK_SYNC
 231         for (i = 0; i < NUM_ROUNDS; i++)
 232                 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
 233                        t[i].rt, t[i].master, t[i].diff, t[i].lat);
 234 #endif
 235
 236         printk(KERN_INFO "CPU %d: synchronized TICK with master CPU "
 237                "(last diff %ld cycles, maxerr %lu cycles)\n",
 238                smp_processor_id(), delta, rt);
 239 }
 240
 241 static void smp_start_sync_tick_client(int cpu);
 242
 243 static void smp_synchronize_one_tick(int cpu)
 244 {
 245         unsigned long flags, i;
 246
 247         go[MASTER] = 0;
 248
 249         smp_start_sync_tick_client(cpu);
 250
 251         /* wait for client to be ready */
 252         while (!go[MASTER])
 253                 rmb();
 254
 255         /* now let the client proceed into his loop */
 256         go[MASTER] = 0;
 257         membar_safe("#StoreLoad");
 258
 259         spin_lock_irqsave(&itc_sync_lock, flags);
 260         {
 261                 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; i++) {
 262                         while (!go[MASTER])
 263                                 rmb();
 264                         go[MASTER] = 0;
 265                         wmb();
 266                         go[SLAVE] = tick_ops->get_tick();
 267                         membar_safe("#StoreLoad");
 268                 }
 269         }
 270         spin_unlock_irqrestore(&itc_sync_lock, flags);
 271 }
 272
 273 #if defined(CONFIG_SUN_LDOMS) && defined(CONFIG_HOTPLUG_CPU)
 274 /* XXX Put this in some common place. XXX */
 275 static unsigned long kimage_addr_to_ra(void *p)
 276 {
 277         unsigned long val = (unsigned long) p;
 278
 279         return kern_base + (val - KERNBASE);
 280 }
 281
 282 static void __cpuinit ldom_startcpu_cpuid(unsigned int cpu, unsigned long thread_reg, void **descrp)
 283 {
 284         extern unsigned long sparc64_ttable_tl0;
 285         extern unsigned long kern_locked_tte_data;
 286         struct hvtramp_descr *hdesc;
 287         unsigned long trampoline_ra;
 288         struct trap_per_cpu *tb;
 289         u64 tte_vaddr, tte_data;
 290         unsigned long hv_err;
 291         int i;
 292
 293         hdesc = kzalloc(sizeof(*hdesc) +
 294                         (sizeof(struct hvtramp_mapping) *
 295                          num_kernel_image_mappings - 1),
 296                         GFP_KERNEL);
 297         if (!hdesc) {
 298                 printk(KERN_ERR "ldom_startcpu_cpuid: Cannot allocate "
 299                        "hvtramp_descr.\n");
 300                 return;
 301         }
 302         *descrp = hdesc;
 303
 304         hdesc->cpu = cpu;
 305         hdesc->num_mappings = num_kernel_image_mappings;
 306
 307         tb = &trap_block[cpu];
 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         for (i = 0; i < hdesc->num_mappings; i++) {
 318                 hdesc->maps[i].vaddr = tte_vaddr;
 319                 hdesc->maps[i].tte   = tte_data;
 320                 tte_vaddr += 0x400000;
 321                 tte_data  += 0x400000;
 322         }
 323
 324         trampoline_ra = kimage_addr_to_ra(hv_cpu_startup);
 325
 326         hv_err = sun4v_cpu_start(cpu, trampoline_ra,
 327                                  kimage_addr_to_ra(&sparc64_ttable_tl0),
 328                                  __pa(hdesc));
 329         if (hv_err)
 330                 printk(KERN_ERR "ldom_startcpu_cpuid: sun4v_cpu_start() "
 331                        "gives error %lu\n", hv_err);
 332 }
 333 #endif
 334
 335 extern unsigned long sparc64_cpu_startup;
 336
 337 /* The OBP cpu startup callback truncates the 3rd arg cookie to
 338  * 32-bits (I think) so to be safe we have it read the pointer
 339  * contained here so we work on >4GB machines. -DaveM
 340  */
 341 static struct thread_info *cpu_new_thread = NULL;
 342
 343 static int __cpuinit smp_boot_one_cpu(unsigned int cpu, struct task_struct *idle)
 344 {
 345         unsigned long entry =
 346                 (unsigned long)(&sparc64_cpu_startup);
 347         unsigned long cookie =
 348                 (unsigned long)(&cpu_new_thread);
 349         void *descr = NULL;
 350         int timeout, ret;
 351
 352         callin_flag = 0;
 353         cpu_new_thread = task_thread_info(idle);
 354
 355         if (tlb_type == hypervisor) {
 356 #if defined(CONFIG_SUN_LDOMS) && defined(CONFIG_HOTPLUG_CPU)
 357                 if (ldom_domaining_enabled)
 358                         ldom_startcpu_cpuid(cpu,
 359                                             (unsigned long) cpu_new_thread,
 360                                             &descr);
 361                 else
 362 #endif
 363                         prom_startcpu_cpuid(cpu, entry, cookie);
 364         } else {
 365                 struct device_node *dp = of_find_node_by_cpuid(cpu);
 366
 367                 prom_startcpu(dp->phandle, entry, cookie);
 368         }
 369
 370         for (timeout = 0; timeout < 50000; timeout++) {
 371                 if (callin_flag)
 372                         break;
 373                 udelay(100);
 374         }
 375
 376         if (callin_flag) {
 377                 ret = 0;
 378         } else {
 379                 printk("Processor %d is stuck.\n", cpu);
 380                 ret = -ENODEV;
 381         }
 382         cpu_new_thread = NULL;
 383
 384         kfree(descr);
 385
 386         return ret;
 387 }
 388
 389 static void spitfire_xcall_helper(u64 data0, u64 data1, u64 data2, u64 pstate, unsigned long cpu)
 390 {
 391         u64 result, target;
 392         int stuck, tmp;
 393
 394         if (this_is_starfire) {
 395                 /* map to real upaid */
 396                 cpu = (((cpu & 0x3c) << 1) |
 397                         ((cpu & 0x40) >> 4) |
 398                         (cpu & 0x3));
 399         }
 400
 401         target = (cpu << 14) | 0x70;
 402 again:
 403         /* Ok, this is the real Spitfire Errata #54.
 404          * One must read back from a UDB internal register
 405          * after writes to the UDB interrupt dispatch, but
 406          * before the membar Sync for that write.
 407          * So we use the high UDB control register (ASI 0x7f,
 408          * ADDR 0x20) for the dummy read. -DaveM
 409          */
 410         tmp = 0x40;
 411         __asm__ __volatile__(
 412         "wrpr   %1, %2, %%pstate\n\t"
 413         "stxa   %4, [%0] %3\n\t"
 414         "stxa   %5, [%0+%8] %3\n\t"
 415         "add    %0, %8, %0\n\t"
 416         "stxa   %6, [%0+%8] %3\n\t"
 417         "membar #Sync\n\t"
 418         "stxa   %%g0, [%7] %3\n\t"
 419         "membar #Sync\n\t"
 420         "mov    0x20, %%g1\n\t"
 421         "ldxa   [%%g1] 0x7f, %%g0\n\t"
 422         "membar #Sync"
 423         : "=r" (tmp)
 424         : "r" (pstate), "i" (PSTATE_IE), "i" (ASI_INTR_W),
 425           "r" (data0), "r" (data1), "r" (data2), "r" (target),
 426           "r" (0x10), "0" (tmp)
 427         : "g1");
 428
 429         /* NOTE: PSTATE_IE is still clear. */
 430         stuck = 100000;
 431         do {
 432                 __asm__ __volatile__("ldxa [%%g0] %1, %0"
 433                         : "=r" (result)
 434                         : "i" (ASI_INTR_DISPATCH_STAT));
 435                 if (result == 0) {
 436                         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 437                                              : : "r" (pstate));
 438                         return;
 439                 }
 440                 stuck -= 1;
 441                 if (stuck == 0)
 442                         break;
 443         } while (result & 0x1);
 444         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 445                              : : "r" (pstate));
 446         if (stuck == 0) {
 447                 printk("CPU[%d]: mondo stuckage result[%016llx]\n",
 448                        smp_processor_id(), result);
 449         } else {
 450                 udelay(2);
 451                 goto again;
 452         }
 453 }
 454
 455 static void spitfire_xcall_deliver(struct trap_per_cpu *tb, int cnt)
 456 {
 457         u64 *mondo, data0, data1, data2;
 458         u16 *cpu_list;
 459         u64 pstate;
 460         int i;
 461
 462         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 463         cpu_list = __va(tb->cpu_list_pa);
 464         mondo = __va(tb->cpu_mondo_block_pa);
 465         data0 = mondo[0];
 466         data1 = mondo[1];
 467         data2 = mondo[2];
 468         for (i = 0; i < cnt; i++)
 469                 spitfire_xcall_helper(data0, data1, data2, pstate, cpu_list[i]);
 470 }
 471
 472 /* Cheetah now allows to send the whole 64-bytes of data in the interrupt
 473  * packet, but we have no use for that.  However we do take advantage of
 474  * the new pipelining feature (ie. dispatch to multiple cpus simultaneously).
 475  */
 476 static void cheetah_xcall_deliver(struct trap_per_cpu *tb, int cnt)
 477 {
 478         int nack_busy_id, is_jbus, need_more;
 479         u64 *mondo, pstate, ver, busy_mask;
 480         u16 *cpu_list;
 481
 482         cpu_list = __va(tb->cpu_list_pa);
 483         mondo = __va(tb->cpu_mondo_block_pa);
 484
 485         /* Unfortunately, someone at Sun had the brilliant idea to make the
 486          * busy/nack fields hard-coded by ITID number for this Ultra-III
 487          * derivative processor.
 488          */
 489         __asm__ ("rdpr %%ver, %0" : "=r" (ver));
 490         is_jbus = ((ver >> 32) == __JALAPENO_ID ||
 491                    (ver >> 32) == __SERRANO_ID);
 492
 493         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 494
 495 retry:
 496         need_more = 0;
 497         __asm__ __volatile__("wrpr %0, %1, %%pstate\n\t"
 498                              : : "r" (pstate), "i" (PSTATE_IE));
 499
 500         /* Setup the dispatch data registers. */
 501         __asm__ __volatile__("stxa      %0, [%3] %6\n\t"
 502                              "stxa      %1, [%4] %6\n\t"
 503                              "stxa      %2, [%5] %6\n\t"
 504                              "membar    #Sync\n\t"
 505                              : /* no outputs */
 506                              : "r" (mondo[0]), "r" (mondo[1]), "r" (mondo[2]),
 507                                "r" (0x40), "r" (0x50), "r" (0x60),
 508                                "i" (ASI_INTR_W));
 509
 510         nack_busy_id = 0;
 511         busy_mask = 0;
 512         {
 513                 int i;
 514
 515                 for (i = 0; i < cnt; i++) {
 516                         u64 target, nr;
 517
 518                         nr = cpu_list[i];
 519                         if (nr == 0xffff)
 520                                 continue;
 521
 522                         target = (nr << 14) | 0x70;
 523                         if (is_jbus) {
 524                                 busy_mask |= (0x1UL << (nr * 2));
 525                         } else {
 526                                 target |= (nack_busy_id << 24);
 527                                 busy_mask |= (0x1UL <<
 528                                               (nack_busy_id * 2));
 529                         }
 530                         __asm__ __volatile__(
 531                                 "stxa   %%g0, [%0] %1\n\t"
 532                                 "membar #Sync\n\t"
 533                                 : /* no outputs */
 534                                 : "r" (target), "i" (ASI_INTR_W));
 535                         nack_busy_id++;
 536                         if (nack_busy_id == 32) {
 537                                 need_more = 1;
 538                                 break;
 539                         }
 540                 }
 541         }
 542
 543         /* Now, poll for completion. */
 544         {
 545                 u64 dispatch_stat, nack_mask;
 546                 long stuck;
 547
 548                 stuck = 100000 * nack_busy_id;
 549                 nack_mask = busy_mask << 1;
 550                 do {
 551                         __asm__ __volatile__("ldxa      [%%g0] %1, %0"
 552                                              : "=r" (dispatch_stat)
 553                                              : "i" (ASI_INTR_DISPATCH_STAT));
 554                         if (!(dispatch_stat & (busy_mask | nack_mask))) {
 555                                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 556                                                      : : "r" (pstate));
 557                                 if (unlikely(need_more)) {
 558                                         int i, this_cnt = 0;
 559                                         for (i = 0; i < cnt; i++) {
 560                                                 if (cpu_list[i] == 0xffff)
 561                                                         continue;
 562                                                 cpu_list[i] = 0xffff;
 563                                                 this_cnt++;
 564                                                 if (this_cnt == 32)
 565                                                         break;
 566                                         }
 567                                         goto retry;
 568                                 }
 569                                 return;
 570                         }
 571                         if (!--stuck)
 572                                 break;
 573                 } while (dispatch_stat & busy_mask);
 574
 575                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 576                                      : : "r" (pstate));
 577
 578                 if (dispatch_stat & busy_mask) {
 579                         /* Busy bits will not clear, continue instead
 580                          * of freezing up on this cpu.
 581                          */
 582                         printk("CPU[%d]: mondo stuckage result[%016llx]\n",
 583                                smp_processor_id(), dispatch_stat);
 584                 } else {
 585                         int i, this_busy_nack = 0;
 586
 587                         /* Delay some random time with interrupts enabled
 588                          * to prevent deadlock.
 589                          */
 590                         udelay(2 * nack_busy_id);
 591
 592                         /* Clear out the mask bits for cpus which did not
 593                          * NACK us.
 594                          */
 595                         for (i = 0; i < cnt; i++) {
 596                                 u64 check_mask, nr;
 597
 598                                 nr = cpu_list[i];
 599                                 if (nr == 0xffff)
 600                                         continue;
 601
 602                                 if (is_jbus)
 603                                         check_mask = (0x2UL << (2*nr));
 604                                 else
 605                                         check_mask = (0x2UL <<
 606                                                       this_busy_nack);
 607                                 if ((dispatch_stat & check_mask) == 0)
 608                                         cpu_list[i] = 0xffff;
 609                                 this_busy_nack += 2;
 610                                 if (this_busy_nack == 64)
 611                                         break;
 612                         }
 613
 614                         goto retry;
 615                 }
 616         }
 617 }
 618
 619 /* Multi-cpu list version.  */
 620 static void hypervisor_xcall_deliver(struct trap_per_cpu *tb, int cnt)
 621 {
 622         int retries, this_cpu, prev_sent, i, saw_cpu_error;
 623         unsigned long status;
 624         u16 *cpu_list;
 625
 626         this_cpu = smp_processor_id();
 627
 628         cpu_list = __va(tb->cpu_list_pa);
 629
 630         saw_cpu_error = 0;
 631         retries = 0;
 632         prev_sent = 0;
 633         do {
 634                 int forward_progress, n_sent;
 635
 636                 status = sun4v_cpu_mondo_send(cnt,
 637                                               tb->cpu_list_pa,
 638                                               tb->cpu_mondo_block_pa);
 639
 640                 /* HV_EOK means all cpus received the xcall, we're done.  */
 641                 if (likely(status == HV_EOK))
 642                         break;
 643
 644                 /* First, see if we made any forward progress.
 645                  *
 646                  * The hypervisor indicates successful sends by setting
 647                  * cpu list entries to the value 0xffff.
 648                  */
 649                 n_sent = 0;
 650                 for (i = 0; i < cnt; i++) {
 651                         if (likely(cpu_list[i] == 0xffff))
 652                                 n_sent++;
 653                 }
 654
 655                 forward_progress = 0;
 656                 if (n_sent > prev_sent)
 657                         forward_progress = 1;
 658
 659                 prev_sent = n_sent;
 660
 661                 /* If we get a HV_ECPUERROR, then one or more of the cpus
 662                  * in the list are in error state.  Use the cpu_state()
 663                  * hypervisor call to find out which cpus are in error state.
 664                  */
 665                 if (unlikely(status == HV_ECPUERROR)) {
 666                         for (i = 0; i < cnt; i++) {
 667                                 long err;
 668                                 u16 cpu;
 669
 670                                 cpu = cpu_list[i];
 671                                 if (cpu == 0xffff)
 672                                         continue;
 673
 674                                 err = sun4v_cpu_state(cpu);
 675                                 if (err == HV_CPU_STATE_ERROR) {
 676                                         saw_cpu_error = (cpu + 1);
 677                                         cpu_list[i] = 0xffff;
 678                                 }
 679                         }
 680                 } else if (unlikely(status != HV_EWOULDBLOCK))
 681                         goto fatal_mondo_error;
 682
 683                 /* Don't bother rewriting the CPU list, just leave the
 684                  * 0xffff and non-0xffff entries in there and the
 685                  * hypervisor will do the right thing.
 686                  *
 687                  * Only advance timeout state if we didn't make any
 688                  * forward progress.
 689                  */
 690                 if (unlikely(!forward_progress)) {
 691                         if (unlikely(++retries > 10000))
 692                                 goto fatal_mondo_timeout;
 693
 694                         /* Delay a little bit to let other cpus catch up
 695                          * on their cpu mondo queue work.
 696                          */
 697                         udelay(2 * cnt);
 698                 }
 699         } while (1);
 700
 701         if (unlikely(saw_cpu_error))
 702                 goto fatal_mondo_cpu_error;
 703
 704         return;
 705
 706 fatal_mondo_cpu_error:
 707         printk(KERN_CRIT "CPU[%d]: SUN4V mondo cpu error, some target cpus "
 708                "(including %d) were in error state\n",
 709                this_cpu, saw_cpu_error - 1);
 710         return;
 711
 712 fatal_mondo_timeout:
 713         printk(KERN_CRIT "CPU[%d]: SUN4V mondo timeout, no forward "
 714                " progress after %d retries.\n",
 715                this_cpu, retries);
 716         goto dump_cpu_list_and_out;
 717
 718 fatal_mondo_error:
 719         printk(KERN_CRIT "CPU[%d]: Unexpected SUN4V mondo error %lu\n",
 720                this_cpu, status);
 721         printk(KERN_CRIT "CPU[%d]: Args were cnt(%d) cpulist_pa(%lx) "
 722                "mondo_block_pa(%lx)\n",
 723                this_cpu, cnt, tb->cpu_list_pa, tb->cpu_mondo_block_pa);
 724
 725 dump_cpu_list_and_out:
 726         printk(KERN_CRIT "CPU[%d]: CPU list [ ", this_cpu);
 727         for (i = 0; i < cnt; i++)
 728                 printk("%u ", cpu_list[i]);
 729         printk("]\n");
 730 }
 731
 732 static void (*xcall_deliver_impl)(struct trap_per_cpu *, int);
 733
 734 static void xcall_deliver(u64 data0, u64 data1, u64 data2, const cpumask_t *mask)
 735 {
 736         struct trap_per_cpu *tb;
 737         int this_cpu, i, cnt;
 738         unsigned long flags;
 739         u16 *cpu_list;
 740         u64 *mondo;
 741
 742         /* We have to do this whole thing with interrupts fully disabled.
 743          * Otherwise if we send an xcall from interrupt context it will
 744          * corrupt both our mondo block and cpu list state.
 745          *
 746          * One consequence of this is that we cannot use timeout mechanisms
 747          * that depend upon interrupts being delivered locally.  So, for
 748          * example, we cannot sample jiffies and expect it to advance.
 749          *
 750          * Fortunately, udelay() uses %stick/%tick so we can use that.
 751          */
 752         local_irq_save(flags);
 753
 754         this_cpu = smp_processor_id();
 755         tb = &trap_block[this_cpu];
 756
 757         mondo = __va(tb->cpu_mondo_block_pa);
 758         mondo[0] = data0;
 759         mondo[1] = data1;
 760         mondo[2] = data2;
 761         wmb();
 762
 763         cpu_list = __va(tb->cpu_list_pa);
 764
 765         /* Setup the initial cpu list.  */
 766         cnt = 0;
 767         for_each_cpu(i, mask) {
 768                 if (i == this_cpu || !cpu_online(i))
 769                         continue;
 770                 cpu_list[cnt++] = i;
 771         }
 772
 773         if (cnt)
 774                 xcall_deliver_impl(tb, cnt);
 775
 776         local_irq_restore(flags);
 777 }
 778
 779 /* Send cross call to all processors mentioned in MASK_P
 780  * except self.  Really, there are only two cases currently,
 781  * "cpu_online_mask" and "mm_cpumask(mm)".
 782  */
 783 static void smp_cross_call_masked(unsigned long *func, u32 ctx, u64 data1, u64 data2, const cpumask_t *mask)
 784 {
 785         u64 data0 = (((u64)ctx)<<32 | (((u64)func) & 0xffffffff));
 786
 787         xcall_deliver(data0, data1, data2, mask);
 788 }
 789
 790 /* Send cross call to all processors except self. */
 791 static void smp_cross_call(unsigned long *func, u32 ctx, u64 data1, u64 data2)
 792 {
 793         smp_cross_call_masked(func, ctx, data1, data2, cpu_online_mask);
 794 }
 795
 796 extern unsigned long xcall_sync_tick;
 797
 798 static void smp_start_sync_tick_client(int cpu)
 799 {
 800         xcall_deliver((u64) &xcall_sync_tick, 0, 0,
 801                       cpumask_of(cpu));
 802 }
 803
 804 extern unsigned long xcall_call_function;
 805
 806 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
 807 {
 808         xcall_deliver((u64) &xcall_call_function, 0, 0, mask);
 809 }
 810
 811 extern unsigned long xcall_call_function_single;
 812
 813 void arch_send_call_function_single_ipi(int cpu)
 814 {
 815         xcall_deliver((u64) &xcall_call_function_single, 0, 0,
 816                       cpumask_of(cpu));
 817 }
 818
 819 void __irq_entry smp_call_function_client(int irq, struct pt_regs *regs)
 820 {
 821         clear_softint(1 << irq);
 822         generic_smp_call_function_interrupt();
 823 }
 824
 825 void __irq_entry smp_call_function_single_client(int irq, struct pt_regs *regs)
 826 {
 827         clear_softint(1 << irq);
 828         generic_smp_call_function_single_interrupt();
 829 }
 830
 831 static void tsb_sync(void *info)
 832 {
 833         struct trap_per_cpu *tp = &trap_block[raw_smp_processor_id()];
 834         struct mm_struct *mm = info;
 835
 836         /* It is not valid to test "current->active_mm == mm" here.
 837          *
 838          * The value of "current" is not changed atomically with
 839          * switch_mm().  But that's OK, we just need to check the
 840          * current cpu's trap block PGD physical address.
 841          */
 842         if (tp->pgd_paddr == __pa(mm->pgd))
 843                 tsb_context_switch(mm);
 844 }
 845
 846 void smp_tsb_sync(struct mm_struct *mm)
 847 {
 848         smp_call_function_many(mm_cpumask(mm), tsb_sync, mm, 1);
 849 }
 850
 851 extern unsigned long xcall_flush_tlb_mm;
 852 extern unsigned long xcall_flush_tlb_pending;
 853 extern unsigned long xcall_flush_tlb_kernel_range;
 854 extern unsigned long xcall_fetch_glob_regs;
 855 extern unsigned long xcall_fetch_glob_pmu;
 856 extern unsigned long xcall_fetch_glob_pmu_n4;
 857 extern unsigned long xcall_receive_signal;
 858 extern unsigned long xcall_new_mmu_context_version;
 859 #ifdef CONFIG_KGDB
 860 extern unsigned long xcall_kgdb_capture;
 861 #endif
 862
 863 #ifdef DCACHE_ALIASING_POSSIBLE
 864 extern unsigned long xcall_flush_dcache_page_cheetah;
 865 #endif
 866 extern unsigned long xcall_flush_dcache_page_spitfire;
 867
 868 #ifdef CONFIG_DEBUG_DCFLUSH
 869 extern atomic_t dcpage_flushes;
 870 extern atomic_t dcpage_flushes_xcall;
 871 #endif
 872
 873 static inline void __local_flush_dcache_page(struct page *page)
 874 {
 875 #ifdef DCACHE_ALIASING_POSSIBLE
 876         __flush_dcache_page(page_address(page),
 877                             ((tlb_type == spitfire) &&
 878                              page_mapping(page) != NULL));
 879 #else
 880         if (page_mapping(page) != NULL &&
 881             tlb_type == spitfire)
 882                 __flush_icache_page(__pa(page_address(page)));
 883 #endif
 884 }
 885
 886 void smp_flush_dcache_page_impl(struct page *page, int cpu)
 887 {
 888         int this_cpu;
 889
 890         if (tlb_type == hypervisor)
 891                 return;
 892
 893 #ifdef CONFIG_DEBUG_DCFLUSH
 894         atomic_inc(&dcpage_flushes);
 895 #endif
 896
 897         this_cpu = get_cpu();
 898
 899         if (cpu == this_cpu) {
 900                 __local_flush_dcache_page(page);
 901         } else if (cpu_online(cpu)) {
 902                 void *pg_addr = page_address(page);
 903                 u64 data0 = 0;
 904
 905                 if (tlb_type == spitfire) {
 906                         data0 = ((u64)&xcall_flush_dcache_page_spitfire);
 907                         if (page_mapping(page) != NULL)
 908                                 data0 |= ((u64)1 << 32);
 909                 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 910 #ifdef DCACHE_ALIASING_POSSIBLE
 911                         data0 = ((u64)&xcall_flush_dcache_page_cheetah);
 912 #endif
 913                 }
 914                 if (data0) {
 915                         xcall_deliver(data0, __pa(pg_addr),
 916                                       (u64) pg_addr, cpumask_of(cpu));
 917 #ifdef CONFIG_DEBUG_DCFLUSH
 918                         atomic_inc(&dcpage_flushes_xcall);
 919 #endif
 920                 }
 921         }
 922
 923         put_cpu();
 924 }
 925
 926 void flush_dcache_page_all(struct mm_struct *mm, struct page *page)
 927 {
 928         void *pg_addr;
 929         u64 data0;
 930
 931         if (tlb_type == hypervisor)
 932                 return;
 933
 934         preempt_disable();
 935
 936 #ifdef CONFIG_DEBUG_DCFLUSH
 937         atomic_inc(&dcpage_flushes);
 938 #endif
 939         data0 = 0;
 940         pg_addr = page_address(page);
 941         if (tlb_type == spitfire) {
 942                 data0 = ((u64)&xcall_flush_dcache_page_spitfire);
 943                 if (page_mapping(page) != NULL)
 944                         data0 |= ((u64)1 << 32);
 945         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 946 #ifdef DCACHE_ALIASING_POSSIBLE
 947                 data0 = ((u64)&xcall_flush_dcache_page_cheetah);
 948 #endif
 949         }
 950         if (data0) {
 951                 xcall_deliver(data0, __pa(pg_addr),
 952                               (u64) pg_addr, cpu_online_mask);
 953 #ifdef CONFIG_DEBUG_DCFLUSH
 954                 atomic_inc(&dcpage_flushes_xcall);
 955 #endif
 956         }
 957         __local_flush_dcache_page(page);
 958
 959         preempt_enable();
 960 }
 961
 962 void __irq_entry smp_new_mmu_context_version_client(int irq, struct pt_regs *regs)
 963 {
 964         struct mm_struct *mm;
 965         unsigned long flags;
 966
 967         clear_softint(1 << irq);
 968
 969         /* See if we need to allocate a new TLB context because
 970          * the version of the one we are using is now out of date.
 971          */
 972         mm = current->active_mm;
 973         if (unlikely(!mm || (mm == &init_mm)))
 974                 return;
 975
 976         spin_lock_irqsave(&mm->context.lock, flags);
 977
 978         if (unlikely(!CTX_VALID(mm->context)))
 979                 get_new_mmu_context(mm);
 980
 981         spin_unlock_irqrestore(&mm->context.lock, flags);
 982
 983         load_secondary_context(mm);
 984         __flush_tlb_mm(CTX_HWBITS(mm->context),
 985                        SECONDARY_CONTEXT);
 986 }
 987
 988 void smp_new_mmu_context_version(void)
 989 {
 990         smp_cross_call(&xcall_new_mmu_context_version, 0, 0, 0);
 991 }
 992
 993 #ifdef CONFIG_KGDB
 994 void kgdb_roundup_cpus(unsigned long flags)
 995 {
 996         smp_cross_call(&xcall_kgdb_capture, 0, 0, 0);
 997 }
 998 #endif
 999
1000 void smp_fetch_global_regs(void)
1001 {
1002         smp_cross_call(&xcall_fetch_glob_regs, 0, 0, 0);
1003 }
1004
1005 void smp_fetch_global_pmu(void)
1006 {
1007         if (tlb_type == hypervisor &&
1008             sun4v_chip_type >= SUN4V_CHIP_NIAGARA4)
1009                 smp_cross_call(&xcall_fetch_glob_pmu_n4, 0, 0, 0);
1010         else
1011                 smp_cross_call(&xcall_fetch_glob_pmu, 0, 0, 0);
1012 }
1013
1014 /* We know that the window frames of the user have been flushed
1015  * to the stack before we get here because all callers of us
1016  * are flush_tlb_*() routines, and these run after flush_cache_*()
1017  * which performs the flushw.
1018  *
1019  * The SMP TLB coherency scheme we use works as follows:
1020  *
1021  * 1) mm->cpu_vm_mask is a bit mask of which cpus an address
1022  *    space has (potentially) executed on, this is the heuristic
1023  *    we use to avoid doing cross calls.
1024  *
1025  *    Also, for flushing from kswapd and also for clones, we
1026  *    use cpu_vm_mask as the list of cpus to make run the TLB.
1027  *
1028  * 2) TLB context numbers are shared globally across all processors
1029  *    in the system, this allows us to play several games to avoid
1030  *    cross calls.
1031  *
1032  *    One invariant is that when a cpu switches to a process, and
1033  *    that processes tsk->active_mm->cpu_vm_mask does not have the
1034  *    current cpu's bit set, that tlb context is flushed locally.
1035  *
1036  *    If the address space is non-shared (ie. mm->count == 1) we avoid
1037  *    cross calls when we want to flush the currently running process's
1038  *    tlb state.  This is done by clearing all cpu bits except the current
1039  *    processor's in current->mm->cpu_vm_mask and performing the
1040  *    flush locally only.  This will force any subsequent cpus which run
1041  *    this task to flush the context from the local tlb if the process
1042  *    migrates to another cpu (again).
1043  *
1044  * 3) For shared address spaces (threads) and swapping we bite the
1045  *    bullet for most cases and perform the cross call (but only to
1046  *    the cpus listed in cpu_vm_mask).
1047  *
1048  *    The performance gain from "optimizing" away the cross call for threads is
1049  *    questionable (in theory the big win for threads is the massive sharing of
1050  *    address space state across processors).
1051  */
1052
1053 /* This currently is only used by the hugetlb arch pre-fault
1054  * hook on UltraSPARC-III+ and later when changing the pagesize
1055  * bits of the context register for an address space.
1056  */
1057 void smp_flush_tlb_mm(struct mm_struct *mm)
1058 {
1059         u32 ctx = CTX_HWBITS(mm->context);
1060         int cpu = get_cpu();
1061
1062         if (atomic_read(&mm->mm_users) == 1) {
1063                 cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
1064                 goto local_flush_and_out;
1065         }
1066
1067         smp_cross_call_masked(&xcall_flush_tlb_mm,
1068                               ctx, 0, 0,
1069                               mm_cpumask(mm));
1070
1071 local_flush_and_out:
1072         __flush_tlb_mm(ctx, SECONDARY_CONTEXT);
1073
1074         put_cpu();
1075 }
1076
1077 void smp_flush_tlb_pending(struct mm_struct *mm, unsigned long nr, unsigned long *vaddrs)
1078 {
1079         u32 ctx = CTX_HWBITS(mm->context);
1080         int cpu = get_cpu();
1081
1082         if (mm == current->mm && atomic_read(&mm->mm_users) == 1)
1083                 cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
1084         else
1085                 smp_cross_call_masked(&xcall_flush_tlb_pending,
1086                                       ctx, nr, (unsigned long) vaddrs,
1087                                       mm_cpumask(mm));
1088
1089         __flush_tlb_pending(ctx, nr, vaddrs);
1090
1091         put_cpu();
1092 }
1093
1094 void smp_flush_tlb_kernel_range(unsigned long start, unsigned long end)
1095 {
1096         start &= PAGE_MASK;
1097         end    = PAGE_ALIGN(end);
1098         if (start != end) {
1099                 smp_cross_call(&xcall_flush_tlb_kernel_range,
1100                                0, start, end);
1101
1102                 __flush_tlb_kernel_range(start, end);
1103         }
1104 }
1105
1106 /* CPU capture. */
1107 /* #define CAPTURE_DEBUG */
1108 extern unsigned long xcall_capture;
1109
1110 static atomic_t smp_capture_depth = ATOMIC_INIT(0);
1111 static atomic_t smp_capture_registry = ATOMIC_INIT(0);
1112 static unsigned long penguins_are_doing_time;
1113
1114 void smp_capture(void)
1115 {
1116         int result = atomic_add_ret(1, &smp_capture_depth);
1117
1118         if (result == 1) {
1119                 int ncpus = num_online_cpus();
1120
1121 #ifdef CAPTURE_DEBUG
1122                 printk("CPU[%d]: Sending penguins to jail...",
1123                        smp_processor_id());
1124 #endif
1125                 penguins_are_doing_time = 1;
1126                 atomic_inc(&smp_capture_registry);
1127                 smp_cross_call(&xcall_capture, 0, 0, 0);
1128                 while (atomic_read(&smp_capture_registry) != ncpus)
1129                         rmb();
1130 #ifdef CAPTURE_DEBUG
1131                 printk("done\n");
1132 #endif
1133         }
1134 }
1135
1136 void smp_release(void)
1137 {
1138         if (atomic_dec_and_test(&smp_capture_depth)) {
1139 #ifdef CAPTURE_DEBUG
1140                 printk("CPU[%d]: Giving pardon to "
1141                        "imprisoned penguins\n",
1142                        smp_processor_id());
1143 #endif
1144                 penguins_are_doing_time = 0;
1145                 membar_safe("#StoreLoad");
1146                 atomic_dec(&smp_capture_registry);
1147         }
1148 }
1149
1150 /* Imprisoned penguins run with %pil == PIL_NORMAL_MAX, but PSTATE_IE
1151  * set, so they can service tlb flush xcalls...
1152  */
1153 extern void prom_world(int);
1154
1155 void __irq_entry smp_penguin_jailcell(int irq, struct pt_regs *regs)
1156 {
1157         clear_softint(1 << irq);
1158
1159         preempt_disable();
1160
1161         __asm__ __volatile__("flushw");
1162         prom_world(1);
1163         atomic_inc(&smp_capture_registry);
1164         membar_safe("#StoreLoad");
1165         while (penguins_are_doing_time)
1166                 rmb();
1167         atomic_dec(&smp_capture_registry);
1168         prom_world(0);
1169
1170         preempt_enable();
1171 }
1172
1173 /* /proc/profile writes can call this, don't __init it please. */
1174 int setup_profiling_timer(unsigned int multiplier)
1175 {
1176         return -EINVAL;
1177 }
1178
1179 void __init smp_prepare_cpus(unsigned int max_cpus)
1180 {
1181 }
1182
1183 void __devinit smp_prepare_boot_cpu(void)
1184 {
1185 }
1186
1187 void __init smp_setup_processor_id(void)
1188 {
1189         if (tlb_type == spitfire)
1190                 xcall_deliver_impl = spitfire_xcall_deliver;
1191         else if (tlb_type == cheetah || tlb_type == cheetah_plus)
1192                 xcall_deliver_impl = cheetah_xcall_deliver;
1193         else
1194                 xcall_deliver_impl = hypervisor_xcall_deliver;
1195 }
1196
1197 void __devinit smp_fill_in_sib_core_maps(void)
1198 {
1199         unsigned int i;
1200
1201         for_each_present_cpu(i) {
1202                 unsigned int j;
1203
1204                 cpumask_clear(&cpu_core_map[i]);
1205                 if (cpu_data(i).core_id == 0) {
1206                         cpumask_set_cpu(i, &cpu_core_map[i]);
1207                         continue;
1208                 }
1209
1210                 for_each_present_cpu(j) {
1211                         if (cpu_data(i).core_id ==
1212                             cpu_data(j).core_id)
1213                                 cpumask_set_cpu(j, &cpu_core_map[i]);
1214                 }
1215         }
1216
1217         for_each_present_cpu(i) {
1218                 unsigned int j;
1219
1220                 cpumask_clear(&per_cpu(cpu_sibling_map, i));
1221                 if (cpu_data(i).proc_id == -1) {
1222                         cpumask_set_cpu(i, &per_cpu(cpu_sibling_map, i));
1223                         continue;
1224                 }
1225
1226                 for_each_present_cpu(j) {
1227                         if (cpu_data(i).proc_id ==
1228                             cpu_data(j).proc_id)
1229                                 cpumask_set_cpu(j, &per_cpu(cpu_sibling_map, i));
1230                 }
1231         }
1232 }
1233
1234 int __cpuinit __cpu_up(unsigned int cpu, struct task_struct *tidle)
1235 {
1236         int ret = smp_boot_one_cpu(cpu, tidle);
1237
1238         if (!ret) {
1239                 cpumask_set_cpu(cpu, &smp_commenced_mask);
1240                 while (!cpu_online(cpu))
1241                         mb();
1242                 if (!cpu_online(cpu)) {
1243                         ret = -ENODEV;
1244                 } else {
1245                         /* On SUN4V, writes to %tick and %stick are
1246                          * not allowed.
1247                          */
1248                         if (tlb_type != hypervisor)
1249                                 smp_synchronize_one_tick(cpu);
1250                 }
1251         }
1252         return ret;
1253 }
1254
1255 #ifdef CONFIG_HOTPLUG_CPU
1256 void cpu_play_dead(void)
1257 {
1258         int cpu = smp_processor_id();
1259         unsigned long pstate;
1260
1261         idle_task_exit();
1262
1263         if (tlb_type == hypervisor) {
1264                 struct trap_per_cpu *tb = &trap_block[cpu];
1265
1266                 sun4v_cpu_qconf(HV_CPU_QUEUE_CPU_MONDO,
1267                                 tb->cpu_mondo_pa, 0);
1268                 sun4v_cpu_qconf(HV_CPU_QUEUE_DEVICE_MONDO,
1269                                 tb->dev_mondo_pa, 0);
1270                 sun4v_cpu_qconf(HV_CPU_QUEUE_RES_ERROR,
1271                                 tb->resum_mondo_pa, 0);
1272                 sun4v_cpu_qconf(HV_CPU_QUEUE_NONRES_ERROR,
1273                                 tb->nonresum_mondo_pa, 0);
1274         }
1275
1276         cpumask_clear_cpu(cpu, &smp_commenced_mask);
1277         membar_safe("#Sync");
1278
1279         local_irq_disable();
1280
1281         __asm__ __volatile__(
1282                 "rdpr   %%pstate, %0\n\t"
1283                 "wrpr   %0, %1, %%pstate"
1284                 : "=r" (pstate)
1285                 : "i" (PSTATE_IE));
1286
1287         while (1)
1288                 barrier();
1289 }
1290
1291 int __cpu_disable(void)
1292 {
1293         int cpu = smp_processor_id();
1294         cpuinfo_sparc *c;
1295         int i;
1296
1297         for_each_cpu(i, &cpu_core_map[cpu])
1298                 cpumask_clear_cpu(cpu, &cpu_core_map[i]);
1299         cpumask_clear(&cpu_core_map[cpu]);
1300
1301         for_each_cpu(i, &per_cpu(cpu_sibling_map, cpu))
1302                 cpumask_clear_cpu(cpu, &per_cpu(cpu_sibling_map, i));
1303         cpumask_clear(&per_cpu(cpu_sibling_map, cpu));
1304
1305         c = &cpu_data(cpu);
1306
1307         c->core_id = 0;
1308         c->proc_id = -1;
1309
1310         smp_wmb();
1311
1312         /* Make sure no interrupts point to this cpu.  */
1313         fixup_irqs();
1314
1315         local_irq_enable();
1316         mdelay(1);
1317         local_irq_disable();
1318
1319         set_cpu_online(cpu, false);
1320
1321         cpu_map_rebuild();
1322
1323         return 0;
1324 }
1325
1326 void __cpu_die(unsigned int cpu)
1327 {
1328         int i;
1329
1330         for (i = 0; i < 100; i++) {
1331                 smp_rmb();
1332                 if (!cpumask_test_cpu(cpu, &smp_commenced_mask))
1333                         break;
1334                 msleep(100);
1335         }
1336         if (cpumask_test_cpu(cpu, &smp_commenced_mask)) {
1337                 printk(KERN_ERR "CPU %u didn't die...\n", cpu);
1338         } else {
1339 #if defined(CONFIG_SUN_LDOMS)
1340                 unsigned long hv_err;
1341                 int limit = 100;
1342
1343                 do {
1344                         hv_err = sun4v_cpu_stop(cpu);
1345                         if (hv_err == HV_EOK) {
1346                                 set_cpu_present(cpu, false);
1347                                 break;
1348                         }
1349                 } while (--limit > 0);
1350                 if (limit <= 0) {
1351                         printk(KERN_ERR "sun4v_cpu_stop() fails err=%lu\n",
1352                                hv_err);
1353                 }
1354 #endif
1355         }
1356 }
1357 #endif
1358
1359 void __init smp_cpus_done(unsigned int max_cpus)
1360 {
1361         pcr_arch_init();
1362 }
1363
1364 void smp_send_reschedule(int cpu)
1365 {
1366         xcall_deliver((u64) &xcall_receive_signal, 0, 0,
1367                       cpumask_of(cpu));
1368 }
1369
1370 void __irq_entry smp_receive_signal_client(int irq, struct pt_regs *regs)
1371 {
1372         clear_softint(1 << irq);
1373         scheduler_ipi();
1374 }
1375
1376 /* This is a nop because we capture all other cpus
1377  * anyways when making the PROM active.
1378  */
1379 void smp_send_stop(void)
1380 {
1381 }
1382
1383 /**
1384  * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu
1385  * @cpu: cpu to allocate for
1386  * @size: size allocation in bytes
1387  * @align: alignment
1388  *
1389  * Allocate @size bytes aligned at @align for cpu @cpu.  This wrapper
1390  * does the right thing for NUMA regardless of the current
1391  * configuration.
1392  *
1393  * RETURNS:
1394  * Pointer to the allocated area on success, NULL on failure.
1395  */
1396 static void * __init pcpu_alloc_bootmem(unsigned int cpu, size_t size,
1397                                         size_t align)
1398 {
1399         const unsigned long goal = __pa(MAX_DMA_ADDRESS);
1400 #ifdef CONFIG_NEED_MULTIPLE_NODES
1401         int node = cpu_to_node(cpu);
1402         void *ptr;
1403
1404         if (!node_online(node) || !NODE_DATA(node)) {
1405                 ptr = __alloc_bootmem(size, align, goal);
1406                 pr_info("cpu %d has no node %d or node-local memory\n",
1407                         cpu, node);
1408                 pr_debug("per cpu data for cpu%d %lu bytes at %016lx\n",
1409                          cpu, size, __pa(ptr));
1410         } else {
1411                 ptr = __alloc_bootmem_node(NODE_DATA(node),
1412                                            size, align, goal);
1413                 pr_debug("per cpu data for cpu%d %lu bytes on node%d at "
1414                          "%016lx\n", cpu, size, node, __pa(ptr));
1415         }
1416         return ptr;
1417 #else
1418         return __alloc_bootmem(size, align, goal);
1419 #endif
1420 }
1421
1422 static void __init pcpu_free_bootmem(void *ptr, size_t size)
1423 {
1424         free_bootmem(__pa(ptr), size);
1425 }
1426
1427 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
1428 {
1429         if (cpu_to_node(from) == cpu_to_node(to))
1430                 return LOCAL_DISTANCE;
1431         else
1432                 return REMOTE_DISTANCE;
1433 }
1434
1435 static void __init pcpu_populate_pte(unsigned long addr)
1436 {
1437         pgd_t *pgd = pgd_offset_k(addr);
1438         pud_t *pud;
1439         pmd_t *pmd;
1440
1441         pud = pud_offset(pgd, addr);
1442         if (pud_none(*pud)) {
1443                 pmd_t *new;
1444
1445                 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1446                 pud_populate(&init_mm, pud, new);
1447         }
1448
1449         pmd = pmd_offset(pud, addr);
1450         if (!pmd_present(*pmd)) {
1451                 pte_t *new;
1452
1453                 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1454                 pmd_populate_kernel(&init_mm, pmd, new);
1455         }
1456 }
1457
1458 void __init setup_per_cpu_areas(void)
1459 {
1460         unsigned long delta;
1461         unsigned int cpu;
1462         int rc = -EINVAL;
1463
1464         if (pcpu_chosen_fc != PCPU_FC_PAGE) {
1465                 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
1466                                             PERCPU_DYNAMIC_RESERVE, 4 << 20,
1467                                             pcpu_cpu_distance,
1468                                             pcpu_alloc_bootmem,
1469                                             pcpu_free_bootmem);
1470                 if (rc)
1471                         pr_warning("PERCPU: %s allocator failed (%d), "
1472                                    "falling back to page size\n",
1473                                    pcpu_fc_names[pcpu_chosen_fc], rc);
1474         }
1475         if (rc < 0)
1476                 rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE,
1477                                            pcpu_alloc_bootmem,
1478                                            pcpu_free_bootmem,
1479                                            pcpu_populate_pte);
1480         if (rc < 0)
1481                 panic("cannot initialize percpu area (err=%d)", rc);
1482
1483         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1484         for_each_possible_cpu(cpu)
1485                 __per_cpu_offset(cpu) = delta + pcpu_unit_offsets[cpu];
1486
1487         /* Setup %g5 for the boot cpu.  */
1488         __local_per_cpu_offset = __per_cpu_offset(smp_processor_id());
1489
1490         of_fill_in_cpu_data();
1491         if (tlb_type == hypervisor)
1492                 mdesc_fill_in_cpu_data(cpu_all_mask);
1493 }