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 #include <linux/kgdb.h>
  29
  30 #include <asm/head.h>
  31 #include <asm/ptrace.h>
  32 #include <linux/atomic.h>
  33 #include <asm/tlbflush.h>
  34 #include <asm/mmu_context.h>
  35 #include <asm/cpudata.h>
  36 #include <asm/hvtramp.h>
  37 #include <asm/io.h>
  38 #include <asm/timer.h>
  39 #include <asm/setup.h>
  40
  41 #include <asm/irq.h>
  42 #include <asm/irq_regs.h>
  43 #include <asm/page.h>
  44 #include <asm/pgtable.h>
  45 #include <asm/oplib.h>
  46 #include <asm/uaccess.h>
  47 #include <asm/starfire.h>
  48 #include <asm/tlb.h>
  49 #include <asm/sections.h>
  50 #include <asm/prom.h>
  51 #include <asm/mdesc.h>
  52 #include <asm/ldc.h>
  53 #include <asm/hypervisor.h>
  54 #include <asm/pcr.h>
  55
  56 #include "cpumap.h"
  57 #include "kernel.h"
  58
  59 DEFINE_PER_CPU(cpumask_t, cpu_sibling_map) = CPU_MASK_NONE;
  60 cpumask_t cpu_core_map[NR_CPUS] __read_mostly =
  61         { [0 ... NR_CPUS-1] = CPU_MASK_NONE };
  62
  63 cpumask_t cpu_core_sib_map[NR_CPUS] __read_mostly = {
  64         [0 ... NR_CPUS-1] = CPU_MASK_NONE };
  65
  66 cpumask_t cpu_core_sib_cache_map[NR_CPUS] __read_mostly = {
  67         [0 ... NR_CPUS - 1] = CPU_MASK_NONE };
  68
  69 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
  70 EXPORT_SYMBOL(cpu_core_map);
  71 EXPORT_SYMBOL(cpu_core_sib_map);
  72 EXPORT_SYMBOL(cpu_core_sib_cache_map);
  73
  74 static cpumask_t smp_commenced_mask;
  75
  76 void smp_info(struct seq_file *m)
  77 {
  78         int i;
  79
  80         seq_printf(m, "State:\n");
  81         for_each_online_cpu(i)
  82                 seq_printf(m, "CPU%d:\t\tonline\n", i);
  83 }
  84
  85 void smp_bogo(struct seq_file *m)
  86 {
  87         int i;
  88
  89         for_each_online_cpu(i)
  90                 seq_printf(m,
  91                            "Cpu%dClkTck\t: %016lx\n",
  92                            i, cpu_data(i).clock_tick);
  93 }
  94
  95 extern void setup_sparc64_timer(void);
  96
  97 static volatile unsigned long callin_flag = 0;
  98
  99 void smp_callin(void)
 100 {
 101         int cpuid = hard_smp_processor_id();
 102
 103         __local_per_cpu_offset = __per_cpu_offset(cpuid);
 104
 105         if (tlb_type == hypervisor)
 106                 sun4v_ktsb_register();
 107
 108         __flush_tlb_all();
 109
 110         setup_sparc64_timer();
 111
 112         if (cheetah_pcache_forced_on)
 113                 cheetah_enable_pcache();
 114
 115         callin_flag = 1;
 116         __asm__ __volatile__("membar #Sync\n\t"
 117                              "flush  %%g6" : : : "memory");
 118
 119         /* Clear this or we will die instantly when we
 120          * schedule back to this idler...
 121          */
 122         current_thread_info()->new_child = 0;
 123
 124         /* Attach to the address space of init_task. */
 125         atomic_inc(&init_mm.mm_count);
 126         current->active_mm = &init_mm;
 127
 128         /* inform the notifiers about the new cpu */
 129         notify_cpu_starting(cpuid);
 130
 131         while (!cpumask_test_cpu(cpuid, &smp_commenced_mask))
 132                 rmb();
 133
 134         set_cpu_online(cpuid, true);
 135
 136         /* idle thread is expected to have preempt disabled */
 137         preempt_disable();
 138
 139         local_irq_enable();
 140
 141         cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
 142 }
 143
 144 void cpu_panic(void)
 145 {
 146         printk("CPU[%d]: Returns from cpu_idle!\n", smp_processor_id());
 147         panic("SMP bolixed\n");
 148 }
 149
 150 /* This tick register synchronization scheme is taken entirely from
 151  * the ia64 port, see arch/ia64/kernel/smpboot.c for details and credit.
 152  *
 153  * The only change I've made is to rework it so that the master
 154  * initiates the synchonization instead of the slave. -DaveM
 155  */
 156
 157 #define MASTER  0
 158 #define SLAVE   (SMP_CACHE_BYTES/sizeof(unsigned long))
 159
 160 #define NUM_ROUNDS      64      /* magic value */
 161 #define NUM_ITERS       5       /* likewise */
 162
 163 static DEFINE_RAW_SPINLOCK(itc_sync_lock);
 164 static unsigned long go[SLAVE + 1];
 165
 166 #define DEBUG_TICK_SYNC 0
 167
 168 static inline long get_delta (long *rt, long *master)
 169 {
 170         unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
 171         unsigned long tcenter, t0, t1, tm;
 172         unsigned long i;
 173
 174         for (i = 0; i < NUM_ITERS; i++) {
 175                 t0 = tick_ops->get_tick();
 176                 go[MASTER] = 1;
 177                 membar_safe("#StoreLoad");
 178                 while (!(tm = go[SLAVE]))
 179                         rmb();
 180                 go[SLAVE] = 0;
 181                 wmb();
 182                 t1 = tick_ops->get_tick();
 183
 184                 if (t1 - t0 < best_t1 - best_t0)
 185                         best_t0 = t0, best_t1 = t1, best_tm = tm;
 186         }
 187
 188         *rt = best_t1 - best_t0;
 189         *master = best_tm - best_t0;
 190
 191         /* average best_t0 and best_t1 without overflow: */
 192         tcenter = (best_t0/2 + best_t1/2);
 193         if (best_t0 % 2 + best_t1 % 2 == 2)
 194                 tcenter++;
 195         return tcenter - best_tm;
 196 }
 197
 198 void smp_synchronize_tick_client(void)
 199 {
 200         long i, delta, adj, adjust_latency = 0, done = 0;
 201         unsigned long flags, rt, master_time_stamp;
 202 #if DEBUG_TICK_SYNC
 203         struct {
 204                 long rt;        /* roundtrip time */
 205                 long master;    /* master's timestamp */
 206                 long diff;      /* difference between midpoint and master's timestamp */
 207                 long lat;       /* estimate of itc adjustment latency */
 208         } t[NUM_ROUNDS];
 209 #endif
 210
 211         go[MASTER] = 1;
 212
 213         while (go[MASTER])
 214                 rmb();
 215
 216         local_irq_save(flags);
 217         {
 218                 for (i = 0; i < NUM_ROUNDS; i++) {
 219                         delta = get_delta(&rt, &master_time_stamp);
 220                         if (delta == 0)
 221                                 done = 1;       /* let's lock on to this... */
 222
 223                         if (!done) {
 224                                 if (i > 0) {
 225                                         adjust_latency += -delta;
 226                                         adj = -delta + adjust_latency/4;
 227                                 } else
 228                                         adj = -delta;
 229
 230                                 tick_ops->add_tick(adj);
 231                         }
 232 #if DEBUG_TICK_SYNC
 233                         t[i].rt = rt;
 234                         t[i].master = master_time_stamp;
 235                         t[i].diff = delta;
 236                         t[i].lat = adjust_latency/4;
 237 #endif
 238                 }
 239         }
 240         local_irq_restore(flags);
 241
 242 #if DEBUG_TICK_SYNC
 243         for (i = 0; i < NUM_ROUNDS; i++)
 244                 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
 245                        t[i].rt, t[i].master, t[i].diff, t[i].lat);
 246 #endif
 247
 248         printk(KERN_INFO "CPU %d: synchronized TICK with master CPU "
 249                "(last diff %ld cycles, maxerr %lu cycles)\n",
 250                smp_processor_id(), delta, rt);
 251 }
 252
 253 static void smp_start_sync_tick_client(int cpu);
 254
 255 static void smp_synchronize_one_tick(int cpu)
 256 {
 257         unsigned long flags, i;
 258
 259         go[MASTER] = 0;
 260
 261         smp_start_sync_tick_client(cpu);
 262
 263         /* wait for client to be ready */
 264         while (!go[MASTER])
 265                 rmb();
 266
 267         /* now let the client proceed into his loop */
 268         go[MASTER] = 0;
 269         membar_safe("#StoreLoad");
 270
 271         raw_spin_lock_irqsave(&itc_sync_lock, flags);
 272         {
 273                 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; i++) {
 274                         while (!go[MASTER])
 275                                 rmb();
 276                         go[MASTER] = 0;
 277                         wmb();
 278                         go[SLAVE] = tick_ops->get_tick();
 279                         membar_safe("#StoreLoad");
 280                 }
 281         }
 282         raw_spin_unlock_irqrestore(&itc_sync_lock, flags);
 283 }
 284
 285 #if defined(CONFIG_SUN_LDOMS) && defined(CONFIG_HOTPLUG_CPU)
 286 static void ldom_startcpu_cpuid(unsigned int cpu, unsigned long thread_reg,
 287                                 void **descrp)
 288 {
 289         extern unsigned long sparc64_ttable_tl0;
 290         extern unsigned long kern_locked_tte_data;
 291         struct hvtramp_descr *hdesc;
 292         unsigned long trampoline_ra;
 293         struct trap_per_cpu *tb;
 294         u64 tte_vaddr, tte_data;
 295         unsigned long hv_err;
 296         int i;
 297
 298         hdesc = kzalloc(sizeof(*hdesc) +
 299                         (sizeof(struct hvtramp_mapping) *
 300                          num_kernel_image_mappings - 1),
 301                         GFP_KERNEL);
 302         if (!hdesc) {
 303                 printk(KERN_ERR "ldom_startcpu_cpuid: Cannot allocate "
 304                        "hvtramp_descr.\n");
 305                 return;
 306         }
 307         *descrp = hdesc;
 308
 309         hdesc->cpu = cpu;
 310         hdesc->num_mappings = num_kernel_image_mappings;
 311
 312         tb = &trap_block[cpu];
 313
 314         hdesc->fault_info_va = (unsigned long) &tb->fault_info;
 315         hdesc->fault_info_pa = kimage_addr_to_ra(&tb->fault_info);
 316
 317         hdesc->thread_reg = thread_reg;
 318
 319         tte_vaddr = (unsigned long) KERNBASE;
 320         tte_data = kern_locked_tte_data;
 321
 322         for (i = 0; i < hdesc->num_mappings; i++) {
 323                 hdesc->maps[i].vaddr = tte_vaddr;
 324                 hdesc->maps[i].tte   = tte_data;
 325                 tte_vaddr += 0x400000;
 326                 tte_data  += 0x400000;
 327         }
 328
 329         trampoline_ra = kimage_addr_to_ra(hv_cpu_startup);
 330
 331         hv_err = sun4v_cpu_start(cpu, trampoline_ra,
 332                                  kimage_addr_to_ra(&sparc64_ttable_tl0),
 333                                  __pa(hdesc));
 334         if (hv_err)
 335                 printk(KERN_ERR "ldom_startcpu_cpuid: sun4v_cpu_start() "
 336                        "gives error %lu\n", hv_err);
 337 }
 338 #endif
 339
 340 extern unsigned long sparc64_cpu_startup;
 341
 342 /* The OBP cpu startup callback truncates the 3rd arg cookie to
 343  * 32-bits (I think) so to be safe we have it read the pointer
 344  * contained here so we work on >4GB machines. -DaveM
 345  */
 346 static struct thread_info *cpu_new_thread = NULL;
 347
 348 static int smp_boot_one_cpu(unsigned int cpu, struct task_struct *idle)
 349 {
 350         unsigned long entry =
 351                 (unsigned long)(&sparc64_cpu_startup);
 352         unsigned long cookie =
 353                 (unsigned long)(&cpu_new_thread);
 354         void *descr = NULL;
 355         int timeout, ret;
 356
 357         callin_flag = 0;
 358         cpu_new_thread = task_thread_info(idle);
 359
 360         if (tlb_type == hypervisor) {
 361 #if defined(CONFIG_SUN_LDOMS) && defined(CONFIG_HOTPLUG_CPU)
 362                 if (ldom_domaining_enabled)
 363                         ldom_startcpu_cpuid(cpu,
 364                                             (unsigned long) cpu_new_thread,
 365                                             &descr);
 366                 else
 367 #endif
 368                         prom_startcpu_cpuid(cpu, entry, cookie);
 369         } else {
 370                 struct device_node *dp = of_find_node_by_cpuid(cpu);
 371
 372                 prom_startcpu(dp->phandle, entry, cookie);
 373         }
 374
 375         for (timeout = 0; timeout < 50000; timeout++) {
 376                 if (callin_flag)
 377                         break;
 378                 udelay(100);
 379         }
 380
 381         if (callin_flag) {
 382                 ret = 0;
 383         } else {
 384                 printk("Processor %d is stuck.\n", cpu);
 385                 ret = -ENODEV;
 386         }
 387         cpu_new_thread = NULL;
 388
 389         kfree(descr);
 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_mask" and "mm_cpumask(mm)".
 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_mask);
 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));
 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));
 822 }
 823
 824 void __irq_entry smp_call_function_client(int irq, struct pt_regs *regs)
 825 {
 826         clear_softint(1 << irq);
 827         irq_enter();
 828         generic_smp_call_function_interrupt();
 829         irq_exit();
 830 }
 831
 832 void __irq_entry smp_call_function_single_client(int irq, struct pt_regs *regs)
 833 {
 834         clear_softint(1 << irq);
 835         irq_enter();
 836         generic_smp_call_function_single_interrupt();
 837         irq_exit();
 838 }
 839
 840 static void tsb_sync(void *info)
 841 {
 842         struct trap_per_cpu *tp = &trap_block[raw_smp_processor_id()];
 843         struct mm_struct *mm = info;
 844
 845         /* It is not valid to test "current->active_mm == mm" here.
 846          *
 847          * The value of "current" is not changed atomically with
 848          * switch_mm().  But that's OK, we just need to check the
 849          * current cpu's trap block PGD physical address.
 850          */
 851         if (tp->pgd_paddr == __pa(mm->pgd))
 852                 tsb_context_switch(mm);
 853 }
 854
 855 void smp_tsb_sync(struct mm_struct *mm)
 856 {
 857         smp_call_function_many(mm_cpumask(mm), tsb_sync, mm, 1);
 858 }
 859
 860 extern unsigned long xcall_flush_tlb_mm;
 861 extern unsigned long xcall_flush_tlb_page;
 862 extern unsigned long xcall_flush_tlb_kernel_range;
 863 extern unsigned long xcall_fetch_glob_regs;
 864 extern unsigned long xcall_fetch_glob_pmu;
 865 extern unsigned long xcall_fetch_glob_pmu_n4;
 866 extern unsigned long xcall_receive_signal;
 867 extern unsigned long xcall_new_mmu_context_version;
 868 #ifdef CONFIG_KGDB
 869 extern unsigned long xcall_kgdb_capture;
 870 #endif
 871
 872 #ifdef DCACHE_ALIASING_POSSIBLE
 873 extern unsigned long xcall_flush_dcache_page_cheetah;
 874 #endif
 875 extern unsigned long xcall_flush_dcache_page_spitfire;
 876
 877 static inline void __local_flush_dcache_page(struct page *page)
 878 {
 879 #ifdef DCACHE_ALIASING_POSSIBLE
 880         __flush_dcache_page(page_address(page),
 881                             ((tlb_type == spitfire) &&
 882                              page_mapping(page) != NULL));
 883 #else
 884         if (page_mapping(page) != NULL &&
 885             tlb_type == spitfire)
 886                 __flush_icache_page(__pa(page_address(page)));
 887 #endif
 888 }
 889
 890 void smp_flush_dcache_page_impl(struct page *page, int cpu)
 891 {
 892         int this_cpu;
 893
 894         if (tlb_type == hypervisor)
 895                 return;
 896
 897 #ifdef CONFIG_DEBUG_DCFLUSH
 898         atomic_inc(&dcpage_flushes);
 899 #endif
 900
 901         this_cpu = get_cpu();
 902
 903         if (cpu == this_cpu) {
 904                 __local_flush_dcache_page(page);
 905         } else if (cpu_online(cpu)) {
 906                 void *pg_addr = page_address(page);
 907                 u64 data0 = 0;
 908
 909                 if (tlb_type == spitfire) {
 910                         data0 = ((u64)&xcall_flush_dcache_page_spitfire);
 911                         if (page_mapping(page) != NULL)
 912                                 data0 |= ((u64)1 << 32);
 913                 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 914 #ifdef DCACHE_ALIASING_POSSIBLE
 915                         data0 = ((u64)&xcall_flush_dcache_page_cheetah);
 916 #endif
 917                 }
 918                 if (data0) {
 919                         xcall_deliver(data0, __pa(pg_addr),
 920                                       (u64) pg_addr, cpumask_of(cpu));
 921 #ifdef CONFIG_DEBUG_DCFLUSH
 922                         atomic_inc(&dcpage_flushes_xcall);
 923 #endif
 924                 }
 925         }
 926
 927         put_cpu();
 928 }
 929
 930 void flush_dcache_page_all(struct mm_struct *mm, struct page *page)
 931 {
 932         void *pg_addr;
 933         u64 data0;
 934
 935         if (tlb_type == hypervisor)
 936                 return;
 937
 938         preempt_disable();
 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_mask);
 957 #ifdef CONFIG_DEBUG_DCFLUSH
 958                 atomic_inc(&dcpage_flushes_xcall);
 959 #endif
 960         }
 961         __local_flush_dcache_page(page);
 962
 963         preempt_enable();
 964 }
 965
 966 void __irq_entry 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 void smp_fetch_global_pmu(void)
1010 {
1011         if (tlb_type == hypervisor &&
1012             sun4v_chip_type >= SUN4V_CHIP_NIAGARA4)
1013                 smp_cross_call(&xcall_fetch_glob_pmu_n4, 0, 0, 0);
1014         else
1015                 smp_cross_call(&xcall_fetch_glob_pmu, 0, 0, 0);
1016 }
1017
1018 /* We know that the window frames of the user have been flushed
1019  * to the stack before we get here because all callers of us
1020  * are flush_tlb_*() routines, and these run after flush_cache_*()
1021  * which performs the flushw.
1022  *
1023  * The SMP TLB coherency scheme we use works as follows:
1024  *
1025  * 1) mm->cpu_vm_mask is a bit mask of which cpus an address
1026  *    space has (potentially) executed on, this is the heuristic
1027  *    we use to avoid doing cross calls.
1028  *
1029  *    Also, for flushing from kswapd and also for clones, we
1030  *    use cpu_vm_mask as the list of cpus to make run the TLB.
1031  *
1032  * 2) TLB context numbers are shared globally across all processors
1033  *    in the system, this allows us to play several games to avoid
1034  *    cross calls.
1035  *
1036  *    One invariant is that when a cpu switches to a process, and
1037  *    that processes tsk->active_mm->cpu_vm_mask does not have the
1038  *    current cpu's bit set, that tlb context is flushed locally.
1039  *
1040  *    If the address space is non-shared (ie. mm->count == 1) we avoid
1041  *    cross calls when we want to flush the currently running process's
1042  *    tlb state.  This is done by clearing all cpu bits except the current
1043  *    processor's in current->mm->cpu_vm_mask and performing the
1044  *    flush locally only.  This will force any subsequent cpus which run
1045  *    this task to flush the context from the local tlb if the process
1046  *    migrates to another cpu (again).
1047  *
1048  * 3) For shared address spaces (threads) and swapping we bite the
1049  *    bullet for most cases and perform the cross call (but only to
1050  *    the cpus listed in cpu_vm_mask).
1051  *
1052  *    The performance gain from "optimizing" away the cross call for threads is
1053  *    questionable (in theory the big win for threads is the massive sharing of
1054  *    address space state across processors).
1055  */
1056
1057 /* This currently is only used by the hugetlb arch pre-fault
1058  * hook on UltraSPARC-III+ and later when changing the pagesize
1059  * bits of the context register for an address space.
1060  */
1061 void smp_flush_tlb_mm(struct mm_struct *mm)
1062 {
1063         u32 ctx = CTX_HWBITS(mm->context);
1064         int cpu = get_cpu();
1065
1066         if (atomic_read(&mm->mm_users) == 1) {
1067                 cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
1068                 goto local_flush_and_out;
1069         }
1070
1071         smp_cross_call_masked(&xcall_flush_tlb_mm,
1072                               ctx, 0, 0,
1073                               mm_cpumask(mm));
1074
1075 local_flush_and_out:
1076         __flush_tlb_mm(ctx, SECONDARY_CONTEXT);
1077
1078         put_cpu();
1079 }
1080
1081 struct tlb_pending_info {
1082         unsigned long ctx;
1083         unsigned long nr;
1084         unsigned long *vaddrs;
1085 };
1086
1087 static void tlb_pending_func(void *info)
1088 {
1089         struct tlb_pending_info *t = info;
1090
1091         __flush_tlb_pending(t->ctx, t->nr, t->vaddrs);
1092 }
1093
1094 void smp_flush_tlb_pending(struct mm_struct *mm, unsigned long nr, unsigned long *vaddrs)
1095 {
1096         u32 ctx = CTX_HWBITS(mm->context);
1097         struct tlb_pending_info info;
1098         int cpu = get_cpu();
1099
1100         info.ctx = ctx;
1101         info.nr = nr;
1102         info.vaddrs = vaddrs;
1103
1104         if (mm == current->mm && atomic_read(&mm->mm_users) == 1)
1105                 cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
1106         else
1107                 smp_call_function_many(mm_cpumask(mm), tlb_pending_func,
1108                                        &info, 1);
1109
1110         __flush_tlb_pending(ctx, nr, vaddrs);
1111
1112         put_cpu();
1113 }
1114
1115 void smp_flush_tlb_page(struct mm_struct *mm, unsigned long vaddr)
1116 {
1117         unsigned long context = CTX_HWBITS(mm->context);
1118         int cpu = get_cpu();
1119
1120         if (mm == current->mm && atomic_read(&mm->mm_users) == 1)
1121                 cpumask_copy(mm_cpumask(mm), cpumask_of(cpu));
1122         else
1123                 smp_cross_call_masked(&xcall_flush_tlb_page,
1124                                       context, vaddr, 0,
1125                                       mm_cpumask(mm));
1126         __flush_tlb_page(context, vaddr);
1127
1128         put_cpu();
1129 }
1130
1131 void smp_flush_tlb_kernel_range(unsigned long start, unsigned long end)
1132 {
1133         start &= PAGE_MASK;
1134         end    = PAGE_ALIGN(end);
1135         if (start != end) {
1136                 smp_cross_call(&xcall_flush_tlb_kernel_range,
1137                                0, start, end);
1138
1139                 __flush_tlb_kernel_range(start, end);
1140         }
1141 }
1142
1143 /* CPU capture. */
1144 /* #define CAPTURE_DEBUG */
1145 extern unsigned long xcall_capture;
1146
1147 static atomic_t smp_capture_depth = ATOMIC_INIT(0);
1148 static atomic_t smp_capture_registry = ATOMIC_INIT(0);
1149 static unsigned long penguins_are_doing_time;
1150
1151 void smp_capture(void)
1152 {
1153         int result = atomic_add_return(1, &smp_capture_depth);
1154
1155         if (result == 1) {
1156                 int ncpus = num_online_cpus();
1157
1158 #ifdef CAPTURE_DEBUG
1159                 printk("CPU[%d]: Sending penguins to jail...",
1160                        smp_processor_id());
1161 #endif
1162                 penguins_are_doing_time = 1;
1163                 atomic_inc(&smp_capture_registry);
1164                 smp_cross_call(&xcall_capture, 0, 0, 0);
1165                 while (atomic_read(&smp_capture_registry) != ncpus)
1166                         rmb();
1167 #ifdef CAPTURE_DEBUG
1168                 printk("done\n");
1169 #endif
1170         }
1171 }
1172
1173 void smp_release(void)
1174 {
1175         if (atomic_dec_and_test(&smp_capture_depth)) {
1176 #ifdef CAPTURE_DEBUG
1177                 printk("CPU[%d]: Giving pardon to "
1178                        "imprisoned penguins\n",
1179                        smp_processor_id());
1180 #endif
1181                 penguins_are_doing_time = 0;
1182                 membar_safe("#StoreLoad");
1183                 atomic_dec(&smp_capture_registry);
1184         }
1185 }
1186
1187 /* Imprisoned penguins run with %pil == PIL_NORMAL_MAX, but PSTATE_IE
1188  * set, so they can service tlb flush xcalls...
1189  */
1190 extern void prom_world(int);
1191
1192 void __irq_entry smp_penguin_jailcell(int irq, struct pt_regs *regs)
1193 {
1194         clear_softint(1 << irq);
1195
1196         preempt_disable();
1197
1198         __asm__ __volatile__("flushw");
1199         prom_world(1);
1200         atomic_inc(&smp_capture_registry);
1201         membar_safe("#StoreLoad");
1202         while (penguins_are_doing_time)
1203                 rmb();
1204         atomic_dec(&smp_capture_registry);
1205         prom_world(0);
1206
1207         preempt_enable();
1208 }
1209
1210 /* /proc/profile writes can call this, don't __init it please. */
1211 int setup_profiling_timer(unsigned int multiplier)
1212 {
1213         return -EINVAL;
1214 }
1215
1216 void __init smp_prepare_cpus(unsigned int max_cpus)
1217 {
1218 }
1219
1220 void smp_prepare_boot_cpu(void)
1221 {
1222 }
1223
1224 void __init smp_setup_processor_id(void)
1225 {
1226         if (tlb_type == spitfire)
1227                 xcall_deliver_impl = spitfire_xcall_deliver;
1228         else if (tlb_type == cheetah || tlb_type == cheetah_plus)
1229                 xcall_deliver_impl = cheetah_xcall_deliver;
1230         else
1231                 xcall_deliver_impl = hypervisor_xcall_deliver;
1232 }
1233
1234 void __init smp_fill_in_cpu_possible_map(void)
1235 {
1236         int possible_cpus = num_possible_cpus();
1237         int i;
1238
1239         if (possible_cpus > nr_cpu_ids)
1240                 possible_cpus = nr_cpu_ids;
1241
1242         for (i = 0; i < possible_cpus; i++)
1243                 set_cpu_possible(i, true);
1244         for (; i < NR_CPUS; i++)
1245                 set_cpu_possible(i, false);
1246 }
1247
1248 void smp_fill_in_sib_core_maps(void)
1249 {
1250         unsigned int i;
1251
1252         for_each_present_cpu(i) {
1253                 unsigned int j;
1254
1255                 cpumask_clear(&cpu_core_map[i]);
1256                 if (cpu_data(i).core_id == 0) {
1257                         cpumask_set_cpu(i, &cpu_core_map[i]);
1258                         continue;
1259                 }
1260
1261                 for_each_present_cpu(j) {
1262                         if (cpu_data(i).core_id ==
1263                             cpu_data(j).core_id)
1264                                 cpumask_set_cpu(j, &cpu_core_map[i]);
1265                 }
1266         }
1267
1268         for_each_present_cpu(i)  {
1269                 unsigned int j;
1270
1271                 for_each_present_cpu(j)  {
1272                         if (cpu_data(i).max_cache_id ==
1273                             cpu_data(j).max_cache_id)
1274                                 cpumask_set_cpu(j, &cpu_core_sib_cache_map[i]);
1275
1276                         if (cpu_data(i).sock_id == cpu_data(j).sock_id)
1277                                 cpumask_set_cpu(j, &cpu_core_sib_map[i]);
1278                 }
1279         }
1280
1281         for_each_present_cpu(i) {
1282                 unsigned int j;
1283
1284                 cpumask_clear(&per_cpu(cpu_sibling_map, i));
1285                 if (cpu_data(i).proc_id == -1) {
1286                         cpumask_set_cpu(i, &per_cpu(cpu_sibling_map, i));
1287                         continue;
1288                 }
1289
1290                 for_each_present_cpu(j) {
1291                         if (cpu_data(i).proc_id ==
1292                             cpu_data(j).proc_id)
1293                                 cpumask_set_cpu(j, &per_cpu(cpu_sibling_map, i));
1294                 }
1295         }
1296 }
1297
1298 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
1299 {
1300         int ret = smp_boot_one_cpu(cpu, tidle);
1301
1302         if (!ret) {
1303                 cpumask_set_cpu(cpu, &smp_commenced_mask);
1304                 while (!cpu_online(cpu))
1305                         mb();
1306                 if (!cpu_online(cpu)) {
1307                         ret = -ENODEV;
1308                 } else {
1309                         /* On SUN4V, writes to %tick and %stick are
1310                          * not allowed.
1311                          */
1312                         if (tlb_type != hypervisor)
1313                                 smp_synchronize_one_tick(cpu);
1314                 }
1315         }
1316         return ret;
1317 }
1318
1319 #ifdef CONFIG_HOTPLUG_CPU
1320 void cpu_play_dead(void)
1321 {
1322         int cpu = smp_processor_id();
1323         unsigned long pstate;
1324
1325         idle_task_exit();
1326
1327         if (tlb_type == hypervisor) {
1328                 struct trap_per_cpu *tb = &trap_block[cpu];
1329
1330                 sun4v_cpu_qconf(HV_CPU_QUEUE_CPU_MONDO,
1331                                 tb->cpu_mondo_pa, 0);
1332                 sun4v_cpu_qconf(HV_CPU_QUEUE_DEVICE_MONDO,
1333                                 tb->dev_mondo_pa, 0);
1334                 sun4v_cpu_qconf(HV_CPU_QUEUE_RES_ERROR,
1335                                 tb->resum_mondo_pa, 0);
1336                 sun4v_cpu_qconf(HV_CPU_QUEUE_NONRES_ERROR,
1337                                 tb->nonresum_mondo_pa, 0);
1338         }
1339
1340         cpumask_clear_cpu(cpu, &smp_commenced_mask);
1341         membar_safe("#Sync");
1342
1343         local_irq_disable();
1344
1345         __asm__ __volatile__(
1346                 "rdpr   %%pstate, %0\n\t"
1347                 "wrpr   %0, %1, %%pstate"
1348                 : "=r" (pstate)
1349                 : "i" (PSTATE_IE));
1350
1351         while (1)
1352                 barrier();
1353 }
1354
1355 int __cpu_disable(void)
1356 {
1357         int cpu = smp_processor_id();
1358         cpuinfo_sparc *c;
1359         int i;
1360
1361         for_each_cpu(i, &cpu_core_map[cpu])
1362                 cpumask_clear_cpu(cpu, &cpu_core_map[i]);
1363         cpumask_clear(&cpu_core_map[cpu]);
1364
1365         for_each_cpu(i, &per_cpu(cpu_sibling_map, cpu))
1366                 cpumask_clear_cpu(cpu, &per_cpu(cpu_sibling_map, i));
1367         cpumask_clear(&per_cpu(cpu_sibling_map, cpu));
1368
1369         c = &cpu_data(cpu);
1370
1371         c->core_id = 0;
1372         c->proc_id = -1;
1373
1374         smp_wmb();
1375
1376         /* Make sure no interrupts point to this cpu.  */
1377         fixup_irqs();
1378
1379         local_irq_enable();
1380         mdelay(1);
1381         local_irq_disable();
1382
1383         set_cpu_online(cpu, false);
1384
1385         cpu_map_rebuild();
1386
1387         return 0;
1388 }
1389
1390 void __cpu_die(unsigned int cpu)
1391 {
1392         int i;
1393
1394         for (i = 0; i < 100; i++) {
1395                 smp_rmb();
1396                 if (!cpumask_test_cpu(cpu, &smp_commenced_mask))
1397                         break;
1398                 msleep(100);
1399         }
1400         if (cpumask_test_cpu(cpu, &smp_commenced_mask)) {
1401                 printk(KERN_ERR "CPU %u didn't die...\n", cpu);
1402         } else {
1403 #if defined(CONFIG_SUN_LDOMS)
1404                 unsigned long hv_err;
1405                 int limit = 100;
1406
1407                 do {
1408                         hv_err = sun4v_cpu_stop(cpu);
1409                         if (hv_err == HV_EOK) {
1410                                 set_cpu_present(cpu, false);
1411                                 break;
1412                         }
1413                 } while (--limit > 0);
1414                 if (limit <= 0) {
1415                         printk(KERN_ERR "sun4v_cpu_stop() fails err=%lu\n",
1416                                hv_err);
1417                 }
1418 #endif
1419         }
1420 }
1421 #endif
1422
1423 void __init smp_cpus_done(unsigned int max_cpus)
1424 {
1425 }
1426
1427 void smp_send_reschedule(int cpu)
1428 {
1429         if (cpu == smp_processor_id()) {
1430                 WARN_ON_ONCE(preemptible());
1431                 set_softint(1 << PIL_SMP_RECEIVE_SIGNAL);
1432         } else {
1433                 xcall_deliver((u64) &xcall_receive_signal,
1434                               0, 0, cpumask_of(cpu));
1435         }
1436 }
1437
1438 void __irq_entry smp_receive_signal_client(int irq, struct pt_regs *regs)
1439 {
1440         clear_softint(1 << irq);
1441         scheduler_ipi();
1442 }
1443
1444 static void stop_this_cpu(void *dummy)
1445 {
1446         prom_stopself();
1447 }
1448
1449 void smp_send_stop(void)
1450 {
1451         int cpu;
1452
1453         if (tlb_type == hypervisor) {
1454                 for_each_online_cpu(cpu) {
1455                         if (cpu == smp_processor_id())
1456                                 continue;
1457 #ifdef CONFIG_SUN_LDOMS
1458                         if (ldom_domaining_enabled) {
1459                                 unsigned long hv_err;
1460                                 hv_err = sun4v_cpu_stop(cpu);
1461                                 if (hv_err)
1462                                         printk(KERN_ERR "sun4v_cpu_stop() "
1463                                                "failed err=%lu\n", hv_err);
1464                         } else
1465 #endif
1466                                 prom_stopcpu_cpuid(cpu);
1467                 }
1468         } else
1469                 smp_call_function(stop_this_cpu, NULL, 0);
1470 }
1471
1472 /**
1473  * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu
1474  * @cpu: cpu to allocate for
1475  * @size: size allocation in bytes
1476  * @align: alignment
1477  *
1478  * Allocate @size bytes aligned at @align for cpu @cpu.  This wrapper
1479  * does the right thing for NUMA regardless of the current
1480  * configuration.
1481  *
1482  * RETURNS:
1483  * Pointer to the allocated area on success, NULL on failure.
1484  */
1485 static void * __init pcpu_alloc_bootmem(unsigned int cpu, size_t size,
1486                                         size_t align)
1487 {
1488         const unsigned long goal = __pa(MAX_DMA_ADDRESS);
1489 #ifdef CONFIG_NEED_MULTIPLE_NODES
1490         int node = cpu_to_node(cpu);
1491         void *ptr;
1492
1493         if (!node_online(node) || !NODE_DATA(node)) {
1494                 ptr = __alloc_bootmem(size, align, goal);
1495                 pr_info("cpu %d has no node %d or node-local memory\n",
1496                         cpu, node);
1497                 pr_debug("per cpu data for cpu%d %lu bytes at %016lx\n",
1498                          cpu, size, __pa(ptr));
1499         } else {
1500                 ptr = __alloc_bootmem_node(NODE_DATA(node),
1501                                            size, align, goal);
1502                 pr_debug("per cpu data for cpu%d %lu bytes on node%d at "
1503                          "%016lx\n", cpu, size, node, __pa(ptr));
1504         }
1505         return ptr;
1506 #else
1507         return __alloc_bootmem(size, align, goal);
1508 #endif
1509 }
1510
1511 static void __init pcpu_free_bootmem(void *ptr, size_t size)
1512 {
1513         free_bootmem(__pa(ptr), size);
1514 }
1515
1516 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
1517 {
1518         if (cpu_to_node(from) == cpu_to_node(to))
1519                 return LOCAL_DISTANCE;
1520         else
1521                 return REMOTE_DISTANCE;
1522 }
1523
1524 static void __init pcpu_populate_pte(unsigned long addr)
1525 {
1526         pgd_t *pgd = pgd_offset_k(addr);
1527         pud_t *pud;
1528         pmd_t *pmd;
1529
1530         if (pgd_none(*pgd)) {
1531                 pud_t *new;
1532
1533                 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1534                 pgd_populate(&init_mm, pgd, new);
1535         }
1536
1537         pud = pud_offset(pgd, addr);
1538         if (pud_none(*pud)) {
1539                 pmd_t *new;
1540
1541                 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1542                 pud_populate(&init_mm, pud, new);
1543         }
1544
1545         pmd = pmd_offset(pud, addr);
1546         if (!pmd_present(*pmd)) {
1547                 pte_t *new;
1548
1549                 new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1550                 pmd_populate_kernel(&init_mm, pmd, new);
1551         }
1552 }
1553
1554 void __init setup_per_cpu_areas(void)
1555 {
1556         unsigned long delta;
1557         unsigned int cpu;
1558         int rc = -EINVAL;
1559
1560         if (pcpu_chosen_fc != PCPU_FC_PAGE) {
1561                 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
1562                                             PERCPU_DYNAMIC_RESERVE, 4 << 20,
1563                                             pcpu_cpu_distance,
1564                                             pcpu_alloc_bootmem,
1565                                             pcpu_free_bootmem);
1566                 if (rc)
1567                         pr_warning("PERCPU: %s allocator failed (%d), "
1568                                    "falling back to page size\n",
1569                                    pcpu_fc_names[pcpu_chosen_fc], rc);
1570         }
1571         if (rc < 0)
1572                 rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE,
1573                                            pcpu_alloc_bootmem,
1574                                            pcpu_free_bootmem,
1575                                            pcpu_populate_pte);
1576         if (rc < 0)
1577                 panic("cannot initialize percpu area (err=%d)", rc);
1578
1579         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1580         for_each_possible_cpu(cpu)
1581                 __per_cpu_offset(cpu) = delta + pcpu_unit_offsets[cpu];
1582
1583         /* Setup %g5 for the boot cpu.  */
1584         __local_per_cpu_offset = __per_cpu_offset(smp_processor_id());
1585
1586         of_fill_in_cpu_data();
1587         if (tlb_type == hypervisor)
1588                 mdesc_fill_in_cpu_data(cpu_all_mask);
1589 }