arch/sparc/kernel/smp_64.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /* smp.c: Sparc64 SMP support.
   3  *
   4  * Copyright (C) 1997, 2007, 2008 David S. Miller (davem@davemloft.net)
   5  */
   6
   7 #include <linux/export.h>
   8 #include <linux/kernel.h>
   9 #include <linux/sched/mm.h>
  10 #include <linux/sched/hotplug.h>
  11 #include <linux/mm.h>
  12 #include <linux/pagemap.h>
  13 #include <linux/threads.h>
  14 #include <linux/smp.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/kernel_stat.h>
  17 #include <linux/delay.h>
  18 #include <linux/init.h>
  19 #include <linux/spinlock.h>
  20 #include <linux/fs.h>
  21 #include <linux/seq_file.h>
  22 #include <linux/cache.h>
  23 #include <linux/jiffies.h>
  24 #include <linux/profile.h>
  25 #include <linux/memblock.h>
  26 #include <linux/vmalloc.h>
  27 #include <linux/ftrace.h>
  28 #include <linux/cpu.h>
  29 #include <linux/slab.h>
  30 #include <linux/kgdb.h>
  31
  32 #include <asm/head.h>
  33 #include <asm/ptrace.h>
  34 #include <linux/atomic.h>
  35 #include <asm/tlbflush.h>
  36 #include <asm/mmu_context.h>
  37 #include <asm/cpudata.h>
  38 #include <asm/hvtramp.h>
  39 #include <asm/io.h>
  40 #include <asm/timer.h>
  41 #include <asm/setup.h>
  42
  43 #include <asm/irq.h>
  44 #include <asm/irq_regs.h>
  45 #include <asm/page.h>
  46 #include <asm/oplib.h>
  47 #include <linux/uaccess.h>
  48 #include <asm/starfire.h>
  49 #include <asm/tlb.h>
  50 #include <asm/pgalloc.h>
  51 #include <asm/sections.h>
  52 #include <asm/prom.h>
  53 #include <asm/mdesc.h>
  54 #include <asm/ldc.h>
  55 #include <asm/hypervisor.h>
  56 #include <asm/pcr.h>
  57
  58 #include "cpumap.h"
  59 #include "kernel.h"
  60
  61 DEFINE_PER_CPU(cpumask_t, cpu_sibling_map) = CPU_MASK_NONE;
  62 cpumask_t cpu_core_map[NR_CPUS] __read_mostly =
  63         { [0 ... NR_CPUS-1] = CPU_MASK_NONE };
  64
  65 cpumask_t cpu_core_sib_map[NR_CPUS] __read_mostly = {
  66         [0 ... NR_CPUS-1] = CPU_MASK_NONE };
  67
  68 cpumask_t cpu_core_sib_cache_map[NR_CPUS] __read_mostly = {
  69         [0 ... NR_CPUS - 1] = CPU_MASK_NONE };
  70
  71 EXPORT_PER_CPU_SYMBOL(cpu_sibling_map);
  72 EXPORT_SYMBOL(cpu_core_map);
  73 EXPORT_SYMBOL(cpu_core_sib_map);
  74 EXPORT_SYMBOL(cpu_core_sib_cache_map);
  75
  76 static cpumask_t smp_commenced_mask;
  77
  78 static DEFINE_PER_CPU(bool, poke);
  79 static bool cpu_poke;
  80
  81 void smp_info(struct seq_file *m)
  82 {
  83         int i;
  84
  85         seq_printf(m, "State:\n");
  86         for_each_online_cpu(i)
  87                 seq_printf(m, "CPU%d:\t\tonline\n", i);
  88 }
  89
  90 void smp_bogo(struct seq_file *m)
  91 {
  92         int i;
  93
  94         for_each_online_cpu(i)
  95                 seq_printf(m,
  96                            "Cpu%dClkTck\t: %016lx\n",
  97                            i, cpu_data(i).clock_tick);
  98 }
  99
 100 extern void setup_sparc64_timer(void);
 101
 102 static volatile unsigned long callin_flag = 0;
 103
 104 void smp_callin(void)
 105 {
 106         int cpuid = hard_smp_processor_id();
 107
 108         __local_per_cpu_offset = __per_cpu_offset(cpuid);
 109
 110         if (tlb_type == hypervisor)
 111                 sun4v_ktsb_register();
 112
 113         __flush_tlb_all();
 114
 115         setup_sparc64_timer();
 116
 117         if (cheetah_pcache_forced_on)
 118                 cheetah_enable_pcache();
 119
 120         callin_flag = 1;
 121         __asm__ __volatile__("membar #Sync\n\t"
 122                              "flush  %%g6" : : : "memory");
 123
 124         /* Clear this or we will die instantly when we
 125          * schedule back to this idler...
 126          */
 127         current_thread_info()->new_child = 0;
 128
 129         /* Attach to the address space of init_task. */
 130         mmgrab(&init_mm);
 131         current->active_mm = &init_mm;
 132
 133         /* inform the notifiers about the new cpu */
 134         notify_cpu_starting(cpuid);
 135
 136         while (!cpumask_test_cpu(cpuid, &smp_commenced_mask))
 137                 rmb();
 138
 139         set_cpu_online(cpuid, true);
 140
 141         local_irq_enable();
 142
 143         cpu_startup_entry(CPUHP_AP_ONLINE_IDLE);
 144 }
 145
 146 void cpu_panic(void)
 147 {
 148         printk("CPU[%d]: Returns from cpu_idle!\n", smp_processor_id());
 149         panic("SMP bolixed\n");
 150 }
 151
 152 /* This tick register synchronization scheme is taken entirely from
 153  * the ia64 port, see arch/ia64/kernel/smpboot.c for details and credit.
 154  *
 155  * The only change I've made is to rework it so that the master
 156  * initiates the synchonization instead of the slave. -DaveM
 157  */
 158
 159 #define MASTER  0
 160 #define SLAVE   (SMP_CACHE_BYTES/sizeof(unsigned long))
 161
 162 #define NUM_ROUNDS      64      /* magic value */
 163 #define NUM_ITERS       5       /* likewise */
 164
 165 static DEFINE_RAW_SPINLOCK(itc_sync_lock);
 166 static unsigned long go[SLAVE + 1];
 167
 168 #define DEBUG_TICK_SYNC 0
 169
 170 static inline long get_delta (long *rt, long *master)
 171 {
 172         unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
 173         unsigned long tcenter, t0, t1, tm;
 174         unsigned long i;
 175
 176         for (i = 0; i < NUM_ITERS; i++) {
 177                 t0 = tick_ops->get_tick();
 178                 go[MASTER] = 1;
 179                 membar_safe("#StoreLoad");
 180                 while (!(tm = go[SLAVE]))
 181                         rmb();
 182                 go[SLAVE] = 0;
 183                 wmb();
 184                 t1 = tick_ops->get_tick();
 185
 186                 if (t1 - t0 < best_t1 - best_t0)
 187                         best_t0 = t0, best_t1 = t1, best_tm = tm;
 188         }
 189
 190         *rt = best_t1 - best_t0;
 191         *master = best_tm - best_t0;
 192
 193         /* average best_t0 and best_t1 without overflow: */
 194         tcenter = (best_t0/2 + best_t1/2);
 195         if (best_t0 % 2 + best_t1 % 2 == 2)
 196                 tcenter++;
 197         return tcenter - best_tm;
 198 }
 199
 200 void smp_synchronize_tick_client(void)
 201 {
 202         long i, delta, adj, adjust_latency = 0, done = 0;
 203         unsigned long flags, rt, master_time_stamp;
 204 #if DEBUG_TICK_SYNC
 205         struct {
 206                 long rt;        /* roundtrip time */
 207                 long master;    /* master's timestamp */
 208                 long diff;      /* difference between midpoint and master's timestamp */
 209                 long lat;       /* estimate of itc adjustment latency */
 210         } t[NUM_ROUNDS];
 211 #endif
 212
 213         go[MASTER] = 1;
 214
 215         while (go[MASTER])
 216                 rmb();
 217
 218         local_irq_save(flags);
 219         {
 220                 for (i = 0; i < NUM_ROUNDS; i++) {
 221                         delta = get_delta(&rt, &master_time_stamp);
 222                         if (delta == 0)
 223                                 done = 1;       /* let's lock on to this... */
 224
 225                         if (!done) {
 226                                 if (i > 0) {
 227                                         adjust_latency += -delta;
 228                                         adj = -delta + adjust_latency/4;
 229                                 } else
 230                                         adj = -delta;
 231
 232                                 tick_ops->add_tick(adj);
 233                         }
 234 #if DEBUG_TICK_SYNC
 235                         t[i].rt = rt;
 236                         t[i].master = master_time_stamp;
 237                         t[i].diff = delta;
 238                         t[i].lat = adjust_latency/4;
 239 #endif
 240                 }
 241         }
 242         local_irq_restore(flags);
 243
 244 #if DEBUG_TICK_SYNC
 245         for (i = 0; i < NUM_ROUNDS; i++)
 246                 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
 247                        t[i].rt, t[i].master, t[i].diff, t[i].lat);
 248 #endif
 249
 250         printk(KERN_INFO "CPU %d: synchronized TICK with master CPU "
 251                "(last diff %ld cycles, maxerr %lu cycles)\n",
 252                smp_processor_id(), delta, rt);
 253 }
 254
 255 static void smp_start_sync_tick_client(int cpu);
 256
 257 static void smp_synchronize_one_tick(int cpu)
 258 {
 259         unsigned long flags, i;
 260
 261         go[MASTER] = 0;
 262
 263         smp_start_sync_tick_client(cpu);
 264
 265         /* wait for client to be ready */
 266         while (!go[MASTER])
 267                 rmb();
 268
 269         /* now let the client proceed into his loop */
 270         go[MASTER] = 0;
 271         membar_safe("#StoreLoad");
 272
 273         raw_spin_lock_irqsave(&itc_sync_lock, flags);
 274         {
 275                 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; i++) {
 276                         while (!go[MASTER])
 277                                 rmb();
 278                         go[MASTER] = 0;
 279                         wmb();
 280                         go[SLAVE] = tick_ops->get_tick();
 281                         membar_safe("#StoreLoad");
 282                 }
 283         }
 284         raw_spin_unlock_irqrestore(&itc_sync_lock, flags);
 285 }
 286
 287 #if defined(CONFIG_SUN_LDOMS) && defined(CONFIG_HOTPLUG_CPU)
 288 static void ldom_startcpu_cpuid(unsigned int cpu, unsigned long thread_reg,
 289                                 void **descrp)
 290 {
 291         extern unsigned long sparc64_ttable_tl0;
 292         extern unsigned long kern_locked_tte_data;
 293         struct hvtramp_descr *hdesc;
 294         unsigned long trampoline_ra;
 295         struct trap_per_cpu *tb;
 296         u64 tte_vaddr, tte_data;
 297         unsigned long hv_err;
 298         int i;
 299
 300         hdesc = kzalloc(sizeof(*hdesc) +
 301                         (sizeof(struct hvtramp_mapping) *
 302                          num_kernel_image_mappings - 1),
 303                         GFP_KERNEL);
 304         if (!hdesc) {
 305                 printk(KERN_ERR "ldom_startcpu_cpuid: Cannot allocate "
 306                        "hvtramp_descr.\n");
 307                 return;
 308         }
 309         *descrp = hdesc;
 310
 311         hdesc->cpu = cpu;
 312         hdesc->num_mappings = num_kernel_image_mappings;
 313
 314         tb = &trap_block[cpu];
 315
 316         hdesc->fault_info_va = (unsigned long) &tb->fault_info;
 317         hdesc->fault_info_pa = kimage_addr_to_ra(&tb->fault_info);
 318
 319         hdesc->thread_reg = thread_reg;
 320
 321         tte_vaddr = (unsigned long) KERNBASE;
 322         tte_data = kern_locked_tte_data;
 323
 324         for (i = 0; i < hdesc->num_mappings; i++) {
 325                 hdesc->maps[i].vaddr = tte_vaddr;
 326                 hdesc->maps[i].tte   = tte_data;
 327                 tte_vaddr += 0x400000;
 328                 tte_data  += 0x400000;
 329         }
 330
 331         trampoline_ra = kimage_addr_to_ra(hv_cpu_startup);
 332
 333         hv_err = sun4v_cpu_start(cpu, trampoline_ra,
 334                                  kimage_addr_to_ra(&sparc64_ttable_tl0),
 335                                  __pa(hdesc));
 336         if (hv_err)
 337                 printk(KERN_ERR "ldom_startcpu_cpuid: sun4v_cpu_start() "
 338                        "gives error %lu\n", hv_err);
 339 }
 340 #endif
 341
 342 extern unsigned long sparc64_cpu_startup;
 343
 344 /* The OBP cpu startup callback truncates the 3rd arg cookie to
 345  * 32-bits (I think) so to be safe we have it read the pointer
 346  * contained here so we work on >4GB machines. -DaveM
 347  */
 348 static struct thread_info *cpu_new_thread = NULL;
 349
 350 static int smp_boot_one_cpu(unsigned int cpu, struct task_struct *idle)
 351 {
 352         unsigned long entry =
 353                 (unsigned long)(&sparc64_cpu_startup);
 354         unsigned long cookie =
 355                 (unsigned long)(&cpu_new_thread);
 356         void *descr = NULL;
 357         int timeout, ret;
 358
 359         callin_flag = 0;
 360         cpu_new_thread = task_thread_info(idle);
 361
 362         if (tlb_type == hypervisor) {
 363 #if defined(CONFIG_SUN_LDOMS) && defined(CONFIG_HOTPLUG_CPU)
 364                 if (ldom_domaining_enabled)
 365                         ldom_startcpu_cpuid(cpu,
 366                                             (unsigned long) cpu_new_thread,
 367                                             &descr);
 368                 else
 369 #endif
 370                         prom_startcpu_cpuid(cpu, entry, cookie);
 371         } else {
 372                 struct device_node *dp = of_find_node_by_cpuid(cpu);
 373
 374                 prom_startcpu(dp->phandle, entry, cookie);
 375         }
 376
 377         for (timeout = 0; timeout < 50000; timeout++) {
 378                 if (callin_flag)
 379                         break;
 380                 udelay(100);
 381         }
 382
 383         if (callin_flag) {
 384                 ret = 0;
 385         } else {
 386                 printk("Processor %d is stuck.\n", cpu);
 387                 ret = -ENODEV;
 388         }
 389         cpu_new_thread = NULL;
 390
 391         kfree(descr);
 392
 393         return ret;
 394 }
 395
 396 static void spitfire_xcall_helper(u64 data0, u64 data1, u64 data2, u64 pstate, unsigned long cpu)
 397 {
 398         u64 result, target;
 399         int stuck, tmp;
 400
 401         if (this_is_starfire) {
 402                 /* map to real upaid */
 403                 cpu = (((cpu & 0x3c) << 1) |
 404                         ((cpu & 0x40) >> 4) |
 405                         (cpu & 0x3));
 406         }
 407
 408         target = (cpu << 14) | 0x70;
 409 again:
 410         /* Ok, this is the real Spitfire Errata #54.
 411          * One must read back from a UDB internal register
 412          * after writes to the UDB interrupt dispatch, but
 413          * before the membar Sync for that write.
 414          * So we use the high UDB control register (ASI 0x7f,
 415          * ADDR 0x20) for the dummy read. -DaveM
 416          */
 417         tmp = 0x40;
 418         __asm__ __volatile__(
 419         "wrpr   %1, %2, %%pstate\n\t"
 420         "stxa   %4, [%0] %3\n\t"
 421         "stxa   %5, [%0+%8] %3\n\t"
 422         "add    %0, %8, %0\n\t"
 423         "stxa   %6, [%0+%8] %3\n\t"
 424         "membar #Sync\n\t"
 425         "stxa   %%g0, [%7] %3\n\t"
 426         "membar #Sync\n\t"
 427         "mov    0x20, %%g1\n\t"
 428         "ldxa   [%%g1] 0x7f, %%g0\n\t"
 429         "membar #Sync"
 430         : "=r" (tmp)
 431         : "r" (pstate), "i" (PSTATE_IE), "i" (ASI_INTR_W),
 432           "r" (data0), "r" (data1), "r" (data2), "r" (target),
 433           "r" (0x10), "0" (tmp)
 434         : "g1");
 435
 436         /* NOTE: PSTATE_IE is still clear. */
 437         stuck = 100000;
 438         do {
 439                 __asm__ __volatile__("ldxa [%%g0] %1, %0"
 440                         : "=r" (result)
 441                         : "i" (ASI_INTR_DISPATCH_STAT));
 442                 if (result == 0) {
 443                         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 444                                              : : "r" (pstate));
 445                         return;
 446                 }
 447                 stuck -= 1;
 448                 if (stuck == 0)
 449                         break;
 450         } while (result & 0x1);
 451         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 452                              : : "r" (pstate));
 453         if (stuck == 0) {
 454                 printk("CPU[%d]: mondo stuckage result[%016llx]\n",
 455                        smp_processor_id(), result);
 456         } else {
 457                 udelay(2);
 458                 goto again;
 459         }
 460 }
 461
 462 static void spitfire_xcall_deliver(struct trap_per_cpu *tb, int cnt)
 463 {
 464         u64 *mondo, data0, data1, data2;
 465         u16 *cpu_list;
 466         u64 pstate;
 467         int i;
 468
 469         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 470         cpu_list = __va(tb->cpu_list_pa);
 471         mondo = __va(tb->cpu_mondo_block_pa);
 472         data0 = mondo[0];
 473         data1 = mondo[1];
 474         data2 = mondo[2];
 475         for (i = 0; i < cnt; i++)
 476                 spitfire_xcall_helper(data0, data1, data2, pstate, cpu_list[i]);
 477 }
 478
 479 /* Cheetah now allows to send the whole 64-bytes of data in the interrupt
 480  * packet, but we have no use for that.  However we do take advantage of
 481  * the new pipelining feature (ie. dispatch to multiple cpus simultaneously).
 482  */
 483 static void cheetah_xcall_deliver(struct trap_per_cpu *tb, int cnt)
 484 {
 485         int nack_busy_id, is_jbus, need_more;
 486         u64 *mondo, pstate, ver, busy_mask;
 487         u16 *cpu_list;
 488
 489         cpu_list = __va(tb->cpu_list_pa);
 490         mondo = __va(tb->cpu_mondo_block_pa);
 491
 492         /* Unfortunately, someone at Sun had the brilliant idea to make the
 493          * busy/nack fields hard-coded by ITID number for this Ultra-III
 494          * derivative processor.
 495          */
 496         __asm__ ("rdpr %%ver, %0" : "=r" (ver));
 497         is_jbus = ((ver >> 32) == __JALAPENO_ID ||
 498                    (ver >> 32) == __SERRANO_ID);
 499
 500         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 501
 502 retry:
 503         need_more = 0;
 504         __asm__ __volatile__("wrpr %0, %1, %%pstate\n\t"
 505                              : : "r" (pstate), "i" (PSTATE_IE));
 506
 507         /* Setup the dispatch data registers. */
 508         __asm__ __volatile__("stxa      %0, [%3] %6\n\t"
 509                              "stxa      %1, [%4] %6\n\t"
 510                              "stxa      %2, [%5] %6\n\t"
 511                              "membar    #Sync\n\t"
 512                              : /* no outputs */
 513                              : "r" (mondo[0]), "r" (mondo[1]), "r" (mondo[2]),
 514                                "r" (0x40), "r" (0x50), "r" (0x60),
 515                                "i" (ASI_INTR_W));
 516
 517         nack_busy_id = 0;
 518         busy_mask = 0;
 519         {
 520                 int i;
 521
 522                 for (i = 0; i < cnt; i++) {
 523                         u64 target, nr;
 524
 525                         nr = cpu_list[i];
 526                         if (nr == 0xffff)
 527                                 continue;
 528
 529                         target = (nr << 14) | 0x70;
 530                         if (is_jbus) {
 531                                 busy_mask |= (0x1UL << (nr * 2));
 532                         } else {
 533                                 target |= (nack_busy_id << 24);
 534                                 busy_mask |= (0x1UL <<
 535                                               (nack_busy_id * 2));
 536                         }
 537                         __asm__ __volatile__(
 538                                 "stxa   %%g0, [%0] %1\n\t"
 539                                 "membar #Sync\n\t"
 540                                 : /* no outputs */
 541                                 : "r" (target), "i" (ASI_INTR_W));
 542                         nack_busy_id++;
 543                         if (nack_busy_id == 32) {
 544                                 need_more = 1;
 545                                 break;
 546                         }
 547                 }
 548         }
 549
 550         /* Now, poll for completion. */
 551         {
 552                 u64 dispatch_stat, nack_mask;
 553                 long stuck;
 554
 555                 stuck = 100000 * nack_busy_id;
 556                 nack_mask = busy_mask << 1;
 557                 do {
 558                         __asm__ __volatile__("ldxa      [%%g0] %1, %0"
 559                                              : "=r" (dispatch_stat)
 560                                              : "i" (ASI_INTR_DISPATCH_STAT));
 561                         if (!(dispatch_stat & (busy_mask | nack_mask))) {
 562                                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 563                                                      : : "r" (pstate));
 564                                 if (unlikely(need_more)) {
 565                                         int i, this_cnt = 0;
 566                                         for (i = 0; i < cnt; i++) {
 567                                                 if (cpu_list[i] == 0xffff)
 568                                                         continue;
 569                                                 cpu_list[i] = 0xffff;
 570                                                 this_cnt++;
 571                                                 if (this_cnt == 32)
 572                                                         break;
 573                                         }
 574                                         goto retry;
 575                                 }
 576                                 return;
 577                         }
 578                         if (!--stuck)
 579                                 break;
 580                 } while (dispatch_stat & busy_mask);
 581
 582                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 583                                      : : "r" (pstate));
 584
 585                 if (dispatch_stat & busy_mask) {
 586                         /* Busy bits will not clear, continue instead
 587                          * of freezing up on this cpu.
 588                          */
 589                         printk("CPU[%d]: mondo stuckage result[%016llx]\n",
 590                                smp_processor_id(), dispatch_stat);
 591                 } else {
 592                         int i, this_busy_nack = 0;
 593
 594                         /* Delay some random time with interrupts enabled
 595                          * to prevent deadlock.
 596                          */
 597                         udelay(2 * nack_busy_id);
 598
 599                         /* Clear out the mask bits for cpus which did not
 600                          * NACK us.
 601                          */
 602                         for (i = 0; i < cnt; i++) {
 603                                 u64 check_mask, nr;
 604
 605                                 nr = cpu_list[i];
 606                                 if (nr == 0xffff)
 607                                         continue;
 608
 609                                 if (is_jbus)
 610                                         check_mask = (0x2UL << (2*nr));
 611                                 else
 612                                         check_mask = (0x2UL <<
 613                                                       this_busy_nack);
 614                                 if ((dispatch_stat & check_mask) == 0)
 615                                         cpu_list[i] = 0xffff;
 616                                 this_busy_nack += 2;
 617                                 if (this_busy_nack == 64)
 618                                         break;
 619                         }
 620
 621                         goto retry;
 622                 }
 623         }
 624 }
 625
 626 #define CPU_MONDO_COUNTER(cpuid)        (cpu_mondo_counter[cpuid])
 627 #define MONDO_USEC_WAIT_MIN             2
 628 #define MONDO_USEC_WAIT_MAX             100
 629 #define MONDO_RETRY_LIMIT               500000
 630
 631 /* Multi-cpu list version.
 632  *
 633  * Deliver xcalls to 'cnt' number of cpus in 'cpu_list'.
 634  * Sometimes not all cpus receive the mondo, requiring us to re-send
 635  * the mondo until all cpus have received, or cpus are truly stuck
 636  * unable to receive mondo, and we timeout.
 637  * Occasionally a target cpu strand is borrowed briefly by hypervisor to
 638  * perform guest service, such as PCIe error handling. Consider the
 639  * service time, 1 second overall wait is reasonable for 1 cpu.
 640  * Here two in-between mondo check wait time are defined: 2 usec for
 641  * single cpu quick turn around and up to 100usec for large cpu count.
 642  * Deliver mondo to large number of cpus could take longer, we adjusts
 643  * the retry count as long as target cpus are making forward progress.
 644  */
 645 static void hypervisor_xcall_deliver(struct trap_per_cpu *tb, int cnt)
 646 {
 647         int this_cpu, tot_cpus, prev_sent, i, rem;
 648         int usec_wait, retries, tot_retries;
 649         u16 first_cpu = 0xffff;
 650         unsigned long xc_rcvd = 0;
 651         unsigned long status;
 652         int ecpuerror_id = 0;
 653         int enocpu_id = 0;
 654         u16 *cpu_list;
 655         u16 cpu;
 656
 657         this_cpu = smp_processor_id();
 658         cpu_list = __va(tb->cpu_list_pa);
 659         usec_wait = cnt * MONDO_USEC_WAIT_MIN;
 660         if (usec_wait > MONDO_USEC_WAIT_MAX)
 661                 usec_wait = MONDO_USEC_WAIT_MAX;
 662         retries = tot_retries = 0;
 663         tot_cpus = cnt;
 664         prev_sent = 0;
 665
 666         do {
 667                 int n_sent, mondo_delivered, target_cpu_busy;
 668
 669                 status = sun4v_cpu_mondo_send(cnt,
 670                                               tb->cpu_list_pa,
 671                                               tb->cpu_mondo_block_pa);
 672
 673                 /* HV_EOK means all cpus received the xcall, we're done.  */
 674                 if (likely(status == HV_EOK))
 675                         goto xcall_done;
 676
 677                 /* If not these non-fatal errors, panic */
 678                 if (unlikely((status != HV_EWOULDBLOCK) &&
 679                         (status != HV_ECPUERROR) &&
 680                         (status != HV_ENOCPU)))
 681                         goto fatal_errors;
 682
 683                 /* First, see if we made any forward progress.
 684                  *
 685                  * Go through the cpu_list, count the target cpus that have
 686                  * received our mondo (n_sent), and those that did not (rem).
 687                  * Re-pack cpu_list with the cpus remain to be retried in the
 688                  * front - this simplifies tracking the truly stalled cpus.
 689                  *
 690                  * The hypervisor indicates successful sends by setting
 691                  * cpu list entries to the value 0xffff.
 692                  *
 693                  * EWOULDBLOCK means some target cpus did not receive the
 694                  * mondo and retry usually helps.
 695                  *
 696                  * ECPUERROR means at least one target cpu is in error state,
 697                  * it's usually safe to skip the faulty cpu and retry.
 698                  *
 699                  * ENOCPU means one of the target cpu doesn't belong to the
 700                  * domain, perhaps offlined which is unexpected, but not
 701                  * fatal and it's okay to skip the offlined cpu.
 702                  */
 703                 rem = 0;
 704                 n_sent = 0;
 705                 for (i = 0; i < cnt; i++) {
 706                         cpu = cpu_list[i];
 707                         if (likely(cpu == 0xffff)) {
 708                                 n_sent++;
 709                         } else if ((status == HV_ECPUERROR) &&
 710                                 (sun4v_cpu_state(cpu) == HV_CPU_STATE_ERROR)) {
 711                                 ecpuerror_id = cpu + 1;
 712                         } else if (status == HV_ENOCPU && !cpu_online(cpu)) {
 713                                 enocpu_id = cpu + 1;
 714                         } else {
 715                                 cpu_list[rem++] = cpu;
 716                         }
 717                 }
 718
 719                 /* No cpu remained, we're done. */
 720                 if (rem == 0)
 721                         break;
 722
 723                 /* Otherwise, update the cpu count for retry. */
 724                 cnt = rem;
 725
 726                 /* Record the overall number of mondos received by the
 727                  * first of the remaining cpus.
 728                  */
 729                 if (first_cpu != cpu_list[0]) {
 730                         first_cpu = cpu_list[0];
 731                         xc_rcvd = CPU_MONDO_COUNTER(first_cpu);
 732                 }
 733
 734                 /* Was any mondo delivered successfully? */
 735                 mondo_delivered = (n_sent > prev_sent);
 736                 prev_sent = n_sent;
 737
 738                 /* or, was any target cpu busy processing other mondos? */
 739                 target_cpu_busy = (xc_rcvd < CPU_MONDO_COUNTER(first_cpu));
 740                 xc_rcvd = CPU_MONDO_COUNTER(first_cpu);
 741
 742                 /* Retry count is for no progress. If we're making progress,
 743                  * reset the retry count.
 744                  */
 745                 if (likely(mondo_delivered || target_cpu_busy)) {
 746                         tot_retries += retries;
 747                         retries = 0;
 748                 } else if (unlikely(retries > MONDO_RETRY_LIMIT)) {
 749                         goto fatal_mondo_timeout;
 750                 }
 751
 752                 /* Delay a little bit to let other cpus catch up on
 753                  * their cpu mondo queue work.
 754                  */
 755                 if (!mondo_delivered)
 756                         udelay(usec_wait);
 757
 758                 retries++;
 759         } while (1);
 760
 761 xcall_done:
 762         if (unlikely(ecpuerror_id > 0)) {
 763                 pr_crit("CPU[%d]: SUN4V mondo cpu error, target cpu(%d) was in error state\n",
 764                        this_cpu, ecpuerror_id - 1);
 765         } else if (unlikely(enocpu_id > 0)) {
 766                 pr_crit("CPU[%d]: SUN4V mondo cpu error, target cpu(%d) does not belong to the domain\n",
 767                        this_cpu, enocpu_id - 1);
 768         }
 769         return;
 770
 771 fatal_errors:
 772         /* fatal errors include bad alignment, etc */
 773         pr_crit("CPU[%d]: Args were cnt(%d) cpulist_pa(%lx) mondo_block_pa(%lx)\n",
 774                this_cpu, tot_cpus, tb->cpu_list_pa, tb->cpu_mondo_block_pa);
 775         panic("Unexpected SUN4V mondo error %lu\n", status);
 776
 777 fatal_mondo_timeout:
 778         /* some cpus being non-responsive to the cpu mondo */
 779         pr_crit("CPU[%d]: SUN4V mondo timeout, cpu(%d) made no forward progress after %d retries. Total target cpus(%d).\n",
 780                this_cpu, first_cpu, (tot_retries + retries), tot_cpus);
 781         panic("SUN4V mondo timeout panic\n");
 782 }
 783
 784 static void (*xcall_deliver_impl)(struct trap_per_cpu *, int);
 785
 786 static void xcall_deliver(u64 data0, u64 data1, u64 data2, const cpumask_t *mask)
 787 {
 788         struct trap_per_cpu *tb;
 789         int this_cpu, i, cnt;
 790         unsigned long flags;
 791         u16 *cpu_list;
 792         u64 *mondo;
 793
 794         /* We have to do this whole thing with interrupts fully disabled.
 795          * Otherwise if we send an xcall from interrupt context it will
 796          * corrupt both our mondo block and cpu list state.
 797          *
 798          * One consequence of this is that we cannot use timeout mechanisms
 799          * that depend upon interrupts being delivered locally.  So, for
 800          * example, we cannot sample jiffies and expect it to advance.
 801          *
 802          * Fortunately, udelay() uses %stick/%tick so we can use that.
 803          */
 804         local_irq_save(flags);
 805
 806         this_cpu = smp_processor_id();
 807         tb = &trap_block[this_cpu];
 808
 809         mondo = __va(tb->cpu_mondo_block_pa);
 810         mondo[0] = data0;
 811         mondo[1] = data1;
 812         mondo[2] = data2;
 813         wmb();
 814
 815         cpu_list = __va(tb->cpu_list_pa);
 816
 817         /* Setup the initial cpu list.  */
 818         cnt = 0;
 819         for_each_cpu(i, mask) {
 820                 if (i == this_cpu || !cpu_online(i))
 821                         continue;
 822                 cpu_list[cnt++] = i;
 823         }
 824
 825         if (cnt)
 826                 xcall_deliver_impl(tb, cnt);
 827
 828         local_irq_restore(flags);
 829 }
 830
 831 /* Send cross call to all processors mentioned in MASK_P
 832  * except self.  Really, there are only two cases currently,
 833  * "cpu_online_mask" and "mm_cpumask(mm)".
 834  */
 835 static void smp_cross_call_masked(unsigned long *func, u32 ctx, u64 data1, u64 data2, const cpumask_t *mask)
 836 {
 837         u64 data0 = (((u64)ctx)<<32 | (((u64)func) & 0xffffffff));
 838
 839         xcall_deliver(data0, data1, data2, mask);
 840 }
 841
 842 /* Send cross call to all processors except self. */
 843 static void smp_cross_call(unsigned long *func, u32 ctx, u64 data1, u64 data2)
 844 {
 845         smp_cross_call_masked(func, ctx, data1, data2, cpu_online_mask);
 846 }
 847
 848 extern unsigned long xcall_sync_tick;
 849
 850 static void smp_start_sync_tick_client(int cpu)
 851 {
 852         xcall_deliver((u64) &xcall_sync_tick, 0, 0,
 853                       cpumask_of(cpu));
 854 }
 855
 856 extern unsigned long xcall_call_function;
 857
 858 void arch_send_call_function_ipi_mask(const struct cpumask *mask)
 859 {
 860         xcall_deliver((u64) &xcall_call_function, 0, 0, mask);
 861 }
 862
 863 extern unsigned long xcall_call_function_single;
 864
 865 void arch_send_call_function_single_ipi(int cpu)
 866 {
 867         xcall_deliver((u64) &xcall_call_function_single, 0, 0,
 868                       cpumask_of(cpu));
 869 }
 870
 871 void __irq_entry smp_call_function_client(int irq, struct pt_regs *regs)
 872 {
 873         clear_softint(1 << irq);
 874         irq_enter();
 875         generic_smp_call_function_interrupt();
 876         irq_exit();
 877 }
 878
 879 void __irq_entry smp_call_function_single_client(int irq, struct pt_regs *regs)
 880 {
 881         clear_softint(1 << irq);
 882         irq_enter();
 883         generic_smp_call_function_single_interrupt();
 884         irq_exit();
 885 }
 886
 887 static void tsb_sync(void *info)
 888 {
 889         struct trap_per_cpu *tp = &trap_block[raw_smp_processor_id()];
 890         struct mm_struct *mm = info;
 891
 892         /* It is not valid to test "current->active_mm == mm" here.
 893          *
 894          * The value of "current" is not changed atomically with
 895          * switch_mm().  But that's OK, we just need to check the
 896          * current cpu's trap block PGD physical address.
 897          */
 898         if (tp->pgd_paddr == __pa(mm->pgd))
 899                 tsb_context_switch(mm);
 900 }
 901
 902 void smp_tsb_sync(struct mm_struct *mm)
 903 {
 904         smp_call_function_many(mm_cpumask(mm), tsb_sync, mm, 1);
 905 }
 906
 907 extern unsigned long xcall_flush_tlb_mm;
 908 extern unsigned long xcall_flush_tlb_page;
 909 extern unsigned long xcall_flush_tlb_kernel_range;
 910 extern unsigned long xcall_fetch_glob_regs;
 911 extern unsigned long xcall_fetch_glob_pmu;
 912 extern unsigned long xcall_fetch_glob_pmu_n4;
 913 extern unsigned long xcall_receive_signal;
 914 extern unsigned long xcall_new_mmu_context_version;
 915 #ifdef CONFIG_KGDB
 916 extern unsigned long xcall_kgdb_capture;
 917 #endif
 918
 919 #ifdef DCACHE_ALIASING_POSSIBLE
 920 extern unsigned long xcall_flush_dcache_page_cheetah;
 921 #endif
 922 extern unsigned long xcall_flush_dcache_page_spitfire;
 923
 924 static inline void __local_flush_dcache_page(struct page *page)
 925 {
 926 #ifdef DCACHE_ALIASING_POSSIBLE
 927         __flush_dcache_page(page_address(page),
 928                             ((tlb_type == spitfire) &&
 929                              page_mapping_file(page) != NULL));
 930 #else
 931         if (page_mapping_file(page) != NULL &&
 932             tlb_type == spitfire)
 933                 __flush_icache_page(__pa(page_address(page)));
 934 #endif
 935 }
 936
 937 void smp_flush_dcache_page_impl(struct page *page, int cpu)
 938 {
 939         int this_cpu;
 940
 941         if (tlb_type == hypervisor)
 942                 return;
 943
 944 #ifdef CONFIG_DEBUG_DCFLUSH
 945         atomic_inc(&dcpage_flushes);
 946 #endif
 947
 948         this_cpu = get_cpu();
 949
 950         if (cpu == this_cpu) {
 951                 __local_flush_dcache_page(page);
 952         } else if (cpu_online(cpu)) {
 953                 void *pg_addr = page_address(page);
 954                 u64 data0 = 0;
 955
 956                 if (tlb_type == spitfire) {
 957                         data0 = ((u64)&xcall_flush_dcache_page_spitfire);
 958                         if (page_mapping_file(page) != NULL)
 959                                 data0 |= ((u64)1 << 32);
 960                 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 961 #ifdef DCACHE_ALIASING_POSSIBLE
 962                         data0 = ((u64)&xcall_flush_dcache_page_cheetah);
 963 #endif
 964                 }
 965                 if (data0) {
 966                         xcall_deliver(data0, __pa(pg_addr),
 967                                       (u64) pg_addr, cpumask_of(cpu));
 968 #ifdef CONFIG_DEBUG_DCFLUSH
 969                         atomic_inc(&dcpage_flushes_xcall);
 970 #endif
 971                 }
 972         }
 973
 974         put_cpu();
 975 }
 976
 977 void flush_dcache_page_all(struct mm_struct *mm, struct page *page)
 978 {
 979         void *pg_addr;
 980         u64 data0;
 981
 982         if (tlb_type == hypervisor)
 983                 return;
 984
 985         preempt_disable();
 986
 987 #ifdef CONFIG_DEBUG_DCFLUSH
 988         atomic_inc(&dcpage_flushes);
 989 #endif
 990         data0 = 0;
 991         pg_addr = page_address(page);
 992         if (tlb_type == spitfire) {
 993                 data0 = ((u64)&xcall_flush_dcache_page_spitfire);
 994                 if (page_mapping_file(page) != NULL)
 995                         data0 |= ((u64)1 << 32);
 996         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 997 #ifdef DCACHE_ALIASING_POSSIBLE
 998                 data0 = ((u64)&xcall_flush_dcache_page_cheetah);
 999 #endif
1000         }
1001         if (data0) {
1002                 xcall_deliver(data0, __pa(pg_addr),
1003                               (u64) pg_addr, cpu_online_mask);
1004 #ifdef CONFIG_DEBUG_DCFLUSH
1005                 atomic_inc(&dcpage_flushes_xcall);
1006 #endif
1007         }
1008         __local_flush_dcache_page(page);
1009
1010         preempt_enable();
1011 }
1012
1013 #ifdef CONFIG_KGDB
1014 void kgdb_roundup_cpus(void)
1015 {
1016         smp_cross_call(&xcall_kgdb_capture, 0, 0, 0);
1017 }
1018 #endif
1019
1020 void smp_fetch_global_regs(void)
1021 {
1022         smp_cross_call(&xcall_fetch_glob_regs, 0, 0, 0);
1023 }
1024
1025 void smp_fetch_global_pmu(void)
1026 {
1027         if (tlb_type == hypervisor &&
1028             sun4v_chip_type >= SUN4V_CHIP_NIAGARA4)
1029                 smp_cross_call(&xcall_fetch_glob_pmu_n4, 0, 0, 0);
1030         else
1031                 smp_cross_call(&xcall_fetch_glob_pmu, 0, 0, 0);
1032 }
1033
1034 /* We know that the window frames of the user have been flushed
1035  * to the stack before we get here because all callers of us
1036  * are flush_tlb_*() routines, and these run after flush_cache_*()
1037  * which performs the flushw.
1038  *
1039  * mm->cpu_vm_mask is a bit mask of which cpus an address
1040  * space has (potentially) executed on, this is the heuristic
1041  * we use to limit cross calls.
1042  */
1043
1044 /* This currently is only used by the hugetlb arch pre-fault
1045  * hook on UltraSPARC-III+ and later when changing the pagesize
1046  * bits of the context register for an address space.
1047  */
1048 void smp_flush_tlb_mm(struct mm_struct *mm)
1049 {
1050         u32 ctx = CTX_HWBITS(mm->context);
1051
1052         get_cpu();
1053
1054         smp_cross_call_masked(&xcall_flush_tlb_mm,
1055                               ctx, 0, 0,
1056                               mm_cpumask(mm));
1057
1058         __flush_tlb_mm(ctx, SECONDARY_CONTEXT);
1059
1060         put_cpu();
1061 }
1062
1063 struct tlb_pending_info {
1064         unsigned long ctx;
1065         unsigned long nr;
1066         unsigned long *vaddrs;
1067 };
1068
1069 static void tlb_pending_func(void *info)
1070 {
1071         struct tlb_pending_info *t = info;
1072
1073         __flush_tlb_pending(t->ctx, t->nr, t->vaddrs);
1074 }
1075
1076 void smp_flush_tlb_pending(struct mm_struct *mm, unsigned long nr, unsigned long *vaddrs)
1077 {
1078         u32 ctx = CTX_HWBITS(mm->context);
1079         struct tlb_pending_info info;
1080
1081         get_cpu();
1082
1083         info.ctx = ctx;
1084         info.nr = nr;
1085         info.vaddrs = vaddrs;
1086
1087         smp_call_function_many(mm_cpumask(mm), tlb_pending_func,
1088                                &info, 1);
1089
1090         __flush_tlb_pending(ctx, nr, vaddrs);
1091
1092         put_cpu();
1093 }
1094
1095 void smp_flush_tlb_page(struct mm_struct *mm, unsigned long vaddr)
1096 {
1097         unsigned long context = CTX_HWBITS(mm->context);
1098
1099         get_cpu();
1100
1101         smp_cross_call_masked(&xcall_flush_tlb_page,
1102                               context, vaddr, 0,
1103                               mm_cpumask(mm));
1104
1105         __flush_tlb_page(context, vaddr);
1106
1107         put_cpu();
1108 }
1109
1110 void smp_flush_tlb_kernel_range(unsigned long start, unsigned long end)
1111 {
1112         start &= PAGE_MASK;
1113         end    = PAGE_ALIGN(end);
1114         if (start != end) {
1115                 smp_cross_call(&xcall_flush_tlb_kernel_range,
1116                                0, start, end);
1117
1118                 __flush_tlb_kernel_range(start, end);
1119         }
1120 }
1121
1122 /* CPU capture. */
1123 /* #define CAPTURE_DEBUG */
1124 extern unsigned long xcall_capture;
1125
1126 static atomic_t smp_capture_depth = ATOMIC_INIT(0);
1127 static atomic_t smp_capture_registry = ATOMIC_INIT(0);
1128 static unsigned long penguins_are_doing_time;
1129
1130 void smp_capture(void)
1131 {
1132         int result = atomic_add_return(1, &smp_capture_depth);
1133
1134         if (result == 1) {
1135                 int ncpus = num_online_cpus();
1136
1137 #ifdef CAPTURE_DEBUG
1138                 printk("CPU[%d]: Sending penguins to jail...",
1139                        smp_processor_id());
1140 #endif
1141                 penguins_are_doing_time = 1;
1142                 atomic_inc(&smp_capture_registry);
1143                 smp_cross_call(&xcall_capture, 0, 0, 0);
1144                 while (atomic_read(&smp_capture_registry) != ncpus)
1145                         rmb();
1146 #ifdef CAPTURE_DEBUG
1147                 printk("done\n");
1148 #endif
1149         }
1150 }
1151
1152 void smp_release(void)
1153 {
1154         if (atomic_dec_and_test(&smp_capture_depth)) {
1155 #ifdef CAPTURE_DEBUG
1156                 printk("CPU[%d]: Giving pardon to "
1157                        "imprisoned penguins\n",
1158                        smp_processor_id());
1159 #endif
1160                 penguins_are_doing_time = 0;
1161                 membar_safe("#StoreLoad");
1162                 atomic_dec(&smp_capture_registry);
1163         }
1164 }
1165
1166 /* Imprisoned penguins run with %pil == PIL_NORMAL_MAX, but PSTATE_IE
1167  * set, so they can service tlb flush xcalls...
1168  */
1169 extern void prom_world(int);
1170
1171 void __irq_entry smp_penguin_jailcell(int irq, struct pt_regs *regs)
1172 {
1173         clear_softint(1 << irq);
1174
1175         preempt_disable();
1176
1177         __asm__ __volatile__("flushw");
1178         prom_world(1);
1179         atomic_inc(&smp_capture_registry);
1180         membar_safe("#StoreLoad");
1181         while (penguins_are_doing_time)
1182                 rmb();
1183         atomic_dec(&smp_capture_registry);
1184         prom_world(0);
1185
1186         preempt_enable();
1187 }
1188
1189 /* /proc/profile writes can call this, don't __init it please. */
1190 int setup_profiling_timer(unsigned int multiplier)
1191 {
1192         return -EINVAL;
1193 }
1194
1195 void __init smp_prepare_cpus(unsigned int max_cpus)
1196 {
1197 }
1198
1199 void smp_prepare_boot_cpu(void)
1200 {
1201 }
1202
1203 void __init smp_setup_processor_id(void)
1204 {
1205         if (tlb_type == spitfire)
1206                 xcall_deliver_impl = spitfire_xcall_deliver;
1207         else if (tlb_type == cheetah || tlb_type == cheetah_plus)
1208                 xcall_deliver_impl = cheetah_xcall_deliver;
1209         else
1210                 xcall_deliver_impl = hypervisor_xcall_deliver;
1211 }
1212
1213 void __init smp_fill_in_cpu_possible_map(void)
1214 {
1215         int possible_cpus = num_possible_cpus();
1216         int i;
1217
1218         if (possible_cpus > nr_cpu_ids)
1219                 possible_cpus = nr_cpu_ids;
1220
1221         for (i = 0; i < possible_cpus; i++)
1222                 set_cpu_possible(i, true);
1223         for (; i < NR_CPUS; i++)
1224                 set_cpu_possible(i, false);
1225 }
1226
1227 void smp_fill_in_sib_core_maps(void)
1228 {
1229         unsigned int i;
1230
1231         for_each_present_cpu(i) {
1232                 unsigned int j;
1233
1234                 cpumask_clear(&cpu_core_map[i]);
1235                 if (cpu_data(i).core_id == 0) {
1236                         cpumask_set_cpu(i, &cpu_core_map[i]);
1237                         continue;
1238                 }
1239
1240                 for_each_present_cpu(j) {
1241                         if (cpu_data(i).core_id ==
1242                             cpu_data(j).core_id)
1243                                 cpumask_set_cpu(j, &cpu_core_map[i]);
1244                 }
1245         }
1246
1247         for_each_present_cpu(i)  {
1248                 unsigned int j;
1249
1250                 for_each_present_cpu(j)  {
1251                         if (cpu_data(i).max_cache_id ==
1252                             cpu_data(j).max_cache_id)
1253                                 cpumask_set_cpu(j, &cpu_core_sib_cache_map[i]);
1254
1255                         if (cpu_data(i).sock_id == cpu_data(j).sock_id)
1256                                 cpumask_set_cpu(j, &cpu_core_sib_map[i]);
1257                 }
1258         }
1259
1260         for_each_present_cpu(i) {
1261                 unsigned int j;
1262
1263                 cpumask_clear(&per_cpu(cpu_sibling_map, i));
1264                 if (cpu_data(i).proc_id == -1) {
1265                         cpumask_set_cpu(i, &per_cpu(cpu_sibling_map, i));
1266                         continue;
1267                 }
1268
1269                 for_each_present_cpu(j) {
1270                         if (cpu_data(i).proc_id ==
1271                             cpu_data(j).proc_id)
1272                                 cpumask_set_cpu(j, &per_cpu(cpu_sibling_map, i));
1273                 }
1274         }
1275 }
1276
1277 int __cpu_up(unsigned int cpu, struct task_struct *tidle)
1278 {
1279         int ret = smp_boot_one_cpu(cpu, tidle);
1280
1281         if (!ret) {
1282                 cpumask_set_cpu(cpu, &smp_commenced_mask);
1283                 while (!cpu_online(cpu))
1284                         mb();
1285                 if (!cpu_online(cpu)) {
1286                         ret = -ENODEV;
1287                 } else {
1288                         /* On SUN4V, writes to %tick and %stick are
1289                          * not allowed.
1290                          */
1291                         if (tlb_type != hypervisor)
1292                                 smp_synchronize_one_tick(cpu);
1293                 }
1294         }
1295         return ret;
1296 }
1297
1298 #ifdef CONFIG_HOTPLUG_CPU
1299 void cpu_play_dead(void)
1300 {
1301         int cpu = smp_processor_id();
1302         unsigned long pstate;
1303
1304         idle_task_exit();
1305
1306         if (tlb_type == hypervisor) {
1307                 struct trap_per_cpu *tb = &trap_block[cpu];
1308
1309                 sun4v_cpu_qconf(HV_CPU_QUEUE_CPU_MONDO,
1310                                 tb->cpu_mondo_pa, 0);
1311                 sun4v_cpu_qconf(HV_CPU_QUEUE_DEVICE_MONDO,
1312                                 tb->dev_mondo_pa, 0);
1313                 sun4v_cpu_qconf(HV_CPU_QUEUE_RES_ERROR,
1314                                 tb->resum_mondo_pa, 0);
1315                 sun4v_cpu_qconf(HV_CPU_QUEUE_NONRES_ERROR,
1316                                 tb->nonresum_mondo_pa, 0);
1317         }
1318
1319         cpumask_clear_cpu(cpu, &smp_commenced_mask);
1320         membar_safe("#Sync");
1321
1322         local_irq_disable();
1323
1324         __asm__ __volatile__(
1325                 "rdpr   %%pstate, %0\n\t"
1326                 "wrpr   %0, %1, %%pstate"
1327                 : "=r" (pstate)
1328                 : "i" (PSTATE_IE));
1329
1330         while (1)
1331                 barrier();
1332 }
1333
1334 int __cpu_disable(void)
1335 {
1336         int cpu = smp_processor_id();
1337         cpuinfo_sparc *c;
1338         int i;
1339
1340         for_each_cpu(i, &cpu_core_map[cpu])
1341                 cpumask_clear_cpu(cpu, &cpu_core_map[i]);
1342         cpumask_clear(&cpu_core_map[cpu]);
1343
1344         for_each_cpu(i, &per_cpu(cpu_sibling_map, cpu))
1345                 cpumask_clear_cpu(cpu, &per_cpu(cpu_sibling_map, i));
1346         cpumask_clear(&per_cpu(cpu_sibling_map, cpu));
1347
1348         c = &cpu_data(cpu);
1349
1350         c->core_id = 0;
1351         c->proc_id = -1;
1352
1353         smp_wmb();
1354
1355         /* Make sure no interrupts point to this cpu.  */
1356         fixup_irqs();
1357
1358         local_irq_enable();
1359         mdelay(1);
1360         local_irq_disable();
1361
1362         set_cpu_online(cpu, false);
1363
1364         cpu_map_rebuild();
1365
1366         return 0;
1367 }
1368
1369 void __cpu_die(unsigned int cpu)
1370 {
1371         int i;
1372
1373         for (i = 0; i < 100; i++) {
1374                 smp_rmb();
1375                 if (!cpumask_test_cpu(cpu, &smp_commenced_mask))
1376                         break;
1377                 msleep(100);
1378         }
1379         if (cpumask_test_cpu(cpu, &smp_commenced_mask)) {
1380                 printk(KERN_ERR "CPU %u didn't die...\n", cpu);
1381         } else {
1382 #if defined(CONFIG_SUN_LDOMS)
1383                 unsigned long hv_err;
1384                 int limit = 100;
1385
1386                 do {
1387                         hv_err = sun4v_cpu_stop(cpu);
1388                         if (hv_err == HV_EOK) {
1389                                 set_cpu_present(cpu, false);
1390                                 break;
1391                         }
1392                 } while (--limit > 0);
1393                 if (limit <= 0) {
1394                         printk(KERN_ERR "sun4v_cpu_stop() fails err=%lu\n",
1395                                hv_err);
1396                 }
1397 #endif
1398         }
1399 }
1400 #endif
1401
1402 void __init smp_cpus_done(unsigned int max_cpus)
1403 {
1404 }
1405
1406 static void send_cpu_ipi(int cpu)
1407 {
1408         xcall_deliver((u64) &xcall_receive_signal,
1409                         0, 0, cpumask_of(cpu));
1410 }
1411
1412 void scheduler_poke(void)
1413 {
1414         if (!cpu_poke)
1415                 return;
1416
1417         if (!__this_cpu_read(poke))
1418                 return;
1419
1420         __this_cpu_write(poke, false);
1421         set_softint(1 << PIL_SMP_RECEIVE_SIGNAL);
1422 }
1423
1424 static unsigned long send_cpu_poke(int cpu)
1425 {
1426         unsigned long hv_err;
1427
1428         per_cpu(poke, cpu) = true;
1429         hv_err = sun4v_cpu_poke(cpu);
1430         if (hv_err != HV_EOK) {
1431                 per_cpu(poke, cpu) = false;
1432                 pr_err_ratelimited("%s: sun4v_cpu_poke() fails err=%lu\n",
1433                                     __func__, hv_err);
1434         }
1435
1436         return hv_err;
1437 }
1438
1439 void smp_send_reschedule(int cpu)
1440 {
1441         if (cpu == smp_processor_id()) {
1442                 WARN_ON_ONCE(preemptible());
1443                 set_softint(1 << PIL_SMP_RECEIVE_SIGNAL);
1444                 return;
1445         }
1446
1447         /* Use cpu poke to resume idle cpu if supported. */
1448         if (cpu_poke && idle_cpu(cpu)) {
1449                 unsigned long ret;
1450
1451                 ret = send_cpu_poke(cpu);
1452                 if (ret == HV_EOK)
1453                         return;
1454         }
1455
1456         /* Use IPI in following cases:
1457          * - cpu poke not supported
1458          * - cpu not idle
1459          * - send_cpu_poke() returns with error
1460          */
1461         send_cpu_ipi(cpu);
1462 }
1463
1464 void smp_init_cpu_poke(void)
1465 {
1466         unsigned long major;
1467         unsigned long minor;
1468         int ret;
1469
1470         if (tlb_type != hypervisor)
1471                 return;
1472
1473         ret = sun4v_hvapi_get(HV_GRP_CORE, &major, &minor);
1474         if (ret) {
1475                 pr_debug("HV_GRP_CORE is not registered\n");
1476                 return;
1477         }
1478
1479         if (major == 1 && minor >= 6) {
1480                 /* CPU POKE is registered. */
1481                 cpu_poke = true;
1482                 return;
1483         }
1484
1485         pr_debug("CPU_POKE not supported\n");
1486 }
1487
1488 void __irq_entry smp_receive_signal_client(int irq, struct pt_regs *regs)
1489 {
1490         clear_softint(1 << irq);
1491         scheduler_ipi();
1492 }
1493
1494 static void stop_this_cpu(void *dummy)
1495 {
1496         set_cpu_online(smp_processor_id(), false);
1497         prom_stopself();
1498 }
1499
1500 void smp_send_stop(void)
1501 {
1502         int cpu;
1503
1504         if (tlb_type == hypervisor) {
1505                 int this_cpu = smp_processor_id();
1506 #ifdef CONFIG_SERIAL_SUNHV
1507                 sunhv_migrate_hvcons_irq(this_cpu);
1508 #endif
1509                 for_each_online_cpu(cpu) {
1510                         if (cpu == this_cpu)
1511                                 continue;
1512
1513                         set_cpu_online(cpu, false);
1514 #ifdef CONFIG_SUN_LDOMS
1515                         if (ldom_domaining_enabled) {
1516                                 unsigned long hv_err;
1517                                 hv_err = sun4v_cpu_stop(cpu);
1518                                 if (hv_err)
1519                                         printk(KERN_ERR "sun4v_cpu_stop() "
1520                                                "failed err=%lu\n", hv_err);
1521                         } else
1522 #endif
1523                                 prom_stopcpu_cpuid(cpu);
1524                 }
1525         } else
1526                 smp_call_function(stop_this_cpu, NULL, 0);
1527 }
1528
1529 /**
1530  * pcpu_alloc_bootmem - NUMA friendly alloc_bootmem wrapper for percpu
1531  * @cpu: cpu to allocate for
1532  * @size: size allocation in bytes
1533  * @align: alignment
1534  *
1535  * Allocate @size bytes aligned at @align for cpu @cpu.  This wrapper
1536  * does the right thing for NUMA regardless of the current
1537  * configuration.
1538  *
1539  * RETURNS:
1540  * Pointer to the allocated area on success, NULL on failure.
1541  */
1542 static void * __init pcpu_alloc_bootmem(unsigned int cpu, size_t size,
1543                                         size_t align)
1544 {
1545         const unsigned long goal = __pa(MAX_DMA_ADDRESS);
1546 #ifdef CONFIG_NUMA
1547         int node = cpu_to_node(cpu);
1548         void *ptr;
1549
1550         if (!node_online(node) || !NODE_DATA(node)) {
1551                 ptr = memblock_alloc_from(size, align, goal);
1552                 pr_info("cpu %d has no node %d or node-local memory\n",
1553                         cpu, node);
1554                 pr_debug("per cpu data for cpu%d %lu bytes at %016lx\n",
1555                          cpu, size, __pa(ptr));
1556         } else {
1557                 ptr = memblock_alloc_try_nid(size, align, goal,
1558                                              MEMBLOCK_ALLOC_ACCESSIBLE, node);
1559                 pr_debug("per cpu data for cpu%d %lu bytes on node%d at "
1560                          "%016lx\n", cpu, size, node, __pa(ptr));
1561         }
1562         return ptr;
1563 #else
1564         return memblock_alloc_from(size, align, goal);
1565 #endif
1566 }
1567
1568 static void __init pcpu_free_bootmem(void *ptr, size_t size)
1569 {
1570         memblock_free(__pa(ptr), size);
1571 }
1572
1573 static int __init pcpu_cpu_distance(unsigned int from, unsigned int to)
1574 {
1575         if (cpu_to_node(from) == cpu_to_node(to))
1576                 return LOCAL_DISTANCE;
1577         else
1578                 return REMOTE_DISTANCE;
1579 }
1580
1581 static void __init pcpu_populate_pte(unsigned long addr)
1582 {
1583         pgd_t *pgd = pgd_offset_k(addr);
1584         p4d_t *p4d;
1585         pud_t *pud;
1586         pmd_t *pmd;
1587
1588         if (pgd_none(*pgd)) {
1589                 pud_t *new;
1590
1591                 new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1592                 if (!new)
1593                         goto err_alloc;
1594                 pgd_populate(&init_mm, pgd, new);
1595         }
1596
1597         p4d = p4d_offset(pgd, addr);
1598         if (p4d_none(*p4d)) {
1599                 pud_t *new;
1600
1601                 new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1602                 if (!new)
1603                         goto err_alloc;
1604                 p4d_populate(&init_mm, p4d, new);
1605         }
1606
1607         pud = pud_offset(p4d, addr);
1608         if (pud_none(*pud)) {
1609                 pmd_t *new;
1610
1611                 new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1612                 if (!new)
1613                         goto err_alloc;
1614                 pud_populate(&init_mm, pud, new);
1615         }
1616
1617         pmd = pmd_offset(pud, addr);
1618         if (!pmd_present(*pmd)) {
1619                 pte_t *new;
1620
1621                 new = memblock_alloc_from(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1622                 if (!new)
1623                         goto err_alloc;
1624                 pmd_populate_kernel(&init_mm, pmd, new);
1625         }
1626
1627         return;
1628
1629 err_alloc:
1630         panic("%s: Failed to allocate %lu bytes align=%lx from=%lx\n",
1631               __func__, PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1632 }
1633
1634 void __init setup_per_cpu_areas(void)
1635 {
1636         unsigned long delta;
1637         unsigned int cpu;
1638         int rc = -EINVAL;
1639
1640         if (pcpu_chosen_fc != PCPU_FC_PAGE) {
1641                 rc = pcpu_embed_first_chunk(PERCPU_MODULE_RESERVE,
1642                                             PERCPU_DYNAMIC_RESERVE, 4 << 20,
1643                                             pcpu_cpu_distance,
1644                                             pcpu_alloc_bootmem,
1645                                             pcpu_free_bootmem);
1646                 if (rc)
1647                         pr_warn("PERCPU: %s allocator failed (%d), "
1648                                 "falling back to page size\n",
1649                                 pcpu_fc_names[pcpu_chosen_fc], rc);
1650         }
1651         if (rc < 0)
1652                 rc = pcpu_page_first_chunk(PERCPU_MODULE_RESERVE,
1653                                            pcpu_alloc_bootmem,
1654                                            pcpu_free_bootmem,
1655                                            pcpu_populate_pte);
1656         if (rc < 0)
1657                 panic("cannot initialize percpu area (err=%d)", rc);
1658
1659         delta = (unsigned long)pcpu_base_addr - (unsigned long)__per_cpu_start;
1660         for_each_possible_cpu(cpu)
1661                 __per_cpu_offset(cpu) = delta + pcpu_unit_offsets[cpu];
1662
1663         /* Setup %g5 for the boot cpu.  */
1664         __local_per_cpu_offset = __per_cpu_offset(smp_processor_id());
1665
1666         of_fill_in_cpu_data();
1667         if (tlb_type == hypervisor)
1668                 mdesc_fill_in_cpu_data(cpu_all_mask);
1669 }