git.oblomov.eu Git - linux-2.6/blob - arch/sparc64/kernel/smp.c

   1 /* smp.c: Sparc64 SMP support.
   2  *
   3  * Copyright (C) 1997 David S. Miller (davem@caip.rutgers.edu)
   4  */
   5
   6 #include <linux/module.h>
   7 #include <linux/kernel.h>
   8 #include <linux/sched.h>
   9 #include <linux/mm.h>
  10 #include <linux/pagemap.h>
  11 #include <linux/threads.h>
  12 #include <linux/smp.h>
  13 #include <linux/smp_lock.h>
  14 #include <linux/interrupt.h>
  15 #include <linux/kernel_stat.h>
  16 #include <linux/delay.h>
  17 #include <linux/init.h>
  18 #include <linux/spinlock.h>
  19 #include <linux/fs.h>
  20 #include <linux/seq_file.h>
  21 #include <linux/cache.h>
  22 #include <linux/jiffies.h>
  23 #include <linux/profile.h>
  24 #include <linux/bootmem.h>
  25
  26 #include <asm/head.h>
  27 #include <asm/ptrace.h>
  28 #include <asm/atomic.h>
  29 #include <asm/tlbflush.h>
  30 #include <asm/mmu_context.h>
  31 #include <asm/cpudata.h>
  32
  33 #include <asm/irq.h>
  34 #include <asm/page.h>
  35 #include <asm/pgtable.h>
  36 #include <asm/oplib.h>
  37 #include <asm/uaccess.h>
  38 #include <asm/timer.h>
  39 #include <asm/starfire.h>
  40 #include <asm/tlb.h>
  41 #include <asm/sections.h>
  42
  43 extern void calibrate_delay(void);
  44
  45 /* Please don't make this stuff initdata!!!  --DaveM */
  46 static unsigned char boot_cpu_id;
  47
  48 cpumask_t cpu_online_map __read_mostly = CPU_MASK_NONE;
  49 cpumask_t phys_cpu_present_map __read_mostly = CPU_MASK_NONE;
  50 static cpumask_t smp_commenced_mask;
  51 static cpumask_t cpu_callout_map;
  52
  53 void smp_info(struct seq_file *m)
  54 {
  55         int i;
  56
  57         seq_printf(m, "State:\n");
  58         for (i = 0; i < NR_CPUS; i++) {
  59                 if (cpu_online(i))
  60                         seq_printf(m,
  61                                    "CPU%d:\t\tonline\n", i);
  62         }
  63 }
  64
  65 void smp_bogo(struct seq_file *m)
  66 {
  67         int i;
  68
  69         for (i = 0; i < NR_CPUS; i++)
  70                 if (cpu_online(i))
  71                         seq_printf(m,
  72                                    "Cpu%dBogo\t: %lu.%02lu\n"
  73                                    "Cpu%dClkTck\t: %016lx\n",
  74                                    i, cpu_data(i).udelay_val / (500000/HZ),
  75                                    (cpu_data(i).udelay_val / (5000/HZ)) % 100,
  76                                    i, cpu_data(i).clock_tick);
  77 }
  78
  79 void __init smp_store_cpu_info(int id)
  80 {
  81         int cpu_node;
  82
  83         /* multiplier and counter set by
  84            smp_setup_percpu_timer()  */
  85         cpu_data(id).udelay_val                 = loops_per_jiffy;
  86
  87         cpu_find_by_mid(id, &cpu_node);
  88         cpu_data(id).clock_tick = prom_getintdefault(cpu_node,
  89                                                      "clock-frequency", 0);
  90
  91         cpu_data(id).idle_volume                = 1;
  92
  93         cpu_data(id).dcache_size = prom_getintdefault(cpu_node, "dcache-size",
  94                                                       16 * 1024);
  95         cpu_data(id).dcache_line_size =
  96                 prom_getintdefault(cpu_node, "dcache-line-size", 32);
  97         cpu_data(id).icache_size = prom_getintdefault(cpu_node, "icache-size",
  98                                                       16 * 1024);
  99         cpu_data(id).icache_line_size =
 100                 prom_getintdefault(cpu_node, "icache-line-size", 32);
 101         cpu_data(id).ecache_size = prom_getintdefault(cpu_node, "ecache-size",
 102                                                       4 * 1024 * 1024);
 103         cpu_data(id).ecache_line_size =
 104                 prom_getintdefault(cpu_node, "ecache-line-size", 64);
 105         printk("CPU[%d]: Caches "
 106                "D[sz(%d):line_sz(%d)] "
 107                "I[sz(%d):line_sz(%d)] "
 108                "E[sz(%d):line_sz(%d)]\n",
 109                id,
 110                cpu_data(id).dcache_size, cpu_data(id).dcache_line_size,
 111                cpu_data(id).icache_size, cpu_data(id).icache_line_size,
 112                cpu_data(id).ecache_size, cpu_data(id).ecache_line_size);
 113 }
 114
 115 static void smp_setup_percpu_timer(void);
 116
 117 static volatile unsigned long callin_flag = 0;
 118
 119 void __init smp_callin(void)
 120 {
 121         int cpuid = hard_smp_processor_id();
 122
 123         __local_per_cpu_offset = __per_cpu_offset(cpuid);
 124
 125         if (tlb_type == hypervisor) {
 126                 sun4v_register_fault_status();
 127                 sun4v_ktsb_register();
 128         }
 129
 130         __flush_tlb_all();
 131
 132         smp_setup_percpu_timer();
 133
 134         if (cheetah_pcache_forced_on)
 135                 cheetah_enable_pcache();
 136
 137         local_irq_enable();
 138
 139         calibrate_delay();
 140         smp_store_cpu_info(cpuid);
 141         callin_flag = 1;
 142         __asm__ __volatile__("membar #Sync\n\t"
 143                              "flush  %%g6" : : : "memory");
 144
 145         /* Clear this or we will die instantly when we
 146          * schedule back to this idler...
 147          */
 148         current_thread_info()->new_child = 0;
 149
 150         /* Attach to the address space of init_task. */
 151         atomic_inc(&init_mm.mm_count);
 152         current->active_mm = &init_mm;
 153
 154         while (!cpu_isset(cpuid, smp_commenced_mask))
 155                 rmb();
 156
 157         cpu_set(cpuid, cpu_online_map);
 158
 159         /* idle thread is expected to have preempt disabled */
 160         preempt_disable();
 161 }
 162
 163 void cpu_panic(void)
 164 {
 165         printk("CPU[%d]: Returns from cpu_idle!\n", smp_processor_id());
 166         panic("SMP bolixed\n");
 167 }
 168
 169 static unsigned long current_tick_offset __read_mostly;
 170
 171 /* This tick register synchronization scheme is taken entirely from
 172  * the ia64 port, see arch/ia64/kernel/smpboot.c for details and credit.
 173  *
 174  * The only change I've made is to rework it so that the master
 175  * initiates the synchonization instead of the slave. -DaveM
 176  */
 177
 178 #define MASTER  0
 179 #define SLAVE   (SMP_CACHE_BYTES/sizeof(unsigned long))
 180
 181 #define NUM_ROUNDS      64      /* magic value */
 182 #define NUM_ITERS       5       /* likewise */
 183
 184 static DEFINE_SPINLOCK(itc_sync_lock);
 185 static unsigned long go[SLAVE + 1];
 186
 187 #define DEBUG_TICK_SYNC 0
 188
 189 static inline long get_delta (long *rt, long *master)
 190 {
 191         unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
 192         unsigned long tcenter, t0, t1, tm;
 193         unsigned long i;
 194
 195         for (i = 0; i < NUM_ITERS; i++) {
 196                 t0 = tick_ops->get_tick();
 197                 go[MASTER] = 1;
 198                 membar_storeload();
 199                 while (!(tm = go[SLAVE]))
 200                         rmb();
 201                 go[SLAVE] = 0;
 202                 wmb();
 203                 t1 = tick_ops->get_tick();
 204
 205                 if (t1 - t0 < best_t1 - best_t0)
 206                         best_t0 = t0, best_t1 = t1, best_tm = tm;
 207         }
 208
 209         *rt = best_t1 - best_t0;
 210         *master = best_tm - best_t0;
 211
 212         /* average best_t0 and best_t1 without overflow: */
 213         tcenter = (best_t0/2 + best_t1/2);
 214         if (best_t0 % 2 + best_t1 % 2 == 2)
 215                 tcenter++;
 216         return tcenter - best_tm;
 217 }
 218
 219 void smp_synchronize_tick_client(void)
 220 {
 221         long i, delta, adj, adjust_latency = 0, done = 0;
 222         unsigned long flags, rt, master_time_stamp, bound;
 223 #if DEBUG_TICK_SYNC
 224         struct {
 225                 long rt;        /* roundtrip time */
 226                 long master;    /* master's timestamp */
 227                 long diff;      /* difference between midpoint and master's timestamp */
 228                 long lat;       /* estimate of itc adjustment latency */
 229         } t[NUM_ROUNDS];
 230 #endif
 231
 232         go[MASTER] = 1;
 233
 234         while (go[MASTER])
 235                 rmb();
 236
 237         local_irq_save(flags);
 238         {
 239                 for (i = 0; i < NUM_ROUNDS; i++) {
 240                         delta = get_delta(&rt, &master_time_stamp);
 241                         if (delta == 0) {
 242                                 done = 1;       /* let's lock on to this... */
 243                                 bound = rt;
 244                         }
 245
 246                         if (!done) {
 247                                 if (i > 0) {
 248                                         adjust_latency += -delta;
 249                                         adj = -delta + adjust_latency/4;
 250                                 } else
 251                                         adj = -delta;
 252
 253                                 tick_ops->add_tick(adj, current_tick_offset);
 254                         }
 255 #if DEBUG_TICK_SYNC
 256                         t[i].rt = rt;
 257                         t[i].master = master_time_stamp;
 258                         t[i].diff = delta;
 259                         t[i].lat = adjust_latency/4;
 260 #endif
 261                 }
 262         }
 263         local_irq_restore(flags);
 264
 265 #if DEBUG_TICK_SYNC
 266         for (i = 0; i < NUM_ROUNDS; i++)
 267                 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
 268                        t[i].rt, t[i].master, t[i].diff, t[i].lat);
 269 #endif
 270
 271         printk(KERN_INFO "CPU %d: synchronized TICK with master CPU (last diff %ld cycles,"
 272                "maxerr %lu cycles)\n", smp_processor_id(), delta, rt);
 273 }
 274
 275 static void smp_start_sync_tick_client(int cpu);
 276
 277 static void smp_synchronize_one_tick(int cpu)
 278 {
 279         unsigned long flags, i;
 280
 281         go[MASTER] = 0;
 282
 283         smp_start_sync_tick_client(cpu);
 284
 285         /* wait for client to be ready */
 286         while (!go[MASTER])
 287                 rmb();
 288
 289         /* now let the client proceed into his loop */
 290         go[MASTER] = 0;
 291         membar_storeload();
 292
 293         spin_lock_irqsave(&itc_sync_lock, flags);
 294         {
 295                 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; i++) {
 296                         while (!go[MASTER])
 297                                 rmb();
 298                         go[MASTER] = 0;
 299                         wmb();
 300                         go[SLAVE] = tick_ops->get_tick();
 301                         membar_storeload();
 302                 }
 303         }
 304         spin_unlock_irqrestore(&itc_sync_lock, flags);
 305 }
 306
 307 extern unsigned long sparc64_cpu_startup;
 308
 309 /* The OBP cpu startup callback truncates the 3rd arg cookie to
 310  * 32-bits (I think) so to be safe we have it read the pointer
 311  * contained here so we work on >4GB machines. -DaveM
 312  */
 313 static struct thread_info *cpu_new_thread = NULL;
 314
 315 static int __devinit smp_boot_one_cpu(unsigned int cpu)
 316 {
 317         unsigned long entry =
 318                 (unsigned long)(&sparc64_cpu_startup);
 319         unsigned long cookie =
 320                 (unsigned long)(&cpu_new_thread);
 321         struct task_struct *p;
 322         int timeout, ret, cpu_node;
 323
 324         p = fork_idle(cpu);
 325         callin_flag = 0;
 326         cpu_new_thread = task_thread_info(p);
 327         cpu_set(cpu, cpu_callout_map);
 328
 329         cpu_find_by_mid(cpu, &cpu_node);
 330         prom_startcpu(cpu_node, entry, cookie);
 331
 332         for (timeout = 0; timeout < 5000000; timeout++) {
 333                 if (callin_flag)
 334                         break;
 335                 udelay(100);
 336         }
 337         if (callin_flag) {
 338                 ret = 0;
 339         } else {
 340                 printk("Processor %d is stuck.\n", cpu);
 341                 cpu_clear(cpu, cpu_callout_map);
 342                 ret = -ENODEV;
 343         }
 344         cpu_new_thread = NULL;
 345
 346         return ret;
 347 }
 348
 349 static void spitfire_xcall_helper(u64 data0, u64 data1, u64 data2, u64 pstate, unsigned long cpu)
 350 {
 351         u64 result, target;
 352         int stuck, tmp;
 353
 354         if (this_is_starfire) {
 355                 /* map to real upaid */
 356                 cpu = (((cpu & 0x3c) << 1) |
 357                         ((cpu & 0x40) >> 4) |
 358                         (cpu & 0x3));
 359         }
 360
 361         target = (cpu << 14) | 0x70;
 362 again:
 363         /* Ok, this is the real Spitfire Errata #54.
 364          * One must read back from a UDB internal register
 365          * after writes to the UDB interrupt dispatch, but
 366          * before the membar Sync for that write.
 367          * So we use the high UDB control register (ASI 0x7f,
 368          * ADDR 0x20) for the dummy read. -DaveM
 369          */
 370         tmp = 0x40;
 371         __asm__ __volatile__(
 372         "wrpr   %1, %2, %%pstate\n\t"
 373         "stxa   %4, [%0] %3\n\t"
 374         "stxa   %5, [%0+%8] %3\n\t"
 375         "add    %0, %8, %0\n\t"
 376         "stxa   %6, [%0+%8] %3\n\t"
 377         "membar #Sync\n\t"
 378         "stxa   %%g0, [%7] %3\n\t"
 379         "membar #Sync\n\t"
 380         "mov    0x20, %%g1\n\t"
 381         "ldxa   [%%g1] 0x7f, %%g0\n\t"
 382         "membar #Sync"
 383         : "=r" (tmp)
 384         : "r" (pstate), "i" (PSTATE_IE), "i" (ASI_INTR_W),
 385           "r" (data0), "r" (data1), "r" (data2), "r" (target),
 386           "r" (0x10), "0" (tmp)
 387         : "g1");
 388
 389         /* NOTE: PSTATE_IE is still clear. */
 390         stuck = 100000;
 391         do {
 392                 __asm__ __volatile__("ldxa [%%g0] %1, %0"
 393                         : "=r" (result)
 394                         : "i" (ASI_INTR_DISPATCH_STAT));
 395                 if (result == 0) {
 396                         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 397                                              : : "r" (pstate));
 398                         return;
 399                 }
 400                 stuck -= 1;
 401                 if (stuck == 0)
 402                         break;
 403         } while (result & 0x1);
 404         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 405                              : : "r" (pstate));
 406         if (stuck == 0) {
 407                 printk("CPU[%d]: mondo stuckage result[%016lx]\n",
 408                        smp_processor_id(), result);
 409         } else {
 410                 udelay(2);
 411                 goto again;
 412         }
 413 }
 414
 415 static __inline__ void spitfire_xcall_deliver(u64 data0, u64 data1, u64 data2, cpumask_t mask)
 416 {
 417         u64 pstate;
 418         int i;
 419
 420         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 421         for_each_cpu_mask(i, mask)
 422                 spitfire_xcall_helper(data0, data1, data2, pstate, i);
 423 }
 424
 425 /* Cheetah now allows to send the whole 64-bytes of data in the interrupt
 426  * packet, but we have no use for that.  However we do take advantage of
 427  * the new pipelining feature (ie. dispatch to multiple cpus simultaneously).
 428  */
 429 static void cheetah_xcall_deliver(u64 data0, u64 data1, u64 data2, cpumask_t mask)
 430 {
 431         u64 pstate, ver;
 432         int nack_busy_id, is_jbus;
 433
 434         if (cpus_empty(mask))
 435                 return;
 436
 437         /* Unfortunately, someone at Sun had the brilliant idea to make the
 438          * busy/nack fields hard-coded by ITID number for this Ultra-III
 439          * derivative processor.
 440          */
 441         __asm__ ("rdpr %%ver, %0" : "=r" (ver));
 442         is_jbus = ((ver >> 32) == __JALAPENO_ID ||
 443                    (ver >> 32) == __SERRANO_ID);
 444
 445         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 446
 447 retry:
 448         __asm__ __volatile__("wrpr %0, %1, %%pstate\n\t"
 449                              : : "r" (pstate), "i" (PSTATE_IE));
 450
 451         /* Setup the dispatch data registers. */
 452         __asm__ __volatile__("stxa      %0, [%3] %6\n\t"
 453                              "stxa      %1, [%4] %6\n\t"
 454                              "stxa      %2, [%5] %6\n\t"
 455                              "membar    #Sync\n\t"
 456                              : /* no outputs */
 457                              : "r" (data0), "r" (data1), "r" (data2),
 458                                "r" (0x40), "r" (0x50), "r" (0x60),
 459                                "i" (ASI_INTR_W));
 460
 461         nack_busy_id = 0;
 462         {
 463                 int i;
 464
 465                 for_each_cpu_mask(i, mask) {
 466                         u64 target = (i << 14) | 0x70;
 467
 468                         if (!is_jbus)
 469                                 target |= (nack_busy_id << 24);
 470                         __asm__ __volatile__(
 471                                 "stxa   %%g0, [%0] %1\n\t"
 472                                 "membar #Sync\n\t"
 473                                 : /* no outputs */
 474                                 : "r" (target), "i" (ASI_INTR_W));
 475                         nack_busy_id++;
 476                 }
 477         }
 478
 479         /* Now, poll for completion. */
 480         {
 481                 u64 dispatch_stat;
 482                 long stuck;
 483
 484                 stuck = 100000 * nack_busy_id;
 485                 do {
 486                         __asm__ __volatile__("ldxa      [%%g0] %1, %0"
 487                                              : "=r" (dispatch_stat)
 488                                              : "i" (ASI_INTR_DISPATCH_STAT));
 489                         if (dispatch_stat == 0UL) {
 490                                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 491                                                      : : "r" (pstate));
 492                                 return;
 493                         }
 494                         if (!--stuck)
 495                                 break;
 496                 } while (dispatch_stat & 0x5555555555555555UL);
 497
 498                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 499                                      : : "r" (pstate));
 500
 501                 if ((dispatch_stat & ~(0x5555555555555555UL)) == 0) {
 502                         /* Busy bits will not clear, continue instead
 503                          * of freezing up on this cpu.
 504                          */
 505                         printk("CPU[%d]: mondo stuckage result[%016lx]\n",
 506                                smp_processor_id(), dispatch_stat);
 507                 } else {
 508                         int i, this_busy_nack = 0;
 509
 510                         /* Delay some random time with interrupts enabled
 511                          * to prevent deadlock.
 512                          */
 513                         udelay(2 * nack_busy_id);
 514
 515                         /* Clear out the mask bits for cpus which did not
 516                          * NACK us.
 517                          */
 518                         for_each_cpu_mask(i, mask) {
 519                                 u64 check_mask;
 520
 521                                 if (is_jbus)
 522                                         check_mask = (0x2UL << (2*i));
 523                                 else
 524                                         check_mask = (0x2UL <<
 525                                                       this_busy_nack);
 526                                 if ((dispatch_stat & check_mask) == 0)
 527                                         cpu_clear(i, mask);
 528                                 this_busy_nack += 2;
 529                         }
 530
 531                         goto retry;
 532                 }
 533         }
 534 }
 535
 536 #if 0
 537 /* Multi-cpu list version.  */
 538 static int init_cpu_list(u16 *list, cpumask_t mask)
 539 {
 540         int i, cnt;
 541
 542         cnt = 0;
 543         for_each_cpu_mask(i, mask)
 544                 list[cnt++] = i;
 545
 546         return cnt;
 547 }
 548
 549 static int update_cpu_list(u16 *list, int orig_cnt, cpumask_t mask)
 550 {
 551         int i;
 552
 553         for (i = 0; i < orig_cnt; i++) {
 554                 if (list[i] == 0xffff)
 555                         cpu_clear(i, mask);
 556         }
 557
 558         return init_cpu_list(list, mask);
 559 }
 560
 561 static void hypervisor_xcall_deliver(u64 data0, u64 data1, u64 data2, cpumask_t mask)
 562 {
 563         int this_cpu = get_cpu();
 564         struct trap_per_cpu *tb = &trap_block[this_cpu];
 565         u64 *mondo = __va(tb->cpu_mondo_block_pa);
 566         u16 *cpu_list = __va(tb->cpu_list_pa);
 567         int cnt, retries;
 568
 569         mondo[0] = data0;
 570         mondo[1] = data1;
 571         mondo[2] = data2;
 572         wmb();
 573
 574         retries = 0;
 575         cnt = init_cpu_list(cpu_list, mask);
 576         do {
 577                 register unsigned long func __asm__("%o5");
 578                 register unsigned long arg0 __asm__("%o0");
 579                 register unsigned long arg1 __asm__("%o1");
 580                 register unsigned long arg2 __asm__("%o2");
 581
 582                 func = HV_FAST_CPU_MONDO_SEND;
 583                 arg0 = cnt;
 584                 arg1 = tb->cpu_list_pa;
 585                 arg2 = tb->cpu_mondo_block_pa;
 586
 587                 __asm__ __volatile__("ta        %8"
 588                                      : "=&r" (func), "=&r" (arg0),
 589                                        "=&r" (arg1), "=&r" (arg2)
 590                                      : "0" (func), "1" (arg0),
 591                                        "2" (arg1), "3" (arg2),
 592                                        "i" (HV_FAST_TRAP)
 593                                      : "memory");
 594                 if (likely(arg0 == HV_EOK))
 595                         break;
 596
 597                 if (unlikely(++retries > 100)) {
 598                         printk("CPU[%d]: sun4v mondo error %lu\n",
 599                                this_cpu, func);
 600                         break;
 601                 }
 602
 603                 cnt = update_cpu_list(cpu_list, cnt, mask);
 604
 605                 udelay(2 * cnt);
 606         } while (1);
 607
 608         put_cpu();
 609 }
 610 #else
 611 /* Single-cpu list version.  */
 612 static void hypervisor_xcall_deliver(u64 data0, u64 data1, u64 data2, cpumask_t mask)
 613 {
 614         int this_cpu = get_cpu();
 615         struct trap_per_cpu *tb = &trap_block[this_cpu];
 616         u64 *mondo = __va(tb->cpu_mondo_block_pa);
 617         u16 *cpu_list = __va(tb->cpu_list_pa);
 618         int i;
 619
 620         mondo[0] = data0;
 621         mondo[1] = data1;
 622         mondo[2] = data2;
 623         wmb();
 624
 625         for_each_cpu_mask(i, mask) {
 626                 int retries = 0;
 627
 628                 do {
 629                         register unsigned long func __asm__("%o5");
 630                         register unsigned long arg0 __asm__("%o0");
 631                         register unsigned long arg1 __asm__("%o1");
 632                         register unsigned long arg2 __asm__("%o2");
 633
 634                         cpu_list[0] = i;
 635                         func = HV_FAST_CPU_MONDO_SEND;
 636                         arg0 = 1;
 637                         arg1 = tb->cpu_list_pa;
 638                         arg2 = tb->cpu_mondo_block_pa;
 639
 640                         __asm__ __volatile__("ta        %8"
 641                                              : "=&r" (func), "=&r" (arg0),
 642                                                "=&r" (arg1), "=&r" (arg2)
 643                                              : "0" (func), "1" (arg0),
 644                                                "2" (arg1), "3" (arg2),
 645                                                "i" (HV_FAST_TRAP)
 646                                              : "memory");
 647                         if (likely(arg0 == HV_EOK))
 648                                 break;
 649
 650                         if (unlikely(++retries > 100)) {
 651                                 printk("CPU[%d]: sun4v mondo error %lu\n",
 652                                        this_cpu, func);
 653                                 break;
 654                         }
 655
 656                         udelay(2 * i);
 657                 } while (1);
 658         }
 659
 660         put_cpu();
 661 }
 662 #endif
 663
 664 /* Send cross call to all processors mentioned in MASK
 665  * except self.
 666  */
 667 static void smp_cross_call_masked(unsigned long *func, u32 ctx, u64 data1, u64 data2, cpumask_t mask)
 668 {
 669         u64 data0 = (((u64)ctx)<<32 | (((u64)func) & 0xffffffff));
 670         int this_cpu = get_cpu();
 671
 672         cpus_and(mask, mask, cpu_online_map);
 673         cpu_clear(this_cpu, mask);
 674
 675         if (tlb_type == spitfire)
 676                 spitfire_xcall_deliver(data0, data1, data2, mask);
 677         else if (tlb_type == cheetah || tlb_type == cheetah_plus)
 678                 cheetah_xcall_deliver(data0, data1, data2, mask);
 679         else
 680                 hypervisor_xcall_deliver(data0, data1, data2, mask);
 681         /* NOTE: Caller runs local copy on master. */
 682
 683         put_cpu();
 684 }
 685
 686 extern unsigned long xcall_sync_tick;
 687
 688 static void smp_start_sync_tick_client(int cpu)
 689 {
 690         cpumask_t mask = cpumask_of_cpu(cpu);
 691
 692         smp_cross_call_masked(&xcall_sync_tick,
 693                               0, 0, 0, mask);
 694 }
 695
 696 /* Send cross call to all processors except self. */
 697 #define smp_cross_call(func, ctx, data1, data2) \
 698         smp_cross_call_masked(func, ctx, data1, data2, cpu_online_map)
 699
 700 struct call_data_struct {
 701         void (*func) (void *info);
 702         void *info;
 703         atomic_t finished;
 704         int wait;
 705 };
 706
 707 static DEFINE_SPINLOCK(call_lock);
 708 static struct call_data_struct *call_data;
 709
 710 extern unsigned long xcall_call_function;
 711
 712 /*
 713  * You must not call this function with disabled interrupts or from a
 714  * hardware interrupt handler or from a bottom half handler.
 715  */
 716 static int smp_call_function_mask(void (*func)(void *info), void *info,
 717                                   int nonatomic, int wait, cpumask_t mask)
 718 {
 719         struct call_data_struct data;
 720         int cpus = cpus_weight(mask) - 1;
 721         long timeout;
 722
 723         if (!cpus)
 724                 return 0;
 725
 726         /* Can deadlock when called with interrupts disabled */
 727         WARN_ON(irqs_disabled());
 728
 729         data.func = func;
 730         data.info = info;
 731         atomic_set(&data.finished, 0);
 732         data.wait = wait;
 733
 734         spin_lock(&call_lock);
 735
 736         call_data = &data;
 737
 738         smp_cross_call_masked(&xcall_call_function, 0, 0, 0, mask);
 739
 740         /*
 741          * Wait for other cpus to complete function or at
 742          * least snap the call data.
 743          */
 744         timeout = 1000000;
 745         while (atomic_read(&data.finished) != cpus) {
 746                 if (--timeout <= 0)
 747                         goto out_timeout;
 748                 barrier();
 749                 udelay(1);
 750         }
 751
 752         spin_unlock(&call_lock);
 753
 754         return 0;
 755
 756 out_timeout:
 757         spin_unlock(&call_lock);
 758         printk("XCALL: Remote cpus not responding, ncpus=%ld finished=%ld\n",
 759                (long) num_online_cpus() - 1L,
 760                (long) atomic_read(&data.finished));
 761         return 0;
 762 }
 763
 764 int smp_call_function(void (*func)(void *info), void *info,
 765                       int nonatomic, int wait)
 766 {
 767         return smp_call_function_mask(func, info, nonatomic, wait,
 768                                       cpu_online_map);
 769 }
 770
 771 void smp_call_function_client(int irq, struct pt_regs *regs)
 772 {
 773         void (*func) (void *info) = call_data->func;
 774         void *info = call_data->info;
 775
 776         clear_softint(1 << irq);
 777         if (call_data->wait) {
 778                 /* let initiator proceed only after completion */
 779                 func(info);
 780                 atomic_inc(&call_data->finished);
 781         } else {
 782                 /* let initiator proceed after getting data */
 783                 atomic_inc(&call_data->finished);
 784                 func(info);
 785         }
 786 }
 787
 788 static void tsb_sync(void *info)
 789 {
 790         struct mm_struct *mm = info;
 791
 792         if (current->active_mm == mm)
 793                 tsb_context_switch(mm);
 794 }
 795
 796 void smp_tsb_sync(struct mm_struct *mm)
 797 {
 798         smp_call_function_mask(tsb_sync, mm, 0, 1, mm->cpu_vm_mask);
 799 }
 800
 801 extern unsigned long xcall_flush_tlb_mm;
 802 extern unsigned long xcall_flush_tlb_pending;
 803 extern unsigned long xcall_flush_tlb_kernel_range;
 804 extern unsigned long xcall_report_regs;
 805 extern unsigned long xcall_receive_signal;
 806
 807 #ifdef DCACHE_ALIASING_POSSIBLE
 808 extern unsigned long xcall_flush_dcache_page_cheetah;
 809 #endif
 810 extern unsigned long xcall_flush_dcache_page_spitfire;
 811
 812 #ifdef CONFIG_DEBUG_DCFLUSH
 813 extern atomic_t dcpage_flushes;
 814 extern atomic_t dcpage_flushes_xcall;
 815 #endif
 816
 817 static __inline__ void __local_flush_dcache_page(struct page *page)
 818 {
 819 #ifdef DCACHE_ALIASING_POSSIBLE
 820         __flush_dcache_page(page_address(page),
 821                             ((tlb_type == spitfire) &&
 822                              page_mapping(page) != NULL));
 823 #else
 824         if (page_mapping(page) != NULL &&
 825             tlb_type == spitfire)
 826                 __flush_icache_page(__pa(page_address(page)));
 827 #endif
 828 }
 829
 830 void smp_flush_dcache_page_impl(struct page *page, int cpu)
 831 {
 832         cpumask_t mask = cpumask_of_cpu(cpu);
 833         int this_cpu;
 834
 835         if (tlb_type == hypervisor)
 836                 return;
 837
 838 #ifdef CONFIG_DEBUG_DCFLUSH
 839         atomic_inc(&dcpage_flushes);
 840 #endif
 841
 842         this_cpu = get_cpu();
 843
 844         if (cpu == this_cpu) {
 845                 __local_flush_dcache_page(page);
 846         } else if (cpu_online(cpu)) {
 847                 void *pg_addr = page_address(page);
 848                 u64 data0;
 849
 850                 if (tlb_type == spitfire) {
 851                         data0 =
 852                                 ((u64)&xcall_flush_dcache_page_spitfire);
 853                         if (page_mapping(page) != NULL)
 854                                 data0 |= ((u64)1 << 32);
 855                         spitfire_xcall_deliver(data0,
 856                                                __pa(pg_addr),
 857                                                (u64) pg_addr,
 858                                                mask);
 859                 } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 860 #ifdef DCACHE_ALIASING_POSSIBLE
 861                         data0 =
 862                                 ((u64)&xcall_flush_dcache_page_cheetah);
 863                         cheetah_xcall_deliver(data0,
 864                                               __pa(pg_addr),
 865                                               0, mask);
 866 #endif
 867                 }
 868 #ifdef CONFIG_DEBUG_DCFLUSH
 869                 atomic_inc(&dcpage_flushes_xcall);
 870 #endif
 871         }
 872
 873         put_cpu();
 874 }
 875
 876 void flush_dcache_page_all(struct mm_struct *mm, struct page *page)
 877 {
 878         void *pg_addr = page_address(page);
 879         cpumask_t mask = cpu_online_map;
 880         u64 data0;
 881         int this_cpu;
 882
 883         if (tlb_type == hypervisor)
 884                 return;
 885
 886         this_cpu = get_cpu();
 887
 888         cpu_clear(this_cpu, mask);
 889
 890 #ifdef CONFIG_DEBUG_DCFLUSH
 891         atomic_inc(&dcpage_flushes);
 892 #endif
 893         if (cpus_empty(mask))
 894                 goto flush_self;
 895         if (tlb_type == spitfire) {
 896                 data0 = ((u64)&xcall_flush_dcache_page_spitfire);
 897                 if (page_mapping(page) != NULL)
 898                         data0 |= ((u64)1 << 32);
 899                 spitfire_xcall_deliver(data0,
 900                                        __pa(pg_addr),
 901                                        (u64) pg_addr,
 902                                        mask);
 903         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 904 #ifdef DCACHE_ALIASING_POSSIBLE
 905                 data0 = ((u64)&xcall_flush_dcache_page_cheetah);
 906                 cheetah_xcall_deliver(data0,
 907                                       __pa(pg_addr),
 908                                       0, mask);
 909 #endif
 910         }
 911 #ifdef CONFIG_DEBUG_DCFLUSH
 912         atomic_inc(&dcpage_flushes_xcall);
 913 #endif
 914  flush_self:
 915         __local_flush_dcache_page(page);
 916
 917         put_cpu();
 918 }
 919
 920 void smp_receive_signal(int cpu)
 921 {
 922         cpumask_t mask = cpumask_of_cpu(cpu);
 923
 924         if (cpu_online(cpu)) {
 925                 u64 data0 = (((u64)&xcall_receive_signal) & 0xffffffff);
 926
 927                 if (tlb_type == spitfire)
 928                         spitfire_xcall_deliver(data0, 0, 0, mask);
 929                 else if (tlb_type == cheetah || tlb_type == cheetah_plus)
 930                         cheetah_xcall_deliver(data0, 0, 0, mask);
 931                 else if (tlb_type == hypervisor)
 932                         hypervisor_xcall_deliver(data0, 0, 0, mask);
 933         }
 934 }
 935
 936 void smp_receive_signal_client(int irq, struct pt_regs *regs)
 937 {
 938         /* Just return, rtrap takes care of the rest. */
 939         clear_softint(1 << irq);
 940 }
 941
 942 void smp_report_regs(void)
 943 {
 944         smp_cross_call(&xcall_report_regs, 0, 0, 0);
 945 }
 946
 947 /* We know that the window frames of the user have been flushed
 948  * to the stack before we get here because all callers of us
 949  * are flush_tlb_*() routines, and these run after flush_cache_*()
 950  * which performs the flushw.
 951  *
 952  * The SMP TLB coherency scheme we use works as follows:
 953  *
 954  * 1) mm->cpu_vm_mask is a bit mask of which cpus an address
 955  *    space has (potentially) executed on, this is the heuristic
 956  *    we use to avoid doing cross calls.
 957  *
 958  *    Also, for flushing from kswapd and also for clones, we
 959  *    use cpu_vm_mask as the list of cpus to make run the TLB.
 960  *
 961  * 2) TLB context numbers are shared globally across all processors
 962  *    in the system, this allows us to play several games to avoid
 963  *    cross calls.
 964  *
 965  *    One invariant is that when a cpu switches to a process, and
 966  *    that processes tsk->active_mm->cpu_vm_mask does not have the
 967  *    current cpu's bit set, that tlb context is flushed locally.
 968  *
 969  *    If the address space is non-shared (ie. mm->count == 1) we avoid
 970  *    cross calls when we want to flush the currently running process's
 971  *    tlb state.  This is done by clearing all cpu bits except the current
 972  *    processor's in current->active_mm->cpu_vm_mask and performing the
 973  *    flush locally only.  This will force any subsequent cpus which run
 974  *    this task to flush the context from the local tlb if the process
 975  *    migrates to another cpu (again).
 976  *
 977  * 3) For shared address spaces (threads) and swapping we bite the
 978  *    bullet for most cases and perform the cross call (but only to
 979  *    the cpus listed in cpu_vm_mask).
 980  *
 981  *    The performance gain from "optimizing" away the cross call for threads is
 982  *    questionable (in theory the big win for threads is the massive sharing of
 983  *    address space state across processors).
 984  */
 985
 986 /* This currently is only used by the hugetlb arch pre-fault
 987  * hook on UltraSPARC-III+ and later when changing the pagesize
 988  * bits of the context register for an address space.
 989  */
 990 void smp_flush_tlb_mm(struct mm_struct *mm)
 991 {
 992         u32 ctx = CTX_HWBITS(mm->context);
 993         int cpu = get_cpu();
 994
 995         if (atomic_read(&mm->mm_users) == 1) {
 996                 mm->cpu_vm_mask = cpumask_of_cpu(cpu);
 997                 goto local_flush_and_out;
 998         }
 999
1000         smp_cross_call_masked(&xcall_flush_tlb_mm,
1001                               ctx, 0, 0,
1002                               mm->cpu_vm_mask);
1003
1004 local_flush_and_out:
1005         __flush_tlb_mm(ctx, SECONDARY_CONTEXT);
1006
1007         put_cpu();
1008 }
1009
1010 void smp_flush_tlb_pending(struct mm_struct *mm, unsigned long nr, unsigned long *vaddrs)
1011 {
1012         u32 ctx = CTX_HWBITS(mm->context);
1013         int cpu = get_cpu();
1014
1015         if (mm == current->active_mm && atomic_read(&mm->mm_users) == 1)
1016                 mm->cpu_vm_mask = cpumask_of_cpu(cpu);
1017         else
1018                 smp_cross_call_masked(&xcall_flush_tlb_pending,
1019                                       ctx, nr, (unsigned long) vaddrs,
1020                                       mm->cpu_vm_mask);
1021
1022         __flush_tlb_pending(ctx, nr, vaddrs);
1023
1024         put_cpu();
1025 }
1026
1027 void smp_flush_tlb_kernel_range(unsigned long start, unsigned long end)
1028 {
1029         start &= PAGE_MASK;
1030         end    = PAGE_ALIGN(end);
1031         if (start != end) {
1032                 smp_cross_call(&xcall_flush_tlb_kernel_range,
1033                                0, start, end);
1034
1035                 __flush_tlb_kernel_range(start, end);
1036         }
1037 }
1038
1039 /* CPU capture. */
1040 /* #define CAPTURE_DEBUG */
1041 extern unsigned long xcall_capture;
1042
1043 static atomic_t smp_capture_depth = ATOMIC_INIT(0);
1044 static atomic_t smp_capture_registry = ATOMIC_INIT(0);
1045 static unsigned long penguins_are_doing_time;
1046
1047 void smp_capture(void)
1048 {
1049         int result = atomic_add_ret(1, &smp_capture_depth);
1050
1051         if (result == 1) {
1052                 int ncpus = num_online_cpus();
1053
1054 #ifdef CAPTURE_DEBUG
1055                 printk("CPU[%d]: Sending penguins to jail...",
1056                        smp_processor_id());
1057 #endif
1058                 penguins_are_doing_time = 1;
1059                 membar_storestore_loadstore();
1060                 atomic_inc(&smp_capture_registry);
1061                 smp_cross_call(&xcall_capture, 0, 0, 0);
1062                 while (atomic_read(&smp_capture_registry) != ncpus)
1063                         rmb();
1064 #ifdef CAPTURE_DEBUG
1065                 printk("done\n");
1066 #endif
1067         }
1068 }
1069
1070 void smp_release(void)
1071 {
1072         if (atomic_dec_and_test(&smp_capture_depth)) {
1073 #ifdef CAPTURE_DEBUG
1074                 printk("CPU[%d]: Giving pardon to "
1075                        "imprisoned penguins\n",
1076                        smp_processor_id());
1077 #endif
1078                 penguins_are_doing_time = 0;
1079                 membar_storeload_storestore();
1080                 atomic_dec(&smp_capture_registry);
1081         }
1082 }
1083
1084 /* Imprisoned penguins run with %pil == 15, but PSTATE_IE set, so they
1085  * can service tlb flush xcalls...
1086  */
1087 extern void prom_world(int);
1088
1089 void smp_penguin_jailcell(int irq, struct pt_regs *regs)
1090 {
1091         clear_softint(1 << irq);
1092
1093         preempt_disable();
1094
1095         __asm__ __volatile__("flushw");
1096         prom_world(1);
1097         atomic_inc(&smp_capture_registry);
1098         membar_storeload_storestore();
1099         while (penguins_are_doing_time)
1100                 rmb();
1101         atomic_dec(&smp_capture_registry);
1102         prom_world(0);
1103
1104         preempt_enable();
1105 }
1106
1107 #define prof_multiplier(__cpu)          cpu_data(__cpu).multiplier
1108 #define prof_counter(__cpu)             cpu_data(__cpu).counter
1109
1110 void smp_percpu_timer_interrupt(struct pt_regs *regs)
1111 {
1112         unsigned long compare, tick, pstate;
1113         int cpu = smp_processor_id();
1114         int user = user_mode(regs);
1115
1116         /*
1117          * Check for level 14 softint.
1118          */
1119         {
1120                 unsigned long tick_mask = tick_ops->softint_mask;
1121
1122                 if (!(get_softint() & tick_mask)) {
1123                         extern void handler_irq(int, struct pt_regs *);
1124
1125                         handler_irq(14, regs);
1126                         return;
1127                 }
1128                 clear_softint(tick_mask);
1129         }
1130
1131         do {
1132                 profile_tick(CPU_PROFILING, regs);
1133                 if (!--prof_counter(cpu)) {
1134                         irq_enter();
1135
1136                         if (cpu == boot_cpu_id) {
1137                                 kstat_this_cpu.irqs[0]++;
1138                                 timer_tick_interrupt(regs);
1139                         }
1140
1141                         update_process_times(user);
1142
1143                         irq_exit();
1144
1145                         prof_counter(cpu) = prof_multiplier(cpu);
1146                 }
1147
1148                 /* Guarantee that the following sequences execute
1149                  * uninterrupted.
1150                  */
1151                 __asm__ __volatile__("rdpr      %%pstate, %0\n\t"
1152                                      "wrpr      %0, %1, %%pstate"
1153                                      : "=r" (pstate)
1154                                      : "i" (PSTATE_IE));
1155
1156                 compare = tick_ops->add_compare(current_tick_offset);
1157                 tick = tick_ops->get_tick();
1158
1159                 /* Restore PSTATE_IE. */
1160                 __asm__ __volatile__("wrpr      %0, 0x0, %%pstate"
1161                                      : /* no outputs */
1162                                      : "r" (pstate));
1163         } while (time_after_eq(tick, compare));
1164 }
1165
1166 static void __init smp_setup_percpu_timer(void)
1167 {
1168         int cpu = smp_processor_id();
1169         unsigned long pstate;
1170
1171         prof_counter(cpu) = prof_multiplier(cpu) = 1;
1172
1173         /* Guarantee that the following sequences execute
1174          * uninterrupted.
1175          */
1176         __asm__ __volatile__("rdpr      %%pstate, %0\n\t"
1177                              "wrpr      %0, %1, %%pstate"
1178                              : "=r" (pstate)
1179                              : "i" (PSTATE_IE));
1180
1181         tick_ops->init_tick(current_tick_offset);
1182
1183         /* Restore PSTATE_IE. */
1184         __asm__ __volatile__("wrpr      %0, 0x0, %%pstate"
1185                              : /* no outputs */
1186                              : "r" (pstate));
1187 }
1188
1189 void __init smp_tick_init(void)
1190 {
1191         boot_cpu_id = hard_smp_processor_id();
1192         current_tick_offset = timer_tick_offset;
1193
1194         cpu_set(boot_cpu_id, cpu_online_map);
1195         prof_counter(boot_cpu_id) = prof_multiplier(boot_cpu_id) = 1;
1196 }
1197
1198 /* /proc/profile writes can call this, don't __init it please. */
1199 static DEFINE_SPINLOCK(prof_setup_lock);
1200
1201 int setup_profiling_timer(unsigned int multiplier)
1202 {
1203         unsigned long flags;
1204         int i;
1205
1206         if ((!multiplier) || (timer_tick_offset / multiplier) < 1000)
1207                 return -EINVAL;
1208
1209         spin_lock_irqsave(&prof_setup_lock, flags);
1210         for (i = 0; i < NR_CPUS; i++)
1211                 prof_multiplier(i) = multiplier;
1212         current_tick_offset = (timer_tick_offset / multiplier);
1213         spin_unlock_irqrestore(&prof_setup_lock, flags);
1214
1215         return 0;
1216 }
1217
1218 /* Constrain the number of cpus to max_cpus.  */
1219 void __init smp_prepare_cpus(unsigned int max_cpus)
1220 {
1221         if (num_possible_cpus() > max_cpus) {
1222                 int instance, mid;
1223
1224                 instance = 0;
1225                 while (!cpu_find_by_instance(instance, NULL, &mid)) {
1226                         if (mid != boot_cpu_id) {
1227                                 cpu_clear(mid, phys_cpu_present_map);
1228                                 if (num_possible_cpus() <= max_cpus)
1229                                         break;
1230                         }
1231                         instance++;
1232                 }
1233         }
1234
1235         smp_store_cpu_info(boot_cpu_id);
1236 }
1237
1238 /* Set this up early so that things like the scheduler can init
1239  * properly.  We use the same cpu mask for both the present and
1240  * possible cpu map.
1241  */
1242 void __init smp_setup_cpu_possible_map(void)
1243 {
1244         int instance, mid;
1245
1246         instance = 0;
1247         while (!cpu_find_by_instance(instance, NULL, &mid)) {
1248                 if (mid < NR_CPUS)
1249                         cpu_set(mid, phys_cpu_present_map);
1250                 instance++;
1251         }
1252 }
1253
1254 void __devinit smp_prepare_boot_cpu(void)
1255 {
1256         int cpu = hard_smp_processor_id();
1257
1258         if (cpu >= NR_CPUS) {
1259                 prom_printf("Serious problem, boot cpu id >= NR_CPUS\n");
1260                 prom_halt();
1261         }
1262
1263         current_thread_info()->cpu = cpu;
1264         __local_per_cpu_offset = __per_cpu_offset(cpu);
1265
1266         cpu_set(smp_processor_id(), cpu_online_map);
1267         cpu_set(smp_processor_id(), phys_cpu_present_map);
1268 }
1269
1270 int __devinit __cpu_up(unsigned int cpu)
1271 {
1272         int ret = smp_boot_one_cpu(cpu);
1273
1274         if (!ret) {
1275                 cpu_set(cpu, smp_commenced_mask);
1276                 while (!cpu_isset(cpu, cpu_online_map))
1277                         mb();
1278                 if (!cpu_isset(cpu, cpu_online_map)) {
1279                         ret = -ENODEV;
1280                 } else {
1281                         /* On SUN4V, writes to %tick and %stick are
1282                          * not allowed.
1283                          */
1284                         if (tlb_type != hypervisor)
1285                                 smp_synchronize_one_tick(cpu);
1286                 }
1287         }
1288         return ret;
1289 }
1290
1291 void __init smp_cpus_done(unsigned int max_cpus)
1292 {
1293         unsigned long bogosum = 0;
1294         int i;
1295
1296         for (i = 0; i < NR_CPUS; i++) {
1297                 if (cpu_online(i))
1298                         bogosum += cpu_data(i).udelay_val;
1299         }
1300         printk("Total of %ld processors activated "
1301                "(%lu.%02lu BogoMIPS).\n",
1302                (long) num_online_cpus(),
1303                bogosum/(500000/HZ),
1304                (bogosum/(5000/HZ))%100);
1305 }
1306
1307 void smp_send_reschedule(int cpu)
1308 {
1309         smp_receive_signal(cpu);
1310 }
1311
1312 /* This is a nop because we capture all other cpus
1313  * anyways when making the PROM active.
1314  */
1315 void smp_send_stop(void)
1316 {
1317 }
1318
1319 unsigned long __per_cpu_base __read_mostly;
1320 unsigned long __per_cpu_shift __read_mostly;
1321
1322 EXPORT_SYMBOL(__per_cpu_base);
1323 EXPORT_SYMBOL(__per_cpu_shift);
1324
1325 void __init setup_per_cpu_areas(void)
1326 {
1327         unsigned long goal, size, i;
1328         char *ptr;
1329
1330         /* Copy section for each CPU (we discard the original) */
1331         goal = ALIGN(__per_cpu_end - __per_cpu_start, SMP_CACHE_BYTES);
1332 #ifdef CONFIG_MODULES
1333         if (goal < PERCPU_ENOUGH_ROOM)
1334                 goal = PERCPU_ENOUGH_ROOM;
1335 #endif
1336         __per_cpu_shift = 0;
1337         for (size = 1UL; size < goal; size <<= 1UL)
1338                 __per_cpu_shift++;
1339
1340         ptr = alloc_bootmem(size * NR_CPUS);
1341
1342         __per_cpu_base = ptr - __per_cpu_start;
1343
1344         for (i = 0; i < NR_CPUS; i++, ptr += size)
1345                 memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start);
1346 }