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
  42 extern int linux_num_cpus;
  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 = CPU_MASK_NONE;
  49 cpumask_t phys_cpu_present_map = 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).pgcache_size               = 0;
  92         cpu_data(id).pte_cache[0]               = NULL;
  93         cpu_data(id).pte_cache[1]               = NULL;
  94         cpu_data(id).pgd_cache                  = NULL;
  95         cpu_data(id).idle_volume                = 1;
  96 }
  97
  98 static void smp_setup_percpu_timer(void);
  99
 100 static volatile unsigned long callin_flag = 0;
 101
 102 extern void inherit_locked_prom_mappings(int save_p);
 103
 104 static inline void cpu_setup_percpu_base(unsigned long cpu_id)
 105 {
 106         __asm__ __volatile__("mov       %0, %%g5\n\t"
 107                              "stxa      %0, [%1] %2\n\t"
 108                              "membar    #Sync"
 109                              : /* no outputs */
 110                              : "r" (__per_cpu_offset(cpu_id)),
 111                                "r" (TSB_REG), "i" (ASI_IMMU));
 112 }
 113
 114 void __init smp_callin(void)
 115 {
 116         int cpuid = hard_smp_processor_id();
 117
 118         inherit_locked_prom_mappings(0);
 119
 120         __flush_tlb_all();
 121
 122         cpu_setup_percpu_base(cpuid);
 123
 124         smp_setup_percpu_timer();
 125
 126         if (cheetah_pcache_forced_on)
 127                 cheetah_enable_pcache();
 128
 129         local_irq_enable();
 130
 131         calibrate_delay();
 132         smp_store_cpu_info(cpuid);
 133         callin_flag = 1;
 134         __asm__ __volatile__("membar #Sync\n\t"
 135                              "flush  %%g6" : : : "memory");
 136
 137         /* Clear this or we will die instantly when we
 138          * schedule back to this idler...
 139          */
 140         clear_thread_flag(TIF_NEWCHILD);
 141
 142         /* Attach to the address space of init_task. */
 143         atomic_inc(&init_mm.mm_count);
 144         current->active_mm = &init_mm;
 145
 146         while (!cpu_isset(cpuid, smp_commenced_mask))
 147                 membar("#LoadLoad");
 148
 149         cpu_set(cpuid, cpu_online_map);
 150 }
 151
 152 void cpu_panic(void)
 153 {
 154         printk("CPU[%d]: Returns from cpu_idle!\n", smp_processor_id());
 155         panic("SMP bolixed\n");
 156 }
 157
 158 static unsigned long current_tick_offset;
 159
 160 /* This tick register synchronization scheme is taken entirely from
 161  * the ia64 port, see arch/ia64/kernel/smpboot.c for details and credit.
 162  *
 163  * The only change I've made is to rework it so that the master
 164  * initiates the synchonization instead of the slave. -DaveM
 165  */
 166
 167 #define MASTER  0
 168 #define SLAVE   (SMP_CACHE_BYTES/sizeof(unsigned long))
 169
 170 #define NUM_ROUNDS      64      /* magic value */
 171 #define NUM_ITERS       5       /* likewise */
 172
 173 static DEFINE_SPINLOCK(itc_sync_lock);
 174 static unsigned long go[SLAVE + 1];
 175
 176 #define DEBUG_TICK_SYNC 0
 177
 178 static inline long get_delta (long *rt, long *master)
 179 {
 180         unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
 181         unsigned long tcenter, t0, t1, tm;
 182         unsigned long i;
 183
 184         for (i = 0; i < NUM_ITERS; i++) {
 185                 t0 = tick_ops->get_tick();
 186                 go[MASTER] = 1;
 187                 membar("#StoreLoad");
 188                 while (!(tm = go[SLAVE]))
 189                         membar("#LoadLoad");
 190                 go[SLAVE] = 0;
 191                 membar("#StoreStore");
 192                 t1 = tick_ops->get_tick();
 193
 194                 if (t1 - t0 < best_t1 - best_t0)
 195                         best_t0 = t0, best_t1 = t1, best_tm = tm;
 196         }
 197
 198         *rt = best_t1 - best_t0;
 199         *master = best_tm - best_t0;
 200
 201         /* average best_t0 and best_t1 without overflow: */
 202         tcenter = (best_t0/2 + best_t1/2);
 203         if (best_t0 % 2 + best_t1 % 2 == 2)
 204                 tcenter++;
 205         return tcenter - best_tm;
 206 }
 207
 208 void smp_synchronize_tick_client(void)
 209 {
 210         long i, delta, adj, adjust_latency = 0, done = 0;
 211         unsigned long flags, rt, master_time_stamp, bound;
 212 #if DEBUG_TICK_SYNC
 213         struct {
 214                 long rt;        /* roundtrip time */
 215                 long master;    /* master's timestamp */
 216                 long diff;      /* difference between midpoint and master's timestamp */
 217                 long lat;       /* estimate of itc adjustment latency */
 218         } t[NUM_ROUNDS];
 219 #endif
 220
 221         go[MASTER] = 1;
 222
 223         while (go[MASTER])
 224                 membar("#LoadLoad");
 225
 226         local_irq_save(flags);
 227         {
 228                 for (i = 0; i < NUM_ROUNDS; i++) {
 229                         delta = get_delta(&rt, &master_time_stamp);
 230                         if (delta == 0) {
 231                                 done = 1;       /* let's lock on to this... */
 232                                 bound = rt;
 233                         }
 234
 235                         if (!done) {
 236                                 if (i > 0) {
 237                                         adjust_latency += -delta;
 238                                         adj = -delta + adjust_latency/4;
 239                                 } else
 240                                         adj = -delta;
 241
 242                                 tick_ops->add_tick(adj, current_tick_offset);
 243                         }
 244 #if DEBUG_TICK_SYNC
 245                         t[i].rt = rt;
 246                         t[i].master = master_time_stamp;
 247                         t[i].diff = delta;
 248                         t[i].lat = adjust_latency/4;
 249 #endif
 250                 }
 251         }
 252         local_irq_restore(flags);
 253
 254 #if DEBUG_TICK_SYNC
 255         for (i = 0; i < NUM_ROUNDS; i++)
 256                 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
 257                        t[i].rt, t[i].master, t[i].diff, t[i].lat);
 258 #endif
 259
 260         printk(KERN_INFO "CPU %d: synchronized TICK with master CPU (last diff %ld cycles,"
 261                "maxerr %lu cycles)\n", smp_processor_id(), delta, rt);
 262 }
 263
 264 static void smp_start_sync_tick_client(int cpu);
 265
 266 static void smp_synchronize_one_tick(int cpu)
 267 {
 268         unsigned long flags, i;
 269
 270         go[MASTER] = 0;
 271
 272         smp_start_sync_tick_client(cpu);
 273
 274         /* wait for client to be ready */
 275         while (!go[MASTER])
 276                 membar("#LoadLoad");
 277
 278         /* now let the client proceed into his loop */
 279         go[MASTER] = 0;
 280         membar("#StoreLoad");
 281
 282         spin_lock_irqsave(&itc_sync_lock, flags);
 283         {
 284                 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; i++) {
 285                         while (!go[MASTER])
 286                                 membar("#LoadLoad");
 287                         go[MASTER] = 0;
 288                         membar("#StoreStore");
 289                         go[SLAVE] = tick_ops->get_tick();
 290                         membar("#StoreLoad");
 291                 }
 292         }
 293         spin_unlock_irqrestore(&itc_sync_lock, flags);
 294 }
 295
 296 extern unsigned long sparc64_cpu_startup;
 297
 298 /* The OBP cpu startup callback truncates the 3rd arg cookie to
 299  * 32-bits (I think) so to be safe we have it read the pointer
 300  * contained here so we work on >4GB machines. -DaveM
 301  */
 302 static struct thread_info *cpu_new_thread = NULL;
 303
 304 static int __devinit smp_boot_one_cpu(unsigned int cpu)
 305 {
 306         unsigned long entry =
 307                 (unsigned long)(&sparc64_cpu_startup);
 308         unsigned long cookie =
 309                 (unsigned long)(&cpu_new_thread);
 310         struct task_struct *p;
 311         int timeout, ret, cpu_node;
 312
 313         p = fork_idle(cpu);
 314         callin_flag = 0;
 315         cpu_new_thread = p->thread_info;
 316         cpu_set(cpu, cpu_callout_map);
 317
 318         cpu_find_by_mid(cpu, &cpu_node);
 319         prom_startcpu(cpu_node, entry, cookie);
 320
 321         for (timeout = 0; timeout < 5000000; timeout++) {
 322                 if (callin_flag)
 323                         break;
 324                 udelay(100);
 325         }
 326         if (callin_flag) {
 327                 ret = 0;
 328         } else {
 329                 printk("Processor %d is stuck.\n", cpu);
 330                 cpu_clear(cpu, cpu_callout_map);
 331                 ret = -ENODEV;
 332         }
 333         cpu_new_thread = NULL;
 334
 335         return ret;
 336 }
 337
 338 static void spitfire_xcall_helper(u64 data0, u64 data1, u64 data2, u64 pstate, unsigned long cpu)
 339 {
 340         u64 result, target;
 341         int stuck, tmp;
 342
 343         if (this_is_starfire) {
 344                 /* map to real upaid */
 345                 cpu = (((cpu & 0x3c) << 1) |
 346                         ((cpu & 0x40) >> 4) |
 347                         (cpu & 0x3));
 348         }
 349
 350         target = (cpu << 14) | 0x70;
 351 again:
 352         /* Ok, this is the real Spitfire Errata #54.
 353          * One must read back from a UDB internal register
 354          * after writes to the UDB interrupt dispatch, but
 355          * before the membar Sync for that write.
 356          * So we use the high UDB control register (ASI 0x7f,
 357          * ADDR 0x20) for the dummy read. -DaveM
 358          */
 359         tmp = 0x40;
 360         __asm__ __volatile__(
 361         "wrpr   %1, %2, %%pstate\n\t"
 362         "stxa   %4, [%0] %3\n\t"
 363         "stxa   %5, [%0+%8] %3\n\t"
 364         "add    %0, %8, %0\n\t"
 365         "stxa   %6, [%0+%8] %3\n\t"
 366         "membar #Sync\n\t"
 367         "stxa   %%g0, [%7] %3\n\t"
 368         "membar #Sync\n\t"
 369         "mov    0x20, %%g1\n\t"
 370         "ldxa   [%%g1] 0x7f, %%g0\n\t"
 371         "membar #Sync"
 372         : "=r" (tmp)
 373         : "r" (pstate), "i" (PSTATE_IE), "i" (ASI_INTR_W),
 374           "r" (data0), "r" (data1), "r" (data2), "r" (target),
 375           "r" (0x10), "0" (tmp)
 376         : "g1");
 377
 378         /* NOTE: PSTATE_IE is still clear. */
 379         stuck = 100000;
 380         do {
 381                 __asm__ __volatile__("ldxa [%%g0] %1, %0"
 382                         : "=r" (result)
 383                         : "i" (ASI_INTR_DISPATCH_STAT));
 384                 if (result == 0) {
 385                         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 386                                              : : "r" (pstate));
 387                         return;
 388                 }
 389                 stuck -= 1;
 390                 if (stuck == 0)
 391                         break;
 392         } while (result & 0x1);
 393         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 394                              : : "r" (pstate));
 395         if (stuck == 0) {
 396                 printk("CPU[%d]: mondo stuckage result[%016lx]\n",
 397                        smp_processor_id(), result);
 398         } else {
 399                 udelay(2);
 400                 goto again;
 401         }
 402 }
 403
 404 static __inline__ void spitfire_xcall_deliver(u64 data0, u64 data1, u64 data2, cpumask_t mask)
 405 {
 406         u64 pstate;
 407         int i;
 408
 409         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 410         for_each_cpu_mask(i, mask)
 411                 spitfire_xcall_helper(data0, data1, data2, pstate, i);
 412 }
 413
 414 /* Cheetah now allows to send the whole 64-bytes of data in the interrupt
 415  * packet, but we have no use for that.  However we do take advantage of
 416  * the new pipelining feature (ie. dispatch to multiple cpus simultaneously).
 417  */
 418 static void cheetah_xcall_deliver(u64 data0, u64 data1, u64 data2, cpumask_t mask)
 419 {
 420         u64 pstate, ver;
 421         int nack_busy_id, is_jalapeno;
 422
 423         if (cpus_empty(mask))
 424                 return;
 425
 426         /* Unfortunately, someone at Sun had the brilliant idea to make the
 427          * busy/nack fields hard-coded by ITID number for this Ultra-III
 428          * derivative processor.
 429          */
 430         __asm__ ("rdpr %%ver, %0" : "=r" (ver));
 431         is_jalapeno = ((ver >> 32) == 0x003e0016);
 432
 433         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 434
 435 retry:
 436         __asm__ __volatile__("wrpr %0, %1, %%pstate\n\t"
 437                              : : "r" (pstate), "i" (PSTATE_IE));
 438
 439         /* Setup the dispatch data registers. */
 440         __asm__ __volatile__("stxa      %0, [%3] %6\n\t"
 441                              "stxa      %1, [%4] %6\n\t"
 442                              "stxa      %2, [%5] %6\n\t"
 443                              "membar    #Sync\n\t"
 444                              : /* no outputs */
 445                              : "r" (data0), "r" (data1), "r" (data2),
 446                                "r" (0x40), "r" (0x50), "r" (0x60),
 447                                "i" (ASI_INTR_W));
 448
 449         nack_busy_id = 0;
 450         {
 451                 int i;
 452
 453                 for_each_cpu_mask(i, mask) {
 454                         u64 target = (i << 14) | 0x70;
 455
 456                         if (!is_jalapeno)
 457                                 target |= (nack_busy_id << 24);
 458                         __asm__ __volatile__(
 459                                 "stxa   %%g0, [%0] %1\n\t"
 460                                 "membar #Sync\n\t"
 461                                 : /* no outputs */
 462                                 : "r" (target), "i" (ASI_INTR_W));
 463                         nack_busy_id++;
 464                 }
 465         }
 466
 467         /* Now, poll for completion. */
 468         {
 469                 u64 dispatch_stat;
 470                 long stuck;
 471
 472                 stuck = 100000 * nack_busy_id;
 473                 do {
 474                         __asm__ __volatile__("ldxa      [%%g0] %1, %0"
 475                                              : "=r" (dispatch_stat)
 476                                              : "i" (ASI_INTR_DISPATCH_STAT));
 477                         if (dispatch_stat == 0UL) {
 478                                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 479                                                      : : "r" (pstate));
 480                                 return;
 481                         }
 482                         if (!--stuck)
 483                                 break;
 484                 } while (dispatch_stat & 0x5555555555555555UL);
 485
 486                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 487                                      : : "r" (pstate));
 488
 489                 if ((dispatch_stat & ~(0x5555555555555555UL)) == 0) {
 490                         /* Busy bits will not clear, continue instead
 491                          * of freezing up on this cpu.
 492                          */
 493                         printk("CPU[%d]: mondo stuckage result[%016lx]\n",
 494                                smp_processor_id(), dispatch_stat);
 495                 } else {
 496                         int i, this_busy_nack = 0;
 497
 498                         /* Delay some random time with interrupts enabled
 499                          * to prevent deadlock.
 500                          */
 501                         udelay(2 * nack_busy_id);
 502
 503                         /* Clear out the mask bits for cpus which did not
 504                          * NACK us.
 505                          */
 506                         for_each_cpu_mask(i, mask) {
 507                                 u64 check_mask;
 508
 509                                 if (is_jalapeno)
 510                                         check_mask = (0x2UL << (2*i));
 511                                 else
 512                                         check_mask = (0x2UL <<
 513                                                       this_busy_nack);
 514                                 if ((dispatch_stat & check_mask) == 0)
 515                                         cpu_clear(i, mask);
 516                                 this_busy_nack += 2;
 517                         }
 518
 519                         goto retry;
 520                 }
 521         }
 522 }
 523
 524 /* Send cross call to all processors mentioned in MASK
 525  * except self.
 526  */
 527 static void smp_cross_call_masked(unsigned long *func, u32 ctx, u64 data1, u64 data2, cpumask_t mask)
 528 {
 529         u64 data0 = (((u64)ctx)<<32 | (((u64)func) & 0xffffffff));
 530         int this_cpu = get_cpu();
 531
 532         cpus_and(mask, mask, cpu_online_map);
 533         cpu_clear(this_cpu, mask);
 534
 535         if (tlb_type == spitfire)
 536                 spitfire_xcall_deliver(data0, data1, data2, mask);
 537         else
 538                 cheetah_xcall_deliver(data0, data1, data2, mask);
 539         /* NOTE: Caller runs local copy on master. */
 540
 541         put_cpu();
 542 }
 543
 544 extern unsigned long xcall_sync_tick;
 545
 546 static void smp_start_sync_tick_client(int cpu)
 547 {
 548         cpumask_t mask = cpumask_of_cpu(cpu);
 549
 550         smp_cross_call_masked(&xcall_sync_tick,
 551                               0, 0, 0, mask);
 552 }
 553
 554 /* Send cross call to all processors except self. */
 555 #define smp_cross_call(func, ctx, data1, data2) \
 556         smp_cross_call_masked(func, ctx, data1, data2, cpu_online_map)
 557
 558 struct call_data_struct {
 559         void (*func) (void *info);
 560         void *info;
 561         atomic_t finished;
 562         int wait;
 563 };
 564
 565 static DEFINE_SPINLOCK(call_lock);
 566 static struct call_data_struct *call_data;
 567
 568 extern unsigned long xcall_call_function;
 569
 570 /*
 571  * You must not call this function with disabled interrupts or from a
 572  * hardware interrupt handler or from a bottom half handler.
 573  */
 574 int smp_call_function(void (*func)(void *info), void *info,
 575                       int nonatomic, int wait)
 576 {
 577         struct call_data_struct data;
 578         int cpus = num_online_cpus() - 1;
 579         long timeout;
 580
 581         if (!cpus)
 582                 return 0;
 583
 584         /* Can deadlock when called with interrupts disabled */
 585         WARN_ON(irqs_disabled());
 586
 587         data.func = func;
 588         data.info = info;
 589         atomic_set(&data.finished, 0);
 590         data.wait = wait;
 591
 592         spin_lock(&call_lock);
 593
 594         call_data = &data;
 595
 596         smp_cross_call(&xcall_call_function, 0, 0, 0);
 597
 598         /*
 599          * Wait for other cpus to complete function or at
 600          * least snap the call data.
 601          */
 602         timeout = 1000000;
 603         while (atomic_read(&data.finished) != cpus) {
 604                 if (--timeout <= 0)
 605                         goto out_timeout;
 606                 barrier();
 607                 udelay(1);
 608         }
 609
 610         spin_unlock(&call_lock);
 611
 612         return 0;
 613
 614 out_timeout:
 615         spin_unlock(&call_lock);
 616         printk("XCALL: Remote cpus not responding, ncpus=%ld finished=%ld\n",
 617                (long) num_online_cpus() - 1L,
 618                (long) atomic_read(&data.finished));
 619         return 0;
 620 }
 621
 622 void smp_call_function_client(int irq, struct pt_regs *regs)
 623 {
 624         void (*func) (void *info) = call_data->func;
 625         void *info = call_data->info;
 626
 627         clear_softint(1 << irq);
 628         if (call_data->wait) {
 629                 /* let initiator proceed only after completion */
 630                 func(info);
 631                 atomic_inc(&call_data->finished);
 632         } else {
 633                 /* let initiator proceed after getting data */
 634                 atomic_inc(&call_data->finished);
 635                 func(info);
 636         }
 637 }
 638
 639 extern unsigned long xcall_flush_tlb_mm;
 640 extern unsigned long xcall_flush_tlb_pending;
 641 extern unsigned long xcall_flush_tlb_kernel_range;
 642 extern unsigned long xcall_flush_tlb_all_spitfire;
 643 extern unsigned long xcall_flush_tlb_all_cheetah;
 644 extern unsigned long xcall_report_regs;
 645 extern unsigned long xcall_receive_signal;
 646
 647 #ifdef DCACHE_ALIASING_POSSIBLE
 648 extern unsigned long xcall_flush_dcache_page_cheetah;
 649 #endif
 650 extern unsigned long xcall_flush_dcache_page_spitfire;
 651
 652 #ifdef CONFIG_DEBUG_DCFLUSH
 653 extern atomic_t dcpage_flushes;
 654 extern atomic_t dcpage_flushes_xcall;
 655 #endif
 656
 657 static __inline__ void __local_flush_dcache_page(struct page *page)
 658 {
 659 #ifdef DCACHE_ALIASING_POSSIBLE
 660         __flush_dcache_page(page_address(page),
 661                             ((tlb_type == spitfire) &&
 662                              page_mapping(page) != NULL));
 663 #else
 664         if (page_mapping(page) != NULL &&
 665             tlb_type == spitfire)
 666                 __flush_icache_page(__pa(page_address(page)));
 667 #endif
 668 }
 669
 670 void smp_flush_dcache_page_impl(struct page *page, int cpu)
 671 {
 672         cpumask_t mask = cpumask_of_cpu(cpu);
 673         int this_cpu = get_cpu();
 674
 675 #ifdef CONFIG_DEBUG_DCFLUSH
 676         atomic_inc(&dcpage_flushes);
 677 #endif
 678         if (cpu == this_cpu) {
 679                 __local_flush_dcache_page(page);
 680         } else if (cpu_online(cpu)) {
 681                 void *pg_addr = page_address(page);
 682                 u64 data0;
 683
 684                 if (tlb_type == spitfire) {
 685                         data0 =
 686                                 ((u64)&xcall_flush_dcache_page_spitfire);
 687                         if (page_mapping(page) != NULL)
 688                                 data0 |= ((u64)1 << 32);
 689                         spitfire_xcall_deliver(data0,
 690                                                __pa(pg_addr),
 691                                                (u64) pg_addr,
 692                                                mask);
 693                 } else {
 694 #ifdef DCACHE_ALIASING_POSSIBLE
 695                         data0 =
 696                                 ((u64)&xcall_flush_dcache_page_cheetah);
 697                         cheetah_xcall_deliver(data0,
 698                                               __pa(pg_addr),
 699                                               0, mask);
 700 #endif
 701                 }
 702 #ifdef CONFIG_DEBUG_DCFLUSH
 703                 atomic_inc(&dcpage_flushes_xcall);
 704 #endif
 705         }
 706
 707         put_cpu();
 708 }
 709
 710 void flush_dcache_page_all(struct mm_struct *mm, struct page *page)
 711 {
 712         void *pg_addr = page_address(page);
 713         cpumask_t mask = cpu_online_map;
 714         u64 data0;
 715         int this_cpu = get_cpu();
 716
 717         cpu_clear(this_cpu, mask);
 718
 719 #ifdef CONFIG_DEBUG_DCFLUSH
 720         atomic_inc(&dcpage_flushes);
 721 #endif
 722         if (cpus_empty(mask))
 723                 goto flush_self;
 724         if (tlb_type == spitfire) {
 725                 data0 = ((u64)&xcall_flush_dcache_page_spitfire);
 726                 if (page_mapping(page) != NULL)
 727                         data0 |= ((u64)1 << 32);
 728                 spitfire_xcall_deliver(data0,
 729                                        __pa(pg_addr),
 730                                        (u64) pg_addr,
 731                                        mask);
 732         } else {
 733 #ifdef DCACHE_ALIASING_POSSIBLE
 734                 data0 = ((u64)&xcall_flush_dcache_page_cheetah);
 735                 cheetah_xcall_deliver(data0,
 736                                       __pa(pg_addr),
 737                                       0, mask);
 738 #endif
 739         }
 740 #ifdef CONFIG_DEBUG_DCFLUSH
 741         atomic_inc(&dcpage_flushes_xcall);
 742 #endif
 743  flush_self:
 744         __local_flush_dcache_page(page);
 745
 746         put_cpu();
 747 }
 748
 749 void smp_receive_signal(int cpu)
 750 {
 751         cpumask_t mask = cpumask_of_cpu(cpu);
 752
 753         if (cpu_online(cpu)) {
 754                 u64 data0 = (((u64)&xcall_receive_signal) & 0xffffffff);
 755
 756                 if (tlb_type == spitfire)
 757                         spitfire_xcall_deliver(data0, 0, 0, mask);
 758                 else
 759                         cheetah_xcall_deliver(data0, 0, 0, mask);
 760         }
 761 }
 762
 763 void smp_receive_signal_client(int irq, struct pt_regs *regs)
 764 {
 765         /* Just return, rtrap takes care of the rest. */
 766         clear_softint(1 << irq);
 767 }
 768
 769 void smp_report_regs(void)
 770 {
 771         smp_cross_call(&xcall_report_regs, 0, 0, 0);
 772 }
 773
 774 void smp_flush_tlb_all(void)
 775 {
 776         if (tlb_type == spitfire)
 777                 smp_cross_call(&xcall_flush_tlb_all_spitfire, 0, 0, 0);
 778         else
 779                 smp_cross_call(&xcall_flush_tlb_all_cheetah, 0, 0, 0);
 780         __flush_tlb_all();
 781 }
 782
 783 /* We know that the window frames of the user have been flushed
 784  * to the stack before we get here because all callers of us
 785  * are flush_tlb_*() routines, and these run after flush_cache_*()
 786  * which performs the flushw.
 787  *
 788  * The SMP TLB coherency scheme we use works as follows:
 789  *
 790  * 1) mm->cpu_vm_mask is a bit mask of which cpus an address
 791  *    space has (potentially) executed on, this is the heuristic
 792  *    we use to avoid doing cross calls.
 793  *
 794  *    Also, for flushing from kswapd and also for clones, we
 795  *    use cpu_vm_mask as the list of cpus to make run the TLB.
 796  *
 797  * 2) TLB context numbers are shared globally across all processors
 798  *    in the system, this allows us to play several games to avoid
 799  *    cross calls.
 800  *
 801  *    One invariant is that when a cpu switches to a process, and
 802  *    that processes tsk->active_mm->cpu_vm_mask does not have the
 803  *    current cpu's bit set, that tlb context is flushed locally.
 804  *
 805  *    If the address space is non-shared (ie. mm->count == 1) we avoid
 806  *    cross calls when we want to flush the currently running process's
 807  *    tlb state.  This is done by clearing all cpu bits except the current
 808  *    processor's in current->active_mm->cpu_vm_mask and performing the
 809  *    flush locally only.  This will force any subsequent cpus which run
 810  *    this task to flush the context from the local tlb if the process
 811  *    migrates to another cpu (again).
 812  *
 813  * 3) For shared address spaces (threads) and swapping we bite the
 814  *    bullet for most cases and perform the cross call (but only to
 815  *    the cpus listed in cpu_vm_mask).
 816  *
 817  *    The performance gain from "optimizing" away the cross call for threads is
 818  *    questionable (in theory the big win for threads is the massive sharing of
 819  *    address space state across processors).
 820  */
 821 void smp_flush_tlb_mm(struct mm_struct *mm)
 822 {
 823         /*
 824          * This code is called from two places, dup_mmap and exit_mmap. In the
 825          * former case, we really need a flush. In the later case, the callers
 826          * are single threaded exec_mmap (really need a flush), multithreaded
 827          * exec_mmap case (do not need to flush, since the caller gets a new
 828          * context via activate_mm), and all other callers of mmput() whence
 829          * the flush can be optimized since the associated threads are dead and
 830          * the mm is being torn down (__exit_mm and other mmput callers) or the
 831          * owning thread is dissociating itself from the mm. The
 832          * (atomic_read(&mm->mm_users) == 0) check ensures real work is done
 833          * for single thread exec and dup_mmap cases. An alternate check might
 834          * have been (current->mm != mm).
 835          *                                              Kanoj Sarcar
 836          */
 837         if (atomic_read(&mm->mm_users) == 0)
 838                 return;
 839
 840         {
 841                 u32 ctx = CTX_HWBITS(mm->context);
 842                 int cpu = get_cpu();
 843
 844                 if (atomic_read(&mm->mm_users) == 1) {
 845                         mm->cpu_vm_mask = cpumask_of_cpu(cpu);
 846                         goto local_flush_and_out;
 847                 }
 848
 849                 smp_cross_call_masked(&xcall_flush_tlb_mm,
 850                                       ctx, 0, 0,
 851                                       mm->cpu_vm_mask);
 852
 853         local_flush_and_out:
 854                 __flush_tlb_mm(ctx, SECONDARY_CONTEXT);
 855
 856                 put_cpu();
 857         }
 858 }
 859
 860 void smp_flush_tlb_pending(struct mm_struct *mm, unsigned long nr, unsigned long *vaddrs)
 861 {
 862         u32 ctx = CTX_HWBITS(mm->context);
 863         int cpu = get_cpu();
 864
 865         if (mm == current->active_mm && atomic_read(&mm->mm_users) == 1) {
 866                 mm->cpu_vm_mask = cpumask_of_cpu(cpu);
 867                 goto local_flush_and_out;
 868         } else {
 869                 /* This optimization is not valid.  Normally
 870                  * we will be holding the page_table_lock, but
 871                  * there is an exception which is copy_page_range()
 872                  * when forking.  The lock is held during the individual
 873                  * page table updates in the parent, but not at the
 874                  * top level, which is where we are invoked.
 875                  */
 876                 if (0) {
 877                         cpumask_t this_cpu_mask = cpumask_of_cpu(cpu);
 878
 879                         /* By virtue of running under the mm->page_table_lock,
 880                          * and mmu_context.h:switch_mm doing the same, the
 881                          * following operation is safe.
 882                          */
 883                         if (cpus_equal(mm->cpu_vm_mask, this_cpu_mask))
 884                                 goto local_flush_and_out;
 885                 }
 886         }
 887
 888         smp_cross_call_masked(&xcall_flush_tlb_pending,
 889                               ctx, nr, (unsigned long) vaddrs,
 890                               mm->cpu_vm_mask);
 891
 892 local_flush_and_out:
 893         __flush_tlb_pending(ctx, nr, vaddrs);
 894
 895         put_cpu();
 896 }
 897
 898 void smp_flush_tlb_kernel_range(unsigned long start, unsigned long end)
 899 {
 900         start &= PAGE_MASK;
 901         end    = PAGE_ALIGN(end);
 902         if (start != end) {
 903                 smp_cross_call(&xcall_flush_tlb_kernel_range,
 904                                0, start, end);
 905
 906                 __flush_tlb_kernel_range(start, end);
 907         }
 908 }
 909
 910 /* CPU capture. */
 911 /* #define CAPTURE_DEBUG */
 912 extern unsigned long xcall_capture;
 913
 914 static atomic_t smp_capture_depth = ATOMIC_INIT(0);
 915 static atomic_t smp_capture_registry = ATOMIC_INIT(0);
 916 static unsigned long penguins_are_doing_time;
 917
 918 void smp_capture(void)
 919 {
 920         int result = atomic_add_ret(1, &smp_capture_depth);
 921
 922         if (result == 1) {
 923                 int ncpus = num_online_cpus();
 924
 925 #ifdef CAPTURE_DEBUG
 926                 printk("CPU[%d]: Sending penguins to jail...",
 927                        smp_processor_id());
 928 #endif
 929                 penguins_are_doing_time = 1;
 930                 membar("#StoreStore | #LoadStore");
 931                 atomic_inc(&smp_capture_registry);
 932                 smp_cross_call(&xcall_capture, 0, 0, 0);
 933                 while (atomic_read(&smp_capture_registry) != ncpus)
 934                         membar("#LoadLoad");
 935 #ifdef CAPTURE_DEBUG
 936                 printk("done\n");
 937 #endif
 938         }
 939 }
 940
 941 void smp_release(void)
 942 {
 943         if (atomic_dec_and_test(&smp_capture_depth)) {
 944 #ifdef CAPTURE_DEBUG
 945                 printk("CPU[%d]: Giving pardon to "
 946                        "imprisoned penguins\n",
 947                        smp_processor_id());
 948 #endif
 949                 penguins_are_doing_time = 0;
 950                 membar("#StoreStore | #StoreLoad");
 951                 atomic_dec(&smp_capture_registry);
 952         }
 953 }
 954
 955 /* Imprisoned penguins run with %pil == 15, but PSTATE_IE set, so they
 956  * can service tlb flush xcalls...
 957  */
 958 extern void prom_world(int);
 959 extern void save_alternate_globals(unsigned long *);
 960 extern void restore_alternate_globals(unsigned long *);
 961 void smp_penguin_jailcell(int irq, struct pt_regs *regs)
 962 {
 963         unsigned long global_save[24];
 964
 965         clear_softint(1 << irq);
 966
 967         preempt_disable();
 968
 969         __asm__ __volatile__("flushw");
 970         save_alternate_globals(global_save);
 971         prom_world(1);
 972         atomic_inc(&smp_capture_registry);
 973         membar("#StoreLoad | #StoreStore");
 974         while (penguins_are_doing_time)
 975                 membar("#LoadLoad");
 976         restore_alternate_globals(global_save);
 977         atomic_dec(&smp_capture_registry);
 978         prom_world(0);
 979
 980         preempt_enable();
 981 }
 982
 983 extern unsigned long xcall_promstop;
 984
 985 void smp_promstop_others(void)
 986 {
 987         smp_cross_call(&xcall_promstop, 0, 0, 0);
 988 }
 989
 990 #define prof_multiplier(__cpu)          cpu_data(__cpu).multiplier
 991 #define prof_counter(__cpu)             cpu_data(__cpu).counter
 992
 993 void smp_percpu_timer_interrupt(struct pt_regs *regs)
 994 {
 995         unsigned long compare, tick, pstate;
 996         int cpu = smp_processor_id();
 997         int user = user_mode(regs);
 998
 999         /*
1000          * Check for level 14 softint.
1001          */
1002         {
1003                 unsigned long tick_mask = tick_ops->softint_mask;
1004
1005                 if (!(get_softint() & tick_mask)) {
1006                         extern void handler_irq(int, struct pt_regs *);
1007
1008                         handler_irq(14, regs);
1009                         return;
1010                 }
1011                 clear_softint(tick_mask);
1012         }
1013
1014         do {
1015                 profile_tick(CPU_PROFILING, regs);
1016                 if (!--prof_counter(cpu)) {
1017                         irq_enter();
1018
1019                         if (cpu == boot_cpu_id) {
1020                                 kstat_this_cpu.irqs[0]++;
1021                                 timer_tick_interrupt(regs);
1022                         }
1023
1024                         update_process_times(user);
1025
1026                         irq_exit();
1027
1028                         prof_counter(cpu) = prof_multiplier(cpu);
1029                 }
1030
1031                 /* Guarantee that the following sequences execute
1032                  * uninterrupted.
1033                  */
1034                 __asm__ __volatile__("rdpr      %%pstate, %0\n\t"
1035                                      "wrpr      %0, %1, %%pstate"
1036                                      : "=r" (pstate)
1037                                      : "i" (PSTATE_IE));
1038
1039                 compare = tick_ops->add_compare(current_tick_offset);
1040                 tick = tick_ops->get_tick();
1041
1042                 /* Restore PSTATE_IE. */
1043                 __asm__ __volatile__("wrpr      %0, 0x0, %%pstate"
1044                                      : /* no outputs */
1045                                      : "r" (pstate));
1046         } while (time_after_eq(tick, compare));
1047 }
1048
1049 static void __init smp_setup_percpu_timer(void)
1050 {
1051         int cpu = smp_processor_id();
1052         unsigned long pstate;
1053
1054         prof_counter(cpu) = prof_multiplier(cpu) = 1;
1055
1056         /* Guarantee that the following sequences execute
1057          * uninterrupted.
1058          */
1059         __asm__ __volatile__("rdpr      %%pstate, %0\n\t"
1060                              "wrpr      %0, %1, %%pstate"
1061                              : "=r" (pstate)
1062                              : "i" (PSTATE_IE));
1063
1064         tick_ops->init_tick(current_tick_offset);
1065
1066         /* Restore PSTATE_IE. */
1067         __asm__ __volatile__("wrpr      %0, 0x0, %%pstate"
1068                              : /* no outputs */
1069                              : "r" (pstate));
1070 }
1071
1072 void __init smp_tick_init(void)
1073 {
1074         boot_cpu_id = hard_smp_processor_id();
1075         current_tick_offset = timer_tick_offset;
1076
1077         cpu_set(boot_cpu_id, cpu_online_map);
1078         prof_counter(boot_cpu_id) = prof_multiplier(boot_cpu_id) = 1;
1079 }
1080
1081 /* /proc/profile writes can call this, don't __init it please. */
1082 static DEFINE_SPINLOCK(prof_setup_lock);
1083
1084 int setup_profiling_timer(unsigned int multiplier)
1085 {
1086         unsigned long flags;
1087         int i;
1088
1089         if ((!multiplier) || (timer_tick_offset / multiplier) < 1000)
1090                 return -EINVAL;
1091
1092         spin_lock_irqsave(&prof_setup_lock, flags);
1093         for (i = 0; i < NR_CPUS; i++)
1094                 prof_multiplier(i) = multiplier;
1095         current_tick_offset = (timer_tick_offset / multiplier);
1096         spin_unlock_irqrestore(&prof_setup_lock, flags);
1097
1098         return 0;
1099 }
1100
1101 void __init smp_prepare_cpus(unsigned int max_cpus)
1102 {
1103         int instance, mid;
1104
1105         instance = 0;
1106         while (!cpu_find_by_instance(instance, NULL, &mid)) {
1107                 if (mid < max_cpus)
1108                         cpu_set(mid, phys_cpu_present_map);
1109                 instance++;
1110         }
1111
1112         if (num_possible_cpus() > max_cpus) {
1113                 instance = 0;
1114                 while (!cpu_find_by_instance(instance, NULL, &mid)) {
1115                         if (mid != boot_cpu_id) {
1116                                 cpu_clear(mid, phys_cpu_present_map);
1117                                 if (num_possible_cpus() <= max_cpus)
1118                                         break;
1119                         }
1120                         instance++;
1121                 }
1122         }
1123
1124         smp_store_cpu_info(boot_cpu_id);
1125 }
1126
1127 void __devinit smp_prepare_boot_cpu(void)
1128 {
1129         if (hard_smp_processor_id() >= NR_CPUS) {
1130                 prom_printf("Serious problem, boot cpu id >= NR_CPUS\n");
1131                 prom_halt();
1132         }
1133
1134         current_thread_info()->cpu = hard_smp_processor_id();
1135
1136         cpu_set(smp_processor_id(), cpu_online_map);
1137         cpu_set(smp_processor_id(), phys_cpu_present_map);
1138 }
1139
1140 int __devinit __cpu_up(unsigned int cpu)
1141 {
1142         int ret = smp_boot_one_cpu(cpu);
1143
1144         if (!ret) {
1145                 cpu_set(cpu, smp_commenced_mask);
1146                 while (!cpu_isset(cpu, cpu_online_map))
1147                         mb();
1148                 if (!cpu_isset(cpu, cpu_online_map)) {
1149                         ret = -ENODEV;
1150                 } else {
1151                         smp_synchronize_one_tick(cpu);
1152                 }
1153         }
1154         return ret;
1155 }
1156
1157 void __init smp_cpus_done(unsigned int max_cpus)
1158 {
1159         unsigned long bogosum = 0;
1160         int i;
1161
1162         for (i = 0; i < NR_CPUS; i++) {
1163                 if (cpu_online(i))
1164                         bogosum += cpu_data(i).udelay_val;
1165         }
1166         printk("Total of %ld processors activated "
1167                "(%lu.%02lu BogoMIPS).\n",
1168                (long) num_online_cpus(),
1169                bogosum/(500000/HZ),
1170                (bogosum/(5000/HZ))%100);
1171 }
1172
1173 /* This needn't do anything as we do not sleep the cpu
1174  * inside of the idler task, so an interrupt is not needed
1175  * to get a clean fast response.
1176  *
1177  * XXX Reverify this assumption... -DaveM
1178  *
1179  * Addendum: We do want it to do something for the signal
1180  *           delivery case, we detect that by just seeing
1181  *           if we are trying to send this to an idler or not.
1182  */
1183 void smp_send_reschedule(int cpu)
1184 {
1185         if (cpu_data(cpu).idle_volume == 0)
1186                 smp_receive_signal(cpu);
1187 }
1188
1189 /* This is a nop because we capture all other cpus
1190  * anyways when making the PROM active.
1191  */
1192 void smp_send_stop(void)
1193 {
1194 }
1195
1196 unsigned long __per_cpu_base;
1197 unsigned long __per_cpu_shift;
1198
1199 EXPORT_SYMBOL(__per_cpu_base);
1200 EXPORT_SYMBOL(__per_cpu_shift);
1201
1202 void __init setup_per_cpu_areas(void)
1203 {
1204         unsigned long goal, size, i;
1205         char *ptr;
1206         /* Created by linker magic */
1207         extern char __per_cpu_start[], __per_cpu_end[];
1208
1209         /* Copy section for each CPU (we discard the original) */
1210         goal = ALIGN(__per_cpu_end - __per_cpu_start, PAGE_SIZE);
1211
1212 #ifdef CONFIG_MODULES
1213         if (goal < PERCPU_ENOUGH_ROOM)
1214                 goal = PERCPU_ENOUGH_ROOM;
1215 #endif
1216         __per_cpu_shift = 0;
1217         for (size = 1UL; size < goal; size <<= 1UL)
1218                 __per_cpu_shift++;
1219
1220         /* Make sure the resulting __per_cpu_base value
1221          * will fit in the 43-bit sign extended IMMU
1222          * TSB register.
1223          */
1224         ptr = __alloc_bootmem(size * NR_CPUS, PAGE_SIZE,
1225                               (unsigned long) __per_cpu_start);
1226
1227         __per_cpu_base = ptr - __per_cpu_start;
1228
1229         if ((__per_cpu_shift < PAGE_SHIFT) ||
1230             (__per_cpu_base & ~PAGE_MASK) ||
1231             (__per_cpu_base != (((long) __per_cpu_base << 20) >> 20))) {
1232                 prom_printf("PER_CPU: Invalid layout, "
1233                             "ptr[%p] shift[%lx] base[%lx]\n",
1234                             ptr, __per_cpu_shift, __per_cpu_base);
1235                 prom_halt();
1236         }
1237
1238         for (i = 0; i < NR_CPUS; i++, ptr += size)
1239                 memcpy(ptr, __per_cpu_start, __per_cpu_end - __per_cpu_start);
1240
1241         /* Finally, load in the boot cpu's base value.
1242          * We abuse the IMMU TSB register for trap handler
1243          * entry and exit loading of %g5.  That is why it
1244          * has to be page aligned.
1245          */
1246         cpu_setup_percpu_base(hard_smp_processor_id());
1247 }