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/kernel.h>
   7 #include <linux/sched.h>
   8 #include <linux/mm.h>
   9 #include <linux/pagemap.h>
  10 #include <linux/threads.h>
  11 #include <linux/smp.h>
  12 #include <linux/smp_lock.h>
  13 #include <linux/interrupt.h>
  14 #include <linux/kernel_stat.h>
  15 #include <linux/delay.h>
  16 #include <linux/init.h>
  17 #include <linux/spinlock.h>
  18 #include <linux/fs.h>
  19 #include <linux/seq_file.h>
  20 #include <linux/cache.h>
  21 #include <linux/jiffies.h>
  22
  23 #include <asm/head.h>
  24 #include <asm/ptrace.h>
  25 #include <asm/atomic.h>
  26 #include <asm/tlbflush.h>
  27 #include <asm/mmu_context.h>
  28 #include <asm/cpudata.h>
  29
  30 #include <asm/irq.h>
  31 #include <asm/page.h>
  32 #include <asm/pgtable.h>
  33 #include <asm/oplib.h>
  34 #include <asm/hardirq.h>
  35 #include <asm/uaccess.h>
  36 #include <asm/timer.h>
  37 #include <asm/starfire.h>
  38
  39 extern int linux_num_cpus;
  40 extern void calibrate_delay(void);
  41
  42 /* Please don't make this stuff initdata!!!  --DaveM */
  43 static unsigned char boot_cpu_id;
  44
  45 cpumask_t cpu_online_map = CPU_MASK_NONE;
  46 cpumask_t phys_cpu_present_map = CPU_MASK_NONE;
  47 static cpumask_t smp_commenced_mask;
  48 static cpumask_t cpu_callout_map;
  49
  50 void smp_info(struct seq_file *m)
  51 {
  52         int i;
  53
  54         seq_printf(m, "State:\n");
  55         for (i = 0; i < NR_CPUS; i++) {
  56                 if (cpu_online(i))
  57                         seq_printf(m,
  58                                    "CPU%d:\t\tonline\n", i);
  59         }
  60 }
  61
  62 void smp_bogo(struct seq_file *m)
  63 {
  64         int i;
  65
  66         for (i = 0; i < NR_CPUS; i++)
  67                 if (cpu_online(i))
  68                         seq_printf(m,
  69                                    "Cpu%dBogo\t: %lu.%02lu\n"
  70                                    "Cpu%dClkTck\t: %016lx\n",
  71                                    i, cpu_data(i).udelay_val / (500000/HZ),
  72                                    (cpu_data(i).udelay_val / (5000/HZ)) % 100,
  73                                    i, cpu_data(i).clock_tick);
  74 }
  75
  76 void __init smp_store_cpu_info(int id)
  77 {
  78         int cpu_node;
  79
  80         /* multiplier and counter set by
  81            smp_setup_percpu_timer()  */
  82         cpu_data(id).udelay_val                 = loops_per_jiffy;
  83
  84         cpu_find_by_mid(id, &cpu_node);
  85         cpu_data(id).clock_tick = prom_getintdefault(cpu_node,
  86                                                      "clock-frequency", 0);
  87
  88         cpu_data(id).pgcache_size               = 0;
  89         cpu_data(id).pte_cache[0]               = NULL;
  90         cpu_data(id).pte_cache[1]               = NULL;
  91         cpu_data(id).pgdcache_size              = 0;
  92         cpu_data(id).pgd_cache                  = NULL;
  93         cpu_data(id).idle_volume                = 1;
  94 }
  95
  96 static void smp_setup_percpu_timer(void);
  97
  98 static volatile unsigned long callin_flag = 0;
  99
 100 extern void inherit_locked_prom_mappings(int save_p);
 101
 102 void __init smp_callin(void)
 103 {
 104         int cpuid = hard_smp_processor_id();
 105         extern int bigkernel;
 106         extern unsigned long kern_locked_tte_data;
 107
 108         if (bigkernel) {
 109                 prom_dtlb_load(sparc64_highest_locked_tlbent()-1,
 110                         kern_locked_tte_data + 0x400000, KERNBASE + 0x400000);
 111                 prom_itlb_load(sparc64_highest_locked_tlbent()-1,
 112                         kern_locked_tte_data + 0x400000, KERNBASE + 0x400000);
 113         }
 114
 115         inherit_locked_prom_mappings(0);
 116
 117         __flush_tlb_all();
 118
 119         smp_setup_percpu_timer();
 120
 121         local_irq_enable();
 122
 123         calibrate_delay();
 124         smp_store_cpu_info(cpuid);
 125         callin_flag = 1;
 126         __asm__ __volatile__("membar #Sync\n\t"
 127                              "flush  %%g6" : : : "memory");
 128
 129         /* Clear this or we will die instantly when we
 130          * schedule back to this idler...
 131          */
 132         clear_thread_flag(TIF_NEWCHILD);
 133
 134         /* Attach to the address space of init_task. */
 135         atomic_inc(&init_mm.mm_count);
 136         current->active_mm = &init_mm;
 137
 138         while (!cpu_isset(cpuid, smp_commenced_mask))
 139                 membar("#LoadLoad");
 140
 141         cpu_set(cpuid, cpu_online_map);
 142 }
 143
 144 void cpu_panic(void)
 145 {
 146         printk("CPU[%d]: Returns from cpu_idle!\n", smp_processor_id());
 147         panic("SMP bolixed\n");
 148 }
 149
 150 static unsigned long current_tick_offset;
 151
 152 /* This tick register synchronization scheme is taken entirely from
 153  * the ia64 port, see arch/ia64/kernel/smpboot.c for details and credit.
 154  *
 155  * The only change I've made is to rework it so that the master
 156  * initiates the synchonization instead of the slave. -DaveM
 157  */
 158
 159 #define MASTER  0
 160 #define SLAVE   (SMP_CACHE_BYTES/sizeof(unsigned long))
 161
 162 #define NUM_ROUNDS      64      /* magic value */
 163 #define NUM_ITERS       5       /* likewise */
 164
 165 static spinlock_t itc_sync_lock = SPIN_LOCK_UNLOCKED;
 166 static unsigned long go[SLAVE + 1];
 167
 168 #define DEBUG_TICK_SYNC 0
 169
 170 static inline long get_delta (long *rt, long *master)
 171 {
 172         unsigned long best_t0 = 0, best_t1 = ~0UL, best_tm = 0;
 173         unsigned long tcenter, t0, t1, tm;
 174         unsigned long i;
 175
 176         for (i = 0; i < NUM_ITERS; i++) {
 177                 t0 = tick_ops->get_tick();
 178                 go[MASTER] = 1;
 179                 membar("#StoreLoad");
 180                 while (!(tm = go[SLAVE]))
 181                         membar("#LoadLoad");
 182                 go[SLAVE] = 0;
 183                 membar("#StoreStore");
 184                 t1 = tick_ops->get_tick();
 185
 186                 if (t1 - t0 < best_t1 - best_t0)
 187                         best_t0 = t0, best_t1 = t1, best_tm = tm;
 188         }
 189
 190         *rt = best_t1 - best_t0;
 191         *master = best_tm - best_t0;
 192
 193         /* average best_t0 and best_t1 without overflow: */
 194         tcenter = (best_t0/2 + best_t1/2);
 195         if (best_t0 % 2 + best_t1 % 2 == 2)
 196                 tcenter++;
 197         return tcenter - best_tm;
 198 }
 199
 200 void smp_synchronize_tick_client(void)
 201 {
 202         long i, delta, adj, adjust_latency = 0, done = 0;
 203         unsigned long flags, rt, master_time_stamp, bound;
 204 #if DEBUG_TICK_SYNC
 205         struct {
 206                 long rt;        /* roundtrip time */
 207                 long master;    /* master's timestamp */
 208                 long diff;      /* difference between midpoint and master's timestamp */
 209                 long lat;       /* estimate of itc adjustment latency */
 210         } t[NUM_ROUNDS];
 211 #endif
 212
 213         go[MASTER] = 1;
 214
 215         while (go[MASTER])
 216                 membar("#LoadLoad");
 217
 218         local_irq_save(flags);
 219         {
 220                 for (i = 0; i < NUM_ROUNDS; i++) {
 221                         delta = get_delta(&rt, &master_time_stamp);
 222                         if (delta == 0) {
 223                                 done = 1;       /* let's lock on to this... */
 224                                 bound = rt;
 225                         }
 226
 227                         if (!done) {
 228                                 if (i > 0) {
 229                                         adjust_latency += -delta;
 230                                         adj = -delta + adjust_latency/4;
 231                                 } else
 232                                         adj = -delta;
 233
 234                                 tick_ops->add_tick(adj, current_tick_offset);
 235                         }
 236 #if DEBUG_TICK_SYNC
 237                         t[i].rt = rt;
 238                         t[i].master = master_time_stamp;
 239                         t[i].diff = delta;
 240                         t[i].lat = adjust_latency/4;
 241 #endif
 242                 }
 243         }
 244         local_irq_restore(flags);
 245
 246 #if DEBUG_TICK_SYNC
 247         for (i = 0; i < NUM_ROUNDS; i++)
 248                 printk("rt=%5ld master=%5ld diff=%5ld adjlat=%5ld\n",
 249                        t[i].rt, t[i].master, t[i].diff, t[i].lat);
 250 #endif
 251
 252         printk(KERN_INFO "CPU %d: synchronized TICK with master CPU (last diff %ld cycles,"
 253                "maxerr %lu cycles)\n", smp_processor_id(), delta, rt);
 254 }
 255
 256 static void smp_start_sync_tick_client(int cpu);
 257
 258 static void smp_synchronize_one_tick(int cpu)
 259 {
 260         unsigned long flags, i;
 261
 262         go[MASTER] = 0;
 263
 264         smp_start_sync_tick_client(cpu);
 265
 266         /* wait for client to be ready */
 267         while (!go[MASTER])
 268                 membar("#LoadLoad");
 269
 270         /* now let the client proceed into his loop */
 271         go[MASTER] = 0;
 272         membar("#StoreLoad");
 273
 274         spin_lock_irqsave(&itc_sync_lock, flags);
 275         {
 276                 for (i = 0; i < NUM_ROUNDS*NUM_ITERS; i++) {
 277                         while (!go[MASTER])
 278                                 membar("#LoadLoad");
 279                         go[MASTER] = 0;
 280                         membar("#StoreStore");
 281                         go[SLAVE] = tick_ops->get_tick();
 282                         membar("#StoreLoad");
 283                 }
 284         }
 285         spin_unlock_irqrestore(&itc_sync_lock, flags);
 286 }
 287
 288 extern unsigned long sparc64_cpu_startup;
 289
 290 /* The OBP cpu startup callback truncates the 3rd arg cookie to
 291  * 32-bits (I think) so to be safe we have it read the pointer
 292  * contained here so we work on >4GB machines. -DaveM
 293  */
 294 static struct thread_info *cpu_new_thread = NULL;
 295
 296 static int __devinit smp_boot_one_cpu(unsigned int cpu)
 297 {
 298         unsigned long entry =
 299                 (unsigned long)(&sparc64_cpu_startup);
 300         unsigned long cookie =
 301                 (unsigned long)(&cpu_new_thread);
 302         struct task_struct *p;
 303         int timeout, ret, cpu_node;
 304
 305         kernel_thread(NULL, NULL, CLONE_IDLETASK);
 306
 307         p = prev_task(&init_task);
 308
 309         init_idle(p, cpu);
 310
 311         unhash_process(p);
 312
 313         callin_flag = 0;
 314         cpu_new_thread = p->thread_info;
 315         cpu_set(cpu, cpu_callout_map);
 316
 317         cpu_find_by_mid(cpu, &cpu_node);
 318         prom_startcpu(cpu_node, entry, cookie);
 319
 320         for (timeout = 0; timeout < 5000000; timeout++) {
 321                 if (callin_flag)
 322                         break;
 323                 udelay(100);
 324         }
 325         if (callin_flag) {
 326                 ret = 0;
 327         } else {
 328                 printk("Processor %d is stuck.\n", cpu);
 329                 cpu_clear(cpu, cpu_callout_map);
 330                 ret = -ENODEV;
 331         }
 332         cpu_new_thread = NULL;
 333
 334         return ret;
 335 }
 336
 337 static void spitfire_xcall_helper(u64 data0, u64 data1, u64 data2, u64 pstate, unsigned long cpu)
 338 {
 339         u64 result, target;
 340         int stuck, tmp;
 341
 342         if (this_is_starfire) {
 343                 /* map to real upaid */
 344                 cpu = (((cpu & 0x3c) << 1) |
 345                         ((cpu & 0x40) >> 4) |
 346                         (cpu & 0x3));
 347         }
 348
 349         target = (cpu << 14) | 0x70;
 350 again:
 351         /* Ok, this is the real Spitfire Errata #54.
 352          * One must read back from a UDB internal register
 353          * after writes to the UDB interrupt dispatch, but
 354          * before the membar Sync for that write.
 355          * So we use the high UDB control register (ASI 0x7f,
 356          * ADDR 0x20) for the dummy read. -DaveM
 357          */
 358         tmp = 0x40;
 359         __asm__ __volatile__(
 360         "wrpr   %1, %2, %%pstate\n\t"
 361         "stxa   %4, [%0] %3\n\t"
 362         "stxa   %5, [%0+%8] %3\n\t"
 363         "add    %0, %8, %0\n\t"
 364         "stxa   %6, [%0+%8] %3\n\t"
 365         "membar #Sync\n\t"
 366         "stxa   %%g0, [%7] %3\n\t"
 367         "membar #Sync\n\t"
 368         "mov    0x20, %%g1\n\t"
 369         "ldxa   [%%g1] 0x7f, %%g0\n\t"
 370         "membar #Sync"
 371         : "=r" (tmp)
 372         : "r" (pstate), "i" (PSTATE_IE), "i" (ASI_INTR_W),
 373           "r" (data0), "r" (data1), "r" (data2), "r" (target),
 374           "r" (0x10), "0" (tmp)
 375         : "g1");
 376
 377         /* NOTE: PSTATE_IE is still clear. */
 378         stuck = 100000;
 379         do {
 380                 __asm__ __volatile__("ldxa [%%g0] %1, %0"
 381                         : "=r" (result)
 382                         : "i" (ASI_INTR_DISPATCH_STAT));
 383                 if (result == 0) {
 384                         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 385                                              : : "r" (pstate));
 386                         return;
 387                 }
 388                 stuck -= 1;
 389                 if (stuck == 0)
 390                         break;
 391         } while (result & 0x1);
 392         __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 393                              : : "r" (pstate));
 394         if (stuck == 0) {
 395                 printk("CPU[%d]: mondo stuckage result[%016lx]\n",
 396                        smp_processor_id(), result);
 397         } else {
 398                 udelay(2);
 399                 goto again;
 400         }
 401 }
 402
 403 static __inline__ void spitfire_xcall_deliver(u64 data0, u64 data1, u64 data2, cpumask_t mask)
 404 {
 405         u64 pstate;
 406         int i;
 407
 408         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 409         for_each_cpu_mask(i, mask)
 410                 spitfire_xcall_helper(data0, data1, data2, pstate, i);
 411 }
 412
 413 /* Cheetah now allows to send the whole 64-bytes of data in the interrupt
 414  * packet, but we have no use for that.  However we do take advantage of
 415  * the new pipelining feature (ie. dispatch to multiple cpus simultaneously).
 416  */
 417 static void cheetah_xcall_deliver(u64 data0, u64 data1, u64 data2, cpumask_t mask)
 418 {
 419         u64 pstate, ver;
 420         int nack_busy_id, is_jalapeno;
 421
 422         if (cpus_empty(mask))
 423                 return;
 424
 425         /* Unfortunately, someone at Sun had the brilliant idea to make the
 426          * busy/nack fields hard-coded by ITID number for this Ultra-III
 427          * derivative processor.
 428          */
 429         __asm__ ("rdpr %%ver, %0" : "=r" (ver));
 430         is_jalapeno = ((ver >> 32) == 0x003e0016);
 431
 432         __asm__ __volatile__("rdpr %%pstate, %0" : "=r" (pstate));
 433
 434 retry:
 435         __asm__ __volatile__("wrpr %0, %1, %%pstate\n\t"
 436                              : : "r" (pstate), "i" (PSTATE_IE));
 437
 438         /* Setup the dispatch data registers. */
 439         __asm__ __volatile__("stxa      %0, [%3] %6\n\t"
 440                              "stxa      %1, [%4] %6\n\t"
 441                              "stxa      %2, [%5] %6\n\t"
 442                              "membar    #Sync\n\t"
 443                              : /* no outputs */
 444                              : "r" (data0), "r" (data1), "r" (data2),
 445                                "r" (0x40), "r" (0x50), "r" (0x60),
 446                                "i" (ASI_INTR_W));
 447
 448         nack_busy_id = 0;
 449         {
 450                 int i;
 451
 452                 for_each_cpu_mask(i, mask) {
 453                         u64 target = (i << 14) | 0x70;
 454
 455                         if (!is_jalapeno)
 456                                 target |= (nack_busy_id << 24);
 457                         __asm__ __volatile__(
 458                                 "stxa   %%g0, [%0] %1\n\t"
 459                                 "membar #Sync\n\t"
 460                                 : /* no outputs */
 461                                 : "r" (target), "i" (ASI_INTR_W));
 462                         nack_busy_id++;
 463                 }
 464         }
 465
 466         /* Now, poll for completion. */
 467         {
 468                 u64 dispatch_stat;
 469                 long stuck;
 470
 471                 stuck = 100000 * nack_busy_id;
 472                 do {
 473                         __asm__ __volatile__("ldxa      [%%g0] %1, %0"
 474                                              : "=r" (dispatch_stat)
 475                                              : "i" (ASI_INTR_DISPATCH_STAT));
 476                         if (dispatch_stat == 0UL) {
 477                                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 478                                                      : : "r" (pstate));
 479                                 return;
 480                         }
 481                         if (!--stuck)
 482                                 break;
 483                 } while (dispatch_stat & 0x5555555555555555UL);
 484
 485                 __asm__ __volatile__("wrpr %0, 0x0, %%pstate"
 486                                      : : "r" (pstate));
 487
 488                 if ((dispatch_stat & ~(0x5555555555555555UL)) == 0) {
 489                         /* Busy bits will not clear, continue instead
 490                          * of freezing up on this cpu.
 491                          */
 492                         printk("CPU[%d]: mondo stuckage result[%016lx]\n",
 493                                smp_processor_id(), dispatch_stat);
 494                 } else {
 495                         int i, this_busy_nack = 0;
 496
 497                         /* Delay some random time with interrupts enabled
 498                          * to prevent deadlock.
 499                          */
 500                         udelay(2 * nack_busy_id);
 501
 502                         /* Clear out the mask bits for cpus which did not
 503                          * NACK us.
 504                          */
 505                         for_each_cpu_mask(i, mask) {
 506                                 u64 check_mask;
 507
 508                                 if (is_jalapeno)
 509                                         check_mask = (0x2UL << (2*i));
 510                                 else
 511                                         check_mask = (0x2UL <<
 512                                                       this_busy_nack);
 513                                 if ((dispatch_stat & check_mask) == 0)
 514                                         cpu_clear(i, mask);
 515                                 this_busy_nack += 2;
 516                         }
 517
 518                         goto retry;
 519                 }
 520         }
 521 }
 522
 523 /* Send cross call to all processors mentioned in MASK
 524  * except self.
 525  */
 526 static void smp_cross_call_masked(unsigned long *func, u32 ctx, u64 data1, u64 data2, cpumask_t mask)
 527 {
 528         u64 data0 = (((u64)ctx)<<32 | (((u64)func) & 0xffffffff));
 529         int this_cpu = get_cpu();
 530
 531         cpus_and(mask, mask, cpu_online_map);
 532         cpu_clear(this_cpu, mask);
 533
 534         if (tlb_type == spitfire)
 535                 spitfire_xcall_deliver(data0, data1, data2, mask);
 536         else
 537                 cheetah_xcall_deliver(data0, data1, data2, mask);
 538         /* NOTE: Caller runs local copy on master. */
 539
 540         put_cpu();
 541 }
 542
 543 extern unsigned long xcall_sync_tick;
 544
 545 static void smp_start_sync_tick_client(int cpu)
 546 {
 547         cpumask_t mask = cpumask_of_cpu(cpu);
 548
 549         smp_cross_call_masked(&xcall_sync_tick,
 550                               0, 0, 0, mask);
 551 }
 552
 553 /* Send cross call to all processors except self. */
 554 #define smp_cross_call(func, ctx, data1, data2) \
 555         smp_cross_call_masked(func, ctx, data1, data2, cpu_online_map)
 556
 557 struct call_data_struct {
 558         void (*func) (void *info);
 559         void *info;
 560         atomic_t finished;
 561         int wait;
 562 };
 563
 564 static spinlock_t call_lock = SPIN_LOCK_UNLOCKED;
 565 static struct call_data_struct *call_data;
 566
 567 extern unsigned long xcall_call_function;
 568
 569 /*
 570  * You must not call this function with disabled interrupts or from a
 571  * hardware interrupt handler or from a bottom half handler.
 572  */
 573 int smp_call_function(void (*func)(void *info), void *info,
 574                       int nonatomic, int wait)
 575 {
 576         struct call_data_struct data;
 577         int cpus = num_online_cpus() - 1;
 578         long timeout;
 579
 580         if (!cpus)
 581                 return 0;
 582
 583         /* Can deadlock when called with interrupts disabled */
 584         WARN_ON(irqs_disabled());
 585
 586         data.func = func;
 587         data.info = info;
 588         atomic_set(&data.finished, 0);
 589         data.wait = wait;
 590
 591         spin_lock(&call_lock);
 592
 593         call_data = &data;
 594
 595         smp_cross_call(&xcall_call_function, 0, 0, 0);
 596
 597         /*
 598          * Wait for other cpus to complete function or at
 599          * least snap the call data.
 600          */
 601         timeout = 1000000;
 602         while (atomic_read(&data.finished) != cpus) {
 603                 if (--timeout <= 0)
 604                         goto out_timeout;
 605                 barrier();
 606                 udelay(1);
 607         }
 608
 609         spin_unlock(&call_lock);
 610
 611         return 0;
 612
 613 out_timeout:
 614         spin_unlock(&call_lock);
 615         printk("XCALL: Remote cpus not responding, ncpus=%ld finished=%ld\n",
 616                (long) num_online_cpus() - 1L,
 617                (long) atomic_read(&data.finished));
 618         return 0;
 619 }
 620
 621 void smp_call_function_client(int irq, struct pt_regs *regs)
 622 {
 623         void (*func) (void *info) = call_data->func;
 624         void *info = call_data->info;
 625
 626         clear_softint(1 << irq);
 627         if (call_data->wait) {
 628                 /* let initiator proceed only after completion */
 629                 func(info);
 630                 atomic_inc(&call_data->finished);
 631         } else {
 632                 /* let initiator proceed after getting data */
 633                 atomic_inc(&call_data->finished);
 634                 func(info);
 635         }
 636 }
 637
 638 extern unsigned long xcall_flush_tlb_page;
 639 extern unsigned long xcall_flush_tlb_mm;
 640 extern unsigned long xcall_flush_tlb_range;
 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 extern unsigned long xcall_flush_dcache_page_cheetah;
 647 extern unsigned long xcall_flush_dcache_page_spitfire;
 648
 649 #ifdef CONFIG_DEBUG_DCFLUSH
 650 extern atomic_t dcpage_flushes;
 651 extern atomic_t dcpage_flushes_xcall;
 652 #endif
 653
 654 static __inline__ void __local_flush_dcache_page(struct page *page)
 655 {
 656 #if (L1DCACHE_SIZE > PAGE_SIZE)
 657         __flush_dcache_page(page_address(page),
 658                             ((tlb_type == spitfire) &&
 659                              page_mapping(page) != NULL));
 660 #else
 661         if (page_mapping(page) != NULL &&
 662             tlb_type == spitfire)
 663                 __flush_icache_page(__pa(page_address(page)));
 664 #endif
 665 }
 666
 667 void smp_flush_dcache_page_impl(struct page *page, int cpu)
 668 {
 669         cpumask_t mask = cpumask_of_cpu(cpu);
 670         int this_cpu = get_cpu();
 671
 672 #ifdef CONFIG_DEBUG_DCFLUSH
 673         atomic_inc(&dcpage_flushes);
 674 #endif
 675         if (cpu == this_cpu) {
 676                 __local_flush_dcache_page(page);
 677         } else if (cpu_online(cpu)) {
 678                 void *pg_addr = page_address(page);
 679                 u64 data0;
 680
 681                 if (tlb_type == spitfire) {
 682                         data0 =
 683                                 ((u64)&xcall_flush_dcache_page_spitfire);
 684                         if (page_mapping(page) != NULL)
 685                                 data0 |= ((u64)1 << 32);
 686                         spitfire_xcall_deliver(data0,
 687                                                __pa(pg_addr),
 688                                                (u64) pg_addr,
 689                                                mask);
 690                 } else {
 691                         data0 =
 692                                 ((u64)&xcall_flush_dcache_page_cheetah);
 693                         cheetah_xcall_deliver(data0,
 694                                               __pa(pg_addr),
 695                                               0, mask);
 696                 }
 697 #ifdef CONFIG_DEBUG_DCFLUSH
 698                 atomic_inc(&dcpage_flushes_xcall);
 699 #endif
 700         }
 701
 702         put_cpu();
 703 }
 704
 705 void flush_dcache_page_all(struct mm_struct *mm, struct page *page)
 706 {
 707         void *pg_addr = page_address(page);
 708         cpumask_t mask = cpu_online_map;
 709         u64 data0;
 710         int this_cpu = get_cpu();
 711
 712         cpu_clear(this_cpu, mask);
 713
 714 #ifdef CONFIG_DEBUG_DCFLUSH
 715         atomic_inc(&dcpage_flushes);
 716 #endif
 717         if (cpus_empty(mask))
 718                 goto flush_self;
 719         if (tlb_type == spitfire) {
 720                 data0 = ((u64)&xcall_flush_dcache_page_spitfire);
 721                 if (page_mapping(page) != NULL)
 722                         data0 |= ((u64)1 << 32);
 723                 spitfire_xcall_deliver(data0,
 724                                        __pa(pg_addr),
 725                                        (u64) pg_addr,
 726                                        mask);
 727         } else {
 728                 data0 = ((u64)&xcall_flush_dcache_page_cheetah);
 729                 cheetah_xcall_deliver(data0,
 730                                       __pa(pg_addr),
 731                                       0, mask);
 732         }
 733 #ifdef CONFIG_DEBUG_DCFLUSH
 734         atomic_inc(&dcpage_flushes_xcall);
 735 #endif
 736  flush_self:
 737         __local_flush_dcache_page(page);
 738
 739         put_cpu();
 740 }
 741
 742 void smp_receive_signal(int cpu)
 743 {
 744         cpumask_t mask = cpumask_of_cpu(cpu);
 745
 746         if (cpu_online(cpu)) {
 747                 u64 data0 = (((u64)&xcall_receive_signal) & 0xffffffff);
 748
 749                 if (tlb_type == spitfire)
 750                         spitfire_xcall_deliver(data0, 0, 0, mask);
 751                 else
 752                         cheetah_xcall_deliver(data0, 0, 0, mask);
 753         }
 754 }
 755
 756 void smp_receive_signal_client(int irq, struct pt_regs *regs)
 757 {
 758         /* Just return, rtrap takes care of the rest. */
 759         clear_softint(1 << irq);
 760 }
 761
 762 void smp_report_regs(void)
 763 {
 764         smp_cross_call(&xcall_report_regs, 0, 0, 0);
 765 }
 766
 767 void smp_flush_tlb_all(void)
 768 {
 769         if (tlb_type == spitfire)
 770                 smp_cross_call(&xcall_flush_tlb_all_spitfire, 0, 0, 0);
 771         else
 772                 smp_cross_call(&xcall_flush_tlb_all_cheetah, 0, 0, 0);
 773         __flush_tlb_all();
 774 }
 775
 776 /* We know that the window frames of the user have been flushed
 777  * to the stack before we get here because all callers of us
 778  * are flush_tlb_*() routines, and these run after flush_cache_*()
 779  * which performs the flushw.
 780  *
 781  * The SMP TLB coherency scheme we use works as follows:
 782  *
 783  * 1) mm->cpu_vm_mask is a bit mask of which cpus an address
 784  *    space has (potentially) executed on, this is the heuristic
 785  *    we use to avoid doing cross calls.
 786  *
 787  *    Also, for flushing from kswapd and also for clones, we
 788  *    use cpu_vm_mask as the list of cpus to make run the TLB.
 789  *
 790  * 2) TLB context numbers are shared globally across all processors
 791  *    in the system, this allows us to play several games to avoid
 792  *    cross calls.
 793  *
 794  *    One invariant is that when a cpu switches to a process, and
 795  *    that processes tsk->active_mm->cpu_vm_mask does not have the
 796  *    current cpu's bit set, that tlb context is flushed locally.
 797  *
 798  *    If the address space is non-shared (ie. mm->count == 1) we avoid
 799  *    cross calls when we want to flush the currently running process's
 800  *    tlb state.  This is done by clearing all cpu bits except the current
 801  *    processor's in current->active_mm->cpu_vm_mask and performing the
 802  *    flush locally only.  This will force any subsequent cpus which run
 803  *    this task to flush the context from the local tlb if the process
 804  *    migrates to another cpu (again).
 805  *
 806  * 3) For shared address spaces (threads) and swapping we bite the
 807  *    bullet for most cases and perform the cross call (but only to
 808  *    the cpus listed in cpu_vm_mask).
 809  *
 810  *    The performance gain from "optimizing" away the cross call for threads is
 811  *    questionable (in theory the big win for threads is the massive sharing of
 812  *    address space state across processors).
 813  */
 814 void smp_flush_tlb_mm(struct mm_struct *mm)
 815 {
 816         /*
 817          * This code is called from two places, dup_mmap and exit_mmap. In the
 818          * former case, we really need a flush. In the later case, the callers
 819          * are single threaded exec_mmap (really need a flush), multithreaded
 820          * exec_mmap case (do not need to flush, since the caller gets a new
 821          * context via activate_mm), and all other callers of mmput() whence
 822          * the flush can be optimized since the associated threads are dead and
 823          * the mm is being torn down (__exit_mm and other mmput callers) or the
 824          * owning thread is dissociating itself from the mm. The
 825          * (atomic_read(&mm->mm_users) == 0) check ensures real work is done
 826          * for single thread exec and dup_mmap cases. An alternate check might
 827          * have been (current->mm != mm).
 828          *                                              Kanoj Sarcar
 829          */
 830         if (atomic_read(&mm->mm_users) == 0)
 831                 return;
 832
 833         {
 834                 u32 ctx = CTX_HWBITS(mm->context);
 835                 int cpu = get_cpu();
 836
 837                 if (atomic_read(&mm->mm_users) == 1) {
 838                         /* See smp_flush_tlb_page for info about this. */
 839                         mm->cpu_vm_mask = cpumask_of_cpu(cpu);
 840                         goto local_flush_and_out;
 841                 }
 842
 843                 smp_cross_call_masked(&xcall_flush_tlb_mm,
 844                                       ctx, 0, 0,
 845                                       mm->cpu_vm_mask);
 846
 847         local_flush_and_out:
 848                 __flush_tlb_mm(ctx, SECONDARY_CONTEXT);
 849
 850                 put_cpu();
 851         }
 852 }
 853
 854 void smp_flush_tlb_range(struct mm_struct *mm, unsigned long start,
 855                          unsigned long end)
 856 {
 857         u32 ctx = CTX_HWBITS(mm->context);
 858         int cpu = get_cpu();
 859
 860         start &= PAGE_MASK;
 861         end    = PAGE_ALIGN(end);
 862
 863         if (mm == current->active_mm && atomic_read(&mm->mm_users) == 1) {
 864                 mm->cpu_vm_mask = cpumask_of_cpu(cpu);
 865                 goto local_flush_and_out;
 866         }
 867
 868         smp_cross_call_masked(&xcall_flush_tlb_range,
 869                               ctx, start, end,
 870                               mm->cpu_vm_mask);
 871
 872  local_flush_and_out:
 873         __flush_tlb_range(ctx, start, SECONDARY_CONTEXT,
 874                           end, PAGE_SIZE, (end-start));
 875
 876         put_cpu();
 877 }
 878
 879 void smp_flush_tlb_kernel_range(unsigned long start, unsigned long end)
 880 {
 881         start &= PAGE_MASK;
 882         end    = PAGE_ALIGN(end);
 883         if (start != end) {
 884                 smp_cross_call(&xcall_flush_tlb_kernel_range,
 885                                0, start, end);
 886
 887                 __flush_tlb_kernel_range(start, end);
 888         }
 889 }
 890
 891 void smp_flush_tlb_page(struct mm_struct *mm, unsigned long page)
 892 {
 893         {
 894                 u32 ctx = CTX_HWBITS(mm->context);
 895                 int cpu = get_cpu();
 896
 897                 page &= PAGE_MASK;
 898                 if (mm == current->active_mm &&
 899                     atomic_read(&mm->mm_users) == 1) {
 900                         /* By virtue of being the current address space, and
 901                          * having the only reference to it, the following
 902                          * operation is safe.
 903                          *
 904                          * It would not be a win to perform the xcall tlb
 905                          * flush in this case, because even if we switch back
 906                          * to one of the other processors in cpu_vm_mask it
 907                          * is almost certain that all TLB entries for this
 908                          * context will be replaced by the time that happens.
 909                          */
 910                         mm->cpu_vm_mask = cpumask_of_cpu(cpu);
 911                         goto local_flush_and_out;
 912                 } else {
 913                         cpumask_t this_cpu_mask = cpumask_of_cpu(cpu);
 914
 915                         /* By virtue of running under the mm->page_table_lock,
 916                          * and mmu_context.h:switch_mm doing the same, the
 917                          * following operation is safe.
 918                          */
 919                         if (cpus_equal(mm->cpu_vm_mask, this_cpu_mask))
 920                                 goto local_flush_and_out;
 921                 }
 922
 923                 /* OK, we have to actually perform the cross call.  Most
 924                  * likely this is a cloned mm or kswapd is kicking out pages
 925                  * for a task which has run recently on another cpu.
 926                  */
 927                 smp_cross_call_masked(&xcall_flush_tlb_page,
 928                                       ctx, page, 0,
 929                                       mm->cpu_vm_mask);
 930                 if (!cpu_isset(cpu, mm->cpu_vm_mask))
 931                         return;
 932
 933         local_flush_and_out:
 934                 __flush_tlb_page(ctx, page, SECONDARY_CONTEXT);
 935
 936                 put_cpu();
 937         }
 938 }
 939
 940 /* CPU capture. */
 941 /* #define CAPTURE_DEBUG */
 942 extern unsigned long xcall_capture;
 943
 944 static atomic_t smp_capture_depth = ATOMIC_INIT(0);
 945 static atomic_t smp_capture_registry = ATOMIC_INIT(0);
 946 static unsigned long penguins_are_doing_time;
 947
 948 void smp_capture(void)
 949 {
 950         int result = __atomic_add(1, &smp_capture_depth);
 951
 952         membar("#StoreStore | #LoadStore");
 953         if (result == 1) {
 954                 int ncpus = num_online_cpus();
 955
 956 #ifdef CAPTURE_DEBUG
 957                 printk("CPU[%d]: Sending penguins to jail...",
 958                        smp_processor_id());
 959 #endif
 960                 penguins_are_doing_time = 1;
 961                 membar("#StoreStore | #LoadStore");
 962                 atomic_inc(&smp_capture_registry);
 963                 smp_cross_call(&xcall_capture, 0, 0, 0);
 964                 while (atomic_read(&smp_capture_registry) != ncpus)
 965                         membar("#LoadLoad");
 966 #ifdef CAPTURE_DEBUG
 967                 printk("done\n");
 968 #endif
 969         }
 970 }
 971
 972 void smp_release(void)
 973 {
 974         if (atomic_dec_and_test(&smp_capture_depth)) {
 975 #ifdef CAPTURE_DEBUG
 976                 printk("CPU[%d]: Giving pardon to "
 977                        "imprisoned penguins\n",
 978                        smp_processor_id());
 979 #endif
 980                 penguins_are_doing_time = 0;
 981                 membar("#StoreStore | #StoreLoad");
 982                 atomic_dec(&smp_capture_registry);
 983         }
 984 }
 985
 986 /* Imprisoned penguins run with %pil == 15, but PSTATE_IE set, so they
 987  * can service tlb flush xcalls...
 988  */
 989 extern void prom_world(int);
 990 extern void save_alternate_globals(unsigned long *);
 991 extern void restore_alternate_globals(unsigned long *);
 992 void smp_penguin_jailcell(int irq, struct pt_regs *regs)
 993 {
 994         unsigned long global_save[24];
 995
 996         clear_softint(1 << irq);
 997
 998         preempt_disable();
 999
1000         __asm__ __volatile__("flushw");
1001         save_alternate_globals(global_save);
1002         prom_world(1);
1003         atomic_inc(&smp_capture_registry);
1004         membar("#StoreLoad | #StoreStore");
1005         while (penguins_are_doing_time)
1006                 membar("#LoadLoad");
1007         restore_alternate_globals(global_save);
1008         atomic_dec(&smp_capture_registry);
1009         prom_world(0);
1010
1011         preempt_enable();
1012 }
1013
1014 extern unsigned long xcall_promstop;
1015
1016 void smp_promstop_others(void)
1017 {
1018         smp_cross_call(&xcall_promstop, 0, 0, 0);
1019 }
1020
1021 extern void sparc64_do_profile(struct pt_regs *regs);
1022
1023 #define prof_multiplier(__cpu)          cpu_data(__cpu).multiplier
1024 #define prof_counter(__cpu)             cpu_data(__cpu).counter
1025
1026 void smp_percpu_timer_interrupt(struct pt_regs *regs)
1027 {
1028         unsigned long compare, tick, pstate;
1029         int cpu = smp_processor_id();
1030         int user = user_mode(regs);
1031
1032         /*
1033          * Check for level 14 softint.
1034          */
1035         {
1036                 unsigned long tick_mask = tick_ops->softint_mask;
1037
1038                 if (!(get_softint() & tick_mask)) {
1039                         extern void handler_irq(int, struct pt_regs *);
1040
1041                         handler_irq(14, regs);
1042                         return;
1043                 }
1044                 clear_softint(tick_mask);
1045         }
1046
1047         do {
1048                 sparc64_do_profile(regs);
1049                 if (!--prof_counter(cpu)) {
1050                         irq_enter();
1051
1052                         if (cpu == boot_cpu_id) {
1053                                 kstat_this_cpu.irqs[0]++;
1054                                 timer_tick_interrupt(regs);
1055                         }
1056
1057                         update_process_times(user);
1058
1059                         irq_exit();
1060
1061                         prof_counter(cpu) = prof_multiplier(cpu);
1062                 }
1063
1064                 /* Guarantee that the following sequences execute
1065                  * uninterrupted.
1066                  */
1067                 __asm__ __volatile__("rdpr      %%pstate, %0\n\t"
1068                                      "wrpr      %0, %1, %%pstate"
1069                                      : "=r" (pstate)
1070                                      : "i" (PSTATE_IE));
1071
1072                 compare = tick_ops->add_compare(current_tick_offset);
1073                 tick = tick_ops->get_tick();
1074
1075                 /* Restore PSTATE_IE. */
1076                 __asm__ __volatile__("wrpr      %0, 0x0, %%pstate"
1077                                      : /* no outputs */
1078                                      : "r" (pstate));
1079         } while (time_after_eq(tick, compare));
1080 }
1081
1082 static void __init smp_setup_percpu_timer(void)
1083 {
1084         int cpu = smp_processor_id();
1085         unsigned long pstate;
1086
1087         prof_counter(cpu) = prof_multiplier(cpu) = 1;
1088
1089         /* Guarantee that the following sequences execute
1090          * uninterrupted.
1091          */
1092         __asm__ __volatile__("rdpr      %%pstate, %0\n\t"
1093                              "wrpr      %0, %1, %%pstate"
1094                              : "=r" (pstate)
1095                              : "i" (PSTATE_IE));
1096
1097         tick_ops->init_tick(current_tick_offset);
1098
1099         /* Restore PSTATE_IE. */
1100         __asm__ __volatile__("wrpr      %0, 0x0, %%pstate"
1101                              : /* no outputs */
1102                              : "r" (pstate));
1103 }
1104
1105 void __init smp_tick_init(void)
1106 {
1107         boot_cpu_id = hard_smp_processor_id();
1108         current_tick_offset = timer_tick_offset;
1109
1110         cpu_set(boot_cpu_id, cpu_online_map);
1111         prof_counter(boot_cpu_id) = prof_multiplier(boot_cpu_id) = 1;
1112 }
1113
1114 cycles_t cacheflush_time;
1115 unsigned long cache_decay_ticks;
1116
1117 extern unsigned long cheetah_tune_scheduling(void);
1118
1119 static void __init smp_tune_scheduling(void)
1120 {
1121         unsigned long orig_flush_base, flush_base, flags, *p;
1122         unsigned int ecache_size, order;
1123         cycles_t tick1, tick2, raw;
1124         int cpu_node;
1125
1126         /* Approximate heuristic for SMP scheduling.  It is an
1127          * estimation of the time it takes to flush the L2 cache
1128          * on the local processor.
1129          *
1130          * The ia32 chooses to use the L1 cache flush time instead,
1131          * and I consider this complete nonsense.  The Ultra can service
1132          * a miss to the L1 with a hit to the L2 in 7 or 8 cycles, and
1133          * L2 misses are what create extra bus traffic (ie. the "cost"
1134          * of moving a process from one cpu to another).
1135          */
1136         printk("SMP: Calibrating ecache flush... ");
1137         if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1138                 cacheflush_time = cheetah_tune_scheduling();
1139                 goto report;
1140         }
1141
1142         cpu_find_by_instance(0, &cpu_node, NULL);
1143         ecache_size = prom_getintdefault(cpu_node,
1144                                          "ecache-size", (512 * 1024));
1145         if (ecache_size > (4 * 1024 * 1024))
1146                 ecache_size = (4 * 1024 * 1024);
1147         orig_flush_base = flush_base =
1148                 __get_free_pages(GFP_KERNEL, order = get_order(ecache_size));
1149
1150         if (flush_base != 0UL) {
1151                 local_irq_save(flags);
1152
1153                 /* Scan twice the size once just to get the TLB entries
1154                  * loaded and make sure the second scan measures pure misses.
1155                  */
1156                 for (p = (unsigned long *)flush_base;
1157                      ((unsigned long)p) < (flush_base + (ecache_size<<1));
1158                      p += (64 / sizeof(unsigned long)))
1159                         *((volatile unsigned long *)p);
1160
1161                 tick1 = tick_ops->get_tick();
1162
1163                 __asm__ __volatile__("1:\n\t"
1164                                      "ldx       [%0 + 0x000], %%g1\n\t"
1165                                      "ldx       [%0 + 0x040], %%g2\n\t"
1166                                      "ldx       [%0 + 0x080], %%g3\n\t"
1167                                      "ldx       [%0 + 0x0c0], %%g5\n\t"
1168                                      "add       %0, 0x100, %0\n\t"
1169                                      "cmp       %0, %2\n\t"
1170                                      "bne,pt    %%xcc, 1b\n\t"
1171                                      " nop"
1172                                      : "=&r" (flush_base)
1173                                      : "0" (flush_base),
1174                                        "r" (flush_base + ecache_size)
1175                                      : "g1", "g2", "g3", "g5");
1176
1177                 tick2 = tick_ops->get_tick();
1178
1179                 local_irq_restore(flags);
1180
1181                 raw = (tick2 - tick1);
1182
1183                 /* Dampen it a little, considering two processes
1184                  * sharing the cache and fitting.
1185                  */
1186                 cacheflush_time = (raw - (raw >> 2));
1187
1188                 free_pages(orig_flush_base, order);
1189         } else {
1190                 cacheflush_time = ((ecache_size << 2) +
1191                                    (ecache_size << 1));
1192         }
1193 report:
1194         /* Convert ticks/sticks to jiffies. */
1195         cache_decay_ticks = cacheflush_time / timer_tick_offset;
1196         if (cache_decay_ticks < 1)
1197                 cache_decay_ticks = 1;
1198
1199         printk("Using heuristic of %ld cycles, %ld ticks.\n",
1200                cacheflush_time, cache_decay_ticks);
1201 }
1202
1203 /* /proc/profile writes can call this, don't __init it please. */
1204 static spinlock_t prof_setup_lock = SPIN_LOCK_UNLOCKED;
1205
1206 int setup_profiling_timer(unsigned int multiplier)
1207 {
1208         unsigned long flags;
1209         int i;
1210
1211         if ((!multiplier) || (timer_tick_offset / multiplier) < 1000)
1212                 return -EINVAL;
1213
1214         spin_lock_irqsave(&prof_setup_lock, flags);
1215         for (i = 0; i < NR_CPUS; i++)
1216                 prof_multiplier(i) = multiplier;
1217         current_tick_offset = (timer_tick_offset / multiplier);
1218         spin_unlock_irqrestore(&prof_setup_lock, flags);
1219
1220         return 0;
1221 }
1222
1223 void __init smp_prepare_cpus(unsigned int max_cpus)
1224 {
1225         int instance, mid;
1226
1227         instance = 0;
1228         while (!cpu_find_by_instance(instance, NULL, &mid)) {
1229                 if (mid < max_cpus)
1230                         cpu_set(mid, phys_cpu_present_map);
1231                 instance++;
1232         }
1233
1234         if (num_possible_cpus() > max_cpus) {
1235                 instance = 0;
1236                 while (!cpu_find_by_instance(instance, NULL, &mid)) {
1237                         if (mid != boot_cpu_id) {
1238                                 cpu_clear(mid, phys_cpu_present_map);
1239                                 if (num_possible_cpus() <= max_cpus)
1240                                         break;
1241                         }
1242                         instance++;
1243                 }
1244         }
1245
1246         smp_store_cpu_info(boot_cpu_id);
1247 }
1248
1249 void __devinit smp_prepare_boot_cpu(void)
1250 {
1251         if (hard_smp_processor_id() >= NR_CPUS) {
1252                 prom_printf("Serious problem, boot cpu id >= NR_CPUS\n");
1253                 prom_halt();
1254         }
1255
1256         current_thread_info()->cpu = hard_smp_processor_id();
1257         cpu_set(smp_processor_id(), cpu_online_map);
1258         cpu_set(smp_processor_id(), phys_cpu_present_map);
1259 }
1260
1261 int __devinit __cpu_up(unsigned int cpu)
1262 {
1263         int ret = smp_boot_one_cpu(cpu);
1264
1265         if (!ret) {
1266                 cpu_set(cpu, smp_commenced_mask);
1267                 while (!cpu_isset(cpu, cpu_online_map))
1268                         mb();
1269                 if (!cpu_isset(cpu, cpu_online_map)) {
1270                         ret = -ENODEV;
1271                 } else {
1272                         smp_synchronize_one_tick(cpu);
1273                 }
1274         }
1275         return ret;
1276 }
1277
1278 void __init smp_cpus_done(unsigned int max_cpus)
1279 {
1280         unsigned long bogosum = 0;
1281         int i;
1282
1283         for (i = 0; i < NR_CPUS; i++) {
1284                 if (cpu_online(i))
1285                         bogosum += cpu_data(i).udelay_val;
1286         }
1287         printk("Total of %ld processors activated "
1288                "(%lu.%02lu BogoMIPS).\n",
1289                (long) num_online_cpus(),
1290                bogosum/(500000/HZ),
1291                (bogosum/(5000/HZ))%100);
1292
1293         /* We want to run this with all the other cpus spinning
1294          * in the kernel.
1295          */
1296         smp_tune_scheduling();
1297 }
1298
1299 /* This needn't do anything as we do not sleep the cpu
1300  * inside of the idler task, so an interrupt is not needed
1301  * to get a clean fast response.
1302  *
1303  * XXX Reverify this assumption... -DaveM
1304  *
1305  * Addendum: We do want it to do something for the signal
1306  *           delivery case, we detect that by just seeing
1307  *           if we are trying to send this to an idler or not.
1308  */
1309 void smp_send_reschedule(int cpu)
1310 {
1311         if (cpu_data(cpu).idle_volume == 0)
1312                 smp_receive_signal(cpu);
1313 }
1314
1315 /* This is a nop because we capture all other cpus
1316  * anyways when making the PROM active.
1317  */
1318 void smp_send_stop(void)
1319 {
1320 }
1321