arch/sparc/kernel/sun4d_smp.c

   1 /* sun4d_smp.c: Sparc SS1000/SC2000 SMP support.
   2  *
   3  * Copyright (C) 1998 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
   4  *
   5  * Based on sun4m's smp.c, which is:
   6  * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu)
   7  */
   8
   9 #include <asm/head.h>
  10
  11 #include <linux/kernel.h>
  12 #include <linux/sched.h>
  13 #include <linux/threads.h>
  14 #include <linux/smp.h>
  15 #include <linux/smp_lock.h>
  16 #include <linux/interrupt.h>
  17 #include <linux/kernel_stat.h>
  18 #include <linux/init.h>
  19 #include <linux/spinlock.h>
  20 #include <linux/mm.h>
  21 #include <linux/swap.h>
  22 #include <linux/profile.h>
  23
  24 #include <asm/ptrace.h>
  25 #include <asm/atomic.h>
  26
  27 #include <asm/delay.h>
  28 #include <asm/irq.h>
  29 #include <asm/page.h>
  30 #include <asm/pgalloc.h>
  31 #include <asm/pgtable.h>
  32 #include <asm/oplib.h>
  33 #include <asm/sbus.h>
  34 #include <asm/sbi.h>
  35 #include <asm/tlbflush.h>
  36 #include <asm/cacheflush.h>
  37 #include <asm/cpudata.h>
  38
  39 #define IRQ_CROSS_CALL          15
  40
  41 extern ctxd_t *srmmu_ctx_table_phys;
  42
  43 extern void calibrate_delay(void);
  44
  45 extern volatile int smp_processors_ready;
  46 extern int smp_num_cpus;
  47 static int smp_highest_cpu;
  48 extern int smp_threads_ready;
  49 extern volatile unsigned long cpu_callin_map[NR_CPUS];
  50 extern struct cpuinfo_sparc cpu_data[NR_CPUS];
  51 extern unsigned char boot_cpu_id;
  52 extern int smp_activated;
  53 extern volatile int __cpu_number_map[NR_CPUS];
  54 extern volatile int __cpu_logical_map[NR_CPUS];
  55 extern volatile unsigned long ipi_count;
  56 extern volatile int smp_process_available;
  57 extern volatile int smp_commenced;
  58 extern int __smp4d_processor_id(void);
  59
  60 /* #define SMP_DEBUG */
  61
  62 #ifdef SMP_DEBUG
  63 #define SMP_PRINTK(x)   printk x
  64 #else
  65 #define SMP_PRINTK(x)
  66 #endif
  67
  68 static inline unsigned long swap(volatile unsigned long *ptr, unsigned long val)
  69 {
  70         __asm__ __volatile__("swap [%1], %0\n\t" :
  71                              "=&r" (val), "=&r" (ptr) :
  72                              "0" (val), "1" (ptr));
  73         return val;
  74 }
  75
  76 static void smp_setup_percpu_timer(void);
  77 extern void cpu_probe(void);
  78 extern void sun4d_distribute_irqs(void);
  79
  80 void __init smp4d_callin(void)
  81 {
  82         int cpuid = hard_smp4d_processor_id();
  83         extern spinlock_t sun4d_imsk_lock;
  84         unsigned long flags;
  85
  86         /* Show we are alive */
  87         cpu_leds[cpuid] = 0x6;
  88         show_leds(cpuid);
  89
  90         /* Enable level15 interrupt, disable level14 interrupt for now */
  91         cc_set_imsk((cc_get_imsk() & ~0x8000) | 0x4000);
  92
  93         local_flush_cache_all();
  94         local_flush_tlb_all();
  95
  96         /*
  97          * Unblock the master CPU _only_ when the scheduler state
  98          * of all secondary CPUs will be up-to-date, so after
  99          * the SMP initialization the master will be just allowed
 100          * to call the scheduler code.
 101          */
 102         /* Get our local ticker going. */
 103         smp_setup_percpu_timer();
 104
 105         calibrate_delay();
 106         smp_store_cpu_info(cpuid);
 107         local_flush_cache_all();
 108         local_flush_tlb_all();
 109
 110         /* Allow master to continue. */
 111         swap((unsigned long *)&cpu_callin_map[cpuid], 1);
 112         local_flush_cache_all();
 113         local_flush_tlb_all();
 114
 115         cpu_probe();
 116
 117         while((unsigned long)current_set[cpuid] < PAGE_OFFSET)
 118                 barrier();
 119
 120         while(current_set[cpuid]->cpu != cpuid)
 121                 barrier();
 122
 123         /* Fix idle thread fields. */
 124         __asm__ __volatile__("ld [%0], %%g6\n\t"
 125                              "sta %%g6, [%%g0] %1\n\t"
 126                              : : "r" (&current_set[cpuid]), "i" (ASI_M_VIKING_TMP2)
 127                              : "memory" /* paranoid */);
 128
 129         cpu_leds[cpuid] = 0x9;
 130         show_leds(cpuid);
 131
 132         /* Attach to the address space of init_task. */
 133         atomic_inc(&init_mm.mm_count);
 134         current->active_mm = &init_mm;
 135
 136         local_flush_cache_all();
 137         local_flush_tlb_all();
 138
 139         local_irq_enable();     /* We don't allow PIL 14 yet */
 140
 141         while(!smp_commenced)
 142                 barrier();
 143
 144         spin_lock_irqsave(&sun4d_imsk_lock, flags);
 145         cc_set_imsk(cc_get_imsk() & ~0x4000); /* Allow PIL 14 as well */
 146         spin_unlock_irqrestore(&sun4d_imsk_lock, flags);
 147 }
 148
 149 extern int cpu_idle(void *unused);
 150 extern void init_IRQ(void);
 151 extern void cpu_panic(void);
 152
 153 /*
 154  *      Cycle through the processors asking the PROM to start each one.
 155  */
 156
 157 extern struct linux_prom_registers smp_penguin_ctable;
 158 extern unsigned long trapbase_cpu1[];
 159 extern unsigned long trapbase_cpu2[];
 160 extern unsigned long trapbase_cpu3[];
 161
 162 void __init smp4d_boot_cpus(void)
 163 {
 164         int cpucount = 0;
 165         int i, mid;
 166
 167         printk("Entering SMP Mode...\n");
 168
 169         if (boot_cpu_id)
 170                 current_set[0] = NULL;
 171
 172         local_irq_enable();
 173         cpus_clear(cpu_present_map);
 174
 175         /* XXX This whole thing has to go.  See sparc64. */
 176         for (i = 0; !cpu_find_by_instance(i, NULL, &mid); i++)
 177                 cpu_set(mid, cpu_present_map);
 178         SMP_PRINTK(("cpu_present_map %08lx\n", cpus_addr(cpu_present_map)[0]));
 179         for(i=0; i < NR_CPUS; i++)
 180                 __cpu_number_map[i] = -1;
 181         for(i=0; i < NR_CPUS; i++)
 182                 __cpu_logical_map[i] = -1;
 183         __cpu_number_map[boot_cpu_id] = 0;
 184         __cpu_logical_map[0] = boot_cpu_id;
 185         current_thread_info()->cpu = boot_cpu_id;
 186         smp_store_cpu_info(boot_cpu_id);
 187         smp_setup_percpu_timer();
 188         local_flush_cache_all();
 189         if (cpu_find_by_instance(1, NULL, NULL))
 190                 return;  /* Not an MP box. */
 191         SMP_PRINTK(("Iterating over CPUs\n"));
 192         for(i = 0; i < NR_CPUS; i++) {
 193                 if(i == boot_cpu_id)
 194                         continue;
 195
 196                 if (cpu_isset(i, cpu_present_map)) {
 197                         extern unsigned long sun4d_cpu_startup;
 198                         unsigned long *entry = &sun4d_cpu_startup;
 199                         struct task_struct *p;
 200                         int timeout;
 201                         int no;
 202
 203                         /* Cook up an idler for this guy. */
 204                         p = fork_idle(i);
 205                         cpucount++;
 206                         current_set[i] = p->thread_info;
 207                         for (no = 0; !cpu_find_by_instance(no, NULL, &mid)
 208                                      && mid != i; no++) ;
 209
 210                         /*
 211                          * Initialize the contexts table
 212                          * Since the call to prom_startcpu() trashes the structure,
 213                          * we need to re-initialize it for each cpu
 214                          */
 215                         smp_penguin_ctable.which_io = 0;
 216                         smp_penguin_ctable.phys_addr = (unsigned int) srmmu_ctx_table_phys;
 217                         smp_penguin_ctable.reg_size = 0;
 218
 219                         /* whirrr, whirrr, whirrrrrrrrr... */
 220                         SMP_PRINTK(("Starting CPU %d at %p task %d node %08x\n", i, entry, cpucount, cpu_data(no).prom_node));
 221                         local_flush_cache_all();
 222                         prom_startcpu(cpu_data(no).prom_node,
 223                                       &smp_penguin_ctable, 0, (char *)entry);
 224
 225                         SMP_PRINTK(("prom_startcpu returned :)\n"));
 226
 227                         /* wheee... it's going... */
 228                         for(timeout = 0; timeout < 10000; timeout++) {
 229                                 if(cpu_callin_map[i])
 230                                         break;
 231                                 udelay(200);
 232                         }
 233
 234                         if(cpu_callin_map[i]) {
 235                                 /* Another "Red Snapper". */
 236                                 __cpu_number_map[i] = cpucount;
 237                                 __cpu_logical_map[cpucount] = i;
 238                         } else {
 239                                 cpucount--;
 240                                 printk("Processor %d is stuck.\n", i);
 241                         }
 242                 }
 243                 if(!(cpu_callin_map[i])) {
 244                         cpu_clear(i, cpu_present_map);
 245                         __cpu_number_map[i] = -1;
 246                 }
 247         }
 248         local_flush_cache_all();
 249         if(cpucount == 0) {
 250                 printk("Error: only one Processor found.\n");
 251                 cpu_present_map = cpumask_of_cpu(hard_smp4d_processor_id());
 252         } else {
 253                 unsigned long bogosum = 0;
 254
 255                 for(i = 0; i < NR_CPUS; i++) {
 256                         if (cpu_isset(i, cpu_present_map)) {
 257                                 bogosum += cpu_data(i).udelay_val;
 258                                 smp_highest_cpu = i;
 259                         }
 260                 }
 261                 SMP_PRINTK(("Total of %d Processors activated (%lu.%02lu BogoMIPS).\n", cpucount + 1, bogosum/(500000/HZ), (bogosum/(5000/HZ))%100));
 262                 printk("Total of %d Processors activated (%lu.%02lu BogoMIPS).\n",
 263                        cpucount + 1,
 264                        bogosum/(500000/HZ),
 265                        (bogosum/(5000/HZ))%100);
 266                 smp_activated = 1;
 267                 smp_num_cpus = cpucount + 1;
 268         }
 269
 270         /* Free unneeded trap tables */
 271         ClearPageReserved(virt_to_page(trapbase_cpu1));
 272         set_page_count(virt_to_page(trapbase_cpu1), 1);
 273         free_page((unsigned long)trapbase_cpu1);
 274         totalram_pages++;
 275         num_physpages++;
 276
 277         ClearPageReserved(virt_to_page(trapbase_cpu2));
 278         set_page_count(virt_to_page(trapbase_cpu2), 1);
 279         free_page((unsigned long)trapbase_cpu2);
 280         totalram_pages++;
 281         num_physpages++;
 282
 283         ClearPageReserved(virt_to_page(trapbase_cpu3));
 284         set_page_count(virt_to_page(trapbase_cpu3), 1);
 285         free_page((unsigned long)trapbase_cpu3);
 286         totalram_pages++;
 287         num_physpages++;
 288
 289         /* Ok, they are spinning and ready to go. */
 290         smp_processors_ready = 1;
 291         sun4d_distribute_irqs();
 292 }
 293
 294 static struct smp_funcall {
 295         smpfunc_t func;
 296         unsigned long arg1;
 297         unsigned long arg2;
 298         unsigned long arg3;
 299         unsigned long arg4;
 300         unsigned long arg5;
 301         unsigned char processors_in[NR_CPUS];  /* Set when ipi entered. */
 302         unsigned char processors_out[NR_CPUS]; /* Set when ipi exited. */
 303 } ccall_info __attribute__((aligned(8)));
 304
 305 static spinlock_t cross_call_lock = SPIN_LOCK_UNLOCKED;
 306
 307 /* Cross calls must be serialized, at least currently. */
 308 void smp4d_cross_call(smpfunc_t func, unsigned long arg1, unsigned long arg2,
 309                     unsigned long arg3, unsigned long arg4, unsigned long arg5)
 310 {
 311         if(smp_processors_ready) {
 312                 register int high = smp_highest_cpu;
 313                 unsigned long flags;
 314
 315                 spin_lock_irqsave(&cross_call_lock, flags);
 316
 317                 {
 318                         /* If you make changes here, make sure gcc generates proper code... */
 319                         register smpfunc_t f asm("i0") = func;
 320                         register unsigned long a1 asm("i1") = arg1;
 321                         register unsigned long a2 asm("i2") = arg2;
 322                         register unsigned long a3 asm("i3") = arg3;
 323                         register unsigned long a4 asm("i4") = arg4;
 324                         register unsigned long a5 asm("i5") = arg5;
 325
 326                         __asm__ __volatile__(
 327                                 "std %0, [%6]\n\t"
 328                                 "std %2, [%6 + 8]\n\t"
 329                                 "std %4, [%6 + 16]\n\t" : :
 330                                 "r"(f), "r"(a1), "r"(a2), "r"(a3), "r"(a4), "r"(a5),
 331                                 "r" (&ccall_info.func));
 332                 }
 333
 334                 /* Init receive/complete mapping, plus fire the IPI's off. */
 335                 {
 336                         cpumask_t mask;
 337                         register int i;
 338
 339                         mask = cpumask_of_cpu(hard_smp4d_processor_id());
 340                         cpus_andnot(mask, cpu_present_map, mask);
 341                         for(i = 0; i <= high; i++) {
 342                                 if (cpu_isset(i, mask)) {
 343                                         ccall_info.processors_in[i] = 0;
 344                                         ccall_info.processors_out[i] = 0;
 345                                         sun4d_send_ipi(i, IRQ_CROSS_CALL);
 346                                 }
 347                         }
 348                 }
 349
 350                 {
 351                         register int i;
 352
 353                         i = 0;
 354                         do {
 355                                 while(!ccall_info.processors_in[i])
 356                                         barrier();
 357                         } while(++i <= high);
 358
 359                         i = 0;
 360                         do {
 361                                 while(!ccall_info.processors_out[i])
 362                                         barrier();
 363                         } while(++i <= high);
 364                 }
 365
 366                 spin_unlock_irqrestore(&cross_call_lock, flags);
 367         }
 368 }
 369
 370 /* Running cross calls. */
 371 void smp4d_cross_call_irq(void)
 372 {
 373         int i = hard_smp4d_processor_id();
 374
 375         ccall_info.processors_in[i] = 1;
 376         ccall_info.func(ccall_info.arg1, ccall_info.arg2, ccall_info.arg3,
 377                         ccall_info.arg4, ccall_info.arg5);
 378         ccall_info.processors_out[i] = 1;
 379 }
 380
 381 static int smp4d_stop_cpu_sender;
 382
 383 static void smp4d_stop_cpu(void)
 384 {
 385         int me = hard_smp4d_processor_id();
 386
 387         if (me != smp4d_stop_cpu_sender)
 388                 while(1) barrier();
 389 }
 390
 391 /* Cross calls, in order to work efficiently and atomically do all
 392  * the message passing work themselves, only stopcpu and reschedule
 393  * messages come through here.
 394  */
 395 void smp4d_message_pass(int target, int msg, unsigned long data, int wait)
 396 {
 397         int me = hard_smp4d_processor_id();
 398
 399         SMP_PRINTK(("smp4d_message_pass %d %d %08lx %d\n", target, msg, data, wait));
 400         if (msg == MSG_STOP_CPU && target == MSG_ALL_BUT_SELF) {
 401                 unsigned long flags;
 402                 static spinlock_t stop_cpu_lock = SPIN_LOCK_UNLOCKED;
 403                 spin_lock_irqsave(&stop_cpu_lock, flags);
 404                 smp4d_stop_cpu_sender = me;
 405                 smp4d_cross_call((smpfunc_t)smp4d_stop_cpu, 0, 0, 0, 0, 0);
 406                 spin_unlock_irqrestore(&stop_cpu_lock, flags);
 407         }
 408         printk("Yeeee, trying to send SMP msg(%d) to %d on cpu %d\n", msg, target, me);
 409         panic("Bogon SMP message pass.");
 410 }
 411
 412 void smp4d_percpu_timer_interrupt(struct pt_regs *regs)
 413 {
 414         int cpu = hard_smp4d_processor_id();
 415         static int cpu_tick[NR_CPUS];
 416         static char led_mask[] = { 0xe, 0xd, 0xb, 0x7, 0xb, 0xd };
 417
 418         bw_get_prof_limit(cpu);
 419         bw_clear_intr_mask(0, 1);       /* INTR_TABLE[0] & 1 is Profile IRQ */
 420
 421         cpu_tick[cpu]++;
 422         if (!(cpu_tick[cpu] & 15)) {
 423                 if (cpu_tick[cpu] == 0x60)
 424                         cpu_tick[cpu] = 0;
 425                 cpu_leds[cpu] = led_mask[cpu_tick[cpu] >> 4];
 426                 show_leds(cpu);
 427         }
 428
 429         profile_tick(CPU_PROFILING, regs);
 430
 431         if(!--prof_counter(cpu)) {
 432                 int user = user_mode(regs);
 433
 434                 irq_enter();
 435                 update_process_times(user);
 436                 irq_exit();
 437
 438                 prof_counter(cpu) = prof_multiplier(cpu);
 439         }
 440 }
 441
 442 extern unsigned int lvl14_resolution;
 443
 444 static void __init smp_setup_percpu_timer(void)
 445 {
 446         int cpu = hard_smp4d_processor_id();
 447
 448         prof_counter(cpu) = prof_multiplier(cpu) = 1;
 449         load_profile_irq(cpu, lvl14_resolution);
 450 }
 451
 452 void __init smp4d_blackbox_id(unsigned *addr)
 453 {
 454         int rd = *addr & 0x3e000000;
 455
 456         addr[0] = 0xc0800800 | rd;              /* lda [%g0] ASI_M_VIKING_TMP1, reg */
 457         addr[1] = 0x01000000;                   /* nop */
 458         addr[2] = 0x01000000;                   /* nop */
 459 }
 460
 461 void __init smp4d_blackbox_current(unsigned *addr)
 462 {
 463         /* We have a nice Linux current register :) */
 464         int rd = addr[1] & 0x3e000000;
 465
 466         addr[0] = 0x10800006;                   /* b .+24 */
 467         addr[1] = 0xc0800820 | rd;              /* lda [%g0] ASI_M_VIKING_TMP2, reg */
 468 }
 469
 470 void __init sun4d_init_smp(void)
 471 {
 472         int i;
 473         extern unsigned int patchme_store_new_current[];
 474         extern unsigned int t_nmi[], linux_trap_ipi15_sun4d[], linux_trap_ipi15_sun4m[];
 475
 476         /* Store current into Linux current register :) */
 477         __asm__ __volatile__("sta %%g6, [%%g0] %0" : : "i"(ASI_M_VIKING_TMP2));
 478
 479         /* Patch switch_to */
 480         patchme_store_new_current[0] = (patchme_store_new_current[0] & 0x3e000000) | 0xc0a00820;
 481
 482         /* Patch ipi15 trap table */
 483         t_nmi[1] = t_nmi[1] + (linux_trap_ipi15_sun4d - linux_trap_ipi15_sun4m);
 484
 485         /* And set btfixup... */
 486         BTFIXUPSET_BLACKBOX(hard_smp_processor_id, smp4d_blackbox_id);
 487         BTFIXUPSET_BLACKBOX(load_current, smp4d_blackbox_current);
 488         BTFIXUPSET_CALL(smp_cross_call, smp4d_cross_call, BTFIXUPCALL_NORM);
 489         BTFIXUPSET_CALL(smp_message_pass, smp4d_message_pass, BTFIXUPCALL_NORM);
 490         BTFIXUPSET_CALL(__hard_smp_processor_id, __smp4d_processor_id, BTFIXUPCALL_NORM);
 491
 492         for (i = 0; i < NR_CPUS; i++) {
 493                 ccall_info.processors_in[i] = 1;
 494                 ccall_info.processors_out[i] = 1;
 495         }
 496 }