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
  23 #include <asm/ptrace.h>
  24 #include <asm/atomic.h>
  25
  26 #include <asm/delay.h>
  27 #include <asm/irq.h>
  28 #include <asm/page.h>
  29 #include <asm/pgalloc.h>
  30 #include <asm/pgtable.h>
  31 #include <asm/oplib.h>
  32 #include <asm/hardirq.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                         kernel_thread(start_secondary, NULL, CLONE_IDLETASK);
 205
 206                         cpucount++;
 207
 208                         p = prev_task(&init_task);
 209
 210                         init_idle(p, i);
 211
 212                         current_set[i] = p->thread_info;
 213
 214                         unhash_process(p);
 215
 216                         for (no = 0; !cpu_find_by_instance(no, NULL, &mid)
 217                                      && mid != i; no++) ;
 218
 219                         /*
 220                          * Initialize the contexts table
 221                          * Since the call to prom_startcpu() trashes the structure,
 222                          * we need to re-initialize it for each cpu
 223                          */
 224                         smp_penguin_ctable.which_io = 0;
 225                         smp_penguin_ctable.phys_addr = (unsigned int) srmmu_ctx_table_phys;
 226                         smp_penguin_ctable.reg_size = 0;
 227
 228                         /* whirrr, whirrr, whirrrrrrrrr... */
 229                         SMP_PRINTK(("Starting CPU %d at %p task %d node %08x\n", i, entry, cpucount, cpu_data(no).prom_node));
 230                         local_flush_cache_all();
 231                         prom_startcpu(cpu_data(no).prom_node,
 232                                       &smp_penguin_ctable, 0, (char *)entry);
 233
 234                         SMP_PRINTK(("prom_startcpu returned :)\n"));
 235
 236                         /* wheee... it's going... */
 237                         for(timeout = 0; timeout < 10000; timeout++) {
 238                                 if(cpu_callin_map[i])
 239                                         break;
 240                                 udelay(200);
 241                         }
 242
 243                         if(cpu_callin_map[i]) {
 244                                 /* Another "Red Snapper". */
 245                                 __cpu_number_map[i] = cpucount;
 246                                 __cpu_logical_map[cpucount] = i;
 247                         } else {
 248                                 cpucount--;
 249                                 printk("Processor %d is stuck.\n", i);
 250                         }
 251                 }
 252                 if(!(cpu_callin_map[i])) {
 253                         cpu_clear(i, cpu_present_map);
 254                         __cpu_number_map[i] = -1;
 255                 }
 256         }
 257         local_flush_cache_all();
 258         if(cpucount == 0) {
 259                 printk("Error: only one Processor found.\n");
 260                 cpu_present_map = cpumask_of_cpu(hard_smp4d_processor_id());
 261         } else {
 262                 unsigned long bogosum = 0;
 263
 264                 for(i = 0; i < NR_CPUS; i++) {
 265                         if (cpu_isset(i, cpu_present_map)) {
 266                                 bogosum += cpu_data(i).udelay_val;
 267                                 smp_highest_cpu = i;
 268                         }
 269                 }
 270                 SMP_PRINTK(("Total of %d Processors activated (%lu.%02lu BogoMIPS).\n", cpucount + 1, bogosum/(500000/HZ), (bogosum/(5000/HZ))%100));
 271                 printk("Total of %d Processors activated (%lu.%02lu BogoMIPS).\n",
 272                        cpucount + 1,
 273                        bogosum/(500000/HZ),
 274                        (bogosum/(5000/HZ))%100);
 275                 smp_activated = 1;
 276                 smp_num_cpus = cpucount + 1;
 277         }
 278
 279         /* Free unneeded trap tables */
 280         ClearPageReserved(virt_to_page(trapbase_cpu1));
 281         set_page_count(virt_to_page(trapbase_cpu1), 1);
 282         free_page((unsigned long)trapbase_cpu1);
 283         totalram_pages++;
 284         num_physpages++;
 285
 286         ClearPageReserved(virt_to_page(trapbase_cpu2));
 287         set_page_count(virt_to_page(trapbase_cpu2), 1);
 288         free_page((unsigned long)trapbase_cpu2);
 289         totalram_pages++;
 290         num_physpages++;
 291
 292         ClearPageReserved(virt_to_page(trapbase_cpu3));
 293         set_page_count(virt_to_page(trapbase_cpu3), 1);
 294         free_page((unsigned long)trapbase_cpu3);
 295         totalram_pages++;
 296         num_physpages++;
 297
 298         /* Ok, they are spinning and ready to go. */
 299         smp_processors_ready = 1;
 300         sun4d_distribute_irqs();
 301 }
 302
 303 static struct smp_funcall {
 304         smpfunc_t func;
 305         unsigned long arg1;
 306         unsigned long arg2;
 307         unsigned long arg3;
 308         unsigned long arg4;
 309         unsigned long arg5;
 310         unsigned char processors_in[NR_CPUS];  /* Set when ipi entered. */
 311         unsigned char processors_out[NR_CPUS]; /* Set when ipi exited. */
 312 } ccall_info __attribute__((aligned(8)));
 313
 314 static spinlock_t cross_call_lock = SPIN_LOCK_UNLOCKED;
 315
 316 /* Cross calls must be serialized, at least currently. */
 317 void smp4d_cross_call(smpfunc_t func, unsigned long arg1, unsigned long arg2,
 318                     unsigned long arg3, unsigned long arg4, unsigned long arg5)
 319 {
 320         if(smp_processors_ready) {
 321                 register int high = smp_highest_cpu;
 322                 unsigned long flags;
 323
 324                 spin_lock_irqsave(&cross_call_lock, flags);
 325
 326                 {
 327                         /* If you make changes here, make sure gcc generates proper code... */
 328                         register smpfunc_t f asm("i0") = func;
 329                         register unsigned long a1 asm("i1") = arg1;
 330                         register unsigned long a2 asm("i2") = arg2;
 331                         register unsigned long a3 asm("i3") = arg3;
 332                         register unsigned long a4 asm("i4") = arg4;
 333                         register unsigned long a5 asm("i5") = arg5;
 334
 335                         __asm__ __volatile__(
 336                                 "std %0, [%6]\n\t"
 337                                 "std %2, [%6 + 8]\n\t"
 338                                 "std %4, [%6 + 16]\n\t" : :
 339                                 "r"(f), "r"(a1), "r"(a2), "r"(a3), "r"(a4), "r"(a5),
 340                                 "r" (&ccall_info.func));
 341                 }
 342
 343                 /* Init receive/complete mapping, plus fire the IPI's off. */
 344                 {
 345                         cpumask_t mask;
 346                         register int i;
 347
 348                         mask = cpumask_of_cpu(hard_smp4d_processor_id());
 349                         cpus_andnot(mask, cpu_present_map, mask);
 350                         for(i = 0; i <= high; i++) {
 351                                 if (cpu_isset(i, mask)) {
 352                                         ccall_info.processors_in[i] = 0;
 353                                         ccall_info.processors_out[i] = 0;
 354                                         sun4d_send_ipi(i, IRQ_CROSS_CALL);
 355                                 }
 356                         }
 357                 }
 358
 359                 {
 360                         register int i;
 361
 362                         i = 0;
 363                         do {
 364                                 while(!ccall_info.processors_in[i])
 365                                         barrier();
 366                         } while(++i <= high);
 367
 368                         i = 0;
 369                         do {
 370                                 while(!ccall_info.processors_out[i])
 371                                         barrier();
 372                         } while(++i <= high);
 373                 }
 374
 375                 spin_unlock_irqrestore(&cross_call_lock, flags);
 376         }
 377 }
 378
 379 /* Running cross calls. */
 380 void smp4d_cross_call_irq(void)
 381 {
 382         int i = hard_smp4d_processor_id();
 383
 384         ccall_info.processors_in[i] = 1;
 385         ccall_info.func(ccall_info.arg1, ccall_info.arg2, ccall_info.arg3,
 386                         ccall_info.arg4, ccall_info.arg5);
 387         ccall_info.processors_out[i] = 1;
 388 }
 389
 390 static int smp4d_stop_cpu_sender;
 391
 392 static void smp4d_stop_cpu(void)
 393 {
 394         int me = hard_smp4d_processor_id();
 395
 396         if (me != smp4d_stop_cpu_sender)
 397                 while(1) barrier();
 398 }
 399
 400 /* Cross calls, in order to work efficiently and atomically do all
 401  * the message passing work themselves, only stopcpu and reschedule
 402  * messages come through here.
 403  */
 404 void smp4d_message_pass(int target, int msg, unsigned long data, int wait)
 405 {
 406         int me = hard_smp4d_processor_id();
 407
 408         SMP_PRINTK(("smp4d_message_pass %d %d %08lx %d\n", target, msg, data, wait));
 409         if (msg == MSG_STOP_CPU && target == MSG_ALL_BUT_SELF) {
 410                 unsigned long flags;
 411                 static spinlock_t stop_cpu_lock = SPIN_LOCK_UNLOCKED;
 412                 spin_lock_irqsave(&stop_cpu_lock, flags);
 413                 smp4d_stop_cpu_sender = me;
 414                 smp4d_cross_call((smpfunc_t)smp4d_stop_cpu, 0, 0, 0, 0, 0);
 415                 spin_unlock_irqrestore(&stop_cpu_lock, flags);
 416         }
 417         printk("Yeeee, trying to send SMP msg(%d) to %d on cpu %d\n", msg, target, me);
 418         panic("Bogon SMP message pass.");
 419 }
 420
 421 extern void sparc_do_profile(unsigned long pc, unsigned long o7);
 422
 423 void smp4d_percpu_timer_interrupt(struct pt_regs *regs)
 424 {
 425         int cpu = hard_smp4d_processor_id();
 426         static int cpu_tick[NR_CPUS];
 427         static char led_mask[] = { 0xe, 0xd, 0xb, 0x7, 0xb, 0xd };
 428
 429         bw_get_prof_limit(cpu);
 430         bw_clear_intr_mask(0, 1);       /* INTR_TABLE[0] & 1 is Profile IRQ */
 431
 432         cpu_tick[cpu]++;
 433         if (!(cpu_tick[cpu] & 15)) {
 434                 if (cpu_tick[cpu] == 0x60)
 435                         cpu_tick[cpu] = 0;
 436                 cpu_leds[cpu] = led_mask[cpu_tick[cpu] >> 4];
 437                 show_leds(cpu);
 438         }
 439
 440         if(!user_mode(regs))
 441                 sparc_do_profile(regs->pc, regs->u_regs[UREG_RETPC]);
 442
 443         if(!--prof_counter(cpu)) {
 444                 int user = user_mode(regs);
 445
 446                 irq_enter();
 447                 update_process_times(user);
 448                 irq_exit();
 449
 450                 prof_counter(cpu) = prof_multiplier(cpu);
 451         }
 452 }
 453
 454 extern unsigned int lvl14_resolution;
 455
 456 static void __init smp_setup_percpu_timer(void)
 457 {
 458         int cpu = hard_smp4d_processor_id();
 459
 460         prof_counter(cpu) = prof_multiplier(cpu) = 1;
 461         load_profile_irq(cpu, lvl14_resolution);
 462 }
 463
 464 void __init smp4d_blackbox_id(unsigned *addr)
 465 {
 466         int rd = *addr & 0x3e000000;
 467
 468         addr[0] = 0xc0800800 | rd;              /* lda [%g0] ASI_M_VIKING_TMP1, reg */
 469         addr[1] = 0x01000000;                   /* nop */
 470         addr[2] = 0x01000000;                   /* nop */
 471 }
 472
 473 void __init smp4d_blackbox_current(unsigned *addr)
 474 {
 475         /* We have a nice Linux current register :) */
 476         int rd = addr[1] & 0x3e000000;
 477
 478         addr[0] = 0x10800006;                   /* b .+24 */
 479         addr[1] = 0xc0800820 | rd;              /* lda [%g0] ASI_M_VIKING_TMP2, reg */
 480 }
 481
 482 void __init sun4d_init_smp(void)
 483 {
 484         int i;
 485         extern unsigned int patchme_store_new_current[];
 486         extern unsigned int t_nmi[], linux_trap_ipi15_sun4d[], linux_trap_ipi15_sun4m[];
 487
 488         /* Store current into Linux current register :) */
 489         __asm__ __volatile__("sta %%g6, [%%g0] %0" : : "i"(ASI_M_VIKING_TMP2));
 490
 491         /* Patch switch_to */
 492         patchme_store_new_current[0] = (patchme_store_new_current[0] & 0x3e000000) | 0xc0a00820;
 493
 494         /* Patch ipi15 trap table */
 495         t_nmi[1] = t_nmi[1] + (linux_trap_ipi15_sun4d - linux_trap_ipi15_sun4m);
 496
 497         /* And set btfixup... */
 498         BTFIXUPSET_BLACKBOX(hard_smp_processor_id, smp4d_blackbox_id);
 499         BTFIXUPSET_BLACKBOX(load_current, smp4d_blackbox_current);
 500         BTFIXUPSET_CALL(smp_cross_call, smp4d_cross_call, BTFIXUPCALL_NORM);
 501         BTFIXUPSET_CALL(smp_message_pass, smp4d_message_pass, BTFIXUPCALL_NORM);
 502         BTFIXUPSET_CALL(__hard_smp_processor_id, __smp4d_processor_id, BTFIXUPCALL_NORM);
 503
 504         for (i = 0; i < NR_CPUS; i++) {
 505                 ccall_info.processors_in[i] = 1;
 506                 ccall_info.processors_out[i] = 1;
 507         }
 508 }