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                              : : "r" (&current_set[cpuid])
 126                              : "memory" /* paranoid */);
 127
 128         cpu_leds[cpuid] = 0x9;
 129         show_leds(cpuid);
 130
 131         /* Attach to the address space of init_task. */
 132         atomic_inc(&init_mm.mm_count);
 133         current->active_mm = &init_mm;
 134
 135         local_flush_cache_all();
 136         local_flush_tlb_all();
 137
 138         local_irq_enable();     /* We don't allow PIL 14 yet */
 139
 140         while(!smp_commenced)
 141                 barrier();
 142
 143         spin_lock_irqsave(&sun4d_imsk_lock, flags);
 144         cc_set_imsk(cc_get_imsk() & ~0x4000); /* Allow PIL 14 as well */
 145         spin_unlock_irqrestore(&sun4d_imsk_lock, flags);
 146 }
 147
 148 extern void init_IRQ(void);
 149 extern void cpu_panic(void);
 150
 151 /*
 152  *      Cycle through the processors asking the PROM to start each one.
 153  */
 154
 155 extern struct linux_prom_registers smp_penguin_ctable;
 156 extern unsigned long trapbase_cpu1[];
 157 extern unsigned long trapbase_cpu2[];
 158 extern unsigned long trapbase_cpu3[];
 159
 160 void __init smp4d_boot_cpus(void)
 161 {
 162         int cpucount = 0;
 163         int i, mid;
 164
 165         printk("Entering SMP Mode...\n");
 166
 167         if (boot_cpu_id)
 168                 current_set[0] = NULL;
 169
 170         local_irq_enable();
 171         cpus_clear(cpu_present_map);
 172
 173         /* XXX This whole thing has to go.  See sparc64. */
 174         for (i = 0; !cpu_find_by_instance(i, NULL, &mid); i++)
 175                 cpu_set(mid, cpu_present_map);
 176         SMP_PRINTK(("cpu_present_map %08lx\n", cpus_addr(cpu_present_map)[0]));
 177         for(i=0; i < NR_CPUS; i++)
 178                 __cpu_number_map[i] = -1;
 179         for(i=0; i < NR_CPUS; i++)
 180                 __cpu_logical_map[i] = -1;
 181         __cpu_number_map[boot_cpu_id] = 0;
 182         __cpu_logical_map[0] = boot_cpu_id;
 183         current_thread_info()->cpu = boot_cpu_id;
 184         smp_store_cpu_info(boot_cpu_id);
 185         smp_setup_percpu_timer();
 186         local_flush_cache_all();
 187         if (cpu_find_by_instance(1, NULL, NULL))
 188                 return;  /* Not an MP box. */
 189         SMP_PRINTK(("Iterating over CPUs\n"));
 190         for(i = 0; i < NR_CPUS; i++) {
 191                 if(i == boot_cpu_id)
 192                         continue;
 193
 194                 if (cpu_isset(i, cpu_present_map)) {
 195                         extern unsigned long sun4d_cpu_startup;
 196                         unsigned long *entry = &sun4d_cpu_startup;
 197                         struct task_struct *p;
 198                         int timeout;
 199                         int no;
 200
 201                         /* Cook up an idler for this guy. */
 202                         p = fork_idle(i);
 203                         cpucount++;
 204                         current_set[i] = p->thread_info;
 205                         for (no = 0; !cpu_find_by_instance(no, NULL, &mid)
 206                                      && mid != i; no++) ;
 207
 208                         /*
 209                          * Initialize the contexts table
 210                          * Since the call to prom_startcpu() trashes the structure,
 211                          * we need to re-initialize it for each cpu
 212                          */
 213                         smp_penguin_ctable.which_io = 0;
 214                         smp_penguin_ctable.phys_addr = (unsigned int) srmmu_ctx_table_phys;
 215                         smp_penguin_ctable.reg_size = 0;
 216
 217                         /* whirrr, whirrr, whirrrrrrrrr... */
 218                         SMP_PRINTK(("Starting CPU %d at %p task %d node %08x\n", i, entry, cpucount, cpu_data(no).prom_node));
 219                         local_flush_cache_all();
 220                         prom_startcpu(cpu_data(no).prom_node,
 221                                       &smp_penguin_ctable, 0, (char *)entry);
 222
 223                         SMP_PRINTK(("prom_startcpu returned :)\n"));
 224
 225                         /* wheee... it's going... */
 226                         for(timeout = 0; timeout < 10000; timeout++) {
 227                                 if(cpu_callin_map[i])
 228                                         break;
 229                                 udelay(200);
 230                         }
 231
 232                         if(cpu_callin_map[i]) {
 233                                 /* Another "Red Snapper". */
 234                                 __cpu_number_map[i] = cpucount;
 235                                 __cpu_logical_map[cpucount] = i;
 236                         } else {
 237                                 cpucount--;
 238                                 printk("Processor %d is stuck.\n", i);
 239                         }
 240                 }
 241                 if(!(cpu_callin_map[i])) {
 242                         cpu_clear(i, cpu_present_map);
 243                         __cpu_number_map[i] = -1;
 244                 }
 245         }
 246         local_flush_cache_all();
 247         if(cpucount == 0) {
 248                 printk("Error: only one Processor found.\n");
 249                 cpu_present_map = cpumask_of_cpu(hard_smp4d_processor_id());
 250         } else {
 251                 unsigned long bogosum = 0;
 252
 253                 for(i = 0; i < NR_CPUS; i++) {
 254                         if (cpu_isset(i, cpu_present_map)) {
 255                                 bogosum += cpu_data(i).udelay_val;
 256                                 smp_highest_cpu = i;
 257                         }
 258                 }
 259                 SMP_PRINTK(("Total of %d Processors activated (%lu.%02lu BogoMIPS).\n", cpucount + 1, bogosum/(500000/HZ), (bogosum/(5000/HZ))%100));
 260                 printk("Total of %d Processors activated (%lu.%02lu BogoMIPS).\n",
 261                        cpucount + 1,
 262                        bogosum/(500000/HZ),
 263                        (bogosum/(5000/HZ))%100);
 264                 smp_activated = 1;
 265                 smp_num_cpus = cpucount + 1;
 266         }
 267
 268         /* Free unneeded trap tables */
 269         ClearPageReserved(virt_to_page(trapbase_cpu1));
 270         set_page_count(virt_to_page(trapbase_cpu1), 1);
 271         free_page((unsigned long)trapbase_cpu1);
 272         totalram_pages++;
 273         num_physpages++;
 274
 275         ClearPageReserved(virt_to_page(trapbase_cpu2));
 276         set_page_count(virt_to_page(trapbase_cpu2), 1);
 277         free_page((unsigned long)trapbase_cpu2);
 278         totalram_pages++;
 279         num_physpages++;
 280
 281         ClearPageReserved(virt_to_page(trapbase_cpu3));
 282         set_page_count(virt_to_page(trapbase_cpu3), 1);
 283         free_page((unsigned long)trapbase_cpu3);
 284         totalram_pages++;
 285         num_physpages++;
 286
 287         /* Ok, they are spinning and ready to go. */
 288         smp_processors_ready = 1;
 289         sun4d_distribute_irqs();
 290 }
 291
 292 static struct smp_funcall {
 293         smpfunc_t func;
 294         unsigned long arg1;
 295         unsigned long arg2;
 296         unsigned long arg3;
 297         unsigned long arg4;
 298         unsigned long arg5;
 299         unsigned char processors_in[NR_CPUS];  /* Set when ipi entered. */
 300         unsigned char processors_out[NR_CPUS]; /* Set when ipi exited. */
 301 } ccall_info __attribute__((aligned(8)));
 302
 303 static DEFINE_SPINLOCK(cross_call_lock);
 304
 305 /* Cross calls must be serialized, at least currently. */
 306 void smp4d_cross_call(smpfunc_t func, unsigned long arg1, unsigned long arg2,
 307                     unsigned long arg3, unsigned long arg4, unsigned long arg5)
 308 {
 309         if(smp_processors_ready) {
 310                 register int high = smp_highest_cpu;
 311                 unsigned long flags;
 312
 313                 spin_lock_irqsave(&cross_call_lock, flags);
 314
 315                 {
 316                         /* If you make changes here, make sure gcc generates proper code... */
 317                         register smpfunc_t f asm("i0") = func;
 318                         register unsigned long a1 asm("i1") = arg1;
 319                         register unsigned long a2 asm("i2") = arg2;
 320                         register unsigned long a3 asm("i3") = arg3;
 321                         register unsigned long a4 asm("i4") = arg4;
 322                         register unsigned long a5 asm("i5") = arg5;
 323
 324                         __asm__ __volatile__(
 325                                 "std %0, [%6]\n\t"
 326                                 "std %2, [%6 + 8]\n\t"
 327                                 "std %4, [%6 + 16]\n\t" : :
 328                                 "r"(f), "r"(a1), "r"(a2), "r"(a3), "r"(a4), "r"(a5),
 329                                 "r" (&ccall_info.func));
 330                 }
 331
 332                 /* Init receive/complete mapping, plus fire the IPI's off. */
 333                 {
 334                         cpumask_t mask;
 335                         register int i;
 336
 337                         mask = cpumask_of_cpu(hard_smp4d_processor_id());
 338                         cpus_andnot(mask, cpu_present_map, mask);
 339                         for(i = 0; i <= high; i++) {
 340                                 if (cpu_isset(i, mask)) {
 341                                         ccall_info.processors_in[i] = 0;
 342                                         ccall_info.processors_out[i] = 0;
 343                                         sun4d_send_ipi(i, IRQ_CROSS_CALL);
 344                                 }
 345                         }
 346                 }
 347
 348                 {
 349                         register int i;
 350
 351                         i = 0;
 352                         do {
 353                                 while(!ccall_info.processors_in[i])
 354                                         barrier();
 355                         } while(++i <= high);
 356
 357                         i = 0;
 358                         do {
 359                                 while(!ccall_info.processors_out[i])
 360                                         barrier();
 361                         } while(++i <= high);
 362                 }
 363
 364                 spin_unlock_irqrestore(&cross_call_lock, flags);
 365         }
 366 }
 367
 368 /* Running cross calls. */
 369 void smp4d_cross_call_irq(void)
 370 {
 371         int i = hard_smp4d_processor_id();
 372
 373         ccall_info.processors_in[i] = 1;
 374         ccall_info.func(ccall_info.arg1, ccall_info.arg2, ccall_info.arg3,
 375                         ccall_info.arg4, ccall_info.arg5);
 376         ccall_info.processors_out[i] = 1;
 377 }
 378
 379 static int smp4d_stop_cpu_sender;
 380
 381 static void smp4d_stop_cpu(void)
 382 {
 383         int me = hard_smp4d_processor_id();
 384
 385         if (me != smp4d_stop_cpu_sender)
 386                 while(1) barrier();
 387 }
 388
 389 /* Cross calls, in order to work efficiently and atomically do all
 390  * the message passing work themselves, only stopcpu and reschedule
 391  * messages come through here.
 392  */
 393 void smp4d_message_pass(int target, int msg, unsigned long data, int wait)
 394 {
 395         int me = hard_smp4d_processor_id();
 396
 397         SMP_PRINTK(("smp4d_message_pass %d %d %08lx %d\n", target, msg, data, wait));
 398         if (msg == MSG_STOP_CPU && target == MSG_ALL_BUT_SELF) {
 399                 unsigned long flags;
 400                 static DEFINE_SPINLOCK(stop_cpu_lock);
 401                 spin_lock_irqsave(&stop_cpu_lock, flags);
 402                 smp4d_stop_cpu_sender = me;
 403                 smp4d_cross_call((smpfunc_t)smp4d_stop_cpu, 0, 0, 0, 0, 0);
 404                 spin_unlock_irqrestore(&stop_cpu_lock, flags);
 405         }
 406         printk("Yeeee, trying to send SMP msg(%d) to %d on cpu %d\n", msg, target, me);
 407         panic("Bogon SMP message pass.");
 408 }
 409
 410 void smp4d_percpu_timer_interrupt(struct pt_regs *regs)
 411 {
 412         int cpu = hard_smp4d_processor_id();
 413         static int cpu_tick[NR_CPUS];
 414         static char led_mask[] = { 0xe, 0xd, 0xb, 0x7, 0xb, 0xd };
 415
 416         bw_get_prof_limit(cpu);
 417         bw_clear_intr_mask(0, 1);       /* INTR_TABLE[0] & 1 is Profile IRQ */
 418
 419         cpu_tick[cpu]++;
 420         if (!(cpu_tick[cpu] & 15)) {
 421                 if (cpu_tick[cpu] == 0x60)
 422                         cpu_tick[cpu] = 0;
 423                 cpu_leds[cpu] = led_mask[cpu_tick[cpu] >> 4];
 424                 show_leds(cpu);
 425         }
 426
 427         profile_tick(CPU_PROFILING, regs);
 428
 429         if(!--prof_counter(cpu)) {
 430                 int user = user_mode(regs);
 431
 432                 irq_enter();
 433                 update_process_times(user);
 434                 irq_exit();
 435
 436                 prof_counter(cpu) = prof_multiplier(cpu);
 437         }
 438 }
 439
 440 extern unsigned int lvl14_resolution;
 441
 442 static void __init smp_setup_percpu_timer(void)
 443 {
 444         int cpu = hard_smp4d_processor_id();
 445
 446         prof_counter(cpu) = prof_multiplier(cpu) = 1;
 447         load_profile_irq(cpu, lvl14_resolution);
 448 }
 449
 450 void __init smp4d_blackbox_id(unsigned *addr)
 451 {
 452         int rd = *addr & 0x3e000000;
 453
 454         addr[0] = 0xc0800800 | rd;              /* lda [%g0] ASI_M_VIKING_TMP1, reg */
 455         addr[1] = 0x01000000;                   /* nop */
 456         addr[2] = 0x01000000;                   /* nop */
 457 }
 458
 459 void __init smp4d_blackbox_current(unsigned *addr)
 460 {
 461         int rd = *addr & 0x3e000000;
 462
 463         addr[0] = 0xc0800800 | rd;              /* lda [%g0] ASI_M_VIKING_TMP1, reg */
 464         addr[2] = 0x81282002 | rd | (rd >> 11); /* sll reg, 2, reg */
 465         addr[4] = 0x01000000;                   /* nop */
 466 }
 467
 468 void __init sun4d_init_smp(void)
 469 {
 470         int i;
 471         extern unsigned int t_nmi[], linux_trap_ipi15_sun4d[], linux_trap_ipi15_sun4m[];
 472
 473         /* Patch ipi15 trap table */
 474         t_nmi[1] = t_nmi[1] + (linux_trap_ipi15_sun4d - linux_trap_ipi15_sun4m);
 475
 476         /* And set btfixup... */
 477         BTFIXUPSET_BLACKBOX(hard_smp_processor_id, smp4d_blackbox_id);
 478         BTFIXUPSET_BLACKBOX(load_current, smp4d_blackbox_current);
 479         BTFIXUPSET_CALL(smp_cross_call, smp4d_cross_call, BTFIXUPCALL_NORM);
 480         BTFIXUPSET_CALL(smp_message_pass, smp4d_message_pass, BTFIXUPCALL_NORM);
 481         BTFIXUPSET_CALL(__hard_smp_processor_id, __smp4d_processor_id, BTFIXUPCALL_NORM);
 482
 483         for (i = 0; i < NR_CPUS; i++) {
 484                 ccall_info.processors_in[i] = 1;
 485                 ccall_info.processors_out[i] = 1;
 486         }
 487 }