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
  22 #include <asm/ptrace.h>
  23 #include <asm/atomic.h>
  24
  25 #include <asm/delay.h>
  26 #include <asm/irq.h>
  27 #include <asm/page.h>
  28 #include <asm/pgalloc.h>
  29 #include <asm/pgtable.h>
  30 #include <asm/oplib.h>
  31 #include <asm/hardirq.h>
  32 #include <asm/softirq.h>
  33 #include <asm/sbus.h>
  34 #include <asm/sbi.h>
  35
  36 #define __KERNEL_SYSCALLS__
  37 #include <linux/unistd.h>
  38
  39 #define IRQ_CROSS_CALL          15
  40
  41 extern ctxd_t *srmmu_ctx_table_phys;
  42 extern int linux_num_cpus;
  43
  44 extern void calibrate_delay(void);
  45
  46 extern struct task_struct *current_set[NR_CPUS];
  47 extern volatile int smp_processors_ready;
  48 extern unsigned long cpu_present_map;
  49 extern int smp_num_cpus;
  50 static int smp_highest_cpu;
  51 extern int smp_threads_ready;
  52 extern unsigned char mid_xlate[NR_CPUS];
  53 extern volatile unsigned long cpu_callin_map[NR_CPUS];
  54 extern unsigned long smp_proc_in_lock[NR_CPUS];
  55 extern struct cpuinfo_sparc cpu_data[NR_CPUS];
  56 extern unsigned long cpu_offset[NR_CPUS];
  57 extern unsigned char boot_cpu_id;
  58 extern int smp_activated;
  59 extern volatile int __cpu_number_map[NR_CPUS];
  60 extern volatile int __cpu_logical_map[NR_CPUS];
  61 extern volatile unsigned long ipi_count;
  62 extern volatile int smp_process_available;
  63 extern volatile int smp_commenced;
  64 extern int __smp4d_processor_id(void);
  65
  66 extern unsigned long totalram_pages;
  67
  68 /* #define SMP_DEBUG */
  69
  70 #ifdef SMP_DEBUG
  71 #define SMP_PRINTK(x)   printk x
  72 #else
  73 #define SMP_PRINTK(x)
  74 #endif
  75
  76 static inline unsigned long swap(volatile unsigned long *ptr, unsigned long val)
  77 {
  78         __asm__ __volatile__("swap [%1], %0\n\t" :
  79                              "=&r" (val), "=&r" (ptr) :
  80                              "0" (val), "1" (ptr));
  81         return val;
  82 }
  83
  84 static void smp_setup_percpu_timer(void);
  85 extern void cpu_probe(void);
  86 extern void sun4d_distribute_irqs(void);
  87
  88 void __init smp4d_callin(void)
  89 {
  90         int cpuid = hard_smp4d_processor_id();
  91         extern spinlock_t sun4d_imsk_lock;
  92         unsigned long flags;
  93
  94         /* Show we are alive */
  95         cpu_leds[cpuid] = 0x6;
  96         show_leds(cpuid);
  97
  98         /* Enable level15 interrupt, disable level14 interrupt for now */
  99         cc_set_imsk((cc_get_imsk() & ~0x8000) | 0x4000);
 100
 101         local_flush_cache_all();
 102         local_flush_tlb_all();
 103
 104         /*
 105          * Unblock the master CPU _only_ when the scheduler state
 106          * of all secondary CPUs will be up-to-date, so after
 107          * the SMP initialization the master will be just allowed
 108          * to call the scheduler code.
 109          */
 110         init_idle();
 111
 112         /* Get our local ticker going. */
 113         smp_setup_percpu_timer();
 114
 115         calibrate_delay();
 116         smp_store_cpu_info(cpuid);
 117         local_flush_cache_all();
 118         local_flush_tlb_all();
 119
 120         /* Allow master to continue. */
 121         swap((unsigned long *)&cpu_callin_map[cpuid], 1);
 122         local_flush_cache_all();
 123         local_flush_tlb_all();
 124
 125         cpu_probe();
 126
 127         while((unsigned long)current_set[cpuid] < PAGE_OFFSET)
 128                 barrier();
 129
 130         while(current_set[cpuid]->processor != cpuid)
 131                 barrier();
 132
 133         /* Fix idle thread fields. */
 134         __asm__ __volatile__("ld [%0], %%g6\n\t"
 135                              : : "r" (&current_set[cpuid])
 136                              : "memory" /* paranoid */);
 137
 138         cpu_leds[cpuid] = 0x9;
 139         show_leds(cpuid);
 140
 141         /* Attach to the address space of init_task. */
 142         atomic_inc(&init_mm.mm_count);
 143         current->active_mm = &init_mm;
 144
 145         local_flush_cache_all();
 146         local_flush_tlb_all();
 147
 148         __sti();        /* We don't allow PIL 14 yet */
 149
 150         while(!smp_commenced)
 151                 barrier();
 152
 153         spin_lock_irqsave(&sun4d_imsk_lock, flags);
 154         cc_set_imsk(cc_get_imsk() & ~0x4000); /* Allow PIL 14 as well */
 155         spin_unlock_irqrestore(&sun4d_imsk_lock, flags);
 156 }
 157
 158 extern int cpu_idle(void *unused);
 159 extern void init_IRQ(void);
 160 extern void cpu_panic(void);
 161 extern int start_secondary(void *unused);
 162
 163 /*
 164  *      Cycle through the processors asking the PROM to start each one.
 165  */
 166
 167 extern struct prom_cpuinfo linux_cpus[NR_CPUS];
 168 extern struct linux_prom_registers smp_penguin_ctable;
 169 extern unsigned long trapbase_cpu1[];
 170 extern unsigned long trapbase_cpu2[];
 171 extern unsigned long trapbase_cpu3[];
 172
 173 void __init smp4d_boot_cpus(void)
 174 {
 175         int cpucount = 0;
 176         int i = 0;
 177
 178         printk("Entering SMP Mode...\n");
 179
 180         for (i = 0; i < NR_CPUS; i++)
 181                 cpu_offset[i] = (char *)&cpu_data[i] - (char *)&cpu_data;
 182
 183         if (boot_cpu_id)
 184                 current_set[0] = NULL;
 185
 186         __sti();
 187         cpu_present_map = 0;
 188         for(i=0; i < linux_num_cpus; i++)
 189                 cpu_present_map |= (1<<linux_cpus[i].mid);
 190         SMP_PRINTK(("cpu_present_map %08lx\n", cpu_present_map));
 191         for(i=0; i < NR_CPUS; i++)
 192                 __cpu_number_map[i] = -1;
 193         for(i=0; i < NR_CPUS; i++)
 194                 __cpu_logical_map[i] = -1;
 195         for(i=0; i < NR_CPUS; i++)
 196                 mid_xlate[i] = i;
 197         __cpu_number_map[boot_cpu_id] = 0;
 198         __cpu_logical_map[0] = boot_cpu_id;
 199         current->processor = boot_cpu_id;
 200         smp_store_cpu_info(boot_cpu_id);
 201         smp_setup_percpu_timer();
 202         init_idle();
 203         local_flush_cache_all();
 204         if(linux_num_cpus == 1)
 205                 return;  /* Not an MP box. */
 206         SMP_PRINTK(("Iterating over CPUs\n"));
 207         for(i = 0; i < NR_CPUS; i++) {
 208                 if(i == boot_cpu_id)
 209                         continue;
 210
 211                 if(cpu_present_map & (1 << i)) {
 212                         extern unsigned long sun4d_cpu_startup;
 213                         unsigned long *entry = &sun4d_cpu_startup;
 214                         struct task_struct *p;
 215                         int timeout;
 216                         int no;
 217
 218                         /* Cook up an idler for this guy. */
 219                         kernel_thread(start_secondary, NULL, CLONE_PID);
 220
 221                         cpucount++;
 222
 223                         p = init_task.prev_task;
 224                         init_tasks[i] = p;
 225
 226                         p->processor = i;
 227                         p->cpus_runnable = 1 << i; /* we schedule the first task manually */
 228
 229                         current_set[i] = p;
 230
 231                         del_from_runqueue(p);
 232                         unhash_process(p);
 233
 234                         for (no = 0; no < linux_num_cpus; no++)
 235                                 if (linux_cpus[no].mid == i)
 236                                         break;
 237
 238                         /*
 239                          * Initialize the contexts table
 240                          * Since the call to prom_startcpu() trashes the structure,
 241                          * we need to re-initialize it for each cpu
 242                          */
 243                         smp_penguin_ctable.which_io = 0;
 244                         smp_penguin_ctable.phys_addr = (unsigned int) srmmu_ctx_table_phys;
 245                         smp_penguin_ctable.reg_size = 0;
 246
 247                         /* whirrr, whirrr, whirrrrrrrrr... */
 248                         SMP_PRINTK(("Starting CPU %d at %p task %d node %08x\n", i, entry, cpucount, linux_cpus[no].prom_node));
 249                         local_flush_cache_all();
 250                         prom_startcpu(linux_cpus[no].prom_node,
 251                                       &smp_penguin_ctable, 0, (char *)entry);
 252
 253                         SMP_PRINTK(("prom_startcpu returned :)\n"));
 254
 255                         /* wheee... it's going... */
 256                         for(timeout = 0; timeout < 10000; timeout++) {
 257                                 if(cpu_callin_map[i])
 258                                         break;
 259                                 udelay(200);
 260                         }
 261
 262                         if(cpu_callin_map[i]) {
 263                                 /* Another "Red Snapper". */
 264                                 __cpu_number_map[i] = cpucount;
 265                                 __cpu_logical_map[cpucount] = i;
 266                         } else {
 267                                 cpucount--;
 268                                 printk("Processor %d is stuck.\n", i);
 269                         }
 270                 }
 271                 if(!(cpu_callin_map[i])) {
 272                         cpu_present_map &= ~(1 << i);
 273                         __cpu_number_map[i] = -1;
 274                 }
 275         }
 276         local_flush_cache_all();
 277         if(cpucount == 0) {
 278                 printk("Error: only one Processor found.\n");
 279                 cpu_present_map = (1 << hard_smp4d_processor_id());
 280         } else {
 281                 unsigned long bogosum = 0;
 282
 283                 for(i = 0; i < NR_CPUS; i++) {
 284                         if(cpu_present_map & (1 << i)) {
 285                                 bogosum += cpu_data[i].udelay_val;
 286                                 smp_highest_cpu = i;
 287                         }
 288                 }
 289                 SMP_PRINTK(("Total of %d Processors activated (%lu.%02lu BogoMIPS).\n", cpucount + 1, bogosum/(500000/HZ), (bogosum/(5000/HZ))%100));
 290                 printk("Total of %d Processors activated (%lu.%02lu BogoMIPS).\n",
 291                        cpucount + 1,
 292                        bogosum/(500000/HZ),
 293                        (bogosum/(5000/HZ))%100);
 294                 smp_activated = 1;
 295                 smp_num_cpus = cpucount + 1;
 296         }
 297
 298         /* Free unneeded trap tables */
 299         ClearPageReserved(virt_to_page(trapbase_cpu1));
 300         set_page_count(virt_to_page(trapbase_cpu1), 1);
 301         free_page((unsigned long)trapbase_cpu1);
 302         totalram_pages++;
 303         num_physpages++;
 304
 305         ClearPageReserved(virt_to_page(trapbase_cpu2));
 306         set_page_count(virt_to_page(trapbase_cpu2), 1);
 307         free_page((unsigned long)trapbase_cpu2);
 308         totalram_pages++;
 309         num_physpages++;
 310
 311         ClearPageReserved(virt_to_page(trapbase_cpu3));
 312         set_page_count(virt_to_page(trapbase_cpu3), 1);
 313         free_page((unsigned long)trapbase_cpu3);
 314         totalram_pages++;
 315         num_physpages++;
 316
 317         /* Ok, they are spinning and ready to go. */
 318         smp_processors_ready = 1;
 319         sun4d_distribute_irqs();
 320 }
 321
 322 static struct smp_funcall {
 323         smpfunc_t func;
 324         unsigned long arg1;
 325         unsigned long arg2;
 326         unsigned long arg3;
 327         unsigned long arg4;
 328         unsigned long arg5;
 329         unsigned char processors_in[NR_CPUS];  /* Set when ipi entered. */
 330         unsigned char processors_out[NR_CPUS]; /* Set when ipi exited. */
 331 } ccall_info __attribute__((aligned(8)));
 332
 333 static spinlock_t cross_call_lock = SPIN_LOCK_UNLOCKED;
 334
 335 /* Cross calls must be serialized, at least currently. */
 336 void smp4d_cross_call(smpfunc_t func, unsigned long arg1, unsigned long arg2,
 337                     unsigned long arg3, unsigned long arg4, unsigned long arg5)
 338 {
 339         if(smp_processors_ready) {
 340                 register int high = smp_highest_cpu;
 341                 unsigned long flags;
 342
 343                 spin_lock_irqsave(&cross_call_lock, flags);
 344
 345                 {
 346                         /* If you make changes here, make sure gcc generates proper code... */
 347                         smpfunc_t f asm("i0") = func;
 348                         unsigned long a1 asm("i1") = arg1;
 349                         unsigned long a2 asm("i2") = arg2;
 350                         unsigned long a3 asm("i3") = arg3;
 351                         unsigned long a4 asm("i4") = arg4;
 352                         unsigned long a5 asm("i5") = arg5;
 353
 354                         __asm__ __volatile__(
 355                                 "std %0, [%6]\n\t"
 356                                 "std %2, [%6 + 8]\n\t"
 357                                 "std %4, [%6 + 16]\n\t" : :
 358                                 "r"(f), "r"(a1), "r"(a2), "r"(a3), "r"(a4), "r"(a5),
 359                                 "r" (&ccall_info.func));
 360                 }
 361
 362                 /* Init receive/complete mapping, plus fire the IPI's off. */
 363                 {
 364                         register unsigned long mask;
 365                         register int i;
 366
 367                         mask = (cpu_present_map & ~(1 << hard_smp4d_processor_id()));
 368                         for(i = 0; i <= high; i++) {
 369                                 if(mask & (1 << i)) {
 370                                         ccall_info.processors_in[i] = 0;
 371                                         ccall_info.processors_out[i] = 0;
 372                                         sun4d_send_ipi(i, IRQ_CROSS_CALL);
 373                                 }
 374                         }
 375                 }
 376
 377                 {
 378                         register int i;
 379
 380                         i = 0;
 381                         do {
 382                                 while(!ccall_info.processors_in[i])
 383                                         barrier();
 384                         } while(++i <= high);
 385
 386                         i = 0;
 387                         do {
 388                                 while(!ccall_info.processors_out[i])
 389                                         barrier();
 390                         } while(++i <= high);
 391                 }
 392
 393                 spin_unlock_irqrestore(&cross_call_lock, flags);
 394         }
 395 }
 396
 397 /* Running cross calls. */
 398 void smp4d_cross_call_irq(void)
 399 {
 400         int i = hard_smp4d_processor_id();
 401
 402         ccall_info.processors_in[i] = 1;
 403         ccall_info.func(ccall_info.arg1, ccall_info.arg2, ccall_info.arg3,
 404                         ccall_info.arg4, ccall_info.arg5);
 405         ccall_info.processors_out[i] = 1;
 406 }
 407
 408 static int smp4d_stop_cpu_sender;
 409
 410 static void smp4d_stop_cpu(void)
 411 {
 412         int me = hard_smp4d_processor_id();
 413
 414         if (me != smp4d_stop_cpu_sender)
 415                 while(1) barrier();
 416 }
 417
 418 /* Cross calls, in order to work efficiently and atomically do all
 419  * the message passing work themselves, only stopcpu and reschedule
 420  * messages come through here.
 421  */
 422 void smp4d_message_pass(int target, int msg, unsigned long data, int wait)
 423 {
 424         int me = hard_smp4d_processor_id();
 425
 426         SMP_PRINTK(("smp4d_message_pass %d %d %08lx %d\n", target, msg, data, wait));
 427         if (msg == MSG_STOP_CPU && target == MSG_ALL_BUT_SELF) {
 428                 unsigned long flags;
 429                 static spinlock_t stop_cpu_lock = SPIN_LOCK_UNLOCKED;
 430                 spin_lock_irqsave(&stop_cpu_lock, flags);
 431                 smp4d_stop_cpu_sender = me;
 432                 smp4d_cross_call((smpfunc_t)smp4d_stop_cpu, 0, 0, 0, 0, 0);
 433                 spin_unlock_irqrestore(&stop_cpu_lock, flags);
 434         }
 435         printk("Yeeee, trying to send SMP msg(%d) to %d on cpu %d\n", msg, target, me);
 436         panic("Bogon SMP message pass.");
 437 }
 438
 439 extern unsigned int prof_multiplier[NR_CPUS];
 440 extern unsigned int prof_counter[NR_CPUS];
 441
 442 extern void sparc_do_profile(unsigned long pc, unsigned long o7);
 443
 444 void smp4d_percpu_timer_interrupt(struct pt_regs *regs)
 445 {
 446         int cpu = hard_smp4d_processor_id();
 447         static int cpu_tick[NR_CPUS];
 448         static char led_mask[] = { 0xe, 0xd, 0xb, 0x7, 0xb, 0xd };
 449
 450         bw_get_prof_limit(cpu);
 451         bw_clear_intr_mask(0, 1);       /* INTR_TABLE[0] & 1 is Profile IRQ */
 452
 453         cpu_tick[cpu]++;
 454         if (!(cpu_tick[cpu] & 15)) {
 455                 if (cpu_tick[cpu] == 0x60)
 456                         cpu_tick[cpu] = 0;
 457                 cpu_leds[cpu] = led_mask[cpu_tick[cpu] >> 4];
 458                 show_leds(cpu);
 459         }
 460
 461         if(!user_mode(regs))
 462                 sparc_do_profile(regs->pc, regs->u_regs[UREG_RETPC]);
 463
 464         if(!--prof_counter[cpu]) {
 465                 int user = user_mode(regs);
 466
 467                 irq_enter(cpu, 0);
 468                 update_process_times(user);
 469                 irq_exit(cpu, 0);
 470
 471                 prof_counter[cpu] = prof_multiplier[cpu];
 472         }
 473 }
 474
 475 extern unsigned int lvl14_resolution;
 476
 477 static void __init smp_setup_percpu_timer(void)
 478 {
 479         int cpu = hard_smp4d_processor_id();
 480
 481         prof_counter[cpu] = prof_multiplier[cpu] = 1;
 482         load_profile_irq(cpu, lvl14_resolution);
 483 }
 484
 485 void __init smp4d_blackbox_id(unsigned *addr)
 486 {
 487         int rd = *addr & 0x3e000000;
 488
 489         addr[0] = 0xc0800800 | rd;              /* lda [%g0] ASI_M_VIKING_TMP1, reg */
 490         addr[1] = 0x01000000;                   /* nop */
 491         addr[2] = 0x01000000;                   /* nop */
 492 }
 493
 494 void __init smp4d_blackbox_current(unsigned *addr)
 495 {
 496         int rd = *addr & 0x3e000000;
 497
 498         addr[0] = 0xc0800800 | rd;              /* lda [%g0] ASI_M_VIKING_TMP1, reg */
 499         addr[2] = 0x81282002 | rd | (rd >> 11); /* sll reg, 2, reg */
 500         addr[4] = 0x01000000;                   /* nop */
 501 }
 502
 503 void __init sun4d_init_smp(void)
 504 {
 505         int i;
 506         extern unsigned int t_nmi[], linux_trap_ipi15_sun4d[], linux_trap_ipi15_sun4m[];
 507
 508         /* Patch ipi15 trap table */
 509         t_nmi[1] = t_nmi[1] + (linux_trap_ipi15_sun4d - linux_trap_ipi15_sun4m);
 510
 511         /* And set btfixup... */
 512         BTFIXUPSET_BLACKBOX(smp_processor_id, smp4d_blackbox_id);
 513         BTFIXUPSET_BLACKBOX(load_current, smp4d_blackbox_current);
 514         BTFIXUPSET_CALL(smp_cross_call, smp4d_cross_call, BTFIXUPCALL_NORM);
 515         BTFIXUPSET_CALL(smp_message_pass, smp4d_message_pass, BTFIXUPCALL_NORM);
 516         BTFIXUPSET_CALL(__smp_processor_id, __smp4d_processor_id, BTFIXUPCALL_NORM);
 517
 518         for (i = 0; i < NR_CPUS; i++) {
 519                 ccall_info.processors_in[i] = 1;
 520                 ccall_info.processors_out[i] = 1;
 521         }
 522 }