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/interrupt.h>
  16 #include <linux/kernel_stat.h>
  17 #include <linux/init.h>
  18 #include <linux/spinlock.h>
  19 #include <linux/mm.h>
  20 #include <linux/swap.h>
  21 #include <linux/profile.h>
  22 #include <linux/delay.h>
  23
  24 #include <asm/ptrace.h>
  25 #include <asm/atomic.h>
  26 #include <asm/irq_regs.h>
  27
  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/sbi.h>
  34 #include <asm/tlbflush.h>
  35 #include <asm/cacheflush.h>
  36 #include <asm/cpudata.h>
  37
  38 #include "irq.h"
  39 #define IRQ_CROSS_CALL          15
  40
  41 extern ctxd_t *srmmu_ctx_table_phys;
  42
  43 static volatile int smp_processors_ready = 0;
  44 static int smp_highest_cpu;
  45 extern volatile unsigned long cpu_callin_map[NR_CPUS];
  46 extern cpuinfo_sparc cpu_data[NR_CPUS];
  47 extern unsigned char boot_cpu_id;
  48 extern volatile int smp_process_available;
  49
  50 extern cpumask_t smp_commenced_mask;
  51
  52 extern int __smp4d_processor_id(void);
  53
  54 /* #define SMP_DEBUG */
  55
  56 #ifdef SMP_DEBUG
  57 #define SMP_PRINTK(x)   printk x
  58 #else
  59 #define SMP_PRINTK(x)
  60 #endif
  61
  62 static inline unsigned long swap(volatile unsigned long *ptr, unsigned long val)
  63 {
  64         __asm__ __volatile__("swap [%1], %0\n\t" :
  65                              "=&r" (val), "=&r" (ptr) :
  66                              "0" (val), "1" (ptr));
  67         return val;
  68 }
  69
  70 static void smp_setup_percpu_timer(void);
  71 extern void cpu_probe(void);
  72 extern void sun4d_distribute_irqs(void);
  73
  74 static unsigned char cpu_leds[32];
  75
  76 static inline void show_leds(int cpuid)
  77 {
  78         cpuid &= 0x1e;
  79         __asm__ __volatile__ ("stba %0, [%1] %2" : :
  80                               "r" ((cpu_leds[cpuid] << 4) | cpu_leds[cpuid+1]),
  81                               "r" (ECSR_BASE(cpuid) | BB_LEDS),
  82                               "i" (ASI_M_CTL));
  83 }
  84
  85 void __init smp4d_callin(void)
  86 {
  87         int cpuid = hard_smp4d_processor_id();
  88         extern spinlock_t sun4d_imsk_lock;
  89         unsigned long flags;
  90
  91         /* Show we are alive */
  92         cpu_leds[cpuid] = 0x6;
  93         show_leds(cpuid);
  94
  95         /* Enable level15 interrupt, disable level14 interrupt for now */
  96         cc_set_imsk((cc_get_imsk() & ~0x8000) | 0x4000);
  97
  98         local_flush_cache_all();
  99         local_flush_tlb_all();
 100
 101         /*
 102          * Unblock the master CPU _only_ when the scheduler state
 103          * of all secondary CPUs will be up-to-date, so after
 104          * the SMP initialization the master will be just allowed
 105          * to call the scheduler code.
 106          */
 107         /* Get our local ticker going. */
 108         smp_setup_percpu_timer();
 109
 110         calibrate_delay();
 111         smp_store_cpu_info(cpuid);
 112         local_flush_cache_all();
 113         local_flush_tlb_all();
 114
 115         /* Allow master to continue. */
 116         swap((unsigned long *)&cpu_callin_map[cpuid], 1);
 117         local_flush_cache_all();
 118         local_flush_tlb_all();
 119
 120         cpu_probe();
 121
 122         while((unsigned long)current_set[cpuid] < PAGE_OFFSET)
 123                 barrier();
 124
 125         while(current_set[cpuid]->cpu != cpuid)
 126                 barrier();
 127
 128         /* Fix idle thread fields. */
 129         __asm__ __volatile__("ld [%0], %%g6\n\t"
 130                              : : "r" (&current_set[cpuid])
 131                              : "memory" /* paranoid */);
 132
 133         cpu_leds[cpuid] = 0x9;
 134         show_leds(cpuid);
 135
 136         /* Attach to the address space of init_task. */
 137         atomic_inc(&init_mm.mm_count);
 138         current->active_mm = &init_mm;
 139
 140         local_flush_cache_all();
 141         local_flush_tlb_all();
 142
 143         local_irq_enable();     /* We don't allow PIL 14 yet */
 144
 145         while (!cpu_isset(cpuid, smp_commenced_mask))
 146                 barrier();
 147
 148         spin_lock_irqsave(&sun4d_imsk_lock, flags);
 149         cc_set_imsk(cc_get_imsk() & ~0x4000); /* Allow PIL 14 as well */
 150         spin_unlock_irqrestore(&sun4d_imsk_lock, flags);
 151         cpu_set(cpuid, cpu_online_map);
 152
 153 }
 154
 155 extern void init_IRQ(void);
 156 extern void cpu_panic(void);
 157
 158 /*
 159  *      Cycle through the processors asking the PROM to start each one.
 160  */
 161
 162 extern struct linux_prom_registers smp_penguin_ctable;
 163 extern unsigned long trapbase_cpu1[];
 164 extern unsigned long trapbase_cpu2[];
 165 extern unsigned long trapbase_cpu3[];
 166
 167 void __init smp4d_boot_cpus(void)
 168 {
 169         if (boot_cpu_id)
 170                 current_set[0] = NULL;
 171         smp_setup_percpu_timer();
 172         local_flush_cache_all();
 173 }
 174
 175 int __cpuinit smp4d_boot_one_cpu(int i)
 176 {
 177                         extern unsigned long sun4d_cpu_startup;
 178                         unsigned long *entry = &sun4d_cpu_startup;
 179                         struct task_struct *p;
 180                         int timeout;
 181                         int cpu_node;
 182
 183                         cpu_find_by_instance(i, &cpu_node,NULL);
 184                         /* Cook up an idler for this guy. */
 185                         p = fork_idle(i);
 186                         current_set[i] = task_thread_info(p);
 187
 188                         /*
 189                          * Initialize the contexts table
 190                          * Since the call to prom_startcpu() trashes the structure,
 191                          * we need to re-initialize it for each cpu
 192                          */
 193                         smp_penguin_ctable.which_io = 0;
 194                         smp_penguin_ctable.phys_addr = (unsigned int) srmmu_ctx_table_phys;
 195                         smp_penguin_ctable.reg_size = 0;
 196
 197                         /* whirrr, whirrr, whirrrrrrrrr... */
 198                         SMP_PRINTK(("Starting CPU %d at %p \n", i, entry));
 199                         local_flush_cache_all();
 200                         prom_startcpu(cpu_node,
 201                                       &smp_penguin_ctable, 0, (char *)entry);
 202
 203                         SMP_PRINTK(("prom_startcpu returned :)\n"));
 204
 205                         /* wheee... it's going... */
 206                         for(timeout = 0; timeout < 10000; timeout++) {
 207                                 if(cpu_callin_map[i])
 208                                         break;
 209                                 udelay(200);
 210                         }
 211
 212         if (!(cpu_callin_map[i])) {
 213                 printk("Processor %d is stuck.\n", i);
 214                 return -ENODEV;
 215
 216         }
 217         local_flush_cache_all();
 218         return 0;
 219 }
 220
 221 void __init smp4d_smp_done(void)
 222 {
 223         int i, first;
 224         int *prev;
 225
 226         /* setup cpu list for irq rotation */
 227         first = 0;
 228         prev = &first;
 229         for (i = 0; i < NR_CPUS; i++)
 230                 if (cpu_online(i)) {
 231                         *prev = i;
 232                         prev = &cpu_data(i).next;
 233                 }
 234         *prev = first;
 235         local_flush_cache_all();
 236
 237         /* Free unneeded trap tables */
 238         ClearPageReserved(virt_to_page(trapbase_cpu1));
 239         init_page_count(virt_to_page(trapbase_cpu1));
 240         free_page((unsigned long)trapbase_cpu1);
 241         totalram_pages++;
 242         num_physpages++;
 243
 244         ClearPageReserved(virt_to_page(trapbase_cpu2));
 245         init_page_count(virt_to_page(trapbase_cpu2));
 246         free_page((unsigned long)trapbase_cpu2);
 247         totalram_pages++;
 248         num_physpages++;
 249
 250         ClearPageReserved(virt_to_page(trapbase_cpu3));
 251         init_page_count(virt_to_page(trapbase_cpu3));
 252         free_page((unsigned long)trapbase_cpu3);
 253         totalram_pages++;
 254         num_physpages++;
 255
 256         /* Ok, they are spinning and ready to go. */
 257         smp_processors_ready = 1;
 258         sun4d_distribute_irqs();
 259 }
 260
 261 static struct smp_funcall {
 262         smpfunc_t func;
 263         unsigned long arg1;
 264         unsigned long arg2;
 265         unsigned long arg3;
 266         unsigned long arg4;
 267         unsigned long arg5;
 268         unsigned char processors_in[NR_CPUS];  /* Set when ipi entered. */
 269         unsigned char processors_out[NR_CPUS]; /* Set when ipi exited. */
 270 } ccall_info __attribute__((aligned(8)));
 271
 272 static DEFINE_SPINLOCK(cross_call_lock);
 273
 274 /* Cross calls must be serialized, at least currently. */
 275 static void smp4d_cross_call(smpfunc_t func, cpumask_t mask, unsigned long arg1,
 276                              unsigned long arg2, unsigned long arg3,
 277                              unsigned long arg4)
 278 {
 279         if(smp_processors_ready) {
 280                 register int high = smp_highest_cpu;
 281                 unsigned long flags;
 282
 283                 spin_lock_irqsave(&cross_call_lock, flags);
 284
 285                 {
 286                         /* If you make changes here, make sure gcc generates proper code... */
 287                         register smpfunc_t f asm("i0") = func;
 288                         register unsigned long a1 asm("i1") = arg1;
 289                         register unsigned long a2 asm("i2") = arg2;
 290                         register unsigned long a3 asm("i3") = arg3;
 291                         register unsigned long a4 asm("i4") = arg4;
 292                         register unsigned long a5 asm("i5") = 0;
 293
 294                         __asm__ __volatile__(
 295                                 "std %0, [%6]\n\t"
 296                                 "std %2, [%6 + 8]\n\t"
 297                                 "std %4, [%6 + 16]\n\t" : :
 298                                 "r"(f), "r"(a1), "r"(a2), "r"(a3), "r"(a4), "r"(a5),
 299                                 "r" (&ccall_info.func));
 300                 }
 301
 302                 /* Init receive/complete mapping, plus fire the IPI's off. */
 303                 {
 304                         register int i;
 305
 306                         cpu_clear(smp_processor_id(), mask);
 307                         cpus_and(mask, cpu_online_map, mask);
 308                         for(i = 0; i <= high; i++) {
 309                                 if (cpu_isset(i, mask)) {
 310                                         ccall_info.processors_in[i] = 0;
 311                                         ccall_info.processors_out[i] = 0;
 312                                         sun4d_send_ipi(i, IRQ_CROSS_CALL);
 313                                 }
 314                         }
 315                 }
 316
 317                 {
 318                         register int i;
 319
 320                         i = 0;
 321                         do {
 322                                 if (!cpu_isset(i, mask))
 323                                         continue;
 324                                 while(!ccall_info.processors_in[i])
 325                                         barrier();
 326                         } while(++i <= high);
 327
 328                         i = 0;
 329                         do {
 330                                 if (!cpu_isset(i, mask))
 331                                         continue;
 332                                 while(!ccall_info.processors_out[i])
 333                                         barrier();
 334                         } while(++i <= high);
 335                 }
 336
 337                 spin_unlock_irqrestore(&cross_call_lock, flags);
 338         }
 339 }
 340
 341 /* Running cross calls. */
 342 void smp4d_cross_call_irq(void)
 343 {
 344         int i = hard_smp4d_processor_id();
 345
 346         ccall_info.processors_in[i] = 1;
 347         ccall_info.func(ccall_info.arg1, ccall_info.arg2, ccall_info.arg3,
 348                         ccall_info.arg4, ccall_info.arg5);
 349         ccall_info.processors_out[i] = 1;
 350 }
 351
 352 void smp4d_percpu_timer_interrupt(struct pt_regs *regs)
 353 {
 354         struct pt_regs *old_regs;
 355         int cpu = hard_smp4d_processor_id();
 356         static int cpu_tick[NR_CPUS];
 357         static char led_mask[] = { 0xe, 0xd, 0xb, 0x7, 0xb, 0xd };
 358
 359         old_regs = set_irq_regs(regs);
 360         bw_get_prof_limit(cpu);
 361         bw_clear_intr_mask(0, 1);       /* INTR_TABLE[0] & 1 is Profile IRQ */
 362
 363         cpu_tick[cpu]++;
 364         if (!(cpu_tick[cpu] & 15)) {
 365                 if (cpu_tick[cpu] == 0x60)
 366                         cpu_tick[cpu] = 0;
 367                 cpu_leds[cpu] = led_mask[cpu_tick[cpu] >> 4];
 368                 show_leds(cpu);
 369         }
 370
 371         profile_tick(CPU_PROFILING);
 372
 373         if(!--prof_counter(cpu)) {
 374                 int user = user_mode(regs);
 375
 376                 irq_enter();
 377                 update_process_times(user);
 378                 irq_exit();
 379
 380                 prof_counter(cpu) = prof_multiplier(cpu);
 381         }
 382         set_irq_regs(old_regs);
 383 }
 384
 385 extern unsigned int lvl14_resolution;
 386
 387 static void __init smp_setup_percpu_timer(void)
 388 {
 389         int cpu = hard_smp4d_processor_id();
 390
 391         prof_counter(cpu) = prof_multiplier(cpu) = 1;
 392         load_profile_irq(cpu, lvl14_resolution);
 393 }
 394
 395 void __init smp4d_blackbox_id(unsigned *addr)
 396 {
 397         int rd = *addr & 0x3e000000;
 398
 399         addr[0] = 0xc0800800 | rd;              /* lda [%g0] ASI_M_VIKING_TMP1, reg */
 400         addr[1] = 0x01000000;                   /* nop */
 401         addr[2] = 0x01000000;                   /* nop */
 402 }
 403
 404 void __init smp4d_blackbox_current(unsigned *addr)
 405 {
 406         int rd = *addr & 0x3e000000;
 407
 408         addr[0] = 0xc0800800 | rd;              /* lda [%g0] ASI_M_VIKING_TMP1, reg */
 409         addr[2] = 0x81282002 | rd | (rd >> 11); /* sll reg, 2, reg */
 410         addr[4] = 0x01000000;                   /* nop */
 411 }
 412
 413 void __init sun4d_init_smp(void)
 414 {
 415         int i;
 416         extern unsigned int t_nmi[], linux_trap_ipi15_sun4d[], linux_trap_ipi15_sun4m[];
 417
 418         /* Patch ipi15 trap table */
 419         t_nmi[1] = t_nmi[1] + (linux_trap_ipi15_sun4d - linux_trap_ipi15_sun4m);
 420
 421         /* And set btfixup... */
 422         BTFIXUPSET_BLACKBOX(hard_smp_processor_id, smp4d_blackbox_id);
 423         BTFIXUPSET_BLACKBOX(load_current, smp4d_blackbox_current);
 424         BTFIXUPSET_CALL(smp_cross_call, smp4d_cross_call, BTFIXUPCALL_NORM);
 425         BTFIXUPSET_CALL(__hard_smp_processor_id, __smp4d_processor_id, BTFIXUPCALL_NORM);
 426
 427         for (i = 0; i < NR_CPUS; i++) {
 428                 ccall_info.processors_in[i] = 1;
 429                 ccall_info.processors_out[i] = 1;
 430         }
 431 }