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