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 #include <linux/cpu.h>
  24
  25 #include <asm/ptrace.h>
  26 #include <asm/atomic.h>
  27 #include <asm/irq_regs.h>
  28
  29 #include <asm/irq.h>
  30 #include <asm/page.h>
  31 #include <asm/pgalloc.h>
  32 #include <asm/pgtable.h>
  33 #include <asm/oplib.h>
  34 #include <asm/sbi.h>
  35 #include <asm/tlbflush.h>
  36 #include <asm/cacheflush.h>
  37 #include <asm/cpudata.h>
  38
  39 #include "irq.h"
  40 #define IRQ_CROSS_CALL          15
  41
  42 extern ctxd_t *srmmu_ctx_table_phys;
  43
  44 static volatile int smp_processors_ready = 0;
  45 static int smp_highest_cpu;
  46 extern volatile unsigned long cpu_callin_map[NR_CPUS];
  47 extern cpuinfo_sparc cpu_data[NR_CPUS];
  48 extern unsigned char boot_cpu_id;
  49 extern volatile int smp_process_available;
  50
  51 extern cpumask_t smp_commenced_mask;
  52
  53 extern int __smp4d_processor_id(void);
  54
  55 /* #define SMP_DEBUG */
  56
  57 #ifdef SMP_DEBUG
  58 #define SMP_PRINTK(x)   printk x
  59 #else
  60 #define SMP_PRINTK(x)
  61 #endif
  62
  63 static inline unsigned long sun4d_swap(volatile unsigned long *ptr, unsigned long val)
  64 {
  65         __asm__ __volatile__("swap [%1], %0\n\t" :
  66                              "=&r" (val), "=&r" (ptr) :
  67                              "0" (val), "1" (ptr));
  68         return val;
  69 }
  70
  71 static void smp_setup_percpu_timer(void);
  72 extern void cpu_probe(void);
  73 extern void sun4d_distribute_irqs(void);
  74
  75 static unsigned char cpu_leds[32];
  76
  77 static inline void show_leds(int cpuid)
  78 {
  79         cpuid &= 0x1e;
  80         __asm__ __volatile__ ("stba %0, [%1] %2" : :
  81                               "r" ((cpu_leds[cpuid] << 4) | cpu_leds[cpuid+1]),
  82                               "r" (ECSR_BASE(cpuid) | BB_LEDS),
  83                               "i" (ASI_M_CTL));
  84 }
  85
  86 void __cpuinit smp4d_callin(void)
  87 {
  88         int cpuid = hard_smp4d_processor_id();
  89         extern spinlock_t sun4d_imsk_lock;
  90         unsigned long flags;
  91
  92         /* Show we are alive */
  93         cpu_leds[cpuid] = 0x6;
  94         show_leds(cpuid);
  95
  96         /* Enable level15 interrupt, disable level14 interrupt for now */
  97         cc_set_imsk((cc_get_imsk() & ~0x8000) | 0x4000);
  98
  99         local_flush_cache_all();
 100         local_flush_tlb_all();
 101
 102         notify_cpu_starting(cpuid);
 103         /*
 104          * Unblock the master CPU _only_ when the scheduler state
 105          * of all secondary CPUs will be up-to-date, so after
 106          * the SMP initialization the master will be just allowed
 107          * to call the scheduler code.
 108          */
 109         /* Get our local ticker going. */
 110         smp_setup_percpu_timer();
 111
 112         calibrate_delay();
 113         smp_store_cpu_info(cpuid);
 114         local_flush_cache_all();
 115         local_flush_tlb_all();
 116
 117         /* Allow master to continue. */
 118         sun4d_swap((unsigned long *)&cpu_callin_map[cpuid], 1);
 119         local_flush_cache_all();
 120         local_flush_tlb_all();
 121
 122         cpu_probe();
 123
 124         while((unsigned long)current_set[cpuid] < PAGE_OFFSET)
 125                 barrier();
 126
 127         while(current_set[cpuid]->cpu != cpuid)
 128                 barrier();
 129
 130         /* Fix idle thread fields. */
 131         __asm__ __volatile__("ld [%0], %%g6\n\t"
 132                              : : "r" (&current_set[cpuid])
 133                              : "memory" /* paranoid */);
 134
 135         cpu_leds[cpuid] = 0x9;
 136         show_leds(cpuid);
 137
 138         /* Attach to the address space of init_task. */
 139         atomic_inc(&init_mm.mm_count);
 140         current->active_mm = &init_mm;
 141
 142         local_flush_cache_all();
 143         local_flush_tlb_all();
 144
 145         local_irq_enable();     /* We don't allow PIL 14 yet */
 146
 147         while (!cpu_isset(cpuid, smp_commenced_mask))
 148                 barrier();
 149
 150         spin_lock_irqsave(&sun4d_imsk_lock, flags);
 151         cc_set_imsk(cc_get_imsk() & ~0x4000); /* Allow PIL 14 as well */
 152         spin_unlock_irqrestore(&sun4d_imsk_lock, flags);
 153         set_cpu_online(cpuid, true);
 154
 155 }
 156
 157 extern void init_IRQ(void);
 158 extern void cpu_panic(void);
 159
 160 /*
 161  *      Cycle through the processors asking the PROM to start each one.
 162  */
 163
 164 extern struct linux_prom_registers smp_penguin_ctable;
 165
 166 void __init smp4d_boot_cpus(void)
 167 {
 168         if (boot_cpu_id)
 169                 current_set[0] = NULL;
 170         smp_setup_percpu_timer();
 171         local_flush_cache_all();
 172 }
 173
 174 int __cpuinit smp4d_boot_one_cpu(int i)
 175 {
 176                         extern unsigned long sun4d_cpu_startup;
 177                         unsigned long *entry = &sun4d_cpu_startup;
 178                         struct task_struct *p;
 179                         int timeout;
 180                         int cpu_node;
 181
 182                         cpu_find_by_instance(i, &cpu_node,NULL);
 183                         /* Cook up an idler for this guy. */
 184                         p = fork_idle(i);
 185                         current_set[i] = task_thread_info(p);
 186
 187                         /*
 188                          * Initialize the contexts table
 189                          * Since the call to prom_startcpu() trashes the structure,
 190                          * we need to re-initialize it for each cpu
 191                          */
 192                         smp_penguin_ctable.which_io = 0;
 193                         smp_penguin_ctable.phys_addr = (unsigned int) srmmu_ctx_table_phys;
 194                         smp_penguin_ctable.reg_size = 0;
 195
 196                         /* whirrr, whirrr, whirrrrrrrrr... */
 197                         SMP_PRINTK(("Starting CPU %d at %p\n", i, entry));
 198                         local_flush_cache_all();
 199                         prom_startcpu(cpu_node,
 200                                       &smp_penguin_ctable, 0, (char *)entry);
 201
 202                         SMP_PRINTK(("prom_startcpu returned :)\n"));
 203
 204                         /* wheee... it's going... */
 205                         for(timeout = 0; timeout < 10000; timeout++) {
 206                                 if(cpu_callin_map[i])
 207                                         break;
 208                                 udelay(200);
 209                         }
 210
 211         if (!(cpu_callin_map[i])) {
 212                 printk("Processor %d is stuck.\n", i);
 213                 return -ENODEV;
 214
 215         }
 216         local_flush_cache_all();
 217         return 0;
 218 }
 219
 220 void __init smp4d_smp_done(void)
 221 {
 222         int i, first;
 223         int *prev;
 224
 225         /* setup cpu list for irq rotation */
 226         first = 0;
 227         prev = &first;
 228         for_each_online_cpu(i) {
 229                 *prev = i;
 230                 prev = &cpu_data(i).next;
 231         }
 232         *prev = first;
 233         local_flush_cache_all();
 234
 235         /* Ok, they are spinning and ready to go. */
 236         smp_processors_ready = 1;
 237         sun4d_distribute_irqs();
 238 }
 239
 240 static struct smp_funcall {
 241         smpfunc_t func;
 242         unsigned long arg1;
 243         unsigned long arg2;
 244         unsigned long arg3;
 245         unsigned long arg4;
 246         unsigned long arg5;
 247         unsigned char processors_in[NR_CPUS];  /* Set when ipi entered. */
 248         unsigned char processors_out[NR_CPUS]; /* Set when ipi exited. */
 249 } ccall_info __attribute__((aligned(8)));
 250
 251 static DEFINE_SPINLOCK(cross_call_lock);
 252
 253 /* Cross calls must be serialized, at least currently. */
 254 static void smp4d_cross_call(smpfunc_t func, cpumask_t mask, unsigned long arg1,
 255                              unsigned long arg2, unsigned long arg3,
 256                              unsigned long arg4)
 257 {
 258         if(smp_processors_ready) {
 259                 register int high = smp_highest_cpu;
 260                 unsigned long flags;
 261
 262                 spin_lock_irqsave(&cross_call_lock, flags);
 263
 264                 {
 265                         /* If you make changes here, make sure gcc generates proper code... */
 266                         register smpfunc_t f asm("i0") = func;
 267                         register unsigned long a1 asm("i1") = arg1;
 268                         register unsigned long a2 asm("i2") = arg2;
 269                         register unsigned long a3 asm("i3") = arg3;
 270                         register unsigned long a4 asm("i4") = arg4;
 271                         register unsigned long a5 asm("i5") = 0;
 272
 273                         __asm__ __volatile__(
 274                                 "std %0, [%6]\n\t"
 275                                 "std %2, [%6 + 8]\n\t"
 276                                 "std %4, [%6 + 16]\n\t" : :
 277                                 "r"(f), "r"(a1), "r"(a2), "r"(a3), "r"(a4), "r"(a5),
 278                                 "r" (&ccall_info.func));
 279                 }
 280
 281                 /* Init receive/complete mapping, plus fire the IPI's off. */
 282                 {
 283                         register int i;
 284
 285                         cpu_clear(smp_processor_id(), mask);
 286                         cpus_and(mask, cpu_online_map, mask);
 287                         for(i = 0; i <= high; i++) {
 288                                 if (cpu_isset(i, mask)) {
 289                                         ccall_info.processors_in[i] = 0;
 290                                         ccall_info.processors_out[i] = 0;
 291                                         sun4d_send_ipi(i, IRQ_CROSS_CALL);
 292                                 }
 293                         }
 294                 }
 295
 296                 {
 297                         register int i;
 298
 299                         i = 0;
 300                         do {
 301                                 if (!cpu_isset(i, mask))
 302                                         continue;
 303                                 while(!ccall_info.processors_in[i])
 304                                         barrier();
 305                         } while(++i <= high);
 306
 307                         i = 0;
 308                         do {
 309                                 if (!cpu_isset(i, mask))
 310                                         continue;
 311                                 while(!ccall_info.processors_out[i])
 312                                         barrier();
 313                         } while(++i <= high);
 314                 }
 315
 316                 spin_unlock_irqrestore(&cross_call_lock, flags);
 317         }
 318 }
 319
 320 /* Running cross calls. */
 321 void smp4d_cross_call_irq(void)
 322 {
 323         int i = hard_smp4d_processor_id();
 324
 325         ccall_info.processors_in[i] = 1;
 326         ccall_info.func(ccall_info.arg1, ccall_info.arg2, ccall_info.arg3,
 327                         ccall_info.arg4, ccall_info.arg5);
 328         ccall_info.processors_out[i] = 1;
 329 }
 330
 331 void smp4d_percpu_timer_interrupt(struct pt_regs *regs)
 332 {
 333         struct pt_regs *old_regs;
 334         int cpu = hard_smp4d_processor_id();
 335         static int cpu_tick[NR_CPUS];
 336         static char led_mask[] = { 0xe, 0xd, 0xb, 0x7, 0xb, 0xd };
 337
 338         old_regs = set_irq_regs(regs);
 339         bw_get_prof_limit(cpu);
 340         bw_clear_intr_mask(0, 1);       /* INTR_TABLE[0] & 1 is Profile IRQ */
 341
 342         cpu_tick[cpu]++;
 343         if (!(cpu_tick[cpu] & 15)) {
 344                 if (cpu_tick[cpu] == 0x60)
 345                         cpu_tick[cpu] = 0;
 346                 cpu_leds[cpu] = led_mask[cpu_tick[cpu] >> 4];
 347                 show_leds(cpu);
 348         }
 349
 350         profile_tick(CPU_PROFILING);
 351
 352         if(!--prof_counter(cpu)) {
 353                 int user = user_mode(regs);
 354
 355                 irq_enter();
 356                 update_process_times(user);
 357                 irq_exit();
 358
 359                 prof_counter(cpu) = prof_multiplier(cpu);
 360         }
 361         set_irq_regs(old_regs);
 362 }
 363
 364 extern unsigned int lvl14_resolution;
 365
 366 static void __cpuinit smp_setup_percpu_timer(void)
 367 {
 368         int cpu = hard_smp4d_processor_id();
 369
 370         prof_counter(cpu) = prof_multiplier(cpu) = 1;
 371         load_profile_irq(cpu, lvl14_resolution);
 372 }
 373
 374 void __init smp4d_blackbox_id(unsigned *addr)
 375 {
 376         int rd = *addr & 0x3e000000;
 377
 378         addr[0] = 0xc0800800 | rd;              /* lda [%g0] ASI_M_VIKING_TMP1, reg */
 379         addr[1] = 0x01000000;                   /* nop */
 380         addr[2] = 0x01000000;                   /* nop */
 381 }
 382
 383 void __init smp4d_blackbox_current(unsigned *addr)
 384 {
 385         int rd = *addr & 0x3e000000;
 386
 387         addr[0] = 0xc0800800 | rd;              /* lda [%g0] ASI_M_VIKING_TMP1, reg */
 388         addr[2] = 0x81282002 | rd | (rd >> 11); /* sll reg, 2, reg */
 389         addr[4] = 0x01000000;                   /* nop */
 390 }
 391
 392 void __init sun4d_init_smp(void)
 393 {
 394         int i;
 395         extern unsigned int t_nmi[], linux_trap_ipi15_sun4d[], linux_trap_ipi15_sun4m[];
 396
 397         /* Patch ipi15 trap table */
 398         t_nmi[1] = t_nmi[1] + (linux_trap_ipi15_sun4d - linux_trap_ipi15_sun4m);
 399
 400         /* And set btfixup... */
 401         BTFIXUPSET_BLACKBOX(hard_smp_processor_id, smp4d_blackbox_id);
 402         BTFIXUPSET_BLACKBOX(load_current, smp4d_blackbox_current);
 403         BTFIXUPSET_CALL(smp_cross_call, smp4d_cross_call, BTFIXUPCALL_NORM);
 404         BTFIXUPSET_CALL(__hard_smp_processor_id, __smp4d_processor_id, BTFIXUPCALL_NORM);
 405
 406         for (i = 0; i < NR_CPUS; i++) {
 407                 ccall_info.processors_in[i] = 1;
 408                 ccall_info.processors_out[i] = 1;
 409         }
 410 }