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 = 0;
  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                              "sta %%g6, [%%g0] %1\n\t"
 136                              : : "r" (&current_set[cpuid]), "i" (ASI_M_VIKING_TMP2)
 137                              : "memory" /* paranoid */);
 138
 139         cpu_leds[cpuid] = 0x9;
 140         show_leds(cpuid);
 141
 142         /* Attach to the address space of init_task. */
 143         atomic_inc(&init_mm.mm_count);
 144         current->active_mm = &init_mm;
 145
 146         local_flush_cache_all();
 147         local_flush_tlb_all();
 148
 149         __sti();        /* We don't allow PIL 14 yet */
 150
 151         while(!smp_commenced)
 152                 barrier();
 153
 154         spin_lock_irqsave(&sun4d_imsk_lock, flags);
 155         cc_set_imsk(cc_get_imsk() & ~0x4000); /* Allow PIL 14 as well */
 156         spin_unlock_irqrestore(&sun4d_imsk_lock, flags);
 157 }
 158
 159 extern int cpu_idle(void *unused);
 160 extern void init_IRQ(void);
 161 extern void cpu_panic(void);
 162 extern int start_secondary(void *unused);
 163
 164 /*
 165  *      Cycle through the processors asking the PROM to start each one.
 166  */
 167
 168 extern struct prom_cpuinfo linux_cpus[NR_CPUS];
 169 extern struct linux_prom_registers smp_penguin_ctable;
 170 extern unsigned long trapbase_cpu1[];
 171 extern unsigned long trapbase_cpu2[];
 172 extern unsigned long trapbase_cpu3[];
 173
 174 void __init smp4d_boot_cpus(void)
 175 {
 176         int cpucount = 0;
 177         int i = 0;
 178
 179         printk("Entering SMP Mode...\n");
 180
 181         for (i = 0; i < NR_CPUS; i++)
 182                 cpu_offset[i] = (char *)&cpu_data[i] - (char *)&cpu_data;
 183
 184         if (boot_cpu_id)
 185                 current_set[0] = NULL;
 186
 187         __sti();
 188         cpu_present_map = 0;
 189         for(i=0; i < linux_num_cpus; i++)
 190                 cpu_present_map |= (1<<linux_cpus[i].mid);
 191         SMP_PRINTK(("cpu_present_map %08lx\n", cpu_present_map));
 192         for(i=0; i < NR_CPUS; i++)
 193                 __cpu_number_map[i] = -1;
 194         for(i=0; i < NR_CPUS; i++)
 195                 __cpu_logical_map[i] = -1;
 196         for(i=0; i < NR_CPUS; i++)
 197                 mid_xlate[i] = i;
 198         __cpu_number_map[boot_cpu_id] = 0;
 199         __cpu_logical_map[0] = boot_cpu_id;
 200         current->processor = boot_cpu_id;
 201         smp_store_cpu_info(boot_cpu_id);
 202         smp_setup_percpu_timer();
 203         init_idle();
 204         local_flush_cache_all();
 205         if(linux_num_cpus == 1)
 206                 return;  /* Not an MP box. */
 207         SMP_PRINTK(("Iterating over CPUs\n"));
 208         for(i = 0; i < NR_CPUS; i++) {
 209                 if(i == boot_cpu_id)
 210                         continue;
 211
 212                 if(cpu_present_map & (1 << i)) {
 213                         extern unsigned long sun4d_cpu_startup;
 214                         unsigned long *entry = &sun4d_cpu_startup;
 215                         struct task_struct *p;
 216                         int timeout;
 217                         int no;
 218
 219                         /* Cook up an idler for this guy. */
 220                         kernel_thread(start_secondary, NULL, CLONE_PID);
 221
 222                         cpucount++;
 223
 224                         p = init_task.prev_task;
 225                         init_tasks[i] = p;
 226
 227                         p->processor = i;
 228                         p->has_cpu = 1; /* we schedule the first task manually */
 229
 230                         current_set[i] = p;
 231
 232                         del_from_runqueue(p);
 233                         unhash_process(p);
 234
 235                         for (no = 0; no < linux_num_cpus; no++)
 236                                 if (linux_cpus[no].mid == i)
 237                                         break;
 238
 239                         /*
 240                          * Initialize the contexts table
 241                          * Since the call to prom_startcpu() trashes the structure,
 242                          * we need to re-initialize it for each cpu
 243                          */
 244                         smp_penguin_ctable.which_io = 0;
 245                         smp_penguin_ctable.phys_addr = (unsigned int) srmmu_ctx_table_phys;
 246                         smp_penguin_ctable.reg_size = 0;
 247
 248                         /* whirrr, whirrr, whirrrrrrrrr... */
 249                         SMP_PRINTK(("Starting CPU %d at %p task %d node %08x\n", i, entry, cpucount, linux_cpus[no].prom_node));
 250                         local_flush_cache_all();
 251                         prom_startcpu(linux_cpus[no].prom_node,
 252                                       &smp_penguin_ctable, 0, (char *)entry);
 253
 254                         SMP_PRINTK(("prom_startcpu returned :)\n"));
 255
 256                         /* wheee... it's going... */
 257                         for(timeout = 0; timeout < 10000; timeout++) {
 258                                 if(cpu_callin_map[i])
 259                                         break;
 260                                 udelay(200);
 261                         }
 262
 263                         if(cpu_callin_map[i]) {
 264                                 /* Another "Red Snapper". */
 265                                 __cpu_number_map[i] = cpucount;
 266                                 __cpu_logical_map[cpucount] = i;
 267                         } else {
 268                                 cpucount--;
 269                                 printk("Processor %d is stuck.\n", i);
 270                         }
 271                 }
 272                 if(!(cpu_callin_map[i])) {
 273                         cpu_present_map &= ~(1 << i);
 274                         __cpu_number_map[i] = -1;
 275                 }
 276         }
 277         local_flush_cache_all();
 278         if(cpucount == 0) {
 279                 printk("Error: only one Processor found.\n");
 280                 cpu_present_map = (1 << hard_smp4d_processor_id());
 281         } else {
 282                 unsigned long bogosum = 0;
 283
 284                 for(i = 0; i < NR_CPUS; i++) {
 285                         if(cpu_present_map & (1 << i)) {
 286                                 bogosum += cpu_data[i].udelay_val;
 287                                 smp_highest_cpu = i;
 288                         }
 289                 }
 290                 SMP_PRINTK(("Total of %d Processors activated (%lu.%02lu BogoMIPS).\n", cpucount + 1, (bogosum + 2500)/500000, ((bogosum + 2500)/5000)%100));
 291                 printk("Total of %d Processors activated (%lu.%02lu BogoMIPS).\n",
 292                        cpucount + 1,
 293                        (bogosum + 2500)/500000,
 294                        ((bogosum + 2500)/5000)%100);
 295                 smp_activated = 1;
 296                 smp_num_cpus = cpucount + 1;
 297         }
 298
 299         /* Free unneeded trap tables */
 300         ClearPageReserved(virt_to_page(trapbase_cpu1));
 301         set_page_count(virt_to_page(trapbase_cpu1), 1);
 302         free_page((unsigned long)trapbase_cpu1);
 303         totalram_pages++;
 304         num_physpages++;
 305
 306         ClearPageReserved(virt_to_page(trapbase_cpu2));
 307         set_page_count(virt_to_page(trapbase_cpu2), 1);
 308         free_page((unsigned long)trapbase_cpu2);
 309         totalram_pages++;
 310         num_physpages++;
 311
 312         ClearPageReserved(virt_to_page(trapbase_cpu3));
 313         set_page_count(virt_to_page(trapbase_cpu3), 1);
 314         free_page((unsigned long)trapbase_cpu3);
 315         totalram_pages++;
 316         num_physpages++;
 317
 318         /* Ok, they are spinning and ready to go. */
 319         smp_processors_ready = 1;
 320         sun4d_distribute_irqs();
 321 }
 322
 323 static struct smp_funcall {
 324         smpfunc_t func;
 325         unsigned long arg1;
 326         unsigned long arg2;
 327         unsigned long arg3;
 328         unsigned long arg4;
 329         unsigned long arg5;
 330         unsigned char processors_in[NR_CPUS];  /* Set when ipi entered. */
 331         unsigned char processors_out[NR_CPUS]; /* Set when ipi exited. */
 332 } ccall_info __attribute__((aligned(8)));
 333
 334 static spinlock_t cross_call_lock = SPIN_LOCK_UNLOCKED;
 335
 336 /* Cross calls must be serialized, at least currently. */
 337 void smp4d_cross_call(smpfunc_t func, unsigned long arg1, unsigned long arg2,
 338                     unsigned long arg3, unsigned long arg4, unsigned long arg5)
 339 {
 340         if(smp_processors_ready) {
 341                 register int high = smp_highest_cpu;
 342                 unsigned long flags;
 343
 344                 spin_lock_irqsave(&cross_call_lock, flags);
 345
 346                 {
 347                         /* If you make changes here, make sure gcc generates proper code... */
 348                         smpfunc_t f asm("i0") = func;
 349                         unsigned long a1 asm("i1") = arg1;
 350                         unsigned long a2 asm("i2") = arg2;
 351                         unsigned long a3 asm("i3") = arg3;
 352                         unsigned long a4 asm("i4") = arg4;
 353                         unsigned long a5 asm("i5") = arg5;
 354
 355                         __asm__ __volatile__("
 356                                 std %0, [%6]
 357                                 std %2, [%6 + 8]
 358                                 std %4, [%6 + 16]" : :
 359                                 "r"(f), "r"(a1), "r"(a2), "r"(a3), "r"(a4), "r"(a5),
 360                                 "r" (&ccall_info.func));
 361                 }
 362
 363                 /* Init receive/complete mapping, plus fire the IPI's off. */
 364                 {
 365                         register unsigned long mask;
 366                         register int i;
 367
 368                         mask = (cpu_present_map & ~(1 << hard_smp4d_processor_id()));
 369                         for(i = 0; i <= high; i++) {
 370                                 if(mask & (1 << i)) {
 371                                         ccall_info.processors_in[i] = 0;
 372                                         ccall_info.processors_out[i] = 0;
 373                                         sun4d_send_ipi(i, IRQ_CROSS_CALL);
 374                                 }
 375                         }
 376                 }
 377
 378                 {
 379                         register int i;
 380
 381                         i = 0;
 382                         do {
 383                                 while(!ccall_info.processors_in[i])
 384                                         barrier();
 385                         } while(++i <= high);
 386
 387                         i = 0;
 388                         do {
 389                                 while(!ccall_info.processors_out[i])
 390                                         barrier();
 391                         } while(++i <= high);
 392                 }
 393
 394                 spin_unlock_irqrestore(&cross_call_lock, flags);
 395         }
 396 }
 397
 398 /* Running cross calls. */
 399 void smp4d_cross_call_irq(void)
 400 {
 401         int i = hard_smp4d_processor_id();
 402
 403         ccall_info.processors_in[i] = 1;
 404         ccall_info.func(ccall_info.arg1, ccall_info.arg2, ccall_info.arg3,
 405                         ccall_info.arg4, ccall_info.arg5);
 406         ccall_info.processors_out[i] = 1;
 407 }
 408
 409 static int smp4d_stop_cpu_sender;
 410
 411 static void smp4d_stop_cpu(void)
 412 {
 413         int me = hard_smp4d_processor_id();
 414
 415         if (me != smp4d_stop_cpu_sender)
 416                 while(1) barrier();
 417 }
 418
 419 /* Cross calls, in order to work efficiently and atomically do all
 420  * the message passing work themselves, only stopcpu and reschedule
 421  * messages come through here.
 422  */
 423 void smp4d_message_pass(int target, int msg, unsigned long data, int wait)
 424 {
 425         int me = hard_smp4d_processor_id();
 426
 427         SMP_PRINTK(("smp4d_message_pass %d %d %08lx %d\n", target, msg, data, wait));
 428         if (msg == MSG_STOP_CPU && target == MSG_ALL_BUT_SELF) {
 429                 unsigned long flags;
 430                 static spinlock_t stop_cpu_lock = SPIN_LOCK_UNLOCKED;
 431                 spin_lock_irqsave(&stop_cpu_lock, flags);
 432                 smp4d_stop_cpu_sender = me;
 433                 smp4d_cross_call((smpfunc_t)smp4d_stop_cpu, 0, 0, 0, 0, 0);
 434                 spin_unlock_irqrestore(&stop_cpu_lock, flags);
 435         }
 436         printk("Yeeee, trying to send SMP msg(%d) to %d on cpu %d\n", msg, target, me);
 437         panic("Bogon SMP message pass.");
 438 }
 439
 440 extern unsigned int prof_multiplier[NR_CPUS];
 441 extern unsigned int prof_counter[NR_CPUS];
 442
 443 extern void sparc_do_profile(unsigned long pc, unsigned long o7);
 444
 445 void smp4d_percpu_timer_interrupt(struct pt_regs *regs)
 446 {
 447         int cpu = hard_smp4d_processor_id();
 448         static int cpu_tick[NR_CPUS];
 449         static char led_mask[] = { 0xe, 0xd, 0xb, 0x7, 0xb, 0xd };
 450
 451         bw_get_prof_limit(cpu);
 452         bw_clear_intr_mask(0, 1);       /* INTR_TABLE[0] & 1 is Profile IRQ */
 453
 454         cpu_tick[cpu]++;
 455         if (!(cpu_tick[cpu] & 15)) {
 456                 if (cpu_tick[cpu] == 0x60)
 457                         cpu_tick[cpu] = 0;
 458                 cpu_leds[cpu] = led_mask[cpu_tick[cpu] >> 4];
 459                 show_leds(cpu);
 460         }
 461
 462         if(!user_mode(regs))
 463                 sparc_do_profile(regs->pc, regs->u_regs[UREG_RETPC]);
 464
 465         if(!--prof_counter[cpu]) {
 466                 int user = user_mode(regs);
 467
 468                 irq_enter(cpu, 0);
 469                 update_process_times(user);
 470                 irq_exit(cpu, 0);
 471
 472                 prof_counter[cpu] = prof_multiplier[cpu];
 473         }
 474 }
 475
 476 extern unsigned int lvl14_resolution;
 477
 478 static void __init smp_setup_percpu_timer(void)
 479 {
 480         int cpu = hard_smp4d_processor_id();
 481
 482         prof_counter[cpu] = prof_multiplier[cpu] = 1;
 483         load_profile_irq(cpu, lvl14_resolution);
 484 }
 485
 486 void __init smp4d_blackbox_id(unsigned *addr)
 487 {
 488         int rd = *addr & 0x3e000000;
 489
 490         addr[0] = 0xc0800800 | rd;              /* lda [%g0] ASI_M_VIKING_TMP1, reg */
 491         addr[1] = 0x01000000;                   /* nop */
 492         addr[2] = 0x01000000;                   /* nop */
 493 }
 494
 495 void __init smp4d_blackbox_current(unsigned *addr)
 496 {
 497         /* We have a nice Linux current register :) */
 498         int rd = addr[1] & 0x3e000000;
 499
 500         addr[0] = 0x10800006;                   /* b .+24 */
 501         addr[1] = 0xc0800820 | rd;              /* lda [%g0] ASI_M_VIKING_TMP2, reg */
 502 }
 503
 504 void __init sun4d_init_smp(void)
 505 {
 506         int i;
 507         extern unsigned int patchme_store_new_current[];
 508         extern unsigned int t_nmi[], linux_trap_ipi15_sun4d[], linux_trap_ipi15_sun4m[];
 509
 510         /* Store current into Linux current register :) */
 511         __asm__ __volatile__("sta %%g6, [%%g0] %0" : : "i"(ASI_M_VIKING_TMP2));
 512
 513         /* Patch switch_to */
 514         patchme_store_new_current[0] = (patchme_store_new_current[0] & 0x3e000000) | 0xc0a00820;
 515
 516         /* Patch ipi15 trap table */
 517         t_nmi[1] = t_nmi[1] + (linux_trap_ipi15_sun4d - linux_trap_ipi15_sun4m);
 518
 519         /* And set btfixup... */
 520         BTFIXUPSET_BLACKBOX(smp_processor_id, smp4d_blackbox_id);
 521         BTFIXUPSET_BLACKBOX(load_current, smp4d_blackbox_current);
 522         BTFIXUPSET_CALL(smp_cross_call, smp4d_cross_call, BTFIXUPCALL_NORM);
 523         BTFIXUPSET_CALL(smp_message_pass, smp4d_message_pass, BTFIXUPCALL_NORM);
 524         BTFIXUPSET_CALL(__smp_processor_id, __smp4d_processor_id, BTFIXUPCALL_NORM);
 525
 526         for (i = 0; i < NR_CPUS; i++) {
 527                 ccall_info.processors_in[i] = 1;
 528                 ccall_info.processors_out[i] = 1;
 529         }
 530 }