arch/sparc/kernel/leon_smp.c

   1 /* leon_smp.c: Sparc-Leon SMP support.
   2  *
   3  * based on sun4m_smp.c
   4  * Copyright (C) 1996 David S. Miller (davem@caip.rutgers.edu)
   5  * Copyright (C) 2009 Daniel Hellstrom (daniel@gaisler.com) Aeroflex Gaisler AB
   6  * Copyright (C) 2009 Konrad Eisele (konrad@gaisler.com) Aeroflex Gaisler AB
   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/of.h>
  18 #include <linux/init.h>
  19 #include <linux/spinlock.h>
  20 #include <linux/mm.h>
  21 #include <linux/swap.h>
  22 #include <linux/profile.h>
  23 #include <linux/pm.h>
  24 #include <linux/delay.h>
  25 #include <linux/gfp.h>
  26 #include <linux/cpu.h>
  27
  28 #include <asm/cacheflush.h>
  29 #include <asm/tlbflush.h>
  30
  31 #include <asm/ptrace.h>
  32 #include <linux/atomic.h>
  33 #include <asm/irq_regs.h>
  34 #include <asm/traps.h>
  35
  36 #include <asm/delay.h>
  37 #include <asm/irq.h>
  38 #include <asm/page.h>
  39 #include <asm/pgalloc.h>
  40 #include <asm/pgtable.h>
  41 #include <asm/oplib.h>
  42 #include <asm/cpudata.h>
  43 #include <asm/asi.h>
  44 #include <asm/leon.h>
  45 #include <asm/leon_amba.h>
  46
  47 #include "kernel.h"
  48
  49 #ifdef CONFIG_SPARC_LEON
  50
  51 #include "irq.h"
  52
  53 extern ctxd_t *srmmu_ctx_table_phys;
  54 static int smp_processors_ready;
  55 extern volatile unsigned long cpu_callin_map[NR_CPUS];
  56 extern cpumask_t smp_commenced_mask;
  57 void __init leon_configure_cache_smp(void);
  58 static void leon_ipi_init(void);
  59
  60 /* IRQ number of LEON IPIs */
  61 int leon_ipi_irq = LEON3_IRQ_IPI_DEFAULT;
  62
  63 static inline unsigned long do_swap(volatile unsigned long *ptr,
  64                                     unsigned long val)
  65 {
  66         __asm__ __volatile__("swapa [%2] %3, %0\n\t" : "=&r"(val)
  67                              : "0"(val), "r"(ptr), "i"(ASI_LEON_DCACHE_MISS)
  68                              : "memory");
  69         return val;
  70 }
  71
  72 static void smp_setup_percpu_timer(void);
  73
  74 void __cpuinit leon_callin(void)
  75 {
  76         int cpuid = hard_smpleon_processor_id();
  77
  78         local_flush_cache_all();
  79         local_flush_tlb_all();
  80         leon_configure_cache_smp();
  81
  82         notify_cpu_starting(cpuid);
  83
  84         /* Get our local ticker going. */
  85         smp_setup_percpu_timer();
  86
  87         calibrate_delay();
  88         smp_store_cpu_info(cpuid);
  89
  90         local_flush_cache_all();
  91         local_flush_tlb_all();
  92
  93         /*
  94          * Unblock the master CPU _only_ when the scheduler state
  95          * of all secondary CPUs will be up-to-date, so after
  96          * the SMP initialization the master will be just allowed
  97          * to call the scheduler code.
  98          * Allow master to continue.
  99          */
 100         do_swap(&cpu_callin_map[cpuid], 1);
 101
 102         local_flush_cache_all();
 103         local_flush_tlb_all();
 104
 105         /* Fix idle thread fields. */
 106         __asm__ __volatile__("ld [%0], %%g6\n\t" : : "r"(&current_set[cpuid])
 107                              : "memory" /* paranoid */);
 108
 109         /* Attach to the address space of init_task. */
 110         atomic_inc(&init_mm.mm_count);
 111         current->active_mm = &init_mm;
 112
 113         while (!cpumask_test_cpu(cpuid, &smp_commenced_mask))
 114                 mb();
 115
 116         local_irq_enable();
 117         set_cpu_online(cpuid, true);
 118 }
 119
 120 /*
 121  *      Cycle through the processors asking the PROM to start each one.
 122  */
 123
 124 extern struct linux_prom_registers smp_penguin_ctable;
 125
 126 void __init leon_configure_cache_smp(void)
 127 {
 128         unsigned long cfg = sparc_leon3_get_dcachecfg();
 129         int me = smp_processor_id();
 130
 131         if (ASI_LEON3_SYSCTRL_CFG_SSIZE(cfg) > 4) {
 132                 printk(KERN_INFO "Note: SMP with snooping only works on 4k cache, found %dk(0x%x) on cpu %d, disabling caches\n",
 133                      (unsigned int)ASI_LEON3_SYSCTRL_CFG_SSIZE(cfg),
 134                      (unsigned int)cfg, (unsigned int)me);
 135                 sparc_leon3_disable_cache();
 136         } else {
 137                 if (cfg & ASI_LEON3_SYSCTRL_CFG_SNOOPING) {
 138                         sparc_leon3_enable_snooping();
 139                 } else {
 140                         printk(KERN_INFO "Note: You have to enable snooping in the vhdl model cpu %d, disabling caches\n",
 141                              me);
 142                         sparc_leon3_disable_cache();
 143                 }
 144         }
 145
 146         local_flush_cache_all();
 147         local_flush_tlb_all();
 148 }
 149
 150 void leon_smp_setbroadcast(unsigned int mask)
 151 {
 152         int broadcast =
 153             ((LEON3_BYPASS_LOAD_PA(&(leon3_irqctrl_regs->mpstatus)) >>
 154               LEON3_IRQMPSTATUS_BROADCAST) & 1);
 155         if (!broadcast) {
 156                 prom_printf("######## !!!! The irqmp-ctrl must have broadcast enabled, smp wont work !!!!! ####### nr cpus: %d\n",
 157                      leon_smp_nrcpus());
 158                 if (leon_smp_nrcpus() > 1) {
 159                         BUG();
 160                 } else {
 161                         prom_printf("continue anyway\n");
 162                         return;
 163                 }
 164         }
 165         LEON_BYPASS_STORE_PA(&(leon3_irqctrl_regs->mpbroadcast), mask);
 166 }
 167
 168 unsigned int leon_smp_getbroadcast(void)
 169 {
 170         unsigned int mask;
 171         mask = LEON_BYPASS_LOAD_PA(&(leon3_irqctrl_regs->mpbroadcast));
 172         return mask;
 173 }
 174
 175 int leon_smp_nrcpus(void)
 176 {
 177         int nrcpu =
 178             ((LEON3_BYPASS_LOAD_PA(&(leon3_irqctrl_regs->mpstatus)) >>
 179               LEON3_IRQMPSTATUS_CPUNR) & 0xf) + 1;
 180         return nrcpu;
 181 }
 182
 183 void __init leon_boot_cpus(void)
 184 {
 185         int nrcpu = leon_smp_nrcpus();
 186         int me = smp_processor_id();
 187
 188         /* Setup IPI */
 189         leon_ipi_init();
 190
 191         printk(KERN_INFO "%d:(%d:%d) cpus mpirq at 0x%x\n", (unsigned int)me,
 192                (unsigned int)nrcpu, (unsigned int)NR_CPUS,
 193                (unsigned int)&(leon3_irqctrl_regs->mpstatus));
 194
 195         leon_enable_irq_cpu(LEON3_IRQ_CROSS_CALL, me);
 196         leon_enable_irq_cpu(LEON3_IRQ_TICKER, me);
 197         leon_enable_irq_cpu(leon_ipi_irq, me);
 198
 199         leon_smp_setbroadcast(1 << LEON3_IRQ_TICKER);
 200
 201         leon_configure_cache_smp();
 202         smp_setup_percpu_timer();
 203         local_flush_cache_all();
 204
 205 }
 206
 207 int __cpuinit leon_boot_one_cpu(int i)
 208 {
 209
 210         struct task_struct *p;
 211         int timeout;
 212
 213         /* Cook up an idler for this guy. */
 214         p = fork_idle(i);
 215
 216         current_set[i] = task_thread_info(p);
 217
 218         /* See trampoline.S:leon_smp_cpu_startup for details...
 219          * Initialize the contexts table
 220          * Since the call to prom_startcpu() trashes the structure,
 221          * we need to re-initialize it for each cpu
 222          */
 223         smp_penguin_ctable.which_io = 0;
 224         smp_penguin_ctable.phys_addr = (unsigned int)srmmu_ctx_table_phys;
 225         smp_penguin_ctable.reg_size = 0;
 226
 227         /* whirrr, whirrr, whirrrrrrrrr... */
 228         printk(KERN_INFO "Starting CPU %d : (irqmp: 0x%x)\n", (unsigned int)i,
 229                (unsigned int)&leon3_irqctrl_regs->mpstatus);
 230         local_flush_cache_all();
 231
 232         /* Make sure all IRQs are of from the start for this new CPU */
 233         LEON_BYPASS_STORE_PA(&leon3_irqctrl_regs->mask[i], 0);
 234
 235         /* Wake one CPU */
 236         LEON_BYPASS_STORE_PA(&(leon3_irqctrl_regs->mpstatus), 1 << i);
 237
 238         /* wheee... it's going... */
 239         for (timeout = 0; timeout < 10000; timeout++) {
 240                 if (cpu_callin_map[i])
 241                         break;
 242                 udelay(200);
 243         }
 244         printk(KERN_INFO "Started CPU %d\n", (unsigned int)i);
 245
 246         if (!(cpu_callin_map[i])) {
 247                 printk(KERN_ERR "Processor %d is stuck.\n", i);
 248                 return -ENODEV;
 249         } else {
 250                 leon_enable_irq_cpu(LEON3_IRQ_CROSS_CALL, i);
 251                 leon_enable_irq_cpu(LEON3_IRQ_TICKER, i);
 252                 leon_enable_irq_cpu(leon_ipi_irq, i);
 253         }
 254
 255         local_flush_cache_all();
 256         return 0;
 257 }
 258
 259 void __init leon_smp_done(void)
 260 {
 261
 262         int i, first;
 263         int *prev;
 264
 265         /* setup cpu list for irq rotation */
 266         first = 0;
 267         prev = &first;
 268         for (i = 0; i < NR_CPUS; i++) {
 269                 if (cpu_online(i)) {
 270                         *prev = i;
 271                         prev = &cpu_data(i).next;
 272                 }
 273         }
 274         *prev = first;
 275         local_flush_cache_all();
 276
 277         /* Free unneeded trap tables */
 278         if (!cpu_present(1)) {
 279                 ClearPageReserved(virt_to_page(&trapbase_cpu1));
 280                 init_page_count(virt_to_page(&trapbase_cpu1));
 281                 free_page((unsigned long)&trapbase_cpu1);
 282                 totalram_pages++;
 283                 num_physpages++;
 284         }
 285         if (!cpu_present(2)) {
 286                 ClearPageReserved(virt_to_page(&trapbase_cpu2));
 287                 init_page_count(virt_to_page(&trapbase_cpu2));
 288                 free_page((unsigned long)&trapbase_cpu2);
 289                 totalram_pages++;
 290                 num_physpages++;
 291         }
 292         if (!cpu_present(3)) {
 293                 ClearPageReserved(virt_to_page(&trapbase_cpu3));
 294                 init_page_count(virt_to_page(&trapbase_cpu3));
 295                 free_page((unsigned long)&trapbase_cpu3);
 296                 totalram_pages++;
 297                 num_physpages++;
 298         }
 299         /* Ok, they are spinning and ready to go. */
 300         smp_processors_ready = 1;
 301
 302 }
 303
 304 void leon_irq_rotate(int cpu)
 305 {
 306 }
 307
 308 struct leon_ipi_work {
 309         int single;
 310         int msk;
 311         int resched;
 312 };
 313
 314 static DEFINE_PER_CPU_SHARED_ALIGNED(struct leon_ipi_work, leon_ipi_work);
 315
 316 /* Initialize IPIs on the LEON, in order to save IRQ resources only one IRQ
 317  * is used for all three types of IPIs.
 318  */
 319 static void __init leon_ipi_init(void)
 320 {
 321         int cpu, len;
 322         struct leon_ipi_work *work;
 323         struct property *pp;
 324         struct device_node *rootnp;
 325         struct tt_entry *trap_table;
 326         unsigned long flags;
 327
 328         /* Find IPI IRQ or stick with default value */
 329         rootnp = of_find_node_by_path("/ambapp0");
 330         if (rootnp) {
 331                 pp = of_find_property(rootnp, "ipi_num", &len);
 332                 if (pp && (*(int *)pp->value))
 333                         leon_ipi_irq = *(int *)pp->value;
 334         }
 335         printk(KERN_INFO "leon: SMP IPIs at IRQ %d\n", leon_ipi_irq);
 336
 337         /* Adjust so that we jump directly to smpleon_ipi */
 338         local_irq_save(flags);
 339         trap_table = &sparc_ttable[SP_TRAP_IRQ1 + (leon_ipi_irq - 1)];
 340         trap_table->inst_three += smpleon_ipi - real_irq_entry;
 341         local_flush_cache_all();
 342         local_irq_restore(flags);
 343
 344         for_each_possible_cpu(cpu) {
 345                 work = &per_cpu(leon_ipi_work, cpu);
 346                 work->single = work->msk = work->resched = 0;
 347         }
 348 }
 349
 350 static void leon_ipi_single(int cpu)
 351 {
 352         struct leon_ipi_work *work = &per_cpu(leon_ipi_work, cpu);
 353
 354         /* Mark work */
 355         work->single = 1;
 356
 357         /* Generate IRQ on the CPU */
 358         set_cpu_int(cpu, leon_ipi_irq);
 359 }
 360
 361 static void leon_ipi_mask_one(int cpu)
 362 {
 363         struct leon_ipi_work *work = &per_cpu(leon_ipi_work, cpu);
 364
 365         /* Mark work */
 366         work->msk = 1;
 367
 368         /* Generate IRQ on the CPU */
 369         set_cpu_int(cpu, leon_ipi_irq);
 370 }
 371
 372 static void leon_ipi_resched(int cpu)
 373 {
 374         struct leon_ipi_work *work = &per_cpu(leon_ipi_work, cpu);
 375
 376         /* Mark work */
 377         work->resched = 1;
 378
 379         /* Generate IRQ on the CPU (any IRQ will cause resched) */
 380         set_cpu_int(cpu, leon_ipi_irq);
 381 }
 382
 383 void leonsmp_ipi_interrupt(void)
 384 {
 385         struct leon_ipi_work *work = &__get_cpu_var(leon_ipi_work);
 386
 387         if (work->single) {
 388                 work->single = 0;
 389                 smp_call_function_single_interrupt();
 390         }
 391         if (work->msk) {
 392                 work->msk = 0;
 393                 smp_call_function_interrupt();
 394         }
 395         if (work->resched) {
 396                 work->resched = 0;
 397                 smp_resched_interrupt();
 398         }
 399 }
 400
 401 static struct smp_funcall {
 402         smpfunc_t func;
 403         unsigned long arg1;
 404         unsigned long arg2;
 405         unsigned long arg3;
 406         unsigned long arg4;
 407         unsigned long arg5;
 408         unsigned long processors_in[NR_CPUS];   /* Set when ipi entered. */
 409         unsigned long processors_out[NR_CPUS];  /* Set when ipi exited. */
 410 } ccall_info;
 411
 412 static DEFINE_SPINLOCK(cross_call_lock);
 413
 414 /* Cross calls must be serialized, at least currently. */
 415 static void leon_cross_call(smpfunc_t func, cpumask_t mask, unsigned long arg1,
 416                             unsigned long arg2, unsigned long arg3,
 417                             unsigned long arg4)
 418 {
 419         if (smp_processors_ready) {
 420                 register int high = NR_CPUS - 1;
 421                 unsigned long flags;
 422
 423                 spin_lock_irqsave(&cross_call_lock, flags);
 424
 425                 {
 426                         /* If you make changes here, make sure gcc generates proper code... */
 427                         register smpfunc_t f asm("i0") = func;
 428                         register unsigned long a1 asm("i1") = arg1;
 429                         register unsigned long a2 asm("i2") = arg2;
 430                         register unsigned long a3 asm("i3") = arg3;
 431                         register unsigned long a4 asm("i4") = arg4;
 432                         register unsigned long a5 asm("i5") = 0;
 433
 434                         __asm__ __volatile__("std %0, [%6]\n\t"
 435                                              "std %2, [%6 + 8]\n\t"
 436                                              "std %4, [%6 + 16]\n\t" : :
 437                                              "r"(f), "r"(a1), "r"(a2), "r"(a3),
 438                                              "r"(a4), "r"(a5),
 439                                              "r"(&ccall_info.func));
 440                 }
 441
 442                 /* Init receive/complete mapping, plus fire the IPI's off. */
 443                 {
 444                         register int i;
 445
 446                         cpumask_clear_cpu(smp_processor_id(), &mask);
 447                         cpumask_and(&mask, cpu_online_mask, &mask);
 448                         for (i = 0; i <= high; i++) {
 449                                 if (cpumask_test_cpu(i, &mask)) {
 450                                         ccall_info.processors_in[i] = 0;
 451                                         ccall_info.processors_out[i] = 0;
 452                                         set_cpu_int(i, LEON3_IRQ_CROSS_CALL);
 453
 454                                 }
 455                         }
 456                 }
 457
 458                 {
 459                         register int i;
 460
 461                         i = 0;
 462                         do {
 463                                 if (!cpumask_test_cpu(i, &mask))
 464                                         continue;
 465
 466                                 while (!ccall_info.processors_in[i])
 467                                         barrier();
 468                         } while (++i <= high);
 469
 470                         i = 0;
 471                         do {
 472                                 if (!cpumask_test_cpu(i, &mask))
 473                                         continue;
 474
 475                                 while (!ccall_info.processors_out[i])
 476                                         barrier();
 477                         } while (++i <= high);
 478                 }
 479
 480                 spin_unlock_irqrestore(&cross_call_lock, flags);
 481         }
 482 }
 483
 484 /* Running cross calls. */
 485 void leon_cross_call_irq(void)
 486 {
 487         int i = smp_processor_id();
 488
 489         ccall_info.processors_in[i] = 1;
 490         ccall_info.func(ccall_info.arg1, ccall_info.arg2, ccall_info.arg3,
 491                         ccall_info.arg4, ccall_info.arg5);
 492         ccall_info.processors_out[i] = 1;
 493 }
 494
 495 irqreturn_t leon_percpu_timer_interrupt(int irq, void *unused)
 496 {
 497         int cpu = smp_processor_id();
 498
 499         leon_clear_profile_irq(cpu);
 500
 501         profile_tick(CPU_PROFILING);
 502
 503         if (!--prof_counter(cpu)) {
 504                 int user = user_mode(get_irq_regs());
 505
 506                 update_process_times(user);
 507
 508                 prof_counter(cpu) = prof_multiplier(cpu);
 509         }
 510
 511         return IRQ_HANDLED;
 512 }
 513
 514 static void __init smp_setup_percpu_timer(void)
 515 {
 516         int cpu = smp_processor_id();
 517
 518         prof_counter(cpu) = prof_multiplier(cpu) = 1;
 519 }
 520
 521 void __init leon_blackbox_id(unsigned *addr)
 522 {
 523         int rd = *addr & 0x3e000000;
 524         int rs1 = rd >> 11;
 525
 526         /* patch places where ___b_hard_smp_processor_id appears */
 527         addr[0] = 0x81444000 | rd;      /* rd %asr17, reg */
 528         addr[1] = 0x8130201c | rd | rs1;        /* srl reg, 0x1c, reg */
 529         addr[2] = 0x01000000;   /* nop */
 530 }
 531
 532 void __init leon_blackbox_current(unsigned *addr)
 533 {
 534         int rd = *addr & 0x3e000000;
 535         int rs1 = rd >> 11;
 536
 537         /* patch LOAD_CURRENT macro where ___b_load_current appears */
 538         addr[0] = 0x81444000 | rd;      /* rd %asr17, reg */
 539         addr[2] = 0x8130201c | rd | rs1;        /* srl reg, 0x1c, reg */
 540         addr[4] = 0x81282002 | rd | rs1;        /* sll reg, 0x2, reg */
 541
 542 }
 543
 544 void __init leon_init_smp(void)
 545 {
 546         /* Patch ipi15 trap table */
 547         t_nmi[1] = t_nmi[1] + (linux_trap_ipi15_leon - linux_trap_ipi15_sun4m);
 548
 549         BTFIXUPSET_BLACKBOX(hard_smp_processor_id, leon_blackbox_id);
 550         BTFIXUPSET_BLACKBOX(load_current, leon_blackbox_current);
 551         BTFIXUPSET_CALL(smp_cross_call, leon_cross_call, BTFIXUPCALL_NORM);
 552         BTFIXUPSET_CALL(__hard_smp_processor_id, __leon_processor_id,
 553                         BTFIXUPCALL_NORM);
 554         BTFIXUPSET_CALL(smp_ipi_resched, leon_ipi_resched, BTFIXUPCALL_NORM);
 555         BTFIXUPSET_CALL(smp_ipi_single, leon_ipi_single, BTFIXUPCALL_NORM);
 556         BTFIXUPSET_CALL(smp_ipi_mask_one, leon_ipi_mask_one, BTFIXUPCALL_NORM);
 557 }
 558
 559 #endif /* CONFIG_SPARC_LEON */