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