arch/blackfin/mach-common/smp.c

   1 /*
   2  * IPI management based on arch/arm/kernel/smp.c (Copyright 2002 ARM Limited)
   3  *
   4  * Copyright 2007-2009 Analog Devices Inc.
   5  *                         Philippe Gerum <rpm@xenomai.org>
   6  *
   7  * Licensed under the GPL-2.
   8  */
   9
  10 #include <linux/module.h>
  11 #include <linux/delay.h>
  12 #include <linux/init.h>
  13 #include <linux/spinlock.h>
  14 #include <linux/sched.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/cache.h>
  17 #include <linux/profile.h>
  18 #include <linux/errno.h>
  19 #include <linux/mm.h>
  20 #include <linux/cpu.h>
  21 #include <linux/smp.h>
  22 #include <linux/cpumask.h>
  23 #include <linux/seq_file.h>
  24 #include <linux/irq.h>
  25 #include <linux/slab.h>
  26 #include <asm/atomic.h>
  27 #include <asm/cacheflush.h>
  28 #include <asm/mmu_context.h>
  29 #include <asm/pgtable.h>
  30 #include <asm/pgalloc.h>
  31 #include <asm/processor.h>
  32 #include <asm/ptrace.h>
  33 #include <asm/cpu.h>
  34 #include <asm/time.h>
  35 #include <linux/err.h>
  36
  37 /*
  38  * Anomaly notes:
  39  * 05000120 - we always define corelock as 32-bit integer in L2
  40  */
  41 struct corelock_slot corelock __attribute__ ((__section__(".l2.bss")));
  42
  43 void __cpuinitdata *init_retx_coreb, *init_saved_retx_coreb,
  44         *init_saved_seqstat_coreb, *init_saved_icplb_fault_addr_coreb,
  45         *init_saved_dcplb_fault_addr_coreb;
  46
  47 #define BFIN_IPI_RESCHEDULE   0
  48 #define BFIN_IPI_CALL_FUNC    1
  49 #define BFIN_IPI_CPU_STOP     2
  50
  51 struct blackfin_flush_data {
  52         unsigned long start;
  53         unsigned long end;
  54 };
  55
  56 void *secondary_stack;
  57
  58
  59 struct smp_call_struct {
  60         void (*func)(void *info);
  61         void *info;
  62         int wait;
  63         cpumask_t *waitmask;
  64 };
  65
  66 static struct blackfin_flush_data smp_flush_data;
  67
  68 static DEFINE_SPINLOCK(stop_lock);
  69
  70 struct ipi_message {
  71         unsigned long type;
  72         struct smp_call_struct call_struct;
  73 };
  74
  75 /* A magic number - stress test shows this is safe for common cases */
  76 #define BFIN_IPI_MSGQ_LEN 5
  77
  78 /* Simple FIFO buffer, overflow leads to panic */
  79 struct ipi_message_queue {
  80         spinlock_t lock;
  81         unsigned long count;
  82         unsigned long head; /* head of the queue */
  83         struct ipi_message ipi_message[BFIN_IPI_MSGQ_LEN];
  84 };
  85
  86 static DEFINE_PER_CPU(struct ipi_message_queue, ipi_msg_queue);
  87
  88 static void ipi_cpu_stop(unsigned int cpu)
  89 {
  90         spin_lock(&stop_lock);
  91         printk(KERN_CRIT "CPU%u: stopping\n", cpu);
  92         dump_stack();
  93         spin_unlock(&stop_lock);
  94
  95         cpu_clear(cpu, cpu_online_map);
  96
  97         local_irq_disable();
  98
  99         while (1)
 100                 SSYNC();
 101 }
 102
 103 static void ipi_flush_icache(void *info)
 104 {
 105         struct blackfin_flush_data *fdata = info;
 106
 107         /* Invalidate the memory holding the bounds of the flushed region. */
 108         invalidate_dcache_range((unsigned long)fdata,
 109                 (unsigned long)fdata + sizeof(*fdata));
 110
 111         flush_icache_range(fdata->start, fdata->end);
 112 }
 113
 114 static void ipi_call_function(unsigned int cpu, struct ipi_message *msg)
 115 {
 116         int wait;
 117         void (*func)(void *info);
 118         void *info;
 119         func = msg->call_struct.func;
 120         info = msg->call_struct.info;
 121         wait = msg->call_struct.wait;
 122         func(info);
 123         if (wait) {
 124 #ifdef __ARCH_SYNC_CORE_DCACHE
 125                 /*
 126                  * 'wait' usually means synchronization between CPUs.
 127                  * Invalidate D cache in case shared data was changed
 128                  * by func() to ensure cache coherence.
 129                  */
 130                 resync_core_dcache();
 131 #endif
 132                 cpu_clear(cpu, *msg->call_struct.waitmask);
 133         }
 134 }
 135
 136 /* Use IRQ_SUPPLE_0 to request reschedule.
 137  * When returning from interrupt to user space,
 138  * there is chance to reschedule */
 139 static irqreturn_t ipi_handler_int0(int irq, void *dev_instance)
 140 {
 141         unsigned int cpu = smp_processor_id();
 142
 143         platform_clear_ipi(cpu, IRQ_SUPPLE_0);
 144         return IRQ_HANDLED;
 145 }
 146
 147 static irqreturn_t ipi_handler_int1(int irq, void *dev_instance)
 148 {
 149         struct ipi_message *msg;
 150         struct ipi_message_queue *msg_queue;
 151         unsigned int cpu = smp_processor_id();
 152         unsigned long flags;
 153
 154         platform_clear_ipi(cpu, IRQ_SUPPLE_1);
 155
 156         msg_queue = &__get_cpu_var(ipi_msg_queue);
 157
 158         spin_lock_irqsave(&msg_queue->lock, flags);
 159
 160         while (msg_queue->count) {
 161                 msg = &msg_queue->ipi_message[msg_queue->head];
 162                 switch (msg->type) {
 163                 case BFIN_IPI_CALL_FUNC:
 164                         spin_unlock_irqrestore(&msg_queue->lock, flags);
 165                         ipi_call_function(cpu, msg);
 166                         spin_lock_irqsave(&msg_queue->lock, flags);
 167                         break;
 168                 case BFIN_IPI_CPU_STOP:
 169                         spin_unlock_irqrestore(&msg_queue->lock, flags);
 170                         ipi_cpu_stop(cpu);
 171                         spin_lock_irqsave(&msg_queue->lock, flags);
 172                         break;
 173                 default:
 174                         printk(KERN_CRIT "CPU%u: Unknown IPI message 0x%lx\n",
 175                                cpu, msg->type);
 176                         break;
 177                 }
 178                 msg_queue->head++;
 179                 msg_queue->head %= BFIN_IPI_MSGQ_LEN;
 180                 msg_queue->count--;
 181         }
 182         spin_unlock_irqrestore(&msg_queue->lock, flags);
 183         return IRQ_HANDLED;
 184 }
 185
 186 static void ipi_queue_init(void)
 187 {
 188         unsigned int cpu;
 189         struct ipi_message_queue *msg_queue;
 190         for_each_possible_cpu(cpu) {
 191                 msg_queue = &per_cpu(ipi_msg_queue, cpu);
 192                 spin_lock_init(&msg_queue->lock);
 193                 msg_queue->count = 0;
 194                 msg_queue->head = 0;
 195         }
 196 }
 197
 198 static inline void smp_send_message(cpumask_t callmap, unsigned long type,
 199                                         void (*func) (void *info), void *info, int wait)
 200 {
 201         unsigned int cpu;
 202         struct ipi_message_queue *msg_queue;
 203         struct ipi_message *msg;
 204         unsigned long flags, next_msg;
 205         cpumask_t waitmask = callmap; /* waitmask is shared by all cpus */
 206
 207         for_each_cpu_mask(cpu, callmap) {
 208                 msg_queue = &per_cpu(ipi_msg_queue, cpu);
 209                 spin_lock_irqsave(&msg_queue->lock, flags);
 210                 if (msg_queue->count < BFIN_IPI_MSGQ_LEN) {
 211                         next_msg = (msg_queue->head + msg_queue->count)
 212                                         % BFIN_IPI_MSGQ_LEN;
 213                         msg = &msg_queue->ipi_message[next_msg];
 214                         msg->type = type;
 215                         if (type == BFIN_IPI_CALL_FUNC) {
 216                                 msg->call_struct.func = func;
 217                                 msg->call_struct.info = info;
 218                                 msg->call_struct.wait = wait;
 219                                 msg->call_struct.waitmask = &waitmask;
 220                         }
 221                         msg_queue->count++;
 222                 } else
 223                         panic("IPI message queue overflow\n");
 224                 spin_unlock_irqrestore(&msg_queue->lock, flags);
 225                 platform_send_ipi_cpu(cpu, IRQ_SUPPLE_1);
 226         }
 227
 228         if (wait) {
 229                 while (!cpus_empty(waitmask))
 230                         blackfin_dcache_invalidate_range(
 231                                 (unsigned long)(&waitmask),
 232                                 (unsigned long)(&waitmask));
 233 #ifdef __ARCH_SYNC_CORE_DCACHE
 234                 /*
 235                  * Invalidate D cache in case shared data was changed by
 236                  * other processors to ensure cache coherence.
 237                  */
 238                 resync_core_dcache();
 239 #endif
 240         }
 241 }
 242
 243 int smp_call_function(void (*func)(void *info), void *info, int wait)
 244 {
 245         cpumask_t callmap;
 246
 247         preempt_disable();
 248         callmap = cpu_online_map;
 249         cpu_clear(smp_processor_id(), callmap);
 250         if (!cpus_empty(callmap))
 251                 smp_send_message(callmap, BFIN_IPI_CALL_FUNC, func, info, wait);
 252
 253         preempt_enable();
 254
 255         return 0;
 256 }
 257 EXPORT_SYMBOL_GPL(smp_call_function);
 258
 259 int smp_call_function_single(int cpuid, void (*func) (void *info), void *info,
 260                                 int wait)
 261 {
 262         unsigned int cpu = cpuid;
 263         cpumask_t callmap;
 264
 265         if (cpu_is_offline(cpu))
 266                 return 0;
 267         cpus_clear(callmap);
 268         cpu_set(cpu, callmap);
 269
 270         smp_send_message(callmap, BFIN_IPI_CALL_FUNC, func, info, wait);
 271
 272         return 0;
 273 }
 274 EXPORT_SYMBOL_GPL(smp_call_function_single);
 275
 276 void smp_send_reschedule(int cpu)
 277 {
 278         /* simply trigger an ipi */
 279         if (cpu_is_offline(cpu))
 280                 return;
 281         platform_send_ipi_cpu(cpu, IRQ_SUPPLE_0);
 282
 283         return;
 284 }
 285
 286 void smp_send_stop(void)
 287 {
 288         cpumask_t callmap;
 289
 290         preempt_disable();
 291         callmap = cpu_online_map;
 292         cpu_clear(smp_processor_id(), callmap);
 293         if (!cpus_empty(callmap))
 294                 smp_send_message(callmap, BFIN_IPI_CPU_STOP, NULL, NULL, 0);
 295
 296         preempt_enable();
 297
 298         return;
 299 }
 300
 301 int __cpuinit __cpu_up(unsigned int cpu)
 302 {
 303         int ret;
 304         static struct task_struct *idle;
 305
 306         if (idle)
 307                 free_task(idle);
 308
 309         idle = fork_idle(cpu);
 310         if (IS_ERR(idle)) {
 311                 printk(KERN_ERR "CPU%u: fork() failed\n", cpu);
 312                 return PTR_ERR(idle);
 313         }
 314
 315         secondary_stack = task_stack_page(idle) + THREAD_SIZE;
 316
 317         ret = platform_boot_secondary(cpu, idle);
 318
 319         secondary_stack = NULL;
 320
 321         return ret;
 322 }
 323
 324 static void __cpuinit setup_secondary(unsigned int cpu)
 325 {
 326         unsigned long ilat;
 327
 328         bfin_write_IMASK(0);
 329         CSYNC();
 330         ilat = bfin_read_ILAT();
 331         CSYNC();
 332         bfin_write_ILAT(ilat);
 333         CSYNC();
 334
 335         /* Enable interrupt levels IVG7-15. IARs have been already
 336          * programmed by the boot CPU.  */
 337         bfin_irq_flags |= IMASK_IVG15 |
 338             IMASK_IVG14 | IMASK_IVG13 | IMASK_IVG12 | IMASK_IVG11 |
 339             IMASK_IVG10 | IMASK_IVG9 | IMASK_IVG8 | IMASK_IVG7 | IMASK_IVGHW;
 340 }
 341
 342 void __cpuinit secondary_start_kernel(void)
 343 {
 344         unsigned int cpu = smp_processor_id();
 345         struct mm_struct *mm = &init_mm;
 346
 347         if (_bfin_swrst & SWRST_DBL_FAULT_B) {
 348                 printk(KERN_EMERG "CoreB Recovering from DOUBLE FAULT event\n");
 349 #ifdef CONFIG_DEBUG_DOUBLEFAULT
 350                 printk(KERN_EMERG " While handling exception (EXCAUSE = 0x%x) at %pF\n",
 351                         (int)init_saved_seqstat_coreb & SEQSTAT_EXCAUSE, init_saved_retx_coreb);
 352                 printk(KERN_NOTICE "   DCPLB_FAULT_ADDR: %pF\n", init_saved_dcplb_fault_addr_coreb);
 353                 printk(KERN_NOTICE "   ICPLB_FAULT_ADDR: %pF\n", init_saved_icplb_fault_addr_coreb);
 354 #endif
 355                 printk(KERN_NOTICE " The instruction at %pF caused a double exception\n",
 356                         init_retx_coreb);
 357         }
 358
 359         /*
 360          * We want the D-cache to be enabled early, in case the atomic
 361          * support code emulates cache coherence (see
 362          * __ARCH_SYNC_CORE_DCACHE).
 363          */
 364         init_exception_vectors();
 365
 366         local_irq_disable();
 367
 368         /* Attach the new idle task to the global mm. */
 369         atomic_inc(&mm->mm_users);
 370         atomic_inc(&mm->mm_count);
 371         current->active_mm = mm;
 372
 373         preempt_disable();
 374
 375         setup_secondary(cpu);
 376
 377         platform_secondary_init(cpu);
 378
 379         /* setup local core timer */
 380         bfin_local_timer_setup();
 381
 382         local_irq_enable();
 383
 384         bfin_setup_caches(cpu);
 385
 386         /*
 387          * Calibrate loops per jiffy value.
 388          * IRQs need to be enabled here - D-cache can be invalidated
 389          * in timer irq handler, so core B can read correct jiffies.
 390          */
 391         calibrate_delay();
 392
 393         cpu_idle();
 394 }
 395
 396 void __init smp_prepare_boot_cpu(void)
 397 {
 398 }
 399
 400 void __init smp_prepare_cpus(unsigned int max_cpus)
 401 {
 402         platform_prepare_cpus(max_cpus);
 403         ipi_queue_init();
 404         platform_request_ipi(IRQ_SUPPLE_0, ipi_handler_int0);
 405         platform_request_ipi(IRQ_SUPPLE_1, ipi_handler_int1);
 406 }
 407
 408 void __init smp_cpus_done(unsigned int max_cpus)
 409 {
 410         unsigned long bogosum = 0;
 411         unsigned int cpu;
 412
 413         for_each_online_cpu(cpu)
 414                 bogosum += loops_per_jiffy;
 415
 416         printk(KERN_INFO "SMP: Total of %d processors activated "
 417                "(%lu.%02lu BogoMIPS).\n",
 418                num_online_cpus(),
 419                bogosum / (500000/HZ),
 420                (bogosum / (5000/HZ)) % 100);
 421 }
 422
 423 void smp_icache_flush_range_others(unsigned long start, unsigned long end)
 424 {
 425         smp_flush_data.start = start;
 426         smp_flush_data.end = end;
 427
 428         if (smp_call_function(&ipi_flush_icache, &smp_flush_data, 0))
 429                 printk(KERN_WARNING "SMP: failed to run I-cache flush request on other CPUs\n");
 430 }
 431 EXPORT_SYMBOL_GPL(smp_icache_flush_range_others);
 432
 433 #ifdef __ARCH_SYNC_CORE_ICACHE
 434 unsigned long icache_invld_count[NR_CPUS];
 435 void resync_core_icache(void)
 436 {
 437         unsigned int cpu = get_cpu();
 438         blackfin_invalidate_entire_icache();
 439         icache_invld_count[cpu]++;
 440         put_cpu();
 441 }
 442 EXPORT_SYMBOL(resync_core_icache);
 443 #endif
 444
 445 #ifdef __ARCH_SYNC_CORE_DCACHE
 446 unsigned long dcache_invld_count[NR_CPUS];
 447 unsigned long barrier_mask __attribute__ ((__section__(".l2.bss")));
 448
 449 void resync_core_dcache(void)
 450 {
 451         unsigned int cpu = get_cpu();
 452         blackfin_invalidate_entire_dcache();
 453         dcache_invld_count[cpu]++;
 454         put_cpu();
 455 }
 456 EXPORT_SYMBOL(resync_core_dcache);
 457 #endif
 458
 459 #ifdef CONFIG_HOTPLUG_CPU
 460 int __cpuexit __cpu_disable(void)
 461 {
 462         unsigned int cpu = smp_processor_id();
 463
 464         if (cpu == 0)
 465                 return -EPERM;
 466
 467         set_cpu_online(cpu, false);
 468         return 0;
 469 }
 470
 471 static DECLARE_COMPLETION(cpu_killed);
 472
 473 int __cpuexit __cpu_die(unsigned int cpu)
 474 {
 475         return wait_for_completion_timeout(&cpu_killed, 5000);
 476 }
 477
 478 void cpu_die(void)
 479 {
 480         complete(&cpu_killed);
 481
 482         atomic_dec(&init_mm.mm_users);
 483         atomic_dec(&init_mm.mm_count);
 484
 485         local_irq_disable();
 486         platform_cpu_die();
 487 }
 488 #endif