arch/ia64/kernel/smp.c

   1 /*
   2  * SMP Support
   3  *
   4  * Copyright (C) 1999 Walt Drummond <drummond@valinux.com>
   5  * Copyright (C) 1999, 2001, 2003 David Mosberger-Tang <davidm@hpl.hp.com>
   6  *
   7  * Lots of stuff stolen from arch/alpha/kernel/smp.c
   8  *
   9  * 01/05/16 Rohit Seth <rohit.seth@intel.com>  IA64-SMP functions. Reorganized
  10  * the existing code (on the lines of x86 port).
  11  * 00/09/11 David Mosberger <davidm@hpl.hp.com> Do loops_per_jiffy
  12  * calibration on each CPU.
  13  * 00/08/23 Asit Mallick <asit.k.mallick@intel.com> fixed logical processor id
  14  * 00/03/31 Rohit Seth <rohit.seth@intel.com>   Fixes for Bootstrap Processor
  15  * & cpu_online_map now gets done here (instead of setup.c)
  16  * 99/10/05 davidm      Update to bring it in sync with new command-line processing
  17  *  scheme.
  18  * 10/13/00 Goutham Rao <goutham.rao@intel.com> Updated smp_call_function and
  19  *              smp_call_function_single to resend IPI on timeouts
  20  */
  21 #include <linux/module.h>
  22 #include <linux/kernel.h>
  23 #include <linux/sched.h>
  24 #include <linux/init.h>
  25 #include <linux/interrupt.h>
  26 #include <linux/smp.h>
  27 #include <linux/kernel_stat.h>
  28 #include <linux/mm.h>
  29 #include <linux/cache.h>
  30 #include <linux/delay.h>
  31 #include <linux/efi.h>
  32 #include <linux/bitops.h>
  33 #include <linux/kexec.h>
  34
  35 #include <asm/atomic.h>
  36 #include <asm/current.h>
  37 #include <asm/delay.h>
  38 #include <asm/machvec.h>
  39 #include <asm/io.h>
  40 #include <asm/irq.h>
  41 #include <asm/page.h>
  42 #include <asm/pgalloc.h>
  43 #include <asm/pgtable.h>
  44 #include <asm/processor.h>
  45 #include <asm/ptrace.h>
  46 #include <asm/sal.h>
  47 #include <asm/system.h>
  48 #include <asm/tlbflush.h>
  49 #include <asm/unistd.h>
  50 #include <asm/mca.h>
  51
  52 /*
  53  * Note: alignment of 4 entries/cacheline was empirically determined
  54  * to be a good tradeoff between hot cachelines & spreading the array
  55  * across too many cacheline.
  56  */
  57 static struct local_tlb_flush_counts {
  58         unsigned int count;
  59 } __attribute__((__aligned__(32))) local_tlb_flush_counts[NR_CPUS];
  60
  61 static DEFINE_PER_CPU(unsigned int, shadow_flush_counts[NR_CPUS]) ____cacheline_aligned;
  62
  63
  64 /*
  65  * Structure and data for smp_call_function(). This is designed to minimise static memory
  66  * requirements. It also looks cleaner.
  67  */
  68 static  __cacheline_aligned DEFINE_SPINLOCK(call_lock);
  69
  70 struct call_data_struct {
  71         void (*func) (void *info);
  72         void *info;
  73         long wait;
  74         atomic_t started;
  75         atomic_t finished;
  76 };
  77
  78 static volatile struct call_data_struct *call_data;
  79
  80 #define IPI_CALL_FUNC           0
  81 #define IPI_CPU_STOP            1
  82 #define IPI_KDUMP_CPU_STOP      3
  83
  84 /* This needs to be cacheline aligned because it is written to by *other* CPUs.  */
  85 static DEFINE_PER_CPU(u64, ipi_operation) ____cacheline_aligned;
  86
  87 extern void cpu_halt (void);
  88
  89 void
  90 lock_ipi_calllock(void)
  91 {
  92         spin_lock_irq(&call_lock);
  93 }
  94
  95 void
  96 unlock_ipi_calllock(void)
  97 {
  98         spin_unlock_irq(&call_lock);
  99 }
 100
 101 static void
 102 stop_this_cpu (void)
 103 {
 104         /*
 105          * Remove this CPU:
 106          */
 107         cpu_clear(smp_processor_id(), cpu_online_map);
 108         max_xtp();
 109         local_irq_disable();
 110         cpu_halt();
 111 }
 112
 113 void
 114 cpu_die(void)
 115 {
 116         max_xtp();
 117         local_irq_disable();
 118         cpu_halt();
 119         /* Should never be here */
 120         BUG();
 121         for (;;);
 122 }
 123
 124 irqreturn_t
 125 handle_IPI (int irq, void *dev_id)
 126 {
 127         int this_cpu = get_cpu();
 128         unsigned long *pending_ipis = &__ia64_per_cpu_var(ipi_operation);
 129         unsigned long ops;
 130
 131         mb();   /* Order interrupt and bit testing. */
 132         while ((ops = xchg(pending_ipis, 0)) != 0) {
 133                 mb();   /* Order bit clearing and data access. */
 134                 do {
 135                         unsigned long which;
 136
 137                         which = ffz(~ops);
 138                         ops &= ~(1 << which);
 139
 140                         switch (which) {
 141                               case IPI_CALL_FUNC:
 142                               {
 143                                       struct call_data_struct *data;
 144                                       void (*func)(void *info);
 145                                       void *info;
 146                                       int wait;
 147
 148                                       /* release the 'pointer lock' */
 149                                       data = (struct call_data_struct *) call_data;
 150                                       func = data->func;
 151                                       info = data->info;
 152                                       wait = data->wait;
 153
 154                                       mb();
 155                                       atomic_inc(&data->started);
 156                                       /*
 157                                        * At this point the structure may be gone unless
 158                                        * wait is true.
 159                                        */
 160                                       (*func)(info);
 161
 162                                       /* Notify the sending CPU that the task is done.  */
 163                                       mb();
 164                                       if (wait)
 165                                               atomic_inc(&data->finished);
 166                               }
 167                               break;
 168
 169                               case IPI_CPU_STOP:
 170                                 stop_this_cpu();
 171                                 break;
 172 #ifdef CONFIG_KEXEC
 173                               case IPI_KDUMP_CPU_STOP:
 174                                 unw_init_running(kdump_cpu_freeze, NULL);
 175                                 break;
 176 #endif
 177                               default:
 178                                 printk(KERN_CRIT "Unknown IPI on CPU %d: %lu\n", this_cpu, which);
 179                                 break;
 180                         }
 181                 } while (ops);
 182                 mb();   /* Order data access and bit testing. */
 183         }
 184         put_cpu();
 185         return IRQ_HANDLED;
 186 }
 187
 188 /*
 189  * Called with preeemption disabled.
 190  */
 191 static inline void
 192 send_IPI_single (int dest_cpu, int op)
 193 {
 194         set_bit(op, &per_cpu(ipi_operation, dest_cpu));
 195         platform_send_ipi(dest_cpu, IA64_IPI_VECTOR, IA64_IPI_DM_INT, 0);
 196 }
 197
 198 /*
 199  * Called with preeemption disabled.
 200  */
 201 static inline void
 202 send_IPI_allbutself (int op)
 203 {
 204         unsigned int i;
 205
 206         for_each_online_cpu(i) {
 207                 if (i != smp_processor_id())
 208                         send_IPI_single(i, op);
 209         }
 210 }
 211
 212 /*
 213  * Called with preeemption disabled.
 214  */
 215 static inline void
 216 send_IPI_all (int op)
 217 {
 218         int i;
 219
 220         for_each_online_cpu(i) {
 221                 send_IPI_single(i, op);
 222         }
 223 }
 224
 225 /*
 226  * Called with preeemption disabled.
 227  */
 228 static inline void
 229 send_IPI_self (int op)
 230 {
 231         send_IPI_single(smp_processor_id(), op);
 232 }
 233
 234 #ifdef CONFIG_KEXEC
 235 void
 236 kdump_smp_send_stop(void)
 237 {
 238         send_IPI_allbutself(IPI_KDUMP_CPU_STOP);
 239 }
 240
 241 void
 242 kdump_smp_send_init(void)
 243 {
 244         unsigned int cpu, self_cpu;
 245         self_cpu = smp_processor_id();
 246         for_each_online_cpu(cpu) {
 247                 if (cpu != self_cpu) {
 248                         if(kdump_status[cpu] == 0)
 249                                 platform_send_ipi(cpu, 0, IA64_IPI_DM_INIT, 0);
 250                 }
 251         }
 252 }
 253 #endif
 254 /*
 255  * Called with preeemption disabled.
 256  */
 257 void
 258 smp_send_reschedule (int cpu)
 259 {
 260         platform_send_ipi(cpu, IA64_IPI_RESCHEDULE, IA64_IPI_DM_INT, 0);
 261 }
 262
 263 /*
 264  * Called with preeemption disabled.
 265  */
 266 static void
 267 smp_send_local_flush_tlb (int cpu)
 268 {
 269         platform_send_ipi(cpu, IA64_IPI_LOCAL_TLB_FLUSH, IA64_IPI_DM_INT, 0);
 270 }
 271
 272 void
 273 smp_local_flush_tlb(void)
 274 {
 275         /*
 276          * Use atomic ops. Otherwise, the load/increment/store sequence from
 277          * a "++" operation can have the line stolen between the load & store.
 278          * The overhead of the atomic op in negligible in this case & offers
 279          * significant benefit for the brief periods where lots of cpus
 280          * are simultaneously flushing TLBs.
 281          */
 282         ia64_fetchadd(1, &local_tlb_flush_counts[smp_processor_id()].count, acq);
 283         local_flush_tlb_all();
 284 }
 285
 286 #define FLUSH_DELAY     5 /* Usec backoff to eliminate excessive cacheline bouncing */
 287
 288 void
 289 smp_flush_tlb_cpumask(cpumask_t xcpumask)
 290 {
 291         unsigned int *counts = __ia64_per_cpu_var(shadow_flush_counts);
 292         cpumask_t cpumask = xcpumask;
 293         int mycpu, cpu, flush_mycpu = 0;
 294
 295         preempt_disable();
 296         mycpu = smp_processor_id();
 297
 298         for_each_cpu_mask(cpu, cpumask)
 299                 counts[cpu] = local_tlb_flush_counts[cpu].count;
 300
 301         mb();
 302         for_each_cpu_mask(cpu, cpumask) {
 303                 if (cpu == mycpu)
 304                         flush_mycpu = 1;
 305                 else
 306                         smp_send_local_flush_tlb(cpu);
 307         }
 308
 309         if (flush_mycpu)
 310                 smp_local_flush_tlb();
 311
 312         for_each_cpu_mask(cpu, cpumask)
 313                 while(counts[cpu] == local_tlb_flush_counts[cpu].count)
 314                         udelay(FLUSH_DELAY);
 315
 316         preempt_enable();
 317 }
 318
 319 void
 320 smp_flush_tlb_all (void)
 321 {
 322         on_each_cpu((void (*)(void *))local_flush_tlb_all, NULL, 1, 1);
 323 }
 324
 325 void
 326 smp_flush_tlb_mm (struct mm_struct *mm)
 327 {
 328         preempt_disable();
 329         /* this happens for the common case of a single-threaded fork():  */
 330         if (likely(mm == current->active_mm && atomic_read(&mm->mm_users) == 1))
 331         {
 332                 local_finish_flush_tlb_mm(mm);
 333                 preempt_enable();
 334                 return;
 335         }
 336
 337         preempt_enable();
 338         /*
 339          * We could optimize this further by using mm->cpu_vm_mask to track which CPUs
 340          * have been running in the address space.  It's not clear that this is worth the
 341          * trouble though: to avoid races, we have to raise the IPI on the target CPU
 342          * anyhow, and once a CPU is interrupted, the cost of local_flush_tlb_all() is
 343          * rather trivial.
 344          */
 345         on_each_cpu((void (*)(void *))local_finish_flush_tlb_mm, mm, 1, 1);
 346 }
 347
 348 /*
 349  * Run a function on another CPU
 350  *  <func>      The function to run. This must be fast and non-blocking.
 351  *  <info>      An arbitrary pointer to pass to the function.
 352  *  <nonatomic> Currently unused.
 353  *  <wait>      If true, wait until function has completed on other CPUs.
 354  *  [RETURNS]   0 on success, else a negative status code.
 355  *
 356  * Does not return until the remote CPU is nearly ready to execute <func>
 357  * or is or has executed.
 358  */
 359
 360 int
 361 smp_call_function_single (int cpuid, void (*func) (void *info), void *info, int nonatomic,
 362                           int wait)
 363 {
 364         struct call_data_struct data;
 365         int cpus = 1;
 366         int me = get_cpu(); /* prevent preemption and reschedule on another processor */
 367
 368         if (cpuid == me) {
 369                 printk(KERN_INFO "%s: trying to call self\n", __FUNCTION__);
 370                 put_cpu();
 371                 return -EBUSY;
 372         }
 373
 374         data.func = func;
 375         data.info = info;
 376         atomic_set(&data.started, 0);
 377         data.wait = wait;
 378         if (wait)
 379                 atomic_set(&data.finished, 0);
 380
 381         spin_lock_bh(&call_lock);
 382
 383         call_data = &data;
 384         mb();   /* ensure store to call_data precedes setting of IPI_CALL_FUNC */
 385         send_IPI_single(cpuid, IPI_CALL_FUNC);
 386
 387         /* Wait for response */
 388         while (atomic_read(&data.started) != cpus)
 389                 cpu_relax();
 390
 391         if (wait)
 392                 while (atomic_read(&data.finished) != cpus)
 393                         cpu_relax();
 394         call_data = NULL;
 395
 396         spin_unlock_bh(&call_lock);
 397         put_cpu();
 398         return 0;
 399 }
 400 EXPORT_SYMBOL(smp_call_function_single);
 401
 402 /*
 403  * this function sends a 'generic call function' IPI to all other CPUs
 404  * in the system.
 405  */
 406
 407 /*
 408  *  [SUMMARY]   Run a function on all other CPUs.
 409  *  <func>      The function to run. This must be fast and non-blocking.
 410  *  <info>      An arbitrary pointer to pass to the function.
 411  *  <nonatomic> currently unused.
 412  *  <wait>      If true, wait (atomically) until function has completed on other CPUs.
 413  *  [RETURNS]   0 on success, else a negative status code.
 414  *
 415  * Does not return until remote CPUs are nearly ready to execute <func> or are or have
 416  * executed.
 417  *
 418  * You must not call this function with disabled interrupts or from a
 419  * hardware interrupt handler or from a bottom half handler.
 420  */
 421 int
 422 smp_call_function (void (*func) (void *info), void *info, int nonatomic, int wait)
 423 {
 424         struct call_data_struct data;
 425         int cpus;
 426
 427         spin_lock(&call_lock);
 428         cpus = num_online_cpus() - 1;
 429         if (!cpus) {
 430                 spin_unlock(&call_lock);
 431                 return 0;
 432         }
 433
 434         /* Can deadlock when called with interrupts disabled */
 435         WARN_ON(irqs_disabled());
 436
 437         data.func = func;
 438         data.info = info;
 439         atomic_set(&data.started, 0);
 440         data.wait = wait;
 441         if (wait)
 442                 atomic_set(&data.finished, 0);
 443
 444         call_data = &data;
 445         mb();   /* ensure store to call_data precedes setting of IPI_CALL_FUNC */
 446         send_IPI_allbutself(IPI_CALL_FUNC);
 447
 448         /* Wait for response */
 449         while (atomic_read(&data.started) != cpus)
 450                 cpu_relax();
 451
 452         if (wait)
 453                 while (atomic_read(&data.finished) != cpus)
 454                         cpu_relax();
 455         call_data = NULL;
 456
 457         spin_unlock(&call_lock);
 458         return 0;
 459 }
 460 EXPORT_SYMBOL(smp_call_function);
 461
 462 /*
 463  * this function calls the 'stop' function on all other CPUs in the system.
 464  */
 465 void
 466 smp_send_stop (void)
 467 {
 468         send_IPI_allbutself(IPI_CPU_STOP);
 469 }
 470
 471 int __init
 472 setup_profiling_timer (unsigned int multiplier)
 473 {
 474         return -EINVAL;
 475 }