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 #define IPI_CALL_FUNC           0
  64 #define IPI_CPU_STOP            1
  65 #define IPI_CALL_FUNC_SINGLE    2
  66 #define IPI_KDUMP_CPU_STOP      3
  67
  68 /* This needs to be cacheline aligned because it is written to by *other* CPUs.  */
  69 static DEFINE_PER_CPU_SHARED_ALIGNED(u64, ipi_operation);
  70
  71 extern void cpu_halt (void);
  72
  73 static void
  74 stop_this_cpu(void)
  75 {
  76         /*
  77          * Remove this CPU:
  78          */
  79         cpu_clear(smp_processor_id(), cpu_online_map);
  80         max_xtp();
  81         local_irq_disable();
  82         cpu_halt();
  83 }
  84
  85 void
  86 cpu_die(void)
  87 {
  88         max_xtp();
  89         local_irq_disable();
  90         cpu_halt();
  91         /* Should never be here */
  92         BUG();
  93         for (;;);
  94 }
  95
  96 irqreturn_t
  97 handle_IPI (int irq, void *dev_id)
  98 {
  99         int this_cpu = get_cpu();
 100         unsigned long *pending_ipis = &__ia64_per_cpu_var(ipi_operation);
 101         unsigned long ops;
 102
 103         mb();   /* Order interrupt and bit testing. */
 104         while ((ops = xchg(pending_ipis, 0)) != 0) {
 105                 mb();   /* Order bit clearing and data access. */
 106                 do {
 107                         unsigned long which;
 108
 109                         which = ffz(~ops);
 110                         ops &= ~(1 << which);
 111
 112                         switch (which) {
 113                         case IPI_CPU_STOP:
 114                                 stop_this_cpu();
 115                                 break;
 116                         case IPI_CALL_FUNC:
 117                                 generic_smp_call_function_interrupt();
 118                                 break;
 119                         case IPI_CALL_FUNC_SINGLE:
 120                                 generic_smp_call_function_single_interrupt();
 121                                 break;
 122 #ifdef CONFIG_KEXEC
 123                         case IPI_KDUMP_CPU_STOP:
 124                                 unw_init_running(kdump_cpu_freeze, NULL);
 125                                 break;
 126 #endif
 127                         default:
 128                                 printk(KERN_CRIT "Unknown IPI on CPU %d: %lu\n",
 129                                                 this_cpu, which);
 130                                 break;
 131                         }
 132                 } while (ops);
 133                 mb();   /* Order data access and bit testing. */
 134         }
 135         put_cpu();
 136         return IRQ_HANDLED;
 137 }
 138
 139
 140
 141 /*
 142  * Called with preemption disabled.
 143  */
 144 static inline void
 145 send_IPI_single (int dest_cpu, int op)
 146 {
 147         set_bit(op, &per_cpu(ipi_operation, dest_cpu));
 148         platform_send_ipi(dest_cpu, IA64_IPI_VECTOR, IA64_IPI_DM_INT, 0);
 149 }
 150
 151 /*
 152  * Called with preemption disabled.
 153  */
 154 static inline void
 155 send_IPI_allbutself (int op)
 156 {
 157         unsigned int i;
 158
 159         for_each_online_cpu(i) {
 160                 if (i != smp_processor_id())
 161                         send_IPI_single(i, op);
 162         }
 163 }
 164
 165 /*
 166  * Called with preemption disabled.
 167  */
 168 static inline void
 169 send_IPI_mask(cpumask_t mask, int op)
 170 {
 171         unsigned int cpu;
 172
 173         for_each_cpu_mask(cpu, mask) {
 174                         send_IPI_single(cpu, op);
 175         }
 176 }
 177
 178 /*
 179  * Called with preemption disabled.
 180  */
 181 static inline void
 182 send_IPI_all (int op)
 183 {
 184         int i;
 185
 186         for_each_online_cpu(i) {
 187                 send_IPI_single(i, op);
 188         }
 189 }
 190
 191 /*
 192  * Called with preemption disabled.
 193  */
 194 static inline void
 195 send_IPI_self (int op)
 196 {
 197         send_IPI_single(smp_processor_id(), op);
 198 }
 199
 200 #ifdef CONFIG_KEXEC
 201 void
 202 kdump_smp_send_stop(void)
 203 {
 204         send_IPI_allbutself(IPI_KDUMP_CPU_STOP);
 205 }
 206
 207 void
 208 kdump_smp_send_init(void)
 209 {
 210         unsigned int cpu, self_cpu;
 211         self_cpu = smp_processor_id();
 212         for_each_online_cpu(cpu) {
 213                 if (cpu != self_cpu) {
 214                         if(kdump_status[cpu] == 0)
 215                                 platform_send_ipi(cpu, 0, IA64_IPI_DM_INIT, 0);
 216                 }
 217         }
 218 }
 219 #endif
 220 /*
 221  * Called with preemption disabled.
 222  */
 223 void
 224 smp_send_reschedule (int cpu)
 225 {
 226         platform_send_ipi(cpu, IA64_IPI_RESCHEDULE, IA64_IPI_DM_INT, 0);
 227 }
 228
 229 /*
 230  * Called with preemption disabled.
 231  */
 232 static void
 233 smp_send_local_flush_tlb (int cpu)
 234 {
 235         platform_send_ipi(cpu, IA64_IPI_LOCAL_TLB_FLUSH, IA64_IPI_DM_INT, 0);
 236 }
 237
 238 void
 239 smp_local_flush_tlb(void)
 240 {
 241         /*
 242          * Use atomic ops. Otherwise, the load/increment/store sequence from
 243          * a "++" operation can have the line stolen between the load & store.
 244          * The overhead of the atomic op in negligible in this case & offers
 245          * significant benefit for the brief periods where lots of cpus
 246          * are simultaneously flushing TLBs.
 247          */
 248         ia64_fetchadd(1, &local_tlb_flush_counts[smp_processor_id()].count, acq);
 249         local_flush_tlb_all();
 250 }
 251
 252 #define FLUSH_DELAY     5 /* Usec backoff to eliminate excessive cacheline bouncing */
 253
 254 void
 255 smp_flush_tlb_cpumask(cpumask_t xcpumask)
 256 {
 257         unsigned int *counts = __ia64_per_cpu_var(shadow_flush_counts);
 258         cpumask_t cpumask = xcpumask;
 259         int mycpu, cpu, flush_mycpu = 0;
 260
 261         preempt_disable();
 262         mycpu = smp_processor_id();
 263
 264         for_each_cpu_mask(cpu, cpumask)
 265                 counts[cpu] = local_tlb_flush_counts[cpu].count;
 266
 267         mb();
 268         for_each_cpu_mask(cpu, cpumask) {
 269                 if (cpu == mycpu)
 270                         flush_mycpu = 1;
 271                 else
 272                         smp_send_local_flush_tlb(cpu);
 273         }
 274
 275         if (flush_mycpu)
 276                 smp_local_flush_tlb();
 277
 278         for_each_cpu_mask(cpu, cpumask)
 279                 while(counts[cpu] == local_tlb_flush_counts[cpu].count)
 280                         udelay(FLUSH_DELAY);
 281
 282         preempt_enable();
 283 }
 284
 285 void
 286 smp_flush_tlb_all (void)
 287 {
 288         on_each_cpu((void (*)(void *))local_flush_tlb_all, NULL, 1);
 289 }
 290
 291 void
 292 smp_flush_tlb_mm (struct mm_struct *mm)
 293 {
 294         preempt_disable();
 295         /* this happens for the common case of a single-threaded fork():  */
 296         if (likely(mm == current->active_mm && atomic_read(&mm->mm_users) == 1))
 297         {
 298                 local_finish_flush_tlb_mm(mm);
 299                 preempt_enable();
 300                 return;
 301         }
 302
 303         preempt_enable();
 304         /*
 305          * We could optimize this further by using mm->cpu_vm_mask to track which CPUs
 306          * have been running in the address space.  It's not clear that this is worth the
 307          * trouble though: to avoid races, we have to raise the IPI on the target CPU
 308          * anyhow, and once a CPU is interrupted, the cost of local_flush_tlb_all() is
 309          * rather trivial.
 310          */
 311         on_each_cpu((void (*)(void *))local_finish_flush_tlb_mm, mm, 1);
 312 }
 313
 314 void arch_send_call_function_single_ipi(int cpu)
 315 {
 316         send_IPI_single(cpu, IPI_CALL_FUNC_SINGLE);
 317 }
 318
 319 void arch_send_call_function_ipi(cpumask_t mask)
 320 {
 321         send_IPI_mask(mask, IPI_CALL_FUNC);
 322 }
 323
 324 /*
 325  * this function calls the 'stop' function on all other CPUs in the system.
 326  */
 327 void
 328 smp_send_stop (void)
 329 {
 330         send_IPI_allbutself(IPI_CPU_STOP);
 331 }
 332
 333 int
 334 setup_profiling_timer (unsigned int multiplier)
 335 {
 336         return -EINVAL;
 337 }