Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 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 | |
| 34 | #include <asm/atomic.h> |
| 35 | #include <asm/current.h> |
| 36 | #include <asm/delay.h> |
| 37 | #include <asm/machvec.h> |
| 38 | #include <asm/io.h> |
| 39 | #include <asm/irq.h> |
| 40 | #include <asm/page.h> |
| 41 | #include <asm/pgalloc.h> |
| 42 | #include <asm/pgtable.h> |
| 43 | #include <asm/processor.h> |
| 44 | #include <asm/ptrace.h> |
| 45 | #include <asm/sal.h> |
| 46 | #include <asm/system.h> |
| 47 | #include <asm/tlbflush.h> |
| 48 | #include <asm/unistd.h> |
| 49 | #include <asm/mca.h> |
| 50 | |
| 51 | /* |
| 52 | * Structure and data for smp_call_function(). This is designed to minimise static memory |
| 53 | * requirements. It also looks cleaner. |
| 54 | */ |
| 55 | static __cacheline_aligned DEFINE_SPINLOCK(call_lock); |
| 56 | |
| 57 | struct call_data_struct { |
| 58 | void (*func) (void *info); |
| 59 | void *info; |
| 60 | long wait; |
| 61 | atomic_t started; |
| 62 | atomic_t finished; |
| 63 | }; |
| 64 | |
| 65 | static volatile struct call_data_struct *call_data; |
| 66 | |
| 67 | #define IPI_CALL_FUNC 0 |
| 68 | #define IPI_CPU_STOP 1 |
| 69 | |
| 70 | /* This needs to be cacheline aligned because it is written to by *other* CPUs. */ |
| 71 | static DEFINE_PER_CPU(u64, ipi_operation) ____cacheline_aligned; |
| 72 | |
| 73 | extern void cpu_halt (void); |
| 74 | |
| 75 | void |
| 76 | lock_ipi_calllock(void) |
| 77 | { |
| 78 | spin_lock_irq(&call_lock); |
| 79 | } |
| 80 | |
| 81 | void |
| 82 | unlock_ipi_calllock(void) |
| 83 | { |
| 84 | spin_unlock_irq(&call_lock); |
| 85 | } |
| 86 | |
| 87 | static void |
| 88 | stop_this_cpu (void) |
| 89 | { |
| 90 | /* |
| 91 | * Remove this CPU: |
| 92 | */ |
| 93 | cpu_clear(smp_processor_id(), cpu_online_map); |
| 94 | max_xtp(); |
| 95 | local_irq_disable(); |
| 96 | cpu_halt(); |
| 97 | } |
| 98 | |
| 99 | void |
| 100 | cpu_die(void) |
| 101 | { |
| 102 | max_xtp(); |
| 103 | local_irq_disable(); |
| 104 | cpu_halt(); |
| 105 | /* Should never be here */ |
| 106 | BUG(); |
| 107 | for (;;); |
| 108 | } |
| 109 | |
| 110 | irqreturn_t |
| 111 | handle_IPI (int irq, void *dev_id, struct pt_regs *regs) |
| 112 | { |
| 113 | int this_cpu = get_cpu(); |
| 114 | unsigned long *pending_ipis = &__ia64_per_cpu_var(ipi_operation); |
| 115 | unsigned long ops; |
| 116 | |
| 117 | mb(); /* Order interrupt and bit testing. */ |
| 118 | while ((ops = xchg(pending_ipis, 0)) != 0) { |
| 119 | mb(); /* Order bit clearing and data access. */ |
| 120 | do { |
| 121 | unsigned long which; |
| 122 | |
| 123 | which = ffz(~ops); |
| 124 | ops &= ~(1 << which); |
| 125 | |
| 126 | switch (which) { |
| 127 | case IPI_CALL_FUNC: |
| 128 | { |
| 129 | struct call_data_struct *data; |
| 130 | void (*func)(void *info); |
| 131 | void *info; |
| 132 | int wait; |
| 133 | |
| 134 | /* release the 'pointer lock' */ |
| 135 | data = (struct call_data_struct *) call_data; |
| 136 | func = data->func; |
| 137 | info = data->info; |
| 138 | wait = data->wait; |
| 139 | |
| 140 | mb(); |
| 141 | atomic_inc(&data->started); |
| 142 | /* |
| 143 | * At this point the structure may be gone unless |
| 144 | * wait is true. |
| 145 | */ |
| 146 | (*func)(info); |
| 147 | |
| 148 | /* Notify the sending CPU that the task is done. */ |
| 149 | mb(); |
| 150 | if (wait) |
| 151 | atomic_inc(&data->finished); |
| 152 | } |
| 153 | break; |
| 154 | |
| 155 | case IPI_CPU_STOP: |
| 156 | stop_this_cpu(); |
| 157 | break; |
| 158 | |
| 159 | default: |
| 160 | printk(KERN_CRIT "Unknown IPI on CPU %d: %lu\n", this_cpu, which); |
| 161 | break; |
| 162 | } |
| 163 | } while (ops); |
| 164 | mb(); /* Order data access and bit testing. */ |
| 165 | } |
| 166 | put_cpu(); |
| 167 | return IRQ_HANDLED; |
| 168 | } |
| 169 | |
| 170 | /* |
| 171 | * Called with preeemption disabled. |
| 172 | */ |
| 173 | static inline void |
| 174 | send_IPI_single (int dest_cpu, int op) |
| 175 | { |
| 176 | set_bit(op, &per_cpu(ipi_operation, dest_cpu)); |
| 177 | platform_send_ipi(dest_cpu, IA64_IPI_VECTOR, IA64_IPI_DM_INT, 0); |
| 178 | } |
| 179 | |
| 180 | /* |
| 181 | * Called with preeemption disabled. |
| 182 | */ |
| 183 | static inline void |
| 184 | send_IPI_allbutself (int op) |
| 185 | { |
| 186 | unsigned int i; |
| 187 | |
| 188 | for (i = 0; i < NR_CPUS; i++) { |
| 189 | if (cpu_online(i) && i != smp_processor_id()) |
| 190 | send_IPI_single(i, op); |
| 191 | } |
| 192 | } |
| 193 | |
| 194 | /* |
| 195 | * Called with preeemption disabled. |
| 196 | */ |
| 197 | static inline void |
| 198 | send_IPI_all (int op) |
| 199 | { |
| 200 | int i; |
| 201 | |
| 202 | for (i = 0; i < NR_CPUS; i++) |
| 203 | if (cpu_online(i)) |
| 204 | send_IPI_single(i, op); |
| 205 | } |
| 206 | |
| 207 | /* |
| 208 | * Called with preeemption disabled. |
| 209 | */ |
| 210 | static inline void |
| 211 | send_IPI_self (int op) |
| 212 | { |
| 213 | send_IPI_single(smp_processor_id(), op); |
| 214 | } |
| 215 | |
| 216 | /* |
| 217 | * Called with preeemption disabled. |
| 218 | */ |
| 219 | void |
| 220 | smp_send_reschedule (int cpu) |
| 221 | { |
| 222 | platform_send_ipi(cpu, IA64_IPI_RESCHEDULE, IA64_IPI_DM_INT, 0); |
| 223 | } |
| 224 | |
| 225 | void |
| 226 | smp_flush_tlb_all (void) |
| 227 | { |
| 228 | on_each_cpu((void (*)(void *))local_flush_tlb_all, NULL, 1, 1); |
| 229 | } |
| 230 | |
| 231 | void |
| 232 | smp_flush_tlb_mm (struct mm_struct *mm) |
| 233 | { |
| 234 | /* this happens for the common case of a single-threaded fork(): */ |
| 235 | if (likely(mm == current->active_mm && atomic_read(&mm->mm_users) == 1)) |
| 236 | { |
| 237 | local_finish_flush_tlb_mm(mm); |
| 238 | return; |
| 239 | } |
| 240 | |
| 241 | /* |
| 242 | * We could optimize this further by using mm->cpu_vm_mask to track which CPUs |
| 243 | * have been running in the address space. It's not clear that this is worth the |
| 244 | * trouble though: to avoid races, we have to raise the IPI on the target CPU |
| 245 | * anyhow, and once a CPU is interrupted, the cost of local_flush_tlb_all() is |
| 246 | * rather trivial. |
| 247 | */ |
| 248 | on_each_cpu((void (*)(void *))local_finish_flush_tlb_mm, mm, 1, 1); |
| 249 | } |
| 250 | |
| 251 | /* |
| 252 | * Run a function on another CPU |
| 253 | * <func> The function to run. This must be fast and non-blocking. |
| 254 | * <info> An arbitrary pointer to pass to the function. |
| 255 | * <nonatomic> Currently unused. |
| 256 | * <wait> If true, wait until function has completed on other CPUs. |
| 257 | * [RETURNS] 0 on success, else a negative status code. |
| 258 | * |
| 259 | * Does not return until the remote CPU is nearly ready to execute <func> |
| 260 | * or is or has executed. |
| 261 | */ |
| 262 | |
| 263 | int |
| 264 | smp_call_function_single (int cpuid, void (*func) (void *info), void *info, int nonatomic, |
| 265 | int wait) |
| 266 | { |
| 267 | struct call_data_struct data; |
| 268 | int cpus = 1; |
| 269 | int me = get_cpu(); /* prevent preemption and reschedule on another processor */ |
| 270 | |
| 271 | if (cpuid == me) { |
Christophe Lucas | 52a0de2 | 2005-06-20 14:50:00 -0700 | [diff] [blame^] | 272 | printk(KERN_INFO "%s: trying to call self\n", __FUNCTION__); |
Linus Torvalds | 1da177e | 2005-04-16 15:20:36 -0700 | [diff] [blame] | 273 | put_cpu(); |
| 274 | return -EBUSY; |
| 275 | } |
| 276 | |
| 277 | data.func = func; |
| 278 | data.info = info; |
| 279 | atomic_set(&data.started, 0); |
| 280 | data.wait = wait; |
| 281 | if (wait) |
| 282 | atomic_set(&data.finished, 0); |
| 283 | |
| 284 | spin_lock_bh(&call_lock); |
| 285 | |
| 286 | call_data = &data; |
| 287 | mb(); /* ensure store to call_data precedes setting of IPI_CALL_FUNC */ |
| 288 | send_IPI_single(cpuid, IPI_CALL_FUNC); |
| 289 | |
| 290 | /* Wait for response */ |
| 291 | while (atomic_read(&data.started) != cpus) |
| 292 | cpu_relax(); |
| 293 | |
| 294 | if (wait) |
| 295 | while (atomic_read(&data.finished) != cpus) |
| 296 | cpu_relax(); |
| 297 | call_data = NULL; |
| 298 | |
| 299 | spin_unlock_bh(&call_lock); |
| 300 | put_cpu(); |
| 301 | return 0; |
| 302 | } |
| 303 | EXPORT_SYMBOL(smp_call_function_single); |
| 304 | |
| 305 | /* |
| 306 | * this function sends a 'generic call function' IPI to all other CPUs |
| 307 | * in the system. |
| 308 | */ |
| 309 | |
| 310 | /* |
| 311 | * [SUMMARY] Run a function on all other CPUs. |
| 312 | * <func> The function to run. This must be fast and non-blocking. |
| 313 | * <info> An arbitrary pointer to pass to the function. |
| 314 | * <nonatomic> currently unused. |
| 315 | * <wait> If true, wait (atomically) until function has completed on other CPUs. |
| 316 | * [RETURNS] 0 on success, else a negative status code. |
| 317 | * |
| 318 | * Does not return until remote CPUs are nearly ready to execute <func> or are or have |
| 319 | * executed. |
| 320 | * |
| 321 | * You must not call this function with disabled interrupts or from a |
| 322 | * hardware interrupt handler or from a bottom half handler. |
| 323 | */ |
| 324 | int |
| 325 | smp_call_function (void (*func) (void *info), void *info, int nonatomic, int wait) |
| 326 | { |
| 327 | struct call_data_struct data; |
| 328 | int cpus = num_online_cpus()-1; |
| 329 | |
| 330 | if (!cpus) |
| 331 | return 0; |
| 332 | |
| 333 | /* Can deadlock when called with interrupts disabled */ |
| 334 | WARN_ON(irqs_disabled()); |
| 335 | |
| 336 | data.func = func; |
| 337 | data.info = info; |
| 338 | atomic_set(&data.started, 0); |
| 339 | data.wait = wait; |
| 340 | if (wait) |
| 341 | atomic_set(&data.finished, 0); |
| 342 | |
| 343 | spin_lock(&call_lock); |
| 344 | |
| 345 | call_data = &data; |
| 346 | mb(); /* ensure store to call_data precedes setting of IPI_CALL_FUNC */ |
| 347 | send_IPI_allbutself(IPI_CALL_FUNC); |
| 348 | |
| 349 | /* Wait for response */ |
| 350 | while (atomic_read(&data.started) != cpus) |
| 351 | cpu_relax(); |
| 352 | |
| 353 | if (wait) |
| 354 | while (atomic_read(&data.finished) != cpus) |
| 355 | cpu_relax(); |
| 356 | call_data = NULL; |
| 357 | |
| 358 | spin_unlock(&call_lock); |
| 359 | return 0; |
| 360 | } |
| 361 | EXPORT_SYMBOL(smp_call_function); |
| 362 | |
| 363 | /* |
| 364 | * this function calls the 'stop' function on all other CPUs in the system. |
| 365 | */ |
| 366 | void |
| 367 | smp_send_stop (void) |
| 368 | { |
| 369 | send_IPI_allbutself(IPI_CPU_STOP); |
| 370 | } |
| 371 | |
| 372 | int __init |
| 373 | setup_profiling_timer (unsigned int multiplier) |
| 374 | { |
| 375 | return -EINVAL; |
| 376 | } |