| 1 | /* $NetBSD: subr_ipi.c,v 1.3 2015/01/18 23:16:35 rmind Exp $ */ |
| 2 | |
| 3 | /*- |
| 4 | * Copyright (c) 2014 The NetBSD Foundation, Inc. |
| 5 | * All rights reserved. |
| 6 | * |
| 7 | * This code is derived from software contributed to The NetBSD Foundation |
| 8 | * by Mindaugas Rasiukevicius. |
| 9 | * |
| 10 | * Redistribution and use in source and binary forms, with or without |
| 11 | * modification, are permitted provided that the following conditions |
| 12 | * are met: |
| 13 | * 1. Redistributions of source code must retain the above copyright |
| 14 | * notice, this list of conditions and the following disclaimer. |
| 15 | * 2. Redistributions in binary form must reproduce the above copyright |
| 16 | * notice, this list of conditions and the following disclaimer in the |
| 17 | * documentation and/or other materials provided with the distribution. |
| 18 | * |
| 19 | * THIS SOFTWARE IS PROVIDED BY THE NETBSD FOUNDATION, INC. AND CONTRIBUTORS |
| 20 | * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED |
| 21 | * TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR |
| 22 | * PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE FOUNDATION OR CONTRIBUTORS |
| 23 | * BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR |
| 24 | * CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF |
| 25 | * SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS |
| 26 | * INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN |
| 27 | * CONTRACT, STRICT LIABILITY, OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) |
| 28 | * ARISING IN ANY WAY OUT OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE |
| 29 | * POSSIBILITY OF SUCH DAMAGE. |
| 30 | */ |
| 31 | |
| 32 | /* |
| 33 | * Inter-processor interrupt (IPI) interface: asynchronous IPIs to |
| 34 | * invoke functions with a constant argument and synchronous IPIs |
| 35 | * with the cross-call support. |
| 36 | */ |
| 37 | |
| 38 | #include <sys/cdefs.h> |
| 39 | __KERNEL_RCSID(0, "$NetBSD: subr_ipi.c,v 1.3 2015/01/18 23:16:35 rmind Exp $" ); |
| 40 | |
| 41 | #include <sys/param.h> |
| 42 | #include <sys/types.h> |
| 43 | |
| 44 | #include <sys/atomic.h> |
| 45 | #include <sys/evcnt.h> |
| 46 | #include <sys/cpu.h> |
| 47 | #include <sys/ipi.h> |
| 48 | #include <sys/intr.h> |
| 49 | #include <sys/kcpuset.h> |
| 50 | #include <sys/kmem.h> |
| 51 | #include <sys/lock.h> |
| 52 | #include <sys/mutex.h> |
| 53 | |
| 54 | /* |
| 55 | * An array of the IPI handlers used for asynchronous invocation. |
| 56 | * The lock protects the slot allocation. |
| 57 | */ |
| 58 | |
| 59 | typedef struct { |
| 60 | ipi_func_t func; |
| 61 | void * arg; |
| 62 | } ipi_intr_t; |
| 63 | |
| 64 | static kmutex_t ipi_mngmt_lock; |
| 65 | static ipi_intr_t ipi_intrs[IPI_MAXREG] __cacheline_aligned; |
| 66 | |
| 67 | /* |
| 68 | * Per-CPU mailbox for IPI messages: it is a single cache line storing |
| 69 | * up to IPI_MSG_MAX messages. This interface is built on top of the |
| 70 | * synchronous IPIs. |
| 71 | */ |
| 72 | |
| 73 | #define IPI_MSG_SLOTS (CACHE_LINE_SIZE / sizeof(ipi_msg_t *)) |
| 74 | #define IPI_MSG_MAX IPI_MSG_SLOTS |
| 75 | |
| 76 | typedef struct { |
| 77 | ipi_msg_t * msg[IPI_MSG_SLOTS]; |
| 78 | } ipi_mbox_t; |
| 79 | |
| 80 | |
| 81 | /* Mailboxes for the synchronous IPIs. */ |
| 82 | static ipi_mbox_t * ipi_mboxes __read_mostly; |
| 83 | static struct evcnt ipi_mboxfull_ev __cacheline_aligned; |
| 84 | static void ipi_msg_cpu_handler(void *); |
| 85 | |
| 86 | /* Handler for the synchronous IPIs - it must be zero. */ |
| 87 | #define IPI_SYNCH_ID 0 |
| 88 | |
| 89 | #ifndef MULTIPROCESSOR |
| 90 | #define cpu_ipi(ci) KASSERT(ci == NULL) |
| 91 | #endif |
| 92 | |
| 93 | void |
| 94 | ipi_sysinit(void) |
| 95 | { |
| 96 | const size_t len = ncpu * sizeof(ipi_mbox_t); |
| 97 | |
| 98 | /* Initialise the per-CPU bit fields. */ |
| 99 | for (u_int i = 0; i < ncpu; i++) { |
| 100 | struct cpu_info *ci = cpu_lookup(i); |
| 101 | memset(&ci->ci_ipipend, 0, sizeof(ci->ci_ipipend)); |
| 102 | } |
| 103 | mutex_init(&ipi_mngmt_lock, MUTEX_DEFAULT, IPL_NONE); |
| 104 | memset(ipi_intrs, 0, sizeof(ipi_intrs)); |
| 105 | |
| 106 | /* Allocate per-CPU IPI mailboxes. */ |
| 107 | ipi_mboxes = kmem_zalloc(len, KM_SLEEP); |
| 108 | KASSERT(ipi_mboxes != NULL); |
| 109 | |
| 110 | /* |
| 111 | * Register the handler for synchronous IPIs. This mechanism |
| 112 | * is built on top of the asynchronous interface. Slot zero is |
| 113 | * reserved permanently; it is also handy to use zero as a failure |
| 114 | * for other registers (as it is potentially less error-prone). |
| 115 | */ |
| 116 | ipi_intrs[IPI_SYNCH_ID].func = ipi_msg_cpu_handler; |
| 117 | |
| 118 | evcnt_attach_dynamic(&ipi_mboxfull_ev, EVCNT_TYPE_MISC, NULL, |
| 119 | "ipi" , "full" ); |
| 120 | } |
| 121 | |
| 122 | /* |
| 123 | * ipi_register: register an asynchronous IPI handler. |
| 124 | * |
| 125 | * => Returns IPI ID which is greater than zero; on failure - zero. |
| 126 | */ |
| 127 | u_int |
| 128 | ipi_register(ipi_func_t func, void *arg) |
| 129 | { |
| 130 | mutex_enter(&ipi_mngmt_lock); |
| 131 | for (u_int i = 0; i < IPI_MAXREG; i++) { |
| 132 | if (ipi_intrs[i].func == NULL) { |
| 133 | /* Register the function. */ |
| 134 | ipi_intrs[i].func = func; |
| 135 | ipi_intrs[i].arg = arg; |
| 136 | mutex_exit(&ipi_mngmt_lock); |
| 137 | |
| 138 | KASSERT(i != IPI_SYNCH_ID); |
| 139 | return i; |
| 140 | } |
| 141 | } |
| 142 | mutex_exit(&ipi_mngmt_lock); |
| 143 | printf("WARNING: ipi_register: table full, increase IPI_MAXREG\n" ); |
| 144 | return 0; |
| 145 | } |
| 146 | |
| 147 | /* |
| 148 | * ipi_unregister: release the IPI handler given the ID. |
| 149 | */ |
| 150 | void |
| 151 | ipi_unregister(u_int ipi_id) |
| 152 | { |
| 153 | ipi_msg_t ipimsg = { .func = (ipi_func_t)nullop }; |
| 154 | |
| 155 | KASSERT(ipi_id != IPI_SYNCH_ID); |
| 156 | KASSERT(ipi_id < IPI_MAXREG); |
| 157 | |
| 158 | /* Release the slot. */ |
| 159 | mutex_enter(&ipi_mngmt_lock); |
| 160 | KASSERT(ipi_intrs[ipi_id].func != NULL); |
| 161 | ipi_intrs[ipi_id].func = NULL; |
| 162 | |
| 163 | /* Ensure that there are no IPIs in flight. */ |
| 164 | kpreempt_disable(); |
| 165 | ipi_broadcast(&ipimsg); |
| 166 | ipi_wait(&ipimsg); |
| 167 | kpreempt_enable(); |
| 168 | mutex_exit(&ipi_mngmt_lock); |
| 169 | } |
| 170 | |
| 171 | /* |
| 172 | * ipi_trigger: asynchronously send an IPI to the specified CPU. |
| 173 | */ |
| 174 | void |
| 175 | ipi_trigger(u_int ipi_id, struct cpu_info *ci) |
| 176 | { |
| 177 | const u_int i = ipi_id >> IPI_BITW_SHIFT; |
| 178 | const uint32_t bitm = 1U << (ipi_id & IPI_BITW_MASK); |
| 179 | |
| 180 | KASSERT(ipi_id < IPI_MAXREG); |
| 181 | KASSERT(kpreempt_disabled()); |
| 182 | KASSERT(curcpu() != ci); |
| 183 | |
| 184 | /* Mark as pending and send an IPI. */ |
| 185 | if (membar_consumer(), (ci->ci_ipipend[i] & bitm) == 0) { |
| 186 | atomic_or_32(&ci->ci_ipipend[i], bitm); |
| 187 | cpu_ipi(ci); |
| 188 | } |
| 189 | } |
| 190 | |
| 191 | /* |
| 192 | * ipi_trigger_multi: same as ipi_trigger() but sends to the multiple |
| 193 | * CPUs given the target CPU set. |
| 194 | */ |
| 195 | void |
| 196 | ipi_trigger_multi(u_int ipi_id, const kcpuset_t *target) |
| 197 | { |
| 198 | const cpuid_t selfid = cpu_index(curcpu()); |
| 199 | CPU_INFO_ITERATOR cii; |
| 200 | struct cpu_info *ci; |
| 201 | |
| 202 | KASSERT(kpreempt_disabled()); |
| 203 | KASSERT(target != NULL); |
| 204 | |
| 205 | for (CPU_INFO_FOREACH(cii, ci)) { |
| 206 | const cpuid_t cpuid = cpu_index(ci); |
| 207 | |
| 208 | if (!kcpuset_isset(target, cpuid) || cpuid == selfid) { |
| 209 | continue; |
| 210 | } |
| 211 | ipi_trigger(ipi_id, ci); |
| 212 | } |
| 213 | if (kcpuset_isset(target, selfid)) { |
| 214 | int s = splhigh(); |
| 215 | ipi_cpu_handler(); |
| 216 | splx(s); |
| 217 | } |
| 218 | } |
| 219 | |
| 220 | /* |
| 221 | * put_msg: insert message into the mailbox. |
| 222 | */ |
| 223 | static inline void |
| 224 | put_msg(ipi_mbox_t *mbox, ipi_msg_t *msg) |
| 225 | { |
| 226 | int count = SPINLOCK_BACKOFF_MIN; |
| 227 | again: |
| 228 | for (u_int i = 0; i < IPI_MSG_MAX; i++) { |
| 229 | if (__predict_true(mbox->msg[i] == NULL) && |
| 230 | atomic_cas_ptr(&mbox->msg[i], NULL, msg) == NULL) { |
| 231 | return; |
| 232 | } |
| 233 | } |
| 234 | |
| 235 | /* All slots are full: we have to spin-wait. */ |
| 236 | ipi_mboxfull_ev.ev_count++; |
| 237 | SPINLOCK_BACKOFF(count); |
| 238 | goto again; |
| 239 | } |
| 240 | |
| 241 | /* |
| 242 | * ipi_cpu_handler: the IPI handler. |
| 243 | */ |
| 244 | void |
| 245 | ipi_cpu_handler(void) |
| 246 | { |
| 247 | struct cpu_info * const ci = curcpu(); |
| 248 | |
| 249 | /* |
| 250 | * Handle asynchronous IPIs: inspect per-CPU bit field, extract |
| 251 | * IPI ID numbers and execute functions in those slots. |
| 252 | */ |
| 253 | for (u_int i = 0; i < IPI_BITWORDS; i++) { |
| 254 | uint32_t pending, bit; |
| 255 | |
| 256 | if (ci->ci_ipipend[i] == 0) { |
| 257 | continue; |
| 258 | } |
| 259 | pending = atomic_swap_32(&ci->ci_ipipend[i], 0); |
| 260 | #ifndef __HAVE_ATOMIC_AS_MEMBAR |
| 261 | membar_producer(); |
| 262 | #endif |
| 263 | while ((bit = ffs(pending)) != 0) { |
| 264 | const u_int ipi_id = (i << IPI_BITW_SHIFT) | --bit; |
| 265 | ipi_intr_t *ipi_hdl = &ipi_intrs[ipi_id]; |
| 266 | |
| 267 | pending &= ~(1U << bit); |
| 268 | KASSERT(ipi_hdl->func != NULL); |
| 269 | ipi_hdl->func(ipi_hdl->arg); |
| 270 | } |
| 271 | } |
| 272 | } |
| 273 | |
| 274 | /* |
| 275 | * ipi_msg_cpu_handler: handle synchronous IPIs - iterate mailbox, |
| 276 | * execute the passed functions and acknowledge the messages. |
| 277 | */ |
| 278 | static void |
| 279 | ipi_msg_cpu_handler(void *arg __unused) |
| 280 | { |
| 281 | const struct cpu_info * const ci = curcpu(); |
| 282 | ipi_mbox_t *mbox = &ipi_mboxes[cpu_index(ci)]; |
| 283 | |
| 284 | for (u_int i = 0; i < IPI_MSG_MAX; i++) { |
| 285 | ipi_msg_t *msg; |
| 286 | |
| 287 | /* Get the message. */ |
| 288 | if ((msg = mbox->msg[i]) == NULL) { |
| 289 | continue; |
| 290 | } |
| 291 | mbox->msg[i] = NULL; |
| 292 | |
| 293 | /* Execute the handler. */ |
| 294 | KASSERT(msg->func); |
| 295 | msg->func(msg->arg); |
| 296 | |
| 297 | /* Ack the request. */ |
| 298 | atomic_dec_uint(&msg->_pending); |
| 299 | } |
| 300 | } |
| 301 | |
| 302 | /* |
| 303 | * ipi_unicast: send an IPI to a single CPU. |
| 304 | * |
| 305 | * => The CPU must be remote; must not be local. |
| 306 | * => The caller must ipi_wait() on the message for completion. |
| 307 | */ |
| 308 | void |
| 309 | ipi_unicast(ipi_msg_t *msg, struct cpu_info *ci) |
| 310 | { |
| 311 | const cpuid_t id = cpu_index(ci); |
| 312 | |
| 313 | KASSERT(msg->func != NULL); |
| 314 | KASSERT(kpreempt_disabled()); |
| 315 | KASSERT(curcpu() != ci); |
| 316 | |
| 317 | msg->_pending = 1; |
| 318 | membar_producer(); |
| 319 | |
| 320 | put_msg(&ipi_mboxes[id], msg); |
| 321 | ipi_trigger(IPI_SYNCH_ID, ci); |
| 322 | } |
| 323 | |
| 324 | /* |
| 325 | * ipi_multicast: send an IPI to each CPU in the specified set. |
| 326 | * |
| 327 | * => The caller must ipi_wait() on the message for completion. |
| 328 | */ |
| 329 | void |
| 330 | ipi_multicast(ipi_msg_t *msg, const kcpuset_t *target) |
| 331 | { |
| 332 | const struct cpu_info * const self = curcpu(); |
| 333 | CPU_INFO_ITERATOR cii; |
| 334 | struct cpu_info *ci; |
| 335 | u_int local; |
| 336 | |
| 337 | KASSERT(msg->func != NULL); |
| 338 | KASSERT(kpreempt_disabled()); |
| 339 | |
| 340 | local = !!kcpuset_isset(target, cpu_index(self)); |
| 341 | msg->_pending = kcpuset_countset(target) - local; |
| 342 | membar_producer(); |
| 343 | |
| 344 | for (CPU_INFO_FOREACH(cii, ci)) { |
| 345 | cpuid_t id; |
| 346 | |
| 347 | if (__predict_false(ci == self)) { |
| 348 | continue; |
| 349 | } |
| 350 | id = cpu_index(ci); |
| 351 | if (!kcpuset_isset(target, id)) { |
| 352 | continue; |
| 353 | } |
| 354 | put_msg(&ipi_mboxes[id], msg); |
| 355 | ipi_trigger(IPI_SYNCH_ID, ci); |
| 356 | } |
| 357 | if (local) { |
| 358 | msg->func(msg->arg); |
| 359 | } |
| 360 | } |
| 361 | |
| 362 | /* |
| 363 | * ipi_broadcast: send an IPI to all CPUs. |
| 364 | * |
| 365 | * => The caller must ipi_wait() on the message for completion. |
| 366 | */ |
| 367 | void |
| 368 | ipi_broadcast(ipi_msg_t *msg) |
| 369 | { |
| 370 | const struct cpu_info * const self = curcpu(); |
| 371 | CPU_INFO_ITERATOR cii; |
| 372 | struct cpu_info *ci; |
| 373 | |
| 374 | KASSERT(msg->func != NULL); |
| 375 | KASSERT(kpreempt_disabled()); |
| 376 | |
| 377 | msg->_pending = ncpu - 1; |
| 378 | membar_producer(); |
| 379 | |
| 380 | /* Broadcast IPIs for remote CPUs. */ |
| 381 | for (CPU_INFO_FOREACH(cii, ci)) { |
| 382 | cpuid_t id; |
| 383 | |
| 384 | if (__predict_false(ci == self)) { |
| 385 | continue; |
| 386 | } |
| 387 | id = cpu_index(ci); |
| 388 | put_msg(&ipi_mboxes[id], msg); |
| 389 | ipi_trigger(IPI_SYNCH_ID, ci); |
| 390 | } |
| 391 | |
| 392 | /* Finally, execute locally. */ |
| 393 | msg->func(msg->arg); |
| 394 | } |
| 395 | |
| 396 | /* |
| 397 | * ipi_wait: spin-wait until the message is processed. |
| 398 | */ |
| 399 | void |
| 400 | ipi_wait(ipi_msg_t *msg) |
| 401 | { |
| 402 | int count = SPINLOCK_BACKOFF_MIN; |
| 403 | |
| 404 | while (msg->_pending) { |
| 405 | KASSERT(msg->_pending < ncpu); |
| 406 | SPINLOCK_BACKOFF(count); |
| 407 | } |
| 408 | } |
| 409 | |