lib/librte_eal/linuxapp/eal/eal_interrupts.c

   1 /*-
   2  *   BSD LICENSE
   3  *
   4  *   Copyright(c) 2010-2012 Intel Corporation. All rights reserved.
   5  *   All rights reserved.
   6  *
   7  *   Redistribution and use in source and binary forms, with or without
   8  *   modification, are permitted provided that the following conditions
   9  *   are met:
  10  *
  11  *     * Redistributions of source code must retain the above copyright
  12  *       notice, this list of conditions and the following disclaimer.
  13  *     * Redistributions in binary form must reproduce the above copyright
  14  *       notice, this list of conditions and the following disclaimer in
  15  *       the documentation and/or other materials provided with the
  16  *       distribution.
  17  *     * Neither the name of Intel Corporation nor the names of its
  18  *       contributors may be used to endorse or promote products derived
  19  *       from this software without specific prior written permission.
  20  *
  21  *   THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  *   "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  *   LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
  24  *   A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
  25  *   OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
  26  *   SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
  27  *   LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
  28  *   DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
  29  *   THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
  30  *   (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
  31  *   OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
  32  *
  33  *  version: DPDK.L.1.2.3-3
  34  */
  35
  36 #include <stdio.h>
  37 #include <stdint.h>
  38 #include <stdlib.h>
  39 #include <pthread.h>
  40 #include <sys/queue.h>
  41 #include <malloc.h>
  42 #include <stdarg.h>
  43 #include <unistd.h>
  44 #include <string.h>
  45 #include <errno.h>
  46 #include <inttypes.h>
  47 #include <sys/epoll.h>
  48 #include <sys/signalfd.h>
  49
  50 #include <rte_common.h>
  51 #include <rte_interrupts.h>
  52 #include <rte_memory.h>
  53 #include <rte_memzone.h>
  54 #include <rte_launch.h>
  55 #include <rte_tailq.h>
  56 #include <rte_eal.h>
  57 #include <rte_per_lcore.h>
  58 #include <rte_lcore.h>
  59 #include <rte_atomic.h>
  60 #include <rte_branch_prediction.h>
  61 #include <rte_ring.h>
  62 #include <rte_debug.h>
  63 #include <rte_log.h>
  64 #include <rte_mempool.h>
  65 #include <rte_pci.h>
  66 #include <rte_malloc.h>
  67 #include <rte_errno.h>
  68 #include <rte_spinlock.h>
  69
  70 #include "eal_private.h"
  71
  72 #define EAL_INTR_EPOLL_WAIT_FOREVER (-1)
  73
  74 /**
  75  * union for pipe fds.
  76  */
  77 union intr_pipefds{
  78         struct {
  79                 int pipefd[2];
  80         };
  81         struct {
  82                 int readfd;
  83                 int writefd;
  84         };
  85 };
  86
  87 /**
  88  * union buffer for reading on different devices
  89  */
  90 union rte_intr_read_buffer {
  91         int uio_intr_count;              /* for uio device */
  92         uint64_t timerfd_num;            /* for timerfd */
  93         char charbuf[16];                /* for others */
  94 };
  95
  96 TAILQ_HEAD(rte_intr_cb_list, rte_intr_callback);
  97 TAILQ_HEAD(rte_intr_source_list, rte_intr_source);
  98
  99 struct rte_intr_callback {
 100         TAILQ_ENTRY(rte_intr_callback) next;
 101         rte_intr_callback_fn cb_fn;  /**< callback address */
 102         void *cb_arg;                /**< parameter for callback */
 103 };
 104
 105 struct rte_intr_source {
 106         TAILQ_ENTRY(rte_intr_source) next;
 107         struct rte_intr_handle intr_handle; /**< interrupt handle */
 108         struct rte_intr_cb_list callbacks;  /**< user callbacks */
 109 };
 110
 111 /* global spinlock for interrupt data operation */
 112 static rte_spinlock_t intr_lock = RTE_SPINLOCK_INITIALIZER;
 113
 114 /* union buffer for pipe read/write */
 115 static union intr_pipefds intr_pipe;
 116
 117 /* interrupt sources list */
 118 static struct rte_intr_source_list intr_sources;
 119
 120 /* interrupt handling thread */
 121 static pthread_t intr_thread;
 122
 123 int
 124 rte_intr_callback_register(struct rte_intr_handle *intr_handle,
 125                         rte_intr_callback_fn cb, void *cb_arg)
 126 {
 127         int ret = -1;
 128         struct rte_intr_source *src;
 129         int wake_thread = 0;
 130
 131         /* first do parameter checking */
 132         if (intr_handle == NULL || intr_handle->fd < 0 || cb == NULL) {
 133                 RTE_LOG(ERR, EAL,
 134                         "Registering with invalid input parameter\n");
 135                 return -EINVAL;
 136         }
 137
 138         /* allocate a new interrupt callback entity */
 139         struct rte_intr_callback *callback =
 140                 rte_zmalloc("interrupt callback list",
 141                                 sizeof(*callback), 0);
 142         if (callback == NULL) {
 143                 RTE_LOG(ERR, EAL, "Can not allocate memory\n");
 144                 return -ENOMEM;
 145         }
 146         callback->cb_fn = cb;
 147         callback->cb_arg = cb_arg;
 148
 149         rte_spinlock_lock(&intr_lock);
 150
 151         /* check if there is at least one callback registered for the fd */
 152         TAILQ_FOREACH(src, &intr_sources, next)
 153         if (src->intr_handle.fd == intr_handle->fd) {
 154                 if (src->callbacks.tqh_first == NULL)
 155                         /* we had no interrupts for this */
 156                         wake_thread = 1;
 157
 158                 TAILQ_INSERT_TAIL(&(src->callbacks), callback, next);
 159                 break;
 160         }
 161
 162         /* No callback registered for this fd */
 163         if (src == NULL){
 164                 /* no existing callbacks for this - add new source */
 165                 src = rte_zmalloc("interrupt source list", sizeof(*src), 0);
 166                 if (src == NULL){
 167                         RTE_LOG(ERR, EAL, "Can not allocate memory\n");
 168                         ret = -ENOMEM;
 169                         goto error;
 170                 }
 171                 src->intr_handle = *intr_handle;
 172                 TAILQ_INIT(&src->callbacks);
 173
 174                 TAILQ_INSERT_TAIL(&intr_sources, src, next);
 175                 TAILQ_INSERT_TAIL(&(src->callbacks), callback, next);
 176                 wake_thread = 1;
 177         }
 178
 179         rte_spinlock_unlock(&intr_lock);
 180         /**
 181          * check if need to notify the pipe fd waited by epoll_wait to
 182          * rebuild the wait list.
 183          */
 184         if (wake_thread)
 185                 if (write(intr_pipe.writefd, "1", 1) < 0)
 186                         return -EPIPE;
 187
 188         return 0;
 189
 190 error:
 191         rte_spinlock_unlock(&intr_lock);
 192
 193         return ret;
 194 }
 195
 196 int
 197 rte_intr_callback_unregister(struct rte_intr_handle *intr_handle,
 198                         rte_intr_callback_fn cb_fn, void *cb_arg)
 199 {
 200         int ret = -1;
 201         struct rte_intr_source *src;
 202         struct rte_intr_callback *cb;
 203
 204         /* do parameter checking first */
 205         if (intr_handle == NULL || intr_handle->fd < 0) {
 206                 RTE_LOG(ERR, EAL,
 207                 "Unregistering with invalid input parameter\n");
 208                 return -EINVAL;
 209         }
 210
 211         rte_spinlock_lock(&intr_lock);
 212
 213         /* check if the insterrupt source for the fd is existent */
 214         TAILQ_FOREACH(src, &intr_sources, next)
 215                 if (src->intr_handle.fd == intr_handle->fd)
 216                         break;
 217
 218         /* No interrupt source registered for the fd */
 219         if (src == NULL) {
 220                 ret = -ENOENT;
 221                 goto error;
 222         }
 223
 224         ret = 0;
 225         TAILQ_FOREACH(cb, &src->callbacks, next) {
 226                 if (cb->cb_fn != cb_fn)
 227                         continue;
 228                 if (cb_arg == (void *)-1 || cb->cb_arg == cb_arg) {
 229                         TAILQ_REMOVE(&src->callbacks, cb, next);
 230                         rte_free(cb);
 231                         ret ++;
 232                 }
 233
 234                 if (src->callbacks.tqh_first == NULL) {
 235                         TAILQ_REMOVE(&intr_sources, src, next);
 236                         rte_free(src);
 237                 }
 238         }
 239
 240         /* notify the pipe fd waited by epoll_wait to rebuild the wait list */
 241         if (write(intr_pipe.writefd, "1", 1) < 0) {
 242                 ret = -EPIPE;
 243                 goto error;
 244         }
 245
 246         rte_spinlock_unlock(&intr_lock);
 247
 248         return ret;
 249
 250 error:
 251         rte_spinlock_unlock(&intr_lock);
 252
 253         return ret;
 254 }
 255
 256 int
 257 rte_intr_enable(struct rte_intr_handle *intr_handle)
 258 {
 259         const int value = 1;
 260
 261         if (!intr_handle || intr_handle->fd < 0)
 262                 return -1;
 263
 264         switch (intr_handle->type){
 265         /* write to the uio fd to enable the interrupt */
 266         case RTE_INTR_HANDLE_UIO:
 267                 if (write(intr_handle->fd, &value, sizeof(value)) < 0) {
 268                         RTE_LOG(ERR, EAL,
 269                                 "Error enabling interrupts for fd %d\n",
 270                                                         intr_handle->fd);
 271                         return -1;
 272                 }
 273                 break;
 274         /* not used at this moment */
 275         case RTE_INTR_HANDLE_ALARM:
 276                 return -1;
 277         /* unkown handle type */
 278         default:
 279                 RTE_LOG(ERR, EAL,
 280                         "Unknown handle type of fd %d\n",
 281                                         intr_handle->fd);
 282                 return -1;
 283         }
 284
 285         return 0;
 286 }
 287
 288 int
 289 rte_intr_disable(struct rte_intr_handle *intr_handle)
 290 {
 291         const int value = 0;
 292
 293         if (!intr_handle || intr_handle->fd < 0)
 294                 return -1;
 295
 296         switch (intr_handle->type){
 297         /* write to the uio fd to disable the interrupt */
 298         case RTE_INTR_HANDLE_UIO:
 299                 if (write(intr_handle->fd, &value, sizeof(value)) < 0){
 300                         RTE_LOG(ERR, EAL,
 301                                 "Error enabling interrupts for fd %d\n",
 302                                                         intr_handle->fd);
 303                         return -1;
 304                 }
 305                 break;
 306         /* not used at this moment */
 307         case RTE_INTR_HANDLE_ALARM:
 308                 return -1;
 309         /* unkown handle type */
 310         default:
 311                 RTE_LOG(ERR, EAL,
 312                         "Unknown handle type of fd %d\n",
 313                                         intr_handle->fd);
 314                 return -1;
 315         }
 316
 317         return 0;
 318 }
 319
 320 static int
 321 eal_intr_process_interrupts(struct epoll_event *events, int nfds)
 322 {
 323         int n, i, active_cb, bytes_read;
 324         struct rte_intr_source *src;
 325         struct rte_intr_callback *cb;
 326         union rte_intr_read_buffer buf;
 327         struct rte_intr_callback active_cbs[32];
 328
 329         for (n = 0; n < nfds; n++) {
 330                 /**
 331                  * if the pipe fd is ready to read, return out to
 332                  * rebuild the wait list.
 333                  */
 334                 if (events[n].data.fd == intr_pipe.readfd){
 335                         int r = read(intr_pipe.readfd, buf.charbuf,
 336                                         sizeof(buf.charbuf));
 337                         RTE_SET_USED(r);
 338                         return -1;
 339                 }
 340                 rte_spinlock_lock(&intr_lock);
 341                 TAILQ_FOREACH(src, &intr_sources, next)
 342                         if (src->intr_handle.fd ==
 343                                         events[n].data.fd)
 344                                 break;
 345                 if (src == NULL){
 346                         rte_spinlock_unlock(&intr_lock);
 347                         continue;
 348                 }
 349
 350                 /* for this source, make a copy of all the callbacks,
 351                  * then unlock the lock, so the callbacks can
 352                  * themselves manipulate the list for future
 353                  * instances.
 354                  */
 355                 active_cb = 0;
 356                 memset(active_cbs, 0, sizeof(active_cbs));
 357                 TAILQ_FOREACH(cb, &src->callbacks, next)
 358                         active_cbs[active_cb++] = *cb;
 359                 rte_spinlock_unlock(&intr_lock);
 360
 361                 /* set the length to be read dor different handle type */
 362                 switch (src->intr_handle.type) {
 363                 case RTE_INTR_HANDLE_UIO:
 364                         bytes_read = 4;
 365                         break;
 366                 case RTE_INTR_HANDLE_ALARM:
 367                         bytes_read = sizeof(uint64_t);
 368                         break;
 369                 default:
 370                         bytes_read = 1;
 371                         break;
 372                 }
 373                 /**
 374                  * read out to clear the ready-to-be-read flag
 375                  * for epoll_wait.
 376                  */
 377                 bytes_read = read(events[n].data.fd, &buf, bytes_read);
 378                 if (bytes_read < 0) {
 379                         RTE_LOG(ERR, EAL, "Error reading from file descriptor"
 380                                 " %d, error: %d\n", events[n].data.fd, errno);
 381                         continue;
 382                 }
 383                 else if (bytes_read == 0) {
 384                         RTE_LOG(ERR, EAL,
 385                                 "Read nothing from file descriptor %d.\n",
 386                                                         events[n].data.fd);
 387                         continue;
 388                 }
 389                 /**
 390                  * Finally, call all callbacks from the copy
 391                  * we made earlier.
 392                  */
 393                 for (i = 0; i < active_cb; i++) {
 394                         if (active_cbs[i].cb_fn == NULL)
 395                                 continue;
 396                         active_cbs[i].cb_fn(&src->intr_handle,
 397                                         active_cbs[i].cb_arg);
 398                 }
 399         }
 400
 401         return 0;
 402 }
 403
 404 /**
 405  * It handles all the interrupts.
 406  *
 407  * @param pfd
 408  *  epoll file descriptor.
 409  * @param totalfds
 410  *  The number of file descriptors added in epoll.
 411  *
 412  * @return
 413  *  void
 414  */
 415 static void
 416 eal_intr_handle_interrupts(int pfd, unsigned totalfds)
 417 {
 418         struct epoll_event events[totalfds];
 419         int nfds = 0;
 420
 421         for(;;) {
 422                 nfds = epoll_wait(pfd, events, totalfds,
 423                         EAL_INTR_EPOLL_WAIT_FOREVER);
 424                 /* epoll_wait fail */
 425                 if (nfds < 0) {
 426                         if (errno == EINTR)
 427                                 continue;
 428                         RTE_LOG(ERR, EAL,
 429                                 "epoll_wait returns with fail\n");
 430                         return;
 431                 }
 432                 /* epoll_wait timeout, will never happens here */
 433                 else if (nfds == 0)
 434                         continue;
 435                 /* epoll_wait has at least one fd ready to read */
 436                 if (eal_intr_process_interrupts(events, nfds) < 0)
 437                         return;
 438         }
 439 }
 440
 441 /**
 442  * It builds/rebuilds up the epoll file descriptor with all the
 443  * file descriptors being waited on. Then handles the interrupts.
 444  *
 445  * @param arg
 446  *  pointer. (unused)
 447  *
 448  * @return
 449  *  never return;
 450  */
 451 static __attribute__((noreturn)) void *
 452 eal_intr_thread_main(__rte_unused void *arg)
 453 {
 454         struct epoll_event ev;
 455
 456         /* host thread, never break out */
 457         for (;;) {
 458                 /* build up the epoll fd with all descriptors we are to
 459                  * wait on then pass it to the handle_interrupts function
 460                  */
 461                 static struct epoll_event pipe_event = {
 462                         .events = EPOLLIN | EPOLLPRI,
 463                 };
 464                 struct rte_intr_source *src;
 465                 unsigned numfds = 0;
 466
 467                 /* create epoll fd */
 468                 int pfd = epoll_create(1);
 469                 if (pfd < 0)
 470                         rte_panic("Cannot create epoll instance\n");
 471
 472                 pipe_event.data.fd = intr_pipe.readfd;
 473                 /**
 474                  * add pipe fd into wait list, this pipe is used to
 475                  * rebuild the wait list.
 476                  */
 477                 if (epoll_ctl(pfd, EPOLL_CTL_ADD, intr_pipe.readfd,
 478                                                 &pipe_event) < 0) {
 479                         rte_panic("Error adding fd to %d epoll_ctl, %s\n",
 480                                         intr_pipe.readfd, strerror(errno));
 481                 }
 482                 numfds++;
 483
 484                 rte_spinlock_lock(&intr_lock);
 485
 486                 TAILQ_FOREACH(src, &intr_sources, next) {
 487                         if (src->callbacks.tqh_first == NULL)
 488                                 continue; /* skip those with no callbacks */
 489                         ev.events = EPOLLIN | EPOLLPRI;
 490                         ev.data.fd = src->intr_handle.fd;
 491
 492                         /**
 493                          * add all the uio device file descriptor
 494                          * into wait list.
 495                          */
 496                         if (epoll_ctl(pfd, EPOLL_CTL_ADD,
 497                                         src->intr_handle.fd, &ev) < 0){
 498                                 rte_panic("Error adding fd %d epoll_ctl, %s\n",
 499                                         src->intr_handle.fd, strerror(errno));
 500                         }
 501                         else
 502                                 numfds++;
 503                 }
 504                 rte_spinlock_unlock(&intr_lock);
 505                 /* serve the interrupt */
 506                 eal_intr_handle_interrupts(pfd, numfds);
 507
 508                 /**
 509                  * when we return, we need to rebuild the
 510                  * list of fds to monitor.
 511                  */
 512                 close(pfd);
 513         }
 514 }
 515
 516 int
 517 rte_eal_intr_init(void)
 518 {
 519         int ret = 0;
 520
 521         /* init the global interrupt source head */
 522         TAILQ_INIT(&intr_sources);
 523
 524         /**
 525          * create a pipe which will be waited by epoll and notified to
 526          * rebuild the wait list of epoll.
 527          */
 528         if (pipe(intr_pipe.pipefd) < 0)
 529                 return -1;
 530
 531         /* create the host thread to wait/handle the interrupt */
 532         ret = pthread_create(&intr_thread, NULL,
 533                         eal_intr_thread_main, NULL);
 534         if (ret != 0)
 535                 RTE_LOG(ERR, EAL,
 536                         "Failed to create thread for interrupt handling\n");
 537
 538         return -ret;
 539 }
 540