lib/librte_eal/linuxapp/eal/eal_interrupts.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2010-2014 Intel Corporation
   3  */
   4
   5 #include <stdio.h>
   6 #include <stdint.h>
   7 #include <stdlib.h>
   8 #include <pthread.h>
   9 #include <sys/queue.h>
  10 #include <stdarg.h>
  11 #include <unistd.h>
  12 #include <string.h>
  13 #include <errno.h>
  14 #include <inttypes.h>
  15 #include <sys/epoll.h>
  16 #include <sys/signalfd.h>
  17 #include <sys/ioctl.h>
  18 #include <sys/eventfd.h>
  19 #include <assert.h>
  20 #include <stdbool.h>
  21
  22 #include <rte_common.h>
  23 #include <rte_interrupts.h>
  24 #include <rte_memory.h>
  25 #include <rte_launch.h>
  26 #include <rte_eal.h>
  27 #include <rte_per_lcore.h>
  28 #include <rte_lcore.h>
  29 #include <rte_atomic.h>
  30 #include <rte_branch_prediction.h>
  31 #include <rte_debug.h>
  32 #include <rte_log.h>
  33 #include <rte_malloc.h>
  34 #include <rte_errno.h>
  35 #include <rte_spinlock.h>
  36 #include <rte_pause.h>
  37
  38 #include "eal_private.h"
  39 #include "eal_vfio.h"
  40 #include "eal_thread.h"
  41
  42 #define EAL_INTR_EPOLL_WAIT_FOREVER (-1)
  43 #define NB_OTHER_INTR               1
  44
  45 static RTE_DEFINE_PER_LCORE(int, _epfd) = -1; /**< epoll fd per thread */
  46
  47 /**
  48  * union for pipe fds.
  49  */
  50 union intr_pipefds{
  51         struct {
  52                 int pipefd[2];
  53         };
  54         struct {
  55                 int readfd;
  56                 int writefd;
  57         };
  58 };
  59
  60 /**
  61  * union buffer for reading on different devices
  62  */
  63 union rte_intr_read_buffer {
  64         int uio_intr_count;              /* for uio device */
  65 #ifdef VFIO_PRESENT
  66         uint64_t vfio_intr_count;        /* for vfio device */
  67 #endif
  68         uint64_t timerfd_num;            /* for timerfd */
  69         char charbuf[16];                /* for others */
  70 };
  71
  72 TAILQ_HEAD(rte_intr_cb_list, rte_intr_callback);
  73 TAILQ_HEAD(rte_intr_source_list, rte_intr_source);
  74
  75 struct rte_intr_callback {
  76         TAILQ_ENTRY(rte_intr_callback) next;
  77         rte_intr_callback_fn cb_fn;  /**< callback address */
  78         void *cb_arg;                /**< parameter for callback */
  79 };
  80
  81 struct rte_intr_source {
  82         TAILQ_ENTRY(rte_intr_source) next;
  83         struct rte_intr_handle intr_handle; /**< interrupt handle */
  84         struct rte_intr_cb_list callbacks;  /**< user callbacks */
  85         uint32_t active;
  86 };
  87
  88 /* global spinlock for interrupt data operation */
  89 static rte_spinlock_t intr_lock = RTE_SPINLOCK_INITIALIZER;
  90
  91 /* union buffer for pipe read/write */
  92 static union intr_pipefds intr_pipe;
  93
  94 /* interrupt sources list */
  95 static struct rte_intr_source_list intr_sources;
  96
  97 /* interrupt handling thread */
  98 static pthread_t intr_thread;
  99
 100 /* VFIO interrupts */
 101 #ifdef VFIO_PRESENT
 102
 103 #define IRQ_SET_BUF_LEN  (sizeof(struct vfio_irq_set) + sizeof(int))
 104 /* irq set buffer length for queue interrupts and LSC interrupt */
 105 #define MSIX_IRQ_SET_BUF_LEN (sizeof(struct vfio_irq_set) + \
 106                               sizeof(int) * (RTE_MAX_RXTX_INTR_VEC_ID + 1))
 107
 108 /* enable legacy (INTx) interrupts */
 109 static int
 110 vfio_enable_intx(const struct rte_intr_handle *intr_handle) {
 111         struct vfio_irq_set *irq_set;
 112         char irq_set_buf[IRQ_SET_BUF_LEN];
 113         int len, ret;
 114         int *fd_ptr;
 115
 116         len = sizeof(irq_set_buf);
 117
 118         /* enable INTx */
 119         irq_set = (struct vfio_irq_set *) irq_set_buf;
 120         irq_set->argsz = len;
 121         irq_set->count = 1;
 122         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
 123         irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
 124         irq_set->start = 0;
 125         fd_ptr = (int *) &irq_set->data;
 126         *fd_ptr = intr_handle->fd;
 127
 128         ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 129
 130         if (ret) {
 131                 RTE_LOG(ERR, EAL, "Error enabling INTx interrupts for fd %d\n",
 132                                                 intr_handle->fd);
 133                 return -1;
 134         }
 135
 136         /* unmask INTx after enabling */
 137         memset(irq_set, 0, len);
 138         len = sizeof(struct vfio_irq_set);
 139         irq_set->argsz = len;
 140         irq_set->count = 1;
 141         irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK;
 142         irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
 143         irq_set->start = 0;
 144
 145         ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 146
 147         if (ret) {
 148                 RTE_LOG(ERR, EAL, "Error unmasking INTx interrupts for fd %d\n",
 149                                                 intr_handle->fd);
 150                 return -1;
 151         }
 152         return 0;
 153 }
 154
 155 /* disable legacy (INTx) interrupts */
 156 static int
 157 vfio_disable_intx(const struct rte_intr_handle *intr_handle) {
 158         struct vfio_irq_set *irq_set;
 159         char irq_set_buf[IRQ_SET_BUF_LEN];
 160         int len, ret;
 161
 162         len = sizeof(struct vfio_irq_set);
 163
 164         /* mask interrupts before disabling */
 165         irq_set = (struct vfio_irq_set *) irq_set_buf;
 166         irq_set->argsz = len;
 167         irq_set->count = 1;
 168         irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK;
 169         irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
 170         irq_set->start = 0;
 171
 172         ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 173
 174         if (ret) {
 175                 RTE_LOG(ERR, EAL, "Error masking INTx interrupts for fd %d\n",
 176                                                 intr_handle->fd);
 177                 return -1;
 178         }
 179
 180         /* disable INTx*/
 181         memset(irq_set, 0, len);
 182         irq_set->argsz = len;
 183         irq_set->count = 0;
 184         irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER;
 185         irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
 186         irq_set->start = 0;
 187
 188         ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 189
 190         if (ret) {
 191                 RTE_LOG(ERR, EAL,
 192                         "Error disabling INTx interrupts for fd %d\n", intr_handle->fd);
 193                 return -1;
 194         }
 195         return 0;
 196 }
 197
 198 /* enable MSI interrupts */
 199 static int
 200 vfio_enable_msi(const struct rte_intr_handle *intr_handle) {
 201         int len, ret;
 202         char irq_set_buf[IRQ_SET_BUF_LEN];
 203         struct vfio_irq_set *irq_set;
 204         int *fd_ptr;
 205
 206         len = sizeof(irq_set_buf);
 207
 208         irq_set = (struct vfio_irq_set *) irq_set_buf;
 209         irq_set->argsz = len;
 210         irq_set->count = 1;
 211         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
 212         irq_set->index = VFIO_PCI_MSI_IRQ_INDEX;
 213         irq_set->start = 0;
 214         fd_ptr = (int *) &irq_set->data;
 215         *fd_ptr = intr_handle->fd;
 216
 217         ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 218
 219         if (ret) {
 220                 RTE_LOG(ERR, EAL, "Error enabling MSI interrupts for fd %d\n",
 221                                                 intr_handle->fd);
 222                 return -1;
 223         }
 224         return 0;
 225 }
 226
 227 /* disable MSI interrupts */
 228 static int
 229 vfio_disable_msi(const struct rte_intr_handle *intr_handle) {
 230         struct vfio_irq_set *irq_set;
 231         char irq_set_buf[IRQ_SET_BUF_LEN];
 232         int len, ret;
 233
 234         len = sizeof(struct vfio_irq_set);
 235
 236         irq_set = (struct vfio_irq_set *) irq_set_buf;
 237         irq_set->argsz = len;
 238         irq_set->count = 0;
 239         irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER;
 240         irq_set->index = VFIO_PCI_MSI_IRQ_INDEX;
 241         irq_set->start = 0;
 242
 243         ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 244
 245         if (ret)
 246                 RTE_LOG(ERR, EAL,
 247                         "Error disabling MSI interrupts for fd %d\n", intr_handle->fd);
 248
 249         return ret;
 250 }
 251
 252 /* enable MSI-X interrupts */
 253 static int
 254 vfio_enable_msix(const struct rte_intr_handle *intr_handle) {
 255         int len, ret;
 256         char irq_set_buf[MSIX_IRQ_SET_BUF_LEN];
 257         struct vfio_irq_set *irq_set;
 258         int *fd_ptr;
 259
 260         len = sizeof(irq_set_buf);
 261
 262         irq_set = (struct vfio_irq_set *) irq_set_buf;
 263         irq_set->argsz = len;
 264         /* 0 < irq_set->count < RTE_MAX_RXTX_INTR_VEC_ID + 1 */
 265         irq_set->count = intr_handle->max_intr ?
 266                 (intr_handle->max_intr > RTE_MAX_RXTX_INTR_VEC_ID + 1 ?
 267                 RTE_MAX_RXTX_INTR_VEC_ID + 1 : intr_handle->max_intr) : 1;
 268         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
 269         irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
 270         irq_set->start = 0;
 271         fd_ptr = (int *) &irq_set->data;
 272         /* INTR vector offset 0 reserve for non-efds mapping */
 273         fd_ptr[RTE_INTR_VEC_ZERO_OFFSET] = intr_handle->fd;
 274         memcpy(&fd_ptr[RTE_INTR_VEC_RXTX_OFFSET], intr_handle->efds,
 275                 sizeof(*intr_handle->efds) * intr_handle->nb_efd);
 276
 277         ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 278
 279         if (ret) {
 280                 RTE_LOG(ERR, EAL, "Error enabling MSI-X interrupts for fd %d\n",
 281                                                 intr_handle->fd);
 282                 return -1;
 283         }
 284
 285         return 0;
 286 }
 287
 288 /* disable MSI-X interrupts */
 289 static int
 290 vfio_disable_msix(const struct rte_intr_handle *intr_handle) {
 291         struct vfio_irq_set *irq_set;
 292         char irq_set_buf[MSIX_IRQ_SET_BUF_LEN];
 293         int len, ret;
 294
 295         len = sizeof(struct vfio_irq_set);
 296
 297         irq_set = (struct vfio_irq_set *) irq_set_buf;
 298         irq_set->argsz = len;
 299         irq_set->count = 0;
 300         irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER;
 301         irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
 302         irq_set->start = 0;
 303
 304         ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 305
 306         if (ret)
 307                 RTE_LOG(ERR, EAL,
 308                         "Error disabling MSI-X interrupts for fd %d\n", intr_handle->fd);
 309
 310         return ret;
 311 }
 312 #endif
 313
 314 static int
 315 uio_intx_intr_disable(const struct rte_intr_handle *intr_handle)
 316 {
 317         unsigned char command_high;
 318
 319         /* use UIO config file descriptor for uio_pci_generic */
 320         if (pread(intr_handle->uio_cfg_fd, &command_high, 1, 5) != 1) {
 321                 RTE_LOG(ERR, EAL,
 322                         "Error reading interrupts status for fd %d\n",
 323                         intr_handle->uio_cfg_fd);
 324                 return -1;
 325         }
 326         /* disable interrupts */
 327         command_high |= 0x4;
 328         if (pwrite(intr_handle->uio_cfg_fd, &command_high, 1, 5) != 1) {
 329                 RTE_LOG(ERR, EAL,
 330                         "Error disabling interrupts for fd %d\n",
 331                         intr_handle->uio_cfg_fd);
 332                 return -1;
 333         }
 334
 335         return 0;
 336 }
 337
 338 static int
 339 uio_intx_intr_enable(const struct rte_intr_handle *intr_handle)
 340 {
 341         unsigned char command_high;
 342
 343         /* use UIO config file descriptor for uio_pci_generic */
 344         if (pread(intr_handle->uio_cfg_fd, &command_high, 1, 5) != 1) {
 345                 RTE_LOG(ERR, EAL,
 346                         "Error reading interrupts status for fd %d\n",
 347                         intr_handle->uio_cfg_fd);
 348                 return -1;
 349         }
 350         /* enable interrupts */
 351         command_high &= ~0x4;
 352         if (pwrite(intr_handle->uio_cfg_fd, &command_high, 1, 5) != 1) {
 353                 RTE_LOG(ERR, EAL,
 354                         "Error enabling interrupts for fd %d\n",
 355                         intr_handle->uio_cfg_fd);
 356                 return -1;
 357         }
 358
 359         return 0;
 360 }
 361
 362 static int
 363 uio_intr_disable(const struct rte_intr_handle *intr_handle)
 364 {
 365         const int value = 0;
 366
 367         if (write(intr_handle->fd, &value, sizeof(value)) < 0) {
 368                 RTE_LOG(ERR, EAL,
 369                         "Error disabling interrupts for fd %d (%s)\n",
 370                         intr_handle->fd, strerror(errno));
 371                 return -1;
 372         }
 373         return 0;
 374 }
 375
 376 static int
 377 uio_intr_enable(const struct rte_intr_handle *intr_handle)
 378 {
 379         const int value = 1;
 380
 381         if (write(intr_handle->fd, &value, sizeof(value)) < 0) {
 382                 RTE_LOG(ERR, EAL,
 383                         "Error enabling interrupts for fd %d (%s)\n",
 384                         intr_handle->fd, strerror(errno));
 385                 return -1;
 386         }
 387         return 0;
 388 }
 389
 390 int
 391 rte_intr_callback_register(const struct rte_intr_handle *intr_handle,
 392                         rte_intr_callback_fn cb, void *cb_arg)
 393 {
 394         int ret, wake_thread;
 395         struct rte_intr_source *src;
 396         struct rte_intr_callback *callback;
 397
 398         wake_thread = 0;
 399
 400         /* first do parameter checking */
 401         if (intr_handle == NULL || intr_handle->fd < 0 || cb == NULL) {
 402                 RTE_LOG(ERR, EAL,
 403                         "Registering with invalid input parameter\n");
 404                 return -EINVAL;
 405         }
 406
 407         /* allocate a new interrupt callback entity */
 408         callback = rte_zmalloc("interrupt callback list",
 409                                 sizeof(*callback), 0);
 410         if (callback == NULL) {
 411                 RTE_LOG(ERR, EAL, "Can not allocate memory\n");
 412                 return -ENOMEM;
 413         }
 414         callback->cb_fn = cb;
 415         callback->cb_arg = cb_arg;
 416
 417         rte_spinlock_lock(&intr_lock);
 418
 419         /* check if there is at least one callback registered for the fd */
 420         TAILQ_FOREACH(src, &intr_sources, next) {
 421                 if (src->intr_handle.fd == intr_handle->fd) {
 422                         /* we had no interrupts for this */
 423                         if TAILQ_EMPTY(&src->callbacks)
 424                                 wake_thread = 1;
 425
 426                         TAILQ_INSERT_TAIL(&(src->callbacks), callback, next);
 427                         ret = 0;
 428                         break;
 429                 }
 430         }
 431
 432         /* no existing callbacks for this - add new source */
 433         if (src == NULL) {
 434                 if ((src = rte_zmalloc("interrupt source list",
 435                                 sizeof(*src), 0)) == NULL) {
 436                         RTE_LOG(ERR, EAL, "Can not allocate memory\n");
 437                         rte_free(callback);
 438                         ret = -ENOMEM;
 439                 } else {
 440                         src->intr_handle = *intr_handle;
 441                         TAILQ_INIT(&src->callbacks);
 442                         TAILQ_INSERT_TAIL(&(src->callbacks), callback, next);
 443                         TAILQ_INSERT_TAIL(&intr_sources, src, next);
 444                         wake_thread = 1;
 445                         ret = 0;
 446                 }
 447         }
 448
 449         rte_spinlock_unlock(&intr_lock);
 450
 451         /**
 452          * check if need to notify the pipe fd waited by epoll_wait to
 453          * rebuild the wait list.
 454          */
 455         if (wake_thread)
 456                 if (write(intr_pipe.writefd, "1", 1) < 0)
 457                         return -EPIPE;
 458
 459         return ret;
 460 }
 461
 462 int
 463 rte_intr_callback_unregister(const struct rte_intr_handle *intr_handle,
 464                         rte_intr_callback_fn cb_fn, void *cb_arg)
 465 {
 466         int ret;
 467         struct rte_intr_source *src;
 468         struct rte_intr_callback *cb, *next;
 469
 470         /* do parameter checking first */
 471         if (intr_handle == NULL || intr_handle->fd < 0) {
 472                 RTE_LOG(ERR, EAL,
 473                 "Unregistering with invalid input parameter\n");
 474                 return -EINVAL;
 475         }
 476
 477         rte_spinlock_lock(&intr_lock);
 478
 479         /* check if the insterrupt source for the fd is existent */
 480         TAILQ_FOREACH(src, &intr_sources, next)
 481                 if (src->intr_handle.fd == intr_handle->fd)
 482                         break;
 483
 484         /* No interrupt source registered for the fd */
 485         if (src == NULL) {
 486                 ret = -ENOENT;
 487
 488         /* interrupt source has some active callbacks right now. */
 489         } else if (src->active != 0) {
 490                 ret = -EAGAIN;
 491
 492         /* ok to remove. */
 493         } else {
 494                 ret = 0;
 495
 496                 /*walk through the callbacks and remove all that match. */
 497                 for (cb = TAILQ_FIRST(&src->callbacks); cb != NULL; cb = next) {
 498
 499                         next = TAILQ_NEXT(cb, next);
 500
 501                         if (cb->cb_fn == cb_fn && (cb_arg == (void *)-1 ||
 502                                         cb->cb_arg == cb_arg)) {
 503                                 TAILQ_REMOVE(&src->callbacks, cb, next);
 504                                 rte_free(cb);
 505                                 ret++;
 506                         }
 507                 }
 508
 509                 /* all callbacks for that source are removed. */
 510                 if (TAILQ_EMPTY(&src->callbacks)) {
 511                         TAILQ_REMOVE(&intr_sources, src, next);
 512                         rte_free(src);
 513                 }
 514         }
 515
 516         rte_spinlock_unlock(&intr_lock);
 517
 518         /* notify the pipe fd waited by epoll_wait to rebuild the wait list */
 519         if (ret >= 0 && write(intr_pipe.writefd, "1", 1) < 0) {
 520                 ret = -EPIPE;
 521         }
 522
 523         return ret;
 524 }
 525
 526 int
 527 rte_intr_enable(const struct rte_intr_handle *intr_handle)
 528 {
 529         if (intr_handle && intr_handle->type == RTE_INTR_HANDLE_VDEV)
 530                 return 0;
 531
 532         if (!intr_handle || intr_handle->fd < 0 || intr_handle->uio_cfg_fd < 0)
 533                 return -1;
 534
 535         switch (intr_handle->type){
 536         /* write to the uio fd to enable the interrupt */
 537         case RTE_INTR_HANDLE_UIO:
 538                 if (uio_intr_enable(intr_handle))
 539                         return -1;
 540                 break;
 541         case RTE_INTR_HANDLE_UIO_INTX:
 542                 if (uio_intx_intr_enable(intr_handle))
 543                         return -1;
 544                 break;
 545         /* not used at this moment */
 546         case RTE_INTR_HANDLE_ALARM:
 547                 return -1;
 548 #ifdef VFIO_PRESENT
 549         case RTE_INTR_HANDLE_VFIO_MSIX:
 550                 if (vfio_enable_msix(intr_handle))
 551                         return -1;
 552                 break;
 553         case RTE_INTR_HANDLE_VFIO_MSI:
 554                 if (vfio_enable_msi(intr_handle))
 555                         return -1;
 556                 break;
 557         case RTE_INTR_HANDLE_VFIO_LEGACY:
 558                 if (vfio_enable_intx(intr_handle))
 559                         return -1;
 560                 break;
 561 #endif
 562         /* unknown handle type */
 563         default:
 564                 RTE_LOG(ERR, EAL,
 565                         "Unknown handle type of fd %d\n",
 566                                         intr_handle->fd);
 567                 return -1;
 568         }
 569
 570         return 0;
 571 }
 572
 573 int
 574 rte_intr_disable(const struct rte_intr_handle *intr_handle)
 575 {
 576         if (intr_handle && intr_handle->type == RTE_INTR_HANDLE_VDEV)
 577                 return 0;
 578
 579         if (!intr_handle || intr_handle->fd < 0 || intr_handle->uio_cfg_fd < 0)
 580                 return -1;
 581
 582         switch (intr_handle->type){
 583         /* write to the uio fd to disable the interrupt */
 584         case RTE_INTR_HANDLE_UIO:
 585                 if (uio_intr_disable(intr_handle))
 586                         return -1;
 587                 break;
 588         case RTE_INTR_HANDLE_UIO_INTX:
 589                 if (uio_intx_intr_disable(intr_handle))
 590                         return -1;
 591                 break;
 592         /* not used at this moment */
 593         case RTE_INTR_HANDLE_ALARM:
 594                 return -1;
 595 #ifdef VFIO_PRESENT
 596         case RTE_INTR_HANDLE_VFIO_MSIX:
 597                 if (vfio_disable_msix(intr_handle))
 598                         return -1;
 599                 break;
 600         case RTE_INTR_HANDLE_VFIO_MSI:
 601                 if (vfio_disable_msi(intr_handle))
 602                         return -1;
 603                 break;
 604         case RTE_INTR_HANDLE_VFIO_LEGACY:
 605                 if (vfio_disable_intx(intr_handle))
 606                         return -1;
 607                 break;
 608 #endif
 609         /* unknown handle type */
 610         default:
 611                 RTE_LOG(ERR, EAL,
 612                         "Unknown handle type of fd %d\n",
 613                                         intr_handle->fd);
 614                 return -1;
 615         }
 616
 617         return 0;
 618 }
 619
 620 static int
 621 eal_intr_process_interrupts(struct epoll_event *events, int nfds)
 622 {
 623         bool call = false;
 624         int n, bytes_read;
 625         struct rte_intr_source *src;
 626         struct rte_intr_callback *cb;
 627         union rte_intr_read_buffer buf;
 628         struct rte_intr_callback active_cb;
 629
 630         for (n = 0; n < nfds; n++) {
 631
 632                 /**
 633                  * if the pipe fd is ready to read, return out to
 634                  * rebuild the wait list.
 635                  */
 636                 if (events[n].data.fd == intr_pipe.readfd){
 637                         int r = read(intr_pipe.readfd, buf.charbuf,
 638                                         sizeof(buf.charbuf));
 639                         RTE_SET_USED(r);
 640                         return -1;
 641                 }
 642                 rte_spinlock_lock(&intr_lock);
 643                 TAILQ_FOREACH(src, &intr_sources, next)
 644                         if (src->intr_handle.fd ==
 645                                         events[n].data.fd)
 646                                 break;
 647                 if (src == NULL){
 648                         rte_spinlock_unlock(&intr_lock);
 649                         continue;
 650                 }
 651
 652                 /* mark this interrupt source as active and release the lock. */
 653                 src->active = 1;
 654                 rte_spinlock_unlock(&intr_lock);
 655
 656                 /* set the length to be read dor different handle type */
 657                 switch (src->intr_handle.type) {
 658                 case RTE_INTR_HANDLE_UIO:
 659                 case RTE_INTR_HANDLE_UIO_INTX:
 660                         bytes_read = sizeof(buf.uio_intr_count);
 661                         break;
 662                 case RTE_INTR_HANDLE_ALARM:
 663                         bytes_read = sizeof(buf.timerfd_num);
 664                         break;
 665 #ifdef VFIO_PRESENT
 666                 case RTE_INTR_HANDLE_VFIO_MSIX:
 667                 case RTE_INTR_HANDLE_VFIO_MSI:
 668                 case RTE_INTR_HANDLE_VFIO_LEGACY:
 669                         bytes_read = sizeof(buf.vfio_intr_count);
 670                         break;
 671 #endif
 672                 case RTE_INTR_HANDLE_VDEV:
 673                 case RTE_INTR_HANDLE_EXT:
 674                         bytes_read = 0;
 675                         call = true;
 676                         break;
 677
 678                 default:
 679                         bytes_read = 1;
 680                         break;
 681                 }
 682
 683                 if (bytes_read > 0) {
 684                         /**
 685                          * read out to clear the ready-to-be-read flag
 686                          * for epoll_wait.
 687                          */
 688                         bytes_read = read(events[n].data.fd, &buf, bytes_read);
 689                         if (bytes_read < 0) {
 690                                 if (errno == EINTR || errno == EWOULDBLOCK)
 691                                         continue;
 692
 693                                 RTE_LOG(ERR, EAL, "Error reading from file "
 694                                         "descriptor %d: %s\n",
 695                                         events[n].data.fd,
 696                                         strerror(errno));
 697                         } else if (bytes_read == 0)
 698                                 RTE_LOG(ERR, EAL, "Read nothing from file "
 699                                         "descriptor %d\n", events[n].data.fd);
 700                         else
 701                                 call = true;
 702                 }
 703
 704                 /* grab a lock, again to call callbacks and update status. */
 705                 rte_spinlock_lock(&intr_lock);
 706
 707                 if (call) {
 708
 709                         /* Finally, call all callbacks. */
 710                         TAILQ_FOREACH(cb, &src->callbacks, next) {
 711
 712                                 /* make a copy and unlock. */
 713                                 active_cb = *cb;
 714                                 rte_spinlock_unlock(&intr_lock);
 715
 716                                 /* call the actual callback */
 717                                 active_cb.cb_fn(active_cb.cb_arg);
 718
 719                                 /*get the lock back. */
 720                                 rte_spinlock_lock(&intr_lock);
 721                         }
 722                 }
 723
 724                 /* we done with that interrupt source, release it. */
 725                 src->active = 0;
 726                 rte_spinlock_unlock(&intr_lock);
 727         }
 728
 729         return 0;
 730 }
 731
 732 /**
 733  * It handles all the interrupts.
 734  *
 735  * @param pfd
 736  *  epoll file descriptor.
 737  * @param totalfds
 738  *  The number of file descriptors added in epoll.
 739  *
 740  * @return
 741  *  void
 742  */
 743 static void
 744 eal_intr_handle_interrupts(int pfd, unsigned totalfds)
 745 {
 746         struct epoll_event events[totalfds];
 747         int nfds = 0;
 748
 749         for(;;) {
 750                 nfds = epoll_wait(pfd, events, totalfds,
 751                         EAL_INTR_EPOLL_WAIT_FOREVER);
 752                 /* epoll_wait fail */
 753                 if (nfds < 0) {
 754                         if (errno == EINTR)
 755                                 continue;
 756                         RTE_LOG(ERR, EAL,
 757                                 "epoll_wait returns with fail\n");
 758                         return;
 759                 }
 760                 /* epoll_wait timeout, will never happens here */
 761                 else if (nfds == 0)
 762                         continue;
 763                 /* epoll_wait has at least one fd ready to read */
 764                 if (eal_intr_process_interrupts(events, nfds) < 0)
 765                         return;
 766         }
 767 }
 768
 769 /**
 770  * It builds/rebuilds up the epoll file descriptor with all the
 771  * file descriptors being waited on. Then handles the interrupts.
 772  *
 773  * @param arg
 774  *  pointer. (unused)
 775  *
 776  * @return
 777  *  never return;
 778  */
 779 static __attribute__((noreturn)) void *
 780 eal_intr_thread_main(__rte_unused void *arg)
 781 {
 782         struct epoll_event ev;
 783
 784         /* host thread, never break out */
 785         for (;;) {
 786                 /* build up the epoll fd with all descriptors we are to
 787                  * wait on then pass it to the handle_interrupts function
 788                  */
 789                 static struct epoll_event pipe_event = {
 790                         .events = EPOLLIN | EPOLLPRI,
 791                 };
 792                 struct rte_intr_source *src;
 793                 unsigned numfds = 0;
 794
 795                 /* create epoll fd */
 796                 int pfd = epoll_create(1);
 797                 if (pfd < 0)
 798                         rte_panic("Cannot create epoll instance\n");
 799
 800                 pipe_event.data.fd = intr_pipe.readfd;
 801                 /**
 802                  * add pipe fd into wait list, this pipe is used to
 803                  * rebuild the wait list.
 804                  */
 805                 if (epoll_ctl(pfd, EPOLL_CTL_ADD, intr_pipe.readfd,
 806                                                 &pipe_event) < 0) {
 807                         rte_panic("Error adding fd to %d epoll_ctl, %s\n",
 808                                         intr_pipe.readfd, strerror(errno));
 809                 }
 810                 numfds++;
 811
 812                 rte_spinlock_lock(&intr_lock);
 813
 814                 TAILQ_FOREACH(src, &intr_sources, next) {
 815                         if (src->callbacks.tqh_first == NULL)
 816                                 continue; /* skip those with no callbacks */
 817                         ev.events = EPOLLIN | EPOLLPRI | EPOLLRDHUP | EPOLLHUP;
 818                         ev.data.fd = src->intr_handle.fd;
 819
 820                         /**
 821                          * add all the uio device file descriptor
 822                          * into wait list.
 823                          */
 824                         if (epoll_ctl(pfd, EPOLL_CTL_ADD,
 825                                         src->intr_handle.fd, &ev) < 0){
 826                                 rte_panic("Error adding fd %d epoll_ctl, %s\n",
 827                                         src->intr_handle.fd, strerror(errno));
 828                         }
 829                         else
 830                                 numfds++;
 831                 }
 832                 rte_spinlock_unlock(&intr_lock);
 833                 /* serve the interrupt */
 834                 eal_intr_handle_interrupts(pfd, numfds);
 835
 836                 /**
 837                  * when we return, we need to rebuild the
 838                  * list of fds to monitor.
 839                  */
 840                 close(pfd);
 841         }
 842 }
 843
 844 int
 845 rte_eal_intr_init(void)
 846 {
 847         int ret = 0, ret_1 = 0;
 848         char thread_name[RTE_MAX_THREAD_NAME_LEN];
 849
 850         /* init the global interrupt source head */
 851         TAILQ_INIT(&intr_sources);
 852
 853         /**
 854          * create a pipe which will be waited by epoll and notified to
 855          * rebuild the wait list of epoll.
 856          */
 857         if (pipe(intr_pipe.pipefd) < 0) {
 858                 rte_errno = errno;
 859                 return -1;
 860         }
 861
 862         /* create the host thread to wait/handle the interrupt */
 863         ret = pthread_create(&intr_thread, NULL,
 864                         eal_intr_thread_main, NULL);
 865         if (ret != 0) {
 866                 rte_errno = ret;
 867                 RTE_LOG(ERR, EAL,
 868                         "Failed to create thread for interrupt handling\n");
 869         } else {
 870                 /* Set thread_name for aid in debugging. */
 871                 snprintf(thread_name, RTE_MAX_THREAD_NAME_LEN,
 872                         "eal-intr-thread");
 873                 ret_1 = rte_thread_setname(intr_thread, thread_name);
 874                 if (ret_1 != 0)
 875                         RTE_LOG(DEBUG, EAL,
 876                         "Failed to set thread name for interrupt handling\n");
 877         }
 878
 879         return -ret;
 880 }
 881
 882 static void
 883 eal_intr_proc_rxtx_intr(int fd, const struct rte_intr_handle *intr_handle)
 884 {
 885         union rte_intr_read_buffer buf;
 886         int bytes_read = 0;
 887         int nbytes;
 888
 889         switch (intr_handle->type) {
 890         case RTE_INTR_HANDLE_UIO:
 891         case RTE_INTR_HANDLE_UIO_INTX:
 892                 bytes_read = sizeof(buf.uio_intr_count);
 893                 break;
 894 #ifdef VFIO_PRESENT
 895         case RTE_INTR_HANDLE_VFIO_MSIX:
 896         case RTE_INTR_HANDLE_VFIO_MSI:
 897         case RTE_INTR_HANDLE_VFIO_LEGACY:
 898                 bytes_read = sizeof(buf.vfio_intr_count);
 899                 break;
 900 #endif
 901         case RTE_INTR_HANDLE_VDEV:
 902                 bytes_read = intr_handle->efd_counter_size;
 903                 /* For vdev, number of bytes to read is set by driver */
 904                 break;
 905         case RTE_INTR_HANDLE_EXT:
 906                 return;
 907         default:
 908                 bytes_read = 1;
 909                 RTE_LOG(INFO, EAL, "unexpected intr type\n");
 910                 break;
 911         }
 912
 913         /**
 914          * read out to clear the ready-to-be-read flag
 915          * for epoll_wait.
 916          */
 917         if (bytes_read == 0)
 918                 return;
 919         do {
 920                 nbytes = read(fd, &buf, bytes_read);
 921                 if (nbytes < 0) {
 922                         if (errno == EINTR || errno == EWOULDBLOCK ||
 923                             errno == EAGAIN)
 924                                 continue;
 925                         RTE_LOG(ERR, EAL,
 926                                 "Error reading from fd %d: %s\n",
 927                                 fd, strerror(errno));
 928                 } else if (nbytes == 0)
 929                         RTE_LOG(ERR, EAL, "Read nothing from fd %d\n", fd);
 930                 return;
 931         } while (1);
 932 }
 933
 934 static int
 935 eal_epoll_process_event(struct epoll_event *evs, unsigned int n,
 936                         struct rte_epoll_event *events)
 937 {
 938         unsigned int i, count = 0;
 939         struct rte_epoll_event *rev;
 940
 941         for (i = 0; i < n; i++) {
 942                 rev = evs[i].data.ptr;
 943                 if (!rev || !rte_atomic32_cmpset(&rev->status, RTE_EPOLL_VALID,
 944                                                  RTE_EPOLL_EXEC))
 945                         continue;
 946
 947                 events[count].status        = RTE_EPOLL_VALID;
 948                 events[count].fd            = rev->fd;
 949                 events[count].epfd          = rev->epfd;
 950                 events[count].epdata.event  = rev->epdata.event;
 951                 events[count].epdata.data   = rev->epdata.data;
 952                 if (rev->epdata.cb_fun)
 953                         rev->epdata.cb_fun(rev->fd,
 954                                            rev->epdata.cb_arg);
 955
 956                 rte_compiler_barrier();
 957                 rev->status = RTE_EPOLL_VALID;
 958                 count++;
 959         }
 960         return count;
 961 }
 962
 963 static inline int
 964 eal_init_tls_epfd(void)
 965 {
 966         int pfd = epoll_create(255);
 967
 968         if (pfd < 0) {
 969                 RTE_LOG(ERR, EAL,
 970                         "Cannot create epoll instance\n");
 971                 return -1;
 972         }
 973         return pfd;
 974 }
 975
 976 int
 977 rte_intr_tls_epfd(void)
 978 {
 979         if (RTE_PER_LCORE(_epfd) == -1)
 980                 RTE_PER_LCORE(_epfd) = eal_init_tls_epfd();
 981
 982         return RTE_PER_LCORE(_epfd);
 983 }
 984
 985 int
 986 rte_epoll_wait(int epfd, struct rte_epoll_event *events,
 987                int maxevents, int timeout)
 988 {
 989         struct epoll_event evs[maxevents];
 990         int rc;
 991
 992         if (!events) {
 993                 RTE_LOG(ERR, EAL, "rte_epoll_event can't be NULL\n");
 994                 return -1;
 995         }
 996
 997         /* using per thread epoll fd */
 998         if (epfd == RTE_EPOLL_PER_THREAD)
 999                 epfd = rte_intr_tls_epfd();
1000
1001         while (1) {
1002                 rc = epoll_wait(epfd, evs, maxevents, timeout);
1003                 if (likely(rc > 0)) {
1004                         /* epoll_wait has at least one fd ready to read */
1005                         rc = eal_epoll_process_event(evs, rc, events);
1006                         break;
1007                 } else if (rc < 0) {
1008                         if (errno == EINTR)
1009                                 continue;
1010                         /* epoll_wait fail */
1011                         RTE_LOG(ERR, EAL, "epoll_wait returns with fail %s\n",
1012                                 strerror(errno));
1013                         rc = -1;
1014                         break;
1015                 } else {
1016                         /* rc == 0, epoll_wait timed out */
1017                         break;
1018                 }
1019         }
1020
1021         return rc;
1022 }
1023
1024 static inline void
1025 eal_epoll_data_safe_free(struct rte_epoll_event *ev)
1026 {
1027         while (!rte_atomic32_cmpset(&ev->status, RTE_EPOLL_VALID,
1028                                     RTE_EPOLL_INVALID))
1029                 while (ev->status != RTE_EPOLL_VALID)
1030                         rte_pause();
1031         memset(&ev->epdata, 0, sizeof(ev->epdata));
1032         ev->fd = -1;
1033         ev->epfd = -1;
1034 }
1035
1036 int
1037 rte_epoll_ctl(int epfd, int op, int fd,
1038               struct rte_epoll_event *event)
1039 {
1040         struct epoll_event ev;
1041
1042         if (!event) {
1043                 RTE_LOG(ERR, EAL, "rte_epoll_event can't be NULL\n");
1044                 return -1;
1045         }
1046
1047         /* using per thread epoll fd */
1048         if (epfd == RTE_EPOLL_PER_THREAD)
1049                 epfd = rte_intr_tls_epfd();
1050
1051         if (op == EPOLL_CTL_ADD) {
1052                 event->status = RTE_EPOLL_VALID;
1053                 event->fd = fd;  /* ignore fd in event */
1054                 event->epfd = epfd;
1055                 ev.data.ptr = (void *)event;
1056         }
1057
1058         ev.events = event->epdata.event;
1059         if (epoll_ctl(epfd, op, fd, &ev) < 0) {
1060                 RTE_LOG(ERR, EAL, "Error op %d fd %d epoll_ctl, %s\n",
1061                         op, fd, strerror(errno));
1062                 if (op == EPOLL_CTL_ADD)
1063                         /* rollback status when CTL_ADD fail */
1064                         event->status = RTE_EPOLL_INVALID;
1065                 return -1;
1066         }
1067
1068         if (op == EPOLL_CTL_DEL && event->status != RTE_EPOLL_INVALID)
1069                 eal_epoll_data_safe_free(event);
1070
1071         return 0;
1072 }
1073
1074 int
1075 rte_intr_rx_ctl(struct rte_intr_handle *intr_handle, int epfd,
1076                 int op, unsigned int vec, void *data)
1077 {
1078         struct rte_epoll_event *rev;
1079         struct rte_epoll_data *epdata;
1080         int epfd_op;
1081         unsigned int efd_idx;
1082         int rc = 0;
1083
1084         efd_idx = (vec >= RTE_INTR_VEC_RXTX_OFFSET) ?
1085                 (vec - RTE_INTR_VEC_RXTX_OFFSET) : vec;
1086
1087         if (!intr_handle || intr_handle->nb_efd == 0 ||
1088             efd_idx >= intr_handle->nb_efd) {
1089                 RTE_LOG(ERR, EAL, "Wrong intr vector number.\n");
1090                 return -EPERM;
1091         }
1092
1093         switch (op) {
1094         case RTE_INTR_EVENT_ADD:
1095                 epfd_op = EPOLL_CTL_ADD;
1096                 rev = &intr_handle->elist[efd_idx];
1097                 if (rev->status != RTE_EPOLL_INVALID) {
1098                         RTE_LOG(INFO, EAL, "Event already been added.\n");
1099                         return -EEXIST;
1100                 }
1101
1102                 /* attach to intr vector fd */
1103                 epdata = &rev->epdata;
1104                 epdata->event  = EPOLLIN | EPOLLPRI | EPOLLET;
1105                 epdata->data   = data;
1106                 epdata->cb_fun = (rte_intr_event_cb_t)eal_intr_proc_rxtx_intr;
1107                 epdata->cb_arg = (void *)intr_handle;
1108                 rc = rte_epoll_ctl(epfd, epfd_op,
1109                                    intr_handle->efds[efd_idx], rev);
1110                 if (!rc)
1111                         RTE_LOG(DEBUG, EAL,
1112                                 "efd %d associated with vec %d added on epfd %d"
1113                                 "\n", rev->fd, vec, epfd);
1114                 else
1115                         rc = -EPERM;
1116                 break;
1117         case RTE_INTR_EVENT_DEL:
1118                 epfd_op = EPOLL_CTL_DEL;
1119                 rev = &intr_handle->elist[efd_idx];
1120                 if (rev->status == RTE_EPOLL_INVALID) {
1121                         RTE_LOG(INFO, EAL, "Event does not exist.\n");
1122                         return -EPERM;
1123                 }
1124
1125                 rc = rte_epoll_ctl(rev->epfd, epfd_op, rev->fd, rev);
1126                 if (rc)
1127                         rc = -EPERM;
1128                 break;
1129         default:
1130                 RTE_LOG(ERR, EAL, "event op type mismatch\n");
1131                 rc = -EPERM;
1132         }
1133
1134         return rc;
1135 }
1136
1137 void
1138 rte_intr_free_epoll_fd(struct rte_intr_handle *intr_handle)
1139 {
1140         uint32_t i;
1141         struct rte_epoll_event *rev;
1142
1143         for (i = 0; i < intr_handle->nb_efd; i++) {
1144                 rev = &intr_handle->elist[i];
1145                 if (rev->status == RTE_EPOLL_INVALID)
1146                         continue;
1147                 if (rte_epoll_ctl(rev->epfd, EPOLL_CTL_DEL, rev->fd, rev)) {
1148                         /* force free if the entry valid */
1149                         eal_epoll_data_safe_free(rev);
1150                         rev->status = RTE_EPOLL_INVALID;
1151                 }
1152         }
1153 }
1154
1155 int
1156 rte_intr_efd_enable(struct rte_intr_handle *intr_handle, uint32_t nb_efd)
1157 {
1158         uint32_t i;
1159         int fd;
1160         uint32_t n = RTE_MIN(nb_efd, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
1161
1162         assert(nb_efd != 0);
1163
1164         if (intr_handle->type == RTE_INTR_HANDLE_VFIO_MSIX) {
1165                 for (i = 0; i < n; i++) {
1166                         fd = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC);
1167                         if (fd < 0) {
1168                                 RTE_LOG(ERR, EAL,
1169                                         "can't setup eventfd, error %i (%s)\n",
1170                                         errno, strerror(errno));
1171                                 return -errno;
1172                         }
1173                         intr_handle->efds[i] = fd;
1174                 }
1175                 intr_handle->nb_efd   = n;
1176                 intr_handle->max_intr = NB_OTHER_INTR + n;
1177         } else if (intr_handle->type == RTE_INTR_HANDLE_VDEV) {
1178                 /* only check, initialization would be done in vdev driver.*/
1179                 if (intr_handle->efd_counter_size >
1180                     sizeof(union rte_intr_read_buffer)) {
1181                         RTE_LOG(ERR, EAL, "the efd_counter_size is oversized");
1182                         return -EINVAL;
1183                 }
1184         } else {
1185                 intr_handle->efds[0]  = intr_handle->fd;
1186                 intr_handle->nb_efd   = RTE_MIN(nb_efd, 1U);
1187                 intr_handle->max_intr = NB_OTHER_INTR;
1188         }
1189
1190         return 0;
1191 }
1192
1193 void
1194 rte_intr_efd_disable(struct rte_intr_handle *intr_handle)
1195 {
1196         uint32_t i;
1197
1198         rte_intr_free_epoll_fd(intr_handle);
1199         if (intr_handle->max_intr > intr_handle->nb_efd) {
1200                 for (i = 0; i < intr_handle->nb_efd; i++)
1201                         close(intr_handle->efds[i]);
1202         }
1203         intr_handle->nb_efd = 0;
1204         intr_handle->max_intr = 0;
1205 }
1206
1207 int
1208 rte_intr_dp_is_en(struct rte_intr_handle *intr_handle)
1209 {
1210         return !(!intr_handle->nb_efd);
1211 }
1212
1213 int
1214 rte_intr_allow_others(struct rte_intr_handle *intr_handle)
1215 {
1216         if (!rte_intr_dp_is_en(intr_handle))
1217                 return 1;
1218         else
1219                 return !!(intr_handle->max_intr - intr_handle->nb_efd);
1220 }
1221
1222 int
1223 rte_intr_cap_multiple(struct rte_intr_handle *intr_handle)
1224 {
1225         if (intr_handle->type == RTE_INTR_HANDLE_VFIO_MSIX)
1226                 return 1;
1227
1228         if (intr_handle->type == RTE_INTR_HANDLE_VDEV)
1229                 return 1;
1230
1231         return 0;
1232 }