lib/librte_eal/linux/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_errno.h>
  34 #include <rte_spinlock.h>
  35 #include <rte_pause.h>
  36 #include <rte_vfio.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         uint8_t pending_delete;      /**< delete after callback is called */
  80         rte_intr_unregister_callback_fn ucb_fn; /**< fn to call before cb is deleted */
  81 };
  82
  83 struct rte_intr_source {
  84         TAILQ_ENTRY(rte_intr_source) next;
  85         struct rte_intr_handle intr_handle; /**< interrupt handle */
  86         struct rte_intr_cb_list callbacks;  /**< user callbacks */
  87         uint32_t active;
  88 };
  89
  90 /* global spinlock for interrupt data operation */
  91 static rte_spinlock_t intr_lock = RTE_SPINLOCK_INITIALIZER;
  92
  93 /* union buffer for pipe read/write */
  94 static union intr_pipefds intr_pipe;
  95
  96 /* interrupt sources list */
  97 static struct rte_intr_source_list intr_sources;
  98
  99 /* interrupt handling thread */
 100 static pthread_t intr_thread;
 101
 102 /* VFIO interrupts */
 103 #ifdef VFIO_PRESENT
 104
 105 #define IRQ_SET_BUF_LEN  (sizeof(struct vfio_irq_set) + sizeof(int))
 106 /* irq set buffer length for queue interrupts and LSC interrupt */
 107 #define MSIX_IRQ_SET_BUF_LEN (sizeof(struct vfio_irq_set) + \
 108                               sizeof(int) * (RTE_MAX_RXTX_INTR_VEC_ID + 1))
 109
 110 /* enable legacy (INTx) interrupts */
 111 static int
 112 vfio_enable_intx(const struct rte_intr_handle *intr_handle) {
 113         struct vfio_irq_set *irq_set;
 114         char irq_set_buf[IRQ_SET_BUF_LEN];
 115         int len, ret;
 116         int *fd_ptr;
 117
 118         len = sizeof(irq_set_buf);
 119
 120         /* enable INTx */
 121         irq_set = (struct vfio_irq_set *) irq_set_buf;
 122         irq_set->argsz = len;
 123         irq_set->count = 1;
 124         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
 125         irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
 126         irq_set->start = 0;
 127         fd_ptr = (int *) &irq_set->data;
 128         *fd_ptr = intr_handle->fd;
 129
 130         ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 131
 132         if (ret) {
 133                 RTE_LOG(ERR, EAL, "Error enabling INTx interrupts for fd %d\n",
 134                                                 intr_handle->fd);
 135                 return -1;
 136         }
 137
 138         /* unmask INTx after enabling */
 139         memset(irq_set, 0, len);
 140         len = sizeof(struct vfio_irq_set);
 141         irq_set->argsz = len;
 142         irq_set->count = 1;
 143         irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK;
 144         irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
 145         irq_set->start = 0;
 146
 147         ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 148
 149         if (ret) {
 150                 RTE_LOG(ERR, EAL, "Error unmasking INTx interrupts for fd %d\n",
 151                                                 intr_handle->fd);
 152                 return -1;
 153         }
 154         return 0;
 155 }
 156
 157 /* disable legacy (INTx) interrupts */
 158 static int
 159 vfio_disable_intx(const struct rte_intr_handle *intr_handle) {
 160         struct vfio_irq_set *irq_set;
 161         char irq_set_buf[IRQ_SET_BUF_LEN];
 162         int len, ret;
 163
 164         len = sizeof(struct vfio_irq_set);
 165
 166         /* mask interrupts before disabling */
 167         irq_set = (struct vfio_irq_set *) irq_set_buf;
 168         irq_set->argsz = len;
 169         irq_set->count = 1;
 170         irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK;
 171         irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
 172         irq_set->start = 0;
 173
 174         ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 175
 176         if (ret) {
 177                 RTE_LOG(ERR, EAL, "Error masking INTx interrupts for fd %d\n",
 178                                                 intr_handle->fd);
 179                 return -1;
 180         }
 181
 182         /* disable INTx*/
 183         memset(irq_set, 0, len);
 184         irq_set->argsz = len;
 185         irq_set->count = 0;
 186         irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER;
 187         irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
 188         irq_set->start = 0;
 189
 190         ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 191
 192         if (ret) {
 193                 RTE_LOG(ERR, EAL,
 194                         "Error disabling INTx interrupts for fd %d\n", intr_handle->fd);
 195                 return -1;
 196         }
 197         return 0;
 198 }
 199
 200 /* unmask/ack legacy (INTx) interrupts */
 201 static int
 202 vfio_ack_intx(const struct rte_intr_handle *intr_handle)
 203 {
 204         struct vfio_irq_set irq_set;
 205
 206         /* unmask INTx */
 207         memset(&irq_set, 0, sizeof(irq_set));
 208         irq_set.argsz = sizeof(irq_set);
 209         irq_set.count = 1;
 210         irq_set.flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK;
 211         irq_set.index = VFIO_PCI_INTX_IRQ_INDEX;
 212         irq_set.start = 0;
 213
 214         if (ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, &irq_set)) {
 215                 RTE_LOG(ERR, EAL, "Error unmasking INTx interrupts for fd %d\n",
 216                         intr_handle->fd);
 217                 return -1;
 218         }
 219         return 0;
 220 }
 221
 222 /* enable MSI interrupts */
 223 static int
 224 vfio_enable_msi(const struct rte_intr_handle *intr_handle) {
 225         int len, ret;
 226         char irq_set_buf[IRQ_SET_BUF_LEN];
 227         struct vfio_irq_set *irq_set;
 228         int *fd_ptr;
 229
 230         len = sizeof(irq_set_buf);
 231
 232         irq_set = (struct vfio_irq_set *) irq_set_buf;
 233         irq_set->argsz = len;
 234         irq_set->count = 1;
 235         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
 236         irq_set->index = VFIO_PCI_MSI_IRQ_INDEX;
 237         irq_set->start = 0;
 238         fd_ptr = (int *) &irq_set->data;
 239         *fd_ptr = intr_handle->fd;
 240
 241         ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 242
 243         if (ret) {
 244                 RTE_LOG(ERR, EAL, "Error enabling MSI interrupts for fd %d\n",
 245                                                 intr_handle->fd);
 246                 return -1;
 247         }
 248         return 0;
 249 }
 250
 251 /* disable MSI interrupts */
 252 static int
 253 vfio_disable_msi(const struct rte_intr_handle *intr_handle) {
 254         struct vfio_irq_set *irq_set;
 255         char irq_set_buf[IRQ_SET_BUF_LEN];
 256         int len, ret;
 257
 258         len = sizeof(struct vfio_irq_set);
 259
 260         irq_set = (struct vfio_irq_set *) irq_set_buf;
 261         irq_set->argsz = len;
 262         irq_set->count = 0;
 263         irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER;
 264         irq_set->index = VFIO_PCI_MSI_IRQ_INDEX;
 265         irq_set->start = 0;
 266
 267         ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 268
 269         if (ret)
 270                 RTE_LOG(ERR, EAL,
 271                         "Error disabling MSI interrupts for fd %d\n", intr_handle->fd);
 272
 273         return ret;
 274 }
 275
 276 /* enable MSI-X interrupts */
 277 static int
 278 vfio_enable_msix(const struct rte_intr_handle *intr_handle) {
 279         int len, ret;
 280         char irq_set_buf[MSIX_IRQ_SET_BUF_LEN];
 281         struct vfio_irq_set *irq_set;
 282         int *fd_ptr;
 283
 284         len = sizeof(irq_set_buf);
 285
 286         irq_set = (struct vfio_irq_set *) irq_set_buf;
 287         irq_set->argsz = len;
 288         /* 0 < irq_set->count < RTE_MAX_RXTX_INTR_VEC_ID + 1 */
 289         irq_set->count = intr_handle->max_intr ?
 290                 (intr_handle->max_intr > RTE_MAX_RXTX_INTR_VEC_ID + 1 ?
 291                 RTE_MAX_RXTX_INTR_VEC_ID + 1 : intr_handle->max_intr) : 1;
 292         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
 293         irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
 294         irq_set->start = 0;
 295         fd_ptr = (int *) &irq_set->data;
 296         /* INTR vector offset 0 reserve for non-efds mapping */
 297         fd_ptr[RTE_INTR_VEC_ZERO_OFFSET] = intr_handle->fd;
 298         memcpy(&fd_ptr[RTE_INTR_VEC_RXTX_OFFSET], intr_handle->efds,
 299                 sizeof(*intr_handle->efds) * intr_handle->nb_efd);
 300
 301         ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 302
 303         if (ret) {
 304                 RTE_LOG(ERR, EAL, "Error enabling MSI-X interrupts for fd %d\n",
 305                                                 intr_handle->fd);
 306                 return -1;
 307         }
 308
 309         return 0;
 310 }
 311
 312 /* disable MSI-X interrupts */
 313 static int
 314 vfio_disable_msix(const struct rte_intr_handle *intr_handle) {
 315         struct vfio_irq_set *irq_set;
 316         char irq_set_buf[MSIX_IRQ_SET_BUF_LEN];
 317         int len, ret;
 318
 319         len = sizeof(struct vfio_irq_set);
 320
 321         irq_set = (struct vfio_irq_set *) irq_set_buf;
 322         irq_set->argsz = len;
 323         irq_set->count = 0;
 324         irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER;
 325         irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
 326         irq_set->start = 0;
 327
 328         ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 329
 330         if (ret)
 331                 RTE_LOG(ERR, EAL,
 332                         "Error disabling MSI-X interrupts for fd %d\n", intr_handle->fd);
 333
 334         return ret;
 335 }
 336
 337 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
 338 /* enable req notifier */
 339 static int
 340 vfio_enable_req(const struct rte_intr_handle *intr_handle)
 341 {
 342         int len, ret;
 343         char irq_set_buf[IRQ_SET_BUF_LEN];
 344         struct vfio_irq_set *irq_set;
 345         int *fd_ptr;
 346
 347         len = sizeof(irq_set_buf);
 348
 349         irq_set = (struct vfio_irq_set *) irq_set_buf;
 350         irq_set->argsz = len;
 351         irq_set->count = 1;
 352         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
 353                          VFIO_IRQ_SET_ACTION_TRIGGER;
 354         irq_set->index = VFIO_PCI_REQ_IRQ_INDEX;
 355         irq_set->start = 0;
 356         fd_ptr = (int *) &irq_set->data;
 357         *fd_ptr = intr_handle->fd;
 358
 359         ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 360
 361         if (ret) {
 362                 RTE_LOG(ERR, EAL, "Error enabling req interrupts for fd %d\n",
 363                                                 intr_handle->fd);
 364                 return -1;
 365         }
 366
 367         return 0;
 368 }
 369
 370 /* disable req notifier */
 371 static int
 372 vfio_disable_req(const struct rte_intr_handle *intr_handle)
 373 {
 374         struct vfio_irq_set *irq_set;
 375         char irq_set_buf[IRQ_SET_BUF_LEN];
 376         int len, ret;
 377
 378         len = sizeof(struct vfio_irq_set);
 379
 380         irq_set = (struct vfio_irq_set *) irq_set_buf;
 381         irq_set->argsz = len;
 382         irq_set->count = 0;
 383         irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER;
 384         irq_set->index = VFIO_PCI_REQ_IRQ_INDEX;
 385         irq_set->start = 0;
 386
 387         ret = ioctl(intr_handle->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 388
 389         if (ret)
 390                 RTE_LOG(ERR, EAL, "Error disabling req interrupts for fd %d\n",
 391                         intr_handle->fd);
 392
 393         return ret;
 394 }
 395 #endif
 396 #endif
 397
 398 static int
 399 uio_intx_intr_disable(const struct rte_intr_handle *intr_handle)
 400 {
 401         unsigned char command_high;
 402
 403         /* use UIO config file descriptor for uio_pci_generic */
 404         if (pread(intr_handle->uio_cfg_fd, &command_high, 1, 5) != 1) {
 405                 RTE_LOG(ERR, EAL,
 406                         "Error reading interrupts status for fd %d\n",
 407                         intr_handle->uio_cfg_fd);
 408                 return -1;
 409         }
 410         /* disable interrupts */
 411         command_high |= 0x4;
 412         if (pwrite(intr_handle->uio_cfg_fd, &command_high, 1, 5) != 1) {
 413                 RTE_LOG(ERR, EAL,
 414                         "Error disabling interrupts for fd %d\n",
 415                         intr_handle->uio_cfg_fd);
 416                 return -1;
 417         }
 418
 419         return 0;
 420 }
 421
 422 static int
 423 uio_intx_intr_enable(const struct rte_intr_handle *intr_handle)
 424 {
 425         unsigned char command_high;
 426
 427         /* use UIO config file descriptor for uio_pci_generic */
 428         if (pread(intr_handle->uio_cfg_fd, &command_high, 1, 5) != 1) {
 429                 RTE_LOG(ERR, EAL,
 430                         "Error reading interrupts status for fd %d\n",
 431                         intr_handle->uio_cfg_fd);
 432                 return -1;
 433         }
 434         /* enable interrupts */
 435         command_high &= ~0x4;
 436         if (pwrite(intr_handle->uio_cfg_fd, &command_high, 1, 5) != 1) {
 437                 RTE_LOG(ERR, EAL,
 438                         "Error enabling interrupts for fd %d\n",
 439                         intr_handle->uio_cfg_fd);
 440                 return -1;
 441         }
 442
 443         return 0;
 444 }
 445
 446 static int
 447 uio_intr_disable(const struct rte_intr_handle *intr_handle)
 448 {
 449         const int value = 0;
 450
 451         if (write(intr_handle->fd, &value, sizeof(value)) < 0) {
 452                 RTE_LOG(ERR, EAL,
 453                         "Error disabling interrupts for fd %d (%s)\n",
 454                         intr_handle->fd, strerror(errno));
 455                 return -1;
 456         }
 457         return 0;
 458 }
 459
 460 static int
 461 uio_intr_enable(const struct rte_intr_handle *intr_handle)
 462 {
 463         const int value = 1;
 464
 465         if (write(intr_handle->fd, &value, sizeof(value)) < 0) {
 466                 RTE_LOG(ERR, EAL,
 467                         "Error enabling interrupts for fd %d (%s)\n",
 468                         intr_handle->fd, strerror(errno));
 469                 return -1;
 470         }
 471         return 0;
 472 }
 473
 474 int
 475 rte_intr_callback_register(const struct rte_intr_handle *intr_handle,
 476                         rte_intr_callback_fn cb, void *cb_arg)
 477 {
 478         int ret, wake_thread;
 479         struct rte_intr_source *src;
 480         struct rte_intr_callback *callback;
 481
 482         wake_thread = 0;
 483
 484         /* first do parameter checking */
 485         if (intr_handle == NULL || intr_handle->fd < 0 || cb == NULL) {
 486                 RTE_LOG(ERR, EAL,
 487                         "Registering with invalid input parameter\n");
 488                 return -EINVAL;
 489         }
 490
 491         /* allocate a new interrupt callback entity */
 492         callback = calloc(1, sizeof(*callback));
 493         if (callback == NULL) {
 494                 RTE_LOG(ERR, EAL, "Can not allocate memory\n");
 495                 return -ENOMEM;
 496         }
 497         callback->cb_fn = cb;
 498         callback->cb_arg = cb_arg;
 499         callback->pending_delete = 0;
 500         callback->ucb_fn = NULL;
 501
 502         rte_spinlock_lock(&intr_lock);
 503
 504         /* check if there is at least one callback registered for the fd */
 505         TAILQ_FOREACH(src, &intr_sources, next) {
 506                 if (src->intr_handle.fd == intr_handle->fd) {
 507                         /* we had no interrupts for this */
 508                         if (TAILQ_EMPTY(&src->callbacks))
 509                                 wake_thread = 1;
 510
 511                         TAILQ_INSERT_TAIL(&(src->callbacks), callback, next);
 512                         ret = 0;
 513                         break;
 514                 }
 515         }
 516
 517         /* no existing callbacks for this - add new source */
 518         if (src == NULL) {
 519                 src = calloc(1, sizeof(*src));
 520                 if (src == NULL) {
 521                         RTE_LOG(ERR, EAL, "Can not allocate memory\n");
 522                         free(callback);
 523                         ret = -ENOMEM;
 524                 } else {
 525                         src->intr_handle = *intr_handle;
 526                         TAILQ_INIT(&src->callbacks);
 527                         TAILQ_INSERT_TAIL(&(src->callbacks), callback, next);
 528                         TAILQ_INSERT_TAIL(&intr_sources, src, next);
 529                         wake_thread = 1;
 530                         ret = 0;
 531                 }
 532         }
 533
 534         rte_spinlock_unlock(&intr_lock);
 535
 536         /**
 537          * check if need to notify the pipe fd waited by epoll_wait to
 538          * rebuild the wait list.
 539          */
 540         if (wake_thread)
 541                 if (write(intr_pipe.writefd, "1", 1) < 0)
 542                         return -EPIPE;
 543
 544         return ret;
 545 }
 546
 547 int
 548 rte_intr_callback_unregister_pending(const struct rte_intr_handle *intr_handle,
 549                                 rte_intr_callback_fn cb_fn, void *cb_arg,
 550                                 rte_intr_unregister_callback_fn ucb_fn)
 551 {
 552         int ret;
 553         struct rte_intr_source *src;
 554         struct rte_intr_callback *cb, *next;
 555
 556         /* do parameter checking first */
 557         if (intr_handle == NULL || intr_handle->fd < 0) {
 558                 RTE_LOG(ERR, EAL,
 559                 "Unregistering with invalid input parameter\n");
 560                 return -EINVAL;
 561         }
 562
 563         rte_spinlock_lock(&intr_lock);
 564
 565         /* check if the insterrupt source for the fd is existent */
 566         TAILQ_FOREACH(src, &intr_sources, next)
 567                 if (src->intr_handle.fd == intr_handle->fd)
 568                         break;
 569
 570         /* No interrupt source registered for the fd */
 571         if (src == NULL) {
 572                 ret = -ENOENT;
 573
 574         /* only usable if the source is active */
 575         } else if (src->active == 0) {
 576                 ret = -EAGAIN;
 577
 578         } else {
 579                 ret = 0;
 580
 581                 /* walk through the callbacks and mark all that match. */
 582                 for (cb = TAILQ_FIRST(&src->callbacks); cb != NULL; cb = next) {
 583                         next = TAILQ_NEXT(cb, next);
 584                         if (cb->cb_fn == cb_fn && (cb_arg == (void *)-1 ||
 585                                         cb->cb_arg == cb_arg)) {
 586                                 cb->pending_delete = 1;
 587                                 cb->ucb_fn = ucb_fn;
 588                                 ret++;
 589                         }
 590                 }
 591         }
 592
 593         rte_spinlock_unlock(&intr_lock);
 594
 595         return ret;
 596 }
 597
 598 int
 599 rte_intr_callback_unregister(const struct rte_intr_handle *intr_handle,
 600                         rte_intr_callback_fn cb_fn, void *cb_arg)
 601 {
 602         int ret;
 603         struct rte_intr_source *src;
 604         struct rte_intr_callback *cb, *next;
 605
 606         /* do parameter checking first */
 607         if (intr_handle == NULL || intr_handle->fd < 0) {
 608                 RTE_LOG(ERR, EAL,
 609                 "Unregistering with invalid input parameter\n");
 610                 return -EINVAL;
 611         }
 612
 613         rte_spinlock_lock(&intr_lock);
 614
 615         /* check if the insterrupt source for the fd is existent */
 616         TAILQ_FOREACH(src, &intr_sources, next)
 617                 if (src->intr_handle.fd == intr_handle->fd)
 618                         break;
 619
 620         /* No interrupt source registered for the fd */
 621         if (src == NULL) {
 622                 ret = -ENOENT;
 623
 624         /* interrupt source has some active callbacks right now. */
 625         } else if (src->active != 0) {
 626                 ret = -EAGAIN;
 627
 628         /* ok to remove. */
 629         } else {
 630                 ret = 0;
 631
 632                 /*walk through the callbacks and remove all that match. */
 633                 for (cb = TAILQ_FIRST(&src->callbacks); cb != NULL; cb = next) {
 634
 635                         next = TAILQ_NEXT(cb, next);
 636
 637                         if (cb->cb_fn == cb_fn && (cb_arg == (void *)-1 ||
 638                                         cb->cb_arg == cb_arg)) {
 639                                 TAILQ_REMOVE(&src->callbacks, cb, next);
 640                                 free(cb);
 641                                 ret++;
 642                         }
 643                 }
 644
 645                 /* all callbacks for that source are removed. */
 646                 if (TAILQ_EMPTY(&src->callbacks)) {
 647                         TAILQ_REMOVE(&intr_sources, src, next);
 648                         free(src);
 649                 }
 650         }
 651
 652         rte_spinlock_unlock(&intr_lock);
 653
 654         /* notify the pipe fd waited by epoll_wait to rebuild the wait list */
 655         if (ret >= 0 && write(intr_pipe.writefd, "1", 1) < 0) {
 656                 ret = -EPIPE;
 657         }
 658
 659         return ret;
 660 }
 661
 662 int
 663 rte_intr_enable(const struct rte_intr_handle *intr_handle)
 664 {
 665         if (intr_handle && intr_handle->type == RTE_INTR_HANDLE_VDEV)
 666                 return 0;
 667
 668         if (!intr_handle || intr_handle->fd < 0 || intr_handle->uio_cfg_fd < 0)
 669                 return -1;
 670
 671         switch (intr_handle->type){
 672         /* write to the uio fd to enable the interrupt */
 673         case RTE_INTR_HANDLE_UIO:
 674                 if (uio_intr_enable(intr_handle))
 675                         return -1;
 676                 break;
 677         case RTE_INTR_HANDLE_UIO_INTX:
 678                 if (uio_intx_intr_enable(intr_handle))
 679                         return -1;
 680                 break;
 681         /* not used at this moment */
 682         case RTE_INTR_HANDLE_ALARM:
 683                 return -1;
 684 #ifdef VFIO_PRESENT
 685         case RTE_INTR_HANDLE_VFIO_MSIX:
 686                 if (vfio_enable_msix(intr_handle))
 687                         return -1;
 688                 break;
 689         case RTE_INTR_HANDLE_VFIO_MSI:
 690                 if (vfio_enable_msi(intr_handle))
 691                         return -1;
 692                 break;
 693         case RTE_INTR_HANDLE_VFIO_LEGACY:
 694                 if (vfio_enable_intx(intr_handle))
 695                         return -1;
 696                 break;
 697 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
 698         case RTE_INTR_HANDLE_VFIO_REQ:
 699                 if (vfio_enable_req(intr_handle))
 700                         return -1;
 701                 break;
 702 #endif
 703 #endif
 704         /* not used at this moment */
 705         case RTE_INTR_HANDLE_DEV_EVENT:
 706                 return -1;
 707         /* unknown handle type */
 708         default:
 709                 RTE_LOG(ERR, EAL,
 710                         "Unknown handle type of fd %d\n",
 711                                         intr_handle->fd);
 712                 return -1;
 713         }
 714
 715         return 0;
 716 }
 717
 718 /**
 719  * PMD generally calls this function at the end of its IRQ callback.
 720  * Internally, it unmasks the interrupt if possible.
 721  *
 722  * For INTx, unmasking is required as the interrupt is auto-masked prior to
 723  * invoking callback.
 724  *
 725  * For MSI/MSI-X, unmasking is typically not needed as the interrupt is not
 726  * auto-masked. In fact, for interrupt handle types VFIO_MSIX and VFIO_MSI,
 727  * this function is no-op.
 728  */
 729 int
 730 rte_intr_ack(const struct rte_intr_handle *intr_handle)
 731 {
 732         if (intr_handle && intr_handle->type == RTE_INTR_HANDLE_VDEV)
 733                 return 0;
 734
 735         if (!intr_handle || intr_handle->fd < 0 || intr_handle->uio_cfg_fd < 0)
 736                 return -1;
 737
 738         switch (intr_handle->type) {
 739         /* Both acking and enabling are same for UIO */
 740         case RTE_INTR_HANDLE_UIO:
 741                 if (uio_intr_enable(intr_handle))
 742                         return -1;
 743                 break;
 744         case RTE_INTR_HANDLE_UIO_INTX:
 745                 if (uio_intx_intr_enable(intr_handle))
 746                         return -1;
 747                 break;
 748         /* not used at this moment */
 749         case RTE_INTR_HANDLE_ALARM:
 750                 return -1;
 751 #ifdef VFIO_PRESENT
 752         /* VFIO MSI* is implicitly acked unlike INTx, nothing to do */
 753         case RTE_INTR_HANDLE_VFIO_MSIX:
 754         case RTE_INTR_HANDLE_VFIO_MSI:
 755                 return 0;
 756         case RTE_INTR_HANDLE_VFIO_LEGACY:
 757                 if (vfio_ack_intx(intr_handle))
 758                         return -1;
 759                 break;
 760 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
 761         case RTE_INTR_HANDLE_VFIO_REQ:
 762                 return -1;
 763 #endif
 764 #endif
 765         /* not used at this moment */
 766         case RTE_INTR_HANDLE_DEV_EVENT:
 767                 return -1;
 768         /* unknown handle type */
 769         default:
 770                 RTE_LOG(ERR, EAL, "Unknown handle type of fd %d\n",
 771                         intr_handle->fd);
 772                 return -1;
 773         }
 774
 775         return 0;
 776 }
 777
 778 int
 779 rte_intr_disable(const struct rte_intr_handle *intr_handle)
 780 {
 781         if (intr_handle && intr_handle->type == RTE_INTR_HANDLE_VDEV)
 782                 return 0;
 783
 784         if (!intr_handle || intr_handle->fd < 0 || intr_handle->uio_cfg_fd < 0)
 785                 return -1;
 786
 787         switch (intr_handle->type){
 788         /* write to the uio fd to disable the interrupt */
 789         case RTE_INTR_HANDLE_UIO:
 790                 if (uio_intr_disable(intr_handle))
 791                         return -1;
 792                 break;
 793         case RTE_INTR_HANDLE_UIO_INTX:
 794                 if (uio_intx_intr_disable(intr_handle))
 795                         return -1;
 796                 break;
 797         /* not used at this moment */
 798         case RTE_INTR_HANDLE_ALARM:
 799                 return -1;
 800 #ifdef VFIO_PRESENT
 801         case RTE_INTR_HANDLE_VFIO_MSIX:
 802                 if (vfio_disable_msix(intr_handle))
 803                         return -1;
 804                 break;
 805         case RTE_INTR_HANDLE_VFIO_MSI:
 806                 if (vfio_disable_msi(intr_handle))
 807                         return -1;
 808                 break;
 809         case RTE_INTR_HANDLE_VFIO_LEGACY:
 810                 if (vfio_disable_intx(intr_handle))
 811                         return -1;
 812                 break;
 813 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
 814         case RTE_INTR_HANDLE_VFIO_REQ:
 815                 if (vfio_disable_req(intr_handle))
 816                         return -1;
 817                 break;
 818 #endif
 819 #endif
 820         /* not used at this moment */
 821         case RTE_INTR_HANDLE_DEV_EVENT:
 822                 return -1;
 823         /* unknown handle type */
 824         default:
 825                 RTE_LOG(ERR, EAL,
 826                         "Unknown handle type of fd %d\n",
 827                                         intr_handle->fd);
 828                 return -1;
 829         }
 830
 831         return 0;
 832 }
 833
 834 static int
 835 eal_intr_process_interrupts(struct epoll_event *events, int nfds)
 836 {
 837         bool call = false;
 838         int n, bytes_read, rv;
 839         struct rte_intr_source *src;
 840         struct rte_intr_callback *cb, *next;
 841         union rte_intr_read_buffer buf;
 842         struct rte_intr_callback active_cb;
 843
 844         for (n = 0; n < nfds; n++) {
 845
 846                 /**
 847                  * if the pipe fd is ready to read, return out to
 848                  * rebuild the wait list.
 849                  */
 850                 if (events[n].data.fd == intr_pipe.readfd){
 851                         int r = read(intr_pipe.readfd, buf.charbuf,
 852                                         sizeof(buf.charbuf));
 853                         RTE_SET_USED(r);
 854                         return -1;
 855                 }
 856                 rte_spinlock_lock(&intr_lock);
 857                 TAILQ_FOREACH(src, &intr_sources, next)
 858                         if (src->intr_handle.fd ==
 859                                         events[n].data.fd)
 860                                 break;
 861                 if (src == NULL){
 862                         rte_spinlock_unlock(&intr_lock);
 863                         continue;
 864                 }
 865
 866                 /* mark this interrupt source as active and release the lock. */
 867                 src->active = 1;
 868                 rte_spinlock_unlock(&intr_lock);
 869
 870                 /* set the length to be read dor different handle type */
 871                 switch (src->intr_handle.type) {
 872                 case RTE_INTR_HANDLE_UIO:
 873                 case RTE_INTR_HANDLE_UIO_INTX:
 874                         bytes_read = sizeof(buf.uio_intr_count);
 875                         break;
 876                 case RTE_INTR_HANDLE_ALARM:
 877                         bytes_read = sizeof(buf.timerfd_num);
 878                         break;
 879 #ifdef VFIO_PRESENT
 880                 case RTE_INTR_HANDLE_VFIO_MSIX:
 881                 case RTE_INTR_HANDLE_VFIO_MSI:
 882                 case RTE_INTR_HANDLE_VFIO_LEGACY:
 883                         bytes_read = sizeof(buf.vfio_intr_count);
 884                         break;
 885 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
 886                 case RTE_INTR_HANDLE_VFIO_REQ:
 887                         bytes_read = 0;
 888                         call = true;
 889                         break;
 890 #endif
 891 #endif
 892                 case RTE_INTR_HANDLE_VDEV:
 893                 case RTE_INTR_HANDLE_EXT:
 894                         bytes_read = 0;
 895                         call = true;
 896                         break;
 897                 case RTE_INTR_HANDLE_DEV_EVENT:
 898                         bytes_read = 0;
 899                         call = true;
 900                         break;
 901                 default:
 902                         bytes_read = 1;
 903                         break;
 904                 }
 905
 906                 if (bytes_read > 0) {
 907                         /**
 908                          * read out to clear the ready-to-be-read flag
 909                          * for epoll_wait.
 910                          */
 911                         bytes_read = read(events[n].data.fd, &buf, bytes_read);
 912                         if (bytes_read < 0) {
 913                                 if (errno == EINTR || errno == EWOULDBLOCK)
 914                                         continue;
 915
 916                                 RTE_LOG(ERR, EAL, "Error reading from file "
 917                                         "descriptor %d: %s\n",
 918                                         events[n].data.fd,
 919                                         strerror(errno));
 920                                 /*
 921                                  * The device is unplugged or buggy, remove
 922                                  * it as an interrupt source and return to
 923                                  * force the wait list to be rebuilt.
 924                                  */
 925                                 rte_spinlock_lock(&intr_lock);
 926                                 TAILQ_REMOVE(&intr_sources, src, next);
 927                                 rte_spinlock_unlock(&intr_lock);
 928
 929                                 for (cb = TAILQ_FIRST(&src->callbacks); cb;
 930                                                         cb = next) {
 931                                         next = TAILQ_NEXT(cb, next);
 932                                         TAILQ_REMOVE(&src->callbacks, cb, next);
 933                                         free(cb);
 934                                 }
 935                                 free(src);
 936                                 return -1;
 937                         } else if (bytes_read == 0)
 938                                 RTE_LOG(ERR, EAL, "Read nothing from file "
 939                                         "descriptor %d\n", events[n].data.fd);
 940                         else
 941                                 call = true;
 942                 }
 943
 944                 /* grab a lock, again to call callbacks and update status. */
 945                 rte_spinlock_lock(&intr_lock);
 946
 947                 if (call) {
 948
 949                         /* Finally, call all callbacks. */
 950                         TAILQ_FOREACH(cb, &src->callbacks, next) {
 951
 952                                 /* make a copy and unlock. */
 953                                 active_cb = *cb;
 954                                 rte_spinlock_unlock(&intr_lock);
 955
 956                                 /* call the actual callback */
 957                                 active_cb.cb_fn(active_cb.cb_arg);
 958
 959                                 /*get the lock back. */
 960                                 rte_spinlock_lock(&intr_lock);
 961                         }
 962                 }
 963                 /* we done with that interrupt source, release it. */
 964                 src->active = 0;
 965
 966                 rv = 0;
 967
 968                 /* check if any callback are supposed to be removed */
 969                 for (cb = TAILQ_FIRST(&src->callbacks); cb != NULL; cb = next) {
 970                         next = TAILQ_NEXT(cb, next);
 971                         if (cb->pending_delete) {
 972                                 TAILQ_REMOVE(&src->callbacks, cb, next);
 973                                 if (cb->ucb_fn)
 974                                         cb->ucb_fn(&src->intr_handle, cb->cb_arg);
 975                                 free(cb);
 976                                 rv++;
 977                         }
 978                 }
 979
 980                 /* all callbacks for that source are removed. */
 981                 if (TAILQ_EMPTY(&src->callbacks)) {
 982                         TAILQ_REMOVE(&intr_sources, src, next);
 983                         free(src);
 984                 }
 985
 986                 /* notify the pipe fd waited by epoll_wait to rebuild the wait list */
 987                 if (rv >= 0 && write(intr_pipe.writefd, "1", 1) < 0) {
 988                         rte_spinlock_unlock(&intr_lock);
 989                         return -EPIPE;
 990                 }
 991
 992                 rte_spinlock_unlock(&intr_lock);
 993         }
 994
 995         return 0;
 996 }
 997
 998 /**
 999  * It handles all the interrupts.
1000  *
1001  * @param pfd
1002  *  epoll file descriptor.
1003  * @param totalfds
1004  *  The number of file descriptors added in epoll.
1005  *
1006  * @return
1007  *  void
1008  */
1009 static void
1010 eal_intr_handle_interrupts(int pfd, unsigned totalfds)
1011 {
1012         struct epoll_event events[totalfds];
1013         int nfds = 0;
1014
1015         for(;;) {
1016                 nfds = epoll_wait(pfd, events, totalfds,
1017                         EAL_INTR_EPOLL_WAIT_FOREVER);
1018                 /* epoll_wait fail */
1019                 if (nfds < 0) {
1020                         if (errno == EINTR)
1021                                 continue;
1022                         RTE_LOG(ERR, EAL,
1023                                 "epoll_wait returns with fail\n");
1024                         return;
1025                 }
1026                 /* epoll_wait timeout, will never happens here */
1027                 else if (nfds == 0)
1028                         continue;
1029                 /* epoll_wait has at least one fd ready to read */
1030                 if (eal_intr_process_interrupts(events, nfds) < 0)
1031                         return;
1032         }
1033 }
1034
1035 /**
1036  * It builds/rebuilds up the epoll file descriptor with all the
1037  * file descriptors being waited on. Then handles the interrupts.
1038  *
1039  * @param arg
1040  *  pointer. (unused)
1041  *
1042  * @return
1043  *  never return;
1044  */
1045 static __attribute__((noreturn)) void *
1046 eal_intr_thread_main(__rte_unused void *arg)
1047 {
1048         /* host thread, never break out */
1049         for (;;) {
1050                 /* build up the epoll fd with all descriptors we are to
1051                  * wait on then pass it to the handle_interrupts function
1052                  */
1053                 static struct epoll_event pipe_event = {
1054                         .events = EPOLLIN | EPOLLPRI,
1055                 };
1056                 struct rte_intr_source *src;
1057                 unsigned numfds = 0;
1058
1059                 /* create epoll fd */
1060                 int pfd = epoll_create(1);
1061                 if (pfd < 0)
1062                         rte_panic("Cannot create epoll instance\n");
1063
1064                 pipe_event.data.fd = intr_pipe.readfd;
1065                 /**
1066                  * add pipe fd into wait list, this pipe is used to
1067                  * rebuild the wait list.
1068                  */
1069                 if (epoll_ctl(pfd, EPOLL_CTL_ADD, intr_pipe.readfd,
1070                                                 &pipe_event) < 0) {
1071                         rte_panic("Error adding fd to %d epoll_ctl, %s\n",
1072                                         intr_pipe.readfd, strerror(errno));
1073                 }
1074                 numfds++;
1075
1076                 rte_spinlock_lock(&intr_lock);
1077
1078                 TAILQ_FOREACH(src, &intr_sources, next) {
1079                         struct epoll_event ev;
1080
1081                         if (src->callbacks.tqh_first == NULL)
1082                                 continue; /* skip those with no callbacks */
1083                         memset(&ev, 0, sizeof(ev));
1084                         ev.events = EPOLLIN | EPOLLPRI | EPOLLRDHUP | EPOLLHUP;
1085                         ev.data.fd = src->intr_handle.fd;
1086
1087                         /**
1088                          * add all the uio device file descriptor
1089                          * into wait list.
1090                          */
1091                         if (epoll_ctl(pfd, EPOLL_CTL_ADD,
1092                                         src->intr_handle.fd, &ev) < 0){
1093                                 rte_panic("Error adding fd %d epoll_ctl, %s\n",
1094                                         src->intr_handle.fd, strerror(errno));
1095                         }
1096                         else
1097                                 numfds++;
1098                 }
1099                 rte_spinlock_unlock(&intr_lock);
1100                 /* serve the interrupt */
1101                 eal_intr_handle_interrupts(pfd, numfds);
1102
1103                 /**
1104                  * when we return, we need to rebuild the
1105                  * list of fds to monitor.
1106                  */
1107                 close(pfd);
1108         }
1109 }
1110
1111 int
1112 rte_eal_intr_init(void)
1113 {
1114         int ret = 0;
1115
1116         /* init the global interrupt source head */
1117         TAILQ_INIT(&intr_sources);
1118
1119         /**
1120          * create a pipe which will be waited by epoll and notified to
1121          * rebuild the wait list of epoll.
1122          */
1123         if (pipe(intr_pipe.pipefd) < 0) {
1124                 rte_errno = errno;
1125                 return -1;
1126         }
1127
1128         /* create the host thread to wait/handle the interrupt */
1129         ret = rte_ctrl_thread_create(&intr_thread, "eal-intr-thread", NULL,
1130                         eal_intr_thread_main, NULL);
1131         if (ret != 0) {
1132                 rte_errno = -ret;
1133                 RTE_LOG(ERR, EAL,
1134                         "Failed to create thread for interrupt handling\n");
1135         }
1136
1137         return ret;
1138 }
1139
1140 static void
1141 eal_intr_proc_rxtx_intr(int fd, const struct rte_intr_handle *intr_handle)
1142 {
1143         union rte_intr_read_buffer buf;
1144         int bytes_read = 0;
1145         int nbytes;
1146
1147         switch (intr_handle->type) {
1148         case RTE_INTR_HANDLE_UIO:
1149         case RTE_INTR_HANDLE_UIO_INTX:
1150                 bytes_read = sizeof(buf.uio_intr_count);
1151                 break;
1152 #ifdef VFIO_PRESENT
1153         case RTE_INTR_HANDLE_VFIO_MSIX:
1154         case RTE_INTR_HANDLE_VFIO_MSI:
1155         case RTE_INTR_HANDLE_VFIO_LEGACY:
1156                 bytes_read = sizeof(buf.vfio_intr_count);
1157                 break;
1158 #endif
1159         case RTE_INTR_HANDLE_VDEV:
1160                 bytes_read = intr_handle->efd_counter_size;
1161                 /* For vdev, number of bytes to read is set by driver */
1162                 break;
1163         case RTE_INTR_HANDLE_EXT:
1164                 return;
1165         default:
1166                 bytes_read = 1;
1167                 RTE_LOG(INFO, EAL, "unexpected intr type\n");
1168                 break;
1169         }
1170
1171         /**
1172          * read out to clear the ready-to-be-read flag
1173          * for epoll_wait.
1174          */
1175         if (bytes_read == 0)
1176                 return;
1177         do {
1178                 nbytes = read(fd, &buf, bytes_read);
1179                 if (nbytes < 0) {
1180                         if (errno == EINTR || errno == EWOULDBLOCK ||
1181                             errno == EAGAIN)
1182                                 continue;
1183                         RTE_LOG(ERR, EAL,
1184                                 "Error reading from fd %d: %s\n",
1185                                 fd, strerror(errno));
1186                 } else if (nbytes == 0)
1187                         RTE_LOG(ERR, EAL, "Read nothing from fd %d\n", fd);
1188                 return;
1189         } while (1);
1190 }
1191
1192 static int
1193 eal_epoll_process_event(struct epoll_event *evs, unsigned int n,
1194                         struct rte_epoll_event *events)
1195 {
1196         unsigned int i, count = 0;
1197         struct rte_epoll_event *rev;
1198
1199         for (i = 0; i < n; i++) {
1200                 rev = evs[i].data.ptr;
1201                 if (!rev || !rte_atomic32_cmpset(&rev->status, RTE_EPOLL_VALID,
1202                                                  RTE_EPOLL_EXEC))
1203                         continue;
1204
1205                 events[count].status        = RTE_EPOLL_VALID;
1206                 events[count].fd            = rev->fd;
1207                 events[count].epfd          = rev->epfd;
1208                 events[count].epdata.event  = rev->epdata.event;
1209                 events[count].epdata.data   = rev->epdata.data;
1210                 if (rev->epdata.cb_fun)
1211                         rev->epdata.cb_fun(rev->fd,
1212                                            rev->epdata.cb_arg);
1213
1214                 rte_compiler_barrier();
1215                 rev->status = RTE_EPOLL_VALID;
1216                 count++;
1217         }
1218         return count;
1219 }
1220
1221 static inline int
1222 eal_init_tls_epfd(void)
1223 {
1224         int pfd = epoll_create(255);
1225
1226         if (pfd < 0) {
1227                 RTE_LOG(ERR, EAL,
1228                         "Cannot create epoll instance\n");
1229                 return -1;
1230         }
1231         return pfd;
1232 }
1233
1234 int
1235 rte_intr_tls_epfd(void)
1236 {
1237         if (RTE_PER_LCORE(_epfd) == -1)
1238                 RTE_PER_LCORE(_epfd) = eal_init_tls_epfd();
1239
1240         return RTE_PER_LCORE(_epfd);
1241 }
1242
1243 int
1244 rte_epoll_wait(int epfd, struct rte_epoll_event *events,
1245                int maxevents, int timeout)
1246 {
1247         struct epoll_event evs[maxevents];
1248         int rc;
1249
1250         if (!events) {
1251                 RTE_LOG(ERR, EAL, "rte_epoll_event can't be NULL\n");
1252                 return -1;
1253         }
1254
1255         /* using per thread epoll fd */
1256         if (epfd == RTE_EPOLL_PER_THREAD)
1257                 epfd = rte_intr_tls_epfd();
1258
1259         while (1) {
1260                 rc = epoll_wait(epfd, evs, maxevents, timeout);
1261                 if (likely(rc > 0)) {
1262                         /* epoll_wait has at least one fd ready to read */
1263                         rc = eal_epoll_process_event(evs, rc, events);
1264                         break;
1265                 } else if (rc < 0) {
1266                         if (errno == EINTR)
1267                                 continue;
1268                         /* epoll_wait fail */
1269                         RTE_LOG(ERR, EAL, "epoll_wait returns with fail %s\n",
1270                                 strerror(errno));
1271                         rc = -1;
1272                         break;
1273                 } else {
1274                         /* rc == 0, epoll_wait timed out */
1275                         break;
1276                 }
1277         }
1278
1279         return rc;
1280 }
1281
1282 static inline void
1283 eal_epoll_data_safe_free(struct rte_epoll_event *ev)
1284 {
1285         while (!rte_atomic32_cmpset(&ev->status, RTE_EPOLL_VALID,
1286                                     RTE_EPOLL_INVALID))
1287                 while (ev->status != RTE_EPOLL_VALID)
1288                         rte_pause();
1289         memset(&ev->epdata, 0, sizeof(ev->epdata));
1290         ev->fd = -1;
1291         ev->epfd = -1;
1292 }
1293
1294 int
1295 rte_epoll_ctl(int epfd, int op, int fd,
1296               struct rte_epoll_event *event)
1297 {
1298         struct epoll_event ev;
1299
1300         if (!event) {
1301                 RTE_LOG(ERR, EAL, "rte_epoll_event can't be NULL\n");
1302                 return -1;
1303         }
1304
1305         /* using per thread epoll fd */
1306         if (epfd == RTE_EPOLL_PER_THREAD)
1307                 epfd = rte_intr_tls_epfd();
1308
1309         if (op == EPOLL_CTL_ADD) {
1310                 event->status = RTE_EPOLL_VALID;
1311                 event->fd = fd;  /* ignore fd in event */
1312                 event->epfd = epfd;
1313                 ev.data.ptr = (void *)event;
1314         }
1315
1316         ev.events = event->epdata.event;
1317         if (epoll_ctl(epfd, op, fd, &ev) < 0) {
1318                 RTE_LOG(ERR, EAL, "Error op %d fd %d epoll_ctl, %s\n",
1319                         op, fd, strerror(errno));
1320                 if (op == EPOLL_CTL_ADD)
1321                         /* rollback status when CTL_ADD fail */
1322                         event->status = RTE_EPOLL_INVALID;
1323                 return -1;
1324         }
1325
1326         if (op == EPOLL_CTL_DEL && event->status != RTE_EPOLL_INVALID)
1327                 eal_epoll_data_safe_free(event);
1328
1329         return 0;
1330 }
1331
1332 int
1333 rte_intr_rx_ctl(struct rte_intr_handle *intr_handle, int epfd,
1334                 int op, unsigned int vec, void *data)
1335 {
1336         struct rte_epoll_event *rev;
1337         struct rte_epoll_data *epdata;
1338         int epfd_op;
1339         unsigned int efd_idx;
1340         int rc = 0;
1341
1342         efd_idx = (vec >= RTE_INTR_VEC_RXTX_OFFSET) ?
1343                 (vec - RTE_INTR_VEC_RXTX_OFFSET) : vec;
1344
1345         if (!intr_handle || intr_handle->nb_efd == 0 ||
1346             efd_idx >= intr_handle->nb_efd) {
1347                 RTE_LOG(ERR, EAL, "Wrong intr vector number.\n");
1348                 return -EPERM;
1349         }
1350
1351         switch (op) {
1352         case RTE_INTR_EVENT_ADD:
1353                 epfd_op = EPOLL_CTL_ADD;
1354                 rev = &intr_handle->elist[efd_idx];
1355                 if (rev->status != RTE_EPOLL_INVALID) {
1356                         RTE_LOG(INFO, EAL, "Event already been added.\n");
1357                         return -EEXIST;
1358                 }
1359
1360                 /* attach to intr vector fd */
1361                 epdata = &rev->epdata;
1362                 epdata->event  = EPOLLIN | EPOLLPRI | EPOLLET;
1363                 epdata->data   = data;
1364                 epdata->cb_fun = (rte_intr_event_cb_t)eal_intr_proc_rxtx_intr;
1365                 epdata->cb_arg = (void *)intr_handle;
1366                 rc = rte_epoll_ctl(epfd, epfd_op,
1367                                    intr_handle->efds[efd_idx], rev);
1368                 if (!rc)
1369                         RTE_LOG(DEBUG, EAL,
1370                                 "efd %d associated with vec %d added on epfd %d"
1371                                 "\n", rev->fd, vec, epfd);
1372                 else
1373                         rc = -EPERM;
1374                 break;
1375         case RTE_INTR_EVENT_DEL:
1376                 epfd_op = EPOLL_CTL_DEL;
1377                 rev = &intr_handle->elist[efd_idx];
1378                 if (rev->status == RTE_EPOLL_INVALID) {
1379                         RTE_LOG(INFO, EAL, "Event does not exist.\n");
1380                         return -EPERM;
1381                 }
1382
1383                 rc = rte_epoll_ctl(rev->epfd, epfd_op, rev->fd, rev);
1384                 if (rc)
1385                         rc = -EPERM;
1386                 break;
1387         default:
1388                 RTE_LOG(ERR, EAL, "event op type mismatch\n");
1389                 rc = -EPERM;
1390         }
1391
1392         return rc;
1393 }
1394
1395 void
1396 rte_intr_free_epoll_fd(struct rte_intr_handle *intr_handle)
1397 {
1398         uint32_t i;
1399         struct rte_epoll_event *rev;
1400
1401         for (i = 0; i < intr_handle->nb_efd; i++) {
1402                 rev = &intr_handle->elist[i];
1403                 if (rev->status == RTE_EPOLL_INVALID)
1404                         continue;
1405                 if (rte_epoll_ctl(rev->epfd, EPOLL_CTL_DEL, rev->fd, rev)) {
1406                         /* force free if the entry valid */
1407                         eal_epoll_data_safe_free(rev);
1408                         rev->status = RTE_EPOLL_INVALID;
1409                 }
1410         }
1411 }
1412
1413 int
1414 rte_intr_efd_enable(struct rte_intr_handle *intr_handle, uint32_t nb_efd)
1415 {
1416         uint32_t i;
1417         int fd;
1418         uint32_t n = RTE_MIN(nb_efd, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
1419
1420         assert(nb_efd != 0);
1421
1422         if (intr_handle->type == RTE_INTR_HANDLE_VFIO_MSIX) {
1423                 for (i = 0; i < n; i++) {
1424                         fd = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC);
1425                         if (fd < 0) {
1426                                 RTE_LOG(ERR, EAL,
1427                                         "can't setup eventfd, error %i (%s)\n",
1428                                         errno, strerror(errno));
1429                                 return -errno;
1430                         }
1431                         intr_handle->efds[i] = fd;
1432                 }
1433                 intr_handle->nb_efd   = n;
1434                 intr_handle->max_intr = NB_OTHER_INTR + n;
1435         } else if (intr_handle->type == RTE_INTR_HANDLE_VDEV) {
1436                 /* only check, initialization would be done in vdev driver.*/
1437                 if (intr_handle->efd_counter_size >
1438                     sizeof(union rte_intr_read_buffer)) {
1439                         RTE_LOG(ERR, EAL, "the efd_counter_size is oversized");
1440                         return -EINVAL;
1441                 }
1442         } else {
1443                 intr_handle->efds[0]  = intr_handle->fd;
1444                 intr_handle->nb_efd   = RTE_MIN(nb_efd, 1U);
1445                 intr_handle->max_intr = NB_OTHER_INTR;
1446         }
1447
1448         return 0;
1449 }
1450
1451 void
1452 rte_intr_efd_disable(struct rte_intr_handle *intr_handle)
1453 {
1454         uint32_t i;
1455
1456         rte_intr_free_epoll_fd(intr_handle);
1457         if (intr_handle->max_intr > intr_handle->nb_efd) {
1458                 for (i = 0; i < intr_handle->nb_efd; i++)
1459                         close(intr_handle->efds[i]);
1460         }
1461         intr_handle->nb_efd = 0;
1462         intr_handle->max_intr = 0;
1463 }
1464
1465 int
1466 rte_intr_dp_is_en(struct rte_intr_handle *intr_handle)
1467 {
1468         return !(!intr_handle->nb_efd);
1469 }
1470
1471 int
1472 rte_intr_allow_others(struct rte_intr_handle *intr_handle)
1473 {
1474         if (!rte_intr_dp_is_en(intr_handle))
1475                 return 1;
1476         else
1477                 return !!(intr_handle->max_intr - intr_handle->nb_efd);
1478 }
1479
1480 int
1481 rte_intr_cap_multiple(struct rte_intr_handle *intr_handle)
1482 {
1483         if (intr_handle->type == RTE_INTR_HANDLE_VFIO_MSIX)
1484                 return 1;
1485
1486         if (intr_handle->type == RTE_INTR_HANDLE_VDEV)
1487                 return 1;
1488
1489         return 0;
1490 }