lib/librte_eal/linux/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_branch_prediction.h>
  30 #include <rte_debug.h>
  31 #include <rte_log.h>
  32 #include <rte_errno.h>
  33 #include <rte_spinlock.h>
  34 #include <rte_pause.h>
  35 #include <rte_vfio.h>
  36 #include <rte_eal_trace.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                         ret = -EPIPE;
 543
 544         rte_eal_trace_intr_callback_register(intr_handle, cb, cb_arg, ret);
 545         return ret;
 546 }
 547
 548 int
 549 rte_intr_callback_unregister_pending(const struct rte_intr_handle *intr_handle,
 550                                 rte_intr_callback_fn cb_fn, void *cb_arg,
 551                                 rte_intr_unregister_callback_fn ucb_fn)
 552 {
 553         int ret;
 554         struct rte_intr_source *src;
 555         struct rte_intr_callback *cb, *next;
 556
 557         /* do parameter checking first */
 558         if (intr_handle == NULL || intr_handle->fd < 0) {
 559                 RTE_LOG(ERR, EAL,
 560                 "Unregistering with invalid input parameter\n");
 561                 return -EINVAL;
 562         }
 563
 564         rte_spinlock_lock(&intr_lock);
 565
 566         /* check if the insterrupt source for the fd is existent */
 567         TAILQ_FOREACH(src, &intr_sources, next)
 568                 if (src->intr_handle.fd == intr_handle->fd)
 569                         break;
 570
 571         /* No interrupt source registered for the fd */
 572         if (src == NULL) {
 573                 ret = -ENOENT;
 574
 575         /* only usable if the source is active */
 576         } else if (src->active == 0) {
 577                 ret = -EAGAIN;
 578
 579         } else {
 580                 ret = 0;
 581
 582                 /* walk through the callbacks and mark all that match. */
 583                 for (cb = TAILQ_FIRST(&src->callbacks); cb != NULL; cb = next) {
 584                         next = TAILQ_NEXT(cb, next);
 585                         if (cb->cb_fn == cb_fn && (cb_arg == (void *)-1 ||
 586                                         cb->cb_arg == cb_arg)) {
 587                                 cb->pending_delete = 1;
 588                                 cb->ucb_fn = ucb_fn;
 589                                 ret++;
 590                         }
 591                 }
 592         }
 593
 594         rte_spinlock_unlock(&intr_lock);
 595
 596         return ret;
 597 }
 598
 599 int
 600 rte_intr_callback_unregister(const struct rte_intr_handle *intr_handle,
 601                         rte_intr_callback_fn cb_fn, void *cb_arg)
 602 {
 603         int ret;
 604         struct rte_intr_source *src;
 605         struct rte_intr_callback *cb, *next;
 606
 607         /* do parameter checking first */
 608         if (intr_handle == NULL || intr_handle->fd < 0) {
 609                 RTE_LOG(ERR, EAL,
 610                 "Unregistering with invalid input parameter\n");
 611                 return -EINVAL;
 612         }
 613
 614         rte_spinlock_lock(&intr_lock);
 615
 616         /* check if the insterrupt source for the fd is existent */
 617         TAILQ_FOREACH(src, &intr_sources, next)
 618                 if (src->intr_handle.fd == intr_handle->fd)
 619                         break;
 620
 621         /* No interrupt source registered for the fd */
 622         if (src == NULL) {
 623                 ret = -ENOENT;
 624
 625         /* interrupt source has some active callbacks right now. */
 626         } else if (src->active != 0) {
 627                 ret = -EAGAIN;
 628
 629         /* ok to remove. */
 630         } else {
 631                 ret = 0;
 632
 633                 /*walk through the callbacks and remove all that match. */
 634                 for (cb = TAILQ_FIRST(&src->callbacks); cb != NULL; cb = next) {
 635
 636                         next = TAILQ_NEXT(cb, next);
 637
 638                         if (cb->cb_fn == cb_fn && (cb_arg == (void *)-1 ||
 639                                         cb->cb_arg == cb_arg)) {
 640                                 TAILQ_REMOVE(&src->callbacks, cb, next);
 641                                 free(cb);
 642                                 ret++;
 643                         }
 644                 }
 645
 646                 /* all callbacks for that source are removed. */
 647                 if (TAILQ_EMPTY(&src->callbacks)) {
 648                         TAILQ_REMOVE(&intr_sources, src, next);
 649                         free(src);
 650                 }
 651         }
 652
 653         rte_spinlock_unlock(&intr_lock);
 654
 655         /* notify the pipe fd waited by epoll_wait to rebuild the wait list */
 656         if (ret >= 0 && write(intr_pipe.writefd, "1", 1) < 0) {
 657                 ret = -EPIPE;
 658         }
 659
 660         rte_eal_trace_intr_callback_unregister(intr_handle, cb_fn, cb_arg,
 661                 ret);
 662         return ret;
 663 }
 664
 665 int
 666 rte_intr_enable(const struct rte_intr_handle *intr_handle)
 667 {
 668         int rc = 0;
 669
 670         if (intr_handle == NULL)
 671                 return -1;
 672
 673         if (intr_handle->type == RTE_INTR_HANDLE_VDEV) {
 674                 rc = 0;
 675                 goto out;
 676         }
 677
 678         if (intr_handle->fd < 0 || intr_handle->uio_cfg_fd < 0) {
 679                 rc = -1;
 680                 goto out;
 681         }
 682
 683         switch (intr_handle->type){
 684         /* write to the uio fd to enable the interrupt */
 685         case RTE_INTR_HANDLE_UIO:
 686                 if (uio_intr_enable(intr_handle))
 687                         rc = -1;
 688                 break;
 689         case RTE_INTR_HANDLE_UIO_INTX:
 690                 if (uio_intx_intr_enable(intr_handle))
 691                         rc = -1;
 692                 break;
 693         /* not used at this moment */
 694         case RTE_INTR_HANDLE_ALARM:
 695                 rc = -1;
 696                 break;
 697 #ifdef VFIO_PRESENT
 698         case RTE_INTR_HANDLE_VFIO_MSIX:
 699                 if (vfio_enable_msix(intr_handle))
 700                         rc = -1;
 701                 break;
 702         case RTE_INTR_HANDLE_VFIO_MSI:
 703                 if (vfio_enable_msi(intr_handle))
 704                         rc = -1;
 705                 break;
 706         case RTE_INTR_HANDLE_VFIO_LEGACY:
 707                 if (vfio_enable_intx(intr_handle))
 708                         rc = -1;
 709                 break;
 710 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
 711         case RTE_INTR_HANDLE_VFIO_REQ:
 712                 if (vfio_enable_req(intr_handle))
 713                         rc = -1;
 714                 break;
 715 #endif
 716 #endif
 717         /* not used at this moment */
 718         case RTE_INTR_HANDLE_DEV_EVENT:
 719                 rc = -1;
 720                 break;
 721         /* unknown handle type */
 722         default:
 723                 RTE_LOG(ERR, EAL,
 724                         "Unknown handle type of fd %d\n",
 725                                         intr_handle->fd);
 726                 rc = -1;
 727                 break;
 728         }
 729 out:
 730         rte_eal_trace_intr_enable(intr_handle, rc);
 731         return rc;
 732 }
 733
 734 /**
 735  * PMD generally calls this function at the end of its IRQ callback.
 736  * Internally, it unmasks the interrupt if possible.
 737  *
 738  * For INTx, unmasking is required as the interrupt is auto-masked prior to
 739  * invoking callback.
 740  *
 741  * For MSI/MSI-X, unmasking is typically not needed as the interrupt is not
 742  * auto-masked. In fact, for interrupt handle types VFIO_MSIX and VFIO_MSI,
 743  * this function is no-op.
 744  */
 745 int
 746 rte_intr_ack(const struct rte_intr_handle *intr_handle)
 747 {
 748         if (intr_handle && intr_handle->type == RTE_INTR_HANDLE_VDEV)
 749                 return 0;
 750
 751         if (!intr_handle || intr_handle->fd < 0 || intr_handle->uio_cfg_fd < 0)
 752                 return -1;
 753
 754         switch (intr_handle->type) {
 755         /* Both acking and enabling are same for UIO */
 756         case RTE_INTR_HANDLE_UIO:
 757                 if (uio_intr_enable(intr_handle))
 758                         return -1;
 759                 break;
 760         case RTE_INTR_HANDLE_UIO_INTX:
 761                 if (uio_intx_intr_enable(intr_handle))
 762                         return -1;
 763                 break;
 764         /* not used at this moment */
 765         case RTE_INTR_HANDLE_ALARM:
 766                 return -1;
 767 #ifdef VFIO_PRESENT
 768         /* VFIO MSI* is implicitly acked unlike INTx, nothing to do */
 769         case RTE_INTR_HANDLE_VFIO_MSIX:
 770         case RTE_INTR_HANDLE_VFIO_MSI:
 771                 return 0;
 772         case RTE_INTR_HANDLE_VFIO_LEGACY:
 773                 if (vfio_ack_intx(intr_handle))
 774                         return -1;
 775                 break;
 776 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
 777         case RTE_INTR_HANDLE_VFIO_REQ:
 778                 return -1;
 779 #endif
 780 #endif
 781         /* not used at this moment */
 782         case RTE_INTR_HANDLE_DEV_EVENT:
 783                 return -1;
 784         /* unknown handle type */
 785         default:
 786                 RTE_LOG(ERR, EAL, "Unknown handle type of fd %d\n",
 787                         intr_handle->fd);
 788                 return -1;
 789         }
 790
 791         return 0;
 792 }
 793
 794 int
 795 rte_intr_disable(const struct rte_intr_handle *intr_handle)
 796 {
 797         int rc = 0;
 798
 799         if (intr_handle == NULL)
 800                 return -1;
 801
 802         if (intr_handle->type == RTE_INTR_HANDLE_VDEV) {
 803                 rc = 0;
 804                 goto out;
 805         }
 806
 807         if (intr_handle->fd < 0 || intr_handle->uio_cfg_fd < 0) {
 808                 rc = -1;
 809                 goto out;
 810         }
 811
 812         switch (intr_handle->type){
 813         /* write to the uio fd to disable the interrupt */
 814         case RTE_INTR_HANDLE_UIO:
 815                 if (uio_intr_disable(intr_handle))
 816                         rc = -1;
 817                 break;
 818         case RTE_INTR_HANDLE_UIO_INTX:
 819                 if (uio_intx_intr_disable(intr_handle))
 820                         rc = -1;
 821                 break;
 822         /* not used at this moment */
 823         case RTE_INTR_HANDLE_ALARM:
 824                 rc = -1;
 825                 break;
 826 #ifdef VFIO_PRESENT
 827         case RTE_INTR_HANDLE_VFIO_MSIX:
 828                 if (vfio_disable_msix(intr_handle))
 829                         rc = -1;
 830                 break;
 831         case RTE_INTR_HANDLE_VFIO_MSI:
 832                 if (vfio_disable_msi(intr_handle))
 833                         rc = -1;
 834                 break;
 835         case RTE_INTR_HANDLE_VFIO_LEGACY:
 836                 if (vfio_disable_intx(intr_handle))
 837                         rc = -1;
 838                 break;
 839 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
 840         case RTE_INTR_HANDLE_VFIO_REQ:
 841                 if (vfio_disable_req(intr_handle))
 842                         rc = -1;
 843                 break;
 844 #endif
 845 #endif
 846         /* not used at this moment */
 847         case RTE_INTR_HANDLE_DEV_EVENT:
 848                 rc = -1;
 849                 break;
 850         /* unknown handle type */
 851         default:
 852                 RTE_LOG(ERR, EAL,
 853                         "Unknown handle type of fd %d\n",
 854                                         intr_handle->fd);
 855                 rc = -1;
 856                 break;
 857         }
 858 out:
 859         rte_eal_trace_intr_disable(intr_handle, rc);
 860         return rc;
 861 }
 862
 863 static int
 864 eal_intr_process_interrupts(struct epoll_event *events, int nfds)
 865 {
 866         bool call = false;
 867         int n, bytes_read, rv;
 868         struct rte_intr_source *src;
 869         struct rte_intr_callback *cb, *next;
 870         union rte_intr_read_buffer buf;
 871         struct rte_intr_callback active_cb;
 872
 873         for (n = 0; n < nfds; n++) {
 874
 875                 /**
 876                  * if the pipe fd is ready to read, return out to
 877                  * rebuild the wait list.
 878                  */
 879                 if (events[n].data.fd == intr_pipe.readfd){
 880                         int r = read(intr_pipe.readfd, buf.charbuf,
 881                                         sizeof(buf.charbuf));
 882                         RTE_SET_USED(r);
 883                         return -1;
 884                 }
 885                 rte_spinlock_lock(&intr_lock);
 886                 TAILQ_FOREACH(src, &intr_sources, next)
 887                         if (src->intr_handle.fd ==
 888                                         events[n].data.fd)
 889                                 break;
 890                 if (src == NULL){
 891                         rte_spinlock_unlock(&intr_lock);
 892                         continue;
 893                 }
 894
 895                 /* mark this interrupt source as active and release the lock. */
 896                 src->active = 1;
 897                 rte_spinlock_unlock(&intr_lock);
 898
 899                 /* set the length to be read dor different handle type */
 900                 switch (src->intr_handle.type) {
 901                 case RTE_INTR_HANDLE_UIO:
 902                 case RTE_INTR_HANDLE_UIO_INTX:
 903                         bytes_read = sizeof(buf.uio_intr_count);
 904                         break;
 905                 case RTE_INTR_HANDLE_ALARM:
 906                         bytes_read = sizeof(buf.timerfd_num);
 907                         break;
 908 #ifdef VFIO_PRESENT
 909                 case RTE_INTR_HANDLE_VFIO_MSIX:
 910                 case RTE_INTR_HANDLE_VFIO_MSI:
 911                 case RTE_INTR_HANDLE_VFIO_LEGACY:
 912                         bytes_read = sizeof(buf.vfio_intr_count);
 913                         break;
 914 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
 915                 case RTE_INTR_HANDLE_VFIO_REQ:
 916                         bytes_read = 0;
 917                         call = true;
 918                         break;
 919 #endif
 920 #endif
 921                 case RTE_INTR_HANDLE_VDEV:
 922                 case RTE_INTR_HANDLE_EXT:
 923                         bytes_read = 0;
 924                         call = true;
 925                         break;
 926                 case RTE_INTR_HANDLE_DEV_EVENT:
 927                         bytes_read = 0;
 928                         call = true;
 929                         break;
 930                 default:
 931                         bytes_read = 1;
 932                         break;
 933                 }
 934
 935                 if (bytes_read > 0) {
 936                         /**
 937                          * read out to clear the ready-to-be-read flag
 938                          * for epoll_wait.
 939                          */
 940                         bytes_read = read(events[n].data.fd, &buf, bytes_read);
 941                         if (bytes_read < 0) {
 942                                 if (errno == EINTR || errno == EWOULDBLOCK)
 943                                         continue;
 944
 945                                 RTE_LOG(ERR, EAL, "Error reading from file "
 946                                         "descriptor %d: %s\n",
 947                                         events[n].data.fd,
 948                                         strerror(errno));
 949                                 /*
 950                                  * The device is unplugged or buggy, remove
 951                                  * it as an interrupt source and return to
 952                                  * force the wait list to be rebuilt.
 953                                  */
 954                                 rte_spinlock_lock(&intr_lock);
 955                                 TAILQ_REMOVE(&intr_sources, src, next);
 956                                 rte_spinlock_unlock(&intr_lock);
 957
 958                                 for (cb = TAILQ_FIRST(&src->callbacks); cb;
 959                                                         cb = next) {
 960                                         next = TAILQ_NEXT(cb, next);
 961                                         TAILQ_REMOVE(&src->callbacks, cb, next);
 962                                         free(cb);
 963                                 }
 964                                 free(src);
 965                                 return -1;
 966                         } else if (bytes_read == 0)
 967                                 RTE_LOG(ERR, EAL, "Read nothing from file "
 968                                         "descriptor %d\n", events[n].data.fd);
 969                         else
 970                                 call = true;
 971                 }
 972
 973                 /* grab a lock, again to call callbacks and update status. */
 974                 rte_spinlock_lock(&intr_lock);
 975
 976                 if (call) {
 977
 978                         /* Finally, call all callbacks. */
 979                         TAILQ_FOREACH(cb, &src->callbacks, next) {
 980
 981                                 /* make a copy and unlock. */
 982                                 active_cb = *cb;
 983                                 rte_spinlock_unlock(&intr_lock);
 984
 985                                 /* call the actual callback */
 986                                 active_cb.cb_fn(active_cb.cb_arg);
 987
 988                                 /*get the lock back. */
 989                                 rte_spinlock_lock(&intr_lock);
 990                         }
 991                 }
 992                 /* we done with that interrupt source, release it. */
 993                 src->active = 0;
 994
 995                 rv = 0;
 996
 997                 /* check if any callback are supposed to be removed */
 998                 for (cb = TAILQ_FIRST(&src->callbacks); cb != NULL; cb = next) {
 999                         next = TAILQ_NEXT(cb, next);
1000                         if (cb->pending_delete) {
1001                                 TAILQ_REMOVE(&src->callbacks, cb, next);
1002                                 if (cb->ucb_fn)
1003                                         cb->ucb_fn(&src->intr_handle, cb->cb_arg);
1004                                 free(cb);
1005                                 rv++;
1006                         }
1007                 }
1008
1009                 /* all callbacks for that source are removed. */
1010                 if (TAILQ_EMPTY(&src->callbacks)) {
1011                         TAILQ_REMOVE(&intr_sources, src, next);
1012                         free(src);
1013                 }
1014
1015                 /* notify the pipe fd waited by epoll_wait to rebuild the wait list */
1016                 if (rv > 0 && write(intr_pipe.writefd, "1", 1) < 0) {
1017                         rte_spinlock_unlock(&intr_lock);
1018                         return -EPIPE;
1019                 }
1020
1021                 rte_spinlock_unlock(&intr_lock);
1022         }
1023
1024         return 0;
1025 }
1026
1027 /**
1028  * It handles all the interrupts.
1029  *
1030  * @param pfd
1031  *  epoll file descriptor.
1032  * @param totalfds
1033  *  The number of file descriptors added in epoll.
1034  *
1035  * @return
1036  *  void
1037  */
1038 static void
1039 eal_intr_handle_interrupts(int pfd, unsigned totalfds)
1040 {
1041         struct epoll_event events[totalfds];
1042         int nfds = 0;
1043
1044         for(;;) {
1045                 nfds = epoll_wait(pfd, events, totalfds,
1046                         EAL_INTR_EPOLL_WAIT_FOREVER);
1047                 /* epoll_wait fail */
1048                 if (nfds < 0) {
1049                         if (errno == EINTR)
1050                                 continue;
1051                         RTE_LOG(ERR, EAL,
1052                                 "epoll_wait returns with fail\n");
1053                         return;
1054                 }
1055                 /* epoll_wait timeout, will never happens here */
1056                 else if (nfds == 0)
1057                         continue;
1058                 /* epoll_wait has at least one fd ready to read */
1059                 if (eal_intr_process_interrupts(events, nfds) < 0)
1060                         return;
1061         }
1062 }
1063
1064 /**
1065  * It builds/rebuilds up the epoll file descriptor with all the
1066  * file descriptors being waited on. Then handles the interrupts.
1067  *
1068  * @param arg
1069  *  pointer. (unused)
1070  *
1071  * @return
1072  *  never return;
1073  */
1074 static __rte_noreturn void *
1075 eal_intr_thread_main(__rte_unused void *arg)
1076 {
1077         /* host thread, never break out */
1078         for (;;) {
1079                 /* build up the epoll fd with all descriptors we are to
1080                  * wait on then pass it to the handle_interrupts function
1081                  */
1082                 static struct epoll_event pipe_event = {
1083                         .events = EPOLLIN | EPOLLPRI,
1084                 };
1085                 struct rte_intr_source *src;
1086                 unsigned numfds = 0;
1087
1088                 /* create epoll fd */
1089                 int pfd = epoll_create(1);
1090                 if (pfd < 0)
1091                         rte_panic("Cannot create epoll instance\n");
1092
1093                 pipe_event.data.fd = intr_pipe.readfd;
1094                 /**
1095                  * add pipe fd into wait list, this pipe is used to
1096                  * rebuild the wait list.
1097                  */
1098                 if (epoll_ctl(pfd, EPOLL_CTL_ADD, intr_pipe.readfd,
1099                                                 &pipe_event) < 0) {
1100                         rte_panic("Error adding fd to %d epoll_ctl, %s\n",
1101                                         intr_pipe.readfd, strerror(errno));
1102                 }
1103                 numfds++;
1104
1105                 rte_spinlock_lock(&intr_lock);
1106
1107                 TAILQ_FOREACH(src, &intr_sources, next) {
1108                         struct epoll_event ev;
1109
1110                         if (src->callbacks.tqh_first == NULL)
1111                                 continue; /* skip those with no callbacks */
1112                         memset(&ev, 0, sizeof(ev));
1113                         ev.events = EPOLLIN | EPOLLPRI | EPOLLRDHUP | EPOLLHUP;
1114                         ev.data.fd = src->intr_handle.fd;
1115
1116                         /**
1117                          * add all the uio device file descriptor
1118                          * into wait list.
1119                          */
1120                         if (epoll_ctl(pfd, EPOLL_CTL_ADD,
1121                                         src->intr_handle.fd, &ev) < 0){
1122                                 rte_panic("Error adding fd %d epoll_ctl, %s\n",
1123                                         src->intr_handle.fd, strerror(errno));
1124                         }
1125                         else
1126                                 numfds++;
1127                 }
1128                 rte_spinlock_unlock(&intr_lock);
1129                 /* serve the interrupt */
1130                 eal_intr_handle_interrupts(pfd, numfds);
1131
1132                 /**
1133                  * when we return, we need to rebuild the
1134                  * list of fds to monitor.
1135                  */
1136                 close(pfd);
1137         }
1138 }
1139
1140 int
1141 rte_eal_intr_init(void)
1142 {
1143         int ret = 0;
1144
1145         /* init the global interrupt source head */
1146         TAILQ_INIT(&intr_sources);
1147
1148         /**
1149          * create a pipe which will be waited by epoll and notified to
1150          * rebuild the wait list of epoll.
1151          */
1152         if (pipe(intr_pipe.pipefd) < 0) {
1153                 rte_errno = errno;
1154                 return -1;
1155         }
1156
1157         /* create the host thread to wait/handle the interrupt */
1158         ret = rte_ctrl_thread_create(&intr_thread, "eal-intr-thread", NULL,
1159                         eal_intr_thread_main, NULL);
1160         if (ret != 0) {
1161                 rte_errno = -ret;
1162                 RTE_LOG(ERR, EAL,
1163                         "Failed to create thread for interrupt handling\n");
1164         }
1165
1166         return ret;
1167 }
1168
1169 static void
1170 eal_intr_proc_rxtx_intr(int fd, const struct rte_intr_handle *intr_handle)
1171 {
1172         union rte_intr_read_buffer buf;
1173         int bytes_read = 0;
1174         int nbytes;
1175
1176         switch (intr_handle->type) {
1177         case RTE_INTR_HANDLE_UIO:
1178         case RTE_INTR_HANDLE_UIO_INTX:
1179                 bytes_read = sizeof(buf.uio_intr_count);
1180                 break;
1181 #ifdef VFIO_PRESENT
1182         case RTE_INTR_HANDLE_VFIO_MSIX:
1183         case RTE_INTR_HANDLE_VFIO_MSI:
1184         case RTE_INTR_HANDLE_VFIO_LEGACY:
1185                 bytes_read = sizeof(buf.vfio_intr_count);
1186                 break;
1187 #endif
1188         case RTE_INTR_HANDLE_VDEV:
1189                 bytes_read = intr_handle->efd_counter_size;
1190                 /* For vdev, number of bytes to read is set by driver */
1191                 break;
1192         case RTE_INTR_HANDLE_EXT:
1193                 return;
1194         default:
1195                 bytes_read = 1;
1196                 RTE_LOG(INFO, EAL, "unexpected intr type\n");
1197                 break;
1198         }
1199
1200         /**
1201          * read out to clear the ready-to-be-read flag
1202          * for epoll_wait.
1203          */
1204         if (bytes_read == 0)
1205                 return;
1206         do {
1207                 nbytes = read(fd, &buf, bytes_read);
1208                 if (nbytes < 0) {
1209                         if (errno == EINTR || errno == EWOULDBLOCK ||
1210                             errno == EAGAIN)
1211                                 continue;
1212                         RTE_LOG(ERR, EAL,
1213                                 "Error reading from fd %d: %s\n",
1214                                 fd, strerror(errno));
1215                 } else if (nbytes == 0)
1216                         RTE_LOG(ERR, EAL, "Read nothing from fd %d\n", fd);
1217                 return;
1218         } while (1);
1219 }
1220
1221 static int
1222 eal_epoll_process_event(struct epoll_event *evs, unsigned int n,
1223                         struct rte_epoll_event *events)
1224 {
1225         unsigned int i, count = 0;
1226         struct rte_epoll_event *rev;
1227         uint32_t valid_status;
1228
1229         for (i = 0; i < n; i++) {
1230                 rev = evs[i].data.ptr;
1231                 valid_status =  RTE_EPOLL_VALID;
1232                 /* ACQUIRE memory ordering here pairs with RELEASE
1233                  * ordering below acting as a lock to synchronize
1234                  * the event data updating.
1235                  */
1236                 if (!rev || !__atomic_compare_exchange_n(&rev->status,
1237                                     &valid_status, RTE_EPOLL_EXEC, 0,
1238                                     __ATOMIC_ACQUIRE, __ATOMIC_RELAXED))
1239                         continue;
1240
1241                 events[count].status        = RTE_EPOLL_VALID;
1242                 events[count].fd            = rev->fd;
1243                 events[count].epfd          = rev->epfd;
1244                 events[count].epdata.event  = evs[i].events;
1245                 events[count].epdata.data   = rev->epdata.data;
1246                 if (rev->epdata.cb_fun)
1247                         rev->epdata.cb_fun(rev->fd,
1248                                            rev->epdata.cb_arg);
1249
1250                 /* the status update should be observed after
1251                  * the other fields change.
1252                  */
1253                 __atomic_store_n(&rev->status, RTE_EPOLL_VALID,
1254                                 __ATOMIC_RELEASE);
1255                 count++;
1256         }
1257         return count;
1258 }
1259
1260 static inline int
1261 eal_init_tls_epfd(void)
1262 {
1263         int pfd = epoll_create(255);
1264
1265         if (pfd < 0) {
1266                 RTE_LOG(ERR, EAL,
1267                         "Cannot create epoll instance\n");
1268                 return -1;
1269         }
1270         return pfd;
1271 }
1272
1273 int
1274 rte_intr_tls_epfd(void)
1275 {
1276         if (RTE_PER_LCORE(_epfd) == -1)
1277                 RTE_PER_LCORE(_epfd) = eal_init_tls_epfd();
1278
1279         return RTE_PER_LCORE(_epfd);
1280 }
1281
1282 static int
1283 eal_epoll_wait(int epfd, struct rte_epoll_event *events,
1284                int maxevents, int timeout, bool interruptible)
1285 {
1286         struct epoll_event evs[maxevents];
1287         int rc;
1288
1289         if (!events) {
1290                 RTE_LOG(ERR, EAL, "rte_epoll_event can't be NULL\n");
1291                 return -1;
1292         }
1293
1294         /* using per thread epoll fd */
1295         if (epfd == RTE_EPOLL_PER_THREAD)
1296                 epfd = rte_intr_tls_epfd();
1297
1298         while (1) {
1299                 rc = epoll_wait(epfd, evs, maxevents, timeout);
1300                 if (likely(rc > 0)) {
1301                         /* epoll_wait has at least one fd ready to read */
1302                         rc = eal_epoll_process_event(evs, rc, events);
1303                         break;
1304                 } else if (rc < 0) {
1305                         if (errno == EINTR) {
1306                                 if (interruptible)
1307                                         return -1;
1308                                 else
1309                                         continue;
1310                         }
1311                         /* epoll_wait fail */
1312                         RTE_LOG(ERR, EAL, "epoll_wait returns with fail %s\n",
1313                                 strerror(errno));
1314                         rc = -1;
1315                         break;
1316                 } else {
1317                         /* rc == 0, epoll_wait timed out */
1318                         break;
1319                 }
1320         }
1321
1322         return rc;
1323 }
1324
1325 int
1326 rte_epoll_wait(int epfd, struct rte_epoll_event *events,
1327                int maxevents, int timeout)
1328 {
1329         return eal_epoll_wait(epfd, events, maxevents, timeout, false);
1330 }
1331
1332 int
1333 rte_epoll_wait_interruptible(int epfd, struct rte_epoll_event *events,
1334                              int maxevents, int timeout)
1335 {
1336         return eal_epoll_wait(epfd, events, maxevents, timeout, true);
1337 }
1338
1339 static inline void
1340 eal_epoll_data_safe_free(struct rte_epoll_event *ev)
1341 {
1342         uint32_t valid_status = RTE_EPOLL_VALID;
1343
1344         while (!__atomic_compare_exchange_n(&ev->status, &valid_status,
1345                     RTE_EPOLL_INVALID, 0, __ATOMIC_ACQUIRE, __ATOMIC_RELAXED)) {
1346                 while (__atomic_load_n(&ev->status,
1347                                 __ATOMIC_RELAXED) != RTE_EPOLL_VALID)
1348                         rte_pause();
1349                 valid_status = RTE_EPOLL_VALID;
1350         }
1351         memset(&ev->epdata, 0, sizeof(ev->epdata));
1352         ev->fd = -1;
1353         ev->epfd = -1;
1354 }
1355
1356 int
1357 rte_epoll_ctl(int epfd, int op, int fd,
1358               struct rte_epoll_event *event)
1359 {
1360         struct epoll_event ev;
1361
1362         if (!event) {
1363                 RTE_LOG(ERR, EAL, "rte_epoll_event can't be NULL\n");
1364                 return -1;
1365         }
1366
1367         /* using per thread epoll fd */
1368         if (epfd == RTE_EPOLL_PER_THREAD)
1369                 epfd = rte_intr_tls_epfd();
1370
1371         if (op == EPOLL_CTL_ADD) {
1372                 __atomic_store_n(&event->status, RTE_EPOLL_VALID,
1373                                 __ATOMIC_RELAXED);
1374                 event->fd = fd;  /* ignore fd in event */
1375                 event->epfd = epfd;
1376                 ev.data.ptr = (void *)event;
1377         }
1378
1379         ev.events = event->epdata.event;
1380         if (epoll_ctl(epfd, op, fd, &ev) < 0) {
1381                 RTE_LOG(ERR, EAL, "Error op %d fd %d epoll_ctl, %s\n",
1382                         op, fd, strerror(errno));
1383                 if (op == EPOLL_CTL_ADD)
1384                         /* rollback status when CTL_ADD fail */
1385                         __atomic_store_n(&event->status, RTE_EPOLL_INVALID,
1386                                         __ATOMIC_RELAXED);
1387                 return -1;
1388         }
1389
1390         if (op == EPOLL_CTL_DEL && __atomic_load_n(&event->status,
1391                         __ATOMIC_RELAXED) != RTE_EPOLL_INVALID)
1392                 eal_epoll_data_safe_free(event);
1393
1394         return 0;
1395 }
1396
1397 int
1398 rte_intr_rx_ctl(struct rte_intr_handle *intr_handle, int epfd,
1399                 int op, unsigned int vec, void *data)
1400 {
1401         struct rte_epoll_event *rev;
1402         struct rte_epoll_data *epdata;
1403         int epfd_op;
1404         unsigned int efd_idx;
1405         int rc = 0;
1406
1407         efd_idx = (vec >= RTE_INTR_VEC_RXTX_OFFSET) ?
1408                 (vec - RTE_INTR_VEC_RXTX_OFFSET) : vec;
1409
1410         if (!intr_handle || intr_handle->nb_efd == 0 ||
1411             efd_idx >= intr_handle->nb_efd) {
1412                 RTE_LOG(ERR, EAL, "Wrong intr vector number.\n");
1413                 return -EPERM;
1414         }
1415
1416         switch (op) {
1417         case RTE_INTR_EVENT_ADD:
1418                 epfd_op = EPOLL_CTL_ADD;
1419                 rev = &intr_handle->elist[efd_idx];
1420                 if (__atomic_load_n(&rev->status,
1421                                 __ATOMIC_RELAXED) != RTE_EPOLL_INVALID) {
1422                         RTE_LOG(INFO, EAL, "Event already been added.\n");
1423                         return -EEXIST;
1424                 }
1425
1426                 /* attach to intr vector fd */
1427                 epdata = &rev->epdata;
1428                 epdata->event  = EPOLLIN | EPOLLPRI | EPOLLET;
1429                 epdata->data   = data;
1430                 epdata->cb_fun = (rte_intr_event_cb_t)eal_intr_proc_rxtx_intr;
1431                 epdata->cb_arg = (void *)intr_handle;
1432                 rc = rte_epoll_ctl(epfd, epfd_op,
1433                                    intr_handle->efds[efd_idx], rev);
1434                 if (!rc)
1435                         RTE_LOG(DEBUG, EAL,
1436                                 "efd %d associated with vec %d added on epfd %d"
1437                                 "\n", rev->fd, vec, epfd);
1438                 else
1439                         rc = -EPERM;
1440                 break;
1441         case RTE_INTR_EVENT_DEL:
1442                 epfd_op = EPOLL_CTL_DEL;
1443                 rev = &intr_handle->elist[efd_idx];
1444                 if (__atomic_load_n(&rev->status,
1445                                 __ATOMIC_RELAXED) == RTE_EPOLL_INVALID) {
1446                         RTE_LOG(INFO, EAL, "Event does not exist.\n");
1447                         return -EPERM;
1448                 }
1449
1450                 rc = rte_epoll_ctl(rev->epfd, epfd_op, rev->fd, rev);
1451                 if (rc)
1452                         rc = -EPERM;
1453                 break;
1454         default:
1455                 RTE_LOG(ERR, EAL, "event op type mismatch\n");
1456                 rc = -EPERM;
1457         }
1458
1459         return rc;
1460 }
1461
1462 void
1463 rte_intr_free_epoll_fd(struct rte_intr_handle *intr_handle)
1464 {
1465         uint32_t i;
1466         struct rte_epoll_event *rev;
1467
1468         for (i = 0; i < intr_handle->nb_efd; i++) {
1469                 rev = &intr_handle->elist[i];
1470                 if (__atomic_load_n(&rev->status,
1471                                 __ATOMIC_RELAXED) == RTE_EPOLL_INVALID)
1472                         continue;
1473                 if (rte_epoll_ctl(rev->epfd, EPOLL_CTL_DEL, rev->fd, rev)) {
1474                         /* force free if the entry valid */
1475                         eal_epoll_data_safe_free(rev);
1476                 }
1477         }
1478 }
1479
1480 int
1481 rte_intr_efd_enable(struct rte_intr_handle *intr_handle, uint32_t nb_efd)
1482 {
1483         uint32_t i;
1484         int fd;
1485         uint32_t n = RTE_MIN(nb_efd, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
1486
1487         assert(nb_efd != 0);
1488
1489         if (intr_handle->type == RTE_INTR_HANDLE_VFIO_MSIX) {
1490                 for (i = 0; i < n; i++) {
1491                         fd = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC);
1492                         if (fd < 0) {
1493                                 RTE_LOG(ERR, EAL,
1494                                         "can't setup eventfd, error %i (%s)\n",
1495                                         errno, strerror(errno));
1496                                 return -errno;
1497                         }
1498                         intr_handle->efds[i] = fd;
1499                 }
1500                 intr_handle->nb_efd   = n;
1501                 intr_handle->max_intr = NB_OTHER_INTR + n;
1502         } else if (intr_handle->type == RTE_INTR_HANDLE_VDEV) {
1503                 /* only check, initialization would be done in vdev driver.*/
1504                 if (intr_handle->efd_counter_size >
1505                     sizeof(union rte_intr_read_buffer)) {
1506                         RTE_LOG(ERR, EAL, "the efd_counter_size is oversized");
1507                         return -EINVAL;
1508                 }
1509         } else {
1510                 intr_handle->efds[0]  = intr_handle->fd;
1511                 intr_handle->nb_efd   = RTE_MIN(nb_efd, 1U);
1512                 intr_handle->max_intr = NB_OTHER_INTR;
1513         }
1514
1515         return 0;
1516 }
1517
1518 void
1519 rte_intr_efd_disable(struct rte_intr_handle *intr_handle)
1520 {
1521         uint32_t i;
1522
1523         rte_intr_free_epoll_fd(intr_handle);
1524         if (intr_handle->max_intr > intr_handle->nb_efd) {
1525                 for (i = 0; i < intr_handle->nb_efd; i++)
1526                         close(intr_handle->efds[i]);
1527         }
1528         intr_handle->nb_efd = 0;
1529         intr_handle->max_intr = 0;
1530 }
1531
1532 int
1533 rte_intr_dp_is_en(struct rte_intr_handle *intr_handle)
1534 {
1535         return !(!intr_handle->nb_efd);
1536 }
1537
1538 int
1539 rte_intr_allow_others(struct rte_intr_handle *intr_handle)
1540 {
1541         if (!rte_intr_dp_is_en(intr_handle))
1542                 return 1;
1543         else
1544                 return !!(intr_handle->max_intr - intr_handle->nb_efd);
1545 }
1546
1547 int
1548 rte_intr_cap_multiple(struct rte_intr_handle *intr_handle)
1549 {
1550         if (intr_handle->type == RTE_INTR_HANDLE_VFIO_MSIX)
1551                 return 1;
1552
1553         if (intr_handle->type == RTE_INTR_HANDLE_VDEV)
1554                 return 1;
1555
1556         return 0;
1557 }
1558
1559 int rte_thread_is_intr(void)
1560 {
1561         return pthread_equal(intr_thread, pthread_self());
1562 }