lib/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_callback_unregister_sync(const struct rte_intr_handle *intr_handle,
 667                         rte_intr_callback_fn cb_fn, void *cb_arg)
 668 {
 669         int ret = 0;
 670
 671         while ((ret = rte_intr_callback_unregister(intr_handle, cb_fn, cb_arg)) == -EAGAIN)
 672                 rte_pause();
 673
 674         return ret;
 675 }
 676
 677 int
 678 rte_intr_enable(const struct rte_intr_handle *intr_handle)
 679 {
 680         int rc = 0;
 681
 682         if (intr_handle == NULL)
 683                 return -1;
 684
 685         if (intr_handle->type == RTE_INTR_HANDLE_VDEV) {
 686                 rc = 0;
 687                 goto out;
 688         }
 689
 690         if (intr_handle->fd < 0 || intr_handle->uio_cfg_fd < 0) {
 691                 rc = -1;
 692                 goto out;
 693         }
 694
 695         switch (intr_handle->type){
 696         /* write to the uio fd to enable the interrupt */
 697         case RTE_INTR_HANDLE_UIO:
 698                 if (uio_intr_enable(intr_handle))
 699                         rc = -1;
 700                 break;
 701         case RTE_INTR_HANDLE_UIO_INTX:
 702                 if (uio_intx_intr_enable(intr_handle))
 703                         rc = -1;
 704                 break;
 705         /* not used at this moment */
 706         case RTE_INTR_HANDLE_ALARM:
 707                 rc = -1;
 708                 break;
 709 #ifdef VFIO_PRESENT
 710         case RTE_INTR_HANDLE_VFIO_MSIX:
 711                 if (vfio_enable_msix(intr_handle))
 712                         rc = -1;
 713                 break;
 714         case RTE_INTR_HANDLE_VFIO_MSI:
 715                 if (vfio_enable_msi(intr_handle))
 716                         rc = -1;
 717                 break;
 718         case RTE_INTR_HANDLE_VFIO_LEGACY:
 719                 if (vfio_enable_intx(intr_handle))
 720                         rc = -1;
 721                 break;
 722 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
 723         case RTE_INTR_HANDLE_VFIO_REQ:
 724                 if (vfio_enable_req(intr_handle))
 725                         rc = -1;
 726                 break;
 727 #endif
 728 #endif
 729         /* not used at this moment */
 730         case RTE_INTR_HANDLE_DEV_EVENT:
 731                 rc = -1;
 732                 break;
 733         /* unknown handle type */
 734         default:
 735                 RTE_LOG(ERR, EAL,
 736                         "Unknown handle type of fd %d\n",
 737                                         intr_handle->fd);
 738                 rc = -1;
 739                 break;
 740         }
 741 out:
 742         rte_eal_trace_intr_enable(intr_handle, rc);
 743         return rc;
 744 }
 745
 746 /**
 747  * PMD generally calls this function at the end of its IRQ callback.
 748  * Internally, it unmasks the interrupt if possible.
 749  *
 750  * For INTx, unmasking is required as the interrupt is auto-masked prior to
 751  * invoking callback.
 752  *
 753  * For MSI/MSI-X, unmasking is typically not needed as the interrupt is not
 754  * auto-masked. In fact, for interrupt handle types VFIO_MSIX and VFIO_MSI,
 755  * this function is no-op.
 756  */
 757 int
 758 rte_intr_ack(const struct rte_intr_handle *intr_handle)
 759 {
 760         if (intr_handle && intr_handle->type == RTE_INTR_HANDLE_VDEV)
 761                 return 0;
 762
 763         if (!intr_handle || intr_handle->fd < 0 || intr_handle->uio_cfg_fd < 0)
 764                 return -1;
 765
 766         switch (intr_handle->type) {
 767         /* Both acking and enabling are same for UIO */
 768         case RTE_INTR_HANDLE_UIO:
 769                 if (uio_intr_enable(intr_handle))
 770                         return -1;
 771                 break;
 772         case RTE_INTR_HANDLE_UIO_INTX:
 773                 if (uio_intx_intr_enable(intr_handle))
 774                         return -1;
 775                 break;
 776         /* not used at this moment */
 777         case RTE_INTR_HANDLE_ALARM:
 778                 return -1;
 779 #ifdef VFIO_PRESENT
 780         /* VFIO MSI* is implicitly acked unlike INTx, nothing to do */
 781         case RTE_INTR_HANDLE_VFIO_MSIX:
 782         case RTE_INTR_HANDLE_VFIO_MSI:
 783                 return 0;
 784         case RTE_INTR_HANDLE_VFIO_LEGACY:
 785                 if (vfio_ack_intx(intr_handle))
 786                         return -1;
 787                 break;
 788 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
 789         case RTE_INTR_HANDLE_VFIO_REQ:
 790                 return -1;
 791 #endif
 792 #endif
 793         /* not used at this moment */
 794         case RTE_INTR_HANDLE_DEV_EVENT:
 795                 return -1;
 796         /* unknown handle type */
 797         default:
 798                 RTE_LOG(ERR, EAL, "Unknown handle type of fd %d\n",
 799                         intr_handle->fd);
 800                 return -1;
 801         }
 802
 803         return 0;
 804 }
 805
 806 int
 807 rte_intr_disable(const struct rte_intr_handle *intr_handle)
 808 {
 809         int rc = 0;
 810
 811         if (intr_handle == NULL)
 812                 return -1;
 813
 814         if (intr_handle->type == RTE_INTR_HANDLE_VDEV) {
 815                 rc = 0;
 816                 goto out;
 817         }
 818
 819         if (intr_handle->fd < 0 || intr_handle->uio_cfg_fd < 0) {
 820                 rc = -1;
 821                 goto out;
 822         }
 823
 824         switch (intr_handle->type){
 825         /* write to the uio fd to disable the interrupt */
 826         case RTE_INTR_HANDLE_UIO:
 827                 if (uio_intr_disable(intr_handle))
 828                         rc = -1;
 829                 break;
 830         case RTE_INTR_HANDLE_UIO_INTX:
 831                 if (uio_intx_intr_disable(intr_handle))
 832                         rc = -1;
 833                 break;
 834         /* not used at this moment */
 835         case RTE_INTR_HANDLE_ALARM:
 836                 rc = -1;
 837                 break;
 838 #ifdef VFIO_PRESENT
 839         case RTE_INTR_HANDLE_VFIO_MSIX:
 840                 if (vfio_disable_msix(intr_handle))
 841                         rc = -1;
 842                 break;
 843         case RTE_INTR_HANDLE_VFIO_MSI:
 844                 if (vfio_disable_msi(intr_handle))
 845                         rc = -1;
 846                 break;
 847         case RTE_INTR_HANDLE_VFIO_LEGACY:
 848                 if (vfio_disable_intx(intr_handle))
 849                         rc = -1;
 850                 break;
 851 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
 852         case RTE_INTR_HANDLE_VFIO_REQ:
 853                 if (vfio_disable_req(intr_handle))
 854                         rc = -1;
 855                 break;
 856 #endif
 857 #endif
 858         /* not used at this moment */
 859         case RTE_INTR_HANDLE_DEV_EVENT:
 860                 rc = -1;
 861                 break;
 862         /* unknown handle type */
 863         default:
 864                 RTE_LOG(ERR, EAL,
 865                         "Unknown handle type of fd %d\n",
 866                                         intr_handle->fd);
 867                 rc = -1;
 868                 break;
 869         }
 870 out:
 871         rte_eal_trace_intr_disable(intr_handle, rc);
 872         return rc;
 873 }
 874
 875 static int
 876 eal_intr_process_interrupts(struct epoll_event *events, int nfds)
 877 {
 878         bool call = false;
 879         int n, bytes_read, rv;
 880         struct rte_intr_source *src;
 881         struct rte_intr_callback *cb, *next;
 882         union rte_intr_read_buffer buf;
 883         struct rte_intr_callback active_cb;
 884
 885         for (n = 0; n < nfds; n++) {
 886
 887                 /**
 888                  * if the pipe fd is ready to read, return out to
 889                  * rebuild the wait list.
 890                  */
 891                 if (events[n].data.fd == intr_pipe.readfd){
 892                         int r = read(intr_pipe.readfd, buf.charbuf,
 893                                         sizeof(buf.charbuf));
 894                         RTE_SET_USED(r);
 895                         return -1;
 896                 }
 897                 rte_spinlock_lock(&intr_lock);
 898                 TAILQ_FOREACH(src, &intr_sources, next)
 899                         if (src->intr_handle.fd ==
 900                                         events[n].data.fd)
 901                                 break;
 902                 if (src == NULL){
 903                         rte_spinlock_unlock(&intr_lock);
 904                         continue;
 905                 }
 906
 907                 /* mark this interrupt source as active and release the lock. */
 908                 src->active = 1;
 909                 rte_spinlock_unlock(&intr_lock);
 910
 911                 /* set the length to be read dor different handle type */
 912                 switch (src->intr_handle.type) {
 913                 case RTE_INTR_HANDLE_UIO:
 914                 case RTE_INTR_HANDLE_UIO_INTX:
 915                         bytes_read = sizeof(buf.uio_intr_count);
 916                         break;
 917                 case RTE_INTR_HANDLE_ALARM:
 918                         bytes_read = sizeof(buf.timerfd_num);
 919                         break;
 920 #ifdef VFIO_PRESENT
 921                 case RTE_INTR_HANDLE_VFIO_MSIX:
 922                 case RTE_INTR_HANDLE_VFIO_MSI:
 923                 case RTE_INTR_HANDLE_VFIO_LEGACY:
 924                         bytes_read = sizeof(buf.vfio_intr_count);
 925                         break;
 926 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
 927                 case RTE_INTR_HANDLE_VFIO_REQ:
 928                         bytes_read = 0;
 929                         call = true;
 930                         break;
 931 #endif
 932 #endif
 933                 case RTE_INTR_HANDLE_VDEV:
 934                 case RTE_INTR_HANDLE_EXT:
 935                         bytes_read = 0;
 936                         call = true;
 937                         break;
 938                 case RTE_INTR_HANDLE_DEV_EVENT:
 939                         bytes_read = 0;
 940                         call = true;
 941                         break;
 942                 default:
 943                         bytes_read = 1;
 944                         break;
 945                 }
 946
 947                 if (bytes_read > 0) {
 948                         /**
 949                          * read out to clear the ready-to-be-read flag
 950                          * for epoll_wait.
 951                          */
 952                         bytes_read = read(events[n].data.fd, &buf, bytes_read);
 953                         if (bytes_read < 0) {
 954                                 if (errno == EINTR || errno == EWOULDBLOCK)
 955                                         continue;
 956
 957                                 RTE_LOG(ERR, EAL, "Error reading from file "
 958                                         "descriptor %d: %s\n",
 959                                         events[n].data.fd,
 960                                         strerror(errno));
 961                                 /*
 962                                  * The device is unplugged or buggy, remove
 963                                  * it as an interrupt source and return to
 964                                  * force the wait list to be rebuilt.
 965                                  */
 966                                 rte_spinlock_lock(&intr_lock);
 967                                 TAILQ_REMOVE(&intr_sources, src, next);
 968                                 rte_spinlock_unlock(&intr_lock);
 969
 970                                 for (cb = TAILQ_FIRST(&src->callbacks); cb;
 971                                                         cb = next) {
 972                                         next = TAILQ_NEXT(cb, next);
 973                                         TAILQ_REMOVE(&src->callbacks, cb, next);
 974                                         free(cb);
 975                                 }
 976                                 free(src);
 977                                 return -1;
 978                         } else if (bytes_read == 0)
 979                                 RTE_LOG(ERR, EAL, "Read nothing from file "
 980                                         "descriptor %d\n", events[n].data.fd);
 981                         else
 982                                 call = true;
 983                 }
 984
 985                 /* grab a lock, again to call callbacks and update status. */
 986                 rte_spinlock_lock(&intr_lock);
 987
 988                 if (call) {
 989
 990                         /* Finally, call all callbacks. */
 991                         TAILQ_FOREACH(cb, &src->callbacks, next) {
 992
 993                                 /* make a copy and unlock. */
 994                                 active_cb = *cb;
 995                                 rte_spinlock_unlock(&intr_lock);
 996
 997                                 /* call the actual callback */
 998                                 active_cb.cb_fn(active_cb.cb_arg);
 999
1000                                 /*get the lock back. */
1001                                 rte_spinlock_lock(&intr_lock);
1002                         }
1003                 }
1004                 /* we done with that interrupt source, release it. */
1005                 src->active = 0;
1006
1007                 rv = 0;
1008
1009                 /* check if any callback are supposed to be removed */
1010                 for (cb = TAILQ_FIRST(&src->callbacks); cb != NULL; cb = next) {
1011                         next = TAILQ_NEXT(cb, next);
1012                         if (cb->pending_delete) {
1013                                 TAILQ_REMOVE(&src->callbacks, cb, next);
1014                                 if (cb->ucb_fn)
1015                                         cb->ucb_fn(&src->intr_handle, cb->cb_arg);
1016                                 free(cb);
1017                                 rv++;
1018                         }
1019                 }
1020
1021                 /* all callbacks for that source are removed. */
1022                 if (TAILQ_EMPTY(&src->callbacks)) {
1023                         TAILQ_REMOVE(&intr_sources, src, next);
1024                         free(src);
1025                 }
1026
1027                 /* notify the pipe fd waited by epoll_wait to rebuild the wait list */
1028                 if (rv > 0 && write(intr_pipe.writefd, "1", 1) < 0) {
1029                         rte_spinlock_unlock(&intr_lock);
1030                         return -EPIPE;
1031                 }
1032
1033                 rte_spinlock_unlock(&intr_lock);
1034         }
1035
1036         return 0;
1037 }
1038
1039 /**
1040  * It handles all the interrupts.
1041  *
1042  * @param pfd
1043  *  epoll file descriptor.
1044  * @param totalfds
1045  *  The number of file descriptors added in epoll.
1046  *
1047  * @return
1048  *  void
1049  */
1050 static void
1051 eal_intr_handle_interrupts(int pfd, unsigned totalfds)
1052 {
1053         struct epoll_event events[totalfds];
1054         int nfds = 0;
1055
1056         for(;;) {
1057                 nfds = epoll_wait(pfd, events, totalfds,
1058                         EAL_INTR_EPOLL_WAIT_FOREVER);
1059                 /* epoll_wait fail */
1060                 if (nfds < 0) {
1061                         if (errno == EINTR)
1062                                 continue;
1063                         RTE_LOG(ERR, EAL,
1064                                 "epoll_wait returns with fail\n");
1065                         return;
1066                 }
1067                 /* epoll_wait timeout, will never happens here */
1068                 else if (nfds == 0)
1069                         continue;
1070                 /* epoll_wait has at least one fd ready to read */
1071                 if (eal_intr_process_interrupts(events, nfds) < 0)
1072                         return;
1073         }
1074 }
1075
1076 /**
1077  * It builds/rebuilds up the epoll file descriptor with all the
1078  * file descriptors being waited on. Then handles the interrupts.
1079  *
1080  * @param arg
1081  *  pointer. (unused)
1082  *
1083  * @return
1084  *  never return;
1085  */
1086 static __rte_noreturn void *
1087 eal_intr_thread_main(__rte_unused void *arg)
1088 {
1089         /* host thread, never break out */
1090         for (;;) {
1091                 /* build up the epoll fd with all descriptors we are to
1092                  * wait on then pass it to the handle_interrupts function
1093                  */
1094                 static struct epoll_event pipe_event = {
1095                         .events = EPOLLIN | EPOLLPRI,
1096                 };
1097                 struct rte_intr_source *src;
1098                 unsigned numfds = 0;
1099
1100                 /* create epoll fd */
1101                 int pfd = epoll_create(1);
1102                 if (pfd < 0)
1103                         rte_panic("Cannot create epoll instance\n");
1104
1105                 pipe_event.data.fd = intr_pipe.readfd;
1106                 /**
1107                  * add pipe fd into wait list, this pipe is used to
1108                  * rebuild the wait list.
1109                  */
1110                 if (epoll_ctl(pfd, EPOLL_CTL_ADD, intr_pipe.readfd,
1111                                                 &pipe_event) < 0) {
1112                         rte_panic("Error adding fd to %d epoll_ctl, %s\n",
1113                                         intr_pipe.readfd, strerror(errno));
1114                 }
1115                 numfds++;
1116
1117                 rte_spinlock_lock(&intr_lock);
1118
1119                 TAILQ_FOREACH(src, &intr_sources, next) {
1120                         struct epoll_event ev;
1121
1122                         if (src->callbacks.tqh_first == NULL)
1123                                 continue; /* skip those with no callbacks */
1124                         memset(&ev, 0, sizeof(ev));
1125                         ev.events = EPOLLIN | EPOLLPRI | EPOLLRDHUP | EPOLLHUP;
1126                         ev.data.fd = src->intr_handle.fd;
1127
1128                         /**
1129                          * add all the uio device file descriptor
1130                          * into wait list.
1131                          */
1132                         if (epoll_ctl(pfd, EPOLL_CTL_ADD,
1133                                         src->intr_handle.fd, &ev) < 0){
1134                                 rte_panic("Error adding fd %d epoll_ctl, %s\n",
1135                                         src->intr_handle.fd, strerror(errno));
1136                         }
1137                         else
1138                                 numfds++;
1139                 }
1140                 rte_spinlock_unlock(&intr_lock);
1141                 /* serve the interrupt */
1142                 eal_intr_handle_interrupts(pfd, numfds);
1143
1144                 /**
1145                  * when we return, we need to rebuild the
1146                  * list of fds to monitor.
1147                  */
1148                 close(pfd);
1149         }
1150 }
1151
1152 int
1153 rte_eal_intr_init(void)
1154 {
1155         int ret = 0;
1156
1157         /* init the global interrupt source head */
1158         TAILQ_INIT(&intr_sources);
1159
1160         /**
1161          * create a pipe which will be waited by epoll and notified to
1162          * rebuild the wait list of epoll.
1163          */
1164         if (pipe(intr_pipe.pipefd) < 0) {
1165                 rte_errno = errno;
1166                 return -1;
1167         }
1168
1169         /* create the host thread to wait/handle the interrupt */
1170         ret = rte_ctrl_thread_create(&intr_thread, "eal-intr-thread", NULL,
1171                         eal_intr_thread_main, NULL);
1172         if (ret != 0) {
1173                 rte_errno = -ret;
1174                 RTE_LOG(ERR, EAL,
1175                         "Failed to create thread for interrupt handling\n");
1176         }
1177
1178         return ret;
1179 }
1180
1181 static void
1182 eal_intr_proc_rxtx_intr(int fd, const struct rte_intr_handle *intr_handle)
1183 {
1184         union rte_intr_read_buffer buf;
1185         int bytes_read = 0;
1186         int nbytes;
1187
1188         switch (intr_handle->type) {
1189         case RTE_INTR_HANDLE_UIO:
1190         case RTE_INTR_HANDLE_UIO_INTX:
1191                 bytes_read = sizeof(buf.uio_intr_count);
1192                 break;
1193 #ifdef VFIO_PRESENT
1194         case RTE_INTR_HANDLE_VFIO_MSIX:
1195         case RTE_INTR_HANDLE_VFIO_MSI:
1196         case RTE_INTR_HANDLE_VFIO_LEGACY:
1197                 bytes_read = sizeof(buf.vfio_intr_count);
1198                 break;
1199 #endif
1200         case RTE_INTR_HANDLE_VDEV:
1201                 bytes_read = intr_handle->efd_counter_size;
1202                 /* For vdev, number of bytes to read is set by driver */
1203                 break;
1204         case RTE_INTR_HANDLE_EXT:
1205                 return;
1206         default:
1207                 bytes_read = 1;
1208                 RTE_LOG(INFO, EAL, "unexpected intr type\n");
1209                 break;
1210         }
1211
1212         /**
1213          * read out to clear the ready-to-be-read flag
1214          * for epoll_wait.
1215          */
1216         if (bytes_read == 0)
1217                 return;
1218         do {
1219                 nbytes = read(fd, &buf, bytes_read);
1220                 if (nbytes < 0) {
1221                         if (errno == EINTR || errno == EWOULDBLOCK ||
1222                             errno == EAGAIN)
1223                                 continue;
1224                         RTE_LOG(ERR, EAL,
1225                                 "Error reading from fd %d: %s\n",
1226                                 fd, strerror(errno));
1227                 } else if (nbytes == 0)
1228                         RTE_LOG(ERR, EAL, "Read nothing from fd %d\n", fd);
1229                 return;
1230         } while (1);
1231 }
1232
1233 static int
1234 eal_epoll_process_event(struct epoll_event *evs, unsigned int n,
1235                         struct rte_epoll_event *events)
1236 {
1237         unsigned int i, count = 0;
1238         struct rte_epoll_event *rev;
1239         uint32_t valid_status;
1240
1241         for (i = 0; i < n; i++) {
1242                 rev = evs[i].data.ptr;
1243                 valid_status =  RTE_EPOLL_VALID;
1244                 /* ACQUIRE memory ordering here pairs with RELEASE
1245                  * ordering below acting as a lock to synchronize
1246                  * the event data updating.
1247                  */
1248                 if (!rev || !__atomic_compare_exchange_n(&rev->status,
1249                                     &valid_status, RTE_EPOLL_EXEC, 0,
1250                                     __ATOMIC_ACQUIRE, __ATOMIC_RELAXED))
1251                         continue;
1252
1253                 events[count].status        = RTE_EPOLL_VALID;
1254                 events[count].fd            = rev->fd;
1255                 events[count].epfd          = rev->epfd;
1256                 events[count].epdata.event  = evs[i].events;
1257                 events[count].epdata.data   = rev->epdata.data;
1258                 if (rev->epdata.cb_fun)
1259                         rev->epdata.cb_fun(rev->fd,
1260                                            rev->epdata.cb_arg);
1261
1262                 /* the status update should be observed after
1263                  * the other fields change.
1264                  */
1265                 __atomic_store_n(&rev->status, RTE_EPOLL_VALID,
1266                                 __ATOMIC_RELEASE);
1267                 count++;
1268         }
1269         return count;
1270 }
1271
1272 static inline int
1273 eal_init_tls_epfd(void)
1274 {
1275         int pfd = epoll_create(255);
1276
1277         if (pfd < 0) {
1278                 RTE_LOG(ERR, EAL,
1279                         "Cannot create epoll instance\n");
1280                 return -1;
1281         }
1282         return pfd;
1283 }
1284
1285 int
1286 rte_intr_tls_epfd(void)
1287 {
1288         if (RTE_PER_LCORE(_epfd) == -1)
1289                 RTE_PER_LCORE(_epfd) = eal_init_tls_epfd();
1290
1291         return RTE_PER_LCORE(_epfd);
1292 }
1293
1294 static int
1295 eal_epoll_wait(int epfd, struct rte_epoll_event *events,
1296                int maxevents, int timeout, bool interruptible)
1297 {
1298         struct epoll_event evs[maxevents];
1299         int rc;
1300
1301         if (!events) {
1302                 RTE_LOG(ERR, EAL, "rte_epoll_event can't be NULL\n");
1303                 return -1;
1304         }
1305
1306         /* using per thread epoll fd */
1307         if (epfd == RTE_EPOLL_PER_THREAD)
1308                 epfd = rte_intr_tls_epfd();
1309
1310         while (1) {
1311                 rc = epoll_wait(epfd, evs, maxevents, timeout);
1312                 if (likely(rc > 0)) {
1313                         /* epoll_wait has at least one fd ready to read */
1314                         rc = eal_epoll_process_event(evs, rc, events);
1315                         break;
1316                 } else if (rc < 0) {
1317                         if (errno == EINTR) {
1318                                 if (interruptible)
1319                                         return -1;
1320                                 else
1321                                         continue;
1322                         }
1323                         /* epoll_wait fail */
1324                         RTE_LOG(ERR, EAL, "epoll_wait returns with fail %s\n",
1325                                 strerror(errno));
1326                         rc = -1;
1327                         break;
1328                 } else {
1329                         /* rc == 0, epoll_wait timed out */
1330                         break;
1331                 }
1332         }
1333
1334         return rc;
1335 }
1336
1337 int
1338 rte_epoll_wait(int epfd, struct rte_epoll_event *events,
1339                int maxevents, int timeout)
1340 {
1341         return eal_epoll_wait(epfd, events, maxevents, timeout, false);
1342 }
1343
1344 int
1345 rte_epoll_wait_interruptible(int epfd, struct rte_epoll_event *events,
1346                              int maxevents, int timeout)
1347 {
1348         return eal_epoll_wait(epfd, events, maxevents, timeout, true);
1349 }
1350
1351 static inline void
1352 eal_epoll_data_safe_free(struct rte_epoll_event *ev)
1353 {
1354         uint32_t valid_status = RTE_EPOLL_VALID;
1355
1356         while (!__atomic_compare_exchange_n(&ev->status, &valid_status,
1357                     RTE_EPOLL_INVALID, 0, __ATOMIC_ACQUIRE, __ATOMIC_RELAXED)) {
1358                 while (__atomic_load_n(&ev->status,
1359                                 __ATOMIC_RELAXED) != RTE_EPOLL_VALID)
1360                         rte_pause();
1361                 valid_status = RTE_EPOLL_VALID;
1362         }
1363         memset(&ev->epdata, 0, sizeof(ev->epdata));
1364         ev->fd = -1;
1365         ev->epfd = -1;
1366 }
1367
1368 int
1369 rte_epoll_ctl(int epfd, int op, int fd,
1370               struct rte_epoll_event *event)
1371 {
1372         struct epoll_event ev;
1373
1374         if (!event) {
1375                 RTE_LOG(ERR, EAL, "rte_epoll_event can't be NULL\n");
1376                 return -1;
1377         }
1378
1379         /* using per thread epoll fd */
1380         if (epfd == RTE_EPOLL_PER_THREAD)
1381                 epfd = rte_intr_tls_epfd();
1382
1383         if (op == EPOLL_CTL_ADD) {
1384                 __atomic_store_n(&event->status, RTE_EPOLL_VALID,
1385                                 __ATOMIC_RELAXED);
1386                 event->fd = fd;  /* ignore fd in event */
1387                 event->epfd = epfd;
1388                 ev.data.ptr = (void *)event;
1389         }
1390
1391         ev.events = event->epdata.event;
1392         if (epoll_ctl(epfd, op, fd, &ev) < 0) {
1393                 RTE_LOG(ERR, EAL, "Error op %d fd %d epoll_ctl, %s\n",
1394                         op, fd, strerror(errno));
1395                 if (op == EPOLL_CTL_ADD)
1396                         /* rollback status when CTL_ADD fail */
1397                         __atomic_store_n(&event->status, RTE_EPOLL_INVALID,
1398                                         __ATOMIC_RELAXED);
1399                 return -1;
1400         }
1401
1402         if (op == EPOLL_CTL_DEL && __atomic_load_n(&event->status,
1403                         __ATOMIC_RELAXED) != RTE_EPOLL_INVALID)
1404                 eal_epoll_data_safe_free(event);
1405
1406         return 0;
1407 }
1408
1409 int
1410 rte_intr_rx_ctl(struct rte_intr_handle *intr_handle, int epfd,
1411                 int op, unsigned int vec, void *data)
1412 {
1413         struct rte_epoll_event *rev;
1414         struct rte_epoll_data *epdata;
1415         int epfd_op;
1416         unsigned int efd_idx;
1417         int rc = 0;
1418
1419         efd_idx = (vec >= RTE_INTR_VEC_RXTX_OFFSET) ?
1420                 (vec - RTE_INTR_VEC_RXTX_OFFSET) : vec;
1421
1422         if (!intr_handle || intr_handle->nb_efd == 0 ||
1423             efd_idx >= intr_handle->nb_efd) {
1424                 RTE_LOG(ERR, EAL, "Wrong intr vector number.\n");
1425                 return -EPERM;
1426         }
1427
1428         switch (op) {
1429         case RTE_INTR_EVENT_ADD:
1430                 epfd_op = EPOLL_CTL_ADD;
1431                 rev = &intr_handle->elist[efd_idx];
1432                 if (__atomic_load_n(&rev->status,
1433                                 __ATOMIC_RELAXED) != RTE_EPOLL_INVALID) {
1434                         RTE_LOG(INFO, EAL, "Event already been added.\n");
1435                         return -EEXIST;
1436                 }
1437
1438                 /* attach to intr vector fd */
1439                 epdata = &rev->epdata;
1440                 epdata->event  = EPOLLIN | EPOLLPRI | EPOLLET;
1441                 epdata->data   = data;
1442                 epdata->cb_fun = (rte_intr_event_cb_t)eal_intr_proc_rxtx_intr;
1443                 epdata->cb_arg = (void *)intr_handle;
1444                 rc = rte_epoll_ctl(epfd, epfd_op,
1445                                    intr_handle->efds[efd_idx], rev);
1446                 if (!rc)
1447                         RTE_LOG(DEBUG, EAL,
1448                                 "efd %d associated with vec %d added on epfd %d"
1449                                 "\n", rev->fd, vec, epfd);
1450                 else
1451                         rc = -EPERM;
1452                 break;
1453         case RTE_INTR_EVENT_DEL:
1454                 epfd_op = EPOLL_CTL_DEL;
1455                 rev = &intr_handle->elist[efd_idx];
1456                 if (__atomic_load_n(&rev->status,
1457                                 __ATOMIC_RELAXED) == RTE_EPOLL_INVALID) {
1458                         RTE_LOG(INFO, EAL, "Event does not exist.\n");
1459                         return -EPERM;
1460                 }
1461
1462                 rc = rte_epoll_ctl(rev->epfd, epfd_op, rev->fd, rev);
1463                 if (rc)
1464                         rc = -EPERM;
1465                 break;
1466         default:
1467                 RTE_LOG(ERR, EAL, "event op type mismatch\n");
1468                 rc = -EPERM;
1469         }
1470
1471         return rc;
1472 }
1473
1474 void
1475 rte_intr_free_epoll_fd(struct rte_intr_handle *intr_handle)
1476 {
1477         uint32_t i;
1478         struct rte_epoll_event *rev;
1479
1480         for (i = 0; i < intr_handle->nb_efd; i++) {
1481                 rev = &intr_handle->elist[i];
1482                 if (__atomic_load_n(&rev->status,
1483                                 __ATOMIC_RELAXED) == RTE_EPOLL_INVALID)
1484                         continue;
1485                 if (rte_epoll_ctl(rev->epfd, EPOLL_CTL_DEL, rev->fd, rev)) {
1486                         /* force free if the entry valid */
1487                         eal_epoll_data_safe_free(rev);
1488                 }
1489         }
1490 }
1491
1492 int
1493 rte_intr_efd_enable(struct rte_intr_handle *intr_handle, uint32_t nb_efd)
1494 {
1495         uint32_t i;
1496         int fd;
1497         uint32_t n = RTE_MIN(nb_efd, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
1498
1499         assert(nb_efd != 0);
1500
1501         if (intr_handle->type == RTE_INTR_HANDLE_VFIO_MSIX) {
1502                 for (i = 0; i < n; i++) {
1503                         fd = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC);
1504                         if (fd < 0) {
1505                                 RTE_LOG(ERR, EAL,
1506                                         "can't setup eventfd, error %i (%s)\n",
1507                                         errno, strerror(errno));
1508                                 return -errno;
1509                         }
1510                         intr_handle->efds[i] = fd;
1511                 }
1512                 intr_handle->nb_efd   = n;
1513                 intr_handle->max_intr = NB_OTHER_INTR + n;
1514         } else if (intr_handle->type == RTE_INTR_HANDLE_VDEV) {
1515                 /* only check, initialization would be done in vdev driver.*/
1516                 if (intr_handle->efd_counter_size >
1517                     sizeof(union rte_intr_read_buffer)) {
1518                         RTE_LOG(ERR, EAL, "the efd_counter_size is oversized");
1519                         return -EINVAL;
1520                 }
1521         } else {
1522                 intr_handle->efds[0]  = intr_handle->fd;
1523                 intr_handle->nb_efd   = RTE_MIN(nb_efd, 1U);
1524                 intr_handle->max_intr = NB_OTHER_INTR;
1525         }
1526
1527         return 0;
1528 }
1529
1530 void
1531 rte_intr_efd_disable(struct rte_intr_handle *intr_handle)
1532 {
1533         uint32_t i;
1534
1535         rte_intr_free_epoll_fd(intr_handle);
1536         if (intr_handle->max_intr > intr_handle->nb_efd) {
1537                 for (i = 0; i < intr_handle->nb_efd; i++)
1538                         close(intr_handle->efds[i]);
1539         }
1540         intr_handle->nb_efd = 0;
1541         intr_handle->max_intr = 0;
1542 }
1543
1544 int
1545 rte_intr_dp_is_en(struct rte_intr_handle *intr_handle)
1546 {
1547         return !(!intr_handle->nb_efd);
1548 }
1549
1550 int
1551 rte_intr_allow_others(struct rte_intr_handle *intr_handle)
1552 {
1553         if (!rte_intr_dp_is_en(intr_handle))
1554                 return 1;
1555         else
1556                 return !!(intr_handle->max_intr - intr_handle->nb_efd);
1557 }
1558
1559 int
1560 rte_intr_cap_multiple(struct rte_intr_handle *intr_handle)
1561 {
1562         if (intr_handle->type == RTE_INTR_HANDLE_VFIO_MSIX)
1563                 return 1;
1564
1565         if (intr_handle->type == RTE_INTR_HANDLE_VDEV)
1566                 return 1;
1567
1568         return 0;
1569 }
1570
1571 int rte_thread_is_intr(void)
1572 {
1573         return pthread_equal(intr_thread, pthread_self());
1574 }