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, vfio_dev_fd;
 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 = rte_intr_fd_get(intr_handle);
 129
 130         vfio_dev_fd = rte_intr_dev_fd_get(intr_handle);
 131         ret = ioctl(vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 132
 133         if (ret) {
 134                 RTE_LOG(ERR, EAL, "Error enabling INTx interrupts for fd %d\n",
 135                         rte_intr_fd_get(intr_handle));
 136                 return -1;
 137         }
 138
 139         /* unmask INTx after enabling */
 140         memset(irq_set, 0, len);
 141         len = sizeof(struct vfio_irq_set);
 142         irq_set->argsz = len;
 143         irq_set->count = 1;
 144         irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK;
 145         irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
 146         irq_set->start = 0;
 147
 148         ret = ioctl(vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 149
 150         if (ret) {
 151                 RTE_LOG(ERR, EAL, "Error unmasking INTx interrupts for fd %d\n",
 152                         rte_intr_fd_get(intr_handle));
 153                 return -1;
 154         }
 155         return 0;
 156 }
 157
 158 /* disable legacy (INTx) interrupts */
 159 static int
 160 vfio_disable_intx(const struct rte_intr_handle *intr_handle) {
 161         struct vfio_irq_set *irq_set;
 162         char irq_set_buf[IRQ_SET_BUF_LEN];
 163         int len, ret, vfio_dev_fd;
 164
 165         len = sizeof(struct vfio_irq_set);
 166
 167         /* mask interrupts before disabling */
 168         irq_set = (struct vfio_irq_set *) irq_set_buf;
 169         irq_set->argsz = len;
 170         irq_set->count = 1;
 171         irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_MASK;
 172         irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
 173         irq_set->start = 0;
 174
 175         vfio_dev_fd = rte_intr_dev_fd_get(intr_handle);
 176         ret = ioctl(vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 177
 178         if (ret) {
 179                 RTE_LOG(ERR, EAL, "Error masking INTx interrupts for fd %d\n",
 180                         rte_intr_fd_get(intr_handle));
 181                 return -1;
 182         }
 183
 184         /* disable INTx*/
 185         memset(irq_set, 0, len);
 186         irq_set->argsz = len;
 187         irq_set->count = 0;
 188         irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER;
 189         irq_set->index = VFIO_PCI_INTX_IRQ_INDEX;
 190         irq_set->start = 0;
 191
 192         ret = ioctl(vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 193
 194         if (ret) {
 195                 RTE_LOG(ERR, EAL, "Error disabling INTx interrupts for fd %d\n",
 196                         rte_intr_fd_get(intr_handle));
 197                 return -1;
 198         }
 199         return 0;
 200 }
 201
 202 /* unmask/ack legacy (INTx) interrupts */
 203 static int
 204 vfio_ack_intx(const struct rte_intr_handle *intr_handle)
 205 {
 206         struct vfio_irq_set irq_set;
 207         int vfio_dev_fd;
 208
 209         /* unmask INTx */
 210         memset(&irq_set, 0, sizeof(irq_set));
 211         irq_set.argsz = sizeof(irq_set);
 212         irq_set.count = 1;
 213         irq_set.flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_UNMASK;
 214         irq_set.index = VFIO_PCI_INTX_IRQ_INDEX;
 215         irq_set.start = 0;
 216
 217         vfio_dev_fd = rte_intr_dev_fd_get(intr_handle);
 218         if (ioctl(vfio_dev_fd, VFIO_DEVICE_SET_IRQS, &irq_set)) {
 219                 RTE_LOG(ERR, EAL, "Error unmasking INTx interrupts for fd %d\n",
 220                         rte_intr_fd_get(intr_handle));
 221                 return -1;
 222         }
 223         return 0;
 224 }
 225
 226 /* enable MSI interrupts */
 227 static int
 228 vfio_enable_msi(const struct rte_intr_handle *intr_handle) {
 229         int len, ret;
 230         char irq_set_buf[IRQ_SET_BUF_LEN];
 231         struct vfio_irq_set *irq_set;
 232         int *fd_ptr, vfio_dev_fd;
 233
 234         len = sizeof(irq_set_buf);
 235
 236         irq_set = (struct vfio_irq_set *) irq_set_buf;
 237         irq_set->argsz = len;
 238         irq_set->count = 1;
 239         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
 240         irq_set->index = VFIO_PCI_MSI_IRQ_INDEX;
 241         irq_set->start = 0;
 242         fd_ptr = (int *) &irq_set->data;
 243         *fd_ptr = rte_intr_fd_get(intr_handle);
 244
 245         vfio_dev_fd = rte_intr_dev_fd_get(intr_handle);
 246         ret = ioctl(vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 247
 248         if (ret) {
 249                 RTE_LOG(ERR, EAL, "Error enabling MSI interrupts for fd %d\n",
 250                         rte_intr_fd_get(intr_handle));
 251                 return -1;
 252         }
 253         return 0;
 254 }
 255
 256 /* disable MSI interrupts */
 257 static int
 258 vfio_disable_msi(const struct rte_intr_handle *intr_handle) {
 259         struct vfio_irq_set *irq_set;
 260         char irq_set_buf[IRQ_SET_BUF_LEN];
 261         int len, ret, vfio_dev_fd;
 262
 263         len = sizeof(struct vfio_irq_set);
 264
 265         irq_set = (struct vfio_irq_set *) irq_set_buf;
 266         irq_set->argsz = len;
 267         irq_set->count = 0;
 268         irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER;
 269         irq_set->index = VFIO_PCI_MSI_IRQ_INDEX;
 270         irq_set->start = 0;
 271
 272         vfio_dev_fd = rte_intr_dev_fd_get(intr_handle);
 273         ret = ioctl(vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 274         if (ret)
 275                 RTE_LOG(ERR, EAL, "Error disabling MSI interrupts for fd %d\n",
 276                         rte_intr_fd_get(intr_handle));
 277
 278         return ret;
 279 }
 280
 281 /* enable MSI-X interrupts */
 282 static int
 283 vfio_enable_msix(const struct rte_intr_handle *intr_handle) {
 284         int len, ret;
 285         char irq_set_buf[MSIX_IRQ_SET_BUF_LEN];
 286         struct vfio_irq_set *irq_set;
 287         int *fd_ptr, vfio_dev_fd, i;
 288
 289         len = sizeof(irq_set_buf);
 290
 291         irq_set = (struct vfio_irq_set *) irq_set_buf;
 292         irq_set->argsz = len;
 293         /* 0 < irq_set->count < RTE_MAX_RXTX_INTR_VEC_ID + 1 */
 294         irq_set->count = rte_intr_max_intr_get(intr_handle) ?
 295                 (rte_intr_max_intr_get(intr_handle) >
 296                  RTE_MAX_RXTX_INTR_VEC_ID + 1 ? RTE_MAX_RXTX_INTR_VEC_ID + 1 :
 297                  rte_intr_max_intr_get(intr_handle)) : 1;
 298
 299         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD | VFIO_IRQ_SET_ACTION_TRIGGER;
 300         irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
 301         irq_set->start = 0;
 302         fd_ptr = (int *) &irq_set->data;
 303         /* INTR vector offset 0 reserve for non-efds mapping */
 304         fd_ptr[RTE_INTR_VEC_ZERO_OFFSET] = rte_intr_fd_get(intr_handle);
 305         for (i = 0; i < rte_intr_nb_efd_get(intr_handle); i++) {
 306                 fd_ptr[RTE_INTR_VEC_RXTX_OFFSET + i] =
 307                         rte_intr_efds_index_get(intr_handle, i);
 308         }
 309
 310         vfio_dev_fd = rte_intr_dev_fd_get(intr_handle);
 311         ret = ioctl(vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 312
 313         if (ret) {
 314                 RTE_LOG(ERR, EAL, "Error enabling MSI-X interrupts for fd %d\n",
 315                         rte_intr_fd_get(intr_handle));
 316                 return -1;
 317         }
 318
 319         return 0;
 320 }
 321
 322 /* disable MSI-X interrupts */
 323 static int
 324 vfio_disable_msix(const struct rte_intr_handle *intr_handle) {
 325         struct vfio_irq_set *irq_set;
 326         char irq_set_buf[MSIX_IRQ_SET_BUF_LEN];
 327         int len, ret, vfio_dev_fd;
 328
 329         len = sizeof(struct vfio_irq_set);
 330
 331         irq_set = (struct vfio_irq_set *) irq_set_buf;
 332         irq_set->argsz = len;
 333         irq_set->count = 0;
 334         irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER;
 335         irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
 336         irq_set->start = 0;
 337
 338         vfio_dev_fd = rte_intr_dev_fd_get(intr_handle);
 339         ret = ioctl(vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 340
 341         if (ret)
 342                 RTE_LOG(ERR, EAL, "Error disabling MSI-X interrupts for fd %d\n",
 343                         rte_intr_fd_get(intr_handle));
 344
 345         return ret;
 346 }
 347
 348 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
 349 /* enable req notifier */
 350 static int
 351 vfio_enable_req(const struct rte_intr_handle *intr_handle)
 352 {
 353         int len, ret;
 354         char irq_set_buf[IRQ_SET_BUF_LEN];
 355         struct vfio_irq_set *irq_set;
 356         int *fd_ptr, vfio_dev_fd;
 357
 358         len = sizeof(irq_set_buf);
 359
 360         irq_set = (struct vfio_irq_set *) irq_set_buf;
 361         irq_set->argsz = len;
 362         irq_set->count = 1;
 363         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
 364                          VFIO_IRQ_SET_ACTION_TRIGGER;
 365         irq_set->index = VFIO_PCI_REQ_IRQ_INDEX;
 366         irq_set->start = 0;
 367         fd_ptr = (int *) &irq_set->data;
 368         *fd_ptr = rte_intr_fd_get(intr_handle);
 369
 370         vfio_dev_fd = rte_intr_dev_fd_get(intr_handle);
 371         ret = ioctl(vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 372
 373         if (ret) {
 374                 RTE_LOG(ERR, EAL, "Error enabling req interrupts for fd %d\n",
 375                         rte_intr_fd_get(intr_handle));
 376                 return -1;
 377         }
 378
 379         return 0;
 380 }
 381
 382 /* disable req notifier */
 383 static int
 384 vfio_disable_req(const struct rte_intr_handle *intr_handle)
 385 {
 386         struct vfio_irq_set *irq_set;
 387         char irq_set_buf[IRQ_SET_BUF_LEN];
 388         int len, ret, vfio_dev_fd;
 389
 390         len = sizeof(struct vfio_irq_set);
 391
 392         irq_set = (struct vfio_irq_set *) irq_set_buf;
 393         irq_set->argsz = len;
 394         irq_set->count = 0;
 395         irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER;
 396         irq_set->index = VFIO_PCI_REQ_IRQ_INDEX;
 397         irq_set->start = 0;
 398
 399         vfio_dev_fd = rte_intr_dev_fd_get(intr_handle);
 400         ret = ioctl(vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 401
 402         if (ret)
 403                 RTE_LOG(ERR, EAL, "Error disabling req interrupts for fd %d\n",
 404                         rte_intr_fd_get(intr_handle));
 405
 406         return ret;
 407 }
 408 #endif
 409 #endif
 410
 411 static int
 412 uio_intx_intr_disable(const struct rte_intr_handle *intr_handle)
 413 {
 414         unsigned char command_high;
 415         int uio_cfg_fd;
 416
 417         /* use UIO config file descriptor for uio_pci_generic */
 418         uio_cfg_fd = rte_intr_dev_fd_get(intr_handle);
 419         if (uio_cfg_fd < 0 || pread(uio_cfg_fd, &command_high, 1, 5) != 1) {
 420                 RTE_LOG(ERR, EAL,
 421                         "Error reading interrupts status for fd %d\n",
 422                         uio_cfg_fd);
 423                 return -1;
 424         }
 425         /* disable interrupts */
 426         command_high |= 0x4;
 427         if (pwrite(uio_cfg_fd, &command_high, 1, 5) != 1) {
 428                 RTE_LOG(ERR, EAL,
 429                         "Error disabling interrupts for fd %d\n",
 430                         uio_cfg_fd);
 431                 return -1;
 432         }
 433
 434         return 0;
 435 }
 436
 437 static int
 438 uio_intx_intr_enable(const struct rte_intr_handle *intr_handle)
 439 {
 440         unsigned char command_high;
 441         int uio_cfg_fd;
 442
 443         /* use UIO config file descriptor for uio_pci_generic */
 444         uio_cfg_fd = rte_intr_dev_fd_get(intr_handle);
 445         if (uio_cfg_fd < 0 || pread(uio_cfg_fd, &command_high, 1, 5) != 1) {
 446                 RTE_LOG(ERR, EAL,
 447                         "Error reading interrupts status for fd %d\n",
 448                         uio_cfg_fd);
 449                 return -1;
 450         }
 451         /* enable interrupts */
 452         command_high &= ~0x4;
 453         if (pwrite(uio_cfg_fd, &command_high, 1, 5) != 1) {
 454                 RTE_LOG(ERR, EAL,
 455                         "Error enabling interrupts for fd %d\n",
 456                         uio_cfg_fd);
 457                 return -1;
 458         }
 459
 460         return 0;
 461 }
 462
 463 static int
 464 uio_intr_disable(const struct rte_intr_handle *intr_handle)
 465 {
 466         const int value = 0;
 467
 468         if (rte_intr_fd_get(intr_handle) < 0 ||
 469             write(rte_intr_fd_get(intr_handle), &value, sizeof(value)) < 0) {
 470                 RTE_LOG(ERR, EAL, "Error disabling interrupts for fd %d (%s)\n",
 471                         rte_intr_fd_get(intr_handle), strerror(errno));
 472                 return -1;
 473         }
 474         return 0;
 475 }
 476
 477 static int
 478 uio_intr_enable(const struct rte_intr_handle *intr_handle)
 479 {
 480         const int value = 1;
 481
 482         if (rte_intr_fd_get(intr_handle) < 0 ||
 483             write(rte_intr_fd_get(intr_handle), &value, sizeof(value)) < 0) {
 484                 RTE_LOG(ERR, EAL, "Error enabling interrupts for fd %d (%s)\n",
 485                         rte_intr_fd_get(intr_handle), strerror(errno));
 486                 return -1;
 487         }
 488         return 0;
 489 }
 490
 491 int
 492 rte_intr_callback_register(const struct rte_intr_handle *intr_handle,
 493                         rte_intr_callback_fn cb, void *cb_arg)
 494 {
 495         int ret, wake_thread;
 496         struct rte_intr_source *src;
 497         struct rte_intr_callback *callback;
 498
 499         wake_thread = 0;
 500
 501         /* first do parameter checking */
 502         if (rte_intr_fd_get(intr_handle) < 0 || cb == NULL) {
 503                 RTE_LOG(ERR, EAL, "Registering with invalid input parameter\n");
 504                 return -EINVAL;
 505         }
 506
 507         /* allocate a new interrupt callback entity */
 508         callback = calloc(1, sizeof(*callback));
 509         if (callback == NULL) {
 510                 RTE_LOG(ERR, EAL, "Can not allocate memory\n");
 511                 return -ENOMEM;
 512         }
 513         callback->cb_fn = cb;
 514         callback->cb_arg = cb_arg;
 515         callback->pending_delete = 0;
 516         callback->ucb_fn = NULL;
 517
 518         rte_spinlock_lock(&intr_lock);
 519
 520         /* check if there is at least one callback registered for the fd */
 521         TAILQ_FOREACH(src, &intr_sources, next) {
 522                 if (rte_intr_fd_get(src->intr_handle) == rte_intr_fd_get(intr_handle)) {
 523                         /* we had no interrupts for this */
 524                         if (TAILQ_EMPTY(&src->callbacks))
 525                                 wake_thread = 1;
 526
 527                         TAILQ_INSERT_TAIL(&(src->callbacks), callback, next);
 528                         ret = 0;
 529                         break;
 530                 }
 531         }
 532
 533         /* no existing callbacks for this - add new source */
 534         if (src == NULL) {
 535                 src = calloc(1, sizeof(*src));
 536                 if (src == NULL) {
 537                         RTE_LOG(ERR, EAL, "Can not allocate memory\n");
 538                         ret = -ENOMEM;
 539                         free(callback);
 540                         callback = NULL;
 541                 } else {
 542                         src->intr_handle = rte_intr_instance_dup(intr_handle);
 543                         if (src->intr_handle == NULL) {
 544                                 RTE_LOG(ERR, EAL, "Can not create intr instance\n");
 545                                 ret = -ENOMEM;
 546                                 free(callback);
 547                                 callback = NULL;
 548                                 free(src);
 549                                 src = NULL;
 550                         } else {
 551                                 TAILQ_INIT(&src->callbacks);
 552                                 TAILQ_INSERT_TAIL(&(src->callbacks), callback,
 553                                                   next);
 554                                 TAILQ_INSERT_TAIL(&intr_sources, src, next);
 555                                 wake_thread = 1;
 556                                 ret = 0;
 557                         }
 558                 }
 559         }
 560
 561         rte_spinlock_unlock(&intr_lock);
 562
 563         /**
 564          * check if need to notify the pipe fd waited by epoll_wait to
 565          * rebuild the wait list.
 566          */
 567         if (wake_thread)
 568                 if (write(intr_pipe.writefd, "1", 1) < 0)
 569                         ret = -EPIPE;
 570
 571         rte_eal_trace_intr_callback_register(intr_handle, cb, cb_arg, ret);
 572         return ret;
 573 }
 574
 575 int
 576 rte_intr_callback_unregister_pending(const struct rte_intr_handle *intr_handle,
 577                                 rte_intr_callback_fn cb_fn, void *cb_arg,
 578                                 rte_intr_unregister_callback_fn ucb_fn)
 579 {
 580         int ret;
 581         struct rte_intr_source *src;
 582         struct rte_intr_callback *cb, *next;
 583
 584         /* do parameter checking first */
 585         if (rte_intr_fd_get(intr_handle) < 0) {
 586                 RTE_LOG(ERR, EAL, "Unregistering with invalid input parameter\n");
 587                 return -EINVAL;
 588         }
 589
 590         rte_spinlock_lock(&intr_lock);
 591
 592         /* check if the insterrupt source for the fd is existent */
 593         TAILQ_FOREACH(src, &intr_sources, next) {
 594                 if (rte_intr_fd_get(src->intr_handle) == rte_intr_fd_get(intr_handle))
 595                         break;
 596         }
 597
 598         /* No interrupt source registered for the fd */
 599         if (src == NULL) {
 600                 ret = -ENOENT;
 601
 602         /* only usable if the source is active */
 603         } else if (src->active == 0) {
 604                 ret = -EAGAIN;
 605
 606         } else {
 607                 ret = 0;
 608
 609                 /* walk through the callbacks and mark all that match. */
 610                 for (cb = TAILQ_FIRST(&src->callbacks); cb != NULL; cb = next) {
 611                         next = TAILQ_NEXT(cb, next);
 612                         if (cb->cb_fn == cb_fn && (cb_arg == (void *)-1 ||
 613                                         cb->cb_arg == cb_arg)) {
 614                                 cb->pending_delete = 1;
 615                                 cb->ucb_fn = ucb_fn;
 616                                 ret++;
 617                         }
 618                 }
 619         }
 620
 621         rte_spinlock_unlock(&intr_lock);
 622
 623         return ret;
 624 }
 625
 626 int
 627 rte_intr_callback_unregister(const struct rte_intr_handle *intr_handle,
 628                         rte_intr_callback_fn cb_fn, void *cb_arg)
 629 {
 630         int ret;
 631         struct rte_intr_source *src;
 632         struct rte_intr_callback *cb, *next;
 633
 634         /* do parameter checking first */
 635         if (rte_intr_fd_get(intr_handle) < 0) {
 636                 RTE_LOG(ERR, EAL, "Unregistering with invalid input parameter\n");
 637                 return -EINVAL;
 638         }
 639
 640         rte_spinlock_lock(&intr_lock);
 641
 642         /* check if the insterrupt source for the fd is existent */
 643         TAILQ_FOREACH(src, &intr_sources, next)
 644                 if (rte_intr_fd_get(src->intr_handle) == rte_intr_fd_get(intr_handle))
 645                         break;
 646
 647         /* No interrupt source registered for the fd */
 648         if (src == NULL) {
 649                 ret = -ENOENT;
 650
 651         /* interrupt source has some active callbacks right now. */
 652         } else if (src->active != 0) {
 653                 ret = -EAGAIN;
 654
 655         /* ok to remove. */
 656         } else {
 657                 ret = 0;
 658
 659                 /*walk through the callbacks and remove all that match. */
 660                 for (cb = TAILQ_FIRST(&src->callbacks); cb != NULL; cb = next) {
 661
 662                         next = TAILQ_NEXT(cb, next);
 663
 664                         if (cb->cb_fn == cb_fn && (cb_arg == (void *)-1 ||
 665                                         cb->cb_arg == cb_arg)) {
 666                                 TAILQ_REMOVE(&src->callbacks, cb, next);
 667                                 free(cb);
 668                                 ret++;
 669                         }
 670                 }
 671
 672                 /* all callbacks for that source are removed. */
 673                 if (TAILQ_EMPTY(&src->callbacks)) {
 674                         TAILQ_REMOVE(&intr_sources, src, next);
 675                         rte_intr_instance_free(src->intr_handle);
 676                         free(src);
 677                 }
 678         }
 679
 680         rte_spinlock_unlock(&intr_lock);
 681
 682         /* notify the pipe fd waited by epoll_wait to rebuild the wait list */
 683         if (ret >= 0 && write(intr_pipe.writefd, "1", 1) < 0) {
 684                 ret = -EPIPE;
 685         }
 686
 687         rte_eal_trace_intr_callback_unregister(intr_handle, cb_fn, cb_arg,
 688                 ret);
 689         return ret;
 690 }
 691
 692 int
 693 rte_intr_callback_unregister_sync(const struct rte_intr_handle *intr_handle,
 694                         rte_intr_callback_fn cb_fn, void *cb_arg)
 695 {
 696         int ret = 0;
 697
 698         while ((ret = rte_intr_callback_unregister(intr_handle, cb_fn, cb_arg)) == -EAGAIN)
 699                 rte_pause();
 700
 701         return ret;
 702 }
 703
 704 int
 705 rte_intr_enable(const struct rte_intr_handle *intr_handle)
 706 {
 707         int rc = 0, uio_cfg_fd;
 708
 709         if (intr_handle == NULL)
 710                 return -1;
 711
 712         if (rte_intr_type_get(intr_handle) == RTE_INTR_HANDLE_VDEV) {
 713                 rc = 0;
 714                 goto out;
 715         }
 716
 717         uio_cfg_fd = rte_intr_dev_fd_get(intr_handle);
 718         if (rte_intr_fd_get(intr_handle) < 0 || uio_cfg_fd < 0) {
 719                 rc = -1;
 720                 goto out;
 721         }
 722
 723         switch (rte_intr_type_get(intr_handle)) {
 724         /* write to the uio fd to enable the interrupt */
 725         case RTE_INTR_HANDLE_UIO:
 726                 if (uio_intr_enable(intr_handle))
 727                         rc = -1;
 728                 break;
 729         case RTE_INTR_HANDLE_UIO_INTX:
 730                 if (uio_intx_intr_enable(intr_handle))
 731                         rc = -1;
 732                 break;
 733         /* not used at this moment */
 734         case RTE_INTR_HANDLE_ALARM:
 735                 rc = -1;
 736                 break;
 737 #ifdef VFIO_PRESENT
 738         case RTE_INTR_HANDLE_VFIO_MSIX:
 739                 if (vfio_enable_msix(intr_handle))
 740                         rc = -1;
 741                 break;
 742         case RTE_INTR_HANDLE_VFIO_MSI:
 743                 if (vfio_enable_msi(intr_handle))
 744                         rc = -1;
 745                 break;
 746         case RTE_INTR_HANDLE_VFIO_LEGACY:
 747                 if (vfio_enable_intx(intr_handle))
 748                         rc = -1;
 749                 break;
 750 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
 751         case RTE_INTR_HANDLE_VFIO_REQ:
 752                 if (vfio_enable_req(intr_handle))
 753                         rc = -1;
 754                 break;
 755 #endif
 756 #endif
 757         /* not used at this moment */
 758         case RTE_INTR_HANDLE_DEV_EVENT:
 759                 rc = -1;
 760                 break;
 761         /* unknown handle type */
 762         default:
 763                 RTE_LOG(ERR, EAL, "Unknown handle type of fd %d\n",
 764                         rte_intr_fd_get(intr_handle));
 765                 rc = -1;
 766                 break;
 767         }
 768 out:
 769         rte_eal_trace_intr_enable(intr_handle, rc);
 770         return rc;
 771 }
 772
 773 /**
 774  * PMD generally calls this function at the end of its IRQ callback.
 775  * Internally, it unmasks the interrupt if possible.
 776  *
 777  * For INTx, unmasking is required as the interrupt is auto-masked prior to
 778  * invoking callback.
 779  *
 780  * For MSI/MSI-X, unmasking is typically not needed as the interrupt is not
 781  * auto-masked. In fact, for interrupt handle types VFIO_MSIX and VFIO_MSI,
 782  * this function is no-op.
 783  */
 784 int
 785 rte_intr_ack(const struct rte_intr_handle *intr_handle)
 786 {
 787         int uio_cfg_fd;
 788
 789         if (rte_intr_type_get(intr_handle) == RTE_INTR_HANDLE_VDEV)
 790                 return 0;
 791
 792         uio_cfg_fd = rte_intr_dev_fd_get(intr_handle);
 793         if (rte_intr_fd_get(intr_handle) < 0 || uio_cfg_fd < 0)
 794                 return -1;
 795
 796         switch (rte_intr_type_get(intr_handle)) {
 797         /* Both acking and enabling are same for UIO */
 798         case RTE_INTR_HANDLE_UIO:
 799                 if (uio_intr_enable(intr_handle))
 800                         return -1;
 801                 break;
 802         case RTE_INTR_HANDLE_UIO_INTX:
 803                 if (uio_intx_intr_enable(intr_handle))
 804                         return -1;
 805                 break;
 806         /* not used at this moment */
 807         case RTE_INTR_HANDLE_ALARM:
 808                 return -1;
 809 #ifdef VFIO_PRESENT
 810         /* VFIO MSI* is implicitly acked unlike INTx, nothing to do */
 811         case RTE_INTR_HANDLE_VFIO_MSIX:
 812         case RTE_INTR_HANDLE_VFIO_MSI:
 813                 return 0;
 814         case RTE_INTR_HANDLE_VFIO_LEGACY:
 815                 if (vfio_ack_intx(intr_handle))
 816                         return -1;
 817                 break;
 818 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
 819         case RTE_INTR_HANDLE_VFIO_REQ:
 820                 return -1;
 821 #endif
 822 #endif
 823         /* not used at this moment */
 824         case RTE_INTR_HANDLE_DEV_EVENT:
 825                 return -1;
 826         /* unknown handle type */
 827         default:
 828                 RTE_LOG(ERR, EAL, "Unknown handle type of fd %d\n",
 829                         rte_intr_fd_get(intr_handle));
 830                 return -1;
 831         }
 832
 833         return 0;
 834 }
 835
 836 int
 837 rte_intr_disable(const struct rte_intr_handle *intr_handle)
 838 {
 839         int rc = 0, uio_cfg_fd;
 840
 841         if (intr_handle == NULL)
 842                 return -1;
 843
 844         if (rte_intr_type_get(intr_handle) == RTE_INTR_HANDLE_VDEV) {
 845                 rc = 0;
 846                 goto out;
 847         }
 848
 849         uio_cfg_fd = rte_intr_dev_fd_get(intr_handle);
 850         if (rte_intr_fd_get(intr_handle) < 0 || uio_cfg_fd < 0) {
 851                 rc = -1;
 852                 goto out;
 853         }
 854
 855         switch (rte_intr_type_get(intr_handle)) {
 856         /* write to the uio fd to disable the interrupt */
 857         case RTE_INTR_HANDLE_UIO:
 858                 if (uio_intr_disable(intr_handle))
 859                         rc = -1;
 860                 break;
 861         case RTE_INTR_HANDLE_UIO_INTX:
 862                 if (uio_intx_intr_disable(intr_handle))
 863                         rc = -1;
 864                 break;
 865         /* not used at this moment */
 866         case RTE_INTR_HANDLE_ALARM:
 867                 rc = -1;
 868                 break;
 869 #ifdef VFIO_PRESENT
 870         case RTE_INTR_HANDLE_VFIO_MSIX:
 871                 if (vfio_disable_msix(intr_handle))
 872                         rc = -1;
 873                 break;
 874         case RTE_INTR_HANDLE_VFIO_MSI:
 875                 if (vfio_disable_msi(intr_handle))
 876                         rc = -1;
 877                 break;
 878         case RTE_INTR_HANDLE_VFIO_LEGACY:
 879                 if (vfio_disable_intx(intr_handle))
 880                         rc = -1;
 881                 break;
 882 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
 883         case RTE_INTR_HANDLE_VFIO_REQ:
 884                 if (vfio_disable_req(intr_handle))
 885                         rc = -1;
 886                 break;
 887 #endif
 888 #endif
 889         /* not used at this moment */
 890         case RTE_INTR_HANDLE_DEV_EVENT:
 891                 rc = -1;
 892                 break;
 893         /* unknown handle type */
 894         default:
 895                 RTE_LOG(ERR, EAL, "Unknown handle type of fd %d\n",
 896                         rte_intr_fd_get(intr_handle));
 897                 rc = -1;
 898                 break;
 899         }
 900 out:
 901         rte_eal_trace_intr_disable(intr_handle, rc);
 902         return rc;
 903 }
 904
 905 static int
 906 eal_intr_process_interrupts(struct epoll_event *events, int nfds)
 907 {
 908         bool call = false;
 909         int n, bytes_read, rv;
 910         struct rte_intr_source *src;
 911         struct rte_intr_callback *cb, *next;
 912         union rte_intr_read_buffer buf;
 913         struct rte_intr_callback active_cb;
 914
 915         for (n = 0; n < nfds; n++) {
 916
 917                 /**
 918                  * if the pipe fd is ready to read, return out to
 919                  * rebuild the wait list.
 920                  */
 921                 if (events[n].data.fd == intr_pipe.readfd){
 922                         int r = read(intr_pipe.readfd, buf.charbuf,
 923                                         sizeof(buf.charbuf));
 924                         RTE_SET_USED(r);
 925                         return -1;
 926                 }
 927                 rte_spinlock_lock(&intr_lock);
 928                 TAILQ_FOREACH(src, &intr_sources, next)
 929                         if (rte_intr_fd_get(src->intr_handle) == events[n].data.fd)
 930                                 break;
 931                 if (src == NULL){
 932                         rte_spinlock_unlock(&intr_lock);
 933                         continue;
 934                 }
 935
 936                 /* mark this interrupt source as active and release the lock. */
 937                 src->active = 1;
 938                 rte_spinlock_unlock(&intr_lock);
 939
 940                 /* set the length to be read dor different handle type */
 941                 switch (rte_intr_type_get(src->intr_handle)) {
 942                 case RTE_INTR_HANDLE_UIO:
 943                 case RTE_INTR_HANDLE_UIO_INTX:
 944                         bytes_read = sizeof(buf.uio_intr_count);
 945                         break;
 946                 case RTE_INTR_HANDLE_ALARM:
 947                         bytes_read = sizeof(buf.timerfd_num);
 948                         break;
 949 #ifdef VFIO_PRESENT
 950                 case RTE_INTR_HANDLE_VFIO_MSIX:
 951                 case RTE_INTR_HANDLE_VFIO_MSI:
 952                 case RTE_INTR_HANDLE_VFIO_LEGACY:
 953                         bytes_read = sizeof(buf.vfio_intr_count);
 954                         break;
 955 #ifdef HAVE_VFIO_DEV_REQ_INTERFACE
 956                 case RTE_INTR_HANDLE_VFIO_REQ:
 957                         bytes_read = 0;
 958                         call = true;
 959                         break;
 960 #endif
 961 #endif
 962                 case RTE_INTR_HANDLE_VDEV:
 963                 case RTE_INTR_HANDLE_EXT:
 964                         bytes_read = 0;
 965                         call = true;
 966                         break;
 967                 case RTE_INTR_HANDLE_DEV_EVENT:
 968                         bytes_read = 0;
 969                         call = true;
 970                         break;
 971                 default:
 972                         bytes_read = 1;
 973                         break;
 974                 }
 975
 976                 if (bytes_read > 0) {
 977                         /**
 978                          * read out to clear the ready-to-be-read flag
 979                          * for epoll_wait.
 980                          */
 981                         bytes_read = read(events[n].data.fd, &buf, bytes_read);
 982                         if (bytes_read < 0) {
 983                                 if (errno == EINTR || errno == EWOULDBLOCK)
 984                                         continue;
 985
 986                                 RTE_LOG(ERR, EAL, "Error reading from file "
 987                                         "descriptor %d: %s\n",
 988                                         events[n].data.fd,
 989                                         strerror(errno));
 990                                 /*
 991                                  * The device is unplugged or buggy, remove
 992                                  * it as an interrupt source and return to
 993                                  * force the wait list to be rebuilt.
 994                                  */
 995                                 rte_spinlock_lock(&intr_lock);
 996                                 TAILQ_REMOVE(&intr_sources, src, next);
 997                                 rte_spinlock_unlock(&intr_lock);
 998
 999                                 for (cb = TAILQ_FIRST(&src->callbacks); cb;
1000                                                         cb = next) {
1001                                         next = TAILQ_NEXT(cb, next);
1002                                         TAILQ_REMOVE(&src->callbacks, cb, next);
1003                                         free(cb);
1004                                 }
1005                                 rte_intr_instance_free(src->intr_handle);
1006                                 free(src);
1007                                 return -1;
1008                         } else if (bytes_read == 0)
1009                                 RTE_LOG(ERR, EAL, "Read nothing from file "
1010                                         "descriptor %d\n", events[n].data.fd);
1011                         else
1012                                 call = true;
1013                 }
1014
1015                 /* grab a lock, again to call callbacks and update status. */
1016                 rte_spinlock_lock(&intr_lock);
1017
1018                 if (call) {
1019
1020                         /* Finally, call all callbacks. */
1021                         TAILQ_FOREACH(cb, &src->callbacks, next) {
1022
1023                                 /* make a copy and unlock. */
1024                                 active_cb = *cb;
1025                                 rte_spinlock_unlock(&intr_lock);
1026
1027                                 /* call the actual callback */
1028                                 active_cb.cb_fn(active_cb.cb_arg);
1029
1030                                 /*get the lock back. */
1031                                 rte_spinlock_lock(&intr_lock);
1032                         }
1033                 }
1034                 /* we done with that interrupt source, release it. */
1035                 src->active = 0;
1036
1037                 rv = 0;
1038
1039                 /* check if any callback are supposed to be removed */
1040                 for (cb = TAILQ_FIRST(&src->callbacks); cb != NULL; cb = next) {
1041                         next = TAILQ_NEXT(cb, next);
1042                         if (cb->pending_delete) {
1043                                 TAILQ_REMOVE(&src->callbacks, cb, next);
1044                                 if (cb->ucb_fn)
1045                                         cb->ucb_fn(src->intr_handle, cb->cb_arg);
1046                                 free(cb);
1047                                 rv++;
1048                         }
1049                 }
1050
1051                 /* all callbacks for that source are removed. */
1052                 if (TAILQ_EMPTY(&src->callbacks)) {
1053                         TAILQ_REMOVE(&intr_sources, src, next);
1054                         rte_intr_instance_free(src->intr_handle);
1055                         free(src);
1056                 }
1057
1058                 /* notify the pipe fd waited by epoll_wait to rebuild the wait list */
1059                 if (rv > 0 && write(intr_pipe.writefd, "1", 1) < 0) {
1060                         rte_spinlock_unlock(&intr_lock);
1061                         return -EPIPE;
1062                 }
1063
1064                 rte_spinlock_unlock(&intr_lock);
1065         }
1066
1067         return 0;
1068 }
1069
1070 /**
1071  * It handles all the interrupts.
1072  *
1073  * @param pfd
1074  *  epoll file descriptor.
1075  * @param totalfds
1076  *  The number of file descriptors added in epoll.
1077  *
1078  * @return
1079  *  void
1080  */
1081 static void
1082 eal_intr_handle_interrupts(int pfd, unsigned totalfds)
1083 {
1084         struct epoll_event events[totalfds];
1085         int nfds = 0;
1086
1087         for(;;) {
1088                 nfds = epoll_wait(pfd, events, totalfds,
1089                         EAL_INTR_EPOLL_WAIT_FOREVER);
1090                 /* epoll_wait fail */
1091                 if (nfds < 0) {
1092                         if (errno == EINTR)
1093                                 continue;
1094                         RTE_LOG(ERR, EAL,
1095                                 "epoll_wait returns with fail\n");
1096                         return;
1097                 }
1098                 /* epoll_wait timeout, will never happens here */
1099                 else if (nfds == 0)
1100                         continue;
1101                 /* epoll_wait has at least one fd ready to read */
1102                 if (eal_intr_process_interrupts(events, nfds) < 0)
1103                         return;
1104         }
1105 }
1106
1107 /**
1108  * It builds/rebuilds up the epoll file descriptor with all the
1109  * file descriptors being waited on. Then handles the interrupts.
1110  *
1111  * @param arg
1112  *  pointer. (unused)
1113  *
1114  * @return
1115  *  never return;
1116  */
1117 static __rte_noreturn void *
1118 eal_intr_thread_main(__rte_unused void *arg)
1119 {
1120         /* host thread, never break out */
1121         for (;;) {
1122                 /* build up the epoll fd with all descriptors we are to
1123                  * wait on then pass it to the handle_interrupts function
1124                  */
1125                 static struct epoll_event pipe_event = {
1126                         .events = EPOLLIN | EPOLLPRI,
1127                 };
1128                 struct rte_intr_source *src;
1129                 unsigned numfds = 0;
1130
1131                 /* create epoll fd */
1132                 int pfd = epoll_create(1);
1133                 if (pfd < 0)
1134                         rte_panic("Cannot create epoll instance\n");
1135
1136                 pipe_event.data.fd = intr_pipe.readfd;
1137                 /**
1138                  * add pipe fd into wait list, this pipe is used to
1139                  * rebuild the wait list.
1140                  */
1141                 if (epoll_ctl(pfd, EPOLL_CTL_ADD, intr_pipe.readfd,
1142                                                 &pipe_event) < 0) {
1143                         rte_panic("Error adding fd to %d epoll_ctl, %s\n",
1144                                         intr_pipe.readfd, strerror(errno));
1145                 }
1146                 numfds++;
1147
1148                 rte_spinlock_lock(&intr_lock);
1149
1150                 TAILQ_FOREACH(src, &intr_sources, next) {
1151                         struct epoll_event ev;
1152
1153                         if (src->callbacks.tqh_first == NULL)
1154                                 continue; /* skip those with no callbacks */
1155                         memset(&ev, 0, sizeof(ev));
1156                         ev.events = EPOLLIN | EPOLLPRI | EPOLLRDHUP | EPOLLHUP;
1157                         ev.data.fd = rte_intr_fd_get(src->intr_handle);
1158
1159                         /**
1160                          * add all the uio device file descriptor
1161                          * into wait list.
1162                          */
1163                         if (epoll_ctl(pfd, EPOLL_CTL_ADD,
1164                                         rte_intr_fd_get(src->intr_handle), &ev) < 0) {
1165                                 rte_panic("Error adding fd %d epoll_ctl, %s\n",
1166                                         rte_intr_fd_get(src->intr_handle),
1167                                         strerror(errno));
1168                         }
1169                         else
1170                                 numfds++;
1171                 }
1172                 rte_spinlock_unlock(&intr_lock);
1173                 /* serve the interrupt */
1174                 eal_intr_handle_interrupts(pfd, numfds);
1175
1176                 /**
1177                  * when we return, we need to rebuild the
1178                  * list of fds to monitor.
1179                  */
1180                 close(pfd);
1181         }
1182 }
1183
1184 int
1185 rte_eal_intr_init(void)
1186 {
1187         int ret = 0;
1188
1189         /* init the global interrupt source head */
1190         TAILQ_INIT(&intr_sources);
1191
1192         /**
1193          * create a pipe which will be waited by epoll and notified to
1194          * rebuild the wait list of epoll.
1195          */
1196         if (pipe(intr_pipe.pipefd) < 0) {
1197                 rte_errno = errno;
1198                 return -1;
1199         }
1200
1201         /* create the host thread to wait/handle the interrupt */
1202         ret = rte_ctrl_thread_create(&intr_thread, "eal-intr-thread", NULL,
1203                         eal_intr_thread_main, NULL);
1204         if (ret != 0) {
1205                 rte_errno = -ret;
1206                 RTE_LOG(ERR, EAL,
1207                         "Failed to create thread for interrupt handling\n");
1208         }
1209
1210         return ret;
1211 }
1212
1213 static void
1214 eal_intr_proc_rxtx_intr(int fd, const struct rte_intr_handle *intr_handle)
1215 {
1216         union rte_intr_read_buffer buf;
1217         int bytes_read = 0;
1218         int nbytes;
1219
1220         switch (rte_intr_type_get(intr_handle)) {
1221         case RTE_INTR_HANDLE_UIO:
1222         case RTE_INTR_HANDLE_UIO_INTX:
1223                 bytes_read = sizeof(buf.uio_intr_count);
1224                 break;
1225 #ifdef VFIO_PRESENT
1226         case RTE_INTR_HANDLE_VFIO_MSIX:
1227         case RTE_INTR_HANDLE_VFIO_MSI:
1228         case RTE_INTR_HANDLE_VFIO_LEGACY:
1229                 bytes_read = sizeof(buf.vfio_intr_count);
1230                 break;
1231 #endif
1232         case RTE_INTR_HANDLE_VDEV:
1233                 bytes_read = rte_intr_efd_counter_size_get(intr_handle);
1234                 /* For vdev, number of bytes to read is set by driver */
1235                 break;
1236         case RTE_INTR_HANDLE_EXT:
1237                 return;
1238         default:
1239                 bytes_read = 1;
1240                 RTE_LOG(INFO, EAL, "unexpected intr type\n");
1241                 break;
1242         }
1243
1244         /**
1245          * read out to clear the ready-to-be-read flag
1246          * for epoll_wait.
1247          */
1248         if (bytes_read == 0)
1249                 return;
1250         do {
1251                 nbytes = read(fd, &buf, bytes_read);
1252                 if (nbytes < 0) {
1253                         if (errno == EINTR || errno == EWOULDBLOCK ||
1254                             errno == EAGAIN)
1255                                 continue;
1256                         RTE_LOG(ERR, EAL,
1257                                 "Error reading from fd %d: %s\n",
1258                                 fd, strerror(errno));
1259                 } else if (nbytes == 0)
1260                         RTE_LOG(ERR, EAL, "Read nothing from fd %d\n", fd);
1261                 return;
1262         } while (1);
1263 }
1264
1265 static int
1266 eal_epoll_process_event(struct epoll_event *evs, unsigned int n,
1267                         struct rte_epoll_event *events)
1268 {
1269         unsigned int i, count = 0;
1270         struct rte_epoll_event *rev;
1271         uint32_t valid_status;
1272
1273         for (i = 0; i < n; i++) {
1274                 rev = evs[i].data.ptr;
1275                 valid_status =  RTE_EPOLL_VALID;
1276                 /* ACQUIRE memory ordering here pairs with RELEASE
1277                  * ordering below acting as a lock to synchronize
1278                  * the event data updating.
1279                  */
1280                 if (!rev || !__atomic_compare_exchange_n(&rev->status,
1281                                     &valid_status, RTE_EPOLL_EXEC, 0,
1282                                     __ATOMIC_ACQUIRE, __ATOMIC_RELAXED))
1283                         continue;
1284
1285                 events[count].status        = RTE_EPOLL_VALID;
1286                 events[count].fd            = rev->fd;
1287                 events[count].epfd          = rev->epfd;
1288                 events[count].epdata.event  = evs[i].events;
1289                 events[count].epdata.data   = rev->epdata.data;
1290                 if (rev->epdata.cb_fun)
1291                         rev->epdata.cb_fun(rev->fd,
1292                                            rev->epdata.cb_arg);
1293
1294                 /* the status update should be observed after
1295                  * the other fields change.
1296                  */
1297                 __atomic_store_n(&rev->status, RTE_EPOLL_VALID,
1298                                 __ATOMIC_RELEASE);
1299                 count++;
1300         }
1301         return count;
1302 }
1303
1304 static inline int
1305 eal_init_tls_epfd(void)
1306 {
1307         int pfd = epoll_create(255);
1308
1309         if (pfd < 0) {
1310                 RTE_LOG(ERR, EAL,
1311                         "Cannot create epoll instance\n");
1312                 return -1;
1313         }
1314         return pfd;
1315 }
1316
1317 int
1318 rte_intr_tls_epfd(void)
1319 {
1320         if (RTE_PER_LCORE(_epfd) == -1)
1321                 RTE_PER_LCORE(_epfd) = eal_init_tls_epfd();
1322
1323         return RTE_PER_LCORE(_epfd);
1324 }
1325
1326 static int
1327 eal_epoll_wait(int epfd, struct rte_epoll_event *events,
1328                int maxevents, int timeout, bool interruptible)
1329 {
1330         struct epoll_event evs[maxevents];
1331         int rc;
1332
1333         if (!events) {
1334                 RTE_LOG(ERR, EAL, "rte_epoll_event can't be NULL\n");
1335                 return -1;
1336         }
1337
1338         /* using per thread epoll fd */
1339         if (epfd == RTE_EPOLL_PER_THREAD)
1340                 epfd = rte_intr_tls_epfd();
1341
1342         while (1) {
1343                 rc = epoll_wait(epfd, evs, maxevents, timeout);
1344                 if (likely(rc > 0)) {
1345                         /* epoll_wait has at least one fd ready to read */
1346                         rc = eal_epoll_process_event(evs, rc, events);
1347                         break;
1348                 } else if (rc < 0) {
1349                         if (errno == EINTR) {
1350                                 if (interruptible)
1351                                         return -1;
1352                                 else
1353                                         continue;
1354                         }
1355                         /* epoll_wait fail */
1356                         RTE_LOG(ERR, EAL, "epoll_wait returns with fail %s\n",
1357                                 strerror(errno));
1358                         rc = -1;
1359                         break;
1360                 } else {
1361                         /* rc == 0, epoll_wait timed out */
1362                         break;
1363                 }
1364         }
1365
1366         return rc;
1367 }
1368
1369 int
1370 rte_epoll_wait(int epfd, struct rte_epoll_event *events,
1371                int maxevents, int timeout)
1372 {
1373         return eal_epoll_wait(epfd, events, maxevents, timeout, false);
1374 }
1375
1376 int
1377 rte_epoll_wait_interruptible(int epfd, struct rte_epoll_event *events,
1378                              int maxevents, int timeout)
1379 {
1380         return eal_epoll_wait(epfd, events, maxevents, timeout, true);
1381 }
1382
1383 static inline void
1384 eal_epoll_data_safe_free(struct rte_epoll_event *ev)
1385 {
1386         uint32_t valid_status = RTE_EPOLL_VALID;
1387
1388         while (!__atomic_compare_exchange_n(&ev->status, &valid_status,
1389                     RTE_EPOLL_INVALID, 0, __ATOMIC_ACQUIRE, __ATOMIC_RELAXED)) {
1390                 while (__atomic_load_n(&ev->status,
1391                                 __ATOMIC_RELAXED) != RTE_EPOLL_VALID)
1392                         rte_pause();
1393                 valid_status = RTE_EPOLL_VALID;
1394         }
1395         memset(&ev->epdata, 0, sizeof(ev->epdata));
1396         ev->fd = -1;
1397         ev->epfd = -1;
1398 }
1399
1400 int
1401 rte_epoll_ctl(int epfd, int op, int fd,
1402               struct rte_epoll_event *event)
1403 {
1404         struct epoll_event ev;
1405
1406         if (!event) {
1407                 RTE_LOG(ERR, EAL, "rte_epoll_event can't be NULL\n");
1408                 return -1;
1409         }
1410
1411         /* using per thread epoll fd */
1412         if (epfd == RTE_EPOLL_PER_THREAD)
1413                 epfd = rte_intr_tls_epfd();
1414
1415         if (op == EPOLL_CTL_ADD) {
1416                 __atomic_store_n(&event->status, RTE_EPOLL_VALID,
1417                                 __ATOMIC_RELAXED);
1418                 event->fd = fd;  /* ignore fd in event */
1419                 event->epfd = epfd;
1420                 ev.data.ptr = (void *)event;
1421         }
1422
1423         ev.events = event->epdata.event;
1424         if (epoll_ctl(epfd, op, fd, &ev) < 0) {
1425                 RTE_LOG(ERR, EAL, "Error op %d fd %d epoll_ctl, %s\n",
1426                         op, fd, strerror(errno));
1427                 if (op == EPOLL_CTL_ADD)
1428                         /* rollback status when CTL_ADD fail */
1429                         __atomic_store_n(&event->status, RTE_EPOLL_INVALID,
1430                                         __ATOMIC_RELAXED);
1431                 return -1;
1432         }
1433
1434         if (op == EPOLL_CTL_DEL && __atomic_load_n(&event->status,
1435                         __ATOMIC_RELAXED) != RTE_EPOLL_INVALID)
1436                 eal_epoll_data_safe_free(event);
1437
1438         return 0;
1439 }
1440
1441 int
1442 rte_intr_rx_ctl(struct rte_intr_handle *intr_handle, int epfd,
1443                 int op, unsigned int vec, void *data)
1444 {
1445         struct rte_epoll_event *rev;
1446         struct rte_epoll_data *epdata;
1447         int epfd_op;
1448         unsigned int efd_idx;
1449         int rc = 0;
1450
1451         efd_idx = (vec >= RTE_INTR_VEC_RXTX_OFFSET) ?
1452                 (vec - RTE_INTR_VEC_RXTX_OFFSET) : vec;
1453
1454         if (intr_handle == NULL || rte_intr_nb_efd_get(intr_handle) == 0 ||
1455                         efd_idx >= (unsigned int)rte_intr_nb_efd_get(intr_handle)) {
1456                 RTE_LOG(ERR, EAL, "Wrong intr vector number.\n");
1457                 return -EPERM;
1458         }
1459
1460         switch (op) {
1461         case RTE_INTR_EVENT_ADD:
1462                 epfd_op = EPOLL_CTL_ADD;
1463                 rev = rte_intr_elist_index_get(intr_handle, efd_idx);
1464                 if (__atomic_load_n(&rev->status,
1465                                 __ATOMIC_RELAXED) != RTE_EPOLL_INVALID) {
1466                         RTE_LOG(INFO, EAL, "Event already been added.\n");
1467                         return -EEXIST;
1468                 }
1469
1470                 /* attach to intr vector fd */
1471                 epdata = &rev->epdata;
1472                 epdata->event  = EPOLLIN | EPOLLPRI | EPOLLET;
1473                 epdata->data   = data;
1474                 epdata->cb_fun = (rte_intr_event_cb_t)eal_intr_proc_rxtx_intr;
1475                 epdata->cb_arg = (void *)intr_handle;
1476                 rc = rte_epoll_ctl(epfd, epfd_op,
1477                         rte_intr_efds_index_get(intr_handle, efd_idx), rev);
1478                 if (!rc)
1479                         RTE_LOG(DEBUG, EAL,
1480                                 "efd %d associated with vec %d added on epfd %d"
1481                                 "\n", rev->fd, vec, epfd);
1482                 else
1483                         rc = -EPERM;
1484                 break;
1485         case RTE_INTR_EVENT_DEL:
1486                 epfd_op = EPOLL_CTL_DEL;
1487                 rev = rte_intr_elist_index_get(intr_handle, efd_idx);
1488                 if (__atomic_load_n(&rev->status,
1489                                 __ATOMIC_RELAXED) == RTE_EPOLL_INVALID) {
1490                         RTE_LOG(INFO, EAL, "Event does not exist.\n");
1491                         return -EPERM;
1492                 }
1493
1494                 rc = rte_epoll_ctl(rev->epfd, epfd_op, rev->fd, rev);
1495                 if (rc)
1496                         rc = -EPERM;
1497                 break;
1498         default:
1499                 RTE_LOG(ERR, EAL, "event op type mismatch\n");
1500                 rc = -EPERM;
1501         }
1502
1503         return rc;
1504 }
1505
1506 void
1507 rte_intr_free_epoll_fd(struct rte_intr_handle *intr_handle)
1508 {
1509         uint32_t i;
1510         struct rte_epoll_event *rev;
1511
1512         for (i = 0; i < (uint32_t)rte_intr_nb_efd_get(intr_handle); i++) {
1513                 rev = rte_intr_elist_index_get(intr_handle, i);
1514                 if (__atomic_load_n(&rev->status,
1515                                 __ATOMIC_RELAXED) == RTE_EPOLL_INVALID)
1516                         continue;
1517                 if (rte_epoll_ctl(rev->epfd, EPOLL_CTL_DEL, rev->fd, rev)) {
1518                         /* force free if the entry valid */
1519                         eal_epoll_data_safe_free(rev);
1520                 }
1521         }
1522 }
1523
1524 int
1525 rte_intr_efd_enable(struct rte_intr_handle *intr_handle, uint32_t nb_efd)
1526 {
1527         uint32_t i;
1528         int fd;
1529         uint32_t n = RTE_MIN(nb_efd, (uint32_t)RTE_MAX_RXTX_INTR_VEC_ID);
1530
1531         assert(nb_efd != 0);
1532
1533         if (rte_intr_type_get(intr_handle) == RTE_INTR_HANDLE_VFIO_MSIX) {
1534                 for (i = 0; i < n; i++) {
1535                         fd = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC);
1536                         if (fd < 0) {
1537                                 RTE_LOG(ERR, EAL,
1538                                         "can't setup eventfd, error %i (%s)\n",
1539                                         errno, strerror(errno));
1540                                 return -errno;
1541                         }
1542
1543                         if (rte_intr_efds_index_set(intr_handle, i, fd))
1544                                 return -rte_errno;
1545                 }
1546
1547                 if (rte_intr_nb_efd_set(intr_handle, n))
1548                         return -rte_errno;
1549
1550                 if (rte_intr_max_intr_set(intr_handle, NB_OTHER_INTR + n))
1551                         return -rte_errno;
1552         } else if (rte_intr_type_get(intr_handle) == RTE_INTR_HANDLE_VDEV) {
1553                 /* only check, initialization would be done in vdev driver.*/
1554                 if ((uint64_t)rte_intr_efd_counter_size_get(intr_handle) >
1555                     sizeof(union rte_intr_read_buffer)) {
1556                         RTE_LOG(ERR, EAL, "the efd_counter_size is oversized");
1557                         return -EINVAL;
1558                 }
1559         } else {
1560                 if (rte_intr_efds_index_set(intr_handle, 0, rte_intr_fd_get(intr_handle)))
1561                         return -rte_errno;
1562                 if (rte_intr_nb_efd_set(intr_handle, RTE_MIN(nb_efd, 1U)))
1563                         return -rte_errno;
1564                 if (rte_intr_max_intr_set(intr_handle, NB_OTHER_INTR))
1565                         return -rte_errno;
1566         }
1567
1568         return 0;
1569 }
1570
1571 void
1572 rte_intr_efd_disable(struct rte_intr_handle *intr_handle)
1573 {
1574         uint32_t i;
1575
1576         rte_intr_free_epoll_fd(intr_handle);
1577         if (rte_intr_max_intr_get(intr_handle) > rte_intr_nb_efd_get(intr_handle)) {
1578                 for (i = 0; i < (uint32_t)rte_intr_nb_efd_get(intr_handle); i++)
1579                         close(rte_intr_efds_index_get(intr_handle, i));
1580         }
1581         rte_intr_nb_efd_set(intr_handle, 0);
1582         rte_intr_max_intr_set(intr_handle, 0);
1583 }
1584
1585 int
1586 rte_intr_dp_is_en(struct rte_intr_handle *intr_handle)
1587 {
1588         return !(!rte_intr_nb_efd_get(intr_handle));
1589 }
1590
1591 int
1592 rte_intr_allow_others(struct rte_intr_handle *intr_handle)
1593 {
1594         if (!rte_intr_dp_is_en(intr_handle))
1595                 return 1;
1596         else
1597                 return !!(rte_intr_max_intr_get(intr_handle) -
1598                                 rte_intr_nb_efd_get(intr_handle));
1599 }
1600
1601 int
1602 rte_intr_cap_multiple(struct rte_intr_handle *intr_handle)
1603 {
1604         if (rte_intr_type_get(intr_handle) == RTE_INTR_HANDLE_VFIO_MSIX)
1605                 return 1;
1606
1607         if (rte_intr_type_get(intr_handle) == RTE_INTR_HANDLE_VDEV)
1608                 return 1;
1609
1610         return 0;
1611 }
1612
1613 int rte_thread_is_intr(void)
1614 {
1615         return pthread_equal(intr_thread, pthread_self());
1616 }