drivers/net/ifc/ifcvf_vdpa.c

   1 /* SPDX-License-Identifier: BSD-3-Clause
   2  * Copyright(c) 2018 Intel Corporation
   3  */
   4
   5 #include <unistd.h>
   6 #include <pthread.h>
   7 #include <fcntl.h>
   8 #include <sys/ioctl.h>
   9 #include <sys/epoll.h>
  10 #include <linux/virtio_net.h>
  11 #include <stdbool.h>
  12
  13 #include <rte_malloc.h>
  14 #include <rte_memory.h>
  15 #include <rte_bus_pci.h>
  16 #include <rte_vhost.h>
  17 #include <rte_vdpa.h>
  18 #include <rte_vfio.h>
  19 #include <rte_spinlock.h>
  20 #include <rte_log.h>
  21 #include <rte_kvargs.h>
  22 #include <rte_devargs.h>
  23
  24 #include "base/ifcvf.h"
  25
  26 #define DRV_LOG(level, fmt, args...) \
  27         rte_log(RTE_LOG_ ## level, ifcvf_vdpa_logtype, \
  28                 "IFCVF %s(): " fmt "\n", __func__, ##args)
  29
  30 #ifndef PAGE_SIZE
  31 #define PAGE_SIZE 4096
  32 #endif
  33
  34 #define IFCVF_USED_RING_LEN(size) \
  35         ((size) * sizeof(struct vring_used_elem) + sizeof(uint16_t) * 3)
  36
  37 #define IFCVF_VDPA_MODE         "vdpa"
  38 #define IFCVF_SW_FALLBACK_LM    "sw-live-migration"
  39
  40 static const char * const ifcvf_valid_arguments[] = {
  41         IFCVF_VDPA_MODE,
  42         IFCVF_SW_FALLBACK_LM,
  43         NULL
  44 };
  45
  46 static int ifcvf_vdpa_logtype;
  47
  48 struct ifcvf_internal {
  49         struct rte_vdpa_dev_addr dev_addr;
  50         struct rte_pci_device *pdev;
  51         struct ifcvf_hw hw;
  52         int vfio_container_fd;
  53         int vfio_group_fd;
  54         int vfio_dev_fd;
  55         pthread_t tid;  /* thread for notify relay */
  56         int epfd;
  57         int vid;
  58         int did;
  59         uint16_t max_queues;
  60         uint64_t features;
  61         rte_atomic32_t started;
  62         rte_atomic32_t dev_attached;
  63         rte_atomic32_t running;
  64         rte_spinlock_t lock;
  65         bool sw_lm;
  66         bool sw_fallback_running;
  67         /* mediated vring for sw fallback */
  68         struct vring m_vring[IFCVF_MAX_QUEUES * 2];
  69 };
  70
  71 struct internal_list {
  72         TAILQ_ENTRY(internal_list) next;
  73         struct ifcvf_internal *internal;
  74 };
  75
  76 TAILQ_HEAD(internal_list_head, internal_list);
  77 static struct internal_list_head internal_list =
  78         TAILQ_HEAD_INITIALIZER(internal_list);
  79
  80 static pthread_mutex_t internal_list_lock = PTHREAD_MUTEX_INITIALIZER;
  81
  82 static struct internal_list *
  83 find_internal_resource_by_did(int did)
  84 {
  85         int found = 0;
  86         struct internal_list *list;
  87
  88         pthread_mutex_lock(&internal_list_lock);
  89
  90         TAILQ_FOREACH(list, &internal_list, next) {
  91                 if (did == list->internal->did) {
  92                         found = 1;
  93                         break;
  94                 }
  95         }
  96
  97         pthread_mutex_unlock(&internal_list_lock);
  98
  99         if (!found)
 100                 return NULL;
 101
 102         return list;
 103 }
 104
 105 static struct internal_list *
 106 find_internal_resource_by_dev(struct rte_pci_device *pdev)
 107 {
 108         int found = 0;
 109         struct internal_list *list;
 110
 111         pthread_mutex_lock(&internal_list_lock);
 112
 113         TAILQ_FOREACH(list, &internal_list, next) {
 114                 if (pdev == list->internal->pdev) {
 115                         found = 1;
 116                         break;
 117                 }
 118         }
 119
 120         pthread_mutex_unlock(&internal_list_lock);
 121
 122         if (!found)
 123                 return NULL;
 124
 125         return list;
 126 }
 127
 128 static int
 129 ifcvf_vfio_setup(struct ifcvf_internal *internal)
 130 {
 131         struct rte_pci_device *dev = internal->pdev;
 132         char devname[RTE_DEV_NAME_MAX_LEN] = {0};
 133         int iommu_group_num;
 134         int i;
 135
 136         internal->vfio_dev_fd = -1;
 137         internal->vfio_group_fd = -1;
 138         internal->vfio_container_fd = -1;
 139
 140         rte_pci_device_name(&dev->addr, devname, RTE_DEV_NAME_MAX_LEN);
 141         rte_vfio_get_group_num(rte_pci_get_sysfs_path(), devname,
 142                         &iommu_group_num);
 143
 144         internal->vfio_container_fd = rte_vfio_container_create();
 145         if (internal->vfio_container_fd < 0)
 146                 return -1;
 147
 148         internal->vfio_group_fd = rte_vfio_container_group_bind(
 149                         internal->vfio_container_fd, iommu_group_num);
 150         if (internal->vfio_group_fd < 0)
 151                 goto err;
 152
 153         if (rte_pci_map_device(dev))
 154                 goto err;
 155
 156         internal->vfio_dev_fd = dev->intr_handle.vfio_dev_fd;
 157
 158         for (i = 0; i < RTE_MIN(PCI_MAX_RESOURCE, IFCVF_PCI_MAX_RESOURCE);
 159                         i++) {
 160                 internal->hw.mem_resource[i].addr =
 161                         internal->pdev->mem_resource[i].addr;
 162                 internal->hw.mem_resource[i].phys_addr =
 163                         internal->pdev->mem_resource[i].phys_addr;
 164                 internal->hw.mem_resource[i].len =
 165                         internal->pdev->mem_resource[i].len;
 166         }
 167
 168         return 0;
 169
 170 err:
 171         rte_vfio_container_destroy(internal->vfio_container_fd);
 172         return -1;
 173 }
 174
 175 static int
 176 ifcvf_dma_map(struct ifcvf_internal *internal, int do_map)
 177 {
 178         uint32_t i;
 179         int ret;
 180         struct rte_vhost_memory *mem = NULL;
 181         int vfio_container_fd;
 182
 183         ret = rte_vhost_get_mem_table(internal->vid, &mem);
 184         if (ret < 0) {
 185                 DRV_LOG(ERR, "failed to get VM memory layout.");
 186                 goto exit;
 187         }
 188
 189         vfio_container_fd = internal->vfio_container_fd;
 190
 191         for (i = 0; i < mem->nregions; i++) {
 192                 struct rte_vhost_mem_region *reg;
 193
 194                 reg = &mem->regions[i];
 195                 DRV_LOG(INFO, "%s, region %u: HVA 0x%" PRIx64 ", "
 196                         "GPA 0x%" PRIx64 ", size 0x%" PRIx64 ".",
 197                         do_map ? "DMA map" : "DMA unmap", i,
 198                         reg->host_user_addr, reg->guest_phys_addr, reg->size);
 199
 200                 if (do_map) {
 201                         ret = rte_vfio_container_dma_map(vfio_container_fd,
 202                                 reg->host_user_addr, reg->guest_phys_addr,
 203                                 reg->size);
 204                         if (ret < 0) {
 205                                 DRV_LOG(ERR, "DMA map failed.");
 206                                 goto exit;
 207                         }
 208                 } else {
 209                         ret = rte_vfio_container_dma_unmap(vfio_container_fd,
 210                                 reg->host_user_addr, reg->guest_phys_addr,
 211                                 reg->size);
 212                         if (ret < 0) {
 213                                 DRV_LOG(ERR, "DMA unmap failed.");
 214                                 goto exit;
 215                         }
 216                 }
 217         }
 218
 219 exit:
 220         if (mem)
 221                 free(mem);
 222         return ret;
 223 }
 224
 225 static uint64_t
 226 hva_to_gpa(int vid, uint64_t hva)
 227 {
 228         struct rte_vhost_memory *mem = NULL;
 229         struct rte_vhost_mem_region *reg;
 230         uint32_t i;
 231         uint64_t gpa = 0;
 232
 233         if (rte_vhost_get_mem_table(vid, &mem) < 0)
 234                 goto exit;
 235
 236         for (i = 0; i < mem->nregions; i++) {
 237                 reg = &mem->regions[i];
 238
 239                 if (hva >= reg->host_user_addr &&
 240                                 hva < reg->host_user_addr + reg->size) {
 241                         gpa = hva - reg->host_user_addr + reg->guest_phys_addr;
 242                         break;
 243                 }
 244         }
 245
 246 exit:
 247         if (mem)
 248                 free(mem);
 249         return gpa;
 250 }
 251
 252 static int
 253 vdpa_ifcvf_start(struct ifcvf_internal *internal)
 254 {
 255         struct ifcvf_hw *hw = &internal->hw;
 256         int i, nr_vring;
 257         int vid;
 258         struct rte_vhost_vring vq;
 259         uint64_t gpa;
 260
 261         vid = internal->vid;
 262         nr_vring = rte_vhost_get_vring_num(vid);
 263         rte_vhost_get_negotiated_features(vid, &hw->req_features);
 264
 265         for (i = 0; i < nr_vring; i++) {
 266                 rte_vhost_get_vhost_vring(vid, i, &vq);
 267                 gpa = hva_to_gpa(vid, (uint64_t)(uintptr_t)vq.desc);
 268                 if (gpa == 0) {
 269                         DRV_LOG(ERR, "Fail to get GPA for descriptor ring.");
 270                         return -1;
 271                 }
 272                 hw->vring[i].desc = gpa;
 273
 274                 gpa = hva_to_gpa(vid, (uint64_t)(uintptr_t)vq.avail);
 275                 if (gpa == 0) {
 276                         DRV_LOG(ERR, "Fail to get GPA for available ring.");
 277                         return -1;
 278                 }
 279                 hw->vring[i].avail = gpa;
 280
 281                 gpa = hva_to_gpa(vid, (uint64_t)(uintptr_t)vq.used);
 282                 if (gpa == 0) {
 283                         DRV_LOG(ERR, "Fail to get GPA for used ring.");
 284                         return -1;
 285                 }
 286                 hw->vring[i].used = gpa;
 287
 288                 hw->vring[i].size = vq.size;
 289                 rte_vhost_get_vring_base(vid, i, &hw->vring[i].last_avail_idx,
 290                                 &hw->vring[i].last_used_idx);
 291         }
 292         hw->nr_vring = i;
 293
 294         return ifcvf_start_hw(&internal->hw);
 295 }
 296
 297 static void
 298 vdpa_ifcvf_stop(struct ifcvf_internal *internal)
 299 {
 300         struct ifcvf_hw *hw = &internal->hw;
 301         uint32_t i;
 302         int vid;
 303         uint64_t features;
 304         uint64_t log_base, log_size;
 305         uint64_t len;
 306
 307         vid = internal->vid;
 308         ifcvf_stop_hw(hw);
 309
 310         for (i = 0; i < hw->nr_vring; i++)
 311                 rte_vhost_set_vring_base(vid, i, hw->vring[i].last_avail_idx,
 312                                 hw->vring[i].last_used_idx);
 313
 314         if (internal->sw_lm)
 315                 return;
 316
 317         rte_vhost_get_negotiated_features(vid, &features);
 318         if (RTE_VHOST_NEED_LOG(features)) {
 319                 ifcvf_disable_logging(hw);
 320                 rte_vhost_get_log_base(internal->vid, &log_base, &log_size);
 321                 rte_vfio_container_dma_unmap(internal->vfio_container_fd,
 322                                 log_base, IFCVF_LOG_BASE, log_size);
 323                 /*
 324                  * IFCVF marks dirty memory pages for only packet buffer,
 325                  * SW helps to mark the used ring as dirty after device stops.
 326                  */
 327                 for (i = 0; i < hw->nr_vring; i++) {
 328                         len = IFCVF_USED_RING_LEN(hw->vring[i].size);
 329                         rte_vhost_log_used_vring(vid, i, 0, len);
 330                 }
 331         }
 332 }
 333
 334 #define MSIX_IRQ_SET_BUF_LEN (sizeof(struct vfio_irq_set) + \
 335                 sizeof(int) * (IFCVF_MAX_QUEUES * 2 + 1))
 336 static int
 337 vdpa_enable_vfio_intr(struct ifcvf_internal *internal)
 338 {
 339         int ret;
 340         uint32_t i, nr_vring;
 341         char irq_set_buf[MSIX_IRQ_SET_BUF_LEN];
 342         struct vfio_irq_set *irq_set;
 343         int *fd_ptr;
 344         struct rte_vhost_vring vring;
 345
 346         nr_vring = rte_vhost_get_vring_num(internal->vid);
 347
 348         irq_set = (struct vfio_irq_set *)irq_set_buf;
 349         irq_set->argsz = sizeof(irq_set_buf);
 350         irq_set->count = nr_vring + 1;
 351         irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
 352                          VFIO_IRQ_SET_ACTION_TRIGGER;
 353         irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
 354         irq_set->start = 0;
 355         fd_ptr = (int *)&irq_set->data;
 356         fd_ptr[RTE_INTR_VEC_ZERO_OFFSET] = internal->pdev->intr_handle.fd;
 357
 358         for (i = 0; i < nr_vring; i++) {
 359                 rte_vhost_get_vhost_vring(internal->vid, i, &vring);
 360                 fd_ptr[RTE_INTR_VEC_RXTX_OFFSET + i] = vring.callfd;
 361         }
 362
 363         ret = ioctl(internal->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 364         if (ret) {
 365                 DRV_LOG(ERR, "Error enabling MSI-X interrupts: %s",
 366                                 strerror(errno));
 367                 return -1;
 368         }
 369
 370         return 0;
 371 }
 372
 373 static int
 374 vdpa_disable_vfio_intr(struct ifcvf_internal *internal)
 375 {
 376         int ret;
 377         char irq_set_buf[MSIX_IRQ_SET_BUF_LEN];
 378         struct vfio_irq_set *irq_set;
 379
 380         irq_set = (struct vfio_irq_set *)irq_set_buf;
 381         irq_set->argsz = sizeof(irq_set_buf);
 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_MSIX_IRQ_INDEX;
 385         irq_set->start = 0;
 386
 387         ret = ioctl(internal->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
 388         if (ret) {
 389                 DRV_LOG(ERR, "Error disabling MSI-X interrupts: %s",
 390                                 strerror(errno));
 391                 return -1;
 392         }
 393
 394         return 0;
 395 }
 396
 397 static void *
 398 notify_relay(void *arg)
 399 {
 400         int i, kickfd, epfd, nfds = 0;
 401         uint32_t qid, q_num;
 402         struct epoll_event events[IFCVF_MAX_QUEUES * 2];
 403         struct epoll_event ev;
 404         uint64_t buf;
 405         int nbytes;
 406         struct rte_vhost_vring vring;
 407         struct ifcvf_internal *internal = (struct ifcvf_internal *)arg;
 408         struct ifcvf_hw *hw = &internal->hw;
 409
 410         q_num = rte_vhost_get_vring_num(internal->vid);
 411
 412         epfd = epoll_create(IFCVF_MAX_QUEUES * 2);
 413         if (epfd < 0) {
 414                 DRV_LOG(ERR, "failed to create epoll instance.");
 415                 return NULL;
 416         }
 417         internal->epfd = epfd;
 418
 419         for (qid = 0; qid < q_num; qid++) {
 420                 ev.events = EPOLLIN | EPOLLPRI;
 421                 rte_vhost_get_vhost_vring(internal->vid, qid, &vring);
 422                 ev.data.u64 = qid | (uint64_t)vring.kickfd << 32;
 423                 if (epoll_ctl(epfd, EPOLL_CTL_ADD, vring.kickfd, &ev) < 0) {
 424                         DRV_LOG(ERR, "epoll add error: %s", strerror(errno));
 425                         return NULL;
 426                 }
 427         }
 428
 429         for (;;) {
 430                 nfds = epoll_wait(epfd, events, q_num, -1);
 431                 if (nfds < 0) {
 432                         if (errno == EINTR)
 433                                 continue;
 434                         DRV_LOG(ERR, "epoll_wait return fail\n");
 435                         return NULL;
 436                 }
 437
 438                 for (i = 0; i < nfds; i++) {
 439                         qid = events[i].data.u32;
 440                         kickfd = (uint32_t)(events[i].data.u64 >> 32);
 441                         do {
 442                                 nbytes = read(kickfd, &buf, 8);
 443                                 if (nbytes < 0) {
 444                                         if (errno == EINTR ||
 445                                             errno == EWOULDBLOCK ||
 446                                             errno == EAGAIN)
 447                                                 continue;
 448                                         DRV_LOG(INFO, "Error reading "
 449                                                 "kickfd: %s",
 450                                                 strerror(errno));
 451                                 }
 452                                 break;
 453                         } while (1);
 454
 455                         ifcvf_notify_queue(hw, qid);
 456                 }
 457         }
 458
 459         return NULL;
 460 }
 461
 462 static int
 463 setup_notify_relay(struct ifcvf_internal *internal)
 464 {
 465         int ret;
 466
 467         ret = pthread_create(&internal->tid, NULL, notify_relay,
 468                         (void *)internal);
 469         if (ret) {
 470                 DRV_LOG(ERR, "failed to create notify relay pthread.");
 471                 return -1;
 472         }
 473         return 0;
 474 }
 475
 476 static int
 477 unset_notify_relay(struct ifcvf_internal *internal)
 478 {
 479         void *status;
 480
 481         if (internal->tid) {
 482                 pthread_cancel(internal->tid);
 483                 pthread_join(internal->tid, &status);
 484         }
 485         internal->tid = 0;
 486
 487         if (internal->epfd >= 0)
 488                 close(internal->epfd);
 489         internal->epfd = -1;
 490
 491         return 0;
 492 }
 493
 494 static int
 495 update_datapath(struct ifcvf_internal *internal)
 496 {
 497         int ret;
 498
 499         rte_spinlock_lock(&internal->lock);
 500
 501         if (!rte_atomic32_read(&internal->running) &&
 502             (rte_atomic32_read(&internal->started) &&
 503              rte_atomic32_read(&internal->dev_attached))) {
 504                 ret = ifcvf_dma_map(internal, 1);
 505                 if (ret)
 506                         goto err;
 507
 508                 ret = vdpa_enable_vfio_intr(internal);
 509                 if (ret)
 510                         goto err;
 511
 512                 ret = vdpa_ifcvf_start(internal);
 513                 if (ret)
 514                         goto err;
 515
 516                 ret = setup_notify_relay(internal);
 517                 if (ret)
 518                         goto err;
 519
 520                 rte_atomic32_set(&internal->running, 1);
 521         } else if (rte_atomic32_read(&internal->running) &&
 522                    (!rte_atomic32_read(&internal->started) ||
 523                     !rte_atomic32_read(&internal->dev_attached))) {
 524                 ret = unset_notify_relay(internal);
 525                 if (ret)
 526                         goto err;
 527
 528                 vdpa_ifcvf_stop(internal);
 529
 530                 ret = vdpa_disable_vfio_intr(internal);
 531                 if (ret)
 532                         goto err;
 533
 534                 ret = ifcvf_dma_map(internal, 0);
 535                 if (ret)
 536                         goto err;
 537
 538                 rte_atomic32_set(&internal->running, 0);
 539         }
 540
 541         rte_spinlock_unlock(&internal->lock);
 542         return 0;
 543 err:
 544         rte_spinlock_unlock(&internal->lock);
 545         return ret;
 546 }
 547
 548 static int
 549 m_ifcvf_start(struct ifcvf_internal *internal)
 550 {
 551         struct ifcvf_hw *hw = &internal->hw;
 552         uint32_t i, nr_vring;
 553         int vid, ret;
 554         struct rte_vhost_vring vq;
 555         void *vring_buf;
 556         uint64_t m_vring_iova = IFCVF_MEDIATED_VRING;
 557         uint64_t size;
 558         uint64_t gpa;
 559
 560         vid = internal->vid;
 561         nr_vring = rte_vhost_get_vring_num(vid);
 562         rte_vhost_get_negotiated_features(vid, &hw->req_features);
 563
 564         for (i = 0; i < nr_vring; i++) {
 565                 rte_vhost_get_vhost_vring(vid, i, &vq);
 566
 567                 size = RTE_ALIGN_CEIL(vring_size(vq.size, PAGE_SIZE),
 568                                 PAGE_SIZE);
 569                 vring_buf = rte_zmalloc("ifcvf", size, PAGE_SIZE);
 570                 vring_init(&internal->m_vring[i], vq.size, vring_buf,
 571                                 PAGE_SIZE);
 572
 573                 ret = rte_vfio_container_dma_map(internal->vfio_container_fd,
 574                         (uint64_t)(uintptr_t)vring_buf, m_vring_iova, size);
 575                 if (ret < 0) {
 576                         DRV_LOG(ERR, "mediated vring DMA map failed.");
 577                         goto error;
 578                 }
 579
 580                 gpa = hva_to_gpa(vid, (uint64_t)(uintptr_t)vq.desc);
 581                 if (gpa == 0) {
 582                         DRV_LOG(ERR, "Fail to get GPA for descriptor ring.");
 583                         return -1;
 584                 }
 585                 hw->vring[i].desc = gpa;
 586
 587                 hw->vring[i].avail = m_vring_iova +
 588                         (char *)internal->m_vring[i].avail -
 589                         (char *)internal->m_vring[i].desc;
 590
 591                 hw->vring[i].used = m_vring_iova +
 592                         (char *)internal->m_vring[i].used -
 593                         (char *)internal->m_vring[i].desc;
 594
 595                 hw->vring[i].size = vq.size;
 596
 597                 rte_vhost_get_vring_base(vid, i, &hw->vring[i].last_avail_idx,
 598                                 &hw->vring[i].last_used_idx);
 599
 600                 m_vring_iova += size;
 601         }
 602         hw->nr_vring = nr_vring;
 603
 604         return ifcvf_start_hw(&internal->hw);
 605
 606 error:
 607         for (i = 0; i < nr_vring; i++)
 608                 if (internal->m_vring[i].desc)
 609                         rte_free(internal->m_vring[i].desc);
 610
 611         return -1;
 612 }
 613
 614 static int
 615 m_ifcvf_stop(struct ifcvf_internal *internal)
 616 {
 617         int vid;
 618         uint32_t i;
 619         struct rte_vhost_vring vq;
 620         struct ifcvf_hw *hw = &internal->hw;
 621         uint64_t m_vring_iova = IFCVF_MEDIATED_VRING;
 622         uint64_t size, len;
 623
 624         vid = internal->vid;
 625         ifcvf_stop_hw(hw);
 626
 627         for (i = 0; i < hw->nr_vring; i++) {
 628                 rte_vhost_get_vhost_vring(vid, i, &vq);
 629                 len = IFCVF_USED_RING_LEN(vq.size);
 630                 rte_vhost_log_used_vring(vid, i, 0, len);
 631
 632                 size = RTE_ALIGN_CEIL(vring_size(vq.size, PAGE_SIZE),
 633                                 PAGE_SIZE);
 634                 rte_vfio_container_dma_unmap(internal->vfio_container_fd,
 635                         (uint64_t)(uintptr_t)internal->m_vring[i].desc,
 636                         m_vring_iova, size);
 637
 638                 rte_vhost_set_vring_base(vid, i, hw->vring[i].last_avail_idx,
 639                                 hw->vring[i].last_used_idx);
 640                 rte_free(internal->m_vring[i].desc);
 641                 m_vring_iova += size;
 642         }
 643
 644         return 0;
 645 }
 646
 647 static int
 648 m_enable_vfio_intr(struct ifcvf_internal *internal)
 649 {
 650         uint32_t nr_vring;
 651         struct rte_intr_handle *intr_handle = &internal->pdev->intr_handle;
 652         int ret;
 653
 654         nr_vring = rte_vhost_get_vring_num(internal->vid);
 655
 656         ret = rte_intr_efd_enable(intr_handle, nr_vring);
 657         if (ret)
 658                 return -1;
 659
 660         ret = rte_intr_enable(intr_handle);
 661         if (ret)
 662                 return -1;
 663
 664         return 0;
 665 }
 666
 667 static void
 668 m_disable_vfio_intr(struct ifcvf_internal *internal)
 669 {
 670         struct rte_intr_handle *intr_handle = &internal->pdev->intr_handle;
 671
 672         rte_intr_efd_disable(intr_handle);
 673         rte_intr_disable(intr_handle);
 674 }
 675
 676 static void
 677 update_avail_ring(struct ifcvf_internal *internal, uint16_t qid)
 678 {
 679         rte_vdpa_relay_vring_avail(internal->vid, qid, &internal->m_vring[qid]);
 680         ifcvf_notify_queue(&internal->hw, qid);
 681 }
 682
 683 static void
 684 update_used_ring(struct ifcvf_internal *internal, uint16_t qid)
 685 {
 686         rte_vdpa_relay_vring_used(internal->vid, qid, &internal->m_vring[qid]);
 687         rte_vhost_vring_call(internal->vid, qid);
 688 }
 689
 690 static void *
 691 vring_relay(void *arg)
 692 {
 693         int i, vid, epfd, fd, nfds;
 694         struct ifcvf_internal *internal = (struct ifcvf_internal *)arg;
 695         struct rte_vhost_vring vring;
 696         struct rte_intr_handle *intr_handle;
 697         uint16_t qid, q_num;
 698         struct epoll_event events[IFCVF_MAX_QUEUES * 4];
 699         struct epoll_event ev;
 700         int nbytes;
 701         uint64_t buf;
 702
 703         vid = internal->vid;
 704         q_num = rte_vhost_get_vring_num(vid);
 705         /* prepare the mediated vring */
 706         for (qid = 0; qid < q_num; qid++) {
 707                 rte_vhost_get_vring_base(vid, qid,
 708                                 &internal->m_vring[qid].avail->idx,
 709                                 &internal->m_vring[qid].used->idx);
 710                 rte_vdpa_relay_vring_avail(vid, qid, &internal->m_vring[qid]);
 711         }
 712
 713         /* add notify fd and interrupt fd to epoll */
 714         epfd = epoll_create(IFCVF_MAX_QUEUES * 2);
 715         if (epfd < 0) {
 716                 DRV_LOG(ERR, "failed to create epoll instance.");
 717                 return NULL;
 718         }
 719         internal->epfd = epfd;
 720
 721         for (qid = 0; qid < q_num; qid++) {
 722                 ev.events = EPOLLIN | EPOLLPRI;
 723                 rte_vhost_get_vhost_vring(vid, qid, &vring);
 724                 ev.data.u64 = qid << 1 | (uint64_t)vring.kickfd << 32;
 725                 if (epoll_ctl(epfd, EPOLL_CTL_ADD, vring.kickfd, &ev) < 0) {
 726                         DRV_LOG(ERR, "epoll add error: %s", strerror(errno));
 727                         return NULL;
 728                 }
 729         }
 730
 731         intr_handle = &internal->pdev->intr_handle;
 732         for (qid = 0; qid < q_num; qid++) {
 733                 ev.events = EPOLLIN | EPOLLPRI;
 734                 ev.data.u64 = 1 | qid << 1 |
 735                         (uint64_t)intr_handle->efds[qid] << 32;
 736                 if (epoll_ctl(epfd, EPOLL_CTL_ADD, intr_handle->efds[qid], &ev)
 737                                 < 0) {
 738                         DRV_LOG(ERR, "epoll add error: %s", strerror(errno));
 739                         return NULL;
 740                 }
 741         }
 742
 743         /* start relay with a first kick */
 744         for (qid = 0; qid < q_num; qid++)
 745                 ifcvf_notify_queue(&internal->hw, qid);
 746
 747         /* listen to the events and react accordingly */
 748         for (;;) {
 749                 nfds = epoll_wait(epfd, events, q_num * 2, -1);
 750                 if (nfds < 0) {
 751                         if (errno == EINTR)
 752                                 continue;
 753                         DRV_LOG(ERR, "epoll_wait return fail\n");
 754                         return NULL;
 755                 }
 756
 757                 for (i = 0; i < nfds; i++) {
 758                         fd = (uint32_t)(events[i].data.u64 >> 32);
 759                         do {
 760                                 nbytes = read(fd, &buf, 8);
 761                                 if (nbytes < 0) {
 762                                         if (errno == EINTR ||
 763                                             errno == EWOULDBLOCK ||
 764                                             errno == EAGAIN)
 765                                                 continue;
 766                                         DRV_LOG(INFO, "Error reading "
 767                                                 "kickfd: %s",
 768                                                 strerror(errno));
 769                                 }
 770                                 break;
 771                         } while (1);
 772
 773                         qid = events[i].data.u32 >> 1;
 774
 775                         if (events[i].data.u32 & 1)
 776                                 update_used_ring(internal, qid);
 777                         else
 778                                 update_avail_ring(internal, qid);
 779                 }
 780         }
 781
 782         return NULL;
 783 }
 784
 785 static int
 786 setup_vring_relay(struct ifcvf_internal *internal)
 787 {
 788         int ret;
 789
 790         ret = pthread_create(&internal->tid, NULL, vring_relay,
 791                         (void *)internal);
 792         if (ret) {
 793                 DRV_LOG(ERR, "failed to create ring relay pthread.");
 794                 return -1;
 795         }
 796         return 0;
 797 }
 798
 799 static int
 800 unset_vring_relay(struct ifcvf_internal *internal)
 801 {
 802         void *status;
 803
 804         if (internal->tid) {
 805                 pthread_cancel(internal->tid);
 806                 pthread_join(internal->tid, &status);
 807         }
 808         internal->tid = 0;
 809
 810         if (internal->epfd >= 0)
 811                 close(internal->epfd);
 812         internal->epfd = -1;
 813
 814         return 0;
 815 }
 816
 817 static int
 818 ifcvf_sw_fallback_switchover(struct ifcvf_internal *internal)
 819 {
 820         int ret;
 821
 822         /* stop the direct IO data path */
 823         unset_notify_relay(internal);
 824         vdpa_ifcvf_stop(internal);
 825         vdpa_disable_vfio_intr(internal);
 826
 827         ret = rte_vhost_host_notifier_ctrl(internal->vid, false);
 828         if (ret && ret != -ENOTSUP)
 829                 goto error;
 830
 831         /* set up interrupt for interrupt relay */
 832         ret = m_enable_vfio_intr(internal);
 833         if (ret)
 834                 goto unmap;
 835
 836         /* config the VF */
 837         ret = m_ifcvf_start(internal);
 838         if (ret)
 839                 goto unset_intr;
 840
 841         /* set up vring relay thread */
 842         ret = setup_vring_relay(internal);
 843         if (ret)
 844                 goto stop_vf;
 845
 846         internal->sw_fallback_running = true;
 847
 848         return 0;
 849
 850 stop_vf:
 851         m_ifcvf_stop(internal);
 852 unset_intr:
 853         m_disable_vfio_intr(internal);
 854 unmap:
 855         ifcvf_dma_map(internal, 0);
 856 error:
 857         return -1;
 858 }
 859
 860 static int
 861 ifcvf_dev_config(int vid)
 862 {
 863         int did;
 864         struct internal_list *list;
 865         struct ifcvf_internal *internal;
 866
 867         did = rte_vhost_get_vdpa_device_id(vid);
 868         list = find_internal_resource_by_did(did);
 869         if (list == NULL) {
 870                 DRV_LOG(ERR, "Invalid device id: %d", did);
 871                 return -1;
 872         }
 873
 874         internal = list->internal;
 875         internal->vid = vid;
 876         rte_atomic32_set(&internal->dev_attached, 1);
 877         update_datapath(internal);
 878
 879         if (rte_vhost_host_notifier_ctrl(vid, true) != 0)
 880                 DRV_LOG(NOTICE, "vDPA (%d): software relay is used.", did);
 881
 882         return 0;
 883 }
 884
 885 static int
 886 ifcvf_dev_close(int vid)
 887 {
 888         int did;
 889         struct internal_list *list;
 890         struct ifcvf_internal *internal;
 891
 892         did = rte_vhost_get_vdpa_device_id(vid);
 893         list = find_internal_resource_by_did(did);
 894         if (list == NULL) {
 895                 DRV_LOG(ERR, "Invalid device id: %d", did);
 896                 return -1;
 897         }
 898
 899         internal = list->internal;
 900
 901         if (internal->sw_fallback_running) {
 902                 /* unset ring relay */
 903                 unset_vring_relay(internal);
 904
 905                 /* reset VF */
 906                 m_ifcvf_stop(internal);
 907
 908                 /* remove interrupt setting */
 909                 m_disable_vfio_intr(internal);
 910
 911                 /* unset DMA map for guest memory */
 912                 ifcvf_dma_map(internal, 0);
 913
 914                 internal->sw_fallback_running = false;
 915         } else {
 916                 rte_atomic32_set(&internal->dev_attached, 0);
 917                 update_datapath(internal);
 918         }
 919
 920         return 0;
 921 }
 922
 923 static int
 924 ifcvf_set_features(int vid)
 925 {
 926         uint64_t features;
 927         int did;
 928         struct internal_list *list;
 929         struct ifcvf_internal *internal;
 930         uint64_t log_base, log_size;
 931
 932         did = rte_vhost_get_vdpa_device_id(vid);
 933         list = find_internal_resource_by_did(did);
 934         if (list == NULL) {
 935                 DRV_LOG(ERR, "Invalid device id: %d", did);
 936                 return -1;
 937         }
 938
 939         internal = list->internal;
 940         rte_vhost_get_negotiated_features(vid, &features);
 941
 942         if (!RTE_VHOST_NEED_LOG(features))
 943                 return 0;
 944
 945         if (internal->sw_lm) {
 946                 ifcvf_sw_fallback_switchover(internal);
 947         } else {
 948                 rte_vhost_get_log_base(vid, &log_base, &log_size);
 949                 rte_vfio_container_dma_map(internal->vfio_container_fd,
 950                                 log_base, IFCVF_LOG_BASE, log_size);
 951                 ifcvf_enable_logging(&internal->hw, IFCVF_LOG_BASE, log_size);
 952         }
 953
 954         return 0;
 955 }
 956
 957 static int
 958 ifcvf_get_vfio_group_fd(int vid)
 959 {
 960         int did;
 961         struct internal_list *list;
 962
 963         did = rte_vhost_get_vdpa_device_id(vid);
 964         list = find_internal_resource_by_did(did);
 965         if (list == NULL) {
 966                 DRV_LOG(ERR, "Invalid device id: %d", did);
 967                 return -1;
 968         }
 969
 970         return list->internal->vfio_group_fd;
 971 }
 972
 973 static int
 974 ifcvf_get_vfio_device_fd(int vid)
 975 {
 976         int did;
 977         struct internal_list *list;
 978
 979         did = rte_vhost_get_vdpa_device_id(vid);
 980         list = find_internal_resource_by_did(did);
 981         if (list == NULL) {
 982                 DRV_LOG(ERR, "Invalid device id: %d", did);
 983                 return -1;
 984         }
 985
 986         return list->internal->vfio_dev_fd;
 987 }
 988
 989 static int
 990 ifcvf_get_notify_area(int vid, int qid, uint64_t *offset, uint64_t *size)
 991 {
 992         int did;
 993         struct internal_list *list;
 994         struct ifcvf_internal *internal;
 995         struct vfio_region_info reg = { .argsz = sizeof(reg) };
 996         int ret;
 997
 998         did = rte_vhost_get_vdpa_device_id(vid);
 999         list = find_internal_resource_by_did(did);
1000         if (list == NULL) {
1001                 DRV_LOG(ERR, "Invalid device id: %d", did);
1002                 return -1;
1003         }
1004
1005         internal = list->internal;
1006
1007         reg.index = ifcvf_get_notify_region(&internal->hw);
1008         ret = ioctl(internal->vfio_dev_fd, VFIO_DEVICE_GET_REGION_INFO, &reg);
1009         if (ret) {
1010                 DRV_LOG(ERR, "Get not get device region info: %s",
1011                                 strerror(errno));
1012                 return -1;
1013         }
1014
1015         *offset = ifcvf_get_queue_notify_off(&internal->hw, qid) + reg.offset;
1016         *size = 0x1000;
1017
1018         return 0;
1019 }
1020
1021 static int
1022 ifcvf_get_queue_num(int did, uint32_t *queue_num)
1023 {
1024         struct internal_list *list;
1025
1026         list = find_internal_resource_by_did(did);
1027         if (list == NULL) {
1028                 DRV_LOG(ERR, "Invalid device id: %d", did);
1029                 return -1;
1030         }
1031
1032         *queue_num = list->internal->max_queues;
1033
1034         return 0;
1035 }
1036
1037 static int
1038 ifcvf_get_vdpa_features(int did, uint64_t *features)
1039 {
1040         struct internal_list *list;
1041
1042         list = find_internal_resource_by_did(did);
1043         if (list == NULL) {
1044                 DRV_LOG(ERR, "Invalid device id: %d", did);
1045                 return -1;
1046         }
1047
1048         *features = list->internal->features;
1049
1050         return 0;
1051 }
1052
1053 #define VDPA_SUPPORTED_PROTOCOL_FEATURES \
1054                 (1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK | \
1055                  1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ | \
1056                  1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD | \
1057                  1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER | \
1058                  1ULL << VHOST_USER_PROTOCOL_F_LOG_SHMFD)
1059 static int
1060 ifcvf_get_protocol_features(int did __rte_unused, uint64_t *features)
1061 {
1062         *features = VDPA_SUPPORTED_PROTOCOL_FEATURES;
1063         return 0;
1064 }
1065
1066 static struct rte_vdpa_dev_ops ifcvf_ops = {
1067         .get_queue_num = ifcvf_get_queue_num,
1068         .get_features = ifcvf_get_vdpa_features,
1069         .get_protocol_features = ifcvf_get_protocol_features,
1070         .dev_conf = ifcvf_dev_config,
1071         .dev_close = ifcvf_dev_close,
1072         .set_vring_state = NULL,
1073         .set_features = ifcvf_set_features,
1074         .migration_done = NULL,
1075         .get_vfio_group_fd = ifcvf_get_vfio_group_fd,
1076         .get_vfio_device_fd = ifcvf_get_vfio_device_fd,
1077         .get_notify_area = ifcvf_get_notify_area,
1078 };
1079
1080 static inline int
1081 open_int(const char *key __rte_unused, const char *value, void *extra_args)
1082 {
1083         uint16_t *n = extra_args;
1084
1085         if (value == NULL || extra_args == NULL)
1086                 return -EINVAL;
1087
1088         *n = (uint16_t)strtoul(value, NULL, 0);
1089         if (*n == USHRT_MAX && errno == ERANGE)
1090                 return -1;
1091
1092         return 0;
1093 }
1094
1095 static int
1096 ifcvf_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1097                 struct rte_pci_device *pci_dev)
1098 {
1099         uint64_t features;
1100         struct ifcvf_internal *internal = NULL;
1101         struct internal_list *list = NULL;
1102         int vdpa_mode = 0;
1103         int sw_fallback_lm = 0;
1104         struct rte_kvargs *kvlist = NULL;
1105         int ret = 0;
1106
1107         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1108                 return 0;
1109
1110         kvlist = rte_kvargs_parse(pci_dev->device.devargs->args,
1111                         ifcvf_valid_arguments);
1112         if (kvlist == NULL)
1113                 return 1;
1114
1115         /* probe only when vdpa mode is specified */
1116         if (rte_kvargs_count(kvlist, IFCVF_VDPA_MODE) == 0) {
1117                 rte_kvargs_free(kvlist);
1118                 return 1;
1119         }
1120
1121         ret = rte_kvargs_process(kvlist, IFCVF_VDPA_MODE, &open_int,
1122                         &vdpa_mode);
1123         if (ret < 0 || vdpa_mode == 0) {
1124                 rte_kvargs_free(kvlist);
1125                 return 1;
1126         }
1127
1128         list = rte_zmalloc("ifcvf", sizeof(*list), 0);
1129         if (list == NULL)
1130                 goto error;
1131
1132         internal = rte_zmalloc("ifcvf", sizeof(*internal), 0);
1133         if (internal == NULL)
1134                 goto error;
1135
1136         internal->pdev = pci_dev;
1137         rte_spinlock_init(&internal->lock);
1138
1139         if (ifcvf_vfio_setup(internal) < 0) {
1140                 DRV_LOG(ERR, "failed to setup device %s", pci_dev->name);
1141                 goto error;
1142         }
1143
1144         if (ifcvf_init_hw(&internal->hw, internal->pdev) < 0) {
1145                 DRV_LOG(ERR, "failed to init device %s", pci_dev->name);
1146                 goto error;
1147         }
1148
1149         internal->max_queues = IFCVF_MAX_QUEUES;
1150         features = ifcvf_get_features(&internal->hw);
1151         internal->features = (features &
1152                 ~(1ULL << VIRTIO_F_IOMMU_PLATFORM)) |
1153                 (1ULL << VIRTIO_NET_F_GUEST_ANNOUNCE) |
1154                 (1ULL << VIRTIO_NET_F_CTRL_VQ) |
1155                 (1ULL << VIRTIO_NET_F_STATUS) |
1156                 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES) |
1157                 (1ULL << VHOST_F_LOG_ALL);
1158
1159         internal->dev_addr.pci_addr = pci_dev->addr;
1160         internal->dev_addr.type = PCI_ADDR;
1161         list->internal = internal;
1162
1163         if (rte_kvargs_count(kvlist, IFCVF_SW_FALLBACK_LM)) {
1164                 ret = rte_kvargs_process(kvlist, IFCVF_SW_FALLBACK_LM,
1165                                 &open_int, &sw_fallback_lm);
1166                 if (ret < 0)
1167                         goto error;
1168         }
1169         internal->sw_lm = sw_fallback_lm;
1170
1171         internal->did = rte_vdpa_register_device(&internal->dev_addr,
1172                                 &ifcvf_ops);
1173         if (internal->did < 0) {
1174                 DRV_LOG(ERR, "failed to register device %s", pci_dev->name);
1175                 goto error;
1176         }
1177
1178         pthread_mutex_lock(&internal_list_lock);
1179         TAILQ_INSERT_TAIL(&internal_list, list, next);
1180         pthread_mutex_unlock(&internal_list_lock);
1181
1182         rte_atomic32_set(&internal->started, 1);
1183         update_datapath(internal);
1184
1185         rte_kvargs_free(kvlist);
1186         return 0;
1187
1188 error:
1189         rte_kvargs_free(kvlist);
1190         rte_free(list);
1191         rte_free(internal);
1192         return -1;
1193 }
1194
1195 static int
1196 ifcvf_pci_remove(struct rte_pci_device *pci_dev)
1197 {
1198         struct ifcvf_internal *internal;
1199         struct internal_list *list;
1200
1201         if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1202                 return 0;
1203
1204         list = find_internal_resource_by_dev(pci_dev);
1205         if (list == NULL) {
1206                 DRV_LOG(ERR, "Invalid device: %s", pci_dev->name);
1207                 return -1;
1208         }
1209
1210         internal = list->internal;
1211         rte_atomic32_set(&internal->started, 0);
1212         update_datapath(internal);
1213
1214         rte_pci_unmap_device(internal->pdev);
1215         rte_vfio_container_destroy(internal->vfio_container_fd);
1216         rte_vdpa_unregister_device(internal->did);
1217
1218         pthread_mutex_lock(&internal_list_lock);
1219         TAILQ_REMOVE(&internal_list, list, next);
1220         pthread_mutex_unlock(&internal_list_lock);
1221
1222         rte_free(list);
1223         rte_free(internal);
1224
1225         return 0;
1226 }
1227
1228 /*
1229  * IFCVF has the same vendor ID and device ID as virtio net PCI
1230  * device, with its specific subsystem vendor ID and device ID.
1231  */
1232 static const struct rte_pci_id pci_id_ifcvf_map[] = {
1233         { .class_id = RTE_CLASS_ANY_ID,
1234           .vendor_id = IFCVF_VENDOR_ID,
1235           .device_id = IFCVF_DEVICE_ID,
1236           .subsystem_vendor_id = IFCVF_SUBSYS_VENDOR_ID,
1237           .subsystem_device_id = IFCVF_SUBSYS_DEVICE_ID,
1238         },
1239
1240         { .vendor_id = 0, /* sentinel */
1241         },
1242 };
1243
1244 static struct rte_pci_driver rte_ifcvf_vdpa = {
1245         .id_table = pci_id_ifcvf_map,
1246         .drv_flags = 0,
1247         .probe = ifcvf_pci_probe,
1248         .remove = ifcvf_pci_remove,
1249 };
1250
1251 RTE_PMD_REGISTER_PCI(net_ifcvf, rte_ifcvf_vdpa);
1252 RTE_PMD_REGISTER_PCI_TABLE(net_ifcvf, pci_id_ifcvf_map);
1253 RTE_PMD_REGISTER_KMOD_DEP(net_ifcvf, "* vfio-pci");
1254
1255 RTE_INIT(ifcvf_vdpa_init_log)
1256 {
1257         ifcvf_vdpa_logtype = rte_log_register("pmd.net.ifcvf_vdpa");
1258         if (ifcvf_vdpa_logtype >= 0)
1259                 rte_log_set_level(ifcvf_vdpa_logtype, RTE_LOG_NOTICE);
1260 }