1 /* SPDX-License-Identifier: BSD-3-Clause
2 * Copyright(c) 2018 Intel Corporation
10 #include <sys/epoll.h>
11 #include <linux/virtio_net.h>
14 #include <rte_malloc.h>
15 #include <rte_memory.h>
16 #include <rte_bus_pci.h>
17 #include <rte_vhost.h>
20 #include <rte_spinlock.h>
22 #include <rte_kvargs.h>
23 #include <rte_devargs.h>
25 #include "base/ifcvf.h"
27 #define DRV_LOG(level, fmt, args...) \
28 rte_log(RTE_LOG_ ## level, ifcvf_vdpa_logtype, \
29 "IFCVF %s(): " fmt "\n", __func__, ##args)
32 #define PAGE_SIZE 4096
35 #define IFCVF_USED_RING_LEN(size) \
36 ((size) * sizeof(struct vring_used_elem) + sizeof(uint16_t) * 3)
38 #define IFCVF_VDPA_MODE "vdpa"
39 #define IFCVF_SW_FALLBACK_LM "sw-live-migration"
41 static const char * const ifcvf_valid_arguments[] = {
47 static int ifcvf_vdpa_logtype;
49 struct ifcvf_internal {
50 struct rte_pci_device *pdev;
52 int vfio_container_fd;
55 pthread_t tid; /* thread for notify relay */
61 rte_atomic32_t started;
62 rte_atomic32_t dev_attached;
63 rte_atomic32_t running;
66 bool sw_fallback_running;
67 /* mediated vring for sw fallback */
68 struct vring m_vring[IFCVF_MAX_QUEUES * 2];
69 /* eventfd for used ring interrupt */
70 int intr_fd[IFCVF_MAX_QUEUES * 2];
73 struct internal_list {
74 TAILQ_ENTRY(internal_list) next;
75 struct ifcvf_internal *internal;
78 TAILQ_HEAD(internal_list_head, internal_list);
79 static struct internal_list_head internal_list =
80 TAILQ_HEAD_INITIALIZER(internal_list);
82 static pthread_mutex_t internal_list_lock = PTHREAD_MUTEX_INITIALIZER;
84 static void update_used_ring(struct ifcvf_internal *internal, uint16_t qid);
86 static struct internal_list *
87 find_internal_resource_by_did(int did)
90 struct internal_list *list;
92 pthread_mutex_lock(&internal_list_lock);
94 TAILQ_FOREACH(list, &internal_list, next) {
95 if (did == list->internal->did) {
101 pthread_mutex_unlock(&internal_list_lock);
109 static struct internal_list *
110 find_internal_resource_by_dev(struct rte_pci_device *pdev)
113 struct internal_list *list;
115 pthread_mutex_lock(&internal_list_lock);
117 TAILQ_FOREACH(list, &internal_list, next) {
118 if (!rte_pci_addr_cmp(&pdev->addr,
119 &list->internal->pdev->addr)) {
125 pthread_mutex_unlock(&internal_list_lock);
134 ifcvf_vfio_setup(struct ifcvf_internal *internal)
136 struct rte_pci_device *dev = internal->pdev;
137 char devname[RTE_DEV_NAME_MAX_LEN] = {0};
141 internal->vfio_dev_fd = -1;
142 internal->vfio_group_fd = -1;
143 internal->vfio_container_fd = -1;
145 rte_pci_device_name(&dev->addr, devname, RTE_DEV_NAME_MAX_LEN);
146 ret = rte_vfio_get_group_num(rte_pci_get_sysfs_path(), devname,
149 DRV_LOG(ERR, "%s failed to get IOMMU group", devname);
153 internal->vfio_container_fd = rte_vfio_container_create();
154 if (internal->vfio_container_fd < 0)
157 internal->vfio_group_fd = rte_vfio_container_group_bind(
158 internal->vfio_container_fd, iommu_group_num);
159 if (internal->vfio_group_fd < 0)
162 if (rte_pci_map_device(dev))
165 internal->vfio_dev_fd = dev->intr_handle.vfio_dev_fd;
167 for (i = 0; i < RTE_MIN(PCI_MAX_RESOURCE, IFCVF_PCI_MAX_RESOURCE);
169 internal->hw.mem_resource[i].addr =
170 internal->pdev->mem_resource[i].addr;
171 internal->hw.mem_resource[i].phys_addr =
172 internal->pdev->mem_resource[i].phys_addr;
173 internal->hw.mem_resource[i].len =
174 internal->pdev->mem_resource[i].len;
180 rte_vfio_container_destroy(internal->vfio_container_fd);
185 ifcvf_dma_map(struct ifcvf_internal *internal, int do_map)
189 struct rte_vhost_memory *mem = NULL;
190 int vfio_container_fd;
192 ret = rte_vhost_get_mem_table(internal->vid, &mem);
194 DRV_LOG(ERR, "failed to get VM memory layout.");
198 vfio_container_fd = internal->vfio_container_fd;
200 for (i = 0; i < mem->nregions; i++) {
201 struct rte_vhost_mem_region *reg;
203 reg = &mem->regions[i];
204 DRV_LOG(INFO, "%s, region %u: HVA 0x%" PRIx64 ", "
205 "GPA 0x%" PRIx64 ", size 0x%" PRIx64 ".",
206 do_map ? "DMA map" : "DMA unmap", i,
207 reg->host_user_addr, reg->guest_phys_addr, reg->size);
210 ret = rte_vfio_container_dma_map(vfio_container_fd,
211 reg->host_user_addr, reg->guest_phys_addr,
214 DRV_LOG(ERR, "DMA map failed.");
218 ret = rte_vfio_container_dma_unmap(vfio_container_fd,
219 reg->host_user_addr, reg->guest_phys_addr,
222 DRV_LOG(ERR, "DMA unmap failed.");
235 hva_to_gpa(int vid, uint64_t hva)
237 struct rte_vhost_memory *mem = NULL;
238 struct rte_vhost_mem_region *reg;
242 if (rte_vhost_get_mem_table(vid, &mem) < 0)
245 for (i = 0; i < mem->nregions; i++) {
246 reg = &mem->regions[i];
248 if (hva >= reg->host_user_addr &&
249 hva < reg->host_user_addr + reg->size) {
250 gpa = hva - reg->host_user_addr + reg->guest_phys_addr;
262 vdpa_ifcvf_start(struct ifcvf_internal *internal)
264 struct ifcvf_hw *hw = &internal->hw;
267 struct rte_vhost_vring vq;
271 nr_vring = rte_vhost_get_vring_num(vid);
272 rte_vhost_get_negotiated_features(vid, &hw->req_features);
274 for (i = 0; i < nr_vring; i++) {
275 rte_vhost_get_vhost_vring(vid, i, &vq);
276 gpa = hva_to_gpa(vid, (uint64_t)(uintptr_t)vq.desc);
278 DRV_LOG(ERR, "Fail to get GPA for descriptor ring.");
281 hw->vring[i].desc = gpa;
283 gpa = hva_to_gpa(vid, (uint64_t)(uintptr_t)vq.avail);
285 DRV_LOG(ERR, "Fail to get GPA for available ring.");
288 hw->vring[i].avail = gpa;
290 gpa = hva_to_gpa(vid, (uint64_t)(uintptr_t)vq.used);
292 DRV_LOG(ERR, "Fail to get GPA for used ring.");
295 hw->vring[i].used = gpa;
297 hw->vring[i].size = vq.size;
298 rte_vhost_get_vring_base(vid, i, &hw->vring[i].last_avail_idx,
299 &hw->vring[i].last_used_idx);
303 return ifcvf_start_hw(&internal->hw);
307 vdpa_ifcvf_stop(struct ifcvf_internal *internal)
309 struct ifcvf_hw *hw = &internal->hw;
312 uint64_t features = 0;
313 uint64_t log_base = 0, log_size = 0;
319 for (i = 0; i < hw->nr_vring; i++)
320 rte_vhost_set_vring_base(vid, i, hw->vring[i].last_avail_idx,
321 hw->vring[i].last_used_idx);
326 rte_vhost_get_negotiated_features(vid, &features);
327 if (RTE_VHOST_NEED_LOG(features)) {
328 ifcvf_disable_logging(hw);
329 rte_vhost_get_log_base(internal->vid, &log_base, &log_size);
330 rte_vfio_container_dma_unmap(internal->vfio_container_fd,
331 log_base, IFCVF_LOG_BASE, log_size);
333 * IFCVF marks dirty memory pages for only packet buffer,
334 * SW helps to mark the used ring as dirty after device stops.
336 for (i = 0; i < hw->nr_vring; i++) {
337 len = IFCVF_USED_RING_LEN(hw->vring[i].size);
338 rte_vhost_log_used_vring(vid, i, 0, len);
343 #define MSIX_IRQ_SET_BUF_LEN (sizeof(struct vfio_irq_set) + \
344 sizeof(int) * (IFCVF_MAX_QUEUES * 2 + 1))
346 vdpa_enable_vfio_intr(struct ifcvf_internal *internal, bool m_rx)
349 uint32_t i, nr_vring;
350 char irq_set_buf[MSIX_IRQ_SET_BUF_LEN];
351 struct vfio_irq_set *irq_set;
353 struct rte_vhost_vring vring;
358 nr_vring = rte_vhost_get_vring_num(internal->vid);
360 irq_set = (struct vfio_irq_set *)irq_set_buf;
361 irq_set->argsz = sizeof(irq_set_buf);
362 irq_set->count = nr_vring + 1;
363 irq_set->flags = VFIO_IRQ_SET_DATA_EVENTFD |
364 VFIO_IRQ_SET_ACTION_TRIGGER;
365 irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
367 fd_ptr = (int *)&irq_set->data;
368 fd_ptr[RTE_INTR_VEC_ZERO_OFFSET] = internal->pdev->intr_handle.fd;
370 for (i = 0; i < nr_vring; i++)
371 internal->intr_fd[i] = -1;
373 for (i = 0; i < nr_vring; i++) {
374 rte_vhost_get_vhost_vring(internal->vid, i, &vring);
375 fd_ptr[RTE_INTR_VEC_RXTX_OFFSET + i] = vring.callfd;
376 if ((i & 1) == 0 && m_rx == true) {
377 fd = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC);
379 DRV_LOG(ERR, "can't setup eventfd: %s",
383 internal->intr_fd[i] = fd;
384 fd_ptr[RTE_INTR_VEC_RXTX_OFFSET + i] = fd;
388 ret = ioctl(internal->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
390 DRV_LOG(ERR, "Error enabling MSI-X interrupts: %s",
399 vdpa_disable_vfio_intr(struct ifcvf_internal *internal)
402 uint32_t i, nr_vring;
403 char irq_set_buf[MSIX_IRQ_SET_BUF_LEN];
404 struct vfio_irq_set *irq_set;
406 irq_set = (struct vfio_irq_set *)irq_set_buf;
407 irq_set->argsz = sizeof(irq_set_buf);
409 irq_set->flags = VFIO_IRQ_SET_DATA_NONE | VFIO_IRQ_SET_ACTION_TRIGGER;
410 irq_set->index = VFIO_PCI_MSIX_IRQ_INDEX;
413 nr_vring = rte_vhost_get_vring_num(internal->vid);
414 for (i = 0; i < nr_vring; i++) {
415 if (internal->intr_fd[i] >= 0)
416 close(internal->intr_fd[i]);
417 internal->intr_fd[i] = -1;
420 ret = ioctl(internal->vfio_dev_fd, VFIO_DEVICE_SET_IRQS, irq_set);
422 DRV_LOG(ERR, "Error disabling MSI-X interrupts: %s",
431 notify_relay(void *arg)
433 int i, kickfd, epfd, nfds = 0;
435 struct epoll_event events[IFCVF_MAX_QUEUES * 2];
436 struct epoll_event ev;
439 struct rte_vhost_vring vring;
440 struct ifcvf_internal *internal = (struct ifcvf_internal *)arg;
441 struct ifcvf_hw *hw = &internal->hw;
443 q_num = rte_vhost_get_vring_num(internal->vid);
445 epfd = epoll_create(IFCVF_MAX_QUEUES * 2);
447 DRV_LOG(ERR, "failed to create epoll instance.");
450 internal->epfd = epfd;
453 for (qid = 0; qid < q_num; qid++) {
454 ev.events = EPOLLIN | EPOLLPRI;
455 rte_vhost_get_vhost_vring(internal->vid, qid, &vring);
456 ev.data.u64 = qid | (uint64_t)vring.kickfd << 32;
457 if (epoll_ctl(epfd, EPOLL_CTL_ADD, vring.kickfd, &ev) < 0) {
458 DRV_LOG(ERR, "epoll add error: %s", strerror(errno));
464 nfds = epoll_wait(epfd, events, q_num, -1);
468 DRV_LOG(ERR, "epoll_wait return fail\n");
472 for (i = 0; i < nfds; i++) {
473 qid = events[i].data.u32;
474 kickfd = (uint32_t)(events[i].data.u64 >> 32);
476 nbytes = read(kickfd, &buf, 8);
478 if (errno == EINTR ||
479 errno == EWOULDBLOCK ||
482 DRV_LOG(INFO, "Error reading "
489 ifcvf_notify_queue(hw, qid);
497 setup_notify_relay(struct ifcvf_internal *internal)
501 ret = pthread_create(&internal->tid, NULL, notify_relay,
504 DRV_LOG(ERR, "failed to create notify relay pthread.");
511 unset_notify_relay(struct ifcvf_internal *internal)
516 pthread_cancel(internal->tid);
517 pthread_join(internal->tid, &status);
521 if (internal->epfd >= 0)
522 close(internal->epfd);
529 update_datapath(struct ifcvf_internal *internal)
533 rte_spinlock_lock(&internal->lock);
535 if (!rte_atomic32_read(&internal->running) &&
536 (rte_atomic32_read(&internal->started) &&
537 rte_atomic32_read(&internal->dev_attached))) {
538 ret = ifcvf_dma_map(internal, 1);
542 ret = vdpa_enable_vfio_intr(internal, 0);
546 ret = vdpa_ifcvf_start(internal);
550 ret = setup_notify_relay(internal);
554 rte_atomic32_set(&internal->running, 1);
555 } else if (rte_atomic32_read(&internal->running) &&
556 (!rte_atomic32_read(&internal->started) ||
557 !rte_atomic32_read(&internal->dev_attached))) {
558 ret = unset_notify_relay(internal);
562 vdpa_ifcvf_stop(internal);
564 ret = vdpa_disable_vfio_intr(internal);
568 ret = ifcvf_dma_map(internal, 0);
572 rte_atomic32_set(&internal->running, 0);
575 rte_spinlock_unlock(&internal->lock);
578 rte_spinlock_unlock(&internal->lock);
583 m_ifcvf_start(struct ifcvf_internal *internal)
585 struct ifcvf_hw *hw = &internal->hw;
586 uint32_t i, nr_vring;
588 struct rte_vhost_vring vq;
590 uint64_t m_vring_iova = IFCVF_MEDIATED_VRING;
594 memset(&vq, 0, sizeof(vq));
596 nr_vring = rte_vhost_get_vring_num(vid);
597 rte_vhost_get_negotiated_features(vid, &hw->req_features);
599 for (i = 0; i < nr_vring; i++) {
600 rte_vhost_get_vhost_vring(vid, i, &vq);
602 size = RTE_ALIGN_CEIL(vring_size(vq.size, PAGE_SIZE),
604 vring_buf = rte_zmalloc("ifcvf", size, PAGE_SIZE);
605 vring_init(&internal->m_vring[i], vq.size, vring_buf,
608 ret = rte_vfio_container_dma_map(internal->vfio_container_fd,
609 (uint64_t)(uintptr_t)vring_buf, m_vring_iova, size);
611 DRV_LOG(ERR, "mediated vring DMA map failed.");
615 gpa = hva_to_gpa(vid, (uint64_t)(uintptr_t)vq.desc);
617 DRV_LOG(ERR, "Fail to get GPA for descriptor ring.");
620 hw->vring[i].desc = gpa;
622 gpa = hva_to_gpa(vid, (uint64_t)(uintptr_t)vq.avail);
624 DRV_LOG(ERR, "Fail to get GPA for available ring.");
627 hw->vring[i].avail = gpa;
629 /* Direct I/O for Tx queue, relay for Rx queue */
631 gpa = hva_to_gpa(vid, (uint64_t)(uintptr_t)vq.used);
633 DRV_LOG(ERR, "Fail to get GPA for used ring.");
636 hw->vring[i].used = gpa;
638 hw->vring[i].used = m_vring_iova +
639 (char *)internal->m_vring[i].used -
640 (char *)internal->m_vring[i].desc;
643 hw->vring[i].size = vq.size;
645 rte_vhost_get_vring_base(vid, i,
646 &internal->m_vring[i].avail->idx,
647 &internal->m_vring[i].used->idx);
649 rte_vhost_get_vring_base(vid, i, &hw->vring[i].last_avail_idx,
650 &hw->vring[i].last_used_idx);
652 m_vring_iova += size;
654 hw->nr_vring = nr_vring;
656 return ifcvf_start_hw(&internal->hw);
659 for (i = 0; i < nr_vring; i++)
660 if (internal->m_vring[i].desc)
661 rte_free(internal->m_vring[i].desc);
667 m_ifcvf_stop(struct ifcvf_internal *internal)
671 struct rte_vhost_vring vq;
672 struct ifcvf_hw *hw = &internal->hw;
673 uint64_t m_vring_iova = IFCVF_MEDIATED_VRING;
679 for (i = 0; i < hw->nr_vring; i++) {
680 /* synchronize remaining new used entries if any */
682 update_used_ring(internal, i);
684 rte_vhost_get_vhost_vring(vid, i, &vq);
685 len = IFCVF_USED_RING_LEN(vq.size);
686 rte_vhost_log_used_vring(vid, i, 0, len);
688 size = RTE_ALIGN_CEIL(vring_size(vq.size, PAGE_SIZE),
690 rte_vfio_container_dma_unmap(internal->vfio_container_fd,
691 (uint64_t)(uintptr_t)internal->m_vring[i].desc,
694 rte_vhost_set_vring_base(vid, i, hw->vring[i].last_avail_idx,
695 hw->vring[i].last_used_idx);
696 rte_free(internal->m_vring[i].desc);
697 m_vring_iova += size;
704 update_used_ring(struct ifcvf_internal *internal, uint16_t qid)
706 rte_vdpa_relay_vring_used(internal->vid, qid, &internal->m_vring[qid]);
707 rte_vhost_vring_call(internal->vid, qid);
711 vring_relay(void *arg)
713 int i, vid, epfd, fd, nfds;
714 struct ifcvf_internal *internal = (struct ifcvf_internal *)arg;
715 struct rte_vhost_vring vring;
717 struct epoll_event events[IFCVF_MAX_QUEUES * 4];
718 struct epoll_event ev;
723 q_num = rte_vhost_get_vring_num(vid);
725 /* add notify fd and interrupt fd to epoll */
726 epfd = epoll_create(IFCVF_MAX_QUEUES * 2);
728 DRV_LOG(ERR, "failed to create epoll instance.");
731 internal->epfd = epfd;
734 for (qid = 0; qid < q_num; qid++) {
735 ev.events = EPOLLIN | EPOLLPRI;
736 rte_vhost_get_vhost_vring(vid, qid, &vring);
737 ev.data.u64 = qid << 1 | (uint64_t)vring.kickfd << 32;
738 if (epoll_ctl(epfd, EPOLL_CTL_ADD, vring.kickfd, &ev) < 0) {
739 DRV_LOG(ERR, "epoll add error: %s", strerror(errno));
744 for (qid = 0; qid < q_num; qid += 2) {
745 ev.events = EPOLLIN | EPOLLPRI;
746 /* leave a flag to mark it's for interrupt */
747 ev.data.u64 = 1 | qid << 1 |
748 (uint64_t)internal->intr_fd[qid] << 32;
749 if (epoll_ctl(epfd, EPOLL_CTL_ADD, internal->intr_fd[qid], &ev)
751 DRV_LOG(ERR, "epoll add error: %s", strerror(errno));
754 update_used_ring(internal, qid);
757 /* start relay with a first kick */
758 for (qid = 0; qid < q_num; qid++)
759 ifcvf_notify_queue(&internal->hw, qid);
761 /* listen to the events and react accordingly */
763 nfds = epoll_wait(epfd, events, q_num * 2, -1);
767 DRV_LOG(ERR, "epoll_wait return fail\n");
771 for (i = 0; i < nfds; i++) {
772 fd = (uint32_t)(events[i].data.u64 >> 32);
774 nbytes = read(fd, &buf, 8);
776 if (errno == EINTR ||
777 errno == EWOULDBLOCK ||
780 DRV_LOG(INFO, "Error reading "
787 qid = events[i].data.u32 >> 1;
789 if (events[i].data.u32 & 1)
790 update_used_ring(internal, qid);
792 ifcvf_notify_queue(&internal->hw, qid);
800 setup_vring_relay(struct ifcvf_internal *internal)
804 ret = pthread_create(&internal->tid, NULL, vring_relay,
807 DRV_LOG(ERR, "failed to create ring relay pthread.");
814 unset_vring_relay(struct ifcvf_internal *internal)
819 pthread_cancel(internal->tid);
820 pthread_join(internal->tid, &status);
824 if (internal->epfd >= 0)
825 close(internal->epfd);
832 ifcvf_sw_fallback_switchover(struct ifcvf_internal *internal)
835 int vid = internal->vid;
837 /* stop the direct IO data path */
838 unset_notify_relay(internal);
839 vdpa_ifcvf_stop(internal);
840 vdpa_disable_vfio_intr(internal);
842 ret = rte_vhost_host_notifier_ctrl(vid, false);
843 if (ret && ret != -ENOTSUP)
846 /* set up interrupt for interrupt relay */
847 ret = vdpa_enable_vfio_intr(internal, 1);
852 ret = m_ifcvf_start(internal);
856 /* set up vring relay thread */
857 ret = setup_vring_relay(internal);
861 rte_vhost_host_notifier_ctrl(vid, true);
863 internal->sw_fallback_running = true;
868 m_ifcvf_stop(internal);
870 vdpa_disable_vfio_intr(internal);
872 ifcvf_dma_map(internal, 0);
878 ifcvf_dev_config(int vid)
881 struct internal_list *list;
882 struct ifcvf_internal *internal;
884 did = rte_vhost_get_vdpa_device_id(vid);
885 list = find_internal_resource_by_did(did);
887 DRV_LOG(ERR, "Invalid device id: %d", did);
891 internal = list->internal;
893 rte_atomic32_set(&internal->dev_attached, 1);
894 update_datapath(internal);
896 if (rte_vhost_host_notifier_ctrl(vid, true) != 0)
897 DRV_LOG(NOTICE, "vDPA (%d): software relay is used.", did);
903 ifcvf_dev_close(int vid)
906 struct internal_list *list;
907 struct ifcvf_internal *internal;
909 did = rte_vhost_get_vdpa_device_id(vid);
910 list = find_internal_resource_by_did(did);
912 DRV_LOG(ERR, "Invalid device id: %d", did);
916 internal = list->internal;
918 if (internal->sw_fallback_running) {
919 /* unset ring relay */
920 unset_vring_relay(internal);
923 m_ifcvf_stop(internal);
925 /* remove interrupt setting */
926 vdpa_disable_vfio_intr(internal);
928 /* unset DMA map for guest memory */
929 ifcvf_dma_map(internal, 0);
931 internal->sw_fallback_running = false;
933 rte_atomic32_set(&internal->dev_attached, 0);
934 update_datapath(internal);
941 ifcvf_set_features(int vid)
943 uint64_t features = 0;
945 struct internal_list *list;
946 struct ifcvf_internal *internal;
947 uint64_t log_base = 0, log_size = 0;
949 did = rte_vhost_get_vdpa_device_id(vid);
950 list = find_internal_resource_by_did(did);
952 DRV_LOG(ERR, "Invalid device id: %d", did);
956 internal = list->internal;
957 rte_vhost_get_negotiated_features(vid, &features);
959 if (!RTE_VHOST_NEED_LOG(features))
962 if (internal->sw_lm) {
963 ifcvf_sw_fallback_switchover(internal);
965 rte_vhost_get_log_base(vid, &log_base, &log_size);
966 rte_vfio_container_dma_map(internal->vfio_container_fd,
967 log_base, IFCVF_LOG_BASE, log_size);
968 ifcvf_enable_logging(&internal->hw, IFCVF_LOG_BASE, log_size);
975 ifcvf_get_vfio_group_fd(int vid)
978 struct internal_list *list;
980 did = rte_vhost_get_vdpa_device_id(vid);
981 list = find_internal_resource_by_did(did);
983 DRV_LOG(ERR, "Invalid device id: %d", did);
987 return list->internal->vfio_group_fd;
991 ifcvf_get_vfio_device_fd(int vid)
994 struct internal_list *list;
996 did = rte_vhost_get_vdpa_device_id(vid);
997 list = find_internal_resource_by_did(did);
999 DRV_LOG(ERR, "Invalid device id: %d", did);
1003 return list->internal->vfio_dev_fd;
1007 ifcvf_get_notify_area(int vid, int qid, uint64_t *offset, uint64_t *size)
1010 struct internal_list *list;
1011 struct ifcvf_internal *internal;
1012 struct vfio_region_info reg = { .argsz = sizeof(reg) };
1015 did = rte_vhost_get_vdpa_device_id(vid);
1016 list = find_internal_resource_by_did(did);
1018 DRV_LOG(ERR, "Invalid device id: %d", did);
1022 internal = list->internal;
1024 reg.index = ifcvf_get_notify_region(&internal->hw);
1025 ret = ioctl(internal->vfio_dev_fd, VFIO_DEVICE_GET_REGION_INFO, ®);
1027 DRV_LOG(ERR, "Get not get device region info: %s",
1032 *offset = ifcvf_get_queue_notify_off(&internal->hw, qid) + reg.offset;
1039 ifcvf_get_queue_num(int did, uint32_t *queue_num)
1041 struct internal_list *list;
1043 list = find_internal_resource_by_did(did);
1045 DRV_LOG(ERR, "Invalid device id: %d", did);
1049 *queue_num = list->internal->max_queues;
1055 ifcvf_get_vdpa_features(int did, uint64_t *features)
1057 struct internal_list *list;
1059 list = find_internal_resource_by_did(did);
1061 DRV_LOG(ERR, "Invalid device id: %d", did);
1065 *features = list->internal->features;
1070 #define VDPA_SUPPORTED_PROTOCOL_FEATURES \
1071 (1ULL << VHOST_USER_PROTOCOL_F_REPLY_ACK | \
1072 1ULL << VHOST_USER_PROTOCOL_F_SLAVE_REQ | \
1073 1ULL << VHOST_USER_PROTOCOL_F_SLAVE_SEND_FD | \
1074 1ULL << VHOST_USER_PROTOCOL_F_HOST_NOTIFIER | \
1075 1ULL << VHOST_USER_PROTOCOL_F_LOG_SHMFD)
1077 ifcvf_get_protocol_features(int did __rte_unused, uint64_t *features)
1079 *features = VDPA_SUPPORTED_PROTOCOL_FEATURES;
1083 static struct rte_vdpa_dev_ops ifcvf_ops = {
1084 .get_queue_num = ifcvf_get_queue_num,
1085 .get_features = ifcvf_get_vdpa_features,
1086 .get_protocol_features = ifcvf_get_protocol_features,
1087 .dev_conf = ifcvf_dev_config,
1088 .dev_close = ifcvf_dev_close,
1089 .set_vring_state = NULL,
1090 .set_features = ifcvf_set_features,
1091 .migration_done = NULL,
1092 .get_vfio_group_fd = ifcvf_get_vfio_group_fd,
1093 .get_vfio_device_fd = ifcvf_get_vfio_device_fd,
1094 .get_notify_area = ifcvf_get_notify_area,
1098 open_int(const char *key __rte_unused, const char *value, void *extra_args)
1100 uint16_t *n = extra_args;
1102 if (value == NULL || extra_args == NULL)
1105 *n = (uint16_t)strtoul(value, NULL, 0);
1106 if (*n == USHRT_MAX && errno == ERANGE)
1113 ifcvf_pci_probe(struct rte_pci_driver *pci_drv __rte_unused,
1114 struct rte_pci_device *pci_dev)
1117 struct ifcvf_internal *internal = NULL;
1118 struct internal_list *list = NULL;
1120 int sw_fallback_lm = 0;
1121 struct rte_kvargs *kvlist = NULL;
1124 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1127 if (!pci_dev->device.devargs)
1130 kvlist = rte_kvargs_parse(pci_dev->device.devargs->args,
1131 ifcvf_valid_arguments);
1135 /* probe only when vdpa mode is specified */
1136 if (rte_kvargs_count(kvlist, IFCVF_VDPA_MODE) == 0) {
1137 rte_kvargs_free(kvlist);
1141 ret = rte_kvargs_process(kvlist, IFCVF_VDPA_MODE, &open_int,
1143 if (ret < 0 || vdpa_mode == 0) {
1144 rte_kvargs_free(kvlist);
1148 list = rte_zmalloc("ifcvf", sizeof(*list), 0);
1152 internal = rte_zmalloc("ifcvf", sizeof(*internal), 0);
1153 if (internal == NULL)
1156 internal->pdev = pci_dev;
1157 rte_spinlock_init(&internal->lock);
1159 if (ifcvf_vfio_setup(internal) < 0) {
1160 DRV_LOG(ERR, "failed to setup device %s", pci_dev->name);
1164 if (ifcvf_init_hw(&internal->hw, internal->pdev) < 0) {
1165 DRV_LOG(ERR, "failed to init device %s", pci_dev->name);
1169 internal->max_queues = IFCVF_MAX_QUEUES;
1170 features = ifcvf_get_features(&internal->hw);
1171 internal->features = (features &
1172 ~(1ULL << VIRTIO_F_IOMMU_PLATFORM)) |
1173 (1ULL << VIRTIO_NET_F_GUEST_ANNOUNCE) |
1174 (1ULL << VIRTIO_NET_F_CTRL_VQ) |
1175 (1ULL << VIRTIO_NET_F_STATUS) |
1176 (1ULL << VHOST_USER_F_PROTOCOL_FEATURES) |
1177 (1ULL << VHOST_F_LOG_ALL);
1179 list->internal = internal;
1181 if (rte_kvargs_count(kvlist, IFCVF_SW_FALLBACK_LM)) {
1182 ret = rte_kvargs_process(kvlist, IFCVF_SW_FALLBACK_LM,
1183 &open_int, &sw_fallback_lm);
1187 internal->sw_lm = sw_fallback_lm;
1189 internal->did = rte_vdpa_register_device(&pci_dev->device, &ifcvf_ops);
1190 if (internal->did < 0) {
1191 DRV_LOG(ERR, "failed to register device %s", pci_dev->name);
1195 pthread_mutex_lock(&internal_list_lock);
1196 TAILQ_INSERT_TAIL(&internal_list, list, next);
1197 pthread_mutex_unlock(&internal_list_lock);
1199 rte_atomic32_set(&internal->started, 1);
1200 update_datapath(internal);
1202 rte_kvargs_free(kvlist);
1206 rte_kvargs_free(kvlist);
1213 ifcvf_pci_remove(struct rte_pci_device *pci_dev)
1215 struct ifcvf_internal *internal;
1216 struct internal_list *list;
1218 if (rte_eal_process_type() != RTE_PROC_PRIMARY)
1221 list = find_internal_resource_by_dev(pci_dev);
1223 DRV_LOG(ERR, "Invalid device: %s", pci_dev->name);
1227 internal = list->internal;
1228 rte_atomic32_set(&internal->started, 0);
1229 update_datapath(internal);
1231 rte_pci_unmap_device(internal->pdev);
1232 rte_vfio_container_destroy(internal->vfio_container_fd);
1233 rte_vdpa_unregister_device(internal->did);
1235 pthread_mutex_lock(&internal_list_lock);
1236 TAILQ_REMOVE(&internal_list, list, next);
1237 pthread_mutex_unlock(&internal_list_lock);
1246 * IFCVF has the same vendor ID and device ID as virtio net PCI
1247 * device, with its specific subsystem vendor ID and device ID.
1249 static const struct rte_pci_id pci_id_ifcvf_map[] = {
1250 { .class_id = RTE_CLASS_ANY_ID,
1251 .vendor_id = IFCVF_VENDOR_ID,
1252 .device_id = IFCVF_DEVICE_ID,
1253 .subsystem_vendor_id = IFCVF_SUBSYS_VENDOR_ID,
1254 .subsystem_device_id = IFCVF_SUBSYS_DEVICE_ID,
1257 { .vendor_id = 0, /* sentinel */
1261 static struct rte_pci_driver rte_ifcvf_vdpa = {
1262 .id_table = pci_id_ifcvf_map,
1264 .probe = ifcvf_pci_probe,
1265 .remove = ifcvf_pci_remove,
1268 RTE_PMD_REGISTER_PCI(net_ifcvf, rte_ifcvf_vdpa);
1269 RTE_PMD_REGISTER_PCI_TABLE(net_ifcvf, pci_id_ifcvf_map);
1270 RTE_PMD_REGISTER_KMOD_DEP(net_ifcvf, "* vfio-pci");
1272 RTE_INIT(ifcvf_vdpa_init_log)
1274 ifcvf_vdpa_logtype = rte_log_register("pmd.net.ifcvf_vdpa");
1275 if (ifcvf_vdpa_logtype >= 0)
1276 rte_log_set_level(ifcvf_vdpa_logtype, RTE_LOG_NOTICE);