11e0a84eaSCindy Lu /* 21e0a84eaSCindy Lu * vhost-vdpa.c 31e0a84eaSCindy Lu * 41e0a84eaSCindy Lu * Copyright(c) 2017-2018 Intel Corporation. 51e0a84eaSCindy Lu * Copyright(c) 2020 Red Hat, Inc. 61e0a84eaSCindy Lu * 71e0a84eaSCindy Lu * This work is licensed under the terms of the GNU GPL, version 2 or later. 81e0a84eaSCindy Lu * See the COPYING file in the top-level directory. 91e0a84eaSCindy Lu * 101e0a84eaSCindy Lu */ 111e0a84eaSCindy Lu 121e0a84eaSCindy Lu #include "qemu/osdep.h" 131e0a84eaSCindy Lu #include "clients.h" 14bd907ae4SEugenio Pérez #include "hw/virtio/virtio-net.h" 151e0a84eaSCindy Lu #include "net/vhost_net.h" 161e0a84eaSCindy Lu #include "net/vhost-vdpa.h" 171e0a84eaSCindy Lu #include "hw/virtio/vhost-vdpa.h" 181e0a84eaSCindy Lu #include "qemu/config-file.h" 191e0a84eaSCindy Lu #include "qemu/error-report.h" 20bd907ae4SEugenio Pérez #include "qemu/log.h" 21bd907ae4SEugenio Pérez #include "qemu/memalign.h" 221e0a84eaSCindy Lu #include "qemu/option.h" 231e0a84eaSCindy Lu #include "qapi/error.h" 2440237840SJason Wang #include <linux/vhost.h> 251e0a84eaSCindy Lu #include <sys/ioctl.h> 261e0a84eaSCindy Lu #include <err.h> 271e0a84eaSCindy Lu #include "standard-headers/linux/virtio_net.h" 281e0a84eaSCindy Lu #include "monitor/monitor.h" 291e0a84eaSCindy Lu #include "hw/virtio/vhost.h" 301e0a84eaSCindy Lu 311e0a84eaSCindy Lu /* Todo:need to add the multiqueue support here */ 321e0a84eaSCindy Lu typedef struct VhostVDPAState { 331e0a84eaSCindy Lu NetClientState nc; 341e0a84eaSCindy Lu struct vhost_vdpa vhost_vdpa; 351e0a84eaSCindy Lu VHostNetState *vhost_net; 362df4dd31SEugenio Pérez 372df4dd31SEugenio Pérez /* Control commands shadow buffers */ 382df4dd31SEugenio Pérez void *cvq_cmd_out_buffer, *cvq_cmd_in_buffer; 391e0a84eaSCindy Lu bool started; 401e0a84eaSCindy Lu } VhostVDPAState; 411e0a84eaSCindy Lu 421e0a84eaSCindy Lu const int vdpa_feature_bits[] = { 431e0a84eaSCindy Lu VIRTIO_F_NOTIFY_ON_EMPTY, 441e0a84eaSCindy Lu VIRTIO_RING_F_INDIRECT_DESC, 451e0a84eaSCindy Lu VIRTIO_RING_F_EVENT_IDX, 461e0a84eaSCindy Lu VIRTIO_F_ANY_LAYOUT, 471e0a84eaSCindy Lu VIRTIO_F_VERSION_1, 481e0a84eaSCindy Lu VIRTIO_NET_F_CSUM, 491e0a84eaSCindy Lu VIRTIO_NET_F_GUEST_CSUM, 501e0a84eaSCindy Lu VIRTIO_NET_F_GSO, 511e0a84eaSCindy Lu VIRTIO_NET_F_GUEST_TSO4, 521e0a84eaSCindy Lu VIRTIO_NET_F_GUEST_TSO6, 531e0a84eaSCindy Lu VIRTIO_NET_F_GUEST_ECN, 541e0a84eaSCindy Lu VIRTIO_NET_F_GUEST_UFO, 551e0a84eaSCindy Lu VIRTIO_NET_F_HOST_TSO4, 561e0a84eaSCindy Lu VIRTIO_NET_F_HOST_TSO6, 571e0a84eaSCindy Lu VIRTIO_NET_F_HOST_ECN, 581e0a84eaSCindy Lu VIRTIO_NET_F_HOST_UFO, 591e0a84eaSCindy Lu VIRTIO_NET_F_MRG_RXBUF, 601e0a84eaSCindy Lu VIRTIO_NET_F_MTU, 6140237840SJason Wang VIRTIO_NET_F_CTRL_RX, 6240237840SJason Wang VIRTIO_NET_F_CTRL_RX_EXTRA, 6340237840SJason Wang VIRTIO_NET_F_CTRL_VLAN, 6440237840SJason Wang VIRTIO_NET_F_GUEST_ANNOUNCE, 6540237840SJason Wang VIRTIO_NET_F_CTRL_MAC_ADDR, 6640237840SJason Wang VIRTIO_NET_F_RSS, 6740237840SJason Wang VIRTIO_NET_F_MQ, 6840237840SJason Wang VIRTIO_NET_F_CTRL_VQ, 691e0a84eaSCindy Lu VIRTIO_F_IOMMU_PLATFORM, 701e0a84eaSCindy Lu VIRTIO_F_RING_PACKED, 710145c393SAndrew Melnychenko VIRTIO_NET_F_RSS, 720145c393SAndrew Melnychenko VIRTIO_NET_F_HASH_REPORT, 731e0a84eaSCindy Lu VIRTIO_NET_F_GUEST_ANNOUNCE, 749aa47eddSSi-Wei Liu VIRTIO_NET_F_STATUS, 751e0a84eaSCindy Lu VHOST_INVALID_FEATURE_BIT 761e0a84eaSCindy Lu }; 771e0a84eaSCindy Lu 781576dbb5SEugenio Pérez /** Supported device specific feature bits with SVQ */ 791576dbb5SEugenio Pérez static const uint64_t vdpa_svq_device_features = 801576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_CSUM) | 811576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_GUEST_CSUM) | 821576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_MTU) | 831576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_MAC) | 841576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_GUEST_TSO4) | 851576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_GUEST_TSO6) | 861576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_GUEST_ECN) | 871576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_GUEST_UFO) | 881576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_HOST_TSO4) | 891576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_HOST_TSO6) | 901576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_HOST_ECN) | 911576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_HOST_UFO) | 921576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_MRG_RXBUF) | 931576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_STATUS) | 941576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_CTRL_VQ) | 951576dbb5SEugenio Pérez BIT_ULL(VIRTIO_F_ANY_LAYOUT) | 961576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_CTRL_MAC_ADDR) | 971576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_RSC_EXT) | 981576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_STANDBY); 991576dbb5SEugenio Pérez 1001e0a84eaSCindy Lu VHostNetState *vhost_vdpa_get_vhost_net(NetClientState *nc) 1011e0a84eaSCindy Lu { 1021e0a84eaSCindy Lu VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 1031e0a84eaSCindy Lu assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 1041e0a84eaSCindy Lu return s->vhost_net; 1051e0a84eaSCindy Lu } 1061e0a84eaSCindy Lu 1071e0a84eaSCindy Lu static int vhost_vdpa_net_check_device_id(struct vhost_net *net) 1081e0a84eaSCindy Lu { 1091e0a84eaSCindy Lu uint32_t device_id; 1101e0a84eaSCindy Lu int ret; 1111e0a84eaSCindy Lu struct vhost_dev *hdev; 1121e0a84eaSCindy Lu 1131e0a84eaSCindy Lu hdev = (struct vhost_dev *)&net->dev; 1141e0a84eaSCindy Lu ret = hdev->vhost_ops->vhost_get_device_id(hdev, &device_id); 1151e0a84eaSCindy Lu if (device_id != VIRTIO_ID_NET) { 1161e0a84eaSCindy Lu return -ENOTSUP; 1171e0a84eaSCindy Lu } 1181e0a84eaSCindy Lu return ret; 1191e0a84eaSCindy Lu } 1201e0a84eaSCindy Lu 12140237840SJason Wang static int vhost_vdpa_add(NetClientState *ncs, void *be, 12240237840SJason Wang int queue_pair_index, int nvqs) 1231e0a84eaSCindy Lu { 1241e0a84eaSCindy Lu VhostNetOptions options; 1251e0a84eaSCindy Lu struct vhost_net *net = NULL; 1261e0a84eaSCindy Lu VhostVDPAState *s; 1271e0a84eaSCindy Lu int ret; 1281e0a84eaSCindy Lu 1291e0a84eaSCindy Lu options.backend_type = VHOST_BACKEND_TYPE_VDPA; 1301e0a84eaSCindy Lu assert(ncs->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 1311e0a84eaSCindy Lu s = DO_UPCAST(VhostVDPAState, nc, ncs); 1321e0a84eaSCindy Lu options.net_backend = ncs; 1331e0a84eaSCindy Lu options.opaque = be; 1341e0a84eaSCindy Lu options.busyloop_timeout = 0; 13540237840SJason Wang options.nvqs = nvqs; 1361e0a84eaSCindy Lu 1371e0a84eaSCindy Lu net = vhost_net_init(&options); 1381e0a84eaSCindy Lu if (!net) { 1391e0a84eaSCindy Lu error_report("failed to init vhost_net for queue"); 140a97ef87aSJason Wang goto err_init; 1411e0a84eaSCindy Lu } 1421e0a84eaSCindy Lu s->vhost_net = net; 1431e0a84eaSCindy Lu ret = vhost_vdpa_net_check_device_id(net); 1441e0a84eaSCindy Lu if (ret) { 145a97ef87aSJason Wang goto err_check; 1461e0a84eaSCindy Lu } 1471e0a84eaSCindy Lu return 0; 148a97ef87aSJason Wang err_check: 1491e0a84eaSCindy Lu vhost_net_cleanup(net); 150ab36edcfSJason Wang g_free(net); 151a97ef87aSJason Wang err_init: 1521e0a84eaSCindy Lu return -1; 1531e0a84eaSCindy Lu } 1541e0a84eaSCindy Lu 1551e0a84eaSCindy Lu static void vhost_vdpa_cleanup(NetClientState *nc) 1561e0a84eaSCindy Lu { 1571e0a84eaSCindy Lu VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 1581576dbb5SEugenio Pérez struct vhost_dev *dev = &s->vhost_net->dev; 1591e0a84eaSCindy Lu 1602df4dd31SEugenio Pérez qemu_vfree(s->cvq_cmd_out_buffer); 1612df4dd31SEugenio Pérez qemu_vfree(s->cvq_cmd_in_buffer); 1621576dbb5SEugenio Pérez if (dev->vq_index + dev->nvqs == dev->vq_index_end) { 1631576dbb5SEugenio Pérez g_clear_pointer(&s->vhost_vdpa.iova_tree, vhost_iova_tree_delete); 1641576dbb5SEugenio Pérez } 1651e0a84eaSCindy Lu if (s->vhost_net) { 1661e0a84eaSCindy Lu vhost_net_cleanup(s->vhost_net); 1671e0a84eaSCindy Lu g_free(s->vhost_net); 1681e0a84eaSCindy Lu s->vhost_net = NULL; 1691e0a84eaSCindy Lu } 17057b3a7d8SCindy Lu if (s->vhost_vdpa.device_fd >= 0) { 17157b3a7d8SCindy Lu qemu_close(s->vhost_vdpa.device_fd); 17257b3a7d8SCindy Lu s->vhost_vdpa.device_fd = -1; 17357b3a7d8SCindy Lu } 1741e0a84eaSCindy Lu } 1751e0a84eaSCindy Lu 1761e0a84eaSCindy Lu static bool vhost_vdpa_has_vnet_hdr(NetClientState *nc) 1771e0a84eaSCindy Lu { 1781e0a84eaSCindy Lu assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 1791e0a84eaSCindy Lu 1801e0a84eaSCindy Lu return true; 1811e0a84eaSCindy Lu } 1821e0a84eaSCindy Lu 1831e0a84eaSCindy Lu static bool vhost_vdpa_has_ufo(NetClientState *nc) 1841e0a84eaSCindy Lu { 1851e0a84eaSCindy Lu assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 1861e0a84eaSCindy Lu VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 1871e0a84eaSCindy Lu uint64_t features = 0; 1881e0a84eaSCindy Lu features |= (1ULL << VIRTIO_NET_F_HOST_UFO); 1891e0a84eaSCindy Lu features = vhost_net_get_features(s->vhost_net, features); 1901e0a84eaSCindy Lu return !!(features & (1ULL << VIRTIO_NET_F_HOST_UFO)); 1911e0a84eaSCindy Lu 1921e0a84eaSCindy Lu } 1931e0a84eaSCindy Lu 194ee8a1c63SKevin Wolf static bool vhost_vdpa_check_peer_type(NetClientState *nc, ObjectClass *oc, 195ee8a1c63SKevin Wolf Error **errp) 196ee8a1c63SKevin Wolf { 197ee8a1c63SKevin Wolf const char *driver = object_class_get_name(oc); 198ee8a1c63SKevin Wolf 199ee8a1c63SKevin Wolf if (!g_str_has_prefix(driver, "virtio-net-")) { 200ee8a1c63SKevin Wolf error_setg(errp, "vhost-vdpa requires frontend driver virtio-net-*"); 201ee8a1c63SKevin Wolf return false; 202ee8a1c63SKevin Wolf } 203ee8a1c63SKevin Wolf 204ee8a1c63SKevin Wolf return true; 205ee8a1c63SKevin Wolf } 206ee8a1c63SKevin Wolf 207846a1e85SEugenio Pérez /** Dummy receive in case qemu falls back to userland tap networking */ 208846a1e85SEugenio Pérez static ssize_t vhost_vdpa_receive(NetClientState *nc, const uint8_t *buf, 209846a1e85SEugenio Pérez size_t size) 210846a1e85SEugenio Pérez { 211846a1e85SEugenio Pérez return 0; 212846a1e85SEugenio Pérez } 213846a1e85SEugenio Pérez 2141e0a84eaSCindy Lu static NetClientInfo net_vhost_vdpa_info = { 2151e0a84eaSCindy Lu .type = NET_CLIENT_DRIVER_VHOST_VDPA, 2161e0a84eaSCindy Lu .size = sizeof(VhostVDPAState), 217846a1e85SEugenio Pérez .receive = vhost_vdpa_receive, 2181e0a84eaSCindy Lu .cleanup = vhost_vdpa_cleanup, 2191e0a84eaSCindy Lu .has_vnet_hdr = vhost_vdpa_has_vnet_hdr, 2201e0a84eaSCindy Lu .has_ufo = vhost_vdpa_has_ufo, 221ee8a1c63SKevin Wolf .check_peer_type = vhost_vdpa_check_peer_type, 2221e0a84eaSCindy Lu }; 2231e0a84eaSCindy Lu 2242df4dd31SEugenio Pérez static void vhost_vdpa_cvq_unmap_buf(struct vhost_vdpa *v, void *addr) 2252df4dd31SEugenio Pérez { 2262df4dd31SEugenio Pérez VhostIOVATree *tree = v->iova_tree; 2272df4dd31SEugenio Pérez DMAMap needle = { 2282df4dd31SEugenio Pérez /* 2292df4dd31SEugenio Pérez * No need to specify size or to look for more translations since 2302df4dd31SEugenio Pérez * this contiguous chunk was allocated by us. 2312df4dd31SEugenio Pérez */ 2322df4dd31SEugenio Pérez .translated_addr = (hwaddr)(uintptr_t)addr, 2332df4dd31SEugenio Pérez }; 2342df4dd31SEugenio Pérez const DMAMap *map = vhost_iova_tree_find_iova(tree, &needle); 2352df4dd31SEugenio Pérez int r; 2362df4dd31SEugenio Pérez 2372df4dd31SEugenio Pérez if (unlikely(!map)) { 2382df4dd31SEugenio Pérez error_report("Cannot locate expected map"); 2392df4dd31SEugenio Pérez return; 2402df4dd31SEugenio Pérez } 2412df4dd31SEugenio Pérez 2422df4dd31SEugenio Pérez r = vhost_vdpa_dma_unmap(v, map->iova, map->size + 1); 2432df4dd31SEugenio Pérez if (unlikely(r != 0)) { 2442df4dd31SEugenio Pérez error_report("Device cannot unmap: %s(%d)", g_strerror(r), r); 2452df4dd31SEugenio Pérez } 2462df4dd31SEugenio Pérez 24769292a8eSEugenio Pérez vhost_iova_tree_remove(tree, *map); 2482df4dd31SEugenio Pérez } 2492df4dd31SEugenio Pérez 2502df4dd31SEugenio Pérez static size_t vhost_vdpa_net_cvq_cmd_len(void) 2512df4dd31SEugenio Pérez { 2522df4dd31SEugenio Pérez /* 2532df4dd31SEugenio Pérez * MAC_TABLE_SET is the ctrl command that produces the longer out buffer. 2542df4dd31SEugenio Pérez * In buffer is always 1 byte, so it should fit here 2552df4dd31SEugenio Pérez */ 2562df4dd31SEugenio Pérez return sizeof(struct virtio_net_ctrl_hdr) + 2572df4dd31SEugenio Pérez 2 * sizeof(struct virtio_net_ctrl_mac) + 2582df4dd31SEugenio Pérez MAC_TABLE_ENTRIES * ETH_ALEN; 2592df4dd31SEugenio Pérez } 2602df4dd31SEugenio Pérez 2612df4dd31SEugenio Pérez static size_t vhost_vdpa_net_cvq_cmd_page_len(void) 2622df4dd31SEugenio Pérez { 2632df4dd31SEugenio Pérez return ROUND_UP(vhost_vdpa_net_cvq_cmd_len(), qemu_real_host_page_size()); 2642df4dd31SEugenio Pérez } 2652df4dd31SEugenio Pérez 2667a7f87e9SEugenio Pérez /** Map CVQ buffer. */ 2677a7f87e9SEugenio Pérez static int vhost_vdpa_cvq_map_buf(struct vhost_vdpa *v, void *buf, size_t size, 2687a7f87e9SEugenio Pérez bool write) 2692df4dd31SEugenio Pérez { 2702df4dd31SEugenio Pérez DMAMap map = {}; 2712df4dd31SEugenio Pérez int r; 2722df4dd31SEugenio Pérez 2732df4dd31SEugenio Pérez map.translated_addr = (hwaddr)(uintptr_t)buf; 2747a7f87e9SEugenio Pérez map.size = size - 1; 2752df4dd31SEugenio Pérez map.perm = write ? IOMMU_RW : IOMMU_RO, 2762df4dd31SEugenio Pérez r = vhost_iova_tree_map_alloc(v->iova_tree, &map); 2772df4dd31SEugenio Pérez if (unlikely(r != IOVA_OK)) { 2782df4dd31SEugenio Pérez error_report("Cannot map injected element"); 2797a7f87e9SEugenio Pérez return r; 2802df4dd31SEugenio Pérez } 2812df4dd31SEugenio Pérez 2822df4dd31SEugenio Pérez r = vhost_vdpa_dma_map(v, map.iova, vhost_vdpa_net_cvq_cmd_page_len(), buf, 2832df4dd31SEugenio Pérez !write); 2842df4dd31SEugenio Pérez if (unlikely(r < 0)) { 2852df4dd31SEugenio Pérez goto dma_map_err; 2862df4dd31SEugenio Pérez } 2872df4dd31SEugenio Pérez 2887a7f87e9SEugenio Pérez return 0; 2892df4dd31SEugenio Pérez 2902df4dd31SEugenio Pérez dma_map_err: 29169292a8eSEugenio Pérez vhost_iova_tree_remove(v->iova_tree, map); 2927a7f87e9SEugenio Pérez return r; 2932df4dd31SEugenio Pérez } 2942df4dd31SEugenio Pérez 2957a7f87e9SEugenio Pérez static int vhost_vdpa_net_cvq_start(NetClientState *nc) 2962df4dd31SEugenio Pérez { 2977a7f87e9SEugenio Pérez VhostVDPAState *s; 2987a7f87e9SEugenio Pérez int r; 2992df4dd31SEugenio Pérez 3007a7f87e9SEugenio Pérez assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 3017a7f87e9SEugenio Pérez 3027a7f87e9SEugenio Pérez s = DO_UPCAST(VhostVDPAState, nc, nc); 3037a7f87e9SEugenio Pérez if (!s->vhost_vdpa.shadow_vqs_enabled) { 3047a7f87e9SEugenio Pérez return 0; 3052df4dd31SEugenio Pérez } 3062df4dd31SEugenio Pérez 3077a7f87e9SEugenio Pérez r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer, 3087a7f87e9SEugenio Pérez vhost_vdpa_net_cvq_cmd_page_len(), false); 3097a7f87e9SEugenio Pérez if (unlikely(r < 0)) { 3107a7f87e9SEugenio Pérez return r; 3117a7f87e9SEugenio Pérez } 3127a7f87e9SEugenio Pérez 3137a7f87e9SEugenio Pérez r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->cvq_cmd_in_buffer, 3147a7f87e9SEugenio Pérez vhost_vdpa_net_cvq_cmd_page_len(), true); 3157a7f87e9SEugenio Pérez if (unlikely(r < 0)) { 3162df4dd31SEugenio Pérez vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer); 3172df4dd31SEugenio Pérez } 3182df4dd31SEugenio Pérez 3197a7f87e9SEugenio Pérez return r; 3207a7f87e9SEugenio Pérez } 3217a7f87e9SEugenio Pérez 3227a7f87e9SEugenio Pérez static void vhost_vdpa_net_cvq_stop(NetClientState *nc) 3237a7f87e9SEugenio Pérez { 3247a7f87e9SEugenio Pérez VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 3257a7f87e9SEugenio Pérez 3267a7f87e9SEugenio Pérez assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 3277a7f87e9SEugenio Pérez 3287a7f87e9SEugenio Pérez if (s->vhost_vdpa.shadow_vqs_enabled) { 3297a7f87e9SEugenio Pérez vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer); 3307a7f87e9SEugenio Pérez vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_in_buffer); 3317a7f87e9SEugenio Pérez } 3322df4dd31SEugenio Pérez } 3332df4dd31SEugenio Pérez 334*be4278b6SEugenio Pérez static ssize_t vhost_vdpa_net_cvq_add(VhostVDPAState *s, size_t out_len, 335*be4278b6SEugenio Pérez size_t in_len) 336*be4278b6SEugenio Pérez { 337*be4278b6SEugenio Pérez /* Buffers for the device */ 338*be4278b6SEugenio Pérez const struct iovec out = { 339*be4278b6SEugenio Pérez .iov_base = s->cvq_cmd_out_buffer, 340*be4278b6SEugenio Pérez .iov_len = out_len, 341*be4278b6SEugenio Pérez }; 342*be4278b6SEugenio Pérez const struct iovec in = { 343*be4278b6SEugenio Pérez .iov_base = s->cvq_cmd_in_buffer, 344*be4278b6SEugenio Pérez .iov_len = sizeof(virtio_net_ctrl_ack), 345*be4278b6SEugenio Pérez }; 346*be4278b6SEugenio Pérez VhostShadowVirtqueue *svq = g_ptr_array_index(s->vhost_vdpa.shadow_vqs, 0); 347*be4278b6SEugenio Pérez int r; 348*be4278b6SEugenio Pérez 349*be4278b6SEugenio Pérez r = vhost_svq_add(svq, &out, 1, &in, 1, NULL); 350*be4278b6SEugenio Pérez if (unlikely(r != 0)) { 351*be4278b6SEugenio Pérez if (unlikely(r == -ENOSPC)) { 352*be4278b6SEugenio Pérez qemu_log_mask(LOG_GUEST_ERROR, "%s: No space on device queue\n", 353*be4278b6SEugenio Pérez __func__); 354*be4278b6SEugenio Pérez } 355*be4278b6SEugenio Pérez return r; 356*be4278b6SEugenio Pérez } 357*be4278b6SEugenio Pérez 358*be4278b6SEugenio Pérez /* 359*be4278b6SEugenio Pérez * We can poll here since we've had BQL from the time we sent the 360*be4278b6SEugenio Pérez * descriptor. Also, we need to take the answer before SVQ pulls by itself, 361*be4278b6SEugenio Pérez * when BQL is released 362*be4278b6SEugenio Pérez */ 363*be4278b6SEugenio Pérez return vhost_svq_poll(svq); 364*be4278b6SEugenio Pérez } 365*be4278b6SEugenio Pérez 366f8972b56SEugenio Pérez static NetClientInfo net_vhost_vdpa_cvq_info = { 367f8972b56SEugenio Pérez .type = NET_CLIENT_DRIVER_VHOST_VDPA, 368f8972b56SEugenio Pérez .size = sizeof(VhostVDPAState), 369f8972b56SEugenio Pérez .receive = vhost_vdpa_receive, 3707a7f87e9SEugenio Pérez .start = vhost_vdpa_net_cvq_start, 3717a7f87e9SEugenio Pérez .stop = vhost_vdpa_net_cvq_stop, 372f8972b56SEugenio Pérez .cleanup = vhost_vdpa_cleanup, 373f8972b56SEugenio Pérez .has_vnet_hdr = vhost_vdpa_has_vnet_hdr, 374f8972b56SEugenio Pérez .has_ufo = vhost_vdpa_has_ufo, 375f8972b56SEugenio Pérez .check_peer_type = vhost_vdpa_check_peer_type, 376f8972b56SEugenio Pérez }; 377f8972b56SEugenio Pérez 3782df4dd31SEugenio Pérez /** 3792df4dd31SEugenio Pérez * Do not forward commands not supported by SVQ. Otherwise, the device could 3802df4dd31SEugenio Pérez * accept it and qemu would not know how to update the device model. 3812df4dd31SEugenio Pérez */ 3827a7f87e9SEugenio Pérez static bool vhost_vdpa_net_cvq_validate_cmd(const void *out_buf, size_t len) 3832df4dd31SEugenio Pérez { 3842df4dd31SEugenio Pérez struct virtio_net_ctrl_hdr ctrl; 3852df4dd31SEugenio Pérez 3867a7f87e9SEugenio Pérez if (unlikely(len < sizeof(ctrl))) { 3872df4dd31SEugenio Pérez qemu_log_mask(LOG_GUEST_ERROR, 3887a7f87e9SEugenio Pérez "%s: invalid legnth of out buffer %zu\n", __func__, len); 3892df4dd31SEugenio Pérez return false; 3902df4dd31SEugenio Pérez } 3912df4dd31SEugenio Pérez 3927a7f87e9SEugenio Pérez memcpy(&ctrl, out_buf, sizeof(ctrl)); 3932df4dd31SEugenio Pérez switch (ctrl.class) { 3942df4dd31SEugenio Pérez case VIRTIO_NET_CTRL_MAC: 3952df4dd31SEugenio Pérez switch (ctrl.cmd) { 3962df4dd31SEugenio Pérez case VIRTIO_NET_CTRL_MAC_ADDR_SET: 3972df4dd31SEugenio Pérez return true; 3982df4dd31SEugenio Pérez default: 3992df4dd31SEugenio Pérez qemu_log_mask(LOG_GUEST_ERROR, "%s: invalid mac cmd %u\n", 4002df4dd31SEugenio Pérez __func__, ctrl.cmd); 4012df4dd31SEugenio Pérez }; 4022df4dd31SEugenio Pérez break; 4032df4dd31SEugenio Pérez default: 4042df4dd31SEugenio Pérez qemu_log_mask(LOG_GUEST_ERROR, "%s: invalid control class %u\n", 4052df4dd31SEugenio Pérez __func__, ctrl.class); 4062df4dd31SEugenio Pérez }; 4072df4dd31SEugenio Pérez 4082df4dd31SEugenio Pérez return false; 4092df4dd31SEugenio Pérez } 4102df4dd31SEugenio Pérez 4112df4dd31SEugenio Pérez /** 4122df4dd31SEugenio Pérez * Validate and copy control virtqueue commands. 4132df4dd31SEugenio Pérez * 4142df4dd31SEugenio Pérez * Following QEMU guidelines, we offer a copy of the buffers to the device to 4152df4dd31SEugenio Pérez * prevent TOCTOU bugs. 416bd907ae4SEugenio Pérez */ 417bd907ae4SEugenio Pérez static int vhost_vdpa_net_handle_ctrl_avail(VhostShadowVirtqueue *svq, 418bd907ae4SEugenio Pérez VirtQueueElement *elem, 419bd907ae4SEugenio Pérez void *opaque) 420bd907ae4SEugenio Pérez { 4212df4dd31SEugenio Pérez VhostVDPAState *s = opaque; 422*be4278b6SEugenio Pérez size_t in_len; 423bd907ae4SEugenio Pérez virtio_net_ctrl_ack status = VIRTIO_NET_ERR; 4247a7f87e9SEugenio Pérez /* Out buffer sent to both the vdpa device and the device model */ 4257a7f87e9SEugenio Pérez struct iovec out = { 4267a7f87e9SEugenio Pérez .iov_base = s->cvq_cmd_out_buffer, 4277a7f87e9SEugenio Pérez }; 4282df4dd31SEugenio Pérez /* in buffer used for device model */ 4292df4dd31SEugenio Pérez const struct iovec in = { 4302df4dd31SEugenio Pérez .iov_base = &status, 4312df4dd31SEugenio Pérez .iov_len = sizeof(status), 4322df4dd31SEugenio Pérez }; 433*be4278b6SEugenio Pérez ssize_t dev_written = -EINVAL; 4342df4dd31SEugenio Pérez bool ok; 435bd907ae4SEugenio Pérez 4367a7f87e9SEugenio Pérez out.iov_len = iov_to_buf(elem->out_sg, elem->out_num, 0, 4377a7f87e9SEugenio Pérez s->cvq_cmd_out_buffer, 4387a7f87e9SEugenio Pérez vhost_vdpa_net_cvq_cmd_len()); 4397a7f87e9SEugenio Pérez ok = vhost_vdpa_net_cvq_validate_cmd(s->cvq_cmd_out_buffer, out.iov_len); 4402df4dd31SEugenio Pérez if (unlikely(!ok)) { 4412df4dd31SEugenio Pérez goto out; 4422df4dd31SEugenio Pérez } 443bd907ae4SEugenio Pérez 444*be4278b6SEugenio Pérez dev_written = vhost_vdpa_net_cvq_add(s, out.iov_len, sizeof(status)); 445*be4278b6SEugenio Pérez if (unlikely(dev_written < 0)) { 446bd907ae4SEugenio Pérez goto out; 447bd907ae4SEugenio Pérez } 448bd907ae4SEugenio Pérez 449bd907ae4SEugenio Pérez if (unlikely(dev_written < sizeof(status))) { 450bd907ae4SEugenio Pérez error_report("Insufficient written data (%zu)", dev_written); 4512df4dd31SEugenio Pérez goto out; 4522df4dd31SEugenio Pérez } 4532df4dd31SEugenio Pérez 4547a7f87e9SEugenio Pérez memcpy(&status, s->cvq_cmd_in_buffer, sizeof(status)); 4552df4dd31SEugenio Pérez if (status != VIRTIO_NET_OK) { 456*be4278b6SEugenio Pérez return VIRTIO_NET_ERR; 4572df4dd31SEugenio Pérez } 4582df4dd31SEugenio Pérez 4592df4dd31SEugenio Pérez status = VIRTIO_NET_ERR; 4607a7f87e9SEugenio Pérez virtio_net_handle_ctrl_iov(svq->vdev, &in, 1, &out, 1); 4612df4dd31SEugenio Pérez if (status != VIRTIO_NET_OK) { 4622df4dd31SEugenio Pérez error_report("Bad CVQ processing in model"); 463bd907ae4SEugenio Pérez } 464bd907ae4SEugenio Pérez 465bd907ae4SEugenio Pérez out: 466bd907ae4SEugenio Pérez in_len = iov_from_buf(elem->in_sg, elem->in_num, 0, &status, 467bd907ae4SEugenio Pérez sizeof(status)); 468bd907ae4SEugenio Pérez if (unlikely(in_len < sizeof(status))) { 469bd907ae4SEugenio Pérez error_report("Bad device CVQ written length"); 470bd907ae4SEugenio Pérez } 471bd907ae4SEugenio Pérez vhost_svq_push_elem(svq, elem, MIN(in_len, sizeof(status))); 472bd907ae4SEugenio Pérez g_free(elem); 473*be4278b6SEugenio Pérez return dev_written < 0 ? dev_written : 0; 474bd907ae4SEugenio Pérez } 475bd907ae4SEugenio Pérez 476bd907ae4SEugenio Pérez static const VhostShadowVirtqueueOps vhost_vdpa_net_svq_ops = { 477bd907ae4SEugenio Pérez .avail_handler = vhost_vdpa_net_handle_ctrl_avail, 478bd907ae4SEugenio Pérez }; 479bd907ae4SEugenio Pérez 480654790b6SJason Wang static NetClientState *net_vhost_vdpa_init(NetClientState *peer, 481654790b6SJason Wang const char *device, 482654790b6SJason Wang const char *name, 48340237840SJason Wang int vdpa_device_fd, 48440237840SJason Wang int queue_pair_index, 48540237840SJason Wang int nvqs, 4861576dbb5SEugenio Pérez bool is_datapath, 4871576dbb5SEugenio Pérez bool svq, 4881576dbb5SEugenio Pérez VhostIOVATree *iova_tree) 4891e0a84eaSCindy Lu { 4901e0a84eaSCindy Lu NetClientState *nc = NULL; 4911e0a84eaSCindy Lu VhostVDPAState *s; 4921e0a84eaSCindy Lu int ret = 0; 4931e0a84eaSCindy Lu assert(name); 49440237840SJason Wang if (is_datapath) { 49540237840SJason Wang nc = qemu_new_net_client(&net_vhost_vdpa_info, peer, device, 49640237840SJason Wang name); 49740237840SJason Wang } else { 498f8972b56SEugenio Pérez nc = qemu_new_net_control_client(&net_vhost_vdpa_cvq_info, peer, 49940237840SJason Wang device, name); 50040237840SJason Wang } 50156e6f594SJason Wang snprintf(nc->info_str, sizeof(nc->info_str), TYPE_VHOST_VDPA); 5021e0a84eaSCindy Lu s = DO_UPCAST(VhostVDPAState, nc, nc); 5037327813dSJason Wang 5041e0a84eaSCindy Lu s->vhost_vdpa.device_fd = vdpa_device_fd; 50540237840SJason Wang s->vhost_vdpa.index = queue_pair_index; 5061576dbb5SEugenio Pérez s->vhost_vdpa.shadow_vqs_enabled = svq; 5071576dbb5SEugenio Pérez s->vhost_vdpa.iova_tree = iova_tree; 508bd907ae4SEugenio Pérez if (!is_datapath) { 5092df4dd31SEugenio Pérez s->cvq_cmd_out_buffer = qemu_memalign(qemu_real_host_page_size(), 5102df4dd31SEugenio Pérez vhost_vdpa_net_cvq_cmd_page_len()); 5112df4dd31SEugenio Pérez memset(s->cvq_cmd_out_buffer, 0, vhost_vdpa_net_cvq_cmd_page_len()); 5122df4dd31SEugenio Pérez s->cvq_cmd_in_buffer = qemu_memalign(qemu_real_host_page_size(), 5132df4dd31SEugenio Pérez vhost_vdpa_net_cvq_cmd_page_len()); 5142df4dd31SEugenio Pérez memset(s->cvq_cmd_in_buffer, 0, vhost_vdpa_net_cvq_cmd_page_len()); 5152df4dd31SEugenio Pérez 516bd907ae4SEugenio Pérez s->vhost_vdpa.shadow_vq_ops = &vhost_vdpa_net_svq_ops; 517bd907ae4SEugenio Pérez s->vhost_vdpa.shadow_vq_ops_opaque = s; 5181576dbb5SEugenio Pérez error_setg(&s->vhost_vdpa.migration_blocker, 5191576dbb5SEugenio Pérez "Migration disabled: vhost-vdpa uses CVQ."); 520bd907ae4SEugenio Pérez } 52140237840SJason Wang ret = vhost_vdpa_add(nc, (void *)&s->vhost_vdpa, queue_pair_index, nvqs); 52274af5eecSJason Wang if (ret) { 52374af5eecSJason Wang qemu_del_net_client(nc); 524654790b6SJason Wang return NULL; 52574af5eecSJason Wang } 526654790b6SJason Wang return nc; 5271e0a84eaSCindy Lu } 5281e0a84eaSCindy Lu 5291576dbb5SEugenio Pérez static int vhost_vdpa_get_iova_range(int fd, 5301576dbb5SEugenio Pérez struct vhost_vdpa_iova_range *iova_range) 5311576dbb5SEugenio Pérez { 5321576dbb5SEugenio Pérez int ret = ioctl(fd, VHOST_VDPA_GET_IOVA_RANGE, iova_range); 5331576dbb5SEugenio Pérez 5341576dbb5SEugenio Pérez return ret < 0 ? -errno : 0; 5351576dbb5SEugenio Pérez } 5361576dbb5SEugenio Pérez 5378170ab3fSEugenio Pérez static int vhost_vdpa_get_features(int fd, uint64_t *features, Error **errp) 5388170ab3fSEugenio Pérez { 5398170ab3fSEugenio Pérez int ret = ioctl(fd, VHOST_GET_FEATURES, features); 5408170ab3fSEugenio Pérez if (unlikely(ret < 0)) { 5418170ab3fSEugenio Pérez error_setg_errno(errp, errno, 5428170ab3fSEugenio Pérez "Fail to query features from vhost-vDPA device"); 5438170ab3fSEugenio Pérez } 5448170ab3fSEugenio Pérez return ret; 5458170ab3fSEugenio Pérez } 5468170ab3fSEugenio Pérez 5478170ab3fSEugenio Pérez static int vhost_vdpa_get_max_queue_pairs(int fd, uint64_t features, 5488170ab3fSEugenio Pérez int *has_cvq, Error **errp) 54940237840SJason Wang { 55040237840SJason Wang unsigned long config_size = offsetof(struct vhost_vdpa_config, buf); 551cd523a41SStefano Garzarella g_autofree struct vhost_vdpa_config *config = NULL; 55240237840SJason Wang __virtio16 *max_queue_pairs; 55340237840SJason Wang int ret; 55440237840SJason Wang 55540237840SJason Wang if (features & (1 << VIRTIO_NET_F_CTRL_VQ)) { 55640237840SJason Wang *has_cvq = 1; 55740237840SJason Wang } else { 55840237840SJason Wang *has_cvq = 0; 55940237840SJason Wang } 56040237840SJason Wang 56140237840SJason Wang if (features & (1 << VIRTIO_NET_F_MQ)) { 56240237840SJason Wang config = g_malloc0(config_size + sizeof(*max_queue_pairs)); 56340237840SJason Wang config->off = offsetof(struct virtio_net_config, max_virtqueue_pairs); 56440237840SJason Wang config->len = sizeof(*max_queue_pairs); 56540237840SJason Wang 56640237840SJason Wang ret = ioctl(fd, VHOST_VDPA_GET_CONFIG, config); 56740237840SJason Wang if (ret) { 56840237840SJason Wang error_setg(errp, "Fail to get config from vhost-vDPA device"); 56940237840SJason Wang return -ret; 57040237840SJason Wang } 57140237840SJason Wang 57240237840SJason Wang max_queue_pairs = (__virtio16 *)&config->buf; 57340237840SJason Wang 57440237840SJason Wang return lduw_le_p(max_queue_pairs); 57540237840SJason Wang } 57640237840SJason Wang 57740237840SJason Wang return 1; 57840237840SJason Wang } 57940237840SJason Wang 5801e0a84eaSCindy Lu int net_init_vhost_vdpa(const Netdev *netdev, const char *name, 5811e0a84eaSCindy Lu NetClientState *peer, Error **errp) 5821e0a84eaSCindy Lu { 5831e0a84eaSCindy Lu const NetdevVhostVDPAOptions *opts; 5848170ab3fSEugenio Pérez uint64_t features; 585654790b6SJason Wang int vdpa_device_fd; 586eb3cb751SEugenio Pérez g_autofree NetClientState **ncs = NULL; 5871576dbb5SEugenio Pérez g_autoptr(VhostIOVATree) iova_tree = NULL; 588eb3cb751SEugenio Pérez NetClientState *nc; 589aed5da45SEugenio Pérez int queue_pairs, r, i = 0, has_cvq = 0; 5901e0a84eaSCindy Lu 5911e0a84eaSCindy Lu assert(netdev->type == NET_CLIENT_DRIVER_VHOST_VDPA); 5921e0a84eaSCindy Lu opts = &netdev->u.vhost_vdpa; 593c8295404SEugenio Pérez if (!opts->vhostdev) { 594c8295404SEugenio Pérez error_setg(errp, "vdpa character device not specified with vhostdev"); 595c8295404SEugenio Pérez return -1; 596c8295404SEugenio Pérez } 5977327813dSJason Wang 5980351152bSEugenio Pérez vdpa_device_fd = qemu_open(opts->vhostdev, O_RDWR, errp); 5997327813dSJason Wang if (vdpa_device_fd == -1) { 6007327813dSJason Wang return -errno; 6017327813dSJason Wang } 6027327813dSJason Wang 6038170ab3fSEugenio Pérez r = vhost_vdpa_get_features(vdpa_device_fd, &features, errp); 6048170ab3fSEugenio Pérez if (unlikely(r < 0)) { 605aed5da45SEugenio Pérez goto err; 6068170ab3fSEugenio Pérez } 6078170ab3fSEugenio Pérez 6088170ab3fSEugenio Pérez queue_pairs = vhost_vdpa_get_max_queue_pairs(vdpa_device_fd, features, 60940237840SJason Wang &has_cvq, errp); 61040237840SJason Wang if (queue_pairs < 0) { 6117327813dSJason Wang qemu_close(vdpa_device_fd); 61240237840SJason Wang return queue_pairs; 6137327813dSJason Wang } 6147327813dSJason Wang 6151576dbb5SEugenio Pérez if (opts->x_svq) { 6161576dbb5SEugenio Pérez struct vhost_vdpa_iova_range iova_range; 6171576dbb5SEugenio Pérez 6181576dbb5SEugenio Pérez uint64_t invalid_dev_features = 6191576dbb5SEugenio Pérez features & ~vdpa_svq_device_features & 6201576dbb5SEugenio Pérez /* Transport are all accepted at this point */ 6211576dbb5SEugenio Pérez ~MAKE_64BIT_MASK(VIRTIO_TRANSPORT_F_START, 6221576dbb5SEugenio Pérez VIRTIO_TRANSPORT_F_END - VIRTIO_TRANSPORT_F_START); 6231576dbb5SEugenio Pérez 6241576dbb5SEugenio Pérez if (invalid_dev_features) { 6251576dbb5SEugenio Pérez error_setg(errp, "vdpa svq does not work with features 0x%" PRIx64, 6261576dbb5SEugenio Pérez invalid_dev_features); 6271576dbb5SEugenio Pérez goto err_svq; 6281576dbb5SEugenio Pérez } 6291576dbb5SEugenio Pérez 6301576dbb5SEugenio Pérez vhost_vdpa_get_iova_range(vdpa_device_fd, &iova_range); 6311576dbb5SEugenio Pérez iova_tree = vhost_iova_tree_new(iova_range.first, iova_range.last); 6321576dbb5SEugenio Pérez } 6331576dbb5SEugenio Pérez 63440237840SJason Wang ncs = g_malloc0(sizeof(*ncs) * queue_pairs); 63540237840SJason Wang 63640237840SJason Wang for (i = 0; i < queue_pairs; i++) { 63740237840SJason Wang ncs[i] = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name, 6381576dbb5SEugenio Pérez vdpa_device_fd, i, 2, true, opts->x_svq, 6391576dbb5SEugenio Pérez iova_tree); 64040237840SJason Wang if (!ncs[i]) 64140237840SJason Wang goto err; 64240237840SJason Wang } 64340237840SJason Wang 64440237840SJason Wang if (has_cvq) { 64540237840SJason Wang nc = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name, 6461576dbb5SEugenio Pérez vdpa_device_fd, i, 1, false, 6471576dbb5SEugenio Pérez opts->x_svq, iova_tree); 64840237840SJason Wang if (!nc) 64940237840SJason Wang goto err; 65040237840SJason Wang } 65140237840SJason Wang 6521576dbb5SEugenio Pérez /* iova_tree ownership belongs to last NetClientState */ 6531576dbb5SEugenio Pérez g_steal_pointer(&iova_tree); 654654790b6SJason Wang return 0; 65540237840SJason Wang 65640237840SJason Wang err: 65740237840SJason Wang if (i) { 6589bd05507SSi-Wei Liu for (i--; i >= 0; i--) { 6599bd05507SSi-Wei Liu qemu_del_net_client(ncs[i]); 6609bd05507SSi-Wei Liu } 66140237840SJason Wang } 6621576dbb5SEugenio Pérez 6631576dbb5SEugenio Pérez err_svq: 66440237840SJason Wang qemu_close(vdpa_device_fd); 66540237840SJason Wang 66640237840SJason Wang return -1; 6671e0a84eaSCindy Lu } 668