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" 2969498430SEugenio Pérez #include "migration/migration.h" 3069498430SEugenio Pérez #include "migration/misc.h" 311e0a84eaSCindy Lu #include "hw/virtio/vhost.h" 321e0a84eaSCindy Lu 331e0a84eaSCindy Lu /* Todo:need to add the multiqueue support here */ 341e0a84eaSCindy Lu typedef struct VhostVDPAState { 351e0a84eaSCindy Lu NetClientState nc; 361e0a84eaSCindy Lu struct vhost_vdpa vhost_vdpa; 3769498430SEugenio Pérez Notifier migration_state; 381e0a84eaSCindy Lu VHostNetState *vhost_net; 392df4dd31SEugenio Pérez 402df4dd31SEugenio Pérez /* Control commands shadow buffers */ 4117fb889fSEugenio Pérez void *cvq_cmd_out_buffer; 4217fb889fSEugenio Pérez virtio_net_ctrl_ack *status; 4317fb889fSEugenio Pérez 447f211a28SEugenio Pérez /* The device always have SVQ enabled */ 457f211a28SEugenio Pérez bool always_svq; 46152128d6SEugenio Pérez 47152128d6SEugenio Pérez /* The device can isolate CVQ in its own ASID */ 48152128d6SEugenio Pérez bool cvq_isolated; 49152128d6SEugenio Pérez 501e0a84eaSCindy Lu bool started; 511e0a84eaSCindy Lu } VhostVDPAState; 521e0a84eaSCindy Lu 532875a0caSHawkins Jiawei /* 542875a0caSHawkins Jiawei * The array is sorted alphabetically in ascending order, 552875a0caSHawkins Jiawei * with the exception of VHOST_INVALID_FEATURE_BIT, 562875a0caSHawkins Jiawei * which should always be the last entry. 572875a0caSHawkins Jiawei */ 581e0a84eaSCindy Lu const int vdpa_feature_bits[] = { 591e0a84eaSCindy Lu VIRTIO_F_ANY_LAYOUT, 602875a0caSHawkins Jiawei VIRTIO_F_IOMMU_PLATFORM, 612875a0caSHawkins Jiawei VIRTIO_F_NOTIFY_ON_EMPTY, 622875a0caSHawkins Jiawei VIRTIO_F_RING_PACKED, 632875a0caSHawkins Jiawei VIRTIO_F_RING_RESET, 641e0a84eaSCindy Lu VIRTIO_F_VERSION_1, 651e0a84eaSCindy Lu VIRTIO_NET_F_CSUM, 6651e84244SEugenio Pérez VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, 672875a0caSHawkins Jiawei VIRTIO_NET_F_CTRL_MAC_ADDR, 6840237840SJason Wang VIRTIO_NET_F_CTRL_RX, 6940237840SJason Wang VIRTIO_NET_F_CTRL_RX_EXTRA, 7040237840SJason Wang VIRTIO_NET_F_CTRL_VLAN, 7140237840SJason Wang VIRTIO_NET_F_CTRL_VQ, 722875a0caSHawkins Jiawei VIRTIO_NET_F_GSO, 732875a0caSHawkins Jiawei VIRTIO_NET_F_GUEST_CSUM, 742875a0caSHawkins Jiawei VIRTIO_NET_F_GUEST_ECN, 752875a0caSHawkins Jiawei VIRTIO_NET_F_GUEST_TSO4, 762875a0caSHawkins Jiawei VIRTIO_NET_F_GUEST_TSO6, 772875a0caSHawkins Jiawei VIRTIO_NET_F_GUEST_UFO, 780145c393SAndrew Melnychenko VIRTIO_NET_F_HASH_REPORT, 792875a0caSHawkins Jiawei VIRTIO_NET_F_HOST_ECN, 802875a0caSHawkins Jiawei VIRTIO_NET_F_HOST_TSO4, 812875a0caSHawkins Jiawei VIRTIO_NET_F_HOST_TSO6, 822875a0caSHawkins Jiawei VIRTIO_NET_F_HOST_UFO, 832875a0caSHawkins Jiawei VIRTIO_NET_F_MQ, 842875a0caSHawkins Jiawei VIRTIO_NET_F_MRG_RXBUF, 852875a0caSHawkins Jiawei VIRTIO_NET_F_MTU, 862875a0caSHawkins Jiawei VIRTIO_NET_F_RSS, 879aa47eddSSi-Wei Liu VIRTIO_NET_F_STATUS, 882875a0caSHawkins Jiawei VIRTIO_RING_F_EVENT_IDX, 892875a0caSHawkins Jiawei VIRTIO_RING_F_INDIRECT_DESC, 902875a0caSHawkins Jiawei 912875a0caSHawkins Jiawei /* VHOST_INVALID_FEATURE_BIT should always be the last entry */ 921e0a84eaSCindy Lu VHOST_INVALID_FEATURE_BIT 931e0a84eaSCindy Lu }; 941e0a84eaSCindy Lu 951576dbb5SEugenio Pérez /** Supported device specific feature bits with SVQ */ 961576dbb5SEugenio Pérez static const uint64_t vdpa_svq_device_features = 971576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_CSUM) | 981576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_GUEST_CSUM) | 994b4a1378SHawkins Jiawei BIT_ULL(VIRTIO_NET_F_CTRL_GUEST_OFFLOADS) | 1001576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_MTU) | 1011576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_MAC) | 1021576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_GUEST_TSO4) | 1031576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_GUEST_TSO6) | 1041576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_GUEST_ECN) | 1051576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_GUEST_UFO) | 1061576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_HOST_TSO4) | 1071576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_HOST_TSO6) | 1081576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_HOST_ECN) | 1091576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_HOST_UFO) | 1101576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_MRG_RXBUF) | 1111576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_STATUS) | 1121576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_CTRL_VQ) | 11372b99a87SEugenio Pérez BIT_ULL(VIRTIO_NET_F_MQ) | 1141576dbb5SEugenio Pérez BIT_ULL(VIRTIO_F_ANY_LAYOUT) | 1151576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_CTRL_MAC_ADDR) | 116609ab4c3SEugenio Pérez /* VHOST_F_LOG_ALL is exposed by SVQ */ 117609ab4c3SEugenio Pérez BIT_ULL(VHOST_F_LOG_ALL) | 1181576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_RSC_EXT) | 1190d74e2b7SEugenio Pérez BIT_ULL(VIRTIO_NET_F_STANDBY) | 1200d74e2b7SEugenio Pérez BIT_ULL(VIRTIO_NET_F_SPEED_DUPLEX); 1211576dbb5SEugenio Pérez 122c1a10086SEugenio Pérez #define VHOST_VDPA_NET_CVQ_ASID 1 123c1a10086SEugenio Pérez 1241e0a84eaSCindy Lu VHostNetState *vhost_vdpa_get_vhost_net(NetClientState *nc) 1251e0a84eaSCindy Lu { 1261e0a84eaSCindy Lu VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 1271e0a84eaSCindy Lu assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 1281e0a84eaSCindy Lu return s->vhost_net; 1291e0a84eaSCindy Lu } 1301e0a84eaSCindy Lu 131915bf6ccSEugenio Pérez static size_t vhost_vdpa_net_cvq_cmd_len(void) 132915bf6ccSEugenio Pérez { 133915bf6ccSEugenio Pérez /* 134915bf6ccSEugenio Pérez * MAC_TABLE_SET is the ctrl command that produces the longer out buffer. 135915bf6ccSEugenio Pérez * In buffer is always 1 byte, so it should fit here 136915bf6ccSEugenio Pérez */ 137915bf6ccSEugenio Pérez return sizeof(struct virtio_net_ctrl_hdr) + 138915bf6ccSEugenio Pérez 2 * sizeof(struct virtio_net_ctrl_mac) + 139915bf6ccSEugenio Pérez MAC_TABLE_ENTRIES * ETH_ALEN; 140915bf6ccSEugenio Pérez } 141915bf6ccSEugenio Pérez 142915bf6ccSEugenio Pérez static size_t vhost_vdpa_net_cvq_cmd_page_len(void) 143915bf6ccSEugenio Pérez { 144915bf6ccSEugenio Pérez return ROUND_UP(vhost_vdpa_net_cvq_cmd_len(), qemu_real_host_page_size()); 145915bf6ccSEugenio Pérez } 146915bf6ccSEugenio Pérez 14736e46472SEugenio Pérez static bool vhost_vdpa_net_valid_svq_features(uint64_t features, Error **errp) 14836e46472SEugenio Pérez { 14936e46472SEugenio Pérez uint64_t invalid_dev_features = 15036e46472SEugenio Pérez features & ~vdpa_svq_device_features & 15136e46472SEugenio Pérez /* Transport are all accepted at this point */ 15236e46472SEugenio Pérez ~MAKE_64BIT_MASK(VIRTIO_TRANSPORT_F_START, 15336e46472SEugenio Pérez VIRTIO_TRANSPORT_F_END - VIRTIO_TRANSPORT_F_START); 15436e46472SEugenio Pérez 15536e46472SEugenio Pérez if (invalid_dev_features) { 15636e46472SEugenio Pérez error_setg(errp, "vdpa svq does not work with features 0x%" PRIx64, 15736e46472SEugenio Pérez invalid_dev_features); 158258a0394SEugenio Pérez return false; 15936e46472SEugenio Pérez } 16036e46472SEugenio Pérez 161258a0394SEugenio Pérez return vhost_svq_valid_features(features, errp); 16236e46472SEugenio Pérez } 16336e46472SEugenio Pérez 1641e0a84eaSCindy Lu static int vhost_vdpa_net_check_device_id(struct vhost_net *net) 1651e0a84eaSCindy Lu { 1661e0a84eaSCindy Lu uint32_t device_id; 1671e0a84eaSCindy Lu int ret; 1681e0a84eaSCindy Lu struct vhost_dev *hdev; 1691e0a84eaSCindy Lu 1701e0a84eaSCindy Lu hdev = (struct vhost_dev *)&net->dev; 1711e0a84eaSCindy Lu ret = hdev->vhost_ops->vhost_get_device_id(hdev, &device_id); 1721e0a84eaSCindy Lu if (device_id != VIRTIO_ID_NET) { 1731e0a84eaSCindy Lu return -ENOTSUP; 1741e0a84eaSCindy Lu } 1751e0a84eaSCindy Lu return ret; 1761e0a84eaSCindy Lu } 1771e0a84eaSCindy Lu 17840237840SJason Wang static int vhost_vdpa_add(NetClientState *ncs, void *be, 17940237840SJason Wang int queue_pair_index, int nvqs) 1801e0a84eaSCindy Lu { 1811e0a84eaSCindy Lu VhostNetOptions options; 1821e0a84eaSCindy Lu struct vhost_net *net = NULL; 1831e0a84eaSCindy Lu VhostVDPAState *s; 1841e0a84eaSCindy Lu int ret; 1851e0a84eaSCindy Lu 1861e0a84eaSCindy Lu options.backend_type = VHOST_BACKEND_TYPE_VDPA; 1871e0a84eaSCindy Lu assert(ncs->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 1881e0a84eaSCindy Lu s = DO_UPCAST(VhostVDPAState, nc, ncs); 1891e0a84eaSCindy Lu options.net_backend = ncs; 1901e0a84eaSCindy Lu options.opaque = be; 1911e0a84eaSCindy Lu options.busyloop_timeout = 0; 19240237840SJason Wang options.nvqs = nvqs; 1931e0a84eaSCindy Lu 1941e0a84eaSCindy Lu net = vhost_net_init(&options); 1951e0a84eaSCindy Lu if (!net) { 1961e0a84eaSCindy Lu error_report("failed to init vhost_net for queue"); 197a97ef87aSJason Wang goto err_init; 1981e0a84eaSCindy Lu } 1991e0a84eaSCindy Lu s->vhost_net = net; 2001e0a84eaSCindy Lu ret = vhost_vdpa_net_check_device_id(net); 2011e0a84eaSCindy Lu if (ret) { 202a97ef87aSJason Wang goto err_check; 2031e0a84eaSCindy Lu } 2041e0a84eaSCindy Lu return 0; 205a97ef87aSJason Wang err_check: 2061e0a84eaSCindy Lu vhost_net_cleanup(net); 207ab36edcfSJason Wang g_free(net); 208a97ef87aSJason Wang err_init: 2091e0a84eaSCindy Lu return -1; 2101e0a84eaSCindy Lu } 2111e0a84eaSCindy Lu 2121e0a84eaSCindy Lu static void vhost_vdpa_cleanup(NetClientState *nc) 2131e0a84eaSCindy Lu { 2141e0a84eaSCindy Lu VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 2151e0a84eaSCindy Lu 216a0d7215eSAni Sinha /* 217a0d7215eSAni Sinha * If a peer NIC is attached, do not cleanup anything. 218a0d7215eSAni Sinha * Cleanup will happen as a part of qemu_cleanup() -> net_cleanup() 219a0d7215eSAni Sinha * when the guest is shutting down. 220a0d7215eSAni Sinha */ 221a0d7215eSAni Sinha if (nc->peer && nc->peer->info->type == NET_CLIENT_DRIVER_NIC) { 222a0d7215eSAni Sinha return; 223a0d7215eSAni Sinha } 224babf8b87SEugenio Pérez munmap(s->cvq_cmd_out_buffer, vhost_vdpa_net_cvq_cmd_page_len()); 225babf8b87SEugenio Pérez munmap(s->status, vhost_vdpa_net_cvq_cmd_page_len()); 2261e0a84eaSCindy Lu if (s->vhost_net) { 2271e0a84eaSCindy Lu vhost_net_cleanup(s->vhost_net); 2281e0a84eaSCindy Lu g_free(s->vhost_net); 2291e0a84eaSCindy Lu s->vhost_net = NULL; 2301e0a84eaSCindy Lu } 23157b3a7d8SCindy Lu if (s->vhost_vdpa.device_fd >= 0) { 23257b3a7d8SCindy Lu qemu_close(s->vhost_vdpa.device_fd); 23357b3a7d8SCindy Lu s->vhost_vdpa.device_fd = -1; 23457b3a7d8SCindy Lu } 2351e0a84eaSCindy Lu } 2361e0a84eaSCindy Lu 2371e0a84eaSCindy Lu static bool vhost_vdpa_has_vnet_hdr(NetClientState *nc) 2381e0a84eaSCindy Lu { 2391e0a84eaSCindy Lu assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 2401e0a84eaSCindy Lu 2411e0a84eaSCindy Lu return true; 2421e0a84eaSCindy Lu } 2431e0a84eaSCindy Lu 2441e0a84eaSCindy Lu static bool vhost_vdpa_has_ufo(NetClientState *nc) 2451e0a84eaSCindy Lu { 2461e0a84eaSCindy Lu assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 2471e0a84eaSCindy Lu VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 2481e0a84eaSCindy Lu uint64_t features = 0; 2491e0a84eaSCindy Lu features |= (1ULL << VIRTIO_NET_F_HOST_UFO); 2501e0a84eaSCindy Lu features = vhost_net_get_features(s->vhost_net, features); 2511e0a84eaSCindy Lu return !!(features & (1ULL << VIRTIO_NET_F_HOST_UFO)); 2521e0a84eaSCindy Lu 2531e0a84eaSCindy Lu } 2541e0a84eaSCindy Lu 255ee8a1c63SKevin Wolf static bool vhost_vdpa_check_peer_type(NetClientState *nc, ObjectClass *oc, 256ee8a1c63SKevin Wolf Error **errp) 257ee8a1c63SKevin Wolf { 258ee8a1c63SKevin Wolf const char *driver = object_class_get_name(oc); 259ee8a1c63SKevin Wolf 260ee8a1c63SKevin Wolf if (!g_str_has_prefix(driver, "virtio-net-")) { 261ee8a1c63SKevin Wolf error_setg(errp, "vhost-vdpa requires frontend driver virtio-net-*"); 262ee8a1c63SKevin Wolf return false; 263ee8a1c63SKevin Wolf } 264ee8a1c63SKevin Wolf 265ee8a1c63SKevin Wolf return true; 266ee8a1c63SKevin Wolf } 267ee8a1c63SKevin Wolf 268846a1e85SEugenio Pérez /** Dummy receive in case qemu falls back to userland tap networking */ 269846a1e85SEugenio Pérez static ssize_t vhost_vdpa_receive(NetClientState *nc, const uint8_t *buf, 270846a1e85SEugenio Pérez size_t size) 271846a1e85SEugenio Pérez { 272bc5add1dSSi-Wei Liu return size; 273846a1e85SEugenio Pérez } 274846a1e85SEugenio Pérez 27500ef422eSEugenio Pérez /** From any vdpa net client, get the netclient of the first queue pair */ 27600ef422eSEugenio Pérez static VhostVDPAState *vhost_vdpa_net_first_nc_vdpa(VhostVDPAState *s) 27700ef422eSEugenio Pérez { 27800ef422eSEugenio Pérez NICState *nic = qemu_get_nic(s->nc.peer); 27900ef422eSEugenio Pérez NetClientState *nc0 = qemu_get_peer(nic->ncs, 0); 28000ef422eSEugenio Pérez 28100ef422eSEugenio Pérez return DO_UPCAST(VhostVDPAState, nc, nc0); 28200ef422eSEugenio Pérez } 28300ef422eSEugenio Pérez 28469498430SEugenio Pérez static void vhost_vdpa_net_log_global_enable(VhostVDPAState *s, bool enable) 28569498430SEugenio Pérez { 28669498430SEugenio Pérez struct vhost_vdpa *v = &s->vhost_vdpa; 28769498430SEugenio Pérez VirtIONet *n; 28869498430SEugenio Pérez VirtIODevice *vdev; 28969498430SEugenio Pérez int data_queue_pairs, cvq, r; 29069498430SEugenio Pérez 29169498430SEugenio Pérez /* We are only called on the first data vqs and only if x-svq is not set */ 29269498430SEugenio Pérez if (s->vhost_vdpa.shadow_vqs_enabled == enable) { 29369498430SEugenio Pérez return; 29469498430SEugenio Pérez } 29569498430SEugenio Pérez 29669498430SEugenio Pérez vdev = v->dev->vdev; 29769498430SEugenio Pérez n = VIRTIO_NET(vdev); 29869498430SEugenio Pérez if (!n->vhost_started) { 29969498430SEugenio Pérez return; 30069498430SEugenio Pérez } 30169498430SEugenio Pérez 30269498430SEugenio Pérez data_queue_pairs = n->multiqueue ? n->max_queue_pairs : 1; 30369498430SEugenio Pérez cvq = virtio_vdev_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) ? 30469498430SEugenio Pérez n->max_ncs - n->max_queue_pairs : 0; 30569498430SEugenio Pérez /* 30669498430SEugenio Pérez * TODO: vhost_net_stop does suspend, get_base and reset. We can be smarter 30769498430SEugenio Pérez * in the future and resume the device if read-only operations between 30869498430SEugenio Pérez * suspend and reset goes wrong. 30969498430SEugenio Pérez */ 31069498430SEugenio Pérez vhost_net_stop(vdev, n->nic->ncs, data_queue_pairs, cvq); 31169498430SEugenio Pérez 31269498430SEugenio Pérez /* Start will check migration setup_or_active to configure or not SVQ */ 31369498430SEugenio Pérez r = vhost_net_start(vdev, n->nic->ncs, data_queue_pairs, cvq); 31469498430SEugenio Pérez if (unlikely(r < 0)) { 31569498430SEugenio Pérez error_report("unable to start vhost net: %s(%d)", g_strerror(-r), -r); 31669498430SEugenio Pérez } 31769498430SEugenio Pérez } 31869498430SEugenio Pérez 31969498430SEugenio Pérez static void vdpa_net_migration_state_notifier(Notifier *notifier, void *data) 32069498430SEugenio Pérez { 32169498430SEugenio Pérez MigrationState *migration = data; 32269498430SEugenio Pérez VhostVDPAState *s = container_of(notifier, VhostVDPAState, 32369498430SEugenio Pérez migration_state); 32469498430SEugenio Pérez 32569498430SEugenio Pérez if (migration_in_setup(migration)) { 32669498430SEugenio Pérez vhost_vdpa_net_log_global_enable(s, true); 32769498430SEugenio Pérez } else if (migration_has_failed(migration)) { 32869498430SEugenio Pérez vhost_vdpa_net_log_global_enable(s, false); 32969498430SEugenio Pérez } 33069498430SEugenio Pérez } 33169498430SEugenio Pérez 33200ef422eSEugenio Pérez static void vhost_vdpa_net_data_start_first(VhostVDPAState *s) 33300ef422eSEugenio Pérez { 33400ef422eSEugenio Pérez struct vhost_vdpa *v = &s->vhost_vdpa; 33500ef422eSEugenio Pérez 33669498430SEugenio Pérez add_migration_state_change_notifier(&s->migration_state); 33700ef422eSEugenio Pérez if (v->shadow_vqs_enabled) { 33800ef422eSEugenio Pérez v->iova_tree = vhost_iova_tree_new(v->iova_range.first, 33900ef422eSEugenio Pérez v->iova_range.last); 34000ef422eSEugenio Pérez } 34100ef422eSEugenio Pérez } 34200ef422eSEugenio Pérez 34300ef422eSEugenio Pérez static int vhost_vdpa_net_data_start(NetClientState *nc) 34400ef422eSEugenio Pérez { 34500ef422eSEugenio Pérez VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 34600ef422eSEugenio Pérez struct vhost_vdpa *v = &s->vhost_vdpa; 34700ef422eSEugenio Pérez 34800ef422eSEugenio Pérez assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 34900ef422eSEugenio Pérez 35069498430SEugenio Pérez if (s->always_svq || 35169498430SEugenio Pérez migration_is_setup_or_active(migrate_get_current()->state)) { 35269498430SEugenio Pérez v->shadow_vqs_enabled = true; 35369498430SEugenio Pérez v->shadow_data = true; 35469498430SEugenio Pérez } else { 35569498430SEugenio Pérez v->shadow_vqs_enabled = false; 35669498430SEugenio Pérez v->shadow_data = false; 35769498430SEugenio Pérez } 35869498430SEugenio Pérez 35900ef422eSEugenio Pérez if (v->index == 0) { 36000ef422eSEugenio Pérez vhost_vdpa_net_data_start_first(s); 36100ef422eSEugenio Pérez return 0; 36200ef422eSEugenio Pérez } 36300ef422eSEugenio Pérez 36400ef422eSEugenio Pérez if (v->shadow_vqs_enabled) { 36500ef422eSEugenio Pérez VhostVDPAState *s0 = vhost_vdpa_net_first_nc_vdpa(s); 36600ef422eSEugenio Pérez v->iova_tree = s0->vhost_vdpa.iova_tree; 36700ef422eSEugenio Pérez } 36800ef422eSEugenio Pérez 36900ef422eSEugenio Pérez return 0; 37000ef422eSEugenio Pérez } 37100ef422eSEugenio Pérez 37200ef422eSEugenio Pérez static void vhost_vdpa_net_client_stop(NetClientState *nc) 37300ef422eSEugenio Pérez { 37400ef422eSEugenio Pérez VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 37500ef422eSEugenio Pérez struct vhost_dev *dev; 37600ef422eSEugenio Pérez 37700ef422eSEugenio Pérez assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 37800ef422eSEugenio Pérez 37969498430SEugenio Pérez if (s->vhost_vdpa.index == 0) { 38069498430SEugenio Pérez remove_migration_state_change_notifier(&s->migration_state); 38169498430SEugenio Pérez } 38269498430SEugenio Pérez 38300ef422eSEugenio Pérez dev = s->vhost_vdpa.dev; 38400ef422eSEugenio Pérez if (dev->vq_index + dev->nvqs == dev->vq_index_end) { 38500ef422eSEugenio Pérez g_clear_pointer(&s->vhost_vdpa.iova_tree, vhost_iova_tree_delete); 38600ef422eSEugenio Pérez } 38700ef422eSEugenio Pérez } 38800ef422eSEugenio Pérez 3891e0a84eaSCindy Lu static NetClientInfo net_vhost_vdpa_info = { 3901e0a84eaSCindy Lu .type = NET_CLIENT_DRIVER_VHOST_VDPA, 3911e0a84eaSCindy Lu .size = sizeof(VhostVDPAState), 392846a1e85SEugenio Pérez .receive = vhost_vdpa_receive, 39300ef422eSEugenio Pérez .start = vhost_vdpa_net_data_start, 39400ef422eSEugenio Pérez .stop = vhost_vdpa_net_client_stop, 3951e0a84eaSCindy Lu .cleanup = vhost_vdpa_cleanup, 3961e0a84eaSCindy Lu .has_vnet_hdr = vhost_vdpa_has_vnet_hdr, 3971e0a84eaSCindy Lu .has_ufo = vhost_vdpa_has_ufo, 398ee8a1c63SKevin Wolf .check_peer_type = vhost_vdpa_check_peer_type, 3991e0a84eaSCindy Lu }; 4001e0a84eaSCindy Lu 401152128d6SEugenio Pérez static int64_t vhost_vdpa_get_vring_group(int device_fd, unsigned vq_index, 402152128d6SEugenio Pérez Error **errp) 403c1a10086SEugenio Pérez { 404c1a10086SEugenio Pérez struct vhost_vring_state state = { 405c1a10086SEugenio Pérez .index = vq_index, 406c1a10086SEugenio Pérez }; 407c1a10086SEugenio Pérez int r = ioctl(device_fd, VHOST_VDPA_GET_VRING_GROUP, &state); 408c1a10086SEugenio Pérez 409c1a10086SEugenio Pérez if (unlikely(r < 0)) { 4100f2bb0bfSEugenio Pérez r = -errno; 411152128d6SEugenio Pérez error_setg_errno(errp, errno, "Cannot get VQ %u group", vq_index); 412c1a10086SEugenio Pérez return r; 413c1a10086SEugenio Pérez } 414c1a10086SEugenio Pérez 415c1a10086SEugenio Pérez return state.num; 416c1a10086SEugenio Pérez } 417c1a10086SEugenio Pérez 418c1a10086SEugenio Pérez static int vhost_vdpa_set_address_space_id(struct vhost_vdpa *v, 419c1a10086SEugenio Pérez unsigned vq_group, 420c1a10086SEugenio Pérez unsigned asid_num) 421c1a10086SEugenio Pérez { 422c1a10086SEugenio Pérez struct vhost_vring_state asid = { 423c1a10086SEugenio Pérez .index = vq_group, 424c1a10086SEugenio Pérez .num = asid_num, 425c1a10086SEugenio Pérez }; 426c1a10086SEugenio Pérez int r; 427c1a10086SEugenio Pérez 428c1a10086SEugenio Pérez r = ioctl(v->device_fd, VHOST_VDPA_SET_GROUP_ASID, &asid); 429c1a10086SEugenio Pérez if (unlikely(r < 0)) { 430c1a10086SEugenio Pérez error_report("Can't set vq group %u asid %u, errno=%d (%s)", 431c1a10086SEugenio Pérez asid.index, asid.num, errno, g_strerror(errno)); 432c1a10086SEugenio Pérez } 433c1a10086SEugenio Pérez return r; 434c1a10086SEugenio Pérez } 435c1a10086SEugenio Pérez 4362df4dd31SEugenio Pérez static void vhost_vdpa_cvq_unmap_buf(struct vhost_vdpa *v, void *addr) 4372df4dd31SEugenio Pérez { 4382df4dd31SEugenio Pérez VhostIOVATree *tree = v->iova_tree; 4392df4dd31SEugenio Pérez DMAMap needle = { 4402df4dd31SEugenio Pérez /* 4412df4dd31SEugenio Pérez * No need to specify size or to look for more translations since 4422df4dd31SEugenio Pérez * this contiguous chunk was allocated by us. 4432df4dd31SEugenio Pérez */ 4442df4dd31SEugenio Pérez .translated_addr = (hwaddr)(uintptr_t)addr, 4452df4dd31SEugenio Pérez }; 4462df4dd31SEugenio Pérez const DMAMap *map = vhost_iova_tree_find_iova(tree, &needle); 4472df4dd31SEugenio Pérez int r; 4482df4dd31SEugenio Pérez 4492df4dd31SEugenio Pérez if (unlikely(!map)) { 4502df4dd31SEugenio Pérez error_report("Cannot locate expected map"); 4512df4dd31SEugenio Pérez return; 4522df4dd31SEugenio Pérez } 4532df4dd31SEugenio Pérez 454cd831ed5SEugenio Pérez r = vhost_vdpa_dma_unmap(v, v->address_space_id, map->iova, map->size + 1); 4552df4dd31SEugenio Pérez if (unlikely(r != 0)) { 4562df4dd31SEugenio Pérez error_report("Device cannot unmap: %s(%d)", g_strerror(r), r); 4572df4dd31SEugenio Pérez } 4582df4dd31SEugenio Pérez 45969292a8eSEugenio Pérez vhost_iova_tree_remove(tree, *map); 4602df4dd31SEugenio Pérez } 4612df4dd31SEugenio Pérez 4627a7f87e9SEugenio Pérez /** Map CVQ buffer. */ 4637a7f87e9SEugenio Pérez static int vhost_vdpa_cvq_map_buf(struct vhost_vdpa *v, void *buf, size_t size, 4647a7f87e9SEugenio Pérez bool write) 4652df4dd31SEugenio Pérez { 4662df4dd31SEugenio Pérez DMAMap map = {}; 4672df4dd31SEugenio Pérez int r; 4682df4dd31SEugenio Pérez 4692df4dd31SEugenio Pérez map.translated_addr = (hwaddr)(uintptr_t)buf; 4707a7f87e9SEugenio Pérez map.size = size - 1; 4712df4dd31SEugenio Pérez map.perm = write ? IOMMU_RW : IOMMU_RO, 4722df4dd31SEugenio Pérez r = vhost_iova_tree_map_alloc(v->iova_tree, &map); 4732df4dd31SEugenio Pérez if (unlikely(r != IOVA_OK)) { 4742df4dd31SEugenio Pérez error_report("Cannot map injected element"); 4757a7f87e9SEugenio Pérez return r; 4762df4dd31SEugenio Pérez } 4772df4dd31SEugenio Pérez 478cd831ed5SEugenio Pérez r = vhost_vdpa_dma_map(v, v->address_space_id, map.iova, 479cd831ed5SEugenio Pérez vhost_vdpa_net_cvq_cmd_page_len(), buf, !write); 4802df4dd31SEugenio Pérez if (unlikely(r < 0)) { 4812df4dd31SEugenio Pérez goto dma_map_err; 4822df4dd31SEugenio Pérez } 4832df4dd31SEugenio Pérez 4847a7f87e9SEugenio Pérez return 0; 4852df4dd31SEugenio Pérez 4862df4dd31SEugenio Pérez dma_map_err: 48769292a8eSEugenio Pérez vhost_iova_tree_remove(v->iova_tree, map); 4887a7f87e9SEugenio Pérez return r; 4892df4dd31SEugenio Pérez } 4902df4dd31SEugenio Pérez 4917a7f87e9SEugenio Pérez static int vhost_vdpa_net_cvq_start(NetClientState *nc) 4922df4dd31SEugenio Pérez { 49300ef422eSEugenio Pérez VhostVDPAState *s, *s0; 494c1a10086SEugenio Pérez struct vhost_vdpa *v; 495c1a10086SEugenio Pérez int64_t cvq_group; 496152128d6SEugenio Pérez int r; 497152128d6SEugenio Pérez Error *err = NULL; 4982df4dd31SEugenio Pérez 4997a7f87e9SEugenio Pérez assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 5007a7f87e9SEugenio Pérez 5017a7f87e9SEugenio Pérez s = DO_UPCAST(VhostVDPAState, nc, nc); 502c1a10086SEugenio Pérez v = &s->vhost_vdpa; 503c1a10086SEugenio Pérez 50469498430SEugenio Pérez s0 = vhost_vdpa_net_first_nc_vdpa(s); 50569498430SEugenio Pérez v->shadow_data = s0->vhost_vdpa.shadow_vqs_enabled; 506c1a10086SEugenio Pérez v->shadow_vqs_enabled = s->always_svq; 507c1a10086SEugenio Pérez s->vhost_vdpa.address_space_id = VHOST_VDPA_GUEST_PA_ASID; 508c1a10086SEugenio Pérez 50969498430SEugenio Pérez if (s->vhost_vdpa.shadow_data) { 510c1a10086SEugenio Pérez /* SVQ is already configured for all virtqueues */ 511c1a10086SEugenio Pérez goto out; 512c1a10086SEugenio Pérez } 513c1a10086SEugenio Pérez 514c1a10086SEugenio Pérez /* 515c1a10086SEugenio Pérez * If we early return in these cases SVQ will not be enabled. The migration 516c1a10086SEugenio Pérez * will be blocked as long as vhost-vdpa backends will not offer _F_LOG. 517c1a10086SEugenio Pérez */ 518152128d6SEugenio Pérez if (!vhost_vdpa_net_valid_svq_features(v->dev->features, NULL)) { 519c1a10086SEugenio Pérez return 0; 520c1a10086SEugenio Pérez } 521c1a10086SEugenio Pérez 522152128d6SEugenio Pérez if (!s->cvq_isolated) { 523152128d6SEugenio Pérez return 0; 524152128d6SEugenio Pérez } 525152128d6SEugenio Pérez 526152128d6SEugenio Pérez cvq_group = vhost_vdpa_get_vring_group(v->device_fd, 527152128d6SEugenio Pérez v->dev->vq_index_end - 1, 528152128d6SEugenio Pérez &err); 529c1a10086SEugenio Pérez if (unlikely(cvq_group < 0)) { 530152128d6SEugenio Pérez error_report_err(err); 531c1a10086SEugenio Pérez return cvq_group; 532c1a10086SEugenio Pérez } 533c1a10086SEugenio Pérez 534c1a10086SEugenio Pérez r = vhost_vdpa_set_address_space_id(v, cvq_group, VHOST_VDPA_NET_CVQ_ASID); 535c1a10086SEugenio Pérez if (unlikely(r < 0)) { 536c1a10086SEugenio Pérez return r; 537c1a10086SEugenio Pérez } 538c1a10086SEugenio Pérez 539c1a10086SEugenio Pérez v->shadow_vqs_enabled = true; 540c1a10086SEugenio Pérez s->vhost_vdpa.address_space_id = VHOST_VDPA_NET_CVQ_ASID; 541c1a10086SEugenio Pérez 542c1a10086SEugenio Pérez out: 5437a7f87e9SEugenio Pérez if (!s->vhost_vdpa.shadow_vqs_enabled) { 5447a7f87e9SEugenio Pérez return 0; 5452df4dd31SEugenio Pérez } 5462df4dd31SEugenio Pérez 54700ef422eSEugenio Pérez if (s0->vhost_vdpa.iova_tree) { 54800ef422eSEugenio Pérez /* 54900ef422eSEugenio Pérez * SVQ is already configured for all virtqueues. Reuse IOVA tree for 55000ef422eSEugenio Pérez * simplicity, whether CVQ shares ASID with guest or not, because: 55100ef422eSEugenio Pérez * - Memory listener need access to guest's memory addresses allocated 55200ef422eSEugenio Pérez * in the IOVA tree. 55300ef422eSEugenio Pérez * - There should be plenty of IOVA address space for both ASID not to 55400ef422eSEugenio Pérez * worry about collisions between them. Guest's translations are 55500ef422eSEugenio Pérez * still validated with virtio virtqueue_pop so there is no risk for 55600ef422eSEugenio Pérez * the guest to access memory that it shouldn't. 55700ef422eSEugenio Pérez * 55800ef422eSEugenio Pérez * To allocate a iova tree per ASID is doable but it complicates the 55900ef422eSEugenio Pérez * code and it is not worth it for the moment. 56000ef422eSEugenio Pérez */ 56100ef422eSEugenio Pérez v->iova_tree = s0->vhost_vdpa.iova_tree; 56200ef422eSEugenio Pérez } else { 56300ef422eSEugenio Pérez v->iova_tree = vhost_iova_tree_new(v->iova_range.first, 56400ef422eSEugenio Pérez v->iova_range.last); 56500ef422eSEugenio Pérez } 56600ef422eSEugenio Pérez 5677a7f87e9SEugenio Pérez r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer, 5687a7f87e9SEugenio Pérez vhost_vdpa_net_cvq_cmd_page_len(), false); 5697a7f87e9SEugenio Pérez if (unlikely(r < 0)) { 5707a7f87e9SEugenio Pérez return r; 5717a7f87e9SEugenio Pérez } 5727a7f87e9SEugenio Pérez 57317fb889fSEugenio Pérez r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->status, 5747a7f87e9SEugenio Pérez vhost_vdpa_net_cvq_cmd_page_len(), true); 5757a7f87e9SEugenio Pérez if (unlikely(r < 0)) { 5762df4dd31SEugenio Pérez vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer); 5772df4dd31SEugenio Pérez } 5782df4dd31SEugenio Pérez 5797a7f87e9SEugenio Pérez return r; 5807a7f87e9SEugenio Pérez } 5817a7f87e9SEugenio Pérez 5827a7f87e9SEugenio Pérez static void vhost_vdpa_net_cvq_stop(NetClientState *nc) 5837a7f87e9SEugenio Pérez { 5847a7f87e9SEugenio Pérez VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 5857a7f87e9SEugenio Pérez 5867a7f87e9SEugenio Pérez assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 5877a7f87e9SEugenio Pérez 5887a7f87e9SEugenio Pérez if (s->vhost_vdpa.shadow_vqs_enabled) { 5897a7f87e9SEugenio Pérez vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer); 59017fb889fSEugenio Pérez vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->status); 591c1a10086SEugenio Pérez } 59200ef422eSEugenio Pérez 59300ef422eSEugenio Pérez vhost_vdpa_net_client_stop(nc); 5942df4dd31SEugenio Pérez } 5952df4dd31SEugenio Pérez 596be4278b6SEugenio Pérez static ssize_t vhost_vdpa_net_cvq_add(VhostVDPAState *s, size_t out_len, 597be4278b6SEugenio Pérez size_t in_len) 598be4278b6SEugenio Pérez { 599be4278b6SEugenio Pérez /* Buffers for the device */ 600be4278b6SEugenio Pérez const struct iovec out = { 601be4278b6SEugenio Pérez .iov_base = s->cvq_cmd_out_buffer, 602be4278b6SEugenio Pérez .iov_len = out_len, 603be4278b6SEugenio Pérez }; 604be4278b6SEugenio Pérez const struct iovec in = { 60517fb889fSEugenio Pérez .iov_base = s->status, 606be4278b6SEugenio Pérez .iov_len = sizeof(virtio_net_ctrl_ack), 607be4278b6SEugenio Pérez }; 608be4278b6SEugenio Pérez VhostShadowVirtqueue *svq = g_ptr_array_index(s->vhost_vdpa.shadow_vqs, 0); 609be4278b6SEugenio Pérez int r; 610be4278b6SEugenio Pérez 611be4278b6SEugenio Pérez r = vhost_svq_add(svq, &out, 1, &in, 1, NULL); 612be4278b6SEugenio Pérez if (unlikely(r != 0)) { 613be4278b6SEugenio Pérez if (unlikely(r == -ENOSPC)) { 614be4278b6SEugenio Pérez qemu_log_mask(LOG_GUEST_ERROR, "%s: No space on device queue\n", 615be4278b6SEugenio Pérez __func__); 616be4278b6SEugenio Pérez } 617be4278b6SEugenio Pérez return r; 618be4278b6SEugenio Pérez } 619be4278b6SEugenio Pérez 620be4278b6SEugenio Pérez /* 621be4278b6SEugenio Pérez * We can poll here since we've had BQL from the time we sent the 622be4278b6SEugenio Pérez * descriptor. Also, we need to take the answer before SVQ pulls by itself, 623be4278b6SEugenio Pérez * when BQL is released 624be4278b6SEugenio Pérez */ 625be4278b6SEugenio Pérez return vhost_svq_poll(svq); 626be4278b6SEugenio Pérez } 627be4278b6SEugenio Pérez 628f73c0c43SEugenio Pérez static ssize_t vhost_vdpa_net_load_cmd(VhostVDPAState *s, uint8_t class, 6292848c6aaSHawkins Jiawei uint8_t cmd, const struct iovec *data_sg, 6302848c6aaSHawkins Jiawei size_t data_num) 631f73c0c43SEugenio Pérez { 632f73c0c43SEugenio Pérez const struct virtio_net_ctrl_hdr ctrl = { 633f73c0c43SEugenio Pérez .class = class, 634f73c0c43SEugenio Pérez .cmd = cmd, 635f73c0c43SEugenio Pérez }; 6362848c6aaSHawkins Jiawei size_t data_size = iov_size(data_sg, data_num); 637f73c0c43SEugenio Pérez 638f73c0c43SEugenio Pérez assert(data_size < vhost_vdpa_net_cvq_cmd_page_len() - sizeof(ctrl)); 639f73c0c43SEugenio Pérez 6402848c6aaSHawkins Jiawei /* pack the CVQ command header */ 641f73c0c43SEugenio Pérez memcpy(s->cvq_cmd_out_buffer, &ctrl, sizeof(ctrl)); 642f73c0c43SEugenio Pérez 6432848c6aaSHawkins Jiawei /* pack the CVQ command command-specific-data */ 6442848c6aaSHawkins Jiawei iov_to_buf(data_sg, data_num, 0, 6452848c6aaSHawkins Jiawei s->cvq_cmd_out_buffer + sizeof(ctrl), data_size); 6462848c6aaSHawkins Jiawei 6472848c6aaSHawkins Jiawei return vhost_vdpa_net_cvq_add(s, data_size + sizeof(ctrl), 648f73c0c43SEugenio Pérez sizeof(virtio_net_ctrl_ack)); 649f73c0c43SEugenio Pérez } 650f73c0c43SEugenio Pérez 651f73c0c43SEugenio Pérez static int vhost_vdpa_net_load_mac(VhostVDPAState *s, const VirtIONet *n) 652f73c0c43SEugenio Pérez { 65302d3bf09SHawkins Jiawei if (virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_MAC_ADDR)) { 6542848c6aaSHawkins Jiawei const struct iovec data = { 6552848c6aaSHawkins Jiawei .iov_base = (void *)n->mac, 6562848c6aaSHawkins Jiawei .iov_len = sizeof(n->mac), 6572848c6aaSHawkins Jiawei }; 658f73c0c43SEugenio Pérez ssize_t dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_MAC, 659f73c0c43SEugenio Pérez VIRTIO_NET_CTRL_MAC_ADDR_SET, 6602848c6aaSHawkins Jiawei &data, 1); 661f73c0c43SEugenio Pérez if (unlikely(dev_written < 0)) { 662f73c0c43SEugenio Pérez return dev_written; 663f73c0c43SEugenio Pérez } 664b479bc3cSHawkins Jiawei if (*s->status != VIRTIO_NET_OK) { 665b479bc3cSHawkins Jiawei return -EIO; 666b479bc3cSHawkins Jiawei } 667f73c0c43SEugenio Pérez } 668f73c0c43SEugenio Pérez 6690ddcecb8SHawkins Jiawei /* 6700ddcecb8SHawkins Jiawei * According to VirtIO standard, "The device MUST have an 6710ddcecb8SHawkins Jiawei * empty MAC filtering table on reset.". 6720ddcecb8SHawkins Jiawei * 6730ddcecb8SHawkins Jiawei * Therefore, there is no need to send this CVQ command if the 6740ddcecb8SHawkins Jiawei * driver also sets an empty MAC filter table, which aligns with 6750ddcecb8SHawkins Jiawei * the device's defaults. 6760ddcecb8SHawkins Jiawei * 6770ddcecb8SHawkins Jiawei * Note that the device's defaults can mismatch the driver's 6780ddcecb8SHawkins Jiawei * configuration only at live migration. 6790ddcecb8SHawkins Jiawei */ 6800ddcecb8SHawkins Jiawei if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_RX) || 6810ddcecb8SHawkins Jiawei n->mac_table.in_use == 0) { 6820ddcecb8SHawkins Jiawei return 0; 6830ddcecb8SHawkins Jiawei } 6840ddcecb8SHawkins Jiawei 6850ddcecb8SHawkins Jiawei uint32_t uni_entries = n->mac_table.first_multi, 6860ddcecb8SHawkins Jiawei uni_macs_size = uni_entries * ETH_ALEN, 6870ddcecb8SHawkins Jiawei mul_entries = n->mac_table.in_use - uni_entries, 6880ddcecb8SHawkins Jiawei mul_macs_size = mul_entries * ETH_ALEN; 6890ddcecb8SHawkins Jiawei struct virtio_net_ctrl_mac uni = { 6900ddcecb8SHawkins Jiawei .entries = cpu_to_le32(uni_entries), 6910ddcecb8SHawkins Jiawei }; 6920ddcecb8SHawkins Jiawei struct virtio_net_ctrl_mac mul = { 6930ddcecb8SHawkins Jiawei .entries = cpu_to_le32(mul_entries), 6940ddcecb8SHawkins Jiawei }; 6950ddcecb8SHawkins Jiawei const struct iovec data[] = { 6960ddcecb8SHawkins Jiawei { 6970ddcecb8SHawkins Jiawei .iov_base = &uni, 6980ddcecb8SHawkins Jiawei .iov_len = sizeof(uni), 6990ddcecb8SHawkins Jiawei }, { 7000ddcecb8SHawkins Jiawei .iov_base = n->mac_table.macs, 7010ddcecb8SHawkins Jiawei .iov_len = uni_macs_size, 7020ddcecb8SHawkins Jiawei }, { 7030ddcecb8SHawkins Jiawei .iov_base = &mul, 7040ddcecb8SHawkins Jiawei .iov_len = sizeof(mul), 7050ddcecb8SHawkins Jiawei }, { 7060ddcecb8SHawkins Jiawei .iov_base = &n->mac_table.macs[uni_macs_size], 7070ddcecb8SHawkins Jiawei .iov_len = mul_macs_size, 7080ddcecb8SHawkins Jiawei }, 7090ddcecb8SHawkins Jiawei }; 7100ddcecb8SHawkins Jiawei ssize_t dev_written = vhost_vdpa_net_load_cmd(s, 7110ddcecb8SHawkins Jiawei VIRTIO_NET_CTRL_MAC, 7120ddcecb8SHawkins Jiawei VIRTIO_NET_CTRL_MAC_TABLE_SET, 7130ddcecb8SHawkins Jiawei data, ARRAY_SIZE(data)); 7140ddcecb8SHawkins Jiawei if (unlikely(dev_written < 0)) { 7150ddcecb8SHawkins Jiawei return dev_written; 7160ddcecb8SHawkins Jiawei } 7170ddcecb8SHawkins Jiawei if (*s->status != VIRTIO_NET_OK) { 7180ddcecb8SHawkins Jiawei return -EIO; 7190ddcecb8SHawkins Jiawei } 7200ddcecb8SHawkins Jiawei 721f73c0c43SEugenio Pérez return 0; 722f73c0c43SEugenio Pérez } 723f73c0c43SEugenio Pérez 724f64c7cdaSEugenio Pérez static int vhost_vdpa_net_load_mq(VhostVDPAState *s, 725f64c7cdaSEugenio Pérez const VirtIONet *n) 726f64c7cdaSEugenio Pérez { 727f64c7cdaSEugenio Pérez struct virtio_net_ctrl_mq mq; 728f64c7cdaSEugenio Pérez ssize_t dev_written; 729f64c7cdaSEugenio Pérez 73002d3bf09SHawkins Jiawei if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_MQ)) { 731f64c7cdaSEugenio Pérez return 0; 732f64c7cdaSEugenio Pérez } 733f64c7cdaSEugenio Pérez 734f64c7cdaSEugenio Pérez mq.virtqueue_pairs = cpu_to_le16(n->curr_queue_pairs); 7352848c6aaSHawkins Jiawei const struct iovec data = { 7362848c6aaSHawkins Jiawei .iov_base = &mq, 7372848c6aaSHawkins Jiawei .iov_len = sizeof(mq), 7382848c6aaSHawkins Jiawei }; 739f64c7cdaSEugenio Pérez dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_MQ, 7402848c6aaSHawkins Jiawei VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, 7412848c6aaSHawkins Jiawei &data, 1); 742f64c7cdaSEugenio Pérez if (unlikely(dev_written < 0)) { 743f64c7cdaSEugenio Pérez return dev_written; 744f64c7cdaSEugenio Pérez } 745f45fd95eSHawkins Jiawei if (*s->status != VIRTIO_NET_OK) { 746f45fd95eSHawkins Jiawei return -EIO; 747f45fd95eSHawkins Jiawei } 748f64c7cdaSEugenio Pérez 749f45fd95eSHawkins Jiawei return 0; 750f64c7cdaSEugenio Pérez } 751f64c7cdaSEugenio Pérez 7520b58d368SHawkins Jiawei static int vhost_vdpa_net_load_offloads(VhostVDPAState *s, 7530b58d368SHawkins Jiawei const VirtIONet *n) 7540b58d368SHawkins Jiawei { 7550b58d368SHawkins Jiawei uint64_t offloads; 7560b58d368SHawkins Jiawei ssize_t dev_written; 7570b58d368SHawkins Jiawei 7580b58d368SHawkins Jiawei if (!virtio_vdev_has_feature(&n->parent_obj, 7590b58d368SHawkins Jiawei VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)) { 7600b58d368SHawkins Jiawei return 0; 7610b58d368SHawkins Jiawei } 7620b58d368SHawkins Jiawei 7630b58d368SHawkins Jiawei if (n->curr_guest_offloads == virtio_net_supported_guest_offloads(n)) { 7640b58d368SHawkins Jiawei /* 7650b58d368SHawkins Jiawei * According to VirtIO standard, "Upon feature negotiation 7660b58d368SHawkins Jiawei * corresponding offload gets enabled to preserve 7670b58d368SHawkins Jiawei * backward compatibility.". 7680b58d368SHawkins Jiawei * 7690b58d368SHawkins Jiawei * Therefore, there is no need to send this CVQ command if the 7700b58d368SHawkins Jiawei * driver also enables all supported offloads, which aligns with 7710b58d368SHawkins Jiawei * the device's defaults. 7720b58d368SHawkins Jiawei * 7730b58d368SHawkins Jiawei * Note that the device's defaults can mismatch the driver's 7740b58d368SHawkins Jiawei * configuration only at live migration. 7750b58d368SHawkins Jiawei */ 7760b58d368SHawkins Jiawei return 0; 7770b58d368SHawkins Jiawei } 7780b58d368SHawkins Jiawei 7790b58d368SHawkins Jiawei offloads = cpu_to_le64(n->curr_guest_offloads); 7802848c6aaSHawkins Jiawei const struct iovec data = { 7812848c6aaSHawkins Jiawei .iov_base = &offloads, 7822848c6aaSHawkins Jiawei .iov_len = sizeof(offloads), 7832848c6aaSHawkins Jiawei }; 7840b58d368SHawkins Jiawei dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_GUEST_OFFLOADS, 7850b58d368SHawkins Jiawei VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, 7862848c6aaSHawkins Jiawei &data, 1); 7870b58d368SHawkins Jiawei if (unlikely(dev_written < 0)) { 7880b58d368SHawkins Jiawei return dev_written; 7890b58d368SHawkins Jiawei } 7906f348071SHawkins Jiawei if (*s->status != VIRTIO_NET_OK) { 7916f348071SHawkins Jiawei return -EIO; 7926f348071SHawkins Jiawei } 7930b58d368SHawkins Jiawei 7946f348071SHawkins Jiawei return 0; 7950b58d368SHawkins Jiawei } 7960b58d368SHawkins Jiawei 797b12f907eSHawkins Jiawei static int vhost_vdpa_net_load_rx_mode(VhostVDPAState *s, 798b12f907eSHawkins Jiawei uint8_t cmd, 799b12f907eSHawkins Jiawei uint8_t on) 800b12f907eSHawkins Jiawei { 801b12f907eSHawkins Jiawei const struct iovec data = { 802b12f907eSHawkins Jiawei .iov_base = &on, 803b12f907eSHawkins Jiawei .iov_len = sizeof(on), 804b12f907eSHawkins Jiawei }; 805b12f907eSHawkins Jiawei return vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_RX, 806b12f907eSHawkins Jiawei cmd, &data, 1); 807b12f907eSHawkins Jiawei } 808b12f907eSHawkins Jiawei 809b12f907eSHawkins Jiawei static int vhost_vdpa_net_load_rx(VhostVDPAState *s, 810b12f907eSHawkins Jiawei const VirtIONet *n) 811b12f907eSHawkins Jiawei { 812b12f907eSHawkins Jiawei ssize_t dev_written; 813b12f907eSHawkins Jiawei 814b12f907eSHawkins Jiawei if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_RX)) { 815b12f907eSHawkins Jiawei return 0; 816b12f907eSHawkins Jiawei } 817b12f907eSHawkins Jiawei 818b12f907eSHawkins Jiawei /* 819b12f907eSHawkins Jiawei * According to virtio_net_reset(), device turns promiscuous mode 820b12f907eSHawkins Jiawei * on by default. 821b12f907eSHawkins Jiawei * 822b12f907eSHawkins Jiawei * Addtionally, according to VirtIO standard, "Since there are 823b12f907eSHawkins Jiawei * no guarantees, it can use a hash filter or silently switch to 824b12f907eSHawkins Jiawei * allmulti or promiscuous mode if it is given too many addresses.". 825b12f907eSHawkins Jiawei * QEMU marks `n->mac_table.uni_overflow` if guest sets too many 826b12f907eSHawkins Jiawei * non-multicast MAC addresses, indicating that promiscuous mode 827b12f907eSHawkins Jiawei * should be enabled. 828b12f907eSHawkins Jiawei * 829b12f907eSHawkins Jiawei * Therefore, QEMU should only send this CVQ command if the 830b12f907eSHawkins Jiawei * `n->mac_table.uni_overflow` is not marked and `n->promisc` is off, 831b12f907eSHawkins Jiawei * which sets promiscuous mode on, different from the device's defaults. 832b12f907eSHawkins Jiawei * 833b12f907eSHawkins Jiawei * Note that the device's defaults can mismatch the driver's 834b12f907eSHawkins Jiawei * configuration only at live migration. 835b12f907eSHawkins Jiawei */ 836b12f907eSHawkins Jiawei if (!n->mac_table.uni_overflow && !n->promisc) { 837b12f907eSHawkins Jiawei dev_written = vhost_vdpa_net_load_rx_mode(s, 838b12f907eSHawkins Jiawei VIRTIO_NET_CTRL_RX_PROMISC, 0); 839b12f907eSHawkins Jiawei if (unlikely(dev_written < 0)) { 840b12f907eSHawkins Jiawei return dev_written; 841b12f907eSHawkins Jiawei } 842b12f907eSHawkins Jiawei if (*s->status != VIRTIO_NET_OK) { 843b12f907eSHawkins Jiawei return -EIO; 844b12f907eSHawkins Jiawei } 845b12f907eSHawkins Jiawei } 846b12f907eSHawkins Jiawei 847b12f907eSHawkins Jiawei /* 848b12f907eSHawkins Jiawei * According to virtio_net_reset(), device turns all-multicast mode 849b12f907eSHawkins Jiawei * off by default. 850b12f907eSHawkins Jiawei * 851b12f907eSHawkins Jiawei * According to VirtIO standard, "Since there are no guarantees, 852b12f907eSHawkins Jiawei * it can use a hash filter or silently switch to allmulti or 853b12f907eSHawkins Jiawei * promiscuous mode if it is given too many addresses.". QEMU marks 854b12f907eSHawkins Jiawei * `n->mac_table.multi_overflow` if guest sets too many 855b12f907eSHawkins Jiawei * non-multicast MAC addresses. 856b12f907eSHawkins Jiawei * 857b12f907eSHawkins Jiawei * Therefore, QEMU should only send this CVQ command if the 858b12f907eSHawkins Jiawei * `n->mac_table.multi_overflow` is marked or `n->allmulti` is on, 859b12f907eSHawkins Jiawei * which sets all-multicast mode on, different from the device's defaults. 860b12f907eSHawkins Jiawei * 861b12f907eSHawkins Jiawei * Note that the device's defaults can mismatch the driver's 862b12f907eSHawkins Jiawei * configuration only at live migration. 863b12f907eSHawkins Jiawei */ 864b12f907eSHawkins Jiawei if (n->mac_table.multi_overflow || n->allmulti) { 865b12f907eSHawkins Jiawei dev_written = vhost_vdpa_net_load_rx_mode(s, 866b12f907eSHawkins Jiawei VIRTIO_NET_CTRL_RX_ALLMULTI, 1); 867b12f907eSHawkins Jiawei if (unlikely(dev_written < 0)) { 868b12f907eSHawkins Jiawei return dev_written; 869b12f907eSHawkins Jiawei } 870b12f907eSHawkins Jiawei if (*s->status != VIRTIO_NET_OK) { 871b12f907eSHawkins Jiawei return -EIO; 872b12f907eSHawkins Jiawei } 873b12f907eSHawkins Jiawei } 874b12f907eSHawkins Jiawei 875b12f907eSHawkins Jiawei return 0; 876b12f907eSHawkins Jiawei } 877b12f907eSHawkins Jiawei 878dd036d8dSEugenio Pérez static int vhost_vdpa_net_load(NetClientState *nc) 879dd036d8dSEugenio Pérez { 880dd036d8dSEugenio Pérez VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 881f73c0c43SEugenio Pérez struct vhost_vdpa *v = &s->vhost_vdpa; 882dd036d8dSEugenio Pérez const VirtIONet *n; 883f73c0c43SEugenio Pérez int r; 884dd036d8dSEugenio Pérez 885dd036d8dSEugenio Pérez assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 886dd036d8dSEugenio Pérez 887dd036d8dSEugenio Pérez if (!v->shadow_vqs_enabled) { 888dd036d8dSEugenio Pérez return 0; 889dd036d8dSEugenio Pérez } 890dd036d8dSEugenio Pérez 891dd036d8dSEugenio Pérez n = VIRTIO_NET(v->dev->vdev); 892f73c0c43SEugenio Pérez r = vhost_vdpa_net_load_mac(s, n); 893f73c0c43SEugenio Pérez if (unlikely(r < 0)) { 894f73c0c43SEugenio Pérez return r; 895dd036d8dSEugenio Pérez } 896f64c7cdaSEugenio Pérez r = vhost_vdpa_net_load_mq(s, n); 897f64c7cdaSEugenio Pérez if (unlikely(r)) { 898f64c7cdaSEugenio Pérez return r; 899f64c7cdaSEugenio Pérez } 9000b58d368SHawkins Jiawei r = vhost_vdpa_net_load_offloads(s, n); 9010b58d368SHawkins Jiawei if (unlikely(r)) { 9020b58d368SHawkins Jiawei return r; 9030b58d368SHawkins Jiawei } 904b12f907eSHawkins Jiawei r = vhost_vdpa_net_load_rx(s, n); 905b12f907eSHawkins Jiawei if (unlikely(r)) { 906b12f907eSHawkins Jiawei return r; 907b12f907eSHawkins Jiawei } 908dd036d8dSEugenio Pérez 909dd036d8dSEugenio Pérez return 0; 910dd036d8dSEugenio Pérez } 911dd036d8dSEugenio Pérez 912f8972b56SEugenio Pérez static NetClientInfo net_vhost_vdpa_cvq_info = { 913f8972b56SEugenio Pérez .type = NET_CLIENT_DRIVER_VHOST_VDPA, 914f8972b56SEugenio Pérez .size = sizeof(VhostVDPAState), 915f8972b56SEugenio Pérez .receive = vhost_vdpa_receive, 9167a7f87e9SEugenio Pérez .start = vhost_vdpa_net_cvq_start, 917dd036d8dSEugenio Pérez .load = vhost_vdpa_net_load, 9187a7f87e9SEugenio Pérez .stop = vhost_vdpa_net_cvq_stop, 919f8972b56SEugenio Pérez .cleanup = vhost_vdpa_cleanup, 920f8972b56SEugenio Pérez .has_vnet_hdr = vhost_vdpa_has_vnet_hdr, 921f8972b56SEugenio Pérez .has_ufo = vhost_vdpa_has_ufo, 922f8972b56SEugenio Pérez .check_peer_type = vhost_vdpa_check_peer_type, 923f8972b56SEugenio Pérez }; 924f8972b56SEugenio Pérez 9252df4dd31SEugenio Pérez /** 9262df4dd31SEugenio Pérez * Validate and copy control virtqueue commands. 9272df4dd31SEugenio Pérez * 9282df4dd31SEugenio Pérez * Following QEMU guidelines, we offer a copy of the buffers to the device to 9292df4dd31SEugenio Pérez * prevent TOCTOU bugs. 930bd907ae4SEugenio Pérez */ 931bd907ae4SEugenio Pérez static int vhost_vdpa_net_handle_ctrl_avail(VhostShadowVirtqueue *svq, 932bd907ae4SEugenio Pérez VirtQueueElement *elem, 933bd907ae4SEugenio Pérez void *opaque) 934bd907ae4SEugenio Pérez { 9352df4dd31SEugenio Pérez VhostVDPAState *s = opaque; 936be4278b6SEugenio Pérez size_t in_len; 937*45c41018SHawkins Jiawei const struct virtio_net_ctrl_hdr *ctrl; 938bd907ae4SEugenio Pérez virtio_net_ctrl_ack status = VIRTIO_NET_ERR; 9397a7f87e9SEugenio Pérez /* Out buffer sent to both the vdpa device and the device model */ 9407a7f87e9SEugenio Pérez struct iovec out = { 9417a7f87e9SEugenio Pérez .iov_base = s->cvq_cmd_out_buffer, 9427a7f87e9SEugenio Pérez }; 9432df4dd31SEugenio Pérez /* in buffer used for device model */ 9442df4dd31SEugenio Pérez const struct iovec in = { 9452df4dd31SEugenio Pérez .iov_base = &status, 9462df4dd31SEugenio Pérez .iov_len = sizeof(status), 9472df4dd31SEugenio Pérez }; 948be4278b6SEugenio Pérez ssize_t dev_written = -EINVAL; 949bd907ae4SEugenio Pérez 9507a7f87e9SEugenio Pérez out.iov_len = iov_to_buf(elem->out_sg, elem->out_num, 0, 9517a7f87e9SEugenio Pérez s->cvq_cmd_out_buffer, 9527a7f87e9SEugenio Pérez vhost_vdpa_net_cvq_cmd_len()); 953*45c41018SHawkins Jiawei 954*45c41018SHawkins Jiawei ctrl = s->cvq_cmd_out_buffer; 955*45c41018SHawkins Jiawei if (ctrl->class == VIRTIO_NET_CTRL_ANNOUNCE) { 9563f9a3eebSEugenio Pérez /* 9573f9a3eebSEugenio Pérez * Guest announce capability is emulated by qemu, so don't forward to 9583f9a3eebSEugenio Pérez * the device. 9593f9a3eebSEugenio Pérez */ 9603f9a3eebSEugenio Pérez dev_written = sizeof(status); 9613f9a3eebSEugenio Pérez *s->status = VIRTIO_NET_OK; 9623f9a3eebSEugenio Pérez } else { 963be4278b6SEugenio Pérez dev_written = vhost_vdpa_net_cvq_add(s, out.iov_len, sizeof(status)); 964be4278b6SEugenio Pérez if (unlikely(dev_written < 0)) { 965bd907ae4SEugenio Pérez goto out; 966bd907ae4SEugenio Pérez } 9673f9a3eebSEugenio Pérez } 968bd907ae4SEugenio Pérez 969bd907ae4SEugenio Pérez if (unlikely(dev_written < sizeof(status))) { 970bd907ae4SEugenio Pérez error_report("Insufficient written data (%zu)", dev_written); 9712df4dd31SEugenio Pérez goto out; 9722df4dd31SEugenio Pérez } 9732df4dd31SEugenio Pérez 97417fb889fSEugenio Pérez if (*s->status != VIRTIO_NET_OK) { 975d45243bcSEugenio Pérez goto out; 9762df4dd31SEugenio Pérez } 9772df4dd31SEugenio Pérez 9782df4dd31SEugenio Pérez status = VIRTIO_NET_ERR; 9797a7f87e9SEugenio Pérez virtio_net_handle_ctrl_iov(svq->vdev, &in, 1, &out, 1); 9802df4dd31SEugenio Pérez if (status != VIRTIO_NET_OK) { 9812df4dd31SEugenio Pérez error_report("Bad CVQ processing in model"); 982bd907ae4SEugenio Pérez } 983bd907ae4SEugenio Pérez 984bd907ae4SEugenio Pérez out: 985bd907ae4SEugenio Pérez in_len = iov_from_buf(elem->in_sg, elem->in_num, 0, &status, 986bd907ae4SEugenio Pérez sizeof(status)); 987bd907ae4SEugenio Pérez if (unlikely(in_len < sizeof(status))) { 988bd907ae4SEugenio Pérez error_report("Bad device CVQ written length"); 989bd907ae4SEugenio Pérez } 990bd907ae4SEugenio Pérez vhost_svq_push_elem(svq, elem, MIN(in_len, sizeof(status))); 991031b1abaSHawkins Jiawei /* 992031b1abaSHawkins Jiawei * `elem` belongs to vhost_vdpa_net_handle_ctrl_avail() only when 993031b1abaSHawkins Jiawei * the function successfully forwards the CVQ command, indicated 994031b1abaSHawkins Jiawei * by a non-negative value of `dev_written`. Otherwise, it still 995031b1abaSHawkins Jiawei * belongs to SVQ. 996031b1abaSHawkins Jiawei * This function should only free the `elem` when it owns. 997031b1abaSHawkins Jiawei */ 998031b1abaSHawkins Jiawei if (dev_written >= 0) { 999bd907ae4SEugenio Pérez g_free(elem); 1000031b1abaSHawkins Jiawei } 1001be4278b6SEugenio Pérez return dev_written < 0 ? dev_written : 0; 1002bd907ae4SEugenio Pérez } 1003bd907ae4SEugenio Pérez 1004bd907ae4SEugenio Pérez static const VhostShadowVirtqueueOps vhost_vdpa_net_svq_ops = { 1005bd907ae4SEugenio Pérez .avail_handler = vhost_vdpa_net_handle_ctrl_avail, 1006bd907ae4SEugenio Pérez }; 1007bd907ae4SEugenio Pérez 1008152128d6SEugenio Pérez /** 1009152128d6SEugenio Pérez * Probe if CVQ is isolated 1010152128d6SEugenio Pérez * 1011152128d6SEugenio Pérez * @device_fd The vdpa device fd 1012152128d6SEugenio Pérez * @features Features offered by the device. 1013152128d6SEugenio Pérez * @cvq_index The control vq pair index 1014152128d6SEugenio Pérez * 1015152128d6SEugenio Pérez * Returns <0 in case of failure, 0 if false and 1 if true. 1016152128d6SEugenio Pérez */ 1017152128d6SEugenio Pérez static int vhost_vdpa_probe_cvq_isolation(int device_fd, uint64_t features, 1018152128d6SEugenio Pérez int cvq_index, Error **errp) 1019152128d6SEugenio Pérez { 1020152128d6SEugenio Pérez uint64_t backend_features; 1021152128d6SEugenio Pérez int64_t cvq_group; 1022152128d6SEugenio Pérez uint8_t status = VIRTIO_CONFIG_S_ACKNOWLEDGE | 1023152128d6SEugenio Pérez VIRTIO_CONFIG_S_DRIVER | 1024152128d6SEugenio Pérez VIRTIO_CONFIG_S_FEATURES_OK; 1025152128d6SEugenio Pérez int r; 1026152128d6SEugenio Pérez 1027152128d6SEugenio Pérez ERRP_GUARD(); 1028152128d6SEugenio Pérez 1029152128d6SEugenio Pérez r = ioctl(device_fd, VHOST_GET_BACKEND_FEATURES, &backend_features); 1030152128d6SEugenio Pérez if (unlikely(r < 0)) { 1031152128d6SEugenio Pérez error_setg_errno(errp, errno, "Cannot get vdpa backend_features"); 1032152128d6SEugenio Pérez return r; 1033152128d6SEugenio Pérez } 1034152128d6SEugenio Pérez 1035152128d6SEugenio Pérez if (!(backend_features & BIT_ULL(VHOST_BACKEND_F_IOTLB_ASID))) { 1036152128d6SEugenio Pérez return 0; 1037152128d6SEugenio Pérez } 1038152128d6SEugenio Pérez 1039152128d6SEugenio Pérez r = ioctl(device_fd, VHOST_SET_FEATURES, &features); 1040152128d6SEugenio Pérez if (unlikely(r)) { 1041152128d6SEugenio Pérez error_setg_errno(errp, errno, "Cannot set features"); 1042152128d6SEugenio Pérez } 1043152128d6SEugenio Pérez 1044152128d6SEugenio Pérez r = ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status); 1045152128d6SEugenio Pérez if (unlikely(r)) { 1046152128d6SEugenio Pérez error_setg_errno(errp, -r, "Cannot set device features"); 1047152128d6SEugenio Pérez goto out; 1048152128d6SEugenio Pérez } 1049152128d6SEugenio Pérez 1050152128d6SEugenio Pérez cvq_group = vhost_vdpa_get_vring_group(device_fd, cvq_index, errp); 1051152128d6SEugenio Pérez if (unlikely(cvq_group < 0)) { 1052152128d6SEugenio Pérez if (cvq_group != -ENOTSUP) { 1053152128d6SEugenio Pérez r = cvq_group; 1054152128d6SEugenio Pérez goto out; 1055152128d6SEugenio Pérez } 1056152128d6SEugenio Pérez 1057152128d6SEugenio Pérez /* 1058152128d6SEugenio Pérez * The kernel report VHOST_BACKEND_F_IOTLB_ASID if the vdpa frontend 1059152128d6SEugenio Pérez * support ASID even if the parent driver does not. The CVQ cannot be 1060152128d6SEugenio Pérez * isolated in this case. 1061152128d6SEugenio Pérez */ 1062152128d6SEugenio Pérez error_free(*errp); 1063152128d6SEugenio Pérez *errp = NULL; 1064152128d6SEugenio Pérez r = 0; 1065152128d6SEugenio Pérez goto out; 1066152128d6SEugenio Pérez } 1067152128d6SEugenio Pérez 1068152128d6SEugenio Pérez for (int i = 0; i < cvq_index; ++i) { 1069152128d6SEugenio Pérez int64_t group = vhost_vdpa_get_vring_group(device_fd, i, errp); 1070152128d6SEugenio Pérez if (unlikely(group < 0)) { 1071152128d6SEugenio Pérez r = group; 1072152128d6SEugenio Pérez goto out; 1073152128d6SEugenio Pérez } 1074152128d6SEugenio Pérez 1075152128d6SEugenio Pérez if (group == (int64_t)cvq_group) { 1076152128d6SEugenio Pérez r = 0; 1077152128d6SEugenio Pérez goto out; 1078152128d6SEugenio Pérez } 1079152128d6SEugenio Pérez } 1080152128d6SEugenio Pérez 1081152128d6SEugenio Pérez r = 1; 1082152128d6SEugenio Pérez 1083152128d6SEugenio Pérez out: 1084152128d6SEugenio Pérez status = 0; 1085152128d6SEugenio Pérez ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status); 1086152128d6SEugenio Pérez return r; 1087152128d6SEugenio Pérez } 1088152128d6SEugenio Pérez 1089654790b6SJason Wang static NetClientState *net_vhost_vdpa_init(NetClientState *peer, 1090654790b6SJason Wang const char *device, 1091654790b6SJason Wang const char *name, 109240237840SJason Wang int vdpa_device_fd, 109340237840SJason Wang int queue_pair_index, 109440237840SJason Wang int nvqs, 10951576dbb5SEugenio Pérez bool is_datapath, 10961576dbb5SEugenio Pérez bool svq, 10975c1ebd4cSEugenio Pérez struct vhost_vdpa_iova_range iova_range, 1098152128d6SEugenio Pérez uint64_t features, 1099152128d6SEugenio Pérez Error **errp) 11001e0a84eaSCindy Lu { 11011e0a84eaSCindy Lu NetClientState *nc = NULL; 11021e0a84eaSCindy Lu VhostVDPAState *s; 11031e0a84eaSCindy Lu int ret = 0; 11041e0a84eaSCindy Lu assert(name); 1105152128d6SEugenio Pérez int cvq_isolated; 1106152128d6SEugenio Pérez 110740237840SJason Wang if (is_datapath) { 110840237840SJason Wang nc = qemu_new_net_client(&net_vhost_vdpa_info, peer, device, 110940237840SJason Wang name); 111040237840SJason Wang } else { 1111152128d6SEugenio Pérez cvq_isolated = vhost_vdpa_probe_cvq_isolation(vdpa_device_fd, features, 1112152128d6SEugenio Pérez queue_pair_index * 2, 1113152128d6SEugenio Pérez errp); 1114152128d6SEugenio Pérez if (unlikely(cvq_isolated < 0)) { 1115152128d6SEugenio Pérez return NULL; 1116152128d6SEugenio Pérez } 1117152128d6SEugenio Pérez 1118f8972b56SEugenio Pérez nc = qemu_new_net_control_client(&net_vhost_vdpa_cvq_info, peer, 111940237840SJason Wang device, name); 112040237840SJason Wang } 112153b85d95SLaurent Vivier qemu_set_info_str(nc, TYPE_VHOST_VDPA); 11221e0a84eaSCindy Lu s = DO_UPCAST(VhostVDPAState, nc, nc); 11237327813dSJason Wang 11241e0a84eaSCindy Lu s->vhost_vdpa.device_fd = vdpa_device_fd; 112540237840SJason Wang s->vhost_vdpa.index = queue_pair_index; 11267f211a28SEugenio Pérez s->always_svq = svq; 112769498430SEugenio Pérez s->migration_state.notify = vdpa_net_migration_state_notifier; 11281576dbb5SEugenio Pérez s->vhost_vdpa.shadow_vqs_enabled = svq; 1129a585fad2SEugenio Pérez s->vhost_vdpa.iova_range = iova_range; 11306188d78aSEugenio Pérez s->vhost_vdpa.shadow_data = svq; 11315c1ebd4cSEugenio Pérez if (queue_pair_index == 0) { 11325c1ebd4cSEugenio Pérez vhost_vdpa_net_valid_svq_features(features, 11335c1ebd4cSEugenio Pérez &s->vhost_vdpa.migration_blocker); 11345c1ebd4cSEugenio Pérez } else if (!is_datapath) { 1135babf8b87SEugenio Pérez s->cvq_cmd_out_buffer = mmap(NULL, vhost_vdpa_net_cvq_cmd_page_len(), 1136babf8b87SEugenio Pérez PROT_READ | PROT_WRITE, 1137babf8b87SEugenio Pérez MAP_SHARED | MAP_ANONYMOUS, -1, 0); 1138babf8b87SEugenio Pérez s->status = mmap(NULL, vhost_vdpa_net_cvq_cmd_page_len(), 1139babf8b87SEugenio Pérez PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANONYMOUS, 1140babf8b87SEugenio Pérez -1, 0); 11412df4dd31SEugenio Pérez 1142bd907ae4SEugenio Pérez s->vhost_vdpa.shadow_vq_ops = &vhost_vdpa_net_svq_ops; 1143bd907ae4SEugenio Pérez s->vhost_vdpa.shadow_vq_ops_opaque = s; 1144152128d6SEugenio Pérez s->cvq_isolated = cvq_isolated; 11459c363cf6SEugenio Pérez 11469c363cf6SEugenio Pérez /* 11478bc0049eSEugenio Pérez * TODO: We cannot migrate devices with CVQ and no x-svq enabled as 11488bc0049eSEugenio Pérez * there is no way to set the device state (MAC, MQ, etc) before 11498bc0049eSEugenio Pérez * starting the datapath. 11509c363cf6SEugenio Pérez * 11519c363cf6SEugenio Pérez * Migration blocker ownership now belongs to s->vhost_vdpa. 11529c363cf6SEugenio Pérez */ 11538bc0049eSEugenio Pérez if (!svq) { 11549c363cf6SEugenio Pérez error_setg(&s->vhost_vdpa.migration_blocker, 11559c363cf6SEugenio Pérez "net vdpa cannot migrate with CVQ feature"); 1156bd907ae4SEugenio Pérez } 11578bc0049eSEugenio Pérez } 115840237840SJason Wang ret = vhost_vdpa_add(nc, (void *)&s->vhost_vdpa, queue_pair_index, nvqs); 115974af5eecSJason Wang if (ret) { 116074af5eecSJason Wang qemu_del_net_client(nc); 1161654790b6SJason Wang return NULL; 116274af5eecSJason Wang } 1163654790b6SJason Wang return nc; 11641e0a84eaSCindy Lu } 11651e0a84eaSCindy Lu 11668170ab3fSEugenio Pérez static int vhost_vdpa_get_features(int fd, uint64_t *features, Error **errp) 11678170ab3fSEugenio Pérez { 11688170ab3fSEugenio Pérez int ret = ioctl(fd, VHOST_GET_FEATURES, features); 11698170ab3fSEugenio Pérez if (unlikely(ret < 0)) { 11708170ab3fSEugenio Pérez error_setg_errno(errp, errno, 11718170ab3fSEugenio Pérez "Fail to query features from vhost-vDPA device"); 11728170ab3fSEugenio Pérez } 11738170ab3fSEugenio Pérez return ret; 11748170ab3fSEugenio Pérez } 11758170ab3fSEugenio Pérez 11768170ab3fSEugenio Pérez static int vhost_vdpa_get_max_queue_pairs(int fd, uint64_t features, 11778170ab3fSEugenio Pérez int *has_cvq, Error **errp) 117840237840SJason Wang { 117940237840SJason Wang unsigned long config_size = offsetof(struct vhost_vdpa_config, buf); 1180cd523a41SStefano Garzarella g_autofree struct vhost_vdpa_config *config = NULL; 118140237840SJason Wang __virtio16 *max_queue_pairs; 118240237840SJason Wang int ret; 118340237840SJason Wang 118440237840SJason Wang if (features & (1 << VIRTIO_NET_F_CTRL_VQ)) { 118540237840SJason Wang *has_cvq = 1; 118640237840SJason Wang } else { 118740237840SJason Wang *has_cvq = 0; 118840237840SJason Wang } 118940237840SJason Wang 119040237840SJason Wang if (features & (1 << VIRTIO_NET_F_MQ)) { 119140237840SJason Wang config = g_malloc0(config_size + sizeof(*max_queue_pairs)); 119240237840SJason Wang config->off = offsetof(struct virtio_net_config, max_virtqueue_pairs); 119340237840SJason Wang config->len = sizeof(*max_queue_pairs); 119440237840SJason Wang 119540237840SJason Wang ret = ioctl(fd, VHOST_VDPA_GET_CONFIG, config); 119640237840SJason Wang if (ret) { 119740237840SJason Wang error_setg(errp, "Fail to get config from vhost-vDPA device"); 119840237840SJason Wang return -ret; 119940237840SJason Wang } 120040237840SJason Wang 120140237840SJason Wang max_queue_pairs = (__virtio16 *)&config->buf; 120240237840SJason Wang 120340237840SJason Wang return lduw_le_p(max_queue_pairs); 120440237840SJason Wang } 120540237840SJason Wang 120640237840SJason Wang return 1; 120740237840SJason Wang } 120840237840SJason Wang 12091e0a84eaSCindy Lu int net_init_vhost_vdpa(const Netdev *netdev, const char *name, 12101e0a84eaSCindy Lu NetClientState *peer, Error **errp) 12111e0a84eaSCindy Lu { 12121e0a84eaSCindy Lu const NetdevVhostVDPAOptions *opts; 12138170ab3fSEugenio Pérez uint64_t features; 1214654790b6SJason Wang int vdpa_device_fd; 1215eb3cb751SEugenio Pérez g_autofree NetClientState **ncs = NULL; 1216a585fad2SEugenio Pérez struct vhost_vdpa_iova_range iova_range; 1217eb3cb751SEugenio Pérez NetClientState *nc; 1218aed5da45SEugenio Pérez int queue_pairs, r, i = 0, has_cvq = 0; 12191e0a84eaSCindy Lu 12201e0a84eaSCindy Lu assert(netdev->type == NET_CLIENT_DRIVER_VHOST_VDPA); 12211e0a84eaSCindy Lu opts = &netdev->u.vhost_vdpa; 12227480874aSMarkus Armbruster if (!opts->vhostdev && !opts->vhostfd) { 12238801ccd0SSi-Wei Liu error_setg(errp, 12248801ccd0SSi-Wei Liu "vhost-vdpa: neither vhostdev= nor vhostfd= was specified"); 1225c8295404SEugenio Pérez return -1; 1226c8295404SEugenio Pérez } 12277327813dSJason Wang 12287480874aSMarkus Armbruster if (opts->vhostdev && opts->vhostfd) { 12298801ccd0SSi-Wei Liu error_setg(errp, 12308801ccd0SSi-Wei Liu "vhost-vdpa: vhostdev= and vhostfd= are mutually exclusive"); 12318801ccd0SSi-Wei Liu return -1; 12328801ccd0SSi-Wei Liu } 12338801ccd0SSi-Wei Liu 12347480874aSMarkus Armbruster if (opts->vhostdev) { 12350351152bSEugenio Pérez vdpa_device_fd = qemu_open(opts->vhostdev, O_RDWR, errp); 12367327813dSJason Wang if (vdpa_device_fd == -1) { 12377327813dSJason Wang return -errno; 12387327813dSJason Wang } 12395107fd3eSPeter Maydell } else { 12405107fd3eSPeter Maydell /* has_vhostfd */ 12418801ccd0SSi-Wei Liu vdpa_device_fd = monitor_fd_param(monitor_cur(), opts->vhostfd, errp); 12428801ccd0SSi-Wei Liu if (vdpa_device_fd == -1) { 12438801ccd0SSi-Wei Liu error_prepend(errp, "vhost-vdpa: unable to parse vhostfd: "); 12448801ccd0SSi-Wei Liu return -1; 12458801ccd0SSi-Wei Liu } 12468801ccd0SSi-Wei Liu } 12477327813dSJason Wang 12488170ab3fSEugenio Pérez r = vhost_vdpa_get_features(vdpa_device_fd, &features, errp); 12498170ab3fSEugenio Pérez if (unlikely(r < 0)) { 1250aed5da45SEugenio Pérez goto err; 12518170ab3fSEugenio Pérez } 12528170ab3fSEugenio Pérez 12538170ab3fSEugenio Pérez queue_pairs = vhost_vdpa_get_max_queue_pairs(vdpa_device_fd, features, 125440237840SJason Wang &has_cvq, errp); 125540237840SJason Wang if (queue_pairs < 0) { 12567327813dSJason Wang qemu_close(vdpa_device_fd); 125740237840SJason Wang return queue_pairs; 12587327813dSJason Wang } 12597327813dSJason Wang 1260bf7a2ad8SLongpeng r = vhost_vdpa_get_iova_range(vdpa_device_fd, &iova_range); 1261bf7a2ad8SLongpeng if (unlikely(r < 0)) { 1262bf7a2ad8SLongpeng error_setg(errp, "vhost-vdpa: get iova range failed: %s", 1263bf7a2ad8SLongpeng strerror(-r)); 1264bf7a2ad8SLongpeng goto err; 1265bf7a2ad8SLongpeng } 1266bf7a2ad8SLongpeng 126700ef422eSEugenio Pérez if (opts->x_svq && !vhost_vdpa_net_valid_svq_features(features, errp)) { 126800ef422eSEugenio Pérez goto err; 12691576dbb5SEugenio Pérez } 12701576dbb5SEugenio Pérez 127140237840SJason Wang ncs = g_malloc0(sizeof(*ncs) * queue_pairs); 127240237840SJason Wang 127340237840SJason Wang for (i = 0; i < queue_pairs; i++) { 127440237840SJason Wang ncs[i] = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name, 12751576dbb5SEugenio Pérez vdpa_device_fd, i, 2, true, opts->x_svq, 1276152128d6SEugenio Pérez iova_range, features, errp); 127740237840SJason Wang if (!ncs[i]) 127840237840SJason Wang goto err; 127940237840SJason Wang } 128040237840SJason Wang 128140237840SJason Wang if (has_cvq) { 128240237840SJason Wang nc = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name, 12831576dbb5SEugenio Pérez vdpa_device_fd, i, 1, false, 1284152128d6SEugenio Pérez opts->x_svq, iova_range, features, errp); 128540237840SJason Wang if (!nc) 128640237840SJason Wang goto err; 128740237840SJason Wang } 128840237840SJason Wang 1289654790b6SJason Wang return 0; 129040237840SJason Wang 129140237840SJason Wang err: 129240237840SJason Wang if (i) { 12939bd05507SSi-Wei Liu for (i--; i >= 0; i--) { 12949bd05507SSi-Wei Liu qemu_del_net_client(ncs[i]); 12959bd05507SSi-Wei Liu } 129640237840SJason Wang } 12971576dbb5SEugenio Pérez 129840237840SJason Wang qemu_close(vdpa_device_fd); 129940237840SJason Wang 130040237840SJason Wang return -1; 13011e0a84eaSCindy Lu } 1302