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, 789da16849SAndrew Melnychenko VIRTIO_NET_F_GUEST_USO4, 799da16849SAndrew Melnychenko VIRTIO_NET_F_GUEST_USO6, 800145c393SAndrew Melnychenko VIRTIO_NET_F_HASH_REPORT, 812875a0caSHawkins Jiawei VIRTIO_NET_F_HOST_ECN, 822875a0caSHawkins Jiawei VIRTIO_NET_F_HOST_TSO4, 832875a0caSHawkins Jiawei VIRTIO_NET_F_HOST_TSO6, 842875a0caSHawkins Jiawei VIRTIO_NET_F_HOST_UFO, 859da16849SAndrew Melnychenko VIRTIO_NET_F_HOST_USO, 862875a0caSHawkins Jiawei VIRTIO_NET_F_MQ, 872875a0caSHawkins Jiawei VIRTIO_NET_F_MRG_RXBUF, 882875a0caSHawkins Jiawei VIRTIO_NET_F_MTU, 892875a0caSHawkins Jiawei VIRTIO_NET_F_RSS, 909aa47eddSSi-Wei Liu VIRTIO_NET_F_STATUS, 912875a0caSHawkins Jiawei VIRTIO_RING_F_EVENT_IDX, 922875a0caSHawkins Jiawei VIRTIO_RING_F_INDIRECT_DESC, 932875a0caSHawkins Jiawei 942875a0caSHawkins Jiawei /* VHOST_INVALID_FEATURE_BIT should always be the last entry */ 951e0a84eaSCindy Lu VHOST_INVALID_FEATURE_BIT 961e0a84eaSCindy Lu }; 971e0a84eaSCindy Lu 981576dbb5SEugenio Pérez /** Supported device specific feature bits with SVQ */ 991576dbb5SEugenio Pérez static const uint64_t vdpa_svq_device_features = 1001576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_CSUM) | 1011576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_GUEST_CSUM) | 1024b4a1378SHawkins Jiawei BIT_ULL(VIRTIO_NET_F_CTRL_GUEST_OFFLOADS) | 1031576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_MTU) | 1041576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_MAC) | 1051576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_GUEST_TSO4) | 1061576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_GUEST_TSO6) | 1071576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_GUEST_ECN) | 1081576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_GUEST_UFO) | 1091576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_HOST_TSO4) | 1101576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_HOST_TSO6) | 1111576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_HOST_ECN) | 1121576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_HOST_UFO) | 1131576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_MRG_RXBUF) | 1141576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_STATUS) | 1151576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_CTRL_VQ) | 116ea6eec49SHawkins Jiawei BIT_ULL(VIRTIO_NET_F_CTRL_RX) | 117e213c45aSHawkins Jiawei BIT_ULL(VIRTIO_NET_F_CTRL_VLAN) | 118d669b7bbSHawkins Jiawei BIT_ULL(VIRTIO_NET_F_CTRL_RX_EXTRA) | 11972b99a87SEugenio Pérez BIT_ULL(VIRTIO_NET_F_MQ) | 1201576dbb5SEugenio Pérez BIT_ULL(VIRTIO_F_ANY_LAYOUT) | 1211576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_CTRL_MAC_ADDR) | 122609ab4c3SEugenio Pérez /* VHOST_F_LOG_ALL is exposed by SVQ */ 123609ab4c3SEugenio Pérez BIT_ULL(VHOST_F_LOG_ALL) | 1241576dbb5SEugenio Pérez BIT_ULL(VIRTIO_NET_F_RSC_EXT) | 1250d74e2b7SEugenio Pérez BIT_ULL(VIRTIO_NET_F_STANDBY) | 1260d74e2b7SEugenio Pérez BIT_ULL(VIRTIO_NET_F_SPEED_DUPLEX); 1271576dbb5SEugenio Pérez 128c1a10086SEugenio Pérez #define VHOST_VDPA_NET_CVQ_ASID 1 129c1a10086SEugenio Pérez 1301e0a84eaSCindy Lu VHostNetState *vhost_vdpa_get_vhost_net(NetClientState *nc) 1311e0a84eaSCindy Lu { 1321e0a84eaSCindy Lu VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 1331e0a84eaSCindy Lu assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 1341e0a84eaSCindy Lu return s->vhost_net; 1351e0a84eaSCindy Lu } 1361e0a84eaSCindy Lu 137915bf6ccSEugenio Pérez static size_t vhost_vdpa_net_cvq_cmd_len(void) 138915bf6ccSEugenio Pérez { 139915bf6ccSEugenio Pérez /* 140915bf6ccSEugenio Pérez * MAC_TABLE_SET is the ctrl command that produces the longer out buffer. 141915bf6ccSEugenio Pérez * In buffer is always 1 byte, so it should fit here 142915bf6ccSEugenio Pérez */ 143915bf6ccSEugenio Pérez return sizeof(struct virtio_net_ctrl_hdr) + 144915bf6ccSEugenio Pérez 2 * sizeof(struct virtio_net_ctrl_mac) + 145915bf6ccSEugenio Pérez MAC_TABLE_ENTRIES * ETH_ALEN; 146915bf6ccSEugenio Pérez } 147915bf6ccSEugenio Pérez 148915bf6ccSEugenio Pérez static size_t vhost_vdpa_net_cvq_cmd_page_len(void) 149915bf6ccSEugenio Pérez { 150915bf6ccSEugenio Pérez return ROUND_UP(vhost_vdpa_net_cvq_cmd_len(), qemu_real_host_page_size()); 151915bf6ccSEugenio Pérez } 152915bf6ccSEugenio Pérez 15336e46472SEugenio Pérez static bool vhost_vdpa_net_valid_svq_features(uint64_t features, Error **errp) 15436e46472SEugenio Pérez { 15536e46472SEugenio Pérez uint64_t invalid_dev_features = 15636e46472SEugenio Pérez features & ~vdpa_svq_device_features & 15736e46472SEugenio Pérez /* Transport are all accepted at this point */ 15836e46472SEugenio Pérez ~MAKE_64BIT_MASK(VIRTIO_TRANSPORT_F_START, 15936e46472SEugenio Pérez VIRTIO_TRANSPORT_F_END - VIRTIO_TRANSPORT_F_START); 16036e46472SEugenio Pérez 16136e46472SEugenio Pérez if (invalid_dev_features) { 16236e46472SEugenio Pérez error_setg(errp, "vdpa svq does not work with features 0x%" PRIx64, 16336e46472SEugenio Pérez invalid_dev_features); 164258a0394SEugenio Pérez return false; 16536e46472SEugenio Pérez } 16636e46472SEugenio Pérez 167258a0394SEugenio Pérez return vhost_svq_valid_features(features, errp); 16836e46472SEugenio Pérez } 16936e46472SEugenio Pérez 1701e0a84eaSCindy Lu static int vhost_vdpa_net_check_device_id(struct vhost_net *net) 1711e0a84eaSCindy Lu { 1721e0a84eaSCindy Lu uint32_t device_id; 1731e0a84eaSCindy Lu int ret; 1741e0a84eaSCindy Lu struct vhost_dev *hdev; 1751e0a84eaSCindy Lu 1761e0a84eaSCindy Lu hdev = (struct vhost_dev *)&net->dev; 1771e0a84eaSCindy Lu ret = hdev->vhost_ops->vhost_get_device_id(hdev, &device_id); 1781e0a84eaSCindy Lu if (device_id != VIRTIO_ID_NET) { 1791e0a84eaSCindy Lu return -ENOTSUP; 1801e0a84eaSCindy Lu } 1811e0a84eaSCindy Lu return ret; 1821e0a84eaSCindy Lu } 1831e0a84eaSCindy Lu 18440237840SJason Wang static int vhost_vdpa_add(NetClientState *ncs, void *be, 18540237840SJason Wang int queue_pair_index, int nvqs) 1861e0a84eaSCindy Lu { 1871e0a84eaSCindy Lu VhostNetOptions options; 1881e0a84eaSCindy Lu struct vhost_net *net = NULL; 1891e0a84eaSCindy Lu VhostVDPAState *s; 1901e0a84eaSCindy Lu int ret; 1911e0a84eaSCindy Lu 1921e0a84eaSCindy Lu options.backend_type = VHOST_BACKEND_TYPE_VDPA; 1931e0a84eaSCindy Lu assert(ncs->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 1941e0a84eaSCindy Lu s = DO_UPCAST(VhostVDPAState, nc, ncs); 1951e0a84eaSCindy Lu options.net_backend = ncs; 1961e0a84eaSCindy Lu options.opaque = be; 1971e0a84eaSCindy Lu options.busyloop_timeout = 0; 19840237840SJason Wang options.nvqs = nvqs; 1991e0a84eaSCindy Lu 2001e0a84eaSCindy Lu net = vhost_net_init(&options); 2011e0a84eaSCindy Lu if (!net) { 2021e0a84eaSCindy Lu error_report("failed to init vhost_net for queue"); 203a97ef87aSJason Wang goto err_init; 2041e0a84eaSCindy Lu } 2051e0a84eaSCindy Lu s->vhost_net = net; 2061e0a84eaSCindy Lu ret = vhost_vdpa_net_check_device_id(net); 2071e0a84eaSCindy Lu if (ret) { 208a97ef87aSJason Wang goto err_check; 2091e0a84eaSCindy Lu } 2101e0a84eaSCindy Lu return 0; 211a97ef87aSJason Wang err_check: 2121e0a84eaSCindy Lu vhost_net_cleanup(net); 213ab36edcfSJason Wang g_free(net); 214a97ef87aSJason Wang err_init: 2151e0a84eaSCindy Lu return -1; 2161e0a84eaSCindy Lu } 2171e0a84eaSCindy Lu 2181e0a84eaSCindy Lu static void vhost_vdpa_cleanup(NetClientState *nc) 2191e0a84eaSCindy Lu { 2201e0a84eaSCindy Lu VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 2211e0a84eaSCindy Lu 222a0d7215eSAni Sinha /* 223a0d7215eSAni Sinha * If a peer NIC is attached, do not cleanup anything. 224a0d7215eSAni Sinha * Cleanup will happen as a part of qemu_cleanup() -> net_cleanup() 225a0d7215eSAni Sinha * when the guest is shutting down. 226a0d7215eSAni Sinha */ 227a0d7215eSAni Sinha if (nc->peer && nc->peer->info->type == NET_CLIENT_DRIVER_NIC) { 228a0d7215eSAni Sinha return; 229a0d7215eSAni Sinha } 230babf8b87SEugenio Pérez munmap(s->cvq_cmd_out_buffer, vhost_vdpa_net_cvq_cmd_page_len()); 231babf8b87SEugenio Pérez munmap(s->status, vhost_vdpa_net_cvq_cmd_page_len()); 2321e0a84eaSCindy Lu if (s->vhost_net) { 2331e0a84eaSCindy Lu vhost_net_cleanup(s->vhost_net); 2341e0a84eaSCindy Lu g_free(s->vhost_net); 2351e0a84eaSCindy Lu s->vhost_net = NULL; 2361e0a84eaSCindy Lu } 23757b3a7d8SCindy Lu if (s->vhost_vdpa.device_fd >= 0) { 23857b3a7d8SCindy Lu qemu_close(s->vhost_vdpa.device_fd); 23957b3a7d8SCindy Lu s->vhost_vdpa.device_fd = -1; 24057b3a7d8SCindy Lu } 2411e0a84eaSCindy Lu } 2421e0a84eaSCindy Lu 2431e0a84eaSCindy Lu static bool vhost_vdpa_has_vnet_hdr(NetClientState *nc) 2441e0a84eaSCindy Lu { 2451e0a84eaSCindy Lu assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 2461e0a84eaSCindy Lu 2471e0a84eaSCindy Lu return true; 2481e0a84eaSCindy Lu } 2491e0a84eaSCindy Lu 2501e0a84eaSCindy Lu static bool vhost_vdpa_has_ufo(NetClientState *nc) 2511e0a84eaSCindy Lu { 2521e0a84eaSCindy Lu assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 2531e0a84eaSCindy Lu VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 2541e0a84eaSCindy Lu uint64_t features = 0; 2551e0a84eaSCindy Lu features |= (1ULL << VIRTIO_NET_F_HOST_UFO); 2561e0a84eaSCindy Lu features = vhost_net_get_features(s->vhost_net, features); 2571e0a84eaSCindy Lu return !!(features & (1ULL << VIRTIO_NET_F_HOST_UFO)); 2581e0a84eaSCindy Lu 2591e0a84eaSCindy Lu } 2601e0a84eaSCindy Lu 261ee8a1c63SKevin Wolf static bool vhost_vdpa_check_peer_type(NetClientState *nc, ObjectClass *oc, 262ee8a1c63SKevin Wolf Error **errp) 263ee8a1c63SKevin Wolf { 264ee8a1c63SKevin Wolf const char *driver = object_class_get_name(oc); 265ee8a1c63SKevin Wolf 266ee8a1c63SKevin Wolf if (!g_str_has_prefix(driver, "virtio-net-")) { 267ee8a1c63SKevin Wolf error_setg(errp, "vhost-vdpa requires frontend driver virtio-net-*"); 268ee8a1c63SKevin Wolf return false; 269ee8a1c63SKevin Wolf } 270ee8a1c63SKevin Wolf 271ee8a1c63SKevin Wolf return true; 272ee8a1c63SKevin Wolf } 273ee8a1c63SKevin Wolf 274846a1e85SEugenio Pérez /** Dummy receive in case qemu falls back to userland tap networking */ 275846a1e85SEugenio Pérez static ssize_t vhost_vdpa_receive(NetClientState *nc, const uint8_t *buf, 276846a1e85SEugenio Pérez size_t size) 277846a1e85SEugenio Pérez { 278bc5add1dSSi-Wei Liu return size; 279846a1e85SEugenio Pérez } 280846a1e85SEugenio Pérez 28100ef422eSEugenio Pérez /** From any vdpa net client, get the netclient of the first queue pair */ 28200ef422eSEugenio Pérez static VhostVDPAState *vhost_vdpa_net_first_nc_vdpa(VhostVDPAState *s) 28300ef422eSEugenio Pérez { 28400ef422eSEugenio Pérez NICState *nic = qemu_get_nic(s->nc.peer); 28500ef422eSEugenio Pérez NetClientState *nc0 = qemu_get_peer(nic->ncs, 0); 28600ef422eSEugenio Pérez 28700ef422eSEugenio Pérez return DO_UPCAST(VhostVDPAState, nc, nc0); 28800ef422eSEugenio Pérez } 28900ef422eSEugenio Pérez 29069498430SEugenio Pérez static void vhost_vdpa_net_log_global_enable(VhostVDPAState *s, bool enable) 29169498430SEugenio Pérez { 29269498430SEugenio Pérez struct vhost_vdpa *v = &s->vhost_vdpa; 29369498430SEugenio Pérez VirtIONet *n; 29469498430SEugenio Pérez VirtIODevice *vdev; 29569498430SEugenio Pérez int data_queue_pairs, cvq, r; 29669498430SEugenio Pérez 29769498430SEugenio Pérez /* We are only called on the first data vqs and only if x-svq is not set */ 29869498430SEugenio Pérez if (s->vhost_vdpa.shadow_vqs_enabled == enable) { 29969498430SEugenio Pérez return; 30069498430SEugenio Pérez } 30169498430SEugenio Pérez 30269498430SEugenio Pérez vdev = v->dev->vdev; 30369498430SEugenio Pérez n = VIRTIO_NET(vdev); 30469498430SEugenio Pérez if (!n->vhost_started) { 30569498430SEugenio Pérez return; 30669498430SEugenio Pérez } 30769498430SEugenio Pérez 30869498430SEugenio Pérez data_queue_pairs = n->multiqueue ? n->max_queue_pairs : 1; 30969498430SEugenio Pérez cvq = virtio_vdev_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) ? 31069498430SEugenio Pérez n->max_ncs - n->max_queue_pairs : 0; 31169498430SEugenio Pérez /* 31269498430SEugenio Pérez * TODO: vhost_net_stop does suspend, get_base and reset. We can be smarter 31369498430SEugenio Pérez * in the future and resume the device if read-only operations between 31469498430SEugenio Pérez * suspend and reset goes wrong. 31569498430SEugenio Pérez */ 31669498430SEugenio Pérez vhost_net_stop(vdev, n->nic->ncs, data_queue_pairs, cvq); 31769498430SEugenio Pérez 31869498430SEugenio Pérez /* Start will check migration setup_or_active to configure or not SVQ */ 31969498430SEugenio Pérez r = vhost_net_start(vdev, n->nic->ncs, data_queue_pairs, cvq); 32069498430SEugenio Pérez if (unlikely(r < 0)) { 32169498430SEugenio Pérez error_report("unable to start vhost net: %s(%d)", g_strerror(-r), -r); 32269498430SEugenio Pérez } 32369498430SEugenio Pérez } 32469498430SEugenio Pérez 32569498430SEugenio Pérez static void vdpa_net_migration_state_notifier(Notifier *notifier, void *data) 32669498430SEugenio Pérez { 32769498430SEugenio Pérez MigrationState *migration = data; 32869498430SEugenio Pérez VhostVDPAState *s = container_of(notifier, VhostVDPAState, 32969498430SEugenio Pérez migration_state); 33069498430SEugenio Pérez 33169498430SEugenio Pérez if (migration_in_setup(migration)) { 33269498430SEugenio Pérez vhost_vdpa_net_log_global_enable(s, true); 33369498430SEugenio Pérez } else if (migration_has_failed(migration)) { 33469498430SEugenio Pérez vhost_vdpa_net_log_global_enable(s, false); 33569498430SEugenio Pérez } 33669498430SEugenio Pérez } 33769498430SEugenio Pérez 33800ef422eSEugenio Pérez static void vhost_vdpa_net_data_start_first(VhostVDPAState *s) 33900ef422eSEugenio Pérez { 34000ef422eSEugenio Pérez struct vhost_vdpa *v = &s->vhost_vdpa; 34100ef422eSEugenio Pérez 34269498430SEugenio Pérez add_migration_state_change_notifier(&s->migration_state); 34300ef422eSEugenio Pérez if (v->shadow_vqs_enabled) { 34400ef422eSEugenio Pérez v->iova_tree = vhost_iova_tree_new(v->iova_range.first, 34500ef422eSEugenio Pérez v->iova_range.last); 34600ef422eSEugenio Pérez } 34700ef422eSEugenio Pérez } 34800ef422eSEugenio Pérez 34900ef422eSEugenio Pérez static int vhost_vdpa_net_data_start(NetClientState *nc) 35000ef422eSEugenio Pérez { 35100ef422eSEugenio Pérez VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 35200ef422eSEugenio Pérez struct vhost_vdpa *v = &s->vhost_vdpa; 35300ef422eSEugenio Pérez 35400ef422eSEugenio Pérez assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 35500ef422eSEugenio Pérez 35669498430SEugenio Pérez if (s->always_svq || 35769498430SEugenio Pérez migration_is_setup_or_active(migrate_get_current()->state)) { 35869498430SEugenio Pérez v->shadow_vqs_enabled = true; 35969498430SEugenio Pérez v->shadow_data = true; 36069498430SEugenio Pérez } else { 36169498430SEugenio Pérez v->shadow_vqs_enabled = false; 36269498430SEugenio Pérez v->shadow_data = false; 36369498430SEugenio Pérez } 36469498430SEugenio Pérez 36500ef422eSEugenio Pérez if (v->index == 0) { 36600ef422eSEugenio Pérez vhost_vdpa_net_data_start_first(s); 36700ef422eSEugenio Pérez return 0; 36800ef422eSEugenio Pérez } 36900ef422eSEugenio Pérez 37000ef422eSEugenio Pérez if (v->shadow_vqs_enabled) { 37100ef422eSEugenio Pérez VhostVDPAState *s0 = vhost_vdpa_net_first_nc_vdpa(s); 37200ef422eSEugenio Pérez v->iova_tree = s0->vhost_vdpa.iova_tree; 37300ef422eSEugenio Pérez } 37400ef422eSEugenio Pérez 37500ef422eSEugenio Pérez return 0; 37600ef422eSEugenio Pérez } 37700ef422eSEugenio Pérez 37800ef422eSEugenio Pérez static void vhost_vdpa_net_client_stop(NetClientState *nc) 37900ef422eSEugenio Pérez { 38000ef422eSEugenio Pérez VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 38100ef422eSEugenio Pérez struct vhost_dev *dev; 38200ef422eSEugenio Pérez 38300ef422eSEugenio Pérez assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 38400ef422eSEugenio Pérez 38569498430SEugenio Pérez if (s->vhost_vdpa.index == 0) { 38669498430SEugenio Pérez remove_migration_state_change_notifier(&s->migration_state); 38769498430SEugenio Pérez } 38869498430SEugenio Pérez 38900ef422eSEugenio Pérez dev = s->vhost_vdpa.dev; 39000ef422eSEugenio Pérez if (dev->vq_index + dev->nvqs == dev->vq_index_end) { 39100ef422eSEugenio Pérez g_clear_pointer(&s->vhost_vdpa.iova_tree, vhost_iova_tree_delete); 39200ef422eSEugenio Pérez } 39300ef422eSEugenio Pérez } 39400ef422eSEugenio Pérez 3951e0a84eaSCindy Lu static NetClientInfo net_vhost_vdpa_info = { 3961e0a84eaSCindy Lu .type = NET_CLIENT_DRIVER_VHOST_VDPA, 3971e0a84eaSCindy Lu .size = sizeof(VhostVDPAState), 398846a1e85SEugenio Pérez .receive = vhost_vdpa_receive, 39900ef422eSEugenio Pérez .start = vhost_vdpa_net_data_start, 40000ef422eSEugenio Pérez .stop = vhost_vdpa_net_client_stop, 4011e0a84eaSCindy Lu .cleanup = vhost_vdpa_cleanup, 4021e0a84eaSCindy Lu .has_vnet_hdr = vhost_vdpa_has_vnet_hdr, 4031e0a84eaSCindy Lu .has_ufo = vhost_vdpa_has_ufo, 404ee8a1c63SKevin Wolf .check_peer_type = vhost_vdpa_check_peer_type, 4051e0a84eaSCindy Lu }; 4061e0a84eaSCindy Lu 407152128d6SEugenio Pérez static int64_t vhost_vdpa_get_vring_group(int device_fd, unsigned vq_index, 408152128d6SEugenio Pérez Error **errp) 409c1a10086SEugenio Pérez { 410c1a10086SEugenio Pérez struct vhost_vring_state state = { 411c1a10086SEugenio Pérez .index = vq_index, 412c1a10086SEugenio Pérez }; 413c1a10086SEugenio Pérez int r = ioctl(device_fd, VHOST_VDPA_GET_VRING_GROUP, &state); 414c1a10086SEugenio Pérez 415c1a10086SEugenio Pérez if (unlikely(r < 0)) { 4160f2bb0bfSEugenio Pérez r = -errno; 417152128d6SEugenio Pérez error_setg_errno(errp, errno, "Cannot get VQ %u group", vq_index); 418c1a10086SEugenio Pérez return r; 419c1a10086SEugenio Pérez } 420c1a10086SEugenio Pérez 421c1a10086SEugenio Pérez return state.num; 422c1a10086SEugenio Pérez } 423c1a10086SEugenio Pérez 424c1a10086SEugenio Pérez static int vhost_vdpa_set_address_space_id(struct vhost_vdpa *v, 425c1a10086SEugenio Pérez unsigned vq_group, 426c1a10086SEugenio Pérez unsigned asid_num) 427c1a10086SEugenio Pérez { 428c1a10086SEugenio Pérez struct vhost_vring_state asid = { 429c1a10086SEugenio Pérez .index = vq_group, 430c1a10086SEugenio Pérez .num = asid_num, 431c1a10086SEugenio Pérez }; 432c1a10086SEugenio Pérez int r; 433c1a10086SEugenio Pérez 434c1a10086SEugenio Pérez r = ioctl(v->device_fd, VHOST_VDPA_SET_GROUP_ASID, &asid); 435c1a10086SEugenio Pérez if (unlikely(r < 0)) { 436c1a10086SEugenio Pérez error_report("Can't set vq group %u asid %u, errno=%d (%s)", 437c1a10086SEugenio Pérez asid.index, asid.num, errno, g_strerror(errno)); 438c1a10086SEugenio Pérez } 439c1a10086SEugenio Pérez return r; 440c1a10086SEugenio Pérez } 441c1a10086SEugenio Pérez 4422df4dd31SEugenio Pérez static void vhost_vdpa_cvq_unmap_buf(struct vhost_vdpa *v, void *addr) 4432df4dd31SEugenio Pérez { 4442df4dd31SEugenio Pérez VhostIOVATree *tree = v->iova_tree; 4452df4dd31SEugenio Pérez DMAMap needle = { 4462df4dd31SEugenio Pérez /* 4472df4dd31SEugenio Pérez * No need to specify size or to look for more translations since 4482df4dd31SEugenio Pérez * this contiguous chunk was allocated by us. 4492df4dd31SEugenio Pérez */ 4502df4dd31SEugenio Pérez .translated_addr = (hwaddr)(uintptr_t)addr, 4512df4dd31SEugenio Pérez }; 4522df4dd31SEugenio Pérez const DMAMap *map = vhost_iova_tree_find_iova(tree, &needle); 4532df4dd31SEugenio Pérez int r; 4542df4dd31SEugenio Pérez 4552df4dd31SEugenio Pérez if (unlikely(!map)) { 4562df4dd31SEugenio Pérez error_report("Cannot locate expected map"); 4572df4dd31SEugenio Pérez return; 4582df4dd31SEugenio Pérez } 4592df4dd31SEugenio Pérez 460cd831ed5SEugenio Pérez r = vhost_vdpa_dma_unmap(v, v->address_space_id, map->iova, map->size + 1); 4612df4dd31SEugenio Pérez if (unlikely(r != 0)) { 4622df4dd31SEugenio Pérez error_report("Device cannot unmap: %s(%d)", g_strerror(r), r); 4632df4dd31SEugenio Pérez } 4642df4dd31SEugenio Pérez 46569292a8eSEugenio Pérez vhost_iova_tree_remove(tree, *map); 4662df4dd31SEugenio Pérez } 4672df4dd31SEugenio Pérez 4687a7f87e9SEugenio Pérez /** Map CVQ buffer. */ 4697a7f87e9SEugenio Pérez static int vhost_vdpa_cvq_map_buf(struct vhost_vdpa *v, void *buf, size_t size, 4707a7f87e9SEugenio Pérez bool write) 4712df4dd31SEugenio Pérez { 4722df4dd31SEugenio Pérez DMAMap map = {}; 4732df4dd31SEugenio Pérez int r; 4742df4dd31SEugenio Pérez 4752df4dd31SEugenio Pérez map.translated_addr = (hwaddr)(uintptr_t)buf; 4767a7f87e9SEugenio Pérez map.size = size - 1; 4772df4dd31SEugenio Pérez map.perm = write ? IOMMU_RW : IOMMU_RO, 4782df4dd31SEugenio Pérez r = vhost_iova_tree_map_alloc(v->iova_tree, &map); 4792df4dd31SEugenio Pérez if (unlikely(r != IOVA_OK)) { 4802df4dd31SEugenio Pérez error_report("Cannot map injected element"); 4817a7f87e9SEugenio Pérez return r; 4822df4dd31SEugenio Pérez } 4832df4dd31SEugenio Pérez 484cd831ed5SEugenio Pérez r = vhost_vdpa_dma_map(v, v->address_space_id, map.iova, 485cd831ed5SEugenio Pérez vhost_vdpa_net_cvq_cmd_page_len(), buf, !write); 4862df4dd31SEugenio Pérez if (unlikely(r < 0)) { 4872df4dd31SEugenio Pérez goto dma_map_err; 4882df4dd31SEugenio Pérez } 4892df4dd31SEugenio Pérez 4907a7f87e9SEugenio Pérez return 0; 4912df4dd31SEugenio Pérez 4922df4dd31SEugenio Pérez dma_map_err: 49369292a8eSEugenio Pérez vhost_iova_tree_remove(v->iova_tree, map); 4947a7f87e9SEugenio Pérez return r; 4952df4dd31SEugenio Pérez } 4962df4dd31SEugenio Pérez 4977a7f87e9SEugenio Pérez static int vhost_vdpa_net_cvq_start(NetClientState *nc) 4982df4dd31SEugenio Pérez { 49900ef422eSEugenio Pérez VhostVDPAState *s, *s0; 500c1a10086SEugenio Pérez struct vhost_vdpa *v; 501c1a10086SEugenio Pérez int64_t cvq_group; 502152128d6SEugenio Pérez int r; 503152128d6SEugenio Pérez Error *err = NULL; 5042df4dd31SEugenio Pérez 5057a7f87e9SEugenio Pérez assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 5067a7f87e9SEugenio Pérez 5077a7f87e9SEugenio Pérez s = DO_UPCAST(VhostVDPAState, nc, nc); 508c1a10086SEugenio Pérez v = &s->vhost_vdpa; 509c1a10086SEugenio Pérez 51069498430SEugenio Pérez s0 = vhost_vdpa_net_first_nc_vdpa(s); 51169498430SEugenio Pérez v->shadow_data = s0->vhost_vdpa.shadow_vqs_enabled; 512b40eba9cSEugenio Pérez v->shadow_vqs_enabled = s0->vhost_vdpa.shadow_vqs_enabled; 513c1a10086SEugenio Pérez s->vhost_vdpa.address_space_id = VHOST_VDPA_GUEST_PA_ASID; 514c1a10086SEugenio Pérez 51569498430SEugenio Pérez if (s->vhost_vdpa.shadow_data) { 516c1a10086SEugenio Pérez /* SVQ is already configured for all virtqueues */ 517c1a10086SEugenio Pérez goto out; 518c1a10086SEugenio Pérez } 519c1a10086SEugenio Pérez 520c1a10086SEugenio Pérez /* 521c1a10086SEugenio Pérez * If we early return in these cases SVQ will not be enabled. The migration 522c1a10086SEugenio Pérez * will be blocked as long as vhost-vdpa backends will not offer _F_LOG. 523c1a10086SEugenio Pérez */ 524152128d6SEugenio Pérez if (!vhost_vdpa_net_valid_svq_features(v->dev->features, NULL)) { 525c1a10086SEugenio Pérez return 0; 526c1a10086SEugenio Pérez } 527c1a10086SEugenio Pérez 528152128d6SEugenio Pérez if (!s->cvq_isolated) { 529152128d6SEugenio Pérez return 0; 530152128d6SEugenio Pérez } 531152128d6SEugenio Pérez 532152128d6SEugenio Pérez cvq_group = vhost_vdpa_get_vring_group(v->device_fd, 533152128d6SEugenio Pérez v->dev->vq_index_end - 1, 534152128d6SEugenio Pérez &err); 535c1a10086SEugenio Pérez if (unlikely(cvq_group < 0)) { 536152128d6SEugenio Pérez error_report_err(err); 537c1a10086SEugenio Pérez return cvq_group; 538c1a10086SEugenio Pérez } 539c1a10086SEugenio Pérez 540c1a10086SEugenio Pérez r = vhost_vdpa_set_address_space_id(v, cvq_group, VHOST_VDPA_NET_CVQ_ASID); 541c1a10086SEugenio Pérez if (unlikely(r < 0)) { 542c1a10086SEugenio Pérez return r; 543c1a10086SEugenio Pérez } 544c1a10086SEugenio Pérez 545c1a10086SEugenio Pérez v->shadow_vqs_enabled = true; 546c1a10086SEugenio Pérez s->vhost_vdpa.address_space_id = VHOST_VDPA_NET_CVQ_ASID; 547c1a10086SEugenio Pérez 548c1a10086SEugenio Pérez out: 5497a7f87e9SEugenio Pérez if (!s->vhost_vdpa.shadow_vqs_enabled) { 5507a7f87e9SEugenio Pérez return 0; 5512df4dd31SEugenio Pérez } 5522df4dd31SEugenio Pérez 55300ef422eSEugenio Pérez if (s0->vhost_vdpa.iova_tree) { 55400ef422eSEugenio Pérez /* 55500ef422eSEugenio Pérez * SVQ is already configured for all virtqueues. Reuse IOVA tree for 55600ef422eSEugenio Pérez * simplicity, whether CVQ shares ASID with guest or not, because: 55700ef422eSEugenio Pérez * - Memory listener need access to guest's memory addresses allocated 55800ef422eSEugenio Pérez * in the IOVA tree. 55900ef422eSEugenio Pérez * - There should be plenty of IOVA address space for both ASID not to 56000ef422eSEugenio Pérez * worry about collisions between them. Guest's translations are 56100ef422eSEugenio Pérez * still validated with virtio virtqueue_pop so there is no risk for 56200ef422eSEugenio Pérez * the guest to access memory that it shouldn't. 56300ef422eSEugenio Pérez * 56400ef422eSEugenio Pérez * To allocate a iova tree per ASID is doable but it complicates the 56500ef422eSEugenio Pérez * code and it is not worth it for the moment. 56600ef422eSEugenio Pérez */ 56700ef422eSEugenio Pérez v->iova_tree = s0->vhost_vdpa.iova_tree; 56800ef422eSEugenio Pérez } else { 56900ef422eSEugenio Pérez v->iova_tree = vhost_iova_tree_new(v->iova_range.first, 57000ef422eSEugenio Pérez v->iova_range.last); 57100ef422eSEugenio Pérez } 57200ef422eSEugenio Pérez 5737a7f87e9SEugenio Pérez r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer, 5747a7f87e9SEugenio Pérez vhost_vdpa_net_cvq_cmd_page_len(), false); 5757a7f87e9SEugenio Pérez if (unlikely(r < 0)) { 5767a7f87e9SEugenio Pérez return r; 5777a7f87e9SEugenio Pérez } 5787a7f87e9SEugenio Pérez 57917fb889fSEugenio Pérez r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->status, 5807a7f87e9SEugenio Pérez vhost_vdpa_net_cvq_cmd_page_len(), true); 5817a7f87e9SEugenio Pérez if (unlikely(r < 0)) { 5822df4dd31SEugenio Pérez vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer); 5832df4dd31SEugenio Pérez } 5842df4dd31SEugenio Pérez 5857a7f87e9SEugenio Pérez return r; 5867a7f87e9SEugenio Pérez } 5877a7f87e9SEugenio Pérez 5887a7f87e9SEugenio Pérez static void vhost_vdpa_net_cvq_stop(NetClientState *nc) 5897a7f87e9SEugenio Pérez { 5907a7f87e9SEugenio Pérez VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 5917a7f87e9SEugenio Pérez 5927a7f87e9SEugenio Pérez assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 5937a7f87e9SEugenio Pérez 5947a7f87e9SEugenio Pérez if (s->vhost_vdpa.shadow_vqs_enabled) { 5957a7f87e9SEugenio Pérez vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer); 59617fb889fSEugenio Pérez vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->status); 597c1a10086SEugenio Pérez } 59800ef422eSEugenio Pérez 59900ef422eSEugenio Pérez vhost_vdpa_net_client_stop(nc); 6002df4dd31SEugenio Pérez } 6012df4dd31SEugenio Pérez 602be4278b6SEugenio Pérez static ssize_t vhost_vdpa_net_cvq_add(VhostVDPAState *s, size_t out_len, 603be4278b6SEugenio Pérez size_t in_len) 604be4278b6SEugenio Pérez { 605be4278b6SEugenio Pérez /* Buffers for the device */ 606be4278b6SEugenio Pérez const struct iovec out = { 607be4278b6SEugenio Pérez .iov_base = s->cvq_cmd_out_buffer, 608be4278b6SEugenio Pérez .iov_len = out_len, 609be4278b6SEugenio Pérez }; 610be4278b6SEugenio Pérez const struct iovec in = { 61117fb889fSEugenio Pérez .iov_base = s->status, 612be4278b6SEugenio Pérez .iov_len = sizeof(virtio_net_ctrl_ack), 613be4278b6SEugenio Pérez }; 614be4278b6SEugenio Pérez VhostShadowVirtqueue *svq = g_ptr_array_index(s->vhost_vdpa.shadow_vqs, 0); 615be4278b6SEugenio Pérez int r; 616be4278b6SEugenio Pérez 617be4278b6SEugenio Pérez r = vhost_svq_add(svq, &out, 1, &in, 1, NULL); 618be4278b6SEugenio Pérez if (unlikely(r != 0)) { 619be4278b6SEugenio Pérez if (unlikely(r == -ENOSPC)) { 620be4278b6SEugenio Pérez qemu_log_mask(LOG_GUEST_ERROR, "%s: No space on device queue\n", 621be4278b6SEugenio Pérez __func__); 622be4278b6SEugenio Pérez } 623be4278b6SEugenio Pérez return r; 624be4278b6SEugenio Pérez } 625be4278b6SEugenio Pérez 626be4278b6SEugenio Pérez /* 627be4278b6SEugenio Pérez * We can poll here since we've had BQL from the time we sent the 628be4278b6SEugenio Pérez * descriptor. Also, we need to take the answer before SVQ pulls by itself, 629be4278b6SEugenio Pérez * when BQL is released 630be4278b6SEugenio Pérez */ 631be4278b6SEugenio Pérez return vhost_svq_poll(svq); 632be4278b6SEugenio Pérez } 633be4278b6SEugenio Pérez 634f73c0c43SEugenio Pérez static ssize_t vhost_vdpa_net_load_cmd(VhostVDPAState *s, uint8_t class, 6352848c6aaSHawkins Jiawei uint8_t cmd, const struct iovec *data_sg, 6362848c6aaSHawkins Jiawei size_t data_num) 637f73c0c43SEugenio Pérez { 638f73c0c43SEugenio Pérez const struct virtio_net_ctrl_hdr ctrl = { 639f73c0c43SEugenio Pérez .class = class, 640f73c0c43SEugenio Pérez .cmd = cmd, 641f73c0c43SEugenio Pérez }; 6422848c6aaSHawkins Jiawei size_t data_size = iov_size(data_sg, data_num); 643f73c0c43SEugenio Pérez 644f73c0c43SEugenio Pérez assert(data_size < vhost_vdpa_net_cvq_cmd_page_len() - sizeof(ctrl)); 645f73c0c43SEugenio Pérez 6462848c6aaSHawkins Jiawei /* pack the CVQ command header */ 647f73c0c43SEugenio Pérez memcpy(s->cvq_cmd_out_buffer, &ctrl, sizeof(ctrl)); 648f73c0c43SEugenio Pérez 6492848c6aaSHawkins Jiawei /* pack the CVQ command command-specific-data */ 6502848c6aaSHawkins Jiawei iov_to_buf(data_sg, data_num, 0, 6512848c6aaSHawkins Jiawei s->cvq_cmd_out_buffer + sizeof(ctrl), data_size); 6522848c6aaSHawkins Jiawei 6532848c6aaSHawkins Jiawei return vhost_vdpa_net_cvq_add(s, data_size + sizeof(ctrl), 654f73c0c43SEugenio Pérez sizeof(virtio_net_ctrl_ack)); 655f73c0c43SEugenio Pérez } 656f73c0c43SEugenio Pérez 657f73c0c43SEugenio Pérez static int vhost_vdpa_net_load_mac(VhostVDPAState *s, const VirtIONet *n) 658f73c0c43SEugenio Pérez { 65902d3bf09SHawkins Jiawei if (virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_MAC_ADDR)) { 6602848c6aaSHawkins Jiawei const struct iovec data = { 6612848c6aaSHawkins Jiawei .iov_base = (void *)n->mac, 6622848c6aaSHawkins Jiawei .iov_len = sizeof(n->mac), 6632848c6aaSHawkins Jiawei }; 664f73c0c43SEugenio Pérez ssize_t dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_MAC, 665f73c0c43SEugenio Pérez VIRTIO_NET_CTRL_MAC_ADDR_SET, 6662848c6aaSHawkins Jiawei &data, 1); 667f73c0c43SEugenio Pérez if (unlikely(dev_written < 0)) { 668f73c0c43SEugenio Pérez return dev_written; 669f73c0c43SEugenio Pérez } 670b479bc3cSHawkins Jiawei if (*s->status != VIRTIO_NET_OK) { 671b479bc3cSHawkins Jiawei return -EIO; 672b479bc3cSHawkins Jiawei } 673f73c0c43SEugenio Pérez } 674f73c0c43SEugenio Pérez 6750ddcecb8SHawkins Jiawei /* 6760ddcecb8SHawkins Jiawei * According to VirtIO standard, "The device MUST have an 6770ddcecb8SHawkins Jiawei * empty MAC filtering table on reset.". 6780ddcecb8SHawkins Jiawei * 6790ddcecb8SHawkins Jiawei * Therefore, there is no need to send this CVQ command if the 6800ddcecb8SHawkins Jiawei * driver also sets an empty MAC filter table, which aligns with 6810ddcecb8SHawkins Jiawei * the device's defaults. 6820ddcecb8SHawkins Jiawei * 6830ddcecb8SHawkins Jiawei * Note that the device's defaults can mismatch the driver's 6840ddcecb8SHawkins Jiawei * configuration only at live migration. 6850ddcecb8SHawkins Jiawei */ 6860ddcecb8SHawkins Jiawei if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_RX) || 6870ddcecb8SHawkins Jiawei n->mac_table.in_use == 0) { 6880ddcecb8SHawkins Jiawei return 0; 6890ddcecb8SHawkins Jiawei } 6900ddcecb8SHawkins Jiawei 6910ddcecb8SHawkins Jiawei uint32_t uni_entries = n->mac_table.first_multi, 6920ddcecb8SHawkins Jiawei uni_macs_size = uni_entries * ETH_ALEN, 6930ddcecb8SHawkins Jiawei mul_entries = n->mac_table.in_use - uni_entries, 6940ddcecb8SHawkins Jiawei mul_macs_size = mul_entries * ETH_ALEN; 6950ddcecb8SHawkins Jiawei struct virtio_net_ctrl_mac uni = { 6960ddcecb8SHawkins Jiawei .entries = cpu_to_le32(uni_entries), 6970ddcecb8SHawkins Jiawei }; 6980ddcecb8SHawkins Jiawei struct virtio_net_ctrl_mac mul = { 6990ddcecb8SHawkins Jiawei .entries = cpu_to_le32(mul_entries), 7000ddcecb8SHawkins Jiawei }; 7010ddcecb8SHawkins Jiawei const struct iovec data[] = { 7020ddcecb8SHawkins Jiawei { 7030ddcecb8SHawkins Jiawei .iov_base = &uni, 7040ddcecb8SHawkins Jiawei .iov_len = sizeof(uni), 7050ddcecb8SHawkins Jiawei }, { 7060ddcecb8SHawkins Jiawei .iov_base = n->mac_table.macs, 7070ddcecb8SHawkins Jiawei .iov_len = uni_macs_size, 7080ddcecb8SHawkins Jiawei }, { 7090ddcecb8SHawkins Jiawei .iov_base = &mul, 7100ddcecb8SHawkins Jiawei .iov_len = sizeof(mul), 7110ddcecb8SHawkins Jiawei }, { 7120ddcecb8SHawkins Jiawei .iov_base = &n->mac_table.macs[uni_macs_size], 7130ddcecb8SHawkins Jiawei .iov_len = mul_macs_size, 7140ddcecb8SHawkins Jiawei }, 7150ddcecb8SHawkins Jiawei }; 7160ddcecb8SHawkins Jiawei ssize_t dev_written = vhost_vdpa_net_load_cmd(s, 7170ddcecb8SHawkins Jiawei VIRTIO_NET_CTRL_MAC, 7180ddcecb8SHawkins Jiawei VIRTIO_NET_CTRL_MAC_TABLE_SET, 7190ddcecb8SHawkins Jiawei data, ARRAY_SIZE(data)); 7200ddcecb8SHawkins Jiawei if (unlikely(dev_written < 0)) { 7210ddcecb8SHawkins Jiawei return dev_written; 7220ddcecb8SHawkins Jiawei } 7230ddcecb8SHawkins Jiawei if (*s->status != VIRTIO_NET_OK) { 7240ddcecb8SHawkins Jiawei return -EIO; 7250ddcecb8SHawkins Jiawei } 7260ddcecb8SHawkins Jiawei 727f73c0c43SEugenio Pérez return 0; 728f73c0c43SEugenio Pérez } 729f73c0c43SEugenio Pérez 730f64c7cdaSEugenio Pérez static int vhost_vdpa_net_load_mq(VhostVDPAState *s, 731f64c7cdaSEugenio Pérez const VirtIONet *n) 732f64c7cdaSEugenio Pérez { 733f64c7cdaSEugenio Pérez struct virtio_net_ctrl_mq mq; 734f64c7cdaSEugenio Pérez ssize_t dev_written; 735f64c7cdaSEugenio Pérez 73602d3bf09SHawkins Jiawei if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_MQ)) { 737f64c7cdaSEugenio Pérez return 0; 738f64c7cdaSEugenio Pérez } 739f64c7cdaSEugenio Pérez 740f64c7cdaSEugenio Pérez mq.virtqueue_pairs = cpu_to_le16(n->curr_queue_pairs); 7412848c6aaSHawkins Jiawei const struct iovec data = { 7422848c6aaSHawkins Jiawei .iov_base = &mq, 7432848c6aaSHawkins Jiawei .iov_len = sizeof(mq), 7442848c6aaSHawkins Jiawei }; 745f64c7cdaSEugenio Pérez dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_MQ, 7462848c6aaSHawkins Jiawei VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, 7472848c6aaSHawkins Jiawei &data, 1); 748f64c7cdaSEugenio Pérez if (unlikely(dev_written < 0)) { 749f64c7cdaSEugenio Pérez return dev_written; 750f64c7cdaSEugenio Pérez } 751f45fd95eSHawkins Jiawei if (*s->status != VIRTIO_NET_OK) { 752f45fd95eSHawkins Jiawei return -EIO; 753f45fd95eSHawkins Jiawei } 754f64c7cdaSEugenio Pérez 755f45fd95eSHawkins Jiawei return 0; 756f64c7cdaSEugenio Pérez } 757f64c7cdaSEugenio Pérez 7580b58d368SHawkins Jiawei static int vhost_vdpa_net_load_offloads(VhostVDPAState *s, 7590b58d368SHawkins Jiawei const VirtIONet *n) 7600b58d368SHawkins Jiawei { 7610b58d368SHawkins Jiawei uint64_t offloads; 7620b58d368SHawkins Jiawei ssize_t dev_written; 7630b58d368SHawkins Jiawei 7640b58d368SHawkins Jiawei if (!virtio_vdev_has_feature(&n->parent_obj, 7650b58d368SHawkins Jiawei VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)) { 7660b58d368SHawkins Jiawei return 0; 7670b58d368SHawkins Jiawei } 7680b58d368SHawkins Jiawei 7690b58d368SHawkins Jiawei if (n->curr_guest_offloads == virtio_net_supported_guest_offloads(n)) { 7700b58d368SHawkins Jiawei /* 7710b58d368SHawkins Jiawei * According to VirtIO standard, "Upon feature negotiation 7720b58d368SHawkins Jiawei * corresponding offload gets enabled to preserve 7730b58d368SHawkins Jiawei * backward compatibility.". 7740b58d368SHawkins Jiawei * 7750b58d368SHawkins Jiawei * Therefore, there is no need to send this CVQ command if the 7760b58d368SHawkins Jiawei * driver also enables all supported offloads, which aligns with 7770b58d368SHawkins Jiawei * the device's defaults. 7780b58d368SHawkins Jiawei * 7790b58d368SHawkins Jiawei * Note that the device's defaults can mismatch the driver's 7800b58d368SHawkins Jiawei * configuration only at live migration. 7810b58d368SHawkins Jiawei */ 7820b58d368SHawkins Jiawei return 0; 7830b58d368SHawkins Jiawei } 7840b58d368SHawkins Jiawei 7850b58d368SHawkins Jiawei offloads = cpu_to_le64(n->curr_guest_offloads); 7862848c6aaSHawkins Jiawei const struct iovec data = { 7872848c6aaSHawkins Jiawei .iov_base = &offloads, 7882848c6aaSHawkins Jiawei .iov_len = sizeof(offloads), 7892848c6aaSHawkins Jiawei }; 7900b58d368SHawkins Jiawei dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_GUEST_OFFLOADS, 7910b58d368SHawkins Jiawei VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, 7922848c6aaSHawkins Jiawei &data, 1); 7930b58d368SHawkins Jiawei if (unlikely(dev_written < 0)) { 7940b58d368SHawkins Jiawei return dev_written; 7950b58d368SHawkins Jiawei } 7966f348071SHawkins Jiawei if (*s->status != VIRTIO_NET_OK) { 7976f348071SHawkins Jiawei return -EIO; 7986f348071SHawkins Jiawei } 7990b58d368SHawkins Jiawei 8006f348071SHawkins Jiawei return 0; 8010b58d368SHawkins Jiawei } 8020b58d368SHawkins Jiawei 803b12f907eSHawkins Jiawei static int vhost_vdpa_net_load_rx_mode(VhostVDPAState *s, 804b12f907eSHawkins Jiawei uint8_t cmd, 805b12f907eSHawkins Jiawei uint8_t on) 806b12f907eSHawkins Jiawei { 807b12f907eSHawkins Jiawei const struct iovec data = { 808b12f907eSHawkins Jiawei .iov_base = &on, 809b12f907eSHawkins Jiawei .iov_len = sizeof(on), 810b12f907eSHawkins Jiawei }; 811b12f907eSHawkins Jiawei return vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_RX, 812b12f907eSHawkins Jiawei cmd, &data, 1); 813b12f907eSHawkins Jiawei } 814b12f907eSHawkins Jiawei 815b12f907eSHawkins Jiawei static int vhost_vdpa_net_load_rx(VhostVDPAState *s, 816b12f907eSHawkins Jiawei const VirtIONet *n) 817b12f907eSHawkins Jiawei { 818b12f907eSHawkins Jiawei ssize_t dev_written; 819b12f907eSHawkins Jiawei 820b12f907eSHawkins Jiawei if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_RX)) { 821b12f907eSHawkins Jiawei return 0; 822b12f907eSHawkins Jiawei } 823b12f907eSHawkins Jiawei 824b12f907eSHawkins Jiawei /* 825b12f907eSHawkins Jiawei * According to virtio_net_reset(), device turns promiscuous mode 826b12f907eSHawkins Jiawei * on by default. 827b12f907eSHawkins Jiawei * 8280a19d879SMichael Tokarev * Additionally, according to VirtIO standard, "Since there are 829b12f907eSHawkins Jiawei * no guarantees, it can use a hash filter or silently switch to 830b12f907eSHawkins Jiawei * allmulti or promiscuous mode if it is given too many addresses.". 831b12f907eSHawkins Jiawei * QEMU marks `n->mac_table.uni_overflow` if guest sets too many 832b12f907eSHawkins Jiawei * non-multicast MAC addresses, indicating that promiscuous mode 833b12f907eSHawkins Jiawei * should be enabled. 834b12f907eSHawkins Jiawei * 835b12f907eSHawkins Jiawei * Therefore, QEMU should only send this CVQ command if the 836b12f907eSHawkins Jiawei * `n->mac_table.uni_overflow` is not marked and `n->promisc` is off, 837b12f907eSHawkins Jiawei * which sets promiscuous mode on, different from the device's defaults. 838b12f907eSHawkins Jiawei * 839b12f907eSHawkins Jiawei * Note that the device's defaults can mismatch the driver's 840b12f907eSHawkins Jiawei * configuration only at live migration. 841b12f907eSHawkins Jiawei */ 842b12f907eSHawkins Jiawei if (!n->mac_table.uni_overflow && !n->promisc) { 843b12f907eSHawkins Jiawei dev_written = vhost_vdpa_net_load_rx_mode(s, 844b12f907eSHawkins Jiawei VIRTIO_NET_CTRL_RX_PROMISC, 0); 845b12f907eSHawkins Jiawei if (unlikely(dev_written < 0)) { 846b12f907eSHawkins Jiawei return dev_written; 847b12f907eSHawkins Jiawei } 848b12f907eSHawkins Jiawei if (*s->status != VIRTIO_NET_OK) { 849b12f907eSHawkins Jiawei return -EIO; 850b12f907eSHawkins Jiawei } 851b12f907eSHawkins Jiawei } 852b12f907eSHawkins Jiawei 853b12f907eSHawkins Jiawei /* 854b12f907eSHawkins Jiawei * According to virtio_net_reset(), device turns all-multicast mode 855b12f907eSHawkins Jiawei * off by default. 856b12f907eSHawkins Jiawei * 857b12f907eSHawkins Jiawei * According to VirtIO standard, "Since there are no guarantees, 858b12f907eSHawkins Jiawei * it can use a hash filter or silently switch to allmulti or 859b12f907eSHawkins Jiawei * promiscuous mode if it is given too many addresses.". QEMU marks 860b12f907eSHawkins Jiawei * `n->mac_table.multi_overflow` if guest sets too many 861b12f907eSHawkins Jiawei * non-multicast MAC addresses. 862b12f907eSHawkins Jiawei * 863b12f907eSHawkins Jiawei * Therefore, QEMU should only send this CVQ command if the 864b12f907eSHawkins Jiawei * `n->mac_table.multi_overflow` is marked or `n->allmulti` is on, 865b12f907eSHawkins Jiawei * which sets all-multicast mode on, different from the device's defaults. 866b12f907eSHawkins Jiawei * 867b12f907eSHawkins Jiawei * Note that the device's defaults can mismatch the driver's 868b12f907eSHawkins Jiawei * configuration only at live migration. 869b12f907eSHawkins Jiawei */ 870b12f907eSHawkins Jiawei if (n->mac_table.multi_overflow || n->allmulti) { 871b12f907eSHawkins Jiawei dev_written = vhost_vdpa_net_load_rx_mode(s, 872b12f907eSHawkins Jiawei VIRTIO_NET_CTRL_RX_ALLMULTI, 1); 873b12f907eSHawkins Jiawei if (unlikely(dev_written < 0)) { 874b12f907eSHawkins Jiawei return dev_written; 875b12f907eSHawkins Jiawei } 876b12f907eSHawkins Jiawei if (*s->status != VIRTIO_NET_OK) { 877b12f907eSHawkins Jiawei return -EIO; 878b12f907eSHawkins Jiawei } 879b12f907eSHawkins Jiawei } 880b12f907eSHawkins Jiawei 8814fd180c7SHawkins Jiawei if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_RX_EXTRA)) { 8824fd180c7SHawkins Jiawei return 0; 8834fd180c7SHawkins Jiawei } 8844fd180c7SHawkins Jiawei 8854fd180c7SHawkins Jiawei /* 8864fd180c7SHawkins Jiawei * According to virtio_net_reset(), device turns all-unicast mode 8874fd180c7SHawkins Jiawei * off by default. 8884fd180c7SHawkins Jiawei * 8894fd180c7SHawkins Jiawei * Therefore, QEMU should only send this CVQ command if the driver 8904fd180c7SHawkins Jiawei * sets all-unicast mode on, different from the device's defaults. 8914fd180c7SHawkins Jiawei * 8924fd180c7SHawkins Jiawei * Note that the device's defaults can mismatch the driver's 8934fd180c7SHawkins Jiawei * configuration only at live migration. 8944fd180c7SHawkins Jiawei */ 8954fd180c7SHawkins Jiawei if (n->alluni) { 8964fd180c7SHawkins Jiawei dev_written = vhost_vdpa_net_load_rx_mode(s, 8974fd180c7SHawkins Jiawei VIRTIO_NET_CTRL_RX_ALLUNI, 1); 8984fd180c7SHawkins Jiawei if (dev_written < 0) { 8994fd180c7SHawkins Jiawei return dev_written; 9004fd180c7SHawkins Jiawei } 9014fd180c7SHawkins Jiawei if (*s->status != VIRTIO_NET_OK) { 9024fd180c7SHawkins Jiawei return -EIO; 9034fd180c7SHawkins Jiawei } 9044fd180c7SHawkins Jiawei } 9054fd180c7SHawkins Jiawei 9064fd180c7SHawkins Jiawei /* 9074fd180c7SHawkins Jiawei * According to virtio_net_reset(), device turns non-multicast mode 9084fd180c7SHawkins Jiawei * off by default. 9094fd180c7SHawkins Jiawei * 9104fd180c7SHawkins Jiawei * Therefore, QEMU should only send this CVQ command if the driver 9114fd180c7SHawkins Jiawei * sets non-multicast mode on, different from the device's defaults. 9124fd180c7SHawkins Jiawei * 9134fd180c7SHawkins Jiawei * Note that the device's defaults can mismatch the driver's 9144fd180c7SHawkins Jiawei * configuration only at live migration. 9154fd180c7SHawkins Jiawei */ 9164fd180c7SHawkins Jiawei if (n->nomulti) { 9174fd180c7SHawkins Jiawei dev_written = vhost_vdpa_net_load_rx_mode(s, 9184fd180c7SHawkins Jiawei VIRTIO_NET_CTRL_RX_NOMULTI, 1); 9194fd180c7SHawkins Jiawei if (dev_written < 0) { 9204fd180c7SHawkins Jiawei return dev_written; 9214fd180c7SHawkins Jiawei } 9224fd180c7SHawkins Jiawei if (*s->status != VIRTIO_NET_OK) { 9234fd180c7SHawkins Jiawei return -EIO; 9244fd180c7SHawkins Jiawei } 9254fd180c7SHawkins Jiawei } 9264fd180c7SHawkins Jiawei 9274fd180c7SHawkins Jiawei /* 9284fd180c7SHawkins Jiawei * According to virtio_net_reset(), device turns non-unicast mode 9294fd180c7SHawkins Jiawei * off by default. 9304fd180c7SHawkins Jiawei * 9314fd180c7SHawkins Jiawei * Therefore, QEMU should only send this CVQ command if the driver 9324fd180c7SHawkins Jiawei * sets non-unicast mode on, different from the device's defaults. 9334fd180c7SHawkins Jiawei * 9344fd180c7SHawkins Jiawei * Note that the device's defaults can mismatch the driver's 9354fd180c7SHawkins Jiawei * configuration only at live migration. 9364fd180c7SHawkins Jiawei */ 9374fd180c7SHawkins Jiawei if (n->nouni) { 9384fd180c7SHawkins Jiawei dev_written = vhost_vdpa_net_load_rx_mode(s, 9394fd180c7SHawkins Jiawei VIRTIO_NET_CTRL_RX_NOUNI, 1); 9404fd180c7SHawkins Jiawei if (dev_written < 0) { 9414fd180c7SHawkins Jiawei return dev_written; 9424fd180c7SHawkins Jiawei } 9434fd180c7SHawkins Jiawei if (*s->status != VIRTIO_NET_OK) { 9444fd180c7SHawkins Jiawei return -EIO; 9454fd180c7SHawkins Jiawei } 9464fd180c7SHawkins Jiawei } 9474fd180c7SHawkins Jiawei 9484fd180c7SHawkins Jiawei /* 9494fd180c7SHawkins Jiawei * According to virtio_net_reset(), device turns non-broadcast mode 9504fd180c7SHawkins Jiawei * off by default. 9514fd180c7SHawkins Jiawei * 9524fd180c7SHawkins Jiawei * Therefore, QEMU should only send this CVQ command if the driver 9534fd180c7SHawkins Jiawei * sets non-broadcast mode on, different from the device's defaults. 9544fd180c7SHawkins Jiawei * 9554fd180c7SHawkins Jiawei * Note that the device's defaults can mismatch the driver's 9564fd180c7SHawkins Jiawei * configuration only at live migration. 9574fd180c7SHawkins Jiawei */ 9584fd180c7SHawkins Jiawei if (n->nobcast) { 9594fd180c7SHawkins Jiawei dev_written = vhost_vdpa_net_load_rx_mode(s, 9604fd180c7SHawkins Jiawei VIRTIO_NET_CTRL_RX_NOBCAST, 1); 9614fd180c7SHawkins Jiawei if (dev_written < 0) { 9624fd180c7SHawkins Jiawei return dev_written; 9634fd180c7SHawkins Jiawei } 9644fd180c7SHawkins Jiawei if (*s->status != VIRTIO_NET_OK) { 9654fd180c7SHawkins Jiawei return -EIO; 9664fd180c7SHawkins Jiawei } 9674fd180c7SHawkins Jiawei } 9684fd180c7SHawkins Jiawei 969b12f907eSHawkins Jiawei return 0; 970b12f907eSHawkins Jiawei } 971b12f907eSHawkins Jiawei 9728f7e9967SHawkins Jiawei static int vhost_vdpa_net_load_single_vlan(VhostVDPAState *s, 9738f7e9967SHawkins Jiawei const VirtIONet *n, 9748f7e9967SHawkins Jiawei uint16_t vid) 9758f7e9967SHawkins Jiawei { 9768f7e9967SHawkins Jiawei const struct iovec data = { 9778f7e9967SHawkins Jiawei .iov_base = &vid, 9788f7e9967SHawkins Jiawei .iov_len = sizeof(vid), 9798f7e9967SHawkins Jiawei }; 9808f7e9967SHawkins Jiawei ssize_t dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_VLAN, 9818f7e9967SHawkins Jiawei VIRTIO_NET_CTRL_VLAN_ADD, 9828f7e9967SHawkins Jiawei &data, 1); 9838f7e9967SHawkins Jiawei if (unlikely(dev_written < 0)) { 9848f7e9967SHawkins Jiawei return dev_written; 9858f7e9967SHawkins Jiawei } 9868f7e9967SHawkins Jiawei if (unlikely(*s->status != VIRTIO_NET_OK)) { 9878f7e9967SHawkins Jiawei return -EIO; 9888f7e9967SHawkins Jiawei } 9898f7e9967SHawkins Jiawei 9908f7e9967SHawkins Jiawei return 0; 9918f7e9967SHawkins Jiawei } 9928f7e9967SHawkins Jiawei 9938f7e9967SHawkins Jiawei static int vhost_vdpa_net_load_vlan(VhostVDPAState *s, 9948f7e9967SHawkins Jiawei const VirtIONet *n) 9958f7e9967SHawkins Jiawei { 9968f7e9967SHawkins Jiawei int r; 9978f7e9967SHawkins Jiawei 9988f7e9967SHawkins Jiawei if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_VLAN)) { 9998f7e9967SHawkins Jiawei return 0; 10008f7e9967SHawkins Jiawei } 10018f7e9967SHawkins Jiawei 10028f7e9967SHawkins Jiawei for (int i = 0; i < MAX_VLAN >> 5; i++) { 10038f7e9967SHawkins Jiawei for (int j = 0; n->vlans[i] && j <= 0x1f; j++) { 10048f7e9967SHawkins Jiawei if (n->vlans[i] & (1U << j)) { 10058f7e9967SHawkins Jiawei r = vhost_vdpa_net_load_single_vlan(s, n, (i << 5) + j); 10068f7e9967SHawkins Jiawei if (unlikely(r != 0)) { 10078f7e9967SHawkins Jiawei return r; 10088f7e9967SHawkins Jiawei } 10098f7e9967SHawkins Jiawei } 10108f7e9967SHawkins Jiawei } 10118f7e9967SHawkins Jiawei } 10128f7e9967SHawkins Jiawei 10138f7e9967SHawkins Jiawei return 0; 10148f7e9967SHawkins Jiawei } 10158f7e9967SHawkins Jiawei 1016*f3fada59SEugenio Pérez static int vhost_vdpa_net_cvq_load(NetClientState *nc) 1017dd036d8dSEugenio Pérez { 1018dd036d8dSEugenio Pérez VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); 1019f73c0c43SEugenio Pérez struct vhost_vdpa *v = &s->vhost_vdpa; 1020dd036d8dSEugenio Pérez const VirtIONet *n; 1021f73c0c43SEugenio Pérez int r; 1022dd036d8dSEugenio Pérez 1023dd036d8dSEugenio Pérez assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); 1024dd036d8dSEugenio Pérez 1025dd036d8dSEugenio Pérez if (!v->shadow_vqs_enabled) { 1026dd036d8dSEugenio Pérez return 0; 1027dd036d8dSEugenio Pérez } 1028dd036d8dSEugenio Pérez 1029dd036d8dSEugenio Pérez n = VIRTIO_NET(v->dev->vdev); 1030f73c0c43SEugenio Pérez r = vhost_vdpa_net_load_mac(s, n); 1031f73c0c43SEugenio Pérez if (unlikely(r < 0)) { 1032f73c0c43SEugenio Pérez return r; 1033dd036d8dSEugenio Pérez } 1034f64c7cdaSEugenio Pérez r = vhost_vdpa_net_load_mq(s, n); 1035f64c7cdaSEugenio Pérez if (unlikely(r)) { 1036f64c7cdaSEugenio Pérez return r; 1037f64c7cdaSEugenio Pérez } 10380b58d368SHawkins Jiawei r = vhost_vdpa_net_load_offloads(s, n); 10390b58d368SHawkins Jiawei if (unlikely(r)) { 10400b58d368SHawkins Jiawei return r; 10410b58d368SHawkins Jiawei } 1042b12f907eSHawkins Jiawei r = vhost_vdpa_net_load_rx(s, n); 1043b12f907eSHawkins Jiawei if (unlikely(r)) { 1044b12f907eSHawkins Jiawei return r; 1045b12f907eSHawkins Jiawei } 10468f7e9967SHawkins Jiawei r = vhost_vdpa_net_load_vlan(s, n); 10478f7e9967SHawkins Jiawei if (unlikely(r)) { 10488f7e9967SHawkins Jiawei return r; 10498f7e9967SHawkins Jiawei } 1050dd036d8dSEugenio Pérez 1051dd036d8dSEugenio Pérez return 0; 1052dd036d8dSEugenio Pérez } 1053dd036d8dSEugenio Pérez 1054f8972b56SEugenio Pérez static NetClientInfo net_vhost_vdpa_cvq_info = { 1055f8972b56SEugenio Pérez .type = NET_CLIENT_DRIVER_VHOST_VDPA, 1056f8972b56SEugenio Pérez .size = sizeof(VhostVDPAState), 1057f8972b56SEugenio Pérez .receive = vhost_vdpa_receive, 10587a7f87e9SEugenio Pérez .start = vhost_vdpa_net_cvq_start, 1059*f3fada59SEugenio Pérez .load = vhost_vdpa_net_cvq_load, 10607a7f87e9SEugenio Pérez .stop = vhost_vdpa_net_cvq_stop, 1061f8972b56SEugenio Pérez .cleanup = vhost_vdpa_cleanup, 1062f8972b56SEugenio Pérez .has_vnet_hdr = vhost_vdpa_has_vnet_hdr, 1063f8972b56SEugenio Pérez .has_ufo = vhost_vdpa_has_ufo, 1064f8972b56SEugenio Pérez .check_peer_type = vhost_vdpa_check_peer_type, 1065f8972b56SEugenio Pérez }; 1066f8972b56SEugenio Pérez 1067fee364e4SHawkins Jiawei /* 1068fee364e4SHawkins Jiawei * Forward the excessive VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command to 1069fee364e4SHawkins Jiawei * vdpa device. 1070fee364e4SHawkins Jiawei * 1071fee364e4SHawkins Jiawei * Considering that QEMU cannot send the entire filter table to the 1072fee364e4SHawkins Jiawei * vdpa device, it should send the VIRTIO_NET_CTRL_RX_PROMISC CVQ 1073fee364e4SHawkins Jiawei * command to enable promiscuous mode to receive all packets, 1074fee364e4SHawkins Jiawei * according to VirtIO standard, "Since there are no guarantees, 1075fee364e4SHawkins Jiawei * it can use a hash filter or silently switch to allmulti or 1076fee364e4SHawkins Jiawei * promiscuous mode if it is given too many addresses.". 1077fee364e4SHawkins Jiawei * 1078fee364e4SHawkins Jiawei * Since QEMU ignores MAC addresses beyond `MAC_TABLE_ENTRIES` and 1079fee364e4SHawkins Jiawei * marks `n->mac_table.x_overflow` accordingly, it should have 1080fee364e4SHawkins Jiawei * the same effect on the device model to receive 1081fee364e4SHawkins Jiawei * (`MAC_TABLE_ENTRIES` + 1) or more non-multicast MAC addresses. 1082fee364e4SHawkins Jiawei * The same applies to multicast MAC addresses. 1083fee364e4SHawkins Jiawei * 1084fee364e4SHawkins Jiawei * Therefore, QEMU can provide the device model with a fake 1085fee364e4SHawkins Jiawei * VIRTIO_NET_CTRL_MAC_TABLE_SET command with (`MAC_TABLE_ENTRIES` + 1) 1086fee364e4SHawkins Jiawei * non-multicast MAC addresses and (`MAC_TABLE_ENTRIES` + 1) multicast 1087fee364e4SHawkins Jiawei * MAC addresses. This ensures that the device model marks 1088fee364e4SHawkins Jiawei * `n->mac_table.uni_overflow` and `n->mac_table.multi_overflow`, 1089fee364e4SHawkins Jiawei * allowing all packets to be received, which aligns with the 1090fee364e4SHawkins Jiawei * state of the vdpa device. 1091fee364e4SHawkins Jiawei */ 1092fee364e4SHawkins Jiawei static int vhost_vdpa_net_excessive_mac_filter_cvq_add(VhostVDPAState *s, 1093fee364e4SHawkins Jiawei VirtQueueElement *elem, 1094fee364e4SHawkins Jiawei struct iovec *out) 1095fee364e4SHawkins Jiawei { 1096fee364e4SHawkins Jiawei struct virtio_net_ctrl_mac mac_data, *mac_ptr; 1097fee364e4SHawkins Jiawei struct virtio_net_ctrl_hdr *hdr_ptr; 1098fee364e4SHawkins Jiawei uint32_t cursor; 1099fee364e4SHawkins Jiawei ssize_t r; 1100fee364e4SHawkins Jiawei 1101fee364e4SHawkins Jiawei /* parse the non-multicast MAC address entries from CVQ command */ 1102fee364e4SHawkins Jiawei cursor = sizeof(*hdr_ptr); 1103fee364e4SHawkins Jiawei r = iov_to_buf(elem->out_sg, elem->out_num, cursor, 1104fee364e4SHawkins Jiawei &mac_data, sizeof(mac_data)); 1105fee364e4SHawkins Jiawei if (unlikely(r != sizeof(mac_data))) { 1106fee364e4SHawkins Jiawei /* 1107fee364e4SHawkins Jiawei * If the CVQ command is invalid, we should simulate the vdpa device 1108fee364e4SHawkins Jiawei * to reject the VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command 1109fee364e4SHawkins Jiawei */ 1110fee364e4SHawkins Jiawei *s->status = VIRTIO_NET_ERR; 1111fee364e4SHawkins Jiawei return sizeof(*s->status); 1112fee364e4SHawkins Jiawei } 1113fee364e4SHawkins Jiawei cursor += sizeof(mac_data) + le32_to_cpu(mac_data.entries) * ETH_ALEN; 1114fee364e4SHawkins Jiawei 1115fee364e4SHawkins Jiawei /* parse the multicast MAC address entries from CVQ command */ 1116fee364e4SHawkins Jiawei r = iov_to_buf(elem->out_sg, elem->out_num, cursor, 1117fee364e4SHawkins Jiawei &mac_data, sizeof(mac_data)); 1118fee364e4SHawkins Jiawei if (r != sizeof(mac_data)) { 1119fee364e4SHawkins Jiawei /* 1120fee364e4SHawkins Jiawei * If the CVQ command is invalid, we should simulate the vdpa device 1121fee364e4SHawkins Jiawei * to reject the VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command 1122fee364e4SHawkins Jiawei */ 1123fee364e4SHawkins Jiawei *s->status = VIRTIO_NET_ERR; 1124fee364e4SHawkins Jiawei return sizeof(*s->status); 1125fee364e4SHawkins Jiawei } 1126fee364e4SHawkins Jiawei cursor += sizeof(mac_data) + le32_to_cpu(mac_data.entries) * ETH_ALEN; 1127fee364e4SHawkins Jiawei 1128fee364e4SHawkins Jiawei /* validate the CVQ command */ 1129fee364e4SHawkins Jiawei if (iov_size(elem->out_sg, elem->out_num) != cursor) { 1130fee364e4SHawkins Jiawei /* 1131fee364e4SHawkins Jiawei * If the CVQ command is invalid, we should simulate the vdpa device 1132fee364e4SHawkins Jiawei * to reject the VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command 1133fee364e4SHawkins Jiawei */ 1134fee364e4SHawkins Jiawei *s->status = VIRTIO_NET_ERR; 1135fee364e4SHawkins Jiawei return sizeof(*s->status); 1136fee364e4SHawkins Jiawei } 1137fee364e4SHawkins Jiawei 1138fee364e4SHawkins Jiawei /* 1139fee364e4SHawkins Jiawei * According to VirtIO standard, "Since there are no guarantees, 1140fee364e4SHawkins Jiawei * it can use a hash filter or silently switch to allmulti or 1141fee364e4SHawkins Jiawei * promiscuous mode if it is given too many addresses.". 1142fee364e4SHawkins Jiawei * 1143fee364e4SHawkins Jiawei * Therefore, considering that QEMU is unable to send the entire 1144fee364e4SHawkins Jiawei * filter table to the vdpa device, it should send the 1145fee364e4SHawkins Jiawei * VIRTIO_NET_CTRL_RX_PROMISC CVQ command to enable promiscuous mode 1146fee364e4SHawkins Jiawei */ 1147fee364e4SHawkins Jiawei r = vhost_vdpa_net_load_rx_mode(s, VIRTIO_NET_CTRL_RX_PROMISC, 1); 1148fee364e4SHawkins Jiawei if (unlikely(r < 0)) { 1149fee364e4SHawkins Jiawei return r; 1150fee364e4SHawkins Jiawei } 1151fee364e4SHawkins Jiawei if (*s->status != VIRTIO_NET_OK) { 1152fee364e4SHawkins Jiawei return sizeof(*s->status); 1153fee364e4SHawkins Jiawei } 1154fee364e4SHawkins Jiawei 1155fee364e4SHawkins Jiawei /* 1156fee364e4SHawkins Jiawei * QEMU should also send a fake VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ 1157fee364e4SHawkins Jiawei * command to the device model, including (`MAC_TABLE_ENTRIES` + 1) 1158fee364e4SHawkins Jiawei * non-multicast MAC addresses and (`MAC_TABLE_ENTRIES` + 1) 1159fee364e4SHawkins Jiawei * multicast MAC addresses. 1160fee364e4SHawkins Jiawei * 1161fee364e4SHawkins Jiawei * By doing so, the device model can mark `n->mac_table.uni_overflow` 1162fee364e4SHawkins Jiawei * and `n->mac_table.multi_overflow`, enabling all packets to be 1163fee364e4SHawkins Jiawei * received, which aligns with the state of the vdpa device. 1164fee364e4SHawkins Jiawei */ 1165fee364e4SHawkins Jiawei cursor = 0; 1166fee364e4SHawkins Jiawei uint32_t fake_uni_entries = MAC_TABLE_ENTRIES + 1, 1167fee364e4SHawkins Jiawei fake_mul_entries = MAC_TABLE_ENTRIES + 1, 1168fee364e4SHawkins Jiawei fake_cvq_size = sizeof(struct virtio_net_ctrl_hdr) + 1169fee364e4SHawkins Jiawei sizeof(mac_data) + fake_uni_entries * ETH_ALEN + 1170fee364e4SHawkins Jiawei sizeof(mac_data) + fake_mul_entries * ETH_ALEN; 1171fee364e4SHawkins Jiawei 1172fee364e4SHawkins Jiawei assert(fake_cvq_size < vhost_vdpa_net_cvq_cmd_page_len()); 1173fee364e4SHawkins Jiawei out->iov_len = fake_cvq_size; 1174fee364e4SHawkins Jiawei 1175fee364e4SHawkins Jiawei /* pack the header for fake CVQ command */ 1176fee364e4SHawkins Jiawei hdr_ptr = out->iov_base + cursor; 1177fee364e4SHawkins Jiawei hdr_ptr->class = VIRTIO_NET_CTRL_MAC; 1178fee364e4SHawkins Jiawei hdr_ptr->cmd = VIRTIO_NET_CTRL_MAC_TABLE_SET; 1179fee364e4SHawkins Jiawei cursor += sizeof(*hdr_ptr); 1180fee364e4SHawkins Jiawei 1181fee364e4SHawkins Jiawei /* 1182fee364e4SHawkins Jiawei * Pack the non-multicast MAC addresses part for fake CVQ command. 1183fee364e4SHawkins Jiawei * 1184fee364e4SHawkins Jiawei * According to virtio_net_handle_mac(), QEMU doesn't verify the MAC 11850a19d879SMichael Tokarev * addresses provided in CVQ command. Therefore, only the entries 1186fee364e4SHawkins Jiawei * field need to be prepared in the CVQ command. 1187fee364e4SHawkins Jiawei */ 1188fee364e4SHawkins Jiawei mac_ptr = out->iov_base + cursor; 1189fee364e4SHawkins Jiawei mac_ptr->entries = cpu_to_le32(fake_uni_entries); 1190fee364e4SHawkins Jiawei cursor += sizeof(*mac_ptr) + fake_uni_entries * ETH_ALEN; 1191fee364e4SHawkins Jiawei 1192fee364e4SHawkins Jiawei /* 1193fee364e4SHawkins Jiawei * Pack the multicast MAC addresses part for fake CVQ command. 1194fee364e4SHawkins Jiawei * 1195fee364e4SHawkins Jiawei * According to virtio_net_handle_mac(), QEMU doesn't verify the MAC 11960a19d879SMichael Tokarev * addresses provided in CVQ command. Therefore, only the entries 1197fee364e4SHawkins Jiawei * field need to be prepared in the CVQ command. 1198fee364e4SHawkins Jiawei */ 1199fee364e4SHawkins Jiawei mac_ptr = out->iov_base + cursor; 1200fee364e4SHawkins Jiawei mac_ptr->entries = cpu_to_le32(fake_mul_entries); 1201fee364e4SHawkins Jiawei 1202fee364e4SHawkins Jiawei /* 1203fee364e4SHawkins Jiawei * Simulating QEMU poll a vdpa device used buffer 1204fee364e4SHawkins Jiawei * for VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command 1205fee364e4SHawkins Jiawei */ 1206fee364e4SHawkins Jiawei return sizeof(*s->status); 1207fee364e4SHawkins Jiawei } 1208fee364e4SHawkins Jiawei 12092df4dd31SEugenio Pérez /** 12102df4dd31SEugenio Pérez * Validate and copy control virtqueue commands. 12112df4dd31SEugenio Pérez * 12122df4dd31SEugenio Pérez * Following QEMU guidelines, we offer a copy of the buffers to the device to 12132df4dd31SEugenio Pérez * prevent TOCTOU bugs. 1214bd907ae4SEugenio Pérez */ 1215bd907ae4SEugenio Pérez static int vhost_vdpa_net_handle_ctrl_avail(VhostShadowVirtqueue *svq, 1216bd907ae4SEugenio Pérez VirtQueueElement *elem, 1217bd907ae4SEugenio Pérez void *opaque) 1218bd907ae4SEugenio Pérez { 12192df4dd31SEugenio Pérez VhostVDPAState *s = opaque; 1220be4278b6SEugenio Pérez size_t in_len; 122145c41018SHawkins Jiawei const struct virtio_net_ctrl_hdr *ctrl; 1222bd907ae4SEugenio Pérez virtio_net_ctrl_ack status = VIRTIO_NET_ERR; 12237a7f87e9SEugenio Pérez /* Out buffer sent to both the vdpa device and the device model */ 12247a7f87e9SEugenio Pérez struct iovec out = { 12257a7f87e9SEugenio Pérez .iov_base = s->cvq_cmd_out_buffer, 12267a7f87e9SEugenio Pérez }; 12272df4dd31SEugenio Pérez /* in buffer used for device model */ 12282df4dd31SEugenio Pérez const struct iovec in = { 12292df4dd31SEugenio Pérez .iov_base = &status, 12302df4dd31SEugenio Pérez .iov_len = sizeof(status), 12312df4dd31SEugenio Pérez }; 1232be4278b6SEugenio Pérez ssize_t dev_written = -EINVAL; 1233bd907ae4SEugenio Pérez 12347a7f87e9SEugenio Pérez out.iov_len = iov_to_buf(elem->out_sg, elem->out_num, 0, 12357a7f87e9SEugenio Pérez s->cvq_cmd_out_buffer, 1236fee364e4SHawkins Jiawei vhost_vdpa_net_cvq_cmd_page_len()); 123745c41018SHawkins Jiawei 123845c41018SHawkins Jiawei ctrl = s->cvq_cmd_out_buffer; 123945c41018SHawkins Jiawei if (ctrl->class == VIRTIO_NET_CTRL_ANNOUNCE) { 12403f9a3eebSEugenio Pérez /* 12413f9a3eebSEugenio Pérez * Guest announce capability is emulated by qemu, so don't forward to 12423f9a3eebSEugenio Pérez * the device. 12433f9a3eebSEugenio Pérez */ 12443f9a3eebSEugenio Pérez dev_written = sizeof(status); 12453f9a3eebSEugenio Pérez *s->status = VIRTIO_NET_OK; 1246fee364e4SHawkins Jiawei } else if (unlikely(ctrl->class == VIRTIO_NET_CTRL_MAC && 1247fee364e4SHawkins Jiawei ctrl->cmd == VIRTIO_NET_CTRL_MAC_TABLE_SET && 1248fee364e4SHawkins Jiawei iov_size(elem->out_sg, elem->out_num) > out.iov_len)) { 1249fee364e4SHawkins Jiawei /* 1250fee364e4SHawkins Jiawei * Due to the size limitation of the out buffer sent to the vdpa device, 1251fee364e4SHawkins Jiawei * which is determined by vhost_vdpa_net_cvq_cmd_page_len(), excessive 1252fee364e4SHawkins Jiawei * MAC addresses set by the driver for the filter table can cause 1253fee364e4SHawkins Jiawei * truncation of the CVQ command in QEMU. As a result, the vdpa device 1254fee364e4SHawkins Jiawei * rejects the flawed CVQ command. 1255fee364e4SHawkins Jiawei * 1256fee364e4SHawkins Jiawei * Therefore, QEMU must handle this situation instead of sending 12570a19d879SMichael Tokarev * the CVQ command directly. 1258fee364e4SHawkins Jiawei */ 1259fee364e4SHawkins Jiawei dev_written = vhost_vdpa_net_excessive_mac_filter_cvq_add(s, elem, 1260fee364e4SHawkins Jiawei &out); 1261fee364e4SHawkins Jiawei if (unlikely(dev_written < 0)) { 1262fee364e4SHawkins Jiawei goto out; 1263fee364e4SHawkins Jiawei } 12643f9a3eebSEugenio Pérez } else { 1265be4278b6SEugenio Pérez dev_written = vhost_vdpa_net_cvq_add(s, out.iov_len, sizeof(status)); 1266be4278b6SEugenio Pérez if (unlikely(dev_written < 0)) { 1267bd907ae4SEugenio Pérez goto out; 1268bd907ae4SEugenio Pérez } 12693f9a3eebSEugenio Pérez } 1270bd907ae4SEugenio Pérez 1271bd907ae4SEugenio Pérez if (unlikely(dev_written < sizeof(status))) { 1272bd907ae4SEugenio Pérez error_report("Insufficient written data (%zu)", dev_written); 12732df4dd31SEugenio Pérez goto out; 12742df4dd31SEugenio Pérez } 12752df4dd31SEugenio Pérez 127617fb889fSEugenio Pérez if (*s->status != VIRTIO_NET_OK) { 1277d45243bcSEugenio Pérez goto out; 12782df4dd31SEugenio Pérez } 12792df4dd31SEugenio Pérez 12802df4dd31SEugenio Pérez status = VIRTIO_NET_ERR; 12817a7f87e9SEugenio Pérez virtio_net_handle_ctrl_iov(svq->vdev, &in, 1, &out, 1); 12822df4dd31SEugenio Pérez if (status != VIRTIO_NET_OK) { 12832df4dd31SEugenio Pérez error_report("Bad CVQ processing in model"); 1284bd907ae4SEugenio Pérez } 1285bd907ae4SEugenio Pérez 1286bd907ae4SEugenio Pérez out: 1287bd907ae4SEugenio Pérez in_len = iov_from_buf(elem->in_sg, elem->in_num, 0, &status, 1288bd907ae4SEugenio Pérez sizeof(status)); 1289bd907ae4SEugenio Pérez if (unlikely(in_len < sizeof(status))) { 1290bd907ae4SEugenio Pérez error_report("Bad device CVQ written length"); 1291bd907ae4SEugenio Pérez } 1292bd907ae4SEugenio Pérez vhost_svq_push_elem(svq, elem, MIN(in_len, sizeof(status))); 1293031b1abaSHawkins Jiawei /* 1294031b1abaSHawkins Jiawei * `elem` belongs to vhost_vdpa_net_handle_ctrl_avail() only when 1295031b1abaSHawkins Jiawei * the function successfully forwards the CVQ command, indicated 1296031b1abaSHawkins Jiawei * by a non-negative value of `dev_written`. Otherwise, it still 1297031b1abaSHawkins Jiawei * belongs to SVQ. 1298031b1abaSHawkins Jiawei * This function should only free the `elem` when it owns. 1299031b1abaSHawkins Jiawei */ 1300031b1abaSHawkins Jiawei if (dev_written >= 0) { 1301bd907ae4SEugenio Pérez g_free(elem); 1302031b1abaSHawkins Jiawei } 1303be4278b6SEugenio Pérez return dev_written < 0 ? dev_written : 0; 1304bd907ae4SEugenio Pérez } 1305bd907ae4SEugenio Pérez 1306bd907ae4SEugenio Pérez static const VhostShadowVirtqueueOps vhost_vdpa_net_svq_ops = { 1307bd907ae4SEugenio Pérez .avail_handler = vhost_vdpa_net_handle_ctrl_avail, 1308bd907ae4SEugenio Pérez }; 1309bd907ae4SEugenio Pérez 1310152128d6SEugenio Pérez /** 1311152128d6SEugenio Pérez * Probe if CVQ is isolated 1312152128d6SEugenio Pérez * 1313152128d6SEugenio Pérez * @device_fd The vdpa device fd 1314152128d6SEugenio Pérez * @features Features offered by the device. 1315152128d6SEugenio Pérez * @cvq_index The control vq pair index 1316152128d6SEugenio Pérez * 1317152128d6SEugenio Pérez * Returns <0 in case of failure, 0 if false and 1 if true. 1318152128d6SEugenio Pérez */ 1319152128d6SEugenio Pérez static int vhost_vdpa_probe_cvq_isolation(int device_fd, uint64_t features, 1320152128d6SEugenio Pérez int cvq_index, Error **errp) 1321152128d6SEugenio Pérez { 1322152128d6SEugenio Pérez uint64_t backend_features; 1323152128d6SEugenio Pérez int64_t cvq_group; 1324152128d6SEugenio Pérez uint8_t status = VIRTIO_CONFIG_S_ACKNOWLEDGE | 1325152128d6SEugenio Pérez VIRTIO_CONFIG_S_DRIVER | 1326152128d6SEugenio Pérez VIRTIO_CONFIG_S_FEATURES_OK; 1327152128d6SEugenio Pérez int r; 1328152128d6SEugenio Pérez 1329152128d6SEugenio Pérez ERRP_GUARD(); 1330152128d6SEugenio Pérez 1331152128d6SEugenio Pérez r = ioctl(device_fd, VHOST_GET_BACKEND_FEATURES, &backend_features); 1332152128d6SEugenio Pérez if (unlikely(r < 0)) { 1333152128d6SEugenio Pérez error_setg_errno(errp, errno, "Cannot get vdpa backend_features"); 1334152128d6SEugenio Pérez return r; 1335152128d6SEugenio Pérez } 1336152128d6SEugenio Pérez 1337152128d6SEugenio Pérez if (!(backend_features & BIT_ULL(VHOST_BACKEND_F_IOTLB_ASID))) { 1338152128d6SEugenio Pérez return 0; 1339152128d6SEugenio Pérez } 1340152128d6SEugenio Pérez 1341152128d6SEugenio Pérez r = ioctl(device_fd, VHOST_SET_FEATURES, &features); 1342152128d6SEugenio Pérez if (unlikely(r)) { 1343152128d6SEugenio Pérez error_setg_errno(errp, errno, "Cannot set features"); 1344152128d6SEugenio Pérez } 1345152128d6SEugenio Pérez 1346152128d6SEugenio Pérez r = ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status); 1347152128d6SEugenio Pérez if (unlikely(r)) { 1348152128d6SEugenio Pérez error_setg_errno(errp, -r, "Cannot set device features"); 1349152128d6SEugenio Pérez goto out; 1350152128d6SEugenio Pérez } 1351152128d6SEugenio Pérez 1352152128d6SEugenio Pérez cvq_group = vhost_vdpa_get_vring_group(device_fd, cvq_index, errp); 1353152128d6SEugenio Pérez if (unlikely(cvq_group < 0)) { 1354152128d6SEugenio Pérez if (cvq_group != -ENOTSUP) { 1355152128d6SEugenio Pérez r = cvq_group; 1356152128d6SEugenio Pérez goto out; 1357152128d6SEugenio Pérez } 1358152128d6SEugenio Pérez 1359152128d6SEugenio Pérez /* 1360152128d6SEugenio Pérez * The kernel report VHOST_BACKEND_F_IOTLB_ASID if the vdpa frontend 1361152128d6SEugenio Pérez * support ASID even if the parent driver does not. The CVQ cannot be 1362152128d6SEugenio Pérez * isolated in this case. 1363152128d6SEugenio Pérez */ 1364152128d6SEugenio Pérez error_free(*errp); 1365152128d6SEugenio Pérez *errp = NULL; 1366152128d6SEugenio Pérez r = 0; 1367152128d6SEugenio Pérez goto out; 1368152128d6SEugenio Pérez } 1369152128d6SEugenio Pérez 1370152128d6SEugenio Pérez for (int i = 0; i < cvq_index; ++i) { 1371152128d6SEugenio Pérez int64_t group = vhost_vdpa_get_vring_group(device_fd, i, errp); 1372152128d6SEugenio Pérez if (unlikely(group < 0)) { 1373152128d6SEugenio Pérez r = group; 1374152128d6SEugenio Pérez goto out; 1375152128d6SEugenio Pérez } 1376152128d6SEugenio Pérez 1377152128d6SEugenio Pérez if (group == (int64_t)cvq_group) { 1378152128d6SEugenio Pérez r = 0; 1379152128d6SEugenio Pérez goto out; 1380152128d6SEugenio Pérez } 1381152128d6SEugenio Pérez } 1382152128d6SEugenio Pérez 1383152128d6SEugenio Pérez r = 1; 1384152128d6SEugenio Pérez 1385152128d6SEugenio Pérez out: 1386152128d6SEugenio Pérez status = 0; 1387152128d6SEugenio Pérez ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status); 1388152128d6SEugenio Pérez return r; 1389152128d6SEugenio Pérez } 1390152128d6SEugenio Pérez 1391654790b6SJason Wang static NetClientState *net_vhost_vdpa_init(NetClientState *peer, 1392654790b6SJason Wang const char *device, 1393654790b6SJason Wang const char *name, 139440237840SJason Wang int vdpa_device_fd, 139540237840SJason Wang int queue_pair_index, 139640237840SJason Wang int nvqs, 13971576dbb5SEugenio Pérez bool is_datapath, 13981576dbb5SEugenio Pérez bool svq, 13995c1ebd4cSEugenio Pérez struct vhost_vdpa_iova_range iova_range, 1400152128d6SEugenio Pérez uint64_t features, 1401152128d6SEugenio Pérez Error **errp) 14021e0a84eaSCindy Lu { 14031e0a84eaSCindy Lu NetClientState *nc = NULL; 14041e0a84eaSCindy Lu VhostVDPAState *s; 14051e0a84eaSCindy Lu int ret = 0; 14061e0a84eaSCindy Lu assert(name); 1407152128d6SEugenio Pérez int cvq_isolated; 1408152128d6SEugenio Pérez 140940237840SJason Wang if (is_datapath) { 141040237840SJason Wang nc = qemu_new_net_client(&net_vhost_vdpa_info, peer, device, 141140237840SJason Wang name); 141240237840SJason Wang } else { 1413152128d6SEugenio Pérez cvq_isolated = vhost_vdpa_probe_cvq_isolation(vdpa_device_fd, features, 1414152128d6SEugenio Pérez queue_pair_index * 2, 1415152128d6SEugenio Pérez errp); 1416152128d6SEugenio Pérez if (unlikely(cvq_isolated < 0)) { 1417152128d6SEugenio Pérez return NULL; 1418152128d6SEugenio Pérez } 1419152128d6SEugenio Pérez 1420f8972b56SEugenio Pérez nc = qemu_new_net_control_client(&net_vhost_vdpa_cvq_info, peer, 142140237840SJason Wang device, name); 142240237840SJason Wang } 142353b85d95SLaurent Vivier qemu_set_info_str(nc, TYPE_VHOST_VDPA); 14241e0a84eaSCindy Lu s = DO_UPCAST(VhostVDPAState, nc, nc); 14257327813dSJason Wang 14261e0a84eaSCindy Lu s->vhost_vdpa.device_fd = vdpa_device_fd; 142740237840SJason Wang s->vhost_vdpa.index = queue_pair_index; 14287f211a28SEugenio Pérez s->always_svq = svq; 142969498430SEugenio Pérez s->migration_state.notify = vdpa_net_migration_state_notifier; 14301576dbb5SEugenio Pérez s->vhost_vdpa.shadow_vqs_enabled = svq; 1431a585fad2SEugenio Pérez s->vhost_vdpa.iova_range = iova_range; 14326188d78aSEugenio Pérez s->vhost_vdpa.shadow_data = svq; 14335c1ebd4cSEugenio Pérez if (queue_pair_index == 0) { 14345c1ebd4cSEugenio Pérez vhost_vdpa_net_valid_svq_features(features, 14355c1ebd4cSEugenio Pérez &s->vhost_vdpa.migration_blocker); 14365c1ebd4cSEugenio Pérez } else if (!is_datapath) { 1437babf8b87SEugenio Pérez s->cvq_cmd_out_buffer = mmap(NULL, vhost_vdpa_net_cvq_cmd_page_len(), 1438babf8b87SEugenio Pérez PROT_READ | PROT_WRITE, 1439babf8b87SEugenio Pérez MAP_SHARED | MAP_ANONYMOUS, -1, 0); 1440babf8b87SEugenio Pérez s->status = mmap(NULL, vhost_vdpa_net_cvq_cmd_page_len(), 1441babf8b87SEugenio Pérez PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANONYMOUS, 1442babf8b87SEugenio Pérez -1, 0); 14432df4dd31SEugenio Pérez 1444bd907ae4SEugenio Pérez s->vhost_vdpa.shadow_vq_ops = &vhost_vdpa_net_svq_ops; 1445bd907ae4SEugenio Pérez s->vhost_vdpa.shadow_vq_ops_opaque = s; 1446152128d6SEugenio Pérez s->cvq_isolated = cvq_isolated; 14479c363cf6SEugenio Pérez 14489c363cf6SEugenio Pérez /* 14498bc0049eSEugenio Pérez * TODO: We cannot migrate devices with CVQ and no x-svq enabled as 14508bc0049eSEugenio Pérez * there is no way to set the device state (MAC, MQ, etc) before 14518bc0049eSEugenio Pérez * starting the datapath. 14529c363cf6SEugenio Pérez * 14539c363cf6SEugenio Pérez * Migration blocker ownership now belongs to s->vhost_vdpa. 14549c363cf6SEugenio Pérez */ 14558bc0049eSEugenio Pérez if (!svq) { 14569c363cf6SEugenio Pérez error_setg(&s->vhost_vdpa.migration_blocker, 14579c363cf6SEugenio Pérez "net vdpa cannot migrate with CVQ feature"); 1458bd907ae4SEugenio Pérez } 14598bc0049eSEugenio Pérez } 146040237840SJason Wang ret = vhost_vdpa_add(nc, (void *)&s->vhost_vdpa, queue_pair_index, nvqs); 146174af5eecSJason Wang if (ret) { 146274af5eecSJason Wang qemu_del_net_client(nc); 1463654790b6SJason Wang return NULL; 146474af5eecSJason Wang } 1465654790b6SJason Wang return nc; 14661e0a84eaSCindy Lu } 14671e0a84eaSCindy Lu 14688170ab3fSEugenio Pérez static int vhost_vdpa_get_features(int fd, uint64_t *features, Error **errp) 14698170ab3fSEugenio Pérez { 14708170ab3fSEugenio Pérez int ret = ioctl(fd, VHOST_GET_FEATURES, features); 14718170ab3fSEugenio Pérez if (unlikely(ret < 0)) { 14728170ab3fSEugenio Pérez error_setg_errno(errp, errno, 14738170ab3fSEugenio Pérez "Fail to query features from vhost-vDPA device"); 14748170ab3fSEugenio Pérez } 14758170ab3fSEugenio Pérez return ret; 14768170ab3fSEugenio Pérez } 14778170ab3fSEugenio Pérez 14788170ab3fSEugenio Pérez static int vhost_vdpa_get_max_queue_pairs(int fd, uint64_t features, 14798170ab3fSEugenio Pérez int *has_cvq, Error **errp) 148040237840SJason Wang { 148140237840SJason Wang unsigned long config_size = offsetof(struct vhost_vdpa_config, buf); 1482cd523a41SStefano Garzarella g_autofree struct vhost_vdpa_config *config = NULL; 148340237840SJason Wang __virtio16 *max_queue_pairs; 148440237840SJason Wang int ret; 148540237840SJason Wang 148640237840SJason Wang if (features & (1 << VIRTIO_NET_F_CTRL_VQ)) { 148740237840SJason Wang *has_cvq = 1; 148840237840SJason Wang } else { 148940237840SJason Wang *has_cvq = 0; 149040237840SJason Wang } 149140237840SJason Wang 149240237840SJason Wang if (features & (1 << VIRTIO_NET_F_MQ)) { 149340237840SJason Wang config = g_malloc0(config_size + sizeof(*max_queue_pairs)); 149440237840SJason Wang config->off = offsetof(struct virtio_net_config, max_virtqueue_pairs); 149540237840SJason Wang config->len = sizeof(*max_queue_pairs); 149640237840SJason Wang 149740237840SJason Wang ret = ioctl(fd, VHOST_VDPA_GET_CONFIG, config); 149840237840SJason Wang if (ret) { 149940237840SJason Wang error_setg(errp, "Fail to get config from vhost-vDPA device"); 150040237840SJason Wang return -ret; 150140237840SJason Wang } 150240237840SJason Wang 150340237840SJason Wang max_queue_pairs = (__virtio16 *)&config->buf; 150440237840SJason Wang 150540237840SJason Wang return lduw_le_p(max_queue_pairs); 150640237840SJason Wang } 150740237840SJason Wang 150840237840SJason Wang return 1; 150940237840SJason Wang } 151040237840SJason Wang 15111e0a84eaSCindy Lu int net_init_vhost_vdpa(const Netdev *netdev, const char *name, 15121e0a84eaSCindy Lu NetClientState *peer, Error **errp) 15131e0a84eaSCindy Lu { 15141e0a84eaSCindy Lu const NetdevVhostVDPAOptions *opts; 15158170ab3fSEugenio Pérez uint64_t features; 1516654790b6SJason Wang int vdpa_device_fd; 1517eb3cb751SEugenio Pérez g_autofree NetClientState **ncs = NULL; 1518a585fad2SEugenio Pérez struct vhost_vdpa_iova_range iova_range; 1519eb3cb751SEugenio Pérez NetClientState *nc; 1520aed5da45SEugenio Pérez int queue_pairs, r, i = 0, has_cvq = 0; 15211e0a84eaSCindy Lu 15221e0a84eaSCindy Lu assert(netdev->type == NET_CLIENT_DRIVER_VHOST_VDPA); 15231e0a84eaSCindy Lu opts = &netdev->u.vhost_vdpa; 15247480874aSMarkus Armbruster if (!opts->vhostdev && !opts->vhostfd) { 15258801ccd0SSi-Wei Liu error_setg(errp, 15268801ccd0SSi-Wei Liu "vhost-vdpa: neither vhostdev= nor vhostfd= was specified"); 1527c8295404SEugenio Pérez return -1; 1528c8295404SEugenio Pérez } 15297327813dSJason Wang 15307480874aSMarkus Armbruster if (opts->vhostdev && opts->vhostfd) { 15318801ccd0SSi-Wei Liu error_setg(errp, 15328801ccd0SSi-Wei Liu "vhost-vdpa: vhostdev= and vhostfd= are mutually exclusive"); 15338801ccd0SSi-Wei Liu return -1; 15348801ccd0SSi-Wei Liu } 15358801ccd0SSi-Wei Liu 15367480874aSMarkus Armbruster if (opts->vhostdev) { 15370351152bSEugenio Pérez vdpa_device_fd = qemu_open(opts->vhostdev, O_RDWR, errp); 15387327813dSJason Wang if (vdpa_device_fd == -1) { 15397327813dSJason Wang return -errno; 15407327813dSJason Wang } 15415107fd3eSPeter Maydell } else { 15425107fd3eSPeter Maydell /* has_vhostfd */ 15438801ccd0SSi-Wei Liu vdpa_device_fd = monitor_fd_param(monitor_cur(), opts->vhostfd, errp); 15448801ccd0SSi-Wei Liu if (vdpa_device_fd == -1) { 15458801ccd0SSi-Wei Liu error_prepend(errp, "vhost-vdpa: unable to parse vhostfd: "); 15468801ccd0SSi-Wei Liu return -1; 15478801ccd0SSi-Wei Liu } 15488801ccd0SSi-Wei Liu } 15497327813dSJason Wang 15508170ab3fSEugenio Pérez r = vhost_vdpa_get_features(vdpa_device_fd, &features, errp); 15518170ab3fSEugenio Pérez if (unlikely(r < 0)) { 1552aed5da45SEugenio Pérez goto err; 15538170ab3fSEugenio Pérez } 15548170ab3fSEugenio Pérez 15558170ab3fSEugenio Pérez queue_pairs = vhost_vdpa_get_max_queue_pairs(vdpa_device_fd, features, 155640237840SJason Wang &has_cvq, errp); 155740237840SJason Wang if (queue_pairs < 0) { 15587327813dSJason Wang qemu_close(vdpa_device_fd); 155940237840SJason Wang return queue_pairs; 15607327813dSJason Wang } 15617327813dSJason Wang 1562bf7a2ad8SLongpeng r = vhost_vdpa_get_iova_range(vdpa_device_fd, &iova_range); 1563bf7a2ad8SLongpeng if (unlikely(r < 0)) { 1564bf7a2ad8SLongpeng error_setg(errp, "vhost-vdpa: get iova range failed: %s", 1565bf7a2ad8SLongpeng strerror(-r)); 1566bf7a2ad8SLongpeng goto err; 1567bf7a2ad8SLongpeng } 1568bf7a2ad8SLongpeng 156900ef422eSEugenio Pérez if (opts->x_svq && !vhost_vdpa_net_valid_svq_features(features, errp)) { 157000ef422eSEugenio Pérez goto err; 15711576dbb5SEugenio Pérez } 15721576dbb5SEugenio Pérez 157340237840SJason Wang ncs = g_malloc0(sizeof(*ncs) * queue_pairs); 157440237840SJason Wang 157540237840SJason Wang for (i = 0; i < queue_pairs; i++) { 157640237840SJason Wang ncs[i] = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name, 15771576dbb5SEugenio Pérez vdpa_device_fd, i, 2, true, opts->x_svq, 1578152128d6SEugenio Pérez iova_range, features, errp); 157940237840SJason Wang if (!ncs[i]) 158040237840SJason Wang goto err; 158140237840SJason Wang } 158240237840SJason Wang 158340237840SJason Wang if (has_cvq) { 158440237840SJason Wang nc = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name, 15851576dbb5SEugenio Pérez vdpa_device_fd, i, 1, false, 1586152128d6SEugenio Pérez opts->x_svq, iova_range, features, errp); 158740237840SJason Wang if (!nc) 158840237840SJason Wang goto err; 158940237840SJason Wang } 159040237840SJason Wang 1591654790b6SJason Wang return 0; 159240237840SJason Wang 159340237840SJason Wang err: 159440237840SJason Wang if (i) { 15959bd05507SSi-Wei Liu for (i--; i >= 0; i--) { 15969bd05507SSi-Wei Liu qemu_del_net_client(ncs[i]); 15979bd05507SSi-Wei Liu } 159840237840SJason Wang } 15991576dbb5SEugenio Pérez 160040237840SJason Wang qemu_close(vdpa_device_fd); 160140237840SJason Wang 160240237840SJason Wang return -1; 16031e0a84eaSCindy Lu } 1604