xref: /openbmc/qemu/net/vhost-vdpa.c (revision 259ebed4)
1 /*
2  * vhost-vdpa.c
3  *
4  * Copyright(c) 2017-2018 Intel Corporation.
5  * Copyright(c) 2020 Red Hat, Inc.
6  *
7  * This work is licensed under the terms of the GNU GPL, version 2 or later.
8  * See the COPYING file in the top-level directory.
9  *
10  */
11 
12 #include "qemu/osdep.h"
13 #include "clients.h"
14 #include "hw/virtio/virtio-net.h"
15 #include "net/vhost_net.h"
16 #include "net/vhost-vdpa.h"
17 #include "hw/virtio/vhost-vdpa.h"
18 #include "qemu/config-file.h"
19 #include "qemu/error-report.h"
20 #include "qemu/log.h"
21 #include "qemu/memalign.h"
22 #include "qemu/option.h"
23 #include "qapi/error.h"
24 #include <linux/vhost.h>
25 #include <sys/ioctl.h>
26 #include <err.h>
27 #include "standard-headers/linux/virtio_net.h"
28 #include "monitor/monitor.h"
29 #include "migration/migration.h"
30 #include "migration/misc.h"
31 #include "hw/virtio/vhost.h"
32 
33 /* Todo:need to add the multiqueue support here */
34 typedef struct VhostVDPAState {
35     NetClientState nc;
36     struct vhost_vdpa vhost_vdpa;
37     Notifier migration_state;
38     VHostNetState *vhost_net;
39 
40     /* Control commands shadow buffers */
41     void *cvq_cmd_out_buffer;
42     virtio_net_ctrl_ack *status;
43 
44     /* The device always have SVQ enabled */
45     bool always_svq;
46 
47     /* The device can isolate CVQ in its own ASID */
48     bool cvq_isolated;
49 
50     bool started;
51 } VhostVDPAState;
52 
53 /*
54  * The array is sorted alphabetically in ascending order,
55  * with the exception of VHOST_INVALID_FEATURE_BIT,
56  * which should always be the last entry.
57  */
58 const int vdpa_feature_bits[] = {
59     VIRTIO_F_ANY_LAYOUT,
60     VIRTIO_F_IOMMU_PLATFORM,
61     VIRTIO_F_NOTIFY_ON_EMPTY,
62     VIRTIO_F_RING_PACKED,
63     VIRTIO_F_RING_RESET,
64     VIRTIO_F_VERSION_1,
65     VIRTIO_NET_F_CSUM,
66     VIRTIO_NET_F_CTRL_GUEST_OFFLOADS,
67     VIRTIO_NET_F_CTRL_MAC_ADDR,
68     VIRTIO_NET_F_CTRL_RX,
69     VIRTIO_NET_F_CTRL_RX_EXTRA,
70     VIRTIO_NET_F_CTRL_VLAN,
71     VIRTIO_NET_F_CTRL_VQ,
72     VIRTIO_NET_F_GSO,
73     VIRTIO_NET_F_GUEST_CSUM,
74     VIRTIO_NET_F_GUEST_ECN,
75     VIRTIO_NET_F_GUEST_TSO4,
76     VIRTIO_NET_F_GUEST_TSO6,
77     VIRTIO_NET_F_GUEST_UFO,
78     VIRTIO_NET_F_GUEST_USO4,
79     VIRTIO_NET_F_GUEST_USO6,
80     VIRTIO_NET_F_HASH_REPORT,
81     VIRTIO_NET_F_HOST_ECN,
82     VIRTIO_NET_F_HOST_TSO4,
83     VIRTIO_NET_F_HOST_TSO6,
84     VIRTIO_NET_F_HOST_UFO,
85     VIRTIO_NET_F_HOST_USO,
86     VIRTIO_NET_F_MQ,
87     VIRTIO_NET_F_MRG_RXBUF,
88     VIRTIO_NET_F_MTU,
89     VIRTIO_NET_F_RSS,
90     VIRTIO_NET_F_STATUS,
91     VIRTIO_RING_F_EVENT_IDX,
92     VIRTIO_RING_F_INDIRECT_DESC,
93 
94     /* VHOST_INVALID_FEATURE_BIT should always be the last entry */
95     VHOST_INVALID_FEATURE_BIT
96 };
97 
98 /** Supported device specific feature bits with SVQ */
99 static const uint64_t vdpa_svq_device_features =
100     BIT_ULL(VIRTIO_NET_F_CSUM) |
101     BIT_ULL(VIRTIO_NET_F_GUEST_CSUM) |
102     BIT_ULL(VIRTIO_NET_F_CTRL_GUEST_OFFLOADS) |
103     BIT_ULL(VIRTIO_NET_F_MTU) |
104     BIT_ULL(VIRTIO_NET_F_MAC) |
105     BIT_ULL(VIRTIO_NET_F_GUEST_TSO4) |
106     BIT_ULL(VIRTIO_NET_F_GUEST_TSO6) |
107     BIT_ULL(VIRTIO_NET_F_GUEST_ECN) |
108     BIT_ULL(VIRTIO_NET_F_GUEST_UFO) |
109     BIT_ULL(VIRTIO_NET_F_HOST_TSO4) |
110     BIT_ULL(VIRTIO_NET_F_HOST_TSO6) |
111     BIT_ULL(VIRTIO_NET_F_HOST_ECN) |
112     BIT_ULL(VIRTIO_NET_F_HOST_UFO) |
113     BIT_ULL(VIRTIO_NET_F_MRG_RXBUF) |
114     BIT_ULL(VIRTIO_NET_F_STATUS) |
115     BIT_ULL(VIRTIO_NET_F_CTRL_VQ) |
116     BIT_ULL(VIRTIO_NET_F_CTRL_RX) |
117     BIT_ULL(VIRTIO_NET_F_CTRL_VLAN) |
118     BIT_ULL(VIRTIO_NET_F_CTRL_RX_EXTRA) |
119     BIT_ULL(VIRTIO_NET_F_MQ) |
120     BIT_ULL(VIRTIO_F_ANY_LAYOUT) |
121     BIT_ULL(VIRTIO_NET_F_CTRL_MAC_ADDR) |
122     /* VHOST_F_LOG_ALL is exposed by SVQ */
123     BIT_ULL(VHOST_F_LOG_ALL) |
124     BIT_ULL(VIRTIO_NET_F_RSC_EXT) |
125     BIT_ULL(VIRTIO_NET_F_STANDBY) |
126     BIT_ULL(VIRTIO_NET_F_SPEED_DUPLEX);
127 
128 #define VHOST_VDPA_NET_CVQ_ASID 1
129 
130 VHostNetState *vhost_vdpa_get_vhost_net(NetClientState *nc)
131 {
132     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
133     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
134     return s->vhost_net;
135 }
136 
137 static size_t vhost_vdpa_net_cvq_cmd_len(void)
138 {
139     /*
140      * MAC_TABLE_SET is the ctrl command that produces the longer out buffer.
141      * In buffer is always 1 byte, so it should fit here
142      */
143     return sizeof(struct virtio_net_ctrl_hdr) +
144            2 * sizeof(struct virtio_net_ctrl_mac) +
145            MAC_TABLE_ENTRIES * ETH_ALEN;
146 }
147 
148 static size_t vhost_vdpa_net_cvq_cmd_page_len(void)
149 {
150     return ROUND_UP(vhost_vdpa_net_cvq_cmd_len(), qemu_real_host_page_size());
151 }
152 
153 static bool vhost_vdpa_net_valid_svq_features(uint64_t features, Error **errp)
154 {
155     uint64_t invalid_dev_features =
156         features & ~vdpa_svq_device_features &
157         /* Transport are all accepted at this point */
158         ~MAKE_64BIT_MASK(VIRTIO_TRANSPORT_F_START,
159                          VIRTIO_TRANSPORT_F_END - VIRTIO_TRANSPORT_F_START);
160 
161     if (invalid_dev_features) {
162         error_setg(errp, "vdpa svq does not work with features 0x%" PRIx64,
163                    invalid_dev_features);
164         return false;
165     }
166 
167     return vhost_svq_valid_features(features, errp);
168 }
169 
170 static int vhost_vdpa_net_check_device_id(struct vhost_net *net)
171 {
172     uint32_t device_id;
173     int ret;
174     struct vhost_dev *hdev;
175 
176     hdev = (struct vhost_dev *)&net->dev;
177     ret = hdev->vhost_ops->vhost_get_device_id(hdev, &device_id);
178     if (device_id != VIRTIO_ID_NET) {
179         return -ENOTSUP;
180     }
181     return ret;
182 }
183 
184 static int vhost_vdpa_add(NetClientState *ncs, void *be,
185                           int queue_pair_index, int nvqs)
186 {
187     VhostNetOptions options;
188     struct vhost_net *net = NULL;
189     VhostVDPAState *s;
190     int ret;
191 
192     options.backend_type = VHOST_BACKEND_TYPE_VDPA;
193     assert(ncs->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
194     s = DO_UPCAST(VhostVDPAState, nc, ncs);
195     options.net_backend = ncs;
196     options.opaque      = be;
197     options.busyloop_timeout = 0;
198     options.nvqs = nvqs;
199 
200     net = vhost_net_init(&options);
201     if (!net) {
202         error_report("failed to init vhost_net for queue");
203         goto err_init;
204     }
205     s->vhost_net = net;
206     ret = vhost_vdpa_net_check_device_id(net);
207     if (ret) {
208         goto err_check;
209     }
210     return 0;
211 err_check:
212     vhost_net_cleanup(net);
213     g_free(net);
214 err_init:
215     return -1;
216 }
217 
218 static void vhost_vdpa_cleanup(NetClientState *nc)
219 {
220     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
221 
222     /*
223      * If a peer NIC is attached, do not cleanup anything.
224      * Cleanup will happen as a part of qemu_cleanup() -> net_cleanup()
225      * when the guest is shutting down.
226      */
227     if (nc->peer && nc->peer->info->type == NET_CLIENT_DRIVER_NIC) {
228         return;
229     }
230     munmap(s->cvq_cmd_out_buffer, vhost_vdpa_net_cvq_cmd_page_len());
231     munmap(s->status, vhost_vdpa_net_cvq_cmd_page_len());
232     if (s->vhost_net) {
233         vhost_net_cleanup(s->vhost_net);
234         g_free(s->vhost_net);
235         s->vhost_net = NULL;
236     }
237      if (s->vhost_vdpa.device_fd >= 0) {
238         qemu_close(s->vhost_vdpa.device_fd);
239         s->vhost_vdpa.device_fd = -1;
240     }
241 }
242 
243 static bool vhost_vdpa_has_vnet_hdr(NetClientState *nc)
244 {
245     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
246 
247     return true;
248 }
249 
250 static bool vhost_vdpa_has_ufo(NetClientState *nc)
251 {
252     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
253     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
254     uint64_t features = 0;
255     features |= (1ULL << VIRTIO_NET_F_HOST_UFO);
256     features = vhost_net_get_features(s->vhost_net, features);
257     return !!(features & (1ULL << VIRTIO_NET_F_HOST_UFO));
258 
259 }
260 
261 static bool vhost_vdpa_check_peer_type(NetClientState *nc, ObjectClass *oc,
262                                        Error **errp)
263 {
264     const char *driver = object_class_get_name(oc);
265 
266     if (!g_str_has_prefix(driver, "virtio-net-")) {
267         error_setg(errp, "vhost-vdpa requires frontend driver virtio-net-*");
268         return false;
269     }
270 
271     return true;
272 }
273 
274 /** Dummy receive in case qemu falls back to userland tap networking */
275 static ssize_t vhost_vdpa_receive(NetClientState *nc, const uint8_t *buf,
276                                   size_t size)
277 {
278     return size;
279 }
280 
281 /** From any vdpa net client, get the netclient of the first queue pair */
282 static VhostVDPAState *vhost_vdpa_net_first_nc_vdpa(VhostVDPAState *s)
283 {
284     NICState *nic = qemu_get_nic(s->nc.peer);
285     NetClientState *nc0 = qemu_get_peer(nic->ncs, 0);
286 
287     return DO_UPCAST(VhostVDPAState, nc, nc0);
288 }
289 
290 static void vhost_vdpa_net_log_global_enable(VhostVDPAState *s, bool enable)
291 {
292     struct vhost_vdpa *v = &s->vhost_vdpa;
293     VirtIONet *n;
294     VirtIODevice *vdev;
295     int data_queue_pairs, cvq, r;
296 
297     /* We are only called on the first data vqs and only if x-svq is not set */
298     if (s->vhost_vdpa.shadow_vqs_enabled == enable) {
299         return;
300     }
301 
302     vdev = v->dev->vdev;
303     n = VIRTIO_NET(vdev);
304     if (!n->vhost_started) {
305         return;
306     }
307 
308     data_queue_pairs = n->multiqueue ? n->max_queue_pairs : 1;
309     cvq = virtio_vdev_has_feature(vdev, VIRTIO_NET_F_CTRL_VQ) ?
310                                   n->max_ncs - n->max_queue_pairs : 0;
311     /*
312      * TODO: vhost_net_stop does suspend, get_base and reset. We can be smarter
313      * in the future and resume the device if read-only operations between
314      * suspend and reset goes wrong.
315      */
316     vhost_net_stop(vdev, n->nic->ncs, data_queue_pairs, cvq);
317 
318     /* Start will check migration setup_or_active to configure or not SVQ */
319     r = vhost_net_start(vdev, n->nic->ncs, data_queue_pairs, cvq);
320     if (unlikely(r < 0)) {
321         error_report("unable to start vhost net: %s(%d)", g_strerror(-r), -r);
322     }
323 }
324 
325 static void vdpa_net_migration_state_notifier(Notifier *notifier, void *data)
326 {
327     MigrationState *migration = data;
328     VhostVDPAState *s = container_of(notifier, VhostVDPAState,
329                                      migration_state);
330 
331     if (migration_in_setup(migration)) {
332         vhost_vdpa_net_log_global_enable(s, true);
333     } else if (migration_has_failed(migration)) {
334         vhost_vdpa_net_log_global_enable(s, false);
335     }
336 }
337 
338 static void vhost_vdpa_net_data_start_first(VhostVDPAState *s)
339 {
340     struct vhost_vdpa *v = &s->vhost_vdpa;
341 
342     migration_add_notifier(&s->migration_state,
343                            vdpa_net_migration_state_notifier);
344     if (v->shadow_vqs_enabled) {
345         v->iova_tree = vhost_iova_tree_new(v->iova_range.first,
346                                            v->iova_range.last);
347     }
348 }
349 
350 static int vhost_vdpa_net_data_start(NetClientState *nc)
351 {
352     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
353     struct vhost_vdpa *v = &s->vhost_vdpa;
354 
355     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
356 
357     if (s->always_svq ||
358         migration_is_setup_or_active(migrate_get_current()->state)) {
359         v->shadow_vqs_enabled = true;
360         v->shadow_data = true;
361     } else {
362         v->shadow_vqs_enabled = false;
363         v->shadow_data = false;
364     }
365 
366     if (v->index == 0) {
367         vhost_vdpa_net_data_start_first(s);
368         return 0;
369     }
370 
371     if (v->shadow_vqs_enabled) {
372         VhostVDPAState *s0 = vhost_vdpa_net_first_nc_vdpa(s);
373         v->iova_tree = s0->vhost_vdpa.iova_tree;
374     }
375 
376     return 0;
377 }
378 
379 static int vhost_vdpa_net_data_load(NetClientState *nc)
380 {
381     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
382     struct vhost_vdpa *v = &s->vhost_vdpa;
383     bool has_cvq = v->dev->vq_index_end % 2;
384 
385     if (has_cvq) {
386         return 0;
387     }
388 
389     for (int i = 0; i < v->dev->nvqs; ++i) {
390         vhost_vdpa_set_vring_ready(v, i + v->dev->vq_index);
391     }
392     return 0;
393 }
394 
395 static void vhost_vdpa_net_client_stop(NetClientState *nc)
396 {
397     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
398     struct vhost_dev *dev;
399 
400     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
401 
402     if (s->vhost_vdpa.index == 0) {
403         migration_remove_notifier(&s->migration_state);
404     }
405 
406     dev = s->vhost_vdpa.dev;
407     if (dev->vq_index + dev->nvqs == dev->vq_index_end) {
408         g_clear_pointer(&s->vhost_vdpa.iova_tree, vhost_iova_tree_delete);
409     } else {
410         s->vhost_vdpa.iova_tree = NULL;
411     }
412 }
413 
414 static NetClientInfo net_vhost_vdpa_info = {
415         .type = NET_CLIENT_DRIVER_VHOST_VDPA,
416         .size = sizeof(VhostVDPAState),
417         .receive = vhost_vdpa_receive,
418         .start = vhost_vdpa_net_data_start,
419         .load = vhost_vdpa_net_data_load,
420         .stop = vhost_vdpa_net_client_stop,
421         .cleanup = vhost_vdpa_cleanup,
422         .has_vnet_hdr = vhost_vdpa_has_vnet_hdr,
423         .has_ufo = vhost_vdpa_has_ufo,
424         .check_peer_type = vhost_vdpa_check_peer_type,
425 };
426 
427 static int64_t vhost_vdpa_get_vring_group(int device_fd, unsigned vq_index,
428                                           Error **errp)
429 {
430     struct vhost_vring_state state = {
431         .index = vq_index,
432     };
433     int r = ioctl(device_fd, VHOST_VDPA_GET_VRING_GROUP, &state);
434 
435     if (unlikely(r < 0)) {
436         r = -errno;
437         error_setg_errno(errp, errno, "Cannot get VQ %u group", vq_index);
438         return r;
439     }
440 
441     return state.num;
442 }
443 
444 static int vhost_vdpa_set_address_space_id(struct vhost_vdpa *v,
445                                            unsigned vq_group,
446                                            unsigned asid_num)
447 {
448     struct vhost_vring_state asid = {
449         .index = vq_group,
450         .num = asid_num,
451     };
452     int r;
453 
454     r = ioctl(v->device_fd, VHOST_VDPA_SET_GROUP_ASID, &asid);
455     if (unlikely(r < 0)) {
456         error_report("Can't set vq group %u asid %u, errno=%d (%s)",
457                      asid.index, asid.num, errno, g_strerror(errno));
458     }
459     return r;
460 }
461 
462 static void vhost_vdpa_cvq_unmap_buf(struct vhost_vdpa *v, void *addr)
463 {
464     VhostIOVATree *tree = v->iova_tree;
465     DMAMap needle = {
466         /*
467          * No need to specify size or to look for more translations since
468          * this contiguous chunk was allocated by us.
469          */
470         .translated_addr = (hwaddr)(uintptr_t)addr,
471     };
472     const DMAMap *map = vhost_iova_tree_find_iova(tree, &needle);
473     int r;
474 
475     if (unlikely(!map)) {
476         error_report("Cannot locate expected map");
477         return;
478     }
479 
480     r = vhost_vdpa_dma_unmap(v, v->address_space_id, map->iova, map->size + 1);
481     if (unlikely(r != 0)) {
482         error_report("Device cannot unmap: %s(%d)", g_strerror(r), r);
483     }
484 
485     vhost_iova_tree_remove(tree, *map);
486 }
487 
488 /** Map CVQ buffer. */
489 static int vhost_vdpa_cvq_map_buf(struct vhost_vdpa *v, void *buf, size_t size,
490                                   bool write)
491 {
492     DMAMap map = {};
493     int r;
494 
495     map.translated_addr = (hwaddr)(uintptr_t)buf;
496     map.size = size - 1;
497     map.perm = write ? IOMMU_RW : IOMMU_RO,
498     r = vhost_iova_tree_map_alloc(v->iova_tree, &map);
499     if (unlikely(r != IOVA_OK)) {
500         error_report("Cannot map injected element");
501         return r;
502     }
503 
504     r = vhost_vdpa_dma_map(v, v->address_space_id, map.iova,
505                            vhost_vdpa_net_cvq_cmd_page_len(), buf, !write);
506     if (unlikely(r < 0)) {
507         goto dma_map_err;
508     }
509 
510     return 0;
511 
512 dma_map_err:
513     vhost_iova_tree_remove(v->iova_tree, map);
514     return r;
515 }
516 
517 static int vhost_vdpa_net_cvq_start(NetClientState *nc)
518 {
519     VhostVDPAState *s, *s0;
520     struct vhost_vdpa *v;
521     int64_t cvq_group;
522     int r;
523     Error *err = NULL;
524 
525     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
526 
527     s = DO_UPCAST(VhostVDPAState, nc, nc);
528     v = &s->vhost_vdpa;
529 
530     s0 = vhost_vdpa_net_first_nc_vdpa(s);
531     v->shadow_data = s0->vhost_vdpa.shadow_vqs_enabled;
532     v->shadow_vqs_enabled = s0->vhost_vdpa.shadow_vqs_enabled;
533     s->vhost_vdpa.address_space_id = VHOST_VDPA_GUEST_PA_ASID;
534 
535     if (s->vhost_vdpa.shadow_data) {
536         /* SVQ is already configured for all virtqueues */
537         goto out;
538     }
539 
540     /*
541      * If we early return in these cases SVQ will not be enabled. The migration
542      * will be blocked as long as vhost-vdpa backends will not offer _F_LOG.
543      */
544     if (!vhost_vdpa_net_valid_svq_features(v->dev->features, NULL)) {
545         return 0;
546     }
547 
548     if (!s->cvq_isolated) {
549         return 0;
550     }
551 
552     cvq_group = vhost_vdpa_get_vring_group(v->device_fd,
553                                            v->dev->vq_index_end - 1,
554                                            &err);
555     if (unlikely(cvq_group < 0)) {
556         error_report_err(err);
557         return cvq_group;
558     }
559 
560     r = vhost_vdpa_set_address_space_id(v, cvq_group, VHOST_VDPA_NET_CVQ_ASID);
561     if (unlikely(r < 0)) {
562         return r;
563     }
564 
565     v->shadow_vqs_enabled = true;
566     s->vhost_vdpa.address_space_id = VHOST_VDPA_NET_CVQ_ASID;
567 
568 out:
569     if (!s->vhost_vdpa.shadow_vqs_enabled) {
570         return 0;
571     }
572 
573     if (s0->vhost_vdpa.iova_tree) {
574         /*
575          * SVQ is already configured for all virtqueues.  Reuse IOVA tree for
576          * simplicity, whether CVQ shares ASID with guest or not, because:
577          * - Memory listener need access to guest's memory addresses allocated
578          *   in the IOVA tree.
579          * - There should be plenty of IOVA address space for both ASID not to
580          *   worry about collisions between them.  Guest's translations are
581          *   still validated with virtio virtqueue_pop so there is no risk for
582          *   the guest to access memory that it shouldn't.
583          *
584          * To allocate a iova tree per ASID is doable but it complicates the
585          * code and it is not worth it for the moment.
586          */
587         v->iova_tree = s0->vhost_vdpa.iova_tree;
588     } else {
589         v->iova_tree = vhost_iova_tree_new(v->iova_range.first,
590                                            v->iova_range.last);
591     }
592 
593     r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer,
594                                vhost_vdpa_net_cvq_cmd_page_len(), false);
595     if (unlikely(r < 0)) {
596         return r;
597     }
598 
599     r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->status,
600                                vhost_vdpa_net_cvq_cmd_page_len(), true);
601     if (unlikely(r < 0)) {
602         vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer);
603     }
604 
605     return r;
606 }
607 
608 static void vhost_vdpa_net_cvq_stop(NetClientState *nc)
609 {
610     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
611 
612     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
613 
614     if (s->vhost_vdpa.shadow_vqs_enabled) {
615         vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer);
616         vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->status);
617     }
618 
619     vhost_vdpa_net_client_stop(nc);
620 }
621 
622 static ssize_t vhost_vdpa_net_cvq_add(VhostVDPAState *s,
623                                     const struct iovec *out_sg, size_t out_num,
624                                     const struct iovec *in_sg, size_t in_num)
625 {
626     VhostShadowVirtqueue *svq = g_ptr_array_index(s->vhost_vdpa.shadow_vqs, 0);
627     int r;
628 
629     r = vhost_svq_add(svq, out_sg, out_num, in_sg, in_num, NULL);
630     if (unlikely(r != 0)) {
631         if (unlikely(r == -ENOSPC)) {
632             qemu_log_mask(LOG_GUEST_ERROR, "%s: No space on device queue\n",
633                           __func__);
634         }
635     }
636 
637     return r;
638 }
639 
640 /*
641  * Convenience wrapper to poll SVQ for multiple control commands.
642  *
643  * Caller should hold the BQL when invoking this function, and should take
644  * the answer before SVQ pulls by itself when BQL is released.
645  */
646 static ssize_t vhost_vdpa_net_svq_poll(VhostVDPAState *s, size_t cmds_in_flight)
647 {
648     VhostShadowVirtqueue *svq = g_ptr_array_index(s->vhost_vdpa.shadow_vqs, 0);
649     return vhost_svq_poll(svq, cmds_in_flight);
650 }
651 
652 static void vhost_vdpa_net_load_cursor_reset(VhostVDPAState *s,
653                                              struct iovec *out_cursor,
654                                              struct iovec *in_cursor)
655 {
656     /* reset the cursor of the output buffer for the device */
657     out_cursor->iov_base = s->cvq_cmd_out_buffer;
658     out_cursor->iov_len = vhost_vdpa_net_cvq_cmd_page_len();
659 
660     /* reset the cursor of the in buffer for the device */
661     in_cursor->iov_base = s->status;
662     in_cursor->iov_len = vhost_vdpa_net_cvq_cmd_page_len();
663 }
664 
665 /*
666  * Poll SVQ for multiple pending control commands and check the device's ack.
667  *
668  * Caller should hold the BQL when invoking this function.
669  *
670  * @s: The VhostVDPAState
671  * @len: The length of the pending status shadow buffer
672  */
673 static ssize_t vhost_vdpa_net_svq_flush(VhostVDPAState *s, size_t len)
674 {
675     /* device uses a one-byte length ack for each control command */
676     ssize_t dev_written = vhost_vdpa_net_svq_poll(s, len);
677     if (unlikely(dev_written != len)) {
678         return -EIO;
679     }
680 
681     /* check the device's ack */
682     for (int i = 0; i < len; ++i) {
683         if (s->status[i] != VIRTIO_NET_OK) {
684             return -EIO;
685         }
686     }
687     return 0;
688 }
689 
690 static ssize_t vhost_vdpa_net_load_cmd(VhostVDPAState *s,
691                                        struct iovec *out_cursor,
692                                        struct iovec *in_cursor, uint8_t class,
693                                        uint8_t cmd, const struct iovec *data_sg,
694                                        size_t data_num)
695 {
696     const struct virtio_net_ctrl_hdr ctrl = {
697         .class = class,
698         .cmd = cmd,
699     };
700     size_t data_size = iov_size(data_sg, data_num), cmd_size;
701     struct iovec out, in;
702     ssize_t r;
703     unsigned dummy_cursor_iov_cnt;
704     VhostShadowVirtqueue *svq = g_ptr_array_index(s->vhost_vdpa.shadow_vqs, 0);
705 
706     assert(data_size < vhost_vdpa_net_cvq_cmd_page_len() - sizeof(ctrl));
707     cmd_size = sizeof(ctrl) + data_size;
708     if (vhost_svq_available_slots(svq) < 2 ||
709         iov_size(out_cursor, 1) < cmd_size) {
710         /*
711          * It is time to flush all pending control commands if SVQ is full
712          * or control commands shadow buffers are full.
713          *
714          * We can poll here since we've had BQL from the time
715          * we sent the descriptor.
716          */
717         r = vhost_vdpa_net_svq_flush(s, in_cursor->iov_base -
718                                      (void *)s->status);
719         if (unlikely(r < 0)) {
720             return r;
721         }
722 
723         vhost_vdpa_net_load_cursor_reset(s, out_cursor, in_cursor);
724     }
725 
726     /* pack the CVQ command header */
727     iov_from_buf(out_cursor, 1, 0, &ctrl, sizeof(ctrl));
728     /* pack the CVQ command command-specific-data */
729     iov_to_buf(data_sg, data_num, 0,
730                out_cursor->iov_base + sizeof(ctrl), data_size);
731 
732     /* extract the required buffer from the cursor for output */
733     iov_copy(&out, 1, out_cursor, 1, 0, cmd_size);
734     /* extract the required buffer from the cursor for input */
735     iov_copy(&in, 1, in_cursor, 1, 0, sizeof(*s->status));
736 
737     r = vhost_vdpa_net_cvq_add(s, &out, 1, &in, 1);
738     if (unlikely(r < 0)) {
739         return r;
740     }
741 
742     /* iterate the cursors */
743     dummy_cursor_iov_cnt = 1;
744     iov_discard_front(&out_cursor, &dummy_cursor_iov_cnt, cmd_size);
745     dummy_cursor_iov_cnt = 1;
746     iov_discard_front(&in_cursor, &dummy_cursor_iov_cnt, sizeof(*s->status));
747 
748     return 0;
749 }
750 
751 static int vhost_vdpa_net_load_mac(VhostVDPAState *s, const VirtIONet *n,
752                                    struct iovec *out_cursor,
753                                    struct iovec *in_cursor)
754 {
755     if (virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_MAC_ADDR)) {
756         const struct iovec data = {
757             .iov_base = (void *)n->mac,
758             .iov_len = sizeof(n->mac),
759         };
760         ssize_t r = vhost_vdpa_net_load_cmd(s, out_cursor, in_cursor,
761                                             VIRTIO_NET_CTRL_MAC,
762                                             VIRTIO_NET_CTRL_MAC_ADDR_SET,
763                                             &data, 1);
764         if (unlikely(r < 0)) {
765             return r;
766         }
767     }
768 
769     /*
770      * According to VirtIO standard, "The device MUST have an
771      * empty MAC filtering table on reset.".
772      *
773      * Therefore, there is no need to send this CVQ command if the
774      * driver also sets an empty MAC filter table, which aligns with
775      * the device's defaults.
776      *
777      * Note that the device's defaults can mismatch the driver's
778      * configuration only at live migration.
779      */
780     if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_RX) ||
781         n->mac_table.in_use == 0) {
782         return 0;
783     }
784 
785     uint32_t uni_entries = n->mac_table.first_multi,
786              uni_macs_size = uni_entries * ETH_ALEN,
787              mul_entries = n->mac_table.in_use - uni_entries,
788              mul_macs_size = mul_entries * ETH_ALEN;
789     struct virtio_net_ctrl_mac uni = {
790         .entries = cpu_to_le32(uni_entries),
791     };
792     struct virtio_net_ctrl_mac mul = {
793         .entries = cpu_to_le32(mul_entries),
794     };
795     const struct iovec data[] = {
796         {
797             .iov_base = &uni,
798             .iov_len = sizeof(uni),
799         }, {
800             .iov_base = n->mac_table.macs,
801             .iov_len = uni_macs_size,
802         }, {
803             .iov_base = &mul,
804             .iov_len = sizeof(mul),
805         }, {
806             .iov_base = &n->mac_table.macs[uni_macs_size],
807             .iov_len = mul_macs_size,
808         },
809     };
810     ssize_t r = vhost_vdpa_net_load_cmd(s, out_cursor, in_cursor,
811                                         VIRTIO_NET_CTRL_MAC,
812                                         VIRTIO_NET_CTRL_MAC_TABLE_SET,
813                                         data, ARRAY_SIZE(data));
814     if (unlikely(r < 0)) {
815         return r;
816     }
817 
818     return 0;
819 }
820 
821 static int vhost_vdpa_net_load_mq(VhostVDPAState *s,
822                                   const VirtIONet *n,
823                                   struct iovec *out_cursor,
824                                   struct iovec *in_cursor)
825 {
826     struct virtio_net_ctrl_mq mq;
827     ssize_t r;
828 
829     if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_MQ)) {
830         return 0;
831     }
832 
833     mq.virtqueue_pairs = cpu_to_le16(n->curr_queue_pairs);
834     const struct iovec data = {
835         .iov_base = &mq,
836         .iov_len = sizeof(mq),
837     };
838     r = vhost_vdpa_net_load_cmd(s, out_cursor, in_cursor,
839                                 VIRTIO_NET_CTRL_MQ,
840                                 VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET,
841                                 &data, 1);
842     if (unlikely(r < 0)) {
843         return r;
844     }
845 
846     return 0;
847 }
848 
849 static int vhost_vdpa_net_load_offloads(VhostVDPAState *s,
850                                         const VirtIONet *n,
851                                         struct iovec *out_cursor,
852                                         struct iovec *in_cursor)
853 {
854     uint64_t offloads;
855     ssize_t r;
856 
857     if (!virtio_vdev_has_feature(&n->parent_obj,
858                                  VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)) {
859         return 0;
860     }
861 
862     if (n->curr_guest_offloads == virtio_net_supported_guest_offloads(n)) {
863         /*
864          * According to VirtIO standard, "Upon feature negotiation
865          * corresponding offload gets enabled to preserve
866          * backward compatibility.".
867          *
868          * Therefore, there is no need to send this CVQ command if the
869          * driver also enables all supported offloads, which aligns with
870          * the device's defaults.
871          *
872          * Note that the device's defaults can mismatch the driver's
873          * configuration only at live migration.
874          */
875         return 0;
876     }
877 
878     offloads = cpu_to_le64(n->curr_guest_offloads);
879     const struct iovec data = {
880         .iov_base = &offloads,
881         .iov_len = sizeof(offloads),
882     };
883     r = vhost_vdpa_net_load_cmd(s, out_cursor, in_cursor,
884                                 VIRTIO_NET_CTRL_GUEST_OFFLOADS,
885                                 VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET,
886                                 &data, 1);
887     if (unlikely(r < 0)) {
888         return r;
889     }
890 
891     return 0;
892 }
893 
894 static int vhost_vdpa_net_load_rx_mode(VhostVDPAState *s,
895                                        struct iovec *out_cursor,
896                                        struct iovec *in_cursor,
897                                        uint8_t cmd,
898                                        uint8_t on)
899 {
900     const struct iovec data = {
901         .iov_base = &on,
902         .iov_len = sizeof(on),
903     };
904     ssize_t r;
905 
906     r = vhost_vdpa_net_load_cmd(s, out_cursor, in_cursor,
907                                 VIRTIO_NET_CTRL_RX, cmd, &data, 1);
908     if (unlikely(r < 0)) {
909         return r;
910     }
911 
912     return 0;
913 }
914 
915 static int vhost_vdpa_net_load_rx(VhostVDPAState *s,
916                                   const VirtIONet *n,
917                                   struct iovec *out_cursor,
918                                   struct iovec *in_cursor)
919 {
920     ssize_t r;
921 
922     if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_RX)) {
923         return 0;
924     }
925 
926     /*
927      * According to virtio_net_reset(), device turns promiscuous mode
928      * on by default.
929      *
930      * Additionally, according to VirtIO standard, "Since there are
931      * no guarantees, it can use a hash filter or silently switch to
932      * allmulti or promiscuous mode if it is given too many addresses.".
933      * QEMU marks `n->mac_table.uni_overflow` if guest sets too many
934      * non-multicast MAC addresses, indicating that promiscuous mode
935      * should be enabled.
936      *
937      * Therefore, QEMU should only send this CVQ command if the
938      * `n->mac_table.uni_overflow` is not marked and `n->promisc` is off,
939      * which sets promiscuous mode on, different from the device's defaults.
940      *
941      * Note that the device's defaults can mismatch the driver's
942      * configuration only at live migration.
943      */
944     if (!n->mac_table.uni_overflow && !n->promisc) {
945         r = vhost_vdpa_net_load_rx_mode(s, out_cursor, in_cursor,
946                                         VIRTIO_NET_CTRL_RX_PROMISC, 0);
947         if (unlikely(r < 0)) {
948             return r;
949         }
950     }
951 
952     /*
953      * According to virtio_net_reset(), device turns all-multicast mode
954      * off by default.
955      *
956      * According to VirtIO standard, "Since there are no guarantees,
957      * it can use a hash filter or silently switch to allmulti or
958      * promiscuous mode if it is given too many addresses.". QEMU marks
959      * `n->mac_table.multi_overflow` if guest sets too many
960      * non-multicast MAC addresses.
961      *
962      * Therefore, QEMU should only send this CVQ command if the
963      * `n->mac_table.multi_overflow` is marked or `n->allmulti` is on,
964      * which sets all-multicast mode on, different from the device's defaults.
965      *
966      * Note that the device's defaults can mismatch the driver's
967      * configuration only at live migration.
968      */
969     if (n->mac_table.multi_overflow || n->allmulti) {
970         r = vhost_vdpa_net_load_rx_mode(s, out_cursor, in_cursor,
971                                         VIRTIO_NET_CTRL_RX_ALLMULTI, 1);
972         if (unlikely(r < 0)) {
973             return r;
974         }
975     }
976 
977     if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_RX_EXTRA)) {
978         return 0;
979     }
980 
981     /*
982      * According to virtio_net_reset(), device turns all-unicast mode
983      * off by default.
984      *
985      * Therefore, QEMU should only send this CVQ command if the driver
986      * sets all-unicast mode on, different from the device's defaults.
987      *
988      * Note that the device's defaults can mismatch the driver's
989      * configuration only at live migration.
990      */
991     if (n->alluni) {
992         r = vhost_vdpa_net_load_rx_mode(s, out_cursor, in_cursor,
993                                         VIRTIO_NET_CTRL_RX_ALLUNI, 1);
994         if (r < 0) {
995             return r;
996         }
997     }
998 
999     /*
1000      * According to virtio_net_reset(), device turns non-multicast mode
1001      * off by default.
1002      *
1003      * Therefore, QEMU should only send this CVQ command if the driver
1004      * sets non-multicast mode on, different from the device's defaults.
1005      *
1006      * Note that the device's defaults can mismatch the driver's
1007      * configuration only at live migration.
1008      */
1009     if (n->nomulti) {
1010         r = vhost_vdpa_net_load_rx_mode(s, out_cursor, in_cursor,
1011                                         VIRTIO_NET_CTRL_RX_NOMULTI, 1);
1012         if (r < 0) {
1013             return r;
1014         }
1015     }
1016 
1017     /*
1018      * According to virtio_net_reset(), device turns non-unicast mode
1019      * off by default.
1020      *
1021      * Therefore, QEMU should only send this CVQ command if the driver
1022      * sets non-unicast mode on, different from the device's defaults.
1023      *
1024      * Note that the device's defaults can mismatch the driver's
1025      * configuration only at live migration.
1026      */
1027     if (n->nouni) {
1028         r = vhost_vdpa_net_load_rx_mode(s, out_cursor, in_cursor,
1029                                         VIRTIO_NET_CTRL_RX_NOUNI, 1);
1030         if (r < 0) {
1031             return r;
1032         }
1033     }
1034 
1035     /*
1036      * According to virtio_net_reset(), device turns non-broadcast mode
1037      * off by default.
1038      *
1039      * Therefore, QEMU should only send this CVQ command if the driver
1040      * sets non-broadcast mode on, different from the device's defaults.
1041      *
1042      * Note that the device's defaults can mismatch the driver's
1043      * configuration only at live migration.
1044      */
1045     if (n->nobcast) {
1046         r = vhost_vdpa_net_load_rx_mode(s, out_cursor, in_cursor,
1047                                         VIRTIO_NET_CTRL_RX_NOBCAST, 1);
1048         if (r < 0) {
1049             return r;
1050         }
1051     }
1052 
1053     return 0;
1054 }
1055 
1056 static int vhost_vdpa_net_load_single_vlan(VhostVDPAState *s,
1057                                            const VirtIONet *n,
1058                                            struct iovec *out_cursor,
1059                                            struct iovec *in_cursor,
1060                                            uint16_t vid)
1061 {
1062     const struct iovec data = {
1063         .iov_base = &vid,
1064         .iov_len = sizeof(vid),
1065     };
1066     ssize_t r = vhost_vdpa_net_load_cmd(s, out_cursor, in_cursor,
1067                                         VIRTIO_NET_CTRL_VLAN,
1068                                         VIRTIO_NET_CTRL_VLAN_ADD,
1069                                         &data, 1);
1070     if (unlikely(r < 0)) {
1071         return r;
1072     }
1073 
1074     return 0;
1075 }
1076 
1077 static int vhost_vdpa_net_load_vlan(VhostVDPAState *s,
1078                                     const VirtIONet *n,
1079                                     struct iovec *out_cursor,
1080                                     struct iovec *in_cursor)
1081 {
1082     int r;
1083 
1084     if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_VLAN)) {
1085         return 0;
1086     }
1087 
1088     for (int i = 0; i < MAX_VLAN >> 5; i++) {
1089         for (int j = 0; n->vlans[i] && j <= 0x1f; j++) {
1090             if (n->vlans[i] & (1U << j)) {
1091                 r = vhost_vdpa_net_load_single_vlan(s, n, out_cursor,
1092                                                     in_cursor, (i << 5) + j);
1093                 if (unlikely(r != 0)) {
1094                     return r;
1095                 }
1096             }
1097         }
1098     }
1099 
1100     return 0;
1101 }
1102 
1103 static int vhost_vdpa_net_cvq_load(NetClientState *nc)
1104 {
1105     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
1106     struct vhost_vdpa *v = &s->vhost_vdpa;
1107     const VirtIONet *n;
1108     int r;
1109     struct iovec out_cursor, in_cursor;
1110 
1111     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
1112 
1113     vhost_vdpa_set_vring_ready(v, v->dev->vq_index);
1114 
1115     if (v->shadow_vqs_enabled) {
1116         n = VIRTIO_NET(v->dev->vdev);
1117         vhost_vdpa_net_load_cursor_reset(s, &out_cursor, &in_cursor);
1118         r = vhost_vdpa_net_load_mac(s, n, &out_cursor, &in_cursor);
1119         if (unlikely(r < 0)) {
1120             return r;
1121         }
1122         r = vhost_vdpa_net_load_mq(s, n, &out_cursor, &in_cursor);
1123         if (unlikely(r)) {
1124             return r;
1125         }
1126         r = vhost_vdpa_net_load_offloads(s, n, &out_cursor, &in_cursor);
1127         if (unlikely(r)) {
1128             return r;
1129         }
1130         r = vhost_vdpa_net_load_rx(s, n, &out_cursor, &in_cursor);
1131         if (unlikely(r)) {
1132             return r;
1133         }
1134         r = vhost_vdpa_net_load_vlan(s, n, &out_cursor, &in_cursor);
1135         if (unlikely(r)) {
1136             return r;
1137         }
1138 
1139         /*
1140          * We need to poll and check all pending device's used buffers.
1141          *
1142          * We can poll here since we've had BQL from the time
1143          * we sent the descriptor.
1144          */
1145         r = vhost_vdpa_net_svq_flush(s, in_cursor.iov_base - (void *)s->status);
1146         if (unlikely(r)) {
1147             return r;
1148         }
1149     }
1150 
1151     for (int i = 0; i < v->dev->vq_index; ++i) {
1152         vhost_vdpa_set_vring_ready(v, i);
1153     }
1154 
1155     return 0;
1156 }
1157 
1158 static NetClientInfo net_vhost_vdpa_cvq_info = {
1159     .type = NET_CLIENT_DRIVER_VHOST_VDPA,
1160     .size = sizeof(VhostVDPAState),
1161     .receive = vhost_vdpa_receive,
1162     .start = vhost_vdpa_net_cvq_start,
1163     .load = vhost_vdpa_net_cvq_load,
1164     .stop = vhost_vdpa_net_cvq_stop,
1165     .cleanup = vhost_vdpa_cleanup,
1166     .has_vnet_hdr = vhost_vdpa_has_vnet_hdr,
1167     .has_ufo = vhost_vdpa_has_ufo,
1168     .check_peer_type = vhost_vdpa_check_peer_type,
1169 };
1170 
1171 /*
1172  * Forward the excessive VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command to
1173  * vdpa device.
1174  *
1175  * Considering that QEMU cannot send the entire filter table to the
1176  * vdpa device, it should send the VIRTIO_NET_CTRL_RX_PROMISC CVQ
1177  * command to enable promiscuous mode to receive all packets,
1178  * according to VirtIO standard, "Since there are no guarantees,
1179  * it can use a hash filter or silently switch to allmulti or
1180  * promiscuous mode if it is given too many addresses.".
1181  *
1182  * Since QEMU ignores MAC addresses beyond `MAC_TABLE_ENTRIES` and
1183  * marks `n->mac_table.x_overflow` accordingly, it should have
1184  * the same effect on the device model to receive
1185  * (`MAC_TABLE_ENTRIES` + 1) or more non-multicast MAC addresses.
1186  * The same applies to multicast MAC addresses.
1187  *
1188  * Therefore, QEMU can provide the device model with a fake
1189  * VIRTIO_NET_CTRL_MAC_TABLE_SET command with (`MAC_TABLE_ENTRIES` + 1)
1190  * non-multicast MAC addresses and (`MAC_TABLE_ENTRIES` + 1) multicast
1191  * MAC addresses. This ensures that the device model marks
1192  * `n->mac_table.uni_overflow` and `n->mac_table.multi_overflow`,
1193  * allowing all packets to be received, which aligns with the
1194  * state of the vdpa device.
1195  */
1196 static int vhost_vdpa_net_excessive_mac_filter_cvq_add(VhostVDPAState *s,
1197                                                        VirtQueueElement *elem,
1198                                                        struct iovec *out,
1199                                                        const struct iovec *in)
1200 {
1201     struct virtio_net_ctrl_mac mac_data, *mac_ptr;
1202     struct virtio_net_ctrl_hdr *hdr_ptr;
1203     uint32_t cursor;
1204     ssize_t r;
1205     uint8_t on = 1;
1206 
1207     /* parse the non-multicast MAC address entries from CVQ command */
1208     cursor = sizeof(*hdr_ptr);
1209     r = iov_to_buf(elem->out_sg, elem->out_num, cursor,
1210                    &mac_data, sizeof(mac_data));
1211     if (unlikely(r != sizeof(mac_data))) {
1212         /*
1213          * If the CVQ command is invalid, we should simulate the vdpa device
1214          * to reject the VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command
1215          */
1216         *s->status = VIRTIO_NET_ERR;
1217         return sizeof(*s->status);
1218     }
1219     cursor += sizeof(mac_data) + le32_to_cpu(mac_data.entries) * ETH_ALEN;
1220 
1221     /* parse the multicast MAC address entries from CVQ command */
1222     r = iov_to_buf(elem->out_sg, elem->out_num, cursor,
1223                    &mac_data, sizeof(mac_data));
1224     if (r != sizeof(mac_data)) {
1225         /*
1226          * If the CVQ command is invalid, we should simulate the vdpa device
1227          * to reject the VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command
1228          */
1229         *s->status = VIRTIO_NET_ERR;
1230         return sizeof(*s->status);
1231     }
1232     cursor += sizeof(mac_data) + le32_to_cpu(mac_data.entries) * ETH_ALEN;
1233 
1234     /* validate the CVQ command */
1235     if (iov_size(elem->out_sg, elem->out_num) != cursor) {
1236         /*
1237          * If the CVQ command is invalid, we should simulate the vdpa device
1238          * to reject the VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command
1239          */
1240         *s->status = VIRTIO_NET_ERR;
1241         return sizeof(*s->status);
1242     }
1243 
1244     /*
1245      * According to VirtIO standard, "Since there are no guarantees,
1246      * it can use a hash filter or silently switch to allmulti or
1247      * promiscuous mode if it is given too many addresses.".
1248      *
1249      * Therefore, considering that QEMU is unable to send the entire
1250      * filter table to the vdpa device, it should send the
1251      * VIRTIO_NET_CTRL_RX_PROMISC CVQ command to enable promiscuous mode
1252      */
1253     hdr_ptr = out->iov_base;
1254     out->iov_len = sizeof(*hdr_ptr) + sizeof(on);
1255 
1256     hdr_ptr->class = VIRTIO_NET_CTRL_RX;
1257     hdr_ptr->cmd = VIRTIO_NET_CTRL_RX_PROMISC;
1258     iov_from_buf(out, 1, sizeof(*hdr_ptr), &on, sizeof(on));
1259     r = vhost_vdpa_net_cvq_add(s, out, 1, in, 1);
1260     if (unlikely(r < 0)) {
1261         return r;
1262     }
1263 
1264     /*
1265      * We can poll here since we've had BQL from the time
1266      * we sent the descriptor.
1267      */
1268     r = vhost_vdpa_net_svq_poll(s, 1);
1269     if (unlikely(r < sizeof(*s->status))) {
1270         return r;
1271     }
1272     if (*s->status != VIRTIO_NET_OK) {
1273         return sizeof(*s->status);
1274     }
1275 
1276     /*
1277      * QEMU should also send a fake VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ
1278      * command to the device model, including (`MAC_TABLE_ENTRIES` + 1)
1279      * non-multicast MAC addresses and (`MAC_TABLE_ENTRIES` + 1)
1280      * multicast MAC addresses.
1281      *
1282      * By doing so, the device model can mark `n->mac_table.uni_overflow`
1283      * and `n->mac_table.multi_overflow`, enabling all packets to be
1284      * received, which aligns with the state of the vdpa device.
1285      */
1286     cursor = 0;
1287     uint32_t fake_uni_entries = MAC_TABLE_ENTRIES + 1,
1288              fake_mul_entries = MAC_TABLE_ENTRIES + 1,
1289              fake_cvq_size = sizeof(struct virtio_net_ctrl_hdr) +
1290                              sizeof(mac_data) + fake_uni_entries * ETH_ALEN +
1291                              sizeof(mac_data) + fake_mul_entries * ETH_ALEN;
1292 
1293     assert(fake_cvq_size < vhost_vdpa_net_cvq_cmd_page_len());
1294     out->iov_len = fake_cvq_size;
1295 
1296     /* pack the header for fake CVQ command */
1297     hdr_ptr = out->iov_base + cursor;
1298     hdr_ptr->class = VIRTIO_NET_CTRL_MAC;
1299     hdr_ptr->cmd = VIRTIO_NET_CTRL_MAC_TABLE_SET;
1300     cursor += sizeof(*hdr_ptr);
1301 
1302     /*
1303      * Pack the non-multicast MAC addresses part for fake CVQ command.
1304      *
1305      * According to virtio_net_handle_mac(), QEMU doesn't verify the MAC
1306      * addresses provided in CVQ command. Therefore, only the entries
1307      * field need to be prepared in the CVQ command.
1308      */
1309     mac_ptr = out->iov_base + cursor;
1310     mac_ptr->entries = cpu_to_le32(fake_uni_entries);
1311     cursor += sizeof(*mac_ptr) + fake_uni_entries * ETH_ALEN;
1312 
1313     /*
1314      * Pack the multicast MAC addresses part for fake CVQ command.
1315      *
1316      * According to virtio_net_handle_mac(), QEMU doesn't verify the MAC
1317      * addresses provided in CVQ command. Therefore, only the entries
1318      * field need to be prepared in the CVQ command.
1319      */
1320     mac_ptr = out->iov_base + cursor;
1321     mac_ptr->entries = cpu_to_le32(fake_mul_entries);
1322 
1323     /*
1324      * Simulating QEMU poll a vdpa device used buffer
1325      * for VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command
1326      */
1327     return sizeof(*s->status);
1328 }
1329 
1330 /**
1331  * Validate and copy control virtqueue commands.
1332  *
1333  * Following QEMU guidelines, we offer a copy of the buffers to the device to
1334  * prevent TOCTOU bugs.
1335  */
1336 static int vhost_vdpa_net_handle_ctrl_avail(VhostShadowVirtqueue *svq,
1337                                             VirtQueueElement *elem,
1338                                             void *opaque)
1339 {
1340     VhostVDPAState *s = opaque;
1341     size_t in_len;
1342     const struct virtio_net_ctrl_hdr *ctrl;
1343     virtio_net_ctrl_ack status = VIRTIO_NET_ERR;
1344     /* Out buffer sent to both the vdpa device and the device model */
1345     struct iovec out = {
1346         .iov_base = s->cvq_cmd_out_buffer,
1347     };
1348     /* in buffer used for device model */
1349     const struct iovec model_in = {
1350         .iov_base = &status,
1351         .iov_len = sizeof(status),
1352     };
1353     /* in buffer used for vdpa device */
1354     const struct iovec vdpa_in = {
1355         .iov_base = s->status,
1356         .iov_len = sizeof(*s->status),
1357     };
1358     ssize_t dev_written = -EINVAL;
1359 
1360     out.iov_len = iov_to_buf(elem->out_sg, elem->out_num, 0,
1361                              s->cvq_cmd_out_buffer,
1362                              vhost_vdpa_net_cvq_cmd_page_len());
1363 
1364     ctrl = s->cvq_cmd_out_buffer;
1365     if (ctrl->class == VIRTIO_NET_CTRL_ANNOUNCE) {
1366         /*
1367          * Guest announce capability is emulated by qemu, so don't forward to
1368          * the device.
1369          */
1370         dev_written = sizeof(status);
1371         *s->status = VIRTIO_NET_OK;
1372     } else if (unlikely(ctrl->class == VIRTIO_NET_CTRL_MAC &&
1373                         ctrl->cmd == VIRTIO_NET_CTRL_MAC_TABLE_SET &&
1374                         iov_size(elem->out_sg, elem->out_num) > out.iov_len)) {
1375         /*
1376          * Due to the size limitation of the out buffer sent to the vdpa device,
1377          * which is determined by vhost_vdpa_net_cvq_cmd_page_len(), excessive
1378          * MAC addresses set by the driver for the filter table can cause
1379          * truncation of the CVQ command in QEMU. As a result, the vdpa device
1380          * rejects the flawed CVQ command.
1381          *
1382          * Therefore, QEMU must handle this situation instead of sending
1383          * the CVQ command directly.
1384          */
1385         dev_written = vhost_vdpa_net_excessive_mac_filter_cvq_add(s, elem,
1386                                                             &out, &vdpa_in);
1387         if (unlikely(dev_written < 0)) {
1388             goto out;
1389         }
1390     } else {
1391         ssize_t r;
1392         r = vhost_vdpa_net_cvq_add(s, &out, 1, &vdpa_in, 1);
1393         if (unlikely(r < 0)) {
1394             dev_written = r;
1395             goto out;
1396         }
1397 
1398         /*
1399          * We can poll here since we've had BQL from the time
1400          * we sent the descriptor.
1401          */
1402         dev_written = vhost_vdpa_net_svq_poll(s, 1);
1403     }
1404 
1405     if (unlikely(dev_written < sizeof(status))) {
1406         error_report("Insufficient written data (%zu)", dev_written);
1407         goto out;
1408     }
1409 
1410     if (*s->status != VIRTIO_NET_OK) {
1411         goto out;
1412     }
1413 
1414     status = VIRTIO_NET_ERR;
1415     virtio_net_handle_ctrl_iov(svq->vdev, &model_in, 1, &out, 1);
1416     if (status != VIRTIO_NET_OK) {
1417         error_report("Bad CVQ processing in model");
1418     }
1419 
1420 out:
1421     in_len = iov_from_buf(elem->in_sg, elem->in_num, 0, &status,
1422                           sizeof(status));
1423     if (unlikely(in_len < sizeof(status))) {
1424         error_report("Bad device CVQ written length");
1425     }
1426     vhost_svq_push_elem(svq, elem, MIN(in_len, sizeof(status)));
1427     /*
1428      * `elem` belongs to vhost_vdpa_net_handle_ctrl_avail() only when
1429      * the function successfully forwards the CVQ command, indicated
1430      * by a non-negative value of `dev_written`. Otherwise, it still
1431      * belongs to SVQ.
1432      * This function should only free the `elem` when it owns.
1433      */
1434     if (dev_written >= 0) {
1435         g_free(elem);
1436     }
1437     return dev_written < 0 ? dev_written : 0;
1438 }
1439 
1440 static const VhostShadowVirtqueueOps vhost_vdpa_net_svq_ops = {
1441     .avail_handler = vhost_vdpa_net_handle_ctrl_avail,
1442 };
1443 
1444 /**
1445  * Probe if CVQ is isolated
1446  *
1447  * @device_fd         The vdpa device fd
1448  * @features          Features offered by the device.
1449  * @cvq_index         The control vq pair index
1450  *
1451  * Returns <0 in case of failure, 0 if false and 1 if true.
1452  */
1453 static int vhost_vdpa_probe_cvq_isolation(int device_fd, uint64_t features,
1454                                           int cvq_index, Error **errp)
1455 {
1456     uint64_t backend_features;
1457     int64_t cvq_group;
1458     uint8_t status = VIRTIO_CONFIG_S_ACKNOWLEDGE |
1459                      VIRTIO_CONFIG_S_DRIVER;
1460     int r;
1461 
1462     ERRP_GUARD();
1463 
1464     r = ioctl(device_fd, VHOST_GET_BACKEND_FEATURES, &backend_features);
1465     if (unlikely(r < 0)) {
1466         error_setg_errno(errp, errno, "Cannot get vdpa backend_features");
1467         return r;
1468     }
1469 
1470     if (!(backend_features & BIT_ULL(VHOST_BACKEND_F_IOTLB_ASID))) {
1471         return 0;
1472     }
1473 
1474     r = ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status);
1475     if (unlikely(r)) {
1476         error_setg_errno(errp, -r, "Cannot set device status");
1477         goto out;
1478     }
1479 
1480     r = ioctl(device_fd, VHOST_SET_FEATURES, &features);
1481     if (unlikely(r)) {
1482         error_setg_errno(errp, -r, "Cannot set features");
1483         goto out;
1484     }
1485 
1486     status |= VIRTIO_CONFIG_S_FEATURES_OK;
1487     r = ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status);
1488     if (unlikely(r)) {
1489         error_setg_errno(errp, -r, "Cannot set device status");
1490         goto out;
1491     }
1492 
1493     cvq_group = vhost_vdpa_get_vring_group(device_fd, cvq_index, errp);
1494     if (unlikely(cvq_group < 0)) {
1495         if (cvq_group != -ENOTSUP) {
1496             r = cvq_group;
1497             goto out;
1498         }
1499 
1500         /*
1501          * The kernel report VHOST_BACKEND_F_IOTLB_ASID if the vdpa frontend
1502          * support ASID even if the parent driver does not.  The CVQ cannot be
1503          * isolated in this case.
1504          */
1505         error_free(*errp);
1506         *errp = NULL;
1507         r = 0;
1508         goto out;
1509     }
1510 
1511     for (int i = 0; i < cvq_index; ++i) {
1512         int64_t group = vhost_vdpa_get_vring_group(device_fd, i, errp);
1513         if (unlikely(group < 0)) {
1514             r = group;
1515             goto out;
1516         }
1517 
1518         if (group == (int64_t)cvq_group) {
1519             r = 0;
1520             goto out;
1521         }
1522     }
1523 
1524     r = 1;
1525 
1526 out:
1527     status = 0;
1528     ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status);
1529     return r;
1530 }
1531 
1532 static NetClientState *net_vhost_vdpa_init(NetClientState *peer,
1533                                        const char *device,
1534                                        const char *name,
1535                                        int vdpa_device_fd,
1536                                        int queue_pair_index,
1537                                        int nvqs,
1538                                        bool is_datapath,
1539                                        bool svq,
1540                                        struct vhost_vdpa_iova_range iova_range,
1541                                        uint64_t features,
1542                                        Error **errp)
1543 {
1544     NetClientState *nc = NULL;
1545     VhostVDPAState *s;
1546     int ret = 0;
1547     assert(name);
1548     int cvq_isolated = 0;
1549 
1550     if (is_datapath) {
1551         nc = qemu_new_net_client(&net_vhost_vdpa_info, peer, device,
1552                                  name);
1553     } else {
1554         cvq_isolated = vhost_vdpa_probe_cvq_isolation(vdpa_device_fd, features,
1555                                                       queue_pair_index * 2,
1556                                                       errp);
1557         if (unlikely(cvq_isolated < 0)) {
1558             return NULL;
1559         }
1560 
1561         nc = qemu_new_net_control_client(&net_vhost_vdpa_cvq_info, peer,
1562                                          device, name);
1563     }
1564     qemu_set_info_str(nc, TYPE_VHOST_VDPA);
1565     s = DO_UPCAST(VhostVDPAState, nc, nc);
1566 
1567     s->vhost_vdpa.device_fd = vdpa_device_fd;
1568     s->vhost_vdpa.index = queue_pair_index;
1569     s->always_svq = svq;
1570     s->migration_state.notify = NULL;
1571     s->vhost_vdpa.shadow_vqs_enabled = svq;
1572     s->vhost_vdpa.iova_range = iova_range;
1573     s->vhost_vdpa.shadow_data = svq;
1574     if (queue_pair_index == 0) {
1575         vhost_vdpa_net_valid_svq_features(features,
1576                                           &s->vhost_vdpa.migration_blocker);
1577     } else if (!is_datapath) {
1578         s->cvq_cmd_out_buffer = mmap(NULL, vhost_vdpa_net_cvq_cmd_page_len(),
1579                                      PROT_READ | PROT_WRITE,
1580                                      MAP_SHARED | MAP_ANONYMOUS, -1, 0);
1581         s->status = mmap(NULL, vhost_vdpa_net_cvq_cmd_page_len(),
1582                          PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANONYMOUS,
1583                          -1, 0);
1584 
1585         s->vhost_vdpa.shadow_vq_ops = &vhost_vdpa_net_svq_ops;
1586         s->vhost_vdpa.shadow_vq_ops_opaque = s;
1587         s->cvq_isolated = cvq_isolated;
1588     }
1589     ret = vhost_vdpa_add(nc, (void *)&s->vhost_vdpa, queue_pair_index, nvqs);
1590     if (ret) {
1591         qemu_del_net_client(nc);
1592         return NULL;
1593     }
1594     return nc;
1595 }
1596 
1597 static int vhost_vdpa_get_features(int fd, uint64_t *features, Error **errp)
1598 {
1599     int ret = ioctl(fd, VHOST_GET_FEATURES, features);
1600     if (unlikely(ret < 0)) {
1601         error_setg_errno(errp, errno,
1602                          "Fail to query features from vhost-vDPA device");
1603     }
1604     return ret;
1605 }
1606 
1607 static int vhost_vdpa_get_max_queue_pairs(int fd, uint64_t features,
1608                                           int *has_cvq, Error **errp)
1609 {
1610     unsigned long config_size = offsetof(struct vhost_vdpa_config, buf);
1611     g_autofree struct vhost_vdpa_config *config = NULL;
1612     __virtio16 *max_queue_pairs;
1613     int ret;
1614 
1615     if (features & (1 << VIRTIO_NET_F_CTRL_VQ)) {
1616         *has_cvq = 1;
1617     } else {
1618         *has_cvq = 0;
1619     }
1620 
1621     if (features & (1 << VIRTIO_NET_F_MQ)) {
1622         config = g_malloc0(config_size + sizeof(*max_queue_pairs));
1623         config->off = offsetof(struct virtio_net_config, max_virtqueue_pairs);
1624         config->len = sizeof(*max_queue_pairs);
1625 
1626         ret = ioctl(fd, VHOST_VDPA_GET_CONFIG, config);
1627         if (ret) {
1628             error_setg(errp, "Fail to get config from vhost-vDPA device");
1629             return -ret;
1630         }
1631 
1632         max_queue_pairs = (__virtio16 *)&config->buf;
1633 
1634         return lduw_le_p(max_queue_pairs);
1635     }
1636 
1637     return 1;
1638 }
1639 
1640 int net_init_vhost_vdpa(const Netdev *netdev, const char *name,
1641                         NetClientState *peer, Error **errp)
1642 {
1643     const NetdevVhostVDPAOptions *opts;
1644     uint64_t features;
1645     int vdpa_device_fd;
1646     g_autofree NetClientState **ncs = NULL;
1647     struct vhost_vdpa_iova_range iova_range;
1648     NetClientState *nc;
1649     int queue_pairs, r, i = 0, has_cvq = 0;
1650 
1651     assert(netdev->type == NET_CLIENT_DRIVER_VHOST_VDPA);
1652     opts = &netdev->u.vhost_vdpa;
1653     if (!opts->vhostdev && !opts->vhostfd) {
1654         error_setg(errp,
1655                    "vhost-vdpa: neither vhostdev= nor vhostfd= was specified");
1656         return -1;
1657     }
1658 
1659     if (opts->vhostdev && opts->vhostfd) {
1660         error_setg(errp,
1661                    "vhost-vdpa: vhostdev= and vhostfd= are mutually exclusive");
1662         return -1;
1663     }
1664 
1665     if (opts->vhostdev) {
1666         vdpa_device_fd = qemu_open(opts->vhostdev, O_RDWR, errp);
1667         if (vdpa_device_fd == -1) {
1668             return -errno;
1669         }
1670     } else {
1671         /* has_vhostfd */
1672         vdpa_device_fd = monitor_fd_param(monitor_cur(), opts->vhostfd, errp);
1673         if (vdpa_device_fd == -1) {
1674             error_prepend(errp, "vhost-vdpa: unable to parse vhostfd: ");
1675             return -1;
1676         }
1677     }
1678 
1679     r = vhost_vdpa_get_features(vdpa_device_fd, &features, errp);
1680     if (unlikely(r < 0)) {
1681         goto err;
1682     }
1683 
1684     queue_pairs = vhost_vdpa_get_max_queue_pairs(vdpa_device_fd, features,
1685                                                  &has_cvq, errp);
1686     if (queue_pairs < 0) {
1687         qemu_close(vdpa_device_fd);
1688         return queue_pairs;
1689     }
1690 
1691     r = vhost_vdpa_get_iova_range(vdpa_device_fd, &iova_range);
1692     if (unlikely(r < 0)) {
1693         error_setg(errp, "vhost-vdpa: get iova range failed: %s",
1694                    strerror(-r));
1695         goto err;
1696     }
1697 
1698     if (opts->x_svq && !vhost_vdpa_net_valid_svq_features(features, errp)) {
1699         goto err;
1700     }
1701 
1702     ncs = g_malloc0(sizeof(*ncs) * queue_pairs);
1703 
1704     for (i = 0; i < queue_pairs; i++) {
1705         ncs[i] = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name,
1706                                      vdpa_device_fd, i, 2, true, opts->x_svq,
1707                                      iova_range, features, errp);
1708         if (!ncs[i])
1709             goto err;
1710     }
1711 
1712     if (has_cvq) {
1713         nc = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name,
1714                                  vdpa_device_fd, i, 1, false,
1715                                  opts->x_svq, iova_range, features, errp);
1716         if (!nc)
1717             goto err;
1718     }
1719 
1720     return 0;
1721 
1722 err:
1723     if (i) {
1724         for (i--; i >= 0; i--) {
1725             qemu_del_net_client(ncs[i]);
1726         }
1727     }
1728 
1729     qemu_close(vdpa_device_fd);
1730 
1731     return -1;
1732 }
1733