xref: /openbmc/qemu/net/vhost-vdpa.c (revision 00ef422e9fbfef1fb40447b08826db0951d788dd)
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 "hw/virtio/vhost.h"
30 
31 /* Todo:need to add the multiqueue support here */
32 typedef struct VhostVDPAState {
33     NetClientState nc;
34     struct vhost_vdpa vhost_vdpa;
35     VHostNetState *vhost_net;
36 
37     /* Control commands shadow buffers */
38     void *cvq_cmd_out_buffer;
39     virtio_net_ctrl_ack *status;
40 
41     /* The device always have SVQ enabled */
42     bool always_svq;
43     bool started;
44 } VhostVDPAState;
45 
46 const int vdpa_feature_bits[] = {
47     VIRTIO_F_NOTIFY_ON_EMPTY,
48     VIRTIO_RING_F_INDIRECT_DESC,
49     VIRTIO_RING_F_EVENT_IDX,
50     VIRTIO_F_ANY_LAYOUT,
51     VIRTIO_F_VERSION_1,
52     VIRTIO_NET_F_CSUM,
53     VIRTIO_NET_F_GUEST_CSUM,
54     VIRTIO_NET_F_GSO,
55     VIRTIO_NET_F_GUEST_TSO4,
56     VIRTIO_NET_F_GUEST_TSO6,
57     VIRTIO_NET_F_GUEST_ECN,
58     VIRTIO_NET_F_GUEST_UFO,
59     VIRTIO_NET_F_HOST_TSO4,
60     VIRTIO_NET_F_HOST_TSO6,
61     VIRTIO_NET_F_HOST_ECN,
62     VIRTIO_NET_F_HOST_UFO,
63     VIRTIO_NET_F_MRG_RXBUF,
64     VIRTIO_NET_F_MTU,
65     VIRTIO_NET_F_CTRL_RX,
66     VIRTIO_NET_F_CTRL_RX_EXTRA,
67     VIRTIO_NET_F_CTRL_VLAN,
68     VIRTIO_NET_F_CTRL_MAC_ADDR,
69     VIRTIO_NET_F_RSS,
70     VIRTIO_NET_F_MQ,
71     VIRTIO_NET_F_CTRL_VQ,
72     VIRTIO_F_IOMMU_PLATFORM,
73     VIRTIO_F_RING_PACKED,
74     VIRTIO_F_RING_RESET,
75     VIRTIO_NET_F_RSS,
76     VIRTIO_NET_F_HASH_REPORT,
77     VIRTIO_NET_F_STATUS,
78     VHOST_INVALID_FEATURE_BIT
79 };
80 
81 /** Supported device specific feature bits with SVQ */
82 static const uint64_t vdpa_svq_device_features =
83     BIT_ULL(VIRTIO_NET_F_CSUM) |
84     BIT_ULL(VIRTIO_NET_F_GUEST_CSUM) |
85     BIT_ULL(VIRTIO_NET_F_MTU) |
86     BIT_ULL(VIRTIO_NET_F_MAC) |
87     BIT_ULL(VIRTIO_NET_F_GUEST_TSO4) |
88     BIT_ULL(VIRTIO_NET_F_GUEST_TSO6) |
89     BIT_ULL(VIRTIO_NET_F_GUEST_ECN) |
90     BIT_ULL(VIRTIO_NET_F_GUEST_UFO) |
91     BIT_ULL(VIRTIO_NET_F_HOST_TSO4) |
92     BIT_ULL(VIRTIO_NET_F_HOST_TSO6) |
93     BIT_ULL(VIRTIO_NET_F_HOST_ECN) |
94     BIT_ULL(VIRTIO_NET_F_HOST_UFO) |
95     BIT_ULL(VIRTIO_NET_F_MRG_RXBUF) |
96     BIT_ULL(VIRTIO_NET_F_STATUS) |
97     BIT_ULL(VIRTIO_NET_F_CTRL_VQ) |
98     BIT_ULL(VIRTIO_NET_F_MQ) |
99     BIT_ULL(VIRTIO_F_ANY_LAYOUT) |
100     BIT_ULL(VIRTIO_NET_F_CTRL_MAC_ADDR) |
101     BIT_ULL(VIRTIO_NET_F_RSC_EXT) |
102     BIT_ULL(VIRTIO_NET_F_STANDBY);
103 
104 #define VHOST_VDPA_NET_CVQ_ASID 1
105 
106 VHostNetState *vhost_vdpa_get_vhost_net(NetClientState *nc)
107 {
108     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
109     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
110     return s->vhost_net;
111 }
112 
113 static bool vhost_vdpa_net_valid_svq_features(uint64_t features, Error **errp)
114 {
115     uint64_t invalid_dev_features =
116         features & ~vdpa_svq_device_features &
117         /* Transport are all accepted at this point */
118         ~MAKE_64BIT_MASK(VIRTIO_TRANSPORT_F_START,
119                          VIRTIO_TRANSPORT_F_END - VIRTIO_TRANSPORT_F_START);
120 
121     if (invalid_dev_features) {
122         error_setg(errp, "vdpa svq does not work with features 0x%" PRIx64,
123                    invalid_dev_features);
124         return false;
125     }
126 
127     return vhost_svq_valid_features(features, errp);
128 }
129 
130 static int vhost_vdpa_net_check_device_id(struct vhost_net *net)
131 {
132     uint32_t device_id;
133     int ret;
134     struct vhost_dev *hdev;
135 
136     hdev = (struct vhost_dev *)&net->dev;
137     ret = hdev->vhost_ops->vhost_get_device_id(hdev, &device_id);
138     if (device_id != VIRTIO_ID_NET) {
139         return -ENOTSUP;
140     }
141     return ret;
142 }
143 
144 static int vhost_vdpa_add(NetClientState *ncs, void *be,
145                           int queue_pair_index, int nvqs)
146 {
147     VhostNetOptions options;
148     struct vhost_net *net = NULL;
149     VhostVDPAState *s;
150     int ret;
151 
152     options.backend_type = VHOST_BACKEND_TYPE_VDPA;
153     assert(ncs->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
154     s = DO_UPCAST(VhostVDPAState, nc, ncs);
155     options.net_backend = ncs;
156     options.opaque      = be;
157     options.busyloop_timeout = 0;
158     options.nvqs = nvqs;
159 
160     net = vhost_net_init(&options);
161     if (!net) {
162         error_report("failed to init vhost_net for queue");
163         goto err_init;
164     }
165     s->vhost_net = net;
166     ret = vhost_vdpa_net_check_device_id(net);
167     if (ret) {
168         goto err_check;
169     }
170     return 0;
171 err_check:
172     vhost_net_cleanup(net);
173     g_free(net);
174 err_init:
175     return -1;
176 }
177 
178 static void vhost_vdpa_cleanup(NetClientState *nc)
179 {
180     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
181 
182     qemu_vfree(s->cvq_cmd_out_buffer);
183     qemu_vfree(s->status);
184     if (s->vhost_net) {
185         vhost_net_cleanup(s->vhost_net);
186         g_free(s->vhost_net);
187         s->vhost_net = NULL;
188     }
189      if (s->vhost_vdpa.device_fd >= 0) {
190         qemu_close(s->vhost_vdpa.device_fd);
191         s->vhost_vdpa.device_fd = -1;
192     }
193 }
194 
195 static bool vhost_vdpa_has_vnet_hdr(NetClientState *nc)
196 {
197     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
198 
199     return true;
200 }
201 
202 static bool vhost_vdpa_has_ufo(NetClientState *nc)
203 {
204     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
205     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
206     uint64_t features = 0;
207     features |= (1ULL << VIRTIO_NET_F_HOST_UFO);
208     features = vhost_net_get_features(s->vhost_net, features);
209     return !!(features & (1ULL << VIRTIO_NET_F_HOST_UFO));
210 
211 }
212 
213 static bool vhost_vdpa_check_peer_type(NetClientState *nc, ObjectClass *oc,
214                                        Error **errp)
215 {
216     const char *driver = object_class_get_name(oc);
217 
218     if (!g_str_has_prefix(driver, "virtio-net-")) {
219         error_setg(errp, "vhost-vdpa requires frontend driver virtio-net-*");
220         return false;
221     }
222 
223     return true;
224 }
225 
226 /** Dummy receive in case qemu falls back to userland tap networking */
227 static ssize_t vhost_vdpa_receive(NetClientState *nc, const uint8_t *buf,
228                                   size_t size)
229 {
230     return size;
231 }
232 
233 /** From any vdpa net client, get the netclient of the first queue pair */
234 static VhostVDPAState *vhost_vdpa_net_first_nc_vdpa(VhostVDPAState *s)
235 {
236     NICState *nic = qemu_get_nic(s->nc.peer);
237     NetClientState *nc0 = qemu_get_peer(nic->ncs, 0);
238 
239     return DO_UPCAST(VhostVDPAState, nc, nc0);
240 }
241 
242 static void vhost_vdpa_net_data_start_first(VhostVDPAState *s)
243 {
244     struct vhost_vdpa *v = &s->vhost_vdpa;
245 
246     if (v->shadow_vqs_enabled) {
247         v->iova_tree = vhost_iova_tree_new(v->iova_range.first,
248                                            v->iova_range.last);
249     }
250 }
251 
252 static int vhost_vdpa_net_data_start(NetClientState *nc)
253 {
254     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
255     struct vhost_vdpa *v = &s->vhost_vdpa;
256 
257     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
258 
259     if (v->index == 0) {
260         vhost_vdpa_net_data_start_first(s);
261         return 0;
262     }
263 
264     if (v->shadow_vqs_enabled) {
265         VhostVDPAState *s0 = vhost_vdpa_net_first_nc_vdpa(s);
266         v->iova_tree = s0->vhost_vdpa.iova_tree;
267     }
268 
269     return 0;
270 }
271 
272 static void vhost_vdpa_net_client_stop(NetClientState *nc)
273 {
274     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
275     struct vhost_dev *dev;
276 
277     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
278 
279     dev = s->vhost_vdpa.dev;
280     if (dev->vq_index + dev->nvqs == dev->vq_index_end) {
281         g_clear_pointer(&s->vhost_vdpa.iova_tree, vhost_iova_tree_delete);
282     }
283 }
284 
285 static NetClientInfo net_vhost_vdpa_info = {
286         .type = NET_CLIENT_DRIVER_VHOST_VDPA,
287         .size = sizeof(VhostVDPAState),
288         .receive = vhost_vdpa_receive,
289         .start = vhost_vdpa_net_data_start,
290         .stop = vhost_vdpa_net_client_stop,
291         .cleanup = vhost_vdpa_cleanup,
292         .has_vnet_hdr = vhost_vdpa_has_vnet_hdr,
293         .has_ufo = vhost_vdpa_has_ufo,
294         .check_peer_type = vhost_vdpa_check_peer_type,
295 };
296 
297 static int64_t vhost_vdpa_get_vring_group(int device_fd, unsigned vq_index)
298 {
299     struct vhost_vring_state state = {
300         .index = vq_index,
301     };
302     int r = ioctl(device_fd, VHOST_VDPA_GET_VRING_GROUP, &state);
303 
304     if (unlikely(r < 0)) {
305         error_report("Cannot get VQ %u group: %s", vq_index,
306                      g_strerror(errno));
307         return r;
308     }
309 
310     return state.num;
311 }
312 
313 static int vhost_vdpa_set_address_space_id(struct vhost_vdpa *v,
314                                            unsigned vq_group,
315                                            unsigned asid_num)
316 {
317     struct vhost_vring_state asid = {
318         .index = vq_group,
319         .num = asid_num,
320     };
321     int r;
322 
323     r = ioctl(v->device_fd, VHOST_VDPA_SET_GROUP_ASID, &asid);
324     if (unlikely(r < 0)) {
325         error_report("Can't set vq group %u asid %u, errno=%d (%s)",
326                      asid.index, asid.num, errno, g_strerror(errno));
327     }
328     return r;
329 }
330 
331 static void vhost_vdpa_cvq_unmap_buf(struct vhost_vdpa *v, void *addr)
332 {
333     VhostIOVATree *tree = v->iova_tree;
334     DMAMap needle = {
335         /*
336          * No need to specify size or to look for more translations since
337          * this contiguous chunk was allocated by us.
338          */
339         .translated_addr = (hwaddr)(uintptr_t)addr,
340     };
341     const DMAMap *map = vhost_iova_tree_find_iova(tree, &needle);
342     int r;
343 
344     if (unlikely(!map)) {
345         error_report("Cannot locate expected map");
346         return;
347     }
348 
349     r = vhost_vdpa_dma_unmap(v, v->address_space_id, map->iova, map->size + 1);
350     if (unlikely(r != 0)) {
351         error_report("Device cannot unmap: %s(%d)", g_strerror(r), r);
352     }
353 
354     vhost_iova_tree_remove(tree, *map);
355 }
356 
357 static size_t vhost_vdpa_net_cvq_cmd_len(void)
358 {
359     /*
360      * MAC_TABLE_SET is the ctrl command that produces the longer out buffer.
361      * In buffer is always 1 byte, so it should fit here
362      */
363     return sizeof(struct virtio_net_ctrl_hdr) +
364            2 * sizeof(struct virtio_net_ctrl_mac) +
365            MAC_TABLE_ENTRIES * ETH_ALEN;
366 }
367 
368 static size_t vhost_vdpa_net_cvq_cmd_page_len(void)
369 {
370     return ROUND_UP(vhost_vdpa_net_cvq_cmd_len(), qemu_real_host_page_size());
371 }
372 
373 /** Map CVQ buffer. */
374 static int vhost_vdpa_cvq_map_buf(struct vhost_vdpa *v, void *buf, size_t size,
375                                   bool write)
376 {
377     DMAMap map = {};
378     int r;
379 
380     map.translated_addr = (hwaddr)(uintptr_t)buf;
381     map.size = size - 1;
382     map.perm = write ? IOMMU_RW : IOMMU_RO,
383     r = vhost_iova_tree_map_alloc(v->iova_tree, &map);
384     if (unlikely(r != IOVA_OK)) {
385         error_report("Cannot map injected element");
386         return r;
387     }
388 
389     r = vhost_vdpa_dma_map(v, v->address_space_id, map.iova,
390                            vhost_vdpa_net_cvq_cmd_page_len(), buf, !write);
391     if (unlikely(r < 0)) {
392         goto dma_map_err;
393     }
394 
395     return 0;
396 
397 dma_map_err:
398     vhost_iova_tree_remove(v->iova_tree, map);
399     return r;
400 }
401 
402 static int vhost_vdpa_net_cvq_start(NetClientState *nc)
403 {
404     VhostVDPAState *s, *s0;
405     struct vhost_vdpa *v;
406     uint64_t backend_features;
407     int64_t cvq_group;
408     int cvq_index, r;
409 
410     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
411 
412     s = DO_UPCAST(VhostVDPAState, nc, nc);
413     v = &s->vhost_vdpa;
414 
415     v->shadow_data = s->always_svq;
416     v->shadow_vqs_enabled = s->always_svq;
417     s->vhost_vdpa.address_space_id = VHOST_VDPA_GUEST_PA_ASID;
418 
419     if (s->always_svq) {
420         /* SVQ is already configured for all virtqueues */
421         goto out;
422     }
423 
424     /*
425      * If we early return in these cases SVQ will not be enabled. The migration
426      * will be blocked as long as vhost-vdpa backends will not offer _F_LOG.
427      *
428      * Calling VHOST_GET_BACKEND_FEATURES as they are not available in v->dev
429      * yet.
430      */
431     r = ioctl(v->device_fd, VHOST_GET_BACKEND_FEATURES, &backend_features);
432     if (unlikely(r < 0)) {
433         error_report("Cannot get vdpa backend_features: %s(%d)",
434             g_strerror(errno), errno);
435         return -1;
436     }
437     if (!(backend_features & BIT_ULL(VHOST_BACKEND_F_IOTLB_ASID)) ||
438         !vhost_vdpa_net_valid_svq_features(v->dev->features, NULL)) {
439         return 0;
440     }
441 
442     /*
443      * Check if all the virtqueues of the virtio device are in a different vq
444      * than the last vq. VQ group of last group passed in cvq_group.
445      */
446     cvq_index = v->dev->vq_index_end - 1;
447     cvq_group = vhost_vdpa_get_vring_group(v->device_fd, cvq_index);
448     if (unlikely(cvq_group < 0)) {
449         return cvq_group;
450     }
451     for (int i = 0; i < cvq_index; ++i) {
452         int64_t group = vhost_vdpa_get_vring_group(v->device_fd, i);
453 
454         if (unlikely(group < 0)) {
455             return group;
456         }
457 
458         if (group == cvq_group) {
459             return 0;
460         }
461     }
462 
463     r = vhost_vdpa_set_address_space_id(v, cvq_group, VHOST_VDPA_NET_CVQ_ASID);
464     if (unlikely(r < 0)) {
465         return r;
466     }
467 
468     v->shadow_vqs_enabled = true;
469     s->vhost_vdpa.address_space_id = VHOST_VDPA_NET_CVQ_ASID;
470 
471 out:
472     if (!s->vhost_vdpa.shadow_vqs_enabled) {
473         return 0;
474     }
475 
476     s0 = vhost_vdpa_net_first_nc_vdpa(s);
477     if (s0->vhost_vdpa.iova_tree) {
478         /*
479          * SVQ is already configured for all virtqueues.  Reuse IOVA tree for
480          * simplicity, whether CVQ shares ASID with guest or not, because:
481          * - Memory listener need access to guest's memory addresses allocated
482          *   in the IOVA tree.
483          * - There should be plenty of IOVA address space for both ASID not to
484          *   worry about collisions between them.  Guest's translations are
485          *   still validated with virtio virtqueue_pop so there is no risk for
486          *   the guest to access memory that it shouldn't.
487          *
488          * To allocate a iova tree per ASID is doable but it complicates the
489          * code and it is not worth it for the moment.
490          */
491         v->iova_tree = s0->vhost_vdpa.iova_tree;
492     } else {
493         v->iova_tree = vhost_iova_tree_new(v->iova_range.first,
494                                            v->iova_range.last);
495     }
496 
497     r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer,
498                                vhost_vdpa_net_cvq_cmd_page_len(), false);
499     if (unlikely(r < 0)) {
500         return r;
501     }
502 
503     r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->status,
504                                vhost_vdpa_net_cvq_cmd_page_len(), true);
505     if (unlikely(r < 0)) {
506         vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer);
507     }
508 
509     return r;
510 }
511 
512 static void vhost_vdpa_net_cvq_stop(NetClientState *nc)
513 {
514     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
515 
516     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
517 
518     if (s->vhost_vdpa.shadow_vqs_enabled) {
519         vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer);
520         vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->status);
521     }
522 
523     vhost_vdpa_net_client_stop(nc);
524 }
525 
526 static ssize_t vhost_vdpa_net_cvq_add(VhostVDPAState *s, size_t out_len,
527                                       size_t in_len)
528 {
529     /* Buffers for the device */
530     const struct iovec out = {
531         .iov_base = s->cvq_cmd_out_buffer,
532         .iov_len = out_len,
533     };
534     const struct iovec in = {
535         .iov_base = s->status,
536         .iov_len = sizeof(virtio_net_ctrl_ack),
537     };
538     VhostShadowVirtqueue *svq = g_ptr_array_index(s->vhost_vdpa.shadow_vqs, 0);
539     int r;
540 
541     r = vhost_svq_add(svq, &out, 1, &in, 1, NULL);
542     if (unlikely(r != 0)) {
543         if (unlikely(r == -ENOSPC)) {
544             qemu_log_mask(LOG_GUEST_ERROR, "%s: No space on device queue\n",
545                           __func__);
546         }
547         return r;
548     }
549 
550     /*
551      * We can poll here since we've had BQL from the time we sent the
552      * descriptor. Also, we need to take the answer before SVQ pulls by itself,
553      * when BQL is released
554      */
555     return vhost_svq_poll(svq);
556 }
557 
558 static ssize_t vhost_vdpa_net_load_cmd(VhostVDPAState *s, uint8_t class,
559                                        uint8_t cmd, const void *data,
560                                        size_t data_size)
561 {
562     const struct virtio_net_ctrl_hdr ctrl = {
563         .class = class,
564         .cmd = cmd,
565     };
566 
567     assert(data_size < vhost_vdpa_net_cvq_cmd_page_len() - sizeof(ctrl));
568 
569     memcpy(s->cvq_cmd_out_buffer, &ctrl, sizeof(ctrl));
570     memcpy(s->cvq_cmd_out_buffer + sizeof(ctrl), data, data_size);
571 
572     return vhost_vdpa_net_cvq_add(s, sizeof(ctrl) + data_size,
573                                   sizeof(virtio_net_ctrl_ack));
574 }
575 
576 static int vhost_vdpa_net_load_mac(VhostVDPAState *s, const VirtIONet *n)
577 {
578     uint64_t features = n->parent_obj.guest_features;
579     if (features & BIT_ULL(VIRTIO_NET_F_CTRL_MAC_ADDR)) {
580         ssize_t dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_MAC,
581                                                   VIRTIO_NET_CTRL_MAC_ADDR_SET,
582                                                   n->mac, sizeof(n->mac));
583         if (unlikely(dev_written < 0)) {
584             return dev_written;
585         }
586 
587         return *s->status != VIRTIO_NET_OK;
588     }
589 
590     return 0;
591 }
592 
593 static int vhost_vdpa_net_load_mq(VhostVDPAState *s,
594                                   const VirtIONet *n)
595 {
596     struct virtio_net_ctrl_mq mq;
597     uint64_t features = n->parent_obj.guest_features;
598     ssize_t dev_written;
599 
600     if (!(features & BIT_ULL(VIRTIO_NET_F_MQ))) {
601         return 0;
602     }
603 
604     mq.virtqueue_pairs = cpu_to_le16(n->curr_queue_pairs);
605     dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_MQ,
606                                           VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, &mq,
607                                           sizeof(mq));
608     if (unlikely(dev_written < 0)) {
609         return dev_written;
610     }
611 
612     return *s->status != VIRTIO_NET_OK;
613 }
614 
615 static int vhost_vdpa_net_load(NetClientState *nc)
616 {
617     VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc);
618     struct vhost_vdpa *v = &s->vhost_vdpa;
619     const VirtIONet *n;
620     int r;
621 
622     assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA);
623 
624     if (!v->shadow_vqs_enabled) {
625         return 0;
626     }
627 
628     n = VIRTIO_NET(v->dev->vdev);
629     r = vhost_vdpa_net_load_mac(s, n);
630     if (unlikely(r < 0)) {
631         return r;
632     }
633     r = vhost_vdpa_net_load_mq(s, n);
634     if (unlikely(r)) {
635         return r;
636     }
637 
638     return 0;
639 }
640 
641 static NetClientInfo net_vhost_vdpa_cvq_info = {
642     .type = NET_CLIENT_DRIVER_VHOST_VDPA,
643     .size = sizeof(VhostVDPAState),
644     .receive = vhost_vdpa_receive,
645     .start = vhost_vdpa_net_cvq_start,
646     .load = vhost_vdpa_net_load,
647     .stop = vhost_vdpa_net_cvq_stop,
648     .cleanup = vhost_vdpa_cleanup,
649     .has_vnet_hdr = vhost_vdpa_has_vnet_hdr,
650     .has_ufo = vhost_vdpa_has_ufo,
651     .check_peer_type = vhost_vdpa_check_peer_type,
652 };
653 
654 /**
655  * Validate and copy control virtqueue commands.
656  *
657  * Following QEMU guidelines, we offer a copy of the buffers to the device to
658  * prevent TOCTOU bugs.
659  */
660 static int vhost_vdpa_net_handle_ctrl_avail(VhostShadowVirtqueue *svq,
661                                             VirtQueueElement *elem,
662                                             void *opaque)
663 {
664     VhostVDPAState *s = opaque;
665     size_t in_len;
666     virtio_net_ctrl_ack status = VIRTIO_NET_ERR;
667     /* Out buffer sent to both the vdpa device and the device model */
668     struct iovec out = {
669         .iov_base = s->cvq_cmd_out_buffer,
670     };
671     /* in buffer used for device model */
672     const struct iovec in = {
673         .iov_base = &status,
674         .iov_len = sizeof(status),
675     };
676     ssize_t dev_written = -EINVAL;
677 
678     out.iov_len = iov_to_buf(elem->out_sg, elem->out_num, 0,
679                              s->cvq_cmd_out_buffer,
680                              vhost_vdpa_net_cvq_cmd_len());
681     if (*(uint8_t *)s->cvq_cmd_out_buffer == VIRTIO_NET_CTRL_ANNOUNCE) {
682         /*
683          * Guest announce capability is emulated by qemu, so don't forward to
684          * the device.
685          */
686         dev_written = sizeof(status);
687         *s->status = VIRTIO_NET_OK;
688     } else {
689         dev_written = vhost_vdpa_net_cvq_add(s, out.iov_len, sizeof(status));
690         if (unlikely(dev_written < 0)) {
691             goto out;
692         }
693     }
694 
695     if (unlikely(dev_written < sizeof(status))) {
696         error_report("Insufficient written data (%zu)", dev_written);
697         goto out;
698     }
699 
700     if (*s->status != VIRTIO_NET_OK) {
701         return VIRTIO_NET_ERR;
702     }
703 
704     status = VIRTIO_NET_ERR;
705     virtio_net_handle_ctrl_iov(svq->vdev, &in, 1, &out, 1);
706     if (status != VIRTIO_NET_OK) {
707         error_report("Bad CVQ processing in model");
708     }
709 
710 out:
711     in_len = iov_from_buf(elem->in_sg, elem->in_num, 0, &status,
712                           sizeof(status));
713     if (unlikely(in_len < sizeof(status))) {
714         error_report("Bad device CVQ written length");
715     }
716     vhost_svq_push_elem(svq, elem, MIN(in_len, sizeof(status)));
717     g_free(elem);
718     return dev_written < 0 ? dev_written : 0;
719 }
720 
721 static const VhostShadowVirtqueueOps vhost_vdpa_net_svq_ops = {
722     .avail_handler = vhost_vdpa_net_handle_ctrl_avail,
723 };
724 
725 static NetClientState *net_vhost_vdpa_init(NetClientState *peer,
726                                        const char *device,
727                                        const char *name,
728                                        int vdpa_device_fd,
729                                        int queue_pair_index,
730                                        int nvqs,
731                                        bool is_datapath,
732                                        bool svq,
733                                        struct vhost_vdpa_iova_range iova_range)
734 {
735     NetClientState *nc = NULL;
736     VhostVDPAState *s;
737     int ret = 0;
738     assert(name);
739     if (is_datapath) {
740         nc = qemu_new_net_client(&net_vhost_vdpa_info, peer, device,
741                                  name);
742     } else {
743         nc = qemu_new_net_control_client(&net_vhost_vdpa_cvq_info, peer,
744                                          device, name);
745     }
746     qemu_set_info_str(nc, TYPE_VHOST_VDPA);
747     s = DO_UPCAST(VhostVDPAState, nc, nc);
748 
749     s->vhost_vdpa.device_fd = vdpa_device_fd;
750     s->vhost_vdpa.index = queue_pair_index;
751     s->always_svq = svq;
752     s->vhost_vdpa.shadow_vqs_enabled = svq;
753     s->vhost_vdpa.iova_range = iova_range;
754     s->vhost_vdpa.shadow_data = svq;
755     if (!is_datapath) {
756         s->cvq_cmd_out_buffer = qemu_memalign(qemu_real_host_page_size(),
757                                             vhost_vdpa_net_cvq_cmd_page_len());
758         memset(s->cvq_cmd_out_buffer, 0, vhost_vdpa_net_cvq_cmd_page_len());
759         s->status = qemu_memalign(qemu_real_host_page_size(),
760                                   vhost_vdpa_net_cvq_cmd_page_len());
761         memset(s->status, 0, vhost_vdpa_net_cvq_cmd_page_len());
762 
763         s->vhost_vdpa.shadow_vq_ops = &vhost_vdpa_net_svq_ops;
764         s->vhost_vdpa.shadow_vq_ops_opaque = s;
765     }
766     ret = vhost_vdpa_add(nc, (void *)&s->vhost_vdpa, queue_pair_index, nvqs);
767     if (ret) {
768         qemu_del_net_client(nc);
769         return NULL;
770     }
771     return nc;
772 }
773 
774 static int vhost_vdpa_get_features(int fd, uint64_t *features, Error **errp)
775 {
776     int ret = ioctl(fd, VHOST_GET_FEATURES, features);
777     if (unlikely(ret < 0)) {
778         error_setg_errno(errp, errno,
779                          "Fail to query features from vhost-vDPA device");
780     }
781     return ret;
782 }
783 
784 static int vhost_vdpa_get_max_queue_pairs(int fd, uint64_t features,
785                                           int *has_cvq, Error **errp)
786 {
787     unsigned long config_size = offsetof(struct vhost_vdpa_config, buf);
788     g_autofree struct vhost_vdpa_config *config = NULL;
789     __virtio16 *max_queue_pairs;
790     int ret;
791 
792     if (features & (1 << VIRTIO_NET_F_CTRL_VQ)) {
793         *has_cvq = 1;
794     } else {
795         *has_cvq = 0;
796     }
797 
798     if (features & (1 << VIRTIO_NET_F_MQ)) {
799         config = g_malloc0(config_size + sizeof(*max_queue_pairs));
800         config->off = offsetof(struct virtio_net_config, max_virtqueue_pairs);
801         config->len = sizeof(*max_queue_pairs);
802 
803         ret = ioctl(fd, VHOST_VDPA_GET_CONFIG, config);
804         if (ret) {
805             error_setg(errp, "Fail to get config from vhost-vDPA device");
806             return -ret;
807         }
808 
809         max_queue_pairs = (__virtio16 *)&config->buf;
810 
811         return lduw_le_p(max_queue_pairs);
812     }
813 
814     return 1;
815 }
816 
817 int net_init_vhost_vdpa(const Netdev *netdev, const char *name,
818                         NetClientState *peer, Error **errp)
819 {
820     const NetdevVhostVDPAOptions *opts;
821     uint64_t features;
822     int vdpa_device_fd;
823     g_autofree NetClientState **ncs = NULL;
824     struct vhost_vdpa_iova_range iova_range;
825     NetClientState *nc;
826     int queue_pairs, r, i = 0, has_cvq = 0;
827 
828     assert(netdev->type == NET_CLIENT_DRIVER_VHOST_VDPA);
829     opts = &netdev->u.vhost_vdpa;
830     if (!opts->vhostdev && !opts->vhostfd) {
831         error_setg(errp,
832                    "vhost-vdpa: neither vhostdev= nor vhostfd= was specified");
833         return -1;
834     }
835 
836     if (opts->vhostdev && opts->vhostfd) {
837         error_setg(errp,
838                    "vhost-vdpa: vhostdev= and vhostfd= are mutually exclusive");
839         return -1;
840     }
841 
842     if (opts->vhostdev) {
843         vdpa_device_fd = qemu_open(opts->vhostdev, O_RDWR, errp);
844         if (vdpa_device_fd == -1) {
845             return -errno;
846         }
847     } else {
848         /* has_vhostfd */
849         vdpa_device_fd = monitor_fd_param(monitor_cur(), opts->vhostfd, errp);
850         if (vdpa_device_fd == -1) {
851             error_prepend(errp, "vhost-vdpa: unable to parse vhostfd: ");
852             return -1;
853         }
854     }
855 
856     r = vhost_vdpa_get_features(vdpa_device_fd, &features, errp);
857     if (unlikely(r < 0)) {
858         goto err;
859     }
860 
861     queue_pairs = vhost_vdpa_get_max_queue_pairs(vdpa_device_fd, features,
862                                                  &has_cvq, errp);
863     if (queue_pairs < 0) {
864         qemu_close(vdpa_device_fd);
865         return queue_pairs;
866     }
867 
868     r = vhost_vdpa_get_iova_range(vdpa_device_fd, &iova_range);
869     if (unlikely(r < 0)) {
870         error_setg(errp, "vhost-vdpa: get iova range failed: %s",
871                    strerror(-r));
872         goto err;
873     }
874 
875     if (opts->x_svq && !vhost_vdpa_net_valid_svq_features(features, errp)) {
876         goto err;
877     }
878 
879     ncs = g_malloc0(sizeof(*ncs) * queue_pairs);
880 
881     for (i = 0; i < queue_pairs; i++) {
882         ncs[i] = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name,
883                                      vdpa_device_fd, i, 2, true, opts->x_svq,
884                                      iova_range);
885         if (!ncs[i])
886             goto err;
887     }
888 
889     if (has_cvq) {
890         nc = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name,
891                                  vdpa_device_fd, i, 1, false,
892                                  opts->x_svq, iova_range);
893         if (!nc)
894             goto err;
895     }
896 
897     return 0;
898 
899 err:
900     if (i) {
901         for (i--; i >= 0; i--) {
902             qemu_del_net_client(ncs[i]);
903         }
904     }
905 
906     qemu_close(vdpa_device_fd);
907 
908     return -1;
909 }
910