]>
Commit | Line | Data |
---|---|---|
1e0a84ea CL |
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" | |
bd907ae4 | 14 | #include "hw/virtio/virtio-net.h" |
1e0a84ea CL |
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" | |
bd907ae4 EP |
20 | #include "qemu/log.h" |
21 | #include "qemu/memalign.h" | |
1e0a84ea CL |
22 | #include "qemu/option.h" |
23 | #include "qapi/error.h" | |
40237840 | 24 | #include <linux/vhost.h> |
1e0a84ea CL |
25 | #include <sys/ioctl.h> |
26 | #include <err.h> | |
27 | #include "standard-headers/linux/virtio_net.h" | |
28 | #include "monitor/monitor.h" | |
69498430 EP |
29 | #include "migration/migration.h" |
30 | #include "migration/misc.h" | |
1e0a84ea CL |
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; | |
69498430 | 37 | Notifier migration_state; |
1e0a84ea | 38 | VHostNetState *vhost_net; |
2df4dd31 EP |
39 | |
40 | /* Control commands shadow buffers */ | |
17fb889f EP |
41 | void *cvq_cmd_out_buffer; |
42 | virtio_net_ctrl_ack *status; | |
43 | ||
7f211a28 EP |
44 | /* The device always have SVQ enabled */ |
45 | bool always_svq; | |
152128d6 EP |
46 | |
47 | /* The device can isolate CVQ in its own ASID */ | |
48 | bool cvq_isolated; | |
49 | ||
1e0a84ea CL |
50 | bool started; |
51 | } VhostVDPAState; | |
52 | ||
2875a0ca HJ |
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 | */ | |
1e0a84ea | 58 | const int vdpa_feature_bits[] = { |
1e0a84ea | 59 | VIRTIO_F_ANY_LAYOUT, |
2875a0ca HJ |
60 | VIRTIO_F_IOMMU_PLATFORM, |
61 | VIRTIO_F_NOTIFY_ON_EMPTY, | |
62 | VIRTIO_F_RING_PACKED, | |
63 | VIRTIO_F_RING_RESET, | |
1e0a84ea CL |
64 | VIRTIO_F_VERSION_1, |
65 | VIRTIO_NET_F_CSUM, | |
51e84244 | 66 | VIRTIO_NET_F_CTRL_GUEST_OFFLOADS, |
2875a0ca HJ |
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, | |
1e0a84ea | 72 | VIRTIO_NET_F_GSO, |
2875a0ca HJ |
73 | VIRTIO_NET_F_GUEST_CSUM, |
74 | VIRTIO_NET_F_GUEST_ECN, | |
1e0a84ea CL |
75 | VIRTIO_NET_F_GUEST_TSO4, |
76 | VIRTIO_NET_F_GUEST_TSO6, | |
1e0a84ea | 77 | VIRTIO_NET_F_GUEST_UFO, |
9da16849 AM |
78 | VIRTIO_NET_F_GUEST_USO4, |
79 | VIRTIO_NET_F_GUEST_USO6, | |
2875a0ca HJ |
80 | VIRTIO_NET_F_HASH_REPORT, |
81 | VIRTIO_NET_F_HOST_ECN, | |
1e0a84ea CL |
82 | VIRTIO_NET_F_HOST_TSO4, |
83 | VIRTIO_NET_F_HOST_TSO6, | |
1e0a84ea | 84 | VIRTIO_NET_F_HOST_UFO, |
9da16849 | 85 | VIRTIO_NET_F_HOST_USO, |
2875a0ca | 86 | VIRTIO_NET_F_MQ, |
1e0a84ea CL |
87 | VIRTIO_NET_F_MRG_RXBUF, |
88 | VIRTIO_NET_F_MTU, | |
0145c393 | 89 | VIRTIO_NET_F_RSS, |
9aa47edd | 90 | VIRTIO_NET_F_STATUS, |
2875a0ca HJ |
91 | VIRTIO_RING_F_EVENT_IDX, |
92 | VIRTIO_RING_F_INDIRECT_DESC, | |
93 | ||
94 | /* VHOST_INVALID_FEATURE_BIT should always be the last entry */ | |
1e0a84ea CL |
95 | VHOST_INVALID_FEATURE_BIT |
96 | }; | |
97 | ||
1576dbb5 EP |
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) | | |
4b4a1378 | 102 | BIT_ULL(VIRTIO_NET_F_CTRL_GUEST_OFFLOADS) | |
1576dbb5 EP |
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) | | |
ea6eec49 | 116 | BIT_ULL(VIRTIO_NET_F_CTRL_RX) | |
e213c45a | 117 | BIT_ULL(VIRTIO_NET_F_CTRL_VLAN) | |
d669b7bb | 118 | BIT_ULL(VIRTIO_NET_F_CTRL_RX_EXTRA) | |
72b99a87 | 119 | BIT_ULL(VIRTIO_NET_F_MQ) | |
1576dbb5 EP |
120 | BIT_ULL(VIRTIO_F_ANY_LAYOUT) | |
121 | BIT_ULL(VIRTIO_NET_F_CTRL_MAC_ADDR) | | |
609ab4c3 EP |
122 | /* VHOST_F_LOG_ALL is exposed by SVQ */ |
123 | BIT_ULL(VHOST_F_LOG_ALL) | | |
1576dbb5 | 124 | BIT_ULL(VIRTIO_NET_F_RSC_EXT) | |
0d74e2b7 EP |
125 | BIT_ULL(VIRTIO_NET_F_STANDBY) | |
126 | BIT_ULL(VIRTIO_NET_F_SPEED_DUPLEX); | |
1576dbb5 | 127 | |
c1a10086 EP |
128 | #define VHOST_VDPA_NET_CVQ_ASID 1 |
129 | ||
1e0a84ea CL |
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 | ||
915bf6cc EP |
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 | ||
36e46472 EP |
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); | |
258a0394 | 164 | return false; |
36e46472 EP |
165 | } |
166 | ||
258a0394 | 167 | return vhost_svq_valid_features(features, errp); |
36e46472 EP |
168 | } |
169 | ||
1e0a84ea CL |
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 | ||
40237840 JW |
184 | static int vhost_vdpa_add(NetClientState *ncs, void *be, |
185 | int queue_pair_index, int nvqs) | |
1e0a84ea CL |
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; | |
40237840 | 198 | options.nvqs = nvqs; |
1e0a84ea CL |
199 | |
200 | net = vhost_net_init(&options); | |
201 | if (!net) { | |
202 | error_report("failed to init vhost_net for queue"); | |
a97ef87a | 203 | goto err_init; |
1e0a84ea | 204 | } |
1e0a84ea CL |
205 | s->vhost_net = net; |
206 | ret = vhost_vdpa_net_check_device_id(net); | |
207 | if (ret) { | |
a97ef87a | 208 | goto err_check; |
1e0a84ea CL |
209 | } |
210 | return 0; | |
a97ef87a JW |
211 | err_check: |
212 | vhost_net_cleanup(net); | |
213 | g_free(net); | |
214 | err_init: | |
1e0a84ea CL |
215 | return -1; |
216 | } | |
217 | ||
218 | static void vhost_vdpa_cleanup(NetClientState *nc) | |
219 | { | |
220 | VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); | |
221 | ||
a0d7215e AS |
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 | } | |
babf8b87 EP |
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()); | |
1e0a84ea CL |
232 | if (s->vhost_net) { |
233 | vhost_net_cleanup(s->vhost_net); | |
234 | g_free(s->vhost_net); | |
235 | s->vhost_net = NULL; | |
236 | } | |
57b3a7d8 CL |
237 | if (s->vhost_vdpa.device_fd >= 0) { |
238 | qemu_close(s->vhost_vdpa.device_fd); | |
239 | s->vhost_vdpa.device_fd = -1; | |
240 | } | |
1e0a84ea CL |
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 | ||
ee8a1c63 KW |
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 | ||
846a1e85 EP |
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 | { | |
bc5add1d | 278 | return size; |
846a1e85 EP |
279 | } |
280 | ||
00ef422e EP |
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 | ||
69498430 EP |
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 | ||
00ef422e EP |
338 | static void vhost_vdpa_net_data_start_first(VhostVDPAState *s) |
339 | { | |
340 | struct vhost_vdpa *v = &s->vhost_vdpa; | |
341 | ||
69498430 | 342 | add_migration_state_change_notifier(&s->migration_state); |
00ef422e EP |
343 | if (v->shadow_vqs_enabled) { |
344 | v->iova_tree = vhost_iova_tree_new(v->iova_range.first, | |
345 | v->iova_range.last); | |
346 | } | |
347 | } | |
348 | ||
349 | static int vhost_vdpa_net_data_start(NetClientState *nc) | |
350 | { | |
351 | VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); | |
352 | struct vhost_vdpa *v = &s->vhost_vdpa; | |
353 | ||
354 | assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); | |
355 | ||
69498430 EP |
356 | if (s->always_svq || |
357 | migration_is_setup_or_active(migrate_get_current()->state)) { | |
358 | v->shadow_vqs_enabled = true; | |
359 | v->shadow_data = true; | |
360 | } else { | |
361 | v->shadow_vqs_enabled = false; | |
362 | v->shadow_data = false; | |
363 | } | |
364 | ||
00ef422e EP |
365 | if (v->index == 0) { |
366 | vhost_vdpa_net_data_start_first(s); | |
367 | return 0; | |
368 | } | |
369 | ||
370 | if (v->shadow_vqs_enabled) { | |
371 | VhostVDPAState *s0 = vhost_vdpa_net_first_nc_vdpa(s); | |
372 | v->iova_tree = s0->vhost_vdpa.iova_tree; | |
373 | } | |
374 | ||
375 | return 0; | |
376 | } | |
377 | ||
378 | static void vhost_vdpa_net_client_stop(NetClientState *nc) | |
379 | { | |
380 | VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); | |
381 | struct vhost_dev *dev; | |
382 | ||
383 | assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); | |
384 | ||
69498430 EP |
385 | if (s->vhost_vdpa.index == 0) { |
386 | remove_migration_state_change_notifier(&s->migration_state); | |
387 | } | |
388 | ||
00ef422e EP |
389 | dev = s->vhost_vdpa.dev; |
390 | if (dev->vq_index + dev->nvqs == dev->vq_index_end) { | |
391 | g_clear_pointer(&s->vhost_vdpa.iova_tree, vhost_iova_tree_delete); | |
392 | } | |
393 | } | |
394 | ||
1e0a84ea CL |
395 | static NetClientInfo net_vhost_vdpa_info = { |
396 | .type = NET_CLIENT_DRIVER_VHOST_VDPA, | |
397 | .size = sizeof(VhostVDPAState), | |
846a1e85 | 398 | .receive = vhost_vdpa_receive, |
00ef422e EP |
399 | .start = vhost_vdpa_net_data_start, |
400 | .stop = vhost_vdpa_net_client_stop, | |
1e0a84ea CL |
401 | .cleanup = vhost_vdpa_cleanup, |
402 | .has_vnet_hdr = vhost_vdpa_has_vnet_hdr, | |
403 | .has_ufo = vhost_vdpa_has_ufo, | |
ee8a1c63 | 404 | .check_peer_type = vhost_vdpa_check_peer_type, |
1e0a84ea CL |
405 | }; |
406 | ||
152128d6 EP |
407 | static int64_t vhost_vdpa_get_vring_group(int device_fd, unsigned vq_index, |
408 | Error **errp) | |
c1a10086 EP |
409 | { |
410 | struct vhost_vring_state state = { | |
411 | .index = vq_index, | |
412 | }; | |
413 | int r = ioctl(device_fd, VHOST_VDPA_GET_VRING_GROUP, &state); | |
414 | ||
415 | if (unlikely(r < 0)) { | |
0f2bb0bf | 416 | r = -errno; |
152128d6 | 417 | error_setg_errno(errp, errno, "Cannot get VQ %u group", vq_index); |
c1a10086 EP |
418 | return r; |
419 | } | |
420 | ||
421 | return state.num; | |
422 | } | |
423 | ||
424 | static int vhost_vdpa_set_address_space_id(struct vhost_vdpa *v, | |
425 | unsigned vq_group, | |
426 | unsigned asid_num) | |
427 | { | |
428 | struct vhost_vring_state asid = { | |
429 | .index = vq_group, | |
430 | .num = asid_num, | |
431 | }; | |
432 | int r; | |
433 | ||
434 | r = ioctl(v->device_fd, VHOST_VDPA_SET_GROUP_ASID, &asid); | |
435 | if (unlikely(r < 0)) { | |
436 | error_report("Can't set vq group %u asid %u, errno=%d (%s)", | |
437 | asid.index, asid.num, errno, g_strerror(errno)); | |
438 | } | |
439 | return r; | |
440 | } | |
441 | ||
2df4dd31 EP |
442 | static void vhost_vdpa_cvq_unmap_buf(struct vhost_vdpa *v, void *addr) |
443 | { | |
444 | VhostIOVATree *tree = v->iova_tree; | |
445 | DMAMap needle = { | |
446 | /* | |
447 | * No need to specify size or to look for more translations since | |
448 | * this contiguous chunk was allocated by us. | |
449 | */ | |
450 | .translated_addr = (hwaddr)(uintptr_t)addr, | |
451 | }; | |
452 | const DMAMap *map = vhost_iova_tree_find_iova(tree, &needle); | |
453 | int r; | |
454 | ||
455 | if (unlikely(!map)) { | |
456 | error_report("Cannot locate expected map"); | |
457 | return; | |
458 | } | |
459 | ||
cd831ed5 | 460 | r = vhost_vdpa_dma_unmap(v, v->address_space_id, map->iova, map->size + 1); |
2df4dd31 EP |
461 | if (unlikely(r != 0)) { |
462 | error_report("Device cannot unmap: %s(%d)", g_strerror(r), r); | |
463 | } | |
464 | ||
69292a8e | 465 | vhost_iova_tree_remove(tree, *map); |
2df4dd31 EP |
466 | } |
467 | ||
7a7f87e9 EP |
468 | /** Map CVQ buffer. */ |
469 | static int vhost_vdpa_cvq_map_buf(struct vhost_vdpa *v, void *buf, size_t size, | |
470 | bool write) | |
2df4dd31 EP |
471 | { |
472 | DMAMap map = {}; | |
473 | int r; | |
474 | ||
2df4dd31 | 475 | map.translated_addr = (hwaddr)(uintptr_t)buf; |
7a7f87e9 | 476 | map.size = size - 1; |
2df4dd31 EP |
477 | map.perm = write ? IOMMU_RW : IOMMU_RO, |
478 | r = vhost_iova_tree_map_alloc(v->iova_tree, &map); | |
479 | if (unlikely(r != IOVA_OK)) { | |
480 | error_report("Cannot map injected element"); | |
7a7f87e9 | 481 | return r; |
2df4dd31 EP |
482 | } |
483 | ||
cd831ed5 EP |
484 | r = vhost_vdpa_dma_map(v, v->address_space_id, map.iova, |
485 | vhost_vdpa_net_cvq_cmd_page_len(), buf, !write); | |
2df4dd31 EP |
486 | if (unlikely(r < 0)) { |
487 | goto dma_map_err; | |
488 | } | |
489 | ||
7a7f87e9 | 490 | return 0; |
2df4dd31 EP |
491 | |
492 | dma_map_err: | |
69292a8e | 493 | vhost_iova_tree_remove(v->iova_tree, map); |
7a7f87e9 | 494 | return r; |
2df4dd31 EP |
495 | } |
496 | ||
7a7f87e9 | 497 | static int vhost_vdpa_net_cvq_start(NetClientState *nc) |
2df4dd31 | 498 | { |
00ef422e | 499 | VhostVDPAState *s, *s0; |
c1a10086 | 500 | struct vhost_vdpa *v; |
c1a10086 | 501 | int64_t cvq_group; |
152128d6 EP |
502 | int r; |
503 | Error *err = NULL; | |
2df4dd31 | 504 | |
7a7f87e9 EP |
505 | assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); |
506 | ||
507 | s = DO_UPCAST(VhostVDPAState, nc, nc); | |
c1a10086 EP |
508 | v = &s->vhost_vdpa; |
509 | ||
69498430 EP |
510 | s0 = vhost_vdpa_net_first_nc_vdpa(s); |
511 | v->shadow_data = s0->vhost_vdpa.shadow_vqs_enabled; | |
c1a10086 EP |
512 | v->shadow_vqs_enabled = s->always_svq; |
513 | s->vhost_vdpa.address_space_id = VHOST_VDPA_GUEST_PA_ASID; | |
514 | ||
69498430 | 515 | if (s->vhost_vdpa.shadow_data) { |
c1a10086 EP |
516 | /* SVQ is already configured for all virtqueues */ |
517 | goto out; | |
518 | } | |
519 | ||
520 | /* | |
521 | * If we early return in these cases SVQ will not be enabled. The migration | |
522 | * will be blocked as long as vhost-vdpa backends will not offer _F_LOG. | |
c1a10086 | 523 | */ |
152128d6 EP |
524 | if (!vhost_vdpa_net_valid_svq_features(v->dev->features, NULL)) { |
525 | return 0; | |
c1a10086 | 526 | } |
152128d6 EP |
527 | |
528 | if (!s->cvq_isolated) { | |
c1a10086 EP |
529 | return 0; |
530 | } | |
531 | ||
152128d6 EP |
532 | cvq_group = vhost_vdpa_get_vring_group(v->device_fd, |
533 | v->dev->vq_index_end - 1, | |
534 | &err); | |
c1a10086 | 535 | if (unlikely(cvq_group < 0)) { |
152128d6 | 536 | error_report_err(err); |
c1a10086 EP |
537 | return cvq_group; |
538 | } | |
c1a10086 EP |
539 | |
540 | r = vhost_vdpa_set_address_space_id(v, cvq_group, VHOST_VDPA_NET_CVQ_ASID); | |
541 | if (unlikely(r < 0)) { | |
542 | return r; | |
543 | } | |
544 | ||
c1a10086 EP |
545 | v->shadow_vqs_enabled = true; |
546 | s->vhost_vdpa.address_space_id = VHOST_VDPA_NET_CVQ_ASID; | |
547 | ||
548 | out: | |
7a7f87e9 EP |
549 | if (!s->vhost_vdpa.shadow_vqs_enabled) { |
550 | return 0; | |
2df4dd31 EP |
551 | } |
552 | ||
00ef422e EP |
553 | if (s0->vhost_vdpa.iova_tree) { |
554 | /* | |
555 | * SVQ is already configured for all virtqueues. Reuse IOVA tree for | |
556 | * simplicity, whether CVQ shares ASID with guest or not, because: | |
557 | * - Memory listener need access to guest's memory addresses allocated | |
558 | * in the IOVA tree. | |
559 | * - There should be plenty of IOVA address space for both ASID not to | |
560 | * worry about collisions between them. Guest's translations are | |
561 | * still validated with virtio virtqueue_pop so there is no risk for | |
562 | * the guest to access memory that it shouldn't. | |
563 | * | |
564 | * To allocate a iova tree per ASID is doable but it complicates the | |
565 | * code and it is not worth it for the moment. | |
566 | */ | |
567 | v->iova_tree = s0->vhost_vdpa.iova_tree; | |
568 | } else { | |
569 | v->iova_tree = vhost_iova_tree_new(v->iova_range.first, | |
570 | v->iova_range.last); | |
571 | } | |
572 | ||
7a7f87e9 EP |
573 | r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer, |
574 | vhost_vdpa_net_cvq_cmd_page_len(), false); | |
575 | if (unlikely(r < 0)) { | |
576 | return r; | |
577 | } | |
578 | ||
17fb889f | 579 | r = vhost_vdpa_cvq_map_buf(&s->vhost_vdpa, s->status, |
7a7f87e9 EP |
580 | vhost_vdpa_net_cvq_cmd_page_len(), true); |
581 | if (unlikely(r < 0)) { | |
2df4dd31 | 582 | vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer); |
2df4dd31 EP |
583 | } |
584 | ||
7a7f87e9 EP |
585 | return r; |
586 | } | |
587 | ||
588 | static void vhost_vdpa_net_cvq_stop(NetClientState *nc) | |
589 | { | |
590 | VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); | |
591 | ||
592 | assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); | |
593 | ||
594 | if (s->vhost_vdpa.shadow_vqs_enabled) { | |
595 | vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->cvq_cmd_out_buffer); | |
17fb889f | 596 | vhost_vdpa_cvq_unmap_buf(&s->vhost_vdpa, s->status); |
7a7f87e9 | 597 | } |
00ef422e EP |
598 | |
599 | vhost_vdpa_net_client_stop(nc); | |
2df4dd31 EP |
600 | } |
601 | ||
be4278b6 EP |
602 | static ssize_t vhost_vdpa_net_cvq_add(VhostVDPAState *s, size_t out_len, |
603 | size_t in_len) | |
604 | { | |
605 | /* Buffers for the device */ | |
606 | const struct iovec out = { | |
607 | .iov_base = s->cvq_cmd_out_buffer, | |
608 | .iov_len = out_len, | |
609 | }; | |
610 | const struct iovec in = { | |
17fb889f | 611 | .iov_base = s->status, |
be4278b6 EP |
612 | .iov_len = sizeof(virtio_net_ctrl_ack), |
613 | }; | |
614 | VhostShadowVirtqueue *svq = g_ptr_array_index(s->vhost_vdpa.shadow_vqs, 0); | |
615 | int r; | |
616 | ||
617 | r = vhost_svq_add(svq, &out, 1, &in, 1, NULL); | |
618 | if (unlikely(r != 0)) { | |
619 | if (unlikely(r == -ENOSPC)) { | |
620 | qemu_log_mask(LOG_GUEST_ERROR, "%s: No space on device queue\n", | |
621 | __func__); | |
622 | } | |
623 | return r; | |
624 | } | |
625 | ||
626 | /* | |
627 | * We can poll here since we've had BQL from the time we sent the | |
628 | * descriptor. Also, we need to take the answer before SVQ pulls by itself, | |
629 | * when BQL is released | |
630 | */ | |
631 | return vhost_svq_poll(svq); | |
632 | } | |
633 | ||
f73c0c43 | 634 | static ssize_t vhost_vdpa_net_load_cmd(VhostVDPAState *s, uint8_t class, |
2848c6aa HJ |
635 | uint8_t cmd, const struct iovec *data_sg, |
636 | size_t data_num) | |
f73c0c43 EP |
637 | { |
638 | const struct virtio_net_ctrl_hdr ctrl = { | |
639 | .class = class, | |
640 | .cmd = cmd, | |
641 | }; | |
2848c6aa | 642 | size_t data_size = iov_size(data_sg, data_num); |
f73c0c43 EP |
643 | |
644 | assert(data_size < vhost_vdpa_net_cvq_cmd_page_len() - sizeof(ctrl)); | |
645 | ||
2848c6aa | 646 | /* pack the CVQ command header */ |
f73c0c43 | 647 | memcpy(s->cvq_cmd_out_buffer, &ctrl, sizeof(ctrl)); |
f73c0c43 | 648 | |
2848c6aa HJ |
649 | /* pack the CVQ command command-specific-data */ |
650 | iov_to_buf(data_sg, data_num, 0, | |
651 | s->cvq_cmd_out_buffer + sizeof(ctrl), data_size); | |
652 | ||
653 | return vhost_vdpa_net_cvq_add(s, data_size + sizeof(ctrl), | |
f73c0c43 EP |
654 | sizeof(virtio_net_ctrl_ack)); |
655 | } | |
656 | ||
657 | static int vhost_vdpa_net_load_mac(VhostVDPAState *s, const VirtIONet *n) | |
658 | { | |
02d3bf09 | 659 | if (virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_MAC_ADDR)) { |
2848c6aa HJ |
660 | const struct iovec data = { |
661 | .iov_base = (void *)n->mac, | |
662 | .iov_len = sizeof(n->mac), | |
663 | }; | |
f73c0c43 EP |
664 | ssize_t dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_MAC, |
665 | VIRTIO_NET_CTRL_MAC_ADDR_SET, | |
2848c6aa | 666 | &data, 1); |
f73c0c43 EP |
667 | if (unlikely(dev_written < 0)) { |
668 | return dev_written; | |
669 | } | |
b479bc3c HJ |
670 | if (*s->status != VIRTIO_NET_OK) { |
671 | return -EIO; | |
672 | } | |
f73c0c43 EP |
673 | } |
674 | ||
0ddcecb8 HJ |
675 | /* |
676 | * According to VirtIO standard, "The device MUST have an | |
677 | * empty MAC filtering table on reset.". | |
678 | * | |
679 | * Therefore, there is no need to send this CVQ command if the | |
680 | * driver also sets an empty MAC filter table, which aligns with | |
681 | * the device's defaults. | |
682 | * | |
683 | * Note that the device's defaults can mismatch the driver's | |
684 | * configuration only at live migration. | |
685 | */ | |
686 | if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_RX) || | |
687 | n->mac_table.in_use == 0) { | |
688 | return 0; | |
689 | } | |
690 | ||
691 | uint32_t uni_entries = n->mac_table.first_multi, | |
692 | uni_macs_size = uni_entries * ETH_ALEN, | |
693 | mul_entries = n->mac_table.in_use - uni_entries, | |
694 | mul_macs_size = mul_entries * ETH_ALEN; | |
695 | struct virtio_net_ctrl_mac uni = { | |
696 | .entries = cpu_to_le32(uni_entries), | |
697 | }; | |
698 | struct virtio_net_ctrl_mac mul = { | |
699 | .entries = cpu_to_le32(mul_entries), | |
700 | }; | |
701 | const struct iovec data[] = { | |
702 | { | |
703 | .iov_base = &uni, | |
704 | .iov_len = sizeof(uni), | |
705 | }, { | |
706 | .iov_base = n->mac_table.macs, | |
707 | .iov_len = uni_macs_size, | |
708 | }, { | |
709 | .iov_base = &mul, | |
710 | .iov_len = sizeof(mul), | |
711 | }, { | |
712 | .iov_base = &n->mac_table.macs[uni_macs_size], | |
713 | .iov_len = mul_macs_size, | |
714 | }, | |
715 | }; | |
716 | ssize_t dev_written = vhost_vdpa_net_load_cmd(s, | |
717 | VIRTIO_NET_CTRL_MAC, | |
718 | VIRTIO_NET_CTRL_MAC_TABLE_SET, | |
719 | data, ARRAY_SIZE(data)); | |
720 | if (unlikely(dev_written < 0)) { | |
721 | return dev_written; | |
722 | } | |
723 | if (*s->status != VIRTIO_NET_OK) { | |
724 | return -EIO; | |
725 | } | |
726 | ||
f73c0c43 EP |
727 | return 0; |
728 | } | |
729 | ||
f64c7cda EP |
730 | static int vhost_vdpa_net_load_mq(VhostVDPAState *s, |
731 | const VirtIONet *n) | |
732 | { | |
733 | struct virtio_net_ctrl_mq mq; | |
f64c7cda EP |
734 | ssize_t dev_written; |
735 | ||
02d3bf09 | 736 | if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_MQ)) { |
f64c7cda EP |
737 | return 0; |
738 | } | |
739 | ||
740 | mq.virtqueue_pairs = cpu_to_le16(n->curr_queue_pairs); | |
2848c6aa HJ |
741 | const struct iovec data = { |
742 | .iov_base = &mq, | |
743 | .iov_len = sizeof(mq), | |
744 | }; | |
f64c7cda | 745 | dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_MQ, |
2848c6aa HJ |
746 | VIRTIO_NET_CTRL_MQ_VQ_PAIRS_SET, |
747 | &data, 1); | |
f64c7cda EP |
748 | if (unlikely(dev_written < 0)) { |
749 | return dev_written; | |
750 | } | |
f45fd95e HJ |
751 | if (*s->status != VIRTIO_NET_OK) { |
752 | return -EIO; | |
753 | } | |
f64c7cda | 754 | |
f45fd95e | 755 | return 0; |
f64c7cda EP |
756 | } |
757 | ||
0b58d368 HJ |
758 | static int vhost_vdpa_net_load_offloads(VhostVDPAState *s, |
759 | const VirtIONet *n) | |
760 | { | |
761 | uint64_t offloads; | |
762 | ssize_t dev_written; | |
763 | ||
764 | if (!virtio_vdev_has_feature(&n->parent_obj, | |
765 | VIRTIO_NET_F_CTRL_GUEST_OFFLOADS)) { | |
766 | return 0; | |
767 | } | |
768 | ||
769 | if (n->curr_guest_offloads == virtio_net_supported_guest_offloads(n)) { | |
770 | /* | |
771 | * According to VirtIO standard, "Upon feature negotiation | |
772 | * corresponding offload gets enabled to preserve | |
773 | * backward compatibility.". | |
774 | * | |
775 | * Therefore, there is no need to send this CVQ command if the | |
776 | * driver also enables all supported offloads, which aligns with | |
777 | * the device's defaults. | |
778 | * | |
779 | * Note that the device's defaults can mismatch the driver's | |
780 | * configuration only at live migration. | |
781 | */ | |
782 | return 0; | |
783 | } | |
784 | ||
785 | offloads = cpu_to_le64(n->curr_guest_offloads); | |
2848c6aa HJ |
786 | const struct iovec data = { |
787 | .iov_base = &offloads, | |
788 | .iov_len = sizeof(offloads), | |
789 | }; | |
0b58d368 HJ |
790 | dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_GUEST_OFFLOADS, |
791 | VIRTIO_NET_CTRL_GUEST_OFFLOADS_SET, | |
2848c6aa | 792 | &data, 1); |
0b58d368 HJ |
793 | if (unlikely(dev_written < 0)) { |
794 | return dev_written; | |
795 | } | |
6f348071 HJ |
796 | if (*s->status != VIRTIO_NET_OK) { |
797 | return -EIO; | |
798 | } | |
0b58d368 | 799 | |
6f348071 | 800 | return 0; |
0b58d368 HJ |
801 | } |
802 | ||
b12f907e HJ |
803 | static int vhost_vdpa_net_load_rx_mode(VhostVDPAState *s, |
804 | uint8_t cmd, | |
805 | uint8_t on) | |
806 | { | |
807 | const struct iovec data = { | |
808 | .iov_base = &on, | |
809 | .iov_len = sizeof(on), | |
810 | }; | |
811 | return vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_RX, | |
812 | cmd, &data, 1); | |
813 | } | |
814 | ||
815 | static int vhost_vdpa_net_load_rx(VhostVDPAState *s, | |
816 | const VirtIONet *n) | |
817 | { | |
818 | ssize_t dev_written; | |
819 | ||
820 | if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_RX)) { | |
821 | return 0; | |
822 | } | |
823 | ||
824 | /* | |
825 | * According to virtio_net_reset(), device turns promiscuous mode | |
826 | * on by default. | |
827 | * | |
0a19d879 | 828 | * Additionally, according to VirtIO standard, "Since there are |
b12f907e HJ |
829 | * no guarantees, it can use a hash filter or silently switch to |
830 | * allmulti or promiscuous mode if it is given too many addresses.". | |
831 | * QEMU marks `n->mac_table.uni_overflow` if guest sets too many | |
832 | * non-multicast MAC addresses, indicating that promiscuous mode | |
833 | * should be enabled. | |
834 | * | |
835 | * Therefore, QEMU should only send this CVQ command if the | |
836 | * `n->mac_table.uni_overflow` is not marked and `n->promisc` is off, | |
837 | * which sets promiscuous mode on, different from the device's defaults. | |
838 | * | |
839 | * Note that the device's defaults can mismatch the driver's | |
840 | * configuration only at live migration. | |
841 | */ | |
842 | if (!n->mac_table.uni_overflow && !n->promisc) { | |
843 | dev_written = vhost_vdpa_net_load_rx_mode(s, | |
844 | VIRTIO_NET_CTRL_RX_PROMISC, 0); | |
845 | if (unlikely(dev_written < 0)) { | |
846 | return dev_written; | |
847 | } | |
848 | if (*s->status != VIRTIO_NET_OK) { | |
849 | return -EIO; | |
850 | } | |
851 | } | |
852 | ||
853 | /* | |
854 | * According to virtio_net_reset(), device turns all-multicast mode | |
855 | * off by default. | |
856 | * | |
857 | * According to VirtIO standard, "Since there are no guarantees, | |
858 | * it can use a hash filter or silently switch to allmulti or | |
859 | * promiscuous mode if it is given too many addresses.". QEMU marks | |
860 | * `n->mac_table.multi_overflow` if guest sets too many | |
861 | * non-multicast MAC addresses. | |
862 | * | |
863 | * Therefore, QEMU should only send this CVQ command if the | |
864 | * `n->mac_table.multi_overflow` is marked or `n->allmulti` is on, | |
865 | * which sets all-multicast mode on, different from the device's defaults. | |
866 | * | |
867 | * Note that the device's defaults can mismatch the driver's | |
868 | * configuration only at live migration. | |
869 | */ | |
870 | if (n->mac_table.multi_overflow || n->allmulti) { | |
871 | dev_written = vhost_vdpa_net_load_rx_mode(s, | |
872 | VIRTIO_NET_CTRL_RX_ALLMULTI, 1); | |
873 | if (unlikely(dev_written < 0)) { | |
874 | return dev_written; | |
875 | } | |
876 | if (*s->status != VIRTIO_NET_OK) { | |
877 | return -EIO; | |
878 | } | |
879 | } | |
880 | ||
4fd180c7 HJ |
881 | if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_RX_EXTRA)) { |
882 | return 0; | |
883 | } | |
884 | ||
885 | /* | |
886 | * According to virtio_net_reset(), device turns all-unicast mode | |
887 | * off by default. | |
888 | * | |
889 | * Therefore, QEMU should only send this CVQ command if the driver | |
890 | * sets all-unicast mode on, different from the device's defaults. | |
891 | * | |
892 | * Note that the device's defaults can mismatch the driver's | |
893 | * configuration only at live migration. | |
894 | */ | |
895 | if (n->alluni) { | |
896 | dev_written = vhost_vdpa_net_load_rx_mode(s, | |
897 | VIRTIO_NET_CTRL_RX_ALLUNI, 1); | |
898 | if (dev_written < 0) { | |
899 | return dev_written; | |
900 | } | |
901 | if (*s->status != VIRTIO_NET_OK) { | |
902 | return -EIO; | |
903 | } | |
904 | } | |
905 | ||
906 | /* | |
907 | * According to virtio_net_reset(), device turns non-multicast mode | |
908 | * off by default. | |
909 | * | |
910 | * Therefore, QEMU should only send this CVQ command if the driver | |
911 | * sets non-multicast mode on, different from the device's defaults. | |
912 | * | |
913 | * Note that the device's defaults can mismatch the driver's | |
914 | * configuration only at live migration. | |
915 | */ | |
916 | if (n->nomulti) { | |
917 | dev_written = vhost_vdpa_net_load_rx_mode(s, | |
918 | VIRTIO_NET_CTRL_RX_NOMULTI, 1); | |
919 | if (dev_written < 0) { | |
920 | return dev_written; | |
921 | } | |
922 | if (*s->status != VIRTIO_NET_OK) { | |
923 | return -EIO; | |
924 | } | |
925 | } | |
926 | ||
927 | /* | |
928 | * According to virtio_net_reset(), device turns non-unicast mode | |
929 | * off by default. | |
930 | * | |
931 | * Therefore, QEMU should only send this CVQ command if the driver | |
932 | * sets non-unicast mode on, different from the device's defaults. | |
933 | * | |
934 | * Note that the device's defaults can mismatch the driver's | |
935 | * configuration only at live migration. | |
936 | */ | |
937 | if (n->nouni) { | |
938 | dev_written = vhost_vdpa_net_load_rx_mode(s, | |
939 | VIRTIO_NET_CTRL_RX_NOUNI, 1); | |
940 | if (dev_written < 0) { | |
941 | return dev_written; | |
942 | } | |
943 | if (*s->status != VIRTIO_NET_OK) { | |
944 | return -EIO; | |
945 | } | |
946 | } | |
947 | ||
948 | /* | |
949 | * According to virtio_net_reset(), device turns non-broadcast mode | |
950 | * off by default. | |
951 | * | |
952 | * Therefore, QEMU should only send this CVQ command if the driver | |
953 | * sets non-broadcast mode on, different from the device's defaults. | |
954 | * | |
955 | * Note that the device's defaults can mismatch the driver's | |
956 | * configuration only at live migration. | |
957 | */ | |
958 | if (n->nobcast) { | |
959 | dev_written = vhost_vdpa_net_load_rx_mode(s, | |
960 | VIRTIO_NET_CTRL_RX_NOBCAST, 1); | |
961 | if (dev_written < 0) { | |
962 | return dev_written; | |
963 | } | |
964 | if (*s->status != VIRTIO_NET_OK) { | |
965 | return -EIO; | |
966 | } | |
967 | } | |
968 | ||
b12f907e HJ |
969 | return 0; |
970 | } | |
971 | ||
8f7e9967 HJ |
972 | static int vhost_vdpa_net_load_single_vlan(VhostVDPAState *s, |
973 | const VirtIONet *n, | |
974 | uint16_t vid) | |
975 | { | |
976 | const struct iovec data = { | |
977 | .iov_base = &vid, | |
978 | .iov_len = sizeof(vid), | |
979 | }; | |
980 | ssize_t dev_written = vhost_vdpa_net_load_cmd(s, VIRTIO_NET_CTRL_VLAN, | |
981 | VIRTIO_NET_CTRL_VLAN_ADD, | |
982 | &data, 1); | |
983 | if (unlikely(dev_written < 0)) { | |
984 | return dev_written; | |
985 | } | |
986 | if (unlikely(*s->status != VIRTIO_NET_OK)) { | |
987 | return -EIO; | |
988 | } | |
989 | ||
990 | return 0; | |
991 | } | |
992 | ||
993 | static int vhost_vdpa_net_load_vlan(VhostVDPAState *s, | |
994 | const VirtIONet *n) | |
995 | { | |
996 | int r; | |
997 | ||
998 | if (!virtio_vdev_has_feature(&n->parent_obj, VIRTIO_NET_F_CTRL_VLAN)) { | |
999 | return 0; | |
1000 | } | |
1001 | ||
1002 | for (int i = 0; i < MAX_VLAN >> 5; i++) { | |
1003 | for (int j = 0; n->vlans[i] && j <= 0x1f; j++) { | |
1004 | if (n->vlans[i] & (1U << j)) { | |
1005 | r = vhost_vdpa_net_load_single_vlan(s, n, (i << 5) + j); | |
1006 | if (unlikely(r != 0)) { | |
1007 | return r; | |
1008 | } | |
1009 | } | |
1010 | } | |
1011 | } | |
1012 | ||
1013 | return 0; | |
1014 | } | |
1015 | ||
dd036d8d EP |
1016 | static int vhost_vdpa_net_load(NetClientState *nc) |
1017 | { | |
1018 | VhostVDPAState *s = DO_UPCAST(VhostVDPAState, nc, nc); | |
f73c0c43 | 1019 | struct vhost_vdpa *v = &s->vhost_vdpa; |
dd036d8d | 1020 | const VirtIONet *n; |
f73c0c43 | 1021 | int r; |
dd036d8d EP |
1022 | |
1023 | assert(nc->info->type == NET_CLIENT_DRIVER_VHOST_VDPA); | |
1024 | ||
1025 | if (!v->shadow_vqs_enabled) { | |
1026 | return 0; | |
1027 | } | |
1028 | ||
1029 | n = VIRTIO_NET(v->dev->vdev); | |
f73c0c43 EP |
1030 | r = vhost_vdpa_net_load_mac(s, n); |
1031 | if (unlikely(r < 0)) { | |
1032 | return r; | |
dd036d8d | 1033 | } |
f64c7cda EP |
1034 | r = vhost_vdpa_net_load_mq(s, n); |
1035 | if (unlikely(r)) { | |
1036 | return r; | |
1037 | } | |
0b58d368 HJ |
1038 | r = vhost_vdpa_net_load_offloads(s, n); |
1039 | if (unlikely(r)) { | |
1040 | return r; | |
1041 | } | |
b12f907e HJ |
1042 | r = vhost_vdpa_net_load_rx(s, n); |
1043 | if (unlikely(r)) { | |
1044 | return r; | |
1045 | } | |
8f7e9967 HJ |
1046 | r = vhost_vdpa_net_load_vlan(s, n); |
1047 | if (unlikely(r)) { | |
1048 | return r; | |
1049 | } | |
dd036d8d EP |
1050 | |
1051 | return 0; | |
1052 | } | |
1053 | ||
f8972b56 EP |
1054 | static NetClientInfo net_vhost_vdpa_cvq_info = { |
1055 | .type = NET_CLIENT_DRIVER_VHOST_VDPA, | |
1056 | .size = sizeof(VhostVDPAState), | |
1057 | .receive = vhost_vdpa_receive, | |
7a7f87e9 | 1058 | .start = vhost_vdpa_net_cvq_start, |
dd036d8d | 1059 | .load = vhost_vdpa_net_load, |
7a7f87e9 | 1060 | .stop = vhost_vdpa_net_cvq_stop, |
f8972b56 EP |
1061 | .cleanup = vhost_vdpa_cleanup, |
1062 | .has_vnet_hdr = vhost_vdpa_has_vnet_hdr, | |
1063 | .has_ufo = vhost_vdpa_has_ufo, | |
1064 | .check_peer_type = vhost_vdpa_check_peer_type, | |
1065 | }; | |
1066 | ||
fee364e4 HJ |
1067 | /* |
1068 | * Forward the excessive VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command to | |
1069 | * vdpa device. | |
1070 | * | |
1071 | * Considering that QEMU cannot send the entire filter table to the | |
1072 | * vdpa device, it should send the VIRTIO_NET_CTRL_RX_PROMISC CVQ | |
1073 | * command to enable promiscuous mode to receive all packets, | |
1074 | * according to VirtIO standard, "Since there are no guarantees, | |
1075 | * it can use a hash filter or silently switch to allmulti or | |
1076 | * promiscuous mode if it is given too many addresses.". | |
1077 | * | |
1078 | * Since QEMU ignores MAC addresses beyond `MAC_TABLE_ENTRIES` and | |
1079 | * marks `n->mac_table.x_overflow` accordingly, it should have | |
1080 | * the same effect on the device model to receive | |
1081 | * (`MAC_TABLE_ENTRIES` + 1) or more non-multicast MAC addresses. | |
1082 | * The same applies to multicast MAC addresses. | |
1083 | * | |
1084 | * Therefore, QEMU can provide the device model with a fake | |
1085 | * VIRTIO_NET_CTRL_MAC_TABLE_SET command with (`MAC_TABLE_ENTRIES` + 1) | |
1086 | * non-multicast MAC addresses and (`MAC_TABLE_ENTRIES` + 1) multicast | |
1087 | * MAC addresses. This ensures that the device model marks | |
1088 | * `n->mac_table.uni_overflow` and `n->mac_table.multi_overflow`, | |
1089 | * allowing all packets to be received, which aligns with the | |
1090 | * state of the vdpa device. | |
1091 | */ | |
1092 | static int vhost_vdpa_net_excessive_mac_filter_cvq_add(VhostVDPAState *s, | |
1093 | VirtQueueElement *elem, | |
1094 | struct iovec *out) | |
1095 | { | |
1096 | struct virtio_net_ctrl_mac mac_data, *mac_ptr; | |
1097 | struct virtio_net_ctrl_hdr *hdr_ptr; | |
1098 | uint32_t cursor; | |
1099 | ssize_t r; | |
1100 | ||
1101 | /* parse the non-multicast MAC address entries from CVQ command */ | |
1102 | cursor = sizeof(*hdr_ptr); | |
1103 | r = iov_to_buf(elem->out_sg, elem->out_num, cursor, | |
1104 | &mac_data, sizeof(mac_data)); | |
1105 | if (unlikely(r != sizeof(mac_data))) { | |
1106 | /* | |
1107 | * If the CVQ command is invalid, we should simulate the vdpa device | |
1108 | * to reject the VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command | |
1109 | */ | |
1110 | *s->status = VIRTIO_NET_ERR; | |
1111 | return sizeof(*s->status); | |
1112 | } | |
1113 | cursor += sizeof(mac_data) + le32_to_cpu(mac_data.entries) * ETH_ALEN; | |
1114 | ||
1115 | /* parse the multicast MAC address entries from CVQ command */ | |
1116 | r = iov_to_buf(elem->out_sg, elem->out_num, cursor, | |
1117 | &mac_data, sizeof(mac_data)); | |
1118 | if (r != sizeof(mac_data)) { | |
1119 | /* | |
1120 | * If the CVQ command is invalid, we should simulate the vdpa device | |
1121 | * to reject the VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command | |
1122 | */ | |
1123 | *s->status = VIRTIO_NET_ERR; | |
1124 | return sizeof(*s->status); | |
1125 | } | |
1126 | cursor += sizeof(mac_data) + le32_to_cpu(mac_data.entries) * ETH_ALEN; | |
1127 | ||
1128 | /* validate the CVQ command */ | |
1129 | if (iov_size(elem->out_sg, elem->out_num) != cursor) { | |
1130 | /* | |
1131 | * If the CVQ command is invalid, we should simulate the vdpa device | |
1132 | * to reject the VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command | |
1133 | */ | |
1134 | *s->status = VIRTIO_NET_ERR; | |
1135 | return sizeof(*s->status); | |
1136 | } | |
1137 | ||
1138 | /* | |
1139 | * According to VirtIO standard, "Since there are no guarantees, | |
1140 | * it can use a hash filter or silently switch to allmulti or | |
1141 | * promiscuous mode if it is given too many addresses.". | |
1142 | * | |
1143 | * Therefore, considering that QEMU is unable to send the entire | |
1144 | * filter table to the vdpa device, it should send the | |
1145 | * VIRTIO_NET_CTRL_RX_PROMISC CVQ command to enable promiscuous mode | |
1146 | */ | |
1147 | r = vhost_vdpa_net_load_rx_mode(s, VIRTIO_NET_CTRL_RX_PROMISC, 1); | |
1148 | if (unlikely(r < 0)) { | |
1149 | return r; | |
1150 | } | |
1151 | if (*s->status != VIRTIO_NET_OK) { | |
1152 | return sizeof(*s->status); | |
1153 | } | |
1154 | ||
1155 | /* | |
1156 | * QEMU should also send a fake VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ | |
1157 | * command to the device model, including (`MAC_TABLE_ENTRIES` + 1) | |
1158 | * non-multicast MAC addresses and (`MAC_TABLE_ENTRIES` + 1) | |
1159 | * multicast MAC addresses. | |
1160 | * | |
1161 | * By doing so, the device model can mark `n->mac_table.uni_overflow` | |
1162 | * and `n->mac_table.multi_overflow`, enabling all packets to be | |
1163 | * received, which aligns with the state of the vdpa device. | |
1164 | */ | |
1165 | cursor = 0; | |
1166 | uint32_t fake_uni_entries = MAC_TABLE_ENTRIES + 1, | |
1167 | fake_mul_entries = MAC_TABLE_ENTRIES + 1, | |
1168 | fake_cvq_size = sizeof(struct virtio_net_ctrl_hdr) + | |
1169 | sizeof(mac_data) + fake_uni_entries * ETH_ALEN + | |
1170 | sizeof(mac_data) + fake_mul_entries * ETH_ALEN; | |
1171 | ||
1172 | assert(fake_cvq_size < vhost_vdpa_net_cvq_cmd_page_len()); | |
1173 | out->iov_len = fake_cvq_size; | |
1174 | ||
1175 | /* pack the header for fake CVQ command */ | |
1176 | hdr_ptr = out->iov_base + cursor; | |
1177 | hdr_ptr->class = VIRTIO_NET_CTRL_MAC; | |
1178 | hdr_ptr->cmd = VIRTIO_NET_CTRL_MAC_TABLE_SET; | |
1179 | cursor += sizeof(*hdr_ptr); | |
1180 | ||
1181 | /* | |
1182 | * Pack the non-multicast MAC addresses part for fake CVQ command. | |
1183 | * | |
1184 | * According to virtio_net_handle_mac(), QEMU doesn't verify the MAC | |
0a19d879 | 1185 | * addresses provided in CVQ command. Therefore, only the entries |
fee364e4 HJ |
1186 | * field need to be prepared in the CVQ command. |
1187 | */ | |
1188 | mac_ptr = out->iov_base + cursor; | |
1189 | mac_ptr->entries = cpu_to_le32(fake_uni_entries); | |
1190 | cursor += sizeof(*mac_ptr) + fake_uni_entries * ETH_ALEN; | |
1191 | ||
1192 | /* | |
1193 | * Pack the multicast MAC addresses part for fake CVQ command. | |
1194 | * | |
1195 | * According to virtio_net_handle_mac(), QEMU doesn't verify the MAC | |
0a19d879 | 1196 | * addresses provided in CVQ command. Therefore, only the entries |
fee364e4 HJ |
1197 | * field need to be prepared in the CVQ command. |
1198 | */ | |
1199 | mac_ptr = out->iov_base + cursor; | |
1200 | mac_ptr->entries = cpu_to_le32(fake_mul_entries); | |
1201 | ||
1202 | /* | |
1203 | * Simulating QEMU poll a vdpa device used buffer | |
1204 | * for VIRTIO_NET_CTRL_MAC_TABLE_SET CVQ command | |
1205 | */ | |
1206 | return sizeof(*s->status); | |
1207 | } | |
1208 | ||
2df4dd31 EP |
1209 | /** |
1210 | * Validate and copy control virtqueue commands. | |
1211 | * | |
1212 | * Following QEMU guidelines, we offer a copy of the buffers to the device to | |
1213 | * prevent TOCTOU bugs. | |
bd907ae4 EP |
1214 | */ |
1215 | static int vhost_vdpa_net_handle_ctrl_avail(VhostShadowVirtqueue *svq, | |
1216 | VirtQueueElement *elem, | |
1217 | void *opaque) | |
1218 | { | |
2df4dd31 | 1219 | VhostVDPAState *s = opaque; |
be4278b6 | 1220 | size_t in_len; |
45c41018 | 1221 | const struct virtio_net_ctrl_hdr *ctrl; |
bd907ae4 | 1222 | virtio_net_ctrl_ack status = VIRTIO_NET_ERR; |
7a7f87e9 EP |
1223 | /* Out buffer sent to both the vdpa device and the device model */ |
1224 | struct iovec out = { | |
1225 | .iov_base = s->cvq_cmd_out_buffer, | |
1226 | }; | |
2df4dd31 EP |
1227 | /* in buffer used for device model */ |
1228 | const struct iovec in = { | |
1229 | .iov_base = &status, | |
1230 | .iov_len = sizeof(status), | |
1231 | }; | |
be4278b6 | 1232 | ssize_t dev_written = -EINVAL; |
2df4dd31 | 1233 | |
7a7f87e9 EP |
1234 | out.iov_len = iov_to_buf(elem->out_sg, elem->out_num, 0, |
1235 | s->cvq_cmd_out_buffer, | |
fee364e4 | 1236 | vhost_vdpa_net_cvq_cmd_page_len()); |
45c41018 HJ |
1237 | |
1238 | ctrl = s->cvq_cmd_out_buffer; | |
1239 | if (ctrl->class == VIRTIO_NET_CTRL_ANNOUNCE) { | |
3f9a3eeb EP |
1240 | /* |
1241 | * Guest announce capability is emulated by qemu, so don't forward to | |
1242 | * the device. | |
1243 | */ | |
1244 | dev_written = sizeof(status); | |
1245 | *s->status = VIRTIO_NET_OK; | |
fee364e4 HJ |
1246 | } else if (unlikely(ctrl->class == VIRTIO_NET_CTRL_MAC && |
1247 | ctrl->cmd == VIRTIO_NET_CTRL_MAC_TABLE_SET && | |
1248 | iov_size(elem->out_sg, elem->out_num) > out.iov_len)) { | |
1249 | /* | |
1250 | * Due to the size limitation of the out buffer sent to the vdpa device, | |
1251 | * which is determined by vhost_vdpa_net_cvq_cmd_page_len(), excessive | |
1252 | * MAC addresses set by the driver for the filter table can cause | |
1253 | * truncation of the CVQ command in QEMU. As a result, the vdpa device | |
1254 | * rejects the flawed CVQ command. | |
1255 | * | |
1256 | * Therefore, QEMU must handle this situation instead of sending | |
0a19d879 | 1257 | * the CVQ command directly. |
fee364e4 HJ |
1258 | */ |
1259 | dev_written = vhost_vdpa_net_excessive_mac_filter_cvq_add(s, elem, | |
1260 | &out); | |
1261 | if (unlikely(dev_written < 0)) { | |
1262 | goto out; | |
1263 | } | |
3f9a3eeb EP |
1264 | } else { |
1265 | dev_written = vhost_vdpa_net_cvq_add(s, out.iov_len, sizeof(status)); | |
1266 | if (unlikely(dev_written < 0)) { | |
1267 | goto out; | |
1268 | } | |
bd907ae4 EP |
1269 | } |
1270 | ||
bd907ae4 EP |
1271 | if (unlikely(dev_written < sizeof(status))) { |
1272 | error_report("Insufficient written data (%zu)", dev_written); | |
2df4dd31 EP |
1273 | goto out; |
1274 | } | |
1275 | ||
17fb889f | 1276 | if (*s->status != VIRTIO_NET_OK) { |
d45243bc | 1277 | goto out; |
2df4dd31 EP |
1278 | } |
1279 | ||
1280 | status = VIRTIO_NET_ERR; | |
7a7f87e9 | 1281 | virtio_net_handle_ctrl_iov(svq->vdev, &in, 1, &out, 1); |
2df4dd31 EP |
1282 | if (status != VIRTIO_NET_OK) { |
1283 | error_report("Bad CVQ processing in model"); | |
bd907ae4 EP |
1284 | } |
1285 | ||
1286 | out: | |
1287 | in_len = iov_from_buf(elem->in_sg, elem->in_num, 0, &status, | |
1288 | sizeof(status)); | |
1289 | if (unlikely(in_len < sizeof(status))) { | |
1290 | error_report("Bad device CVQ written length"); | |
1291 | } | |
1292 | vhost_svq_push_elem(svq, elem, MIN(in_len, sizeof(status))); | |
031b1aba HJ |
1293 | /* |
1294 | * `elem` belongs to vhost_vdpa_net_handle_ctrl_avail() only when | |
1295 | * the function successfully forwards the CVQ command, indicated | |
1296 | * by a non-negative value of `dev_written`. Otherwise, it still | |
1297 | * belongs to SVQ. | |
1298 | * This function should only free the `elem` when it owns. | |
1299 | */ | |
1300 | if (dev_written >= 0) { | |
1301 | g_free(elem); | |
1302 | } | |
be4278b6 | 1303 | return dev_written < 0 ? dev_written : 0; |
bd907ae4 EP |
1304 | } |
1305 | ||
1306 | static const VhostShadowVirtqueueOps vhost_vdpa_net_svq_ops = { | |
1307 | .avail_handler = vhost_vdpa_net_handle_ctrl_avail, | |
1308 | }; | |
1309 | ||
152128d6 EP |
1310 | /** |
1311 | * Probe if CVQ is isolated | |
1312 | * | |
1313 | * @device_fd The vdpa device fd | |
1314 | * @features Features offered by the device. | |
1315 | * @cvq_index The control vq pair index | |
1316 | * | |
1317 | * Returns <0 in case of failure, 0 if false and 1 if true. | |
1318 | */ | |
1319 | static int vhost_vdpa_probe_cvq_isolation(int device_fd, uint64_t features, | |
1320 | int cvq_index, Error **errp) | |
1321 | { | |
1322 | uint64_t backend_features; | |
1323 | int64_t cvq_group; | |
1324 | uint8_t status = VIRTIO_CONFIG_S_ACKNOWLEDGE | | |
1325 | VIRTIO_CONFIG_S_DRIVER | | |
1326 | VIRTIO_CONFIG_S_FEATURES_OK; | |
1327 | int r; | |
1328 | ||
1329 | ERRP_GUARD(); | |
1330 | ||
1331 | r = ioctl(device_fd, VHOST_GET_BACKEND_FEATURES, &backend_features); | |
1332 | if (unlikely(r < 0)) { | |
1333 | error_setg_errno(errp, errno, "Cannot get vdpa backend_features"); | |
1334 | return r; | |
1335 | } | |
1336 | ||
1337 | if (!(backend_features & BIT_ULL(VHOST_BACKEND_F_IOTLB_ASID))) { | |
1338 | return 0; | |
1339 | } | |
1340 | ||
1341 | r = ioctl(device_fd, VHOST_SET_FEATURES, &features); | |
1342 | if (unlikely(r)) { | |
1343 | error_setg_errno(errp, errno, "Cannot set features"); | |
1344 | } | |
1345 | ||
1346 | r = ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status); | |
1347 | if (unlikely(r)) { | |
1348 | error_setg_errno(errp, -r, "Cannot set device features"); | |
1349 | goto out; | |
1350 | } | |
1351 | ||
1352 | cvq_group = vhost_vdpa_get_vring_group(device_fd, cvq_index, errp); | |
1353 | if (unlikely(cvq_group < 0)) { | |
1354 | if (cvq_group != -ENOTSUP) { | |
1355 | r = cvq_group; | |
1356 | goto out; | |
1357 | } | |
1358 | ||
1359 | /* | |
1360 | * The kernel report VHOST_BACKEND_F_IOTLB_ASID if the vdpa frontend | |
1361 | * support ASID even if the parent driver does not. The CVQ cannot be | |
1362 | * isolated in this case. | |
1363 | */ | |
1364 | error_free(*errp); | |
1365 | *errp = NULL; | |
1366 | r = 0; | |
1367 | goto out; | |
1368 | } | |
1369 | ||
1370 | for (int i = 0; i < cvq_index; ++i) { | |
1371 | int64_t group = vhost_vdpa_get_vring_group(device_fd, i, errp); | |
1372 | if (unlikely(group < 0)) { | |
1373 | r = group; | |
1374 | goto out; | |
1375 | } | |
1376 | ||
1377 | if (group == (int64_t)cvq_group) { | |
1378 | r = 0; | |
1379 | goto out; | |
1380 | } | |
1381 | } | |
1382 | ||
1383 | r = 1; | |
1384 | ||
1385 | out: | |
1386 | status = 0; | |
1387 | ioctl(device_fd, VHOST_VDPA_SET_STATUS, &status); | |
1388 | return r; | |
1389 | } | |
1390 | ||
654790b6 | 1391 | static NetClientState *net_vhost_vdpa_init(NetClientState *peer, |
a585fad2 EP |
1392 | const char *device, |
1393 | const char *name, | |
1394 | int vdpa_device_fd, | |
1395 | int queue_pair_index, | |
1396 | int nvqs, | |
1397 | bool is_datapath, | |
1398 | bool svq, | |
5c1ebd4c | 1399 | struct vhost_vdpa_iova_range iova_range, |
152128d6 EP |
1400 | uint64_t features, |
1401 | Error **errp) | |
1e0a84ea CL |
1402 | { |
1403 | NetClientState *nc = NULL; | |
1404 | VhostVDPAState *s; | |
1e0a84ea CL |
1405 | int ret = 0; |
1406 | assert(name); | |
152128d6 EP |
1407 | int cvq_isolated; |
1408 | ||
40237840 JW |
1409 | if (is_datapath) { |
1410 | nc = qemu_new_net_client(&net_vhost_vdpa_info, peer, device, | |
1411 | name); | |
1412 | } else { | |
152128d6 EP |
1413 | cvq_isolated = vhost_vdpa_probe_cvq_isolation(vdpa_device_fd, features, |
1414 | queue_pair_index * 2, | |
1415 | errp); | |
1416 | if (unlikely(cvq_isolated < 0)) { | |
1417 | return NULL; | |
1418 | } | |
1419 | ||
f8972b56 | 1420 | nc = qemu_new_net_control_client(&net_vhost_vdpa_cvq_info, peer, |
40237840 JW |
1421 | device, name); |
1422 | } | |
53b85d95 | 1423 | qemu_set_info_str(nc, TYPE_VHOST_VDPA); |
1e0a84ea | 1424 | s = DO_UPCAST(VhostVDPAState, nc, nc); |
7327813d | 1425 | |
1e0a84ea | 1426 | s->vhost_vdpa.device_fd = vdpa_device_fd; |
40237840 | 1427 | s->vhost_vdpa.index = queue_pair_index; |
7f211a28 | 1428 | s->always_svq = svq; |
69498430 | 1429 | s->migration_state.notify = vdpa_net_migration_state_notifier; |
1576dbb5 | 1430 | s->vhost_vdpa.shadow_vqs_enabled = svq; |
a585fad2 | 1431 | s->vhost_vdpa.iova_range = iova_range; |
6188d78a | 1432 | s->vhost_vdpa.shadow_data = svq; |
5c1ebd4c EP |
1433 | if (queue_pair_index == 0) { |
1434 | vhost_vdpa_net_valid_svq_features(features, | |
1435 | &s->vhost_vdpa.migration_blocker); | |
1436 | } else if (!is_datapath) { | |
babf8b87 EP |
1437 | s->cvq_cmd_out_buffer = mmap(NULL, vhost_vdpa_net_cvq_cmd_page_len(), |
1438 | PROT_READ | PROT_WRITE, | |
1439 | MAP_SHARED | MAP_ANONYMOUS, -1, 0); | |
1440 | s->status = mmap(NULL, vhost_vdpa_net_cvq_cmd_page_len(), | |
1441 | PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANONYMOUS, | |
1442 | -1, 0); | |
2df4dd31 | 1443 | |
bd907ae4 EP |
1444 | s->vhost_vdpa.shadow_vq_ops = &vhost_vdpa_net_svq_ops; |
1445 | s->vhost_vdpa.shadow_vq_ops_opaque = s; | |
152128d6 | 1446 | s->cvq_isolated = cvq_isolated; |
9c363cf6 EP |
1447 | |
1448 | /* | |
8bc0049e EP |
1449 | * TODO: We cannot migrate devices with CVQ and no x-svq enabled as |
1450 | * there is no way to set the device state (MAC, MQ, etc) before | |
1451 | * starting the datapath. | |
9c363cf6 EP |
1452 | * |
1453 | * Migration blocker ownership now belongs to s->vhost_vdpa. | |
1454 | */ | |
8bc0049e EP |
1455 | if (!svq) { |
1456 | error_setg(&s->vhost_vdpa.migration_blocker, | |
1457 | "net vdpa cannot migrate with CVQ feature"); | |
1458 | } | |
bd907ae4 | 1459 | } |
40237840 | 1460 | ret = vhost_vdpa_add(nc, (void *)&s->vhost_vdpa, queue_pair_index, nvqs); |
74af5eec | 1461 | if (ret) { |
74af5eec | 1462 | qemu_del_net_client(nc); |
654790b6 | 1463 | return NULL; |
74af5eec | 1464 | } |
654790b6 | 1465 | return nc; |
1e0a84ea CL |
1466 | } |
1467 | ||
8170ab3f EP |
1468 | static int vhost_vdpa_get_features(int fd, uint64_t *features, Error **errp) |
1469 | { | |
1470 | int ret = ioctl(fd, VHOST_GET_FEATURES, features); | |
1471 | if (unlikely(ret < 0)) { | |
1472 | error_setg_errno(errp, errno, | |
1473 | "Fail to query features from vhost-vDPA device"); | |
1474 | } | |
1475 | return ret; | |
1476 | } | |
1477 | ||
1478 | static int vhost_vdpa_get_max_queue_pairs(int fd, uint64_t features, | |
1479 | int *has_cvq, Error **errp) | |
40237840 JW |
1480 | { |
1481 | unsigned long config_size = offsetof(struct vhost_vdpa_config, buf); | |
cd523a41 | 1482 | g_autofree struct vhost_vdpa_config *config = NULL; |
40237840 | 1483 | __virtio16 *max_queue_pairs; |
40237840 JW |
1484 | int ret; |
1485 | ||
40237840 JW |
1486 | if (features & (1 << VIRTIO_NET_F_CTRL_VQ)) { |
1487 | *has_cvq = 1; | |
1488 | } else { | |
1489 | *has_cvq = 0; | |
1490 | } | |
1491 | ||
1492 | if (features & (1 << VIRTIO_NET_F_MQ)) { | |
1493 | config = g_malloc0(config_size + sizeof(*max_queue_pairs)); | |
1494 | config->off = offsetof(struct virtio_net_config, max_virtqueue_pairs); | |
1495 | config->len = sizeof(*max_queue_pairs); | |
1496 | ||
1497 | ret = ioctl(fd, VHOST_VDPA_GET_CONFIG, config); | |
1498 | if (ret) { | |
1499 | error_setg(errp, "Fail to get config from vhost-vDPA device"); | |
1500 | return -ret; | |
1501 | } | |
1502 | ||
1503 | max_queue_pairs = (__virtio16 *)&config->buf; | |
1504 | ||
1505 | return lduw_le_p(max_queue_pairs); | |
1506 | } | |
1507 | ||
1508 | return 1; | |
1509 | } | |
1510 | ||
1e0a84ea CL |
1511 | int net_init_vhost_vdpa(const Netdev *netdev, const char *name, |
1512 | NetClientState *peer, Error **errp) | |
1513 | { | |
1514 | const NetdevVhostVDPAOptions *opts; | |
8170ab3f | 1515 | uint64_t features; |
654790b6 | 1516 | int vdpa_device_fd; |
eb3cb751 | 1517 | g_autofree NetClientState **ncs = NULL; |
a585fad2 | 1518 | struct vhost_vdpa_iova_range iova_range; |
eb3cb751 | 1519 | NetClientState *nc; |
aed5da45 | 1520 | int queue_pairs, r, i = 0, has_cvq = 0; |
1e0a84ea CL |
1521 | |
1522 | assert(netdev->type == NET_CLIENT_DRIVER_VHOST_VDPA); | |
1523 | opts = &netdev->u.vhost_vdpa; | |
7480874a | 1524 | if (!opts->vhostdev && !opts->vhostfd) { |
8801ccd0 SWL |
1525 | error_setg(errp, |
1526 | "vhost-vdpa: neither vhostdev= nor vhostfd= was specified"); | |
c8295404 EP |
1527 | return -1; |
1528 | } | |
7327813d | 1529 | |
7480874a | 1530 | if (opts->vhostdev && opts->vhostfd) { |
8801ccd0 SWL |
1531 | error_setg(errp, |
1532 | "vhost-vdpa: vhostdev= and vhostfd= are mutually exclusive"); | |
1533 | return -1; | |
1534 | } | |
1535 | ||
7480874a | 1536 | if (opts->vhostdev) { |
8801ccd0 SWL |
1537 | vdpa_device_fd = qemu_open(opts->vhostdev, O_RDWR, errp); |
1538 | if (vdpa_device_fd == -1) { | |
1539 | return -errno; | |
1540 | } | |
5107fd3e PM |
1541 | } else { |
1542 | /* has_vhostfd */ | |
8801ccd0 SWL |
1543 | vdpa_device_fd = monitor_fd_param(monitor_cur(), opts->vhostfd, errp); |
1544 | if (vdpa_device_fd == -1) { | |
1545 | error_prepend(errp, "vhost-vdpa: unable to parse vhostfd: "); | |
1546 | return -1; | |
1547 | } | |
7327813d JW |
1548 | } |
1549 | ||
8170ab3f EP |
1550 | r = vhost_vdpa_get_features(vdpa_device_fd, &features, errp); |
1551 | if (unlikely(r < 0)) { | |
aed5da45 | 1552 | goto err; |
8170ab3f EP |
1553 | } |
1554 | ||
1555 | queue_pairs = vhost_vdpa_get_max_queue_pairs(vdpa_device_fd, features, | |
40237840 JW |
1556 | &has_cvq, errp); |
1557 | if (queue_pairs < 0) { | |
7327813d | 1558 | qemu_close(vdpa_device_fd); |
40237840 JW |
1559 | return queue_pairs; |
1560 | } | |
1561 | ||
bf7a2ad8 LM |
1562 | r = vhost_vdpa_get_iova_range(vdpa_device_fd, &iova_range); |
1563 | if (unlikely(r < 0)) { | |
1564 | error_setg(errp, "vhost-vdpa: get iova range failed: %s", | |
1565 | strerror(-r)); | |
1566 | goto err; | |
1567 | } | |
1568 | ||
00ef422e EP |
1569 | if (opts->x_svq && !vhost_vdpa_net_valid_svq_features(features, errp)) { |
1570 | goto err; | |
1576dbb5 EP |
1571 | } |
1572 | ||
40237840 JW |
1573 | ncs = g_malloc0(sizeof(*ncs) * queue_pairs); |
1574 | ||
1575 | for (i = 0; i < queue_pairs; i++) { | |
1576 | ncs[i] = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name, | |
1576dbb5 | 1577 | vdpa_device_fd, i, 2, true, opts->x_svq, |
152128d6 | 1578 | iova_range, features, errp); |
40237840 JW |
1579 | if (!ncs[i]) |
1580 | goto err; | |
7327813d JW |
1581 | } |
1582 | ||
40237840 JW |
1583 | if (has_cvq) { |
1584 | nc = net_vhost_vdpa_init(peer, TYPE_VHOST_VDPA, name, | |
1576dbb5 | 1585 | vdpa_device_fd, i, 1, false, |
152128d6 | 1586 | opts->x_svq, iova_range, features, errp); |
40237840 JW |
1587 | if (!nc) |
1588 | goto err; | |
1589 | } | |
1590 | ||
654790b6 | 1591 | return 0; |
40237840 JW |
1592 | |
1593 | err: | |
1594 | if (i) { | |
9bd05507 SWL |
1595 | for (i--; i >= 0; i--) { |
1596 | qemu_del_net_client(ncs[i]); | |
1597 | } | |
40237840 | 1598 | } |
1576dbb5 | 1599 | |
40237840 | 1600 | qemu_close(vdpa_device_fd); |
40237840 JW |
1601 | |
1602 | return -1; | |
1e0a84ea | 1603 | } |