| 1 | /* |
| 2 | * vhost shadow virtqueue |
| 3 | * |
| 4 | * SPDX-FileCopyrightText: Red Hat, Inc. 2021 |
| 5 | * SPDX-FileContributor: Author: Eugenio Pérez <eperezma@redhat.com> |
| 6 | * |
| 7 | * SPDX-License-Identifier: GPL-2.0-or-later |
| 8 | */ |
| 9 | |
| 10 | #include "qemu/osdep.h" |
| 11 | #include "hw/virtio/vhost-shadow-virtqueue.h" |
| 12 | |
| 13 | #include "qemu/error-report.h" |
| 14 | #include "qapi/error.h" |
| 15 | #include "qemu/iov.h" |
| 16 | #include "qemu/main-loop.h" |
| 17 | #include "qemu/log.h" |
| 18 | #include "qemu/memalign.h" |
| 19 | #include "linux-headers/linux/vhost.h" |
| 20 | |
| 21 | #define VIRTIO_RING_NOT_IN_BATCH UINT16_MAX |
| 22 | |
| 23 | /** |
| 24 | * Validate the transport device features that both guests can use with the SVQ |
| 25 | * and SVQs can use with the device. |
| 26 | * |
| 27 | * @dev_features: The features |
| 28 | * @errp: Error pointer |
| 29 | */ |
| 30 | bool vhost_svq_valid_features(uint64_t features, Error **errp) |
| 31 | { |
| 32 | bool ok = true; |
| 33 | uint64_t svq_features = features; |
| 34 | |
| 35 | for (uint64_t b = VIRTIO_TRANSPORT_F_START; b <= VIRTIO_TRANSPORT_F_END; |
| 36 | ++b) { |
| 37 | switch (b) { |
| 38 | case VIRTIO_F_ANY_LAYOUT: |
| 39 | case VIRTIO_RING_F_EVENT_IDX: |
| 40 | case VIRTIO_RING_F_INDIRECT_DESC: |
| 41 | case VIRTIO_F_IN_ORDER: |
| 42 | continue; |
| 43 | |
| 44 | case VIRTIO_F_ACCESS_PLATFORM: |
| 45 | /* SVQ trust in the host's IOMMU to translate addresses */ |
| 46 | case VIRTIO_F_VERSION_1: |
| 47 | /* SVQ trust that the guest vring is little endian */ |
| 48 | if (!(svq_features & BIT_ULL(b))) { |
| 49 | svq_features |= BIT_ULL(b); |
| 50 | ok = false; |
| 51 | } |
| 52 | continue; |
| 53 | |
| 54 | default: |
| 55 | if (svq_features & BIT_ULL(b)) { |
| 56 | svq_features &= ~BIT_ULL(b); |
| 57 | ok = false; |
| 58 | } |
| 59 | } |
| 60 | } |
| 61 | |
| 62 | if (!ok) { |
| 63 | error_setg(errp, "SVQ Invalid device feature flags, offer: 0x%"PRIx64 |
| 64 | ", ok: 0x%"PRIx64, features, svq_features); |
| 65 | } |
| 66 | return ok; |
| 67 | } |
| 68 | |
| 69 | /** |
| 70 | * Number of descriptors that the SVQ can make available from the guest. |
| 71 | * |
| 72 | * @svq: The svq |
| 73 | */ |
| 74 | uint16_t vhost_svq_available_slots(const VhostShadowVirtqueue *svq) |
| 75 | { |
| 76 | return svq->num_free; |
| 77 | } |
| 78 | |
| 79 | /** |
| 80 | * Translate addresses between the qemu's virtual address and the SVQ IOVA |
| 81 | * |
| 82 | * @svq: Shadow VirtQueue |
| 83 | * @vaddr: Translated IOVA addresses |
| 84 | * @iovec: Source qemu's VA addresses |
| 85 | * @num: Length of iovec and minimum length of vaddr |
| 86 | * @gpas: Descriptors' GPAs, if backed by guest memory |
| 87 | */ |
| 88 | static bool vhost_svq_translate_addr(const VhostShadowVirtqueue *svq, |
| 89 | hwaddr *addrs, const struct iovec *iovec, |
| 90 | size_t num, const hwaddr *gpas) |
| 91 | { |
| 92 | if (num == 0) { |
| 93 | return true; |
| 94 | } |
| 95 | |
| 96 | for (size_t i = 0; i < num; ++i) { |
| 97 | Int128 needle_last, map_last; |
| 98 | size_t off; |
| 99 | const DMAMap *map; |
| 100 | DMAMap needle; |
| 101 | |
| 102 | /* Check if the descriptor is backed by guest memory */ |
| 103 | if (gpas) { |
| 104 | /* Search the GPA->IOVA tree */ |
| 105 | needle = (DMAMap) { |
| 106 | .translated_addr = gpas[i], |
| 107 | .size = iovec[i].iov_len, |
| 108 | }; |
| 109 | map = vhost_iova_tree_find_gpa(svq->iova_tree, &needle); |
| 110 | } else { |
| 111 | /* Search the IOVA->HVA tree */ |
| 112 | needle = (DMAMap) { |
| 113 | .translated_addr = (hwaddr)(uintptr_t)iovec[i].iov_base, |
| 114 | .size = iovec[i].iov_len, |
| 115 | }; |
| 116 | map = vhost_iova_tree_find_iova(svq->iova_tree, &needle); |
| 117 | } |
| 118 | |
| 119 | /* |
| 120 | * Map cannot be NULL since iova map contains all guest space and |
| 121 | * qemu already has a physical address mapped |
| 122 | */ |
| 123 | if (unlikely(!map)) { |
| 124 | qemu_log_mask(LOG_GUEST_ERROR, |
| 125 | "Invalid address 0x%"HWADDR_PRIx" given by guest", |
| 126 | needle.translated_addr); |
| 127 | return false; |
| 128 | } |
| 129 | |
| 130 | off = needle.translated_addr - map->translated_addr; |
| 131 | addrs[i] = map->iova + off; |
| 132 | |
| 133 | needle_last = int128_add(int128_make64(needle.translated_addr), |
| 134 | int128_makes64(iovec[i].iov_len - 1)); |
| 135 | map_last = int128_make64(map->translated_addr + map->size); |
| 136 | if (unlikely(int128_gt(needle_last, map_last))) { |
| 137 | qemu_log_mask(LOG_GUEST_ERROR, |
| 138 | "Guest buffer expands over iova range"); |
| 139 | return false; |
| 140 | } |
| 141 | } |
| 142 | |
| 143 | return true; |
| 144 | } |
| 145 | |
| 146 | /** |
| 147 | * Get the next descriptor in the chain in SVQ vring from a descriptor id |
| 148 | * |
| 149 | * @svq Shadow Virtqueue |
| 150 | * @id ID of the descriptor |
| 151 | * |
| 152 | * Return the id of the next descriptor. |
| 153 | */ |
| 154 | static uint16_t vhost_svq_next_desc(const VhostShadowVirtqueue *svq, |
| 155 | uint16_t id) |
| 156 | { |
| 157 | if (virtio_vdev_has_feature(svq->vdev, VIRTIO_F_IN_ORDER)) { |
| 158 | return (id == svq->vring.num) ? 0 : ++id; |
| 159 | } else { |
| 160 | return svq->desc_state[id].next; |
| 161 | } |
| 162 | } |
| 163 | |
| 164 | /** |
| 165 | * Updates the SVQ free_head member after adding them to the SVQ avail ring. |
| 166 | * The new free_head is the next descriptor that SVQ will make available by |
| 167 | * forwarding a new guest descriptor. |
| 168 | * |
| 169 | * @svq Shadow Virtqueue |
| 170 | * @num Number of descriptors added |
| 171 | * @id ID of the last descriptor added to the SVQ avail ring. |
| 172 | */ |
| 173 | static void vhost_svq_update_free_head(VhostShadowVirtqueue *svq, |
| 174 | size_t num, uint16_t id) |
| 175 | { |
| 176 | if (virtio_vdev_has_feature(svq->vdev, VIRTIO_F_IN_ORDER)) { |
| 177 | svq->free_head += num; |
| 178 | if (svq->free_head >= svq->vring.num) { |
| 179 | svq->free_head -= svq->vring.num; |
| 180 | } |
| 181 | } else { |
| 182 | svq->free_head = vhost_svq_next_desc(svq, id); |
| 183 | } |
| 184 | } |
| 185 | |
| 186 | /** |
| 187 | * Write descriptors to SVQ vring |
| 188 | * |
| 189 | * @svq: The shadow virtqueue |
| 190 | * @sg: Cache for hwaddr |
| 191 | * @iovec: The iovec from the guest |
| 192 | * @num: iovec length |
| 193 | * @addr: Descriptors' GPAs, if backed by guest memory |
| 194 | * @more_descs: True if more descriptors come in the chain |
| 195 | * @write: True if they are writeable descriptors |
| 196 | * |
| 197 | * Return true if success, false otherwise and print error. |
| 198 | */ |
| 199 | static bool vhost_svq_vring_write_descs(VhostShadowVirtqueue *svq, hwaddr *sg, |
| 200 | const struct iovec *iovec, size_t num, |
| 201 | const hwaddr *addr, bool more_descs, |
| 202 | bool write) |
| 203 | { |
| 204 | uint16_t i = svq->free_head, last = svq->free_head; |
| 205 | unsigned n; |
| 206 | uint16_t flags = write ? cpu_to_le16(VRING_DESC_F_WRITE) : 0; |
| 207 | vring_desc_t *descs = svq->vring.desc; |
| 208 | bool ok; |
| 209 | |
| 210 | if (num == 0) { |
| 211 | return true; |
| 212 | } |
| 213 | |
| 214 | ok = vhost_svq_translate_addr(svq, sg, iovec, num, addr); |
| 215 | if (unlikely(!ok)) { |
| 216 | return false; |
| 217 | } |
| 218 | |
| 219 | for (n = 0; n < num; n++) { |
| 220 | uint16_t next = vhost_svq_next_desc(svq, i); |
| 221 | |
| 222 | if (more_descs || (n + 1 < num)) { |
| 223 | descs[i].flags = flags | cpu_to_le16(VRING_DESC_F_NEXT); |
| 224 | descs[i].next = cpu_to_le16(next); |
| 225 | } else { |
| 226 | descs[i].flags = flags; |
| 227 | } |
| 228 | descs[i].addr = cpu_to_le64(sg[n]); |
| 229 | descs[i].len = cpu_to_le32(iovec[n].iov_len); |
| 230 | |
| 231 | last = i; |
| 232 | i = next; |
| 233 | } |
| 234 | |
| 235 | vhost_svq_update_free_head(svq, num, last); |
| 236 | return true; |
| 237 | } |
| 238 | |
| 239 | static bool vhost_svq_add_split(VhostShadowVirtqueue *svq, |
| 240 | const struct iovec *out_sg, size_t out_num, |
| 241 | const hwaddr *out_addr, |
| 242 | const struct iovec *in_sg, size_t in_num, |
| 243 | const hwaddr *in_addr, unsigned *head) |
| 244 | { |
| 245 | unsigned avail_idx; |
| 246 | vring_avail_t *avail = svq->vring.avail; |
| 247 | bool ok; |
| 248 | g_autofree hwaddr *sgs = g_new(hwaddr, MAX(out_num, in_num)); |
| 249 | |
| 250 | *head = svq->free_head; |
| 251 | |
| 252 | /* We need some descriptors here */ |
| 253 | if (unlikely(!out_num && !in_num)) { |
| 254 | qemu_log_mask(LOG_GUEST_ERROR, |
| 255 | "Guest provided element with no descriptors"); |
| 256 | return false; |
| 257 | } |
| 258 | |
| 259 | ok = vhost_svq_vring_write_descs(svq, sgs, out_sg, out_num, out_addr, |
| 260 | in_num > 0, false); |
| 261 | if (unlikely(!ok)) { |
| 262 | return false; |
| 263 | } |
| 264 | |
| 265 | ok = vhost_svq_vring_write_descs(svq, sgs, in_sg, in_num, in_addr, false, |
| 266 | true); |
| 267 | if (unlikely(!ok)) { |
| 268 | return false; |
| 269 | } |
| 270 | |
| 271 | /* |
| 272 | * Put the entry in the available array (but don't update avail->idx until |
| 273 | * they do sync). |
| 274 | */ |
| 275 | avail_idx = svq->shadow_avail_idx & (svq->vring.num - 1); |
| 276 | avail->ring[avail_idx] = cpu_to_le16(*head); |
| 277 | svq->shadow_avail_idx++; |
| 278 | |
| 279 | /* Update the avail index after write the descriptor */ |
| 280 | smp_wmb(); |
| 281 | avail->idx = cpu_to_le16(svq->shadow_avail_idx); |
| 282 | |
| 283 | return true; |
| 284 | } |
| 285 | |
| 286 | static void vhost_svq_kick(VhostShadowVirtqueue *svq) |
| 287 | { |
| 288 | bool needs_kick; |
| 289 | |
| 290 | /* |
| 291 | * We need to expose the available array entries before checking the used |
| 292 | * flags |
| 293 | */ |
| 294 | smp_mb(); |
| 295 | |
| 296 | if (virtio_vdev_has_feature(svq->vdev, VIRTIO_RING_F_EVENT_IDX)) { |
| 297 | uint16_t avail_event = le16_to_cpu( |
| 298 | *(uint16_t *)(&svq->vring.used->ring[svq->vring.num])); |
| 299 | needs_kick = vring_need_event(avail_event, svq->shadow_avail_idx, svq->shadow_avail_idx - 1); |
| 300 | } else { |
| 301 | needs_kick = |
| 302 | !(svq->vring.used->flags & cpu_to_le16(VRING_USED_F_NO_NOTIFY)); |
| 303 | } |
| 304 | |
| 305 | if (!needs_kick) { |
| 306 | return; |
| 307 | } |
| 308 | |
| 309 | event_notifier_set(&svq->hdev_kick); |
| 310 | } |
| 311 | |
| 312 | /** |
| 313 | * Add an element to a SVQ. |
| 314 | * |
| 315 | * Return -EINVAL if element is invalid, -ENOSPC if dev queue is full |
| 316 | */ |
| 317 | int vhost_svq_add(VhostShadowVirtqueue *svq, const struct iovec *out_sg, |
| 318 | size_t out_num, const hwaddr *out_addr, |
| 319 | const struct iovec *in_sg, size_t in_num, |
| 320 | const hwaddr *in_addr, VirtQueueElement *elem) |
| 321 | { |
| 322 | unsigned qemu_head; |
| 323 | unsigned ndescs = in_num + out_num; |
| 324 | bool ok; |
| 325 | |
| 326 | if (unlikely(ndescs > vhost_svq_available_slots(svq))) { |
| 327 | return -ENOSPC; |
| 328 | } |
| 329 | |
| 330 | ok = vhost_svq_add_split(svq, out_sg, out_num, out_addr, in_sg, in_num, |
| 331 | in_addr, &qemu_head); |
| 332 | if (unlikely(!ok)) { |
| 333 | return -EINVAL; |
| 334 | } |
| 335 | |
| 336 | svq->num_free -= ndescs; |
| 337 | svq->desc_state[qemu_head].elem = elem; |
| 338 | svq->desc_state[qemu_head].ndescs = ndescs; |
| 339 | if (virtio_vdev_has_feature(svq->vdev, VIRTIO_F_IN_ORDER)) { |
| 340 | svq->desc_state[qemu_head].in_bytes = iov_size(in_sg, in_num); |
| 341 | } |
| 342 | vhost_svq_kick(svq); |
| 343 | return 0; |
| 344 | } |
| 345 | |
| 346 | /* Convenience wrapper to add a guest's element to SVQ */ |
| 347 | static int vhost_svq_add_element(VhostShadowVirtqueue *svq, |
| 348 | VirtQueueElement *elem) |
| 349 | { |
| 350 | return vhost_svq_add(svq, elem->out_sg, elem->out_num, elem->out_addr, |
| 351 | elem->in_sg, elem->in_num, elem->in_addr, elem); |
| 352 | } |
| 353 | |
| 354 | /** |
| 355 | * Forward available buffers. |
| 356 | * |
| 357 | * @svq: Shadow VirtQueue |
| 358 | * |
| 359 | * Note that this function does not guarantee that all guest's available |
| 360 | * buffers are available to the device in SVQ avail ring. The guest may have |
| 361 | * exposed a GPA / GIOVA contiguous buffer, but it may not be contiguous in |
| 362 | * qemu vaddr. |
| 363 | * |
| 364 | * If that happens, guest's kick notifications will be disabled until the |
| 365 | * device uses some buffers. |
| 366 | */ |
| 367 | static void vhost_handle_guest_kick(VhostShadowVirtqueue *svq) |
| 368 | { |
| 369 | /* Clear event notifier */ |
| 370 | event_notifier_test_and_clear(&svq->svq_kick); |
| 371 | |
| 372 | /* Forward to the device as many available buffers as possible */ |
| 373 | do { |
| 374 | virtio_queue_set_notification(svq->vq, false); |
| 375 | |
| 376 | while (true) { |
| 377 | g_autofree VirtQueueElement *elem = NULL; |
| 378 | int r; |
| 379 | |
| 380 | if (svq->next_guest_avail_elem) { |
| 381 | elem = g_steal_pointer(&svq->next_guest_avail_elem); |
| 382 | } else { |
| 383 | elem = virtqueue_pop(svq->vq, sizeof(*elem)); |
| 384 | } |
| 385 | |
| 386 | if (!elem) { |
| 387 | break; |
| 388 | } |
| 389 | |
| 390 | if (svq->ops) { |
| 391 | r = svq->ops->avail_handler(svq, elem, svq->ops_opaque); |
| 392 | } else { |
| 393 | r = vhost_svq_add_element(svq, elem); |
| 394 | } |
| 395 | if (unlikely(r != 0)) { |
| 396 | if (r == -ENOSPC) { |
| 397 | /* |
| 398 | * This condition is possible since a contiguous buffer in |
| 399 | * GPA does not imply a contiguous buffer in qemu's VA |
| 400 | * scatter-gather segments. If that happens, the buffer |
| 401 | * exposed to the device needs to be a chain of descriptors |
| 402 | * at this moment. |
| 403 | * |
| 404 | * SVQ cannot hold more available buffers if we are here: |
| 405 | * queue the current guest descriptor and ignore kicks |
| 406 | * until some elements are used. |
| 407 | */ |
| 408 | svq->next_guest_avail_elem = g_steal_pointer(&elem); |
| 409 | } |
| 410 | |
| 411 | /* VQ is full or broken, just return and ignore kicks */ |
| 412 | return; |
| 413 | } |
| 414 | /* elem belongs to SVQ or external caller now */ |
| 415 | elem = NULL; |
| 416 | } |
| 417 | |
| 418 | virtio_queue_set_notification(svq->vq, true); |
| 419 | } while (!virtio_queue_empty(svq->vq)); |
| 420 | } |
| 421 | |
| 422 | /** |
| 423 | * Handle guest's kick. |
| 424 | * |
| 425 | * @n: guest kick event notifier, the one that guest set to notify svq. |
| 426 | */ |
| 427 | static void vhost_handle_guest_kick_notifier(EventNotifier *n) |
| 428 | { |
| 429 | VhostShadowVirtqueue *svq = container_of(n, VhostShadowVirtqueue, svq_kick); |
| 430 | event_notifier_test_and_clear(n); |
| 431 | vhost_handle_guest_kick(svq); |
| 432 | } |
| 433 | |
| 434 | static bool vhost_svq_more_used(VhostShadowVirtqueue *svq) |
| 435 | { |
| 436 | uint16_t *used_idx = &svq->vring.used->idx; |
| 437 | |
| 438 | if (virtio_vdev_has_feature(svq->vdev, VIRTIO_F_IN_ORDER) && |
| 439 | svq->batch_last.id != VIRTIO_RING_NOT_IN_BATCH) { |
| 440 | return true; |
| 441 | } |
| 442 | |
| 443 | if (svq->last_used_idx != svq->shadow_used_idx) { |
| 444 | return true; |
| 445 | } |
| 446 | |
| 447 | svq->shadow_used_idx = le16_to_cpu(*(volatile uint16_t *)used_idx); |
| 448 | |
| 449 | return svq->last_used_idx != svq->shadow_used_idx; |
| 450 | } |
| 451 | |
| 452 | /** |
| 453 | * Enable vhost device calls after disable them. |
| 454 | * |
| 455 | * @svq: The svq |
| 456 | * |
| 457 | * It returns false if there are pending used buffers from the vhost device, |
| 458 | * avoiding the possible races between SVQ checking for more work and enabling |
| 459 | * callbacks. True if SVQ used vring has no more pending buffers. |
| 460 | */ |
| 461 | static bool vhost_svq_enable_notification(VhostShadowVirtqueue *svq) |
| 462 | { |
| 463 | if (virtio_vdev_has_feature(svq->vdev, VIRTIO_RING_F_EVENT_IDX)) { |
| 464 | uint16_t *used_event = (uint16_t *)&svq->vring.avail->ring[svq->vring.num]; |
| 465 | *used_event = cpu_to_le16(svq->shadow_used_idx); |
| 466 | } else { |
| 467 | svq->vring.avail->flags &= ~cpu_to_le16(VRING_AVAIL_F_NO_INTERRUPT); |
| 468 | } |
| 469 | |
| 470 | /* Make sure the event is enabled before the read of used_idx */ |
| 471 | smp_mb(); |
| 472 | return !vhost_svq_more_used(svq); |
| 473 | } |
| 474 | |
| 475 | static void vhost_svq_disable_notification(VhostShadowVirtqueue *svq) |
| 476 | { |
| 477 | /* |
| 478 | * No need to disable notification in the event idx case, since used event |
| 479 | * index is already an index too far away. |
| 480 | */ |
| 481 | if (!virtio_vdev_has_feature(svq->vdev, VIRTIO_RING_F_EVENT_IDX)) { |
| 482 | svq->vring.avail->flags |= cpu_to_le16(VRING_AVAIL_F_NO_INTERRUPT); |
| 483 | } |
| 484 | } |
| 485 | |
| 486 | /* |
| 487 | * Gets the next buffer id and moves forward the used idx, so the next time |
| 488 | * SVQ calls this function will get the next one. |
| 489 | * |
| 490 | * @svq: Shadow VirtQueue |
| 491 | * @len: Consumed length by the device. |
| 492 | * |
| 493 | * Return the next descriptor consumed by the device. |
| 494 | */ |
| 495 | static uint16_t vhost_svq_get_last_used_split(VhostShadowVirtqueue *svq, |
| 496 | uint32_t *len) |
| 497 | { |
| 498 | const vring_used_t *used = svq->vring.used; |
| 499 | uint16_t last_used = svq->last_used_idx++ & (svq->vring.num - 1); |
| 500 | |
| 501 | *len = le32_to_cpu(used->ring[last_used].len); |
| 502 | return le32_to_cpu(used->ring[last_used].id); |
| 503 | } |
| 504 | |
| 505 | /* |
| 506 | * Gets the next buffer id and moves forward the used idx, so the next time |
| 507 | * SVQ calls this function will get the next one. IN_ORDER version |
| 508 | * |
| 509 | * @svq: Shadow VirtQueue |
| 510 | * @len: Consumed length by the device. |
| 511 | * |
| 512 | * Return the next descriptor consumed by the device. |
| 513 | */ |
| 514 | static int32_t vhost_svq_get_last_used_split_in_order( |
| 515 | VhostShadowVirtqueue *svq, |
| 516 | uint32_t *len) |
| 517 | { |
| 518 | unsigned num = svq->vring.num; |
| 519 | const vring_used_t *used = svq->vring.used; |
| 520 | uint16_t last_used = svq->last_used & (num - 1); |
| 521 | uint16_t last_used_idx = svq->last_used_idx & (num - 1); |
| 522 | |
| 523 | if (svq->batch_last.id == VIRTIO_RING_NOT_IN_BATCH) { |
| 524 | svq->batch_last.id = le32_to_cpu(used->ring[last_used_idx].id); |
| 525 | svq->batch_last.len = le32_to_cpu(used->ring[last_used_idx].len); |
| 526 | } |
| 527 | |
| 528 | if (unlikely(last_used >= num)) { |
| 529 | qemu_log_mask(LOG_GUEST_ERROR, "Device %s says index %u is used", |
| 530 | svq->vdev->name, last_used); |
| 531 | return -1; |
| 532 | } |
| 533 | |
| 534 | if (svq->batch_last.id == last_used) { |
| 535 | svq->batch_last.id = VIRTIO_RING_NOT_IN_BATCH; |
| 536 | *len = svq->batch_last.len; |
| 537 | } else { |
| 538 | *len = svq->desc_state[last_used].in_bytes; |
| 539 | } |
| 540 | |
| 541 | svq->last_used += svq->desc_state[last_used].ndescs; |
| 542 | svq->last_used_idx++; |
| 543 | return last_used; |
| 544 | } |
| 545 | |
| 546 | static uint16_t vhost_svq_last_desc_of_chain(const VhostShadowVirtqueue *svq, |
| 547 | uint16_t num, uint16_t i) |
| 548 | { |
| 549 | for (uint16_t j = 0; j < (num - 1); ++j) { |
| 550 | i = vhost_svq_next_desc(svq, i); |
| 551 | } |
| 552 | |
| 553 | return i; |
| 554 | } |
| 555 | |
| 556 | G_GNUC_WARN_UNUSED_RESULT |
| 557 | static VirtQueueElement *vhost_svq_detach_buf_split(VhostShadowVirtqueue *svq, |
| 558 | uint16_t id) |
| 559 | { |
| 560 | uint16_t num = svq->desc_state[id].ndescs; |
| 561 | uint16_t last_used_chain = vhost_svq_last_desc_of_chain(svq, num, id); |
| 562 | |
| 563 | svq->desc_state[last_used_chain].next = svq->free_head; |
| 564 | svq->free_head = id; |
| 565 | |
| 566 | return g_steal_pointer(&svq->desc_state[id].elem); |
| 567 | } |
| 568 | |
| 569 | G_GNUC_WARN_UNUSED_RESULT |
| 570 | static VirtQueueElement *vhost_svq_detach_buf_split_in_order( |
| 571 | VhostShadowVirtqueue *svq, |
| 572 | uint16_t id) |
| 573 | { |
| 574 | return g_steal_pointer(&svq->desc_state[id].elem); |
| 575 | } |
| 576 | |
| 577 | /* |
| 578 | * Return the descriptor id (and the chain of ids) to the free list |
| 579 | * |
| 580 | * @svq: Shadow Virtqueue |
| 581 | * @id: Id of the buffer to return. |
| 582 | * |
| 583 | * Return the element associated to the buffer if any. |
| 584 | */ |
| 585 | G_GNUC_WARN_UNUSED_RESULT |
| 586 | static VirtQueueElement *vhost_svq_detach_buf(VhostShadowVirtqueue *svq, |
| 587 | uint16_t id) |
| 588 | { |
| 589 | if (virtio_vdev_has_feature(svq->vdev, VIRTIO_F_IN_ORDER)) { |
| 590 | return vhost_svq_detach_buf_split_in_order(svq, id); |
| 591 | } else { |
| 592 | return vhost_svq_detach_buf_split(svq, id); |
| 593 | } |
| 594 | } |
| 595 | |
| 596 | G_GNUC_WARN_UNUSED_RESULT |
| 597 | static VirtQueueElement *vhost_svq_get_buf(VhostShadowVirtqueue *svq, |
| 598 | uint32_t *len) |
| 599 | { |
| 600 | uint16_t last_used; |
| 601 | |
| 602 | if (!vhost_svq_more_used(svq)) { |
| 603 | return NULL; |
| 604 | } |
| 605 | |
| 606 | /* Only get used array entries after they have been exposed by dev */ |
| 607 | smp_rmb(); |
| 608 | |
| 609 | if (virtio_vdev_has_feature(svq->vdev, VIRTIO_F_IN_ORDER)) { |
| 610 | int32_t r; |
| 611 | r = vhost_svq_get_last_used_split_in_order(svq, len); |
| 612 | if (r < 0) { |
| 613 | return NULL; |
| 614 | } |
| 615 | |
| 616 | last_used = r; |
| 617 | } else { |
| 618 | last_used = vhost_svq_get_last_used_split(svq, len); |
| 619 | } |
| 620 | |
| 621 | if (unlikely(last_used >= svq->vring.num)) { |
| 622 | qemu_log_mask(LOG_GUEST_ERROR, "Device %s says index %u is used", |
| 623 | svq->vdev->name, last_used); |
| 624 | return NULL; |
| 625 | } |
| 626 | |
| 627 | if (unlikely(!svq->desc_state[last_used].ndescs)) { |
| 628 | qemu_log_mask(LOG_GUEST_ERROR, |
| 629 | "Device %s says index %u is used, but it was not available", |
| 630 | svq->vdev->name, last_used); |
| 631 | return NULL; |
| 632 | } |
| 633 | |
| 634 | svq->num_free += svq->desc_state[last_used].ndescs; |
| 635 | svq->desc_state[last_used].ndescs = 0; |
| 636 | return vhost_svq_detach_buf(svq, last_used); |
| 637 | } |
| 638 | |
| 639 | /** |
| 640 | * Push an element to SVQ, returning it to the guest. |
| 641 | */ |
| 642 | void vhost_svq_push_elem(VhostShadowVirtqueue *svq, |
| 643 | const VirtQueueElement *elem, uint32_t len) |
| 644 | { |
| 645 | virtqueue_push(svq->vq, elem, len); |
| 646 | if (svq->next_guest_avail_elem) { |
| 647 | /* |
| 648 | * Avail ring was full when vhost_svq_flush was called, so it's a |
| 649 | * good moment to make more descriptors available if possible. |
| 650 | */ |
| 651 | vhost_handle_guest_kick(svq); |
| 652 | } |
| 653 | } |
| 654 | |
| 655 | static void vhost_svq_flush(VhostShadowVirtqueue *svq, |
| 656 | bool check_for_avail_queue) |
| 657 | { |
| 658 | VirtQueue *vq = svq->vq; |
| 659 | |
| 660 | /* Forward as many used buffers as possible. */ |
| 661 | do { |
| 662 | unsigned i = 0; |
| 663 | |
| 664 | vhost_svq_disable_notification(svq); |
| 665 | while (true) { |
| 666 | uint32_t len; |
| 667 | g_autofree VirtQueueElement *elem = vhost_svq_get_buf(svq, &len); |
| 668 | if (!elem) { |
| 669 | break; |
| 670 | } |
| 671 | |
| 672 | if (unlikely(i >= svq->vring.num)) { |
| 673 | qemu_log_mask(LOG_GUEST_ERROR, |
| 674 | "More than %u used buffers obtained in a %u size SVQ", |
| 675 | i, svq->vring.num); |
| 676 | virtqueue_fill(vq, elem, len, i); |
| 677 | virtqueue_flush(vq, i); |
| 678 | return; |
| 679 | } |
| 680 | virtqueue_fill(vq, elem, len, i++); |
| 681 | } |
| 682 | |
| 683 | virtqueue_flush(vq, i); |
| 684 | event_notifier_set(&svq->svq_call); |
| 685 | |
| 686 | if (check_for_avail_queue && svq->next_guest_avail_elem) { |
| 687 | /* |
| 688 | * Avail ring was full when vhost_svq_flush was called, so it's a |
| 689 | * good moment to make more descriptors available if possible. |
| 690 | */ |
| 691 | vhost_handle_guest_kick(svq); |
| 692 | } |
| 693 | } while (!vhost_svq_enable_notification(svq)); |
| 694 | } |
| 695 | |
| 696 | /** |
| 697 | * Poll the SVQ to wait for the device to use the specified number |
| 698 | * of elements and return the total length written by the device. |
| 699 | * |
| 700 | * This function race with main event loop SVQ polling, so extra |
| 701 | * synchronization is needed. |
| 702 | * |
| 703 | * @svq: The svq |
| 704 | * @num: The number of elements that need to be used |
| 705 | */ |
| 706 | size_t vhost_svq_poll(VhostShadowVirtqueue *svq, size_t num) |
| 707 | { |
| 708 | size_t len = 0; |
| 709 | |
| 710 | while (num--) { |
| 711 | g_autofree VirtQueueElement *elem = NULL; |
| 712 | int64_t start_us = g_get_monotonic_time(); |
| 713 | uint32_t r = 0; |
| 714 | |
| 715 | do { |
| 716 | if (vhost_svq_more_used(svq)) { |
| 717 | break; |
| 718 | } |
| 719 | |
| 720 | if (unlikely(g_get_monotonic_time() - start_us > 10e6)) { |
| 721 | return len; |
| 722 | } |
| 723 | } while (true); |
| 724 | |
| 725 | elem = vhost_svq_get_buf(svq, &r); |
| 726 | len += r; |
| 727 | } |
| 728 | |
| 729 | return len; |
| 730 | } |
| 731 | |
| 732 | /** |
| 733 | * Forward used buffers. |
| 734 | * |
| 735 | * @n: hdev call event notifier, the one that device set to notify svq. |
| 736 | * |
| 737 | * Note that we are not making any buffers available in the loop, there is no |
| 738 | * way that it runs more than virtqueue size times. |
| 739 | */ |
| 740 | static void vhost_svq_handle_call(EventNotifier *n) |
| 741 | { |
| 742 | VhostShadowVirtqueue *svq = container_of(n, VhostShadowVirtqueue, |
| 743 | hdev_call); |
| 744 | event_notifier_test_and_clear(n); |
| 745 | vhost_svq_flush(svq, true); |
| 746 | } |
| 747 | |
| 748 | /** |
| 749 | * Set the call notifier for the SVQ to call the guest |
| 750 | * |
| 751 | * @svq: Shadow virtqueue |
| 752 | * @call_fd: call notifier |
| 753 | * |
| 754 | * Called on BQL context. |
| 755 | */ |
| 756 | void vhost_svq_set_svq_call_fd(VhostShadowVirtqueue *svq, int call_fd) |
| 757 | { |
| 758 | if (call_fd == VHOST_FILE_UNBIND) { |
| 759 | /* |
| 760 | * Fail event_notifier_set if called handling device call. |
| 761 | * |
| 762 | * SVQ still needs device notifications, since it needs to keep |
| 763 | * forwarding used buffers even with the unbind. |
| 764 | */ |
| 765 | memset(&svq->svq_call, 0, sizeof(svq->svq_call)); |
| 766 | } else { |
| 767 | event_notifier_init_fd(&svq->svq_call, call_fd); |
| 768 | } |
| 769 | } |
| 770 | |
| 771 | /** |
| 772 | * Get the shadow vq vring address. |
| 773 | * @svq: Shadow virtqueue |
| 774 | * @addr: Destination to store address |
| 775 | */ |
| 776 | void vhost_svq_get_vring_addr(const VhostShadowVirtqueue *svq, |
| 777 | struct vhost_vring_addr *addr) |
| 778 | { |
| 779 | addr->desc_user_addr = (uint64_t)(uintptr_t)svq->vring.desc; |
| 780 | addr->avail_user_addr = (uint64_t)(uintptr_t)svq->vring.avail; |
| 781 | addr->used_user_addr = (uint64_t)(uintptr_t)svq->vring.used; |
| 782 | } |
| 783 | |
| 784 | size_t vhost_svq_driver_area_size(const VhostShadowVirtqueue *svq) |
| 785 | { |
| 786 | size_t desc_size = sizeof(vring_desc_t) * svq->vring.num; |
| 787 | size_t avail_size = offsetof(vring_avail_t, ring[svq->vring.num]) + |
| 788 | sizeof(uint16_t); |
| 789 | |
| 790 | return ROUND_UP(desc_size + avail_size, qemu_real_host_page_size()); |
| 791 | } |
| 792 | |
| 793 | size_t vhost_svq_device_area_size(const VhostShadowVirtqueue *svq) |
| 794 | { |
| 795 | size_t used_size = offsetof(vring_used_t, ring[svq->vring.num]) + |
| 796 | sizeof(uint16_t); |
| 797 | return ROUND_UP(used_size, qemu_real_host_page_size()); |
| 798 | } |
| 799 | |
| 800 | /** |
| 801 | * Set a new file descriptor for the guest to kick the SVQ and notify for avail |
| 802 | * |
| 803 | * @svq: The svq |
| 804 | * @svq_kick_fd: The svq kick fd |
| 805 | * |
| 806 | * Note that the SVQ will never close the old file descriptor. |
| 807 | */ |
| 808 | void vhost_svq_set_svq_kick_fd(VhostShadowVirtqueue *svq, int svq_kick_fd) |
| 809 | { |
| 810 | EventNotifier *svq_kick = &svq->svq_kick; |
| 811 | bool poll_stop = VHOST_FILE_UNBIND != event_notifier_get_fd(svq_kick); |
| 812 | bool poll_start = svq_kick_fd != VHOST_FILE_UNBIND; |
| 813 | |
| 814 | if (poll_stop) { |
| 815 | event_notifier_set_handler(svq_kick, NULL); |
| 816 | } |
| 817 | |
| 818 | event_notifier_init_fd(svq_kick, svq_kick_fd); |
| 819 | /* |
| 820 | * event_notifier_set_handler already checks for guest's notifications if |
| 821 | * they arrive at the new file descriptor in the switch, so there is no |
| 822 | * need to explicitly check for them. |
| 823 | */ |
| 824 | if (poll_start) { |
| 825 | event_notifier_set(svq_kick); |
| 826 | event_notifier_set_handler(svq_kick, vhost_handle_guest_kick_notifier); |
| 827 | } |
| 828 | } |
| 829 | |
| 830 | /** |
| 831 | * Start the shadow virtqueue operation. |
| 832 | * |
| 833 | * @svq: Shadow Virtqueue |
| 834 | * @vdev: VirtIO device |
| 835 | * @vq: Virtqueue to shadow |
| 836 | * @iova_tree: Tree to perform descriptors translations |
| 837 | */ |
| 838 | void vhost_svq_start(VhostShadowVirtqueue *svq, VirtIODevice *vdev, |
| 839 | VirtQueue *vq, VhostIOVATree *iova_tree) |
| 840 | { |
| 841 | size_t desc_size; |
| 842 | |
| 843 | event_notifier_set_handler(&svq->hdev_call, vhost_svq_handle_call); |
| 844 | svq->next_guest_avail_elem = NULL; |
| 845 | svq->shadow_avail_idx = 0; |
| 846 | svq->shadow_used_idx = 0; |
| 847 | memset(&svq->batch_last, 0, sizeof(svq->batch_last)); |
| 848 | svq->last_used = 0; |
| 849 | svq->last_used_idx = 0; |
| 850 | svq->vdev = vdev; |
| 851 | svq->vq = vq; |
| 852 | svq->iova_tree = iova_tree; |
| 853 | |
| 854 | svq->vring.num = virtio_queue_get_num(vdev, virtio_get_queue_index(vq)); |
| 855 | svq->num_free = svq->vring.num; |
| 856 | svq->vring.desc = mmap(NULL, vhost_svq_driver_area_size(svq), |
| 857 | PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANONYMOUS, |
| 858 | -1, 0); |
| 859 | desc_size = sizeof(vring_desc_t) * svq->vring.num; |
| 860 | svq->vring.avail = (void *)((char *)svq->vring.desc + desc_size); |
| 861 | svq->vring.used = mmap(NULL, vhost_svq_device_area_size(svq), |
| 862 | PROT_READ | PROT_WRITE, MAP_SHARED | MAP_ANONYMOUS, |
| 863 | -1, 0); |
| 864 | svq->desc_state = g_new0(SVQDescState, svq->vring.num); |
| 865 | if (virtio_vdev_has_feature(svq->vdev, VIRTIO_F_IN_ORDER)) { |
| 866 | svq->batch_last.id = VIRTIO_RING_NOT_IN_BATCH; |
| 867 | } else { |
| 868 | for (unsigned i = 0; i < svq->vring.num - 1; i++) { |
| 869 | svq->desc_state[i].next = i + 1; |
| 870 | } |
| 871 | } |
| 872 | } |
| 873 | |
| 874 | /** |
| 875 | * Stop the shadow virtqueue operation. |
| 876 | * @svq: Shadow Virtqueue |
| 877 | */ |
| 878 | void vhost_svq_stop(VhostShadowVirtqueue *svq) |
| 879 | { |
| 880 | vhost_svq_set_svq_kick_fd(svq, VHOST_FILE_UNBIND); |
| 881 | g_autofree VirtQueueElement *next_avail_elem = NULL; |
| 882 | |
| 883 | if (!svq->vq) { |
| 884 | return; |
| 885 | } |
| 886 | |
| 887 | /* Send all pending used descriptors to guest */ |
| 888 | vhost_svq_flush(svq, false); |
| 889 | |
| 890 | for (unsigned i = 0; i < svq->vring.num; ++i) { |
| 891 | g_autofree VirtQueueElement *elem = NULL; |
| 892 | elem = g_steal_pointer(&svq->desc_state[i].elem); |
| 893 | if (elem) { |
| 894 | /* |
| 895 | * TODO: This is ok for networking, but other kinds of devices |
| 896 | * might have problems with just unpop these. |
| 897 | */ |
| 898 | virtqueue_unpop(svq->vq, elem, 0); |
| 899 | } |
| 900 | } |
| 901 | |
| 902 | next_avail_elem = g_steal_pointer(&svq->next_guest_avail_elem); |
| 903 | if (next_avail_elem) { |
| 904 | virtqueue_unpop(svq->vq, next_avail_elem, 0); |
| 905 | } |
| 906 | svq->vq = NULL; |
| 907 | g_free(svq->desc_state); |
| 908 | munmap(svq->vring.desc, vhost_svq_driver_area_size(svq)); |
| 909 | munmap(svq->vring.used, vhost_svq_device_area_size(svq)); |
| 910 | event_notifier_set_handler(&svq->hdev_call, NULL); |
| 911 | } |
| 912 | |
| 913 | /** |
| 914 | * Creates vhost shadow virtqueue, and instructs the vhost device to use the |
| 915 | * shadow methods and file descriptors. |
| 916 | * |
| 917 | * @ops: SVQ owner callbacks |
| 918 | * @ops_opaque: ops opaque pointer |
| 919 | */ |
| 920 | VhostShadowVirtqueue *vhost_svq_new(const VhostShadowVirtqueueOps *ops, |
| 921 | void *ops_opaque) |
| 922 | { |
| 923 | VhostShadowVirtqueue *svq = g_new0(VhostShadowVirtqueue, 1); |
| 924 | |
| 925 | event_notifier_init_fd(&svq->svq_kick, VHOST_FILE_UNBIND); |
| 926 | svq->ops = ops; |
| 927 | svq->ops_opaque = ops_opaque; |
| 928 | return svq; |
| 929 | } |
| 930 | |
| 931 | /** |
| 932 | * Free the resources of the shadow virtqueue. |
| 933 | * |
| 934 | * @pvq: gpointer to SVQ so it can be used by autofree functions. |
| 935 | */ |
| 936 | void vhost_svq_free(gpointer pvq) |
| 937 | { |
| 938 | VhostShadowVirtqueue *vq = pvq; |
| 939 | vhost_svq_stop(vq); |
| 940 | g_free(vq); |
| 941 | } |