| 1 | /* |
| 2 | * VDUSE (vDPA Device in Userspace) library |
| 3 | * |
| 4 | * Copyright (C) 2022 Bytedance Inc. and/or its affiliates. All rights reserved. |
| 5 | * Portions of codes and concepts borrowed from libvhost-user.c, so: |
| 6 | * Copyright IBM, Corp. 2007 |
| 7 | * Copyright (c) 2016 Red Hat, Inc. |
| 8 | * |
| 9 | * Author: |
| 10 | * Xie Yongji <xieyongji@bytedance.com> |
| 11 | * Anthony Liguori <aliguori@us.ibm.com> |
| 12 | * Marc-André Lureau <mlureau@redhat.com> |
| 13 | * Victor Kaplansky <victork@redhat.com> |
| 14 | * |
| 15 | * This work is licensed under the terms of the GNU GPL, version 2 or |
| 16 | * later. See the COPYING file in the top-level directory. |
| 17 | */ |
| 18 | |
| 19 | #ifndef _GNU_SOURCE |
| 20 | #define _GNU_SOURCE |
| 21 | #endif |
| 22 | |
| 23 | #include <stdlib.h> |
| 24 | #include <stdio.h> |
| 25 | #include <stdbool.h> |
| 26 | #include <stddef.h> |
| 27 | #include <errno.h> |
| 28 | #include <string.h> |
| 29 | #include <assert.h> |
| 30 | #include <endian.h> |
| 31 | #include <unistd.h> |
| 32 | #include <limits.h> |
| 33 | #include <fcntl.h> |
| 34 | #include <inttypes.h> |
| 35 | |
| 36 | #include <sys/ioctl.h> |
| 37 | #include <sys/eventfd.h> |
| 38 | #include <sys/mman.h> |
| 39 | |
| 40 | #include "include/atomic.h" |
| 41 | #include "linux-headers/linux/virtio_ring.h" |
| 42 | #include "linux-headers/linux/virtio_config.h" |
| 43 | #include "linux-headers/linux/vduse.h" |
| 44 | #include "libvduse.h" |
| 45 | |
| 46 | #define VDUSE_VQ_ALIGN 4096 |
| 47 | #define MAX_IOVA_REGIONS 256 |
| 48 | |
| 49 | #define LOG_ALIGNMENT 64 |
| 50 | |
| 51 | /* Round number down to multiple */ |
| 52 | #define ALIGN_DOWN(n, m) ((n) / (m) * (m)) |
| 53 | |
| 54 | /* Round number up to multiple */ |
| 55 | #define ALIGN_UP(n, m) ALIGN_DOWN((n) + (m) - 1, (m)) |
| 56 | |
| 57 | #ifndef unlikely |
| 58 | #define unlikely(x) __builtin_expect(!!(x), 0) |
| 59 | #endif |
| 60 | |
| 61 | typedef struct VduseDescStateSplit { |
| 62 | uint8_t inflight; |
| 63 | uint8_t padding[5]; |
| 64 | uint16_t next; |
| 65 | uint64_t counter; |
| 66 | } VduseDescStateSplit; |
| 67 | |
| 68 | typedef struct VduseVirtqLogInflight { |
| 69 | uint64_t features; |
| 70 | uint16_t version; |
| 71 | uint16_t desc_num; |
| 72 | uint16_t last_batch_head; |
| 73 | uint16_t used_idx; |
| 74 | VduseDescStateSplit desc[]; |
| 75 | } VduseVirtqLogInflight; |
| 76 | |
| 77 | typedef struct VduseVirtqLog { |
| 78 | VduseVirtqLogInflight inflight; |
| 79 | } VduseVirtqLog; |
| 80 | |
| 81 | typedef struct VduseVirtqInflightDesc { |
| 82 | uint16_t index; |
| 83 | uint64_t counter; |
| 84 | } VduseVirtqInflightDesc; |
| 85 | |
| 86 | typedef struct VduseRing { |
| 87 | unsigned int num; |
| 88 | uint64_t desc_addr; |
| 89 | uint64_t avail_addr; |
| 90 | uint64_t used_addr; |
| 91 | struct vring_desc *desc; |
| 92 | struct vring_avail *avail; |
| 93 | struct vring_used *used; |
| 94 | } VduseRing; |
| 95 | |
| 96 | struct VduseVirtq { |
| 97 | VduseRing vring; |
| 98 | uint16_t last_avail_idx; |
| 99 | uint16_t shadow_avail_idx; |
| 100 | uint16_t used_idx; |
| 101 | uint16_t signalled_used; |
| 102 | bool signalled_used_valid; |
| 103 | int index; |
| 104 | unsigned int inuse; |
| 105 | bool ready; |
| 106 | int fd; |
| 107 | VduseDev *dev; |
| 108 | VduseVirtqInflightDesc *resubmit_list; |
| 109 | uint16_t resubmit_num; |
| 110 | uint64_t counter; |
| 111 | VduseVirtqLog *log; |
| 112 | }; |
| 113 | |
| 114 | typedef struct VduseIovaRegion { |
| 115 | uint64_t iova; |
| 116 | uint64_t size; |
| 117 | uint64_t mmap_offset; |
| 118 | uint64_t mmap_addr; |
| 119 | } VduseIovaRegion; |
| 120 | |
| 121 | struct VduseDev { |
| 122 | VduseVirtq *vqs; |
| 123 | VduseIovaRegion regions[MAX_IOVA_REGIONS]; |
| 124 | int num_regions; |
| 125 | char *name; |
| 126 | uint32_t device_id; |
| 127 | uint32_t vendor_id; |
| 128 | uint16_t num_queues; |
| 129 | uint16_t queue_size; |
| 130 | uint64_t features; |
| 131 | const VduseOps *ops; |
| 132 | int fd; |
| 133 | int ctrl_fd; |
| 134 | void *priv; |
| 135 | void *log; |
| 136 | }; |
| 137 | |
| 138 | static inline size_t vduse_vq_log_size(uint16_t queue_size) |
| 139 | { |
| 140 | return ALIGN_UP(sizeof(VduseDescStateSplit) * queue_size + |
| 141 | sizeof(VduseVirtqLogInflight), LOG_ALIGNMENT); |
| 142 | } |
| 143 | |
| 144 | static void *vduse_log_get(const char *filename, size_t size) |
| 145 | { |
| 146 | void *ptr = MAP_FAILED; |
| 147 | int fd; |
| 148 | |
| 149 | fd = open(filename, O_RDWR | O_CREAT, 0600); |
| 150 | if (fd == -1) { |
| 151 | return MAP_FAILED; |
| 152 | } |
| 153 | |
| 154 | if (ftruncate(fd, size) == -1) { |
| 155 | goto out; |
| 156 | } |
| 157 | |
| 158 | ptr = mmap(0, size, PROT_READ | PROT_WRITE, MAP_SHARED, fd, 0); |
| 159 | |
| 160 | out: |
| 161 | close(fd); |
| 162 | return ptr; |
| 163 | } |
| 164 | |
| 165 | static inline bool has_feature(uint64_t features, unsigned int fbit) |
| 166 | { |
| 167 | assert(fbit < 64); |
| 168 | return !!(features & (1ULL << fbit)); |
| 169 | } |
| 170 | |
| 171 | static inline bool vduse_dev_has_feature(VduseDev *dev, unsigned int fbit) |
| 172 | { |
| 173 | return has_feature(dev->features, fbit); |
| 174 | } |
| 175 | |
| 176 | uint64_t vduse_get_virtio_features(void) |
| 177 | { |
| 178 | return (1ULL << VIRTIO_F_IOMMU_PLATFORM) | |
| 179 | (1ULL << VIRTIO_F_VERSION_1) | |
| 180 | (1ULL << VIRTIO_F_NOTIFY_ON_EMPTY) | |
| 181 | (1ULL << VIRTIO_RING_F_EVENT_IDX) | |
| 182 | (1ULL << VIRTIO_RING_F_INDIRECT_DESC); |
| 183 | } |
| 184 | |
| 185 | VduseDev *vduse_queue_get_dev(VduseVirtq *vq) |
| 186 | { |
| 187 | return vq->dev; |
| 188 | } |
| 189 | |
| 190 | int vduse_queue_get_fd(VduseVirtq *vq) |
| 191 | { |
| 192 | return vq->fd; |
| 193 | } |
| 194 | |
| 195 | void *vduse_dev_get_priv(VduseDev *dev) |
| 196 | { |
| 197 | return dev->priv; |
| 198 | } |
| 199 | |
| 200 | VduseVirtq *vduse_dev_get_queue(VduseDev *dev, int index) |
| 201 | { |
| 202 | return &dev->vqs[index]; |
| 203 | } |
| 204 | |
| 205 | int vduse_dev_get_fd(VduseDev *dev) |
| 206 | { |
| 207 | return dev->fd; |
| 208 | } |
| 209 | |
| 210 | static int vduse_inject_irq(VduseDev *dev, int index) |
| 211 | { |
| 212 | return ioctl(dev->fd, VDUSE_VQ_INJECT_IRQ, &index); |
| 213 | } |
| 214 | |
| 215 | static int inflight_desc_compare(const void *a, const void *b) |
| 216 | { |
| 217 | VduseVirtqInflightDesc *desc0 = (VduseVirtqInflightDesc *)a, |
| 218 | *desc1 = (VduseVirtqInflightDesc *)b; |
| 219 | |
| 220 | if (desc1->counter > desc0->counter && |
| 221 | (desc1->counter - desc0->counter) < VIRTQUEUE_MAX_SIZE * 2) { |
| 222 | return 1; |
| 223 | } |
| 224 | |
| 225 | return -1; |
| 226 | } |
| 227 | |
| 228 | static int vduse_queue_check_inflights(VduseVirtq *vq) |
| 229 | { |
| 230 | int i = 0; |
| 231 | VduseDev *dev = vq->dev; |
| 232 | |
| 233 | vq->used_idx = le16toh(vq->vring.used->idx); |
| 234 | vq->resubmit_num = 0; |
| 235 | vq->resubmit_list = NULL; |
| 236 | vq->counter = 0; |
| 237 | |
| 238 | if (unlikely(vq->log->inflight.used_idx != vq->used_idx)) { |
| 239 | if (vq->log->inflight.last_batch_head > VIRTQUEUE_MAX_SIZE) { |
| 240 | return -1; |
| 241 | } |
| 242 | |
| 243 | vq->log->inflight.desc[vq->log->inflight.last_batch_head].inflight = 0; |
| 244 | |
| 245 | barrier(); |
| 246 | |
| 247 | vq->log->inflight.used_idx = vq->used_idx; |
| 248 | } |
| 249 | |
| 250 | for (i = 0; i < vq->log->inflight.desc_num; i++) { |
| 251 | if (vq->log->inflight.desc[i].inflight == 1) { |
| 252 | vq->inuse++; |
| 253 | } |
| 254 | } |
| 255 | |
| 256 | vq->shadow_avail_idx = vq->last_avail_idx = vq->inuse + vq->used_idx; |
| 257 | |
| 258 | if (vq->inuse) { |
| 259 | vq->resubmit_list = calloc(vq->inuse, sizeof(VduseVirtqInflightDesc)); |
| 260 | if (!vq->resubmit_list) { |
| 261 | return -1; |
| 262 | } |
| 263 | |
| 264 | for (i = 0; i < vq->log->inflight.desc_num; i++) { |
| 265 | if (vq->log->inflight.desc[i].inflight) { |
| 266 | vq->resubmit_list[vq->resubmit_num].index = i; |
| 267 | vq->resubmit_list[vq->resubmit_num].counter = |
| 268 | vq->log->inflight.desc[i].counter; |
| 269 | vq->resubmit_num++; |
| 270 | } |
| 271 | } |
| 272 | |
| 273 | if (vq->resubmit_num > 1) { |
| 274 | qsort(vq->resubmit_list, vq->resubmit_num, |
| 275 | sizeof(VduseVirtqInflightDesc), inflight_desc_compare); |
| 276 | } |
| 277 | vq->counter = vq->resubmit_list[0].counter + 1; |
| 278 | } |
| 279 | |
| 280 | vduse_inject_irq(dev, vq->index); |
| 281 | |
| 282 | return 0; |
| 283 | } |
| 284 | |
| 285 | static int vduse_queue_inflight_get(VduseVirtq *vq, int desc_idx) |
| 286 | { |
| 287 | vq->log->inflight.desc[desc_idx].counter = vq->counter++; |
| 288 | |
| 289 | barrier(); |
| 290 | |
| 291 | vq->log->inflight.desc[desc_idx].inflight = 1; |
| 292 | |
| 293 | return 0; |
| 294 | } |
| 295 | |
| 296 | static int vduse_queue_inflight_pre_put(VduseVirtq *vq, int desc_idx) |
| 297 | { |
| 298 | vq->log->inflight.last_batch_head = desc_idx; |
| 299 | |
| 300 | return 0; |
| 301 | } |
| 302 | |
| 303 | static int vduse_queue_inflight_post_put(VduseVirtq *vq, int desc_idx) |
| 304 | { |
| 305 | vq->log->inflight.desc[desc_idx].inflight = 0; |
| 306 | |
| 307 | barrier(); |
| 308 | |
| 309 | vq->log->inflight.used_idx = vq->used_idx; |
| 310 | |
| 311 | return 0; |
| 312 | } |
| 313 | |
| 314 | static void vduse_iova_remove_region(VduseDev *dev, uint64_t start, |
| 315 | uint64_t last) |
| 316 | { |
| 317 | int i; |
| 318 | |
| 319 | if (last == start) { |
| 320 | return; |
| 321 | } |
| 322 | |
| 323 | for (i = 0; i < MAX_IOVA_REGIONS; i++) { |
| 324 | if (!dev->regions[i].mmap_addr) { |
| 325 | continue; |
| 326 | } |
| 327 | |
| 328 | if (start <= dev->regions[i].iova && |
| 329 | last >= (dev->regions[i].iova + dev->regions[i].size - 1)) { |
| 330 | munmap((void *)(uintptr_t)dev->regions[i].mmap_addr, |
| 331 | dev->regions[i].mmap_offset + dev->regions[i].size); |
| 332 | dev->regions[i].mmap_addr = 0; |
| 333 | dev->num_regions--; |
| 334 | } |
| 335 | } |
| 336 | } |
| 337 | |
| 338 | static int vduse_iova_add_region(VduseDev *dev, int fd, |
| 339 | uint64_t offset, uint64_t start, |
| 340 | uint64_t last, int prot) |
| 341 | { |
| 342 | int i; |
| 343 | uint64_t size = last - start + 1; |
| 344 | void *mmap_addr = mmap(0, size + offset, prot, MAP_SHARED, fd, 0); |
| 345 | |
| 346 | if (mmap_addr == MAP_FAILED) { |
| 347 | close(fd); |
| 348 | return -EINVAL; |
| 349 | } |
| 350 | |
| 351 | for (i = 0; i < MAX_IOVA_REGIONS; i++) { |
| 352 | if (!dev->regions[i].mmap_addr) { |
| 353 | dev->regions[i].mmap_addr = (uint64_t)(uintptr_t)mmap_addr; |
| 354 | dev->regions[i].mmap_offset = offset; |
| 355 | dev->regions[i].iova = start; |
| 356 | dev->regions[i].size = size; |
| 357 | dev->num_regions++; |
| 358 | break; |
| 359 | } |
| 360 | } |
| 361 | assert(i < MAX_IOVA_REGIONS); |
| 362 | close(fd); |
| 363 | |
| 364 | return 0; |
| 365 | } |
| 366 | |
| 367 | static int perm_to_prot(uint8_t perm) |
| 368 | { |
| 369 | int prot = 0; |
| 370 | |
| 371 | switch (perm) { |
| 372 | case VDUSE_ACCESS_WO: |
| 373 | prot |= PROT_WRITE; |
| 374 | break; |
| 375 | case VDUSE_ACCESS_RO: |
| 376 | prot |= PROT_READ; |
| 377 | break; |
| 378 | case VDUSE_ACCESS_RW: |
| 379 | prot |= PROT_READ | PROT_WRITE; |
| 380 | break; |
| 381 | default: |
| 382 | break; |
| 383 | } |
| 384 | |
| 385 | return prot; |
| 386 | } |
| 387 | |
| 388 | static inline void *iova_to_va(VduseDev *dev, uint64_t *plen, uint64_t iova) |
| 389 | { |
| 390 | int i, ret; |
| 391 | struct vduse_iotlb_entry entry; |
| 392 | |
| 393 | for (i = 0; i < MAX_IOVA_REGIONS; i++) { |
| 394 | VduseIovaRegion *r = &dev->regions[i]; |
| 395 | |
| 396 | if (!r->mmap_addr) { |
| 397 | continue; |
| 398 | } |
| 399 | |
| 400 | if ((iova >= r->iova) && (iova < (r->iova + r->size))) { |
| 401 | if ((iova + *plen) > (r->iova + r->size)) { |
| 402 | *plen = r->iova + r->size - iova; |
| 403 | } |
| 404 | return (void *)(uintptr_t)(iova - r->iova + |
| 405 | r->mmap_addr + r->mmap_offset); |
| 406 | } |
| 407 | } |
| 408 | |
| 409 | entry.start = iova; |
| 410 | entry.last = iova + 1; |
| 411 | ret = ioctl(dev->fd, VDUSE_IOTLB_GET_FD, &entry); |
| 412 | if (ret < 0) { |
| 413 | return NULL; |
| 414 | } |
| 415 | |
| 416 | if (!vduse_iova_add_region(dev, ret, entry.offset, entry.start, |
| 417 | entry.last, perm_to_prot(entry.perm))) { |
| 418 | return iova_to_va(dev, plen, iova); |
| 419 | } |
| 420 | |
| 421 | return NULL; |
| 422 | } |
| 423 | |
| 424 | static inline uint16_t vring_avail_flags(VduseVirtq *vq) |
| 425 | { |
| 426 | return le16toh(vq->vring.avail->flags); |
| 427 | } |
| 428 | |
| 429 | static inline uint16_t vring_avail_idx(VduseVirtq *vq) |
| 430 | { |
| 431 | vq->shadow_avail_idx = le16toh(vq->vring.avail->idx); |
| 432 | |
| 433 | return vq->shadow_avail_idx; |
| 434 | } |
| 435 | |
| 436 | static inline uint16_t vring_avail_ring(VduseVirtq *vq, int i) |
| 437 | { |
| 438 | return le16toh(vq->vring.avail->ring[i]); |
| 439 | } |
| 440 | |
| 441 | static inline uint16_t vring_get_used_event(VduseVirtq *vq) |
| 442 | { |
| 443 | return vring_avail_ring(vq, vq->vring.num); |
| 444 | } |
| 445 | |
| 446 | static bool vduse_queue_get_head(VduseVirtq *vq, unsigned int idx, |
| 447 | unsigned int *head) |
| 448 | { |
| 449 | /* |
| 450 | * Grab the next descriptor number they're advertising, and increment |
| 451 | * the index we've seen. |
| 452 | */ |
| 453 | *head = vring_avail_ring(vq, idx % vq->vring.num); |
| 454 | |
| 455 | /* If their number is silly, that's a fatal mistake. */ |
| 456 | if (*head >= vq->vring.num) { |
| 457 | fprintf(stderr, "Guest says index %u is available\n", *head); |
| 458 | return false; |
| 459 | } |
| 460 | |
| 461 | return true; |
| 462 | } |
| 463 | |
| 464 | static int |
| 465 | vduse_queue_read_indirect_desc(VduseDev *dev, struct vring_desc *desc, |
| 466 | uint64_t addr, size_t len) |
| 467 | { |
| 468 | char *dst_desc = (char *)desc; |
| 469 | uint64_t read_len; |
| 470 | void *ori_desc; |
| 471 | |
| 472 | if (len > (VIRTQUEUE_MAX_SIZE * sizeof(struct vring_desc))) { |
| 473 | return -1; |
| 474 | } |
| 475 | |
| 476 | if (len == 0) { |
| 477 | return -1; |
| 478 | } |
| 479 | |
| 480 | while (len) { |
| 481 | read_len = len; |
| 482 | ori_desc = iova_to_va(dev, &read_len, addr); |
| 483 | if (!ori_desc) { |
| 484 | return -1; |
| 485 | } |
| 486 | |
| 487 | memcpy(dst_desc, ori_desc, read_len); |
| 488 | len -= read_len; |
| 489 | addr += read_len; |
| 490 | dst_desc += read_len; |
| 491 | } |
| 492 | |
| 493 | return 0; |
| 494 | } |
| 495 | |
| 496 | enum { |
| 497 | VIRTQUEUE_READ_DESC_ERROR = -1, |
| 498 | VIRTQUEUE_READ_DESC_DONE = 0, /* end of chain */ |
| 499 | VIRTQUEUE_READ_DESC_MORE = 1, /* more buffers in chain */ |
| 500 | }; |
| 501 | |
| 502 | static int vduse_queue_read_next_desc(struct vring_desc *desc, int i, |
| 503 | unsigned int max, unsigned int *next) |
| 504 | { |
| 505 | /* If this descriptor says it doesn't chain, we're done. */ |
| 506 | if (!(le16toh(desc[i].flags) & VRING_DESC_F_NEXT)) { |
| 507 | return VIRTQUEUE_READ_DESC_DONE; |
| 508 | } |
| 509 | |
| 510 | /* Check they're not leading us off end of descriptors. */ |
| 511 | *next = desc[i].next; |
| 512 | /* Make sure compiler knows to grab that: we don't want it changing! */ |
| 513 | smp_wmb(); |
| 514 | |
| 515 | if (*next >= max) { |
| 516 | fprintf(stderr, "Desc next is %u\n", *next); |
| 517 | return VIRTQUEUE_READ_DESC_ERROR; |
| 518 | } |
| 519 | |
| 520 | return VIRTQUEUE_READ_DESC_MORE; |
| 521 | } |
| 522 | |
| 523 | /* |
| 524 | * Fetch avail_idx from VQ memory only when we really need to know if |
| 525 | * guest has added some buffers. |
| 526 | */ |
| 527 | static bool vduse_queue_empty(VduseVirtq *vq) |
| 528 | { |
| 529 | if (unlikely(!vq->vring.avail)) { |
| 530 | return true; |
| 531 | } |
| 532 | |
| 533 | if (vq->shadow_avail_idx != vq->last_avail_idx) { |
| 534 | return false; |
| 535 | } |
| 536 | |
| 537 | return vring_avail_idx(vq) == vq->last_avail_idx; |
| 538 | } |
| 539 | |
| 540 | static bool vduse_queue_should_notify(VduseVirtq *vq) |
| 541 | { |
| 542 | VduseDev *dev = vq->dev; |
| 543 | uint16_t old, new; |
| 544 | bool v; |
| 545 | |
| 546 | /* We need to expose used array entries before checking used event. */ |
| 547 | smp_mb(); |
| 548 | |
| 549 | /* Always notify when queue is empty (when feature acknowledge) */ |
| 550 | if (vduse_dev_has_feature(dev, VIRTIO_F_NOTIFY_ON_EMPTY) && |
| 551 | !vq->inuse && vduse_queue_empty(vq)) { |
| 552 | return true; |
| 553 | } |
| 554 | |
| 555 | if (!vduse_dev_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) { |
| 556 | return !(vring_avail_flags(vq) & VRING_AVAIL_F_NO_INTERRUPT); |
| 557 | } |
| 558 | |
| 559 | v = vq->signalled_used_valid; |
| 560 | vq->signalled_used_valid = true; |
| 561 | old = vq->signalled_used; |
| 562 | new = vq->signalled_used = vq->used_idx; |
| 563 | return !v || vring_need_event(vring_get_used_event(vq), new, old); |
| 564 | } |
| 565 | |
| 566 | void vduse_queue_notify(VduseVirtq *vq) |
| 567 | { |
| 568 | VduseDev *dev = vq->dev; |
| 569 | |
| 570 | if (unlikely(!vq->vring.avail)) { |
| 571 | return; |
| 572 | } |
| 573 | |
| 574 | if (!vduse_queue_should_notify(vq)) { |
| 575 | return; |
| 576 | } |
| 577 | |
| 578 | if (vduse_inject_irq(dev, vq->index) < 0) { |
| 579 | fprintf(stderr, "Error inject irq for vq %d: %s\n", |
| 580 | vq->index, strerror(errno)); |
| 581 | } |
| 582 | } |
| 583 | |
| 584 | static inline void vring_set_avail_event(VduseVirtq *vq, uint16_t val) |
| 585 | { |
| 586 | uint16_t val_le = htole16(val); |
| 587 | memcpy(&vq->vring.used->ring[vq->vring.num], &val_le, sizeof(uint16_t)); |
| 588 | } |
| 589 | |
| 590 | static bool vduse_queue_map_single_desc(VduseVirtq *vq, unsigned int *p_num_sg, |
| 591 | struct iovec *iov, unsigned int max_num_sg, |
| 592 | bool is_write, uint64_t pa, size_t sz) |
| 593 | { |
| 594 | unsigned num_sg = *p_num_sg; |
| 595 | VduseDev *dev = vq->dev; |
| 596 | |
| 597 | assert(num_sg <= max_num_sg); |
| 598 | |
| 599 | if (!sz) { |
| 600 | fprintf(stderr, "virtio: zero sized buffers are not allowed\n"); |
| 601 | return false; |
| 602 | } |
| 603 | |
| 604 | while (sz) { |
| 605 | uint64_t len = sz; |
| 606 | |
| 607 | if (num_sg == max_num_sg) { |
| 608 | fprintf(stderr, |
| 609 | "virtio: too many descriptors in indirect table\n"); |
| 610 | return false; |
| 611 | } |
| 612 | |
| 613 | iov[num_sg].iov_base = iova_to_va(dev, &len, pa); |
| 614 | if (iov[num_sg].iov_base == NULL) { |
| 615 | fprintf(stderr, "virtio: invalid address for buffers\n"); |
| 616 | return false; |
| 617 | } |
| 618 | iov[num_sg++].iov_len = len; |
| 619 | sz -= len; |
| 620 | pa += len; |
| 621 | } |
| 622 | |
| 623 | *p_num_sg = num_sg; |
| 624 | return true; |
| 625 | } |
| 626 | |
| 627 | static void *vduse_queue_alloc_element(size_t sz, unsigned out_num, |
| 628 | unsigned in_num) |
| 629 | { |
| 630 | VduseVirtqElement *elem; |
| 631 | size_t in_sg_ofs = ALIGN_UP(sz, __alignof__(elem->in_sg[0])); |
| 632 | size_t out_sg_ofs = in_sg_ofs + in_num * sizeof(elem->in_sg[0]); |
| 633 | size_t out_sg_end = out_sg_ofs + out_num * sizeof(elem->out_sg[0]); |
| 634 | |
| 635 | assert(sz >= sizeof(VduseVirtqElement)); |
| 636 | elem = malloc(out_sg_end); |
| 637 | if (!elem) { |
| 638 | return NULL; |
| 639 | } |
| 640 | elem->out_num = out_num; |
| 641 | elem->in_num = in_num; |
| 642 | elem->in_sg = (void *)elem + in_sg_ofs; |
| 643 | elem->out_sg = (void *)elem + out_sg_ofs; |
| 644 | return elem; |
| 645 | } |
| 646 | |
| 647 | static void *vduse_queue_map_desc(VduseVirtq *vq, unsigned int idx, size_t sz) |
| 648 | { |
| 649 | struct vring_desc *desc = vq->vring.desc; |
| 650 | VduseDev *dev = vq->dev; |
| 651 | uint64_t desc_addr, read_len; |
| 652 | unsigned int desc_len; |
| 653 | unsigned int max = vq->vring.num; |
| 654 | unsigned int i = idx; |
| 655 | VduseVirtqElement *elem; |
| 656 | struct iovec iov[VIRTQUEUE_MAX_SIZE]; |
| 657 | struct vring_desc desc_buf[VIRTQUEUE_MAX_SIZE]; |
| 658 | unsigned int out_num = 0, in_num = 0; |
| 659 | int rc; |
| 660 | |
| 661 | if (le16toh(desc[i].flags) & VRING_DESC_F_INDIRECT) { |
| 662 | if (le32toh(desc[i].len) % sizeof(struct vring_desc)) { |
| 663 | fprintf(stderr, "Invalid size for indirect buffer table\n"); |
| 664 | return NULL; |
| 665 | } |
| 666 | |
| 667 | /* loop over the indirect descriptor table */ |
| 668 | desc_addr = le64toh(desc[i].addr); |
| 669 | desc_len = le32toh(desc[i].len); |
| 670 | max = desc_len / sizeof(struct vring_desc); |
| 671 | read_len = desc_len; |
| 672 | desc = iova_to_va(dev, &read_len, desc_addr); |
| 673 | if (unlikely(desc && read_len != desc_len)) { |
| 674 | /* Failed to use zero copy */ |
| 675 | desc = NULL; |
| 676 | if (!vduse_queue_read_indirect_desc(dev, desc_buf, |
| 677 | desc_addr, |
| 678 | desc_len)) { |
| 679 | desc = desc_buf; |
| 680 | } |
| 681 | } |
| 682 | if (!desc) { |
| 683 | fprintf(stderr, "Invalid indirect buffer table\n"); |
| 684 | return NULL; |
| 685 | } |
| 686 | i = 0; |
| 687 | } |
| 688 | |
| 689 | /* Collect all the descriptors */ |
| 690 | do { |
| 691 | if (le16toh(desc[i].flags) & VRING_DESC_F_WRITE) { |
| 692 | if (!vduse_queue_map_single_desc(vq, &in_num, iov + out_num, |
| 693 | VIRTQUEUE_MAX_SIZE - out_num, |
| 694 | true, le64toh(desc[i].addr), |
| 695 | le32toh(desc[i].len))) { |
| 696 | return NULL; |
| 697 | } |
| 698 | } else { |
| 699 | if (in_num) { |
| 700 | fprintf(stderr, "Incorrect order for descriptors\n"); |
| 701 | return NULL; |
| 702 | } |
| 703 | if (!vduse_queue_map_single_desc(vq, &out_num, iov, |
| 704 | VIRTQUEUE_MAX_SIZE, false, |
| 705 | le64toh(desc[i].addr), |
| 706 | le32toh(desc[i].len))) { |
| 707 | return NULL; |
| 708 | } |
| 709 | } |
| 710 | |
| 711 | /* If we've got too many, that implies a descriptor loop. */ |
| 712 | if ((in_num + out_num) > max) { |
| 713 | fprintf(stderr, "Looped descriptor\n"); |
| 714 | return NULL; |
| 715 | } |
| 716 | rc = vduse_queue_read_next_desc(desc, i, max, &i); |
| 717 | } while (rc == VIRTQUEUE_READ_DESC_MORE); |
| 718 | |
| 719 | if (rc == VIRTQUEUE_READ_DESC_ERROR) { |
| 720 | fprintf(stderr, "read descriptor error\n"); |
| 721 | return NULL; |
| 722 | } |
| 723 | |
| 724 | /* Now copy what we have collected and mapped */ |
| 725 | elem = vduse_queue_alloc_element(sz, out_num, in_num); |
| 726 | if (!elem) { |
| 727 | fprintf(stderr, "read descriptor error\n"); |
| 728 | return NULL; |
| 729 | } |
| 730 | elem->index = idx; |
| 731 | for (i = 0; i < out_num; i++) { |
| 732 | elem->out_sg[i] = iov[i]; |
| 733 | } |
| 734 | for (i = 0; i < in_num; i++) { |
| 735 | elem->in_sg[i] = iov[out_num + i]; |
| 736 | } |
| 737 | |
| 738 | return elem; |
| 739 | } |
| 740 | |
| 741 | void *vduse_queue_pop(VduseVirtq *vq, size_t sz) |
| 742 | { |
| 743 | unsigned int head; |
| 744 | VduseVirtqElement *elem; |
| 745 | VduseDev *dev = vq->dev; |
| 746 | int i; |
| 747 | |
| 748 | if (unlikely(!vq->vring.avail)) { |
| 749 | return NULL; |
| 750 | } |
| 751 | |
| 752 | if (unlikely(vq->resubmit_list && vq->resubmit_num > 0)) { |
| 753 | i = (--vq->resubmit_num); |
| 754 | elem = vduse_queue_map_desc(vq, vq->resubmit_list[i].index, sz); |
| 755 | |
| 756 | if (!vq->resubmit_num) { |
| 757 | free(vq->resubmit_list); |
| 758 | vq->resubmit_list = NULL; |
| 759 | } |
| 760 | |
| 761 | return elem; |
| 762 | } |
| 763 | |
| 764 | if (vduse_queue_empty(vq)) { |
| 765 | return NULL; |
| 766 | } |
| 767 | /* Needed after virtio_queue_empty() */ |
| 768 | smp_rmb(); |
| 769 | |
| 770 | if (vq->inuse >= vq->vring.num) { |
| 771 | fprintf(stderr, "Virtqueue size exceeded: %d\n", vq->inuse); |
| 772 | return NULL; |
| 773 | } |
| 774 | |
| 775 | if (!vduse_queue_get_head(vq, vq->last_avail_idx++, &head)) { |
| 776 | return NULL; |
| 777 | } |
| 778 | |
| 779 | if (vduse_dev_has_feature(dev, VIRTIO_RING_F_EVENT_IDX)) { |
| 780 | vring_set_avail_event(vq, vq->last_avail_idx); |
| 781 | } |
| 782 | |
| 783 | elem = vduse_queue_map_desc(vq, head, sz); |
| 784 | |
| 785 | if (!elem) { |
| 786 | return NULL; |
| 787 | } |
| 788 | |
| 789 | vq->inuse++; |
| 790 | |
| 791 | vduse_queue_inflight_get(vq, head); |
| 792 | |
| 793 | return elem; |
| 794 | } |
| 795 | |
| 796 | static inline void vring_used_write(VduseVirtq *vq, |
| 797 | struct vring_used_elem *uelem, int i) |
| 798 | { |
| 799 | struct vring_used *used = vq->vring.used; |
| 800 | |
| 801 | used->ring[i] = *uelem; |
| 802 | } |
| 803 | |
| 804 | static void vduse_queue_fill(VduseVirtq *vq, const VduseVirtqElement *elem, |
| 805 | unsigned int len, unsigned int idx) |
| 806 | { |
| 807 | struct vring_used_elem uelem; |
| 808 | |
| 809 | if (unlikely(!vq->vring.used)) { |
| 810 | return; |
| 811 | } |
| 812 | |
| 813 | idx = (idx + vq->used_idx) % vq->vring.num; |
| 814 | |
| 815 | uelem.id = htole32(elem->index); |
| 816 | uelem.len = htole32(len); |
| 817 | vring_used_write(vq, &uelem, idx); |
| 818 | } |
| 819 | |
| 820 | static inline void vring_used_idx_set(VduseVirtq *vq, uint16_t val) |
| 821 | { |
| 822 | vq->vring.used->idx = htole16(val); |
| 823 | vq->used_idx = val; |
| 824 | } |
| 825 | |
| 826 | static void vduse_queue_flush(VduseVirtq *vq, unsigned int count) |
| 827 | { |
| 828 | uint16_t old, new; |
| 829 | |
| 830 | if (unlikely(!vq->vring.used)) { |
| 831 | return; |
| 832 | } |
| 833 | |
| 834 | /* Make sure buffer is written before we update index. */ |
| 835 | smp_wmb(); |
| 836 | |
| 837 | old = vq->used_idx; |
| 838 | new = old + count; |
| 839 | vring_used_idx_set(vq, new); |
| 840 | vq->inuse -= count; |
| 841 | if (unlikely((int16_t)(new - vq->signalled_used) < (uint16_t)(new - old))) { |
| 842 | vq->signalled_used_valid = false; |
| 843 | } |
| 844 | } |
| 845 | |
| 846 | void vduse_queue_push(VduseVirtq *vq, const VduseVirtqElement *elem, |
| 847 | unsigned int len) |
| 848 | { |
| 849 | vduse_queue_fill(vq, elem, len, 0); |
| 850 | vduse_queue_inflight_pre_put(vq, elem->index); |
| 851 | vduse_queue_flush(vq, 1); |
| 852 | vduse_queue_inflight_post_put(vq, elem->index); |
| 853 | } |
| 854 | |
| 855 | static int vduse_queue_update_vring(VduseVirtq *vq, uint64_t desc_addr, |
| 856 | uint64_t avail_addr, uint64_t used_addr) |
| 857 | { |
| 858 | struct VduseDev *dev = vq->dev; |
| 859 | uint64_t len; |
| 860 | |
| 861 | len = sizeof(struct vring_desc); |
| 862 | vq->vring.desc = iova_to_va(dev, &len, desc_addr); |
| 863 | if (len != sizeof(struct vring_desc)) { |
| 864 | return -EINVAL; |
| 865 | } |
| 866 | |
| 867 | len = sizeof(struct vring_avail); |
| 868 | vq->vring.avail = iova_to_va(dev, &len, avail_addr); |
| 869 | if (len != sizeof(struct vring_avail)) { |
| 870 | return -EINVAL; |
| 871 | } |
| 872 | |
| 873 | len = sizeof(struct vring_used); |
| 874 | vq->vring.used = iova_to_va(dev, &len, used_addr); |
| 875 | if (len != sizeof(struct vring_used)) { |
| 876 | return -EINVAL; |
| 877 | } |
| 878 | |
| 879 | if (!vq->vring.desc || !vq->vring.avail || !vq->vring.used) { |
| 880 | fprintf(stderr, "Failed to get vq[%d] iova mapping\n", vq->index); |
| 881 | return -EINVAL; |
| 882 | } |
| 883 | |
| 884 | return 0; |
| 885 | } |
| 886 | |
| 887 | static void vduse_queue_enable(VduseVirtq *vq) |
| 888 | { |
| 889 | struct VduseDev *dev = vq->dev; |
| 890 | struct vduse_vq_info vq_info; |
| 891 | struct vduse_vq_eventfd vq_eventfd; |
| 892 | int fd; |
| 893 | |
| 894 | vq_info.index = vq->index; |
| 895 | if (ioctl(dev->fd, VDUSE_VQ_GET_INFO, &vq_info)) { |
| 896 | fprintf(stderr, "Failed to get vq[%d] info: %s\n", |
| 897 | vq->index, strerror(errno)); |
| 898 | return; |
| 899 | } |
| 900 | |
| 901 | if (!vq_info.ready) { |
| 902 | return; |
| 903 | } |
| 904 | |
| 905 | if (vq_info.num > VIRTQUEUE_MAX_SIZE) { |
| 906 | fprintf(stderr, "vq[%d] vring num %u exceeds max %u\n", |
| 907 | vq->index, vq_info.num, VIRTQUEUE_MAX_SIZE); |
| 908 | return; |
| 909 | } |
| 910 | vq->vring.num = vq_info.num; |
| 911 | vq->vring.desc_addr = vq_info.desc_addr; |
| 912 | vq->vring.avail_addr = vq_info.driver_addr; |
| 913 | vq->vring.used_addr = vq_info.device_addr; |
| 914 | |
| 915 | if (vduse_queue_update_vring(vq, vq_info.desc_addr, |
| 916 | vq_info.driver_addr, vq_info.device_addr)) { |
| 917 | fprintf(stderr, "Failed to update vring for vq[%d]\n", vq->index); |
| 918 | return; |
| 919 | } |
| 920 | |
| 921 | fd = eventfd(0, EFD_NONBLOCK | EFD_CLOEXEC); |
| 922 | if (fd < 0) { |
| 923 | fprintf(stderr, "Failed to init eventfd for vq[%d]\n", vq->index); |
| 924 | return; |
| 925 | } |
| 926 | |
| 927 | vq_eventfd.index = vq->index; |
| 928 | vq_eventfd.fd = fd; |
| 929 | if (ioctl(dev->fd, VDUSE_VQ_SETUP_KICKFD, &vq_eventfd)) { |
| 930 | fprintf(stderr, "Failed to setup kick fd for vq[%d]\n", vq->index); |
| 931 | close(fd); |
| 932 | return; |
| 933 | } |
| 934 | |
| 935 | vq->fd = fd; |
| 936 | vq->signalled_used_valid = false; |
| 937 | vq->ready = true; |
| 938 | |
| 939 | if (vduse_queue_check_inflights(vq)) { |
| 940 | fprintf(stderr, "Failed to check inflights for vq[%d]\n", vq->index); |
| 941 | close(fd); |
| 942 | return; |
| 943 | } |
| 944 | |
| 945 | dev->ops->enable_queue(dev, vq); |
| 946 | } |
| 947 | |
| 948 | static void vduse_queue_disable(VduseVirtq *vq) |
| 949 | { |
| 950 | struct VduseDev *dev = vq->dev; |
| 951 | struct vduse_vq_eventfd eventfd; |
| 952 | |
| 953 | if (!vq->ready) { |
| 954 | return; |
| 955 | } |
| 956 | |
| 957 | dev->ops->disable_queue(dev, vq); |
| 958 | |
| 959 | eventfd.index = vq->index; |
| 960 | eventfd.fd = VDUSE_EVENTFD_DEASSIGN; |
| 961 | ioctl(dev->fd, VDUSE_VQ_SETUP_KICKFD, &eventfd); |
| 962 | close(vq->fd); |
| 963 | |
| 964 | assert(vq->inuse == 0); |
| 965 | |
| 966 | vq->vring.num = 0; |
| 967 | vq->vring.desc_addr = 0; |
| 968 | vq->vring.avail_addr = 0; |
| 969 | vq->vring.used_addr = 0; |
| 970 | vq->vring.desc = 0; |
| 971 | vq->vring.avail = 0; |
| 972 | vq->vring.used = 0; |
| 973 | vq->ready = false; |
| 974 | vq->fd = -1; |
| 975 | } |
| 976 | |
| 977 | static void vduse_dev_start_dataplane(VduseDev *dev) |
| 978 | { |
| 979 | int i; |
| 980 | |
| 981 | if (ioctl(dev->fd, VDUSE_DEV_GET_FEATURES, &dev->features)) { |
| 982 | fprintf(stderr, "Failed to get features: %s\n", strerror(errno)); |
| 983 | return; |
| 984 | } |
| 985 | assert(vduse_dev_has_feature(dev, VIRTIO_F_VERSION_1)); |
| 986 | |
| 987 | for (i = 0; i < dev->num_queues; i++) { |
| 988 | vduse_queue_enable(&dev->vqs[i]); |
| 989 | } |
| 990 | } |
| 991 | |
| 992 | static void vduse_dev_stop_dataplane(VduseDev *dev) |
| 993 | { |
| 994 | size_t log_size = dev->num_queues * vduse_vq_log_size(VIRTQUEUE_MAX_SIZE); |
| 995 | int i; |
| 996 | |
| 997 | for (i = 0; i < dev->num_queues; i++) { |
| 998 | vduse_queue_disable(&dev->vqs[i]); |
| 999 | } |
| 1000 | if (dev->log) { |
| 1001 | memset(dev->log, 0, log_size); |
| 1002 | } |
| 1003 | dev->features = 0; |
| 1004 | vduse_iova_remove_region(dev, 0, ULONG_MAX); |
| 1005 | } |
| 1006 | |
| 1007 | int vduse_dev_handler(VduseDev *dev) |
| 1008 | { |
| 1009 | struct vduse_dev_request req; |
| 1010 | struct vduse_dev_response resp = { 0 }; |
| 1011 | VduseVirtq *vq; |
| 1012 | int i, ret; |
| 1013 | |
| 1014 | ret = read(dev->fd, &req, sizeof(req)); |
| 1015 | if (ret != sizeof(req)) { |
| 1016 | fprintf(stderr, "Read request error [%d]: %s\n", |
| 1017 | ret, strerror(errno)); |
| 1018 | return -errno; |
| 1019 | } |
| 1020 | resp.request_id = req.request_id; |
| 1021 | |
| 1022 | switch (req.type) { |
| 1023 | case VDUSE_GET_VQ_STATE: |
| 1024 | vq = &dev->vqs[req.vq_state.index]; |
| 1025 | resp.vq_state.split.avail_index = vq->last_avail_idx; |
| 1026 | resp.result = VDUSE_REQ_RESULT_OK; |
| 1027 | break; |
| 1028 | case VDUSE_SET_STATUS: |
| 1029 | if (req.s.status & VIRTIO_CONFIG_S_DRIVER_OK) { |
| 1030 | vduse_dev_start_dataplane(dev); |
| 1031 | } else if (req.s.status == 0) { |
| 1032 | vduse_dev_stop_dataplane(dev); |
| 1033 | } |
| 1034 | resp.result = VDUSE_REQ_RESULT_OK; |
| 1035 | break; |
| 1036 | case VDUSE_UPDATE_IOTLB: |
| 1037 | /* The iova will be updated by iova_to_va() later, so just remove it */ |
| 1038 | vduse_iova_remove_region(dev, req.iova.start, req.iova.last); |
| 1039 | for (i = 0; i < dev->num_queues; i++) { |
| 1040 | vq = &dev->vqs[i]; |
| 1041 | if (vq->ready) { |
| 1042 | if (vduse_queue_update_vring(vq, vq->vring.desc_addr, |
| 1043 | vq->vring.avail_addr, |
| 1044 | vq->vring.used_addr)) { |
| 1045 | fprintf(stderr, "Failed to update vring for vq[%d]\n", |
| 1046 | vq->index); |
| 1047 | } |
| 1048 | } |
| 1049 | } |
| 1050 | resp.result = VDUSE_REQ_RESULT_OK; |
| 1051 | break; |
| 1052 | default: |
| 1053 | resp.result = VDUSE_REQ_RESULT_FAILED; |
| 1054 | break; |
| 1055 | } |
| 1056 | |
| 1057 | ret = write(dev->fd, &resp, sizeof(resp)); |
| 1058 | if (ret != sizeof(resp)) { |
| 1059 | fprintf(stderr, "Write request %d error [%d]: %s\n", |
| 1060 | req.type, ret, strerror(errno)); |
| 1061 | return -errno; |
| 1062 | } |
| 1063 | return 0; |
| 1064 | } |
| 1065 | |
| 1066 | int vduse_dev_update_config(VduseDev *dev, uint32_t size, |
| 1067 | uint32_t offset, char *buffer) |
| 1068 | { |
| 1069 | int ret; |
| 1070 | struct vduse_config_data *data; |
| 1071 | |
| 1072 | data = malloc(offsetof(struct vduse_config_data, buffer) + size); |
| 1073 | if (!data) { |
| 1074 | return -ENOMEM; |
| 1075 | } |
| 1076 | |
| 1077 | data->offset = offset; |
| 1078 | data->length = size; |
| 1079 | memcpy(data->buffer, buffer, size); |
| 1080 | |
| 1081 | ret = ioctl(dev->fd, VDUSE_DEV_SET_CONFIG, data); |
| 1082 | free(data); |
| 1083 | |
| 1084 | if (ret) { |
| 1085 | return -errno; |
| 1086 | } |
| 1087 | |
| 1088 | if (ioctl(dev->fd, VDUSE_DEV_INJECT_CONFIG_IRQ)) { |
| 1089 | return -errno; |
| 1090 | } |
| 1091 | |
| 1092 | return 0; |
| 1093 | } |
| 1094 | |
| 1095 | int vduse_dev_setup_queue(VduseDev *dev, int index, int max_size) |
| 1096 | { |
| 1097 | VduseVirtq *vq = &dev->vqs[index]; |
| 1098 | struct vduse_vq_config vq_config = { 0 }; |
| 1099 | |
| 1100 | if (max_size > VIRTQUEUE_MAX_SIZE) { |
| 1101 | return -EINVAL; |
| 1102 | } |
| 1103 | |
| 1104 | vq_config.index = vq->index; |
| 1105 | vq_config.max_size = max_size; |
| 1106 | |
| 1107 | if (ioctl(dev->fd, VDUSE_VQ_SETUP, &vq_config)) { |
| 1108 | return -errno; |
| 1109 | } |
| 1110 | |
| 1111 | vduse_queue_enable(vq); |
| 1112 | |
| 1113 | return 0; |
| 1114 | } |
| 1115 | |
| 1116 | int vduse_set_reconnect_log_file(VduseDev *dev, const char *filename) |
| 1117 | { |
| 1118 | |
| 1119 | size_t log_size = dev->num_queues * vduse_vq_log_size(VIRTQUEUE_MAX_SIZE); |
| 1120 | void *log; |
| 1121 | int i; |
| 1122 | |
| 1123 | dev->log = log = vduse_log_get(filename, log_size); |
| 1124 | if (log == MAP_FAILED) { |
| 1125 | fprintf(stderr, "Failed to get vduse log\n"); |
| 1126 | return -EINVAL; |
| 1127 | } |
| 1128 | |
| 1129 | for (i = 0; i < dev->num_queues; i++) { |
| 1130 | dev->vqs[i].log = log; |
| 1131 | dev->vqs[i].log->inflight.desc_num = VIRTQUEUE_MAX_SIZE; |
| 1132 | log = (void *)((char *)log + vduse_vq_log_size(VIRTQUEUE_MAX_SIZE)); |
| 1133 | } |
| 1134 | |
| 1135 | return 0; |
| 1136 | } |
| 1137 | |
| 1138 | static int vduse_dev_init_vqs(VduseDev *dev, uint16_t num_queues) |
| 1139 | { |
| 1140 | VduseVirtq *vqs; |
| 1141 | int i; |
| 1142 | |
| 1143 | vqs = calloc(sizeof(VduseVirtq), num_queues); |
| 1144 | if (!vqs) { |
| 1145 | return -ENOMEM; |
| 1146 | } |
| 1147 | |
| 1148 | for (i = 0; i < num_queues; i++) { |
| 1149 | vqs[i].index = i; |
| 1150 | vqs[i].dev = dev; |
| 1151 | vqs[i].fd = -1; |
| 1152 | } |
| 1153 | dev->vqs = vqs; |
| 1154 | |
| 1155 | return 0; |
| 1156 | } |
| 1157 | |
| 1158 | static int vduse_dev_init(VduseDev *dev, const char *name, |
| 1159 | uint16_t num_queues, const VduseOps *ops, |
| 1160 | void *priv) |
| 1161 | { |
| 1162 | char *dev_path, *dev_name; |
| 1163 | int ret, fd; |
| 1164 | |
| 1165 | dev_path = malloc(strlen(name) + strlen("/dev/vduse/") + 1); |
| 1166 | if (!dev_path) { |
| 1167 | return -ENOMEM; |
| 1168 | } |
| 1169 | sprintf(dev_path, "/dev/vduse/%s", name); |
| 1170 | |
| 1171 | fd = open(dev_path, O_RDWR); |
| 1172 | free(dev_path); |
| 1173 | if (fd < 0) { |
| 1174 | fprintf(stderr, "Failed to open vduse dev %s: %s\n", |
| 1175 | name, strerror(errno)); |
| 1176 | return -errno; |
| 1177 | } |
| 1178 | |
| 1179 | if (ioctl(fd, VDUSE_DEV_GET_FEATURES, &dev->features)) { |
| 1180 | fprintf(stderr, "Failed to get features: %s\n", strerror(errno)); |
| 1181 | close(fd); |
| 1182 | return -errno; |
| 1183 | } |
| 1184 | |
| 1185 | dev_name = strdup(name); |
| 1186 | if (!dev_name) { |
| 1187 | close(fd); |
| 1188 | return -ENOMEM; |
| 1189 | } |
| 1190 | |
| 1191 | ret = vduse_dev_init_vqs(dev, num_queues); |
| 1192 | if (ret) { |
| 1193 | free(dev_name); |
| 1194 | close(fd); |
| 1195 | return ret; |
| 1196 | } |
| 1197 | |
| 1198 | dev->name = dev_name; |
| 1199 | dev->num_queues = num_queues; |
| 1200 | dev->fd = fd; |
| 1201 | dev->ops = ops; |
| 1202 | dev->priv = priv; |
| 1203 | |
| 1204 | return 0; |
| 1205 | } |
| 1206 | |
| 1207 | static inline bool vduse_name_is_invalid(const char *name) |
| 1208 | { |
| 1209 | return strlen(name) >= VDUSE_NAME_MAX || strstr(name, ".."); |
| 1210 | } |
| 1211 | |
| 1212 | VduseDev *vduse_dev_create_by_fd(int fd, uint16_t num_queues, |
| 1213 | const VduseOps *ops, void *priv) |
| 1214 | { |
| 1215 | VduseDev *dev; |
| 1216 | int ret; |
| 1217 | |
| 1218 | if (!ops || !ops->enable_queue || !ops->disable_queue) { |
| 1219 | fprintf(stderr, "Invalid parameter for vduse\n"); |
| 1220 | return NULL; |
| 1221 | } |
| 1222 | |
| 1223 | dev = calloc(sizeof(VduseDev), 1); |
| 1224 | if (!dev) { |
| 1225 | fprintf(stderr, "Failed to allocate vduse device\n"); |
| 1226 | return NULL; |
| 1227 | } |
| 1228 | |
| 1229 | if (ioctl(fd, VDUSE_DEV_GET_FEATURES, &dev->features)) { |
| 1230 | fprintf(stderr, "Failed to get features: %s\n", strerror(errno)); |
| 1231 | free(dev); |
| 1232 | return NULL; |
| 1233 | } |
| 1234 | |
| 1235 | ret = vduse_dev_init_vqs(dev, num_queues); |
| 1236 | if (ret) { |
| 1237 | fprintf(stderr, "Failed to init vqs\n"); |
| 1238 | free(dev); |
| 1239 | return NULL; |
| 1240 | } |
| 1241 | |
| 1242 | dev->num_queues = num_queues; |
| 1243 | dev->fd = fd; |
| 1244 | dev->ops = ops; |
| 1245 | dev->priv = priv; |
| 1246 | |
| 1247 | return dev; |
| 1248 | } |
| 1249 | |
| 1250 | VduseDev *vduse_dev_create_by_name(const char *name, uint16_t num_queues, |
| 1251 | const VduseOps *ops, void *priv) |
| 1252 | { |
| 1253 | VduseDev *dev; |
| 1254 | int ret; |
| 1255 | |
| 1256 | if (!name || vduse_name_is_invalid(name) || !ops || |
| 1257 | !ops->enable_queue || !ops->disable_queue) { |
| 1258 | fprintf(stderr, "Invalid parameter for vduse\n"); |
| 1259 | return NULL; |
| 1260 | } |
| 1261 | |
| 1262 | dev = calloc(sizeof(VduseDev), 1); |
| 1263 | if (!dev) { |
| 1264 | fprintf(stderr, "Failed to allocate vduse device\n"); |
| 1265 | return NULL; |
| 1266 | } |
| 1267 | |
| 1268 | ret = vduse_dev_init(dev, name, num_queues, ops, priv); |
| 1269 | if (ret < 0) { |
| 1270 | fprintf(stderr, "Failed to init vduse device %s: %s\n", |
| 1271 | name, strerror(-ret)); |
| 1272 | free(dev); |
| 1273 | return NULL; |
| 1274 | } |
| 1275 | |
| 1276 | return dev; |
| 1277 | } |
| 1278 | |
| 1279 | VduseDev *vduse_dev_create(const char *name, uint32_t device_id, |
| 1280 | uint32_t vendor_id, uint64_t features, |
| 1281 | uint16_t num_queues, uint32_t config_size, |
| 1282 | char *config, const VduseOps *ops, void *priv) |
| 1283 | { |
| 1284 | VduseDev *dev; |
| 1285 | int ret, ctrl_fd; |
| 1286 | uint64_t version; |
| 1287 | struct vduse_dev_config *dev_config; |
| 1288 | size_t size = offsetof(struct vduse_dev_config, config); |
| 1289 | |
| 1290 | if (!name || vduse_name_is_invalid(name) || |
| 1291 | !has_feature(features, VIRTIO_F_VERSION_1) || !config || |
| 1292 | !config_size || !ops || !ops->enable_queue || !ops->disable_queue) { |
| 1293 | fprintf(stderr, "Invalid parameter for vduse\n"); |
| 1294 | return NULL; |
| 1295 | } |
| 1296 | |
| 1297 | dev = calloc(sizeof(VduseDev), 1); |
| 1298 | if (!dev) { |
| 1299 | fprintf(stderr, "Failed to allocate vduse device\n"); |
| 1300 | return NULL; |
| 1301 | } |
| 1302 | |
| 1303 | ctrl_fd = open("/dev/vduse/control", O_RDWR); |
| 1304 | if (ctrl_fd < 0) { |
| 1305 | fprintf(stderr, "Failed to open /dev/vduse/control: %s\n", |
| 1306 | strerror(errno)); |
| 1307 | goto err_ctrl; |
| 1308 | } |
| 1309 | |
| 1310 | version = VDUSE_API_VERSION; |
| 1311 | if (ioctl(ctrl_fd, VDUSE_SET_API_VERSION, &version)) { |
| 1312 | fprintf(stderr, "Failed to set api version %" PRIu64 ": %s\n", |
| 1313 | version, strerror(errno)); |
| 1314 | goto err_dev; |
| 1315 | } |
| 1316 | |
| 1317 | dev_config = calloc(size + config_size, 1); |
| 1318 | if (!dev_config) { |
| 1319 | fprintf(stderr, "Failed to allocate config space\n"); |
| 1320 | goto err_dev; |
| 1321 | } |
| 1322 | |
| 1323 | assert(!vduse_name_is_invalid(name)); |
| 1324 | strcpy(dev_config->name, name); |
| 1325 | dev_config->device_id = device_id; |
| 1326 | dev_config->vendor_id = vendor_id; |
| 1327 | dev_config->features = features; |
| 1328 | dev_config->vq_num = num_queues; |
| 1329 | dev_config->vq_align = VDUSE_VQ_ALIGN; |
| 1330 | dev_config->config_size = config_size; |
| 1331 | memcpy(dev_config->config, config, config_size); |
| 1332 | |
| 1333 | ret = ioctl(ctrl_fd, VDUSE_CREATE_DEV, dev_config); |
| 1334 | free(dev_config); |
| 1335 | if (ret && errno != EEXIST) { |
| 1336 | fprintf(stderr, "Failed to create vduse device %s: %s\n", |
| 1337 | name, strerror(errno)); |
| 1338 | goto err_dev; |
| 1339 | } |
| 1340 | dev->ctrl_fd = ctrl_fd; |
| 1341 | |
| 1342 | ret = vduse_dev_init(dev, name, num_queues, ops, priv); |
| 1343 | if (ret < 0) { |
| 1344 | fprintf(stderr, "Failed to init vduse device %s: %s\n", |
| 1345 | name, strerror(-ret)); |
| 1346 | goto err; |
| 1347 | } |
| 1348 | |
| 1349 | return dev; |
| 1350 | err: |
| 1351 | ioctl(ctrl_fd, VDUSE_DESTROY_DEV, name); |
| 1352 | err_dev: |
| 1353 | close(ctrl_fd); |
| 1354 | err_ctrl: |
| 1355 | free(dev); |
| 1356 | |
| 1357 | return NULL; |
| 1358 | } |
| 1359 | |
| 1360 | int vduse_dev_destroy(VduseDev *dev) |
| 1361 | { |
| 1362 | size_t log_size = dev->num_queues * vduse_vq_log_size(VIRTQUEUE_MAX_SIZE); |
| 1363 | int i, ret = 0; |
| 1364 | |
| 1365 | if (dev->log) { |
| 1366 | munmap(dev->log, log_size); |
| 1367 | } |
| 1368 | for (i = 0; i < dev->num_queues; i++) { |
| 1369 | free(dev->vqs[i].resubmit_list); |
| 1370 | } |
| 1371 | free(dev->vqs); |
| 1372 | if (dev->fd >= 0) { |
| 1373 | close(dev->fd); |
| 1374 | dev->fd = -1; |
| 1375 | } |
| 1376 | if (dev->ctrl_fd >= 0) { |
| 1377 | if (ioctl(dev->ctrl_fd, VDUSE_DESTROY_DEV, dev->name)) { |
| 1378 | ret = -errno; |
| 1379 | } |
| 1380 | close(dev->ctrl_fd); |
| 1381 | dev->ctrl_fd = -1; |
| 1382 | } |
| 1383 | free(dev->name); |
| 1384 | free(dev); |
| 1385 | |
| 1386 | return ret; |
| 1387 | } |