1 /* 2 * Virtio Block Device 3 * 4 * Copyright IBM, Corp. 2007 5 * 6 * Authors: 7 * Anthony Liguori <aliguori@us.ibm.com> 8 * 9 * This work is licensed under the terms of the GNU GPL, version 2. See 10 * the COPYING file in the top-level directory. 11 * 12 */ 13 14 #include "qemu/osdep.h" 15 #include "qemu/defer-call.h" 16 #include "qapi/error.h" 17 #include "qemu/iov.h" 18 #include "qemu/module.h" 19 #include "qemu/error-report.h" 20 #include "qemu/main-loop.h" 21 #include "block/block_int.h" 22 #include "trace.h" 23 #include "hw/block/block.h" 24 #include "hw/qdev-properties.h" 25 #include "sysemu/blockdev.h" 26 #include "sysemu/block-ram-registrar.h" 27 #include "sysemu/sysemu.h" 28 #include "sysemu/runstate.h" 29 #include "hw/virtio/virtio-blk.h" 30 #include "scsi/constants.h" 31 #ifdef __linux__ 32 # include <scsi/sg.h> 33 #endif 34 #include "hw/virtio/virtio-bus.h" 35 #include "migration/qemu-file-types.h" 36 #include "hw/virtio/virtio-access.h" 37 #include "hw/virtio/virtio-blk-common.h" 38 #include "qemu/coroutine.h" 39 40 static void virtio_blk_init_request(VirtIOBlock *s, VirtQueue *vq, 41 VirtIOBlockReq *req) 42 { 43 req->dev = s; 44 req->vq = vq; 45 req->qiov.size = 0; 46 req->in_len = 0; 47 req->next = NULL; 48 req->mr_next = NULL; 49 } 50 51 static void virtio_blk_free_request(VirtIOBlockReq *req) 52 { 53 g_free(req); 54 } 55 56 static void virtio_blk_req_complete(VirtIOBlockReq *req, unsigned char status) 57 { 58 VirtIOBlock *s = req->dev; 59 VirtIODevice *vdev = VIRTIO_DEVICE(s); 60 61 trace_virtio_blk_req_complete(vdev, req, status); 62 63 stb_p(&req->in->status, status); 64 iov_discard_undo(&req->inhdr_undo); 65 iov_discard_undo(&req->outhdr_undo); 66 virtqueue_push(req->vq, &req->elem, req->in_len); 67 if (s->ioeventfd_started && !s->ioeventfd_disabled) { 68 virtio_notify_irqfd(vdev, req->vq); 69 } else { 70 virtio_notify(vdev, req->vq); 71 } 72 } 73 74 static int virtio_blk_handle_rw_error(VirtIOBlockReq *req, int error, 75 bool is_read, bool acct_failed) 76 { 77 VirtIOBlock *s = req->dev; 78 BlockErrorAction action = blk_get_error_action(s->blk, is_read, error); 79 80 if (action == BLOCK_ERROR_ACTION_STOP) { 81 /* Break the link as the next request is going to be parsed from the 82 * ring again. Otherwise we may end up doing a double completion! */ 83 req->mr_next = NULL; 84 85 WITH_QEMU_LOCK_GUARD(&s->rq_lock) { 86 req->next = s->rq; 87 s->rq = req; 88 } 89 } else if (action == BLOCK_ERROR_ACTION_REPORT) { 90 virtio_blk_req_complete(req, VIRTIO_BLK_S_IOERR); 91 if (acct_failed) { 92 block_acct_failed(blk_get_stats(s->blk), &req->acct); 93 } 94 virtio_blk_free_request(req); 95 } 96 97 blk_error_action(s->blk, action, is_read, error); 98 return action != BLOCK_ERROR_ACTION_IGNORE; 99 } 100 101 static void virtio_blk_rw_complete(void *opaque, int ret) 102 { 103 VirtIOBlockReq *next = opaque; 104 VirtIOBlock *s = next->dev; 105 VirtIODevice *vdev = VIRTIO_DEVICE(s); 106 107 while (next) { 108 VirtIOBlockReq *req = next; 109 next = req->mr_next; 110 trace_virtio_blk_rw_complete(vdev, req, ret); 111 112 if (req->qiov.nalloc != -1) { 113 /* If nalloc is != -1 req->qiov is a local copy of the original 114 * external iovec. It was allocated in submit_requests to be 115 * able to merge requests. */ 116 qemu_iovec_destroy(&req->qiov); 117 } 118 119 if (ret) { 120 int p = virtio_ldl_p(VIRTIO_DEVICE(s), &req->out.type); 121 bool is_read = !(p & VIRTIO_BLK_T_OUT); 122 /* Note that memory may be dirtied on read failure. If the 123 * virtio request is not completed here, as is the case for 124 * BLOCK_ERROR_ACTION_STOP, the memory may not be copied 125 * correctly during live migration. While this is ugly, 126 * it is acceptable because the device is free to write to 127 * the memory until the request is completed (which will 128 * happen on the other side of the migration). 129 */ 130 if (virtio_blk_handle_rw_error(req, -ret, is_read, true)) { 131 continue; 132 } 133 } 134 135 virtio_blk_req_complete(req, VIRTIO_BLK_S_OK); 136 block_acct_done(blk_get_stats(s->blk), &req->acct); 137 virtio_blk_free_request(req); 138 } 139 } 140 141 static void virtio_blk_flush_complete(void *opaque, int ret) 142 { 143 VirtIOBlockReq *req = opaque; 144 VirtIOBlock *s = req->dev; 145 146 if (ret && virtio_blk_handle_rw_error(req, -ret, 0, true)) { 147 return; 148 } 149 150 virtio_blk_req_complete(req, VIRTIO_BLK_S_OK); 151 block_acct_done(blk_get_stats(s->blk), &req->acct); 152 virtio_blk_free_request(req); 153 } 154 155 static void virtio_blk_discard_write_zeroes_complete(void *opaque, int ret) 156 { 157 VirtIOBlockReq *req = opaque; 158 VirtIOBlock *s = req->dev; 159 bool is_write_zeroes = (virtio_ldl_p(VIRTIO_DEVICE(s), &req->out.type) & 160 ~VIRTIO_BLK_T_BARRIER) == VIRTIO_BLK_T_WRITE_ZEROES; 161 162 if (ret && virtio_blk_handle_rw_error(req, -ret, false, is_write_zeroes)) { 163 return; 164 } 165 166 virtio_blk_req_complete(req, VIRTIO_BLK_S_OK); 167 if (is_write_zeroes) { 168 block_acct_done(blk_get_stats(s->blk), &req->acct); 169 } 170 virtio_blk_free_request(req); 171 } 172 173 #ifdef __linux__ 174 175 typedef struct { 176 VirtIOBlockReq *req; 177 struct sg_io_hdr hdr; 178 } VirtIOBlockIoctlReq; 179 180 static void virtio_blk_ioctl_complete(void *opaque, int status) 181 { 182 VirtIOBlockIoctlReq *ioctl_req = opaque; 183 VirtIOBlockReq *req = ioctl_req->req; 184 VirtIOBlock *s = req->dev; 185 VirtIODevice *vdev = VIRTIO_DEVICE(s); 186 struct virtio_scsi_inhdr *scsi; 187 struct sg_io_hdr *hdr; 188 189 scsi = (void *)req->elem.in_sg[req->elem.in_num - 2].iov_base; 190 191 if (status) { 192 status = VIRTIO_BLK_S_UNSUPP; 193 virtio_stl_p(vdev, &scsi->errors, 255); 194 goto out; 195 } 196 197 hdr = &ioctl_req->hdr; 198 /* 199 * From SCSI-Generic-HOWTO: "Some lower level drivers (e.g. ide-scsi) 200 * clear the masked_status field [hence status gets cleared too, see 201 * block/scsi_ioctl.c] even when a CHECK_CONDITION or COMMAND_TERMINATED 202 * status has occurred. However they do set DRIVER_SENSE in driver_status 203 * field. Also a (sb_len_wr > 0) indicates there is a sense buffer. 204 */ 205 if (hdr->status == 0 && hdr->sb_len_wr > 0) { 206 hdr->status = CHECK_CONDITION; 207 } 208 209 virtio_stl_p(vdev, &scsi->errors, 210 hdr->status | (hdr->msg_status << 8) | 211 (hdr->host_status << 16) | (hdr->driver_status << 24)); 212 virtio_stl_p(vdev, &scsi->residual, hdr->resid); 213 virtio_stl_p(vdev, &scsi->sense_len, hdr->sb_len_wr); 214 virtio_stl_p(vdev, &scsi->data_len, hdr->dxfer_len); 215 216 out: 217 virtio_blk_req_complete(req, status); 218 virtio_blk_free_request(req); 219 g_free(ioctl_req); 220 } 221 222 #endif 223 224 static VirtIOBlockReq *virtio_blk_get_request(VirtIOBlock *s, VirtQueue *vq) 225 { 226 VirtIOBlockReq *req = virtqueue_pop(vq, sizeof(VirtIOBlockReq)); 227 228 if (req) { 229 virtio_blk_init_request(s, vq, req); 230 } 231 return req; 232 } 233 234 static int virtio_blk_handle_scsi_req(VirtIOBlockReq *req) 235 { 236 int status = VIRTIO_BLK_S_OK; 237 struct virtio_scsi_inhdr *scsi = NULL; 238 VirtIOBlock *blk = req->dev; 239 VirtIODevice *vdev = VIRTIO_DEVICE(blk); 240 VirtQueueElement *elem = &req->elem; 241 242 #ifdef __linux__ 243 int i; 244 VirtIOBlockIoctlReq *ioctl_req; 245 BlockAIOCB *acb; 246 #endif 247 248 /* 249 * We require at least one output segment each for the virtio_blk_outhdr 250 * and the SCSI command block. 251 * 252 * We also at least require the virtio_blk_inhdr, the virtio_scsi_inhdr 253 * and the sense buffer pointer in the input segments. 254 */ 255 if (elem->out_num < 2 || elem->in_num < 3) { 256 status = VIRTIO_BLK_S_IOERR; 257 goto fail; 258 } 259 260 /* 261 * The scsi inhdr is placed in the second-to-last input segment, just 262 * before the regular inhdr. 263 */ 264 scsi = (void *)elem->in_sg[elem->in_num - 2].iov_base; 265 266 if (!virtio_has_feature(blk->host_features, VIRTIO_BLK_F_SCSI)) { 267 status = VIRTIO_BLK_S_UNSUPP; 268 goto fail; 269 } 270 271 /* 272 * No support for bidirection commands yet. 273 */ 274 if (elem->out_num > 2 && elem->in_num > 3) { 275 status = VIRTIO_BLK_S_UNSUPP; 276 goto fail; 277 } 278 279 #ifdef __linux__ 280 ioctl_req = g_new0(VirtIOBlockIoctlReq, 1); 281 ioctl_req->req = req; 282 ioctl_req->hdr.interface_id = 'S'; 283 ioctl_req->hdr.cmd_len = elem->out_sg[1].iov_len; 284 ioctl_req->hdr.cmdp = elem->out_sg[1].iov_base; 285 ioctl_req->hdr.dxfer_len = 0; 286 287 if (elem->out_num > 2) { 288 /* 289 * If there are more than the minimally required 2 output segments 290 * there is write payload starting from the third iovec. 291 */ 292 ioctl_req->hdr.dxfer_direction = SG_DXFER_TO_DEV; 293 ioctl_req->hdr.iovec_count = elem->out_num - 2; 294 295 for (i = 0; i < ioctl_req->hdr.iovec_count; i++) { 296 ioctl_req->hdr.dxfer_len += elem->out_sg[i + 2].iov_len; 297 } 298 299 ioctl_req->hdr.dxferp = elem->out_sg + 2; 300 301 } else if (elem->in_num > 3) { 302 /* 303 * If we have more than 3 input segments the guest wants to actually 304 * read data. 305 */ 306 ioctl_req->hdr.dxfer_direction = SG_DXFER_FROM_DEV; 307 ioctl_req->hdr.iovec_count = elem->in_num - 3; 308 for (i = 0; i < ioctl_req->hdr.iovec_count; i++) { 309 ioctl_req->hdr.dxfer_len += elem->in_sg[i].iov_len; 310 } 311 312 ioctl_req->hdr.dxferp = elem->in_sg; 313 } else { 314 /* 315 * Some SCSI commands don't actually transfer any data. 316 */ 317 ioctl_req->hdr.dxfer_direction = SG_DXFER_NONE; 318 } 319 320 ioctl_req->hdr.sbp = elem->in_sg[elem->in_num - 3].iov_base; 321 ioctl_req->hdr.mx_sb_len = elem->in_sg[elem->in_num - 3].iov_len; 322 323 acb = blk_aio_ioctl(blk->blk, SG_IO, &ioctl_req->hdr, 324 virtio_blk_ioctl_complete, ioctl_req); 325 if (!acb) { 326 g_free(ioctl_req); 327 status = VIRTIO_BLK_S_UNSUPP; 328 goto fail; 329 } 330 return -EINPROGRESS; 331 #else 332 abort(); 333 #endif 334 335 fail: 336 /* Just put anything nonzero so that the ioctl fails in the guest. */ 337 if (scsi) { 338 virtio_stl_p(vdev, &scsi->errors, 255); 339 } 340 return status; 341 } 342 343 static void virtio_blk_handle_scsi(VirtIOBlockReq *req) 344 { 345 int status; 346 347 status = virtio_blk_handle_scsi_req(req); 348 if (status != -EINPROGRESS) { 349 virtio_blk_req_complete(req, status); 350 virtio_blk_free_request(req); 351 } 352 } 353 354 static inline void submit_requests(VirtIOBlock *s, MultiReqBuffer *mrb, 355 int start, int num_reqs, int niov) 356 { 357 BlockBackend *blk = s->blk; 358 QEMUIOVector *qiov = &mrb->reqs[start]->qiov; 359 int64_t sector_num = mrb->reqs[start]->sector_num; 360 bool is_write = mrb->is_write; 361 BdrvRequestFlags flags = 0; 362 363 if (num_reqs > 1) { 364 int i; 365 struct iovec *tmp_iov = qiov->iov; 366 int tmp_niov = qiov->niov; 367 368 /* mrb->reqs[start]->qiov was initialized from external so we can't 369 * modify it here. We need to initialize it locally and then add the 370 * external iovecs. */ 371 qemu_iovec_init(qiov, niov); 372 373 for (i = 0; i < tmp_niov; i++) { 374 qemu_iovec_add(qiov, tmp_iov[i].iov_base, tmp_iov[i].iov_len); 375 } 376 377 for (i = start + 1; i < start + num_reqs; i++) { 378 qemu_iovec_concat(qiov, &mrb->reqs[i]->qiov, 0, 379 mrb->reqs[i]->qiov.size); 380 mrb->reqs[i - 1]->mr_next = mrb->reqs[i]; 381 } 382 383 trace_virtio_blk_submit_multireq(VIRTIO_DEVICE(mrb->reqs[start]->dev), 384 mrb, start, num_reqs, 385 sector_num << BDRV_SECTOR_BITS, 386 qiov->size, is_write); 387 block_acct_merge_done(blk_get_stats(blk), 388 is_write ? BLOCK_ACCT_WRITE : BLOCK_ACCT_READ, 389 num_reqs - 1); 390 } 391 392 if (blk_ram_registrar_ok(&s->blk_ram_registrar)) { 393 flags |= BDRV_REQ_REGISTERED_BUF; 394 } 395 396 if (is_write) { 397 blk_aio_pwritev(blk, sector_num << BDRV_SECTOR_BITS, qiov, 398 flags, virtio_blk_rw_complete, 399 mrb->reqs[start]); 400 } else { 401 blk_aio_preadv(blk, sector_num << BDRV_SECTOR_BITS, qiov, 402 flags, virtio_blk_rw_complete, 403 mrb->reqs[start]); 404 } 405 } 406 407 static int multireq_compare(const void *a, const void *b) 408 { 409 const VirtIOBlockReq *req1 = *(VirtIOBlockReq **)a, 410 *req2 = *(VirtIOBlockReq **)b; 411 412 /* 413 * Note that we can't simply subtract sector_num1 from sector_num2 414 * here as that could overflow the return value. 415 */ 416 if (req1->sector_num > req2->sector_num) { 417 return 1; 418 } else if (req1->sector_num < req2->sector_num) { 419 return -1; 420 } else { 421 return 0; 422 } 423 } 424 425 static void virtio_blk_submit_multireq(VirtIOBlock *s, MultiReqBuffer *mrb) 426 { 427 int i = 0, start = 0, num_reqs = 0, niov = 0, nb_sectors = 0; 428 uint32_t max_transfer; 429 int64_t sector_num = 0; 430 431 if (mrb->num_reqs == 1) { 432 submit_requests(s, mrb, 0, 1, -1); 433 mrb->num_reqs = 0; 434 return; 435 } 436 437 max_transfer = blk_get_max_transfer(mrb->reqs[0]->dev->blk); 438 439 qsort(mrb->reqs, mrb->num_reqs, sizeof(*mrb->reqs), 440 &multireq_compare); 441 442 for (i = 0; i < mrb->num_reqs; i++) { 443 VirtIOBlockReq *req = mrb->reqs[i]; 444 if (num_reqs > 0) { 445 /* 446 * NOTE: We cannot merge the requests in below situations: 447 * 1. requests are not sequential 448 * 2. merge would exceed maximum number of IOVs 449 * 3. merge would exceed maximum transfer length of backend device 450 */ 451 if (sector_num + nb_sectors != req->sector_num || 452 niov > blk_get_max_iov(s->blk) - req->qiov.niov || 453 req->qiov.size > max_transfer || 454 nb_sectors > (max_transfer - 455 req->qiov.size) / BDRV_SECTOR_SIZE) { 456 submit_requests(s, mrb, start, num_reqs, niov); 457 num_reqs = 0; 458 } 459 } 460 461 if (num_reqs == 0) { 462 sector_num = req->sector_num; 463 nb_sectors = niov = 0; 464 start = i; 465 } 466 467 nb_sectors += req->qiov.size / BDRV_SECTOR_SIZE; 468 niov += req->qiov.niov; 469 num_reqs++; 470 } 471 472 submit_requests(s, mrb, start, num_reqs, niov); 473 mrb->num_reqs = 0; 474 } 475 476 static void virtio_blk_handle_flush(VirtIOBlockReq *req, MultiReqBuffer *mrb) 477 { 478 VirtIOBlock *s = req->dev; 479 480 block_acct_start(blk_get_stats(s->blk), &req->acct, 0, 481 BLOCK_ACCT_FLUSH); 482 483 /* 484 * Make sure all outstanding writes are posted to the backing device. 485 */ 486 if (mrb->is_write && mrb->num_reqs > 0) { 487 virtio_blk_submit_multireq(s, mrb); 488 } 489 blk_aio_flush(s->blk, virtio_blk_flush_complete, req); 490 } 491 492 static bool virtio_blk_sect_range_ok(VirtIOBlock *dev, 493 uint64_t sector, size_t size) 494 { 495 uint64_t nb_sectors = size >> BDRV_SECTOR_BITS; 496 uint64_t total_sectors; 497 498 if (nb_sectors > BDRV_REQUEST_MAX_SECTORS) { 499 return false; 500 } 501 if (sector & dev->sector_mask) { 502 return false; 503 } 504 if (size % dev->conf.conf.logical_block_size) { 505 return false; 506 } 507 blk_get_geometry(dev->blk, &total_sectors); 508 if (sector > total_sectors || nb_sectors > total_sectors - sector) { 509 return false; 510 } 511 return true; 512 } 513 514 static uint8_t virtio_blk_handle_discard_write_zeroes(VirtIOBlockReq *req, 515 struct virtio_blk_discard_write_zeroes *dwz_hdr, bool is_write_zeroes) 516 { 517 VirtIOBlock *s = req->dev; 518 VirtIODevice *vdev = VIRTIO_DEVICE(s); 519 uint64_t sector; 520 uint32_t num_sectors, flags, max_sectors; 521 uint8_t err_status; 522 int bytes; 523 524 sector = virtio_ldq_p(vdev, &dwz_hdr->sector); 525 num_sectors = virtio_ldl_p(vdev, &dwz_hdr->num_sectors); 526 flags = virtio_ldl_p(vdev, &dwz_hdr->flags); 527 max_sectors = is_write_zeroes ? s->conf.max_write_zeroes_sectors : 528 s->conf.max_discard_sectors; 529 530 /* 531 * max_sectors is at most BDRV_REQUEST_MAX_SECTORS, this check 532 * make us sure that "num_sectors << BDRV_SECTOR_BITS" can fit in 533 * the integer variable. 534 */ 535 if (unlikely(num_sectors > max_sectors)) { 536 err_status = VIRTIO_BLK_S_IOERR; 537 goto err; 538 } 539 540 bytes = num_sectors << BDRV_SECTOR_BITS; 541 542 if (unlikely(!virtio_blk_sect_range_ok(s, sector, bytes))) { 543 err_status = VIRTIO_BLK_S_IOERR; 544 goto err; 545 } 546 547 /* 548 * The device MUST set the status byte to VIRTIO_BLK_S_UNSUPP for discard 549 * and write zeroes commands if any unknown flag is set. 550 */ 551 if (unlikely(flags & ~VIRTIO_BLK_WRITE_ZEROES_FLAG_UNMAP)) { 552 err_status = VIRTIO_BLK_S_UNSUPP; 553 goto err; 554 } 555 556 if (is_write_zeroes) { /* VIRTIO_BLK_T_WRITE_ZEROES */ 557 int blk_aio_flags = 0; 558 559 if (flags & VIRTIO_BLK_WRITE_ZEROES_FLAG_UNMAP) { 560 blk_aio_flags |= BDRV_REQ_MAY_UNMAP; 561 } 562 563 block_acct_start(blk_get_stats(s->blk), &req->acct, bytes, 564 BLOCK_ACCT_WRITE); 565 566 blk_aio_pwrite_zeroes(s->blk, sector << BDRV_SECTOR_BITS, 567 bytes, blk_aio_flags, 568 virtio_blk_discard_write_zeroes_complete, req); 569 } else { /* VIRTIO_BLK_T_DISCARD */ 570 /* 571 * The device MUST set the status byte to VIRTIO_BLK_S_UNSUPP for 572 * discard commands if the unmap flag is set. 573 */ 574 if (unlikely(flags & VIRTIO_BLK_WRITE_ZEROES_FLAG_UNMAP)) { 575 err_status = VIRTIO_BLK_S_UNSUPP; 576 goto err; 577 } 578 579 blk_aio_pdiscard(s->blk, sector << BDRV_SECTOR_BITS, bytes, 580 virtio_blk_discard_write_zeroes_complete, req); 581 } 582 583 return VIRTIO_BLK_S_OK; 584 585 err: 586 if (is_write_zeroes) { 587 block_acct_invalid(blk_get_stats(s->blk), BLOCK_ACCT_WRITE); 588 } 589 return err_status; 590 } 591 592 typedef struct ZoneCmdData { 593 VirtIOBlockReq *req; 594 struct iovec *in_iov; 595 unsigned in_num; 596 union { 597 struct { 598 unsigned int nr_zones; 599 BlockZoneDescriptor *zones; 600 } zone_report_data; 601 struct { 602 int64_t offset; 603 } zone_append_data; 604 }; 605 } ZoneCmdData; 606 607 /* 608 * check zoned_request: error checking before issuing requests. If all checks 609 * passed, return true. 610 * append: true if only zone append requests issued. 611 */ 612 static bool check_zoned_request(VirtIOBlock *s, int64_t offset, int64_t len, 613 bool append, uint8_t *status) { 614 BlockDriverState *bs = blk_bs(s->blk); 615 int index; 616 617 if (!virtio_has_feature(s->host_features, VIRTIO_BLK_F_ZONED)) { 618 *status = VIRTIO_BLK_S_UNSUPP; 619 return false; 620 } 621 622 if (offset < 0 || len < 0 || len > (bs->total_sectors << BDRV_SECTOR_BITS) 623 || offset > (bs->total_sectors << BDRV_SECTOR_BITS) - len) { 624 *status = VIRTIO_BLK_S_ZONE_INVALID_CMD; 625 return false; 626 } 627 628 if (append) { 629 if (bs->bl.write_granularity) { 630 if ((offset % bs->bl.write_granularity) != 0) { 631 *status = VIRTIO_BLK_S_ZONE_UNALIGNED_WP; 632 return false; 633 } 634 } 635 636 index = offset / bs->bl.zone_size; 637 if (BDRV_ZT_IS_CONV(bs->wps->wp[index])) { 638 *status = VIRTIO_BLK_S_ZONE_INVALID_CMD; 639 return false; 640 } 641 642 if (len / 512 > bs->bl.max_append_sectors) { 643 if (bs->bl.max_append_sectors == 0) { 644 *status = VIRTIO_BLK_S_UNSUPP; 645 } else { 646 *status = VIRTIO_BLK_S_ZONE_INVALID_CMD; 647 } 648 return false; 649 } 650 } 651 return true; 652 } 653 654 static void virtio_blk_zone_report_complete(void *opaque, int ret) 655 { 656 ZoneCmdData *data = opaque; 657 VirtIOBlockReq *req = data->req; 658 VirtIODevice *vdev = VIRTIO_DEVICE(req->dev); 659 struct iovec *in_iov = data->in_iov; 660 unsigned in_num = data->in_num; 661 int64_t zrp_size, n, j = 0; 662 int64_t nz = data->zone_report_data.nr_zones; 663 int8_t err_status = VIRTIO_BLK_S_OK; 664 struct virtio_blk_zone_report zrp_hdr = (struct virtio_blk_zone_report) { 665 .nr_zones = cpu_to_le64(nz), 666 }; 667 668 trace_virtio_blk_zone_report_complete(vdev, req, nz, ret); 669 if (ret) { 670 err_status = VIRTIO_BLK_S_ZONE_INVALID_CMD; 671 goto out; 672 } 673 674 zrp_size = sizeof(struct virtio_blk_zone_report) 675 + sizeof(struct virtio_blk_zone_descriptor) * nz; 676 n = iov_from_buf(in_iov, in_num, 0, &zrp_hdr, sizeof(zrp_hdr)); 677 if (n != sizeof(zrp_hdr)) { 678 virtio_error(vdev, "Driver provided input buffer that is too small!"); 679 err_status = VIRTIO_BLK_S_ZONE_INVALID_CMD; 680 goto out; 681 } 682 683 for (size_t i = sizeof(zrp_hdr); i < zrp_size; 684 i += sizeof(struct virtio_blk_zone_descriptor), ++j) { 685 struct virtio_blk_zone_descriptor desc = 686 (struct virtio_blk_zone_descriptor) { 687 .z_start = cpu_to_le64(data->zone_report_data.zones[j].start 688 >> BDRV_SECTOR_BITS), 689 .z_cap = cpu_to_le64(data->zone_report_data.zones[j].cap 690 >> BDRV_SECTOR_BITS), 691 .z_wp = cpu_to_le64(data->zone_report_data.zones[j].wp 692 >> BDRV_SECTOR_BITS), 693 }; 694 695 switch (data->zone_report_data.zones[j].type) { 696 case BLK_ZT_CONV: 697 desc.z_type = VIRTIO_BLK_ZT_CONV; 698 break; 699 case BLK_ZT_SWR: 700 desc.z_type = VIRTIO_BLK_ZT_SWR; 701 break; 702 case BLK_ZT_SWP: 703 desc.z_type = VIRTIO_BLK_ZT_SWP; 704 break; 705 default: 706 g_assert_not_reached(); 707 } 708 709 switch (data->zone_report_data.zones[j].state) { 710 case BLK_ZS_RDONLY: 711 desc.z_state = VIRTIO_BLK_ZS_RDONLY; 712 break; 713 case BLK_ZS_OFFLINE: 714 desc.z_state = VIRTIO_BLK_ZS_OFFLINE; 715 break; 716 case BLK_ZS_EMPTY: 717 desc.z_state = VIRTIO_BLK_ZS_EMPTY; 718 break; 719 case BLK_ZS_CLOSED: 720 desc.z_state = VIRTIO_BLK_ZS_CLOSED; 721 break; 722 case BLK_ZS_FULL: 723 desc.z_state = VIRTIO_BLK_ZS_FULL; 724 break; 725 case BLK_ZS_EOPEN: 726 desc.z_state = VIRTIO_BLK_ZS_EOPEN; 727 break; 728 case BLK_ZS_IOPEN: 729 desc.z_state = VIRTIO_BLK_ZS_IOPEN; 730 break; 731 case BLK_ZS_NOT_WP: 732 desc.z_state = VIRTIO_BLK_ZS_NOT_WP; 733 break; 734 default: 735 g_assert_not_reached(); 736 } 737 738 /* TODO: it takes O(n^2) time complexity. Optimizations required. */ 739 n = iov_from_buf(in_iov, in_num, i, &desc, sizeof(desc)); 740 if (n != sizeof(desc)) { 741 virtio_error(vdev, "Driver provided input buffer " 742 "for descriptors that is too small!"); 743 err_status = VIRTIO_BLK_S_ZONE_INVALID_CMD; 744 } 745 } 746 747 out: 748 virtio_blk_req_complete(req, err_status); 749 virtio_blk_free_request(req); 750 g_free(data->zone_report_data.zones); 751 g_free(data); 752 } 753 754 static void virtio_blk_handle_zone_report(VirtIOBlockReq *req, 755 struct iovec *in_iov, 756 unsigned in_num) 757 { 758 VirtIOBlock *s = req->dev; 759 VirtIODevice *vdev = VIRTIO_DEVICE(s); 760 unsigned int nr_zones; 761 ZoneCmdData *data; 762 int64_t zone_size, offset; 763 uint8_t err_status; 764 765 if (req->in_len < sizeof(struct virtio_blk_inhdr) + 766 sizeof(struct virtio_blk_zone_report) + 767 sizeof(struct virtio_blk_zone_descriptor)) { 768 virtio_error(vdev, "in buffer too small for zone report"); 769 return; 770 } 771 772 /* start byte offset of the zone report */ 773 offset = virtio_ldq_p(vdev, &req->out.sector) << BDRV_SECTOR_BITS; 774 if (!check_zoned_request(s, offset, 0, false, &err_status)) { 775 goto out; 776 } 777 nr_zones = (req->in_len - sizeof(struct virtio_blk_inhdr) - 778 sizeof(struct virtio_blk_zone_report)) / 779 sizeof(struct virtio_blk_zone_descriptor); 780 trace_virtio_blk_handle_zone_report(vdev, req, 781 offset >> BDRV_SECTOR_BITS, nr_zones); 782 783 zone_size = sizeof(BlockZoneDescriptor) * nr_zones; 784 data = g_malloc(sizeof(ZoneCmdData)); 785 data->req = req; 786 data->in_iov = in_iov; 787 data->in_num = in_num; 788 data->zone_report_data.nr_zones = nr_zones; 789 data->zone_report_data.zones = g_malloc(zone_size), 790 791 blk_aio_zone_report(s->blk, offset, &data->zone_report_data.nr_zones, 792 data->zone_report_data.zones, 793 virtio_blk_zone_report_complete, data); 794 return; 795 out: 796 virtio_blk_req_complete(req, err_status); 797 virtio_blk_free_request(req); 798 } 799 800 static void virtio_blk_zone_mgmt_complete(void *opaque, int ret) 801 { 802 VirtIOBlockReq *req = opaque; 803 VirtIOBlock *s = req->dev; 804 VirtIODevice *vdev = VIRTIO_DEVICE(s); 805 int8_t err_status = VIRTIO_BLK_S_OK; 806 trace_virtio_blk_zone_mgmt_complete(vdev, req,ret); 807 808 if (ret) { 809 err_status = VIRTIO_BLK_S_ZONE_INVALID_CMD; 810 } 811 812 virtio_blk_req_complete(req, err_status); 813 virtio_blk_free_request(req); 814 } 815 816 static int virtio_blk_handle_zone_mgmt(VirtIOBlockReq *req, BlockZoneOp op) 817 { 818 VirtIOBlock *s = req->dev; 819 VirtIODevice *vdev = VIRTIO_DEVICE(s); 820 BlockDriverState *bs = blk_bs(s->blk); 821 int64_t offset = virtio_ldq_p(vdev, &req->out.sector) << BDRV_SECTOR_BITS; 822 uint64_t len; 823 uint64_t capacity = bs->total_sectors << BDRV_SECTOR_BITS; 824 uint8_t err_status = VIRTIO_BLK_S_OK; 825 826 uint32_t type = virtio_ldl_p(vdev, &req->out.type); 827 if (type == VIRTIO_BLK_T_ZONE_RESET_ALL) { 828 /* Entire drive capacity */ 829 offset = 0; 830 len = capacity; 831 trace_virtio_blk_handle_zone_reset_all(vdev, req, 0, 832 bs->total_sectors); 833 } else { 834 if (bs->bl.zone_size > capacity - offset) { 835 /* The zoned device allows the last smaller zone. */ 836 len = capacity - bs->bl.zone_size * (bs->bl.nr_zones - 1); 837 } else { 838 len = bs->bl.zone_size; 839 } 840 trace_virtio_blk_handle_zone_mgmt(vdev, req, op, 841 offset >> BDRV_SECTOR_BITS, 842 len >> BDRV_SECTOR_BITS); 843 } 844 845 if (!check_zoned_request(s, offset, len, false, &err_status)) { 846 goto out; 847 } 848 849 blk_aio_zone_mgmt(s->blk, op, offset, len, 850 virtio_blk_zone_mgmt_complete, req); 851 852 return 0; 853 out: 854 virtio_blk_req_complete(req, err_status); 855 virtio_blk_free_request(req); 856 return err_status; 857 } 858 859 static void virtio_blk_zone_append_complete(void *opaque, int ret) 860 { 861 ZoneCmdData *data = opaque; 862 VirtIOBlockReq *req = data->req; 863 VirtIODevice *vdev = VIRTIO_DEVICE(req->dev); 864 int64_t append_sector, n; 865 uint8_t err_status = VIRTIO_BLK_S_OK; 866 867 if (ret) { 868 err_status = VIRTIO_BLK_S_ZONE_INVALID_CMD; 869 goto out; 870 } 871 872 virtio_stq_p(vdev, &append_sector, 873 data->zone_append_data.offset >> BDRV_SECTOR_BITS); 874 n = iov_from_buf(data->in_iov, data->in_num, 0, &append_sector, 875 sizeof(append_sector)); 876 if (n != sizeof(append_sector)) { 877 virtio_error(vdev, "Driver provided input buffer less than size of " 878 "append_sector"); 879 err_status = VIRTIO_BLK_S_ZONE_INVALID_CMD; 880 goto out; 881 } 882 trace_virtio_blk_zone_append_complete(vdev, req, append_sector, ret); 883 884 out: 885 virtio_blk_req_complete(req, err_status); 886 virtio_blk_free_request(req); 887 g_free(data); 888 } 889 890 static int virtio_blk_handle_zone_append(VirtIOBlockReq *req, 891 struct iovec *out_iov, 892 struct iovec *in_iov, 893 uint64_t out_num, 894 unsigned in_num) { 895 VirtIOBlock *s = req->dev; 896 VirtIODevice *vdev = VIRTIO_DEVICE(s); 897 uint8_t err_status = VIRTIO_BLK_S_OK; 898 899 int64_t offset = virtio_ldq_p(vdev, &req->out.sector) << BDRV_SECTOR_BITS; 900 int64_t len = iov_size(out_iov, out_num); 901 ZoneCmdData *data; 902 903 trace_virtio_blk_handle_zone_append(vdev, req, offset >> BDRV_SECTOR_BITS); 904 if (!check_zoned_request(s, offset, len, true, &err_status)) { 905 goto out; 906 } 907 908 data = g_malloc(sizeof(ZoneCmdData)); 909 data->req = req; 910 data->in_iov = in_iov; 911 data->in_num = in_num; 912 data->zone_append_data.offset = offset; 913 qemu_iovec_init_external(&req->qiov, out_iov, out_num); 914 915 block_acct_start(blk_get_stats(s->blk), &req->acct, len, 916 BLOCK_ACCT_ZONE_APPEND); 917 918 blk_aio_zone_append(s->blk, &data->zone_append_data.offset, &req->qiov, 0, 919 virtio_blk_zone_append_complete, data); 920 return 0; 921 922 out: 923 virtio_blk_req_complete(req, err_status); 924 virtio_blk_free_request(req); 925 return err_status; 926 } 927 928 static int virtio_blk_handle_request(VirtIOBlockReq *req, MultiReqBuffer *mrb) 929 { 930 uint32_t type; 931 struct iovec *in_iov = req->elem.in_sg; 932 struct iovec *out_iov = req->elem.out_sg; 933 unsigned in_num = req->elem.in_num; 934 unsigned out_num = req->elem.out_num; 935 VirtIOBlock *s = req->dev; 936 VirtIODevice *vdev = VIRTIO_DEVICE(s); 937 938 if (req->elem.out_num < 1 || req->elem.in_num < 1) { 939 virtio_error(vdev, "virtio-blk missing headers"); 940 return -1; 941 } 942 943 if (unlikely(iov_to_buf(out_iov, out_num, 0, &req->out, 944 sizeof(req->out)) != sizeof(req->out))) { 945 virtio_error(vdev, "virtio-blk request outhdr too short"); 946 return -1; 947 } 948 949 iov_discard_front_undoable(&out_iov, &out_num, sizeof(req->out), 950 &req->outhdr_undo); 951 952 if (in_iov[in_num - 1].iov_len < sizeof(struct virtio_blk_inhdr)) { 953 virtio_error(vdev, "virtio-blk request inhdr too short"); 954 iov_discard_undo(&req->outhdr_undo); 955 return -1; 956 } 957 958 /* We always touch the last byte, so just see how big in_iov is. */ 959 req->in_len = iov_size(in_iov, in_num); 960 req->in = (void *)in_iov[in_num - 1].iov_base 961 + in_iov[in_num - 1].iov_len 962 - sizeof(struct virtio_blk_inhdr); 963 iov_discard_back_undoable(in_iov, &in_num, sizeof(struct virtio_blk_inhdr), 964 &req->inhdr_undo); 965 966 type = virtio_ldl_p(vdev, &req->out.type); 967 968 /* VIRTIO_BLK_T_OUT defines the command direction. VIRTIO_BLK_T_BARRIER 969 * is an optional flag. Although a guest should not send this flag if 970 * not negotiated we ignored it in the past. So keep ignoring it. */ 971 switch (type & ~(VIRTIO_BLK_T_OUT | VIRTIO_BLK_T_BARRIER)) { 972 case VIRTIO_BLK_T_IN: 973 { 974 bool is_write = type & VIRTIO_BLK_T_OUT; 975 req->sector_num = virtio_ldq_p(vdev, &req->out.sector); 976 977 if (is_write) { 978 qemu_iovec_init_external(&req->qiov, out_iov, out_num); 979 trace_virtio_blk_handle_write(vdev, req, req->sector_num, 980 req->qiov.size / BDRV_SECTOR_SIZE); 981 } else { 982 qemu_iovec_init_external(&req->qiov, in_iov, in_num); 983 trace_virtio_blk_handle_read(vdev, req, req->sector_num, 984 req->qiov.size / BDRV_SECTOR_SIZE); 985 } 986 987 if (!virtio_blk_sect_range_ok(s, req->sector_num, req->qiov.size)) { 988 virtio_blk_req_complete(req, VIRTIO_BLK_S_IOERR); 989 block_acct_invalid(blk_get_stats(s->blk), 990 is_write ? BLOCK_ACCT_WRITE : BLOCK_ACCT_READ); 991 virtio_blk_free_request(req); 992 return 0; 993 } 994 995 block_acct_start(blk_get_stats(s->blk), &req->acct, req->qiov.size, 996 is_write ? BLOCK_ACCT_WRITE : BLOCK_ACCT_READ); 997 998 /* merge would exceed maximum number of requests or IO direction 999 * changes */ 1000 if (mrb->num_reqs > 0 && (mrb->num_reqs == VIRTIO_BLK_MAX_MERGE_REQS || 1001 is_write != mrb->is_write || 1002 !s->conf.request_merging)) { 1003 virtio_blk_submit_multireq(s, mrb); 1004 } 1005 1006 assert(mrb->num_reqs < VIRTIO_BLK_MAX_MERGE_REQS); 1007 mrb->reqs[mrb->num_reqs++] = req; 1008 mrb->is_write = is_write; 1009 break; 1010 } 1011 case VIRTIO_BLK_T_FLUSH: 1012 virtio_blk_handle_flush(req, mrb); 1013 break; 1014 case VIRTIO_BLK_T_ZONE_REPORT: 1015 virtio_blk_handle_zone_report(req, in_iov, in_num); 1016 break; 1017 case VIRTIO_BLK_T_ZONE_OPEN: 1018 virtio_blk_handle_zone_mgmt(req, BLK_ZO_OPEN); 1019 break; 1020 case VIRTIO_BLK_T_ZONE_CLOSE: 1021 virtio_blk_handle_zone_mgmt(req, BLK_ZO_CLOSE); 1022 break; 1023 case VIRTIO_BLK_T_ZONE_FINISH: 1024 virtio_blk_handle_zone_mgmt(req, BLK_ZO_FINISH); 1025 break; 1026 case VIRTIO_BLK_T_ZONE_RESET: 1027 virtio_blk_handle_zone_mgmt(req, BLK_ZO_RESET); 1028 break; 1029 case VIRTIO_BLK_T_ZONE_RESET_ALL: 1030 virtio_blk_handle_zone_mgmt(req, BLK_ZO_RESET); 1031 break; 1032 case VIRTIO_BLK_T_SCSI_CMD: 1033 virtio_blk_handle_scsi(req); 1034 break; 1035 case VIRTIO_BLK_T_GET_ID: 1036 { 1037 /* 1038 * NB: per existing s/n string convention the string is 1039 * terminated by '\0' only when shorter than buffer. 1040 */ 1041 const char *serial = s->conf.serial ? s->conf.serial : ""; 1042 size_t size = MIN(strlen(serial) + 1, 1043 MIN(iov_size(in_iov, in_num), 1044 VIRTIO_BLK_ID_BYTES)); 1045 iov_from_buf(in_iov, in_num, 0, serial, size); 1046 virtio_blk_req_complete(req, VIRTIO_BLK_S_OK); 1047 virtio_blk_free_request(req); 1048 break; 1049 } 1050 case VIRTIO_BLK_T_ZONE_APPEND & ~VIRTIO_BLK_T_OUT: 1051 /* 1052 * Passing out_iov/out_num and in_iov/in_num is not safe 1053 * to access req->elem.out_sg directly because it may be 1054 * modified by virtio_blk_handle_request(). 1055 */ 1056 virtio_blk_handle_zone_append(req, out_iov, in_iov, out_num, in_num); 1057 break; 1058 /* 1059 * VIRTIO_BLK_T_DISCARD and VIRTIO_BLK_T_WRITE_ZEROES are defined with 1060 * VIRTIO_BLK_T_OUT flag set. We masked this flag in the switch statement, 1061 * so we must mask it for these requests, then we will check if it is set. 1062 */ 1063 case VIRTIO_BLK_T_DISCARD & ~VIRTIO_BLK_T_OUT: 1064 case VIRTIO_BLK_T_WRITE_ZEROES & ~VIRTIO_BLK_T_OUT: 1065 { 1066 struct virtio_blk_discard_write_zeroes dwz_hdr; 1067 size_t out_len = iov_size(out_iov, out_num); 1068 bool is_write_zeroes = (type & ~VIRTIO_BLK_T_BARRIER) == 1069 VIRTIO_BLK_T_WRITE_ZEROES; 1070 uint8_t err_status; 1071 1072 /* 1073 * Unsupported if VIRTIO_BLK_T_OUT is not set or the request contains 1074 * more than one segment. 1075 */ 1076 if (unlikely(!(type & VIRTIO_BLK_T_OUT) || 1077 out_len > sizeof(dwz_hdr))) { 1078 virtio_blk_req_complete(req, VIRTIO_BLK_S_UNSUPP); 1079 virtio_blk_free_request(req); 1080 return 0; 1081 } 1082 1083 if (unlikely(iov_to_buf(out_iov, out_num, 0, &dwz_hdr, 1084 sizeof(dwz_hdr)) != sizeof(dwz_hdr))) { 1085 iov_discard_undo(&req->inhdr_undo); 1086 iov_discard_undo(&req->outhdr_undo); 1087 virtio_error(vdev, "virtio-blk discard/write_zeroes header" 1088 " too short"); 1089 return -1; 1090 } 1091 1092 err_status = virtio_blk_handle_discard_write_zeroes(req, &dwz_hdr, 1093 is_write_zeroes); 1094 if (err_status != VIRTIO_BLK_S_OK) { 1095 virtio_blk_req_complete(req, err_status); 1096 virtio_blk_free_request(req); 1097 } 1098 1099 break; 1100 } 1101 default: 1102 virtio_blk_req_complete(req, VIRTIO_BLK_S_UNSUPP); 1103 virtio_blk_free_request(req); 1104 } 1105 return 0; 1106 } 1107 1108 void virtio_blk_handle_vq(VirtIOBlock *s, VirtQueue *vq) 1109 { 1110 VirtIOBlockReq *req; 1111 MultiReqBuffer mrb = {}; 1112 bool suppress_notifications = virtio_queue_get_notification(vq); 1113 1114 defer_call_begin(); 1115 1116 do { 1117 if (suppress_notifications) { 1118 virtio_queue_set_notification(vq, 0); 1119 } 1120 1121 while ((req = virtio_blk_get_request(s, vq))) { 1122 if (virtio_blk_handle_request(req, &mrb)) { 1123 virtqueue_detach_element(req->vq, &req->elem, 0); 1124 virtio_blk_free_request(req); 1125 break; 1126 } 1127 } 1128 1129 if (suppress_notifications) { 1130 virtio_queue_set_notification(vq, 1); 1131 } 1132 } while (!virtio_queue_empty(vq)); 1133 1134 if (mrb.num_reqs) { 1135 virtio_blk_submit_multireq(s, &mrb); 1136 } 1137 1138 defer_call_end(); 1139 } 1140 1141 static void virtio_blk_handle_output(VirtIODevice *vdev, VirtQueue *vq) 1142 { 1143 VirtIOBlock *s = (VirtIOBlock *)vdev; 1144 1145 if (!s->ioeventfd_disabled && !s->ioeventfd_started) { 1146 /* Some guests kick before setting VIRTIO_CONFIG_S_DRIVER_OK so start 1147 * ioeventfd here instead of waiting for .set_status(). 1148 */ 1149 virtio_device_start_ioeventfd(vdev); 1150 if (!s->ioeventfd_disabled) { 1151 return; 1152 } 1153 } 1154 1155 virtio_blk_handle_vq(s, vq); 1156 } 1157 1158 static void virtio_blk_dma_restart_bh(void *opaque) 1159 { 1160 VirtIOBlockReq *req = opaque; 1161 VirtIOBlock *s = req->dev; /* we're called with at least one request */ 1162 1163 MultiReqBuffer mrb = {}; 1164 1165 while (req) { 1166 VirtIOBlockReq *next = req->next; 1167 if (virtio_blk_handle_request(req, &mrb)) { 1168 /* Device is now broken and won't do any processing until it gets 1169 * reset. Already queued requests will be lost: let's purge them. 1170 */ 1171 while (req) { 1172 next = req->next; 1173 virtqueue_detach_element(req->vq, &req->elem, 0); 1174 virtio_blk_free_request(req); 1175 req = next; 1176 } 1177 break; 1178 } 1179 req = next; 1180 } 1181 1182 if (mrb.num_reqs) { 1183 virtio_blk_submit_multireq(s, &mrb); 1184 } 1185 1186 /* Paired with inc in virtio_blk_dma_restart_cb() */ 1187 blk_dec_in_flight(s->conf.conf.blk); 1188 } 1189 1190 static void virtio_blk_dma_restart_cb(void *opaque, bool running, 1191 RunState state) 1192 { 1193 VirtIOBlock *s = opaque; 1194 uint16_t num_queues = s->conf.num_queues; 1195 g_autofree VirtIOBlockReq **vq_rq = NULL; 1196 VirtIOBlockReq *rq; 1197 1198 if (!running) { 1199 return; 1200 } 1201 1202 /* Split the device-wide s->rq request list into per-vq request lists */ 1203 vq_rq = g_new0(VirtIOBlockReq *, num_queues); 1204 1205 WITH_QEMU_LOCK_GUARD(&s->rq_lock) { 1206 rq = s->rq; 1207 s->rq = NULL; 1208 } 1209 1210 while (rq) { 1211 VirtIOBlockReq *next = rq->next; 1212 uint16_t idx = virtio_get_queue_index(rq->vq); 1213 1214 /* Only num_queues vqs were created so vq_rq[idx] is within bounds */ 1215 assert(idx < num_queues); 1216 rq->next = vq_rq[idx]; 1217 vq_rq[idx] = rq; 1218 rq = next; 1219 } 1220 1221 /* Schedule a BH to submit the requests in each vq's AioContext */ 1222 for (uint16_t i = 0; i < num_queues; i++) { 1223 if (!vq_rq[i]) { 1224 continue; 1225 } 1226 1227 /* Paired with dec in virtio_blk_dma_restart_bh() */ 1228 blk_inc_in_flight(s->conf.conf.blk); 1229 1230 aio_bh_schedule_oneshot(s->vq_aio_context[i], 1231 virtio_blk_dma_restart_bh, 1232 vq_rq[i]); 1233 } 1234 } 1235 1236 static void virtio_blk_reset(VirtIODevice *vdev) 1237 { 1238 VirtIOBlock *s = VIRTIO_BLK(vdev); 1239 VirtIOBlockReq *req; 1240 1241 /* Dataplane has stopped... */ 1242 assert(!s->ioeventfd_started); 1243 1244 /* ...but requests may still be in flight. */ 1245 blk_drain(s->blk); 1246 1247 /* We drop queued requests after blk_drain() because blk_drain() itself can 1248 * produce them. */ 1249 WITH_QEMU_LOCK_GUARD(&s->rq_lock) { 1250 while (s->rq) { 1251 req = s->rq; 1252 s->rq = req->next; 1253 1254 /* No other threads can access req->vq here */ 1255 virtqueue_detach_element(req->vq, &req->elem, 0); 1256 1257 virtio_blk_free_request(req); 1258 } 1259 } 1260 1261 blk_set_enable_write_cache(s->blk, s->original_wce); 1262 } 1263 1264 /* coalesce internal state, copy to pci i/o region 0 1265 */ 1266 static void virtio_blk_update_config(VirtIODevice *vdev, uint8_t *config) 1267 { 1268 VirtIOBlock *s = VIRTIO_BLK(vdev); 1269 BlockConf *conf = &s->conf.conf; 1270 BlockDriverState *bs = blk_bs(s->blk); 1271 struct virtio_blk_config blkcfg; 1272 uint64_t capacity; 1273 int64_t length; 1274 int blk_size = conf->logical_block_size; 1275 1276 blk_get_geometry(s->blk, &capacity); 1277 memset(&blkcfg, 0, sizeof(blkcfg)); 1278 virtio_stq_p(vdev, &blkcfg.capacity, capacity); 1279 virtio_stl_p(vdev, &blkcfg.seg_max, 1280 s->conf.seg_max_adjust ? s->conf.queue_size - 2 : 128 - 2); 1281 virtio_stw_p(vdev, &blkcfg.geometry.cylinders, conf->cyls); 1282 virtio_stl_p(vdev, &blkcfg.blk_size, blk_size); 1283 virtio_stw_p(vdev, &blkcfg.min_io_size, conf->min_io_size / blk_size); 1284 virtio_stl_p(vdev, &blkcfg.opt_io_size, conf->opt_io_size / blk_size); 1285 blkcfg.geometry.heads = conf->heads; 1286 /* 1287 * We must ensure that the block device capacity is a multiple of 1288 * the logical block size. If that is not the case, let's use 1289 * sector_mask to adopt the geometry to have a correct picture. 1290 * For those devices where the capacity is ok for the given geometry 1291 * we don't touch the sector value of the geometry, since some devices 1292 * (like s390 dasd) need a specific value. Here the capacity is already 1293 * cyls*heads*secs*blk_size and the sector value is not block size 1294 * divided by 512 - instead it is the amount of blk_size blocks 1295 * per track (cylinder). 1296 */ 1297 length = blk_getlength(s->blk); 1298 if (length > 0 && length / conf->heads / conf->secs % blk_size) { 1299 blkcfg.geometry.sectors = conf->secs & ~s->sector_mask; 1300 } else { 1301 blkcfg.geometry.sectors = conf->secs; 1302 } 1303 blkcfg.size_max = 0; 1304 blkcfg.physical_block_exp = get_physical_block_exp(conf); 1305 blkcfg.alignment_offset = 0; 1306 blkcfg.wce = blk_enable_write_cache(s->blk); 1307 virtio_stw_p(vdev, &blkcfg.num_queues, s->conf.num_queues); 1308 if (virtio_has_feature(s->host_features, VIRTIO_BLK_F_DISCARD)) { 1309 uint32_t discard_granularity = conf->discard_granularity; 1310 if (discard_granularity == -1 || !s->conf.report_discard_granularity) { 1311 discard_granularity = blk_size; 1312 } 1313 virtio_stl_p(vdev, &blkcfg.max_discard_sectors, 1314 s->conf.max_discard_sectors); 1315 virtio_stl_p(vdev, &blkcfg.discard_sector_alignment, 1316 discard_granularity >> BDRV_SECTOR_BITS); 1317 /* 1318 * We support only one segment per request since multiple segments 1319 * are not widely used and there are no userspace APIs that allow 1320 * applications to submit multiple segments in a single call. 1321 */ 1322 virtio_stl_p(vdev, &blkcfg.max_discard_seg, 1); 1323 } 1324 if (virtio_has_feature(s->host_features, VIRTIO_BLK_F_WRITE_ZEROES)) { 1325 virtio_stl_p(vdev, &blkcfg.max_write_zeroes_sectors, 1326 s->conf.max_write_zeroes_sectors); 1327 blkcfg.write_zeroes_may_unmap = 1; 1328 virtio_stl_p(vdev, &blkcfg.max_write_zeroes_seg, 1); 1329 } 1330 if (bs->bl.zoned != BLK_Z_NONE) { 1331 switch (bs->bl.zoned) { 1332 case BLK_Z_HM: 1333 blkcfg.zoned.model = VIRTIO_BLK_Z_HM; 1334 break; 1335 case BLK_Z_HA: 1336 blkcfg.zoned.model = VIRTIO_BLK_Z_HA; 1337 break; 1338 default: 1339 g_assert_not_reached(); 1340 } 1341 1342 virtio_stl_p(vdev, &blkcfg.zoned.zone_sectors, 1343 bs->bl.zone_size / 512); 1344 virtio_stl_p(vdev, &blkcfg.zoned.max_active_zones, 1345 bs->bl.max_active_zones); 1346 virtio_stl_p(vdev, &blkcfg.zoned.max_open_zones, 1347 bs->bl.max_open_zones); 1348 virtio_stl_p(vdev, &blkcfg.zoned.write_granularity, blk_size); 1349 virtio_stl_p(vdev, &blkcfg.zoned.max_append_sectors, 1350 bs->bl.max_append_sectors); 1351 } else { 1352 blkcfg.zoned.model = VIRTIO_BLK_Z_NONE; 1353 } 1354 memcpy(config, &blkcfg, s->config_size); 1355 } 1356 1357 static void virtio_blk_set_config(VirtIODevice *vdev, const uint8_t *config) 1358 { 1359 VirtIOBlock *s = VIRTIO_BLK(vdev); 1360 struct virtio_blk_config blkcfg; 1361 1362 memcpy(&blkcfg, config, s->config_size); 1363 1364 blk_set_enable_write_cache(s->blk, blkcfg.wce != 0); 1365 } 1366 1367 static uint64_t virtio_blk_get_features(VirtIODevice *vdev, uint64_t features, 1368 Error **errp) 1369 { 1370 VirtIOBlock *s = VIRTIO_BLK(vdev); 1371 1372 /* Firstly sync all virtio-blk possible supported features */ 1373 features |= s->host_features; 1374 1375 virtio_add_feature(&features, VIRTIO_BLK_F_SEG_MAX); 1376 virtio_add_feature(&features, VIRTIO_BLK_F_GEOMETRY); 1377 virtio_add_feature(&features, VIRTIO_BLK_F_TOPOLOGY); 1378 virtio_add_feature(&features, VIRTIO_BLK_F_BLK_SIZE); 1379 if (virtio_has_feature(features, VIRTIO_F_VERSION_1)) { 1380 if (virtio_has_feature(s->host_features, VIRTIO_BLK_F_SCSI)) { 1381 error_setg(errp, "Please set scsi=off for virtio-blk devices in order to use virtio 1.0"); 1382 return 0; 1383 } 1384 } else { 1385 virtio_clear_feature(&features, VIRTIO_F_ANY_LAYOUT); 1386 virtio_add_feature(&features, VIRTIO_BLK_F_SCSI); 1387 } 1388 1389 if (blk_enable_write_cache(s->blk) || 1390 (s->conf.x_enable_wce_if_config_wce && 1391 virtio_has_feature(features, VIRTIO_BLK_F_CONFIG_WCE))) { 1392 virtio_add_feature(&features, VIRTIO_BLK_F_WCE); 1393 } 1394 if (!blk_is_writable(s->blk)) { 1395 virtio_add_feature(&features, VIRTIO_BLK_F_RO); 1396 } 1397 if (s->conf.num_queues > 1) { 1398 virtio_add_feature(&features, VIRTIO_BLK_F_MQ); 1399 } 1400 1401 return features; 1402 } 1403 1404 static void virtio_blk_set_status(VirtIODevice *vdev, uint8_t status) 1405 { 1406 VirtIOBlock *s = VIRTIO_BLK(vdev); 1407 1408 if (!(status & (VIRTIO_CONFIG_S_DRIVER | VIRTIO_CONFIG_S_DRIVER_OK))) { 1409 assert(!s->ioeventfd_started); 1410 } 1411 1412 if (!(status & VIRTIO_CONFIG_S_DRIVER_OK)) { 1413 return; 1414 } 1415 1416 /* A guest that supports VIRTIO_BLK_F_CONFIG_WCE must be able to send 1417 * cache flushes. Thus, the "auto writethrough" behavior is never 1418 * necessary for guests that support the VIRTIO_BLK_F_CONFIG_WCE feature. 1419 * Leaving it enabled would break the following sequence: 1420 * 1421 * Guest started with "-drive cache=writethrough" 1422 * Guest sets status to 0 1423 * Guest sets DRIVER bit in status field 1424 * Guest reads host features (WCE=0, CONFIG_WCE=1) 1425 * Guest writes guest features (WCE=0, CONFIG_WCE=1) 1426 * Guest writes 1 to the WCE configuration field (writeback mode) 1427 * Guest sets DRIVER_OK bit in status field 1428 * 1429 * s->blk would erroneously be placed in writethrough mode. 1430 */ 1431 if (!virtio_vdev_has_feature(vdev, VIRTIO_BLK_F_CONFIG_WCE)) { 1432 blk_set_enable_write_cache(s->blk, 1433 virtio_vdev_has_feature(vdev, 1434 VIRTIO_BLK_F_WCE)); 1435 } 1436 } 1437 1438 static void virtio_blk_save_device(VirtIODevice *vdev, QEMUFile *f) 1439 { 1440 VirtIOBlock *s = VIRTIO_BLK(vdev); 1441 1442 WITH_QEMU_LOCK_GUARD(&s->rq_lock) { 1443 VirtIOBlockReq *req = s->rq; 1444 1445 while (req) { 1446 qemu_put_sbyte(f, 1); 1447 1448 if (s->conf.num_queues > 1) { 1449 qemu_put_be32(f, virtio_get_queue_index(req->vq)); 1450 } 1451 1452 qemu_put_virtqueue_element(vdev, f, &req->elem); 1453 req = req->next; 1454 } 1455 } 1456 1457 qemu_put_sbyte(f, 0); 1458 } 1459 1460 static int virtio_blk_load_device(VirtIODevice *vdev, QEMUFile *f, 1461 int version_id) 1462 { 1463 VirtIOBlock *s = VIRTIO_BLK(vdev); 1464 1465 while (qemu_get_sbyte(f)) { 1466 unsigned nvqs = s->conf.num_queues; 1467 unsigned vq_idx = 0; 1468 VirtIOBlockReq *req; 1469 1470 if (nvqs > 1) { 1471 vq_idx = qemu_get_be32(f); 1472 1473 if (vq_idx >= nvqs) { 1474 error_report("Invalid virtqueue index in request list: %#x", 1475 vq_idx); 1476 return -EINVAL; 1477 } 1478 } 1479 1480 req = qemu_get_virtqueue_element(vdev, f, sizeof(VirtIOBlockReq)); 1481 virtio_blk_init_request(s, virtio_get_queue(vdev, vq_idx), req); 1482 1483 WITH_QEMU_LOCK_GUARD(&s->rq_lock) { 1484 req->next = s->rq; 1485 s->rq = req; 1486 } 1487 } 1488 1489 return 0; 1490 } 1491 1492 static void virtio_resize_cb(void *opaque) 1493 { 1494 VirtIODevice *vdev = opaque; 1495 1496 assert(qemu_get_current_aio_context() == qemu_get_aio_context()); 1497 virtio_notify_config(vdev); 1498 } 1499 1500 static void virtio_blk_resize(void *opaque) 1501 { 1502 VirtIODevice *vdev = VIRTIO_DEVICE(opaque); 1503 1504 /* 1505 * virtio_notify_config() needs to acquire the BQL, 1506 * so it can't be called from an iothread. Instead, schedule 1507 * it to be run in the main context BH. 1508 */ 1509 aio_bh_schedule_oneshot(qemu_get_aio_context(), virtio_resize_cb, vdev); 1510 } 1511 1512 static void virtio_blk_ioeventfd_detach(VirtIOBlock *s) 1513 { 1514 VirtIODevice *vdev = VIRTIO_DEVICE(s); 1515 1516 for (uint16_t i = 0; i < s->conf.num_queues; i++) { 1517 VirtQueue *vq = virtio_get_queue(vdev, i); 1518 virtio_queue_aio_detach_host_notifier(vq, s->vq_aio_context[i]); 1519 } 1520 } 1521 1522 static void virtio_blk_ioeventfd_attach(VirtIOBlock *s) 1523 { 1524 VirtIODevice *vdev = VIRTIO_DEVICE(s); 1525 1526 for (uint16_t i = 0; i < s->conf.num_queues; i++) { 1527 VirtQueue *vq = virtio_get_queue(vdev, i); 1528 virtio_queue_aio_attach_host_notifier(vq, s->vq_aio_context[i]); 1529 } 1530 } 1531 1532 /* Suspend virtqueue ioeventfd processing during drain */ 1533 static void virtio_blk_drained_begin(void *opaque) 1534 { 1535 VirtIOBlock *s = opaque; 1536 1537 if (s->ioeventfd_started) { 1538 virtio_blk_ioeventfd_detach(s); 1539 } 1540 } 1541 1542 /* Resume virtqueue ioeventfd processing after drain */ 1543 static void virtio_blk_drained_end(void *opaque) 1544 { 1545 VirtIOBlock *s = opaque; 1546 1547 if (s->ioeventfd_started) { 1548 virtio_blk_ioeventfd_attach(s); 1549 } 1550 } 1551 1552 static const BlockDevOps virtio_block_ops = { 1553 .resize_cb = virtio_blk_resize, 1554 .drained_begin = virtio_blk_drained_begin, 1555 .drained_end = virtio_blk_drained_end, 1556 }; 1557 1558 static bool 1559 validate_iothread_vq_mapping_list(IOThreadVirtQueueMappingList *list, 1560 uint16_t num_queues, Error **errp) 1561 { 1562 g_autofree unsigned long *vqs = bitmap_new(num_queues); 1563 g_autoptr(GHashTable) iothreads = 1564 g_hash_table_new(g_str_hash, g_str_equal); 1565 1566 for (IOThreadVirtQueueMappingList *node = list; node; node = node->next) { 1567 const char *name = node->value->iothread; 1568 uint16List *vq; 1569 1570 if (!iothread_by_id(name)) { 1571 error_setg(errp, "IOThread \"%s\" object does not exist", name); 1572 return false; 1573 } 1574 1575 if (!g_hash_table_add(iothreads, (gpointer)name)) { 1576 error_setg(errp, 1577 "duplicate IOThread name \"%s\" in iothread-vq-mapping", 1578 name); 1579 return false; 1580 } 1581 1582 if (node != list) { 1583 if (!!node->value->vqs != !!list->value->vqs) { 1584 error_setg(errp, "either all items in iothread-vq-mapping " 1585 "must have vqs or none of them must have it"); 1586 return false; 1587 } 1588 } 1589 1590 for (vq = node->value->vqs; vq; vq = vq->next) { 1591 if (vq->value >= num_queues) { 1592 error_setg(errp, "vq index %u for IOThread \"%s\" must be " 1593 "less than num_queues %u in iothread-vq-mapping", 1594 vq->value, name, num_queues); 1595 return false; 1596 } 1597 1598 if (test_and_set_bit(vq->value, vqs)) { 1599 error_setg(errp, "cannot assign vq %u to IOThread \"%s\" " 1600 "because it is already assigned", vq->value, name); 1601 return false; 1602 } 1603 } 1604 } 1605 1606 if (list->value->vqs) { 1607 for (uint16_t i = 0; i < num_queues; i++) { 1608 if (!test_bit(i, vqs)) { 1609 error_setg(errp, 1610 "missing vq %u IOThread assignment in iothread-vq-mapping", 1611 i); 1612 return false; 1613 } 1614 } 1615 } 1616 1617 return true; 1618 } 1619 1620 /** 1621 * apply_iothread_vq_mapping: 1622 * @iothread_vq_mapping_list: The mapping of virtqueues to IOThreads. 1623 * @vq_aio_context: The array of AioContext pointers to fill in. 1624 * @num_queues: The length of @vq_aio_context. 1625 * @errp: If an error occurs, a pointer to the area to store the error. 1626 * 1627 * Fill in the AioContext for each virtqueue in the @vq_aio_context array given 1628 * the iothread-vq-mapping parameter in @iothread_vq_mapping_list. 1629 * 1630 * Returns: %true on success, %false on failure. 1631 **/ 1632 static bool apply_iothread_vq_mapping( 1633 IOThreadVirtQueueMappingList *iothread_vq_mapping_list, 1634 AioContext **vq_aio_context, 1635 uint16_t num_queues, 1636 Error **errp) 1637 { 1638 IOThreadVirtQueueMappingList *node; 1639 size_t num_iothreads = 0; 1640 size_t cur_iothread = 0; 1641 1642 if (!validate_iothread_vq_mapping_list(iothread_vq_mapping_list, 1643 num_queues, errp)) { 1644 return false; 1645 } 1646 1647 for (node = iothread_vq_mapping_list; node; node = node->next) { 1648 num_iothreads++; 1649 } 1650 1651 for (node = iothread_vq_mapping_list; node; node = node->next) { 1652 IOThread *iothread = iothread_by_id(node->value->iothread); 1653 AioContext *ctx = iothread_get_aio_context(iothread); 1654 1655 /* Released in virtio_blk_vq_aio_context_cleanup() */ 1656 object_ref(OBJECT(iothread)); 1657 1658 if (node->value->vqs) { 1659 uint16List *vq; 1660 1661 /* Explicit vq:IOThread assignment */ 1662 for (vq = node->value->vqs; vq; vq = vq->next) { 1663 assert(vq->value < num_queues); 1664 vq_aio_context[vq->value] = ctx; 1665 } 1666 } else { 1667 /* Round-robin vq:IOThread assignment */ 1668 for (unsigned i = cur_iothread; i < num_queues; 1669 i += num_iothreads) { 1670 vq_aio_context[i] = ctx; 1671 } 1672 } 1673 1674 cur_iothread++; 1675 } 1676 1677 return true; 1678 } 1679 1680 /* Context: BQL held */ 1681 static bool virtio_blk_vq_aio_context_init(VirtIOBlock *s, Error **errp) 1682 { 1683 VirtIODevice *vdev = VIRTIO_DEVICE(s); 1684 VirtIOBlkConf *conf = &s->conf; 1685 BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev))); 1686 VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(qbus); 1687 1688 if (conf->iothread && conf->iothread_vq_mapping_list) { 1689 error_setg(errp, 1690 "iothread and iothread-vq-mapping properties cannot be set " 1691 "at the same time"); 1692 return false; 1693 } 1694 1695 if (conf->iothread || conf->iothread_vq_mapping_list) { 1696 if (!k->set_guest_notifiers || !k->ioeventfd_assign) { 1697 error_setg(errp, 1698 "device is incompatible with iothread " 1699 "(transport does not support notifiers)"); 1700 return false; 1701 } 1702 if (!virtio_device_ioeventfd_enabled(vdev)) { 1703 error_setg(errp, "ioeventfd is required for iothread"); 1704 return false; 1705 } 1706 1707 /* 1708 * If ioeventfd is (re-)enabled while the guest is running there could 1709 * be block jobs that can conflict. 1710 */ 1711 if (blk_op_is_blocked(conf->conf.blk, BLOCK_OP_TYPE_DATAPLANE, errp)) { 1712 error_prepend(errp, "cannot start virtio-blk ioeventfd: "); 1713 return false; 1714 } 1715 } 1716 1717 s->vq_aio_context = g_new(AioContext *, conf->num_queues); 1718 1719 if (conf->iothread_vq_mapping_list) { 1720 if (!apply_iothread_vq_mapping(conf->iothread_vq_mapping_list, 1721 s->vq_aio_context, 1722 conf->num_queues, 1723 errp)) { 1724 g_free(s->vq_aio_context); 1725 s->vq_aio_context = NULL; 1726 return false; 1727 } 1728 } else if (conf->iothread) { 1729 AioContext *ctx = iothread_get_aio_context(conf->iothread); 1730 for (unsigned i = 0; i < conf->num_queues; i++) { 1731 s->vq_aio_context[i] = ctx; 1732 } 1733 1734 /* Released in virtio_blk_vq_aio_context_cleanup() */ 1735 object_ref(OBJECT(conf->iothread)); 1736 } else { 1737 AioContext *ctx = qemu_get_aio_context(); 1738 for (unsigned i = 0; i < conf->num_queues; i++) { 1739 s->vq_aio_context[i] = ctx; 1740 } 1741 } 1742 1743 return true; 1744 } 1745 1746 /* Context: BQL held */ 1747 static void virtio_blk_vq_aio_context_cleanup(VirtIOBlock *s) 1748 { 1749 VirtIOBlkConf *conf = &s->conf; 1750 1751 assert(!s->ioeventfd_started); 1752 1753 if (conf->iothread_vq_mapping_list) { 1754 IOThreadVirtQueueMappingList *node; 1755 1756 for (node = conf->iothread_vq_mapping_list; node; node = node->next) { 1757 IOThread *iothread = iothread_by_id(node->value->iothread); 1758 object_unref(OBJECT(iothread)); 1759 } 1760 } 1761 1762 if (conf->iothread) { 1763 object_unref(OBJECT(conf->iothread)); 1764 } 1765 1766 g_free(s->vq_aio_context); 1767 s->vq_aio_context = NULL; 1768 } 1769 1770 /* Context: BQL held */ 1771 static int virtio_blk_start_ioeventfd(VirtIODevice *vdev) 1772 { 1773 VirtIOBlock *s = VIRTIO_BLK(vdev); 1774 BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(s))); 1775 VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(qbus); 1776 unsigned i; 1777 unsigned nvqs = s->conf.num_queues; 1778 Error *local_err = NULL; 1779 int r; 1780 1781 if (s->ioeventfd_started || s->ioeventfd_starting) { 1782 return 0; 1783 } 1784 1785 s->ioeventfd_starting = true; 1786 1787 /* Set up guest notifier (irq) */ 1788 r = k->set_guest_notifiers(qbus->parent, nvqs, true); 1789 if (r != 0) { 1790 error_report("virtio-blk failed to set guest notifier (%d), " 1791 "ensure -accel kvm is set.", r); 1792 goto fail_guest_notifiers; 1793 } 1794 1795 /* 1796 * Batch all the host notifiers in a single transaction to avoid 1797 * quadratic time complexity in address_space_update_ioeventfds(). 1798 */ 1799 memory_region_transaction_begin(); 1800 1801 /* Set up virtqueue notify */ 1802 for (i = 0; i < nvqs; i++) { 1803 r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), i, true); 1804 if (r != 0) { 1805 int j = i; 1806 1807 fprintf(stderr, "virtio-blk failed to set host notifier (%d)\n", r); 1808 while (i--) { 1809 virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), i, false); 1810 } 1811 1812 /* 1813 * The transaction expects the ioeventfds to be open when it 1814 * commits. Do it now, before the cleanup loop. 1815 */ 1816 memory_region_transaction_commit(); 1817 1818 while (j--) { 1819 virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), j); 1820 } 1821 goto fail_host_notifiers; 1822 } 1823 } 1824 1825 memory_region_transaction_commit(); 1826 1827 /* 1828 * Try to change the AioContext so that block jobs and other operations can 1829 * co-locate their activity in the same AioContext. If it fails, nevermind. 1830 */ 1831 assert(nvqs > 0); /* enforced during ->realize() */ 1832 r = blk_set_aio_context(s->conf.conf.blk, s->vq_aio_context[0], 1833 &local_err); 1834 if (r < 0) { 1835 warn_report_err(local_err); 1836 } 1837 1838 /* 1839 * These fields must be visible to the IOThread when it processes the 1840 * virtqueue, otherwise it will think ioeventfd has not started yet. 1841 * 1842 * Make sure ->ioeventfd_started is false when blk_set_aio_context() is 1843 * called above so that draining does not cause the host notifier to be 1844 * detached/attached prematurely. 1845 */ 1846 s->ioeventfd_starting = false; 1847 s->ioeventfd_started = true; 1848 smp_wmb(); /* paired with aio_notify_accept() on the read side */ 1849 1850 /* Get this show started by hooking up our callbacks */ 1851 for (i = 0; i < nvqs; i++) { 1852 VirtQueue *vq = virtio_get_queue(vdev, i); 1853 AioContext *ctx = s->vq_aio_context[i]; 1854 1855 /* Kick right away to begin processing requests already in vring */ 1856 event_notifier_set(virtio_queue_get_host_notifier(vq)); 1857 1858 if (!blk_in_drain(s->conf.conf.blk)) { 1859 virtio_queue_aio_attach_host_notifier(vq, ctx); 1860 } 1861 } 1862 return 0; 1863 1864 fail_host_notifiers: 1865 k->set_guest_notifiers(qbus->parent, nvqs, false); 1866 fail_guest_notifiers: 1867 s->ioeventfd_disabled = true; 1868 s->ioeventfd_starting = false; 1869 return -ENOSYS; 1870 } 1871 1872 /* Stop notifications for new requests from guest. 1873 * 1874 * Context: BH in IOThread 1875 */ 1876 static void virtio_blk_ioeventfd_stop_vq_bh(void *opaque) 1877 { 1878 VirtQueue *vq = opaque; 1879 EventNotifier *host_notifier = virtio_queue_get_host_notifier(vq); 1880 1881 virtio_queue_aio_detach_host_notifier(vq, qemu_get_current_aio_context()); 1882 1883 /* 1884 * Test and clear notifier after disabling event, in case poll callback 1885 * didn't have time to run. 1886 */ 1887 virtio_queue_host_notifier_read(host_notifier); 1888 } 1889 1890 /* Context: BQL held */ 1891 static void virtio_blk_stop_ioeventfd(VirtIODevice *vdev) 1892 { 1893 VirtIOBlock *s = VIRTIO_BLK(vdev); 1894 BusState *qbus = qdev_get_parent_bus(DEVICE(s)); 1895 VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(qbus); 1896 unsigned i; 1897 unsigned nvqs = s->conf.num_queues; 1898 1899 if (!s->ioeventfd_started || s->ioeventfd_stopping) { 1900 return; 1901 } 1902 1903 /* Better luck next time. */ 1904 if (s->ioeventfd_disabled) { 1905 s->ioeventfd_disabled = false; 1906 s->ioeventfd_started = false; 1907 return; 1908 } 1909 s->ioeventfd_stopping = true; 1910 1911 if (!blk_in_drain(s->conf.conf.blk)) { 1912 for (i = 0; i < nvqs; i++) { 1913 VirtQueue *vq = virtio_get_queue(vdev, i); 1914 AioContext *ctx = s->vq_aio_context[i]; 1915 1916 aio_wait_bh_oneshot(ctx, virtio_blk_ioeventfd_stop_vq_bh, vq); 1917 } 1918 } 1919 1920 /* 1921 * Batch all the host notifiers in a single transaction to avoid 1922 * quadratic time complexity in address_space_update_ioeventfds(). 1923 */ 1924 memory_region_transaction_begin(); 1925 1926 for (i = 0; i < nvqs; i++) { 1927 virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), i, false); 1928 } 1929 1930 /* 1931 * The transaction expects the ioeventfds to be open when it 1932 * commits. Do it now, before the cleanup loop. 1933 */ 1934 memory_region_transaction_commit(); 1935 1936 for (i = 0; i < nvqs; i++) { 1937 virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), i); 1938 } 1939 1940 /* 1941 * Set ->ioeventfd_started to false before draining so that host notifiers 1942 * are not detached/attached anymore. 1943 */ 1944 s->ioeventfd_started = false; 1945 1946 /* Wait for virtio_blk_dma_restart_bh() and in flight I/O to complete */ 1947 blk_drain(s->conf.conf.blk); 1948 1949 /* 1950 * Try to switch bs back to the QEMU main loop. If other users keep the 1951 * BlockBackend in the iothread, that's ok 1952 */ 1953 blk_set_aio_context(s->conf.conf.blk, qemu_get_aio_context(), NULL); 1954 1955 /* Clean up guest notifier (irq) */ 1956 k->set_guest_notifiers(qbus->parent, nvqs, false); 1957 1958 s->ioeventfd_stopping = false; 1959 } 1960 1961 static void virtio_blk_device_realize(DeviceState *dev, Error **errp) 1962 { 1963 VirtIODevice *vdev = VIRTIO_DEVICE(dev); 1964 VirtIOBlock *s = VIRTIO_BLK(dev); 1965 VirtIOBlkConf *conf = &s->conf; 1966 BlockDriverState *bs; 1967 Error *err = NULL; 1968 unsigned i; 1969 1970 if (!conf->conf.blk) { 1971 error_setg(errp, "drive property not set"); 1972 return; 1973 } 1974 if (!blk_is_inserted(conf->conf.blk)) { 1975 error_setg(errp, "Device needs media, but drive is empty"); 1976 return; 1977 } 1978 if (conf->num_queues == VIRTIO_BLK_AUTO_NUM_QUEUES) { 1979 conf->num_queues = 1; 1980 } 1981 if (!conf->num_queues) { 1982 error_setg(errp, "num-queues property must be larger than 0"); 1983 return; 1984 } 1985 if (conf->queue_size <= 2) { 1986 error_setg(errp, "invalid queue-size property (%" PRIu16 "), " 1987 "must be > 2", conf->queue_size); 1988 return; 1989 } 1990 if (!is_power_of_2(conf->queue_size) || 1991 conf->queue_size > VIRTQUEUE_MAX_SIZE) { 1992 error_setg(errp, "invalid queue-size property (%" PRIu16 "), " 1993 "must be a power of 2 (max %d)", 1994 conf->queue_size, VIRTQUEUE_MAX_SIZE); 1995 return; 1996 } 1997 1998 if (!blkconf_apply_backend_options(&conf->conf, 1999 !blk_supports_write_perm(conf->conf.blk), 2000 true, errp)) { 2001 return; 2002 } 2003 s->original_wce = blk_enable_write_cache(conf->conf.blk); 2004 if (!blkconf_geometry(&conf->conf, NULL, 65535, 255, 255, errp)) { 2005 return; 2006 } 2007 2008 if (!blkconf_blocksizes(&conf->conf, errp)) { 2009 return; 2010 } 2011 2012 bs = blk_bs(conf->conf.blk); 2013 if (bs->bl.zoned != BLK_Z_NONE) { 2014 virtio_add_feature(&s->host_features, VIRTIO_BLK_F_ZONED); 2015 if (bs->bl.zoned == BLK_Z_HM) { 2016 virtio_clear_feature(&s->host_features, VIRTIO_BLK_F_DISCARD); 2017 } 2018 } 2019 2020 if (virtio_has_feature(s->host_features, VIRTIO_BLK_F_DISCARD) && 2021 (!conf->max_discard_sectors || 2022 conf->max_discard_sectors > BDRV_REQUEST_MAX_SECTORS)) { 2023 error_setg(errp, "invalid max-discard-sectors property (%" PRIu32 ")" 2024 ", must be between 1 and %d", 2025 conf->max_discard_sectors, (int)BDRV_REQUEST_MAX_SECTORS); 2026 return; 2027 } 2028 2029 if (virtio_has_feature(s->host_features, VIRTIO_BLK_F_WRITE_ZEROES) && 2030 (!conf->max_write_zeroes_sectors || 2031 conf->max_write_zeroes_sectors > BDRV_REQUEST_MAX_SECTORS)) { 2032 error_setg(errp, "invalid max-write-zeroes-sectors property (%" PRIu32 2033 "), must be between 1 and %d", 2034 conf->max_write_zeroes_sectors, 2035 (int)BDRV_REQUEST_MAX_SECTORS); 2036 return; 2037 } 2038 2039 s->config_size = virtio_get_config_size(&virtio_blk_cfg_size_params, 2040 s->host_features); 2041 virtio_init(vdev, VIRTIO_ID_BLOCK, s->config_size); 2042 2043 qemu_mutex_init(&s->rq_lock); 2044 2045 s->blk = conf->conf.blk; 2046 s->rq = NULL; 2047 s->sector_mask = (s->conf.conf.logical_block_size / BDRV_SECTOR_SIZE) - 1; 2048 2049 for (i = 0; i < conf->num_queues; i++) { 2050 virtio_add_queue(vdev, conf->queue_size, virtio_blk_handle_output); 2051 } 2052 qemu_coroutine_inc_pool_size(conf->num_queues * conf->queue_size / 2); 2053 2054 /* Don't start ioeventfd if transport does not support notifiers. */ 2055 if (!virtio_device_ioeventfd_enabled(vdev)) { 2056 s->ioeventfd_disabled = true; 2057 } 2058 2059 virtio_blk_vq_aio_context_init(s, &err); 2060 if (err != NULL) { 2061 error_propagate(errp, err); 2062 for (i = 0; i < conf->num_queues; i++) { 2063 virtio_del_queue(vdev, i); 2064 } 2065 virtio_cleanup(vdev); 2066 return; 2067 } 2068 2069 /* 2070 * This must be after virtio_init() so virtio_blk_dma_restart_cb() gets 2071 * called after ->start_ioeventfd() has already set blk's AioContext. 2072 */ 2073 s->change = 2074 qdev_add_vm_change_state_handler(dev, virtio_blk_dma_restart_cb, s); 2075 2076 blk_ram_registrar_init(&s->blk_ram_registrar, s->blk); 2077 blk_set_dev_ops(s->blk, &virtio_block_ops, s); 2078 2079 blk_iostatus_enable(s->blk); 2080 2081 add_boot_device_lchs(dev, "/disk@0,0", 2082 conf->conf.lcyls, 2083 conf->conf.lheads, 2084 conf->conf.lsecs); 2085 } 2086 2087 static void virtio_blk_device_unrealize(DeviceState *dev) 2088 { 2089 VirtIODevice *vdev = VIRTIO_DEVICE(dev); 2090 VirtIOBlock *s = VIRTIO_BLK(dev); 2091 VirtIOBlkConf *conf = &s->conf; 2092 unsigned i; 2093 2094 blk_drain(s->blk); 2095 del_boot_device_lchs(dev, "/disk@0,0"); 2096 virtio_blk_vq_aio_context_cleanup(s); 2097 for (i = 0; i < conf->num_queues; i++) { 2098 virtio_del_queue(vdev, i); 2099 } 2100 qemu_coroutine_dec_pool_size(conf->num_queues * conf->queue_size / 2); 2101 qemu_mutex_destroy(&s->rq_lock); 2102 blk_ram_registrar_destroy(&s->blk_ram_registrar); 2103 qemu_del_vm_change_state_handler(s->change); 2104 blockdev_mark_auto_del(s->blk); 2105 virtio_cleanup(vdev); 2106 } 2107 2108 static void virtio_blk_instance_init(Object *obj) 2109 { 2110 VirtIOBlock *s = VIRTIO_BLK(obj); 2111 2112 device_add_bootindex_property(obj, &s->conf.conf.bootindex, 2113 "bootindex", "/disk@0,0", 2114 DEVICE(obj)); 2115 } 2116 2117 static const VMStateDescription vmstate_virtio_blk = { 2118 .name = "virtio-blk", 2119 .minimum_version_id = 2, 2120 .version_id = 2, 2121 .fields = (const VMStateField[]) { 2122 VMSTATE_VIRTIO_DEVICE, 2123 VMSTATE_END_OF_LIST() 2124 }, 2125 }; 2126 2127 static Property virtio_blk_properties[] = { 2128 DEFINE_BLOCK_PROPERTIES(VirtIOBlock, conf.conf), 2129 DEFINE_BLOCK_ERROR_PROPERTIES(VirtIOBlock, conf.conf), 2130 DEFINE_BLOCK_CHS_PROPERTIES(VirtIOBlock, conf.conf), 2131 DEFINE_PROP_STRING("serial", VirtIOBlock, conf.serial), 2132 DEFINE_PROP_BIT64("config-wce", VirtIOBlock, host_features, 2133 VIRTIO_BLK_F_CONFIG_WCE, true), 2134 #ifdef __linux__ 2135 DEFINE_PROP_BIT64("scsi", VirtIOBlock, host_features, 2136 VIRTIO_BLK_F_SCSI, false), 2137 #endif 2138 DEFINE_PROP_BIT("request-merging", VirtIOBlock, conf.request_merging, 0, 2139 true), 2140 DEFINE_PROP_UINT16("num-queues", VirtIOBlock, conf.num_queues, 2141 VIRTIO_BLK_AUTO_NUM_QUEUES), 2142 DEFINE_PROP_UINT16("queue-size", VirtIOBlock, conf.queue_size, 256), 2143 DEFINE_PROP_BOOL("seg-max-adjust", VirtIOBlock, conf.seg_max_adjust, true), 2144 DEFINE_PROP_LINK("iothread", VirtIOBlock, conf.iothread, TYPE_IOTHREAD, 2145 IOThread *), 2146 DEFINE_PROP_IOTHREAD_VQ_MAPPING_LIST("iothread-vq-mapping", VirtIOBlock, 2147 conf.iothread_vq_mapping_list), 2148 DEFINE_PROP_BIT64("discard", VirtIOBlock, host_features, 2149 VIRTIO_BLK_F_DISCARD, true), 2150 DEFINE_PROP_BOOL("report-discard-granularity", VirtIOBlock, 2151 conf.report_discard_granularity, true), 2152 DEFINE_PROP_BIT64("write-zeroes", VirtIOBlock, host_features, 2153 VIRTIO_BLK_F_WRITE_ZEROES, true), 2154 DEFINE_PROP_UINT32("max-discard-sectors", VirtIOBlock, 2155 conf.max_discard_sectors, BDRV_REQUEST_MAX_SECTORS), 2156 DEFINE_PROP_UINT32("max-write-zeroes-sectors", VirtIOBlock, 2157 conf.max_write_zeroes_sectors, BDRV_REQUEST_MAX_SECTORS), 2158 DEFINE_PROP_BOOL("x-enable-wce-if-config-wce", VirtIOBlock, 2159 conf.x_enable_wce_if_config_wce, true), 2160 DEFINE_PROP_END_OF_LIST(), 2161 }; 2162 2163 static void virtio_blk_class_init(ObjectClass *klass, void *data) 2164 { 2165 DeviceClass *dc = DEVICE_CLASS(klass); 2166 VirtioDeviceClass *vdc = VIRTIO_DEVICE_CLASS(klass); 2167 2168 device_class_set_props(dc, virtio_blk_properties); 2169 dc->vmsd = &vmstate_virtio_blk; 2170 set_bit(DEVICE_CATEGORY_STORAGE, dc->categories); 2171 vdc->realize = virtio_blk_device_realize; 2172 vdc->unrealize = virtio_blk_device_unrealize; 2173 vdc->get_config = virtio_blk_update_config; 2174 vdc->set_config = virtio_blk_set_config; 2175 vdc->get_features = virtio_blk_get_features; 2176 vdc->set_status = virtio_blk_set_status; 2177 vdc->reset = virtio_blk_reset; 2178 vdc->save = virtio_blk_save_device; 2179 vdc->load = virtio_blk_load_device; 2180 vdc->start_ioeventfd = virtio_blk_start_ioeventfd; 2181 vdc->stop_ioeventfd = virtio_blk_stop_ioeventfd; 2182 } 2183 2184 static const TypeInfo virtio_blk_info = { 2185 .name = TYPE_VIRTIO_BLK, 2186 .parent = TYPE_VIRTIO_DEVICE, 2187 .instance_size = sizeof(VirtIOBlock), 2188 .instance_init = virtio_blk_instance_init, 2189 .class_init = virtio_blk_class_init, 2190 }; 2191 2192 static void virtio_register_types(void) 2193 { 2194 type_register_static(&virtio_blk_info); 2195 } 2196 2197 type_init(virtio_register_types) 2198