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_ioeventfd_attach(VirtIOBlock *s); 41 42 static void virtio_blk_init_request(VirtIOBlock *s, VirtQueue *vq, 43 VirtIOBlockReq *req) 44 { 45 req->dev = s; 46 req->vq = vq; 47 req->qiov.size = 0; 48 req->in_len = 0; 49 req->next = NULL; 50 req->mr_next = NULL; 51 } 52 53 static void virtio_blk_free_request(VirtIOBlockReq *req) 54 { 55 g_free(req); 56 } 57 58 static void virtio_blk_req_complete(VirtIOBlockReq *req, unsigned char status) 59 { 60 VirtIOBlock *s = req->dev; 61 VirtIODevice *vdev = VIRTIO_DEVICE(s); 62 63 trace_virtio_blk_req_complete(vdev, req, status); 64 65 stb_p(&req->in->status, status); 66 iov_discard_undo(&req->inhdr_undo); 67 iov_discard_undo(&req->outhdr_undo); 68 virtqueue_push(req->vq, &req->elem, req->in_len); 69 if (qemu_in_iothread()) { 70 virtio_notify_irqfd(vdev, req->vq); 71 } else { 72 virtio_notify(vdev, req->vq); 73 } 74 } 75 76 static int virtio_blk_handle_rw_error(VirtIOBlockReq *req, int error, 77 bool is_read, bool acct_failed) 78 { 79 VirtIOBlock *s = req->dev; 80 BlockErrorAction action = blk_get_error_action(s->blk, is_read, error); 81 82 if (action == BLOCK_ERROR_ACTION_STOP) { 83 /* Break the link as the next request is going to be parsed from the 84 * ring again. Otherwise we may end up doing a double completion! */ 85 req->mr_next = NULL; 86 87 WITH_QEMU_LOCK_GUARD(&s->rq_lock) { 88 req->next = s->rq; 89 s->rq = req; 90 } 91 } else if (action == BLOCK_ERROR_ACTION_REPORT) { 92 virtio_blk_req_complete(req, VIRTIO_BLK_S_IOERR); 93 if (acct_failed) { 94 block_acct_failed(blk_get_stats(s->blk), &req->acct); 95 } 96 virtio_blk_free_request(req); 97 } 98 99 blk_error_action(s->blk, action, is_read, error); 100 return action != BLOCK_ERROR_ACTION_IGNORE; 101 } 102 103 static void virtio_blk_rw_complete(void *opaque, int ret) 104 { 105 VirtIOBlockReq *next = opaque; 106 VirtIOBlock *s = next->dev; 107 VirtIODevice *vdev = VIRTIO_DEVICE(s); 108 109 while (next) { 110 VirtIOBlockReq *req = next; 111 next = req->mr_next; 112 trace_virtio_blk_rw_complete(vdev, req, ret); 113 114 if (req->qiov.nalloc != -1) { 115 /* If nalloc is != -1 req->qiov is a local copy of the original 116 * external iovec. It was allocated in submit_requests to be 117 * able to merge requests. */ 118 qemu_iovec_destroy(&req->qiov); 119 } 120 121 if (ret) { 122 int p = virtio_ldl_p(VIRTIO_DEVICE(s), &req->out.type); 123 bool is_read = !(p & VIRTIO_BLK_T_OUT); 124 /* Note that memory may be dirtied on read failure. If the 125 * virtio request is not completed here, as is the case for 126 * BLOCK_ERROR_ACTION_STOP, the memory may not be copied 127 * correctly during live migration. While this is ugly, 128 * it is acceptable because the device is free to write to 129 * the memory until the request is completed (which will 130 * happen on the other side of the migration). 131 */ 132 if (virtio_blk_handle_rw_error(req, -ret, is_read, true)) { 133 continue; 134 } 135 } 136 137 virtio_blk_req_complete(req, VIRTIO_BLK_S_OK); 138 block_acct_done(blk_get_stats(s->blk), &req->acct); 139 virtio_blk_free_request(req); 140 } 141 } 142 143 static void virtio_blk_flush_complete(void *opaque, int ret) 144 { 145 VirtIOBlockReq *req = opaque; 146 VirtIOBlock *s = req->dev; 147 148 if (ret && virtio_blk_handle_rw_error(req, -ret, 0, true)) { 149 return; 150 } 151 152 virtio_blk_req_complete(req, VIRTIO_BLK_S_OK); 153 block_acct_done(blk_get_stats(s->blk), &req->acct); 154 virtio_blk_free_request(req); 155 } 156 157 static void virtio_blk_discard_write_zeroes_complete(void *opaque, int ret) 158 { 159 VirtIOBlockReq *req = opaque; 160 VirtIOBlock *s = req->dev; 161 bool is_write_zeroes = (virtio_ldl_p(VIRTIO_DEVICE(s), &req->out.type) & 162 ~VIRTIO_BLK_T_BARRIER) == VIRTIO_BLK_T_WRITE_ZEROES; 163 164 if (ret && virtio_blk_handle_rw_error(req, -ret, false, is_write_zeroes)) { 165 return; 166 } 167 168 virtio_blk_req_complete(req, VIRTIO_BLK_S_OK); 169 if (is_write_zeroes) { 170 block_acct_done(blk_get_stats(s->blk), &req->acct); 171 } 172 virtio_blk_free_request(req); 173 } 174 175 #ifdef __linux__ 176 177 typedef struct { 178 VirtIOBlockReq *req; 179 struct sg_io_hdr hdr; 180 } VirtIOBlockIoctlReq; 181 182 static void virtio_blk_ioctl_complete(void *opaque, int status) 183 { 184 VirtIOBlockIoctlReq *ioctl_req = opaque; 185 VirtIOBlockReq *req = ioctl_req->req; 186 VirtIOBlock *s = req->dev; 187 VirtIODevice *vdev = VIRTIO_DEVICE(s); 188 struct virtio_scsi_inhdr *scsi; 189 struct sg_io_hdr *hdr; 190 191 scsi = (void *)req->elem.in_sg[req->elem.in_num - 2].iov_base; 192 193 if (status) { 194 status = VIRTIO_BLK_S_UNSUPP; 195 virtio_stl_p(vdev, &scsi->errors, 255); 196 goto out; 197 } 198 199 hdr = &ioctl_req->hdr; 200 /* 201 * From SCSI-Generic-HOWTO: "Some lower level drivers (e.g. ide-scsi) 202 * clear the masked_status field [hence status gets cleared too, see 203 * block/scsi_ioctl.c] even when a CHECK_CONDITION or COMMAND_TERMINATED 204 * status has occurred. However they do set DRIVER_SENSE in driver_status 205 * field. Also a (sb_len_wr > 0) indicates there is a sense buffer. 206 */ 207 if (hdr->status == 0 && hdr->sb_len_wr > 0) { 208 hdr->status = CHECK_CONDITION; 209 } 210 211 virtio_stl_p(vdev, &scsi->errors, 212 hdr->status | (hdr->msg_status << 8) | 213 (hdr->host_status << 16) | (hdr->driver_status << 24)); 214 virtio_stl_p(vdev, &scsi->residual, hdr->resid); 215 virtio_stl_p(vdev, &scsi->sense_len, hdr->sb_len_wr); 216 virtio_stl_p(vdev, &scsi->data_len, hdr->dxfer_len); 217 218 out: 219 virtio_blk_req_complete(req, status); 220 virtio_blk_free_request(req); 221 g_free(ioctl_req); 222 } 223 224 #endif 225 226 static VirtIOBlockReq *virtio_blk_get_request(VirtIOBlock *s, VirtQueue *vq) 227 { 228 VirtIOBlockReq *req = virtqueue_pop(vq, sizeof(VirtIOBlockReq)); 229 230 if (req) { 231 virtio_blk_init_request(s, vq, req); 232 } 233 return req; 234 } 235 236 static int virtio_blk_handle_scsi_req(VirtIOBlockReq *req) 237 { 238 int status = VIRTIO_BLK_S_OK; 239 struct virtio_scsi_inhdr *scsi = NULL; 240 VirtIOBlock *blk = req->dev; 241 VirtIODevice *vdev = VIRTIO_DEVICE(blk); 242 VirtQueueElement *elem = &req->elem; 243 244 #ifdef __linux__ 245 int i; 246 VirtIOBlockIoctlReq *ioctl_req; 247 BlockAIOCB *acb; 248 #endif 249 250 /* 251 * We require at least one output segment each for the virtio_blk_outhdr 252 * and the SCSI command block. 253 * 254 * We also at least require the virtio_blk_inhdr, the virtio_scsi_inhdr 255 * and the sense buffer pointer in the input segments. 256 */ 257 if (elem->out_num < 2 || elem->in_num < 3) { 258 status = VIRTIO_BLK_S_IOERR; 259 goto fail; 260 } 261 262 /* 263 * The scsi inhdr is placed in the second-to-last input segment, just 264 * before the regular inhdr. 265 */ 266 scsi = (void *)elem->in_sg[elem->in_num - 2].iov_base; 267 268 if (!virtio_has_feature(blk->host_features, VIRTIO_BLK_F_SCSI)) { 269 status = VIRTIO_BLK_S_UNSUPP; 270 goto fail; 271 } 272 273 /* 274 * No support for bidirection commands yet. 275 */ 276 if (elem->out_num > 2 && elem->in_num > 3) { 277 status = VIRTIO_BLK_S_UNSUPP; 278 goto fail; 279 } 280 281 #ifdef __linux__ 282 ioctl_req = g_new0(VirtIOBlockIoctlReq, 1); 283 ioctl_req->req = req; 284 ioctl_req->hdr.interface_id = 'S'; 285 ioctl_req->hdr.cmd_len = elem->out_sg[1].iov_len; 286 ioctl_req->hdr.cmdp = elem->out_sg[1].iov_base; 287 ioctl_req->hdr.dxfer_len = 0; 288 289 if (elem->out_num > 2) { 290 /* 291 * If there are more than the minimally required 2 output segments 292 * there is write payload starting from the third iovec. 293 */ 294 ioctl_req->hdr.dxfer_direction = SG_DXFER_TO_DEV; 295 ioctl_req->hdr.iovec_count = elem->out_num - 2; 296 297 for (i = 0; i < ioctl_req->hdr.iovec_count; i++) { 298 ioctl_req->hdr.dxfer_len += elem->out_sg[i + 2].iov_len; 299 } 300 301 ioctl_req->hdr.dxferp = elem->out_sg + 2; 302 303 } else if (elem->in_num > 3) { 304 /* 305 * If we have more than 3 input segments the guest wants to actually 306 * read data. 307 */ 308 ioctl_req->hdr.dxfer_direction = SG_DXFER_FROM_DEV; 309 ioctl_req->hdr.iovec_count = elem->in_num - 3; 310 for (i = 0; i < ioctl_req->hdr.iovec_count; i++) { 311 ioctl_req->hdr.dxfer_len += elem->in_sg[i].iov_len; 312 } 313 314 ioctl_req->hdr.dxferp = elem->in_sg; 315 } else { 316 /* 317 * Some SCSI commands don't actually transfer any data. 318 */ 319 ioctl_req->hdr.dxfer_direction = SG_DXFER_NONE; 320 } 321 322 ioctl_req->hdr.sbp = elem->in_sg[elem->in_num - 3].iov_base; 323 ioctl_req->hdr.mx_sb_len = elem->in_sg[elem->in_num - 3].iov_len; 324 325 acb = blk_aio_ioctl(blk->blk, SG_IO, &ioctl_req->hdr, 326 virtio_blk_ioctl_complete, ioctl_req); 327 if (!acb) { 328 g_free(ioctl_req); 329 status = VIRTIO_BLK_S_UNSUPP; 330 goto fail; 331 } 332 return -EINPROGRESS; 333 #else 334 abort(); 335 #endif 336 337 fail: 338 /* Just put anything nonzero so that the ioctl fails in the guest. */ 339 if (scsi) { 340 virtio_stl_p(vdev, &scsi->errors, 255); 341 } 342 return status; 343 } 344 345 static void virtio_blk_handle_scsi(VirtIOBlockReq *req) 346 { 347 int status; 348 349 status = virtio_blk_handle_scsi_req(req); 350 if (status != -EINPROGRESS) { 351 virtio_blk_req_complete(req, status); 352 virtio_blk_free_request(req); 353 } 354 } 355 356 static inline void submit_requests(VirtIOBlock *s, MultiReqBuffer *mrb, 357 int start, int num_reqs, int niov) 358 { 359 BlockBackend *blk = s->blk; 360 QEMUIOVector *qiov = &mrb->reqs[start]->qiov; 361 int64_t sector_num = mrb->reqs[start]->sector_num; 362 bool is_write = mrb->is_write; 363 BdrvRequestFlags flags = 0; 364 365 if (num_reqs > 1) { 366 int i; 367 struct iovec *tmp_iov = qiov->iov; 368 int tmp_niov = qiov->niov; 369 370 /* mrb->reqs[start]->qiov was initialized from external so we can't 371 * modify it here. We need to initialize it locally and then add the 372 * external iovecs. */ 373 qemu_iovec_init(qiov, niov); 374 375 for (i = 0; i < tmp_niov; i++) { 376 qemu_iovec_add(qiov, tmp_iov[i].iov_base, tmp_iov[i].iov_len); 377 } 378 379 for (i = start + 1; i < start + num_reqs; i++) { 380 qemu_iovec_concat(qiov, &mrb->reqs[i]->qiov, 0, 381 mrb->reqs[i]->qiov.size); 382 mrb->reqs[i - 1]->mr_next = mrb->reqs[i]; 383 } 384 385 trace_virtio_blk_submit_multireq(VIRTIO_DEVICE(mrb->reqs[start]->dev), 386 mrb, start, num_reqs, 387 sector_num << BDRV_SECTOR_BITS, 388 qiov->size, is_write); 389 block_acct_merge_done(blk_get_stats(blk), 390 is_write ? BLOCK_ACCT_WRITE : BLOCK_ACCT_READ, 391 num_reqs - 1); 392 } 393 394 if (blk_ram_registrar_ok(&s->blk_ram_registrar)) { 395 flags |= BDRV_REQ_REGISTERED_BUF; 396 } 397 398 if (is_write) { 399 blk_aio_pwritev(blk, sector_num << BDRV_SECTOR_BITS, qiov, 400 flags, virtio_blk_rw_complete, 401 mrb->reqs[start]); 402 } else { 403 blk_aio_preadv(blk, sector_num << BDRV_SECTOR_BITS, qiov, 404 flags, virtio_blk_rw_complete, 405 mrb->reqs[start]); 406 } 407 } 408 409 static int multireq_compare(const void *a, const void *b) 410 { 411 const VirtIOBlockReq *req1 = *(VirtIOBlockReq **)a, 412 *req2 = *(VirtIOBlockReq **)b; 413 414 /* 415 * Note that we can't simply subtract sector_num1 from sector_num2 416 * here as that could overflow the return value. 417 */ 418 if (req1->sector_num > req2->sector_num) { 419 return 1; 420 } else if (req1->sector_num < req2->sector_num) { 421 return -1; 422 } else { 423 return 0; 424 } 425 } 426 427 static void virtio_blk_submit_multireq(VirtIOBlock *s, MultiReqBuffer *mrb) 428 { 429 int i = 0, start = 0, num_reqs = 0, niov = 0, nb_sectors = 0; 430 uint32_t max_transfer; 431 int64_t sector_num = 0; 432 433 if (mrb->num_reqs == 1) { 434 submit_requests(s, mrb, 0, 1, -1); 435 mrb->num_reqs = 0; 436 return; 437 } 438 439 max_transfer = blk_get_max_transfer(mrb->reqs[0]->dev->blk); 440 441 qsort(mrb->reqs, mrb->num_reqs, sizeof(*mrb->reqs), 442 &multireq_compare); 443 444 for (i = 0; i < mrb->num_reqs; i++) { 445 VirtIOBlockReq *req = mrb->reqs[i]; 446 if (num_reqs > 0) { 447 /* 448 * NOTE: We cannot merge the requests in below situations: 449 * 1. requests are not sequential 450 * 2. merge would exceed maximum number of IOVs 451 * 3. merge would exceed maximum transfer length of backend device 452 */ 453 if (sector_num + nb_sectors != req->sector_num || 454 niov > blk_get_max_iov(s->blk) - req->qiov.niov || 455 req->qiov.size > max_transfer || 456 nb_sectors > (max_transfer - 457 req->qiov.size) / BDRV_SECTOR_SIZE) { 458 submit_requests(s, mrb, start, num_reqs, niov); 459 num_reqs = 0; 460 } 461 } 462 463 if (num_reqs == 0) { 464 sector_num = req->sector_num; 465 nb_sectors = niov = 0; 466 start = i; 467 } 468 469 nb_sectors += req->qiov.size / BDRV_SECTOR_SIZE; 470 niov += req->qiov.niov; 471 num_reqs++; 472 } 473 474 submit_requests(s, mrb, start, num_reqs, niov); 475 mrb->num_reqs = 0; 476 } 477 478 static void virtio_blk_handle_flush(VirtIOBlockReq *req, MultiReqBuffer *mrb) 479 { 480 VirtIOBlock *s = req->dev; 481 482 block_acct_start(blk_get_stats(s->blk), &req->acct, 0, 483 BLOCK_ACCT_FLUSH); 484 485 /* 486 * Make sure all outstanding writes are posted to the backing device. 487 */ 488 if (mrb->is_write && mrb->num_reqs > 0) { 489 virtio_blk_submit_multireq(s, mrb); 490 } 491 blk_aio_flush(s->blk, virtio_blk_flush_complete, req); 492 } 493 494 static bool virtio_blk_sect_range_ok(VirtIOBlock *dev, 495 uint64_t sector, size_t size) 496 { 497 uint64_t nb_sectors = size >> BDRV_SECTOR_BITS; 498 uint64_t total_sectors; 499 500 if (nb_sectors > BDRV_REQUEST_MAX_SECTORS) { 501 return false; 502 } 503 if (sector & dev->sector_mask) { 504 return false; 505 } 506 if (size % dev->conf.conf.logical_block_size) { 507 return false; 508 } 509 blk_get_geometry(dev->blk, &total_sectors); 510 if (sector > total_sectors || nb_sectors > total_sectors - sector) { 511 return false; 512 } 513 return true; 514 } 515 516 static uint8_t virtio_blk_handle_discard_write_zeroes(VirtIOBlockReq *req, 517 struct virtio_blk_discard_write_zeroes *dwz_hdr, bool is_write_zeroes) 518 { 519 VirtIOBlock *s = req->dev; 520 VirtIODevice *vdev = VIRTIO_DEVICE(s); 521 uint64_t sector; 522 uint32_t num_sectors, flags, max_sectors; 523 uint8_t err_status; 524 int bytes; 525 526 sector = virtio_ldq_p(vdev, &dwz_hdr->sector); 527 num_sectors = virtio_ldl_p(vdev, &dwz_hdr->num_sectors); 528 flags = virtio_ldl_p(vdev, &dwz_hdr->flags); 529 max_sectors = is_write_zeroes ? s->conf.max_write_zeroes_sectors : 530 s->conf.max_discard_sectors; 531 532 /* 533 * max_sectors is at most BDRV_REQUEST_MAX_SECTORS, this check 534 * make us sure that "num_sectors << BDRV_SECTOR_BITS" can fit in 535 * the integer variable. 536 */ 537 if (unlikely(num_sectors > max_sectors)) { 538 err_status = VIRTIO_BLK_S_IOERR; 539 goto err; 540 } 541 542 bytes = num_sectors << BDRV_SECTOR_BITS; 543 544 if (unlikely(!virtio_blk_sect_range_ok(s, sector, bytes))) { 545 err_status = VIRTIO_BLK_S_IOERR; 546 goto err; 547 } 548 549 /* 550 * The device MUST set the status byte to VIRTIO_BLK_S_UNSUPP for discard 551 * and write zeroes commands if any unknown flag is set. 552 */ 553 if (unlikely(flags & ~VIRTIO_BLK_WRITE_ZEROES_FLAG_UNMAP)) { 554 err_status = VIRTIO_BLK_S_UNSUPP; 555 goto err; 556 } 557 558 if (is_write_zeroes) { /* VIRTIO_BLK_T_WRITE_ZEROES */ 559 int blk_aio_flags = 0; 560 561 if (flags & VIRTIO_BLK_WRITE_ZEROES_FLAG_UNMAP) { 562 blk_aio_flags |= BDRV_REQ_MAY_UNMAP; 563 } 564 565 block_acct_start(blk_get_stats(s->blk), &req->acct, bytes, 566 BLOCK_ACCT_WRITE); 567 568 blk_aio_pwrite_zeroes(s->blk, sector << BDRV_SECTOR_BITS, 569 bytes, blk_aio_flags, 570 virtio_blk_discard_write_zeroes_complete, req); 571 } else { /* VIRTIO_BLK_T_DISCARD */ 572 /* 573 * The device MUST set the status byte to VIRTIO_BLK_S_UNSUPP for 574 * discard commands if the unmap flag is set. 575 */ 576 if (unlikely(flags & VIRTIO_BLK_WRITE_ZEROES_FLAG_UNMAP)) { 577 err_status = VIRTIO_BLK_S_UNSUPP; 578 goto err; 579 } 580 581 blk_aio_pdiscard(s->blk, sector << BDRV_SECTOR_BITS, bytes, 582 virtio_blk_discard_write_zeroes_complete, req); 583 } 584 585 return VIRTIO_BLK_S_OK; 586 587 err: 588 if (is_write_zeroes) { 589 block_acct_invalid(blk_get_stats(s->blk), BLOCK_ACCT_WRITE); 590 } 591 return err_status; 592 } 593 594 typedef struct ZoneCmdData { 595 VirtIOBlockReq *req; 596 struct iovec *in_iov; 597 unsigned in_num; 598 union { 599 struct { 600 unsigned int nr_zones; 601 BlockZoneDescriptor *zones; 602 } zone_report_data; 603 struct { 604 int64_t offset; 605 } zone_append_data; 606 }; 607 } ZoneCmdData; 608 609 /* 610 * check zoned_request: error checking before issuing requests. If all checks 611 * passed, return true. 612 * append: true if only zone append requests issued. 613 */ 614 static bool check_zoned_request(VirtIOBlock *s, int64_t offset, int64_t len, 615 bool append, uint8_t *status) { 616 BlockDriverState *bs = blk_bs(s->blk); 617 int index; 618 619 if (!virtio_has_feature(s->host_features, VIRTIO_BLK_F_ZONED)) { 620 *status = VIRTIO_BLK_S_UNSUPP; 621 return false; 622 } 623 624 if (offset < 0 || len < 0 || len > (bs->total_sectors << BDRV_SECTOR_BITS) 625 || offset > (bs->total_sectors << BDRV_SECTOR_BITS) - len) { 626 *status = VIRTIO_BLK_S_ZONE_INVALID_CMD; 627 return false; 628 } 629 630 if (append) { 631 if (bs->bl.write_granularity) { 632 if ((offset % bs->bl.write_granularity) != 0) { 633 *status = VIRTIO_BLK_S_ZONE_UNALIGNED_WP; 634 return false; 635 } 636 } 637 638 index = offset / bs->bl.zone_size; 639 if (BDRV_ZT_IS_CONV(bs->wps->wp[index])) { 640 *status = VIRTIO_BLK_S_ZONE_INVALID_CMD; 641 return false; 642 } 643 644 if (len / 512 > bs->bl.max_append_sectors) { 645 if (bs->bl.max_append_sectors == 0) { 646 *status = VIRTIO_BLK_S_UNSUPP; 647 } else { 648 *status = VIRTIO_BLK_S_ZONE_INVALID_CMD; 649 } 650 return false; 651 } 652 } 653 return true; 654 } 655 656 static void virtio_blk_zone_report_complete(void *opaque, int ret) 657 { 658 ZoneCmdData *data = opaque; 659 VirtIOBlockReq *req = data->req; 660 VirtIODevice *vdev = VIRTIO_DEVICE(req->dev); 661 struct iovec *in_iov = data->in_iov; 662 unsigned in_num = data->in_num; 663 int64_t zrp_size, n, j = 0; 664 int64_t nz = data->zone_report_data.nr_zones; 665 int8_t err_status = VIRTIO_BLK_S_OK; 666 struct virtio_blk_zone_report zrp_hdr = (struct virtio_blk_zone_report) { 667 .nr_zones = cpu_to_le64(nz), 668 }; 669 670 trace_virtio_blk_zone_report_complete(vdev, req, nz, ret); 671 if (ret) { 672 err_status = VIRTIO_BLK_S_ZONE_INVALID_CMD; 673 goto out; 674 } 675 676 zrp_size = sizeof(struct virtio_blk_zone_report) 677 + sizeof(struct virtio_blk_zone_descriptor) * nz; 678 n = iov_from_buf(in_iov, in_num, 0, &zrp_hdr, sizeof(zrp_hdr)); 679 if (n != sizeof(zrp_hdr)) { 680 virtio_error(vdev, "Driver provided input buffer that is too small!"); 681 err_status = VIRTIO_BLK_S_ZONE_INVALID_CMD; 682 goto out; 683 } 684 685 for (size_t i = sizeof(zrp_hdr); i < zrp_size; 686 i += sizeof(struct virtio_blk_zone_descriptor), ++j) { 687 struct virtio_blk_zone_descriptor desc = 688 (struct virtio_blk_zone_descriptor) { 689 .z_start = cpu_to_le64(data->zone_report_data.zones[j].start 690 >> BDRV_SECTOR_BITS), 691 .z_cap = cpu_to_le64(data->zone_report_data.zones[j].cap 692 >> BDRV_SECTOR_BITS), 693 .z_wp = cpu_to_le64(data->zone_report_data.zones[j].wp 694 >> BDRV_SECTOR_BITS), 695 }; 696 697 switch (data->zone_report_data.zones[j].type) { 698 case BLK_ZT_CONV: 699 desc.z_type = VIRTIO_BLK_ZT_CONV; 700 break; 701 case BLK_ZT_SWR: 702 desc.z_type = VIRTIO_BLK_ZT_SWR; 703 break; 704 case BLK_ZT_SWP: 705 desc.z_type = VIRTIO_BLK_ZT_SWP; 706 break; 707 default: 708 g_assert_not_reached(); 709 } 710 711 switch (data->zone_report_data.zones[j].state) { 712 case BLK_ZS_RDONLY: 713 desc.z_state = VIRTIO_BLK_ZS_RDONLY; 714 break; 715 case BLK_ZS_OFFLINE: 716 desc.z_state = VIRTIO_BLK_ZS_OFFLINE; 717 break; 718 case BLK_ZS_EMPTY: 719 desc.z_state = VIRTIO_BLK_ZS_EMPTY; 720 break; 721 case BLK_ZS_CLOSED: 722 desc.z_state = VIRTIO_BLK_ZS_CLOSED; 723 break; 724 case BLK_ZS_FULL: 725 desc.z_state = VIRTIO_BLK_ZS_FULL; 726 break; 727 case BLK_ZS_EOPEN: 728 desc.z_state = VIRTIO_BLK_ZS_EOPEN; 729 break; 730 case BLK_ZS_IOPEN: 731 desc.z_state = VIRTIO_BLK_ZS_IOPEN; 732 break; 733 case BLK_ZS_NOT_WP: 734 desc.z_state = VIRTIO_BLK_ZS_NOT_WP; 735 break; 736 default: 737 g_assert_not_reached(); 738 } 739 740 /* TODO: it takes O(n^2) time complexity. Optimizations required. */ 741 n = iov_from_buf(in_iov, in_num, i, &desc, sizeof(desc)); 742 if (n != sizeof(desc)) { 743 virtio_error(vdev, "Driver provided input buffer " 744 "for descriptors that is too small!"); 745 err_status = VIRTIO_BLK_S_ZONE_INVALID_CMD; 746 } 747 } 748 749 out: 750 virtio_blk_req_complete(req, err_status); 751 virtio_blk_free_request(req); 752 g_free(data->zone_report_data.zones); 753 g_free(data); 754 } 755 756 static void virtio_blk_handle_zone_report(VirtIOBlockReq *req, 757 struct iovec *in_iov, 758 unsigned in_num) 759 { 760 VirtIOBlock *s = req->dev; 761 VirtIODevice *vdev = VIRTIO_DEVICE(s); 762 unsigned int nr_zones; 763 ZoneCmdData *data; 764 int64_t zone_size, offset; 765 uint8_t err_status; 766 767 if (req->in_len < sizeof(struct virtio_blk_inhdr) + 768 sizeof(struct virtio_blk_zone_report) + 769 sizeof(struct virtio_blk_zone_descriptor)) { 770 virtio_error(vdev, "in buffer too small for zone report"); 771 err_status = VIRTIO_BLK_S_ZONE_INVALID_CMD; 772 goto out; 773 } 774 775 /* start byte offset of the zone report */ 776 offset = virtio_ldq_p(vdev, &req->out.sector) << BDRV_SECTOR_BITS; 777 if (!check_zoned_request(s, offset, 0, false, &err_status)) { 778 goto out; 779 } 780 nr_zones = (req->in_len - sizeof(struct virtio_blk_inhdr) - 781 sizeof(struct virtio_blk_zone_report)) / 782 sizeof(struct virtio_blk_zone_descriptor); 783 trace_virtio_blk_handle_zone_report(vdev, req, 784 offset >> BDRV_SECTOR_BITS, nr_zones); 785 786 zone_size = sizeof(BlockZoneDescriptor) * nr_zones; 787 data = g_malloc(sizeof(ZoneCmdData)); 788 data->req = req; 789 data->in_iov = in_iov; 790 data->in_num = in_num; 791 data->zone_report_data.nr_zones = nr_zones; 792 data->zone_report_data.zones = g_malloc(zone_size), 793 794 blk_aio_zone_report(s->blk, offset, &data->zone_report_data.nr_zones, 795 data->zone_report_data.zones, 796 virtio_blk_zone_report_complete, data); 797 return; 798 out: 799 virtio_blk_req_complete(req, err_status); 800 virtio_blk_free_request(req); 801 } 802 803 static void virtio_blk_zone_mgmt_complete(void *opaque, int ret) 804 { 805 VirtIOBlockReq *req = opaque; 806 VirtIOBlock *s = req->dev; 807 VirtIODevice *vdev = VIRTIO_DEVICE(s); 808 int8_t err_status = VIRTIO_BLK_S_OK; 809 trace_virtio_blk_zone_mgmt_complete(vdev, req,ret); 810 811 if (ret) { 812 err_status = VIRTIO_BLK_S_ZONE_INVALID_CMD; 813 } 814 815 virtio_blk_req_complete(req, err_status); 816 virtio_blk_free_request(req); 817 } 818 819 static int virtio_blk_handle_zone_mgmt(VirtIOBlockReq *req, BlockZoneOp op) 820 { 821 VirtIOBlock *s = req->dev; 822 VirtIODevice *vdev = VIRTIO_DEVICE(s); 823 BlockDriverState *bs = blk_bs(s->blk); 824 int64_t offset = virtio_ldq_p(vdev, &req->out.sector) << BDRV_SECTOR_BITS; 825 uint64_t len; 826 uint64_t capacity = bs->total_sectors << BDRV_SECTOR_BITS; 827 uint8_t err_status = VIRTIO_BLK_S_OK; 828 829 uint32_t type = virtio_ldl_p(vdev, &req->out.type); 830 if (type == VIRTIO_BLK_T_ZONE_RESET_ALL) { 831 /* Entire drive capacity */ 832 offset = 0; 833 len = capacity; 834 trace_virtio_blk_handle_zone_reset_all(vdev, req, 0, 835 bs->total_sectors); 836 } else { 837 if (bs->bl.zone_size > capacity - offset) { 838 /* The zoned device allows the last smaller zone. */ 839 len = capacity - bs->bl.zone_size * (bs->bl.nr_zones - 1); 840 } else { 841 len = bs->bl.zone_size; 842 } 843 trace_virtio_blk_handle_zone_mgmt(vdev, req, op, 844 offset >> BDRV_SECTOR_BITS, 845 len >> BDRV_SECTOR_BITS); 846 } 847 848 if (!check_zoned_request(s, offset, len, false, &err_status)) { 849 goto out; 850 } 851 852 blk_aio_zone_mgmt(s->blk, op, offset, len, 853 virtio_blk_zone_mgmt_complete, req); 854 855 return 0; 856 out: 857 virtio_blk_req_complete(req, err_status); 858 virtio_blk_free_request(req); 859 return err_status; 860 } 861 862 static void virtio_blk_zone_append_complete(void *opaque, int ret) 863 { 864 ZoneCmdData *data = opaque; 865 VirtIOBlockReq *req = data->req; 866 VirtIODevice *vdev = VIRTIO_DEVICE(req->dev); 867 int64_t append_sector, n; 868 uint8_t err_status = VIRTIO_BLK_S_OK; 869 870 if (ret) { 871 err_status = VIRTIO_BLK_S_ZONE_INVALID_CMD; 872 goto out; 873 } 874 875 virtio_stq_p(vdev, &append_sector, 876 data->zone_append_data.offset >> BDRV_SECTOR_BITS); 877 n = iov_from_buf(data->in_iov, data->in_num, 0, &append_sector, 878 sizeof(append_sector)); 879 if (n != sizeof(append_sector)) { 880 virtio_error(vdev, "Driver provided input buffer less than size of " 881 "append_sector"); 882 err_status = VIRTIO_BLK_S_ZONE_INVALID_CMD; 883 goto out; 884 } 885 trace_virtio_blk_zone_append_complete(vdev, req, append_sector, ret); 886 887 out: 888 virtio_blk_req_complete(req, err_status); 889 virtio_blk_free_request(req); 890 g_free(data); 891 } 892 893 static int virtio_blk_handle_zone_append(VirtIOBlockReq *req, 894 struct iovec *out_iov, 895 struct iovec *in_iov, 896 uint64_t out_num, 897 unsigned in_num) { 898 VirtIOBlock *s = req->dev; 899 VirtIODevice *vdev = VIRTIO_DEVICE(s); 900 uint8_t err_status = VIRTIO_BLK_S_OK; 901 902 int64_t offset = virtio_ldq_p(vdev, &req->out.sector) << BDRV_SECTOR_BITS; 903 int64_t len = iov_size(out_iov, out_num); 904 ZoneCmdData *data; 905 906 trace_virtio_blk_handle_zone_append(vdev, req, offset >> BDRV_SECTOR_BITS); 907 if (!check_zoned_request(s, offset, len, true, &err_status)) { 908 goto out; 909 } 910 911 data = g_malloc(sizeof(ZoneCmdData)); 912 data->req = req; 913 data->in_iov = in_iov; 914 data->in_num = in_num; 915 data->zone_append_data.offset = offset; 916 qemu_iovec_init_external(&req->qiov, out_iov, out_num); 917 918 block_acct_start(blk_get_stats(s->blk), &req->acct, len, 919 BLOCK_ACCT_ZONE_APPEND); 920 921 blk_aio_zone_append(s->blk, &data->zone_append_data.offset, &req->qiov, 0, 922 virtio_blk_zone_append_complete, data); 923 return 0; 924 925 out: 926 virtio_blk_req_complete(req, err_status); 927 virtio_blk_free_request(req); 928 return err_status; 929 } 930 931 static int virtio_blk_handle_request(VirtIOBlockReq *req, MultiReqBuffer *mrb) 932 { 933 uint32_t type; 934 struct iovec *in_iov = req->elem.in_sg; 935 struct iovec *out_iov = req->elem.out_sg; 936 unsigned in_num = req->elem.in_num; 937 unsigned out_num = req->elem.out_num; 938 VirtIOBlock *s = req->dev; 939 VirtIODevice *vdev = VIRTIO_DEVICE(s); 940 941 if (req->elem.out_num < 1 || req->elem.in_num < 1) { 942 virtio_error(vdev, "virtio-blk missing headers"); 943 return -1; 944 } 945 946 if (unlikely(iov_to_buf(out_iov, out_num, 0, &req->out, 947 sizeof(req->out)) != sizeof(req->out))) { 948 virtio_error(vdev, "virtio-blk request outhdr too short"); 949 return -1; 950 } 951 952 iov_discard_front_undoable(&out_iov, &out_num, sizeof(req->out), 953 &req->outhdr_undo); 954 955 if (in_iov[in_num - 1].iov_len < sizeof(struct virtio_blk_inhdr)) { 956 virtio_error(vdev, "virtio-blk request inhdr too short"); 957 iov_discard_undo(&req->outhdr_undo); 958 return -1; 959 } 960 961 /* We always touch the last byte, so just see how big in_iov is. */ 962 req->in_len = iov_size(in_iov, in_num); 963 req->in = (void *)in_iov[in_num - 1].iov_base 964 + in_iov[in_num - 1].iov_len 965 - sizeof(struct virtio_blk_inhdr); 966 iov_discard_back_undoable(in_iov, &in_num, sizeof(struct virtio_blk_inhdr), 967 &req->inhdr_undo); 968 969 type = virtio_ldl_p(vdev, &req->out.type); 970 971 /* VIRTIO_BLK_T_OUT defines the command direction. VIRTIO_BLK_T_BARRIER 972 * is an optional flag. Although a guest should not send this flag if 973 * not negotiated we ignored it in the past. So keep ignoring it. */ 974 switch (type & ~(VIRTIO_BLK_T_OUT | VIRTIO_BLK_T_BARRIER)) { 975 case VIRTIO_BLK_T_IN: 976 { 977 bool is_write = type & VIRTIO_BLK_T_OUT; 978 req->sector_num = virtio_ldq_p(vdev, &req->out.sector); 979 980 if (is_write) { 981 qemu_iovec_init_external(&req->qiov, out_iov, out_num); 982 trace_virtio_blk_handle_write(vdev, req, req->sector_num, 983 req->qiov.size / BDRV_SECTOR_SIZE); 984 } else { 985 qemu_iovec_init_external(&req->qiov, in_iov, in_num); 986 trace_virtio_blk_handle_read(vdev, req, req->sector_num, 987 req->qiov.size / BDRV_SECTOR_SIZE); 988 } 989 990 if (!virtio_blk_sect_range_ok(s, req->sector_num, req->qiov.size)) { 991 virtio_blk_req_complete(req, VIRTIO_BLK_S_IOERR); 992 block_acct_invalid(blk_get_stats(s->blk), 993 is_write ? BLOCK_ACCT_WRITE : BLOCK_ACCT_READ); 994 virtio_blk_free_request(req); 995 return 0; 996 } 997 998 block_acct_start(blk_get_stats(s->blk), &req->acct, req->qiov.size, 999 is_write ? BLOCK_ACCT_WRITE : BLOCK_ACCT_READ); 1000 1001 /* merge would exceed maximum number of requests or IO direction 1002 * changes */ 1003 if (mrb->num_reqs > 0 && (mrb->num_reqs == VIRTIO_BLK_MAX_MERGE_REQS || 1004 is_write != mrb->is_write || 1005 !s->conf.request_merging)) { 1006 virtio_blk_submit_multireq(s, mrb); 1007 } 1008 1009 assert(mrb->num_reqs < VIRTIO_BLK_MAX_MERGE_REQS); 1010 mrb->reqs[mrb->num_reqs++] = req; 1011 mrb->is_write = is_write; 1012 break; 1013 } 1014 case VIRTIO_BLK_T_FLUSH: 1015 virtio_blk_handle_flush(req, mrb); 1016 break; 1017 case VIRTIO_BLK_T_ZONE_REPORT: 1018 virtio_blk_handle_zone_report(req, in_iov, in_num); 1019 break; 1020 case VIRTIO_BLK_T_ZONE_OPEN: 1021 virtio_blk_handle_zone_mgmt(req, BLK_ZO_OPEN); 1022 break; 1023 case VIRTIO_BLK_T_ZONE_CLOSE: 1024 virtio_blk_handle_zone_mgmt(req, BLK_ZO_CLOSE); 1025 break; 1026 case VIRTIO_BLK_T_ZONE_FINISH: 1027 virtio_blk_handle_zone_mgmt(req, BLK_ZO_FINISH); 1028 break; 1029 case VIRTIO_BLK_T_ZONE_RESET: 1030 virtio_blk_handle_zone_mgmt(req, BLK_ZO_RESET); 1031 break; 1032 case VIRTIO_BLK_T_ZONE_RESET_ALL: 1033 virtio_blk_handle_zone_mgmt(req, BLK_ZO_RESET); 1034 break; 1035 case VIRTIO_BLK_T_SCSI_CMD: 1036 virtio_blk_handle_scsi(req); 1037 break; 1038 case VIRTIO_BLK_T_GET_ID: 1039 { 1040 /* 1041 * NB: per existing s/n string convention the string is 1042 * terminated by '\0' only when shorter than buffer. 1043 */ 1044 const char *serial = s->conf.serial ? s->conf.serial : ""; 1045 size_t size = MIN(strlen(serial) + 1, 1046 MIN(iov_size(in_iov, in_num), 1047 VIRTIO_BLK_ID_BYTES)); 1048 iov_from_buf(in_iov, in_num, 0, serial, size); 1049 virtio_blk_req_complete(req, VIRTIO_BLK_S_OK); 1050 virtio_blk_free_request(req); 1051 break; 1052 } 1053 case VIRTIO_BLK_T_ZONE_APPEND & ~VIRTIO_BLK_T_OUT: 1054 /* 1055 * Passing out_iov/out_num and in_iov/in_num is not safe 1056 * to access req->elem.out_sg directly because it may be 1057 * modified by virtio_blk_handle_request(). 1058 */ 1059 virtio_blk_handle_zone_append(req, out_iov, in_iov, out_num, in_num); 1060 break; 1061 /* 1062 * VIRTIO_BLK_T_DISCARD and VIRTIO_BLK_T_WRITE_ZEROES are defined with 1063 * VIRTIO_BLK_T_OUT flag set. We masked this flag in the switch statement, 1064 * so we must mask it for these requests, then we will check if it is set. 1065 */ 1066 case VIRTIO_BLK_T_DISCARD & ~VIRTIO_BLK_T_OUT: 1067 case VIRTIO_BLK_T_WRITE_ZEROES & ~VIRTIO_BLK_T_OUT: 1068 { 1069 struct virtio_blk_discard_write_zeroes dwz_hdr; 1070 size_t out_len = iov_size(out_iov, out_num); 1071 bool is_write_zeroes = (type & ~VIRTIO_BLK_T_BARRIER) == 1072 VIRTIO_BLK_T_WRITE_ZEROES; 1073 uint8_t err_status; 1074 1075 /* 1076 * Unsupported if VIRTIO_BLK_T_OUT is not set or the request contains 1077 * more than one segment. 1078 */ 1079 if (unlikely(!(type & VIRTIO_BLK_T_OUT) || 1080 out_len > sizeof(dwz_hdr))) { 1081 virtio_blk_req_complete(req, VIRTIO_BLK_S_UNSUPP); 1082 virtio_blk_free_request(req); 1083 return 0; 1084 } 1085 1086 if (unlikely(iov_to_buf(out_iov, out_num, 0, &dwz_hdr, 1087 sizeof(dwz_hdr)) != sizeof(dwz_hdr))) { 1088 iov_discard_undo(&req->inhdr_undo); 1089 iov_discard_undo(&req->outhdr_undo); 1090 virtio_error(vdev, "virtio-blk discard/write_zeroes header" 1091 " too short"); 1092 return -1; 1093 } 1094 1095 err_status = virtio_blk_handle_discard_write_zeroes(req, &dwz_hdr, 1096 is_write_zeroes); 1097 if (err_status != VIRTIO_BLK_S_OK) { 1098 virtio_blk_req_complete(req, err_status); 1099 virtio_blk_free_request(req); 1100 } 1101 1102 break; 1103 } 1104 default: 1105 virtio_blk_req_complete(req, VIRTIO_BLK_S_UNSUPP); 1106 virtio_blk_free_request(req); 1107 } 1108 return 0; 1109 } 1110 1111 void virtio_blk_handle_vq(VirtIOBlock *s, VirtQueue *vq) 1112 { 1113 VirtIOBlockReq *req; 1114 MultiReqBuffer mrb = {}; 1115 bool suppress_notifications = virtio_queue_get_notification(vq); 1116 1117 defer_call_begin(); 1118 1119 do { 1120 if (suppress_notifications) { 1121 virtio_queue_set_notification(vq, 0); 1122 } 1123 1124 while ((req = virtio_blk_get_request(s, vq))) { 1125 if (virtio_blk_handle_request(req, &mrb)) { 1126 virtqueue_detach_element(req->vq, &req->elem, 0); 1127 virtio_blk_free_request(req); 1128 break; 1129 } 1130 } 1131 1132 if (suppress_notifications) { 1133 virtio_queue_set_notification(vq, 1); 1134 } 1135 } while (!virtio_queue_empty(vq)); 1136 1137 if (mrb.num_reqs) { 1138 virtio_blk_submit_multireq(s, &mrb); 1139 } 1140 1141 defer_call_end(); 1142 } 1143 1144 static void virtio_blk_handle_output(VirtIODevice *vdev, VirtQueue *vq) 1145 { 1146 VirtIOBlock *s = (VirtIOBlock *)vdev; 1147 1148 if (!s->ioeventfd_disabled && !s->ioeventfd_started) { 1149 /* Some guests kick before setting VIRTIO_CONFIG_S_DRIVER_OK so start 1150 * ioeventfd here instead of waiting for .set_status(). 1151 */ 1152 virtio_device_start_ioeventfd(vdev); 1153 if (!s->ioeventfd_disabled) { 1154 return; 1155 } 1156 } 1157 1158 virtio_blk_handle_vq(s, vq); 1159 } 1160 1161 static void virtio_blk_dma_restart_bh(void *opaque) 1162 { 1163 VirtIOBlockReq *req = opaque; 1164 VirtIOBlock *s = req->dev; /* we're called with at least one request */ 1165 1166 MultiReqBuffer mrb = {}; 1167 1168 while (req) { 1169 VirtIOBlockReq *next = req->next; 1170 if (virtio_blk_handle_request(req, &mrb)) { 1171 /* Device is now broken and won't do any processing until it gets 1172 * reset. Already queued requests will be lost: let's purge them. 1173 */ 1174 while (req) { 1175 next = req->next; 1176 virtqueue_detach_element(req->vq, &req->elem, 0); 1177 virtio_blk_free_request(req); 1178 req = next; 1179 } 1180 break; 1181 } 1182 req = next; 1183 } 1184 1185 if (mrb.num_reqs) { 1186 virtio_blk_submit_multireq(s, &mrb); 1187 } 1188 1189 /* Paired with inc in virtio_blk_dma_restart_cb() */ 1190 blk_dec_in_flight(s->conf.conf.blk); 1191 } 1192 1193 static void virtio_blk_dma_restart_cb(void *opaque, bool running, 1194 RunState state) 1195 { 1196 VirtIOBlock *s = opaque; 1197 uint16_t num_queues = s->conf.num_queues; 1198 g_autofree VirtIOBlockReq **vq_rq = NULL; 1199 VirtIOBlockReq *rq; 1200 1201 if (!running) { 1202 return; 1203 } 1204 1205 /* Split the device-wide s->rq request list into per-vq request lists */ 1206 vq_rq = g_new0(VirtIOBlockReq *, num_queues); 1207 1208 WITH_QEMU_LOCK_GUARD(&s->rq_lock) { 1209 rq = s->rq; 1210 s->rq = NULL; 1211 } 1212 1213 while (rq) { 1214 VirtIOBlockReq *next = rq->next; 1215 uint16_t idx = virtio_get_queue_index(rq->vq); 1216 1217 /* Only num_queues vqs were created so vq_rq[idx] is within bounds */ 1218 assert(idx < num_queues); 1219 rq->next = vq_rq[idx]; 1220 vq_rq[idx] = rq; 1221 rq = next; 1222 } 1223 1224 /* Schedule a BH to submit the requests in each vq's AioContext */ 1225 for (uint16_t i = 0; i < num_queues; i++) { 1226 if (!vq_rq[i]) { 1227 continue; 1228 } 1229 1230 /* Paired with dec in virtio_blk_dma_restart_bh() */ 1231 blk_inc_in_flight(s->conf.conf.blk); 1232 1233 aio_bh_schedule_oneshot(s->vq_aio_context[i], 1234 virtio_blk_dma_restart_bh, 1235 vq_rq[i]); 1236 } 1237 } 1238 1239 static void virtio_blk_reset(VirtIODevice *vdev) 1240 { 1241 VirtIOBlock *s = VIRTIO_BLK(vdev); 1242 VirtIOBlockReq *req; 1243 1244 /* Dataplane has stopped... */ 1245 assert(!s->ioeventfd_started); 1246 1247 /* ...but requests may still be in flight. */ 1248 blk_drain(s->blk); 1249 1250 /* We drop queued requests after blk_drain() because blk_drain() itself can 1251 * produce them. */ 1252 WITH_QEMU_LOCK_GUARD(&s->rq_lock) { 1253 while (s->rq) { 1254 req = s->rq; 1255 s->rq = req->next; 1256 1257 /* No other threads can access req->vq here */ 1258 virtqueue_detach_element(req->vq, &req->elem, 0); 1259 1260 virtio_blk_free_request(req); 1261 } 1262 } 1263 1264 blk_set_enable_write_cache(s->blk, s->original_wce); 1265 } 1266 1267 /* coalesce internal state, copy to pci i/o region 0 1268 */ 1269 static void virtio_blk_update_config(VirtIODevice *vdev, uint8_t *config) 1270 { 1271 VirtIOBlock *s = VIRTIO_BLK(vdev); 1272 BlockConf *conf = &s->conf.conf; 1273 BlockDriverState *bs = blk_bs(s->blk); 1274 struct virtio_blk_config blkcfg; 1275 uint64_t capacity; 1276 int64_t length; 1277 int blk_size = conf->logical_block_size; 1278 1279 blk_get_geometry(s->blk, &capacity); 1280 memset(&blkcfg, 0, sizeof(blkcfg)); 1281 virtio_stq_p(vdev, &blkcfg.capacity, capacity); 1282 virtio_stl_p(vdev, &blkcfg.seg_max, 1283 s->conf.seg_max_adjust ? s->conf.queue_size - 2 : 128 - 2); 1284 virtio_stw_p(vdev, &blkcfg.geometry.cylinders, conf->cyls); 1285 virtio_stl_p(vdev, &blkcfg.blk_size, blk_size); 1286 virtio_stw_p(vdev, &blkcfg.min_io_size, conf->min_io_size / blk_size); 1287 virtio_stl_p(vdev, &blkcfg.opt_io_size, conf->opt_io_size / blk_size); 1288 blkcfg.geometry.heads = conf->heads; 1289 /* 1290 * We must ensure that the block device capacity is a multiple of 1291 * the logical block size. If that is not the case, let's use 1292 * sector_mask to adopt the geometry to have a correct picture. 1293 * For those devices where the capacity is ok for the given geometry 1294 * we don't touch the sector value of the geometry, since some devices 1295 * (like s390 dasd) need a specific value. Here the capacity is already 1296 * cyls*heads*secs*blk_size and the sector value is not block size 1297 * divided by 512 - instead it is the amount of blk_size blocks 1298 * per track (cylinder). 1299 */ 1300 length = blk_getlength(s->blk); 1301 if (length > 0 && length / conf->heads / conf->secs % blk_size) { 1302 blkcfg.geometry.sectors = conf->secs & ~s->sector_mask; 1303 } else { 1304 blkcfg.geometry.sectors = conf->secs; 1305 } 1306 blkcfg.size_max = 0; 1307 blkcfg.physical_block_exp = get_physical_block_exp(conf); 1308 blkcfg.alignment_offset = 0; 1309 blkcfg.wce = blk_enable_write_cache(s->blk); 1310 virtio_stw_p(vdev, &blkcfg.num_queues, s->conf.num_queues); 1311 if (virtio_has_feature(s->host_features, VIRTIO_BLK_F_DISCARD)) { 1312 uint32_t discard_granularity = conf->discard_granularity; 1313 if (discard_granularity == -1 || !s->conf.report_discard_granularity) { 1314 discard_granularity = blk_size; 1315 } 1316 virtio_stl_p(vdev, &blkcfg.max_discard_sectors, 1317 s->conf.max_discard_sectors); 1318 virtio_stl_p(vdev, &blkcfg.discard_sector_alignment, 1319 discard_granularity >> BDRV_SECTOR_BITS); 1320 /* 1321 * We support only one segment per request since multiple segments 1322 * are not widely used and there are no userspace APIs that allow 1323 * applications to submit multiple segments in a single call. 1324 */ 1325 virtio_stl_p(vdev, &blkcfg.max_discard_seg, 1); 1326 } 1327 if (virtio_has_feature(s->host_features, VIRTIO_BLK_F_WRITE_ZEROES)) { 1328 virtio_stl_p(vdev, &blkcfg.max_write_zeroes_sectors, 1329 s->conf.max_write_zeroes_sectors); 1330 blkcfg.write_zeroes_may_unmap = 1; 1331 virtio_stl_p(vdev, &blkcfg.max_write_zeroes_seg, 1); 1332 } 1333 if (bs->bl.zoned != BLK_Z_NONE) { 1334 switch (bs->bl.zoned) { 1335 case BLK_Z_HM: 1336 blkcfg.zoned.model = VIRTIO_BLK_Z_HM; 1337 break; 1338 case BLK_Z_HA: 1339 blkcfg.zoned.model = VIRTIO_BLK_Z_HA; 1340 break; 1341 default: 1342 g_assert_not_reached(); 1343 } 1344 1345 virtio_stl_p(vdev, &blkcfg.zoned.zone_sectors, 1346 bs->bl.zone_size / 512); 1347 virtio_stl_p(vdev, &blkcfg.zoned.max_active_zones, 1348 bs->bl.max_active_zones); 1349 virtio_stl_p(vdev, &blkcfg.zoned.max_open_zones, 1350 bs->bl.max_open_zones); 1351 virtio_stl_p(vdev, &blkcfg.zoned.write_granularity, blk_size); 1352 virtio_stl_p(vdev, &blkcfg.zoned.max_append_sectors, 1353 bs->bl.max_append_sectors); 1354 } else { 1355 blkcfg.zoned.model = VIRTIO_BLK_Z_NONE; 1356 } 1357 memcpy(config, &blkcfg, s->config_size); 1358 } 1359 1360 static void virtio_blk_set_config(VirtIODevice *vdev, const uint8_t *config) 1361 { 1362 VirtIOBlock *s = VIRTIO_BLK(vdev); 1363 struct virtio_blk_config blkcfg; 1364 1365 memcpy(&blkcfg, config, s->config_size); 1366 1367 blk_set_enable_write_cache(s->blk, blkcfg.wce != 0); 1368 } 1369 1370 static uint64_t virtio_blk_get_features(VirtIODevice *vdev, uint64_t features, 1371 Error **errp) 1372 { 1373 VirtIOBlock *s = VIRTIO_BLK(vdev); 1374 1375 /* Firstly sync all virtio-blk possible supported features */ 1376 features |= s->host_features; 1377 1378 virtio_add_feature(&features, VIRTIO_BLK_F_SEG_MAX); 1379 virtio_add_feature(&features, VIRTIO_BLK_F_GEOMETRY); 1380 virtio_add_feature(&features, VIRTIO_BLK_F_TOPOLOGY); 1381 virtio_add_feature(&features, VIRTIO_BLK_F_BLK_SIZE); 1382 if (virtio_has_feature(features, VIRTIO_F_VERSION_1)) { 1383 if (virtio_has_feature(s->host_features, VIRTIO_BLK_F_SCSI)) { 1384 error_setg(errp, "Please set scsi=off for virtio-blk devices in order to use virtio 1.0"); 1385 return 0; 1386 } 1387 } else { 1388 virtio_clear_feature(&features, VIRTIO_F_ANY_LAYOUT); 1389 virtio_add_feature(&features, VIRTIO_BLK_F_SCSI); 1390 } 1391 1392 if (blk_enable_write_cache(s->blk) || 1393 (s->conf.x_enable_wce_if_config_wce && 1394 virtio_has_feature(features, VIRTIO_BLK_F_CONFIG_WCE))) { 1395 virtio_add_feature(&features, VIRTIO_BLK_F_WCE); 1396 } 1397 if (!blk_is_writable(s->blk)) { 1398 virtio_add_feature(&features, VIRTIO_BLK_F_RO); 1399 } 1400 if (s->conf.num_queues > 1) { 1401 virtio_add_feature(&features, VIRTIO_BLK_F_MQ); 1402 } 1403 1404 return features; 1405 } 1406 1407 static void virtio_blk_set_status(VirtIODevice *vdev, uint8_t status) 1408 { 1409 VirtIOBlock *s = VIRTIO_BLK(vdev); 1410 1411 if (!(status & (VIRTIO_CONFIG_S_DRIVER | VIRTIO_CONFIG_S_DRIVER_OK))) { 1412 assert(!s->ioeventfd_started); 1413 } 1414 1415 if (!(status & VIRTIO_CONFIG_S_DRIVER_OK)) { 1416 return; 1417 } 1418 1419 /* A guest that supports VIRTIO_BLK_F_CONFIG_WCE must be able to send 1420 * cache flushes. Thus, the "auto writethrough" behavior is never 1421 * necessary for guests that support the VIRTIO_BLK_F_CONFIG_WCE feature. 1422 * Leaving it enabled would break the following sequence: 1423 * 1424 * Guest started with "-drive cache=writethrough" 1425 * Guest sets status to 0 1426 * Guest sets DRIVER bit in status field 1427 * Guest reads host features (WCE=0, CONFIG_WCE=1) 1428 * Guest writes guest features (WCE=0, CONFIG_WCE=1) 1429 * Guest writes 1 to the WCE configuration field (writeback mode) 1430 * Guest sets DRIVER_OK bit in status field 1431 * 1432 * s->blk would erroneously be placed in writethrough mode. 1433 */ 1434 if (!virtio_vdev_has_feature(vdev, VIRTIO_BLK_F_CONFIG_WCE)) { 1435 blk_set_enable_write_cache(s->blk, 1436 virtio_vdev_has_feature(vdev, 1437 VIRTIO_BLK_F_WCE)); 1438 } 1439 } 1440 1441 static void virtio_blk_save_device(VirtIODevice *vdev, QEMUFile *f) 1442 { 1443 VirtIOBlock *s = VIRTIO_BLK(vdev); 1444 1445 WITH_QEMU_LOCK_GUARD(&s->rq_lock) { 1446 VirtIOBlockReq *req = s->rq; 1447 1448 while (req) { 1449 qemu_put_sbyte(f, 1); 1450 1451 if (s->conf.num_queues > 1) { 1452 qemu_put_be32(f, virtio_get_queue_index(req->vq)); 1453 } 1454 1455 qemu_put_virtqueue_element(vdev, f, &req->elem); 1456 req = req->next; 1457 } 1458 } 1459 1460 qemu_put_sbyte(f, 0); 1461 } 1462 1463 static int virtio_blk_load_device(VirtIODevice *vdev, QEMUFile *f, 1464 int version_id) 1465 { 1466 VirtIOBlock *s = VIRTIO_BLK(vdev); 1467 1468 while (qemu_get_sbyte(f)) { 1469 unsigned nvqs = s->conf.num_queues; 1470 unsigned vq_idx = 0; 1471 VirtIOBlockReq *req; 1472 1473 if (nvqs > 1) { 1474 vq_idx = qemu_get_be32(f); 1475 1476 if (vq_idx >= nvqs) { 1477 error_report("Invalid virtqueue index in request list: %#x", 1478 vq_idx); 1479 return -EINVAL; 1480 } 1481 } 1482 1483 req = qemu_get_virtqueue_element(vdev, f, sizeof(VirtIOBlockReq)); 1484 virtio_blk_init_request(s, virtio_get_queue(vdev, vq_idx), req); 1485 1486 WITH_QEMU_LOCK_GUARD(&s->rq_lock) { 1487 req->next = s->rq; 1488 s->rq = req; 1489 } 1490 } 1491 1492 return 0; 1493 } 1494 1495 static void virtio_resize_cb(void *opaque) 1496 { 1497 VirtIODevice *vdev = opaque; 1498 1499 assert(qemu_get_current_aio_context() == qemu_get_aio_context()); 1500 virtio_notify_config(vdev); 1501 } 1502 1503 static void virtio_blk_resize(void *opaque) 1504 { 1505 VirtIODevice *vdev = VIRTIO_DEVICE(opaque); 1506 1507 /* 1508 * virtio_notify_config() needs to acquire the BQL, 1509 * so it can't be called from an iothread. Instead, schedule 1510 * it to be run in the main context BH. 1511 */ 1512 aio_bh_schedule_oneshot(qemu_get_aio_context(), virtio_resize_cb, vdev); 1513 } 1514 1515 static void virtio_blk_ioeventfd_detach(VirtIOBlock *s) 1516 { 1517 VirtIODevice *vdev = VIRTIO_DEVICE(s); 1518 1519 for (uint16_t i = 0; i < s->conf.num_queues; i++) { 1520 VirtQueue *vq = virtio_get_queue(vdev, i); 1521 virtio_queue_aio_detach_host_notifier(vq, s->vq_aio_context[i]); 1522 } 1523 } 1524 1525 static void virtio_blk_ioeventfd_attach(VirtIOBlock *s) 1526 { 1527 VirtIODevice *vdev = VIRTIO_DEVICE(s); 1528 1529 for (uint16_t i = 0; i < s->conf.num_queues; i++) { 1530 VirtQueue *vq = virtio_get_queue(vdev, i); 1531 virtio_queue_aio_attach_host_notifier(vq, s->vq_aio_context[i]); 1532 } 1533 } 1534 1535 /* Suspend virtqueue ioeventfd processing during drain */ 1536 static void virtio_blk_drained_begin(void *opaque) 1537 { 1538 VirtIOBlock *s = opaque; 1539 1540 if (s->ioeventfd_started) { 1541 virtio_blk_ioeventfd_detach(s); 1542 } 1543 } 1544 1545 /* Resume virtqueue ioeventfd processing after drain */ 1546 static void virtio_blk_drained_end(void *opaque) 1547 { 1548 VirtIOBlock *s = opaque; 1549 1550 if (s->ioeventfd_started) { 1551 virtio_blk_ioeventfd_attach(s); 1552 } 1553 } 1554 1555 static const BlockDevOps virtio_block_ops = { 1556 .resize_cb = virtio_blk_resize, 1557 .drained_begin = virtio_blk_drained_begin, 1558 .drained_end = virtio_blk_drained_end, 1559 }; 1560 1561 static bool 1562 validate_iothread_vq_mapping_list(IOThreadVirtQueueMappingList *list, 1563 uint16_t num_queues, Error **errp) 1564 { 1565 g_autofree unsigned long *vqs = bitmap_new(num_queues); 1566 g_autoptr(GHashTable) iothreads = 1567 g_hash_table_new(g_str_hash, g_str_equal); 1568 1569 for (IOThreadVirtQueueMappingList *node = list; node; node = node->next) { 1570 const char *name = node->value->iothread; 1571 uint16List *vq; 1572 1573 if (!iothread_by_id(name)) { 1574 error_setg(errp, "IOThread \"%s\" object does not exist", name); 1575 return false; 1576 } 1577 1578 if (!g_hash_table_add(iothreads, (gpointer)name)) { 1579 error_setg(errp, 1580 "duplicate IOThread name \"%s\" in iothread-vq-mapping", 1581 name); 1582 return false; 1583 } 1584 1585 if (node != list) { 1586 if (!!node->value->vqs != !!list->value->vqs) { 1587 error_setg(errp, "either all items in iothread-vq-mapping " 1588 "must have vqs or none of them must have it"); 1589 return false; 1590 } 1591 } 1592 1593 for (vq = node->value->vqs; vq; vq = vq->next) { 1594 if (vq->value >= num_queues) { 1595 error_setg(errp, "vq index %u for IOThread \"%s\" must be " 1596 "less than num_queues %u in iothread-vq-mapping", 1597 vq->value, name, num_queues); 1598 return false; 1599 } 1600 1601 if (test_and_set_bit(vq->value, vqs)) { 1602 error_setg(errp, "cannot assign vq %u to IOThread \"%s\" " 1603 "because it is already assigned", vq->value, name); 1604 return false; 1605 } 1606 } 1607 } 1608 1609 if (list->value->vqs) { 1610 for (uint16_t i = 0; i < num_queues; i++) { 1611 if (!test_bit(i, vqs)) { 1612 error_setg(errp, 1613 "missing vq %u IOThread assignment in iothread-vq-mapping", 1614 i); 1615 return false; 1616 } 1617 } 1618 } 1619 1620 return true; 1621 } 1622 1623 /** 1624 * apply_iothread_vq_mapping: 1625 * @iothread_vq_mapping_list: The mapping of virtqueues to IOThreads. 1626 * @vq_aio_context: The array of AioContext pointers to fill in. 1627 * @num_queues: The length of @vq_aio_context. 1628 * @errp: If an error occurs, a pointer to the area to store the error. 1629 * 1630 * Fill in the AioContext for each virtqueue in the @vq_aio_context array given 1631 * the iothread-vq-mapping parameter in @iothread_vq_mapping_list. 1632 * 1633 * Returns: %true on success, %false on failure. 1634 **/ 1635 static bool apply_iothread_vq_mapping( 1636 IOThreadVirtQueueMappingList *iothread_vq_mapping_list, 1637 AioContext **vq_aio_context, 1638 uint16_t num_queues, 1639 Error **errp) 1640 { 1641 IOThreadVirtQueueMappingList *node; 1642 size_t num_iothreads = 0; 1643 size_t cur_iothread = 0; 1644 1645 if (!validate_iothread_vq_mapping_list(iothread_vq_mapping_list, 1646 num_queues, errp)) { 1647 return false; 1648 } 1649 1650 for (node = iothread_vq_mapping_list; node; node = node->next) { 1651 num_iothreads++; 1652 } 1653 1654 for (node = iothread_vq_mapping_list; node; node = node->next) { 1655 IOThread *iothread = iothread_by_id(node->value->iothread); 1656 AioContext *ctx = iothread_get_aio_context(iothread); 1657 1658 /* Released in virtio_blk_vq_aio_context_cleanup() */ 1659 object_ref(OBJECT(iothread)); 1660 1661 if (node->value->vqs) { 1662 uint16List *vq; 1663 1664 /* Explicit vq:IOThread assignment */ 1665 for (vq = node->value->vqs; vq; vq = vq->next) { 1666 assert(vq->value < num_queues); 1667 vq_aio_context[vq->value] = ctx; 1668 } 1669 } else { 1670 /* Round-robin vq:IOThread assignment */ 1671 for (unsigned i = cur_iothread; i < num_queues; 1672 i += num_iothreads) { 1673 vq_aio_context[i] = ctx; 1674 } 1675 } 1676 1677 cur_iothread++; 1678 } 1679 1680 return true; 1681 } 1682 1683 /* Context: BQL held */ 1684 static bool virtio_blk_vq_aio_context_init(VirtIOBlock *s, Error **errp) 1685 { 1686 ERRP_GUARD(); 1687 VirtIODevice *vdev = VIRTIO_DEVICE(s); 1688 VirtIOBlkConf *conf = &s->conf; 1689 BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(vdev))); 1690 VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(qbus); 1691 1692 if (conf->iothread && conf->iothread_vq_mapping_list) { 1693 error_setg(errp, 1694 "iothread and iothread-vq-mapping properties cannot be set " 1695 "at the same time"); 1696 return false; 1697 } 1698 1699 if (conf->iothread || conf->iothread_vq_mapping_list) { 1700 if (!k->set_guest_notifiers || !k->ioeventfd_assign) { 1701 error_setg(errp, 1702 "device is incompatible with iothread " 1703 "(transport does not support notifiers)"); 1704 return false; 1705 } 1706 if (!virtio_device_ioeventfd_enabled(vdev)) { 1707 error_setg(errp, "ioeventfd is required for iothread"); 1708 return false; 1709 } 1710 1711 /* 1712 * If ioeventfd is (re-)enabled while the guest is running there could 1713 * be block jobs that can conflict. 1714 */ 1715 if (blk_op_is_blocked(conf->conf.blk, BLOCK_OP_TYPE_DATAPLANE, errp)) { 1716 error_prepend(errp, "cannot start virtio-blk ioeventfd: "); 1717 return false; 1718 } 1719 } 1720 1721 s->vq_aio_context = g_new(AioContext *, conf->num_queues); 1722 1723 if (conf->iothread_vq_mapping_list) { 1724 if (!apply_iothread_vq_mapping(conf->iothread_vq_mapping_list, 1725 s->vq_aio_context, 1726 conf->num_queues, 1727 errp)) { 1728 g_free(s->vq_aio_context); 1729 s->vq_aio_context = NULL; 1730 return false; 1731 } 1732 } else if (conf->iothread) { 1733 AioContext *ctx = iothread_get_aio_context(conf->iothread); 1734 for (unsigned i = 0; i < conf->num_queues; i++) { 1735 s->vq_aio_context[i] = ctx; 1736 } 1737 1738 /* Released in virtio_blk_vq_aio_context_cleanup() */ 1739 object_ref(OBJECT(conf->iothread)); 1740 } else { 1741 AioContext *ctx = qemu_get_aio_context(); 1742 for (unsigned i = 0; i < conf->num_queues; i++) { 1743 s->vq_aio_context[i] = ctx; 1744 } 1745 } 1746 1747 return true; 1748 } 1749 1750 /* Context: BQL held */ 1751 static void virtio_blk_vq_aio_context_cleanup(VirtIOBlock *s) 1752 { 1753 VirtIOBlkConf *conf = &s->conf; 1754 1755 assert(!s->ioeventfd_started); 1756 1757 if (conf->iothread_vq_mapping_list) { 1758 IOThreadVirtQueueMappingList *node; 1759 1760 for (node = conf->iothread_vq_mapping_list; node; node = node->next) { 1761 IOThread *iothread = iothread_by_id(node->value->iothread); 1762 object_unref(OBJECT(iothread)); 1763 } 1764 } 1765 1766 if (conf->iothread) { 1767 object_unref(OBJECT(conf->iothread)); 1768 } 1769 1770 g_free(s->vq_aio_context); 1771 s->vq_aio_context = NULL; 1772 } 1773 1774 /* Context: BQL held */ 1775 static int virtio_blk_start_ioeventfd(VirtIODevice *vdev) 1776 { 1777 VirtIOBlock *s = VIRTIO_BLK(vdev); 1778 BusState *qbus = BUS(qdev_get_parent_bus(DEVICE(s))); 1779 VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(qbus); 1780 unsigned i; 1781 unsigned nvqs = s->conf.num_queues; 1782 Error *local_err = NULL; 1783 int r; 1784 1785 if (s->ioeventfd_started || s->ioeventfd_starting) { 1786 return 0; 1787 } 1788 1789 s->ioeventfd_starting = true; 1790 1791 /* Set up guest notifier (irq) */ 1792 r = k->set_guest_notifiers(qbus->parent, nvqs, true); 1793 if (r != 0) { 1794 error_report("virtio-blk failed to set guest notifier (%d), " 1795 "ensure -accel kvm is set.", r); 1796 goto fail_guest_notifiers; 1797 } 1798 1799 /* 1800 * Batch all the host notifiers in a single transaction to avoid 1801 * quadratic time complexity in address_space_update_ioeventfds(). 1802 */ 1803 memory_region_transaction_begin(); 1804 1805 /* Set up virtqueue notify */ 1806 for (i = 0; i < nvqs; i++) { 1807 r = virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), i, true); 1808 if (r != 0) { 1809 int j = i; 1810 1811 fprintf(stderr, "virtio-blk failed to set host notifier (%d)\n", r); 1812 while (i--) { 1813 virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), i, false); 1814 } 1815 1816 /* 1817 * The transaction expects the ioeventfds to be open when it 1818 * commits. Do it now, before the cleanup loop. 1819 */ 1820 memory_region_transaction_commit(); 1821 1822 while (j--) { 1823 virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), j); 1824 } 1825 goto fail_host_notifiers; 1826 } 1827 } 1828 1829 memory_region_transaction_commit(); 1830 1831 /* 1832 * Try to change the AioContext so that block jobs and other operations can 1833 * co-locate their activity in the same AioContext. If it fails, nevermind. 1834 */ 1835 assert(nvqs > 0); /* enforced during ->realize() */ 1836 r = blk_set_aio_context(s->conf.conf.blk, s->vq_aio_context[0], 1837 &local_err); 1838 if (r < 0) { 1839 warn_report_err(local_err); 1840 } 1841 1842 /* 1843 * These fields must be visible to the IOThread when it processes the 1844 * virtqueue, otherwise it will think ioeventfd has not started yet. 1845 * 1846 * Make sure ->ioeventfd_started is false when blk_set_aio_context() is 1847 * called above so that draining does not cause the host notifier to be 1848 * detached/attached prematurely. 1849 */ 1850 s->ioeventfd_starting = false; 1851 s->ioeventfd_started = true; 1852 smp_wmb(); /* paired with aio_notify_accept() on the read side */ 1853 1854 /* 1855 * Get this show started by hooking up our callbacks. If drained now, 1856 * virtio_blk_drained_end() will do this later. 1857 * Attaching the notifier also kicks the virtqueues, processing any requests 1858 * they may already have. 1859 */ 1860 if (!blk_in_drain(s->conf.conf.blk)) { 1861 virtio_blk_ioeventfd_attach(s); 1862 } 1863 return 0; 1864 1865 fail_host_notifiers: 1866 k->set_guest_notifiers(qbus->parent, nvqs, false); 1867 fail_guest_notifiers: 1868 s->ioeventfd_disabled = true; 1869 s->ioeventfd_starting = false; 1870 return -ENOSYS; 1871 } 1872 1873 /* Stop notifications for new requests from guest. 1874 * 1875 * Context: BH in IOThread 1876 */ 1877 static void virtio_blk_ioeventfd_stop_vq_bh(void *opaque) 1878 { 1879 VirtQueue *vq = opaque; 1880 EventNotifier *host_notifier = virtio_queue_get_host_notifier(vq); 1881 1882 virtio_queue_aio_detach_host_notifier(vq, qemu_get_current_aio_context()); 1883 1884 /* 1885 * Test and clear notifier after disabling event, in case poll callback 1886 * didn't have time to run. 1887 */ 1888 virtio_queue_host_notifier_read(host_notifier); 1889 } 1890 1891 /* Context: BQL held */ 1892 static void virtio_blk_stop_ioeventfd(VirtIODevice *vdev) 1893 { 1894 VirtIOBlock *s = VIRTIO_BLK(vdev); 1895 BusState *qbus = qdev_get_parent_bus(DEVICE(s)); 1896 VirtioBusClass *k = VIRTIO_BUS_GET_CLASS(qbus); 1897 unsigned i; 1898 unsigned nvqs = s->conf.num_queues; 1899 1900 if (!s->ioeventfd_started || s->ioeventfd_stopping) { 1901 return; 1902 } 1903 1904 /* Better luck next time. */ 1905 if (s->ioeventfd_disabled) { 1906 s->ioeventfd_disabled = false; 1907 s->ioeventfd_started = false; 1908 return; 1909 } 1910 s->ioeventfd_stopping = true; 1911 1912 if (!blk_in_drain(s->conf.conf.blk)) { 1913 for (i = 0; i < nvqs; i++) { 1914 VirtQueue *vq = virtio_get_queue(vdev, i); 1915 AioContext *ctx = s->vq_aio_context[i]; 1916 1917 aio_wait_bh_oneshot(ctx, virtio_blk_ioeventfd_stop_vq_bh, vq); 1918 } 1919 } 1920 1921 /* 1922 * Batch all the host notifiers in a single transaction to avoid 1923 * quadratic time complexity in address_space_update_ioeventfds(). 1924 */ 1925 memory_region_transaction_begin(); 1926 1927 for (i = 0; i < nvqs; i++) { 1928 virtio_bus_set_host_notifier(VIRTIO_BUS(qbus), i, false); 1929 } 1930 1931 /* 1932 * The transaction expects the ioeventfds to be open when it 1933 * commits. Do it now, before the cleanup loop. 1934 */ 1935 memory_region_transaction_commit(); 1936 1937 for (i = 0; i < nvqs; i++) { 1938 virtio_bus_cleanup_host_notifier(VIRTIO_BUS(qbus), i); 1939 } 1940 1941 /* 1942 * Set ->ioeventfd_started to false before draining so that host notifiers 1943 * are not detached/attached anymore. 1944 */ 1945 s->ioeventfd_started = false; 1946 1947 /* Wait for virtio_blk_dma_restart_bh() and in flight I/O to complete */ 1948 blk_drain(s->conf.conf.blk); 1949 1950 /* 1951 * Try to switch bs back to the QEMU main loop. If other users keep the 1952 * BlockBackend in the iothread, that's ok 1953 */ 1954 blk_set_aio_context(s->conf.conf.blk, qemu_get_aio_context(), NULL); 1955 1956 /* Clean up guest notifier (irq) */ 1957 k->set_guest_notifiers(qbus->parent, nvqs, false); 1958 1959 s->ioeventfd_stopping = false; 1960 } 1961 1962 static void virtio_blk_device_realize(DeviceState *dev, Error **errp) 1963 { 1964 VirtIODevice *vdev = VIRTIO_DEVICE(dev); 1965 VirtIOBlock *s = VIRTIO_BLK(dev); 1966 VirtIOBlkConf *conf = &s->conf; 1967 BlockDriverState *bs; 1968 Error *err = NULL; 1969 unsigned i; 1970 1971 if (!conf->conf.blk) { 1972 error_setg(errp, "drive property not set"); 1973 return; 1974 } 1975 if (!blk_is_inserted(conf->conf.blk)) { 1976 error_setg(errp, "Device needs media, but drive is empty"); 1977 return; 1978 } 1979 if (conf->num_queues == VIRTIO_BLK_AUTO_NUM_QUEUES) { 1980 conf->num_queues = 1; 1981 } 1982 if (!conf->num_queues) { 1983 error_setg(errp, "num-queues property must be larger than 0"); 1984 return; 1985 } 1986 if (conf->queue_size <= 2) { 1987 error_setg(errp, "invalid queue-size property (%" PRIu16 "), " 1988 "must be > 2", conf->queue_size); 1989 return; 1990 } 1991 if (!is_power_of_2(conf->queue_size) || 1992 conf->queue_size > VIRTQUEUE_MAX_SIZE) { 1993 error_setg(errp, "invalid queue-size property (%" PRIu16 "), " 1994 "must be a power of 2 (max %d)", 1995 conf->queue_size, VIRTQUEUE_MAX_SIZE); 1996 return; 1997 } 1998 1999 if (!blkconf_apply_backend_options(&conf->conf, 2000 !blk_supports_write_perm(conf->conf.blk), 2001 true, errp)) { 2002 return; 2003 } 2004 s->original_wce = blk_enable_write_cache(conf->conf.blk); 2005 if (!blkconf_geometry(&conf->conf, NULL, 65535, 255, 255, errp)) { 2006 return; 2007 } 2008 2009 if (!blkconf_blocksizes(&conf->conf, errp)) { 2010 return; 2011 } 2012 2013 bs = blk_bs(conf->conf.blk); 2014 if (bs->bl.zoned != BLK_Z_NONE) { 2015 virtio_add_feature(&s->host_features, VIRTIO_BLK_F_ZONED); 2016 if (bs->bl.zoned == BLK_Z_HM) { 2017 virtio_clear_feature(&s->host_features, VIRTIO_BLK_F_DISCARD); 2018 } 2019 } 2020 2021 if (virtio_has_feature(s->host_features, VIRTIO_BLK_F_DISCARD) && 2022 (!conf->max_discard_sectors || 2023 conf->max_discard_sectors > BDRV_REQUEST_MAX_SECTORS)) { 2024 error_setg(errp, "invalid max-discard-sectors property (%" PRIu32 ")" 2025 ", must be between 1 and %d", 2026 conf->max_discard_sectors, (int)BDRV_REQUEST_MAX_SECTORS); 2027 return; 2028 } 2029 2030 if (virtio_has_feature(s->host_features, VIRTIO_BLK_F_WRITE_ZEROES) && 2031 (!conf->max_write_zeroes_sectors || 2032 conf->max_write_zeroes_sectors > BDRV_REQUEST_MAX_SECTORS)) { 2033 error_setg(errp, "invalid max-write-zeroes-sectors property (%" PRIu32 2034 "), must be between 1 and %d", 2035 conf->max_write_zeroes_sectors, 2036 (int)BDRV_REQUEST_MAX_SECTORS); 2037 return; 2038 } 2039 2040 s->config_size = virtio_get_config_size(&virtio_blk_cfg_size_params, 2041 s->host_features); 2042 virtio_init(vdev, VIRTIO_ID_BLOCK, s->config_size); 2043 2044 qemu_mutex_init(&s->rq_lock); 2045 2046 s->blk = conf->conf.blk; 2047 s->rq = NULL; 2048 s->sector_mask = (s->conf.conf.logical_block_size / BDRV_SECTOR_SIZE) - 1; 2049 2050 for (i = 0; i < conf->num_queues; i++) { 2051 virtio_add_queue(vdev, conf->queue_size, virtio_blk_handle_output); 2052 } 2053 qemu_coroutine_inc_pool_size(conf->num_queues * conf->queue_size / 2); 2054 2055 /* Don't start ioeventfd if transport does not support notifiers. */ 2056 if (!virtio_device_ioeventfd_enabled(vdev)) { 2057 s->ioeventfd_disabled = true; 2058 } 2059 2060 virtio_blk_vq_aio_context_init(s, &err); 2061 if (err != NULL) { 2062 error_propagate(errp, err); 2063 for (i = 0; i < conf->num_queues; i++) { 2064 virtio_del_queue(vdev, i); 2065 } 2066 virtio_cleanup(vdev); 2067 return; 2068 } 2069 2070 /* 2071 * This must be after virtio_init() so virtio_blk_dma_restart_cb() gets 2072 * called after ->start_ioeventfd() has already set blk's AioContext. 2073 */ 2074 s->change = 2075 qdev_add_vm_change_state_handler(dev, virtio_blk_dma_restart_cb, s); 2076 2077 blk_ram_registrar_init(&s->blk_ram_registrar, s->blk); 2078 blk_set_dev_ops(s->blk, &virtio_block_ops, s); 2079 2080 blk_iostatus_enable(s->blk); 2081 2082 add_boot_device_lchs(dev, "/disk@0,0", 2083 conf->conf.lcyls, 2084 conf->conf.lheads, 2085 conf->conf.lsecs); 2086 } 2087 2088 static void virtio_blk_device_unrealize(DeviceState *dev) 2089 { 2090 VirtIODevice *vdev = VIRTIO_DEVICE(dev); 2091 VirtIOBlock *s = VIRTIO_BLK(dev); 2092 VirtIOBlkConf *conf = &s->conf; 2093 unsigned i; 2094 2095 blk_drain(s->blk); 2096 del_boot_device_lchs(dev, "/disk@0,0"); 2097 virtio_blk_vq_aio_context_cleanup(s); 2098 for (i = 0; i < conf->num_queues; i++) { 2099 virtio_del_queue(vdev, i); 2100 } 2101 qemu_coroutine_dec_pool_size(conf->num_queues * conf->queue_size / 2); 2102 qemu_mutex_destroy(&s->rq_lock); 2103 blk_ram_registrar_destroy(&s->blk_ram_registrar); 2104 qemu_del_vm_change_state_handler(s->change); 2105 blockdev_mark_auto_del(s->blk); 2106 virtio_cleanup(vdev); 2107 } 2108 2109 static void virtio_blk_instance_init(Object *obj) 2110 { 2111 VirtIOBlock *s = VIRTIO_BLK(obj); 2112 2113 device_add_bootindex_property(obj, &s->conf.conf.bootindex, 2114 "bootindex", "/disk@0,0", 2115 DEVICE(obj)); 2116 } 2117 2118 static const VMStateDescription vmstate_virtio_blk = { 2119 .name = "virtio-blk", 2120 .minimum_version_id = 2, 2121 .version_id = 2, 2122 .fields = (const VMStateField[]) { 2123 VMSTATE_VIRTIO_DEVICE, 2124 VMSTATE_END_OF_LIST() 2125 }, 2126 }; 2127 2128 static Property virtio_blk_properties[] = { 2129 DEFINE_BLOCK_PROPERTIES(VirtIOBlock, conf.conf), 2130 DEFINE_BLOCK_ERROR_PROPERTIES(VirtIOBlock, conf.conf), 2131 DEFINE_BLOCK_CHS_PROPERTIES(VirtIOBlock, conf.conf), 2132 DEFINE_PROP_STRING("serial", VirtIOBlock, conf.serial), 2133 DEFINE_PROP_BIT64("config-wce", VirtIOBlock, host_features, 2134 VIRTIO_BLK_F_CONFIG_WCE, true), 2135 #ifdef __linux__ 2136 DEFINE_PROP_BIT64("scsi", VirtIOBlock, host_features, 2137 VIRTIO_BLK_F_SCSI, false), 2138 #endif 2139 DEFINE_PROP_BIT("request-merging", VirtIOBlock, conf.request_merging, 0, 2140 true), 2141 DEFINE_PROP_UINT16("num-queues", VirtIOBlock, conf.num_queues, 2142 VIRTIO_BLK_AUTO_NUM_QUEUES), 2143 DEFINE_PROP_UINT16("queue-size", VirtIOBlock, conf.queue_size, 256), 2144 DEFINE_PROP_BOOL("seg-max-adjust", VirtIOBlock, conf.seg_max_adjust, true), 2145 DEFINE_PROP_LINK("iothread", VirtIOBlock, conf.iothread, TYPE_IOTHREAD, 2146 IOThread *), 2147 DEFINE_PROP_IOTHREAD_VQ_MAPPING_LIST("iothread-vq-mapping", VirtIOBlock, 2148 conf.iothread_vq_mapping_list), 2149 DEFINE_PROP_BIT64("discard", VirtIOBlock, host_features, 2150 VIRTIO_BLK_F_DISCARD, true), 2151 DEFINE_PROP_BOOL("report-discard-granularity", VirtIOBlock, 2152 conf.report_discard_granularity, true), 2153 DEFINE_PROP_BIT64("write-zeroes", VirtIOBlock, host_features, 2154 VIRTIO_BLK_F_WRITE_ZEROES, true), 2155 DEFINE_PROP_UINT32("max-discard-sectors", VirtIOBlock, 2156 conf.max_discard_sectors, BDRV_REQUEST_MAX_SECTORS), 2157 DEFINE_PROP_UINT32("max-write-zeroes-sectors", VirtIOBlock, 2158 conf.max_write_zeroes_sectors, BDRV_REQUEST_MAX_SECTORS), 2159 DEFINE_PROP_BOOL("x-enable-wce-if-config-wce", VirtIOBlock, 2160 conf.x_enable_wce_if_config_wce, true), 2161 DEFINE_PROP_END_OF_LIST(), 2162 }; 2163 2164 static void virtio_blk_class_init(ObjectClass *klass, void *data) 2165 { 2166 DeviceClass *dc = DEVICE_CLASS(klass); 2167 VirtioDeviceClass *vdc = VIRTIO_DEVICE_CLASS(klass); 2168 2169 device_class_set_props(dc, virtio_blk_properties); 2170 dc->vmsd = &vmstate_virtio_blk; 2171 set_bit(DEVICE_CATEGORY_STORAGE, dc->categories); 2172 vdc->realize = virtio_blk_device_realize; 2173 vdc->unrealize = virtio_blk_device_unrealize; 2174 vdc->get_config = virtio_blk_update_config; 2175 vdc->set_config = virtio_blk_set_config; 2176 vdc->get_features = virtio_blk_get_features; 2177 vdc->set_status = virtio_blk_set_status; 2178 vdc->reset = virtio_blk_reset; 2179 vdc->save = virtio_blk_save_device; 2180 vdc->load = virtio_blk_load_device; 2181 vdc->start_ioeventfd = virtio_blk_start_ioeventfd; 2182 vdc->stop_ioeventfd = virtio_blk_stop_ioeventfd; 2183 } 2184 2185 static const TypeInfo virtio_blk_info = { 2186 .name = TYPE_VIRTIO_BLK, 2187 .parent = TYPE_VIRTIO_DEVICE, 2188 .instance_size = sizeof(VirtIOBlock), 2189 .instance_init = virtio_blk_instance_init, 2190 .class_init = virtio_blk_class_init, 2191 }; 2192 2193 static void virtio_register_types(void) 2194 { 2195 type_register_static(&virtio_blk_info); 2196 } 2197 2198 type_init(virtio_register_types) 2199