hw/virtio-blk.c

   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-common.h"
  15 #include "qemu-error.h"
  16 #include "trace.h"
  17 #include "blockdev.h"
  18 #include "virtio-blk.h"
  19 #include "scsi-defs.h"
  20 #ifdef __linux__
  21 # include <scsi/sg.h>
  22 #endif
  23
  24 typedef struct VirtIOBlock
  25 {
  26     VirtIODevice vdev;
  27     BlockDriverState *bs;
  28     VirtQueue *vq;
  29     void *rq;
  30     QEMUBH *bh;
  31     BlockConf *conf;
  32     char *serial;
  33     unsigned short sector_mask;
  34     DeviceState *qdev;
  35 } VirtIOBlock;
  36
  37 static VirtIOBlock *to_virtio_blk(VirtIODevice *vdev)
  38 {
  39     return (VirtIOBlock *)vdev;
  40 }
  41
  42 typedef struct VirtIOBlockReq
  43 {
  44     VirtIOBlock *dev;
  45     VirtQueueElement elem;
  46     struct virtio_blk_inhdr *in;
  47     struct virtio_blk_outhdr *out;
  48     struct virtio_scsi_inhdr *scsi;
  49     QEMUIOVector qiov;
  50     struct VirtIOBlockReq *next;
  51     BlockAcctCookie acct;
  52 } VirtIOBlockReq;
  53
  54 static void virtio_blk_req_complete(VirtIOBlockReq *req, int status)
  55 {
  56     VirtIOBlock *s = req->dev;
  57
  58     trace_virtio_blk_req_complete(req, status);
  59
  60     stb_p(&req->in->status, status);
  61     virtqueue_push(s->vq, &req->elem, req->qiov.size + sizeof(*req->in));
  62     virtio_notify(&s->vdev, s->vq);
  63 }
  64
  65 static int virtio_blk_handle_rw_error(VirtIOBlockReq *req, int error,
  66     int is_read)
  67 {
  68     BlockErrorAction action = bdrv_get_on_error(req->dev->bs, is_read);
  69     VirtIOBlock *s = req->dev;
  70
  71     if (action == BLOCK_ERR_IGNORE) {
  72         bdrv_mon_event(s->bs, BDRV_ACTION_IGNORE, is_read);
  73         return 0;
  74     }
  75
  76     if ((error == ENOSPC && action == BLOCK_ERR_STOP_ENOSPC)
  77             || action == BLOCK_ERR_STOP_ANY) {
  78         req->next = s->rq;
  79         s->rq = req;
  80         bdrv_mon_event(s->bs, BDRV_ACTION_STOP, is_read);
  81         vm_stop(RUN_STATE_IO_ERROR);
  82         bdrv_iostatus_set_err(s->bs, error);
  83     } else {
  84         virtio_blk_req_complete(req, VIRTIO_BLK_S_IOERR);
  85         bdrv_acct_done(s->bs, &req->acct);
  86         g_free(req);
  87         bdrv_mon_event(s->bs, BDRV_ACTION_REPORT, is_read);
  88     }
  89
  90     return 1;
  91 }
  92
  93 static void virtio_blk_rw_complete(void *opaque, int ret)
  94 {
  95     VirtIOBlockReq *req = opaque;
  96
  97     trace_virtio_blk_rw_complete(req, ret);
  98
  99     if (ret) {
 100         int is_read = !(ldl_p(&req->out->type) & VIRTIO_BLK_T_OUT);
 101         if (virtio_blk_handle_rw_error(req, -ret, is_read))
 102             return;
 103     }
 104
 105     virtio_blk_req_complete(req, VIRTIO_BLK_S_OK);
 106     bdrv_acct_done(req->dev->bs, &req->acct);
 107     g_free(req);
 108 }
 109
 110 static void virtio_blk_flush_complete(void *opaque, int ret)
 111 {
 112     VirtIOBlockReq *req = opaque;
 113
 114     if (ret) {
 115         if (virtio_blk_handle_rw_error(req, -ret, 0)) {
 116             return;
 117         }
 118     }
 119
 120     virtio_blk_req_complete(req, VIRTIO_BLK_S_OK);
 121     bdrv_acct_done(req->dev->bs, &req->acct);
 122     g_free(req);
 123 }
 124
 125 static VirtIOBlockReq *virtio_blk_alloc_request(VirtIOBlock *s)
 126 {
 127     VirtIOBlockReq *req = g_malloc(sizeof(*req));
 128     req->dev = s;
 129     req->qiov.size = 0;
 130     req->next = NULL;
 131     return req;
 132 }
 133
 134 static VirtIOBlockReq *virtio_blk_get_request(VirtIOBlock *s)
 135 {
 136     VirtIOBlockReq *req = virtio_blk_alloc_request(s);
 137
 138     if (req != NULL) {
 139         if (!virtqueue_pop(s->vq, &req->elem)) {
 140             g_free(req);
 141             return NULL;
 142         }
 143     }
 144
 145     return req;
 146 }
 147
 148 #ifdef __linux__
 149 static void virtio_blk_handle_scsi(VirtIOBlockReq *req)
 150 {
 151     struct sg_io_hdr hdr;
 152     int ret;
 153     int status;
 154     int i;
 155
 156     /*
 157      * We require at least one output segment each for the virtio_blk_outhdr
 158      * and the SCSI command block.
 159      *
 160      * We also at least require the virtio_blk_inhdr, the virtio_scsi_inhdr
 161      * and the sense buffer pointer in the input segments.
 162      */
 163     if (req->elem.out_num < 2 || req->elem.in_num < 3) {
 164         virtio_blk_req_complete(req, VIRTIO_BLK_S_IOERR);
 165         g_free(req);
 166         return;
 167     }
 168
 169     /*
 170      * No support for bidirection commands yet.
 171      */
 172     if (req->elem.out_num > 2 && req->elem.in_num > 3) {
 173         virtio_blk_req_complete(req, VIRTIO_BLK_S_UNSUPP);
 174         g_free(req);
 175         return;
 176     }
 177
 178     /*
 179      * The scsi inhdr is placed in the second-to-last input segment, just
 180      * before the regular inhdr.
 181      */
 182     req->scsi = (void *)req->elem.in_sg[req->elem.in_num - 2].iov_base;
 183
 184     memset(&hdr, 0, sizeof(struct sg_io_hdr));
 185     hdr.interface_id = 'S';
 186     hdr.cmd_len = req->elem.out_sg[1].iov_len;
 187     hdr.cmdp = req->elem.out_sg[1].iov_base;
 188     hdr.dxfer_len = 0;
 189
 190     if (req->elem.out_num > 2) {
 191         /*
 192          * If there are more than the minimally required 2 output segments
 193          * there is write payload starting from the third iovec.
 194          */
 195         hdr.dxfer_direction = SG_DXFER_TO_DEV;
 196         hdr.iovec_count = req->elem.out_num - 2;
 197
 198         for (i = 0; i < hdr.iovec_count; i++)
 199             hdr.dxfer_len += req->elem.out_sg[i + 2].iov_len;
 200
 201         hdr.dxferp = req->elem.out_sg + 2;
 202
 203     } else if (req->elem.in_num > 3) {
 204         /*
 205          * If we have more than 3 input segments the guest wants to actually
 206          * read data.
 207          */
 208         hdr.dxfer_direction = SG_DXFER_FROM_DEV;
 209         hdr.iovec_count = req->elem.in_num - 3;
 210         for (i = 0; i < hdr.iovec_count; i++)
 211             hdr.dxfer_len += req->elem.in_sg[i].iov_len;
 212
 213         hdr.dxferp = req->elem.in_sg;
 214     } else {
 215         /*
 216          * Some SCSI commands don't actually transfer any data.
 217          */
 218         hdr.dxfer_direction = SG_DXFER_NONE;
 219     }
 220
 221     hdr.sbp = req->elem.in_sg[req->elem.in_num - 3].iov_base;
 222     hdr.mx_sb_len = req->elem.in_sg[req->elem.in_num - 3].iov_len;
 223
 224     ret = bdrv_ioctl(req->dev->bs, SG_IO, &hdr);
 225     if (ret) {
 226         status = VIRTIO_BLK_S_UNSUPP;
 227         hdr.status = ret;
 228         hdr.resid = hdr.dxfer_len;
 229     } else if (hdr.status) {
 230         status = VIRTIO_BLK_S_IOERR;
 231     } else {
 232         status = VIRTIO_BLK_S_OK;
 233     }
 234
 235     /*
 236      * From SCSI-Generic-HOWTO: "Some lower level drivers (e.g. ide-scsi)
 237      * clear the masked_status field [hence status gets cleared too, see
 238      * block/scsi_ioctl.c] even when a CHECK_CONDITION or COMMAND_TERMINATED
 239      * status has occurred.  However they do set DRIVER_SENSE in driver_status
 240      * field. Also a (sb_len_wr > 0) indicates there is a sense buffer.
 241      */
 242     if (hdr.status == 0 && hdr.sb_len_wr > 0) {
 243         hdr.status = CHECK_CONDITION;
 244     }
 245
 246     stl_p(&req->scsi->errors,
 247           hdr.status | (hdr.msg_status << 8) |
 248           (hdr.host_status << 16) | (hdr.driver_status << 24));
 249     stl_p(&req->scsi->residual, hdr.resid);
 250     stl_p(&req->scsi->sense_len, hdr.sb_len_wr);
 251     stl_p(&req->scsi->data_len, hdr.dxfer_len);
 252
 253     virtio_blk_req_complete(req, status);
 254     g_free(req);
 255 }
 256 #else
 257 static void virtio_blk_handle_scsi(VirtIOBlockReq *req)
 258 {
 259     virtio_blk_req_complete(req, VIRTIO_BLK_S_UNSUPP);
 260     g_free(req);
 261 }
 262 #endif /* __linux__ */
 263
 264 typedef struct MultiReqBuffer {
 265     BlockRequest        blkreq[32];
 266     unsigned int        num_writes;
 267 } MultiReqBuffer;
 268
 269 static void virtio_submit_multiwrite(BlockDriverState *bs, MultiReqBuffer *mrb)
 270 {
 271     int i, ret;
 272
 273     if (!mrb->num_writes) {
 274         return;
 275     }
 276
 277     ret = bdrv_aio_multiwrite(bs, mrb->blkreq, mrb->num_writes);
 278     if (ret != 0) {
 279         for (i = 0; i < mrb->num_writes; i++) {
 280             if (mrb->blkreq[i].error) {
 281                 virtio_blk_rw_complete(mrb->blkreq[i].opaque, -EIO);
 282             }
 283         }
 284     }
 285
 286     mrb->num_writes = 0;
 287 }
 288
 289 static void virtio_blk_handle_flush(VirtIOBlockReq *req, MultiReqBuffer *mrb)
 290 {
 291     BlockDriverAIOCB *acb;
 292
 293     bdrv_acct_start(req->dev->bs, &req->acct, 0, BDRV_ACCT_FLUSH);
 294
 295     /*
 296      * Make sure all outstanding writes are posted to the backing device.
 297      */
 298     virtio_submit_multiwrite(req->dev->bs, mrb);
 299
 300     acb = bdrv_aio_flush(req->dev->bs, virtio_blk_flush_complete, req);
 301     if (!acb) {
 302         virtio_blk_flush_complete(req, -EIO);
 303     }
 304 }
 305
 306 static void virtio_blk_handle_write(VirtIOBlockReq *req, MultiReqBuffer *mrb)
 307 {
 308     BlockRequest *blkreq;
 309     uint64_t sector;
 310
 311     sector = ldq_p(&req->out->sector);
 312
 313     bdrv_acct_start(req->dev->bs, &req->acct, req->qiov.size, BDRV_ACCT_WRITE);
 314
 315     trace_virtio_blk_handle_write(req, sector, req->qiov.size / 512);
 316
 317     if (sector & req->dev->sector_mask) {
 318         virtio_blk_rw_complete(req, -EIO);
 319         return;
 320     }
 321     if (req->qiov.size % req->dev->conf->logical_block_size) {
 322         virtio_blk_rw_complete(req, -EIO);
 323         return;
 324     }
 325
 326     if (mrb->num_writes == 32) {
 327         virtio_submit_multiwrite(req->dev->bs, mrb);
 328     }
 329
 330     blkreq = &mrb->blkreq[mrb->num_writes];
 331     blkreq->sector = sector;
 332     blkreq->nb_sectors = req->qiov.size / BDRV_SECTOR_SIZE;
 333     blkreq->qiov = &req->qiov;
 334     blkreq->cb = virtio_blk_rw_complete;
 335     blkreq->opaque = req;
 336     blkreq->error = 0;
 337
 338     mrb->num_writes++;
 339 }
 340
 341 static void virtio_blk_handle_read(VirtIOBlockReq *req)
 342 {
 343     BlockDriverAIOCB *acb;
 344     uint64_t sector;
 345
 346     sector = ldq_p(&req->out->sector);
 347
 348     bdrv_acct_start(req->dev->bs, &req->acct, req->qiov.size, BDRV_ACCT_READ);
 349
 350     if (sector & req->dev->sector_mask) {
 351         virtio_blk_rw_complete(req, -EIO);
 352         return;
 353     }
 354     if (req->qiov.size % req->dev->conf->logical_block_size) {
 355         virtio_blk_rw_complete(req, -EIO);
 356         return;
 357     }
 358
 359     acb = bdrv_aio_readv(req->dev->bs, sector, &req->qiov,
 360                          req->qiov.size / BDRV_SECTOR_SIZE,
 361                          virtio_blk_rw_complete, req);
 362     if (!acb) {
 363         virtio_blk_rw_complete(req, -EIO);
 364     }
 365 }
 366
 367 static void virtio_blk_handle_request(VirtIOBlockReq *req,
 368     MultiReqBuffer *mrb)
 369 {
 370     uint32_t type;
 371
 372     if (req->elem.out_num < 1 || req->elem.in_num < 1) {
 373         error_report("virtio-blk missing headers");
 374         exit(1);
 375     }
 376
 377     if (req->elem.out_sg[0].iov_len < sizeof(*req->out) ||
 378         req->elem.in_sg[req->elem.in_num - 1].iov_len < sizeof(*req->in)) {
 379         error_report("virtio-blk header not in correct element");
 380         exit(1);
 381     }
 382
 383     req->out = (void *)req->elem.out_sg[0].iov_base;
 384     req->in = (void *)req->elem.in_sg[req->elem.in_num - 1].iov_base;
 385
 386     type = ldl_p(&req->out->type);
 387
 388     if (type & VIRTIO_BLK_T_FLUSH) {
 389         virtio_blk_handle_flush(req, mrb);
 390     } else if (type & VIRTIO_BLK_T_SCSI_CMD) {
 391         virtio_blk_handle_scsi(req);
 392     } else if (type & VIRTIO_BLK_T_GET_ID) {
 393         VirtIOBlock *s = req->dev;
 394
 395         /*
 396          * NB: per existing s/n string convention the string is
 397          * terminated by '\0' only when shorter than buffer.
 398          */
 399         strncpy(req->elem.in_sg[0].iov_base,
 400                 s->serial ? s->serial : "",
 401                 MIN(req->elem.in_sg[0].iov_len, VIRTIO_BLK_ID_BYTES));
 402         virtio_blk_req_complete(req, VIRTIO_BLK_S_OK);
 403         g_free(req);
 404     } else if (type & VIRTIO_BLK_T_OUT) {
 405         qemu_iovec_init_external(&req->qiov, &req->elem.out_sg[1],
 406                                  req->elem.out_num - 1);
 407         virtio_blk_handle_write(req, mrb);
 408     } else {
 409         qemu_iovec_init_external(&req->qiov, &req->elem.in_sg[0],
 410                                  req->elem.in_num - 1);
 411         virtio_blk_handle_read(req);
 412     }
 413 }
 414
 415 static void virtio_blk_handle_output(VirtIODevice *vdev, VirtQueue *vq)
 416 {
 417     VirtIOBlock *s = to_virtio_blk(vdev);
 418     VirtIOBlockReq *req;
 419     MultiReqBuffer mrb = {
 420         .num_writes = 0,
 421     };
 422
 423     while ((req = virtio_blk_get_request(s))) {
 424         virtio_blk_handle_request(req, &mrb);
 425     }
 426
 427     virtio_submit_multiwrite(s->bs, &mrb);
 428
 429     /*
 430      * FIXME: Want to check for completions before returning to guest mode,
 431      * so cached reads and writes are reported as quickly as possible. But
 432      * that should be done in the generic block layer.
 433      */
 434 }
 435
 436 static void virtio_blk_dma_restart_bh(void *opaque)
 437 {
 438     VirtIOBlock *s = opaque;
 439     VirtIOBlockReq *req = s->rq;
 440     MultiReqBuffer mrb = {
 441         .num_writes = 0,
 442     };
 443
 444     qemu_bh_delete(s->bh);
 445     s->bh = NULL;
 446
 447     s->rq = NULL;
 448
 449     while (req) {
 450         virtio_blk_handle_request(req, &mrb);
 451         req = req->next;
 452     }
 453
 454     virtio_submit_multiwrite(s->bs, &mrb);
 455 }
 456
 457 static void virtio_blk_dma_restart_cb(void *opaque, int running,
 458                                       RunState state)
 459 {
 460     VirtIOBlock *s = opaque;
 461
 462     if (!running)
 463         return;
 464
 465     if (!s->bh) {
 466         s->bh = qemu_bh_new(virtio_blk_dma_restart_bh, s);
 467         qemu_bh_schedule(s->bh);
 468     }
 469 }
 470
 471 static void virtio_blk_reset(VirtIODevice *vdev)
 472 {
 473     /*
 474      * This should cancel pending requests, but can't do nicely until there
 475      * are per-device request lists.
 476      */
 477     qemu_aio_flush();
 478 }
 479
 480 /* coalesce internal state, copy to pci i/o region 0
 481  */
 482 static void virtio_blk_update_config(VirtIODevice *vdev, uint8_t *config)
 483 {
 484     VirtIOBlock *s = to_virtio_blk(vdev);
 485     struct virtio_blk_config blkcfg;
 486     uint64_t capacity;
 487     int cylinders, heads, secs;
 488
 489     bdrv_get_geometry(s->bs, &capacity);
 490     bdrv_get_geometry_hint(s->bs, &cylinders, &heads, &secs);
 491     memset(&blkcfg, 0, sizeof(blkcfg));
 492     stq_raw(&blkcfg.capacity, capacity);
 493     stl_raw(&blkcfg.seg_max, 128 - 2);
 494     stw_raw(&blkcfg.cylinders, cylinders);
 495     blkcfg.heads = heads;
 496     blkcfg.sectors = secs & ~s->sector_mask;
 497     blkcfg.blk_size = s->conf->logical_block_size;
 498     blkcfg.size_max = 0;
 499     blkcfg.physical_block_exp = get_physical_block_exp(s->conf);
 500     blkcfg.alignment_offset = 0;
 501     blkcfg.min_io_size = s->conf->min_io_size / blkcfg.blk_size;
 502     blkcfg.opt_io_size = s->conf->opt_io_size / blkcfg.blk_size;
 503     memcpy(config, &blkcfg, sizeof(struct virtio_blk_config));
 504 }
 505
 506 static uint32_t virtio_blk_get_features(VirtIODevice *vdev, uint32_t features)
 507 {
 508     VirtIOBlock *s = to_virtio_blk(vdev);
 509
 510     features |= (1 << VIRTIO_BLK_F_SEG_MAX);
 511     features |= (1 << VIRTIO_BLK_F_GEOMETRY);
 512     features |= (1 << VIRTIO_BLK_F_TOPOLOGY);
 513     features |= (1 << VIRTIO_BLK_F_BLK_SIZE);
 514
 515     if (bdrv_enable_write_cache(s->bs))
 516         features |= (1 << VIRTIO_BLK_F_WCACHE);
 517
 518     if (bdrv_is_read_only(s->bs))
 519         features |= 1 << VIRTIO_BLK_F_RO;
 520
 521     return features;
 522 }
 523
 524 static void virtio_blk_save(QEMUFile *f, void *opaque)
 525 {
 526     VirtIOBlock *s = opaque;
 527     VirtIOBlockReq *req = s->rq;
 528
 529     virtio_save(&s->vdev, f);
 530
 531     while (req) {
 532         qemu_put_sbyte(f, 1);
 533         qemu_put_buffer(f, (unsigned char*)&req->elem, sizeof(req->elem));
 534         req = req->next;
 535     }
 536     qemu_put_sbyte(f, 0);
 537 }
 538
 539 static int virtio_blk_load(QEMUFile *f, void *opaque, int version_id)
 540 {
 541     VirtIOBlock *s = opaque;
 542
 543     if (version_id != 2)
 544         return -EINVAL;
 545
 546     virtio_load(&s->vdev, f);
 547     while (qemu_get_sbyte(f)) {
 548         VirtIOBlockReq *req = virtio_blk_alloc_request(s);
 549         qemu_get_buffer(f, (unsigned char*)&req->elem, sizeof(req->elem));
 550         req->next = s->rq;
 551         s->rq = req;
 552
 553         virtqueue_map_sg(req->elem.in_sg, req->elem.in_addr,
 554             req->elem.in_num, 1);
 555         virtqueue_map_sg(req->elem.out_sg, req->elem.out_addr,
 556             req->elem.out_num, 0);
 557     }
 558
 559     return 0;
 560 }
 561
 562 static void virtio_blk_resize(void *opaque)
 563 {
 564     VirtIOBlock *s = opaque;
 565
 566     virtio_notify_config(&s->vdev);
 567 }
 568
 569 static const BlockDevOps virtio_block_ops = {
 570     .resize_cb = virtio_blk_resize,
 571 };
 572
 573 VirtIODevice *virtio_blk_init(DeviceState *dev, BlockConf *conf,
 574                               char **serial)
 575 {
 576     VirtIOBlock *s;
 577     int cylinders, heads, secs;
 578     static int virtio_blk_id;
 579     DriveInfo *dinfo;
 580
 581     if (!conf->bs) {
 582         error_report("virtio-blk-pci: drive property not set");
 583         return NULL;
 584     }
 585     if (!bdrv_is_inserted(conf->bs)) {
 586         error_report("Device needs media, but drive is empty");
 587         return NULL;
 588     }
 589
 590     if (!*serial) {
 591         /* try to fall back to value set with legacy -drive serial=... */
 592         dinfo = drive_get_by_blockdev(conf->bs);
 593         if (*dinfo->serial) {
 594             *serial = strdup(dinfo->serial);
 595         }
 596     }
 597
 598     s = (VirtIOBlock *)virtio_common_init("virtio-blk", VIRTIO_ID_BLOCK,
 599                                           sizeof(struct virtio_blk_config),
 600                                           sizeof(VirtIOBlock));
 601
 602     s->vdev.get_config = virtio_blk_update_config;
 603     s->vdev.get_features = virtio_blk_get_features;
 604     s->vdev.reset = virtio_blk_reset;
 605     s->bs = conf->bs;
 606     s->conf = conf;
 607     s->serial = *serial;
 608     s->rq = NULL;
 609     s->sector_mask = (s->conf->logical_block_size / BDRV_SECTOR_SIZE) - 1;
 610     bdrv_guess_geometry(s->bs, &cylinders, &heads, &secs);
 611
 612     s->vq = virtio_add_queue(&s->vdev, 128, virtio_blk_handle_output);
 613
 614     qemu_add_vm_change_state_handler(virtio_blk_dma_restart_cb, s);
 615     s->qdev = dev;
 616     register_savevm(dev, "virtio-blk", virtio_blk_id++, 2,
 617                     virtio_blk_save, virtio_blk_load, s);
 618     bdrv_set_dev_ops(s->bs, &virtio_block_ops, s);
 619     bdrv_set_buffer_alignment(s->bs, conf->logical_block_size);
 620
 621     bdrv_iostatus_enable(s->bs);
 622     add_boot_device_path(conf->bootindex, dev, "/disk@0,0");
 623
 624     return &s->vdev;
 625 }
 626
 627 void virtio_blk_exit(VirtIODevice *vdev)
 628 {
 629     VirtIOBlock *s = to_virtio_blk(vdev);
 630     unregister_savevm(s->qdev, "virtio-blk", s);
 631     virtio_cleanup(vdev);
 632 }