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