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