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     if ((req->dev->vdev.guest_features & (1 << VIRTIO_BLK_F_SCSI)) == 0) {
 157         virtio_blk_req_complete(req, VIRTIO_BLK_S_UNSUPP);
 158         g_free(req);
 159         return;
 160     }
 161
 162     /*
 163      * We require at least one output segment each for the virtio_blk_outhdr
 164      * and the SCSI command block.
 165      *
 166      * We also at least require the virtio_blk_inhdr, the virtio_scsi_inhdr
 167      * and the sense buffer pointer in the input segments.
 168      */
 169     if (req->elem.out_num < 2 || req->elem.in_num < 3) {
 170         virtio_blk_req_complete(req, VIRTIO_BLK_S_IOERR);
 171         g_free(req);
 172         return;
 173     }
 174
 175     /*
 176      * No support for bidirection commands yet.
 177      */
 178     if (req->elem.out_num > 2 && req->elem.in_num > 3) {
 179         virtio_blk_req_complete(req, VIRTIO_BLK_S_UNSUPP);
 180         g_free(req);
 181         return;
 182     }
 183
 184     /*
 185      * The scsi inhdr is placed in the second-to-last input segment, just
 186      * before the regular inhdr.
 187      */
 188     req->scsi = (void *)req->elem.in_sg[req->elem.in_num - 2].iov_base;
 189
 190     memset(&hdr, 0, sizeof(struct sg_io_hdr));
 191     hdr.interface_id = 'S';
 192     hdr.cmd_len = req->elem.out_sg[1].iov_len;
 193     hdr.cmdp = req->elem.out_sg[1].iov_base;
 194     hdr.dxfer_len = 0;
 195
 196     if (req->elem.out_num > 2) {
 197         /*
 198          * If there are more than the minimally required 2 output segments
 199          * there is write payload starting from the third iovec.
 200          */
 201         hdr.dxfer_direction = SG_DXFER_TO_DEV;
 202         hdr.iovec_count = req->elem.out_num - 2;
 203
 204         for (i = 0; i < hdr.iovec_count; i++)
 205             hdr.dxfer_len += req->elem.out_sg[i + 2].iov_len;
 206
 207         hdr.dxferp = req->elem.out_sg + 2;
 208
 209     } else if (req->elem.in_num > 3) {
 210         /*
 211          * If we have more than 3 input segments the guest wants to actually
 212          * read data.
 213          */
 214         hdr.dxfer_direction = SG_DXFER_FROM_DEV;
 215         hdr.iovec_count = req->elem.in_num - 3;
 216         for (i = 0; i < hdr.iovec_count; i++)
 217             hdr.dxfer_len += req->elem.in_sg[i].iov_len;
 218
 219         hdr.dxferp = req->elem.in_sg;
 220     } else {
 221         /*
 222          * Some SCSI commands don't actually transfer any data.
 223          */
 224         hdr.dxfer_direction = SG_DXFER_NONE;
 225     }
 226
 227     hdr.sbp = req->elem.in_sg[req->elem.in_num - 3].iov_base;
 228     hdr.mx_sb_len = req->elem.in_sg[req->elem.in_num - 3].iov_len;
 229
 230     ret = bdrv_ioctl(req->dev->bs, SG_IO, &hdr);
 231     if (ret) {
 232         status = VIRTIO_BLK_S_UNSUPP;
 233         hdr.status = ret;
 234         hdr.resid = hdr.dxfer_len;
 235     } else if (hdr.status) {
 236         status = VIRTIO_BLK_S_IOERR;
 237     } else {
 238         status = VIRTIO_BLK_S_OK;
 239     }
 240
 241     /*
 242      * From SCSI-Generic-HOWTO: "Some lower level drivers (e.g. ide-scsi)
 243      * clear the masked_status field [hence status gets cleared too, see
 244      * block/scsi_ioctl.c] even when a CHECK_CONDITION or COMMAND_TERMINATED
 245      * status has occurred.  However they do set DRIVER_SENSE in driver_status
 246      * field. Also a (sb_len_wr > 0) indicates there is a sense buffer.
 247      */
 248     if (hdr.status == 0 && hdr.sb_len_wr > 0) {
 249         hdr.status = CHECK_CONDITION;
 250     }
 251
 252     stl_p(&req->scsi->errors,
 253           hdr.status | (hdr.msg_status << 8) |
 254           (hdr.host_status << 16) | (hdr.driver_status << 24));
 255     stl_p(&req->scsi->residual, hdr.resid);
 256     stl_p(&req->scsi->sense_len, hdr.sb_len_wr);
 257     stl_p(&req->scsi->data_len, hdr.dxfer_len);
 258
 259     virtio_blk_req_complete(req, status);
 260     g_free(req);
 261 }
 262 #else
 263 static void virtio_blk_handle_scsi(VirtIOBlockReq *req)
 264 {
 265     virtio_blk_req_complete(req, VIRTIO_BLK_S_UNSUPP);
 266     g_free(req);
 267 }
 268 #endif /* __linux__ */
 269
 270 typedef struct MultiReqBuffer {
 271     BlockRequest        blkreq[32];
 272     unsigned int        num_writes;
 273 } MultiReqBuffer;
 274
 275 static void virtio_submit_multiwrite(BlockDriverState *bs, MultiReqBuffer *mrb)
 276 {
 277     int i, ret;
 278
 279     if (!mrb->num_writes) {
 280         return;
 281     }
 282
 283     ret = bdrv_aio_multiwrite(bs, mrb->blkreq, mrb->num_writes);
 284     if (ret != 0) {
 285         for (i = 0; i < mrb->num_writes; i++) {
 286             if (mrb->blkreq[i].error) {
 287                 virtio_blk_rw_complete(mrb->blkreq[i].opaque, -EIO);
 288             }
 289         }
 290     }
 291
 292     mrb->num_writes = 0;
 293 }
 294
 295 static void virtio_blk_handle_flush(VirtIOBlockReq *req, MultiReqBuffer *mrb)
 296 {
 297     bdrv_acct_start(req->dev->bs, &req->acct, 0, BDRV_ACCT_FLUSH);
 298
 299     /*
 300      * Make sure all outstanding writes are posted to the backing device.
 301      */
 302     virtio_submit_multiwrite(req->dev->bs, mrb);
 303     bdrv_aio_flush(req->dev->bs, virtio_blk_flush_complete, req);
 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     uint64_t sector;
 344
 345     sector = ldq_p(&req->out->sector);
 346
 347     bdrv_acct_start(req->dev->bs, &req->acct, req->qiov.size, BDRV_ACCT_READ);
 348
 349     if (sector & req->dev->sector_mask) {
 350         virtio_blk_rw_complete(req, -EIO);
 351         return;
 352     }
 353     if (req->qiov.size % req->dev->conf->logical_block_size) {
 354         virtio_blk_rw_complete(req, -EIO);
 355         return;
 356     }
 357     bdrv_aio_readv(req->dev->bs, sector, &req->qiov,
 358                    req->qiov.size / BDRV_SECTOR_SIZE,
 359                    virtio_blk_rw_complete, req);
 360 }
 361
 362 static void virtio_blk_handle_request(VirtIOBlockReq *req,
 363     MultiReqBuffer *mrb)
 364 {
 365     uint32_t type;
 366
 367     if (req->elem.out_num < 1 || req->elem.in_num < 1) {
 368         error_report("virtio-blk missing headers");
 369         exit(1);
 370     }
 371
 372     if (req->elem.out_sg[0].iov_len < sizeof(*req->out) ||
 373         req->elem.in_sg[req->elem.in_num - 1].iov_len < sizeof(*req->in)) {
 374         error_report("virtio-blk header not in correct element");
 375         exit(1);
 376     }
 377
 378     req->out = (void *)req->elem.out_sg[0].iov_base;
 379     req->in = (void *)req->elem.in_sg[req->elem.in_num - 1].iov_base;
 380
 381     type = ldl_p(&req->out->type);
 382
 383     if (type & VIRTIO_BLK_T_FLUSH) {
 384         virtio_blk_handle_flush(req, mrb);
 385     } else if (type & VIRTIO_BLK_T_SCSI_CMD) {
 386         virtio_blk_handle_scsi(req);
 387     } else if (type & VIRTIO_BLK_T_GET_ID) {
 388         VirtIOBlock *s = req->dev;
 389
 390         /*
 391          * NB: per existing s/n string convention the string is
 392          * terminated by '\0' only when shorter than buffer.
 393          */
 394         strncpy(req->elem.in_sg[0].iov_base,
 395                 s->serial ? s->serial : "",
 396                 MIN(req->elem.in_sg[0].iov_len, VIRTIO_BLK_ID_BYTES));
 397         virtio_blk_req_complete(req, VIRTIO_BLK_S_OK);
 398         g_free(req);
 399     } else if (type & VIRTIO_BLK_T_OUT) {
 400         qemu_iovec_init_external(&req->qiov, &req->elem.out_sg[1],
 401                                  req->elem.out_num - 1);
 402         virtio_blk_handle_write(req, mrb);
 403     } else {
 404         qemu_iovec_init_external(&req->qiov, &req->elem.in_sg[0],
 405                                  req->elem.in_num - 1);
 406         virtio_blk_handle_read(req);
 407     }
 408 }
 409
 410 static void virtio_blk_handle_output(VirtIODevice *vdev, VirtQueue *vq)
 411 {
 412     VirtIOBlock *s = to_virtio_blk(vdev);
 413     VirtIOBlockReq *req;
 414     MultiReqBuffer mrb = {
 415         .num_writes = 0,
 416     };
 417
 418     while ((req = virtio_blk_get_request(s))) {
 419         virtio_blk_handle_request(req, &mrb);
 420     }
 421
 422     virtio_submit_multiwrite(s->bs, &mrb);
 423
 424     /*
 425      * FIXME: Want to check for completions before returning to guest mode,
 426      * so cached reads and writes are reported as quickly as possible. But
 427      * that should be done in the generic block layer.
 428      */
 429 }
 430
 431 static void virtio_blk_dma_restart_bh(void *opaque)
 432 {
 433     VirtIOBlock *s = opaque;
 434     VirtIOBlockReq *req = s->rq;
 435     MultiReqBuffer mrb = {
 436         .num_writes = 0,
 437     };
 438
 439     qemu_bh_delete(s->bh);
 440     s->bh = NULL;
 441
 442     s->rq = NULL;
 443
 444     while (req) {
 445         virtio_blk_handle_request(req, &mrb);
 446         req = req->next;
 447     }
 448
 449     virtio_submit_multiwrite(s->bs, &mrb);
 450 }
 451
 452 static void virtio_blk_dma_restart_cb(void *opaque, int running,
 453                                       RunState state)
 454 {
 455     VirtIOBlock *s = opaque;
 456
 457     if (!running)
 458         return;
 459
 460     if (!s->bh) {
 461         s->bh = qemu_bh_new(virtio_blk_dma_restart_bh, s);
 462         qemu_bh_schedule(s->bh);
 463     }
 464 }
 465
 466 static void virtio_blk_reset(VirtIODevice *vdev)
 467 {
 468     /*
 469      * This should cancel pending requests, but can't do nicely until there
 470      * are per-device request lists.
 471      */
 472     bdrv_drain_all();
 473 }
 474
 475 /* coalesce internal state, copy to pci i/o region 0
 476  */
 477 static void virtio_blk_update_config(VirtIODevice *vdev, uint8_t *config)
 478 {
 479     VirtIOBlock *s = to_virtio_blk(vdev);
 480     struct virtio_blk_config blkcfg;
 481     uint64_t capacity;
 482     int cylinders, heads, secs;
 483     int blk_size = s->conf->logical_block_size;
 484
 485     bdrv_get_geometry(s->bs, &capacity);
 486     bdrv_get_geometry_hint(s->bs, &cylinders, &heads, &secs);
 487     memset(&blkcfg, 0, sizeof(blkcfg));
 488     stq_raw(&blkcfg.capacity, capacity);
 489     stl_raw(&blkcfg.seg_max, 128 - 2);
 490     stw_raw(&blkcfg.cylinders, cylinders);
 491     stl_raw(&blkcfg.blk_size, blk_size);
 492     stw_raw(&blkcfg.min_io_size, s->conf->min_io_size / blk_size);
 493     stw_raw(&blkcfg.opt_io_size, s->conf->opt_io_size / blk_size);
 494     blkcfg.heads = heads;
 495     blkcfg.sectors = secs & ~s->sector_mask;
 496     blkcfg.size_max = 0;
 497     blkcfg.physical_block_exp = get_physical_block_exp(s->conf);
 498     blkcfg.alignment_offset = 0;
 499     memcpy(config, &blkcfg, sizeof(struct virtio_blk_config));
 500 }
 501
 502 static uint32_t virtio_blk_get_features(VirtIODevice *vdev, uint32_t features)
 503 {
 504     VirtIOBlock *s = to_virtio_blk(vdev);
 505
 506     features |= (1 << VIRTIO_BLK_F_SEG_MAX);
 507     features |= (1 << VIRTIO_BLK_F_GEOMETRY);
 508     features |= (1 << VIRTIO_BLK_F_TOPOLOGY);
 509     features |= (1 << VIRTIO_BLK_F_BLK_SIZE);
 510
 511     if (bdrv_enable_write_cache(s->bs))
 512         features |= (1 << VIRTIO_BLK_F_WCACHE);
 513
 514     if (bdrv_is_read_only(s->bs))
 515         features |= 1 << VIRTIO_BLK_F_RO;
 516
 517     return features;
 518 }
 519
 520 static void virtio_blk_save(QEMUFile *f, void *opaque)
 521 {
 522     VirtIOBlock *s = opaque;
 523     VirtIOBlockReq *req = s->rq;
 524
 525     virtio_save(&s->vdev, f);
 526
 527     while (req) {
 528         qemu_put_sbyte(f, 1);
 529         qemu_put_buffer(f, (unsigned char*)&req->elem, sizeof(req->elem));
 530         req = req->next;
 531     }
 532     qemu_put_sbyte(f, 0);
 533 }
 534
 535 static int virtio_blk_load(QEMUFile *f, void *opaque, int version_id)
 536 {
 537     VirtIOBlock *s = opaque;
 538
 539     if (version_id != 2)
 540         return -EINVAL;
 541
 542     virtio_load(&s->vdev, f);
 543     while (qemu_get_sbyte(f)) {
 544         VirtIOBlockReq *req = virtio_blk_alloc_request(s);
 545         qemu_get_buffer(f, (unsigned char*)&req->elem, sizeof(req->elem));
 546         req->next = s->rq;
 547         s->rq = req;
 548
 549         virtqueue_map_sg(req->elem.in_sg, req->elem.in_addr,
 550             req->elem.in_num, 1);
 551         virtqueue_map_sg(req->elem.out_sg, req->elem.out_addr,
 552             req->elem.out_num, 0);
 553     }
 554
 555     return 0;
 556 }
 557
 558 static void virtio_blk_resize(void *opaque)
 559 {
 560     VirtIOBlock *s = opaque;
 561
 562     virtio_notify_config(&s->vdev);
 563 }
 564
 565 static const BlockDevOps virtio_block_ops = {
 566     .resize_cb = virtio_blk_resize,
 567 };
 568
 569 VirtIODevice *virtio_blk_init(DeviceState *dev, BlockConf *conf,
 570                               char **serial)
 571 {
 572     VirtIOBlock *s;
 573     int cylinders, heads, secs;
 574     static int virtio_blk_id;
 575     DriveInfo *dinfo;
 576
 577     if (!conf->bs) {
 578         error_report("drive property not set");
 579         return NULL;
 580     }
 581     if (!bdrv_is_inserted(conf->bs)) {
 582         error_report("Device needs media, but drive is empty");
 583         return NULL;
 584     }
 585
 586     if (!*serial) {
 587         /* try to fall back to value set with legacy -drive serial=... */
 588         dinfo = drive_get_by_blockdev(conf->bs);
 589         if (*dinfo->serial) {
 590             *serial = strdup(dinfo->serial);
 591         }
 592     }
 593
 594     s = (VirtIOBlock *)virtio_common_init("virtio-blk", VIRTIO_ID_BLOCK,
 595                                           sizeof(struct virtio_blk_config),
 596                                           sizeof(VirtIOBlock));
 597
 598     s->vdev.get_config = virtio_blk_update_config;
 599     s->vdev.get_features = virtio_blk_get_features;
 600     s->vdev.reset = virtio_blk_reset;
 601     s->bs = conf->bs;
 602     s->conf = conf;
 603     s->serial = *serial;
 604     s->rq = NULL;
 605     s->sector_mask = (s->conf->logical_block_size / BDRV_SECTOR_SIZE) - 1;
 606     bdrv_guess_geometry(s->bs, &cylinders, &heads, &secs);
 607
 608     s->vq = virtio_add_queue(&s->vdev, 128, virtio_blk_handle_output);
 609
 610     qemu_add_vm_change_state_handler(virtio_blk_dma_restart_cb, s);
 611     s->qdev = dev;
 612     register_savevm(dev, "virtio-blk", virtio_blk_id++, 2,
 613                     virtio_blk_save, virtio_blk_load, s);
 614     bdrv_set_dev_ops(s->bs, &virtio_block_ops, s);
 615     bdrv_set_buffer_alignment(s->bs, conf->logical_block_size);
 616
 617     bdrv_iostatus_enable(s->bs);
 618     add_boot_device_path(conf->bootindex, dev, "/disk@0,0");
 619
 620     return &s->vdev;
 621 }
 622
 623 void virtio_blk_exit(VirtIODevice *vdev)
 624 {
 625     VirtIOBlock *s = to_virtio_blk(vdev);
 626     unregister_savevm(s->qdev, "virtio-blk", s);
 627     virtio_cleanup(vdev);
 628 }