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     VirtIOBlkConf *blk;
  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_emit_qmp_error_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_emit_qmp_error_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_emit_qmp_error_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 static void virtio_blk_handle_scsi(VirtIOBlockReq *req)
 149 {
 150     int ret;
 151     int status = VIRTIO_BLK_S_OK;
 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      * The scsi inhdr is placed in the second-to-last input segment, just
 169      * before the regular inhdr.
 170      */
 171     req->scsi = (void *)req->elem.in_sg[req->elem.in_num - 2].iov_base;
 172
 173     if (!req->dev->blk->scsi) {
 174         status = VIRTIO_BLK_S_UNSUPP;
 175         goto fail;
 176     }
 177
 178     /*
 179      * No support for bidirection commands yet.
 180      */
 181     if (req->elem.out_num > 2 && req->elem.in_num > 3) {
 182         status = VIRTIO_BLK_S_UNSUPP;
 183         goto fail;
 184     }
 185
 186 #ifdef __linux__
 187     struct sg_io_hdr hdr;
 188     memset(&hdr, 0, sizeof(struct sg_io_hdr));
 189     hdr.interface_id = 'S';
 190     hdr.cmd_len = req->elem.out_sg[1].iov_len;
 191     hdr.cmdp = req->elem.out_sg[1].iov_base;
 192     hdr.dxfer_len = 0;
 193
 194     if (req->elem.out_num > 2) {
 195         /*
 196          * If there are more than the minimally required 2 output segments
 197          * there is write payload starting from the third iovec.
 198          */
 199         hdr.dxfer_direction = SG_DXFER_TO_DEV;
 200         hdr.iovec_count = req->elem.out_num - 2;
 201
 202         for (i = 0; i < hdr.iovec_count; i++)
 203             hdr.dxfer_len += req->elem.out_sg[i + 2].iov_len;
 204
 205         hdr.dxferp = req->elem.out_sg + 2;
 206
 207     } else if (req->elem.in_num > 3) {
 208         /*
 209          * If we have more than 3 input segments the guest wants to actually
 210          * read data.
 211          */
 212         hdr.dxfer_direction = SG_DXFER_FROM_DEV;
 213         hdr.iovec_count = req->elem.in_num - 3;
 214         for (i = 0; i < hdr.iovec_count; i++)
 215             hdr.dxfer_len += req->elem.in_sg[i].iov_len;
 216
 217         hdr.dxferp = req->elem.in_sg;
 218     } else {
 219         /*
 220          * Some SCSI commands don't actually transfer any data.
 221          */
 222         hdr.dxfer_direction = SG_DXFER_NONE;
 223     }
 224
 225     hdr.sbp = req->elem.in_sg[req->elem.in_num - 3].iov_base;
 226     hdr.mx_sb_len = req->elem.in_sg[req->elem.in_num - 3].iov_len;
 227
 228     ret = bdrv_ioctl(req->dev->bs, SG_IO, &hdr);
 229     if (ret) {
 230         status = VIRTIO_BLK_S_UNSUPP;
 231         goto fail;
 232     }
 233
 234     /*
 235      * From SCSI-Generic-HOWTO: "Some lower level drivers (e.g. ide-scsi)
 236      * clear the masked_status field [hence status gets cleared too, see
 237      * block/scsi_ioctl.c] even when a CHECK_CONDITION or COMMAND_TERMINATED
 238      * status has occurred.  However they do set DRIVER_SENSE in driver_status
 239      * field. Also a (sb_len_wr > 0) indicates there is a sense buffer.
 240      */
 241     if (hdr.status == 0 && hdr.sb_len_wr > 0) {
 242         hdr.status = CHECK_CONDITION;
 243     }
 244
 245     stl_p(&req->scsi->errors,
 246           hdr.status | (hdr.msg_status << 8) |
 247           (hdr.host_status << 16) | (hdr.driver_status << 24));
 248     stl_p(&req->scsi->residual, hdr.resid);
 249     stl_p(&req->scsi->sense_len, hdr.sb_len_wr);
 250     stl_p(&req->scsi->data_len, hdr.dxfer_len);
 251
 252     virtio_blk_req_complete(req, status);
 253     g_free(req);
 254 #else
 255     abort();
 256 #endif
 257
 258 fail:
 259     /* Just put anything nonzero so that the ioctl fails in the guest.  */
 260     stl_p(&req->scsi->errors, 255);
 261     virtio_blk_req_complete(req, status);
 262     g_free(req);
 263 }
 264
 265 typedef struct MultiReqBuffer {
 266     BlockRequest        blkreq[32];
 267     unsigned int        num_writes;
 268 } MultiReqBuffer;
 269
 270 static void virtio_submit_multiwrite(BlockDriverState *bs, MultiReqBuffer *mrb)
 271 {
 272     int i, ret;
 273
 274     if (!mrb->num_writes) {
 275         return;
 276     }
 277
 278     ret = bdrv_aio_multiwrite(bs, mrb->blkreq, mrb->num_writes);
 279     if (ret != 0) {
 280         for (i = 0; i < mrb->num_writes; i++) {
 281             if (mrb->blkreq[i].error) {
 282                 virtio_blk_rw_complete(mrb->blkreq[i].opaque, -EIO);
 283             }
 284         }
 285     }
 286
 287     mrb->num_writes = 0;
 288 }
 289
 290 static void virtio_blk_handle_flush(VirtIOBlockReq *req, MultiReqBuffer *mrb)
 291 {
 292     bdrv_acct_start(req->dev->bs, &req->acct, 0, BDRV_ACCT_FLUSH);
 293
 294     /*
 295      * Make sure all outstanding writes are posted to the backing device.
 296      */
 297     virtio_submit_multiwrite(req->dev->bs, mrb);
 298     bdrv_aio_flush(req->dev->bs, virtio_blk_flush_complete, req);
 299 }
 300
 301 static void virtio_blk_handle_write(VirtIOBlockReq *req, MultiReqBuffer *mrb)
 302 {
 303     BlockRequest *blkreq;
 304     uint64_t sector;
 305
 306     sector = ldq_p(&req->out->sector);
 307
 308     bdrv_acct_start(req->dev->bs, &req->acct, req->qiov.size, BDRV_ACCT_WRITE);
 309
 310     trace_virtio_blk_handle_write(req, sector, req->qiov.size / 512);
 311
 312     if (sector & req->dev->sector_mask) {
 313         virtio_blk_rw_complete(req, -EIO);
 314         return;
 315     }
 316     if (req->qiov.size % req->dev->conf->logical_block_size) {
 317         virtio_blk_rw_complete(req, -EIO);
 318         return;
 319     }
 320
 321     if (mrb->num_writes == 32) {
 322         virtio_submit_multiwrite(req->dev->bs, mrb);
 323     }
 324
 325     blkreq = &mrb->blkreq[mrb->num_writes];
 326     blkreq->sector = sector;
 327     blkreq->nb_sectors = req->qiov.size / BDRV_SECTOR_SIZE;
 328     blkreq->qiov = &req->qiov;
 329     blkreq->cb = virtio_blk_rw_complete;
 330     blkreq->opaque = req;
 331     blkreq->error = 0;
 332
 333     mrb->num_writes++;
 334 }
 335
 336 static void virtio_blk_handle_read(VirtIOBlockReq *req)
 337 {
 338     uint64_t sector;
 339
 340     sector = ldq_p(&req->out->sector);
 341
 342     bdrv_acct_start(req->dev->bs, &req->acct, req->qiov.size, BDRV_ACCT_READ);
 343
 344     trace_virtio_blk_handle_read(req, sector, req->qiov.size / 512);
 345
 346     if (sector & req->dev->sector_mask) {
 347         virtio_blk_rw_complete(req, -EIO);
 348         return;
 349     }
 350     if (req->qiov.size % req->dev->conf->logical_block_size) {
 351         virtio_blk_rw_complete(req, -EIO);
 352         return;
 353     }
 354     bdrv_aio_readv(req->dev->bs, sector, &req->qiov,
 355                    req->qiov.size / BDRV_SECTOR_SIZE,
 356                    virtio_blk_rw_complete, req);
 357 }
 358
 359 static void virtio_blk_handle_request(VirtIOBlockReq *req,
 360     MultiReqBuffer *mrb)
 361 {
 362     uint32_t type;
 363
 364     if (req->elem.out_num < 1 || req->elem.in_num < 1) {
 365         error_report("virtio-blk missing headers");
 366         exit(1);
 367     }
 368
 369     if (req->elem.out_sg[0].iov_len < sizeof(*req->out) ||
 370         req->elem.in_sg[req->elem.in_num - 1].iov_len < sizeof(*req->in)) {
 371         error_report("virtio-blk header not in correct element");
 372         exit(1);
 373     }
 374
 375     req->out = (void *)req->elem.out_sg[0].iov_base;
 376     req->in = (void *)req->elem.in_sg[req->elem.in_num - 1].iov_base;
 377
 378     type = ldl_p(&req->out->type);
 379
 380     if (type & VIRTIO_BLK_T_FLUSH) {
 381         virtio_blk_handle_flush(req, mrb);
 382     } else if (type & VIRTIO_BLK_T_SCSI_CMD) {
 383         virtio_blk_handle_scsi(req);
 384     } else if (type & VIRTIO_BLK_T_GET_ID) {
 385         VirtIOBlock *s = req->dev;
 386
 387         /*
 388          * NB: per existing s/n string convention the string is
 389          * terminated by '\0' only when shorter than buffer.
 390          */
 391         strncpy(req->elem.in_sg[0].iov_base,
 392                 s->blk->serial ? s->blk->serial : "",
 393                 MIN(req->elem.in_sg[0].iov_len, VIRTIO_BLK_ID_BYTES));
 394         virtio_blk_req_complete(req, VIRTIO_BLK_S_OK);
 395         g_free(req);
 396     } else if (type & VIRTIO_BLK_T_OUT) {
 397         qemu_iovec_init_external(&req->qiov, &req->elem.out_sg[1],
 398                                  req->elem.out_num - 1);
 399         virtio_blk_handle_write(req, mrb);
 400     } else {
 401         qemu_iovec_init_external(&req->qiov, &req->elem.in_sg[0],
 402                                  req->elem.in_num - 1);
 403         virtio_blk_handle_read(req);
 404     }
 405 }
 406
 407 static void virtio_blk_handle_output(VirtIODevice *vdev, VirtQueue *vq)
 408 {
 409     VirtIOBlock *s = to_virtio_blk(vdev);
 410     VirtIOBlockReq *req;
 411     MultiReqBuffer mrb = {
 412         .num_writes = 0,
 413     };
 414
 415     while ((req = virtio_blk_get_request(s))) {
 416         virtio_blk_handle_request(req, &mrb);
 417     }
 418
 419     virtio_submit_multiwrite(s->bs, &mrb);
 420
 421     /*
 422      * FIXME: Want to check for completions before returning to guest mode,
 423      * so cached reads and writes are reported as quickly as possible. But
 424      * that should be done in the generic block layer.
 425      */
 426 }
 427
 428 static void virtio_blk_dma_restart_bh(void *opaque)
 429 {
 430     VirtIOBlock *s = opaque;
 431     VirtIOBlockReq *req = s->rq;
 432     MultiReqBuffer mrb = {
 433         .num_writes = 0,
 434     };
 435
 436     qemu_bh_delete(s->bh);
 437     s->bh = NULL;
 438
 439     s->rq = NULL;
 440
 441     while (req) {
 442         virtio_blk_handle_request(req, &mrb);
 443         req = req->next;
 444     }
 445
 446     virtio_submit_multiwrite(s->bs, &mrb);
 447 }
 448
 449 static void virtio_blk_dma_restart_cb(void *opaque, int running,
 450                                       RunState state)
 451 {
 452     VirtIOBlock *s = opaque;
 453
 454     if (!running)
 455         return;
 456
 457     if (!s->bh) {
 458         s->bh = qemu_bh_new(virtio_blk_dma_restart_bh, s);
 459         qemu_bh_schedule(s->bh);
 460     }
 461 }
 462
 463 static void virtio_blk_reset(VirtIODevice *vdev)
 464 {
 465     /*
 466      * This should cancel pending requests, but can't do nicely until there
 467      * are per-device request lists.
 468      */
 469     bdrv_drain_all();
 470 }
 471
 472 /* coalesce internal state, copy to pci i/o region 0
 473  */
 474 static void virtio_blk_update_config(VirtIODevice *vdev, uint8_t *config)
 475 {
 476     VirtIOBlock *s = to_virtio_blk(vdev);
 477     struct virtio_blk_config blkcfg;
 478     uint64_t capacity;
 479     int cylinders, heads, secs;
 480     int blk_size = s->conf->logical_block_size;
 481
 482     bdrv_get_geometry(s->bs, &capacity);
 483     bdrv_get_geometry_hint(s->bs, &cylinders, &heads, &secs);
 484     memset(&blkcfg, 0, sizeof(blkcfg));
 485     stq_raw(&blkcfg.capacity, capacity);
 486     stl_raw(&blkcfg.seg_max, 128 - 2);
 487     stw_raw(&blkcfg.cylinders, cylinders);
 488     stl_raw(&blkcfg.blk_size, blk_size);
 489     stw_raw(&blkcfg.min_io_size, s->conf->min_io_size / blk_size);
 490     stw_raw(&blkcfg.opt_io_size, s->conf->opt_io_size / blk_size);
 491     blkcfg.heads = heads;
 492     blkcfg.sectors = secs & ~s->sector_mask;
 493     blkcfg.size_max = 0;
 494     blkcfg.physical_block_exp = get_physical_block_exp(s->conf);
 495     blkcfg.alignment_offset = 0;
 496     memcpy(config, &blkcfg, sizeof(struct virtio_blk_config));
 497 }
 498
 499 static uint32_t virtio_blk_get_features(VirtIODevice *vdev, uint32_t features)
 500 {
 501     VirtIOBlock *s = to_virtio_blk(vdev);
 502
 503     features |= (1 << VIRTIO_BLK_F_SEG_MAX);
 504     features |= (1 << VIRTIO_BLK_F_GEOMETRY);
 505     features |= (1 << VIRTIO_BLK_F_TOPOLOGY);
 506     features |= (1 << VIRTIO_BLK_F_BLK_SIZE);
 507     features |= (1 << VIRTIO_BLK_F_SCSI);
 508
 509     if (bdrv_enable_write_cache(s->bs))
 510         features |= (1 << VIRTIO_BLK_F_WCACHE);
 511
 512     if (bdrv_is_read_only(s->bs))
 513         features |= 1 << VIRTIO_BLK_F_RO;
 514
 515     return features;
 516 }
 517
 518 static void virtio_blk_save(QEMUFile *f, void *opaque)
 519 {
 520     VirtIOBlock *s = opaque;
 521     VirtIOBlockReq *req = s->rq;
 522
 523     virtio_save(&s->vdev, f);
 524
 525     while (req) {
 526         qemu_put_sbyte(f, 1);
 527         qemu_put_buffer(f, (unsigned char*)&req->elem, sizeof(req->elem));
 528         req = req->next;
 529     }
 530     qemu_put_sbyte(f, 0);
 531 }
 532
 533 static int virtio_blk_load(QEMUFile *f, void *opaque, int version_id)
 534 {
 535     VirtIOBlock *s = opaque;
 536
 537     if (version_id != 2)
 538         return -EINVAL;
 539
 540     virtio_load(&s->vdev, f);
 541     while (qemu_get_sbyte(f)) {
 542         VirtIOBlockReq *req = virtio_blk_alloc_request(s);
 543         qemu_get_buffer(f, (unsigned char*)&req->elem, sizeof(req->elem));
 544         req->next = s->rq;
 545         s->rq = req;
 546
 547         virtqueue_map_sg(req->elem.in_sg, req->elem.in_addr,
 548             req->elem.in_num, 1);
 549         virtqueue_map_sg(req->elem.out_sg, req->elem.out_addr,
 550             req->elem.out_num, 0);
 551     }
 552
 553     return 0;
 554 }
 555
 556 static void virtio_blk_resize(void *opaque)
 557 {
 558     VirtIOBlock *s = opaque;
 559
 560     virtio_notify_config(&s->vdev);
 561 }
 562
 563 static const BlockDevOps virtio_block_ops = {
 564     .resize_cb = virtio_blk_resize,
 565 };
 566
 567 VirtIODevice *virtio_blk_init(DeviceState *dev, VirtIOBlkConf *blk)
 568 {
 569     VirtIOBlock *s;
 570     int cylinders, heads, secs;
 571     static int virtio_blk_id;
 572     DriveInfo *dinfo;
 573
 574     if (!blk->conf.bs) {
 575         error_report("drive property not set");
 576         return NULL;
 577     }
 578     if (!bdrv_is_inserted(blk->conf.bs)) {
 579         error_report("Device needs media, but drive is empty");
 580         return NULL;
 581     }
 582
 583     if (!blk->serial) {
 584         /* try to fall back to value set with legacy -drive serial=... */
 585         dinfo = drive_get_by_blockdev(blk->conf.bs);
 586         if (*dinfo->serial) {
 587             blk->serial = strdup(dinfo->serial);
 588         }
 589     }
 590
 591     s = (VirtIOBlock *)virtio_common_init("virtio-blk", VIRTIO_ID_BLOCK,
 592                                           sizeof(struct virtio_blk_config),
 593                                           sizeof(VirtIOBlock));
 594
 595     s->vdev.get_config = virtio_blk_update_config;
 596     s->vdev.get_features = virtio_blk_get_features;
 597     s->vdev.reset = virtio_blk_reset;
 598     s->bs = blk->conf.bs;
 599     s->conf = &blk->conf;
 600     s->blk = blk;
 601     s->rq = NULL;
 602     s->sector_mask = (s->conf->logical_block_size / BDRV_SECTOR_SIZE) - 1;
 603     bdrv_guess_geometry(s->bs, &cylinders, &heads, &secs);
 604
 605     s->vq = virtio_add_queue(&s->vdev, 128, virtio_blk_handle_output);
 606
 607     qemu_add_vm_change_state_handler(virtio_blk_dma_restart_cb, s);
 608     s->qdev = dev;
 609     register_savevm(dev, "virtio-blk", virtio_blk_id++, 2,
 610                     virtio_blk_save, virtio_blk_load, s);
 611     bdrv_set_dev_ops(s->bs, &virtio_block_ops, s);
 612     bdrv_set_buffer_alignment(s->bs, s->conf->logical_block_size);
 613
 614     bdrv_iostatus_enable(s->bs);
 615     add_boot_device_path(s->conf->bootindex, dev, "/disk@0,0");
 616
 617     return &s->vdev;
 618 }
 619
 620 void virtio_blk_exit(VirtIODevice *vdev)
 621 {
 622     VirtIOBlock *s = to_virtio_blk(vdev);
 623     unregister_savevm(s->qdev, "virtio-blk", s);
 624     blockdev_mark_auto_del(s->bs);
 625     virtio_cleanup(vdev);
 626 }