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     bdrv_acct_start(req->dev->bs, &req->acct, 0, BDRV_ACCT_FLUSH);
 292
 293     /*
 294      * Make sure all outstanding writes are posted to the backing device.
 295      */
 296     virtio_submit_multiwrite(req->dev->bs, mrb);
 297     bdrv_aio_flush(req->dev->bs, virtio_blk_flush_complete, req);
 298 }
 299
 300 static void virtio_blk_handle_write(VirtIOBlockReq *req, MultiReqBuffer *mrb)
 301 {
 302     BlockRequest *blkreq;
 303     uint64_t sector;
 304
 305     sector = ldq_p(&req->out->sector);
 306
 307     bdrv_acct_start(req->dev->bs, &req->acct, req->qiov.size, BDRV_ACCT_WRITE);
 308
 309     trace_virtio_blk_handle_write(req, sector, req->qiov.size / 512);
 310
 311     if (sector & req->dev->sector_mask) {
 312         virtio_blk_rw_complete(req, -EIO);
 313         return;
 314     }
 315     if (req->qiov.size % req->dev->conf->logical_block_size) {
 316         virtio_blk_rw_complete(req, -EIO);
 317         return;
 318     }
 319
 320     if (mrb->num_writes == 32) {
 321         virtio_submit_multiwrite(req->dev->bs, mrb);
 322     }
 323
 324     blkreq = &mrb->blkreq[mrb->num_writes];
 325     blkreq->sector = sector;
 326     blkreq->nb_sectors = req->qiov.size / BDRV_SECTOR_SIZE;
 327     blkreq->qiov = &req->qiov;
 328     blkreq->cb = virtio_blk_rw_complete;
 329     blkreq->opaque = req;
 330     blkreq->error = 0;
 331
 332     mrb->num_writes++;
 333 }
 334
 335 static void virtio_blk_handle_read(VirtIOBlockReq *req)
 336 {
 337     uint64_t sector;
 338
 339     sector = ldq_p(&req->out->sector);
 340
 341     bdrv_acct_start(req->dev->bs, &req->acct, req->qiov.size, BDRV_ACCT_READ);
 342
 343     if (sector & req->dev->sector_mask) {
 344         virtio_blk_rw_complete(req, -EIO);
 345         return;
 346     }
 347     if (req->qiov.size % req->dev->conf->logical_block_size) {
 348         virtio_blk_rw_complete(req, -EIO);
 349         return;
 350     }
 351     bdrv_aio_readv(req->dev->bs, sector, &req->qiov,
 352                    req->qiov.size / BDRV_SECTOR_SIZE,
 353                    virtio_blk_rw_complete, req);
 354 }
 355
 356 static void virtio_blk_handle_request(VirtIOBlockReq *req,
 357     MultiReqBuffer *mrb)
 358 {
 359     uint32_t type;
 360
 361     if (req->elem.out_num < 1 || req->elem.in_num < 1) {
 362         error_report("virtio-blk missing headers");
 363         exit(1);
 364     }
 365
 366     if (req->elem.out_sg[0].iov_len < sizeof(*req->out) ||
 367         req->elem.in_sg[req->elem.in_num - 1].iov_len < sizeof(*req->in)) {
 368         error_report("virtio-blk header not in correct element");
 369         exit(1);
 370     }
 371
 372     req->out = (void *)req->elem.out_sg[0].iov_base;
 373     req->in = (void *)req->elem.in_sg[req->elem.in_num - 1].iov_base;
 374
 375     type = ldl_p(&req->out->type);
 376
 377     if (type & VIRTIO_BLK_T_FLUSH) {
 378         virtio_blk_handle_flush(req, mrb);
 379     } else if (type & VIRTIO_BLK_T_SCSI_CMD) {
 380         virtio_blk_handle_scsi(req);
 381     } else if (type & VIRTIO_BLK_T_GET_ID) {
 382         VirtIOBlock *s = req->dev;
 383
 384         /*
 385          * NB: per existing s/n string convention the string is
 386          * terminated by '\0' only when shorter than buffer.
 387          */
 388         strncpy(req->elem.in_sg[0].iov_base,
 389                 s->serial ? s->serial : "",
 390                 MIN(req->elem.in_sg[0].iov_len, VIRTIO_BLK_ID_BYTES));
 391         virtio_blk_req_complete(req, VIRTIO_BLK_S_OK);
 392         g_free(req);
 393     } else if (type & VIRTIO_BLK_T_OUT) {
 394         qemu_iovec_init_external(&req->qiov, &req->elem.out_sg[1],
 395                                  req->elem.out_num - 1);
 396         virtio_blk_handle_write(req, mrb);
 397     } else {
 398         qemu_iovec_init_external(&req->qiov, &req->elem.in_sg[0],
 399                                  req->elem.in_num - 1);
 400         virtio_blk_handle_read(req);
 401     }
 402 }
 403
 404 static void virtio_blk_handle_output(VirtIODevice *vdev, VirtQueue *vq)
 405 {
 406     VirtIOBlock *s = to_virtio_blk(vdev);
 407     VirtIOBlockReq *req;
 408     MultiReqBuffer mrb = {
 409         .num_writes = 0,
 410     };
 411
 412     while ((req = virtio_blk_get_request(s))) {
 413         virtio_blk_handle_request(req, &mrb);
 414     }
 415
 416     virtio_submit_multiwrite(s->bs, &mrb);
 417
 418     /*
 419      * FIXME: Want to check for completions before returning to guest mode,
 420      * so cached reads and writes are reported as quickly as possible. But
 421      * that should be done in the generic block layer.
 422      */
 423 }
 424
 425 static void virtio_blk_dma_restart_bh(void *opaque)
 426 {
 427     VirtIOBlock *s = opaque;
 428     VirtIOBlockReq *req = s->rq;
 429     MultiReqBuffer mrb = {
 430         .num_writes = 0,
 431     };
 432
 433     qemu_bh_delete(s->bh);
 434     s->bh = NULL;
 435
 436     s->rq = NULL;
 437
 438     while (req) {
 439         virtio_blk_handle_request(req, &mrb);
 440         req = req->next;
 441     }
 442
 443     virtio_submit_multiwrite(s->bs, &mrb);
 444 }
 445
 446 static void virtio_blk_dma_restart_cb(void *opaque, int running,
 447                                       RunState state)
 448 {
 449     VirtIOBlock *s = opaque;
 450
 451     if (!running)
 452         return;
 453
 454     if (!s->bh) {
 455         s->bh = qemu_bh_new(virtio_blk_dma_restart_bh, s);
 456         qemu_bh_schedule(s->bh);
 457     }
 458 }
 459
 460 static void virtio_blk_reset(VirtIODevice *vdev)
 461 {
 462     /*
 463      * This should cancel pending requests, but can't do nicely until there
 464      * are per-device request lists.
 465      */
 466     bdrv_drain_all();
 467 }
 468
 469 /* coalesce internal state, copy to pci i/o region 0
 470  */
 471 static void virtio_blk_update_config(VirtIODevice *vdev, uint8_t *config)
 472 {
 473     VirtIOBlock *s = to_virtio_blk(vdev);
 474     struct virtio_blk_config blkcfg;
 475     uint64_t capacity;
 476     int cylinders, heads, secs;
 477     int blk_size = s->conf->logical_block_size;
 478
 479     bdrv_get_geometry(s->bs, &capacity);
 480     bdrv_get_geometry_hint(s->bs, &cylinders, &heads, &secs);
 481     memset(&blkcfg, 0, sizeof(blkcfg));
 482     stq_raw(&blkcfg.capacity, capacity);
 483     stl_raw(&blkcfg.seg_max, 128 - 2);
 484     stw_raw(&blkcfg.cylinders, cylinders);
 485     stl_raw(&blkcfg.blk_size, blk_size);
 486     stw_raw(&blkcfg.min_io_size, s->conf->min_io_size / blk_size);
 487     stw_raw(&blkcfg.opt_io_size, s->conf->opt_io_size / blk_size);
 488     blkcfg.heads = heads;
 489     blkcfg.sectors = secs & ~s->sector_mask;
 490     blkcfg.size_max = 0;
 491     blkcfg.physical_block_exp = get_physical_block_exp(s->conf);
 492     blkcfg.alignment_offset = 0;
 493     memcpy(config, &blkcfg, sizeof(struct virtio_blk_config));
 494 }
 495
 496 static uint32_t virtio_blk_get_features(VirtIODevice *vdev, uint32_t features)
 497 {
 498     VirtIOBlock *s = to_virtio_blk(vdev);
 499
 500     features |= (1 << VIRTIO_BLK_F_SEG_MAX);
 501     features |= (1 << VIRTIO_BLK_F_GEOMETRY);
 502     features |= (1 << VIRTIO_BLK_F_TOPOLOGY);
 503     features |= (1 << VIRTIO_BLK_F_BLK_SIZE);
 504
 505     if (bdrv_enable_write_cache(s->bs))
 506         features |= (1 << VIRTIO_BLK_F_WCACHE);
 507
 508     if (bdrv_is_read_only(s->bs))
 509         features |= 1 << VIRTIO_BLK_F_RO;
 510
 511     return features;
 512 }
 513
 514 static void virtio_blk_save(QEMUFile *f, void *opaque)
 515 {
 516     VirtIOBlock *s = opaque;
 517     VirtIOBlockReq *req = s->rq;
 518
 519     virtio_save(&s->vdev, f);
 520
 521     while (req) {
 522         qemu_put_sbyte(f, 1);
 523         qemu_put_buffer(f, (unsigned char*)&req->elem, sizeof(req->elem));
 524         req = req->next;
 525     }
 526     qemu_put_sbyte(f, 0);
 527 }
 528
 529 static int virtio_blk_load(QEMUFile *f, void *opaque, int version_id)
 530 {
 531     VirtIOBlock *s = opaque;
 532
 533     if (version_id != 2)
 534         return -EINVAL;
 535
 536     virtio_load(&s->vdev, f);
 537     while (qemu_get_sbyte(f)) {
 538         VirtIOBlockReq *req = virtio_blk_alloc_request(s);
 539         qemu_get_buffer(f, (unsigned char*)&req->elem, sizeof(req->elem));
 540         req->next = s->rq;
 541         s->rq = req;
 542
 543         virtqueue_map_sg(req->elem.in_sg, req->elem.in_addr,
 544             req->elem.in_num, 1);
 545         virtqueue_map_sg(req->elem.out_sg, req->elem.out_addr,
 546             req->elem.out_num, 0);
 547     }
 548
 549     return 0;
 550 }
 551
 552 static void virtio_blk_resize(void *opaque)
 553 {
 554     VirtIOBlock *s = opaque;
 555
 556     virtio_notify_config(&s->vdev);
 557 }
 558
 559 static const BlockDevOps virtio_block_ops = {
 560     .resize_cb = virtio_blk_resize,
 561 };
 562
 563 VirtIODevice *virtio_blk_init(DeviceState *dev, BlockConf *conf,
 564                               char **serial)
 565 {
 566     VirtIOBlock *s;
 567     int cylinders, heads, secs;
 568     static int virtio_blk_id;
 569     DriveInfo *dinfo;
 570
 571     if (!conf->bs) {
 572         error_report("virtio-blk-pci: drive property not set");
 573         return NULL;
 574     }
 575     if (!bdrv_is_inserted(conf->bs)) {
 576         error_report("Device needs media, but drive is empty");
 577         return NULL;
 578     }
 579
 580     if (!*serial) {
 581         /* try to fall back to value set with legacy -drive serial=... */
 582         dinfo = drive_get_by_blockdev(conf->bs);
 583         if (*dinfo->serial) {
 584             *serial = strdup(dinfo->serial);
 585         }
 586     }
 587
 588     s = (VirtIOBlock *)virtio_common_init("virtio-blk", VIRTIO_ID_BLOCK,
 589                                           sizeof(struct virtio_blk_config),
 590                                           sizeof(VirtIOBlock));
 591
 592     s->vdev.get_config = virtio_blk_update_config;
 593     s->vdev.get_features = virtio_blk_get_features;
 594     s->vdev.reset = virtio_blk_reset;
 595     s->bs = conf->bs;
 596     s->conf = conf;
 597     s->serial = *serial;
 598     s->rq = NULL;
 599     s->sector_mask = (s->conf->logical_block_size / BDRV_SECTOR_SIZE) - 1;
 600     bdrv_guess_geometry(s->bs, &cylinders, &heads, &secs);
 601
 602     s->vq = virtio_add_queue(&s->vdev, 128, virtio_blk_handle_output);
 603
 604     qemu_add_vm_change_state_handler(virtio_blk_dma_restart_cb, s);
 605     s->qdev = dev;
 606     register_savevm(dev, "virtio-blk", virtio_blk_id++, 2,
 607                     virtio_blk_save, virtio_blk_load, s);
 608     bdrv_set_dev_ops(s->bs, &virtio_block_ops, s);
 609     bdrv_set_buffer_alignment(s->bs, conf->logical_block_size);
 610
 611     bdrv_iostatus_enable(s->bs);
 612     add_boot_device_path(conf->bootindex, dev, "/disk@0,0");
 613
 614     return &s->vdev;
 615 }
 616
 617 void virtio_blk_exit(VirtIODevice *vdev)
 618 {
 619     VirtIOBlock *s = to_virtio_blk(vdev);
 620     unregister_savevm(s->qdev, "virtio-blk", s);
 621     virtio_cleanup(vdev);
 622 }