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