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     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     req->scsi->errors = hdr.status;
 227     req->scsi->residual = hdr.resid;
 228     req->scsi->sense_len = hdr.sb_len_wr;
 229     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_req_complete(req, VIRTIO_BLK_S_IOERR);
 277     }
 278 }
 279
 280 static void virtio_blk_handle_write(VirtIOBlockReq *req, MultiReqBuffer *mrb)
 281 {
 282     BlockRequest *blkreq;
 283
 284     trace_virtio_blk_handle_write(req, req->out->sector, req->qiov.size / 512);
 285
 286     if (req->out->sector & req->dev->sector_mask) {
 287         virtio_blk_rw_complete(req, -EIO);
 288         return;
 289     }
 290
 291     if (mrb->num_writes == 32) {
 292         virtio_submit_multiwrite(req->dev->bs, mrb);
 293     }
 294
 295     blkreq = &mrb->blkreq[mrb->num_writes];
 296     blkreq->sector = req->out->sector;
 297     blkreq->nb_sectors = req->qiov.size / BDRV_SECTOR_SIZE;
 298     blkreq->qiov = &req->qiov;
 299     blkreq->cb = virtio_blk_rw_complete;
 300     blkreq->opaque = req;
 301     blkreq->error = 0;
 302
 303     mrb->num_writes++;
 304 }
 305
 306 static void virtio_blk_handle_read(VirtIOBlockReq *req)
 307 {
 308     BlockDriverAIOCB *acb;
 309
 310     if (req->out->sector & req->dev->sector_mask) {
 311         virtio_blk_rw_complete(req, -EIO);
 312         return;
 313     }
 314
 315     acb = bdrv_aio_readv(req->dev->bs, req->out->sector, &req->qiov,
 316                          req->qiov.size / BDRV_SECTOR_SIZE,
 317                          virtio_blk_rw_complete, req);
 318     if (!acb) {
 319         virtio_blk_rw_complete(req, -EIO);
 320     }
 321 }
 322
 323 static void virtio_blk_handle_request(VirtIOBlockReq *req,
 324     MultiReqBuffer *mrb)
 325 {
 326     if (req->elem.out_num < 1 || req->elem.in_num < 1) {
 327         fprintf(stderr, "virtio-blk missing headers\n");
 328         exit(1);
 329     }
 330
 331     if (req->elem.out_sg[0].iov_len < sizeof(*req->out) ||
 332         req->elem.in_sg[req->elem.in_num - 1].iov_len < sizeof(*req->in)) {
 333         fprintf(stderr, "virtio-blk header not in correct element\n");
 334         exit(1);
 335     }
 336
 337     req->out = (void *)req->elem.out_sg[0].iov_base;
 338     req->in = (void *)req->elem.in_sg[req->elem.in_num - 1].iov_base;
 339
 340     if (req->out->type & VIRTIO_BLK_T_FLUSH) {
 341         virtio_blk_handle_flush(req, mrb);
 342     } else if (req->out->type & VIRTIO_BLK_T_SCSI_CMD) {
 343         virtio_blk_handle_scsi(req);
 344     } else if (req->out->type & VIRTIO_BLK_T_GET_ID) {
 345         VirtIOBlock *s = req->dev;
 346
 347         memcpy(req->elem.in_sg[0].iov_base, s->sn,
 348                MIN(req->elem.in_sg[0].iov_len, sizeof(s->sn)));
 349         virtio_blk_req_complete(req, VIRTIO_BLK_S_OK);
 350     } else if (req->out->type & VIRTIO_BLK_T_OUT) {
 351         qemu_iovec_init_external(&req->qiov, &req->elem.out_sg[1],
 352                                  req->elem.out_num - 1);
 353         virtio_blk_handle_write(req, mrb);
 354     } else {
 355         qemu_iovec_init_external(&req->qiov, &req->elem.in_sg[0],
 356                                  req->elem.in_num - 1);
 357         virtio_blk_handle_read(req);
 358     }
 359 }
 360
 361 static void virtio_blk_handle_output(VirtIODevice *vdev, VirtQueue *vq)
 362 {
 363     VirtIOBlock *s = to_virtio_blk(vdev);
 364     VirtIOBlockReq *req;
 365     MultiReqBuffer mrb = {
 366         .num_writes = 0,
 367     };
 368
 369     while ((req = virtio_blk_get_request(s))) {
 370         virtio_blk_handle_request(req, &mrb);
 371     }
 372
 373     virtio_submit_multiwrite(s->bs, &mrb);
 374
 375     /*
 376      * FIXME: Want to check for completions before returning to guest mode,
 377      * so cached reads and writes are reported as quickly as possible. But
 378      * that should be done in the generic block layer.
 379      */
 380 }
 381
 382 static void virtio_blk_dma_restart_bh(void *opaque)
 383 {
 384     VirtIOBlock *s = opaque;
 385     VirtIOBlockReq *req = s->rq;
 386     MultiReqBuffer mrb = {
 387         .num_writes = 0,
 388     };
 389
 390     qemu_bh_delete(s->bh);
 391     s->bh = NULL;
 392
 393     s->rq = NULL;
 394
 395     while (req) {
 396         virtio_blk_handle_request(req, &mrb);
 397         req = req->next;
 398     }
 399
 400     virtio_submit_multiwrite(s->bs, &mrb);
 401 }
 402
 403 static void virtio_blk_dma_restart_cb(void *opaque, int running, int reason)
 404 {
 405     VirtIOBlock *s = opaque;
 406
 407     if (!running)
 408         return;
 409
 410     if (!s->bh) {
 411         s->bh = qemu_bh_new(virtio_blk_dma_restart_bh, s);
 412         qemu_bh_schedule(s->bh);
 413     }
 414 }
 415
 416 static void virtio_blk_reset(VirtIODevice *vdev)
 417 {
 418     /*
 419      * This should cancel pending requests, but can't do nicely until there
 420      * are per-device request lists.
 421      */
 422     qemu_aio_flush();
 423 }
 424
 425 /* coalesce internal state, copy to pci i/o region 0
 426  */
 427 static void virtio_blk_update_config(VirtIODevice *vdev, uint8_t *config)
 428 {
 429     VirtIOBlock *s = to_virtio_blk(vdev);
 430     struct virtio_blk_config blkcfg;
 431     uint64_t capacity;
 432     int cylinders, heads, secs;
 433
 434     bdrv_get_geometry(s->bs, &capacity);
 435     bdrv_get_geometry_hint(s->bs, &cylinders, &heads, &secs);
 436     memset(&blkcfg, 0, sizeof(blkcfg));
 437     stq_raw(&blkcfg.capacity, capacity);
 438     stl_raw(&blkcfg.seg_max, 128 - 2);
 439     stw_raw(&blkcfg.cylinders, cylinders);
 440     blkcfg.heads = heads;
 441     blkcfg.sectors = secs & ~s->sector_mask;
 442     blkcfg.blk_size = s->conf->logical_block_size;
 443     blkcfg.size_max = 0;
 444     blkcfg.physical_block_exp = get_physical_block_exp(s->conf);
 445     blkcfg.alignment_offset = 0;
 446     blkcfg.min_io_size = s->conf->min_io_size / blkcfg.blk_size;
 447     blkcfg.opt_io_size = s->conf->opt_io_size / blkcfg.blk_size;
 448     memcpy(config, &blkcfg, sizeof(struct virtio_blk_config));
 449 }
 450
 451 static uint32_t virtio_blk_get_features(VirtIODevice *vdev, uint32_t features)
 452 {
 453     VirtIOBlock *s = to_virtio_blk(vdev);
 454
 455     features |= (1 << VIRTIO_BLK_F_SEG_MAX);
 456     features |= (1 << VIRTIO_BLK_F_GEOMETRY);
 457     features |= (1 << VIRTIO_BLK_F_TOPOLOGY);
 458     features |= (1 << VIRTIO_BLK_F_BLK_SIZE);
 459
 460     if (bdrv_enable_write_cache(s->bs))
 461         features |= (1 << VIRTIO_BLK_F_WCACHE);
 462
 463     if (bdrv_is_read_only(s->bs))
 464         features |= 1 << VIRTIO_BLK_F_RO;
 465
 466     return features;
 467 }
 468
 469 static void virtio_blk_save(QEMUFile *f, void *opaque)
 470 {
 471     VirtIOBlock *s = opaque;
 472     VirtIOBlockReq *req = s->rq;
 473
 474     virtio_save(&s->vdev, f);
 475
 476     while (req) {
 477         qemu_put_sbyte(f, 1);
 478         qemu_put_buffer(f, (unsigned char*)&req->elem, sizeof(req->elem));
 479         req = req->next;
 480     }
 481     qemu_put_sbyte(f, 0);
 482 }
 483
 484 static int virtio_blk_load(QEMUFile *f, void *opaque, int version_id)
 485 {
 486     VirtIOBlock *s = opaque;
 487
 488     if (version_id != 2)
 489         return -EINVAL;
 490
 491     virtio_load(&s->vdev, f);
 492     while (qemu_get_sbyte(f)) {
 493         VirtIOBlockReq *req = virtio_blk_alloc_request(s);
 494         qemu_get_buffer(f, (unsigned char*)&req->elem, sizeof(req->elem));
 495         req->next = s->rq;
 496         s->rq = req;
 497
 498         virtqueue_map_sg(req->elem.in_sg, req->elem.in_addr,
 499             req->elem.in_num, 1);
 500         virtqueue_map_sg(req->elem.out_sg, req->elem.out_addr,
 501             req->elem.out_num, 0);
 502     }
 503
 504     return 0;
 505 }
 506
 507 VirtIODevice *virtio_blk_init(DeviceState *dev, BlockConf *conf)
 508 {
 509     VirtIOBlock *s;
 510     int cylinders, heads, secs;
 511     static int virtio_blk_id;
 512     DriveInfo *dinfo;
 513
 514     if (!conf->bs) {
 515         error_report("virtio-blk-pci: drive property not set");
 516         return NULL;
 517     }
 518     if (!bdrv_is_inserted(conf->bs)) {
 519         error_report("Device needs media, but drive is empty");
 520         return NULL;
 521     }
 522
 523     s = (VirtIOBlock *)virtio_common_init("virtio-blk", VIRTIO_ID_BLOCK,
 524                                           sizeof(struct virtio_blk_config),
 525                                           sizeof(VirtIOBlock));
 526
 527     s->vdev.get_config = virtio_blk_update_config;
 528     s->vdev.get_features = virtio_blk_get_features;
 529     s->vdev.reset = virtio_blk_reset;
 530     s->bs = conf->bs;
 531     s->conf = conf;
 532     s->rq = NULL;
 533     s->sector_mask = (s->conf->logical_block_size / BDRV_SECTOR_SIZE) - 1;
 534     bdrv_guess_geometry(s->bs, &cylinders, &heads, &secs);
 535
 536     /* NB: per existing s/n string convention the string is terminated
 537      * by '\0' only when less than sizeof (s->sn)
 538      */
 539     dinfo = drive_get_by_blockdev(s->bs);
 540     strncpy(s->sn, dinfo->serial, sizeof (s->sn));
 541
 542     s->vq = virtio_add_queue(&s->vdev, 128, virtio_blk_handle_output);
 543
 544     qemu_add_vm_change_state_handler(virtio_blk_dma_restart_cb, s);
 545     s->qdev = dev;
 546     register_savevm(dev, "virtio-blk", virtio_blk_id++, 2,
 547                     virtio_blk_save, virtio_blk_load, s);
 548     bdrv_set_removable(s->bs, 0);
 549     s->bs->buffer_alignment = conf->logical_block_size;
 550
 551     return &s->vdev;
 552 }
 553
 554 void virtio_blk_exit(VirtIODevice *vdev)
 555 {
 556     VirtIOBlock *s = to_virtio_blk(vdev);
 557     unregister_savevm(s->qdev, "virtio-blk", s);
 558 }