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 "virtio-blk.h"
  17 #ifdef __linux__
  18 # include <scsi/sg.h>
  19 #endif
  20
  21 typedef struct VirtIOBlock
  22 {
  23     VirtIODevice vdev;
  24     BlockDriverState *bs;
  25     VirtQueue *vq;
  26     void *rq;
  27     QEMUBH *bh;
  28     BlockConf *conf;
  29     unsigned short sector_mask;
  30     char sn[BLOCK_SERIAL_STRLEN];
  31 } VirtIOBlock;
  32
  33 static VirtIOBlock *to_virtio_blk(VirtIODevice *vdev)
  34 {
  35     return (VirtIOBlock *)vdev;
  36 }
  37
  38 typedef struct VirtIOBlockReq
  39 {
  40     VirtIOBlock *dev;
  41     VirtQueueElement elem;
  42     struct virtio_blk_inhdr *in;
  43     struct virtio_blk_outhdr *out;
  44     struct virtio_scsi_inhdr *scsi;
  45     QEMUIOVector qiov;
  46     struct VirtIOBlockReq *next;
  47 } VirtIOBlockReq;
  48
  49 static void virtio_blk_req_complete(VirtIOBlockReq *req, int status)
  50 {
  51     VirtIOBlock *s = req->dev;
  52
  53     req->in->status = status;
  54     virtqueue_push(s->vq, &req->elem, req->qiov.size + sizeof(*req->in));
  55     virtio_notify(&s->vdev, s->vq);
  56
  57     qemu_free(req);
  58 }
  59
  60 static int virtio_blk_handle_rw_error(VirtIOBlockReq *req, int error,
  61     int is_read)
  62 {
  63     BlockErrorAction action = bdrv_get_on_error(req->dev->bs, is_read);
  64     VirtIOBlock *s = req->dev;
  65
  66     if (action == BLOCK_ERR_IGNORE) {
  67         bdrv_mon_event(s->bs, BDRV_ACTION_IGNORE, is_read);
  68         return 0;
  69     }
  70
  71     if ((error == ENOSPC && action == BLOCK_ERR_STOP_ENOSPC)
  72             || action == BLOCK_ERR_STOP_ANY) {
  73         req->next = s->rq;
  74         s->rq = req;
  75         bdrv_mon_event(s->bs, BDRV_ACTION_STOP, is_read);
  76         vm_stop(0);
  77     } else {
  78         virtio_blk_req_complete(req, VIRTIO_BLK_S_IOERR);
  79         bdrv_mon_event(s->bs, BDRV_ACTION_REPORT, is_read);
  80     }
  81
  82     return 1;
  83 }
  84
  85 static void virtio_blk_rw_complete(void *opaque, int ret)
  86 {
  87     VirtIOBlockReq *req = opaque;
  88
  89     if (ret) {
  90         int is_read = !(req->out->type & VIRTIO_BLK_T_OUT);
  91         if (virtio_blk_handle_rw_error(req, -ret, is_read))
  92             return;
  93     }
  94
  95     virtio_blk_req_complete(req, VIRTIO_BLK_S_OK);
  96 }
  97
  98 static void virtio_blk_flush_complete(void *opaque, int ret)
  99 {
 100     VirtIOBlockReq *req = opaque;
 101
 102     virtio_blk_req_complete(req, ret ? VIRTIO_BLK_S_IOERR : VIRTIO_BLK_S_OK);
 103 }
 104
 105 static VirtIOBlockReq *virtio_blk_alloc_request(VirtIOBlock *s)
 106 {
 107     VirtIOBlockReq *req = qemu_malloc(sizeof(*req));
 108     req->dev = s;
 109     req->qiov.size = 0;
 110     req->next = NULL;
 111     return req;
 112 }
 113
 114 static VirtIOBlockReq *virtio_blk_get_request(VirtIOBlock *s)
 115 {
 116     VirtIOBlockReq *req = virtio_blk_alloc_request(s);
 117
 118     if (req != NULL) {
 119         if (!virtqueue_pop(s->vq, &req->elem)) {
 120             qemu_free(req);
 121             return NULL;
 122         }
 123     }
 124
 125     return req;
 126 }
 127
 128 #ifdef __linux__
 129 static void virtio_blk_handle_scsi(VirtIOBlockReq *req)
 130 {
 131     struct sg_io_hdr hdr;
 132     int ret;
 133     int status;
 134     int i;
 135
 136     /*
 137      * We require at least one output segment each for the virtio_blk_outhdr
 138      * and the SCSI command block.
 139      *
 140      * We also at least require the virtio_blk_inhdr, the virtio_scsi_inhdr
 141      * and the sense buffer pointer in the input segments.
 142      */
 143     if (req->elem.out_num < 2 || req->elem.in_num < 3) {
 144         virtio_blk_req_complete(req, VIRTIO_BLK_S_IOERR);
 145         return;
 146     }
 147
 148     /*
 149      * No support for bidirection commands yet.
 150      */
 151     if (req->elem.out_num > 2 && req->elem.in_num > 3) {
 152         virtio_blk_req_complete(req, VIRTIO_BLK_S_UNSUPP);
 153         return;
 154     }
 155
 156     /*
 157      * The scsi inhdr is placed in the second-to-last input segment, just
 158      * before the regular inhdr.
 159      */
 160     req->scsi = (void *)req->elem.in_sg[req->elem.in_num - 2].iov_base;
 161
 162     memset(&hdr, 0, sizeof(struct sg_io_hdr));
 163     hdr.interface_id = 'S';
 164     hdr.cmd_len = req->elem.out_sg[1].iov_len;
 165     hdr.cmdp = req->elem.out_sg[1].iov_base;
 166     hdr.dxfer_len = 0;
 167
 168     if (req->elem.out_num > 2) {
 169         /*
 170          * If there are more than the minimally required 2 output segments
 171          * there is write payload starting from the third iovec.
 172          */
 173         hdr.dxfer_direction = SG_DXFER_TO_DEV;
 174         hdr.iovec_count = req->elem.out_num - 2;
 175
 176         for (i = 0; i < hdr.iovec_count; i++)
 177             hdr.dxfer_len += req->elem.out_sg[i + 2].iov_len;
 178
 179         hdr.dxferp = req->elem.out_sg + 2;
 180
 181     } else if (req->elem.in_num > 3) {
 182         /*
 183          * If we have more than 3 input segments the guest wants to actually
 184          * read data.
 185          */
 186         hdr.dxfer_direction = SG_DXFER_FROM_DEV;
 187         hdr.iovec_count = req->elem.in_num - 3;
 188         for (i = 0; i < hdr.iovec_count; i++)
 189             hdr.dxfer_len += req->elem.in_sg[i].iov_len;
 190
 191         hdr.dxferp = req->elem.in_sg;
 192     } else {
 193         /*
 194          * Some SCSI commands don't actually transfer any data.
 195          */
 196         hdr.dxfer_direction = SG_DXFER_NONE;
 197     }
 198
 199     hdr.sbp = req->elem.in_sg[req->elem.in_num - 3].iov_base;
 200     hdr.mx_sb_len = req->elem.in_sg[req->elem.in_num - 3].iov_len;
 201
 202     ret = bdrv_ioctl(req->dev->bs, SG_IO, &hdr);
 203     if (ret) {
 204         status = VIRTIO_BLK_S_UNSUPP;
 205         hdr.status = ret;
 206         hdr.resid = hdr.dxfer_len;
 207     } else if (hdr.status) {
 208         status = VIRTIO_BLK_S_IOERR;
 209     } else {
 210         status = VIRTIO_BLK_S_OK;
 211     }
 212
 213     req->scsi->errors = hdr.status;
 214     req->scsi->residual = hdr.resid;
 215     req->scsi->sense_len = hdr.sb_len_wr;
 216     req->scsi->data_len = hdr.dxfer_len;
 217
 218     virtio_blk_req_complete(req, status);
 219 }
 220 #else
 221 static void virtio_blk_handle_scsi(VirtIOBlockReq *req)
 222 {
 223     virtio_blk_req_complete(req, VIRTIO_BLK_S_UNSUPP);
 224 }
 225 #endif /* __linux__ */
 226
 227 typedef struct MultiReqBuffer {
 228     BlockRequest        blkreq[32];
 229     unsigned int        num_writes;
 230 } MultiReqBuffer;
 231
 232 static void virtio_submit_multiwrite(BlockDriverState *bs, MultiReqBuffer *mrb)
 233 {
 234     int i, ret;
 235
 236     if (!mrb->num_writes) {
 237         return;
 238     }
 239
 240     ret = bdrv_aio_multiwrite(bs, mrb->blkreq, mrb->num_writes);
 241     if (ret != 0) {
 242         for (i = 0; i < mrb->num_writes; i++) {
 243             if (mrb->blkreq[i].error) {
 244                 virtio_blk_rw_complete(mrb->blkreq[i].opaque, -EIO);
 245             }
 246         }
 247     }
 248
 249     mrb->num_writes = 0;
 250 }
 251
 252 static void virtio_blk_handle_flush(VirtIOBlockReq *req, MultiReqBuffer *mrb)
 253 {
 254     BlockDriverAIOCB *acb;
 255
 256     /*
 257      * Make sure all outstanding writes are posted to the backing device.
 258      */
 259     virtio_submit_multiwrite(req->dev->bs, mrb);
 260
 261     acb = bdrv_aio_flush(req->dev->bs, virtio_blk_flush_complete, req);
 262     if (!acb) {
 263         virtio_blk_req_complete(req, VIRTIO_BLK_S_IOERR);
 264     }
 265 }
 266
 267 static void virtio_blk_handle_write(VirtIOBlockReq *req, MultiReqBuffer *mrb)
 268 {
 269     BlockRequest *blkreq;
 270
 271     if (req->out->sector & req->dev->sector_mask) {
 272         virtio_blk_rw_complete(req, -EIO);
 273         return;
 274     }
 275
 276     if (mrb->num_writes == 32) {
 277         virtio_submit_multiwrite(req->dev->bs, mrb);
 278     }
 279
 280     blkreq = &mrb->blkreq[mrb->num_writes];
 281     blkreq->sector = req->out->sector;
 282     blkreq->nb_sectors = req->qiov.size / BDRV_SECTOR_SIZE;
 283     blkreq->qiov = &req->qiov;
 284     blkreq->cb = virtio_blk_rw_complete;
 285     blkreq->opaque = req;
 286     blkreq->error = 0;
 287
 288     mrb->num_writes++;
 289 }
 290
 291 static void virtio_blk_handle_read(VirtIOBlockReq *req)
 292 {
 293     BlockDriverAIOCB *acb;
 294
 295     if (req->out->sector & req->dev->sector_mask) {
 296         virtio_blk_rw_complete(req, -EIO);
 297         return;
 298     }
 299
 300     acb = bdrv_aio_readv(req->dev->bs, req->out->sector, &req->qiov,
 301                          req->qiov.size / BDRV_SECTOR_SIZE,
 302                          virtio_blk_rw_complete, req);
 303     if (!acb) {
 304         virtio_blk_rw_complete(req, -EIO);
 305     }
 306 }
 307
 308 static void virtio_blk_handle_request(VirtIOBlockReq *req,
 309     MultiReqBuffer *mrb)
 310 {
 311     if (req->elem.out_num < 1 || req->elem.in_num < 1) {
 312         fprintf(stderr, "virtio-blk missing headers\n");
 313         exit(1);
 314     }
 315
 316     if (req->elem.out_sg[0].iov_len < sizeof(*req->out) ||
 317         req->elem.in_sg[req->elem.in_num - 1].iov_len < sizeof(*req->in)) {
 318         fprintf(stderr, "virtio-blk header not in correct element\n");
 319         exit(1);
 320     }
 321
 322     req->out = (void *)req->elem.out_sg[0].iov_base;
 323     req->in = (void *)req->elem.in_sg[req->elem.in_num - 1].iov_base;
 324
 325     if (req->out->type & VIRTIO_BLK_T_FLUSH) {
 326         virtio_blk_handle_flush(req, mrb);
 327     } else if (req->out->type & VIRTIO_BLK_T_SCSI_CMD) {
 328         virtio_blk_handle_scsi(req);
 329     } else if (req->out->type & VIRTIO_BLK_T_GET_ID) {
 330         VirtIOBlock *s = req->dev;
 331
 332         memcpy(req->elem.in_sg[0].iov_base, s->sn,
 333                MIN(req->elem.in_sg[0].iov_len, sizeof(s->sn)));
 334         virtio_blk_req_complete(req, VIRTIO_BLK_S_OK);
 335     } else if (req->out->type & VIRTIO_BLK_T_OUT) {
 336         qemu_iovec_init_external(&req->qiov, &req->elem.out_sg[1],
 337                                  req->elem.out_num - 1);
 338         virtio_blk_handle_write(req, mrb);
 339     } else {
 340         qemu_iovec_init_external(&req->qiov, &req->elem.in_sg[0],
 341                                  req->elem.in_num - 1);
 342         virtio_blk_handle_read(req);
 343     }
 344 }
 345
 346 static void virtio_blk_handle_output(VirtIODevice *vdev, VirtQueue *vq)
 347 {
 348     VirtIOBlock *s = to_virtio_blk(vdev);
 349     VirtIOBlockReq *req;
 350     MultiReqBuffer mrb = {
 351         .num_writes = 0,
 352     };
 353
 354     while ((req = virtio_blk_get_request(s))) {
 355         virtio_blk_handle_request(req, &mrb);
 356     }
 357
 358     virtio_submit_multiwrite(s->bs, &mrb);
 359
 360     /*
 361      * FIXME: Want to check for completions before returning to guest mode,
 362      * so cached reads and writes are reported as quickly as possible. But
 363      * that should be done in the generic block layer.
 364      */
 365 }
 366
 367 static void virtio_blk_dma_restart_bh(void *opaque)
 368 {
 369     VirtIOBlock *s = opaque;
 370     VirtIOBlockReq *req = s->rq;
 371     MultiReqBuffer mrb = {
 372         .num_writes = 0,
 373     };
 374
 375     qemu_bh_delete(s->bh);
 376     s->bh = NULL;
 377
 378     s->rq = NULL;
 379
 380     while (req) {
 381         virtio_blk_handle_request(req, &mrb);
 382         req = req->next;
 383     }
 384
 385     virtio_submit_multiwrite(s->bs, &mrb);
 386 }
 387
 388 static void virtio_blk_dma_restart_cb(void *opaque, int running, int reason)
 389 {
 390     VirtIOBlock *s = opaque;
 391
 392     if (!running)
 393         return;
 394
 395     if (!s->bh) {
 396         s->bh = qemu_bh_new(virtio_blk_dma_restart_bh, s);
 397         qemu_bh_schedule(s->bh);
 398     }
 399 }
 400
 401 static void virtio_blk_reset(VirtIODevice *vdev)
 402 {
 403     /*
 404      * This should cancel pending requests, but can't do nicely until there
 405      * are per-device request lists.
 406      */
 407     qemu_aio_flush();
 408 }
 409
 410 /* coalesce internal state, copy to pci i/o region 0
 411  */
 412 static void virtio_blk_update_config(VirtIODevice *vdev, uint8_t *config)
 413 {
 414     VirtIOBlock *s = to_virtio_blk(vdev);
 415     struct virtio_blk_config blkcfg;
 416     uint64_t capacity;
 417     int cylinders, heads, secs;
 418
 419     bdrv_get_geometry(s->bs, &capacity);
 420     bdrv_get_geometry_hint(s->bs, &cylinders, &heads, &secs);
 421     memset(&blkcfg, 0, sizeof(blkcfg));
 422     stq_raw(&blkcfg.capacity, capacity);
 423     stl_raw(&blkcfg.seg_max, 128 - 2);
 424     stw_raw(&blkcfg.cylinders, cylinders);
 425     blkcfg.heads = heads;
 426     blkcfg.sectors = secs & ~s->sector_mask;
 427     blkcfg.blk_size = s->conf->logical_block_size;
 428     blkcfg.size_max = 0;
 429     blkcfg.physical_block_exp = get_physical_block_exp(s->conf);
 430     blkcfg.alignment_offset = 0;
 431     blkcfg.min_io_size = s->conf->min_io_size / blkcfg.blk_size;
 432     blkcfg.opt_io_size = s->conf->opt_io_size / blkcfg.blk_size;
 433     memcpy(config, &blkcfg, sizeof(struct virtio_blk_config));
 434 }
 435
 436 static uint32_t virtio_blk_get_features(VirtIODevice *vdev, uint32_t features)
 437 {
 438     VirtIOBlock *s = to_virtio_blk(vdev);
 439
 440     features |= (1 << VIRTIO_BLK_F_SEG_MAX);
 441     features |= (1 << VIRTIO_BLK_F_GEOMETRY);
 442     features |= (1 << VIRTIO_BLK_F_TOPOLOGY);
 443     features |= (1 << VIRTIO_BLK_F_BLK_SIZE);
 444
 445     if (bdrv_enable_write_cache(s->bs))
 446         features |= (1 << VIRTIO_BLK_F_WCACHE);
 447
 448     if (bdrv_is_read_only(s->bs))
 449         features |= 1 << VIRTIO_BLK_F_RO;
 450
 451     return features;
 452 }
 453
 454 static void virtio_blk_save(QEMUFile *f, void *opaque)
 455 {
 456     VirtIOBlock *s = opaque;
 457     VirtIOBlockReq *req = s->rq;
 458
 459     virtio_save(&s->vdev, f);
 460
 461     while (req) {
 462         qemu_put_sbyte(f, 1);
 463         qemu_put_buffer(f, (unsigned char*)&req->elem, sizeof(req->elem));
 464         req = req->next;
 465     }
 466     qemu_put_sbyte(f, 0);
 467 }
 468
 469 static int virtio_blk_load(QEMUFile *f, void *opaque, int version_id)
 470 {
 471     VirtIOBlock *s = opaque;
 472
 473     if (version_id != 2)
 474         return -EINVAL;
 475
 476     virtio_load(&s->vdev, f);
 477     while (qemu_get_sbyte(f)) {
 478         VirtIOBlockReq *req = virtio_blk_alloc_request(s);
 479         qemu_get_buffer(f, (unsigned char*)&req->elem, sizeof(req->elem));
 480         req->next = s->rq;
 481         s->rq = req;
 482     }
 483
 484     return 0;
 485 }
 486
 487 VirtIODevice *virtio_blk_init(DeviceState *dev, BlockConf *conf)
 488 {
 489     VirtIOBlock *s;
 490     int cylinders, heads, secs;
 491     static int virtio_blk_id;
 492     DriveInfo *dinfo;
 493
 494     if (!conf->bs) {
 495         error_report("virtio-blk-pci: drive property not set");
 496         return NULL;
 497     }
 498     if (!bdrv_is_inserted(conf->bs)) {
 499         error_report("Device needs media, but drive is empty");
 500         return NULL;
 501     }
 502
 503     s = (VirtIOBlock *)virtio_common_init("virtio-blk", VIRTIO_ID_BLOCK,
 504                                           sizeof(struct virtio_blk_config),
 505                                           sizeof(VirtIOBlock));
 506
 507     s->vdev.get_config = virtio_blk_update_config;
 508     s->vdev.get_features = virtio_blk_get_features;
 509     s->vdev.reset = virtio_blk_reset;
 510     s->bs = conf->bs;
 511     s->conf = conf;
 512     s->rq = NULL;
 513     s->sector_mask = (s->conf->logical_block_size / BDRV_SECTOR_SIZE) - 1;
 514     bdrv_guess_geometry(s->bs, &cylinders, &heads, &secs);
 515
 516     /* NB: per existing s/n string convention the string is terminated
 517      * by '\0' only when less than sizeof (s->sn)
 518      */
 519     dinfo = drive_get_by_blockdev(s->bs);
 520     strncpy(s->sn, dinfo->serial, sizeof (s->sn));
 521
 522     s->vq = virtio_add_queue(&s->vdev, 128, virtio_blk_handle_output);
 523
 524     qemu_add_vm_change_state_handler(virtio_blk_dma_restart_cb, s);
 525     register_savevm(dev, "virtio-blk", virtio_blk_id++, 2,
 526                     virtio_blk_save, virtio_blk_load, s);
 527     bdrv_set_removable(s->bs, 0);
 528
 529     return &s->vdev;
 530 }