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     DeviceState *qdev;
  32 } VirtIOBlock;
  33
  34 static VirtIOBlock *to_virtio_blk(VirtIODevice *vdev)
  35 {
  36     return (VirtIOBlock *)vdev;
  37 }
  38
  39 typedef struct VirtIOBlockReq
  40 {
  41     VirtIOBlock *dev;
  42     VirtQueueElement elem;
  43     struct virtio_blk_inhdr *in;
  44     struct virtio_blk_outhdr *out;
  45     struct virtio_scsi_inhdr *scsi;
  46     QEMUIOVector qiov;
  47     struct VirtIOBlockReq *next;
  48 } VirtIOBlockReq;
  49
  50 static void virtio_blk_req_complete(VirtIOBlockReq *req, int status)
  51 {
  52     VirtIOBlock *s = req->dev;
  53
  54     req->in->status = status;
  55     virtqueue_push(s->vq, &req->elem, req->qiov.size + sizeof(*req->in));
  56     virtio_notify(&s->vdev, s->vq);
  57
  58     qemu_free(req);
  59 }
  60
  61 static int virtio_blk_handle_rw_error(VirtIOBlockReq *req, int error,
  62     int is_read)
  63 {
  64     BlockErrorAction action = bdrv_get_on_error(req->dev->bs, is_read);
  65     VirtIOBlock *s = req->dev;
  66
  67     if (action == BLOCK_ERR_IGNORE) {
  68         bdrv_mon_event(s->bs, BDRV_ACTION_IGNORE, is_read);
  69         return 0;
  70     }
  71
  72     if ((error == ENOSPC && action == BLOCK_ERR_STOP_ENOSPC)
  73             || action == BLOCK_ERR_STOP_ANY) {
  74         req->next = s->rq;
  75         s->rq = req;
  76         bdrv_mon_event(s->bs, BDRV_ACTION_STOP, is_read);
  77         vm_stop(0);
  78     } else {
  79         virtio_blk_req_complete(req, VIRTIO_BLK_S_IOERR);
  80         bdrv_mon_event(s->bs, BDRV_ACTION_REPORT, is_read);
  81     }
  82
  83     return 1;
  84 }
  85
  86 static void virtio_blk_rw_complete(void *opaque, int ret)
  87 {
  88     VirtIOBlockReq *req = opaque;
  89
  90     if (ret) {
  91         int is_read = !(req->out->type & VIRTIO_BLK_T_OUT);
  92         if (virtio_blk_handle_rw_error(req, -ret, is_read))
  93             return;
  94     }
  95
  96     virtio_blk_req_complete(req, VIRTIO_BLK_S_OK);
  97 }
  98
  99 static void virtio_blk_flush_complete(void *opaque, int ret)
 100 {
 101     VirtIOBlockReq *req = opaque;
 102
 103     virtio_blk_req_complete(req, ret ? VIRTIO_BLK_S_IOERR : VIRTIO_BLK_S_OK);
 104 }
 105
 106 static VirtIOBlockReq *virtio_blk_alloc_request(VirtIOBlock *s)
 107 {
 108     VirtIOBlockReq *req = qemu_malloc(sizeof(*req));
 109     req->dev = s;
 110     req->qiov.size = 0;
 111     req->next = NULL;
 112     return req;
 113 }
 114
 115 static VirtIOBlockReq *virtio_blk_get_request(VirtIOBlock *s)
 116 {
 117     VirtIOBlockReq *req = virtio_blk_alloc_request(s);
 118
 119     if (req != NULL) {
 120         if (!virtqueue_pop(s->vq, &req->elem)) {
 121             qemu_free(req);
 122             return NULL;
 123         }
 124     }
 125
 126     return req;
 127 }
 128
 129 #ifdef __linux__
 130 static void virtio_blk_handle_scsi(VirtIOBlockReq *req)
 131 {
 132     struct sg_io_hdr hdr;
 133     int ret;
 134     int status;
 135     int i;
 136
 137     /*
 138      * We require at least one output segment each for the virtio_blk_outhdr
 139      * and the SCSI command block.
 140      *
 141      * We also at least require the virtio_blk_inhdr, the virtio_scsi_inhdr
 142      * and the sense buffer pointer in the input segments.
 143      */
 144     if (req->elem.out_num < 2 || req->elem.in_num < 3) {
 145         virtio_blk_req_complete(req, VIRTIO_BLK_S_IOERR);
 146         return;
 147     }
 148
 149     /*
 150      * No support for bidirection commands yet.
 151      */
 152     if (req->elem.out_num > 2 && req->elem.in_num > 3) {
 153         virtio_blk_req_complete(req, VIRTIO_BLK_S_UNSUPP);
 154         return;
 155     }
 156
 157     /*
 158      * The scsi inhdr is placed in the second-to-last input segment, just
 159      * before the regular inhdr.
 160      */
 161     req->scsi = (void *)req->elem.in_sg[req->elem.in_num - 2].iov_base;
 162
 163     memset(&hdr, 0, sizeof(struct sg_io_hdr));
 164     hdr.interface_id = 'S';
 165     hdr.cmd_len = req->elem.out_sg[1].iov_len;
 166     hdr.cmdp = req->elem.out_sg[1].iov_base;
 167     hdr.dxfer_len = 0;
 168
 169     if (req->elem.out_num > 2) {
 170         /*
 171          * If there are more than the minimally required 2 output segments
 172          * there is write payload starting from the third iovec.
 173          */
 174         hdr.dxfer_direction = SG_DXFER_TO_DEV;
 175         hdr.iovec_count = req->elem.out_num - 2;
 176
 177         for (i = 0; i < hdr.iovec_count; i++)
 178             hdr.dxfer_len += req->elem.out_sg[i + 2].iov_len;
 179
 180         hdr.dxferp = req->elem.out_sg + 2;
 181
 182     } else if (req->elem.in_num > 3) {
 183         /*
 184          * If we have more than 3 input segments the guest wants to actually
 185          * read data.
 186          */
 187         hdr.dxfer_direction = SG_DXFER_FROM_DEV;
 188         hdr.iovec_count = req->elem.in_num - 3;
 189         for (i = 0; i < hdr.iovec_count; i++)
 190             hdr.dxfer_len += req->elem.in_sg[i].iov_len;
 191
 192         hdr.dxferp = req->elem.in_sg;
 193     } else {
 194         /*
 195          * Some SCSI commands don't actually transfer any data.
 196          */
 197         hdr.dxfer_direction = SG_DXFER_NONE;
 198     }
 199
 200     hdr.sbp = req->elem.in_sg[req->elem.in_num - 3].iov_base;
 201     hdr.mx_sb_len = req->elem.in_sg[req->elem.in_num - 3].iov_len;
 202
 203     ret = bdrv_ioctl(req->dev->bs, SG_IO, &hdr);
 204     if (ret) {
 205         status = VIRTIO_BLK_S_UNSUPP;
 206         hdr.status = ret;
 207         hdr.resid = hdr.dxfer_len;
 208     } else if (hdr.status) {
 209         status = VIRTIO_BLK_S_IOERR;
 210     } else {
 211         status = VIRTIO_BLK_S_OK;
 212     }
 213
 214     req->scsi->errors = hdr.status;
 215     req->scsi->residual = hdr.resid;
 216     req->scsi->sense_len = hdr.sb_len_wr;
 217     req->scsi->data_len = hdr.dxfer_len;
 218
 219     virtio_blk_req_complete(req, status);
 220 }
 221 #else
 222 static void virtio_blk_handle_scsi(VirtIOBlockReq *req)
 223 {
 224     virtio_blk_req_complete(req, VIRTIO_BLK_S_UNSUPP);
 225 }
 226 #endif /* __linux__ */
 227
 228 typedef struct MultiReqBuffer {
 229     BlockRequest        blkreq[32];
 230     unsigned int        num_writes;
 231 } MultiReqBuffer;
 232
 233 static void virtio_submit_multiwrite(BlockDriverState *bs, MultiReqBuffer *mrb)
 234 {
 235     int i, ret;
 236
 237     if (!mrb->num_writes) {
 238         return;
 239     }
 240
 241     ret = bdrv_aio_multiwrite(bs, mrb->blkreq, mrb->num_writes);
 242     if (ret != 0) {
 243         for (i = 0; i < mrb->num_writes; i++) {
 244             if (mrb->blkreq[i].error) {
 245                 virtio_blk_rw_complete(mrb->blkreq[i].opaque, -EIO);
 246             }
 247         }
 248     }
 249
 250     mrb->num_writes = 0;
 251 }
 252
 253 static void virtio_blk_handle_flush(VirtIOBlockReq *req, MultiReqBuffer *mrb)
 254 {
 255     BlockDriverAIOCB *acb;
 256
 257     /*
 258      * Make sure all outstanding writes are posted to the backing device.
 259      */
 260     virtio_submit_multiwrite(req->dev->bs, mrb);
 261
 262     acb = bdrv_aio_flush(req->dev->bs, virtio_blk_flush_complete, req);
 263     if (!acb) {
 264         virtio_blk_req_complete(req, VIRTIO_BLK_S_IOERR);
 265     }
 266 }
 267
 268 static void virtio_blk_handle_write(VirtIOBlockReq *req, MultiReqBuffer *mrb)
 269 {
 270     BlockRequest *blkreq;
 271
 272     if (req->out->sector & req->dev->sector_mask) {
 273         virtio_blk_rw_complete(req, -EIO);
 274         return;
 275     }
 276
 277     if (mrb->num_writes == 32) {
 278         virtio_submit_multiwrite(req->dev->bs, mrb);
 279     }
 280
 281     blkreq = &mrb->blkreq[mrb->num_writes];
 282     blkreq->sector = req->out->sector;
 283     blkreq->nb_sectors = req->qiov.size / BDRV_SECTOR_SIZE;
 284     blkreq->qiov = &req->qiov;
 285     blkreq->cb = virtio_blk_rw_complete;
 286     blkreq->opaque = req;
 287     blkreq->error = 0;
 288
 289     mrb->num_writes++;
 290 }
 291
 292 static void virtio_blk_handle_read(VirtIOBlockReq *req)
 293 {
 294     BlockDriverAIOCB *acb;
 295
 296     if (req->out->sector & req->dev->sector_mask) {
 297         virtio_blk_rw_complete(req, -EIO);
 298         return;
 299     }
 300
 301     acb = bdrv_aio_readv(req->dev->bs, req->out->sector, &req->qiov,
 302                          req->qiov.size / BDRV_SECTOR_SIZE,
 303                          virtio_blk_rw_complete, req);
 304     if (!acb) {
 305         virtio_blk_rw_complete(req, -EIO);
 306     }
 307 }
 308
 309 static void virtio_blk_handle_request(VirtIOBlockReq *req,
 310     MultiReqBuffer *mrb)
 311 {
 312     if (req->elem.out_num < 1 || req->elem.in_num < 1) {
 313         fprintf(stderr, "virtio-blk missing headers\n");
 314         exit(1);
 315     }
 316
 317     if (req->elem.out_sg[0].iov_len < sizeof(*req->out) ||
 318         req->elem.in_sg[req->elem.in_num - 1].iov_len < sizeof(*req->in)) {
 319         fprintf(stderr, "virtio-blk header not in correct element\n");
 320         exit(1);
 321     }
 322
 323     req->out = (void *)req->elem.out_sg[0].iov_base;
 324     req->in = (void *)req->elem.in_sg[req->elem.in_num - 1].iov_base;
 325
 326     if (req->out->type & VIRTIO_BLK_T_FLUSH) {
 327         virtio_blk_handle_flush(req, mrb);
 328     } else if (req->out->type & VIRTIO_BLK_T_SCSI_CMD) {
 329         virtio_blk_handle_scsi(req);
 330     } else if (req->out->type & VIRTIO_BLK_T_GET_ID) {
 331         VirtIOBlock *s = req->dev;
 332
 333         memcpy(req->elem.in_sg[0].iov_base, s->sn,
 334                MIN(req->elem.in_sg[0].iov_len, sizeof(s->sn)));
 335         virtio_blk_req_complete(req, VIRTIO_BLK_S_OK);
 336     } else if (req->out->type & VIRTIO_BLK_T_OUT) {
 337         qemu_iovec_init_external(&req->qiov, &req->elem.out_sg[1],
 338                                  req->elem.out_num - 1);
 339         virtio_blk_handle_write(req, mrb);
 340     } else {
 341         qemu_iovec_init_external(&req->qiov, &req->elem.in_sg[0],
 342                                  req->elem.in_num - 1);
 343         virtio_blk_handle_read(req);
 344     }
 345 }
 346
 347 static void virtio_blk_handle_output(VirtIODevice *vdev, VirtQueue *vq)
 348 {
 349     VirtIOBlock *s = to_virtio_blk(vdev);
 350     VirtIOBlockReq *req;
 351     MultiReqBuffer mrb = {
 352         .num_writes = 0,
 353     };
 354
 355     while ((req = virtio_blk_get_request(s))) {
 356         virtio_blk_handle_request(req, &mrb);
 357     }
 358
 359     virtio_submit_multiwrite(s->bs, &mrb);
 360
 361     /*
 362      * FIXME: Want to check for completions before returning to guest mode,
 363      * so cached reads and writes are reported as quickly as possible. But
 364      * that should be done in the generic block layer.
 365      */
 366 }
 367
 368 static void virtio_blk_dma_restart_bh(void *opaque)
 369 {
 370     VirtIOBlock *s = opaque;
 371     VirtIOBlockReq *req = s->rq;
 372     MultiReqBuffer mrb = {
 373         .num_writes = 0,
 374     };
 375
 376     qemu_bh_delete(s->bh);
 377     s->bh = NULL;
 378
 379     s->rq = NULL;
 380
 381     while (req) {
 382         virtio_blk_handle_request(req, &mrb);
 383         req = req->next;
 384     }
 385
 386     virtio_submit_multiwrite(s->bs, &mrb);
 387 }
 388
 389 static void virtio_blk_dma_restart_cb(void *opaque, int running, int reason)
 390 {
 391     VirtIOBlock *s = opaque;
 392
 393     if (!running)
 394         return;
 395
 396     if (!s->bh) {
 397         s->bh = qemu_bh_new(virtio_blk_dma_restart_bh, s);
 398         qemu_bh_schedule(s->bh);
 399     }
 400 }
 401
 402 static void virtio_blk_reset(VirtIODevice *vdev)
 403 {
 404     /*
 405      * This should cancel pending requests, but can't do nicely until there
 406      * are per-device request lists.
 407      */
 408     qemu_aio_flush();
 409 }
 410
 411 /* coalesce internal state, copy to pci i/o region 0
 412  */
 413 static void virtio_blk_update_config(VirtIODevice *vdev, uint8_t *config)
 414 {
 415     VirtIOBlock *s = to_virtio_blk(vdev);
 416     struct virtio_blk_config blkcfg;
 417     uint64_t capacity;
 418     int cylinders, heads, secs;
 419
 420     bdrv_get_geometry(s->bs, &capacity);
 421     bdrv_get_geometry_hint(s->bs, &cylinders, &heads, &secs);
 422     memset(&blkcfg, 0, sizeof(blkcfg));
 423     stq_raw(&blkcfg.capacity, capacity);
 424     stl_raw(&blkcfg.seg_max, 128 - 2);
 425     stw_raw(&blkcfg.cylinders, cylinders);
 426     blkcfg.heads = heads;
 427     blkcfg.sectors = secs & ~s->sector_mask;
 428     blkcfg.blk_size = s->conf->logical_block_size;
 429     blkcfg.size_max = 0;
 430     blkcfg.physical_block_exp = get_physical_block_exp(s->conf);
 431     blkcfg.alignment_offset = 0;
 432     blkcfg.min_io_size = s->conf->min_io_size / blkcfg.blk_size;
 433     blkcfg.opt_io_size = s->conf->opt_io_size / blkcfg.blk_size;
 434     memcpy(config, &blkcfg, sizeof(struct virtio_blk_config));
 435 }
 436
 437 static uint32_t virtio_blk_get_features(VirtIODevice *vdev, uint32_t features)
 438 {
 439     VirtIOBlock *s = to_virtio_blk(vdev);
 440
 441     features |= (1 << VIRTIO_BLK_F_SEG_MAX);
 442     features |= (1 << VIRTIO_BLK_F_GEOMETRY);
 443     features |= (1 << VIRTIO_BLK_F_TOPOLOGY);
 444     features |= (1 << VIRTIO_BLK_F_BLK_SIZE);
 445
 446     if (bdrv_enable_write_cache(s->bs))
 447         features |= (1 << VIRTIO_BLK_F_WCACHE);
 448
 449     if (bdrv_is_read_only(s->bs))
 450         features |= 1 << VIRTIO_BLK_F_RO;
 451
 452     return features;
 453 }
 454
 455 static void virtio_blk_save(QEMUFile *f, void *opaque)
 456 {
 457     VirtIOBlock *s = opaque;
 458     VirtIOBlockReq *req = s->rq;
 459
 460     virtio_save(&s->vdev, f);
 461
 462     while (req) {
 463         qemu_put_sbyte(f, 1);
 464         qemu_put_buffer(f, (unsigned char*)&req->elem, sizeof(req->elem));
 465         req = req->next;
 466     }
 467     qemu_put_sbyte(f, 0);
 468 }
 469
 470 static int virtio_blk_load(QEMUFile *f, void *opaque, int version_id)
 471 {
 472     VirtIOBlock *s = opaque;
 473
 474     if (version_id != 2)
 475         return -EINVAL;
 476
 477     virtio_load(&s->vdev, f);
 478     while (qemu_get_sbyte(f)) {
 479         VirtIOBlockReq *req = virtio_blk_alloc_request(s);
 480         qemu_get_buffer(f, (unsigned char*)&req->elem, sizeof(req->elem));
 481         req->next = s->rq;
 482         s->rq = req;
 483     }
 484
 485     return 0;
 486 }
 487
 488 VirtIODevice *virtio_blk_init(DeviceState *dev, BlockConf *conf)
 489 {
 490     VirtIOBlock *s;
 491     int cylinders, heads, secs;
 492     static int virtio_blk_id;
 493     DriveInfo *dinfo;
 494
 495     if (!conf->bs) {
 496         error_report("virtio-blk-pci: drive property not set");
 497         return NULL;
 498     }
 499     if (!bdrv_is_inserted(conf->bs)) {
 500         error_report("Device needs media, but drive is empty");
 501         return NULL;
 502     }
 503
 504     s = (VirtIOBlock *)virtio_common_init("virtio-blk", VIRTIO_ID_BLOCK,
 505                                           sizeof(struct virtio_blk_config),
 506                                           sizeof(VirtIOBlock));
 507
 508     s->vdev.get_config = virtio_blk_update_config;
 509     s->vdev.get_features = virtio_blk_get_features;
 510     s->vdev.reset = virtio_blk_reset;
 511     s->bs = conf->bs;
 512     s->conf = conf;
 513     s->rq = NULL;
 514     s->sector_mask = (s->conf->logical_block_size / BDRV_SECTOR_SIZE) - 1;
 515     bdrv_guess_geometry(s->bs, &cylinders, &heads, &secs);
 516
 517     /* NB: per existing s/n string convention the string is terminated
 518      * by '\0' only when less than sizeof (s->sn)
 519      */
 520     dinfo = drive_get_by_blockdev(s->bs);
 521     strncpy(s->sn, dinfo->serial, sizeof (s->sn));
 522
 523     s->vq = virtio_add_queue(&s->vdev, 128, virtio_blk_handle_output);
 524
 525     qemu_add_vm_change_state_handler(virtio_blk_dma_restart_cb, s);
 526     s->qdev = dev;
 527     register_savevm(dev, "virtio-blk", virtio_blk_id++, 2,
 528                     virtio_blk_save, virtio_blk_load, s);
 529     bdrv_set_removable(s->bs, 0);
 530
 531     return &s->vdev;
 532 }
 533
 534 void virtio_blk_exit(VirtIODevice *vdev)
 535 {
 536     VirtIOBlock *s = to_virtio_blk(vdev);
 537     unregister_savevm(s->qdev, "virtio-blk", s);
 538 }