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