hw/block/xen_disk.c

   1 /*
   2  *  xen paravirt block device backend
   3  *
   4  *  (c) Gerd Hoffmann <kraxel@redhat.com>
   5  *
   6  *  This program is free software; you can redistribute it and/or modify
   7  *  it under the terms of the GNU General Public License as published by
   8  *  the Free Software Foundation; under version 2 of the License.
   9  *
  10  *  This program is distributed in the hope that it will be useful,
  11  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  *  GNU General Public License for more details.
  14  *
  15  *  You should have received a copy of the GNU General Public License along
  16  *  with this program; if not, see <http://www.gnu.org/licenses/>.
  17  *
  18  *  Contributions after 2012-01-13 are licensed under the terms of the
  19  *  GNU GPL, version 2 or (at your option) any later version.
  20  */
  21
  22 #include <stdio.h>
  23 #include <stdlib.h>
  24 #include <stdarg.h>
  25 #include <string.h>
  26 #include <unistd.h>
  27 #include <signal.h>
  28 #include <inttypes.h>
  29 #include <time.h>
  30 #include <fcntl.h>
  31 #include <errno.h>
  32 #include <sys/ioctl.h>
  33 #include <sys/types.h>
  34 #include <sys/stat.h>
  35 #include <sys/mman.h>
  36 #include <sys/uio.h>
  37
  38 #include "hw/hw.h"
  39 #include "hw/xen/xen_backend.h"
  40 #include "xen_blkif.h"
  41 #include "sysemu/blockdev.h"
  42
  43 /* ------------------------------------------------------------- */
  44
  45 static int batch_maps   = 0;
  46
  47 static int max_requests = 32;
  48
  49 /* ------------------------------------------------------------- */
  50
  51 #define BLOCK_SIZE  512
  52 #define IOCB_COUNT  (BLKIF_MAX_SEGMENTS_PER_REQUEST + 2)
  53
  54 struct PersistentGrant {
  55     void *page;
  56     struct XenBlkDev *blkdev;
  57 };
  58
  59 typedef struct PersistentGrant PersistentGrant;
  60
  61 struct ioreq {
  62     blkif_request_t     req;
  63     int16_t             status;
  64
  65     /* parsed request */
  66     off_t               start;
  67     QEMUIOVector        v;
  68     int                 presync;
  69     int                 postsync;
  70     uint8_t             mapped;
  71
  72     /* grant mapping */
  73     uint32_t            domids[BLKIF_MAX_SEGMENTS_PER_REQUEST];
  74     uint32_t            refs[BLKIF_MAX_SEGMENTS_PER_REQUEST];
  75     int                 prot;
  76     void                *page[BLKIF_MAX_SEGMENTS_PER_REQUEST];
  77     void                *pages;
  78     int                 num_unmap;
  79
  80     /* aio status */
  81     int                 aio_inflight;
  82     int                 aio_errors;
  83
  84     struct XenBlkDev    *blkdev;
  85     QLIST_ENTRY(ioreq)   list;
  86     BlockAcctCookie     acct;
  87 };
  88
  89 struct XenBlkDev {
  90     struct XenDevice    xendev;  /* must be first */
  91     char                *params;
  92     char                *mode;
  93     char                *type;
  94     char                *dev;
  95     char                *devtype;
  96     bool                directiosafe;
  97     const char          *fileproto;
  98     const char          *filename;
  99     int                 ring_ref;
 100     void                *sring;
 101     int64_t             file_blk;
 102     int64_t             file_size;
 103     int                 protocol;
 104     blkif_back_rings_t  rings;
 105     int                 more_work;
 106     int                 cnt_map;
 107
 108     /* request lists */
 109     QLIST_HEAD(inflight_head, ioreq) inflight;
 110     QLIST_HEAD(finished_head, ioreq) finished;
 111     QLIST_HEAD(freelist_head, ioreq) freelist;
 112     int                 requests_total;
 113     int                 requests_inflight;
 114     int                 requests_finished;
 115
 116     /* Persistent grants extension */
 117     gboolean            feature_persistent;
 118     GTree               *persistent_gnts;
 119     unsigned int        persistent_gnt_count;
 120     unsigned int        max_grants;
 121
 122     /* qemu block driver */
 123     DriveInfo           *dinfo;
 124     BlockDriverState    *bs;
 125     QEMUBH              *bh;
 126 };
 127
 128 /* ------------------------------------------------------------- */
 129
 130 static void ioreq_reset(struct ioreq *ioreq)
 131 {
 132     memset(&ioreq->req, 0, sizeof(ioreq->req));
 133     ioreq->status = 0;
 134     ioreq->start = 0;
 135     ioreq->presync = 0;
 136     ioreq->postsync = 0;
 137     ioreq->mapped = 0;
 138
 139     memset(ioreq->domids, 0, sizeof(ioreq->domids));
 140     memset(ioreq->refs, 0, sizeof(ioreq->refs));
 141     ioreq->prot = 0;
 142     memset(ioreq->page, 0, sizeof(ioreq->page));
 143     ioreq->pages = NULL;
 144
 145     ioreq->aio_inflight = 0;
 146     ioreq->aio_errors = 0;
 147
 148     ioreq->blkdev = NULL;
 149     memset(&ioreq->list, 0, sizeof(ioreq->list));
 150     memset(&ioreq->acct, 0, sizeof(ioreq->acct));
 151
 152     qemu_iovec_reset(&ioreq->v);
 153 }
 154
 155 static gint int_cmp(gconstpointer a, gconstpointer b, gpointer user_data)
 156 {
 157     uint ua = GPOINTER_TO_UINT(a);
 158     uint ub = GPOINTER_TO_UINT(b);
 159     return (ua > ub) - (ua < ub);
 160 }
 161
 162 static void destroy_grant(gpointer pgnt)
 163 {
 164     PersistentGrant *grant = pgnt;
 165     XenGnttab gnt = grant->blkdev->xendev.gnttabdev;
 166
 167     if (xc_gnttab_munmap(gnt, grant->page, 1) != 0) {
 168         xen_be_printf(&grant->blkdev->xendev, 0,
 169                       "xc_gnttab_munmap failed: %s\n",
 170                       strerror(errno));
 171     }
 172     grant->blkdev->persistent_gnt_count--;
 173     xen_be_printf(&grant->blkdev->xendev, 3,
 174                   "unmapped grant %p\n", grant->page);
 175     g_free(grant);
 176 }
 177
 178 static struct ioreq *ioreq_start(struct XenBlkDev *blkdev)
 179 {
 180     struct ioreq *ioreq = NULL;
 181
 182     if (QLIST_EMPTY(&blkdev->freelist)) {
 183         if (blkdev->requests_total >= max_requests) {
 184             goto out;
 185         }
 186         /* allocate new struct */
 187         ioreq = g_malloc0(sizeof(*ioreq));
 188         ioreq->blkdev = blkdev;
 189         blkdev->requests_total++;
 190         qemu_iovec_init(&ioreq->v, BLKIF_MAX_SEGMENTS_PER_REQUEST);
 191     } else {
 192         /* get one from freelist */
 193         ioreq = QLIST_FIRST(&blkdev->freelist);
 194         QLIST_REMOVE(ioreq, list);
 195     }
 196     QLIST_INSERT_HEAD(&blkdev->inflight, ioreq, list);
 197     blkdev->requests_inflight++;
 198
 199 out:
 200     return ioreq;
 201 }
 202
 203 static void ioreq_finish(struct ioreq *ioreq)
 204 {
 205     struct XenBlkDev *blkdev = ioreq->blkdev;
 206
 207     QLIST_REMOVE(ioreq, list);
 208     QLIST_INSERT_HEAD(&blkdev->finished, ioreq, list);
 209     blkdev->requests_inflight--;
 210     blkdev->requests_finished++;
 211 }
 212
 213 static void ioreq_release(struct ioreq *ioreq, bool finish)
 214 {
 215     struct XenBlkDev *blkdev = ioreq->blkdev;
 216
 217     QLIST_REMOVE(ioreq, list);
 218     ioreq_reset(ioreq);
 219     ioreq->blkdev = blkdev;
 220     QLIST_INSERT_HEAD(&blkdev->freelist, ioreq, list);
 221     if (finish) {
 222         blkdev->requests_finished--;
 223     } else {
 224         blkdev->requests_inflight--;
 225     }
 226 }
 227
 228 /*
 229  * translate request into iovec + start offset
 230  * do sanity checks along the way
 231  */
 232 static int ioreq_parse(struct ioreq *ioreq)
 233 {
 234     struct XenBlkDev *blkdev = ioreq->blkdev;
 235     uintptr_t mem;
 236     size_t len;
 237     int i;
 238
 239     xen_be_printf(&blkdev->xendev, 3,
 240                   "op %d, nr %d, handle %d, id %" PRId64 ", sector %" PRId64 "\n",
 241                   ioreq->req.operation, ioreq->req.nr_segments,
 242                   ioreq->req.handle, ioreq->req.id, ioreq->req.sector_number);
 243     switch (ioreq->req.operation) {
 244     case BLKIF_OP_READ:
 245         ioreq->prot = PROT_WRITE; /* to memory */
 246         break;
 247     case BLKIF_OP_FLUSH_DISKCACHE:
 248         ioreq->presync = 1;
 249         if (!ioreq->req.nr_segments) {
 250             return 0;
 251         }
 252         /* fall through */
 253     case BLKIF_OP_WRITE:
 254         ioreq->prot = PROT_READ; /* from memory */
 255         break;
 256     default:
 257         xen_be_printf(&blkdev->xendev, 0, "error: unknown operation (%d)\n",
 258                       ioreq->req.operation);
 259         goto err;
 260     };
 261
 262     if (ioreq->req.operation != BLKIF_OP_READ && blkdev->mode[0] != 'w') {
 263         xen_be_printf(&blkdev->xendev, 0, "error: write req for ro device\n");
 264         goto err;
 265     }
 266
 267     ioreq->start = ioreq->req.sector_number * blkdev->file_blk;
 268     for (i = 0; i < ioreq->req.nr_segments; i++) {
 269         if (i == BLKIF_MAX_SEGMENTS_PER_REQUEST) {
 270             xen_be_printf(&blkdev->xendev, 0, "error: nr_segments too big\n");
 271             goto err;
 272         }
 273         if (ioreq->req.seg[i].first_sect > ioreq->req.seg[i].last_sect) {
 274             xen_be_printf(&blkdev->xendev, 0, "error: first > last sector\n");
 275             goto err;
 276         }
 277         if (ioreq->req.seg[i].last_sect * BLOCK_SIZE >= XC_PAGE_SIZE) {
 278             xen_be_printf(&blkdev->xendev, 0, "error: page crossing\n");
 279             goto err;
 280         }
 281
 282         ioreq->domids[i] = blkdev->xendev.dom;
 283         ioreq->refs[i]   = ioreq->req.seg[i].gref;
 284
 285         mem = ioreq->req.seg[i].first_sect * blkdev->file_blk;
 286         len = (ioreq->req.seg[i].last_sect - ioreq->req.seg[i].first_sect + 1) * blkdev->file_blk;
 287         qemu_iovec_add(&ioreq->v, (void*)mem, len);
 288     }
 289     if (ioreq->start + ioreq->v.size > blkdev->file_size) {
 290         xen_be_printf(&blkdev->xendev, 0, "error: access beyond end of file\n");
 291         goto err;
 292     }
 293     return 0;
 294
 295 err:
 296     ioreq->status = BLKIF_RSP_ERROR;
 297     return -1;
 298 }
 299
 300 static void ioreq_unmap(struct ioreq *ioreq)
 301 {
 302     XenGnttab gnt = ioreq->blkdev->xendev.gnttabdev;
 303     int i;
 304
 305     if (ioreq->num_unmap == 0 || ioreq->mapped == 0) {
 306         return;
 307     }
 308     if (batch_maps) {
 309         if (!ioreq->pages) {
 310             return;
 311         }
 312         if (xc_gnttab_munmap(gnt, ioreq->pages, ioreq->num_unmap) != 0) {
 313             xen_be_printf(&ioreq->blkdev->xendev, 0, "xc_gnttab_munmap failed: %s\n",
 314                           strerror(errno));
 315         }
 316         ioreq->blkdev->cnt_map -= ioreq->num_unmap;
 317         ioreq->pages = NULL;
 318     } else {
 319         for (i = 0; i < ioreq->num_unmap; i++) {
 320             if (!ioreq->page[i]) {
 321                 continue;
 322             }
 323             if (xc_gnttab_munmap(gnt, ioreq->page[i], 1) != 0) {
 324                 xen_be_printf(&ioreq->blkdev->xendev, 0, "xc_gnttab_munmap failed: %s\n",
 325                               strerror(errno));
 326             }
 327             ioreq->blkdev->cnt_map--;
 328             ioreq->page[i] = NULL;
 329         }
 330     }
 331     ioreq->mapped = 0;
 332 }
 333
 334 static int ioreq_map(struct ioreq *ioreq)
 335 {
 336     XenGnttab gnt = ioreq->blkdev->xendev.gnttabdev;
 337     uint32_t domids[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 338     uint32_t refs[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 339     void *page[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 340     int i, j, new_maps = 0;
 341     PersistentGrant *grant;
 342     /* domids and refs variables will contain the information necessary
 343      * to map the grants that are needed to fulfill this request.
 344      *
 345      * After mapping the needed grants, the page array will contain the
 346      * memory address of each granted page in the order specified in ioreq
 347      * (disregarding if it's a persistent grant or not).
 348      */
 349
 350     if (ioreq->v.niov == 0 || ioreq->mapped == 1) {
 351         return 0;
 352     }
 353     if (ioreq->blkdev->feature_persistent) {
 354         for (i = 0; i < ioreq->v.niov; i++) {
 355             grant = g_tree_lookup(ioreq->blkdev->persistent_gnts,
 356                                     GUINT_TO_POINTER(ioreq->refs[i]));
 357
 358             if (grant != NULL) {
 359                 page[i] = grant->page;
 360                 xen_be_printf(&ioreq->blkdev->xendev, 3,
 361                               "using persistent-grant %" PRIu32 "\n",
 362                               ioreq->refs[i]);
 363             } else {
 364                     /* Add the grant to the list of grants that
 365                      * should be mapped
 366                      */
 367                     domids[new_maps] = ioreq->domids[i];
 368                     refs[new_maps] = ioreq->refs[i];
 369                     page[i] = NULL;
 370                     new_maps++;
 371             }
 372         }
 373         /* Set the protection to RW, since grants may be reused later
 374          * with a different protection than the one needed for this request
 375          */
 376         ioreq->prot = PROT_WRITE | PROT_READ;
 377     } else {
 378         /* All grants in the request should be mapped */
 379         memcpy(refs, ioreq->refs, sizeof(refs));
 380         memcpy(domids, ioreq->domids, sizeof(domids));
 381         memset(page, 0, sizeof(page));
 382         new_maps = ioreq->v.niov;
 383     }
 384
 385     if (batch_maps && new_maps) {
 386         ioreq->pages = xc_gnttab_map_grant_refs
 387             (gnt, new_maps, domids, refs, ioreq->prot);
 388         if (ioreq->pages == NULL) {
 389             xen_be_printf(&ioreq->blkdev->xendev, 0,
 390                           "can't map %d grant refs (%s, %d maps)\n",
 391                           new_maps, strerror(errno), ioreq->blkdev->cnt_map);
 392             return -1;
 393         }
 394         for (i = 0, j = 0; i < ioreq->v.niov; i++) {
 395             if (page[i] == NULL) {
 396                 page[i] = ioreq->pages + (j++) * XC_PAGE_SIZE;
 397             }
 398         }
 399         ioreq->blkdev->cnt_map += new_maps;
 400     } else if (new_maps)  {
 401         for (i = 0; i < new_maps; i++) {
 402             ioreq->page[i] = xc_gnttab_map_grant_ref
 403                 (gnt, domids[i], refs[i], ioreq->prot);
 404             if (ioreq->page[i] == NULL) {
 405                 xen_be_printf(&ioreq->blkdev->xendev, 0,
 406                               "can't map grant ref %d (%s, %d maps)\n",
 407                               refs[i], strerror(errno), ioreq->blkdev->cnt_map);
 408                 ioreq_unmap(ioreq);
 409                 return -1;
 410             }
 411             ioreq->blkdev->cnt_map++;
 412         }
 413         for (i = 0, j = 0; i < ioreq->v.niov; i++) {
 414             if (page[i] == NULL) {
 415                 page[i] = ioreq->page[j++];
 416             }
 417         }
 418     }
 419     if (ioreq->blkdev->feature_persistent) {
 420         while ((ioreq->blkdev->persistent_gnt_count < ioreq->blkdev->max_grants)
 421               && new_maps) {
 422             /* Go through the list of newly mapped grants and add as many
 423              * as possible to the list of persistently mapped grants.
 424              *
 425              * Since we start at the end of ioreq->page(s), we only need
 426              * to decrease new_maps to prevent this granted pages from
 427              * being unmapped in ioreq_unmap.
 428              */
 429             grant = g_malloc0(sizeof(*grant));
 430             new_maps--;
 431             if (batch_maps) {
 432                 grant->page = ioreq->pages + (new_maps) * XC_PAGE_SIZE;
 433             } else {
 434                 grant->page = ioreq->page[new_maps];
 435             }
 436             grant->blkdev = ioreq->blkdev;
 437             xen_be_printf(&ioreq->blkdev->xendev, 3,
 438                           "adding grant %" PRIu32 " page: %p\n",
 439                           refs[new_maps], grant->page);
 440             g_tree_insert(ioreq->blkdev->persistent_gnts,
 441                           GUINT_TO_POINTER(refs[new_maps]),
 442                           grant);
 443             ioreq->blkdev->persistent_gnt_count++;
 444         }
 445     }
 446     for (i = 0; i < ioreq->v.niov; i++) {
 447         ioreq->v.iov[i].iov_base += (uintptr_t)page[i];
 448     }
 449     ioreq->mapped = 1;
 450     ioreq->num_unmap = new_maps;
 451     return 0;
 452 }
 453
 454 static int ioreq_runio_qemu_aio(struct ioreq *ioreq);
 455
 456 static void qemu_aio_complete(void *opaque, int ret)
 457 {
 458     struct ioreq *ioreq = opaque;
 459
 460     if (ret != 0) {
 461         xen_be_printf(&ioreq->blkdev->xendev, 0, "%s I/O error\n",
 462                       ioreq->req.operation == BLKIF_OP_READ ? "read" : "write");
 463         ioreq->aio_errors++;
 464     }
 465
 466     ioreq->aio_inflight--;
 467     if (ioreq->presync) {
 468         ioreq->presync = 0;
 469         ioreq_runio_qemu_aio(ioreq);
 470         return;
 471     }
 472     if (ioreq->aio_inflight > 0) {
 473         return;
 474     }
 475     if (ioreq->postsync) {
 476         ioreq->postsync = 0;
 477         ioreq->aio_inflight++;
 478         bdrv_aio_flush(ioreq->blkdev->bs, qemu_aio_complete, ioreq);
 479         return;
 480     }
 481
 482     ioreq->status = ioreq->aio_errors ? BLKIF_RSP_ERROR : BLKIF_RSP_OKAY;
 483     ioreq_unmap(ioreq);
 484     ioreq_finish(ioreq);
 485     bdrv_acct_done(ioreq->blkdev->bs, &ioreq->acct);
 486     qemu_bh_schedule(ioreq->blkdev->bh);
 487 }
 488
 489 static int ioreq_runio_qemu_aio(struct ioreq *ioreq)
 490 {
 491     struct XenBlkDev *blkdev = ioreq->blkdev;
 492
 493     if (ioreq->req.nr_segments && ioreq_map(ioreq) == -1) {
 494         goto err_no_map;
 495     }
 496
 497     ioreq->aio_inflight++;
 498     if (ioreq->presync) {
 499         bdrv_aio_flush(ioreq->blkdev->bs, qemu_aio_complete, ioreq);
 500         return 0;
 501     }
 502
 503     switch (ioreq->req.operation) {
 504     case BLKIF_OP_READ:
 505         bdrv_acct_start(blkdev->bs, &ioreq->acct, ioreq->v.size, BDRV_ACCT_READ);
 506         ioreq->aio_inflight++;
 507         bdrv_aio_readv(blkdev->bs, ioreq->start / BLOCK_SIZE,
 508                        &ioreq->v, ioreq->v.size / BLOCK_SIZE,
 509                        qemu_aio_complete, ioreq);
 510         break;
 511     case BLKIF_OP_WRITE:
 512     case BLKIF_OP_FLUSH_DISKCACHE:
 513         if (!ioreq->req.nr_segments) {
 514             break;
 515         }
 516
 517         bdrv_acct_start(blkdev->bs, &ioreq->acct, ioreq->v.size, BDRV_ACCT_WRITE);
 518         ioreq->aio_inflight++;
 519         bdrv_aio_writev(blkdev->bs, ioreq->start / BLOCK_SIZE,
 520                         &ioreq->v, ioreq->v.size / BLOCK_SIZE,
 521                         qemu_aio_complete, ioreq);
 522         break;
 523     default:
 524         /* unknown operation (shouldn't happen -- parse catches this) */
 525         goto err;
 526     }
 527
 528     qemu_aio_complete(ioreq, 0);
 529
 530     return 0;
 531
 532 err:
 533     ioreq_unmap(ioreq);
 534 err_no_map:
 535     ioreq_finish(ioreq);
 536     ioreq->status = BLKIF_RSP_ERROR;
 537     return -1;
 538 }
 539
 540 static int blk_send_response_one(struct ioreq *ioreq)
 541 {
 542     struct XenBlkDev  *blkdev = ioreq->blkdev;
 543     int               send_notify   = 0;
 544     int               have_requests = 0;
 545     blkif_response_t  resp;
 546     void              *dst;
 547
 548     resp.id        = ioreq->req.id;
 549     resp.operation = ioreq->req.operation;
 550     resp.status    = ioreq->status;
 551
 552     /* Place on the response ring for the relevant domain. */
 553     switch (blkdev->protocol) {
 554     case BLKIF_PROTOCOL_NATIVE:
 555         dst = RING_GET_RESPONSE(&blkdev->rings.native, blkdev->rings.native.rsp_prod_pvt);
 556         break;
 557     case BLKIF_PROTOCOL_X86_32:
 558         dst = RING_GET_RESPONSE(&blkdev->rings.x86_32_part,
 559                                 blkdev->rings.x86_32_part.rsp_prod_pvt);
 560         break;
 561     case BLKIF_PROTOCOL_X86_64:
 562         dst = RING_GET_RESPONSE(&blkdev->rings.x86_64_part,
 563                                 blkdev->rings.x86_64_part.rsp_prod_pvt);
 564         break;
 565     default:
 566         dst = NULL;
 567     }
 568     memcpy(dst, &resp, sizeof(resp));
 569     blkdev->rings.common.rsp_prod_pvt++;
 570
 571     RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blkdev->rings.common, send_notify);
 572     if (blkdev->rings.common.rsp_prod_pvt == blkdev->rings.common.req_cons) {
 573         /*
 574          * Tail check for pending requests. Allows frontend to avoid
 575          * notifications if requests are already in flight (lower
 576          * overheads and promotes batching).
 577          */
 578         RING_FINAL_CHECK_FOR_REQUESTS(&blkdev->rings.common, have_requests);
 579     } else if (RING_HAS_UNCONSUMED_REQUESTS(&blkdev->rings.common)) {
 580         have_requests = 1;
 581     }
 582
 583     if (have_requests) {
 584         blkdev->more_work++;
 585     }
 586     return send_notify;
 587 }
 588
 589 /* walk finished list, send outstanding responses, free requests */
 590 static void blk_send_response_all(struct XenBlkDev *blkdev)
 591 {
 592     struct ioreq *ioreq;
 593     int send_notify = 0;
 594
 595     while (!QLIST_EMPTY(&blkdev->finished)) {
 596         ioreq = QLIST_FIRST(&blkdev->finished);
 597         send_notify += blk_send_response_one(ioreq);
 598         ioreq_release(ioreq, true);
 599     }
 600     if (send_notify) {
 601         xen_be_send_notify(&blkdev->xendev);
 602     }
 603 }
 604
 605 static int blk_get_request(struct XenBlkDev *blkdev, struct ioreq *ioreq, RING_IDX rc)
 606 {
 607     switch (blkdev->protocol) {
 608     case BLKIF_PROTOCOL_NATIVE:
 609         memcpy(&ioreq->req, RING_GET_REQUEST(&blkdev->rings.native, rc),
 610                sizeof(ioreq->req));
 611         break;
 612     case BLKIF_PROTOCOL_X86_32:
 613         blkif_get_x86_32_req(&ioreq->req,
 614                              RING_GET_REQUEST(&blkdev->rings.x86_32_part, rc));
 615         break;
 616     case BLKIF_PROTOCOL_X86_64:
 617         blkif_get_x86_64_req(&ioreq->req,
 618                              RING_GET_REQUEST(&blkdev->rings.x86_64_part, rc));
 619         break;
 620     }
 621     return 0;
 622 }
 623
 624 static void blk_handle_requests(struct XenBlkDev *blkdev)
 625 {
 626     RING_IDX rc, rp;
 627     struct ioreq *ioreq;
 628
 629     blkdev->more_work = 0;
 630
 631     rc = blkdev->rings.common.req_cons;
 632     rp = blkdev->rings.common.sring->req_prod;
 633     xen_rmb(); /* Ensure we see queued requests up to 'rp'. */
 634
 635     blk_send_response_all(blkdev);
 636     while (rc != rp) {
 637         /* pull request from ring */
 638         if (RING_REQUEST_CONS_OVERFLOW(&blkdev->rings.common, rc)) {
 639             break;
 640         }
 641         ioreq = ioreq_start(blkdev);
 642         if (ioreq == NULL) {
 643             blkdev->more_work++;
 644             break;
 645         }
 646         blk_get_request(blkdev, ioreq, rc);
 647         blkdev->rings.common.req_cons = ++rc;
 648
 649         /* parse them */
 650         if (ioreq_parse(ioreq) != 0) {
 651             if (blk_send_response_one(ioreq)) {
 652                 xen_be_send_notify(&blkdev->xendev);
 653             }
 654             ioreq_release(ioreq, false);
 655             continue;
 656         }
 657
 658         ioreq_runio_qemu_aio(ioreq);
 659     }
 660
 661     if (blkdev->more_work && blkdev->requests_inflight < max_requests) {
 662         qemu_bh_schedule(blkdev->bh);
 663     }
 664 }
 665
 666 /* ------------------------------------------------------------- */
 667
 668 static void blk_bh(void *opaque)
 669 {
 670     struct XenBlkDev *blkdev = opaque;
 671     blk_handle_requests(blkdev);
 672 }
 673
 674 /*
 675  * We need to account for the grant allocations requiring contiguous
 676  * chunks; the worst case number would be
 677  *     max_req * max_seg + (max_req - 1) * (max_seg - 1) + 1,
 678  * but in order to keep things simple just use
 679  *     2 * max_req * max_seg.
 680  */
 681 #define MAX_GRANTS(max_req, max_seg) (2 * (max_req) * (max_seg))
 682
 683 static void blk_alloc(struct XenDevice *xendev)
 684 {
 685     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
 686
 687     QLIST_INIT(&blkdev->inflight);
 688     QLIST_INIT(&blkdev->finished);
 689     QLIST_INIT(&blkdev->freelist);
 690     blkdev->bh = qemu_bh_new(blk_bh, blkdev);
 691     if (xen_mode != XEN_EMULATE) {
 692         batch_maps = 1;
 693     }
 694     if (xc_gnttab_set_max_grants(xendev->gnttabdev,
 695             MAX_GRANTS(max_requests, BLKIF_MAX_SEGMENTS_PER_REQUEST)) < 0) {
 696         xen_be_printf(xendev, 0, "xc_gnttab_set_max_grants failed: %s\n",
 697                       strerror(errno));
 698     }
 699 }
 700
 701 static int blk_init(struct XenDevice *xendev)
 702 {
 703     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
 704     int info = 0;
 705     char *directiosafe = NULL;
 706
 707     /* read xenstore entries */
 708     if (blkdev->params == NULL) {
 709         char *h = NULL;
 710         blkdev->params = xenstore_read_be_str(&blkdev->xendev, "params");
 711         if (blkdev->params != NULL) {
 712             h = strchr(blkdev->params, ':');
 713         }
 714         if (h != NULL) {
 715             blkdev->fileproto = blkdev->params;
 716             blkdev->filename  = h+1;
 717             *h = 0;
 718         } else {
 719             blkdev->fileproto = "<unset>";
 720             blkdev->filename  = blkdev->params;
 721         }
 722     }
 723     if (!strcmp("aio", blkdev->fileproto)) {
 724         blkdev->fileproto = "raw";
 725     }
 726     if (blkdev->mode == NULL) {
 727         blkdev->mode = xenstore_read_be_str(&blkdev->xendev, "mode");
 728     }
 729     if (blkdev->type == NULL) {
 730         blkdev->type = xenstore_read_be_str(&blkdev->xendev, "type");
 731     }
 732     if (blkdev->dev == NULL) {
 733         blkdev->dev = xenstore_read_be_str(&blkdev->xendev, "dev");
 734     }
 735     if (blkdev->devtype == NULL) {
 736         blkdev->devtype = xenstore_read_be_str(&blkdev->xendev, "device-type");
 737     }
 738     directiosafe = xenstore_read_be_str(&blkdev->xendev, "direct-io-safe");
 739     blkdev->directiosafe = (directiosafe && atoi(directiosafe));
 740
 741     /* do we have all we need? */
 742     if (blkdev->params == NULL ||
 743         blkdev->mode == NULL   ||
 744         blkdev->type == NULL   ||
 745         blkdev->dev == NULL) {
 746         goto out_error;
 747     }
 748
 749     /* read-only ? */
 750     if (strcmp(blkdev->mode, "w")) {
 751         info  |= VDISK_READONLY;
 752     }
 753
 754     /* cdrom ? */
 755     if (blkdev->devtype && !strcmp(blkdev->devtype, "cdrom")) {
 756         info  |= VDISK_CDROM;
 757     }
 758
 759     blkdev->file_blk  = BLOCK_SIZE;
 760
 761     /* fill info
 762      * blk_connect supplies sector-size and sectors
 763      */
 764     xenstore_write_be_int(&blkdev->xendev, "feature-flush-cache", 1);
 765     xenstore_write_be_int(&blkdev->xendev, "feature-persistent", 1);
 766     xenstore_write_be_int(&blkdev->xendev, "info", info);
 767
 768     g_free(directiosafe);
 769     return 0;
 770
 771 out_error:
 772     g_free(blkdev->params);
 773     blkdev->params = NULL;
 774     g_free(blkdev->mode);
 775     blkdev->mode = NULL;
 776     g_free(blkdev->type);
 777     blkdev->type = NULL;
 778     g_free(blkdev->dev);
 779     blkdev->dev = NULL;
 780     g_free(blkdev->devtype);
 781     blkdev->devtype = NULL;
 782     g_free(directiosafe);
 783     blkdev->directiosafe = false;
 784     return -1;
 785 }
 786
 787 static int blk_connect(struct XenDevice *xendev)
 788 {
 789     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
 790     int pers, index, qflags;
 791     bool readonly = true;
 792
 793     /* read-only ? */
 794     if (blkdev->directiosafe) {
 795         qflags = BDRV_O_NOCACHE | BDRV_O_NATIVE_AIO;
 796     } else {
 797         qflags = BDRV_O_CACHE_WB;
 798     }
 799     if (strcmp(blkdev->mode, "w") == 0) {
 800         qflags |= BDRV_O_RDWR;
 801         readonly = false;
 802     }
 803
 804     /* init qemu block driver */
 805     index = (blkdev->xendev.dev - 202 * 256) / 16;
 806     blkdev->dinfo = drive_get(IF_XEN, 0, index);
 807     if (!blkdev->dinfo) {
 808         /* setup via xenbus -> create new block driver instance */
 809         xen_be_printf(&blkdev->xendev, 2, "create new bdrv (xenbus setup)\n");
 810         blkdev->bs = bdrv_new(blkdev->dev);
 811         if (blkdev->bs) {
 812             Error *local_err = NULL;
 813             BlockDriver *drv = bdrv_find_whitelisted_format(blkdev->fileproto,
 814                                                            readonly);
 815             if (bdrv_open(blkdev->bs,
 816                           blkdev->filename, NULL, qflags, drv, &local_err) != 0)
 817             {
 818                 xen_be_printf(&blkdev->xendev, 0, "error: %s\n",
 819                               error_get_pretty(local_err));
 820                 error_free(local_err);
 821                 bdrv_unref(blkdev->bs);
 822                 blkdev->bs = NULL;
 823             }
 824         }
 825         if (!blkdev->bs) {
 826             return -1;
 827         }
 828     } else {
 829         /* setup via qemu cmdline -> already setup for us */
 830         xen_be_printf(&blkdev->xendev, 2, "get configured bdrv (cmdline setup)\n");
 831         blkdev->bs = blkdev->dinfo->bdrv;
 832         /* blkdev->bs is not create by us, we get a reference
 833          * so we can bdrv_unref() unconditionally */
 834         bdrv_ref(blkdev->bs);
 835     }
 836     bdrv_attach_dev_nofail(blkdev->bs, blkdev);
 837     blkdev->file_size = bdrv_getlength(blkdev->bs);
 838     if (blkdev->file_size < 0) {
 839         xen_be_printf(&blkdev->xendev, 1, "bdrv_getlength: %d (%s) | drv %s\n",
 840                       (int)blkdev->file_size, strerror(-blkdev->file_size),
 841                       bdrv_get_format_name(blkdev->bs) ?: "-");
 842         blkdev->file_size = 0;
 843     }
 844
 845     xen_be_printf(xendev, 1, "type \"%s\", fileproto \"%s\", filename \"%s\","
 846                   " size %" PRId64 " (%" PRId64 " MB)\n",
 847                   blkdev->type, blkdev->fileproto, blkdev->filename,
 848                   blkdev->file_size, blkdev->file_size >> 20);
 849
 850     /* Fill in number of sector size and number of sectors */
 851     xenstore_write_be_int(&blkdev->xendev, "sector-size", blkdev->file_blk);
 852     xenstore_write_be_int64(&blkdev->xendev, "sectors",
 853                             blkdev->file_size / blkdev->file_blk);
 854
 855     if (xenstore_read_fe_int(&blkdev->xendev, "ring-ref", &blkdev->ring_ref) == -1) {
 856         return -1;
 857     }
 858     if (xenstore_read_fe_int(&blkdev->xendev, "event-channel",
 859                              &blkdev->xendev.remote_port) == -1) {
 860         return -1;
 861     }
 862     if (xenstore_read_fe_int(&blkdev->xendev, "feature-persistent", &pers)) {
 863         blkdev->feature_persistent = FALSE;
 864     } else {
 865         blkdev->feature_persistent = !!pers;
 866     }
 867
 868     blkdev->protocol = BLKIF_PROTOCOL_NATIVE;
 869     if (blkdev->xendev.protocol) {
 870         if (strcmp(blkdev->xendev.protocol, XEN_IO_PROTO_ABI_X86_32) == 0) {
 871             blkdev->protocol = BLKIF_PROTOCOL_X86_32;
 872         }
 873         if (strcmp(blkdev->xendev.protocol, XEN_IO_PROTO_ABI_X86_64) == 0) {
 874             blkdev->protocol = BLKIF_PROTOCOL_X86_64;
 875         }
 876     }
 877
 878     blkdev->sring = xc_gnttab_map_grant_ref(blkdev->xendev.gnttabdev,
 879                                             blkdev->xendev.dom,
 880                                             blkdev->ring_ref,
 881                                             PROT_READ | PROT_WRITE);
 882     if (!blkdev->sring) {
 883         return -1;
 884     }
 885     blkdev->cnt_map++;
 886
 887     switch (blkdev->protocol) {
 888     case BLKIF_PROTOCOL_NATIVE:
 889     {
 890         blkif_sring_t *sring_native = blkdev->sring;
 891         BACK_RING_INIT(&blkdev->rings.native, sring_native, XC_PAGE_SIZE);
 892         break;
 893     }
 894     case BLKIF_PROTOCOL_X86_32:
 895     {
 896         blkif_x86_32_sring_t *sring_x86_32 = blkdev->sring;
 897
 898         BACK_RING_INIT(&blkdev->rings.x86_32_part, sring_x86_32, XC_PAGE_SIZE);
 899         break;
 900     }
 901     case BLKIF_PROTOCOL_X86_64:
 902     {
 903         blkif_x86_64_sring_t *sring_x86_64 = blkdev->sring;
 904
 905         BACK_RING_INIT(&blkdev->rings.x86_64_part, sring_x86_64, XC_PAGE_SIZE);
 906         break;
 907     }
 908     }
 909
 910     if (blkdev->feature_persistent) {
 911         /* Init persistent grants */
 912         blkdev->max_grants = max_requests * BLKIF_MAX_SEGMENTS_PER_REQUEST;
 913         blkdev->persistent_gnts = g_tree_new_full((GCompareDataFunc)int_cmp,
 914                                              NULL, NULL,
 915                                              (GDestroyNotify)destroy_grant);
 916         blkdev->persistent_gnt_count = 0;
 917     }
 918
 919     xen_be_bind_evtchn(&blkdev->xendev);
 920
 921     xen_be_printf(&blkdev->xendev, 1, "ok: proto %s, ring-ref %d, "
 922                   "remote port %d, local port %d\n",
 923                   blkdev->xendev.protocol, blkdev->ring_ref,
 924                   blkdev->xendev.remote_port, blkdev->xendev.local_port);
 925     return 0;
 926 }
 927
 928 static void blk_disconnect(struct XenDevice *xendev)
 929 {
 930     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
 931
 932     if (blkdev->bs) {
 933         bdrv_detach_dev(blkdev->bs, blkdev);
 934         bdrv_unref(blkdev->bs);
 935         blkdev->bs = NULL;
 936     }
 937     xen_be_unbind_evtchn(&blkdev->xendev);
 938
 939     if (blkdev->sring) {
 940         xc_gnttab_munmap(blkdev->xendev.gnttabdev, blkdev->sring, 1);
 941         blkdev->cnt_map--;
 942         blkdev->sring = NULL;
 943     }
 944 }
 945
 946 static int blk_free(struct XenDevice *xendev)
 947 {
 948     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
 949     struct ioreq *ioreq;
 950
 951     if (blkdev->bs || blkdev->sring) {
 952         blk_disconnect(xendev);
 953     }
 954
 955     /* Free persistent grants */
 956     if (blkdev->feature_persistent) {
 957         g_tree_destroy(blkdev->persistent_gnts);
 958     }
 959
 960     while (!QLIST_EMPTY(&blkdev->freelist)) {
 961         ioreq = QLIST_FIRST(&blkdev->freelist);
 962         QLIST_REMOVE(ioreq, list);
 963         qemu_iovec_destroy(&ioreq->v);
 964         g_free(ioreq);
 965     }
 966
 967     g_free(blkdev->params);
 968     g_free(blkdev->mode);
 969     g_free(blkdev->type);
 970     g_free(blkdev->dev);
 971     g_free(blkdev->devtype);
 972     qemu_bh_delete(blkdev->bh);
 973     return 0;
 974 }
 975
 976 static void blk_event(struct XenDevice *xendev)
 977 {
 978     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
 979
 980     qemu_bh_schedule(blkdev->bh);
 981 }
 982
 983 struct XenDevOps xen_blkdev_ops = {
 984     .size       = sizeof(struct XenBlkDev),
 985     .flags      = DEVOPS_FLAG_NEED_GNTDEV,
 986     .alloc      = blk_alloc,
 987     .init       = blk_init,
 988     .initialise    = blk_connect,
 989     .disconnect = blk_disconnect,
 990     .event      = blk_event,
 991     .free       = blk_free,
 992 };