hw/block/xen_disk.c

   1 /*
   2  *  xen paravirt block device backend
   3  *
   4  *  (c) Gerd Hoffmann <[email protected]>
   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 "qemu/osdep.h"
  23 #include <sys/ioctl.h>
  24 #include <sys/uio.h>
  25
  26 #include "hw/hw.h"
  27 #include "hw/xen/xen_backend.h"
  28 #include "xen_blkif.h"
  29 #include "sysemu/blockdev.h"
  30 #include "sysemu/iothread.h"
  31 #include "sysemu/block-backend.h"
  32 #include "qapi/error.h"
  33 #include "qapi/qmp/qdict.h"
  34 #include "qapi/qmp/qstring.h"
  35 #include "trace.h"
  36
  37 /* ------------------------------------------------------------- */
  38
  39 static int batch_maps   = 0;
  40
  41 /* ------------------------------------------------------------- */
  42
  43 #define BLOCK_SIZE  512
  44 #define IOCB_COUNT  (BLKIF_MAX_SEGMENTS_PER_REQUEST + 2)
  45
  46 struct PersistentGrant {
  47     void *page;
  48     struct XenBlkDev *blkdev;
  49 };
  50
  51 typedef struct PersistentGrant PersistentGrant;
  52
  53 struct PersistentRegion {
  54     void *addr;
  55     int num;
  56 };
  57
  58 typedef struct PersistentRegion PersistentRegion;
  59
  60 struct ioreq {
  61     blkif_request_t     req;
  62     int16_t             status;
  63
  64     /* parsed request */
  65     off_t               start;
  66     QEMUIOVector        v;
  67     int                 presync;
  68     uint8_t             mapped;
  69
  70     /* grant mapping */
  71     uint32_t            domids[BLKIF_MAX_SEGMENTS_PER_REQUEST];
  72     uint32_t            refs[BLKIF_MAX_SEGMENTS_PER_REQUEST];
  73     int                 prot;
  74     void                *page[BLKIF_MAX_SEGMENTS_PER_REQUEST];
  75     void                *pages;
  76     int                 num_unmap;
  77
  78     /* aio status */
  79     int                 aio_inflight;
  80     int                 aio_errors;
  81
  82     struct XenBlkDev    *blkdev;
  83     QLIST_ENTRY(ioreq)   list;
  84     BlockAcctCookie     acct;
  85 };
  86
  87 #define MAX_RING_PAGE_ORDER 4
  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     unsigned int        ring_ref[1 << MAX_RING_PAGE_ORDER];
 100     unsigned int        nr_ring_ref;
 101     void                *sring;
 102     int64_t             file_blk;
 103     int64_t             file_size;
 104     int                 protocol;
 105     blkif_back_rings_t  rings;
 106     int                 more_work;
 107     int                 cnt_map;
 108
 109     /* request lists */
 110     QLIST_HEAD(inflight_head, ioreq) inflight;
 111     QLIST_HEAD(finished_head, ioreq) finished;
 112     QLIST_HEAD(freelist_head, ioreq) freelist;
 113     int                 requests_total;
 114     int                 requests_inflight;
 115     int                 requests_finished;
 116     unsigned int        max_requests;
 117
 118     /* Persistent grants extension */
 119     gboolean            feature_discard;
 120     gboolean            feature_persistent;
 121     GTree               *persistent_gnts;
 122     GSList              *persistent_regions;
 123     unsigned int        persistent_gnt_count;
 124     unsigned int        max_grants;
 125
 126     /* qemu block driver */
 127     DriveInfo           *dinfo;
 128     BlockBackend        *blk;
 129     QEMUBH              *bh;
 130
 131     IOThread            *iothread;
 132     AioContext          *ctx;
 133 };
 134
 135 /* ------------------------------------------------------------- */
 136
 137 static void ioreq_reset(struct ioreq *ioreq)
 138 {
 139     memset(&ioreq->req, 0, sizeof(ioreq->req));
 140     ioreq->status = 0;
 141     ioreq->start = 0;
 142     ioreq->presync = 0;
 143     ioreq->mapped = 0;
 144
 145     memset(ioreq->domids, 0, sizeof(ioreq->domids));
 146     memset(ioreq->refs, 0, sizeof(ioreq->refs));
 147     ioreq->prot = 0;
 148     memset(ioreq->page, 0, sizeof(ioreq->page));
 149     ioreq->pages = NULL;
 150
 151     ioreq->aio_inflight = 0;
 152     ioreq->aio_errors = 0;
 153
 154     ioreq->blkdev = NULL;
 155     memset(&ioreq->list, 0, sizeof(ioreq->list));
 156     memset(&ioreq->acct, 0, sizeof(ioreq->acct));
 157
 158     qemu_iovec_reset(&ioreq->v);
 159 }
 160
 161 static gint int_cmp(gconstpointer a, gconstpointer b, gpointer user_data)
 162 {
 163     uint ua = GPOINTER_TO_UINT(a);
 164     uint ub = GPOINTER_TO_UINT(b);
 165     return (ua > ub) - (ua < ub);
 166 }
 167
 168 static void destroy_grant(gpointer pgnt)
 169 {
 170     PersistentGrant *grant = pgnt;
 171     xengnttab_handle *gnt = grant->blkdev->xendev.gnttabdev;
 172
 173     if (xengnttab_unmap(gnt, grant->page, 1) != 0) {
 174         xen_pv_printf(&grant->blkdev->xendev, 0,
 175                       "xengnttab_unmap failed: %s\n",
 176                       strerror(errno));
 177     }
 178     grant->blkdev->persistent_gnt_count--;
 179     xen_pv_printf(&grant->blkdev->xendev, 3,
 180                   "unmapped grant %p\n", grant->page);
 181     g_free(grant);
 182 }
 183
 184 static void remove_persistent_region(gpointer data, gpointer dev)
 185 {
 186     PersistentRegion *region = data;
 187     struct XenBlkDev *blkdev = dev;
 188     xengnttab_handle *gnt = blkdev->xendev.gnttabdev;
 189
 190     if (xengnttab_unmap(gnt, region->addr, region->num) != 0) {
 191         xen_pv_printf(&blkdev->xendev, 0,
 192                       "xengnttab_unmap region %p failed: %s\n",
 193                       region->addr, strerror(errno));
 194     }
 195     xen_pv_printf(&blkdev->xendev, 3,
 196                   "unmapped grant region %p with %d pages\n",
 197                   region->addr, region->num);
 198     g_free(region);
 199 }
 200
 201 static struct ioreq *ioreq_start(struct XenBlkDev *blkdev)
 202 {
 203     struct ioreq *ioreq = NULL;
 204
 205     if (QLIST_EMPTY(&blkdev->freelist)) {
 206         if (blkdev->requests_total >= blkdev->max_requests) {
 207             goto out;
 208         }
 209         /* allocate new struct */
 210         ioreq = g_malloc0(sizeof(*ioreq));
 211         ioreq->blkdev = blkdev;
 212         blkdev->requests_total++;
 213         qemu_iovec_init(&ioreq->v, BLKIF_MAX_SEGMENTS_PER_REQUEST);
 214     } else {
 215         /* get one from freelist */
 216         ioreq = QLIST_FIRST(&blkdev->freelist);
 217         QLIST_REMOVE(ioreq, list);
 218     }
 219     QLIST_INSERT_HEAD(&blkdev->inflight, ioreq, list);
 220     blkdev->requests_inflight++;
 221
 222 out:
 223     return ioreq;
 224 }
 225
 226 static void ioreq_finish(struct ioreq *ioreq)
 227 {
 228     struct XenBlkDev *blkdev = ioreq->blkdev;
 229
 230     QLIST_REMOVE(ioreq, list);
 231     QLIST_INSERT_HEAD(&blkdev->finished, ioreq, list);
 232     blkdev->requests_inflight--;
 233     blkdev->requests_finished++;
 234 }
 235
 236 static void ioreq_release(struct ioreq *ioreq, bool finish)
 237 {
 238     struct XenBlkDev *blkdev = ioreq->blkdev;
 239
 240     QLIST_REMOVE(ioreq, list);
 241     ioreq_reset(ioreq);
 242     ioreq->blkdev = blkdev;
 243     QLIST_INSERT_HEAD(&blkdev->freelist, ioreq, list);
 244     if (finish) {
 245         blkdev->requests_finished--;
 246     } else {
 247         blkdev->requests_inflight--;
 248     }
 249 }
 250
 251 /*
 252  * translate request into iovec + start offset
 253  * do sanity checks along the way
 254  */
 255 static int ioreq_parse(struct ioreq *ioreq)
 256 {
 257     struct XenBlkDev *blkdev = ioreq->blkdev;
 258     uintptr_t mem;
 259     size_t len;
 260     int i;
 261
 262     xen_pv_printf(&blkdev->xendev, 3,
 263                   "op %d, nr %d, handle %d, id %" PRId64 ", sector %" PRId64 "\n",
 264                   ioreq->req.operation, ioreq->req.nr_segments,
 265                   ioreq->req.handle, ioreq->req.id, ioreq->req.sector_number);
 266     switch (ioreq->req.operation) {
 267     case BLKIF_OP_READ:
 268         ioreq->prot = PROT_WRITE; /* to memory */
 269         break;
 270     case BLKIF_OP_FLUSH_DISKCACHE:
 271         ioreq->presync = 1;
 272         if (!ioreq->req.nr_segments) {
 273             return 0;
 274         }
 275         /* fall through */
 276     case BLKIF_OP_WRITE:
 277         ioreq->prot = PROT_READ; /* from memory */
 278         break;
 279     case BLKIF_OP_DISCARD:
 280         return 0;
 281     default:
 282         xen_pv_printf(&blkdev->xendev, 0, "error: unknown operation (%d)\n",
 283                       ioreq->req.operation);
 284         goto err;
 285     };
 286
 287     if (ioreq->req.operation != BLKIF_OP_READ && blkdev->mode[0] != 'w') {
 288         xen_pv_printf(&blkdev->xendev, 0, "error: write req for ro device\n");
 289         goto err;
 290     }
 291
 292     ioreq->start = ioreq->req.sector_number * blkdev->file_blk;
 293     for (i = 0; i < ioreq->req.nr_segments; i++) {
 294         if (i == BLKIF_MAX_SEGMENTS_PER_REQUEST) {
 295             xen_pv_printf(&blkdev->xendev, 0, "error: nr_segments too big\n");
 296             goto err;
 297         }
 298         if (ioreq->req.seg[i].first_sect > ioreq->req.seg[i].last_sect) {
 299             xen_pv_printf(&blkdev->xendev, 0, "error: first > last sector\n");
 300             goto err;
 301         }
 302         if (ioreq->req.seg[i].last_sect * BLOCK_SIZE >= XC_PAGE_SIZE) {
 303             xen_pv_printf(&blkdev->xendev, 0, "error: page crossing\n");
 304             goto err;
 305         }
 306
 307         ioreq->domids[i] = blkdev->xendev.dom;
 308         ioreq->refs[i]   = ioreq->req.seg[i].gref;
 309
 310         mem = ioreq->req.seg[i].first_sect * blkdev->file_blk;
 311         len = (ioreq->req.seg[i].last_sect - ioreq->req.seg[i].first_sect + 1) * blkdev->file_blk;
 312         qemu_iovec_add(&ioreq->v, (void*)mem, len);
 313     }
 314     if (ioreq->start + ioreq->v.size > blkdev->file_size) {
 315         xen_pv_printf(&blkdev->xendev, 0, "error: access beyond end of file\n");
 316         goto err;
 317     }
 318     return 0;
 319
 320 err:
 321     ioreq->status = BLKIF_RSP_ERROR;
 322     return -1;
 323 }
 324
 325 static void ioreq_unmap(struct ioreq *ioreq)
 326 {
 327     xengnttab_handle *gnt = ioreq->blkdev->xendev.gnttabdev;
 328     int i;
 329
 330     if (ioreq->num_unmap == 0 || ioreq->mapped == 0) {
 331         return;
 332     }
 333     if (batch_maps) {
 334         if (!ioreq->pages) {
 335             return;
 336         }
 337         if (xengnttab_unmap(gnt, ioreq->pages, ioreq->num_unmap) != 0) {
 338             xen_pv_printf(&ioreq->blkdev->xendev, 0,
 339                           "xengnttab_unmap failed: %s\n",
 340                           strerror(errno));
 341         }
 342         ioreq->blkdev->cnt_map -= ioreq->num_unmap;
 343         ioreq->pages = NULL;
 344     } else {
 345         for (i = 0; i < ioreq->num_unmap; i++) {
 346             if (!ioreq->page[i]) {
 347                 continue;
 348             }
 349             if (xengnttab_unmap(gnt, ioreq->page[i], 1) != 0) {
 350                 xen_pv_printf(&ioreq->blkdev->xendev, 0,
 351                               "xengnttab_unmap failed: %s\n",
 352                               strerror(errno));
 353             }
 354             ioreq->blkdev->cnt_map--;
 355             ioreq->page[i] = NULL;
 356         }
 357     }
 358     ioreq->mapped = 0;
 359 }
 360
 361 static int ioreq_map(struct ioreq *ioreq)
 362 {
 363     xengnttab_handle *gnt = ioreq->blkdev->xendev.gnttabdev;
 364     uint32_t domids[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 365     uint32_t refs[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 366     void *page[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 367     int i, j, new_maps = 0;
 368     PersistentGrant *grant;
 369     PersistentRegion *region;
 370     /* domids and refs variables will contain the information necessary
 371      * to map the grants that are needed to fulfill this request.
 372      *
 373      * After mapping the needed grants, the page array will contain the
 374      * memory address of each granted page in the order specified in ioreq
 375      * (disregarding if it's a persistent grant or not).
 376      */
 377
 378     if (ioreq->v.niov == 0 || ioreq->mapped == 1) {
 379         return 0;
 380     }
 381     if (ioreq->blkdev->feature_persistent) {
 382         for (i = 0; i < ioreq->v.niov; i++) {
 383             grant = g_tree_lookup(ioreq->blkdev->persistent_gnts,
 384                                     GUINT_TO_POINTER(ioreq->refs[i]));
 385
 386             if (grant != NULL) {
 387                 page[i] = grant->page;
 388                 xen_pv_printf(&ioreq->blkdev->xendev, 3,
 389                               "using persistent-grant %" PRIu32 "\n",
 390                               ioreq->refs[i]);
 391             } else {
 392                     /* Add the grant to the list of grants that
 393                      * should be mapped
 394                      */
 395                     domids[new_maps] = ioreq->domids[i];
 396                     refs[new_maps] = ioreq->refs[i];
 397                     page[i] = NULL;
 398                     new_maps++;
 399             }
 400         }
 401         /* Set the protection to RW, since grants may be reused later
 402          * with a different protection than the one needed for this request
 403          */
 404         ioreq->prot = PROT_WRITE | PROT_READ;
 405     } else {
 406         /* All grants in the request should be mapped */
 407         memcpy(refs, ioreq->refs, sizeof(refs));
 408         memcpy(domids, ioreq->domids, sizeof(domids));
 409         memset(page, 0, sizeof(page));
 410         new_maps = ioreq->v.niov;
 411     }
 412
 413     if (batch_maps && new_maps) {
 414         ioreq->pages = xengnttab_map_grant_refs
 415             (gnt, new_maps, domids, refs, ioreq->prot);
 416         if (ioreq->pages == NULL) {
 417             xen_pv_printf(&ioreq->blkdev->xendev, 0,
 418                           "can't map %d grant refs (%s, %d maps)\n",
 419                           new_maps, strerror(errno), ioreq->blkdev->cnt_map);
 420             return -1;
 421         }
 422         for (i = 0, j = 0; i < ioreq->v.niov; i++) {
 423             if (page[i] == NULL) {
 424                 page[i] = ioreq->pages + (j++) * XC_PAGE_SIZE;
 425             }
 426         }
 427         ioreq->blkdev->cnt_map += new_maps;
 428     } else if (new_maps)  {
 429         for (i = 0; i < new_maps; i++) {
 430             ioreq->page[i] = xengnttab_map_grant_ref
 431                 (gnt, domids[i], refs[i], ioreq->prot);
 432             if (ioreq->page[i] == NULL) {
 433                 xen_pv_printf(&ioreq->blkdev->xendev, 0,
 434                               "can't map grant ref %d (%s, %d maps)\n",
 435                               refs[i], strerror(errno), ioreq->blkdev->cnt_map);
 436                 ioreq->mapped = 1;
 437                 ioreq_unmap(ioreq);
 438                 return -1;
 439             }
 440             ioreq->blkdev->cnt_map++;
 441         }
 442         for (i = 0, j = 0; i < ioreq->v.niov; i++) {
 443             if (page[i] == NULL) {
 444                 page[i] = ioreq->page[j++];
 445             }
 446         }
 447     }
 448     if (ioreq->blkdev->feature_persistent && new_maps != 0 &&
 449         (!batch_maps || (ioreq->blkdev->persistent_gnt_count + new_maps <=
 450         ioreq->blkdev->max_grants))) {
 451         /*
 452          * If we are using persistent grants and batch mappings only
 453          * add the new maps to the list of persistent grants if the whole
 454          * area can be persistently mapped.
 455          */
 456         if (batch_maps) {
 457             region = g_malloc0(sizeof(*region));
 458             region->addr = ioreq->pages;
 459             region->num = new_maps;
 460             ioreq->blkdev->persistent_regions = g_slist_append(
 461                                             ioreq->blkdev->persistent_regions,
 462                                             region);
 463         }
 464         while ((ioreq->blkdev->persistent_gnt_count < ioreq->blkdev->max_grants)
 465               && new_maps) {
 466             /* Go through the list of newly mapped grants and add as many
 467              * as possible to the list of persistently mapped grants.
 468              *
 469              * Since we start at the end of ioreq->page(s), we only need
 470              * to decrease new_maps to prevent this granted pages from
 471              * being unmapped in ioreq_unmap.
 472              */
 473             grant = g_malloc0(sizeof(*grant));
 474             new_maps--;
 475             if (batch_maps) {
 476                 grant->page = ioreq->pages + (new_maps) * XC_PAGE_SIZE;
 477             } else {
 478                 grant->page = ioreq->page[new_maps];
 479             }
 480             grant->blkdev = ioreq->blkdev;
 481             xen_pv_printf(&ioreq->blkdev->xendev, 3,
 482                           "adding grant %" PRIu32 " page: %p\n",
 483                           refs[new_maps], grant->page);
 484             g_tree_insert(ioreq->blkdev->persistent_gnts,
 485                           GUINT_TO_POINTER(refs[new_maps]),
 486                           grant);
 487             ioreq->blkdev->persistent_gnt_count++;
 488         }
 489         assert(!batch_maps || new_maps == 0);
 490     }
 491     for (i = 0; i < ioreq->v.niov; i++) {
 492         ioreq->v.iov[i].iov_base += (uintptr_t)page[i];
 493     }
 494     ioreq->mapped = 1;
 495     ioreq->num_unmap = new_maps;
 496     return 0;
 497 }
 498
 499 #if CONFIG_XEN_CTRL_INTERFACE_VERSION >= 40800
 500
 501 static void ioreq_free_copy_buffers(struct ioreq *ioreq)
 502 {
 503     int i;
 504
 505     for (i = 0; i < ioreq->v.niov; i++) {
 506         ioreq->page[i] = NULL;
 507     }
 508
 509     qemu_vfree(ioreq->pages);
 510 }
 511
 512 static int ioreq_init_copy_buffers(struct ioreq *ioreq)
 513 {
 514     int i;
 515
 516     if (ioreq->v.niov == 0) {
 517         return 0;
 518     }
 519
 520     ioreq->pages = qemu_memalign(XC_PAGE_SIZE, ioreq->v.niov * XC_PAGE_SIZE);
 521
 522     for (i = 0; i < ioreq->v.niov; i++) {
 523         ioreq->page[i] = ioreq->pages + i * XC_PAGE_SIZE;
 524         ioreq->v.iov[i].iov_base = ioreq->page[i];
 525     }
 526
 527     return 0;
 528 }
 529
 530 static int ioreq_grant_copy(struct ioreq *ioreq)
 531 {
 532     xengnttab_handle *gnt = ioreq->blkdev->xendev.gnttabdev;
 533     xengnttab_grant_copy_segment_t segs[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 534     int i, count, rc;
 535     int64_t file_blk = ioreq->blkdev->file_blk;
 536
 537     if (ioreq->v.niov == 0) {
 538         return 0;
 539     }
 540
 541     count = ioreq->v.niov;
 542
 543     for (i = 0; i < count; i++) {
 544         if (ioreq->req.operation == BLKIF_OP_READ) {
 545             segs[i].flags = GNTCOPY_dest_gref;
 546             segs[i].dest.foreign.ref = ioreq->refs[i];
 547             segs[i].dest.foreign.domid = ioreq->domids[i];
 548             segs[i].dest.foreign.offset = ioreq->req.seg[i].first_sect * file_blk;
 549             segs[i].source.virt = ioreq->v.iov[i].iov_base;
 550         } else {
 551             segs[i].flags = GNTCOPY_source_gref;
 552             segs[i].source.foreign.ref = ioreq->refs[i];
 553             segs[i].source.foreign.domid = ioreq->domids[i];
 554             segs[i].source.foreign.offset = ioreq->req.seg[i].first_sect * file_blk;
 555             segs[i].dest.virt = ioreq->v.iov[i].iov_base;
 556         }
 557         segs[i].len = (ioreq->req.seg[i].last_sect
 558                        - ioreq->req.seg[i].first_sect + 1) * file_blk;
 559     }
 560
 561     rc = xengnttab_grant_copy(gnt, count, segs);
 562
 563     if (rc) {
 564         xen_pv_printf(&ioreq->blkdev->xendev, 0,
 565                       "failed to copy data %d\n", rc);
 566         ioreq->aio_errors++;
 567         return -1;
 568     }
 569
 570     for (i = 0; i < count; i++) {
 571         if (segs[i].status != GNTST_okay) {
 572             xen_pv_printf(&ioreq->blkdev->xendev, 3,
 573                           "failed to copy data %d for gref %d, domid %d\n",
 574                           segs[i].status, ioreq->refs[i], ioreq->domids[i]);
 575             ioreq->aio_errors++;
 576             rc = -1;
 577         }
 578     }
 579
 580     return rc;
 581 }
 582 #else
 583 static void ioreq_free_copy_buffers(struct ioreq *ioreq)
 584 {
 585     abort();
 586 }
 587
 588 static int ioreq_init_copy_buffers(struct ioreq *ioreq)
 589 {
 590     abort();
 591 }
 592
 593 static int ioreq_grant_copy(struct ioreq *ioreq)
 594 {
 595     abort();
 596 }
 597 #endif
 598
 599 static int ioreq_runio_qemu_aio(struct ioreq *ioreq);
 600
 601 static void qemu_aio_complete(void *opaque, int ret)
 602 {
 603     struct ioreq *ioreq = opaque;
 604     struct XenBlkDev *blkdev = ioreq->blkdev;
 605
 606     aio_context_acquire(blkdev->ctx);
 607
 608     if (ret != 0) {
 609         xen_pv_printf(&blkdev->xendev, 0, "%s I/O error\n",
 610                       ioreq->req.operation == BLKIF_OP_READ ? "read" : "write");
 611         ioreq->aio_errors++;
 612     }
 613
 614     ioreq->aio_inflight--;
 615     if (ioreq->presync) {
 616         ioreq->presync = 0;
 617         ioreq_runio_qemu_aio(ioreq);
 618         goto done;
 619     }
 620     if (ioreq->aio_inflight > 0) {
 621         goto done;
 622     }
 623
 624     if (xen_feature_grant_copy) {
 625         switch (ioreq->req.operation) {
 626         case BLKIF_OP_READ:
 627             /* in case of failure ioreq->aio_errors is increased */
 628             if (ret == 0) {
 629                 ioreq_grant_copy(ioreq);
 630             }
 631             ioreq_free_copy_buffers(ioreq);
 632             break;
 633         case BLKIF_OP_WRITE:
 634         case BLKIF_OP_FLUSH_DISKCACHE:
 635             if (!ioreq->req.nr_segments) {
 636                 break;
 637             }
 638             ioreq_free_copy_buffers(ioreq);
 639             break;
 640         default:
 641             break;
 642         }
 643     }
 644
 645     ioreq->status = ioreq->aio_errors ? BLKIF_RSP_ERROR : BLKIF_RSP_OKAY;
 646     if (!xen_feature_grant_copy) {
 647         ioreq_unmap(ioreq);
 648     }
 649     ioreq_finish(ioreq);
 650     switch (ioreq->req.operation) {
 651     case BLKIF_OP_WRITE:
 652     case BLKIF_OP_FLUSH_DISKCACHE:
 653         if (!ioreq->req.nr_segments) {
 654             break;
 655         }
 656     case BLKIF_OP_READ:
 657         if (ioreq->status == BLKIF_RSP_OKAY) {
 658             block_acct_done(blk_get_stats(blkdev->blk), &ioreq->acct);
 659         } else {
 660             block_acct_failed(blk_get_stats(blkdev->blk), &ioreq->acct);
 661         }
 662         break;
 663     case BLKIF_OP_DISCARD:
 664     default:
 665         break;
 666     }
 667     qemu_bh_schedule(blkdev->bh);
 668
 669 done:
 670     aio_context_release(blkdev->ctx);
 671 }
 672
 673 static bool blk_split_discard(struct ioreq *ioreq, blkif_sector_t sector_number,
 674                               uint64_t nr_sectors)
 675 {
 676     struct XenBlkDev *blkdev = ioreq->blkdev;
 677     int64_t byte_offset;
 678     int byte_chunk;
 679     uint64_t byte_remaining, limit;
 680     uint64_t sec_start = sector_number;
 681     uint64_t sec_count = nr_sectors;
 682
 683     /* Wrap around, or overflowing byte limit? */
 684     if (sec_start + sec_count < sec_count ||
 685         sec_start + sec_count > INT64_MAX >> BDRV_SECTOR_BITS) {
 686         return false;
 687     }
 688
 689     limit = BDRV_REQUEST_MAX_SECTORS << BDRV_SECTOR_BITS;
 690     byte_offset = sec_start << BDRV_SECTOR_BITS;
 691     byte_remaining = sec_count << BDRV_SECTOR_BITS;
 692
 693     do {
 694         byte_chunk = byte_remaining > limit ? limit : byte_remaining;
 695         ioreq->aio_inflight++;
 696         blk_aio_pdiscard(blkdev->blk, byte_offset, byte_chunk,
 697                          qemu_aio_complete, ioreq);
 698         byte_remaining -= byte_chunk;
 699         byte_offset += byte_chunk;
 700     } while (byte_remaining > 0);
 701
 702     return true;
 703 }
 704
 705 static int ioreq_runio_qemu_aio(struct ioreq *ioreq)
 706 {
 707     struct XenBlkDev *blkdev = ioreq->blkdev;
 708
 709     if (xen_feature_grant_copy) {
 710         ioreq_init_copy_buffers(ioreq);
 711         if (ioreq->req.nr_segments && (ioreq->req.operation == BLKIF_OP_WRITE ||
 712             ioreq->req.operation == BLKIF_OP_FLUSH_DISKCACHE) &&
 713             ioreq_grant_copy(ioreq)) {
 714                 ioreq_free_copy_buffers(ioreq);
 715                 goto err;
 716         }
 717     } else {
 718         if (ioreq->req.nr_segments && ioreq_map(ioreq)) {
 719             goto err;
 720         }
 721     }
 722
 723     ioreq->aio_inflight++;
 724     if (ioreq->presync) {
 725         blk_aio_flush(ioreq->blkdev->blk, qemu_aio_complete, ioreq);
 726         return 0;
 727     }
 728
 729     switch (ioreq->req.operation) {
 730     case BLKIF_OP_READ:
 731         block_acct_start(blk_get_stats(blkdev->blk), &ioreq->acct,
 732                          ioreq->v.size, BLOCK_ACCT_READ);
 733         ioreq->aio_inflight++;
 734         blk_aio_preadv(blkdev->blk, ioreq->start, &ioreq->v, 0,
 735                        qemu_aio_complete, ioreq);
 736         break;
 737     case BLKIF_OP_WRITE:
 738     case BLKIF_OP_FLUSH_DISKCACHE:
 739         if (!ioreq->req.nr_segments) {
 740             break;
 741         }
 742
 743         block_acct_start(blk_get_stats(blkdev->blk), &ioreq->acct,
 744                          ioreq->v.size,
 745                          ioreq->req.operation == BLKIF_OP_WRITE ?
 746                          BLOCK_ACCT_WRITE : BLOCK_ACCT_FLUSH);
 747         ioreq->aio_inflight++;
 748         blk_aio_pwritev(blkdev->blk, ioreq->start, &ioreq->v, 0,
 749                         qemu_aio_complete, ioreq);
 750         break;
 751     case BLKIF_OP_DISCARD:
 752     {
 753         struct blkif_request_discard *req = (void *)&ioreq->req;
 754         if (!blk_split_discard(ioreq, req->sector_number, req->nr_sectors)) {
 755             goto err;
 756         }
 757         break;
 758     }
 759     default:
 760         /* unknown operation (shouldn't happen -- parse catches this) */
 761         if (!xen_feature_grant_copy) {
 762             ioreq_unmap(ioreq);
 763         }
 764         goto err;
 765     }
 766
 767     qemu_aio_complete(ioreq, 0);
 768
 769     return 0;
 770
 771 err:
 772     ioreq_finish(ioreq);
 773     ioreq->status = BLKIF_RSP_ERROR;
 774     return -1;
 775 }
 776
 777 static int blk_send_response_one(struct ioreq *ioreq)
 778 {
 779     struct XenBlkDev  *blkdev = ioreq->blkdev;
 780     int               send_notify   = 0;
 781     int               have_requests = 0;
 782     blkif_response_t  *resp;
 783
 784     /* Place on the response ring for the relevant domain. */
 785     switch (blkdev->protocol) {
 786     case BLKIF_PROTOCOL_NATIVE:
 787         resp = (blkif_response_t *) RING_GET_RESPONSE(&blkdev->rings.native,
 788                                  blkdev->rings.native.rsp_prod_pvt);
 789         break;
 790     case BLKIF_PROTOCOL_X86_32:
 791         resp = (blkif_response_t *) RING_GET_RESPONSE(&blkdev->rings.x86_32_part,
 792                                  blkdev->rings.x86_32_part.rsp_prod_pvt);
 793         break;
 794     case BLKIF_PROTOCOL_X86_64:
 795         resp = (blkif_response_t *) RING_GET_RESPONSE(&blkdev->rings.x86_64_part,
 796                                  blkdev->rings.x86_64_part.rsp_prod_pvt);
 797         break;
 798     default:
 799         return 0;
 800     }
 801
 802     resp->id        = ioreq->req.id;
 803     resp->operation = ioreq->req.operation;
 804     resp->status    = ioreq->status;
 805
 806     blkdev->rings.common.rsp_prod_pvt++;
 807
 808     RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blkdev->rings.common, send_notify);
 809     if (blkdev->rings.common.rsp_prod_pvt == blkdev->rings.common.req_cons) {
 810         /*
 811          * Tail check for pending requests. Allows frontend to avoid
 812          * notifications if requests are already in flight (lower
 813          * overheads and promotes batching).
 814          */
 815         RING_FINAL_CHECK_FOR_REQUESTS(&blkdev->rings.common, have_requests);
 816     } else if (RING_HAS_UNCONSUMED_REQUESTS(&blkdev->rings.common)) {
 817         have_requests = 1;
 818     }
 819
 820     if (have_requests) {
 821         blkdev->more_work++;
 822     }
 823     return send_notify;
 824 }
 825
 826 /* walk finished list, send outstanding responses, free requests */
 827 static void blk_send_response_all(struct XenBlkDev *blkdev)
 828 {
 829     struct ioreq *ioreq;
 830     int send_notify = 0;
 831
 832     while (!QLIST_EMPTY(&blkdev->finished)) {
 833         ioreq = QLIST_FIRST(&blkdev->finished);
 834         send_notify += blk_send_response_one(ioreq);
 835         ioreq_release(ioreq, true);
 836     }
 837     if (send_notify) {
 838         xen_pv_send_notify(&blkdev->xendev);
 839     }
 840 }
 841
 842 static int blk_get_request(struct XenBlkDev *blkdev, struct ioreq *ioreq, RING_IDX rc)
 843 {
 844     switch (blkdev->protocol) {
 845     case BLKIF_PROTOCOL_NATIVE:
 846         memcpy(&ioreq->req, RING_GET_REQUEST(&blkdev->rings.native, rc),
 847                sizeof(ioreq->req));
 848         break;
 849     case BLKIF_PROTOCOL_X86_32:
 850         blkif_get_x86_32_req(&ioreq->req,
 851                              RING_GET_REQUEST(&blkdev->rings.x86_32_part, rc));
 852         break;
 853     case BLKIF_PROTOCOL_X86_64:
 854         blkif_get_x86_64_req(&ioreq->req,
 855                              RING_GET_REQUEST(&blkdev->rings.x86_64_part, rc));
 856         break;
 857     }
 858     /* Prevent the compiler from accessing the on-ring fields instead. */
 859     barrier();
 860     return 0;
 861 }
 862
 863 static void blk_handle_requests(struct XenBlkDev *blkdev)
 864 {
 865     RING_IDX rc, rp;
 866     struct ioreq *ioreq;
 867
 868     blkdev->more_work = 0;
 869
 870     rc = blkdev->rings.common.req_cons;
 871     rp = blkdev->rings.common.sring->req_prod;
 872     xen_rmb(); /* Ensure we see queued requests up to 'rp'. */
 873
 874     blk_send_response_all(blkdev);
 875     while (rc != rp) {
 876         /* pull request from ring */
 877         if (RING_REQUEST_CONS_OVERFLOW(&blkdev->rings.common, rc)) {
 878             break;
 879         }
 880         ioreq = ioreq_start(blkdev);
 881         if (ioreq == NULL) {
 882             blkdev->more_work++;
 883             break;
 884         }
 885         blk_get_request(blkdev, ioreq, rc);
 886         blkdev->rings.common.req_cons = ++rc;
 887
 888         /* parse them */
 889         if (ioreq_parse(ioreq) != 0) {
 890
 891             switch (ioreq->req.operation) {
 892             case BLKIF_OP_READ:
 893                 block_acct_invalid(blk_get_stats(blkdev->blk),
 894                                    BLOCK_ACCT_READ);
 895                 break;
 896             case BLKIF_OP_WRITE:
 897                 block_acct_invalid(blk_get_stats(blkdev->blk),
 898                                    BLOCK_ACCT_WRITE);
 899                 break;
 900             case BLKIF_OP_FLUSH_DISKCACHE:
 901                 block_acct_invalid(blk_get_stats(blkdev->blk),
 902                                    BLOCK_ACCT_FLUSH);
 903             default:
 904                 break;
 905             };
 906
 907             if (blk_send_response_one(ioreq)) {
 908                 xen_pv_send_notify(&blkdev->xendev);
 909             }
 910             ioreq_release(ioreq, false);
 911             continue;
 912         }
 913
 914         ioreq_runio_qemu_aio(ioreq);
 915     }
 916
 917     if (blkdev->more_work && blkdev->requests_inflight < blkdev->max_requests) {
 918         qemu_bh_schedule(blkdev->bh);
 919     }
 920 }
 921
 922 /* ------------------------------------------------------------- */
 923
 924 static void blk_bh(void *opaque)
 925 {
 926     struct XenBlkDev *blkdev = opaque;
 927
 928     aio_context_acquire(blkdev->ctx);
 929     blk_handle_requests(blkdev);
 930     aio_context_release(blkdev->ctx);
 931 }
 932
 933 static void blk_alloc(struct XenDevice *xendev)
 934 {
 935     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
 936     Error *err = NULL;
 937
 938     trace_xen_disk_alloc(xendev->name);
 939
 940     QLIST_INIT(&blkdev->inflight);
 941     QLIST_INIT(&blkdev->finished);
 942     QLIST_INIT(&blkdev->freelist);
 943
 944     blkdev->iothread = iothread_create(xendev->name, &err);
 945     assert(!err);
 946
 947     blkdev->ctx = iothread_get_aio_context(blkdev->iothread);
 948     blkdev->bh = aio_bh_new(blkdev->ctx, blk_bh, blkdev);
 949
 950     if (xen_mode != XEN_EMULATE) {
 951         batch_maps = 1;
 952     }
 953 }
 954
 955 static void blk_parse_discard(struct XenBlkDev *blkdev)
 956 {
 957     int enable;
 958
 959     blkdev->feature_discard = true;
 960
 961     if (xenstore_read_be_int(&blkdev->xendev, "discard-enable", &enable) == 0) {
 962         blkdev->feature_discard = !!enable;
 963     }
 964
 965     if (blkdev->feature_discard) {
 966         xenstore_write_be_int(&blkdev->xendev, "feature-discard", 1);
 967     }
 968 }
 969
 970 static int blk_init(struct XenDevice *xendev)
 971 {
 972     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
 973     int info = 0;
 974     char *directiosafe = NULL;
 975
 976     trace_xen_disk_init(xendev->name);
 977
 978     /* read xenstore entries */
 979     if (blkdev->params == NULL) {
 980         char *h = NULL;
 981         blkdev->params = xenstore_read_be_str(&blkdev->xendev, "params");
 982         if (blkdev->params != NULL) {
 983             h = strchr(blkdev->params, ':');
 984         }
 985         if (h != NULL) {
 986             blkdev->fileproto = blkdev->params;
 987             blkdev->filename  = h+1;
 988             *h = 0;
 989         } else {
 990             blkdev->fileproto = "<unset>";
 991             blkdev->filename  = blkdev->params;
 992         }
 993     }
 994     if (!strcmp("aio", blkdev->fileproto)) {
 995         blkdev->fileproto = "raw";
 996     }
 997     if (!strcmp("vhd", blkdev->fileproto)) {
 998         blkdev->fileproto = "vpc";
 999     }
1000     if (blkdev->mode == NULL) {
1001         blkdev->mode = xenstore_read_be_str(&blkdev->xendev, "mode");
1002     }
1003     if (blkdev->type == NULL) {
1004         blkdev->type = xenstore_read_be_str(&blkdev->xendev, "type");
1005     }
1006     if (blkdev->dev == NULL) {
1007         blkdev->dev = xenstore_read_be_str(&blkdev->xendev, "dev");
1008     }
1009     if (blkdev->devtype == NULL) {
1010         blkdev->devtype = xenstore_read_be_str(&blkdev->xendev, "device-type");
1011     }
1012     directiosafe = xenstore_read_be_str(&blkdev->xendev, "direct-io-safe");
1013     blkdev->directiosafe = (directiosafe && atoi(directiosafe));
1014
1015     /* do we have all we need? */
1016     if (blkdev->params == NULL ||
1017         blkdev->mode == NULL   ||
1018         blkdev->type == NULL   ||
1019         blkdev->dev == NULL) {
1020         goto out_error;
1021     }
1022
1023     /* read-only ? */
1024     if (strcmp(blkdev->mode, "w")) {
1025         info  |= VDISK_READONLY;
1026     }
1027
1028     /* cdrom ? */
1029     if (blkdev->devtype && !strcmp(blkdev->devtype, "cdrom")) {
1030         info  |= VDISK_CDROM;
1031     }
1032
1033     blkdev->file_blk  = BLOCK_SIZE;
1034
1035     xen_pv_printf(&blkdev->xendev, 3, "grant copy operation %s\n",
1036                   xen_feature_grant_copy ? "enabled" : "disabled");
1037
1038     /* fill info
1039      * blk_connect supplies sector-size and sectors
1040      */
1041     xenstore_write_be_int(&blkdev->xendev, "feature-flush-cache", 1);
1042     xenstore_write_be_int(&blkdev->xendev, "feature-persistent",
1043                           !xen_feature_grant_copy);
1044     xenstore_write_be_int(&blkdev->xendev, "info", info);
1045
1046     xenstore_write_be_int(&blkdev->xendev, "max-ring-page-order",
1047                           MAX_RING_PAGE_ORDER);
1048
1049     blk_parse_discard(blkdev);
1050
1051     g_free(directiosafe);
1052     return 0;
1053
1054 out_error:
1055     g_free(blkdev->params);
1056     blkdev->params = NULL;
1057     g_free(blkdev->mode);
1058     blkdev->mode = NULL;
1059     g_free(blkdev->type);
1060     blkdev->type = NULL;
1061     g_free(blkdev->dev);
1062     blkdev->dev = NULL;
1063     g_free(blkdev->devtype);
1064     blkdev->devtype = NULL;
1065     g_free(directiosafe);
1066     blkdev->directiosafe = false;
1067     return -1;
1068 }
1069
1070 /*
1071  * We need to account for the grant allocations requiring contiguous
1072  * chunks; the worst case number would be
1073  *     max_req * max_seg + (max_req - 1) * (max_seg - 1) + 1,
1074  * but in order to keep things simple just use
1075  *     2 * max_req * max_seg.
1076  */
1077 #define MAX_GRANTS(max_req, max_seg) (2 * (max_req) * (max_seg))
1078
1079 static int blk_connect(struct XenDevice *xendev)
1080 {
1081     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
1082     int pers, index, qflags;
1083     bool readonly = true;
1084     bool writethrough = true;
1085     int order, ring_ref;
1086     unsigned int ring_size, max_grants;
1087     unsigned int i;
1088     uint32_t *domids;
1089
1090     trace_xen_disk_connect(xendev->name);
1091
1092     /* read-only ? */
1093     if (blkdev->directiosafe) {
1094         qflags = BDRV_O_NOCACHE | BDRV_O_NATIVE_AIO;
1095     } else {
1096         qflags = 0;
1097         writethrough = false;
1098     }
1099     if (strcmp(blkdev->mode, "w") == 0) {
1100         qflags |= BDRV_O_RDWR;
1101         readonly = false;
1102     }
1103     if (blkdev->feature_discard) {
1104         qflags |= BDRV_O_UNMAP;
1105     }
1106
1107     /* init qemu block driver */
1108     index = (blkdev->xendev.dev - 202 * 256) / 16;
1109     blkdev->dinfo = drive_get(IF_XEN, 0, index);
1110     if (!blkdev->dinfo) {
1111         Error *local_err = NULL;
1112         QDict *options = NULL;
1113
1114         if (strcmp(blkdev->fileproto, "<unset>")) {
1115             options = qdict_new();
1116             qdict_put_str(options, "driver", blkdev->fileproto);
1117         }
1118
1119         /* setup via xenbus -> create new block driver instance */
1120         xen_pv_printf(&blkdev->xendev, 2, "create new bdrv (xenbus setup)\n");
1121         blkdev->blk = blk_new_open(blkdev->filename, NULL, options,
1122                                    qflags, &local_err);
1123         if (!blkdev->blk) {
1124             xen_pv_printf(&blkdev->xendev, 0, "error: %s\n",
1125                           error_get_pretty(local_err));
1126             error_free(local_err);
1127             return -1;
1128         }
1129         blk_set_enable_write_cache(blkdev->blk, !writethrough);
1130     } else {
1131         /* setup via qemu cmdline -> already setup for us */
1132         xen_pv_printf(&blkdev->xendev, 2,
1133                       "get configured bdrv (cmdline setup)\n");
1134         blkdev->blk = blk_by_legacy_dinfo(blkdev->dinfo);
1135         if (blk_is_read_only(blkdev->blk) && !readonly) {
1136             xen_pv_printf(&blkdev->xendev, 0, "Unexpected read-only drive");
1137             blkdev->blk = NULL;
1138             return -1;
1139         }
1140         /* blkdev->blk is not create by us, we get a reference
1141          * so we can blk_unref() unconditionally */
1142         blk_ref(blkdev->blk);
1143     }
1144     blk_attach_dev_legacy(blkdev->blk, blkdev);
1145     blkdev->file_size = blk_getlength(blkdev->blk);
1146     if (blkdev->file_size < 0) {
1147         BlockDriverState *bs = blk_bs(blkdev->blk);
1148         const char *drv_name = bs ? bdrv_get_format_name(bs) : NULL;
1149         xen_pv_printf(&blkdev->xendev, 1, "blk_getlength: %d (%s) | drv %s\n",
1150                       (int)blkdev->file_size, strerror(-blkdev->file_size),
1151                       drv_name ?: "-");
1152         blkdev->file_size = 0;
1153     }
1154
1155     xen_pv_printf(xendev, 1, "type \"%s\", fileproto \"%s\", filename \"%s\","
1156                   " size %" PRId64 " (%" PRId64 " MB)\n",
1157                   blkdev->type, blkdev->fileproto, blkdev->filename,
1158                   blkdev->file_size, blkdev->file_size >> 20);
1159
1160     /* Fill in number of sector size and number of sectors */
1161     xenstore_write_be_int(&blkdev->xendev, "sector-size", blkdev->file_blk);
1162     xenstore_write_be_int64(&blkdev->xendev, "sectors",
1163                             blkdev->file_size / blkdev->file_blk);
1164
1165     if (xenstore_read_fe_int(&blkdev->xendev, "ring-page-order",
1166                              &order) == -1) {
1167         blkdev->nr_ring_ref = 1;
1168
1169         if (xenstore_read_fe_int(&blkdev->xendev, "ring-ref",
1170                                  &ring_ref) == -1) {
1171             return -1;
1172         }
1173         blkdev->ring_ref[0] = ring_ref;
1174
1175     } else if (order >= 0 && order <= MAX_RING_PAGE_ORDER) {
1176         blkdev->nr_ring_ref = 1 << order;
1177
1178         for (i = 0; i < blkdev->nr_ring_ref; i++) {
1179             char *key;
1180
1181             key = g_strdup_printf("ring-ref%u", i);
1182             if (!key) {
1183                 return -1;
1184             }
1185
1186             if (xenstore_read_fe_int(&blkdev->xendev, key,
1187                                      &ring_ref) == -1) {
1188                 g_free(key);
1189                 return -1;
1190             }
1191             blkdev->ring_ref[i] = ring_ref;
1192
1193             g_free(key);
1194         }
1195     } else {
1196         xen_pv_printf(xendev, 0, "invalid ring-page-order: %d\n",
1197                       order);
1198         return -1;
1199     }
1200
1201     if (xenstore_read_fe_int(&blkdev->xendev, "event-channel",
1202                              &blkdev->xendev.remote_port) == -1) {
1203         return -1;
1204     }
1205     if (xenstore_read_fe_int(&blkdev->xendev, "feature-persistent", &pers)) {
1206         blkdev->feature_persistent = FALSE;
1207     } else {
1208         blkdev->feature_persistent = !!pers;
1209     }
1210
1211     if (!blkdev->xendev.protocol) {
1212         blkdev->protocol = BLKIF_PROTOCOL_NATIVE;
1213     } else if (strcmp(blkdev->xendev.protocol, XEN_IO_PROTO_ABI_NATIVE) == 0) {
1214         blkdev->protocol = BLKIF_PROTOCOL_NATIVE;
1215     } else if (strcmp(blkdev->xendev.protocol, XEN_IO_PROTO_ABI_X86_32) == 0) {
1216         blkdev->protocol = BLKIF_PROTOCOL_X86_32;
1217     } else if (strcmp(blkdev->xendev.protocol, XEN_IO_PROTO_ABI_X86_64) == 0) {
1218         blkdev->protocol = BLKIF_PROTOCOL_X86_64;
1219     } else {
1220         blkdev->protocol = BLKIF_PROTOCOL_NATIVE;
1221     }
1222
1223     ring_size = XC_PAGE_SIZE * blkdev->nr_ring_ref;
1224     switch (blkdev->protocol) {
1225     case BLKIF_PROTOCOL_NATIVE:
1226     {
1227         blkdev->max_requests = __CONST_RING_SIZE(blkif, ring_size);
1228         break;
1229     }
1230     case BLKIF_PROTOCOL_X86_32:
1231     {
1232         blkdev->max_requests = __CONST_RING_SIZE(blkif_x86_32, ring_size);
1233         break;
1234     }
1235     case BLKIF_PROTOCOL_X86_64:
1236     {
1237         blkdev->max_requests = __CONST_RING_SIZE(blkif_x86_64, ring_size);
1238         break;
1239     }
1240     default:
1241         return -1;
1242     }
1243
1244     /* Calculate the maximum number of grants needed by ioreqs */
1245     max_grants = MAX_GRANTS(blkdev->max_requests,
1246                             BLKIF_MAX_SEGMENTS_PER_REQUEST);
1247     /* Add on the number needed for the ring pages */
1248     max_grants += blkdev->nr_ring_ref;
1249
1250     blkdev->xendev.gnttabdev = xengnttab_open(NULL, 0);
1251     if (blkdev->xendev.gnttabdev == NULL) {
1252         xen_pv_printf(xendev, 0, "xengnttab_open failed: %s\n",
1253                       strerror(errno));
1254         return -1;
1255     }
1256     if (xengnttab_set_max_grants(blkdev->xendev.gnttabdev, max_grants)) {
1257         xen_pv_printf(xendev, 0, "xengnttab_set_max_grants failed: %s\n",
1258                       strerror(errno));
1259         return -1;
1260     }
1261
1262     domids = g_new0(uint32_t, blkdev->nr_ring_ref);
1263     for (i = 0; i < blkdev->nr_ring_ref; i++) {
1264         domids[i] = blkdev->xendev.dom;
1265     }
1266
1267     blkdev->sring = xengnttab_map_grant_refs(blkdev->xendev.gnttabdev,
1268                                              blkdev->nr_ring_ref,
1269                                              domids,
1270                                              blkdev->ring_ref,
1271                                              PROT_READ | PROT_WRITE);
1272
1273     g_free(domids);
1274
1275     if (!blkdev->sring) {
1276         return -1;
1277     }
1278
1279     blkdev->cnt_map++;
1280
1281     switch (blkdev->protocol) {
1282     case BLKIF_PROTOCOL_NATIVE:
1283     {
1284         blkif_sring_t *sring_native = blkdev->sring;
1285         BACK_RING_INIT(&blkdev->rings.native, sring_native, ring_size);
1286         break;
1287     }
1288     case BLKIF_PROTOCOL_X86_32:
1289     {
1290         blkif_x86_32_sring_t *sring_x86_32 = blkdev->sring;
1291
1292         BACK_RING_INIT(&blkdev->rings.x86_32_part, sring_x86_32, ring_size);
1293         break;
1294     }
1295     case BLKIF_PROTOCOL_X86_64:
1296     {
1297         blkif_x86_64_sring_t *sring_x86_64 = blkdev->sring;
1298
1299         BACK_RING_INIT(&blkdev->rings.x86_64_part, sring_x86_64, ring_size);
1300         break;
1301     }
1302     }
1303
1304     if (blkdev->feature_persistent) {
1305         /* Init persistent grants */
1306         blkdev->max_grants = blkdev->max_requests *
1307             BLKIF_MAX_SEGMENTS_PER_REQUEST;
1308         blkdev->persistent_gnts = g_tree_new_full((GCompareDataFunc)int_cmp,
1309                                              NULL, NULL,
1310                                              batch_maps ?
1311                                              (GDestroyNotify)g_free :
1312                                              (GDestroyNotify)destroy_grant);
1313         blkdev->persistent_regions = NULL;
1314         blkdev->persistent_gnt_count = 0;
1315     }
1316
1317     blk_set_aio_context(blkdev->blk, blkdev->ctx);
1318
1319     xen_be_bind_evtchn(&blkdev->xendev);
1320
1321     xen_pv_printf(&blkdev->xendev, 1, "ok: proto %s, nr-ring-ref %u, "
1322                   "remote port %d, local port %d\n",
1323                   blkdev->xendev.protocol, blkdev->nr_ring_ref,
1324                   blkdev->xendev.remote_port, blkdev->xendev.local_port);
1325     return 0;
1326 }
1327
1328 static void blk_disconnect(struct XenDevice *xendev)
1329 {
1330     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
1331
1332     trace_xen_disk_disconnect(xendev->name);
1333
1334     aio_context_acquire(blkdev->ctx);
1335
1336     if (blkdev->blk) {
1337         blk_set_aio_context(blkdev->blk, qemu_get_aio_context());
1338         blk_detach_dev(blkdev->blk, blkdev);
1339         blk_unref(blkdev->blk);
1340         blkdev->blk = NULL;
1341     }
1342     xen_pv_unbind_evtchn(&blkdev->xendev);
1343
1344     aio_context_release(blkdev->ctx);
1345
1346     if (blkdev->sring) {
1347         xengnttab_unmap(blkdev->xendev.gnttabdev, blkdev->sring,
1348                         blkdev->nr_ring_ref);
1349         blkdev->cnt_map--;
1350         blkdev->sring = NULL;
1351     }
1352
1353     /*
1354      * Unmap persistent grants before switching to the closed state
1355      * so the frontend can free them.
1356      *
1357      * In the !batch_maps case g_tree_destroy will take care of unmapping
1358      * the grant, but in the batch_maps case we need to iterate over every
1359      * region in persistent_regions and unmap it.
1360      */
1361     if (blkdev->feature_persistent) {
1362         g_tree_destroy(blkdev->persistent_gnts);
1363         assert(batch_maps || blkdev->persistent_gnt_count == 0);
1364         if (batch_maps) {
1365             blkdev->persistent_gnt_count = 0;
1366             g_slist_foreach(blkdev->persistent_regions,
1367                             (GFunc)remove_persistent_region, blkdev);
1368             g_slist_free(blkdev->persistent_regions);
1369         }
1370         blkdev->feature_persistent = false;
1371     }
1372
1373     if (blkdev->xendev.gnttabdev) {
1374         xengnttab_close(blkdev->xendev.gnttabdev);
1375         blkdev->xendev.gnttabdev = NULL;
1376     }
1377 }
1378
1379 static int blk_free(struct XenDevice *xendev)
1380 {
1381     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
1382     struct ioreq *ioreq;
1383
1384     trace_xen_disk_free(xendev->name);
1385
1386     blk_disconnect(xendev);
1387
1388     while (!QLIST_EMPTY(&blkdev->freelist)) {
1389         ioreq = QLIST_FIRST(&blkdev->freelist);
1390         QLIST_REMOVE(ioreq, list);
1391         qemu_iovec_destroy(&ioreq->v);
1392         g_free(ioreq);
1393     }
1394
1395     g_free(blkdev->params);
1396     g_free(blkdev->mode);
1397     g_free(blkdev->type);
1398     g_free(blkdev->dev);
1399     g_free(blkdev->devtype);
1400     qemu_bh_delete(blkdev->bh);
1401     iothread_destroy(blkdev->iothread);
1402     return 0;
1403 }
1404
1405 static void blk_event(struct XenDevice *xendev)
1406 {
1407     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
1408
1409     qemu_bh_schedule(blkdev->bh);
1410 }
1411
1412 struct XenDevOps xen_blkdev_ops = {
1413     .size       = sizeof(struct XenBlkDev),
1414     .alloc      = blk_alloc,
1415     .init       = blk_init,
1416     .initialise    = blk_connect,
1417     .disconnect = blk_disconnect,
1418     .event      = blk_event,
1419     .free       = blk_free,
1420 };