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