qemu/hw/block/xen_disk.c

   1 /*
   2  *  xen paravirt block device backend
   3  *
   4  *  (c) Gerd Hoffmann <kraxel@redhat.com>
   5  *
   6  *  This program is free software; you can redistribute it and/or modify
   7  *  it under the terms of the GNU General Public License as published by
   8  *  the Free Software Foundation; under version 2 of the License.
   9  *
  10  *  This program is distributed in the hope that it will be useful,
  11  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  *  GNU General Public License for more details.
  14  *
  15  *  You should have received a copy of the GNU General Public License along
  16  *  with this program; if not, see <http://www.gnu.org/licenses/>.
  17  *
  18  *  Contributions after 2012-01-13 are licensed under the terms of the
  19  *  GNU GPL, version 2 or (at your option) any later version.
  20  */
  21
  22 #include "qemu/osdep.h"
  23 #include <sys/ioctl.h>
  24 #include <sys/mman.h>
  25 #include <sys/uio.h>
  26
  27 #include "hw/hw.h"
  28 #include "hw/xen/xen_backend.h"
  29 #include "xen_blkif.h"
  30 #include "sysemu/blockdev.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
  36 /* ------------------------------------------------------------- */
  37
  38 static int batch_maps   = 0;
  39
  40 static int max_requests = 32;
  41
  42 /* ------------------------------------------------------------- */
  43
  44 #define BLOCK_SIZE  512
  45 #define IOCB_COUNT  (BLKIF_MAX_SEGMENTS_PER_REQUEST + 2)
  46
  47 struct PersistentGrant {
  48     void *page;
  49     struct XenBlkDev *blkdev;
  50 };
  51
  52 typedef struct PersistentGrant PersistentGrant;
  53
  54 struct PersistentRegion {
  55     void *addr;
  56     int num;
  57 };
  58
  59 typedef struct PersistentRegion PersistentRegion;
  60
  61 struct ioreq {
  62     blkif_request_t     req;
  63     int16_t             status;
  64
  65     /* parsed request */
  66     off_t               start;
  67     QEMUIOVector        v;
  68     int                 presync;
  69     uint8_t             mapped;
  70
  71     /* grant mapping */
  72     uint32_t            domids[BLKIF_MAX_SEGMENTS_PER_REQUEST];
  73     uint32_t            refs[BLKIF_MAX_SEGMENTS_PER_REQUEST];
  74     int                 prot;
  75     void                *page[BLKIF_MAX_SEGMENTS_PER_REQUEST];
  76     void                *pages;
  77     int                 num_unmap;
  78
  79     /* aio status */
  80     int                 aio_inflight;
  81     int                 aio_errors;
  82
  83     struct XenBlkDev    *blkdev;
  84     QLIST_ENTRY(ioreq)   list;
  85     BlockAcctCookie     acct;
  86 };
  87
  88 struct XenBlkDev {
  89     struct XenDevice    xendev;  /* must be first */
  90     char                *params;
  91     char                *mode;
  92     char                *type;
  93     char                *dev;
  94     char                *devtype;
  95     bool                directiosafe;
  96     const char          *fileproto;
  97     const char          *filename;
  98     int                 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
 115     /* Persistent grants extension */
 116     gboolean            feature_discard;
 117     gboolean            feature_persistent;
 118     GTree               *persistent_gnts;
 119     GSList              *persistent_regions;
 120     unsigned int        persistent_gnt_count;
 121     unsigned int        max_grants;
 122
 123     /* qemu block driver */
 124     DriveInfo           *dinfo;
 125     BlockBackend        *blk;
 126     QEMUBH              *bh;
 127 };
 128
 129 /* ------------------------------------------------------------- */
 130
 131 static void ioreq_reset(struct ioreq *ioreq)
 132 {
 133     memset(&ioreq->req, 0, sizeof(ioreq->req));
 134     ioreq->status = 0;
 135     ioreq->start = 0;
 136     ioreq->presync = 0;
 137     ioreq->mapped = 0;
 138
 139     memset(ioreq->domids, 0, sizeof(ioreq->domids));
 140     memset(ioreq->refs, 0, sizeof(ioreq->refs));
 141     ioreq->prot = 0;
 142     memset(ioreq->page, 0, sizeof(ioreq->page));
 143     ioreq->pages = NULL;
 144
 145     ioreq->aio_inflight = 0;
 146     ioreq->aio_errors = 0;
 147
 148     ioreq->blkdev = NULL;
 149     memset(&ioreq->list, 0, sizeof(ioreq->list));
 150     memset(&ioreq->acct, 0, sizeof(ioreq->acct));
 151
 152     qemu_iovec_reset(&ioreq->v);
 153 }
 154
 155 static gint int_cmp(gconstpointer a, gconstpointer b, gpointer user_data)
 156 {
 157     uint ua = GPOINTER_TO_UINT(a);
 158     uint ub = GPOINTER_TO_UINT(b);
 159     return (ua > ub) - (ua < ub);
 160 }
 161
 162 static void destroy_grant(gpointer pgnt)
 163 {
 164     PersistentGrant *grant = pgnt;
 165     xengnttab_handle *gnt = grant->blkdev->xendev.gnttabdev;
 166
 167     if (xengnttab_unmap(gnt, grant->page, 1) != 0) {
 168         xen_be_printf(&grant->blkdev->xendev, 0,
 169                       "xengnttab_unmap failed: %s\n",
 170                       strerror(errno));
 171     }
 172     grant->blkdev->persistent_gnt_count--;
 173     xen_be_printf(&grant->blkdev->xendev, 3,
 174                   "unmapped grant %p\n", grant->page);
 175     g_free(grant);
 176 }
 177
 178 static void remove_persistent_region(gpointer data, gpointer dev)
 179 {
 180     PersistentRegion *region = data;
 181     struct XenBlkDev *blkdev = dev;
 182     xengnttab_handle *gnt = blkdev->xendev.gnttabdev;
 183
 184     if (xengnttab_unmap(gnt, region->addr, region->num) != 0) {
 185         xen_be_printf(&blkdev->xendev, 0,
 186                       "xengnttab_unmap region %p failed: %s\n",
 187                       region->addr, strerror(errno));
 188     }
 189     xen_be_printf(&blkdev->xendev, 3,
 190                   "unmapped grant region %p with %d pages\n",
 191                   region->addr, region->num);
 192     g_free(region);
 193 }
 194
 195 static struct ioreq *ioreq_start(struct XenBlkDev *blkdev)
 196 {
 197     struct ioreq *ioreq = NULL;
 198
 199     if (QLIST_EMPTY(&blkdev->freelist)) {
 200         if (blkdev->requests_total >= max_requests) {
 201             goto out;
 202         }
 203         /* allocate new struct */
 204         ioreq = g_malloc0(sizeof(*ioreq));
 205         ioreq->blkdev = blkdev;
 206         blkdev->requests_total++;
 207         qemu_iovec_init(&ioreq->v, BLKIF_MAX_SEGMENTS_PER_REQUEST);
 208     } else {
 209         /* get one from freelist */
 210         ioreq = QLIST_FIRST(&blkdev->freelist);
 211         QLIST_REMOVE(ioreq, list);
 212     }
 213     QLIST_INSERT_HEAD(&blkdev->inflight, ioreq, list);
 214     blkdev->requests_inflight++;
 215
 216 out:
 217     return ioreq;
 218 }
 219
 220 static void ioreq_finish(struct ioreq *ioreq)
 221 {
 222     struct XenBlkDev *blkdev = ioreq->blkdev;
 223
 224     QLIST_REMOVE(ioreq, list);
 225     QLIST_INSERT_HEAD(&blkdev->finished, ioreq, list);
 226     blkdev->requests_inflight--;
 227     blkdev->requests_finished++;
 228 }
 229
 230 static void ioreq_release(struct ioreq *ioreq, bool finish)
 231 {
 232     struct XenBlkDev *blkdev = ioreq->blkdev;
 233
 234     QLIST_REMOVE(ioreq, list);
 235     ioreq_reset(ioreq);
 236     ioreq->blkdev = blkdev;
 237     QLIST_INSERT_HEAD(&blkdev->freelist, ioreq, list);
 238     if (finish) {
 239         blkdev->requests_finished--;
 240     } else {
 241         blkdev->requests_inflight--;
 242     }
 243 }
 244
 245 /*
 246  * translate request into iovec + start offset
 247  * do sanity checks along the way
 248  */
 249 static int ioreq_parse(struct ioreq *ioreq)
 250 {
 251     struct XenBlkDev *blkdev = ioreq->blkdev;
 252     uintptr_t mem;
 253     size_t len;
 254     int i;
 255
 256     xen_be_printf(&blkdev->xendev, 3,
 257                   "op %d, nr %d, handle %d, id %" PRId64 ", sector %" PRId64 "\n",
 258                   ioreq->req.operation, ioreq->req.nr_segments,
 259                   ioreq->req.handle, ioreq->req.id, ioreq->req.sector_number);
 260     switch (ioreq->req.operation) {
 261     case BLKIF_OP_READ:
 262         ioreq->prot = PROT_WRITE; /* to memory */
 263         break;
 264     case BLKIF_OP_FLUSH_DISKCACHE:
 265         ioreq->presync = 1;
 266         if (!ioreq->req.nr_segments) {
 267             return 0;
 268         }
 269         /* fall through */
 270     case BLKIF_OP_WRITE:
 271         ioreq->prot = PROT_READ; /* from memory */
 272         break;
 273     case BLKIF_OP_DISCARD:
 274         return 0;
 275     default:
 276         xen_be_printf(&blkdev->xendev, 0, "error: unknown operation (%d)\n",
 277                       ioreq->req.operation);
 278         goto err;
 279     };
 280
 281     if (ioreq->req.operation != BLKIF_OP_READ && blkdev->mode[0] != 'w') {
 282         xen_be_printf(&blkdev->xendev, 0, "error: write req for ro device\n");
 283         goto err;
 284     }
 285
 286     ioreq->start = ioreq->req.sector_number * blkdev->file_blk;
 287     for (i = 0; i < ioreq->req.nr_segments; i++) {
 288         if (i == BLKIF_MAX_SEGMENTS_PER_REQUEST) {
 289             xen_be_printf(&blkdev->xendev, 0, "error: nr_segments too big\n");
 290             goto err;
 291         }
 292         if (ioreq->req.seg[i].first_sect > ioreq->req.seg[i].last_sect) {
 293             xen_be_printf(&blkdev->xendev, 0, "error: first > last sector\n");
 294             goto err;
 295         }
 296         if (ioreq->req.seg[i].last_sect * BLOCK_SIZE >= XC_PAGE_SIZE) {
 297             xen_be_printf(&blkdev->xendev, 0, "error: page crossing\n");
 298             goto err;
 299         }
 300
 301         ioreq->domids[i] = blkdev->xendev.dom;
 302         ioreq->refs[i]   = ioreq->req.seg[i].gref;
 303
 304         mem = ioreq->req.seg[i].first_sect * blkdev->file_blk;
 305         len = (ioreq->req.seg[i].last_sect - ioreq->req.seg[i].first_sect + 1) * blkdev->file_blk;
 306         qemu_iovec_add(&ioreq->v, (void*)mem, len);
 307     }
 308     if (ioreq->start + ioreq->v.size > blkdev->file_size) {
 309         xen_be_printf(&blkdev->xendev, 0, "error: access beyond end of file\n");
 310         goto err;
 311     }
 312     return 0;
 313
 314 err:
 315     ioreq->status = BLKIF_RSP_ERROR;
 316     return -1;
 317 }
 318
 319 static void ioreq_unmap(struct ioreq *ioreq)
 320 {
 321     xengnttab_handle *gnt = ioreq->blkdev->xendev.gnttabdev;
 322     int i;
 323
 324     if (ioreq->num_unmap == 0 || ioreq->mapped == 0) {
 325         return;
 326     }
 327     if (batch_maps) {
 328         if (!ioreq->pages) {
 329             return;
 330         }
 331         if (xengnttab_unmap(gnt, ioreq->pages, ioreq->num_unmap) != 0) {
 332             xen_be_printf(&ioreq->blkdev->xendev, 0,
 333                           "xengnttab_unmap failed: %s\n",
 334                           strerror(errno));
 335         }
 336         ioreq->blkdev->cnt_map -= ioreq->num_unmap;
 337         ioreq->pages = NULL;
 338     } else {
 339         for (i = 0; i < ioreq->num_unmap; i++) {
 340             if (!ioreq->page[i]) {
 341                 continue;
 342             }
 343             if (xengnttab_unmap(gnt, ioreq->page[i], 1) != 0) {
 344                 xen_be_printf(&ioreq->blkdev->xendev, 0,
 345                               "xengnttab_unmap failed: %s\n",
 346                               strerror(errno));
 347             }
 348             ioreq->blkdev->cnt_map--;
 349             ioreq->page[i] = NULL;
 350         }
 351     }
 352     ioreq->mapped = 0;
 353 }
 354
 355 static int ioreq_map(struct ioreq *ioreq)
 356 {
 357     xengnttab_handle *gnt = ioreq->blkdev->xendev.gnttabdev;
 358     uint32_t domids[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 359     uint32_t refs[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 360     void *page[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 361     int i, j, new_maps = 0;
 362     PersistentGrant *grant;
 363     PersistentRegion *region;
 364     /* domids and refs variables will contain the information necessary
 365      * to map the grants that are needed to fulfill this request.
 366      *
 367      * After mapping the needed grants, the page array will contain the
 368      * memory address of each granted page in the order specified in ioreq
 369      * (disregarding if it's a persistent grant or not).
 370      */
 371
 372     if (ioreq->v.niov == 0 || ioreq->mapped == 1) {
 373         return 0;
 374     }
 375     if (ioreq->blkdev->feature_persistent) {
 376         for (i = 0; i < ioreq->v.niov; i++) {
 377             grant = g_tree_lookup(ioreq->blkdev->persistent_gnts,
 378                                     GUINT_TO_POINTER(ioreq->refs[i]));
 379
 380             if (grant != NULL) {
 381                 page[i] = grant->page;
 382                 xen_be_printf(&ioreq->blkdev->xendev, 3,
 383                               "using persistent-grant %" PRIu32 "\n",
 384                               ioreq->refs[i]);
 385             } else {
 386                     /* Add the grant to the list of grants that
 387                      * should be mapped
 388                      */
 389                     domids[new_maps] = ioreq->domids[i];
 390                     refs[new_maps] = ioreq->refs[i];
 391                     page[i] = NULL;
 392                     new_maps++;
 393             }
 394         }
 395         /* Set the protection to RW, since grants may be reused later
 396          * with a different protection than the one needed for this request
 397          */
 398         ioreq->prot = PROT_WRITE | PROT_READ;
 399     } else {
 400         /* All grants in the request should be mapped */
 401         memcpy(refs, ioreq->refs, sizeof(refs));
 402         memcpy(domids, ioreq->domids, sizeof(domids));
 403         memset(page, 0, sizeof(page));
 404         new_maps = ioreq->v.niov;
 405     }
 406
 407     if (batch_maps && new_maps) {
 408         ioreq->pages = xengnttab_map_grant_refs
 409             (gnt, new_maps, domids, refs, ioreq->prot);
 410         if (ioreq->pages == NULL) {
 411             xen_be_printf(&ioreq->blkdev->xendev, 0,
 412                           "can't map %d grant refs (%s, %d maps)\n",
 413                           new_maps, strerror(errno), ioreq->blkdev->cnt_map);
 414             return -1;
 415         }
 416         for (i = 0, j = 0; i < ioreq->v.niov; i++) {
 417             if (page[i] == NULL) {
 418                 page[i] = ioreq->pages + (j++) * XC_PAGE_SIZE;
 419             }
 420         }
 421         ioreq->blkdev->cnt_map += new_maps;
 422     } else if (new_maps)  {
 423         for (i = 0; i < new_maps; i++) {
 424             ioreq->page[i] = xengnttab_map_grant_ref
 425                 (gnt, domids[i], refs[i], ioreq->prot);
 426             if (ioreq->page[i] == NULL) {
 427                 xen_be_printf(&ioreq->blkdev->xendev, 0,
 428                               "can't map grant ref %d (%s, %d maps)\n",
 429                               refs[i], strerror(errno), ioreq->blkdev->cnt_map);
 430                 ioreq->mapped = 1;
 431                 ioreq_unmap(ioreq);
 432                 return -1;
 433             }
 434             ioreq->blkdev->cnt_map++;
 435         }
 436         for (i = 0, j = 0; i < ioreq->v.niov; i++) {
 437             if (page[i] == NULL) {
 438                 page[i] = ioreq->page[j++];
 439             }
 440         }
 441     }
 442     if (ioreq->blkdev->feature_persistent && new_maps != 0 &&
 443         (!batch_maps || (ioreq->blkdev->persistent_gnt_count + new_maps <=
 444         ioreq->blkdev->max_grants))) {
 445         /*
 446          * If we are using persistent grants and batch mappings only
 447          * add the new maps to the list of persistent grants if the whole
 448          * area can be persistently mapped.
 449          */
 450         if (batch_maps) {
 451             region = g_malloc0(sizeof(*region));
 452             region->addr = ioreq->pages;
 453             region->num = new_maps;
 454             ioreq->blkdev->persistent_regions = g_slist_append(
 455                                             ioreq->blkdev->persistent_regions,
 456                                             region);
 457         }
 458         while ((ioreq->blkdev->persistent_gnt_count < ioreq->blkdev->max_grants)
 459               && new_maps) {
 460             /* Go through the list of newly mapped grants and add as many
 461              * as possible to the list of persistently mapped grants.
 462              *
 463              * Since we start at the end of ioreq->page(s), we only need
 464              * to decrease new_maps to prevent this granted pages from
 465              * being unmapped in ioreq_unmap.
 466              */
 467             grant = g_malloc0(sizeof(*grant));
 468             new_maps--;
 469             if (batch_maps) {
 470                 grant->page = ioreq->pages + (new_maps) * XC_PAGE_SIZE;
 471             } else {
 472                 grant->page = ioreq->page[new_maps];
 473             }
 474             grant->blkdev = ioreq->blkdev;
 475             xen_be_printf(&ioreq->blkdev->xendev, 3,
 476                           "adding grant %" PRIu32 " page: %p\n",
 477                           refs[new_maps], grant->page);
 478             g_tree_insert(ioreq->blkdev->persistent_gnts,
 479                           GUINT_TO_POINTER(refs[new_maps]),
 480                           grant);
 481             ioreq->blkdev->persistent_gnt_count++;
 482         }
 483         assert(!batch_maps || new_maps == 0);
 484     }
 485     for (i = 0; i < ioreq->v.niov; i++) {
 486         ioreq->v.iov[i].iov_base += (uintptr_t)page[i];
 487     }
 488     ioreq->mapped = 1;
 489     ioreq->num_unmap = new_maps;
 490     return 0;
 491 }
 492
 493 static int ioreq_runio_qemu_aio(struct ioreq *ioreq);
 494
 495 static void qemu_aio_complete(void *opaque, int ret)
 496 {
 497     struct ioreq *ioreq = opaque;
 498
 499     if (ret != 0) {
 500         xen_be_printf(&ioreq->blkdev->xendev, 0, "%s I/O error\n",
 501                       ioreq->req.operation == BLKIF_OP_READ ? "read" : "write");
 502         ioreq->aio_errors++;
 503     }
 504
 505     ioreq->aio_inflight--;
 506     if (ioreq->presync) {
 507         ioreq->presync = 0;
 508         ioreq_runio_qemu_aio(ioreq);
 509         return;
 510     }
 511     if (ioreq->aio_inflight > 0) {
 512         return;
 513     }
 514
 515     ioreq->status = ioreq->aio_errors ? BLKIF_RSP_ERROR : BLKIF_RSP_OKAY;
 516     ioreq_unmap(ioreq);
 517     ioreq_finish(ioreq);
 518     switch (ioreq->req.operation) {
 519     case BLKIF_OP_WRITE:
 520     case BLKIF_OP_FLUSH_DISKCACHE:
 521         if (!ioreq->req.nr_segments) {
 522             break;
 523         }
 524     case BLKIF_OP_READ:
 525         if (ioreq->status == BLKIF_RSP_OKAY) {
 526             block_acct_done(blk_get_stats(ioreq->blkdev->blk), &ioreq->acct);
 527         } else {
 528             block_acct_failed(blk_get_stats(ioreq->blkdev->blk), &ioreq->acct);
 529         }
 530         break;
 531     case BLKIF_OP_DISCARD:
 532     default:
 533         break;
 534     }
 535     qemu_bh_schedule(ioreq->blkdev->bh);
 536 }
 537
 538 static int ioreq_runio_qemu_aio(struct ioreq *ioreq)
 539 {
 540     struct XenBlkDev *blkdev = ioreq->blkdev;
 541
 542     if (ioreq->req.nr_segments && ioreq_map(ioreq) == -1) {
 543         goto err_no_map;
 544     }
 545
 546     ioreq->aio_inflight++;
 547     if (ioreq->presync) {
 548         blk_aio_flush(ioreq->blkdev->blk, qemu_aio_complete, ioreq);
 549         return 0;
 550     }
 551
 552     switch (ioreq->req.operation) {
 553     case BLKIF_OP_READ:
 554         block_acct_start(blk_get_stats(blkdev->blk), &ioreq->acct,
 555                          ioreq->v.size, BLOCK_ACCT_READ);
 556         ioreq->aio_inflight++;
 557         blk_aio_readv(blkdev->blk, ioreq->start / BLOCK_SIZE,
 558                       &ioreq->v, ioreq->v.size / BLOCK_SIZE,
 559                       qemu_aio_complete, ioreq);
 560         break;
 561     case BLKIF_OP_WRITE:
 562     case BLKIF_OP_FLUSH_DISKCACHE:
 563         if (!ioreq->req.nr_segments) {
 564             break;
 565         }
 566
 567         block_acct_start(blk_get_stats(blkdev->blk), &ioreq->acct,
 568                          ioreq->v.size,
 569                          ioreq->req.operation == BLKIF_OP_WRITE ?
 570                          BLOCK_ACCT_WRITE : BLOCK_ACCT_FLUSH);
 571         ioreq->aio_inflight++;
 572         blk_aio_writev(blkdev->blk, ioreq->start / BLOCK_SIZE,
 573                        &ioreq->v, ioreq->v.size / BLOCK_SIZE,
 574                        qemu_aio_complete, ioreq);
 575         break;
 576     case BLKIF_OP_DISCARD:
 577     {
 578         struct blkif_request_discard *discard_req = (void *)&ioreq->req;
 579         ioreq->aio_inflight++;
 580         blk_aio_discard(blkdev->blk,
 581                         discard_req->sector_number, discard_req->nr_sectors,
 582                         qemu_aio_complete, ioreq);
 583         break;
 584     }
 585     default:
 586         /* unknown operation (shouldn't happen -- parse catches this) */
 587         goto err;
 588     }
 589
 590     qemu_aio_complete(ioreq, 0);
 591
 592     return 0;
 593
 594 err:
 595     ioreq_unmap(ioreq);
 596 err_no_map:
 597     ioreq_finish(ioreq);
 598     ioreq->status = BLKIF_RSP_ERROR;
 599     return -1;
 600 }
 601
 602 static int blk_send_response_one(struct ioreq *ioreq)
 603 {
 604     struct XenBlkDev  *blkdev = ioreq->blkdev;
 605     int               send_notify   = 0;
 606     int               have_requests = 0;
 607     blkif_response_t  resp;
 608     void              *dst;
 609
 610     resp.id        = ioreq->req.id;
 611     resp.operation = ioreq->req.operation;
 612     resp.status    = ioreq->status;
 613
 614     /* Place on the response ring for the relevant domain. */
 615     switch (blkdev->protocol) {
 616     case BLKIF_PROTOCOL_NATIVE:
 617         dst = RING_GET_RESPONSE(&blkdev->rings.native, blkdev->rings.native.rsp_prod_pvt);
 618         break;
 619     case BLKIF_PROTOCOL_X86_32:
 620         dst = RING_GET_RESPONSE(&blkdev->rings.x86_32_part,
 621                                 blkdev->rings.x86_32_part.rsp_prod_pvt);
 622         break;
 623     case BLKIF_PROTOCOL_X86_64:
 624         dst = RING_GET_RESPONSE(&blkdev->rings.x86_64_part,
 625                                 blkdev->rings.x86_64_part.rsp_prod_pvt);
 626         break;
 627     default:
 628         dst = NULL;
 629         return 0;
 630     }
 631     memcpy(dst, &resp, sizeof(resp));
 632     blkdev->rings.common.rsp_prod_pvt++;
 633
 634     RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blkdev->rings.common, send_notify);
 635     if (blkdev->rings.common.rsp_prod_pvt == blkdev->rings.common.req_cons) {
 636         /*
 637          * Tail check for pending requests. Allows frontend to avoid
 638          * notifications if requests are already in flight (lower
 639          * overheads and promotes batching).
 640          */
 641         RING_FINAL_CHECK_FOR_REQUESTS(&blkdev->rings.common, have_requests);
 642     } else if (RING_HAS_UNCONSUMED_REQUESTS(&blkdev->rings.common)) {
 643         have_requests = 1;
 644     }
 645
 646     if (have_requests) {
 647         blkdev->more_work++;
 648     }
 649     return send_notify;
 650 }
 651
 652 /* walk finished list, send outstanding responses, free requests */
 653 static void blk_send_response_all(struct XenBlkDev *blkdev)
 654 {
 655     struct ioreq *ioreq;
 656     int send_notify = 0;
 657
 658     while (!QLIST_EMPTY(&blkdev->finished)) {
 659         ioreq = QLIST_FIRST(&blkdev->finished);
 660         send_notify += blk_send_response_one(ioreq);
 661         ioreq_release(ioreq, true);
 662     }
 663     if (send_notify) {
 664         xen_be_send_notify(&blkdev->xendev);
 665     }
 666 }
 667
 668 static int blk_get_request(struct XenBlkDev *blkdev, struct ioreq *ioreq, RING_IDX rc)
 669 {
 670     switch (blkdev->protocol) {
 671     case BLKIF_PROTOCOL_NATIVE:
 672         memcpy(&ioreq->req, RING_GET_REQUEST(&blkdev->rings.native, rc),
 673                sizeof(ioreq->req));
 674         break;
 675     case BLKIF_PROTOCOL_X86_32:
 676         blkif_get_x86_32_req(&ioreq->req,
 677                              RING_GET_REQUEST(&blkdev->rings.x86_32_part, rc));
 678         break;
 679     case BLKIF_PROTOCOL_X86_64:
 680         blkif_get_x86_64_req(&ioreq->req,
 681                              RING_GET_REQUEST(&blkdev->rings.x86_64_part, rc));
 682         break;
 683     }
 684     return 0;
 685 }
 686
 687 static void blk_handle_requests(struct XenBlkDev *blkdev)
 688 {
 689     RING_IDX rc, rp;
 690     struct ioreq *ioreq;
 691
 692     blkdev->more_work = 0;
 693
 694     rc = blkdev->rings.common.req_cons;
 695     rp = blkdev->rings.common.sring->req_prod;
 696     xen_rmb(); /* Ensure we see queued requests up to 'rp'. */
 697
 698     blk_send_response_all(blkdev);
 699     while (rc != rp) {
 700         /* pull request from ring */
 701         if (RING_REQUEST_CONS_OVERFLOW(&blkdev->rings.common, rc)) {
 702             break;
 703         }
 704         ioreq = ioreq_start(blkdev);
 705         if (ioreq == NULL) {
 706             blkdev->more_work++;
 707             break;
 708         }
 709         blk_get_request(blkdev, ioreq, rc);
 710         blkdev->rings.common.req_cons = ++rc;
 711
 712         /* parse them */
 713         if (ioreq_parse(ioreq) != 0) {
 714
 715             switch (ioreq->req.operation) {
 716             case BLKIF_OP_READ:
 717                 block_acct_invalid(blk_get_stats(blkdev->blk),
 718                                    BLOCK_ACCT_READ);
 719                 break;
 720             case BLKIF_OP_WRITE:
 721                 block_acct_invalid(blk_get_stats(blkdev->blk),
 722                                    BLOCK_ACCT_WRITE);
 723                 break;
 724             case BLKIF_OP_FLUSH_DISKCACHE:
 725                 block_acct_invalid(blk_get_stats(blkdev->blk),
 726                                    BLOCK_ACCT_FLUSH);
 727             default:
 728                 break;
 729             };
 730
 731             if (blk_send_response_one(ioreq)) {
 732                 xen_be_send_notify(&blkdev->xendev);
 733             }
 734             ioreq_release(ioreq, false);
 735             continue;
 736         }
 737
 738         ioreq_runio_qemu_aio(ioreq);
 739     }
 740
 741     if (blkdev->more_work && blkdev->requests_inflight < max_requests) {
 742         qemu_bh_schedule(blkdev->bh);
 743     }
 744 }
 745
 746 /* ------------------------------------------------------------- */
 747
 748 static void blk_bh(void *opaque)
 749 {
 750     struct XenBlkDev *blkdev = opaque;
 751     blk_handle_requests(blkdev);
 752 }
 753
 754 /*
 755  * We need to account for the grant allocations requiring contiguous
 756  * chunks; the worst case number would be
 757  *     max_req * max_seg + (max_req - 1) * (max_seg - 1) + 1,
 758  * but in order to keep things simple just use
 759  *     2 * max_req * max_seg.
 760  */
 761 #define MAX_GRANTS(max_req, max_seg) (2 * (max_req) * (max_seg))
 762
 763 static void blk_alloc(struct XenDevice *xendev)
 764 {
 765     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
 766
 767     QLIST_INIT(&blkdev->inflight);
 768     QLIST_INIT(&blkdev->finished);
 769     QLIST_INIT(&blkdev->freelist);
 770     blkdev->bh = qemu_bh_new(blk_bh, blkdev);
 771     if (xen_mode != XEN_EMULATE) {
 772         batch_maps = 1;
 773     }
 774     if (xengnttab_set_max_grants(xendev->gnttabdev,
 775             MAX_GRANTS(max_requests, BLKIF_MAX_SEGMENTS_PER_REQUEST)) < 0) {
 776         xen_be_printf(xendev, 0, "xengnttab_set_max_grants failed: %s\n",
 777                       strerror(errno));
 778     }
 779 }
 780
 781 static void blk_parse_discard(struct XenBlkDev *blkdev)
 782 {
 783     int enable;
 784
 785     blkdev->feature_discard = true;
 786
 787     if (xenstore_read_be_int(&blkdev->xendev, "discard-enable", &enable) == 0) {
 788         blkdev->feature_discard = !!enable;
 789     }
 790
 791     if (blkdev->feature_discard) {
 792         xenstore_write_be_int(&blkdev->xendev, "feature-discard", 1);
 793     }
 794 }
 795
 796 static int blk_init(struct XenDevice *xendev)
 797 {
 798     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
 799     int info = 0;
 800     char *directiosafe = NULL;
 801
 802     /* read xenstore entries */
 803     if (blkdev->params == NULL) {
 804         char *h = NULL;
 805         blkdev->params = xenstore_read_be_str(&blkdev->xendev, "params");
 806         if (blkdev->params != NULL) {
 807             h = strchr(blkdev->params, ':');
 808         }
 809         if (h != NULL) {
 810             blkdev->fileproto = blkdev->params;
 811             blkdev->filename  = h+1;
 812             *h = 0;
 813         } else {
 814             blkdev->fileproto = "<unset>";
 815             blkdev->filename  = blkdev->params;
 816         }
 817     }
 818     if (!strcmp("aio", blkdev->fileproto)) {
 819         blkdev->fileproto = "raw";
 820     }
 821     if (!strcmp("vhd", blkdev->fileproto)) {
 822         blkdev->fileproto = "vpc";
 823     }
 824     if (blkdev->mode == NULL) {
 825         blkdev->mode = xenstore_read_be_str(&blkdev->xendev, "mode");
 826     }
 827     if (blkdev->type == NULL) {
 828         blkdev->type = xenstore_read_be_str(&blkdev->xendev, "type");
 829     }
 830     if (blkdev->dev == NULL) {
 831         blkdev->dev = xenstore_read_be_str(&blkdev->xendev, "dev");
 832     }
 833     if (blkdev->devtype == NULL) {
 834         blkdev->devtype = xenstore_read_be_str(&blkdev->xendev, "device-type");
 835     }
 836     directiosafe = xenstore_read_be_str(&blkdev->xendev, "direct-io-safe");
 837     blkdev->directiosafe = (directiosafe && atoi(directiosafe));
 838
 839     /* do we have all we need? */
 840     if (blkdev->params == NULL ||
 841         blkdev->mode == NULL   ||
 842         blkdev->type == NULL   ||
 843         blkdev->dev == NULL) {
 844         goto out_error;
 845     }
 846
 847     /* read-only ? */
 848     if (strcmp(blkdev->mode, "w")) {
 849         info  |= VDISK_READONLY;
 850     }
 851
 852     /* cdrom ? */
 853     if (blkdev->devtype && !strcmp(blkdev->devtype, "cdrom")) {
 854         info  |= VDISK_CDROM;
 855     }
 856
 857     blkdev->file_blk  = BLOCK_SIZE;
 858
 859     /* fill info
 860      * blk_connect supplies sector-size and sectors
 861      */
 862     xenstore_write_be_int(&blkdev->xendev, "feature-flush-cache", 1);
 863     xenstore_write_be_int(&blkdev->xendev, "feature-persistent", 1);
 864     xenstore_write_be_int(&blkdev->xendev, "info", info);
 865
 866     blk_parse_discard(blkdev);
 867
 868     g_free(directiosafe);
 869     return 0;
 870
 871 out_error:
 872     g_free(blkdev->params);
 873     blkdev->params = NULL;
 874     g_free(blkdev->mode);
 875     blkdev->mode = NULL;
 876     g_free(blkdev->type);
 877     blkdev->type = NULL;
 878     g_free(blkdev->dev);
 879     blkdev->dev = NULL;
 880     g_free(blkdev->devtype);
 881     blkdev->devtype = NULL;
 882     g_free(directiosafe);
 883     blkdev->directiosafe = false;
 884     return -1;
 885 }
 886
 887 static int blk_connect(struct XenDevice *xendev)
 888 {
 889     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
 890     int pers, index, qflags;
 891     bool readonly = true;
 892     bool writethrough = true;
 893
 894     /* read-only ? */
 895     if (blkdev->directiosafe) {
 896         qflags = BDRV_O_NOCACHE | BDRV_O_NATIVE_AIO;
 897     } else {
 898         qflags = 0;
 899         writethrough = false;
 900     }
 901     if (strcmp(blkdev->mode, "w") == 0) {
 902         qflags |= BDRV_O_RDWR;
 903         readonly = false;
 904     }
 905     if (blkdev->feature_discard) {
 906         qflags |= BDRV_O_UNMAP;
 907     }
 908
 909     /* init qemu block driver */
 910     index = (blkdev->xendev.dev - 202 * 256) / 16;
 911     blkdev->dinfo = drive_get(IF_XEN, 0, index);
 912     if (!blkdev->dinfo) {
 913         Error *local_err = NULL;
 914         QDict *options = NULL;
 915
 916         if (strcmp(blkdev->fileproto, "<unset>")) {
 917             options = qdict_new();
 918             qdict_put(options, "driver", qstring_from_str(blkdev->fileproto));
 919         }
 920
 921         /* setup via xenbus -> create new block driver instance */
 922         xen_be_printf(&blkdev->xendev, 2, "create new bdrv (xenbus setup)\n");
 923         blkdev->blk = blk_new_open(blkdev->filename, NULL, options,
 924                                    qflags, &local_err);
 925         if (!blkdev->blk) {
 926             xen_be_printf(&blkdev->xendev, 0, "error: %s\n",
 927                           error_get_pretty(local_err));
 928             error_free(local_err);
 929             return -1;
 930         }
 931         blk_set_enable_write_cache(blkdev->blk, !writethrough);
 932     } else {
 933         /* setup via qemu cmdline -> already setup for us */
 934         xen_be_printf(&blkdev->xendev, 2, "get configured bdrv (cmdline setup)\n");
 935         blkdev->blk = blk_by_legacy_dinfo(blkdev->dinfo);
 936         if (blk_is_read_only(blkdev->blk) && !readonly) {
 937             xen_be_printf(&blkdev->xendev, 0, "Unexpected read-only drive");
 938             blkdev->blk = NULL;
 939             return -1;
 940         }
 941         /* blkdev->blk is not create by us, we get a reference
 942          * so we can blk_unref() unconditionally */
 943         blk_ref(blkdev->blk);
 944     }
 945     blk_attach_dev_nofail(blkdev->blk, blkdev);
 946     blkdev->file_size = blk_getlength(blkdev->blk);
 947     if (blkdev->file_size < 0) {
 948         BlockDriverState *bs = blk_bs(blkdev->blk);
 949         const char *drv_name = bs ? bdrv_get_format_name(bs) : NULL;
 950         xen_be_printf(&blkdev->xendev, 1, "blk_getlength: %d (%s) | drv %s\n",
 951                       (int)blkdev->file_size, strerror(-blkdev->file_size),
 952                       drv_name ?: "-");
 953         blkdev->file_size = 0;
 954     }
 955
 956     xen_be_printf(xendev, 1, "type \"%s\", fileproto \"%s\", filename \"%s\","
 957                   " size %" PRId64 " (%" PRId64 " MB)\n",
 958                   blkdev->type, blkdev->fileproto, blkdev->filename,
 959                   blkdev->file_size, blkdev->file_size >> 20);
 960
 961     /* Fill in number of sector size and number of sectors */
 962     xenstore_write_be_int(&blkdev->xendev, "sector-size", blkdev->file_blk);
 963     xenstore_write_be_int64(&blkdev->xendev, "sectors",
 964                             blkdev->file_size / blkdev->file_blk);
 965
 966     if (xenstore_read_fe_int(&blkdev->xendev, "ring-ref", &blkdev->ring_ref) == -1) {
 967         return -1;
 968     }
 969     if (xenstore_read_fe_int(&blkdev->xendev, "event-channel",
 970                              &blkdev->xendev.remote_port) == -1) {
 971         return -1;
 972     }
 973     if (xenstore_read_fe_int(&blkdev->xendev, "feature-persistent", &pers)) {
 974         blkdev->feature_persistent = FALSE;
 975     } else {
 976         blkdev->feature_persistent = !!pers;
 977     }
 978
 979     blkdev->protocol = BLKIF_PROTOCOL_NATIVE;
 980     if (blkdev->xendev.protocol) {
 981         if (strcmp(blkdev->xendev.protocol, XEN_IO_PROTO_ABI_X86_32) == 0) {
 982             blkdev->protocol = BLKIF_PROTOCOL_X86_32;
 983         }
 984         if (strcmp(blkdev->xendev.protocol, XEN_IO_PROTO_ABI_X86_64) == 0) {
 985             blkdev->protocol = BLKIF_PROTOCOL_X86_64;
 986         }
 987     }
 988
 989     blkdev->sring = xengnttab_map_grant_ref(blkdev->xendev.gnttabdev,
 990                                             blkdev->xendev.dom,
 991                                             blkdev->ring_ref,
 992                                             PROT_READ | PROT_WRITE);
 993     if (!blkdev->sring) {
 994         return -1;
 995     }
 996     blkdev->cnt_map++;
 997
 998     switch (blkdev->protocol) {
 999     case BLKIF_PROTOCOL_NATIVE:
1000     {
1001         blkif_sring_t *sring_native = blkdev->sring;
1002         BACK_RING_INIT(&blkdev->rings.native, sring_native, XC_PAGE_SIZE);
1003         break;
1004     }
1005     case BLKIF_PROTOCOL_X86_32:
1006     {
1007         blkif_x86_32_sring_t *sring_x86_32 = blkdev->sring;
1008
1009         BACK_RING_INIT(&blkdev->rings.x86_32_part, sring_x86_32, XC_PAGE_SIZE);
1010         break;
1011     }
1012     case BLKIF_PROTOCOL_X86_64:
1013     {
1014         blkif_x86_64_sring_t *sring_x86_64 = blkdev->sring;
1015
1016         BACK_RING_INIT(&blkdev->rings.x86_64_part, sring_x86_64, XC_PAGE_SIZE);
1017         break;
1018     }
1019     }
1020
1021     if (blkdev->feature_persistent) {
1022         /* Init persistent grants */
1023         blkdev->max_grants = max_requests * BLKIF_MAX_SEGMENTS_PER_REQUEST;
1024         blkdev->persistent_gnts = g_tree_new_full((GCompareDataFunc)int_cmp,
1025                                              NULL, NULL,
1026                                              batch_maps ?
1027                                              (GDestroyNotify)g_free :
1028                                              (GDestroyNotify)destroy_grant);
1029         blkdev->persistent_regions = NULL;
1030         blkdev->persistent_gnt_count = 0;
1031     }
1032
1033     xen_be_bind_evtchn(&blkdev->xendev);
1034
1035     xen_be_printf(&blkdev->xendev, 1, "ok: proto %s, ring-ref %d, "
1036                   "remote port %d, local port %d\n",
1037                   blkdev->xendev.protocol, blkdev->ring_ref,
1038                   blkdev->xendev.remote_port, blkdev->xendev.local_port);
1039     return 0;
1040 }
1041
1042 static void blk_disconnect(struct XenDevice *xendev)
1043 {
1044     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
1045
1046     if (blkdev->blk) {
1047         blk_detach_dev(blkdev->blk, blkdev);
1048         blk_unref(blkdev->blk);
1049         blkdev->blk = NULL;
1050     }
1051     xen_be_unbind_evtchn(&blkdev->xendev);
1052
1053     if (blkdev->sring) {
1054         xengnttab_unmap(blkdev->xendev.gnttabdev, blkdev->sring, 1);
1055         blkdev->cnt_map--;
1056         blkdev->sring = NULL;
1057     }
1058
1059     /*
1060      * Unmap persistent grants before switching to the closed state
1061      * so the frontend can free them.
1062      *
1063      * In the !batch_maps case g_tree_destroy will take care of unmapping
1064      * the grant, but in the batch_maps case we need to iterate over every
1065      * region in persistent_regions and unmap it.
1066      */
1067     if (blkdev->feature_persistent) {
1068         g_tree_destroy(blkdev->persistent_gnts);
1069         assert(batch_maps || blkdev->persistent_gnt_count == 0);
1070         if (batch_maps) {
1071             blkdev->persistent_gnt_count = 0;
1072             g_slist_foreach(blkdev->persistent_regions,
1073                             (GFunc)remove_persistent_region, blkdev);
1074             g_slist_free(blkdev->persistent_regions);
1075         }
1076         blkdev->feature_persistent = false;
1077     }
1078 }
1079
1080 static int blk_free(struct XenDevice *xendev)
1081 {
1082     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
1083     struct ioreq *ioreq;
1084
1085     if (blkdev->blk || blkdev->sring) {
1086         blk_disconnect(xendev);
1087     }
1088
1089     while (!QLIST_EMPTY(&blkdev->freelist)) {
1090         ioreq = QLIST_FIRST(&blkdev->freelist);
1091         QLIST_REMOVE(ioreq, list);
1092         qemu_iovec_destroy(&ioreq->v);
1093         g_free(ioreq);
1094     }
1095
1096     g_free(blkdev->params);
1097     g_free(blkdev->mode);
1098     g_free(blkdev->type);
1099     g_free(blkdev->dev);
1100     g_free(blkdev->devtype);
1101     qemu_bh_delete(blkdev->bh);
1102     return 0;
1103 }
1104
1105 static void blk_event(struct XenDevice *xendev)
1106 {
1107     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
1108
1109     qemu_bh_schedule(blkdev->bh);
1110 }
1111
1112 struct XenDevOps xen_blkdev_ops = {
1113     .size       = sizeof(struct XenBlkDev),
1114     .flags      = DEVOPS_FLAG_NEED_GNTDEV,
1115     .alloc      = blk_alloc,
1116     .init       = blk_init,
1117     .initialise    = blk_connect,
1118     .disconnect = blk_disconnect,
1119     .event      = blk_event,
1120     .free       = blk_free,
1121 };