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