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