hw/block/xen_disk.c

   1 /*
   2  *  xen paravirt block device backend
   3  *
   4  *  (c) Gerd Hoffmann <[email protected]>
   5  *
   6  *  This program is free software; you can redistribute it and/or modify
   7  *  it under the terms of the GNU General Public License as published by
   8  *  the Free Software Foundation; under version 2 of the License.
   9  *
  10  *  This program is distributed in the hope that it will be useful,
  11  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  *  GNU General Public License for more details.
  14  *
  15  *  You should have received a copy of the GNU General Public License along
  16  *  with this program; if not, see <http://www.gnu.org/licenses/>.
  17  *
  18  *  Contributions after 2012-01-13 are licensed under the terms of the
  19  *  GNU GPL, version 2 or (at your option) any later version.
  20  */
  21
  22 #include "qemu/osdep.h"
  23 #include <sys/ioctl.h>
  24 #include <sys/uio.h>
  25
  26 #include "hw/hw.h"
  27 #include "hw/xen/xen_backend.h"
  28 #include "xen_blkif.h"
  29 #include "sysemu/blockdev.h"
  30 #include "sysemu/block-backend.h"
  31 #include "qapi/error.h"
  32 #include "qapi/qmp/qdict.h"
  33 #include "qapi/qmp/qstring.h"
  34
  35 /* ------------------------------------------------------------- */
  36
  37 static int batch_maps   = 0;
  38
  39 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_handle *gnt = grant->blkdev->xendev.gnttabdev;
 165
 166     if (xengnttab_unmap(gnt, grant->page, 1) != 0) {
 167         xen_be_printf(&grant->blkdev->xendev, 0,
 168                       "xengnttab_unmap 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_handle *gnt = blkdev->xendev.gnttabdev;
 182
 183     if (xengnttab_unmap(gnt, region->addr, region->num) != 0) {
 184         xen_be_printf(&blkdev->xendev, 0,
 185                       "xengnttab_unmap 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_handle *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 (xengnttab_unmap(gnt, ioreq->pages, ioreq->num_unmap) != 0) {
 331             xen_be_printf(&ioreq->blkdev->xendev, 0,
 332                           "xengnttab_unmap failed: %s\n",
 333                           strerror(errno));
 334         }
 335         ioreq->blkdev->cnt_map -= ioreq->num_unmap;
 336         ioreq->pages = NULL;
 337     } else {
 338         for (i = 0; i < ioreq->num_unmap; i++) {
 339             if (!ioreq->page[i]) {
 340                 continue;
 341             }
 342             if (xengnttab_unmap(gnt, ioreq->page[i], 1) != 0) {
 343                 xen_be_printf(&ioreq->blkdev->xendev, 0,
 344                               "xengnttab_unmap failed: %s\n",
 345                               strerror(errno));
 346             }
 347             ioreq->blkdev->cnt_map--;
 348             ioreq->page[i] = NULL;
 349         }
 350     }
 351     ioreq->mapped = 0;
 352 }
 353
 354 static int ioreq_map(struct ioreq *ioreq)
 355 {
 356     xengnttab_handle *gnt = ioreq->blkdev->xendev.gnttabdev;
 357     uint32_t domids[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 358     uint32_t refs[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 359     void *page[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 360     int i, j, new_maps = 0;
 361     PersistentGrant *grant;
 362     PersistentRegion *region;
 363     /* domids and refs variables will contain the information necessary
 364      * to map the grants that are needed to fulfill this request.
 365      *
 366      * After mapping the needed grants, the page array will contain the
 367      * memory address of each granted page in the order specified in ioreq
 368      * (disregarding if it's a persistent grant or not).
 369      */
 370
 371     if (ioreq->v.niov == 0 || ioreq->mapped == 1) {
 372         return 0;
 373     }
 374     if (ioreq->blkdev->feature_persistent) {
 375         for (i = 0; i < ioreq->v.niov; i++) {
 376             grant = g_tree_lookup(ioreq->blkdev->persistent_gnts,
 377                                     GUINT_TO_POINTER(ioreq->refs[i]));
 378
 379             if (grant != NULL) {
 380                 page[i] = grant->page;
 381                 xen_be_printf(&ioreq->blkdev->xendev, 3,
 382                               "using persistent-grant %" PRIu32 "\n",
 383                               ioreq->refs[i]);
 384             } else {
 385                     /* Add the grant to the list of grants that
 386                      * should be mapped
 387                      */
 388                     domids[new_maps] = ioreq->domids[i];
 389                     refs[new_maps] = ioreq->refs[i];
 390                     page[i] = NULL;
 391                     new_maps++;
 392             }
 393         }
 394         /* Set the protection to RW, since grants may be reused later
 395          * with a different protection than the one needed for this request
 396          */
 397         ioreq->prot = PROT_WRITE | PROT_READ;
 398     } else {
 399         /* All grants in the request should be mapped */
 400         memcpy(refs, ioreq->refs, sizeof(refs));
 401         memcpy(domids, ioreq->domids, sizeof(domids));
 402         memset(page, 0, sizeof(page));
 403         new_maps = ioreq->v.niov;
 404     }
 405
 406     if (batch_maps && new_maps) {
 407         ioreq->pages = xengnttab_map_grant_refs
 408             (gnt, new_maps, domids, refs, ioreq->prot);
 409         if (ioreq->pages == NULL) {
 410             xen_be_printf(&ioreq->blkdev->xendev, 0,
 411                           "can't map %d grant refs (%s, %d maps)\n",
 412                           new_maps, strerror(errno), ioreq->blkdev->cnt_map);
 413             return -1;
 414         }
 415         for (i = 0, j = 0; i < ioreq->v.niov; i++) {
 416             if (page[i] == NULL) {
 417                 page[i] = ioreq->pages + (j++) * XC_PAGE_SIZE;
 418             }
 419         }
 420         ioreq->blkdev->cnt_map += new_maps;
 421     } else if (new_maps)  {
 422         for (i = 0; i < new_maps; i++) {
 423             ioreq->page[i] = xengnttab_map_grant_ref
 424                 (gnt, domids[i], refs[i], ioreq->prot);
 425             if (ioreq->page[i] == NULL) {
 426                 xen_be_printf(&ioreq->blkdev->xendev, 0,
 427                               "can't map grant ref %d (%s, %d maps)\n",
 428                               refs[i], strerror(errno), ioreq->blkdev->cnt_map);
 429                 ioreq->mapped = 1;
 430                 ioreq_unmap(ioreq);
 431                 return -1;
 432             }
 433             ioreq->blkdev->cnt_map++;
 434         }
 435         for (i = 0, j = 0; i < ioreq->v.niov; i++) {
 436             if (page[i] == NULL) {
 437                 page[i] = ioreq->page[j++];
 438             }
 439         }
 440     }
 441     if (ioreq->blkdev->feature_persistent && new_maps != 0 &&
 442         (!batch_maps || (ioreq->blkdev->persistent_gnt_count + new_maps <=
 443         ioreq->blkdev->max_grants))) {
 444         /*
 445          * If we are using persistent grants and batch mappings only
 446          * add the new maps to the list of persistent grants if the whole
 447          * area can be persistently mapped.
 448          */
 449         if (batch_maps) {
 450             region = g_malloc0(sizeof(*region));
 451             region->addr = ioreq->pages;
 452             region->num = new_maps;
 453             ioreq->blkdev->persistent_regions = g_slist_append(
 454                                             ioreq->blkdev->persistent_regions,
 455                                             region);
 456         }
 457         while ((ioreq->blkdev->persistent_gnt_count < ioreq->blkdev->max_grants)
 458               && new_maps) {
 459             /* Go through the list of newly mapped grants and add as many
 460              * as possible to the list of persistently mapped grants.
 461              *
 462              * Since we start at the end of ioreq->page(s), we only need
 463              * to decrease new_maps to prevent this granted pages from
 464              * being unmapped in ioreq_unmap.
 465              */
 466             grant = g_malloc0(sizeof(*grant));
 467             new_maps--;
 468             if (batch_maps) {
 469                 grant->page = ioreq->pages + (new_maps) * XC_PAGE_SIZE;
 470             } else {
 471                 grant->page = ioreq->page[new_maps];
 472             }
 473             grant->blkdev = ioreq->blkdev;
 474             xen_be_printf(&ioreq->blkdev->xendev, 3,
 475                           "adding grant %" PRIu32 " page: %p\n",
 476                           refs[new_maps], grant->page);
 477             g_tree_insert(ioreq->blkdev->persistent_gnts,
 478                           GUINT_TO_POINTER(refs[new_maps]),
 479                           grant);
 480             ioreq->blkdev->persistent_gnt_count++;
 481         }
 482         assert(!batch_maps || new_maps == 0);
 483     }
 484     for (i = 0; i < ioreq->v.niov; i++) {
 485         ioreq->v.iov[i].iov_base += (uintptr_t)page[i];
 486     }
 487     ioreq->mapped = 1;
 488     ioreq->num_unmap = new_maps;
 489     return 0;
 490 }
 491
 492 static int ioreq_runio_qemu_aio(struct ioreq *ioreq);
 493
 494 static void qemu_aio_complete(void *opaque, int ret)
 495 {
 496     struct ioreq *ioreq = opaque;
 497
 498     if (ret != 0) {
 499         xen_be_printf(&ioreq->blkdev->xendev, 0, "%s I/O error\n",
 500                       ioreq->req.operation == BLKIF_OP_READ ? "read" : "write");
 501         ioreq->aio_errors++;
 502     }
 503
 504     ioreq->aio_inflight--;
 505     if (ioreq->presync) {
 506         ioreq->presync = 0;
 507         ioreq_runio_qemu_aio(ioreq);
 508         return;
 509     }
 510     if (ioreq->aio_inflight > 0) {
 511         return;
 512     }
 513
 514     ioreq->status = ioreq->aio_errors ? BLKIF_RSP_ERROR : BLKIF_RSP_OKAY;
 515     ioreq_unmap(ioreq);
 516     ioreq_finish(ioreq);
 517     switch (ioreq->req.operation) {
 518     case BLKIF_OP_WRITE:
 519     case BLKIF_OP_FLUSH_DISKCACHE:
 520         if (!ioreq->req.nr_segments) {
 521             break;
 522         }
 523     case BLKIF_OP_READ:
 524         if (ioreq->status == BLKIF_RSP_OKAY) {
 525             block_acct_done(blk_get_stats(ioreq->blkdev->blk), &ioreq->acct);
 526         } else {
 527             block_acct_failed(blk_get_stats(ioreq->blkdev->blk), &ioreq->acct);
 528         }
 529         break;
 530     case BLKIF_OP_DISCARD:
 531     default:
 532         break;
 533     }
 534     qemu_bh_schedule(ioreq->blkdev->bh);
 535 }
 536
 537 static int ioreq_runio_qemu_aio(struct ioreq *ioreq)
 538 {
 539     struct XenBlkDev *blkdev = ioreq->blkdev;
 540
 541     if (ioreq->req.nr_segments && ioreq_map(ioreq) == -1) {
 542         goto err_no_map;
 543     }
 544
 545     ioreq->aio_inflight++;
 546     if (ioreq->presync) {
 547         blk_aio_flush(ioreq->blkdev->blk, qemu_aio_complete, ioreq);
 548         return 0;
 549     }
 550
 551     switch (ioreq->req.operation) {
 552     case BLKIF_OP_READ:
 553         block_acct_start(blk_get_stats(blkdev->blk), &ioreq->acct,
 554                          ioreq->v.size, BLOCK_ACCT_READ);
 555         ioreq->aio_inflight++;
 556         blk_aio_preadv(blkdev->blk, ioreq->start, &ioreq->v, 0,
 557                        qemu_aio_complete, ioreq);
 558         break;
 559     case BLKIF_OP_WRITE:
 560     case BLKIF_OP_FLUSH_DISKCACHE:
 561         if (!ioreq->req.nr_segments) {
 562             break;
 563         }
 564
 565         block_acct_start(blk_get_stats(blkdev->blk), &ioreq->acct,
 566                          ioreq->v.size,
 567                          ioreq->req.operation == BLKIF_OP_WRITE ?
 568                          BLOCK_ACCT_WRITE : BLOCK_ACCT_FLUSH);
 569         ioreq->aio_inflight++;
 570         blk_aio_pwritev(blkdev->blk, ioreq->start, &ioreq->v, 0,
 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     /* Prevent the compiler from accessing the on-ring fields instead. */
 682     barrier();
 683     return 0;
 684 }
 685
 686 static void blk_handle_requests(struct XenBlkDev *blkdev)
 687 {
 688     RING_IDX rc, rp;
 689     struct ioreq *ioreq;
 690
 691     blkdev->more_work = 0;
 692
 693     rc = blkdev->rings.common.req_cons;
 694     rp = blkdev->rings.common.sring->req_prod;
 695     xen_rmb(); /* Ensure we see queued requests up to 'rp'. */
 696
 697     blk_send_response_all(blkdev);
 698     while (rc != rp) {
 699         /* pull request from ring */
 700         if (RING_REQUEST_CONS_OVERFLOW(&blkdev->rings.common, rc)) {
 701             break;
 702         }
 703         ioreq = ioreq_start(blkdev);
 704         if (ioreq == NULL) {
 705             blkdev->more_work++;
 706             break;
 707         }
 708         blk_get_request(blkdev, ioreq, rc);
 709         blkdev->rings.common.req_cons = ++rc;
 710
 711         /* parse them */
 712         if (ioreq_parse(ioreq) != 0) {
 713
 714             switch (ioreq->req.operation) {
 715             case BLKIF_OP_READ:
 716                 block_acct_invalid(blk_get_stats(blkdev->blk),
 717                                    BLOCK_ACCT_READ);
 718                 break;
 719             case BLKIF_OP_WRITE:
 720                 block_acct_invalid(blk_get_stats(blkdev->blk),
 721                                    BLOCK_ACCT_WRITE);
 722                 break;
 723             case BLKIF_OP_FLUSH_DISKCACHE:
 724                 block_acct_invalid(blk_get_stats(blkdev->blk),
 725                                    BLOCK_ACCT_FLUSH);
 726             default:
 727                 break;
 728             };
 729
 730             if (blk_send_response_one(ioreq)) {
 731                 xen_be_send_notify(&blkdev->xendev);
 732             }
 733             ioreq_release(ioreq, false);
 734             continue;
 735         }
 736
 737         ioreq_runio_qemu_aio(ioreq);
 738     }
 739
 740     if (blkdev->more_work && blkdev->requests_inflight < max_requests) {
 741         qemu_bh_schedule(blkdev->bh);
 742     }
 743 }
 744
 745 /* ------------------------------------------------------------- */
 746
 747 static void blk_bh(void *opaque)
 748 {
 749     struct XenBlkDev *blkdev = opaque;
 750     blk_handle_requests(blkdev);
 751 }
 752
 753 /*
 754  * We need to account for the grant allocations requiring contiguous
 755  * chunks; the worst case number would be
 756  *     max_req * max_seg + (max_req - 1) * (max_seg - 1) + 1,
 757  * but in order to keep things simple just use
 758  *     2 * max_req * max_seg.
 759  */
 760 #define MAX_GRANTS(max_req, max_seg) (2 * (max_req) * (max_seg))
 761
 762 static void blk_alloc(struct XenDevice *xendev)
 763 {
 764     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
 765
 766     QLIST_INIT(&blkdev->inflight);
 767     QLIST_INIT(&blkdev->finished);
 768     QLIST_INIT(&blkdev->freelist);
 769     blkdev->bh = qemu_bh_new(blk_bh, blkdev);
 770     if (xen_mode != XEN_EMULATE) {
 771         batch_maps = 1;
 772     }
 773     if (xengnttab_set_max_grants(xendev->gnttabdev,
 774             MAX_GRANTS(max_requests, BLKIF_MAX_SEGMENTS_PER_REQUEST)) < 0) {
 775         xen_be_printf(xendev, 0, "xengnttab_set_max_grants failed: %s\n",
 776                       strerror(errno));
 777     }
 778 }
 779
 780 static void blk_parse_discard(struct XenBlkDev *blkdev)
 781 {
 782     int enable;
 783
 784     blkdev->feature_discard = true;
 785
 786     if (xenstore_read_be_int(&blkdev->xendev, "discard-enable", &enable) == 0) {
 787         blkdev->feature_discard = !!enable;
 788     }
 789
 790     if (blkdev->feature_discard) {
 791         xenstore_write_be_int(&blkdev->xendev, "feature-discard", 1);
 792     }
 793 }
 794
 795 static int blk_init(struct XenDevice *xendev)
 796 {
 797     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
 798     int info = 0;
 799     char *directiosafe = NULL;
 800
 801     /* read xenstore entries */
 802     if (blkdev->params == NULL) {
 803         char *h = NULL;
 804         blkdev->params = xenstore_read_be_str(&blkdev->xendev, "params");
 805         if (blkdev->params != NULL) {
 806             h = strchr(blkdev->params, ':');
 807         }
 808         if (h != NULL) {
 809             blkdev->fileproto = blkdev->params;
 810             blkdev->filename  = h+1;
 811             *h = 0;
 812         } else {
 813             blkdev->fileproto = "<unset>";
 814             blkdev->filename  = blkdev->params;
 815         }
 816     }
 817     if (!strcmp("aio", blkdev->fileproto)) {
 818         blkdev->fileproto = "raw";
 819     }
 820     if (!strcmp("vhd", blkdev->fileproto)) {
 821         blkdev->fileproto = "vpc";
 822     }
 823     if (blkdev->mode == NULL) {
 824         blkdev->mode = xenstore_read_be_str(&blkdev->xendev, "mode");
 825     }
 826     if (blkdev->type == NULL) {
 827         blkdev->type = xenstore_read_be_str(&blkdev->xendev, "type");
 828     }
 829     if (blkdev->dev == NULL) {
 830         blkdev->dev = xenstore_read_be_str(&blkdev->xendev, "dev");
 831     }
 832     if (blkdev->devtype == NULL) {
 833         blkdev->devtype = xenstore_read_be_str(&blkdev->xendev, "device-type");
 834     }
 835     directiosafe = xenstore_read_be_str(&blkdev->xendev, "direct-io-safe");
 836     blkdev->directiosafe = (directiosafe && atoi(directiosafe));
 837
 838     /* do we have all we need? */
 839     if (blkdev->params == NULL ||
 840         blkdev->mode == NULL   ||
 841         blkdev->type == NULL   ||
 842         blkdev->dev == NULL) {
 843         goto out_error;
 844     }
 845
 846     /* read-only ? */
 847     if (strcmp(blkdev->mode, "w")) {
 848         info  |= VDISK_READONLY;
 849     }
 850
 851     /* cdrom ? */
 852     if (blkdev->devtype && !strcmp(blkdev->devtype, "cdrom")) {
 853         info  |= VDISK_CDROM;
 854     }
 855
 856     blkdev->file_blk  = BLOCK_SIZE;
 857
 858     /* fill info
 859      * blk_connect supplies sector-size and sectors
 860      */
 861     xenstore_write_be_int(&blkdev->xendev, "feature-flush-cache", 1);
 862     xenstore_write_be_int(&blkdev->xendev, "feature-persistent", 1);
 863     xenstore_write_be_int(&blkdev->xendev, "info", info);
 864
 865     blk_parse_discard(blkdev);
 866
 867     g_free(directiosafe);
 868     return 0;
 869
 870 out_error:
 871     g_free(blkdev->params);
 872     blkdev->params = NULL;
 873     g_free(blkdev->mode);
 874     blkdev->mode = NULL;
 875     g_free(blkdev->type);
 876     blkdev->type = NULL;
 877     g_free(blkdev->dev);
 878     blkdev->dev = NULL;
 879     g_free(blkdev->devtype);
 880     blkdev->devtype = NULL;
 881     g_free(directiosafe);
 882     blkdev->directiosafe = false;
 883     return -1;
 884 }
 885
 886 static int blk_connect(struct XenDevice *xendev)
 887 {
 888     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
 889     int pers, index, qflags;
 890     bool readonly = true;
 891     bool writethrough = true;
 892
 893     /* read-only ? */
 894     if (blkdev->directiosafe) {
 895         qflags = BDRV_O_NOCACHE | BDRV_O_NATIVE_AIO;
 896     } else {
 897         qflags = 0;
 898         writethrough = false;
 899     }
 900     if (strcmp(blkdev->mode, "w") == 0) {
 901         qflags |= BDRV_O_RDWR;
 902         readonly = false;
 903     }
 904     if (blkdev->feature_discard) {
 905         qflags |= BDRV_O_UNMAP;
 906     }
 907
 908     /* init qemu block driver */
 909     index = (blkdev->xendev.dev - 202 * 256) / 16;
 910     blkdev->dinfo = drive_get(IF_XEN, 0, index);
 911     if (!blkdev->dinfo) {
 912         Error *local_err = NULL;
 913         QDict *options = NULL;
 914
 915         if (strcmp(blkdev->fileproto, "<unset>")) {
 916             options = qdict_new();
 917             qdict_put(options, "driver", qstring_from_str(blkdev->fileproto));
 918         }
 919
 920         /* setup via xenbus -> create new block driver instance */
 921         xen_be_printf(&blkdev->xendev, 2, "create new bdrv (xenbus setup)\n");
 922         blkdev->blk = blk_new_open(blkdev->filename, NULL, options,
 923                                    qflags, &local_err);
 924         if (!blkdev->blk) {
 925             xen_be_printf(&blkdev->xendev, 0, "error: %s\n",
 926                           error_get_pretty(local_err));
 927             error_free(local_err);
 928             return -1;
 929         }
 930         blk_set_enable_write_cache(blkdev->blk, !writethrough);
 931     } else {
 932         /* setup via qemu cmdline -> already setup for us */
 933         xen_be_printf(&blkdev->xendev, 2, "get configured bdrv (cmdline setup)\n");
 934         blkdev->blk = blk_by_legacy_dinfo(blkdev->dinfo);
 935         if (blk_is_read_only(blkdev->blk) && !readonly) {
 936             xen_be_printf(&blkdev->xendev, 0, "Unexpected read-only drive");
 937             blkdev->blk = NULL;
 938             return -1;
 939         }
 940         /* blkdev->blk is not create by us, we get a reference
 941          * so we can blk_unref() unconditionally */
 942         blk_ref(blkdev->blk);
 943     }
 944     blk_attach_dev_nofail(blkdev->blk, blkdev);
 945     blkdev->file_size = blk_getlength(blkdev->blk);
 946     if (blkdev->file_size < 0) {
 947         BlockDriverState *bs = blk_bs(blkdev->blk);
 948         const char *drv_name = bs ? bdrv_get_format_name(bs) : NULL;
 949         xen_be_printf(&blkdev->xendev, 1, "blk_getlength: %d (%s) | drv %s\n",
 950                       (int)blkdev->file_size, strerror(-blkdev->file_size),
 951                       drv_name ?: "-");
 952         blkdev->file_size = 0;
 953     }
 954
 955     xen_be_printf(xendev, 1, "type \"%s\", fileproto \"%s\", filename \"%s\","
 956                   " size %" PRId64 " (%" PRId64 " MB)\n",
 957                   blkdev->type, blkdev->fileproto, blkdev->filename,
 958                   blkdev->file_size, blkdev->file_size >> 20);
 959
 960     /* Fill in number of sector size and number of sectors */
 961     xenstore_write_be_int(&blkdev->xendev, "sector-size", blkdev->file_blk);
 962     xenstore_write_be_int64(&blkdev->xendev, "sectors",
 963                             blkdev->file_size / blkdev->file_blk);
 964
 965     if (xenstore_read_fe_int(&blkdev->xendev, "ring-ref", &blkdev->ring_ref) == -1) {
 966         return -1;
 967     }
 968     if (xenstore_read_fe_int(&blkdev->xendev, "event-channel",
 969                              &blkdev->xendev.remote_port) == -1) {
 970         return -1;
 971     }
 972     if (xenstore_read_fe_int(&blkdev->xendev, "feature-persistent", &pers)) {
 973         blkdev->feature_persistent = FALSE;
 974     } else {
 975         blkdev->feature_persistent = !!pers;
 976     }
 977
 978     blkdev->protocol = BLKIF_PROTOCOL_NATIVE;
 979     if (blkdev->xendev.protocol) {
 980         if (strcmp(blkdev->xendev.protocol, XEN_IO_PROTO_ABI_X86_32) == 0) {
 981             blkdev->protocol = BLKIF_PROTOCOL_X86_32;
 982         }
 983         if (strcmp(blkdev->xendev.protocol, XEN_IO_PROTO_ABI_X86_64) == 0) {
 984             blkdev->protocol = BLKIF_PROTOCOL_X86_64;
 985         }
 986     }
 987
 988     blkdev->sring = xengnttab_map_grant_ref(blkdev->xendev.gnttabdev,
 989                                             blkdev->xendev.dom,
 990                                             blkdev->ring_ref,
 991                                             PROT_READ | PROT_WRITE);
 992     if (!blkdev->sring) {
 993         return -1;
 994     }
 995     blkdev->cnt_map++;
 996
 997     switch (blkdev->protocol) {
 998     case BLKIF_PROTOCOL_NATIVE:
 999     {
1000         blkif_sring_t *sring_native = blkdev->sring;
1001         BACK_RING_INIT(&blkdev->rings.native, sring_native, XC_PAGE_SIZE);
1002         break;
1003     }
1004     case BLKIF_PROTOCOL_X86_32:
1005     {
1006         blkif_x86_32_sring_t *sring_x86_32 = blkdev->sring;
1007
1008         BACK_RING_INIT(&blkdev->rings.x86_32_part, sring_x86_32, XC_PAGE_SIZE);
1009         break;
1010     }
1011     case BLKIF_PROTOCOL_X86_64:
1012     {
1013         blkif_x86_64_sring_t *sring_x86_64 = blkdev->sring;
1014
1015         BACK_RING_INIT(&blkdev->rings.x86_64_part, sring_x86_64, XC_PAGE_SIZE);
1016         break;
1017     }
1018     }
1019
1020     if (blkdev->feature_persistent) {
1021         /* Init persistent grants */
1022         blkdev->max_grants = max_requests * BLKIF_MAX_SEGMENTS_PER_REQUEST;
1023         blkdev->persistent_gnts = g_tree_new_full((GCompareDataFunc)int_cmp,
1024                                              NULL, NULL,
1025                                              batch_maps ?
1026                                              (GDestroyNotify)g_free :
1027                                              (GDestroyNotify)destroy_grant);
1028         blkdev->persistent_regions = NULL;
1029         blkdev->persistent_gnt_count = 0;
1030     }
1031
1032     xen_be_bind_evtchn(&blkdev->xendev);
1033
1034     xen_be_printf(&blkdev->xendev, 1, "ok: proto %s, ring-ref %d, "
1035                   "remote port %d, local port %d\n",
1036                   blkdev->xendev.protocol, blkdev->ring_ref,
1037                   blkdev->xendev.remote_port, blkdev->xendev.local_port);
1038     return 0;
1039 }
1040
1041 static void blk_disconnect(struct XenDevice *xendev)
1042 {
1043     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
1044
1045     if (blkdev->blk) {
1046         blk_detach_dev(blkdev->blk, blkdev);
1047         blk_unref(blkdev->blk);
1048         blkdev->blk = NULL;
1049     }
1050     xen_be_unbind_evtchn(&blkdev->xendev);
1051
1052     if (blkdev->sring) {
1053         xengnttab_unmap(blkdev->xendev.gnttabdev, blkdev->sring, 1);
1054         blkdev->cnt_map--;
1055         blkdev->sring = NULL;
1056     }
1057
1058     /*
1059      * Unmap persistent grants before switching to the closed state
1060      * so the frontend can free them.
1061      *
1062      * In the !batch_maps case g_tree_destroy will take care of unmapping
1063      * the grant, but in the batch_maps case we need to iterate over every
1064      * region in persistent_regions and unmap it.
1065      */
1066     if (blkdev->feature_persistent) {
1067         g_tree_destroy(blkdev->persistent_gnts);
1068         assert(batch_maps || blkdev->persistent_gnt_count == 0);
1069         if (batch_maps) {
1070             blkdev->persistent_gnt_count = 0;
1071             g_slist_foreach(blkdev->persistent_regions,
1072                             (GFunc)remove_persistent_region, blkdev);
1073             g_slist_free(blkdev->persistent_regions);
1074         }
1075         blkdev->feature_persistent = false;
1076     }
1077 }
1078
1079 static int blk_free(struct XenDevice *xendev)
1080 {
1081     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
1082     struct ioreq *ioreq;
1083
1084     if (blkdev->blk || blkdev->sring) {
1085         blk_disconnect(xendev);
1086     }
1087
1088     while (!QLIST_EMPTY(&blkdev->freelist)) {
1089         ioreq = QLIST_FIRST(&blkdev->freelist);
1090         QLIST_REMOVE(ioreq, list);
1091         qemu_iovec_destroy(&ioreq->v);
1092         g_free(ioreq);
1093     }
1094
1095     g_free(blkdev->params);
1096     g_free(blkdev->mode);
1097     g_free(blkdev->type);
1098     g_free(blkdev->dev);
1099     g_free(blkdev->devtype);
1100     qemu_bh_delete(blkdev->bh);
1101     return 0;
1102 }
1103
1104 static void blk_event(struct XenDevice *xendev)
1105 {
1106     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
1107
1108     qemu_bh_schedule(blkdev->bh);
1109 }
1110
1111 struct XenDevOps xen_blkdev_ops = {
1112     .size       = sizeof(struct XenBlkDev),
1113     .flags      = DEVOPS_FLAG_NEED_GNTDEV,
1114     .alloc      = blk_alloc,
1115     .init       = blk_init,
1116     .initialise    = blk_connect,
1117     .disconnect = blk_disconnect,
1118     .event      = blk_event,
1119     .free       = blk_free,
1120 };