qemu/hw/block/xen_disk.c

   1 /*
   2  *  xen paravirt block device backend
   3  *
   4  *  (c) Gerd Hoffmann <kraxel@redhat.com>
   5  *
   6  *  This program is free software; you can redistribute it and/or modify
   7  *  it under the terms of the GNU General Public License as published by
   8  *  the Free Software Foundation; under version 2 of the License.
   9  *
  10  *  This program is distributed in the hope that it will be useful,
  11  *  but WITHOUT ANY WARRANTY; without even the implied warranty of
  12  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  13  *  GNU General Public License for more details.
  14  *
  15  *  You should have received a copy of the GNU General Public License along
  16  *  with this program; if not, see <http://www.gnu.org/licenses/>.
  17  *
  18  *  Contributions after 2012-01-13 are licensed under the terms of the
  19  *  GNU GPL, version 2 or (at your option) any later version.
  20  */
  21
  22 #include <stdio.h>
  23 #include <stdlib.h>
  24 #include <stdarg.h>
  25 #include <string.h>
  26 #include <unistd.h>
  27 #include <signal.h>
  28 #include <inttypes.h>
  29 #include <time.h>
  30 #include <fcntl.h>
  31 #include <errno.h>
  32 #include <sys/ioctl.h>
  33 #include <sys/types.h>
  34 #include <sys/stat.h>
  35 #include <sys/mman.h>
  36 #include <sys/uio.h>
  37
  38 #include "hw/hw.h"
  39 #include "hw/xen/xen_backend.h"
  40 #include "xen_blkif.h"
  41 #include "sysemu/blockdev.h"
  42 #include "sysemu/block-backend.h"
  43 #include "qapi/qmp/qdict.h"
  44 #include "qapi/qmp/qstring.h"
  45
  46 /* ------------------------------------------------------------- */
  47
  48 static int batch_maps   = 0;
  49
  50 static int max_requests = 32;
  51
  52 /* ------------------------------------------------------------- */
  53
  54 #define BLOCK_SIZE  512
  55 #define IOCB_COUNT  (BLKIF_MAX_SEGMENTS_PER_REQUEST + 2)
  56
  57 struct PersistentGrant {
  58     void *page;
  59     struct XenBlkDev *blkdev;
  60 };
  61
  62 typedef struct PersistentGrant PersistentGrant;
  63
  64 struct PersistentRegion {
  65     void *addr;
  66     int num;
  67 };
  68
  69 typedef struct PersistentRegion PersistentRegion;
  70
  71 struct ioreq {
  72     blkif_request_t     req;
  73     int16_t             status;
  74
  75     /* parsed request */
  76     off_t               start;
  77     QEMUIOVector        v;
  78     int                 presync;
  79     int                 postsync;
  80     uint8_t             mapped;
  81
  82     /* grant mapping */
  83     uint32_t            domids[BLKIF_MAX_SEGMENTS_PER_REQUEST];
  84     uint32_t            refs[BLKIF_MAX_SEGMENTS_PER_REQUEST];
  85     int                 prot;
  86     void                *page[BLKIF_MAX_SEGMENTS_PER_REQUEST];
  87     void                *pages;
  88     int                 num_unmap;
  89
  90     /* aio status */
  91     int                 aio_inflight;
  92     int                 aio_errors;
  93
  94     struct XenBlkDev    *blkdev;
  95     QLIST_ENTRY(ioreq)   list;
  96     BlockAcctCookie     acct;
  97 };
  98
  99 struct XenBlkDev {
 100     struct XenDevice    xendev;  /* must be first */
 101     char                *params;
 102     char                *mode;
 103     char                *type;
 104     char                *dev;
 105     char                *devtype;
 106     bool                directiosafe;
 107     const char          *fileproto;
 108     const char          *filename;
 109     int                 ring_ref;
 110     void                *sring;
 111     int64_t             file_blk;
 112     int64_t             file_size;
 113     int                 protocol;
 114     blkif_back_rings_t  rings;
 115     int                 more_work;
 116     int                 cnt_map;
 117
 118     /* request lists */
 119     QLIST_HEAD(inflight_head, ioreq) inflight;
 120     QLIST_HEAD(finished_head, ioreq) finished;
 121     QLIST_HEAD(freelist_head, ioreq) freelist;
 122     int                 requests_total;
 123     int                 requests_inflight;
 124     int                 requests_finished;
 125
 126     /* Persistent grants extension */
 127     gboolean            feature_discard;
 128     gboolean            feature_persistent;
 129     GTree               *persistent_gnts;
 130     GSList              *persistent_regions;
 131     unsigned int        persistent_gnt_count;
 132     unsigned int        max_grants;
 133
 134     /* qemu block driver */
 135     DriveInfo           *dinfo;
 136     BlockBackend        *blk;
 137     QEMUBH              *bh;
 138 };
 139
 140 /* ------------------------------------------------------------- */
 141
 142 static void ioreq_reset(struct ioreq *ioreq)
 143 {
 144     memset(&ioreq->req, 0, sizeof(ioreq->req));
 145     ioreq->status = 0;
 146     ioreq->start = 0;
 147     ioreq->presync = 0;
 148     ioreq->postsync = 0;
 149     ioreq->mapped = 0;
 150
 151     memset(ioreq->domids, 0, sizeof(ioreq->domids));
 152     memset(ioreq->refs, 0, sizeof(ioreq->refs));
 153     ioreq->prot = 0;
 154     memset(ioreq->page, 0, sizeof(ioreq->page));
 155     ioreq->pages = NULL;
 156
 157     ioreq->aio_inflight = 0;
 158     ioreq->aio_errors = 0;
 159
 160     ioreq->blkdev = NULL;
 161     memset(&ioreq->list, 0, sizeof(ioreq->list));
 162     memset(&ioreq->acct, 0, sizeof(ioreq->acct));
 163
 164     qemu_iovec_reset(&ioreq->v);
 165 }
 166
 167 static gint int_cmp(gconstpointer a, gconstpointer b, gpointer user_data)
 168 {
 169     uint ua = GPOINTER_TO_UINT(a);
 170     uint ub = GPOINTER_TO_UINT(b);
 171     return (ua > ub) - (ua < ub);
 172 }
 173
 174 static void destroy_grant(gpointer pgnt)
 175 {
 176     PersistentGrant *grant = pgnt;
 177     XenGnttab gnt = grant->blkdev->xendev.gnttabdev;
 178
 179     if (xc_gnttab_munmap(gnt, grant->page, 1) != 0) {
 180         xen_be_printf(&grant->blkdev->xendev, 0,
 181                       "xc_gnttab_munmap failed: %s\n",
 182                       strerror(errno));
 183     }
 184     grant->blkdev->persistent_gnt_count--;
 185     xen_be_printf(&grant->blkdev->xendev, 3,
 186                   "unmapped grant %p\n", grant->page);
 187     g_free(grant);
 188 }
 189
 190 static void remove_persistent_region(gpointer data, gpointer dev)
 191 {
 192     PersistentRegion *region = data;
 193     struct XenBlkDev *blkdev = dev;
 194     XenGnttab gnt = blkdev->xendev.gnttabdev;
 195
 196     if (xc_gnttab_munmap(gnt, region->addr, region->num) != 0) {
 197         xen_be_printf(&blkdev->xendev, 0,
 198                       "xc_gnttab_munmap region %p failed: %s\n",
 199                       region->addr, strerror(errno));
 200     }
 201     xen_be_printf(&blkdev->xendev, 3,
 202                   "unmapped grant region %p with %d pages\n",
 203                   region->addr, region->num);
 204     g_free(region);
 205 }
 206
 207 static struct ioreq *ioreq_start(struct XenBlkDev *blkdev)
 208 {
 209     struct ioreq *ioreq = NULL;
 210
 211     if (QLIST_EMPTY(&blkdev->freelist)) {
 212         if (blkdev->requests_total >= max_requests) {
 213             goto out;
 214         }
 215         /* allocate new struct */
 216         ioreq = g_malloc0(sizeof(*ioreq));
 217         ioreq->blkdev = blkdev;
 218         blkdev->requests_total++;
 219         qemu_iovec_init(&ioreq->v, BLKIF_MAX_SEGMENTS_PER_REQUEST);
 220     } else {
 221         /* get one from freelist */
 222         ioreq = QLIST_FIRST(&blkdev->freelist);
 223         QLIST_REMOVE(ioreq, list);
 224     }
 225     QLIST_INSERT_HEAD(&blkdev->inflight, ioreq, list);
 226     blkdev->requests_inflight++;
 227
 228 out:
 229     return ioreq;
 230 }
 231
 232 static void ioreq_finish(struct ioreq *ioreq)
 233 {
 234     struct XenBlkDev *blkdev = ioreq->blkdev;
 235
 236     QLIST_REMOVE(ioreq, list);
 237     QLIST_INSERT_HEAD(&blkdev->finished, ioreq, list);
 238     blkdev->requests_inflight--;
 239     blkdev->requests_finished++;
 240 }
 241
 242 static void ioreq_release(struct ioreq *ioreq, bool finish)
 243 {
 244     struct XenBlkDev *blkdev = ioreq->blkdev;
 245
 246     QLIST_REMOVE(ioreq, list);
 247     ioreq_reset(ioreq);
 248     ioreq->blkdev = blkdev;
 249     QLIST_INSERT_HEAD(&blkdev->freelist, ioreq, list);
 250     if (finish) {
 251         blkdev->requests_finished--;
 252     } else {
 253         blkdev->requests_inflight--;
 254     }
 255 }
 256
 257 /*
 258  * translate request into iovec + start offset
 259  * do sanity checks along the way
 260  */
 261 static int ioreq_parse(struct ioreq *ioreq)
 262 {
 263     struct XenBlkDev *blkdev = ioreq->blkdev;
 264     uintptr_t mem;
 265     size_t len;
 266     int i;
 267
 268     xen_be_printf(&blkdev->xendev, 3,
 269                   "op %d, nr %d, handle %d, id %" PRId64 ", sector %" PRId64 "\n",
 270                   ioreq->req.operation, ioreq->req.nr_segments,
 271                   ioreq->req.handle, ioreq->req.id, ioreq->req.sector_number);
 272     switch (ioreq->req.operation) {
 273     case BLKIF_OP_READ:
 274         ioreq->prot = PROT_WRITE; /* to memory */
 275         break;
 276     case BLKIF_OP_FLUSH_DISKCACHE:
 277         ioreq->presync = 1;
 278         if (!ioreq->req.nr_segments) {
 279             return 0;
 280         }
 281         /* fall through */
 282     case BLKIF_OP_WRITE:
 283         ioreq->prot = PROT_READ; /* from memory */
 284         break;
 285     case BLKIF_OP_DISCARD:
 286         return 0;
 287     default:
 288         xen_be_printf(&blkdev->xendev, 0, "error: unknown operation (%d)\n",
 289                       ioreq->req.operation);
 290         goto err;
 291     };
 292
 293     if (ioreq->req.operation != BLKIF_OP_READ && blkdev->mode[0] != 'w') {
 294         xen_be_printf(&blkdev->xendev, 0, "error: write req for ro device\n");
 295         goto err;
 296     }
 297
 298     ioreq->start = ioreq->req.sector_number * blkdev->file_blk;
 299     for (i = 0; i < ioreq->req.nr_segments; i++) {
 300         if (i == BLKIF_MAX_SEGMENTS_PER_REQUEST) {
 301             xen_be_printf(&blkdev->xendev, 0, "error: nr_segments too big\n");
 302             goto err;
 303         }
 304         if (ioreq->req.seg[i].first_sect > ioreq->req.seg[i].last_sect) {
 305             xen_be_printf(&blkdev->xendev, 0, "error: first > last sector\n");
 306             goto err;
 307         }
 308         if (ioreq->req.seg[i].last_sect * BLOCK_SIZE >= XC_PAGE_SIZE) {
 309             xen_be_printf(&blkdev->xendev, 0, "error: page crossing\n");
 310             goto err;
 311         }
 312
 313         ioreq->domids[i] = blkdev->xendev.dom;
 314         ioreq->refs[i]   = ioreq->req.seg[i].gref;
 315
 316         mem = ioreq->req.seg[i].first_sect * blkdev->file_blk;
 317         len = (ioreq->req.seg[i].last_sect - ioreq->req.seg[i].first_sect + 1) * blkdev->file_blk;
 318         qemu_iovec_add(&ioreq->v, (void*)mem, len);
 319     }
 320     if (ioreq->start + ioreq->v.size > blkdev->file_size) {
 321         xen_be_printf(&blkdev->xendev, 0, "error: access beyond end of file\n");
 322         goto err;
 323     }
 324     return 0;
 325
 326 err:
 327     ioreq->status = BLKIF_RSP_ERROR;
 328     return -1;
 329 }
 330
 331 static void ioreq_unmap(struct ioreq *ioreq)
 332 {
 333     XenGnttab gnt = ioreq->blkdev->xendev.gnttabdev;
 334     int i;
 335
 336     if (ioreq->num_unmap == 0 || ioreq->mapped == 0) {
 337         return;
 338     }
 339     if (batch_maps) {
 340         if (!ioreq->pages) {
 341             return;
 342         }
 343         if (xc_gnttab_munmap(gnt, ioreq->pages, ioreq->num_unmap) != 0) {
 344             xen_be_printf(&ioreq->blkdev->xendev, 0, "xc_gnttab_munmap failed: %s\n",
 345                           strerror(errno));
 346         }
 347         ioreq->blkdev->cnt_map -= ioreq->num_unmap;
 348         ioreq->pages = NULL;
 349     } else {
 350         for (i = 0; i < ioreq->num_unmap; i++) {
 351             if (!ioreq->page[i]) {
 352                 continue;
 353             }
 354             if (xc_gnttab_munmap(gnt, ioreq->page[i], 1) != 0) {
 355                 xen_be_printf(&ioreq->blkdev->xendev, 0, "xc_gnttab_munmap failed: %s\n",
 356                               strerror(errno));
 357             }
 358             ioreq->blkdev->cnt_map--;
 359             ioreq->page[i] = NULL;
 360         }
 361     }
 362     ioreq->mapped = 0;
 363 }
 364
 365 static int ioreq_map(struct ioreq *ioreq)
 366 {
 367     XenGnttab gnt = ioreq->blkdev->xendev.gnttabdev;
 368     uint32_t domids[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 369     uint32_t refs[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 370     void *page[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 371     int i, j, new_maps = 0;
 372     PersistentGrant *grant;
 373     PersistentRegion *region;
 374     /* domids and refs variables will contain the information necessary
 375      * to map the grants that are needed to fulfill this request.
 376      *
 377      * After mapping the needed grants, the page array will contain the
 378      * memory address of each granted page in the order specified in ioreq
 379      * (disregarding if it's a persistent grant or not).
 380      */
 381
 382     if (ioreq->v.niov == 0 || ioreq->mapped == 1) {
 383         return 0;
 384     }
 385     if (ioreq->blkdev->feature_persistent) {
 386         for (i = 0; i < ioreq->v.niov; i++) {
 387             grant = g_tree_lookup(ioreq->blkdev->persistent_gnts,
 388                                     GUINT_TO_POINTER(ioreq->refs[i]));
 389
 390             if (grant != NULL) {
 391                 page[i] = grant->page;
 392                 xen_be_printf(&ioreq->blkdev->xendev, 3,
 393                               "using persistent-grant %" PRIu32 "\n",
 394                               ioreq->refs[i]);
 395             } else {
 396                     /* Add the grant to the list of grants that
 397                      * should be mapped
 398                      */
 399                     domids[new_maps] = ioreq->domids[i];
 400                     refs[new_maps] = ioreq->refs[i];
 401                     page[i] = NULL;
 402                     new_maps++;
 403             }
 404         }
 405         /* Set the protection to RW, since grants may be reused later
 406          * with a different protection than the one needed for this request
 407          */
 408         ioreq->prot = PROT_WRITE | PROT_READ;
 409     } else {
 410         /* All grants in the request should be mapped */
 411         memcpy(refs, ioreq->refs, sizeof(refs));
 412         memcpy(domids, ioreq->domids, sizeof(domids));
 413         memset(page, 0, sizeof(page));
 414         new_maps = ioreq->v.niov;
 415     }
 416
 417     if (batch_maps && new_maps) {
 418         ioreq->pages = xc_gnttab_map_grant_refs
 419             (gnt, new_maps, domids, refs, ioreq->prot);
 420         if (ioreq->pages == NULL) {
 421             xen_be_printf(&ioreq->blkdev->xendev, 0,
 422                           "can't map %d grant refs (%s, %d maps)\n",
 423                           new_maps, strerror(errno), ioreq->blkdev->cnt_map);
 424             return -1;
 425         }
 426         for (i = 0, j = 0; i < ioreq->v.niov; i++) {
 427             if (page[i] == NULL) {
 428                 page[i] = ioreq->pages + (j++) * XC_PAGE_SIZE;
 429             }
 430         }
 431         ioreq->blkdev->cnt_map += new_maps;
 432     } else if (new_maps)  {
 433         for (i = 0; i < new_maps; i++) {
 434             ioreq->page[i] = xc_gnttab_map_grant_ref
 435                 (gnt, domids[i], refs[i], ioreq->prot);
 436             if (ioreq->page[i] == NULL) {
 437                 xen_be_printf(&ioreq->blkdev->xendev, 0,
 438                               "can't map grant ref %d (%s, %d maps)\n",
 439                               refs[i], strerror(errno), ioreq->blkdev->cnt_map);
 440                 ioreq->mapped = 1;
 441                 ioreq_unmap(ioreq);
 442                 return -1;
 443             }
 444             ioreq->blkdev->cnt_map++;
 445         }
 446         for (i = 0, j = 0; i < ioreq->v.niov; i++) {
 447             if (page[i] == NULL) {
 448                 page[i] = ioreq->page[j++];
 449             }
 450         }
 451     }
 452     if (ioreq->blkdev->feature_persistent && new_maps != 0 &&
 453         (!batch_maps || (ioreq->blkdev->persistent_gnt_count + new_maps <=
 454         ioreq->blkdev->max_grants))) {
 455         /*
 456          * If we are using persistent grants and batch mappings only
 457          * add the new maps to the list of persistent grants if the whole
 458          * area can be persistently mapped.
 459          */
 460         if (batch_maps) {
 461             region = g_malloc0(sizeof(*region));
 462             region->addr = ioreq->pages;
 463             region->num = new_maps;
 464             ioreq->blkdev->persistent_regions = g_slist_append(
 465                                             ioreq->blkdev->persistent_regions,
 466                                             region);
 467         }
 468         while ((ioreq->blkdev->persistent_gnt_count < ioreq->blkdev->max_grants)
 469               && new_maps) {
 470             /* Go through the list of newly mapped grants and add as many
 471              * as possible to the list of persistently mapped grants.
 472              *
 473              * Since we start at the end of ioreq->page(s), we only need
 474              * to decrease new_maps to prevent this granted pages from
 475              * being unmapped in ioreq_unmap.
 476              */
 477             grant = g_malloc0(sizeof(*grant));
 478             new_maps--;
 479             if (batch_maps) {
 480                 grant->page = ioreq->pages + (new_maps) * XC_PAGE_SIZE;
 481             } else {
 482                 grant->page = ioreq->page[new_maps];
 483             }
 484             grant->blkdev = ioreq->blkdev;
 485             xen_be_printf(&ioreq->blkdev->xendev, 3,
 486                           "adding grant %" PRIu32 " page: %p\n",
 487                           refs[new_maps], grant->page);
 488             g_tree_insert(ioreq->blkdev->persistent_gnts,
 489                           GUINT_TO_POINTER(refs[new_maps]),
 490                           grant);
 491             ioreq->blkdev->persistent_gnt_count++;
 492         }
 493         assert(!batch_maps || new_maps == 0);
 494     }
 495     for (i = 0; i < ioreq->v.niov; i++) {
 496         ioreq->v.iov[i].iov_base += (uintptr_t)page[i];
 497     }
 498     ioreq->mapped = 1;
 499     ioreq->num_unmap = new_maps;
 500     return 0;
 501 }
 502
 503 static int ioreq_runio_qemu_aio(struct ioreq *ioreq);
 504
 505 static void qemu_aio_complete(void *opaque, int ret)
 506 {
 507     struct ioreq *ioreq = opaque;
 508
 509     if (ret != 0) {
 510         xen_be_printf(&ioreq->blkdev->xendev, 0, "%s I/O error\n",
 511                       ioreq->req.operation == BLKIF_OP_READ ? "read" : "write");
 512         ioreq->aio_errors++;
 513     }
 514
 515     ioreq->aio_inflight--;
 516     if (ioreq->presync) {
 517         ioreq->presync = 0;
 518         ioreq_runio_qemu_aio(ioreq);
 519         return;
 520     }
 521     if (ioreq->aio_inflight > 0) {
 522         return;
 523     }
 524     if (ioreq->postsync) {
 525         ioreq->postsync = 0;
 526         ioreq->aio_inflight++;
 527         blk_aio_flush(ioreq->blkdev->blk, qemu_aio_complete, ioreq);
 528         return;
 529     }
 530
 531     ioreq->status = ioreq->aio_errors ? BLKIF_RSP_ERROR : BLKIF_RSP_OKAY;
 532     ioreq_unmap(ioreq);
 533     ioreq_finish(ioreq);
 534     switch (ioreq->req.operation) {
 535     case BLKIF_OP_WRITE:
 536     case BLKIF_OP_FLUSH_DISKCACHE:
 537         if (!ioreq->req.nr_segments) {
 538             break;
 539         }
 540     case BLKIF_OP_READ:
 541         block_acct_done(blk_get_stats(ioreq->blkdev->blk), &ioreq->acct);
 542         break;
 543     case BLKIF_OP_DISCARD:
 544     default:
 545         break;
 546     }
 547     qemu_bh_schedule(ioreq->blkdev->bh);
 548 }
 549
 550 static int ioreq_runio_qemu_aio(struct ioreq *ioreq)
 551 {
 552     struct XenBlkDev *blkdev = ioreq->blkdev;
 553
 554     if (ioreq->req.nr_segments && ioreq_map(ioreq) == -1) {
 555         goto err_no_map;
 556     }
 557
 558     ioreq->aio_inflight++;
 559     if (ioreq->presync) {
 560         blk_aio_flush(ioreq->blkdev->blk, qemu_aio_complete, ioreq);
 561         return 0;
 562     }
 563
 564     switch (ioreq->req.operation) {
 565     case BLKIF_OP_READ:
 566         block_acct_start(blk_get_stats(blkdev->blk), &ioreq->acct,
 567                          ioreq->v.size, BLOCK_ACCT_READ);
 568         ioreq->aio_inflight++;
 569         blk_aio_readv(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_WRITE:
 574     case BLKIF_OP_FLUSH_DISKCACHE:
 575         if (!ioreq->req.nr_segments) {
 576             break;
 577         }
 578
 579         block_acct_start(blk_get_stats(blkdev->blk), &ioreq->acct,
 580                          ioreq->v.size, BLOCK_ACCT_WRITE);
 581         ioreq->aio_inflight++;
 582         blk_aio_writev(blkdev->blk, ioreq->start / BLOCK_SIZE,
 583                        &ioreq->v, ioreq->v.size / BLOCK_SIZE,
 584                        qemu_aio_complete, ioreq);
 585         break;
 586     case BLKIF_OP_DISCARD:
 587     {
 588         struct blkif_request_discard *discard_req = (void *)&ioreq->req;
 589         ioreq->aio_inflight++;
 590         blk_aio_discard(blkdev->blk,
 591                         discard_req->sector_number, discard_req->nr_sectors,
 592                         qemu_aio_complete, ioreq);
 593         break;
 594     }
 595     default:
 596         /* unknown operation (shouldn't happen -- parse catches this) */
 597         goto err;
 598     }
 599
 600     qemu_aio_complete(ioreq, 0);
 601
 602     return 0;
 603
 604 err:
 605     ioreq_unmap(ioreq);
 606 err_no_map:
 607     ioreq_finish(ioreq);
 608     ioreq->status = BLKIF_RSP_ERROR;
 609     return -1;
 610 }
 611
 612 static int blk_send_response_one(struct ioreq *ioreq)
 613 {
 614     struct XenBlkDev  *blkdev = ioreq->blkdev;
 615     int               send_notify   = 0;
 616     int               have_requests = 0;
 617     blkif_response_t  resp;
 618     void              *dst;
 619
 620     resp.id        = ioreq->req.id;
 621     resp.operation = ioreq->req.operation;
 622     resp.status    = ioreq->status;
 623
 624     /* Place on the response ring for the relevant domain. */
 625     switch (blkdev->protocol) {
 626     case BLKIF_PROTOCOL_NATIVE:
 627         dst = RING_GET_RESPONSE(&blkdev->rings.native, blkdev->rings.native.rsp_prod_pvt);
 628         break;
 629     case BLKIF_PROTOCOL_X86_32:
 630         dst = RING_GET_RESPONSE(&blkdev->rings.x86_32_part,
 631                                 blkdev->rings.x86_32_part.rsp_prod_pvt);
 632         break;
 633     case BLKIF_PROTOCOL_X86_64:
 634         dst = RING_GET_RESPONSE(&blkdev->rings.x86_64_part,
 635                                 blkdev->rings.x86_64_part.rsp_prod_pvt);
 636         break;
 637     default:
 638         dst = NULL;
 639         return 0;
 640     }
 641     memcpy(dst, &resp, sizeof(resp));
 642     blkdev->rings.common.rsp_prod_pvt++;
 643
 644     RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blkdev->rings.common, send_notify);
 645     if (blkdev->rings.common.rsp_prod_pvt == blkdev->rings.common.req_cons) {
 646         /*
 647          * Tail check for pending requests. Allows frontend to avoid
 648          * notifications if requests are already in flight (lower
 649          * overheads and promotes batching).
 650          */
 651         RING_FINAL_CHECK_FOR_REQUESTS(&blkdev->rings.common, have_requests);
 652     } else if (RING_HAS_UNCONSUMED_REQUESTS(&blkdev->rings.common)) {
 653         have_requests = 1;
 654     }
 655
 656     if (have_requests) {
 657         blkdev->more_work++;
 658     }
 659     return send_notify;
 660 }
 661
 662 /* walk finished list, send outstanding responses, free requests */
 663 static void blk_send_response_all(struct XenBlkDev *blkdev)
 664 {
 665     struct ioreq *ioreq;
 666     int send_notify = 0;
 667
 668     while (!QLIST_EMPTY(&blkdev->finished)) {
 669         ioreq = QLIST_FIRST(&blkdev->finished);
 670         send_notify += blk_send_response_one(ioreq);
 671         ioreq_release(ioreq, true);
 672     }
 673     if (send_notify) {
 674         xen_be_send_notify(&blkdev->xendev);
 675     }
 676 }
 677
 678 static int blk_get_request(struct XenBlkDev *blkdev, struct ioreq *ioreq, RING_IDX rc)
 679 {
 680     switch (blkdev->protocol) {
 681     case BLKIF_PROTOCOL_NATIVE:
 682         memcpy(&ioreq->req, RING_GET_REQUEST(&blkdev->rings.native, rc),
 683                sizeof(ioreq->req));
 684         break;
 685     case BLKIF_PROTOCOL_X86_32:
 686         blkif_get_x86_32_req(&ioreq->req,
 687                              RING_GET_REQUEST(&blkdev->rings.x86_32_part, rc));
 688         break;
 689     case BLKIF_PROTOCOL_X86_64:
 690         blkif_get_x86_64_req(&ioreq->req,
 691                              RING_GET_REQUEST(&blkdev->rings.x86_64_part, rc));
 692         break;
 693     }
 694     return 0;
 695 }
 696
 697 static void blk_handle_requests(struct XenBlkDev *blkdev)
 698 {
 699     RING_IDX rc, rp;
 700     struct ioreq *ioreq;
 701
 702     blkdev->more_work = 0;
 703
 704     rc = blkdev->rings.common.req_cons;
 705     rp = blkdev->rings.common.sring->req_prod;
 706     xen_rmb(); /* Ensure we see queued requests up to 'rp'. */
 707
 708     blk_send_response_all(blkdev);
 709     while (rc != rp) {
 710         /* pull request from ring */
 711         if (RING_REQUEST_CONS_OVERFLOW(&blkdev->rings.common, rc)) {
 712             break;
 713         }
 714         ioreq = ioreq_start(blkdev);
 715         if (ioreq == NULL) {
 716             blkdev->more_work++;
 717             break;
 718         }
 719         blk_get_request(blkdev, ioreq, rc);
 720         blkdev->rings.common.req_cons = ++rc;
 721
 722         /* parse them */
 723         if (ioreq_parse(ioreq) != 0) {
 724             if (blk_send_response_one(ioreq)) {
 725                 xen_be_send_notify(&blkdev->xendev);
 726             }
 727             ioreq_release(ioreq, false);
 728             continue;
 729         }
 730
 731         ioreq_runio_qemu_aio(ioreq);
 732     }
 733
 734     if (blkdev->more_work && blkdev->requests_inflight < max_requests) {
 735         qemu_bh_schedule(blkdev->bh);
 736     }
 737 }
 738
 739 /* ------------------------------------------------------------- */
 740
 741 static void blk_bh(void *opaque)
 742 {
 743     struct XenBlkDev *blkdev = opaque;
 744     blk_handle_requests(blkdev);
 745 }
 746
 747 /*
 748  * We need to account for the grant allocations requiring contiguous
 749  * chunks; the worst case number would be
 750  *     max_req * max_seg + (max_req - 1) * (max_seg - 1) + 1,
 751  * but in order to keep things simple just use
 752  *     2 * max_req * max_seg.
 753  */
 754 #define MAX_GRANTS(max_req, max_seg) (2 * (max_req) * (max_seg))
 755
 756 static void blk_alloc(struct XenDevice *xendev)
 757 {
 758     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
 759
 760     QLIST_INIT(&blkdev->inflight);
 761     QLIST_INIT(&blkdev->finished);
 762     QLIST_INIT(&blkdev->freelist);
 763     blkdev->bh = qemu_bh_new(blk_bh, blkdev);
 764     if (xen_mode != XEN_EMULATE) {
 765         batch_maps = 1;
 766     }
 767     if (xc_gnttab_set_max_grants(xendev->gnttabdev,
 768             MAX_GRANTS(max_requests, BLKIF_MAX_SEGMENTS_PER_REQUEST)) < 0) {
 769         xen_be_printf(xendev, 0, "xc_gnttab_set_max_grants failed: %s\n",
 770                       strerror(errno));
 771     }
 772 }
 773
 774 static void blk_parse_discard(struct XenBlkDev *blkdev)
 775 {
 776     int enable;
 777
 778     blkdev->feature_discard = true;
 779
 780     if (xenstore_read_be_int(&blkdev->xendev, "discard-enable", &enable) == 0) {
 781         blkdev->feature_discard = !!enable;
 782     }
 783
 784     if (blkdev->feature_discard) {
 785         xenstore_write_be_int(&blkdev->xendev, "feature-discard", 1);
 786     }
 787 }
 788
 789 static int blk_init(struct XenDevice *xendev)
 790 {
 791     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
 792     int info = 0;
 793     char *directiosafe = NULL;
 794
 795     /* read xenstore entries */
 796     if (blkdev->params == NULL) {
 797         char *h = NULL;
 798         blkdev->params = xenstore_read_be_str(&blkdev->xendev, "params");
 799         if (blkdev->params != NULL) {
 800             h = strchr(blkdev->params, ':');
 801         }
 802         if (h != NULL) {
 803             blkdev->fileproto = blkdev->params;
 804             blkdev->filename  = h+1;
 805             *h = 0;
 806         } else {
 807             blkdev->fileproto = "<unset>";
 808             blkdev->filename  = blkdev->params;
 809         }
 810     }
 811     if (!strcmp("aio", blkdev->fileproto)) {
 812         blkdev->fileproto = "raw";
 813     }
 814     if (blkdev->mode == NULL) {
 815         blkdev->mode = xenstore_read_be_str(&blkdev->xendev, "mode");
 816     }
 817     if (blkdev->type == NULL) {
 818         blkdev->type = xenstore_read_be_str(&blkdev->xendev, "type");
 819     }
 820     if (blkdev->dev == NULL) {
 821         blkdev->dev = xenstore_read_be_str(&blkdev->xendev, "dev");
 822     }
 823     if (blkdev->devtype == NULL) {
 824         blkdev->devtype = xenstore_read_be_str(&blkdev->xendev, "device-type");
 825     }
 826     directiosafe = xenstore_read_be_str(&blkdev->xendev, "direct-io-safe");
 827     blkdev->directiosafe = (directiosafe && atoi(directiosafe));
 828
 829     /* do we have all we need? */
 830     if (blkdev->params == NULL ||
 831         blkdev->mode == NULL   ||
 832         blkdev->type == NULL   ||
 833         blkdev->dev == NULL) {
 834         goto out_error;
 835     }
 836
 837     /* read-only ? */
 838     if (strcmp(blkdev->mode, "w")) {
 839         info  |= VDISK_READONLY;
 840     }
 841
 842     /* cdrom ? */
 843     if (blkdev->devtype && !strcmp(blkdev->devtype, "cdrom")) {
 844         info  |= VDISK_CDROM;
 845     }
 846
 847     blkdev->file_blk  = BLOCK_SIZE;
 848
 849     /* fill info
 850      * blk_connect supplies sector-size and sectors
 851      */
 852     xenstore_write_be_int(&blkdev->xendev, "feature-flush-cache", 1);
 853     xenstore_write_be_int(&blkdev->xendev, "feature-persistent", 1);
 854     xenstore_write_be_int(&blkdev->xendev, "info", info);
 855
 856     blk_parse_discard(blkdev);
 857
 858     g_free(directiosafe);
 859     return 0;
 860
 861 out_error:
 862     g_free(blkdev->params);
 863     blkdev->params = NULL;
 864     g_free(blkdev->mode);
 865     blkdev->mode = NULL;
 866     g_free(blkdev->type);
 867     blkdev->type = NULL;
 868     g_free(blkdev->dev);
 869     blkdev->dev = NULL;
 870     g_free(blkdev->devtype);
 871     blkdev->devtype = NULL;
 872     g_free(directiosafe);
 873     blkdev->directiosafe = false;
 874     return -1;
 875 }
 876
 877 static int blk_connect(struct XenDevice *xendev)
 878 {
 879     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
 880     int pers, index, qflags;
 881     bool readonly = true;
 882
 883     /* read-only ? */
 884     if (blkdev->directiosafe) {
 885         qflags = BDRV_O_NOCACHE | BDRV_O_NATIVE_AIO;
 886     } else {
 887         qflags = BDRV_O_CACHE_WB;
 888     }
 889     if (strcmp(blkdev->mode, "w") == 0) {
 890         qflags |= BDRV_O_RDWR;
 891         readonly = false;
 892     }
 893     if (blkdev->feature_discard) {
 894         qflags |= BDRV_O_UNMAP;
 895     }
 896
 897     /* init qemu block driver */
 898     index = (blkdev->xendev.dev - 202 * 256) / 16;
 899     blkdev->dinfo = drive_get(IF_XEN, 0, index);
 900     if (!blkdev->dinfo) {
 901         Error *local_err = NULL;
 902         QDict *options = NULL;
 903
 904         if (strcmp(blkdev->fileproto, "<unset>")) {
 905             options = qdict_new();
 906             qdict_put(options, "driver", qstring_from_str(blkdev->fileproto));
 907         }
 908
 909         /* setup via xenbus -> create new block driver instance */
 910         xen_be_printf(&blkdev->xendev, 2, "create new bdrv (xenbus setup)\n");
 911         blkdev->blk = blk_new_open(blkdev->dev, blkdev->filename, NULL, options,
 912                                    qflags, &local_err);
 913         if (!blkdev->blk) {
 914             xen_be_printf(&blkdev->xendev, 0, "error: %s\n",
 915                           error_get_pretty(local_err));
 916             error_free(local_err);
 917             return -1;
 918         }
 919     } else {
 920         /* setup via qemu cmdline -> already setup for us */
 921         xen_be_printf(&blkdev->xendev, 2, "get configured bdrv (cmdline setup)\n");
 922         blkdev->blk = blk_by_legacy_dinfo(blkdev->dinfo);
 923         if (blk_is_read_only(blkdev->blk) && !readonly) {
 924             xen_be_printf(&blkdev->xendev, 0, "Unexpected read-only drive");
 925             blkdev->blk = NULL;
 926             return -1;
 927         }
 928         /* blkdev->blk is not create by us, we get a reference
 929          * so we can blk_unref() unconditionally */
 930         blk_ref(blkdev->blk);
 931     }
 932     blk_attach_dev_nofail(blkdev->blk, blkdev);
 933     blkdev->file_size = blk_getlength(blkdev->blk);
 934     if (blkdev->file_size < 0) {
 935         xen_be_printf(&blkdev->xendev, 1, "blk_getlength: %d (%s) | drv %s\n",
 936                       (int)blkdev->file_size, strerror(-blkdev->file_size),
 937                       bdrv_get_format_name(blk_bs(blkdev->blk)) ?: "-");
 938         blkdev->file_size = 0;
 939     }
 940
 941     xen_be_printf(xendev, 1, "type \"%s\", fileproto \"%s\", filename \"%s\","
 942                   " size %" PRId64 " (%" PRId64 " MB)\n",
 943                   blkdev->type, blkdev->fileproto, blkdev->filename,
 944                   blkdev->file_size, blkdev->file_size >> 20);
 945
 946     /* Fill in number of sector size and number of sectors */
 947     xenstore_write_be_int(&blkdev->xendev, "sector-size", blkdev->file_blk);
 948     xenstore_write_be_int64(&blkdev->xendev, "sectors",
 949                             blkdev->file_size / blkdev->file_blk);
 950
 951     if (xenstore_read_fe_int(&blkdev->xendev, "ring-ref", &blkdev->ring_ref) == -1) {
 952         return -1;
 953     }
 954     if (xenstore_read_fe_int(&blkdev->xendev, "event-channel",
 955                              &blkdev->xendev.remote_port) == -1) {
 956         return -1;
 957     }
 958     if (xenstore_read_fe_int(&blkdev->xendev, "feature-persistent", &pers)) {
 959         blkdev->feature_persistent = FALSE;
 960     } else {
 961         blkdev->feature_persistent = !!pers;
 962     }
 963
 964     blkdev->protocol = BLKIF_PROTOCOL_NATIVE;
 965     if (blkdev->xendev.protocol) {
 966         if (strcmp(blkdev->xendev.protocol, XEN_IO_PROTO_ABI_X86_32) == 0) {
 967             blkdev->protocol = BLKIF_PROTOCOL_X86_32;
 968         }
 969         if (strcmp(blkdev->xendev.protocol, XEN_IO_PROTO_ABI_X86_64) == 0) {
 970             blkdev->protocol = BLKIF_PROTOCOL_X86_64;
 971         }
 972     }
 973
 974     blkdev->sring = xc_gnttab_map_grant_ref(blkdev->xendev.gnttabdev,
 975                                             blkdev->xendev.dom,
 976                                             blkdev->ring_ref,
 977                                             PROT_READ | PROT_WRITE);
 978     if (!blkdev->sring) {
 979         return -1;
 980     }
 981     blkdev->cnt_map++;
 982
 983     switch (blkdev->protocol) {
 984     case BLKIF_PROTOCOL_NATIVE:
 985     {
 986         blkif_sring_t *sring_native = blkdev->sring;
 987         BACK_RING_INIT(&blkdev->rings.native, sring_native, XC_PAGE_SIZE);
 988         break;
 989     }
 990     case BLKIF_PROTOCOL_X86_32:
 991     {
 992         blkif_x86_32_sring_t *sring_x86_32 = blkdev->sring;
 993
 994         BACK_RING_INIT(&blkdev->rings.x86_32_part, sring_x86_32, XC_PAGE_SIZE);
 995         break;
 996     }
 997     case BLKIF_PROTOCOL_X86_64:
 998     {
 999         blkif_x86_64_sring_t *sring_x86_64 = blkdev->sring;
1000
1001         BACK_RING_INIT(&blkdev->rings.x86_64_part, sring_x86_64, XC_PAGE_SIZE);
1002         break;
1003     }
1004     }
1005
1006     if (blkdev->feature_persistent) {
1007         /* Init persistent grants */
1008         blkdev->max_grants = max_requests * BLKIF_MAX_SEGMENTS_PER_REQUEST;
1009         blkdev->persistent_gnts = g_tree_new_full((GCompareDataFunc)int_cmp,
1010                                              NULL, NULL,
1011                                              batch_maps ?
1012                                              (GDestroyNotify)g_free :
1013                                              (GDestroyNotify)destroy_grant);
1014         blkdev->persistent_regions = NULL;
1015         blkdev->persistent_gnt_count = 0;
1016     }
1017
1018     xen_be_bind_evtchn(&blkdev->xendev);
1019
1020     xen_be_printf(&blkdev->xendev, 1, "ok: proto %s, ring-ref %d, "
1021                   "remote port %d, local port %d\n",
1022                   blkdev->xendev.protocol, blkdev->ring_ref,
1023                   blkdev->xendev.remote_port, blkdev->xendev.local_port);
1024     return 0;
1025 }
1026
1027 static void blk_disconnect(struct XenDevice *xendev)
1028 {
1029     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
1030
1031     if (blkdev->blk) {
1032         blk_detach_dev(blkdev->blk, blkdev);
1033         blk_unref(blkdev->blk);
1034         blkdev->blk = NULL;
1035     }
1036     xen_be_unbind_evtchn(&blkdev->xendev);
1037
1038     if (blkdev->sring) {
1039         xc_gnttab_munmap(blkdev->xendev.gnttabdev, blkdev->sring, 1);
1040         blkdev->cnt_map--;
1041         blkdev->sring = NULL;
1042     }
1043
1044     /*
1045      * Unmap persistent grants before switching to the closed state
1046      * so the frontend can free them.
1047      *
1048      * In the !batch_maps case g_tree_destroy will take care of unmapping
1049      * the grant, but in the batch_maps case we need to iterate over every
1050      * region in persistent_regions and unmap it.
1051      */
1052     if (blkdev->feature_persistent) {
1053         g_tree_destroy(blkdev->persistent_gnts);
1054         assert(batch_maps || blkdev->persistent_gnt_count == 0);
1055         if (batch_maps) {
1056             blkdev->persistent_gnt_count = 0;
1057             g_slist_foreach(blkdev->persistent_regions,
1058                             (GFunc)remove_persistent_region, blkdev);
1059             g_slist_free(blkdev->persistent_regions);
1060         }
1061         blkdev->feature_persistent = false;
1062     }
1063 }
1064
1065 static int blk_free(struct XenDevice *xendev)
1066 {
1067     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
1068     struct ioreq *ioreq;
1069
1070     if (blkdev->blk || blkdev->sring) {
1071         blk_disconnect(xendev);
1072     }
1073
1074     while (!QLIST_EMPTY(&blkdev->freelist)) {
1075         ioreq = QLIST_FIRST(&blkdev->freelist);
1076         QLIST_REMOVE(ioreq, list);
1077         qemu_iovec_destroy(&ioreq->v);
1078         g_free(ioreq);
1079     }
1080
1081     g_free(blkdev->params);
1082     g_free(blkdev->mode);
1083     g_free(blkdev->type);
1084     g_free(blkdev->dev);
1085     g_free(blkdev->devtype);
1086     qemu_bh_delete(blkdev->bh);
1087     return 0;
1088 }
1089
1090 static void blk_event(struct XenDevice *xendev)
1091 {
1092     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
1093
1094     qemu_bh_schedule(blkdev->bh);
1095 }
1096
1097 struct XenDevOps xen_blkdev_ops = {
1098     .size       = sizeof(struct XenBlkDev),
1099     .flags      = DEVOPS_FLAG_NEED_GNTDEV,
1100     .alloc      = blk_alloc,
1101     .init       = blk_init,
1102     .initialise    = blk_connect,
1103     .disconnect = blk_disconnect,
1104     .event      = blk_event,
1105     .free       = blk_free,
1106 };