hw/block/xen_disk.c

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