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