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