hw/block/dataplane/xen-block.c

   1 /*
   2  * Copyright (c) 2018  Citrix Systems Inc.
   3  * (c) Gerd Hoffmann <kraxel@redhat.com>
   4  *
   5  * This program is free software; you can redistribute it and/or modify
   6  * it under the terms of the GNU General Public License as published by
   7  * the Free Software Foundation; under version 2 of the License.
   8  *
   9  * This program is distributed in the hope that it will be useful,
  10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  12  * GNU General Public License for more details.
  13  *
  14  * You should have received a copy of the GNU General Public License along
  15  * with this program; if not, see <http://www.gnu.org/licenses/>.
  16  *
  17  * Contributions after 2012-01-13 are licensed under the terms of the
  18  * GNU GPL, version 2 or (at your option) any later version.
  19  */
  20
  21 struct ioreq {
  22     blkif_request_t     req;
  23     int16_t             status;
  24
  25     /* parsed request */
  26     off_t               start;
  27     QEMUIOVector        v;
  28     void                *buf;
  29     size_t              size;
  30     int                 presync;
  31
  32     /* aio status */
  33     int                 aio_inflight;
  34     int                 aio_errors;
  35
  36     struct XenBlkDev    *blkdev;
  37     QLIST_ENTRY(ioreq)   list;
  38     BlockAcctCookie     acct;
  39 };
  40
  41 #define MAX_RING_PAGE_ORDER 4
  42
  43 struct XenBlkDev {
  44     struct XenLegacyDevice    xendev;  /* must be first */
  45     char                *params;
  46     char                *mode;
  47     char                *type;
  48     char                *dev;
  49     char                *devtype;
  50     bool                directiosafe;
  51     const char          *fileproto;
  52     const char          *filename;
  53     unsigned int        ring_ref[1 << MAX_RING_PAGE_ORDER];
  54     unsigned int        nr_ring_ref;
  55     void                *sring;
  56     int64_t             file_blk;
  57     int64_t             file_size;
  58     int                 protocol;
  59     blkif_back_rings_t  rings;
  60     int                 more_work;
  61
  62     /* request lists */
  63     QLIST_HEAD(inflight_head, ioreq) inflight;
  64     QLIST_HEAD(finished_head, ioreq) finished;
  65     QLIST_HEAD(freelist_head, ioreq) freelist;
  66     int                 requests_total;
  67     int                 requests_inflight;
  68     int                 requests_finished;
  69     unsigned int        max_requests;
  70
  71     gboolean            feature_discard;
  72
  73     /* qemu block driver */
  74     DriveInfo           *dinfo;
  75     BlockBackend        *blk;
  76     QEMUBH              *bh;
  77
  78     IOThread            *iothread;
  79     AioContext          *ctx;
  80 };
  81
  82 static void ioreq_reset(struct ioreq *ioreq)
  83 {
  84     memset(&ioreq->req, 0, sizeof(ioreq->req));
  85     ioreq->status = 0;
  86     ioreq->start = 0;
  87     ioreq->buf = NULL;
  88     ioreq->size = 0;
  89     ioreq->presync = 0;
  90
  91     ioreq->aio_inflight = 0;
  92     ioreq->aio_errors = 0;
  93
  94     ioreq->blkdev = NULL;
  95     memset(&ioreq->list, 0, sizeof(ioreq->list));
  96     memset(&ioreq->acct, 0, sizeof(ioreq->acct));
  97
  98     qemu_iovec_reset(&ioreq->v);
  99 }
 100
 101 static struct ioreq *ioreq_start(struct XenBlkDev *blkdev)
 102 {
 103     struct ioreq *ioreq = NULL;
 104
 105     if (QLIST_EMPTY(&blkdev->freelist)) {
 106         if (blkdev->requests_total >= blkdev->max_requests) {
 107             goto out;
 108         }
 109         /* allocate new struct */
 110         ioreq = g_malloc0(sizeof(*ioreq));
 111         ioreq->blkdev = blkdev;
 112         blkdev->requests_total++;
 113         qemu_iovec_init(&ioreq->v, 1);
 114     } else {
 115         /* get one from freelist */
 116         ioreq = QLIST_FIRST(&blkdev->freelist);
 117         QLIST_REMOVE(ioreq, list);
 118     }
 119     QLIST_INSERT_HEAD(&blkdev->inflight, ioreq, list);
 120     blkdev->requests_inflight++;
 121
 122 out:
 123     return ioreq;
 124 }
 125
 126 static void ioreq_finish(struct ioreq *ioreq)
 127 {
 128     struct XenBlkDev *blkdev = ioreq->blkdev;
 129
 130     QLIST_REMOVE(ioreq, list);
 131     QLIST_INSERT_HEAD(&blkdev->finished, ioreq, list);
 132     blkdev->requests_inflight--;
 133     blkdev->requests_finished++;
 134 }
 135
 136 static void ioreq_release(struct ioreq *ioreq, bool finish)
 137 {
 138     struct XenBlkDev *blkdev = ioreq->blkdev;
 139
 140     QLIST_REMOVE(ioreq, list);
 141     ioreq_reset(ioreq);
 142     ioreq->blkdev = blkdev;
 143     QLIST_INSERT_HEAD(&blkdev->freelist, ioreq, list);
 144     if (finish) {
 145         blkdev->requests_finished--;
 146     } else {
 147         blkdev->requests_inflight--;
 148     }
 149 }
 150
 151 /*
 152  * translate request into iovec + start offset
 153  * do sanity checks along the way
 154  */
 155 static int ioreq_parse(struct ioreq *ioreq)
 156 {
 157     struct XenBlkDev *blkdev = ioreq->blkdev;
 158     struct XenLegacyDevice *xendev = &blkdev->xendev;
 159     size_t len;
 160     int i;
 161
 162     switch (ioreq->req.operation) {
 163     case BLKIF_OP_READ:
 164         break;
 165     case BLKIF_OP_FLUSH_DISKCACHE:
 166         ioreq->presync = 1;
 167         if (!ioreq->req.nr_segments) {
 168             return 0;
 169         }
 170         /* fall through */
 171     case BLKIF_OP_WRITE:
 172         break;
 173     case BLKIF_OP_DISCARD:
 174         return 0;
 175     default:
 176         error_report("error: unknown operation (%d)", ioreq->req.operation);
 177         goto err;
 178     };
 179
 180     if (ioreq->req.operation != BLKIF_OP_READ && blkdev->mode[0] != 'w') {
 181         error_report("error: write req for ro device");
 182         goto err;
 183     }
 184
 185     ioreq->start = ioreq->req.sector_number * blkdev->file_blk;
 186     for (i = 0; i < ioreq->req.nr_segments; i++) {
 187         if (i == BLKIF_MAX_SEGMENTS_PER_REQUEST) {
 188             error_report("error: nr_segments too big");
 189             goto err;
 190         }
 191         if (ioreq->req.seg[i].first_sect > ioreq->req.seg[i].last_sect) {
 192             error_report("error: first > last sector");
 193             goto err;
 194         }
 195         if (ioreq->req.seg[i].last_sect * BLOCK_SIZE >= XC_PAGE_SIZE) {
 196             error_report("error: page crossing");
 197             goto err;
 198         }
 199
 200         len = (ioreq->req.seg[i].last_sect -
 201                ioreq->req.seg[i].first_sect + 1) * blkdev->file_blk;
 202         ioreq->size += len;
 203     }
 204     if (ioreq->start + ioreq->size > blkdev->file_size) {
 205         error_report("error: access beyond end of file");
 206         goto err;
 207     }
 208     return 0;
 209
 210 err:
 211     ioreq->status = BLKIF_RSP_ERROR;
 212     return -1;
 213 }
 214
 215 static int ioreq_grant_copy(struct ioreq *ioreq)
 216 {
 217     struct XenBlkDev *blkdev = ioreq->blkdev;
 218     struct XenLegacyDevice *xendev = &blkdev->xendev;
 219     XenGrantCopySegment segs[BLKIF_MAX_SEGMENTS_PER_REQUEST];
 220     int i, count, rc;
 221     int64_t file_blk = blkdev->file_blk;
 222     bool to_domain = (ioreq->req.operation == BLKIF_OP_READ);
 223     void *virt = ioreq->buf;
 224
 225     if (ioreq->req.nr_segments == 0) {
 226         return 0;
 227     }
 228
 229     count = ioreq->req.nr_segments;
 230
 231     for (i = 0; i < count; i++) {
 232         if (to_domain) {
 233             segs[i].dest.foreign.ref = ioreq->req.seg[i].gref;
 234             segs[i].dest.foreign.offset = ioreq->req.seg[i].first_sect *
 235                 file_blk;
 236             segs[i].source.virt = virt;
 237         } else {
 238             segs[i].source.foreign.ref = ioreq->req.seg[i].gref;
 239             segs[i].source.foreign.offset = ioreq->req.seg[i].first_sect *
 240                 file_blk;
 241             segs[i].dest.virt = virt;
 242         }
 243         segs[i].len = (ioreq->req.seg[i].last_sect
 244                        - ioreq->req.seg[i].first_sect + 1) * file_blk;
 245         virt += segs[i].len;
 246     }
 247
 248     rc = xen_be_copy_grant_refs(xendev, to_domain, segs, count);
 249
 250     if (rc) {
 251         error_report("failed to copy data %d", rc);
 252         ioreq->aio_errors++;
 253         return -1;
 254     }
 255
 256     return rc;
 257 }
 258
 259 static int ioreq_runio_qemu_aio(struct ioreq *ioreq);
 260
 261 static void qemu_aio_complete(void *opaque, int ret)
 262 {
 263     struct ioreq *ioreq = opaque;
 264     struct XenBlkDev *blkdev = ioreq->blkdev;
 265     struct XenLegacyDevice *xendev = &blkdev->xendev;
 266
 267     aio_context_acquire(blkdev->ctx);
 268
 269     if (ret != 0) {
 270         error_report("%s I/O error",
 271                      ioreq->req.operation == BLKIF_OP_READ ?
 272                      "read" : "write");
 273         ioreq->aio_errors++;
 274     }
 275
 276     ioreq->aio_inflight--;
 277     if (ioreq->presync) {
 278         ioreq->presync = 0;
 279         ioreq_runio_qemu_aio(ioreq);
 280         goto done;
 281     }
 282     if (ioreq->aio_inflight > 0) {
 283         goto done;
 284     }
 285
 286     switch (ioreq->req.operation) {
 287     case BLKIF_OP_READ:
 288         /* in case of failure ioreq->aio_errors is increased */
 289         if (ret == 0) {
 290             ioreq_grant_copy(ioreq);
 291         }
 292         qemu_vfree(ioreq->buf);
 293         break;
 294     case BLKIF_OP_WRITE:
 295     case BLKIF_OP_FLUSH_DISKCACHE:
 296         if (!ioreq->req.nr_segments) {
 297             break;
 298         }
 299         qemu_vfree(ioreq->buf);
 300         break;
 301     default:
 302         break;
 303     }
 304
 305     ioreq->status = ioreq->aio_errors ? BLKIF_RSP_ERROR : BLKIF_RSP_OKAY;
 306     ioreq_finish(ioreq);
 307
 308     switch (ioreq->req.operation) {
 309     case BLKIF_OP_WRITE:
 310     case BLKIF_OP_FLUSH_DISKCACHE:
 311         if (!ioreq->req.nr_segments) {
 312             break;
 313         }
 314     case BLKIF_OP_READ:
 315         if (ioreq->status == BLKIF_RSP_OKAY) {
 316             block_acct_done(blk_get_stats(blkdev->blk), &ioreq->acct);
 317         } else {
 318             block_acct_failed(blk_get_stats(blkdev->blk), &ioreq->acct);
 319         }
 320         break;
 321     case BLKIF_OP_DISCARD:
 322     default:
 323         break;
 324     }
 325     qemu_bh_schedule(blkdev->bh);
 326
 327 done:
 328     aio_context_release(blkdev->ctx);
 329 }
 330
 331 static bool blk_split_discard(struct ioreq *ioreq, blkif_sector_t sector_number,
 332                               uint64_t nr_sectors)
 333 {
 334     struct XenBlkDev *blkdev = ioreq->blkdev;
 335     int64_t byte_offset;
 336     int byte_chunk;
 337     uint64_t byte_remaining, limit;
 338     uint64_t sec_start = sector_number;
 339     uint64_t sec_count = nr_sectors;
 340
 341     /* Wrap around, or overflowing byte limit? */
 342     if (sec_start + sec_count < sec_count ||
 343         sec_start + sec_count > INT64_MAX >> BDRV_SECTOR_BITS) {
 344         return false;
 345     }
 346
 347     limit = BDRV_REQUEST_MAX_SECTORS << BDRV_SECTOR_BITS;
 348     byte_offset = sec_start << BDRV_SECTOR_BITS;
 349     byte_remaining = sec_count << BDRV_SECTOR_BITS;
 350
 351     do {
 352         byte_chunk = byte_remaining > limit ? limit : byte_remaining;
 353         ioreq->aio_inflight++;
 354         blk_aio_pdiscard(blkdev->blk, byte_offset, byte_chunk,
 355                          qemu_aio_complete, ioreq);
 356         byte_remaining -= byte_chunk;
 357         byte_offset += byte_chunk;
 358     } while (byte_remaining > 0);
 359
 360     return true;
 361 }
 362
 363 static int ioreq_runio_qemu_aio(struct ioreq *ioreq)
 364 {
 365     struct XenBlkDev *blkdev = ioreq->blkdev;
 366
 367     ioreq->buf = qemu_memalign(XC_PAGE_SIZE, ioreq->size);
 368     if (ioreq->req.nr_segments &&
 369         (ioreq->req.operation == BLKIF_OP_WRITE ||
 370          ioreq->req.operation == BLKIF_OP_FLUSH_DISKCACHE) &&
 371         ioreq_grant_copy(ioreq)) {
 372         qemu_vfree(ioreq->buf);
 373         goto err;
 374     }
 375
 376     ioreq->aio_inflight++;
 377     if (ioreq->presync) {
 378         blk_aio_flush(ioreq->blkdev->blk, qemu_aio_complete, ioreq);
 379         return 0;
 380     }
 381
 382     switch (ioreq->req.operation) {
 383     case BLKIF_OP_READ:
 384         qemu_iovec_add(&ioreq->v, ioreq->buf, ioreq->size);
 385         block_acct_start(blk_get_stats(blkdev->blk), &ioreq->acct,
 386                          ioreq->v.size, BLOCK_ACCT_READ);
 387         ioreq->aio_inflight++;
 388         blk_aio_preadv(blkdev->blk, ioreq->start, &ioreq->v, 0,
 389                        qemu_aio_complete, ioreq);
 390         break;
 391     case BLKIF_OP_WRITE:
 392     case BLKIF_OP_FLUSH_DISKCACHE:
 393         if (!ioreq->req.nr_segments) {
 394             break;
 395         }
 396
 397         qemu_iovec_add(&ioreq->v, ioreq->buf, ioreq->size);
 398         block_acct_start(blk_get_stats(blkdev->blk), &ioreq->acct,
 399                          ioreq->v.size,
 400                          ioreq->req.operation == BLKIF_OP_WRITE ?
 401                          BLOCK_ACCT_WRITE : BLOCK_ACCT_FLUSH);
 402         ioreq->aio_inflight++;
 403         blk_aio_pwritev(blkdev->blk, ioreq->start, &ioreq->v, 0,
 404                         qemu_aio_complete, ioreq);
 405         break;
 406     case BLKIF_OP_DISCARD:
 407     {
 408         struct blkif_request_discard *req = (void *)&ioreq->req;
 409         if (!blk_split_discard(ioreq, req->sector_number, req->nr_sectors)) {
 410             goto err;
 411         }
 412         break;
 413     }
 414     default:
 415         /* unknown operation (shouldn't happen -- parse catches this) */
 416         goto err;
 417     }
 418
 419     qemu_aio_complete(ioreq, 0);
 420
 421     return 0;
 422
 423 err:
 424     ioreq_finish(ioreq);
 425     ioreq->status = BLKIF_RSP_ERROR;
 426     return -1;
 427 }
 428
 429 static int blk_send_response_one(struct ioreq *ioreq)
 430 {
 431     struct XenBlkDev  *blkdev = ioreq->blkdev;
 432     int               send_notify   = 0;
 433     int               have_requests = 0;
 434     blkif_response_t  *resp;
 435
 436     /* Place on the response ring for the relevant domain. */
 437     switch (blkdev->protocol) {
 438     case BLKIF_PROTOCOL_NATIVE:
 439         resp = (blkif_response_t *)RING_GET_RESPONSE(
 440             &blkdev->rings.native,
 441             blkdev->rings.native.rsp_prod_pvt);
 442         break;
 443     case BLKIF_PROTOCOL_X86_32:
 444         resp = (blkif_response_t *)RING_GET_RESPONSE(
 445             &blkdev->rings.x86_32_part,
 446             blkdev->rings.x86_32_part.rsp_prod_pvt);
 447         break;
 448     case BLKIF_PROTOCOL_X86_64:
 449         resp = (blkif_response_t *)RING_GET_RESPONSE(
 450             &blkdev->rings.x86_64_part,
 451             blkdev->rings.x86_64_part.rsp_prod_pvt);
 452         break;
 453     default:
 454         return 0;
 455     }
 456
 457     resp->id        = ioreq->req.id;
 458     resp->operation = ioreq->req.operation;
 459     resp->status    = ioreq->status;
 460
 461     blkdev->rings.common.rsp_prod_pvt++;
 462
 463     RING_PUSH_RESPONSES_AND_CHECK_NOTIFY(&blkdev->rings.common, send_notify);
 464     if (blkdev->rings.common.rsp_prod_pvt == blkdev->rings.common.req_cons) {
 465         /*
 466          * Tail check for pending requests. Allows frontend to avoid
 467          * notifications if requests are already in flight (lower
 468          * overheads and promotes batching).
 469          */
 470         RING_FINAL_CHECK_FOR_REQUESTS(&blkdev->rings.common, have_requests);
 471     } else if (RING_HAS_UNCONSUMED_REQUESTS(&blkdev->rings.common)) {
 472         have_requests = 1;
 473     }
 474
 475     if (have_requests) {
 476         blkdev->more_work++;
 477     }
 478     return send_notify;
 479 }
 480
 481 /* walk finished list, send outstanding responses, free requests */
 482 static void blk_send_response_all(struct XenBlkDev *blkdev)
 483 {
 484     struct ioreq *ioreq;
 485     int send_notify = 0;
 486
 487     while (!QLIST_EMPTY(&blkdev->finished)) {
 488         ioreq = QLIST_FIRST(&blkdev->finished);
 489         send_notify += blk_send_response_one(ioreq);
 490         ioreq_release(ioreq, true);
 491     }
 492     if (send_notify) {
 493         xen_pv_send_notify(&blkdev->xendev);
 494     }
 495 }
 496
 497 static int blk_get_request(struct XenBlkDev *blkdev, struct ioreq *ioreq,
 498                            RING_IDX rc)
 499 {
 500     switch (blkdev->protocol) {
 501     case BLKIF_PROTOCOL_NATIVE:
 502         memcpy(&ioreq->req, RING_GET_REQUEST(&blkdev->rings.native, rc),
 503                sizeof(ioreq->req));
 504         break;
 505     case BLKIF_PROTOCOL_X86_32:
 506         blkif_get_x86_32_req(&ioreq->req,
 507                              RING_GET_REQUEST(&blkdev->rings.x86_32_part, rc));
 508         break;
 509     case BLKIF_PROTOCOL_X86_64:
 510         blkif_get_x86_64_req(&ioreq->req,
 511                              RING_GET_REQUEST(&blkdev->rings.x86_64_part, rc));
 512         break;
 513     }
 514     /* Prevent the compiler from accessing the on-ring fields instead. */
 515     barrier();
 516     return 0;
 517 }
 518
 519 static void blk_handle_requests(struct XenBlkDev *blkdev)
 520 {
 521     RING_IDX rc, rp;
 522     struct ioreq *ioreq;
 523
 524     blkdev->more_work = 0;
 525
 526     rc = blkdev->rings.common.req_cons;
 527     rp = blkdev->rings.common.sring->req_prod;
 528     xen_rmb(); /* Ensure we see queued requests up to 'rp'. */
 529
 530     blk_send_response_all(blkdev);
 531     while (rc != rp) {
 532         /* pull request from ring */
 533         if (RING_REQUEST_CONS_OVERFLOW(&blkdev->rings.common, rc)) {
 534             break;
 535         }
 536         ioreq = ioreq_start(blkdev);
 537         if (ioreq == NULL) {
 538             blkdev->more_work++;
 539             break;
 540         }
 541         blk_get_request(blkdev, ioreq, rc);
 542         blkdev->rings.common.req_cons = ++rc;
 543
 544         /* parse them */
 545         if (ioreq_parse(ioreq) != 0) {
 546
 547             switch (ioreq->req.operation) {
 548             case BLKIF_OP_READ:
 549                 block_acct_invalid(blk_get_stats(blkdev->blk),
 550                                    BLOCK_ACCT_READ);
 551                 break;
 552             case BLKIF_OP_WRITE:
 553                 block_acct_invalid(blk_get_stats(blkdev->blk),
 554                                    BLOCK_ACCT_WRITE);
 555                 break;
 556             case BLKIF_OP_FLUSH_DISKCACHE:
 557                 block_acct_invalid(blk_get_stats(blkdev->blk),
 558                                    BLOCK_ACCT_FLUSH);
 559             default:
 560                 break;
 561             };
 562
 563             if (blk_send_response_one(ioreq)) {
 564                 xen_pv_send_notify(&blkdev->xendev);
 565             }
 566             ioreq_release(ioreq, false);
 567             continue;
 568         }
 569
 570         ioreq_runio_qemu_aio(ioreq);
 571     }
 572
 573     if (blkdev->more_work && blkdev->requests_inflight < blkdev->max_requests) {
 574         qemu_bh_schedule(blkdev->bh);
 575     }
 576 }
 577
 578 static void blk_bh(void *opaque)
 579 {
 580     struct XenBlkDev *blkdev = opaque;
 581
 582     aio_context_acquire(blkdev->ctx);
 583     blk_handle_requests(blkdev);
 584     aio_context_release(blkdev->ctx);
 585 }
 586
 587 static void blk_alloc(struct XenLegacyDevice *xendev)
 588 {
 589     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
 590     Error *err = NULL;
 591
 592     trace_xen_disk_alloc(xendev->name);
 593
 594     QLIST_INIT(&blkdev->inflight);
 595     QLIST_INIT(&blkdev->finished);
 596     QLIST_INIT(&blkdev->freelist);
 597
 598     blkdev->iothread = iothread_create(xendev->name, &err);
 599     assert(!err);
 600
 601     blkdev->ctx = iothread_get_aio_context(blkdev->iothread);
 602     blkdev->bh = aio_bh_new(blkdev->ctx, blk_bh, blkdev);
 603 }
 604
 605 static int blk_free(struct XenLegacyDevice *xendev)
 606 {
 607     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
 608     struct ioreq *ioreq;
 609
 610     trace_xen_disk_free(xendev->name);
 611
 612     blk_disconnect(xendev);
 613
 614     while (!QLIST_EMPTY(&blkdev->freelist)) {
 615         ioreq = QLIST_FIRST(&blkdev->freelist);
 616         QLIST_REMOVE(ioreq, list);
 617         qemu_iovec_destroy(&ioreq->v);
 618         g_free(ioreq);
 619     }
 620
 621     g_free(blkdev->params);
 622     g_free(blkdev->mode);
 623     g_free(blkdev->type);
 624     g_free(blkdev->dev);
 625     g_free(blkdev->devtype);
 626     qemu_bh_delete(blkdev->bh);
 627     iothread_destroy(blkdev->iothread);
 628     return 0;
 629 }
 630
 631 static void blk_event(struct XenLegacyDevice *xendev)
 632 {
 633     struct XenBlkDev *blkdev = container_of(xendev, struct XenBlkDev, xendev);
 634
 635     qemu_bh_schedule(blkdev->bh);
 636 }