block/blkdebug.c

   1 /*
   2  * Block protocol for I/O error injection
   3  *
   4  * Copyright (C) 2016-2017 Red Hat, Inc.
   5  * Copyright (c) 2010 Kevin Wolf <kwolf@redhat.com>
   6  *
   7  * Permission is hereby granted, free of charge, to any person obtaining a copy
   8  * of this software and associated documentation files (the "Software"), to deal
   9  * in the Software without restriction, including without limitation the rights
  10  * to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
  11  * copies of the Software, and to permit persons to whom the Software is
  12  * furnished to do so, subject to the following conditions:
  13  *
  14  * The above copyright notice and this permission notice shall be included in
  15  * all copies or substantial portions of the Software.
  16  *
  17  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  18  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  19  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL
  20  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  21  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  22  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN
  23  * THE SOFTWARE.
  24  */
  25
  26 #include "qemu/osdep.h"
  27 #include "qapi/error.h"
  28 #include "qemu/cutils.h"
  29 #include "qemu/config-file.h"
  30 #include "block/block_int.h"
  31 #include "block/qdict.h"
  32 #include "qemu/module.h"
  33 #include "qemu/option.h"
  34 #include "qapi/qapi-visit-block-core.h"
  35 #include "qapi/qmp/qdict.h"
  36 #include "qapi/qmp/qlist.h"
  37 #include "qapi/qmp/qstring.h"
  38 #include "qapi/qobject-input-visitor.h"
  39 #include "sysemu/qtest.h"
  40
  41 typedef struct BDRVBlkdebugState {
  42     int state;
  43     int new_state;
  44     uint64_t align;
  45     uint64_t max_transfer;
  46     uint64_t opt_write_zero;
  47     uint64_t max_write_zero;
  48     uint64_t opt_discard;
  49     uint64_t max_discard;
  50
  51     uint64_t take_child_perms;
  52     uint64_t unshare_child_perms;
  53
  54     /* For blkdebug_refresh_filename() */
  55     char *config_file;
  56
  57     QLIST_HEAD(, BlkdebugRule) rules[BLKDBG__MAX];
  58     QSIMPLEQ_HEAD(, BlkdebugRule) active_rules;
  59     QLIST_HEAD(, BlkdebugSuspendedReq) suspended_reqs;
  60 } BDRVBlkdebugState;
  61
  62 typedef struct BlkdebugAIOCB {
  63     BlockAIOCB common;
  64     int ret;
  65 } BlkdebugAIOCB;
  66
  67 typedef struct BlkdebugSuspendedReq {
  68     Coroutine *co;
  69     char *tag;
  70     QLIST_ENTRY(BlkdebugSuspendedReq) next;
  71 } BlkdebugSuspendedReq;
  72
  73 enum {
  74     ACTION_INJECT_ERROR,
  75     ACTION_SET_STATE,
  76     ACTION_SUSPEND,
  77 };
  78
  79 typedef struct BlkdebugRule {
  80     BlkdebugEvent event;
  81     int action;
  82     int state;
  83     union {
  84         struct {
  85             uint64_t iotype_mask;
  86             int error;
  87             int immediately;
  88             int once;
  89             int64_t offset;
  90         } inject;
  91         struct {
  92             int new_state;
  93         } set_state;
  94         struct {
  95             char *tag;
  96         } suspend;
  97     } options;
  98     QLIST_ENTRY(BlkdebugRule) next;
  99     QSIMPLEQ_ENTRY(BlkdebugRule) active_next;
 100 } BlkdebugRule;
 101
 102 QEMU_BUILD_BUG_MSG(BLKDEBUG_IO_TYPE__MAX > 64,
 103                    "BlkdebugIOType mask does not fit into an uint64_t");
 104
 105 static QemuOptsList inject_error_opts = {
 106     .name = "inject-error",
 107     .head = QTAILQ_HEAD_INITIALIZER(inject_error_opts.head),
 108     .desc = {
 109         {
 110             .name = "event",
 111             .type = QEMU_OPT_STRING,
 112         },
 113         {
 114             .name = "state",
 115             .type = QEMU_OPT_NUMBER,
 116         },
 117         {
 118             .name = "iotype",
 119             .type = QEMU_OPT_STRING,
 120         },
 121         {
 122             .name = "errno",
 123             .type = QEMU_OPT_NUMBER,
 124         },
 125         {
 126             .name = "sector",
 127             .type = QEMU_OPT_NUMBER,
 128         },
 129         {
 130             .name = "once",
 131             .type = QEMU_OPT_BOOL,
 132         },
 133         {
 134             .name = "immediately",
 135             .type = QEMU_OPT_BOOL,
 136         },
 137         { /* end of list */ }
 138     },
 139 };
 140
 141 static QemuOptsList set_state_opts = {
 142     .name = "set-state",
 143     .head = QTAILQ_HEAD_INITIALIZER(set_state_opts.head),
 144     .desc = {
 145         {
 146             .name = "event",
 147             .type = QEMU_OPT_STRING,
 148         },
 149         {
 150             .name = "state",
 151             .type = QEMU_OPT_NUMBER,
 152         },
 153         {
 154             .name = "new_state",
 155             .type = QEMU_OPT_NUMBER,
 156         },
 157         { /* end of list */ }
 158     },
 159 };
 160
 161 static QemuOptsList *config_groups[] = {
 162     &inject_error_opts,
 163     &set_state_opts,
 164     NULL
 165 };
 166
 167 struct add_rule_data {
 168     BDRVBlkdebugState *s;
 169     int action;
 170 };
 171
 172 static int add_rule(void *opaque, QemuOpts *opts, Error **errp)
 173 {
 174     struct add_rule_data *d = opaque;
 175     BDRVBlkdebugState *s = d->s;
 176     const char *event_name;
 177     int event;
 178     struct BlkdebugRule *rule;
 179     int64_t sector;
 180     BlkdebugIOType iotype;
 181     Error *local_error = NULL;
 182
 183     /* Find the right event for the rule */
 184     event_name = qemu_opt_get(opts, "event");
 185     if (!event_name) {
 186         error_setg(errp, "Missing event name for rule");
 187         return -1;
 188     }
 189     event = qapi_enum_parse(&BlkdebugEvent_lookup, event_name, -1, errp);
 190     if (event < 0) {
 191         return -1;
 192     }
 193
 194     /* Set attributes common for all actions */
 195     rule = g_malloc0(sizeof(*rule));
 196     *rule = (struct BlkdebugRule) {
 197         .event  = event,
 198         .action = d->action,
 199         .state  = qemu_opt_get_number(opts, "state", 0),
 200     };
 201
 202     /* Parse action-specific options */
 203     switch (d->action) {
 204     case ACTION_INJECT_ERROR:
 205         rule->options.inject.error = qemu_opt_get_number(opts, "errno", EIO);
 206         rule->options.inject.once  = qemu_opt_get_bool(opts, "once", 0);
 207         rule->options.inject.immediately =
 208             qemu_opt_get_bool(opts, "immediately", 0);
 209         sector = qemu_opt_get_number(opts, "sector", -1);
 210         rule->options.inject.offset =
 211             sector == -1 ? -1 : sector * BDRV_SECTOR_SIZE;
 212
 213         iotype = qapi_enum_parse(&BlkdebugIOType_lookup,
 214                                  qemu_opt_get(opts, "iotype"),
 215                                  BLKDEBUG_IO_TYPE__MAX, &local_error);
 216         if (local_error) {
 217             error_propagate(errp, local_error);
 218             g_free(rule);
 219             return -1;
 220         }
 221         if (iotype != BLKDEBUG_IO_TYPE__MAX) {
 222             rule->options.inject.iotype_mask = (1ull << iotype);
 223         } else {
 224             /* Apply the default */
 225             rule->options.inject.iotype_mask =
 226                 (1ull << BLKDEBUG_IO_TYPE_READ)
 227                 | (1ull << BLKDEBUG_IO_TYPE_WRITE)
 228                 | (1ull << BLKDEBUG_IO_TYPE_WRITE_ZEROES)
 229                 | (1ull << BLKDEBUG_IO_TYPE_DISCARD)
 230                 | (1ull << BLKDEBUG_IO_TYPE_FLUSH);
 231         }
 232
 233         break;
 234
 235     case ACTION_SET_STATE:
 236         rule->options.set_state.new_state =
 237             qemu_opt_get_number(opts, "new_state", 0);
 238         break;
 239
 240     case ACTION_SUSPEND:
 241         rule->options.suspend.tag =
 242             g_strdup(qemu_opt_get(opts, "tag"));
 243         break;
 244     };
 245
 246     /* Add the rule */
 247     QLIST_INSERT_HEAD(&s->rules[event], rule, next);
 248
 249     return 0;
 250 }
 251
 252 static void remove_rule(BlkdebugRule *rule)
 253 {
 254     switch (rule->action) {
 255     case ACTION_INJECT_ERROR:
 256     case ACTION_SET_STATE:
 257         break;
 258     case ACTION_SUSPEND:
 259         g_free(rule->options.suspend.tag);
 260         break;
 261     }
 262
 263     QLIST_REMOVE(rule, next);
 264     g_free(rule);
 265 }
 266
 267 static int read_config(BDRVBlkdebugState *s, const char *filename,
 268                        QDict *options, Error **errp)
 269 {
 270     FILE *f = NULL;
 271     int ret;
 272     struct add_rule_data d;
 273     Error *local_err = NULL;
 274
 275     if (filename) {
 276         f = fopen(filename, "r");
 277         if (f == NULL) {
 278             error_setg_errno(errp, errno, "Could not read blkdebug config file");
 279             return -errno;
 280         }
 281
 282         ret = qemu_config_parse(f, config_groups, filename, errp);
 283         if (ret < 0) {
 284             goto fail;
 285         }
 286     }
 287
 288     qemu_config_parse_qdict(options, config_groups, &local_err);
 289     if (local_err) {
 290         error_propagate(errp, local_err);
 291         ret = -EINVAL;
 292         goto fail;
 293     }
 294
 295     d.s = s;
 296     d.action = ACTION_INJECT_ERROR;
 297     qemu_opts_foreach(&inject_error_opts, add_rule, &d, &local_err);
 298     if (local_err) {
 299         error_propagate(errp, local_err);
 300         ret = -EINVAL;
 301         goto fail;
 302     }
 303
 304     d.action = ACTION_SET_STATE;
 305     qemu_opts_foreach(&set_state_opts, add_rule, &d, &local_err);
 306     if (local_err) {
 307         error_propagate(errp, local_err);
 308         ret = -EINVAL;
 309         goto fail;
 310     }
 311
 312     ret = 0;
 313 fail:
 314     qemu_opts_reset(&inject_error_opts);
 315     qemu_opts_reset(&set_state_opts);
 316     if (f) {
 317         fclose(f);
 318     }
 319     return ret;
 320 }
 321
 322 /* Valid blkdebug filenames look like blkdebug:path/to/config:path/to/image */
 323 static void blkdebug_parse_filename(const char *filename, QDict *options,
 324                                     Error **errp)
 325 {
 326     const char *c;
 327
 328     /* Parse the blkdebug: prefix */
 329     if (!strstart(filename, "blkdebug:", &filename)) {
 330         /* There was no prefix; therefore, all options have to be already
 331            present in the QDict (except for the filename) */
 332         qdict_put_str(options, "x-image", filename);
 333         return;
 334     }
 335
 336     /* Parse config file path */
 337     c = strchr(filename, ':');
 338     if (c == NULL) {
 339         error_setg(errp, "blkdebug requires both config file and image path");
 340         return;
 341     }
 342
 343     if (c != filename) {
 344         QString *config_path;
 345         config_path = qstring_from_substr(filename, 0, c - filename);
 346         qdict_put(options, "config", config_path);
 347     }
 348
 349     /* TODO Allow multi-level nesting and set file.filename here */
 350     filename = c + 1;
 351     qdict_put_str(options, "x-image", filename);
 352 }
 353
 354 static int blkdebug_parse_perm_list(uint64_t *dest, QDict *options,
 355                                     const char *prefix, Error **errp)
 356 {
 357     int ret = 0;
 358     QDict *subqdict = NULL;
 359     QObject *crumpled_subqdict = NULL;
 360     Visitor *v = NULL;
 361     BlockPermissionList *perm_list = NULL, *element;
 362
 363     *dest = 0;
 364
 365     qdict_extract_subqdict(options, &subqdict, prefix);
 366     if (!qdict_size(subqdict)) {
 367         goto out;
 368     }
 369
 370     crumpled_subqdict = qdict_crumple(subqdict, errp);
 371     if (!crumpled_subqdict) {
 372         ret = -EINVAL;
 373         goto out;
 374     }
 375
 376     v = qobject_input_visitor_new(crumpled_subqdict);
 377     if (!visit_type_BlockPermissionList(v, NULL, &perm_list, errp)) {
 378         ret = -EINVAL;
 379         goto out;
 380     }
 381
 382     for (element = perm_list; element; element = element->next) {
 383         *dest |= bdrv_qapi_perm_to_blk_perm(element->value);
 384     }
 385
 386 out:
 387     qapi_free_BlockPermissionList(perm_list);
 388     visit_free(v);
 389     qobject_unref(subqdict);
 390     qobject_unref(crumpled_subqdict);
 391     return ret;
 392 }
 393
 394 static int blkdebug_parse_perms(BDRVBlkdebugState *s, QDict *options,
 395                                 Error **errp)
 396 {
 397     int ret;
 398
 399     ret = blkdebug_parse_perm_list(&s->take_child_perms, options,
 400                                    "take-child-perms.", errp);
 401     if (ret < 0) {
 402         return ret;
 403     }
 404
 405     ret = blkdebug_parse_perm_list(&s->unshare_child_perms, options,
 406                                    "unshare-child-perms.", errp);
 407     if (ret < 0) {
 408         return ret;
 409     }
 410
 411     return 0;
 412 }
 413
 414 static QemuOptsList runtime_opts = {
 415     .name = "blkdebug",
 416     .head = QTAILQ_HEAD_INITIALIZER(runtime_opts.head),
 417     .desc = {
 418         {
 419             .name = "config",
 420             .type = QEMU_OPT_STRING,
 421             .help = "Path to the configuration file",
 422         },
 423         {
 424             .name = "x-image",
 425             .type = QEMU_OPT_STRING,
 426             .help = "[internal use only, will be removed]",
 427         },
 428         {
 429             .name = "align",
 430             .type = QEMU_OPT_SIZE,
 431             .help = "Required alignment in bytes",
 432         },
 433         {
 434             .name = "max-transfer",
 435             .type = QEMU_OPT_SIZE,
 436             .help = "Maximum transfer size in bytes",
 437         },
 438         {
 439             .name = "opt-write-zero",
 440             .type = QEMU_OPT_SIZE,
 441             .help = "Optimum write zero alignment in bytes",
 442         },
 443         {
 444             .name = "max-write-zero",
 445             .type = QEMU_OPT_SIZE,
 446             .help = "Maximum write zero size in bytes",
 447         },
 448         {
 449             .name = "opt-discard",
 450             .type = QEMU_OPT_SIZE,
 451             .help = "Optimum discard alignment in bytes",
 452         },
 453         {
 454             .name = "max-discard",
 455             .type = QEMU_OPT_SIZE,
 456             .help = "Maximum discard size in bytes",
 457         },
 458         { /* end of list */ }
 459     },
 460 };
 461
 462 static int blkdebug_open(BlockDriverState *bs, QDict *options, int flags,
 463                          Error **errp)
 464 {
 465     BDRVBlkdebugState *s = bs->opaque;
 466     QemuOpts *opts;
 467     Error *local_err = NULL;
 468     int ret;
 469     uint64_t align;
 470
 471     opts = qemu_opts_create(&runtime_opts, NULL, 0, &error_abort);
 472     if (!qemu_opts_absorb_qdict(opts, options, errp)) {
 473         ret = -EINVAL;
 474         goto out;
 475     }
 476
 477     /* Read rules from config file or command line options */
 478     s->config_file = g_strdup(qemu_opt_get(opts, "config"));
 479     ret = read_config(s, s->config_file, options, errp);
 480     if (ret) {
 481         goto out;
 482     }
 483
 484     /* Set initial state */
 485     s->state = 1;
 486
 487     /* Parse permissions modifiers before opening the image file */
 488     ret = blkdebug_parse_perms(s, options, errp);
 489     if (ret < 0) {
 490         goto out;
 491     }
 492
 493     /* Open the image file */
 494     bs->file = bdrv_open_child(qemu_opt_get(opts, "x-image"), options, "image",
 495                                bs, &child_of_bds,
 496                                BDRV_CHILD_FILTERED | BDRV_CHILD_PRIMARY,
 497                                false, &local_err);
 498     if (local_err) {
 499         ret = -EINVAL;
 500         error_propagate(errp, local_err);
 501         goto out;
 502     }
 503
 504     bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED |
 505         (BDRV_REQ_FUA & bs->file->bs->supported_write_flags);
 506     bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED |
 507         ((BDRV_REQ_FUA | BDRV_REQ_MAY_UNMAP | BDRV_REQ_NO_FALLBACK) &
 508             bs->file->bs->supported_zero_flags);
 509     ret = -EINVAL;
 510
 511     /* Set alignment overrides */
 512     s->align = qemu_opt_get_size(opts, "align", 0);
 513     if (s->align && (s->align >= INT_MAX || !is_power_of_2(s->align))) {
 514         error_setg(errp, "Cannot meet constraints with align %" PRIu64,
 515                    s->align);
 516         goto out;
 517     }
 518     align = MAX(s->align, bs->file->bs->bl.request_alignment);
 519
 520     s->max_transfer = qemu_opt_get_size(opts, "max-transfer", 0);
 521     if (s->max_transfer &&
 522         (s->max_transfer >= INT_MAX ||
 523          !QEMU_IS_ALIGNED(s->max_transfer, align))) {
 524         error_setg(errp, "Cannot meet constraints with max-transfer %" PRIu64,
 525                    s->max_transfer);
 526         goto out;
 527     }
 528
 529     s->opt_write_zero = qemu_opt_get_size(opts, "opt-write-zero", 0);
 530     if (s->opt_write_zero &&
 531         (s->opt_write_zero >= INT_MAX ||
 532          !QEMU_IS_ALIGNED(s->opt_write_zero, align))) {
 533         error_setg(errp, "Cannot meet constraints with opt-write-zero %" PRIu64,
 534                    s->opt_write_zero);
 535         goto out;
 536     }
 537
 538     s->max_write_zero = qemu_opt_get_size(opts, "max-write-zero", 0);
 539     if (s->max_write_zero &&
 540         (s->max_write_zero >= INT_MAX ||
 541          !QEMU_IS_ALIGNED(s->max_write_zero,
 542                           MAX(s->opt_write_zero, align)))) {
 543         error_setg(errp, "Cannot meet constraints with max-write-zero %" PRIu64,
 544                    s->max_write_zero);
 545         goto out;
 546     }
 547
 548     s->opt_discard = qemu_opt_get_size(opts, "opt-discard", 0);
 549     if (s->opt_discard &&
 550         (s->opt_discard >= INT_MAX ||
 551          !QEMU_IS_ALIGNED(s->opt_discard, align))) {
 552         error_setg(errp, "Cannot meet constraints with opt-discard %" PRIu64,
 553                    s->opt_discard);
 554         goto out;
 555     }
 556
 557     s->max_discard = qemu_opt_get_size(opts, "max-discard", 0);
 558     if (s->max_discard &&
 559         (s->max_discard >= INT_MAX ||
 560          !QEMU_IS_ALIGNED(s->max_discard,
 561                           MAX(s->opt_discard, align)))) {
 562         error_setg(errp, "Cannot meet constraints with max-discard %" PRIu64,
 563                    s->max_discard);
 564         goto out;
 565     }
 566
 567     bdrv_debug_event(bs, BLKDBG_NONE);
 568
 569     ret = 0;
 570 out:
 571     if (ret < 0) {
 572         g_free(s->config_file);
 573     }
 574     qemu_opts_del(opts);
 575     return ret;
 576 }
 577
 578 static int rule_check(BlockDriverState *bs, uint64_t offset, uint64_t bytes,
 579                       BlkdebugIOType iotype)
 580 {
 581     BDRVBlkdebugState *s = bs->opaque;
 582     BlkdebugRule *rule = NULL;
 583     int error;
 584     bool immediately;
 585
 586     QSIMPLEQ_FOREACH(rule, &s->active_rules, active_next) {
 587         uint64_t inject_offset = rule->options.inject.offset;
 588
 589         if ((inject_offset == -1 ||
 590              (bytes && inject_offset >= offset &&
 591               inject_offset < offset + bytes)) &&
 592             (rule->options.inject.iotype_mask & (1ull << iotype)))
 593         {
 594             break;
 595         }
 596     }
 597
 598     if (!rule || !rule->options.inject.error) {
 599         return 0;
 600     }
 601
 602     immediately = rule->options.inject.immediately;
 603     error = rule->options.inject.error;
 604
 605     if (rule->options.inject.once) {
 606         QSIMPLEQ_REMOVE(&s->active_rules, rule, BlkdebugRule, active_next);
 607         remove_rule(rule);
 608     }
 609
 610     if (!immediately) {
 611         aio_co_schedule(qemu_get_current_aio_context(), qemu_coroutine_self());
 612         qemu_coroutine_yield();
 613     }
 614
 615     return -error;
 616 }
 617
 618 static int coroutine_fn
 619 blkdebug_co_preadv(BlockDriverState *bs, uint64_t offset, uint64_t bytes,
 620                    QEMUIOVector *qiov, int flags)
 621 {
 622     int err;
 623
 624     /* Sanity check block layer guarantees */
 625     assert(QEMU_IS_ALIGNED(offset, bs->bl.request_alignment));
 626     assert(QEMU_IS_ALIGNED(bytes, bs->bl.request_alignment));
 627     if (bs->bl.max_transfer) {
 628         assert(bytes <= bs->bl.max_transfer);
 629     }
 630
 631     err = rule_check(bs, offset, bytes, BLKDEBUG_IO_TYPE_READ);
 632     if (err) {
 633         return err;
 634     }
 635
 636     return bdrv_co_preadv(bs->file, offset, bytes, qiov, flags);
 637 }
 638
 639 static int coroutine_fn
 640 blkdebug_co_pwritev(BlockDriverState *bs, uint64_t offset, uint64_t bytes,
 641                     QEMUIOVector *qiov, int flags)
 642 {
 643     int err;
 644
 645     /* Sanity check block layer guarantees */
 646     assert(QEMU_IS_ALIGNED(offset, bs->bl.request_alignment));
 647     assert(QEMU_IS_ALIGNED(bytes, bs->bl.request_alignment));
 648     if (bs->bl.max_transfer) {
 649         assert(bytes <= bs->bl.max_transfer);
 650     }
 651
 652     err = rule_check(bs, offset, bytes, BLKDEBUG_IO_TYPE_WRITE);
 653     if (err) {
 654         return err;
 655     }
 656
 657     return bdrv_co_pwritev(bs->file, offset, bytes, qiov, flags);
 658 }
 659
 660 static int blkdebug_co_flush(BlockDriverState *bs)
 661 {
 662     int err = rule_check(bs, 0, 0, BLKDEBUG_IO_TYPE_FLUSH);
 663
 664     if (err) {
 665         return err;
 666     }
 667
 668     return bdrv_co_flush(bs->file->bs);
 669 }
 670
 671 static int coroutine_fn blkdebug_co_pwrite_zeroes(BlockDriverState *bs,
 672                                                   int64_t offset, int bytes,
 673                                                   BdrvRequestFlags flags)
 674 {
 675     uint32_t align = MAX(bs->bl.request_alignment,
 676                          bs->bl.pwrite_zeroes_alignment);
 677     int err;
 678
 679     /* Only pass through requests that are larger than requested
 680      * preferred alignment (so that we test the fallback to writes on
 681      * unaligned portions), and check that the block layer never hands
 682      * us anything unaligned that crosses an alignment boundary.  */
 683     if (bytes < align) {
 684         assert(QEMU_IS_ALIGNED(offset, align) ||
 685                QEMU_IS_ALIGNED(offset + bytes, align) ||
 686                DIV_ROUND_UP(offset, align) ==
 687                DIV_ROUND_UP(offset + bytes, align));
 688         return -ENOTSUP;
 689     }
 690     assert(QEMU_IS_ALIGNED(offset, align));
 691     assert(QEMU_IS_ALIGNED(bytes, align));
 692     if (bs->bl.max_pwrite_zeroes) {
 693         assert(bytes <= bs->bl.max_pwrite_zeroes);
 694     }
 695
 696     err = rule_check(bs, offset, bytes, BLKDEBUG_IO_TYPE_WRITE_ZEROES);
 697     if (err) {
 698         return err;
 699     }
 700
 701     return bdrv_co_pwrite_zeroes(bs->file, offset, bytes, flags);
 702 }
 703
 704 static int coroutine_fn blkdebug_co_pdiscard(BlockDriverState *bs,
 705                                              int64_t offset, int bytes)
 706 {
 707     uint32_t align = bs->bl.pdiscard_alignment;
 708     int err;
 709
 710     /* Only pass through requests that are larger than requested
 711      * minimum alignment, and ensure that unaligned requests do not
 712      * cross optimum discard boundaries. */
 713     if (bytes < bs->bl.request_alignment) {
 714         assert(QEMU_IS_ALIGNED(offset, align) ||
 715                QEMU_IS_ALIGNED(offset + bytes, align) ||
 716                DIV_ROUND_UP(offset, align) ==
 717                DIV_ROUND_UP(offset + bytes, align));
 718         return -ENOTSUP;
 719     }
 720     assert(QEMU_IS_ALIGNED(offset, bs->bl.request_alignment));
 721     assert(QEMU_IS_ALIGNED(bytes, bs->bl.request_alignment));
 722     if (align && bytes >= align) {
 723         assert(QEMU_IS_ALIGNED(offset, align));
 724         assert(QEMU_IS_ALIGNED(bytes, align));
 725     }
 726     if (bs->bl.max_pdiscard) {
 727         assert(bytes <= bs->bl.max_pdiscard);
 728     }
 729
 730     err = rule_check(bs, offset, bytes, BLKDEBUG_IO_TYPE_DISCARD);
 731     if (err) {
 732         return err;
 733     }
 734
 735     return bdrv_co_pdiscard(bs->file, offset, bytes);
 736 }
 737
 738 static int coroutine_fn blkdebug_co_block_status(BlockDriverState *bs,
 739                                                  bool want_zero,
 740                                                  int64_t offset,
 741                                                  int64_t bytes,
 742                                                  int64_t *pnum,
 743                                                  int64_t *map,
 744                                                  BlockDriverState **file)
 745 {
 746     int err;
 747
 748     assert(QEMU_IS_ALIGNED(offset | bytes, bs->bl.request_alignment));
 749
 750     err = rule_check(bs, offset, bytes, BLKDEBUG_IO_TYPE_BLOCK_STATUS);
 751     if (err) {
 752         return err;
 753     }
 754
 755     assert(bs->file && bs->file->bs);
 756     *pnum = bytes;
 757     *map = offset;
 758     *file = bs->file->bs;
 759     return BDRV_BLOCK_RAW | BDRV_BLOCK_OFFSET_VALID;
 760 }
 761
 762 static void blkdebug_close(BlockDriverState *bs)
 763 {
 764     BDRVBlkdebugState *s = bs->opaque;
 765     BlkdebugRule *rule, *next;
 766     int i;
 767
 768     for (i = 0; i < BLKDBG__MAX; i++) {
 769         QLIST_FOREACH_SAFE(rule, &s->rules[i], next, next) {
 770             remove_rule(rule);
 771         }
 772     }
 773
 774     g_free(s->config_file);
 775 }
 776
 777 static void suspend_request(BlockDriverState *bs, BlkdebugRule *rule)
 778 {
 779     BDRVBlkdebugState *s = bs->opaque;
 780     BlkdebugSuspendedReq r;
 781
 782     r = (BlkdebugSuspendedReq) {
 783         .co         = qemu_coroutine_self(),
 784         .tag        = g_strdup(rule->options.suspend.tag),
 785     };
 786
 787     remove_rule(rule);
 788     QLIST_INSERT_HEAD(&s->suspended_reqs, &r, next);
 789
 790     if (!qtest_enabled()) {
 791         printf("blkdebug: Suspended request '%s'\n", r.tag);
 792     }
 793     qemu_coroutine_yield();
 794     if (!qtest_enabled()) {
 795         printf("blkdebug: Resuming request '%s'\n", r.tag);
 796     }
 797
 798     QLIST_REMOVE(&r, next);
 799     g_free(r.tag);
 800 }
 801
 802 static bool process_rule(BlockDriverState *bs, struct BlkdebugRule *rule,
 803     bool injected)
 804 {
 805     BDRVBlkdebugState *s = bs->opaque;
 806
 807     /* Only process rules for the current state */
 808     if (rule->state && rule->state != s->state) {
 809         return injected;
 810     }
 811
 812     /* Take the action */
 813     switch (rule->action) {
 814     case ACTION_INJECT_ERROR:
 815         if (!injected) {
 816             QSIMPLEQ_INIT(&s->active_rules);
 817             injected = true;
 818         }
 819         QSIMPLEQ_INSERT_HEAD(&s->active_rules, rule, active_next);
 820         break;
 821
 822     case ACTION_SET_STATE:
 823         s->new_state = rule->options.set_state.new_state;
 824         break;
 825
 826     case ACTION_SUSPEND:
 827         suspend_request(bs, rule);
 828         break;
 829     }
 830     return injected;
 831 }
 832
 833 static void blkdebug_debug_event(BlockDriverState *bs, BlkdebugEvent event)
 834 {
 835     BDRVBlkdebugState *s = bs->opaque;
 836     struct BlkdebugRule *rule, *next;
 837     bool injected;
 838
 839     assert((int)event >= 0 && event < BLKDBG__MAX);
 840
 841     injected = false;
 842     s->new_state = s->state;
 843     QLIST_FOREACH_SAFE(rule, &s->rules[event], next, next) {
 844         injected = process_rule(bs, rule, injected);
 845     }
 846     s->state = s->new_state;
 847 }
 848
 849 static int blkdebug_debug_breakpoint(BlockDriverState *bs, const char *event,
 850                                      const char *tag)
 851 {
 852     BDRVBlkdebugState *s = bs->opaque;
 853     struct BlkdebugRule *rule;
 854     int blkdebug_event;
 855
 856     blkdebug_event = qapi_enum_parse(&BlkdebugEvent_lookup, event, -1, NULL);
 857     if (blkdebug_event < 0) {
 858         return -ENOENT;
 859     }
 860
 861     rule = g_malloc(sizeof(*rule));
 862     *rule = (struct BlkdebugRule) {
 863         .event  = blkdebug_event,
 864         .action = ACTION_SUSPEND,
 865         .state  = 0,
 866         .options.suspend.tag = g_strdup(tag),
 867     };
 868
 869     QLIST_INSERT_HEAD(&s->rules[blkdebug_event], rule, next);
 870
 871     return 0;
 872 }
 873
 874 static int blkdebug_debug_resume(BlockDriverState *bs, const char *tag)
 875 {
 876     BDRVBlkdebugState *s = bs->opaque;
 877     BlkdebugSuspendedReq *r, *next;
 878
 879     QLIST_FOREACH_SAFE(r, &s->suspended_reqs, next, next) {
 880         if (!strcmp(r->tag, tag)) {
 881             qemu_coroutine_enter(r->co);
 882             return 0;
 883         }
 884     }
 885     return -ENOENT;
 886 }
 887
 888 static int blkdebug_debug_remove_breakpoint(BlockDriverState *bs,
 889                                             const char *tag)
 890 {
 891     BDRVBlkdebugState *s = bs->opaque;
 892     BlkdebugSuspendedReq *r, *r_next;
 893     BlkdebugRule *rule, *next;
 894     int i, ret = -ENOENT;
 895
 896     for (i = 0; i < BLKDBG__MAX; i++) {
 897         QLIST_FOREACH_SAFE(rule, &s->rules[i], next, next) {
 898             if (rule->action == ACTION_SUSPEND &&
 899                 !strcmp(rule->options.suspend.tag, tag)) {
 900                 remove_rule(rule);
 901                 ret = 0;
 902             }
 903         }
 904     }
 905     QLIST_FOREACH_SAFE(r, &s->suspended_reqs, next, r_next) {
 906         if (!strcmp(r->tag, tag)) {
 907             qemu_coroutine_enter(r->co);
 908             ret = 0;
 909         }
 910     }
 911     return ret;
 912 }
 913
 914 static bool blkdebug_debug_is_suspended(BlockDriverState *bs, const char *tag)
 915 {
 916     BDRVBlkdebugState *s = bs->opaque;
 917     BlkdebugSuspendedReq *r;
 918
 919     QLIST_FOREACH(r, &s->suspended_reqs, next) {
 920         if (!strcmp(r->tag, tag)) {
 921             return true;
 922         }
 923     }
 924     return false;
 925 }
 926
 927 static int64_t blkdebug_getlength(BlockDriverState *bs)
 928 {
 929     return bdrv_getlength(bs->file->bs);
 930 }
 931
 932 static void blkdebug_refresh_filename(BlockDriverState *bs)
 933 {
 934     BDRVBlkdebugState *s = bs->opaque;
 935     const QDictEntry *e;
 936     int ret;
 937
 938     if (!bs->file->bs->exact_filename[0]) {
 939         return;
 940     }
 941
 942     for (e = qdict_first(bs->full_open_options); e;
 943          e = qdict_next(bs->full_open_options, e))
 944     {
 945         /* Real child options are under "image", but "x-image" may
 946          * contain a filename */
 947         if (strcmp(qdict_entry_key(e), "config") &&
 948             strcmp(qdict_entry_key(e), "image") &&
 949             strcmp(qdict_entry_key(e), "x-image") &&
 950             strcmp(qdict_entry_key(e), "driver"))
 951         {
 952             return;
 953         }
 954     }
 955
 956     ret = snprintf(bs->exact_filename, sizeof(bs->exact_filename),
 957                    "blkdebug:%s:%s",
 958                    s->config_file ?: "", bs->file->bs->exact_filename);
 959     if (ret >= sizeof(bs->exact_filename)) {
 960         /* An overflow makes the filename unusable, so do not report any */
 961         bs->exact_filename[0] = 0;
 962     }
 963 }
 964
 965 static void blkdebug_refresh_limits(BlockDriverState *bs, Error **errp)
 966 {
 967     BDRVBlkdebugState *s = bs->opaque;
 968
 969     if (s->align) {
 970         bs->bl.request_alignment = s->align;
 971     }
 972     if (s->max_transfer) {
 973         bs->bl.max_transfer = s->max_transfer;
 974     }
 975     if (s->opt_write_zero) {
 976         bs->bl.pwrite_zeroes_alignment = s->opt_write_zero;
 977     }
 978     if (s->max_write_zero) {
 979         bs->bl.max_pwrite_zeroes = s->max_write_zero;
 980     }
 981     if (s->opt_discard) {
 982         bs->bl.pdiscard_alignment = s->opt_discard;
 983     }
 984     if (s->max_discard) {
 985         bs->bl.max_pdiscard = s->max_discard;
 986     }
 987 }
 988
 989 static int blkdebug_reopen_prepare(BDRVReopenState *reopen_state,
 990                                    BlockReopenQueue *queue, Error **errp)
 991 {
 992     return 0;
 993 }
 994
 995 static void blkdebug_child_perm(BlockDriverState *bs, BdrvChild *c,
 996                                 BdrvChildRole role,
 997                                 BlockReopenQueue *reopen_queue,
 998                                 uint64_t perm, uint64_t shared,
 999                                 uint64_t *nperm, uint64_t *nshared)
1000 {
1001     BDRVBlkdebugState *s = bs->opaque;
1002
1003     bdrv_default_perms(bs, c, role, reopen_queue,
1004                        perm, shared, nperm, nshared);
1005
1006     *nperm |= s->take_child_perms;
1007     *nshared &= ~s->unshare_child_perms;
1008 }
1009
1010 static const char *const blkdebug_strong_runtime_opts[] = {
1011     "config",
1012     "inject-error.",
1013     "set-state.",
1014     "align",
1015     "max-transfer",
1016     "opt-write-zero",
1017     "max-write-zero",
1018     "opt-discard",
1019     "max-discard",
1020
1021     NULL
1022 };
1023
1024 static BlockDriver bdrv_blkdebug = {
1025     .format_name            = "blkdebug",
1026     .protocol_name          = "blkdebug",
1027     .instance_size          = sizeof(BDRVBlkdebugState),
1028     .is_filter              = true,
1029
1030     .bdrv_parse_filename    = blkdebug_parse_filename,
1031     .bdrv_file_open         = blkdebug_open,
1032     .bdrv_close             = blkdebug_close,
1033     .bdrv_reopen_prepare    = blkdebug_reopen_prepare,
1034     .bdrv_child_perm        = blkdebug_child_perm,
1035
1036     .bdrv_getlength         = blkdebug_getlength,
1037     .bdrv_refresh_filename  = blkdebug_refresh_filename,
1038     .bdrv_refresh_limits    = blkdebug_refresh_limits,
1039
1040     .bdrv_co_preadv         = blkdebug_co_preadv,
1041     .bdrv_co_pwritev        = blkdebug_co_pwritev,
1042     .bdrv_co_flush_to_disk  = blkdebug_co_flush,
1043     .bdrv_co_pwrite_zeroes  = blkdebug_co_pwrite_zeroes,
1044     .bdrv_co_pdiscard       = blkdebug_co_pdiscard,
1045     .bdrv_co_block_status   = blkdebug_co_block_status,
1046
1047     .bdrv_debug_event           = blkdebug_debug_event,
1048     .bdrv_debug_breakpoint      = blkdebug_debug_breakpoint,
1049     .bdrv_debug_remove_breakpoint
1050                                 = blkdebug_debug_remove_breakpoint,
1051     .bdrv_debug_resume          = blkdebug_debug_resume,
1052     .bdrv_debug_is_suspended    = blkdebug_debug_is_suspended,
1053
1054     .strong_runtime_opts        = blkdebug_strong_runtime_opts,
1055 };
1056
1057 static void bdrv_blkdebug_init(void)
1058 {
1059     bdrv_register(&bdrv_blkdebug);
1060 }
1061
1062 block_init(bdrv_blkdebug_init);