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 <[email protected]>
   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 /* All APIs are thread-safe */
  42
  43 typedef struct BDRVBlkdebugState {
  44     /* IN: initialized in blkdebug_open() and never changed */
  45     uint64_t align;
  46     uint64_t max_transfer;
  47     uint64_t opt_write_zero;
  48     uint64_t max_write_zero;
  49     uint64_t opt_discard;
  50     uint64_t max_discard;
  51     char *config_file; /* For blkdebug_refresh_filename() */
  52     /* initialized in blkdebug_parse_perms() */
  53     uint64_t take_child_perms;
  54     uint64_t unshare_child_perms;
  55
  56     /* State. Protected by lock */
  57     int state;
  58     QLIST_HEAD(, BlkdebugRule) rules[BLKDBG__MAX];
  59     QSIMPLEQ_HEAD(, BlkdebugRule) active_rules;
  60     QLIST_HEAD(, BlkdebugSuspendedReq) suspended_reqs;
  61     QemuMutex lock;
  62 } BDRVBlkdebugState;
  63
  64 typedef struct BlkdebugAIOCB {
  65     BlockAIOCB common;
  66     int ret;
  67 } BlkdebugAIOCB;
  68
  69 typedef struct BlkdebugSuspendedReq {
  70     /* IN: initialized in suspend_request() */
  71     Coroutine *co;
  72     char *tag;
  73
  74     /* List entry protected BDRVBlkdebugState's lock */
  75     QLIST_ENTRY(BlkdebugSuspendedReq) next;
  76 } BlkdebugSuspendedReq;
  77
  78 enum {
  79     ACTION_INJECT_ERROR,
  80     ACTION_SET_STATE,
  81     ACTION_SUSPEND,
  82     ACTION__MAX,
  83 };
  84
  85 typedef struct BlkdebugRule {
  86     /* IN: initialized in add_rule() or blkdebug_debug_breakpoint() */
  87     BlkdebugEvent event;
  88     int action;
  89     int state;
  90     union {
  91         struct {
  92             uint64_t iotype_mask;
  93             int error;
  94             int immediately;
  95             int once;
  96             int64_t offset;
  97         } inject;
  98         struct {
  99             int new_state;
 100         } set_state;
 101         struct {
 102             char *tag;
 103         } suspend;
 104     } options;
 105
 106     /* List entries protected BDRVBlkdebugState's lock */
 107     QLIST_ENTRY(BlkdebugRule) next;
 108     QSIMPLEQ_ENTRY(BlkdebugRule) active_next;
 109 } BlkdebugRule;
 110
 111 QEMU_BUILD_BUG_MSG(BLKDEBUG_IO_TYPE__MAX > 64,
 112                    "BlkdebugIOType mask does not fit into an uint64_t");
 113
 114 static QemuOptsList inject_error_opts = {
 115     .name = "inject-error",
 116     .head = QTAILQ_HEAD_INITIALIZER(inject_error_opts.head),
 117     .desc = {
 118         {
 119             .name = "event",
 120             .type = QEMU_OPT_STRING,
 121         },
 122         {
 123             .name = "state",
 124             .type = QEMU_OPT_NUMBER,
 125         },
 126         {
 127             .name = "iotype",
 128             .type = QEMU_OPT_STRING,
 129         },
 130         {
 131             .name = "errno",
 132             .type = QEMU_OPT_NUMBER,
 133         },
 134         {
 135             .name = "sector",
 136             .type = QEMU_OPT_NUMBER,
 137         },
 138         {
 139             .name = "once",
 140             .type = QEMU_OPT_BOOL,
 141         },
 142         {
 143             .name = "immediately",
 144             .type = QEMU_OPT_BOOL,
 145         },
 146         { /* end of list */ }
 147     },
 148 };
 149
 150 static QemuOptsList set_state_opts = {
 151     .name = "set-state",
 152     .head = QTAILQ_HEAD_INITIALIZER(set_state_opts.head),
 153     .desc = {
 154         {
 155             .name = "event",
 156             .type = QEMU_OPT_STRING,
 157         },
 158         {
 159             .name = "state",
 160             .type = QEMU_OPT_NUMBER,
 161         },
 162         {
 163             .name = "new_state",
 164             .type = QEMU_OPT_NUMBER,
 165         },
 166         { /* end of list */ }
 167     },
 168 };
 169
 170 static QemuOptsList *config_groups[] = {
 171     &inject_error_opts,
 172     &set_state_opts,
 173     NULL
 174 };
 175
 176 struct add_rule_data {
 177     BDRVBlkdebugState *s;
 178     int action;
 179 };
 180
 181 static int add_rule(void *opaque, QemuOpts *opts, Error **errp)
 182 {
 183     struct add_rule_data *d = opaque;
 184     BDRVBlkdebugState *s = d->s;
 185     const char *event_name;
 186     int event;
 187     struct BlkdebugRule *rule;
 188     int64_t sector;
 189     BlkdebugIOType iotype;
 190     Error *local_error = NULL;
 191
 192     /* Find the right event for the rule */
 193     event_name = qemu_opt_get(opts, "event");
 194     if (!event_name) {
 195         error_setg(errp, "Missing event name for rule");
 196         return -1;
 197     }
 198     event = qapi_enum_parse(&BlkdebugEvent_lookup, event_name, -1, errp);
 199     if (event < 0) {
 200         return -1;
 201     }
 202
 203     /* Set attributes common for all actions */
 204     rule = g_malloc0(sizeof(*rule));
 205     *rule = (struct BlkdebugRule) {
 206         .event  = event,
 207         .action = d->action,
 208         .state  = qemu_opt_get_number(opts, "state", 0),
 209     };
 210
 211     /* Parse action-specific options */
 212     switch (d->action) {
 213     case ACTION_INJECT_ERROR:
 214         rule->options.inject.error = qemu_opt_get_number(opts, "errno", EIO);
 215         rule->options.inject.once  = qemu_opt_get_bool(opts, "once", 0);
 216         rule->options.inject.immediately =
 217             qemu_opt_get_bool(opts, "immediately", 0);
 218         sector = qemu_opt_get_number(opts, "sector", -1);
 219         rule->options.inject.offset =
 220             sector == -1 ? -1 : sector * BDRV_SECTOR_SIZE;
 221
 222         iotype = qapi_enum_parse(&BlkdebugIOType_lookup,
 223                                  qemu_opt_get(opts, "iotype"),
 224                                  BLKDEBUG_IO_TYPE__MAX, &local_error);
 225         if (local_error) {
 226             error_propagate(errp, local_error);
 227             g_free(rule);
 228             return -1;
 229         }
 230         if (iotype != BLKDEBUG_IO_TYPE__MAX) {
 231             rule->options.inject.iotype_mask = (1ull << iotype);
 232         } else {
 233             /* Apply the default */
 234             rule->options.inject.iotype_mask =
 235                 (1ull << BLKDEBUG_IO_TYPE_READ)
 236                 | (1ull << BLKDEBUG_IO_TYPE_WRITE)
 237                 | (1ull << BLKDEBUG_IO_TYPE_WRITE_ZEROES)
 238                 | (1ull << BLKDEBUG_IO_TYPE_DISCARD)
 239                 | (1ull << BLKDEBUG_IO_TYPE_FLUSH);
 240         }
 241
 242         break;
 243
 244     case ACTION_SET_STATE:
 245         rule->options.set_state.new_state =
 246             qemu_opt_get_number(opts, "new_state", 0);
 247         break;
 248
 249     case ACTION_SUSPEND:
 250         rule->options.suspend.tag =
 251             g_strdup(qemu_opt_get(opts, "tag"));
 252         break;
 253     };
 254
 255     /* Add the rule */
 256     qemu_mutex_lock(&s->lock);
 257     QLIST_INSERT_HEAD(&s->rules[event], rule, next);
 258     qemu_mutex_unlock(&s->lock);
 259
 260     return 0;
 261 }
 262
 263 /* Called with lock held or from .bdrv_close */
 264 static void remove_rule(BlkdebugRule *rule)
 265 {
 266     switch (rule->action) {
 267     case ACTION_INJECT_ERROR:
 268     case ACTION_SET_STATE:
 269         break;
 270     case ACTION_SUSPEND:
 271         g_free(rule->options.suspend.tag);
 272         break;
 273     }
 274
 275     QLIST_REMOVE(rule, next);
 276     g_free(rule);
 277 }
 278
 279 static int read_config(BDRVBlkdebugState *s, const char *filename,
 280                        QDict *options, Error **errp)
 281 {
 282     FILE *f = NULL;
 283     int ret;
 284     struct add_rule_data d;
 285     Error *local_err = NULL;
 286
 287     if (filename) {
 288         f = fopen(filename, "r");
 289         if (f == NULL) {
 290             error_setg_errno(errp, errno, "Could not read blkdebug config file");
 291             return -errno;
 292         }
 293
 294         ret = qemu_config_parse(f, config_groups, filename, errp);
 295         if (ret < 0) {
 296             goto fail;
 297         }
 298     }
 299
 300     qemu_config_parse_qdict(options, config_groups, &local_err);
 301     if (local_err) {
 302         error_propagate(errp, local_err);
 303         ret = -EINVAL;
 304         goto fail;
 305     }
 306
 307     d.s = s;
 308     d.action = ACTION_INJECT_ERROR;
 309     qemu_opts_foreach(&inject_error_opts, add_rule, &d, &local_err);
 310     if (local_err) {
 311         error_propagate(errp, local_err);
 312         ret = -EINVAL;
 313         goto fail;
 314     }
 315
 316     d.action = ACTION_SET_STATE;
 317     qemu_opts_foreach(&set_state_opts, add_rule, &d, &local_err);
 318     if (local_err) {
 319         error_propagate(errp, local_err);
 320         ret = -EINVAL;
 321         goto fail;
 322     }
 323
 324     ret = 0;
 325 fail:
 326     qemu_opts_reset(&inject_error_opts);
 327     qemu_opts_reset(&set_state_opts);
 328     if (f) {
 329         fclose(f);
 330     }
 331     return ret;
 332 }
 333
 334 /* Valid blkdebug filenames look like blkdebug:path/to/config:path/to/image */
 335 static void blkdebug_parse_filename(const char *filename, QDict *options,
 336                                     Error **errp)
 337 {
 338     const char *c;
 339
 340     /* Parse the blkdebug: prefix */
 341     if (!strstart(filename, "blkdebug:", &filename)) {
 342         /* There was no prefix; therefore, all options have to be already
 343            present in the QDict (except for the filename) */
 344         qdict_put_str(options, "x-image", filename);
 345         return;
 346     }
 347
 348     /* Parse config file path */
 349     c = strchr(filename, ':');
 350     if (c == NULL) {
 351         error_setg(errp, "blkdebug requires both config file and image path");
 352         return;
 353     }
 354
 355     if (c != filename) {
 356         QString *config_path;
 357         config_path = qstring_from_substr(filename, 0, c - filename);
 358         qdict_put(options, "config", config_path);
 359     }
 360
 361     /* TODO Allow multi-level nesting and set file.filename here */
 362     filename = c + 1;
 363     qdict_put_str(options, "x-image", filename);
 364 }
 365
 366 static int blkdebug_parse_perm_list(uint64_t *dest, QDict *options,
 367                                     const char *prefix, Error **errp)
 368 {
 369     int ret = 0;
 370     QDict *subqdict = NULL;
 371     QObject *crumpled_subqdict = NULL;
 372     Visitor *v = NULL;
 373     BlockPermissionList *perm_list = NULL, *element;
 374
 375     *dest = 0;
 376
 377     qdict_extract_subqdict(options, &subqdict, prefix);
 378     if (!qdict_size(subqdict)) {
 379         goto out;
 380     }
 381
 382     crumpled_subqdict = qdict_crumple(subqdict, errp);
 383     if (!crumpled_subqdict) {
 384         ret = -EINVAL;
 385         goto out;
 386     }
 387
 388     v = qobject_input_visitor_new(crumpled_subqdict);
 389     if (!visit_type_BlockPermissionList(v, NULL, &perm_list, errp)) {
 390         ret = -EINVAL;
 391         goto out;
 392     }
 393
 394     for (element = perm_list; element; element = element->next) {
 395         *dest |= bdrv_qapi_perm_to_blk_perm(element->value);
 396     }
 397
 398 out:
 399     qapi_free_BlockPermissionList(perm_list);
 400     visit_free(v);
 401     qobject_unref(subqdict);
 402     qobject_unref(crumpled_subqdict);
 403     return ret;
 404 }
 405
 406 static int blkdebug_parse_perms(BDRVBlkdebugState *s, QDict *options,
 407                                 Error **errp)
 408 {
 409     int ret;
 410
 411     ret = blkdebug_parse_perm_list(&s->take_child_perms, options,
 412                                    "take-child-perms.", errp);
 413     if (ret < 0) {
 414         return ret;
 415     }
 416
 417     ret = blkdebug_parse_perm_list(&s->unshare_child_perms, options,
 418                                    "unshare-child-perms.", errp);
 419     if (ret < 0) {
 420         return ret;
 421     }
 422
 423     return 0;
 424 }
 425
 426 static QemuOptsList runtime_opts = {
 427     .name = "blkdebug",
 428     .head = QTAILQ_HEAD_INITIALIZER(runtime_opts.head),
 429     .desc = {
 430         {
 431             .name = "config",
 432             .type = QEMU_OPT_STRING,
 433             .help = "Path to the configuration file",
 434         },
 435         {
 436             .name = "x-image",
 437             .type = QEMU_OPT_STRING,
 438             .help = "[internal use only, will be removed]",
 439         },
 440         {
 441             .name = "align",
 442             .type = QEMU_OPT_SIZE,
 443             .help = "Required alignment in bytes",
 444         },
 445         {
 446             .name = "max-transfer",
 447             .type = QEMU_OPT_SIZE,
 448             .help = "Maximum transfer size in bytes",
 449         },
 450         {
 451             .name = "opt-write-zero",
 452             .type = QEMU_OPT_SIZE,
 453             .help = "Optimum write zero alignment in bytes",
 454         },
 455         {
 456             .name = "max-write-zero",
 457             .type = QEMU_OPT_SIZE,
 458             .help = "Maximum write zero size in bytes",
 459         },
 460         {
 461             .name = "opt-discard",
 462             .type = QEMU_OPT_SIZE,
 463             .help = "Optimum discard alignment in bytes",
 464         },
 465         {
 466             .name = "max-discard",
 467             .type = QEMU_OPT_SIZE,
 468             .help = "Maximum discard size in bytes",
 469         },
 470         { /* end of list */ }
 471     },
 472 };
 473
 474 static int blkdebug_open(BlockDriverState *bs, QDict *options, int flags,
 475                          Error **errp)
 476 {
 477     BDRVBlkdebugState *s = bs->opaque;
 478     QemuOpts *opts;
 479     int ret;
 480     uint64_t align;
 481
 482     qemu_mutex_init(&s->lock);
 483     opts = qemu_opts_create(&runtime_opts, NULL, 0, &error_abort);
 484     if (!qemu_opts_absorb_qdict(opts, options, errp)) {
 485         ret = -EINVAL;
 486         goto out;
 487     }
 488
 489     /* Read rules from config file or command line options */
 490     s->config_file = g_strdup(qemu_opt_get(opts, "config"));
 491     ret = read_config(s, s->config_file, options, errp);
 492     if (ret) {
 493         goto out;
 494     }
 495
 496     /* Set initial state */
 497     s->state = 1;
 498
 499     /* Parse permissions modifiers before opening the image file */
 500     ret = blkdebug_parse_perms(s, options, errp);
 501     if (ret < 0) {
 502         goto out;
 503     }
 504
 505     /* Open the image file */
 506     ret = bdrv_open_file_child(qemu_opt_get(opts, "x-image"), options, "image",
 507                                bs, errp);
 508     if (ret < 0) {
 509         goto out;
 510     }
 511
 512     bs->supported_write_flags = BDRV_REQ_WRITE_UNCHANGED |
 513         (BDRV_REQ_FUA & bs->file->bs->supported_write_flags);
 514     bs->supported_zero_flags = BDRV_REQ_WRITE_UNCHANGED |
 515         ((BDRV_REQ_FUA | BDRV_REQ_MAY_UNMAP | BDRV_REQ_NO_FALLBACK) &
 516             bs->file->bs->supported_zero_flags);
 517     ret = -EINVAL;
 518
 519     /* Set alignment overrides */
 520     s->align = qemu_opt_get_size(opts, "align", 0);
 521     if (s->align && (s->align >= INT_MAX || !is_power_of_2(s->align))) {
 522         error_setg(errp, "Cannot meet constraints with align %" PRIu64,
 523                    s->align);
 524         goto out;
 525     }
 526     align = MAX(s->align, bs->file->bs->bl.request_alignment);
 527
 528     s->max_transfer = qemu_opt_get_size(opts, "max-transfer", 0);
 529     if (s->max_transfer &&
 530         (s->max_transfer >= INT_MAX ||
 531          !QEMU_IS_ALIGNED(s->max_transfer, align))) {
 532         error_setg(errp, "Cannot meet constraints with max-transfer %" PRIu64,
 533                    s->max_transfer);
 534         goto out;
 535     }
 536
 537     s->opt_write_zero = qemu_opt_get_size(opts, "opt-write-zero", 0);
 538     if (s->opt_write_zero &&
 539         (s->opt_write_zero >= INT_MAX ||
 540          !QEMU_IS_ALIGNED(s->opt_write_zero, align))) {
 541         error_setg(errp, "Cannot meet constraints with opt-write-zero %" PRIu64,
 542                    s->opt_write_zero);
 543         goto out;
 544     }
 545
 546     s->max_write_zero = qemu_opt_get_size(opts, "max-write-zero", 0);
 547     if (s->max_write_zero &&
 548         (s->max_write_zero >= INT_MAX ||
 549          !QEMU_IS_ALIGNED(s->max_write_zero,
 550                           MAX(s->opt_write_zero, align)))) {
 551         error_setg(errp, "Cannot meet constraints with max-write-zero %" PRIu64,
 552                    s->max_write_zero);
 553         goto out;
 554     }
 555
 556     s->opt_discard = qemu_opt_get_size(opts, "opt-discard", 0);
 557     if (s->opt_discard &&
 558         (s->opt_discard >= INT_MAX ||
 559          !QEMU_IS_ALIGNED(s->opt_discard, align))) {
 560         error_setg(errp, "Cannot meet constraints with opt-discard %" PRIu64,
 561                    s->opt_discard);
 562         goto out;
 563     }
 564
 565     s->max_discard = qemu_opt_get_size(opts, "max-discard", 0);
 566     if (s->max_discard &&
 567         (s->max_discard >= INT_MAX ||
 568          !QEMU_IS_ALIGNED(s->max_discard,
 569                           MAX(s->opt_discard, align)))) {
 570         error_setg(errp, "Cannot meet constraints with max-discard %" PRIu64,
 571                    s->max_discard);
 572         goto out;
 573     }
 574
 575     bdrv_debug_event(bs, BLKDBG_NONE);
 576
 577     ret = 0;
 578 out:
 579     if (ret < 0) {
 580         qemu_mutex_destroy(&s->lock);
 581         g_free(s->config_file);
 582     }
 583     qemu_opts_del(opts);
 584     return ret;
 585 }
 586
 587 static int rule_check(BlockDriverState *bs, uint64_t offset, uint64_t bytes,
 588                       BlkdebugIOType iotype)
 589 {
 590     BDRVBlkdebugState *s = bs->opaque;
 591     BlkdebugRule *rule = NULL;
 592     int error;
 593     bool immediately;
 594
 595     qemu_mutex_lock(&s->lock);
 596     QSIMPLEQ_FOREACH(rule, &s->active_rules, active_next) {
 597         uint64_t inject_offset = rule->options.inject.offset;
 598
 599         if ((inject_offset == -1 ||
 600              (bytes && inject_offset >= offset &&
 601               inject_offset < offset + bytes)) &&
 602             (rule->options.inject.iotype_mask & (1ull << iotype)))
 603         {
 604             break;
 605         }
 606     }
 607
 608     if (!rule || !rule->options.inject.error) {
 609         qemu_mutex_unlock(&s->lock);
 610         return 0;
 611     }
 612
 613     immediately = rule->options.inject.immediately;
 614     error = rule->options.inject.error;
 615
 616     if (rule->options.inject.once) {
 617         QSIMPLEQ_REMOVE(&s->active_rules, rule, BlkdebugRule, active_next);
 618         remove_rule(rule);
 619     }
 620
 621     qemu_mutex_unlock(&s->lock);
 622     if (!immediately) {
 623         aio_co_schedule(qemu_get_current_aio_context(), qemu_coroutine_self());
 624         qemu_coroutine_yield();
 625     }
 626
 627     return -error;
 628 }
 629
 630 static int coroutine_fn
 631 blkdebug_co_preadv(BlockDriverState *bs, int64_t offset, int64_t bytes,
 632                    QEMUIOVector *qiov, BdrvRequestFlags flags)
 633 {
 634     int err;
 635
 636     /* Sanity check block layer guarantees */
 637     assert(QEMU_IS_ALIGNED(offset, bs->bl.request_alignment));
 638     assert(QEMU_IS_ALIGNED(bytes, bs->bl.request_alignment));
 639     if (bs->bl.max_transfer) {
 640         assert(bytes <= bs->bl.max_transfer);
 641     }
 642
 643     err = rule_check(bs, offset, bytes, BLKDEBUG_IO_TYPE_READ);
 644     if (err) {
 645         return err;
 646     }
 647
 648     return bdrv_co_preadv(bs->file, offset, bytes, qiov, flags);
 649 }
 650
 651 static int coroutine_fn
 652 blkdebug_co_pwritev(BlockDriverState *bs, int64_t offset, int64_t bytes,
 653                     QEMUIOVector *qiov, BdrvRequestFlags flags)
 654 {
 655     int err;
 656
 657     /* Sanity check block layer guarantees */
 658     assert(QEMU_IS_ALIGNED(offset, bs->bl.request_alignment));
 659     assert(QEMU_IS_ALIGNED(bytes, bs->bl.request_alignment));
 660     if (bs->bl.max_transfer) {
 661         assert(bytes <= bs->bl.max_transfer);
 662     }
 663
 664     err = rule_check(bs, offset, bytes, BLKDEBUG_IO_TYPE_WRITE);
 665     if (err) {
 666         return err;
 667     }
 668
 669     return bdrv_co_pwritev(bs->file, offset, bytes, qiov, flags);
 670 }
 671
 672 static int coroutine_fn blkdebug_co_flush(BlockDriverState *bs)
 673 {
 674     int err = rule_check(bs, 0, 0, BLKDEBUG_IO_TYPE_FLUSH);
 675
 676     if (err) {
 677         return err;
 678     }
 679
 680     return bdrv_co_flush(bs->file->bs);
 681 }
 682
 683 static int coroutine_fn blkdebug_co_pwrite_zeroes(BlockDriverState *bs,
 684                                                   int64_t offset, int64_t bytes,
 685                                                   BdrvRequestFlags flags)
 686 {
 687     uint32_t align = MAX(bs->bl.request_alignment,
 688                          bs->bl.pwrite_zeroes_alignment);
 689     int err;
 690
 691     /* Only pass through requests that are larger than requested
 692      * preferred alignment (so that we test the fallback to writes on
 693      * unaligned portions), and check that the block layer never hands
 694      * us anything unaligned that crosses an alignment boundary.  */
 695     if (bytes < align) {
 696         assert(QEMU_IS_ALIGNED(offset, align) ||
 697                QEMU_IS_ALIGNED(offset + bytes, align) ||
 698                DIV_ROUND_UP(offset, align) ==
 699                DIV_ROUND_UP(offset + bytes, align));
 700         return -ENOTSUP;
 701     }
 702     assert(QEMU_IS_ALIGNED(offset, align));
 703     assert(QEMU_IS_ALIGNED(bytes, align));
 704     if (bs->bl.max_pwrite_zeroes) {
 705         assert(bytes <= bs->bl.max_pwrite_zeroes);
 706     }
 707
 708     err = rule_check(bs, offset, bytes, BLKDEBUG_IO_TYPE_WRITE_ZEROES);
 709     if (err) {
 710         return err;
 711     }
 712
 713     return bdrv_co_pwrite_zeroes(bs->file, offset, bytes, flags);
 714 }
 715
 716 static int coroutine_fn blkdebug_co_pdiscard(BlockDriverState *bs,
 717                                              int64_t offset, int64_t bytes)
 718 {
 719     uint32_t align = bs->bl.pdiscard_alignment;
 720     int err;
 721
 722     /* Only pass through requests that are larger than requested
 723      * minimum alignment, and ensure that unaligned requests do not
 724      * cross optimum discard boundaries. */
 725     if (bytes < bs->bl.request_alignment) {
 726         assert(QEMU_IS_ALIGNED(offset, align) ||
 727                QEMU_IS_ALIGNED(offset + bytes, align) ||
 728                DIV_ROUND_UP(offset, align) ==
 729                DIV_ROUND_UP(offset + bytes, align));
 730         return -ENOTSUP;
 731     }
 732     assert(QEMU_IS_ALIGNED(offset, bs->bl.request_alignment));
 733     assert(QEMU_IS_ALIGNED(bytes, bs->bl.request_alignment));
 734     if (align && bytes >= align) {
 735         assert(QEMU_IS_ALIGNED(offset, align));
 736         assert(QEMU_IS_ALIGNED(bytes, align));
 737     }
 738     if (bs->bl.max_pdiscard) {
 739         assert(bytes <= bs->bl.max_pdiscard);
 740     }
 741
 742     err = rule_check(bs, offset, bytes, BLKDEBUG_IO_TYPE_DISCARD);
 743     if (err) {
 744         return err;
 745     }
 746
 747     return bdrv_co_pdiscard(bs->file, offset, bytes);
 748 }
 749
 750 static int coroutine_fn blkdebug_co_block_status(BlockDriverState *bs,
 751                                                  bool want_zero,
 752                                                  int64_t offset,
 753                                                  int64_t bytes,
 754                                                  int64_t *pnum,
 755                                                  int64_t *map,
 756                                                  BlockDriverState **file)
 757 {
 758     int err;
 759
 760     assert(QEMU_IS_ALIGNED(offset | bytes, bs->bl.request_alignment));
 761
 762     err = rule_check(bs, offset, bytes, BLKDEBUG_IO_TYPE_BLOCK_STATUS);
 763     if (err) {
 764         return err;
 765     }
 766
 767     assert(bs->file && bs->file->bs);
 768     *pnum = bytes;
 769     *map = offset;
 770     *file = bs->file->bs;
 771     return BDRV_BLOCK_RAW | BDRV_BLOCK_OFFSET_VALID;
 772 }
 773
 774 static void blkdebug_close(BlockDriverState *bs)
 775 {
 776     BDRVBlkdebugState *s = bs->opaque;
 777     BlkdebugRule *rule, *next;
 778     int i;
 779
 780     for (i = 0; i < BLKDBG__MAX; i++) {
 781         QLIST_FOREACH_SAFE(rule, &s->rules[i], next, next) {
 782             remove_rule(rule);
 783         }
 784     }
 785
 786     g_free(s->config_file);
 787     qemu_mutex_destroy(&s->lock);
 788 }
 789
 790 /* Called with lock held.  */
 791 static void suspend_request(BlockDriverState *bs, BlkdebugRule *rule)
 792 {
 793     BDRVBlkdebugState *s = bs->opaque;
 794     BlkdebugSuspendedReq *r;
 795
 796     r = g_new(BlkdebugSuspendedReq, 1);
 797
 798     r->co         = qemu_coroutine_self();
 799     r->tag        = g_strdup(rule->options.suspend.tag);
 800
 801     remove_rule(rule);
 802     QLIST_INSERT_HEAD(&s->suspended_reqs, r, next);
 803
 804     if (!qtest_enabled()) {
 805         printf("blkdebug: Suspended request '%s'\n", r->tag);
 806     }
 807 }
 808
 809 /* Called with lock held.  */
 810 static void process_rule(BlockDriverState *bs, struct BlkdebugRule *rule,
 811                          int *action_count, int *new_state)
 812 {
 813     BDRVBlkdebugState *s = bs->opaque;
 814
 815     /* Only process rules for the current state */
 816     if (rule->state && rule->state != s->state) {
 817         return;
 818     }
 819
 820     /* Take the action */
 821     action_count[rule->action]++;
 822     switch (rule->action) {
 823     case ACTION_INJECT_ERROR:
 824         if (action_count[ACTION_INJECT_ERROR] == 1) {
 825             QSIMPLEQ_INIT(&s->active_rules);
 826         }
 827         QSIMPLEQ_INSERT_HEAD(&s->active_rules, rule, active_next);
 828         break;
 829
 830     case ACTION_SET_STATE:
 831         *new_state = rule->options.set_state.new_state;
 832         break;
 833
 834     case ACTION_SUSPEND:
 835         suspend_request(bs, rule);
 836         break;
 837     }
 838 }
 839
 840 static void blkdebug_debug_event(BlockDriverState *bs, BlkdebugEvent event)
 841 {
 842     BDRVBlkdebugState *s = bs->opaque;
 843     struct BlkdebugRule *rule, *next;
 844     int new_state;
 845     int actions_count[ACTION__MAX] = { 0 };
 846
 847     assert((int)event >= 0 && event < BLKDBG__MAX);
 848
 849     WITH_QEMU_LOCK_GUARD(&s->lock) {
 850         new_state = s->state;
 851         QLIST_FOREACH_SAFE(rule, &s->rules[event], next, next) {
 852             process_rule(bs, rule, actions_count, &new_state);
 853         }
 854         s->state = new_state;
 855     }
 856
 857     while (actions_count[ACTION_SUSPEND] > 0) {
 858         qemu_coroutine_yield();
 859         actions_count[ACTION_SUSPEND]--;
 860     }
 861 }
 862
 863 static int blkdebug_debug_breakpoint(BlockDriverState *bs, const char *event,
 864                                      const char *tag)
 865 {
 866     BDRVBlkdebugState *s = bs->opaque;
 867     struct BlkdebugRule *rule;
 868     int blkdebug_event;
 869
 870     blkdebug_event = qapi_enum_parse(&BlkdebugEvent_lookup, event, -1, NULL);
 871     if (blkdebug_event < 0) {
 872         return -ENOENT;
 873     }
 874
 875     rule = g_malloc(sizeof(*rule));
 876     *rule = (struct BlkdebugRule) {
 877         .event  = blkdebug_event,
 878         .action = ACTION_SUSPEND,
 879         .state  = 0,
 880         .options.suspend.tag = g_strdup(tag),
 881     };
 882
 883     qemu_mutex_lock(&s->lock);
 884     QLIST_INSERT_HEAD(&s->rules[blkdebug_event], rule, next);
 885     qemu_mutex_unlock(&s->lock);
 886
 887     return 0;
 888 }
 889
 890 /* Called with lock held. May temporarily release lock. */
 891 static int resume_req_by_tag(BDRVBlkdebugState *s, const char *tag, bool all)
 892 {
 893     BlkdebugSuspendedReq *r;
 894
 895 retry:
 896     /*
 897      * No need for _SAFE, since a different coroutine can remove another node
 898      * (not the current one) in this list, and when the current one is removed
 899      * the iteration starts back from beginning anyways.
 900      */
 901     QLIST_FOREACH(r, &s->suspended_reqs, next) {
 902         if (!strcmp(r->tag, tag)) {
 903             Coroutine *co = r->co;
 904
 905             if (!qtest_enabled()) {
 906                 printf("blkdebug: Resuming request '%s'\n", r->tag);
 907             }
 908
 909             QLIST_REMOVE(r, next);
 910             g_free(r->tag);
 911             g_free(r);
 912
 913             qemu_mutex_unlock(&s->lock);
 914             qemu_coroutine_enter(co);
 915             qemu_mutex_lock(&s->lock);
 916
 917             if (all) {
 918                 goto retry;
 919             }
 920             return 0;
 921         }
 922     }
 923     return -ENOENT;
 924 }
 925
 926 static int blkdebug_debug_resume(BlockDriverState *bs, const char *tag)
 927 {
 928     BDRVBlkdebugState *s = bs->opaque;
 929     QEMU_LOCK_GUARD(&s->lock);
 930     return resume_req_by_tag(s, tag, false);
 931 }
 932
 933 static int blkdebug_debug_remove_breakpoint(BlockDriverState *bs,
 934                                             const char *tag)
 935 {
 936     BDRVBlkdebugState *s = bs->opaque;
 937     BlkdebugRule *rule, *next;
 938     int i, ret = -ENOENT;
 939
 940     QEMU_LOCK_GUARD(&s->lock);
 941     for (i = 0; i < BLKDBG__MAX; i++) {
 942         QLIST_FOREACH_SAFE(rule, &s->rules[i], next, next) {
 943             if (rule->action == ACTION_SUSPEND &&
 944                 !strcmp(rule->options.suspend.tag, tag)) {
 945                 remove_rule(rule);
 946                 ret = 0;
 947             }
 948         }
 949     }
 950     if (resume_req_by_tag(s, tag, true) == 0) {
 951         ret = 0;
 952     }
 953     return ret;
 954 }
 955
 956 static bool blkdebug_debug_is_suspended(BlockDriverState *bs, const char *tag)
 957 {
 958     BDRVBlkdebugState *s = bs->opaque;
 959     BlkdebugSuspendedReq *r;
 960
 961     QEMU_LOCK_GUARD(&s->lock);
 962     QLIST_FOREACH(r, &s->suspended_reqs, next) {
 963         if (!strcmp(r->tag, tag)) {
 964             return true;
 965         }
 966     }
 967     return false;
 968 }
 969
 970 static int64_t blkdebug_getlength(BlockDriverState *bs)
 971 {
 972     return bdrv_getlength(bs->file->bs);
 973 }
 974
 975 static void blkdebug_refresh_filename(BlockDriverState *bs)
 976 {
 977     BDRVBlkdebugState *s = bs->opaque;
 978     const QDictEntry *e;
 979     int ret;
 980
 981     if (!bs->file->bs->exact_filename[0]) {
 982         return;
 983     }
 984
 985     for (e = qdict_first(bs->full_open_options); e;
 986          e = qdict_next(bs->full_open_options, e))
 987     {
 988         /* Real child options are under "image", but "x-image" may
 989          * contain a filename */
 990         if (strcmp(qdict_entry_key(e), "config") &&
 991             strcmp(qdict_entry_key(e), "image") &&
 992             strcmp(qdict_entry_key(e), "x-image") &&
 993             strcmp(qdict_entry_key(e), "driver"))
 994         {
 995             return;
 996         }
 997     }
 998
 999     ret = snprintf(bs->exact_filename, sizeof(bs->exact_filename),
1000                    "blkdebug:%s:%s",
1001                    s->config_file ?: "", bs->file->bs->exact_filename);
1002     if (ret >= sizeof(bs->exact_filename)) {
1003         /* An overflow makes the filename unusable, so do not report any */
1004         bs->exact_filename[0] = 0;
1005     }
1006 }
1007
1008 static void blkdebug_refresh_limits(BlockDriverState *bs, Error **errp)
1009 {
1010     BDRVBlkdebugState *s = bs->opaque;
1011
1012     if (s->align) {
1013         bs->bl.request_alignment = s->align;
1014     }
1015     if (s->max_transfer) {
1016         bs->bl.max_transfer = s->max_transfer;
1017     }
1018     if (s->opt_write_zero) {
1019         bs->bl.pwrite_zeroes_alignment = s->opt_write_zero;
1020     }
1021     if (s->max_write_zero) {
1022         bs->bl.max_pwrite_zeroes = s->max_write_zero;
1023     }
1024     if (s->opt_discard) {
1025         bs->bl.pdiscard_alignment = s->opt_discard;
1026     }
1027     if (s->max_discard) {
1028         bs->bl.max_pdiscard = s->max_discard;
1029     }
1030 }
1031
1032 static int blkdebug_reopen_prepare(BDRVReopenState *reopen_state,
1033                                    BlockReopenQueue *queue, Error **errp)
1034 {
1035     return 0;
1036 }
1037
1038 static void blkdebug_child_perm(BlockDriverState *bs, BdrvChild *c,
1039                                 BdrvChildRole role,
1040                                 BlockReopenQueue *reopen_queue,
1041                                 uint64_t perm, uint64_t shared,
1042                                 uint64_t *nperm, uint64_t *nshared)
1043 {
1044     BDRVBlkdebugState *s = bs->opaque;
1045
1046     bdrv_default_perms(bs, c, role, reopen_queue,
1047                        perm, shared, nperm, nshared);
1048
1049     *nperm |= s->take_child_perms;
1050     *nshared &= ~s->unshare_child_perms;
1051 }
1052
1053 static const char *const blkdebug_strong_runtime_opts[] = {
1054     "config",
1055     "inject-error.",
1056     "set-state.",
1057     "align",
1058     "max-transfer",
1059     "opt-write-zero",
1060     "max-write-zero",
1061     "opt-discard",
1062     "max-discard",
1063
1064     NULL
1065 };
1066
1067 static BlockDriver bdrv_blkdebug = {
1068     .format_name            = "blkdebug",
1069     .protocol_name          = "blkdebug",
1070     .instance_size          = sizeof(BDRVBlkdebugState),
1071     .is_filter              = true,
1072
1073     .bdrv_parse_filename    = blkdebug_parse_filename,
1074     .bdrv_file_open         = blkdebug_open,
1075     .bdrv_close             = blkdebug_close,
1076     .bdrv_reopen_prepare    = blkdebug_reopen_prepare,
1077     .bdrv_child_perm        = blkdebug_child_perm,
1078
1079     .bdrv_getlength         = blkdebug_getlength,
1080     .bdrv_refresh_filename  = blkdebug_refresh_filename,
1081     .bdrv_refresh_limits    = blkdebug_refresh_limits,
1082
1083     .bdrv_co_preadv         = blkdebug_co_preadv,
1084     .bdrv_co_pwritev        = blkdebug_co_pwritev,
1085     .bdrv_co_flush_to_disk  = blkdebug_co_flush,
1086     .bdrv_co_pwrite_zeroes  = blkdebug_co_pwrite_zeroes,
1087     .bdrv_co_pdiscard       = blkdebug_co_pdiscard,
1088     .bdrv_co_block_status   = blkdebug_co_block_status,
1089
1090     .bdrv_debug_event           = blkdebug_debug_event,
1091     .bdrv_debug_breakpoint      = blkdebug_debug_breakpoint,
1092     .bdrv_debug_remove_breakpoint
1093                                 = blkdebug_debug_remove_breakpoint,
1094     .bdrv_debug_resume          = blkdebug_debug_resume,
1095     .bdrv_debug_is_suspended    = blkdebug_debug_is_suspended,
1096
1097     .strong_runtime_opts        = blkdebug_strong_runtime_opts,
1098 };
1099
1100 static void bdrv_blkdebug_init(void)
1101 {
1102     bdrv_register(&bdrv_blkdebug);
1103 }
1104
1105 block_init(bdrv_blkdebug_init);