sys/vfs/hammer2/hammer2_cluster.c

   1 /*
   2  * Copyright (c) 2013-2015 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@dragonflybsd.org>
   6  *
   7  * Redistribution and use in source and binary forms, with or without
   8  * modification, are permitted provided that the following conditions
   9  * are met:
  10  *
  11  * 1. Redistributions of source code must retain the above copyright
  12  *    notice, this list of conditions and the following disclaimer.
  13  * 2. Redistributions in binary form must reproduce the above copyright
  14  *    notice, this list of conditions and the following disclaimer in
  15  *    the documentation and/or other materials provided with the
  16  *    distribution.
  17  * 3. Neither the name of The DragonFly Project nor the names of its
  18  *    contributors may be used to endorse or promote products derived
  19  *    from this software without specific, prior written permission.
  20  *
  21  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  22  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  23  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  24  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  25  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  26  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  27  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  28  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  29  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  30  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  31  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  32  * SUCH DAMAGE.
  33  */
  34 /*
  35  * The cluster module collects multiple chains representing the same
  36  * information from different nodes into a single entity.  It allows direct
  37  * access to media data as long as it is not blockref array data (which
  38  * will obviously have to be different at each node).
  39  *
  40  * This module also handles I/O dispatch, status rollup, and various
  41  * mastership arrangements including quorum operations.  It effectively
  42  * presents one topology to the vnops layer.
  43  *
  44  * Many of the API calls mimic chain API calls but operate on clusters
  45  * instead of chains.  Please see hammer2_chain.c for more complete code
  46  * documentation of the API functions.
  47  *
  48  * WARNING! This module is *extremely* complex.  It must issue asynchronous
  49  *          locks and I/O, do quorum and/or master-slave processing, and
  50  *          it must operate properly even if some nodes are broken (which
  51  *          can also mean indefinite locks).
  52  *
  53  *                              CLUSTER OPERATIONS
  54  *
  55  * Cluster operations can be broken down into three pieces:
  56  *
  57  * (1) Chain locking and data retrieval.
  58  *
  59  *      - Most complex functions, quorum management on transaction ids.
  60  *
  61  *      - Locking and data accesses must be internally asynchronous.
  62  *
  63  *      - Validate and manage cache coherency primitives (cache state
  64  *        is stored in chain topologies but must be validated by these
  65  *        functions).
  66  *
  67  * (2) Lookups and Scans
  68  *              hammer2_cluster_lookup()
  69  *              hammer2_cluster_next()
  70  *
  71  *      - Depend on locking & data retrieval functions, but still complex.
  72  *
  73  *      - Must do quorum management on transaction ids.
  74  *
  75  *      - Lookup and Iteration ops Must be internally asynchronous.
  76  *
  77  * (3) Modifying Operations
  78  *              hammer2_cluster_create()
  79  *
  80  *      - Can usually punt on failures, operation continues unless quorum
  81  *        is lost.  If quorum is lost, must wait for resynchronization
  82  *        (depending on the management mode).
  83  *
  84  *      - Must disconnect node on failures (also not flush), remount, and
  85  *        resynchronize.
  86  *
  87  *      - Network links (via kdmsg) are relatively easy to issue as the
  88  *        complex underworkings of hammer2_chain.c don't have to messed
  89  *        with (the protocol is at a higher level than block-level).
  90  *
  91  *      - Multiple local disk nodes (i.e. block devices) are another matter.
  92  *        Chain operations have to be dispatched to per-node threads (xN)
  93  *        because we can't asynchronize potentially very complex chain
  94  *        operations in hammer2_chain.c (it would be a huge mess).
  95  *
  96  *        (these threads are also used to terminate incoming kdmsg ops from
  97  *        other machines).
  98  *
  99  *      - Single-node filesystems do not use threads and will simply call
 100  *        hammer2_chain.c functions directly.  This short-cut is handled
 101  *        at the base of each cluster function.
 102  */
 103 #include <sys/cdefs.h>
 104 #include <sys/param.h>
 105 #include <sys/systm.h>
 106 #include <sys/types.h>
 107 #include <sys/lock.h>
 108 #include <sys/uuid.h>
 109
 110 #include "hammer2.h"
 111
 112 /*
 113  * Returns the bref type of the cluster's foucs.
 114  *
 115  * If the cluster is errored, returns HAMMER2_BREF_TYPE_EMPTY (0).
 116  * The cluster must be locked.
 117  */
 118 uint8_t
 119 hammer2_cluster_type(hammer2_cluster_t *cluster)
 120 {
 121         if (cluster->error == 0) {
 122                 KKASSERT(cluster->focus != NULL);
 123                 return(cluster->focus->bref.type);
 124         }
 125         return 0;
 126 }
 127
 128 /*
 129  * Returns non-zero if the cluster's focus is flagged as being modified.
 130  *
 131  * If the cluster is errored, returns 0.
 132  */
 133 static
 134 int
 135 hammer2_cluster_modified(hammer2_cluster_t *cluster)
 136 {
 137         if (cluster->error == 0) {
 138                 KKASSERT(cluster->focus != NULL);
 139                 return((cluster->focus->flags & HAMMER2_CHAIN_MODIFIED) != 0);
 140         }
 141         return 0;
 142 }
 143
 144 /*
 145  * Returns the bref of the cluster's focus, sans any data-offset information
 146  * (since offset information is per-node and wouldn't be useful).
 147  *
 148  * Callers use this function to access modify_tid, mirror_tid, type,
 149  * key, and keybits.
 150  *
 151  * If the cluster is errored, returns an empty bref.
 152  * The cluster must be locked.
 153  */
 154 void
 155 hammer2_cluster_bref(hammer2_cluster_t *cluster, hammer2_blockref_t *bref)
 156 {
 157         if (cluster->error == 0) {
 158                 KKASSERT(cluster->focus != NULL);
 159                 *bref = cluster->focus->bref;
 160                 bref->data_off = 0;
 161         } else {
 162                 bzero(bref, sizeof(*bref));
 163         }
 164 }
 165
 166 /*
 167  * Create a degenerate cluster with one ref from a single locked chain.
 168  * The returned cluster will be focused on the chain and inherit its
 169  * error state.
 170  *
 171  * The chain's lock and reference are transfered to the new cluster, so
 172  * the caller should not try to unlock the chain separately.
 173  *
 174  * We fake the flags.
 175  */
 176 hammer2_cluster_t *
 177 hammer2_cluster_from_chain(hammer2_chain_t *chain)
 178 {
 179         hammer2_cluster_t *cluster;
 180
 181         cluster = kmalloc(sizeof(*cluster), M_HAMMER2, M_WAITOK | M_ZERO);
 182         cluster->array[0].chain = chain;
 183         cluster->array[0].flags = HAMMER2_CITEM_FEMOD;
 184         cluster->nchains = 1;
 185         cluster->focus = chain;
 186         cluster->focus_index = 0;
 187         cluster->pmp = chain->pmp;
 188         cluster->refs = 1;
 189         cluster->error = chain->error;
 190         cluster->flags = HAMMER2_CLUSTER_LOCKED |
 191                          HAMMER2_CLUSTER_WRHARD |
 192                          HAMMER2_CLUSTER_RDHARD |
 193                          HAMMER2_CLUSTER_MSYNCED |
 194                          HAMMER2_CLUSTER_SSYNCED;
 195
 196         return cluster;
 197 }
 198
 199 /*
 200  * Add a reference to a cluster and its underlying chains.
 201  *
 202  * We must also ref the underlying chains in order to allow ref/unlock
 203  * sequences to later re-lock.
 204  */
 205 void
 206 hammer2_cluster_ref(hammer2_cluster_t *cluster)
 207 {
 208         atomic_add_int(&cluster->refs, 1);
 209 }
 210
 211 /*
 212  * Drop the caller's reference to the cluster.  When the ref count drops to
 213  * zero this function frees the cluster and drops all underlying chains.
 214  *
 215  * In-progress read I/Os are typically detached from the cluster once the
 216  * first one returns (the remaining stay attached to the DIOs but are then
 217  * ignored and drop naturally).
 218  */
 219 void
 220 hammer2_cluster_drop(hammer2_cluster_t *cluster)
 221 {
 222         hammer2_chain_t *chain;
 223         int i;
 224
 225         KKASSERT(cluster->refs > 0);
 226         if (atomic_fetchadd_int(&cluster->refs, -1) == 1) {
 227                 cluster->focus = NULL;          /* safety XXX chg to assert */
 228                 cluster->focus_index = 0;
 229
 230                 for (i = 0; i < cluster->nchains; ++i) {
 231                         chain = cluster->array[i].chain;
 232                         if (chain) {
 233                                 hammer2_chain_drop(chain);
 234                                 cluster->array[i].chain = NULL; /* safety */
 235                         }
 236                 }
 237                 cluster->nchains = 0;                           /* safety */
 238
 239                 kfree(cluster, M_HAMMER2);
 240                 /* cluster is invalid */
 241         }
 242 }
 243
 244 /*
 245  * Lock a cluster.  Cluster must already be referenced.  Focus is maintained.
 246  *
 247  * WARNING! This function expects the caller to handle resolution of the
 248  *          cluster.  We never re-resolve the cluster in this function,
 249  *          because it might be used to temporarily unlock/relock a cparent
 250  *          in an iteration or recursrion, and the cparents elements do not
 251  *          necessarily match.
 252  */
 253 void
 254 hammer2_cluster_lock(hammer2_cluster_t *cluster, int how)
 255 {
 256         hammer2_chain_t *chain;
 257         int i;
 258
 259         /* cannot be on inode-embedded cluster template, must be on copy */
 260         KKASSERT(cluster->refs > 0);
 261         KKASSERT((cluster->flags & HAMMER2_CLUSTER_INODE) == 0);
 262         if (cluster->flags & HAMMER2_CLUSTER_LOCKED) {
 263                 panic("hammer2_cluster_lock: cluster %p already locked!\n",
 264                         cluster);
 265         }
 266         atomic_set_int(&cluster->flags, HAMMER2_CLUSTER_LOCKED);
 267
 268         /*
 269          * Lock chains and resolve state.
 270          */
 271         for (i = 0; i < cluster->nchains; ++i) {
 272                 chain = cluster->array[i].chain;
 273                 if (chain == NULL)
 274                         continue;
 275                 hammer2_chain_lock(chain, how);
 276         }
 277 }
 278
 279 /*
 280  * Calculate the clustering state for the cluster and set its focus.
 281  * This routine must be called with care.  For example, it should not
 282  * normally be called after relocking a non-leaf cluster because parent
 283  * clusters help iterations and each element might be at a slightly different
 284  * indirect node (each node's topology is independently indexed).
 285  *
 286  * HAMMER2_CITEM_FEMOD flags which elements can be modified by normal
 287  * operations.  Typically this is only set on a quorum of MASTERs or
 288  * on a SOFT_MASTER.  Also as a degenerate case on SUPROOT.  If a SOFT_MASTER
 289  * is present, this bit is *not* set on a quorum of MASTERs.  The
 290  * synchronization code ignores this bit, but all hammer2_cluster_*() calls
 291  * that create/modify/delete elements use it.
 292  *
 293  * The chains making up the cluster may be narrowed down based on quorum
 294  * acceptability, and if RESOLVE_RDONLY is specified the chains can be
 295  * narrowed down to a single chain as long as the entire subtopology is known
 296  * to be intact.  So, for example, we can narrow a read-only op to a single
 297  * fast SLAVE but if we focus a CACHE chain we must still retain at least
 298  * a SLAVE to ensure that the subtopology can be accessed.
 299  *
 300  * RESOLVE_RDONLY operations are effectively as-of so the quorum does not need
 301  * to be maintained once the topology is validated as-of the top level of
 302  * the operation.
 303  *
 304  * If a failure occurs the operation must be aborted by higher-level code and
 305  * retried. XXX
 306  */
 307 void
 308 hammer2_cluster_resolve(hammer2_cluster_t *cluster)
 309 {
 310         hammer2_chain_t *chain;
 311         hammer2_chain_t *focus;
 312         hammer2_pfs_t *pmp;
 313         hammer2_tid_t quorum_tid;
 314         hammer2_tid_t last_best_quorum_tid;
 315         int focus_pfs_type;
 316         uint32_t nflags;
 317         int ttlmasters;
 318         int ttlslaves;
 319         int nmasters;
 320         int nslaves;
 321         int nquorum;
 322         int smpresent;
 323         int i;
 324
 325         cluster->error = 0;
 326         cluster->focus = NULL;
 327
 328         focus_pfs_type = 0;
 329         nflags = 0;
 330         ttlmasters = 0;
 331         ttlslaves = 0;
 332         nmasters = 0;
 333         nslaves = 0;
 334
 335         /*
 336          * Calculate quorum
 337          */
 338         pmp = cluster->pmp;
 339         KKASSERT(pmp != NULL || cluster->nchains == 0);
 340         nquorum = pmp ? pmp->pfs_nmasters / 2 + 1 : 0;
 341         smpresent = 0;
 342
 343         /*
 344          * Pass 1
 345          *
 346          * NOTE: A NULL chain is not necessarily an error, it could be
 347          *       e.g. a lookup failure or the end of an iteration.
 348          *       Process normally.
 349          */
 350         for (i = 0; i < cluster->nchains; ++i) {
 351                 chain = cluster->array[i].chain;
 352                 if (chain && chain->error) {
 353                         if (cluster->focus == NULL || cluster->focus == chain) {
 354                                 /* error will be overridden by valid focus */
 355                                 cluster->error = chain->error;
 356                         }
 357
 358                         /*
 359                          * Must count total masters and slaves whether the
 360                          * chain is errored or not.
 361                          */
 362                         switch (cluster->pmp->pfs_types[i]) {
 363                         case HAMMER2_PFSTYPE_MASTER:
 364                                 ++ttlmasters;
 365                                 break;
 366                         case HAMMER2_PFSTYPE_SLAVE:
 367                                 ++ttlslaves;
 368                                 break;
 369                         }
 370                         continue;
 371                 }
 372                 switch (cluster->pmp->pfs_types[i]) {
 373                 case HAMMER2_PFSTYPE_MASTER:
 374                         ++ttlmasters;
 375                         break;
 376                 case HAMMER2_PFSTYPE_SLAVE:
 377                         ++ttlslaves;
 378                         break;
 379                 case HAMMER2_PFSTYPE_SOFT_MASTER:
 380                         nflags |= HAMMER2_CLUSTER_WRSOFT;
 381                         nflags |= HAMMER2_CLUSTER_RDSOFT;
 382                         smpresent = 1;
 383                         break;
 384                 case HAMMER2_PFSTYPE_SOFT_SLAVE:
 385                         nflags |= HAMMER2_CLUSTER_RDSOFT;
 386                         break;
 387                 case HAMMER2_PFSTYPE_SUPROOT:
 388                         /*
 389                          * Degenerate cluster representing the super-root
 390                          * topology on a single device.  Fake stuff so
 391                          * cluster ops work as expected.
 392                          */
 393                         nflags |= HAMMER2_CLUSTER_WRHARD;
 394                         nflags |= HAMMER2_CLUSTER_RDHARD;
 395                         cluster->focus_index = i;
 396                         cluster->focus = chain;
 397                         cluster->error = chain ? chain->error : 0;
 398                         break;
 399                 default:
 400                         break;
 401                 }
 402         }
 403
 404         /*
 405          * Pass 2
 406          *
 407          * Resolve masters.  Calculate nmasters for the highest matching
 408          * TID, if a quorum cannot be attained try the next lower matching
 409          * TID until we exhaust TIDs.
 410          *
 411          * NOTE: A NULL chain is not necessarily an error, it could be
 412          *       e.g. a lookup failure or the end of an iteration.
 413          *       Process normally.
 414          */
 415         last_best_quorum_tid = HAMMER2_TID_MAX;
 416         quorum_tid = 0;         /* fix gcc warning */
 417
 418         while (nmasters < nquorum && last_best_quorum_tid != 0) {
 419                 nmasters = 0;
 420                 quorum_tid = 0;
 421
 422                 for (i = 0; i < cluster->nchains; ++i) {
 423                         if (cluster->pmp->pfs_types[i] !=
 424                             HAMMER2_PFSTYPE_MASTER) {
 425                                 continue;
 426                         }
 427                         chain = cluster->array[i].chain;
 428
 429                         if (cluster->array[i].flags & HAMMER2_CITEM_INVALID) {
 430                                 /*
 431                                  * Invalid as in unsynchronized, cannot be
 432                                  * used to calculate the quorum.
 433                                  */
 434                         } else if (chain == NULL && quorum_tid == 0) {
 435                                 /*
 436                                  * NULL chain on master matches NULL chains
 437                                  * on other masters.
 438                                  */
 439                                 ++nmasters;
 440                         } else if (quorum_tid < last_best_quorum_tid &&
 441                                    chain != NULL &&
 442                                    (quorum_tid < chain->bref.modify_tid ||
 443                                     nmasters == 0)) {
 444                                 /*
 445                                  * Better TID located, reset nmasters count.
 446                                  */
 447                                 nmasters = 1;
 448                                 quorum_tid = chain->bref.modify_tid;
 449                         } else if (chain &&
 450                                    quorum_tid == chain->bref.modify_tid) {
 451                                 /*
 452                                  * TID matches current collection.
 453                                  */
 454                                 ++nmasters;
 455                         }
 456                 }
 457                 if (nmasters >= nquorum)
 458                         break;
 459                 last_best_quorum_tid = quorum_tid;
 460         }
 461
 462         /*
 463          * Pass 3
 464          *
 465          * NOTE: A NULL chain is not necessarily an error, it could be
 466          *       e.g. a lookup failure or the end of an iteration.
 467          *       Process normally.
 468          */
 469         for (i = 0; i < cluster->nchains; ++i) {
 470                 cluster->array[i].flags &= ~HAMMER2_CITEM_FEMOD;
 471                 chain = cluster->array[i].chain;
 472                 if (chain && chain->error) {
 473                         if (cluster->focus == NULL || cluster->focus == chain) {
 474                                 /* error will be overridden by valid focus */
 475                                 cluster->error = chain->error;
 476                         }
 477                         continue;
 478                 }
 479
 480                 switch (cluster->pmp->pfs_types[i]) {
 481                 case HAMMER2_PFSTYPE_MASTER:
 482                         /*
 483                          * We must have enough up-to-date masters to reach
 484                          * a quorum and the master modify_tid must match
 485                          * the quorum's modify_tid.
 486                          *
 487                          * Do not select an errored or out-of-sync master.
 488                          */
 489                         if (cluster->array[i].flags & HAMMER2_CITEM_INVALID) {
 490                                 nflags |= HAMMER2_CLUSTER_UNHARD;
 491                         } else if (nmasters >= nquorum &&
 492                                    (chain == NULL || chain->error == 0) &&
 493                                    ((chain == NULL && quorum_tid == 0) ||
 494                                     (chain != NULL && quorum_tid ==
 495                                                   chain->bref.modify_tid))) {
 496                                 nflags |= HAMMER2_CLUSTER_WRHARD;
 497                                 nflags |= HAMMER2_CLUSTER_RDHARD;
 498                                 if (!smpresent) {
 499                                         cluster->array[i].flags |=
 500                                                         HAMMER2_CITEM_FEMOD;
 501                                 }
 502                                 if (cluster->focus == NULL ||
 503                                     focus_pfs_type == HAMMER2_PFSTYPE_SLAVE) {
 504                                         focus_pfs_type = HAMMER2_PFSTYPE_MASTER;
 505                                         cluster->focus_index = i;
 506                                         cluster->focus = chain; /* NULL ok */
 507                                         cluster->error = chain ? chain->error :
 508                                                                  0;
 509                                 }
 510                         } else if (chain == NULL || chain->error == 0) {
 511                                 nflags |= HAMMER2_CLUSTER_UNHARD;
 512                         }
 513                         break;
 514                 case HAMMER2_PFSTYPE_SLAVE:
 515                         /*
 516                          * We must have enough up-to-date masters to reach
 517                          * a quorum and the slave modify_tid must match the
 518                          * quorum's modify_tid.
 519                          *
 520                          * Do not select an errored slave.
 521                          */
 522                         if (cluster->array[i].flags & HAMMER2_CITEM_INVALID) {
 523                                 nflags |= HAMMER2_CLUSTER_UNHARD;
 524                         } else if (nmasters >= nquorum &&
 525                                    (chain == NULL || chain->error == 0) &&
 526                                    ((chain == NULL && quorum_tid == 0) ||
 527                                     (chain && quorum_tid ==
 528                                               chain->bref.modify_tid))) {
 529                                 ++nslaves;
 530                                 nflags |= HAMMER2_CLUSTER_RDHARD;
 531 #if 0
 532                                 /* XXX optimize for RESOLVE_RDONLY */
 533                                 if (cluster->focus == NULL) {
 534                                         focus_pfs_type = HAMMER2_PFSTYPE_SLAVE;
 535                                         cluster->focus_index = i;
 536                                         cluster->focus = chain; /* NULL ok */
 537                                         cluster->error = chain ? chain->error :
 538                                                                  0;
 539                                 }
 540 #endif
 541                         } else if (chain == NULL || chain->error == 0) {
 542                                 nflags |= HAMMER2_CLUSTER_UNSOFT;
 543                         }
 544                         break;
 545                 case HAMMER2_PFSTYPE_SOFT_MASTER:
 546                         /*
 547                          * Directly mounted soft master always wins.  There
 548                          * should be only one.
 549                          */
 550                         KKASSERT(focus_pfs_type != HAMMER2_PFSTYPE_SOFT_MASTER);
 551                         cluster->focus_index = i;
 552                         cluster->focus = chain;
 553                         cluster->error = chain ? chain->error : 0;
 554                         focus_pfs_type = HAMMER2_PFSTYPE_SOFT_MASTER;
 555                         cluster->array[i].flags |= HAMMER2_CITEM_FEMOD;
 556                         break;
 557                 case HAMMER2_PFSTYPE_SOFT_SLAVE:
 558                         /*
 559                          * Directly mounted soft slave always wins.  There
 560                          * should be only one.
 561                          */
 562                         KKASSERT(focus_pfs_type != HAMMER2_PFSTYPE_SOFT_SLAVE);
 563                         if (focus_pfs_type != HAMMER2_PFSTYPE_SOFT_MASTER) {
 564                                 cluster->focus_index = i;
 565                                 cluster->focus = chain;
 566                                 cluster->error = chain ? chain->error : 0;
 567                                 focus_pfs_type = HAMMER2_PFSTYPE_SOFT_SLAVE;
 568                         }
 569                         break;
 570                 case HAMMER2_PFSTYPE_SUPROOT:
 571                         /*
 572                          * spmp (degenerate case)
 573                          */
 574                         KKASSERT(i == 0);
 575                         cluster->focus_index = i;
 576                         cluster->focus = chain;
 577                         cluster->error = chain ? chain->error : 0;
 578                         focus_pfs_type = HAMMER2_PFSTYPE_SUPROOT;
 579                         cluster->array[i].flags |= HAMMER2_CITEM_FEMOD;
 580                         break;
 581                 default:
 582                         break;
 583                 }
 584         }
 585
 586         /*
 587          * Focus now set, adjust ddflag.  Skip this pass if the focus
 588          * is bad or if we are at the PFS root (the bref won't match at
 589          * the PFS root, obviously).
 590          */
 591         focus = cluster->focus;
 592         if (focus) {
 593                 cluster->ddflag =
 594                         (cluster->focus->bref.type == HAMMER2_BREF_TYPE_INODE);
 595         } else {
 596                 cluster->ddflag = 0;
 597                 goto skip4;
 598         }
 599         if (cluster->focus->flags & HAMMER2_CHAIN_PFSBOUNDARY)
 600                 goto skip4;
 601
 602         /*
 603          * Pass 4
 604          *
 605          * Validate the elements that were not marked invalid.  They should
 606          * match.
 607          */
 608         for (i = 0; i < cluster->nchains; ++i) {
 609                 int ddflag;
 610
 611                 chain = cluster->array[i].chain;
 612
 613                 if (chain == NULL)
 614                         continue;
 615                 if (chain == focus)
 616                         continue;
 617                 if (cluster->array[i].flags & HAMMER2_CITEM_INVALID)
 618                         continue;
 619
 620                 ddflag = (chain->bref.type == HAMMER2_BREF_TYPE_INODE);
 621                 if (chain->bref.type != focus->bref.type ||
 622                     chain->bref.key != focus->bref.key ||
 623                     chain->bref.keybits != focus->bref.keybits ||
 624                     chain->bref.modify_tid != focus->bref.modify_tid ||
 625                     chain->bytes != focus->bytes ||
 626                     ddflag != cluster->ddflag) {
 627                         cluster->array[i].flags |= HAMMER2_CITEM_INVALID;
 628                         if (hammer2_debug & 1)
 629                         kprintf("cluster_resolve: matching modify_tid failed "
 630                                 "bref test: idx=%d type=%02x/%02x "
 631                                 "key=%016jx/%d-%016jx/%d "
 632                                 "mod=%016jx/%016jx bytes=%u/%u\n",
 633                                 i,
 634                                 chain->bref.type, focus->bref.type,
 635                                 chain->bref.key, chain->bref.keybits,
 636                                 focus->bref.key, focus->bref.keybits,
 637                                 chain->bref.modify_tid, focus->bref.modify_tid,
 638                                 chain->bytes, focus->bytes);
 639                         if (hammer2_debug & 0x4000)
 640                                 panic("cluster_resolve");
 641                         /* flag issue and force resync? */
 642                 }
 643         }
 644 skip4:
 645
 646         if (ttlslaves == 0)
 647                 nflags |= HAMMER2_CLUSTER_NOSOFT;
 648         if (ttlmasters == 0)
 649                 nflags |= HAMMER2_CLUSTER_NOHARD;
 650
 651         /*
 652          * Set SSYNCED or MSYNCED for slaves and masters respectively if
 653          * all available nodes (even if 0 are available) are fully
 654          * synchronized.  This is used by the synchronization thread to
 655          * determine if there is work it could potentially accomplish.
 656          */
 657         if (nslaves == ttlslaves)
 658                 nflags |= HAMMER2_CLUSTER_SSYNCED;
 659         if (nmasters == ttlmasters)
 660                 nflags |= HAMMER2_CLUSTER_MSYNCED;
 661
 662         /*
 663          * Determine if the cluster was successfully locked for the
 664          * requested operation and generate an error code.  The cluster
 665          * will not be locked (or ref'd) if an error is returned.
 666          */
 667         atomic_set_int(&cluster->flags, nflags);
 668         atomic_clear_int(&cluster->flags, HAMMER2_CLUSTER_ZFLAGS & ~nflags);
 669 }
 670
 671 /*
 672  * This is used by the XOPS subsystem to calculate the state of
 673  * the collection and tell hammer2_xop_collect() what to do with it.
 674  * The collection can be in various states of desynchronization, the
 675  * caller specifically wants to resolve the passed-in key.
 676  *
 677  * Return values:
 678  *      0               - Quorum agreement, key is valid
 679  *
 680  *      ENOENT          - Quorum agreement, end of scan
 681  *
 682  *      ESRCH           - Quorum agreement, key is INVALID (caller should
 683  *                        skip key).
 684  *
 685  *      EIO             - Quorum agreement but all elements had errors.
 686  *
 687  *      EDEADLK         - No quorum agreement possible for key, a repair
 688  *                        may be needed.  Caller has to decide what to do,
 689  *                        possibly iterating the key or generating an EIO.
 690  *
 691  *      EINPROGRESS     - No quorum agreement yet, but agreement is still
 692  *                        possible if caller waits for more responses.  Caller
 693  *                        should not iterate key.
 694  *
 695  * XXX needs to handle SOFT_MASTER and SOFT_SLAVE
 696  */
 697 int
 698 hammer2_cluster_check(hammer2_cluster_t *cluster, hammer2_key_t key, int flags)
 699 {
 700         hammer2_chain_t *chain;
 701         hammer2_chain_t *focus;
 702         hammer2_pfs_t *pmp;
 703         hammer2_tid_t quorum_tid;
 704         hammer2_tid_t last_best_quorum_tid;
 705         uint32_t nflags;
 706         int ttlmasters;
 707         int ttlslaves;
 708         int nmasters;
 709         int nmasters_keymatch;
 710         int nslaves;
 711         int nquorum;
 712         int umasters;   /* unknown masters (still in progress) */
 713         int smpresent;
 714         int i;
 715
 716         cluster->error = 0;
 717         cluster->focus = NULL;
 718
 719         nflags = 0;
 720         ttlmasters = 0;
 721         ttlslaves = 0;
 722         nmasters = 0;
 723         nmasters_keymatch = 0;
 724         umasters = 0;
 725         nslaves = 0;
 726
 727         /*
 728          * Calculate quorum
 729          */
 730         pmp = cluster->pmp;
 731         KKASSERT(pmp != NULL || cluster->nchains == 0);
 732         nquorum = pmp ? pmp->pfs_nmasters / 2 + 1 : 0;
 733         smpresent = 0;
 734
 735         /*
 736          * Pass 1
 737          *
 738          * NOTE: A NULL chain is not necessarily an error, it could be
 739          *       e.g. a lookup failure or the end of an iteration.
 740          *       Process normally.
 741          */
 742         for (i = 0; i < cluster->nchains; ++i) {
 743                 cluster->array[i].flags &= ~HAMMER2_CITEM_FEMOD;
 744                 cluster->array[i].flags |= HAMMER2_CITEM_INVALID;
 745
 746                 chain = cluster->array[i].chain;
 747                 if (chain && chain->error) {
 748                         if (cluster->focus == NULL || cluster->focus == chain) {
 749                                 /* error will be overridden by valid focus */
 750                                 cluster->error = chain->error;
 751                         }
 752
 753                         /*
 754                          * Must count total masters and slaves whether the
 755                          * chain is errored or not.
 756                          */
 757                         switch (cluster->pmp->pfs_types[i]) {
 758                         case HAMMER2_PFSTYPE_MASTER:
 759                                 ++ttlmasters;
 760                                 break;
 761                         case HAMMER2_PFSTYPE_SLAVE:
 762                                 ++ttlslaves;
 763                                 break;
 764                         }
 765                         continue;
 766                 }
 767                 switch (cluster->pmp->pfs_types[i]) {
 768                 case HAMMER2_PFSTYPE_MASTER:
 769                         ++ttlmasters;
 770                         break;
 771                 case HAMMER2_PFSTYPE_SLAVE:
 772                         ++ttlslaves;
 773                         break;
 774                 case HAMMER2_PFSTYPE_SOFT_MASTER:
 775                         nflags |= HAMMER2_CLUSTER_WRSOFT;
 776                         nflags |= HAMMER2_CLUSTER_RDSOFT;
 777                         smpresent = 1;
 778                         break;
 779                 case HAMMER2_PFSTYPE_SOFT_SLAVE:
 780                         nflags |= HAMMER2_CLUSTER_RDSOFT;
 781                         break;
 782                 case HAMMER2_PFSTYPE_SUPROOT:
 783                         /*
 784                          * Degenerate cluster representing the super-root
 785                          * topology on a single device.  Fake stuff so
 786                          * cluster ops work as expected.
 787                          */
 788                         nflags |= HAMMER2_CLUSTER_WRHARD;
 789                         nflags |= HAMMER2_CLUSTER_RDHARD;
 790                         cluster->focus_index = i;
 791                         cluster->focus = chain;
 792                         cluster->error = chain ? chain->error : 0;
 793                         break;
 794                 default:
 795                         break;
 796                 }
 797         }
 798
 799         /*
 800          * Pass 2
 801          *
 802          * Resolve nmasters             - master nodes fully match
 803          *
 804          * Resolve umasters             - master nodes operation still
 805          *                                in progress
 806          *
 807          * Resolve nmasters_keymatch    - master nodes match the passed-in
 808          *                                key and may or may not match
 809          *                                the quorum-agreed tid.
 810          *
 811          * The quorum-agreed TID is the highest matching TID.
 812          */
 813         last_best_quorum_tid = HAMMER2_TID_MAX;
 814         quorum_tid = 0;         /* fix gcc warning */
 815
 816         while (nmasters < nquorum && last_best_quorum_tid != 0) {
 817                 nmasters = 0;
 818                 quorum_tid = 0;
 819
 820                 for (i = 0; i < cluster->nchains; ++i) {
 821                         /* XXX SOFT smpresent handling */
 822                         if (cluster->pmp->pfs_types[i] !=
 823                             HAMMER2_PFSTYPE_MASTER) {
 824                                 continue;
 825                         }
 826
 827                         chain = cluster->array[i].chain;
 828
 829                         /*
 830                          * Skip elements still in progress.  umasters keeps
 831                          * track of masters that might still be in-progress.
 832                          */
 833                         if (chain == NULL && (cluster->array[i].flags &
 834                                               HAMMER2_CITEM_NULL) == 0) {
 835                                 ++umasters;
 836                                 continue;
 837                         }
 838
 839                         /*
 840                          * Key match?
 841                          */
 842                         if (flags & HAMMER2_CHECK_NULL) {
 843                                 if (chain == NULL) {
 844                                         ++nmasters;
 845                                         ++nmasters_keymatch;
 846                                 }
 847                         } else if (chain &&
 848                                    (key == (hammer2_key_t)-1 ||
 849                                     chain->bref.key == key)) {
 850                                 ++nmasters_keymatch;
 851                                 if (quorum_tid < last_best_quorum_tid &&
 852                                     (quorum_tid < chain->bref.modify_tid ||
 853                                      nmasters == 0)) {
 854                                         /*
 855                                          * Better TID located, reset
 856                                          * nmasters count.
 857                                          */
 858                                         nmasters = 0;
 859                                         quorum_tid = chain->bref.modify_tid;
 860                                 }
 861                                 if (quorum_tid == chain->bref.modify_tid) {
 862                                         /*
 863                                          * TID matches current collection.
 864                                          */
 865                                         ++nmasters;
 866                                         if (chain->error == 0) {
 867                                                 cluster->focus = chain;
 868                                                 cluster->focus_index = i;
 869                                         }
 870                                 }
 871                         }
 872                 }
 873                 if (nmasters >= nquorum)
 874                         break;
 875                 last_best_quorum_tid = quorum_tid;
 876         }
 877
 878         /*
 879         kprintf("nmasters %d/%d nmaster_keymatch=%d umasters=%d\n",
 880                 nmasters, nquorum, nmasters_keymatch, umasters);
 881         */
 882
 883         /*
 884          * Early return if we do not have enough masters.
 885          */
 886         if (nmasters < nquorum) {
 887                 if (nmasters + umasters >= nquorum)
 888                         return EINPROGRESS;
 889                 if (nmasters_keymatch < nquorum)
 890                         return ESRCH;
 891                 return EDEADLK;
 892         }
 893
 894         /*
 895          * Validated end of scan.
 896          */
 897         if (flags & HAMMER2_CHECK_NULL)
 898                 return ENOENT;
 899
 900         /*
 901          * If we have a NULL focus at this point the agreeing quorum all
 902          * had chain errors.
 903          */
 904         if (cluster->focus == NULL)
 905                 return EIO;
 906
 907         /*
 908          * Pass 3
 909          *
 910          * We have quorum agreement, validate elements, not end of scan.
 911          */
 912         for (i = 0; i < cluster->nchains; ++i) {
 913                 chain = cluster->array[i].chain;
 914                 if (chain == NULL ||
 915                     chain->bref.key != key ||
 916                     chain->bref.modify_tid != quorum_tid) {
 917                         continue;
 918                 }
 919
 920                 switch (cluster->pmp->pfs_types[i]) {
 921                 case HAMMER2_PFSTYPE_MASTER:
 922                         cluster->array[i].flags |= HAMMER2_CITEM_FEMOD;
 923                         cluster->array[i].flags &= ~HAMMER2_CITEM_INVALID;
 924                         nflags |= HAMMER2_CLUSTER_WRHARD;
 925                         nflags |= HAMMER2_CLUSTER_RDHARD;
 926                         break;
 927                 case HAMMER2_PFSTYPE_SLAVE:
 928                         /*
 929                          * We must have enough up-to-date masters to reach
 930                          * a quorum and the slave modify_tid must match the
 931                          * quorum's modify_tid.
 932                          *
 933                          * Do not select an errored slave.
 934                          */
 935                         cluster->array[i].flags &= ~HAMMER2_CITEM_INVALID;
 936                         nflags |= HAMMER2_CLUSTER_RDHARD;
 937                         ++nslaves;
 938                         break;
 939                 case HAMMER2_PFSTYPE_SOFT_MASTER:
 940                         /*
 941                          * Directly mounted soft master always wins.  There
 942                          * should be only one.
 943                          */
 944                         cluster->array[i].flags |= HAMMER2_CITEM_FEMOD;
 945                         cluster->array[i].flags &= ~HAMMER2_CITEM_INVALID;
 946                         break;
 947                 case HAMMER2_PFSTYPE_SOFT_SLAVE:
 948                         /*
 949                          * Directly mounted soft slave always wins.  There
 950                          * should be only one.
 951                          *
 952                          * XXX
 953                          */
 954                         cluster->array[i].flags &= ~HAMMER2_CITEM_INVALID;
 955                         break;
 956                 case HAMMER2_PFSTYPE_SUPROOT:
 957                         /*
 958                          * spmp (degenerate case)
 959                          */
 960                         cluster->array[i].flags |= HAMMER2_CITEM_FEMOD;
 961                         cluster->array[i].flags &= ~HAMMER2_CITEM_INVALID;
 962                         break;
 963                 default:
 964                         break;
 965                 }
 966         }
 967
 968         /*
 969          * Focus now set, adjust ddflag.  Skip this pass if the focus
 970          * is bad or if we are at the PFS root (the bref won't match at
 971          * the PFS root, obviously).
 972          */
 973         focus = cluster->focus;
 974         if (focus) {
 975                 cluster->ddflag =
 976                         (cluster->focus->bref.type == HAMMER2_BREF_TYPE_INODE);
 977         } else {
 978                 cluster->ddflag = 0;
 979                 goto skip4;
 980         }
 981         if (cluster->focus->flags & HAMMER2_CHAIN_PFSBOUNDARY)
 982                 goto skip4;
 983
 984         /*
 985          * Pass 4
 986          *
 987          * Validate the elements that were not marked invalid.  They should
 988          * match.
 989          */
 990         for (i = 0; i < cluster->nchains; ++i) {
 991                 int ddflag;
 992
 993                 chain = cluster->array[i].chain;
 994
 995                 if (chain == NULL)
 996                         continue;
 997                 if (chain == focus)
 998                         continue;
 999                 if (cluster->array[i].flags & HAMMER2_CITEM_INVALID)
1000                         continue;
1001
1002                 ddflag = (chain->bref.type == HAMMER2_BREF_TYPE_INODE);
1003                 if (chain->bref.type != focus->bref.type ||
1004                     chain->bref.key != focus->bref.key ||
1005                     chain->bref.keybits != focus->bref.keybits ||
1006                     chain->bref.modify_tid != focus->bref.modify_tid ||
1007                     chain->bytes != focus->bytes ||
1008                     ddflag != cluster->ddflag) {
1009                         cluster->array[i].flags |= HAMMER2_CITEM_INVALID;
1010                         if (hammer2_debug & 1)
1011                         kprintf("cluster_resolve: matching modify_tid failed "
1012                                 "bref test: idx=%d type=%02x/%02x "
1013                                 "key=%016jx/%d-%016jx/%d "
1014                                 "mod=%016jx/%016jx bytes=%u/%u\n",
1015                                 i,
1016                                 chain->bref.type, focus->bref.type,
1017                                 chain->bref.key, chain->bref.keybits,
1018                                 focus->bref.key, focus->bref.keybits,
1019                                 chain->bref.modify_tid, focus->bref.modify_tid,
1020                                 chain->bytes, focus->bytes);
1021                         if (hammer2_debug & 0x4000)
1022                                 panic("cluster_resolve");
1023                         /* flag issue and force resync? */
1024                 }
1025         }
1026 skip4:
1027
1028         if (ttlslaves == 0)
1029                 nflags |= HAMMER2_CLUSTER_NOSOFT;
1030         if (ttlmasters == 0)
1031                 nflags |= HAMMER2_CLUSTER_NOHARD;
1032
1033         /*
1034          * Set SSYNCED or MSYNCED for slaves and masters respectively if
1035          * all available nodes (even if 0 are available) are fully
1036          * synchronized.  This is used by the synchronization thread to
1037          * determine if there is work it could potentially accomplish.
1038          */
1039         if (nslaves == ttlslaves)
1040                 nflags |= HAMMER2_CLUSTER_SSYNCED;
1041         if (nmasters == ttlmasters)
1042                 nflags |= HAMMER2_CLUSTER_MSYNCED;
1043
1044         /*
1045          * Determine if the cluster was successfully locked for the
1046          * requested operation and generate an error code.  The cluster
1047          * will not be locked (or ref'd) if an error is returned.
1048          */
1049         atomic_set_int(&cluster->flags, nflags);
1050         atomic_clear_int(&cluster->flags, HAMMER2_CLUSTER_ZFLAGS & ~nflags);
1051
1052         return 0;
1053 }
1054
1055 /*
1056  * This is used by the sync thread to force non-NULL elements of a copy
1057  * of the pmp->iroot cluster to be good which is required to prime the
1058  * sync.
1059  */
1060 void
1061 hammer2_cluster_forcegood(hammer2_cluster_t *cluster)
1062 {
1063         int i;
1064
1065         for (i = 0; i < cluster->nchains; ++i) {
1066                 if (cluster->array[i].chain)
1067                         cluster->array[i].flags &= ~HAMMER2_CITEM_INVALID;
1068         }
1069 }
1070
1071 /*
1072  * Copy a cluster, returned a ref'd cluster.  All underlying chains
1073  * are also ref'd, but not locked.  Focus state is also copied.
1074  *
1075  * Original cluster does not have to be locked but usually is.
1076  * New cluster will not be flagged as locked.
1077  *
1078  * Callers using this function to initialize a new cluster from an inode
1079  * generally lock and resolve the resulting cluster.
1080  *
1081  * Callers which use this function to save/restore a cluster structure
1082  * generally retain the focus state and do not re-resolve it.  Caller should
1083  * not try to re-resolve internal (cparent) node state during an iteration
1084  * as the individual tracking elements of cparent in an iteration may not
1085  * match even though they are correct.
1086  */
1087 hammer2_cluster_t *
1088 hammer2_cluster_copy(hammer2_cluster_t *ocluster)
1089 {
1090         hammer2_pfs_t *pmp = ocluster->pmp;
1091         hammer2_cluster_t *ncluster;
1092         hammer2_chain_t *chain;
1093         int i;
1094
1095         ncluster = kmalloc(sizeof(*ncluster), M_HAMMER2, M_WAITOK | M_ZERO);
1096         ncluster->pmp = pmp;
1097         ncluster->nchains = ocluster->nchains;
1098         ncluster->refs = 1;
1099
1100         for (i = 0; i < ocluster->nchains; ++i) {
1101                 chain = ocluster->array[i].chain;
1102                 ncluster->array[i].chain = chain;
1103                 ncluster->array[i].flags = ocluster->array[i].flags;
1104                 if (chain)
1105                         hammer2_chain_ref(chain);
1106         }
1107         ncluster->focus_index = ocluster->focus_index;
1108         ncluster->focus = ocluster->focus;
1109         ncluster->flags = ocluster->flags & ~(HAMMER2_CLUSTER_LOCKED |
1110                                               HAMMER2_CLUSTER_INODE);
1111
1112         return (ncluster);
1113 }
1114
1115 /*
1116  * Unlock a cluster.  Refcount and focus is maintained.
1117  */
1118 void
1119 hammer2_cluster_unlock(hammer2_cluster_t *cluster)
1120 {
1121         hammer2_chain_t *chain;
1122         int i;
1123
1124         if ((cluster->flags & HAMMER2_CLUSTER_LOCKED) == 0) {
1125                 kprintf("hammer2_cluster_unlock: cluster %p not locked\n",
1126                         cluster);
1127         }
1128         KKASSERT(cluster->flags & HAMMER2_CLUSTER_LOCKED);
1129         KKASSERT(cluster->refs > 0);
1130         atomic_clear_int(&cluster->flags, HAMMER2_CLUSTER_LOCKED);
1131
1132         for (i = 0; i < cluster->nchains; ++i) {
1133                 chain = cluster->array[i].chain;
1134                 if (chain)
1135                         hammer2_chain_unlock(chain);
1136         }
1137 }
1138
1139 /************************************************************************
1140  *                              CLUSTER I/O                             *
1141  ************************************************************************
1142  *
1143  *
1144  * WARNING! blockref[] array data is not universal.  These functions should
1145  *          only be used to access universal data.
1146  *
1147  * NOTE!    The rdata call will wait for at least one of the chain I/Os to
1148  *          complete if necessary.  The I/O's should have already been
1149  *          initiated by the cluster_lock/chain_lock operation.
1150  *
1151  *          The cluster must already be in a modified state before wdata
1152  *          is called.  The data will already be available for this case.
1153  */
1154 const hammer2_media_data_t *
1155 hammer2_cluster_rdata(hammer2_cluster_t *cluster)
1156 {
1157         KKASSERT(cluster->focus != NULL);
1158         return(cluster->focus->data);
1159 }
1160
1161 hammer2_media_data_t *
1162 hammer2_cluster_wdata(hammer2_cluster_t *cluster)
1163 {
1164         KKASSERT(cluster->focus != NULL);
1165         KKASSERT(hammer2_cluster_modified(cluster));
1166         return(cluster->focus->data);
1167 }