sys/vfs/hammer2/hammer2_vfsops.c

   1 /*
   2  * Copyright (c) 2011-2015 The DragonFly Project.  All rights reserved.
   3  *
   4  * This code is derived from software contributed to The DragonFly Project
   5  * by Matthew Dillon <dillon@backplane.com>
   6  * by Daniel Flores (GSOC 2013 - mentored by Matthew Dillon, compression)
   7  *
   8  * Redistribution and use in source and binary forms, with or without
   9  * modification, are permitted provided that the following conditions
  10  * are met:
  11  *
  12  * 1. Redistributions of source code must retain the above copyright
  13  *    notice, this list of conditions and the following disclaimer.
  14  * 2. Redistributions in binary form must reproduce the above copyright
  15  *    notice, this list of conditions and the following disclaimer in
  16  *    the documentation and/or other materials provided with the
  17  *    distribution.
  18  * 3. Neither the name of The DragonFly Project nor the names of its
  19  *    contributors may be used to endorse or promote products derived
  20  *    from this software without specific, prior written permission.
  21  *
  22  * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
  23  * ``AS IS'' AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
  24  * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
  25  * FOR A PARTICULAR PURPOSE ARE DISCLAIMED.  IN NO EVENT SHALL THE
  26  * COPYRIGHT HOLDERS OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
  27  * INCIDENTAL, SPECIAL, EXEMPLARY OR CONSEQUENTIAL DAMAGES (INCLUDING,
  28  * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
  29  * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
  30  * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
  31  * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
  32  * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
  33  * SUCH DAMAGE.
  34  */
  35 #include <sys/param.h>
  36 #include <sys/systm.h>
  37 #include <sys/kernel.h>
  38 #include <sys/nlookup.h>
  39 #include <sys/vnode.h>
  40 #include <sys/mount.h>
  41 #include <sys/fcntl.h>
  42 #include <sys/buf.h>
  43 #include <sys/uuid.h>
  44 #include <sys/vfsops.h>
  45 #include <sys/sysctl.h>
  46 #include <sys/socket.h>
  47 #include <sys/objcache.h>
  48
  49 #include <sys/proc.h>
  50 #include <sys/namei.h>
  51 #include <sys/mountctl.h>
  52 #include <sys/dirent.h>
  53 #include <sys/uio.h>
  54
  55 #include <sys/mutex.h>
  56 #include <sys/mutex2.h>
  57
  58 #include "hammer2.h"
  59 #include "hammer2_disk.h"
  60 #include "hammer2_mount.h"
  61 #include "hammer2_lz4.h"
  62
  63 #include "zlib/hammer2_zlib.h"
  64
  65 #define REPORT_REFS_ERRORS 1    /* XXX remove me */
  66
  67 MALLOC_DEFINE(M_OBJCACHE, "objcache", "Object Cache");
  68
  69 struct hammer2_sync_info {
  70         int error;
  71         int waitfor;
  72 };
  73
  74 TAILQ_HEAD(hammer2_mntlist, hammer2_dev);
  75 static struct hammer2_mntlist hammer2_mntlist;
  76
  77 struct hammer2_pfslist hammer2_pfslist;
  78 struct lock hammer2_mntlk;
  79
  80 int hammer2_debug;
  81 int hammer2_cluster_read = 4;           /* physical read-ahead */
  82 int hammer2_cluster_write = 0;          /* bdwrite() so later inval works */
  83 int hammer2_dedup_enable = 1;
  84 int hammer2_inval_enable = 0;
  85 int hammer2_flush_pipe = 100;
  86 int hammer2_synchronous_flush = 1;
  87 int hammer2_dio_count;
  88 long hammer2_chain_allocs;
  89 long hammer2_chain_frees;
  90 long hammer2_limit_dirty_chains;
  91 long hammer2_count_modified_chains;
  92 long hammer2_iod_invals;
  93 long hammer2_iod_file_read;
  94 long hammer2_iod_meta_read;
  95 long hammer2_iod_indr_read;
  96 long hammer2_iod_fmap_read;
  97 long hammer2_iod_volu_read;
  98 long hammer2_iod_file_write;
  99 long hammer2_iod_file_wembed;
 100 long hammer2_iod_file_wzero;
 101 long hammer2_iod_file_wdedup;
 102 long hammer2_iod_meta_write;
 103 long hammer2_iod_indr_write;
 104 long hammer2_iod_fmap_write;
 105 long hammer2_iod_volu_write;
 106
 107 MALLOC_DECLARE(M_HAMMER2_CBUFFER);
 108 MALLOC_DEFINE(M_HAMMER2_CBUFFER, "HAMMER2-compbuffer",
 109                 "Buffer used for compression.");
 110
 111 MALLOC_DECLARE(M_HAMMER2_DEBUFFER);
 112 MALLOC_DEFINE(M_HAMMER2_DEBUFFER, "HAMMER2-decompbuffer",
 113                 "Buffer used for decompression.");
 114
 115 SYSCTL_NODE(_vfs, OID_AUTO, hammer2, CTLFLAG_RW, 0, "HAMMER2 filesystem");
 116
 117 SYSCTL_INT(_vfs_hammer2, OID_AUTO, debug, CTLFLAG_RW,
 118            &hammer2_debug, 0, "");
 119 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_read, CTLFLAG_RW,
 120            &hammer2_cluster_read, 0, "");
 121 SYSCTL_INT(_vfs_hammer2, OID_AUTO, cluster_write, CTLFLAG_RW,
 122            &hammer2_cluster_write, 0, "");
 123 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dedup_enable, CTLFLAG_RW,
 124            &hammer2_dedup_enable, 0, "");
 125 SYSCTL_INT(_vfs_hammer2, OID_AUTO, inval_enable, CTLFLAG_RW,
 126            &hammer2_inval_enable, 0, "");
 127 SYSCTL_INT(_vfs_hammer2, OID_AUTO, flush_pipe, CTLFLAG_RW,
 128            &hammer2_flush_pipe, 0, "");
 129 SYSCTL_INT(_vfs_hammer2, OID_AUTO, synchronous_flush, CTLFLAG_RW,
 130            &hammer2_synchronous_flush, 0, "");
 131 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_allocs, CTLFLAG_RW,
 132            &hammer2_chain_allocs, 0, "");
 133 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, chain_frees, CTLFLAG_RW,
 134            &hammer2_chain_frees, 0, "");
 135 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, limit_dirty_chains, CTLFLAG_RW,
 136            &hammer2_limit_dirty_chains, 0, "");
 137 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, count_modified_chains, CTLFLAG_RW,
 138            &hammer2_count_modified_chains, 0, "");
 139 SYSCTL_INT(_vfs_hammer2, OID_AUTO, dio_count, CTLFLAG_RD,
 140            &hammer2_dio_count, 0, "");
 141
 142 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_invals, CTLFLAG_RW,
 143            &hammer2_iod_invals, 0, "");
 144 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_read, CTLFLAG_RW,
 145            &hammer2_iod_file_read, 0, "");
 146 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_read, CTLFLAG_RW,
 147            &hammer2_iod_meta_read, 0, "");
 148 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_read, CTLFLAG_RW,
 149            &hammer2_iod_indr_read, 0, "");
 150 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_read, CTLFLAG_RW,
 151            &hammer2_iod_fmap_read, 0, "");
 152 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_read, CTLFLAG_RW,
 153            &hammer2_iod_volu_read, 0, "");
 154
 155 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_write, CTLFLAG_RW,
 156            &hammer2_iod_file_write, 0, "");
 157 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wembed, CTLFLAG_RW,
 158            &hammer2_iod_file_wembed, 0, "");
 159 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wzero, CTLFLAG_RW,
 160            &hammer2_iod_file_wzero, 0, "");
 161 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_file_wdedup, CTLFLAG_RW,
 162            &hammer2_iod_file_wdedup, 0, "");
 163 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_meta_write, CTLFLAG_RW,
 164            &hammer2_iod_meta_write, 0, "");
 165 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_indr_write, CTLFLAG_RW,
 166            &hammer2_iod_indr_write, 0, "");
 167 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_fmap_write, CTLFLAG_RW,
 168            &hammer2_iod_fmap_write, 0, "");
 169 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, iod_volu_write, CTLFLAG_RW,
 170            &hammer2_iod_volu_write, 0, "");
 171
 172 long hammer2_check_icrc32;
 173 long hammer2_check_xxhash64;
 174 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_icrc32, CTLFLAG_RW,
 175            &hammer2_check_icrc32, 0, "");
 176 SYSCTL_LONG(_vfs_hammer2, OID_AUTO, check_xxhash64, CTLFLAG_RW,
 177            &hammer2_check_xxhash64, 0, "");
 178
 179 static int hammer2_vfs_init(struct vfsconf *conf);
 180 static int hammer2_vfs_uninit(struct vfsconf *vfsp);
 181 static int hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 182                                 struct ucred *cred);
 183 static int hammer2_remount(hammer2_dev_t *, struct mount *, char *,
 184                                 struct vnode *, struct ucred *);
 185 static int hammer2_recovery(hammer2_dev_t *hmp);
 186 static int hammer2_vfs_unmount(struct mount *mp, int mntflags);
 187 static int hammer2_vfs_root(struct mount *mp, struct vnode **vpp);
 188 static int hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp,
 189                                 struct ucred *cred);
 190 static int hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp,
 191                                 struct ucred *cred);
 192 static int hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
 193                                 struct fid *fhp, struct vnode **vpp);
 194 static int hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp);
 195 static int hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
 196                                 int *exflagsp, struct ucred **credanonp);
 197
 198 static int hammer2_install_volume_header(hammer2_dev_t *hmp);
 199 static int hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data);
 200
 201 static void hammer2_update_pmps(hammer2_dev_t *hmp);
 202
 203 static void hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp);
 204 static void hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp,
 205                                 hammer2_dev_t *hmp);
 206
 207 /*
 208  * HAMMER2 vfs operations.
 209  */
 210 static struct vfsops hammer2_vfsops = {
 211         .vfs_init       = hammer2_vfs_init,
 212         .vfs_uninit     = hammer2_vfs_uninit,
 213         .vfs_sync       = hammer2_vfs_sync,
 214         .vfs_mount      = hammer2_vfs_mount,
 215         .vfs_unmount    = hammer2_vfs_unmount,
 216         .vfs_root       = hammer2_vfs_root,
 217         .vfs_statfs     = hammer2_vfs_statfs,
 218         .vfs_statvfs    = hammer2_vfs_statvfs,
 219         .vfs_vget       = hammer2_vfs_vget,
 220         .vfs_vptofh     = hammer2_vfs_vptofh,
 221         .vfs_fhtovp     = hammer2_vfs_fhtovp,
 222         .vfs_checkexp   = hammer2_vfs_checkexp
 223 };
 224
 225 MALLOC_DEFINE(M_HAMMER2, "HAMMER2-mount", "");
 226
 227 VFS_SET(hammer2_vfsops, hammer2, VFCF_MPSAFE);
 228 MODULE_VERSION(hammer2, 1);
 229
 230 static
 231 int
 232 hammer2_vfs_init(struct vfsconf *conf)
 233 {
 234         static struct objcache_malloc_args margs_read;
 235         static struct objcache_malloc_args margs_write;
 236         static struct objcache_malloc_args margs_vop;
 237
 238         int error;
 239
 240         error = 0;
 241
 242         if (HAMMER2_BLOCKREF_BYTES != sizeof(struct hammer2_blockref))
 243                 error = EINVAL;
 244         if (HAMMER2_INODE_BYTES != sizeof(struct hammer2_inode_data))
 245                 error = EINVAL;
 246         if (HAMMER2_VOLUME_BYTES != sizeof(struct hammer2_volume_data))
 247                 error = EINVAL;
 248
 249         if (error)
 250                 kprintf("HAMMER2 structure size mismatch; cannot continue.\n");
 251
 252         margs_read.objsize = 65536;
 253         margs_read.mtype = M_HAMMER2_DEBUFFER;
 254
 255         margs_write.objsize = 32768;
 256         margs_write.mtype = M_HAMMER2_CBUFFER;
 257
 258         margs_vop.objsize = sizeof(hammer2_xop_t);
 259         margs_vop.mtype = M_HAMMER2;
 260
 261         /*
 262          * Note thaht for the XOPS cache we want backing store allocations
 263          * to use M_ZERO.  This is not allowed in objcache_get() (to avoid
 264          * confusion), so use the backing store function that does it.  This
 265          * means that initial XOPS objects are zerod but REUSED objects are
 266          * not.  So we are responsible for cleaning the object up sufficiently
 267          * for our needs before objcache_put()ing it back (typically just the
 268          * FIFO indices).
 269          */
 270         cache_buffer_read = objcache_create(margs_read.mtype->ks_shortdesc,
 271                                 0, 1, NULL, NULL, NULL,
 272                                 objcache_malloc_alloc,
 273                                 objcache_malloc_free,
 274                                 &margs_read);
 275         cache_buffer_write = objcache_create(margs_write.mtype->ks_shortdesc,
 276                                 0, 1, NULL, NULL, NULL,
 277                                 objcache_malloc_alloc,
 278                                 objcache_malloc_free,
 279                                 &margs_write);
 280         cache_xops = objcache_create(margs_vop.mtype->ks_shortdesc,
 281                                 0, 1, NULL, NULL, NULL,
 282                                 objcache_malloc_alloc_zero,
 283                                 objcache_malloc_free,
 284                                 &margs_vop);
 285
 286
 287         lockinit(&hammer2_mntlk, "mntlk", 0, 0);
 288         TAILQ_INIT(&hammer2_mntlist);
 289         TAILQ_INIT(&hammer2_pfslist);
 290
 291         hammer2_limit_dirty_chains = maxvnodes / 10;
 292         if (hammer2_limit_dirty_chains > HAMMER2_LIMIT_DIRTY_CHAINS)
 293                 hammer2_limit_dirty_chains = HAMMER2_LIMIT_DIRTY_CHAINS;
 294
 295         return (error);
 296 }
 297
 298 static
 299 int
 300 hammer2_vfs_uninit(struct vfsconf *vfsp __unused)
 301 {
 302         objcache_destroy(cache_buffer_read);
 303         objcache_destroy(cache_buffer_write);
 304         objcache_destroy(cache_xops);
 305         return 0;
 306 }
 307
 308 /*
 309  * Core PFS allocator.  Used to allocate or reference the pmp structure
 310  * for PFS cluster mounts and the spmp structure for media (hmp) structures.
 311  * The pmp can be passed in or loaded by this function using the chain and
 312  * inode data.
 313  *
 314  * pmp->modify_tid tracks new modify_tid transaction ids for front-end
 315  * transactions.  Note that synchronization does not use this field.
 316  * (typically frontend operations and synchronization cannot run on the
 317  * same PFS node at the same time).
 318  *
 319  * XXX check locking
 320  */
 321 hammer2_pfs_t *
 322 hammer2_pfsalloc(hammer2_chain_t *chain,
 323                  const hammer2_inode_data_t *ripdata,
 324                  hammer2_tid_t modify_tid, hammer2_dev_t *force_local)
 325 {
 326         hammer2_pfs_t *pmp;
 327         hammer2_inode_t *iroot;
 328         int count;
 329         int i;
 330         int j;
 331
 332         pmp = NULL;
 333
 334         /*
 335          * Locate or create the PFS based on the cluster id.  If ripdata
 336          * is NULL this is a spmp which is unique and is always allocated.
 337          *
 338          * If the device is mounted in local mode all PFSs are considered
 339          * independent and not part of any cluster (for debugging only).
 340          */
 341         if (ripdata) {
 342                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 343                         if (force_local != pmp->force_local)
 344                                 continue;
 345                         if (force_local == NULL &&
 346                             bcmp(&pmp->pfs_clid, &ripdata->meta.pfs_clid,
 347                                  sizeof(pmp->pfs_clid)) == 0) {
 348                                         break;
 349                         } else if (force_local && pmp->pfs_names[0] &&
 350                             strcmp(pmp->pfs_names[0], ripdata->filename) == 0) {
 351                                         break;
 352                         }
 353                 }
 354         }
 355
 356         if (pmp == NULL) {
 357                 pmp = kmalloc(sizeof(*pmp), M_HAMMER2, M_WAITOK | M_ZERO);
 358                 pmp->force_local = force_local;
 359                 hammer2_trans_manage_init(pmp);
 360                 kmalloc_create(&pmp->minode, "HAMMER2-inodes");
 361                 kmalloc_create(&pmp->mmsg, "HAMMER2-pfsmsg");
 362                 lockinit(&pmp->lock, "pfslk", 0, 0);
 363                 lockinit(&pmp->lock_nlink, "h2nlink", 0, 0);
 364                 spin_init(&pmp->inum_spin, "hm2pfsalloc_inum");
 365                 spin_init(&pmp->xop_spin, "h2xop");
 366                 spin_init(&pmp->lru_spin, "h2lru");
 367                 RB_INIT(&pmp->inum_tree);
 368                 TAILQ_INIT(&pmp->sideq);
 369                 TAILQ_INIT(&pmp->lru_list);
 370                 spin_init(&pmp->list_spin, "hm2pfsalloc_list");
 371
 372                 /*
 373                  * Distribute backend operations to threads
 374                  */
 375                 for (i = 0; i < HAMMER2_XOPGROUPS; ++i)
 376                         hammer2_xop_group_init(pmp, &pmp->xop_groups[i]);
 377
 378                 /*
 379                  * Save the last media transaction id for the flusher.  Set
 380                  * initial
 381                  */
 382                 if (ripdata)
 383                         pmp->pfs_clid = ripdata->meta.pfs_clid;
 384                 TAILQ_INSERT_TAIL(&hammer2_pfslist, pmp, mntentry);
 385
 386                 /*
 387                  * The synchronization thread may start too early, make
 388                  * sure it stays frozen until we are ready to let it go.
 389                  * XXX
 390                  */
 391                 /*
 392                 pmp->primary_thr.flags = HAMMER2_THREAD_FROZEN |
 393                                          HAMMER2_THREAD_REMASTER;
 394                 */
 395         }
 396
 397         /*
 398          * Create the PFS's root inode and any missing XOP helper threads.
 399          */
 400         if ((iroot = pmp->iroot) == NULL) {
 401                 iroot = hammer2_inode_get(pmp, NULL, NULL, -1);
 402                 if (ripdata)
 403                         iroot->meta = ripdata->meta;
 404                 pmp->iroot = iroot;
 405                 hammer2_inode_ref(iroot);
 406                 hammer2_inode_unlock(iroot);
 407         }
 408
 409         /*
 410          * Stop here if no chain is passed in.
 411          */
 412         if (chain == NULL)
 413                 goto done;
 414
 415         /*
 416          * When a chain is passed in we must add it to the PFS's root
 417          * inode, update pmp->pfs_types[], and update the syncronization
 418          * threads.
 419          *
 420          * When forcing local mode, mark the PFS as a MASTER regardless.
 421          *
 422          * At the moment empty spots can develop due to removals or failures.
 423          * Ultimately we want to re-fill these spots but doing so might
 424          * confused running code. XXX
 425          */
 426         hammer2_inode_ref(iroot);
 427         hammer2_mtx_ex(&iroot->lock);
 428         j = iroot->cluster.nchains;
 429
 430         kprintf("add PFS to pmp %p[%d]\n", pmp, j);
 431
 432         if (j == HAMMER2_MAXCLUSTER) {
 433                 kprintf("hammer2_mount: cluster full!\n");
 434                 /* XXX fatal error? */
 435         } else {
 436                 KKASSERT(chain->pmp == NULL);
 437                 chain->pmp = pmp;
 438                 hammer2_chain_ref(chain);
 439                 iroot->cluster.array[j].chain = chain;
 440                 if (force_local)
 441                         pmp->pfs_types[j] = HAMMER2_PFSTYPE_MASTER;
 442                 else
 443                         pmp->pfs_types[j] = ripdata->meta.pfs_type;
 444                 pmp->pfs_names[j] = kstrdup(ripdata->filename, M_HAMMER2);
 445                 pmp->pfs_hmps[j] = chain->hmp;
 446
 447                 /*
 448                  * If the PFS is already mounted we must account
 449                  * for the mount_count here.
 450                  */
 451                 if (pmp->mp)
 452                         ++chain->hmp->mount_count;
 453
 454                 /*
 455                  * May have to fixup dirty chain tracking.  Previous
 456                  * pmp was NULL so nothing to undo.
 457                  */
 458                 if (chain->flags & HAMMER2_CHAIN_MODIFIED)
 459                         hammer2_pfs_memory_inc(pmp);
 460                 ++j;
 461         }
 462         iroot->cluster.nchains = j;
 463
 464         /*
 465          * Update nmasters from any PFS inode which is part of the cluster.
 466          * It is possible that this will result in a value which is too
 467          * high.  MASTER PFSs are authoritative for pfs_nmasters and will
 468          * override this value later on.
 469          *
 470          * (This informs us of masters that might not currently be
 471          *  discoverable by this mount).
 472          */
 473         if (ripdata && pmp->pfs_nmasters < ripdata->meta.pfs_nmasters) {
 474                 pmp->pfs_nmasters = ripdata->meta.pfs_nmasters;
 475         }
 476
 477         /*
 478          * Count visible masters.  Masters are usually added with
 479          * ripdata->meta.pfs_nmasters set to 1.  This detects when there
 480          * are more (XXX and must update the master inodes).
 481          */
 482         count = 0;
 483         for (i = 0; i < iroot->cluster.nchains; ++i) {
 484                 if (pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER)
 485                         ++count;
 486         }
 487         if (pmp->pfs_nmasters < count)
 488                 pmp->pfs_nmasters = count;
 489
 490         /*
 491          * Create missing synchronization and support threads.
 492          *
 493          * Single-node masters (including snapshots) have nothing to
 494          * synchronize and do not require this thread.
 495          *
 496          * Multi-node masters or any number of soft masters, slaves, copy,
 497          * or other PFS types need the thread.
 498          *
 499          * Each thread is responsible for its particular cluster index.
 500          * We use independent threads so stalls or mismatches related to
 501          * any given target do not affect other targets.
 502          */
 503         for (i = 0; i < iroot->cluster.nchains; ++i) {
 504                 /*
 505                  * Single-node masters (including snapshots) have nothing
 506                  * to synchronize and will make direct xops support calls,
 507                  * thus they do not require this thread.
 508                  *
 509                  * Note that there can be thousands of snapshots.  We do not
 510                  * want to create thousands of threads.
 511                  */
 512                 if (pmp->pfs_nmasters <= 1 &&
 513                     pmp->pfs_types[i] == HAMMER2_PFSTYPE_MASTER) {
 514                         continue;
 515                 }
 516
 517                 /*
 518                  * Sync support thread
 519                  */
 520                 if (pmp->sync_thrs[i].td == NULL) {
 521                         hammer2_thr_create(&pmp->sync_thrs[i], pmp,
 522                                            "h2nod", i, -1,
 523                                            hammer2_primary_sync_thread);
 524                 }
 525         }
 526
 527         /*
 528          * Create missing Xop threads
 529          *
 530          * NOTE: We create helper threads for all mounted PFSs or any
 531          *       PFSs with 2+ nodes (so the sync thread can update them,
 532          *       even if not mounted).
 533          */
 534         if (pmp->mp || iroot->cluster.nchains >= 2)
 535                 hammer2_xop_helper_create(pmp);
 536
 537         hammer2_mtx_unlock(&iroot->lock);
 538         hammer2_inode_drop(iroot);
 539 done:
 540         return pmp;
 541 }
 542
 543 /*
 544  * Deallocate an element of a probed PFS.  If destroying and this is a
 545  * MASTER, adjust nmasters.
 546  *
 547  * This function does not physically destroy the PFS element in its device
 548  * under the super-root  (see hammer2_ioctl_pfs_delete()).
 549  */
 550 void
 551 hammer2_pfsdealloc(hammer2_pfs_t *pmp, int clindex, int destroying)
 552 {
 553         hammer2_inode_t *iroot;
 554         hammer2_chain_t *chain;
 555         int j;
 556
 557         /*
 558          * Cleanup our reference on iroot.  iroot is (should) not be needed
 559          * by the flush code.
 560          */
 561         iroot = pmp->iroot;
 562         if (iroot) {
 563                 /*
 564                  * Stop synchronizing
 565                  *
 566                  * XXX flush after acquiring the iroot lock.
 567                  * XXX clean out the cluster index from all inode structures.
 568                  */
 569                 hammer2_thr_delete(&pmp->sync_thrs[clindex]);
 570
 571                 /*
 572                  * Remove the cluster index from the group.  If destroying
 573                  * the PFS and this is a master, adjust pfs_nmasters.
 574                  */
 575                 hammer2_mtx_ex(&iroot->lock);
 576                 chain = iroot->cluster.array[clindex].chain;
 577                 iroot->cluster.array[clindex].chain = NULL;
 578
 579                 switch(pmp->pfs_types[clindex]) {
 580                 case HAMMER2_PFSTYPE_MASTER:
 581                         if (destroying && pmp->pfs_nmasters > 0)
 582                                 --pmp->pfs_nmasters;
 583                         /* XXX adjust ripdata->meta.pfs_nmasters */
 584                         break;
 585                 default:
 586                         break;
 587                 }
 588                 pmp->pfs_types[clindex] = HAMMER2_PFSTYPE_NONE;
 589
 590                 hammer2_mtx_unlock(&iroot->lock);
 591
 592                 /*
 593                  * Release the chain.
 594                  */
 595                 if (chain) {
 596                         atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 597                         hammer2_chain_drop(chain);
 598                 }
 599
 600                 /*
 601                  * Terminate all XOP threads for the cluster index.
 602                  */
 603                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 604                         hammer2_thr_delete(&pmp->xop_groups[j].thrs[clindex]);
 605         }
 606 }
 607
 608 /*
 609  * Destroy a PFS, typically only occurs after the last mount on a device
 610  * has gone away.
 611  */
 612 static void
 613 hammer2_pfsfree(hammer2_pfs_t *pmp)
 614 {
 615         hammer2_inode_t *iroot;
 616         hammer2_chain_t *chain;
 617         int i;
 618         int j;
 619
 620         /*
 621          * Cleanup our reference on iroot.  iroot is (should) not be needed
 622          * by the flush code.
 623          */
 624         TAILQ_REMOVE(&hammer2_pfslist, pmp, mntentry);
 625
 626         iroot = pmp->iroot;
 627         if (iroot) {
 628                 for (i = 0; i < iroot->cluster.nchains; ++i) {
 629                         hammer2_thr_delete(&pmp->sync_thrs[i]);
 630                         for (j = 0; j < HAMMER2_XOPGROUPS; ++j)
 631                                 hammer2_thr_delete(&pmp->xop_groups[j].thrs[i]);
 632                 }
 633 #if REPORT_REFS_ERRORS
 634                 if (pmp->iroot->refs != 1)
 635                         kprintf("PMP->IROOT %p REFS WRONG %d\n",
 636                                 pmp->iroot, pmp->iroot->refs);
 637 #else
 638                 KKASSERT(pmp->iroot->refs == 1);
 639 #endif
 640                 /* ref for pmp->iroot */
 641                 hammer2_inode_drop(pmp->iroot);
 642                 pmp->iroot = NULL;
 643         }
 644
 645         /*
 646          * Cleanup chains remaining on LRU list.
 647          */
 648         kprintf("pfsfree: %p lrucount=%d\n", pmp, pmp->lru_count);
 649         while ((chain = TAILQ_FIRST(&pmp->lru_list)) != NULL) {
 650                 hammer2_chain_ref(chain);
 651                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
 652                 hammer2_chain_drop(chain);
 653         }
 654
 655         /*
 656          * Free remaining pmp resources
 657          */
 658         kmalloc_destroy(&pmp->mmsg);
 659         kmalloc_destroy(&pmp->minode);
 660
 661         kfree(pmp, M_HAMMER2);
 662 }
 663
 664 /*
 665  * Remove all references to hmp from the pfs list.  Any PFS which becomes
 666  * empty is terminated and freed.
 667  *
 668  * XXX inefficient.
 669  */
 670 static void
 671 hammer2_pfsfree_scan(hammer2_dev_t *hmp)
 672 {
 673         hammer2_pfs_t *pmp;
 674         hammer2_inode_t *iroot;
 675         hammer2_chain_t *rchain;
 676         int didfreeze;
 677         int i;
 678         int j;
 679
 680 again:
 681         TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 682                 if ((iroot = pmp->iroot) == NULL)
 683                         continue;
 684                 if (hmp->spmp == pmp) {
 685                         kprintf("unmount hmp %p remove spmp %p\n",
 686                                 hmp, pmp);
 687                         hmp->spmp = NULL;
 688                 }
 689
 690                 /*
 691                  * Determine if this PFS is affected.  If it is we must
 692                  * freeze all management threads and lock its iroot.
 693                  *
 694                  * Freezing a management thread forces it idle, operations
 695                  * in-progress will be aborted and it will have to start
 696                  * over again when unfrozen, or exit if told to exit.
 697                  */
 698                 for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 699                         if (pmp->pfs_hmps[i] == hmp)
 700                                 break;
 701                 }
 702                 if (i != HAMMER2_MAXCLUSTER) {
 703                         /*
 704                          * Make sure all synchronization threads are locked
 705                          * down.
 706                          */
 707                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 708                                 if (pmp->pfs_hmps[i] == NULL)
 709                                         continue;
 710                                 hammer2_thr_freeze_async(&pmp->sync_thrs[i]);
 711                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 712                                         hammer2_thr_freeze_async(
 713                                                 &pmp->xop_groups[j].thrs[i]);
 714                                 }
 715                         }
 716                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 717                                 if (pmp->pfs_hmps[i] == NULL)
 718                                         continue;
 719                                 hammer2_thr_freeze(&pmp->sync_thrs[i]);
 720                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 721                                         hammer2_thr_freeze(
 722                                                 &pmp->xop_groups[j].thrs[i]);
 723                                 }
 724                         }
 725
 726                         /*
 727                          * Lock the inode and clean out matching chains.
 728                          * Note that we cannot use hammer2_inode_lock_*()
 729                          * here because that would attempt to validate the
 730                          * cluster that we are in the middle of ripping
 731                          * apart.
 732                          *
 733                          * WARNING! We are working directly on the inodes
 734                          *          embedded cluster.
 735                          */
 736                         hammer2_mtx_ex(&iroot->lock);
 737
 738                         /*
 739                          * Remove the chain from matching elements of the PFS.
 740                          */
 741                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 742                                 if (pmp->pfs_hmps[i] != hmp)
 743                                         continue;
 744                                 hammer2_thr_delete(&pmp->sync_thrs[i]);
 745                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 746                                         hammer2_thr_delete(
 747                                                 &pmp->xop_groups[j].thrs[i]);
 748                                 }
 749                                 rchain = iroot->cluster.array[i].chain;
 750                                 iroot->cluster.array[i].chain = NULL;
 751                                 pmp->pfs_types[i] = 0;
 752                                 if (pmp->pfs_names[i]) {
 753                                         kfree(pmp->pfs_names[i], M_HAMMER2);
 754                                         pmp->pfs_names[i] = NULL;
 755                                 }
 756                                 if (rchain) {
 757                                         hammer2_chain_drop(rchain);
 758                                         /* focus hint */
 759                                         if (iroot->cluster.focus == rchain)
 760                                                 iroot->cluster.focus = NULL;
 761                                 }
 762                                 pmp->pfs_hmps[i] = NULL;
 763                         }
 764                         hammer2_mtx_unlock(&iroot->lock);
 765                         didfreeze = 1;  /* remaster, unfreeze down below */
 766                 } else {
 767                         didfreeze = 0;
 768                 }
 769
 770                 /*
 771                  * Cleanup trailing chains.  Gaps may remain.
 772                  */
 773                 for (i = HAMMER2_MAXCLUSTER - 1; i >= 0; --i) {
 774                         if (pmp->pfs_hmps[i])
 775                                 break;
 776                 }
 777                 iroot->cluster.nchains = i + 1;
 778
 779                 /*
 780                  * If the PMP has no elements remaining we can destroy it.
 781                  * (this will transition management threads from frozen->exit).
 782                  */
 783                 if (iroot->cluster.nchains == 0) {
 784                         kprintf("unmount hmp %p last ref to PMP=%p\n",
 785                                 hmp, pmp);
 786                         hammer2_pfsfree(pmp);
 787                         goto again;
 788                 }
 789
 790                 /*
 791                  * If elements still remain we need to set the REMASTER
 792                  * flag and unfreeze it.
 793                  */
 794                 if (didfreeze) {
 795                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 796                                 if (pmp->pfs_hmps[i] == NULL)
 797                                         continue;
 798                                 hammer2_thr_remaster(&pmp->sync_thrs[i]);
 799                                 hammer2_thr_unfreeze(&pmp->sync_thrs[i]);
 800                                 for (j = 0; j < HAMMER2_XOPGROUPS; ++j) {
 801                                         hammer2_thr_remaster(
 802                                                 &pmp->xop_groups[j].thrs[i]);
 803                                         hammer2_thr_unfreeze(
 804                                                 &pmp->xop_groups[j].thrs[i]);
 805                                 }
 806                         }
 807                 }
 808         }
 809 }
 810
 811 /*
 812  * Mount or remount HAMMER2 fileystem from physical media
 813  *
 814  *      mountroot
 815  *              mp              mount point structure
 816  *              path            NULL
 817  *              data            <unused>
 818  *              cred            <unused>
 819  *
 820  *      mount
 821  *              mp              mount point structure
 822  *              path            path to mount point
 823  *              data            pointer to argument structure in user space
 824  *                      volume  volume path (device@LABEL form)
 825  *                      hflags  user mount flags
 826  *              cred            user credentials
 827  *
 828  * RETURNS:     0       Success
 829  *              !0      error number
 830  */
 831 static
 832 int
 833 hammer2_vfs_mount(struct mount *mp, char *path, caddr_t data,
 834                   struct ucred *cred)
 835 {
 836         struct hammer2_mount_info info;
 837         hammer2_pfs_t *pmp;
 838         hammer2_pfs_t *spmp;
 839         hammer2_dev_t *hmp;
 840         hammer2_dev_t *force_local;
 841         hammer2_key_t key_next;
 842         hammer2_key_t key_dummy;
 843         hammer2_key_t lhc;
 844         struct vnode *devvp;
 845         struct nlookupdata nd;
 846         hammer2_chain_t *parent;
 847         hammer2_chain_t *chain;
 848         hammer2_cluster_t *cluster;
 849         const hammer2_inode_data_t *ripdata;
 850         hammer2_blockref_t bref;
 851         struct file *fp;
 852         char devstr[MNAMELEN];
 853         size_t size;
 854         size_t done;
 855         char *dev;
 856         char *label;
 857         int ronly = 1;
 858         int error;
 859         int cache_index;
 860         int i;
 861
 862         hmp = NULL;
 863         pmp = NULL;
 864         dev = NULL;
 865         label = NULL;
 866         devvp = NULL;
 867         cache_index = -1;
 868
 869         kprintf("hammer2_mount\n");
 870
 871         if (path == NULL) {
 872                 /*
 873                  * Root mount
 874                  */
 875                 bzero(&info, sizeof(info));
 876                 info.cluster_fd = -1;
 877                 ksnprintf(devstr, sizeof(devstr), "%s",
 878                           mp->mnt_stat.f_mntfromname);
 879                 kprintf("hammer2_mount: root '%s'\n", devstr);
 880         } else {
 881                 /*
 882                  * Non-root mount or updating a mount
 883                  */
 884                 error = copyin(data, &info, sizeof(info));
 885                 if (error)
 886                         return (error);
 887
 888                 error = copyinstr(info.volume, devstr, MNAMELEN - 1, &done);
 889                 if (error)
 890                         return (error);
 891         }
 892
 893         /*
 894          * Extract device and label, automatically mount @LOCAL if no
 895          * label specified.
 896          */
 897         dev = devstr;
 898         label = strchr(devstr, '@');
 899         if (label && ((label + 1) - dev) > done)
 900                 return (EINVAL);
 901         if (label == NULL || label[1] == 0) {
 902                 label = "LOCAL";        /* not modified after this point */
 903         } else {
 904                 *label = '\0';
 905                 label++;
 906         }
 907
 908         kprintf("hammer2_mount: dev=\"%s\" label=\"%s\"\n",
 909                 dev, label);
 910
 911         if (mp->mnt_flag & MNT_UPDATE) {
 912                 /*
 913                  * Update mount.  Note that pmp->iroot->cluster is
 914                  * an inode-embedded cluster and thus cannot be
 915                  * directly locked.
 916                  *
 917                  * XXX HAMMER2 needs to implement NFS export via
 918                  *     mountctl.
 919                  */
 920                 pmp = MPTOPMP(mp);
 921                 pmp->hflags = info.hflags;
 922                 cluster = &pmp->iroot->cluster;
 923                 for (i = 0; i < cluster->nchains; ++i) {
 924                         if (cluster->array[i].chain == NULL)
 925                                 continue;
 926                         hmp = cluster->array[i].chain->hmp;
 927                         devvp = hmp->devvp;
 928                         error = hammer2_remount(hmp, mp, path,
 929                                                 devvp, cred);
 930                         if (error)
 931                                 break;
 932                 }
 933
 934                 return error;
 935         }
 936
 937         /*
 938          * HMP device mount
 939          *
 940          * If a path is specified and dev is not an empty string, lookup the
 941          * name and verify that it referes to a block device.
 942          *
 943          * If a path is specified and dev is an empty string we fall through
 944          * and locate the label in the hmp search.
 945          */
 946         if (path && *dev != 0) {
 947                 error = nlookup_init(&nd, dev, UIO_SYSSPACE, NLC_FOLLOW);
 948                 if (error == 0)
 949                         error = nlookup(&nd);
 950                 if (error == 0)
 951                         error = cache_vref(&nd.nl_nch, nd.nl_cred, &devvp);
 952                 nlookup_done(&nd);
 953         } else if (path == NULL) {
 954                 /* root mount */
 955                 cdev_t cdev = kgetdiskbyname(dev);
 956                 error = bdevvp(cdev, &devvp);
 957                 if (error)
 958                         kprintf("hammer2: cannot find '%s'\n", dev);
 959         } else {
 960                 /*
 961                  * We will locate the hmp using the label in the hmp loop.
 962                  */
 963                 error = 0;
 964         }
 965
 966         if (error == 0 && devvp) {
 967                 if (vn_isdisk(devvp, &error))
 968                         error = vfs_mountedon(devvp);
 969         }
 970
 971         /*
 972          * Determine if the device has already been mounted.  After this
 973          * check hmp will be non-NULL if we are doing the second or more
 974          * hammer2 mounts from the same device.
 975          */
 976         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
 977         if (devvp) {
 978                 /*
 979                  * Match the device
 980                  */
 981                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
 982                         if (hmp->devvp == devvp)
 983                                 break;
 984                 }
 985         } else if (error == 0) {
 986                 /*
 987                  * Match the label to a pmp already probed.
 988                  */
 989                 TAILQ_FOREACH(pmp, &hammer2_pfslist, mntentry) {
 990                         for (i = 0; i < HAMMER2_MAXCLUSTER; ++i) {
 991                                 if (pmp->pfs_names[i] &&
 992                                     strcmp(pmp->pfs_names[i], label) == 0) {
 993                                         hmp = pmp->pfs_hmps[i];
 994                                         break;
 995                                 }
 996                         }
 997                         if (hmp)
 998                                 break;
 999                 }
1000                 if (hmp == NULL)
1001                         error = ENOENT;
1002         }
1003
1004         /*
1005          * Open the device if this isn't a secondary mount and construct
1006          * the H2 device mount (hmp).
1007          */
1008         if (hmp == NULL) {
1009                 hammer2_chain_t *schain;
1010                 hammer2_xid_t xid;
1011
1012                 if (error == 0 && vcount(devvp) > 0)
1013                         error = EBUSY;
1014
1015                 /*
1016                  * Now open the device
1017                  */
1018                 if (error == 0) {
1019                         ronly = ((mp->mnt_flag & MNT_RDONLY) != 0);
1020                         vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1021                         error = vinvalbuf(devvp, V_SAVE, 0, 0);
1022                         if (error == 0) {
1023                                 error = VOP_OPEN(devvp,
1024                                                  ronly ? FREAD : FREAD | FWRITE,
1025                                                  FSCRED, NULL);
1026                         }
1027                         vn_unlock(devvp);
1028                 }
1029                 if (error && devvp) {
1030                         vrele(devvp);
1031                         devvp = NULL;
1032                 }
1033                 if (error) {
1034                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1035                         return error;
1036                 }
1037                 hmp = kmalloc(sizeof(*hmp), M_HAMMER2, M_WAITOK | M_ZERO);
1038                 ksnprintf(hmp->devrepname, sizeof(hmp->devrepname), "%s", dev);
1039                 hmp->ronly = ronly;
1040                 hmp->devvp = devvp;
1041                 hmp->hflags = info.hflags & HMNT2_DEVFLAGS;
1042                 kmalloc_create(&hmp->mchain, "HAMMER2-chains");
1043                 TAILQ_INSERT_TAIL(&hammer2_mntlist, hmp, mntentry);
1044                 RB_INIT(&hmp->iotree);
1045                 spin_init(&hmp->io_spin, "hm2mount_io");
1046                 spin_init(&hmp->list_spin, "hm2mount_list");
1047                 TAILQ_INIT(&hmp->flushq);
1048
1049                 lockinit(&hmp->vollk, "h2vol", 0, 0);
1050                 lockinit(&hmp->bulklk, "h2bulk", 0, 0);
1051
1052                 /*
1053                  * vchain setup. vchain.data is embedded.
1054                  * vchain.refs is initialized and will never drop to 0.
1055                  *
1056                  * NOTE! voldata is not yet loaded.
1057                  */
1058                 hmp->vchain.hmp = hmp;
1059                 hmp->vchain.refs = 1;
1060                 hmp->vchain.data = (void *)&hmp->voldata;
1061                 hmp->vchain.bref.type = HAMMER2_BREF_TYPE_VOLUME;
1062                 hmp->vchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1063                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1064
1065                 hammer2_chain_core_init(&hmp->vchain);
1066                 /* hmp->vchain.u.xxx is left NULL */
1067
1068                 /*
1069                  * fchain setup.  fchain.data is embedded.
1070                  * fchain.refs is initialized and will never drop to 0.
1071                  *
1072                  * The data is not used but needs to be initialized to
1073                  * pass assertion muster.  We use this chain primarily
1074                  * as a placeholder for the freemap's top-level RBTREE
1075                  * so it does not interfere with the volume's topology
1076                  * RBTREE.
1077                  */
1078                 hmp->fchain.hmp = hmp;
1079                 hmp->fchain.refs = 1;
1080                 hmp->fchain.data = (void *)&hmp->voldata.freemap_blockset;
1081                 hmp->fchain.bref.type = HAMMER2_BREF_TYPE_FREEMAP;
1082                 hmp->fchain.bref.data_off = 0 | HAMMER2_PBUFRADIX;
1083                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1084                 hmp->fchain.bref.methods =
1085                         HAMMER2_ENC_CHECK(HAMMER2_CHECK_FREEMAP) |
1086                         HAMMER2_ENC_COMP(HAMMER2_COMP_NONE);
1087
1088                 hammer2_chain_core_init(&hmp->fchain);
1089                 /* hmp->fchain.u.xxx is left NULL */
1090
1091                 /*
1092                  * Install the volume header and initialize fields from
1093                  * voldata.
1094                  */
1095                 error = hammer2_install_volume_header(hmp);
1096                 if (error) {
1097                         hammer2_unmount_helper(mp, NULL, hmp);
1098                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1099                         hammer2_vfs_unmount(mp, MNT_FORCE);
1100                         return error;
1101                 }
1102
1103                 /*
1104                  * Really important to get these right or flush will get
1105                  * confused.
1106                  */
1107                 hmp->spmp = hammer2_pfsalloc(NULL, NULL, 0, NULL);
1108                 kprintf("alloc spmp %p tid %016jx\n",
1109                         hmp->spmp, hmp->voldata.mirror_tid);
1110                 spmp = hmp->spmp;
1111
1112                 /*
1113                  * Dummy-up vchain and fchain's modify_tid.  mirror_tid
1114                  * is inherited from the volume header.
1115                  */
1116                 xid = 0;
1117                 hmp->vchain.bref.mirror_tid = hmp->voldata.mirror_tid;
1118                 hmp->vchain.bref.modify_tid = hmp->vchain.bref.mirror_tid;
1119                 hmp->vchain.pmp = spmp;
1120                 hmp->fchain.bref.mirror_tid = hmp->voldata.freemap_tid;
1121                 hmp->fchain.bref.modify_tid = hmp->fchain.bref.mirror_tid;
1122                 hmp->fchain.pmp = spmp;
1123
1124                 /*
1125                  * First locate the super-root inode, which is key 0
1126                  * relative to the volume header's blockset.
1127                  *
1128                  * Then locate the root inode by scanning the directory keyspace
1129                  * represented by the label.
1130                  */
1131                 parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1132                 schain = hammer2_chain_lookup(&parent, &key_dummy,
1133                                       HAMMER2_SROOT_KEY, HAMMER2_SROOT_KEY,
1134                                       &cache_index, 0);
1135                 hammer2_chain_lookup_done(parent);
1136                 if (schain == NULL) {
1137                         kprintf("hammer2_mount: invalid super-root\n");
1138                         hammer2_unmount_helper(mp, NULL, hmp);
1139                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1140                         hammer2_vfs_unmount(mp, MNT_FORCE);
1141                         return EINVAL;
1142                 }
1143                 if (schain->error) {
1144                         kprintf("hammer2_mount: error %s reading super-root\n",
1145                                 hammer2_error_str(schain->error));
1146                         hammer2_chain_unlock(schain);
1147                         hammer2_chain_drop(schain);
1148                         schain = NULL;
1149                         hammer2_unmount_helper(mp, NULL, hmp);
1150                         lockmgr(&hammer2_mntlk, LK_RELEASE);
1151                         hammer2_vfs_unmount(mp, MNT_FORCE);
1152                         return EINVAL;
1153                 }
1154
1155                 /*
1156                  * The super-root always uses an inode_tid of 1 when
1157                  * creating PFSs.
1158                  */
1159                 spmp->inode_tid = 1;
1160                 spmp->modify_tid = schain->bref.modify_tid + 1;
1161
1162                 /*
1163                  * Sanity-check schain's pmp and finish initialization.
1164                  * Any chain belonging to the super-root topology should
1165                  * have a NULL pmp (not even set to spmp).
1166                  */
1167                 ripdata = &hammer2_chain_rdata(schain)->ipdata;
1168                 KKASSERT(schain->pmp == NULL);
1169                 spmp->pfs_clid = ripdata->meta.pfs_clid;
1170
1171                 /*
1172                  * Replace the dummy spmp->iroot with a real one.  It's
1173                  * easier to just do a wholesale replacement than to try
1174                  * to update the chain and fixup the iroot fields.
1175                  *
1176                  * The returned inode is locked with the supplied cluster.
1177                  */
1178                 cluster = hammer2_cluster_from_chain(schain);
1179                 hammer2_inode_drop(spmp->iroot);
1180                 spmp->iroot = NULL;
1181                 spmp->iroot = hammer2_inode_get(spmp, NULL, cluster, -1);
1182                 spmp->spmp_hmp = hmp;
1183                 spmp->pfs_types[0] = ripdata->meta.pfs_type;
1184                 spmp->pfs_hmps[0] = hmp;
1185                 hammer2_inode_ref(spmp->iroot);
1186                 hammer2_inode_unlock(spmp->iroot);
1187                 hammer2_cluster_unlock(cluster);
1188                 hammer2_cluster_drop(cluster);
1189                 schain = NULL;
1190                 /* leave spmp->iroot with one ref */
1191
1192                 if ((mp->mnt_flag & MNT_RDONLY) == 0) {
1193                         error = hammer2_recovery(hmp);
1194                         /* XXX do something with error */
1195                 }
1196                 hammer2_update_pmps(hmp);
1197                 hammer2_iocom_init(hmp);
1198
1199                 /*
1200                  * Ref the cluster management messaging descriptor.  The mount
1201                  * program deals with the other end of the communications pipe.
1202                  *
1203                  * Root mounts typically do not supply one.
1204                  */
1205                 if (info.cluster_fd >= 0) {
1206                         fp = holdfp(curproc->p_fd, info.cluster_fd, -1);
1207                         if (fp) {
1208                                 hammer2_cluster_reconnect(hmp, fp);
1209                         } else {
1210                                 kprintf("hammer2_mount: bad cluster_fd!\n");
1211                         }
1212                 }
1213         } else {
1214                 spmp = hmp->spmp;
1215                 if (info.hflags & HMNT2_DEVFLAGS) {
1216                         kprintf("hammer2: Warning: mount flags pertaining "
1217                                 "to the whole device may only be specified "
1218                                 "on the first mount of the device: %08x\n",
1219                                 info.hflags & HMNT2_DEVFLAGS);
1220                 }
1221         }
1222
1223         /*
1224          * Force local mount (disassociate all PFSs from their clusters).
1225          * Used primarily for debugging.
1226          */
1227         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1228
1229         /*
1230          * Lookup the mount point under the media-localized super-root.
1231          * Scanning hammer2_pfslist doesn't help us because it represents
1232          * PFS cluster ids which can aggregate several named PFSs together.
1233          *
1234          * cluster->pmp will incorrectly point to spmp and must be fixed
1235          * up later on.
1236          */
1237         hammer2_inode_lock(spmp->iroot, 0);
1238         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1239         lhc = hammer2_dirhash(label, strlen(label));
1240         chain = hammer2_chain_lookup(&parent, &key_next,
1241                                      lhc, lhc + HAMMER2_DIRHASH_LOMASK,
1242                                      &cache_index, 0);
1243         while (chain) {
1244                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
1245                     strcmp(label, chain->data->ipdata.filename) == 0) {
1246                         break;
1247                 }
1248                 chain = hammer2_chain_next(&parent, chain, &key_next,
1249                                             key_next,
1250                                             lhc + HAMMER2_DIRHASH_LOMASK,
1251                                             &cache_index, 0);
1252         }
1253         if (parent) {
1254                 hammer2_chain_unlock(parent);
1255                 hammer2_chain_drop(parent);
1256         }
1257         hammer2_inode_unlock(spmp->iroot);
1258
1259         /*
1260          * PFS could not be found?
1261          */
1262         if (chain == NULL) {
1263                 kprintf("hammer2_mount: PFS label not found\n");
1264                 hammer2_unmount_helper(mp, NULL, hmp);
1265                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1266                 hammer2_vfs_unmount(mp, MNT_FORCE);
1267
1268                 return EINVAL;
1269         }
1270
1271         /*
1272          * Acquire the pmp structure (it should have already been allocated
1273          * via hammer2_update_pmps() so do not pass cluster in to add to
1274          * available chains).
1275          *
1276          * Check if the cluster has already been mounted.  A cluster can
1277          * only be mounted once, use null mounts to mount additional copies.
1278          */
1279         ripdata = &chain->data->ipdata;
1280         bref = chain->bref;
1281         pmp = hammer2_pfsalloc(NULL, ripdata,
1282                                bref.modify_tid, force_local);
1283         hammer2_chain_unlock(chain);
1284         hammer2_chain_drop(chain);
1285
1286         /*
1287          * Finish the mount
1288          */
1289         kprintf("hammer2_mount hmp=%p pmp=%p\n", hmp, pmp);
1290
1291         if (pmp->mp) {
1292                 kprintf("hammer2_mount: PFS already mounted!\n");
1293                 hammer2_unmount_helper(mp, NULL, hmp);
1294                 lockmgr(&hammer2_mntlk, LK_RELEASE);
1295                 hammer2_vfs_unmount(mp, MNT_FORCE);
1296
1297                 return EBUSY;
1298         }
1299
1300         pmp->hflags = info.hflags;
1301         mp->mnt_flag = MNT_LOCAL;
1302         mp->mnt_kern_flag |= MNTK_ALL_MPSAFE;   /* all entry pts are SMP */
1303         mp->mnt_kern_flag |= MNTK_THR_SYNC;     /* new vsyncscan semantics */
1304
1305         /*
1306          * required mount structure initializations
1307          */
1308         mp->mnt_stat.f_iosize = HAMMER2_PBUFSIZE;
1309         mp->mnt_stat.f_bsize = HAMMER2_PBUFSIZE;
1310
1311         mp->mnt_vstat.f_frsize = HAMMER2_PBUFSIZE;
1312         mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1313
1314         /*
1315          * Optional fields
1316          */
1317         mp->mnt_iosize_max = MAXPHYS;
1318
1319         /*
1320          * Connect up mount pointers.
1321          */
1322         hammer2_mount_helper(mp, pmp);
1323
1324         lockmgr(&hammer2_mntlk, LK_RELEASE);
1325
1326         /*
1327          * Finish setup
1328          */
1329         vfs_getnewfsid(mp);
1330         vfs_add_vnodeops(mp, &hammer2_vnode_vops, &mp->mnt_vn_norm_ops);
1331         vfs_add_vnodeops(mp, &hammer2_spec_vops, &mp->mnt_vn_spec_ops);
1332         vfs_add_vnodeops(mp, &hammer2_fifo_vops, &mp->mnt_vn_fifo_ops);
1333
1334         if (path) {
1335                 copyinstr(info.volume, mp->mnt_stat.f_mntfromname,
1336                           MNAMELEN - 1, &size);
1337                 bzero(mp->mnt_stat.f_mntfromname + size, MNAMELEN - size);
1338         } /* else root mount, already in there */
1339
1340         bzero(mp->mnt_stat.f_mntonname, sizeof(mp->mnt_stat.f_mntonname));
1341         if (path) {
1342                 copyinstr(path, mp->mnt_stat.f_mntonname,
1343                           sizeof(mp->mnt_stat.f_mntonname) - 1,
1344                           &size);
1345         } else {
1346                 /* root mount */
1347                 mp->mnt_stat.f_mntonname[0] = '/';
1348         }
1349
1350         /*
1351          * Initial statfs to prime mnt_stat.
1352          */
1353         hammer2_vfs_statfs(mp, &mp->mnt_stat, cred);
1354
1355         return 0;
1356 }
1357
1358 /*
1359  * Scan PFSs under the super-root and create hammer2_pfs structures.
1360  */
1361 static
1362 void
1363 hammer2_update_pmps(hammer2_dev_t *hmp)
1364 {
1365         const hammer2_inode_data_t *ripdata;
1366         hammer2_chain_t *parent;
1367         hammer2_chain_t *chain;
1368         hammer2_blockref_t bref;
1369         hammer2_dev_t *force_local;
1370         hammer2_pfs_t *spmp;
1371         hammer2_pfs_t *pmp;
1372         hammer2_key_t key_next;
1373         int cache_index = -1;
1374
1375         /*
1376          * Force local mount (disassociate all PFSs from their clusters).
1377          * Used primarily for debugging.
1378          */
1379         force_local = (hmp->hflags & HMNT2_LOCAL) ? hmp : NULL;
1380
1381         /*
1382          * Lookup mount point under the media-localized super-root.
1383          *
1384          * cluster->pmp will incorrectly point to spmp and must be fixed
1385          * up later on.
1386          */
1387         spmp = hmp->spmp;
1388         hammer2_inode_lock(spmp->iroot, 0);
1389         parent = hammer2_inode_chain(spmp->iroot, 0, HAMMER2_RESOLVE_ALWAYS);
1390         chain = hammer2_chain_lookup(&parent, &key_next,
1391                                          HAMMER2_KEY_MIN, HAMMER2_KEY_MAX,
1392                                          &cache_index, 0);
1393         while (chain) {
1394                 if (chain->bref.type != HAMMER2_BREF_TYPE_INODE)
1395                         continue;
1396                 ripdata = &chain->data->ipdata;
1397                 bref = chain->bref;
1398                 kprintf("ADD LOCAL PFS: %s\n", ripdata->filename);
1399
1400                 pmp = hammer2_pfsalloc(chain, ripdata,
1401                                        bref.modify_tid, force_local);
1402                 chain = hammer2_chain_next(&parent, chain, &key_next,
1403                                            key_next, HAMMER2_KEY_MAX,
1404                                            &cache_index, 0);
1405         }
1406         if (parent) {
1407                 hammer2_chain_unlock(parent);
1408                 hammer2_chain_drop(parent);
1409         }
1410         hammer2_inode_unlock(spmp->iroot);
1411 }
1412
1413 static
1414 int
1415 hammer2_remount(hammer2_dev_t *hmp, struct mount *mp, char *path __unused,
1416                 struct vnode *devvp, struct ucred *cred)
1417 {
1418         int error;
1419
1420         if (hmp->ronly && (mp->mnt_kern_flag & MNTK_WANTRDWR)) {
1421                 error = hammer2_recovery(hmp);
1422         } else {
1423                 error = 0;
1424         }
1425         return error;
1426 }
1427
1428 static
1429 int
1430 hammer2_vfs_unmount(struct mount *mp, int mntflags)
1431 {
1432         hammer2_pfs_t *pmp;
1433         int flags;
1434         int error = 0;
1435
1436         pmp = MPTOPMP(mp);
1437
1438         if (pmp == NULL)
1439                 return(0);
1440
1441         lockmgr(&hammer2_mntlk, LK_EXCLUSIVE);
1442
1443         /*
1444          * If mount initialization proceeded far enough we must flush
1445          * its vnodes and sync the underlying mount points.  Three syncs
1446          * are required to fully flush the filesystem (freemap updates lag
1447          * by one flush, and one extra for safety).
1448          */
1449         if (mntflags & MNT_FORCE)
1450                 flags = FORCECLOSE;
1451         else
1452                 flags = 0;
1453         if (pmp->iroot) {
1454                 error = vflush(mp, 0, flags);
1455                 if (error)
1456                         goto failed;
1457                 hammer2_vfs_sync(mp, MNT_WAIT);
1458                 hammer2_vfs_sync(mp, MNT_WAIT);
1459                 hammer2_vfs_sync(mp, MNT_WAIT);
1460         }
1461
1462         /*
1463          * Cleanup the frontend support XOPS threads
1464          */
1465         hammer2_xop_helper_cleanup(pmp);
1466
1467         if (pmp->mp)
1468                 hammer2_unmount_helper(mp, pmp, NULL);
1469
1470         error = 0;
1471 failed:
1472         lockmgr(&hammer2_mntlk, LK_RELEASE);
1473
1474         return (error);
1475 }
1476
1477 /*
1478  * Mount helper, hook the system mount into our PFS.
1479  * The mount lock is held.
1480  *
1481  * We must bump the mount_count on related devices for any
1482  * mounted PFSs.
1483  */
1484 static
1485 void
1486 hammer2_mount_helper(struct mount *mp, hammer2_pfs_t *pmp)
1487 {
1488         hammer2_cluster_t *cluster;
1489         hammer2_chain_t *rchain;
1490         int i;
1491
1492         mp->mnt_data = (qaddr_t)pmp;
1493         pmp->mp = mp;
1494
1495         /*
1496          * After pmp->mp is set we have to adjust hmp->mount_count.
1497          */
1498         cluster = &pmp->iroot->cluster;
1499         for (i = 0; i < cluster->nchains; ++i) {
1500                 rchain = cluster->array[i].chain;
1501                 if (rchain == NULL)
1502                         continue;
1503                 ++rchain->hmp->mount_count;
1504                 kprintf("hammer2_mount hmp=%p ++mount_count=%d\n",
1505                         rchain->hmp, rchain->hmp->mount_count);
1506         }
1507
1508         /*
1509          * Create missing Xop threads
1510          */
1511         hammer2_xop_helper_create(pmp);
1512 }
1513
1514 /*
1515  * Mount helper, unhook the system mount from our PFS.
1516  * The mount lock is held.
1517  *
1518  * If hmp is supplied a mount responsible for being the first to open
1519  * the block device failed and the block device and all PFSs using the
1520  * block device must be cleaned up.
1521  *
1522  * If pmp is supplied multiple devices might be backing the PFS and each
1523  * must be disconnected.  This might not be the last PFS using some of the
1524  * underlying devices.  Also, we have to adjust our hmp->mount_count
1525  * accounting for the devices backing the pmp which is now undergoing an
1526  * unmount.
1527  */
1528 static
1529 void
1530 hammer2_unmount_helper(struct mount *mp, hammer2_pfs_t *pmp, hammer2_dev_t *hmp)
1531 {
1532         hammer2_cluster_t *cluster;
1533         hammer2_chain_t *rchain;
1534         struct vnode *devvp;
1535         int dumpcnt;
1536         int ronly = 0;
1537         int i;
1538
1539         /*
1540          * If no device supplied this is a high-level unmount and we have to
1541          * to disconnect the mount, adjust mount_count, and locate devices
1542          * that might now have no mounts.
1543          */
1544         if (pmp) {
1545                 KKASSERT(hmp == NULL);
1546                 KKASSERT((void *)(intptr_t)mp->mnt_data == pmp);
1547                 pmp->mp = NULL;
1548                 mp->mnt_data = NULL;
1549
1550                 /*
1551                  * After pmp->mp is cleared we have to account for
1552                  * mount_count.
1553                  */
1554                 cluster = &pmp->iroot->cluster;
1555                 for (i = 0; i < cluster->nchains; ++i) {
1556                         rchain = cluster->array[i].chain;
1557                         if (rchain == NULL)
1558                                 continue;
1559                         --rchain->hmp->mount_count;
1560                         kprintf("hammer2_unmount hmp=%p --mount_count=%d\n",
1561                                 rchain->hmp, rchain->hmp->mount_count);
1562                         /* scrapping hmp now may invalidate the pmp */
1563                 }
1564 again:
1565                 TAILQ_FOREACH(hmp, &hammer2_mntlist, mntentry) {
1566                         if (hmp->mount_count == 0) {
1567                                 hammer2_unmount_helper(NULL, NULL, hmp);
1568                                 goto again;
1569                         }
1570                 }
1571                 return;
1572         }
1573
1574         /*
1575          * Try to terminate the block device.  We can't terminate it if
1576          * there are still PFSs referencing it.
1577          */
1578         kprintf("hammer2_unmount hmp=%p mount_count=%d\n",
1579                 hmp, hmp->mount_count);
1580         if (hmp->mount_count)
1581                 return;
1582
1583         hammer2_pfsfree_scan(hmp);
1584         hammer2_dev_exlock(hmp);        /* XXX order */
1585
1586         /*
1587          * Cycle the volume data lock as a safety (probably not needed any
1588          * more).  To ensure everything is out we need to flush at least
1589          * three times.  (1) The running of the sideq can dirty the
1590          * filesystem, (2) A normal flush can dirty the freemap, and
1591          * (3) ensure that the freemap is fully synchronized.
1592          *
1593          * The next mount's recovery scan can clean everything up but we want
1594          * to leave the filesystem in a 100% clean state on a normal unmount.
1595          */
1596 #if 0
1597         hammer2_voldata_lock(hmp);
1598         hammer2_voldata_unlock(hmp);
1599 #endif
1600         hammer2_iocom_uninit(hmp);
1601
1602         if ((hmp->vchain.flags | hmp->fchain.flags) &
1603             HAMMER2_CHAIN_FLUSH_MASK) {
1604                 kprintf("hammer2_unmount: chains left over "
1605                         "after final sync\n");
1606                 kprintf("    vchain %08x\n", hmp->vchain.flags);
1607                 kprintf("    fchain %08x\n", hmp->fchain.flags);
1608
1609                 if (hammer2_debug & 0x0010)
1610                         Debugger("entered debugger");
1611         }
1612
1613         KKASSERT(hmp->spmp == NULL);
1614
1615         /*
1616          * Finish up with the device vnode
1617          */
1618         if ((devvp = hmp->devvp) != NULL) {
1619                 vn_lock(devvp, LK_EXCLUSIVE | LK_RETRY);
1620                 vinvalbuf(devvp, (ronly ? 0 : V_SAVE), 0, 0);
1621                 hmp->devvp = NULL;
1622                 VOP_CLOSE(devvp, (ronly ? FREAD : FREAD|FWRITE), NULL);
1623                 vn_unlock(devvp);
1624                 vrele(devvp);
1625                 devvp = NULL;
1626         }
1627
1628         /*
1629          * Clear vchain/fchain flags that might prevent final cleanup
1630          * of these chains.
1631          */
1632         if (hmp->vchain.flags & HAMMER2_CHAIN_MODIFIED) {
1633                 atomic_add_long(&hammer2_count_modified_chains, -1);
1634                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_MODIFIED);
1635                 hammer2_pfs_memory_wakeup(hmp->vchain.pmp);
1636         }
1637         if (hmp->vchain.flags & HAMMER2_CHAIN_UPDATE) {
1638                 atomic_clear_int(&hmp->vchain.flags, HAMMER2_CHAIN_UPDATE);
1639         }
1640
1641         if (hmp->fchain.flags & HAMMER2_CHAIN_MODIFIED) {
1642                 atomic_add_long(&hammer2_count_modified_chains, -1);
1643                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_MODIFIED);
1644                 hammer2_pfs_memory_wakeup(hmp->fchain.pmp);
1645         }
1646         if (hmp->fchain.flags & HAMMER2_CHAIN_UPDATE) {
1647                 atomic_clear_int(&hmp->fchain.flags, HAMMER2_CHAIN_UPDATE);
1648         }
1649
1650         /*
1651          * Final drop of embedded freemap root chain to
1652          * clean up fchain.core (fchain structure is not
1653          * flagged ALLOCATED so it is cleaned out and then
1654          * left to rot).
1655          */
1656         hammer2_chain_drop(&hmp->fchain);
1657
1658         /*
1659          * Final drop of embedded volume root chain to clean
1660          * up vchain.core (vchain structure is not flagged
1661          * ALLOCATED so it is cleaned out and then left to
1662          * rot).
1663          */
1664         dumpcnt = 50;
1665         hammer2_dump_chain(&hmp->vchain, 0, &dumpcnt, 'v');
1666         dumpcnt = 50;
1667         hammer2_dump_chain(&hmp->fchain, 0, &dumpcnt, 'f');
1668         hammer2_dev_unlock(hmp);
1669         hammer2_chain_drop(&hmp->vchain);
1670
1671         hammer2_io_cleanup(hmp, &hmp->iotree);
1672         if (hmp->iofree_count) {
1673                 kprintf("io_cleanup: %d I/O's left hanging\n",
1674                         hmp->iofree_count);
1675         }
1676
1677         TAILQ_REMOVE(&hammer2_mntlist, hmp, mntentry);
1678         kmalloc_destroy(&hmp->mchain);
1679         kfree(hmp, M_HAMMER2);
1680 }
1681
1682 int
1683 hammer2_vfs_vget(struct mount *mp, struct vnode *dvp,
1684                  ino_t ino, struct vnode **vpp)
1685 {
1686         hammer2_xop_lookup_t *xop;
1687         hammer2_pfs_t *pmp;
1688         hammer2_inode_t *ip;
1689         hammer2_tid_t inum;
1690         int error;
1691
1692         inum = (hammer2_tid_t)ino & HAMMER2_DIRHASH_USERMSK;
1693
1694         error = 0;
1695         pmp = MPTOPMP(mp);
1696
1697         /*
1698          * Easy if we already have it cached
1699          */
1700         ip = hammer2_inode_lookup(pmp, inum);
1701         if (ip) {
1702                 hammer2_inode_lock(ip, HAMMER2_RESOLVE_SHARED);
1703                 *vpp = hammer2_igetv(ip, &error);
1704                 hammer2_inode_unlock(ip);
1705                 hammer2_inode_drop(ip);         /* from lookup */
1706
1707                 return error;
1708         }
1709
1710         /*
1711          * Otherwise we have to find the inode
1712          */
1713         xop = hammer2_xop_alloc(pmp->iroot, 0);
1714         xop->lhc = inum;
1715         hammer2_xop_start(&xop->head, hammer2_xop_lookup);
1716         error = hammer2_xop_collect(&xop->head, 0);
1717
1718         if (error == 0) {
1719                 if (hammer2_cluster_rdata(&xop->head.cluster) == NULL) {
1720                         kprintf("vget: no collect error but also no rdata\n");
1721                         kprintf("xop %p\n", xop);
1722                         while ((hammer2_debug & 0x80000) == 0) {
1723                                 tsleep(xop, PCATCH, "wait", hz * 10);
1724                         }
1725                         ip = NULL;
1726                 } else {
1727                         ip = hammer2_inode_get(pmp, NULL, &xop->head.cluster, -1);
1728                 }
1729         }
1730         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1731
1732         if (ip) {
1733                 *vpp = hammer2_igetv(ip, &error);
1734                 hammer2_inode_unlock(ip);
1735         } else {
1736                 *vpp = NULL;
1737                 error = ENOENT;
1738         }
1739         return (error);
1740 }
1741
1742 static
1743 int
1744 hammer2_vfs_root(struct mount *mp, struct vnode **vpp)
1745 {
1746         hammer2_pfs_t *pmp;
1747         struct vnode *vp;
1748         int error;
1749
1750         pmp = MPTOPMP(mp);
1751         if (pmp->iroot == NULL) {
1752                 *vpp = NULL;
1753                 return EINVAL;
1754         }
1755
1756         error = 0;
1757         hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1758
1759         while (pmp->inode_tid == 0) {
1760                 hammer2_xop_ipcluster_t *xop;
1761                 hammer2_inode_meta_t *meta;
1762
1763                 xop = hammer2_xop_alloc(pmp->iroot, HAMMER2_XOP_MODIFYING);
1764                 hammer2_xop_start(&xop->head, hammer2_xop_ipcluster);
1765                 error = hammer2_xop_collect(&xop->head, 0);
1766
1767                 if (error == 0) {
1768                         meta = &xop->head.cluster.focus->data->ipdata.meta;
1769                         pmp->iroot->meta = *meta;
1770                         pmp->inode_tid = meta->pfs_inum + 1;
1771                         if (pmp->inode_tid < HAMMER2_INODE_START)
1772                                 pmp->inode_tid = HAMMER2_INODE_START;
1773                         pmp->modify_tid =
1774                                 xop->head.cluster.focus->bref.modify_tid + 1;
1775                         kprintf("PFS: Starting inode %jd\n",
1776                                 (intmax_t)pmp->inode_tid);
1777                         kprintf("PMP focus good set nextino=%ld mod=%016jx\n",
1778                                 pmp->inode_tid, pmp->modify_tid);
1779                         wakeup(&pmp->iroot);
1780
1781                         hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1782
1783                         /*
1784                          * Prime the mount info.
1785                          */
1786                         hammer2_vfs_statfs(mp, &mp->mnt_stat, NULL);
1787                         break;
1788                 }
1789
1790                 /*
1791                  * Loop, try again
1792                  */
1793                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
1794                 hammer2_inode_unlock(pmp->iroot);
1795                 error = tsleep(&pmp->iroot, PCATCH, "h2root", hz);
1796                 hammer2_inode_lock(pmp->iroot, HAMMER2_RESOLVE_SHARED);
1797                 if (error == EINTR)
1798                         break;
1799         }
1800
1801         if (error) {
1802                 hammer2_inode_unlock(pmp->iroot);
1803                 *vpp = NULL;
1804         } else {
1805                 vp = hammer2_igetv(pmp->iroot, &error);
1806                 hammer2_inode_unlock(pmp->iroot);
1807                 *vpp = vp;
1808         }
1809
1810         return (error);
1811 }
1812
1813 /*
1814  * Filesystem status
1815  *
1816  * XXX incorporate ipdata->meta.inode_quota and data_quota
1817  */
1818 static
1819 int
1820 hammer2_vfs_statfs(struct mount *mp, struct statfs *sbp, struct ucred *cred)
1821 {
1822         hammer2_pfs_t *pmp;
1823         hammer2_dev_t *hmp;
1824         hammer2_blockref_t bref;
1825         int i;
1826
1827         /*
1828          * NOTE: iroot might not have validated the cluster yet.
1829          */
1830         pmp = MPTOPMP(mp);
1831
1832         mp->mnt_stat.f_files = 0;
1833         mp->mnt_stat.f_ffree = 0;
1834         mp->mnt_stat.f_blocks = 0;
1835         mp->mnt_stat.f_bfree = 0;
1836         mp->mnt_stat.f_bavail = 0;
1837
1838         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1839                 hmp = pmp->pfs_hmps[i];
1840                 if (hmp == NULL)
1841                         continue;
1842                 if (pmp->iroot->cluster.array[i].chain)
1843                         bref = pmp->iroot->cluster.array[i].chain->bref;
1844                 else
1845                         bzero(&bref, sizeof(bref));
1846
1847                 mp->mnt_stat.f_files = bref.embed.stats.inode_count;
1848                 mp->mnt_stat.f_ffree = 0;
1849                 mp->mnt_stat.f_blocks = hmp->voldata.allocator_size /
1850                                         mp->mnt_vstat.f_bsize;
1851                 mp->mnt_stat.f_bfree = hmp->voldata.allocator_free /
1852                                         mp->mnt_vstat.f_bsize;
1853                 mp->mnt_stat.f_bavail = mp->mnt_stat.f_bfree;
1854
1855                 *sbp = mp->mnt_stat;
1856         }
1857         return (0);
1858 }
1859
1860 static
1861 int
1862 hammer2_vfs_statvfs(struct mount *mp, struct statvfs *sbp, struct ucred *cred)
1863 {
1864         hammer2_pfs_t *pmp;
1865         hammer2_dev_t *hmp;
1866         hammer2_blockref_t bref;
1867         int i;
1868
1869         /*
1870          * NOTE: iroot might not have validated the cluster yet.
1871          */
1872         pmp = MPTOPMP(mp);
1873
1874         mp->mnt_vstat.f_bsize = 0;
1875         mp->mnt_vstat.f_files = 0;
1876         mp->mnt_vstat.f_ffree = 0;
1877         mp->mnt_vstat.f_blocks = 0;
1878         mp->mnt_vstat.f_bfree = 0;
1879         mp->mnt_vstat.f_bavail = 0;
1880
1881         for (i = 0; i < pmp->iroot->cluster.nchains; ++i) {
1882                 hmp = pmp->pfs_hmps[i];
1883                 if (hmp == NULL)
1884                         continue;
1885                 if (pmp->iroot->cluster.array[i].chain)
1886                         bref = pmp->iroot->cluster.array[i].chain->bref;
1887                 else
1888                         bzero(&bref, sizeof(bref));
1889
1890                 mp->mnt_vstat.f_bsize = HAMMER2_PBUFSIZE;
1891                 mp->mnt_vstat.f_files = bref.embed.stats.inode_count;
1892                 mp->mnt_vstat.f_ffree = 0;
1893                 mp->mnt_vstat.f_blocks = hmp->voldata.allocator_size /
1894                                         mp->mnt_vstat.f_bsize;
1895                 mp->mnt_vstat.f_bfree = hmp->voldata.allocator_free /
1896                                         mp->mnt_vstat.f_bsize;
1897                 mp->mnt_vstat.f_bavail = mp->mnt_vstat.f_bfree;
1898
1899                 *sbp = mp->mnt_vstat;
1900         }
1901         return (0);
1902 }
1903
1904 /*
1905  * Mount-time recovery (RW mounts)
1906  *
1907  * Updates to the free block table are allowed to lag flushes by one
1908  * transaction.  In case of a crash, then on a fresh mount we must do an
1909  * incremental scan of the last committed transaction id and make sure that
1910  * all related blocks have been marked allocated.
1911  *
1912  * The super-root topology and each PFS has its own transaction id domain,
1913  * so we must track PFS boundary transitions.
1914  */
1915 struct hammer2_recovery_elm {
1916         TAILQ_ENTRY(hammer2_recovery_elm) entry;
1917         hammer2_chain_t *chain;
1918         hammer2_tid_t sync_tid;
1919 };
1920
1921 TAILQ_HEAD(hammer2_recovery_list, hammer2_recovery_elm);
1922
1923 struct hammer2_recovery_info {
1924         struct hammer2_recovery_list list;
1925         hammer2_tid_t   mtid;
1926         int     depth;
1927 };
1928
1929 static int hammer2_recovery_scan(hammer2_dev_t *hmp,
1930                         hammer2_chain_t *parent,
1931                         struct hammer2_recovery_info *info,
1932                         hammer2_tid_t sync_tid);
1933
1934 #define HAMMER2_RECOVERY_MAXDEPTH       10
1935
1936 static
1937 int
1938 hammer2_recovery(hammer2_dev_t *hmp)
1939 {
1940         struct hammer2_recovery_info info;
1941         struct hammer2_recovery_elm *elm;
1942         hammer2_chain_t *parent;
1943         hammer2_tid_t sync_tid;
1944         hammer2_tid_t mirror_tid;
1945         int error;
1946         int cumulative_error = 0;
1947
1948         hammer2_trans_init(hmp->spmp, 0);
1949
1950         sync_tid = hmp->voldata.freemap_tid;
1951         mirror_tid = hmp->voldata.mirror_tid;
1952
1953         kprintf("hammer2 mount \"%s\": ", hmp->devrepname);
1954         if (sync_tid >= mirror_tid) {
1955                 kprintf(" no recovery needed\n");
1956         } else {
1957                 kprintf(" freemap recovery %016jx-%016jx\n",
1958                         sync_tid + 1, mirror_tid);
1959         }
1960
1961         TAILQ_INIT(&info.list);
1962         info.depth = 0;
1963         parent = hammer2_chain_lookup_init(&hmp->vchain, 0);
1964         cumulative_error = hammer2_recovery_scan(hmp, parent, &info, sync_tid);
1965         hammer2_chain_lookup_done(parent);
1966
1967         while ((elm = TAILQ_FIRST(&info.list)) != NULL) {
1968                 TAILQ_REMOVE(&info.list, elm, entry);
1969                 parent = elm->chain;
1970                 sync_tid = elm->sync_tid;
1971                 kfree(elm, M_HAMMER2);
1972
1973                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
1974                 error = hammer2_recovery_scan(hmp, parent, &info,
1975                                               hmp->voldata.freemap_tid);
1976                 hammer2_chain_unlock(parent);
1977                 hammer2_chain_drop(parent);     /* drop elm->chain ref */
1978                 if (error)
1979                         cumulative_error = error;
1980         }
1981         hammer2_trans_done(hmp->spmp);
1982
1983         return cumulative_error;
1984 }
1985
1986 static
1987 int
1988 hammer2_recovery_scan(hammer2_dev_t *hmp, hammer2_chain_t *parent,
1989                       struct hammer2_recovery_info *info,
1990                       hammer2_tid_t sync_tid)
1991 {
1992         const hammer2_inode_data_t *ripdata;
1993         hammer2_chain_t *chain;
1994         hammer2_blockref_t bref;
1995         int cache_index;
1996         int cumulative_error = 0;
1997         int error;
1998         int first;
1999
2000         /*
2001          * Adjust freemap to ensure that the block(s) are marked allocated.
2002          */
2003         if (parent->bref.type != HAMMER2_BREF_TYPE_VOLUME) {
2004                 hammer2_freemap_adjust(hmp, &parent->bref,
2005                                        HAMMER2_FREEMAP_DORECOVER);
2006         }
2007
2008         /*
2009          * Check type for recursive scan
2010          */
2011         switch(parent->bref.type) {
2012         case HAMMER2_BREF_TYPE_VOLUME:
2013                 /* data already instantiated */
2014                 break;
2015         case HAMMER2_BREF_TYPE_INODE:
2016                 /*
2017                  * Must instantiate data for DIRECTDATA test and also
2018                  * for recursion.
2019                  */
2020                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2021                 ripdata = &hammer2_chain_rdata(parent)->ipdata;
2022                 if (ripdata->meta.op_flags & HAMMER2_OPFLAG_DIRECTDATA) {
2023                         /* not applicable to recovery scan */
2024                         hammer2_chain_unlock(parent);
2025                         return 0;
2026                 }
2027                 hammer2_chain_unlock(parent);
2028                 break;
2029         case HAMMER2_BREF_TYPE_INDIRECT:
2030                 /*
2031                  * Must instantiate data for recursion
2032                  */
2033                 hammer2_chain_lock(parent, HAMMER2_RESOLVE_ALWAYS);
2034                 hammer2_chain_unlock(parent);
2035                 break;
2036         case HAMMER2_BREF_TYPE_DIRENT:
2037         case HAMMER2_BREF_TYPE_DATA:
2038         case HAMMER2_BREF_TYPE_FREEMAP:
2039         case HAMMER2_BREF_TYPE_FREEMAP_NODE:
2040         case HAMMER2_BREF_TYPE_FREEMAP_LEAF:
2041                 /* not applicable to recovery scan */
2042                 return 0;
2043                 break;
2044         default:
2045                 return EDOM;
2046         }
2047
2048         /*
2049          * Defer operation if depth limit reached or if we are crossing a
2050          * PFS boundary.
2051          */
2052         if (info->depth >= HAMMER2_RECOVERY_MAXDEPTH) {
2053                 struct hammer2_recovery_elm *elm;
2054
2055                 elm = kmalloc(sizeof(*elm), M_HAMMER2, M_ZERO | M_WAITOK);
2056                 elm->chain = parent;
2057                 elm->sync_tid = sync_tid;
2058                 hammer2_chain_ref(parent);
2059                 TAILQ_INSERT_TAIL(&info->list, elm, entry);
2060                 /* unlocked by caller */
2061
2062                 return(0);
2063         }
2064
2065
2066         /*
2067          * Recursive scan of the last flushed transaction only.  We are
2068          * doing this without pmp assignments so don't leave the chains
2069          * hanging around after we are done with them.
2070          */
2071         cache_index = 0;
2072         chain = NULL;
2073         first = 1;
2074
2075         while (hammer2_chain_scan(parent, &chain, &bref,
2076                                   &first, &cache_index,
2077                                   HAMMER2_LOOKUP_NODATA) != NULL) {
2078                 /*
2079                  * If this is a leaf
2080                  */
2081                 if (chain == NULL) {
2082                         if (bref.mirror_tid > sync_tid) {
2083                                 hammer2_freemap_adjust(hmp, &bref,
2084                                                      HAMMER2_FREEMAP_DORECOVER);
2085                         }
2086                         continue;
2087                 }
2088
2089                 /*
2090                  * This may or may not be a recursive node.
2091                  */
2092                 atomic_set_int(&chain->flags, HAMMER2_CHAIN_RELEASE);
2093                 if (bref.mirror_tid > sync_tid) {
2094                         ++info->depth;
2095                         error = hammer2_recovery_scan(hmp, chain,
2096                                                       info, sync_tid);
2097                         --info->depth;
2098                         if (error)
2099                                 cumulative_error = error;
2100                 }
2101
2102                 /*
2103                  * Flush the recovery at the PFS boundary to stage it for
2104                  * the final flush of the super-root topology.
2105                  */
2106                 if ((bref.flags & HAMMER2_BREF_FLAG_PFSROOT) &&
2107                     (chain->flags & HAMMER2_CHAIN_ONFLUSH)) {
2108                         hammer2_flush(chain, HAMMER2_FLUSH_TOP);
2109                 }
2110         }
2111
2112         return cumulative_error;
2113 }
2114
2115 /*
2116  * Sync a mount point; this is called on a per-mount basis from the
2117  * filesystem syncer process periodically and whenever a user issues
2118  * a sync.
2119  */
2120 int
2121 hammer2_vfs_sync(struct mount *mp, int waitfor)
2122 {
2123         hammer2_xop_flush_t *xop;
2124         struct hammer2_sync_info info;
2125         hammer2_inode_t *iroot;
2126         hammer2_pfs_t *pmp;
2127         int flags;
2128         int error;
2129
2130         pmp = MPTOPMP(mp);
2131         iroot = pmp->iroot;
2132         KKASSERT(iroot);
2133         KKASSERT(iroot->pmp == pmp);
2134
2135         /*
2136          * We can't acquire locks on existing vnodes while in a transaction
2137          * without risking a deadlock.  This assumes that vfsync() can be
2138          * called without the vnode locked (which it can in DragonFly).
2139          * Otherwise we'd have to implement a multi-pass or flag the lock
2140          * failures and retry.
2141          *
2142          * The reclamation code interlocks with the sync list's token
2143          * (by removing the vnode from the scan list) before unlocking
2144          * the inode, giving us time to ref the inode.
2145          */
2146         /*flags = VMSC_GETVP;*/
2147         flags = 0;
2148         if (waitfor & MNT_LAZY)
2149                 flags |= VMSC_ONEPASS;
2150
2151         /*
2152          * Preflush the vnodes using a normal transaction before interlocking
2153          * with a flush transaction.  We do this to try to run as much of
2154          * the compression as possible outside the flush transaction.
2155          *
2156          * For efficiency do an async pass before making sure with a
2157          * synchronous pass on all related buffer cache buffers.
2158          */
2159         hammer2_trans_init(pmp, 0);
2160         info.error = 0;
2161         info.waitfor = MNT_NOWAIT;
2162         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2163         info.waitfor = MNT_WAIT;
2164         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2165         hammer2_trans_done(pmp);
2166
2167         /*
2168          * Start our flush transaction.  This does not return until all
2169          * concurrent transactions have completed and will prevent any
2170          * new transactions from running concurrently, except for the
2171          * buffer cache transactions.
2172          *
2173          * NOTE!  It is still possible for the paging code to push pages
2174          *        out via a UIO_NOCOPY hammer2_vop_write() during the main
2175          *        flush.
2176          */
2177         hammer2_trans_init(pmp, HAMMER2_TRANS_ISFLUSH);
2178         hammer2_inode_run_sideq(pmp);
2179
2180         info.error = 0;
2181         info.waitfor = MNT_NOWAIT;
2182         vsyncscan(mp, flags | VMSC_NOWAIT, hammer2_sync_scan2, &info);
2183         info.waitfor = MNT_WAIT;
2184         vsyncscan(mp, flags, hammer2_sync_scan2, &info);
2185         hammer2_bioq_sync(pmp);
2186
2187         /*
2188          * Use the XOP interface to concurrently flush all nodes to
2189          * synchronize the PFSROOT subtopology to the media.  A standard
2190          * end-of-scan ENOENT error indicates cluster sufficiency.
2191          *
2192          * Note that this flush will not be visible on crash recovery until
2193          * we flush the super-root topology in the next loop.
2194          *
2195          * XXX For now wait for all flushes to complete.
2196          */
2197         if (iroot) {
2198                 xop = hammer2_xop_alloc(iroot, HAMMER2_XOP_MODIFYING);
2199                 hammer2_xop_start(&xop->head, hammer2_inode_xop_flush);
2200                 error = hammer2_xop_collect(&xop->head,
2201                                             HAMMER2_XOP_COLLECT_WAITALL);
2202                 hammer2_xop_retire(&xop->head, HAMMER2_XOPMASK_VOP);
2203                 if (error == ENOENT)
2204                         error = 0;
2205         } else {
2206                 error = 0;
2207         }
2208         hammer2_trans_done(pmp);
2209
2210         return (error);
2211 }
2212
2213 /*
2214  * Sync passes.
2215  *
2216  * Note that we ignore the tranasction mtid we got above.  Instead,
2217  * each vfsync below will ultimately get its own via TRANS_BUFCACHE
2218  * transactions.
2219  */
2220 static int
2221 hammer2_sync_scan2(struct mount *mp, struct vnode *vp, void *data)
2222 {
2223         struct hammer2_sync_info *info = data;
2224         hammer2_inode_t *ip;
2225         int error;
2226
2227         /*
2228          * Degenerate cases.  Note that ip == NULL typically means the
2229          * syncer vnode itself and we don't want to vclrisdirty() in that
2230          * situation.
2231          */
2232         ip = VTOI(vp);
2233         if (ip == NULL) {
2234                 return(0);
2235         }
2236         if (vp->v_type == VNON || vp->v_type == VBAD) {
2237                 vclrisdirty(vp);
2238                 return(0);
2239         }
2240
2241         /*
2242          * VOP_FSYNC will start a new transaction so replicate some code
2243          * here to do it inline (see hammer2_vop_fsync()).
2244          *
2245          * WARNING: The vfsync interacts with the buffer cache and might
2246          *          block, we can't hold the inode lock at that time.
2247          *          However, we MUST ref ip before blocking to ensure that
2248          *          it isn't ripped out from under us (since we do not
2249          *          hold a lock on the vnode).
2250          */
2251         hammer2_inode_ref(ip);
2252         if ((ip->flags & HAMMER2_INODE_MODIFIED) ||
2253             !RB_EMPTY(&vp->v_rbdirty_tree)) {
2254                 vfsync(vp, info->waitfor, 1, NULL, NULL);
2255                 if (ip->flags & (HAMMER2_INODE_RESIZED |
2256                                  HAMMER2_INODE_MODIFIED)) {
2257                         hammer2_inode_lock(ip, 0);
2258                         hammer2_inode_chain_sync(ip);
2259                         hammer2_inode_unlock(ip);
2260                 }
2261         }
2262         if ((ip->flags & HAMMER2_INODE_MODIFIED) == 0 &&
2263             RB_EMPTY(&vp->v_rbdirty_tree)) {
2264                 vclrisdirty(vp);
2265         }
2266
2267         hammer2_inode_drop(ip);
2268 #if 1
2269         error = 0;
2270         if (error)
2271                 info->error = error;
2272 #endif
2273         return(0);
2274 }
2275
2276 static
2277 int
2278 hammer2_vfs_vptofh(struct vnode *vp, struct fid *fhp)
2279 {
2280         hammer2_inode_t *ip;
2281
2282         KKASSERT(MAXFIDSZ >= 16);
2283         ip = VTOI(vp);
2284         fhp->fid_len = offsetof(struct fid, fid_data[16]);
2285         fhp->fid_ext = 0;
2286         ((hammer2_tid_t *)fhp->fid_data)[0] = ip->meta.inum;
2287         ((hammer2_tid_t *)fhp->fid_data)[1] = 0;
2288
2289         return 0;
2290 }
2291
2292 static
2293 int
2294 hammer2_vfs_fhtovp(struct mount *mp, struct vnode *rootvp,
2295                struct fid *fhp, struct vnode **vpp)
2296 {
2297         hammer2_pfs_t *pmp;
2298         hammer2_tid_t inum;
2299         int error;
2300
2301         pmp = MPTOPMP(mp);
2302         inum = ((hammer2_tid_t *)fhp->fid_data)[0] & HAMMER2_DIRHASH_USERMSK;
2303         if (vpp) {
2304                 if (inum == 1)
2305                         error = hammer2_vfs_root(mp, vpp);
2306                 else
2307                         error = hammer2_vfs_vget(mp, NULL, inum, vpp);
2308         } else {
2309                 error = 0;
2310         }
2311         if (error)
2312                 kprintf("fhtovp: %016jx -> %p, %d\n", inum, *vpp, error);
2313         return error;
2314 }
2315
2316 static
2317 int
2318 hammer2_vfs_checkexp(struct mount *mp, struct sockaddr *nam,
2319                  int *exflagsp, struct ucred **credanonp)
2320 {
2321         hammer2_pfs_t *pmp;
2322         struct netcred *np;
2323         int error;
2324
2325         pmp = MPTOPMP(mp);
2326         np = vfs_export_lookup(mp, &pmp->export, nam);
2327         if (np) {
2328                 *exflagsp = np->netc_exflags;
2329                 *credanonp = &np->netc_anon;
2330                 error = 0;
2331         } else {
2332                 error = EACCES;
2333         }
2334         return error;
2335 }
2336
2337 /*
2338  * Support code for hammer2_vfs_mount().  Read, verify, and install the volume
2339  * header into the HMP
2340  *
2341  * XXX read four volhdrs and use the one with the highest TID whos CRC
2342  *     matches.
2343  *
2344  * XXX check iCRCs.
2345  *
2346  * XXX For filesystems w/ less than 4 volhdrs, make sure to not write to
2347  *     nonexistant locations.
2348  *
2349  * XXX Record selected volhdr and ring updates to each of 4 volhdrs
2350  */
2351 static
2352 int
2353 hammer2_install_volume_header(hammer2_dev_t *hmp)
2354 {
2355         hammer2_volume_data_t *vd;
2356         struct buf *bp;
2357         hammer2_crc32_t crc0, crc, bcrc0, bcrc;
2358         int error_reported;
2359         int error;
2360         int valid;
2361         int i;
2362
2363         error_reported = 0;
2364         error = 0;
2365         valid = 0;
2366         bp = NULL;
2367
2368         /*
2369          * There are up to 4 copies of the volume header (syncs iterate
2370          * between them so there is no single master).  We don't trust the
2371          * volu_size field so we don't know precisely how large the filesystem
2372          * is, so depend on the OS to return an error if we go beyond the
2373          * block device's EOF.
2374          */
2375         for (i = 0; i < HAMMER2_NUM_VOLHDRS; i++) {
2376                 error = bread(hmp->devvp, i * HAMMER2_ZONE_BYTES64,
2377                               HAMMER2_VOLUME_BYTES, &bp);
2378                 if (error) {
2379                         brelse(bp);
2380                         bp = NULL;
2381                         continue;
2382                 }
2383
2384                 vd = (struct hammer2_volume_data *) bp->b_data;
2385                 if ((vd->magic != HAMMER2_VOLUME_ID_HBO) &&
2386                     (vd->magic != HAMMER2_VOLUME_ID_ABO)) {
2387                         brelse(bp);
2388                         bp = NULL;
2389                         continue;
2390                 }
2391
2392                 if (vd->magic == HAMMER2_VOLUME_ID_ABO) {
2393                         /* XXX: Reversed-endianness filesystem */
2394                         kprintf("hammer2: reverse-endian filesystem detected");
2395                         brelse(bp);
2396                         bp = NULL;
2397                         continue;
2398                 }
2399
2400                 crc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT0];
2401                 crc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC0_OFF,
2402                                       HAMMER2_VOLUME_ICRC0_SIZE);
2403                 bcrc = vd->icrc_sects[HAMMER2_VOL_ICRC_SECT1];
2404                 bcrc0 = hammer2_icrc32(bp->b_data + HAMMER2_VOLUME_ICRC1_OFF,
2405                                        HAMMER2_VOLUME_ICRC1_SIZE);
2406                 if ((crc0 != crc) || (bcrc0 != bcrc)) {
2407                         kprintf("hammer2 volume header crc "
2408                                 "mismatch copy #%d %08x/%08x\n",
2409                                 i, crc0, crc);
2410                         error_reported = 1;
2411                         brelse(bp);
2412                         bp = NULL;
2413                         continue;
2414                 }
2415                 if (valid == 0 || hmp->voldata.mirror_tid < vd->mirror_tid) {
2416                         valid = 1;
2417                         hmp->voldata = *vd;
2418                         hmp->volhdrno = i;
2419                 }
2420                 brelse(bp);
2421                 bp = NULL;
2422         }
2423         if (valid) {
2424                 hmp->volsync = hmp->voldata;
2425                 error = 0;
2426                 if (error_reported || bootverbose || 1) { /* 1/DEBUG */
2427                         kprintf("hammer2: using volume header #%d\n",
2428                                 hmp->volhdrno);
2429                 }
2430         } else {
2431                 error = EINVAL;
2432                 kprintf("hammer2: no valid volume headers found!\n");
2433         }
2434         return (error);
2435 }
2436
2437 /*
2438  * This handles hysteresis on regular file flushes.  Because the BIOs are
2439  * routed to a thread it is possible for an excessive number to build up
2440  * and cause long front-end stalls long before the runningbuffspace limit
2441  * is hit, so we implement hammer2_flush_pipe to control the
2442  * hysteresis.
2443  *
2444  * This is a particular problem when compression is used.
2445  */
2446 void
2447 hammer2_lwinprog_ref(hammer2_pfs_t *pmp)
2448 {
2449         atomic_add_int(&pmp->count_lwinprog, 1);
2450 }
2451
2452 void
2453 hammer2_lwinprog_drop(hammer2_pfs_t *pmp)
2454 {
2455         int lwinprog;
2456
2457         lwinprog = atomic_fetchadd_int(&pmp->count_lwinprog, -1);
2458         if ((lwinprog & HAMMER2_LWINPROG_WAITING) &&
2459             (lwinprog & HAMMER2_LWINPROG_MASK) <= hammer2_flush_pipe * 2 / 3) {
2460                 atomic_clear_int(&pmp->count_lwinprog,
2461                                  HAMMER2_LWINPROG_WAITING);
2462                 wakeup(&pmp->count_lwinprog);
2463         }
2464         if ((lwinprog & HAMMER2_LWINPROG_WAITING0) &&
2465             (lwinprog & HAMMER2_LWINPROG_MASK) <= 0) {
2466                 atomic_clear_int(&pmp->count_lwinprog,
2467                                  HAMMER2_LWINPROG_WAITING0);
2468                 wakeup(&pmp->count_lwinprog);
2469         }
2470 }
2471
2472 void
2473 hammer2_lwinprog_wait(hammer2_pfs_t *pmp, int flush_pipe)
2474 {
2475         int lwinprog;
2476         int lwflag = (flush_pipe) ? HAMMER2_LWINPROG_WAITING :
2477                                     HAMMER2_LWINPROG_WAITING0;
2478
2479         for (;;) {
2480                 lwinprog = pmp->count_lwinprog;
2481                 cpu_ccfence();
2482                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2483                         break;
2484                 tsleep_interlock(&pmp->count_lwinprog, 0);
2485                 atomic_set_int(&pmp->count_lwinprog, lwflag);
2486                 lwinprog = pmp->count_lwinprog;
2487                 if ((lwinprog & HAMMER2_LWINPROG_MASK) <= flush_pipe)
2488                         break;
2489                 tsleep(&pmp->count_lwinprog, PINTERLOCKED, "h2wpipe", hz);
2490         }
2491 }
2492
2493 /*
2494  * Manage excessive memory resource use for chain and related
2495  * structures.
2496  */
2497 void
2498 hammer2_pfs_memory_wait(hammer2_pfs_t *pmp)
2499 {
2500         uint32_t waiting;
2501         uint32_t count;
2502         uint32_t limit;
2503 #if 0
2504         static int zzticks;
2505 #endif
2506
2507         /*
2508          * Atomic check condition and wait.  Also do an early speedup of
2509          * the syncer to try to avoid hitting the wait.
2510          */
2511         for (;;) {
2512                 waiting = pmp->inmem_dirty_chains;
2513                 cpu_ccfence();
2514                 count = waiting & HAMMER2_DIRTYCHAIN_MASK;
2515
2516                 limit = pmp->mp->mnt_nvnodelistsize / 10;
2517                 if (limit < hammer2_limit_dirty_chains)
2518                         limit = hammer2_limit_dirty_chains;
2519                 if (limit < 1000)
2520                         limit = 1000;
2521
2522 #if 0
2523                 if ((int)(ticks - zzticks) > hz) {
2524                         zzticks = ticks;
2525                         kprintf("count %ld %ld\n", count, limit);
2526                 }
2527 #endif
2528
2529                 /*
2530                  * Block if there are too many dirty chains present, wait
2531                  * for the flush to clean some out.
2532                  */
2533                 if (count > limit) {
2534                         tsleep_interlock(&pmp->inmem_dirty_chains, 0);
2535                         if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2536                                                waiting,
2537                                        waiting | HAMMER2_DIRTYCHAIN_WAITING)) {
2538                                 speedup_syncer(pmp->mp);
2539                                 tsleep(&pmp->inmem_dirty_chains, PINTERLOCKED,
2540                                        "chnmem", hz);
2541                         }
2542                         continue;       /* loop on success or fail */
2543                 }
2544
2545                 /*
2546                  * Try to start an early flush before we are forced to block.
2547                  */
2548                 if (count > limit * 7 / 10)
2549                         speedup_syncer(pmp->mp);
2550                 break;
2551         }
2552 }
2553
2554 void
2555 hammer2_pfs_memory_inc(hammer2_pfs_t *pmp)
2556 {
2557         if (pmp) {
2558                 atomic_add_int(&pmp->inmem_dirty_chains, 1);
2559         }
2560 }
2561
2562 void
2563 hammer2_pfs_memory_wakeup(hammer2_pfs_t *pmp)
2564 {
2565         uint32_t waiting;
2566
2567         if (pmp == NULL)
2568                 return;
2569
2570         for (;;) {
2571                 waiting = pmp->inmem_dirty_chains;
2572                 cpu_ccfence();
2573                 if (atomic_cmpset_int(&pmp->inmem_dirty_chains,
2574                                        waiting,
2575                                        (waiting - 1) &
2576                                         ~HAMMER2_DIRTYCHAIN_WAITING)) {
2577                         break;
2578                 }
2579         }
2580
2581         if (waiting & HAMMER2_DIRTYCHAIN_WAITING)
2582                 wakeup(&pmp->inmem_dirty_chains);
2583 }
2584
2585 /*
2586  * Debugging
2587  */
2588 void
2589 hammer2_dump_chain(hammer2_chain_t *chain, int tab, int *countp, char pfx)
2590 {
2591         hammer2_chain_t *scan;
2592         hammer2_chain_t *parent;
2593
2594         --*countp;
2595         if (*countp == 0) {
2596                 kprintf("%*.*s...\n", tab, tab, "");
2597                 return;
2598         }
2599         if (*countp < 0)
2600                 return;
2601         kprintf("%*.*s%c-chain %p.%d %016jx/%d mir=%016jx\n",
2602                 tab, tab, "", pfx,
2603                 chain, chain->bref.type,
2604                 chain->bref.key, chain->bref.keybits,
2605                 chain->bref.mirror_tid);
2606
2607         kprintf("%*.*s      [%08x] (%s) refs=%d",
2608                 tab, tab, "",
2609                 chain->flags,
2610                 ((chain->bref.type == HAMMER2_BREF_TYPE_INODE &&
2611                 chain->data) ?  (char *)chain->data->ipdata.filename : "?"),
2612                 chain->refs);
2613
2614         parent = chain->parent;
2615         if (parent)
2616                 kprintf("\n%*.*s      p=%p [pflags %08x prefs %d",
2617                         tab, tab, "",
2618                         parent, parent->flags, parent->refs);
2619         if (RB_EMPTY(&chain->core.rbtree)) {
2620                 kprintf("\n");
2621         } else {
2622                 kprintf(" {\n");
2623                 RB_FOREACH(scan, hammer2_chain_tree, &chain->core.rbtree)
2624                         hammer2_dump_chain(scan, tab + 4, countp, 'a');
2625                 if (chain->bref.type == HAMMER2_BREF_TYPE_INODE && chain->data)
2626                         kprintf("%*.*s}(%s)\n", tab, tab, "",
2627                                 chain->data->ipdata.filename);
2628                 else
2629                         kprintf("%*.*s}\n", tab, tab, "");
2630         }
2631 }